Repository navigation
Conversation
82219c7 to
35a6e91
Compare
5d0f77c to
b19f247
Compare
|
Thanks for getting tunneled PCIe working on the M4 Pro. Moving T6040 onto the Intr2AXI match data, with the comment explaining why, is the approach we prefer. Since this is stacked on #2068, the diff also carries that PR's per-SoC pixel-clock selection ( We can run Thunderbolt and PCIe no-regression rows on our M1 Pro, M2, M4 Pro and A18 Pro once #2068 settles. |
|
dcp_fabric_t6040_tunnel_clock() in dcp-fabric.c decides whether apple_t6040_dp_tunnel_rate() drives the tunnel and whether dcp_tunnel_prepare() skips apple_atc_dp_tunnel_open(), and it decides this with of_machine_is_compatible("apple,t6040"/"apple,t6041"). Which clock feeds DP IN 0/1 is a property of the ATC PHY behind the route. It should come from that PHY's compatible or match data, or from a property on the DCP or route node, not from the machine compatible. Until then this should stay a draft. |
In CIO mode the T6040 ATC PHY's AUSPLL feeds the pixel clocks of the router's DP IN adapters: one fixed AUSPLL frequency, a divider select per link rate in TX_DP_CTRL0 (RBR 4, HBR 3, HBR2 1, HBR3 0), the clock switch gate in the cfg bank, and the stop. DP IN 0 uses PCLK1 and switch 1; DP IN 1 uses PCLK2 and switch 2 and needs PCLK1 running as well. appledrm drives it through phy_configure() on the PHY's DP phy, with three new DP options: set_tunnel_rate, tunnel_dpin and tunnel_rate (the DP link rate code, 0 to stop). Which path a tunnel takes follows the PHY behind the route, not the machine: an "apple,t6040-atcphy" PHY clocks the tunnel itself (on T6040 and T6041 alike) and needs no apple_atc_dp_tunnel_open(). The PHY stops the tunnel clocks when it leaves CIO. [petr: taken from fbalazsy/linux j616s/split-display (the tunnel clock, 432f1ec "give the DP tunnel pixel clock a symbol of its own" and 346a417 "give DP IN 1 of a tunnel its PCLK1 prerequisite") onto integration/aurora. Tested on the J614s with a Studio Display (2022) on Thunderbolt up to 5120x2880@60, aurora-silicon#1589.] [petr: per review, the clock goes through phy_configure() instead of an exported apple_t6040_dp_tunnel_rate(), and dcp-fabric selects it by the PHY's compatible instead of of_machine_is_compatible("apple,t6040"/"t6041").] Signed-off-by: František Balázsy <frantisekbalazsy@gmail.com> Signed-off-by: Petr Škuta <20834920+petrskuta@users.noreply.github.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On T6040/T6041 the crossbar's 0x34 is a per-clock UHBR select (bit n set when the link rate is UHBR) and the ATC enable of a DP IN output is at 0x38. The T6030 link wrote the ATC enable to 0x34, which set the UHBR select of an HBR2 stream and left the ATC disabled: the tunnel trained but no pixel data reached the DP IN. DP IN tunnels carry DP 1.x rates only, so clear the UHBR select and set the enable at 0x38; the DP alt mode output keeps the T6030 write. Resume checks the same register. [petr: split out of fbalazsy/linux 5e66575 "drm/apple: T6041 display coprocessors across hibernation and S2R" (with its comment from ed490dd). On the J614s this took the Studio Display's video path from ~480 to ~976,000 packets/s at 1080p60, aurora-silicon#1589.] Signed-off-by: František Balázsy <frantisekbalazsy@gmail.com> Signed-off-by: Petr Škuta <20834920+petrskuta@users.noreply.github.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Give the T6040 ACIO the M4 (T8132) DP IN policy and qualify T6040 and T6041 for DP tunnels. Their DP IN adapters take the t8103 handshake and the M3 flow, as on T8132; the pixel clock is the ATC PHY's AUSPLL, set through phy_configure(). Tested on the J614s (M4 Pro) with a Studio Display (2022) on the left Thunderbolt port: 1920x1080, 2560x1440 and 5120x2880 at 60 Hz, extended desktop, aurora-silicon#1589. Signed-off-by: Petr Škuta <20834920+petrskuta@users.noreply.github.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
dcp_tunnel_crossbar_up() routes the mux at the first
DidChangeLinkConfiguration and enables the connection only for
dcp_t602x_tunnel_route(), which never matches the T6041 crossbar on 26.6.
