drm/nouveau: kill tasks causing the GPU to raise errors

index 373950e..cc6dfe4 100644
@@ -231,6 +231,9 @@ nouveau_channel_alloc(struct drm_device *dev, struct nouveau_channel **chan_ret,
 
 	nouveau_debugfs_channel_init(chan);
 
+	chan->state = NV_CHANNEL_OK;
+	chan->error_action = nouveau_default_channel_error_action;
+
 	NV_INFO(dev, "%s: initialised FIFO %d\n", __func__, channel);
 	*chan_ret = chan;
 	return 0;

@@ -416,6 +419,26 @@ nouveau_ioctl_fifo_free(struct drm_device *dev, void *data,
 	return 0;
 }
 
+int nouveau_handle_state(struct nouveau_channel *chan)
+{
+	if (chan->state == NV_CHANNEL_KILLED)
+		return -EBADFD;
+
+	if (chan->state == NV_CHANNEL_BUS) {
+		NV_INFO(chan->dev, "sending SIGBUS to task %d \"%s\"\n",
+					task_pid_nr(current), current->comm);
+		kill_pid(task_pid(current), SIGBUS, 0);
+		chan->state = NV_CHANNEL_KILL;
+	} else if (chan->state == NV_CHANNEL_KILL) {
+		NV_INFO(chan->dev, "sending SIGKILL to task %d \"%s\"\n",
+					task_pid_nr(current), current->comm);
+		kill_pid(task_pid(current), SIGKILL, 0);
+		chan->state = NV_CHANNEL_KILLED;
+	}
+
+	return -EBADFD;
+}
+
 /***********************************
  * finally, the ioctl table
  ***********************************/

index 9087549..3463780 100644
@@ -115,6 +115,13 @@ MODULE_PARM_DESC(perflvl_wr, "Allow perflvl changes (warning: dangerous!)\n");
 int nouveau_perflvl_wr;
 module_param_named(perflvl_wr, nouveau_perflvl_wr, int, 0400);
 
+MODULE_PARM_DESC(error_action, "Default action on error:\n"
+		"\t\t0 - do nothing\n"
+		"\t\t1 - kill process owning the offending channel (default)");
+int nouveau_default_channel_error_action = NV_KILL_PROCESS;
+module_param_named(error_action, nouveau_default_channel_error_action,
+					int, 0400);
+
 int nouveau_fbpercrtc;
 #if 0
 module_param_named(fbpercrtc, nouveau_fbpercrtc, int, 0400);

index 1c7db64..70a320d 100644
@@ -166,6 +166,13 @@ struct nouveau_channel {
 	struct drm_device *dev;
 	int id;
 
+	enum nouveau_channel_state {
+		NV_CHANNEL_OK = 0,
+		NV_CHANNEL_BUS = 1,
+		NV_CHANNEL_KILL = 2,
+		NV_CHANNEL_KILLED = 3,
+	} state;
+
 	/* owner of this fifo */
 	struct drm_file *file_priv;
 	/* mapping of the fifo itself */

@@ -248,6 +255,11 @@ struct nouveau_channel {
 		char name[32];
 		struct drm_info_list info;
 	} debugfs;
+
+	enum nouveau_channel_error_action {
+		NV_DO_NOTHING = 0,
+		NV_KILL_PROCESS = 1,
+	} error_action;
 };
 
 struct nouveau_instmem_engine {

@@ -740,6 +752,7 @@ extern int nouveau_fbpercrtc;
 extern int nouveau_tv_disable;
 extern char *nouveau_tv_norm;
 extern int nouveau_reg_debug;
+extern int nouveau_default_channel_error_action;
 extern char *nouveau_vbios;
 extern int nouveau_ignorelid;
 extern int nouveau_nofbaccel;

@@ -810,6 +823,7 @@ extern int  nouveau_channel_alloc(struct drm_device *dev,
 				  struct drm_file *file_priv,
 				  uint32_t fb_ctxdma, uint32_t tt_ctxdma);
 extern void nouveau_channel_free(struct nouveau_channel *);
+extern int nouveau_handle_state(struct nouveau_channel *chan);
 
 /* nouveau_object.c */
 extern int  nouveau_gpuobj_early_init(struct drm_device *);

