Skip to content

j514s (MacBook Pro 14" M3 Pro): external modes above one DCP pipe's pixel rate (5K@120/180) are dropped; needs the dispext peer-pipe split macOS uses #434

Open

Description

@JeppeBylov

Summary

On a MacBook Pro 14" M3 Pro (J514S, T6030), external modes above one display pipe's pixel-rate limit never reach Linux. The external DCP firmware drops them itself:

IOMFB removing mode: 5120 x 2160 @ 119, bw: 1316044800 max: 1274019840
IOMFB removing mode: 5120 x 2160 @ 120, bw: 1327104000 max: 1274019840

bw is active pixels per second. max (1,274,019,840 = 0x4bf00000) is the per-pipe default in the firmware. macOS gets past this by splitting one display across two pipes (dispext0 + dispext1, "peer pipe"). Apple rates the M3 Pro's HDMI port for 8K@60 and 4K@240, and both are well above one pipe's limit. No Linux driver drives that split path today, so on Linux the limit is about 1.27 Gpix/s per monitor (for example 5120x2160 at about 115 Hz at most, 4K at about 153 Hz).

I couldn't find an existing issue for this. #45 is about two-tile monitors that use two streams, which is a different case. This is one stream that needs two pipes.

Mac and installed build

  • MacBook Pro 14" M3 Pro, Nov 2023 (apple,j514s, apple,t6030)
  • Omarchy on official packages: linux-aurora 7.2.11.aurora15-1 (7.2.11-15-1-ARCH), m1n1-aurora 1.6.1.aurora22-1, mesa-aurora 26.3.0_devel.20261031, omarchy-mac-boot 20261104-1, Hyprland 0.57.1
  • /etc/m1n1.conf: chosen.asahi,t6030-dcp=1, chosen.asahi,t6030-gpu=1, chosen.asahi,t6030-dcpext=1
  • Monitor: Samsung Odyssey G7 G75F 40" (5120x2160, 21:9, 180 Hz), through an Anker TB5 dock (its HDMI and DisplayPort outputs) and directly on the Mac's HDMI port

Firmware

  • os-fw 14.7, system-fw 27.0
  • DCP firmware studied: t6030dcp.im4p from the stub's Restore/Firmware/dcp/ (plain Mach-O, stripped). The addresses below are firmware vaddrs (file offset = vaddr + 0x1000).

What works today

  • 5120x2160@60 over HDMI: works natively, DSC=NO, stable. The HDMI EDID lists it as CTA VIC 126, and DCP keeps it.
  • Lower-resolution high-refresh modes: 3840x2160@120 and 3440x1440@180 work, both with DSC.

Reproduction, expected and actual

Connect the G75F over HDMI or DisplayPort and enable DCP tracing.

  • HDMI: the EDID offers 5K at 60 Hz (VIC 126), 120 Hz (VIC 193) and 180 Hz (DTD, 2248.75 MHz, preferred). Only 5K@60 reaches userspace.
  • DisplayPort: no 5K mode reaches userspace at all.
    • 5K@120 and 5K@180 are Type 7 timings, which the pipe limit drops.
    • 5K@60 (and 3440x1440@120) appear only in a DisplayID 2.0 Type 10 (formula/CVT-RB) block. Neither is in the firmware's timing catalog (checked with iomfb_parse_mode traces, no parse failures), so the DCP firmware seems to ignore Type 10 timings.
  • Expected: 5120x2160@120 (and @180), as macOS on the same Mac and monitor gets.
  • Actual: 5K@60 over HDMI only. Over DisplayPort, nothing above 3840x2160.

What I found in the DCP firmware

Where modes are dropped. UnifiedPipeline::videointerface_filter_modes (fn 0x19cb80) logs the "removing mode" line. max comes from the pipe object (vtable +0xb98). The same function has a bypass, VFTGMerge conifg. Allowing width %d (0x19cef4), taken when the display allocation merges two VFTGs.

