Merge branch 'linux-5.1' of git://github.com/skeggsb/linux into drm-next

Various fixes/cleanups, along with initial support for SVM features
utilising HMM address-space mirroring and device memory migration.
There's a lot more work to do in these areas, both in terms of
features and efficiency, but these can slowly trickle in later down
the track.

Signed-off-by: Dave Airlie <[e-mail]>
From: Ben Skeggs <[e-mail]>
Link: https://patchwork.freedesktop.org/patch/msgid/[e-mail]

drivers/gpu/drm/nouveau/Kbuild

···
 # DRM - memory management
 nouveau-y += nouveau_bo.o
 nouveau-y += nouveau_gem.o
+nouveau-$(CONFIG_DRM_NOUVEAU_SVM) += nouveau_svm.o
+nouveau-$(CONFIG_DRM_NOUVEAU_SVM) += nouveau_dmem.o
 nouveau-y += nouveau_mem.o
 nouveau-y += nouveau_prime.o
 nouveau-y += nouveau_sgdma.o

drivers/gpu/drm/nouveau/Kconfig

···
 	help
 	  Say Y here if you want to control the backlight of your display
 	  (e.g. a laptop panel).
+
+config DRM_NOUVEAU_SVM
+	bool "(EXPERIMENTAL) Enable SVM (Shared Virtual Memory) support"
+	depends on ARCH_HAS_HMM
+	depends on DRM_NOUVEAU
+	depends on STAGING
+	select HMM_MIRROR
+	select DEVICE_PRIVATE
+	default n
+	help
+	  Say Y here if you want to enable experimental support for
+	  Shared Virtual Memory (SVM).

drivers/gpu/drm/nouveau/dispnv04/crtc.c

···
 #include "nvreg.h"
 #include "nouveau_fbcon.h"
 #include "disp.h"
+#include "nouveau_dma.h"
 
 #include <subdev/bios/pll.h>
 #include <subdev/clk.h>
···
 	return ret;
 }
 
+struct nv04_page_flip_state {
+	struct list_head head;
+	struct drm_pending_vblank_event *event;
+	struct drm_crtc *crtc;
+	int bpp, pitch;
+	u64 offset;
+};
+
+static int
+nv04_finish_page_flip(struct nouveau_channel *chan,
+		      struct nv04_page_flip_state *ps)
+{
+	struct nouveau_fence_chan *fctx = chan->fence;
+	struct nouveau_drm *drm = chan->drm;
+	struct drm_device *dev = drm->dev;
+	struct nv04_page_flip_state *s;
+	unsigned long flags;
+
+	spin_lock_irqsave(&dev->event_lock, flags);
+
+	if (list_empty(&fctx->flip)) {
+		NV_ERROR(drm, "unexpected pageflip\n");
+		spin_unlock_irqrestore(&dev->event_lock, flags);
+		return -EINVAL;
+	}
+
+	s = list_first_entry(&fctx->flip, struct nv04_page_flip_state, head);
+	if (s->event) {
+		drm_crtc_arm_vblank_event(s->crtc, s->event);
+	} else {
+		/* Give up ownership of vblank for page-flipped crtc */
+		drm_crtc_vblank_put(s->crtc);
+	}
+
+	list_del(&s->head);
+	if (ps)
+		*ps = *s;
+	kfree(s);
+
+	spin_unlock_irqrestore(&dev->event_lock, flags);
+	return 0;
+}
+
+int
+nv04_flip_complete(struct nvif_notify *notify)
+{
+	struct nouveau_cli *cli = (void *)notify->object->client;
+	struct nouveau_drm *drm = cli->drm;
+	struct nouveau_channel *chan = drm->channel;
+	struct nv04_page_flip_state state;
+
+	if (!nv04_finish_page_flip(chan, &state)) {
+		nv_set_crtc_base(drm->dev, drm_crtc_index(state.crtc),
+				 state.offset + state.crtc->y *
+				 state.pitch + state.crtc->x *
+				 state.bpp / 8);
+	}
+
+	return NVIF_NOTIFY_KEEP;
+}
+
+static int
+nv04_page_flip_emit(struct nouveau_channel *chan,
+		    struct nouveau_bo *old_bo,
+		    struct nouveau_bo *new_bo,
+		    struct nv04_page_flip_state *s,
+		    struct nouveau_fence **pfence)
+{
+	struct nouveau_fence_chan *fctx = chan->fence;
+	struct nouveau_drm *drm = chan->drm;
+	struct drm_device *dev = drm->dev;
+	unsigned long flags;
+	int ret;
+
+	/* Queue it to the pending list */
+	spin_lock_irqsave(&dev->event_lock, flags);
+	list_add_tail(&s->head, &fctx->flip);
+	spin_unlock_irqrestore(&dev->event_lock, flags);
+
+	/* Synchronize with the old framebuffer */
+	ret = nouveau_fence_sync(old_bo, chan, false, false);
+	if (ret)
+		goto fail;
+
+	/* Emit the pageflip */
+	ret = RING_SPACE(chan, 2);
+	if (ret)
+		goto fail;
+
+	BEGIN_NV04(chan, NvSubSw, NV_SW_PAGE_FLIP, 1);
+	OUT_RING  (chan, 0x00000000);
+	FIRE_RING (chan);
+
+	ret = nouveau_fence_new(chan, false, pfence);
+	if (ret)
+		goto fail;
+
+	return 0;
+fail:
+	spin_lock_irqsave(&dev->event_lock, flags);
+	list_del(&s->head);
+	spin_unlock_irqrestore(&dev->event_lock, flags);
+	return ret;
+}
+
+static int
+nv04_crtc_page_flip(struct drm_crtc *crtc, struct drm_framebuffer *fb,
+		    struct drm_pending_vblank_event *event, u32 flags,
+		    struct drm_modeset_acquire_ctx *ctx)
+{
+	const int swap_interval = (flags & DRM_MODE_PAGE_FLIP_ASYNC) ? 0 : 1;
+	struct drm_device *dev = crtc->dev;
+	struct nouveau_drm *drm = nouveau_drm(dev);
+	struct nouveau_bo *old_bo = nouveau_framebuffer(crtc->primary->fb)->nvbo;
+	struct nouveau_bo *new_bo = nouveau_framebuffer(fb)->nvbo;
+	struct nv04_page_flip_state *s;
+	struct nouveau_channel *chan;
+	struct nouveau_cli *cli;
+	struct nouveau_fence *fence;
+	struct nv04_display *dispnv04 = nv04_display(dev);
+	int head = nouveau_crtc(crtc)->index;
+	int ret;
+
+	chan = drm->channel;
+	if (!chan)
+		return -ENODEV;
+	cli = (void *)chan->user.client;
+
+	s = kzalloc(sizeof(*s), GFP_KERNEL);
+	if (!s)
+		return -ENOMEM;
+
+	if (new_bo != old_bo) {
+		ret = nouveau_bo_pin(new_bo, TTM_PL_FLAG_VRAM, true);
+		if (ret)
+			goto fail_free;
+	}
+
+	mutex_lock(&cli->mutex);
+	ret = ttm_bo_reserve(&new_bo->bo, true, false, NULL);
+	if (ret)
+		goto fail_unpin;
+
+	/* synchronise rendering channel with the kernel's channel */
+	ret = nouveau_fence_sync(new_bo, chan, false, true);
+	if (ret) {
+		ttm_bo_unreserve(&new_bo->bo);
+		goto fail_unpin;
+	}
+
+	if (new_bo != old_bo) {
+		ttm_bo_unreserve(&new_bo->bo);
+
+		ret = ttm_bo_reserve(&old_bo->bo, true, false, NULL);
+		if (ret)
+			goto fail_unpin;
+	}
+
+	/* Initialize a page flip struct */
+	*s = (struct nv04_page_flip_state)
+		{ { }, event, crtc, fb->format->cpp[0] * 8, fb->pitches[0],
+		  new_bo->bo.offset };
+
+	/* Keep vblanks on during flip, for the target crtc of this flip */
+	drm_crtc_vblank_get(crtc);
+
+	/* Emit a page flip */
+	if (swap_interval) {
+		ret = RING_SPACE(chan, 8);
+		if (ret)
+			goto fail_unreserve;
+
+		BEGIN_NV04(chan, NvSubImageBlit, 0x012c, 1);
+		OUT_RING  (chan, 0);
+		BEGIN_NV04(chan, NvSubImageBlit, 0x0134, 1);
+		OUT_RING  (chan, head);
+		BEGIN_NV04(chan, NvSubImageBlit, 0x0100, 1);
+		OUT_RING  (chan, 0);
+		BEGIN_NV04(chan, NvSubImageBlit, 0x0130, 1);
+		OUT_RING  (chan, 0);
+	}
+
+	nouveau_bo_ref(new_bo, &dispnv04->image[head]);
+
+	ret = nv04_page_flip_emit(chan, old_bo, new_bo, s, &fence);
+	if (ret)
+		goto fail_unreserve;
+	mutex_unlock(&cli->mutex);
+
+	/* Update the crtc struct and cleanup */
+	crtc->primary->fb = fb;
+
+	nouveau_bo_fence(old_bo, fence, false);
+	ttm_bo_unreserve(&old_bo->bo);
+	if (old_bo != new_bo)
+		nouveau_bo_unpin(old_bo);
+	nouveau_fence_unref(&fence);
+	return 0;
+
+fail_unreserve:
+	drm_crtc_vblank_put(crtc);
+	ttm_bo_unreserve(&old_bo->bo);
+fail_unpin:
+	mutex_unlock(&cli->mutex);
+	if (old_bo != new_bo)
+		nouveau_bo_unpin(new_bo);
+fail_free:
+	kfree(s);
+	return ret;
+}
+
 static const struct drm_crtc_funcs nv04_crtc_funcs = {
 	.cursor_set = nv04_crtc_cursor_set,
 	.cursor_move = nv04_crtc_cursor_move,
 	.gamma_set = nv_crtc_gamma_set,
 	.set_config = nouveau_crtc_set_config,
-	.page_flip = nouveau_crtc_page_flip,
+	.page_flip = nv04_crtc_page_flip,
 	.destroy = nv_crtc_destroy,
 };
 

