diff options
Diffstat (limited to 'drivers/gpu/drm/i915/i915_request.c')
| -rw-r--r-- | drivers/gpu/drm/i915/i915_request.c | 775 |
1 files changed, 439 insertions, 336 deletions
diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c index a195a92d0105..78a5f5d3c070 100644 --- a/drivers/gpu/drm/i915/i915_request.c +++ b/drivers/gpu/drm/i915/i915_request.c @@ -31,10 +31,13 @@ #include "gem/i915_gem_context.h" #include "gt/intel_context.h" +#include "gt/intel_ring.h" +#include "gt/intel_rps.h" #include "i915_active.h" #include "i915_drv.h" #include "i915_globals.h" +#include "i915_trace.h" #include "intel_pm.h" struct execute_cb { @@ -54,11 +57,13 @@ static struct i915_global_request { static const char *i915_fence_get_driver_name(struct dma_fence *fence) { - return "i915"; + return dev_name(to_request(fence)->i915->drm.dev); } static const char *i915_fence_get_timeline_name(struct dma_fence *fence) { + const struct i915_gem_context *ctx; + /* * The timeline struct (as part of the ppgtt underneath a context) * may be freed when the request is no longer in use by the GPU. @@ -71,7 +76,11 @@ static const char *i915_fence_get_timeline_name(struct dma_fence *fence) if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) return "signaled"; - return to_request(fence)->gem_context->name ?: "[i915]"; + ctx = i915_request_gem_context(to_request(fence)); + if (!ctx) + return "[" DRIVER_NAME "]"; + + return ctx->name; } static bool i915_fence_signaled(struct dma_fence *fence) @@ -119,62 +128,73 @@ const struct dma_fence_ops i915_fence_ops = { .release = i915_fence_release, }; -static inline void -i915_request_remove_from_client(struct i915_request *request) +static void irq_execute_cb(struct irq_work *wrk) { - struct drm_i915_file_private *file_priv; + struct execute_cb *cb = container_of(wrk, typeof(*cb), work); - file_priv = request->file_priv; - if (!file_priv) - return; + i915_sw_fence_complete(cb->fence); + kmem_cache_free(global.slab_execute_cbs, cb); +} - spin_lock(&file_priv->mm.lock); - if (request->file_priv) { - list_del(&request->client_link); - request->file_priv = NULL; - } - spin_unlock(&file_priv->mm.lock); +static void irq_execute_cb_hook(struct irq_work *wrk) +{ + struct execute_cb *cb = container_of(wrk, typeof(*cb), work); + + cb->hook(container_of(cb->fence, struct i915_request, submit), + &cb->signal->fence); + i915_request_put(cb->signal); + + irq_execute_cb(wrk); } -static void advance_ring(struct i915_request *request) +static void __notify_execute_cb(struct i915_request *rq) { - struct intel_ring *ring = request->ring; - unsigned int tail; + struct execute_cb *cb; + + lockdep_assert_held(&rq->lock); + + if (list_empty(&rq->execute_cb)) + return; + + list_for_each_entry(cb, &rq->execute_cb, link) + irq_work_queue(&cb->work); /* - * We know the GPU must have read the request to have - * sent us the seqno + interrupt, so use the position - * of tail of the request to update the last known position - * of the GPU head. + * XXX Rollback on __i915_request_unsubmit() * - * Note this requires that we are always called in request - * completion order. + * In the future, perhaps when we have an active time-slicing scheduler, + * it will be interesting to unsubmit parallel execution and remove + * busywaits from the GPU until their master is restarted. This is + * quite hairy, we have to carefully rollback the fence and do a + * preempt-to-idle cycle on the target engine, all the while the + * master execute_cb may refire. */ - GEM_BUG_ON(!list_is_first(&request->ring_link, &ring->request_list)); - if (list_is_last(&request->ring_link, &ring->request_list)) { - /* - * We may race here with execlists resubmitting this request - * as we retire it. The resubmission will move the ring->tail - * forwards (to request->wa_tail). We either read the - * current value that was written to hw, or the value that - * is just about to be. Either works, if we miss the last two - * noops - they are safe to be replayed on a reset. - */ - tail = READ_ONCE(request->tail); - list_del(&ring->active_link); - } else { - tail = request->postfix; - } - list_del_init(&request->ring_link); + INIT_LIST_HEAD(&rq->execute_cb); +} + +static inline void +remove_from_client(struct i915_request *request) +{ + struct drm_i915_file_private *file_priv; + + if (!READ_ONCE(request->file_priv)) + return; - ring->head = tail; + rcu_read_lock(); + file_priv = xchg(&request->file_priv, NULL); + if (file_priv) { + spin_lock(&file_priv->mm.lock); + list_del(&request->client_link); + spin_unlock(&file_priv->mm.lock); + } + rcu_read_unlock(); } static void free_capture_list(struct i915_request *request) { struct i915_capture_list *capture; - capture = request->capture_list; + capture = fetch_and_zero(&request->capture_list); while (capture) { struct i915_capture_list *next = capture->next; @@ -183,79 +203,83 @@ static void free_capture_list(struct i915_request *request) } } -static bool i915_request_retire(struct i915_request *rq) +static void remove_from_engine(struct i915_request *rq) { - struct i915_active_request *active, *next; + struct intel_engine_cs *engine, *locked; + + /* + * Virtual engines complicate acquiring the engine timeline lock, + * as their rq->engine pointer is not stable until under that + * engine lock. The simple ploy we use is to take the lock then + * check that the rq still belongs to the newly locked engine. + */ + locked = READ_ONCE(rq->engine); + spin_lock_irq(&locked->active.lock); + while (unlikely(locked != (engine = READ_ONCE(rq->engine)))) { + spin_unlock(&locked->active.lock); + spin_lock(&engine->active.lock); + locked = engine; + } + list_del_init(&rq->sched.link); + clear_bit(I915_FENCE_FLAG_PQUEUE, &rq->fence.flags); + clear_bit(I915_FENCE_FLAG_HOLD, &rq->fence.flags); + spin_unlock_irq(&locked->active.lock); +} - lockdep_assert_held(&rq->i915->drm.struct_mutex); +bool i915_request_retire(struct i915_request *rq) +{ if (!i915_request_completed(rq)) return false; - GEM_TRACE("%s fence %llx:%lld, current %d\n", - rq->engine->name, - rq->fence.context, rq->fence.seqno, - hwsp_seqno(rq)); + RQ_TRACE(rq, "\n"); GEM_BUG_ON(!i915_sw_fence_signaled(&rq->submit)); trace_i915_request_retire(rq); - advance_ring(rq); - /* - * Walk through the active list, calling retire on each. This allows - * objects to track their GPU activity and mark themselves as idle - * when their *last* active request is completed (updating state - * tracking lists for eviction, active references for GEM, etc). + * We know the GPU must have read the request to have + * sent us the seqno + interrupt, so use the position + * of tail of the request to update the last known position + * of the GPU head. * - * As the ->retire() may free the node, we decouple it first and - * pass along the auxiliary information (to avoid dereferencing - * the node after the callback). + * Note this requires that we are always called in request + * completion order. */ - list_for_each_entry_safe(active, next, &rq->active_list, link) { - /* - * In microbenchmarks or focusing upon time inside the kernel, - * we may spend an inordinate amount of time simply handling - * the retirement of requests and processing their callbacks. - * Of which, this loop itself is particularly hot due to the - * cache misses when jumping around the list of - * i915_active_request. So we try to keep this loop as - * streamlined as possible and also prefetch the next - * i915_active_request to try and hide the likely cache miss. - */ - prefetchw(next); - - INIT_LIST_HEAD(&active->link); - RCU_INIT_POINTER(active->request, NULL); - - active->retire(active, rq); - } - - local_irq_disable(); + GEM_BUG_ON(!list_is_first(&rq->link, + &i915_request_timeline(rq)->requests)); + rq->ring->head = rq->postfix; - spin_lock(&rq->engine->active.lock); - list_del(&rq->sched.link); - spin_unlock(&rq->engine->active.lock); + /* + * We only loosely track inflight requests across preemption, + * and so we may find ourselves attempting to retire a _completed_ + * request that we have removed from the HW and put back on a run + * queue. + */ + remove_from_engine(rq); - spin_lock(&rq->lock); + spin_lock_irq(&rq->lock); i915_request_mark_complete(rq); if (!i915_request_signaled(rq)) dma_fence_signal_locked(&rq->fence); if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &rq->fence.flags)) i915_request_cancel_breadcrumb(rq); - if (rq->waitboost) { - GEM_BUG_ON(!atomic_read(&rq->i915->gt_pm.rps.num_waiters)); - atomic_dec(&rq->i915->gt_pm.rps.num_waiters); + if (i915_request_has_waitboost(rq)) { + GEM_BUG_ON(!atomic_read(&rq->engine->gt->rps.num_waiters)); + atomic_dec(&rq->engine->gt->rps.num_waiters); } - spin_unlock(&rq->lock); - - local_irq_enable(); - - intel_context_exit(rq->hw_context); - intel_context_unpin(rq->hw_context); + if (!test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags)) { + set_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags); + __notify_execute_cb(rq); + } + GEM_BUG_ON(!list_empty(&rq->execute_cb)); + spin_unlock_irq(&rq->lock); - i915_request_remove_from_client(rq); + remove_from_client(rq); list_del(&rq->link); + intel_context_exit(rq->context); + intel_context_unpin(rq->context); + free_capture_list(rq); i915_sched_node_fini(&rq->sched); i915_request_put(rq); @@ -265,76 +289,24 @@ static bool i915_request_retire(struct i915_request *rq) void i915_request_retire_upto(struct i915_request *rq) { - struct intel_ring *ring = rq->ring; + struct intel_timeline * const tl = i915_request_timeline(rq); struct i915_request *tmp; - GEM_TRACE("%s fence %llx:%lld, current %d\n", - rq->engine->name, - rq->fence.context, rq->fence.seqno, - hwsp_seqno(rq)); + RQ_TRACE(rq, "\n"); - lockdep_assert_held(&rq->i915->drm.struct_mutex); GEM_BUG_ON(!i915_request_completed(rq)); - if (list_empty(&rq->ring_link)) - return; - do { - tmp = list_first_entry(&ring->request_list, - typeof(*tmp), ring_link); + tmp = list_first_entry(&tl->requests, typeof(*tmp), link); } while (i915_request_retire(tmp) && tmp != rq); } -static void irq_execute_cb(struct irq_work *wrk) -{ - struct execute_cb *cb = container_of(wrk, typeof(*cb), work); - - i915_sw_fence_complete(cb->fence); - kmem_cache_free(global.slab_execute_cbs, cb); -} - -static void irq_execute_cb_hook(struct irq_work *wrk) -{ - struct execute_cb *cb = container_of(wrk, typeof(*cb), work); - - cb->hook(container_of(cb->fence, struct i915_request, submit), - &cb->signal->fence); - i915_request_put(cb->signal); - - irq_execute_cb(wrk); -} - -static void __notify_execute_cb(struct i915_request *rq) -{ - struct execute_cb *cb; - - lockdep_assert_held(&rq->lock); - - if (list_empty(&rq->execute_cb)) - return; - - list_for_each_entry(cb, &rq->execute_cb, link) - irq_work_queue(&cb->work); - - /* - * XXX Rollback on __i915_request_unsubmit() - * - * In the future, perhaps when we have an active time-slicing scheduler, - * it will be interesting to unsubmit parallel execution and remove - * busywaits from the GPU until their master is restarted. This is - * quite hairy, we have to carefully rollback the fence and do a - * preempt-to-idle cycle on the target engine, all the while the - * master execute_cb may refire. - */ - INIT_LIST_HEAD(&rq->execute_cb); -} - static int -__i915_request_await_execution(struct i915_request *rq, - struct i915_request *signal, - void (*hook)(struct i915_request *rq, - struct dma_fence *signal), - gfp_t gfp) +__await_execution(struct i915_request *rq, + struct i915_request *signal, + void (*hook)(struct i915_request *rq, + struct dma_fence *signal), + gfp_t gfp) { struct execute_cb *cb; @@ -371,22 +343,41 @@ __i915_request_await_execution(struct i915_request *rq, } spin_unlock_irq(&signal->lock); + /* Copy across semaphore status as we need the same behaviour */ + rq->sched.flags |= signal->sched.flags; return 0; } -void __i915_request_submit(struct i915_request *request) +bool __i915_request_submit(struct i915_request *request) { struct intel_engine_cs *engine = request->engine; + bool result = false; - GEM_TRACE("%s fence %llx:%lld, current %d\n", - engine->name, - request->fence.context, request->fence.seqno, - hwsp_seqno(request)); + RQ_TRACE(request, "\n"); GEM_BUG_ON(!irqs_disabled()); lockdep_assert_held(&engine->active.lock); - if (i915_gem_context_is_banned(request->gem_context)) + /* + * With the advent of preempt-to-busy, we frequently encounter + * requests that we have unsubmitted from HW, but left running + * until the next ack and so have completed in the meantime. On + * resubmission of that completed request, we can skip + * updating the payload, and execlists can even skip submitting + * the request. + * + * We must remove the request from the caller's priority queue, + * and the caller must only call us when the request is in their + * priority queue, under the active.lock. This ensures that the + * request has *not* yet been retired and we can safely move + * the request into the engine->active.list where it will be + * dropped upon retiring. (Otherwise if resubmit a *retired* + * request, this would be a horrible use-after-free.) + */ + if (i915_request_completed(request)) + goto xfer; + + if (intel_context_is_banned(request->context)) i915_request_skip(request, -EIO); /* @@ -409,29 +400,31 @@ void __i915_request_submit(struct i915_request *request) i915_sw_fence_signaled(&request->semaphore)) engine->saturated |= request->sched.semaphores; - /* We may be recursing from the signal callback of another i915 fence */ - spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING); + engine->emit_fini_breadcrumb(request, + request->ring->vaddr + request->postfix); - list_move_tail(&request->sched.link, &engine->active.requests); + trace_i915_request_execute(request); + engine->serial++; + result = true; - GEM_BUG_ON(test_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags)); - set_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags); +xfer: /* We may be recursing from the signal callback of another i915 fence */ + spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING); + + if (!test_and_set_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags)) { + list_move_tail(&request->sched.link, &engine->active.requests); + clear_bit(I915_FENCE_FLAG_PQUEUE, &request->fence.flags); + } if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags) && !