性能测试用例和结果更新

.gitignore

@@ -6,5 +6,6 @@
 *.so
 
 codex/
+results/*
 tests/bin
 zvfs_daemon

README.md

@@ -11,6 +11,35 @@ ZVFS 是一个 **透明用户态文件系统原型**，通过 `LD_PRELOAD` 劫
 
 ---
 
+# 测试方案
+
+```shell
+git clone http://gitlab.0voice.com/lianyiheng/zvfs.git
+cd zvfs
+git submodule update --init --recursive
+
+cd spdk
+sudo ./scripts/pkgdep.sh
+sudo ./configure --with-shared
+sudo make -j
+sudo ./scripts/setup.sh
+
+cd ..
+make && make test
+
+su root
+
+# 测试
+mkdir -p /zvfs/zvfs_sys_fio
+mkdir -p /tmp/zvfs_sys_fio 
+
+SPDK_JSON_CONFIG="$PWD/src/zvfsnvme.json" ./src/daemon/zvfs_daemon
+ZVFS_LD_PRELOAD_VALUE="$PWD/src/libzvfs.so" ./scripts/run_fio_matrix.sh
+
+# 结果：
+results/fio/..../summary.md
+```
+
 # 设计思路
 
 大多数用户态文件系统（如 FUSE）需要修改应用或挂载文件系统。用户态文件系统如果要通过VFS，需要多一到两次额外的用户态/内核态切换。ZVFS 的目标是对应用完全透明：应用按正常方式调用 POSIX API，底层存储路径被悄悄替换掉。
@@ -135,54 +164,89 @@ fcntl fcntl64 ioctl
 > 注：VMware 模拟 NVMe 无法体现 SPDK 轮询模式对中断驱动 I/O 的延迟优势，
 > 以下数据用于评估 hook 层与 IPC 的额外开销，不代表真实硬件上的性能对比。
 
-### 顺序写吞吐
+### fio 4K（psync，30s）
 
-| Block Size | spdk_nvme_perf | ZVFS |
-|---|---|---|
-| 4K | 100 MiB/s | 94 MiB/s |
-| 128K | 1843 MiB/s | 1662 MiB/s |
+测试口径：
 
-ZVFS 达到 **SPDK 原生性能约 90%**。
+- `ioengine=psync`
+- `direct=1`
+- `iodepth=1`
+- `bs=4K`
+- `time_based=1`
+- `runtime=30`
+- `size=512M`
+- `sys`: 普通文件路径
+- `zvfs`: `LD_PRELOAD=./src/libzvfs.so`
 
----
+#### prepare_fill 顺序写带宽
 
-### fio 随机写（16K，psync）
+| | sys | ZVFS |
+|---|---:|---:|
+| 带宽 | 10.92 MiB/s | 14.41 MiB/s |
+| disk util | 99.68% | 5.49% |
 
-| | kernel (psync) | ZVFS |
-|---|---|---|
-| IOPS | 1855 | 1353 |
-| 吞吐 | 28.0 MiB/s | 21.2 MiB/s |
-| avg clat | 492 µs | 692 µs |
-| sys% | 28.6% | 8.4% |
+#### randread_4k
 
-> 当前 ZVFS 在该单线程 `psync` 随机写场景下达到 kernel `psync` 的约 73% IOPS。daemon 内部 `SPDK + reply_q` 已收敛到较稳定范围，剩余主要开销集中在 `client -> daemon` 请求进入阶段。
+| | sys | ZVFS |
+|---|---:|---:|
+| IOPS | 3118.31 | 3685.21 |
+| 吞吐 | 12.18 MiB/s | 14.40 MiB/s |
+| avg clat | 318.31 µs | 268.91 µs |
+| disk util | 99.77% | 0.52% |
+
+#### randwrite_4k
+
+| | sys | ZVFS |
+|---|---:|---:|
+| IOPS | 2883.24 | 3816.78 |
+| 吞吐 | 11.26 MiB/s | 14.91 MiB/s |
+| avg clat | 344.20 µs | 259.53 µs |
+| disk util | 99.80% | 3.97% |
+
+#### randrw_4k（50/50）
+
+| | sys | ZVFS |
+|---|---:|---:|
+| 读 IOPS | 1614.29 | 2652.07 |
+| 写 IOPS | 1605.60 | 2637.78 |
+| 读 avg clat | 306.56 µs | 184.11 µs |
+| 写 avg clat | 309.72 µs | 189.44 µs |
+| disk util | 99.87% | 2.98% |
 
