[11/28] drm/i915: Move list of timelines under its own lock

Message ID	20190128010245.20148-11-chris@chris-wilson.co.uk (mailing list archive)
State	New, archived
Headers	show Return-Path: <intel-gfx-bounces@lists.freedesktop.org> From: Chris Wilson <chris@chris-wilson.co.uk> To: intel-gfx@lists.freedesktop.org Date: Mon, 28 Jan 2019 01:02:28 +0000 Message-Id: <20190128010245.20148-11-chris@chris-wilson.co.uk> In-Reply-To: <20190128010245.20148-1-chris@chris-wilson.co.uk> References: <20190128010245.20148-1-chris@chris-wilson.co.uk> MIME-Version: 1.0 Subject: [Intel-gfx] [PATCH 11/28] drm/i915: Move list of timelines under its own lock Precedence: list Content-Type: text/plain; charset="utf-8" Content-Transfer-Encoding: base64 Errors-To: intel-gfx-bounces@lists.freedesktop.org Sender: "Intel-gfx" <intel-gfx-bounces@lists.freedesktop.org>
Series	[01/28] drm/i915: Wait for a moment before forcibly resetting the device \| expand [01/28] drm/i915: Wait for a moment before forcibly resetting the device [02/28] drm/i915: Rename execlists->queue_priority to preempt_priority_hint [03/28] drm/i915/execlists: Suppress preempting self [04/28] drm/i915/execlists: Suppress redundant preemption [05/28] drm/i915/selftests: Exercise some AB...BA preemption chains [06/28] drm/i915: Stop tracking MRU activity on VMA [07/28] drm/i915: Pull VM lists under the VM mutex. [08/28] drm/i915: Move vma lookup to its own lock [09/28] drm/i915: Always allocate an object/vma for the HWSP [10/28] drm/i915: Add timeline barrier support [11/28] drm/i915: Move list of timelines under its own lock [12/28] drm/i915: Introduce concept of per-timeline (context) HWSP [13/28] drm/i915: Enlarge vma->pin_count [14/28] drm/i915: Allocate a status page for each timeline [15/28] drm/i915: Share per-timeline HWSP using a slab suballocator [16/28] drm/i915: Track the context's seqno in its own timeline HWSP [17/28] drm/i915: Track active timelines [18/28] drm/i915: Identify active requests [19/28] drm/i915: Remove the intel_engine_notify tracepoint [20/28] drm/i915: Replace global breadcrumbs with per-context interrupt tracking [21/28] drm/i915: Drop fake breadcrumb irq [22/28] drm/i915: Generalise GPU activity tracking [23/28] drm/i915: Allocate active tracking nodes from a slabcache [24/28] drm/i915: Pull i915_gem_active into the i915_active family [25/28] drm/i915: Keep timeline HWSP allocated until the system is idle [26/28] drm/i915/execlists: Refactor out can_merge_rq() [27/28] drm/i915: Use HW semaphores for inter-engine synchronisation on gen8+ [28/28] drm/i915: Prioritise non-busywait semaphore workloads

diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index 0133d1da3d3c..8a181b455197 100644 --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h @@ -1975,7 +1975,10 @@ struct drm_i915_private { void (*resume)(struct drm_i915_private *); void (*cleanup_engine)(struct intel_engine_cs *engine); - struct list_head timelines; + struct i915_gt_timelines { + struct mutex mutex; /* protects list, tainted by GPU */ + struct list_head list; + } timelines; struct list_head active_rings; struct list_head closed_vma; diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index 653c7ba4c69f..d68f3fdd8a8e 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -3224,33 +3224,6 @@ i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file) return ret; } -static long wait_for_timeline(struct i915_timeline *tl, - unsigned int flags, long timeout) -{ - struct i915_request *rq; - - rq = i915_gem_active_get_unlocked(&tl->last_request); - if (!rq) - return