[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: 64bit support news



Hi:
 
The main concern is that it looks like swapping doesn't work well with
rumpdisk, perhaps something like rumpdisk or pci-arbiter not managing
to wire themselves completely, and thus getting a hang if any part
gets swapped out. This means I had to disable swap entirely to keep
the box a bit more stable. But then when building packages that are a
bit demanding in memory we end up with lack of memory and other stuck
conditions (the buildd box has only 3G memory).

This problem is not only in 64-bit. It seems to be a rumpdisk generalized problem, because in i386 and in smp I had the same issue.
I want to try to update rumpkernel sources, to add better hardware support in rumpdisk, and then we can test if the problem keeps in the new version.

Also, it looks like we are encountering deadlocks in ext2fs/libdiskfs
that we had not seen (most probably only by sheer luck) on i386.

 This could be the cause of some smp freezes during the boot. We have to check race conditions in rumpdisk and ext2fs.

It necessary to check these servers, to find some possible race conditions and try to fix them


El jue, 1 ago 2024 a las 0:46, Samuel Thibault (<samuel.thibault@gnu.org>) escribió:
Samuel Thibault, le jeu. 01 août 2024 00:14:35 +0200, a ecrit:
> We are then seeing build failures, see
> https://people.debian.org/~sthibault/hurd-amd64/failed.txt

Ah, I forgot to talk about the general stability of the port.

The main concern is that it looks like swapping doesn't work well with
rumpdisk, perhaps something like rumpdisk or pci-arbiter not managing
to wire themselves completely, and thus getting a hang if any part
gets swapped out. This means I had to disable swap entirely to keep
the box a bit more stable. But then when building packages that are a
bit demanding in memory we end up with lack of memory and other stuck
conditions (the buildd box has only 3G memory). Merely e.g. unpacking
texlive-latex-extra can lock the box. For some packages that we really
want, I built them by hand on a vm on a server that has a lot of memory.

Also, it looks like we are encountering deadlocks in ext2fs/libdiskfs
that we had not seen (most probably only by sheer luck) on i386.

All in all, I have to reboot the buildd box several times a day
(while the i386 buildds can compile packages for several weeks before
encountering hangs).

Samuel


Reply to: