Package: src:linux
Version: 6.1.192-1
Severity: critical
Tags: regression upstream
After upgrading to linux-image-6.1.0-55-amd64 (6.1.192-1), two of four bookworm
servers hard-hung: kernel still answers ping, all userspace frozen, ACPI
shutdown ignored, nothing written to the journal. One crashed ~160 s after
boot, the other after ~4 hours. Both are stable again on 6.1.0-54-amd64
(6.1.190-1).
Affected: one Hetzner Cloud KVM guest (virtio) and one bare-metal Intel
i9-13900 with md RAID1 on NVMe + SATA, so this is not hardware-specific. All
filesystems are ext4. The hang occurs from a deletion (rm) path, consistent
with page cache pressure evicting the buddy cache.
Console of the KVM guest:
RIP: 0010:ext4_mb_load_buddy_gfp+0xbc/0x380 [ext4]
Code: ... <48> 8b 00 ...
RAX: 0000000000000000
CR2: 0000000000000000
note: rm[5397] exited with irqs disabled
followed by journald and timesyncd watchdog timeouts and an unkillable system.
This appears to be the known 6.1.y backport bug: 6.1.191 backported the mballoc
folio conversion but kept IS_ERR() checks, while __filemap_get_folio() in 6.1
returns NULL. When the buddy cache pages are evicted, the next allocation or
free dereferences NULL. An upstream 6.1.y fix ("ext4: fix NULL folio
dereference in mballoc") reportedly exists but is not in 6.1.192.
References:
- https://github.com/sporevm/kernels/pull/13
- https://lkml.iu.edu/2603.3/01511.html (same oops bisected to "ext4: convert
bd_bitmap_page to bd_bitmap_folio" during 6.1.167-rc1 review)
Since 6.1.192-1 also carries the CVE-2026-99288 fix, users who roll back to
6.1.190-1 to avoid the hang lose that fix. Could the ext4 fix be cherry-picked
into the next bookworm-security upload?
Thanks.
Ioannis Dressos
Information Technology Department <https://aviet.aero/>
m: +30 698 284 5673
e: [email protected] AVIET LTD
Airways Building
1 Triq Ganni Vassallo
Luqa, Malta LQA1510
www.aviet.aero <https://aviet.aero/>