timeout; - - /* - * "Race-to-idle". - * - * Switching to the kernel context is often used a synchronous - * step prior to idling, e.g. in suspend for flushing all - * current operations to memory before sleeping. These we - * want to complete as quickly as possible to avoid prolonged - * stalls, so allow the gpu to boost to maximum clocks. - */ - if (flags & I915_WAIT_FOR_IDLE_BOOST) - gen6_rps_boost(rq, NULL); - - timeout = i915_request_wait(rq, flags, timeout); - i915_request_put(rq); - - return timeout; -} - static int wait_for_engines(struct drm_i915_private *i915) { if (wait_for(intel_engines_are_idle(i915), I915_IDLE_ENGINES_TIMEOUT)) { @@ -3264,6 +3237,52 @@ static int wait_for_engines(struct drm_i915_private *i915) return 0; } +static long +wait_for_timelines(struct drm_i915_private *i915, + unsigned int flags, long timeout) +{ + struct i915_gt_timelines *gt = &i915->gt.timelines; + struct i915_timeline *tl; + + if (!READ_ONCE(i915->gt.active_requests)) + return timeout; + + mutex_lock(&gt->mutex); + list_for_each_entry(tl, &gt->list, link) { + struct i915_request *rq; + + rq = i915_gem_active_get_unlocked(&tl->last_request); + if (!rq) + continue; + + mutex_unlock(&gt->mutex); + + /* + * "Race-to-idle". + * + * Switching to the kernel context is often used a synchronous + * step prior to idling, e.g. in suspend for flushing all + * current operations to memory before sleeping. These we + * want to complete as quickly as possible to avoid prolonged + * stalls, so allow the gpu to boost to maximum clocks. + */ + if (flags & I915_WAIT_FOR_IDLE_BOOST) + gen6_rps_boost(rq, NULL); + + timeout = i915_request_wait(rq, flags, timeout); + i915_request_put(rq); + if (timeout < 0) + return timeout; + + /* restart after reacquiring the lock */ + mutex_lock(&gt->mutex); + tl = list_entry(&gt->list, typeof(*tl), link); + } + mutex_unlock(&gt->mutex); + + return timeout; +} + int i915_gem_wait_for_idle(struct drm_i915_private *i915, unsigned int flags, long timeout) { @@ -3275,17 +3294,15 @@ int i915_gem_wait_for_idle(struct drm_i915_private *i915, if (!READ_ONCE(i915->gt.awake)) return 0; + timeout = wait_for_timelines(i915, flags, timeout); + if (timeout < 0) + return timeout; + if (flags & I915_WAIT_LOCKED) { - struct i915_timeline *tl; int err; lockdep_assert_held(&i915->drm.struct_mutex); - list_for_each_entry(tl, &i915->gt.timelines, link) { - timeout = wait_for_timeline(tl, flags, timeout); - if (timeout < 0) - return timeout; - } if (GEM_SHOW_DEBUG() && !timeout) { /* Presume that timeout was non-zero to begin with! */ dev_warn(&i915->drm.pdev->dev, @@ -3299,17 +3316,6 @@ int i915_gem_wait_for_idle(struct drm_i915_private *i915, i915_retire_requests(i915); GEM_BUG_ON(i915->gt.active_requests); - } else { - struct intel_engine_cs *engine; - enum intel_engine_id id; - - for_each_engine(engine, i915, id) { - struct i915_timeline *tl = &engine->timeline; - - timeout = wait_for_timeline(tl, flags, timeout); - if (timeout < 0) - return timeout; - } } return 0; @@ -5010,6 +5016,8 @@ int i915_gem_init(struct drm_i915_private *dev_priv) dev_priv->gt.cleanup_engine = intel_engine_cleanup; } + i915_timelines_init(dev_priv); + ret = i915_gem_init_userptr(dev_priv); if (ret) return ret; @@ -5132,8 +5140,10 @@ int i915_gem_init(struct drm_i915_private *dev_priv) err_uc_misc: intel_uc_fini_misc(dev_priv); - if (ret != -EIO) + if (ret != -EIO) { i915_gem_cleanup_userptr(dev_priv); + i915_timelines_fini(dev_priv); + } if (ret == -EIO) { mutex_lock(&dev_priv->drm.struct_mutex); @@ -5184,6 +5194,7 @@ void i915_gem_fini(struct drm_i915_private *dev_priv) intel_uc_fini_misc(dev_priv); i915_gem_cleanup_userptr(dev_priv); + i915_timelines_fini(dev_priv); i915_gem_drain_freed_objects(dev_priv); @@ -5286,7 +5297,6 @@ int i915_gem_init_early(struct drm_i915_private *dev_priv) if (!dev_priv->priorities) goto err_dependencies; - INIT_LIST_HEAD(&dev_priv->gt.timelines); INIT_LIST_HEAD(&dev_priv->gt.active_rings); INIT_LIST_HEAD(&dev_priv->gt.closed_vma); @@ -5330,7 +5340,6 @@ void i915_gem_cleanup_early(struct drm_i915_private *dev_priv) GEM_BUG_ON(!llist_empty(&dev_priv->mm.free_list)); GEM_BUG_ON(atomic_read(&dev_priv->mm.free_count)); WARN_ON(dev_priv->mm.object_count); - WARN_ON(!list_empty(&dev_priv->gt.timelines)); kmem_cache_destroy(dev_priv->priorities); kmem_cache_destroy(dev_priv->dependencies); diff --git a/drivers/gpu/drm/i915/i915_reset.c b/drivers/gpu/drm/i915/i915_reset.c index 99bd3bc336b3..d2dca85a543d 100644 --- a/drivers/gpu/drm/i915/i915_reset.c +++ b/drivers/gpu/drm/i915/i915_reset.c @@ -854,7 +854,8 @@ bool i915_gem_unset_wedged(struct drm_i915_private *i915) * * No more can be submitted until we reset the wedged bit. */ - list_for_each_entry(tl, &i915->gt.timelines, link) { + mutex_lock(&i915->gt.timelines.mutex); + list_for_each_entry(tl, &i915->gt.timelines.list, link) { struct i915_request *rq; long timeout; @@ -876,9 +877,12 @@ bool i915_gem_unset_wedged(struct drm_i915_private *i915) timeout = dma_fence_default_wait(&rq->fence, true, MAX_SCHEDULE_TIMEOUT); i915_request_put(rq); - if (timeout < 0) + if (timeout < 0) { + mutex_unlock(&i915->gt.timelines.mutex); goto unlock; + } } + mutex_unlock(&i915->gt.timelines.mutex); intel_engines_sanitize(i915, false); diff --git a/drivers/gpu/drm/i915/i915_timeline.c b/drivers/gpu/drm/i915/i915_timeline.c index 6d5774cb8504..79ab03a0fdfe 100644 --- a/drivers/gpu/drm/i915/i915_timeline.c +++ b/drivers/gpu/drm/i915/i915_timeline.c @@ -13,7 +13,7 @@ void i915_timeline_init(struct drm_i915_private *i915, struct i915_timeline *timeline, const char *name) { - lockdep_assert_held(&i915->drm.struct_mutex); + struct i915_gt_timelines *gt = &i915->gt.timelines; /* * Ideally we want a set of engines on a single leaf as we expect @@ -23,9 +23,12 @@ void i915_timeline_init(struct drm_i915_private *i915, */ BUILD_BUG_ON(KSYNCMAP < I915_NUM_ENGINES); + timeline->i915 = i915; timeline->name = name; - list_add(&timeline->link, &i915->gt.timelines); + mutex_lock(&gt->mutex); + list_add(&timeline->link, &gt->list); + mutex_unlock(&gt->mutex); /* Called during early_init before we know how many engines there are */ @@ -40,6 +43,17 @@ void i915_timeline_init(struct drm_i915_private *i915, i915_syncmap_init(&timeline->sync); } +void i915_timelines_init(struct drm_i915_private *i915) +{ + struct i915_gt_timelines *gt = &i915->gt.timelines; + + mutex_init(&gt->mutex); + INIT_LIST_HEAD(&gt->list); + + /* via i915_gem_wait_for_idle() */ + i915_gem_shrinker_taints_mutex(i915, &gt->mutex); +} + /** * i915_timelines_park - called when the driver idles * @i915: the drm_i915_private device @@ -52,11 +66,11 @@ void i915_timeline_init(struct drm_i915_private *i915, */ void i915_timelines_park(struct drm_i915_private *i915) { + struct i915_gt_timelines *gt = &i915->gt.timelines; struct i915_timeline *timeline; - lockdep_assert_held(&i915->drm.struct_mutex); - - list_for_each_entry(timeline, &i915->gt.timelines, link) { + mutex_lock(&gt->mutex); + list_for_each_entry(timeline, &gt->list, link) { /* * All known fences are completed so we can scrap * the current sync point tracking and start afresh, @@ -65,16 +79,21 @@ void i915_timelines_park(struct