Repository navigation
Conversation
|
[Joshua's agent] Friendly heads-up for when this board lights up under a Hyprland 0.57 desktop: the |
|
[Joshua's agent] Correcting my heads-up above. Two things I said were wrong for this board. First, the "frames two updates late" symptom is not what stock Hyprland does: it came from our own downstream aquamarine patch (hyprwm/aquamarine#496, closed), which allocates AR24 when XR24 is missing; on a stock stack a plane without XR24 fails to start on that output ("GBM: Failed to allocate a GBM buffer: format XR24 isn't supported by primary backend"). Second, the 1816 fix I linked covers firmware 27.3 only (opaque-X property id 0x4f verified against the 27.3 kernelcache table) and makes no claim for the firmware this board runs; I should not have implied it transfers. What we measured, on our M5 panel with the 27.3 firmware only: buffer on the plane held the frame from two back 34 of 34 with XR24 missing, current 2987 of 2987 once XR24 was offered with X opaque. Whether that applies to 27.0/27.2 on this board is unmeasured; treat it as one data point, not guidance. |
The J704 shares the J813's T8142 SoC, GPU identity (family 0xb, variant 2, one cluster of ten core slots) and 26A428 COMMON26 firmware interface, but not its power data. Add a J704 board on the COMMON26 graph, built from the J813 one with the values that differ on this machine, each taken from the J704 ADT or the init data macOS 27.0 hands its GPU firmware: - 14 perf states (13 GPU states, 338..1620 MHz) with the J704 voltages, and the relative-performance table; - the AON rows as macOS publishes them (RT0g with index 0x28); - region C words, power-policy words (power limits 1300/38110) and the idle standby time, through two new board fields that are empty for the other boards; - no clearing write at hw_data 0x1118, which is the 1620 MHz entry here. Encoded on the host, the J704 HW data differs from the captured macOS data only in runtime pointers; region C and AON are identical. hw::t8142::select() picks the board by the root compatible (apple,j704; everything else keeps the J813 config) and fails the probe if the compatible cannot be read, rather than falling back to J813 values. The new aurora.m5_probe_only=1 parameter selects PROBE, which logs the ID registers and core masks and refuses the GPU before the UAT or the firmware are touched, on either firmware path. Tested on J704 (macOS 27.0 firmware, booted through aura): with m5_probe_only the ID registers and core mask 0x3ff are logged and the GPU is refused; without it both roles acknowledge their InitData and "[drm] Initialized aurora 1.0.0 for 582600000.gpu" registers renderD128. With Mesa aurora-wip, GLES 3.2 and Vulkan run on it: the dEQP GLES2, GLES3, GLES31 and Vulkan caselists (about 1.1 million tests, on the 15 November base) pass with the agx fixes that are being sent separately, and Ubuntu 26.04 runs with this GPU from an NFS root. Assisted-by: LLM Signed-off-by: termg <hello@termg.com>
Describe the J704 GPU like the J813's (t8142-j813-gpu.dtsi): sgx, the gfx and gfx1 ASC mailboxes, pmgr_gpu with ps_gfx, and the uat-*, gfx-data and gfx1-data reservations that the boot loader fills, plus fw-data and fw-text, the firmware DATA and TEXT the driver reads and identifies the build by (both at the macOS 27.0 addresses, as on the J813; the J704 ADT gfx-asc/gfx1-asc segment-ranges give the same TEXT span). Addresses and interrupts match the J704 ADT. The OPP table has 14 entries, since this machine's ADT has perf-state-count 14 (the J813 has 13). Assisted-by: LLM Signed-off-by: termg <hello@termg.com>
The firmware-27 DCP interface asks for its provider clocks by index into the ADT /arm-io/disp0 clock-ids, and fails to start without apple,clock-ids. Describe them as on the J813: the J704 ADT has the same ids, 585 and 431, and entry 0x49 of its clock-frequencies (the one 585's low byte names) is the same 712 MHz. 431 is never asked for. Tested on J704 (macOS 27.0 firmware, booted through aura): the DCP driver binds, "DCP booted", eDP-1 connected and enabled at 3024x1890 inheriting the boot mode, and appledrmfb takes over the console. Assisted-by: LLM Signed-off-by: termg <hello@termg.com>
07d3d71 to
83650c7
Compare
|
Rebased on integration/aurora 59b02d5 (the conflicts were with the 27.2 / 26B123 support). Two changes besides the rebase:
Tested on J704 (rebased branch + #2094 + #2095 + bring-up test trees, macOS 27.0 / 26A428 firmware): the firmware is |
|
The production boot stub pins at macOS 27.2 (26B123), so please retarget this to 27.2. The J704 board data, fw-text range and DCP clocks here are taken and tested under the 27.0 (26A428) stub; please bring the J704 GPU and display up on the 27.2 stub (the 26B123 T8142 graph from #2135 is on integration/aurora now). |
|
Thanks for bringing up the J704 GPU and display. One part doesn't match how we usually select behaviour. This PR already adds the J704 GPU node with its 14 OPPs, and the boot loader fills the GPU node from the ADT. Could the per-board values (perf states, voltages and the other words that differ) come from that node, with the probe refusing a board whose node lacks them? This goes together with the 27.2 retarget question already open here. We can run no-regression rows on our A18 Pro and M4 Pro. |
|
One follow-up to the two open points above, so the 27.2 re-test can be done in a single pass. The display part looks right: When the board data comes from the GPU node and
Please post the dmesg, the mode list and the glmark2 output. We will then run the same driver change on our A18 Pro and M4 Pro to check that nothing regresses. |
|
@term-guy, a follow-up to the review above, about the rest of the J704 rather than this PR. To bring the J704 up to the J813/J815 level we need a few things that only a J704 can give us.
