On Tue, Aug 03, 2021 at 03:29:01PM -0700, Matthew Brost wrote:
> Add a heuristic which checks if over half of the available guc_ids are
> currently consumed by requests not ready to be submitted. If this
> heuristic is true at request creation time (normal guc_id allocation
> location) force all submissions + guc_ids allocations to tasklet.
>
> Signed-off-by: Matthew Brost <matthew.brost@xxxxxxxxx>
> ---
> drivers/gpu/drm/i915/gt/intel_context_types.h | 3 ++
> drivers/gpu/drm/i915/gt/intel_reset.c | 9 ++++
> drivers/gpu/drm/i915/gt/uc/intel_guc.h | 1 +
> .../gpu/drm/i915/gt/uc/intel_guc_submission.c | 53 +++++++++++++++++--
> .../gpu/drm/i915/gt/uc/intel_guc_submission.h | 2 +
> 5 files changed, 65 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/gt/intel_context_types.h b/drivers/gpu/drm/i915/gt/intel_context_types.h
> index 8ed964ef967b..c01530d7dc67 100644
> --- a/drivers/gpu/drm/i915/gt/intel_context_types.h
> +++ b/drivers/gpu/drm/i915/gt/intel_context_types.h
> @@ -188,6 +188,9 @@ struct intel_context {
> /* Number of rq submitted without a guc_id */
> u16 guc_num_rq_submit_no_id;
>
> + /* GuC number of requests not ready */
> + atomic_t guc_num_rq_not_ready;
atomic_t by default is unordered. This needs some giantic comments and
explainers why this is totally ok and we don't need barriers here.
I think good excuse to convert all the docs here into kerneldoc.
-Daniel
> +
> /*
> * GuC ID link - in list when unpinned but guc_id still valid in GuC
> */
> diff --git a/drivers/gpu/drm/i915/gt/intel_reset.c b/drivers/gpu/drm/i915/gt/intel_reset.c
> index 91200c43951f..ea763138197f 100644
> --- a/drivers/gpu/drm/i915/gt/intel_reset.c
> +++ b/drivers/gpu/drm/i915/gt/intel_reset.c
> @@ -22,6 +22,7 @@
> #include "intel_reset.h"
>
> #include "uc/intel_guc.h"
> +#include "uc/intel_guc_submission.h"
>
> #define RESET_MAX_RETRIES 3
>
> @@ -850,6 +851,14 @@ static void nop_submit_request(struct i915_request *request)
> {
> RQ_TRACE(request, "-EIO\n");
>
> + /*
> + * XXX: Kinda ugly to check for GuC submission here but this function is
> + * going away once we switch to the DRM scheduler so we can live with
> + * this for now.
> + */
> + if (intel_engine_uses_guc(request->engine))
> + intel_guc_decr_num_rq_not_ready(request->context);
> +
> request = i915_request_mark_eio(request);
> if (request) {
> i915_request_submit(request);
> diff --git a/drivers/gpu/drm/i915/gt/uc/intel_guc.h b/drivers/gpu/drm/i915/gt/uc/intel_guc.h
> index e76579396efd..917352c9f323 100644
> --- a/drivers/gpu/drm/i915/gt/uc/intel_guc.h
> +++ b/drivers/gpu/drm/i915/gt/uc/intel_guc.h
> @@ -76,6 +76,7 @@ struct intel_guc {
> struct ida guc_ids;
> u32 num_guc_ids;
> u32 max_guc_ids;
> + atomic_t num_guc_ids_not_ready;
> struct list_head guc_id_list_no_ref;
> struct list_head guc_id_list_unpinned;
>
> diff --git a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
> index f42a707f60ca..ba750fc87af1 100644
> --- a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
> +++ b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
> @@ -1384,6 +1384,41 @@ static inline void queue_request(struct i915_sched_engine *sched_engine,
> kick_tasklet(&rq->engine->gt->uc.guc);
> }
>
> +/* Macro to tweak heuristic, using a simple over 50% not ready for now */
> +#define TOO_MANY_GUC_IDS_NOT_READY(avail, consumed) \
> + ((consumed) > (avail) / 2)
> +static bool too_many_guc_ids_not_ready(struct intel_guc *guc,
> + struct intel_context *ce)
> +{
> + u32 available_guc_ids, guc_ids_consumed;
> +
> + available_guc_ids = guc->num_guc_ids;
> + guc_ids_consumed = atomic_read(&guc->num_guc_ids_not_ready);
> +
> + if (TOO_MANY_GUC_IDS_NOT_READY(available_guc_ids, guc_ids_consumed)) {
> + set_and_update_guc_ids_exhausted(guc);
