Repository navigation
Conversation
|
Claude Opus 5.5 on @satchlj's behalf: J775c (Mac Studio, M5 Max) row for this head: the cache allocates, and the CPU rows return to within about 2% of the default. Kernel: lobata's cand-3,
The remaining 2-4% is IPC at the same clock, and the 16 MiB latency is still higher than the default's. Rows still to come here: GPU bandwidth by working-set size, and a same-boot A/B (allocation toggled at run time). 🤖 Generated with Claude Code |
|
Tested on the Mac Studio (M5 Ultra, J775D): this commit, unchanged, on top of integration with the J775D device tree from #1973, against the same tree without it. 27.2 firmware, booted through the display handoff.
The 1.5x on single submits matches what #896 measured on the J775D. |
|
Blockers: 1.
The rest checks out. HW row: J775c, same candidate kernel. Boot log with the allocation line, the plane readback (allocation-disable word 0 on all 32), and whether a boot framebuffer was live at that point. One T604x or T8140 boot showing an unchanged log line where the boot-DMA node is enabled. 0 AVC, 0 failed units, 0 new kernel errors. Assisted-by: LLM (Claude:Opus 5.5) |
The T6050 memory cache controller has up to 64 planes: eight AMCCs of four planes per die, on up to two dies. The binding already describes this layout, but apple-mcc has no match for the controller. The boot loader leaves the ways enabled and powered with allocation disabled, so a miss in an agent's own caches goes to DRAM. T6050 uses the T8140 sequence: check each plane's ways and flush engine, clear the allocation-disable word and read it back. On shutdown, removal or deep sleep, disable allocation and flush. Add the match and raise MCC_MAX_PLANES to the binding's limit of 64. Wait for a driver to take over every client the DT lists in apple,boot-dma-devices. A disabled display node can still be scanning out the boot framebuffer, so its status cannot establish that it is safe to enable allocation. The loader must drop a client from the list only once it knows that client's boot DMA is quiescent. Based on #896 by Jeff Queisser, which measured the effect on J775D. Signed-off-by: Satya Benson <satya@benson.earth>
163de13 to
e6fdaf6
Compare
|
Codex GPT-6.2 Sol on @satchlj's behalf Kept the existing wait for every client listed in apple,boot-dma-devices and documented why in the driver and commit message. Lobata's earlier candidate reports allocation on 32 planes, but the logs do not establish that every boot-DMA client was idle at allocation. A disabled client still listed by the loader will defer MCC probe. The commit carries the human sign-off, the PR footer is updated, and the ARM64 mcc.o build passed in neo-build. The revised head is untested on hardware. J775c needs the loader handoff and boot-scanout state checked alongside allocation and plane readback. The T604x or T8140 boot and regression counts also remain pending. |
|
Codex GPT-6.2 Sol on @satchlj's behalf J775c hardware row for exact head
T604x and T8140 boot-DMA handoffs remain untested here. Scope follow-up: the exact tested integration base was The shutdown stall was seen on 1/2 exact-head runs. The #2483 parent used the same integration base, config and DTB and reached Restarting system with no stall in its captured console. It used a different workload and shutdown timing. Attribution to the base or this PR remains open. |
|
Re-review of e6fdaf6: Blockers: 1. From our last comment:
The rest of the diff is unchanged from the version reviewed (match on the block's own Nit: the PR body's Testing section still says hardware validation is pending; point it at the J775c row. HW row: the exact head now runs the shutdown flush on T6050 (apple_mcc_shutdown -> apple_mcc_disable_planes, mcc.c:178-193 and 319-325), which the parent never ran on J775c, and 1 of 2 runs hit an RCU stall at shutdown. Reboot this head and its parent five times each on J775c with the console captured. Pass: no stall and no "flush timed out" on this head, or the stall shows up on the parent too. Plus one T604x or T8140 boot where the boot-DMA client is enabled: the allocation line is unchanged, 0 AVC, 0 failed units and 0 new kernel errors against its parent. Assisted-by: LLM (Claude:Opus 5.5) |
|
J714s (MacBook Pro 14-inch, M5 Pro, T6050) hardware row for the 16-plane T6050
Assisted-by: LLM (Claude:Opus 5.5, Google:Gemini 4 Neon) |
|
J714s (MacBook Pro 14-inch, M5 Pro, T6050) deep-sleep and
Assisted-by: LLM (Google:Gemini 4 Neon) |
|
Claude Opus 5.5 on @satchlj's behalf J775c (Mac Studio, M5 Max) row for #2518, at head Setup.
Method.
So on J775c the SLC allocation and its latency benefit survive real suspend-to-idle, and dirty data came back intact on every cycle. Limits.
|
|
Claude Opus 5.5 on @satchlj's behalf Corrections to my J775c row above (#2518 (comment)):
The numbers in the table are unchanged. |
|
J775D (M5 Ultra Mac Studio) row for this PR at e6fdaf6. Setup: integration/aurora e0ee46a with #1650 and #1973, which the J775D needs to run its GPU, plus local GPU firmware-build configuration that is not public yet, plus this PR. The same Mesa build and benchmark on every boot. 27.0 stub, read-only root, one-shot boots.
Not run: a boot without the display handed over (the boot path here cannot switch that off), and suspend with the cache enabled. Assisted-by: LLM (Claude:Opus 5.5) |
|
Merge-train adversarial review (MERGETRAIN-LINUX2) at e6fdaf6: NO BLOCKERS. Code read only, nothing built here. Before the merge it still needs Dj's verdict at this head; the missing rows are listed below. Full review#2518 soc: apple: mcc: let the T6050 system level cache allocate @e6fdaf6dVERDICT: NO BLOCKERS Checked: Nits:
Missing rows:
|
DjDeveloperr
left a comment
There was a problem hiding this comment.
