On 2/12/21 1:15 PM, Linus Torvalds wrote:
> On Fri, Feb 12, 2021 at 12:10 PM Jens Axboe <axboe@xxxxxxxxx> wrote:
>>
>> Right, it's exactly the AF_UNIX dgram case. Working on adding some checks
>> that means we'll catch this sort of thing upfront while testing.
>
> You might also just add a comment to the IORING_OP_{SEND,RECV}MSG
> cases to the work-flags.
>
> It doesn't hurt to just mention those kinds of things explicitly.
>
> Because maybe somebody decides that IO_WQ_WORK_FS is very expensive
> for their workload. With the comment they might then be able to say
> "let's set it only for the AF_UNIX case" or some similar optimization.
>
> Yeah, it probably doesn't matter, but just as a policy, I think "we
> got this wrong, so let's clarify" is a good idea.
I'll start with the comment, that's certainly the easiest part and
will help catch changes like that in the future. My other idea was
to add a check in path resolution that catches it, for future safe
guards outside of send/recvmsg. That's obviously a separate change
from the comment, but would be nice to have.
--
Jens Axboe
