Merge pull request 'mesa-panvk-bifrost-video: regenerate 0005 patch from POST-review snapshot' (#84) from claude-noether/marfrit-packages:noether/mesa-panvk-bifrost-video-retrigger into main

Reviewed-on: #84

arch/mesa-panvk-bifrost-video/0005-panvk-bifrost-video-KHR-video-decode-h264.patch

@@ -1,6 +1,6 @@
 diff -urN a/src/panfrost/vulkan/jm/panvk_cmd_buffer.h b/src/panfrost/vulkan/jm/panvk_cmd_buffer.h
 --- a/src/panfrost/vulkan/jm/panvk_cmd_buffer.h	2026-05-21 22:46:57.477785029 +0200
-+++ b/src/panfrost/vulkan/jm/panvk_cmd_buffer.h	2026-05-21 22:47:09.189957157 +0200
++++ b/src/panfrost/vulkan/jm/panvk_cmd_buffer.h	2026-05-22 10:17:41.214043265 +0200
 @@ -88,8 +88,18 @@
        struct panvk_cmd_compute_state compute;
        struct panvk_push_constant_state push_constants;
@@ -22,7 +22,7 @@ diff -urN a/src/panfrost/vulkan/jm/panvk_cmd_buffer.h b/src/panfrost/vulkan/jm/p
  
 diff -urN a/src/panfrost/vulkan/meson.build b/src/panfrost/vulkan/meson.build
 --- a/src/panfrost/vulkan/meson.build	2026-05-21 22:46:59.277811484 +0200
-+++ b/src/panfrost/vulkan/meson.build	2026-05-21 22:47:09.189957157 +0200
++++ b/src/panfrost/vulkan/meson.build	2026-05-22 10:17:41.214043265 +0200
 @@ -41,6 +41,10 @@
    'panvk_device_memory.c',
    'panvk_host_copy.c',
@@ -36,7 +36,7 @@ diff -urN a/src/panfrost/vulkan/meson.build b/src/panfrost/vulkan/meson.build
    'panvk_physical_device.c',
 diff -urN a/src/panfrost/vulkan/panvk_buffer.c b/src/panfrost/vulkan/panvk_buffer.c
 --- a/src/panfrost/vulkan/panvk_buffer.c	2026-05-21 22:46:57.485785147 +0200
-+++ b/src/panfrost/vulkan/panvk_buffer.c	2026-05-21 22:47:09.189957157 +0200
++++ b/src/panfrost/vulkan/panvk_buffer.c	2026-05-22 10:17:41.214043265 +0200
 @@ -88,6 +88,8 @@
           *bind_status->pResult = VK_SUCCESS;
  
@@ -48,7 +48,7 @@ diff -urN a/src/panfrost/vulkan/panvk_buffer.c b/src/panfrost/vulkan/panvk_buffe
  }
 diff -urN a/src/panfrost/vulkan/panvk_buffer.h b/src/panfrost/vulkan/panvk_buffer.h
 --- a/src/panfrost/vulkan/panvk_buffer.h	2026-05-21 22:46:57.485785147 +0200
-+++ b/src/panfrost/vulkan/panvk_buffer.h	2026-05-21 22:47:09.189957157 +0200
++++ b/src/panfrost/vulkan/panvk_buffer.h	2026-05-22 10:17:41.214043265 +0200
 @@ -14,8 +14,14 @@
  
  struct panvk_priv_bo;
@@ -66,7 +66,7 @@ diff -urN a/src/panfrost/vulkan/panvk_buffer.h b/src/panfrost/vulkan/panvk_buffe
  VK_DEFINE_NONDISP_HANDLE_CASTS(panvk_buffer, vk.base, VkBuffer,
 diff -urN a/src/panfrost/vulkan/panvk_device.h b/src/panfrost/vulkan/panvk_device.h
 --- a/src/panfrost/vulkan/panvk_device.h	2026-05-21 22:46:57.489785206 +0200
-+++ b/src/panfrost/vulkan/panvk_device.h	2026-05-21 22:47:09.189957157 +0200
++++ b/src/panfrost/vulkan/panvk_device.h	2026-05-22 10:17:41.214043265 +0200
 @@ -45,6 +45,8 @@
  enum panvk_queue_family {
