drm/syncobj: Allow wait for submit and signal behavior (v5)
authorJason Ekstrand <jason@jlekstrand.net>
Fri, 25 Aug 2017 17:52:24 +0000 (10:52 -0700)
committerDave Airlie <airlied@redhat.com>
Mon, 28 Aug 2017 20:28:17 +0000 (06:28 +1000)
Vulkan VkFence semantics require that the application be able to perform
a CPU wait on work which may not yet have been submitted.  This is
perfectly safe because the CPU wait has a timeout which will get
triggered eventually if no work is ever submitted.  This behavior is
advantageous for multi-threaded workloads because, so long as all of the
threads agree on what fences to use up-front, you don't have the extra
cross-thread synchronization cost of thread A telling thread B that it
has submitted its dependent work and thread B is now free to wait.

Within a single process, this can be implemented in the userspace driver
by doing exactly the same kind of tracking the app would have to do
using posix condition variables or similar.  However, in order for this
to work cross-process (as is required by VK_KHR_external_fence), we need
to handle this in the kernel.

This commit adds a WAIT_FOR_SUBMIT flag to DRM_IOCTL_SYNCOBJ_WAIT which
instructs the IOCTL to wait for the syncobj to have a non-null fence and
then wait on the fence.  Combined with DRM_IOCTL_SYNCOBJ_RESET, you can
easily get the Vulkan behavior.

v2:
 - Fix a bug in the invalid syncobj error path
 - Unify the wait-all and wait-any cases
v3:
 - Unify the timeout == 0 case a bit with the timeout > 0 case
 - Use wait_event_interruptible_timeout
v4:
 - Use proxy fence
v5:
 - Revert to a combination of v2 and v3
 - Don't use proxy fences
 - Don't use wait_event_interruptible_timeout because it just adds an
   extra layer of callbacks

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: Dave Airlie <airlied@redhat.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Christian König <christian.koenig@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
drivers/gpu/drm/drm_syncobj.c
include/uapi/drm/drm.h

index 12db8c9564cd0b8e5a0646ecfa61c99f3f5fa345..cccd3bd194c68390f4f5e0c464b900b71803080d 100644 (file)
@@ -51,6 +51,7 @@
 #include <linux/fs.h>
 #include <linux/anon_inodes.h>
 #include <linux/sync_file.h>
+#include <linux/sched/signal.h>
 
 #include "drm_internal.h"
 #include <drm/drm_syncobj.h>
@@ -88,6 +89,35 @@ static void drm_syncobj_add_callback_locked(struct drm_syncobj *syncobj,
        list_add_tail(&cb->node, &syncobj->cb_list);
 }
 
+static int drm_syncobj_fence_get_or_add_callback(struct drm_syncobj *syncobj,
+                                                struct dma_fence **fence,
+                                                struct drm_syncobj_cb *cb,
+                                                drm_syncobj_func_t func)
+{
+       int ret;
+
+       *fence = drm_syncobj_fence_get(syncobj);
+       if (*fence)
+               return 1;
+
+       spin_lock(&syncobj->lock);
+       /* We've already tried once to get a fence and failed.  Now that we
+        * have the lock, try one more time just to be sure we don't add a
+        * callback when a fence has already been set.
+        */
+       if (syncobj->fence) {
+               *fence = dma_fence_get(syncobj->fence);
+               ret = 1;
+       } else {
+               *fence = NULL;
+               drm_syncobj_add_callback_locked(syncobj, cb, func);
+               ret = 0;
+       }
+       spin_unlock(&syncobj->lock);
+
+       return ret;
+}
+
 /**
  * drm_syncobj_add_callback - adds a callback to syncobj::cb_list
  * @syncobj: Sync object to which to add the callback
@@ -560,6 +590,160 @@ drm_syncobj_fd_to_handle_ioctl(struct drm_device *dev, void *data,
                                        &args->handle);
 }
 
