nir: Add a nir_foreach_uniform_variable helper
[mesa.git] / src / intel / vulkan / anv_queue.c
index dc476470b75e7201069bb78d6fa3be972e8e5f02..7ba64af54c46d580c9fe8fd765f4852bcfaf84d4 100644 (file)
  * This file implements VkQueue, VkFence, and VkSemaphore
  */
 
+#include <errno.h>
 #include <fcntl.h>
 #include <unistd.h>
 
+#include "util/os_file.h"
+
 #include "anv_private.h"
 #include "vk_util.h"
 
@@ -98,17 +101,192 @@ anv_queue_submit_free(struct anv_device *device,
       close(submit->out_fence);
    vk_free(alloc, submit->fences);
    vk_free(alloc, submit->temporary_semaphores);
+   vk_free(alloc, submit->wait_timelines);
+   vk_free(alloc, submit->wait_timeline_values);
+   vk_free(alloc, submit->signal_timelines);
+   vk_free(alloc, submit->signal_timeline_values);
    vk_free(alloc, submit->fence_bos);
    vk_free(alloc, submit);
 }
 
+static bool
+anv_queue_submit_ready_locked(struct anv_queue_submit *submit)
+{
+   for (uint32_t i = 0; i < submit->wait_timeline_count; i++) {
+      if (submit->wait_timeline_values[i] > submit->wait_timelines[i]->highest_pending)
+         return false;
+   }
+
+   return true;
+}
+
 static VkResult
-_anv_queue_submit(struct anv_queue *queue, struct anv_queue_submit **_submit)
+anv_timeline_init(struct anv_device *device,
+                  struct anv_timeline *timeline,
+                  uint64_t initial_value)
 {
-   struct anv_queue_submit *submit = *_submit;
-   VkResult result = anv_queue_execbuf(queue, submit);
+   timeline->highest_past =
+      timeline->highest_pending = initial_value;
+   list_inithead(&timeline->points);
+   list_inithead(&timeline->free_points);
+
+   return VK_SUCCESS;
+}
+
+static void
+anv_timeline_finish(struct anv_device *device,
+                    struct anv_timeline *timeline)
+{
+   list_for_each_entry_safe(struct anv_timeline_point, point,
+                            &timeline->free_points, link) {
+      list_del(&point->link);
+      anv_device_release_bo(device, point->bo);
+      vk_free(&device->vk.alloc, point);
+   }
+   list_for_each_entry_safe(struct anv_timeline_point, point,
+                            &timeline->points, link) {
+      list_del(&point->link);
+      anv_device_release_bo(device, point->bo);
+      vk_free(&device->vk.alloc, point);
+   }
+}
+
+static VkResult
+anv_timeline_add_point_locked(struct anv_device *device,
+                              struct anv_timeline *timeline,
+                              uint64_t value,
+                              struct anv_timeline_point **point)
+{
+   VkResult result = VK_SUCCESS;
+
+   if (list_is_empty(&timeline->free_points)) {
+      *point =
+         vk_zalloc(&device->vk.alloc, sizeof(**point),
+                   8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
+      if (!(*point))
+         result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+      if (result == VK_SUCCESS) {
+         result = anv_device_alloc_bo(device, 4096,
+                                      ANV_BO_ALLOC_EXTERNAL |
+                                      ANV_BO_ALLOC_IMPLICIT_SYNC,
+                                      0 /* explicit_address */,
+                                      &(*point)->bo);
+         if (result != VK_SUCCESS)
+            vk_free(&device->vk.alloc, *point);
+      }
+   } else {
+      *point = list_first_entry(&timeline->free_points,
+                                struct anv_timeline_point, link);
+      list_del(&(*point)->link);
+   }
+
+   if (result == VK_SUCCESS) {
+      (*point)->serial = value;
+      list_addtail(&(*point)->link, &timeline->points);
+   }
+
+   return result;
+}
+
+static VkResult
+anv_timeline_gc_locked(struct anv_device *device,
+                       struct anv_timeline *timeline)
+{
+   list_for_each_entry_safe(struct anv_timeline_point, point,
+                            &timeline->points, link) {
+      /* timeline->higest_pending is only incremented once submission has
+       * happened. If this point has a greater serial, it means the point
+       * hasn't been submitted yet.
+       */
+      if (point->serial > timeline->highest_pending)
+         return VK_SUCCESS;
+
+      /* If someone is waiting on this time point, consider it busy and don't
+       * try to recycle it. There's a slim possibility that it's no longer
+       * busy by the time we look at it but we would be recycling it out from
+       * under a waiter and that can lead to weird races.
+       *
+       * We walk the list in-order so if this time point is still busy so is
+       * every following time point
+       */
+      assert(point->waiting >= 0);
+      if (point->waiting)
+         return VK_SUCCESS;
+
+      /* Garbage collect any signaled point. */
+      VkResult result = anv_device_bo_busy(device, point->bo);
+      if (result == VK_NOT_READY) {
+         /* We walk the list in-order so if this time point is still busy so
+          * is every following time point
+          */
+         return VK_SUCCESS;
+      } else if (result != VK_SUCCESS) {
+         return result;
+      }
+
+      assert(timeline->highest_past < point->serial);
+      timeline->highest_past = point->serial;
+
+      list_del(&point->link);
+      list_add(&point->link, &timeline->free_points);
+   }
+
+   return VK_SUCCESS;
+}
+
+static VkResult anv_queue_submit_add_fence_bo(struct anv_queue_submit *submit,
+                                              struct anv_bo *bo,
+                                              bool signal);
+
+static VkResult
+anv_queue_submit_timeline_locked(struct anv_queue *queue,
+                                 struct anv_queue_submit *submit)
+{
+   VkResult result;
+
+   for (uint32_t i = 0; i < submit->wait_timeline_count; i++) {
+      struct anv_timeline *timeline = submit->wait_timelines[i];
+      uint64_t wait_value = submit->wait_timeline_values[i];
+
+      if (timeline->highest_past >= wait_value)
+         continue;
+
+      list_for_each_entry(struct anv_timeline_point, point, &timeline->points, link) {
+         if (point->serial < wait_value)
+            continue;
+         result = anv_queue_submit_add_fence_bo(submit, point->bo, false);
+         if (result != VK_SUCCESS)
+            return result;
+         break;
+      }
+   }
+   for (uint32_t i = 0; i < submit->signal_timeline_count; i++) {
+      struct anv_timeline *timeline = submit->signal_timelines[i];
+      uint64_t signal_value = submit->signal_timeline_values[i];
+      struct anv_timeline_point *point;
+
+      result = anv_timeline_add_point_locked(queue->device, timeline,
+                                             signal_value, &point);
+      if (result != VK_SUCCESS)
+         return result;
+
+      result = anv_queue_submit_add_fence_bo(submit, point->bo, true);
+      if (result != VK_SUCCESS)
+         return result;
+   }
+
+   result = anv_queue_execbuf_locked(queue, submit);
 