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &request->fence.flags) && !i915_request_enable_breadcrumb(request)) - intel_engine_queue_breadcrumbs(engine); + intel_engine_signal_breadcrumbs(engine); __notify_execute_cb(request); spin_unlock(&request->lock); - engine->emit_fini_breadcrumb(request, - request->ring->vaddr + request->postfix); - - engine->serial++; - - trace_i915_request_execute(request); + return result; } void i915_request_submit(struct i915_request *request) @@ -451,10 +444,7 @@ void __i915_request_unsubmit(struct i915_request *request) { struct intel_engine_cs *engine = request->engine; - GEM_TRACE("%s fence %llx:%lld, current %d\n", - engine->name, - request->fence.context, request->fence.seqno, - hwsp_seqno(request)); + RQ_TRACE(request, "\n"); GEM_BUG_ON(!irqs_disabled()); lockdep_assert_held(&engine->active.lock); @@ -512,6 +502,10 @@ submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state) switch (state) { case FENCE_COMPLETE: trace_i915_request_submit(request); + + if (unlikely(fence->error)) + i915_request_skip(request, fence->error); + /* * We need to serialize use of the submit_request() callback * with its hotplugging performed during an emergency @@ -552,29 +546,28 @@ semaphore_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state) return NOTIFY_DONE; } -static void ring_retire_requests(struct intel_ring *ring) +static void retire_requests(struct intel_timeline *tl) { struct i915_request *rq, *rn; - list_for_each_entry_safe(rq, rn, &ring->request_list, ring_link) + list_for_each_entry_safe(rq, rn, &tl->requests, link) if (!i915_request_retire(rq)) break; } static noinline struct i915_request * -request_alloc_slow(struct intel_context *ce, gfp_t gfp) +request_alloc_slow(struct intel_timeline *tl, gfp_t gfp) { - struct intel_ring *ring = ce->ring; struct i915_request *rq; - if (list_empty(&ring->request_list)) + if (list_empty(&tl->requests)) goto out; if (!gfpflags_allow_blocking(gfp)) goto out; /* Move our oldest request to the slab-cache (if not in use!) */ - rq = list_first_entry(&ring->request_list, typeof(*rq), ring_link); + rq = list_first_entry(&tl->requests, typeof(*rq), link); i915_request_retire(rq); rq = kmem_cache_alloc(global.slab_requests, @@ -583,20 +576,35 @@ request_alloc_slow(struct intel_context *ce, gfp_t gfp) return rq; /* Ratelimit ourselves to prevent oom from malicious clients */ - rq = list_last_entry(&ring->request_list, typeof(*rq), ring_link); + rq = list_last_entry(&tl->requests, typeof(*rq), link); cond_synchronize_rcu(rq->rcustate); /* Retire our old requests in the hope that we free some */ - ring_retire_requests(ring); + retire_requests(tl); out: return kmem_cache_alloc(global.slab_requests, gfp); } +static void __i915_request_ctor(void *arg) +{ + struct i915_request *rq = arg; + + spin_lock_init(&rq->lock); + i915_sched_node_init(&rq->sched); + i915_sw_fence_init(&rq->submit, submit_notify); + i915_sw_fence_init(&rq->semaphore, semaphore_notify); + + rq->file_priv = NULL; + rq->capture_list = NULL; + + INIT_LIST_HEAD(&rq->execute_cb); +} + struct i915_request * __i915_request_create(struct intel_context *ce, gfp_t gfp) { - struct i915_timeline *tl = ce->ring->timeline; + struct intel_timeline *tl = ce->timeline; struct i915_request *rq; u32 seqno; int ret; @@ -638,46 +646,43 @@ __i915_request_create(struct intel_context *ce, gfp_t gfp) rq = kmem_cache_alloc(global.slab_requests, gfp | __GFP_RETRY_MAYFAIL | __GFP_NOWARN); if (unlikely(!rq)) { - rq = request_alloc_slow(ce, gfp); + rq = request_alloc_slow(tl, gfp); if (!rq) { ret = -ENOMEM; goto err_unreserve; } } - ret = i915_timeline_get_seqno(tl, rq, &seqno); + ret = intel_timeline_get_seqno(tl, rq, &seqno); if (ret) goto err_free; rq->i915 = ce->engine->i915; - rq->hw_context = ce; - rq->gem_context = ce->gem_context; + rq->context = ce; rq->engine = ce->engine; rq->ring = ce->ring; - rq->timeline = tl; + rq->execution_mask = ce->engine->mask; + + RCU_INIT_POINTER(rq->timeline, tl); + RCU_INIT_POINTER(rq->hwsp_cacheline, tl->hwsp_cacheline); rq->hwsp_seqno = tl->hwsp_seqno; - rq->hwsp_cacheline = tl->hwsp_cacheline; + rq->rcustate = get_state_synchronize_rcu(); /* acts as smp_mb() */ - spin_lock_init(&rq->lock); dma_fence_init(&rq->fence, &i915_fence_ops, &rq->lock, tl->fence_context, seqno); /* We bump the ref for the fence chain */ - i915_sw_fence_init(&i915_request_get(rq)->submit, submit_notify); - i915_sw_fence_init(&i915_request_get(rq)->semaphore, semaphore_notify); + i915_sw_fence_reinit(&i915_request_get(rq)->submit); + i915_sw_fence_reinit(&i915_request_get(rq)->semaphore); - i915_sched_node_init(&rq->sched); + i915_sched_node_reinit(&rq->sched); - /* No zalloc, must clear what we need by hand */ - rq->file_priv = NULL; + /* No zalloc, everything must be cleared after use */ rq->batch = NULL; - rq->capture_list = NULL; - rq->waitboost = false; - rq->execution_mask = ALL_ENGINES; - - INIT_LIST_HEAD(&rq->active_list); - INIT_LIST_HEAD(&rq->execute_cb); + GEM_BUG_ON(rq->file_priv); + GEM_BUG_ON(rq->capture_list); + GEM_BUG_ON(!list_empty(&rq->execute_cb)); /* * Reserve