drm/nv50/gr: move to exec engine interfaces

This needs a massive cleanup, but to catch bugs from the interface changes
vs the engine code cleanup, this will be done later.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>

8 files changed, 252 insertions, 246 deletions

diff --git a/drivers/gpu/drm/nouveau/nouveau_channel.c b/drivers/gpu/drm/nouveau/nouveau_channel.c
index 3a047ec..8b0a668 100644
--- a/drivers/gpu/drm/nouveau/nouveau_channel.c
+++ b/drivers/gpu/drm/nouveau/nouveau_channel.c
@@ -304,7 +304,8 @@ nouveau_channel_put_unlocked(struct nouveau_channel **pchan)
 
 	/* destroy the engine specific contexts */
 	pfifo->destroy_context(chan);
-	pgraph->destroy_context(chan);
+	if (pgraph->destroy_context)
+		pgraph->destroy_context(chan);
 	for (i = 0; i < NVOBJ_ENGINE_NR; i++) {
 		if (chan->engctx[i])
 			dev_priv->eng[i]->context_del(chan, i);
diff --git a/drivers/gpu/drm/nouveau/nouveau_drv.c b/drivers/gpu/drm/nouveau/nouveau_drv.c
index 30b9e89..ba87173 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drv.c
+++ b/drivers/gpu/drm/nouveau/nouveau_drv.c
@@ -217,7 +217,8 @@ nouveau_pci_suspend(struct pci_dev *pdev, pm_message_t pm_state)
 	pfifo->reassign(dev, false);
 	pfifo->disable(dev);
 	pfifo->unload_context(dev);
-	pgraph->unload_context(dev);
+	if (pgraph->unload_context)
+		pgraph->unload_context(dev);
 
 	ret = pinstmem->suspend(dev);
 	if (ret) {
diff --git a/drivers/gpu/drm/nouveau/nouveau_drv.h b/drivers/gpu/drm/nouveau/nouveau_drv.h
index 6b43cb0..66591d3 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drv.h
+++ b/drivers/gpu/drm/nouveau/nouveau_drv.h
@@ -1189,18 +1189,8 @@ extern void nv40_grctx_init(struct nouveau_grctx *);
 extern void nv40_graph_set_tile_region(struct drm_device *dev, int i);
 
 /* nv50_graph.c */
-extern int  nv50_graph_init(struct drm_device *);
-extern void nv50_graph_takedown(struct drm_device *);
-extern void nv50_graph_fifo_access(struct drm_device *, bool);
-extern struct nouveau_channel *nv50_graph_channel(struct drm_device *);
-extern int  nv50_graph_create_context(struct nouveau_channel *);
-extern void nv50_graph_destroy_context(struct nouveau_channel *);
-extern int  nv50_graph_load_context(struct nouveau_channel *);
-extern int  nv50_graph_unload_context(struct drm_device *);
-extern int  nv50_graph_object_new(struct nouveau_channel *, u32, u16);
+extern int  nv50_graph_create(struct drm_device *);
 extern int  nv50_grctx_init(struct nouveau_grctx *);
-extern void nv50_graph_tlb_flush(struct drm_device *dev);
-extern void nv84_graph_tlb_flush(struct drm_device *dev);
 extern struct nouveau_enum nv50_data_error_names[];
 
 /* nvc0_graph.c */
diff --git a/drivers/gpu/drm/nouveau/nouveau_object.c b/drivers/gpu/drm/nouveau/nouveau_object.c
index 4fb05b6..9ea3ab9 100644
--- a/drivers/gpu/drm/nouveau/nouveau_object.c
+++ b/drivers/gpu/drm/nouveau/nouveau_object.c
@@ -636,18 +636,20 @@ nouveau_gpuobj_gr_new(struct nouveau_channel *chan, u32 handle, int class)
 	return -EINVAL;
 
