vmwgfx: Add fence events

Add a way to send DRM events down the gpu fifo by attaching them to
fence objects. This may be useful for Xserver swapbuffer throttling and
page-flip done notifications.

Bump version to 2.2 to signal the availability of the FENCE_EVENT ioctl.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Jakob Bornecrantz <jakob@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

1 files changed, 516 insertions, 17 deletions

diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
index 5f60be76166e..35d5f61fc7e4 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
@@ -34,13 +34,18 @@ struct vmw_fence_manager {
         int num_fence_objects;
         struct vmw_private *dev_priv;
         spinlock_t lock;
-        u32 next_seqno;
         struct list_head fence_list;
         struct work_struct work;
         u32 user_fence_size;
         u32 fence_size;
+        u32 event_fence_action_size;
         bool fifo_down;
         struct list_head cleanup_list;
+        uint32_t pending_actions[VMW_ACTION_MAX];
+        struct mutex goal_irq_mutex;
+        bool goal_irq_on; /* Protected by @goal_irq_mutex */
+        bool seqno_valid; /* Protected by @lock, and may not be set to true
+                             without the @goal_irq_mutex held. */
 };
 
 struct vmw_user_fence {
@@ -49,8 +54,51 @@ struct vmw_user_fence {
 };
 
 /**
- * vmw_fence_destroy_locked
+ * struct vmw_event_fence_action - fence action that delivers a drm event.
  *
+ * @e: A struct drm_pending_event that controls the event delivery.
+ * @action: A struct vmw_fence_action to hook up to a fence.
+ * @fence: A referenced pointer to the fence to keep it alive while @action
+ * hangs on it.
+ * @dev: Pointer to a struct drm_device so we can access the event stuff.
+ * @kref: Both @e and @action has destructors, so we need to refcount.
+ * @size: Size accounted for this object.
+ * @tv_sec: If non-null, the variable pointed to will be assigned
+ * current time tv_sec val when the fence signals.
+ * @tv_usec: Must be set if @tv_sec is set, and the variable pointed to will
+ * be assigned the current time tv_usec val when the fence signals.
+ */
+struct vmw_event_fence_action {
+        struct drm_pending_event e;
+        struct vmw_fence_action action;
+        struct vmw_fence_obj *fence;
+        struct drm_device *dev;
+        struct kref kref;
+        uint32_t size;
+        uint32_t *tv_sec;
+        uint32_t *tv_usec;
+};
+
+/**
+ * Note on fencing subsystem usage of irqs:
+ * Typically the vmw_fences_update function is called
+ *
+ * a) When a new fence seqno has been submitted by the fifo code.
+ * b) On-demand when we have waiters. Sleeping waiters will switch on the
+ * ANY_FENCE irq and call vmw_fences_update function each time an ANY_FENCE
+ * irq is received. When the last fence waiter is gone, that IRQ is masked
+ * away.
+ *
+ * In situations where there are no waiters and we don't submit any new fences,
+ * fence objects may not be signaled. This is perfectly OK, since there are
+ * no consumers of the signaled data, but that is NOT ok when there are fence
+ * actions attached to a fence. The fencing subsystem then makes use of the
+ * FENCE_GOAL irq and sets the fence goal seqno to that of the next fence
+ * which has an action attached, and each time vmw_fences_update is called,
+ * the subsystem makes sure the fence goal seqno is updated.
+ *
+ * The fence goal seqno irq is on as long as there are unsignaled fence
+ * objects with actions attached to them.
  */
 
 static void vmw_fence_obj_destroy_locked(struct kref *kref)
@@ -85,24 +133,36 @@ static void vmw_fence_work_func(struct work_struct *work)
                 container_of(work, struct vmw_fence_manager, work);
         struct list_head list;
         struct vmw_fence_action *action, *next_action;
+        bool seqno_valid;
 
         do {
                 INIT_LIST_HEAD(&list);
+                mutex_lock(&fman->goal_irq_mutex);
+
                 spin_lock_irq(&fman->lock);
                 list_splice_init(&fman->cleanup_list, &list);
+                seqno_valid = fman->seqno_valid;
                 spin_unlock_irq(&fman->lock);
 