dmesg plus the command output, here or in a new issue, is perfect. Thanks again. |
The T8142 GPU's power data differs between the boards that use the SoC: the J704 (MacBook Pro 14-inch, M5) has fourteen performance states up to 1620 MHz, other voltages, relative power, AON rows and power limits than the J813 (MacBook Air, M5). Describe it in the GPU node so that each board states its own, and require it on apple,agx-t8142: - apple,sram-microvolt: the SRAM voltage of each operating point (the operating points carry the core voltage, one cell, as the boot loader fills them from the ADT's sgx perf-states); - apple,m5-relative-power: each operating point's relative performance, IEEE single bit patterns; - apple,m5-aon-records: the AON rows the firmware reads; - apple,m5-hw-words, apple,m5-region-c-words, apple,m5-policy-words: <offset value> pairs where the board's firmware data differs from the SoC's template in the driver (region C and policy may be empty); - apple,idleoff-standby-timer: the idle standby time, us, the name the G15 boards' power words already use. Assisted-by: LLM Signed-off-by: termg <hello@termg.com>
Give the J813 and J704 GPU nodes the power data the binding now requires, so the driver can take it from the node rather than pick a table by the machine compatible: - t8142-j813-gpu.dtsi: the values the driver's J813 board has carried so far (perf-states-sram of the J813 ADT, the relative power and HW words read out of macOS 27.0's own data on a J813, the RT0g AON row, no region C or policy words, the template's 1250 us standby timer). The J815 includes this file and has run the J813's board until now; it keeps those values. Its operating points stay filled in by the boot loader from its own ADT, which the driver now reads. - t8142-j704-gpu.dtsi: the J704 board's values (its ADT and macOS 27.0's own InitData on a J704), as the driver's J704 board had them: SRAM voltages, relative power, the four AON rows, the HW words without the J813's clearing write at +0x1118, the five region C words, the 54 power policy words and the 2000 us standby timer. Assisted-by: LLM Signed-off-by: termg <hello@termg.com>
hw::t8142::select() picked the GPU power data by the root compatible: the J704 table on apple,j704, the J813's on every other T8142 board. A board that is neither ran with the J813's performance states, voltages and power limits, and the J704 could only run the 26A428 (macOS 27.0) firmware, since the 26B123 (27.2) configuration carried the J813's board. Keep only the SoC's part in the driver: the identity, core layout, firmware builds and a T8142 board template (graph, register windows, MCC mask, render pool, the J813's) with no performance state. The probe reads each board's power data from its GPU node (BoardPower): the frequencies and core voltages from its operating points, the SRAM voltages, relative power, AON rows, HW data, region C and power policy words and the idle standby timer from the properties the binding describes, and completes the firmware build's template with it (Board::with_power). The relative-frequency curve is computed from the operating points with the same IEEE single roundings as before, in integer arithmetic; it matches the floating-point one for the static tables (checked at compile time). The probe fails closed: a GPU node without one of these properties, with malformed ones (counts, alignment, a word over the policy's pointers or over a used state's frequency, operating points the loader did not fill) or with policy words the build's layout has no place for is refused with "M5: GPU node has no valid <property> ..." or "M5: invalid GPU power data in the device tree: <reason>", and no board falls back to another's values. The board is kept for the boot (a rebind must read the same). aurora.m5_probe_only stays as the off-by-default diagnostic. With the board no longer tied to a build, the J704 runs both builds: on 26B123 its words are applied in the 26A428 layout and moved with the rest of the policy block, as the J813's are (every J704 policy word lies in a band the 26B123 layout carries). Encoded on the host, the boards built from the J813 and J704 trees give InitData (HW data bundle with the power policy, region C, AON) byte for byte identical to the static J813 and J704 boards they replace on 26A428, and to those boards on the 26B123 graph; on 26B123 the J704's policy differs from the J813's only at its own words, relocated. Assisted-by: LLM Signed-off-by: termg <hello@termg.com>
|
Thanks, all three done. 1. Board power data from the GPU node, fail closed. Three new commits on top of the branch (no rewrite of the existing ones):
(860fde0, 4450dd1, f41fdbe.)
