diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
index 7b5d221..6c6b655 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
@@ -406,11 +406,9 @@
 		if (unlikely(ret != 0))
 			--dev_priv->num_3d_resources;
 	} else if (unhide_svga) {
-		mutex_lock(&dev_priv->hw_mutex);
 		vmw_write(dev_priv, SVGA_REG_ENABLE,
 			  vmw_read(dev_priv, SVGA_REG_ENABLE) &
 			  ~SVGA_REG_ENABLE_HIDE);
-		mutex_unlock(&dev_priv->hw_mutex);
 	}
 
 	mutex_unlock(&dev_priv->release_mutex);
@@ -433,13 +431,10 @@
 	mutex_lock(&dev_priv->release_mutex);
 	if (unlikely(--dev_priv->num_3d_resources == 0))
 		vmw_release_device(dev_priv);
-	else if (hide_svga) {
-		mutex_lock(&dev_priv->hw_mutex);
+	else if (hide_svga)
 		vmw_write(dev_priv, SVGA_REG_ENABLE,
 			  vmw_read(dev_priv, SVGA_REG_ENABLE) |
 			  SVGA_REG_ENABLE_HIDE);
-		mutex_unlock(&dev_priv->hw_mutex);
-	}
 
 	n3d = (int32_t) dev_priv->num_3d_resources;
 	mutex_unlock(&dev_priv->release_mutex);
@@ -600,12 +595,14 @@
 	dev_priv->dev = dev;
 	dev_priv->vmw_chipset = chipset;
 	dev_priv->last_read_seqno = (uint32_t) -100;
-	mutex_init(&dev_priv->hw_mutex);
 	mutex_init(&dev_priv->cmdbuf_mutex);
 	mutex_init(&dev_priv->release_mutex);
 	mutex_init(&dev_priv->binding_mutex);
 	rwlock_init(&dev_priv->resource_lock);
 	ttm_lock_init(&dev_priv->reservation_sem);
+	spin_lock_init(&dev_priv->hw_lock);
+	spin_lock_init(&dev_priv->waiter_lock);
+	spin_lock_init(&dev_priv->cap_lock);
 
 	for (i = vmw_res_context; i < vmw_res_max; ++i) {
 		idr_init(&dev_priv->res_idr[i]);
@@ -626,14 +623,11 @@
 
 	dev_priv->enable_fb = enable_fbdev;
 
-	mutex_lock(&dev_priv->hw_mutex);
-
 	vmw_write(dev_priv, SVGA_REG_ID, SVGA_ID_2);
 	svga_id = vmw_read(dev_priv, SVGA_REG_ID);
 	if (svga_id != SVGA_ID_2) {
 		ret = -ENOSYS;
 		DRM_ERROR("Unsupported SVGA ID 0x%x\n", svga_id);
-		mutex_unlock(&dev_priv->hw_mutex);
 		goto out_err0;
 	}
 
@@ -683,10 +677,8 @@
 		dev_priv->prim_bb_mem = dev_priv->vram_size;
 
 	ret = vmw_dma_masks(dev_priv);
-	if (unlikely(ret != 0)) {
-		mutex_unlock(&dev_priv->hw_mutex);
+	if (unlikely(ret != 0))
 		goto out_err0;
-	}
 
 	/*
 	 * Limit back buffer size to VRAM size.  Remove this once
@@ -695,8 +687,6 @@
 	if (dev_priv->prim_bb_mem > dev_priv->vram_size)
 		dev_priv->prim_bb_mem = dev_priv->vram_size;
 