+                if (!seqno_valid && fman->goal_irq_on) {
+                        fman->goal_irq_on = false;
+                        vmw_goal_waiter_remove(fman->dev_priv);
+                }
+                mutex_unlock(&fman->goal_irq_mutex);
+
                 if (list_empty(&list))
                         return;
 
                 /*
                  * At this point, only we should be able to manipulate the
                  * list heads of the actions we have on the private list.
+                 * hence fman::lock not held.
                  */
 
                 list_for_each_entry_safe(action, next_action, &list, head) {
                         list_del_init(&action->head);
-                        action->cleanup(action);
+                        if (action->cleanup)
+                                action->cleanup(action);
                 }
         } while (1);
 }
@@ -122,6 +182,9 @@ struct vmw_fence_manager *vmw_fence_manager_init(struct vmw_private *dev_priv)
         fman->fifo_down = true;
         fman->user_fence_size = ttm_round_pot(sizeof(struct vmw_user_fence));
         fman->fence_size = ttm_round_pot(sizeof(struct vmw_fence_obj));
+        fman->event_fence_action_size =
+                ttm_round_pot(sizeof(struct vmw_event_fence_action));
+        mutex_init(&fman->goal_irq_mutex);
 
         return fman;
 }
@@ -214,6 +277,7 @@ void vmw_fences_perform_actions(struct vmw_fence_manager *fman,
 
         list_for_each_entry_safe(action, next_action, list, head) {
                 list_del_init(&action->head);
+                fman->pending_actions[action->type]--;
                 if (action->seq_passed != NULL)
                         action->seq_passed(action);
 
@@ -222,17 +286,101 @@ void vmw_fences_perform_actions(struct vmw_fence_manager *fman,
                  * it will be performed by a worker task.
                  */
 
-                if (action->cleanup != NULL)
-                        list_add_tail(&action->head, &fman->cleanup_list);
+                list_add_tail(&action->head, &fman->cleanup_list);
+        }
+}
+
+/**
+ * vmw_fence_goal_new_locked - Figure out a new device fence goal
+ * seqno if needed.
+ *
+ * @fman: Pointer to a fence manager.
+ * @passed_seqno: The seqno the device currently signals as passed.
+ *
+ * This function should be called with the fence manager lock held.
+ * It is typically called when we have a new passed_seqno, and
+ * we might need to update the fence goal. It checks to see whether
+ * the current fence goal has already passed, and, in that case,
+ * scans through all unsignaled fences to get the next fence object with an
+ * action attached, and sets the seqno of that fence as a new fence goal.
+ *
+ * returns true if the device goal seqno was updated. False otherwise.
+ */
+static bool vmw_fence_goal_new_locked(struct vmw_fence_manager *fman,
+                                      u32 passed_seqno)
+{
+        u32 goal_seqno;
+        __le32 __iomem *fifo_mem;
+        struct vmw_fence_obj *fence;
+
+        if (likely(!fman->seqno_valid))
+                return false;
+
+        fifo_mem = fman->dev_priv->mmio_virt;
+        goal_seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE_GOAL);
+        if (likely(passed_seqno - goal_seqno >= VMW_FENCE_WRAP))
+                return false;
+
+        fman->seqno_valid = false;
+        list_for_each_entry(fence, &fman->fence_list, head) {
+                if (!list_empty(&fence->seq_passed_actions)) {
+                        fman->seqno_valid = true;
+                        iowrite32(fence->seqno,
+                                  fifo_mem + SVGA_FIFO_FENCE_GOAL);
+                        break;
+                }
         }
+
+        return true;
+}
+
+
+/**
+ * vmw_fence_goal_check_locked - Replace the device fence goal seqno if
+ * needed.
+ *
+ * @fence: Pointer to a struct vmw_fence_obj the seqno of which should be
+ * considered as a device fence goal.
+ *
+ * This function should be called with the fence manager lock held.
+ * It is typically called when an action has been attached to a fence to
+ * check whether the seqno of that fence should be used for a fence
+ * goal interrupt. This is typically needed if the current fence goal is
+ * invalid, or has a higher seqno than that of the current fence object.
+ *
+ * returns true if the device goal seqno was updated. False otherwise.
+ */
+static bool vmw_fence_goal_check_locked(struct vmw_fence_obj *fence)
+{
+        u32 goal_seqno;
+        __le32 __iomem *fifo_mem;
+
+        if (fence->signaled & DRM_VMW_FENCE_FLAG_EXEC)
+                return false;
+
+        fifo_mem = fence->fman->dev_priv->mmio_virt;
+        goal_seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE_GOAL);
+        if (likely(fence->fman->seqno_valid &&
+                   goal_seqno - fence->seqno < VMW_FENCE_WRAP))
+                return false;
+
+        iowrite32(fence->seqno, fifo_mem + SVGA_FIFO_FENCE_GOAL);
+        fence->fman->seqno_valid = true;
+
+        return true;
 }
 
