drm/nvc0-/gr: generate grctx template at init time, not first context ctor

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

5 files changed, 460 insertions, 604 deletions

diff --git a/drivers/gpu/drm/nouveau/core/engine/graph/ctxnvc0.c b/drivers/gpu/drm/nouveau/core/engine/graph/ctxnvc0.c
index 57353015329f..b19a406e55d9 100644
--- a/drivers/gpu/drm/nouveau/core/engine/graph/ctxnvc0.c
+++ b/drivers/gpu/drm/nouveau/core/engine/graph/ctxnvc0.c
@@ -35,6 +35,156 @@ nv_icmd(struct drm_device *priv, u32 icmd, u32 data)
 	while (nv_rd32(priv, 0x400700) & 2) {}
 }
 
+int
+nvc0_grctx_init(struct drm_device *priv, struct nvc0_graph_priv *oprv,
+		struct nvc0_grctx *info)
+{
+	struct nouveau_gpuobj *chan;
+	u32 size = (0x80000 + oprv->size + 4095) & ~4095;
+	int ret, i;
+
+	/* allocate memory to for a "channel", which we'll use to generate
+	 * the default context values
+	 */
+	ret = nouveau_gpuobj_new(priv, NULL, size, 0x1000,
+				 NVOBJ_FLAG_ZERO_ALLOC, &info->chan);
+	chan = info->chan;
+	if (ret) {
+		NV_ERROR(priv, "failed to allocate channel memory, %d\n", ret);
+		return ret;
+	}
+
+	/* PGD pointer */
+	nv_wo32(chan, 0x0200, lower_32_bits(chan->addr + 0x1000));
+	nv_wo32(chan, 0x0204, upper_32_bits(chan->addr + 0x1000));
+	nv_wo32(chan, 0x0208, 0xffffffff);
+	nv_wo32(chan, 0x020c, 0x000000ff);
+
+	/* PGT[0] pointer */
+	nv_wo32(chan, 0x1000, 0x00000000);
+	nv_wo32(chan, 0x1004, 0x00000001 | (chan->addr + 0x2000) >> 8);
+
+	/* identity-map the whole "channel" into its own vm */
+	for (i = 0; i < size / 4096; i++) {
+		u64 addr = ((chan->addr + (i * 4096)) >> 8) | 1;
+		nv_wo32(chan, 0x2000 + (i * 8), lower_32_bits(addr));
+		nv_wo32(chan, 0x2004 + (i * 8), upper_32_bits(addr));
+	}
+
+	/* context pointer (virt) */
+	nv_wo32(chan, 0x0210, 0x00080004);
+	nv_wo32(chan, 0x0214, 0x00000000);
+
+	nvimem_flush(priv);
+
+	nv_wr32(priv, 0x100cb8, (chan->addr + 0x1000) >> 8);
+	nv_wr32(priv, 0x100cbc, 0x80000001);
+	nv_wait(priv, 0x100c80, 0x00008000, 0x00008000);
+
+	/* setup default state for mmio list construction */
+	info->dev  = priv;
+	info->data = oprv->mmio_data;
+	info->mmio = oprv->mmio_list;
+	info->addr = 0x2000 + (i * 8);
+	info->priv = oprv;
+	info->buffer_nr = 0;
+
+	if (oprv->firmware) {
+		nv_wr32(priv, 0x409840, 0x00000030);
+		nv_wr32(priv, 0x409500, 0x80000000 | chan->addr >> 12);
+		nv_wr32(priv, 0x409504, 0x00000003);
+		if (!nv_wait(priv, 0x409800, 0x00000010, 0x00000010))
+			NV_ERROR(priv, "load_ctx timeout\n");
+
+		nv_wo32(chan, 0x8001c, 1);
+		nv_wo32(chan, 0x80020, 0);
+		nv_wo32(chan, 0x80028, 0);
+		nv_wo32(chan, 0x8002c, 0);
+		nvimem_flush(priv);
+		return 0;
+	}
+
+	/* HUB_FUC(SET_CHAN) */
+	nv_wr32(priv, 0x409840, 0x80000000);
+	nv_wr32(priv, 0x409500, 0x80000000 | chan->addr >> 12);
+	nv_wr32(priv, 0x409504, 0x00000001);
+	if (!nv_wait(priv, 0x409800, 0x80000000, 0x80000000)) {
+		NV_ERROR(priv, "HUB_SET_CHAN timeout\n");
+		nvc0_graph_ctxctl_debug(priv);
+		nouveau_gpuobj_ref(NULL, &info->chan);
+		return -EBUSY;
+	}
+
+	return 0;
+}
+
+void
+nvc0_grctx_data(struct nvc0_grctx *info, u32 size, u32 align, u32 access)
+{
+	info->buffer[info->buffer_nr]  = info->addr;
+	info->buffer[info->buffer_nr] +=  (align - 1);
+	info->buffer[info->buffer_nr] &= ~(align - 1);
+	info->addr = info->buffer[info->buffer_nr++] + size;
+
+	info->data->size = size;
+	info->data->align = align;
+	info->data->access = access;
+	info->data++;
+}
+
+void
+nvc0_grctx_mmio(struct nvc0_grctx *info, u32 addr, u32 data, u32 shift, u32 buf)
+{
+	info->mmio->addr = addr;
+	info->mmio->data = data;
+	info->mmio->shift = shift;
+	info->mmio->buffer = buf;
+	info->mmio++;
+
+	if (shift)
+		data |= info->buffer[buf] >> shift;
+	nv_wr32(info->dev, addr, data);
+}
+
+int
+nvc0_grctx_fini(struct nvc0_grctx *info)
+{
+	struct nvc0_graph_priv *priv = info->priv;
+	int i;
+
+	if (priv->firmware) {
+		nv_wr32(info->dev, 0x409840, 0x00000003);
+		nv_wr32(info->dev, 0x409500, 0x80000000 | info->chan->addr >> 12);
+		nv_wr32(info->dev, 0x409504, 0x00000009);
+		if (!nv_wait(info->dev, 0x409800, 0x00000001, 0x00000000)) {
+			NV_ERROR(info->dev, "unload_ctx timeout\n");
+			return -EBUSY;
+		}
+
+		goto save;
+	}
+
+	/* HUB_FUC(CTX_SAVE) */
+	nv_wr32(info->dev, 0x409840, 0x80000000);
+	nv_wr32(info->dev, 0x409500, 0x80000000 | info->chan->addr >> 12);
+	nv_wr32(info->dev, 0x409504, 0x00000002);
+	if (!nv_wait(info->dev, 0x409800, 0x80000000, 0x80000000)) {
+		NV_ERROR(info->dev, "HUB_CTX_SAVE timeout\n");
+		nvc0_graph_ctxctl_debug(info->dev);
+		return -EBUSY;
+	}
+
+save:
+	priv->data = kmalloc(priv->size, GFP_KERNEL);
+	if (priv->data) {
+		for (i = 0; i < priv->size; i += 4)
+			priv->data[i / 4] = nv_ro32(info->chan, 0x80000 + i);
+	}
+
+	nouveau_gpuobj_ref(NULL, &info->chan);
+	return priv->data ? 0 : -ENOMEM;
+}
+
 static void
 nvc0_grctx_generate_9097(struct drm_device *priv)
 {
@@ -1779,16 +1929,19 @@ nvc0_grctx_generate_tp(struct drm_device *priv)
 }
 
 int
-nvc0_grctx_generate(struct nouveau_channel *chan)
+nvc0_grctx_generate(struct drm_device *priv)
 {
-	struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
-	struct nvc0_graph_priv *oprv = nv_engine(chan->dev, NVOBJ_ENGINE_GR);
-	struct nvc0_graph_chan *grch = chan->engctx[NVOBJ_ENGINE_GR];
-	struct drm_device *priv = chan->dev;
-	int i, gpc, tp, id;
+	struct drm_nouveau_private *dev_priv = priv->dev_private;
+	struct nvc0_graph_priv *oprv = nv_engine(priv, NVOBJ_ENGINE_GR);
+	struct nvc0_grctx info;
+	int ret, i, gpc, tpc, id;
 	u32 fermi = nvc0_graph_class(priv);
 	u32 r000260, tmp;
 
+	ret = nvc0_grctx_init(priv, oprv, &info);
+	if (ret)
+		return ret;
+
 	r000260 = nv_rd32(priv, 0x000260);
 	nv_wr32(priv, 0x000260, r000260 & ~1);
 	nv_wr32(priv, 0x400208, 0x00000000);
@@ -1808,19 +1961,55 @@ nvc0_grctx_generate(struct nouveau_channel *chan)
 
 	nv_wr32(priv, 0x404154, 0x00000000);
 
