drm/nouveau: Refactor context destruction to avoid a lock ordering issue.

The destroy_context() engine hooks call gpuobj management functions to
release the channel resources, these functions use HARDIRQ-unsafe locks
whereas destroy_context() is called with the HARDIRQ-safe
context_switch_lock held, that's a lock ordering violation.

Push the engine-specific channel destruction logic into destroy_context()
and let the hardware-specific code lock and unlock when it's actually
needed. Change the engine destruction order to avoid a race in the small
gap between pgraph and pfifo context uninitialization.

Reported-by: Marcin Slusarz <marcin.slusarz@gmail.com>
Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>

drivers/gpu/drm/nouveau/nouveau_channel.c

@@ -313,32 +313,20 @@ nouveau_channel_put(struct nouveau_channel **pchan)
 	/* boot it off the hardware */
 	pfifo->reassign(dev, false);
 
-	/* We want to give pgraph a chance to idle and get rid of all potential
-	 * errors. We need to do this before the lock, otherwise the irq handler
-	 * is unable to process them.
+	/* We want to give pgraph a chance to idle and get rid of all
+	 * potential errors. We need to do this without the context
+	 * switch lock held, otherwise the irq handler is unable to
+	 * process them.
 	 */
 	if (pgraph->channel(dev) == chan)
 		nouveau_wait_for_idle(dev);
 
-	spin_lock_irqsave(&dev_priv->context_switch_lock, flags);
-
-	pgraph->fifo_access(dev, false);
-	if (pgraph->channel(dev) == chan)
-		pgraph->unload_context(dev);
-	pgraph->destroy_context(chan);
-	pgraph->fifo_access(dev, true);
-
-	if (pfifo->channel_id(dev) == chan->id) {
-		pfifo->disable(dev);
-		pfifo->unload_context(dev);
-		pfifo->enable(dev);
-	}
+	/* destroy the engine specific contexts */
 	pfifo->destroy_context(chan);
+	pgraph->destroy_context(chan);
 
 	pfifo->reassign(dev, true);
 
-	spin_unlock_irqrestore(&dev_priv->context_switch_lock, flags);
-
 	/* aside from its resources, the channel should now be dead,
 	 * remove it from the channel list
 	 */

drivers/gpu/drm/nouveau/nouveau_drv.h

@@ -998,14 +998,12 @@ extern int  nv04_fifo_unload_context(struct drm_device *);
 extern int  nv10_fifo_init(struct drm_device *);
 extern int  nv10_fifo_channel_id(struct drm_device *);
 extern int  nv10_fifo_create_context(struct nouveau_channel *);
-extern void nv10_fifo_destroy_context(struct nouveau_channel *);
 extern int  nv10_fifo_load_context(struct nouveau_channel *);
 extern int  nv10_fifo_unload_context(struct drm_device *);
 
 /* nv40_fifo.c */
 extern int  nv40_fifo_init(struct drm_device *);
 extern int  nv40_fifo_create_context(struct nouveau_channel *);
-extern void nv40_fifo_destroy_context(struct nouveau_channel *);
 extern int  nv40_fifo_load_context(struct nouveau_channel *);
 extern int  nv40_fifo_unload_context(struct drm_device *);
 

