测试内存池

2026-01-31 15:38:52 +00:00
parent fbdcff6878
commit 003566b69a
29 changed files with 614 additions and 233 deletions
--- a/diskuring/diskuring.c
+++ b/diskuring/diskuring.c
@@ -3,8 +3,16 @@
 #include <poll.h>
 #include <sys/eventfd.h>

+static destroy_queue_t g_destroy_queue = {NULL, PTHREAD_MUTEX_INITIALIZER};
+
+static long long push_to_queue  = 0;
+static long long push_to_sqe    = 0;
+static long long get_from_cqe   = 0;
+static long long release_cnt    = 0;
+
 void task_init(task_t *t)
 {
+    push_to_queue ++;
    pthread_mutex_init(&t->m, NULL);
    pthread_cond_init(&t->cv, NULL);
    t->done = 0;
@@ -59,68 +67,196 @@ static void queue_push(iouring_ctx_t *ctx, task_t *t)
    pthread_mutex_unlock(&ctx->q_m);
 }

+static void queue_push_front(iouring_ctx_t *ctx, task_t *list_head, task_t *list_tail) { 
+    pthread_mutex_lock(&ctx->q_m); 
+    list_tail->next = ctx->q_head;
+    ctx->q_head = list_head;
+    if (!ctx->q_tail) {
+        ctx->q_tail = list_tail;
+    }
+    pthread_cond_signal(&ctx->q_cv); 
+    pthread_mutex_unlock(&ctx->q_m); }
+
 static task_t *queue_pop_all(iouring_ctx_t *ctx)
 {
    pthread_mutex_lock(&ctx->q_m);
-    while (!ctx->stop && ctx->q_head == NULL)
-    {
-        pthread_cond_wait(&ctx->q_cv, &ctx->q_m);
-    }
    task_t *list = ctx->q_head;
    ctx->q_head = ctx->q_tail = NULL;
    pthread_mutex_unlock(&ctx->q_m);
    return list;
 }

+static task_t *queue_pop_n(iouring_ctx_t *ctx, int n)
+{
+    if (n <= 0)
+        return NULL;
+    pthread_mutex_lock(&ctx->q_m);
+    task_t *head = ctx->q_head;
+    if (!head) {
+        pthread_mutex_unlock(&ctx->q_m);
+        return NULL;
+    }
+    task_t *curr = head;
+    task_t *prev = NULL;
+    int count = 0;
+
+    while (curr && count < n) {
+        prev = curr;
+        curr = curr->next;
+        count++;
+    }
+
+    ctx->q_head = curr;
+    if (!curr) {
+        // 队列被取空
+        ctx->q_tail = NULL;
+    }
+    prev->next = NULL;
+    pthread_mutex_unlock(&ctx->q_m);
+    return head;
+}
+
+extern void sync_wakeup();
 static void *worker_main(void *arg)
 {
    iouring_ctx_t *ctx = (iouring_ctx_t *)arg;
+    const int BATCH_SIZE = 256;  // 每次最多准备这么多，防止一次占满 SQ

    while (!ctx->stop)
-    {
-
-        // 1. 阻塞等任务（这里睡觉）
-        task_t *list = queue_pop_all(ctx);
-        if (!list)
-            continue;
-
-        // 2. 提交所有任务
-        task_t *t;
-        int submit_cnt = 0;
-
-        for (t = list; t; t = t->next)
-        {
-            struct io_uring_sqe *sqe = io_uring_get_sqe(&ctx->ring);
-            if (!sqe)
-                break;
-
-            io_uring_prep_writev(sqe, t->fd, t->iovs, t->iovcnt, t->off);
-            sqe->user_data = (uint64_t)(uintptr_t)t;
-            submit_cnt++;
+    {   
+        int cq_count = 0;
+        
+        // ========== 1. 疯狂收割 CQE（必须优先做，释放 in_flight 额度）==========
+        // 使用 while 而不是 if，确保把 CQ 薅干净
+        while (true) {
+            struct io_uring_cqe *cqe;
+            unsigned head;
+            
+            io_uring_for_each_cqe(&ctx->ring, head, cqe) {
+                task_t *done = (task_t *)(uintptr_t)cqe->user_data;
+                
+                // 先减计数（必须在处理前减，否则可能瞬间突破上限）
+                atomic_fetch_sub(&ctx->in_flight, 1);
+                
+                task_finish(done, cqe->res);
+                
+                if (cqe->res < 0) {
+                    fprintf(stderr, "write fail: fd=%d res=%d\n", done->fd, cqe->res);
+                }
+                
+                // 加入销毁队列
+                pthread_mutex_lock(&g_destroy_queue.lock);
+                done->next = g_destroy_queue.head;
+                g_destroy_queue.head = done;
+                pthread_mutex_unlock(&g_destroy_queue.lock);
+                
+                get_from_cqe++;
+                cq_count++;
+            }
+            
+            if (cq_count > 0) {
+                io_uring_cq_advance(&ctx->ring, cq_count);
+                sync_wakeup();
+            }
+            
+            // 如果这次没收满，说明 CQ 空了，退出收割循环
+            if (cq_count == 0) break;
+            cq_count = 0;  // 重置继续薅（可能有新的完成了）
        }

-        io_uring_submit(&ctx->ring);
+        // 检查溢出（保险起见，虽然有了背压不该再溢出）
+        if (*ctx->ring.sq.kflags & IORING_SQ_CQ_OVERFLOW) {
+            fprintf(stderr, "FATAL: CQ overflow detected! Backpressure broken!\n");
+            abort();  // 直接崩溃，说明逻辑有 bug
+        }