    if (result == VK_SUCCESS) {
+      /* Update the pending values in the timeline objects. */
+      for (uint32_t i = 0; i < submit->signal_timeline_count; i++) {
+         struct anv_timeline *timeline = submit->signal_timelines[i];
+         uint64_t signal_value = submit->signal_timeline_values[i];
+
+         assert(signal_value > timeline->highest_pending);
+         timeline->highest_pending = signal_value;
+      }
+
       /* Update signaled semaphores backed by syncfd. */
       for (uint32_t i = 0; i < submit->sync_fd_semaphore_count; i++) {
          struct anv_semaphore *semaphore = submit->sync_fd_semaphores[i];
@@ -119,26 +297,105 @@ _anv_queue_submit(struct anv_queue *queue, struct anv_queue_submit **_submit)
          struct anv_semaphore_impl *impl = &semaphore->permanent;
 
          assert(impl->type == ANV_SEMAPHORE_TYPE_SYNC_FILE);
-         impl->fd = dup(submit->out_fence);
+         impl->fd = os_dupfd_cloexec(submit->out_fence);
+      }
+   } else {
+      /* Unblock any waiter by signaling the points, the application will get
+       * a device lost error code.
+       */
+      for (uint32_t i = 0; i < submit->signal_timeline_count; i++) {
+         struct anv_timeline *timeline = submit->signal_timelines[i];
+         uint64_t signal_value = submit->signal_timeline_values[i];
+
+         assert(signal_value > timeline->highest_pending);
+         timeline->highest_past = timeline->highest_pending = signal_value;
       }
    }
 
    return result;
 }
 
+static VkResult
+anv_queue_submit_deferred_locked(struct anv_queue *queue, uint32_t *advance)
+{
+   VkResult result = VK_SUCCESS;
+
+   /* Go through all the queued submissions and submit then until we find one
+    * that's waiting on a point that hasn't materialized yet.
+    */
+   list_for_each_entry_safe(struct anv_queue_submit, submit,
+                            &queue->queued_submits, link) {
+      if (!anv_queue_submit_ready_locked(submit))
+         break;
+
+      (*advance)++;
+      list_del(&submit->link);
+
+      result = anv_queue_submit_timeline_locked(queue, submit);
+
+      anv_queue_submit_free(queue->device, submit);
+
+      if (result != VK_SUCCESS)
+         break;
+   }
+
+   return result;
+}
+
+static VkResult
+anv_device_submit_deferred_locked(struct anv_device *device)
+{
+   uint32_t advance = 0;
+   return anv_queue_submit_deferred_locked(&device->queue, &advance);
+}
+
+static VkResult
+_anv_queue_submit(struct anv_queue *queue, struct anv_queue_submit **_submit,
+                  bool flush_queue)
+{
+   struct anv_queue_submit *submit = *_submit;
+
+   /* Wait before signal behavior means we might keep alive the
+    * anv_queue_submit object a bit longer, so transfer the ownership to the
+    * anv_queue.
+    */
+   *_submit = NULL;
+
+   pthread_mutex_lock(&queue->device->mutex);
+   list_addtail(&submit->link, &queue->queued_submits);
+   VkResult result = anv_device_submit_deferred_locked(queue->device);
+   if (flush_queue) {
+      while (result == VK_SUCCESS && !list_is_empty(&queue->queued_submits)) {
+         int ret = pthread_cond_wait(&queue->device->queue_submit,
+                                     &queue->device->mutex);
+         if (ret != 0) {
+            result = anv_device_set_lost(queue->device, "wait timeout");
+            break;
+         }
+
+         result = anv_device_submit_deferred_locked(queue->device);
+      }
+   }
+   pthread_mutex_unlock(&queue->device->mutex);
+   return result;
+}
+
 VkResult
 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
 {
-   queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
+   vk_object_base_init(&device->vk, &queue->base, VK_OBJECT_TYPE_QUEUE);
    queue->device = device;
    queue->flags = 0;
 
+   list_inithead(&queue->queued_submits);
+
    return VK_SUCCESS;
 }
 
 void
 anv_queue_finish(struct anv_queue *queue)
 {
+   vk_object_base_finish(&queue->base);
 }
 
 static VkResult
@@ -218,11 +475,81 @@ anv_queue_submit_add_sync_fd_fence(struct anv_queue_submit *submit,
    return VK_SUCCESS;
 }
 
+static VkResult
+anv_queue_submit_add_timeline_wait(struct anv_queue_submit* submit,
+                                   struct anv_device *device,
+                                   struct anv_timeline *timeline,
+                                   uint64_t value)
+{
+   if (submit->wait_timeline_count >= submit->wait_timeline_array_length) {
+      uint32_t new_len = MAX2(submit->wait_timeline_array_length * 2, 64);
+
+      submit->wait_timelines =
+         vk_realloc(submit->alloc,
+                    submit->wait_timelines, new_len * sizeof(*submit->wait_timelines),
+                    8, submit->alloc_scope);
+      if (submit->wait_timelines == NULL)
+         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+      submit->wait_timeline_values =
+         vk_realloc(submit->alloc,
+                    submit->wait_timeline_values, new_len * sizeof(*submit->wait_timeline_values),
+                    8, submit->alloc_scope);
+      if (submit->wait_timeline_values == NULL)
+         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+      submit->wait_timeline_array_length = new_len;
+   }
+
+   submit->wait_timelines[submit->wait_timeline_count] = timeline;
+   submit->wait_timeline_values[submit->wait_timeline_count] = value;
+
+   submit->wait_timeline_count++;
+
+   return VK_SUCCESS;
+}
+
+static VkResult
+anv_queue_submit_add_timeline_signal(struct anv_queue_submit* submit,
+                                     struct anv_device *device,
+                                     struct anv_timeline *timeline,
+                                     uint64_t value)
+{
+   assert(timeline->highest_pending < value);
+
+   if (submit->signal_timeline_count >= submit->signal_timeline_array_length) {
+      uint32_t new_len = MAX2(submit->signal_timeline_array_length * 2, 64);
+
+      submit->signal_timelines =
+         vk_realloc(submit->alloc,
+                    submit->signal_timelines, new_len * sizeof(*submit->signal_timelines),
+                    8, submit->alloc_scope);
+      if (submit->signal_timelines == NULL)
+            return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+      submit->signal_timeline_values =
+         vk_realloc(submit->alloc,
+                    submit->signal_timeline_values, new_len * sizeof(*submit->signal_timeline_values),
+                    8, submit->alloc_scope);
+      if (submit->signal_timeline_values == NULL)
+         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+      submit->signal_timeline_array_length = new_len;
+   }
+
+   submit->signal_timelines[submit->signal_timeline_count] = timeline;
+   submit->signal_timeline_values[submit->signal_timeline_count] = value;
+
+   submit->signal_timeline_count++;
+
+   return VK_SUCCESS;
+}
+
 static struct anv_queue_submit *
-anv_queue_submit_alloc(struct anv_device *device)
+anv_queue_submit_alloc(struct anv_device *device, int perf_query_pass)
 {
-   const VkAllocationCallbacks *alloc = &device->alloc;
-   VkSystemAllocationScope alloc_scope = VK_SYSTEM_ALLOCATION_SCOPE_COMMAND;
+   const VkAllocationCallbacks *alloc = &device->vk.alloc;
+   VkSystemAllocationScope alloc_scope = VK_SYSTEM_ALLOCATION_SCOPE_DEVICE;
 
