Skip to content

soc: apple: sleep: T6050: J716c (M5 Max) deep sleep - #2119

Merged
DjDeveloperr merged 7 commits into
integration/aurorafrom
j716c/sleep-j716c-profile
Nov 17, 2026
Merged

DjDeveloperr merged 7 commits into
integration/aurorafrom
j716c/sleep-j716c-profile

Conversation

@erikwb

@erikwb erikwb commented Nov 17, 2026 •

Copy link
Copy Markdown
Contributor

What this does

Deep sleep (suspend-to-RAM) for the 16" M5 Max MacBook Pro (J716c) in the T6050 sleep driver.

  • dt-bindings: the J716c's apple,t6050-j716c-sleep compatible (with apple,t6050-sleep as
    fallback), and up to seven SPMI buses (the Max's nub-spmi1 and nub-spmi4 after the J714s's
    five).
  • The M5 Max's power inventory (the Pro's words, then the second encoder's and AMCC16-31's),
    checked before any hardware is touched, with the board and CPU topology validated first.
  • The full profile (apple,t6050-j716c-sleep): the J714s's entry and wake over the Max's
    inventory, with the Max's own power words saved at the entry and put back at the wake, and its
    banks and display slots checked against the J716c's addresses before anything is mapped. The
    paths that served the J714s only (the storage controller's late stop and wake, the display
    coprocessors' power readiness and restart power words, the firmware 27 late scope) serve both
    full T6050 profiles. No PMP runs on this board, so the commit takes the PTD entries itself.
  • Room for the Max's two extra SPMI buses in the sleep's interrupt masks.
  • The secondary CPU clusters (MPIDR affinity 1 of 1 and 2) are taken down at the start of the
    commit, before the boot cluster: their cores are released to their last WFI first
    (spin_table_release_cpus()), counted, their requests dropped and their sequencers polled
    done. The clusters' power-down programs run one at a time; dropped together, the P cluster's
    queued behind the boot cluster's, which can't finish while the boot core runs the commit. This
    is past the point of no return and refuses nothing: its waits are bounded on the CPU's counter
    (2 ms for the cores, 5 ms for each sequencer), and if not all cores are counted the cluster
    requests are left for the commit to drop as before. Any failure ends in the engine's watchdog
    reset 5 s after the commit, as without this step. The result is logged in one line after the
    wake.
  • VENC0_SYS, on with no Linux user at the commit, is lowered at the entry; if it does not go off
    the sleep is refused there, and a refusal puts it back.

The J714s's profile is unchanged.

Tested

  • soc: apple: sleep: T6050: J716c (M5 Max) deep sleep #2119 (j716c/sleep-j716c-profile) builds on its own base (Image.gz, modules and dtbs, arm64, LLVM).
  • The kunit tests in drivers/soc/apple/sleep-test.c cover the J716c inventory and the full profile (built; not run separately on hardware).
  • All 17 branches merged on integration/aurora were built as one kernel and run on a J716c (16" M5 Max): one suspend-to-RAM with a USB4 dock attached, about 30 s asleep, woken by the RTC alarm (wake source 76). After the wake all 18 CPUs were online, the dock's router came back in place with its Ethernet up, and the SEP answered RNG reads. The sleep's log showed "secondary clusters down first: cores 6+6, cluster 1 0x1a in 353 us, cluster 2 0x1c01a in 448 us". The PMP was not running in that build.

Dependencies

Branch #2119 (j716c/sleep-j716c-profile). Stacked on #2118 (j716c/spin-table-release-cpus).

Part of the J716c deep sleep set: #2118 (j716c/spin-table-release-cpus), #2119 (j716c/sleep-j716c-profile), #2120 (j716c/sleep-wake-source-names), #2121 (j716c/sleep-sep-auto-wake), #2122 (j716c/sep-t6050-profile), #2123 (j716c/dts-deep-sleep), #2124 (j716c/drm-apple-sleep), #2125 (j716c/dart-sleep), #2126 (j716c/thunderbolt-sleep), #2127 (j716c/usb-sleep), #2128 (j716c/phy-sleep), #2129 (j716c/pcie-sleep), #2130 (j716c/aic-sleep), #2131 (j716c/asahi-sleep), #2132 (j716c/spmi-sleep), #2133 (j716c/macsmc-sleep), #2134 (j716c/nvme-sleep).

🤖 Generated with Claude Code

… buses

The M5 Max J716c uses the T6050 sleep controller too. Add its
compatible, apple,t6050-j716c-sleep, with apple,t6050-sleep as the
fallback.

The M5 Max has two SPMI buses beyond the J714s's five (nub-spmi1 and
nub-spmi4): allow up to seven in apple,spmi-buses and
apple,spmi-sleep-order, after the J714s's five, so that the first five
keep their register-sequence block indices (8 to 12) and the two new
ones take 13 and 14.

Developed with Claude Code.

Signed-off-by: Erik Bourget <erik@ebourget.net>
Keep the Max power-domain inventory distinct and refuse incomplete
profiles before hardware preparation. Validate board and CPU topology
before applying boot register sequences. The J716c entry callbacks
remain unavailable pending qualification of its memory-controller
transitions.

Developed with Claude Code.
Signed-off-by: Erik Bourget <erik@ebourget.net>
The J716c's apple,t6050-j716c-sleep profile was inventory only and
refused at the probe. Give it the J714s's entry and wake on the M5 Max:
the same operations, gate list, SMC policy, CPU layout and display
words, over the Max's power inventory (the Pro's words at the same
indices, then the second encoder's and AMCC16-31's), with the keepalive
tables of the Pro's geometry, which the Max's PMGR has too.

Its map first checks the PMGR banks and the display slots against the
J716c's addresses, and then maps as on the J714s. The Max-only power
words are in none of the lists the entry lowers: saved at the entry,
and put back at the wake if they dropped.

The paths that served the J714s's profile only (the storage
controller's late stop and wake, the display coprocessors' power
readiness and their restart's power words, the firmware 27 late
scope) now serve both full T6050 profiles.

No PMP runs on this board and its report is disabled, so nothing else
owns the PTD entries the commit drops: such a profile takes them
without the report's ownership. The gate still refuses a PMP that runs
unbound, and a board whose report is enabled still has the report give
them up.

The J714s's profile is unchanged, and an inventory profile with an
unavailable reason still refuses.

Developed with Claude Code.

Signed-off-by: Erik Bourget <erik@ebourget.net>
The sleep sets every SPMI bus's interrupt mask to the sleep's response
enables after ANS, and puts the run-time masks back at the wake. It had
room for the J714s's five buses. The M5 Max has two more, nub-spmi1 and
nub-spmi4, behind its other PMICs: take up to seven, after the J714s's
five, so that their indices as register-sequence blocks (8 to 12) do not
move. A DT with five buses, as the J714s's, works as before.

Developed with Claude Code.
Signed-off-by: Erik Bourget <erik@ebourget.net>
… before the boot cluster

The T6050's CPU clusters' power-down programs run one at a time on one
engine. On the J716c, with all three cluster requests dropped in the
commit window, the P cluster's program queued behind the boot
cluster's, which cannot finish while the boot core is still running
the commit, and the sleep did not complete.

For the J716c profile (clusters_down_first in its SoC data), take
clusters 1 and 2 down at the very start of the commit, which is past
every refusal: release their cores to their last WFI ahead of the boot
cluster's (spin_table_release_cpus()), wait for all of them to be
counted in t6050_cpu_down(), drop each cluster's PTD bit and request,
and poll each cluster's sequencer until its program is done. The boot
cluster then goes down last, as before, and all three programs complete.

This is past the point of no return, so it refuses nothing. Its waits
count the CPU's own counter, as timekeeping is stopped by then, and are
bounded well inside the commit's runway: 2 ms for the cores, 5 ms for
each sequencer. If not all cores are counted, the cluster requests are
left up and the commit drops them later, as before. Any failure ends in
the engine's watchdog reset 5 s after the commit, as it would without
this step, so there is nothing to undo.

Once the requests are dropped, nothing in the commit touches those
clusters' blocks: their later request release and their CPM counter
reads in the power-gate rounds are skipped. The results are kept in
the region state and reported in one line after the wake.

Developed with Claude Code.

Signed-off-by: Erik Bourget <erik@ebourget.net>
On the J716c the first video encoder's system power word reads
0x0f0003ff (on, with no automatic state) at the commit, although no
Linux driver uses it and every other media word of the GPU die is off
or under automatic power. The entry's last phase waits for its power
gate.

For the J716c profile (lower_venc0 in its SoC data), lower VENC0_SYS at
the entry, after AUSS, and poll it off. If it does not go off, the
sleep is refused there, before anything irreversible. A refusal puts
it back as it read before; after a sleep the snapshot's replay raises
it again.

Developed with Claude Code.

Signed-off-by: Erik Bourget <erik@ebourget.net>
@erikwb
erikwb force-pushed the j716c/spin-table-release-cpus branch from 3abd1e1 to d127243 Compare November 17, 2026 15:08
@erikwb
erikwb force-pushed the j716c/sleep-j716c-profile branch from cc10822 to 949b24a Compare November 17, 2026 15:08
…f CPU

The early cluster sequence counts present cluster-1 and cluster-2 CPUs
as they execute the final CPU-off callback. A CPU tracked as off after
a failed restart passes the spin-table held check but cannot reach that
callback again. Reject the inconsistent cohort while the final sleep
checks can still return. Preserve the treatment of offline CPUs that
really are parked, and do not change the spin-table masks.

Signed-off-by: Diljit Singh <hey@djdev.me>
@DjDeveloperr
DjDeveloperr changed the base branch from j716c/spin-table-release-cpus to integration/aurora November 17, 2026 19:06
@DjDeveloperr
DjDeveloperr marked this pull request as ready for review November 17, 2026 19:14
@DjDeveloperr
DjDeveloperr merged commit c28f863 into integration/aurora Nov 17, 2026
6 checks passed
@RyanTheTide
RyanTheTide deleted the j716c/sleep-j716c-profile branch November 17, 2026 23:53
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