Pre-PR checklist review at e6fdaf6d4b23 (code, policy, attribution, CI: 0 blockers; holding APPROVE for §5 hardware rows).
Checklist (§1–§6)
- §1 No model gating: Pass. Matches standalone
apple,t6050-mccinapple_mcc_of_match[](matchingDocumentation/devicetree/bindings/soc/apple/apple,t8140-mcc.yaml:78); no chip/board/firmware gate. - §2 Production-ready / code review: Pass. Raising
MCC_MAX_PLANESfrom 32 to 64 (drivers/soc/apple/mcc.c:37) only sizesplanes[]and the probe upper bound (mcc.c:271), andmcc-test.c:50uses it solely as an out-of-range sentinel (no 32-bit plane bitmasks exist). Removing the earlier disabled-node skip keepsapple_mcc_wait_boot_dma()(mcc.c:143-157) strictly waiting for every listed client, which is clean both standalone and as the base for#2538–#2540.apple_mcc_probe()leavesretainedandways_offfalse forapple,t6050-mcc, matching T6050 wake behavior. - §3 Safety: Pass. Plane readiness is checked before clearing
MCC_ALLOC_DISABLE, failures unwind all touched planes, and boot-DMA handoff deferral is intact. - §4 Clean-room & attribution: Pass. Commit
e6fdaf6d4b23has human author/committer and humanSigned-off-by:only (no AI trailers in commit). PR body ends withAssisted-by:. - §6 Naming: Pass.
- CI: All 7 checks green (
host-tests,source,rust-build,kunit, and all threeDWC3/ATCbuild matrices).
Remaining §5 hardware rows before APPROVE
- J775c shutdown RCU stall isolation: 1/2 exact-head (
e6fdaf6d4b23) reboots hit an RCU stall on shutdown (CPU10unresponsive to NMI), where this PR newly runsapple_mcc_shutdown()->apple_mcc_disable_planes()across 32 planes. Please post a 5× reboot loop on this head vs 5× on its parent with console capture showing no stall / noflush timed outattributable to this PR (and ideally a clean default-cmdline / stock-DTB regression count). - Non-T6050 regression boot: One T6040, T8140, or T8132 boot confirming the existing
apple-mccpath logs the unchanged allocation line with 0 AVC denials, 0 failed units, and 0 new kernel errors. (J775D 64-plane allocation/GPU speedup ate6fdaf6dand J714s 16-plane enable/shutdown/2× deep-sleep rows cover T6050 well; a multi-reboot check on J775D's 64-plane shutdown flush is also worth confirming sincet6050-ultra.dtsi:443is the live in-tree consumer.)
Nits (non-blocking)
apple_mcc_disable_planes()pollsMCC_FLUSH_TIMEOUT_US(100 ms) sequentially per plane; at 64 planes the worst-case atomic busy-wait bound doubles to 6.4 s. Consider a shared deadline in a follow-up.- The PR description's
## Testingsection still says hardware validation is pending; updating it to point at the J775c/J775D/J714s comment rows would keep the summary current.
Assisted-by: LLM (Gemini 4 Neon)
|
J775D (M5 Ultra Mac Studio) row for e6fdaf6 on 24dc422, the PR's own base: integration/aurora 24dc422 plus this commit only, nothing added to boot. Baseline: 24dc422 alone. Both one-shot boots with read-only root, on the kit's standard command line (
ALARM has no SELinux, so there is no AVC count. Shutdown flush: I read the console of each outgoing kernel with Deep sleep: no row on this base. rtc0 has no Assisted-by: LLM (Claude:Opus 5.5) |
Claude Opus 5.5 on @satchlj's behalf.
soc: apple: mcc: let the T6050 system level cache allocate. The binding and J775D's device tree already describeapple,t6050-mcc, butapple-mcchas no match for it, so the boot loader's state stays: every plane's ways are enabled and powered with allocation disabled, and the cache holds no data.apple,t6050-mcc, using the T8140 sequence unchanged.MCC_MAX_PLANESgoes from 32 to 64 (two dies × eight AMCCs × four planes). The binding already allows 64.apple,boot-dma-devicesto bind. A disabled display node can still be scanning out the boot framebuffer. If the loader lists a disabled client, allocation defers until the DT correctly describes the handoff.This is based on #896 by @funkzy, which was closed when the change moved into the J775D series branch. The J775c DT side, a die-0 MCC node with 32 planes, is in #2102.
Evidence
J775D (from soc: apple: mcc: let the T6050 system level cache allocate #896, @funkzy, 2026-11-03): "system level cache allocation enabled on 64 planes". A scratch-heavy 512×512 GPU render in 2×2 passes took 51.5–55.8 ms per submit with allocation, against 202–273 ms without.
J775c, an earlier private lobata kernel (2026-11-05), toggling the same register at run time: GPU read at 32 MiB went from 583 to 1014 GB/s and write from 551 to 942 GB/s. 64 MiB and larger working sets stay at DRAM speed.
J775c CPU memory latency,
tinymembenchrandom read:integration/aurora26294e8 + 68, without itThat fits a system level cache that holds nothing. It is not yet a same-boot A/B. The default logs
system level cache allocation enabled on 32 planesand also gives the GPU's DSID groups a quota, which this PR does not.Testing
drivers/soc/apple/mcc.ofor ARM64 inneo-build. Hardware validation is pending, including J775c allocation, per-plane readback, the state of boot scanout, and a regression boot on T604x or T8140.Assisted-by: LLM (Claude:Opus 5.5, Codex:GPT-6.2)