drivers/gpu/drm/nouveau/dispnv04/disp.c

···
 #include "hw.h"
 #include "nouveau_encoder.h"
 #include "nouveau_connector.h"
+#include "nouveau_bo.h"
+
+#include <nvif/if0004.h>
+
+static void
+nv04_display_fini(struct drm_device *dev, bool suspend)
+{
+	struct nv04_display *disp = nv04_display(dev);
+	struct drm_crtc *crtc;
+
+	/* Disable flip completion events. */
+	nvif_notify_put(&disp->flip);
+
+	/* Disable vblank interrupts. */
+	NVWriteCRTC(dev, 0, NV_PCRTC_INTR_EN_0, 0);
+	if (nv_two_heads(dev))
+		NVWriteCRTC(dev, 1, NV_PCRTC_INTR_EN_0, 0);
+
+	if (!suspend)
+		return;
+
+	/* Un-pin FB and cursors so they'll be evicted to system memory. */
+	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
+		struct nouveau_framebuffer *nouveau_fb;
+
+		nouveau_fb = nouveau_framebuffer(crtc->primary->fb);
+		if (!nouveau_fb || !nouveau_fb->nvbo)
+			continue;
+
+		nouveau_bo_unpin(nouveau_fb->nvbo);
+	}
+
+	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
+		struct nouveau_crtc *nv_crtc = nouveau_crtc(crtc);
+		if (nv_crtc->cursor.nvbo) {
+			if (nv_crtc->cursor.set_offset)
+				nouveau_bo_unmap(nv_crtc->cursor.nvbo);
+			nouveau_bo_unpin(nv_crtc->cursor.nvbo);
+		}
+	}
+}
+
+static int
+nv04_display_init(struct drm_device *dev, bool resume, bool runtime)
+{
+	struct nv04_display *disp = nv04_display(dev);
+	struct nouveau_drm *drm = nouveau_drm(dev);
+	struct nouveau_encoder *encoder;
+	struct drm_crtc *crtc;
+	int ret;
+
+	/* meh.. modeset apparently doesn't setup all the regs and depends
+	 * on pre-existing state, for now load the state of the card *before*
+	 * nouveau was loaded, and then do a modeset.
+	 *
+	 * best thing to do probably is to make save/restore routines not
+	 * save/restore "pre-load" state, but more general so we can save
+	 * on suspend too.
+	 */
+	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
+		struct nouveau_crtc *nv_crtc = nouveau_crtc(crtc);
+		nv_crtc->save(&nv_crtc->base);
+	}
+
+	list_for_each_entry(encoder, &dev->mode_config.encoder_list, base.base.head)
+		encoder->enc_save(&encoder->base.base);
+
+	/* Enable flip completion events. */
+	nvif_notify_get(&disp->flip);
+
+	if (!resume)
+		return 0;
+
+	/* Re-pin FB/cursors. */
+	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
+		struct nouveau_framebuffer *nouveau_fb;
+
+		nouveau_fb = nouveau_framebuffer(crtc->primary->fb);
+		if (!nouveau_fb || !nouveau_fb->nvbo)
+			continue;
+
+		ret = nouveau_bo_pin(nouveau_fb->nvbo, TTM_PL_FLAG_VRAM, true);
+		if (ret)
+			NV_ERROR(drm, "Could not pin framebuffer\n");
+	}
+
+	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
+		struct nouveau_crtc *nv_crtc = nouveau_crtc(crtc);
+		if (!nv_crtc->cursor.nvbo)
+			continue;
+
+		ret = nouveau_bo_pin(nv_crtc->cursor.nvbo, TTM_PL_FLAG_VRAM, true);
+		if (!ret && nv_crtc->cursor.set_offset)
+			ret = nouveau_bo_map(nv_crtc->cursor.nvbo);
+		if (ret)
+			NV_ERROR(drm, "Could not pin/map cursor.\n");
+	}
+
+	/* Force CLUT to get re-loaded during modeset. */
+	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
+		struct nouveau_crtc *nv_crtc = nouveau_crtc(crtc);
+
+		nv_crtc->lut.depth = 0;
+	}
+
+	/* This should ensure we don't hit a locking problem when someone
+	 * wakes us up via a connector.  We should never go into suspend
+	 * while the display is on anyways.
+	 */
+	if (runtime)
+		return 0;
+
+	/* Restore mode. */
+	drm_helper_resume_force_mode(dev);
+
+	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
+		struct nouveau_crtc *nv_crtc = nouveau_crtc(crtc);
+
+		if (!nv_crtc->cursor.nvbo)
+			continue;
+
+		if (nv_crtc->cursor.set_offset)
+			nv_crtc->cursor.set_offset(nv_crtc, nv_crtc->cursor.nvbo->bo.offset);
+		nv_crtc->cursor.set_pos(nv_crtc, nv_crtc->cursor_saved_x,
+						 nv_crtc->cursor_saved_y);
+	}
+
+	return 0;
+}
+
+static void
+nv04_display_destroy(struct drm_device *dev)
+{
+	struct nv04_display *disp = nv04_display(dev);
+	struct nouveau_drm *drm = nouveau_drm(dev);
+	struct nouveau_encoder *encoder;
+	struct nouveau_crtc *nv_crtc;
+
+	/* Restore state */
+	list_for_each_entry(encoder, &dev->mode_config.encoder_list, base.base.head)
+		encoder->enc_restore(&encoder->base.base);
+
+	list_for_each_entry(nv_crtc, &dev->mode_config.crtc_list, base.head)
+		nv_crtc->restore(&nv_crtc->base);
+
+	nouveau_hw_save_vga_fonts(dev, 0);
+
+	nvif_notify_fini(&disp->flip);
+
+	nouveau_display(dev)->priv = NULL;
+	kfree(disp);
+
+	nvif_object_unmap(&drm->client.device.object);
+}
 
 int
 nv04_display_create(struct drm_device *dev)
···
 
 	/* Pre-nv50 doesn't support atomic, so don't expose the ioctls */
 	dev->driver->driver_features &= ~DRIVER_ATOMIC;
+
+	/* Request page flip completion event. */
+	if (drm->nvsw.client) {
+		nvif_notify_init(&drm->nvsw, nv04_flip_complete,
+				 false, NV04_NVSW_NTFY_UEVENT,
+				 NULL, 0, 0, &disp->flip);
+	}
 
 	nouveau_hw_save_vga_fonts(dev, 1);
 
···
 	nouveau_overlay_init(dev);
 
 	return 0;
-}
-
-void
-nv04_display_destroy(struct drm_device *dev)
-{
-	struct nv04_display *disp = nv04_display(dev);
-	struct nouveau_drm *drm = nouveau_drm(dev);
-	struct nouveau_encoder *encoder;
-	struct nouveau_crtc *nv_crtc;
-
-	/* Restore state */
-	list_for_each_entry(encoder, &dev->mode_config.encoder_list, base.base.head)
-		encoder->enc_restore(&encoder->base.base);
-
-	list_for_each_entry(nv_crtc, &dev->mode_config.crtc_list, base.head)
-		nv_crtc->restore(&nv_crtc->base);
-
-	nouveau_hw_save_vga_fonts(dev, 0);
-
-	nouveau_display(dev)->priv = NULL;
-	kfree(disp);
-
-	nvif_object_unmap(&drm->client.device.object);
-}
-
-int
-nv04_display_init(struct drm_device *dev)
-{
-	struct nouveau_encoder *encoder;
-	struct nouveau_crtc *crtc;
-
-	/* meh.. modeset apparently doesn't setup all the regs and depends
-	 * on pre-existing state, for now load the state of the card *before*
-	 * nouveau was loaded, and then do a modeset.
-	 *
-	 * best thing to do probably is to make save/restore routines not
-	 * save/restore "pre-load" state, but more general so we can save
-	 * on suspend too.
-	 */
-	list_for_each_entry(crtc, &dev->mode_config.crtc_list, base.head)
-		crtc->save(&crtc->base);
-
-	list_for_each_entry(encoder, &dev->mode_config.encoder_list, base.base.head)
-		encoder->enc_save(&encoder->base.base);
-
-	return 0;
-}
-
-void
-nv04_display_fini(struct drm_device *dev)
-{
-	/* disable vblank interrupts */
-	NVWriteCRTC(dev, 0, NV_PCRTC_INTR_EN_0, 0);
-	if (nv_two_heads(dev))
-		NVWriteCRTC(dev, 1, NV_PCRTC_INTR_EN_0, 0);
 }

drivers/gpu/drm/nouveau/dispnv04/disp.h

···
 	uint32_t saved_vga_font[4][16384];
 	uint32_t dac_users[4];
 	struct nouveau_bo *image[2];
+	struct nvif_notify flip;
 };
 
 static inline struct nv04_display *
···
 
 /* nv04_display.c */
 int nv04_display_create(struct drm_device *);
-void nv04_display_destroy(struct drm_device *);
-int nv04_display_init(struct drm_device *);
-void nv04_display_fini(struct drm_device *);
 