-void vmw_fences_update(struct vmw_fence_manager *fman, u32 seqno)
+void vmw_fences_update(struct vmw_fence_manager *fman)
 {
         unsigned long flags;
         struct vmw_fence_obj *fence, *next_fence;
         struct list_head action_list;
+        bool needs_rerun;
+        uint32_t seqno, new_seqno;
+        __le32 __iomem *fifo_mem = fman->dev_priv->mmio_virt;
 
+        seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE);
+rerun:
         spin_lock_irqsave(&fman->lock, flags);
         list_for_each_entry_safe(fence, next_fence, &fman->fence_list, head) {
                 if (seqno - fence->seqno < VMW_FENCE_WRAP) {
@@ -243,14 +391,30 @@ void vmw_fences_update(struct vmw_fence_manager *fman, u32 seqno)
                                          &action_list);
                         vmw_fences_perform_actions(fman, &action_list);
                         wake_up_all(&fence->queue);
-                }
-
+                } else
+                        break;
         }
+
+        needs_rerun = vmw_fence_goal_new_locked(fman, seqno);
+
         if (!list_empty(&fman->cleanup_list))
                 (void) schedule_work(&fman->work);
         spin_unlock_irqrestore(&fman->lock, flags);
-}
 
+        /*
+         * Rerun if the fence goal seqno was updated, and the
+         * hardware might have raced with that update, so that
+         * we missed a fence_goal irq.
+         */
+
+        if (unlikely(needs_rerun)) {
+                new_seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE);
+                if (new_seqno != seqno) {
+                        seqno = new_seqno;
+                        goto rerun;
+                }
+        }
+}
 
 bool vmw_fence_obj_signaled(struct vmw_fence_obj *fence,
                             uint32_t flags)
@@ -267,14 +431,8 @@ bool vmw_fence_obj_signaled(struct vmw_fence_obj *fence,
         if ((signaled & flags) == flags)
                 return 1;
 
-        if ((signaled & DRM_VMW_FENCE_FLAG_EXEC) == 0) {
-                struct vmw_private *dev_priv = fman->dev_priv;
-                __le32 __iomem *fifo_mem = dev_priv->mmio_virt;
-                u32 seqno;
-
-                seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE);
-                vmw_fences_update(fman, seqno);
-        }
+        if ((signaled & DRM_VMW_FENCE_FLAG_EXEC) == 0)
+                vmw_fences_update(fman);
 
