Skip to content

arm64: dts: apple: describe J714s per-cluster L2 caches - #2099

Draft
DjDeveloperr wants to merge 1 commit into
integration/aurorafrom
djdev/j714s-l2-cache-intg2080
Draft

DjDeveloperr wants to merge 1 commit into
integration/aurorafrom
djdev/j714s-l2-cache-intg2080

Conversation

@DjDeveloperr

@DjDeveloperr DjDeveloperr commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

J714s has a CPU cluster map but no shared L2 cache links. Describe one unified L2 per cluster and point each CPU at its own cache: 8 MiB for CPUs 0–5 and 6–11, and 16 MiB for CPUs 12–17, as supplied by the board's ADT. Cache geometry and L1 sizes remain unspecified because the ADT does not supply them. Only t6050-j714s.dts changes.

Based on merged integration/aurora bf423bf, including #2080. This changes no GPU controller defaults or driver code.

Validation so far:

  • Exact-head standalone DTC check on Fedora: clang 22.1.8 + DTC 1.8.1, private exports of bf423bf and this head, serialized under the existing CI build lock and 6 GiB resource cap. Both compile; no new DTC warning classes. The result contains all 18 CPU cache links and exactly two 8 MiB plus one 16 MiB L2 nodes. Head DTB SHA256: 23b52d100b655421ae4c4e070b84118b4e28d1a151f65073c083d412502f6551. The shared CI worktree was preserved unchanged.
  • The new board DTS blob equals the earlier cache-description change a797051; its parent board DTS is byte-identical to bf423bf's. Diff check passes.
  • Equivalent cache properties added to the frozen intg10 DTB booted twice on J714s. Linux exposes L2 8192K shared by 0–5, 8192K by 6–11, and 16384K by 12–17. Unspecified geometry is reported as such.
  • The timing boot completed two L rounds before the host storage guard requested a yield: zero GPU faults/new kernel lines, and the default was restored and verified. Calibration remained the reference words, with policy/region-C maxima 98884 mW; no drm/asahi: g17: M5: apply the controller defaults on every part #2080 mismatch-part speedup is assumed.

These boots used equivalent additions to an existing DTB, not a DTB compiled from this exact source head. One stock boot with three rounds versus one interrupted L2 boot with two rounds is not a qualified speedup comparison; some rows improve and others regress. Draft pending kernel-build-system/schema checks and complete repeated stock/L2 boots. No performance win or default change is claimed.

The J714s describes its CPU clusters without their shared L2 links.
Describe one unified L2 per cluster with the sizes supplied by the
board description: 8 MiB for clusters 0 and 1 and 16 MiB for cluster 2.
Point each CPU at its cluster cache so cacheinfo and the scheduler can
distinguish the three sharing domains.

The board description supplies no cache geometry or L1 sizes; leave
those unspecified. The change is local to J714s.

Signed-off-by: Diljit Singh <hey@djdev.me>
@RyanTheTide

Copy link
Copy Markdown
Contributor

Blockers: 2.

  1. arch/arm64/boot/dts/apple/t6050-j714s.dts:40-144: the CPU nodes and cpu-map are in t6050-max.dtsi:29-47, which J714c, J716c, J716s, J775c, J873s and K114/K116 also include. The cluster caches belong to the SoC there, as t8103.dtsi, t600x-common.dtsi, t6030.dtsi, t604x.dtsi and t8132.dtsi all do. In the board file, every other T6050 board keeps the "all eighteen share one L2" topology this PR fixes. Move the nodes and next-level-cache links into t6050-max.dtsi (and per die in t6050-ultra.dtsi).
  2. The P-cluster model conflicts with arm64: dts: apple: t6050: give each P core a private L2 in front of the cluster L3 #2574, which describes the same caches in the shared T6050 dtsi. That PR reports CLIDR_EL1 showing three levels on the P cores: a private L2 per core, then a 16 MiB L3 shared by the cluster. This PR describes the 16 MiB as an L2 shared by cpu12-17. Both can't be right. If CLIDR says three levels, the 16 MiB node here gives cacheinfo a cache level the CPU doesn't report. Settle the topology against CLIDR_EL1 on a P core, then land one of the two. If arm64: dts: apple: t6050: give each P core a private L2 in front of the cluster L3 #2574 is right, close this PR in its favour.

HW row: J714s booting a DTB built from this exact head (not an edited intg10 DTB): /sys/devices/system/cpu/cpu{0,6,12}/cache/index*/{level,size,shared_cpu_list} matching CLIDR_EL1, plus the same on one other T6050 board (J775c or J716c) once the nodes are in the shared dtsi. Also 0 AVC, 0 failed units and no new kernel errors.

Assisted-by: LLM (Claude:Opus 5.5)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants