Repository navigation
Update Aurora Edge to integration/aurora e0ee46a1 for the next Neo Edge kernel - #2527
Conversation
Adopt the working external display DT and use the firmware 27.2 link release before RTKit idle. Give each external processor its own power layout and include both in the sleep controller. Both warm restarts pass a dry run; real sleep still resets, with all external power states now matching macOS. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de>
The J813 external display processors complete the warm RTKit handshake after deep sleep, but lose CPU_CONTROL.RUN within 20 ms. The next suspend queues an AP-link release that the stopped processor cannot answer. Set PS_MIN to 4 before raising the external CPU domain and enabling its automatic power management. This matches the measured J813 macOS wake sequence and the existing T8140 external DCP path. On J813, both processors now retain RUN after wake and answer the next suspend's link-release request. The second full sleep still needs further debugging; this fixes the display processor retention failure. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de>
T8142 has no GAPF slice for the DCP DART. Its SPTM (27.2.1) keeps a bit per DART in a shared PS write-disable register instead (/chosen/dart-ps-wr-dis-entries): set, the fabric rejects writes to the power word of the coprocessor CPU behind the DART. The driver clears that bit where it clears GAPF bit 4 on T8140, for a DCP firmware stop, and sets it again after. Only this DART's bit of the register is changed. A GAPF slice is handled as before. Nothing of the shared register's boot value is written back, only this DART's bit. From m0d16l14n1's J815 sleep work (m0d16l14n1#1, 192a56c), taken out on its own for the DPMS firmware stop. Assisted-by: Claude Opus 5.5 Assisted-by: Claude Fable 5.2 Signed-off-by: Kyle Arrington <kyle@arrington.co>
T8142's display power node has the power gate and power state words only: its PMC keeps no count of the powered coprocessors, so there is no pmc-dcp word. From m0d16l14n1's J815 sleep work (m0d16l14n1#1, 192a56c). Assisted-by: Claude Opus 5.5 Assisted-by: Claude Fable 5.2 Signed-off-by: Kyle Arrington <kyle@arrington.co>
The power gate and power state words a DCP firmware stop writes, and the DCP DART's PS write-disable bit that protects DISP_CPU's power word (ADT dart-dcp clock-protection-wr-dis-id 0x1f: pmgr reg[1] + 0x3fdc, bit 6). Assisted-by: Claude Fable 5.2 Signed-off-by: Kyle Arrington <kyle@arrington.co>
On a macOS 27.2 (26B123) stub, the J813 refuses its DCP's
apple,firmware-compat of 27.2.0. It then drives the DCP with the H17P
interface, and the 27.2 DCP firmware crashes on the first call
("ap_link: rpc_callee_gated: link_rpc_lookup callback not found:
'A427'").
The firmware 27.2 method table (iomfb_v27_methods("27.2.0")) drives the
J813's 27.2 DCP: with it the DCP boots, the panel comes up at 2560x1600
and sway presents on it.
Admit 27.2.0 on T8142 as firmware 27, so that probe picks that table.
Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de>
Describe the status word and source mask that the sleep driver owns on the Gen4 SPMI controller. T8142 needs the corresponding peripheral interrupt enabled as well as the RTC alarm programmed to wake the AP. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de>
Programming the RTC compare is insufficient when the Gen4 SPMI peripheral source remains masked. J813 reached PMU stage 0x68 on its second sleep; RTC_CTL showed FIRED, but the fresh aura entry marker never changed. Enabling peripheral source 2 manually allowed a second physical resume, with NUB.SPMI0.SW3 reported and the boot ID preserved. After successfully programming the next alarm, acknowledge only the DT- owned stale sources and enable their route. Restore their previous enable state during wake, preserving other sources. This avoids using an old latched alarm as the next wake event. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de>
The J813 ADT assigns PMU peripheral interrupt 2 to the RTC path. macOS enables it at nub-spmi0 offset 0x400 and acknowledges it at 0x600. Describe that source for the sleep driver. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de>
T8142 has the same per-cluster power-gate enable words as T8132 and T8140. Assisted-by: Claude Fable 5.2 Signed-off-by: Kyle Arrington <kyle@arrington.co>
iBoot leaves the boot core's power-gate enable (P core 0, /arm-io/pmgr reg[34] + 0x60) clear on T8142, so the boot core never powers down in WFI. The node lists the E and P clusters' first words, which apple-cluster-pwrgate sets early in boot. T8132 and T8140 describe the same words; there they are enabled for deep sleep, here for idle power. Tested on a J813 (MacBook Air 13" M5, firmware 27.0), same kernel with and without the node, on battery, SMC PSTR: - PCPU_0's power state word reads off in idle with the node, on without it; - screen off, idle: 2.72 W without, 2.54 W with (12 samples each); - screen on, idle: 3.93 W without, 3.83 W with (3 samples each). All 10 cores come up and the system runs as before. The register layout is from m0d16l14n1's J815 sleep work (m0d16l14n1#1). Assisted-by: Claude Fable 5.2 Signed-off-by: Kyle Arrington <kyle@arrington.co>
The T8142 platform wake retrains the radio port with its reset values. The MSI doorbell address and enable bit programmed at probe are then lost, while the PCI core only restores the endpoint MSI capability. Keep the mapped ports and replay their MSI routing during noirq resume after a firmware-driven sleep. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de>
Pause ring doorbells, request ACIPC host sleep and require its shared-memory acknowledgment before the link goes away. Wait for the wake acknowledgment before restarting polling, and unwind a refused entry. Restore the DMA-backed endpoint MSI capability after PERST; it is programmed outside the PCI core's allocated-vector state. The native 26B123 AirshipDK protocol is verified with the installed J813 firmware. Enable the handshake on J813 only; other boards need qualification.
Describe the J813 SEP, mailbox and power domains, publish its alias for aura's firmware/manifests handoff, and admit J813 to the T8142 profile. Leave sep_dart disabled until the SEP module enables it. Describe the ADT's SID0 DMA aperture directly with no DART offset. The profile fallback gives the DMA API a low address and adds BIT(40) only to the page-table mapping; the SEP consumed that registration without replying. The explicit high window matches the existing J813 Touch ID work (7003518) and the macOS s05 registration trace. On J813, fresh attach with xart_writes=0 provision_keybag=0 trusted_keys=0 filevault_api=0 changes from zero replies at IOVA 0xfffc0000 to all 12 endpoints ready at IOVA 0x100fff80000. The key store opens and hwrng registers. The sleep dry run passes with eight coprocessors; a real commit still resets with no aura wake marker. This is sleep bring-up, not Touch ID or suspend/resume qualification. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de> (cherry picked from commit c1a2048)
The J813 GPU driver is about to identify the firmware build iBoot loaded before it starts it, as the M5 Pro and Max do: it measures both roles' __TEXT, which iBoot loads back to back (ADT gfx-asc and gfx1-asc segment-ranges, 0x50000 + 0x4c000 bytes). Add the region as fw-text. Like fw-data, it is fixed at the address iBoot uses under macOS 27.0 (26A428), 0x100_00e4_4000, until the boot loader fills it in. Under 27.2 (26B123) iBoot loads the TEXT 64 KiB higher, at 0x100_00e5_4000, and the DATA at 0x100_0245_8000, so a 27.2 boot needs a loader that fills both. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de> (cherry picked from commit 1fadec2)
The J813 runs the common 26A428 InitData graph, the J873's interface. macOS 27.2 (26B123) ships a new G17G build (RTKit 3514.41.1), which never acknowledges that graph: the secondary's InitData times out. The 26B123 interface differs from 26A428 in three places, the same three as the J873's G18G build of this release: - both roles expect the InitData root version word with bit 45 set, 0x0c8b_6339_83a0_04c0; - the policy grows by 0x60 and the status config by 8; the other 22 allocations keep their sizes. Everything after the policy in the shared cluster moves by 0x80; the page total and the private cluster (0x6d8 still packs to 0x700) stay as they are; - policy fields +0x2f0c..+0x2f17 move by 0x10 and +0x2fb4..+0x6d9f by 0x18, with a new word at +0x2f28; status-config fields +0x59c..+0x5b3 move by 4 and +0x5b4 to the end by 8, with a new word at +0x5b8 and +0x59c left zero. The new words carry a new accelerator power-controller field, which the generic power and performance setup sets to 5000 and the G17G one leaves alone. The HW data, the roots, the main configs, status B and region C keep their field offsets. Add the 26B123 profile and graph. The 26A428 encoders run unchanged and their policy and status config are moved in place by bands; a nonzero byte no band carries is refused. Compile-time checks keep the bands ascending and inside the new sizes. The board for it is J813_26B123; no configuration selects it yet, and the 26A428 encoding is unchanged. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de> (cherry picked from commit 2bb41c8)
The J813 ran a fixed variant: it started whatever firmware iBoot had loaded on the 26A428 graph. Under the macOS 27.2 stub that is the 26B123 build, which needs its own graph. Identify the loaded build as the M5 Pro and Max do, from the fw-text region: both roles' TEXT (G17G c000, then c010), with the entropy records and the ranges iBoot fills zeroed (the image header's load fields, the boot arguments before each role's GKTS record and the records from it). The 26A428 build keeps the J813 configuration; the 26B123 build runs the board on the 26B123 graph. Both digests reproduce from the stock images in the IPSWs with the same ranges zeroed and were measured on 27.0 (two boots) and 27.2 boots. An unknown build is refused before anything is written to the GPU. Tested on a J813 under the macOS 27.2 stub, with fw-text and fw-data at the 27.2 addresses: both roles acknowledge the 26B123 InitData, the render node registers, and Mesa runs GLES2 16955/0, GLES3 44379/1 (an order-dependent texture-units case that passes alone) and GLES31 37339/0, matching the 27.0 results. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de> (cherry picked from commit 349d61e)
…ops, wake restarts) Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de> (cherry picked from commit 0367ba8)
The firmware 27.2 table has no ID for dfb_ready_query, so boot_5() skips the query on 27.2. The 27.3 trace sends it at A385, one above its 27.0 ID, as 27.2 moved set_create_DFB and vi_set_temperature_hint up by one. Give the 27.2 table A385, so 27.2 boots with the same sequence as 27.0. The J813 ran with this table on 27.2: boot_done() acknowledges the DCP's boot callback only after the query's reply, and the DCP booted and the panel came up. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de> (cherry picked from commit 801cec0)
The M5's DCP announces powerlog-service on ep:20 with only
ep-has-desc-mgr and interface-id. AFK matches such a service by its
announced name, but only for the H17G method profile, and a T8142 on
firmware 27 runs the default profile. So on 27.0 and 27.2 the service is
never bound ("AFK[ep:20]: failed to extract init props: -2"), and the
mNits reports it carries, the loader-level hint that
dcp_backlight_takeover() takes on firmware 27, could never arrive.
Fall back to the announced name for a T8142 on firmware 27 as well: the
H17P-generation processor (hw.firmware_compat) that probe admits as
firmware 27. The service-rebind host test's apple_dcp gets that field.
On 27.2 the service now binds on channel 1. That firmware sent no report
in the first minutes after boot, so the backlight still starts from the
fallback level there.
Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de>
(cherry picked from commit 0e76ef5)
On firmware 27 the first present of a new KMS client (modetest, kmscube) fails with "present dropped: backlight prepare -61" and "swap failed!", and modetest -v then sees no page flip. With the panel backlight active, iomfb_apply_backlight_h17p() drops every present that dcp_backlight_prepare() refuses. dcp_bl_prepare() refuses a present without a surface with -ENODATA, because such a present cannot apply a new level, so the whole present is dropped rather than only the level. Treat a present without a surface as one with nothing to reserve: it carries the last completed level and is sent, and a pending level waits, still dirty, for the next present with a surface. Only the H17P/firmware 27 path asks without a surface; the M4 and 14.7 callers always pass one. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de> (cherry picked from commit f99c37a)
Port the audio power policy from the successful J813 experimental stack: do not apply T8132's forced state-4 list to these no_ps domains. Keep the updated sleep branch's own PMU diagnostics. The old diagnostic hunk is omitted because it uses the same scratch bytes differently. This prerequisite has not been isolated as the cause of a failure. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de> (cherry picked from commit 8dd86e4)
The system-suspend callback skipped every s2r_untested instance even after probe admitted J813 through appledrm.s2r_untested. With the panel active, the DCP kept RUN and the sleep controller refused entry. Use the existing qualified DCP quiesce and restart path on J813 during system sleep. Preserve the explicit opt-in and all existing readiness checks; other unqualified boards keep their prior behavior. Validated active-panel sleeps with and without fbcon bound, followed by three systemctl suspend cycles. The panel, GPU rendering and Bluetooth recovered each time. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de>
After a battery lid suspend on J813, ANS failed its IOP IDLE handshake and SMC faulted at the ANS mailbox. Linux resumed, but ANS could not restart and DCP power-on timed out. The matching native trace idles ANS before the final SMC MBSE/AP IDLE sequence. Defer that sequence until the peer stops have finished. Reserve the existing bounded SMC transport before IRQ shutdown and check ownership before irreversible stops. Release it at platform wake, retaining AP IDLE until macsmc resume_noirq so the original wake-notice timing is preserved. Earlier tests with AP ON in platform wake had Alpha Wi-Fi aborts; preserve the established ordering without claiming that the timing change alone explains those failures. Keep the J813-only diagnostic switch, command sequence and delay floors. Three RTC and two battery lid wakes completed on the candidate stack. The resulting source tree is identical to tested candidate 651007b. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de>
T8142 loses its common PHY and PIPE state across a physical sleep outside genpd's normal transitions. Mark that loss at resume and use the existing baseline/mode restoration path before another controller or Type-C operation. Keep the dummy PHY enabled when parking so the next PIPE handoff has a producer after the boot state is lost. On J813, hot-plug after battery lid wakes timed out in PIPE parking and blocked later suspend during device prepare. With this change the hub and RTL8153 enumerate at USB3 speed after an empty-port deep sleep, establish Ethernet SSH, and pass another suspend preflight. A hub left attached across sleep also re-enumerates. Alpha Wi-Fi still aborts across some physical sleeps; this change does not address that failure. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de>
J813 can resume with Alpha aborted and an immediate wifibt wake while the other devices recover. Send the disconnected host-power request before ACIPC host sleep, and restore ACIPC before the awake request. Match the installed firmware's event-filter limits. Pause ring traffic and background work across the transition and require acknowledgments before restarting producers. Keep the bring-up path behind apple_centauri.alpha_sleep_untested=1. Three cancelled preflights and four physical RTC cycles across two boots passed with active panel, M5 GPU, Bluetooth, Wi-Fi-bound traffic, USB Ethernet, CPUs and direct storage reads. The second boot never initialized the firmware logger. The complete source tree matches candidate 53e28ba. Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de>
Run the T8122 AUX, AUSPLL and lane sequences for T8142. Apply loader-supplied LIOA and common-configuration calibration before AUX, using separately claimed resources. Refuse DP without the complete calibration bundle. Serialize lane validation with Type-C mux setup so DCP sees lanes only after AUX is ready. No board identity or out-of-resource MMIO mapping is needed. Signed-off-by: Dmitriy <noc@pimk.in> (cherry picked from commit df6cd4a in dmitzsaz/linux) Signed-off-by: Nick <nekreshen@yandex.ru> (cherry picked from commit 4e47750)
A port left at reset by the loader retains bit 16 of the nonselected PIPE override. The dummy backend cannot acknowledge its lock while this hold is asserted. Clear it before enabling that backend on T8142, including ports which the loader did not initialize. Signed-off-by: Dmitriy <noc@pimk.in> (cherry picked from commit def4169 in dmitzsaz/linux) Signed-off-by: Nick <nekreshen@yandex.ru> (cherry picked from commit 0db8f37)
A device tree written before DisplayPort was described has no "lioa" window. Probe as before then and leave DisplayPort off, instead of failing the probe and taking USB down with it. Signed-off-by: Nick <nekreshen@yandex.ru> (cherry picked from commit 7292f7d)
… DisplayPort calibration
On T8142 the mode switch refuses every mode with DP AUX while the boot
loader's LIOA and common configuration banks are missing ("DisplayPort
bank calibration missing; DP disabled"). A partner in DP Alt Mode then
loses its USB as well: on a J813 with a dock in pin assignment D, under
a loader without the bank export, the PHY init fails with -EOPNOTSUPP
and the port's dwc3 never probes, where the kernel before the T8142
DisplayPort sequence ran the dock's hub at 10 Gb/s.