Split timings. DCPAVSplitDisplay ("Adding %u-way split timing element ... created from timing ID") and DCPDPVirtualDevice are present.

Peer pipe (in AppleDCPDPTXVirtualPort::handleStart, 0x2cac4):

  • It reads the u32 peer-pipe-cfg and takes the nibble for this dispext's external-index (ADT: dcpext0=1, dcpext1=2): peer = cfg >> (idx*4-4) & 0xf.
  • Default max active pixel rate is 0x4bf00000, the limit above. Default max width is in *0x5b26a8 (6720 per pipe, per earlier M4 Pro reports). All three can be overridden by firmware boot-args dpxbar-max-pixel-width-per-pipe, dpxbar-max-pixel-rate-per-pipe and dpxbar-max-active-pixel-rate-per-pipe.
  • It logs dispext%d: peer pipe dispext%d, max pixel width %d, ....
  • The firmware also has _current.displayAllocation: extraPipes=%d, mainUFP=%d, peerUFP=%d (fn ~0x81cc0). In the ADT, dcpext0 and dcpext1 share dp-switch-ufp-port 0, with dp-switch-ufp-endpoint 0 and 1.
  • The J514S ADT (DeviceTree.j514sap.im4p) has no peer-pipe-cfg.

Firmware boot-args can be set from Linux:

  • The getter (0x2bc80 → 0x2fdc4) builds "<prefix>-<name>". The prefix is the provider's boot-args-prefix, default appledptx, so the full name is appledptx-peer-pipe-cfg.
  • parse_boot_argn (0xc100) reads a bss copy at 0x88ad40 (copied flag 0x8874b8), filled lazily from the _rtk_patchbay tag bArg (0x400 bytes, at dcpextN-data + 0x14ace8, GKTS/grAb header). iBoot leaves it empty.
  • The firmware has already run once by the time Linux starts it, so both the patch bay and the bss copy need patching. /dev/mem writes must be aligned u32.

Experiment and result

I injected appledptx-peer-pipe-cfg=0x12 into both dcpext firmwares before plugging in. The firmware did hold the value after plug-in, but nothing changed. The firmware only ever creates AppleDCPDPTXRemotePort services (ep 0x2a) and never an AppleDCPDPTXVirtualPort, so the code that reads peer-pipe-cfg never runs.

So the open question is what makes the firmware create the VirtualPort or split allocation. Candidates:

  • DPTX apcall 26 GET_VIRTUAL_DEVICE_MODE: dptxep.c answers it with zeros. In my traces the firmware only called 0, 1, 2, 5, 6, 7, 9, 10, 12, 13, 14, 15, 18, 20 and 21, never 26.
  • DPTX apcall 21 SetTiledDisplayHints (144 bytes): Linux echoes the request back with retcode 1. DCPAVServiceEDID::finalizeDisplayHints appears to get the display allocation (extraPipes/mainUFP/peerUFP at obj+0x34c) from the port (vtable +0x300). A reply with a non-zero allocation might be the trigger. I built an experimental appledrm that can dump and patch this reply, but I have no result to report.
  • DCPAVControllerEPClient: it has SetVirtualDeviceMode and GetVirtualDisplayMode. Linux binds the service but sends nothing. This may be AirPlay/Sidecar-style rather than the split.

The rest of the job would be coordinating two dcpext instances on one connector (both slices open and swapping together) plus the crossbar routing.

What would unblock this

An m1n1 hypervisor trace of macOS driving a >1.27 Gpix/s mode on T6030, showing which EPIC services and calls appear on dcpext0/1 and what is returned for apcalls 21 and 26. Any monitor that macOS runs at 4K@>153 Hz, 5K@120 or 8K@60 would do. proxyclient/hv/trace_pmp.py already supports T6030, so the hypervisor runs on this SoC. I don't have a second USB machine for the hypervisor right now, but I'm happy to test patches on J514S with this monitor.

