To make the main difference between the two kernels explicit (it is only
one bullet in the description):

- 6.8 sends at most 1280 KiB per disk write 
(/sys/block/sdb/queue/max_sectors_kb = 1280).
- 7.0 sends writes of up to 4 MiB (max_sectors_kb = 4096). max_hw_sectors_kb is 
32767 on both.
- Every failed command is a 4 MiB WRITE FPDMA QUEUED (52 of 52 in the 
2026-11-11 boot), and every DMAR fault address is 4 MiB-aligned. The fault is 
"DMA Read" because the AHCI controller reads the write data from RAM.
- The IOMMU runs in the same mode (Translated, lazy) on both kernels, so the 
IOMMU itself is not new; the 4 MiB requests are.

Hypothesis (not verified): on 7.0 the 4 MiB write buffers get an IOMMU
mapping (possibly a 2 MiB superpage, or some other PTE format) that this
Haswell VT-d rejects with fault reason 0x0c. The DMA fails, AHCI reports
a host bus error, the write fails, and ext4 remounts / read-only. 6.8
never issues requests that large, so it never hits this.

If it helps, I can reinstall 7.0 and test one of these, each on its own:
1. intel_iommu=sp_off (would confirm or refute the superpage idea)
2. max_sectors_kb capped at 1280 via a udev rule (would confirm that request 
size is the trigger)

(Agent disclosure: Claude Code was used to help me formulate and verify
this bug report and this comment)

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2174419

Title:
  7.0 HWE: DMAR fault "non-zero reserved fields in PTE" on Intel 9
  Series AHCI [8086:8c82] causes SATA host bus errors, failed writes and
  ext4 read-only remount (6.8 unaffected)

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux-hwe-7.0/+bug/2174419/+subscriptions


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

Reply via email to