drm/i915: Replace engine->timeline with a plain list
authorChris Wilson <chris@chris-wilson.co.uk>
Fri, 14 Jun 2019 16:46:06 +0000 (17:46 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Fri, 14 Jun 2019 18:03:40 +0000 (19:03 +0100)
To continue the onslaught of removing the assumption of a global
execution ordering, another casualty is the engine->timeline. Without an
actual timeline to track, it is overkill and we can replace it with a
much less grand plain list. We still need a list of requests inflight,
for the simple purpose of finding inflight requests (for retiring,
resetting, preemption etc).

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190614164606.15633-3-chris@chris-wilson.co.uk
16 files changed:
drivers/gpu/drm/i915/gt/intel_engine.h
drivers/gpu/drm/i915/gt/intel_engine_cs.c
drivers/gpu/drm/i915/gt/intel_engine_types.h
drivers/gpu/drm/i915/gt/intel_lrc.c
drivers/gpu/drm/i915/gt/intel_reset.c
drivers/gpu/drm/i915/gt/intel_ringbuffer.c
drivers/gpu/drm/i915/gt/mock_engine.c
drivers/gpu/drm/i915/i915_gpu_error.c
drivers/gpu/drm/i915/i915_request.c
drivers/gpu/drm/i915/i915_request.h
drivers/gpu/drm/i915/i915_scheduler.c
drivers/gpu/drm/i915/i915_timeline.c
drivers/gpu/drm/i915/i915_timeline.h
drivers/gpu/drm/i915/i915_timeline_types.h
drivers/gpu/drm/i915/intel_guc_submission.c
drivers/gpu/drm/i915/selftests/mock_timeline.c

index 3b5a6d1..2f1c687 100644 (file)
@@ -565,4 +565,10 @@ static inline bool inject_preempt_hang(struct intel_engine_execlists *execlists)
 
 #endif
 
+void intel_engine_init_active(struct intel_engine_cs *engine,
+                             unsigned int subclass);
+#define ENGINE_PHYSICAL        0
+#define ENGINE_MOCK    1
+#define ENGINE_VIRTUAL 2
+
 #endif /* _INTEL_RINGBUFFER_H_ */
index 22242e9..8986929 100644 (file)
@@ -617,14 +617,7 @@ static int intel_engine_setup_common(struct intel_engine_cs *engine)
        if (err)
                return err;
 
-       err = i915_timeline_init(engine->i915,
-                                &engine->timeline,
-                                engine->status_page.vma);
-       if (err)
-               goto err_hwsp;
-
-       i915_timeline_set_subclass(&engine->timeline, TIMELINE_ENGINE);
-
+       intel_engine_init_active(engine, ENGINE_PHYSICAL);
        intel_engine_init_breadcrumbs(engine);
        intel_engine_init_execlists(engine);
        intel_engine_init_hangcheck(engine);
@@ -637,10 +630,6 @@ static int intel_engine_setup_common(struct intel_engine_cs *engine)
                intel_sseu_from_device_info(&RUNTIME_INFO(engine->i915)->sseu);
 
        return 0;
-
-err_hwsp:
-       cleanup_status_page(engine);
-       return err;
 }
 
 /**
@@ -797,6 +786,27 @@ static int pin_context(struct i915_gem_context *ctx,
        return 0;
 }
 
+void
+intel_engine_init_active(struct intel_engine_cs *engine, unsigned int subclass)
+{
+       INIT_LIST_HEAD(&engine->active.requests);
+
+       spin_lock_init(&engine->active.lock);
+       lockdep_set_subclass(&engine->active.lock, subclass);
+
+       /*
+        * Due to an interesting quirk in lockdep's internal debug tracking,
+        * after setting a subclass we must ensure the lock is used. Otherwise,
+        * nr_unused_locks is incremented once too often.
+        */
+#ifdef CONFIG_DEBUG_LOCK_ALLOC
+       local_irq_disable();
+       lock_map_acquire(&engine->active.lock.dep_map);
+       lock_map_release(&engine->active.lock.dep_map);
+       local_irq_enable();
+#endif
+}
+
 /**
  * intel_engines_init_common - initialize cengine state which might require hw access
  * @engine: Engine to initialize.
@@ -860,6 +870,8 @@ err_unpin:
  */
 void intel_engine_cleanup_common(struct intel_engine_cs *engine)
 {
+       GEM_BUG_ON(!list_empty(&engine->active.requests));
+
        cleanup_status_page(engine);
 
        intel_engine_fini_breadcrumbs(engine);
@@ -874,8 +886,6 @@ void intel_engine_cleanup_common(struct intel_engine_cs *engine)
        intel_context_unpin(engine->kernel_context);
        GEM_BUG_ON(!llist_empty(&engine->barrier_tasks));
 
-       i915_timeline_fini(&engine->timeline);
-
        intel_wa_list_free(&engine->ctx_wa_list);
        intel_wa_list_free(&engine->wa_list);
        intel_wa_list_free(&engine->whitelist);
@@ -1482,16 +1492,6 @@ void intel_engine_dump(struct intel_engine_cs *engine,
 
        drm_printf(m, "\tRequests:\n");
 