    struct anv_queue_submit *submit = vk_zalloc(alloc, sizeof(*submit), 8, alloc_scope);
    if (!submit)
@@ -232,6 +559,7 @@ anv_queue_submit_alloc(struct anv_device *device)
    submit->alloc_scope = alloc_scope;
    submit->in_fence = -1;
    submit->out_fence = -1;
+   submit->perf_query_pass = perf_query_pass;
 
    return submit;
 }
@@ -240,12 +568,15 @@ VkResult
 anv_queue_submit_simple_batch(struct anv_queue *queue,
                               struct anv_batch *batch)
 {
+   if (queue->device->no_hw)
+      return VK_SUCCESS;
+
    struct anv_device *device = queue->device;
-   struct anv_queue_submit *submit = anv_queue_submit_alloc(device);
+   struct anv_queue_submit *submit = anv_queue_submit_alloc(device, -1);
    if (!submit)
       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 
-   bool has_syncobj_wait = device->instance->physicalDevice.has_syncobj_wait;
+   bool has_syncobj_wait = device->physical->has_syncobj_wait;
    VkResult result;
    uint32_t syncobj;
    struct anv_bo *batch_bo, *sync_bo;
@@ -263,6 +594,7 @@ anv_queue_submit_simple_batch(struct anv_queue *queue,
       result = anv_device_alloc_bo(device, 4096,
                                    ANV_BO_ALLOC_EXTERNAL |
                                    ANV_BO_ALLOC_IMPLICIT_SYNC,
+                                   0 /* explicit_address */,
                                    &sync_bo);
       if (result != VK_SUCCESS)
          goto err_free_submit;
@@ -287,7 +619,7 @@ anv_queue_submit_simple_batch(struct anv_queue *queue,
       submit->simple_bo_size = size;
    }
 
-   result = _anv_queue_submit(queue, &submit);
+   result = _anv_queue_submit(queue, &submit, true);
 
    if (result == VK_SUCCESS) {
       if (has_syncobj_wait) {
@@ -338,6 +670,9 @@ maybe_transfer_temporary_semaphore(struct anv_queue_submit *submit,
       return VK_SUCCESS;
    }
 
+   /* BO backed timeline semaphores cannot be temporary. */
+   assert(impl->type != ANV_SEMAPHORE_TYPE_TIMELINE);
+
    /*
     * There is a requirement to reset semaphore to their permanent state after
     * submission. From the Vulkan 1.0.53 spec:
@@ -387,12 +722,14 @@ anv_queue_submit(struct anv_queue *queue,
                  const VkSemaphore *out_semaphores,
                  const uint64_t *out_values,
                  uint32_t num_out_semaphores,
-                 VkFence _fence)
+                 struct anv_bo *wsi_signal_bo,
+                 VkFence _fence,
+                 int perf_query_pass)
 {
    ANV_FROM_HANDLE(anv_fence, fence, _fence);
    struct anv_device *device = queue->device;
-   UNUSED struct anv_physical_device *pdevice = &device->instance->physicalDevice;
-   struct anv_queue_submit *submit = anv_queue_submit_alloc(device);
+   UNUSED struct anv_physical_device *pdevice = device->physical;
+   struct anv_queue_submit *submit = anv_queue_submit_alloc(device, perf_query_pass);
    if (!submit)
       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 
@@ -416,6 +753,18 @@ anv_queue_submit(struct anv_queue *queue,
             goto error;
          break;
 
+      case ANV_SEMAPHORE_TYPE_WSI_BO:
+         /* When using a window-system buffer as a semaphore, always enable
+          * EXEC_OBJECT_WRITE.  This gives us a WaR hazard with the display or
+          * compositor's read of the buffer and enforces that we don't start
+          * rendering until they are finished.  This is exactly the
+          * synchronization we want with vkAcquireNextImage.
+          */
+         result = anv_queue_submit_add_fence_bo(submit, impl->bo, true /* signal */);
+         if (result != VK_SUCCESS)
+            goto error;
+         break;
+
       case ANV_SEMAPHORE_TYPE_SYNC_FILE:
          assert(!pdevice->has_syncobj);
          if (submit->in_fence == -1) {
@@ -447,6 +796,14 @@ anv_queue_submit(struct anv_queue *queue,
          break;
       }
 
+      case ANV_SEMAPHORE_TYPE_TIMELINE:
+         result = anv_queue_submit_add_timeline_wait(submit, device,
+                                                     &impl->timeline,
+                                                     in_values ? in_values[i] : 0);
+         if (result != VK_SUCCESS)
+            goto error;
+         break;
+
       default:
          break;
       }
@@ -493,11 +850,25 @@ anv_queue_submit(struct anv_queue *queue,
          break;
       }
 
+      case ANV_SEMAPHORE_TYPE_TIMELINE:
+         result = anv_queue_submit_add_timeline_signal(submit, device,
+                                                       &impl->timeline,
+                                                       out_values ? out_values[i] : 0);
+         if (result != VK_SUCCESS)
+            goto error;
+         break;
+
       default:
          break;
       }
    }
 
+   if (wsi_signal_bo) {
+      result = anv_queue_submit_add_fence_bo(submit, wsi_signal_bo, true /* signal */);
+      if (result != VK_SUCCESS)
+         goto error;
+   }
+
    if (fence) {
       /* Under most circumstances, out fences won't be temporary.  However,
        * the spec does allow it for opaque_fd.  From the Vulkan 1.0.53 spec:
@@ -539,12 +910,21 @@ anv_queue_submit(struct anv_queue *queue,
       }
    }
 