drivers/gpu/drm/nouveau/nouveau_state.c

@@ -137,7 +137,7 @@ static int nouveau_init_engine_ptrs(struct drm_device *dev)
 		engine->fifo.cache_pull		= nv04_fifo_cache_pull;
 		engine->fifo.channel_id		= nv10_fifo_channel_id;
 		engine->fifo.create_context	= nv10_fifo_create_context;
-		engine->fifo.destroy_context	= nv10_fifo_destroy_context;
+		engine->fifo.destroy_context	= nv04_fifo_destroy_context;
 		engine->fifo.load_context	= nv10_fifo_load_context;
 		engine->fifo.unload_context	= nv10_fifo_unload_context;
 		engine->display.early_init	= nv04_display_early_init;
@@ -191,7 +191,7 @@ static int nouveau_init_engine_ptrs(struct drm_device *dev)
 		engine->fifo.cache_pull		= nv04_fifo_cache_pull;
 		engine->fifo.channel_id		= nv10_fifo_channel_id;
 		engine->fifo.create_context	= nv10_fifo_create_context;
-		engine->fifo.destroy_context	= nv10_fifo_destroy_context;
+		engine->fifo.destroy_context	= nv04_fifo_destroy_context;
 		engine->fifo.load_context	= nv10_fifo_load_context;
 		engine->fifo.unload_context	= nv10_fifo_unload_context;
 		engine->display.early_init	= nv04_display_early_init;
@@ -245,7 +245,7 @@ static int nouveau_init_engine_ptrs(struct drm_device *dev)
 		engine->fifo.cache_pull		= nv04_fifo_cache_pull;
 		engine->fifo.channel_id		= nv10_fifo_channel_id;
 		engine->fifo.create_context	= nv10_fifo_create_context;
-		engine->fifo.destroy_context	= nv10_fifo_destroy_context;
+		engine->fifo.destroy_context	= nv04_fifo_destroy_context;
 		engine->fifo.load_context	= nv10_fifo_load_context;
 		engine->fifo.unload_context	= nv10_fifo_unload_context;
 		engine->display.early_init	= nv04_display_early_init;
@@ -302,7 +302,7 @@ static int nouveau_init_engine_ptrs(struct drm_device *dev)
 		engine->fifo.cache_pull		= nv04_fifo_cache_pull;
 		engine->fifo.channel_id		= nv10_fifo_channel_id;
 		engine->fifo.create_context	= nv40_fifo_create_context;
-		engine->fifo.destroy_context	= nv40_fifo_destroy_context;
+		engine->fifo.destroy_context	= nv04_fifo_destroy_context;
 		engine->fifo.load_context	= nv40_fifo_load_context;
 		engine->fifo.unload_context	= nv40_fifo_unload_context;
 		engine->display.early_init	= nv04_display_early_init;

drivers/gpu/drm/nouveau/nv04_fifo.c

@@ -151,10 +151,27 @@ void
 nv04_fifo_destroy_context(struct nouveau_channel *chan)
 {
 	struct drm_device *dev = chan->dev;
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nouveau_fifo_engine *pfifo = &dev_priv->engine.fifo;
+	unsigned long flags;
 
-	nv_wr32(dev, NV04_PFIFO_MODE,
-		nv_rd32(dev, NV04_PFIFO_MODE) & ~(1 << chan->id));
+	spin_lock_irqsave(&dev_priv->context_switch_lock, flags);
+	pfifo->reassign(dev, false);
+
+	/* Unload the context if it's the currently active one */
+	if (pfifo->channel_id(dev) == chan->id) {
+		pfifo->disable(dev);
+		pfifo->unload_context(dev);
+		pfifo->enable(dev);
+	}
+
+	/* Keep it from being rescheduled */
+	nv_mask(dev, NV04_PFIFO_MODE, 1 << chan->id, 0);
+
+	pfifo->reassign(dev, true);
+	spin_unlock_irqrestore(&dev_priv->context_switch_lock, flags);
 
+	/* Free the channel resources */
 	nouveau_gpuobj_ref(NULL, &chan->ramfc);
 }
 

drivers/gpu/drm/nouveau/nv04_graph.c

@@ -412,10 +412,25 @@ int nv04_graph_create_context(struct nouveau_channel *chan)
 
 void nv04_graph_destroy_context(struct nouveau_channel *chan)
 {
+	struct drm_device *dev = chan->dev;
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nouveau_pgraph_engine *pgraph = &dev_priv->engine.graph;
 	struct graph_state *pgraph_ctx = chan->pgraph_ctx;
+	unsigned long flags;
+
+	spin_lock_irqsave(&dev_priv->context_switch_lock, flags);
+	pgraph->fifo_access(dev, false);
+
+	/* Unload the context if it's the currently active one */
+	if (pgraph->channel(dev) == chan)
+		pgraph->unload_context(dev);
 