 /* nv04_crtc.c */
 int nv04_crtc_create(struct drm_device *, int index);
···
 	);
 }
 
+int nv04_flip_complete(struct nvif_notify *);
 #endif

drivers/gpu/drm/nouveau/dispnv50/core.c

···
 		int version;
 		int (*new)(struct nouveau_drm *, s32, struct nv50_core **);
 	} cores[] = {
-		{ TU104_DISP_CORE_CHANNEL_DMA, 0, corec57d_new },
+		{ TU102_DISP_CORE_CHANNEL_DMA, 0, corec57d_new },
 		{ GV100_DISP_CORE_CHANNEL_DMA, 0, corec37d_new },
 		{ GP102_DISP_CORE_CHANNEL_DMA, 0, core917d_new },
 		{ GP100_DISP_CORE_CHANNEL_DMA, 0, core917d_new },

drivers/gpu/drm/nouveau/dispnv50/curs.c

···
 		int version;
 		int (*new)(struct nouveau_drm *, int, s32, struct nv50_wndw **);
 	} curses[] = {
-		{ TU104_DISP_CURSOR, 0, cursc37a_new },
+		{ TU102_DISP_CURSOR, 0, cursc37a_new },
 		{ GV100_DISP_CURSOR, 0, cursc37a_new },
 		{ GK104_DISP_CURSOR, 0, curs907a_new },
 		{ GF110_DISP_CURSOR, 0, curs907a_new },

drivers/gpu/drm/nouveau/dispnv50/disp.c

···
 
 	r = drm_dp_mst_allocate_vcpi(&mstm->mgr, mstc->port, armh->dp.pbn,
 				     armh->dp.tu);
-	WARN_ON(!r);
+	if (!r)
+		DRM_DEBUG_KMS("Failed to allocate VCPI\n");
 
 	if (!mstm->links++)
 		nv50_outp_acquire(mstm->outp);
···
  * Init
  *****************************************************************************/
 
-void
-nv50_display_fini(struct drm_device *dev)
+static void
+nv50_display_fini(struct drm_device *dev, bool suspend)
 {
 	struct nouveau_encoder *nv_encoder;
 	struct drm_encoder *encoder;
···
 	}
 }
 
-int
-nv50_display_init(struct drm_device *dev)
+static int
+nv50_display_init(struct drm_device *dev, bool resume, bool runtime)
 {
 	struct nv50_core *core = nv50_disp(dev)->core;
 	struct drm_encoder *encoder;
···
 	return 0;
 }
 
-void
+static void
 nv50_display_destroy(struct drm_device *dev)
 {
 	struct nv50_disp *disp = nv50_disp(dev);

drivers/gpu/drm/nouveau/dispnv50/wimm.c

···
 		int version;
 		int (*init)(struct nouveau_drm *, s32, struct nv50_wndw *);
 	} wimms[] = {
-		{ TU104_DISP_WINDOW_IMM_CHANNEL_DMA, 0, wimmc37b_init },
+		{ TU102_DISP_WINDOW_IMM_CHANNEL_DMA, 0, wimmc37b_init },
 		{ GV100_DISP_WINDOW_IMM_CHANNEL_DMA, 0, wimmc37b_init },
 		{}
 	};

drivers/gpu/drm/nouveau/dispnv50/wndw.c

···
 		int (*new)(struct nouveau_drm *, enum drm_plane_type,
 			   int, s32, struct nv50_wndw **);
 	} wndws[] = {
-		{ TU104_DISP_WINDOW_CHANNEL_DMA, 0, wndwc57e_new },
+		{ TU102_DISP_WINDOW_CHANNEL_DMA, 0, wndwc57e_new },
 		{ GV100_DISP_WINDOW_CHANNEL_DMA, 0, wndwc37e_new },
 		{}
 	};

drivers/gpu/drm/nouveau/include/nvif/class.h

···
 
 #define VOLTA_USERMODE_A                                             0x0000c361
 
+#define MAXWELL_FAULT_BUFFER_A                        /* clb069.h */ 0x0000b069
+#define VOLTA_FAULT_BUFFER_A                          /* clb069.h */ 0x0000c369
+
 #define NV03_CHANNEL_DMA                              /* cl506b.h */ 0x0000006b
 #define NV10_CHANNEL_DMA                              /* cl506b.h */ 0x0000006e
 #define NV17_CHANNEL_DMA                              /* cl506b.h */ 0x0000176e
···
 #define GP100_DISP                                    /* cl5070.h */ 0x00009770
 #define GP102_DISP                                    /* cl5070.h */ 0x00009870
 #define GV100_DISP                                    /* cl5070.h */ 0x0000c370
-#define TU104_DISP                                    /* cl5070.h */ 0x0000c570
+#define TU102_DISP                                    /* cl5070.h */ 0x0000c570
 
 #define NV31_MPEG                                                    0x00003174
 #define G82_MPEG                                                     0x00008274
···
 #define GF110_DISP_CURSOR                             /* cl507a.h */ 0x0000907a
 #define GK104_DISP_CURSOR                             /* cl507a.h */ 0x0000917a
 #define GV100_DISP_CURSOR                             /* cl507a.h */ 0x0000c37a
-#define TU104_DISP_CURSOR                             /* cl507a.h */ 0x0000c57a
+#define TU102_DISP_CURSOR                             /* cl507a.h */ 0x0000c57a
 
 #define NV50_DISP_OVERLAY                             /* cl507b.h */ 0x0000507b
 #define G82_DISP_OVERLAY                              /* cl507b.h */ 0x0000827b
···
 #define GK104_DISP_OVERLAY                            /* cl507b.h */ 0x0000917b
 
 #define GV100_DISP_WINDOW_IMM_CHANNEL_DMA             /* clc37b.h */ 0x0000c37b
-#define TU104_DISP_WINDOW_IMM_CHANNEL_DMA             /* clc37b.h */ 0x0000c57b
+#define TU102_DISP_WINDOW_IMM_CHANNEL_DMA             /* clc37b.h */ 0x0000c57b
 
 #define NV50_DISP_BASE_CHANNEL_DMA                    /* cl507c.h */ 0x0000507c
 #define G82_DISP_BASE_CHANNEL_DMA                     /* cl507c.h */ 0x0000827c
···
 #define GP100_DISP_CORE_CHANNEL_DMA                   /* cl507d.h */ 0x0000977d
 #define GP102_DISP_CORE_CHANNEL_DMA                   /* cl507d.h */ 0x0000987d
 #define GV100_DISP_CORE_CHANNEL_DMA                   /* cl507d.h */ 0x0000c37d
-#define TU104_DISP_CORE_CHANNEL_DMA                   /* cl507d.h */ 0x0000c57d
+#define TU102_DISP_CORE_CHANNEL_DMA                   /* cl507d.h */ 0x0000c57d
 
 #define NV50_DISP_OVERLAY_CHANNEL_DMA                 /* cl507e.h */ 0x0000507e
 #define G82_DISP_OVERLAY_CHANNEL_DMA                  /* cl507e.h */ 0x0000827e
···
 #define GK104_DISP_OVERLAY_CONTROL_DMA                /* cl507e.h */ 0x0000917e
 
 #define GV100_DISP_WINDOW_CHANNEL_DMA                 /* clc37e.h */ 0x0000c37e
-#define TU104_DISP_WINDOW_CHANNEL_DMA                 /* clc37e.h */ 0x0000c57e
+#define TU102_DISP_WINDOW_CHANNEL_DMA                 /* clc37e.h */ 0x0000c57e
 
