From 2ca10259b4189a433c309054496dd6af1415f992 Mon Sep 17 00:00:00 2001 From: Jens Axboe Date: Thu, 13 Feb 2020 17:17:35 -0700 Subject: io_uring: prune request from overflow list on flush MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Carter reported an issue where he could produce a stall on ring exit, when we're cleaning up requests that match the given file table. For this particular test case, a combination of a few things caused the issue: - The cq ring was overflown - The request being canceled was in the overflow list The combination of the above means that the cq overflow list holds a reference to the request. The request is canceled correctly, but since the overflow list holds a reference to it, the final put won't happen. Since the final put doesn't happen, the request remains in the inflight. Hence we never finish the cancelation flush. Fix this by removing requests from the overflow list if we're canceling them. Cc: stable@vger.kernel.org # 5.5 Reported-by: Carter Li 李通洲 Signed-off-by: Jens Axboe --- fs/io_uring.c | 28 ++++++++++++++++++++++++++++ 1 file changed, 28 insertions(+) (limited to 'fs/io_uring.c') diff --git a/fs/io_uring.c b/fs/io_uring.c index 6d4e20d59729..5a826017ebb8 100644 --- a/fs/io_uring.c +++ b/fs/io_uring.c @@ -481,6 +481,7 @@ enum { REQ_F_TIMEOUT_NOSEQ_BIT, REQ_F_COMP_LOCKED_BIT, REQ_F_NEED_CLEANUP_BIT, + REQ_F_OVERFLOW_BIT, }; enum { @@ -521,6 +522,8 @@ enum { REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT), /* needs cleanup */ REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT), + /* in overflow list */ + REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT), }; /* @@ -1103,6 +1106,7 @@ static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb, list); list_move(&req->list, &list); + req->flags &= ~REQ_F_OVERFLOW; if (cqe) { WRITE_ONCE(cqe->user_data, req->user_data); WRITE_ONCE(cqe->res, req->result); @@ -1155,6 +1159,7 @@ static void io_cqring_fill_event(struct io_kiocb *req, long res) set_bit(0, &ctx->sq_check_overflow); set_bit(0, &ctx->cq_check_overflow); } + req->flags |= REQ_F_OVERFLOW; refcount_inc(&req->refs); req->result = res; list_add_tail(&req->list, &ctx->cq_overflow_list); @@ -6463,6 +6468,29 @@ static void io_uring_cancel_files(struct io_ring_ctx *ctx, if (!cancel_req) break; + if (cancel_req->flags & REQ_F_OVERFLOW) { + spin_lock_irq(&ctx->completion_lock); + list_del(&cancel_req->list); + cancel_req->flags &= ~REQ_F_OVERFLOW; + if (list_empty(&ctx->cq_overflow_list)) { + clear_bit(0, &ctx->sq_check_overflow); + clear_bit(0, &ctx->cq_check_overflow); + } + spin_unlock_irq(&ctx->completion_lock); + + WRITE_ONCE(ctx->rings->cq_overflow, + atomic_inc_return(&ctx->cached_cq_overflow)); + + /* + * Put inflight ref and overflow ref. If that's + * all we had, then we're done with this request. + */ + if (refcount_sub_and_test(2, &cancel_req->refs)) { + io_put_req(cancel_req); + continue; + } + } + io_wq_cancel_work(ctx->io_wq, &cancel_req->work); io_put_req(cancel_req); schedule(); -- cgit v1.2.3