]> git.karo-electronics.de Git - linux-beck.git/commitdiff
drm/i915: Convert non-blocking waits for requests over to using RCU
authorChris Wilson <chris@chris-wilson.co.uk>
Fri, 5 Aug 2016 09:14:07 +0000 (10:14 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Fri, 5 Aug 2016 09:54:34 +0000 (10:54 +0100)
We can completely avoid taking the struct_mutex around the non-blocking
waits by switching over to the RCU request management (trading the mutex
for a RCU read lock and some complex atomic operations). The improvement
is that we gain further contention reduction, and overall the code
become simpler due to the reduced mutex dancing.

v2: Move i915_gem_fault tracepoint back to the start of the function,
before the unlocked wait.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1470388464-28458-2-git-send-email-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_gem.c

index bc414789ed3ded4c7d795b288eb4044aba5ed033..cc6d102f1d48c41cb92fc537c39608b0390e438c 100644 (file)
@@ -349,24 +349,20 @@ i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
        return 0;
 }
 
-/* A nonblocking variant of the above wait. This is a highly dangerous routine
- * as the object state may change during this call.
+/* A nonblocking variant of the above wait. Must be called prior to
+ * acquiring the mutex for the object, as the object state may change
+ * during this call. A reference must be held by the caller for the object.
  */
 static __must_check int
-i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
-                                           struct intel_rps_client *rps,
-                                           bool readonly)
+__unsafe_wait_rendering(struct drm_i915_gem_object *obj,
+                       struct intel_rps_client *rps,
+                       bool readonly)
 {
-       struct drm_device *dev = obj->base.dev;
-       struct drm_i915_gem_request *requests[I915_NUM_ENGINES];
        struct i915_gem_active *active;
        unsigned long active_mask;
-       int ret, i, n = 0;
-
-       lockdep_assert_held(&dev->struct_mutex);
-       GEM_BUG_ON(!to_i915(dev)->mm.interruptible);
+       int idx;
 
-       active_mask = i915_gem_object_get_active(obj);
+       active_mask = __I915_BO_ACTIVE(obj);
        if (!active_mask)
                return 0;
 
@@ -377,25 +373,16 @@ i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
                active = &obj->last_write;
        }
 
-       for_each_active(active_mask, i) {
-               struct drm_i915_gem_request *req;
+       for_each_active(active_mask, idx) {
+               int ret;
 
-               req = i915_gem_active_get(&active[i],
-                                         &obj->base.dev->struct_mutex);
-               if (req)
-                       requests[n++] = req;
+               ret = i915_gem_active_wait_unlocked(&active[idx],
+                                                   true, NULL, rps);
+               if (ret)
+                       return ret;
        }
 
-       mutex_unlock(&dev->struct_mutex);
-       ret = 0;
-       for (i = 0; ret == 0 && i < n; i++)
-               ret = i915_wait_request(requests[i], true, NULL, rps);
-       mutex_lock(&dev->struct_mutex);
-
-       for (i = 0; i < n; i++)
-               i915_gem_request_put(requests[i]);
-
-       return ret;
+       return 0;
 }
 
 static struct intel_rps_client *to_rps_client(struct drm_file *file)
@@ -1467,10 +1454,7 @@ i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
        int ret;
 
        /* Only handle setting domains to types used by the CPU. */
-       if (write_domain & I915_GEM_GPU_DOMAINS)
-               return -EINVAL;
-
-       if (read_domains & I915_GEM_GPU_DOMAINS)
+       if ((write_domain | read_domains) & I915_GEM_GPU_DOMAINS)
                return -EINVAL;
 
        /* Having something in the write domain implies it's in the read
@@ -1479,25 +1463,21 @@ i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
        if (write_domain != 0 && read_domains != write_domain)
                return -EINVAL;
 
-       ret = i915_mutex_lock_interruptible(dev);
-       if (ret)
-               return ret;
-
        obj = i915_gem_object_lookup(file, args->handle);
-       if (!obj) {
-               ret = -ENOENT;
-               goto unlock;
-       }
+       if (!obj)
+               return -ENOENT;
 