     PANVK_QUEUE_FAMILY_GPU,
@@ -102,7 +102,7 @@ diff -urN a/src/panfrost/vulkan/panvk_device.h b/src/panfrost/vulkan/panvk_devic
     struct {
 diff -urN a/src/panfrost/vulkan/panvk_physical_device.c b/src/panfrost/vulkan/panvk_physical_device.c
 --- a/src/panfrost/vulkan/panvk_physical_device.c	2026-05-21 22:46:57.497785323 +0200
-+++ b/src/panfrost/vulkan/panvk_physical_device.c	2026-05-21 22:47:09.189957157 +0200
++++ b/src/panfrost/vulkan/panvk_physical_device.c	2026-05-22 10:17:41.214043265 +0200
 @@ -577,12 +577,22 @@
           .queueFlags = VK_QUEUE_SPARSE_BINDING_BIT,
           .queueCount = 1,
@@ -234,8 +234,8 @@ diff -urN a/src/panfrost/vulkan/panvk_physical_device.c b/src/panfrost/vulkan/pa
 +}
 diff -urN a/src/panfrost/vulkan/panvk_v4l2.c b/src/panfrost/vulkan/panvk_v4l2.c
 --- a/src/panfrost/vulkan/panvk_v4l2.c	1970-01-01 01:00:00.000000000 +0100
-+++ b/src/panfrost/vulkan/panvk_v4l2.c	2026-05-21 22:47:09.189957157 +0200
++++ b/src/panfrost/vulkan/panvk_v4l2.c	2026-05-22 10:17:41.214043265 +0200
-@@ -0,0 +1,569 @@
+@@ -0,0 +1,615 @@
 +/*
 + * panvk-bifrost-video Phase 4 commit 3:
 + *
@@ -250,6 +250,7 @@ diff -urN a/src/panfrost/vulkan/panvk_v4l2.c b/src/panfrost/vulkan/panvk_v4l2.c
 +#include "panvk_video_decode.h"
 +#include "panvk_device.h"
 +
++#include "util/macros.h"
 +#include "vk_alloc.h"
 +#include "vk_log.h"
 +
@@ -417,7 +418,9 @@ diff -urN a/src/panfrost/vulkan/panvk_v4l2.c b/src/panfrost/vulkan/panvk_v4l2.c
 +      mesa_loge("panvk_v4l2: REQBUFS OUTPUT failed: %s", strerror(errno));
 +      return -errno;
 +   }
-+   vs->num_output_buffers = rb.count;
++   /* REQBUFS may round up the count above the request — clamp to our
++    * fixed-size mmap arrays (Phase 5 review: prevents output_map OOB). */
++   vs->num_output_buffers = MIN2(rb.count, 18);
 +   vs->output_next = 0;
 +
 +   /* CAPTURE: MMAP — kernel-allocated, mmap to CPU for copy-out path. */
@@ -430,7 +433,7 @@ diff -urN a/src/panfrost/vulkan/panvk_v4l2.c b/src/panfrost/vulkan/panvk_v4l2.c
 +      mesa_loge("panvk_v4l2: REQBUFS CAPTURE failed: %s", strerror(errno));
 +      return -errno;
 +   }
-+   vs->num_capture_buffers = rb.count;
++   vs->num_capture_buffers = MIN2(rb.count, 18);
 +   vs->capture_next = 0;
 +
 +   return 0;
@@ -788,6 +791,49 @@ diff -urN a/src/panfrost/vulkan/panvk_v4l2.c b/src/panfrost/vulkan/panvk_v4l2.c
 +                          struct vk_device *vk_dev,
 +                          const VkAllocationCallbacks *alloc)
 +{
++   /* Unwind in reverse order of session_init. Each step is guarded by
++    * "have we got far enough to need this" so the function is safe to
++    * call on partially-initialised sessions (the session_init failure
++    * paths jump here via `goto fail`). */
++
++   /* munmap CAPTURE + OUTPUT (no-op for entries left at NULL by an
