drm/nve0: use async copy engine for ttm buffer moves if available

Kepler PFIFO lost the ability to address multiple engines from a single
channel, so we need a separate one for the copy engine.

v2: Marcin Slusarz <marcin.slusarz@gmail.com>
- regression fix: restore hw accelerated buffer copies

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

drivers/gpu/drm/nouveau/nouveau_abi16.c

@@ -264,6 +264,11 @@ nouveau_abi16_ioctl_channel_alloc(ABI16_IOCTL_ARGS)
 	abi16->handles |= (1 << init->channel);
 
 	/* create channel object and initialise dma and fence management */
+	if (device->card_type >= NV_E0) {
+		init->fb_ctxdma_handle = NVE0_CHANNEL_IND_ENGINE_GR;
+		init->tt_ctxdma_handle = 0;
+	}
+
 	ret = nouveau_channel_new(drm, cli, NVDRM_DEVICE, NVDRM_CHAN |
 				  init->channel, init->fb_ctxdma_handle,
 				  init->tt_ctxdma_handle, &chan->chan);

drivers/gpu/drm/nouveau/nouveau_bo.c

@@ -570,6 +570,18 @@ nouveau_bo_move_accel_cleanup(struct nouveau_channel *chan,
 	return ret;
 }
 
+static int
+nve0_bo_move_init(struct nouveau_channel *chan, u32 handle)
+{
+	int ret = RING_SPACE(chan, 2);
+	if (ret == 0) {
+		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
+		OUT_RING  (chan, handle);
+		FIRE_RING (chan);
+	}
+	return ret;
+}
+
 static int
 nve0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
@@ -991,10 +1003,8 @@ nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
 }
 
 void
-nouveau_bo_move_init(struct nouveau_channel *chan)
+nouveau_bo_move_init(struct nouveau_drm *drm)
 {
-	struct nouveau_cli *cli = chan->cli;
-	struct nouveau_drm *drm = chan->drm;
 	static const struct {
 		const char *name;
 		int engine;
@@ -1004,7 +1014,8 @@ nouveau_bo_move_init(struct nouveau_channel *chan)
 			    struct ttm_mem_reg *, struct ttm_mem_reg *);
 		int (*init)(struct nouveau_channel *, u32 handle);
 	} _methods[] = {
-		{  "COPY", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
+		{  "COPY", 0, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
+		{  "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
 		{ "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
 		{ "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
 		{  "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
@@ -1020,14 +1031,22 @@ nouveau_bo_move_init(struct nouveau_channel *chan)
 
 	do {
 		struct nouveau_object *object;
+		struct nouveau_channel *chan;
 		u32 handle = (mthd->engine << 16) | mthd->oclass;
 
-		ret = nouveau_object_new(nv_object(cli), chan->handle, handle,
+		if (mthd->init == nve0_bo_move_init)
+			chan = drm->cechan;
+		else
+			chan = drm->channel;
+		if (chan == NULL)
+			continue;
+
+		ret = nouveau_object_new(nv_object(drm), chan->handle, handle,
 					 mthd->oclass, NULL, 0, &object);
 		if (ret == 0) {
 			ret = mthd->init(chan, handle);
 			if (ret) {
-				nouveau_object_del(nv_object(cli),
+				nouveau_object_del(nv_object(drm),
 						   chan->handle, handle);
 				continue;
 			}

drivers/gpu/drm/nouveau/nouveau_bo.h

@@ -61,7 +61,7 @@ nouveau_bo_ref(struct nouveau_bo *ref, struct nouveau_bo **pnvbo)
 
 extern struct ttm_bo_driver nouveau_bo_driver;
 
-void nouveau_bo_move_init(struct nouveau_channel *);
+void nouveau_bo_move_init(struct nouveau_drm *);
 int  nouveau_bo_new(struct drm_device *, int size, int align, u32 flags,
 		    u32 tile_mode, u32 tile_flags, struct sg_table *sg,
 		    struct nouveau_bo **);

drivers/gpu/drm/nouveau/nouveau_chan.c

@@ -184,7 +184,8 @@ nouveau_channel_prep(struct nouveau_drm *drm, struct nouveau_cli *cli,
 
 int
 nouveau_channel_ind(struct nouveau_drm *drm, struct nouveau_cli *cli,
-		    u32 parent, u32 handle, struct nouveau_channel **pchan)
+		    u32 parent, u32 handle, u32 engine,
+		    struct nouveau_channel **pchan)
 {
 	static const u16 oclasses[] = { 0xa06f, 0x906f, 0x826f, 0x506f, 0 };
 	const u16 *oclass = oclasses;
@@ -202,7 +203,7 @@ nouveau_channel_ind(struct nouveau_drm *drm, struct nouveau_cli *cli,
 	args.pushbuf = chan->push.handle;
 	args.ioffset = 0x10000 + chan->push.vma.offset;
 	args.ilength = 0x02000;
-	args.engine  = NVE0_CHANNEL_IND_ENGINE_GR;
+	args.engine  = engine;
 
 	do {
 		ret = nouveau_object_new(nv_object(cli), parent, handle,
@@ -261,9 +262,6 @@ nouveau_channel_init(struct nouveau_channel *chan, u32 vram, u32 gart)
 	struct nv_dma_class args;
 	int ret, i;
 
-	chan->vram = vram;
-	chan->gart = gart;
-
 	/* allocate dma objects to cover all allowed vram, and gart */
 	if (device->card_type < NV_C0) {
 		if (device->card_type >= NV_50) {
@@ -301,6 +299,9 @@ nouveau_channel_init(struct nouveau_channel *chan, u32 vram, u32 gart)
 					 0x003d, &args, sizeof(args), &object);
 		if (ret)
 			return ret;
+
+		chan->vram = vram;
+		chan->gart = gart;
 	}
 