 ---
 
 ### WRITE 请求端到端延迟分解（单位 µs）
-基于 12 条 `WRITE` trace 样本统计，下面按调用栈层级展开平均耗时。由于四舍五入，父子项相加会有 `±1 µs` 误差。
+```bash
+sudo env \
+  ZVFS_TRACE_LATENCY=1 \
+  LD_PRELOAD="$PWD/src/libzvfs.so" \
+  fio ./zvfs_fio_test/zvfs/randwrite_4k.fio 2> /tmp/zvfs.write.trace.log
+```
+基于 `/tmp/zvfs.write.trace.log` 中 `107946` 条 `WRITE` trace 样本统计，下面按调用栈层级展开平均耗时。由于四舍五入，父子项相加会有 `±1 µs` 误差。
 
 ```text
-total 748
-├─ c2s 317
-│  ├─ send 39
-│  └─ server_rx_wait 278
-├─ server 336
-│  ├─ rx_dispatch 12
-│  ├─ dispatch_spdk 25
-│  ├─ spdk 194
-│  └─ reply_q 103
-│     ├─ spdk_post 11
-│     └─ cq_wait 91
-│        ├─ kick 13
-│        ├─ wake_sched 65
-│        └─ wake_to_tx 12
-└─ s2c 95
-   ├─ resp_wait 83
-   └─ parse 12
+total 256
+├─ c2s 41
+│  ├─ send 7
+│  └─ server_rx_wait 34
+├─ server 154
+│  ├─ rx_dispatch 0
+│  ├─ dispatch_spdk 5
+│  ├─ spdk 138
+│  │  ├─ phase1 0
+│  │  └─ phase2 138
+│  └─ reply_q 10
+│     ├─ spdk_post 0
+│     └─ cq_wait 10
+│        ├─ kick 1
+│        ├─ wake_sched 8
+│        └─ wake_to_tx 0
+└─ s2c 60
+   ├─ resp_wait 60
+   └─ parse 0
 ```
 
-当前 `WRITE` 的主要额外开销已经比较清晰：一是 `c2s / server_rx_wait`，二是 `server` 内部的 `spdk` 与 `reply_q`。在 `reply_q` 中，`wake_sched` 已明显大于 `kick` 和 `wake_to_tx`，说明回包路径的主要损耗不在 `eventfd` 写入本身，而在 reactor 被唤醒后的调度等待。
+现在一次 `WRITE` 平均大约 `256 µs`。其中最耗时的是实际存储写入（`spdk`，约 `138 µs`），其次是请求发给 daemon 和结果返回应用这两段通信等待（`c2s` + `s2c`，约 `101 µs`）。回包队列相关开销（`reply_q`）已经压到约 `10 µs`，不再是主要瓶颈。
 
 ---
 
@@ -243,3 +307,74 @@ PostgreSQL tablespace 通过 symbolic link 访问路径： pg_tblspc/xxx
 ## write 延迟显著高于预期
 
 这次 fio 延迟排查里，最初 `WRITE` 延迟明显高于预期。沿端到端路径加轻量打点后发现问题并不在 SPDK 本体，而是同时叠加了无条件 RMW、VM 中 poller 调度抖动、线程未绑核，以及后期 trace 暴露出来的 reactor 唤醒后核心切换抖动。对应处理是：整块对齐写跳过 read phase、将 reactor/md/io 线程固定到指定 CPU，并把 io 线程数和绑核目标收敛到配置项中。修复后 `dispatch_spdk` 从毫秒级降到几十微秒，`WRITE` 平均延迟也回落到约 700 µs，但剩余尾延迟仍主要表现为请求进入与回包阶段的调度等待。 
+
+---
+
+# 脚本参数
+
+以下脚本都支持通过环境变量 `ZVFS_LD_PRELOAD_VALUE` 指定加载的 so 库：
+
+- `scripts/run_fio_matrix.sh`
+- `scripts/run_pgbench_zvfs.sh`
+- `scripts/run_db_bench_zvfs.sh`
+- `scripts/run_test_hook_api.sh`
+
+示例：
+
+```bash
+sudo env ZVFS_LD_PRELOAD_VALUE="$PWD/src/libzvfs.so" ./scripts/run_fio_matrix.sh
+```
+
+其他脚本同理：
+
+```bash
+sudo env ZVFS_LD_PRELOAD_VALUE="$PWD/src/libzvfs.so" ./scripts/run_pgbench_zvfs.sh
+sudo env ZVFS_LD_PRELOAD_VALUE="$PWD/src/libzvfs.so" ./scripts/run_db_bench_zvfs.sh
+env ZVFS_LD_PRELOAD_VALUE="$PWD/src/libzvfs.so" ./scripts/run_test_hook_api.sh
+```
+
+## 延迟 Trace
+
+`WRITE` / `SYNC_MD` 的端到端阶段打印通过环境变量 `ZVFS_TRACE_LATENCY=1` 打开。  
+打印代码在客户端侧 [`src/spdk_engine/io_engine.c`](/home/lian/share/zvfs/src/spdk_engine/io_engine.c)，输出会写到执行 workload 的进程标准错误。
+
+示例：
+
+```bash
+sudo env \
+  ZVFS_TRACE_LATENCY=1 \
+  LD_PRELOAD="$PWD/src/libzvfs.so" \
+  fio ./zvfs_fio_test/zvfs/randwrite_4k.fio 2> /tmp/zvfs.write.trace.log
+```
+
+筛出 `WRITE` trace：
+
+```bash
+grep '\[zvfs\]\[trace\]\[WRITE\]' /tmp/zvfs.write.trace.log
+```
+
+筛出 `SYNC_MD` trace：
+
+```bash
+grep '\[zvfs\]\[trace\]\[SYNC_MD\]' /tmp/zvfs.write.trace.log
+```
+
+单行输出字段包括：
+
+- `total`
+- `c2s`
+- `send`
+- `server_rx_wait`
+- `rx_dispatch`
+- `dispatch_spdk`
+- `spdk`
+- `phase1`
+- `phase2`
+- `spdk_post`
+- `kick`
+- `wake_sched`
+- `wake_to_tx`
+- `reply_q`
+- `cq_wait`
+
+更新 README 中的 `WRITE 请求端到端延迟分解` 时，可对多条 `[zvfs][trace][WRITE]` 日志按字段取平均后再汇总。