-   result = _anv_queue_submit(queue, &submit);
+   result = _anv_queue_submit(queue, &submit, false);
    if (result != VK_SUCCESS)
       goto error;
 
    if (fence && fence->permanent.type == ANV_FENCE_TYPE_BO) {
-      /* BO fences can't be shared, so they can't be temporary. */
+      /* If we have permanent BO fence, the only type of temporary possible
+       * would be BO_WSI (because BO fences are not shareable). The Vulkan spec
+       * also requires that the fence passed to vkQueueSubmit() be :
+       *
+       *    * unsignaled
+       *    * not be associated with any other queue command that has not yet
+       *      completed execution on that queue
+       *
+       * So the only acceptable type for the temporary is NONE.
+       */
       assert(fence->temporary.type == ANV_FENCE_TYPE_NONE);
 
       /* Once the execbuf has returned, we need to set the fence state to
@@ -575,6 +955,9 @@ VkResult anv_QueueSubmit(
 {
    ANV_FROM_HANDLE(anv_queue, queue, _queue);
 
+   if (queue->device->no_hw)
+      return VK_SUCCESS;
+
    /* Query for device status prior to submitting.  Technically, we don't need
     * to do this.  However, if we have a client that's submitting piles of
     * garbage, we would rather break as early as possible to keep the GPU
@@ -592,7 +975,8 @@ VkResult anv_QueueSubmit(
        * come up with something more efficient but this shouldn't be a
        * common case.
        */
-      result = anv_queue_submit(queue, NULL, NULL, NULL, 0, NULL, NULL, 0, fence);
+      result = anv_queue_submit(queue, NULL, NULL, NULL, 0, NULL, NULL, 0,
+                                NULL, fence, -1);
       goto out;
    }
 
@@ -600,9 +984,19 @@ VkResult anv_QueueSubmit(
       /* Fence for this submit.  NULL for all but the last one */
       VkFence submit_fence = (i == submitCount - 1) ? fence : VK_NULL_HANDLE;
 
+      const struct wsi_memory_signal_submit_info *mem_signal_info =
+         vk_find_struct_const(pSubmits[i].pNext,
+                              WSI_MEMORY_SIGNAL_SUBMIT_INFO_MESA);
+      struct anv_bo *wsi_signal_bo =
+         mem_signal_info && mem_signal_info->memory != VK_NULL_HANDLE ?
+         anv_device_memory_from_handle(mem_signal_info->memory)->bo : NULL;
+
       const VkTimelineSemaphoreSubmitInfoKHR *timeline_info =
          vk_find_struct_const(pSubmits[i].pNext,
                               TIMELINE_SEMAPHORE_SUBMIT_INFO_KHR);
+      const VkPerformanceQuerySubmitInfoKHR *perf_info =
+         vk_find_struct_const(pSubmits[i].pNext,
+                              PERFORMANCE_QUERY_SUBMIT_INFO_KHR);
       const uint64_t *wait_values =
          timeline_info && timeline_info->waitSemaphoreValueCount ?
          timeline_info->pWaitSemaphoreValues : NULL;
@@ -623,7 +1017,9 @@ VkResult anv_QueueSubmit(
                                    pSubmits[i].pSignalSemaphores,
                                    signal_values,
                                    pSubmits[i].signalSemaphoreCount,
-                                   submit_fence);
+                                   wsi_signal_bo,
+                                   submit_fence,
+                                   -1);
          if (result != VK_SUCCESS)
             goto out;
 
@@ -661,7 +1057,8 @@ VkResult anv_QueueSubmit(
          result = anv_queue_submit(queue, cmd_buffer,
                                    in_semaphores, in_values, num_in_semaphores,
                                    out_semaphores, out_values, num_out_semaphores,
-                                   execbuf_fence);
+                                   wsi_signal_bo, execbuf_fence,
+                                   perf_info ? perf_info->counterPassIndex : 0);
          if (result != VK_SUCCESS)
             goto out;
       }
@@ -713,12 +1110,14 @@ VkResult anv_CreateFence(
 
    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO);
 
-   fence = vk_zalloc2(&device->alloc, pAllocator, sizeof(*fence), 8,
+   fence = vk_zalloc2(&device->vk.alloc, pAllocator, sizeof(*fence), 8,
                       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
    if (fence == NULL)
       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 
-   if (device->instance->physicalDevice.has_syncobj_wait) {
+   vk_object_base_init(&device->vk, &fence->base, VK_OBJECT_TYPE_FENCE);
+
+   if (device->physical->has_syncobj_wait) {
       fence->permanent.type = ANV_FENCE_TYPE_SYNCOBJ;
 
       uint32_t create_flags = 0;
@@ -761,6 +1160,10 @@ anv_fence_impl_cleanup(struct anv_device *device,
       anv_bo_pool_free(&device->batch_bo_pool, impl->bo.bo);
       break;
 
+   case ANV_FENCE_TYPE_WSI_BO:
+      anv_device_release_bo(device, impl->bo.bo);
+      break;
+
    case ANV_FENCE_TYPE_SYNCOBJ:
       anv_gem_syncobj_destroy(device, impl->syncobj);
       break;
@@ -776,6 +1179,16 @@ anv_fence_impl_cleanup(struct anv_device *device,
    impl->type = ANV_FENCE_TYPE_NONE;
 }
 
+void
+anv_fence_reset_temporary(struct anv_device *device,
+                          struct anv_fence *fence)
+{
+   if (fence->temporary.type == ANV_FENCE_TYPE_NONE)
+      return;
+
+   anv_fence_impl_cleanup(device, &fence->temporary);
+}
+
 void anv_DestroyFence(
     VkDevice                                    _device,
     VkFence                                     _fence,
@@ -790,7 +1203,8 @@ void anv_DestroyFence(
    anv_fence_impl_cleanup(device, &fence->temporary);
    anv_fence_impl_cleanup(device, &fence->permanent);
 
-   vk_free2(&device->alloc, pAllocator, fence);
+   vk_object_base_finish(&fence->base);
+   vk_free2(&device->vk.alloc, pAllocator, fence);
 }
 
 VkResult anv_ResetFences(
@@ -810,8 +1224,7 @@ VkResult anv_ResetFences(
        *    first restored. The remaining operations described therefore
        *    operate on the restored payload.
        */
-      if (fence->temporary.type != ANV_FENCE_TYPE_NONE)
-         anv_fence_impl_cleanup(device, &fence->temporary);
+      anv_fence_reset_temporary(device, fence);
 
       struct anv_fence_impl *impl = &fence->permanent;
 
