drm/nouveau: implement new VM_BIND uAPI

This commit provides the implementation for the new uapi motivated by the
Vulkan API. It allows user mode drivers (UMDs) to:

1) Initialize a GPU virtual address (VA) space via the new
   DRM_IOCTL_NOUVEAU_VM_INIT ioctl for UMDs to specify the portion of VA
   space managed by the kernel and userspace, respectively.

2) Allocate and free a VA space region as well as bind and unbind memory
   to the GPUs VA space via the new DRM_IOCTL_NOUVEAU_VM_BIND ioctl.
   UMDs can request the named operations to be processed either
   synchronously or asynchronously. It supports DRM syncobjs
   (incl. timelines) as synchronization mechanism. The management of the
   GPU VA mappings is implemented with the DRM GPU VA manager.

3) Execute push buffers with the new DRM_IOCTL_NOUVEAU_EXEC ioctl. The
   execution happens asynchronously. It supports DRM syncobj (incl.
   timelines) as synchronization mechanism. DRM GEM object locking is
   handled with drm_exec.

Both, DRM_IOCTL_NOUVEAU_VM_BIND and DRM_IOCTL_NOUVEAU_EXEC, use the DRM
GPU scheduler for the asynchronous paths.

Reviewed-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Danilo Krummrich <dakr@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230804182406.5222-12-dakr@redhat.com

19 files changed, 3321 insertions, 69 deletions

diff --git a/Documentation/gpu/driver-uapi.rst b/Documentation/gpu/driver-uapi.rst
index 9c7ca6e33a68..c08bcbb95fb3 100644
--- a/Documentation/gpu/driver-uapi.rst
+++ b/Documentation/gpu/driver-uapi.rst
@@ -13,4 +13,7 @@ drm/nouveau uAPI
 VM_BIND / EXEC uAPI
 -------------------
 
+.. kernel-doc:: drivers/gpu/drm/nouveau/nouveau_exec.c
+    :doc: Overview
+
 .. kernel-doc:: include/uapi/drm/nouveau_drm.h
diff --git a/drivers/gpu/drm/nouveau/Kbuild b/drivers/gpu/drm/nouveau/Kbuild
index 5e5617006da5..cf6b3a80c0c8 100644
--- a/drivers/gpu/drm/nouveau/Kbuild
+++ b/drivers/gpu/drm/nouveau/Kbuild
@@ -47,6 +47,9 @@ nouveau-y += nouveau_prime.o
 nouveau-y += nouveau_sgdma.o
 nouveau-y += nouveau_ttm.o
 nouveau-y += nouveau_vmm.o
+nouveau-y += nouveau_exec.o
+nouveau-y += nouveau_sched.o
+nouveau-y += nouveau_uvmm.o
 
 # DRM - modesetting
 nouveau-$(CONFIG_DRM_NOUVEAU_BACKLIGHT) += nouveau_backlight.o
diff --git a/drivers/gpu/drm/nouveau/Kconfig b/drivers/gpu/drm/nouveau/Kconfig
index a70bd65e1400..c52e8096cca4 100644
--- a/drivers/gpu/drm/nouveau/Kconfig
+++ b/drivers/gpu/drm/nouveau/Kconfig
@@ -10,6 +10,8 @@ config DRM_NOUVEAU
 	select DRM_KMS_HELPER
 	select DRM_TTM
 	select DRM_TTM_HELPER
+	select DRM_EXEC
+	select DRM_SCHED
 	select I2C
 	select I2C_ALGOBIT
 	select BACKLIGHT_CLASS_DEVICE if DRM_NOUVEAU_BACKLIGHT
diff --git a/drivers/gpu/drm/nouveau/nouveau_abi16.c b/drivers/gpu/drm/nouveau/nouveau_abi16.c
index 82dab51d8aeb..30afbec9e3b1 100644
--- a/drivers/gpu/drm/nouveau/nouveau_abi16.c
+++ b/drivers/gpu/drm/nouveau/nouveau_abi16.c
@@ -35,6 +35,7 @@
 #include "nouveau_chan.h"
 #include "nouveau_abi16.h"
 #include "nouveau_vmm.h"
+#include "nouveau_sched.h"
 
 static struct nouveau_abi16 *
 nouveau_abi16(struct drm_file *file_priv)
@@ -125,6 +126,17 @@ nouveau_abi16_chan_fini(struct nouveau_abi16 *abi16,
 {
 	struct nouveau_abi16_ntfy *ntfy, *temp;
 