 	/* initialise dma tracking parameters */
@@ -336,15 +337,17 @@ nouveau_channel_init(struct nouveau_channel *chan, u32 vram, u32 gart)
 	/* allocate software object class (used for fences on <= nv05, and
 	 * to signal flip completion), bind it to a subchannel.
 	 */
-	ret = nouveau_object_new(nv_object(client), chan->handle,
-				 NvSw, nouveau_abi16_swclass(chan->drm),
-				 NULL, 0, &object);
-	if (ret)
-		return ret;
+	if (chan != chan->drm->cechan) {
+		ret = nouveau_object_new(nv_object(client), chan->handle,
+					 NvSw, nouveau_abi16_swclass(chan->drm),
+					 NULL, 0, &object);
+		if (ret)
+			return ret;
 
-	swch = (void *)object->parent;
-	swch->flip = nouveau_flip_complete;
-	swch->flip_data = chan;
+		swch = (void *)object->parent;
+		swch->flip = nouveau_flip_complete;
+		swch->flip_data = chan;
+	}
 
 	if (device->card_type < NV_C0) {
 		ret = RING_SPACE(chan, 2);
@@ -362,12 +365,12 @@ nouveau_channel_init(struct nouveau_channel *chan, u32 vram, u32 gart)
 
 int
 nouveau_channel_new(struct nouveau_drm *drm, struct nouveau_cli *cli,
-		    u32 parent, u32 handle, u32 vram, u32 gart,
+		    u32 parent, u32 handle, u32 arg0, u32 arg1,
 		    struct nouveau_channel **pchan)
 {
 	int ret;
 
-	ret = nouveau_channel_ind(drm, cli, parent, handle, pchan);
+	ret = nouveau_channel_ind(drm, cli, parent, handle, arg0, pchan);
 	if (ret) {
 		NV_DEBUG(drm, "ib channel create, %d\n", ret);
 		ret = nouveau_channel_dma(drm, cli, parent, handle, pchan);
@@ -377,7 +380,7 @@ nouveau_channel_new(struct nouveau_drm *drm, struct nouveau_cli *cli,
 		}
 	}
 
-	ret = nouveau_channel_init(*pchan, vram, gart);
+	ret = nouveau_channel_init(*pchan, arg0, arg1);
 	if (ret) {
 		NV_ERROR(drm, "channel failed to initialise, %d\n", ret);
 		nouveau_channel_del(pchan);

drivers/gpu/drm/nouveau/nouveau_chan.h

@@ -39,7 +39,7 @@ struct nouveau_channel {
 
 
 int  nouveau_channel_new(struct nouveau_drm *, struct nouveau_cli *,
-			 u32 parent, u32 handle, u32 vram, u32 gart,
+			 u32 parent, u32 handle, u32 arg0, u32 arg1,
 			 struct nouveau_channel **);
 void nouveau_channel_del(struct nouveau_channel **);
 int  nouveau_channel_idle(struct nouveau_channel *);

drivers/gpu/drm/nouveau/nouveau_drm.c

@@ -109,6 +109,7 @@ nouveau_accel_fini(struct nouveau_drm *drm)
 {
 	nouveau_gpuobj_ref(NULL, &drm->notify);
 	nouveau_channel_del(&drm->channel);
+	nouveau_channel_del(&drm->cechan);
 	if (drm->fence)
 		nouveau_fence(drm)->dtor(drm);
 }
@@ -118,6 +119,7 @@ nouveau_accel_init(struct nouveau_drm *drm)
 {
 	struct nouveau_device *device = nv_device(drm->device);
 	struct nouveau_object *object;
+	u32 arg0, arg1;
 	int ret;
 
 	if (nouveau_noaccel)
@@ -134,8 +136,24 @@ nouveau_accel_init(struct nouveau_drm *drm)
 		return;
 	}
 
+	if (device->card_type >= NV_E0) {
+		ret = nouveau_channel_new(drm, &drm->client, NVDRM_DEVICE,
+					  NVDRM_CHAN + 1,
+					  NVE0_CHANNEL_IND_ENGINE_CE0 |
+					  NVE0_CHANNEL_IND_ENGINE_CE1, 0,
+					  &drm->cechan);
+		if (ret)
+			NV_ERROR(drm, "failed to create ce channel, %d\n", ret);
+
+		arg0 = NVE0_CHANNEL_IND_ENGINE_GR;
+		arg1 = 0;
+	} else {
+		arg0 = NvDmaFB;
+		arg1 = NvDmaTT;
+	}
+
 	ret = nouveau_channel_new(drm, &drm->client, NVDRM_DEVICE, NVDRM_CHAN,
-				  NvDmaFB, NvDmaTT, &drm->channel);
+				  arg0, arg1, &drm->channel);
 	if (ret) {
 		NV_ERROR(drm, "failed to create kernel channel, %d\n", ret);
 		nouveau_accel_fini(drm);
@@ -167,7 +185,7 @@ nouveau_accel_init(struct nouveau_drm *drm)
 	}
 
 
-	nouveau_bo_move_init(drm->channel);
+	nouveau_bo_move_init(drm);
 }
 
 static int __devinit

drivers/gpu/drm/nouveau/nouveau_drm.h

@@ -97,6 +97,7 @@ struct nouveau_drm {
 	void *fence;
 
 	/* context for accelerated drm-internal operations */
+	struct nouveau_channel *cechan;
 	struct nouveau_channel *channel;
 	struct nouveau_gpuobj *notify;
 	struct nouveau_fbdev *fbcon;