On Wed, Nov 24, 2021 at 11:36 PM Dmitry Baryshkov dmitry.baryshkov@linaro.org wrote:
On 28/04/2021 22:36, Rob Clark wrote:
From: Rob Clark robdclark@chromium.org
Currently if userspace manages to fill up the ring faster than the GPU can consume we (a) spin for up to 1sec, and then (b) overwrite the ringbuffer contents from previous submits that the GPU is still busy executing. Which predictably goes rather badly.
Instead, just skip flushing (updating WPTR) and reset ring->next back to where it was before we tried writing the submit into the ringbuffer, and return an error to userspace (which can then try again).
Signed-off-by: Rob Clark robdclark@chromium.org
Rob, you've posted this patch, but never merged it. Should it be merged at some point?
I think it is a bit less needed now, since drm/sched will limit the # of in-flight submits (when I sent that patch, it was before conversion to use drm/sched)
With a bit more locking re-work we could do something more clever like just blocking until there is space in the ringbuffer.. but aren't there quite yet.
BR, -R
drivers/gpu/drm/msm/adreno/a5xx_gpu.c | 3 +++ drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 3 +++ drivers/gpu/drm/msm/adreno/adreno_gpu.c | 24 +++++++++++++++++- drivers/gpu/drm/msm/msm_gem_submit.c | 7 +++++- drivers/gpu/drm/msm/msm_gpu.c | 33 +++++++++++++++++++++++-- drivers/gpu/drm/msm/msm_gpu.h | 2 +- drivers/gpu/drm/msm/msm_ringbuffer.h | 5 ++++ 7 files changed, 72 insertions(+), 5 deletions(-)
diff --git a/drivers/gpu/drm/msm/adreno/a5xx_gpu.c b/drivers/gpu/drm/msm/adreno/a5xx_gpu.c index ce13d49e615b..0c8faad3b328 100644 --- a/drivers/gpu/drm/msm/adreno/a5xx_gpu.c +++ b/drivers/gpu/drm/msm/adreno/a5xx_gpu.c @@ -36,6 +36,9 @@ void a5xx_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring, OUT_RING(ring, upper_32_bits(shadowptr(a5xx_gpu, ring))); }
if (unlikely(ring->overflow))return;spin_lock_irqsave(&ring->preempt_lock, flags); /* Copy the shadow to the actual register */diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c index d553f62f4eeb..4a4728a774c0 100644 --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c @@ -68,6 +68,9 @@ static void a6xx_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring) OUT_RING(ring, upper_32_bits(shadowptr(a6xx_gpu, ring))); }
if (unlikely(ring->overflow))return;spin_lock_irqsave(&ring->preempt_lock, flags); /* Copy the shadow to the actual register */diff --git a/drivers/gpu/drm/msm/adreno/adreno_gpu.c b/drivers/gpu/drm/msm/adreno/adreno_gpu.c index 0f184c3dd9d9..a658777e07b1 100644 --- a/drivers/gpu/drm/msm/adreno/adreno_gpu.c +++ b/drivers/gpu/drm/msm/adreno/adreno_gpu.c @@ -467,6 +467,9 @@ void adreno_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring, u32 reg) { uint32_t wptr;
if (unlikely(ring->overflow))return;/* Copy the shadow to the actual register */ ring->cur = ring->next;@@ -788,12 +791,31 @@ static uint32_t ring_freewords(struct msm_ringbuffer *ring) return (rptr + (size - 1) - wptr) % size; }
+static bool space_avail(struct msm_ringbuffer *ring, uint32_t ndwords) +{
if (ring_freewords(ring) >= ndwords)return true;/* We don't have a good way to know in general when the RPTR has* advanced.. newer things that use CP_WHERE_AM_I to update the* shadow rptr could possibly insert a packet to generate an irq.* But that doesn't cover older GPUs. But if the ringbuffer is* full, it could take a while before it is empty again, so just* insert a blind sleep to avoid a busy loop.*/msleep(1);return false;+}
- void adreno_wait_ring(struct msm_ringbuffer *ring, uint32_t ndwords) {
if (spin_until(ring_freewords(ring) >= ndwords))