++    * earlier-failed mmap loop). */
++   for (unsigned i = 0; i < 18; i++) {
++      if (vs->capture_map[i]) {
++         munmap(vs->capture_map[i], vs->capture_map_size[i]);
++         vs->capture_map[i] = NULL;
++         vs->capture_map_size[i] = 0;
++      }
++      if (vs->output_map[i]) {
++         munmap(vs->output_map[i], vs->output_map_size[i]);
++         vs->output_map[i] = NULL;
++         vs->output_map_size[i] = 0;
++      }
++   }
++
++   if (vs->video_fd >= 0) {
++      /* STREAMOFF (safe to call even if STREAMON never ran — kernel
++       * returns EINVAL which we ignore). */
++      enum v4l2_buf_type t;
++      t = vs->mplane ? V4L2_BUF_TYPE_VIDEO_OUTPUT_MPLANE
++                     : V4L2_BUF_TYPE_VIDEO_OUTPUT;
++      (void) ioctl(vs->video_fd, VIDIOC_STREAMOFF, &t);
++      t = vs->mplane ? V4L2_BUF_TYPE_VIDEO_CAPTURE_MPLANE
++                     : V4L2_BUF_TYPE_VIDEO_CAPTURE;
++      (void) ioctl(vs->video_fd, VIDIOC_STREAMOFF, &t);
++
++      /* Release the kernel buffer queues via REQBUFS count=0. */
++      struct v4l2_requestbuffers rb;
++      memset(&rb, 0, sizeof(rb));
++      rb.memory = V4L2_MEMORY_MMAP;
++      rb.type = vs->mplane ? V4L2_BUF_TYPE_VIDEO_OUTPUT_MPLANE
++                           : V4L2_BUF_TYPE_VIDEO_OUTPUT;
++      (void) ioctl(vs->video_fd, VIDIOC_REQBUFS, &rb);
++      rb.type = vs->mplane ? V4L2_BUF_TYPE_VIDEO_CAPTURE_MPLANE
++                           : V4L2_BUF_TYPE_VIDEO_CAPTURE;
++      (void) ioctl(vs->video_fd, VIDIOC_REQBUFS, &rb);
++   }
++
 +   if (vs->request_fds) {
 +      for (unsigned i = 0; i < vs->num_request_fds; i++)
 +         if (vs->request_fds[i] >= 0)
@@ -807,7 +853,7 @@ diff -urN a/src/panfrost/vulkan/panvk_v4l2.c b/src/panfrost/vulkan/panvk_v4l2.c
 +}
 diff -urN a/src/panfrost/vulkan/panvk_v4l2_h264.c b/src/panfrost/vulkan/panvk_v4l2_h264.c
 --- a/src/panfrost/vulkan/panvk_v4l2_h264.c	1970-01-01 01:00:00.000000000 +0100
-+++ b/src/panfrost/vulkan/panvk_v4l2_h264.c	2026-05-21 22:47:09.189957157 +0200
++++ b/src/panfrost/vulkan/panvk_v4l2_h264.c	2026-05-22 10:17:41.214043265 +0200
 @@ -0,0 +1,478 @@
 +/*
 + * panvk-bifrost-video Phase 4: Vulkan StdVideo H.264 → V4L2 stateless H.264
@@ -1289,7 +1335,7 @@ diff -urN a/src/panfrost/vulkan/panvk_v4l2_h264.c b/src/panfrost/vulkan/panvk_v4
 +}
 diff -urN a/src/panfrost/vulkan/panvk_v4l2_h264_slice_header.c b/src/panfrost/vulkan/panvk_v4l2_h264_slice_header.c
 --- a/src/panfrost/vulkan/panvk_v4l2_h264_slice_header.c	1970-01-01 01:00:00.000000000 +0100
-+++ b/src/panfrost/vulkan/panvk_v4l2_h264_slice_header.c	2026-05-21 22:47:09.189957157 +0200
++++ b/src/panfrost/vulkan/panvk_v4l2_h264_slice_header.c	2026-05-22 10:17:41.214043265 +0200
 @@ -0,0 +1,314 @@
 +/*
 + * H.264 slice header bit-parser implementation.