-       rq = list_first_entry(&engine->timeline.requests,
-                             struct i915_request, link);
-       if (&rq->link != &engine->timeline.requests)
-               print_request(m, rq, "\t\tfirst  ");
-
-       rq = list_last_entry(&engine->timeline.requests,
-                            struct i915_request, link);
-       if (&rq->link != &engine->timeline.requests)
-               print_request(m, rq, "\t\tlast   ");
-
        rq = intel_engine_find_active_request(engine);
        if (rq) {
                print_request(m, rq, "\t\tactive ");
@@ -1572,7 +1572,7 @@ int intel_enable_engine_stats(struct intel_engine_cs *engine)
        if (!intel_engine_supports_stats(engine))
                return -ENODEV;
 
-       spin_lock_irqsave(&engine->timeline.lock, flags);
+       spin_lock_irqsave(&engine->active.lock, flags);
        write_seqlock(&engine->stats.lock);
 
        if (unlikely(engine->stats.enabled == ~0)) {
@@ -1598,7 +1598,7 @@ int intel_enable_engine_stats(struct intel_engine_cs *engine)
 
 unlock:
        write_sequnlock(&engine->stats.lock);
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 
        return err;
 }
@@ -1683,22 +1683,22 @@ intel_engine_find_active_request(struct intel_engine_cs *engine)
         * At all other times, we must assume the GPU is still running, but
         * we only care about the snapshot of this moment.
         */
-       spin_lock_irqsave(&engine->timeline.lock, flags);
-       list_for_each_entry(request, &engine->timeline.requests, link) {
+       spin_lock_irqsave(&engine->active.lock, flags);
+       list_for_each_entry(request, &engine->active.requests, sched.link) {
                if (i915_request_completed(request))
                        continue;
 
                if (!i915_request_started(request))
-                       break;
+                       continue;
 
                /* More than one preemptible request may match! */
                if (!match_ring(request))
-                       break;
+                       continue;
 
                active = request;
                break;
        }
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 
        return active;
 }
index 33a31aa..b2faca8 100644 (file)
@@ -288,7 +288,11 @@ struct intel_engine_cs {
 
        struct intel_ring *buffer;
 
-       struct i915_timeline timeline;
+       struct {
+               spinlock_t lock;
+               struct list_head requests;
+       } active;
+
        struct llist_head barrier_tasks;
 
        struct intel_context *kernel_context; /* pinned */
index d0a5175..c400c66 100644 (file)
@@ -298,8 +298,8 @@ static inline bool need_preempt(const struct intel_engine_cs *engine,
         * Check against the first request in ELSP[1], it will, thanks to the
         * power of PI, be the highest priority of that context.
         */
-       if (!list_is_last(&rq->link, &engine->timeline.requests) &&
-           rq_prio(list_next_entry(rq, link)) > last_prio)
+       if (!list_is_last(&rq->sched.link, &engine->active.requests) &&
+           rq_prio(list_next_entry(rq, sched.link)) > last_prio)
                return true;
 
        if (rb) {
@@ -434,11 +434,11 @@ __unwind_incomplete_requests(struct intel_engine_cs *engine)
        struct list_head *uninitialized_var(pl);
        int prio = I915_PRIORITY_INVALID;
 
-       lockdep_assert_held(&engine->timeline.lock);
+       lockdep_assert_held(&engine->active.lock);
 
        list_for_each_entry_safe_reverse(rq, rn,
-                                        &engine->timeline.requests,
-                                        link) {
+                                        &engine->active.requests,
+                                        sched.link) {
                struct intel_engine_cs *owner;
 
                if (i915_request_completed(rq))
@@ -465,7 +465,7 @@ __unwind_incomplete_requests(struct intel_engine_cs *engine)
                        }
                        GEM_BUG_ON(RB_EMPTY_ROOT(&engine->execlists.queue.rb_root));
 
-                       list_add(&rq->sched.link, pl);
+                       list_move(&rq->sched.link, pl);
                        active = rq;
                } else {
                        rq->engine = owner;
@@ -933,11 +933,11 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
                        rb_entry(rb, typeof(*ve), nodes[engine->id].rb);
                struct i915_request *rq;
 
-               spin_lock(&ve->base.timeline.lock);
+               spin_lock(&ve->base.active.lock);
 
                rq = ve->request;
                if (unlikely(!rq)) { /* lost the race to a sibling */
-                       spin_unlock(&ve->base.timeline.lock);
+                       spin_unlock(&ve->base.active.lock);
                        rb_erase_cached(rb, &execlists->virtual);
                        RB_CLEAR_NODE(rb);
                        rb = rb_first_cached(&execlists->virtual);
@@ -950,13 +950,13 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 
                if (rq_prio(rq) >= queue_prio(execlists)) {
                        if (!virtual_matches(ve, rq, engine)) {
-                               spin_unlock(&ve->base.timeline.lock);
+                               spin_unlock(&ve->base.active.lock);
                                rb = rb_next(rb);
                                continue;
                        }
 
                        if (last && !can_merge_rq(last, rq)) {
-                               spin_unlock(&ve->base.timeline.lock);
+                               spin_unlock(&ve->base.active.lock);
                                return; /* leave this rq for another engine */
                        }
 
@@ -1011,7 +1011,7 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
                        last = rq;
                }
 
-               spin_unlock(&ve->base.timeline.lock);
+               spin_unlock(&ve->base.active.lock);
                break;
        }
 
@@ -1068,8 +1068,6 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
                                GEM_BUG_ON(port_isset(port));
                        }
 
-                       list_del_init(&rq->sched.link);
-
                        __i915_request_submit(rq);
                        trace_i915_request_in(rq, port_index(port, execlists));
 
@@ -1170,7 +1168,7 @@ static void process_csb(struct intel_engine_cs *engine)
        const u8 num_entries = execlists->csb_size;
        u8 head, tail;
 
