]> git.baikalelectronics.ru Git - kernel.git/commitdiff
drm/i915: Track GGTT writes on the vma
authorChris Wilson <chris@chris-wilson.co.uk>
Wed, 6 Dec 2017 12:49:14 +0000 (12:49 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Thu, 7 Dec 2017 14:01:59 +0000 (14:01 +0000)
As writes through the GTT and GGTT PTE updates do not share the same
path, they are not strictly ordered and so we must explicitly flush the
indirect writes prior to modifying the PTE. We do track outstanding GGTT
writes on the object itself, but since the object may have multiple GGTT
vma, that is overly coarse as we can track and flush individual vma as
required.

Whilst here, update the GGTT flushing behaviour for Cannonlake.

v2: Hard-code ring offset to allow use during unload (after RCS may have
been freed, or never existed!)

References: https://bugs.freedesktop.org/show_bug.cgi?id=104002
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171206124914.19960-2-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_gem.c
drivers/gpu/drm/i915/i915_vma.c
drivers/gpu/drm/i915/i915_vma.h

index 02551c781f0ab8688eafbf631174b422f6ac74db..f9386e793c8709f15d463c51ab156d87cac4ef3f 100644 (file)
@@ -3887,6 +3887,8 @@ int __must_check i915_gem_evict_for_node(struct i915_address_space *vm,
                                         unsigned int flags);
 int i915_gem_evict_vm(struct i915_address_space *vm);
 
+void i915_gem_flush_ggtt_writes(struct drm_i915_private *dev_priv);
+
 /* belongs in i915_gem_gtt.h */
 static inline void i915_gem_chipset_flush(struct drm_i915_private *dev_priv)
 {
index 5504be753092ca077d68c337cff5794080700f61..67dc11effc8e55c31fcc80960582cf6da85aacca 100644 (file)
@@ -666,17 +666,13 @@ fb_write_origin(struct drm_i915_gem_object *obj, unsigned int domain)
                obj->frontbuffer_ggtt_origin : ORIGIN_CPU);
 }
 
-static void
-flush_write_domain(struct drm_i915_gem_object *obj, unsigned int flush_domains)
+void i915_gem_flush_ggtt_writes(struct drm_i915_private *dev_priv)
 {
-       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.
+       /*
+        * No actual flushing is required for the GTT write domain for reads
+        * from the GTT 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 the GPU 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
@@ -687,22 +683,46 @@ flush_write_domain(struct drm_i915_gem_object *obj, unsigned int flush_domains)
         * 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).
+        * system agents we cannot reproduce this behaviour, until Cannonlake
+        * that was!).
         */
+
        wmb();
 
+       intel_runtime_pm_get(dev_priv);
+       spin_lock_irq(&dev_priv->uncore.lock);
+
+       POSTING_READ_FW(RING_HEAD(RENDER_RING_BASE));
+
+       spin_unlock_irq(&dev_priv->uncore.lock);
+       intel_runtime_pm_put(dev_priv);
+}
+
+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);
+       struct i915_vma *vma;
+
+       if (!(obj->base.write_domain & flush_domains))
+               return;
+
        switch (obj->base.write_domain) {
        case I915_GEM_DOMAIN_GTT:
-               if (!HAS_LLC(dev_priv)) {
-                       intel_runtime_pm_get(dev_priv);
-                       spin_lock_irq(&dev_priv->uncore.lock);
-                       POSTING_READ_FW(RING_HEAD(dev_priv->engine[RCS]->mmio_base));
-                       spin_unlock_irq(&dev_priv->uncore.lock);
-                       intel_runtime_pm_put(dev_priv);
-               }
+               i915_gem_flush_ggtt_writes(dev_priv);
 
                intel_fb_obj_flush(obj,
                                   fb_write_origin(obj, I915_GEM_DOMAIN_GTT));
+
+               list_for_each_entry(vma, &obj->vma_list, obj_link) {
+                       if (!i915_vma_is_ggtt(vma))
+                               break;
+
+                       if (vma->iomap)
+                               continue;
+
+                       i915_vma_unset_ggtt_write(vma);
+               }
                break;
 
        case I915_GEM_DOMAIN_CPU:
@@ -1965,6 +1985,8 @@ int i915_gem_fault(struct vm_fault *vmf)
                list_add(&obj->userfault_link, &dev_priv->mm.userfault_list);
        GEM_BUG_ON(!obj->userfault_count);
 
