diff options
Diffstat (limited to 'drivers/gpu/drm/i915/i915_request.c')
| -rw-r--r-- | drivers/gpu/drm/i915/i915_request.c | 234 |
1 files changed, 141 insertions, 93 deletions
diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c index e5a55801f753..c0df71d7d0ff 100644 --- a/drivers/gpu/drm/i915/i915_request.c +++ b/drivers/gpu/drm/i915/i915_request.c @@ -290,7 +290,7 @@ bool i915_request_retire(struct i915_request *rq) spin_unlock_irq(&rq->lock); remove_from_client(rq); - list_del_rcu(&rq->link); + __list_del_entry(&rq->link); /* poison neither prev/next (RCU walks) */ intel_context_exit(rq->context); intel_context_unpin(rq->context); @@ -363,6 +363,50 @@ __await_execution(struct i915_request *rq, return 0; } +static bool fatal_error(int error) +{ + switch (error) { + case 0: /* not an error! */ + case -EAGAIN: /* innocent victim of a GT reset (__i915_request_reset) */ + case -ETIMEDOUT: /* waiting for Godot (timer_i915_sw_fence_wake) */ + return false; + default: + return true; + } +} + +void __i915_request_skip(struct i915_request *rq) +{ + GEM_BUG_ON(!fatal_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 + * breadcrumb at the end (so we get the fence notifications). + */ + __i915_request_fill(rq, 0); + rq->infix = rq->postfix; +} + +void i915_request_set_error_once(struct i915_request *rq, int error) +{ + int old; + + GEM_BUG_ON(!IS_ERR_VALUE((long)error)); + + if (i915_request_signaled(rq)) + return; + + old = READ_ONCE(rq->fence.error); + do { + if (fatal_error(old)) + return; + } while (!try_cmpxchg(&rq->fence.error, &old, error)); +} + bool __i915_request_submit(struct i915_request *request) { struct intel_engine_cs *engine = request->engine; @@ -392,8 +436,10 @@ bool __i915_request_submit(struct i915_request *request) if (i915_request_completed(request)) goto xfer; - if (intel_context_is_banned(request->context)) - i915_request_skip(request, -EIO); + if (unlikely(intel_context_is_banned(request->context))) + i915_request_set_error_once(request, -EIO); + if (unlikely(fatal_error(request->fence.error))) + __i915_request_skip(request); /* * Are we using semaphores when the gpu is already saturated? @@ -519,7 +565,7 @@ submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state) trace_i915_request_submit(request); if (unlikely(fence->error)) - i915_request_skip(request, fence->error); + i915_request_set_error_once(request, fence->error); /* * We need to serialize use of the submit_request() callback @@ -542,19 +588,31 @@ submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state) return NOTIFY_DONE; } +static void irq_semaphore_cb(struct irq_work *wrk) +{ + struct i915_request *rq = + container_of(wrk, typeof(*rq), semaphore_work); + + i915_schedule_bump_priority(rq, I915_PRIORITY_NOSEMAPHORE); + i915_request_put(rq); +} + static int __i915_sw_fence_call semaphore_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state) { - struct i915_request *request = - container_of(fence, typeof(*request), semaphore); + struct i915_request *rq = container_of(fence, typeof(*rq), semaphore); switch (state) { case FENCE_COMPLETE: - i915_schedule_bump_priority(request, I915_PRIORITY_NOSEMAPHORE); + if (!(READ_ONCE(rq->sched.attr.priority) & I915_PRIORITY_NOSEMAPHORE)) { + i915_request_get(rq); + init_irq_work(&rq->semaphore_work, irq_semaphore_cb); + irq_work_queue(&rq->semaphore_work); + } break; case FENCE_FREE: - i915_request_put(request); + i915_request_put(rq); break; } @@ -691,6 +749,7 @@ __i915_request_create(struct intel_context *ce, gfp_t gfp) RCU_INIT_POINTER(rq->timeline, tl); RCU_INIT_POINTER(rq->hwsp_cacheline, tl->hwsp_cacheline); rq->hwsp_seqno = tl->hwsp_seqno; + GEM_BUG_ON(i915_request_completed(rq)); rq->rcustate = get_state_synchronize_rcu(); /* acts as smp_mb() */ @@ -791,8 +850,8 @@ i915_request_await_start(struct i915_request *rq, struct i915_request *signal) struct dma_fence *fence; int err; - GEM_BUG_ON(i915_request_timeline(rq) == - rcu_access_pointer(signal->timeline)); + if (i915_request_timeline(rq) == rcu_access_pointer(signal->timeline)) + return 