+	/* When a client exits without waiting for it's queued up jobs to
+	 * finish it might happen that we fault the channel. This is due to
+	 * drm_file_free() calling drm_gem_release() before the postclose()
+	 * callback. Hence, we can't tear down this scheduler entity before
+	 * uvmm mappings are unmapped. Currently, we can't detect this case.
+	 *
+	 * However, this should be rare and harmless, since the channel isn't
+	 * needed anymore.
+	 */
+	nouveau_sched_entity_fini(&chan->sched_entity);
+
 	/* wait for all activity to stop before cleaning up */
 	if (chan->chan)
 		nouveau_channel_idle(chan->chan);
@@ -261,6 +273,13 @@ nouveau_abi16_ioctl_channel_alloc(ABI16_IOCTL_ARGS)
 	if (!drm->channel)
 		return nouveau_abi16_put(abi16, -ENODEV);
 
+	/* If uvmm wasn't initialized until now disable it completely to prevent
+	 * userspace from mixing up UAPIs.
+	 *
+	 * The client lock is already acquired by nouveau_abi16_get().
+	 */
+	__nouveau_cli_disable_uvmm_noinit(cli);
+
 	device = &abi16->device;
 	engine = NV_DEVICE_HOST_RUNLIST_ENGINES_GR;
 
@@ -304,6 +323,11 @@ nouveau_abi16_ioctl_channel_alloc(ABI16_IOCTL_ARGS)
 	if (ret)
 		goto done;
 
+	ret = nouveau_sched_entity_init(&chan->sched_entity, &drm->sched,
+					drm->sched_wq);
+	if (ret)
+		goto done;
+
 	init->channel = chan->chan->chid;
 
 	if (device->info.family >= NV_DEVICE_INFO_V0_TESLA)
diff --git a/drivers/gpu/drm/nouveau/nouveau_abi16.h b/drivers/gpu/drm/nouveau/nouveau_abi16.h
index d5d80d0d9011..9f538486c10e 100644
--- a/drivers/gpu/drm/nouveau/nouveau_abi16.h
+++ b/drivers/gpu/drm/nouveau/nouveau_abi16.h
@@ -26,6 +26,7 @@ struct nouveau_abi16_chan {
 	struct nouveau_bo *ntfy;
 	struct nouveau_vma *ntfy_vma;
 	struct nvkm_mm  heap;
+	struct nouveau_sched_entity sched_entity;
 };
 
 struct nouveau_abi16 {
diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
index e38e448d9632..19cab37ac69c 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bo.c
+++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
@@ -199,7 +199,7 @@ nouveau_bo_fixup_align(struct nouveau_bo *nvbo, int *align, u64 *size)
 
 struct nouveau_bo *
 nouveau_bo_alloc(struct nouveau_cli *cli, u64 *size, int *align, u32 domain,
-		 u32 tile_mode, u32 tile_flags)
+		 u32 tile_mode, u32 tile_flags, bool internal)
 {
 	struct nouveau_drm *drm = cli->drm;
 	struct nouveau_bo *nvbo;
@@ -233,68 +233,103 @@ nouveau_bo_alloc(struct nouveau_cli *cli, u64 *size, int *align, u32 domain,
 			nvbo->force_coherent = true;
 	}
 
-	if (cli->device.info.family >= NV_DEVICE_INFO_V0_FERMI) {
-		nvbo->kind = (tile_flags & 0x0000ff00) >> 8;
-		if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
-			kfree(nvbo);
-			return ERR_PTR(-EINVAL);
+	nvbo->contig = !(tile_flags & NOUVEAU_GEM_TILE_NONCONTIG);
+	if (!nouveau_cli_uvmm(cli) || internal) {
+		/* for BO noVM allocs, don't assign kinds */
+		if (cli->device.info.family >= NV_DEVICE_INFO_V0_FERMI) {
+			nvbo->kind = (tile_flags & 0x0000ff00) >> 8;
+			if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
+				kfree(nvbo);
+				return ERR_PTR(-EINVAL);
+			}
+
+			nvbo->comp = mmu->kind[nvbo->kind] != nvbo->kind;
+		} else if (cli->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
+			nvbo->kind = (tile_flags & 0x00007f00) >> 8;
+			nvbo->comp = (tile_flags & 0x00030000) >> 16;
+			if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
+				kfree(nvbo);
+				return ERR_PTR(-EINVAL);
+			}
+		} else {
+			nvbo->zeta = (tile_flags & 0x00000007);
 		}
+		nvbo->mode = tile_mode;
+
+		/* Determine the desirable target GPU page size for the buffer. */
+		for (i = 0; i < vmm->page_nr; i++) {
+			/* Because we cannot currently allow VMM maps to fail
+			 * during buffer migration, we need to determine page
+			 * size for the buffer up-front, and pre-allocate its
+			 * page tables.
+			 *
+			 * Skip page sizes that can't support needed domains.
+			 */
+			if (cli->device.info.family > NV_DEVICE_INFO_V0_CURIE &&
+			    (domain & NOUVEAU_GEM_DOMAIN_VRAM) && !vmm->page[i].vram)
+				continue;
+			if ((domain & NOUVEAU_GEM_DOMAIN_GART) &&
+			    (!vmm->page[i].host || vmm->page[i].shift > PAGE_SHIFT))
+				continue;
 
