drm/vmwgfx: Move irq bottom half processing to threads
authorThomas Hellstrom <thellstrom@vmware.com>
Thu, 24 Aug 2017 06:06:28 +0000 (08:06 +0200)
committerSinclair Yeh <syeh@vmware.com>
Mon, 28 Aug 2017 15:40:43 +0000 (17:40 +0200)
This gets rid of the irq bottom half tasklets and instead performs the
work needed in process context. We also convert irq-disabling spinlocks to
ordinary spinlocks.

This should decrease system latency for other system components, like
sound for example but has the potential to increase latency for processes
that wait on the GPU.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>
drivers/gpu/drm/vmwgfx/vmwgfx_cmdbuf.c
drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
drivers/gpu/drm/vmwgfx/vmwgfx_irq.c

index 86178796de6c354a32cf1ef35637aa459d8cb7b5..2e5718509c2308addc27614e9d9fa283b9bdaeb0 100644 (file)
@@ -85,7 +85,6 @@ struct vmw_cmdbuf_context {
  * Internal protection.
  * @dheaders: Pool of DMA memory for device command buffer headers with trailing
  * space for inline data. Internal protection.
- * @tasklet: Tasklet struct for irq processing. Immutable.
  * @alloc_queue: Wait queue for processes waiting to allocate command buffer
  * space.
  * @idle_queue: Wait queue for processes waiting for command buffer idle.
@@ -117,7 +116,6 @@ struct vmw_cmdbuf_man {
        spinlock_t lock;
        struct dma_pool *headers;
        struct dma_pool *dheaders;
-       struct tasklet_struct tasklet;
        wait_queue_head_t alloc_queue;
        wait_queue_head_t idle_queue;
        bool irq_on;
@@ -278,9 +276,9 @@ void vmw_cmdbuf_header_free(struct vmw_cmdbuf_header *header)
                vmw_cmdbuf_header_inline_free(header);
                return;
        }
-       spin_lock_bh(&man->lock);
+       spin_lock(&man->lock);
        __vmw_cmdbuf_header_free(header);
-       spin_unlock_bh(&man->lock);
+       spin_unlock(&man->lock);
 }
 
 
@@ -468,20 +466,17 @@ static void vmw_cmdbuf_ctx_add(struct vmw_cmdbuf_man *man,
 }
 
 /**
- * vmw_cmdbuf_man_tasklet - The main part of the command buffer interrupt
- * handler implemented as a tasklet.
+ * vmw_cmdbuf_irqthread - The main part of the command buffer interrupt
+ * handler implemented as a threaded irq task.
  *
- * @data: Tasklet closure. A pointer to the command buffer manager cast to
- * an unsigned long.
+ * @man: Pointer to the command buffer manager.
  *
- * The bottom half (tasklet) of the interrupt handler simply calls into the
+ * The bottom half of the interrupt handler simply calls into the
  * command buffer processor to free finished buffers and submit any
  * queued buffers to hardware.
  */
-static void vmw_cmdbuf_man_tasklet(unsigned long data)
+void vmw_cmdbuf_irqthread(struct vmw_cmdbuf_man *man)
 {
-       struct vmw_cmdbuf_man *man = (struct vmw_cmdbuf_man *) data;
-
        spin_lock(&man->lock);
        vmw_cmdbuf_man_process(man);
        spin_unlock(&man->lock);
@@ -504,7 +499,7 @@ static void vmw_cmdbuf_work_func(struct work_struct *work)
        uint32_t dummy;
        bool restart = false;
 
-       spin_lock_bh(&man->lock);
+       spin_lock(&man->lock);
        list_for_each_entry_safe(entry, next, &man->error, list) {
                restart = true;
                DRM_ERROR("Command buffer error.\n");
@@ -513,7 +508,7 @@ static void vmw_cmdbuf_work_func(struct work_struct *work)
                __vmw_cmdbuf_header_free(entry);
                wake_up_all(&man->idle_queue);
        }
-       spin_unlock_bh(&man->lock);
+       spin_unlock(&man->lock);
 
        if (restart && vmw_cmdbuf_startstop(man, true))
                DRM_ERROR("Failed restarting command buffer context 0.\n");
@@ -536,7 +531,7 @@ static bool vmw_cmdbuf_man_idle(struct vmw_cmdbuf_man *man,
        bool idle = false;
        int i;
 