Related

Not tested

  • macOS refresh rate with this monitor: confirmed at 5120x2160, but I didn't check which rate macOS picks or over which connection.
  • Overriding dpxbar-max-*-per-pipe on its own. A single pipe probably can't run above its rated limit anyway.
  • An EDID emulator that adds a 5K@100–115 timing, which would fit inside one pipe.
  • Apcall 21 reply patching (the module was built but I have no result).
  • Any other M3 Pro, M3 Max or M4 board.

Investigation and write-up done with Claude Code (LLM-assisted). The firmware addresses and log lines come from this machine.

Activity

  1. JeppeBylov commented on Nov 11, 2026

    @JeppeBylov
    Author

    Update: macOS 14 drives 5120x2160@180 on this Mac by splitting it across dispext0 + dispext1

    I installed macOS Sonoma 14.8.9 (23J631) on a second volume of the same J514S and connected the same Samsung Odyssey G75F over DisplayPort. macOS runs it at 5120x2160@180 (system_profiler: "UI Looks like: 5120 x 2160 @ 180.00Hz"). That's 1.99 Gpix/s, well above the 1,274,019,840 px/s per-pipe limit that makes the firmware drop these modes on Linux.

    ioreg -l dumps at 5K@180, 5K@120 and "3840x1620" (with the built-in panel on) show how macOS does it.

    There are two external framebuffers, each half the width:

    dispext0@C0000000 dispext1@C4000000
    DisplayWidth x DisplayHeight 2560 x 2160 2560 x 2160
    DPTimingModeId (180 Hz / 120 Hz) 10078 / 10079 10078 / 10079
    DisplayClock (180 / 120 Hz) 730843776 / 490912512 same
    VideoClock (180 / 120 Hz) 562187520 / 377625024 same
    clockRatio (180 / 120 Hz) 108995 / 162267 same
    PixelClock 935000000 same
    TimingElements 54 entries none
    dp-switch-ufp-port / -endpoint 0 / 0 0 / 1
    • Each pipe drives 2560x2160. At 180 Hz that's about 0.995 Gpix/s per pipe, under the limit.
    • dispext0 keeps the 5K timings. DPTimingModeId = 10000 + the TimingElements ID, and dispext0's catalog keeps ID 78 (5120x2160@180, total 5504x2270) and ID 79 (5120x2160@120, total 5504x2287). Both have IsSplit=No and IsVirtual=No. These are the same DisplayID timings the firmware removes on Linux ("IOMFB removing mode ... max 1274019840"). So the IsSplit flag describes the catalog entry, and the split happens below it.
    • dispext1 is a peer only. It has no catalog and no video/audio interface epics. It does have its own AppleDCPDPTXRemotePortUFP(iop-dcpext1-nub) and two DCPAVControllerProxy. Both pipes sit on the same dp-switch-ufp-port (0), with endpoints 0 and 1. This matches the firmware's displayAllocation: extraPipes / mainUFP / peerUFP strings.
    • No AppleDCPDPTXVirtualPort anywhere. That explains why injecting appledptx-peer-pipe-cfg (described in the issue above) had no effect: macOS doesn't use the VirtualPort path for this.
    • The 3840x1620 "low resolution" choice is still 5K@120 underneath. The clocks match 5K@120 and the image is scaled, so this monitor has no below-limit mode in macOS to compare with.

    What this narrows down

    The per-pipe limit isn't raised anywhere. Instead, the dispext0 firmware has to count dispext1 as a peer pipe on UFP endpoint 1 before it filters modes, and both halves then have to be configured and swapped together. Leads for what macOS does that Linux doesn't:

    1. Bringing up dispext1's DPTX remote port on the same UFP port (endpoint 1) while the display is attached to dispext0.
    2. The display-allocation reply. Linux echoes DPTX apcall 21 (SetTiledDisplayHints) back unchanged. macOS may answer with extraPipes=1 and the peer UFP.
    3. Two framebuffers with one timing, with swaps coordinated across both DCPs.

