Re: [PATCH net] tcp: make sure EPOLLOUT wont be missed

[Eric Dumazet <eric.dumazet@gmail.com>]

On 8/17/19 4:19 PM, Jason Baron wrote:
> 
> 
> On 8/17/19 12:26 AM, Eric Dumazet wrote:
>> As Jason Baron explained in commit 790ba4566c1a ("tcp: set SOCK_NOSPACE
>> under memory pressure"), it is crucial we properly set SOCK_NOSPACE
>> when needed.
>>
>> However, Jason patch had a bug, because the 'nonblocking' status
>> as far as sk_stream_wait_memory() is concerned is governed
>> by MSG_DONTWAIT flag passed at sendmsg() time :
>>
>>     long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
>>
>> So it is very possible that tcp sendmsg() calls sk_stream_wait_memory(),
>> and that sk_stream_wait_memory() returns -EAGAIN with SOCK_NOSPACE
>> cleared, if sk->sk_sndtimeo has been set to a small (but not zero)
>> value.
> 
> Is MSG_DONTWAIT not set in this case? The original patch was intended
> only for the explicit non-blocking case. The epoll manpage says:
> "EPOLLET flag should use nonblocking file descriptors". So the original
> intention was not to impact the blocking case. This seems to me like
> a different use-case.
>

I guess the problem is how we define 'non-blocking' ...

SO_SNDTIMEO can be used by application to implement a variation of non-blocking,
by waiting for a socket event with a short timeout, to maybe recover
from memory pressure conditions in a more efficient way than simply looping.

Note that the man page for epoll() only _suggests_ to use nonblocking file descriptors.

<quote>
       The  suggested  way  to use epoll as an edge-triggered (EPOLLET)
       interface is as follows:

              i   with nonblocking file descriptors; and

              ii  by  waiting  for  an  event  only  after  read(2)  or
                  write(2) return EAGAIN.
</quote>