-       spin_lock_bh(&man->lock);
+       spin_lock(&man->lock);
        vmw_cmdbuf_man_process(man);
        for_each_cmdbuf_ctx(man, i, ctx) {
                if (!list_empty(&ctx->submitted) ||
@@ -548,7 +543,7 @@ static bool vmw_cmdbuf_man_idle(struct vmw_cmdbuf_man *man,
        idle = list_empty(&man->error);
 
 out_unlock:
-       spin_unlock_bh(&man->lock);
+       spin_unlock(&man->lock);
 
        return idle;
 }
@@ -571,7 +566,7 @@ static void __vmw_cmdbuf_cur_flush(struct vmw_cmdbuf_man *man)
        if (!cur)
                return;
 
-       spin_lock_bh(&man->lock);
+       spin_lock(&man->lock);
        if (man->cur_pos == 0) {
                __vmw_cmdbuf_header_free(cur);
                goto out_unlock;
@@ -580,7 +575,7 @@ static void __vmw_cmdbuf_cur_flush(struct vmw_cmdbuf_man *man)
        man->cur->cb_header->length = man->cur_pos;
        vmw_cmdbuf_ctx_add(man, man->cur, SVGA_CB_CONTEXT_0);
 out_unlock:
-       spin_unlock_bh(&man->lock);
+       spin_unlock(&man->lock);
        man->cur = NULL;
        man->cur_pos = 0;
 }
@@ -673,14 +668,14 @@ static bool vmw_cmdbuf_try_alloc(struct vmw_cmdbuf_man *man,
                return true;
  
        memset(info->node, 0, sizeof(*info->node));
-       spin_lock_bh(&man->lock);
+       spin_lock(&man->lock);
        ret = drm_mm_insert_node(&man->mm, info->node, info->page_size);
        if (ret) {
                vmw_cmdbuf_man_process(man);
                ret = drm_mm_insert_node(&man->mm, info->node, info->page_size);
        }
 
-       spin_unlock_bh(&man->lock);
+       spin_unlock(&man->lock);
        info->done = !ret;
 
        return info->done;
@@ -801,9 +796,9 @@ static int vmw_cmdbuf_space_pool(struct vmw_cmdbuf_man *man,
        return 0;
 
 out_no_cb_header:
-       spin_lock_bh(&man->lock);
+       spin_lock(&man->lock);
        drm_mm_remove_node(&header->node);
-       spin_unlock_bh(&man->lock);
+       spin_unlock(&man->lock);
 
        return ret;
 }
@@ -1023,18 +1018,6 @@ void vmw_cmdbuf_commit(struct vmw_cmdbuf_man *man, size_t size,
        vmw_cmdbuf_cur_unlock(man);
 }
 
-/**
- * vmw_cmdbuf_tasklet_schedule - Schedule the interrupt handler bottom half.
- *
- * @man: The command buffer manager.
- */
-void vmw_cmdbuf_tasklet_schedule(struct vmw_cmdbuf_man *man)
-{
-       if (!man)
-               return;
-
-       tasklet_schedule(&man->tasklet);
-}
 
 /**
  * vmw_cmdbuf_send_device_command - Send a command through the device context.
@@ -1059,9 +1042,9 @@ static int vmw_cmdbuf_send_device_command(struct vmw_cmdbuf_man *man,
        memcpy(cmd, command, size);
        header->cb_header->length = size;
        header->cb_context = SVGA_CB_CONTEXT_DEVICE;
-       spin_lock_bh(&man->lock);
+       spin_lock(&man->lock);
        status = vmw_cmdbuf_header_submit(header);
-       spin_unlock_bh(&man->lock);
+       spin_unlock(&man->lock);
        vmw_cmdbuf_header_free(header);
 