         spin_lock_irqsave(&fman->lock, irq_flags);
         signaled = fence->signaled;
@@ -624,3 +782,344 @@ int vmw_fence_obj_unref_ioctl(struct drm_device *dev, void *data,
                                          arg->handle,
                                          TTM_REF_USAGE);
 }
+
+/**
+ * vmw_event_fence_action_destroy
+ *
+ * @kref: The struct kref embedded in a struct vmw_event_fence_action.
+ *
+ * The vmw_event_fence_action destructor that may be called either after
+ * the fence action cleanup, or when the event is delivered.
+ * It frees both the vmw_event_fence_action struct and the actual
+ * event structure copied to user-space.
+ */
+static void vmw_event_fence_action_destroy(struct kref *kref)
+{
+        struct vmw_event_fence_action *eaction =
+                container_of(kref, struct vmw_event_fence_action, kref);
+        struct ttm_mem_global *mem_glob =
+                vmw_mem_glob(vmw_priv(eaction->dev));
+        uint32_t size = eaction->size;
+
+        kfree(eaction->e.event);
+        kfree(eaction);
+        ttm_mem_global_free(mem_glob, size);
+}
+
+
+/**
+ * vmw_event_fence_action_delivered
+ *
+ * @e: The struct drm_pending_event embedded in a struct
+ * vmw_event_fence_action.
+ *
+ * The struct drm_pending_event destructor that is called by drm
+ * once the event is delivered. Since we don't know whether this function
+ * will be called before or after the fence action destructor, we
+ * free a refcount and destroy if it becomes zero.
+ */
+static void vmw_event_fence_action_delivered(struct drm_pending_event *e)
+{
+        struct vmw_event_fence_action *eaction =
+                container_of(e, struct vmw_event_fence_action, e);
+
+        kref_put(&eaction->kref, vmw_event_fence_action_destroy);
+}
+
+
+/**
+ * vmw_event_fence_action_seq_passed
+ *
+ * @action: The struct vmw_fence_action embedded in a struct
+ * vmw_event_fence_action.
+ *
+ * This function is called when the seqno of the fence where @action is
+ * attached has passed. It queues the event on the submitter's event list.
+ * This function is always called from atomic context, and may be called
+ * from irq context. It ups a refcount reflecting that we now have two
+ * destructors.
+ */
+static void vmw_event_fence_action_seq_passed(struct vmw_fence_action *action)
+{
+        struct vmw_event_fence_action *eaction =
+                container_of(action, struct vmw_event_fence_action, action);
+        struct drm_device *dev = eaction->dev;
+        struct drm_file *file_priv = eaction->e.file_priv;
+        unsigned long irq_flags;
+
+        kref_get(&eaction->kref);
+        spin_lock_irqsave(&dev->event_lock, irq_flags);
+
+        if (likely(eaction->tv_sec != NULL)) {
+                struct timeval tv;
+
+                do_gettimeofday(&tv);
+                *eaction->tv_sec = tv.tv_sec;
+                *eaction->tv_usec = tv.tv_usec;
+        }
+
+        list_add_tail(&eaction->e.link, &file_priv->event_list);
+        wake_up_all(&file_priv->event_wait);
+        spin_unlock_irqrestore(&dev->event_lock, irq_flags);
+}
+
+/**
+ * vmw_event_fence_action_cleanup
+ *
+ * @action: The struct vmw_fence_action embedded in a struct
+ * vmw_event_fence_action.
+ *
+ * This function is the struct vmw_fence_action destructor. It's typically
+ * called from a workqueue.
+ */
+static void vmw_event_fence_action_cleanup(struct vmw_fence_action *action)
+{
+        struct vmw_event_fence_action *eaction =
+                container_of(action, struct vmw_event_fence_action, action);
+
+        vmw_fence_obj_unreference(&eaction->fence);
+        kref_put(&eaction->kref, vmw_event_fence_action_destroy);
+}
+
+
+/**
+ * vmw_fence_obj_add_action - Add an action to a fence object.
+ *
+ * @fence - The fence object.
+ * @action - The action to add.
+ *
+ * Note that the action callbacks may be executed before this function
+ * returns.
+ */
+void vmw_fence_obj_add_action(struct vmw_fence_obj *fence,
+                              struct vmw_fence_action *action)
+{
+        struct vmw_fence_manager *fman = fence->fman;
+        unsigned long irq_flags;
+        bool run_update = false;
+
+        mutex_lock(&fman->goal_irq_mutex);
+        spin_lock_irqsave(&fman->lock, irq_flags);
+
+        fman->pending_actions[action->type]++;
