Executive brief
The Linux kernel's RAID1 and RAID10 disk mirroring drivers contain a logic flaw in error handling that can cause system deadlocks under memory pressure. When a mirrored disk read fails and the I/O operation is retried, the code fails to correctly identify that it is in an error recovery path, causing memory allocation requests to be made without emergency reserves. This can starve the system of memory needed to complete the recovery, causing the system to hang or become unresponsive.
Technical details
The vulnerability is a logic error in the RAID1 and RAID10 drivers' read request handling functions. After a failed bio is split and resubmitted, the code previously relied on checking r1_bio (RAID1) or r10_bio->read_slot (RAID10) to detect the error path, but these pointers/values could be NULL or -1 respectively, causing the error condition to go unrecognized. This caused memory allocation to use GFP_NOIO instead of (GFP_NOIO | __GFP_HIGH), preventing emergency memory reserves from being accessed. Under memory pressure, this can trigger a deadlock since the RAID recovery thread itself is blocked waiting for memory. The fix uses md_cloned_bio() as a more reliable indicator of the error path, ensuring high-priority memory allocation flags are applied correctly.
Affected products
- Linux Linux kernel Likely 5.9 and later (based on referenced fixes from commits 689389a06ce7 and 545250f24809)
Timeline
- 2026-08-15: disclosed: Published to NVD
- 2026-05-01: patched: Fix committed upstream (commit 811545e0926d02a6a0b1a1258bb5544777c164d4)
- 2026-07-24: other: Fix included in stable release (commit 20fb582c92fd64e5c8bd83c6176264b2794603b7)