> + return true;
> + }
> +
> + return false;
> +}
> +
> +static void incr_num_rq_not_ready(struct intel_context *ce)
> +{
> + struct intel_guc *guc = ce_to_guc(ce);
> +
> + if (!atomic_fetch_add(1, &ce->guc_num_rq_not_ready))
> + atomic_inc(&guc->num_guc_ids_not_ready);
> +}
> +
> +void intel_guc_decr_num_rq_not_ready(struct intel_context *ce)
> +{
> + struct intel_guc *guc = ce_to_guc(ce);
> +
> + if (atomic_fetch_add(-1, &ce->guc_num_rq_not_ready) == 1)
> + atomic_dec(&guc->num_guc_ids_not_ready);
> +}
> +
> static bool need_tasklet(struct intel_guc *guc, struct intel_context *ce)
> {
> struct i915_sched_engine * const sched_engine =
> @@ -1430,6 +1465,8 @@ static void guc_submit_request(struct i915_request *rq)
> kick_tasklet(guc);
>
> spin_unlock_irqrestore(&sched_engine->lock, flags);
> +
> + intel_guc_decr_num_rq_not_ready(rq->context);
> }
>
> static int new_guc_id(struct intel_guc *guc)
> @@ -2647,10 +2684,13 @@ static int guc_request_alloc(struct i915_request *rq)
> rq->reserved_space -= GUC_REQUEST_SIZE;
>
> /*
> - * guc_ids are exhausted, don't allocate one here, defer to submission
> - * in the tasklet.
> + * guc_ids are exhausted or a heuristic is met indicating too many
> + * guc_ids are waiting on requests with submission dependencies (not
> + * ready to submit). Don't allocate one here, defer to submission in the
> + * tasklet.
> */
> - if (test_and_update_guc_ids_exhausted(guc)) {
> + if (test_and_update_guc_ids_exhausted(guc) ||
> + too_many_guc_ids_not_ready(guc, ce)) {
> set_bit(I915_FENCE_FLAG_GUC_ID_NOT_PINNED, &rq->fence.flags);
> goto out;
> }
> @@ -2684,6 +2724,7 @@ static int guc_request_alloc(struct i915_request *rq)
> */
> set_bit(I915_FENCE_FLAG_GUC_ID_NOT_PINNED, &rq->fence.flags);
> set_and_update_guc_ids_exhausted(guc);
> + incr_num_rq_not_ready(ce);
>
> return 0;
> } else if (unlikely(ret < 0)) {
> @@ -2708,6 +2749,8 @@ static int guc_request_alloc(struct i915_request *rq)
> clear_bit(CONTEXT_LRCA_DIRTY, &ce->flags);
>
> out:
> + incr_num_rq_not_ready(ce);
> +
> /*
> * We block all requests on this context if a G2H is pending for a
> * schedule disable or context deregistration as the GuC will fail a
> @@ -3512,6 +3555,8 @@ void intel_guc_submission_print_info(struct intel_guc *guc,
> drm_printf(p, "GuC submit flags: 0x%04lx\n", guc->flags);
> drm_printf(p, "GuC total number request without guc_id: %d\n",
> guc->total_num_rq_with_no_guc_id);
> + drm_printf(p, "GuC Number GuC IDs not ready: %d\n",
> + atomic_read(&guc->num_guc_ids_not_ready));
> drm_printf(p, "GuC stall reason: %d\n", guc->submission_stall_reason);
> drm_printf(p, "GuC stalled request: %s\n",
> yesno(guc->stalled_rq));
> @@ -3567,6 +3612,8 @@ void intel_guc_submission_print_context_info(struct intel_guc *guc,
> atomic_read(&ce->pin_count));
> drm_printf(p, "\t\tGuC ID Ref Count: %u\n",
> atomic_read(&ce->guc_id_ref));
> + drm_printf(p, "\t\tNumber Requests Not Ready: %u\n",
> + atomic_read(&ce->guc_num_rq_not_ready));
> drm_printf(p, "\t\tSchedule State: 0x%x, 0x%x\n\n",
> ce->guc_state.sched_state,
> atomic_read(&ce->guc_sched_state_no_lock));
> diff --git a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.h b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.h
> index c7ef44fa0c36..17af5e123b09 100644
> --- a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.h
> +++ b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.h
> @@ -51,4 +51,6 @@ static inline bool intel_guc_submission_is_used(struct intel_guc *guc)
> return intel_guc_is_used(guc) && intel_guc_submission_is_wanted(guc);
> }
>
> +void intel_guc_decr_num_rq_not_ready(struct intel_context *ce);
> +
> #endif
> --
> 2.28.0
>
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
