Re: [PATCH v2 05/14] ublk: don't hard code IO_URING_F_UNLOCKED

[Pavel Begunkov <asml.silence@gmail.com>]

On 3/18/24 12:52, Pavel Begunkov wrote:
> On 3/18/24 08:16, Ming Lei wrote:
>> On Mon, Mar 18, 2024 at 12:41:50AM +0000, Pavel Begunkov wrote:
>>> uring_cmd implementations should not try to guess issue_flags, just use
>>> a newly added io_uring_cmd_complete(). We're loosing an optimisation in
>>> the cancellation path in ublk_uring_cmd_cancel_fn(), but the assumption
>>> is that we don't care that much about it.
>>>
>>> Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
>>> Link: https://lore.kernel.org/r/2f7bc9fbc98b11412d10b8fd88e58e35614e3147.1710514702.git.asml.silence@gmail.com
>>> Signed-off-by: Jens Axboe <axboe@kernel.dk>
>>> ---
>>>   drivers/block/ublk_drv.c | 18 ++++++++----------
>>>   1 file changed, 8 insertions(+), 10 deletions(-)
>>>
>>> diff --git a/drivers/block/ublk_drv.c b/drivers/block/ublk_drv.c
>>> index bea3d5cf8a83..97dceecadab2 100644
>>> --- a/drivers/block/ublk_drv.c
>>> +++ b/drivers/block/ublk_drv.c
>>> @@ -1417,8 +1417,7 @@ static bool ublk_abort_requests(struct ublk_device *ub, struct ublk_queue *ubq)
>>>       return true;
>>>   }
>>> -static void ublk_cancel_cmd(struct ublk_queue *ubq, struct ublk_io *io,
>>> -        unsigned int issue_flags)
>>> +static void ublk_cancel_cmd(struct ublk_queue *ubq, struct ublk_io *io)
>>>   {
>>>       bool done;
>>> @@ -1432,15 +1431,14 @@ static void ublk_cancel_cmd(struct ublk_queue *ubq, struct ublk_io *io,
>>>       spin_unlock(&ubq->cancel_lock);
>>>       if (!done)
>>> -        io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0, issue_flags);
>>> +        io_uring_cmd_complete(io->cmd, UBLK_IO_RES_ABORT, 0);
>>>   }
>>>   /*
>>>    * The ublk char device won't be closed when calling cancel fn, so both
>>>    * ublk device and queue are guaranteed to be live
>>>    */
>>> -static void ublk_uring_cmd_cancel_fn(struct io_uring_cmd *cmd,
>>> -        unsigned int issue_flags)
>>> +static void ublk_uring_cmd_cancel_fn(struct io_uring_cmd *cmd)
>>>   {
>>>       struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
>>>       struct ublk_queue *ubq = pdu->ubq;
>>> @@ -1464,7 +1462,7 @@ static void ublk_uring_cmd_cancel_fn(struct io_uring_cmd *cmd,
>>>       io = &ubq->ios[pdu->tag];
>>>       WARN_ON_ONCE(io->cmd != cmd);
>>> -    ublk_cancel_cmd(ubq, io, issue_flags);
>>> +    ublk_cancel_cmd(ubq, io);
>>
>> .cancel_fn is always called with .uring_lock held, so this 'issue_flags' can't
>> be removed, otherwise double task run is caused because .cancel_fn
>> can be called multiple times if the request stays in ctx->cancelable_uring_cmd.
> 
> I see, that's exactly why I was asking whether it can be deferred
> to tw. Let me see if I can get by without that patch, but honestly
> it's a horrible abuse of the ring state. Any ideas how that can be
> cleaned up?

I assume io_uring_try_cancel_uring_cmd() can run in parallel with
completions, so there can be two parallel calls calls to ->uring_cmd
(e.g. io-wq + cancel), which gives me shivers. Also, I'd rather
no cancel in place requests of another task, io_submit_flush_completions()
but it complicates things.

Is there any argument against removing requests from the cancellation
list in io_uring_try_cancel_uring_cmd() before calling ->uring_cmd?

io_uring_try_cancel_uring_cmd() {
	lock();
	for_each_req() {
		remove_req_from_cancel_list(req);
		req->file->uring_cmd();
	}
	unlock();
}

-- 
Pavel Begunkov