    A trace of macOS 14 under the hypervisor would show the exact sequence. macOS 14 doesn't need the SPTM path; my macOS 27 attempt stopped there (no RTBuddySeg in the host ADT). I can't do that trace myself, but I'm happy to test patches on J514S with this monitor, and to share the full ioreg dumps (about 7 MB each).

    Analysis done with Claude Code (LLM-assisted). The numbers above are copied from the ioreg dumps on this machine.

  2. JeppeBylov commented on Nov 11, 2026

    @JeppeBylov
    Author

    Update: trace of macOS 14 setting up the split (J514S, macOS 14.8.9, Samsung G75F over DP)

    I traced macOS 14.8.9 under the m1n1 hypervisor, with both dcpext0 and dcpext1 traced in the same run, while switching the Samsung between 5120x2160@120 and @180. This shows what macOS does that Linux doesn't.

    1. The pipe allocation is in the reply to DPTX apcall 21 (SET_TILED_DISPLAY_HINTS)

    Offsets below are into the payload, after the 64-byte EPIC call header.

    dcpext0's firmware sends a 0x90-byte request:

    • +0x00 = 1
    • +0x10 = ff ff ff
    • +0x38 = 5120, +0x3c = 2160 (the same width/height fields dptxep.c already reads for tiled displays; the tile counts are 0 here)
    • +0x40 = b0 33 09 86 92 82 a4 76, +0x48 = 0x000a0000
    • everything else 0

    macOS echoes the request and fills in one block:

    offset value probable meaning
    +0x64 1 extra pipes
    +0x68 1 peer pipe / peer UFP
    +0x6c 0x1800 (6144) max pixel width per pipe
    +0x70 0x55b62380 (1438000000) max pixel rate per pipe
    +0x74 0x4bf00000 (1274019840) max active pixel rate per pipe: the max in "IOMFB removing mode"

    Linux (dptxport_call_set_tiled_display_hint) copies the request back and sets +0x00 = 1, so this block goes back as zeros. That fits the firmware then assuming a single pipe and dropping 5K@120/180.

    A second request (with +0x00 = 0 and no size) arrives on HPD removal. macOS echoes it unchanged.

    2. Mode set: split IDs = ID + 10000, sent to both DCPs

    1. macOS first tries set_digital_out_mode(94, 78) on dcpext0 alone. The firmware logs returned 80000104.
    2. It then sends set_digital_out_mode(10094, 10078) (5K@180) to both dcpext0 and dcpext1, and (10094, 10079) for 5K@120.
    3. set_display_device(2) goes to both DCPs as well.

    3. The firmware tells each DCP where its slice starts

    The D115 UnifiedPipeline2::set_tiling_state(event, para, val) callbacks include para=8: val=0 on dcpext0 and val=2560 on dcpext1, i.e. the slice's horizontal origin. The other parameters (1, 2, 4, 5, 6, 9, 10, 11) aren't decoded yet. event=2, para=10, val=1 repeats 238 times on dcpext0.

    4. dcpext1 is a pure second pipe

    • Its dcpdptx-port-epic services are announced but get no DPTX apcalls. All link training (link rate, lane count, link config, activate, tiled hints) happens on dcpext0.
    • Swaps go to both DCPs every frame, alternating dcpext0 → dcpext1 (swap_start 846 vs 845 in the capture).

    What I'll try next on Linux

    1. Answer apcall 21 the way macOS does (fill +0x64..+0x77 when the hint carries a size) and check whether timings 10078/10079 show up in dcpext0's catalog, without modesetting. This is a module parameter, off by default.
    2. If they do: mode set with the +10000 IDs on both DCPs, plus a CRTC that feeds two 2560-wide halves and swaps both. This is the real driver work, and probably needs someone who knows the CRTC/dcpext code.

