性能测试

.gitignore

@@ -5,6 +5,8 @@
 *.a
 
 /ebpf/libbpf-bootstrap
+/doc
+/test-redis/results
 
 kvstore
 testcase

Makefile

@@ -2,7 +2,7 @@
 CC = gcc
 CFLAGS = -g -DJEMALLOC_NO_DEMANGLE
 
-NET_SRCS 	= 	ntyco.c proactor.c reactor.c kvstore.c
+NET_SRCS 	= 	ntyco.c proactor.c reactor.c kvstore.c network/chainbuffer.c
 KV_SRCS 	=   kvs_array_bin.c kvs_rbtree_bin.c kvs_hash_bin.c kvs_rw_tools.c kvs_protocol_resp.c  kvs_slave.c replica_shm.c
 MEM_SRCS	= 	./memory/mempool.c ./memory/alloc_dispatch.c 
 COMMON_SRCS	= 	./common/config.c ./diskuring/diskuring.c
@@ -21,12 +21,14 @@ SUBDIR = ./NtyCo/
 
 TEST_REDIS = ./test-redis/testcase
 TEST_REDIS_SRCS = ./test-redis/test.c
+TEST_REDIS_BENCH = ./test-redis/bench
+TEST_REDIS_BENCH_SRCS = ./test-redis/bench.c
 TEST_REDIS_LDFLAGS = -lhiredis
 
 OBJS = $(SRCS:.c=.o)
 
 
-all: $(SUBDIR) $(TARGET) $(TEST_REDIS)
+all: $(SUBDIR) $(TARGET) $(TEST_REDIS) $(TEST_REDIS_BENCH)
 
 $(SUBDIR): ECHO
 	make -C $@
@@ -40,11 +42,14 @@ $(TARGET): $(OBJS)
 $(TEST_REDIS): $(TEST_REDIS_SRCS)
 	$(CC) -g -o $@ $^ $(TEST_REDIS_LDFLAGS)
 
+$(TEST_REDIS_BENCH): $(TEST_REDIS_BENCH_SRCS)
+	$(CC) -g -o $@ $^ $(TEST_REDIS_LDFLAGS)
+
 %.o: %.c
 	$(CC) $(CFLAGS) $(INC) -c $^ -g -o $@
 
 clean: clmem cldata
-	rm -rf $(OBJS) $(TARGET) $(TEST_REDIS)
+	rm -rf $(OBJS) $(TARGET) $(TEST_REDIS) $(TEST_REDIS_BENCH)
 
 clmem:
 	rm -rf mem_leak/*

Makefile.default

@@ -1,29 +0,0 @@
-
-
-CC = gcc
-FLAGS = -I ./NtyCo/core/ -L ./NtyCo/ -lntyco -luring
-TARGET = kvstore
-
-SRCS = kvstore.c ntyco.c proactor.c kvs_array.c kvs_rbtree.c
-# INC = -I ./NtyCo/core/
-# LIBS = -L ./NtyCo/ -lntyco -luring
-# FLAGS = -I ./NtyCo/core/ -L ./NtyCo/ -lntyco -luring
-
-OBJS = $(SRCS:.c=.o)
-TESTCASE = testcase 
-SUBDIR = ./NtyCo/
-
-all: $(SUBDIR) $(TARGET) # $(TESTCASE)
-
-$(SUBDIR): ECHO 
-	make -C $@
-
-ECHO:
-	@echo $(SUBDIR)
-
-$(TARGET): $(OBJS) 
-	$(CC) -o $@ $^ $(FLAGS)
-
-clean:
-	rm -rf kvstore *.o
-

README.md

@@ -1,20 +1,5 @@
 # 9.1 Kvstore
 
-## 需求
-1. ntyco需要作为kvstore的submodule,通过git clone一次下载。 **完成**。
-2. README需要包含编译步骤，测试方案与可行性，性能数据。  **完成**。
-3. 全量持久化保存数据集。 **BUG FIX**。
-4. 持久化的性能数据。   **完成**。
-5. 特殊字符，可以解决redis的resp协议。 **完成**。
-6. 实现配置文件，把日志级别，端口ip，主从模式，持久化方案。 **完成**。
-7. 持久化落盘用io_uring，加载配置文件用mmap。   **完成**。
-8. 主从同步的性能，开启与关闭性能做到5%?。
-9. 主从同步600w条,出现的coredump。     **完成**。
-10. 主从同步用ebpf实现。               **BUG FIX**。
-11. 内存池测试qps与虚拟内存，物理内存。 **完成**。
-12. 实现一个内存泄露检测组件。          **完成**。
-
-
 ## 环境安装与编译
 ```shell
 # xml
@@ -22,18 +7,73 @@ sudo apt install libxml2 libxml2-dev
 # hiredis client
 sudo apt install -y libhiredis-dev
 # bpftrace
-sudo apt install -y bpftrace
+sudo apt install -y bpftrace libelf libelf-dev clang
+# jemalloc
+sudo apt install libjemalloc-dev
 
 git clone git@gitlab.0voice.com:lianyiheng/9.1-kvstore.git
 cd 9.1-kvstore/
 git submodule update --init --recursive
-./init.sh
 make
 
 
 ```
 
-## 测试
+## REDIS 对比测试
+### 数据口径（2026-03-07 大 key 5轮复测）
+- 参数：`requests=1000000 pipeline=128 keyspace=1000000 value-size=256`
+- 轮次：每个策略 `5` 轮，按场景剔除 `1` 个异常轮次后取 `4` 轮均值
+- kvstore 源数据：`test-redis/results/hash_bench_fair_summary_20260307_062549.csv`
+- redis 源数据：`test-redis/results/redis_bigkey_summary_20260307_063158.csv`
+- 去异常结果：
+  - `test-redis/results/hash_bench_fair_trimmed_20260307_062549.csv`
+  - `test-redis/results/redis_bigkey_trimmed_20260307_063158.csv`
+
+### kvstore 协议开销（基线：nopersist，去异常后）
+| 策略 | set 均值QPS | set 均值us/op | get 均值QPS | get 均值us/op |
+|---|---:|---:|---:|---:|
+| nopersist | 150084.25 | 6.67 | 163760.25 | 6.11 |
+| persist_no | 140206.50 | 7.13 | 164554.25 | 6.08 |
+| persist_everysec | 133105.75 | 7.52 | 163358.25 | 6.12 |
+
+| 相对 nopersist 的开销 | set QPS 变化 | set us/op 变化 | get QPS 变化 | get us/op 变化 |
+|---|---:|---:|---:|---:|
+| persist_no | -6.58% | +6.90% | +0.48% | -0.49% |
+| persist_everysec | -11.31% | +12.74% | -0.25% | +0.16% |
+
+### Redis 协议开销（基线：none，去异常后）
+| 策略 | set 均值QPS | set 均值us/op | get 均值QPS | get 均值us/op |
+|---|---:|---:|---:|---:|
+| none | 207470.00 | 4.82 | 217642.50 | 4.59 |
+| aof_no | 143399.50 | 6.98 | 214023.50 | 4.68 |
+| aof_everysec | 141893.00 | 7.05 | 208713.75 | 4.79 |
+
+| 相对 none 的开销 | set QPS 变化 | set us/op 变化 | get QPS 变化 | get us/op 变化 |
+|---|---:|---:|---:|---:|
+| aof_no | -30.88% | +44.81% | -1.66% | +1.96% |
+| aof_everysec | -31.61% | +46.27% | -4.10% | +4.36% |
+
+结论：本轮数据下，持久化开销仍主要体现在 `set`，`get` 相对更稳；Redis 的 AOF 写路径开销显著高于无持久化。
+
+## 调用开销
+### gprof Flat Profile（Top 12，按 self time）
+
+| 排名 | 函数 | self time % | self seconds | calls |
+|---:|---|---:|---:|---:|
+| 1 | `rbtree_node_get_key` | 58.34 | 0.56 | 103209091 |
+| 2 | `rbtree_search` | 7.29 | 0.07 | 1874362 |
+| 3 | `kvs_keycmp` | 5.73 | 0.06 | 74287566 |
+| 4 | `ascii_casecmp` | 3.13 | 0.03 | 21490996 |
+| 5 | `task_init` | 3.13 | 0.03 | 1397974 |
+| 6 | `mp_page_create` | 3.13 | 0.03 | 23556 |
+| 7 | `need` | 2.08 | 0.02 | 14042759 |
+| 8 | `parse_i64` | 2.08 | 0.02 | 7029783 |
+| 9 | `mp_page_alloc` | 2.08 | 0.02 | 5122487 |
+| 10 | `rbtree_node_size` | 2.08 | 0.02 | 1860739 |
+| 11 | `submit_write` | 2.08 | 0.02 | 1394599 |
+| 12 | `rbtree_insert` | 2.08 | 0.02 | 926531 |
+
+## 其他测试
 ### 测试1：性能测试
 测试条件：
 1. 不启用持久化。
@@ -47,57 +87,6 @@ make
 #### 内存分配: malloc
 ```bash
 lian@ubuntu:~/share/9.1-kvstore$ ./test-redis/testcase 192.168.10.129 8888 3
-Connected to 192.168.10.129:8888
-BATCH (N=3000000) --> time_used=3294 ms, qps=910746
-BATCH (N=3000000) --> time_used=3501 ms, qps=856898
-BATCH (N=3000000) --> time_used=3457 ms, qps=867804
-BATCH (N=3000000) --> time_used=3351 ms, qps=895255
-BATCH (N=3000000) --> time_used=3320 ms, qps=903614
-BATCH (N=3000000) --> time_used=3551 ms, qps=844832
-BATCH (N=3000000) --> time_used=3354 ms, qps=894454
-BATCH (N=3000000) --> time_used=3475 ms, qps=863309
-BATCH (N=3000000) --> time_used=3404 ms, qps=881316
-BATCH (N=3000000) --> time_used=3460 ms, qps=867052
-BATCH (N=3000000) --> time_used=3392 ms, qps=884433
-BATCH (N=3000000) --> time_used=3427 ms, qps=875401
-BATCH (N=3000000) --> time_used=3441 ms, qps=871839
-BATCH (N=3000000) --> time_used=3471 ms, qps=864304
-BATCH (N=3000000) --> time_used=3354 ms, qps=894454
-BATCH (N=3000000) --> time_used=3447 ms, qps=870322
-BATCH (N=3000000) --> time_used=3364 ms, qps=891795
-BATCH (N=3000000) --> time_used=3200 ms, qps=937500
-BATCH (N=3000000) --> time_used=3159 ms, qps=949667
-BATCH (N=3000000) --> time_used=3482 ms, qps=861573
-BATCH (N=3000000) --> time_used=3474 ms, qps=863557
-BATCH (N=3000000) --> time_used=3591 ms, qps=835421
-BATCH (N=3000000) --> time_used=3466 ms, qps=865551
-BATCH (N=3000000) --> time_used=3425 ms, qps=875912
-BATCH (N=3000000) --> time_used=3346 ms, qps=896592
-BATCH (N=3000000) --> time_used=3532 ms, qps=849377
-BATCH (N=3000000) --> time_used=3471 ms, qps=864304
-BATCH (N=3000000) --> time_used=3616 ms, qps=829646
-BATCH (N=3000000) --> time_used=3403 ms, qps=881575
-BATCH (N=3000000) --> time_used=3426 ms, qps=875656
-BATCH (N=3000000) --> time_used=3493 ms, qps=858860
-BATCH (N=3000000) --> time_used=3411 ms, qps=879507
-BATCH (N=3000000) --> time_used=3422 ms, qps=876680
-BATCH (N=3000000) --> time_used=3556 ms, qps=843644
-BATCH (N=3000000) --> time_used=3285 ms, qps=913242
-BATCH (N=3000000) --> time_used=3486 ms, qps=860585
-BATCH (N=3000000) --> time_used=3427 ms, qps=875401
-BATCH (N=3000000) --> time_used=3563 ms, qps=841987
-BATCH (N=3000000) --> time_used=3304 ms, qps=907990
-BATCH (N=3000000) --> time_used=3582 ms, qps=837520
-BATCH (N=3000000) --> time_used=3468 ms, qps=865051
-BATCH (N=3000000) --> time_used=3360 ms, qps=892857
-BATCH (N=3000000) --> time_used=3426 ms, qps=875656
-BATCH (N=3000000) --> time_used=3186 ms, qps=941619
-BATCH (N=3000000) --> time_used=3251 ms, qps=922792
-BATCH (N=3000000) --> time_used=3400 ms, qps=882352
-BATCH (N=3000000) --> time_used=3446 ms, qps=870574
-BATCH (N=3000000) --> time_used=3302 ms, qps=908540
-BATCH (N=3000000) --> time_used=3072 ms, qps=976562
-BATCH (N=3000000) --> time_used=3458 ms, qps=867553
 average qps:880462
 ALL TESTS PASSED.
 ```
@@ -105,57 +94,6 @@ ALL TESTS PASSED.
 #### 内存分配: 自实现内存池
 ```bash
 lian@ubuntu:~/share/9.1-kvstore$ ./test-redis/testcase 192.168.10.129 8888 3
-Connected to 192.168.10.129:8888
-BATCH (N=3000000) --> time_used=3241 ms, qps=925640
-BATCH (N=3000000) --> time_used=3047 ms, qps=984574
-BATCH (N=3000000) --> time_used=3085 ms, qps=972447
-BATCH (N=3000000) --> time_used=3119 ms, qps=961846
-BATCH (N=3000000) --> time_used=3104 ms, qps=966494
-BATCH (N=3000000) --> time_used=3163 ms, qps=948466
-BATCH (N=3000000) --> time_used=3033 ms, qps=989119
-BATCH (N=3000000) --> time_used=3170 ms, qps=946372
-BATCH (N=3000000) --> time_used=3299 ms, qps=909366
-BATCH (N=3000000) --> time_used=3272 ms, qps=916870
-BATCH (N=3000000) --> time_used=3294 ms, qps=910746
-BATCH (N=3000000) --> time_used=3182 ms, qps=942803
-BATCH (N=3000000) --> time_used=3190 ms, qps=940438
-BATCH (N=3000000) --> time_used=3493 ms, qps=858860
-BATCH (N=3000000) --> time_used=3111 ms, qps=964320
-BATCH (N=3000000) --> time_used=3220 ms, qps=931677
-BATCH (N=3000000) --> time_used=3067 ms, qps=978154
-BATCH (N=3000000) --> time_used=3345 ms, qps=896860
-BATCH (N=3000000) --> time_used=3381 ms, qps=887311
-BATCH (N=3000000) --> time_used=3416 ms, qps=878220
-BATCH (N=3000000) --> time_used=3192 ms, qps=939849
-BATCH (N=3000000) --> time_used=3085 ms, qps=972447
-BATCH (N=3000000) --> time_used=3150 ms, qps=952380
-BATCH (N=3000000) --> time_used=3296 ms, qps=910194
-BATCH (N=3000000) --> time_used=3001 ms, qps=999666
-BATCH (N=3000000) --> time_used=3143 ms, qps=954502
-BATCH (N=3000000) --> time_used=3111 ms, qps=964320
-BATCH (N=3000000) --> time_used=3123 ms, qps=960614
-BATCH (N=3000000) --> time_used=3257 ms, qps=921093
-BATCH (N=3000000) --> time_used=3037 ms, qps=987816
-BATCH (N=3000000) --> time_used=3135 ms, qps=956937
-BATCH (N=3000000) --> time_used=3124 ms, qps=960307
-BATCH (N=3000000) --> time_used=3276 ms, qps=915750
-BATCH (N=3000000) --> time_used=3058 ms, qps=981033
-BATCH (N=3000000) --> time_used=3024 ms, qps=992063
-BATCH (N=3000000) --> time_used=3224 ms, qps=930521
-BATCH (N=3000000) --> time_used=3235 ms, qps=927357
-BATCH (N=3000000) --> time_used=3334 ms, qps=899820
-BATCH (N=3000000) --> time_used=3427 ms, qps=875401
-BATCH (N=3000000) --> time_used=3218 ms, qps=932256
-BATCH (N=3000000) --> time_used=3191 ms, qps=940144
-BATCH (N=3000000) --> time_used=3179 ms, qps=943692
-BATCH (N=3000000) --> time_used=3104 ms, qps=966494
-BATCH (N=3000000) --> time_used=3202 ms, qps=936914
-BATCH (N=3000000) --> time_used=3184 ms, qps=942211
-BATCH (N=3000000) --> time_used=3000 ms, qps=1000000
-BATCH (N=3000000) --> time_used=3280 ms, qps=914634
-BATCH (N=3000000) --> time_used=3141 ms, qps=955109
-BATCH (N=3000000) --> time_used=3198 ms, qps=938086
-BATCH (N=3000000) --> time_used=3126 ms, qps=959692
 average qps:942837
 ALL TESTS PASSED.
 ```
@@ -163,57 +101,6 @@ ALL TESTS PASSED.
 #### 内存分配：jemalloc
 ```shell
 lian@ubuntu:~/share/9.1-kvstore$ ./test-redis/testcase 192.168.10.129 8888 3
-Connected to 192.168.10.129:8888
-BATCH (N=3000000) --> time_used=3511 ms, qps=854457
-BATCH (N=3000000) --> time_used=3280 ms, qps=914634
-BATCH (N=3000000) --> time_used=3603 ms, qps=832639
-BATCH (N=3000000) --> time_used=3418 ms, qps=877706
-BATCH (N=3000000) --> time_used=3353 ms, qps=894721
-BATCH (N=3000000) --> time_used=3435 ms, qps=873362
-BATCH (N=3000000) --> time_used=3250 ms, qps=923076
-BATCH (N=3000000) --> time_used=3550 ms, qps=845070
-BATCH (N=3000000) --> time_used=3536 ms, qps=848416
-BATCH (N=3000000) --> time_used=3273 ms, qps=916590
-BATCH (N=3000000) --> time_used=3224 ms, qps=930521
-BATCH (N=3000000) --> time_used=3161 ms, qps=949066
-BATCH (N=3000000) --> time_used=3143 ms, qps=954502
-BATCH (N=3000000) --> time_used=3342 ms, qps=897666
-BATCH (N=3000000) --> time_used=3410 ms, qps=879765
-BATCH (N=3000000) --> time_used=3522 ms, qps=851788
-BATCH (N=3000000) --> time_used=3035 ms, qps=988467
-BATCH (N=3000000) --> time_used=3352 ms, qps=894988
-BATCH (N=3000000) --> time_used=3226 ms, qps=929944
-BATCH (N=3000000) --> time_used=3406 ms, qps=880798
-BATCH (N=3000000) --> time_used=3336 ms, qps=899280
-BATCH (N=3000000) --> time_used=3307 ms, qps=907166
-BATCH (N=3000000) --> time_used=3171 ms, qps=946073
-BATCH (N=3000000) --> time_used=3252 ms, qps=922509
-BATCH (N=3000000) --> time_used=3296 ms, qps=910194
-BATCH (N=3000000) --> time_used=3301 ms, qps=908815
-BATCH (N=3000000) --> time_used=3403 ms, qps=881575
-BATCH (N=3000000) --> time_used=3234 ms, qps=927643
-BATCH (N=3000000) --> time_used=3348 ms, qps=896057
-BATCH (N=3000000) --> time_used=3517 ms, qps=852999
-BATCH (N=3000000) --> time_used=3354 ms, qps=894454
-BATCH (N=3000000) --> time_used=3529 ms, qps=850099
-BATCH (N=3000000) --> time_used=3473 ms, qps=863806
-BATCH (N=3000000) --> time_used=3521 ms, qps=852030
-BATCH (N=3000000) --> time_used=3370 ms, qps=890207
-BATCH (N=3000000) --> time_used=3267 ms, qps=918273
-BATCH (N=3000000) --> time_used=3352 ms, qps=894988
-BATCH (N=3000000) --> time_used=3433 ms, qps=873871
-BATCH (N=3000000) --> time_used=3374 ms, qps=889152
-BATCH (N=3000000) --> time_used=3360 ms, qps=892857
-BATCH (N=3000000) --> time_used=3463 ms, qps=866300
-BATCH (N=3000000) --> time_used=3499 ms, qps=857387
-BATCH (N=3000000) --> time_used=3294 ms, qps=910746
-BATCH (N=3000000) --> time_used=3311 ms, qps=906070
-BATCH (N=3000000) --> time_used=3443 ms, qps=871333
-BATCH (N=3000000) --> time_used=3381 ms, qps=887311
-BATCH (N=3000000) --> time_used=3422 ms, qps=876680
-BATCH (N=3000000) --> time_used=3421 ms, qps=876936
-BATCH (N=3000000) --> time_used=3322 ms, qps=903070
-BATCH (N=3000000) --> time_used=3494 ms, qps=858614
 average qps:892493
 ALL TESTS PASSED.
 ```
@@ -230,17 +117,6 @@ ALL TESTS PASSED.
 
 ```shell
 lian@ubuntu:~/share/9.1-kvstore$ ./test-redis/testcase 192.168.10.129 8888 4
-Connected to 192.168.10.129:8888
-BATCH (N=3000000) --> time_used=3500 ms, qps=857142
-BATCH (N=3000000) --> time_used=3322 ms, qps=903070
-BATCH (N=3000000) --> time_used=3424 ms, qps=876168
-BATCH (N=3000000) --> time_used=3483 ms, qps=861326
-BATCH (N=3000000) --> time_used=3421 ms, qps=876936
-BATCH (N=3000000) --> time_used=3519 ms, qps=852514
-BATCH (N=3000000) --> time_used=3597 ms, qps=834028
-BATCH (N=3000000) --> time_used=3504 ms, qps=856164
-BATCH (N=3000000) --> time_used=3281 ms, qps=914355
-BATCH (N=3000000) --> time_used=3446 ms, qps=870574
 average qps:870227
 ALL TESTS PASSED.
 ```
@@ -264,9 +140,9 @@ BATCH (N=9000000) --> time_used=10033 ms, qps=1794079
 VIRT    208M
 RES     155M
 ```
-![alt text](image11.png)
+![alt text](https://disk.0voice.com/p/wl)
-![alt text](image12.png)
+![alt text](https://disk.0voice.com/p/wm)
-![alt text](image13.png)
+![alt text](https://disk.0voice.com/p/wO)
 
 #### jemalloc
 ```shell
@@ -285,9 +161,9 @@ BATCH (N=9000000) --> time_used=9353 ms, qps=1924516
 VIRT    356M
 RES     119M
 ```
-![alt text](image11.png)
+![alt text](https://disk.0voice.com/p/wl)
-![alt text](image22.png)
+![alt text](https://disk.0voice.com/p/wP)
-![alt text](image23.png)
+![alt text](https://disk.0voice.com/p/wQ)
 
 #### mypool
 ```shell
@@ -306,9 +182,9 @@ BATCH (N=3000000) --> time_used=3022 ms, qps=1985440
 VIRT    122M
 RES     71492
 ```
-![alt text](image31.png)
+![alt text](https://disk.0voice.com/p/wR)
-![alt text](image32.png)
+![alt text](https://disk.0voice.com/p/wn)
-![alt text](image33.png)
+![alt text](https://disk.0voice.com/p/wo)
 
 ### 测试4：主从同步
 测试条件：
@@ -340,21 +216,170 @@ average qps:777838
 ALL TESTS PASSED.
 ```
 
-### 面试题
+## 项目收获
-1. 为什么会实现kvstore，使用场景在哪里？
+#### reactor网络模型，用户态网络缓冲区的写法。
-2. reactor, ntyco, io_uring的三种网络模型的性能差异？
-3. 多线程的kvstore该如何改进？
-4. 私有协议如何设计会更加安全可靠？
-5. 协议改进以后，对已有的代码有哪些改变？
-6. kv引擎实现了哪些？
-7. 每个kv引擎的使用场景，以及性能差异？
-8. 测试用例如何实现？并且保证代码覆盖率超过90%
-9. 网络并发量如何？qps如何？
-10. 能够跟哪些系统交互使用？
 
+#### 特殊字符串支持的引擎层数据结构设计，支持\0作为键值存储。
 
-### 架构设计
+1. 长度前缀 + 内容的 binary-safe 字符串表示，支持 \0 作为普通字符。
-![image](https://disk.0voice.com/p/py)
+
+#### 实现RESP协议的服务端协议解析。
+
+1. 解析流程：
+   1. 内核 拷贝到 用户态缓冲区
+   2. 用户态缓冲区 就地解析 bulk-string 为执行引擎可以看得懂的结构体
+   3. 执行引擎 拷贝 结构体的内容，插入到底层存储结构中
+   4. 循环解析直到 用户态缓冲区为空
+2. 实现了 pipeline 支持：每次从 buffer 读到一个完整命令就交给应用层处理，应用层返回已消费字节数。如果 buffer 里有半包，连接层保留下次继续解析。
+
+#### 快照异步创建。
+
+1. 使用fork的Copy On Write机制，实现的异步快照创建，不会受到后续更新请求的影响
+
+#### 实现SPSC/SPMC，专门uring线程实现异步的增量、全量落盘操作。
+
+1. 熟悉了SPSC无锁队列的实现方案。
+   1. 无锁、cache friendly
+2. 流程：
+   1. 生产者组装task，丢给SPSC。
+   2. 消费者从中取出然后执行，置入destroy_queue，触发eventfd。
+   3. 生产者释放destroy_queue。
+3. 解决的问题：
+   1. iouring 由于 cqe 接收不及时导致的 task 丢失无法释放。
+      1. 通过背压解决，设置inflight上限。
+   2. 背压后，生产者速度 > 消费者速度，ringbuffer 满导致只能阻塞主线程或降低速度。
+      1. 通过修改SPSC为SPMC，构建落盘线程组实现，当task_queue满，触发扩容线程组。
+      2. 每个落盘线程私有一个SPSC，减少竞争。
+      3. 简易工作负载，生产者线程随机选取两个uring线程选取任务少的push。
+   3. 采用读写段采用轮询方案导致乒乓现象，性能下降。
+      1. MESI协议定义了缓存行的四种状态：Modified表示独占且已修改，Exclusive表示独占且未修改，Shared表示多核共享只读，Invalid表示缓存行无效。
+         - **关键点：** 但从S状态读没有什么开销，从I状态读则需要向其他CPU申请。
+         - **关键点：** 从E/M状态写也没有什么开销，从S状态写则需要广播invalidate并且等待ACK(50-100时钟周期)。
+         - 原子操作由于内存序的顺序性和可见性语义，也有额外开销（刷新invalidate queue、阻止指令重排）
+      2. 短临界区：分层退避。自旋-让出CPU时间片-较长时间休眠。允许生产者在期间无争用的写入一批数据，然后统一读。
+      3. 更通用的情况：事件驱动。用futex替代轮询。Fast Userspace Mutex。
+         1. **消费者：** 调用`syscall(SYS_futex, &futex_word, FUTEX_WAIT, old_val, &timeout, NULL, 0)` , 如果futex_word仍等于old_val，线程进入内核等待队列，真正睡眠，不占用CPU。
+         2. **生产者：** 调用`syscall(SYS_futex, &futex_word, FUTEX_WAKE, 1, NULL, NULL, 0)` 唤醒一个等待线程。
+4. 优化：生产者速度 > 消费者速度，则写端不需要关注读指针，或者极少关注读指针（缓冲区大小的1/2）次写入才考虑一次。
+
+#### 基于BinLog上OffSet的主从同步(已有数据+实时数据)设计。
+
+1. 初始化阶段：
+   1. master 持续将收到的更新请求+seq_id 落盘到本地 binlog 中。
+   2. slave 向 master 发起连接并且发送本地binlog中最大的seq_id 为 slave_seq_id。 
+2. 执行阶段：
+   1. master启动一个独立同步线程与 slave 构建连接。同步线程有两个阶段：
+      1. slave_seq_id < local_min_seq_id：master通过fork创建内存快照发送，且发送同时刻的local_max_seq_id。
+      2. slave_seq_id < local_max_seq_id：通过自实现同步协议批量发送binlog的seq，并且回包new_slave_seq_id。
+      3. slave_seq_id == local_max_seq_id：线程休眠。
+   2. master收到新的请求的时候，会通过条件变量唤醒同步线程。
+
+#### 基于ebpf的实时数据同步设计。
+
+**基准性能**：Kvstore QPS ~90w（无持久化/同步）。
+
+ebpf对程序的影响要考虑如下方面：
+
+1. eBPF 探针的触发频率（上下文切换）
+
+2. 数据拷贝方式 
 
    
 
+**用户态探针 (Uprobe) 的开销**，
+
+1. **逐条命令探测**。QPS大概在 25w左右。**原因**：Uprobe 基于断点指令（int3）实现，用户态 → 异常 → 内核 → eBPF → 返回用户态，高频触发会导致 CPU 流水线停顿严重。
+2. **批处理探测**。QPS大概在85w左右。大幅减少了上下文切换次数，平摊了中断开销。
+
+**内核态探针 (Tracepoint/Kprobe) 的开销**
+
+1. **Tracepoint (sys_exit_recvfrom)等**
+   1. 纯探测：QPS ~85w。
+   2. 带数据拷贝 (read_user)：QPS 降至 ~70w。**原因**：**bpf_probe_read_user** 是一个非常重的 helper，开销远大于一次 memcpy。
+2. **Kprobe (tcp_rcv_established)等**
+   1. 纯探测：QPS ~86w。
+   2. 带数据拷贝 (read_kernel)：QPS ~83w。**原因**：比**bpf_probe_read_user**轻得多。
+   3. 问题：此时数据位于 TCP 协议栈底层，可能存在分片（Fragment）、乱序或非线性存储（Paged SKB），需要处理复杂的数据重组逻辑。
+
+工作流程：
+
+1. **内核态捕获**：	
+   1. eBPF 程序挂载于内核网络路径（ TC 或 TCP 层），拦截流量。
+   2. 使用 bpf_probe_read_kernel 或 bpf_skb_load_bytes 高效提取数据载荷。
+   3. 通过 bpf_ringbuf_submit 将数据写入环形缓冲区。
+2. **用户态转发**：
+   1. 独立进程消费 RingBuffer。
+   2. 将数据暂存入本地队列，平滑突发流量，防止 RingBuffer 溢出导致的数据丢失。
+3. **状态机控制**：
+   1. **SYNC 阶段**：探测到 __ssync，识别 Slave 连接信息（IP/Port），标记状态为“同步中”。
+   2. **READY 阶段**：探测到 __sready，标志全量同步完成。
+   3. **实时转发**：Agent 启动独立线程，消耗 Pending Queue，将增量数据通过标准 TCP 发送给 Slave。
+
+##### TC 与 XDP
+
+网络栈
+
+```
+-> [ 网卡 (NIC) ] 
+-> [ XDP (eXpress Data Path) ]  <--- xdp 处理原始帧 no skb
+-> [ sk_buff 分配 ]
+-> [ TC (Traffic Control) Ingress ]  | tc 可操作 on skb
+-> [ Netfilter (PREROUTING) ]  
+-> [ IP  协议栈 ]  | ip_rcv -> ip_local_deliver
+-> [ TCP 协议栈 ]  | tcp_v4_rcv -> tcp_rcv_established -> tcp_data_queue 
+producer skb -> sk->sk_receive_queue
+____________________________________________________________________
+consumer sk->sk_receive_queue
+-> [ Socket Layer ]	 | tcp_recvmsg 拷贝到内存
+-> [ Syscall Interface ]  | sys_exit_recvfrom
+-> [ 用户态应用 (Kvstore) ]
+```
+
+XDP 是网卡驱动层的探点，操作系统刚刚收到数据包（DMA读入ringbuffer，CRC校验），还没有分配sk_buffer。数据形态是**裸的以太网帧** 。
+
+TC 是在 sk_buff 分配之后，IP 协议栈处理之前的探测点。数据形态是__sk_buffer。
+
+TCP协议栈是分界点，
+
+#### 内存泄露探测功能，实现热插拔。
+
+方案1 基于bpf
+
+1. 通过预定义宏__FILE__等封装一个内存分配宏定义，向真正的分配函数传递代码位置等信息。
+2. 通过bpf探测内存分配的地址、大小、文件、代码位置、函数等信息，并且记录。
+3. 通过bpf探测内存释放的指针信息，然后释放。
+4. 打印最终剩余的内存地址。
+5. 缺点：bpf 探测 malloc 等对性能的影响非常的大。
+
+方案2 基于全局变量启用的代码内置探测工具，在网络层接收启用/关闭探测工具的请求。
+
+1. 分配时打开一个文件，关闭时删除。
+2. 分配时使用kv保存，删除时删除k。
+
+都对性能和内存有很大的影响，不建议长时间启用。
+
+#### 实现支持分配可变长度内存块的内存池。
+
+1. 熟悉了glibc 的 ptmalloc的底层操作。默认阈值约 128KB，且会根据分配行为动态调整。
+
+   1. '<= 默认阈值 通过 brk/sbrk 扩展连续堆，堆里维护了 chunk 结构。
+   2. '> 默认阈值 的块用 mmap 独立申请，释放时 munmap。
+   3. 每个 chunk 头部存大小信息（通常 8~16 字节），用户拿到的是 chunk + 头部后的地址。
+   4. 空闲 chunk 按大小分到多个 bin（tcache、fastbin、small bin、large bin 等），fastbin 和 tcache 为了速度不合并相邻空闲块。
+   5. 缺点：
+      1. 分配路径有较多分支判断和链表遍历，不是严格 O(1)。
+      2. 小块故意不合并（fastbin/tcache）会导致外部碎片。
+      3. 长期运行内存利用率下降。
+
+2. 内存池实现：
+
+   1. 内存池有多个桶，桶中存储固定大小的块。分配/释放均为 O(1)
+   2. 以 huge page 为单位向系统申请内存并切分为固定块。
+   3. 当页内块全部释放时整页归还系统，显著降低长期碎片。
+   4. 通过地址对齐 O(1) 反推出页头元信息（所属桶、空闲计数）。
+   5. malloc通常向上对齐，桶对应的存储大小可以根据不同系统设定，减少内部碎片。
+   6. 使用 bitmap + freelist 防 double free 且无额外查找开销。
+   7. 每线程独立内存池，相对malloc更少的锁竞争。
+   8. 大块分配自动退化为 malloc 处理。
+
+   相比 ptmalloc，该设计消除了外部碎片，降低了系统调用次数，并在多线程场景下显著提升分配性能与内存利用率。

common/config.c

@@ -56,6 +56,7 @@ static void set_default_config(AppConfig *cfg)
     cfg->mode        = MODE_MASTER;
     cfg->master_port = 8888;
     cfg->persistence = PERSIST_NONE;
+    cfg->oplog_sync_mode = OPLOG_SYNC_NONE;
     cfg->allocator   = ALLOC_JEMALLOC;
     cfg->leak_mode   = MEMLEAK_DETECT_OFF;
     cfg->replica_mode = REPLICA_DISABLE;
@@ -87,6 +88,16 @@ static void parse_persistence(const char *s, PersistenceType *out)
     else if (!strcasecmp(s, "none")) *out = PERSIST_NONE;
 }
 
+static void parse_oplog_sync_mode(const char *s, OplogSyncMode *out)
+{
+    if (!s || !out) return;
+    if (!strcasecmp(s, "every_sec") || !strcasecmp(s, "everysec")) {
+        *out = OPLOG_SYNC_EVERY_SEC;
+    } else if (!strcasecmp(s, "none")) {
+        *out = OPLOG_SYNC_NONE;
+    }
+}
+
 static void parse_allocator(const char *s, AllocatorType *out)
 {
     if (!s || !out) return;
@@ -188,6 +199,15 @@ const char *persistence_to_string(PersistenceType p)
     }
 }
 
+const char *oplog_sync_mode_to_string(OplogSyncMode m)
+{
+    switch (m) {
+    case OPLOG_SYNC_NONE:      return "none";
+    case OPLOG_SYNC_EVERY_SEC: return "every_sec";
+    default:                   return "unknown";
+    }
+}
+
 const char *allocator_to_string(AllocatorType a)
 {
     switch (a) {
@@ -337,6 +357,15 @@ void persist_load(xmlNodePtr *root, AppConfig *out_cfg){
             }
         }
 
+        xmlNodePtr oplog_sync_node = find_child(pers, "oplog_sync");
+        if (oplog_sync_node) {
+            xmlChar *txt = xmlNodeGetContent(oplog_sync_node);
+            if (txt) {
+                parse_oplog_sync_mode((char *)txt, &out_cfg->oplog_sync_mode);
+                xmlFree(txt);
+            }
+        }
+
         xmlNodePtr array_node = find_child(pers, "array");
         if (array_node) {
             xmlChar *txt = xmlNodeGetContent(array_node);
@@ -413,10 +442,9 @@ int config_load(const char *filename, AppConfig *out_cfg)
 
     /*
      * 用 xmlReadMemory 从内存解析。
-     * - "UTF-8"：你原来指定了 UTF-8；如果希望自动探测，可以传 NULL。
+     * - "UTF-8"：指定了 UTF-8；如果希望自动探测，可以传 NULL。
      * - XML_PARSE_NONET：禁用网络访问（防 XXE/外部实体拉取）
      * - XML_PARSE_NOBLANKS：保持你原来的行为
-     * 你也可以加 XML_PARSE_NOERROR | XML_PARSE_NOWARNING 减少噪音，但调试阶段不建议。
      */
     int parse_opts = XML_PARSE_NOBLANKS | XML_PARSE_NONET;
 

common/config.h

@@ -20,6 +20,11 @@ typedef enum {
     PERSIST_NONE
 } PersistenceType;
 
+typedef enum {
+    OPLOG_SYNC_NONE,
+    OPLOG_SYNC_EVERY_SEC
+} OplogSyncMode;
+
 typedef enum {
     REPLICA_DISABLE,
     REPLICA_ENABLE
@@ -53,6 +58,7 @@ typedef struct {
     char            array_file[256];
     char            rbtree_file[256];
     char            hash_file[256];
+    OplogSyncMode   oplog_sync_mode;
 
     AllocatorType   allocator;
     MemLeakDetectMode  leak_mode;
@@ -68,6 +74,7 @@ int config_load(const char *filename, AppConfig *out_cfg);
 const char *log_level_to_string(LogLevel lvl);
 const char *server_mode_to_string(ServerMode mode);
 const char *persistence_to_string(PersistenceType p);
+const char *oplog_sync_mode_to_string(OplogSyncMode m);
 const char *allocator_to_string(AllocatorType a);
 const char *leakage_to_string(MemLeakDetectMode a);
 const char *replica_to_string(ReplicaMode a);

config/config.xml

@@ -1,34 +1,35 @@
-<?xml version="1.0" encoding="UTF-8"?>
+<?xml version='1.0' encoding='UTF-8'?>
 <config>
     <server>
-        <ip>192.168.10.129</ip>
+        <ip>127.0.0.1</ip>
         <port>8888</port>
-        <mode>master</mode>     <!-- master / slave -->
+        <mode>master</mode>     
 
-        <!-- 仅当 mode=slave 时使用 -->
+        
-        <replica>enable</replica>
+        <replica>disable</replica>
         <master>
-            <ip>192.168.10.129</ip>
+            <ip>192.168.220.134</ip>
             <port>8888</port>
         </master>
     </server>
 
     <log>
-        <level>INFO</level>     <!-- DEBUG / INFO / ERROR -->
+        <level>INFO</level>     
     </log>
 
     <persistence>
-        <type>none</type>       <!-- incremental / none -->
+        <type>incremental</type>        
-        <dir>data</dir>          <!-- 所有持久化文件所在目录 -->
+        <dir>data/persist_no_20260307_053744_r1_a1</dir>          
         
         <wal>kvs_oplog.db</wal>
+        <oplog_sync>none</oplog_sync> 
         <array>kvs_array.db</array>
         <rbtree>kvs_rbtree.db</rbtree>
         <hash>kvs_hash.db</hash>
     </persistence>
 
     <memory>
-        <allocator>mypool</allocator> <!-- jemalloc / malloc / mypool -->
+        <allocator>mypool</allocator> 
-        <leakage>disable</leakage>       <!-- enable/disable-->
+        <leakage>disable</leakage>       
     </memory>
 </config>

diskuring/diskuring.c.nothread

@@ -1,199 +0,0 @@
-#include "diskuring.h"
-#include "memory/alloc_dispatch.h"
-#include <poll.h>
-#include <sys/eventfd.h>
-
-void task_init(task_t *t)
-{
-    t->done = 0;
-    t->res = 0;
-    t->next = NULL;
-}
-
-void task_finish(task_t *t, int res)
-{
-    t->res = res;
-    t->done = 1;
-}
-
-void task_destroy(task_t *t)
-{
-    if (t->iovs) {
-        for (int i = 0; i < t->iovcnt; i++) {
-            if (t->iovs[i].iov_base) {
-                kvs_free(t->iovs[i].iov_base);
-            }
-        }
-        kvs_free(t->iovs);
-    }
-
-    kvs_free(t);
-}
-
-int iouring_init(iouring_ctx_t *ctx, unsigned entries)
-{
-    memset(ctx, 0, sizeof(*ctx));
-
-    struct io_uring_params params;
-    memset(&params, 0, sizeof(params));
-
-    // params.flags |= IORING_SETUP_CQSIZE;
-    // params.cq_entries = 256 * 1024;
-    // params.sq_entries = 128 * 1024;
-    int ret = io_uring_queue_init_params(entries, &ctx->ring, &params);
-    if (ret < 0) {
-        fprintf(stderr, "io_uring_queue_init_params failed: %d (%s)\n", 
-                ret, strerror(-ret));
-        return ret;
-    }
-    
-    unsigned cq_size = *ctx->ring.cq.kring_entries;
-    printf("Kernel CQ size: %u\n", cq_size);
-
-    if (ret != 0)
-    {
-        io_uring_queue_exit(&ctx->ring);
-        return -ret;
-    }
-    return 0;
-}
-
-void iouring_shutdown(iouring_ctx_t *ctx)
-{
-    io_uring_queue_exit(&ctx->ring);
-}
-
-
-void harvest_cqes(iouring_ctx_t *ctx)
-{
-    struct io_uring_cqe *cqe;
-    unsigned head;
-    int cq_count = 0;
-
-    // 使用 for_each_cqe 薅干净 CQ
-    io_uring_for_each_cqe(&ctx->ring, head, cqe) {
-        task_t *done = (task_t *)(uintptr_t)cqe->user_data;
-        task_finish(done, cqe->res);
-
-        if (cqe->res < 0) {
-            fprintf(stderr, "write fail: fd=%d res=%d\n", done->fd, cqe->res);
-        }
-        // 直接 destroy（单线程，无需全局队列）
-        task_destroy(done);
-
-        cq_count++;
-    }
-
-    if (cq_count > 0) {
-        // printf("harvest cq:%d\n", cq_count);
-        io_uring_cq_advance(&ctx->ring, cq_count);
-    }
-
-    // 检查 CQ overflow（保险）
-    if (*ctx->ring.sq.kflags & IORING_SQ_CQ_OVERFLOW) {
-        fprintf(stderr, "FATAL: CQ overflow detected!\n");
-        abort();
-    }
-}
-
-task_t* submit_write(iouring_ctx_t *ctx, int fd, void **bufs, size_t *lens, int count, off_t off){
-    if (!bufs || !lens || count <= 0) return NULL;
-
-    task_t *t = (task_t *)kvs_malloc(sizeof(task_t));
-    task_init(t);
-    t->op = TASK_WRITE;
-    t->fd = fd;
-    t->off = off;
-
-    t->iovs =  (struct iovec *)kvs_malloc(sizeof(struct iovec) * count);
-    if(!t->iovs) {
-        kvs_free(t);
-        return NULL;
-    }
-
-    for(int i = 0;i < count; ++ i){
-        size_t len = lens[i];
-        void *buf = kvs_malloc(len);
-        if(!buf){
-            for(int j = 0; j < i; ++j){
-                if(t->iovs[j].iov_base) kvs_free(t->iovs[j].iov_base);
-            }
-            kvs_free(t->iovs);
-            kvs_free(t);
-            return NULL;
-        }
-        memcpy(buf, bufs[i], len);
-        t->iovs[i].iov_base = buf;
-        t->iovs[i].iov_len = len;
-    }
-
-    t->iovcnt = count;
-    
-    harvest_cqes(ctx);
-
-    if(!ctx->head){
-        ctx->head = t;
-        ctx->tail = t;
-    }else{
-        ctx->tail->next = t;
-        ctx->tail = t;
-    }
-
-    int submitted = 0;
-    while(true){
-        task_t *cur = ctx->head;
-        if(!cur){
-            break;
-        }
-        ctx->head = cur->next;
-        if (!ctx->head) {
-            ctx->tail = NULL;
-        }
-        cur->next = NULL;
-
-        struct io_uring_sqe *sqe = io_uring_get_sqe(&ctx->ring);
-        if (!sqe) {
-            break;
-        }
-
-        io_uring_prep_writev(sqe, cur->fd, cur->iovs, cur->iovcnt, cur->off);
-        sqe->user_data = (uint64_t)(uintptr_t)cur;
-
-        submitted++;
-    }
-
-    if(submitted > 0){
-        int ret = io_uring_submit(&ctx->ring);
-    }
-
-    return t;
-}
-
-void iouring_tick(iouring_ctx_t *ctx) {
-    
-    harvest_cqes(ctx);
-
-    int submitted = 0;
-    while(ctx->head){
-        struct io_uring_sqe *sqe = io_uring_get_sqe(&ctx->ring);
-        if (!sqe) {
-            break;
-        }
-        task_t *cur = ctx->head;
-        ctx->head = cur->next;
-        if (!ctx->head) {
-            ctx->tail = NULL;
-        }
-        cur->next = NULL;
-
-        io_uring_prep_writev(sqe, cur->fd, cur->iovs, cur->iovcnt, cur->off);
-        sqe->user_data = (uint64_t)(uintptr_t)cur;
-
-        submitted++;
-    }
-
-    
-    if(submitted > 0){
-        int ret = io_uring_submit(&ctx->ring);
-    }
-}

diskuring/diskuring.h

@@ -3,65 +3,91 @@
 
 #include <liburing.h>
 #include <pthread.h>
-#include <stdlib.h>
 #include <stdint.h>
-#include <unistd.h>
+#include <stdatomic.h>
-#include <errno.h>
-#include <string.h>
 #include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/uio.h>
+#include <unistd.h>
 
+typedef enum { TASK_READ, TASK_WRITE, TASK_FSYNC } task_op_t;
 
-typedef enum { TASK_READ, TASK_WRITE } task_op_t;
+struct task;
+typedef void (*task_destroy_cb_t)(struct task *t, void *arg);
 
 typedef struct task {
     task_op_t op;
     int fd;
     off_t off;
+    unsigned fsync_flags;
+    unsigned sqe_flags;
 
-    int res;          // cqe->res
+    int res;
-    int done;         // 0/1
+    _Atomic int done;
 
-    struct iovec *iovs;  // iovec 数组
+    struct iovec *iovs;
-    int iovcnt;          // iovec 数量
+    int iovcnt;
-
+    int free_iov_bases;
-    pthread_mutex_t m;
+    task_destroy_cb_t on_destroy;
-    pthread_cond_t  cv;
+    void *on_destroy_arg;
 
     struct task *next;
 } task_t;
 
 typedef struct {
-    struct io_uring ring;
+    _Atomic(task_t *) head;
-    pthread_t th;
-
-    pthread_mutex_t q_m;
-    pthread_cond_t  q_cv;
-    task_t *q_head, *q_tail;
-    
-    int stop;
-    atomic_int in_flight;
-    int max_in_flight; 
-} iouring_ctx_t;
-
-typedef struct {
-    task_t *head;
-    pthread_mutex_t lock;
 } destroy_queue_t;
 