 found:
-	switch (oc->engine) {
-	case NVOBJ_ENGINE_SW:
-		return nouveau_gpuobj_sw_new(chan, handle, class);
-	case NVOBJ_ENGINE_GR:
-		if ((dev_priv->card_type >= NV_20 && !chan->ramin_grctx) ||
-		    (dev_priv->card_type  < NV_20 && !chan->pgraph_ctx)) {
-			ret = pgraph->create_context(chan);
-			if (ret)
-				return ret;
-		}
+	if (!dev_priv->eng[oc->engine]) {
+		switch (oc->engine) {
+		case NVOBJ_ENGINE_SW:
+			return nouveau_gpuobj_sw_new(chan, handle, class);
+		case NVOBJ_ENGINE_GR:
+			if ((dev_priv->card_type >= NV_20 && !chan->ramin_grctx) ||
+			    (dev_priv->card_type  < NV_20 && !chan->pgraph_ctx)) {
+				ret = pgraph->create_context(chan);
+				if (ret)
+					return ret;
+			}
 
-		return pgraph->object_new(chan, handle, class);
+			return pgraph->object_new(chan, handle, class);
+		}
 	}
 
 	if (!chan->engctx[oc->engine]) {
diff --git a/drivers/gpu/drm/nouveau/nouveau_state.c b/drivers/gpu/drm/nouveau/nouveau_state.c
index a1a25ea..6fd6736 100644
--- a/drivers/gpu/drm/nouveau/nouveau_state.c
+++ b/drivers/gpu/drm/nouveau/nouveau_state.c
@@ -363,20 +363,10 @@ static int nouveau_init_engine_ptrs(struct drm_device *dev)
 		engine->timer.takedown		= nv04_timer_takedown;
 		engine->fb.init			= nv50_fb_init;
 		engine->fb.takedown		= nv50_fb_takedown;
-		engine->graph.init		= nv50_graph_init;
-		engine->graph.takedown		= nv50_graph_takedown;
-		engine->graph.fifo_access	= nv50_graph_fifo_access;
-		engine->graph.channel		= nv50_graph_channel;
-		engine->graph.create_context	= nv50_graph_create_context;
-		engine->graph.destroy_context	= nv50_graph_destroy_context;
-		engine->graph.load_context	= nv50_graph_load_context;
-		engine->graph.unload_context	= nv50_graph_unload_context;
-		engine->graph.object_new	= nv50_graph_object_new;
-		if (dev_priv->chipset == 0x50 ||
-		    dev_priv->chipset == 0xac)
-			engine->graph.tlb_flush	= nv50_graph_tlb_flush;
-		else
-			engine->graph.tlb_flush	= nv84_graph_tlb_flush;
+		engine->graph.init		= nouveau_stub_init;
+		engine->graph.takedown		= nouveau_stub_takedown;
+		engine->graph.fifo_access	= nvc0_graph_fifo_access;
+		engine->graph.channel		= nvc0_graph_channel;
 		engine->fifo.channels		= 128;
 		engine->fifo.init		= nv50_fifo_init;
 		engine->fifo.takedown		= nv50_fifo_takedown;
@@ -635,6 +625,9 @@ nouveau_card_init(struct drm_device *dev)
 	if (ret)
 		goto out_timer;
 
+	if (dev_priv->card_type == NV_50)
+		nv50_graph_create(dev);
+
 	switch (dev_priv->chipset) {
 	case 0x84:
 	case 0x86:
@@ -712,8 +705,10 @@ out_graph:
 out_engine:
 	if (!nouveau_noaccel) {
 		for (e = e - 1; e >= 0; e--) {
+			if (!dev_priv->eng[e])
+				continue;
 			dev_priv->eng[e]->fini(dev, e);
-			dev_priv->eng[e]->destroy(dev, e);
+			dev_priv->eng[e]->destroy(dev,e );
 		}
 	}
 
diff --git a/drivers/gpu/drm/nouveau/nouveau_vm.h b/drivers/gpu/drm/nouveau/nouveau_vm.h
index 16ffc4c..c48a9fc 100644
--- a/drivers/gpu/drm/nouveau/nouveau_vm.h
+++ b/drivers/gpu/drm/nouveau/nouveau_vm.h
@@ -53,7 +53,6 @@ struct nouveau_vm {
 	int refcount;
 
 	struct list_head pgd_list;
-	atomic_t pgraph_refs;
 	atomic_t engref[16];
 
 	struct nouveau_vm_pgt *pgt;
diff --git a/drivers/gpu/drm/nouveau/nv50_graph.c b/drivers/gpu/drm/nouveau/nv50_graph.c
index eb83179..bffa486 100644
--- a/drivers/gpu/drm/nouveau/nv50_graph.c
+++ b/drivers/gpu/drm/nouveau/nv50_graph.c
@@ -34,8 +34,92 @@
 #include "nouveau_ramht.h"
 #include "nv50_evo.h"
 
