drm/nouveau: allow accelerated buffer moves even when gr isn't present

There's no need to avoid using copy engines if gr init fails for some
reason (usually missing FW, or incomplete bring-up).

It's not terribly useful for an end-user, but it'll slightly speed up
suspend/resume when saving fb contents, and allow for host/ce code to
be validated.

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

1 files changed, 128 insertions, 85 deletions

diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c b/drivers/gpu/drm/nouveau/nouveau_drm.c
index 50025e1ff857..c1e9069bd2ce 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drm.c
+++ b/drivers/gpu/drm/nouveau/nouveau_drm.c
@@ -282,18 +282,134 @@ done:
 }
 
 static void
-nouveau_accel_fini(struct nouveau_drm *drm)
+nouveau_accel_ce_fini(struct nouveau_drm *drm)
+{
+	nouveau_channel_idle(drm->cechan);
+	nvif_object_fini(&drm->ttm.copy);
+	nouveau_channel_del(&drm->cechan);
+}
+
+static void
+nouveau_accel_ce_init(struct nouveau_drm *drm)
+{
+	struct nvif_device *device = &drm->client.device;
+	int ret = 0;
+
+	/* Allocate channel that has access to a (preferably async) copy
+	 * engine, to use for TTM buffer moves.
+	 */
+	if (device->info.family >= NV_DEVICE_INFO_V0_KEPLER) {
+		ret = nouveau_channel_new(drm, device,
+					  nvif_fifo_runlist_ce(device), 0,
+					  true, &drm->cechan);
+	} else
+	if (device->info.chipset >= 0xa3 &&
+	    device->info.chipset != 0xaa &&
+	    device->info.chipset != 0xac) {
+		/* Prior to Kepler, there's only a single runlist, so all
+		 * engines can be accessed from any channel.
+		 *
+		 * We still want to use a separate channel though.
+		 */
+		ret = nouveau_channel_new(drm, device, NvDmaFB, NvDmaTT, false,
+					  &drm->cechan);
+	}
+
+	if (ret)
+		NV_ERROR(drm, "failed to create ce channel, %d\n", ret);
+}
+
+static void
+nouveau_accel_gr_fini(struct nouveau_drm *drm)
 {
 	nouveau_channel_idle(drm->channel);
 	nvif_object_fini(&drm->ntfy);
 	nvkm_gpuobj_del(&drm->notify);
 	nvif_object_fini(&drm->nvsw);
 	nouveau_channel_del(&drm->channel);
+}
 
-	nouveau_channel_idle(drm->cechan);
-	nvif_object_fini(&drm->ttm.copy);
-	nouveau_channel_del(&drm->cechan);
+static void
+nouveau_accel_gr_init(struct nouveau_drm *drm)
+{
+	struct nvif_device *device = &drm->client.device;
+	u32 arg0, arg1;
+	int ret;
+
+	/* Allocate channel that has access to the graphics engine. */
+	if (device->info.family >= NV_DEVICE_INFO_V0_KEPLER) {
+		arg0 = nvif_fifo_runlist(device, NV_DEVICE_INFO_ENGINE_GR);
+		arg1 = 1;
+	} else {
+		arg0 = NvDmaFB;
+		arg1 = NvDmaTT;
+	}
+
+	ret = nouveau_channel_new(drm, device, arg0, arg1, false,
+				  &drm->channel);
+	if (ret) {
+		NV_ERROR(drm, "failed to create kernel channel, %d\n", ret);
+		nouveau_accel_gr_fini(drm);
+		return;
+	}
+
+	/* A SW class is used on pre-NV50 HW to assist with handling the
+	 * synchronisation of page flips, as well as to implement fences
+	 * on TNT/TNT2 HW that lacks any kind of support in host.
+	 */
+	if (device->info.family < NV_DEVICE_INFO_V0_TESLA) {
+		ret = nvif_object_init(&drm->channel->user, NVDRM_NVSW,
+				       nouveau_abi16_swclass(drm), NULL, 0,
+				       &drm->nvsw);
+		if (ret == 0) {
+			ret = RING_SPACE(drm->channel, 2);
+			if (ret == 0) {
+				BEGIN_NV04(drm->channel, NvSubSw, 0, 1);
+				OUT_RING  (drm->channel, drm->nvsw.handle);
+			}
+		}
+
+		if (ret) {
+			NV_ERROR(drm, "failed to allocate sw class, %d\n", ret);
+			nouveau_accel_gr_fini(drm);
+			return;
+		}
+	}
+
+	/* NvMemoryToMemoryFormat requires a notifier ctxdma for some reason,
+	 * even if notification is never requested, so, allocate a ctxdma on
+	 * any GPU where it's possible we'll end up using M2MF for BO moves.
+	 */
+	if (device->info.family < NV_DEVICE_INFO_V0_FERMI) {
+		ret = nvkm_gpuobj_new(nvxx_device(device), 32, 0, false, NULL,
+				      &drm->notify);
+		if (ret) {
+			NV_ERROR(drm, "failed to allocate notifier, %d\n", ret);
+			nouveau_accel_gr_fini(drm);
+			return;
+		}
+
+		ret = nvif_object_init(&drm->channel->user, NvNotify0,
+				       NV_DMA_IN_MEMORY,
+				       &(struct nv_dma_v0) {
+						.target = NV_DMA_V0_TARGET_VRAM,
+						.access = NV_DMA_V0_ACCESS_RDWR,
+						.start = drm->notify->addr,
+						.limit = drm->notify->addr + 31
+				       }, sizeof(struct nv_dma_v0),
+				       &drm->ntfy);
+		if (ret) {
+			nouveau_accel_gr_fini(drm);
+			return;
+		}
+	}
+}
 
+static void
+nouveau_accel_fini(struct nouveau_drm *drm)
+{
+	nouveau_accel_ce_fini(drm);
+	nouveau_accel_gr_fini(drm);
 	if (drm->fence)
 		nouveau_fence(drm)->dtor(drm);
 }
@@ -303,23 +419,16 @@ nouveau_accel_init(struct nouveau_drm *drm)
 {
 	struct nvif_device *device = &drm->client.device;
 	struct nvif_sclass *sclass;
-	u32 arg0, arg1;
 	int ret, i, n;
 
 	if (nouveau_noaccel)
 		return;
 
+	/* Initialise global support for channels, and synchronisation. */
 	ret = nouveau_channels_init(drm);
 	if (ret)
 		return;
 
-	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_VOLTA) {
-		ret = nvif_user_init(device);
-		if (ret)
-			return;
-	}
-
-	/* initialise synchronisation routines */
 	/*XXX: this is crap, but the fence/channel stuff is a little
 	 *     backwards in some places.  this will be fixed.
 	 */
@@ -366,84 +475,18 @@ nouveau_accel_init(struct nouveau_drm *drm)
 		return;
 	}
 
-	if (device->info.family >= NV_DEVICE_INFO_V0_KEPLER) {
-		ret = nouveau_channel_new(drm, &drm->client.device,
-					  nvif_fifo_runlist_ce(device), 0,
-					  true, &drm->cechan);
-		if (ret)
-			NV_ERROR(drm, "failed to create ce channel, %d\n", ret);
-
-		arg0 = nvif_fifo_runlist(device, NV_DEVICE_INFO_ENGINE_GR);
-		arg1 = 1;
-	} else
-	if (device->info.chipset >= 0xa3 &&
-	    device->info.chipset != 0xaa &&
-	    device->info.chipset != 0xac) {
-		ret = nouveau_channel_new(drm, &drm->client.device,
-					  NvDmaFB, NvDmaTT, false,
-					  &drm->cechan);
+	/* Volta requires access to a doorbell register for kickoff. */
+	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_VOLTA) {
+		ret = nvif_user_init(device);
 		if (ret)
-			NV_ERROR(drm, "failed to create ce channel, %d\n", ret);
-
-		arg0 = NvDmaFB;
-		arg1 = NvDmaTT;
-	} else {
-		arg0 = NvDmaFB;
-		arg1 = NvDmaTT;
-	}
-
-	ret = nouveau_channel_new(drm, &drm->client.device,
-				  arg0, arg1, false, &drm->channel);
-	if (ret) {
-		NV_ERROR(drm, "failed to create kernel channel, %d\n", ret);
-		nouveau_accel_fini(drm);
-		return;
-	}
-
-	if (device->info.family < NV_DEVICE_INFO_V0_TESLA) {
-		ret = nvif_object_init(&drm->channel->user, NVDRM_NVSW,
-				       nouveau_abi16_swclass(drm), NULL, 0,
-				       &drm->nvsw);
-		if (ret == 0) {
-			ret = RING_SPACE(drm->channel, 2);
-			if (ret == 0) {
-				BEGIN_NV04(drm->channel, NvSubSw, 0, 1);
-				OUT_RING  (drm->channel, drm->nvsw.handle);
-			}
-		}
-
-		if (ret) {
-			NV_ERROR(drm, "failed to allocate sw class, %d\n", ret);
-			nouveau_accel_fini(drm);
-			return;
-		}
-	}
-
-	if (device->info.family < NV_DEVICE_INFO_V0_FERMI) {
-		ret = nvkm_gpuobj_new(nvxx_device(&drm->client.device), 32, 0,
-				      false, NULL, &drm->notify);
-		if (ret) {
-			NV_ERROR(drm, "failed to allocate notifier, %d\n", ret);
-			nouveau_accel_fini(drm);
-			return;
-		}
-
-		ret = nvif_object_init(&drm->channel->user, NvNotify0,
-				       NV_DMA_IN_MEMORY,
-				       &(struct nv_dma_v0) {
-						.target = NV_DMA_V0_TARGET_VRAM,
-						.access = NV_DMA_V0_ACCESS_RDWR,
-						.start = drm->notify->addr,
-						.limit = drm->notify->addr + 31
-				       }, sizeof(struct nv_dma_v0),
-				       &drm->ntfy);
-		if (ret) {
-			nouveau_accel_fini(drm);
 			return;
-		}
 	}
 
+	/* Allocate channels we need to support various functions. */
+	nouveau_accel_gr_init(drm);
+	nouveau_accel_ce_init(drm);
 
+	/* Initialise accelerated TTM buffer moves. */
 	nouveau_bo_move_init(drm);
 }