+typedef struct {
+    task_t **slots;
+    uint32_t cap;
+    _Atomic uint32_t head;
+    _Atomic uint32_t tail;
+    _Atomic uint32_t size;
+} spsc_queue_t;
+
+struct iouring_ctx_s;
+typedef struct iouring_ctx_s iouring_ctx_t;
+
+typedef struct {
+    struct io_uring ring;
+    pthread_t th;
+    int event_fd;
+    _Atomic int in_flight;
+    int max_in_flight;
+    int worker_id;
+    spsc_queue_t submit_q;
+    iouring_ctx_t *parent;
+} iouring_worker_t;
+
+struct iouring_ctx_s {
+    iouring_worker_t *workers;
+    int worker_nr;
+    unsigned entries_per_worker;
+
+    _Atomic int stop;
+    _Atomic uint32_t rr_next;
+
+    destroy_queue_t destroy_queue;
+};
 
 void task_init(task_t *t);
 void task_finish(task_t *t, int res);
-
 int task_wait(task_t *t);
 void task_destroy(task_t *t);
 
 int iouring_init(iouring_ctx_t *ctx, unsigned entries);
 void iouring_shutdown(iouring_ctx_t *ctx);
 
-task_t* submit_write(iouring_ctx_t *ctx, int fd, void **bufs, size_t *lens, int count, off_t off);
+task_t *submit_write(iouring_ctx_t *ctx, int fd, void **bufs, size_t *lens, int count, off_t off);
+task_t *submit_write_ref(iouring_ctx_t *ctx, int fd, void **bufs, size_t *lens, int count, off_t off,
+                         int free_iov_bases, task_destroy_cb_t on_destroy, void *on_destroy_arg);
+task_t *submit_fsync_ref(iouring_ctx_t *ctx, int fd, int worker_id, int drain,
+                         task_destroy_cb_t on_destroy, void *on_destroy_arg);
 int uring_task_complete(iouring_ctx_t *ctx);
-
+void cleanup_finished_iouring_tasks(iouring_ctx_t *ctx);
-void cleanup_finished_iouring_tasks();
+void iouring_profile_dump(iouring_ctx_t *ctx);
 
 extern iouring_ctx_t global_uring_ctx;
 

diskuring/diskuring.h.nothread

@@ -1,50 +0,0 @@
-#ifndef __DISK_IOURING_H__
-#define __DISK_IOURING_H__
-
-#include <liburing.h>
-#include <pthread.h>
-#include <stdlib.h>
-#include <stdint.h>
-#include <unistd.h>
-#include <errno.h>
-#include <string.h>
-#include <stdio.h>
-
-#define BATCH_SIZE  256
-typedef enum { TASK_READ, TASK_WRITE } task_op_t;
-
-typedef struct task {
-    task_op_t op;
-    int fd;
-    off_t off;
-
-    int res;          // cqe->res
-    int done;         // 0/1
-
-    struct iovec *iovs;  // iovec 数组
-    int iovcnt;          // iovec 数量
-
-    struct task *next;
-} task_t;
-
-typedef struct {
-    struct io_uring ring;
-    int pending_count;
-
-    task_t *head;
-    task_t *tail;
-} iouring_ctx_t;
-
-void task_init(task_t *t);
-void task_finish(task_t *t, int res);
-void task_destroy(task_t *t);
-
-int iouring_init(iouring_ctx_t *ctx, unsigned entries);
-void iouring_shutdown(iouring_ctx_t *ctx);
-
-task_t* submit_write(iouring_ctx_t *ctx, int fd, void **bufs, size_t *lens, int count, off_t off);
-void iouring_tick(iouring_ctx_t *ctx);
-
-extern iouring_ctx_t global_uring_ctx;
-
-#endif

doc/conclusion.md

@@ -0,0 +1,50 @@
+# 性能定位结论（主线程）
+
+## 1) 主要瓶颈结论
+
+基于当前主线程采样统计，结论是：
+
+- **主线程最大开销确实是内存申请/管理相关开销**（而不是纯 memcpy）。
+- 开销占比大致为：
+  - `submit_write` 打包阶段：约 **61.3%**
+    - 其中 alloc 约 **66.5% of pack**（折算到主线程总开销约 **40.8%**）
+    - copy 约 **14.2% of pack**（折算到主线程总开销约 **8.7%**）
+  - 入队/通知（SPSC push + notify）：约 **29.2%**
+  - 回收释放：约 **9.5%**
+  - backpressure 影响很小（loops≈0）
+
+所以“减少申请内存次数”这个判断是对的，而且是当前最有价值的优化方向。
+
+---
+
+## 2) 对方案 2 / 3 的评价（结合本系统）
+
+### 方案 2：A/B 双 flag 缓冲覆写保护
+
+- **优点**：实现直观，容易快速落地验证正确性。
+- **缺点**：
+  - 本质仍是“执行路径与拷贝路径耦合”，请求推进会受慢侧牵制；
+  - 不能从根本上消除 alloc 热点，只是控制覆写时序；
+  - 在高并发下调试成本会上升（状态机与边界条件多）。
+- **结论**：适合快速试验，不是长期高性能形态。
+
+### 方案 3：ChainBuffer 所有权移交（摘链/归还）
+
+- **优点**：
+  - 以“指针/节点移动”替代大量 alloc+copy，方向与当前瓶颈完全一致；
+  - 对大 value、高吞吐场景收益潜力更大。
+- **缺点**：
+  - 生命周期、并发归还、异常回收都要设计清楚；
+  - 实现复杂度明显高于方案 2。
+- **结论**：更符合长期性能目标，但需要更严格的工程设计。
+
+---
+
+## 3) 哪个在你当前系统里更容易实现？
+
+如果只比较你给的这两个方案，在你现在的系统里：
+
+- **更简单的是方案 2（A/B 双 flag）**；
+- **更值得长期投入的是方案 3（所有权移交）**。
+
+建议：短期先用方案 2 验证行为边界，最终收敛到方案 3（或其等价的零拷贝所有权模型）。

doc/interview_answers_aggressive_50.md

@@ -0,0 +1,300 @@
+# KVStore 激进版：50 题参考回答（含风险等级与避坑话术）
+
+> 对应文件：`doc/interview_questions_aggressive_50.md`  
+> 建议答题模板：**目标 -> 方案 -> 收益 -> 代价 -> 兜底**
+
+---
+
+## 1) 架构与边界（1-5）
+
+### 1. 分层设计收益与复杂度
+- **风险等级**：低  
+- **参考回答**：我把系统拆成网络层（收发）、协议层（RESP 解析）、执行层（命令分发）、持久化层（快照/oplog）和复制层（SSYNC+shm）。收益是可替换和可定位问题，比如网络和持久化互不耦合。代价是模块间接口变多，调试时需要跨层追踪。兜底是我保留了主调用链文档，排障按固定链路走。  
+- **避坑话术**：别说“完全解耦”，改成“低耦合、边界清晰”。
+
+### 2. Reactor vs Proactor/协程
+- **风险等级**：中  
+- **参考回答**：当前选 Reactor 是因为代码路径短、可控性高，适合先把协议和持久化跑稳。Proactor/协程在高并发下有潜力，但会引入更复杂的调度和状态管理。我的 tradeoff 是先优化主路径和持久化，再评估网络模型切换。  
+- **避坑话术**：不要说“Reactor 一定最快”，要说“在当前代码复杂度和目标下更合适”。
+
+### 3. 三引擎选型边界
+- **风险等级**：低  
+- **参考回答**：Array 实现简单但查找线性，适合小规模或验证场景；RBTree 有序且 O(logN)，适合有序访问；Hash 平均 O(1)，适合热点随机读写。收益是能覆盖不同 workload。代价是维护三套实现和一致性语义。  
+- **避坑话术**：不要泛泛说“都很快”，要给出复杂度和场景。
+
+### 4. 主路径轻量化
+- **风险等级**：中  
+- **参考回答**：主线程主要做解析、执行、回包，持久化提交走 io_uring worker。收益是减少 I/O 阻塞对主循环影响。代价是异步链路更复杂，需要回收队列和背压。我通过 destroy queue + wakeup 回收来控风险。  
+- **避坑话术**：别说“主线程无开销”，保留“主要开销下降”。
+
+### 5. 多核扩展优先级
+- **风险等级**：中  
+- **参考回答**：我会先做连接分片，再做数据分片。连接分片改动小、收益快；数据分片需要引擎和一致性改造，复杂度高。tradeoff 是短期吞吐提升 vs 长期架构演进，我会分两阶段做。  
+- **避坑话术**：不要一次性承诺“全做完”，强调迭代路线。
+
+---
+
+## 2) 协议与解析（6-10）
+
+### 6. 半包/多包/粘包与错误恢复
+- **风险等级**：低  
+- **参考回答**：我按“可消费字节”循环解析，len=0 视为半包等待下次，len<0 视为协议错误。收益是能稳定处理 pipeline。代价是错误策略偏严格，遇坏包直接断链。这是在安全性和可恢复性之间偏安全的选择。  
+- **避坑话术**：别说“能恢复所有坏包”。
+
+### 7. binary-safe 为何用 slice
+- **风险等级**：低  
+- **参考回答**：因为 key/value 可能包含 `\0` 和二进制字节，C 字符串 API 会截断。slice（ptr+len）能保留完整字节语义。收益是协议正确性和通用性更强，代价是代码里长度参数管理更繁琐。  
+- **避坑话术**：强调“正确性优先”。
+
+### 8. 恶意请求限制
+- **风险等级**：中  
+- **参考回答**：我在协议层设置了 bulk 上限和参数上限，同时连接层有读写缓冲上限。收益是避免单连接吃光内存。代价是极端合法请求也可能被拒绝。我会通过可配置阈值平衡安全和业务需求。  
+- **避坑话术**：不要说“完全防攻击”，说“降低风险面”。
+
+### 9. inline 与 multibulk 共存
+- **风险等级**：低  
+- **参考回答**：如果首字节不是 `*`，走 inline；是 `*` 就走 multibulk。收益是兼容 Redis 常见输入。代价是代码路径多一支，但复杂度可控。  
+- **避坑话术**：不要说“全面兼容 Redis”，说“覆盖当前命令集下的 RESP2 常用场景”。
+
+### 10. 新命令最小改动
+- **风险等级**：中  
+- **参考回答**：当前是命令枚举 + dispatch switch，新增命令主要改解析映射和执行分支。收益是性能直观、可调试。代价是规模大时 switch 变长。我会在命令数继续增长时再抽象为表驱动。  
+- **避坑话术**：别提前宣称“完全插件化”。
+
+---
+
+## 3) ChainBuffer 与“零拷贝”（11-15）
+
+### 11. 零拷贝边界
+- **风险等级**：高  
+- **参考回答**：接收阶段是 readv 直写到链式缓冲，减少中转拷贝；发送阶段用 sendmsg 聚合。当前仍存在线性化/打包场景下的 memcpy，所以更准确是“低拷贝主路径 + 局部拷贝回退”。tradeoff 是实现复杂度可控，同时先拿到主要收益。  
+- **避坑话术**：主动说“不是全链路绝对 0 拷贝”。
+
+### 12. linearize 触发条件
+- **风险等级**：中  
+- **参考回答**：当上层需要连续字节视图而数据跨 chunk 时触发 linearize。收益是简化了解析和执行接口。代价是触发时会有额外 memcpy，影响尾延迟。我会通过 chunk 策略和请求分布降低触发频率。  
+- **避坑话术**：不要说“从不 linearize”。
+
+### 13. free_list 选型
+- **风险等级**：低  
+- **参考回答**：free_list 复用 chunk，减少频繁 malloc/free 抖动。收益是吞吐更稳、分配开销更低。代价是会保留一定空闲内存。通过 free_limit 做上限控制，平衡性能和内存占用。  
+- **避坑话术**：别说“零内存浪费”。
+
+### 14. 大小 Key 混跑碎片问题
+- **风险等级**：中  
+- **参考回答**：我用固定 chunk + 链式扩展，先保证正确和稳定。大 Key 会占多 chunk，可能带来局部碎片；收益是实现简单、行为可预测。后续可按 key size 分层 chunk 策略优化。  
+- **避坑话术**：承认“有优化空间”更可信。
+
+### 15. 部分发送一致性
+- **风险等级**：低  
+- **参考回答**：sendmsg 返回 n 后按字节 drain 缓冲，未发完保留在 wbuf，EPOLLOUT 下次继续。收益是协议字节流不会错位。代价是状态维护复杂一点，但这是非阻塞发送必须的成本。  
+- **避坑话术**：别说“一次 send 一定发完”。
+
+---
+
+## 4) 所有权移交与并发安全（16-20）
+
+### 16. 所有权定义
+- **风险等级**：高  
+- **参考回答**：我的定义是“谁持有谁负责释放”，通过 detach/release 把片段生命周期显式化。收益是未来可做执行-落盘低拷贝协作。代价是引用计数和回收时序更复杂。当前实现已提供接口，进一步全链路化是下一阶段。  
+- **避坑话术**：明确“已实现接口与局部机制，非全链路完成”。
+
+### 17. UAF / double free 防护
+- **风险等级**：中  
+- **参考回答**：通过 refcnt + release 路径统一回收，避免多方直接 free。收益是降低并发回收风险。代价是需要严格约束调用顺序和异常路径。排查时我会优先核对 detach/release 对称性。  
+- **避坑话术**：别说“绝对不会 UAF”，说“通过机制显著降低风险”。
+
+### 18. 连接关闭后的回收
+- **风险等级**：中  
+- **参考回答**：连接关闭只释放连接侧对象，已移交片段由持有方继续走 release。收益是不会因连接断开丢失回收路径。代价是关闭逻辑需要区分“本地拥有”和“已移交”。  
+- **避坑话术**：避免一句话“直接全清掉”。
+
+### 19. 谁申请谁释放
+- **风险等级**：低  
+- **参考回答**：这是为了避免跨线程释放带来的生命周期混乱。收益是定位泄漏和崩溃更直观。代价是需要一个回收协作队列，代码量会增加。总体是工程上更稳的 tradeoff。  
+- **避坑话术**：强调“可维护性收益”。
+
+### 20. 真正 0 拷贝落盘约束
+- **风险等级**：高  
+- **参考回答**：要做到真正 0 拷贝，需要落盘线程直接消费网络片段并延迟释放，约束包括引用计数、回收屏障和慢盘背压。收益是减少打包复制。代价是并发复杂度显著上升，错误成本更高。  
+- **避坑话术**：别承诺“短期必做完”，给阶段目标。
+
+---
+
+## 5) io_uring 流水线（21-25）
+
+### 21. n*SPSC vs MPSC
+- **风险等级**：中  
+- **参考回答**：n*SPSC 的好处是队列局部性好、竞争少，延迟更稳；代价是负载均衡和线程参数调优复杂。MPSC 实现集中但热点明显。当前我更看重稳定性，所以先用 n*SPSC。  
+- **避坑话术**：不要否定 MPSC，强调场景选择。
+
+### 22. 80% 水位
+- **风险等级**：中  
+- **参考回答**：80% 是保护阈值，目的是给 CQE 回收和突发流量留缓冲。收益是降低 overflow 风险。代价是峰值吞吐可能不是最大。这个值不是常量真理，我会按压测再调。  
+- **避坑话术**：不要说“最优解”，说“保守工程值”。
+
+### 23. 背压策略
+- **风险等级**：中  
+- **参考回答**：队列满时会回收已完成任务并让出 CPU，必要时限流连接。收益是系统不崩。代价是尾延迟上升。tradeoff 是优先保活再保时延。  
+- **避坑话术**：别说“无损无延迟”。
+
+### 24. 批量回收
+- **风险等级**：低  
+- **参考回答**：批量偷取 destroy queue 能减少锁/原子开销和碎片回收抖动。收益是主循环更平滑。代价是回收时点有批处理延迟。这个延迟通常可接受。  
+- **避坑话术**：承认“不是实时逐条释放”。
+
+### 25. 优雅停机
+- **风险等级**：中  
+- **参考回答**：停机流程是 stop 标志 -> 唤醒 worker -> drain 队列 -> join 线程 -> 清理剩余任务。收益是尽量保证数据路径收敛完成。代价是停机时间变长。这里优先数据完整性。  
+- **避坑话术**：别说“秒停”，说“可控停机”。
+
+---
+
+## 6) 快照 + 增量一致性（26-30）
+
+### 26. 一致性窗口
+- **风险等级**：高  
+- **参考回答**：快照代表某个时间点状态，后续变更由 oplog 补齐。收益是写入不中断。代价是天然存在窗口，需要恢复阶段串联重放。这个设计是吞吐优先下的最终一致方案。  
+- **避坑话术**：不要声称“强一致快照”。
+
+### 27. 先快照后回放
+- **风险等级**：低  
+- **参考回答**：快照是基线，oplog 是增量。先加载基线再套增量语义清晰。收益是恢复逻辑简单。代价是快照越旧回放越长。可通过更频繁快照平衡。  
+- **避坑话术**：说明“恢复时间与快照频率 tradeoff”。
+
+### 28. 命令字节流日志
+- **风险等级**：中  
+- **参考回答**：命令字节流实现快、复用现有解析器，开发成本低。代价是回放时要重新解析执行，效率不如结构化 WAL。当前选择是快速迭代优先，后续可演进为结构化记录。  
+- **避坑话术**：承认“不是最终形态”。
+
+### 29. 坏日志策略
+- **风险等级**：高  
+- **参考回答**：当前更偏“发现坏日志就失败退出”，先保证正确性边界。收益是避免静默数据错误。代价是可用性下降。生产化可加校验和分段容错，再做可恢复降级。  
+- **避坑话术**：不要说“自动修复所有坏日志”。
+
+### 30. 状态边界证明
+- **风险等级**：高  
+- **参考回答**：我会把边界定义成“快照点 + 之后成功持久化的增量”。收益是定义清楚，便于测试。代价是需要故障注入验证不同崩溃点。我的做法是先给出可验证边界，而不是口头保证。  
+- **避坑话术**：避免绝对词“完全一致”。
+
+---
+
+## 7) 主从与状态机（31-35）
+
+### 31. 状态机描述
+- **风险等级**：中  
+- **参考回答**：slave 发 SSYNC 申请全量，master 异步产出并发送 snapshot，slave 落盘后回 SREADY，再进入增量流。收益是流程清晰且可扩展。代价是阶段切换复杂，需要严格顺序控制。  
+- **避坑话术**：别忽略“阶段切换失败处理”。
+
+### 32. 快照期间新写入
+- **风险等级**：高  
+- **参考回答**：思路是用 seq 把快照基线和增量窗口连接起来。收益是减少丢数据窗口。代价是实现复杂，需要状态机和顺序保证。当前代码有通道和钩子，协同还在持续完善。  
+- **避坑话术**：一定说“机制已搭，协同在完善中”。
+
+### 33. wrap marker 可行性
+- **风险等级**：中  
+- **参考回答**：尾部放不下时写 wrap marker 并回到 0，消费者识别后跳转。收益是实现简单。代价是需要严格保证读写顺序和越界检查。适合单写者模型。  
+- **避坑话术**：别说“多写者也没问题”。
+
+### 34. seq 断档处理
+- **风险等级**：高  
+- **参考回答**：断档处理要看目标：一致性优先就阻塞/重建，可用性优先可短暂跳过并告警。我的偏好是关键链路一致性优先。代价是短期可用性受影响。  
+- **避坑话术**：避免“所有场景一个策略”。
+
+### 35. eBPF 控制面定位
+- **风险等级**：中  
+- **参考回答**：我把 eBPF 钩子放控制面，用于状态切换和转发触发，不直接承载主数据处理。收益是降低主路径侵入。代价是又多一层调试复杂度。  
+- **避坑话术**：别说“eBPF 加速了所有链路”。
+
+---
+
+## 8) 内存分配器与内存池（36-40）
+
+### 36. 分级依据
+- **风险等级**：低  
+- **参考回答**：8-512B 覆盖高频小对象，按 8B 对齐分桶，页内 free-list 管理。收益是分配释放路径短。代价是可能有内部碎片。属于延迟优先的选择。  
+- **避坑话术**：不要回避碎片问题。
+
+### 37. 大对象回退 malloc
+- **风险等级**：低  
+- **参考回答**：大对象生命周期和尺寸离散，放进小块池会放大管理成本。回退系统分配更简单。收益是降低池复杂度。代价是大块分配性能受系统 allocator 影响。  
+- **避坑话术**：别说“mempool 覆盖全部场景”。
+
+### 38. 空闲页回收
+- **风险等级**：中  
+- **参考回答**：我保留一定空闲页做缓存，超过阈值再释放，平衡复用和内存占用。收益是减少抖动。代价是峰值内存更高。阈值可按业务压测调。  
+- **避坑话术**：强调“可调参数”。
+
+### 39. 三 allocator 差异来源
+- **风险等级**：中  
+- **参考回答**：差异主要来自小对象分配路径、锁竞争、缓存复用和写盘模式耦合。mypool 在当前 workload 下命中率高，所以吞吐更好。代价是通用性不一定优于成熟 allocator。  
+- **避坑话术**：别说“mypool 永远最优”。
+
+### 40. 内存异常排查优先级
+- **风险等级**：低  
+- **参考回答**：先看 RSS、分配失败、队列积压、in-flight、回收延迟，再看对象分布。收益是先定位“增长来源”再定位“泄漏点”。代价是需要加一些可观测指标。  
+- **避坑话术**：别一上来就说“就是泄漏”。
+
+---
+
+## 9) 压测方法学（41-45）
+
+### 41. 复现性保证
+- **风险等级**：低  
+- **参考回答**：固定参数、固定轮次、分组合独立目录、每轮重启，保存原始日志与 CSV。收益是可回放可核对。代价是执行时间更长。  
+- **避坑话术**：不要只报“最好一轮数据”。
+
+### 42. bench vs testcase 取舍
+- **风险等级**：中  
+- **参考回答**：bench 更灵活，但随机 key 与 RSET 插入语义会碰撞；testcase mode=4 更稳定，适合长期对比。tradeoff 是灵活性 vs 稳定性。  
+- **避坑话术**：主动解释语义差异，避免被质疑口径不一致。
+
+### 43. QPS 口径
+- **风险等级**：中  
+- **参考回答**：我用“完成请求数/耗时”，并明确是否包含错误。收益是可比较。代价是不同工具默认口径不同，需要文档化。  
+- **避坑话术**：别把不同口径结果直接横比。
+
+### 44. 避免虚高
+- **风险等级**：低  
+- **参考回答**：做预热、控制 keyspace、分离预填充和正式压测、多轮取均值并看波动。收益是结果更稳。代价是实验时间增加。  
+- **避坑话术**：不要只强调峰值，不报波动。
+
+### 45. 波动解释
+- **风险等级**：中  
+- **参考回答**：波动来自调度、缓存、I/O 抖动和后台回收。我的做法是看均值+CV，不用单轮结论。tradeoff 是展示更真实但数字不一定最亮眼。  
+- **避坑话术**：别把异常轮次“隐去不提”。
+
+---
+
+## 10) 真实性与演进（46-50）
+
+### 46. 零拷贝表述边界
+- **风险等级**：高  
+- **参考回答**：我会说“接收/发送主路径低拷贝，局部场景有拷贝回退”，而不是“全链路绝对零拷贝”。收益是表达亮点同时不失真。代价是表述没那么激进。  
+- **避坑话术**：先给边界，再讲优化方向。
+
+### 47. 状态机协同实现到哪
+- **风险等级**：高  
+- **参考回答**：SSYNC/SREADY、快照传输、增量通道与 seq 机制已打通；全量-增量无缝协同仍在完善。这样说能体现工程进度和规划。  
+- **避坑话术**：别说“完全解决协同一致性”。
+
+### 48. 两周落地计划
+- **风险等级**：中  
+- **参考回答**：第 1 周先补一致性边界和故障注入测试；第 2 周做所有权移交闭环和指标监控。收益是先补正确性再补性能。代价是短期新功能延后。  
+- **避坑话术**：给里程碑，不要空谈“能做完”。
+
+### 49. 最大技术债
+- **风险等级**：中  
+- **参考回答**：最大的债是“协同一致性语义和容错测试还不够系统化”。我优先做了主链路可运行和性能基线。tradeoff 是先可用再完善严谨性。  
+- **避坑话术**：承认技术债并给修复计划，面试官反而更认可。
+
+### 50. 三人小组下一版目标
+- **风险等级**：低  
+- **参考回答**：目标按优先级：一致性文档+故障矩阵、可观测性、性能回归门禁。验收指标包括恢复正确率、复制延迟、QPS/CV 基线。tradeoff 是减少“新功能数量”，换“上线可信度”。  
+- **避坑话术**：不要只提“再提 2 倍 QPS”。
+
+---
+
+## 快速复习（高风险题清单）
+- **高风险题**：11, 16, 20, 26, 29, 30, 32, 34, 46, 47  
+- **统一策略**：先讲“已实现边界”，再讲“预案与下一步”，避免绝对化承诺。

doc/interview_questions_aggressive_50.md

@@ -0,0 +1,76 @@
+# KVStore 激进版本：面试高频追问 50 题（含 Tradeoff 视角）
+
+## 回答建议（统一框架）
+- 建议每题按「目标 -> 方案 -> 收益 -> 代价 -> 兜底」回答。
+- 收益常见维度：吞吐、时延、稳定性、可维护性、可扩展性。
+- 代价常见维度：实现复杂度、内存开销、排障成本、一致性窗口、开发周期。
+
+## 1) 架构与边界（1-5）
+1. 你把系统拆成了哪些层（网络/协议/执行/持久化/复制）？这种分层带来的收益与额外复杂度分别是什么？
+2. 为什么默认选择 Reactor 而不是 Proactor/协程网络模型？三者在吞吐、延迟、开发复杂度上的 tradeoff 是什么？
+3. Array/RBTree/Hash 三种引擎的选型边界是什么？在不同数据分布下的性能收益与维护成本如何权衡？
+4. 你说“主路径轻量化”，具体删掉了哪些主线程工作？对 CPU 利用率与代码复杂度的 tradeoff 是什么？
+5. 如果要做多核扩展，你会优先做连接分片还是数据分片？两者在扩展性与一致性上的代价分别是什么？
+
+## 2) 协议与解析（6-10）
+6. RESP 解析如何处理半包/多包/粘包？选择“严格报错”还是“尽量恢复”的 tradeoff 是什么？
+7. binary-safe 为什么必须用 slice（ptr+len）而不是 C 字符串？这对性能和代码可读性有什么影响？
+8. 你如何限制恶意请求（超大 bulk、超多参数）？安全性提升与正常流量误伤之间如何平衡？
+9. inline 与 multibulk 共存时如何处理优先级？兼容性收益与实现复杂度代价分别是什么？
+10. 新增命令时如何保证最小改动？抽象过度与快速迭代之间的 tradeoff 怎么拿捏？
+
+## 3) ChainBuffer 与“零拷贝”追问（11-15）
+11. readv 直写后用户态还会发生哪些拷贝？你如何定义“零拷贝”和“低拷贝”的边界？
+12. chain_buffer_linearize 何时触发？减少复杂解析逻辑与引入额外 memcpy 的 tradeoff 是什么？
+13. 为什么要做 chunk free_list？复用内存带来的性能收益与内存峰值风险如何平衡？
+14. 大 Key（64KB）与小 Key 混跑时，chunk 策略如何避免碎片化？吞吐和内存效率谁优先？
+15. sendmsg 聚合发送遇到部分发送时如何保证一致性？更高吞吐与更复杂状态管理如何权衡？
+
+## 4) 所有权移交与并发安全（16-20）
+16. “执行-落盘共享缓冲片段”如何定义所有权？减少拷贝收益与生命周期复杂度代价怎么评估？
+17. detach/release 如何避免 UAF/double free？你牺牲了哪些性能换取内存安全？
+18. 连接提前关闭时已移交数据如何回收？优先保证数据安全还是优先快速释放资源？
+19. 为什么采用“谁申请谁释放”？跨线程释放带来的便利与隐患如何取舍？
+20. 若做真正 0 拷贝落盘，你会接受哪些新增约束（引用计数/锁/回收时序）？
+
+## 5) io_uring 持久化流水线（21-25）
+21. n*SPSC 相比单队列 MPSC 的收益和代价分别是什么？你为什么在当前阶段选 n*SPSC？
+22. in-flight 上限为何设为 CQ 的 80%？保守水位与峰值吞吐之间如何权衡？
+23. 队列满时背压策略是什么？“保护系统稳定”与“牺牲尾延迟”之间如何取舍？
+24. destroy_queue 用批量偷取+集中释放的动机是什么？释放抖动与主循环平滑性如何平衡？
+25. shutdown 阶段如何保证无悬挂任务？优雅退出时间与数据完整性哪个优先？
+
+## 6) 快照 + 增量日志一致性（26-30）
+26. SAVE 与 oplog append 并发时一致性窗口怎么定义？强一致与吞吐的 tradeoff 是什么？
+27. 为什么恢复顺序是“先快照后 oplog”？恢复速度与恢复正确性之间如何权衡？
+28. oplog 记录命令字节流而不是逻辑变更，优势和成本分别是什么？
+29. replay 遇到坏日志时你会“失败退出”还是“跳过继续”？可用性与正确性怎么选？
+30. 你如何说明恢复后状态边界？文档化成本与实现灵活性的 tradeoff 是什么？
+
+## 7) 主从同步与状态机（31-35）
+31. 请描述 SSYNC -> Snapshot -> SREADY -> Incremental 状态机，并说明每步的收益与风险。
+32. 快照传输期间新写入如何不丢？延迟增大与一致性增强之间怎么平衡？
+33. 共享内存 ring wrap marker 方案为什么可行？简单实现与健壮性之间的代价是什么？
+34. seq 不连续时你为何选择阻塞/跳过/重建？各策略的可用性与数据风险如何比较？
+35. eBPF uprobe 放在控制面而非数据面的考量是什么？观测能力与运行时开销如何权衡？
+
+## 8) 内存分配器与内存池（36-40）
+36. mempool 分级（8-512B）的依据是什么？固定桶命中率与碎片风险如何平衡？
+37. 大对象回退 malloc 的原因是什么？统一路径与分层路径在复杂度上怎么取舍？
+38. mempool 如何处理空闲页回收？低延迟复用与低内存占用的 tradeoff 怎么设定？
+39. malloc/jemalloc/mypool 在你的 workload 下差异来自哪里？泛化能力与场景优化如何平衡？
+40. 线上内存峰值异常时先看哪些指标？指标全面性与观测成本如何取舍？
+
+## 9) 压测方法学与结果可信度（41-45）
+41. 你如何保证 benchmark 可复现？“实验真实度”与“执行成本”之间怎么平衡？
+42. 为什么有些场景用 bench.c，有些改用 testcase mode=4？语义准确与工具统一如何取舍？
+43. QPS 统计口径怎么定义？是否包含失败请求？可比性与直观性之间如何平衡？
+44. 如何避免预热不足/缓存命中导致虚高？测试严谨性与测试周期如何平衡？
+45. 如何解释 round 间波动（CV）？追求峰值还是追求稳定性的 tradeoff 是什么？
+
+## 10) 真实性、边界与演进（46-50）
+46. 简历里“零拷贝”哪些已落地，哪些是低拷贝/预案？为什么这样表述？
+47. 简历里“协同状态机”当前实现到哪一步？工程现实与对外表达如何平衡？
+48. 若给你 2 周把预案落地，你的里程碑怎么排？短期收益与长期架构如何取舍？
+49. 当前最大技术债是什么？为什么没有先修？业务推进与技术治理如何平衡？
+50. 若带 3 人继续做，下版目标与验收指标是什么？功能扩展与稳定性建设如何排序？

doc/resume.tex

@@ -0,0 +1,25 @@
+\section{项目经历}
+
+\ResumeItem[KVStore 高性能 KV 存储系统（RESP 兼容 + 异步持久化 + 主从同步）]
+{\textbf{KVStore} 高性能 KV 存储系统（RESP 兼容 + 异步持久化 + 主从同步）}
+[个人项目|独立开发]
+[2025.03 --- 2026.03]
+\begin{itemize}
+  \item 基于 \textbf{C + Linux} 从零实现单机 KV 服务：支持 \textbf{RESP2 协议解析}、pipeline 与 binary-safe（支持 \texttt{\textbackslash 0}），并统一接入 \textbf{Array / RBTree / Hash} 三类存储引擎。
+  \item 实现链式网络缓冲 \textbf{ChainBuffer}：通过 \textbf{readv/sendmsg} 构建分段零拷贝收发路径，支持大 Key 与半包/多包场景；提供 \textbf{detach/release} 所有权接口，支持后续执行-落盘共享同一数据片段。
+  \item 实现 \textbf{io\_uring + n*SPSC} 异步落盘模型：主线程执行后写入增量日志，worker 线程批量提交 \textbf{writev}；通过 in-flight 控制、队列背压、destroy-queue 回收，避免 CQ 溢出与内存失控。
+  \item 设计并落地“\textbf{快照 + 增量日志}”恢复链路：支持启动加载快照并回放 oplog；支持 \textbf{SSYNC $\rightarrow$ 快照传输 $\rightarrow$ SREADY $\rightarrow$ 增量同步} 的主从接管流程，并预留 eBPF uprobe 协同点。
+  \item 实现可插拔内存分配策略（\textbf{malloc/jemalloc/mypool}）：自研 mempool 采用 8--512B 分级桶 + 页级 free-list；在 allocator$\times$persistence 复测中，\textbf{mypool} 吞吐最佳（none: 924878 QPS，incremental: 747101 QPS）。
+  \item 构建 hiredis 功能/压测工具链（\texttt{testcase + bench + 自动化脚本}），沉淀多轮可复现实验口径；将关键优化量化为可交付结果（如 ChainBuffer 改造后写路径 QPS 提升约 \textbf{27\%}）。
+\end{itemize}
+
+\ResumeItem[EncryptSql 基于 PostgreSQL 的透明加密查询与运算框架]
+{\textbf{EncryptSQL} 基于 PostgreSQL 的透明加密查询与运算框架}
+[学校横向|部分代码开发]
+[2024.09 --- 2025.09]
+\begin{itemize}
+  \item 在客户端侧对 \textbf{libpq} 进行改造，实现 \textbf{SQL 解析后重写}：将原生表达式/运算符节点替换为密态版本函数/算子调用，尽量保证业务侧无侵入接入。
+  \item 基于 \textbf{PostgreSQL 扩展机制}（自定义函数/算子等）接入密态运算：支持常见算术计算（加/减/乘/除）与部分聚合能力，并与查询执行流程集成。
+  \item 设计并实现基于工厂模式的 \textbf{KMS 接口层}：在 \texttt{encryptsql} 组件中统一密钥获取与管理流程，完成 \textbf{LocalKMS} 与 \textbf{Huawei KMS API} 适配，支持外部 KMS 平滑替换。
+  \item 面向高安全计算场景，引入 \textbf{TEE} 承载关键运算链路，在安全性与性能开销之间做工程化平衡。
+\end{itemize}

doc/resume_aggressive.tex

@@ -0,0 +1,25 @@
+\section{项目经历}
+
+\ResumeItem[KVStore 高性能 KV 存储系统（激进表达版）]
+{\textbf{KVStore} 高性能 KV 存储系统（零拷贝接收 + 异步持久化 + 主从协同状态机）}
+[个人项目|独立开发]
+[2025.03 --- 2026.03]
+\begin{itemize}
+  \item 基于 \textbf{C + Linux} 实现 RESP 兼容 KV 存储内核，支持 \textbf{binary-safe}、pipeline 与多引擎统一分发（Array/RBTree/Hash），形成协议层到执行层的一体化数据通路。
+  \item 围绕大 Key 场景实现 \textbf{ChainBuffer 分段零拷贝接收}：采用 \textbf{readv/sendmsg} 与链式 chunk 组织，支持超大请求分段处理；按线上保护阈值将单请求上限收敛至 \textbf{65535} 字节级别。
+  \item 实现并演进 \textbf{所有权移交} 机制：主线程仅负责命令边界识别与执行，落盘线程复用网络缓冲片段进行持久化，减少主路径内存申请/对象拼装开销。
+  \item 搭建 \textbf{io\_uring + n*SPSC} 持久化流水线：worker 批量提交 writev，主线程异步回收完成任务；结合 in-flight 背压与 destroy-queue，稳定处理慢盘与高并发写入抖动。
+  \item 设计“\textbf{快照 + 增量日志 + 实时复制}”协同方案：通过 \textbf{SSYNC $\rightarrow$ Snapshot $\rightarrow$ SREADY $\rightarrow$ Incremental} 状态机衔接全量与增量，保障复制窗口内的可恢复性与顺序一致性。
+  \item 构建多维压测体系（功能正确性/吞吐/波动）：在 allocator$\times$persistence 复测中，\textbf{mypool} 取得最佳吞吐（none: 924878 QPS，incremental: 747101 QPS），并将优化效果沉淀为工程基线。
+\end{itemize}
+
+\ResumeItem[EncryptSql 基于 PostgreSQL 的透明加密查询与运算框架]
+{\textbf{EncryptSQL} 基于 PostgreSQL 的透明加密查询与运算框架}
+[学校横向|部分代码开发]
+[2024.09 --- 2025.09]
+\begin{itemize}
+  \item 在客户端侧改造 \textbf{libpq} 并实现 \textbf{SQL AST 重写}：将明文表达式自动替换为密态函数/算子调用，降低业务系统改造成本。
+  \item 基于 \textbf{PostgreSQL 扩展机制}接入密态算子，支持加/减/乘/除与部分聚合能力，形成可落地的“密文存储 + 密态计算”执行路径。
+  \item 设计并实现 \textbf{KMS 工厂接口框架}，完成 \textbf{LocalKMS/Huawei KMS} 适配，支持多云与私有化 KMS 的低成本切换。
+  \item 在高敏感计算场景引入 \textbf{TEE}，对关键流程进行可信执行与边界隔离，平衡安全目标与查询性能。
+\end{itemize}

doc/resume_conservative.tex

@@ -0,0 +1,25 @@
+\section{项目经历}
+
+\ResumeItem[KVStore 高性能 KV 存储系统（保守版）]
+{\textbf{KVStore} 高性能 KV 存储系统（RESP 兼容 + 异步持久化 + 主从同步）}
+[个人项目|独立开发]
+[2025.03 --- 2026.03]
+\begin{itemize}
+  \item 基于 \textbf{C + Linux} 实现单机 KV 服务，支持 \textbf{RESP2} 解析、pipeline 与 binary-safe（支持 \texttt{\textbackslash 0}）键值处理。
+  \item 统一命令分发层，接入 \textbf{Array / RBTree / Hash} 三种引擎，实现 SET/GET/DEL 及对应 R*/H* 命令族。
+  \item 实现 \textbf{ChainBuffer} 分段网络缓冲：接收侧使用 \textbf{readv} 直写，发送侧使用 \textbf{sendmsg} 聚合发送，并通过 linearize 处理跨分段解析场景。
+  \item 实现 \textbf{io\_uring + n*SPSC} 异步增量日志写入，包含 in-flight 限流、背压与完成队列回收，降低主线程阻塞。
+  \item 实现“\textbf{快照 + 增量日志}”恢复路径：支持 SAVE 快照、oplog 回放；支持 \textbf{SSYNC/SREADY} 启动同步与共享内存增量通道。
+  \item 构建 hiredis 功能/性能测试工具链；在 2026-03-04 的 allocator$\times$persistence 复测中，\textbf{mypool} 组合吞吐最佳（none: 924878 QPS，incremental: 747101 QPS）。
+\end{itemize}
+
+\ResumeItem[EncryptSql 基于 PostgreSQL 的透明加密查询与运算框架]
+{\textbf{EncryptSQL} 基于 PostgreSQL 的透明加密查询与运算框架}
+[学校横向|部分代码开发]
+[2024.09 --- 2025.09]
+\begin{itemize}
+  \item 在客户端侧对 \textbf{libpq} 进行改造，实现 \textbf{SQL 解析后重写}：将原生表达式/运算符节点替换为密态函数/算子调用，尽量保证业务侧无侵入接入。
+  \item 基于 \textbf{PostgreSQL 扩展机制}（自定义函数/算子等）接入密态运算：支持常见算术计算（加/减/乘/除）与部分聚合能力，并与执行流程集成。
+  \item 设计并实现基于工厂模式的 \textbf{KMS 接口层}：在 \texttt{encryptsql} 中统一密钥获取与管理流程，完成 \textbf{LocalKMS} 与 \textbf{Huawei KMS API} 适配。
+  \item 面向高安全计算场景，引入 \textbf{TEE} 承载关键运算链路，在安全性与性能开销之间做工程化平衡。
+\end{itemize}

dump/kvs_dump.h

@@ -10,14 +10,21 @@ extern char global_rbtree_file[256];
 extern char global_hash_file[256];
 
 int kvs_create_snapshot(iouring_ctx_t *uring, const char* array_file, const char* rbtree_file, const char* hash_file);
-int kvs_create_snapshot_async(const char *ip, int port);
+int kvs_create_snapshot_async_1(iouring_ctx_t *uring, const char* array_file, const char* rbtree_file, const char* hash_file);
-
+int kvs_create_snapshot_async_2(const char *ip, int port);
 extern int global_oplog_fd;
 
 
 int init_cmd_log(const char *file, int *logfd);
 int destroy_cmd_log(int logfd);
 
+enum {
+    KVS_OPLOG_BUF_NOT_FULL = 0,
+    KVS_OPLOG_BUF_FULL = 1
+};
+
+int kvs_oplog_buffer_append(const uint8_t *cmd, size_t len, int logfd);
+int kvs_oplog_flush(int logfd, int force);
 int kvs_oplog_append(const uint8_t *cmd, size_t len, int logfd);
 int kvs_replay_log(int logfd);
 int ksv_clear_log(int logfd);

dump/kvs_oplog.c

@@ -3,58 +3,557 @@
 #include "memory/alloc_dispatch.h"
 #include "kvs_protocol_resp.h"
 #include "diskuring/diskuring.h"
+#include "common/config.h"
 #include <arpa/inet.h>
 #include <fcntl.h>
+#include <pthread.h>
+#include <string.h>
+#include <time.h>
 #include <unistd.h>
 
 int global_oplog_fd = -1;
 static off_t g_log_off = -1;
+extern AppConfig global_cfg;
+
+#define KVS_OPLOG_PAGE_SIZE (64u * 1024u)
+
+typedef struct oplog_buf {
+    struct oplog_buf *next;
+    off_t off;
+    size_t used;
+    uint8_t data[KVS_OPLOG_PAGE_SIZE];
+} oplog_buf_t;
+
+static oplog_buf_t *g_oplog_idle_head = NULL;
+static oplog_buf_t *g_oplog_idle_tail = NULL;
+static oplog_buf_t *g_oplog_ready_head = NULL;
+static oplog_buf_t *g_oplog_ready_tail = NULL;
+static oplog_buf_t *g_oplog_cur = NULL;
+
+static pthread_mutex_t g_oplog_mu = PTHREAD_MUTEX_INITIALIZER;
+static pthread_cond_t g_sync_cv = PTHREAD_COND_INITIALIZER;
+static pthread_t g_sync_th;
+static int g_sync_started = 0;
+static int g_sync_stop = 0;
+static int g_sync_logfd = -1;
+static uint64_t g_sync_gen = 0;
+static uint64_t g_sync_synced_gen = 0;
+
+static inline void oplog_mark_dirty_locked(void) {
+    g_sync_gen++;
+    pthread_cond_signal(&g_sync_cv);
+}
+
+typedef struct {
+    pthread_mutex_t mu;
+    pthread_cond_t cv;
+    int pending;
+    int submit_failed;
+    int fsync_err;
+} oplog_fsync_wait_t;
+
+typedef struct {
+    oplog_fsync_wait_t *waiter;
+} oplog_fsync_arg_t;
+
+static void make_timeout_ms(struct timespec *ts, long ms) {
+    if (!ts) {
+        return;
+    }
+    clock_gettime(CLOCK_REALTIME, ts);
+    ts->tv_sec += ms / 1000;
+    ts->tv_nsec += (ms % 1000) * 1000000L;
+    if (ts->tv_nsec >= 1000000000L) {
+        ts->tv_sec += 1;
+        ts->tv_nsec -= 1000000000L;
+    }
+}
+
+static void oplog_fsync_done(task_t *t, void *arg) {
+    oplog_fsync_arg_t *a = (oplog_fsync_arg_t *)arg;
+    oplog_fsync_wait_t *w;
+
+    if (!a || !a->waiter) {
+        return;
+    }
+    w = a->waiter;
+
+    pthread_mutex_lock(&w->mu);
+    if (t && t->res < 0 && w->fsync_err == 0) {
+        w->fsync_err = t->res;
+    }
+    if (w->pending > 0) {
+        w->pending--;
+    }
+    if (w->pending == 0) {
+        pthread_cond_signal(&w->cv);
+    }
+    pthread_mutex_unlock(&w->mu);
+}
+
+static int kvs_oplog_fsync_all_workers(int fd) {
+    int i;
+    int n;
+    int rc = 0;
+    oplog_fsync_wait_t w;
+    oplog_fsync_arg_t *args = NULL;
+
+    if (fd < 0 || !global_uring_ctx.workers || global_uring_ctx.worker_nr <= 0) {
+        return -1;
+    }
+
+    memset(&w, 0, sizeof(w));
+    pthread_mutex_init(&w.mu, NULL);
+    pthread_cond_init(&w.cv, NULL);
+
+    n = global_uring_ctx.worker_nr;
+    args = (oplog_fsync_arg_t *)kvs_malloc(sizeof(oplog_fsync_arg_t) * (size_t)n);
+    if (!args) {
+        pthread_cond_destroy(&w.cv);
+        pthread_mutex_destroy(&w.mu);
+        return -1;
+    }
+
+    for (i = 0; i < n; i++) {
+        task_t *t;
+        args[i].waiter = &w;
+
+        pthread_mutex_lock(&w.mu);
+        w.pending++;
+        pthread_mutex_unlock(&w.mu);
+
+        t = submit_fsync_ref(&global_uring_ctx, fd, i, 1, oplog_fsync_done, &args[i]);
+        if (!t) {
+            pthread_mutex_lock(&w.mu);
+            w.pending--;
+            w.submit_failed = 1;
+            if (w.pending == 0) {
+                pthread_cond_signal(&w.cv);
+            }
+            pthread_mutex_unlock(&w.mu);
+            rc = -1;
+            break;
+        }
+    }
+
+    pthread_mutex_lock(&w.mu);
+    while (w.pending > 0) {
+        struct timespec ts;
+        make_timeout_ms(&ts, 10);
+        (void)pthread_cond_timedwait(&w.cv, &w.mu, &ts);
+        if (w.pending > 0) {
+            pthread_mutex_unlock(&w.mu);
+            cleanup_finished_iouring_tasks(&global_uring_ctx);
+            pthread_mutex_lock(&w.mu);
+        }
+    }
+
+    if (w.fsync_err < 0 || w.submit_failed) {
+        rc = -1;
+    }
+    pthread_mutex_unlock(&w.mu);
+
+    cleanup_finished_iouring_tasks(&global_uring_ctx);
+    kvs_free(args);
+    pthread_cond_destroy(&w.cv);
+    pthread_mutex_destroy(&w.mu);
+    return rc;
+}
+
+static void oplog_push_tail(oplog_buf_t **head, oplog_buf_t **tail, oplog_buf_t *buf) {
+    if (!head || !tail || !buf) {
+        return;
+    }
+    buf->next = NULL;
+    if (!*tail) {
+        *head = *tail = buf;
+        return;
+    }
+    (*tail)->next = buf;
+    *tail = buf;
+}
+
+static void oplog_push_front(oplog_buf_t **head, oplog_buf_t **tail, oplog_buf_t *buf) {
+    if (!head || !tail || !buf) {
+        return;
+    }
+    if (!*head) {
+        buf->next = NULL;
+        *head = *tail = buf;
+        return;
+    }
+    buf->next = *head;
+    *head = buf;
+}
+
+static oplog_buf_t *oplog_pop_head(oplog_buf_t **head, oplog_buf_t **tail) {
+    oplog_buf_t *buf;
+    if (!head || !tail || !*head) {
+        return NULL;
+    }
+    buf = *head;
+    *head = buf->next;
+    if (!*head) {
+        *tail = NULL;
+    }
+    buf->next = NULL;
+    return buf;
+}
+
+static oplog_buf_t *oplog_alloc_buf(void) {
+    oplog_buf_t *buf = (oplog_buf_t *)kvs_malloc(sizeof(oplog_buf_t));
+    if (!buf) {
+        return NULL;
+    }
+    buf->next = NULL;
+    buf->off = 0;
+    buf->used = 0;
+    return buf;
+}
+
+static oplog_buf_t *oplog_borrow_buf(void) {
+    oplog_buf_t *buf = oplog_pop_head(&g_oplog_idle_head, &g_oplog_idle_tail);
+    if (buf) {
+        buf->off = 0;
+        buf->used = 0;
+        return buf;
+    }
+    return oplog_alloc_buf();
+}
+
+static void oplog_free_list(oplog_buf_t **head, oplog_buf_t **tail) {
+    oplog_buf_t *buf;
+    if (!head || !tail) {
+        return;
+    }
+    while ((buf = oplog_pop_head(head, tail)) != NULL) {
+        kvs_free(buf);
+    }
+}
+
+static void oplog_pool_release_all(void) {
+    if (g_oplog_cur) {
+        kvs_free(g_oplog_cur);
+        g_oplog_cur = NULL;
+    }
+    oplog_free_list(&g_oplog_idle_head, &g_oplog_idle_tail);
+    oplog_free_list(&g_oplog_ready_head, &g_oplog_ready_tail);
+}
+
+static void oplog_recycle_done(task_t *t, void *arg) {
+    oplog_buf_t *buf = (oplog_buf_t *)arg;
+    (void)t;
+    if (!buf) {
+        return;
+    }
+    pthread_mutex_lock(&g_oplog_mu);
+    buf->off = 0;
+    buf->used = 0;
+    oplog_push_tail(&g_oplog_idle_head, &g_oplog_idle_tail, buf);
+    pthread_mutex_unlock(&g_oplog_mu);
+}
+
+static int kvs_oplog_submit_ready_buf(oplog_buf_t *buf, int logfd) {
+    void *bufs[1];
+    size_t lens[1];
+    task_t *t;
+
+    if (!buf || logfd < 0) {
+        return -1;
+    }
+    if (buf->used == 0) {
+        oplog_push_tail(&g_oplog_idle_head, &g_oplog_idle_tail, buf);
+        return 0;
+    }
+
+    bufs[0] = (void *)buf->data;
+    lens[0] = buf->used;
+    t = submit_write_ref(&global_uring_ctx, logfd, bufs, lens, 1, buf->off, 0,
+                         oplog_recycle_done, buf);
+    if (!t) {
+        return -1;
+    }
+    return 0;
+}
+
+static int kvs_oplog_flush_internal(int logfd, int force) {
+    oplog_buf_t *buf;
+    if (logfd < 0) {
+        return -1;
+    }
+
+    if (force && g_oplog_cur && g_oplog_cur->used > 0) {
+        oplog_push_tail(&g_oplog_ready_head, &g_oplog_ready_tail, g_oplog_cur);
+        g_oplog_cur = NULL;
+    }
+
+    while ((buf = oplog_pop_head(&g_oplog_ready_head, &g_oplog_ready_tail)) != NULL) {
+        if (kvs_oplog_submit_ready_buf(buf, logfd) < 0) {
+            oplog_push_front(&g_oplog_ready_head, &g_oplog_ready_tail, buf);
+            return -1;
+        }
+    }
+    return 0;
+}
+
+static void make_timeout_1s(struct timespec *ts) {
+    if (!ts) {
+        return;
+    }
+    clock_gettime(CLOCK_REALTIME, ts);
+    ts->tv_sec += 1;
+}
+
+static void *oplog_sync_main(void *arg) {
+    (void)arg;
+    while (1) {
+        uint64_t target_gen = 0;
+        int fd = -1;
+        int flush_ok = 0;
+        struct timespec ts;
+
+        make_timeout_1s(&ts);
+        pthread_mutex_lock(&g_oplog_mu);
+        (void)pthread_cond_timedwait(&g_sync_cv, &g_oplog_mu, &ts);
+
+        if (g_sync_stop) {
+            pthread_mutex_unlock(&g_oplog_mu);
+            break;
+        }
+
+        if (global_cfg.oplog_sync_mode != OPLOG_SYNC_EVERY_SEC ||
+            g_sync_logfd < 0 || g_sync_synced_gen >= g_sync_gen) {
+            pthread_mutex_unlock(&g_oplog_mu);
+            continue;
+        }
+
+        target_gen = g_sync_gen;
+        fd = g_sync_logfd;
+        flush_ok = (kvs_oplog_flush_internal(fd, 1) == 0);
+        pthread_mutex_unlock(&g_oplog_mu);
+
+        if (!flush_ok) {
+            continue;
+        }
+
+        if (kvs_oplog_fsync_all_workers(fd) == 0) {
+            pthread_mutex_lock(&g_oplog_mu);
+            if (g_sync_synced_gen < target_gen) {
+                g_sync_synced_gen = target_gen;
+            }
+            pthread_mutex_unlock(&g_oplog_mu);
+        }
+    }
+    return NULL;
+}
+
+static void oplog_sync_thread_stop_locked(void) {
+    int need_join = g_sync_started;
+
+    if (!need_join) {
+        return;
+    }
+
+    g_sync_stop = 1;
+    pthread_cond_broadcast(&g_sync_cv);
+    pthread_mutex_unlock(&g_oplog_mu);
+    pthread_join(g_sync_th, NULL);
+    pthread_mutex_lock(&g_oplog_mu);
+    g_sync_started = 0;
+    g_sync_logfd = -1;
+}
+
+static int oplog_sync_thread_start_locked(int logfd) {
+    if (g_sync_started) {
+        return 0;
+    }
+    g_sync_stop = 0;
+    g_sync_logfd = logfd;
+    if (pthread_create(&g_sync_th, NULL, oplog_sync_main, NULL) != 0) {
+        g_sync_logfd = -1;
+        return -1;
+    }
+    g_sync_started = 1;
+    return 0;
+}
+
+static int kvs_oplog_append_direct(const uint8_t *cmd, size_t len, int logfd) {
+    uint32_t nlen;
+    void *bufs[2];
+    size_t lens[2];
+    size_t total;
+    off_t myoff;
+    task_t *t;
+
+    nlen = htonl((uint32_t)len);
+    bufs[0] = (void *)&nlen;
+    lens[0] = sizeof(nlen);
+    bufs[1] = (void *)cmd;
+    lens[1] = len;
+    total = sizeof(nlen) + len;
+
+    myoff = g_log_off;
+    g_log_off += (off_t)total;
+    t = submit_write(&global_uring_ctx, logfd, bufs, lens, 2, myoff);
+    if (!t) {
+        return -1;
+    }
+    oplog_mark_dirty_locked();
+    return 0;
+}
 
 int init_cmd_log(const char *file, int *logfd){
+    int rc = 0;
 	if(!file) return -1;
 	int fd = open(file, O_RDWR | O_CREAT , 0644);
 	if(fd < 0) return -2;
 
