AMD VRAM overcommit work is queued for Linux 7.3, with the eviction deadlock fix still outstanding

Sources: pixelcluster.dev write-up, HN discussion
Summary: The author's earlier VRAM management patches are merged upstream and queued for Linux 7.3. The remaining failure is that TTM does not use the drm_exec wound-wait helper, so an -EDEADLCK raised while evicting during command submission is returned as -ENOMEM and fails the submission instead of retrying the transaction. The author rebased and resent the 2024 patchset for that with fixes, and states it still needs more work before it can be merged, so this part is not in 7.3. The post separately measures that display scanout buffers must be physically contiguous in VRAM, which drove up to 4 GiB of eviction to place roughly 32 MiB of scanout data, and describes hard and soft throttle heuristics added to stop eviction ping-pong between a game and the compositor.
Why it matters: Exceeding physical VRAM should cost performance and not stability, and the post shows why it currently costs both.
Follow-up: Track whether the drm_exec conversion for TTM lands after 7.3.