Map the DP modes to the USB the partner keeps, as T8132 does: USB2 next
to four DP lanes (pin assignments C and E), USB3 next to two (D).
Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de>
(cherry picked from commit ca7da9d114f79b533e2bf33834876dbcf982d7c9)
…-20261117 wifi: centauri: recover after DMA was released before physical reset
T6050: a "freeze" with mem_sleep at deep is a suspend-to-idle (no dock teardown, no -67)
apple_smc_dark_wake() runs for every board with an SMC sleep policy, and the J716c has one too. It looks its two bits up by name, so on the J716c a USB-C_plug-only or wifibt-only wake would also go dark, on a board where that was never chosen or tested. Return early unless the wake table is the J714s's. A port's wake with the lid closed is now made dark before this runs (apple_sleep_port_wake()). Count such a wake too, so the re-entry after it waits for the SMC as it does after one of this function's own; leave its log line and its decision to that rule. Signed-off-by: Diljit Singh <hey@djdev.me>
accel/ane: h16, h17: give program memory and DVA back after close
arm64: dts: apple: t6050: describe the MTP only on the laptops
The J714s's third USB-C port, the left-back one next to MagSafe (ATC0),
has no description, so nothing plugged into it works. Add its USB 2/3
chain: the SN201202x hpm0 on nub-spmi-a0 at 0xc with its connector, the
combo PHY with its firmware mailbox, usb-drd0 and its two DARTs, and the
ATC0 power domains.
The values are the left-front port's (ATC1) moved to ATC0: register
blocks less 0x8000000, interrupts less 32 (the PHY mailbox's less 10).
They match the J714sAP ADT, J716c's ATC0 and the T6050 inventory, whose
usb_drd0, usb0_dart, usb0_dart1 and ps_atc0_* nodes this configures; the
DRD's tunables are the same on every port. The PHY's DP clock source
register, which the ADT does not give per port, is ATC1's less 0x10, as
on J716c.
The inventory makes the FAB3 SoC domain ATC0_COMMON's parent. That
domain is not described on J714s, and ATC1's and ATC2's common domains
have no parent here, so drop it rather than hand FAB3 to Linux (J716c
enables ps_fab3_soc instead, which puts it under genpd's unused
power-off).
The connector is typec2, after the right (typec0) and left-front
(typec1) ones. hpm0 joins the other port controllers in
t6050-j714s-sleep.dtsi with apple,sleep-ssps.
This is the debug (DFU) port, as the ADT marks it (hpm0 has
uart-hpm-rids, usb-drd0 the NCM debug gadget's settings), and the boot
loader's debug link: the console on serial0 (the dockchannel UART) and,
on request, its USB proxy, whose USB2 route it sets with one DVEn action
in a short DBMa session. The boot loader has so far kept the port out of
the DT on purpose ("Port 0 carries aura's debug link: never describe
it"), so every node here starts disabled and the port stays the boot
loader's until it hands the port over. Once it does, tipd, the PHY and
the DRD's role switch decide what the USB2 pair carries, and the port's
domains and DARTs are Linux's, including genpd's power-off when unused.