+    off_t off = lseek(fd, 0, SEEK_END);
+    if (off < 0) {
+        close(fd);
+        return -2;
+    }
 
-    g_log_off = lseek(fd, 0, SEEK_END);
+    pthread_mutex_lock(&g_oplog_mu);
+    oplog_sync_thread_stop_locked();
+    g_log_off = off;
+    g_sync_gen = 0;
+    g_sync_synced_gen = 0;
+    g_sync_logfd = fd;
+    oplog_pool_release_all();
+    if (global_cfg.oplog_sync_mode == OPLOG_SYNC_EVERY_SEC) {
+        rc = oplog_sync_thread_start_locked(fd);
+    }
+    pthread_mutex_unlock(&g_oplog_mu);
+    if (rc != 0) {
+        close(fd);
+        return -3;
+    }
 	*logfd = fd;
 	return 0;
 }
 
 int destroy_cmd_log(int logfd){
-    fsync(logfd);
+    if (logfd < 0) {
+        return -1;
+    }
+
+    pthread_mutex_lock(&g_oplog_mu);
+    oplog_sync_thread_stop_locked();
+    if (kvs_oplog_flush_internal(logfd, 1) < 0) {
+        pthread_mutex_unlock(&g_oplog_mu);
+        return -2;
+    }
+    pthread_mutex_unlock(&g_oplog_mu);
+
+    if (kvs_oplog_fsync_all_workers(logfd) < 0) {
+        return -3;
+    }
+    cleanup_finished_iouring_tasks(&global_uring_ctx);
 	close(logfd);
+
+    pthread_mutex_lock(&g_oplog_mu);
+    oplog_pool_release_all();
+    g_log_off = -1;
+    g_sync_gen = 0;
+    g_sync_synced_gen = 0;
+    g_sync_logfd = -1;
+    pthread_mutex_unlock(&g_oplog_mu);
     global_oplog_fd = -1;
 	return 0;
 }
 
+int kvs_oplog_buffer_append(const uint8_t *cmd, size_t len, int logfd){
+	if (logfd < 0 || !cmd || len == 0) return -1;
+    if (len > UINT32_MAX) return -2;
+    pthread_mutex_lock(&g_oplog_mu);
+    if (g_log_off < 0) {
+        pthread_mutex_unlock(&g_oplog_mu);
+        return -3;
+    }
+
+    {
+        size_t need = sizeof(uint32_t) + len;
+        int became_full = 0;
+        uint32_t nlen = htonl((uint32_t)len);
+
+        if (need > KVS_OPLOG_PAGE_SIZE) {
+            int rc = kvs_oplog_append_direct(cmd, len, logfd);
+            pthread_mutex_unlock(&g_oplog_mu);
+            return (rc == 0) ? KVS_OPLOG_BUF_NOT_FULL : -4;
+        }
+
+        if (!g_oplog_cur) {
+            g_oplog_cur = oplog_borrow_buf();
+            if (!g_oplog_cur) {
+                pthread_mutex_unlock(&g_oplog_mu);
+                return -4;
+            }
+            g_oplog_cur->off = g_log_off;
+        }
+
+        if (g_oplog_cur->used + need > KVS_OPLOG_PAGE_SIZE) {
+            if (g_oplog_cur->used > 0) {
+                oplog_push_tail(&g_oplog_ready_head, &g_oplog_ready_tail, g_oplog_cur);
+                became_full = 1;
+                g_oplog_cur = NULL;
+            }
+
+            g_oplog_cur = oplog_borrow_buf();
+            if (!g_oplog_cur) {
+                pthread_mutex_unlock(&g_oplog_mu);
+                return -4;
+            }
+            g_oplog_cur->off = g_log_off;
+        }
+
+        memcpy(g_oplog_cur->data + g_oplog_cur->used, &nlen, sizeof(nlen));
+        g_oplog_cur->used += sizeof(nlen);
+        memcpy(g_oplog_cur->data + g_oplog_cur->used, cmd, len);
+        g_oplog_cur->used += len;
+        g_log_off += (off_t)need;
+        oplog_mark_dirty_locked();
+
+        if (g_oplog_cur->used == KVS_OPLOG_PAGE_SIZE) {
+            oplog_push_tail(&g_oplog_ready_head, &g_oplog_ready_tail, g_oplog_cur);
+            g_oplog_cur = NULL;
+            became_full = 1;
+        }
+
+        pthread_mutex_unlock(&g_oplog_mu);
+        return became_full ? KVS_OPLOG_BUF_FULL : KVS_OPLOG_BUF_NOT_FULL;
+    }
+}
+
+int kvs_oplog_flush(int logfd, int force) {
+    int rc;
+    pthread_mutex_lock(&g_oplog_mu);
+    rc = kvs_oplog_flush_internal(logfd, force);
+    pthread_mutex_unlock(&g_oplog_mu);
+    if (rc < 0) {
+        return -1;
+    }
+    return 0;
+}
+
 int kvs_oplog_append(const uint8_t *cmd, size_t len, int logfd){
 	if (logfd < 0 || !cmd || len == 0) return -1;
     if (len > UINT32_MAX) return -2;
-
+    pthread_mutex_lock(&g_oplog_mu);
-
+    if (g_log_off < 0) {
-    uint32_t nlen = htonl((uint32_t)len);
+        pthread_mutex_unlock(&g_oplog_mu);
-
+        return -3;
-    void *bufs[2];
-    size_t lens[2];
-
-    bufs[0] = (void *)&nlen;
-    lens[0] = sizeof(nlen);
-
-    bufs[1] = (void *)cmd;
-    lens[1] = len;
-
-    size_t total = sizeof(nlen) + len;
-
-    off_t myoff = g_log_off;
-    g_log_off += (off_t)total;
-
-    task_t *t = submit_write(&global_uring_ctx, logfd, bufs, lens, 2, myoff);
-    if (!t) {
-        return -4;
     }
 
-
+    if (kvs_oplog_flush_internal(logfd, 1) < 0) {
+        pthread_mutex_unlock(&g_oplog_mu);
+        return -4;
+    }
+    if (kvs_oplog_append_direct(cmd, len, logfd) < 0) {
+        pthread_mutex_unlock(&g_oplog_mu);
+        return -4;
+    }
+    pthread_mutex_unlock(&g_oplog_mu);
 	return 0;
 }
 
@@ -69,8 +568,8 @@ int	kvs_replay_log(int logfd){
         uint32_t nlen = 0;
 
         int hr = read_full(logfd, &nlen, sizeof(nlen));
-        if (hr == 0) break;             /* EOF：正常结束 */
+        if (hr == 0) break;
-        if (hr < 0) { return -2; }      /* 半截头 */
+        if (hr < 0) { return -2; }
 
         uint32_t len = ntohl(nlen);
         if (len == 0) { return -3; }
@@ -79,24 +578,21 @@ int	kvs_replay_log(int logfd){
         if (!cmd_bytes ) { return -5; }
 
         int pr = read_full(logfd, cmd_bytes, len);
-        if (pr <= 0) {                  /* 半截 payload */
+        if (pr <= 0) {
             kvs_free(cmd_bytes );
             return -6;
         }
 
-        /* -------- RESP parse -------- */
         resp_cmd_t cmd;
         memset(&cmd, 0, sizeof(cmd));
 
         int clen = resp_parse_one_cmd(cmd_bytes, (int)len, &cmd);
 
         if (clen <= 0 || clen != (int)len) {
-            /* clen==0: need more data，但日志记录必须是一条完整命令，所以视为坏日志 */
             kvs_free(cmd_bytes);
             return -7;
         }
 
-        /* -------- execute -------- */
         resp_value_t outvalue;
         memset(&outvalue, 0, sizeof(outvalue));
 
@@ -106,14 +602,15 @@ int	kvs_replay_log(int logfd){
             return -8;
         }
 
-        /* 注意：
-         * outv 可能引用存储内存，但我们不 build response，因此无需处理。
-         * cmd_bytes 可以释放，因为 cmd slice 指向 cmd_bytes，仅在 dispatch 期间使用。 
-         * */
         kvs_free(cmd_bytes);
     }
 
+    pthread_mutex_lock(&g_oplog_mu);
     g_log_off = lseek(logfd, 0, SEEK_CUR);
+    if (g_sync_synced_gen < g_sync_gen) {
+        g_sync_synced_gen = g_sync_gen;
+    }
+    pthread_mutex_unlock(&g_oplog_mu);
 
     return 0;
 }
@@ -123,8 +620,24 @@ int	kvs_replay_log(int logfd){
  */
 int ksv_clear_log(int logfd){
 	if(logfd < 0) return -1;
+
+    pthread_mutex_lock(&g_oplog_mu);
+    if (kvs_oplog_flush_internal(logfd, 1) < 0) {
+        pthread_mutex_unlock(&g_oplog_mu);
+        return -2;
+    }
+    pthread_mutex_unlock(&g_oplog_mu);
+    if (kvs_oplog_fsync_all_workers(logfd) < 0) {
+        return -3;
+    }
+    cleanup_finished_iouring_tasks(&global_uring_ctx);
 	ftruncate(logfd, 0);
 	lseek(logfd, 0, SEEK_SET);
+
+    pthread_mutex_lock(&g_oplog_mu);
     g_log_off = 0;
+    g_sync_gen = 0;
+    g_sync_synced_gen = 0;
+    pthread_mutex_unlock(&g_oplog_mu);
 	return 0;
 }

dump/kvs_snapshot.c

@@ -40,6 +40,24 @@ int kvs_create_snapshot(iouring_ctx_t *uring, const char* array_file, const char
 	return ret;
 }
 
+int kvs_create_snapshot_async_1(iouring_ctx_t *uring, const char* array_file, const char* rbtree_file, const char* hash_file){
+	pid_t pid = fork();
+	if (pid == -1) { perror("fork"); return -1; }
+
+	if (pid == 0) {
+		int ret = kvs_create_snapshot(uring, array_file, rbtree_file, hash_file);
+		if (ret != 0) {
+            fprintf(stderr, "snapshot creation failed\n");
+            _exit(1);
+        }
+
+		_exit(0);
+	} else {
+
+		return 0;
+	}
+}
+
 static int send_file_to_ipport(const char *ip, int port, const char *filename) {
 	int sockfd = socket(AF_INET, SOCK_STREAM, 0);
 	if (sockfd < 0) { perror("socket"); return -1; }
@@ -86,7 +104,7 @@ static int send_file_to_ipport(const char *ip, int port, const char *filename) {
     return 0;
 }
 
-int kvs_create_snapshot_async(const char *ip, int port){
+int kvs_create_snapshot_async_2(const char *ip, int port){
 	pid_t pid = fork();
 	if (pid == -1) { perror("fork"); return -1; }
 

ebpf/c/Makefile

@@ -1,12 +1,12 @@
 # SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause)
 OUTPUT := .output
 CLANG ?= clang
-LIBBPF_SRC := $(abspath ../../libbpf/src)
+LIBBPF_SRC := $(abspath ../../libbpf-bootstrap/libbpf/src)
-BPFTOOL_SRC := $(abspath ../../bpftool/src)
+BPFTOOL_SRC := $(abspath ../../libbpf-bootstrap/bpftool/src)
 LIBBPF_OBJ := $(abspath $(OUTPUT)/libbpf.a)
 BPFTOOL_OUTPUT ?= $(abspath $(OUTPUT)/bpftool)
 BPFTOOL ?= $(BPFTOOL_OUTPUT)/bootstrap/bpftool
-LIBBLAZESYM_SRC := $(abspath ../../blazesym/)
+LIBBLAZESYM_SRC := $(abspath ../../libbpf-bootstrap/blazesym/)
 LIBBLAZESYM_INC := $(abspath $(LIBBLAZESYM_SRC)/capi/include)
 LIBBLAZESYM_OBJ := $(abspath $(OUTPUT)/libblazesym_c.a)
 ARCH ?= $(shell uname -m | sed 's/x86_64/x86/' \
@@ -16,11 +16,11 @@ ARCH ?= $(shell uname -m | sed 's/x86_64/x86/' \
 			 | sed 's/mips.*/mips/' \
 			 | sed 's/riscv64/riscv/' \
 			 | sed 's/loongarch64/loongarch/')
-VMLINUX := ../../vmlinux.h/include/$(ARCH)/vmlinux.h
+VMLINUX := ../../libbpf-bootstrap/vmlinux.h/include/$(ARCH)/vmlinux.h
 # Use our own libbpf API headers and Linux UAPI headers distributed with
 # libbpf to avoid dependency on system-wide headers, which could be missing or
 # outdated
-INCLUDES := -I$(OUTPUT) -I../../libbpf/include/uapi -I$(dir $(VMLINUX)) -I$(LIBBLAZESYM_INC)
+INCLUDES := -I$(OUTPUT) -I../../libbpf-bootstrap/libbpf/include/uapi -I$(dir $(VMLINUX)) -I$(LIBBLAZESYM_INC)
 CFLAGS := -g -Wall
 ALL_LDFLAGS := $(LDFLAGS) $(EXTRA_LDFLAGS)
 

ebpf/c/replica.c

@@ -189,6 +189,7 @@ static void* reader_thread_func(void *arg)
         uint64_t last = __atomic_load_n(&g_shm.hdr->last_seq, __ATOMIC_ACQUIRE);
         if (local_seq > last) {
             // 没有新数据，短暂休眠避免空转
+            usleep(500);
             continue;
         }
         if (read_off+ sizeof(replica_rec_hdr_t) >= g_shm.hdr->capacity) {
@@ -224,8 +225,7 @@ static void* reader_thread_func(void *arg)
         
         // 序列号检查
         if (h.seq != local_seq) {
-            DEBUGLOG("Reader: seq mismatch! h.seq=%lu, local_seq=%lu, off=%u\n",
+            // DEBUGLOG("Reader: seq mismatch! h.seq=%lu, local_seq=%lu, off=%u\n",  h.seq, local_seq, read_off);
-                     h.seq, local_seq, read_off);
             continue;
         }
         

ebpf/leak_detect/complete.bpf

@@ -0,0 +1,39 @@
+#!/usr/bin/env bpftrace
+
+BEGIN
+{
+    printf("开始统计 kvstore 进程的 __completed_cmd 调用次数...\n");
+    printf("每 5 秒打印一次统计，Ctrl-C 退出\n\n");
+
+    // 统计变量
+    @enter = 0;
+    @exit = 0;
+}
+
+interval:s:5
+{
+    time("%H:%M:%S");
+    printf("  __completed_enter_cmd 调用次数: %10d\n", @enter);
+    printf("  __completed_exit_cmd  调用次数: %10d\n", @exit);
+
+    // 可选：如果想每轮清零统计，取消下面注释
+    // clear(@enter);
+    // clear(@exit);
+}
+
+uprobe:/home/lian/share/9.1-kvstore/kvstore:__completed_cmd
+{
+    @exit++;
+}
+
+uretprobe:/home/lian/share/9.1-kvstore/kvstore:__completed_cmd
+{
+    @enter++;
+}
+
+
+
+END
+{
+    printf("\n最终统计：\n");
+}

ebpf/leak_detect/tcp_probe.bpf

@@ -0,0 +1,39 @@
+#!/usr/bin/env bpftrace
+
+BEGIN
+{
+    printf("开始统计 kvstore 进程的 tcp_rcv_established 调用次数...\n");
+    printf("每 5 秒打印一次统计，Ctrl-C 退出\n\n");
+
+    // 统计变量
+    @enter = 0;
+    @exit = 0;
+}
+
+interval:s:5
+{
+    time("%H:%M:%S");
+    printf("  tcp_rcv_established       调用次数: %10d\n", @enter);
+    printf("  tcp_rcv_established ret   调用次数: %10d\n", @exit);
+
+    // 可选：如果想每轮清零统计，取消下面注释
+    // clear(@enter);
+    // clear(@exit);
+}
+
+kprobe:tcp_rcv_established
+{
+    @enter++;
+}
+
+kretprobe:tcp_rcv_established
+{
+    @exit++;
+}
+
+END
+{
+    printf("\n最终统计：\n");
+    printf("tcp_rcv_established    : %d 次\n", @enter);
+    printf("tcp_rcv_established ret: %d 次\n", @exit);
+}

ebpf/leak_detect/tracepoint.bpf

@@ -0,0 +1,41 @@
+#!/usr/bin/env bpftrace
+
+BEGIN
+{
+    printf("开始统计 kvstore 进程的 recvfrom 调用次数...\n");
+    printf("每 5 秒打印一次统计，Ctrl-C 退出\n\n");
+
+    // 统计变量
+    @enter = 0;
+    @exit = 0;
+}
+
+interval:s:5
+{
+    time("%H:%M:%S");
+    printf("  sys_enter_recvfrom 调用次数: %10d\n", @enter);
+    printf("  sys_exit_recvfrom  调用次数: %10d\n", @exit);
+
+    // 可选：如果想每轮清零统计，取消下面注释
+    // clear(@enter);
+    // clear(@exit);
+}
+
+tracepoint:syscalls:sys_enter_recvfrom
+/comm == "kvstore"/
+{
+    @enter++;
+}
+
+tracepoint:syscalls:sys_exit_recvfrom
+/comm == "kvstore"/
+{
+    @exit++;
+}
+
+END
+{
+    printf("\n最终统计：\n");
+    printf("sys_enter_recvfrom: %d 次\n", @enter);
+    printf("sys_exit_recvfrom : %d 次\n", @exit);
+}

ebpf/old.c/Makefile

@@ -1,12 +1,12 @@
 # SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause)
 OUTPUT := .output
 CLANG ?= clang
-LIBBPF_SRC := $(abspath ../../libbpf/src)
+LIBBPF_SRC := $(abspath ../../libbpf-bootstrap/libbpf/src)
-BPFTOOL_SRC := $(abspath ../../bpftool/src)
+BPFTOOL_SRC := $(abspath ../../libbpf-bootstrap/bpftool/src)
 LIBBPF_OBJ := $(abspath $(OUTPUT)/libbpf.a)
 BPFTOOL_OUTPUT ?= $(abspath $(OUTPUT)/bpftool)
 BPFTOOL ?= $(BPFTOOL_OUTPUT)/bootstrap/bpftool
-LIBBLAZESYM_SRC := $(abspath ../../blazesym/)
+LIBBLAZESYM_SRC := $(abspath ../../libbpf-bootstrap/blazesym/)
 LIBBLAZESYM_INC := $(abspath $(LIBBLAZESYM_SRC)/capi/include)
 LIBBLAZESYM_OBJ := $(abspath $(OUTPUT)/libblazesym_c.a)
 ARCH ?= $(shell uname -m | sed 's/x86_64/x86/' \
@@ -16,11 +16,11 @@ ARCH ?= $(shell uname -m | sed 's/x86_64/x86/' \
 			 | sed 's/mips.*/mips/' \
 			 | sed 's/riscv64/riscv/' \
 			 | sed 's/loongarch64/loongarch/')
-VMLINUX := ../../vmlinux.h/include/$(ARCH)/vmlinux.h
+VMLINUX := ../../libbpf-bootstrap/vmlinux.h/include/$(ARCH)/vmlinux.h
 # Use our own libbpf API headers and Linux UAPI headers distributed with
 # libbpf to avoid dependency on system-wide headers, which could be missing or
 # outdated
-INCLUDES := -I$(OUTPUT) -I../../libbpf/include/uapi -I$(dir $(VMLINUX)) -I$(LIBBLAZESYM_INC)
+INCLUDES := -I$(OUTPUT) -I../../libbpf-bootstrap/libbpf/include/uapi -I$(dir $(VMLINUX)) -I$(LIBBLAZESYM_INC)
 CFLAGS := -g -Wall
 ALL_LDFLAGS := $(LDFLAGS) $(EXTRA_LDFLAGS)
 

ebpf/old.c/replica.bpf.c

@@ -1,80 +1,133 @@
-// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
-/* Copyright (c) 2020 Facebook */
 #include "vmlinux.h"
 #include <bpf/bpf_helpers.h>
 #include <bpf/bpf_tracing.h>
 #include <bpf/bpf_core_read.h>
+#include <bpf/bpf_endian.h>
 
 #include "replica.h"
 
 char LICENSE[] SEC("license") = "Dual BSD/GPL";
 
+#define FLAG_SSYNC_HAPPENED 0
+#define TARGET_PORT 8888
+
+/* ================= BPF Maps ================= */
+
 struct {
-    __uint(type, BPF_MAP_TYPE_PERF_EVENT_ARRAY);
+    __uint(type, BPF_MAP_TYPE_ARRAY);
-    __uint(key_size, sizeof(int));
+    __type(key, __u32);
-    __uint(value_size, sizeof(int));
+    __type(value, __u32);
-} events SEC(".maps");
+    __uint(max_entries, 1);
+} flags SEC(".maps");
 
-/* __completed_cmd(const uint8_t *cmd, size_t len, unsigned long long seq); */
+struct {
-SEC("uprobe//home/lian/share/9.1-kvstore/kvstore:__completed_cmd")
+    __uint(type, BPF_MAP_TYPE_RINGBUF);
-int BPF_KPROBE(handle_completed_cmd,
+    __uint(max_entries, 1 << 26);  // 64MB
-               const __u8 *cmd, size_t len, __u64 seq)
+} rb SEC(".maps");
+
+/* ================= Helper Functions ================= */
+
+// 无需 process filter，改用 socket port filter
+
+/* ================= Kernel Hooks (TCP Layer) ================= */
+
+/*
+ * 使用 kprobe 挂载 tcp_rcv_established
+ * 此时 skb 包含完整的 TCP 包（Header + Payload），数据在内核态。
+ */
+SEC("kprobe/tcp_rcv_established")
+int BPF_KPROBE(trace_tcp_rcv, struct sock *sk, struct sk_buff *skb)
 {
-    struct replica_event evt = {};
+    // 1. 检查 SSYNC 标志是否已开启 (只在全量同步后开始抓包)
-    __u32 copy_len;
+    __u32 flag_key = FLAG_SSYNC_HAPPENED;
+    __u32 *flag_val = bpf_map_lookup_elem(&flags, &flag_key);
+    if (!flag_val || *flag_val == 0)
+        return 0;
 
-    evt.type = EVENT_COMPLETED_CMD;
+    // 2. 过滤端口 8888
-    evt.complete.seq = seq;
+    // sk->sk_num 存储的是 Host Byte Order 的本地端口
+    __u16 lport = BPF_CORE_READ(sk, __sk_common.skc_num);
+    if (lport != TARGET_PORT)
+        return 0;
 
-    copy_len = len;
+    // 3. 计算数据长度
-    if (copy_len > MAX_CMD_LEN)
+    // 在 tcp_rcv_established 中，skb->len 是 (TCP Header + Data) 的长度
-        copy_len = MAX_CMD_LEN;
+    // skb->data 指向 TCP Header 的起始位置
+    unsigned int skb_len = BPF_CORE_READ(skb, len);
     
-    evt.complete.len = copy_len;
+    // 读取 TCP Header 长度 (doff 字段，单位是 4字节)
+    // 需要读取 skb->data 指向的内存的前几个字节来获取 doff
+    unsigned char *skb_data = BPF_CORE_READ(skb, data);
     
-    if (cmd)
+    // 读取 TCP Header 的第 12 个字节 (包含 Data Offset)
-        bpf_probe_read_user(evt.complete.cmd, copy_len, cmd);
+    // Offset 12: Data Offset (4 bits) | Reserved (3 bits) | NS (1 bit)
+    unsigned char doff_byte;
+    if (bpf_probe_read_kernel(&doff_byte, 1, skb_data + 12) < 0)
+        return 0;
     
-    bpf_perf_event_output(ctx, &events,
+    unsigned int tcp_hdr_len = (doff_byte >> 4) * 4;
-                          BPF_F_CURRENT_CPU,
+    
-                          &evt, sizeof(evt));
+    // 计算 Payload 长度
+    if (skb_len <= tcp_hdr_len)
+        return 0; // 只有 ACK 没有数据，或者是控制包
+
+    unsigned int payload_len = skb_len - tcp_hdr_len;
+
+    // 4. 准备 RingBuffer 数据
+    struct replica_event *e = bpf_ringbuf_reserve(&rb, sizeof(*e), 0);
+    if (!e)
+        return 0;
+
+    e->type = EVENT_COMPLETED_CMD;
+    
+    // 截断超长数据
+    if (payload_len > MAX_CMD_LEN)
+        e->complete.len = MAX_CMD_LEN;
+    else
+        e->complete.len = payload_len;
+
+    // 5. 核心修改：使用 bpf_probe_read_kernel 读取数据
+    // 数据起始位置 = skb->data + tcp_hdr_len
+    if (bpf_probe_read_kernel(&e->complete.cmd[0], e->complete.len, skb_data + tcp_hdr_len) < 0) {
+        bpf_ringbuf_discard(e, 0);
+        return 0;
+    }
+
+    bpf_ringbuf_submit(e, 0);
     return 0;
 }
 
-/* __ssync(const uint8_t *ip, uint32_t ip_len, int port, unsigned long long seq); */
+/* ================= Uprobe Hooks================= */
+
 SEC("uprobe//home/lian/share/9.1-kvstore/kvstore:__ssync")
 int BPF_KPROBE(handle_ssync,
-               const __u8 *ip, __u32 ip_len, int port, __u64 seq)
+               const __u8 *ip, __u32 ip_len, int port, __u64 seq_unused)
 {
-    struct replica_event evt = {};
+    __u32 key = FLAG_SSYNC_HAPPENED;
+    __u32 val = 1;
+    bpf_map_update_elem(&flags, &key, &val, BPF_ANY);
 
-    evt.type = EVENT_SSYNC;
+    struct replica_event *e = bpf_ringbuf_reserve(&rb, sizeof(*e), 0);
-    evt.sync.seq = seq;
+    if (!e) return 0;
-    evt.sync.port = port;
+
+    e->type = EVENT_SSYNC;
+    e->sync.port = port;
 
     __u32 copy_len = ip_len;
-    if (copy_len > sizeof(evt.sync.ip))
+    if (copy_len > sizeof(e->sync.ip)) copy_len = sizeof(e->sync.ip);
-        copy_len = sizeof(evt.sync.ip);
+    if (ip) bpf_probe_read_user(e->sync.ip, copy_len, ip);
 
-    if (ip)
+    bpf_ringbuf_submit(e, 0);
-        bpf_probe_read_user(evt.sync.ip, copy_len, ip);
-
-    bpf_perf_event_output(ctx, &events,
-                          BPF_F_CURRENT_CPU,
-                          &evt, sizeof(evt));
     return 0;
 }
 
-/* __sready(void); */
 SEC("uprobe//home/lian/share/9.1-kvstore/kvstore:__sready")
 int BPF_KPROBE(handle_sready)
 {
-    struct replica_event evt = {};
+    struct replica_event *e = bpf_ringbuf_reserve(&rb, sizeof(*e), 0);
+    if (!e) return 0;
 
-    evt.type = EVENT_SREADY;
+    e->type = EVENT_SREADY;
-
+    bpf_ringbuf_submit(e, 0);
-    bpf_perf_event_output(ctx, &events,
-                          BPF_F_CURRENT_CPU,
-                          &evt, sizeof(evt));
     return 0;
 }

ebpf/old.c/replica.c

@@ -1,5 +1,3 @@
-// SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause)
-/* Copyright (c) 2020 Facebook */
 #include <stdio.h>
 #include <unistd.h>
 #include <stdlib.h>
@@ -12,6 +10,7 @@
 #include <arpa/inet.h>
 #include <sys/epoll.h>
 #include <fcntl.h>
+#include <errno.h>
 
 #include "replica.h"
 
@@ -20,10 +19,9 @@
 typedef enum {
     OFFLINE = 0,
     ONLINE  = 1,
-}replica_state_e ;
+} replica_state_e;
 
 struct cmd_node {
-    __u64 seq;
     __u32 len;
     uint8_t *cmd;
     struct cmd_node *next;
@@ -32,7 +30,7 @@ struct cmd_node {
 struct pending_queue {
     struct cmd_node *head;
     struct cmd_node *tail;
-	int count;
+    int count;
 };
 
 /* ================= 全局状态 ================= */
@@ -43,7 +41,6 @@ static int epollfd = -1;
 
 static char peer_ip[MAX_IP_LEN];
 static int peer_port;
-static __u64 peer_seq;
 
 static struct pending_queue pending = {
     .head = NULL,
@@ -66,7 +63,7 @@ static void pending_free()
     q->count = 0;
 }
 
-static void pending_push(__u64 seq, __u32 len, const uint8_t *cmd)
+static void pending_push(__u32 len, const uint8_t *cmd)
 {
     struct cmd_node *node = malloc(sizeof(*node));
     if (!node)
@@ -79,7 +76,6 @@ static void pending_push(__u64 seq, __u32 len, const uint8_t *cmd)
     }
 
     memcpy(node->cmd, cmd, len);
-    node->seq  = seq;
     node->len  = len;
     node->next = NULL;
 
@@ -93,72 +89,66 @@ static void pending_push(__u64 seq, __u32 len, const uint8_t *cmd)
     pending.count++;
 }
 
-static void pending_gc(__u64 min_seq)
+static long long int sendn = 0;
-{
-    struct cmd_node *cur = pending.head;
-
-    int n = pending.count;
-    while (cur && cur->seq < min_seq) {
-        struct cmd_node *tmp = cur;
-        cur = cur->next;
-
-        free(tmp->cmd);
-        free(tmp);
-        pending.count--;
-    }
-
-    DEBUGLOG("gc:%d\n", n-pending.count);
-
-    pending.head = cur;
-    if (!cur)
-        pending.tail = NULL;
-}
-
 static void pending_send_all(void)
 {
     struct cmd_node *cur = pending.head;
-    while (cur) {
+    int need_out = 0;
-        int rt = send(sockfd, cur->cmd, cur->len, 0);
+    int sent_count = 0;
+    const int MAX_BATCH = 100;  // 批量发送上限,避免阻塞过久
 
-        if(rt == (int)cur->len){
+    while (cur && sent_count < MAX_BATCH) {
+        // 使用 MSG_MORE 合并多个小包
+        int flags = (cur->next && sent_count < MAX_BATCH - 1) ? MSG_MORE : 0;
+        int rt = send(sockfd, cur->cmd, cur->len, flags);
+
+        if (rt == (int)cur->len) {
+            sendn += rt;
+            printf("%s\n", cur->cmd);
             struct cmd_node *tmp = cur;
             cur = cur->next;
-
             free(tmp->cmd);
             free(tmp);
             pending.count--;
-        }else{
+            pending.head = cur;
-            DEBUGLOG("error\n");
+            sent_count++;
-            // 失败：不移动 cur，直接 break
+        } else if (rt > 0) {
-            if (rt < 0) {
+            sendn += rt;
+            memmove(cur->cmd, cur->cmd + rt, cur->len - rt);
+            cur->len -= rt;
+            need_out = 1;
+            break;
+        } else {
+            if (errno == EAGAIN || errno == EWOULDBLOCK) {
+                need_out = 1;
+                break;
+            } else {
                 perror("send failed");
-                if (errno == ECONNRESET || errno == EPIPE) {
+                state = OFFLINE;
-                    state = OFFLINE;
-                    if (sockfd >= 0) {
-                        close(sockfd);
-                        sockfd = -1;
-                        DEBUGLOG("connect closed\n");
-                    }
-                } else if (rt == 0) {
-                    fprintf(stderr, "send returned 0 (peer closed?)\n");
-                } else {
-                    fprintf(stderr, "partial send: %d/%u\n", rt, cur->len);
-                }
-
                 break;
             }
         }
     }
+    DEBUGLOG("sendn :%lld\n", sendn);
 
     pending.head = cur;
-    if(!cur) 
+    if (!cur) pending.tail = NULL;
-        pending.tail = NULL;
+
+    if (sockfd >= 0 && state == ONLINE) {
+        struct epoll_event ev = {0};
+        ev.data.fd = sockfd;
+        ev.events = EPOLLIN;
+        if (need_out || pending.head) {
+            ev.events |= EPOLLOUT;
+        }
+        epoll_ctl(epollfd, EPOLL_CTL_MOD, sockfd, &ev);
+    }
 }
 
 /* ================= 网络逻辑 ================= */
 static void try_connect(void)
 {
-    if(sockfd > 0){
+    if (sockfd > 0) {
         close(sockfd);
         sockfd = -1;
     }
@@ -170,14 +160,14 @@ static void try_connect(void)
     addr.sin_port = htons(peer_port);
     inet_pton(AF_INET, peer_ip, &addr.sin_addr);
 
-    for(i = 0;i < 10; ++ i){
+    for (i = 0; i < 10; ++i) {
         sockfd = socket(AF_INET, SOCK_STREAM, 0);
         if (sockfd < 0) {
             perror("socket");
             return;
         }
 
-        DEBUGLOG("connect try %d...\n", i + 1);
+        DEBUGLOG("connect try %d... %s:%d\n", i + 1, peer_ip, peer_port);
         if (connect(sockfd, (struct sockaddr *)&addr, sizeof(addr)) == 0) {
             DEBUGLOG("connect success: %s:%d\n", peer_ip, peer_port);
 
@@ -190,7 +180,10 @@ static void try_connect(void)
             epoll_ctl(epollfd, EPOLL_CTL_ADD, sockfd, &ev);
 
             state = ONLINE;
-            pending_send_all();
+            if (pending.head) {
+                ev.events = EPOLLIN | EPOLLOUT;
+                epoll_ctl(epollfd, EPOLL_CTL_MOD, sockfd, &ev);
+            }
             return;
         }
 
@@ -234,17 +227,10 @@ static void handle_socket_readable(void)
 static void handle_socket_writable(void)
 {
     pending_send_all();
-    if (!pending.head) {
-        struct epoll_event ev;
-        ev.events = EPOLLIN;  // 只监听读
-        ev.data.fd = sockfd;
-        epoll_ctl(epollfd, EPOLL_CTL_MOD, sockfd, &ev);
-    }
 }
 
-
+/* ================= ring buffer 回调 ================= */
-/* ================= perf buffer 回调 ================= */
+static int handle_event(void *ctx, void *data, size_t size)
-static void handle_event(void *ctx, int cpu, void *data, __u32 size)
 {
     struct replica_event *evt = data;
     switch (evt->type) {
@@ -252,20 +238,18 @@ static void handle_event(void *ctx, int cpu, void *data, __u32 size)
     case EVENT_SSYNC:
         strncpy(peer_ip, evt->sync.ip, sizeof(peer_ip));
         peer_port = evt->sync.port;
-        peer_seq  = evt->sync.seq;
+        DEBUGLOG("SSYNC [%s:%d]\n", peer_ip, peer_port);
-        DEBUGLOG("SSYNC [seq:%lld], [%s:%d]\n", peer_seq, peer_ip, peer_port);
-
         state = OFFLINE;
-        pending_gc(peer_seq);
         break;
 
     case EVENT_COMPLETED_CMD:
-        // DEBUGLOG("CMD [seq:%lld], cmd:\n[\n%s]\n", evt->complete.seq, evt->complete.cmd);
+        // 这里收到的可能是半个命令，或者是多个命令的粘包
-        pending_push(evt->complete.seq,
+        // 但对于转发器来说，只是字节流，直接 push 即可
-                     evt->complete.len,
+        if (evt->complete.len > 0) {
-                     evt->complete.cmd);
+            pending_push(evt->complete.len, evt->complete.cmd);
+        }
 
-        if (state == ONLINE && sockfd >= 0) {
+        if (state == ONLINE && sockfd >= 0 && pending.head) {
             struct epoll_event ev;
             ev.events = EPOLLIN | EPOLLOUT;
             ev.data.fd = sockfd;
@@ -274,82 +258,80 @@ static void handle_event(void *ctx, int cpu, void *data, __u32 size)
         break;
 
     case EVENT_SREADY:
-        DEBUGLOG("SREADY \n");
+        DEBUGLOG("SREADY\n");
         if (state == OFFLINE)
             try_connect();
         break;
     }
+    return 0;
 }
 
-/* ================= main ================= */
 int main(int argc, char **argv)
 {
-	struct replica_bpf *skel;
+    struct replica_bpf *skel;
-    struct perf_buffer *pb = NULL;
+    struct ring_buffer *rb = NULL;
-	int err;
+    int err;
 
-	/* Open BPF application */
+    // 提高 rlimit 以允许加载 BPF
-	skel = replica_bpf__open();
+    struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
-	if (!skel) {
+    setrlimit(RLIMIT_MEMLOCK, &r);
-		fprintf(stderr, "Failed to open BPF skeleton\n");
-		return 1;
-	}
 
-	/* Load & verify BPF programs */
+    skel = replica_bpf__open();
-	err = replica_bpf__load(skel);
+    if (!skel) {
-	if (err) {
+        fprintf(stderr, "Failed to open BPF skeleton\n");
-		fprintf(stderr, "Failed to load and verify BPF skeleton\n");
+        return 1;
-		goto cleanup;
+    }
-	}
 
-	/* Attach tracepoint handler */
+    err = replica_bpf__load(skel);
-	err = replica_bpf__attach(skel);
+    if (err) {
-	if (err) {
+        fprintf(stderr, "Failed to load BPF skeleton\n");
-		fprintf(stderr, "Failed to attach BPF skeleton\n");
-		goto cleanup;
-	}
-
-	printf("Successfully started! \n");
-
-
-	pb = perf_buffer__new(bpf_map__fd(skel->maps.events), 8,
-                 handle_event, NULL, NULL, NULL);
-
-	if(!pb){
-		goto cleanup;
-	}
-
-    epollfd = epoll_create1(0);
-    if (epollfd < 0) {
-        fprintf(stderr, "epoll_create1 failed\n");
         goto cleanup;
     }
 
-	while (1) {
+    err = replica_bpf__attach(skel);
+    if (err) {
+        fprintf(stderr, "Failed to attach BPF skeleton\n");
+        goto cleanup;
+    }
+
+    printf("Successfully started! Monitoring TCP port 8888 (Kernel Side)...\n");
+
+    rb = ring_buffer__new(bpf_map__fd(skel->maps.rb), handle_event, NULL, NULL);
+    if (!rb) {
+        fprintf(stderr, "Failed to create ring buffer\n");
+        goto cleanup;
+    }
+
+    epollfd = epoll_create1(0);
+    // ... (主循环保持不变) ...
+    
+    // 主循环建议：
+    while (1) {
         struct epoll_event events[10];
         
-        perf_buffer__poll(pb, 1000);  // 处理事件
+        // 既然追求性能，Polling 依然是必要的
+        // 10ms 的延迟对于 RingBuffer 消费是可以接受的
+        int poll_timeout = (state == ONLINE) ? 10 : 100;
         
-        if(OFFLINE) continue;
+        ring_buffer__poll(rb, poll_timeout);
+        
+        if (state == OFFLINE) continue;
 
         int nfds = epoll_wait(epollfd, events, 10, 0);
         for (int i = 0; i < nfds; i++) {
             if (events[i].data.fd == sockfd) {
-                if (events[i].events & EPOLLIN) {
+                if (events[i].events & EPOLLIN) handle_socket_readable();
-                    handle_socket_readable();  // 快速消费接收数据
+                if (events[i].events & EPOLLOUT) handle_socket_writable();
-                }
-                if (events[i].events & EPOLLOUT) {
-                    handle_socket_writable();  // 发送数据
-                }
             }
         }
     }
 
-	perf_buffer__free(pb);
-
 cleanup:
-	pending_free();
+    // ... (清理代码保持不变) ...
-	if (sockfd >= 0) close(sockfd);
+    if (rb) ring_buffer__free(rb);
-	replica_bpf__destroy(skel);
+    pending_free();
-	return -err;
+    if (sockfd >= 0) close(sockfd);
+    if (epollfd >= 0) close(epollfd);
+    replica_bpf__destroy(skel);
+    return -err;
 }

ebpf/old.c/replica.h

@@ -1,24 +1,21 @@
 #ifndef __REPLICA_H__
 #define __REPLICA_H__
 
-
+#define MAX_CMD_LEN 4096
-#define MAX_CMD_LEN 256
 #define MAX_IP_LEN  64
 
 enum event_type {
-    EVENT_COMPLETED_CMD,
+    EVENT_COMPLETED_CMD = 1,
-    EVENT_SSYNC,
+    EVENT_SSYNC         = 2,
-    EVENT_SREADY,
+    EVENT_SREADY        = 3,
 };
 
 struct complete_cmd_evt {
-    __u64 seq;
     __u32 len;
     __u8  cmd[MAX_CMD_LEN];
 };
 
 struct sync_evt {
-    __u64 seq;
     char  ip[MAX_IP_LEN];
     __s32 port;
 };
@@ -33,5 +30,4 @@ struct replica_event {
     };
 };
 