 #define NV50_TESLA                                                   0x00005097
 #define G82_TESLA                                                    0x00008297

drivers/gpu/drm/nouveau/include/nvif/clb069.h

···
+#ifndef __NVIF_CLB069_H__
+#define __NVIF_CLB069_H__
+struct nvif_clb069_v0 {
+	__u8  version;
+	__u8  pad01[3];
+	__u32 entries;
+	__u32 get;
+	__u32 put;
+};
+
+#define NVB069_V0_NTFY_FAULT                                                0x00
+#endif

drivers/gpu/drm/nouveau/include/nvif/if000c.h

···
 struct nvif_vmm_v0 {
 	__u8  version;
 	__u8  page_nr;
-	__u8  pad02[6];
+	__u8  managed;
+	__u8  pad03[5];
 	__u64 addr;
 	__u64 size;
 	__u8  data[];
···
 #define NVIF_VMM_V0_PUT                                                    0x02
 #define NVIF_VMM_V0_MAP                                                    0x03
 #define NVIF_VMM_V0_UNMAP                                                  0x04
+#define NVIF_VMM_V0_PFNMAP                                                 0x05
+#define NVIF_VMM_V0_PFNCLR                                                 0x06
+#define NVIF_VMM_V0_MTHD(i)                                         ((i) + 0x80)
 
 struct nvif_vmm_page_v0 {
 	__u8  version;
···
 	__u8  version;
 	__u8  pad01[7];
 	__u64 addr;
+};
+
+struct nvif_vmm_pfnmap_v0 {
+	__u8  version;
+	__u8  page;
+	__u8  pad02[6];
+	__u64 addr;
+	__u64 size;
+#define NVIF_VMM_PFNMAP_V0_ADDR                           0xfffffffffffff000ULL
+#define NVIF_VMM_PFNMAP_V0_ADDR_SHIFT                                        12
+#define NVIF_VMM_PFNMAP_V0_APER                           0x00000000000000f0ULL
+#define NVIF_VMM_PFNMAP_V0_HOST                           0x0000000000000000ULL
+#define NVIF_VMM_PFNMAP_V0_VRAM                           0x0000000000000010ULL
+#define NVIF_VMM_PFNMAP_V0_W                              0x0000000000000002ULL
+#define NVIF_VMM_PFNMAP_V0_V                              0x0000000000000001ULL
+#define NVIF_VMM_PFNMAP_V0_NONE                           0x0000000000000000ULL
+	__u64 phys[];
+};
+
+struct nvif_vmm_pfnclr_v0 {
+	__u8  version;
+	__u8  pad01[7];
+	__u64 addr;
+	__u64 size;
 };
 #endif

drivers/gpu/drm/nouveau/include/nvif/ifc00d.h

···
 	/* nvif_vmm_vX ... */
 };
 
+struct gp100_vmm_v0 {
+	/* nvif_vmm_vX ... */
+	__u8  version;
+	__u8  fault_replay;
+};
+
 struct gp100_vmm_map_vn {
 	/* nvif_vmm_map_vX ... */
 };
···
 	__u8  ro;
 	__u8  priv;
 	__u8  kind;
+};
+
+#define GP100_VMM_VN_FAULT_REPLAY                         NVIF_VMM_V0_MTHD(0x00)
+#define GP100_VMM_VN_FAULT_CANCEL                         NVIF_VMM_V0_MTHD(0x01)
+
+struct gp100_vmm_fault_replay_vn {
+};
+
+struct gp100_vmm_fault_cancel_v0 {
+	__u8  version;
+	__u8  hub;
+	__u8  gpc;
+	__u8  client;
+	__u8  pad04[4];
+	__u64 inst;
 };
 #endif

drivers/gpu/drm/nouveau/include/nvif/vmm.h

···
 	int page_nr;
 };
 
-int nvif_vmm_init(struct nvif_mmu *, s32 oclass, u64 addr, u64 size,
-		  void *argv, u32 argc, struct nvif_vmm *);
+int nvif_vmm_init(struct nvif_mmu *, s32 oclass, bool managed, u64 addr,
+		  u64 size, void *argv, u32 argc, struct nvif_vmm *);
 void nvif_vmm_fini(struct nvif_vmm *);
 int nvif_vmm_get(struct nvif_vmm *, enum nvif_vmm_get, bool sparse,
 		 u8 page, u8 align, u64 size, struct nvif_vma *);

drivers/gpu/drm/nouveau/include/nvkm/core/device.h

···
 	NVKM_SUBDEV_ICCSENSE,
 	NVKM_SUBDEV_THERM,
 	NVKM_SUBDEV_CLK,
+	NVKM_SUBDEV_GSP,
 	NVKM_SUBDEV_SECBOOT,
 
 	NVKM_ENGINE_BSP,
···
 	struct nvkm_fb *fb;
 	struct nvkm_fuse *fuse;
 	struct nvkm_gpio *gpio;
+	struct nvkm_gsp *gsp;
 	struct nvkm_i2c *i2c;
 	struct nvkm_subdev *ibus;
 	struct nvkm_iccsense *iccsense;
···
 	int (*fb      )(struct nvkm_device *, int idx, struct nvkm_fb **);
 	int (*fuse    )(struct nvkm_device *, int idx, struct nvkm_fuse **);
 	int (*gpio    )(struct nvkm_device *, int idx, struct nvkm_gpio **);
+	int (*gsp     )(struct nvkm_device *, int idx, struct nvkm_gsp **);
 	int (*i2c     )(struct nvkm_device *, int idx, struct nvkm_i2c **);
 	int (*ibus    )(struct nvkm_device *, int idx, struct nvkm_subdev **);
 	int (*iccsense)(struct nvkm_device *, int idx, struct nvkm_iccsense **);

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

···
 int gp100_ce_new(struct nvkm_device *, int, struct nvkm_engine **);
 int gp102_ce_new(struct nvkm_device *, int, struct nvkm_engine **);
 int gv100_ce_new(struct nvkm_device *, int, struct nvkm_engine **);
-int tu104_ce_new(struct nvkm_device *, int, struct nvkm_engine **);
+int tu102_ce_new(struct nvkm_device *, int, struct nvkm_engine **);
 #endif

drivers/gpu/drm/nouveau/include/nvkm/engine/disp.h

···
 int gp100_disp_new(struct nvkm_device *, int, struct nvkm_disp **);
 int gp102_disp_new(struct nvkm_device *, int, struct nvkm_disp **);
 int gv100_disp_new(struct nvkm_device *, int, struct nvkm_disp **);
-int tu104_disp_new(struct nvkm_device *, int, struct nvkm_disp **);
+int tu102_disp_new(struct nvkm_device *, int, struct nvkm_disp **);
 #endif

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

···
 int gp100_fifo_new(struct nvkm_device *, int, struct nvkm_fifo **);
 int gp10b_fifo_new(struct nvkm_device *, int, struct nvkm_fifo **);
 int gv100_fifo_new(struct nvkm_device *, int, struct nvkm_fifo **);
-int tu104_fifo_new(struct nvkm_device *, int, struct nvkm_fifo **);
+int tu102_fifo_new(struct nvkm_device *, int, struct nvkm_fifo **);
 #endif

drivers/gpu/drm/nouveau/include/nvkm/engine/gr.h

···
 
 u64 nvkm_gr_units(struct nvkm_gr *);
 int nvkm_gr_tlb_flush(struct nvkm_gr *);
+int nvkm_gr_ctxsw_pause(struct nvkm_device *);
+int nvkm_gr_ctxsw_resume(struct nvkm_device *);
+u32 nvkm_gr_ctxsw_inst(struct nvkm_device *);
 
 int nv04_gr_new(struct nvkm_device *, int, struct nvkm_gr **);
 int nv10_gr_new(struct nvkm_device *, int, struct nvkm_gr **);

drivers/gpu/drm/nouveau/include/nvkm/engine/nvdec.h

···
 
 struct nvkm_nvdec {
 	struct nvkm_engine engine;
+	u32 addr;
+
 	struct nvkm_falcon *falcon;
 };
 

drivers/gpu/drm/nouveau/include/nvkm/engine/sec2.h

···
 
 struct nvkm_sec2 {
 	struct nvkm_engine engine;
+	u32 addr;
+
 	struct nvkm_falcon *falcon;
 	struct nvkm_msgqueue *queue;
 	struct work_struct work;
 };
 
 int gp102_sec2_new(struct nvkm_device *, int, struct nvkm_sec2 **);
+int tu102_sec2_new(struct nvkm_device *, int, struct nvkm_sec2 **);
 #endif

drivers/gpu/drm/nouveau/include/nvkm/subdev/bar.h

···
 int gk20a_bar_new(struct nvkm_device *, int, struct nvkm_bar **);
 int gm107_bar_new(struct nvkm_device *, int, struct nvkm_bar **);
 int gm20b_bar_new(struct nvkm_device *, int, struct nvkm_bar **);
-int tu104_bar_new(struct nvkm_device *, int, struct nvkm_bar **);
+int tu102_bar_new(struct nvkm_device *, int, struct nvkm_bar **);
 #endif

drivers/gpu/drm/nouveau/include/nvkm/subdev/devinit.h

···
 int gm107_devinit_new(struct nvkm_device *, int, struct nvkm_devinit **);
 int gm200_devinit_new(struct nvkm_device *, int, struct nvkm_devinit **);
 int gv100_devinit_new(struct nvkm_device *, int, struct nvkm_devinit **);
-int tu104_devinit_new(struct nvkm_device *, int, struct nvkm_devinit **);
+int tu102_devinit_new(struct nvkm_device *, int, struct nvkm_devinit **);
 #endif

