On 10/4/22 2:59 AM, Fiona Ebner wrote:
> Am 26.08.22 um 10:21 schrieb Fiona Ebner:
>> Am 11.07.22 um 15:40 schrieb Fabian Ebner:
>>> Am 09.07.22 um 05:39 schrieb Shyam Prasad N:
>>>> On Sat, Jul 9, 2022 at 9:00 AM Shyam Prasad N <nspmangalore@xxxxxxxxx> wrote:
>>>>>
>>>>> On Fri, Jul 8, 2022 at 11:22 PM Enzo Matsumiya <ematsumiya@xxxxxxx> wrote:
>>>>>>
>>>>>> On 07/08, Fabian Ebner wrote:
>>>>>>> (Re-sending without the log from the older kernel, because the mail hit
>>>>>>> the 100000 char limit with that)
>>>>>>>
>>>>>>> Hi,
>>>>>>> it seems that in kernels >= 5.15, io_uring and CIFS don't interact
>>>>>>> nicely sometimes, leading to IO errors. Unfortunately, my reproducer is
>>>>>>> a QEMU VM with a disk on CIFS (original report by one of our users [0]),
>>>>>>> but I can try to cook up something simpler if you want.
>>>>>>>
>>>>>>> Bisecting got me to 8ef12efe26c8 ("io_uring: run regular file
>>>>>>> completions from task_work") being the first bad commit.
>>>>>>>
>>
>> I finally got around to taking another look at this issue (still present
>> in 5.19.3) and I think I've finally figured out the root cause:
>>
>> After commit 8ef12efe26c8, for my reproducer, the write completion is
>> added to task_work with notify_method being TWA_SIGNAL and thus
>> TIF_NOTIFY_SIGNAL is set for the task.
>>
>> After that, if we end up in sk_stream_wait_memory() via sock_sendmsg(),
>> signal_pending(current) will evaluate to true and thus -EINTR is
>> returned all the way up to sock_sendmsg() in smb_send_kvec().
>>
>> Related: in __smb_send_rqst() there too is a signal_pending(current)
>> check leading to the -ERESTARTSYS return value.
>>
>> To verify that this is the cause, I wasn't able to trigger the issue
>> anymore with this hack applied (i.e. excluding the TIF_NOTIFY_SIGNAL check):
>>
>>> diff --git a/net/core/stream.c b/net/core/stream.c
>>> index 06b36c730ce8..58e3825930bb 100644
>>> --- a/net/core/stream.c
>>> +++ b/net/core/stream.c
>>> @@ -134,7 +134,7 @@ int sk_stream_wait_memory(struct sock *sk, long *timeo_p)
>>> goto do_error;
>>> if (!*timeo_p)
>>> goto do_eagain;
>>> - if (signal_pending(current))
>>> + if (task_sigpending(current))
>>> goto do_interrupted;
>>> sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
>>> if (sk_stream_memory_free(sk) && !vm_wait)
>>
>>
>> In __cifs_writev() we have
>>
>>> /*
>>> * If at least one write was successfully sent, then discard any rc
>>> * value from the later writes. If the other write succeeds, then
>>> * we'll end up returning whatever was written. If it fails, then
>>> * we'll get a new rc value from that.
>>> */
>>
>> so it can happen that collect_uncached_write_data() will (correctly)
>> report a short write when calling ctx->iocb->ki_complete().
>>
>> But QEMU's io_uring backend treats a short write as an -ENOSPC error,
>> which also is a bug? Or does the kernel give any guarantees in that
>> direction?
>>
>> Still, it doesn't seem ideal that the "interrupt" happens and in fact
>> __smb_send_rqst() tries to avoid it, but fails to do so, because of the
>> unexpected TIF_NOTIFY_SIGNAL:
>>> /*
>>> * We should not allow signals to interrupt the network send because
>>> * any partial send will cause session reconnects thus increasing
>>> * latency of system calls and overload a server with unnecessary
>>> * requests.
>>> */
>>>
>>> sigfillset(&mask);
>>> sigprocmask(SIG_BLOCK, &mask, &oldmask);
>>
>> Do you have any suggestions for how to proceed?
>>
>
> Ping. The issue is still present in Linux 6.0. Does it make sense to
> also temporarily unset the task's TIF_NOTIFY_SIGNAL here or is that a
> bad idea?
You could try setting up with ring with IORING_SETUP_COOP_TASKRUN,
that'll avoid the TIF_NOTIFY_SIGNAL bits.
--
Jens Axboe
