Kernel fix stops the Intel Xe driver from handing compression storage out as usable VRAM

Sources: mainline commit 818bebeb63dd
Summary: Linus Torvalds authored and committed the fix on 2026-08-20. It stops the Intel Xe driver from publishing the memory range reserved for compression hardware to the VRAM allocator, where that range could be allocated to normal callers. The commit message reports the symptom on a Battlemage G21 with 16 GiB, where a Mesa virtual machine's level-3 page table landed on the affected page on every cold boot, losing the compositor's batch-buffer heap entry and leaving a black screen with gdm restarting in a loop. It also records the debug session behind a one-line round_up to round_down change: 24 debug patches and 18 kernel boots, with a note that an AI did much of the grunt work while repeatedly declaring the bug unsolvable.
Why it matters: Memory owned by the compression hardware was published to the VRAM allocator and silently overwritten whatever landed there, with no page-table entry or GPU submission needed to corrupt it, and the commit carries Cc: stable@kernel.org, so a backport to stable series is expected.