On Tue, Mar 4, 2025 at 3:08 PM Oleg Nesterov <oleg@xxxxxxxxxx> wrote: > > On 03/04, Mateusz Guzik wrote: > > > > @@ -529,10 +529,9 @@ anon_pipe_write(struct kiocb *iocb, struct iov_iter *from) > > > > if (!iov_iter_count(from)) > > break; > > - } > > > > - if (!pipe_full(head, pipe->tail, pipe->max_usage)) > > continue; > > + } > > Reviewed-by: Oleg Nesterov <oleg@xxxxxxxxxx> > thanks > It seems that we can also remove the unnecessary signal_pending() > check, but I need to recheck and we need to cleanup the poll_usage > logic first. > > This will also remove the unnecessary wakeups when the writer is > interrupted by signal/ > [snip] There are many touch ups to do here, I don't have an opinion about this diff. I don't have compiled stats handy, but few months back I asked some people to query pipe writes with dtrace on FreeBSD. Everything is very heavily skewed towards < 128 bytes in total, including tons of 1 bytes (and no, it's not just make). However, there is also quite a bit of absolutely massive multipage ops as well -- north of 64K even. Doing that kind of op in one go is way faster than restarting rep mov every time interleaved with SMAP toggling, which is expensive in its own right. Thus I would argue someone(tm) should do it in Linux, but I don't have immediate plans. Perhaps you would be happy to do it? :) In the meantime, I would hope any refactoring in the area would make the above easier (my patch does, I'm not claiming your does not :P) -- Mateusz Guzik <mjguzik gmail.com>