The boot loader exports the PHY's tunables for ports 1-5 only, so this
port's are in the DT, from the J714sAP ADT's atc-phy0 (and usb-drd0's
PIPE handler table). They are ATC1's except in axi2af, aus40cmn-shm and
cio4pll.
Signed-off-by: Isaac Duarte <31712685+Isaac-Duarte@users.noreply.github.com>
Add the left-back port's (ATC0) USB4/Thunderbolt router and its PCIe tunnel, so a USB4 or Thunderbolt dock works there as on the other two ports: acio0 with its CPU mailbox, DART and NHI, the PCIe-C root port apciec0 and its DART, their power domains, the LSTX pin, the CIO reconfig reset and the link from hpm0's connector. Like the rest of ATC0, they start disabled. Every address and interrupt is the left-front router's (t6050-j714s-left-usb4.dtsi) moved to ATC0, and the result is the same as J716c's ATC0 router and tunnel (t6050-j716c-left-usb4.dtsi), including the LSTX pin (AOP GPIO 77, now exposed with 79 and 81), the PCIe windows, MSI slice 128-159 and PCI domain 1, which nothing else uses here. The inventory's apciec0_dart gives way to the tunnel's, as for ATC1. The AOP GPIOs' reserved ranges, which leave only the routers' LSTX pins, move from t6050-j714s-left-usb4.dtsi to the board file, the one place that knows all three. The router, NHI, DART and tunnel tunables are ATC1's: they are the same on ATC1 and ATC2 here and on J716c's ATC0 and ATC1. The router also gets its four DP IN interrupts (AIC 1603, 1604, 1609, 1610; the inventory's atc0-dpin0..3, as on J716c), so DP IN events need not be polled once a display pipeline is routed to this port. The host DROM and the PHY's CIO80 RX_TOP tunables are the J714sAP ADT's (acio0 thunderbolt-drom, atc-phy0 tunable_LN0/LN1_RX_TOP_CIO_DFLT). The ADT's acio0, acio-cpu0, dart-acio0, apciec0, dart-apciec0 and atc0-dpin0..3 registers, interrupts, LSTX pin (77) and PCIe ranges and domain agree with the values above. Signed-off-by: Isaac Duarte <31712685+Isaac-Duarte@users.noreply.github.com>
On T6050 every Type-C port/pipeline pair gets a DRM connector, numbered pipeline by pipeline, with tunnel-only routes after the others. Userspace keys per-monitor settings on the connector names, so a port whose routes are added to a board later must not take names the existing connectors have. Add a route property, apple,connector-group (0 to 3, default 0): the connectors are created group by group, so a port added to a board in the next group gets the next names. It is meant to be the same for all routes to one port. On SoCs with a connector per port/pipeline pair, the tunnel-only routes' come after the others in each group. Signed-off-by: Isaac Duarte <31712685+Isaac-Duarte@users.noreply.github.com>
apple_probe_typec_ports() creates T6050's port/pipeline connectors pipeline by pipeline, and the tunnel-only pairs in a second pass. A port added to a board takes a place in every pipeline's run, so every later connector is renumbered: on the J714s, a left-back port would turn dcpext2's right-port connector from DP-3 into DP-4, and so on. Create the connectors group by group (apple,connector-group, default 0). The number of passes is two per group, up to the largest group a route has. On T6050 the tunnel-only pairs come after the others in each group: dcp_typec_port_pass() replaces dcp_typec_port_tunnel_only() and gives a pair's pass. Elsewhere a port's routes share its connectors, which take the first pass of the port's largest group. Without the property nothing changes. A malformed group, or one above 3, fails the route's registration. The topology test builds the J714s's fabric (three ports, dcpext0 to dcpext2 with dcpext1 tunnel-only) and checks the connector order: DP-1..DP-6 as before for the right and left-front ports, and the left-back port's pairs as DP-7..DP-9 in group 1 (or a later group), where in group 0 they would take DP-3. It also checks a port's shared connector pass. Signed-off-by: Isaac Duarte <31712685+Isaac-Duarte@users.noreply.github.com>
