diff options
Diffstat (limited to 'drivers/gpu/drm/i915/i915_gem.c')
| -rw-r--r-- | drivers/gpu/drm/i915/i915_gem.c | 232 |
1 files changed, 151 insertions, 81 deletions
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index 532a577ff7a1..f9c6b9b5002c 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -46,8 +46,6 @@ #include <linux/dma-buf.h> static void i915_gem_flush_free_objects(struct drm_i915_private *i915); -static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj); -static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj); static bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj) { @@ -705,6 +703,61 @@ i915_gem_create_ioctl(struct drm_device *dev, void *data, args->size, &args->handle); } +static inline enum fb_op_origin +fb_write_origin(struct drm_i915_gem_object *obj, unsigned int domain) +{ + return (domain == I915_GEM_DOMAIN_GTT ? + obj->frontbuffer_ggtt_origin : ORIGIN_CPU); +} + +static void +flush_write_domain(struct drm_i915_gem_object *obj, unsigned int flush_domains) +{ + struct drm_i915_private *dev_priv = to_i915(obj->base.dev); + + if (!(obj->base.write_domain & flush_domains)) + return; + + /* No actual flushing is required for the GTT write domain. Writes + * to it "immediately" go to main memory as far as we know, so there's + * no chipset flush. It also doesn't land in render cache. + * + * However, we do have to enforce the order so that all writes through + * the GTT land before any writes to the device, such as updates to + * the GATT itself. + * + * We also have to wait a bit for the writes to land from the GTT. + * An uncached read (i.e. mmio) seems to be ideal for the round-trip + * timing. This issue has only been observed when switching quickly + * between GTT writes and CPU reads from inside the kernel on recent hw, + * and it appears to only affect discrete GTT blocks (i.e. on LLC + * system agents we cannot reproduce this behaviour). + */ + wmb(); + + switch (obj->base.write_domain) { + case I915_GEM_DOMAIN_GTT: + if (INTEL_GEN(dev_priv) >= 6 && !HAS_LLC(dev_priv)) { + if (intel_runtime_pm_get_if_in_use(dev_priv)) { + spin_lock_irq(&dev_priv->uncore.lock); + POSTING_READ_FW(RING_ACTHD(dev_priv->engine[RCS]->mmio_base)); + spin_unlock_irq(&dev_priv->uncore.lock); + intel_runtime_pm_put(dev_priv); + } + } + + intel_fb_obj_flush(obj, + fb_write_origin(obj, I915_GEM_DOMAIN_GTT)); + break; + + case I915_GEM_DOMAIN_CPU: + i915_gem_clflush_object(obj, I915_CLFLUSH_SYNC); + break; + } + + obj->base.write_domain = 0; +} + static inline int __copy_to_user_swizzled(char __user *cpu_vaddr, const char *gpu_vaddr, int gpu_offset, @@ -794,7 +847,7 @@ int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj, goto out; } - i915_gem_object_flush_gtt_write_domain(obj); + flush_write_domain(obj, ~I915_GEM_DOMAIN_CPU); /* If we're not in the cpu read domain, set ourself into the gtt * read domain and manually flush cachelines (if required). This @@ -846,7 +899,7 @@ int i915_gem_obj_prepare_shmem_write(struct drm_i915_gem_object *obj, goto out; } - i915_gem_object_flush_gtt_write_domain(obj); + flush_write_domain(obj, ~I915_GEM_DOMAIN_CPU); /* If we're not in the cpu write domain, set ourself into the * gtt write domain and manually flush cachelines (as required). @@ -1501,13 +1554,6 @@ err: return ret; } -static inline enum fb_op_origin -write_origin(struct drm_i915_gem_object *obj, unsigned domain) -{ - return (domain == I915_GEM_DOMAIN_GTT ? - obj->frontbuffer_ggtt_origin : ORIGIN_CPU); -} - static void i915_gem_object_bump_inactive_ggtt(struct drm_i915_gem_object *obj) { struct drm_i915_private *i915; @@ -1591,10 +1637,12 @@ i915_gem_set_domain_ioctl(struct drm_device *dev, void *data, if (err) goto out_unpin; - if (read_domains & I915_GEM_DOMAIN_GTT) - err = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0); + if (read_domains & I915_GEM_DOMAIN_WC) + err = i915_gem_object_set_to_wc_domain(obj, write_domain); + else if (read_domains & I915_GEM_DOMAIN_GTT) + err = i915_gem_object_set_to_gtt_domain(obj, write_domain); else - err = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0); + err = i915_gem_object_set_to_cpu_domain(obj, write_domain); /* And bump the LRU for this access */ i915_gem_object_bump_inactive_ggtt(obj); @@ -1602,7 +1650,8 @@ i915_gem_set_domain_ioctl(struct drm_device *dev, void *data, mutex_unlock(&dev->struct_mutex); if (write_domain != 0) - intel_fb_obj_invalidate(obj, write_origin(obj, write_domain)); + intel_fb_obj_invalidate(obj, + fb_write_origin(obj, write_domain)); out_unpin: i915_gem_object_unpin_pages(obj); @@ -1737,6 +1786,9 @@ static unsigned int tile_row_pages(struct drm_i915_gem_object *obj) * into userspace. (This view is aligned and sized appropriately for * fenced access.) * + * 2 - Recognise WC as a separate cache domain so that we can flush the + * delayed writes via GTT before performing direct access via WC. + * * Restrictions: * * * snoopable objects cannot be accessed via the GTT. It can cause machine @@ -1764,7 +1816,7 @@ static unsigned int tile_row_pages(struct drm_i915_gem_object *obj) */ int i915_gem_mmap_gtt_version(void) { - return 1; + return 2; } static inline struct i915_ggtt_view @@ -3144,6 +3196,7 @@ i915_gem_idle_work_handler(struct work_struct *work) intel_engine_disarm_breadcrumbs(engine); i915_gem_batch_pool_fini(&engine->batch_pool); } + i915_gem_timelines_mark_idle(dev_priv); GEM_BUG_ON(!dev_priv->gt.awake); dev_priv->gt.awake = false; @@ -3320,56 +3373,6 @@ int i915_gem_wait_for_idle(struct drm_i915_private *i915, unsigned int flags) return ret; } -/** Flushes the GTT write domain for the object if it's dirty. */ -static void -i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj) -{ - struct drm_i915_private *dev_priv = to_i915(obj->base.dev); - - if (obj->base.write_domain != I915_GEM_DOMAIN_GTT) - return; - - /* No actual flushing is required for the GTT write domain. Writes - * to it "immediately" go to main memory as far as we know, so there's - * no chipset flush. It also doesn't land in render cache. - * - * However, we do have to enforce the order so that all writes through - * the GTT land before any writes to the device, such as updates to - * the GATT itself. - * - * We also have to wait a bit for the writes to land from the GTT. - * An uncached read (i.e. mmio) seems to be ideal for the round-trip - * timing. This issue has only been observed when switching quickly - * between GTT writes and CPU reads from inside the kernel on recent hw, - * and it appears to only affect discrete GTT blocks (i.e. on LLC - * system agents we cannot reproduce this behaviour). - */ - wmb(); - if (INTEL_GEN(dev_priv) >= 6 && !HAS_LLC(dev_priv)) { - if (intel_runtime_pm_get_if_in_use(dev_priv)) { - spin_lock_irq(&dev_priv->uncore.lock); - POSTING_READ_FW(RING_ACTHD(dev_priv->engine[RCS]->mmio_base)); - spin_unlock_irq(&dev_priv->uncore.lock); - intel_runtime_pm_put(dev_priv); - } - } - - intel_fb_obj_flush(obj, write_origin(obj, I915_GEM_DOMAIN_GTT)); - - obj->base.write_domain = 0; -} - -/** Flushes the CPU write domain for the object if it's dirty. */ -static void -i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj) -{ - if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) - return; - - i915_gem_clflush_object(obj, I915_CLFLUSH_SYNC); - obj->base.write_domain = 0; -} - static void __i915_gem_object_flush_for_display(struct drm_i915_gem_object *obj) { if (obj->base.write_domain != I915_GEM_DOMAIN_CPU && !obj->cache_dirty) @@ -3390,6 +3393,69 @@ void i915_gem_object_flush_if_display(struct drm_i915_gem_object *obj) } /** + * Moves a single object to the WC read, and possibly write domain. + * @obj: object to act on + * @write: ask for write access or read only + * + * This function returns when the move is complete, including waiting on + * flushes to occur. + */ +int +i915_gem_object_set_to_wc_domain(struct drm_i915_gem_object *obj, bool write) +{ + int ret; + + lockdep_assert_held(&obj->base.dev->struct_mutex); + + ret = i915_gem_object_wait(obj, + I915_WAIT_INTERRUPTIBLE | + I915_WAIT_LOCKED | + (write ? I915_WAIT_ALL : 0), + MAX_SCHEDULE_TIMEOUT, + NULL); + if (ret) + return ret; + + if (obj->base.write_domain == I915_GEM_DOMAIN_WC) + return 0; + + /* Flush and acquire obj->pages so that we are coherent through + * direct access in memory with previous cached writes through + * shmemfs and that our cache domain tracking remains valid. + * For example, if the obj->filp was moved to swap without us + * being notified and releasing the pages, we would mistakenly + * continue to assume that the obj remained out of the CPU cached + * domain. + */ + ret = i915_gem_object_pin_pages(obj); + if (ret) + return ret; + + flush_write_domain(obj, ~I915_GEM_DOMAIN_WC); + + /* Serialise direct access to this object with the barriers for + * coherent writes from the GPU, by effectively invalidating the + * WC domain upon first access. + */ + if ((obj->base.read_domains & I915_GEM_DOMAIN_WC) == 0) + mb(); + + /* It should now be out of any other write domains, and we can update + * the domain values for our