-		nvbo->comp = mmu->kind[nvbo->kind] != nvbo->kind;
-	} else
-	if (cli->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
-		nvbo->kind = (tile_flags & 0x00007f00) >> 8;
-		nvbo->comp = (tile_flags & 0x00030000) >> 16;
-		if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
+			/* Select this page size if it's the first that supports
+			 * the potential memory domains, or when it's compatible
+			 * with the requested compression settings.
+			 */
+			if (pi < 0 || !nvbo->comp || vmm->page[i].comp)
+				pi = i;
+
+			/* Stop once the buffer is larger than the current page size. */
+			if (*size >= 1ULL << vmm->page[i].shift)
+				break;
+		}
+
+		if (WARN_ON(pi < 0)) {
 			kfree(nvbo);
 			return ERR_PTR(-EINVAL);
 		}
-	} else {
-		nvbo->zeta = (tile_flags & 0x00000007);
-	}
-	nvbo->mode = tile_mode;
-	nvbo->contig = !(tile_flags & NOUVEAU_GEM_TILE_NONCONTIG);
-
-	/* Determine the desirable target GPU page size for the buffer. */
-	for (i = 0; i < vmm->page_nr; i++) {
-		/* Because we cannot currently allow VMM maps to fail
-		 * during buffer migration, we need to determine page
-		 * size for the buffer up-front, and pre-allocate its
-		 * page tables.
-		 *
-		 * Skip page sizes that can't support needed domains.
-		 */
-		if (cli->device.info.family > NV_DEVICE_INFO_V0_CURIE &&
-		    (domain & NOUVEAU_GEM_DOMAIN_VRAM) && !vmm->page[i].vram)
-			continue;
-		if ((domain & NOUVEAU_GEM_DOMAIN_GART) &&
-		    (!vmm->page[i].host || vmm->page[i].shift > PAGE_SHIFT))
-			continue;
-
-		/* Select this page size if it's the first that supports
-		 * the potential memory domains, or when it's compatible
-		 * with the requested compression settings.
-		 */
-		if (pi < 0 || !nvbo->comp || vmm->page[i].comp)
-			pi = i;
 
-		/* Stop once the buffer is larger than the current page size. */
-		if (*size >= 1ULL << vmm->page[i].shift)
-			break;
-	}
+		/* Disable compression if suitable settings couldn't be found. */
+		if (nvbo->comp && !vmm->page[pi].comp) {
+			if (mmu->object.oclass >= NVIF_CLASS_MMU_GF100)
+				nvbo->kind = mmu->kind[nvbo->kind];
+			nvbo->comp = 0;
+		}
+		nvbo->page = vmm->page[pi].shift;
+	} else {
+		/* reject other tile flags when in VM mode. */
+		if (tile_mode)
+			return ERR_PTR(-EINVAL);
+		if (tile_flags & ~NOUVEAU_GEM_TILE_NONCONTIG)
+			return ERR_PTR(-EINVAL);
 
-	if (WARN_ON(pi < 0)) {
-		kfree(nvbo);
-		return ERR_PTR(-EINVAL);
-	}
+		/* Determine the desirable target GPU page size for the buffer. */
+		for (i = 0; i < vmm->page_nr; i++) {
+			/* Because we cannot currently allow VMM maps to fail
+			 * during buffer migration, we need to determine page
+			 * size for the buffer up-front, and pre-allocate its
+			 * page tables.
+			 *
+			 * Skip page sizes that can't support needed domains.
+			 */
+			if ((domain & NOUVEAU_GEM_DOMAIN_VRAM) && !vmm->page[i].vram)
+				continue;
+			if ((domain & NOUVEAU_GEM_DOMAIN_GART) &&
+			    (!vmm->page[i].host || vmm->page[i].shift > PAGE_SHIFT))
+				continue;
 
-	/* Disable compression if suitable settings couldn't be found. */
-	if (nvbo->comp && !vmm->page[pi].comp) {
-		if (mmu->object.oclass >= NVIF_CLASS_MMU_GF100)
-			nvbo->kind = mmu->kind[nvbo->kind];
-		nvbo->comp = 0;
+			if (pi < 0)
+				pi = i;
+			/* Stop once the buffer is larger than the current page size. */
+			if (*size >= 1ULL << vmm->page[i].shift)
+				break;
+		}
+		if (WARN_ON(pi < 0)) {
+			kfree(nvbo);
+			return ERR_PTR(-EINVAL);
+		}
+		nvbo->page = vmm->page[pi].shift;
 	}
-	nvbo->page = vmm->page[pi].shift;
 
 	nouveau_bo_fixup_align(nvbo, align, size);
 