@@ -850,6 +864,8 @@ extern irqreturn_t nouveau_irq_handler(DRM_IRQ_ARGS);
 extern void        nouveau_irq_preinstall(struct drm_device *);
 extern int         nouveau_irq_postinstall(struct drm_device *);
 extern void        nouveau_irq_uninstall(struct drm_device *);
+extern void        nouveau_broken_channel(struct nouveau_channel *chan);
+extern void        nouveau_broken_channel_id(struct drm_device *dev, int chid);
 
 /* nouveau_sgdma.c */
 extern int nouveau_sgdma_init(struct drm_device *);

index 2e21412..36ecbb7 100644
@@ -149,6 +149,8 @@ nouveau_gem_ioctl_new(struct drm_device *dev, void *data,
 	if (req->channel_hint) {
 		NOUVEAU_GET_USER_CHANNEL_WITH_RETURN(req->channel_hint,
 						     file_priv, chan);
+		if (unlikely(chan->state != NV_CHANNEL_OK))
+			return nouveau_handle_state(chan);
 	}
 
 	if (req->info.domain & NOUVEAU_GEM_DOMAIN_VRAM)

@@ -594,6 +596,8 @@ nouveau_gem_ioctl_pushbuf(struct drm_device *dev, void *data,
 	int i, j, ret = 0, do_reloc = 0;
 
 	NOUVEAU_GET_USER_CHANNEL_WITH_RETURN(req->channel, file_priv, chan);
+	if (unlikely(chan->state != NV_CHANNEL_OK))
+		return nouveau_handle_state(chan);
 
 	req->vram_available = dev_priv->fb_aper_free;
 	req->gart_available = dev_priv->gart_info.aper_free;

229 lines not on this page · raw message

index 2cc59f8..caafbef 100644
@@ -186,6 +186,8 @@ nouveau_ioctl_notifier_alloc(struct drm_device *dev, void *data,
 	int ret;
 
 	NOUVEAU_GET_USER_CHANNEL_WITH_RETURN(na->channel, file_priv, chan);
+	if (unlikely(chan->state != NV_CHANNEL_OK))
+		return nouveau_handle_state(chan);
 
 	ret = nouveau_notifier_alloc(chan, na->handle, na->size, &na->offset);
 	if (ret)

index dd572ad..704705e 100644
@@ -877,6 +877,8 @@ int nouveau_ioctl_grobj_alloc(struct drm_device *dev, void *data,
 	int ret;
 
 	NOUVEAU_GET_USER_CHANNEL_WITH_RETURN(init->channel, file_priv, chan);
+	if (unlikely(chan->state != NV_CHANNEL_OK))
+		return nouveau_handle_state(chan);
 
 	if (init->handle == ~0)
 		return -EINVAL;

index cd1988b..40a795d 100644
@@ -56,9 +56,6 @@ nv50_fb_vm_trap(struct drm_device *dev, int display, const char *name)
 	}
 	nv_wr32(dev, 0x100c90, idx | 0x80000000);
 
-	if (!display)
-		return;
-
 	chinst = (trap[2] << 16) | trap[1];
 	for (ch = 0; ch < dev_priv->engine.fifo.channels; ch++) {
 		struct nouveau_channel *chan = dev_priv->fifos[ch];

@@ -70,9 +67,13 @@ nv50_fb_vm_trap(struct drm_device *dev, int display, const char *name)
 			break;
 	}
 
-	NV_INFO(dev, "%s - VM: Trapped %s at %02x%04x%04x status %08x "
-		     "channel %d (0x%08x)\n",
-		name, (trap[5] & 0x100 ? "read" : "write"),
-		trap[5] & 0xff, trap[4] & 0xffff, trap[3] & 0xffff,
-		trap[0], ch, chinst);
+	if (display) {
+		NV_INFO(dev, "%s - VM: Trapped %s at %02x%04x%04x status %08x "
+			     "channel %d (0x%08x)\n",
+			name, (trap[5] & 0x100 ? "read" : "write"),
+			trap[5] & 0xff, trap[4] & 0xffff, trap[3] & 0xffff,
+			trap[0], ch, chinst);
+	}
+
+	nouveau_broken_channel_id(dev, ch);
 }