@@ -1607,7 +1653,7 @@ diff -urN a/src/panfrost/vulkan/panvk_v4l2_h264_slice_header.c b/src/panfrost/vu
 +}
 diff -urN a/src/panfrost/vulkan/panvk_v4l2_h264_slice_header.h b/src/panfrost/vulkan/panvk_v4l2_h264_slice_header.h
 --- a/src/panfrost/vulkan/panvk_v4l2_h264_slice_header.h	1970-01-01 01:00:00.000000000 +0100
-+++ b/src/panfrost/vulkan/panvk_v4l2_h264_slice_header.h	2026-05-21 22:47:09.189957157 +0200
++++ b/src/panfrost/vulkan/panvk_v4l2_h264_slice_header.h	2026-05-22 10:17:41.214043265 +0200
 @@ -0,0 +1,94 @@
 +/*
 + * H.264 slice header bit-parser for panvk-bifrost-video / V4L2 stateless
@@ -1705,17 +1751,35 @@ diff -urN a/src/panfrost/vulkan/panvk_v4l2_h264_slice_header.h b/src/panfrost/vu
 +#endif /* PANVK_V4L2_H264_SLICE_HEADER_H */
 diff -urN a/src/panfrost/vulkan/panvk_video_decode.c b/src/panfrost/vulkan/panvk_video_decode.c
 --- a/src/panfrost/vulkan/panvk_video_decode.c	1970-01-01 01:00:00.000000000 +0100
-+++ b/src/panfrost/vulkan/panvk_video_decode.c	2026-05-21 22:47:09.189957157 +0200
++++ b/src/panfrost/vulkan/panvk_video_decode.c	2026-05-22 10:17:41.214043265 +0200
-@@ -0,0 +1,362 @@
+@@ -0,0 +1,380 @@
 +/*
-+ * panvk-bifrost-video Phase 4 commit 7b:
++ * panvk-bifrost-video: Vulkan video decode entrypoints (H.264).
-+ * Vulkan-side decode dispatch wired to V4L2 hantro via dmabuf.
 + *
-+ * Phase 1 simplification: cmd_buffer state tracking via DEVICE-level
++ * Drives the V4L2 stateless hantro VPU backend (panvk_v4l2.c) from
-+ * active_video struct (under a mutex). Per-cmdbuf state hand-off is
++ * Vulkan vkCmdDecodeVideoKHR. Decode is synchronous at record time —
-+ * Phase >>1 once arch-agnostic source can access per-arch cmd_buffer
++ * the full V4L2 ioctl dance runs to completion inside the command-
-+ * structs without the include-path gymnastics. This works for
++ * recording call before returning to the application. The queue-side
-+ * single-session decode workloads (mpv, ffmpeg, vk-video-samples).
++ * `driver_submit` is a no-op signal-everything (see panvk_vX_device.c).
++ *
++ * Phase 1 simplifications worth knowing about:
++ *
++ *  - Cmd-buffer state lives at the DEVICE level (`active_video`) under
++ *    a single mutex, NOT per-cmd-buffer. Concurrent video sessions on
++ *    the same device clobber each other. Sufficient for current single-
++ *    session consumers (mpv-fourier, ffmpeg-vulkan-h264, vk-video-
++ *    samples). Spec-compliant multi-session is a Phase >>1 follow-up.