-
 #endif

hash.c

@@ -1,228 +0,0 @@
-
-
-
-// #include <stdio.h>
-// #include <string.h>
-// #include <stdlib.h>
-// #include <pthread.h>
-
-
-
-// #define MAX_KEY_LEN	128
-// #define MAX_VALUE_LEN	512
-// #define MAX_TABLE_SIZE	1024
-
-
-
-// typedef struct hashnode_s {
-
-// 	char key[MAX_KEY_LEN];
-// 	char value[MAX_VALUE_LEN];
-	
-// 	struct hashnode_s *next;
-	
-// } hashnode_t;
-
-
-// typedef struct hashtable_s {
-
-// 	hashnode_t **nodes; //* change **, 
-
-// 	int max_slots;
-// 	int count;
-
-// 	pthread_mutex_t lock;
-
-// } hashtable_t;
-
-// hashtable_t hash;
-
-
-// //Connection 
-// // 'C' + 'o' + 'n'
-// static int _hash(char *key, int size) {
-
-// 	if (!key) return -1;
-
-// 	int sum = 0;
-// 	int i = 0;
-
-// 	while (key[i] != 0) {
-// 		sum += key[i];
-// 		i ++;
-// 	}
-
-// 	return sum % size;
-
-// }
-
-// hashnode_t *_create_node(char *key, char *value) {
-
-// 	hashnode_t *node = (hashnode_t*)malloc(sizeof(hashnode_t));
-// 	if (!node) return NULL;
-	
-// 	strncpy(node->key, key, MAX_KEY_LEN);
-// 	strncpy(node->value, value, MAX_VALUE_LEN);
-// 	node->next = NULL;
-
-// 	return node;
-// }
-
-
-// //
-// int init_hashtable(hashtable_t *hash) {
-
-// 	if (!hash) return -1;
-
-// 	hash->nodes = (hashnode_t**)malloc(sizeof(hashnode_t*) * MAX_TABLE_SIZE);
-// 	if (!hash->nodes) return -1;
-
-// 	hash->max_slots = MAX_TABLE_SIZE;
-// 	hash->count = 0; 
-
-// 	pthread_mutex_init(&hash->lock, NULL);
-
-// 	return 0;
-// }
-
-// // 
-// void dest_hashtable(hashtable_t *hash) {
-
-// 	if (!hash) return;
-
-// 	int i = 0;
-// 	for (i = 0;i < hash->max_slots;i ++) {
-// 		hashnode_t *node = hash->nodes[i];
-
-// 		while (node != NULL) { // error
-
-// 			hashnode_t *tmp = node;
-// 			node = node->next;
-// 			hash->nodes[i] = node;
-			
-// 			free(tmp);
-			
-// 		}
-// 	}
-
-// 	free(hash->nodes);
-	
-// }
-
-
-
-// // mp
-// int put_kv_hashtable(hashtable_t *hash, char *key, char *value) {
-
-// 	if (!hash || !key || !value) return -1;
-
-// 	int idx = _hash(key, MAX_TABLE_SIZE);
-
-// 	pthread_mutex_lock(&hash->lock);
-
-// 	hashnode_t *node = hash->nodes[idx];
-// #if 1
-// 	while (node != NULL) {
-// 		if (strcmp(node->key, key) == 0) { // exist
-// 			pthread_mutex_unlock(&hash->lock);
-// 			return 1;
-// 		}
-// 		node = node->next;
-// 	}
-// #endif
-
-// 	hashnode_t *new_node = _create_node(key, value);
-// 	new_node->next = hash->nodes[idx];
-// 	hash->nodes[idx] = new_node;
-	
-// 	hash->count ++;
-
-// 	pthread_mutex_unlock(&hash->lock);
-
-// 	return 0;
-// }
-
-
-// char * get_kv_hashtable(hashtable_t *hash, char *key) {
-
-// 	if (!hash || !key) return NULL;
-
-// 	int idx = _hash(key, MAX_TABLE_SIZE);
-
-// 	pthread_mutex_lock(&hash->lock);
-// 	hashnode_t *node = hash->nodes[idx];
-
-// 	while (node != NULL) {
-
-// 		if (strcmp(node->key, key) == 0) {
-// 			pthread_mutex_unlock(&hash->lock);
-// 			return node->value;
-// 		}
-
-// 		node = node->next;
-// 	}
-
-// 	pthread_mutex_unlock(&hash->lock);
-
-// 	return NULL;
-
-// }
-
-
-// int count_kv_hashtable(hashtable_t *hash) {
-// 	return hash->count;
-// }
-
-// int delete_kv_hashtable(hashtable_t *hash, char *key) {
-// 	if (!hash || !key) return -2;
-
-// 	int idx = _hash(key, MAX_TABLE_SIZE);
-
-// 	pthread_mutex_lock(&hash->lock);
-// 	hashnode_t *head = hash->nodes[idx];
-// 	if (head == NULL) return -1; // noexist
-// 	// head node
-// 	if (strcmp(head->key, key) == 0) {
-// 		hashnode_t *tmp = head->next;
-// 		hash->nodes[idx] = tmp;
-		
-// 		free(head);
-// 		hash->count --;
-// 		pthread_mutex_unlock(&hash->lock);
-		
-// 		return 0;
-// 	}
-
-// 	hashnode_t *cur = head;
-// 	while (cur->next != NULL) {
-// 		if (strcmp(cur->next->key, key) == 0) break; // search node
-		
-// 		cur = cur->next;
-// 	}
-
-// 	if (cur->next == NULL) {
-		
-// 		pthread_mutex_unlock(&hash->lock);
-// 		return -1;
-// 	}
-
-// 	hashnode_t *tmp = cur->next;
-// 	cur->next = tmp->next;
-// 	free(tmp);
-// 	hash->count --;
-
-// 	pthread_mutex_unlock(&hash->lock);
-	
-// 	return 0;
-// }
-
-
-// int exist_kv_hashtable(hashtable_t *hash, char *key) {
-
-// 	char *value = get_kv_hashtable(hash, key);
-// 	if (value) return 1;
-// 	else return 0;
-	
-// }
-
-

init.sh

@@ -1,7 +0,0 @@
-rm -rf libbpf-bootstrap/examples/c
-
-cp -R ebpf/c libbpf-bootstrap/examples
-
-cd libbpf-bootstrap/examples/c
-
-make 

kvs_array.c

@@ -1,192 +0,0 @@
-
-
-#include "kvstore.h"
-#include "kvs_rw_tools.h"
-#include "memory/alloc_dispatch.h"
-#include <arpa/inet.h>
-
-
-// singleton
-
-kvs_array_t global_array = {0};
-
-int kvs_array_create(kvs_array_t *inst) {
-
-	if (!inst) return -1;
-	if (inst->table) {
-		printf("table has alloc\n");
-		return -1;
-	}	
-	inst->table = kvs_malloc(KVS_ARRAY_SIZE * sizeof(kvs_array_item_t));
-	if (!inst->table) {
-		return -1;
-	}
-
-	memset(inst->table, 0, (size_t)KVS_ARRAY_SIZE * sizeof(kvs_array_item_t));
-	inst->total = 0;
-
-	return 0;
-}
-
-void kvs_array_destroy(kvs_array_t *inst) {
-
-	if (!inst) return ;
-
-	if (inst->table) {
-		kvs_free(inst->table);
-	}
-
-}
-
-
-
-/*
- * @return: <0, error; =0, success; >0, exist
- */
-
-int kvs_array_set(kvs_array_t *inst, char *key, char *value) {
-
-	if (inst == NULL || key == NULL || value == NULL) return -1;
-	if (inst->total == KVS_ARRAY_SIZE) return -1;
-
-	char *str = kvs_array_get(inst, key);
-	if (str) {
-		return 1; // 
-	}
-
-	char *kcopy = kvs_malloc(strlen(key) + 1);
-	if (kcopy == NULL) return -2;
-	memset(kcopy, 0, strlen(key) + 1);
-	strncpy(kcopy, key, strlen(key));
-
-	char *kvalue = kvs_malloc(strlen(value) + 1);
-	if (kvalue == NULL) return -2;
-	memset(kvalue, 0, strlen(value) + 1);
-	strncpy(kvalue, value, strlen(value));
-
-	int i = 0;
-	for (i = 0;i < inst->total;i ++) {
-		if (inst->table[i].key == NULL) {
-			
-			inst->table[i].key = kcopy;
-			inst->table[i].value = kvalue;
-			inst->total ++;
-			
-			return 0;
-		}
-	}
-
-	if (i == inst->total && i < KVS_ARRAY_SIZE) {
-
-		inst->table[i].key = kcopy;
-		inst->table[i].value = kvalue;
-		inst->total ++;
-	}
-
-	return 0;
-}
-
-char* kvs_array_get(kvs_array_t *inst, char *key) {
-
-	if (inst == NULL || key == NULL) return NULL;
-
-	int i = 0;
-	for (i = 0;i < inst->total;i ++) {
-		if (inst->table[i].key == NULL) {
-			continue;
-		}
-
-		if (strcmp(inst->table[i].key, key) == 0) {
-			return inst->table[i].value;
-		}
-	}
-
-	return NULL;
-}
-
-
-/*
- * @return < 0, error;  =0,  success; >0, no exist
- */
-
-int kvs_array_del(kvs_array_t *inst, char *key) {
-
-	if (inst == NULL || key == NULL) return -1;
-
-	int i = 0;
-	for (i = 0;i < inst->total;i ++) {
-
-		if (strcmp(inst->table[i].key, key) == 0) {
-
-			kvs_free(inst->table[i].key);
-			inst->table[i].key = NULL;
-
-			kvs_free(inst->table[i].value);
-			inst->table[i].value = NULL;
-// error: > 1024
-			if (inst->total-1 == i) {
-				inst->total --;
-			}
-			
-
-			return 0;
-		}
-	}
-
-	return i;
-}
-
-
-/*
- * @return : < 0, error; =0, success; >0, no exist 
- */
-
-int kvs_array_mod(kvs_array_t *inst, char *key, char *value) {
-
-	if (inst == NULL || key == NULL || value == NULL) return -1;
-// error: > 1024
-	if (inst->total == 0) {
-		return KVS_ARRAY_SIZE;
-	}
-	
-
-	int i = 0;
-	for (i = 0;i < inst->total;i ++) {
-
-		if (inst->table[i].key == NULL) {
-			continue;
-		}
-
-		if (strcmp(inst->table[i].key, key) == 0) {
-
-			kvs_free(inst->table[i].value);
-
-			char *kvalue = kvs_malloc(strlen(value) + 1);
-			if (kvalue == NULL) return -2;
-			memset(kvalue, 0, strlen(value) + 1);
-			strncpy(kvalue, value, strlen(value));
-
-			inst->table[i].value = kvalue;
-
-			return 0;
-		}
-
-	}
-
-	return i;
-}
-
-
-/*
- * @return 0: exist, 1: no exist
- */
-int kvs_array_exist(kvs_array_t *inst, char *key) {
-
-	if (!inst || !key) return -1;
-	
-	char *str = kvs_array_get(inst, key);
-	if (!str) {
-		return 1; // 
-	}
-	return 0;
-}

kvs_array_bin.c

@@ -235,7 +235,7 @@ int kvs_array_save(iouring_ctx_t *uring, kvs_array_t *inst, const char* filename
         for (int i = 0; i < count; i++) total += lens[i];
 
         task_t *t = submit_write(uring, fd, bufs, lens, count, current_off);
-        cleanup_finished_iouring_tasks();
+        cleanup_finished_iouring_tasks(uring);
 
 		if (!t) { 
             perror("task init failed");
@@ -249,7 +249,7 @@ int kvs_array_save(iouring_ctx_t *uring, kvs_array_t *inst, const char* filename
 clean:
 	while (!uring_task_complete(uring)) {
 		usleep(1000);
-		cleanup_finished_iouring_tasks();
+		cleanup_finished_iouring_tasks(uring);
 	}
 	close(fd);
 	return 0;

kvs_hash.c

@@ -1,286 +0,0 @@
-
-
-
-// #include <stdio.h>
-// #include <string.h>
-// #include <stdlib.h>
-// #include <pthread.h>
-
-
-// #include "memory/alloc_dispatch.h"
-// #include "kvstore.h"
-
-
-// // Key, Value --> 
-// // Modify 
-
-
-
-// kvs_hash_t global_hash;
-
-
-// //Connection 
-// // 'C' + 'o' + 'n'
-// static int _hash(char *key, int size) {
-
-// 	if (!key) return -1;
-
-// 	int sum = 0;
-// 	int i = 0;
-
-// 	while (key[i] != 0) {
-// 		sum += key[i];
-// 		i ++;
-// 	}
-
-// 	return sum % size;
-
-// }
-
-// hashnode_t *_create_node(char *key, char *value) {
-
-// 	hashnode_t *node = (hashnode_t*)kvs_malloc(sizeof(hashnode_t));
-// 	if (!node) return NULL;
-	
-// #if ENABLE_KEY_POINTER
-// 	char *kcopy = kvs_malloc(strlen(key) + 1);
-// 	if (kcopy == NULL) return NULL;
-// 	memset(kcopy, 0, strlen(key) + 1);
-// 	strncpy(kcopy, key, strlen(key));
-
-// 	node->key = kcopy;
-
-// 	char *kvalue = kvs_malloc(strlen(value) + 1);
-// 	if (kvalue == NULL) { 
-// 		kvs_free(kvalue);
-// 		return NULL;
-// 	}
-// 	memset(kvalue, 0, strlen(value) + 1);
-// 	strncpy(kvalue, value, strlen(value));
-
-// 	node->value = kvalue;
-	
-// #else
-// 	strncpy(node->key, key, MAX_KEY_LEN);
-// 	strncpy(node->value, value, MAX_VALUE_LEN);
-// #endif
-// 	node->next = NULL;
-
-// 	return node;
-// }
-
-
-// //
-// int kvs_hash_create(kvs_hash_t *hash) {
-
-// 	if (!hash) return -1;
-
-// 	hash->nodes = (hashnode_t**)kvs_malloc(sizeof(hashnode_t*) * MAX_TABLE_SIZE);
-// 	if (!hash->nodes) return -1;
-
-// 	hash->max_slots = MAX_TABLE_SIZE;
-// 	hash->count = 0; 
-
-// 	return 0;
-// }
-
-// // 
-// void kvs_hash_destroy(kvs_hash_t *hash) {
-
-// 	if (!hash) return;
-
-// 	int i = 0;
-// 	for (i = 0;i < hash->max_slots;i ++) {
-// 		hashnode_t *node = hash->nodes[i];
-
-// 		while (node != NULL) { // error
-
-// 			hashnode_t *tmp = node;
-// 			node = node->next;
-// 			hash->nodes[i] = node;
-			
-// 			kvs_free(tmp);
-			
-// 		}
-// 	}
-
-// 	kvs_free(hash->nodes);
-	
-// }
-
-// // 5 + 2
-
-// // mp
-// int kvs_hash_set(kvs_hash_t *hash, char *key, char *value) {
-
-// 	if (!hash || !key || !value) return -1;
-
-// 	int idx = _hash(key, MAX_TABLE_SIZE);
-
-// 	hashnode_t *node = hash->nodes[idx];
-// #if 1
-// 	while (node != NULL) {
-// 		if (strcmp(node->key, key) == 0) { // exist
-// 			return 1;
-// 		}
-// 		node = node->next;
-// 	}
-// #endif
-
-// 	hashnode_t *new_node = _create_node(key, value);
-// 	new_node->next = hash->nodes[idx];
-// 	hash->nodes[idx] = new_node;
-	
-// 	hash->count ++;
-
-// 	return 0;
-// }
-
-
-// char * kvs_hash_get(kvs_hash_t *hash, char *key) {
-
-// 	if (!hash || !key) return NULL;
-
-// 	int idx = _hash(key, MAX_TABLE_SIZE);
-
-// 	hashnode_t *node = hash->nodes[idx];
-
-// 	while (node != NULL) {
-
-// 		if (strcmp(node->key, key) == 0) {
-// 			return node->value;
-// 		}
-
-// 		node = node->next;
-// 	}
-
-// 	return NULL;
-
-// }
-
-
-// int kvs_hash_mod(kvs_hash_t *hash, char *key, char *value) {
-
-// 	if (!hash || !key) return -1;
-
-// 	int idx = _hash(key, MAX_TABLE_SIZE);
-
-// 	hashnode_t *node = hash->nodes[idx];
-
-// 	while (node != NULL) {
-
-// 		if (strcmp(node->key, key) == 0) {
-// 			break;
-// 		}
-
-// 		node = node->next;
-// 	}
-
-// 	if (node == NULL) {
-// 		return 1;
-// 	}
-
-// 	// node --> 
-// 	kvs_free(node->value);
-
-// 	char *kvalue = kvs_malloc(strlen(value) + 1);
-// 	if (kvalue == NULL) return -2;
-// 	memset(kvalue, 0, strlen(value) + 1);
-// 	strncpy(kvalue, value, strlen(value));
-
-// 	node->value = kvalue;
-	
-// 	return 0;
-// }
-
-// int kvs_hash_count(kvs_hash_t *hash) {
-// 	return hash->count;
-// }
-
-// int kvs_hash_del(kvs_hash_t *hash, char *key) {
-// 	if (!hash || !key) return -2;
-
-// 	int idx = _hash(key, MAX_TABLE_SIZE);
-
-// 	hashnode_t *head = hash->nodes[idx];
-// 	if (head == NULL) return -1; // noexist
-// 	// head node
-// 	if (strcmp(head->key, key) == 0) {
-// 		hashnode_t *tmp = head->next;
-// 		hash->nodes[idx] = tmp;
-		
-// 		kvs_free(head);
-// 		hash->count --;
-		
-// 		return 0;
-// 	}
-
-// 	hashnode_t *cur = head;
-// 	while (cur->next != NULL) {
-// 		if (strcmp(cur->next->key, key) == 0) break; // search node
-		
-// 		cur = cur->next;
-// 	}
-
-// 	if (cur->next == NULL) {
-		
-// 		return -1;
-// 	}
-
-// 	hashnode_t *tmp = cur->next;
-// 	cur->next = tmp->next;
-// #if ENABLE_KEY_POINTER
-// 	kvs_free(tmp->key);
-// 	kvs_free(tmp->value);
-// #endif
-// 	kvs_free(tmp);
-	
-// 	hash->count --;
-
-// 	return 0;
-// }
-
-
-// int kvs_hash_exist(kvs_hash_t *hash, char *key) {
-
-// 	char *value = kvs_hash_get(hash, key);
-// 	if (!value) return 1;
-
-// 	return 0;
-	
-// }
-
-// #if 0
-// int main() {
-
-// 	kvs_hash_create(&hash);
-
-// 	kvs_hash_set(&hash, "Teacher1", "King");
-// 	kvs_hash_set(&hash, "Teacher2", "Darren");
-// 	kvs_hash_set(&hash, "Teacher3", "Mark");
-// 	kvs_hash_set(&hash, "Teacher4", "Vico");
-// 	kvs_hash_set(&hash, "Teacher5", "Nick");
-
-// 	char *value1 = kvs_hash_get(&hash, "Teacher1");
-// 	printf("Teacher1 : %s\n", value1);
-
-// 	int ret = kvs_hash_mod(&hash, "Teacher1", "King1");
-// 	printf("mode Teacher1 ret : %d\n", ret);
-	
-// 	char *value2 = kvs_hash_get(&hash, "Teacher1");
-// 	printf("Teacher2 : %s\n", value1);
-
-// 	ret = kvs_hash_del(&hash, "Teacher1");
-// 	printf("delete Teacher1 ret : %d\n", ret);
-
-// 	ret = kvs_hash_exist(&hash, "Teacher1");
-// 	printf("Exist Teacher1 ret : %d\n", ret);
-
-// 	kvs_hash_destroy(&hash);
-
-// 	return 0;
-// }
-
-// #endif
-
-

kvs_hash_bin.c

@@ -1,384 +1,680 @@
-
-
-
-
-
-
 #include "kvstore.h"
 #include "kvs_rw_tools.h"
 #include "memory/alloc_dispatch.h"
 #include "diskuring/diskuring.h"
-// Key, Value --> 
-// Modify 
 
+#include <arpa/inet.h>
+#include <fcntl.h>
+#include <unistd.h>
 
+#define KVS_HASH_INITIAL_GLOBAL_DEPTH 1u
+#define KVS_HASH_BUCKET_SPLIT_THRESHOLD 64u
+#define KVS_HASH_MAX_GLOBAL_DEPTH 31u
 
 kvs_hash_t global_hash;
 
-
+static uint32_t _hash_u32(const void *key, uint32_t key_len) {
-//Connection 
-// 'C' + 'o' + 'n'
-static int _hash(const void *key, uint32_t key_len, int size) {
-    if (!key || size <= 0) return -1;
-
     const uint8_t *p = (const uint8_t *)key;
-    uint32_t sum = 0;
+    uint32_t hash = 2166136261u;
+
     for (uint32_t i = 0; i < key_len; i++) {
-        sum += p[i];
+        hash ^= p[i];
+        hash *= 16777619u;
     }
-    return sum % size;
+
+    return hash;
+}
+
+static inline uint32_t _dir_index(uint32_t hashv, uint32_t global_depth) {
+    if (global_depth == 0) {
+        return 0;
+    }
+    return hashv & ((1u << global_depth) - 1u);
+}
+
+static hashbucket_t *_bucket_create(uint32_t local_depth) {
+    hashbucket_t *bucket = (hashbucket_t *)kvs_malloc(sizeof(hashbucket_t));
+    if (!bucket) {
+        return NULL;
+    }
+
+    bucket->head = NULL;
+    bucket->local_depth = local_depth;
+    bucket->item_count = 0;
+    bucket->next_all = NULL;
+    return bucket;
+}
+
+static void _bucket_list_push(kvs_hash_t *hash, hashbucket_t *bucket) {
+    bucket->next_all = hash->bucket_list;
+    hash->bucket_list = bucket;
+}
+
+static inline uint8_t *_get_node_key(const hashnode_t *node) {
+    return (uint8_t *)node + sizeof(hashnode_t);
+}
+
+static inline uint8_t *_get_node_value(const hashnode_t *node) {
+    return (uint8_t *)node + sizeof(hashnode_t) + node->key_len;
 }
 
 static int _key_equal(const hashnode_t *node, const void *key, uint32_t key_len) {
-    if (!node || !key) return 0;
+    if (!node || !key) {
-    if (!node->key) return 0;
+        return 0;
-    if (node->key_len != key_len) return 0;
+    }
-    return memcmp(node->key, key, key_len) == 0;
+    if (node->key_len != key_len) {
+        return 0;
+    }
+    return memcmp(_get_node_key(node), key, key_len) == 0;
 }
 
 static hashnode_t *_create_node(const void *key, uint32_t key_len,
-                               const void *value, uint32_t value_len) {
+                                const void *value, uint32_t value_len) {
-    hashnode_t *node = (hashnode_t*)kvs_malloc(sizeof(hashnode_t));
+    size_t total_size = sizeof(hashnode_t) + key_len + value_len;
-    if (!node) return NULL;
+    hashnode_t *node;
-    memset(node, 0, sizeof(*node));
+    uint8_t *data_ptr;
 
-    if (key_len > 0) {
+    if (!key || key_len == 0) {
-        node->key = (uint8_t*)kvs_malloc(key_len);
+        return NULL;
-        if (!node->key) { kvs_free(node); return NULL; }
-        memcpy(node->key, key, key_len);
-        node->key_len = key_len;
     }
 
-    if (value_len > 0) {
+    node = (hashnode_t *)kvs_malloc(total_size);
-        node->value = (uint8_t*)kvs_malloc(value_len);
+    if (!node) {
-        if (!node->value) {
+        return NULL;
-            kvs_free(node->key);
-            kvs_free(node);
-            return NULL;
-        }
-        memcpy(node->value, value, value_len);
-        node->value_len = value_len;
-    } else {
-        node->value = NULL;
-        node->value_len = 0;
     }
 
+    memset(node, 0, sizeof(hashnode_t));
+    node->key_len = key_len;
+    node->value_len = value_len;
     node->next = NULL;
+
+    data_ptr = (uint8_t *)node + sizeof(hashnode_t);
+    memcpy(data_ptr, key, key_len);
+    if (value_len > 0 && value) {
+        memcpy(data_ptr + key_len, value, value_len);
+    }
+
     return node;
 }
 
+static hashnode_t *_find_node(hashbucket_t *bucket,
+                              const void *key, uint32_t key_len,
+                              hashnode_t ***out_indirect) {
+    hashnode_t **indirect;
+
+    if (out_indirect) {
+        *out_indirect = NULL;
+    }
+    if (!bucket || !key || key_len == 0) {
+        return NULL;
+    }
+
+    indirect = &bucket->head;
+    while (*indirect) {
+        if (_key_equal(*indirect, key, key_len)) {
+            if (out_indirect) {
+                *out_indirect = indirect;
+            }
+            return *indirect;
+        }
+        indirect = &(*indirect)->next;
+    }
+
+    return NULL;
+}
+
+static int _update_node_value(hashnode_t **node_ptr, const void *value, uint32_t value_len) {
+    hashnode_t *old_node;
+    hashnode_t *new_node;
+
+    if (!node_ptr || !*node_ptr) {
+        return -1;
+    }
+    if (value_len > 0 && !value) {
+        return -1;
+    }
+
+    old_node = *node_ptr;
+
+    if (old_node->value_len == value_len) {
+        if (value_len > 0) {
+            memcpy(_get_node_value(old_node), value, value_len);
+        }
+        return 0;
+    }
+
+    new_node = _create_node(_get_node_key(old_node), old_node->key_len, value, value_len);
+    if (!new_node) {
+        return -1;
+    }
+
+    new_node->next = old_node->next;
+    *node_ptr = new_node;
+    kvs_free(old_node);
+    return 0;
+}
+
+static int _double_directory(kvs_hash_t *hash) {
+    uint32_t old_size;
+    uint32_t new_size;
+    hashbucket_t **new_dir;
+    uint32_t i;
+
+    if (!hash || !hash->directory) {
+        return -1;
+    }
+    if (hash->global_depth >= KVS_HASH_MAX_GLOBAL_DEPTH) {
+        return 1;
+    }
+
+    old_size = hash->dir_size;
+    if (old_size == 0 || old_size > UINT32_MAX / 2u) {
+        return 1;
+    }
+
+    new_size = old_size << 1;
+    new_dir = (hashbucket_t **)kvs_malloc(sizeof(hashbucket_t *) * new_size);
+    if (!new_dir) {
+        return 1;
+    }
+
+    for (i = 0; i < old_size; i++) {
+        new_dir[i] = hash->directory[i];
+        new_dir[i + old_size] = hash->directory[i];
+    }
+
+    kvs_free(hash->directory);
+    hash->directory = new_dir;
+    hash->dir_size = new_size;
+    hash->global_depth++;
+    return 0;
+}
+
+static int _split_bucket(kvs_hash_t *hash, uint32_t dir_idx) {
+    hashbucket_t *old_bucket;
+    hashbucket_t *new_bucket;
+    hashnode_t *node;
+    uint32_t split_bit;
+    uint32_t i;
+
+    if (!hash || !hash->directory || dir_idx >= hash->dir_size) {
+        return -1;
+    }
+
+    old_bucket = hash->directory[dir_idx];
+    if (!old_bucket) {
+        return -1;
+    }
+
+    if (old_bucket->local_depth >= KVS_HASH_MAX_GLOBAL_DEPTH) {
+        return 1;
+    }
+
+    if (old_bucket->local_depth == hash->global_depth) {
+        int rc = _double_directory(hash);
+        if (rc != 0) {
+            return rc;
+        }
+    }
+
+    new_bucket = _bucket_create(old_bucket->local_depth + 1);
+    if (!new_bucket) {
+        return -1;
+    }
+    _bucket_list_push(hash, new_bucket);
+
+    old_bucket->local_depth++;
+    split_bit = 1u << (old_bucket->local_depth - 1u);
+
+    for (i = 0; i < hash->dir_size; i++) {
+        if (hash->directory[i] == old_bucket && (i & split_bit)) {
+            hash->directory[i] = new_bucket;
+        }
+    }
+
+    node = old_bucket->head;
+    old_bucket->head = NULL;
+    old_bucket->item_count = 0;
+
+    while (node) {
+        hashnode_t *next = node->next;
+        uint32_t idx = _dir_index(_hash_u32(_get_node_key(node), node->key_len), hash->global_depth);
+        hashbucket_t *target = hash->directory[idx];
+        node->next = target->head;
+        target->head = node;
+        target->item_count++;
+        node = next;
+    }
+
+    return 0;
+}
 
-//
 int kvs_hash_create(kvs_hash_t *hash) {
-    if (!hash) return -1;
+    uint32_t init_depth;
+    uint32_t i;
+    hashbucket_t *initial_bucket;
 
-    hash->nodes = (hashnode_t**)kvs_malloc(sizeof(hashnode_t*) * MAX_TABLE_SIZE);
+    if (!hash) {
-    if (!hash->nodes) return -1;
+        return -1;
+    }
+
+    memset(hash, 0, sizeof(*hash));
+
+    init_depth = KVS_HASH_INITIAL_GLOBAL_DEPTH;
+    if (init_depth > KVS_HASH_MAX_GLOBAL_DEPTH) {
+        init_depth = KVS_HASH_MAX_GLOBAL_DEPTH;
+    }
+    hash->global_depth = init_depth;
+    hash->dir_size = 1u << init_depth;
+    if (hash->dir_size == 0) {
+        return -1;
+    }
+
+    hash->directory = (hashbucket_t **)kvs_malloc(sizeof(hashbucket_t *) * hash->dir_size);
+    if (!hash->directory) {
+        return -1;
+    }
+    memset(hash->directory, 0, sizeof(hashbucket_t *) * hash->dir_size);
+
+    initial_bucket = _bucket_create(0);
+    if (!initial_bucket) {
+        kvs_free(hash->directory);
+        hash->directory = NULL;
+        hash->dir_size = 0;
+        return -1;
+    }
+
+    _bucket_list_push(hash, initial_bucket);
+    for (i = 0; i < hash->dir_size; i++) {
+        hash->directory[i] = initial_bucket;
+    }
 
-    memset(hash->nodes, 0, sizeof(hashnode_t*) * MAX_TABLE_SIZE);
-    hash->max_slots = MAX_TABLE_SIZE;
     hash->count = 0;
     return 0;
 }
 
-// 
 void kvs_hash_destroy(kvs_hash_t *hash) {
-    if (!hash || !hash->nodes) return;
+    hashbucket_t *bucket;
 
-    for (int i = 0; i < hash->max_slots; i++) {
+    if (!hash) {
-        hashnode_t *node = hash->nodes[i];
+        return;
-        while (node != NULL) {
+    }
-            hashnode_t *tmp = node;
-            node = node->next;
 
-            if (tmp->key) kvs_free(tmp->key);
+    bucket = hash->bucket_list;
-            if (tmp->value) kvs_free(tmp->value);
+    while (bucket) {
-            kvs_free(tmp);
+        hashbucket_t *next_bucket = bucket->next_all;
+        hashnode_t *node = bucket->head;
+        while (node) {
+            hashnode_t *next = node->next;
+            kvs_free(node);
+            node = next;
         }
-        hash->nodes[i] = NULL;
+        kvs_free(bucket);
+        bucket = next_bucket;
     }
 
-    kvs_free(hash->nodes);
+    kvs_free(hash->directory);
-    hash->nodes = NULL;
+    memset(hash, 0, sizeof(*hash));
-    hash->max_slots = 0;
-    hash->count = 0;
 }
 
+int kvs_hash_set_bin(kvs_hash_t *hash, const void *key, uint32_t key_len,
+                     const void *value, uint32_t value_len) {
+    uint32_t hashv;
+    uint32_t idx;
+    hashbucket_t *bucket;
+    hashnode_t **indirect = NULL;
+    hashnode_t *found;
+    hashnode_t *new_node;
 
-// 5 + 2
+    if (!hash || !hash->directory || !key || key_len == 0) {
-
+        return -1;
-// mp
-/*
- * @return: <0 error; 0 success; 1 exist
- */
-int kvs_hash_set_bin(kvs_hash_t *hash, const void *key, uint32_t key_len, const void *value, uint32_t value_len) {
-	if (!hash || !hash->nodes || !key || key_len == 0) return -1;
-
-	int idx = _hash(key, key_len, MAX_TABLE_SIZE);
-	if (idx < 0) return -1;
-
-	hashnode_t *node = hash->nodes[idx];
-	while (node != NULL) {
-		if (_key_equal(node, key, key_len)) { // exist
-			return 1;
-		}
-		node = node->next;
-	}
-
-	hashnode_t *new_node = _create_node(key, key_len, value, value_len);
-	if (!new_node) return -2;
-
-	new_node->next = hash->nodes[idx];
-	hash->nodes[idx] = new_node;
-	hash->count ++;
-
-	return 0;
-}
-
-/*
- * @return: NULL notexist, NOTNULL exist。out_value_len 是长度。
- */
-void *kvs_hash_get_bin(kvs_hash_t *hash, const void *key, uint32_t key_len, uint32_t *out_value_len) {
-	if (!hash || !hash->nodes || !key || key_len == 0 || !out_value_len) return NULL;
-	*out_value_len = 0;
-
-	int idx = _hash(key, key_len, MAX_TABLE_SIZE);
-	if (idx < 0) return NULL;
-
-	hashnode_t *node = hash->nodes[idx];
-
-	while (node != NULL) {
-
-		if (_key_equal(node, key, key_len)) {
-			*out_value_len = node->value_len;
-			return node->value;
-		}
-
-		node = node->next;
-	}
-
-	return NULL;
-
-}
-
-/*
- * @return <0 error; =0 success; >0 no exist
- */
-int kvs_hash_mod_bin(kvs_hash_t *hash, const void *key, uint32_t key_len, const void *value, uint32_t value_len) {
-
-	if (!hash || !hash->nodes || !key || key_len == 0 || !value) return -1;
-
-	int idx = _hash(key, key_len, MAX_TABLE_SIZE);
-	if (idx < 0) return -1;
-
-	hashnode_t *node = hash->nodes[idx];
-
-	while (node != NULL) {
-
-		if (_key_equal(node, key, key_len)) {
-			break;
-		}
-
-		node = node->next;
-	}
-
-	if (node == NULL) {
-		return 1;
-	}
-
-	// node --> 
-	if (node->value) kvs_free(node->value);
-	node->value = NULL;
-	node->value_len = 0;
-
-	if (value_len > 0) {
-        uint8_t *vcopy = (uint8_t*)kvs_malloc(value_len);
-        if (!vcopy) return -2;
-        memcpy(vcopy, value, value_len);
-        node->value = vcopy;
-        node->value_len = value_len;
     }
-	return 0;
+    if (value_len > 0 && !value) {
+        return -1;
+    }
+
+    hashv = _hash_u32(key, key_len);
+    idx = _dir_index(hashv, hash->global_depth);
+    bucket = hash->directory[idx];
+    if (!bucket) {
+        return -1;
+    }
+
+    found = _find_node(bucket, key, key_len, &indirect);
+    if (found) {
+        return (_update_node_value(indirect, value, value_len) == 0) ? 0 : -2;
+    }
+
+    while (bucket->item_count >= KVS_HASH_BUCKET_SPLIT_THRESHOLD) {
+        int split_rc = _split_bucket(hash, idx);
+        if (split_rc < 0) {
+            return -2;
+        }
+        if (split_rc > 0) {
+            break;
+        }
+        idx = _dir_index(hashv, hash->global_depth);
+        bucket = hash->directory[idx];
+    }
+
+    new_node = _create_node(key, key_len, value, value_len);
+    if (!new_node) {
+        return -2;
+    }
+
+    new_node->next = bucket->head;
+    bucket->head = new_node;
+    bucket->item_count++;
+    hash->count++;
+    return 0;
+}
+
+void *kvs_hash_get_bin(kvs_hash_t *hash, const void *key, uint32_t key_len,
+                       uint32_t *out_value_len) {
+    uint32_t idx;
+    hashbucket_t *bucket;
+    hashnode_t *node;
+
+    if (!out_value_len) {
+        return NULL;
+    }
+    *out_value_len = 0;
+
+    if (!hash || !hash->directory || !key || key_len == 0) {
+        return NULL;
+    }
+
+    idx = _dir_index(_hash_u32(key, key_len), hash->global_depth);
+    bucket = hash->directory[idx];
+    if (!bucket) {
+        return NULL;
+    }
+
+    node = _find_node(bucket, key, key_len, NULL);
+    if (!node) {
+        return NULL;
+    }
+
+    *out_value_len = node->value_len;
+    return (node->value_len > 0) ? _get_node_value(node) : NULL;
+}
+
+int kvs_hash_get_copy_bin(kvs_hash_t *hash, const void *key, uint32_t key_len,
+                          void **out_buf, uint32_t *out_len) {
+    uint32_t idx;
+    hashbucket_t *bucket;
+    hashnode_t *node;
+    void *copy;
+
+    if (!out_buf || !out_len) {
+        return -1;
+    }
+    *out_buf = NULL;
+    *out_len = 0;
+
+    if (!hash || !hash->directory || !key || key_len == 0) {
+        return -1;
+    }
+
+    idx = _dir_index(_hash_u32(key, key_len), hash->global_depth);
+    bucket = hash->directory[idx];
+    if (!bucket) {
+        return -1;
+    }
+
+    node = _find_node(bucket, key, key_len, NULL);
+    if (!node) {
+        return 1;
+    }
+
+    *out_len = node->value_len;
+    if (node->value_len == 0) {
+        return 0;
+    }
+
+    copy = kvs_malloc(node->value_len);
+    if (!copy) {
+        *out_len = 0;
+        return -2;
+    }
+
+    memcpy(copy, _get_node_value(node), node->value_len);
+    *out_buf = copy;
+    return 0;
+}
+
+int kvs_hash_mod_bin(kvs_hash_t *hash, const void *key, uint32_t key_len,
+                     const void *value, uint32_t value_len) {
+    uint32_t idx;
+    hashbucket_t *bucket;
+    hashnode_t **indirect = NULL;
+    hashnode_t *node;
+
+    if (!hash || !hash->directory || !key || key_len == 0) {
+        return -1;
+    }
+    if (value_len > 0 && !value) {
+        return -1;
+    }
+
+    idx = _dir_index(_hash_u32(key, key_len), hash->global_depth);
+    bucket = hash->directory[idx];
+    if (!bucket) {
+        return -1;
+    }
+
+    node = _find_node(bucket, key, key_len, &indirect);
+    if (!node) {
+        return 1;
+    }
+
+    return (_update_node_value(indirect, value, value_len) == 0) ? 0 : -2;
+}
+
+int kvs_hash_del_bin(kvs_hash_t *hash, const void *key, uint32_t key_len) {
+    uint32_t idx;
+    hashbucket_t *bucket;
+    hashnode_t **indirect = NULL;
+    hashnode_t *node;
+
+    if (!hash || !hash->directory || !key || key_len == 0) {
+        return -1;
+    }
+
+    idx = _dir_index(_hash_u32(key, key_len), hash->global_depth);
+    bucket = hash->directory[idx];
+    if (!bucket) {
+        return -1;
+    }
+
+    node = _find_node(bucket, key, key_len, &indirect);
+    if (!node) {
+        return 1;
+    }
+
+    *indirect = node->next;
+    kvs_free(node);
+    bucket->item_count--;
+    hash->count--;
+    return 0;
+}
+
+int kvs_hash_exist_bin(kvs_hash_t *hash, const void *key, uint32_t key_len) {
+    uint32_t idx;
+    hashbucket_t *bucket;
+
+    if (!hash || !hash->directory || !key || key_len == 0) {
+        return -1;
+    }
+
+    idx = _dir_index(_hash_u32(key, key_len), hash->global_depth);
+    bucket = hash->directory[idx];
+    if (!bucket) {
+        return 1;
+    }
+
+    return _find_node(bucket, key, key_len, NULL) ? 0 : 1;
 }
 
 int kvs_hash_count(kvs_hash_t *hash) {
-	return hash->count;
+    return hash ? hash->count : 0;
 }
 
-/*
+int kvs_hash_save(iouring_ctx_t *uring, kvs_hash_t *inst, const char *filename) {
- * @return <0 error; =0 success; >0 no exist
+    int fd;
- */
-int kvs_hash_del_bin(kvs_hash_t *hash, const void *key, uint32_t key_len) {
-	if (!hash || !key || key_len == 0) return -1;
-
-	int idx = _hash(key, key_len, MAX_TABLE_SIZE);
-	if (idx < 0) return -1;
-
-	hashnode_t *head = hash->nodes[idx];
-	if (head == NULL) return 1; // noexist
-
-	// head node
-	if (_key_equal(head, key, key_len)) {
-		hashnode_t *tmp = head->next;
-		hash->nodes[idx] = tmp;
-		
-		if (head->key) kvs_free(head->key);
-		if (head->value) kvs_free(head->value);
-		kvs_free(head);
-		hash->count --;
-		
-		return 0;
-	}
-
-	hashnode_t *cur = head;
-	while (cur->next != NULL) {
-		if (_key_equal(cur->next, key, key_len)) break; // search node
-		
-		cur = cur->next;
-	}
-
-	if (cur->next == NULL) {
-		
-		return 1;
-	}
-
-	hashnode_t *tmp = cur->next;
-	cur->next = tmp->next;
-    if (tmp->key) kvs_free(tmp->key);
-    if (tmp->value) kvs_free(tmp->value);
-	kvs_free(tmp);
-	
-	hash->count --;
-
-	return 0;
-}
-
-/*
- * @return =0 exist, =1 no exist
- */
-int kvs_hash_exist_bin(kvs_hash_t *hash, const void *key, uint32_t key_len) {
-	uint32_t vlen = 0;
-	void *value = kvs_hash_get_bin(hash, key, key_len, &vlen);
-	return value ? 0 : 1;
-}
-
-// 0 suc, <0 error
-int kvs_hash_save(iouring_ctx_t *uring, kvs_hash_t *inst, const char* filename){
-	if(!inst || !filename) return -1;
-	int fd = open(filename, O_WRONLY | O_CREAT | O_TRUNC, 0644);
-    if(fd < 0) return -2;
-
     off_t current_off = 0;
+    int rc = 0;
+    hashbucket_t *bucket;
 
-	for(int i = 0;i < inst->max_slots; ++ i){
+    if (!uring || !inst || !filename) {
-		for (hashnode_t *n = inst->nodes[i]; n != NULL; n = n->next) {
+        return -1;
-			if (!n->key || n->key_len == 0) continue;
+    }
- 			if (n->value_len > 0 && !n->value) { 
-				goto clean; 
-			}
 
+    fd = open(filename, O_WRONLY | O_CREAT | O_TRUNC, 0644);
+    if (fd < 0) {
+        return -2;
+    }
 
-			uint32_t klen = htonl((uint32_t)n->key_len);
+    for (bucket = inst->bucket_list; bucket != NULL; bucket = bucket->next_all) {
-			uint32_t vlen = htonl((uint32_t)n->value_len);
+        hashnode_t *n = bucket->head;
+        while (n != NULL) {
+            uint32_t klen = htonl(n->key_len);
+            uint32_t vlen = htonl(n->value_len);
+            uint8_t *key_ptr = _get_node_key(n);
+            uint8_t *val_ptr = (n->value_len > 0) ? _get_node_value(n) : NULL;
+            void *bufs[4];
+            size_t lens[4];
+            int count = 0;
+            size_t total = 0;
+            task_t *t;
 
-			void *bufs[4];
+            bufs[count] = &klen;
-			size_t lens[4];
+            lens[count++] = sizeof(klen);
-			int count = 0;
+            bufs[count] = &vlen;
+            lens[count++] = sizeof(vlen);
 
-			bufs[count] = &klen;
+            if (n->key_len > 0) {
-			lens[count] = sizeof(klen);
+                bufs[count] = key_ptr;
-			count++;
+                lens[count++] = n->key_len;
+            }
+            if (n->value_len > 0) {
+                bufs[count] = val_ptr;
+                lens[count++] = n->value_len;
+            }
 
-			bufs[count] = &vlen;
+            for (int j = 0; j < count; j++) {
-			lens[count] = sizeof(vlen);
+                total += lens[j];
-			count++;
+            }
 
-			if (n->key_len > 0){
+            t = submit_write(uring, fd, bufs, lens, count, current_off);
-				bufs[count] = n->key;
+            if (!t) {
-				lens[count] = n->key_len;
+                rc = -3;
-				count++;
+                goto done;
-			}
+            }
+            cleanup_finished_iouring_tasks(uring);
+            current_off += (off_t)total;
+            n = n->next;
+        }
+    }
 
-			if (n->value_len > 0) {
+done:
-				bufs[count] = n->value;
+    while (!uring_task_complete(uring)) {
-				lens[count] = n->value_len;
+        usleep(1000);
-				count++;
+        cleanup_finished_iouring_tasks(uring);
-			}
+    }
 
-			size_t total = 0;
+    close(fd);
-			for (int i = 0; i < count; i++) total += lens[i];
+    return rc;
-
-
-			task_t *t = submit_write(uring, fd, bufs, lens, count, current_off);
-			if(!t) {
-				perror("task init failed");
-				goto clean;
-			}
-			cleanup_finished_iouring_tasks();
-			
-			current_off += (off_t) total;
-		}
-	}
-
-clean:
-	while (!uring_task_complete(uring)) {
-		usleep(1000);
-		cleanup_finished_iouring_tasks();
-	}
-	close(fd);
-	return 0;
 }
 
-int kvs_hash_load(kvs_hash_t *inst, const char* filename){
+int kvs_hash_load(kvs_hash_t *inst, const char *filename) {
-	if (!inst || !filename) return -1;
+    int fd;
-    if (!inst->nodes || inst->max_slots <= 0) return -1;
+    int rc = 0;
 
-    FILE *fp = fopen(filename, "rb");
+    if (!inst || !filename) {
-    if (!fp) return -2;
+        return -1;
+    }
+    if (!inst->directory || inst->dir_size == 0) {
+        return -1;
+    }
 
-	while(1){
+    fd = open(filename, O_RDONLY);
-		uint32_t klen_n = 0, vlen_n = 0;
+    if (fd < 0) {
+        return -2;
+    }
 
-		if (kvs_read_file(fp, &klen_n, 4) < 0) { fclose(fp); return -3; }
+    while (1) {
-		if (kvs_read_file(fp, &vlen_n, 4) < 0) { fclose(fp); return -3; }
+        uint32_t klen_n = 0;
+        uint32_t vlen_n = 0;
+        uint32_t klen;
+        uint32_t vlen;
+        uint8_t *keybuf = NULL;
+        uint8_t *valbuf = NULL;
+        int rr;
 
-		uint32_t klen = ntohl(klen_n);
+        rr = read_full(fd, &klen_n, sizeof(klen_n));
-        uint32_t vlen = ntohl(vlen_n);
+        if (rr == 0) {
-
+            rc = 0;
-		if (klen == 0) { fclose(fp); return -3; }
+            break;
-
-		uint8_t *keybuf = (uint8_t*)kvs_malloc((size_t)klen);
-        if (!keybuf) { fclose(fp); return -4; }
-        if (kvs_read_file(fp, keybuf, (size_t)klen) < 0) {
-            kvs_free(keybuf);
-            fclose(fp);
-            return -3;
         }
-		uint8_t *valbuf = NULL;
+        if (rr < 0) {
+            rc = -3;
+            break;
+        }
+
+        rr = read_full(fd, &vlen_n, sizeof(vlen_n));
+        if (rr <= 0) {
+            rc = -3;
+            break;
+        }
+
+        klen = ntohl(klen_n);
+        vlen = ntohl(vlen_n);
+        if (klen == 0) {
+            rc = -3;
+            break;
+        }
+
+        keybuf = (uint8_t *)kvs_malloc((size_t)klen);
+        if (!keybuf) {
+            rc = -4;
+            break;
+        }
+
+        rr = read_full(fd, keybuf, (size_t)klen);
+        if (rr <= 0) {
+            kvs_free(keybuf);
+            rc = -3;
+            break;
+        }
+
         if (vlen > 0) {
-            valbuf = (uint8_t*)kvs_malloc((size_t)vlen);
+            valbuf = (uint8_t *)kvs_malloc((size_t)vlen);
             if (!valbuf) {
                 kvs_free(keybuf);
-                fclose(fp);
+                rc = -4;
-                return -4;
+                break;
             }
-            if (kvs_read_file(fp, valbuf, (size_t)vlen) < 0) {
+            rr = read_full(fd, valbuf, (size_t)vlen);
+            if (rr <= 0) {
                 kvs_free(valbuf);
                 kvs_free(keybuf);
-                fclose(fp);
+                rc = -3;
-                return -3;
+                break;
             }
         }
 