if (spin_until(space_avail(ring, ndwords))) { DRM_DEV_ERROR(ring->gpu->dev->dev, "timeout waiting for space in ringbuffer %d\n", ring->id);ring->overflow = true;}}
/* Get legacy powerlevels from qcom,gpu-pwrlevels and populate the opp table */
diff --git a/drivers/gpu/drm/msm/msm_gem_submit.c b/drivers/gpu/drm/msm/msm_gem_submit.c index 5480852bdeda..4bc669460fda 100644 --- a/drivers/gpu/drm/msm/msm_gem_submit.c +++ b/drivers/gpu/drm/msm/msm_gem_submit.c @@ -683,6 +683,9 @@ int msm_ioctl_gem_submit(struct drm_device *dev, void *data, submitid = atomic_inc_return(&ident) - 1;
ring = gpu->rb[queue->prio];
GEM_WARN_ON(ring->overflow);trace_msm_gpu_submit(pid_nr(pid), ring->id, submitid, args->nr_bos, args->nr_cmds);@@ -829,7 +832,9 @@ int msm_ioctl_gem_submit(struct drm_device *dev, void *data, } }
msm_gpu_submit(gpu, submit);
ret = msm_gpu_submit(gpu, submit);if (ret)goto out; args->fence = submit->fence->seqno;diff --git a/drivers/gpu/drm/msm/msm_gpu.c b/drivers/gpu/drm/msm/msm_gpu.c index ab7c167b0623..7655ad9108c8 100644 --- a/drivers/gpu/drm/msm/msm_gpu.c +++ b/drivers/gpu/drm/msm/msm_gpu.c @@ -787,7 +787,7 @@ void msm_gpu_retire(struct msm_gpu *gpu) }
/* add bo's to gpu's ring, and kick gpu: */ -void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit) +int msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit) { struct drm_device *dev = gpu->dev; struct msm_drm_private *priv = dev->dev_private; @@ -834,9 +834,38 @@ void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit) spin_unlock(&ring->submit_lock);
gpu->funcs->submit(gpu, submit);
priv->lastctx = submit->queue->ctx; hangcheck_timer_reset(gpu);
if (unlikely(ring->overflow)) {/** Reset the ptr back to before the submit, so the GPU* doesn't see a partial submit:*/ring->next = ring->cur;/** Clear the overflow flag, hopefully the next submit on* the ring actually fits*/ring->overflow = false;/** One might be tempted to remove the submit from the* submits list, and drop it's reference (and drop the* active reference for all the bos). But we can't* really signal the fence attached to obj->resv without* disturbing other fences on the timeline. So instead* just leave it and let it retire normally when a* later submit completes.*/return -ENOSPC;}priv->lastctx = submit->queue->ctx;return 0;}
/*
diff --git a/drivers/gpu/drm/msm/msm_gpu.h b/drivers/gpu/drm/msm/msm_gpu.h index d7cd02cd2109..2dd2ef1f8328 100644 --- a/drivers/gpu/drm/msm/msm_gpu.h +++ b/drivers/gpu/drm/msm/msm_gpu.h @@ -302,7 +302,7 @@ int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime, uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs);
void msm_gpu_retire(struct msm_gpu *gpu); -void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit); +int msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit);
int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev, struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs, diff --git a/drivers/gpu/drm/msm/msm_ringbuffer.h b/drivers/gpu/drm/msm/msm_ringbuffer.h index fe55d4a1aa16..d8ad9818c389 100644 --- a/drivers/gpu/drm/msm/msm_ringbuffer.h +++ b/drivers/gpu/drm/msm/msm_ringbuffer.h @@ -40,6 +40,8 @@ struct msm_ringbuffer { struct drm_gem_object *bo; uint32_t *start, *end, *cur, *next;
bool overflow;/* * List of in-flight submits on this ring. Protected by submit_lock. */@@ -69,6 +71,9 @@ void msm_ringbuffer_destroy(struct msm_ringbuffer *ring); static inline void OUT_RING(struct msm_ringbuffer *ring, uint32_t data) {
if (ring->overflow)return;/* * ring->next points to the current command being written - it won't be * committed as ring->cur until the flush-- With best wishes Dmitry