scripts/run_db_bench_zvfs.sh

@@ -1,6 +1,10 @@
 #!/usr/bin/env bash
 set -euo pipefail
-env -u LD_PRELOAD=/home/lian/share/zvfs/src/libzvfs.so -rf /zvfs/rocksdb_manual || true
+
+ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+LD_PRELOAD_PATH="${ZVFS_LD_PRELOAD_VALUE:-${ROOT_DIR}/src/libzvfs.so}"
+
+env LD_PRELOAD="${LD_PRELOAD_PATH}" rm -rf /zvfs/rocksdb_manual || true
 
 # =========================
 # Manual Config (edit here)
@@ -91,7 +95,7 @@ echo "USE_MMAP_READS=$USE_MMAP_READS USE_MMAP_WRITES=$USE_MMAP_WRITES"
 echo "STATISTICS=$STATISTICS STATS_INTERVAL_SECONDS=$STATS_INTERVAL_SECONDS HISTOGRAM=$HISTOGRAM"
 echo
 
-exec env LD_PRELOAD=/home/lian/share/zvfs/src/libzvfs.so "$DB_BENCH_BIN" \
+exec env LD_PRELOAD="${LD_PRELOAD_PATH}" "$DB_BENCH_BIN" \
   --db="$DB_PATH" \
   --benchmarks="$BENCHMARKS" \
   --num="$NUM" \