+struct syncobj_wait_entry {
+       struct task_struct *task;
+       struct dma_fence *fence;
+       struct dma_fence_cb fence_cb;
+       struct drm_syncobj_cb syncobj_cb;
+};
+
+static void syncobj_wait_fence_func(struct dma_fence *fence,
+                                   struct dma_fence_cb *cb)
+{
+       struct syncobj_wait_entry *wait =
+               container_of(cb, struct syncobj_wait_entry, fence_cb);
+
+       wake_up_process(wait->task);
+}
+
+static void syncobj_wait_syncobj_func(struct drm_syncobj *syncobj,
+                                     struct drm_syncobj_cb *cb)
+{
+       struct syncobj_wait_entry *wait =
+               container_of(cb, struct syncobj_wait_entry, syncobj_cb);
+
+       /* This happens inside the syncobj lock */
+       wait->fence = dma_fence_get(syncobj->fence);
+       wake_up_process(wait->task);
+}
+
+static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
+                                                 uint32_t count,
+                                                 uint32_t flags,
+                                                 signed long timeout,
+                                                 uint32_t *idx)
+{
+       struct syncobj_wait_entry *entries;
+       struct dma_fence *fence;
+       signed long ret;
+       uint32_t signaled_count, i;
+
+       entries = kcalloc(count, sizeof(*entries), GFP_KERNEL);
+       if (!entries)
+               return -ENOMEM;
+
+       /* Walk the list of sync objects and initialize entries.  We do
+        * this up-front so that we can properly return -EINVAL if there is
+        * a syncobj with a missing fence and then never have the chance of
+        * returning -EINVAL again.
+        */
+       signaled_count = 0;
+       for (i = 0; i < count; ++i) {
+               entries[i].task = current;
+               entries[i].fence = drm_syncobj_fence_get(syncobjs[i]);
+               if (!entries[i].fence) {
+                       if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
+                               continue;
+                       } else {
+                               ret = -EINVAL;
+                               goto cleanup_entries;
+                       }
+               }
+
+               if (dma_fence_is_signaled(entries[i].fence)) {
+                       if (signaled_count == 0 && idx)
+                               *idx = i;
+                       signaled_count++;
+               }
+       }
+
+       /* Initialize ret to the max of timeout and 1.  That way, the
+        * default return value indicates a successful wait and not a
+        * timeout.
+        */
+       ret = max_t(signed long, timeout, 1);
+
+       if (signaled_count == count ||
+           (signaled_count > 0 &&
+            !(flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL)))
+               goto cleanup_entries;
+
+       /* There's a very annoying laxness in the dma_fence API here, in
+        * that backends are not required to automatically report when a
+        * fence is signaled prior to fence->ops->enable_signaling() being
+        * called.  So here if we fail to match signaled_count, we need to
+        * fallthough and try a 0 timeout wait!
+        */
+
+       if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
+               for (i = 0; i < count; ++i) {
+                       drm_syncobj_fence_get_or_add_callback(syncobjs[i],
+                                                             &entries[i].fence,
+                                                             &entries[i].syncobj_cb,
+                                                             syncobj_wait_syncobj_func);
+               }
+       }
+
+       do {
+               set_current_state(TASK_INTERRUPTIBLE);
+
+               signaled_count = 0;
+               for (i = 0; i < count; ++i) {
+                       fence = entries[i].fence;
+                       if (!fence)
+                               continue;
+
+                       if (dma_fence_is_signaled(fence) ||
+                           (!entries[i].fence_cb.func &&
+                            dma_fence_add_callback(fence,
+                                                   &entries[i].fence_cb,
+                                                   syncobj_wait_fence_func))) {
+                               /* The fence has been signaled */
+                               if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL) {
+                                       signaled_count++;
+                               } else {
+                                       if (idx)
+                                               *idx = i;
+                                       goto done_waiting;
+                               }
+                       }
+               }
+
+               if (signaled_count == count)
+                       goto done_waiting;
+
+               if (timeout == 0) {
+                       /* If we are doing a 0 timeout wait and we got
+                        * here, then we just timed out.
+                        */
+                       ret = 0;
+                       goto done_waiting;
+               }
+
+               ret = schedule_timeout(ret);
+
+               if (ret > 0 && signal_pending(current))
+                       ret = -ERESTARTSYS;
+       } while (ret > 0);
+
+done_waiting:
+       __set_current_state(TASK_RUNNING);
+
+cleanup_entries:
+       for (i = 0; i < count; ++i) {
+               if (entries[i].syncobj_cb.func)
+                       drm_syncobj_remove_callback(syncobjs[i],
+                                                   &entries[i].syncobj_cb);
+               if (entries[i].fence_cb.func)
+                       dma_fence_remove_callback(entries[i].fence,
+                                                 &entries[i].fence_cb);
+               dma_fence_put(entries[i].fence);
+       }
+       kfree(entries);
+
+       return ret;
+}
+
 /**
  * drm_timeout_abs_to_jiffies - calculate jiffies timeout from absolute value
  *
@@ -592,43 +776,19 @@ static signed long drm_timeout_abs_to_jiffies(int64_t timeout_nsec)
        return timeout_jiffies64 + 1;
 }
 