-static int  nv50_graph_register(struct drm_device *);
-static void nv50_graph_isr(struct drm_device *);
+struct nv50_graph_engine {
+	struct nouveau_exec_engine base;
+	u32 ctxprog[512];
+	u32 ctxprog_size;
+	u32 grctx_size;
+};
+
+static void
+nv50_graph_fifo_access(struct drm_device *dev, bool enabled)
+{
+	const uint32_t mask = 0x00010001;
+
+	if (enabled)
+		nv_wr32(dev, 0x400500, nv_rd32(dev, 0x400500) | mask);
+	else
+		nv_wr32(dev, 0x400500, nv_rd32(dev, 0x400500) & ~mask);
+}
+
+static struct nouveau_channel *
+nv50_graph_channel(struct drm_device *dev)
+{
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	uint32_t inst;
+	int i;
+
+	/* Be sure we're not in the middle of a context switch or bad things
+	 * will happen, such as unloading the wrong pgraph context.
+	 */
+	if (!nv_wait(dev, 0x400300, 0x00000001, 0x00000000))
+		NV_ERROR(dev, "Ctxprog is still running\n");
+
+	inst = nv_rd32(dev, NV50_PGRAPH_CTXCTL_CUR);
+	if (!(inst & NV50_PGRAPH_CTXCTL_CUR_LOADED))
+		return NULL;
+	inst = (inst & NV50_PGRAPH_CTXCTL_CUR_INSTANCE) << 12;
+
+	for (i = 0; i < dev_priv->engine.fifo.channels; i++) {
+		struct nouveau_channel *chan = dev_priv->channels.ptr[i];
+
+		if (chan && chan->ramin && chan->ramin->vinst == inst)
+			return chan;
+	}
+
+	return NULL;
+}
+
+static int
+nv50_graph_do_load_context(struct drm_device *dev, uint32_t inst)
+{
+	uint32_t fifo = nv_rd32(dev, 0x400500);
+
+	nv_wr32(dev, 0x400500, fifo & ~1);
+	nv_wr32(dev, 0x400784, inst);
+	nv_wr32(dev, 0x400824, nv_rd32(dev, 0x400824) | 0x40);
+	nv_wr32(dev, 0x400320, nv_rd32(dev, 0x400320) | 0x11);
+	nv_wr32(dev, 0x400040, 0xffffffff);
+	(void)nv_rd32(dev, 0x400040);
+	nv_wr32(dev, 0x400040, 0x00000000);
+	nv_wr32(dev, 0x400304, nv_rd32(dev, 0x400304) | 1);
+
+	if (nouveau_wait_for_idle(dev))
+		nv_wr32(dev, 0x40032c, inst | (1<<31));
+	nv_wr32(dev, 0x400500, fifo);
+
+	return 0;
+}
+
+static int
+nv50_graph_unload_context(struct drm_device *dev)
+{
+	uint32_t inst;
+
+	inst  = nv_rd32(dev, NV50_PGRAPH_CTXCTL_CUR);
+	if (!(inst & NV50_PGRAPH_CTXCTL_CUR_LOADED))
+		return 0;
+	inst &= NV50_PGRAPH_CTXCTL_CUR_INSTANCE;
+
+	nouveau_wait_for_idle(dev);
+	nv_wr32(dev, 0x400784, inst);
+	nv_wr32(dev, 0x400824, nv_rd32(dev, 0x400824) | 0x20);
+	nv_wr32(dev, 0x400304, nv_rd32(dev, 0x400304) | 0x01);
+	nouveau_wait_for_idle(dev);
+
+	nv_wr32(dev, NV50_PGRAPH_CTXCTL_CUR, inst);
+	return 0;
+}
 
 static void
 nv50_graph_init_reset(struct drm_device *dev)
@@ -53,7 +137,6 @@ nv50_graph_init_intr(struct drm_device *dev)
 {
 	NV_DEBUG(dev, "\n");
 