scripts/run_fio_matrix.sh

@@ -0,0 +1,262 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+FIO_CASE_DIR="${ROOT_DIR}/zvfs_fio_test"
+RESULTS_ROOT="${ROOT_DIR}/results/fio"
+TIMESTAMP="$(date +%Y%m%d_%H%M%S)"
+RUN_DIR="${RESULTS_ROOT}/${TIMESTAMP}"
+
+FIO_BIN="${FIO_BIN:-$(command -v fio 2>/dev/null || true)}"
+PYTHON_BIN="${PYTHON_BIN:-$(command -v python3 2>/dev/null || true)}"
+LD_PRELOAD_PATH="${LD_PRELOAD_PATH:-${ROOT_DIR}/src/libzvfs.so}"
+ZVFS_LD_PRELOAD_VALUE="${ZVFS_LD_PRELOAD_VALUE:-${LD_PRELOAD_PATH}}"
+
+SYS_PARENT_DIR="/tmp/zvfs_sys_fio"
+ZVFS_PARENT_DIR="/zvfs/zvfs_sys_fio"
+SYS_TEST_FILE="${SYS_PARENT_DIR}/testfile"
+ZVFS_TEST_FILE="${ZVFS_PARENT_DIR}/testfile"
+
+WORKLOADS=(
+  "prepare_fill"
+  "randread_4k"
+  "randwrite_4k"
+  "randrw_4k"
+)
+
+MODES=(
+  "sys"
+  "zvfs"
+)
+
+require_bin() {
+  local bin_path="$1"
+  local name="$2"
+
+  if [[ -z "${bin_path}" || ! -x "${bin_path}" ]]; then
+    echo "未找到 ${name}，请先安装或通过环境变量指定路径。" >&2
+    exit 1
+  fi
+}
+
+ensure_paths() {
+  mkdir -p "${SYS_PARENT_DIR}"
+
+  if [[ ! -d "${ZVFS_PARENT_DIR}" ]]; then
+    echo "未找到 ZVFS 测试目录：${ZVFS_PARENT_DIR}" >&2
+    echo "请先准备好对应目录，再执行脚本。" >&2
+    exit 1
+  fi
+
+  if [[ -z "${ZVFS_LD_PRELOAD_VALUE}" ]]; then
+    echo "ZVFS_LD_PRELOAD_VALUE 不能为空。" >&2
+    exit 1
+  fi
+
+  local zvfs_preload_file="${ZVFS_LD_PRELOAD_VALUE%%:*}"
+  if [[ ! -f "${zvfs_preload_file}" ]]; then
+    echo "未找到 libzvfs.so：${zvfs_preload_file}" >&2
+    exit 1
+  fi
+}
+
+cleanup_test_file() {
+  local mode="$1"
+
+  if [[ "${mode}" == "zvfs" ]]; then
+    env LD_PRELOAD="${ZVFS_LD_PRELOAD_VALUE}" rm -f "${ZVFS_TEST_FILE}"
+  else
+    rm -f "${SYS_TEST_FILE}"
+  fi
+}
+
+cleanup_all_test_files() {
+  cleanup_test_file "sys"
+  cleanup_test_file "zvfs"
+}
+
+run_case() {
+  local mode="$1"
+  local workload="$2"
+  local fio_file="${FIO_CASE_DIR}/${mode}/${workload}.fio"
+  local output_file="${RUN_DIR}/${mode}_${workload}.json"
+
+  if [[ ! -f "${fio_file}" ]]; then
+    echo "未找到 fio 配置：${fio_file}" >&2
+    exit 1
+  fi
+
+  echo "[run] mode=${mode} workload=${workload}"
+
+  if [[ "${workload}" == "prepare_fill" ]]; then
+    cleanup_test_file "${mode}"
+  fi
+
+  if [[ "${mode}" == "zvfs" ]]; then
+    env LD_PRELOAD="${ZVFS_LD_PRELOAD_VALUE}" \
+      "${FIO_BIN}" --output-format=json --output="${output_file}" "${fio_file}"
+  else
+    "${FIO_BIN}" --output-format=json --output="${output_file}" "${fio_file}"
+  fi
+}
+
+write_summary() {
+  local summary_file="${RUN_DIR}/summary.md"
+  local parser_script
+  parser_script="$(cat <<'PY'
+import json
+import os
+import sys
+
+run_dir = sys.argv[1]
+modes = ["sys", "zvfs"]
+workloads = ["prepare_fill", "randread_4k", "randwrite_4k", "randrw_4k"]
+
+def load_job(path):
+    with open(path, "r", encoding="utf-8") as f:
+        content = f.read()
+    start = content.find("{")
+    if start == -1:
+        raise RuntimeError(f"fio output is not JSON: {path}")
+    data = json.loads(content[start:])
+    jobs = data.get("jobs", [])
+    if not jobs:
+        raise RuntimeError(f"fio output has no jobs: {path}")
+    return data, jobs[0]
+
+def mib_per_sec(io_part):
+    value = io_part.get("bw_bytes", 0)
+    if not value:
+        return "-"
+    return f"{value / 1024 / 1024:.2f}"
+
+def iops(io_part):
+    value = io_part.get("iops", 0)
+    if not value:
+        return "-"
+    return f"{value:.2f}"
+
+def lat_us(io_part):
+    clat_ns = io_part.get("clat_ns", {})
+    value = clat_ns.get("mean", 0)
+    if not value:
+        return "-"
+    return f"{value / 1000:.2f}"
+
+def disk_util(data):
+    parts = []
+    for item in data.get("disk_util", []):
+        name = item.get("name")
+        util = item.get("util")
+        if not name or util is None:
+            continue
+        parts.append(f"{name}={util:.2f}%")
+    return ", ".join(parts) if parts else "-"
+
+lines = [
+    "# FIO Summary",
+    "",
+    f"run_dir: `{run_dir}`",
+    "",
+    "| mode | workload | bw(MiB/s) | iops | avg_lat(us) | read_iops | write_iops | read_avg_lat(us) | write_avg_lat(us) | disk_util |",
+    "| --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | --- |",
+]
+
+for mode in modes:
+    for workload in workloads:
+        path = os.path.join(run_dir, f"{mode}_{workload}.json")
+        data, job = load_job(path)
+        read_part = job.get("read", {})
+        write_part = job.get("write", {})
+        util_col = disk_util(data)
+
+        if workload == "prepare_fill":
+            row = [
+                mode,
+                workload,
+                mib_per_sec(write_part),
+                "-",
+                "-",
+                "-",
+                "-",
+                "-",
+                "-",
+                util_col,
+            ]
+        elif workload == "randread_4k":
+            row = [
+                mode,
+                workload,
+                mib_per_sec(read_part),
+                iops(read_part),
+                lat_us(read_part),
+                "-",
+                "-",
+                "-",
+                "-",
+                util_col,
+            ]
+        elif workload == "randwrite_4k":
+            row = [
+                mode,
+                workload,
+                mib_per_sec(write_part),
+                iops(write_part),
+                lat_us(write_part),
+                "-",
+                "-",
+                "-",
+                "-",
+                util_col,
+            ]
+        else:
+            row = [
+                mode,
+                workload,
+                "-",
+                "-",
+                "-",
+                iops(read_part),
+                iops(write_part),
+                lat_us(read_part),
+                lat_us(write_part),
+                util_col,
+            ]
+
+        lines.append("| " + " | ".join(row) + " |")
+
+print("\n".join(lines))
+PY
+)"
+
+  "${PYTHON_BIN}" -c "${parser_script}" "${RUN_DIR}" > "${summary_file}"
+  echo
+  echo "[summary] ${summary_file}"
+  sed -n '1,200p' "${summary_file}"
+}
+
+main() {
+  require_bin "${FIO_BIN}" "fio"
+  require_bin "${PYTHON_BIN}" "python3"
+  ensure_paths
+
+  mkdir -p "${RUN_DIR}"
+
+  echo "结果目录：${RUN_DIR}"
+  echo "fio 配置目录：${FIO_CASE_DIR}"
+  echo "LD_PRELOAD：${ZVFS_LD_PRELOAD_VALUE}"
+  echo
+
+  cleanup_all_test_files
+
+  for mode in "${MODES[@]}"; do
+    for workload in "${WORKLOADS[@]}"; do
+      run_case "${mode}" "${workload}"
+    done
+  done
+
+  write_summary
+  cleanup_all_test_files
+}
+
+main "$@"

