Skip to content

T6040 (M4 Pro): sixth P-core slots (pcpu0_5 / pcpu1_5) are present — bring-up not demonstrated #1975

Open

Description

@fbalazsy

The common T604x CPU map describes two P clusters of six cores, but T6040 (M4 Pro) only publishes 5+5 P-cores plus 4 E-cores (14 CPUs). The two unpublished cores would be cpu@10105 and cpu@10205. This issue records what was observed on a J616s (16-inch M4 Pro) and proposes a follow-up. Those two cores have not been brought online. No SoC fuse writes should be attempted, and this is not a request to turn the cores on by default.

Observation

In integration/aurora:

  • t604x.dtsi already defines cpu_p05: cpu@10105 and cpu_p15: cpu@10205 (compatible = "apple,brava-p", reg = <0x0 0x10105> / <0x0 0x10205>, spin-table) and the matching core5 entries under both P clusters.
  • t6040.dtsi deletes both nodes and both core5 map entries, with the comment "The M4 Pro has five cores in each P cluster."
  • t6040-pmgr.dtsi still contains pcpu0_5 (power-controller@48) and pcpu1_5 (power-controller@78), status = "disabled", like the other CPU power-controller nodes (Linux does not manage CPU power through genpd).

That is the same 6+6 template used for T6041. A 16-core M4 Max keeps those two CPU nodes. A 14-core M4 Max does not have them at runtime: #914 notes that the 14-core part's MPIDRs match the J616s set, and that the bootloader nops the two missing nodes by MPIDR.

Public teardown reporting (Tecanalie, EE Times Japan, 17 March 2025) states that M4 Pro silicon has 12 P-cores even though the advertised high SKU is 10P+4E, and that Pro and Max share most of the layout as a chopped pair, not as one physical die sold under two names. TechInsights lists distinct die IDs and areas (TM5J93 ~325 mm² vs TM5J94 ~454 mm²). The useful residual is therefore intra-die binning on the Pro (and the same hole on a 14-core Max), not "the Pro is a fused-down Max die."

Evidence (J616s, observed)

Live Linux: 14 CPUs online (10P+4E). Two P clusters of five (cpu@10100–cpu@10104, cpu@10200–cpu@10204) and one E cluster of four. No cpu@10105 / cpu@10205 nodes. Enable method is spin-table, not PSCI. The boot CPU is cpu@10100. MIDR of the published Brava P-cores on this part is 0x611f0551.

Firmware device tree (ADT):

  • CPU nodes cpu0–cpu8, then a hole (no cpu9), then cpu10–cpu14. No cpu15. No third P cluster. cpu-cluster-count = 3.
  • Core-enable bits (function-enable_core): E 0x1, 0x2, 0x4, 0x8; P0 0x10 … 0x100; 0x200 unused; P1 0x400 … 0x4000; 0x8000 unused.
  • AIC #main-cpus = 12 (not the 14 CPUs that are online).
  • No CPU-node property that looks like a core-enable fuse map. fuse-revision = 2 and the usual security-domain fuse cells are not a CPU mask.
  • cpu-impl-reg is published for the 14 cores (0x211050000 … 0x211450000 for P0 cores 0–4, 0x212050000 … 0x212450000 for P1). The patterned 6th-core addresses 0x211550000 / 0x212550000 are not published and have not been shown to be readable.

Live PMGR (pmgr_aux at 0x502800000, compatible apple,t6041-pmgr, already mapped by the pwrstate syscon). Each ecpu* / pcpu* / *cpm node has a 4-byte PS word at an 8-byte stride. Bit layout as in drivers/pmdomain/apple/pmgr-pwrstate.c: PS_TARGET [3:0], PS_ACTUAL [7:4], WAS_PWRGATED (8), DEV_DISABLE (10), BUSY (11), AUTO_ENABLE (28). Levels: 0x0 PWRGATE, 0xf ACTIVE.

Read once on 2026-11-14 (Europe/Bratislava). All CPU-class words were well-formed (not 0xffffffff, no bus error). The +4 word after each of them read as 0. No DEV_DISABLE, BUSY, or reset bit on any CPU word:

