io_uring: defer file table grabbing request cleanup for locked requests
authorJens Axboe <axboe@kernel.dk>
Mon, 10 Aug 2020 16:55:56 +0000 (10:55 -0600)
committerJens Axboe <axboe@kernel.dk>
Mon, 10 Aug 2020 21:19:25 +0000 (15:19 -0600)
If we're in the error path failing links and we have a link that has
grabbed a reference to the fs_struct, then we cannot safely drop our
reference to the table if we already hold the completion lock. This
adds a hardirq dependency to the fs_struct->lock, which it currently
doesn't have.

Defer the final cleanup and free of such requests to avoid adding this
dependency.

Reported-by: syzbot+ef4b654b49ed7ff049bf@syzkaller.appspotmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
fs/io_uring.c

index 56115cb4b9faafd4fe0adb41d11dd7cbf8d7d9a8..5488698189da8e53fe42f44c229494404dfc55d9 100644 (file)
@@ -1108,10 +1108,16 @@ static void __io_commit_cqring(struct io_ring_ctx *ctx)
        }
 }
 
-static void io_req_clean_work(struct io_kiocb *req)
+/*
+ * Returns true if we need to defer file table putting. This can only happen
+ * from the error path with REQ_F_COMP_LOCKED set.
+ */
+static bool io_req_clean_work(struct io_kiocb *req)
 {
        if (!(req->flags & REQ_F_WORK_INITIALIZED))
-               return;
+               return false;
+
+       req->flags &= ~REQ_F_WORK_INITIALIZED;
 
        if (req->work.mm) {
                mmdrop(req->work.mm);
@@ -1124,6 +1130,9 @@ static void io_req_clean_work(struct io_kiocb *req)
        if (req->work.fs) {
                struct fs_struct *fs = req->work.fs;
 
+               if (req->flags & REQ_F_COMP_LOCKED)
+                       return true;
+
                spin_lock(&req->work.fs->lock);
                if (--fs->users)
                        fs = NULL;
@@ -1132,7 +1141,8 @@ static void io_req_clean_work(struct io_kiocb *req)
                        free_fs_struct(fs);
                req->work.fs = NULL;
        }
-       req->flags &= ~REQ_F_WORK_INITIALIZED;
+
+       return false;
 }
 
 static void io_prep_async_work(struct io_kiocb *req)
@@ -1544,7 +1554,7 @@ static inline void io_put_file(struct io_kiocb *req, struct file *file,
                fput(file);
 }
 
-static void io_dismantle_req(struct io_kiocb *req)
+static bool io_dismantle_req(struct io_kiocb *req)
 {
        io_clean_op(req);
 
@@ -1552,7 +1562,6 @@ static void io_dismantle_req(struct io_kiocb *req)
                kfree(req->io);
        if (req->file)
                io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
-       io_req_clean_work(req);
 
        if (req->flags & REQ_F_INFLIGHT) {
                struct io_ring_ctx *ctx = req->ctx;
@@ -1564,15 +1573,15 @@ static void io_dismantle_req(struct io_kiocb *req)
                        wake_up(&ctx->inflight_wait);
                spin_unlock_irqrestore(&ctx->inflight_lock, flags);
        }
+
+       return io_req_clean_work(req);
 }
 
-static void __io_free_req(struct io_kiocb *req)
+static void __io_free_req_finish(struct io_kiocb *req)
 {
-       struct io_ring_ctx *ctx;
+       struct io_ring_ctx *ctx = req->ctx;
 
-       io_dismantle_req(req);
        __io_put_req_task(req);
-       ctx = req->ctx;
        if (likely(!io_is_fallback_req(req)))
                kmem_cache_free(req_cachep, req);
        else
@@ -1580,6 +1589,39 @@ static void __io_free_req(struct io_kiocb *req)
        percpu_ref_put(&ctx->refs);
 }
 
+static void io_req_task_file_table_put(struct callback_head *cb)
+{
+       struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
+       struct fs_struct *fs = req->work.fs;
+
+       spin_lock(&req->work.fs->lock);
+       if (--fs->users)
+               fs = NULL;
+       spin_unlock(&req->work.fs->lock);
+       if (fs)
+               free_fs_struct(fs);
+       req->work.fs = NULL;
+       __io_free_req_finish(req);
+}
+
+static void __io_free_req(struct io_kiocb *req)
+{
+       if (!io_dismantle_req(req)) {
+               __io_free_req_finish(req);
+       } else {
+               int ret;
+
+               init_task_work(&req->task_work, io_req_task_file_table_put);
+               ret = task_work_add(req->task, &req->task_work, TWA_RESUME);
+               if (unlikely(ret)) {
+                       struct task_struct *tsk;
+
+                       tsk = io_wq_get_task(req->ctx->io_wq);
+                       task_work_add(tsk, &req->task_work, 0);
+               }
+       }
+}
+
 static bool io_link_cancel_timeout(struct io_kiocb *req)
 {
        struct io_ring_ctx *ctx = req->ctx;
@@ -1868,7 +1910,7 @@ static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
                req->flags &= ~REQ_F_TASK_PINNED;
        }
 
-       io_dismantle_req(req);
+       WARN_ON_ONCE(io_dismantle_req(req));
        rb->reqs[rb->to_free++] = req;
        if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
                __io_req_free_batch_flush(req->ctx, rb);