scripts/run_pgbench_zvfs.sh

@@ -1,6 +1,8 @@
 #!/usr/bin/env bash
 set -euo pipefail
 
+ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+
 # 简化版 pgbench 测试脚本：
 # 1) 参数都在本文件顶部配置；
 # 2) 直接连接 benchdb；
@@ -21,7 +23,7 @@ PG_INIT_STEPS="dtg"
 PG_SKIP_INIT="0"
 PG_SUPERUSER="postgres"
 USE_LD_PRELOAD="1"
-LD_PRELOAD_PATH="/home/lian/share/zvfs/src/libzvfs.so"
+LD_PRELOAD_PATH="${ZVFS_LD_PRELOAD_VALUE:-${ROOT_DIR}/src/libzvfs.so}"
 
 # 可选：优先使用这个目录；为空时自动从 PATH 里找
 PG_BIN_DIR="/usr/lib/postgresql/12/bin"

scripts/run_test_hook_api.sh

@@ -1 +1,7 @@
-LD_PRELOAD=/home/lian/share/zvfs/src/libzvfs.so ZVFS_TEST_ROOT=/zvfs /home/lian/share/zvfs/tests/bin/hook_api_test
+#!/usr/bin/env bash
+set -euo pipefail
+
+ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+LD_PRELOAD_PATH="${ZVFS_LD_PRELOAD_VALUE:-${ROOT_DIR}/src/libzvfs.so}"
+
+LD_PRELOAD="${LD_PRELOAD_PATH}" ZVFS_TEST_ROOT=/zvfs "${ROOT_DIR}/tests/bin/hook_api_test"

src/hook/zvfs_hook_dir.c

@@ -14,6 +14,7 @@
 #include <errno.h>
 #include <string.h>
 #include <limits.h>
+#include <fcntl.h>
 #include <linux/fs.h>       /* RENAME_EXCHANGE, RENAME_NOREPLACE */
 