…isabled A DCP's Type-C routes are all or nothing, and a disabled PHY or crossbar never registers, so looking it up defers for ever. A failsafe external DCP (T6050's) waits for it in its supplier preflight, dcp_ext_route_suppliers(), until deferred probing gives up (about 10 s after boot), then joins the display subsystem as an unusable pipe with every route disabled. A native T6030 one disables all its routes, and any other fails its probe. One port's disabled PHY so costs every external display on every port. Leave out a route whose PHY or crossbar node is disabled, in the preflight and when the routes register: dcp_typec_route_supplier_disabled() makes that decision for both. Registration also disables the route in the device tree, as is done for all of them on failure, so that its port's controller stops waiting for its mode-switch. Only a disabled node does this. A missing or malformed phys or mux-controls entry, and every lookup error (-EPROBE_DEFER included), defer or fail as before, and a supplier that is merely late is waited for, never dropped. A PHY that is enabled but fails to probe holds every DCP that lists it until deferred probing gives up, as before: a failsafe dcpext then joins as an unusable pipe, with no route. dcp_typec_routes_disable_dev() and the new single-route disable share one helper, which builds and applies one changeset. The topology tests apply an overlay with an available and a disabled PHY and an available and a disabled crossbar. They check the per-route decision, which routes registration skips, and that a skipped route ends up disabled (the test reverts the changeset). They need CONFIG_OF_OVERLAY; tests/topology.kunitconfig has what the topology suite needs: tools/testing/kunit/kunit.py run --arch=arm64 \ --kunitconfig=drivers/gpu/drm/apple/tests/topology.kunitconfig Signed-off-by: Isaac Duarte <31712685+Isaac-Duarte@users.noreply.github.com>
Give the left-back port (ATC0, typec2) the display routes the other two ports have: DP alt mode and USB4 DP tunnels through dcpext0 and dcpext2, and tunnels only through dcpext1. - Each DCP gets route@2 to a new typec2 endpoint, the ATC0 PHY's DP output (DPTX PHY 0) and the left-back crossbar's controls, named typec2 and typec2-dpin0..3, with the crossbar sources it uses on the other ports (dcpext0 0, dcpext1 2, dcpext2 4). - dcpext1's route is apple,tunnel-only: hpm0, like hpm1 and hpm2, has TYPEC_MUX_MAX_DEVS (3) mode-switch providers with the PHY, dcpext0 and dcpext2. - The three routes are in apple,connector-group 1, so their connectors come after the existing six: DP-1..DP-6 keep their ports and pipelines, and the left-back port's are DP-7 (dcpext0), DP-8 (dcpext2) and DP-9 (dcpext1). - The inventory's atc0_dpxbar gets ATC0's always-on and common domains, as atc1_dpxbar, and the PHY the four ATC0_COMMON_DP tunnel clock gates, as J716c's atc0_phy. Like the rest of ATC0, the crossbar and those gates start disabled, so the three routes are skipped until the port is handed over. - Every DCP now lists the left-back PHY and crossbar. A route whose PHY or crossbar node is disabled is left out, but a PHY that is enabled and fails to probe holds every dcpext until deferred probing gives up (about 10 s); they then join as unusable pipes with no route, so the panel works and no external display does, on any port. The header comment says so. Signed-off-by: Isaac Duarte <31712685+Isaac-Duarte@users.noreply.github.com>
Describe the disabled device group a boot producer may hand to Linux only after releasing its debug/USB ownership and supplying per-machine PHY tables. Keep shared providers outside the group. Signed-off-by: Diljit Singh <hey@djdev.me>
The left-back port can still belong to the debug producer, and its static PHY tables do not match every J714s ADT. Remove those tables and the static ownership assertion. Keep the ATC0 devices, power domains and crossbar disabled until the producer exports its own ADT values and enables the complete handoff group. Keep LSTX reserved until that transaction releases it. Existing ports and their shared AUSS, GPIO and CIO reset providers remain available. Older producers safely leave the new port disabled. Signed-off-by: Diljit Singh <hey@djdev.me>