        if (status != SVGA_CB_STATUS_COMPLETED) {
@@ -1226,8 +1209,6 @@ struct vmw_cmdbuf_man *vmw_cmdbuf_man_create(struct vmw_private *dev_priv)
        spin_lock_init(&man->lock);
        mutex_init(&man->cur_mutex);
        mutex_init(&man->space_mutex);
-       tasklet_init(&man->tasklet, vmw_cmdbuf_man_tasklet,
-                    (unsigned long) man);
        man->default_size = VMW_CMDBUF_INLINE_SIZE;
        init_waitqueue_head(&man->alloc_queue);
        init_waitqueue_head(&man->idle_queue);
@@ -1297,7 +1278,6 @@ void vmw_cmdbuf_man_destroy(struct vmw_cmdbuf_man *man)
 
        vmw_generic_waiter_remove(man->dev_priv, SVGA_IRQFLAG_ERROR,
                                  &man->dev_priv->error_waiters);
-       tasklet_kill(&man->tasklet);
        (void) cancel_work_sync(&man->work);
        dma_pool_destroy(man->dheaders);
        dma_pool_destroy(man->headers);
index 5e7e8a3df7c38afc0c7af94ef3e368abed2a6e7a..b74cee295ede03efc4a06db33b65031804fdae8e 100644 (file)
@@ -352,6 +352,12 @@ struct vmw_otable_batch {
        struct ttm_buffer_object *otable_bo;
 };
 
+enum {
+       VMW_IRQTHREAD_FENCE,
+       VMW_IRQTHREAD_CMDBUF,
+       VMW_IRQTHREAD_MAX
+};
+
 struct vmw_private {
        struct ttm_bo_device bdev;
        struct ttm_bo_global_ref bo_global_ref;
@@ -530,6 +536,7 @@ struct vmw_private {
        struct vmw_otable_batch otable_batch;
 
        struct vmw_cmdbuf_man *cman;
+       DECLARE_BITMAP(irqthread_pending, VMW_IRQTHREAD_MAX);
 };
 
 static inline struct vmw_surface *vmw_res_to_srf(struct vmw_resource *res)
@@ -562,24 +569,21 @@ static inline struct vmw_master *vmw_master(struct drm_master *master)
 static inline void vmw_write(struct vmw_private *dev_priv,
                             unsigned int offset, uint32_t value)
 {
-       unsigned long irq_flags;
-
-       spin_lock_irqsave(&dev_priv->hw_lock, irq_flags);
+       spin_lock(&dev_priv->hw_lock);
        outl(offset, dev_priv->io_start + VMWGFX_INDEX_PORT);
        outl(value, dev_priv->io_start + VMWGFX_VALUE_PORT);
-       spin_unlock_irqrestore(&dev_priv->hw_lock, irq_flags);
+       spin_unlock(&dev_priv->hw_lock);
 }
 
 static inline uint32_t vmw_read(struct vmw_private *dev_priv,
                                unsigned int offset)
 {
-       unsigned long irq_flags;
        u32 val;
 
-       spin_lock_irqsave(&dev_priv->hw_lock, irq_flags);
+       spin_lock(&dev_priv->hw_lock);
        outl(offset, dev_priv->io_start + VMWGFX_INDEX_PORT);
        val = inl(dev_priv->io_start + VMWGFX_VALUE_PORT);
-       spin_unlock_irqrestore(&dev_priv->hw_lock, irq_flags);
+       spin_unlock(&dev_priv->hw_lock);
 
        return val;
 }
@@ -1149,13 +1153,13 @@ extern void *vmw_cmdbuf_reserve(struct vmw_cmdbuf_man *man, size_t size,
 extern void vmw_cmdbuf_commit(struct vmw_cmdbuf_man *man, size_t size,
                              struct vmw_cmdbuf_header *header,
                              bool flush);
-extern void vmw_cmdbuf_tasklet_schedule(struct vmw_cmdbuf_man *man);
 extern void *vmw_cmdbuf_alloc(struct vmw_cmdbuf_man *man,
                              size_t size, bool interruptible,
                              struct vmw_cmdbuf_header **p_header);
 extern void vmw_cmdbuf_header_free(struct vmw_cmdbuf_header *header);
 extern int vmw_cmdbuf_cur_flush(struct vmw_cmdbuf_man *man,
                                bool interruptible);
+extern void vmw_cmdbuf_irqthread(struct vmw_cmdbuf_man *man);
 