@@ -848,8 +1261,7 @@ VkResult anv_GetFenceStatus(
 
    switch (impl->type) {
    case ANV_FENCE_TYPE_BO:
-      /* BO fences don't support import/export */
-      assert(fence->temporary.type == ANV_FENCE_TYPE_NONE);
+   case ANV_FENCE_TYPE_WSI_BO:
       switch (impl->bo.state) {
       case ANV_BO_FENCE_STATE_RESET:
          /* If it hasn't even been sent off to the GPU yet, it's not ready */
@@ -898,7 +1310,7 @@ anv_wait_for_syncobj_fences(struct anv_device *device,
                             bool waitAll,
                             uint64_t abs_timeout_ns)
 {
-   uint32_t *syncobjs = vk_zalloc(&device->alloc,
+   uint32_t *syncobjs = vk_zalloc(&device->vk.alloc,
                                   sizeof(*syncobjs) * fenceCount, 8,
                                   VK_SYSTEM_ALLOCATION_SCOPE_COMMAND);
    if (!syncobjs)
@@ -926,7 +1338,7 @@ anv_wait_for_syncobj_fences(struct anv_device *device,
                                  abs_timeout_ns, waitAll);
    } while (ret == -1 && errno == ETIME && anv_gettime_ns() < abs_timeout_ns);
 
-   vk_free(&device->alloc, syncobjs);
+   vk_free(&device->vk.alloc, syncobjs);
 
    if (ret == -1) {
       if (errno == ETIME) {
@@ -955,13 +1367,11 @@ anv_wait_for_bo_fences(struct anv_device *device,
       for (uint32_t i = 0; i < fenceCount; i++) {
          ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
 
-         /* This function assumes that all fences are BO fences and that they
-          * have no temporary state.  Since BO fences will never be exported,
-          * this should be a safe assumption.
-          */
-         assert(fence->permanent.type == ANV_FENCE_TYPE_BO);
-         assert(fence->temporary.type == ANV_FENCE_TYPE_NONE);
-         struct anv_fence_impl *impl = &fence->permanent;
+         struct anv_fence_impl *impl =
+            fence->temporary.type != ANV_FENCE_TYPE_NONE ?
+            &fence->temporary : &fence->permanent;
+         assert(impl->type == ANV_FENCE_TYPE_BO ||
+                impl->type == ANV_FENCE_TYPE_WSI_BO);
 
          switch (impl->bo.state) {
          case ANV_BO_FENCE_STATE_RESET:
@@ -1056,12 +1466,9 @@ done:
 
 static VkResult
 anv_wait_for_wsi_fence(struct anv_device *device,
-                       const VkFence _fence,
+                       struct anv_fence_impl *impl,
                        uint64_t abs_timeout)
 {
-   ANV_FROM_HANDLE(anv_fence, fence, _fence);
-   struct anv_fence_impl *impl = &fence->permanent;
-
    return impl->fence_wsi->wait(impl->fence_wsi, abs_timeout);
 }
 
@@ -1077,8 +1484,13 @@ anv_wait_for_fences(struct anv_device *device,
    if (fenceCount <= 1 || waitAll) {
       for (uint32_t i = 0; i < fenceCount; i++) {
          ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
-         switch (fence->permanent.type) {
+         struct anv_fence_impl *impl =
+            fence->temporary.type != ANV_FENCE_TYPE_NONE ?
+            &fence->temporary : &fence->permanent;
+
+         switch (impl->type) {
          case ANV_FENCE_TYPE_BO:
+         case ANV_FENCE_TYPE_WSI_BO:
             result = anv_wait_for_bo_fences(device, 1, &pFences[i],
                                             true, abs_timeout);
             break;
@@ -1087,7 +1499,7 @@ anv_wait_for_fences(struct anv_device *device,
                                                  true, abs_timeout);
             break;
          case ANV_FENCE_TYPE_WSI:
-            result = anv_wait_for_wsi_fence(device, pFences[i], abs_timeout);
+            result = anv_wait_for_wsi_fence(device, impl, abs_timeout);
             break;
          case ANV_FENCE_TYPE_NONE:
             result = VK_SUCCESS;
@@ -1112,7 +1524,10 @@ static bool anv_all_fences_syncobj(uint32_t fenceCount, const VkFence *pFences)
 {
    for (uint32_t i = 0; i < fenceCount; ++i) {
       ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
-      if (fence->permanent.type != ANV_FENCE_TYPE_SYNCOBJ)
+      struct anv_fence_impl *impl =
+         fence->temporary.type != ANV_FENCE_TYPE_NONE ?
+         &fence->temporary : &fence->permanent;
+      if (impl->type != ANV_FENCE_TYPE_SYNCOBJ)
          return false;
    }
    return true;
@@ -1122,7 +1537,11 @@ static bool anv_all_fences_bo(uint32_t fenceCount, const VkFence *pFences)
 {
    for (uint32_t i = 0; i < fenceCount; ++i) {
       ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
-      if (fence->permanent.type != ANV_FENCE_TYPE_BO)
+      struct anv_fence_impl *impl =
+         fence->temporary.type != ANV_FENCE_TYPE_NONE ?
+         &fence->temporary : &fence->permanent;
+      if (impl->type != ANV_FENCE_TYPE_BO &&
+          impl->type != ANV_FENCE_TYPE_WSI_BO)
          return false;
    }
    return true;
@@ -1137,6 +1556,9 @@ VkResult anv_WaitForFences(
 {
    ANV_FROM_HANDLE(anv_device, device, _device);
 
+   if (device->no_hw)
+      return VK_SUCCESS;
+
    if (anv_device_is_lost(device))
       return VK_ERROR_DEVICE_LOST;
 
@@ -1211,24 +1633,35 @@ VkResult anv_ImportFenceFdKHR(
 
       break;
 
-   case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT:
+   case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT: {
       /* Sync files are a bit tricky.  Because we want to continue using the
        * syncobj implementation of WaitForFences, we don't use the sync file
        * directly but instead import it into a syncobj.
        */
       new_impl.type = ANV_FENCE_TYPE_SYNCOBJ;
 
-      new_impl.syncobj = anv_gem_syncobj_create(device, 0);
+      /* "If handleType is VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT, the
+       *  special value -1 for fd is treated like a valid sync file descriptor
+       *  referring to an object that has already signaled. The import
+       *  operation will succeed and the VkFence will have a temporarily
+       *  imported payload as if a valid file descriptor had been provided."
+       */
+      uint32_t create_flags = 0;
+      if (fd == -1)
+         create_flags |= DRM_SYNCOBJ_CREATE_SIGNALED;
+
+      new_impl.syncobj = anv_gem_syncobj_create(device, create_flags);
       if (!new_impl.syncobj)
          return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 
-      if (anv_gem_syncobj_import_sync_file(device, new_impl.syncobj, fd)) {
+      if (fd != -1 &&
+          anv_gem_syncobj_import_sync_file(device, new_impl.syncobj, fd)) {
          anv_gem_syncobj_destroy(device, new_impl.syncobj);
-         return vk_errorf(device->instance, NULL,
-                          VK_ERROR_INVALID_EXTERNAL_HANDLE,
+         return vk_errorf(device, NULL, VK_ERROR_INVALID_EXTERNAL_HANDLE,
                           "syncobj sync file import failed: %m");
       }
       break;
+   }
 
    default:
       return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
@@ -1309,6 +1742,57 @@ VkResult anv_GetFenceFdKHR(
 
