Re: [regression] BUG: KASAN: use-after-free in lockdep_register_key+0x755/0x8f0

[Kalle Valo]

Kalle Valo <kvalo@xxxxxxxxxx> writes:

> Kalle Valo <kvalo@xxxxxxxxxx> writes:
>
>> Yesterday I run our ath11k regression tests with v6.10-rc1 and our
>> simple ath11k module reload stress started failing reliably with various
>> KASAN errors. The test removes and inserts ath11k and other wireless
>> modules in a loop. Usually I run it at least 100 times, some times even
>> more, and no issues until yesterday.
>>
>> I have verified that the last wireless-next pull request (tag
>> wireless-next-2024-05-08) works without issues and v6.10-rc1 fails
>> always, usually within 50 module reload loops. From this I'm _guessing_
>> that we have a regression outside wireless, most probably introduced
>> between v6.9 and v6.10-rc1. But of course I cannot be sure of anything
>> yet.
>>
>> I see different KASAN warnings and lockdep seems to be always visible in
>> the stack traces. I think I can reproduce the issue within 15 minutes or
>> so. Before I start bisecting has anyone else seen anything similar? Or
>> any suggestions how to debug this further?
>>
>> I have included some crash logs below, they are retrieved using
>> netconsole. Here's a summary of the errors:
>>
>> [ 159.970765] KASAN: maybe wild-memory-access in range
>> [0xbbbbbbbbbbbbbbb8-0xbbbbbbbbbbbbbbbf]
>> [  700.017632] BUG: KASAN: use-after-free in lockdep_register_key+0x755/0x8f0
>> [  224.695821] BUG: KASAN: slab-out-of-bounds in lockdep_register_key+0x755/0x8f0
>> [  259.666542] BUG: KASAN: slab-use-after-free in lockdep_register_key+0x755/0x8f0
>
> I did a bisect and got this:
>
> cf29111d3e4a9ebe1cbe2b431274718506d69f10 is the first bad commit
> commit cf29111d3e4a9ebe1cbe2b431274718506d69f10
> Merge: ed11a28cb709 e6f7d27df5d2
> Author: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
> Date:   Thu May 16 18:14:11 2024 -0500
>
>     Merge branch 'pci/of'
>     
>     - Check for kcalloc() failure and handle it gracefully (Duoming Zhou)
>     
>     * pci/of:
>       PCI: of_property: Return error for int_map allocation failure
>
>  drivers/pci/of_property.c | 2 ++
>  1 file changed, 2 insertions(+)
>
> But that doesn't make any sense to me, I don't even have
> CONFIG_PCI_DYNAMIC_OF_NODES enabled in my .config. I guess I did a
> mistake during bisect, I'm now testing the parents (e6f7d27df5d2 and
> ed11a28cb709) and trying to pinpoint where I did it wrong.

I found my mistake and was able to finish the bisect. This seems to be
the commit causing my problems:

# first bad commit: [7e89efc6e9e402839643cb297bab14055c547f07] PCI: Lock upstream bridge for pci_reset_function()

I verified by reverting that commit on top of v6.10-rc1 and I have not
seen any crashes so far, normally I would have seen it by now. But I
will continue testing the revert just to be sure.

Adding people and lists involved with that commit. Here is my original
report:

https://lore.kernel.org/all/87v82y6wvi.fsf@xxxxxxxxxx/

-- 
https://patchwork.kernel.org/project/linux-wireless/list/

https://wireless.wiki.kernel.org/en/developers/documentation/submittingpatches