git.ipfire.org Git - thirdparty/openwrt.git/blob

   1 From 131564261399a36a5cf2ac2731ed1ceffba93d10 Mon Sep 17 00:00:00 2001
   2 From: =?UTF-8?q?Ma=C3=ADra=20Canal?= <mcanal@igalia.com>
   3 Date: Sat, 22 Feb 2025 14:56:46 -0300
   4 Subject: [PATCH] drm/v3d: Set job pointer to NULL when the job's fence has an
   5  error
   6 MIME-Version: 1.0
   7 Content-Type: text/plain; charset=UTF-8
   8 Content-Transfer-Encoding: 8bit
   9
  10 Similar to commit e4b5ccd392b9 ("drm/v3d: Ensure job pointer is set to
  11 NULL after job completion"), ensure the job pointer is set to `NULL` when
  12 a job's fence has an error. Failing to do so can trigger kernel warnings
  13 in specific scenarios, such as:
  14
  15 1. v3d_csd_job_run() assigns `v3d->csd_job = job`
  16 2. CSD job exceeds hang limit, causing a timeout → v3d_gpu_reset_for_timeout()
  17 3. GPU reset
  18 4. drm_sched_resubmit_jobs() sets the job's fence to `-ECANCELED`.
  19 5. v3d_csd_job_run() detects the fence error and returns NULL, not
  20    submitting the job to the GPU
  21 6. User-space runs `modprobe -r v3d`
  22 7. v3d_gem_destroy()
  23
  24 v3d_gem_destroy() triggers a warning indicating that the CSD job never
  25 ended, as we didn't set `v3d->csd_job` to NULL after the timeout. The same
  26 can also happen to BIN, RENDER, and TFU jobs.
  27
  28 Signed-off-by: Maíra Canal <mcanal@igalia.com>
  29 ---
  30  drivers/gpu/drm/v3d/v3d_sched.c | 18 ++++++++++++++----
  31  1 file changed, 14 insertions(+), 4 deletions(-)
  32
  33 --- a/drivers/gpu/drm/v3d/v3d_sched.c
  34 +++ b/drivers/gpu/drm/v3d/v3d_sched.c
  35 @@ -189,8 +189,12 @@ static struct dma_fence *v3d_bin_job_run
  36         struct dma_fence *fence;
  37         unsigned long irqflags;
  38
  39 -       if (unlikely(job->base.base.s_fence->finished.error))
  40 +       if (unlikely(job->base.base.s_fence->finished.error)) {
  41 +               spin_lock_irqsave(&v3d->job_lock, irqflags);
  42 +               v3d->bin_job = NULL;
  43 +               spin_unlock_irqrestore(&v3d->job_lock, irqflags);
  44                 return NULL;
  45 +       }
  46
  47         /* Lock required around bin_job update vs
  48          * v3d_overflow_mem_work().
  49 @@ -244,8 +248,10 @@ static struct dma_fence *v3d_render_job_
  50         struct drm_device *dev = &v3d->drm;
  51         struct dma_fence *fence;
  52
  53 -       if (unlikely(job->base.base.s_fence->finished.error))
  54 +       if (unlikely(job->base.base.s_fence->finished.error)) {
  55 +               v3d->render_job = NULL;
  56                 return NULL;
  57 +       }
  58
  59         v3d->render_job = job;
  60
  61 @@ -292,8 +298,10 @@ v3d_tfu_job_run(struct drm_sched_job *sc
  62         struct drm_device *dev = &v3d->drm;
  63         struct dma_fence *fence;
  64
  65 -       if (unlikely(job->base.base.s_fence->finished.error))
  66 +       if (unlikely(job->base.base.s_fence->finished.error)) {
  67 +               v3d->tfu_job = NULL;
  68                 return NULL;
  69 +       }
  70
  71         v3d->tfu_job = job;
  72
  73 @@ -337,8 +345,10 @@ v3d_csd_job_run(struct drm_sched_job *sc
  74         struct dma_fence *fence;
  75         int i, csd_cfg0_reg, csd_cfg_reg_count;
  76
  77 -       if (unlikely(job->base.base.s_fence->finished.error))
  78 +       if (unlikely(job->base.base.s_fence->finished.error)) {
  79 +               v3d->csd_job = NULL;
  80                 return NULL;
  81 +       }
  82
  83         v3d->csd_job = job;
  84