 /* ------------------------------------------------------------------ */
@@ -34,22 +35,36 @@
  *        （inode 和 blob 的清理推迟到 close 路径中 ref_count 归零时）
  */
 static void
-zvfs_unlink_path(const char *abspath)
+zvfs_unlink_blob_by_id(uint64_t blob_id)
+{
+    if (blob_id != 0) {
+        (void)blob_delete(blob_id);
+    }
+}
+
+static void
+zvfs_unlink_path(const char *abspath, uint64_t fallback_blob_id)
 {
     /* --- 查 path_cache -------------------------------------------- */
     pthread_mutex_lock(&g_fs.path_mu);
     struct zvfs_path_entry *pe = path_cache_lookup(abspath);
-    if (!pe) {
-        /*
-         * 不在缓存里：该文件可能从未被 open 过（没有 inode 对象）。
-         * 无内存状态需要清理，直接返回。
-         * blob 也不存在（文件从未被 zvfs open 创建），所以安全。
-         */
+    struct zvfs_inode *inode = pe ? pe->inode : NULL;
     pthread_mutex_unlock(&g_fs.path_mu);
+
+    if (!inode && fallback_blob_id != 0) {
+        pthread_mutex_lock(&g_fs.inode_mu);
+        inode = inode_lookup(fallback_blob_id);
+        pthread_mutex_unlock(&g_fs.inode_mu);
+    }
+
+    if (!inode) {
+        /*
+         * 文件可能从未被当前进程 open 过，因此 path_cache / inode_table
+         * 都为空；但只要删前成功读到 xattr，就仍需删除对应 blob。
+         */
+        zvfs_unlink_blob_by_id(fallback_blob_id);
         return;
     }
-    struct zvfs_inode *inode = pe->inode;
-    pthread_mutex_unlock(&g_fs.path_mu);
 
     /* --- 持 inode->mu 决策 ---------------------------------------- */
     pthread_mutex_lock(&inode->mu);
@@ -63,7 +78,7 @@ zvfs_unlink_path(const char *abspath)
          */
         pthread_mutex_unlock(&inode->mu);
 
-        blob_delete(inode->blob_id);
+        zvfs_unlink_blob_by_id(inode->blob_id);
 
         pthread_mutex_lock(&g_fs.inode_mu);
         inode_remove(inode->blob_id);
@@ -90,6 +105,29 @@ zvfs_unlink_path(const char *abspath)
     }
 }
 
+static int
+zvfs_prefetch_unlink_target(const char *abspath, uint64_t *blob_id_out)
+{
+    int fd;
+
+    *blob_id_out = 0;
+
+    fd = real_open(abspath, O_RDONLY);
+    if (fd < 0) {
+        return -1;
+    }
+
+    if (zvfs_xattr_read_blob_id(fd, blob_id_out) < 0) {
+        int saved = errno;
+        real_close(fd);
+        errno = saved;
+        return -1;
+    }
+
+    real_close(fd);
+    return 0;
+}
+
 /* ------------------------------------------------------------------ */
 /* unlink                                                               */
 /* ------------------------------------------------------------------ */
@@ -99,8 +137,16 @@ unlink(const char *path)
 {
     ZVFS_HOOK_ENTER();
     int ret;
+    char abspath[PATH_MAX];
+    uint64_t blob_id = 0;
+    int is_zvfs_path = 0;
 
-    if (ZVFS_IN_HOOK() || !zvfs_is_zvfs_path(path)) {
+    if (!ZVFS_IN_HOOK() &&
+        zvfs_resolve_atpath(AT_FDCWD, path, abspath, sizeof(abspath)) == 0) {
+        is_zvfs_path = zvfs_is_zvfs_path(abspath);
+    }
+
+    if (ZVFS_IN_HOOK() || !is_zvfs_path) {
         ret = real_unlink(path);
         ZVFS_HOOK_LEAVE();
         return ret;
@@ -108,10 +154,16 @@ unlink(const char *path)
 
     zvfs_ensure_init();
 
+    if (zvfs_prefetch_unlink_target(abspath, &blob_id) != 0) {
+        ret = real_unlink(path);
+        ZVFS_HOOK_LEAVE();
+        return ret;
+    }
+
     /* 先让真实 FS 删除文件（xattr 随之消失） */
     ret = real_unlink(path);
     if (ret == 0)
-        zvfs_unlink_path(path);
+        zvfs_unlink_path(abspath, blob_id);
 
     ZVFS_HOOK_LEAVE();
     return ret;
@@ -126,6 +178,7 @@ unlinkat(int dirfd, const char *path, int flags)
 {
     ZVFS_HOOK_ENTER();
     int ret;
+    uint64_t blob_id = 0;
 
     /*
      * AT_REMOVEDIR：rmdir 语义，目录由真实 FS 管理，直接透传。
@@ -151,9 +204,15 @@ unlinkat(int dirfd, const char *path, int flags)
 
     zvfs_ensure_init();
 
+    if (zvfs_prefetch_unlink_target(abspath, &blob_id) != 0) {
+        ret = real_unlinkat(dirfd, path, flags);
+        ZVFS_HOOK_LEAVE();
+        return ret;
+    }
+
     ret = real_unlinkat(dirfd, path, flags);
     if (ret == 0)
-        zvfs_unlink_path(abspath);
+        zvfs_unlink_path(abspath, blob_id);
 