-	/* fuc "mmio list" writes */
-	for (i = 0; i < grch->mmio_nr * 8; i += 8) {
-		u32 reg = nv_ro32(grch->mmio, i + 0);
-		nv_wr32(priv, reg, nv_ro32(grch->mmio, i + 4));
+	/* generate per-context mmio list data */
+	mmio_data(0x002000, 0x0100, NV_MEM_ACCESS_RW | NV_MEM_ACCESS_SYS);
+	mmio_data(0x008000, 0x0100, NV_MEM_ACCESS_RW | NV_MEM_ACCESS_SYS);
+	mmio_data(0x060000, 0x1000, NV_MEM_ACCESS_RW);
+	mmio_list(0x408004, 0x00000000,  8, 0);
+	mmio_list(0x408008, 0x80000018,  0, 0);
+	mmio_list(0x40800c, 0x00000000,  8, 1);
+	mmio_list(0x408010, 0x80000000,  0, 0);
+	mmio_list(0x418810, 0x80000000, 12, 2);
+	mmio_list(0x419848, 0x10000000, 12, 2);
+	mmio_list(0x419004, 0x00000000,  8, 1);
+	mmio_list(0x419008, 0x00000000,  0, 0);
+	mmio_list(0x418808, 0x00000000,  8, 0);
+	mmio_list(0x41880c, 0x80000018,  0, 0);
+	if (dev_priv->chipset != 0xc1) {
+		tmp = 0x02180000;
+		mmio_list(0x405830, tmp, 0, 0);
+		for (gpc = 0; gpc < oprv->gpc_nr; gpc++) {
+			for (tpc = 0; tpc < oprv->tpc_nr[gpc]; tpc++) {
+				u32 reg = TPC_UNIT(gpc, tpc, 0x0520);
+				mmio_list(reg, tmp, 0, 0);
+				tmp += 0x0324;
+			}
+		}
+	} else {
+		tmp = 0x02180000;
+		mmio_list(0x405830, 0x00000218 | tmp, 0, 0);
+		mmio_list(0x4064c4, 0x0086ffff, 0, 0);
+		for (gpc = 0; gpc < oprv->gpc_nr; gpc++) {
+			for (tpc = 0; tpc < oprv->tpc_nr[gpc]; tpc++) {
+				u32 reg = TPC_UNIT(gpc, tpc, 0x0520);
+				mmio_list(reg, 0x10000000 | tmp, 0, 0);
+				tmp += 0x0324;
+			}
+			for (tpc = 0; tpc < oprv->tpc_nr[gpc]; tpc++) {
+				u32 reg = TPC_UNIT(gpc, tpc, 0x0544);
+				mmio_list(reg, tmp, 0, 0);
+				tmp += 0x0324;
+			}
+		}
 	}
 
-	for (tp = 0, id = 0; tp < 4; tp++) {
+	for (tpc = 0, id = 0; tpc < 4; tpc++) {
 		for (gpc = 0; gpc < oprv->gpc_nr; gpc++) {
-			if (tp < oprv->tpc_nr[gpc]) {
-				nv_wr32(priv, TPC_UNIT(gpc, tp, 0x698), id);
-				nv_wr32(priv, TPC_UNIT(gpc, tp, 0x4e8), id);
-				nv_wr32(priv, GPC_UNIT(gpc, 0x0c10 + tp * 4), id);
-				nv_wr32(priv, TPC_UNIT(gpc, tp, 0x088), id);
+			if (tpc < oprv->tpc_nr[gpc]) {
+				nv_wr32(priv, TPC_UNIT(gpc, tpc, 0x698), id);
+				nv_wr32(priv, TPC_UNIT(gpc, tpc, 0x4e8), id);
+				nv_wr32(priv, GPC_UNIT(gpc, 0x0c10 + tpc * 4), id);
+				nv_wr32(priv, TPC_UNIT(gpc, tpc, 0x088), id);
 				id++;
 			}
 
@@ -1843,18 +2032,18 @@ nvc0_grctx_generate(struct nouveau_channel *chan)
 	nv_wr32(priv, 0x40587c, 0x00000000);
 
 	if (1) {
-		u8 tpnr[GPC_MAX], data[TPC_MAX];
+		u8 tpcnr[GPC_MAX], data[TPC_MAX];
 
-		memcpy(tpnr, oprv->tpc_nr, sizeof(oprv->tpc_nr));
+		memcpy(tpcnr, oprv->tpc_nr, sizeof(oprv->tpc_nr));
 		memset(data, 0x1f, sizeof(data));
 
 		gpc = -1;
-		for (tp = 0; tp < oprv->tpc_total; tp++) {
+		for (tpc = 0; tpc < oprv->tpc_total; tpc++) {
 			do {
 				gpc = (gpc + 1) % oprv->gpc_nr;
-			} while (!tpnr[gpc]);
-			tpnr[gpc]--;
-			data[tp] = gpc;
+			} while (!tpcnr[gpc]);
+			tpcnr[gpc]--;
+			data[tpc] = gpc;
 		}
 
 		for (i = 0; i < 4; i++)
@@ -1863,24 +2052,24 @@ nvc0_grctx_generate(struct nouveau_channel *chan)
 
 	if (1) {
 		u32 data[6] = {}, data2[2] = {};
-		u8 tpnr[GPC_MAX];
+		u8 tpcnr[GPC_MAX];
 		u8 shift, ntpcv;
 
 		/* calculate first set of magics */
-		memcpy(tpnr, oprv->tpc_nr, sizeof(oprv->tpc_nr));
+		memcpy(tpcnr, oprv->tpc_nr, sizeof(oprv->tpc_nr));
 
 		gpc = -1;
-		for (tp = 0; tp < oprv->tpc_total; tp++) {
+		for (tpc = 0; tpc < oprv->tpc_total; tpc++) {
 			do {
 				gpc = (gpc + 1) % oprv->gpc_nr;
-			} while (!tpnr[gpc]);
-			tpnr[gpc]--;
+			} while (!tpcnr[gpc]);
+			tpcnr[gpc]--;
 
-			data[tp / 6] |= gpc << ((tp % 6) * 5);
+			data[tpc / 6] |= gpc << ((tpc % 6) * 5);
 		}
 
-		for (; tp < 32; tp++)
-			data[tp / 6] |= 7 << ((tp % 6) * 5);
+		for (; tpc < 32; tpc++)
+			data[tpc / 6] |= 7 << ((tpc % 6) * 5);
 
 		/* and the second... */
 		shift = 0;
@@ -1918,12 +2107,12 @@ nvc0_grctx_generate(struct nouveau_channel *chan)
 	}
 
 	if (1) {
-		u32 tp_mask = 0, tp_set = 0;
-		u8  tpnr[GPC_MAX], a, b;
+		u32 tpc_mask = 0, tpc_set = 0;
+		u8  tpcnr[GPC_MAX], a, b;
 
-		memcpy(tpnr, oprv->tpc_nr, sizeof(oprv->tpc_nr));
+		memcpy(tpcnr, oprv->tpc_nr, sizeof(oprv->tpc_nr));
 		for (gpc = 0; gpc < oprv->gpc_nr; gpc++)
-			tp_mask |= ((1 << oprv->tpc_nr[gpc]) - 1) << (gpc * 8);
+			tpc_mask |= ((1 << oprv->tpc_nr[gpc]) - 1) << (gpc * 8);
 
 		for (i = 0, gpc = -1, b = -1; i < 32; i++) {
 			a = (i * (oprv->tpc_total - 1)) / 32;
@@ -1931,14 +2120,14 @@ nvc0_grctx_generate(struct nouveau_channel *chan)
 				b = a;
 				do {
 					gpc = (gpc + 1) % oprv->gpc_nr;
-				} while (!tpnr[gpc]);
-				tp = oprv->tpc_nr[gpc] - tpnr[gpc]--;
+				} while (!tpcnr[gpc]);
+				tpc = oprv->tpc_nr[gpc] - tpcnr[gpc]--;
 
-				tp_set |= 1 << ((gpc * 8) + tp);
+				tpc_set |= 1 << ((gpc * 8) + tpc);
 			}
 
-			nv_wr32(priv, 0x406800 + (i * 0x20), tp_set);
-			nv_wr32(priv, 0x406c00 + (i * 0x20), tp_set ^ tp_mask);
+			nv_wr32(priv, 0x406800 + (i * 0x20), tpc_set);
+			nv_wr32(priv, 0x406c00 + (i * 0x20), tpc_set ^ tpc_mask);
 		}
 	}
 
