Public bug reported:

## Summary

Since this machine moved to 7.0.0-42/-43 it has hit a kernel panic 4
times in 10 days (~1 in every 10 boots). Each time the kernel Oopses
instantly (Ubuntu's `panic_on_oops=1`), kdump writes a vmcore and the
machine reboots. Two different MM/folio code paths are involved.

Dumps present in `/var/crash`: `202610311135`, `202611010006`,
`202611041154`, `202611101949`.

## Latest Oops (2026-11-10 19:45:52 +0800, uptime 91 s, kernel
7.0.0-43-generic)

```
BUG: unable to handle page fault for address: 000000000002a200
#PF: supervisor read access in kernel mode
#PF: error_code(0x0000) - not-present page
PGD 1455aa067 P4D 1455aa067 PUD 1455a9067 PMD 0
Oops: Oops: 0000 [#1] SMP NOPTI
CPU: 1 UID: 1000 PID: 2913 Comm: Thread-1 (get_m  Tainted: G           OE       
7.0.0-43-generic #43-Ubuntu PREEMPT(lazy)
Tainted: [O]=OOT_MODULE, [E]=UNSIGNED_MODULE
Hardware name: Dell Inc. Precision 5760/0T7J1J, BIOS 1.39.0 03/31/2026
RIP: 0010:lruvec_stat_mod_folio+0x50/0x100
Code: 1c c5 a0 e9 5f a5 e8 cf d1 c7 ff 49 8b 54 24 38 48 89 d0 48 83 e0 f8 83 
e2 02 74 04 48 8b 40 10 49 63 d5 48 85 c0 74 63 66 90 <48> 63 8b 00 a2 02 00 4c 
8b a4 c8 80 06 00 00 49 83 c4 40 49 3b 9c
RSP: 0018:ffffd196ccf7f6f0 EFLAGS: 00010286
RAX: ffff8a2a0816a1c0 RBX: 0000000000000000 RCX: ffff8a2d7260b200
RDX: ffffffffffffffff RSI: 0000000000000011 RDI: fffff27119138640
R13: 00000000ffffffff R14: 0000000000000011 R15: fffff27119138640
CR2: 000000000002a200
Call Trace:
 <TASK>
 folio_remove_rmap_ptes+0x41/0x270
 zap_present_ptes.constprop.0+0x290/0x750
 zap_pte_range+0x1f5/0x780
 zap_pmd_range.isra.0+0xce/0x250
 unmap_page_range+0x23a/0x3d0
 unmap_single_vma+0x96/0x110
 unmap_vmas+0xa5/0x180
 unmap_region+0xa4/0x110
 vms_complete_munmap_vmas+0x1f7/0x280
 do_vmi_align_munmap+0x185/0x210
 do_vmi_munmap+0xd1/0x1b0
 __vm_munmap+0xc1/0x190
 __x64_sys_munmap+0x1b/0x30
 x64_sys_call+0x22a3/0x2390
 do_syscall_64+0x105/0x580
 entry_SYSCALL_64_after_hwframe+0x76/0x7e
 </TASK>
```

Reading of the registers: `RBX=0` and the faulting instruction is
`movslq 0x2a200(%rbx),%rcx`, i.e. a NULL base plus a member offset that
matches `struct mem_cgroup`; `memcg` obtained from the folio was NULL
while the path still dereferenced it. `R13=0xffffffff` (−1) is the stat
delta (`-nr_pages` for a single page). The `folio` pointer itself
(`R15=fffff27119138640`) is a vmemmap address, so this looks like rmap
accounting on a folio that is not charged to any memcg (device-private /
special mapping), reached through the plain `munmap()` →
`zap_pte_range()` path.

The faulting task is a Python thread (`Thread-1 (get_model)`) of a local
FunASR STT daemon; the `munmap()` being torn down is a ~928 MB region
(`RSI=0x39f3e000`), i.e. it happens while that daemon loads model
weights right after boot. The nvidia modules are loaded (taint OE), but
no nvidia frame appears anywhere in the stack.

## Earlier panic on the same kernel family (2026-10-31 11:35)

```
RIP: 0010:detach_extent_buffer_folio+0x19 [btrfs]
CR2=0x0000200000000018   eb->folios[i] garbage
btree_release_folio → filemap_release_folio → shrink_folio_list → balance_pgdat 
→ kswapd
```

So within 10 days two different folio/rmap/reclaim code paths died. Both
look like mm regressions rather than hardware (no MCE, no I/O errors, no
thermal events in dmesg; memory test not run yet, but the two signatures
are in different subsystems).

## System

- Ubuntu 26.04 (resolute), `7.0.0-43-generic` (7.0.14 base), 16 threads, 61 GiB 
RAM, 16 GiB zram
- Dell Precision 5760, BIOS 1.39.0
- root on btrfs (`space_cache=v2`, `discard=async`, single device, 712 GiB used 
/ 236 GiB free)
- nvidia 595.99.02 (OOT, tainted), virtualbox 7.2.6
- Reproducible only statistically: ~1 panic per 10 boots, always during heavy 
allocation/free right after boot (FunASR model loading) or during kswapd reclaim

## Why I cannot bisect downwards

`7.0.0-31-generic` (installed) boots, but DisplayPort-MST outputs are
dead on it: KWin enumerates `eDP-1`, `DP-9`, `DP-11` (the two external
monitors sit on i915 DP-MST), windows are placed on the MST outputs and
the screens stay black. So 7.0.0-31 is not usable here as a control.

## Secondary issue

`kdump-tools` logs `The kernel version is not supported. The
makedumpfile operation may be incomplete.` while saving the vmcore
(makedumpfile in the kdump initrd is older than the 7.0 kernel). dmesg
extraction still works; the compressed dump is 872 MB.

## Questions / request

1. Is this a known regression in 7.0.0-42/-43 (folio/rmap accounting, 
`lruvec_stat_mod_folio`)?
2. Is there a fix in `-proposed` I can test?
3. I can provide the full dmesg, `/proc/vmcore`-derived dump (872 MB) or run 
any debug kernel.

** Affects: linux (Ubuntu)
     Importance: Undecided
         Status: New

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2173885

Title:
  7.0.0-43-generic: NULL-pointer page fault in lruvec_stat_mod_folio()
  during munmap (folio_remove_rmap_ptes -> zap_pte_range), 4 panics in
  10 days

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2173885/+subscriptions


-- 
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs

Reply via email to