On 3/15/24 22:48, Jens Axboe wrote:
If we get a request with IOSQE_ASYNC set, then we first run the prep
async handlers. But if we then fail setting it up and want to post
a CQE with -EINVAL, we use ->done_io. This was previously guarded with
REQ_F_PARTIAL_IO, and the normal setup handlers do set it up before any
potential errors, but we need to cover the async setup too.
You can hit io_req_defer_failed() { opdef->fail(); }
off of an early submission failure path where def->prep has
not yet been called, I don't think the patch will fix the
problem.
->fail() handlers are fragile, maybe we should skip them
if def->prep() wasn't called. Not even compile tested:
diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
index 846d67a9c72e..56eed1490571 100644
--- a/io_uring/io_uring.c
+++ b/io_uring/io_uring.c
@@ -993,7 +993,7 @@ void io_req_defer_failed(struct io_kiocb *req, s32 res)
req_set_fail(req);
io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
- if (def->fail)
+ if ((req->flags & REQ_F_EARLY_FAIL) && def->fail)
def->fail(req);
io_req_complete_defer(req);
}
@@ -2201,8 +2201,7 @@ static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
}
req->flags |= REQ_F_CREDS;
}
-
- return def->prep(req, sqe);
+ return 0;
}
static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
@@ -2250,8 +2249,15 @@ static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
int ret;
ret = io_init_req(ctx, req, sqe);
- if (unlikely(ret))
+ if (unlikely(ret)) {
+fail:
+ req->flags |= REQ_F_EARLY_FAIL;
return io_submit_fail_init(sqe, req, ret);
+ }
+
+ ret = def->prep(req, sqe);
+ if (unlikely(ret))
+ goto fail;
trace_io_uring_submit_req(req);
--
Pavel Begunkov