@@ -2867,5 +3056,5 @@ nvc0_grctx_generate(struct nouveau_channel *chan)
 	nvc0_grctx_generate_90c0(priv);
 
 	nv_wr32(priv, 0x000260, r000260);
-	return 0;
+	return nvc0_grctx_fini(&info);
 }
diff --git a/drivers/gpu/drm/nouveau/core/engine/graph/ctxnve0.c b/drivers/gpu/drm/nouveau/core/engine/graph/ctxnve0.c
index 47bea8a7b2bc..e5503170d68c 100644
--- a/drivers/gpu/drm/nouveau/core/engine/graph/ctxnve0.c
+++ b/drivers/gpu/drm/nouveau/core/engine/graph/ctxnve0.c
@@ -2604,16 +2604,20 @@ nve0_graph_generate_tpcunk(struct drm_device *priv)
 }
 
 int
-nve0_grctx_generate(struct nouveau_channel *chan)
+nve0_grctx_generate(struct drm_device *priv)
 {
-	struct nvc0_graph_priv *oprv = nv_engine(chan->dev, NVOBJ_ENGINE_GR);
-	struct nvc0_graph_chan *grch = chan->engctx[NVOBJ_ENGINE_GR];
-	struct drm_device *priv = chan->dev;
+	struct nvc0_graph_priv *oprv = nv_engine(priv, NVOBJ_ENGINE_GR);
+	struct nvc0_grctx info;
+	int ret, i, gpc, tpc, id;
 	u32 data[6] = {}, data2[2] = {}, tmp;
 	u32 tpc_set = 0, tpc_mask = 0;
+	u32 magic[GPC_MAX][2], offset;
 	u8 tpcnr[GPC_MAX], a, b;
 	u8 shift, ntpcv;
-	int i, gpc, tpc, id;
+
+	ret = nvc0_grctx_init(priv, oprv, &info);
+	if (ret)
+		return ret;
 
 	nv_mask(priv, 0x000260, 0x00000001, 0x00000000);
 	nv_wr32(priv, 0x400204, 0x00000000);
@@ -2636,11 +2640,37 @@ nve0_grctx_generate(struct nouveau_channel *chan)
 
 	nv_wr32(priv, 0x404154, 0x0);
 
-	for (i = 0; i < grch->mmio_nr * 8; i += 8) {
-		u32 reg = nv_ro32(grch->mmio, i + 0);
-		u32 val = nv_ro32(grch->mmio, i + 4);
-		nv_wr32(priv, reg, val);
+	mmio_data(0x003000, 0x0100, NV_MEM_ACCESS_RW | NV_MEM_ACCESS_SYS);
+	mmio_data(0x008000, 0x0100, NV_MEM_ACCESS_RW | NV_MEM_ACCESS_SYS);
+	mmio_data(0x060000, 0x1000, NV_MEM_ACCESS_RW);
+	mmio_list(0x40800c, 0x00000000,  8, 1);
+	mmio_list(0x408010, 0x80000000,  0, 0);
+	mmio_list(0x419004, 0x00000000,  8, 1);
+	mmio_list(0x419008, 0x00000000,  0, 0);
+	mmio_list(0x4064cc, 0x80000000,  0, 0);
+	mmio_list(0x408004, 0x00000000,  8, 0);
+	mmio_list(0x408008, 0x80000030,  0, 0);
+	mmio_list(0x418808, 0x00000000,  8, 0);
+	mmio_list(0x41880c, 0x80000030,  0, 0);
+	mmio_list(0x4064c8, 0x01800600,  0, 0);
+	mmio_list(0x418810, 0x80000000, 12, 2);
+	mmio_list(0x419848, 0x10000000, 12, 2);
+	mmio_list(0x405830, 0x02180648,  0, 0);
+	mmio_list(0x4064c4, 0x0192ffff,  0, 0);
+	for (gpc = 0, offset = 0; gpc < oprv->gpc_nr; gpc++) {
+		u16 magic0 = 0x0218 * oprv->tpc_nr[gpc];
+		u16 magic1 = 0x0648 * oprv->tpc_nr[gpc];
+		magic[gpc][0]  = 0x10000000 | (magic0 << 16) | offset;
+		magic[gpc][1]  = 0x00000000 | (magic1 << 16);
+		offset += 0x0324 * oprv->tpc_nr[gpc];
+	}
+	for (gpc = 0; gpc < oprv->gpc_nr; gpc++) {
+		mmio_list(GPC_UNIT(gpc, 0x30c0), magic[gpc][0], 0, 0);
+		mmio_list(GPC_UNIT(gpc, 0x30e4), magic[gpc][1] | offset, 0, 0);
+		offset += 0x07ff * oprv->tpc_nr[gpc];
 	}
+	mmio_list(0x17e91c, 0x06060609, 0, 0);
+	mmio_list(0x17e920, 0x00090a05, 0, 0);
 
 	nv_wr32(priv, 0x418c6c, 0x1);
 	nv_wr32(priv, 0x41980c, 0x10);
@@ -2758,5 +2788,5 @@ nve0_grctx_generate(struct nouveau_channel *chan)
 	nv_mask(priv, 0x000260, 0x00000001, 0x00000001);
 	nv_wr32(priv, 0x418800, 0x7026860a); //XXX
 	nv_wr32(priv, 0x41be10, 0x00bb8bc7); //XXX
-	return 0;
+	return nvc0_grctx_fini(&info);
 }
diff --git a/drivers/gpu/drm/nouveau/core/engine/graph/nvc0.c b/drivers/gpu/drm/nouveau/core/engine/graph/nvc0.c
index 158d57619e5b..f994d2f7e8d5 100644
--- a/drivers/gpu/drm/nouveau/core/engine/graph/nvc0.c
+++ b/drivers/gpu/drm/nouveau/core/engine/graph/nvc0.c
@@ -48,7 +48,7 @@ nvc0_graph_ctxctl_debug_unit(struct drm_device *dev, u32 base)
 		nv_rd32(dev, base + 0x818), nv_rd32(dev, base + 0x81c));
 }
 
-static void
+void
 nvc0_graph_ctxctl_debug(struct drm_device *dev)
 {
 	u32 gpcnr = nv_rd32(dev, 0x409604) & 0xffff;
@@ -59,259 +59,92 @@ nvc0_graph_ctxctl_debug(struct drm_device *dev)
 		nvc0_graph_ctxctl_debug_unit(dev, 0x502000 + (gpc * 0x8000));
 }
 
-static int
-nvc0_graph_load_context(struct nouveau_channel *chan)
-{
-	struct drm_device *dev = chan->dev;
-
-	nv_wr32(dev, 0x409840, 0x00000030);
-	nv_wr32(dev, 0x409500, 0x80000000 | chan->ramin->addr >> 12);
-	nv_wr32(dev, 0x409504, 0x00000003);
-	if (!nv_wait(dev, 0x409800, 0x00000010, 0x00000010))
-		NV_ERROR(dev, "PGRAPH: load_ctx timeout\n");
-
-	return 0;
-}
-
-static int
-nvc0_graph_unload_context_to(struct drm_device *dev, u64 chan)
-{
-	nv_wr32(dev, 0x409840, 0x00000003);
-	nv_wr32(dev, 0x409500, 0x80000000 | chan >> 12);
-	nv_wr32(dev, 0x409504, 0x00000009);
-	if (!nv_wait(dev, 0x409800, 0x00000001, 0x00000000)) {
-		NV_ERROR(dev, "PGRAPH: unload_ctx timeout\n");
-		return -EBUSY;
-	}
-
-	return 0;
-}
-
-static int
-nvc0_graph_construct_context(struct nouveau_channel *chan)
+int
+nvc0_graph_context_new(struct nouveau_channel *chan, int engine)
 {
-	struct nvc0_graph_priv *priv = nv_engine(chan->dev, NVOBJ_ENGINE_GR);
-	struct nvc0_graph_chan *grch = chan->engctx[NVOBJ_ENGINE_GR];
 	struct drm_device *dev = chan->dev;
+	struct nvc0_graph_priv *priv = nv_engine(dev, engine);
+	struct nvc0_graph_data *data = priv->mmio_data;
+	struct nvc0_graph_mmio *mmio = priv->mmio_list;
+	struct nvc0_graph_chan *grch;
+	struct nouveau_gpuobj *grctx;
 	int ret, i;
-	u32 *ctx;
 
-	ctx = kmalloc(priv->grctx_size, GFP_KERNEL);
-	if (!ctx)
+	grch = kzalloc(sizeof(*grch), GFP_KERNEL);
+	if (!grch)
 		return -ENOMEM;
+	chan->engctx[NVOBJ_ENGINE_GR] = grch;
 