-       lockdep_assert_held(&engine->timeline.lock);
+       lockdep_assert_held(&engine->active.lock);
 
        /*
         * Note that csb_write, csb_status may be either in HWSP or mmio.
@@ -1330,7 +1328,7 @@ static void process_csb(struct intel_engine_cs *engine)
 
 static void __execlists_submission_tasklet(struct intel_engine_cs *const engine)
 {
-       lockdep_assert_held(&engine->timeline.lock);
+       lockdep_assert_held(&engine->active.lock);
 
        process_csb(engine);
        if (!execlists_is_active(&engine->execlists, EXECLISTS_ACTIVE_PREEMPT))
@@ -1351,15 +1349,16 @@ static void execlists_submission_tasklet(unsigned long data)
                  !!intel_wakeref_active(&engine->wakeref),
                  engine->execlists.active);
 
-       spin_lock_irqsave(&engine->timeline.lock, flags);
+       spin_lock_irqsave(&engine->active.lock, flags);
        __execlists_submission_tasklet(engine);
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 }
 
 static void queue_request(struct intel_engine_cs *engine,
                          struct i915_sched_node *node,
                          int prio)
 {
+       GEM_BUG_ON(!list_empty(&node->link));
        list_add_tail(&node->link, i915_sched_lookup_priolist(engine, prio));
 }
 
@@ -1390,7 +1389,7 @@ static void execlists_submit_request(struct i915_request *request)
        unsigned long flags;
 
        /* Will be called from irq-context when using foreign fences. */
-       spin_lock_irqsave(&engine->timeline.lock, flags);
+       spin_lock_irqsave(&engine->active.lock, flags);
 
        queue_request(engine, &request->sched, rq_prio(request));
 
@@ -1399,7 +1398,7 @@ static void execlists_submit_request(struct i915_request *request)
 
        submit_queue(engine, rq_prio(request));
 
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 }
 
 static void __execlists_context_fini(struct intel_context *ce)
@@ -2050,8 +2049,8 @@ static void execlists_reset_prepare(struct intel_engine_cs *engine)
        intel_engine_stop_cs(engine);
 
        /* And flush any current direct submission. */
-       spin_lock_irqsave(&engine->timeline.lock, flags);
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+       spin_lock_irqsave(&engine->active.lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 }
 
 static bool lrc_regs_ok(const struct i915_request *rq)
@@ -2094,11 +2093,11 @@ static void reset_csb_pointers(struct intel_engine_execlists *execlists)
 
 static struct i915_request *active_request(struct i915_request *rq)
 {
-       const struct list_head * const list = &rq->engine->timeline.requests;
+       const struct list_head * const list = &rq->engine->active.requests;
        const struct intel_context * const context = rq->hw_context;
        struct i915_request *active = NULL;
 
-       list_for_each_entry_from_reverse(rq, list, link) {
+       list_for_each_entry_from_reverse(rq, list, sched.link) {
                if (i915_request_completed(rq))
                        break;
 
@@ -2215,11 +2214,11 @@ static void execlists_reset(struct intel_engine_cs *engine, bool stalled)
 
        GEM_TRACE("%s\n", engine->name);
 
-       spin_lock_irqsave(&engine->timeline.lock, flags);
+       spin_lock_irqsave(&engine->active.lock, flags);
 
        __execlists_reset(engine, stalled);
 
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 }
 
 static void nop_submission_tasklet(unsigned long data)
@@ -2250,12 +2249,12 @@ static void execlists_cancel_requests(struct intel_engine_cs *engine)
         * submission's irq state, we also wish to remind ourselves that
         * it is irq state.)
         */
-       spin_lock_irqsave(&engine->timeline.lock, flags);
+       spin_lock_irqsave(&engine->active.lock, flags);
 
        __execlists_reset(engine, true);
 
        /* Mark all executing requests as skipped. */
-       list_for_each_entry(rq, &engine->timeline.requests, link) {
+       list_for_each_entry(rq, &engine->active.requests, sched.link) {
                if (!i915_request_signaled(rq))
                        dma_fence_set_error(&rq->fence, -EIO);
 
@@ -2286,7 +2285,7 @@ static void execlists_cancel_requests(struct intel_engine_cs *engine)
                rb_erase_cached(rb, &execlists->virtual);
                RB_CLEAR_NODE(rb);
 
-               spin_lock(&ve->base.timeline.lock);
+               spin_lock(&ve->base.active.lock);
                if (ve->request) {
                        ve->request->engine = engine;
                        __i915_request_submit(ve->request);
@@ -2295,7 +2294,7 @@ static void execlists_cancel_requests(struct intel_engine_cs *engine)
                        ve->base.execlists.queue_priority_hint = INT_MIN;
                        ve->request = NULL;
                }
-               spin_unlock(&ve->base.timeline.lock);
+               spin_unlock(&ve->base.active.lock);
        }
 
        /* Remaining _unready_ requests will be nop'ed when submitted */
@@ -2307,7 +2306,7 @@ static void execlists_cancel_requests(struct intel_engine_cs *engine)
        GEM_BUG_ON(__tasklet_is_enabled(&execlists->tasklet));
        execlists->tasklet.func = nop_submission_tasklet;
 
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 }
 
 static void execlists_reset_finish(struct intel_engine_cs *engine)
@@ -3010,12 +3009,18 @@ error_deref_obj:
        return ret;
 }
 