0; if (i915_request_started(signal)) return 0; @@ -836,7 +895,7 @@ i915_request_await_start(struct i915_request *rq, struct i915_request *signal) return 0; err = 0; - if (intel_timeline_sync_is_later(i915_request_timeline(rq), fence)) + 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); @@ -860,7 +919,7 @@ already_busywaiting(struct i915_request *rq) * * See the are-we-too-late? check in __i915_request_submit(). */ - return rq->sched.semaphores | rq->engine->saturated; + return rq->sched.semaphores | READ_ONCE(rq->engine->saturated); } static int @@ -918,6 +977,8 @@ emit_semaphore_wait(struct i915_request *to, struct i915_request *from, gfp_t gfp) { + const intel_engine_mask_t mask = READ_ONCE(from->engine)->mask; + if (!intel_context_use_semaphores(to->context)) goto await_fence; @@ -925,7 +986,7 @@ emit_semaphore_wait(struct i915_request *to, goto await_fence; /* Just emit the first semaphore we see as request space is limited. */ - if (already_busywaiting(to) & from->engine->mask) + if (already_busywaiting(to) & mask) goto await_fence; if (i915_request_await_start(to, from) < 0) @@ -938,7 +999,7 @@ emit_semaphore_wait(struct i915_request *to, if (__emit_semaphore_wait(to, from, from->fence.seqno)) goto await_fence; - to->sched.semaphores |= from->engine->mask; + to->sched.semaphores |= mask; to->sched.flags |= I915_SCHED_HAS_SEMAPHORE_CHAIN; return 0; @@ -1081,14 +1142,45 @@ __i915_request_await_execution(struct i915_request *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); + /* + * Wait until the start of this request. + * + * The execution cb fires when we submit the request to HW. But in + * many cases this may be long before the request itself is ready to + * run (consider that we submit 2 requests for the same context, where + * the request of interest is behind an indefinite spinner). So we hook + * up to both to reduce our queues and keep the execution lag minimised + * in the worst case, though we hope that the await_start is elided. + */ + err = i915_request_await_start(to, from); if (err < 0) return err; + /* + * Ensure both start together [after all semaphores in signal] + * + * Now that we are queued to the HW at roughly the same time (thanks + * to the execute cb) and are ready to run at roughly the same time + * (thanks to the await start), our signaler may still be indefinitely + * delayed by waiting on a semaphore from a remote engine. If our + * signaler depends on a semaphore, so indirectly do we, and we do not + * want to start our payload until our signaler also starts theirs. + * So we wait. + * + * However, there is also a second condition for which we need to wait + * for the precise start of the signaler. Consider that the signaler + * was submitted in a chain of requests following another context + * (with just an ordinary intra-engine fence dependency between the + * two). In this case the signaler is queued to HW, but not for + * immediate execution, and so we must wait until it reaches the + * active slot. + */ + if (intel_engine_has_semaphores(to->engine)) { + err = __emit_semaphore_wait(to, from, from->fence.seqno - 1); + 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); @@ -1209,23 +1301,6 @@ i915_request_await_object(struct i915_request *to, return ret; } -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 - * breadcrumb at the end (so we get the fence notifications). - */ - __i915_request_fill(rq, 0); - rq->infix = rq->postfix; -} - static struct i915_request * __i915_request_add_to_timeline(struct i915_request *rq) { @@ -1255,7 +1330,17 @@ __i915_request_add_to_timeline(struct i915_request *rq) 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)) + /* + * The requests are supposed to be kept in order. However, + * we need to be wary in case the timeline->last_request + * is used as a barrier for external modification to this + * context. + */ + GEM_BUG_ON(prev->context == rq->context && + i915_seqno_passed(prev->fence.seqno, + rq->fence.seqno)); + + if (is_power_of_2(READ_ONCE(prev->engine)->mask | rq->engine->mask)) i915_sw_fence_await_sw_fence(&rq->submit, &prev->submit, &rq->submitq); @@ -1329,9 +1414,9 @@ void __i915_request_queue(struct i915_request *rq, * decide whether to preempt the entire chain so that it is ready to * run at the earliest possible convenience. */ - i915_sw_fence_commit(&rq->semaphore); if (attr && rq->engine->schedule) rq->engine->schedule(rq, attr); + i915_sw_fence_commit(&rq->semaphore); i915_sw_fence_commit(&rq->submit); } @@ -1339,39 +1424,23 @@ 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; + struct i915_gem_context *ctx; lockdep_assert_held(&tl->mutex); lockdep_unpin_lock(&tl->mutex, rq->cookie); trace_i915_request_add(rq); + __i915_request_commit(rq); - prev = __i915_request_commit(rq); - - if (rcu_access_pointer(rq->context->gem_context)) - attr = i915_request_gem_context(rq)->sched; + /* XXX placeholder for selftests */ + rcu_read_lock(); + ctx = rcu_dereference(rq->context->gem_context); + if (ctx) + attr = ctx->sched; + rcu_read_unlock(); - /* - * 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; @@ -1379,32 +1448,10 @@ void i915_request_add(struct i915_request *rq) __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 - * has been completed. If the completed request is still here, that - * implies that request retirement is a long way behind submission, - * suggesting that we haven't been retiring frequently enough from - * the combination of retire-before-alloc, waiters and the background - * retirement worker. So if the last request on this timeline was - * already completed, do a catch up pass, flushing the retirement queue - * up to this client. Since we have now moved the heaviest operations - * during retirement onto secondary workers, such as freeing objects - * or contexts, retiring a bunch of requests is mostly list management - * (and cache misses), and so we should not be overly penalizing this - * client by performing excess work, though we may still performing - * 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) && - rcu_access_pointer(prev->timeline) == tl) - i915_request_retire_upto(prev); - mutex_unlock(&tl->mutex); } -static unsigned long local_clock_us(unsigned int *cpu) +static unsigned long local_clock_ns(unsigned int *cpu) { unsigned long t; @@ -1421,7 +1468,7 @@ static unsigned long local_clock_us(unsigned int *cpu) * stop busywaiting, see busywait_stop(). */ *cpu = get_cpu(); - t = local_clock() >> 10; + t = local_clock(); put_cpu(); return t; @@ -1431,15 +1478,15 @@ static bool busywait_stop(unsigned long timeout, unsigned int cpu) { unsigned int this_cpu; - if (time_after(local_clock_us(&this_cpu), timeout)) + if (time_after(local_clock_ns(&this_cpu), timeout)) return true; return this_cpu != cpu; } -static bool __i915_spin_request(const struct i915_request * const rq, - int state, unsigned long timeout_us) +static bool __i915_spin_request(const struct i915_request * const rq, int state) { + unsigned long timeout_ns; unsigned int cpu; /* @@ -1467,7 +1514,8 @@ static bool __i915_spin_request(const struct i915_request * const rq, * takes to sleep on a request, on the order of a microsecond. */ - timeout_us += local_clock_us(&cpu); + timeout_ns = READ_ONCE(rq->engine->props.max_busywait_duration_ns); + timeout_ns += local_clock_ns(&cpu); do { if (i915_request_completed(rq)) return true; @@ -1475,7 +1523,7 @@ static bool __i915_spin_request(const struct i915_request * const rq, if (signal_pending_state(state, current)) break; - if (busywait_stop(timeout_us, cpu)) + if (busywait_stop(timeout_ns, cpu)) break; cpu_relax(); @@ -1561,8 +1609,8 @@ long i915_request_wait(struct i915_request *rq, * completion. That requires having a good predictor for the request * duration, which we currently lack. */ - if (IS_ACTIVE(CONFIG_DRM_I915_SPIN_REQUEST) && - __i915_spin_request(rq, state, CONFIG_DRM_I915_SPIN_REQUEST)) { + if (IS_ACTIVE(CONFIG_DRM_I915_MAX_REQUEST_BUSYWAIT) && + __i915_spin_request(rq, state)) { dma_fence_signal(&rq->fence); goto out; } |