-	if (!nouveau_ctxfw) {
-		nv_wr32(dev, 0x409840, 0x80000000);
-		nv_wr32(dev, 0x409500, 0x80000000 | chan->ramin->addr >> 12);
-		nv_wr32(dev, 0x409504, 0x00000001);
-		if (!nv_wait(dev, 0x409800, 0x80000000, 0x80000000)) {
-			NV_ERROR(dev, "PGRAPH: HUB_SET_CHAN timeout\n");
-			nvc0_graph_ctxctl_debug(dev);
-			ret = -EBUSY;
-			goto err;
-		}
-	} else {
-		nvc0_graph_load_context(chan);
-
-		nv_wo32(grch->grctx, 0x1c, 1);
-		nv_wo32(grch->grctx, 0x20, 0);
-		nv_wo32(grch->grctx, 0x28, 0);
-		nv_wo32(grch->grctx, 0x2c, 0);
-		nvimem_flush(dev);
-	}
-
-	ret = nvc0_grctx_generate(chan);
+	ret = nouveau_gpuobj_new(dev, NULL, priv->size, 256, 0, &grch->grctx);
 	if (ret)
-		goto err;
-
-	if (!nouveau_ctxfw) {
-		nv_wr32(dev, 0x409840, 0x80000000);
-		nv_wr32(dev, 0x409500, 0x80000000 | chan->ramin->addr >> 12);
-		nv_wr32(dev, 0x409504, 0x00000002);
-		if (!nv_wait(dev, 0x409800, 0x80000000, 0x80000000)) {
-			NV_ERROR(dev, "PGRAPH: HUB_CTX_SAVE timeout\n");
-			nvc0_graph_ctxctl_debug(dev);
-			ret = -EBUSY;
-			goto err;
-		}
-	} else {
-		ret = nvc0_graph_unload_context_to(dev, chan->ramin->addr);
-		if (ret)
-			goto err;
-	}
-
-	for (i = 0; i < priv->grctx_size; i += 4)
-		ctx[i / 4] = nv_ro32(grch->grctx, i);
-
-	priv->grctx_vals = ctx;
-	return 0;
-
-err:
-	kfree(ctx);
-	return ret;
-}
-
-static int
-nvc0_graph_create_context_mmio_list(struct nouveau_channel *chan)
-{
-	struct nvc0_graph_priv *priv = nv_engine(chan->dev, NVOBJ_ENGINE_GR);
-	struct nvc0_graph_chan *grch = chan->engctx[NVOBJ_ENGINE_GR];
-	struct drm_device *dev = chan->dev;
-	struct drm_nouveau_private *dev_priv = dev->dev_private;
-	int i = 0, gpc, tp, ret;
+		goto error;
 
-	ret = nouveau_gpuobj_new(dev, NULL, 0x2000, 256, 0, &grch->unk408004);
+	ret = nouveau_gpuobj_map_vm(grch->grctx, chan->vm, NV_MEM_ACCESS_RW |
+				    NV_MEM_ACCESS_SYS, &grch->grctx_vma);
 	if (ret)
 		return ret;
 
-	ret = nouveau_gpuobj_map_vm(grch->unk408004, chan->vm,
-				    NV_MEM_ACCESS_RW | NV_MEM_ACCESS_SYS,
-				    &grch->unk408004_vma);
-	if (ret)
-		return ret;
+	grctx = grch->grctx;
 
-	ret = nouveau_gpuobj_new(dev, NULL, 0x8000, 256, 0, &grch->unk40800c);
+	/* allocate memory for a "mmio list" buffer that's used by the HUB
+	 * fuc to modify some per-context register settings on first load
+	 * of the context.
+	 */
+	ret = nouveau_gpuobj_new(dev, NULL, 0x1000, 0x100, 0, &grch->mmio);
 	if (ret)
 		return ret;
 
-	ret = nouveau_gpuobj_map_vm(grch->unk40800c, chan->vm,
+	ret = nouveau_gpuobj_map_vm(grch->mmio, chan->vm,
 				    NV_MEM_ACCESS_RW | NV_MEM_ACCESS_SYS,
-				    &grch->unk40800c_vma);
+				    &grch->mmio_vma);
 	if (ret)
 		return ret;
 
-	ret = nouveau_gpuobj_new(dev, NULL, 384 * 1024, 4096, 0,
-				 &grch->unk418810);
-	if (ret)
-		return ret;
-
-	ret = nouveau_gpuobj_map_vm(grch->unk418810, chan->vm,
-				    NV_MEM_ACCESS_RW, &grch->unk418810_vma);
-	if (ret)
-		return ret;
-
-	ret = nouveau_gpuobj_new(dev, NULL, 0x1000, 0, 0, &grch->mmio);
-	if (ret)
-		return ret;
+	/* allocate buffers referenced by mmio list */
+	for (i = 0; data->size && i < ARRAY_SIZE(priv->mmio_data); i++) {
+		ret = nouveau_gpuobj_new(dev, NULL, data->size, data->align,
+					 0, &grch->data[i].mem);
+		if (ret)
+			return ret;
 
-	ret = nouveau_gpuobj_map_vm(grch->mmio, chan->vm, NV_MEM_ACCESS_RW |
-				    NV_MEM_ACCESS_SYS, &grch->mmio_vma);
-	if (ret)
-		return ret;
+		ret = nouveau_gpuobj_map_vm(grch->data[i].mem, chan->vm,
+					    data->access,
+					   &grch->data[i].vma);
+		if (ret)
+			return ret;
 
-	nv_wo32(grch->mmio, i++ * 4, 0x00408004);
-	nv_wo32(grch->mmio, i++ * 4, grch->unk408004_vma.offset >> 8);
-	nv_wo32(grch->mmio, i++ * 4, 0x00408008);
-	nv_wo32(grch->mmio, i++ * 4, 0x80000018);
-
-	nv_wo32(grch->mmio, i++ * 4, 0x0040800c);
-	nv_wo32(grch->mmio, i++ * 4, grch->unk40800c_vma.offset >> 8);
-	nv_wo32(grch->mmio, i++ * 4, 0x00408010);
-	nv_wo32(grch->mmio, i++ * 4, 0x80000000);
-
-	nv_wo32(grch->mmio, i++ * 4, 0x00418810);
-	nv_wo32(grch->mmio, i++ * 4, 0x80000000 | grch->unk418810_vma.offset >> 12);
-	nv_wo32(grch->mmio, i++ * 4, 0x00419848);
-	nv_wo32(grch->mmio, i++ * 4, 0x10000000 | grch->unk418810_vma.offset >> 12);
-
-	nv_wo32(grch->mmio, i++ * 4, 0x00419004);
-	nv_wo32(grch->mmio, i++ * 4, grch->unk40800c_vma.offset >> 8);
-	nv_wo32(grch->mmio, i++ * 4, 0x00419008);
-	nv_wo32(grch->mmio, i++ * 4, 0x00000000);
-
-	nv_wo32(grch->mmio, i++ * 4, 0x00418808);
-	nv_wo32(grch->mmio, i++ * 4, grch->unk408004_vma.offset >> 8);
-	nv_wo32(grch->mmio, i++ * 4, 0x0041880c);
-	nv_wo32(grch->mmio, i++ * 4, 0x80000018);
-
-	if (dev_priv->chipset != 0xc1) {
-		u32 magic = 0x02180000;
-		nv_wo32(grch->mmio, i++ * 4, 0x00405830);
-		nv_wo32(grch->mmio, i++ * 4, magic);
-		for (gpc = 0; gpc < priv->gpc_nr; gpc++) {
-			for (tp = 0; tp < priv->tpc_nr[gpc]; tp++) {
-				u32 reg = TPC_UNIT(gpc, tp, 0x520);
-				nv_wo32(grch->mmio, i++ * 4, reg);
-				nv_wo32(grch->mmio, i++ * 4, magic);
-				magic += 0x0324;
-			}
-		}
-	} else {
-		u32 magic = 0x02180000;
-		nv_wo32(grch->mmio, i++ * 4, 0x00405830);
-		nv_wo32(grch->mmio, i++ * 4, magic | 0x0000218);
-		nv_wo32(grch->mmio, i++ * 4, 0x004064c4);
-		nv_wo32(grch->mmio, i++ * 4, 0x0086ffff);
-		for (gpc = 0; gpc < priv->gpc_nr; gpc++) {
-			for (tp = 0; tp < priv->tpc_nr[gpc]; tp++) {
-				u32 reg = TPC_UNIT(gpc, tp, 0x520);
-				nv_wo32(grch->mmio, i++ * 4, reg);
-				nv_wo32(grch->mmio, i++ * 4, (1 << 28) | magic);
-				magic += 0x0324;
-			}
-			for (tp = 0; tp < priv->tpc_nr[gpc]; tp++) {
-				u32 reg = TPC_UNIT(gpc, tp, 0x544);
-				nv_wo32(grch->mmio, i++ * 4, reg);
-				nv_wo32(grch->mmio, i++ * 4, magic);
-				magic += 0x0324;
-			}
-		}
+		data++;
 	}
 