drm_i915_private *i915) */ i915_syncmap_free(&timeline->sync); } + mutex_unlock(&gt->mutex); } void i915_timeline_fini(struct i915_timeline *timeline) { + struct i915_gt_timelines *gt = &timeline->i915->gt.timelines; + GEM_BUG_ON(!list_empty(&timeline->requests)); GEM_BUG_ON(i915_gem_active_isset(&timeline->barrier)); i915_syncmap_free(&timeline->sync); + mutex_lock(&gt->mutex); list_del(&timeline->link); + mutex_unlock(&gt->mutex); } struct i915_timeline * @@ -120,6 +139,15 @@ void __i915_timeline_free(struct kref *kref) kfree(timeline); } +void i915_timelines_fini(struct drm_i915_private *i915) +{ + struct i915_gt_timelines *gt = &i915->gt.timelines; + + GEM_BUG_ON(!list_empty(&gt->list)); + + mutex_destroy(&gt->mutex); +} + #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) #include "selftests/mock_timeline.c" #include "selftests/i915_timeline.c" diff --git a/drivers/gpu/drm/i915/i915_timeline.h b/drivers/gpu/drm/i915/i915_timeline.h index c8d7117bb205..b0df513b6ca3 100644 --- a/drivers/gpu/drm/i915/i915_timeline.h +++ b/drivers/gpu/drm/i915/i915_timeline.h @@ -76,6 +76,7 @@ struct i915_timeline { struct list_head link; const char *name; + struct drm_i915_private *i915; struct kref kref; }; @@ -144,7 +145,9 @@ static inline bool i915_timeline_sync_is_later(struct i915_timeline *tl, return __i915_timeline_sync_is_later(tl, fence->context, fence->seqno); } +void i915_timelines_init(struct drm_i915_private *i915); void i915_timelines_park(struct drm_i915_private *i915); +void i915_timelines_fini(struct drm_i915_private *i915); /** * i915_timeline_set_barrier - orders submission between different timelines diff --git a/drivers/gpu/drm/i915/selftests/mock_gem_device.c b/drivers/gpu/drm/i915/selftests/mock_gem_device.c index 8ab5a2688a0c..14ae46fda49f 100644 --- a/drivers/gpu/drm/i915/selftests/mock_gem_device.c +++ b/drivers/gpu/drm/i915/selftests/mock_gem_device.c @@ -68,13 +68,14 @@ static void mock_device_release(struct drm_device *dev) i915_gem_contexts_fini(i915); mutex_unlock(&i915->drm.struct_mutex); + i915_timelines_fini(i915); + drain_workqueue(i915->wq); i915_gem_drain_freed_objects(i915); mutex_lock(&i915->drm.struct_mutex); mock_fini_ggtt(&i915->ggtt); mutex_unlock(&i915->drm.struct_mutex); - WARN_ON(!list_empty(&i915->gt.timelines)); destroy_workqueue(i915->wq); @@ -226,7 +227,8 @@ struct drm_i915_private *mock_gem_device(void) if (!i915->priorities) goto err_dependencies; - INIT_LIST_HEAD(&i915->gt.timelines); + i915_timelines_init(i915); + INIT_LIST_HEAD(&i915->gt.active_rings); INIT_LIST_HEAD(&i915->gt.closed_vma); @@ -253,6 +255,7 @@ struct drm_i915_private *mock_gem_device(void) i915_gem_contexts_fini(i915); err_unlock: mutex_unlock(&i915->drm.struct_mutex); + i915_timelines_fini(i915); kmem_cache_destroy(i915->priorities); err_dependencies: kmem_cache_destroy(i915->dependencies); diff --git a/drivers/gpu/drm/i915/selftests/mock_timeline.c b/drivers/gpu/drm/i915/selftests/mock_timeline.c index 408113c1cc63..e5659aaa856d 100644 --- a/drivers/gpu/drm/i915/selftests/mock_timeline.c +++ b/drivers/gpu/drm/i915/selftests/mock_timeline.c @@ -10,6 +10,7 @@ void mock_timeline_init(struct i915_timeline *timeline, u64 context) { + timeline->i915 = NULL; timeline->fence_context = context; spin_lock_init(&timeline->lock); @@ -25,5 +26,5 @@ void mock_timeline_init(struct i915_timeline *timeline, u64 context) void mock_timeline_fini(struct i915_timeline *timeline) { - i915_timeline_fini(timeline); + i915_syncmap_free(&timeline->sync); }

[11/28] drm/i915: Move list of timelines under its own lock

Commit Message

Patch