drm/nouveau/ce/gv100-: move method buffer to ce ctx

Didn't really know what this buffer was when initially implemented,
but these days we do, so move it somewhere more appropriate.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drivers/gpu/drm/nouveau/nouveau_abi16.c

@@ -147,6 +147,7 @@ nouveau_abi16_chan_fini(struct nouveau_abi16 *abi16,
 
 	/* destroy channel object, all children will be killed too */
 	if (chan->chan) {
+		nvif_object_dtor(&chan->ce);
 		nouveau_channel_idle(chan->chan);
 		nouveau_channel_del(&chan->chan);
 	}
@@ -325,6 +326,31 @@ nouveau_abi16_ioctl_channel_alloc(ABI16_IOCTL_ARGS)
 		init->nr_subchan = 2;
 	}
 
+	/* Workaround "nvc0" gallium driver using classes it doesn't allocate on
+	 * Kepler and above.  NVKM no longer always sets CE_CTX_VALID as part of
+	 * channel init, now we know what that stuff actually is.
+	 *
+	 * Doesn't matter for Kepler/Pascal, CE context stored in NV_RAMIN.
+	 *
+	 * Userspace was fixed prior to adding Ampere support.
+	 */
+	switch (device->info.family) {
+	case NV_DEVICE_INFO_V0_VOLTA:
+		ret = nvif_object_ctor(&chan->chan->user, "abi16CeWar", 0, VOLTA_DMA_COPY_A,
+				       NULL, 0, &chan->ce);
+		if (ret)
+			goto done;
+		break;
+	case NV_DEVICE_INFO_V0_TURING:
+		ret = nvif_object_ctor(&chan->chan->user, "abi16CeWar", 0, TURING_DMA_COPY_A,
+				       NULL, 0, &chan->ce);
+		if (ret)
+			goto done;
+		break;
+	default:
+		break;
+	}
+
 	/* Named memory object area */
 	ret = nouveau_gem_new(cli, PAGE_SIZE, 0, NOUVEAU_GEM_DOMAIN_GART,
 			      0, 0, &chan->ntfy);

drivers/gpu/drm/nouveau/nouveau_abi16.h

@@ -21,6 +21,7 @@ struct nouveau_abi16_ntfy {
 struct nouveau_abi16_chan {
 	struct list_head head;
 	struct nouveau_channel *chan;
+	struct nvif_object ce;
 	struct list_head notifiers;
 	struct nouveau_bo *ntfy;
 	struct nouveau_vma *ntfy_vma;

drivers/gpu/drm/nouveau/nvkm/engine/ce/gv100.c

@@ -21,11 +21,35 @@
  */
 #include "priv.h"
 
+#include <core/gpuobj.h>
+#include <core/object.h>
+
 #include <nvif/class.h>
 
+static int
+gv100_ce_cclass_bind(struct nvkm_object *object, struct nvkm_gpuobj *parent, int align,
+		     struct nvkm_gpuobj **pgpuobj)
+{
+	struct nvkm_device *device = object->engine->subdev.device;
+	u32 size;
+
+	/* Allocate fault method buffer (magics come from nvgpu). */
+	size = nvkm_rd32(device, 0x104028); /* NV_PCE_PCE_MAP */
+	size = 27 * 5 * (((9 + 1 + 3) * hweight32(size)) + 2);
+	size = roundup(size, PAGE_SIZE);
+
+	return nvkm_gpuobj_new(device, size, align, true, parent, pgpuobj);
+}
+
+const struct nvkm_object_func
+gv100_ce_cclass = {
+	.bind = gv100_ce_cclass_bind,
+};
+
 static const struct nvkm_engine_func
 gv100_ce = {
 	.intr = gp100_ce_intr,
+	.cclass = &gv100_ce_cclass,
 	.sclass = {
 		{ -1, -1, VOLTA_DMA_COPY_A },
 		{}

drivers/gpu/drm/nouveau/nvkm/engine/ce/priv.h

@@ -6,4 +6,6 @@
 void gt215_ce_intr(struct nvkm_falcon *, struct nvkm_fifo_chan *);
 void gk104_ce_intr(struct nvkm_engine *);
 void gp100_ce_intr(struct nvkm_engine *);
+
+extern const struct nvkm_object_func gv100_ce_cclass;
 #endif

drivers/gpu/drm/nouveau/nvkm/engine/ce/tu102.c

@@ -26,6 +26,7 @@
 static const struct nvkm_engine_func
 tu102_ce = {
 	.intr = gp100_ce_intr,
+	.cclass = &gv100_ce_cclass,
 	.sclass = {
 		{ -1, -1, TURING_DMA_COPY_A },
 		{}

drivers/gpu/drm/nouveau/nvkm/engine/fifo/changk104.h

@@ -14,8 +14,6 @@ struct gk104_fifo_chan {
 	struct list_head head;
 	bool killed;
 
-	struct nvkm_memory *mthd;
-
 #define GK104_FIFO_ENGN_SW 15
 	struct gk104_fifo_engn {
 		struct nvkm_gpuobj *inst;

drivers/gpu/drm/nouveau/nvkm/engine/fifo/gpfifogk104.c

@@ -175,13 +175,19 @@ gk104_fifo_gpfifo_engine_ctor(struct nvkm_fifo_chan *base,
 	struct gk104_fifo_engn *engn = gk104_fifo_gpfifo_engine(chan, engine);
 	int ret;
 
