@@ -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>
@@ -104,6 +105,27 @@ 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;
+
+ spin_lock(&syncobj->lock);
+ 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
@@ -526,6 +548,155 @@ drm_syncobj_fd_to_handle_ioctl(struct drm_device *dev, void *data,
&args->handle);
}
+struct syncobj_wait_entry {
+ wait_queue_head_t *wq;
+ 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(wait->wq);
+}
+
+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(wait->wq);
+}
+
+static bool drm_syncobj_array_wait_complete(struct syncobj_wait_entry *entries,
+ uint32_t count,
+ uint32_t flags,
+ uint32_t *idx)
+{
+ struct dma_fence *fence;
+ uint32_t signaled_count;
+ int i;
+
+ 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;
+ return true;
+ }
+ }
+ }
+
+ if (signaled_count == count)
+ return true;
+
+ return false;
+}
+
+static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
+ uint32_t count,
+ uint32_t flags,
+ signed long timeout,
+ uint32_t *idx)
+{
+ wait_queue_head_t syncobj_wait_queue;
+ struct syncobj_wait_entry *entries;
+ 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].wq = &syncobj_wait_queue;
+ 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++;
+ }
+ }
+
+ if (signaled_count == count ||
+ (signaled_count > 0 &&
+ !(flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL))) {
+ ret = max_t(signed long, timeout, 1);
+ 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!
+ */
+
+ init_waitqueue_head(&syncobj_wait_queue);
+
+ 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);
+ }
+ }
+
+ ret = wait_event_interruptible_timeout(
+ syncobj_wait_queue,
+ drm_syncobj_array_wait_complete(entries, count, flags, idx),
+ timeout);
+
+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
*
@@ -558,43 +729,19 @@ 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;
@@ -610,14 +757,15 @@ 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)
@@ -636,27 +784,28 @@ 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);
@@ -719,6 +719,7 @@ 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 */
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 Signed-off-by: Jason Ekstrand <jason@jlekstrand.net> Cc: Dave Airlie <airlied@redhat.com> --- drivers/gpu/drm/drm_syncobj.c | 239 ++++++++++++++++++++++++++++++++++-------- include/uapi/drm/drm.h | 1 + 2 files changed, 195 insertions(+), 45 deletions(-)