@@ -307,18 +342,26 @@ nouveau_bo_init(struct nouveau_bo *nvbo, u64 size, int align, u32 domain,
 {
 	int type = sg ? ttm_bo_type_sg : ttm_bo_type_device;
 	int ret;
+	struct ttm_operation_ctx ctx = {
+		.interruptible = false,
+		.no_wait_gpu = false,
+		.resv = robj,
+	};
 
 	nouveau_bo_placement_set(nvbo, domain, 0);
 	INIT_LIST_HEAD(&nvbo->io_reserve_lru);
 
-	ret = ttm_bo_init_validate(nvbo->bo.bdev, &nvbo->bo, type,
-				   &nvbo->placement, align >> PAGE_SHIFT, false,
+	ret = ttm_bo_init_reserved(nvbo->bo.bdev, &nvbo->bo, type,
+				   &nvbo->placement, align >> PAGE_SHIFT, &ctx,
 				   sg, robj, nouveau_bo_del_ttm);
 	if (ret) {
 		/* ttm will call nouveau_bo_del_ttm if it fails.. */
 		return ret;
 	}
 
+	if (!robj)
+		ttm_bo_unreserve(&nvbo->bo);
+
 	return 0;
 }
 
@@ -332,7 +375,7 @@ nouveau_bo_new(struct nouveau_cli *cli, u64 size, int align,
 	int ret;
 
 	nvbo = nouveau_bo_alloc(cli, &size, &align, domain, tile_mode,
-				tile_flags);
+				tile_flags, true);
 	if (IS_ERR(nvbo))
 		return PTR_ERR(nvbo);
 
@@ -951,6 +994,7 @@ static void nouveau_bo_move_ntfy(struct ttm_buffer_object *bo,
 		list_for_each_entry(vma, &nvbo->vma_list, head) {
 			nouveau_vma_map(vma, mem);
 		}
+		nouveau_uvmm_bo_map_all(nvbo, mem);
 	} else {
 		list_for_each_entry(vma, &nvbo->vma_list, head) {
 			ret = dma_resv_wait_timeout(bo->base.resv,
@@ -959,6 +1003,7 @@ static void nouveau_bo_move_ntfy(struct ttm_buffer_object *bo,
 			WARN_ON(ret <= 0);
 			nouveau_vma_unmap(vma);
 		}
+		nouveau_uvmm_bo_unmap_all(nvbo);
 	}
 
 	if (new_reg)
diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.h b/drivers/gpu/drm/nouveau/nouveau_bo.h
index 774dd93ca76b..07f671cf895e 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bo.h
+++ b/drivers/gpu/drm/nouveau/nouveau_bo.h
@@ -26,6 +26,7 @@ struct nouveau_bo {
 	struct list_head entry;
 	int pbbo_index;
 	bool validate_mapped;
+	bool no_share;
 
 	/* GPU address space is independent of CPU word size */
 	uint64_t offset;
@@ -73,7 +74,7 @@ extern struct ttm_device_funcs nouveau_bo_driver;
 
 void nouveau_bo_move_init(struct nouveau_drm *);
 struct nouveau_bo *nouveau_bo_alloc(struct nouveau_cli *, u64 *size, int *align,
-				    u32 domain, u32 tile_mode, u32 tile_flags);
+				    u32 domain, u32 tile_mode, u32 tile_flags, bool internal);
 int  nouveau_bo_init(struct nouveau_bo *, u64 size, int align, u32 domain,
 		     struct sg_table *sg, struct dma_resv *robj);
 int  nouveau_bo_new(struct nouveau_cli *, u64 size, int align, u32 domain,
diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c b/drivers/gpu/drm/nouveau/nouveau_drm.c
index 8325fcf35c5e..4396f501b16a 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drm.c
+++ b/drivers/gpu/drm/nouveau/nouveau_drm.c
@@ -68,6 +68,9 @@
 #include "nouveau_platform.h"
 #include "nouveau_svm.h"
 #include "nouveau_dmem.h"
+#include "nouveau_exec.h"
+#include "nouveau_uvmm.h"
+#include "nouveau_sched.h"
 