-	nouveau_irq_register(dev, 12, nv50_graph_isr);
 	nv_wr32(dev, NV03_PGRAPH_INTR, 0xffffffff);
 	nv_wr32(dev, 0x400138, 0xffffffff);
 	nv_wr32(dev, NV40_PGRAPH_INTR_EN, 0xffffffff);
@@ -136,34 +219,14 @@ nv50_graph_init_zcull(struct drm_device *dev)
 static int
 nv50_graph_init_ctxctl(struct drm_device *dev)
 {
-	struct drm_nouveau_private *dev_priv = dev->dev_private;
-	struct nouveau_grctx ctx = {};
-	uint32_t *cp;
+	struct nv50_graph_engine *pgraph = nv_engine(dev, NVOBJ_ENGINE_GR);
 	int i;
 
 	NV_DEBUG(dev, "\n");
 
-	cp = kmalloc(512 * 4, GFP_KERNEL);
-	if (!cp) {
-		NV_ERROR(dev, "failed to allocate ctxprog\n");
-		dev_priv->engine.graph.accel_blocked = true;
-		return 0;
-	}
-
-	ctx.dev = dev;
-	ctx.mode = NOUVEAU_GRCTX_PROG;
-	ctx.data = cp;
-	ctx.ctxprog_max = 512;
-	if (!nv50_grctx_init(&ctx)) {
-		dev_priv->engine.graph.grctx_size = ctx.ctxvals_pos * 4;
-
-		nv_wr32(dev, NV40_PGRAPH_CTXCTL_UCODE_INDEX, 0);
-		for (i = 0; i < ctx.ctxprog_len; i++)
-			nv_wr32(dev, NV40_PGRAPH_CTXCTL_UCODE_DATA, cp[i]);
-	} else {
-		dev_priv->engine.graph.accel_blocked = true;
-	}
-	kfree(cp);
+	nv_wr32(dev, NV40_PGRAPH_CTXCTL_UCODE_INDEX, 0);
+	for (i = 0; i < pgraph->ctxprog_size; i++)
+		nv_wr32(dev, NV40_PGRAPH_CTXCTL_UCODE_DATA, pgraph->ctxprog[i]);
 
 	nv_wr32(dev, 0x40008c, 0x00000004); /* HW_CTX_SWITCH_ENABLED */
 	nv_wr32(dev, 0x400320, 4);
@@ -172,8 +235,8 @@ nv50_graph_init_ctxctl(struct drm_device *dev)
 	return 0;
 }
 
-int
-nv50_graph_init(struct drm_device *dev)
+static int
+nv50_graph_init(struct drm_device *dev, int engine)
 {
 	int ret;
 
@@ -187,105 +250,66 @@ nv50_graph_init(struct drm_device *dev)
 	if (ret)
 		return ret;
 
-	ret = nv50_graph_register(dev);
-	if (ret)
-		return ret;
 	nv50_graph_init_intr(dev);
 	return 0;
 }
 
-void
-nv50_graph_takedown(struct drm_device *dev)
+static int
+nv50_graph_fini(struct drm_device *dev, int engine)
 {
 	NV_DEBUG(dev, "\n");
+	nv50_graph_unload_context(dev);
 	nv_wr32(dev, 0x40013c, 0x00000000);
-	nouveau_irq_unregister(dev, 12);
-}
-
-void
-nv50_graph_fifo_access(struct drm_device *dev, bool enabled)
-{
-	const uint32_t mask = 0x00010001;
-
-	if (enabled)
-		nv_wr32(dev, 0x400500, nv_rd32(dev, 0x400500) | mask);
-	else
-		nv_wr32(dev, 0x400500, nv_rd32(dev, 0x400500) & ~mask);
-}
-
-struct nouveau_channel *
-nv50_graph_channel(struct drm_device *dev)
-{
-	struct drm_nouveau_private *dev_priv = dev->dev_private;
-	uint32_t inst;
-	int i;
-
-	/* Be sure we're not in the middle of a context switch or bad things
-	 * will happen, such as unloading the wrong pgraph context.
-	 */
-	if (!nv_wait(dev, 0x400300, 0x00000001, 0x00000000))
-		NV_ERROR(dev, "Ctxprog is still running\n");
-
-	inst = nv_rd32(dev, NV50_PGRAPH_CTXCTL_CUR);
-	if (!(inst & NV50_PGRAPH_CTXCTL_CUR_LOADED))
-		return NULL;
-	inst = (inst & NV50_PGRAPH_CTXCTL_CUR_INSTANCE) << 12;
-
-	for (i = 0; i < dev_priv->engine.fifo.channels; i++) {
-		struct nouveau_channel *chan = dev_priv->channels.ptr[i];
-
-		if (chan && chan->ramin && chan->ramin->vinst == inst)
-			return chan;
-	}
-
-	return NULL;
+	return 0;
 }
 