-	grch->mmio_nr = i / 2;
-	return 0;
-}
-
-static int
-nvc0_graph_context_new(struct nouveau_channel *chan, int engine)
-{
-	struct drm_device *dev = chan->dev;
-	struct nvc0_graph_priv *priv = nv_engine(dev, engine);
-	struct nvc0_graph_chan *grch;
-	struct nouveau_gpuobj *grctx;
-	int ret, i;
-
-	grch = kzalloc(sizeof(*grch), GFP_KERNEL);
-	if (!grch)
-		return -ENOMEM;
-	chan->engctx[NVOBJ_ENGINE_GR] = grch;
+	/* finally, fill in the mmio list and point the context at it */
+	for (i = 0; mmio->addr && i < ARRAY_SIZE(priv->mmio_list); i++) {
+		u32 addr = mmio->addr;
+		u32 data = mmio->data;
 
-	ret = nouveau_gpuobj_new(dev, NULL, priv->grctx_size, 256, 0,
-				 &grch->grctx);
-	if (ret)
-		goto error;
+		if (mmio->shift) {
+			u64 info = grch->data[mmio->buffer].vma.offset;
+			data |= info >> mmio->shift;
+		}
 
-	ret = nouveau_gpuobj_map_vm(grch->grctx, chan->vm, NV_MEM_ACCESS_RW |
-				    NV_MEM_ACCESS_SYS, &grch->grctx_vma);
-	if (ret)
-		return ret;
+		nv_wo32(grch->mmio, grch->mmio_nr++ * 4, addr);
+		nv_wo32(grch->mmio, grch->mmio_nr++ * 4, data);
+		mmio++;
+	}
 
-	grctx = grch->grctx;
-
-	ret = nvc0_graph_create_context_mmio_list(chan);
-	if (ret)
-		goto error;
+	for (i = 0; i < priv->size; i += 4)
+		nv_wo32(grch->grctx, i, priv->data[i / 4]);
 
 	nv_wo32(chan->ramin, 0x0210, lower_32_bits(grch->grctx_vma.offset) | 4);
 	nv_wo32(chan->ramin, 0x0214, upper_32_bits(grch->grctx_vma.offset));
 	nvimem_flush(dev);
 
-	if (!priv->grctx_vals) {
-		ret = nvc0_graph_construct_context(chan);
-		if (ret)
-			goto error;
-	}
-
-	for (i = 0; i < priv->grctx_size; i += 4)
-		nv_wo32(grctx, i, priv->grctx_vals[i / 4]);
-
-	if (!nouveau_ctxfw) {
-		nv_wo32(grctx, 0x00, grch->mmio_nr);
+	if (!priv->firmware) {
+		nv_wo32(grctx, 0x00, grch->mmio_nr / 2);
 		nv_wo32(grctx, 0x04, grch->mmio_vma.offset >> 8);
 	} else {
 		nv_wo32(grctx, 0xf4, 0);
 		nv_wo32(grctx, 0xf8, 0);
-		nv_wo32(grctx, 0x10, grch->mmio_nr);
+		nv_wo32(grctx, 0x10, grch->mmio_nr / 2);
 		nv_wo32(grctx, 0x14, lower_32_bits(grch->mmio_vma.offset));
 		nv_wo32(grctx, 0x18, upper_32_bits(grch->mmio_vma.offset));
 		nv_wo32(grctx, 0x1c, 1);
@@ -327,20 +160,21 @@ error:
 	return ret;
 }
 
-static void
+void
 nvc0_graph_context_del(struct nouveau_channel *chan, int engine)
 {
 	struct nvc0_graph_chan *grch = chan->engctx[engine];
+	int i;
+
+	for (i = 0; i < ARRAY_SIZE(grch->data); i++) {
+		nouveau_gpuobj_unmap(&grch->data[i].vma);
+		nouveau_gpuobj_ref(NULL, &grch->data[i].mem);
+	}
 
 	nouveau_gpuobj_unmap(&grch->mmio_vma);
-	nouveau_gpuobj_unmap(&grch->unk418810_vma);
-	nouveau_gpuobj_unmap(&grch->unk40800c_vma);
-	nouveau_gpuobj_unmap(&grch->unk408004_vma);
-	nouveau_gpuobj_unmap(&grch->grctx_vma);
 	nouveau_gpuobj_ref(NULL, &grch->mmio);
-	nouveau_gpuobj_ref(NULL, &grch->unk418810);
-	nouveau_gpuobj_ref(NULL, &grch->unk40800c);
-	nouveau_gpuobj_ref(NULL, &grch->unk408004);
+
+	nouveau_gpuobj_unmap(&grch->grctx_vma);
 	nouveau_gpuobj_ref(NULL, &grch->grctx);
 	chan->engctx[engine] = NULL;
 }
@@ -517,89 +351,102 @@ nvc0_graph_init_ctxctl(struct drm_device *dev)
 	u32 r000260;
 	int i;
 
-	if (!nouveau_ctxfw) {
-		/* load HUB microcode */
+	if (priv->firmware) {
+		/* load fuc microcode */
 		r000260 = nv_mask(dev, 0x000260, 0x00000001, 0x00000000);
-		nv_wr32(dev, 0x4091c0, 0x01000000);
-		for (i = 0; i < sizeof(nvc0_grhub_data) / 4; i++)
-			nv_wr32(dev, 0x4091c4, nvc0_grhub_data[i]);
-
-		nv_wr32(dev, 0x409180, 0x01000000);
-		for (i = 0; i < sizeof(nvc0_grhub_code) / 4; i++) {
-			if ((i & 0x3f) == 0)
-				nv_wr32(dev, 0x409188, i >> 6);
-			nv_wr32(dev, 0x409184, nvc0_grhub_code[i]);
-		}
-
-		/* load GPC microcode */
-		nv_wr32(dev, 0x41a1c0, 0x01000000);
-		for (i = 0; i < sizeof(nvc0_grgpc_data) / 4; i++)
-			nv_wr32(dev, 0x41a1c4, nvc0_grgpc_data[i]);
-
-		nv_wr32(dev, 0x41a180, 0x01000000);
-		for (i = 0; i < sizeof(nvc0_grgpc_code) / 4; i++) {
-			if ((i & 0x3f) == 0)
-				nv_wr32(dev, 0x41a188, i >> 6);
-			nv_wr32(dev, 0x41a184, nvc0_grgpc_code[i]);
-		}
+		nvc0_graph_init_fuc(dev, 0x409000, &priv->fuc409c,
+						   &priv->fuc409d);
+		nvc0_graph_init_fuc(dev, 0x41a000, &priv->fuc41ac,
+						   &priv->fuc41ad);
 		nv_wr32(dev, 0x000260, r000260);
 
-		/* start HUB ucode running, it'll init the GPCs */
-		nv_wr32(dev, 0x409800, dev_priv->chipset);
+		/* start both of them running */
+		nv_wr32(dev, 0x409840, 0xffffffff);
+		nv_wr32(dev, 0x41a10c, 0x00000000);
 		nv_wr32(dev, 0x40910c, 0x00000000);
+		nv_wr32(dev, 0x41a100, 0x00000002);
 		nv_wr32(dev, 0x409100, 0x00000002);
-		if (!nv_wait(dev, 0x409800, 0x80000000, 0x80000000)) {
-			NV_ERROR(dev, "PGRAPH: HUB_INIT timed out\n");
-			nvc0_graph_ctxctl_debug(dev);
+		if (!nv_wait(dev, 0x409800, 0x00000001, 0x00000001))
+			NV_INFO(dev, "0x409800 wait failed\n");
+
+		nv_wr32(dev, 0x409840, 0xffffffff);
+		nv_wr32(dev, 0x409500, 0x7fffffff);
+		nv_wr32(dev, 0x409504, 0x00000021);
+
+		nv_wr32(dev, 0x409840, 0xffffffff);
+		nv_wr32(dev, 0x409500, 0x00000000);
+		nv_wr32(dev, 0x409504, 0x00000010);
+		if (!nv_wait_ne(dev, 0x409800, 0xffffffff, 0x00000000)) {
+			NV_ERROR(dev, "fuc09 req 0x10 timeout\n");
 			return -EBUSY;
 		}
+		priv->size = nv_rd32(dev, 0x409800);
 