 // Queue semaphore functions
 
+static VkSemaphoreTypeKHR
+get_semaphore_type(const void *pNext, uint64_t *initial_value)
+{
+   const VkSemaphoreTypeCreateInfoKHR *type_info =
+      vk_find_struct_const(pNext, SEMAPHORE_TYPE_CREATE_INFO_KHR);
+
+   if (!type_info)
+      return VK_SEMAPHORE_TYPE_BINARY_KHR;
+
+   if (initial_value)
+      *initial_value = type_info->initialValue;
+   return type_info->semaphoreType;
+}
+
+static VkResult
+binary_semaphore_create(struct anv_device *device,
+                        struct anv_semaphore_impl *impl,
+                        bool exportable)
+{
+   if (device->physical->has_syncobj) {
+      impl->type = ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ;
+      impl->syncobj = anv_gem_syncobj_create(device, 0);
+      if (!impl->syncobj)
+            return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+      return VK_SUCCESS;
+   } else {
+      impl->type = ANV_SEMAPHORE_TYPE_BO;
+      VkResult result =
+         anv_device_alloc_bo(device, 4096,
+                             ANV_BO_ALLOC_EXTERNAL |
+                             ANV_BO_ALLOC_IMPLICIT_SYNC,
+                             0 /* explicit_address */,
+                             &impl->bo);
+      /* If we're going to use this as a fence, we need to *not* have the
+       * EXEC_OBJECT_ASYNC bit set.
+       */
+      assert(!(impl->bo->flags & EXEC_OBJECT_ASYNC));
+      return result;
+   }
+}
+
+static VkResult
+timeline_semaphore_create(struct anv_device *device,
+                          struct anv_semaphore_impl *impl,
+                          uint64_t initial_value)
+{
+   impl->type = ANV_SEMAPHORE_TYPE_TIMELINE;
+   anv_timeline_init(device, &impl->timeline, initial_value);
+   return VK_SUCCESS;
+}
+
 VkResult anv_CreateSemaphore(
     VkDevice                                    _device,
     const VkSemaphoreCreateInfo*                pCreateInfo,
@@ -1320,61 +1804,58 @@ VkResult anv_CreateSemaphore(
 
    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO);
 
-   semaphore = vk_alloc(&device->alloc, sizeof(*semaphore), 8,
+   uint64_t timeline_value = 0;
+   VkSemaphoreTypeKHR sem_type = get_semaphore_type(pCreateInfo->pNext, &timeline_value);
+
+   semaphore = vk_alloc(&device->vk.alloc, sizeof(*semaphore), 8,
                         VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
    if (semaphore == NULL)
       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 
+   vk_object_base_init(&device->vk, &semaphore->base, VK_OBJECT_TYPE_SEMAPHORE);
+
    p_atomic_set(&semaphore->refcount, 1);
 
    const VkExportSemaphoreCreateInfo *export =
       vk_find_struct_const(pCreateInfo->pNext, EXPORT_SEMAPHORE_CREATE_INFO);
     VkExternalSemaphoreHandleTypeFlags handleTypes =
       export ? export->handleTypes : 0;
+   VkResult result;
 
    if (handleTypes == 0) {
-      /* The DRM execbuffer ioctl always execute in-oder so long as you stay
-       * on the same ring.  Since we don't expose the blit engine as a DMA
-       * queue, a dummy no-op semaphore is a perfectly valid implementation.
-       */
-      semaphore->permanent.type = ANV_SEMAPHORE_TYPE_DUMMY;
+      if (sem_type == VK_SEMAPHORE_TYPE_BINARY_KHR)
+         result = binary_semaphore_create(device, &semaphore->permanent, false);
+      else
+         result = timeline_semaphore_create(device, &semaphore->permanent, timeline_value);
+      if (result != VK_SUCCESS) {
+         vk_free2(&device->vk.alloc, pAllocator, semaphore);
+         return result;
+      }
    } else if (handleTypes & VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT) {
       assert(handleTypes == VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT);
-      if (device->instance->physicalDevice.has_syncobj) {
-         semaphore->permanent.type = ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ;
-         semaphore->permanent.syncobj = anv_gem_syncobj_create(device, 0);
-         if (!semaphore->permanent.syncobj) {
-            vk_free2(&device->alloc, pAllocator, semaphore);
-            return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
-         }
-      } else {
-         semaphore->permanent.type = ANV_SEMAPHORE_TYPE_BO;
-         VkResult result = anv_device_alloc_bo(device, 4096,
-                                               ANV_BO_ALLOC_EXTERNAL |
-                                               ANV_BO_ALLOC_IMPLICIT_SYNC,
-                                               &semaphore->permanent.bo);
-         if (result != VK_SUCCESS) {
-            vk_free2(&device->alloc, pAllocator, semaphore);
-            return result;
-         }
-
-         /* If we're going to use this as a fence, we need to *not* have the
-          * EXEC_OBJECT_ASYNC bit set.
-          */
-         assert(!(semaphore->permanent.bo->flags & EXEC_OBJECT_ASYNC));
+      assert(sem_type == VK_SEMAPHORE_TYPE_BINARY_KHR);
+      result = binary_semaphore_create(device, &semaphore->permanent, true);
+      if (result != VK_SUCCESS) {
+         vk_free2(&device->vk.alloc, pAllocator, semaphore);
+         return result;
       }
    } else if (handleTypes & VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT) {
       assert(handleTypes == VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT);
-      if (device->instance->physicalDevice.has_syncobj) {
+      assert(sem_type == VK_SEMAPHORE_TYPE_BINARY_KHR);
+      if (device->physical->has_syncobj) {
          semaphore->permanent.type = ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ;
          semaphore->permanent.syncobj = anv_gem_syncobj_create(device, 0);
+         if (!semaphore->permanent.syncobj) {
+            vk_free2(&device->vk.alloc, pAllocator, semaphore);
+            return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+         }
       } else {
          semaphore->permanent.type = ANV_SEMAPHORE_TYPE_SYNC_FILE;
          semaphore->permanent.fd = -1;
       }
    } else {
       assert(!"Unknown handle type");
-      vk_free2(&device->alloc, pAllocator, semaphore);
+      vk_free2(&device->vk.alloc, pAllocator, semaphore);
       return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
    }
 