+static struct list_head *virtual_queue(struct virtual_engine *ve)
+{
+       return &ve->base.execlists.default_priolist.requests[0];
+}
+
 static void virtual_context_destroy(struct kref *kref)
 {
        struct virtual_engine *ve =
                container_of(kref, typeof(*ve), context.ref);
        unsigned int n;
 
+       GEM_BUG_ON(!list_empty(virtual_queue(ve)));
        GEM_BUG_ON(ve->request);
        GEM_BUG_ON(ve->context.inflight);
 
@@ -3026,13 +3031,13 @@ static void virtual_context_destroy(struct kref *kref)
                if (RB_EMPTY_NODE(node))
                        continue;
 
-               spin_lock_irq(&sibling->timeline.lock);
+               spin_lock_irq(&sibling->active.lock);
 
                /* Detachment is lazily performed in the execlists tasklet */
                if (!RB_EMPTY_NODE(node))
                        rb_erase_cached(node, &sibling->execlists.virtual);
 
-               spin_unlock_irq(&sibling->timeline.lock);
+               spin_unlock_irq(&sibling->active.lock);
        }
        GEM_BUG_ON(__tasklet_is_scheduled(&ve->base.execlists.tasklet));
 
@@ -3040,8 +3045,6 @@ static void virtual_context_destroy(struct kref *kref)
                __execlists_context_fini(&ve->context);
 
        kfree(ve->bonds);
-
-       i915_timeline_fini(&ve->base.timeline);
        kfree(ve);
 }
 
@@ -3161,16 +3164,16 @@ static void virtual_submission_tasklet(unsigned long data)
 
                if (unlikely(!(mask & sibling->mask))) {
                        if (!RB_EMPTY_NODE(&node->rb)) {
-                               spin_lock(&sibling->timeline.lock);
+                               spin_lock(&sibling->active.lock);
                                rb_erase_cached(&node->rb,
                                                &sibling->execlists.virtual);
                                RB_CLEAR_NODE(&node->rb);
-                               spin_unlock(&sibling->timeline.lock);
+                               spin_unlock(&sibling->active.lock);
                        }
                        continue;
                }
 
-               spin_lock(&sibling->timeline.lock);
+               spin_lock(&sibling->active.lock);
 
                if (!RB_EMPTY_NODE(&node->rb)) {
                        /*
@@ -3214,7 +3217,7 @@ submit_engine:
                        tasklet_hi_schedule(&sibling->execlists.tasklet);
                }
 
-               spin_unlock(&sibling->timeline.lock);
+               spin_unlock(&sibling->active.lock);
        }
        local_irq_enable();
 }
@@ -3231,9 +3234,13 @@ static void virtual_submit_request(struct i915_request *rq)
        GEM_BUG_ON(ve->base.submit_request != virtual_submit_request);
 
        GEM_BUG_ON(ve->request);
+       GEM_BUG_ON(!list_empty(virtual_queue(ve)));
+
        ve->base.execlists.queue_priority_hint = rq_prio(rq);
        WRITE_ONCE(ve->request, rq);
 
+       list_move_tail(&rq->sched.link, virtual_queue(ve));
+
        tasklet_schedule(&ve->base.execlists.tasklet);
 }
 
@@ -3297,10 +3304,7 @@ intel_execlists_create_virtual(struct i915_gem_context *ctx,
 
        snprintf(ve->base.name, sizeof(ve->base.name), "virtual");
 
-       err = i915_timeline_init(ctx->i915, &ve->base.timeline, NULL);
-       if (err)
-               goto err_put;
-       i915_timeline_set_subclass(&ve->base.timeline, TIMELINE_VIRTUAL);
+       intel_engine_init_active(&ve->base, ENGINE_VIRTUAL);
 
        intel_engine_init_execlists(&ve->base);
 
@@ -3311,6 +3315,7 @@ intel_execlists_create_virtual(struct i915_gem_context *ctx,
        ve->base.submit_request = virtual_submit_request;
        ve->base.bond_execute = virtual_bond_execute;
 
+       INIT_LIST_HEAD(virtual_queue(ve));
        ve->base.execlists.queue_priority_hint = INT_MIN;
        tasklet_init(&ve->base.execlists.tasklet,
                     virtual_submission_tasklet,
@@ -3465,11 +3470,11 @@ void intel_execlists_show_requests(struct intel_engine_cs *engine,
        unsigned int count;
        struct rb_node *rb;
 
-       spin_lock_irqsave(&engine->timeline.lock, flags);
+       spin_lock_irqsave(&engine->active.lock, flags);
 
        last = NULL;
        count = 0;
-       list_for_each_entry(rq, &engine->timeline.requests, link) {
+       list_for_each_entry(rq, &engine->active.requests, sched.link) {
                if (count++ < max - 1)
                        show_request(m, rq, "\t\tE ");
                else
@@ -3532,7 +3537,7 @@ void intel_execlists_show_requests(struct intel_engine_cs *engine,
                show_request(m, last, "\t\tV ");
        }
 
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 }
 
 void intel_lr_context_reset(struct intel_engine_cs *engine,
index 6e6807a..84c670b 100644 (file)
@@ -49,12 +49,12 @@ static void engine_skip_context(struct i915_request *rq)
        struct intel_engine_cs *engine = rq->engine;
        struct i915_gem_context *hung_ctx = rq->gem_context;
 
-       lockdep_assert_held(&engine->timeline.lock);
+       lockdep_assert_held(&engine->active.lock);
 
        if (!i915_request_is_active(rq))
                return;
 