-		priv->grctx_size = nv_rd32(dev, 0x409804);
-		return 0;
+		nv_wr32(dev, 0x409840, 0xffffffff);
+		nv_wr32(dev, 0x409500, 0x00000000);
+		nv_wr32(dev, 0x409504, 0x00000016);
+		if (!nv_wait_ne(dev, 0x409800, 0xffffffff, 0x00000000)) {
+			NV_ERROR(dev, "fuc09 req 0x16 timeout\n");
+			return -EBUSY;
+		}
+
+		nv_wr32(dev, 0x409840, 0xffffffff);
+		nv_wr32(dev, 0x409500, 0x00000000);
+		nv_wr32(dev, 0x409504, 0x00000025);
+		if (!nv_wait_ne(dev, 0x409800, 0xffffffff, 0x00000000)) {
+			NV_ERROR(dev, "fuc09 req 0x25 timeout\n");
+			return -EBUSY;
+		}
+
+		goto done;
 	}
 
-	/* load fuc microcode */
+	/* load HUB microcode */
 	r000260 = nv_mask(dev, 0x000260, 0x00000001, 0x00000000);
-	nvc0_graph_init_fuc(dev, 0x409000, &priv->fuc409c, &priv->fuc409d);
-	nvc0_graph_init_fuc(dev, 0x41a000, &priv->fuc41ac, &priv->fuc41ad);
+	nv_wr32(dev, 0x4091c0, 0x01000000);
+	for (i = 0; i < sizeof(nvc0_grhub_data) / 4; i++)
+		nv_wr32(dev, 0x4091c4, nvc0_grhub_data[i]);
+
+	nv_wr32(dev, 0x409180, 0x01000000);
+	for (i = 0; i < sizeof(nvc0_grhub_code) / 4; i++) {
+		if ((i & 0x3f) == 0)
+			nv_wr32(dev, 0x409188, i >> 6);
+		nv_wr32(dev, 0x409184, nvc0_grhub_code[i]);
+	}
+
+	/* load GPC microcode */
+	nv_wr32(dev, 0x41a1c0, 0x01000000);
+	for (i = 0; i < sizeof(nvc0_grgpc_data) / 4; i++)
+		nv_wr32(dev, 0x41a1c4, nvc0_grgpc_data[i]);
+
+	nv_wr32(dev, 0x41a180, 0x01000000);
+	for (i = 0; i < sizeof(nvc0_grgpc_code) / 4; i++) {
+		if ((i & 0x3f) == 0)
+			nv_wr32(dev, 0x41a188, i >> 6);
+		nv_wr32(dev, 0x41a184, nvc0_grgpc_code[i]);
+	}
 	nv_wr32(dev, 0x000260, r000260);
 
-	/* start both of them running */
-	nv_wr32(dev, 0x409840, 0xffffffff);
-	nv_wr32(dev, 0x41a10c, 0x00000000);
+	/* start HUB ucode running, it'll init the GPCs */
+	nv_wr32(dev, 0x409800, dev_priv->chipset);
 	nv_wr32(dev, 0x40910c, 0x00000000);
-	nv_wr32(dev, 0x41a100, 0x00000002);
 	nv_wr32(dev, 0x409100, 0x00000002);
-	if (!nv_wait(dev, 0x409800, 0x00000001, 0x00000001))
-		NV_INFO(dev, "0x409800 wait failed\n");
-
-	nv_wr32(dev, 0x409840, 0xffffffff);
-	nv_wr32(dev, 0x409500, 0x7fffffff);
-	nv_wr32(dev, 0x409504, 0x00000021);
-
-	nv_wr32(dev, 0x409840, 0xffffffff);
-	nv_wr32(dev, 0x409500, 0x00000000);
-	nv_wr32(dev, 0x409504, 0x00000010);
-	if (!nv_wait_ne(dev, 0x409800, 0xffffffff, 0x00000000)) {
-		NV_ERROR(dev, "fuc09 req 0x10 timeout\n");
+	if (!nv_wait(dev, 0x409800, 0x80000000, 0x80000000)) {
+		NV_ERROR(dev, "PGRAPH: HUB_INIT timed out\n");
+		nvc0_graph_ctxctl_debug(dev);
 		return -EBUSY;
 	}
-	priv->grctx_size = nv_rd32(dev, 0x409800);
 
-	nv_wr32(dev, 0x409840, 0xffffffff);
-	nv_wr32(dev, 0x409500, 0x00000000);
-	nv_wr32(dev, 0x409504, 0x00000016);
-	if (!nv_wait_ne(dev, 0x409800, 0xffffffff, 0x00000000)) {
-		NV_ERROR(dev, "fuc09 req 0x16 timeout\n");
-		return -EBUSY;
-	}
+	priv->size = nv_rd32(dev, 0x409804);
+done:
+	if (priv->data == NULL) {
+		int ret = nvc0_grctx_generate(dev);
+		if (ret) {
+			NV_ERROR(dev, "PGRAPH: failed to construct context\n");
+			return ret;
+		}
 
-	nv_wr32(dev, 0x409840, 0xffffffff);
-	nv_wr32(dev, 0x409500, 0x00000000);
-	nv_wr32(dev, 0x409504, 0x00000025);
-	if (!nv_wait_ne(dev, 0x409800, 0xffffffff, 0x00000000)) {
-		NV_ERROR(dev, "fuc09 req 0x25 timeout\n");
-		return -EBUSY;
+		return 1;
 	}
 
 	return 0;
@@ -610,6 +457,7 @@ nvc0_graph_init(struct drm_device *dev, int engine)
 {
 	int ret;
 
+reset:
 	nv_mask(dev, 0x000200, 0x18001000, 0x00000000);
 	nv_mask(dev, 0x000200, 0x18001000, 0x18001000);
 
@@ -636,8 +484,11 @@ nvc0_graph_init(struct drm_device *dev, int engine)
 	nv_wr32(dev, 0x400054, 0x34ce3464);
 
 	ret = nvc0_graph_init_ctxctl(dev);
-	if (ret)
+	if (ret) {
+		if (ret == 1)
+			goto reset;
 		return ret;
+	}
 
 	return 0;
 }
@@ -784,20 +635,18 @@ nvc0_graph_destroy(struct drm_device *dev, int engine)
 {
 	struct nvc0_graph_priv *priv = nv_engine(dev, engine);
 
-	if (nouveau_ctxfw) {
-		nvc0_graph_destroy_fw(&priv->fuc409c);
-		nvc0_graph_destroy_fw(&priv->fuc409d);
-		nvc0_graph_destroy_fw(&priv->fuc41ac);
-		nvc0_graph_destroy_fw(&priv->fuc41ad);
-	}
+	nvc0_graph_destroy_fw(&priv->fuc409c);
+	nvc0_graph_destroy_fw(&priv->fuc409d);
+	nvc0_graph_destroy_fw(&priv->fuc41ac);
+	nvc0_graph_destroy_fw(&priv->fuc41ad);
 
 	nouveau_irq_unregister(dev, 12);
 
 	nouveau_gpuobj_ref(NULL, &priv->unk4188b8);
 	nouveau_gpuobj_ref(NULL, &priv->unk4188b4);
 
-	if (priv->grctx_vals)
-		kfree(priv->grctx_vals);
+	if (priv->data)
+		kfree(priv->data);
 
 	NVOBJ_ENGINE_DEL(dev, GR);
 	kfree(priv);
@@ -840,6 +689,7 @@ nvc0_graph_create(struct drm_device *dev)
 			ret = 0;
 			goto error;
 		}
+		priv->firmware = true;
 	}
 
 	ret = nouveau_gpuobj_new(dev, NULL, 0x1000, 256, 0, &priv->unk4188b4);
diff --git a/drivers/gpu/drm/nouveau/core/engine/graph/nvc0.h b/drivers/gpu/drm/nouveau/core/engine/graph/nvc0.h
index ba9b80fd8958..30ea3ab135c6 100644
--- a/drivers/gpu/drm/nouveau/core/engine/graph/nvc0.h
+++ b/drivers/gpu/drm/nouveau/core/engine/graph/nvc0.h
@@ -66,27 +66,28 @@ struct nvc0_graph_priv {
 	u8 tpc_nr[GPC_MAX];
 	u8 tpc_total;
 