drivers/gpu/drm/nouveau/include/nvkm/subdev/fault.h

···
 	struct nvkm_event event;
 
 	struct nvkm_notify nrpfb;
+
+	struct nvkm_device_oclass user;
 };
 
 struct nvkm_fault_data {
···
 
 int gp100_fault_new(struct nvkm_device *, int, struct nvkm_fault **);
 int gv100_fault_new(struct nvkm_device *, int, struct nvkm_fault **);
-int tu104_fault_new(struct nvkm_device *, int, struct nvkm_fault **);
+int tu102_fault_new(struct nvkm_device *, int, struct nvkm_fault **);
 #endif

drivers/gpu/drm/nouveau/include/nvkm/subdev/gsp.h

···
+#ifndef __NVKM_GSP_H__
+#define __NVKM_GSP_H__
+#define nvkm_gsp(p) container_of((p), struct nvkm_gsp, subdev)
+#include <core/subdev.h>
+
+struct nvkm_gsp {
+	struct nvkm_subdev subdev;
+	u32 addr;
+
+	struct nvkm_falcon *falcon;
+};
+
+int gv100_gsp_new(struct nvkm_device *, int, struct nvkm_gsp **);
+#endif

drivers/gpu/drm/nouveau/include/nvkm/subdev/mc.h

···
 int gk20a_mc_new(struct nvkm_device *, int, struct nvkm_mc **);
 int gp100_mc_new(struct nvkm_device *, int, struct nvkm_mc **);
 int gp10b_mc_new(struct nvkm_device *, int, struct nvkm_mc **);
-int tu104_mc_new(struct nvkm_device *, int, struct nvkm_mc **);
+int tu102_mc_new(struct nvkm_device *, int, struct nvkm_mc **);
 #endif

drivers/gpu/drm/nouveau/include/nvkm/subdev/mmu.h

···
 	bool part:1; /* Region was split from an allocated region by map(). */
 	bool user:1; /* Region user-allocated. */
 	bool busy:1; /* Region busy (for temporarily preventing user access). */
+	bool mapped:1; /* Region contains valid pages. */
 	struct nvkm_memory *memory; /* Memory currently mapped into VMA. */
 	struct nvkm_tags *tags; /* Compression tag reference. */
 };
···
 
 	dma_addr_t null;
 	void *nullp;
+
+	bool replay;
 };
 
 int nvkm_vmm_new(struct nvkm_device *, u64 addr, u64 size, void *argv, u32 argc,
···
 	struct nvkm_mm_node *mem;
 	struct scatterlist *sgl;
 	dma_addr_t *dma;
+	u64 *pfn;
 	u64 off;
 
 	const struct nvkm_vmm_page *page;
···
 int gp100_mmu_new(struct nvkm_device *, int, struct nvkm_mmu **);
 int gp10b_mmu_new(struct nvkm_device *, int, struct nvkm_mmu **);
 int gv100_mmu_new(struct nvkm_device *, int, struct nvkm_mmu **);
-int tu104_mmu_new(struct nvkm_device *, int, struct nvkm_mmu **);
+int tu102_mmu_new(struct nvkm_device *, int, struct nvkm_mmu **);
 #endif

drivers/gpu/drm/nouveau/include/nvkm/subdev/top.h

···
 	struct list_head device;
 };
 
+u32 nvkm_top_addr(struct nvkm_device *, enum nvkm_devidx);
 u32 nvkm_top_reset(struct nvkm_device *, enum nvkm_devidx);
 u32 nvkm_top_intr(struct nvkm_device *, u32 intr, u64 *subdevs);
 u32 nvkm_top_intr_mask(struct nvkm_device *, enum nvkm_devidx);

drivers/gpu/drm/nouveau/include/nvkm/subdev/volt.h

···
 
 int nv40_volt_new(struct nvkm_device *, int, struct nvkm_volt **);
 int gf100_volt_new(struct nvkm_device *, int, struct nvkm_volt **);
+int gf117_volt_new(struct nvkm_device *, int, struct nvkm_volt **);
 int gk104_volt_new(struct nvkm_device *, int, struct nvkm_volt **);
 int gk20a_volt_new(struct nvkm_device *, int, struct nvkm_volt **);
 int gm20b_volt_new(struct nvkm_device *, int, struct nvkm_volt **);

drivers/gpu/drm/nouveau/nouveau_abi16.c

···
 			WARN_ON(1);
 			break;
 		}
+		break;
 	case NOUVEAU_GETPARAM_FB_SIZE:
 		getparam->value = drm->gem.vram_available;
 		break;
···
 		goto done;
 
 	if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
-		ret = nouveau_vma_new(chan->ntfy, &cli->vmm, &chan->ntfy_vma);
+		ret = nouveau_vma_new(chan->ntfy, chan->chan->vmm,
+				      &chan->ntfy_vma);
 		if (ret)
 			goto done;
 	}

drivers/gpu/drm/nouveau/nouveau_bo.c

···
 	struct nouveau_drm *drm = cli->drm;
 	struct nouveau_bo *nvbo;
 	struct nvif_mmu *mmu = &cli->mmu;
-	struct nvif_vmm *vmm = &cli->vmm.vmm;
+	struct nvif_vmm *vmm = cli->svm.cli ? &cli->svm.vmm : &cli->vmm.vmm;
 	size_t acc_size;
 	int type = ttm_bo_type_device;
 	int ret, i, pi = -1;
···
 		if (drm->client.mem->oclass < NVIF_CLASS_MEM_NV50 || !mem->kind)
 			/* untiled */
 			break;
-		/* fallthrough, tiled memory */
+		/* fall through - tiled memory */
 	case TTM_PL_VRAM:
 		reg->bus.offset = reg->start << PAGE_SHIFT;
 		reg->bus.base = device->func->resource_addr(device, 1);

drivers/gpu/drm/nouveau/nouveau_chan.c

···
 #include "nouveau_fence.h"
 #include "nouveau_abi16.h"
 #include "nouveau_vmm.h"
+#include "nouveau_svm.h"
 
 MODULE_PARM_DESC(vram_pushbuf, "Create DMA push buffers in VRAM");
 int nouveau_vram_pushbuf;
···
 
 		if (chan->fence)
 			nouveau_fence(chan->drm)->context_del(chan);
+
+		if (cli)
+			nouveau_svmm_part(chan->vmm->svmm, chan->inst);
+
 		nvif_object_fini(&chan->nvsw);
 		nvif_object_fini(&chan->gart);
 		nvif_object_fini(&chan->vram);
···
 
 	chan->device = device;
 	chan->drm = drm;
+	chan->vmm = cli->svm.cli ? &cli->svm : &cli->vmm;
 	atomic_set(&chan->killed, 0);
 
 	/* allocate memory for dma push buffer */
···
 	chan->push.addr = chan->push.buffer->bo.offset;
 
 	if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
-		ret = nouveau_vma_new(chan->push.buffer, &cli->vmm,
+		ret = nouveau_vma_new(chan->push.buffer, chan->vmm,
 				      &chan->push.vma);
 		if (ret) {
 			nouveau_channel_del(pchan);
···
 		args.target = NV_DMA_V0_TARGET_VM;
 		args.access = NV_DMA_V0_ACCESS_VM;
 		args.start = 0;
-		args.limit = cli->vmm.vmm.limit - 1;
+		args.limit = chan->vmm->vmm.limit - 1;
 	} else
 	if (chan->push.buffer->bo.mem.mem_type == TTM_PL_VRAM) {
 		if (device->info.family == NV_DEVICE_INFO_V0_TNT) {
···
 			args.target = NV_DMA_V0_TARGET_VM;
 			args.access = NV_DMA_V0_ACCESS_RDWR;
 			args.start = 0;
-			args.limit = cli->vmm.vmm.limit - 1;
+			args.limit = chan->vmm->vmm.limit - 1;
 		}
 	}
 
