[Intel-gfx] [PATCH 1/4] drm/i915: Differentiate between sw write location into ring and last hw read
Mika Kuoppala
mika.kuoppala at linux.intel.com
Mon Apr 24 12:21:56 UTC 2017
Chris Wilson <chris at chris-wilson.co.uk> writes:
> We need to keep track of the last location we ask the hw to read up to
> (RING_TAIL) separately from our last write location into the ring, so
> that in the event of a GPU reset we do not tell the HW to proceed into
> a partially written request (which can happen if that request is waiting
> for an external signal before being executed).
But is it so that this can happen also without external signal?
Our submit is already async and waiting for the prev. And we have
already pushed the tail into the partial 'current'.
-Mika
>
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=100144
> Testcase: igt/gem_exec_fence/await-hang
> Fixes: 821ed7df6e2a ("drm/i915: Update reset path to fix incomplete requests")
> Fixes: d55ac5bf97c6 ("drm/i915: Defer transfer onto execution timeline to actual hw submission")
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> Cc: Mika Kuoppala <mika.kuoppala at intel.com>
> ---
> drivers/gpu/drm/i915/i915_gem_request.c | 2 +-
> drivers/gpu/drm/i915/i915_guc_submission.c | 4 +--
> drivers/gpu/drm/i915/intel_lrc.c | 6 ++---
> drivers/gpu/drm/i915/intel_ringbuffer.c | 42 +++++++++++++++++++-----------
> drivers/gpu/drm/i915/intel_ringbuffer.h | 17 +++++++++++-
> 5 files changed, 48 insertions(+), 23 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_gem_request.c b/drivers/gpu/drm/i915/i915_gem_request.c
> index e2ec42b2bf24..99146de1dc93 100644
> --- a/drivers/gpu/drm/i915/i915_gem_request.c
> +++ b/drivers/gpu/drm/i915/i915_gem_request.c
> @@ -651,7 +651,7 @@ i915_gem_request_alloc(struct intel_engine_cs *engine,
> * GPU processing the request, we never over-estimate the
> * position of the head.
> */
> - req->head = req->ring->tail;
> + req->head = req->ring->emit;
>
> /* Check that we didn't interrupt ourselves with a new request */
> GEM_BUG_ON(req->timeline->seqno != req->fence.seqno);
> diff --git a/drivers/gpu/drm/i915/i915_guc_submission.c b/drivers/gpu/drm/i915/i915_guc_submission.c
> index 1642fff9cf13..ab5140ba108d 100644
> --- a/drivers/gpu/drm/i915/i915_guc_submission.c
> +++ b/drivers/gpu/drm/i915/i915_guc_submission.c
> @@ -480,9 +480,7 @@ static void guc_wq_item_append(struct i915_guc_client *client,
> GEM_BUG_ON(freespace < wqi_size);
>
> /* The GuC firmware wants the tail index in QWords, not bytes */
> - tail = rq->tail;
> - assert_ring_tail_valid(rq->ring, rq->tail);
> - tail >>= 3;
> + tail = intel_ring_set_tail(rq->ring, rq->tail) >> 3;
> GEM_BUG_ON(tail > WQ_RING_TAIL_MAX);
>
> /* For now workqueue item is 4 DWs; workqueue buffer is 2 pages. So we
> diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
> index 7df278fe492e..d3612969098f 100644
> --- a/drivers/gpu/drm/i915/intel_lrc.c
> +++ b/drivers/gpu/drm/i915/intel_lrc.c
> @@ -326,8 +326,7 @@ static u64 execlists_update_context(struct drm_i915_gem_request *rq)
> rq->ctx->ppgtt ?: rq->i915->mm.aliasing_ppgtt;
> u32 *reg_state = ce->lrc_reg_state;
>
> - assert_ring_tail_valid(rq->ring, rq->tail);
> - reg_state[CTX_RING_TAIL+1] = rq->tail;
> + reg_state[CTX_RING_TAIL+1] = intel_ring_set_tail(rq->ring, rq->tail);
>
> /* True 32b PPGTT with dynamic page allocation: update PDP
> * registers and point the unallocated PDPs to scratch page.