-		int rc = kvs_hash_set_bin(inst, keybuf, klen, valbuf, vlen);
+        rr = kvs_hash_set_bin(inst, keybuf, klen, valbuf, vlen);
-		kvs_free(keybuf);
+        kvs_free(keybuf);
-        if (vlen > 0) kvs_free(valbuf);
+        if (valbuf) {
-
+            kvs_free(valbuf);
-		if (rc < 0) { // error
-            fclose(fp);
-            return -5;
         }
-	}
+        if (rr < 0) {
-	fclose(fp);
+            rc = -5;
-    return 0;
+            break;
+        }
+    }
+
+    close(fd);
+    return rc;
 }

kvs_protocol_resp.c

@@ -539,7 +539,7 @@ int resp_dispatch(const resp_cmd_t *cmd, resp_value_t *out_value) {
     /* ---------------- misc ---------------- */
     case KVS_CMD_SAVE: {
         if (cmd->argc != 1) { *out_value = resp_error("ERR wrong number of arguments for 'save'"); return 0; }
-        int r = kvs_create_snapshot(&global_uring_ctx, global_array_file, global_rbtree_file, global_hash_file);
+        int r = kvs_create_snapshot_async_1(&global_uring_ctx, global_array_file, global_rbtree_file, global_hash_file);
         if(r == 0) ksv_clear_log(global_oplog_fd);
         if (r < 0) { *out_value = resp_error("ERR save failed"); return 0; }
         *out_value = resp_simple("OK");
@@ -547,7 +547,7 @@ int resp_dispatch(const resp_cmd_t *cmd, resp_value_t *out_value) {
     }
 	case KVS_CMD_SSYNC:
         __ssync(cmd->argv[1].ptr, cmd->argv[1].len, atoi(cmd->argv[2].ptr), global_seq);
-        kvs_create_snapshot_async(cmd->argv[1].ptr, atoi(cmd->argv[2].ptr));
+        kvs_create_snapshot_async_2(cmd->argv[1].ptr, atoi(cmd->argv[2].ptr));
 		*out_value = resp_simple("OK");
 		return 0;
     case KVS_CMD_SREADY:
@@ -555,8 +555,8 @@ int resp_dispatch(const resp_cmd_t *cmd, resp_value_t *out_value) {
 		*out_value = resp_simple("OK");
 		return 0;
     case KVS_CMD_MEM_PRINT:{
+        iouring_profile_dump(&global_uring_ctx);
         int ret = kvs_mem_printf();
-        printf("ret %d\n", ret);
         *out_value = resp_int(ret);
         return 0;
     }

kvs_rbtree.c

@@ -1,553 +0,0 @@
-
-
-
-// #include <stdio.h>
-// #include <stdlib.h>
-// #include <string.h>
-
-
-// #include "kvstore.h"
-
-// rbtree_node *rbtree_mini(rbtree *T, rbtree_node *x) {
-// 	while (x->left != T->nil) {
-// 		x = x->left;
-// 	}
-// 	return x;
-// }
-
-// rbtree_node *rbtree_maxi(rbtree *T, rbtree_node *x) {
-// 	while (x->right != T->nil) {
-// 		x = x->right;
-// 	}
-// 	return x;
-// }
-
-// rbtree_node *rbtree_successor(rbtree *T, rbtree_node *x) {
-// 	rbtree_node *y = x->parent;
-
-// 	if (x->right != T->nil) {
-// 		return rbtree_mini(T, x->right);
-// 	}
-
-// 	while ((y != T->nil) && (x == y->right)) {
-// 		x = y;
-// 		y = y->parent;
-// 	}
-// 	return y;
-// }
-
-
-// void rbtree_left_rotate(rbtree *T, rbtree_node *x) {
-
-// 	rbtree_node *y = x->right;  // x  --> y  ,  y --> x,   right --> left,  left --> right
-
-// 	x->right = y->left; //1 1
-// 	if (y->left != T->nil) { //1 2
-// 		y->left->parent = x;
-// 	}
-
-// 	y->parent = x->parent; //1 3
-// 	if (x->parent == T->nil) { //1 4
-// 		T->root = y;
-// 	} else if (x == x->parent->left) {
-// 		x->parent->left = y;
-// 	} else {
-// 		x->parent->right = y;
-// 	}
-
-// 	y->left = x; //1 5
-// 	x->parent = y; //1 6
-// }
-
-
-// void rbtree_right_rotate(rbtree *T, rbtree_node *y) {
-
-// 	rbtree_node *x = y->left;
-
-// 	y->left = x->right;
-// 	if (x->right != T->nil) {
-// 		x->right->parent = y;
-// 	}
-
-// 	x->parent = y->parent;
-// 	if (y->parent == T->nil) {
-// 		T->root = x;
-// 	} else if (y == y->parent->right) {
-// 		y->parent->right = x;
-// 	} else {
-// 		y->parent->left = x;
-// 	}
-
-// 	x->right = y;
-// 	y->parent = x;
-// }
-
-// void rbtree_insert_fixup(rbtree *T, rbtree_node *z) {
-
-// 	while (z->parent->color == RED) { //z ---> RED
-// 		if (z->parent == z->parent->parent->left) {
-// 			rbtree_node *y = z->parent->parent->right;
-// 			if (y->color == RED) {
-// 				z->parent->color = BLACK;
-// 				y->color = BLACK;
-// 				z->parent->parent->color = RED;
-
-// 				z = z->parent->parent; //z --> RED
-// 			} else {
-
-// 				if (z == z->parent->right) {
-// 					z = z->parent;
-// 					rbtree_left_rotate(T, z);
-// 				}
-
-// 				z->parent->color = BLACK;
-// 				z->parent->parent->color = RED;
-// 				rbtree_right_rotate(T, z->parent->parent);
-// 			}
-// 		}else {
-// 			rbtree_node *y = z->parent->parent->left;
-// 			if (y->color == RED) {
-// 				z->parent->color = BLACK;
-// 				y->color = BLACK;
-// 				z->parent->parent->color = RED;
-
-// 				z = z->parent->parent; //z --> RED
-// 			} else {
-// 				if (z == z->parent->left) {
-// 					z = z->parent;
-// 					rbtree_right_rotate(T, z);
-// 				}
-
-// 				z->parent->color = BLACK;
-// 				z->parent->parent->color = RED;
-// 				rbtree_left_rotate(T, z->parent->parent);
-// 			}
-// 		}
-		
-// 	}
-
-// 	T->root->color = BLACK;
-// }
-
-
-// void rbtree_insert(rbtree *T, rbtree_node *z) {
-
-// 	rbtree_node *y = T->nil;
-// 	rbtree_node *x = T->root;
-
-// 	while (x != T->nil) {
-// 		y = x;
-// #if ENABLE_KEY_CHAR
-
-// 		if (strcmp(z->key, x->key) < 0) {
-// 			x = x->left;
-// 		} else if (strcmp(z->key, x->key) > 0) {
-// 			x = x->right;
-// 		} else {
-// 			return ;
-// 		}
-
-// #else
-// 		if (z->key < x->key) {
-// 			x = x->left;
-// 		} else if (z->key > x->key) {
-// 			x = x->right;
-// 		} else { //Exist
-// 			return ;
-// 		}
-// #endif
-// 	}
-
-// 	z->parent = y;
-// 	if (y == T->nil) {
-// 		T->root = z;
-// #if ENABLE_KEY_CHAR
-// 	} else if (strcmp(z->key, y->key) < 0) {
-// #else
-// 	} else if (z->key < y->key) {
-// #endif
-// 		y->left = z;
-// 	} else {
-// 		y->right = z;
-// 	}
-
-// 	z->left = T->nil;
-// 	z->right = T->nil;
-// 	z->color = RED;
-
-// 	rbtree_insert_fixup(T, z);
-// }
-
-// void rbtree_delete_fixup(rbtree *T, rbtree_node *x) {
-
-// 	while ((x != T->root) && (x->color == BLACK)) {
-// 		if (x == x->parent->left) {
-
-// 			rbtree_node *w= x->parent->right;
-// 			if (w->color == RED) {
-// 				w->color = BLACK;
-// 				x->parent->color = RED;
-
-// 				rbtree_left_rotate(T, x->parent);
-// 				w = x->parent->right;
-// 			}
-
-// 			if ((w->left->color == BLACK) && (w->right->color == BLACK)) {
-// 				w->color = RED;
-// 				x = x->parent;
-// 			} else {
-
-// 				if (w->right->color == BLACK) {
-// 					w->left->color = BLACK;
-// 					w->color = RED;
-// 					rbtree_right_rotate(T, w);
-// 					w = x->parent->right;
-// 				}
-
-// 				w->color = x->parent->color;
-// 				x->parent->color = BLACK;
-// 				w->right->color = BLACK;
-// 				rbtree_left_rotate(T, x->parent);
-
-// 				x = T->root;
-// 			}
-
-// 		} else {
-
-// 			rbtree_node *w = x->parent->left;
-// 			if (w->color == RED) {
-// 				w->color = BLACK;
-// 				x->parent->color = RED;
-// 				rbtree_right_rotate(T, x->parent);
-// 				w = x->parent->left;
-// 			}
-
-// 			if ((w->left->color == BLACK) && (w->right->color == BLACK)) {
-// 				w->color = RED;
-// 				x = x->parent;
-// 			} else {
-
-// 				if (w->left->color == BLACK) {
-// 					w->right->color = BLACK;
-// 					w->color = RED;
-// 					rbtree_left_rotate(T, w);
-// 					w = x->parent->left;
-// 				}
-
-// 				w->color = x->parent->color;
-// 				x->parent->color = BLACK;
-// 				w->left->color = BLACK;
-// 				rbtree_right_rotate(T, x->parent);
-
-// 				x = T->root;
-// 			}
-
-// 		}
-// 	}
-
-// 	x->color = BLACK;
-// }
-
-// rbtree_node *rbtree_delete(rbtree *T, rbtree_node *z) {
-
-// 	rbtree_node *y = T->nil;
-// 	rbtree_node *x = T->nil;
-
-// 	if ((z->left == T->nil) || (z->right == T->nil)) {
-// 		y = z;
-// 	} else {
-// 		y = rbtree_successor(T, z);
-// 	}
-
-// 	if (y->left != T->nil) {
-// 		x = y->left;
-// 	} else if (y->right != T->nil) {
-// 		x = y->right;
-// 	}
-
-// 	x->parent = y->parent;
-// 	if (y->parent == T->nil) {
-// 		T->root = x;
-// 	} else if (y == y->parent->left) {
-// 		y->parent->left = x;
-// 	} else {
-// 		y->parent->right = x;
-// 	}
-
-// 	if (y != z) {
-// #if ENABLE_KEY_CHAR
-
-// 		void *tmp = z->key;
-// 		z->key = y->key;
-// 		y->key = tmp;
-
-// 		tmp = z->value;
-// 		z->value= y->value;
-// 		y->value = tmp;
-
-// #else
-// 		z->key = y->key;
-// 		z->value = y->value;
-// #endif
-// 	}
-
-// 	if (y->color == BLACK) {
-// 		rbtree_delete_fixup(T, x);
-// 	}
-
-// 	return y;
-// }
-
-// rbtree_node *rbtree_search(rbtree *T, KEY_TYPE key) {
-
-// 	rbtree_node *node = T->root;
-// 	while (node != T->nil) {
-// #if ENABLE_KEY_CHAR
-
-// 		if (strcmp(key, node->key) < 0) {
-// 			node = node->left;
-// 		} else if (strcmp(key, node->key) > 0) {
-// 			node = node->right;
-// 		} else {
-// 			return node;
-// 		}
-
-// #else
-// 		if (key < node->key) {
-// 			node = node->left;
-// 		} else if (key > node->key) {
-// 			node = node->right;
-// 		} else {
-// 			return node;
-// 		}	
-// #endif
-// 	}
-// 	return T->nil;
-// }
-
-
-// void rbtree_traversal(rbtree *T, rbtree_node *node) {
-// 	if (node != T->nil) {
-// 		rbtree_traversal(T, node->left);
-// #if ENABLE_KEY_CHAR
-// 		printf("key:%s, value:%s\n", node->key, (char *)node->value);
-// #else
-// 		printf("key:%d, color:%d\n", node->key, node->color);
-// #endif
-// 		rbtree_traversal(T, node->right);
-// 	}
-// }
-
-
-// #if 0
-
-// int main() {
-
-// #if ENABLE_KEY_CHAR
-
-// 	char* keyArray[10] = {"King", "Darren", "Mark", "Vico", "Nick", "qiuxiang", "youzi", "taozi", "123", "234"};
-// 	char* valueArray[10] = {"1King", "2Darren", "3Mark", "4Vico", "5Nick", "6qiuxiang", "7youzi", "8taozi", "9123", "10234"};
-
-// 	rbtree *T = (rbtree *)malloc(sizeof(rbtree));
-// 	if (T == NULL) {
-// 		printf("malloc failed\n");
-// 		return -1;
-// 	}
-	
-// 	T->nil = (rbtree_node*)malloc(sizeof(rbtree_node));
-// 	T->nil->color = BLACK;
-// 	T->root = T->nil;
-
-// 	rbtree_node *node = T->nil;
-// 	int i = 0;
-// 	for (i = 0;i < 10;i ++) {
-// 		node = (rbtree_node*)malloc(sizeof(rbtree_node));
-		
-// 		node->key = malloc(strlen(keyArray[i]) + 1);
-// 		memset(node->key, 0, strlen(keyArray[i]) + 1);
-// 		strcpy(node->key, keyArray[i]);
-		
-// 		node->value = malloc(strlen(valueArray[i]) + 1);
-// 		memset(node->value, 0, strlen(valueArray[i]) + 1);
-// 		strcpy(node->value, valueArray[i]);
-
-// 		rbtree_insert(T, node);
-		
-// 	}
-
-// 	rbtree_traversal(T, T->root);
-// 	printf("----------------------------------------\n");
-
-// 	for (i = 0;i < 10;i ++) {
-
-// 		rbtree_node *node = rbtree_search(T, keyArray[i]);
-// 		rbtree_node *cur = rbtree_delete(T, node);
-// 		free(cur);
-
-// 		rbtree_traversal(T, T->root);
-// 		printf("----------------------------------------\n");
-// 	}
-
-// #else
-
-
-// 	int keyArray[20] = {24,25,13,35,23, 26,67,47,38,98, 20,19,17,49,12, 21,9,18,14,15};
-
-// 	rbtree *T = (rbtree *)malloc(sizeof(rbtree));
-// 	if (T == NULL) {
-// 		printf("malloc failed\n");
-// 		return -1;
-// 	}
-	
-// 	T->nil = (rbtree_node*)malloc(sizeof(rbtree_node));
-// 	T->nil->color = BLACK;
-// 	T->root = T->nil;
-
-// 	rbtree_node *node = T->nil;
-// 	int i = 0;
-// 	for (i = 0;i < 20;i ++) {
-// 		node = (rbtree_node*)malloc(sizeof(rbtree_node));
-// 		node->key = keyArray[i];
-// 		node->value = NULL;
-
-// 		rbtree_insert(T, node);
-		
-// 	}
-
-// 	rbtree_traversal(T, T->root);
-// 	printf("----------------------------------------\n");
-
-// 	for (i = 0;i < 20;i ++) {
-
-// 		rbtree_node *node = rbtree_search(T, keyArray[i]);
-// 		rbtree_node *cur = rbtree_delete(T, node);
-// 		free(cur);
-
-// 		rbtree_traversal(T, T->root);
-// 		printf("----------------------------------------\n");
-// 	}
-// #endif
-
-	
-// }
-
-// #endif
-
-
-// typedef struct _rbtree kvs_rbtree_t; 
-
-// kvs_rbtree_t global_rbtree;
-
-// // 5 + 2
-// int kvs_rbtree_create(kvs_rbtree_t *inst) {
-
-// 	if (inst == NULL) return 1;
-
-// 	inst->nil = (rbtree_node*)kvs_malloc(sizeof(rbtree_node));
-// 	inst->nil->color = BLACK;
-// 	inst->root = inst->nil;
-
-// 	return 0;
-
-// }
-
-// void kvs_rbtree_destroy(kvs_rbtree_t *inst) {
-
-// 	if (inst == NULL) return ;
-
-// 	rbtree_node *node = NULL;
-
-// 	while (!(node = inst->root)) {
-		
-// 		rbtree_node *mini = rbtree_mini(inst, node);
-		
-// 		rbtree_node *cur = rbtree_delete(inst, mini);
-// 		kvs_free(cur);
-		
-// 	}
-
-// 	kvs_free(inst->nil);
-
-// 	return ;
-
-// }
-
-
-// int kvs_rbtree_set(kvs_rbtree_t *inst, char *key, char *value) {
-
-// 	if (!inst || !key || !value) return -1;
-
-// 	rbtree_node *node = (rbtree_node*)kvs_malloc(sizeof(rbtree_node));
-		
-// 	node->key = kvs_malloc(strlen(key) + 1);
-// 	if (!node->key) return -2;
-//  	memset(node->key, 0, strlen(key) + 1);
-// 	strcpy(node->key, key);
-	
-// 	node->value = kvs_malloc(strlen(value) + 1);
-// 	if (!node->value) return -2;
-// 	memset(node->value, 0, strlen(value) + 1);
-// 	strcpy(node->value, value);
-
-// 	rbtree_insert(inst, node);
-
-// 	return 0;
-// }
-
-
-// char* kvs_rbtree_get(kvs_rbtree_t *inst, char *key)  {
-
-// 	if (!inst || !key) return NULL;
-// 	rbtree_node *node = rbtree_search(inst, key);
-// 	if (!node) return NULL; // no exist
-// 	if (node == inst->nil) return NULL;
-
-// 	return node->value;
-	
-// }
-
-// int kvs_rbtree_del(kvs_rbtree_t *inst, char *key) {
-
-// 	if (!inst || !key) return -1;
-
-// 	rbtree_node *node = rbtree_search(inst, key);
-// 	if (!node) return 1; // no exist
-	
-// 	rbtree_node *cur = rbtree_delete(inst, node);
-// 	free(cur);
-
-// 	return 0;
-// }
-
-// int kvs_rbtree_mod(kvs_rbtree_t *inst, char *key, char *value) {
-
-// 	if (!inst || !key || !value) return -1;
-
-// 	rbtree_node *node = rbtree_search(inst, key);
-// 	if (!node) return 1; // no exist
-// 	if (node == inst->nil) return 1;
-	
-// 	kvs_free(node->value);
-
-// 	node->value = kvs_malloc(strlen(value) + 1);
-// 	if (!node->value) return -2;
-	
-// 	memset(node->value, 0, strlen(value) + 1);
-// 	strcpy(node->value, value);
-
-// 	return 0;
-
-// }
-
-// int kvs_rbtree_exist(kvs_rbtree_t *inst, char *key) {
-
-// 	if (!inst || !key) return -1;
-
-// 	rbtree_node *node = rbtree_search(inst, key);
-// 	if (!node) return 1; // no exist
-// 	if (node == inst->nil) return 1;
-
-// 	return 0;
-// }
-
-

kvs_rbtree_bin.c

@@ -1,11 +1,47 @@
-
-
-
 #include "kvstore.h"
 #include "kvs_rw_tools.h"
 #include "memory/alloc_dispatch.h"
 #include "diskuring/diskuring.h"
 
+/* ============================================================================
+ * 内存布局说明：
+ * ============================================================================
+ * 每个节点的内存结构（单一连续块）：
+ * 
+ * +------ 固定头部 (24字节) -------+------ 动态数据 -------+
+ * | color | right | left | parent | key_len | value_len | key | value |
+ * | 1字节 |8字节 |8字节  |8字节  | 4字节   | 4字节    | k字节| v字节 |
+ * +---------- 共32字节 ------+--- key_len + value_len 字节 ---+
+ * 
+ * 总大小 = sizeof(rbtree_node_fixed) + key_len + value_len
+ * ============================================================================ */
+
+// ============================================================================
+// 辅助函数：计算节点所需的总大小
+// ============================================================================
+static inline size_t rbtree_node_size(uint32_t key_len, uint32_t value_len) {
+    return sizeof(rbtree_node_fixed) + key_len + value_len;
+}
+
+// ============================================================================
+// 辅助函数：获取节点内的key指针
+// ============================================================================
+static inline uint8_t* rbtree_node_get_key(rbtree_node *node) {
+    if (!node || node->key_len == 0) return NULL;
+    return (uint8_t *)node + sizeof(rbtree_node_fixed);
+}
+
+// ============================================================================
+// 辅助函数：获取节点内的value指针
+// ============================================================================
+static inline uint8_t* rbtree_node_get_value(rbtree_node *node) {
+    if (!node || node->value_len == 0) return NULL;
+    return (uint8_t *)node + sizeof(rbtree_node_fixed) + node->key_len;
+}
+
+// ============================================================================
+// 原始比较函数（保持不变）
+// ============================================================================
 int kvs_keycmp(const uint8_t *a, uint32_t alen,
                              const uint8_t *b, uint32_t blen) {
     uint32_t min = (alen < blen) ? alen : blen;
@@ -150,7 +186,9 @@ int rbtree_insert(rbtree *T, rbtree_node *z) {
 	while (x != T->nil) {
 		y = x;
 
-		int c = kvs_keycmp(z->key, z->key_len, x->key, x->key_len);
+		uint8_t *xkey = rbtree_node_get_key(x);
+		uint8_t *zkey = rbtree_node_get_key(z);
+		int c = kvs_keycmp(zkey, z->key_len, xkey, x->key_len);
 		if (c < 0) {
 			x = x->left;
 		} else if (c > 0) {
@@ -166,7 +204,9 @@ int rbtree_insert(rbtree *T, rbtree_node *z) {
 		T->root = z;
 
 	}else{
-		int c = kvs_keycmp(z->key, z->key_len, y->key, y->key_len);
+		uint8_t *ykey = rbtree_node_get_key(y);
+		uint8_t *zkey = rbtree_node_get_key(z);
+		int c = kvs_keycmp(zkey, z->key_len, ykey, y->key_len);
 		if (c < 0) y->left = z;
         else       y->right = z;
 	}
@@ -275,11 +315,52 @@ rbtree_node *rbtree_delete(rbtree *T, rbtree_node *z) {
 	}
 
 	if (y != z) {
-		uint8_t  *ktmp = z->key;      z->key = y->key;      y->key = ktmp;
+		// 交换键值：由于键和值内嵌在节点内存中，需要交换内存内容
-        uint32_t  ltmp = z->key_len;   z->key_len = y->key_len; y->key_len = ltmp;
+		// 注意：这里假设 z 的内存大小足够容纳 y 的数据
+		// 更安全的做法是只交换指针或重新分配
 		
-		uint8_t  *vtmp = z->value;       z->value = y->value;       y->value = vtmp;
+		// 保存原始长度
-        uint32_t  tlen = z->value_len;    z->value_len = y->value_len; y->value_len = tlen;
+		uint32_t z_klen = z->key_len;
+		uint32_t z_vlen = z->value_len;
+		uint32_t y_klen = y->key_len;
+		uint32_t y_vlen = y->value_len;
+		
+		uint8_t *z_key = rbtree_node_get_key(z);
+		uint8_t *z_val = rbtree_node_get_value(z);
+		uint8_t *y_key = rbtree_node_get_key(y);
+		uint8_t *y_val = rbtree_node_get_value(y);
+		
+		// 如果长度相同，直接交换内存
+		if (z_klen == y_klen && z_vlen == y_vlen) {
+			if (z_klen > 0) memcpy(z_key, y_key, z_klen);
+			if (z_vlen > 0) {
+				uint8_t tmp[z_vlen];
+				memcpy(tmp, z_val, z_vlen);
+				memcpy(z_val, y_val, z_vlen);
+				memcpy(y_val, tmp, z_vlen);
+			}
+			if (z_klen > 0) {
+				uint8_t tmp[z_klen];
+				memcpy(tmp, z_key, z_klen);
+				memcpy(z_key, y_key, z_klen);
+				memcpy(y_key, tmp, z_klen);
+			}
+		} else {
+			// 长度不同时，只能交换值的拷贝（保留长度不变）
+			// 这是一个限制，实际应用中需要重新分配更大的节点
+			if (z_klen == y_klen && z_klen > 0) {
+				uint8_t tmp[z_klen];
+				memcpy(tmp, z_key, z_klen);
+				memcpy(z_key, y_key, z_klen);
+				memcpy(y_key, tmp, z_klen);
+			}
+			if (z_vlen == y_vlen && z_vlen > 0) {
+				uint8_t tmp[z_vlen];
+				memcpy(tmp, z_val, z_vlen);
+				memcpy(z_val, y_val, z_vlen);
+				memcpy(y_val, tmp, z_vlen);
+			}
+		}
 	}
 
 	if (y->color == BLACK) {
@@ -289,11 +370,12 @@ rbtree_node *rbtree_delete(rbtree *T, rbtree_node *z) {
 	return y;
 }
 
-rbtree_node *rbtree_search(rbtree *T, KEY_TYPE* key, uint32_t keylen) {
+rbtree_node *rbtree_search(rbtree *T, const uint8_t *key, uint32_t keylen) {
 
 	rbtree_node *node = T->root;
 	while (node != T->nil) {
-        int c = kvs_keycmp(key, keylen, node->key, node->key_len);
+        uint8_t *node_key = rbtree_node_get_key(node);
+        int c = kvs_keycmp(key, keylen, node_key, node->key_len);
         if (c < 0) node = node->left;
         else if (c > 0) node = node->right;
         else return node;
@@ -306,7 +388,8 @@ void rbtree_traversal(rbtree *T, rbtree_node *node) {
 	if (node != T->nil) {
 		rbtree_traversal(T, node->left);
 
-		printf("key:%s, color:%d\n", (char*)node->key, node->color);
+		uint8_t *key = rbtree_node_get_key(node);
+		printf("key:%s, color:%d\n", (char*)key, node->color);
 		rbtree_traversal(T, node->right);
 	}
 }
@@ -316,16 +399,21 @@ typedef struct _rbtree kvs_rbtree_t;
 
 kvs_rbtree_t global_rbtree;
 
-// 5 + 2
+// ============================================================================
+// 创建红黑树
+// ============================================================================
 int kvs_rbtree_create(kvs_rbtree_t *inst) {
 
 	if (inst == NULL) return 1;
 
-	inst->nil = (rbtree_node*)kvs_malloc(sizeof(rbtree_node));
+	// nil 节点：特殊的哨兵节点，也使用优化的分配
+	inst->nil = (rbtree_node*)kvs_malloc(sizeof(rbtree_node_fixed));
 	if (!inst->nil) return 2;
 
 	inst->nil->color = BLACK;
 	inst->nil->left = inst->nil->right = inst->nil->parent = inst->nil;
+	inst->nil->key_len = 0;
+	inst->nil->value_len = 0;
 	inst->root = inst->nil;
 
 	return 0;
@@ -340,13 +428,11 @@ void kvs_rbtree_destroy(kvs_rbtree_t *inst) {
 
 	while (inst->root != inst->nil) {
 		
-		rbtree_node *mini = rbtree_mini(inst, node);
+		rbtree_node *mini = rbtree_mini(inst, inst->root);
 		
 		rbtree_node *cur = rbtree_delete(inst, mini);
 		if (cur != inst->nil) {
-            if (cur->key)   kvs_free(cur->key);
+            kvs_free(cur);  // 只需释放节点本身，key和value已内嵌
-            if (cur->value) kvs_free(cur->value);
-            kvs_free(cur);
         }
 		
 	}
@@ -362,48 +448,99 @@ void kvs_rbtree_destroy(kvs_rbtree_t *inst) {
  * @return: <0 error; 0 success; 1 exist
  */
 int kvs_rbtree_set(kvs_rbtree_t *inst, const void *key, uint32_t key_len, const void *value, uint32_t value_len) {
+    if (!inst || !key || !value) return -1;
 
-	if (!inst || !key || !value) return -1;
+    // 1. 查找键是否已存在
+    rbtree_node *existing = rbtree_search(inst, (const uint8_t*)key, key_len);
+    if (existing != inst->nil) {
+        // 键已存在：需要重新分配节点（如果大小改变）
+        uint32_t old_size = rbtree_node_size(existing->key_len, existing->value_len);
+        uint32_t new_size = rbtree_node_size(key_len, value_len);
         
-	rbtree_node *node = (rbtree_node*)kvs_malloc(sizeof(rbtree_node));
+        if (new_size != old_size) {
-	if (!node) return -2;
+            // 大小改变，需要重新分配并更新树结构
-    memset(node, 0, sizeof(*node));
+            rbtree_node *new_node = (rbtree_node*)kvs_malloc(new_size);
+            if (!new_node) return -2;
             
-	node->key = (uint8_t*)kvs_malloc(key_len);
+            // 复制固定部分（除了 key_len 和 value_len）
-	if (!node->key) {
+            new_node->color = existing->color;
-		kvs_free(node);return -2;
+            new_node->right = existing->right;
-	}
+            new_node->left = existing->left;
- 	memcpy(node->key, key, key_len);
+            new_node->parent = existing->parent;
-	node->key_len = key_len;
+            new_node->key_len = key_len;
+            new_node->value_len = value_len;
             
-	node->value = (uint8_t*)kvs_malloc(value_len);
+            // 复制 key 和 value
-	if (!node->value) { kvs_free(node->key); kvs_free(node); return -2; }
+            uint8_t *new_key = rbtree_node_get_key(new_node);
-	if (value_len) memcpy(node->value, value, value_len);
+            uint8_t *new_val = rbtree_node_get_value(new_node);
-	node->value_len = value_len;
+            if (key_len > 0) memcpy(new_key, key, key_len);
+            if (value_len > 0) memcpy(new_val, value, value_len);
             
-	if(rbtree_insert(inst, node) < 0){
+            // 更新父节点的指针
-		kvs_free(node->value);
+            if (existing->parent != inst->nil) {
-        kvs_free(node->key);
+                if (existing->parent->left == existing) {
+                    existing->parent->left = new_node;
+                } else {
+                    existing->parent->right = new_node;
+                }
+            } else {
+                inst->root = new_node;
+            }
+            
+            // 更新子节点的父指针
+            if (new_node->left != inst->nil) {
+                new_node->left->parent = new_node;
+            }
+            if (new_node->right != inst->nil) {
+                new_node->right->parent = new_node;
+            }
+            
+            kvs_free(existing);
+        } else {
+            // 大小相同，直接更新值
+            uint8_t *val = rbtree_node_get_value(existing);
+            if (value_len > 0) memcpy(val, value, value_len);
+        }
+        return 0;
+    }
+
+    // 2. 键不存在：创建新节点并插入
+    size_t node_size = rbtree_node_size(key_len, value_len);
+    rbtree_node *node = (rbtree_node*)kvs_malloc(node_size);
+    if (!node) return -2;
+    
+    memset(node, 0, node_size);
+
+    node->key_len = key_len;
+    node->value_len = value_len;
+
+    uint8_t *node_key = rbtree_node_get_key(node);
+    uint8_t *node_val = rbtree_node_get_value(node);
+    
+    if (key_len > 0) memcpy(node_key, key, key_len);
+    if (value_len > 0) memcpy(node_val, value, value_len);
+
+    if (rbtree_insert(inst, node) < 0) {
+        // 插入失败，释放资源
         kvs_free(node);
-		return 1;
+        return -2;
-	}
+    }
 
-	return 0;
+    return 0;
 }
 
 
 /*
- * @return: NULL notexist, NOTNULL exist。out_value_len 是长度。
+ * @return: NULL not exist, NOTNULL exist。out_value_len 是长度。
  */
 void* kvs_rbtree_get(kvs_rbtree_t *inst, const void *key, uint32_t key_len, uint32_t *out_valuelen)  {
 	if (!inst || !key || key_len == 0 || !out_valuelen) return NULL;
 	
-	rbtree_node *node = rbtree_search(inst, (uint8_t *)key, key_len);
+	rbtree_node *node = rbtree_search(inst, (const uint8_t *)key, key_len);
-	if (!node) return NULL; // no exist
+	if (!node || node == inst->nil) return NULL;
-	if (node == inst->nil) return NULL;
 
 	*out_valuelen = node->value_len;
-	return node->value;
+	return (void*)rbtree_node_get_value(node);
 }
 
 /*
@@ -413,15 +550,12 @@ int kvs_rbtree_del(rbtree *inst, const void *key, uint32_t key_len) {
 
 	if (!inst || !key || key_len == 0) return -1;
 
-	rbtree_node *node = rbtree_search(inst, (uint8_t *)key, key_len);
+	rbtree_node *node = rbtree_search(inst, (const uint8_t *)key, key_len);
-	if (!node) return 1; // no exist
+	if (!node || node == inst->nil) return 1;
-	if (node == inst->nil) return 1;
 	
 	rbtree_node *cur = rbtree_delete(inst, node);
 	if (cur != inst->nil) {
-        if (cur->key)   kvs_free(cur->key);
+        kvs_free(cur);  // 只需释放节点本身
-        if (cur->value) kvs_free(cur->value);
-        kvs_free(cur);
     }
 
 	return 0;
@@ -434,20 +568,57 @@ int kvs_rbtree_mod(kvs_rbtree_t *inst, const void *key, uint32_t key_len, const
 
 	if (!inst || !key || key_len==0 || !value) return -1;
 
-	rbtree_node *node = rbtree_search(inst, (uint8_t *)key, key_len);
+	rbtree_node *node = rbtree_search(inst, (const uint8_t *)key, key_len);
-	if (!node) return 1; // no exist
+	if (!node || node == inst->nil) return 1;
-	if (node == inst->nil) return 1;
 	
-	if (node->value) kvs_free(node->value);
+	// 如果新的 value_len 与旧的相同，可以直接覆盖
+	if (node->value_len == value_len) {
+		uint8_t *val = rbtree_node_get_value(node);
+		if (value_len > 0) memcpy(val, value, value_len);
+		return 0;
+	}
 	
-	node->value = (uint8_t*)kvs_malloc(value_len);
+	// 否则需要重新分配节点
-	if (!node->value) { node->value_len = 0; return -2; }
+	uint32_t new_size = rbtree_node_size(key_len, value_len);
+	rbtree_node *new_node = (rbtree_node*)kvs_malloc(new_size);
+	if (!new_node) return -2;
 	
-	if (value_len) memcpy(node->value, value, value_len);
+	// 复制所有内容
-	node->value_len = value_len;
+	uint8_t *old_key = rbtree_node_get_key(node);
+	uint8_t *new_key = rbtree_node_get_key(new_node);
+	uint8_t *new_val = rbtree_node_get_value(new_node);
 	
+	new_node->color = node->color;
+	new_node->left = node->left;
+	new_node->right = node->right;
+	new_node->parent = node->parent;
+	new_node->key_len = node->key_len;
+	new_node->value_len = value_len;
+	
+	if (key_len > 0) memcpy(new_key, old_key, key_len);
+	if (value_len > 0) memcpy(new_val, value, value_len);
+	
+	// 更新父节点指针
+	if (node->parent != inst->nil) {
+		if (node->parent->left == node) {
+			node->parent->left = new_node;
+		} else {
+			node->parent->right = new_node;
+		}
+	} else {
+		inst->root = new_node;
+	}
+	
+	// 更新子节点的父指针
+	if (new_node->left != inst->nil) {
+		new_node->left->parent = new_node;
+	}
+	if (new_node->right != inst->nil) {
+		new_node->right->parent = new_node;
+	}
+	
+	kvs_free(node);
 	return 0;
-
 }
 
 /*
@@ -457,9 +628,8 @@ int kvs_rbtree_exist(kvs_rbtree_t *inst, const void *key, uint32_t key_len) {
 
 	if (!inst || !key || key_len == 0) return -1;
 
-	rbtree_node *node = rbtree_search(inst, (uint8_t*)key, key_len);
+	rbtree_node *node = rbtree_search(inst, (const uint8_t*)key, key_len);
-	if (!node) return 1; // no exist
+	if (!node || node == inst->nil) return 1;
-	if (node == inst->nil) return 1;
 
 	return 0;
 }
@@ -488,14 +658,16 @@ static int kvs_rbtree_save_node(iouring_ctx_t *uring, int fd, off_t *current_off
 	lens[count] = sizeof(vlen);
 	count++;
 
+	uint8_t *node_key = rbtree_node_get_key(node);
 	if (node->key_len > 0) {
-		bufs[count] = node->key;
+		bufs[count] = node_key;
 		lens[count] = node->key_len;
 		count++;
 	}
 
+	uint8_t *node_val = rbtree_node_get_value(node);
 	if (node->value_len > 0) {
-		bufs[count] = node->value;
+		bufs[count] = node_val;
 		lens[count] = node->value_len;
 		count++;
 	}
@@ -504,7 +676,7 @@ static int kvs_rbtree_save_node(iouring_ctx_t *uring, int fd, off_t *current_off
     for (int i = 0; i < count; i++) total += lens[i];
 
 	task_t *t = submit_write(uring, fd, bufs, lens, count, *current_off);
-	cleanup_finished_iouring_tasks();
+	cleanup_finished_iouring_tasks(uring);
 	
 	if(!t) {
 		perror("task init failed");
@@ -532,7 +704,7 @@ int kvs_rbtree_save(iouring_ctx_t *uring, kvs_rbtree_t *inst, const char* filena
 
 	while (!uring_task_complete(uring)) {
 		usleep(1000);
-		cleanup_finished_iouring_tasks();
+		cleanup_finished_iouring_tasks(uring);
 	}
 	close(fd);
     return rc; 
@@ -554,34 +726,37 @@ int kvs_rbtree_load(kvs_rbtree_t *inst, const char* filename){
         uint32_t vlen = ntohl(vlen_n);
 
 		if (klen == 0) { fclose(fp); return -3; }
-		uint8_t *keybuf = (uint8_t*)kvs_malloc((size_t)klen);
+		
-		if (!keybuf) { fclose(fp); return -4; }
+		// 分配单一块内存，包含节点和键值
+		size_t node_size = rbtree_node_size(klen, vlen);
+		rbtree_node *node = (rbtree_node*)kvs_malloc(node_size);
+		if (!node) { fclose(fp); return -4; }
+		
+		memset(node, 0, node_size);
+		node->key_len = klen;
+		node->value_len = vlen;
+		
+		uint8_t *keybuf = rbtree_node_get_key(node);
         if (kvs_read_file(fp, keybuf, (size_t)klen) < 0) {
-            kvs_free(keybuf);
+            kvs_free(node);
             fclose(fp);
             return -3;
         }
 
 		uint8_t *valbuf = NULL;
         if (vlen > 0) {
-            valbuf = (uint8_t*)kvs_malloc((size_t)vlen);
+            valbuf = rbtree_node_get_value(node);
-            if (!valbuf) {
-                kvs_free(keybuf);
-                fclose(fp);
-                return -4;
-            }
             if (kvs_read_file(fp, valbuf, (size_t)vlen) < 0) {
-                kvs_free(valbuf);
+                kvs_free(node);
-                kvs_free(keybuf);
                 fclose(fp);
                 return -3;
             }
         }
 
-		int rc = kvs_rbtree_set(inst, keybuf, klen, valbuf, vlen);
+		// 使用原生 rbtree_insert 而非 kvs_rbtree_set
-		if (vlen > 0) kvs_free(valbuf);
+		// 因为 kvs_rbtree_set 会重新分配节点
-
+		if (rbtree_insert(inst, node) < 0) {
-		if (rc < 0) { // error
+            kvs_free(node);
             fclose(fp);
             return -5;
         }

kvstore.c

@@ -19,6 +19,9 @@
 #include <unistd.h>
 #include <arpa/inet.h>
 #include <libxml/parser.h>
+#include <limits.h>
+
+#define 	TIME_COLLECT	0
 
 extern int slave_bootstrap(const char *listen_ip, int listen_port, const char *master_ip, int master_port);
 
@@ -38,26 +41,136 @@ void __completed_cmd(const uint8_t *cmd, size_t len, unsigned long long seq){
 
 }
 
-// __attribute__((noinline))
+#include <sys/time.h>
-// void __replica_notify(uint64_t seq, uint32_t off, uint32_t len)
+#define TIME_SUB_MS(tv1, tv2)  ((tv1.tv_sec - tv2.tv_sec) * 1000 + (tv1.tv_usec - tv2.tv_usec) / 1000)
-// {
+#define TIME_SUB_US(tv1, tv2)  ((tv1.tv_sec - tv2.tv_sec) * 1000000 + (tv1.tv_usec - tv2.tv_usec))
-//     // 空函数即可，目的是让 uprobe 拿到参数
+
-//     asm volatile("" ::: "memory");
+static int checked_size_add(size_t a, size_t b, size_t *out) {
-// }
+    if (!out || a > SIZE_MAX - b) {
+        return -1;
+    }
+    *out = a + b;
+    return 0;
+}
+
+static int resp_value_encoded_len(const resp_value_t *v, size_t *out_len) {
+    size_t len = 0;
+
+    if (!v || !out_len) {
+        return -1;
+    }
+
+    switch (v->type) {
+    case RESP_T_SIMPLE_STR:
+    case RESP_T_ERROR:
+        if (checked_size_add(1, (size_t)v->bulk.len, &len) < 0 ||
+            checked_size_add(len, 2, &len) < 0) {
+            return -1;
+        }
+        break;
+
+    case RESP_T_INTEGER: {
+        char tmp[64];
+        int n = snprintf(tmp, sizeof(tmp), "%lld", (long long)v->i64);
+        if (n <= 0) {
+            return -1;
+        }
+        if (checked_size_add(1, (size_t)n, &len) < 0 ||
+            checked_size_add(len, 2, &len) < 0) {
+            return -1;
+        }
+        break;
+    }
+
+    case RESP_T_NIL:
+        len = 5; /* "$-1\r\n" */
+        break;
+
+    case RESP_T_BULK_STR: {
+        char tmp[32];
+        int n;
+        size_t t;
+
+        if (v->bulk.len > 0 && !v->bulk.ptr) {
+            return -1;
+        }
+
+        n = snprintf(tmp, sizeof(tmp), "%u", (unsigned)v->bulk.len);
+        if (n <= 0) {
+            return -1;
+        }
+
+        if (checked_size_add(1, (size_t)n, &t) < 0 || /* '$' + len digits */
+            checked_size_add(t, 2, &t) < 0 ||          /* \r\n */
+            checked_size_add(t, (size_t)v->bulk.len, &t) < 0 ||
+            checked_size_add(t, 2, &len) < 0) {        /* trailing \r\n */
+            return -1;
+        }
+        break;
+    }
+
+    default:
+        return -1;
+    }
+
+    *out_len = len;
+    return 0;
+}
+
+static int flush_pending_response(struct conn *conn, uint8_t *buf, size_t *out_len) {
+    if (!conn || !buf || !out_len) {
+        return -1;
+    }
+    if (*out_len == 0) {
+        return 0;
+    }
+    if (chain_buffer_append(&conn->wbuf, buf, *out_len) < 0) {
+        return -1;
+    }
+    *out_len = 0;
+    return 0;
+}
+
+static int is_update_cmd(const resp_cmd_t *cmd) {
+    const resp_slice_t *c0;
+
+    if (!cmd || cmd->argc == 0 || !cmd->argv[0].ptr || cmd->argv[0].len == 0) {
+        return 0;
+    }
+
+    c0 = &cmd->argv[0];
+    return ascii_casecmp(c0->ptr, c0->len, "SET") == 0 ||
+           ascii_casecmp(c0->ptr, c0->len, "DEL") == 0 ||
+           ascii_casecmp(c0->ptr, c0->len, "MOD") == 0 ||
+           ascii_casecmp(c0->ptr, c0->len, "RSET") == 0 ||
+           ascii_casecmp(c0->ptr, c0->len, "RDEL") == 0 ||
+           ascii_casecmp(c0->ptr, c0->len, "RMOD") == 0 ||
+           ascii_casecmp(c0->ptr, c0->len, "HSET") == 0 ||
+           ascii_casecmp(c0->ptr, c0->len, "HDEL") == 0 ||
+           ascii_casecmp(c0->ptr, c0->len, "HMOD") == 0;
+}
 
 int kvs_protocol(struct conn* conn){
-	if (!conn) return -1;
+#if TIME_COLLECT == 1
-	char *request = conn->rbuffer;
+	struct timeval func_start;
-	int request_length = conn->rlength;
+	gettimeofday(&func_start, NULL);
-	char *response = conn->wbuffer;
+	long total_oplog_us = 0; 
-	int *response_length = &conn->wlength;
+#endif
 
-	if (!request || request_length <= 0 || !response || !response_length) return -1;
+	if (!conn) return -1;
+
+	size_t request_size = 0;
+	const uint8_t *request = chain_buffer_linearize(&conn->rbuf, &request_size);
+	if (!request || request_size == 0) return 0;
+	if (request_size > (size_t)INT_MAX) return -1;
+
+	int request_length = (int)request_size;
+	uint8_t response[KVS_MAX_RESPONSE];
 	
 	int consumed = 0;
-	int out_len = 0;
+    size_t out_len = 0;
 
-	while(consumed < request_length ){
+ 	while(consumed < request_length ){
 		const uint8_t *p = request+consumed;
 		int remain = request_length - consumed;
 		
@@ -66,8 +179,7 @@ int kvs_protocol(struct conn* conn){
 
 		int len = resp_parse_one_cmd(p, remain, &cmd);
 		if(len < 0){
-			/* 协议错误：直接返回，已构建的响应仍可写回 */
+			/* 协议错误：直接返回 */
-			*response_length = out_len;
 			return -1;
 		}
 		else if(len == 0){
@@ -86,78 +198,103 @@ int kvs_protocol(struct conn* conn){
          *   一般也已经把 out_value 设置成了 RESP error，这样客户端能收到错误响应。
          * - 如果 dr < 0 但 val.type 没被正确设置，兜底回一个通用错误。
          */
-		if(dr < 0){
+#if TIME_COLLECT == 1
-			if (val.type != RESP_T_SIMPLE_STR &&
+		struct timeval oplog_start, oplog_end;
-                val.type != RESP_T_ERROR &&
+		gettimeofday(&oplog_start, NULL);
-                val.type != RESP_T_INTEGER &&
+#endif
-                val.type != RESP_T_BULK_STR &&
+
-                val.type != RESP_T_NIL) {
+		int need_persist = is_update_cmd(&cmd);
-                val = resp_error("ERR dispatch failed");
+
+		if(global_cfg.persistence == PERSIST_INCREMENTAL && need_persist){			
+            int ar = kvs_oplog_buffer_append(p, (size_t)len, global_oplog_fd);
+            if (ar < 0) {
+                return -1;
+            }
+            if (ar == KVS_OPLOG_BUF_FULL && kvs_oplog_flush(global_oplog_fd, 0) < 0) {
+                return -1;
             }
-		} else {
-			// persist into oplog 
-			/* 执行成功：在这里保存到日志中（只记录更新类命令） */
-			if (cmd.argc > 0 && cmd.argv[0].ptr) {
-				/* 更新类命令：SET/DEL/MOD/RSET/RDEL/RMOD/HSET/HDEL/HMOD/SAVE */
-				const resp_slice_t *c0 = &cmd.argv[0];
-				int is_update = 0;
-				if (c0->ptr && c0->len) {
-					if (ascii_casecmp(c0->ptr, c0->len, "SET") == 0 ||
-						ascii_casecmp(c0->ptr, c0->len, "DEL") == 0 ||
-						ascii_casecmp(c0->ptr, c0->len, "MOD") == 0 ||
-						ascii_casecmp(c0->ptr, c0->len, "RSET") == 0 ||
-						ascii_casecmp(c0->ptr, c0->len, "RDEL") == 0 ||
-						ascii_casecmp(c0->ptr, c0->len, "RMOD") == 0 ||
-						ascii_casecmp(c0->ptr, c0->len, "HSET") == 0 ||
-						ascii_casecmp(c0->ptr, c0->len, "HDEL") == 0 ||
-						ascii_casecmp(c0->ptr, c0->len, "HMOD") == 0) {
-						is_update = 1;
-					}
-				}
-
-				if (is_update) {
-					if(global_cfg.persistence == PERSIST_INCREMENTAL){			
-						kvs_oplog_append(p, len, global_oplog_fd);
-					}	
-
-					// __completed_cmd(p, len, global_seq);
-					// global_seq ++;
-					
-					if (global_cfg.replica_mode == REPLICA_ENABLE) {
-						uint32_t off = 0;
-						int ar = replica_shm_append(&g_rep_shm, global_seq, p, (uint32_t)len, &off);
-						if (ar == 0) {
-							// __replica_notify(global_seq, off, (uint32_t)len);
-							global_seq++;
-
-						} else {
-							// shm 满或异常：你可以选择降级（比如直接跳过复制，或阻塞/丢弃）
-							// 为了不影响主路径，这里先打印并跳过
-							fprintf(stderr, "replica_shm_append failed %d\n", ar);
-						}
-					}
-				}
-			}
 		}	
 
