Re: [regression] oops on heavy compilations ("kernel BUG at mm/zswap.c:1005!" and "Oops: invalid opcode: 0000")

[Takero Funaki <flintglass@gmail.com>]

2024年8月24日(土) 0:07 Piotr Oniszczuk <piotr.oniszczuk@gmail.com>:
>
>
>
> > Wiadomość napisana przez Matthew Wilcox <willy@infradead.org> w dniu 23.08.2024, o godz. 15:13:
> >
> > I wouldn't be surprised if this were dodgy ram.
>
>
> Well - that was my initial hypothesis.
>
> in fact i had few of them. Ranked (and ordered) like this:
> 1. downstream kernel patches
> 2. hw (ram) issue
> 3. kernel bug
>
> So full history was:
> -build myself archlinux 6.10.2 kernel; upgrade builder OS (only kernel; nothing else)
> -run normal devel process and (to my surprise) discover interrupted CI/CD builds by kernel oops
> -downgrade to 6.8.2 and done 4 full builds (full takes 8..9h of constant 12c/24/t compile). all good.
> -prepare vanilla 6.10.6 (to exclude potential downstream (ArchLinux) root causes)
> -run normal devel process and still discover oops
> -make sure hw is ok by week of test with 6.8.2 (recompiling for 3 architectures on 4 OS (3 in kvm). This was almost 5 full days of 12c/24 compiling. All good
> -because last steep was all good - decide to go to you :-)
>
> sure - this is possible that 6.8.2 had luck with my ram and 6.10.6 had no luck….but i personally don’t believe this is a case….
>
> btw: we can go with elimination strategy.
> So what i need to change/disable to be closer to finding root cause?
> swap?
> now it is swapfile on system nvme
>
>

Hello,

I’m encountering a similar crash and trace in the issue I posted on Bugzilla:
https://bugzilla.kernel.org/show_bug.cgi?id=219154

If this is the same issue caused by virtio_net corrupting memory, you
should be able to reproduce the crash by sending data to the VM over
virtio interface while it is actively allocating memory (e.g., using
iperf3 -s on the VM and running iperf -c from another host).

In my case, as Thorsten suggested, reverting bisected commit
f9dac92ba908 (virtio_ring: enable premapped mode regardless of
use_dma_api) along with two related commits in this series resolved
the issue:
https://lore.kernel.org/all/7774ac707743ad8ce3afeacbd4bee63ac96dd927.1723617902.git.mst@redhat.com/