> @@ -2054,7 +2053,8 @@ void intel_lr_context_resume(struct drm_i915_private *dev_priv)
> ce->state->obj->mm.dirty = true;
> i915_gem_object_unpin_map(ce->state->obj);
>
> - ce->ring->head = ce->ring->tail = 0;
> + GEM_BUG_ON(!list_empty(&ce->ring->request_list));
> + ce->ring->head = ce->ring->tail = ce->ring->emit = 0;
> intel_ring_update_space(ce->ring);
> }
> }
> diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c
> index 32afac6c754f..8c874996c617 100644
> --- a/drivers/gpu/drm/i915/intel_ringbuffer.c
> +++ b/drivers/gpu/drm/i915/intel_ringbuffer.c
> @@ -49,7 +49,7 @@ static int __intel_ring_space(int head, int tail, int size)
>
> void intel_ring_update_space(struct intel_ring *ring)
> {
> - ring->space = __intel_ring_space(ring->head, ring->tail, ring->size);
> + ring->space = __intel_ring_space(ring->head, ring->emit, ring->size);
> }
>
> static int
> @@ -774,8 +774,8 @@ static void i9xx_submit_request(struct drm_i915_gem_request *request)
>
> i915_gem_request_submit(request);
>
> - assert_ring_tail_valid(request->ring, request->tail);
> - I915_WRITE_TAIL(request->engine, request->tail);
> + I915_WRITE_TAIL(request->engine,
> + intel_ring_set_tail(request->ring, request->tail));
> }
>
> static void i9xx_emit_breadcrumb(struct drm_i915_gem_request *req, u32 *cs)
> @@ -1324,6 +1324,12 @@ void intel_ring_unpin(struct intel_ring *ring)
> GEM_BUG_ON(!ring->vma);
> GEM_BUG_ON(!ring->vaddr);
>
> + /* Discard any unused bytes beyond that submitted to hw. */
> + GEM_BUG_ON(!list_empty(&ring->request_list));
> + ring->head = ring->tail;
> + ring->emit = ring->tail;
> + intel_ring_update_space(ring);
> +
> if (i915_vma_is_map_and_fenceable(ring->vma))
> i915_vma_unpin_iomap(ring->vma);
> else
> @@ -1555,8 +1561,14 @@ void intel_legacy_submission_resume(struct drm_i915_private *dev_priv)
> struct intel_engine_cs *engine;
> enum intel_engine_id id;
>
> - for_each_engine(engine, dev_priv, id)
> - engine->buffer->head = engine->buffer->tail;
> + /* Restart from the beginning of the rings for convenience */
> + for_each_engine(engine, dev_priv, id) {
> + struct intel_ring *ring = engine->buffer;
> +
> + GEM_BUG_ON(!list_empty(&ring->request_list));
> + ring->head = ring->tail = ring->emit = 0;
> + intel_ring_update_space(ring);
> + }
> }
>
> static int ring_request_alloc(struct drm_i915_gem_request *request)
> @@ -1609,7 +1621,7 @@ static int wait_for_space(struct drm_i915_gem_request *req, int bytes)
> unsigned space;
>
> /* Would completion of this request free enough space? */
> - space = __intel_ring_space(target->postfix, ring->tail,
> + space = __intel_ring_space(target->postfix, ring->emit,
> ring->size);
> if (space >= bytes)
> break;
> @@ -1634,8 +1646,8 @@ static int wait_for_space(struct drm_i915_gem_request *req, int bytes)
> u32 *intel_ring_begin(struct drm_i915_gem_request *req, int num_dwords)
> {
> struct intel_ring *ring = req->ring;
> - int remain_actual = ring->size - ring->tail;
> - int remain_usable = ring->effective_size - ring->tail;
> + int remain_actual = ring->size - ring->emit;
> + int remain_usable = ring->effective_size - ring->emit;
> int bytes = num_dwords * sizeof(u32);
> int total_bytes, wait_bytes;
> bool need_wrap = false;
> @@ -1671,17 +1683,17 @@ u32 *intel_ring_begin(struct drm_i915_gem_request *req, int num_dwords)
>
> if (unlikely(need_wrap)) {
> GEM_BUG_ON(remain_actual > ring->space);
> - GEM_BUG_ON(ring->tail + remain_actual > ring->size);
> + GEM_BUG_ON(ring->emit + remain_actual > ring->size);
>
> /* Fill the tail with MI_NOOP */
> - memset(ring->vaddr + ring->tail, 0, remain_actual);
> - ring->tail = 0;
> + memset(ring->vaddr + ring->emit, 0, remain_actual);
> + ring->emit = 0;
> ring->space -= remain_actual;
> }
>
> - GEM_BUG_ON(ring->tail > ring->size - bytes);
> - cs = ring->vaddr + ring->tail;
> - ring->tail += bytes;
> + GEM_BUG_ON(ring->emit > ring->size - bytes);
> + cs = ring->vaddr + ring->emit;
> + ring->emit += bytes;
> ring->space -= bytes;
> GEM_BUG_ON(ring->space < 0);
>
> @@ -1692,7 +1704,7 @@ u32 *intel_ring_begin(struct drm_i915_gem_request *req, int num_dwords)
> int intel_ring_cacheline_align(struct drm_i915_gem_request *req)
> {
> int num_dwords =
> - (req->ring->tail & (CACHELINE_BYTES - 1)) / sizeof(uint32_t);
> + (req->ring->emit & (CACHELINE_BYTES - 1)) / sizeof(uint32_t);
> u32 *cs;
>
> if (num_dwords == 0)
> diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.h b/drivers/gpu/drm/i915/intel_ringbuffer.h
> index 00d36aa4e26d..31998ea46fb8 100644
> --- a/drivers/gpu/drm/i915/intel_ringbuffer.h
> +++ b/drivers/gpu/drm/i915/intel_ringbuffer.h
> @@ -143,6 +143,7 @@ struct intel_ring {
>
> u32 head;
> u32 tail;
> + u32 emit;
>
> int space;
> int size;
> @@ -517,7 +518,7 @@ intel_ring_advance(struct drm_i915_gem_request *req, u32 *cs)
> * reserved for the command packet (i.e. the value passed to
> * intel_ring_begin()).
> */
> - GEM_BUG_ON((req->ring->vaddr + req->ring->tail) != cs);
> + GEM_BUG_ON((req->ring->vaddr + req->ring->emit) != cs);
> }
>
> static inline u32
> @@ -546,6 +547,20 @@ assert_ring_tail_valid(const struct intel_ring *ring, unsigned int tail)
> GEM_BUG_ON(tail >= ring->size);
> }
>
> +static inline unsigned int
> +intel_ring_set_tail(struct intel_ring *ring, unsigned int tail)
> +{
> + /* Whilst writes to the tail are strictly order, there is no
> + * serialisation between readers and the writers. The tail may be
> + * read by i915_gem_request_retire() just as it is being updated
> + * by execlists, as although the breadcrumb is complete, the context
> + * switch hasn't been seen.
> + */
> + assert_ring_tail_valid(ring, tail);
> + ring->tail = tail;
> + return tail;
> +}
> +
> void intel_ring_update_space(struct intel_ring *ring);
>
> void intel_engine_init_global_seqno(struct intel_engine_cs *engine, u32 seqno);
> --
> 2.11.0
More information about the Intel-gfx
mailing list