 DECLARE_DYNDBG_CLASSMAP(drm_debug_classes, DD_CLASS_TYPE_DISJOINT_BITS, 0,
 			"DRM_UT_CORE",
@@ -196,6 +199,8 @@ nouveau_cli_fini(struct nouveau_cli *cli)
 	WARN_ON(!list_empty(&cli->worker));
 
 	usif_client_fini(cli);
+	nouveau_uvmm_fini(&cli->uvmm);
+	nouveau_sched_entity_fini(&cli->sched_entity);
 	nouveau_vmm_fini(&cli->svm);
 	nouveau_vmm_fini(&cli->vmm);
 	nvif_mmu_dtor(&cli->mmu);
@@ -301,6 +306,12 @@ nouveau_cli_init(struct nouveau_drm *drm, const char *sname,
 	}
 
 	cli->mem = &mems[ret];
+
+	ret = nouveau_sched_entity_init(&cli->sched_entity, &drm->sched,
+					drm->sched_wq);
+	if (ret)
+		goto done;
+
 	return 0;
 done:
 	if (ret)
@@ -568,10 +579,14 @@ nouveau_drm_device_init(struct drm_device *dev)
 	nvif_parent_ctor(&nouveau_parent, &drm->parent);
 	drm->master.base.object.parent = &drm->parent;
 
-	ret = nouveau_cli_init(drm, "DRM-master", &drm->master);
+	ret = nouveau_sched_init(drm);
 	if (ret)
 		goto fail_alloc;
 
+	ret = nouveau_cli_init(drm, "DRM-master", &drm->master);
+	if (ret)
+		goto fail_sched;
+
 	ret = nouveau_cli_init(drm, "DRM", &drm->client);
 	if (ret)
 		goto fail_master;
@@ -628,7 +643,6 @@ nouveau_drm_device_init(struct drm_device *dev)
 	}
 
 	return 0;
-
 fail_dispinit:
 	nouveau_display_destroy(dev);
 fail_dispctor:
@@ -641,6 +655,8 @@ fail_ttm:
 	nouveau_cli_fini(&drm->client);
 fail_master:
 	nouveau_cli_fini(&drm->master);
+fail_sched:
+	nouveau_sched_fini(drm);
 fail_alloc:
 	nvif_parent_dtor(&drm->parent);
 	kfree(drm);
@@ -692,6 +708,8 @@ nouveau_drm_device_fini(struct drm_device *dev)
 	}
 	mutex_unlock(&drm->clients_lock);
 
+	nouveau_sched_fini(drm);
+
 	nouveau_cli_fini(&drm->client);
 	nouveau_cli_fini(&drm->master);
 	nvif_parent_dtor(&drm->parent);
@@ -1193,6 +1211,9 @@ nouveau_ioctls[] = {
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_CPU_PREP, nouveau_gem_ioctl_cpu_prep, DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_CPU_FINI, nouveau_gem_ioctl_cpu_fini, DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_INFO, nouveau_gem_ioctl_info, DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(NOUVEAU_VM_INIT, nouveau_uvmm_ioctl_vm_init, DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(NOUVEAU_VM_BIND, nouveau_uvmm_ioctl_vm_bind, DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(NOUVEAU_EXEC, nouveau_exec_ioctl_exec, DRM_RENDER_ALLOW),
 };
 
 long
@@ -1240,6 +1261,8 @@ nouveau_driver_fops = {
 static struct drm_driver
 driver_stub = {
 	.driver_features = DRIVER_GEM |
+			   DRIVER_SYNCOBJ | DRIVER_SYNCOBJ_TIMELINE |
+			   DRIVER_GEM_GPUVA |
 			   DRIVER_MODESET |
 			   DRIVER_RENDER,
 	.open = nouveau_drm_open,
diff --git a/drivers/gpu/drm/nouveau/nouveau_drv.h b/drivers/gpu/drm/nouveau/nouveau_drv.h
index d28236021971..1fe17ff95f5e 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drv.h
+++ b/drivers/gpu/drm/nouveau/nouveau_drv.h
@@ -10,8 +10,8 @@
 #define DRIVER_DATE		"20120801"
 
 #define DRIVER_MAJOR		1
-#define DRIVER_MINOR		3
-#define DRIVER_PATCHLEVEL	1
+#define DRIVER_MINOR		4
+#define DRIVER_PATCHLEVEL	0
 
 /*
  * 1.1.1:
@@ -63,7 +63,9 @@ struct platform_device;
 
 #include "nouveau_fence.h"
 #include "nouveau_bios.h"
+#include "nouveau_sched.h"
 #include "nouveau_vmm.h"
+#include "nouveau_uvmm.h"
 
 struct nouveau_drm_tile {
 	struct nouveau_fence *fence;
@@ -91,6 +93,10 @@ struct nouveau_cli {
 	struct nvif_mmu mmu;
 	struct nouveau_vmm vmm;
 	struct nouveau_vmm svm;
+	struct nouveau_uvmm uvmm;
+
+	struct nouveau_sched_entity sched_entity;
+
 	const struct nvif_mclass *mem;
 
 	struct list_head head;
@@ -112,15 +118,59 @@ struct nouveau_cli_work {
 	struct dma_fence_cb cb;
 };
 