···
 nouveau_channel_ind(struct nouveau_drm *drm, struct nvif_device *device,
 		    u64 runlist, bool priv, struct nouveau_channel **pchan)
 {
-	struct nouveau_cli *cli = (void *)device->object.client;
 	static const u16 oclasses[] = { TURING_CHANNEL_GPFIFO_A,
 					VOLTA_CHANNEL_GPFIFO_A,
 					PASCAL_CHANNEL_GPFIFO_A,
···
 			args.volta.ilength = 0x02000;
 			args.volta.ioffset = 0x10000 + chan->push.addr;
 			args.volta.runlist = runlist;
-			args.volta.vmm = nvif_handle(&cli->vmm.vmm.object);
+			args.volta.vmm = nvif_handle(&chan->vmm->vmm.object);
 			args.volta.priv = priv;
 			size = sizeof(args.volta);
 		} else
···
 			args.kepler.ilength = 0x02000;
 			args.kepler.ioffset = 0x10000 + chan->push.addr;
 			args.kepler.runlist = runlist;
-			args.kepler.vmm = nvif_handle(&cli->vmm.vmm.object);
+			args.kepler.vmm = nvif_handle(&chan->vmm->vmm.object);
 			args.kepler.priv = priv;
 			size = sizeof(args.kepler);
 		} else
···
 			args.fermi.version = 0;
 			args.fermi.ilength = 0x02000;
 			args.fermi.ioffset = 0x10000 + chan->push.addr;
-			args.fermi.vmm = nvif_handle(&cli->vmm.vmm.object);
+			args.fermi.vmm = nvif_handle(&chan->vmm->vmm.object);
 			size = sizeof(args.fermi);
 		} else {
 			args.nv50.version = 0;
 			args.nv50.ilength = 0x02000;
 			args.nv50.ioffset = 0x10000 + chan->push.addr;
 			args.nv50.pushbuf = nvif_handle(&chan->push.ctxdma);
-			args.nv50.vmm = nvif_handle(&cli->vmm.vmm.object);
+			args.nv50.vmm = nvif_handle(&chan->vmm->vmm.object);
 			size = sizeof(args.nv50);
 		}
 
···
 nouveau_channel_init(struct nouveau_channel *chan, u32 vram, u32 gart)
 {
 	struct nvif_device *device = chan->device;
-	struct nouveau_cli *cli = (void *)chan->user.client;
 	struct nouveau_drm *drm = chan->drm;
 	struct nv_dma_v0 args = {};
 	int ret, i;
···
 			args.target = NV_DMA_V0_TARGET_VM;
 			args.access = NV_DMA_V0_ACCESS_VM;
 			args.start = 0;
-			args.limit = cli->vmm.vmm.limit - 1;
+			args.limit = chan->vmm->vmm.limit - 1;
 		} else {
 			args.target = NV_DMA_V0_TARGET_VRAM;
 			args.access = NV_DMA_V0_ACCESS_RDWR;
···
 			args.target = NV_DMA_V0_TARGET_VM;
 			args.access = NV_DMA_V0_ACCESS_VM;
 			args.start = 0;
-			args.limit = cli->vmm.vmm.limit - 1;
+			args.limit = chan->vmm->vmm.limit - 1;
 		} else
 		if (chan->drm->agp.bridge) {
 			args.target = NV_DMA_V0_TARGET_AGP;
···
 			args.target = NV_DMA_V0_TARGET_VM;
 			args.access = NV_DMA_V0_ACCESS_RDWR;
 			args.start = 0;
-			args.limit = cli->vmm.vmm.limit - 1;
+			args.limit = chan->vmm->vmm.limit - 1;
 		}
 
 		ret = nvif_object_init(&chan->user, gart, NV_DMA_IN_MEMORY,
···
 		NV_PRINTK(err, cli, "channel failed to initialise, %d\n", ret);
 		nouveau_channel_del(pchan);
 	}
+
+	ret = nouveau_svmm_join((*pchan)->vmm->svmm, (*pchan)->inst);
+	if (ret)
+		nouveau_channel_del(pchan);
 
 done:
 	cli->base.super = super;

drivers/gpu/drm/nouveau/nouveau_chan.h

···
 struct nouveau_channel {
 	struct nvif_device *device;
 	struct nouveau_drm *drm;
+	struct nouveau_vmm *vmm;
 
 	int chid;
 	u64 inst;

drivers/gpu/drm/nouveau/nouveau_display.c

···
 #include <drm/drm_fb_helper.h>
 #include <drm/drm_probe_helper.h>
 
-#include <nvif/class.h>
-
 #include "nouveau_fbcon.h"
-#include "dispnv04/hw.h"
 #include "nouveau_crtc.h"
-#include "nouveau_dma.h"
 #include "nouveau_gem.h"
 #include "nouveau_connector.h"
 #include "nv50_display.h"
 
-#include "nouveau_fence.h"
-
+#include <nvif/class.h>
 #include <nvif/cl0046.h>
 #include <nvif/event.h>
 
···
 #endif
 
 int
-nouveau_display_init(struct drm_device *dev)
+nouveau_display_init(struct drm_device *dev, bool resume, bool runtime)
 {
 	struct nouveau_display *disp = nouveau_display(dev);
-	struct nouveau_drm *drm = nouveau_drm(dev);
 	struct drm_connector *connector;
 	struct drm_connector_list_iter conn_iter;
 	int ret;
 
-	ret = disp->init(dev);
+	ret = disp->init(dev, resume, runtime);
 	if (ret)
 		return ret;
 
···
 	}
 	drm_connector_list_iter_end(&conn_iter);
 
-	/* enable flip completion events */
-	nvif_notify_get(&drm->flip);
 	return ret;
 }
 
···
 			drm_helper_force_disable_all(dev);
 	}
 
-	/* disable flip completion events */
-	nvif_notify_put(&drm->flip);
-
 	/* disable hotplug interrupts */
 	drm_connector_list_iter_begin(dev, &conn_iter);
 	nouveau_for_each_non_mst_connector_iter(connector, &conn_iter) {
···
 		cancel_work_sync(&drm->hpd_work);
 
 	drm_kms_helper_poll_disable(dev);
-	disp->fini(dev);
+	disp->fini(dev, suspend);
 }
 
 static void
···
 nouveau_display_suspend(struct drm_device *dev, bool runtime)
 {
 	struct nouveau_display *disp = nouveau_display(dev);
-	struct drm_crtc *crtc;
 
 	if (drm_drv_uses_atomic_modeset(dev)) {
 		if (!runtime) {
···
 				return ret;
 			}
 		}
-
-		nouveau_display_fini(dev, true, runtime);
-		return 0;
 	}
 
 	nouveau_display_fini(dev, true, runtime);
-
-	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
-		struct nouveau_framebuffer *nouveau_fb;
-
-		nouveau_fb = nouveau_framebuffer(crtc->primary->fb);
-		if (!nouveau_fb || !nouveau_fb->nvbo)
-			continue;
-
-		nouveau_bo_unpin(nouveau_fb->nvbo);
-	}
-
-	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
-		struct nouveau_crtc *nv_crtc = nouveau_crtc(crtc);
-		if (nv_crtc->cursor.nvbo) {
-			if (nv_crtc->cursor.set_offset)
-				nouveau_bo_unmap(nv_crtc->cursor.nvbo);
-			nouveau_bo_unpin(nv_crtc->cursor.nvbo);
-		}
-	}
-
 	return 0;
 }
 
···
 nouveau_display_resume(struct drm_device *dev, bool runtime)
 {
 	struct nouveau_display *disp = nouveau_display(dev);
-	struct nouveau_drm *drm = nouveau_drm(dev);
-	struct drm_crtc *crtc;
-	int ret;
+
+	nouveau_display_init(dev, true, runtime);
 