-       list_for_each_entry_continue(rq, &engine->timeline.requests, link)
+       list_for_each_entry_continue(rq, &engine->active.requests, sched.link)
                if (rq->gem_context == hung_ctx)
                        i915_request_skip(rq, -EIO);
 }
@@ -130,7 +130,7 @@ void i915_reset_request(struct i915_request *rq, bool guilty)
                  rq->fence.seqno,
                  yesno(guilty));
 
-       lockdep_assert_held(&rq->engine->timeline.lock);
+       lockdep_assert_held(&rq->engine->active.lock);
        GEM_BUG_ON(i915_request_completed(rq));
 
        if (guilty) {
@@ -785,10 +785,10 @@ static void nop_submit_request(struct i915_request *request)
                  engine->name, request->fence.context, request->fence.seqno);
        dma_fence_set_error(&request->fence, -EIO);
 
-       spin_lock_irqsave(&engine->timeline.lock, flags);
+       spin_lock_irqsave(&engine->active.lock, flags);
        __i915_request_submit(request);
        i915_request_mark_complete(request);
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 
        intel_engine_queue_breadcrumbs(engine);
 }
index cc901ed..019bf03 100644 (file)
@@ -730,14 +730,13 @@ static void reset_prepare(struct intel_engine_cs *engine)
 
 static void reset_ring(struct intel_engine_cs *engine, bool stalled)
 {
-       struct i915_timeline *tl = &engine->timeline;
        struct i915_request *pos, *rq;
        unsigned long flags;
        u32 head;
 
        rq = NULL;
-       spin_lock_irqsave(&tl->lock, flags);
-       list_for_each_entry(pos, &tl->requests, link) {
+       spin_lock_irqsave(&engine->active.lock, flags);
+       list_for_each_entry(pos, &engine->active.requests, sched.link) {
                if (!i915_request_completed(pos)) {
                        rq = pos;
                        break;
@@ -791,7 +790,7 @@ static void reset_ring(struct intel_engine_cs *engine, bool stalled)
        }
        engine->buffer->head = intel_ring_wrap(engine->buffer, head);
 
-       spin_unlock_irqrestore(&tl->lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 }
 
 static void reset_finish(struct intel_engine_cs *engine)
@@ -877,10 +876,10 @@ static void cancel_requests(struct intel_engine_cs *engine)
        struct i915_request *request;
        unsigned long flags;
 
-       spin_lock_irqsave(&engine->timeline.lock, flags);
+       spin_lock_irqsave(&engine->active.lock, flags);
 
        /* Mark all submitted requests as skipped. */
-       list_for_each_entry(request, &engine->timeline.requests, link) {
+       list_for_each_entry(request, &engine->active.requests, sched.link) {
                if (!i915_request_signaled(request))
                        dma_fence_set_error(&request->fence, -EIO);
 
@@ -889,7 +888,7 @@ static void cancel_requests(struct intel_engine_cs *engine)
 
        /* Remaining _unready_ requests will be nop'ed when submitted */
 
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 }
 
 static void i9xx_submit_request(struct i915_request *request)
@@ -1267,8 +1266,6 @@ intel_engine_create_ring(struct intel_engine_cs *engine,
 
        GEM_BUG_ON(!is_power_of_2(size));
        GEM_BUG_ON(RING_CTL_SIZE(size) & ~RING_NR_PAGES);
-       GEM_BUG_ON(timeline == &engine->timeline);
-       lockdep_assert_held(&engine->i915->drm.struct_mutex);
 
        ring = kzalloc(sizeof(*ring), GFP_KERNEL);
        if (!ring)
index d1ef515..086801b 100644 (file)
@@ -229,17 +229,17 @@ static void mock_cancel_requests(struct intel_engine_cs *engine)
        struct i915_request *request;
        unsigned long flags;
 
-       spin_lock_irqsave(&engine->timeline.lock, flags);
+       spin_lock_irqsave(&engine->active.lock, flags);
 
        /* Mark all submitted requests as skipped. */
-       list_for_each_entry(request, &engine->timeline.requests, sched.link) {
+       list_for_each_entry(request, &engine->active.requests, sched.link) {
                if (!i915_request_signaled(request))
                        dma_fence_set_error(&request->fence, -EIO);
 
                i915_request_mark_complete(request);
        }
 
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 }
 
 struct intel_engine_cs *mock_engine(struct drm_i915_private *i915,
@@ -285,28 +285,23 @@ int mock_engine_init(struct intel_engine_cs *engine)
        struct drm_i915_private *i915 = engine->i915;
        int err;
 
+       intel_engine_init_active(engine, ENGINE_MOCK);
        intel_engine_init_breadcrumbs(engine);
        intel_engine_init_execlists(engine);
        intel_engine_init__pm(engine);
 
-       if (i915_timeline_init(i915, &engine->timeline, NULL))
-               goto err_breadcrumbs;
-       i915_timeline_set_subclass(&engine->timeline, TIMELINE_ENGINE);
-
        engine->kernel_context =
                i915_gem_context_get_engine(i915->kernel_context, engine->id);
        if (IS_ERR(engine->kernel_context))
-               goto err_timeline;
+               goto err_breadcrumbs;
 
        err = intel_context_pin(engine->kernel_context);
        intel_context_put(engine->kernel_context);
        if (err)
-               goto err_timeline;
+               goto err_breadcrumbs;
 
        return 0;
 