    I can share the raw trace (about 26 MB of console log, plus excerpts) if that helps.

    Trace analysis done with Claude Code (LLM-assisted). The values are copied from the captured messages. The trace was made on my own hardware; nothing was sent to the m1n1/Asahi projects.

  3. JeppeBylov commented on Nov 11, 2026

    @JeppeBylov
    Author

    Update: answering apcall 21 like macOS makes the firmware offer 5K@120/180 on Linux

    I tested the allocation reply from my previous comment on Linux. It works for the mode list.

    Setup: J514S, linux-aurora 7.2.11.aurora15-1 (b798f2b). appledrm was rebuilt with the patch below, which adds a runtime switch appledrm.split_alloc (off by default). The Samsung G75F is on DisplayPort through an Anker TB5 dock, with the compositor pinned to a 4K mode so nothing switched to the new modes.

    With split_alloc=Y and a replug, the kernel log shows:

    apple-dcp 2d6c00000.dcp: DPTXPort: EXPERIMENT split_alloc: offered a peer pipe for 5120x2160 (data 144, reply 144)
    apple-dcp 2d6c00000.dcp: RTKit: syslog message: nifiedPipeline.cpp:9945: VFTGMerge conifg. Allowing width 5120
    apple-dcp 2d6c00000.dcp: RTKit: syslog message: nifiedPipeline.cpp:9945: VFTGMerge conifg. Allowing width 5120
    apple-dcp 2d6c00000.dcp: RTKit: syslog message: nifiedPipeline.cpp:9945: VFTGMerge conifg. Allowing width 3840
    apple-dcp 2d6c00000.dcp: DPTXPort: EXPERIMENT split_alloc: hint not used (data 144, reply 144)
    apple-dcp 2d6c00000.dcp: DPTXPort: EXPERIMENT split_alloc: offered a peer pipe for 5120x2160 (data 144, reply 144)
    apple-dcp 2d6c00000.dcp: RTKit: syslog message: nifiedPipeline.cpp:9945: VFTGMerge conifg. Allowing width 5120
    apple-dcp 2d6c00000.dcp: RTKit: syslog message: nifiedPipeline.cpp:9945: VFTGMerge conifg. Allowing width 5120
    apple-dcp 2d6c00000.dcp: RTKit: syslog message: nifiedPipeline.cpp:9945: VFTGMerge conifg. Allowing width 3840
    
    • The firmware takes the VFTGMerge dual-pipe exception in videointerface_filter_modes (the one at 0x19cef4 in the issue description).
    • There are no more "IOMFB removing mode: 5120 x 2160 ..." lines.
    • The compositor's mode list for the monitor now includes 5120x2160@179.99 and 5120x2160@120.
    • Nothing crashed. The monitor kept running 3840x2160@120 single-pipe. The second hint ("hint not used") is the HPD-removal variant without a size, which is echoed as before.

    I have not modeset to the new modes. From the macOS trace, that needs set_digital_out_mode(10000 + colour, 10000 + timing) and set_display_device(2) on both dcpext0 and dcpext1, plus one CRTC feeding two 2560-wide halves with swaps on both DCPs. I'll start on the first part locally (also behind a switch). Pointers from anyone who knows the dcpext/CRTC code would be welcome.

    Patch (local experiment, off by default)
    --- a/drivers/gpu/drm/apple/dptxep.c
    +++ b/drivers/gpu/drm/apple/dptxep.c
    @@ -100,6 +100,21 @@
     #define DPTX_TILE_HINT_MIN_SIZE	0x40
     
