linux-ptl: rebase to 7.2.2.arch1, carry all patches + Xe opportunistic compaction

[spencerbull]

Summary

Rebase linux-ptl from 7.1.8.arch1-2 to Arch's current base linux 7.2.2.arch1-1 (kernel.org 7.2.2 tarball + v7.2.2-arch1 patch), and fold in PR #136 (Brost v6 mm+xe opportunistic compaction) so it ships with the new base.

Every carried patch was analyzed against the 7.2.2 tree and independent lore/mirror research before deciding keep/drop: none of the five are upstream in 7.2.2 (or master through the 7.3 merge window), so all adds are kept.

Patch-by-patch analysis

Patch	7.2.2 status	Action
0020 alpm PR_ALPM gating	Not upstream; bug still reachable — 7.2 moved PR_ALPM_CTL programming under intel_alpm_is_alpm_aux_less(), which is also true for LOBF-only (no Panel Replay) configs	Retargeted onto the new block (&& crtc_state->has_panel_replay)
0028 XPS 14/16 Panel Replay ALPM lag workaround	Not upstream; upstream doubled down the other way — 7.2 now ships Panel Replay disable quirks for both 1028:0db9 (XPS 14) and 1028:0dba (XPS 16, new in 7.2)	Reworked: repoint both upstream entries at the keep-enabled workaround quirk instead of carrying our own 0dba entry; commit message updated
0029 drm/edid DisplayID adaptive-sync monitor range	Not upstream; drm_edid surfaces byte-identical 7.1.8→7.2.2	Carried; plus a review-driven fix (below)
1001/1002 opportunistic compaction (from #136)	v6 (2026-06-17) is still the latest revision on lore — no v7; nothing merged to Linus/mm/drm-xe-next; discussion stalled after Brost's 2026-07-31 reply to Chinner	Carried as the complete b4 v6 patches with byte-identical +/- payload lines, context-refreshed for 7.2 vmscan drift (2 hunks, pure context: success→need_rotate rename, vmpressure() order arg, new balance_pgdat tracepoint)

config.x86_64 is Arch stock 7.2.2.arch1-1 (our previous config was byte-identical to Arch stock 7.1.8, so there were no local deltas to carry).

Review-driven fix in 0029

Deep review found the adaptive-sync parser selected the descriptor with the widest min→max span. DisplayID 2.0 defines descriptor 0 as the panel's primary operating range; later descriptors are alternate operation modes, so widest-span could feed an unsustainable min into intel_vrr on multi-descriptor panels (low-end flicker risk). Now selects the first valid descriptor. Single-descriptor panels (the XPS OLEDs) see no behavior change.

Review + build verification

• Retarget done as per-patch commits on a pristine 7.2.2-arch1 baseline; regenerated patch files apply with --fuzz=0, zero offsets, and reproduce the reviewed tree byte-for-byte. Original b4 mail headers/trailers preserved on 1001/1002.
• Two review rounds (independent reviewer agents):
  • Round 1 (rebase correctness): all hunks verified landed in the right functions; exhaustive cancel-site audit of the new PSR delayed work (no missed teardown path); DEVICE_ID_ANY quirk matching intact; all shrink_slab() callers/signature audit clean; all seven touched objects compiled warning-free with CONFIG_WERROR=y forced.
  • Round 2 (upstream-interaction deep dive + fix verification): zero rebase-specific regressions. drm_edid/mm/shrinker/xe/TTM surfaces are byte-identical 7.1.8→7.2.2; i915 deltas verified non-interacting (LOBF/PR structurally mutually exclusive; new Wa_16025596647 machinery excludes PR; DC3CO path upstream-dead).
• Docker build (./bin/build --package linux-ptl): checksums + GPG verified in-container, all six patches applied, 1 built / 0 failed. Produced linux-ptl-7.2.2.arch1-1 + headers; package contains vmlinuz, i915.ko.zst, xe.ko.zst; prepared version 7.2.2-arch1-1-ptl.

Known items deliberately not addressed here

• Two findings inside the v6 opportunistic-compaction design (an nr_deferred clamp that can discard xe backup-reclaim debt on opportunistic passes, and a hint-unaware runtime-PM force-wake) are upstream-series issues, present since 7.1.8 — per the linux-ptl: fix Xe reclaim stalls under fragmentation #136 review principle we carry the reviewed v6 verbatim rather than forking it downstream. They belong in the lore thread; if xe backup reclaim ever looks anemic under THP pressure, the deferred clamp is the first suspect.
• Watch for a reworked v7+ of the series (Chinner is pushing a compaction-aware shrinker architecture instead — if that lands, 1001/1002 need a rethink, not a rebase).
• Runtime validation on XPS 14/16 hardware still applies: un-disabling Panel Replay newly exercises 7.2's PANEL_REPLAY_CONFIG3/coasting-vtotal paths on these LGD panels, and the flush-exit trades some power under sustained animation for the idle savings.

Test plan

• makepkg source verification + prepare() (host and docker)
• Full docker package build, both packages produced
• Boot + A/B on XPS 14 DA14260 / XPS 16 DA16260 (Panel Replay idle power, cursor lag, VRR range, Xe reclaim under 5K load)