-	mutex_unlock(&dev_priv->hw_mutex);
-
 	vmw_print_capabilities(dev_priv->capabilities);
 
 	if (dev_priv->capabilities & SVGA_CAP_GMR2) {
@@ -1160,9 +1150,7 @@
 		if (unlikely(ret != 0))
 			return ret;
 		vmw_kms_save_vga(dev_priv);
-		mutex_lock(&dev_priv->hw_mutex);
 		vmw_write(dev_priv, SVGA_REG_TRACES, 0);
-		mutex_unlock(&dev_priv->hw_mutex);
 	}
 
 	if (active) {
@@ -1196,9 +1184,7 @@
 	if (!dev_priv->enable_fb) {
 		vmw_kms_restore_vga(dev_priv);
 		vmw_3d_resource_dec(dev_priv, true);
-		mutex_lock(&dev_priv->hw_mutex);
 		vmw_write(dev_priv, SVGA_REG_TRACES, 1);
-		mutex_unlock(&dev_priv->hw_mutex);
 	}
 	return ret;
 }
@@ -1233,9 +1219,7 @@
 			DRM_ERROR("Unable to clean VRAM on master drop.\n");
 		vmw_kms_restore_vga(dev_priv);
 		vmw_3d_resource_dec(dev_priv, true);
-		mutex_lock(&dev_priv->hw_mutex);
 		vmw_write(dev_priv, SVGA_REG_TRACES, 1);
-		mutex_unlock(&dev_priv->hw_mutex);
 	}
 
 	dev_priv->active_master = &dev_priv->fbdev_master;
@@ -1367,10 +1351,8 @@
 	struct drm_device *dev = pci_get_drvdata(pdev);
 	struct vmw_private *dev_priv = vmw_priv(dev);
 
-	mutex_lock(&dev_priv->hw_mutex);
 	vmw_write(dev_priv, SVGA_REG_ID, SVGA_ID_2);
 	(void) vmw_read(dev_priv, SVGA_REG_ID);
-	mutex_unlock(&dev_priv->hw_mutex);
 
 	/**
 	 * Reclaim 3d reference held by fbdev and potentially
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
index 4ee799b..d26a6da 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
@@ -399,7 +399,8 @@
 	uint32_t memory_size;
 	bool has_gmr;
 	bool has_mob;
-	struct mutex hw_mutex;
+	spinlock_t hw_lock;
+	spinlock_t cap_lock;
 
 	/*
 	 * VGA registers.
@@ -449,8 +450,9 @@
 	atomic_t marker_seq;
 	wait_queue_head_t fence_queue;
 	wait_queue_head_t fifo_queue;
-	int fence_queue_waiters; /* Protected by hw_mutex */
-	int goal_queue_waiters; /* Protected by hw_mutex */
+	spinlock_t waiter_lock;
+	int fence_queue_waiters; /* Protected by waiter_lock */
+	int goal_queue_waiters; /* Protected by waiter_lock */
 	atomic_t fifo_queue_waiters;
 	uint32_t last_read_seqno;
 	spinlock_t irq_lock;
@@ -553,20 +555,35 @@
 	return (struct vmw_master *) master->driver_priv;
 }
 
+/*
+ * The locking here is fine-grained, so that it is performed once
+ * for every read- and write operation. This is of course costly, but we
+ * don't perform much register access in the timing critical paths anyway.
+ * Instead we have the extra benefit of being sure that we don't forget
+ * the hw lock around register accesses.
+ */
 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);
 	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);
 }
 
 static inline uint32_t vmw_read(struct vmw_private *dev_priv,
 				unsigned int offset)
 {
-	uint32_t val;
+	unsigned long irq_flags;
+	u32 val;
 
+	spin_lock_irqsave(&dev_priv->hw_lock, irq_flags);
 	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);
+
 	return val;
 }
 
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
index b7594cb..945f1e0 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
@@ -35,7 +35,7 @@
 	struct vmw_private *dev_priv;
 	spinlock_t lock;
 	struct list_head fence_list;
-	struct work_struct work, ping_work;
+	struct work_struct work;
 	u32 user_fence_size;
 	u32 fence_size;
 	u32 event_fence_action_size;
@@ -134,14 +134,6 @@
 	return "svga";
 }
 
-static void vmw_fence_ping_func(struct work_struct *work)
-{
-	struct vmw_fence_manager *fman =
-		container_of(work, struct vmw_fence_manager, ping_work);
-
-	vmw_fifo_ping_host(fman->dev_priv, SVGA_SYNC_GENERIC);
-}
-
 static bool vmw_fence_enable_signaling(struct fence *f)
 {
 	struct vmw_fence_obj *fence =
@@ -155,11 +147,7 @@
 	if (seqno - fence->base.seqno < VMW_FENCE_WRAP)
 		return false;
 