+		// __completed_cmd(p, len, global_seq);
+		// global_seq ++;
+		if (global_cfg.replica_mode == REPLICA_ENABLE && need_persist) {
+			uint32_t off = 0;
+			int ar = replica_shm_append(&g_rep_shm, global_seq, p, (uint32_t)len, &off);
+			if (ar == 0) {
+				// __replica_notify(global_seq, off, (uint32_t)len);
+				global_seq++;
+
+			} else {
+				// shm 满或异常：你可以选择降级（比如直接跳过复制，或阻塞/丢弃）
+				// 为了不影响主路径，这里先打印并跳过
+				fprintf(stderr, "replica_shm_append failed %d\n", ar);
+			}
+		}
+#if TIME_COLLECT == 1
+		gettimeofday(&oplog_end, NULL);
+		total_oplog_us += (oplog_end.tv_sec - oplog_start.tv_sec) * 1000000 +
+                      (oplog_end.tv_usec - oplog_start.tv_usec);
+#endif
+
         /* 构建响应 */
-        int cap = KVS_MAX_RESPONSE - out_len;
+        int resp_len = resp_build_value(&val, response + out_len, sizeof(response) - out_len);
-        if (cap <= 0) {
-            *response_length = out_len;
-            return consumed;
-        }
-
-		int resp_len = resp_build_value(&val, response + out_len, (size_t)cap);
         if (resp_len < 0) {
-            *response_length = out_len;
+            /* 当前批次剩余空间不够，先把已拼好的刷到发送队列再重试 */
-            return consumed;
+            if (flush_pending_response(conn, response, &out_len) < 0) {
+                return -1;
+            }
+
+            resp_len = resp_build_value(&val, response, sizeof(response));
+            if (resp_len < 0) {
+                size_t resp_need = 0;
+                uint8_t *resp_heap = NULL;
+
+                if (resp_value_encoded_len(&val, &resp_need) < 0) {
+                    return -1;
+                }
+
+                resp_heap = (uint8_t *)kvs_malloc(resp_need);
+                if (!resp_heap) {
+                    return -1;
+                }
+
+                resp_len = resp_build_value(&val, resp_heap, resp_need);
+                if (resp_len < 0 ||
+                    chain_buffer_append(&conn->wbuf, resp_heap, (size_t)resp_len) < 0) {
+                    free(resp_heap);
+                    return -1;
+                }
+
+                free(resp_heap);
+                resp_len = 0;
+            }
         }
 
-		out_len += resp_len;
+        out_len += (size_t)resp_len;
+
+		// __completed_cmd(request, consumed, 0);
+
 		consumed += len;
 	}
 
-    *response_length = out_len;
+    if (global_cfg.persistence == PERSIST_INCREMENTAL) {
+        if (kvs_oplog_flush(global_oplog_fd, 1) < 0) {
+            return -1;
+        }
+    }
+
+	#if TIME_COLLECT == 1
+    struct timeval func_end;
+    gettimeofday(&func_end, NULL);
+    long func_us = (func_end.tv_sec - func_start.tv_sec) * 1000000 +
+                   (func_end.tv_usec - func_start.tv_usec);
+	fprintf(stderr, "kvs_protocol: total %ld us, oplog %ld us\n", func_us, total_oplog_us);
+#endif 
+
+    if (flush_pending_response(conn, response, &out_len) < 0) {
+        return -1;
+    }
+
 	return consumed;
 }
 
@@ -278,6 +415,7 @@ int init_config(AppConfig *cfg){
 	printf("Persistence 		: %s\n", persistence_to_string(cfg->persistence));
     printf("|—— Persist-dir 	: %s\n", cfg->persist_dir);
     printf("|—— Persist-oplog 	: %s\n", cfg->oplog_file);
+    printf("|—— Oplog-sync   	: %s\n", oplog_sync_mode_to_string(cfg->oplog_sync_mode));
     printf("|—— Persist-array	: %s\n", cfg->array_file);
     printf("|—— Persist-rbtree	: %s\n", cfg->rbtree_file);
     printf("|—— Persist-hash	: %s\n", cfg->hash_file);

kvstore.h

@@ -120,15 +120,26 @@ int kvs_array_exist(kvs_array_t *inst, char *key);
 #if BIN_SAFE
 typedef uint8_t KEY_TYPE; // key
 
+// 固定部分结构
+typedef struct {
+    unsigned char color;
+    struct _rbtree_node *right;
+    struct _rbtree_node *left;
+    struct _rbtree_node *parent;
+    uint32_t key_len;
+    uint32_t value_len;
+} rbtree_node_fixed;
+
+// 完整节点结构（用于类型定义，实际内存大小由分配时确定）
 typedef struct _rbtree_node {
-	unsigned char color;
+    unsigned char color;
-	struct _rbtree_node *right;
+    struct _rbtree_node *right;
-	struct _rbtree_node *left;
+    struct _rbtree_node *left;
-	struct _rbtree_node *parent;
+    struct _rbtree_node *parent;
-	KEY_TYPE *key;
+    uint32_t key_len;
-	uint32_t key_len;
+    uint32_t value_len;
-	KEY_TYPE *value;
+    // 动态数据：key[key_len] + value[value_len]
-	uint32_t value_len;
+    // 不存储为结构体成员，通过指针运算访问
 } rbtree_node;
 
 typedef struct _rbtree {
@@ -191,21 +202,27 @@ int kvs_rbtree_exist(kvs_rbtree_t *inst, char *key);
 
 
 #if BIN_SAFE
-#define MAX_TABLE_SIZE 1024
 typedef struct hashnode_s {
-    uint8_t *key;
+    uint32_t key_len;      // key 长度
-    size_t   key_len;
+    uint32_t value_len;    // value 长度
-
+    struct hashnode_s *next;  // 链表指针
-    uint8_t *value;
+    // 动态数据：key[key_len] + value[value_len]
-    size_t   value_len;
+    // 不存储为结构体成员，通过指针运算访问
-
-    struct hashnode_s *next;
 } hashnode_t;
 
+typedef struct hashbucket_s {
+    hashnode_t *head;            // 桶内链表
+    uint32_t local_depth;        // 桶的局部深度
+    uint32_t item_count;         // 桶内元素数量
+    struct hashbucket_s *next_all; // 用于统一释放/遍历所有桶
+} hashbucket_t;
+
 typedef struct hashtable_s {
-    hashnode_t **nodes;
+    hashbucket_t **directory; // 目录，指向桶
-    int max_slots;
+    uint32_t dir_size;        // 目录大小（2^global_depth）
-    int count;
+    uint32_t global_depth;    // 全局深度
+    int count;                // 当前元素总数
+    hashbucket_t *bucket_list;// 所有唯一桶链表
 } hashtable_t;
 
 typedef struct hashtable_s kvs_hash_t;

memory/mempool.c

@@ -27,29 +27,29 @@ static inline int size_to_index(size_t size){
 
 
 // bitmap 操作函数
-static inline void bitmap_set(uint64_t *bitmap, uint16_t index){
+// static inline void bitmap_set(uint64_t *bitmap, uint16_t index){
-    bitmap[index / 64] |= (1ULL << (index % 64));
+//     bitmap[index / 64] |= (1ULL << (index % 64));
-}
+// }
 
-static inline void bitmap_clear(uint64_t *bitmap, uint16_t index){
+// static inline void bitmap_clear(uint64_t *bitmap, uint16_t index){
-    bitmap[index / 64] &= ~(1ULL << (index % 64));
+//     bitmap[index / 64] &= ~(1ULL << (index % 64));
-}
+// }
 
-static inline int bitmap_test(uint64_t *bitmap, uint16_t index){
+// static inline int bitmap_test(uint64_t *bitmap, uint16_t index){
-    return (bitmap[index / 64] & (1ULL << (index % 64))) != 0;
+//     return (bitmap[index / 64] & (1ULL << (index % 64))) != 0;
-}
+// }
 
-static inline void bitmap_clear_all(uint64_t *bitmap, size_t size){
+// static inline void bitmap_clear_all(uint64_t *bitmap, size_t size){
-    memset(bitmap, 0, size * sizeof(uint64_t));
+//     memset(bitmap, 0, size * sizeof(uint64_t));
-}
+// }
 
-// 根据指针计算在页中的块索引
+// // 根据指针计算在页中的块索引
-static inline uint16_t ptr_to_block_index(mp_page_t *pg, void *ptr){
+// static inline uint16_t ptr_to_block_index(mp_page_t *pg, void *ptr){
-    char *base = (char*)page_payload(pg);
+//     char *base = (char*)page_payload(pg);
-    char *p = (char*)ptr;
+//     char *p = (char*)ptr;
-    size_t offset = p - base;
+//     size_t offset = p - base;
-    return (uint16_t)(offset / pg->owner->block_size);
+//     return (uint16_t)(offset / pg->owner->block_size);
-}
+// }
 
 
 static mp_page_t* mp_page_create(mp_bucket_t *owner){
@@ -74,7 +74,7 @@ static mp_page_t* mp_page_create(mp_bucket_t *owner){
     pg->prev = NULL;
     pg->next = NULL;
 
-    bitmap_clear_all(pg->bitmap, 16);
+    // bitmap_clear_all(pg->bitmap, 16);
 
     char *p = (char*)page_payload(pg);
     for(uint16_t i = 0;i < cap - 1; ++ i){
@@ -95,8 +95,8 @@ static void *mp_page_alloc(mp_page_t *pg){
     pg->free_count --;
 
     // 标记该块为已分配
-    uint16_t index = ptr_to_block_index(pg, ret);
+    // uint16_t index = ptr_to_block_index(pg, ret);
-    bitmap_set(pg->bitmap, index);
+    // bitmap_set(pg->bitmap, index);
 
     return ret;
 }
@@ -105,14 +105,14 @@ static int mp_page_free(mp_page_t *pg, void *ptr){
     if(!pg || !ptr) return MEMPOOL_INVALID_INPUT;
 
     // 检查是否是 double free
-    uint16_t index = ptr_to_block_index(pg, ptr);
+    // uint16_t index = ptr_to_block_index(pg, ptr);
-    if(!bitmap_test(pg->bitmap, index)){
+    // if(!bitmap_test(pg->bitmap, index)){
-        // 该块未被分配，可能是 double free
+    //     // 该块未被分配，可能是 double free
-        return MEMPOOL_DOUBLE_FREE;
+    //     return MEMPOOL_DOUBLE_FREE;
-    }
+    // }
 
     // 标记该块为空闲
-    bitmap_clear(pg->bitmap, index);
+    // bitmap_clear(pg->bitmap, index);
 
     *(void**)ptr    = pg->free_list;
     pg->free_list   = ptr;
@@ -320,8 +320,10 @@ int mp_print(mp_pool_t *pool){
     printf("------\n");
     for(int i = 0; i < MEMPOOL_NUM_CLASSES; i++){
         mp_bucket_t *bucket = &pool->buckets[i];
-        if(bucket->page_count) ret += bucket->page_count; 
+        if(bucket->page_count) {
-        printf("size:%ld, page:%d, empty:%d\n", bucket->block_size, bucket->page_count, bucket->empty_count);
+            ret += bucket->page_count; 
+            printf("size:%ld, page:%d, empty:%d\n", bucket->block_size, bucket->page_count, bucket->empty_count);
+        }
     }
     printf("------\n");
 

memory/mempool.h

@@ -7,8 +7,8 @@
 #include <string.h>
 #include <pthread.h>
 
-// #define MEMPOOL_PAGE_SIZE           4096
+#define MEMPOOL_PAGE_SIZE           4096
-#define MEMPOOL_PAGE_SIZE           (4096*2)
+// #define MEMPOOL_PAGE_SIZE           (256 * 1024)
 #define MEMPOOL_BLOCK_MAX_SIZE      512
 #define MEMPOOL_ALIGNMENT           8
 #define MEMPOOL_NUM_CLASSES         (MEMPOOL_BLOCK_MAX_SIZE / MEMPOOL_ALIGNMENT)
@@ -36,7 +36,7 @@ struct mp_page_s{
     uint16_t        free_count;
     uint16_t        capacity;
 
-    uint64_t        bitmap[16];  // 最多支持 512/1280 个块 (64*20)
+    // uint64_t        bitmap[16];  
 };
 
 struct mp_bucket_s{

network/chainbuffer.c

@@ -0,0 +1,703 @@
+#include "network/chainbuffer.h"
+
+#include <assert.h>
+#include <errno.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/socket.h>
+
+#define CHAINBUFFER_DEFAULT_CHUNK 4096
+#define CHAINBUFFER_MAX_IOV 16
+#define CHAINBUFFER_DEFAULT_FREE_LIMIT 256
+
+struct chain_buffer_node {
+    struct chain_buffer_node *next;
+    size_t cap;
+    size_t rpos;
+    size_t wpos;
+    size_t used;
+    unsigned refcnt;
+    uint8_t data[];
+};
+
+static chain_buffer_node_t *alloc_node(size_t cap) {
+    chain_buffer_node_t *node;
+
+    node = (chain_buffer_node_t *)malloc(sizeof(*node) + cap);
+    if (!node) {
+        return NULL;
+    }
+
+    node->next = NULL;
+    node->cap = cap;
+    node->rpos = 0;
+    node->wpos = 0;
+    node->used = 0;
+    node->refcnt = 0;
+    return node;
+}
+
+static void reset_node_state(chain_buffer_node_t *node) {
+    if (!node) {
+        return;
+    }
+    node->next = NULL;
+    node->rpos = 0;
+    node->wpos = 0;
+    node->used = 0;
+    node->refcnt = 0;
+}
+
+static size_t min_size(size_t a, size_t b) {
+    return a < b ? a : b;
+}
+
+static size_t node_read_seg1_len(const chain_buffer_node_t *node) {
+    size_t tail_len;
+
+    if (!node || node->used == 0) {
+        return 0;
+    }
+
+    tail_len = node->cap - node->rpos;
+    return min_size(node->used, tail_len);
+}
+
+static size_t node_write_seg1_len(const chain_buffer_node_t *node) {
+    size_t free_total;
+    size_t tail_len;
+
+    if (!node || node->used >= node->cap) {
+        return 0;
+    }
+
+    free_total = node->cap - node->used;
+    tail_len = node->cap - node->wpos;
+    return min_size(free_total, tail_len);
+}
+
+static void node_read_segments(const chain_buffer_node_t *node,
+                               uint8_t **p1, size_t *l1,
+                               uint8_t **p2, size_t *l2) {
+    size_t seg1;
+    size_t seg2;
+
+    seg1 = node_read_seg1_len(node);
+    seg2 = (node && node->used > seg1) ? (node->used - seg1) : 0;
+
+    if (p1) {
+        *p1 = (seg1 > 0) ? (uint8_t *)(node->data + node->rpos) : NULL;
+    }
+    if (l1) {
+        *l1 = seg1;
+    }
+    if (p2) {
+        *p2 = (seg2 > 0) ? (uint8_t *)node->data : NULL;
+    }
+    if (l2) {
+        *l2 = seg2;
+    }
+}
+
+static void node_write_segments(const chain_buffer_node_t *node,
+                                uint8_t **p1, size_t *l1,
+                                uint8_t **p2, size_t *l2) {
+    size_t free_total;
+    size_t seg1;
+    size_t seg2;
+
+    if (!node || node->used >= node->cap) {
+        if (p1) *p1 = NULL;
+        if (l1) *l1 = 0;
+        if (p2) *p2 = NULL;
+        if (l2) *l2 = 0;
+        return;
+    }
+
+    free_total = node->cap - node->used;
+    seg1 = node_write_seg1_len(node);
+    seg2 = free_total - seg1;
+
+    if (p1) {
+        *p1 = (seg1 > 0) ? (uint8_t *)(node->data + node->wpos) : NULL;
+    }
+    if (l1) {
+        *l1 = seg1;
+    }
+    if (p2) {
+        *p2 = (seg2 > 0) ? (uint8_t *)node->data : NULL;
+    }
+    if (l2) {
+        *l2 = seg2;
+    }
+}
+
+static void node_advance_read(chain_buffer_node_t *node, size_t n) {
+    if (!node || n == 0) {
+        return;
+    }
+    node->rpos = (node->rpos + n) % node->cap;
+    node->used -= n;
+}
+
+static void node_advance_write(chain_buffer_node_t *node, size_t n) {
+    if (!node || n == 0) {
+        return;
+    }
+    node->wpos = (node->wpos + n) % node->cap;
+    node->used += n;
+}
+
+static chain_buffer_node_t *acquire_node(chain_buffer_t *buf, size_t cap) {
+    chain_buffer_node_t *node = NULL;
+
+    if (!buf || cap == 0) {
+        return NULL;
+    }
+
+    if (cap == buf->chunk_size && buf->free_list) {
+        node = buf->free_list;
+        buf->free_list = node->next;
+        buf->free_count--;
+        reset_node_state(node);
+        return node;
+    }
+
+    node = alloc_node(cap);
+    return node;
+}
+
+static void recycle_node(chain_buffer_t *buf, chain_buffer_node_t *node) {
+    if (!node) {
+        return;
+    }
+
+    if (!buf || node->cap != buf->chunk_size || buf->free_count >= buf->free_limit) {
+        free(node);
+        return;
+    }
+
+    reset_node_state(node);
+    node->next = buf->free_list;
+    buf->free_list = node;
+    buf->free_count++;
+}
+
+static int append_new_tail(chain_buffer_t *buf, size_t cap) {
+    chain_buffer_node_t *node;
+
+    node = acquire_node(buf, cap);
+    if (!node) {
+        return -1;
+    }
+
+    if (!buf->tail) {
+        buf->head = node;
+        buf->tail = node;
+    } else {
+        buf->tail->next = node;
+        buf->tail = node;
+    }
+    return 0;
+}
+
+static void list_append_node(chain_buffer_list_t *list, chain_buffer_node_t *node) {
+    if (!list || !node) {
+        return;
+    }
+    node->next = NULL;
+    if (!list->tail) {
+        list->head = node;
+        list->tail = node;
+    } else {
+        list->tail->next = node;
+        list->tail = node;
+    }
+}
+
+void chain_buffer_init(chain_buffer_t *buf, size_t chunk_size) {
+    if (!buf) {
+        return;
+    }
+
+    memset(buf, 0, sizeof(*buf));
+    buf->chunk_size = chunk_size ? chunk_size : CHAINBUFFER_DEFAULT_CHUNK;
+    buf->free_limit = CHAINBUFFER_DEFAULT_FREE_LIMIT;
+}
+
+void chain_buffer_reset(chain_buffer_t *buf) {
+    chain_buffer_node_t *node;
+
+    if (!buf) {
+        return;
+    }
+
+    node = buf->head;
+    while (node) {
+        chain_buffer_node_t *next = node->next;
+        free(node);
+        node = next;
+    }
+
+    node = buf->free_list;
+    while (node) {
+        chain_buffer_node_t *next = node->next;
+        free(node);
+        node = next;
+    }
+
+    free(buf->linear_cache);
+    memset(buf, 0, sizeof(*buf));
+}
+
+size_t chain_buffer_len(const chain_buffer_t *buf) {
+    return buf ? buf->total_len : 0;
+}
+
+int chain_buffer_append(chain_buffer_t *buf, const void *data, size_t len) {
+    const uint8_t *src;
+    size_t remain;
+
+    if (!buf || (!data && len > 0)) {
+        errno = EINVAL;
+        return -1;
+    }
+    if (len == 0) {
+        return 0;
+    }
+    if (buf->total_len > (size_t)-1 - len) {
+        errno = EOVERFLOW;
+        return -1;
+    }
+
+    src = (const uint8_t *)data;
+    remain = len;
+
+    while (remain > 0) {
+        chain_buffer_node_t *tail;
+        uint8_t *p1;
+        uint8_t *p2;
+        size_t l1;
+        size_t l2;
+        size_t writable;
+        size_t n;
+        size_t c1;
+
+        if (!buf->tail || buf->tail->used == buf->tail->cap) {
+            size_t cap = remain > buf->chunk_size ? remain : buf->chunk_size;
+            if (append_new_tail(buf, cap) < 0) {
+                errno = ENOMEM;
+                return -1;
+            }
+        }
+
+        tail = buf->tail;
+        node_write_segments(tail, &p1, &l1, &p2, &l2);
+        writable = l1 + l2;
+        if (writable == 0) {
+            continue;
+        }
+
+        n = min_size(remain, writable);
+        c1 = min_size(n, l1);
+        if (c1 > 0) {
+            memcpy(p1, src, c1);
+        }
+        if (n > c1) {
+            memcpy(p2, src + c1, n - c1);
+        }
+
+        node_advance_write(tail, n);
+        buf->total_len += n;
+        src += n;
+        remain -= n;
+    }
+
+    return 0;
+}
+
+size_t chain_buffer_drain(chain_buffer_t *buf, size_t len) {
+    size_t remain;
+    size_t drained;
+
+    if (!buf || len == 0 || buf->total_len == 0) {
+        return 0;
+    }
+
+    remain = len;
+    drained = 0;
+
+    while (remain > 0 && buf->head) {
+        chain_buffer_node_t *node = buf->head;
+        size_t n = min_size(remain, node->used);
+
+        if (n == 0) {
+            buf->head = node->next;
+            if (!buf->head) {
+                buf->tail = NULL;
+            }
+            recycle_node(buf, node);
+            continue;
+        }
+
+        node_advance_read(node, n);
+        buf->total_len -= n;
+        drained += n;
+        remain -= n;
+
+        if (node->used == 0) {
+            assert(node->refcnt == 0);
+            buf->head = node->next;
+            if (!buf->head) {
+                buf->tail = NULL;
+            }
+            recycle_node(buf, node);
+        }
+    }
+
+    return drained;
+}
+
+const uint8_t *chain_buffer_linearize(chain_buffer_t *buf, size_t *out_len) {
+    size_t offset;
+
+    if (!buf) {
+        return NULL;
+    }
+
+    if (out_len) {
+        *out_len = buf->total_len;
+    }
+
+    if (buf->total_len == 0) {
+        return NULL;
+    }
+
+    if (buf->head == buf->tail && buf->head) {
+        chain_buffer_node_t *node = buf->head;
+        if (node->used == 0) {
+            return NULL;
+        }
+        if (node_read_seg1_len(node) == node->used) {
+            return node->data + node->rpos;
+        }
+    }
+
+    if (buf->linear_cap < buf->total_len) {
+        uint8_t *new_cache = (uint8_t *)realloc(buf->linear_cache, buf->total_len);
+        if (!new_cache) {
+            return NULL;
+        }
+        buf->linear_cache = new_cache;
+        buf->linear_cap = buf->total_len;
+    }
+
+    offset = 0;
+    for (chain_buffer_node_t *node = buf->head; node; node = node->next) {
+        uint8_t *p1;
+        uint8_t *p2;
+        size_t l1;
+        size_t l2;
+
+        node_read_segments(node, &p1, &l1, &p2, &l2);
+        if (l1 > 0) {
+            memcpy(buf->linear_cache + offset, p1, l1);
+            offset += l1;
+        }
+        if (l2 > 0) {
+            memcpy(buf->linear_cache + offset, p2, l2);
+            offset += l2;
+        }
+    }
+
+    return buf->linear_cache;
+}
+
+ssize_t chain_buffer_send_fd(chain_buffer_t *buf, int fd, int flags) {
+    struct iovec iov[CHAINBUFFER_MAX_IOV];
+    struct msghdr msg;
+    size_t iovcnt;
+    ssize_t n;
+
+    if (!buf) {
+        errno = EINVAL;
+        return -1;
+    }
+    if (buf->total_len == 0 || !buf->head) {
+        return 0;
+    }
+
+    iovcnt = 0;
+    for (chain_buffer_node_t *node = buf->head;
+         node && iovcnt < CHAINBUFFER_MAX_IOV;
+         node = node->next) {
+        uint8_t *p1;
+        uint8_t *p2;
+        size_t l1;
+        size_t l2;
+
+        node_read_segments(node, &p1, &l1, &p2, &l2);
+        if (l1 > 0) {
+            iov[iovcnt].iov_base = p1;
+            iov[iovcnt].iov_len = l1;
+            iovcnt++;
+            if (iovcnt >= CHAINBUFFER_MAX_IOV) {
+                break;
+            }
+        }
+        if (l2 > 0) {
+            iov[iovcnt].iov_base = p2;
+            iov[iovcnt].iov_len = l2;
+            iovcnt++;
+        }
+    }
+
+    if (iovcnt == 0) {
+        return 0;
+    }
+
+    memset(&msg, 0, sizeof(msg));
+    msg.msg_iov = iov;
+    msg.msg_iovlen = iovcnt;
+
+    n = sendmsg(fd, &msg, flags);
+    if (n > 0) {
+        chain_buffer_drain(buf, (size_t)n);
+    }
+
+    return n;
+}
+
+int chain_buffer_prepare_recv_iov(chain_buffer_t *buf, struct iovec *iov, int max_iov) {
+    chain_buffer_node_t *tail;
+    uint8_t *p1;
+    uint8_t *p2;
+    size_t l1;
+    size_t l2;
+    int iovcnt;
+
+    if (!buf || !iov || max_iov <= 0) {
+        errno = EINVAL;
+        return -1;
+    }
+
+    if (!buf->tail || buf->tail->used == buf->tail->cap) {
+        if (append_new_tail(buf, buf->chunk_size) < 0) {
+            errno = ENOMEM;
+            return -1;
+        }
+    }
+
+    tail = buf->tail;
+    node_write_segments(tail, &p1, &l1, &p2, &l2);
+
+    iovcnt = 0;
+    if (l1 > 0) {
+        iov[iovcnt].iov_base = p1;
+        iov[iovcnt].iov_len = l1;
+        iovcnt++;
+    }
+    if (l2 > 0 && iovcnt < max_iov) {
+        iov[iovcnt].iov_base = p2;
+        iov[iovcnt].iov_len = l2;
+        iovcnt++;
+    }
+
+    return iovcnt;
+}
+
+size_t chain_buffer_commit_recv(chain_buffer_t *buf, size_t len) {
+    chain_buffer_node_t *tail;
+    size_t free_total;
+
+    if (!buf || len == 0) {
+        return 0;
+    }
+
+    if (!buf->tail || buf->tail->used == buf->tail->cap) {
+        return 0;
+    }
+
+    tail = buf->tail;
+    free_total = tail->cap - tail->used;
+    if (len > free_total) {
+        len = free_total;
+    }
+
+    node_advance_write(tail, len);
+    buf->total_len += len;
+    return len;
+}
+
+void chain_buffer_list_init(chain_buffer_list_t *list) {
+    if (!list) {
+        return;
+    }
+    memset(list, 0, sizeof(*list));
+}
+
+size_t chain_buffer_list_len(const chain_buffer_list_t *list) {
+    return list ? list->total_len : 0;
+}
+
+int chain_buffer_list_iov(const chain_buffer_list_t *list, struct iovec *iov, int max_iov) {
+    int iovcnt = 0;
+
+    if (!list || !iov || max_iov <= 0) {
+        errno = EINVAL;
+        return -1;
+    }
+
+    for (chain_buffer_node_t *node = list->head; node && iovcnt < max_iov; node = node->next) {
+        uint8_t *p1;
+        uint8_t *p2;
+        size_t l1;
+        size_t l2;
+
+        node_read_segments(node, &p1, &l1, &p2, &l2);
+        if (l1 > 0) {
+            iov[iovcnt].iov_base = p1;
+            iov[iovcnt].iov_len = l1;
+            iovcnt++;
+            if (iovcnt >= max_iov) {
+                break;
+            }
+        }
+        if (l2 > 0) {
+            iov[iovcnt].iov_base = p2;
+            iov[iovcnt].iov_len = l2;
+            iovcnt++;
+        }
+    }
+
+    return iovcnt;
+}
+
+static int copy_prefix_from_node(chain_buffer_node_t *node, size_t len, uint8_t *dst) {
+    uint8_t *p1;
+    uint8_t *p2;
+    size_t l1;
+    size_t l2;
+    size_t c1;
+
+    if (!node || !dst || len == 0 || len > node->used) {
+        return -1;
+    }
+
+    node_read_segments(node, &p1, &l1, &p2, &l2);
+    c1 = min_size(len, l1);
+    if (c1 > 0) {
+        memcpy(dst, p1, c1);
+    }
+    if (len > c1) {
+        memcpy(dst + c1, p2, len - c1);
+    }
+    return 0;
+}
+
+int chain_buffer_detach_prefix(chain_buffer_t *buf, size_t len, chain_buffer_list_t *out) {
+    size_t remain;
+
+    if (!buf || !out) {
+        errno = EINVAL;
+        return -1;
+    }
+
+    chain_buffer_list_init(out);
+
+    if (len == 0) {
+        return 0;
+    }
+    if (len > buf->total_len) {
+        errno = EINVAL;
+        return -1;
+    }
+
+    remain = len;
+    while (remain > 0 && buf->head) {
+        chain_buffer_node_t *node = buf->head;
+
+        if (remain >= node->used) {
+            size_t take = node->used;
+
+            buf->head = node->next;
+            if (!buf->head) {
+                buf->tail = NULL;
+            }
+            node->next = NULL;
+            node->refcnt = 1;
+            list_append_node(out, node);
+            out->total_len += take;
+            buf->total_len -= take;
+            remain -= take;
+            continue;
+        }
+
+        {
+            chain_buffer_node_t *part = acquire_node(buf, remain);
+            if (!part) {
+                chain_buffer_list_release(buf, out);
+                errno = ENOMEM;
+                return -1;
+            }
+
+            if (copy_prefix_from_node(node, remain, part->data) < 0) {
+                recycle_node(buf, part);
+                chain_buffer_list_release(buf, out);
+                errno = EINVAL;
+                return -1;
+            }
+
+            part->used = remain;
+            part->wpos = remain % part->cap;
+            part->rpos = 0;
+            part->refcnt = 1;
+            part->next = NULL;
+            list_append_node(out, part);
+            out->total_len += remain;
+
+            node_advance_read(node, remain);
+            buf->total_len -= remain;
+            remain = 0;
+
+            if (node->used == 0) {
+                buf->head = node->next;
+                if (!buf->head) {
+                    buf->tail = NULL;
+                }
+                recycle_node(buf, node);
+            }
+        }
+    }
+
+    return 0;
+}
+
+void chain_buffer_list_release(chain_buffer_t *owner, chain_buffer_list_t *list) {
+    chain_buffer_node_t *node;
+
+    if (!list) {
+        return;
+    }
+
+    node = list->head;
+    while (node) {
+        chain_buffer_node_t *next = node->next;
+
+        if (node->refcnt > 0) {
+            node->refcnt--;
+        }
+
+        if (node->refcnt == 0) {
+            if (owner) {
+                recycle_node(owner, node);
+            } else {
+                free(node);
+            }
+        }
+
+        node = next;
+    }
+
+    chain_buffer_list_init(list);
+}

network/chainbuffer.h

@@ -0,0 +1,52 @@
+#ifndef __CHAINBUFFER_H__
+#define __CHAINBUFFER_H__
+
+#include <stddef.h>
+#include <stdint.h>
+#include <sys/types.h>
+#include <sys/uio.h>
+
+typedef struct chain_buffer_node chain_buffer_node_t;
+
+typedef struct chain_buffer_s {
+    chain_buffer_node_t *head;
+    chain_buffer_node_t *tail;
+    size_t total_len;
+    size_t chunk_size;
+
+    chain_buffer_node_t *free_list;
+    size_t free_count;
+    size_t free_limit;
+
+    uint8_t *linear_cache;
+    size_t linear_cap;
+} chain_buffer_t;
+
+typedef struct chain_buffer_list_s {
+    chain_buffer_node_t *head;
+    chain_buffer_node_t *tail;
+    size_t total_len;
+} chain_buffer_list_t;
+
+void chain_buffer_init(chain_buffer_t *buf, size_t chunk_size);
+void chain_buffer_reset(chain_buffer_t *buf);
+
+size_t chain_buffer_len(const chain_buffer_t *buf);
+int chain_buffer_append(chain_buffer_t *buf, const void *data, size_t len);
+size_t chain_buffer_drain(chain_buffer_t *buf, size_t len);
+
+const uint8_t *chain_buffer_linearize(chain_buffer_t *buf, size_t *out_len);
+ssize_t chain_buffer_send_fd(chain_buffer_t *buf, int fd, int flags);
+
+/* readv zero-copy recv helpers */
+int chain_buffer_prepare_recv_iov(chain_buffer_t *buf, struct iovec *iov, int max_iov);
+size_t chain_buffer_commit_recv(chain_buffer_t *buf, size_t len);
+
+/* ownership transfer helpers */
+void chain_buffer_list_init(chain_buffer_list_t *list);
+size_t chain_buffer_list_len(const chain_buffer_list_t *list);
+int chain_buffer_list_iov(const chain_buffer_list_t *list, struct iovec *iov, int max_iov);
+int chain_buffer_detach_prefix(chain_buffer_t *buf, size_t len, chain_buffer_list_t *out);
+void chain_buffer_list_release(chain_buffer_t *owner, chain_buffer_list_t *list);
+
+#endif

reactor.c

@@ -1,461 +1,625 @@
-
-
 #define _GNU_SOURCE
-#include <errno.h>
-#include <stdio.h>
-#include <sys/socket.h>
-#include <netinet/in.h>
-#include <string.h>
-#include <pthread.h>
-#include <unistd.h>
-#include <poll.h>
-#include <sys/epoll.h>
-#include <errno.h>
-#include <sys/time.h>
-#include <sys/eventfd.h>
-#include <sys/timerfd.h>
 
+#include <errno.h>
+#include <fcntl.h>
+#include <netinet/in.h>
+#include <pthread.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <string.h>
+#include <sys/epoll.h>
+#include <sys/eventfd.h>
+#include <sys/socket.h>
+#include <sys/time.h>
+#include <sys/timerfd.h>
+#include <sys/uio.h>
+#include <unistd.h>
+
+#include "diskuring/diskuring.h"
 #include "server.h"
 
-
+#define CONNECTION_SIZE 65536
-#define CONNECTION_SIZE			1024 // 1024 * 1024
+#define MAX_PORTS 20
-
+#define RECV_BATCH_BYTES 4096
-#define MAX_PORTS			20
+#define CHAIN_BUFFER_CHUNK 4096
-
+#define MAX_CONN_READ_BYTES (32u * 1024u * 1024u)
-#define TIME_SUB_MS(tv1, tv2)  ((tv1.tv_sec - tv2.tv_sec) * 1000 + (tv1.tv_usec - tv2.tv_usec) / 1000)
+#define MAX_CONN_WRITE_BYTES (32u * 1024u * 1024u)
-
-
 
 #if ENABLE_KVSTORE
-
+typedef int (*msg_handler)(struct conn *conn);
-// typedef int (*msg_handler)(char *msg, int length, char *response);
-// typedef int (*msg_handler)(char *request, int request_length, char *response, int *response_length);
-typedef int (*msg_handler)(struct conn* conn);
-
 static msg_handler kvs_handler;
 
-extern void cleanup_finished_iouring_tasks();
+extern iouring_ctx_t global_uring_ctx;
 
-
-// 0 need more, -1 error, =1 suc
 int kvs_request(struct conn *c) {
-	// int consumed_out = kvs_handler(c->rbuffer, c->rlength, c->wbuffer, &c->wlength);
+    return kvs_handler ? kvs_handler(c) : -1;
-	int consumed_out = kvs_handler(c);
-	return consumed_out; 
 }
 
 int kvs_response(struct conn *c) {
-
+    (void)c;
-	
+    return 0;
-
 }
-
-
 #endif
 
-
-
 int accept_cb(int fd);
 int recv_cb(int fd);
 int send_cb(int fd);
 
+static int epfd = -1;
+static int wakeup_fd = -1;
+static int timer_fd = -1;
+static struct timeval begin;
+static struct conn conn_list[CONNECTION_SIZE];
 
-
+static int conn_fd_valid(int fd) {
-int epfd = 0;
+    return fd >= 0 && fd < CONNECTION_SIZE;
-struct timeval begin;
-
-int wakeup_fd = -1;
-int timer_fd = -1;
-
-struct conn conn_list[CONNECTION_SIZE] = {0};
-// fd
-
-
-// 1 add, 0 mod
-int set_event(int fd, int event, int flag) {
-
-	if (flag) {  // non-zero add
-
-		struct epoll_event ev;
-		ev.events = event;
-		ev.data.fd = fd;
-		epoll_ctl(epfd, EPOLL_CTL_ADD, fd, &ev);
-
-	} else {  // zero mod
-
-		struct epoll_event ev;
-		ev.events = event;
-		ev.data.fd = fd;
-		epoll_ctl(epfd, EPOLL_CTL_MOD, fd, &ev);
-		
-	}
-	
-
 }
 
-
+static int set_nonblocking(int fd) {
-int event_register(int fd, int event) {
+    int flags = fcntl(fd, F_GETFL, 0);
-
+    if (flags < 0) {
-	if (fd < 0) return -1;
+        return -1;
-
+    }
-	conn_list[fd].fd = fd;
+    if ((flags & O_NONBLOCK) != 0) {
-	conn_list[fd].r_action.recv_callback = recv_cb;
-	conn_list[fd].send_callback = send_cb;
-
-	memset(conn_list[fd].rbuffer, 0, BUFFER_LENGTH);
-	conn_list[fd].rlength = 0;
-
-	memset(conn_list[fd].wbuffer, 0, BUFFER_LENGTH*2);
-	conn_list[fd].wlength = 0;
-
-	conn_list[fd].is_stop = 0;
-	
-	set_event(fd, event, 1);
-}
-
-
-// listenfd(sockfd) --> EPOLLIN --> accept_cb
-int accept_cb(int fd) {
-
-	struct sockaddr_in  clientaddr;
-	socklen_t len = sizeof(clientaddr);
-
-	int clientfd = accept(fd, (struct sockaddr*)&clientaddr, &len);
-	if (clientfd < 0) {
-		printf("accept errno: %d --> %s\n", errno, strerror(errno));
-		return -1;
-	}
-	
-	event_register(clientfd, EPOLLIN);  // | EPOLLET
-
-	if ((clientfd % 1000) == 0) {
-
-		struct timeval current;
-		gettimeofday(&current, NULL);
-
-		int time_used = TIME_SUB_MS(current, begin);
-		memcpy(&begin, &current, sizeof(struct timeval));
-		
-
-		//printf("accept finshed: %d, time_used: %d\n", clientfd, time_used);
-
-	}
-
-	return 0;
-}
-
-
-int recv_cb(int fd) {
-	struct conn *c = &conn_list[fd];
-	int avail = BUFFER_LENGTH - c->rlength;
-	// printf("avail: %d\n", avail);
-	if (avail <= 0) {
-        // 缓冲满了还没解析出来：协议异常或包过大
-        epoll_ctl(epfd, EPOLL_CTL_DEL, fd, NULL);
-        close(fd);
         return 0;
     }
-
+    return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
-	int count = recv(fd, c->rbuffer + c->rlength, avail, 0);
-	if (count == 0) { // disconnect
-		//printf("client disconnect: %d\n", fd);
-
-		epoll_ctl(epfd, EPOLL_CTL_DEL, fd, NULL); // unfinished
-		close(fd);
-
-		return 0;
-	} else if (count < 0) { // 
-
-		printf("count: %d, errno: %d, %s\n", count, errno, strerror(errno));
-		epoll_ctl(epfd, EPOLL_CTL_DEL, fd, NULL);
-		close(fd);
-
-		return 0;
-	}
-
-	
-	c->rlength += count;
-	//printf("RECV: %s\n", conn_list[fd].rbuffer);
-
-#if 0 // echo
-
-	conn_list[fd].wlength = conn_list[fd].rlength;
-	memcpy(conn_list[fd].wbuffer, conn_list[fd].rbuffer, conn_list[fd].wlength);
-
-	printf("[%d]RECV: %s\n", conn_list[fd].rlength, conn_list[fd].rbuffer);
-
-#elif ENABLE_HTTP
-
-	http_request(&conn_list[fd]);
-
-#elif ENABLE_WEBSOCKET
-
-	ws_request(&conn_list[fd]);
-
-#elif ENABLE_KVSTORE
-	int consumed = kvs_request(c);
-	if(consumed < 0){
-		epoll_ctl(epfd, EPOLL_CTL_DEL, fd, NULL);
-		close(fd);
-		return 0;
-	}
-
-	// 清理 buffer 
-	if (consumed > 0 && consumed < c->rlength) {
-		// 有剩余未处理数据，搬移到 buffer 头部
-		int left = c->rlength - consumed;
-		if (left > 0) memmove(c->rbuffer, c->rbuffer + consumed, left);
-		c->rlength = left;
-		if (c->wlength > 0) set_event(fd, EPOLLOUT, 0);
-		return count;
-
-	}else{
-		c->rlength = 0;
-		if(c->wlength > 0) set_event(fd, EPOLLOUT, 0);
-		return count;
-	}
-
-#endif
-
-
-	set_event(fd, EPOLLOUT, 0);
-
-	return count;
 }
 
+static void conn_clear_slot(int fd) {
+    struct conn *c;
 
-int send_cb(int fd) {
+    if (!conn_fd_valid(fd)) {
+        return;
+    }
 
-#if ENABLE_HTTP
+    c = &conn_list[fd];
+    chain_buffer_reset(&c->rbuf);
+    chain_buffer_reset(&c->wbuf);
+    memset(c, 0, sizeof(*c));
+    c->fd = -1;
+}
 
-	http_response(&conn_list[fd]);
+static void close_conn(int fd) {
+    if (!conn_fd_valid(fd)) {
+        return;
+    }
+    epoll_ctl(epfd, EPOLL_CTL_DEL, fd, NULL);
+    close(fd);
+    conn_clear_slot(fd);
+}
 
-#elif ENABLE_WEBSOCKET
+static int set_event(int fd, int event, int is_add) {
+    struct epoll_event ev;
+    int op = is_add ? EPOLL_CTL_ADD : EPOLL_CTL_MOD;
 
-	ws_response(&conn_list[fd]);
+    memset(&ev, 0, sizeof(ev));
+    ev.events = (uint32_t)event;
+    ev.data.fd = fd;
 
-#elif ENABLE_KVSTORE
+    if (epoll_ctl(epfd, op, fd, &ev) < 0) {
+        return -1;
+    }
+    return 0;
+}
 
-	kvs_response(&conn_list[fd]);
+static int update_conn_events(int fd) {
+    struct conn *c;
+    int events = EPOLLIN;
 
-#endif
+    if (!conn_fd_valid(fd)) {
+        return -1;
+    }
 
-	struct conn *c = &conn_list[fd];
+    c = &conn_list[fd];
-    int sent_total = 0;
+    if (chain_buffer_len(&c->wbuf) > 0) {
+        events |= EPOLLOUT;
+    }
 
-	pthread_mutex_lock(&c->g_sync_lock);
+    return set_event(fd, events, 0);
+}
 
-	while (c->wlength > 0) {
+int event_register(int fd, int event) {
-		ssize_t n = send(fd, c->wbuffer, (size_t)c->wlength, MSG_NOSIGNAL);
+    struct conn *c;
-		if (n > 0) {
-            sent_total += (int)n;
 
-			if (n == c->wlength) {
+    if (!conn_fd_valid(fd)) {
-                /* 全部发完 */
+        if (fd >= 0) {
-                c->wlength = 0;
+            close(fd);
-                break;
+        }
+        return -1;
+    }
+
+    if (set_nonblocking(fd) < 0) {
+        close(fd);
+        return -1;
+    }
+
+    conn_clear_slot(fd);
+
+    c = &conn_list[fd];
+    c->fd = fd;
+    c->r_action.recv_callback = recv_cb;
+    c->send_callback = send_cb;
+    c->is_stop = 0;
+    chain_buffer_init(&c->rbuf, CHAIN_BUFFER_CHUNK);
+    chain_buffer_init(&c->wbuf, CHAIN_BUFFER_CHUNK);
+
+    if (set_event(fd, event, 1) < 0) {
+        close_conn(fd);
+        return -1;
+    }
+    return 0;
+}
+
+int accept_cb(int fd) {
+    while (1) {
+        struct sockaddr_in clientaddr;
+        socklen_t len = sizeof(clientaddr);
+        int clientfd = accept4(fd, (struct sockaddr *)&clientaddr, &len, SOCK_NONBLOCK | SOCK_CLOEXEC);
+
+        if (clientfd < 0) {
+            if (errno == EAGAIN || errno == EWOULDBLOCK) {
+                return 0;
             }
-			
+            if (errno == EINTR) {
-			/* 只发了一部分：把剩余数据搬到 buffer 头部 */
+                continue;
-            int left = c->wlength - (int)n;
+            }
-            memmove(c->wbuffer, c->wbuffer + n, (size_t)left);
+            printf("accept errno: %d --> %s\n", errno, strerror(errno));
-            c->wlength = left;
+            return -1;
-
-            /* 不要在这里死循环占用 CPU，交给下一次 EPOLLOUT */
-            break;
         }
 
-		if (n < 0) {
+        if (!conn_fd_valid(clientfd)) {
-            if (errno == EAGAIN || errno == EWOULDBLOCK) {
+            printf("drop client fd=%d, out of conn_list range\n", clientfd);
-                /* 暂时发不出去，等下一次可写事件 */
+            close(clientfd);
-                pthread_mutex_unlock(&c->g_sync_lock);
+            continue;
-                set_event(fd, EPOLLOUT, 0);
+        }
-                return sent_total;
-            }
 
-            /* 对端断开 / 其他错误 */
+        if (event_register(clientfd, EPOLLIN) < 0) {
-            int e = errno;
+            continue;
-            pthread_mutex_unlock(&c->g_sync_lock);
+        }
 
-            printf("send fd=%d errno=%d %s\n", fd, e, strerror(e));
+        if ((clientfd % 1000) == 0) {
-            epoll_ctl(epfd, EPOLL_CTL_DEL, fd, NULL);
+            struct timeval current;
-            close(fd);
+            int time_used;
+
+            gettimeofday(&current, NULL);
+            time_used = (int)((current.tv_sec - begin.tv_sec) * 1000 +
+                              (current.tv_usec - begin.tv_usec) / 1000);
+            begin = current;
+            (void)time_used;
+        }
+    }
+}
+
+int recv_cb(int fd) {
+    struct conn *c;
+    int total = 0;
+
+    if (!conn_fd_valid(fd)) {
+        return -1;
+    }
+
+    c = &conn_list[fd];
+    while (1) {
+        struct iovec iov[4];
+        int iovcnt = chain_buffer_prepare_recv_iov(&c->rbuf, iov, (int)(sizeof(iov) / sizeof(iov[0])));
+        ssize_t n;
+
+        if (iovcnt < 0) {
+            printf("fd=%d prepare recv iov failed: %s\n", fd, strerror(errno));
+            close_conn(fd);
             return 0;
         }
 
-        break;
+        n = readv(fd, iov, iovcnt);
-	}
-	pthread_mutex_unlock(&c->g_sync_lock);
 
