By applying the same logic as for wait-ioctl, we can query whether a
request has completed without holding struct_mutex. The biggest impact
system-wide is removing the flush_active and the contention that causes.
Testcase: igt/gem_busy
Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
Cc: Akash Goel <akash.goel@xxxxxxxxx>
---
drivers/gpu/drm/i915/i915_gem.c | 115 +++++++++++++++++++++++++++++-----------
1 file changed, 85 insertions(+), 30 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 4af64d864587..a4f949038d50 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -3463,49 +3463,104 @@ i915_gem_object_ggtt_unpin_view(struct drm_i915_gem_object *obj,
i915_vma_unpin(i915_gem_obj_to_ggtt_view(obj, view));
}
+static __always_inline unsigned
+__busy_read_flag(const struct drm_i915_gem_request *request)
+{
+ return 0x10000 << request->engine->exec_id;
+}
+
+static __always_inline unsigned
+__busy_write_flag(const struct drm_i915_gem_request *request)
+{
+ return request->engine->exec_id;
+}
+
+static __always_inline unsigned
+__busy_flag(const struct i915_gem_active *active,
+ unsigned (*flag)(const struct drm_i915_gem_request *))
+{
+ do {
+ struct drm_i915_gem_request *request;
+ unsigned busy;
+
+ request = rcu_dereference(active->__request);
+ if (!request || i915_gem_request_completed(request))
+ return 0;
+
+ busy = flag(request);
+ if (request == rcu_dereference(active->__request))
+ return busy;
+ } while (1);
+}
+
+static inline unsigned
+busy_read_flag(const struct i915_gem_active *active)
+{
+ return __busy_flag(active, __busy_read_flag);
+}
+
+static inline unsigned
+busy_write_flag(const struct i915_gem_active *active)
+{
+ return __busy_flag(active, __busy_write_flag);
+}
+
int
i915_gem_busy_ioctl(struct drm_device *dev, void *data,
struct drm_file *file)
{
struct drm_i915_gem_busy *args = data;
struct drm_i915_gem_object *obj;
- int ret;
-
- ret = i915_mutex_lock_interruptible(dev);
- if (ret)
- return ret;
+ unsigned long active;
obj = i915_gem_object_lookup(file, args->handle);
- if (!obj) {
- ret = -ENOENT;
- goto unlock;
- }
+ if (!obj)
+ return -ENOENT;
- /* Count all active objects as busy, even if they are currently not used
- * by the gpu. Users of this interface expect objects to eventually
- * become non-busy without any further actions.
- */
args->busy = 0;
- if (i915_gem_object_is_active(obj)) {
- struct drm_i915_gem_request *req;
- int i;
+ active = __I915_BO_ACTIVE(obj);
+ if (active) {
+ int idx;
- for (i = 0; i < I915_NUM_ENGINES; i++) {
- req = i915_gem_active_peek(&obj->last_read[i],
- &obj->base.dev->struct_mutex);
- if (req)
- args->busy |= 1 << (16 + req->engine->exec_id);
- }
- req = i915_gem_active_peek(&obj->last_write,
- &obj->base.dev->struct_mutex);
- if (req)
- args->busy |= req->engine->exec_id;
+ /* Yes, the lookups are intentionally racy.
+ *
+ * Even though we guard the pointer lookup by RCU, that only
+ * guarantees that the pointer and its contents remain
+ * dereferencable and does *not* mean that the request we
+ * have is the same as the one being tracked by the object.
+ *
+ * Consider that we lookup the request just as it is being
+ * retired and free. We take a local copy of the pointer,
+ * but before we add its engine into the busy set, the other
+ * thread reallocates it and assigns it to a task on another
+ * engine with a fresh and incomplete seqno.
+ *
+ * So after we lookup the engine's id, we double check that
+ * the active request is the same and only then do we add it
+ * into the busy set.
+ */
+ rcu_read_lock();
+
+ for_each_active(active, idx)
+ args->busy |= busy_read_flag(&obj->last_read[idx]);
+
+ /* For ABI sanity, we only care that the write engine is in
+ * the set of read engines. This is ensured by the ordering
+ * of setting last_read/last_write in i915_vma_move_to_active,
+ * and then in reverse in retire.
+ *
+ * We don't care that the set of active read/write engines
+ * may change during construction of the result, as it is
+ * equally liable to change before userspace can inspect
+ * the result.
+ */
+ args->busy |= busy_write_flag(&obj->last_write);
+
+ rcu_read_unlock();
}
- i915_gem_object_put(obj);
-unlock:
- mutex_unlock(&dev->struct_mutex);
- return ret;
+ i915_gem_object_put_unlocked(obj);
+ return 0;
}
int
--
2.8.1
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
