Re: hunting memory corruption bug in 6.18.x

[Nikola Ciprich <nikola.ciprich@linuxbox.cz>]

Some more news.

In the meantime, we’ve had the same crash on another of our own
cluster nodes. This time it was booted with tlbi=ipi, so that
didn’t help either, exactly as Borislav said.

(Just a side note: it seems quite hard to check which TLB
invalidation mechanism is in use. There is no message in dmesg,
and invlpgb is deliberately not displayed in the /proc/cpuinfo flags,
so apart from measuring with perf, I simply have to trust that
the kernel parameter does what it should for my kernel release.)

So if I understand it correctly, we can now completely rule out
an AMD firmware bug, right?

Since I didn’t have anything else to try, and we started having
crashes on our own cluster nodes, where I can be a bit bolder,
I built 7.3-rc6, deployed it on two problematic AMD nodes, and
started a migration torture loop.

Over the weekend, there were ~70 migrations of ~160 VMs between
those two nodes without a single issue.

Today I upgraded both boxes to rc7 and started the migrations again.

I can’t be sure yet, but it really seems the problems are gone in 7.3.

The question is: what now? (I still think bisecting would be unreliable.)

For those interested, I’ve tidied up the shared files a little and
added some more content to https://storage.linuxbox.cz/index.php/s/6rcp8oWCGzAqPEY:

6.18.55-lb9.01 – 6.18.55 patched source, patches and binaries
yyyymmdd-xxxxx – collected crash data for various occurrences
kernel-6.18.x-problems.xlsx – problems overview

Maybe it will be of some help.

BR

nik


On Fri, Oct 09, 2026 at 09:00:42AM +0100, David Laight wrote:
> On Thu, 8 Oct 2026 11:23:56 -0700
> Borislav Petkov <bp@alien8.de> wrote:
> 
> > On Thu, Oct 08, 2026 at 02:29:36PM +0200, Nikola Ciprich wrote:
> > > [11402.940943] BUG: unable to handle page fault for address: ffffffff0c93001c
> > > [11402.942273] #PF: supervisor read access in kernel mode
> > > [11402.943629] #PF: error_code(0x0000) - not-present page
> > > [11402.945025] PGD 6d6f83a067 P4D 6d6f83b067 PUD 0
> > > [11402.946469] Oops: Oops: 0000 [#1] SMP NOPTI
> > > [11402.947950] CPU: 23 UID: 0 PID: 704950 Comm: servercare-moni Kdump: loaded Tainted: G            E       6.18.55lb9.01 #1 PREEMPT(voluntary)
> > > [11402.951254] Tainted: [E]=UNSIGNED_MODULE
> > > [11402.952913] Hardware name: ASUSTeK COMPUTER INC. RS720A-E12-RS12/K14PP-D24 Series, BIOS 1201 08/25/2023
> > > [11402.956569] RIP: 0010:__d_lookup_rcu+0x4d/0xe0
> > > [11402.958452] Code: 48 8d 04 c2 f6 07 02 0f 85 a0 00 00 00 48 8b 10 48 89 d0 48 83 e0 fe 48 83 fa 01 77 0d e9 80 00 00 00 48 8b 00 48 85 c0 74 78 <44> 8b 58 fc 48 39 78 10 75 ee 48 83 78 08  
> > 
> > What is that kernel?
> > 
> > 6.18.55lb9.01
> > 
> > The other machine has a 6.18.20lb9.03-something one.
> > 
> > How can I look at the vmlinux you're running and the sources?
> > 
> > rIP points to:
> > 
> > [11402.958452] Code: 48 8d 04 c2 f6 07 02 0f 85 a0 00 00 00 48 8b 10 48 89 d0 48 83 e0 fe 48 83 fa 01 77 0d e9 80 00 00 00 48 8b 00 48 85 c0 74 78 <44> 8b 58 fc 48 39 78 10 75 ee 48 83 78 08
> > All code
> > ========
> >    0:   48 8d 04 c2             lea    (%rdx,%rax,8),%rax
> >    4:   f6 07 02                testb  $0x2,(%rdi)
> >    7:   0f 85 a0 00 00 00       jne    0xad
> >    d:   48 8b 10                mov    (%rax),%rdx
> >   10:   48 89 d0                mov    %rdx,%rax
> >   13:   48 83 e0 fe             and    $0xfffffffffffffffe,%rax
> >   17:   48 83 fa 01             cmp    $0x1,%rdx
> >   1b:   77 0d                   ja     0x2a
> >   1d:   e9 80 00 00 00          jmp    0xa2
> 
> That looks like the same hlist loop top as in the other failure.
> IIRC the fault has the error address in both %rax and %rdx which
> means the invalid value cane from the top of the hash list, not
> from following the linked list.
> 
> >   22:   48 8b 00                mov    (%rax),%rax
> >   25:   48 85 c0                test   %rax,%rax
> >   28:   74 78                   je     0xa2
> >   2a:*  44 8b 58 fc             mov    -0x4(%rax),%r11d         <-- trapping instruction
> >   2e:   48 39 78 10             cmp    %rdi,0x10(%rax)
> 
> That is the hash compare instruction (that fails in the other function).
> Not sure why it reads offset -4 first though - the compiler will have
> reordered it from after a condition (just to slow the code down!).
> I'd guess there is a container_of() lurking.
> 
> David
> 
> >   32:   75 ee                   jne    0x22
> >   34:   48                      rex.W
> >   35:   83                      .byte 0x83
> >   36:   78 08                   js     0x40
> > 
> > I need to be able to pinpoint it back to the source.
> > 
> > I asked the last time:
> > 
> > "Just to rule out any other issues which got fixed in the meantime, can you try
> > mainline Linux and see if you can reproduce your observation with it?
> > 
> > If so, you could share your crash core along with debug kernels yadda yadda so
> > that I can poke at it.
> > 
> > And before you do, make sure you have the latest BIOS and microcode installed on
> > that machine.
> > 
> > Also, where can I find full dmesg and /proc/cpuinfo from those machines which
> > trigger this?"
> > 
> > But still nothing.
> > 
> > Imagine this issue has been fixed upstream but you don't have the fix in your
> > kernels and we're basically chasing the same thing again...
> > 
> > Sorry, but I have lost my debugging crystal ball which can help me guess what
> > the machine does. :\
> > 
> > > so we now know this didn't fixed it. however I didn't have tlbi=ipi set, so I'll
> > > now try this.  
> > 
> > That won't help either but if you wanna try it.
> > 
> > > any ideas on this new info?  
> > 
> > Yes, see above.
> > 
> > Bottomline is: without sufficient debugging data and up-to-date hardware,
> > there's not a lot I can do.
> > 
> > Thx.
> > 
> 

-- 
Ing. Nikola CIPRICH
technický ředitel

+420 591 166 214
+420 777 093 799
nikola.ciprich@linuxbox.cz

www.linuxbox.cz