     ZVFS_HOOK_LEAVE();
     return ret;
@@ -170,6 +229,7 @@ remove(const char *path)
     int ret;
     char abspath[PATH_MAX];
     int is_zvfs_path = 0;
+    uint64_t blob_id = 0;
 
     if (!ZVFS_IN_HOOK() &&
         zvfs_resolve_atpath(AT_FDCWD, path, abspath, sizeof(abspath)) == 0) {
@@ -184,13 +244,19 @@ remove(const char *path)
 
     zvfs_ensure_init();
 
+    if (zvfs_prefetch_unlink_target(abspath, &blob_id) != 0) {
+        ret = real_remove(path);
+        ZVFS_HOOK_LEAVE();
+        return ret;
+    }
+
     ret = real_remove(path);
     if (ret == 0) {
         /*
          * remove 既可删文件也可删空目录。
          * 若是目录，path_cache 不会命中，zvfs_unlink_path 会安全 no-op。
          */
-        zvfs_unlink_path(abspath);
+        zvfs_unlink_path(abspath, blob_id);
     }
 
     ZVFS_HOOK_LEAVE();
@@ -221,7 +287,7 @@ zvfs_rename_paths(const char *oldabs, const char *newabs)
          * real_rename 已经把它从磁盘上删掉了，
          * 走和 unlink 一样的延迟/立即 blob_delete 逻辑。
          */
-        zvfs_unlink_path(newabs);
+        zvfs_unlink_path(newabs, 0);
     }
 
     /* 把 oldpath 的缓存条目 rename 到 newpath */

src/hook/zvfs_hook_fd.c

@@ -207,12 +207,37 @@ zvfs_open_impl(int real_fd, const char *abspath, int flags, mode_t mode)
     struct zvfs_inode *inode = NULL;
     uint64_t blob_id = 0;
     uint64_t handle_id = 0;
+    int create_new = 0;
 
     zvfs_debug_open_log(abspath, NULL,
                         "open_impl enter real_fd=%d path=%s flags=0x%x mode=%#o",
                         real_fd, zvfs_dbg_str(abspath), flags, (unsigned)mode);
 
     if (flags & O_CREAT) {
+        /*
+         * O_CREAT does not imply the file is newly created.
+         * fio, for example, may open an existing file with O_CREAT again
+         * during the worker phase. Only create a new blob when the backing
+         * file does not already carry a ZVFS blob_id xattr.
+         */
+        if (zvfs_xattr_read_blob_id(real_fd, &blob_id) == 0) {
+            create_new = 0;
+        } else if (errno == ENODATA
+#ifdef ENOATTR
+                   || errno == ENOATTR
+#endif
+        ) {
+            create_new = 1;
+            blob_id = 0;
+        } else {
+            zvfs_debug_open_log(abspath, NULL,
+                                "open_impl xattr probe fail errno=%d(%s)",
+                                errno, strerror(errno));
+            goto fail;
+        }
+    }
+
+    if (create_new) {
         /* ---- 创建路径 -------------------------------------------- */
 
         /* 1. 创建 blob */

src/hook/zvfs_hook_seek.c

@@ -239,24 +239,35 @@ fallocate(int fd, int mode, off_t offset, off_t len)
         return -1;
     }
 
-    /* FALLOC_FL_KEEP_SIZE：预分配但不改变文件逻辑大小，直接返回 0 */
-    if (mode & FALLOC_FL_KEEP_SIZE) {
-        ZVFS_HOOK_LEAVE();
-        return 0;
-    }
+    uint64_t new_end = (uint64_t)offset + (uint64_t)len;
+    uint64_t alloc_end = new_end;
 
     /*
-     * 普通 fallocate（mode == 0）：
-     * 确保 [offset, offset+len) 范围内的空间被"分配"。
-     * zvfs 的语义：把 logical_size 扩展到 max(logical_size, offset+len)。
-     * 不提前 blob_resize，因为 SPDK cluster 按写入时分配更高效。
+     * ZVFS 读路径会按 blob 当前 cluster 大小做越界检查，因此这里不能只
+     * 更新 logical_size/st_size，必须同步把 blob capacity 扩到目标范围。
+     *
+     * 对 KEEP_SIZE：
+     *   预分配空间，但不改变文件逻辑大小。
+     * 对普通 fallocate：
+     *   预分配空间，并把逻辑大小扩展到 offset + len。
      */
-    uint64_t new_end = (uint64_t)offset + (uint64_t)len;
+    pthread_mutex_lock(&of->inode->mu);
+    if (alloc_end < of->inode->logical_size)
+        alloc_end = of->inode->logical_size;
+    pthread_mutex_unlock(&of->inode->mu);
 