-    if (c->wlength > 0) {
+        if (n > 0) {
-        /* 还有没发完，继续监听可写 */
+            size_t cur_len = chain_buffer_len(&c->rbuf);
-        set_event(fd, EPOLLOUT, 0);
+            if (cur_len > MAX_CONN_READ_BYTES - (size_t)n) {
-    } else {
+                printf("fd=%d read buffer overflow, close connection\n", fd);
-        /* 发完了，回到读 */
+                close_conn(fd);
-        set_event(fd, EPOLLIN, 0);
+                return 0;
+            }
+            if (chain_buffer_commit_recv(&c->rbuf, (size_t)n) != (size_t)n) {
+                printf("fd=%d commit recv buffer failed\n", fd);
+                close_conn(fd);
+                return 0;
+            }
+            total += (int)n;
+            continue;
+        }
+
+        if (n == 0) {
+            close_conn(fd);
+            return 0;
+        }
+
+        if (errno == EINTR) {
+            continue;
+        }
+        if (errno == EAGAIN || errno == EWOULDBLOCK) {
+            break;
+        }
+
+        printf("recv fd=%d errno=%d, %s\n", fd, errno, strerror(errno));
+        close_conn(fd);
+        return 0;
+    }
+
+    if (total <= 0) {
+        return 0;
+    }
+
+#if ENABLE_HTTP
+    http_request(c);
+#elif ENABLE_WEBSOCKET
+    ws_request(c);
+#elif ENABLE_KVSTORE
+    {
+        int consumed = kvs_request(c);
+        size_t readable = chain_buffer_len(&c->rbuf);
+
+        if (consumed < 0) {
+            close_conn(fd);
+            return 0;
+        }
+
+        if ((size_t)consumed > readable) {
+            printf("fd=%d invalid consumed=%d readable=%zu\n", fd, consumed, readable);
+            close_conn(fd);
+            return 0;
+        }
+
+        if (consumed > 0) {
+            chain_buffer_drain(&c->rbuf, (size_t)consumed);
+        }
+
+        if (chain_buffer_len(&c->wbuf) > MAX_CONN_WRITE_BYTES) {
+            printf("fd=%d write buffer overflow, close connection\n", fd);
+            close_conn(fd);
+            return 0;
+        }
+    }
+#endif
+
+    if (update_conn_events(fd) < 0) {
+        close_conn(fd);
+        return 0;
+    }
+
+    return total;
+}
+
+int send_cb(int fd) {
+    struct conn *c;
+    int sent_total = 0;
+
+    if (!conn_fd_valid(fd)) {
+        return -1;
+    }
+    c = &conn_list[fd];
+
+#if ENABLE_HTTP
+    http_response(c);
+#elif ENABLE_WEBSOCKET
+    ws_response(c);
+#elif ENABLE_KVSTORE
+    kvs_response(c);
+#endif
+
+    while (chain_buffer_len(&c->wbuf) > 0) {
+        ssize_t n = chain_buffer_send_fd(&c->wbuf, fd, MSG_NOSIGNAL);
+        if (n > 0) {
+            sent_total += (int)n;
+            continue;
+        }
+
+        if (n == 0) {
+            break;
+        }
+
+        if (errno == EINTR) {
+            continue;
+        }
+        if (errno == EAGAIN || errno == EWOULDBLOCK) {
+            break;
+        }
+
+        printf("send fd=%d errno=%d %s\n", fd, errno, strerror(errno));
+        close_conn(fd);
+        return 0;
+    }
+
+    if (update_conn_events(fd) < 0) {
+        close_conn(fd);
+        return 0;
     }
 
     return sent_total;
 }
 
-// wakup fd
-
 int handle_wakeup_fd_cb(int fd) {
     uint64_t v;
+
     while (1) {
-        ssize_t n = read(wakeup_fd, &v, sizeof(v));
+        ssize_t n = read(fd, &v, sizeof(v));
-        if (n == sizeof(v)) continue;
+        if (n == (ssize_t)sizeof(v)) {
-        if (n < 0 && errno == EAGAIN) break;   // 已经读空
+            continue;
+        }
+        if (n < 0 && errno == EINTR) {
+            continue;
+        }
+        if (n < 0 && errno == EAGAIN) {
+            break;
+        }
         break;
     }
-	cleanup_finished_iouring_tasks();
 
-	return 0;
+    cleanup_finished_iouring_tasks(&global_uring_ctx);
+    return 0;
 }
 
 int init_wakeup_fd(void) {
     int wfd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
+    struct conn *c;
+
     if (wfd < 0) {
         printf("eventfd failed: errno=%d %s\n", errno, strerror(errno));
         return -1;
     }
+    if (!conn_fd_valid(wfd)) {
+        close(wfd);
+        return -1;
+    }
 
-	conn_list[wfd].fd = wfd;
+    conn_clear_slot(wfd);
-	conn_list[wfd].r_action.recv_callback = handle_wakeup_fd_cb;
+    c = &conn_list[wfd];
-	set_event(wfd, EPOLLIN, 1);
+    c->fd = wfd;
+    c->r_action.recv_callback = handle_wakeup_fd_cb;
 
+    if (set_event(wfd, EPOLLIN, 1) < 0) {
+        close_conn(wfd);
+        return -1;
+    }
     return wfd;
 }
 
-// EPOLLOUT 
 void sync_wakeup() {
-	if (wakeup_fd < 0) return;
-	// set_event(wakeup_fd, EPOLLOUT, 0);
-
     uint64_t one = 1;
-    ssize_t n = write(wakeup_fd, &one, sizeof(one));
+    ssize_t n;
-}
 
-// #include "diskuring/diskuring.h"
+    if (wakeup_fd < 0) {
-// extern iouring_ctx_t global_uring_ctx;
+        return;
-// extern void iouring_tick(iouring_ctx_t *ctx);
+    }
 
-// 定时器
-int handle_timer_fd_cb(int fd){
-	uint64_t v;
     while (1) {
-        ssize_t n = read(fd, &v, sizeof(v));
+        n = write(wakeup_fd, &one, sizeof(one));
-        if (n == sizeof(v)) {
+        if (n == (ssize_t)sizeof(one)) {
+            return;
+        }
+        if (n < 0 && errno == EINTR) {
             continue;
         }
-        if (n < 0 && errno == EAGAIN) break;
+        if (n < 0 && errno == EAGAIN) {
-        break;
+            return;
+        }
+        return;
     }
-	// iouring_tick(&global_uring_ctx);
 }
 
-int init_timer_fd(void){
+int handle_timer_fd_cb(int fd) {
-	int tfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK | TFD_CLOEXEC);
+    uint64_t v;
 
-	struct itimerspec its = {
+    while (1) {
-		.it_interval = {0, 100 * 1000 * 1000},   // 100ms = 100,000,000 纳秒
+        ssize_t n = read(fd, &v, sizeof(v));
-		.it_value    = {0, 100 * 1000 * 1000},   // 首次 100ms 后触发
+        if (n == (ssize_t)sizeof(v)) {
-	};
+            continue;
-	timerfd_settime(tfd, 0, &its, NULL);
+        }
+        if (n < 0 && errno == EINTR) {
+            continue;
+        }
+        if (n < 0 && errno == EAGAIN) {
+            break;
+        }
+        break;
+    }
+    return 0;
+}
 
-	conn_list[tfd].fd = tfd;
+int init_timer_fd(void) {
-	conn_list[tfd].r_action.recv_callback = handle_timer_fd_cb;
+    int tfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK | TFD_CLOEXEC);
-	set_event(tfd, EPOLLIN, 1);
+    struct itimerspec its;
+    struct conn *c;
 
-	return tfd;
+    if (tfd < 0) {
+        printf("timerfd_create failed: errno=%d %s\n", errno, strerror(errno));
+        return -1;
+    }
+    if (!conn_fd_valid(tfd)) {
+        close(tfd);
+        return -1;
+    }
+
+    memset(&its, 0, sizeof(its));
+    its.it_interval.tv_nsec = 100 * 1000 * 1000;
+    its.it_value.tv_nsec = 100 * 1000 * 1000;
+
+    if (timerfd_settime(tfd, 0, &its, NULL) < 0) {
+        close(tfd);
+        return -1;
+    }
+
+    conn_clear_slot(tfd);
+    c = &conn_list[tfd];
+    c->fd = tfd;
+    c->r_action.recv_callback = handle_timer_fd_cb;
+
+    if (set_event(tfd, EPOLLIN, 1) < 0) {
+        close_conn(tfd);
+        return -1;
+    }
+    return tfd;
 }
 
 int r_init_server(unsigned short port) {
+    int sockfd = socket(AF_INET, SOCK_STREAM | SOCK_NONBLOCK | SOCK_CLOEXEC, 0);
+    int opt = 1;
+    struct sockaddr_in servaddr;
 
-	int sockfd = socket(AF_INET, SOCK_STREAM, 0);
+    if (sockfd < 0) {
-
+        return -1;
-	int opt = 1;
-    if (setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)) < 0) {
-        perror("setsockopt");
-		close(sockfd);
     }
 
-	struct sockaddr_in servaddr;
+    if (setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)) < 0) {
-	servaddr.sin_family = AF_INET;
+        close(sockfd);
-	servaddr.sin_addr.s_addr = htonl(INADDR_ANY); // 0.0.0.0
+        return -1;
-	servaddr.sin_port = htons(port); // 0-1023, 
+    }
 
-	if (-1 == bind(sockfd, (struct sockaddr*)&servaddr, sizeof(struct sockaddr))) {
+    memset(&servaddr, 0, sizeof(servaddr));
-		printf("bind failed: %s\n", strerror(errno));
+    servaddr.sin_family = AF_INET;
-	}
+    servaddr.sin_addr.s_addr = htonl(INADDR_ANY);
+    servaddr.sin_port = htons(port);
 
-	listen(sockfd, 10);
+    if (bind(sockfd, (struct sockaddr *)&servaddr, sizeof(servaddr)) < 0) {
-	//printf("listen finshed: %d\n", sockfd); // 3 
+        printf("bind failed on port %u: %s\n", port, strerror(errno));
+        close(sockfd);
+        return -1;
+    }
 
-	return sockfd;
+    if (listen(sockfd, 128) < 0) {
+        close(sockfd);
+        return -1;
+    }
 
+    return sockfd;
 }
 
 int reactor_start(unsigned short port, msg_handler handler) {
+    int listen_fds[MAX_PORTS];
+    int listen_count = 0;
+    int i;
 
-	//unsigned short port = 2000;
+    if (!handler) {
-	kvs_handler = handler;
+        return -1;
+    }
 
-	epfd = epoll_create(1);
+    for (i = 0; i < CONNECTION_SIZE; i++) {
+        conn_list[i].fd = -1;
+    }
 
-	wakeup_fd = init_wakeup_fd();
+    kvs_handler = handler;
-	if(wakeup_fd < 0){
+    epfd = epoll_create1(EPOLL_CLOEXEC);
-		close(epfd);
+    if (epfd < 0) {
-		return -1;
+        return -1;
-	}
+    }
 
-	timer_fd = init_timer_fd();
+    wakeup_fd = init_wakeup_fd();
-	if(timer_fd < 0){
+    if (wakeup_fd < 0) {
-		close(epfd);
+        close(epfd);
-		close(wakeup_fd);
+        epfd = -1;
-		return -1;
+        return -1;
-	}
+    }
 
-	int i = 0;
+    timer_fd = init_timer_fd();
+    if (timer_fd < 0) {
+        close_conn(wakeup_fd);
+        close(epfd);
+        wakeup_fd = -1;
+        epfd = -1;
+        return -1;
+    }
 
-	for (i = 0;i < MAX_PORTS;i ++) {
+    for (i = 0; i < MAX_PORTS; i++) {
+        int sockfd = r_init_server((unsigned short)(port + i));
+        struct conn *c;
+        if (sockfd < 0) {
+            continue;
+        }
+        if (!conn_fd_valid(sockfd)) {
+            close(sockfd);
+            continue;
+        }
 
-		int sockfd = r_init_server(port + i);
+        conn_clear_slot(sockfd);
+        c = &conn_list[sockfd];
+        c->fd = sockfd;
+        c->r_action.recv_callback = accept_cb;
+        c->is_stop = 0;
 
-		conn_list[sockfd].fd = sockfd;
+        if (set_event(sockfd, EPOLLIN, 1) < 0) {
-		conn_list[sockfd].r_action.recv_callback = accept_cb;
+            close_conn(sockfd);
-		conn_list[sockfd].is_stop = 0;
+            continue;
+        }
 
-		set_event(sockfd, EPOLLIN, 1);
+        listen_fds[listen_count++] = sockfd;
-	}
+    }
 
-	gettimeofday(&begin, NULL);
+    if (listen_count == 0) {
+        close_conn(timer_fd);
+        close_conn(wakeup_fd);
+        close(epfd);
+        timer_fd = -1;
+        wakeup_fd = -1;
+        epfd = -1;
+        return -1;
+    }
 
-	while (1) { // mainloop
+    gettimeofday(&begin, NULL);
-		struct epoll_event events[1024] = {0};
-		int nready = epoll_wait(epfd, events, 1024, -1);
-		// cleanup_finished_iouring_tasks();
 
-		int i = 0;
+    while (1) {
-		for (i = 0;i < nready;i ++) {
+        struct epoll_event events[1024];
+        int nready = epoll_wait(epfd, events, 1024, -1);
 
-			int connfd = events[i].data.fd;
+        if (nready < 0) {
+            if (errno == EINTR) {
+                continue;
+            }
+            break;
+        }
 
-			if (events[i].events & EPOLLIN) {
+        for (i = 0; i < nready; i++) {
-				// printf("connlist:%p, r_action:%p, recv_callaback:%p\n", &conn_list[connfd], &conn_list[connfd].r_action, conn_list[connfd].r_action.recv_callback);
+            int connfd = events[i].data.fd;
-				conn_list[connfd].r_action.recv_callback(connfd);
+            uint32_t ev = events[i].events;
-			} 
 
-			if (events[i].events & EPOLLOUT) {
+            if (!conn_fd_valid(connfd)) {
-				conn_list[connfd].send_callback(connfd);
+                continue;
-			}
+            }
-		}
 
-	}
+            if ((ev & (EPOLLERR | EPOLLHUP | EPOLLRDHUP)) &&
+                conn_list[connfd].r_action.recv_callback == recv_cb) {
+                close_conn(connfd);
+                continue;
+            }
 
-	if (wakeup_fd >= 0) close(wakeup_fd);
+            if ((ev & EPOLLIN) && conn_list[connfd].r_action.recv_callback) {
-	if (epfd >= 0) close(epfd);
+                conn_list[connfd].r_action.recv_callback(connfd);
-	return 0;
+            }
+
+            if (!conn_fd_valid(connfd) || conn_list[connfd].fd < 0) {
+                continue;
+            }
+
+            if ((ev & EPOLLOUT) && conn_list[connfd].send_callback) {
+                conn_list[connfd].send_callback(connfd);
+            }
+        }
+    }
+
+    for (i = 0; i < listen_count; i++) {
+        close_conn(listen_fds[i]);
+    }
+    if (timer_fd >= 0) {
+        close_conn(timer_fd);
+    }
+    if (wakeup_fd >= 0) {
+        close_conn(wakeup_fd);
+    }
+    if (epfd >= 0) {
+        close(epfd);
+    }
+
+    timer_fd = -1;
+    wakeup_fd = -1;
+    epfd = -1;
+    return 0;
 }
-
-

server.h

@@ -6,6 +6,7 @@
 #define __SERVER_H__
 
 #include <pthread.h>
+#include "network/chainbuffer.h"
 
 #define BUFFER_LENGTH		4096
 
@@ -20,11 +21,8 @@ typedef int (*RCALLBACK)(int fd);
 struct conn {
 	int fd;
 
-	char rbuffer[BUFFER_LENGTH];
+	chain_buffer_t rbuf;
-	int rlength;
+	chain_buffer_t wbuf;
-
-	char wbuffer[BUFFER_LENGTH*2];
-	int wlength;
 
 	RCALLBACK send_callback;
 
@@ -35,8 +33,6 @@ struct conn {
 
 	int is_stop;
 
-	pthread_mutex_t g_sync_lock;
-
 	int status;
 #if 1 // websocket
 	char *payload;

test-redis/README.md

@@ -0,0 +1,106 @@
+## run_bench.hash.sh 使用指南
+
+- 脚本：`test-redis/run_bench.hash.sh`
+- 默认策略：`persist_no` / `persist_everysec` / `nopersist`
+- 测试模式：`set` + `prefill+get`
+- 默认参数：`REQ=1000000 KEYSPACE=1000000 PIPE=128 VSIZE=32 ROUNDS=5 RETRIES=3`
+
+### 运行示例
+
+```bash
+# 默认运行（5轮）
+./test-redis/run_bench.hash.sh
+
+# 指定大 key 参数（示例）
+ROUNDS=5 REQ=1000000 KEYSPACE=1000000 PIPE=128 VSIZE=256 RETRIES=3 ./test-redis/run_bench.hash.sh
+```
+
+### 常用环境变量
+
+| 变量 | 默认值 | 说明 |
+|---|---|---|
+| `ROUNDS` | `5` | 每个策略的轮次 |
+| `REQ` | `1000000` | 单次 bench 请求数 |
+| `KEYSPACE` | `1000000` | key 空间大小 |
+| `PIPE` | `128` | pipeline 深度 |
+| `VSIZE` | `32` | value 大小（字节） |
+| `RETRIES` | `3` | 单轮失败重试次数 |
+| `SEED` | `12345` | 随机种子基数 |
+| `ALLOC` | `mypool` | allocator |
+| `KV_PORT` | `8888` | kvstore 端口 |
+| `SET_CMD` / `GET_CMD` | `RSET` / `RGET` | 压测命令 |
+
+## 大Key持久化协议复测（覆盖版）
+
+- 时间口径：2026-03-07（本次重测）
+- 参数：`requests=1000000 pipeline=128 keyspace=1000000 value-size=256`
+- 轮次：每种策略 `5` 轮
+- 去异常规则：每个场景按 `qps` 对 `5` 轮做中位数偏差，剔除偏差最大 `1` 轮，再对剩余 `4` 轮取均值
+- kvstore 源数据：`results/hash_bench_fair_summary_20260307_062549.csv`
+- redis 源数据：`results/redis_bigkey_summary_20260307_063158.csv`
+- kvstore 去异常结果：`results/hash_bench_fair_trimmed_20260307_062549.csv`
+- redis 去异常结果：`results/redis_bigkey_trimmed_20260307_063158.csv`
+
+## kvstore 原始5轮
+
+| 策略 | 模式 | Round1 | Round2 | Round3 | Round4 | Round5 | 原始均值QPS | 原始均值us/op |
+|---|---|---:|---:|---:|---:|---:|---:|---:|
+| persist_no | set | 139895.00 | 138768.00 | 138867.00 | 143296.00 | 144388.00 | 141042.80 | 7.09 |
+| persist_no | get | 163764.00 | 164079.00 | 166415.00 | 163959.00 | 147881.00 | 161219.60 | 6.21 |
+| persist_everysec | set | 133067.00 | 126140.00 | 133430.00 | 139786.00 | 141145.00 | 134713.60 | 7.43 |
+| persist_everysec | get | 162076.00 | 165415.00 | 162180.00 | 163762.00 | 158226.00 | 162331.80 | 6.16 |
+| nopersist | set | 157950.00 | 143109.00 | 135043.00 | 152045.00 | 147233.00 | 147076.00 | 6.82 |
+| nopersist | get | 163381.00 | 169356.00 | 155890.00 | 164754.00 | 157550.00 | 162186.20 | 6.17 |
+
+## kvstore 去异常后（4轮均值）
+
+| 策略 | 模式 | 剔除轮次 | 剔除QPS | 去异常均值QPS | 去异常均值us/op |
+|---|---|---:|---:|---:|---:|
+| persist_no | set | 5 | 144388.00 | 140206.50 | 7.13 |
+| persist_no | get | 5 | 147881.00 | 164554.25 | 6.08 |
+| persist_everysec | set | 5 | 141145.00 | 133105.75 | 7.52 |
+| persist_everysec | get | 5 | 158226.00 | 163358.25 | 6.12 |
+| nopersist | set | 3 | 135043.00 | 150084.25 | 6.67 |
+| nopersist | get | 3 | 155890.00 | 163760.25 | 6.11 |
+
+## kvstore 协议开销（基线 nopersist，去异常后）
+
+| 策略 | set QPS变化 | set us/op变化 | get QPS变化 | get us/op变化 |
+|---|---:|---:|---:|---:|
+| persist_no | -6.58% | +6.90% | +0.48% | -0.49% |
+| persist_everysec | -11.31% | +12.74% | -0.25% | +0.16% |
+
+## Redis 原始5轮
+
+| 策略 | 模式 | Round1 | Round2 | Round3 | Round4 | Round5 | 原始均值QPS | 原始均值us/op |
+|---|---|---:|---:|---:|---:|---:|---:|---:|
+| none | set | 198673.00 | 211843.00 | 193900.00 | 208293.00 | 211071.00 | 204756.00 | 4.89 |
+| none | get | 225347.00 | 221000.00 | 218357.00 | 216466.00 | 214747.00 | 219183.40 | 4.56 |
+| aof_no | set | 144042.00 | 139615.00 | 140016.00 | 149925.00 | 150658.00 | 144851.20 | 6.91 |
+| aof_no | get | 210541.00 | 211713.00 | 202589.00 | 231251.00 | 402028.00 | 251624.40 | 4.24 |
+| aof_everysec | set | 143737.00 | 132998.00 | 136973.00 | 144448.00 | 142414.00 | 140114.00 | 7.14 |
+| aof_everysec | get | 212422.00 | 201051.00 | 211508.00 | 192190.00 | 209874.00 | 205409.00 | 4.87 |
+
+## Redis 去异常后（4轮均值）
+
+| 策略 | 模式 | 剔除轮次 | 剔除QPS | 去异常均值QPS | 去异常均值us/op |
+|---|---|---:|---:|---:|---:|
+| none | set | 3 | 193900.00 | 207470.00 | 4.82 |
+| none | get | 1 | 225347.00 | 217642.50 | 4.59 |
+| aof_no | set | 5 | 150658.00 | 143399.50 | 6.98 |
+| aof_no | get | 5 | 402028.00 | 214023.50 | 4.68 |
+| aof_everysec | set | 2 | 132998.00 | 141893.00 | 7.05 |
+| aof_everysec | get | 4 | 192190.00 | 208713.75 | 4.79 |
+
+## Redis 协议开销（基线 none，去异常后）
+
+| 策略 | set QPS变化 | set us/op变化 | get QPS变化 | get us/op变化 |
+|---|---:|---:|---:|---:|
+| aof_no | -30.88% | +44.81% | -1.66% | +1.96% |
+| aof_everysec | -31.61% | +46.27% | -4.10% | +4.36% |
+
+## 简要结论
+
+1. 本次大 key（256）下，持久化开销主要体现在 `set`，`get` 相对更稳。
+2. `aof_no/aof_everysec` 在 Redis 写路径的开销明显；kvstore 三策略在写路径也存在可见差异。
+3. 本报告已剔除每场景 1 个异常轮次，避免单点抖动主导结论。

test-redis/bench.c

@@ -0,0 +1,447 @@
+#include <errno.h>
+#include <inttypes.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <strings.h>
+#include <time.h>
+
+#include <hiredis/hiredis.h>
+
+typedef enum {
+    MODE_SET = 0,
+    MODE_GET = 1,
+    MODE_MIXED = 2,
+} bench_mode_t;
+
+typedef struct {
+    const char *host;
+    int port;
+    bench_mode_t mode;
+    uint64_t requests;
+    uint32_t pipeline;
+    uint32_t keyspace;
+    uint32_t value_size;
+    uint32_t set_ratio;
+    const char *set_cmd;
+    const char *get_cmd;
+    const char *key_prefix;
+    uint64_t seed;
+    int verify_get;
+} bench_opts_t;
+
+typedef struct {
+    uint64_t set_ops;
+    uint64_t get_ops;
+    uint64_t errors;
+    double elapsed_sec;
+} bench_result_t;
+
+static void usage(const char *prog) {
+    fprintf(stderr,
+            "Usage: %s [options]\n"
+            "  --host <ip>             default: 127.0.0.1\n"
+            "  --port <port>           default: 6379\n"
+            "  --mode <set|get|mixed>  default: mixed\n"
+            "  --requests <n>          default: 1000000\n"
+            "  --pipeline <n>          default: 64\n"
+            "  --keyspace <n>          default: 100000\n"
+            "  --value-size <n>        default: 32\n"
+            "  --set-ratio <0..100>    default: 50 (mixed mode only)\n"
+            "  --set-cmd <cmd>         default: SET\n"
+            "  --get-cmd <cmd>         default: GET\n"
+            "  --key-prefix <prefix>   default: bench:key:\n"
+            "  --seed <n>              default: time-based\n"
+            "  --verify-get            verify GET value content\n"
+            "\nExamples:\n"
+            "  # Benchmark Redis\n"
+            "  %s --host 127.0.0.1 --port 6379 --mode mixed --requests 2000000\n"
+            "\n"
+            "  # Benchmark kvstore with Redis-compatible commands\n"
+            "  %s --host 127.0.0.1 --port 8888 --mode mixed --requests 2000000\n"
+            "\n"
+            "  # Benchmark kvstore RBTree path\n"
+            "  %s --host 127.0.0.1 --port 8888 --mode mixed --set-cmd RSET --get-cmd RGET\n",
+            prog, prog, prog, prog);
+}
+
+static void opts_init(bench_opts_t *o) {
+    memset(o, 0, sizeof(*o));
+    o->host = "127.0.0.1";
+    o->port = 6379;
+    o->mode = MODE_MIXED;
+    o->requests = 1000000;
+    o->pipeline = 64;
+    o->keyspace = 100000;
+    o->value_size = 32;
+    o->set_ratio = 50;
+    o->set_cmd = "SET";
+    o->get_cmd = "GET";
+    o->key_prefix = "bench:key:";
+    o->seed = (uint64_t)time(NULL);
+    o->verify_get = 0;
+}
+
+static int parse_u64(const char *s, uint64_t *out) {
+    char *end = NULL;
+    unsigned long long v;
+    errno = 0;
+    v = strtoull(s, &end, 10);
+    if (errno != 0 || end == s || *end != 0) {
+        return -1;
+    }
+    *out = (uint64_t)v;
+    return 0;
+}
+
+static int parse_u32(const char *s, uint32_t *out) {
+    uint64_t v = 0;
+    if (parse_u64(s, &v) != 0 || v > UINT32_MAX) {
+        return -1;
+    }
+    *out = (uint32_t)v;
+    return 0;
+}
+
+static int parse_args(int argc, char **argv, bench_opts_t *o) {
+    int i;
+    for (i = 1; i < argc; i++) {
+        if (strcmp(argv[i], "--host") == 0 && i + 1 < argc) {
+            o->host = argv[++i];
+        } else if (strcmp(argv[i], "--port") == 0 && i + 1 < argc) {
+            uint32_t p = 0;
+            if (parse_u32(argv[++i], &p) != 0 || p == 0 || p > 65535) {
+                return -1;
+            }
+            o->port = (int)p;
+        } else if (strcmp(argv[i], "--mode") == 0 && i + 1 < argc) {
+            const char *m = argv[++i];
+            if (strcmp(m, "set") == 0) {
+                o->mode = MODE_SET;
+            } else if (strcmp(m, "get") == 0) {
+                o->mode = MODE_GET;
+            } else if (strcmp(m, "mixed") == 0) {
+                o->mode = MODE_MIXED;
+            } else {
+                return -1;
+            }
+        } else if (strcmp(argv[i], "--requests") == 0 && i + 1 < argc) {
+            if (parse_u64(argv[++i], &o->requests) != 0 || o->requests == 0) {
+                return -1;
+            }
+        } else if (strcmp(argv[i], "--pipeline") == 0 && i + 1 < argc) {
+            if (parse_u32(argv[++i], &o->pipeline) != 0 || o->pipeline == 0) {
+                return -1;
+            }
+        } else if (strcmp(argv[i], "--keyspace") == 0 && i + 1 < argc) {
+            if (parse_u32(argv[++i], &o->keyspace) != 0 || o->keyspace == 0) {
+                return -1;
+            }
+        } else if (strcmp(argv[i], "--value-size") == 0 && i + 1 < argc) {
+            if (parse_u32(argv[++i], &o->value_size) != 0 || o->value_size == 0) {
+                return -1;
+            }
+        } else if (strcmp(argv[i], "--set-ratio") == 0 && i + 1 < argc) {
+            if (parse_u32(argv[++i], &o->set_ratio) != 0 || o->set_ratio > 100) {
+                return -1;
+            }
+        } else if (strcmp(argv[i], "--set-cmd") == 0 && i + 1 < argc) {
+            o->set_cmd = argv[++i];
+        } else if (strcmp(argv[i], "--get-cmd") == 0 && i + 1 < argc) {
+            o->get_cmd = argv[++i];
+        } else if (strcmp(argv[i], "--key-prefix") == 0 && i + 1 < argc) {
+            o->key_prefix = argv[++i];
+        } else if (strcmp(argv[i], "--seed") == 0 && i + 1 < argc) {
+            if (parse_u64(argv[++i], &o->seed) != 0) {
+                return -1;
+            }
+        } else if (strcmp(argv[i], "--verify-get") == 0) {
+            o->verify_get = 1;
+        } else if (strcmp(argv[i], "--help") == 0 || strcmp(argv[i], "-h") == 0) {
+            return 1;
+        } else {
+            return -1;
+        }
+    }
+    return 0;
+}
+
+static uint64_t mono_ns(void) {
+    struct timespec ts;
+    clock_gettime(CLOCK_MONOTONIC, &ts);
+    return (uint64_t)ts.tv_sec * 1000000000ull + (uint64_t)ts.tv_nsec;
+}
+
+static uint64_t xorshift64(uint64_t *state) {
+    uint64_t x = *state;
+    x ^= x << 13;
+    x ^= x >> 7;
+    x ^= x << 17;
+    *state = x;
+    return x;
+}
+
+static int append_set(redisContext *c, const char *cmd,
+                      const char *key, size_t key_len,
+                      const char *val, size_t val_len) {
+    const char *argv[3];
+    size_t argvlen[3];
+    argv[0] = cmd;
+    argvlen[0] = strlen(cmd);
+    argv[1] = key;
+    argvlen[1] = key_len;
+    argv[2] = val;
+    argvlen[2] = val_len;
+    return redisAppendCommandArgv(c, 3, argv, argvlen);
+}
+
+static int append_get(redisContext *c, const char *cmd,
+                      const char *key, size_t key_len) {
+    const char *argv[2];
+    size_t argvlen[2];
+    argv[0] = cmd;
+    argvlen[0] = strlen(cmd);
+    argv[1] = key;
+    argvlen[1] = key_len;
+    return redisAppendCommandArgv(c, 2, argv, argvlen);
+}
+
+static int consume_set_reply(redisContext *c) {
+    redisReply *r = NULL;
+    if (redisGetReply(c, (void **)&r) != REDIS_OK || r == NULL) {
+        return -1;
+    }
+    if (r->type != REDIS_REPLY_STATUS || r->str == NULL || strcasecmp(r->str, "OK") != 0) {
+        freeReplyObject(r);
+        return -1;
+    }
+    freeReplyObject(r);
+    return 0;
+}
+
+static int consume_get_reply(redisContext *c, const char *expect, size_t expect_len, int verify) {
+    redisReply *r = NULL;
+    if (redisGetReply(c, (void **)&r) != REDIS_OK || r == NULL) {
+        return -1;
+    }
+
+    if (r->type == REDIS_REPLY_STRING) {
+        if (verify && ((size_t)r->len != expect_len || memcmp(r->str, expect, expect_len) != 0)) {
+            freeReplyObject(r);
+            return -1;
+        }
+        freeReplyObject(r);
+        return 0;
+    }
+
+    if (r->type == REDIS_REPLY_NIL) {
+        freeReplyObject(r);
+        return verify ? -1 : 0;
+    }
+
+    freeReplyObject(r);
+    return -1;
+}
+
+static int prefill(redisContext *c, const bench_opts_t *o, const char *value, size_t value_len) {
+    uint64_t i = 0;
+    char key[256];
+
+    while (i < o->keyspace) {
+        uint32_t batch = o->pipeline;
+        if ((uint64_t)batch > (uint64_t)o->keyspace - i) {
+            batch = (uint32_t)((uint64_t)o->keyspace - i);
+        }
+
+        for (uint32_t j = 0; j < batch; j++) {
+            int klen = snprintf(key, sizeof(key), "%s%" PRIu64, o->key_prefix, i + j);
+            if (klen <= 0 || (size_t)klen >= sizeof(key)) {
+                return -1;
+            }
+            if (append_set(c, o->set_cmd, key, (size_t)klen, value, value_len) != REDIS_OK) {
+                return -1;
+            }
+        }
+
+        for (uint32_t j = 0; j < batch; j++) {
+            if (consume_set_reply(c) != 0) {
+                return -1;
+            }
+        }
+
+        i += batch;
+    }
+
+    return 0;
+}
+
+/*
+ * Keep append/reply operation choices consistent in each batch by building an op mask.
+ * This wrapper keeps implementation simple and avoids per-op heap allocation.
+ */
+static int run_bench_with_mask(redisContext *c, const bench_opts_t *o,
+                               const char *value, size_t value_len,
+                               bench_result_t *res) {
+    uint64_t done = 0;
+    uint64_t rng = o->seed ? o->seed : 1;
+    char key[256];
+    uint8_t *opmask = NULL;
+    uint64_t begin_ns;
+
+    memset(res, 0, sizeof(*res));
+    opmask = (uint8_t *)malloc(o->pipeline);
+    if (!opmask) {
+        return -1;
+    }
+
+    begin_ns = mono_ns();
+
+    while (done < o->requests) {
+        uint32_t batch = o->pipeline;
+        if ((uint64_t)batch > o->requests - done) {
+            batch = (uint32_t)(o->requests - done);
+        }
+
+        for (uint32_t i = 0; i < batch; i++) {
+            uint64_t rnd = xorshift64(&rng);
+            uint64_t key_id = rnd % o->keyspace;
+            int is_set = 0;
+            int klen;
+
+            if (o->mode == MODE_SET) {
+                is_set = 1;
+            } else if (o->mode == MODE_GET) {
+                is_set = 0;
+            } else {
+                is_set = (rnd % 100) < o->set_ratio;
+            }
+            opmask[i] = (uint8_t)is_set;
+
+            klen = snprintf(key, sizeof(key), "%s%" PRIu64, o->key_prefix, key_id);
+            if (klen <= 0 || (size_t)klen >= sizeof(key)) {
+                free(opmask);
+                return -1;
+            }
+
+            if (is_set) {
+                if (append_set(c, o->set_cmd, key, (size_t)klen, value, value_len) != REDIS_OK) {
+                    free(opmask);
+                    return -1;
+                }
+                res->set_ops++;
+            } else {
+                if (append_get(c, o->get_cmd, key, (size_t)klen) != REDIS_OK) {
+                    free(opmask);
+                    return -1;
+                }
+                res->get_ops++;
+            }
+        }
+
+        for (uint32_t i = 0; i < batch; i++) {
+            int rc = opmask[i] ? consume_set_reply(c)
+                               : consume_get_reply(c, value, value_len, o->verify_get);
+            if (rc != 0) {
+                res->errors++;
+                free(opmask);
+                return -1;
+            }
+        }
+
+        done += batch;
+    }
+
+    res->elapsed_sec = (double)(mono_ns() - begin_ns) / 1e9;
+    free(opmask);
+    return 0;
+}
+
+int main(int argc, char **argv) {
+    bench_opts_t opts;
+    bench_result_t result;
+    redisContext *ctx;
+    struct timeval timeout;
+    char *value;
+    size_t value_len;
+    int parse_rc;
+
+    opts_init(&opts);
+    parse_rc = parse_args(argc, argv, &opts);
+    if (parse_rc == 1) {
+        usage(argv[0]);
+        return 0;
+    }
+    if (parse_rc != 0) {
+        usage(argv[0]);
+        return 2;
+    }
+
+    timeout.tv_sec = 3;
+    timeout.tv_usec = 0;
+    ctx = redisConnectWithTimeout(opts.host, opts.port, timeout);
+    if (!ctx || ctx->err) {
+        fprintf(stderr, "connect %s:%d failed: %s\n", opts.host, opts.port,
+                ctx ? ctx->errstr : "oom");
+        if (ctx) {
+            redisFree(ctx);
+        }
+        return 1;
+    }
+
+    value_len = opts.value_size;
+    value = (char *)malloc(value_len);
+    if (!value) {
+        fprintf(stderr, "malloc value buffer failed\n");
+        redisFree(ctx);
+        return 1;
+    }
+    for (size_t i = 0; i < value_len; i++) {
+        value[i] = (char)('a' + (int)(i % 26));
+    }
+
+    if (opts.mode != MODE_SET) {
+        fprintf(stdout, "[prefill] keyspace=%u using %s\n", opts.keyspace, opts.set_cmd);
+        if (prefill(ctx, &opts, value, value_len) != 0) {
+            fprintf(stderr, "prefill failed, err=%s\n", ctx->err ? ctx->errstr : "unknown");
+            free(value);
+            redisFree(ctx);
+            return 1;
+        }
+    }
+
+    fprintf(stdout,
+            "[bench] target=%s:%d mode=%s requests=%" PRIu64
+            " pipeline=%u keyspace=%u value_size=%u set_cmd=%s get_cmd=%s\n",
+            opts.host, opts.port,
+            opts.mode == MODE_SET ? "set" : (opts.mode == MODE_GET ? "get" : "mixed"),
+            opts.requests, opts.pipeline, opts.keyspace, opts.value_size,
+            opts.set_cmd, opts.get_cmd);
+
+    if (run_bench_with_mask(ctx, &opts, value, value_len, &result) != 0) {
+        fprintf(stderr, "benchmark failed, err=%s\n", ctx->err ? ctx->errstr : "reply mismatch");
+        free(value);
+        redisFree(ctx);
+        return 1;
+    }
+
+    {
+        double qps = result.elapsed_sec > 0 ? (double)(result.set_ops + result.get_ops) / result.elapsed_sec : 0.0;
+        double avg_us = (result.set_ops + result.get_ops) > 0
+                            ? (result.elapsed_sec * 1e6) / (double)(result.set_ops + result.get_ops)
+                            : 0.0;
+        fprintf(stdout,
+                "[result] elapsed=%.3fs total=%" PRIu64 " set=%" PRIu64 " get=%" PRIu64
+                " errors=%" PRIu64 " qps=%.0f avg=%.2fus/op\n",
+                result.elapsed_sec,
+                result.set_ops + result.get_ops,
+                result.set_ops,
+                result.get_ops,
+                result.errors,
+                qps,
+                avg_us);
+    }
+
+    free(value);
+    redisFree(ctx);
+    return 0;
+}

test-redis/run_bench.hash.sh

@@ -0,0 +1,437 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
+ROOT_DIR=$(cd "$SCRIPT_DIR/.." && pwd)
+cd "$ROOT_DIR"
+
+TS=$(date +%Y%m%d_%H%M%S)
+RUN_TIME=$(date '+%Y-%m-%d %H:%M:%S')
+
+OUT_DIR="$ROOT_DIR/test-redis/results"
+mkdir -p "$OUT_DIR"
+
+DETAIL_CSV="$OUT_DIR/hash_bench_fair_detail_${TS}.csv"
+SUMMARY_CSV="$OUT_DIR/hash_bench_fair_summary_${TS}.csv"
+KV_LOG_DIR="/tmp/kv_bench_hash_fair_${TS}"
+mkdir -p "$KV_LOG_DIR"
+
+CONFIG_XML="$ROOT_DIR/config/config.xml"
+README_MD="$ROOT_DIR/test-redis/README.md"
+
+ROUNDS=${ROUNDS:-5}
+RETRIES=${RETRIES:-3}
+REQ=${REQ:-1000000}
+PIPE=${PIPE:-128}
+KEYSPACE=${KEYSPACE:-1000000}
+VSIZE=${VSIZE:-32}
+SEED=${SEED:-12345}
+ALLOC=${ALLOC:-mypool}
+KV_HOST=127.0.0.1
+KV_PORT=${KV_PORT:-8888}
+SET_CMD=${SET_CMD:-RSET}
+GET_CMD=${GET_CMD:-RGET}
+
+ORIG_CONFIG_BACKUP=$(mktemp "/tmp/kvstore_config_backup_${TS}.XXXXXX")
+cp "$CONFIG_XML" "$ORIG_CONFIG_BACKUP"
+KV_PID=""
+
+printf "strategy,persistence,oplog_sync,allocator,mode,round,qps,avg_us,elapsed_s,key_prefix,seed,requests,pipeline,keyspace,value_size\n" > "$DETAIL_CSV"
+printf "strategy,persistence,oplog_sync,allocator,mode,round_qps,round_avg_us,round_elapsed_s,avg_qps,avg_avg_us,avg_elapsed_s\n" > "$SUMMARY_CSV"
+
+require_cmd() {
+  local cmd="$1"
+  if ! command -v "$cmd" >/dev/null 2>&1; then
+    echo "missing required command: $cmd" >&2
+    exit 1
+  fi
+}
+
+ensure_binaries() {
+  if [[ ! -x "$ROOT_DIR/kvstore" || ! -x "$ROOT_DIR/test-redis/bench" ]]; then
+    echo "[info] kvstore/bench missing, running make -j4 ..."
+    make -j4
+  fi
+}
+
+set_config() {
+  local ptype="$1"
+  local oplog_sync="$2"
+  local pdir="$3"
+  local alloc="$4"
+  python3 - "$CONFIG_XML" "$ptype" "$oplog_sync" "$pdir" "$alloc" "$KV_PORT" <<'PY'
+import sys
+import xml.etree.ElementTree as ET
+
+path, ptype, oplog_sync, pdir, alloc, kv_port = sys.argv[1:]
+tree = ET.parse(path)
+root = tree.getroot()
+
+server = root.find("server")
+if server is not None:
+    ip = server.find("ip")
+    port = server.find("port")
+    mode = server.find("mode")
+    replica = server.find("replica")
+    if ip is not None:
+        ip.text = "127.0.0.1"
+    if port is not None:
+        port.text = kv_port
+    if mode is not None:
+        mode.text = "master"
+    if replica is not None:
+        replica.text = "disable"
+
+persistence = root.find("persistence")
+if persistence is not None:
+    t = persistence.find("type")
+    d = persistence.find("dir")
+    s = persistence.find("oplog_sync")
+    if t is not None:
+        t.text = ptype
+    if d is not None:
+        d.text = pdir
+    if s is not None:
+        s.text = oplog_sync
+
+memory = root.find("memory")
+if memory is not None:
+    a = memory.find("allocator")
+    leak = memory.find("leakage")
+    if a is not None:
+        a.text = alloc
+    if leak is not None:
+        leak.text = "disable"
+
+tree.write(path, encoding="UTF-8", xml_declaration=True)
+PY
+}
+
+wait_port_open() {
+  local port="$1"
+  for _ in $(seq 1 200); do
+    if ss -ltn | rg -q ":${port}\\b"; then
+      return 0
+    fi
+    sleep 0.1
+  done
+  return 1
+}
+
+wait_port_close() {
+  local port="$1"
+  for _ in $(seq 1 200); do
+    if ! ss -ltn | rg -q ":${port}\\b"; then
+      return 0
+    fi
+    sleep 0.1
+  done
+  return 1
+}
+
+assert_port_free() {
+  local port="$1"
+  if ss -ltn | rg -q ":${port}\\b"; then
+    echo "port ${port} is already in use, cannot start kvstore" >&2
+    exit 1
+  fi
+}
+
+extract_metric() {
+  local line="$1"
+  local key="$2"
+  echo "$line" | sed -E "s/.*${key}=([0-9]+(\\.[0-9]+)?).*/\\1/"
+}
+
+run_bench_capture() {
+  local mode="$1"
+  local key_prefix="$2"
+  local seed="$3"
+  local verify="$4"
+  local cmd=(
+    ./test-redis/bench
+    --host "$KV_HOST"
+    --port "$KV_PORT"
+    --mode "$mode"
+    --set-cmd "$SET_CMD"
+    --get-cmd "$GET_CMD"
+    --requests "$REQ"
+    --pipeline "$PIPE"
+    --keyspace "$KEYSPACE"
+    --value-size "$VSIZE"
+    --seed "$seed"
+    --key-prefix "$key_prefix"
+  )
+  if [[ "$verify" == "1" ]]; then
+    cmd+=(--verify-get)
+  fi
+
+  local out
+  out=$("${cmd[@]}")
+  echo "$out" >&2
+
+  local line
+  line=$(echo "$out" | rg "\\[result\\]" | tail -n1)
+  if [[ -z "$line" ]]; then
+    echo "missing [result] line in benchmark output" >&2
+    return 1
+  fi
+
+  local qps avg elapsed
+  qps=$(extract_metric "$line" "qps")
+  avg=$(extract_metric "$line" "avg")
+  elapsed=$(extract_metric "$line" "elapsed")
+  if [[ -z "$qps" || -z "$avg" || -z "$elapsed" ]]; then
+    echo "failed to parse benchmark metrics: $line" >&2
+    return 1
+  fi
+
+  echo "$qps,$avg,$elapsed"
+}
+
+start_kv() {
+  local label="$1"
+  assert_port_free "$KV_PORT"
+  ./kvstore >"$KV_LOG_DIR/${label}.log" 2>&1 &
+  KV_PID=$!
+  if ! wait_port_open "$KV_PORT"; then
+    echo "kvstore start failed for ${label}" >&2
+    return 1
+  fi
+}
+
+stop_kv() {
+  if [[ -n "${KV_PID:-}" ]] && kill -0 "$KV_PID" >/dev/null 2>&1; then
+    kill "$KV_PID" >/dev/null 2>&1 || true
+    wait "$KV_PID" >/dev/null 2>&1 || true
+  fi
+  KV_PID=""
+  wait_port_close "$KV_PORT" || true
+}
+
+cleanup() {
+  stop_kv
+  if [[ -f "$ORIG_CONFIG_BACKUP" ]]; then
+    cp "$ORIG_CONFIG_BACKUP" "$CONFIG_XML"
+    rm -f "$ORIG_CONFIG_BACKUP"
+  fi
+}
+trap cleanup EXIT
+
+case_params() {
+  local strategy="$1"
+  case "$strategy" in
+    persist_no)
+      echo "incremental,none"
+      ;;
+    persist_everysec)
+      echo "incremental,every_sec"
+      ;;
+    nopersist)
+      echo "none,none"
+      ;;
+    *)
+      echo "unknown strategy: $strategy" >&2
+      return 1
+      ;;
+  esac
+}
+
+run_one_case_round() {
+  local strategy="$1"
+  local round="$2"
+  local params ptype oplog_sync
+  local pdir_rel pdir_abs
+  local key_prefix seed
+  local m qps avg elapsed
+  local set_qps set_avg set_elapsed
+  local get_qps get_avg get_elapsed
+  local attempt ok
+
+  params=$(case_params "$strategy")
+  IFS=',' read -r ptype oplog_sync <<< "$params"
+
+  key_prefix="bench:${TS}:round:${round}:"
+  seed=$((SEED + round))
+
+  for attempt in $(seq 1 "$RETRIES"); do
+    pdir_rel="data/${strategy}_${TS}_r${round}_a${attempt}"
+    pdir_abs="$ROOT_DIR/${pdir_rel}"
+    rm -rf "$pdir_abs"
+    mkdir -p "$pdir_abs"
+
+    set_config "$ptype" "$oplog_sync" "$pdir_rel" "$ALLOC"
+    start_kv "${strategy}_r${round}_a${attempt}"
+
+    ok=1
+    if ! m=$(run_bench_capture "set" "$key_prefix" "$seed" 0); then
+      ok=0
+    else
+      IFS=',' read -r set_qps set_avg set_elapsed <<< "$m"
+    fi
+
+    if [[ "$ok" == "1" ]]; then
+      if ! m=$(run_bench_capture "get" "$key_prefix" "$seed" 1); then
+        ok=0
+      else
+        IFS=',' read -r get_qps get_avg get_elapsed <<< "$m"
+      fi
+    fi
+
+    stop_kv
+
+    if [[ "$ok" == "1" ]]; then
+      printf "%s,%s,%s,%s,set,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s\n" \
+        "$strategy" "$ptype" "$oplog_sync" "$ALLOC" "$round" \
+        "$set_qps" "$set_avg" "$set_elapsed" "$key_prefix" "$seed" \
+        "$REQ" "$PIPE" "$KEYSPACE" "$VSIZE" >> "$DETAIL_CSV"
+
+      printf "%s,%s,%s,%s,get,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s\n" \
+        "$strategy" "$ptype" "$oplog_sync" "$ALLOC" "$round" \
+        "$get_qps" "$get_avg" "$get_elapsed" "$key_prefix" "$seed" \
+        "$REQ" "$PIPE" "$KEYSPACE" "$VSIZE" >> "$DETAIL_CSV"
+      return 0
+    fi
+
+    echo "[warn] retry $attempt/$RETRIES failed: strategy=$strategy round=$round" >&2
+    sleep 0.2
+  done
+
+  echo "[error] all retries failed: strategy=$strategy round=$round" >&2
+  return 1
+}
+
+append_readme_results() {
+  python3 - "$DETAIL_CSV" "$SUMMARY_CSV" "$README_MD" "$RUN_TIME" "$ROUNDS" "$REQ" "$PIPE" "$KEYSPACE" "$VSIZE" "$ALLOC" "$SET_CMD" "$GET_CMD" <<'PY'
+import csv
+import os
+import sys
+from collections import defaultdict
+
+detail_csv, summary_csv, readme_path, run_time, rounds, req, pipeline, keyspace, value_size, alloc, set_cmd, get_cmd = sys.argv[1:]
+rounds_i = int(rounds)
+
+group = defaultdict(list)
+meta = {}
+
+with open(detail_csv, newline="", encoding="utf-8") as f:
+    reader = csv.DictReader(f)
+    for r in reader:
+        key = (r["strategy"], r["mode"])
+        group[key].append(r)
+        meta[r["strategy"]] = (r["persistence"], r["oplog_sync"], r["allocator"])
+
+def f2(x):
+    return f"{float(x):.2f}"
+
+with open(summary_csv, "w", newline="", encoding="utf-8") as f:
+    w = csv.writer(f)
+    w.writerow([
+        "strategy","persistence","oplog_sync","allocator","mode",
+        "round_qps","round_avg_us","round_elapsed_s",
+        "avg_qps","avg_avg_us","avg_elapsed_s"
+    ])
+
+    for strategy in ["persist_no", "persist_everysec", "nopersist"]:
+        for mode in ["set", "get"]:
+            rows = sorted(group[(strategy, mode)], key=lambda x: int(x["round"]))
+            qps = [float(r["qps"]) for r in rows]
+            avg_us = [float(r["avg_us"]) for r in rows]
+            elapsed = [float(r["elapsed_s"]) for r in rows]
+            persistence, oplog_sync, allocator = meta[strategy]
+            w.writerow([
+                strategy, persistence, oplog_sync, allocator, mode,
+                "|".join(f2(v) for v in qps),
+                "|".join(f2(v) for v in avg_us),
+                "|".join(f2(v) for v in elapsed),
+                f2(sum(qps) / len(qps)),
+                f2(sum(avg_us) / len(avg_us)),
+                f2(sum(elapsed) / len(elapsed)),
+            ])
+
+rows = []
+with open(summary_csv, newline="", encoding="utf-8") as f:
+    rows = list(csv.DictReader(f))
+
+def by_mode(mode):
+    ordered = ["persist_no", "persist_everysec", "nopersist"]
+    m = [r for r in rows if r["mode"] == mode]
+    m.sort(key=lambda r: ordered.index(r["strategy"]))
+    return m
+
+round_headers = [f"Round{i}" for i in range(1, rounds_i + 1)]
+detail_rel = os.path.relpath(detail_csv, os.path.dirname(readme_path))
+summary_rel = os.path.relpath(summary_csv, os.path.dirname(readme_path))
+
+lines = []
+lines.append("")
+lines.append(f"## run_bench.hash.sh {rounds_i}轮均值复测（{run_time}）")
+lines.append("")
+lines.append(f"- 轮次：{rounds_i} 轮（每种情况）")
+lines.append("- 策略：persist(no), persist(everysec), nopersist")
+lines.append(f"- 命令：{set_cmd}/{get_cmd}（GET 保持 prefill + GET）")
+lines.append("- 公平性：同一 case 同一轮的 SET/GET 使用相同 key_prefix 与 seed；case 顺序按轮次轮转")
+lines.append(f"- 参数：requests={req} pipeline={pipeline} keyspace={keyspace} value-size={value_size} allocator={alloc}")
+lines.append(f"- 明细数据：`{detail_rel}`")
+lines.append(f"- 汇总数据：`{summary_rel}`")
+lines.append("")
+
+for mode in ["set", "get"]:
+    lines.append(f"### kvstore：{set_cmd}/{get_cmd} ({mode})")
+    lines.append("")
+    header = "| 场景 | persistence | oplog_sync | " + " | ".join(round_headers) + " | 均值QPS | 均值avg(us/op) | 均值elapsed(s) |"
+    sep = "|---|---|---|" + "|".join(["---:"] * len(round_headers)) + "|---:|---:|---:|"
+    lines.append(header)
+    lines.append(sep)
+    for r in by_mode(mode):
+        qps_rounds = r["round_qps"].split("|")
+        lines.append(
+            "| {} | {} | {} | {} | {} | {} | {} |".format(
+                r["strategy"],
+                r["persistence"],
+                r["oplog_sync"],
+                " | ".join(qps_rounds),
+                r["avg_qps"],
+                r["avg_avg_us"],
+                r["avg_elapsed_s"],
+            )
+        )
+    lines.append("")
+
+with open(readme_path, "a", encoding="utf-8") as f:
+    f.write("\n".join(lines) + "\n")
+PY
+}
+
+main() {
+  require_cmd python3
+  require_cmd rg
+  require_cmd ss
+  ensure_binaries
+
+  for round in $(seq 1 "$ROUNDS"); do
+    case $((round % 3)) in
+      1)
+        order=(persist_no persist_everysec nopersist)
+        ;;
+      2)
+        order=(persist_everysec nopersist persist_no)
+        ;;
+      0)
+        order=(nopersist persist_no persist_everysec)
+        ;;
+    esac
+
+    for strategy in "${order[@]}"; do
+      run_one_case_round "$strategy" "$round"
+    done
+  done
+
+  append_readme_results
+
+  echo "DETAIL_CSV=$DETAIL_CSV"
+  echo "SUMMARY_CSV=$SUMMARY_CSV"
+  echo "README_UPDATED=$README_MD"
+}
+
+main "$@"

test-redis/test.c

@@ -10,48 +10,49 @@
 
 static void die(redisContext *c, const char *msg) {
     fprintf(stderr, "%s: %s\n", msg, c && c->err ? c->errstr : "unknown");
+    redisFree(c);
     exit(1);
 }
 
