imposing memory limits in FreeBSD 8

[Joe Schaefer]

On Sat, Apr 9, 2011 at 3:06 PM, Artem Belevich <art at freebsd.org> wrote:
> On Sat, Apr 9, 2011 at 10:15 AM, Joe Schaefer <joesuf4 at gmail.com> wrote:
>> While I am thrilled about the newfound zfs stability that upgrading to 8
>> has brought, one of the things that seems to have been dropped is
>> support for process memory limits.  I have a few servers that occasionally
>> run out of swap due to runaway httpd daemons, and the ulimit -m settings
>> in the startup scripts we use stopped working upon upgrading from FreeBSD 6.
>>
>> I've tried fiddling with the daemon class in login.conf to no avail
>> either.  About
>> the only thing I haven't tried is running httpd under djb's softlimit
>> executable.
>> Here's my daemon class in login.conf:
>>
>> daemon:\
>>        :memoryuse=1g:\
>>        :datasize=1g:\
>>        :stacksize=1g:\
>>        :tc=default:
>>
>> and proof that `limits` groks the config:
>>
>> # limits -eHC daemon
>> ulimit -t unlimited;
>> ulimit -f unlimited;
>> ulimit -d 1048576;
>> ulimit -s 1048576;
>> ulimit -c unlimited;
>> ulimit -m 1048576;
>> ulimit -l unlimited;
>> ulimit -u unlimited;
>> ulimit -n unlimited;
>> ulimit -b unlimited;
>> ulimit -v unlimited;
>> ulimit -p unlimited;
>> ulimit -w unlimited;
>>
>> Any tips from admins who have successfully imposed memory constraints in 8.x?
>
> If I recall it correctly, in -8 malloc defaults to mmap for memory
> allocations, so RLIMIT_DATA no longer applies.
> You have to set RLIMIT_VMEM, but be careful as that would include
> everything mmapped in even if it does not use much of that. rpc.statd
> is one example of that -- it mmaps in ~256M but has only ~400K
> resident set size.
>
> Another option would be to make malloc() switch back to  sbrk() with
> MALLOC_OPTIONS=D. This way datasize limit will still be in effect.

Thanks for the tip.   My concern is with runaway processes that are pushing the
server into swap, so it's pretty easy to pick them out based on what top reports
for their SIZE.  I'll try the vmem limit and let you know how that works out.