offset label raw target actual other bits
0x00 ecpu_0 0x00000100 0 PWRGATE 0 PWRGATE WAS_PWRGATED
0x08 ecpu_1 0x000001f0 0 PWRGATE 0xf ACTIVE WAS_PWRGATED
0x10 ecpu_2 0x00000100 0 PWRGATE 0 PWRGATE WAS_PWRGATED
0x18 ecpu_3 0x00000100 0 PWRGATE 0 PWRGATE WAS_PWRGATED
0x20 pcpu0_0 0x00000100 0 PWRGATE 0 PWRGATE WAS_PWRGATED
0x28 pcpu0_1 0x00000100 0 PWRGATE 0 PWRGATE WAS_PWRGATED
0x30 pcpu0_2 0x00000100 0 PWRGATE 0 PWRGATE WAS_PWRGATED
0x38 pcpu0_3 0x000001f0 0 PWRGATE 0xf ACTIVE WAS_PWRGATED
0x40 pcpu0_4 0x00000100 0 PWRGATE 0 PWRGATE WAS_PWRGATED
0x48 pcpu0_5 0x00000100 0 PWRGATE 0 PWRGATE WAS_PWRGATED
0x50 pcpu1_0 0x00000100 0 PWRGATE 0 PWRGATE WAS_PWRGATED
0x58 pcpu1_1 0x00000100 0 PWRGATE 0 PWRGATE WAS_PWRGATED
0x60 pcpu1_2 0x00000100 0 PWRGATE 0 PWRGATE WAS_PWRGATED
0x68 pcpu1_3 0x00000100 0 PWRGATE 0 PWRGATE WAS_PWRGATED
0x70 pcpu1_4 0x00000100 0 PWRGATE 0 PWRGATE WAS_PWRGATED
0x78 pcpu1_5 0x00000100 0 PWRGATE 0 PWRGATE WAS_PWRGATED
0x80 ecpm 0x000001f0 0 PWRGATE 0xf ACTIVE WAS_PWRGATED
0x88 pcpm0 0x000001f0 0 PWRGATE 0xf ACTIVE WAS_PWRGATED
0x90 pcpm1 0x00000100 0 PWRGATE 0 PWRGATE WAS_PWRGATED

pcpu0_5 / pcpu1_5 look identical in that snapshot to pcpu0_4 / pcpu1_4 and to most of the published cores. That is not, by itself, a fuse-off signature.

PS_TARGET is not what keeps the Linux CPUs on: all 14 were online with TARGET=0 on every CPU PS word. PS_ACTUAL is also not a reliable "this core is running" flag — only some published cores showed ACTUAL=ACTIVE, and pcpm1 has been seen both ACTIVE and PWRGATE on different boots while cluster 1 still had five cores online.

Verified vs not

  • Verified: the sixth-core PMGR slots exist in the already-mapped window and return the same kind of PS word as the fifth-core slots. The firmware CPU list and enable-bit mask skip those two cores. The in-tree DT already has the 6+6 template and deletes the two nodes on T6040.
  • Not verified: that either unpublished core can fetch, execute, or stay online. Tecanalie's "12 P-cores on Pro silicon" has not been re-checked on this particular package. The unpublished cpu-impl-reg windows have not been shown to be readable.
  • Tried, did not bring a core up: on 2026-11-15, with an operator present and an explicit go, a bootloader-proxy test wrote PS_TARGET=0xf (no AUTO_ENABLE) to pcpu0_5 only. The address was the one the ADT gives for PCPU0_5 (+0x48). The word stayed 0x00000100 (the write did not stick; no DEV_DISABLE or BUSY). Neighbour pcpu0_4 was 0x00000100 before the write. The SoC stayed alive; the test aborted without a start-bit write and without running a payload. Ordinary non-CPU domains in the same syscon do accept and hold TARGET (for example venc0_sys at 0x1f0003ff). CPU-class words stay at TARGET=0 even when the corresponding cores are online. The bootloader brings published cores up with RVBAR plus the start-bit registers at pmgr[0] + 0x88000, not with per-core PS_TARGET. Do not retry PS_TARGET as the bring-up method.
  • Not run: start-bit bring-up of cpu@10105. That is the remaining candidate path. It needs an operator at the machine and a new explicit go. It can hang the cluster or the SoC and may need a hard power-cycle.
  • Not done, and should not be done: any SoC efuse read or write. There was no bound apple-efuse nvmem device in the sysfs used for this work.
  • Not a safe PR: removing the /delete-node/ lines in t6040.dtsi would be default-on at boot, and bring-up has not been demonstrated. This is an issue rather than a draft PR for that reason.

Proposed next steps

  1. Read-only, from the bootloader proxy: read the unpublished cpu-impl-reg at 0x211550000 (and later 0x212550000). Report whether the window is readable and whether RVBAR_LOCK is set. Stop if the read hangs or the lock is set.
  2. Only with an explicit go and someone at the machine: start-bit bring-up of one core (pcpu0_5 / MPIDR 0x10105) using the same secondary-start path the bootloader already uses for published T6040 cores, a short-lived marker payload, a timeout, and a stop-bit plus power-cycle rollback plan. Success looks like a 15th CPU, a Brava-P MIDR (published cores on this part are 0x611f0551), and MPIDR affinity 0x10105.
  3. If that works: repeat for cpu@10205, then a default-off / opt-in DT or bootloader path. Do not turn this on for all T6040 boards by default. cpufreq, topology, and park-on-sleep would need to follow.
  4. A 16-core T6041 is a useful comparison for the live cpu@10105 / cpu@10205 path. 14-core Max units have the same hole at runtime.

Risks

  • A start-bit or CPU_ON of an MPIDR the firmware never advertised can hang the cluster or the SoC. If the machine stops responding there is no software rollback; a forced power-cycle is the way back. Do this from the bootloader with a timeout, not from a live desktop session.
  • Do not write SoC fuses. Do not probe unmapped MMIO (for example unused AMCC bases); a status read of an unmapped address can also hang the bus.
  • pcpu1_5 is the poorer first target if its cluster manager is gated.