+	/* Free the context resources */
 	kfree(pgraph_ctx);
 	chan->pgraph_ctx = NULL;
+
+	pgraph->fifo_access(dev, true);
+	spin_unlock_irqrestore(&dev_priv->context_switch_lock, flags);
 }
 
 int nv04_graph_load_context(struct nouveau_channel *chan)

drivers/gpu/drm/nouveau/nv10_fifo.c

@@ -73,17 +73,6 @@ nv10_fifo_create_context(struct nouveau_channel *chan)
 	return 0;
 }
 
-void
-nv10_fifo_destroy_context(struct nouveau_channel *chan)
-{
-	struct drm_device *dev = chan->dev;
-
-	nv_wr32(dev, NV04_PFIFO_MODE,
-			nv_rd32(dev, NV04_PFIFO_MODE) & ~(1 << chan->id));
-
-	nouveau_gpuobj_ref(NULL, &chan->ramfc);
-}
-
 static void
 nv10_fifo_do_load_context(struct drm_device *dev, int chid)
 {

drivers/gpu/drm/nouveau/nv10_graph.c

@@ -875,10 +875,25 @@ int nv10_graph_create_context(struct nouveau_channel *chan)
 
 void nv10_graph_destroy_context(struct nouveau_channel *chan)
 {
+	struct drm_device *dev = chan->dev;
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nouveau_pgraph_engine *pgraph = &dev_priv->engine.graph;
 	struct graph_state *pgraph_ctx = chan->pgraph_ctx;
+	unsigned long flags;
+
+	spin_lock_irqsave(&dev_priv->context_switch_lock, flags);
+	pgraph->fifo_access(dev, false);
+
+	/* Unload the context if it's the currently active one */
+	if (pgraph->channel(dev) == chan)
+		pgraph->unload_context(dev);
 
+	/* Free the context resources */
 	kfree(pgraph_ctx);
 	chan->pgraph_ctx = NULL;
+
+	pgraph->fifo_access(dev, true);
+	spin_unlock_irqrestore(&dev_priv->context_switch_lock, flags);
 }
 
 void

drivers/gpu/drm/nouveau/nv20_graph.c

@@ -425,9 +425,21 @@ nv20_graph_destroy_context(struct nouveau_channel *chan)
 	struct drm_device *dev = chan->dev;
 	struct drm_nouveau_private *dev_priv = dev->dev_private;
 	struct nouveau_pgraph_engine *pgraph = &dev_priv->engine.graph;
+	unsigned long flags;
 
-	nouveau_gpuobj_ref(NULL, &chan->ramin_grctx);
+	spin_lock_irqsave(&dev_priv->context_switch_lock, flags);
+	pgraph->fifo_access(dev, false);
+
+	/* Unload the context if it's the currently active one */
+	if (pgraph->channel(dev) == chan)
+		pgraph->unload_context(dev);
+
+	pgraph->fifo_access(dev, true);
+	spin_unlock_irqrestore(&dev_priv->context_switch_lock, flags);
+
+	/* Free the context resources */
 	nv_wo32(pgraph->ctx_table, chan->id * 4, 0);
+	nouveau_gpuobj_ref(NULL, &chan->ramin_grctx);
 }
 
 int

drivers/gpu/drm/nouveau/nv40_fifo.c

@@ -70,17 +70,6 @@ nv40_fifo_create_context(struct nouveau_channel *chan)
 	return 0;
 }
 