+static inline struct nouveau_uvmm *
+nouveau_cli_uvmm(struct nouveau_cli *cli)
+{
+	if (!cli || !cli->uvmm.vmm.cli)
+		return NULL;
+
+	return &cli->uvmm;
+}
+
+static inline struct nouveau_uvmm *
+nouveau_cli_uvmm_locked(struct nouveau_cli *cli)
+{
+	struct nouveau_uvmm *uvmm;
+
+	mutex_lock(&cli->mutex);
+	uvmm = nouveau_cli_uvmm(cli);
+	mutex_unlock(&cli->mutex);
+
+	return uvmm;
+}
+
 static inline struct nouveau_vmm *
 nouveau_cli_vmm(struct nouveau_cli *cli)
 {
+	struct nouveau_uvmm *uvmm;
+
+	uvmm = nouveau_cli_uvmm(cli);
+	if (uvmm)
+		return &uvmm->vmm;
+
 	if (cli->svm.cli)
 		return &cli->svm;
 
 	return &cli->vmm;
 }
 
+static inline void
+__nouveau_cli_disable_uvmm_noinit(struct nouveau_cli *cli)
+{
+	struct nouveau_uvmm *uvmm = nouveau_cli_uvmm(cli);
+
+	if (!uvmm)
+		cli->uvmm.disabled = true;
+}
+
+static inline void
+nouveau_cli_disable_uvmm_noinit(struct nouveau_cli *cli)
+{
+	mutex_lock(&cli->mutex);
+	__nouveau_cli_disable_uvmm_noinit(cli);
+	mutex_unlock(&cli->mutex);
+}
+
 void nouveau_cli_work_queue(struct nouveau_cli *, struct dma_fence *,
 			    struct nouveau_cli_work *);
 
@@ -257,6 +307,10 @@ struct nouveau_drm {
 		struct mutex lock;
 		bool component_registered;
 	} audio;
+
+	struct drm_gpu_scheduler sched;
+	struct workqueue_struct *sched_wq;
+
 };
 