-int
-nv50_graph_create_context(struct nouveau_channel *chan)
+static int
+nv50_graph_context_new(struct nouveau_channel *chan, int engine)
 {
 	struct drm_device *dev = chan->dev;
 	struct drm_nouveau_private *dev_priv = dev->dev_private;
 	struct nouveau_gpuobj *ramin = chan->ramin;
-	struct nouveau_pgraph_engine *pgraph = &dev_priv->engine.graph;
+	struct nouveau_gpuobj *grctx = NULL;
+	struct nv50_graph_engine *pgraph = nv_engine(dev, engine);
 	struct nouveau_grctx ctx = {};
 	int hdr, ret;
 
 	NV_DEBUG(dev, "ch%d\n", chan->id);
 
-	ret = nouveau_gpuobj_new(dev, chan, pgraph->grctx_size, 0,
+	ret = nouveau_gpuobj_new(dev, NULL, pgraph->grctx_size, 0,
 				 NVOBJ_FLAG_ZERO_ALLOC |
-				 NVOBJ_FLAG_ZERO_FREE, &chan->ramin_grctx);
+				 NVOBJ_FLAG_ZERO_FREE, &grctx);
 	if (ret)
 		return ret;
 
 	hdr = (dev_priv->chipset == 0x50) ? 0x200 : 0x20;
 	nv_wo32(ramin, hdr + 0x00, 0x00190002);
-	nv_wo32(ramin, hdr + 0x04, chan->ramin_grctx->vinst +
-				   pgraph->grctx_size - 1);
-	nv_wo32(ramin, hdr + 0x08, chan->ramin_grctx->vinst);
+	nv_wo32(ramin, hdr + 0x04, grctx->vinst + grctx->size - 1);
+	nv_wo32(ramin, hdr + 0x08, grctx->vinst);
 	nv_wo32(ramin, hdr + 0x0c, 0);
 	nv_wo32(ramin, hdr + 0x10, 0);
 	nv_wo32(ramin, hdr + 0x14, 0x00010000);
 
 	ctx.dev = chan->dev;
 	ctx.mode = NOUVEAU_GRCTX_VALS;
-	ctx.data = chan->ramin_grctx;
+	ctx.data = grctx;
 	nv50_grctx_init(&ctx);
 
-	nv_wo32(chan->ramin_grctx, 0x00000, chan->ramin->vinst >> 12);
+	nv_wo32(grctx, 0x00000, chan->ramin->vinst >> 12);
 
 	dev_priv->engine.instmem.flush(dev);
-	atomic_inc(&chan->vm->pgraph_refs);
+
+	atomic_inc(&chan->vm->engref[NVOBJ_ENGINE_GR]);
+	chan->engctx[NVOBJ_ENGINE_GR] = grctx;
 	return 0;
 }
 
-void
-nv50_graph_destroy_context(struct nouveau_channel *chan)
+static void
+nv50_graph_context_del(struct nouveau_channel *chan, int engine)
 {
+	struct nouveau_gpuobj *grctx = chan->engctx[engine];
 	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 nouveau_fifo_engine *pfifo = &dev_priv->engine.fifo;
 	int i, hdr = (dev_priv->chipset == 0x50) ? 0x200 : 0x20;
 	unsigned long flags;
@@ -297,76 +321,28 @@ nv50_graph_destroy_context(struct nouveau_channel *chan)
 
 	spin_lock_irqsave(&dev_priv->context_switch_lock, flags);
 	pfifo->reassign(dev, false);
-	pgraph->fifo_access(dev, false);
+	nv50_graph_fifo_access(dev, false);
 