space in the ring buffer for all the commands required to @@ -715,7 +720,6 @@ err_unwind: ce->ring->emit = rq->head; /* Make sure we didn't add ourselves to external state before freeing */ - GEM_BUG_ON(!list_empty(&rq->active_list)); GEM_BUG_ON(!list_empty(&rq->sched.signalers_list)); GEM_BUG_ON(!list_empty(&rq->sched.waiters_list)); @@ -730,15 +734,15 @@ struct i915_request * i915_request_create(struct intel_context *ce) { struct i915_request *rq; - int err; + struct intel_timeline *tl; - err = intel_context_timeline_lock(ce); - if (err) - return ERR_PTR(err); + tl = intel_context_timeline_lock(ce); + if (IS_ERR(tl)) + return ERR_CAST(tl); /* Move our oldest request to the slab-cache (if not in use!) */ - rq = list_first_entry(&ce->ring->request_list, typeof(*rq), ring_link); - if (!list_is_last(&rq->ring_link, &ce->ring->request_list)) + rq = list_first_entry(&tl->requests, typeof(*rq), link); + if (!list_is_last(&rq->link, &tl->requests)) i915_request_retire(rq); intel_context_enter(ce); @@ -748,28 +752,58 @@ i915_request_create(struct intel_context *ce) goto err_unlock; /* Check that we do not interrupt ourselves with a new request */ - rq->cookie = lockdep_pin_lock(&ce->ring->timeline->mutex); + rq->cookie = lockdep_pin_lock(&tl->mutex); return rq; err_unlock: - intel_context_timeline_unlock(ce); + intel_context_timeline_unlock(tl); return rq; } static int i915_request_await_start(struct i915_request *rq, struct i915_request *signal) { - if (list_is_first(&signal->ring_link, &signal->ring->request_list)) - return 0; + struct dma_fence *fence; + int err; + + GEM_BUG_ON(i915_request_timeline(rq) == + rcu_access_pointer(signal->timeline)); + + fence = NULL; + rcu_read_lock(); + spin_lock_irq(&signal->lock); + if (!i915_request_started(signal) && + !list_is_first(&signal->link, + &rcu_dereference(signal->timeline)->requests)) { + struct i915_request *prev = list_prev_entry(signal, link); - signal = list_prev_entry(signal, ring_link); - if (i915_timeline_sync_is_later(rq->timeline, &signal->fence)) + /* + * Peek at the request before us in the timeline. That + * request will only be valid before it is retired, so + * after acquiring a reference to it, confirm that it is + * still part of the signaler's timeline. + */ + if (i915_request_get_rcu(prev)) { + if (list_next_entry(prev, link) == signal) + fence = &prev->fence; + else + i915_request_put(prev); + } + } + spin_unlock_irq(&signal->lock); + rcu_read_unlock(); + if (!fence) return 0; - return i915_sw_fence_await_dma_fence(&rq->submit, - &signal->fence, 0, - I915_FENCE_GFP); + err = 0; + if (intel_timeline_sync_is_later(i915_request_timeline(rq), fence)) + err = i915_sw_fence_await_dma_fence(&rq->submit, + fence, 0, + I915_FENCE_GFP); + dma_fence_put(fence); + + return err; } static intel_engine_mask_t @@ -791,38 +825,27 @@ already_busywaiting(struct i915_request *rq) } static int -emit_semaphore_wait(struct i915_request *to, - struct i915_request *from, - gfp_t gfp) +__emit_semaphore_wait(struct i915_request *to, + struct i915_request *from, + u32 seqno) { + const int has_token = INTEL_GEN(to->i915) >= 12; u32 hwsp_offset; + int len, err; u32 *cs; - int err; - GEM_BUG_ON(!from->timeline->has_initial_breadcrumb); GEM_BUG_ON(INTEL_GEN(to->i915) < 8); - /* Just emit the first semaphore we see as request space is limited. */ - if (already_busywaiting(to) & from->engine->mask) - return i915_sw_fence_await_dma_fence(&to->submit, - &from->fence, 0, - I915_FENCE_GFP); - - err = i915_request_await_start(to, from); - if (err < 0) - return err; - - /* Only submit our spinner after the signaler is running! */ - err = __i915_request_await_execution(to, from, NULL, gfp); - if (err) - return err; - /* We need to pin the signaler's HWSP until we are finished reading. */ - err = i915_timeline_read_hwsp(from, to, &hwsp_offset); + err = intel_timeline_read_hwsp(from, to, &hwsp_offset); if (err) return err; - cs = intel_ring_begin(to, 4); + len = 4; + if (has_token) + len += 2; + + cs = intel_ring_begin(to, len); if (IS_ERR(cs)) return PTR_ERR(cs); @@ -834,18 +857,50 @@ emit_semaphore_wait(struct i915_request *to, * (post-wrap) values than they were expecting (and so wait * forever). */ - *cs++ = MI_SEMAPHORE_WAIT | - MI_SEMAPHORE_GLOBAL_GTT | - MI_SEMAPHORE_POLL | - MI_SEMAPHORE_SAD_GTE_SDD; - *cs++ = from->fence.seqno; + *cs++ = (MI_SEMAPHORE_WAIT | + MI_SEMAPHORE_GLOBAL_GTT | + MI_SEMAPHORE_POLL | + MI_SEMAPHORE_SAD_GTE_SDD) + + has_token; + *cs++ = seqno; *cs++ = hwsp_offset; *cs++ = 0; + if (has_token) { + *cs++ = 0; + *cs++ = MI_NOOP; + } intel_ring_advance(to, cs); + return 0; +} + +static int +emit_semaphore_wait(struct i915_request *to, + struct i915_request *from, + gfp_t gfp) +{ + /* Just emit the first semaphore we see as request