@@ -1396,11 +1877,17 @@ anv_semaphore_impl_cleanup(struct anv_device *device,
       break;
 
    case ANV_SEMAPHORE_TYPE_BO:
+   case ANV_SEMAPHORE_TYPE_WSI_BO:
       anv_device_release_bo(device, impl->bo);
       break;
 
    case ANV_SEMAPHORE_TYPE_SYNC_FILE:
-      close(impl->fd);
+      if (impl->fd >= 0)
+         close(impl->fd);
+      break;
+
+   case ANV_SEMAPHORE_TYPE_TIMELINE:
+      anv_timeline_finish(device, &impl->timeline);
       break;
 
    case ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ:
@@ -1440,7 +1927,9 @@ anv_semaphore_unref(struct anv_device *device, struct anv_semaphore *semaphore)
 
    anv_semaphore_impl_cleanup(device, &semaphore->temporary);
    anv_semaphore_impl_cleanup(device, &semaphore->permanent);
-   vk_free(&device->alloc, semaphore);
+
+   vk_object_base_finish(&semaphore->base);
+   vk_free(&device->vk.alloc, semaphore);
 }
 
 void anv_DestroySemaphore(
@@ -1464,8 +1953,14 @@ void anv_GetPhysicalDeviceExternalSemaphoreProperties(
 {
    ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
 
+   VkSemaphoreTypeKHR sem_type =
+      get_semaphore_type(pExternalSemaphoreInfo->pNext, NULL);
+
    switch (pExternalSemaphoreInfo->handleType) {
    case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT:
+      /* Timeline semaphores are not exportable. */
+      if (sem_type == VK_SEMAPHORE_TYPE_TIMELINE_KHR)
+         break;
       pExternalSemaphoreProperties->exportFromImportedHandleTypes =
          VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT;
       pExternalSemaphoreProperties->compatibleHandleTypes =
@@ -1476,17 +1971,18 @@ void anv_GetPhysicalDeviceExternalSemaphoreProperties(
       return;
 
    case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT:
-      if (device->has_exec_fence) {
-         pExternalSemaphoreProperties->exportFromImportedHandleTypes =
-            VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT;
-         pExternalSemaphoreProperties->compatibleHandleTypes =
-            VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT;
-         pExternalSemaphoreProperties->externalSemaphoreFeatures =
-            VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT |
-            VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT;
-         return;
-      }
-      break;
+      if (sem_type == VK_SEMAPHORE_TYPE_TIMELINE_KHR)
+         break;
+      if (!device->has_exec_fence)
+         break;
+      pExternalSemaphoreProperties->exportFromImportedHandleTypes =
+         VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT;
+      pExternalSemaphoreProperties->compatibleHandleTypes =
+         VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT;
+      pExternalSemaphoreProperties->externalSemaphoreFeatures =
+         VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT |
+         VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT;
+      return;
 
    default:
       break;
@@ -1511,7 +2007,7 @@ VkResult anv_ImportSemaphoreFdKHR(
 
    switch (pImportSemaphoreFdInfo->handleType) {
    case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT:
-      if (device->instance->physicalDevice.has_syncobj) {
+      if (device->physical->has_syncobj) {
          new_impl.type = ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ;
 
          new_impl.syncobj = anv_gem_syncobj_fd_to_handle(device, fd);
@@ -1523,6 +2019,7 @@ VkResult anv_ImportSemaphoreFdKHR(
          VkResult result = anv_device_import_bo(device, fd,
                                                 ANV_BO_ALLOC_EXTERNAL |
                                                 ANV_BO_ALLOC_IMPLICIT_SYNC,
+                                                0 /* client_address */,
                                                 &new_impl.bo);
          if (result != VK_SUCCESS)
             return result;
@@ -1551,7 +2048,7 @@ VkResult anv_ImportSemaphoreFdKHR(
       break;
 
    case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT:
-      if (device->instance->physicalDevice.has_syncobj) {
+      if (device->physical->has_syncobj) {
          new_impl = (struct anv_semaphore_impl) {
             .type = ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ,
             .syncobj = anv_gem_syncobj_create(device, 0),
@@ -1560,8 +2057,7 @@ VkResult anv_ImportSemaphoreFdKHR(
             return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
          if (anv_gem_syncobj_import_sync_file(device, new_impl.syncobj, fd)) {
             anv_gem_syncobj_destroy(device, new_impl.syncobj);
-            return vk_errorf(device->instance, NULL,
-                             VK_ERROR_INVALID_EXTERNAL_HANDLE,
+            return vk_errorf(device, NULL, VK_ERROR_INVALID_EXTERNAL_HANDLE,
                              "syncobj sync file import failed: %m");
          }
          /* Ownership of the FD is transfered to Anv. Since we don't need it
@@ -1684,3 +2180,226 @@ VkResult anv_GetSemaphoreFdKHR(
 