-	if (pgraph->channel(dev) == chan)
-		pgraph->unload_context(dev);
+	if (nv50_graph_channel(dev) == chan)
+		nv50_graph_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);
+	nv50_graph_fifo_access(dev, true);
 	pfifo->reassign(dev, true);
 	spin_unlock_irqrestore(&dev_priv->context_switch_lock, flags);
 
-	nouveau_gpuobj_ref(NULL, &chan->ramin_grctx);
+	nouveau_gpuobj_ref(NULL, &grctx);
 
-	atomic_dec(&chan->vm->pgraph_refs);
+	atomic_dec(&chan->vm->engref[engine]);
+	chan->engctx[engine] = NULL;
 }
 
 static int
-nv50_graph_do_load_context(struct drm_device *dev, uint32_t inst)
-{
-	uint32_t fifo = nv_rd32(dev, 0x400500);
-
-	nv_wr32(dev, 0x400500, fifo & ~1);
-	nv_wr32(dev, 0x400784, inst);
-	nv_wr32(dev, 0x400824, nv_rd32(dev, 0x400824) | 0x40);
-	nv_wr32(dev, 0x400320, nv_rd32(dev, 0x400320) | 0x11);
-	nv_wr32(dev, 0x400040, 0xffffffff);
-	(void)nv_rd32(dev, 0x400040);
-	nv_wr32(dev, 0x400040, 0x00000000);
-	nv_wr32(dev, 0x400304, nv_rd32(dev, 0x400304) | 1);
-
-	if (nouveau_wait_for_idle(dev))
-		nv_wr32(dev, 0x40032c, inst | (1<<31));
-	nv_wr32(dev, 0x400500, fifo);
-
-	return 0;
-}
-
-int
-nv50_graph_load_context(struct nouveau_channel *chan)
-{
-	uint32_t inst = chan->ramin->vinst >> 12;
-
-	NV_DEBUG(chan->dev, "ch%d\n", chan->id);
-	return nv50_graph_do_load_context(chan->dev, inst);
-}
-
-int
-nv50_graph_unload_context(struct drm_device *dev)
-{
-	uint32_t inst;
-
-	inst  = nv_rd32(dev, NV50_PGRAPH_CTXCTL_CUR);
-	if (!(inst & NV50_PGRAPH_CTXCTL_CUR_LOADED))
-		return 0;
-	inst &= NV50_PGRAPH_CTXCTL_CUR_INSTANCE;
-
-	nouveau_wait_for_idle(dev);
-	nv_wr32(dev, 0x400784, inst);
-	nv_wr32(dev, 0x400824, nv_rd32(dev, 0x400824) | 0x20);
-	nv_wr32(dev, 0x400304, nv_rd32(dev, 0x400304) | 0x01);
-	nouveau_wait_for_idle(dev);
-
-	nv_wr32(dev, NV50_PGRAPH_CTXCTL_CUR, inst);
-	return 0;
-}
-
-int
-nv50_graph_object_new(struct nouveau_channel *chan, u32 handle, u16 class)
+nv50_graph_object_new(struct nouveau_channel *chan, int engine,
+		      u32 handle, u16 class)
 {
 	struct drm_device *dev = chan->dev;
 	struct drm_nouveau_private *dev_priv = dev->dev_private;
@@ -468,68 +444,15 @@ nv50_graph_nvsw_mthd_page_flip(struct nouveau_channel *chan,
 	return 0;
 }
 