++ *
++ *  - Source bitstream is read via `src_buf->mem->addr.host`, i.e. the
++ *    bound VkDeviceMemory's CPU mapping. Works because panvk-bifrost
++ *    only exposes HOST_VISIBLE memory types; an app that bound the
++ *    bitstream buffer to non-HOST_VISIBLE memory would get a logged
++ *    error and a silent decode skip (CmdDecodeVideoKHR is void, so we
++ *    have no clean error-return path). VkPhysicalDeviceVideo*
++ *    constraints would be the right place to make this contractual.
++ *
++ *  - Requires `PAN_I_WANT_A_BROKEN_VULKAN_DRIVER=1` (mesa-upstream gate
++ *    on panvk-on-Bifrost which is not conformant).
 + *
 + * SPDX-License-Identifier: MIT
 + */
@@ -1929,10 +1993,10 @@ diff -urN a/src/panfrost/vulkan/panvk_video_decode.c b/src/panfrost/vulkan/panvk
 +    * `tv_sec * 1e9 + tv_usec * 1e3`). Sub-microsecond bits are dropped, so
 +    * any high-resolution stamp (e.g. a 64-bit pointer cast) makes the
 +    * lookup miss and P/B frames decode against zero references. Use a
-+    * monotonic per-session counter in microseconds (i.e. * 1000 ns).
++    * per-session monotonic counter in microseconds (i.e. * 1000 ns) so
++    * concurrent sessions sharing /dev/video1 don't collide on stamp.
 +    */
-+   static uint32_t panvk_video_ts_counter = 0;
++   const uint64_t output_ts = ((uint64_t)++vs->ts_counter) * 1000ULL;
-+   const uint64_t output_ts = ((uint64_t)++panvk_video_ts_counter) * 1000ULL;
 +   uint32_t dst_dpb_slot = pDecodeInfo->pSetupReferenceSlot
 +      ? (uint32_t) pDecodeInfo->pSetupReferenceSlot->slotIndex : 0u;
 +
@@ -2071,8 +2135,8 @@ diff -urN a/src/panfrost/vulkan/panvk_video_decode.c b/src/panfrost/vulkan/panvk
 +}
 diff -urN a/src/panfrost/vulkan/panvk_video_decode.h b/src/panfrost/vulkan/panvk_video_decode.h
 --- a/src/panfrost/vulkan/panvk_video_decode.h	1970-01-01 01:00:00.000000000 +0100
-+++ b/src/panfrost/vulkan/panvk_video_decode.h	2026-05-21 22:47:09.189957157 +0200
++++ b/src/panfrost/vulkan/panvk_video_decode.h	2026-05-22 10:17:41.214043265 +0200
-@@ -0,0 +1,114 @@
+@@ -0,0 +1,124 @@
 +/*
 + * panvk-bifrost-video Phase 4 commit 3: extended for V4L2 state.
 + *
@@ -2103,12 +2167,22 @@ diff -urN a/src/panfrost/vulkan/panvk_video_decode.h b/src/panfrost/vulkan/panvk
 +   struct v4l2_format fmt_output;
 +   struct v4l2_format fmt_capture;
 +
-+   /* Request fd pool. PANVK_V4L2_REQUEST_FD_COUNT entries. */
++   /* Request fd pool. PANVK_V4L2_REQUEST_FD_COUNT entries.