     /*
    + * LOCAL EXPERIMENT (not upstream): macOS 14 on J514S answers this call
    + * with a pipe allocation at +0x64..+0x77 (seen in a hypervisor trace:
    + * 1, 1, 6144, 1438000000, 1274019840). The DCP then offers 5K@120/180 as
    + * split timings (ID + 10000). Off by default; can be flipped at runtime and
    + * takes effect on the next hotplug.
    + */
    +static bool split_alloc;
    +module_param(split_alloc, bool, 0644);
    +MODULE_PARM_DESC(split_alloc,
    +		 "EXPERIMENT: offer a dcpext peer pipe in SetTiledDisplayHints replies like macOS 14");
    +
    +#define DPTX_SPLIT_ALLOC_OFF	0x64
    +#define DPTX_SPLIT_ALLOC_END	0x78
    +
    +/*
      * A Thunderbolt DP tunnel uses the same CORE|ATC|DIE|CONNECTED target as a
      * direct alt-mode PHY: CORE is the DFP port (0 = dpphy, 1/2 = dpin0/dpin1)
      * and ATC the route's own ATC index.
    @@ -661,6 +676,30 @@
     	}
     
     	reply->retcode = cpu_to_le32(1);
    +
    +	/* Only for a single-stream sink that reports its size (not a 2-tile LG). */
    +	if (READ_ONCE(split_alloc) && data_size >= DPTX_SPLIT_ALLOC_END &&
    +	    reply_size >= DPTX_SPLIT_ALLOC_END &&
    +	    get_unaligned_le32(hint + DPTX_TILE_HINT_TILES_H) <= 1 &&
    +	    get_unaligned_le32(hint + DPTX_TILE_HINT_WIDTH) &&
    +	    get_unaligned_le32(hint + DPTX_TILE_HINT_HEIGHT)) {
    +		u8 *r = reply_;
    +
    +		put_unaligned_le32(1, r + DPTX_SPLIT_ALLOC_OFF + 0x00);
    +		put_unaligned_le32(1, r + DPTX_SPLIT_ALLOC_OFF + 0x04);
    +		put_unaligned_le32(0x1800, r + DPTX_SPLIT_ALLOC_OFF + 0x08);
    +		put_unaligned_le32(0x55b62380, r + DPTX_SPLIT_ALLOC_OFF + 0x0c);
    +		put_unaligned_le32(0x4bf00000, r + DPTX_SPLIT_ALLOC_OFF + 0x10);
    +		dev_info(service->ep->dcp->dev,
    +			 "DPTXPort: EXPERIMENT split_alloc: offered a peer pipe for %ux%u (data %zu, reply %zu)\n",
    +			 get_unaligned_le32(hint + DPTX_TILE_HINT_WIDTH),
    +			 get_unaligned_le32(hint + DPTX_TILE_HINT_HEIGHT),
    +			 data_size, reply_size);
    +	} else if (READ_ONCE(split_alloc)) {
    +		dev_info(service->ep->dcp->dev,
    +			 "DPTXPort: EXPERIMENT split_alloc: hint not used (data %zu, reply %zu)\n",
    +			 data_size, reply_size);
    +	}
     	return 0;
     }
     

    Patch and analysis done with Claude Code (LLM-assisted), tested on my J514S.

  4. JeppeBylov commented on Nov 11, 2026

    @JeppeBylov
    Author

    The host side of the split mode set, from the macOS 14 trace

    Here's what macOS sends when it switches the G75F to 5120x2160@180. In macOS naming, MAIN = dcpext0 (owns the DPTX link) and HELPER = dcpext1. On my Linux boot the roles are reversed: the monitor is on dcp@2d6c00000 (alias dcpext1), and the free pipe is dcp@2d2c00000. The free pipe's IOMFB is never opened on Linux (no A401/A455).

    macOS runs this round twice before the first swap:

