Re: aarch64 reliability issues

[email protected] (Michael van Elst) writes:

> [email protected] (Greg Troxel) writes:
>
>>  There's no pressure mechanism to free vnodes from zfs under memory
>>  pressure, and those vnodes still hold refs to other memory.
>
> Yes. That's also true for any other filesystem.

Agreed, but with other filesystems, it seems that the vnodes are just
vnodes, and there are some number of them such that total RAM caused by
vnodes is maybe 5%, waving hands wildly.  With ZFS, it seems (this is
now paged out of my head) other structures can be hung off the znode or
some such, so that signficant ram in the ARC is unfreeable.

>>The workaround is to set kern.maxvnodes low. and to be gentle on the
>>combination of dirty pages and memory pressure.  Pro Tip: Don't run
>>syncthing on GB of stuff while doing heavy builds.
>
> I have a local patch to add pressure to maxvnodes when there is
> a memory shortage. But it's a bit difficult to exercise.

That's very interesting.  I would be interested in trying it, and I
suspect some of the other zfs-using people would too, if you are
inclined to post it (tech-kern?).