[PATCH] io_uring/uring_cmd: skip inline completion cleanup if unlocked

[PATCH] io_uring/uring_cmd: skip inline completion cleanup if unlocked

[Jens Axboe]

If the call path to __io_uring_cmd_done() is not locked, then we
cannot recycle the uring_cmd to our allocation cache. Check for
that and skip it, and let the normal locked completion flushing
do the cleanup.

Signed-off-by: Jens Axboe <axboe@kernel.dk>

---

This effectively defeats proper cache recyling for uring_cmd opcodes,
with the fix it's working fine again.

diff --git a/io_uring/uring_cmd.c b/io_uring/uring_cmd.c
index 42be1be5b132..35e2aa8b9446 100644
--- a/io_uring/uring_cmd.c
+++ b/io_uring/uring_cmd.c
@@ -166,7 +166,9 @@ void __io_uring_cmd_done(struct io_uring_cmd *ioucmd, s32 ret, u64 res2,
 			req->cqe.flags |= IORING_CQE_F_32;
 		io_req_set_cqe32_extra(req, res2, 0);
 	}
-	io_req_uring_cleanup(req, issue_flags);
+	/* defer cleanup if not locked, otherwise cache recyling is skipped */
+	if (!(issue_flags & IO_URING_F_UNLOCKED))
+		io_req_uring_cleanup(req, issue_flags);
 	if (req->flags & REQ_F_IOPOLL) {
 		/* order with io_iopoll_req_issued() checking ->iopoll_complete */
 		smp_store_release(&req->iopoll_completed, 1);
@@ -211,6 +213,7 @@ int io_uring_cmd_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
 	ac = io_uring_alloc_async_data(&req->ctx->cmd_cache, req);
 	if (!ac)
 		return -ENOMEM;
+	req->flags |= REQ_F_NEED_CLEANUP;
 	ioucmd->sqe = sqe;
 	return 0;
 }

-- 
Jens Axboe

Re: [PATCH] io_uring/uring_cmd: skip inline completion cleanup if unlocked

[Caleb Sander Mateos]

On Wed, May 6, 2026 at 4:04 AM Jens Axboe <axboe@kernel.dk> wrote:
>
> If the call path to __io_uring_cmd_done() is not locked, then we
> cannot recycle the uring_cmd to our allocation cache. Check for
> that and skip it, and let the normal locked completion flushing
> do the cleanup.
>
> Signed-off-by: Jens Axboe <axboe@kernel.dk>
>
> ---
>
> This effectively defeats proper cache recyling for uring_cmd opcodes,
> with the fix it's working fine again.
>
> diff --git a/io_uring/uring_cmd.c b/io_uring/uring_cmd.c
> index 42be1be5b132..35e2aa8b9446 100644
> --- a/io_uring/uring_cmd.c
> +++ b/io_uring/uring_cmd.c
> @@ -166,7 +166,9 @@ void __io_uring_cmd_done(struct io_uring_cmd *ioucmd, s32 ret, u64 res2,
>                         req->cqe.flags |= IORING_CQE_F_32;
>                 io_req_set_cqe32_extra(req, res2, 0);
>         }
> -       io_req_uring_cleanup(req, issue_flags);
> +       /* defer cleanup if not locked, otherwise cache recyling is skipped */

"recycling"?

> +       if (!(issue_flags & IO_URING_F_UNLOCKED))
> +               io_req_uring_cleanup(req, issue_flags);

Doesn't io_req_uring_cleanup() already check this?

Best,
Caleb

>         if (req->flags & REQ_F_IOPOLL) {
>                 /* order with io_iopoll_req_issued() checking ->iopoll_complete */
>                 smp_store_release(&req->iopoll_completed, 1);
> @@ -211,6 +213,7 @@ int io_uring_cmd_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
>         ac = io_uring_alloc_async_data(&req->ctx->cmd_cache, req);
>         if (!ac)
>                 return -ENOMEM;
> +       req->flags |= REQ_F_NEED_CLEANUP;
>         ioucmd->sqe = sqe;
>         return 0;
>  }
>
> --
> Jens Axboe
>
>

Re: [PATCH] io_uring/uring_cmd: skip inline completion cleanup if unlocked

[Jens Axboe]

On 5/6/26 10:14 AM, Caleb Sander Mateos wrote:
> On Wed, May 6, 2026 at 4:04?AM Jens Axboe <axboe@kernel.dk> wrote:
>>
>> If the call path to __io_uring_cmd_done() is not locked, then we
>> cannot recycle the uring_cmd to our allocation cache. Check for
>> that and skip it, and let the normal locked completion flushing
>> do the cleanup.
>>
>> Signed-off-by: Jens Axboe <axboe@kernel.dk>
>>
>> ---
>>
>> This effectively defeats proper cache recyling for uring_cmd opcodes,
>> with the fix it's working fine again.
>>
>> diff --git a/io_uring/uring_cmd.c b/io_uring/uring_cmd.c
>> index 42be1be5b132..35e2aa8b9446 100644
>> --- a/io_uring/uring_cmd.c
>> +++ b/io_uring/uring_cmd.c
>> @@ -166,7 +166,9 @@ void __io_uring_cmd_done(struct io_uring_cmd *ioucmd, s32 ret, u64 res2,
>>                         req->cqe.flags |= IORING_CQE_F_32;
>>                 io_req_set_cqe32_extra(req, res2, 0);
>>         }
>> -       io_req_uring_cleanup(req, issue_flags);
>> +       /* defer cleanup if not locked, otherwise cache recyling is skipped */
> 
> "recycling"?
> 
>> +       if (!(issue_flags & IO_URING_F_UNLOCKED))
>> +               io_req_uring_cleanup(req, issue_flags);
> 
> Doesn't io_req_uring_cleanup() already check this?

True it does, I think all we need is:

>> @@ -211,6 +213,7 @@ int io_uring_cmd_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
>>         ac = io_uring_alloc_async_data(&req->ctx->cmd_cache, req);
>>         if (!ac)
>>                 return -ENOMEM;
>> +       req->flags |= REQ_F_NEED_CLEANUP;
>>         ioucmd->sqe = sqe;
>>         return 0;
>>  }

to ensure it's called later too. I'll update it, thanks!

-- 
Jens Axboe