+        if (fence->signaled & DRM_VMW_FENCE_FLAG_EXEC) {
+                struct list_head action_list;
+
+                INIT_LIST_HEAD(&action_list);
+                list_add_tail(&action->head, &action_list);
+                vmw_fences_perform_actions(fman, &action_list);
+        } else {
+                list_add_tail(&action->head, &fence->seq_passed_actions);
+
+                /*
+                 * This function may set fman::seqno_valid, so it must
+                 * be run with the goal_irq_mutex held.
+                 */
+                run_update = vmw_fence_goal_check_locked(fence);
+        }
+
+        spin_unlock_irqrestore(&fman->lock, irq_flags);
+
+        if (run_update) {
+                if (!fman->goal_irq_on) {
+                        fman->goal_irq_on = true;
+                        vmw_goal_waiter_add(fman->dev_priv);
+                }
+                vmw_fences_update(fman);
+        }
+        mutex_unlock(&fman->goal_irq_mutex);
+
+}
+
+/**
+ * vmw_event_fence_action_create - Post an event for sending when a fence
+ * object seqno has passed.
+ *
+ * @file_priv: The file connection on which the event should be posted.
+ * @fence: The fence object on which to post the event.
+ * @event: Event to be posted. This event should've been alloced
+ * using k[mz]alloc, and should've been completely initialized.
+ * @interruptible: Interruptible waits if possible.
+ *
+ * As a side effect, the object pointed to by @event may have been
+ * freed when this function returns. If this function returns with
+ * an error code, the caller needs to free that object.
+ */
+
+int vmw_event_fence_action_create(struct drm_file *file_priv,
+                                  struct vmw_fence_obj *fence,
+                                  struct drm_event *event,
+                                  uint32_t *tv_sec,
+                                  uint32_t *tv_usec,
+                                  bool interruptible)
+{
+        struct vmw_event_fence_action *eaction =
+                kzalloc(sizeof(*eaction), GFP_KERNEL);
+        struct ttm_mem_global *mem_glob =
+                vmw_mem_glob(fence->fman->dev_priv);
+        struct vmw_fence_manager *fman = fence->fman;
+        uint32_t size = fman->event_fence_action_size +
+                ttm_round_pot(event->length);
+        int ret;
+
+        /*
+         * Account for internal structure size as well as the
+         * event size itself.
+         */
+
+        ret = ttm_mem_global_alloc(mem_glob, size, false, interruptible);
+        if (unlikely(ret != 0))
+                return ret;
+
+        eaction = kzalloc(sizeof(*eaction), GFP_KERNEL);
+        if (unlikely(eaction == NULL)) {
+                ttm_mem_global_free(mem_glob, size);
+                return -ENOMEM;
+        }
+
+        eaction->e.event = event;
+        eaction->e.file_priv = file_priv;
+        eaction->e.destroy = vmw_event_fence_action_delivered;
+
+        eaction->action.seq_passed = vmw_event_fence_action_seq_passed;
+        eaction->action.cleanup = vmw_event_fence_action_cleanup;
+        eaction->action.type = VMW_ACTION_EVENT;
+
+        eaction->fence = vmw_fence_obj_reference(fence);
+        eaction->dev = fman->dev_priv->dev;
+        eaction->size = size;
+        eaction->tv_sec = tv_sec;
+        eaction->tv_usec = tv_usec;
+
+        kref_init(&eaction->kref);
+        vmw_fence_obj_add_action(fence, &eaction->action);
+
+        return 0;
+}
+
+int vmw_fence_event_ioctl(struct drm_device *dev, void *data,
+                          struct drm_file *file_priv)
+{
+        struct vmw_private *dev_priv = vmw_priv(dev);
+        struct drm_vmw_fence_event_arg *arg =
+                (struct drm_vmw_fence_event_arg *) data;
+        struct vmw_fence_obj *fence = NULL;
+        struct vmw_fpriv *vmw_fp = vmw_fpriv(file_priv);
+        struct drm_vmw_fence_rep __user *user_fence_rep =
+                (struct drm_vmw_fence_rep __user *)(unsigned long)
+                arg->fence_rep;
+        uint32_t handle;
+        unsigned long irq_flags;
+        struct drm_vmw_event_fence *event;
+        int ret;
+
+        /*
+         * Look up an existing fence object,
+         * and if user-space wants a new reference,