 static inline struct nouveau_drm *
diff --git a/drivers/gpu/drm/nouveau/nouveau_exec.c b/drivers/gpu/drm/nouveau/nouveau_exec.c
new file mode 100644
index 000000000000..42d9dd43ea02
--- /dev/null
+++ b/drivers/gpu/drm/nouveau/nouveau_exec.c
@@ -0,0 +1,411 @@
+// SPDX-License-Identifier: MIT
+
+#include <drm/drm_exec.h>
+
+#include "nouveau_drv.h"
+#include "nouveau_gem.h"
+#include "nouveau_mem.h"
+#include "nouveau_dma.h"
+#include "nouveau_exec.h"
+#include "nouveau_abi16.h"
+#include "nouveau_chan.h"
+#include "nouveau_sched.h"
+#include "nouveau_uvmm.h"
+
+/**
+ * DOC: Overview
+ *
+ * Nouveau's VM_BIND / EXEC UAPI consists of three ioctls: DRM_NOUVEAU_VM_INIT,
+ * DRM_NOUVEAU_VM_BIND and DRM_NOUVEAU_EXEC.
+ *
+ * In order to use the UAPI firstly a user client must initialize the VA space
+ * using the DRM_NOUVEAU_VM_INIT ioctl specifying which region of the VA space
+ * should be managed by the kernel and which by the UMD.
+ *
+ * The DRM_NOUVEAU_VM_BIND ioctl provides clients an interface to manage the
+ * userspace-managable portion of the VA space. It provides operations to map
+ * and unmap memory. Mappings may be flagged as sparse. Sparse mappings are not
+ * backed by a GEM object and the kernel will ignore GEM handles provided
+ * alongside a sparse mapping.
+ *
+ * Userspace may request memory backed mappings either within or outside of the
+ * bounds (but not crossing those bounds) of a previously mapped sparse
+ * mapping. Subsequently requested memory backed mappings within a sparse
+ * mapping will take precedence over the corresponding range of the sparse
+ * mapping. If such memory backed mappings are unmapped the kernel will make
+ * sure that the corresponding sparse mapping will take their place again.
+ * Requests to unmap a sparse mapping that still contains memory backed mappings
+ * will result in those memory backed mappings being unmapped first.
+ *
+ * Unmap requests are not bound to the range of existing mappings and can even
+ * overlap the bounds of sparse mappings. For such a request the kernel will
+ * make sure to unmap all memory backed mappings within the given range,
+ * splitting up memory backed mappings which are only partially contained
+ * within the given range. Unmap requests with the sparse flag set must match
+ * the range of a previously mapped sparse mapping exactly though.
+ *
+ * While the kernel generally permits arbitrary sequences and ranges of memory
+ * backed mappings being mapped and unmapped, either within a single or multiple
+ * VM_BIND ioctl calls, there are some restrictions for sparse mappings.
+ *
+ * The kernel does not permit to:
+ *   - unmap non-existent sparse mappings
+ *   - unmap a sparse mapping and map a new sparse mapping overlapping the range
+ *     of the previously unmapped sparse mapping within the same VM_BIND ioctl
+ *   - unmap a sparse mapping and map new memory backed mappings overlapping the
+ *     range of the previously unmapped sparse mapping within the same VM_BIND
+ *     ioctl
+ *
+ * When using the VM_BIND ioctl to request the kernel to map memory to a given
+ * virtual address in the GPU's VA space there is no guarantee that the actual
+ * mappings are created in the GPU's MMU. If the given memory is swapped out
+ * at the time the bind operation is executed the kernel will stash the mapping
+ * details into it's internal alloctor and create the actual MMU mappings once
+ * the memory is swapped back in. While this is transparent for userspace, it is
+ * guaranteed that all the backing memory is swapped back in and all the memory
+ * mappings, as requested by userspace previously, are actually mapped once the
+ * DRM_NOUVEAU_EXEC ioctl is called to submit an exec job.
+ *
+ * A VM_BIND job can be executed either synchronously or asynchronously. If
+ * exectued asynchronously, userspace may provide a list of syncobjs this job
+ * will wait for and/or a list of syncobj the kernel will signal once the
+ * VM_BIND job finished execution. If executed synchronously the ioctl will
+ * block until the bind job is finished. For synchronous jobs the kernel will
+ * not permit any syncobjs submitted to the kernel.
+ *
+ * To execute a push buffer the UAPI provides the DRM_NOUVEAU_EXEC ioctl. EXEC
+ * jobs are always executed asynchronously, and, equal to VM_BIND jobs, provide
+ * the option to synchronize them with syncobjs.
+ *
+ * Besides that, EXEC jobs can be scheduled for a specified channel to execute on.
+ *
+ * Since VM_BIND jobs update the GPU's VA space on job submit, EXEC jobs do have
+ * an up to date view of the VA space. However, the actual mappings might still
+ * be pending. Hence, EXEC jobs require to have the particular fences - of
+ * the corresponding VM_BIND jobs they depent on - attached to them.
+ */
+
+static int
+nouveau_exec_job_submit(struct nouveau_job *job)
+{
+	struct nouveau_exec_job *exec_job = to_nouveau_exec_job(job);
+	struct nouveau_cli *cli = job->cli;