-static int
-nv50_graph_register(struct drm_device *dev)
-{
-	struct drm_nouveau_private *dev_priv = dev->dev_private;
-
-	if (dev_priv->engine.graph.registered)
-		return 0;
-
-	NVOBJ_CLASS(dev, 0x506e, SW); /* nvsw */
-	NVOBJ_MTHD (dev, 0x506e, 0x018c, nv50_graph_nvsw_dma_vblsem);
-	NVOBJ_MTHD (dev, 0x506e, 0x0400, nv50_graph_nvsw_vblsem_offset);
-	NVOBJ_MTHD (dev, 0x506e, 0x0404, nv50_graph_nvsw_vblsem_release_val);
-	NVOBJ_MTHD (dev, 0x506e, 0x0408, nv50_graph_nvsw_vblsem_release);
-	NVOBJ_MTHD (dev, 0x506e, 0x0500, nv50_graph_nvsw_mthd_page_flip);
-
-	NVOBJ_CLASS(dev, 0x0030, GR); /* null */
-	NVOBJ_CLASS(dev, 0x5039, GR); /* m2mf */
-	NVOBJ_CLASS(dev, 0x502d, GR); /* 2d */
-
-	/* tesla */
-	if (dev_priv->chipset == 0x50)
-		NVOBJ_CLASS(dev, 0x5097, GR); /* tesla (nv50) */
-	else
-	if (dev_priv->chipset < 0xa0)
-		NVOBJ_CLASS(dev, 0x8297, GR); /* tesla (nv8x/nv9x) */
-	else {
-		switch (dev_priv->chipset) {
-		case 0xa0:
-		case 0xaa:
-		case 0xac:
-			NVOBJ_CLASS(dev, 0x8397, GR);
-			break;
-		case 0xa3:
-		case 0xa5:
-		case 0xa8:
-			NVOBJ_CLASS(dev, 0x8597, GR);
-			break;
-		case 0xaf:
-			NVOBJ_CLASS(dev, 0x8697, GR);
-			break;
-		}
-	}
-
-	/* compute */
-	NVOBJ_CLASS(dev, 0x50c0, GR);
-	if (dev_priv->chipset  > 0xa0 &&
-	    dev_priv->chipset != 0xaa &&
-	    dev_priv->chipset != 0xac)
-		NVOBJ_CLASS(dev, 0x85c0, GR);
-
-	dev_priv->engine.graph.registered = true;
-	return 0;
-}
 
-void
-nv50_graph_tlb_flush(struct drm_device *dev)
+static void
+nv50_graph_tlb_flush(struct drm_device *dev, int engine)
 {
 	nv50_vm_flush_engine(dev, 0);
 }
 