 	if (drm_drv_uses_atomic_modeset(dev)) {
-		nouveau_display_init(dev);
 		if (disp->suspend) {
 			drm_atomic_helper_resume(dev, disp->suspend);
 			disp->suspend = NULL;
 		}
 		return;
 	}
-
-	/* re-pin fb/cursors */
-	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
-		struct nouveau_framebuffer *nouveau_fb;
-
-		nouveau_fb = nouveau_framebuffer(crtc->primary->fb);
-		if (!nouveau_fb || !nouveau_fb->nvbo)
-			continue;
-
-		ret = nouveau_bo_pin(nouveau_fb->nvbo, TTM_PL_FLAG_VRAM, true);
-		if (ret)
-			NV_ERROR(drm, "Could not pin framebuffer\n");
-	}
-
-	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
-		struct nouveau_crtc *nv_crtc = nouveau_crtc(crtc);
-		if (!nv_crtc->cursor.nvbo)
-			continue;
-
-		ret = nouveau_bo_pin(nv_crtc->cursor.nvbo, TTM_PL_FLAG_VRAM, true);
-		if (!ret && nv_crtc->cursor.set_offset)
-			ret = nouveau_bo_map(nv_crtc->cursor.nvbo);
-		if (ret)
-			NV_ERROR(drm, "Could not pin/map cursor.\n");
-	}
-
-	nouveau_display_init(dev);
-
-	/* Force CLUT to get re-loaded during modeset */
-	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
-		struct nouveau_crtc *nv_crtc = nouveau_crtc(crtc);
-
-		nv_crtc->lut.depth = 0;
-	}
-
-	/* This should ensure we don't hit a locking problem when someone
-	 * wakes us up via a connector.  We should never go into suspend
-	 * while the display is on anyways.
-	 */
-	if (runtime)
-		return;
-
-	drm_helper_resume_force_mode(dev);
-
-	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
-		struct nouveau_crtc *nv_crtc = nouveau_crtc(crtc);
-
-		if (!nv_crtc->cursor.nvbo)
-			continue;
-
-		if (nv_crtc->cursor.set_offset)
-			nv_crtc->cursor.set_offset(nv_crtc, nv_crtc->cursor.nvbo->bo.offset);
-		nv_crtc->cursor.set_pos(nv_crtc, nv_crtc->cursor_saved_x,
-						 nv_crtc->cursor_saved_y);
-	}
-}
-
-static int
-nouveau_page_flip_emit(struct nouveau_channel *chan,
-		       struct nouveau_bo *old_bo,
-		       struct nouveau_bo *new_bo,
-		       struct nouveau_page_flip_state *s,
-		       struct nouveau_fence **pfence)
-{
-	struct nouveau_fence_chan *fctx = chan->fence;
-	struct nouveau_drm *drm = chan->drm;
-	struct drm_device *dev = drm->dev;
-	unsigned long flags;
-	int ret;
-
-	/* Queue it to the pending list */
-	spin_lock_irqsave(&dev->event_lock, flags);
-	list_add_tail(&s->head, &fctx->flip);
-	spin_unlock_irqrestore(&dev->event_lock, flags);
-
-	/* Synchronize with the old framebuffer */
-	ret = nouveau_fence_sync(old_bo, chan, false, false);
-	if (ret)
-		goto fail;
-
-	/* Emit the pageflip */
-	ret = RING_SPACE(chan, 2);
-	if (ret)
-		goto fail;
-
-	BEGIN_NV04(chan, NvSubSw, NV_SW_PAGE_FLIP, 1);
-	OUT_RING  (chan, 0x00000000);
-	FIRE_RING (chan);
-
-	ret = nouveau_fence_new(chan, false, pfence);
-	if (ret)
-		goto fail;
-
-	return 0;
-fail:
-	spin_lock_irqsave(&dev->event_lock, flags);
-	list_del(&s->head);
-	spin_unlock_irqrestore(&dev->event_lock, flags);
-	return ret;
-}
-
-int
-nouveau_crtc_page_flip(struct drm_crtc *crtc, struct drm_framebuffer *fb,
-		       struct drm_pending_vblank_event *event, u32 flags,
-		       struct drm_modeset_acquire_ctx *ctx)
-{
-	const int swap_interval = (flags & DRM_MODE_PAGE_FLIP_ASYNC) ? 0 : 1;
-	struct drm_device *dev = crtc->dev;
-	struct nouveau_drm *drm = nouveau_drm(dev);
-	struct nouveau_bo *old_bo = nouveau_framebuffer(crtc->primary->fb)->nvbo;
-	struct nouveau_bo *new_bo = nouveau_framebuffer(fb)->nvbo;
-	struct nouveau_page_flip_state *s;
-	struct nouveau_channel *chan;
-	struct nouveau_cli *cli;
-	struct nouveau_fence *fence;
-	struct nv04_display *dispnv04 = nv04_display(dev);
-	int head = nouveau_crtc(crtc)->index;
-	int ret;
-
-	chan = drm->channel;
-	if (!chan)
-		return -ENODEV;
-	cli = (void *)chan->user.client;
-
-	s = kzalloc(sizeof(*s), GFP_KERNEL);
-	if (!s)
-		return -ENOMEM;
-
-	if (new_bo != old_bo) {
-		ret = nouveau_bo_pin(new_bo, TTM_PL_FLAG_VRAM, true);
-		if (ret)
-			goto fail_free;
-	}
-
-	mutex_lock(&cli->mutex);
-	ret = ttm_bo_reserve(&new_bo->bo, true, false, NULL);
-	if (ret)
-		goto fail_unpin;
-
-	/* synchronise rendering channel with the kernel's channel */
-	ret = nouveau_fence_sync(new_bo, chan, false, true);
-	if (ret) {
-		ttm_bo_unreserve(&new_bo->bo);
-		goto fail_unpin;
-	}
-
-	if (new_bo != old_bo) {
-		ttm_bo_unreserve(&new_bo->bo);
-
-		ret = ttm_bo_reserve(&old_bo->bo, true, false, NULL);
-		if (ret)
-			goto fail_unpin;
-	}
-
-	/* Initialize a page flip struct */
-	*s = (struct nouveau_page_flip_state)
-		{ { }, event, crtc, fb->format->cpp[0] * 8, fb->pitches[0],
-		  new_bo->bo.offset };
-
-	/* Keep vblanks on during flip, for the target crtc of this flip */
-	drm_crtc_vblank_get(crtc);
-
-	/* Emit a page flip */
-	if (swap_interval) {
-		ret = RING_SPACE(chan, 8);
-		if (ret)
-			goto fail_unreserve;
-
-		BEGIN_NV04(chan, NvSubImageBlit, 0x012c, 1);
-		OUT_RING  (chan, 0);
-		BEGIN_NV04(chan, NvSubImageBlit, 0x0134, 1);
-		OUT_RING  (chan, head);
-		BEGIN_NV04(chan, NvSubImageBlit, 0x0100, 1);
-		OUT_RING  (chan, 0);
-		BEGIN_NV04(chan, NvSubImageBlit, 0x0130, 1);
-		OUT_RING  (chan, 0);
-	}
-
-	nouveau_bo_ref(new_bo, &dispnv04->image[head]);
-
-	ret = nouveau_page_flip_emit(chan, old_bo, new_bo, s, &fence);
-	if (ret)
-		goto fail_unreserve;
-	mutex_unlock(&cli->mutex);
-
-	/* Update the crtc struct and cleanup */
-	crtc->primary->fb = fb;
-
-	nouveau_bo_fence(old_bo, fence, false);
-	ttm_bo_unreserve(&old_bo->bo);
-	if (old_bo != new_bo)
-		nouveau_bo_unpin(old_bo);
-	nouveau_fence_unref(&fence);
-	return 0;
-
-fail_unreserve:
-	drm_crtc_vblank_put(crtc);
-	ttm_bo_unreserve(&old_bo->bo);
-fail_unpin:
-	mutex_unlock(&cli->mutex);
-	if (old_bo != new_bo)
-		nouveau_bo_unpin(new_bo);
-fail_free:
-	kfree(s);
-	return ret;
-}
-
-int
-nouveau_finish_page_flip(struct nouveau_channel *chan,
-			 struct nouveau_page_flip_state *ps)
-{
-	struct nouveau_fence_chan *fctx = chan->fence;
-	struct nouveau_drm *drm = chan->drm;
-	struct drm_device *dev = drm->dev;
-	struct nouveau_page_flip_state *s;
-	unsigned long flags;
-
-	spin_lock_irqsave(&dev->event_lock, flags);
-
-	if (list_empty(&fctx->flip)) {
-		NV_ERROR(drm, "unexpected pageflip\n");
-		spin_unlock_irqrestore(&dev->event_lock, flags);
-		return -EINVAL;
-	}
-
-	s = list_first_entry(&fctx->flip, struct nouveau_page_flip_state, head);
-	if (s->event) {
-		drm_crtc_arm_vblank_event(s->crtc, s->event);
-	} else {
-		/* Give up ownership of vblank for page-flipped crtc */
-		drm_crtc_vblank_put(s->crtc);
-	}
-
-	list_del(&s->head);
-	if (ps)
-		*ps = *s;
-	kfree(s);
-
-	spin_unlock_irqrestore(&dev->event_lock, flags);
-	return 0;
-}
-
-int
-nouveau_flip_complete(struct nvif_notify *notify)
-{
-	struct nouveau_drm *drm = container_of(notify, typeof(*drm), flip);
-	struct nouveau_channel *chan = drm->channel;
-	struct nouveau_page_flip_state state;
-
-	if (!nouveau_finish_page_flip(chan, &state)) {
-		nv_set_crtc_base(drm->dev, drm_crtc_index(state.crtc),
-				 state.offset + state.crtc->y *
-				 state.pitch + state.crtc->x *
-				 state.bpp / 8);
-	}
-
-	return NVIF_NOTIFY_KEEP;
 }
 
 int

drivers/gpu/drm/nouveau/nouveau_display.h

···
 			    const struct drm_mode_fb_cmd2 *,
 			    struct nouveau_bo *, struct nouveau_framebuffer **);
 
-struct nouveau_page_flip_state {
-	struct list_head head;
-	struct drm_pending_vblank_event *event;
-	struct drm_crtc *crtc;
-	int bpp, pitch;
-	u64 offset;
-};
-
 struct nouveau_display {
 	void *priv;
 	void (*dtor)(struct drm_device *);
-	int  (*init)(struct drm_device *);
-	void (*fini)(struct drm_device *);
+	int  (*init)(struct drm_device *, bool resume, bool runtime);
+	void (*fini)(struct drm_device *, bool suspend);
 
 	struct nvif_disp disp;
 