-	if (mutex_trylock(&dev_priv->hw_mutex)) {
-		vmw_fifo_ping_host_locked(dev_priv, SVGA_SYNC_GENERIC);
-		mutex_unlock(&dev_priv->hw_mutex);
-	} else
-		schedule_work(&fman->ping_work);
+	vmw_fifo_ping_host(dev_priv, SVGA_SYNC_GENERIC);
 
 	return true;
 }
@@ -305,7 +293,6 @@
 	INIT_LIST_HEAD(&fman->fence_list);
 	INIT_LIST_HEAD(&fman->cleanup_list);
 	INIT_WORK(&fman->work, &vmw_fence_work_func);
-	INIT_WORK(&fman->ping_work, &vmw_fence_ping_func);
 	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));
@@ -323,7 +310,6 @@
 	bool lists_empty;
 
 	(void) cancel_work_sync(&fman->work);
-	(void) cancel_work_sync(&fman->ping_work);
 
 	spin_lock_irqsave(&fman->lock, irq_flags);
 	lists_empty = list_empty(&fman->fence_list) &&
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c b/drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c
index 09e10ae..39f2b03 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c
@@ -44,10 +44,10 @@
 		if (!dev_priv->has_mob)
 			return false;
 
-		mutex_lock(&dev_priv->hw_mutex);
+		spin_lock(&dev_priv->cap_lock);
 		vmw_write(dev_priv, SVGA_REG_DEV_CAP, SVGA3D_DEVCAP_3D);
 		result = vmw_read(dev_priv, SVGA_REG_DEV_CAP);
-		mutex_unlock(&dev_priv->hw_mutex);
+		spin_unlock(&dev_priv->cap_lock);
 
 		return (result != 0);
 	}
@@ -120,7 +120,6 @@
 	DRM_INFO("height %d\n", vmw_read(dev_priv, SVGA_REG_HEIGHT));
 	DRM_INFO("bpp %d\n", vmw_read(dev_priv, SVGA_REG_BITS_PER_PIXEL));
 
-	mutex_lock(&dev_priv->hw_mutex);
 	dev_priv->enable_state = vmw_read(dev_priv, SVGA_REG_ENABLE);
 	dev_priv->config_done_state = vmw_read(dev_priv, SVGA_REG_CONFIG_DONE);
 	dev_priv->traces_state = vmw_read(dev_priv, SVGA_REG_TRACES);
@@ -143,7 +142,6 @@
 	mb();
 
 	vmw_write(dev_priv, SVGA_REG_CONFIG_DONE, 1);
-	mutex_unlock(&dev_priv->hw_mutex);
 
 	max = ioread32(fifo_mem + SVGA_FIFO_MAX);
 	min = ioread32(fifo_mem  + SVGA_FIFO_MIN);
@@ -160,31 +158,28 @@
 	return vmw_fifo_send_fence(dev_priv, &dummy);
 }
 
-void vmw_fifo_ping_host_locked(struct vmw_private *dev_priv, uint32_t reason)
+void vmw_fifo_ping_host(struct vmw_private *dev_priv, uint32_t reason)
 {
 	__le32 __iomem *fifo_mem = dev_priv->mmio_virt;
+	static DEFINE_SPINLOCK(ping_lock);
+	unsigned long irq_flags;
 
+	/*
+	 * The ping_lock is needed because we don't have an atomic
+	 * test-and-set of the SVGA_FIFO_BUSY register.
+	 */
+	spin_lock_irqsave(&ping_lock, irq_flags);
 	if (unlikely(ioread32(fifo_mem + SVGA_FIFO_BUSY) == 0)) {
 		iowrite32(1, fifo_mem + SVGA_FIFO_BUSY);
 		vmw_write(dev_priv, SVGA_REG_SYNC, reason);
 	}
-}
-
-void vmw_fifo_ping_host(struct vmw_private *dev_priv, uint32_t reason)
-{
-	mutex_lock(&dev_priv->hw_mutex);
-
-	vmw_fifo_ping_host_locked(dev_priv, reason);
-
-	mutex_unlock(&dev_priv->hw_mutex);
+	spin_unlock_irqrestore(&ping_lock, irq_flags);
 }
 
 void vmw_fifo_release(struct vmw_private *dev_priv, struct vmw_fifo_state *fifo)
 {
 	__le32 __iomem *fifo_mem = dev_priv->mmio_virt;
 
