Re: wip/glusterfs to pkgsrc filesystems/glusterfs

[Greg Troxel <gdt%lexort.com@localhost>]

Alexander Schreiber <als%thangorodrim.ch@localhost> writes:

>> Is this new in glusterfs, that it doesn't work on i386, or just new that
>> you figured it out and are documenting it?  For distributed/remote
>> filesystems, I tend to feel its important for them to run everywhere
>> since in my view that's a major part of suitability -- but I realize
>> that if upstream is 64-bit only, it's next to impossible to tilt against
>> that.
>
> Support for 32bit was effectively dropped by glusterfs a while ago,
> there is https://github.com/gluster/glusterfs/issues/702 from 2019 that
> starts proposing this. I know that glusterfs 8.2 (current pkgsrc) works
> on NetBSD/i386, but somewhere between 8.2 and 10.3 (the one I'm working
> on for pkgsrc), 32bit support effectively disappeared. You can still
> _build_ it on 32bit, but it will die with SIGILL very early after
> startup due to
> https://github.com/gluster/glusterfs/issues/3911#issuecomment-1451711686
> (found via trawling through backtraces and then through left behind
> WRKOBJDIR contents due to the way userspace-rcu is integrated at
> build time).
>
> There are some sporadic discussions about "hey, why does this crash on
> 32bit?", but I do not expect anyone to actually put any work into
> making it work on !64bit again, partially because that would involve
> some ugly hackery (a bunch of internal things went from 32 to 64 bit
> and they do atomic updates on those).
>
> Even so, there were reports (from older versions) that 32bit glusterfs
> appears noticeably slower than 64bit.

I do wonder if it's possible to make userspace-rcu work on 32 bit
processors via using libatomic64, even if it's slower.

> My workaround for 32bit machines is to mount the glusterfs filesystem
> from a 64bit machine via sshfs from the 32bit machine. Not fast either,
> but at last it works at all (yes, I still have some 32bit hardware in
> active use with both NetBSD and Linux).

I have quite a few 32-bit machines running, 2 RPI3 (yes I know I could
run 64 on that hardware), 2 desktops, and a notebook that does builds
for the desktop.  To me, if I have to use ssh, then I might as well just
use ssh.

So I guess we have two paths:

  Update and accept that anybody using gluster on 32-bit CPUs is going to
  be broken.

  Version gluster and add gluster8 with the old code and make the new
  code gluster10.

(In either case the new code package's DESCR needs a note that it is
unusuable on 32-bit systems, so that people don't decide to rely on the
marketing copy and later find out it doesn't work.)

Do we know if anyone else is using gluster on 32-bit CPUs?

(Personally, this causes me to prune gluster from my TODO list, as
desupporting 32-bit CPUs does not seem reasonable.)