+       i915_vma_set_ggtt_write(vma);
+
 err_fence:
        i915_vma_unpin_fence(vma);
 err_unpin:
index 1013403fcfea692fe65daddf99a557cc7caca86d..0ebd75693505ea3b39ecef1e38d987f6e3c8dcc1 100644 (file)
@@ -322,6 +322,7 @@ void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
        if (err)
                goto err_unpin;
 
+       i915_vma_set_ggtt_write(vma);
        return ptr;
 
 err_unpin:
@@ -330,12 +331,24 @@ err:
        return IO_ERR_PTR(err);
 }
 
+void i915_vma_flush_writes(struct i915_vma *vma)
+{
+       if (!i915_vma_has_ggtt_write(vma))
+               return;
+
+       i915_gem_flush_ggtt_writes(vma->vm->i915);
+
+       i915_vma_unset_ggtt_write(vma);
+}
+
 void i915_vma_unpin_iomap(struct i915_vma *vma)
 {
        lockdep_assert_held(&vma->obj->base.dev->struct_mutex);
 
        GEM_BUG_ON(vma->iomap == NULL);
 
+       i915_vma_flush_writes(vma);
+
        i915_vma_unpin_fence(vma);
        i915_vma_unpin(vma);
 }
@@ -792,6 +805,15 @@ int i915_vma_unbind(struct i915_vma *vma)
        GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
 
        if (i915_vma_is_map_and_fenceable(vma)) {
+               /*
+                * Check that we have flushed all writes through the GGTT
+                * before the unbind, other due to non-strict nature of those
+                * indirect writes they may end up referencing the GGTT PTE
+                * after the unbind.
+                */
+               i915_vma_flush_writes(vma);
+               GEM_BUG_ON(i915_vma_has_ggtt_write(vma));
+
                /* release the fence reg _after_ flushing */
                ret = i915_vma_put_fence(vma);
                if (ret)
index 1e2bc9b3c3ac19a4790222eb151765050e264d49..f636243eb8f72773a3551bf989ce0f90a44489a4 100644 (file)
@@ -90,6 +90,7 @@ struct i915_vma {
 #define I915_VMA_CLOSED                BIT(10)
 #define I915_VMA_USERFAULT_BIT 11
 #define I915_VMA_USERFAULT     BIT(I915_VMA_USERFAULT_BIT)
+#define I915_VMA_GGTT_WRITE    BIT(12)
 
        unsigned int active;
        struct i915_gem_active last_read[I915_NUM_ENGINES];
@@ -138,6 +139,24 @@ static inline bool i915_vma_is_ggtt(const struct i915_vma *vma)
        return vma->flags & I915_VMA_GGTT;
 }
 
+static inline bool i915_vma_has_ggtt_write(const struct i915_vma *vma)
+{
+       return vma->flags & I915_VMA_GGTT_WRITE;
+}
+
+static inline void i915_vma_set_ggtt_write(struct i915_vma *vma)
+{
+       GEM_BUG_ON(!i915_vma_is_ggtt(vma));
+       vma->flags |= I915_VMA_GGTT_WRITE;
+}
+
+static inline void i915_vma_unset_ggtt_write(struct i915_vma *vma)
+{
+       vma->flags &= ~I915_VMA_GGTT_WRITE;
+}
+
+void i915_vma_flush_writes(struct i915_vma *vma);
+
 static inline bool i915_vma_is_map_and_fenceable(const struct i915_vma *vma)
 {
        return vma->flags & I915_VMA_CAN_FENCE;