-void
-nv84_graph_tlb_flush(struct drm_device *dev)
+static void
+nv84_graph_tlb_flush(struct drm_device *dev, int engine)
 {
 	struct drm_nouveau_private *dev_priv = dev->dev_private;
 	struct nouveau_timer_engine *ptimer = &dev_priv->engine.timer;
@@ -1098,3 +1021,102 @@ nv50_graph_isr(struct drm_device *dev)
 	if (nv_rd32(dev, 0x400824) & (1 << 31))
 		nv_wr32(dev, 0x400824, nv_rd32(dev, 0x400824) & ~(1 << 31));
 }
+
+static void
+nv50_graph_destroy(struct drm_device *dev, int engine)
+{
+	struct nv50_graph_engine *pgraph = nv_engine(dev, engine);
+
+	NVOBJ_ENGINE_DEL(dev, GR);
+
+	nouveau_irq_unregister(dev, 12);
+	kfree(pgraph);
+}
+
+int
+nv50_graph_create(struct drm_device *dev)
+{
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nv50_graph_engine *pgraph;
+	struct nouveau_grctx ctx = {};
+	int ret;
+
+	pgraph = kzalloc(sizeof(*pgraph),GFP_KERNEL);
+	if (!pgraph)
+		return -ENOMEM;
+
+	ctx.dev = dev;
+	ctx.mode = NOUVEAU_GRCTX_PROG;
+	ctx.data = pgraph->ctxprog;
+	ctx.ctxprog_max = ARRAY_SIZE(pgraph->ctxprog);
+
+	ret = nv50_grctx_init(&ctx);
+	if (ret) {
+		NV_ERROR(dev, "PGRAPH: ctxprog build failed\n");
+		dev_priv->engine.graph.accel_blocked = true;
+		kfree(pgraph);
+		return 0;
+	}
+
+	pgraph->grctx_size = ctx.ctxvals_pos * 4;
+	pgraph->ctxprog_size = ctx.ctxprog_len;
+
+	pgraph->base.destroy = nv50_graph_destroy;
+	pgraph->base.init = nv50_graph_init;
+	pgraph->base.fini = nv50_graph_fini;
+	pgraph->base.context_new = nv50_graph_context_new;
+	pgraph->base.context_del = nv50_graph_context_del;
+	pgraph->base.object_new = nv50_graph_object_new;
+	if (dev_priv->chipset == 0x50 || dev_priv->chipset == 0xac)
+		pgraph->base.tlb_flush = nv50_graph_tlb_flush;
+	else
+		pgraph->base.tlb_flush = nv84_graph_tlb_flush;
+
+	nouveau_irq_register(dev, 12, nv50_graph_isr);
+
+	/* NVSW really doesn't live here... */
+	NVOBJ_CLASS(dev, 0x506e, SW); /* nvsw */
+	NVOBJ_MTHD (dev, 0x506e, 0x018c, nv50_graph_nvsw_dma_vblsem);
+	NVOBJ_MTHD (dev, 0x506e, 0x0400, nv50_graph_nvsw_vblsem_offset);
+	NVOBJ_MTHD (dev, 0x506e, 0x0404, nv50_graph_nvsw_vblsem_release_val);
+	NVOBJ_MTHD (dev, 0x506e, 0x0408, nv50_graph_nvsw_vblsem_release);
+	NVOBJ_MTHD (dev, 0x506e, 0x0500, nv50_graph_nvsw_mthd_page_flip);
+
+	NVOBJ_ENGINE_ADD(dev, GR, &pgraph->base);
+	NVOBJ_CLASS(dev, 0x0030, GR); /* null */
+	NVOBJ_CLASS(dev, 0x5039, GR); /* m2mf */
+	NVOBJ_CLASS(dev, 0x502d, GR); /* 2d */
+
+	/* tesla */
+	if (dev_priv->chipset == 0x50)
+		NVOBJ_CLASS(dev, 0x5097, GR); /* tesla (nv50) */
+	else
+	if (dev_priv->chipset < 0xa0)
+		NVOBJ_CLASS(dev, 0x8297, GR); /* tesla (nv8x/nv9x) */
+	else {
+		switch (dev_priv->chipset) {
+		case 0xa0:
+		case 0xaa:
+		case 0xac:
+			NVOBJ_CLASS(dev, 0x8397, GR);
+			break;
+		case 0xa3:
+		case 0xa5:
+		case 0xa8:
+			NVOBJ_CLASS(dev, 0x8597, GR);
+			break;
+		case 0xaf:
+			NVOBJ_CLASS(dev, 0x8697, GR);
+			break;
+		}
+	}
+
+	/* compute */
+	NVOBJ_CLASS(dev, 0x50c0, GR);
+	if (dev_priv->chipset  > 0xa0 &&
+	    dev_priv->chipset != 0xaa &&
+	    dev_priv->chipset != 0xac)
+		NVOBJ_CLASS(dev, 0x85c0, GR);
+
+	return 0;
+}
diff --git a/drivers/gpu/drm/nouveau/nv50_vm.c b/drivers/gpu/drm/nouveau/nv50_vm.c
index 1f58b0d..1a0dd49 100644
--- a/drivers/gpu/drm/nouveau/nv50_vm.c
+++ b/drivers/gpu/drm/nouveau/nv50_vm.c
@@ -151,7 +151,6 @@ nv50_vm_flush(struct nouveau_vm *vm)
 	struct drm_nouveau_private *dev_priv = vm->dev->dev_private;
 	struct nouveau_instmem_engine *pinstmem = &dev_priv->engine.instmem;
 	struct nouveau_fifo_engine *pfifo = &dev_priv->engine.fifo;
-	struct nouveau_pgraph_engine *pgraph = &dev_priv->engine.graph;
 	int i;
 
 	pinstmem->flush(vm->dev);
@@ -163,9 +162,6 @@ nv50_vm_flush(struct nouveau_vm *vm)
 	}
 
 	pfifo->tlb_flush(vm->dev);
-
-	if (atomic_read(&vm->pgraph_refs))
-		pgraph->tlb_flush(vm->dev);
 	for (i = 0; i < NVOBJ_ENGINE_NR; i++) {
 		if (atomic_read(&vm->engref[i]))
 			dev_priv->eng[i]->tlb_flush(vm->dev, i);