++    * Size of request_fd_used[] is bounded by the same compile-time max;
++    * keep them coupled to avoid silent overflow if the pool grows. */
++#define PANVK_VIDEO_REQUEST_FD_MAX 32
 +   int *request_fds;
-+   bool request_fd_used[32];     /* tracks per-fd "ever queued" → REINIT before reuse */
++   bool request_fd_used[PANVK_VIDEO_REQUEST_FD_MAX];
 +   unsigned num_request_fds;
 +   uint32_t request_fd_next;     /* round-robin index */
 +
++   /* Per-session V4L2 buffer-identity counter. Multiplied by 1000 ns at
++    * QBUF time so the stamp round-trips losslessly through (tv_sec,
++    * tv_usec) — hantro's reflist builder matches dpb[i].reference_ts
++    * against the kernel-side OUTPUT timestamp. Per-session (not process-
++    * global) so concurrent sessions sharing /dev/video1 don't collide. */
++   uint32_t ts_counter;
++
 +   /* DPB slotIndex → V4L2 reference_ts mapping (Phase 1 D5) */
 +   struct {
 +      bool valid;
@@ -2189,7 +2263,7 @@ diff -urN a/src/panfrost/vulkan/panvk_video_decode.h b/src/panfrost/vulkan/panvk
 +#endif /* PANVK_VIDEO_DECODE_H */
 diff -urN a/src/panfrost/vulkan/panvk_vX_device.c b/src/panfrost/vulkan/panvk_vX_device.c
 --- a/src/panfrost/vulkan/panvk_vX_device.c	2026-05-21 22:46:57.505785441 +0200
-+++ b/src/panfrost/vulkan/panvk_vX_device.c	2026-05-21 22:47:09.189957157 +0200
++++ b/src/panfrost/vulkan/panvk_vX_device.c	2026-05-22 10:17:41.214043265 +0200
 @@ -203,6 +203,27 @@
     }
  }
@@ -2372,14 +2446,27 @@ diff -urN a/src/panfrost/vulkan/panvk_vX_device.c b/src/panfrost/vulkan/panvk_vX
        qf->queues =
 diff -urN a/src/panfrost/vulkan/panvk_vX_physical_device.c b/src/panfrost/vulkan/panvk_vX_physical_device.c
 --- a/src/panfrost/vulkan/panvk_vX_physical_device.c	2026-05-21 22:46:59.273811425 +0200
-+++ b/src/panfrost/vulkan/panvk_vX_physical_device.c	2026-05-21 22:47:09.189957157 +0200
++++ b/src/panfrost/vulkan/panvk_vX_physical_device.c	2026-05-22 10:17:41.214043265 +0200
-@@ -170,6 +170,9 @@
+@@ -12,6 +12,7 @@
+ #include <sys/sysmacros.h>
+ 
+ #include "git_sha1.h"
++#include "panvk_video_decode.h"
+ 
+ #include "vk_android.h"
+ #include "vk_device.h"
+@@ -170,6 +171,14 @@
        .EXT_queue_family_foreign = true,
        .EXT_robustness2 = true,
        .EXT_transform_feedback = PAN_ARCH < 9,   /* iter13: JM-class only for now */
-+      .KHR_video_queue = PAN_ARCH < 9,        /* panvk-bifrost-video Phase 4 commit 1 */
++      /* Video extensions are advertised only when (a) we're on a Bifrost
-+      .KHR_video_decode_queue = PAN_ARCH < 9, /* hantro V4L2-stateless backend */
++       * arch (PAN_ARCH < 9) AND (b) a hantro VPU is reachable on the
-+      .KHR_video_decode_h264 = PAN_ARCH < 9,  /* H.264 only initially */
++       * expected V4L2 nodes — otherwise CreateVideoSessionKHR would
++       * succeed at the panvk layer and then fail at v4l2_open_fds, giving
++       * the app a misleading capability claim. */
++      .KHR_video_queue = PAN_ARCH < 9 && panvk_v4l2_probe_hantro(),
++      .KHR_video_decode_queue = PAN_ARCH < 9 && panvk_v4l2_probe_hantro(),
++      .KHR_video_decode_h264 = PAN_ARCH < 9 && panvk_v4l2_probe_hantro(),
        .EXT_sampler_filter_minmax = PAN_ARCH >= 10,
        .EXT_scalar_block_layout = true,
        .EXT_separate_stencil_usage = true,