-err_timeline:
-       i915_timeline_fini(&engine->timeline);
 err_breadcrumbs:
        intel_engine_fini_breadcrumbs(engine);
        return -ENOMEM;
@@ -340,7 +335,6 @@ void mock_engine_free(struct intel_engine_cs *engine)
        intel_context_unpin(engine->kernel_context);
 
        intel_engine_fini_breadcrumbs(engine);
-       i915_timeline_fini(&engine->timeline);
 
        kfree(engine);
 }
index 26c9c05..f411e32 100644 (file)
@@ -1275,7 +1275,7 @@ static void engine_record_requests(struct intel_engine_cs *engine,
 
        count = 0;
        request = first;
-       list_for_each_entry_from(request, &engine->timeline.requests, link)
+       list_for_each_entry_from(request, &engine->active.requests, sched.link)
                count++;
        if (!count)
                return;
@@ -1288,7 +1288,8 @@ static void engine_record_requests(struct intel_engine_cs *engine,
 
        count = 0;
        request = first;
-       list_for_each_entry_from(request, &engine->timeline.requests, link) {
+       list_for_each_entry_from(request,
+                                &engine->active.requests, sched.link) {
                if (count >= ee->num_requests) {
                        /*
                         * If the ring request list was changed in
index c99136f..9819483 100644 (file)
@@ -232,9 +232,9 @@ static bool i915_request_retire(struct i915_request *rq)
 
        local_irq_disable();
 
-       spin_lock(&rq->engine->timeline.lock);
-       list_del(&rq->link);
-       spin_unlock(&rq->engine->timeline.lock);
+       spin_lock(&rq->engine->active.lock);
+       list_del(&rq->sched.link);
+       spin_unlock(&rq->engine->active.lock);
 
        spin_lock(&rq->lock);
        i915_request_mark_complete(rq);
@@ -254,6 +254,7 @@ static bool i915_request_retire(struct i915_request *rq)
        intel_context_unpin(rq->hw_context);
 
        i915_request_remove_from_client(rq);
+       list_del(&rq->link);
 
        free_capture_list(rq);
        i915_sched_node_fini(&rq->sched);
@@ -373,28 +374,17 @@ __i915_request_await_execution(struct i915_request *rq,
        return 0;
 }
 
-static void move_to_timeline(struct i915_request *request,
-                            struct i915_timeline *timeline)
-{
-       GEM_BUG_ON(request->timeline == &request->engine->timeline);
-       lockdep_assert_held(&request->engine->timeline.lock);
-
-       spin_lock(&request->timeline->lock);
-       list_move_tail(&request->link, &timeline->requests);
-       spin_unlock(&request->timeline->lock);
-}
-
 void __i915_request_submit(struct i915_request *request)
 {
        struct intel_engine_cs *engine = request->engine;
 
-       GEM_TRACE("%s fence %llx:%lld -> current %d\n",
+       GEM_TRACE("%s fence %llx:%lld, current %d\n",
                  engine->name,
                  request->fence.context, request->fence.seqno,
                  hwsp_seqno(request));
 
        GEM_BUG_ON(!irqs_disabled());
-       lockdep_assert_held(&engine->timeline.lock);
+       lockdep_assert_held(&engine->active.lock);
 
        if (i915_gem_context_is_banned(request->gem_context))
                i915_request_skip(request, -EIO);
@@ -422,6 +412,8 @@ void __i915_request_submit(struct i915_request *request)
        /* We may be recursing from the signal callback of another i915 fence */
        spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING);
 
+       list_move_tail(&request->sched.link, &engine->active.requests);
+
        GEM_BUG_ON(test_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags));
        set_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags);
 
@@ -437,9 +429,6 @@ void __i915_request_submit(struct i915_request *request)
        engine->emit_fini_breadcrumb(request,
                                     request->ring->vaddr + request->postfix);
 
-       /* Transfer from per-context onto the global per-engine timeline */
-       move_to_timeline(request, &engine->timeline);
-
        engine->serial++;
 
        trace_i915_request_execute(request);
@@ -451,11 +440,11 @@ void i915_request_submit(struct i915_request *request)
        unsigned long flags;
 
        /* Will be called from irq-context when using foreign fences. */
-       spin_lock_irqsave(&engine->timeline.lock, flags);
+       spin_lock_irqsave(&engine->active.lock, flags);
 
        __i915_request_submit(request);
 
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 }
 
 void __i915_request_unsubmit(struct i915_request *request)
@@ -468,7 +457,7 @@ void __i915_request_unsubmit(struct i915_request *request)
                  hwsp_seqno(request));
 
        GEM_BUG_ON(!irqs_disabled());
-       lockdep_assert_held(&engine->timeline.lock);
+       lockdep_assert_held(&engine->active.lock);
 
        /*
         * Only unwind in reverse order, required so that the per-context list
@@ -486,9 +475,6 @@ void __i915_request_unsubmit(struct i915_request *request)
 
        spin_unlock(&request->lock);
 
-       /* Transfer back from the global per-engine timeline to per-context */
-       move_to_timeline(request, request->timeline);
-
        /* We've already spun, don't charge on resubmitting. */
        if (request->sched.semaphores && i915_request_started(request)) {
                request->sched.attr.priority |= I915_PRIORITY_NOSEMAPHORE;
@@ -510,11 +496,11 @@ void i915_request_unsubmit(struct i915_request *request)
        unsigned long flags;
 