    return VK_SUCCESS;
 }
+
+VkResult anv_GetSemaphoreCounterValue(
+    VkDevice                                    _device,
+    VkSemaphore                                 _semaphore,
+    uint64_t*                                   pValue)
+{
+   ANV_FROM_HANDLE(anv_device, device, _device);
+   ANV_FROM_HANDLE(anv_semaphore, semaphore, _semaphore);
+
+   struct anv_semaphore_impl *impl =
+      semaphore->temporary.type != ANV_SEMAPHORE_TYPE_NONE ?
+      &semaphore->temporary : &semaphore->permanent;
+
+   switch (impl->type) {
+   case ANV_SEMAPHORE_TYPE_TIMELINE: {
+      pthread_mutex_lock(&device->mutex);
+      anv_timeline_gc_locked(device, &impl->timeline);
+      *pValue = impl->timeline.highest_past;
+      pthread_mutex_unlock(&device->mutex);
+      return VK_SUCCESS;
+   }
+
+   default:
+      unreachable("Invalid semaphore type");
+   }
+}
+
+static VkResult
+anv_timeline_wait_locked(struct anv_device *device,
+                         struct anv_timeline *timeline,
+                         uint64_t serial, uint64_t abs_timeout_ns)
+{
+   /* Wait on the queue_submit condition variable until the timeline has a
+    * time point pending that's at least as high as serial.
+    */
+   while (timeline->highest_pending < serial) {
+      struct timespec abstime = {
+         .tv_sec = abs_timeout_ns / NSEC_PER_SEC,
+         .tv_nsec = abs_timeout_ns % NSEC_PER_SEC,
+      };
+
+      int ret = pthread_cond_timedwait(&device->queue_submit,
+                                       &device->mutex, &abstime);
+      assert(ret != EINVAL);
+      if (anv_gettime_ns() >= abs_timeout_ns &&
+          timeline->highest_pending < serial)
+         return VK_TIMEOUT;
+   }
+
+   while (1) {
+      VkResult result = anv_timeline_gc_locked(device, timeline);
+      if (result != VK_SUCCESS)
+         return result;
+
+      if (timeline->highest_past >= serial)
+         return VK_SUCCESS;
+
+      /* If we got here, our earliest time point has a busy BO */
+      struct anv_timeline_point *point =
+         list_first_entry(&timeline->points,
+                          struct anv_timeline_point, link);
+
+      /* Drop the lock while we wait. */
+      point->waiting++;
+      pthread_mutex_unlock(&device->mutex);
+
+      result = anv_device_wait(device, point->bo,
+                               anv_get_relative_timeout(abs_timeout_ns));
+
+      /* Pick the mutex back up */
+      pthread_mutex_lock(&device->mutex);
+      point->waiting--;
+
+      /* This covers both VK_TIMEOUT and VK_ERROR_DEVICE_LOST */
+      if (result != VK_SUCCESS)
+         return result;
+   }
+}
+
+static VkResult
+anv_timelines_wait(struct anv_device *device,
+                   struct anv_timeline **timelines,
+                   const uint64_t *serials,
+                   uint32_t n_timelines,
+                   bool wait_all,
+                   uint64_t abs_timeout_ns)
+{
+   if (!wait_all && n_timelines > 1) {
+      pthread_mutex_lock(&device->mutex);
+
+      while (1) {
+         VkResult result;
+         for (uint32_t i = 0; i < n_timelines; i++) {
+            result =
+               anv_timeline_wait_locked(device, timelines[i], serials[i], 0);
+            if (result != VK_TIMEOUT)
+               break;
+         }
+
+         if (result != VK_TIMEOUT ||
+             anv_gettime_ns() >= abs_timeout_ns) {
+            pthread_mutex_unlock(&device->mutex);
+            return result;
+         }
+
+         /* If none of them are ready do a short wait so we don't completely
+          * spin while holding the lock. The 10us is completely arbitrary.
+          */
+         uint64_t abs_short_wait_ns =
+            anv_get_absolute_timeout(
+               MIN2((anv_gettime_ns() - abs_timeout_ns) / 10, 10 * 1000));
+         struct timespec abstime = {
+            .tv_sec = abs_short_wait_ns / NSEC_PER_SEC,
+            .tv_nsec = abs_short_wait_ns % NSEC_PER_SEC,
+         };
+         ASSERTED int ret;
+         ret = pthread_cond_timedwait(&device->queue_submit,
+                                      &device->mutex, &abstime);
+         assert(ret != EINVAL);
+      }
+   } else {
+      VkResult result = VK_SUCCESS;
+      pthread_mutex_lock(&device->mutex);
+      for (uint32_t i = 0; i < n_timelines; i++) {
+         result =
+            anv_timeline_wait_locked(device, timelines[i],
+                                     serials[i], abs_timeout_ns);
+         if (result != VK_SUCCESS)
+            break;
+      }
+      pthread_mutex_unlock(&device->mutex);
+      return result;
+   }
+}
+
+VkResult anv_WaitSemaphores(
+    VkDevice                                    _device,
+    const VkSemaphoreWaitInfoKHR*               pWaitInfo,
+    uint64_t                                    timeout)
+{
+   ANV_FROM_HANDLE(anv_device, device, _device);
+
+   if (device->no_hw)
+      return VK_SUCCESS;
+
+   struct anv_timeline **timelines =
+      vk_alloc(&device->vk.alloc,
+               pWaitInfo->semaphoreCount * sizeof(*timelines),
+               8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND);
+   if (!timelines)
+      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+   uint64_t *values = vk_alloc(&device->vk.alloc,
+                               pWaitInfo->semaphoreCount * sizeof(*values),
+                               8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND);
+   if (!values) {
+      vk_free(&device->vk.alloc, timelines);
+      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+   }
+
+   uint32_t handle_count = 0;
+   for (uint32_t i = 0; i < pWaitInfo->semaphoreCount; i++) {
+      ANV_FROM_HANDLE(anv_semaphore, semaphore, pWaitInfo->pSemaphores[i]);
+      struct anv_semaphore_impl *impl =
+         semaphore->temporary.type != ANV_SEMAPHORE_TYPE_NONE ?
+         &semaphore->temporary : &semaphore->permanent;
+
+      assert(impl->type == ANV_SEMAPHORE_TYPE_TIMELINE);
+
+      if (pWaitInfo->pValues[i] == 0)
+         continue;
+
+      timelines[handle_count] = &impl->timeline;
+      values[handle_count] = pWaitInfo->pValues[i];
+      handle_count++;
+   }
+
+   VkResult result = VK_SUCCESS;
+   if (handle_count > 0) {
+      result = anv_timelines_wait(device, timelines, values, handle_count,
+                                  !(pWaitInfo->flags & VK_SEMAPHORE_WAIT_ANY_BIT_KHR),
+                                  anv_get_absolute_timeout(timeout));
+   }
+
+   vk_free(&device->vk.alloc, timelines);
+   vk_free(&device->vk.alloc, values);
+
+   return result;
+}
+
+VkResult anv_SignalSemaphore(
+    VkDevice                                    _device,
+    const VkSemaphoreSignalInfoKHR*             pSignalInfo)
+{
+   ANV_FROM_HANDLE(anv_device, device, _device);
+   ANV_FROM_HANDLE(anv_semaphore, semaphore, pSignalInfo->semaphore);
+
+   struct anv_semaphore_impl *impl =
+      semaphore->temporary.type != ANV_SEMAPHORE_TYPE_NONE ?
+      &semaphore->temporary : &semaphore->permanent;
+
+   switch (impl->type) {
+   case ANV_SEMAPHORE_TYPE_TIMELINE: {
+      pthread_mutex_lock(&device->mutex);
+
+      VkResult result = anv_timeline_gc_locked(device, &impl->timeline);
+
+      assert(pSignalInfo->value > impl->timeline.highest_pending);
+
+      impl->timeline.highest_pending = impl->timeline.highest_past = pSignalInfo->value;
+
+      if (result == VK_SUCCESS)
+         result = anv_device_submit_deferred_locked(device);
+
+      pthread_cond_broadcast(&device->queue_submit);
+      pthread_mutex_unlock(&device->mutex);
+      return result;
+   }
+
+   default:
+      unreachable("Invalid semaphore type");
+   }
+}