changes. + */ + GEM_BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_WC) != 0); + obj->base.read_domains |= I915_GEM_DOMAIN_WC; + if (write) { + obj->base.read_domains = I915_GEM_DOMAIN_WC; + obj->base.write_domain = I915_GEM_DOMAIN_WC; + obj->mm.dirty = true; + } + + i915_gem_object_unpin_pages(obj); + return 0; +} + +/** * Moves a single object to the GTT read, and possibly write domain. * @obj: object to act on * @write: ask for write access or read only @@ -3428,7 +3494,7 @@ i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write) if (ret) return ret; - i915_gem_object_flush_cpu_write_domain(obj); + flush_write_domain(obj, ~I915_GEM_DOMAIN_GTT); /* Serialise direct access to this object with the barriers for * coherent writes from the GPU, by effectively invalidating the @@ -3802,7 +3868,7 @@ i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write) if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) return 0; - i915_gem_object_flush_gtt_write_domain(obj); + flush_write_domain(obj, ~I915_GEM_DOMAIN_CPU); /* Flush the CPU cache if it's still invalid. */ if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) { @@ -3996,7 +4062,7 @@ __busy_set_if_active(const struct dma_fence *fence, if (i915_gem_request_completed(rq)) return 0; - return flag(rq->engine->exec_id); + return flag(rq->engine->uabi_id); } static __always_inline unsigned int @@ -4195,7 +4261,7 @@ i915_gem_object_create(struct drm_i915_private *dev_priv, u64 size) * catch if we ever need to fix it. In the meantime, if you do spot * such a local variable, please consider fixing! */ - if (WARN_ON(size >> PAGE_SHIFT > INT_MAX)) + if (size >> PAGE_SHIFT > INT_MAX) return ERR_PTR(-E2BIG); if (overflows_type(size, obj->base.size)) @@ -4302,6 +4368,8 @@ static void __i915_gem_free_objects(struct drm_i915_private *i915, intel_runtime_pm_put(i915); mutex_unlock(&i915->drm.struct_mutex); + cond_resched(); + llist_for_each_entry_safe(obj, on, freed, freed) { GEM_BUG_ON(obj->bind_count); GEM_BUG_ON(atomic_read(&obj->frontbuffer_bits)); @@ -4349,8 +4417,11 @@ static void __i915_gem_free_work(struct work_struct *work) * unbound now. */ - while ((freed = llist_del_all(&i915->mm.free_list))) + while ((freed = llist_del_all(&i915->mm.free_list))) { __i915_gem_free_objects(i915, freed); + if (need_resched()) + break; + } } static void __i915_gem_free_object_rcu(struct rcu_head *head) @@ -4415,10 +4486,9 @@ void i915_gem_sanitize(struct drm_i915_private *i915) * try to take over. The only way to remove the earlier state * is by resetting. However, resetting on earlier gen is tricky as * it may impact the display and we are uncertain about the stability - * of the reset, so we only reset recent machines with logical - * context support (that must be reset to remove any stray contexts). + * of the reset, so this could be applied to even earlier gen. */ - if (HAS_HW_CONTEXTS(i915)) { + if (INTEL_GEN(i915) >= 5) { int reset = intel_gpu_reset(i915, ALL_ENGINES); WARN_ON(reset && reset != -ENODEV); } @@ -4676,7 +4746,7 @@ int i915_gem_init(struct drm_i915_private *dev_priv) mutex_lock(&dev_priv->drm.struct_mutex); - i915_gem_clflush_init(dev_priv); + dev_priv->mm.unordered_timeline = dma_fence_context_alloc(1); if (!i915.enable_execlists) { dev_priv->gt.resume = intel_legacy_submission_resume; @@ -4822,8 +4892,6 @@ i915_gem_load_init(struct drm_i915_private *dev_priv) init_waitqueue_head(&dev_priv->pending_flip_queue); - dev_priv->mm.interruptible = true; - atomic_set(&dev_priv->mm.bsd_engine_dispatch_index, 0); spin_lock_init(&dev_priv->fb_tracking.lock); @@ -4864,9 +4932,10 @@ void i915_gem_load_cleanup(struct drm_i915_private *dev_priv) int i915_gem_freeze(struct drm_i915_private *dev_priv) { - mutex_lock(&dev_priv->drm.struct_mutex); + /* Discard all purgeable objects, let userspace recover those as + * required after resuming. + */ i915_gem_shrink_all(dev_priv); - mutex_unlock(&dev_priv->drm.struct_mutex); return 0; } @@ -4891,12 +4960,13 @@ int i915_gem_freeze_late(struct drm_i915_private *dev_priv) * we update that state just before writing out the image. * * To try and reduce the hibernation image, we manually shrink - * the objects as well. + * the objects as well, see i915_gem_freeze() */ - mutex_lock(&dev_priv->drm.struct_mutex); i915_gem_shrink(dev_priv, -1UL, I915_SHRINK_UNBOUND); + i915_gem_drain_freed_objects(dev_priv); + mutex_lock(&dev_priv->drm.struct_mutex); for (p = phases; *p; p++) { list_for_each_entry(obj, *p, global_link) { obj->base.read_domains = I915_GEM_DOMAIN_CPU; |