-static int drm_syncobj_wait_fences(struct drm_device *dev,
-                                  struct drm_file *file_private,
-                                  struct drm_syncobj_wait *wait,
-                                  struct dma_fence **fences)
+static int drm_syncobj_array_wait(struct drm_device *dev,
+                                 struct drm_file *file_private,
+                                 struct drm_syncobj_wait *wait,
+                                 struct drm_syncobj **syncobjs)
 {
        signed long timeout = drm_timeout_abs_to_jiffies(wait->timeout_nsec);
        signed long ret = 0;
        uint32_t first = ~0;
 
-       if (wait->flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL) {
-               uint32_t i;
-               for (i = 0; i < wait->count_handles; i++) {
-                       ret = dma_fence_wait_timeout(fences[i], true, timeout);
-
-                       /* Various dma_fence wait callbacks will return
-                        * ENOENT to indicate that the fence has already
-                        * been signaled.  We need to sanitize this to 0 so
-                        * we don't return early and the client doesn't see
-                        * an unexpected error.
-                        */
-                       if (ret == -ENOENT)
-                               ret = 0;
-
-                       if (ret < 0)
-                               return ret;
-                       if (ret == 0)
-                               break;
-                       timeout = ret;
-               }
-               first = 0;
-       } else {
-               ret = dma_fence_wait_any_timeout(fences,
-                                                wait->count_handles,
-                                                true, timeout,
-                                                &first);
-       }
-
+       ret = drm_syncobj_array_wait_timeout(syncobjs,
+                                            wait->count_handles,
+                                            wait->flags,
+                                            timeout, &first);
        if (ret < 0)
                return ret;
 
@@ -644,14 +804,15 @@ drm_syncobj_wait_ioctl(struct drm_device *dev, void *data,
 {
        struct drm_syncobj_wait *args = data;
        uint32_t *handles;
-       struct dma_fence **fences;
+       struct drm_syncobj **syncobjs;
        int ret = 0;
        uint32_t i;
 
        if (!drm_core_check_feature(dev, DRIVER_SYNCOBJ))
                return -ENODEV;
 
-       if (args->flags != 0 && args->flags != DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL)
+       if (args->flags & ~(DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL |
+                           DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT))
                return -EINVAL;
 
        if (args->count_handles == 0)
@@ -670,27 +831,28 @@ drm_syncobj_wait_ioctl(struct drm_device *dev, void *data,
                goto err_free_handles;
        }
 
-       fences = kcalloc(args->count_handles,
-                        sizeof(struct dma_fence *), GFP_KERNEL);
-       if (!fences) {
+       syncobjs = kcalloc(args->count_handles,
+                          sizeof(struct drm_syncobj *), GFP_KERNEL);
+       if (!syncobjs) {
                ret = -ENOMEM;
                goto err_free_handles;
        }
 
        for (i = 0; i < args->count_handles; i++) {
-               ret = drm_syncobj_find_fence(file_private, handles[i],
-                                            &fences[i]);
-               if (ret)
+               syncobjs[i] = drm_syncobj_find(file_private, handles[i]);
+               if (!syncobjs[i]) {
+                       ret = -ENOENT;
                        goto err_free_fence_array;
+               }
        }
 
-       ret = drm_syncobj_wait_fences(dev, file_private,
-                                     args, fences);
+       ret = drm_syncobj_array_wait(dev, file_private,
+                                    args, syncobjs);
 
 err_free_fence_array:
-       for (i = 0; i < args->count_handles; i++)
-               dma_fence_put(fences[i]);
-       kfree(fences);
+       while (i-- > 0)
+               drm_syncobj_put(syncobjs[i]);
+       kfree(syncobjs);
 err_free_handles:
        kfree(handles);
 
index ade7f68d32b530d1e67607bac3e707b6b386abd0..4c746597225ec8fdb3863680ed2bf8b05f17e4c6 100644 (file)
@@ -720,6 +720,7 @@ struct drm_syncobj_handle {
 };
 
 #define DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL (1 << 0)
+#define DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT (1 << 1)
 struct drm_syncobj_wait {
        __u64 handles;
        /* absolute timeout */