8.2 + apache == a LOT of sigprocmask

[Daniil Cherednik]

Hi.
I was trying to understand a cause for this problem, and made an ugly hack:
diff -u ./rtld_lock.c.orig ./rtld_lock.c
--- ./rtld_lock.c.orig 2011-11-15 07:56:14.000000000 +0000
+++ ./rtld_lock.c 2011-11-15 07:54:42.000000000 +0000
@@ -118,7 +118,7 @@
sigset_t tmp_oldsigmask;

for ( ; ; ) {
- sigprocmask(SIG_BLOCK, &fullsigmask, &tmp_oldsigmask);
+// sigprocmask(SIG_BLOCK, &fullsigmask, &tmp_oldsigmask);
if (atomic_cmpset_acq_int(&l->lock, 0, WAFLAG))
break;
sigprocmask(SIG_SETMASK, &tmp_oldsigmask, NULL);
@@ -135,7 +135,7 @@
atomic_add_rel_int(&l->lock, -RC_INCR);
else {
atomic_add_rel_int(&l->lock, -WAFLAG);
- sigprocmask(SIG_SETMASK, &oldsigmask, NULL);
+// sigprocmask(SIG_SETMASK, &oldsigmask, NULL);
}
}

this reduced number of syscalls, but I am not sure about stability and 
correctness of this hack.
And performance problems of apache remained (this hack only gave 5-10% 
increase of performance).
Also I found problem in longjmp syscalls. In FreeBSD we are doing this 
syscalls from gen/setjmp.S without condition check,
for example in Linux we are doing it only if stack has been saved.
Linux code:
if (env[0].__mask_was_saved)
/* Restore the saved signal mask. */
(void) __sigprocmask (SIG_SETMASK, &env[0].__saved_mask,
(sigset_t *) NULL);

After all that I was trying to compare perfomance of return from fork() 
in Linux and FreeBSD (see 
http://lists.freebsd.org/pipermail/freebsd-hackers/2011-October/036705.html) 
and fork() in FreeBSD was slower.

I am not trying to start a holy war, but I really need to increase 
performance of our hosting in FreeBSD.

On 15.11.2011 01:59, Doug Barton wrote:
> On 11/14/2011 12:56, John Baldwin wrote:
>> On Monday, November 14, 2011 3:31:43 pm Doug Barton wrote:
>>> Trying to track down a load problem we're seeing on 8.2-RELEASE-p4 i386
>>> in a busy web hosting environment I came across the following post:
>>>
>>> http://lists.freebsd.org/pipermail/freebsd-questions/2011-
>> October/234520.html
>>> That basically describes what we're seeing as well, including the
>>> "doesn't happen on Linux" part.
>>>
>>> Does anyone have any ideas about this?
>>>
>>> With incredibly similar stuff running on 7.x we didn't see this problem,
>>> so it seems to be something new in 8.
>> I suspect it has to do with some of the changes to rtld such that it now
>> always blocks signals while resolving symbols (or something along those
>> lines IIRC).  It makes throwing exceptions slow as well.
> The calls to sigprocmask() in rtld seem to be doing what you suggest
> here, but they involve setting and restoring the mask. In my followup
> post I pasted what we're seeing, which is different, and much more
> voluminous. For example, 13,500 calls in 30 seconds from a single apache
> worker process.
>
> Although this does seem to explain why our test cases have more calls
> when compiled with C++ than they do when compiled with C. :)
>
> Thanks for the response in any case.
>
>
> Doug
>


-- 

С уважением,
Daniil Cherednik
.masterhost