Properties without an ADT source are explicit arrays in the board .dts. A node that is missing any of them, or has malformed ones, is refused with Equivalence: the J704 DTB's GPU node matches the previous static J704 board exactly:
The template's non-power fields equal the old J813 base. So on 26A428 the J704's InitData is unchanged. Fail closed on hardware: the same kernel on the same 27.2 install, booted with the J704 DTB minus The probe stops before the MMU and RTKit are touched. No GPU card and no renderD128 appear, and the display (card1) still runs. The other refusal paths (malformed counts or offsets) were checked by reading the code. 2. 27.2 (26B123) stub. No J704-specific code is needed for 26B123: the probe picks 3. J704 on 27.2. Mac17,2 J704, fresh install on the macOS 27.2 (26B123) stub. Chain: aura -> U-Boot (u-boot#55) -> GRUB -> this branch (with the Fedora DTS on top) -> Fedora KDE on the GPU.
One cosmetic thing, not from this PR: with no AI-assisted, tested by me on the hardware. |
|
Thanks, this is what we hoped for: the board's GPU power data now comes from its GPU node, the root compatible is gone from On the installer: yes, please open the J704 26B123 pin as an aurora-installer PR so it is reviewed on its own. The quieter firmware-build log also fits best as a separate small change. We will run the no-regression rows on our A18 Pro and M4 Pro on f41fdbe and post them here. |
|
The new m5_probe_only module parameter (asahi_core.rs, read in driver.rs and hw::t8142::select()) adds a bring-up mode that identifies the GPU and then refuses it. That is lab tooling rather than driver behaviour, and it gives the module a second start path that production never takes. Please drop the parameter and the PROBE selection, and keep the ID register and core mask dump as a dev_dbg or a debugfs file if it is still useful. Until then this should stay a draft. |
The m5_probe_only module parameter made the T8142 probe identify the GPU and then refuse it, before the UAT or the firmware were touched. It was a bring-up aid that gave the driver a second start path production never takes. Remove the parameter, the PROBE configuration it selected and hw::t8142::select(): the 26A428 path now always starts with CONFIG, which takes the fields PROBE used to supply through BASE. The ID registers and core masks are still logged by get_g17_id(), as on every G17 GPU. Assisted-by: LLM Signed-off-by: termg <hello@termg.com>
|
Done in c466b17 ("drm/asahi: g17: M5: Drop the identification-only probe"), on top so it is easy to review; I can squash it into 1549699 if you prefer. The parameter, About dt-checks: the three new warnings are the existing ones on the t6021-j475c, t6022-j180d and t6022-j475d GPU nodes ("... do not match any of the regexes"). Now that the binding defines AI-assisted (Assisted-by: LLM). |
… review The J704's caution still said nobody on the team has one and that aura and this installer have not run on it. It has been installed with this installer on the macOS 27.2 (26B123) stub and runs Fedora KDE, but its GPU and internal display need aurora-silicon/linux#2093, which is not yet in the distribution kernel. Say that, in supported-models.json and in the installer's Stage 1 caution. Assisted-by: LLM Signed-off-by: termg <hello@termg.com>
|
Thanks, c466b17 does what we asked: The dt-checks explanation makes sense: the three warnings are the existing t6021-j475c, t6022-j180d and t6022-j475d lists, which change only because the binding now defines When you're ready, please take it out of draft so CI runs on this head. |
|
The Assisted-by: LLM (Claude:Opus 5.5) |
Adds the MacBook Pro 14" M5 (J704) to the G17 driver and its device tree. The J704 shares the J813's T8142 SoC, GPU identity and 26A428 COMMON26 firmware interface, but not its power data.
hw::t8142::select()picks the board by the root compatible(
apple,j704; everything else keeps the J813) and fails the probe if it cannot be read.aurora.m5_probe_only=1logs the ID registers and refuses the GPU before the UAT or firmware are touched.t8142-j813-gpu.dtsi, addresses and interrupts from the J704 ADT; 14 OPPs (perf-state-count 14).apple,clock-ids585/431 as on the J813; without them the firmware-27 DCP does not start.Needs aura with the J704 perf-state fix (aurora-silicon/aura#276) to fill the GPU nodes.
Testing on J704 (macOS 27.0 / 26A428 firmware, booted through aura, this branch on integration/aurora 24b3230): the GPU initializes every boot, t1-t4 (DRM info, Vulkan compute, GLES and Vulkan rendering) pass, the dEQP runs listed in aurora-silicon/mesa#397 pass, the internal panel runs through DCP (eDP-1 3024x1890) with backlight, and Ubuntu 26.04 runs with the GPU from an NFS root (t1-t4 pass there too).
Known problem, not caused by this board data as far as I can tell: with more than four
concurrent clients running geometry/tessellation work, the firmware occasionally restarts the
GPU (it recovers; the affected render is lost). Details in omacom/linux-aurora#493.
Written with an AI assistant (noted in each commit) and tested on the hardware by me.