-        // 3. 等全部完成
-        for (int i = 0; i < submit_cnt; ++i)
-        {
+        // ========== 2. 计算还能提交多少 ==========
+        int current_in_flight = atomic_load(&ctx->in_flight);
+        int available_slots = ctx->max_in_flight - current_in_flight;
+        
+        if (available_slots <= 0) {
+            // 满了！不能取新任务，必须等待 CQE（忙等或阻塞等）
+            
+            // 方案 B：阻塞等 CQE（推荐）
            struct io_uring_cqe *cqe;
-            io_uring_wait_cqe(&ctx->ring, &cqe);
-
-            task_t *done = (task_t *)(uintptr_t)cqe->user_data;
-            task_finish(done, cqe->res);
-
-            // todo: 失败应该通知主线程，提供一个链表，把失败的taskpush进去。主线程用定时器任务定时处理失败信息然后destory。
-            // 暂时用打印说明失败
-            if(done->res < 0){
-                printf("uring failed: fd:%d, offset:%ld\n", done->fd, done->off);
+            int ret = io_uring_wait_cqe(&ctx->ring, &cqe);
+            if (ret == 0 && !ctx->stop) {
+                // 收到一个 CQE，回循环开头处理
+                continue;
            }
-            task_destroy(done);
+            continue;
+        }

-            io_uring_cqe_seen(&ctx->ring, cqe);
+        // ========== 3. 从任务队列取任务（只取 available_slots 个）==========
+        
+        task_t *task_list = queue_pop_n(ctx, available_slots);  
+        if (!task_list) {
+            if (!ctx->stop && atomic_load(&ctx->in_flight) > 0) {
+                int ret = io_uring_submit_and_wait(&ctx->ring, 1);
+                continue;   
+            }
+            // 没任务，等待条件变量
+            pthread_mutex_lock(&ctx->q_m);
+            while (ctx->q_head == NULL && !ctx->stop) {
+                pthread_cond_wait(&ctx->q_cv, &ctx->q_m);
+            }
+            pthread_mutex_unlock(&ctx->q_m);
+            continue;
+        }
+
+        // ========== 4. 准备 SQE（受限于 available_slots）==========
+        int batch_count = 0;
+        task_t *curr = task_list;
+        task_t *prev = NULL;
+        task_t *submitted_head = task_list;  // 记录这次实际要提交的部分
+        task_t *remaining_head = NULL;       // 装不下的部分
+        
+        while (curr && batch_count < available_slots) {
+            struct io_uring_sqe *sqe = io_uring_get_sqe(&ctx->ring);
+            
+            if (!sqe) {
+                // SQ 满了（这种情况在控制 inflight 后很少见，但保险起见）
+                break;
+            }
+            
+            io_uring_prep_writev(sqe, curr->fd, curr->iovs, curr->iovcnt, curr->off);
+            sqe->user_data = (uint64_t)(uintptr_t)curr;
+            
+            batch_count++;
+            prev = curr;
+            curr = curr->next;
+        }
+        
+        // 断开链表：已准备的 和 未准备的
+        if (prev) {
+            prev->next = NULL;  // 已提交的部分结尾
+        }
+        remaining_head = curr;  // 剩下的部分（如果有）
+
+        // ========== 5. 提交并增加计数 ==========
+        if (batch_count > 0) {
+            int submitted = io_uring_submit(&ctx->ring);
+            if (submitted != batch_count) {
+                fprintf(stderr, "CRITICAL: prep %d but submit %d\n", batch_count, submitted);
+                // 这种情况很严重，说明 ring 损坏了，建议 abort
+                abort();
+            }
+            
+            atomic_fetch_add(&ctx->in_flight, submitted);
+            push_to_sqe += submitted;
+        }
+
+        // ========== 6. 把没提交的任务塞回队列头部（保持顺序）==========
+        if (remaining_head) {
+            task_t *tail = remaining_head;
+            while (tail->next) tail = tail->next;
+
+            queue_push_front(ctx, remaining_head, tail);
        }
    }

+    printf("exit uring, stop: %d, inflight: %d\n", ctx->stop, 
+           atomic_load(&ctx->in_flight));
    return NULL;
 }

@@ -135,12 +271,26 @@ int iouring_init(iouring_ctx_t *ctx, unsigned entries)
    memset(ctx, 0, sizeof(*ctx));
    pthread_mutex_init(&ctx->q_m, NULL);
    pthread_cond_init(&ctx->q_cv, NULL);
+    ctx->stop = 0;

    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)
+    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);
+
+    ctx->max_in_flight = (cq_size * 8) / 10;
+    atomic_init(&ctx->in_flight, 0);

    ret = pthread_create(&ctx->th, NULL, worker_main, ctx);
    if (ret != 0)
@@ -201,3 +351,23 @@ task_t* submit_write(iouring_ctx_t *ctx, int fd, void **bufs, size_t *lens, int
    queue_push(ctx, t);
    return t;
 }
+
+// 主线程定期调用此函数清理
+void cleanup_finished_iouring_tasks() {
+    pthread_mutex_lock(&g_destroy_queue.lock);
+    task_t *list = g_destroy_queue.head;
+    g_destroy_queue.head = NULL;
+    pthread_mutex_unlock(&g_destroy_queue.lock);
+    
+    int cnt = 0;
+    while (list) {
+        cnt ++;
+        task_t *next = list->next;
+        task_destroy(list);  // 在主线程执行销毁
+        list = next;
+    }
+    // printf("clean: %d\n\n", cnt);
+    // mp_print();
+    release_cnt += cnt;
+    // printf("push:%lld, sqe:%lld, cqe:%lld, rls:%lld\n", push_to_queue, push_to_sqe, get_from_cqe, release_cnt);
+}