-	mutex_lock(&dev_priv->hw_mutex);
-
 	vmw_write(dev_priv, SVGA_REG_SYNC, SVGA_SYNC_GENERIC);
 	while (vmw_read(dev_priv, SVGA_REG_BUSY) != 0)
 		;
@@ -198,7 +193,6 @@
 	vmw_write(dev_priv, SVGA_REG_TRACES,
 		  dev_priv->traces_state);
 
-	mutex_unlock(&dev_priv->hw_mutex);
 	vmw_marker_queue_takedown(&fifo->marker_queue);
 
 	if (likely(fifo->static_buffer != NULL)) {
@@ -271,7 +265,7 @@
 		return vmw_fifo_wait_noirq(dev_priv, bytes,
 					   interruptible, timeout);
 
-	mutex_lock(&dev_priv->hw_mutex);
+	spin_lock(&dev_priv->waiter_lock);
 	if (atomic_add_return(1, &dev_priv->fifo_queue_waiters) > 0) {
 		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
 		outl(SVGA_IRQFLAG_FIFO_PROGRESS,
@@ -280,7 +274,7 @@
 		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
 	}
-	mutex_unlock(&dev_priv->hw_mutex);
+	spin_unlock(&dev_priv->waiter_lock);
 
 	if (interruptible)
 		ret = wait_event_interruptible_timeout
@@ -296,14 +290,14 @@
 	else if (likely(ret > 0))
 		ret = 0;
 
-	mutex_lock(&dev_priv->hw_mutex);
+	spin_lock(&dev_priv->waiter_lock);
 	if (atomic_dec_and_test(&dev_priv->fifo_queue_waiters)) {
 		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
 		dev_priv->irq_mask &= ~SVGA_IRQFLAG_FIFO_PROGRESS;
 		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
 	}
-	mutex_unlock(&dev_priv->hw_mutex);
+	spin_unlock(&dev_priv->waiter_lock);
 
 	return ret;
 }
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_ioctl.c b/drivers/gpu/drm/vmwgfx/vmwgfx_ioctl.c
index 37881ec..69c8ce23 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_ioctl.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_ioctl.c
@@ -135,13 +135,13 @@
 		(pair_offset + max_size * sizeof(SVGA3dCapPair)) / sizeof(u32);
 	compat_cap->header.type = SVGA3DCAPS_RECORD_DEVCAPS;
 
-	mutex_lock(&dev_priv->hw_mutex);
+	spin_lock(&dev_priv->cap_lock);
 	for (i = 0; i < max_size; ++i) {
 		vmw_write(dev_priv, SVGA_REG_DEV_CAP, i);
 		compat_cap->pairs[i][0] = i;
 		compat_cap->pairs[i][1] = vmw_read(dev_priv, SVGA_REG_DEV_CAP);
 	}
-	mutex_unlock(&dev_priv->hw_mutex);
+	spin_unlock(&dev_priv->cap_lock);
 
 	return 0;
 }
@@ -191,12 +191,12 @@
 		if (num > SVGA3D_DEVCAP_MAX)
 			num = SVGA3D_DEVCAP_MAX;
 
-		mutex_lock(&dev_priv->hw_mutex);
+		spin_lock(&dev_priv->cap_lock);
 		for (i = 0; i < num; ++i) {
 			vmw_write(dev_priv, SVGA_REG_DEV_CAP, i);
 			*bounce32++ = vmw_read(dev_priv, SVGA_REG_DEV_CAP);
 		}
-		mutex_unlock(&dev_priv->hw_mutex);
+		spin_unlock(&dev_priv->cap_lock);
 	} else if (gb_objects) {
 		ret = vmw_fill_compat_cap(dev_priv, bounce, size);
 		if (unlikely(ret != 0))
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_irq.c b/drivers/gpu/drm/vmwgfx/vmwgfx_irq.c
index 0c42376..9fe9827 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_irq.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_irq.c
@@ -62,13 +62,8 @@
 
 static bool vmw_fifo_idle(struct vmw_private *dev_priv, uint32_t seqno)
 {
-	uint32_t busy;
 
