Skip to content

J516S/G15: OrcaSlicer-associated rendering exhausts PB (4095), GPU fail-stop on Wayland and XWayland #495

Open

Description

@ronaldex

Mac and installed build

  • Board / SoC: apple,j516s / apple,t6030 (M3 Pro / G15).
  • OS: Arch Linux ARM rolling with Omarchy 4.5.0-3, omarchy-mac 0.2.3-2.
  • Running custom kernel: 7.2.11-g15-bed544-358-389, also confirmed in the four incident boot journals. Installed stock linux-aurora/headers package: 7.2.11.aurora14-1 (not the running kernel).
  • Custom source: bed544bdf93a496ebf70ba180923bab12b6e01cf + drm/apple: time-stamp T6030 swaps on 14.7 for 120 Hz #358 7359621dca7ff8b62f064f9b516a7284c8b3a59d + drm/asahi: g15: budget compute flights while another VM waits #389 runtime 6aeb9169fcd2f71c5ea77a8c3fd47ed063d14e10 and host test bed2d45390f9cf3e994570cd563a778bb993da81.
  • This build retains display swap timing/120 Hz, G15 compute sharing, and J516S/fw14.7 admission pins. drm/apple: time-stamp T6030 swaps on 14.7 for 120 Hz #358 is display timing; drm/asahi: g15: budget compute flights while another VM waits #389 budgets compute, not PB growth, and is not a PB fix.
  • Mesa/Vulkan: 1:26.3.0_devel.20261031.231483.a1969111475-1.
  • m1n1 1.6.1.aurora22-1; U-Boot 2026.07.aurora6-1; omarchy-aurora-firmware 20261014-7.
  • These package/file values were collected after reboot; they are not a complete historical crash-time package inventory. No renderer probe was run for this report.

Both firmware versions and provisioning

Read from chosen firmware fields: asahi,os-fw-version: 14.7; asahi,system-fw-version: unknown (literal). Provisioning/calibration was not investigated; no per-unit firmware, calibration, serials or raw device tree attached.

Reproduction, expected result, and actual result

Not a minimal or deterministic reproducer. These are retrospective incident observations during OrcaSlicer use, not a controlled pass/fail series.

On 16 November 2026 (CET), kernel PB growth reached 4095 blocks, subsequent growth was left unanswered after two refusals, and the G15 host deadline led to ETIMEDOUT fail-stop. Seen at 10:00:23, 19:18:09, 19:21:16 and 20:56:49 in separate boots of the same custom release. Native OrcaSlicer cores exist at 10:00:23 (PID194157) and 19:21:16 (PID8381). The reporter describes crashes both during slicing and while rotating the model after slicing, using a fairly complex 3D-print model. The exact action for each individual timestamp, total number of successful trials, and minimal triggering scene remain unknown.

A new XWayland launcher was created at 20:55:49. It selects GDK_BACKEND=x11 and WEBKIT_DISABLE_DMABUF_RENDERER=1, with other settings matching the current GPU launcher. At 20:56:42 ctx20 hit the same PB limit; at 20:56:49 the GPU fail-stopped and Xwayland/Hyprland aborted, followed by shell/session fallout. A reboot followed. The reporter's description covers slicing and post-slice model rotation; it does not identify the exact action at this timestamp. This trial did not establish a safe XWayland workaround.

Expected: ordinary slicer rendering should complete without GPU-wide failure; a bad job should ideally be contained to its client. Actual: GPU marked unusable until reboot, taking down the graphical session. No claim about saved/unsaved document loss has been verified.

Proposed future procedure, NOT executed: narrow the reported slicing/post-slice rotation workflow to exact steps and a disposable non-private minimal scene; save work and agree a recovery path and one supervised attempt; record backend, kernel, launcher settings and action sequence; stop immediately on PB refusal/hang. Do not automatically loop workloads or change driver limits to force a reproduction. A sanitized local evidence kit is retained.