The core chosen schema is closed, so an extra schema cannot make its vendor property valid. Move the group into an explicitly compatible firmware node with a required device list and a closed binding. Keep the existing ownership and disabled-device rules intact. Signed-off-by: Diljit Singh <hey@djdev.me>
Nothing in the J714s ATC0 handoff is specific to that board or port: it is a group of a Type-C port's devices that stays disabled until the boot producer releases the port it may be using. Rename the binding to apple,atc-handoff so other boards whose producer's USB port sits on an ATC can use it, and name the port's PHY with apple,phy, so the producer does not have to find it among the devices. Signed-off-by: Isaac Duarte <31712685+Isaac-Duarte@users.noreply.github.com>
#2469 kept the GPU domain raise at the sleep commit's stop on the T8142 and on every apple,agx-g17s part. The G17S M5 boards (J714s, J716c) never run it: they stop the GPU through their own M5 commit stop, which this doesn't touch. On them the flag only made asahi_g17_commit_finish() print "GPU domain at the commit stop: 0x0, raise to full 0" from fields the M5 stop never fills, which reads as a raise that didn't happen. The other G17S boards (K114c, K116s) take the generic stop, where the raise was never measured; their sleep was tested before it came in. Keep the raise to the T8142, the SoC it was made for (the J815's GPU gates itself between frames and its registers may not answer below full power): the other G17S boards go back to the stop they were tested with, and the M5 boards stop printing the empty line. Signed-off-by: Diljit Singh <hey@djdev.me>
media: apple: isp, arm64: dts: t6050-j714s: cut the camera's grain 5x as macOS does
HID: magicmouse: keep host click feedback across system sleep
arm64: dts: apple: t6050-j714s: the left-back USB-C port (USB, USB4, DisplayPort)
ASoC: tas2764: Bring an SDZ group back from a deep sleep as from power-on
…il-closed-20261119 ASoC: tas2764: Fail an SDZ group closed when its restore after a reset fails
soc: apple: sleep, mfd: macsmc: T6050: a port's wake with the lid closed is a dark wake
drm/apple: allow asynchronous PM for the J714s internal DCP
usb: typec: tipd, dwc3: apple: J714s: replug a DP alt mode hub whose SuperSpeed side didn't come back
…-wake J714s: take port-only and wifibt-only SMC wakes as dark wakes (macsmc, dwc3, thunderbolt, ane)
drm/asahi: g17: keep the commit-stop raise to the T8142
Neo J700 qualification of the Edge alpha.2 pair (11-20/21)Pair: kernel-aurora 7.3.0-43 (linux
M5 J714s: rgate35c PASSED on the same signed kernel + Mesa 26.3.0-35 (DSO Owner qualification JSON ( |
|
Correction to the qualification summary above: the Vulkan CTS was not unpatched. It was
All other results are unchanged. That includes |
Bring
aurora-edgetointegration/aurorae0ee46a18e33for the next Neo Edge kernel.aurora-edgeis at6020f67d(the 7.3.0-42 Edge, source6f618bcc). This head isintegration/auroraas of 2026-11-20 05:0x UTC, 385 commits ahead. The edge-only merge commit adds nothing to the tree, so the merged tree ise0ee46a1's.Why now, for the Neo (J700, T8140):
integration/aurora1473d156, an ancestor of this head, passed a 24/24 Neo deep-sleep soak on 11-19/20. That soak used the shipped Mesa and matched the 7.3.0-41 default. This head adds #2240, #2507, #2510, #2233, #2026, #2017, #1996, #2515, #2445, #2077, #2274, #2231 and #2521, mostly T6050 work. The Neo-path files they touch aresoc/apple/sleep.c,rtkit.c,kernel/power/suspend.c,dwc3-apple.c, tipd,macsmc.candg17/sleep.c(#2521 keeps the commit-stop raise on the T8142 only; the T8140 never took it).Before merge, the exact Fedora 44 RPM pair built from this head and the Mesa Edge candidate (aurora-silicon/mesa
djdev/edge-neo-mesa-20261120) gets:Results will be posted here.
🤖 Generated with Claude Code