-	mutex_lock(&dev_priv->hw_mutex);
-	busy = vmw_read(dev_priv, SVGA_REG_BUSY);
-	mutex_unlock(&dev_priv->hw_mutex);
-
-	return (busy == 0);
+	return (vmw_read(dev_priv, SVGA_REG_BUSY) == 0);
 }
 
 void vmw_update_seqno(struct vmw_private *dev_priv,
@@ -184,7 +179,7 @@
 
 void vmw_seqno_waiter_add(struct vmw_private *dev_priv)
 {
-	mutex_lock(&dev_priv->hw_mutex);
+	spin_lock(&dev_priv->waiter_lock);
 	if (dev_priv->fence_queue_waiters++ == 0) {
 		unsigned long irq_flags;
 
@@ -195,12 +190,12 @@
 		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
 	}
-	mutex_unlock(&dev_priv->hw_mutex);
+	spin_unlock(&dev_priv->waiter_lock);
 }
 
 void vmw_seqno_waiter_remove(struct vmw_private *dev_priv)
 {
-	mutex_lock(&dev_priv->hw_mutex);
+	spin_lock(&dev_priv->waiter_lock);
 	if (--dev_priv->fence_queue_waiters == 0) {
 		unsigned long irq_flags;
 
@@ -209,13 +204,13 @@
 		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
 	}
-	mutex_unlock(&dev_priv->hw_mutex);
+	spin_unlock(&dev_priv->waiter_lock);
 }
 
 
 void vmw_goal_waiter_add(struct vmw_private *dev_priv)
 {
-	mutex_lock(&dev_priv->hw_mutex);
+	spin_lock(&dev_priv->waiter_lock);
 	if (dev_priv->goal_queue_waiters++ == 0) {
 		unsigned long irq_flags;
 
@@ -226,12 +221,12 @@
 		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
 	}
-	mutex_unlock(&dev_priv->hw_mutex);
+	spin_unlock(&dev_priv->waiter_lock);
 }
 
 void vmw_goal_waiter_remove(struct vmw_private *dev_priv)
 {
-	mutex_lock(&dev_priv->hw_mutex);
+	spin_lock(&dev_priv->waiter_lock);
 	if (--dev_priv->goal_queue_waiters == 0) {
 		unsigned long irq_flags;
 
@@ -240,7 +235,7 @@
 		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
 	}
-	mutex_unlock(&dev_priv->hw_mutex);
+	spin_unlock(&dev_priv->waiter_lock);
 }
 
 int vmw_wait_seqno(struct vmw_private *dev_priv,
@@ -315,9 +310,7 @@
 	if (!(dev_priv->capabilities & SVGA_CAP_IRQMASK))
 		return;
 
-	mutex_lock(&dev_priv->hw_mutex);
 	vmw_write(dev_priv, SVGA_REG_IRQMASK, 0);
-	mutex_unlock(&dev_priv->hw_mutex);
 
 	status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
 	outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_kms.c b/drivers/gpu/drm/vmwgfx/vmwgfx_kms.c
index 3725b52..8725b79 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_kms.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_kms.c
@@ -1828,9 +1828,7 @@
 	struct vmw_private *dev_priv = vmw_priv(dev);
 	struct vmw_display_unit *du = vmw_connector_to_du(connector);
 
-	mutex_lock(&dev_priv->hw_mutex);
 	num_displays = vmw_read(dev_priv, SVGA_REG_NUM_DISPLAYS);
-	mutex_unlock(&dev_priv->hw_mutex);
 
 	return ((vmw_connector_to_du(connector)->unit < num_displays &&
 		 du->pref_active) ?