 
 /**
index b8bc5bc7de7e0d4616cbbfe6fbee1d309aa55360..c812570ff159fb5dd75f30bce4c1f573386a80f1 100644 (file)
@@ -114,12 +114,11 @@ static void vmw_fence_obj_destroy(struct dma_fence *f)
                container_of(f, struct vmw_fence_obj, base);
 
        struct vmw_fence_manager *fman = fman_from_fence(fence);
-       unsigned long irq_flags;
 
-       spin_lock_irqsave(&fman->lock, irq_flags);
+       spin_lock(&fman->lock);
        list_del_init(&fence->head);
        --fman->num_fence_objects;
-       spin_unlock_irqrestore(&fman->lock, irq_flags);
+       spin_unlock(&fman->lock);
        fence->destroy(fence);
 }
 
@@ -252,10 +251,10 @@ static void vmw_fence_work_func(struct work_struct *work)
                INIT_LIST_HEAD(&list);
                mutex_lock(&fman->goal_irq_mutex);
 
-               spin_lock_irq(&fman->lock);
+               spin_lock(&fman->lock);
                list_splice_init(&fman->cleanup_list, &list);
                seqno_valid = fman->seqno_valid;
-               spin_unlock_irq(&fman->lock);
+               spin_unlock(&fman->lock);
 
                if (!seqno_valid && fman->goal_irq_on) {
                        fman->goal_irq_on = false;
@@ -305,15 +304,14 @@ struct vmw_fence_manager *vmw_fence_manager_init(struct vmw_private *dev_priv)
 
 void vmw_fence_manager_takedown(struct vmw_fence_manager *fman)
 {
-       unsigned long irq_flags;
        bool lists_empty;
 
        (void) cancel_work_sync(&fman->work);
 
-       spin_lock_irqsave(&fman->lock, irq_flags);
+       spin_lock(&fman->lock);
        lists_empty = list_empty(&fman->fence_list) &&
                list_empty(&fman->cleanup_list);
-       spin_unlock_irqrestore(&fman->lock, irq_flags);
+       spin_unlock(&fman->lock);
 
        BUG_ON(!lists_empty);
        kfree(fman);
@@ -323,7 +321,6 @@ static int vmw_fence_obj_init(struct vmw_fence_manager *fman,
                              struct vmw_fence_obj *fence, u32 seqno,
                              void (*destroy) (struct vmw_fence_obj *fence))
 {
-       unsigned long irq_flags;
        int ret = 0;
 
        dma_fence_init(&fence->base, &vmw_fence_ops, &fman->lock,
@@ -331,7 +328,7 @@ static int vmw_fence_obj_init(struct vmw_fence_manager *fman,
        INIT_LIST_HEAD(&fence->seq_passed_actions);
        fence->destroy = destroy;
 
-       spin_lock_irqsave(&fman->lock, irq_flags);
+       spin_lock(&fman->lock);
        if (unlikely(fman->fifo_down)) {
                ret = -EBUSY;
                goto out_unlock;
@@ -340,7 +337,7 @@ static int vmw_fence_obj_init(struct vmw_fence_manager *fman,
        ++fman->num_fence_objects;
 
 out_unlock:
-       spin_unlock_irqrestore(&fman->lock, irq_flags);
+       spin_unlock(&fman->lock);
        return ret;
 
 }
@@ -489,11 +486,9 @@ rerun:
 
 void vmw_fences_update(struct vmw_fence_manager *fman)
 {
-       unsigned long irq_flags;
-
-       spin_lock_irqsave(&fman->lock, irq_flags);
+       spin_lock(&fman->lock);
        __vmw_fences_update(fman);
-       spin_unlock_irqrestore(&fman->lock, irq_flags);
+       spin_unlock(&fman->lock);
 }
 
 bool vmw_fence_obj_signaled(struct vmw_fence_obj *fence)
@@ -663,14 +658,14 @@ void vmw_fence_fifo_down(struct vmw_fence_manager *fman)
         * restart when we've released the fman->lock.
         */
 
