Re: zfs questions

[Takashi YAMAMOTO <yamt9999%gmail.com@localhost>]

hi,

On Sat, Feb 21, 2026 at 9:30 PM Brad Spencer <brad%anduin.eldar.org@localhost> wrote:
>
> Takashi YAMAMOTO <yamt9999%gmail.com@localhost> writes:
>
> > On Sat, Feb 21, 2026 at 9:34 AM Brad Spencer <brad%anduin.eldar.org@localhost> wrote:
> >>
> >> Greg Troxel <gdt%lexort.com@localhost> writes:
> >>
> >> > Takashi YAMAMOTO <yamt9999%gmail.com@localhost> writes:
> >> >
> >> >> * is anyone using zfs in netbsd seriously? :-)
> >> >
> >> > I am using it on my desktop, and a on a dom0 hosting pkg builders.
> >> >
> >> > It mostly works, except that one must be careful not to simultaneously
> >> >
> >> >   run low on memory
> >> >   run programs that dirty mmap'd pages (syncthing!)
> >> >
> >> > and probably also not
> >> >
> >> >   run really low on memory such that there is a lot of paging
> >> >
> >>
> >> This is a reaonsably summary of the quirks.
> >
> > what's the symptom of the issues? do you have PRs?
>
> The system will hang up, either very hard or just decay into a unusable
> state.  I reduced the maxvnodes is a very low value and that made it
> sometimes possible to use ps and one will notice one or more processes
> that are claiming that some form of memory isn't available, flt_noram5
> is pretty common, but not the only state I have seen.  These processes
> will not be killable.
>
> I don't think I created a PR specifically about what I see.  kern/57558
> likely covers it, however.
>
> I trigger the situation by doing 3 or 4 parallel builds of the OS,
> "build.sh -someargs -j2", one a DOMU with 2 vcpus and 16GB (it is
> assumed that the system will page hard at times, which is fine).
> However, even if I do only 1 build over and over (actually only 3 or 4
> times), the system will eventually put itself into the same state.
> Something builds up over time or memory fragments in a manor that makes
> it unusable, I suspect.  The system is backed up using zfs snapshots and
> even if I am not doing builds the system will decay into the same state,
> but that tends to take a very long time (many months).  The daily cron
> jobs are mostly disabled as the ones that use find (hunting down core
> files, etc) can hang up the system when they cross into the zfs
> filesets, especially if they happen to run when a build it going on.
>
> Fairly easy to trigger, it just takes a while for everything to grind to
> a halt so one has to be patience about it all.

didn't the last patch in PR/57558 improve the situation for you?
when it happens, what are relevant threads (pgdaemon and others) doing?

>
>
> Currently on a Xen DOMU PVH guest you can't get a kernel crash dump. so
> that is a frustration point to figuring out what is happening.
>
>
> (Been dealing with this for a long time, and a number of others know
> about it)
>
>
>
>
>
> --
> Brad Spencer - brad%anduin.eldar.org@localhost
>