        /* Try to flush the object off the GPU without holding the lock.
         * We will repeat the flush holding the lock in the normal manner
         * to catch cases where we are gazumped.
         */
-       ret = i915_gem_object_wait_rendering__nonblocking(obj,
-                                                         to_rps_client(file),
-                                                         !write_domain);
+       ret = __unsafe_wait_rendering(obj, to_rps_client(file), !write_domain);
+       if (ret)
+               goto err;
+
+       ret = i915_mutex_lock_interruptible(dev);
        if (ret)
-               goto unref;
+               goto err;
 
        if (read_domains & I915_GEM_DOMAIN_GTT)
                ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
@@ -1507,11 +1487,13 @@ i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
        if (write_domain != 0)
                intel_fb_obj_invalidate(obj, write_origin(obj, write_domain));
 
-unref:
        i915_gem_object_put(obj);
-unlock:
        mutex_unlock(&dev->struct_mutex);
        return ret;
+
+err:
+       i915_gem_object_put_unlocked(obj);
+       return ret;
 }
 
 /**
@@ -1648,36 +1630,36 @@ int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
        struct drm_i915_private *dev_priv = to_i915(dev);
        struct i915_ggtt *ggtt = &dev_priv->ggtt;
        struct i915_ggtt_view view = i915_ggtt_view_normal;
+       bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
        pgoff_t page_offset;
        unsigned long pfn;
-       int ret = 0;
-       bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
-
-       intel_runtime_pm_get(dev_priv);
+       int ret;
 
        /* We don't use vmf->pgoff since that has the fake offset */
        page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
                PAGE_SHIFT;
 
-       ret = i915_mutex_lock_interruptible(dev);
-       if (ret)
-               goto out;
-
        trace_i915_gem_object_fault(obj, page_offset, true, write);
 
        /* Try to flush the object off the GPU first without holding the lock.
-        * Upon reacquiring the lock, we will perform our sanity checks and then
+        * Upon acquiring the lock, we will perform our sanity checks and then
         * repeat the flush holding the lock in the normal manner to catch cases
         * where we are gazumped.
         */
-       ret = i915_gem_object_wait_rendering__nonblocking(obj, NULL, !write);
+       ret = __unsafe_wait_rendering(obj, NULL, !write);
        if (ret)
-               goto unlock;
+               goto err;
+
+       intel_runtime_pm_get(dev_priv);
+
+       ret = i915_mutex_lock_interruptible(dev);
+       if (ret)
+               goto err_rpm;
 
        /* Access to snoopable pages through the GTT is incoherent. */
        if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev)) {
                ret = -EFAULT;
-               goto unlock;
+               goto err_unlock;
        }
 
        /* Use a partial view if the object is bigger than the aperture. */
@@ -1698,15 +1680,15 @@ int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
        /* Now pin it into the GTT if needed */
        ret = i915_gem_object_ggtt_pin(obj, &view, 0, 0, PIN_MAPPABLE);
        if (ret)
-               goto unlock;
+               goto err_unlock;
 
        ret = i915_gem_object_set_to_gtt_domain(obj, write);
        if (ret)
-               goto unpin;
+               goto err_unpin;
 
        ret = i915_gem_object_get_fence(obj);
        if (ret)
-               goto unpin;
+               goto err_unpin;
 
        /* Finally, remap it using the new GTT offset */
        pfn = ggtt->mappable_base +
@@ -1751,11 +1733,13 @@ int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
                                            (unsigned long)vmf->virtual_address,
                                            pfn + page_offset);
        }
-unpin:
+err_unpin:
        i915_gem_object_ggtt_unpin_view(obj, &view);
-unlock:
+err_unlock:
        mutex_unlock(&dev->struct_mutex);
-out:
+err_rpm:
+       intel_runtime_pm_put(dev_priv);
+err:
        switch (ret) {
        case -EIO:
                /*
@@ -1796,8 +1780,6 @@ out:
                ret = VM_FAULT_SIGBUS;
                break;
        }
-
-       intel_runtime_pm_put(dev_priv);
        return ret;
 }