ZFS on 10-STABLE r281159: programs, accessing ZFS pauses for minutes in state [*kmem arena]

[Don Lewis]

On 30 Jul, Konstantin Belousov wrote:
> On Thu, Jul 30, 2015 at 02:30:08PM +0300, Lev Serebryakov wrote:
>> Hello Freebsd-fs,
>> 
>> 
>>   I'm migrating my NAS from geom_raid5 + UFS to ZFS raidz. My main storage
>> is 5x2Tb HDDs. Additionaly, I have 2x3Tb HDDs attached to hold my data when
>> I re-make my main storage.
>> 
>>  So, I have now two ZFS pools:
>> 
>> ztemp mirror ada0 ada1 [both are 3Tb HDDS]
>> zstor raidz ada3 ada4 ada5 ada6 ada7 [all of them are 2Tb]
>> 
>>  ztemp contain one filesystem with 2.1Tb of my data. ztemp was populated
>>  with my data from old geom_raid5 + UFS installation via "rsync" and it was
>>  FAST (HDD-speed).
>> 
>>  zstor contains several empty file systems (one per user), like:
>> 
>> zstor/home/lev
>> zstor/home/sveta
>> zstor/home/nsvn
>> zstor/home/torrents
>> zstor/home/storage
>> 
>>  Deduplication IS TURNED OFF. atime is turned off. Record size set to 1M as
>> I have a lot of big files (movies, RAW photo from DSLR, etc). Compression is
>> turned off.
>> 
>>  When I try to copy all my data from temporary HDDs (ztemp pool) to my new
>> shiny RIAD (zstor pool) with
>> 
>> cd /ztemp/fs && rsync -avH lev sveta nsvn storage /usr/home/
>> 
>>  rsync pauses for tens of minutes (!) after several hundreds of files. ^T
>> and top shows state "[*kmem arena]". When I stop rsync with ^C and try to do
>> "zfs list" it waits forever, in state "[*kmem arena]" again.
> Show the output of sysctl debug.vmem_check.
> 
>> 
>>  This server is equipped with 6GiB of RAM.
>> 
>>  It looks FreeBSD contains bug about year ago which leads to this behavior,
>> but mailing lists says, that it was fixed in r272221, 10 months ago.

I think I may have gotten bitten by this yesterday on a fairly recent
10.2-PRERELEASE machine with 8 GB of RAM.  It's nominally a zfs-only
machine, but I had some data on a couple of UFS drives that I needed to
copy over to a zfs filesystem.  I connected one of the drives to a sata
to usb adapater and plugged it into the machine, then ran rsync to
transfer the contents of a ~100 GB filesystem.  I had a number of active
programs running, including a rather bloated firefox process that had
gobbled lots of ram.  In my case, arc stayed small (< 1 GB), inactive
memory was a couple of GB, and several GB of data got pushed to swap.
Free memory got very low, bouncing around in the 10's of MB for a while
before the machine locked.  It wasn't totally dead because my X11
desktop is configured in focus follow mouse mode and I could see the
window focus change when I moved the mouse around.  Eventually I did
something to provoke the window manager and/or the Xorg server into
locking up as well.  I wasn't able to switch to console mode.  I
eventually gave up and hit the reset button.

%sysctl debug.vmem_check
sysctl: unknown oid 'debug.vmem_check': No such file or directory

With the same set of processes running, but no UFS, this is what top
says about memory usage:

Mem: 1156M Active, 3403M Inact, 1682M Wired, 31M Cache, 1631M Free
ARC: 1129M Total, 588M MFU, 492M MRU, 54K Anon, 10M Header, 39M Other
Swap: 40G Total, 40G Free