Re: [PATCH] af_unix: limit unix_tot_inflight

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

Le mercredi 24 novembre 2010 à 15:44 +0100, Andi Kleen a écrit :
> Eric Dumazet <eric.dumazet@gmail.com> writes:
> >
> > diff --git a/net/unix/garbage.c b/net/unix/garbage.c
> > index c8df6fd..40df93d 100644
> > --- a/net/unix/garbage.c
> > +++ b/net/unix/garbage.c
> > @@ -259,9 +259,16 @@ static void inc_inflight_move_tail(struct unix_sock *u)
> >  }
> >  
> >  static bool gc_in_progress = false;
> > +#define UNIX_INFLIGHT_TRIGGER_GC 16000
> 
> It would be better to define this as a percentage of
> lowmem.
> 

I knew somebody would suggest this ;)

Hmm, why bother ?

Do you think 16000 is too big ? Too small ?

1) What would be the percentage of memory ? 1%, 0.001 % ?

  On a 16TB machine, a percentage will still give huge latencies to the
poor guy that hit the unix_gc().

With 16000, the max latency I had was 11.5 ms (on an Intel E5540
@2.53GHz), instead of more than 2000 ms

I guess it would make more sense to limit to the size of cpu cache
anyway.


2) We currently allocate 4096 bytes (on x86_64) to store one file
pointer, or 2048 bytes on x86_32.

But we can store in it up to 255 files.

 I posted a patch to shrink this to 32 or 16 bytes. Should we then
change the heuristic ?

3) Really who needs more than 16000 inflight unix files ?

  (inflight unix files means : af_unix file descriptors that were sent
(sendfd()) through af_unix, not yet garbage collected.).


4) If we autotune a limit at boot time as a lowmem percentage, some guys
then want a /proc/sys/net/core/max_unix_inflight sysctl , just for
completeness. One extra sysctl... 

I cant see valid uses but programs designed to stress our stack.