+    if (blob_resize(of->handle_id, alloc_end) < 0) {
+        errno = EIO;
+        ZVFS_HOOK_LEAVE();
+        return -1;
+    }
+
+    if ((mode & FALLOC_FL_KEEP_SIZE) == 0) {
         pthread_mutex_lock(&of->inode->mu);
         if (new_end > of->inode->logical_size)
             inode_update_size(of->inode, fd, new_end);
         pthread_mutex_unlock(&of->inode->mu);
+    }
 
     ZVFS_HOOK_LEAVE();
     return 0;
@@ -293,6 +304,17 @@ posix_fallocate(int fd, off_t offset, off_t len)
     if (offset < 0 || len <= 0) { ZVFS_HOOK_LEAVE(); return EINVAL; }
 
     uint64_t new_end = (uint64_t)offset + (uint64_t)len;
+    uint64_t alloc_end = new_end;
+
+    pthread_mutex_lock(&of->inode->mu);
+    if (alloc_end < of->inode->logical_size)
+        alloc_end = of->inode->logical_size;
+    pthread_mutex_unlock(&of->inode->mu);
+
+    if (blob_resize(of->handle_id, alloc_end) < 0) {
+        ZVFS_HOOK_LEAVE();
+        return errno ? errno : EIO;
+    }
 
     pthread_mutex_lock(&of->inode->mu);
     if (new_end > of->inode->logical_size)

zvfs_fio_test/sys/prepare_fill.fio

@@ -0,0 +1,13 @@
+[global]
+ioengine=psync
+direct=1
+thread=1
+group_reporting=1
+iodepth=1
+bs=4k
+size=512M
+filename=/tmp/zvfs_sys_fio/testfile
+
+[prepare_fill]
+description=Sequential prefill for bandwidth
+rw=write

zvfs_fio_test/sys/randread_4k.fio

@@ -0,0 +1,15 @@
+[global]
+ioengine=psync
+direct=1
+thread=1
+group_reporting=1
+iodepth=1
+bs=4k
+size=512M
+time_based=1
+runtime=30
+filename=/tmp/zvfs_sys_fio/testfile
+
+[randread_4k]
+description=4K random read for IOPS and average latency
+rw=randread

zvfs_fio_test/sys/randrw_4k.fio

@@ -0,0 +1,16 @@
+[global]
+ioengine=psync
+direct=1
+thread=1
+group_reporting=1
+iodepth=1
+bs=4k
+size=512M
+time_based=1
+runtime=30
+filename=/tmp/zvfs_sys_fio/testfile
+
+[randrw_4k]
+description=4K random read write for per direction IOPS and average latency
+rw=randrw
+rwmixread=50

zvfs_fio_test/sys/randwrite_4k.fio

@@ -0,0 +1,15 @@
+[global]
+ioengine=psync
+direct=1
+thread=1
+group_reporting=1
+iodepth=1
+bs=4k
+size=512M
+time_based=1
+runtime=30
+filename=/tmp/zvfs_sys_fio/testfile
+
+[randwrite_4k]
+description=4K random write for IOPS and average latency
+rw=randwrite

zvfs_fio_test/zvfs/prepare_fill.fio

@@ -0,0 +1,13 @@
+[global]
+ioengine=psync
+direct=1
+thread=1
+group_reporting=1
+iodepth=1
+bs=4k
+size=512M
+filename=/zvfs/zvfs_sys_fio/testfile
+
+[prepare_fill]
+description=Sequential prefill for bandwidth
+rw=write

zvfs_fio_test/zvfs/randread_4k.fio

@@ -0,0 +1,15 @@
+[global]
+ioengine=psync
+direct=1
+thread=1
+group_reporting=1
+iodepth=1
+bs=4k
+size=512M
+time_based=1
+runtime=30
+filename=/zvfs/zvfs_sys_fio/testfile
+
+[randread_4k]
+description=4K random read for IOPS and average latency
+rw=randread

zvfs_fio_test/zvfs/randrw_4k.fio

@@ -0,0 +1,16 @@
+[global]
+ioengine=psync
+direct=1
+thread=1
+group_reporting=1
+iodepth=1
+bs=4k
+size=512M
+time_based=1
+runtime=30
+filename=/zvfs/zvfs_sys_fio/testfile
+
+[randrw_4k]
+description=4K random read write for per direction IOPS and average latency
+rw=randrw
+rwmixread=50

zvfs_fio_test/zvfs/randwrite_4k.fio

@@ -0,0 +1,15 @@
+[global]
+ioengine=psync
+direct=1
+thread=1
+group_reporting=1
+iodepth=1
+bs=4k
+size=512M
+time_based=1
+runtime=30
+filename=/zvfs/zvfs_sys_fio/testfile
+
+[randwrite_4k]
+description=4K random write for IOPS and average latency
+rw=randwrite