-       spin_lock_irq(&fman->lock);
+       spin_lock(&fman->lock);
        fman->fifo_down = true;
        while (!list_empty(&fman->fence_list)) {
                struct vmw_fence_obj *fence =
                        list_entry(fman->fence_list.prev, struct vmw_fence_obj,
                                   head);
                dma_fence_get(&fence->base);
-               spin_unlock_irq(&fman->lock);
+               spin_unlock(&fman->lock);
 
                ret = vmw_fence_obj_wait(fence, false, false,
                                         VMW_FENCE_WAIT_TIMEOUT);
@@ -686,18 +681,16 @@ void vmw_fence_fifo_down(struct vmw_fence_manager *fman)
 
                BUG_ON(!list_empty(&fence->head));
                dma_fence_put(&fence->base);
-               spin_lock_irq(&fman->lock);
+               spin_lock(&fman->lock);
        }
-       spin_unlock_irq(&fman->lock);
+       spin_unlock(&fman->lock);
 }
 
 void vmw_fence_fifo_up(struct vmw_fence_manager *fman)
 {
-       unsigned long irq_flags;
-
-       spin_lock_irqsave(&fman->lock, irq_flags);
+       spin_lock(&fman->lock);
        fman->fifo_down = false;
-       spin_unlock_irqrestore(&fman->lock, irq_flags);
+       spin_unlock(&fman->lock);
 }
 
 
@@ -812,9 +805,9 @@ int vmw_fence_obj_signaled_ioctl(struct drm_device *dev, void *data,
        arg->signaled = vmw_fence_obj_signaled(fence);
 
        arg->signaled_flags = arg->flags;
-       spin_lock_irq(&fman->lock);
+       spin_lock(&fman->lock);
        arg->passed_seqno = dev_priv->last_read_seqno;
-       spin_unlock_irq(&fman->lock);
+       spin_unlock(&fman->lock);
 
        ttm_base_object_unref(&base);
 
@@ -841,8 +834,7 @@ int vmw_fence_obj_unref_ioctl(struct drm_device *dev, void *data,
  *
  * 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.
+ * This function is always called from atomic context.
  */
 static void vmw_event_fence_action_seq_passed(struct vmw_fence_action *action)
 {
@@ -851,13 +843,13 @@ static void vmw_event_fence_action_seq_passed(struct vmw_fence_action *action)
        struct drm_device *dev = eaction->dev;
        struct drm_pending_event *event = eaction->event;
        struct drm_file *file_priv;
-       unsigned long irq_flags;
+
 
        if (unlikely(event == NULL))
                return;
 
        file_priv = event->file_priv;
-       spin_lock_irqsave(&dev->event_lock, irq_flags);
+       spin_lock_irq(&dev->event_lock);
 
        if (likely(eaction->tv_sec != NULL)) {
                struct timeval tv;
@@ -869,7 +861,7 @@ static void vmw_event_fence_action_seq_passed(struct vmw_fence_action *action)
 
        drm_send_event_locked(dev, eaction->event);
        eaction->event = NULL;
-       spin_unlock_irqrestore(&dev->event_lock, irq_flags);
+       spin_unlock_irq(&dev->event_lock);
 }
 
 /**
@@ -904,11 +896,10 @@ static void vmw_fence_obj_add_action(struct vmw_fence_obj *fence,
                              struct vmw_fence_action *action)
 {
        struct vmw_fence_manager *fman = fman_from_fence(fence);
-       unsigned long irq_flags;
        bool run_update = false;
 
        mutex_lock(&fman->goal_irq_mutex);
-       spin_lock_irqsave(&fman->lock, irq_flags);
+       spin_lock(&fman->lock);
 
        fman->pending_actions[action->type]++;
        if (dma_fence_is_signaled_locked(&fence->base)) {
@@ -927,7 +918,7 @@ static void vmw_fence_obj_add_action(struct vmw_fence_obj *fence,
                run_update = vmw_fence_goal_check_locked(fence);
        }
 
-       spin_unlock_irqrestore(&fman->lock, irq_flags);
+       spin_unlock(&fman->lock);
 
        if (run_update) {
                if (!fman->goal_irq_on) {
index 9b519c4b4ec28e699140c6f05295e277cdae89a3..b9239ba067c484da3f6d453ae02d088e376fd685 100644 (file)
 
 #define VMW_FENCE_WRAP (1 << 24)
 