        /* Will be called from irq-context when using foreign fences. */
-       spin_lock_irqsave(&engine->timeline.lock, flags);
+       spin_lock_irqsave(&engine->active.lock, flags);
 
        __i915_request_unsubmit(request);
 
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 }
 
 static int __i915_sw_fence_call
@@ -669,7 +655,6 @@ __i915_request_create(struct intel_context *ce, gfp_t gfp)
        rq->engine = ce->engine;
        rq->ring = ce->ring;
        rq->timeline = tl;
-       GEM_BUG_ON(rq->timeline == &ce->engine->timeline);
        rq->hwsp_seqno = tl->hwsp_seqno;
        rq->hwsp_cacheline = tl->hwsp_cacheline;
        rq->rcustate = get_state_synchronize_rcu(); /* acts as smp_mb() */
@@ -1136,9 +1121,7 @@ __i915_request_add_to_timeline(struct i915_request *rq)
                                                         0);
        }
 
-       spin_lock_irq(&timeline->lock);
        list_add_tail(&rq->link, &timeline->requests);
-       spin_unlock_irq(&timeline->lock);
 
        /*
         * Make sure that no request gazumped us - if it was allocated after
index c9f7d07..edbbdfe 100644 (file)
@@ -217,7 +217,7 @@ struct i915_request {
 
        bool waitboost;
 
-       /** engine->request_list entry for this request */
+       /** timeline->request entry for this request */
        struct list_head link;
 
        /** ring->request_list entry for this request */
index 78ceb56..2e9b38b 100644 (file)
@@ -77,7 +77,7 @@ i915_sched_lookup_priolist(struct intel_engine_cs *engine, int prio)
        bool first = true;
        int idx, i;
 
-       lockdep_assert_held(&engine->timeline.lock);
+       lockdep_assert_held(&engine->active.lock);
        assert_priolists(execlists);
 
        /* buckets sorted from highest [in slot 0] to lowest priority */
@@ -162,9 +162,9 @@ sched_lock_engine(const struct i915_sched_node *node,
         * check that the rq still belongs to the newly locked engine.
         */
        while (locked != (engine = READ_ONCE(rq->engine))) {
-               spin_unlock(&locked->timeline.lock);
+               spin_unlock(&locked->active.lock);
                memset(cache, 0, sizeof(*cache));
-               spin_lock(&engine->timeline.lock);
+               spin_lock(&engine->active.lock);
                locked = engine;
        }
 
@@ -189,7 +189,7 @@ static void kick_submission(struct intel_engine_cs *engine, int prio)
         * tasklet, i.e. we have not change the priority queue
         * sufficiently to oust the running context.
         */
-       if (inflight && !i915_scheduler_need_preempt(prio, rq_prio(inflight)))
+       if (!inflight || !i915_scheduler_need_preempt(prio, rq_prio(inflight)))
                return;
 
        tasklet_hi_schedule(&engine->execlists.tasklet);
@@ -278,7 +278,7 @@ static void __i915_schedule(struct i915_sched_node *node,
 
        memset(&cache, 0, sizeof(cache));
        engine = node_to_request(node)->engine;
-       spin_lock(&engine->timeline.lock);
+       spin_lock(&engine->active.lock);
 
        /* Fifo and depth-first replacement ensure our deps execute before us */
        engine = sched_lock_engine(node, engine, &cache);
@@ -287,7 +287,7 @@ static void __i915_schedule(struct i915_sched_node *node,
 
                node = dep->signaler;
                engine = sched_lock_engine(node, engine, &cache);
-               lockdep_assert_held(&engine->timeline.lock);
+               lockdep_assert_held(&engine->active.lock);
 
                /* Recheck after acquiring the engine->timeline.lock */
                if (prio <= node->attr.priority || node_signaled(node))
@@ -296,14 +296,8 @@ static void __i915_schedule(struct i915_sched_node *node,
                GEM_BUG_ON(node_to_request(node)->engine != engine);
 
                node->attr.priority = prio;
-               if (!list_empty(&node->link)) {
-                       GEM_BUG_ON(intel_engine_is_virtual(engine));
-                       if (!cache.priolist)
-                               cache.priolist =
-                                       i915_sched_lookup_priolist(engine,
-                                                                  prio);
-                       list_move_tail(&node->link, cache.priolist);
-               } else {
+
+               if (list_empty(&node->link)) {
                        /*
                         * If the request is not in the priolist queue because
                         * it is not yet runnable, then it doesn't contribute
@@ -312,8 +306,16 @@ static void __i915_schedule(struct i915_sched_node *node,
                         * queue; but in that case we may still need to reorder
                         * the inflight requests.
                         */
-                       if (!i915_sw_fence_done(&node_to_request(node)->submit))
-                               continue;
+                       continue;
+               }
+
+               if (!intel_engine_is_virtual(engine) &&
+                   !i915_request_is_active(node_to_request(node))) {
+                       if (!cache.priolist)
+                               cache.priolist =
+                                       i915_sched_lookup_priolist(engine,
+                                                                  prio);
+                       list_move_tail(&node->link, cache.priolist);
                }
 
                if (prio <= engine->execlists.queue_priority_hint)
@@ -325,7 +327,7 @@ static void __i915_schedule(struct i915_sched_node *node,
                kick_submission(engine, prio);
        }
 
-       spin_unlock(&engine->timeline.lock);
+       spin_unlock(&engine->active.lock);
 }
 
 void i915_schedule(struct i915_request *rq, const struct i915_sched_attr *attr)
@@ -439,8 +441,6 @@ void i915_sched_node_fini(struct i915_sched_node *node)
 {
        struct i915_dependency *dep, *tmp;
 