+         * add one.
+         */
+        if (arg->handle) {
+                struct ttm_base_object *base =
+                        ttm_base_object_lookup(vmw_fp->tfile, arg->handle);
+
+                if (unlikely(base == NULL)) {
+                        DRM_ERROR("Fence event invalid fence object handle "
+                                  "0x%08lx.\n",
+                                  (unsigned long)arg->handle);
+                        return -EINVAL;
+                }
+                fence = &(container_of(base, struct vmw_user_fence,
+                                       base)->fence);
+                (void) vmw_fence_obj_reference(fence);
+
+                if (user_fence_rep != NULL) {
+                        bool existed;
+
+                        ret = ttm_ref_object_add(vmw_fp->tfile, base,
+                                                 TTM_REF_USAGE, &existed);
+                        if (unlikely(ret != 0)) {
+                                DRM_ERROR("Failed to reference a fence "
+                                          "object.\n");
+                                goto out_no_ref_obj;
+                        }
+                        handle = base->hash.key;
+                }
+                ttm_base_object_unref(&base);
+        }
+
+        /*
+         * Create a new fence object.
+         */
+        if (!fence) {
+                ret = vmw_execbuf_fence_commands(file_priv, dev_priv,
+                                                 &fence,
+                                                 (user_fence_rep) ?
+                                                 &handle : NULL);
+                if (unlikely(ret != 0)) {
+                        DRM_ERROR("Fence event failed to create fence.\n");
+                        return ret;
+                }
+        }
+
+        BUG_ON(fence == NULL);
+
+        spin_lock_irqsave(&dev->event_lock, irq_flags);
+
+        ret = (file_priv->event_space < sizeof(*event)) ? -EBUSY : 0;
+        if (likely(ret == 0))
+                file_priv->event_space -= sizeof(*event);
+
+        spin_unlock_irqrestore(&dev->event_lock, irq_flags);
+
+        if (unlikely(ret != 0)) {
+                DRM_ERROR("Failed to allocate event space for this file.\n");
+                goto out_no_event_space;
+        }
+
+        event = kzalloc(sizeof(*event), GFP_KERNEL);
+        if (unlikely(event == NULL)) {
+                DRM_ERROR("Failed to allocate an event.\n");
+                goto out_no_event;
+        }
+
+        event->base.type = DRM_VMW_EVENT_FENCE_SIGNALED;
+        event->base.length = sizeof(*event);
+        event->user_data = arg->user_data;
+
+        if (arg->flags & DRM_VMW_FE_FLAG_REQ_TIME)
+                ret = vmw_event_fence_action_create(file_priv, fence,
+                                                    &event->base,
+                                                    &event->tv_sec,
+                                                    &event->tv_usec,
+                                                    true);
+        else
+                ret = vmw_event_fence_action_create(file_priv, fence,
+                                                    &event->base,
+                                                    NULL,
+                                                    NULL,
+                                                    true);
+
+        if (unlikely(ret != 0)) {
+                if (ret != -ERESTARTSYS)
+                        DRM_ERROR("Failed to attach event to fence.\n");
+                goto out_no_attach;
+        }
+
+        vmw_execbuf_copy_fence_user(dev_priv, vmw_fp, 0, user_fence_rep, fence,
+                                    handle);
+        vmw_fence_obj_unreference(&fence);
+        return 0;
+out_no_attach:
+        kfree(event);
+out_no_event:
+        spin_lock_irqsave(&dev->event_lock, irq_flags);
+        file_priv->event_space += sizeof(*event);
+        spin_unlock_irqrestore(&dev->event_lock, irq_flags);
+out_no_event_space:
+        if (user_fence_rep != NULL)
+                ttm_ref_object_base_unref(vmw_fpriv(file_priv)->tfile,
+                                          handle, TTM_REF_USAGE);
+out_no_ref_obj:
+        vmw_fence_obj_unreference(&fence);
+        return ret;
+}