+/**
+ * vmw_thread_fn - Deferred (process context) irq handler
+ *
+ * @irq: irq number
+ * @arg: Closure argument. Pointer to a struct drm_device cast to void *
+ *
+ * This function implements the deferred part of irq processing.
+ * The function is guaranteed to run at least once after the
+ * vmw_irq_handler has returned with IRQ_WAKE_THREAD.
+ *
+ */
+static irqreturn_t vmw_thread_fn(int irq, void *arg)
+{
+       struct drm_device *dev = (struct drm_device *)arg;
+       struct vmw_private *dev_priv = vmw_priv(dev);
+       irqreturn_t ret = IRQ_NONE;
+
+       if (test_and_clear_bit(VMW_IRQTHREAD_FENCE,
+                              dev_priv->irqthread_pending)) {
+               vmw_fences_update(dev_priv->fman);
+               wake_up_all(&dev_priv->fence_queue);
+               ret = IRQ_HANDLED;
+       }
+
+       if (test_and_clear_bit(VMW_IRQTHREAD_CMDBUF,
+                              dev_priv->irqthread_pending)) {
+               vmw_cmdbuf_irqthread(dev_priv->cman);
+               ret = IRQ_HANDLED;
+       }
+
+       return ret;
+}
+
+/**
+ * vmw_irq_handler irq handler
+ *
+ * @irq: irq number
+ * @arg: Closure argument. Pointer to a struct drm_device cast to void *
+ *
+ * This function implements the quick part of irq processing.
+ * The function performs fast actions like clearing the device interrupt
+ * flags and also reasonably quick actions like waking processes waiting for
+ * FIFO space. Other IRQ actions are deferred to the IRQ thread.
+ */
 static irqreturn_t vmw_irq_handler(int irq, void *arg)
 {
        struct drm_device *dev = (struct drm_device *)arg;
        struct vmw_private *dev_priv = vmw_priv(dev);
        uint32_t status, masked_status;
+       irqreturn_t ret = IRQ_HANDLED;
 
        status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
        masked_status = status & READ_ONCE(dev_priv->irq_mask);
@@ -45,20 +90,21 @@ static irqreturn_t vmw_irq_handler(int irq, void *arg)
        if (!status)
                return IRQ_NONE;
 
-       if (masked_status & (SVGA_IRQFLAG_ANY_FENCE |
-                            SVGA_IRQFLAG_FENCE_GOAL)) {
-               vmw_fences_update(dev_priv->fman);
-               wake_up_all(&dev_priv->fence_queue);
-       }
-
        if (masked_status & SVGA_IRQFLAG_FIFO_PROGRESS)
                wake_up_all(&dev_priv->fifo_queue);
 
-       if (masked_status & (SVGA_IRQFLAG_COMMAND_BUFFER |
-                            SVGA_IRQFLAG_ERROR))
-               vmw_cmdbuf_tasklet_schedule(dev_priv->cman);
+       if ((masked_status & (SVGA_IRQFLAG_ANY_FENCE |
+                             SVGA_IRQFLAG_FENCE_GOAL)) &&
+           !test_and_set_bit(VMW_IRQTHREAD_FENCE, dev_priv->irqthread_pending))
+               ret = IRQ_WAKE_THREAD;
+
+       if ((masked_status & (SVGA_IRQFLAG_COMMAND_BUFFER |
+                             SVGA_IRQFLAG_ERROR)) &&
+           !test_and_set_bit(VMW_IRQTHREAD_CMDBUF,
+                             dev_priv->irqthread_pending))
+               ret = IRQ_WAKE_THREAD;
 
-       return IRQ_HANDLED;
+       return ret;
 }
 
 static bool vmw_fifo_idle(struct vmw_private *dev_priv, uint32_t seqno)
@@ -326,7 +372,7 @@ int vmw_irq_install(struct drm_device *dev, int irq)
 
        vmw_irq_preinstall(dev);
 
-       ret = request_threaded_irq(irq, vmw_irq_handler, NULL,
+       ret = request_threaded_irq(irq, vmw_irq_handler, vmw_thread_fn,
                                   IRQF_SHARED, VMWGFX_DRIVER_NAME, dev);
        if (ret < 0)
                return ret;