Re: RFC: How ZFS handles arc memory use

On Wed, Mar 04, 2026 at 11:03:42AM +0100, Alexander Leidinger wrote:
> Am 2026-03-03 23:45, schrieb Doug Ambrisko:
> > On Tue, Mar 03, 2026 at 02:25:11PM -0800, Rick Macklem wrote:
> > | On Tue, Mar 3, 2026 at 12:33 PM Doug Ambrisko <ambrisko@ambrisko.com>
> > wrote:
> > | >
> > | > On Sun, Nov 02, 2025 at 11:48:06AM +0100, Alexander Leidinger wrote:
> > | > | Am 2025-10-29 22:06, schrieb Doug Ambrisko:
> > | > | > It seems around the switch to OpenZFS I would have arc clean
> > task
> > | > | > running
> > | > | > 100% on a core.  I use nullfs on my laptop to map my shared ZFS
> > /data
> > | > | > partiton into a few vnet instances.  Over night or so I would
> > get into
> > | > | > this issue.  I found that I had a bunch of vnodes being held by
> > other
> > | > | > layers.  My solution was to reduce kern.maxvnodes and
> > vfs.zfs.arc.max so
> > | > | > the ARC cache stayed reasonable without killing other
> > applications.
> > | > | >
> > | > | > That is why a while back I added the vnode count to mount -v so
> > that
> > | > | > I could see the usage of vnodes for each mount point.  I made a
> > script
> > | > | > to report on things:
> > | > |
> > | > | Do you see this also with the nullfs mount option "nocache"?
> > | >
> > | > I seems to have run into this issue with nocache
> > | >   /data/jail/current/usr/local/etc/cups
> > /data/jail/current-other/usr/local/etc/cups     nullfs rw,nocache 0 0
> > | >   /data/jail/current/usr/local/etc/sane.d
> > /data/jail/current-other/usr/local/etc/sane.d   nullfs rw,nocache 0 0
> > | >   /data/jail/current/usr/local/www
> > /data/jail/current-other/usr/local/www          nullfs rw,nocache 0 0
> > | >   /data/jail/current/usr/local/etc/nginx
> > /data/jail/current-other/usr/local/etc/nginx    nullfs rw,nocache 0 0
> > | >   /data/jail/current/tftpboot
> > /data/jail/current-other/tftpboot               nullfs rw,nocache 0 0
> > | >   /data/jail/current/usr/local/lib/grub
> > /data/jail/current-other/usr/local/lib/grub     nullfs rw,nocache 0 0
> > | >   /data/jail
> > /data/jail/current-other/data/jail              nullfs rw,nocache 0 0
> > | >   /data/jail
> > /data/jail/current/data/jail                    nullfs rw,nocache 0 0
> > | >
> > | > After a while (a couple of months or more).  My laptop was running
> > slow
> > | > with a high load.  The perodic find was running slow.  arc_prunee
> > was
> > | > spinning.  When I reduced the number of vnodes then things got
> > better.
> > | > My vfs.zfs.arc_max is 1073741824 so that I have memory for other
> > things.
> > | >
> > | > nocache does help taking longer to get into this situation.
> > | Have any of you guys tried increasing vfs.zfs.arc.free_target?
> > |
> > | If I understand the code correctly, when freemem <
> > vfs.zfs.arc.free_target
> > | the reaper thread (the one that does uma_zone_reclaim() to return
> > pages
> > | to the system from the uma keg that the arc uses) should be activated.
> > 
> > I haven't tried that.  I set:
> > 	kern.maxvnodes
> > 	vfs.zfs.arc.min
> > 	vfs.zfs.arc.max
> > 	vfs.zfs.prefetch.disable=1
> > 
> > I need to make sure kern.maxvnodes is small enough so it doesn't thrash
> > when vfs.zfs.arc.max set to 1G.  The issues tend to take a while to
> > happen.  On the plus side I can adjust these when I hit them mostly by
> > reducing kern.maxvnodes without having to do a reboot.
> 
> There was this commit recently_
> https://cgit.freebsd.org/src/commit/sys/fs/nullfs?id=8b64d46fab87af3ae062901312187f3a04ad2d67
> 
> I have not checked if this race condition could result in anything related
> to what we see. From the commit message I can not deduct if this could for
> example lead to a (even temporary) resource leak which may explain this
> behavior. Mark, what is the high-level result of this race condition you
> fixed in nullfs? At first look at the commit log I would rather assume
> vnodes of the lower FS could rather be freed more early and not at all
> because of the race condition.

The high-level result would be a lock leak and presumably an eventual
deadlock or crash.  In an INVARIANTS kernel you'd get an assertion
failure.  I doubt the bug can be responsible for the issues reported in
this thread.