A fix committed by Linus Torvalds on August 21 UTC stops Intel’s Xe driver from offering reserved flat CCS storage to the VRAM allocator. On the reported Battlemage G21 system, that overlap corrupted a page table and prevented the desktop from starting.
Follow the memory boundary
The commit and explanation describe compression metadata overlapping memory offered for ordinary allocations. On Torvalds’s 16 GiB G21 card, the affected allocation was a Mesa page table; GDM kept restarting after the compositor failed.
The actual diff replaces rounding up to 128 KiB with rounding down to 4 KiB. It also replaces an equality check with a check that the CCS region does not overlap GSM. Saying the entire patch only swaps two function names leaves out both changes.
A small example makes the error visible
Imagine a separate, fictional memory map in which reserved storage starts at 130 KiB. An allocator with 4 KiB pages may safely use complete pages below 128 KiB. The page spanning 128-132 KiB straddles the boundary and must be excluded.
Rounding 130 KiB upward to 128 KiB alignment gives 256 KiB. Publishing that as the usable end incorrectly includes the reserved interval from 130 to 256 KiB. Rounding down to 4 KiB alignment gives 128 KiB, sacrificing 2 KiB of otherwise unreserved space to keep whole pages safe. These small addresses illustrate the invariant; they are not the G21’s actual addresses.
The required relationship is published usable end ≤ actual reserved start. That is a different question from allocating a buffer large enough for a request, where rounding its size upward is often appropriate.
Test the invariant, not the helper's name
A useful boundary test set for this fictional allocator includes a reservation starting exactly at 128 KiB, one byte above it and one byte below it. The largest safe whole-page ends are respectively 128 KiB, 128 KiB and 124 KiB. Testing only aligned addresses would miss the important distinction.
The broader review technique is to state which side owns the boundary, whether the end is inclusive or exclusive, and what alignment the consumer needs. A plausible-looking rounded address is not itself evidence of non-overlap.
Keep the AI anecdote in proportion
Torvalds reports 24 instrumentation patches and 18 kernel boots, credits AI assistance, and says it repeatedly suggested giving up. That is an account of an assisted debugging session, not a comparative benchmark of coding models. The durable evidence here is the reproduced failure mechanism and the committed fix. PeopleAreGeek did not reproduce the GPU fault.
Inspect actual diff: down to 4 KiB plus new non-overlap assertion; correct flat CCS meaning and demonstrate boundary invariant with original arithmetic.