Its clocks then stay off until a retrain runs the up path again. 1080p
and 1440p got video only because the firmware retrained after the mode
set; 5120x2880 keeps the trained link ("Forcing link status to true for
synched-pipe config") and stayed dark with the crossbar's FIFO and pixel
clocks off.
Also enable the connection at the first link up for the T6041 crossbar.
Tested on the J614s with a Studio Display (2022): 5120x2880@60,
aurora-silicon#1589.
Signed-off-by: Petr Škuta <20834920+petrskuta@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The tunnel's AUSPLL start sets CLKOUT_MASTER 0x808 and refuses to start
while either bit is set ('already in use'), but the stop cleared only
0x8. After the first stop, e.g. a DPMS off of a Studio Display, the
tunnel clock could not start again:
DP tunnel: AUSPLL already in use (master 0x81d54, cmd 0x10000018)
DP tunnel pixel clock (rate 0x14) failed: -16
and the display stayed dark after DPMS on. Clear both outputs.
Signed-off-by: Petr Škuta <20834920+petrskuta@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The host test extracts dcp_route_clock_rate() from the display fabric and builds it on its own. The fabric now sets the tunnel clock through phy_configure() when the route's PHY is a T6040 ATC PHY, so the test needs those helpers and a model of the PHY operation. Extract dcp_phy_clocks_tunnel() and dcp_phy_tunnel_rate() with the other fabric functions, give the model PHY an OF node and stub phy_configure() with the DP tunnel options. Add a case that runs a T6040 ATC PHY on the T6040, T6041 and T6030 machines through DP IN 0 and 1: the rate reaches the PHY with set_tunnel_rate alone and the right adapter, neither the ATC nor the T6050 provider is called, and a clock that fails to start leaves the route's rate alone. [petr: reworked for phy_configure() and the PHY compatible.] Signed-off-by: František Balázsy <frantisekbalazsy@gmail.com> Signed-off-by: Petr Škuta <20834920+petrskuta@users.noreply.github.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
t6040_pciec_hw was modelled on T8132: the kernel_init (oe-fabric) cold
init, the T8132 port reset values and SID-indexed RID2SID slots. The
ADT's apciec,t6040 has no oe-fabric region (config, rc, port0, fabric,
debug, ioa, ltssm, intr2axi), so cold init failed mapping it (-ENODATA).
Its fabric, rc, debug, config and DART tunables are T6050's with small
value differences.
Make it an Intr2AXI controller with T6050's no_port_reset, and allocate
RID2SID slots as T6050 does: with iommu-map-mask 0xff the Studio
Display's xHCI (RID 0x300) and its Thunderbolt bridge (RID 0x100) share
SID 0, and an SID-indexed slot failed the xHCI's enable with a silent
-EBUSY ("Can't enable PCI device, BIOS handoff failed").
Tested on the J614s with a Studio Display (2022) on the left port: cold
init completes (link up), the display's Titan Ridge bridges and xHCI
enumerate and xhci_hcd binds, aurora-silicon#1589.
Signed-off-by: Petr Škuta <20834920+petrskuta@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Enable the left port's tunneled PCIe (usb4_0_pcie_tunnel, apciec0_dart, usb4_0_pcie) and keep the ACIO's apple,pcie-tunnel, so a Thunderbolt device's PCIe functions reach a root port. No boot loader hands the port over, so the kernel cold-initializes it (apple,pciec-kernel-init). From the j614s ADT (macOS 26 IODeviceTree) apciec0 and dart-apciec0: - the controller's fabric, rc and debug tunables and the port's config tunables; - the DART's tunables and its DMA window: vm-base 1 TiB + 2 GiB, vm-size 0x7ff00000, the first 16 KiB a dead mapping; - bypass streams 253 and 255 (bypass-253, bypass-255): the port's own writes, its MSIs turned into AIC writes, use them. Without the bypass no MSI arrived (xHCI "Command Aborted") and the DART fault IRQ stormed. As on T6050's PCIe-C DART: full error reporting, CTC flush and the ATC0 PCIe power domain. Tested on the J614s with a Studio Display (2022): its USB2 and USB3 Gen2 hubs, speakers and microphones (snd-usb-audio), camera (uvcvideo), HID and a USB-C port (a USB 3.1 drive at 5 Gb/s) work, aurora-silicon#1589. Signed-off-by: Petr Škuta <20834920+petrskuta@users.noreply.github.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The ADT's apciec0-2 and dart-apciec0-2 carry identical tunables, DMA windows and bypass streams, so describe the PCIe-C of the left port once in t6040-j614s-pciec.dtsi and use it for all three ports. Port 0 is unchanged; ports 1 and 2 keep their ACIO's apple,pcie-tunnel now. Signed-off-by: Petr Škuta <20834920+petrskuta@users.noreply.github.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
b19f247 to
ead9c21
Compare
Depends on #2068. This branch is stacked on it, so only the last three commits are new here.
Tunneled PCIe now works on all three Thunderbolt ports of the J614s (T6040), the first M4 Pro/Max PCIe-C on hardware as far as I know. A Studio Display's xHCI enumerates behind it, and with it the display's USB2/USB3 Gen2 hubs, speakers, microphones, camera, HID interfaces and USB-C ports. Findings are in #1589.
Commits
PCI: apple: T6040 PCIe-C is an Intr2AXI controller.
t6040_pciec_hwwas modelled on T8132, but the ADT'sapciec,t6040has nooe-fabricregion, so kernel_init cold init failed (-ENODATA). Its tunables are T6050's with small differences, so it now takes the T6050 path:tunnel_intr2axi,no_port_reset, and RID2SID slots allocated rather than indexed by SID. Withiommu-map-mask = <0xff>the xHCI (RID 0x300) and the display's bridge (RID 0x100) share SID 0, and the indexed slot failed the xHCI's enable with a silent-EBUSY.arm64: dts: apple: t6040-j614s: PCIe-C on the left Thunderbolt port. Enables the tunnel, DART and controller with
apple,pciec-kernel-init, plus everything from the J614s ADT (macOS 26 IODeviceTree,apciec0/dart-apciec0):apple,bypass-sids = <253 255>(bypass-253,bypass-255), as on T6050's PCIe-C DART. The port's own writes, its MSIs turned into AIC writes, use those streams. Without the bypass no MSI arrived (xHCICommand Abortedon every device) and the DART fault IRQ stormed.Full error reporting, CTC flush and the ATC0 PCIe power domain are also as on T6050.
arm64: dts: apple: t6040-j614s: PCIe-C on all three Thunderbolt ports. The ADT's
apciec0-2anddart-apciec0-2carry identical tunables, DMA windows and bypass streams, so the description moves tot6040-j614s-pciec.dtsi(oneJ614S_PCIEC(n)template used for ports 0, 1 and 2). Port 0 is unchanged.Testing
apple,j614s apple,t6040, os-fw 26.6.2, system-fw 27.2. This series onintegration/auroracc8cdb428. Hyprland 0.57.2.status 0x3, link up), domain 0001 has the root port106b:1018, the display's Titan Ridge bridges8086:15efand xHCI8086:15f0, andxhci_hcdbinds.05ac:101d, USB3 Gen2 hub05ac:101e(10 Gb/s), Studio Display05ac:1114.snd-usb-audio, PipeWire playback): pass.uvcvideo, UVC 1.50, MJPEG 1920x1080, live preview): pass.🤖 Generated with Claude Code