Graphics or session details (if relevant)

OrcaSlicer local extracted installation 2.5.1 (path identity; no app started to query version). Hyprland 0.56.2-7, Xwayland 24.1.14-1. Active GL renderer string, complete launch-time environment and GPU opt-in/module parameters were not captured.

Native OrcaSlicer PID8381: SIGSEGV at 19:21:16, top libgallium-26.3.0-devel.so + 0x96c518; stack continues through dri_flush, EGL, wxGLCanvasEGL::SwapBuffers, GLCanvas3D::_render_frame, idle handling. Morning PID194157 has the same top offset but reaches _render_scene / _render_frame / mouse handling through a different stack path.

Xwayland PID1897: SIGABRT at 20:56:49, journal reports original signal 11 before abort; saved stack includes libgallium +0x96c514. Hyprland PID1632 also SIGABRT, libgallium +0x96c514. These are not OrcaSlicer cores. Offsets are adjacent, not proven to name the same function.

The current native launcher was modified after the native crashes; their WebKit DMA-BUF setting is unknown. The XWayland trial's launcher disables WebKit DMA-BUF. Neither this uncontrolled comparison nor the shared failure establishes DMA-BUF causality or rules out every presentation-path bug.

Relevant error lines

Selected journal lines from the explicit XWayland-trial boot, 2026-11-16 CET; device prefix and address-valued fields omitted, not a continuous log:

20:56:42 PB growth request 403 ctx 20 position 0: +1 blocks (4095 of 4095), DEVCTL 7: 0
20:56:42 PB growth request 404 ctx 20 position 0: +0 blocks (4095 of 4095), DEVCTL 7: 0
20:56:42 PB growth request 406 ctx 20 left unanswered: 2 refusals at the maximum (4095 blocks); the host deadline applies
20:56:47 G15: hung, requesting preemption: ctx=20 ... errors=0 busy=1 fault=0x0 5000ms
20:56:49 G15: prefix FAILED: ctx=20 ... errors=0 busy=1 fault=0x0 7000ms
20:56:49 G15: fail-stop (ETIMEDOUT); GPU unusable until reboot

Earlier incident excerpts and selected module-relative backtraces are retained locally and can be supplied as sanitized text. No OOM-kill messages matched the bounded kernel windows; no claim that all memory pressure is excluded.

Observed: PB cap/refusals precede host timeout and GPU-wide lost-device fallout; native OrcaSlicer crashes coincide with two earlier fail-stops.

Inference: OrcaSlicer-associated render activity may exercise a shared G15 PB exhaustion/recovery problem, with Mesa faults as fallout. Neither the root defect nor ctx20's owner is proven. A Mesa/application submission issue could contribute; correlation is not context ownership or proof of where the fix belongs.

Related issues, pull requests, or code examples

  • #440: related GPU-wide G15 fail-stop/recovery outcome, but J516C/T6031 compute-prefill deadline, not this T6030 PB-growth sequence. No established duplicate root cause.
  • #182: on-demand PB, cap, idle reclaim/client-boundary reinit; already merged into bed544, not a missing fix. Its T6030 behavior was untested in that report. It ports aurora-silicon/linux#363.
  • Read-only issue searches for "Orca" (the original search term), "parameter buffer" and "4095" found no matches; not exhaustive deduplication. Kernel integration provenance and kernel-visible PB sequence motivate this initial target, not a proven kernel-only cause.

What was not tested

  • Workload reported: a fairly complex 3D-print model, with failures during slicing or post-slice model rotation. No known minimal model, shortest proven steps, per-incident action mapping, frequency denominator, or known-good/first-bad build.
  • No fresh workload reproduction in preparing this kit; no controlled native-versus-XWayland or DMA-BUF A/B.
  • No software-renderer comparison, stock/newer matched-stack comparison, module-parameter experiment, or driver patch applied.
  • No proven OrcaSlicer ownership of ctx20, full crash-time environment, crash-time RAM measurement, or fully symbolized Mesa functions.
  • No public model/file attached. Model names, raw cores, full journals, home/host identifiers, boot IDs, network data and private environment are intentionally excluded.