space is limited. */ + if (already_busywaiting(to) & from->engine->mask) + goto await_fence; + + if (i915_request_await_start(to, from) < 0) + goto await_fence; + + /* Only submit our spinner after the signaler is running! */ + if (__await_execution(to, from, NULL, gfp)) + goto await_fence; + + if (__emit_semaphore_wait(to, from, from->fence.seqno)) + goto await_fence; + to->sched.semaphores |= from->engine->mask; to->sched.flags |= I915_SCHED_HAS_SEMAPHORE_CHAIN; return 0; + +await_fence: + return i915_sw_fence_await_dma_fence(&to->submit, + &from->fence, 0, + I915_FENCE_GFP); } static int @@ -865,18 +920,16 @@ i915_request_await_request(struct i915_request *to, struct i915_request *from) return ret; } - if (to->engine == from->engine) { + if (to->engine == from->engine) ret = i915_sw_fence_await_sw_fence_gfp(&to->submit, &from->submit, I915_FENCE_GFP); - } else if (intel_engine_has_semaphores(to->engine) && - to->gem_context->sched.priority >= I915_PRIORITY_NORMAL) { + else if (intel_context_use_semaphores(to->context)) ret = emit_semaphore_wait(to, from, I915_FENCE_GFP); - } else { + else ret = i915_sw_fence_await_dma_fence(&to->submit, &from->fence, 0, I915_FENCE_GFP); - } if (ret < 0) return ret; @@ -916,8 +969,10 @@ i915_request_await_dma_fence(struct i915_request *rq, struct dma_fence *fence) do { fence = *child++; - if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) + if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) { + i915_sw_fence_set_error_once(&rq->submit, fence->error); continue; + } /* * Requests on the same timeline are explicitly ordered, along @@ -928,27 +983,80 @@ i915_request_await_dma_fence(struct i915_request *rq, struct dma_fence *fence) continue; /* Squash repeated waits to the same timelines */ - if (fence->context != rq->i915->mm.unordered_timeline && - i915_timeline_sync_is_later(rq->timeline, fence)) + if (fence->context && + intel_timeline_sync_is_later(i915_request_timeline(rq), + fence)) continue; if (dma_fence_is_i915(fence)) ret = i915_request_await_request(rq, to_request(fence)); else ret = i915_sw_fence_await_dma_fence(&rq->submit, fence, - I915_FENCE_TIMEOUT, + fence->context ? I915_FENCE_TIMEOUT : 0, I915_FENCE_GFP); if (ret < 0) return ret; /* Record the latest fence used against each timeline */ - if (fence->context != rq->i915->mm.unordered_timeline) - i915_timeline_sync_set(rq->timeline, fence); + if (fence->context) + intel_timeline_sync_set(i915_request_timeline(rq), + fence); } while (--nchild); return 0; } +static bool intel_timeline_sync_has_start(struct intel_timeline *tl, + struct dma_fence *fence) +{ + return __intel_timeline_sync_is_later(tl, + fence->context, + fence->seqno - 1); +} + +static int intel_timeline_sync_set_start(struct intel_timeline *tl, + const struct dma_fence *fence) +{ + return __intel_timeline_sync_set(tl, fence->context, fence->seqno - 1); +} + +static int +__i915_request_await_execution(struct i915_request *to, + struct i915_request *from, + void (*hook)(struct i915_request *rq, + struct dma_fence *signal)) +{ + int err; + + /* Submit both requests at the same time */ + err = __await_execution(to, from, hook, I915_FENCE_GFP); + if (err) + return err; + + /* Squash repeated depenendices to the same timelines */ + if (intel_timeline_sync_has_start(i915_request_timeline(to), + &from->fence)) + return 0; + + /* Ensure both start together [after all semaphores in signal] */ + if (intel_engine_has_semaphores(to->engine)) + err = __emit_semaphore_wait(to, from, from->fence.seqno - 1); + else + err = i915_request_await_start(to, from); + if (err < 0) + return err; + + /* Couple the dependency tree for PI on this exposed to->fence */ + if (to->engine->schedule) { + err = i915_sched_node_add_dependency(&to->sched, &from->sched); + if (err < 0) + return err; + } + + return intel_timeline_sync_set_start(i915_request_timeline(to), + &from->fence); +} + int i915_request_await_execution(struct i915_request *rq, struct dma_fence *fence, @@ -971,8 +1079,10 @@ i915_request_await_execution(struct i915_request *rq, do { fence = *child++; - if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) + if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) { + i915_sw_fence_set_error_once(&rq->submit, fence->error); continue; + } /* * We don't squash repeated fence dependencies here as we @@ -982,8 +1092,7 @@ i915_request_await_execution(struct i915_request *rq, if (dma_fence_is_i915(fence)) ret = __i915_request_await_execution(rq, to_request(fence), - hook, - I915_FENCE_GFP); + hook); else ret = i915_sw_fence_await_dma_fence(&rq->submit, fence, I915_FENCE_TIMEOUT, @@ -1027,7 +1136,7 @@ i915_request_await_object(struct i915_request *to, struct dma_fence **shared; unsigned int count, i; - ret = reservation_object_get_fences_rcu(obj->base.resv, + ret = dma_resv_get_fences_rcu(obj->base.resv, &excl, &count, &shared); if (ret) return ret; @@ -1044,7 +1153,7 @@ i915_request_await_object(struct i915_request *to, dma_fence_put(shared[i]); kfree(shared); } else { - excl = reservation_object_get_excl_rcu(obj->base.resv); + excl = dma_resv_get_excl_rcu(obj->base.resv); } if (excl) { @@ -1065,6 +1174,9 @@ void i915_request_skip(struct i915_request *rq, int error) GEM_BUG_ON(!IS_ERR_VALUE((long)error)); dma_fence_set_error(&rq->fence, error); + if (rq->infix == rq->postfix) + return; + /* * As this request likely depends on state from the lost * context, clear out all the user operations leaving the @@ -1076,12 +1188,13 @@ void i915_request_skip(struct i915_request *rq, int error) head = 0; } memset(vaddr + head, 0, rq->postfix - head); + rq->infix = rq->postfix; } static struct i915_request * __i915_request_add_to_timeline(struct i915_request *rq) { - struct i915_timeline *timeline = rq->timeline; + struct intel_timeline *timeline = i915_request_timeline(rq); struct i915_request *prev; /* @@ -1104,7 +1217,8 @@ __i915_request_add_to_timeline(struct i915_request *rq) * precludes optimising to use semaphores serialisation of a single * timeline across engines. */ - prev = rcu_dereference_protected(timeline->last_request.request, 1); + prev = to_request(__i915_active_fence_set(&timeline->last_request, + &rq->fence)); if (prev && !i915_request_completed(prev)) { if (is_power_of_2(prev->engine->mask | rq->engine->mask)) i915_sw_fence_await_sw_fence(&rq->submit, @@ -1129,7 +1243,6 @@ __i915_request_add_to_timeline(struct i915_request *rq) * us, the timeline will hold its seqno which is later than ours. */ GEM_BUG_ON(timeline->seqno != rq->fence.seqno); - __i915_active_request_set(&timeline->last_request, rq); return prev; } @@ -1143,11 +1256,9 @@ struct i915_request *__i915_request_commit(struct i915_request *rq) { struct intel_engine_cs *engine = rq->engine; struct intel_ring *ring = rq->ring; - struct i915_request *prev; u32 *cs; - GEM_TRACE("%s fence %llx:%lld\n", - engine->name, rq->fence.context, rq->fence.seqno); + RQ_TRACE(rq, "\n"); /* * To ensure that this call will not fail, space for its emissions @@ -1156,6 +1267,7 @@ struct i915_request *__i915_request_commit(struct i915_request *rq) */ GEM_BUG_ON(rq->reserved_space > ring->space); rq->reserved_space = 0; + rq->emitted_jiffies = jiffies; /* * Record the position of the start of the breadcrumb so that @@ -1167,13 +1279,12 @@ struct i915_request *__i915_request_commit(struct i915_request *rq) GEM_BUG_ON(IS_ERR(cs)); rq->postfix = intel_ring_offset(rq, cs); - prev = __i915_request_add_to_timeline(rq); - - list_add_tail(&rq->ring_link, &ring->request_list); - if (list_is_first(&rq->ring_link, &ring->request_list)) - list_add(&ring->active_link, &rq->i915->gt.active_rings); - rq->emitted_jiffies = jiffies; + return __i915_request_add_to_timeline(rq); +} +void __i915_request_queue(struct i915_request *rq, + const struct i915_sched_attr *attr) +{ /* * Let the backend know a new request has arrived that may need * to adjust the existing execution schedule due to a high priority @@ -1185,56 +1296,56 @@ struct i915_request *__i915_request_commit(struct i915_request *rq) * decide whether to preempt the entire chain so that it is ready to * run at the earliest possible convenience. */ - local_bh_disable(); i915_sw_fence_commit(&rq->semaphore); - rcu_read_lock(); /* RCU serialisation for set-wedged protection */ - if (engine->schedule) { - struct i915_sched_attr attr = rq->gem_context->sched; - - /* - * Boost actual workloads past semaphores! - * - * With semaphores we spin on one engine waiting for another, - * simply to reduce the latency of starting our work when - * the signaler completes. However, if there is any other - * work that we could be doing on this engine instead, that - * is better utilisation and will reduce the overall duration - * of the current work. To avoid PI boosting a semaphore - * far in the distance past over useful work, we keep a history - * of any semaphore use along our dependency chain. - */ - if (!(rq->sched.flags & I915_SCHED_HAS_SEMAPHORE_CHAIN)) - attr.priority |= I915_PRIORITY_NOSEMAPHORE; - - /* - * Boost priorities to new clients (new request flows). - * - * Allow interactive/synchronous clients to jump ahead of - * the bulk clients. (FQ_CODEL) - */ - if (list_empty(&rq->sched.signalers_list)) - attr.priority |= I915_PRIORITY_WAIT; - - engine->schedule(rq, &attr); - } - rcu_read_unlock(); + if (attr && rq->engine->schedule) + rq->engine->schedule(rq, attr); i915_sw_fence_commit(&rq->submit); - local_bh_enable(); /* Kick the execlists tasklet if just scheduled */ - - return prev; } void i915_request_add(struct i915_request *rq) { + struct intel_timeline * const tl = i915_request_timeline(rq); + struct i915_sched_attr attr = {}; struct i915_request *prev; - lockdep_assert_held(&rq->timeline->mutex); - lockdep_unpin_lock(&rq->timeline->mutex, rq->cookie); + lockdep_assert_held(&tl->mutex); + lockdep_unpin_lock(&tl->mutex, rq->cookie); trace_i915_request_add(rq); prev = __i915_request_commit(rq); + if (rcu_access_pointer(rq->context->gem_context)) + attr = i915_request_gem_context(rq)->sched; + + /* + * Boost actual workloads past semaphores! + * + * With semaphores we spin on one engine waiting for another, + * simply to reduce the latency of starting our work when + * the signaler completes. However, if there is any other + * work that we could be doing on this engine instead, that + * is better utilisation and will reduce the overall duration + * of the current work. To avoid PI boosting a semaphore + * far in the distance past over useful work, we keep a history + * of any semaphore use along our dependency chain. + */ + if (!(rq->sched.flags & I915_SCHED_HAS_SEMAPHORE_CHAIN)) + attr.priority |= I915_PRIORITY_NOSEMAPHORE; + + /* + * Boost priorities to new clients (new request flows). + * + * Allow interactive/synchronous clients to jump ahead of + * the bulk clients. (FQ_CODEL) + */ + if (list_empty(&rq->sched.signalers_list)) + attr.priority |= I915_PRIORITY_WAIT; + + local_bh_disable(); + __i915_request_queue(rq, &attr); + local_bh_enable(); /* Kick the execlists tasklet if just scheduled */ + /* * In typical scenarios, we do not expect the previous request on * the timeline to be still tracked by timeline->last_request if it @@ -1252,10 +1363,12 @@ void i915_request_add(struct i915_request *rq) * work on behalf of others -- but instead we should benefit from * improved resource management. (Well, that's the theory at least.) */ - if (prev && i915_request_completed(prev)) + if (prev && + i915_request_completed(prev) && + rcu_access_pointer(prev->timeline) == tl) i915_request_retire_upto(prev); - mutex_unlock(&rq->timeline->mutex); + mutex_unlock(&tl->mutex); } static unsigned long local_clock_us(unsigned int *cpu) @@ -1390,8 +1503,7 @@ long i915_request_wait(struct i915_request *rq, * serialise wait/reset with an explicit lock, we do want * lockdep to detect potential dependency cycles. */ - mutex_acquire(&rq->i915->gpu_error.wedge_mutex.dep_map, - 0, 0, _THIS_IP_); + mutex_acquire(&rq->engine->gt->reset.mutex.dep_map, 0, 0, _THIS_IP_); /* * Optimistic spin before touching IRQs. @@ -1416,7 +1528,7 @@ long i915_request_wait(struct i915_request *rq, * completion. That requires having a good predictor for the request * duration, which we currently lack. */ - if (CONFIG_DRM_I915_SPIN_REQUEST && + if (IS_ACTIVE(CONFIG_DRM_I915_SPIN_REQUEST) && __i915_spin_request(rq, state, CONFIG_DRM_I915_SPIN_REQUEST)) { dma_fence_signal(&rq->fence); goto out; @@ -1436,7 +1548,7 @@ long i915_request_wait(struct i915_request *rq, */ if (flags & I915_WAIT_PRIORITY) { if (!i915_request_started(rq) && INTEL_GEN(rq->i915) >= 6) - gen6_rps_boost(rq); + intel_rps_boost(rq); i915_schedule_bump_priority(rq, I915_PRIORITY_WAIT); } @@ -1447,8 +1559,10 @@ long i915_request_wait(struct i915_request *rq, for (;;) { set_current_state(state); - if (i915_request_completed(rq)) + if (i915_request_completed(rq)) { + dma_fence_signal(&rq->fence); break; + } if (signal_pending_state(state, current)) { timeout = -ERESTARTSYS; @@ -1460,6 +1574,7 @@ long i915_request_wait(struct i915_request *rq, break; } + intel_engine_flush_submission(rq->engine); timeout = io_schedule_timeout(timeout); } __set_current_state(TASK_RUNNING); @@ -1467,27 +1582,11 @@ long i915_request_wait(struct i915_request *rq, dma_fence_remove_callback(&rq->fence, &wait.cb); out: - mutex_release(&rq->i915->gpu_error.wedge_mutex.dep_map, 0, _THIS_IP_); + mutex_release(&rq->engine->gt->reset.mutex.dep_map, _THIS_IP_); trace_i915_request_wait_end(rq); return timeout; } -bool i915_retire_requests(struct drm_i915_private *i915) -{ - struct intel_ring *ring, *tmp; - - lockdep_assert_held(&i915->drm.struct_mutex); - - list_for_each_entry_safe(ring, tmp, - &i915->gt.active_rings, active_link) { - intel_ring_get(ring); /* last rq holds reference! */ - ring_retire_requests(ring); - intel_ring_put(ring); - } - - return !list_empty(&i915->gt.active_rings); -} - #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) #include "selftests/mock_request.c" #include "selftests/i915_request.c" @@ -1514,10 +1613,14 @@ static struct i915_global_request global = { { int __init i915_global_request_init(void) { - global.slab_requests = KMEM_CACHE(i915_request, - SLAB_HWCACHE_ALIGN | - SLAB_RECLAIM_ACCOUNT | - SLAB_TYPESAFE_BY_RCU); + global.slab_requests = + kmem_cache_create("i915_request", + sizeof(struct i915_request), + __alignof__(struct i915_request), + SLAB_HWCACHE_ALIGN | + SLAB_RECLAIM_ACCOUNT | + SLAB_TYPESAFE_BY_RCU, + __i915_request_ctor); if (!global.slab_requests) return -ENOMEM; |