Activity

  1. fbalazsy commented on Nov 15, 2026

    @fbalazsy
    ContributorAuthor

    A few more firmware-device-tree observations from a J616s (16-inch M4 Pro), still read-only. They do not show that the two unpublished cores can run.

    The /cpus node on this part has:

    • max_cpus = 16
    • cpu-cluster-count = 3
    • e-core-count = 4
    • p-core-count = 10

    The container is sized for a 16-CPU (12P+4E) layout, while this SKU only publishes 14 CPU nodes and states 10 P-cores. That matches the 6+6 P-core template already in the common tree, with two cores omitted.

    There is no cpu-fuse-revision and no per-CPU fuse-map property. fuse-revision = 2 is on /arm-io and is not a core-enable mask. So the firmware tree itself does not spell out an efuse that turns those two cores off; it also does not prove the opposite. The unused enable bits remain 0x200 and 0x8000 (cpu-id 9 and 15).

    The PMGR node is compatible pmgr1,t6041 even on T6040, and it publishes misc-cores-offset = 557056 (0x88000). That is the start-bit window mentioned earlier, now as an official property rather than only a loader constant. acc-impl-reg is per cluster and shared by every published core of that cluster: E 0x210f00000, P0 0x211f00000, P1 0x212f00000.

    A caution for anyone who tries a start-bit bring-up later: a status bit of the form 1 << (4 * cluster + core) aliases on these 5-core P clusters. P0 core 4 and P1 core 0 share a bit; P0 core 5 would share a bit with the already-online P1 core 1. The firmware enable bits do not alias — they follow cpu-id. The per-cluster start strobe is at misc-cores-offset + 0x8 + 4 * cluster. Do not use PS_TARGET as the bring-up method, and do not write SoC fuses.

  2. fbalazsy commented on Nov 15, 2026

    @fbalazsy
    ContributorAuthor

    @AxxzyWasTaken, a short question for your J616C, only if it's easy to read (read-only, nothing to write):

    1. The two 32-bit words at 0x508374010 and 0x508338280 (PMGR). On T604x they carry per-core disable bits: E [15:12], P0 [21:16], P1 [27:22] in the first; E [3:0], P0 [13:8], P1 [21:16] in the second. On a 14-core part I'd expect one P bit per cluster. Which positions are set on yours?
    2. The misc-cores word at 0x502888004 (misc-cores-offset + 4).
    3. The per-core slots at 0x50282c090 and 0x50282c0f0 (cpu-id 9 and 15 by the 0x10 spacing from aura#73). Do they read like the other slots?

    If you've ever seen the same words on a 16-core M4 Max, that would be the most useful comparison. Thanks!

  3. RyanTheTide commented on Nov 19, 2026

    @RyanTheTide
    Contributor

    Step 1 (read-only) result, J616s, 2026-11-19. Short version: the 6th P0-core implementation window does not respond. The read hangs, so this core is not usable from the handoff state. The /delete-node/ in t6040.dtsi stays.

    How: aura's boot proxy (Stage 2, over the KIS debug link). The exception guard was on, every read was logged before it was issued, and no writes were made: no start bits, no PS_TARGET, no fuse access.

    • All 14 published cpu-impl-reg windows (E 0x210050000-0x210350000, P0 0x211050000-0x211450000, P1 0x212050000-0x212450000) read back immediately with no exception. Each has RVBAR_LOCK (bit 0) set and the RVBAR pointing at the loader base. Raw words: E 0x0010010005de8001, P0 0x0110010005de8001, P1 0x0210010005de8001. The bits above the address field follow the cluster.
    • 0x211550000 (P0 core 5): the read64 never completed. There was no reply in 240 s and no SError/abort was reported. The access stalled. I stopped there as planned, so 0x212550000 and the PMGR words from the comment above were not read. The machine came back with a USB-PD VDM reboot (~50 s to Linux) and is fine.
    • The firmware's pmgr node does list 0x211550000 and 0x212550000 (size 0x10000) among its register ranges on this part. The window is declared, but nothing behind it answers. /cpus on this boot: max_cpus 16, cpu-cluster-count 3, e-core-count 4, p-core-count 10, no cpu9/cpu15.

    How to read it: the loader reads each published core's RVBAR before it starts that core, so published windows answer before their core runs. The 6th core's window does not answer at all. That fits a harvested/fused-off core, or a power domain that never comes up without a start-bit sequence. It is not yet proof of a fuse. Either way, nothing changes for Linux: 14 CPUs is the advertised configuration (10P+4E), and there is no DT change to make.

    Still open, and only with someone at the machine: step 2 (one start-bit trial on cpu@10105 with a hard power-cycle plan). A useful positive control would be the same read of 0x211550000 on a 16-core M4 Max, where that core is published. If anyone has one, a read-only value would help.

    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

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions