CVE-2026-93264: RDMA/efa: Fix PBL chunk length computation
In the Linux kernel, the following vulnerability has been resolved:
RDMA/efa: Fix PBL chunk length computation
On register MR, when creating the PBL, if it's an indirect PBL we create a chunk list to hold the PBL pages pointers. Each chunk is 4KB in size and can hold 510 addresses (EFAPTRSPERCHUNK) and has a 12-byte control buffer at the end of it holding the next chunk's pointer and its length.
If the PBL number of pages is a multiple of EFAPTRSPERCHUNK, the calculated last chunk length is wrongly computed as 0, even though that chunk is fully populated with 510 real page pointers. This wrong length is used both to DMA map the chunk and is propagated to the device, causing the device to see the chunk as empty and reject the memory registration.
Fix the calculation so it will be performed only if the number of pages isn't a multiple of EFAPTRSPERCHUNK, if it is, its already handled in the above loop correctly. Also prevent out-of-bounds reach in the chunks array in such scenario.
Affected Software
Event History
Frequently Asked Questions
Which systems are exposed to this issue?
Systems using the Linux kernel EFA RDMA driver and registering memory regions that require an indirect PBL are affected when the PBL page count is an exact multiple of 510. The issue occurs during memory registration rather than for all EFA driver activity.
What is the practical impact when the condition is triggered?
The final PBL chunk can be treated as empty because its length is computed as zero. The device can then reject the memory registration, and the faulty handling can also reach beyond the chunks array.
How can I determine whether a failure may be related to this issue?
Investigate failed EFA memory-registration operations where the indirect PBL contains a page count divisible by 510. This condition causes the final fully populated chunk to be supplied with a zero length.
Is there a workaround if an updated kernel cannot be deployed immediately?
Avoid memory-registration layouts that produce an indirect PBL page count exactly divisible by 510, where feasible. The corrected kernel logic preserves the full length for such final chunks and prevents the out-of-bounds array access.