    Order (one round; macOS repeats the whole round twice before the first swap):
     1. helper A363(1)                       ? "you are a slave/peer pipe"
     2. main/helper A472 setPowerState(0)
     3. main/helper A359(0)                  removes DPTimingModeId/DisplayWidth/DisplayHeight props
     4. helper A363(1)
     5. main A361 getVideoDataForIDs -> full timing (htotal 5504, pclk 2248750000, vtotal 2270 ...)
     6. main AND helper A362 (0x7c in)        HALF timing computed by host: htotal 2752, hsync/porches /2,
                                              active 2560, pclk 1124375000; vertical unchanged
     7. main/helper A472 setPowerState(1)
     8. main A357(1?), helper A357(0/2?)      (parse unclear; values 0,1,2 seen)
     9. helper A358(1)
    10. main A411 set_digital_out_mode(10094, 10078)   (colour/timing ID + 10000)
    11. main A366, A364 -> helper A365 (0x20 in: DSC PPS-like "12 00 0e 4e 00 00 0d 40 ... 00 fd 43 14 ...")
    12. main A367 -> helper A368 (4 in: 0x80 = DSC bpp?)       fw: "calling getDSCBitsPerPixel"
    13. main A369 -> helper A370 (8/12 in: 0x40000080)        fw: "calling getDSCslicesize"/"setDSCslicesize"
    14. helper A411 set_digital_out_mode(10094, 10078)
    15. main A437 get_block_dcp; main/helper A409 set_display_device(2)
    ... second round (2-15 again), then A406 swap_start on main then helper, every frame.
    

    The A362 half timing is computed by the host from the A361 reply. For 5K@180:

    • A361 (MAIN): htotal 0x1580 (5504), pixel clock 0x860933b0 (2248750000)
    • A362 (both): htotal 0xac0 (2752), hsync/porches halved, pixel clock 0x430499d8 (1124375000); vertical unchanged

    The DSC values are copied from MAIN to HELPER (A364/A366 → A365, A367 → A368, A369 → A370).

    Then every frame swaps on MAIN, then on HELPER (A406 on both). The firmware tells each pipe its slice origin through D115 set_tiling_state(para=8): 0 on MAIN, 2560 on HELPER.

    Still to decode: the replies of A364/A366/A367/A369, the meaning of A357/A358/A363, the A362 layout, and each half's swap rectangles. I'll post those when I have them.

    From a hypervisor trace on my own hardware; analysis done with Claude Code (LLM-assisted).

  5. JeppeBylov commented on Nov 11, 2026

    @JeppeBylov
    Author

    It works: 5120x2160@180 on Linux, over HDMI and over a Thunderbolt dock

    PR: #445 (three commits, all behind module parameters that are off by default).

    On my J514S with the Samsung Odyssey G75F, Linux now runs 5120x2160@120 and @179.985:

    • through the Mac's HDMI port
    • through an Anker TB5 dock over DisplayPort
    • from boot, with all three switches in /etc/modprobe.d

    What was missing, in order:

    1. The pipe allocation in the apcall 21 reply (see my earlier comments). With it, the firmware keeps the 5K timings.
    2. The host side of the two-pipe mode set. MAIN's A411 waits for HELPER's. The host relays the DSC setup, answers D115/D116 from a shared tiling table, sends paired swaps with the same IOVA and an offset source rect for HELPER, and powers both pipes down together. Doing that one after the other kills MAIN.
    3. The role direction. The merge only works with dcpext0 as MAIN, as on macOS. With dcpext1 as MAIN, every firmware call succeeds but the screen stays black and the sink drops HPD after ~30 s. The display crossbar is not involved: an MMIO trace of macOS shows it routes only MAIN in split mode (dpin0 → state 0, MUX_SELECT 0), the same as the Linux single-pipe state. On a J514S, HDMI uses dcpext0, but a Thunderbolt display lands on dcpext1, so the PR adds an opt-in placement override (split_dock_main).

    Tested:

    • 120 ↔ 180 Hz while running
    • DPMS off/on, several cycles
    • the dock's link retrain
    • helper pipe started on demand
    • 200+ fps, ~5 ms frame time, no seam

    Not tested:

    • suspend: refused while an external link is up, as before
    • other boards
    • audio over the dock

    Details are in the PR description.

    LLM-assisted (Claude Code), tested on my own hardware.

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