+	struct nouveau_uvmm *uvmm = nouveau_cli_uvmm(cli);
+	struct drm_exec *exec = &job->exec;
+	struct drm_gem_object *obj;
+	unsigned long index;
+	int ret;
+
+	ret = nouveau_fence_new(&exec_job->fence);
+	if (ret)
+		return ret;
+
+	nouveau_uvmm_lock(uvmm);
+	drm_exec_init(exec, DRM_EXEC_INTERRUPTIBLE_WAIT |
+			    DRM_EXEC_IGNORE_DUPLICATES);
+	drm_exec_until_all_locked(exec) {
+		struct drm_gpuva *va;
+
+		drm_gpuva_for_each_va(va, &uvmm->umgr) {
+			if (unlikely(va == &uvmm->umgr.kernel_alloc_node))
+				continue;
+
+			ret = drm_exec_prepare_obj(exec, va->gem.obj, 1);
+			drm_exec_retry_on_contention(exec);
+			if (ret)
+				goto err_uvmm_unlock;
+		}
+	}
+	nouveau_uvmm_unlock(uvmm);
+
+	drm_exec_for_each_locked_object(exec, index, obj) {
+		struct nouveau_bo *nvbo = nouveau_gem_object(obj);
+
+		ret = nouveau_bo_validate(nvbo, true, false);
+		if (ret)
+			goto err_exec_fini;
+	}
+
+	return 0;
+
+err_uvmm_unlock:
+	nouveau_uvmm_unlock(uvmm);
+err_exec_fini:
+	drm_exec_fini(exec);
+	return ret;
+
+}
+
+static void
+nouveau_exec_job_armed_submit(struct nouveau_job *job)
+{
+	struct drm_exec *exec = &job->exec;
+	struct drm_gem_object *obj;
+	unsigned long index;
+
+	drm_exec_for_each_locked_object(exec, index, obj)
+		dma_resv_add_fence(obj->resv, job->done_fence, job->resv_usage);
+
+	drm_exec_fini(exec);
+}
+
+static struct dma_fence *
+nouveau_exec_job_run(struct nouveau_job *job)
+{
+	struct nouveau_exec_job *exec_job = to_nouveau_exec_job(job);
+	struct nouveau_channel *chan = exec_job->chan;
+	struct nouveau_fence *fence = exec_job->fence;
+	int i, ret;
+
+	ret = nouveau_dma_wait(chan, exec_job->push.count + 1, 16);
+	if (ret) {
+		NV_PRINTK(err, job->cli, "nv50cal_space: %d\n", ret);
+		return ERR_PTR(ret);
+	}
+
+	for (i = 0; i < exec_job->push.count; i++) {
+		nv50_dma_push(chan, exec_job->push.s[i].va,
+			      exec_job->push.s[i].va_len);
+	}
+
+	ret = nouveau_fence_emit(fence, chan);
+	if (ret) {
+		NV_PRINTK(err, job->cli, "error fencing pushbuf: %d\n", ret);
+		WIND_RING(chan);
+		return ERR_PTR(ret);
+	}
+
+	exec_job->fence = NULL;
+
+	return &fence->base;
+}
+
+static void
+nouveau_exec_job_free(struct nouveau_job *job)
+{
+	struct nouveau_exec_job *exec_job = to_nouveau_exec_job(job);
+
+	nouveau_job_free(job);
+
+	nouveau_fence_unref(&exec_job->fence);
+	kfree(exec_job->push.s);
+	kfree(exec_job);
+}
+
+static enum drm_gpu_sched_stat
+nouveau_exec_job_timeout(struct nouveau_job *job)
+{
+	struct nouveau_exec_job *exec_job = to_nouveau_exec_job(job);
+	struct nouveau_channel *chan = exec_job->chan;
+
+	if (unlikely(!atomic_read(&chan->killed)))
+		nouveau_channel_kill(chan);
+
+	NV_PRINTK(warn, job->cli, "job timeout, channel %d killed!\n",
+		  chan->chid);
+
+	nouveau_sched_entity_fini(job->entity);
+
+	return DRM_GPU_SCHED_STAT_ENODEV;
+}
+
+static struct nouveau_job_ops nouveau_exec_job_ops = {
+	.submit = nouveau_exec_job_submit,
+	.armed_submit = nouveau_exec_job_armed_submit,
+	.run = nouveau_exec_job_run,
+	.free = nouveau_exec_job_free,
+	.timeout = nouveau_exec_job_timeout,
+};
+
+int
+nouveau_exec_job_init(struct nouveau_exec_job **pjob,
+		      struct nouveau_exec_job_args *__args)
+{
+	struct nouveau_exec_job *job;
+	struct nouveau_job_args args = {};
+	int ret;
+
+	job = *pjob = kzalloc(sizeof(*job), GFP_KERNEL);
+	if (!job)
+		return -ENOMEM;
+
+	job->push.count = __args->push.count;
+	if (__args->push.count) {
+		job->push.s = kmemdup(__args->push.s,
+				      sizeof(*__args->push.s) *
+				      __args->push.count,
+				      GFP_KERNEL);
+		if (!job->push.s) {
+			ret = -ENOMEM;
+			goto err_free_job;
+		}
+	}
+
+	job->chan = __args->chan;
+
+	args.sched_entity = __args->sched_entity;
+	args.file_priv = __args->file_priv;
+
+	args.in_sync.count = __args->in_sync.count;
+	args.in_sync.s = __args->in_sync.s;
+
+	args.out_sync.count = __args->out_sync.count;
+	args.out_sync.s = __args->out_sync.s;
+
+	args.ops = &nouveau_exec_job_ops;
+	args.resv_usage = DMA_RESV_USAGE_WRITE;
+
+	ret = nouveau_job_init(&job->base, &args);
+	if (ret)
+		goto err_free_pushs;
+
+	return 0;
+
+err_free_pushs:
+	kfree(job->push.s);
+err_free_job:
+	kfree(job);
+	*pjob = NULL;
+
+	return ret;
+}
+
+static int
+nouveau_exec(struct nouveau_exec_job_args *args)
+{
+	struct nouveau_exec_job *job;
+	int ret;
+
+	ret = nouveau_exec_job_init(&job, args);
+	if (ret)
+		return ret;
+
+	ret = nouveau_job_submit(&job->base);
+	if (ret)
+		goto err_job_fini;
+
+	return 0;
+
+err_job_fini:
+	nouveau_job_fini(&job->base);
+	return ret;
+}
+
+static int
+nouveau_exec_ucopy(struct nouveau_exec_job_args *args,
+		   struct drm_nouveau_exec __user *req)
+{
+	struct drm_nouveau_sync **s;
+	u32 inc = req->wait_count;
+	u64 ins = req->wait_ptr;
+	u32 outc = req->sig_count;
+	u64 outs = req->sig_ptr;
+	u32 pushc = req->push_count;
+	u64 pushs = req->push_ptr;
+	int ret;
+
+	if (pushc) {
+		args->push.count = pushc;
+		args->push.s = u_memcpya(pushs, pushc, sizeof(*args->push.s));