-void
-nv40_fifo_destroy_context(struct nouveau_channel *chan)
-{
-	struct drm_device *dev = chan->dev;
-
-	nv_wr32(dev, NV04_PFIFO_MODE,
-		nv_rd32(dev, NV04_PFIFO_MODE) & ~(1 << chan->id));
-
-	nouveau_gpuobj_ref(NULL, &chan->ramfc);
-}
-
 static void
 nv40_fifo_do_load_context(struct drm_device *dev, int chid)
 {

drivers/gpu/drm/nouveau/nv40_graph.c

@@ -79,6 +79,22 @@ nv40_graph_create_context(struct nouveau_channel *chan)
 void
 nv40_graph_destroy_context(struct nouveau_channel *chan)
 {
+	struct drm_device *dev = chan->dev;
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nouveau_pgraph_engine *pgraph = &dev_priv->engine.graph;
+	unsigned long flags;
+
+	spin_lock_irqsave(&dev_priv->context_switch_lock, flags);
+	pgraph->fifo_access(dev, false);
+
+	/* Unload the context if it's the currently active one */
+	if (pgraph->channel(dev) == chan)
+		pgraph->unload_context(dev);
+
+	pgraph->fifo_access(dev, true);
+	spin_unlock_irqrestore(&dev_priv->context_switch_lock, flags);
+
+	/* Free the context resources */
 	nouveau_gpuobj_ref(NULL, &chan->ramin_grctx);
 }
 

drivers/gpu/drm/nouveau/nv50_fifo.c

@@ -292,10 +292,23 @@ void
 nv50_fifo_destroy_context(struct nouveau_channel *chan)
 {
 	struct drm_device *dev = chan->dev;
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nouveau_fifo_engine *pfifo = &dev_priv->engine.fifo;
 	struct nouveau_gpuobj *ramfc = NULL;
+	unsigned long flags;
 
 	NV_DEBUG(dev, "ch%d\n", chan->id);
 
+	spin_lock_irqsave(&dev_priv->context_switch_lock, flags);
+	pfifo->reassign(dev, false);
+
+	/* Unload the context if it's the currently active one */
+	if (pfifo->channel_id(dev) == chan->id) {
+		pfifo->disable(dev);
+		pfifo->unload_context(dev);
+		pfifo->enable(dev);
+	}
+
 	/* This will ensure the channel is seen as disabled. */
 	nouveau_gpuobj_ref(chan->ramfc, &ramfc);
 	nouveau_gpuobj_ref(NULL, &chan->ramfc);
@@ -306,6 +319,10 @@ nv50_fifo_destroy_context(struct nouveau_channel *chan)
 		nv50_fifo_channel_disable(dev, 127);
 	nv50_fifo_playlist_update(dev);
 
+	pfifo->reassign(dev, true);
+	spin_unlock_irqrestore(&dev_priv->context_switch_lock, flags);
+
+	/* Free the channel resources */
 	nouveau_gpuobj_ref(NULL, &ramfc);
 	nouveau_gpuobj_ref(NULL, &chan->cache);
 }

drivers/gpu/drm/nouveau/nv50_graph.c

@@ -242,17 +242,28 @@ nv50_graph_destroy_context(struct nouveau_channel *chan)
 {
 	struct drm_device *dev = chan->dev;
 	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nouveau_pgraph_engine *pgraph = &dev_priv->engine.graph;
 	int i, hdr = (dev_priv->chipset == 0x50) ? 0x200 : 0x20;
+	unsigned long flags;
 
 	NV_DEBUG(dev, "ch%d\n", chan->id);
 
 	if (!chan->ramin)
 		return;
 
+	spin_lock_irqsave(&dev_priv->context_switch_lock, flags);
+	pgraph->fifo_access(dev, false);
+
+	if (pgraph->channel(dev) == chan)
+		pgraph->unload_context(dev);
+
 	for (i = hdr; i < hdr + 24; i += 4)
 		nv_wo32(chan->ramin, i, 0);
 	dev_priv->engine.instmem.flush(dev);
 
+	pgraph->fifo_access(dev, true);
+	spin_unlock_irqrestore(&dev_priv->context_switch_lock, flags);
+
 	nouveau_gpuobj_ref(NULL, &chan->ramin_grctx);
 }