-       GEM_BUG_ON(!list_empty(&node->link));
-
        spin_lock_irq(&schedule_lock);
 
        /*
index 000e1a9..c311ce9 100644 (file)
@@ -251,7 +251,6 @@ int i915_timeline_init(struct drm_i915_private *i915,
 
        timeline->fence_context = dma_fence_context_alloc(1);
 
-       spin_lock_init(&timeline->lock);
        mutex_init(&timeline->mutex);
 
        INIT_ACTIVE_REQUEST(&timeline->last_request);
index 27668a1..36e5e5a 100644 (file)
@@ -36,25 +36,6 @@ int i915_timeline_init(struct drm_i915_private *i915,
                       struct i915_vma *hwsp);
 void i915_timeline_fini(struct i915_timeline *tl);
 
-static inline void
-i915_timeline_set_subclass(struct i915_timeline *timeline,
-                          unsigned int subclass)
-{
-       lockdep_set_subclass(&timeline->lock, subclass);
-
-       /*
-        * Due to an interesting quirk in lockdep's internal debug tracking,
-        * after setting a subclass we must ensure the lock is used. Otherwise,
-        * nr_unused_locks is incremented once too often.
-        */
-#ifdef CONFIG_DEBUG_LOCK_ALLOC
-       local_irq_disable();
-       lock_map_acquire(&timeline->lock.dep_map);
-       lock_map_release(&timeline->lock.dep_map);
-       local_irq_enable();
-#endif
-}
-
 struct i915_timeline *
 i915_timeline_create(struct drm_i915_private *i915,
                     struct i915_vma *global_hwsp);
index 1688705..fce5cb4 100644 (file)
@@ -23,10 +23,6 @@ struct i915_timeline {
        u64 fence_context;
        u32 seqno;
 
-       spinlock_t lock;
-#define TIMELINE_CLIENT 0 /* default subclass */
-#define TIMELINE_ENGINE 1
-#define TIMELINE_VIRTUAL 2
        struct mutex mutex; /* protects the flow of requests */
 
        unsigned int pin_count;
index 97f6970..db531eb 100644 (file)
@@ -740,7 +740,7 @@ static bool __guc_dequeue(struct intel_engine_cs *engine)
        bool submit = false;
        struct rb_node *rb;
 
-       lockdep_assert_held(&engine->timeline.lock);
+       lockdep_assert_held(&engine->active.lock);
 
        if (port_isset(port)) {
                if (intel_engine_has_preemption(engine)) {
@@ -822,7 +822,7 @@ static void guc_submission_tasklet(unsigned long data)
        struct i915_request *rq;
        unsigned long flags;
 
-       spin_lock_irqsave(&engine->timeline.lock, flags);
+       spin_lock_irqsave(&engine->active.lock, flags);
 
        rq = port_request(port);
        while (rq && i915_request_completed(rq)) {
@@ -847,7 +847,7 @@ static void guc_submission_tasklet(unsigned long data)
        if (!execlists_is_active(execlists, EXECLISTS_ACTIVE_PREEMPT))
                guc_dequeue(engine);
 
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 }
 
 static void guc_reset_prepare(struct intel_engine_cs *engine)
@@ -884,7 +884,7 @@ static void guc_reset(struct intel_engine_cs *engine, bool stalled)
        struct i915_request *rq;
        unsigned long flags;
 
-       spin_lock_irqsave(&engine->timeline.lock, flags);
+       spin_lock_irqsave(&engine->active.lock, flags);
 
        execlists_cancel_port_requests(execlists);
 
@@ -900,7 +900,7 @@ static void guc_reset(struct intel_engine_cs *engine, bool stalled)
        intel_lr_context_reset(engine, rq->hw_context, rq->head, stalled);
 
 out_unlock:
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 }
 
 static void guc_cancel_requests(struct intel_engine_cs *engine)
@@ -926,13 +926,13 @@ static void guc_cancel_requests(struct intel_engine_cs *engine)
         * submission's irq state, we also wish to remind ourselves that
         * it is irq state.)
         */
-       spin_lock_irqsave(&engine->timeline.lock, flags);
+       spin_lock_irqsave(&engine->active.lock, flags);
 
        /* Cancel the requests on the HW and clear the ELSP tracker. */
        execlists_cancel_port_requests(execlists);
 
        /* Mark all executing requests as skipped. */
-       list_for_each_entry(rq, &engine->timeline.requests, link) {
+       list_for_each_entry(rq, &engine->active.requests, sched.link) {
                if (!i915_request_signaled(rq))
                        dma_fence_set_error(&rq->fence, -EIO);
 
@@ -961,7 +961,7 @@ static void guc_cancel_requests(struct intel_engine_cs *engine)
        execlists->queue = RB_ROOT_CACHED;
        GEM_BUG_ON(port_isset(execlists->port));
 
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+       spin_unlock_irqrestore(&engine->active.lock, flags);
 }
 
 static void guc_reset_finish(struct intel_engine_cs *engine)
index e084476..65b52be 100644 (file)
@@ -13,7 +13,6 @@ void mock_timeline_init(struct i915_timeline *timeline, u64 context)
        timeline->i915 = NULL;
        timeline->fence_context = context;
 
-       spin_lock_init(&timeline->lock);
        mutex_init(&timeline->mutex);
 
        INIT_ACTIVE_REQUEST(&timeline->last_request);