Activity

  1. changed the title [-]J516S/G15: Orca-associated rendering exhausts PB (4095), GPU fail-stop on Wayland and XWayland[/-] [+]J516S/G15: OrcaSlicer-associated rendering exhausts PB (4095), GPU fail-stop on Wayland and XWayland[/+] on Nov 16, 2026
  2. ronaldex commented on Nov 16, 2026

    @ronaldex
    Author

    Update: this tentative workaround was followed by another PB-cap / GPU fail-stop crash at 21:19 CET. GALLIUM_THREAD=0 is NOT a reliable workaround; see the later correction in this issue.

    Follow-up: successful same-model trial with GALLIUM_THREAD=0

    The reporter tested a separate launcher using native Wayland (GDK_BACKEND=wayland) and GPU rendering, with GALLIUM_THREAD=0 added. WEBKIT_DISABLE_DMABUF_RENDERER=1 was retained; software rendering was not forced. No kernel/Mesa change was made for this trial.

    Using the same fairly complex model involved in the earlier failures, the reporter varied the workload, sliced it multiple times, and rotated the model, without experiencing a crash. Exact duration and trial count were not recorded, and no kernel-log/PB monitoring or renderer probe was captured for this successful trial.

    This is a promising user-reported workaround, not a demonstrated root cause or long-term stability guarantee. It does not establish which internal driver path changed or prove that PB exhaustion cannot recur. The ordinary OrcaSlicer launcher has now been configured with the same GALLIUM_THREAD=0 setting for continued use.

  3. ronaldex commented on Nov 16, 2026

    @ronaldex
    Author

    Correction: GALLIUM_THREAD=0 did not prevent recurrence

    The initially successful same-model trials were followed by another GPU-wide failure on 16 November 2026 at 21:19 CET. Please do not treat GALLIUM_THREAD=0 as a reliable workaround.

    The ordinary launcher had been configured with GALLIUM_THREAD=0, GDK_BACKEND=wayland and WEBKIT_DISABLE_DMABUF_RENDERER=1, with GPU rendering enabled. The reporter subsequently experienced another crash using this configuration. The launcher still contains those settings; the crashed process's actual environment was not independently captured.

    Selected sanitized kernel lines from that incident:

    21:19:41 PB growth request 403 ctx 8 position 0: +1 blocks (4095 of 4095), DEVCTL 7: 0
    21:19:41 PB growth request 404 ctx 8 position 0: +0 blocks (4095 of 4095), DEVCTL 7: 0
    21:19:41 PB growth request 406 ctx 8 left unanswered: 2 refusals at the maximum (4095 blocks); the host deadline applies
    21:19:46 G15: hung, requesting preemption: ctx=8 ... errors=0 busy=1 fault=0x0 5000ms
    21:19:48 G15: prefix FAILED: ctx=8 ... errors=0 busy=1 fault=0x0 7000ms
    21:19:48 G15: fail-stop (ETIMEDOUT); GPU unusable until reboot
    

    The OrcaSlicer core records SIGSEGV at 21:19:48, again at libgallium-26.3.0-devel.so + 0x96c518, followed by dri_flush, EGL, wxGLCanvasEGL::SwapBuffers and GLCanvas3D::_render_frame. Hyprland also aborted; a reboot followed.

    This is the same observed PB-cap/refusal/timeout/fail-stop pattern as before, not evidence of a separate failure mode. The earlier successful trials remain observations, but do not establish a durable workaround, a lower failure rate, or a threading-related root cause. Exact triggering action for this new incident and ownership of ctx8 remain unverified. No additional workload reproduction or launcher change was performed when recording this correction.

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