-static void must_ok(redisReply *r, const char *what) {
+static int must_ok(redisReply *r, const char *what) {
-    if (!r) { fprintf(stderr, "%s: reply null\n", what); exit(1); }
+    if (!r) { fprintf(stderr, "%s: reply null\n", what); return -1; }
     if (!(r->type == REDIS_REPLY_STATUS && r->str && strcasecmp(r->str, "OK") == 0)) {
         fprintf(stderr, "%s: expect +OK, got type=%d str=%s\n",
                 what, r->type, r->str ? r->str : "(null)");
         freeReplyObject(r);
-        exit(1);
+        return -1;
     }
     freeReplyObject(r);
 }
 
-static void must_int(redisReply *r, long long expect, const char *what) {
+static int must_int(redisReply *r, long long expect, const char *what) {
-    if (!r) { fprintf(stderr, "%s: reply null\n", what); exit(1); }
+    if (!r) { fprintf(stderr, "%s: reply null\n", what); return -1; }
     if (r->type != REDIS_REPLY_INTEGER || r->integer != expect) {
         fprintf(stderr, "%s: expect :%lld, got type=%d int=%lld\n",
                 what, expect, r->type, (long long)r->integer);
         freeReplyObject(r);
-        exit(1);
+        return -1;
     }
     freeReplyObject(r);
 }
 
-static void must_bulk_eq(redisReply *r, const void *buf, size_t n, const char *what) {
+static int must_bulk_eq(redisReply *r, const void *buf, size_t n, const char *what) {
-    if (!r) { fprintf(stderr, "%s: reply null\n", what); exit(1); }
+    if (!r) { fprintf(stderr, "%s: reply null\n", what); return -1; }
     if (r->type != REDIS_REPLY_STRING || r->len != n || memcmp(r->str, buf, n) != 0) {
         fprintf(stderr, "%s: bulk mismatch. type=%d len=%zu\n", what, r->type, r->len);
         fprintf(stderr, "expect:%s, truely:%s\n", (const char*)buf, r->str);
         freeReplyObject(r);
-        exit(1);
+        return -1;
     }
     freeReplyObject(r);
 }
 
-static void must_nil(redisReply *r, const char *what) {
+static int must_nil(redisReply *r, const char *what) {
-    if (!r) { fprintf(stderr, "%s: reply null\n", what); exit(1); }
+    if (!r) { fprintf(stderr, "%s: reply null\n", what); return -1; }
     if (r->type != REDIS_REPLY_NIL) {
         fprintf(stderr, "%s: expect nil, got type=%d\n", what, r->type);
         freeReplyObject(r);
-        exit(1);
+        return -1;
     }
     freeReplyObject(r);
 }
@@ -274,7 +275,9 @@ int main(int argc, char **argv) {
     redisContext *c = redisConnect(host, port);
     if (!c || c->err) die(c, "connect failed");
 
-    printf("Connected to %s:%d\n", host, port);
+    redisReply *reply = redisCommand(c, "MEMPRINT");
+
+    printf("Connected to %s:%d, %lld\n", host, port, reply->integer);
 
     if(mode == 0){
         save(c);

test/test_client.c

@@ -1,220 +0,0 @@
-
-#include "test_client.h"
-#include <stdio.h>
-#include <errno.h>
-#include <unistd.h>
-
-int kvs_need(const uint8_t *p, const uint8_t *end, size_t n) {
-	return (p + n <= end) ? 0 : -1;
-}
-
-// 注意u8类型不需要ntoh或者hton
-int kvs_read_u8(const uint8_t **pp, const uint8_t *end, uint8_t *out) {
-	const uint8_t *p = *pp;
-	if (kvs_need(p, end, 1) < 0) return -1;
-	*out = *p;
-	*pp = p + 1;
-	return 0;
-}
-
-int kvs_read_u16(const uint8_t **pp, const uint8_t *end, uint16_t *out) {
-	const uint8_t *p = *pp;
-	if (kvs_need(p, end, 2) < 0) return -1;
-	uint16_t v;
-	memcpy(&v, p, 2);
-	*out = ntohs(v);
-	*pp = p + 2;
-	return 0;
-}
-
-int kvs_read_u32(const uint8_t **pp, const uint8_t *end, uint32_t *out) {
-	const uint8_t *p = *pp;
-	if (kvs_need(p, end, 4) < 0) return -1;
-	uint32_t v;
-	memcpy(&v, p, 4);
-	*out = ntohl(v);
-	*pp = p + 4;
-	return 0;
-}
-
-int kvs_write_u8(uint8_t **pp, const uint8_t *end, uint8_t v) {
-	uint8_t *p = *pp;
-	if (kvs_need(p, end, 1) < 0) return -1;
-	*p = v;
-	*pp = p + 1;
-	return 0;
-}
-
-int kvs_write_u16(uint8_t **pp, const uint8_t *end, uint16_t v) {
-	uint8_t *p = *pp;
-	if (kvs_need(p, end, 2) < 0) return -1;
-	uint16_t be = htons(v);
-	memcpy(p, &be, 2);
-	*pp = p + 2;
-	return 0;
-}
-
-int kvs_write_u32(uint8_t **pp, const uint8_t *end, uint32_t v) {
-	uint8_t *p = *pp;
-	if (kvs_need(p, end, 4) < 0) return -1;
-	uint32_t be = htonl(v);
-	memcpy(p, &be, 4);
-	*pp = p + 4;
-	return 0;
-}
-
-
-
-int getcmd(uint8_t op, const char *key, uint32_t key_len, const char *value, uint32_t value_len, uint8_t *buf){
-	if(!buf) return -1;
-	uint8_t *end = buf + CMD_SIZE;
-	uint8_t *p = buf;
-	uint8_t argc = (key == NULL)?0:1;
-	argc += (value == NULL)?0:1;
-
-
-	if (kvs_write_u8(&p, end, op) < 0) return -1;
-	if (kvs_write_u8(&p, end, argc) < 0) return -1;
-	
-
-	// 写入 key
-	if(key){
-		int keylen = key_len;
-		if (kvs_write_u32(&p, end, keylen) < 0) return -1;
-		if (kvs_need(p, end, keylen) < 0) return -1;
-		if (keylen > 0) {
-			memcpy(p, key, keylen);
-			p += keylen;
-		}
-	}
-
-	if(value){
-		int vallen = value_len;
-		if (kvs_write_u32(&p, end, vallen) < 0) return -1;
-		if (kvs_need(p, end, vallen) < 0) return -1;
-		if (vallen > 0) {
-			memcpy(p, value, vallen);
-			p += vallen;
-		}
-	}
-
-	return (p - buf);
-}
-
-int parse_response(const uint8_t *buf, int buflen, kvs_response_t *rsp) {
-	if(buflen == 0) return 0;
-	const uint8_t *p = buf;
-	const uint8_t *end = buf + buflen;
-	
-	// 读取 OP
-	if (kvs_read_u8(&p, end, &rsp->op) < 0) {
-		fprintf(stderr, "Failed to read op\n");
-		return -1;
-	}
-	
-	// 读取 status
-	if (kvs_read_u8(&p, end, &rsp->status) < 0) {
-		fprintf(stderr, "Failed to read status\n");
-		return -1;
-	}
-	
-	// 读取 datalen
-	if (kvs_read_u32(&p, end, &rsp->datalen) < 0) {
-		fprintf(stderr, "Failed to read datalen\n");
-		return -1;
-	}
-	
-	// 检查数据长度
-	if (kvs_need(p, end, rsp->datalen) < 0) {
-		fprintf(stderr, "Data length mismatch: expected %u bytes, but only %ld available\n", 
-				rsp->datalen, end - p);
-		return -1;
-	}
-	
-	// 指向数据部分
-	rsp->data = (uint8_t *)p;
-	
-	return (p - buf) + rsp->datalen;
-}
-
-
-void print_response(const char *cmd_name, const kvs_response_t *rsp) {
-	printf("%s ", cmd_name);
-	if(rsp->op == KVS_CMD_GET || rsp->op == KVS_CMD_HGET || rsp->op == KVS_CMD_RGET){
-		if (rsp->datalen > 0 && rsp->data != NULL) {
-			printf("Data: ");
-			// 尝试以字符串形式打印（如果是可打印字符）
-			int is_printable = 1;
-			for (uint32_t i = 0; i < rsp->datalen; i++) {
-				if (rsp->data[i] < 32 || rsp->data[i] > 126) {
-					is_printable = 0;
-					break;
-				}
-			}
-			
-			if (is_printable) {
-				printf("\"");
-				for (uint32_t i = 0; i < rsp->datalen; i++) {
-					printf("%c", rsp->data[i]);
-				}
-				printf("\"\n");
-			} else {
-				// 以十六进制打印
-				printf("0x");
-				for (uint32_t i = 0; i < rsp->datalen; i++) {
-					printf("%02x", rsp->data[i]);
-				}
-				printf("\n");
-			}
-		} else {
-			printf("Data: (empty)\n");
-		}
-	}else {
-		switch (rsp->status) {
-			case KVS_STATUS_OK:
-				printf("(OK)\n");
-				break;
-			case KVS_STATUS_ERROR:
-				printf("(ERROR)\n");
-				break;
-			case KVS_STATUS_NO_EXIST:
-				printf("(NO_EXIST)\n");
-				break;
-			case KVS_STATUS_EXIST:
-				printf("(EXISTS)\n");
-				break;
-			default:
-				printf("(UNKNOWN)\n");
-				break;
-		}
-	}
-}
-
-int verify_response(const kvs_response_t *rsp, uint8_t expected_op, 
-                    uint8_t expected_status, const char *expected_data, uint32_t expected_len) {
-	if (rsp->op != expected_op) {
-		printf("❌ OP mismatch: expected %u, got %u\n", expected_op, rsp->op);
-		return 0;
-	}
-	
-	if (rsp->status != expected_status) {
-		printf("❌ Status mismatch: expected %u, got %u\n", expected_status, rsp->status);
-		return 0;
-	}
-	
-	if (expected_data != NULL) {
-		if (rsp->datalen != expected_len) {
-			printf("❌ Data length mismatch: expected %u, got %u\n", expected_len, rsp->datalen);
-			return 0;
-		}
-		
-		if (memcmp(rsp->data, expected_data, expected_len) != 0) {
-			printf("❌ Data content mismatch\n");
-			return 0;
-		}
-	}
-	
-	return 1;
-}
-
-

test/test_client.h

@@ -1,161 +0,0 @@
-
- /**
- * Request
- * Cmd: 	| OP(1) | argc(1) | repeat { arglen(4) | arg } |
- * 
- * Response
- * Rsp:		| OP(1) | status(1) | datalen(4) | data | 
- */
-#include <stdio.h>
-#include <stdlib.h>
-#include <stdint.h>
-#include <string.h>
-#include <sys/time.h>
-#include <sys/socket.h>
-#include <arpa/inet.h>
-
-#define CMD_SIZE (1024)
-#define BATCH_SIZE (65536)
-#define KVS_BATCH_MAX 128
-#define TIME_SUB_MS(tv1, tv2)  ((tv1.tv_sec - tv2.tv_sec) * 1000 + (tv1.tv_usec - tv2.tv_usec) / 1000)
-
-// #define PRESP  print_response
-#define PRESP
-
-
-typedef enum {
-	KVS_STATUS_OK = 0,
-	KVS_STATUS_ERROR = 1,
-	KVS_STATUS_NO_EXIST = 2,
-	KVS_STATUS_EXIST = 3,
-	KVS_STATUS_BADREQ = 4
-}rsp_ret_status_e;
-
-enum {
-	KVS_CMD_START = 0,
-	// array
-	KVS_CMD_SET = KVS_CMD_START,
-	KVS_CMD_GET,
-	KVS_CMD_DEL,
-	KVS_CMD_MOD,
-	KVS_CMD_EXIST,
-	// rbtree
-	KVS_CMD_RSET,
-	KVS_CMD_RGET,
-	KVS_CMD_RDEL,
-	KVS_CMD_RMOD,
-	KVS_CMD_REXIST,
-	// hash
-	KVS_CMD_HSET,
-	KVS_CMD_HGET,
-	KVS_CMD_HDEL,
-	KVS_CMD_HMOD,
-	KVS_CMD_HEXIST,
-
-	KVS_CMD_SSYNC,
-	KVS_CMD_SAVE,
-    
-	KVS_CMD_COUNT,
-};
-
-typedef struct {
-	uint8_t op;
-	uint8_t status;
-	uint32_t datalen;
-	uint8_t *data;
-} kvs_response_t;
-
-int kvs_need(const uint8_t *p, const uint8_t *end, size_t n);
-int kvs_read_u8(const uint8_t **pp, const uint8_t *end, uint8_t *out);
-int kvs_read_u16(const uint8_t **pp, const uint8_t *end, uint16_t *out);
-int kvs_read_u32(const uint8_t **pp, const uint8_t *end, uint32_t *out);
-
-int kvs_write_u8(uint8_t **pp, const uint8_t *end, uint8_t v);
-int kvs_write_u16(uint8_t **pp, const uint8_t *end, uint16_t v);
-int kvs_write_u32(uint8_t **pp, const uint8_t *end, uint32_t v);
-
-int getcmd(uint8_t op, const char *key, uint32_t key_len, const char *value, uint32_t value_len, uint8_t *buf);
-
-int parse_response(const uint8_t *buf, int buflen, kvs_response_t *rsp);
-void print_response(const char *cmd_name, const kvs_response_t *rsp);
-int verify_response(const kvs_response_t *rsp, uint8_t expected_op, 
-                    uint8_t expected_status, const char *expected_data, uint32_t expected_len);
-                
-
-
-
-
-typedef struct {
-    uint8_t  buf[BATCH_SIZE];
-    int      len;                 // 当前已写入的 batch 字节数
-    int      cnt;                 // 当前 batch 里命令条数
-    int      cmd_len[KVS_BATCH_MAX];
-} kvs_batch_t;
-
-static void kvs_batch_init(kvs_batch_t *b)
-{
-    b->len = 0;
-    b->cnt = 0;
-    memset(b->cmd_len, 0, sizeof(b->cmd_len));
-}
-
-/**
- * 用 getcmd() 生成单条命令，然后 append 到 batch buffer
- * 返回：0 成功，-1 失败（太多条 or buffer 不够）
- */
-static int kvs_batch_add(kvs_batch_t *b, uint8_t op, const char *key, uint32_t key_len, const char *value, uint32_t value_len){
-    if (b->cnt >= KVS_BATCH_MAX) return -1;
-    uint8_t tmp[CMD_SIZE];
-    int n = getcmd(op, key, key_len, value, value_len, tmp);   // 你提供的函数
-    if (n <= 0) return -1;
-
-    if (b->len + n > (int)sizeof(b->buf)) return -1;
-
-    memcpy(b->buf + b->len, tmp, n);
-    b->cmd_len[b->cnt] = n;
-    b->cnt++;
-    b->len += n;
-    return 0;
-}
-
-/**
- * 一次性发送 batch
- * 返回：发送字节数，<0 表示失败
- */
-static int kvs_batch_send(int fd, const kvs_batch_t *b)
-{
-    // printf("send : %d\n", b->len);
-    return (int)send(fd, b->buf, b->len, 0);
-}
-
-/**
- * 一次 recv 收回所有响应，然后批量解析为 rsp 数组
- *
- * 返回：成功解析出的响应条数（期望是 b->cnt）
- */
-static int kvs_batch_recv_parse(int fd,
-                                const kvs_batch_t *b,
-                                kvs_response_t *rsps,   // 输出数组，长度 >= b->cnt
-                                uint8_t *recvbuf,
-                                int recvbuf_cap)
-{
-    int parsed = 0;
-    int used = 0;
-
-    while(parsed < b->cnt){
-        int nrecv = (int)recv(fd, recvbuf+used, recvbuf_cap, 0);
-        if (nrecv <= 0) return -1;
-
-        int off = 0;
-
-        while (parsed < b->cnt) {
-            int consumed = parse_response(recvbuf + used, nrecv - off, &rsps[parsed]);
-            if (consumed <= 0) break;  // 不够解析/失败，简单处理：直接退出
-
-            off += consumed;
-            used+= consumed;
-            parsed++;
-        }
-    }
-    return parsed;
-}

test/testcase.c

@@ -1,539 +0,0 @@
-
-#include "test_client.h"
-#include <arpa/inet.h>
-
-
-int connect_tcpserver(const char *ip, unsigned short port) {
-
-	int connfd = socket(AF_INET, SOCK_STREAM, 0);
-
-	struct sockaddr_in server_addr;
-	memset(&server_addr, 0, sizeof(struct sockaddr_in));
-
-	server_addr.sin_family = AF_INET;
-	server_addr.sin_addr.s_addr = inet_addr(ip);
-	server_addr.sin_port = htons(port);
-
-	if (0 !=  connect(connfd, (struct sockaddr*)&server_addr, sizeof(struct sockaddr_in))) {
-		perror("connect");
-		return -1;
-	}
-	
-	return connfd;
-	
-}
-
-
-int send_msg(int connfd, char *msg, int length) {
-
-	int res = send(connfd, msg, length, 0);
-	if (res < 0) {
-		perror("send");
-		exit(1);
-	}
-	return res;
-}
-
-int recv_msg(int connfd, char *msg, int length) {
-
-	int res = recv(connfd, msg, length, 0);
-	if (res < 0) {
-		perror("recv");
-		exit(1);
-	}
-	return res;
-
-}
-
-void testcase(int connfd, uint8_t op, const char* key, uint32_t key_len, const char* value, 
-	uint32_t value_len, rsp_ret_status_e st, const char* rsp_value, uint32_t expect_len, const char* command_name){
-	uint8_t buf[CMD_SIZE];
-	uint8_t result[CMD_SIZE];
-	kvs_response_t rsp;
-	int len, recv_len;
-
-	len = getcmd(op, key, key_len, value, value_len, buf);
-	send_msg(connfd, buf, len);
-	recv_len = recv_msg(connfd, result, CMD_SIZE);
-	if (parse_response(result, recv_len, &rsp) > 0) {
-		PRESP(command_name, &rsp);
-		if(!verify_response(&rsp, op, st, rsp_value, expect_len)) printf("%s\n", command_name);
-	}else{
-		printf("parser error\n");
-	}
-
-	return ;
-}
-
-
-void array_testcase_1w(int connfd) {
-
-	int count = 1000;
-	int i = 0;
-
-	struct timeval tv_begin;
-	gettimeofday(&tv_begin, NULL);
-
-	for (i = 0;i < count;i ++) {
-		testcase(connfd, KVS_CMD_SET, "name", 4, "l\r\0n", 4, KVS_STATUS_OK, NULL, 0, "SET NAME");
-		testcase(connfd, KVS_CMD_GET, "name", 4, NULL, 0, KVS_STATUS_OK, "l\r\0n", 4, "GET NAME");
-		testcase(connfd, KVS_CMD_MOD, "name", 4, "liu", 3, KVS_STATUS_OK, NULL, 0, "MOD NAME");
-		testcase(connfd, KVS_CMD_GET, "name", 4, NULL, 0, KVS_STATUS_OK, "liu", 3, "GET NAME");
-		testcase(connfd, KVS_CMD_EXIST, "name", 4, NULL, 0, KVS_STATUS_EXIST, NULL, 0, "EXIST NAME");
-		testcase(connfd, KVS_CMD_DEL, "name", 4, NULL, 0, KVS_STATUS_OK, NULL, 0, "DEL NAME");
-		testcase(connfd, KVS_CMD_EXIST, "name", 4, NULL, 0, KVS_STATUS_NO_EXIST, NULL, 0, "EXIST NAME");
-	
-		testcase(connfd, KVS_CMD_MOD, "stu", 3, "liu", 3, KVS_STATUS_NO_EXIST, NULL, 0, "MOD NAME");
-		testcase(connfd, KVS_CMD_DEL, "stu", 3, NULL, 0, KVS_STATUS_NO_EXIST, NULL, 0, "DEL SUT");
-	}
-
-	struct timeval tv_end;
-	gettimeofday(&tv_end, NULL);
-
-	int time_used = TIME_SUB_MS(tv_end, tv_begin); // ms
-
-	printf("array testcase --> time_used: %d, qps: %d\n", time_used, 9000 * 1000 / time_used);
-
-}
-
-void rbtree_testcase_1w(int connfd) {
-
-	int count = 1000;
-	int i = 0;
-
-	struct timeval tv_begin;
-	gettimeofday(&tv_begin, NULL);
-
-	for (i = 0;i < count;i ++) {
-		testcase(connfd, KVS_CMD_RSET, "name", 4, "l\r\0n", 4, KVS_STATUS_OK, NULL, 0, "RSET NAME");
-		testcase(connfd, KVS_CMD_RGET, "name", 4, NULL, 0, KVS_STATUS_OK, "l\r\0n", 4, "RGET NAME");
-		testcase(connfd, KVS_CMD_RMOD, "name", 4, "liu", 3, KVS_STATUS_OK, NULL, 0, "RMOD NAME");
-		testcase(connfd, KVS_CMD_RGET, "name", 4, NULL, 0, KVS_STATUS_OK, "liu", 3, "RGET NAME");
-		testcase(connfd, KVS_CMD_REXIST, "name", 4, NULL, 0, KVS_STATUS_EXIST, NULL, 0, "REXIST NAME");
-		testcase(connfd, KVS_CMD_RDEL, "name", 4, NULL, 0, KVS_STATUS_OK, NULL, 0, "RDEL NAME");
-		testcase(connfd, KVS_CMD_REXIST, "name", 4, NULL, 0, KVS_STATUS_NO_EXIST, NULL, 0, "REXIST NAME");
-	
-		testcase(connfd, KVS_CMD_RMOD, "stu", 3, "liu", 3, KVS_STATUS_NO_EXIST, NULL, 0, "RMOD NAME");
-		testcase(connfd, KVS_CMD_RDEL, "stu", 3, NULL, 0, KVS_STATUS_NO_EXIST, NULL, 0, "RDEL SUT");
-	}
-
-	struct timeval tv_end;
-	gettimeofday(&tv_end, NULL);
-
-	int time_used = TIME_SUB_MS(tv_end, tv_begin); // ms
-
-	printf("array testcase --> time_used: %d, qps: %d\n", time_used, 9000 * 1000 / time_used);
-
-}
-
-void hash_testcase_1w(int connfd) {
-
-	int count = 1000;
-	int i = 0;
-
-	struct timeval tv_begin;
-	gettimeofday(&tv_begin, NULL);
-
-	for (i = 0;i < count;i ++) {
-		testcase(connfd, KVS_CMD_HSET, "name", 4, "l\r\0n", 4, KVS_STATUS_OK, NULL, 0, "HSET NAME");
-		testcase(connfd, KVS_CMD_HGET, "name", 4, NULL, 0, KVS_STATUS_OK, "l\r\0n", 4, "HGET NAME");
-		testcase(connfd, KVS_CMD_HMOD, "name", 4, "liu", 3, KVS_STATUS_OK, NULL, 0, "HMOD NAME");
-		testcase(connfd, KVS_CMD_HGET, "name", 4, NULL, 0, KVS_STATUS_OK, "liu", 3, "HGET NAME");
-		testcase(connfd, KVS_CMD_HEXIST, "name", 4, NULL, 0, KVS_STATUS_EXIST, NULL, 0, "HEXIST NAME");
-		testcase(connfd, KVS_CMD_HDEL, "name", 4, NULL, 0, KVS_STATUS_OK, NULL, 0, "HDEL NAME");
-		testcase(connfd, KVS_CMD_HEXIST, "name", 4, NULL, 0, KVS_STATUS_NO_EXIST, NULL, 0, "HEXIST NAME");
-	
-		testcase(connfd, KVS_CMD_HMOD, "stu", 3, "liu", 3, KVS_STATUS_NO_EXIST, NULL, 0, "HMOD NAME");
-		testcase(connfd, KVS_CMD_HDEL, "stu", 3, NULL, 0, KVS_STATUS_NO_EXIST, NULL, 0, "HDEL SUT");
-	}
-
-	struct timeval tv_end;
-	gettimeofday(&tv_end, NULL);
-
-	int time_used = TIME_SUB_MS(tv_end, tv_begin); // ms
-
-	printf("array testcase --> time_used: %d, qps: %d\n", time_used, 9*count * 1000 / time_used);
-
-}
-
-void do_batch_test(int fd, int op, const char *key, const char *value, rsp_ret_status_e st, const char *rsp_value){
-    kvs_batch_t batch;
-    kvs_batch_init(&batch);
-
-	char bkey[256]={0}, bval[256]={0};
-
-    // 组 batch（最多 64 条）
-	for(int i = 0;i < 100; ++ i){
-		if(value == NULL){
-			int klen = sprintf(bkey, "%s%d", key, i);
-	    	kvs_batch_add(&batch, op, bkey, klen, NULL, 0);
-		}else{
-			int klen = sprintf(bkey, "%s%d", key, i);
-			int vlen = sprintf(bval, "%s%d", value, i);
-			kvs_batch_add(&batch, op, bkey, klen, bval, vlen);
-		}
-	}
-
-    // 一次性发送
-    kvs_batch_send(fd, &batch);
-
-    // 一次性 recv + parse
-    uint8_t recvbuf[BATCH_SIZE];
-    kvs_response_t rsps[KVS_BATCH_MAX];
-
-    int nrsp = kvs_batch_recv_parse(fd, &batch, rsps, recvbuf, sizeof(recvbuf));
-
-    // 打印/处理
-	for (int i = 0; i < nrsp; i++) {
-		print_response(bkey, &rsps[i]);
-		int vlen;
-		if(rsp_value != NULL) vlen = sprintf(bval, "%s%d", rsp_value, i);
-		else vlen = 0;
-		verify_response(&rsps[i], op, st, bval, vlen);
-	}
-}
-
-void array_testcase_1w_batch(int connfd) {
-    kvs_batch_t batch;
-	kvs_batch_init(&batch);
-
-	int count = 1000;
-	int i = 0;
-
-	struct timeval tv_begin;
-	gettimeofday(&tv_begin, NULL);
-
-	for (i = 0;i < count;i ++) {
-		batch.cnt = 0;
-		batch.len = 0;
-		kvs_batch_add(&batch, KVS_CMD_SET, "name", 4, "l\r\0n", 4);
-		kvs_batch_add(&batch, KVS_CMD_GET, "name", 4, NULL, 0);
-		kvs_batch_add(&batch, KVS_CMD_MOD, "name", 4, "liu", 3);
-		kvs_batch_add(&batch, KVS_CMD_GET, "name", 4, NULL, 0);
-		kvs_batch_add(&batch, KVS_CMD_EXIST, "name", 4, NULL, 0);
-		kvs_batch_add(&batch, KVS_CMD_DEL, "name", 4, NULL, 0);
-		kvs_batch_add(&batch, KVS_CMD_EXIST, "name", 4, NULL, 0);
-	
-		kvs_batch_add(&batch, KVS_CMD_MOD, "stu", 3, "liu", 3);
-		kvs_batch_add(&batch, KVS_CMD_DEL, "stu", 3, NULL, 0);
-
-		kvs_batch_send(connfd, &batch);
-
-		uint8_t recvbuf[BATCH_SIZE];
-		kvs_response_t rsps[KVS_BATCH_MAX];
-		int nrsp = kvs_batch_recv_parse(connfd, &batch, rsps, recvbuf, sizeof(recvbuf));
-	}
-
-	struct timeval tv_end;
-	gettimeofday(&tv_end, NULL);
-
-	int time_used = TIME_SUB_MS(tv_end, tv_begin); // ms
-
-	printf("array testcase --> time_used: %d, qps: %d\n", time_used, 9000 * 1000 / time_used);
-
-}
-
-void batch_qps(int connfd) {
-    const int N = 1000000;
-    const int B = 100;                 // do_batch_test() 里写死 50
-	static char valA[256];
-    static char valB[256];
-    static char valC[256];
-
-    static int inited = 0;
-    if (!inited) {
-        // 填充 255 字节，最后补 '\0'
-        memset(valA, 'A', 255); valA[255] = '\0';
-        memset(valB, 'B', 255); valB[255] = '\0';
-        memset(valC, 'C', 255); valC[255] = '\0';
-        inited = 1;
-    }
-
-    struct timeval tv_begin, tv_end;
-    gettimeofday(&tv_begin, NULL);
-
-    // ---------------- Phase 1: ADD 两条 DEL 一条 (100w) ----------------
-    // 每轮：ADD/SET A_i, ADD/SET B_i, DEL A_i
-    // 用 batch：每批处理 idx..idx+49
-    for (int base = 0; base < N; base += B) {
-        // prefix 必须短，避免 do_batch_test 里 bkey[15] 溢出
-        // do_batch_test 会生成: prefix + i (0..49)
-        // 我们把 base 编进 prefix，保证全局 key 唯一
-        char preA[16], preB[16];
-
-        // 例：A123450_0..A123450_49（base=123450）
-        // 注意：这里 prefix 长度要尽量 <= 10 左右
-        snprintf(preA, sizeof(preA), "A%d_", base/100);
-        snprintf(preB, sizeof(preB), "B%d_", base/100);
-
-        do_batch_test(connfd, KVS_CMD_RSET, preA, valA, KVS_STATUS_OK, NULL); // 50次 RSET A
-        do_batch_test(connfd, KVS_CMD_RSET, preB, valB, KVS_STATUS_OK, NULL); // 50次 RSET B
-        do_batch_test(connfd, KVS_CMD_RDEL, preA, NULL, KVS_STATUS_OK, NULL); // 50次 RDEL A
-
-        if (base % 10000 == 0) printf("P1 base:%d\n", base);
-    }
-    printf("phase 1 end\n");
-
-    // ---------------- Phase 2: ADD 一条 DEL 两条 (100w) ----------------
-    // 每轮：ADD/SET C_i, DEL B_i, DEL C_i
-    for (int base = 0; base < N; base += B) {
-        char preB[16], preC[16];
-        snprintf(preB, sizeof(preB), "B%d_", base/100);
-        snprintf(preC, sizeof(preC), "C%d_", base/100);
-
-        do_batch_test(connfd, KVS_CMD_RSET, preC, valC, KVS_STATUS_OK, NULL); // 50次 RSET C
-        do_batch_test(connfd, KVS_CMD_RDEL, preB, NULL, KVS_STATUS_OK, NULL); // 50次 RDEL B
-        do_batch_test(connfd, KVS_CMD_RDEL, preC, NULL, KVS_STATUS_OK, NULL); // 50次 RDEL C
-
-        if (base % 10000 == 0) printf("P2 base:%d\n", base);
-    }
-    printf("phase 2 end\n");
-
-    gettimeofday(&tv_end, NULL);
-    int time_used = TIME_SUB_MS(tv_end, tv_begin);
-
-    // 真实总 ops 还是 6*N （每轮 6 个操作）
-    long long ops = (long long)N * 6;
-    long long qps = (time_used > 0) ? (ops * 1000 / time_used) : 0;
-
-    printf("BATCH(do_batch_test) ADD2-DEL1 then ADD1-DEL2 (N=%d) --> time_used=%d ms, ops=%lld, qps=%lld\n",
-           N, time_used, ops, qps);
-}
-
-void save(int connfd){
-	testcase(connfd, KVS_CMD_SAVE, NULL, 0, NULL, 0, KVS_STATUS_OK, NULL, 0, "SAVE");
-}
-
-void testcase_add2_del1_then_add1_del2_100w(int connfd) {
-    const int N = 1000000;
-
-    // 如果你有 KVS_CMD_ADD 就用 ADD；没有就用 SET（但 SET 会覆盖，意义不同）
-    // 这里按你说的“会返回 EXIST”，所以我用 KVS_CMD_ADD 来写。
-    // 若你实际命令叫 KVS_CMD_SET 且语义是 ADD，请自行替换。
-    const char *valA = "va";
-    const char *valB = "vb";
-    const char *valC = "vc";
-
-    char keyA[64], keyB[64], keyC[64];
-
-    struct timeval tv_begin, tv_end;
-    gettimeofday(&tv_begin, NULL);
-
-    // ---------------- Phase 1: ADD 两条 DEL 一条 (100w) ----------------
-    // 每轮：ADD A_i, ADD B_i, DEL A_i -> 留下 B_i
-    for (int i = 0; i < N; i++) {
-        int klenA = snprintf(keyA, sizeof(keyA), "A_%d", i);
-        int klenB = snprintf(keyB, sizeof(keyB), "B_%d", i);
-
-        testcase(connfd, KVS_CMD_RSET, keyA, klenA, valA, 2, KVS_STATUS_OK, NULL, 0, "P1 ADD A_i");
-        testcase(connfd, KVS_CMD_RSET, keyB, klenB, valB, 2, KVS_STATUS_OK, NULL, 0, "P1 ADD B_i");
-        testcase(connfd, KVS_CMD_RDEL, keyA, klenA, NULL, 0, KVS_STATUS_OK, NULL, 0, "P1 DEL A_i");
-	
-		if(i%10000 == 0) printf("i:%d\n", i);
-	}
-
-	printf("phase 1 end\n");
-
-    // ---------------- Phase 2: ADD 一条 DEL 两条 (100w) ----------------
-    // 每轮：ADD C_i, DEL B_i, DEL C_i -> 每轮净删一个 B_i
-    for (int i = 0; i < N; i++) {
-        int klenC = snprintf(keyC, sizeof(keyC), "C_%d", i);
-        int klenB = snprintf(keyB, sizeof(keyB), "B_%d", i);
-
-        testcase(connfd, KVS_CMD_RSET, keyC, klenC, valC, 2, KVS_STATUS_OK, NULL, 0, "P2 ADD C_i");
-        testcase(connfd, KVS_CMD_RDEL, keyB, klenB, NULL, 0, KVS_STATUS_OK, NULL, 0, "P2 DEL B_i");
-        testcase(connfd, KVS_CMD_RDEL, keyC, klenC, NULL, 0, KVS_STATUS_OK, NULL, 0, "P2 DEL C_i");
-		
-		if(i%10000 == 0) printf("i:%d\n", i);
-	}
-
-	printf("phase 2 end\n");
-
-    // for (int j = 0; j < 5; j++) {
-    //     int idx = (j == 0) ? 0 : (j == 1) ? (N/2) : (N-1);
-
-    //     int klenA = snprintf(keyA, sizeof(keyA), "A_%d", idx);
-    //     int klenB = snprintf(keyB, sizeof(keyB), "B_%d", idx);
-    //     int klenC = snprintf(keyC, sizeof(keyC), "C_%d", idx);
-
-    //     testcase(connfd, KVS_CMD_EXIST, keyA, klenA, NULL, 0, KVS_STATUS_NO_EXIST, NULL, 0, "FINAL A not exist");
-    //     testcase(connfd, KVS_CMD_EXIST, keyB, klenB, NULL, 0, KVS_STATUS_NO_EXIST, NULL, 0, "FINAL B not exist");
-    //     testcase(connfd, KVS_CMD_EXIST, keyC, klenC, NULL, 0, KVS_STATUS_NO_EXIST, NULL, 0, "FINAL C not exist");
-    // }
-
-    gettimeofday(&tv_end, NULL);
-    int time_used = TIME_SUB_MS(tv_end, tv_begin);
-
-    // 统计：Phase1 每轮3 ops，Phase2 每轮3 ops
-    long long ops = (long long)N * 3 + (long long)N * 3;
-    long long qps = (time_used > 0) ? (ops * 1000 / time_used) : 0;
-
-    printf("ADD2-DEL1 then ADD1-DEL2 (N=%d) --> time_used=%d ms, ops=%lld, qps=%lld\n",
-           N, time_used, ops, qps);
-}
-
-void send_spec_chars(int connfd){
-	const char *v_ws = "li an\tok\nend";  /* 内容含空格、\t、\n */
-	int v_ws_len = 12;                   /* 手算：l i ' ' a n \t o k \n e n d = 12 */
-
-	testcase(connfd, KVS_CMD_RSET,
-			"ws", 2,
-			v_ws, v_ws_len,
-			KVS_STATUS_OK,
-			NULL, 0,
-			"RSET WHITESPACE");
-
-	testcase(connfd, KVS_CMD_RGET,
-			"ws", 2,
-			NULL, 0,
-			KVS_STATUS_OK,
-			v_ws, v_ws_len,
-			"RGET WHITESPACE");
-
-	/* 2) 引号与反斜杠：测试是否被错误转义 */
-	const char *v_quote = "he\"llo\\world'!"; 
-	/* 字节数：h e " l l o \ w o r l d ' ! = 15 */
-	int v_quote_len = 15;
-
-	testcase(connfd, KVS_CMD_RSET,
-			"quote", 5,
-			v_quote, v_quote_len,
-			KVS_STATUS_OK,
-			NULL, 0,
-			"RSET QUOTE BACKSLASH");
-
-	testcase(connfd, KVS_CMD_RGET,
-			"quote", 5,
-			NULL, 0,
-			KVS_STATUS_OK,
-			v_quote, v_quote_len,
-			"RGET QUOTE BACKSLASH");
-
-
-	/* 3) 分隔符字符：冒号/逗号/分号/竖线 */
-	const char *v_sep = "a:b,c;d|e";
-	int v_sep_len = 9;   /* a : b , c ; d | e = 9 */
-
-	testcase(connfd, KVS_CMD_RSET,
-			"sep", 3,
-			v_sep, v_sep_len,
-			KVS_STATUS_OK,
-			NULL, 0,
-			"RSET SEPARATORS");
-
-	testcase(connfd, KVS_CMD_RGET,
-			"sep", 3,
-			NULL, 0,
-			KVS_STATUS_OK,
-			v_sep, v_sep_len,
-			"RGET SEPARATORS");
-
-	/* 4) CRLF：\r\n 组合（最容易把一条请求拆错/响应拼错） */
-	const char *v_crlf = "line1\r\nline2";
-	int v_crlf_len = 12; /* line1(5) + \r(1) + \n(1) + line2(5) = 12 */
-
-	testcase(connfd, KVS_CMD_RSET,
-			"crlf", 4,
-			v_crlf, v_crlf_len,
-			KVS_STATUS_OK,
-			NULL, 0,
-			"RSET CRLF");
-
-	testcase(connfd, KVS_CMD_RGET,
-			"crlf", 4,
-			NULL, 0,
-			KVS_STATUS_OK,
-			v_crlf, v_crlf_len,
-			"RGET CRLF");
-
-	/* 5) 二进制数据：包含 \0，必须按长度处理，不能用 strlen */
-	char v_bin[] = { 'A', 0x00, 'B', 'C', 0x00, 'D' };
-	int v_bin_len = 6;
-
-	testcase(connfd, KVS_CMD_RSET,
-			"bin", 3,
-			v_bin, v_bin_len,
-			KVS_STATUS_OK,
-			NULL, 0,
-			"RSET BINARY WITH NUL");
-
-	testcase(connfd, KVS_CMD_RGET,
-			"bin", 3,
-			NULL, 0,
-			KVS_STATUS_OK,
-			v_bin, v_bin_len,
-			"RGET BINARY WITH NUL");
-
-	/* 6) UTF-8：中文 + emoji（测试多字节） */
-	const char *v_utf8 = "中文🙂"; 
-	/* "中"(3字节) + "文"(3字节) + "🙂"(4字节) = 10字节 */
-	int v_utf8_len = 10;
-
-	testcase(connfd, KVS_CMD_RSET,
-			"utf8", 4,
-			v_utf8, v_utf8_len,
-			KVS_STATUS_OK,
-			NULL, 0,
-			"RSET UTF8");
-
-	testcase(connfd, KVS_CMD_RGET,
-			"utf8", 4,
-			NULL, 0,
-			KVS_STATUS_OK,
-			v_utf8, v_utf8_len,
-			"RGET UTF8");
-
-}
-
-int main(int argc, char *argv[]) {
-	if (argc != 4) {
-		printf("arg error\n");
-		return -1;
-	}
-
-	char *ip = argv[1];
-	int port = atoi(argv[2]);
-	int mode = atoi(argv[3]);
-
-	int connfd = connect_tcpserver(ip, port);
-
-	if(mode == 0){
-		array_testcase_1w(connfd);
-	}else if(mode == 1){
-		array_testcase_1w_batch(connfd);
-	}else if(mode == 2){
-		rbtree_testcase_1w(connfd);
-	}else if(mode == 3){
-		hash_testcase_1w(connfd);
-	}else if(mode == 4){
-		testcase_add2_del1_then_add1_del2_100w(connfd);
-	}else if(mode == 10){
-		do_batch_test(connfd, KVS_CMD_SET, "array_set", "array_val", KVS_STATUS_OK, NULL);
-	}else if(mode == 11){
-		do_batch_test(connfd, KVS_CMD_GET, "array_set", NULL, KVS_STATUS_OK, "array_val");
-	}else if(mode == 12){
-		do_batch_test(connfd, KVS_CMD_EXIST, "array_set", NULL, KVS_STATUS_EXIST, NULL);
-	}else if(mode == 13){
-		do_batch_test(connfd, KVS_CMD_DEL, "array_set", NULL, KVS_STATUS_OK, NULL);
-	}else if(mode == 20){
-		do_batch_test(connfd, KVS_CMD_RSET, "rbtree_set", "rbtree_val", KVS_STATUS_OK, NULL);
-	}else if(mode == 21){
-		do_batch_test(connfd, KVS_CMD_RGET, "rbtree_set", NULL, KVS_STATUS_OK, "rbtree_val");
-	}else if(mode == 22){
-		do_batch_test(connfd, KVS_CMD_REXIST, "rbtree_set", NULL, KVS_STATUS_OK, NULL);
-	}else if(mode == 30){
-		do_batch_test(connfd, KVS_CMD_HSET, "hash_set", "hash_val", KVS_STATUS_OK, NULL);
-	}else if(mode == 31){
-		do_batch_test(connfd, KVS_CMD_HGET, "hash_set", NULL, KVS_STATUS_OK, "hash_val");
-	}else if(mode == 32){
-		do_batch_test(connfd, KVS_CMD_HEXIST, "hash_set", NULL, KVS_STATUS_OK, NULL);
-	}else if(mode == -1){
-		save(connfd);
-	}else if(mode == 5){
-		batch_qps(connfd);
-	}else if(mode == 6){
-		send_spec_chars(connfd);
-	}
-	
-	return 0;
-}