-	u32  grctx_size;
-	u32 *grctx_vals;
 	struct nouveau_gpuobj *unk4188b4;
 	struct nouveau_gpuobj *unk4188b8;
 
+	struct nvc0_graph_data mmio_data[4];
+	struct nvc0_graph_mmio mmio_list[4096/8];
+	u32  size;
+	u32 *data;
+
 	u8 magic_not_rop_nr;
 };
 
 struct nvc0_graph_chan {
 	struct nouveau_gpuobj *grctx;
 	struct nouveau_vma     grctx_vma;
-	struct nouveau_gpuobj *unk408004; /* 0x418808 too */
-	struct nouveau_vma     unk408004_vma;
-	struct nouveau_gpuobj *unk40800c; /* 0x419004 too */
-	struct nouveau_vma     unk40800c_vma;
-	struct nouveau_gpuobj *unk418810; /* 0x419848 too */
-	struct nouveau_vma     unk418810_vma;
 
 	struct nouveau_gpuobj *mmio;
 	struct nouveau_vma     mmio_vma;
 	int mmio_nr;
+	struct {
+		struct nouveau_gpuobj *mem;
+		struct nouveau_vma vma;
+	} data[4];
 };
 
 static inline u32
@@ -124,6 +125,7 @@ nv_mthd(struct drm_device *priv, u32 class, u32 mthd, u32 data)
 }
 
 struct nvc0_grctx {
+	struct drm_device *dev;
 	struct nvc0_graph_priv *priv;
 	struct nvc0_graph_data *data;
 	struct nvc0_graph_mmio *mmio;
@@ -133,13 +135,14 @@ struct nvc0_grctx {
 	u64 addr;
 };
 
-int  nvc0_grctx_generate(struct nouveau_channel *);
-int  nvc0_grctx_init(struct nvc0_graph_priv *, struct nvc0_grctx *);
+int  nvc0_grctx_generate(struct drm_device *);
+int  nvc0_grctx_init(struct drm_device *, struct nvc0_graph_priv *,
+		     struct nvc0_grctx *);
 void nvc0_grctx_data(struct nvc0_grctx *, u32, u32, u32);
 void nvc0_grctx_mmio(struct nvc0_grctx *, u32, u32, u32, u32);
 int  nvc0_grctx_fini(struct nvc0_grctx *);
 
-int  nve0_grctx_generate(struct nouveau_channel *);
+int  nve0_grctx_generate(struct drm_device *);
 
 #define mmio_data(s,a,p) nvc0_grctx_data(&info, (s), (a), (p))
 #define mmio_list(r,d,s,b) nvc0_grctx_mmio(&info, (r), (d), (s), (b))
@@ -154,4 +157,9 @@ int  nvc0_graph_context_ctor(struct nouveau_object *, struct nouveau_object *,
 			     struct nouveau_object **);
 void nvc0_graph_context_dtor(struct nouveau_object *);
 
+void nvc0_graph_ctxctl_debug(struct drm_device *);
+
+int  nvc0_graph_context_new(struct nouveau_channel *, int);
+void nvc0_graph_context_del(struct nouveau_channel *, int);
+
 #endif
diff --git a/drivers/gpu/drm/nouveau/core/engine/graph/nve0.c b/drivers/gpu/drm/nouveau/core/engine/graph/nve0.c
index 2bd55d8d10a7..a3a4ee7c0b2e 100644
--- a/drivers/gpu/drm/nouveau/core/engine/graph/nve0.c
+++ b/drivers/gpu/drm/nouveau/core/engine/graph/nve0.c
@@ -58,243 +58,6 @@ nve0_graph_ctxctl_debug(struct drm_device *dev)
 }
 