-	if (!gk104_fifo_gpfifo_engine_addr(engine))
+	if (!gk104_fifo_gpfifo_engine_addr(engine)) {
+		if (engine->subdev.type != NVKM_ENGINE_CE ||
+		    engine->subdev.device->card_type < GV100)
 			return 0;
+	}
 
 	ret = nvkm_object_bind(object, NULL, 0, &engn->inst);
 	if (ret)
 		return ret;
 
+	if (!gk104_fifo_gpfifo_engine_addr(engine))
+		return 0;
+
 	ret = nvkm_vmm_get(chan->base.vmm, 12, engn->inst->size, &engn->vma);
 	if (ret)
 		return ret;
@@ -231,7 +237,6 @@ void *
 gk104_fifo_gpfifo_dtor(struct nvkm_fifo_chan *base)
 {
 	struct gk104_fifo_chan *chan = gk104_fifo_chan(base);
-	nvkm_memory_unref(&chan->mthd);
 	kfree(chan->cgrp);
 	return chan;
 }

drivers/gpu/drm/nouveau/nvkm/engine/fifo/gpfifogv100.c

@@ -70,8 +70,17 @@ gv100_fifo_gpfifo_engine_fini(struct nvkm_fifo_chan *base,
 	struct nvkm_gpuobj *inst = chan->base.inst;
 	int ret;
 
-	if (engine->subdev.type == NVKM_ENGINE_CE)
-		return gk104_fifo_gpfifo_kick(chan);
+	if (engine->subdev.type == NVKM_ENGINE_CE) {
+		ret = gv100_fifo_gpfifo_engine_valid(chan, true, false);
+		if (ret && suspend)
+			return ret;
+
+		nvkm_kmap(inst);
+		nvkm_wo32(chan->base.inst, 0x220, 0x00000000);
+		nvkm_wo32(chan->base.inst, 0x224, 0x00000000);
+		nvkm_done(inst);
+		return ret;
+	}
 
 	ret = gv100_fifo_gpfifo_engine_valid(chan, false, false);
 	if (ret && suspend)
@@ -92,8 +101,16 @@ gv100_fifo_gpfifo_engine_init(struct nvkm_fifo_chan *base,
 	struct gk104_fifo_engn *engn = gk104_fifo_gpfifo_engine(chan, engine);
 	struct nvkm_gpuobj *inst = chan->base.inst;
 
-	if (engine->subdev.type == NVKM_ENGINE_CE)
-		return 0;
+	if (engine->subdev.type == NVKM_ENGINE_CE) {
+		const u64 bar2 = nvkm_memory_bar2(engn->inst->memory);
+
+		nvkm_kmap(inst);
+		nvkm_wo32(chan->base.inst, 0x220, lower_32_bits(bar2));
+		nvkm_wo32(chan->base.inst, 0x224, upper_32_bits(bar2));
+		nvkm_done(inst);
+
+		return gv100_fifo_gpfifo_engine_valid(chan, true, true);
+	}
 
 	nvkm_kmap(inst);
 	nvkm_wo32(inst, 0x210, lower_32_bits(engn->vma->addr) | 0x00000004);
@@ -123,11 +140,9 @@ gv100_fifo_gpfifo_new_(const struct nvkm_fifo_chan_func *func,
 		       u32 *token, const struct nvkm_oclass *oclass,
 		       struct nvkm_object **pobject)
 {
-	struct nvkm_device *device = fifo->base.engine.subdev.device;
 	struct gk104_fifo_chan *chan;
 	int runlist = ffs(*runlists) -1, ret, i;
-	u64 usermem, mthd;
-	u32 size;
+	u64 usermem;
 
 	if (!vmm || runlist < 0 || runlist >= fifo->runlist_nr)
 		return -EINVAL;
@@ -173,20 +188,6 @@ gv100_fifo_gpfifo_new_(const struct nvkm_fifo_chan_func *func,
 	nvkm_done(fifo->user.mem);
 	usermem = nvkm_memory_addr(fifo->user.mem) + usermem;
 
-	/* Allocate fault method buffer (magics come from nvgpu). */
-	size = nvkm_rd32(device, 0x104028); /* NV_PCE_PCE_MAP */
-	size = 27 * 5 * (((9 + 1 + 3) * hweight32(size)) + 2);
-	size = roundup(size, PAGE_SIZE);
-
-	ret = nvkm_memory_new(device, NVKM_MEM_TARGET_INST, size, 0x1000, true,
-			      &chan->mthd);
-	if (ret)
-		return ret;
-
-	mthd = nvkm_memory_bar2(chan->mthd);
-	if (mthd == ~0ULL)
-		return -EFAULT;
-
 	/* RAMFC */
 	nvkm_kmap(chan->base.inst);
 	nvkm_wo32(chan->base.inst, 0x008, lower_32_bits(usermem));
@@ -203,10 +204,8 @@ gv100_fifo_gpfifo_new_(const struct nvkm_fifo_chan_func *func,
 	nvkm_wo32(chan->base.inst, 0x0f4, 0x00001000);
 	nvkm_wo32(chan->base.inst, 0x0f8, 0x10003080);
 	nvkm_mo32(chan->base.inst, 0x218, 0x00000000, 0x00000000);
-	nvkm_wo32(chan->base.inst, 0x220, lower_32_bits(mthd));
-	nvkm_wo32(chan->base.inst, 0x224, upper_32_bits(mthd));
 	nvkm_done(chan->base.inst);
-	return gv100_fifo_gpfifo_engine_valid(chan, true, true);
+	return 0;
 }
 
 int