 static int
-nve0_graph_load_context(struct nouveau_channel *chan)
-{
-	struct drm_device *dev = chan->dev;
-
-	nv_wr32(dev, 0x409840, 0x00000030);
-	nv_wr32(dev, 0x409500, 0x80000000 | chan->ramin->addr >> 12);
-	nv_wr32(dev, 0x409504, 0x00000003);
-	if (!nv_wait(dev, 0x409800, 0x00000010, 0x00000010))
-		NV_ERROR(dev, "PGRAPH: load_ctx timeout\n");
-
-	return 0;
-}
-
-static int
-nve0_graph_unload_context_to(struct drm_device *dev, u64 chan)
-{
-	nv_wr32(dev, 0x409840, 0x00000003);
-	nv_wr32(dev, 0x409500, 0x80000000 | chan >> 12);
-	nv_wr32(dev, 0x409504, 0x00000009);
-	if (!nv_wait(dev, 0x409800, 0x00000001, 0x00000000)) {
-		NV_ERROR(dev, "PGRAPH: unload_ctx timeout\n");
-		return -EBUSY;
-	}
-
-	return 0;
-}
-
-static int
-nve0_graph_construct_context(struct nouveau_channel *chan)
-{
-	struct nvc0_graph_priv *priv = nv_engine(chan->dev, NVOBJ_ENGINE_GR);
-	struct nvc0_graph_chan *grch = chan->engctx[NVOBJ_ENGINE_GR];
-	struct drm_device *dev = chan->dev;
-	int ret, i;
-	u32 *ctx;
-
-	ctx = kmalloc(priv->grctx_size, GFP_KERNEL);
-	if (!ctx)
-		return -ENOMEM;
-
-	nve0_graph_load_context(chan);
-
-	nv_wo32(grch->grctx, 0x1c, 1);
-	nv_wo32(grch->grctx, 0x20, 0);
-	nv_wo32(grch->grctx, 0x28, 0);
-	nv_wo32(grch->grctx, 0x2c, 0);
-	nvimem_flush(dev);
-
-	ret = nve0_grctx_generate(chan);
-	if (ret)
-		goto err;
-
-	ret = nve0_graph_unload_context_to(dev, chan->ramin->addr);
-	if (ret)
-		goto err;
-
-	for (i = 0; i < priv->grctx_size; i += 4)
-		ctx[i / 4] = nv_ro32(grch->grctx, i);
-
-	priv->grctx_vals = ctx;
-	return 0;
-
-err:
-	kfree(ctx);
-	return ret;
-}
-
-static int
-nve0_graph_create_context_mmio_list(struct nouveau_channel *chan)
-{
-	struct nvc0_graph_priv *priv = nv_engine(chan->dev, NVOBJ_ENGINE_GR);
-	struct nvc0_graph_chan *grch = chan->engctx[NVOBJ_ENGINE_GR];
-	struct drm_device *dev = chan->dev;
-	u32 magic[GPC_MAX][2];
-	u16 offset = 0x0000;
-	int gpc;
-	int ret;
-
-	ret = nouveau_gpuobj_new(dev, NULL, 0x3000, 256, 0, &grch->unk408004);
-	if (ret)
-		return ret;
-
-	ret = nouveau_gpuobj_map_vm(grch->unk408004, chan->vm,
-				    NV_MEM_ACCESS_RW | NV_MEM_ACCESS_SYS,
-				    &grch->unk408004_vma);
-	if (ret)
-		return ret;
-
-	ret = nouveau_gpuobj_new(dev, NULL, 0x8000, 256, 0, &grch->unk40800c);
-	if (ret)
-		return ret;
-
-	ret = nouveau_gpuobj_map_vm(grch->unk40800c, chan->vm,
-				    NV_MEM_ACCESS_RW | NV_MEM_ACCESS_SYS,
-				    &grch->unk40800c_vma);
-	if (ret)
-		return ret;
-
-	ret = nouveau_gpuobj_new(dev, NULL, 384 * 1024, 4096, 0,
-				 &grch->unk418810);
-	if (ret)
-		return ret;
-
-	ret = nouveau_gpuobj_map_vm(grch->unk418810, chan->vm,
-				    NV_MEM_ACCESS_RW, &grch->unk418810_vma);
-	if (ret)
-		return ret;
-
-	ret = nouveau_gpuobj_new(dev, NULL, 0x1000, 0, 0, &grch->mmio);
-	if (ret)
-		return ret;
-
-	ret = nouveau_gpuobj_map_vm(grch->mmio, chan->vm,
-				    NV_MEM_ACCESS_RW | NV_MEM_ACCESS_SYS,
-				    &grch->mmio_vma);
-	if (ret)
-		return ret;
-
-#define mmio(r,v) do {                                                         \
-	nv_wo32(grch->mmio, (grch->mmio_nr * 8) + 0, (r));                     \
-	nv_wo32(grch->mmio, (grch->mmio_nr * 8) + 4, (v));                     \
-	grch->mmio_nr++;                                                       \
-} while (0)
-	mmio(0x40800c, grch->unk40800c_vma.offset >> 8);
-	mmio(0x408010, 0x80000000);
-	mmio(0x419004, grch->unk40800c_vma.offset >> 8);
-	mmio(0x419008, 0x00000000);
-	mmio(0x4064cc, 0x80000000);
-	mmio(0x408004, grch->unk408004_vma.offset >> 8);
-	mmio(0x408008, 0x80000030);
-	mmio(0x418808, grch->unk408004_vma.offset >> 8);
-	mmio(0x41880c, 0x80000030);
-	mmio(0x4064c8, 0x01800600);
-	mmio(0x418810, 0x80000000 | grch->unk418810_vma.offset >> 12);
-	mmio(0x419848, 0x10000000 | grch->unk418810_vma.offset >> 12);
-	mmio(0x405830, 0x02180648);
-	mmio(0x4064c4, 0x0192ffff);
-
-	for (gpc = 0; gpc < priv->gpc_nr; gpc++) {
-		u16 magic0 = 0x0218 * priv->tpc_nr[gpc];
-		u16 magic1 = 0x0648 * priv->tpc_nr[gpc];
-		magic[gpc][0]  = 0x10000000 | (magic0 << 16) | offset;
-		magic[gpc][1]  = 0x00000000 | (magic1 << 16);
-		offset += 0x0324 * priv->tpc_nr[gpc];
-	}
-
-	for (gpc = 0; gpc < priv->gpc_nr; gpc++) {
-		mmio(GPC_UNIT(gpc, 0x30c0), magic[gpc][0]);
-		mmio(GPC_UNIT(gpc, 0x30e4), magic[gpc][1] | offset);
-		offset += 0x07ff * priv->tpc_nr[gpc];
-	}
-
-	mmio(0x17e91c, 0x06060609);
-	mmio(0x17e920, 0x00090a05);
-#undef mmio
-	return 0;
-}
-
-static int
-nve0_graph_context_new(struct nouveau_channel *chan, int engine)
-{
-	struct drm_device *dev = chan->dev;
-	struct nvc0_graph_priv *priv = nv_engine(dev, engine);
-	struct nvc0_graph_chan *grch;
-	struct nouveau_gpuobj *grctx;
-	int ret, i;
-
-	grch = kzalloc(sizeof(*grch), GFP_KERNEL);
-	if (!grch)
-		return -ENOMEM;
-	chan->engctx[NVOBJ_ENGINE_GR] = grch;
-
-	ret = nouveau_gpuobj_new(dev, NULL, priv->grctx_size, 256, 0,
-				 &grch->grctx);
-	if (ret)
-		goto error;
-
-	ret = nouveau_gpuobj_map_vm(grch->grctx, chan->vm, NV_MEM_ACCESS_RW |
-				    NV_MEM_ACCESS_SYS, &grch->grctx_vma);
-	if (ret)
-		return ret;
-
-	grctx = grch->grctx;
-
-	ret = nve0_graph_create_context_mmio_list(chan);
-	if (ret)
-		goto error;
-
-	nv_wo32(chan->ramin, 0x0210, lower_32_bits(grch->grctx_vma.offset) | 4);
-	nv_wo32(chan->ramin, 0x0214, upper_32_bits(grch->grctx_vma.offset));
-	nvimem_flush(dev);
-
-	if (!priv->grctx_vals) {
-		ret = nve0_graph_construct_context(chan);
-		if (ret)
-			goto error;
-	}
-
-	for (i = 0; i < priv->grctx_size; i += 4)
-		nv_wo32(grctx, i, priv->grctx_vals[i / 4]);
-	nv_wo32(grctx, 0xf4, 0);
-	nv_wo32(grctx, 0xf8, 0);
-	nv_wo32(grctx, 0x10, grch->mmio_nr);
-	nv_wo32(grctx, 0x14, lower_32_bits(grch->mmio_vma.offset));
-	nv_wo32(grctx, 0x18, upper_32_bits(grch->mmio_vma.offset));
-	nv_wo32(grctx, 0x1c, 1);
-	nv_wo32(grctx, 0x20, 0);
-	nv_wo32(grctx, 0x28, 0);
-	nv_wo32(grctx, 0x2c, 0);
-
-	nvimem_flush(dev);
-	return 0;
-
-error:
-	priv->base.context_del(chan, engine);
-	return ret;
-}
-
-static void
-nve0_graph_context_del(struct nouveau_channel *chan, int engine)
-{
-	struct nvc0_graph_chan *grch = chan->engctx[engine];
-
-	nouveau_gpuobj_unmap(&grch->mmio_vma);
-	nouveau_gpuobj_unmap(&grch->unk418810_vma);
-	nouveau_gpuobj_unmap(&grch->unk40800c_vma);
-	nouveau_gpuobj_unmap(&grch->unk408004_vma);
-	nouveau_gpuobj_unmap(&grch->grctx_vma);
-	nouveau_gpuobj_ref(NULL, &grch->mmio);
-	nouveau_gpuobj_ref(NULL, &grch->unk418810);
-	nouveau_gpuobj_ref(NULL, &grch->unk40800c);
-	nouveau_gpuobj_ref(NULL, &grch->unk408004);
-	nouveau_gpuobj_ref(NULL, &grch->grctx);
-	chan->engctx[engine] = NULL;
-}
-
-static int
 nve0_graph_object_new(struct nouveau_channel *chan, int engine,
 		      u32 handle, u16 class)
 {
@@ -487,7 +250,7 @@ nve0_graph_init_ctxctl(struct drm_device *dev)
 		NV_ERROR(dev, "fuc09 req 0x10 timeout\n");
 		return -EBUSY;
 	}
-	priv->grctx_size = nv_rd32(dev, 0x409800);
+	priv->size = nv_rd32(dev, 0x409800);
 
 	nv_wr32(dev, 0x409840, 0xffffffff);
 	nv_wr32(dev, 0x409500, 0x00000000);
@@ -534,6 +297,17 @@ nve0_graph_init_ctxctl(struct drm_device *dev)
 	nv_wr32(dev, 0x409614, 0x00000070);
 	nv_wr32(dev, 0x409614, 0x00000770);
 	nv_wr32(dev, 0x40802c, 0x00000001);
+
+	if (priv->data == NULL) {
+		int ret = nve0_grctx_generate(dev);
+		if (ret) {
+			NV_ERROR(dev, "PGRAPH: failed to construct context\n");
+			return ret;
+		}
+
+		return 1;
+	}
+
 	return 0;
 }
 
@@ -542,6 +316,7 @@ nve0_graph_init(struct drm_device *dev, int engine)
 {
 	int ret;
 
+reset:
 	nv_mask(dev, 0x000200, 0x18001000, 0x00000000);
 	nv_mask(dev, 0x000200, 0x18001000, 0x18001000);
 
@@ -566,8 +341,11 @@ nve0_graph_init(struct drm_device *dev, int engine)
 	nv_wr32(dev, 0x400054, 0x34ce3464);
 
 	ret = nve0_graph_init_ctxctl(dev);
-	if (ret)
+	if (ret) {
+		if (ret == 1)
+			goto reset;
 		return ret;
+	}
 
 	return 0;
 }
@@ -758,8 +536,8 @@ nve0_graph_destroy(struct drm_device *dev, int engine)
 	nouveau_gpuobj_ref(NULL, &priv->unk4188b8);
 	nouveau_gpuobj_ref(NULL, &priv->unk4188b4);
 
-	if (priv->grctx_vals)
-		kfree(priv->grctx_vals);
+	if (priv->data)
+		kfree(priv->data);
 
 	NVOBJ_ENGINE_DEL(dev, GR);
 	kfree(priv);
@@ -786,8 +564,8 @@ nve0_graph_create(struct drm_device *dev)
 	priv->base.destroy = nve0_graph_destroy;
 	priv->base.init = nve0_graph_init;
 	priv->base.fini = nve0_graph_fini;
-	priv->base.context_new = nve0_graph_context_new;
-	priv->base.context_del = nve0_graph_context_del;
+	priv->base.context_new = nvc0_graph_context_new;
+	priv->base.context_del = nvc0_graph_context_del;
 	priv->base.object_new = nve0_graph_object_new;
 
 	NVOBJ_ENGINE_ADD(dev, GR, &priv->base);
@@ -801,6 +579,7 @@ nve0_graph_create(struct drm_device *dev)
 		ret = 0;
 		goto error;
 	}
+	priv->firmware = true;
 
 	ret = nouveau_gpuobj_new(dev, NULL, 0x1000, 256, 0, &priv->unk4188b4);
 	if (ret)