···
 
 int  nouveau_display_create(struct drm_device *dev);
 void nouveau_display_destroy(struct drm_device *dev);
-int  nouveau_display_init(struct drm_device *dev);
+int  nouveau_display_init(struct drm_device *dev, bool resume, bool runtime);
 void nouveau_display_fini(struct drm_device *dev, bool suspend, bool runtime);
 int  nouveau_display_suspend(struct drm_device *dev, bool runtime);
 void nouveau_display_resume(struct drm_device *dev, bool runtime);
···
 bool  nouveau_display_scanoutpos(struct drm_device *, unsigned int,
 				 bool, int *, int *, ktime_t *,
 				 ktime_t *, const struct drm_display_mode *);
-
-int  nouveau_crtc_page_flip(struct drm_crtc *crtc, struct drm_framebuffer *fb,
-			    struct drm_pending_vblank_event *event,
-			    uint32_t page_flip_flags,
-			    struct drm_modeset_acquire_ctx *ctx);
-int  nouveau_finish_page_flip(struct nouveau_channel *,
-			      struct nouveau_page_flip_state *);
 
 int  nouveau_display_dumb_create(struct drm_file *, struct drm_device *,
 				 struct drm_mode_create_dumb *args);

drivers/gpu/drm/nouveau/nouveau_dmem.c

···
+/*
+ * Copyright 2018 Red Hat Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+#include "nouveau_dmem.h"
+#include "nouveau_drv.h"
+#include "nouveau_chan.h"
+#include "nouveau_dma.h"
+#include "nouveau_mem.h"
+#include "nouveau_bo.h"
+
+#include <nvif/class.h>
+#include <nvif/object.h>
+#include <nvif/if500b.h>
+#include <nvif/if900b.h>
+
+#include <linux/sched/mm.h>
+#include <linux/hmm.h>
+
+/*
+ * FIXME: this is ugly right now we are using TTM to allocate vram and we pin
+ * it in vram while in use. We likely want to overhaul memory management for
+ * nouveau to be more page like (not necessarily with system page size but a
+ * bigger page size) at lowest level and have some shim layer on top that would
+ * provide the same functionality as TTM.
+ */
+#define DMEM_CHUNK_SIZE (2UL << 20)
+#define DMEM_CHUNK_NPAGES (DMEM_CHUNK_SIZE >> PAGE_SHIFT)
+
+struct nouveau_migrate;
+
+enum nouveau_aper {
+	NOUVEAU_APER_VIRT,
+	NOUVEAU_APER_VRAM,
+	NOUVEAU_APER_HOST,
+};
+
+typedef int (*nouveau_migrate_copy_t)(struct nouveau_drm *drm, u64 npages,
+				      enum nouveau_aper, u64 dst_addr,
+				      enum nouveau_aper, u64 src_addr);
+
+struct nouveau_dmem_chunk {
+	struct list_head list;
+	struct nouveau_bo *bo;
+	struct nouveau_drm *drm;
+	unsigned long pfn_first;
+	unsigned long callocated;
+	unsigned long bitmap[BITS_TO_LONGS(DMEM_CHUNK_NPAGES)];
+	spinlock_t lock;
+};
+
+struct nouveau_dmem_migrate {
+	nouveau_migrate_copy_t copy_func;
+	struct nouveau_channel *chan;
+};
+
+struct nouveau_dmem {
+	struct hmm_devmem *devmem;
+	struct nouveau_dmem_migrate migrate;
+	struct list_head chunk_free;
+	struct list_head chunk_full;
+	struct list_head chunk_empty;
+	struct mutex mutex;
+};
+
+struct nouveau_dmem_fault {
+	struct nouveau_drm *drm;
+	struct nouveau_fence *fence;
+	dma_addr_t *dma;
+	unsigned long npages;
+};
+
+struct nouveau_migrate {
+	struct vm_area_struct *vma;
+	struct nouveau_drm *drm;
+	struct nouveau_fence *fence;
+	unsigned long npages;
+	dma_addr_t *dma;
+	unsigned long dma_nr;
+};
+
+static void
+nouveau_dmem_free(struct hmm_devmem *devmem, struct page *page)
+{
+	struct nouveau_dmem_chunk *chunk;
+	struct nouveau_drm *drm;
+	unsigned long idx;
+
+	chunk = (void *)hmm_devmem_page_get_drvdata(page);
+	idx = page_to_pfn(page) - chunk->pfn_first;
+	drm = chunk->drm;
+
+	/*
+	 * FIXME:
+	 *
+	 * This is really a bad example, we need to overhaul nouveau memory
+	 * management to be more page focus and allow lighter locking scheme
+	 * to be use in the process.
+	 */
+	spin_lock(&chunk->lock);
+	clear_bit(idx, chunk->bitmap);
+	WARN_ON(!chunk->callocated);
+	chunk->callocated--;
+	/*
+	 * FIXME when chunk->callocated reach 0 we should add the chunk to
+	 * a reclaim list so that it can be freed in case of memory pressure.
+	 */
+	spin_unlock(&chunk->lock);
+}
+
+static void
+nouveau_dmem_fault_alloc_and_copy(struct vm_area_struct *vma,
+				  const unsigned long *src_pfns,
+				  unsigned long *dst_pfns,
+				  unsigned long start,
+				  unsigned long end,
+				  void *private)
+{
+	struct nouveau_dmem_fault *fault = private;
+	struct nouveau_drm *drm = fault->drm;
+	struct device *dev = drm->dev->dev;
+	unsigned long addr, i, npages = 0;
+	nouveau_migrate_copy_t copy;
+	int ret;
+
+
+	/* First allocate new memory */
+	for (addr = start, i = 0; addr < end; addr += PAGE_SIZE, i++) {
+		struct page *dpage, *spage;
+
+		dst_pfns[i] = 0;
+		spage = migrate_pfn_to_page(src_pfns[i]);
+		if (!spage || !(src_pfns[i] & MIGRATE_PFN_MIGRATE))
+			continue;
+
+		dpage = hmm_vma_alloc_locked_page(vma, addr);
+		if (!dpage) {
+			dst_pfns[i] = MIGRATE_PFN_ERROR;
+			continue;
+		}
+
+		dst_pfns[i] = migrate_pfn(page_to_pfn(dpage)) |
+			      MIGRATE_PFN_LOCKED;
+		npages++;
+	}
+
+	/* Allocate storage for DMA addresses, so we can unmap later. */
+	fault->dma = kmalloc(sizeof(*fault->dma) * npages, GFP_KERNEL);
+	if (!fault->dma)
+		goto error;
+
+	/* Copy things over */
+	copy = drm->dmem->migrate.copy_func;
+	for (addr = start, i = 0; addr < end; addr += PAGE_SIZE, i++) {
+		struct nouveau_dmem_chunk *chunk;
+		struct page *spage, *dpage;
+		u64 src_addr, dst_addr;
+
+		dpage = migrate_pfn_to_page(dst_pfns[i]);
+		if (!dpage || dst_pfns[i] == MIGRATE_PFN_ERROR)
+			continue;
+
+		spage = migrate_pfn_to_page(src_pfns[i]);
+		if (!spage || !(src_pfns[i] & MIGRATE_PFN_MIGRATE)) {
+			dst_pfns[i] = MIGRATE_PFN_ERROR;
+			__free_page(dpage);
+			continue;
+		}
+
+		fault->dma[fault->npages] =
+			dma_map_page_attrs(dev, dpage, 0, PAGE_SIZE,
+					   PCI_DMA_BIDIRECTIONAL,
+					   DMA_ATTR_SKIP_CPU_SYNC);
+		if (dma_mapping_error(dev, fault->dma[fault->npages])) {
+			dst_pfns[i] = MIGRATE_PFN_ERROR;
+			__free_page(dpage);
+			continue;
+		}
+
+		dst_addr = fault->dma[fault->npages++];
+
+		chunk = (void *)hmm_devmem_page_get_drvdata(spage);
+		src_addr = page_to_pfn(spage) - chunk->pfn_first;
+		src_addr = (src_addr << PAGE_SHIFT) + chunk->bo->bo.offset;
+
+		ret = copy(drm, 1, NOUVEAU_APER_HOST, dst_addr,
+				   NOUVEAU_APER_VRAM, src_addr);
+		if (ret) {
+			dst_pfns[i] = MIGRATE_PFN_ERROR;
+			__free_page(dpage);
+			continue;
+		}
+	}
+
+	nouveau_fence_new(drm->dmem->migrate.chan, false, &fault->fence);
+
+	return;
+
+error:
+	for (addr = start, i = 0; addr < end; addr += PAGE_SIZE, ++i) {
+		struct page *page;
+
+		if (!dst_pfns[i] || dst_pfns[i] == MIGRATE_PFN_ERROR)
+			continue;
+
+		page = migrate_pfn_to_page(dst_pfns[i]);
+		dst_pfns[i] = MIGRATE_PFN_ERROR;
+		if (page == NULL)
+			continue;
+
+		__free_page(page);
+	}
+}
+
+void nouveau_dmem_fault_finalize_and_map(struct vm_area_struct *vma,
+					 const unsigned long *src_pfns,
+					 const unsigned long *dst_pfns,
+					 unsigned long start,
+					 unsigned long end,
+					 void *private)
+{
+	struct nouveau_dmem_fault *fault = private;
+	struct nouveau_drm *drm = fault->drm;
+
+	if (fault->fence) {
+		nouveau_fence_wait(fault->fence, true, false);
+		nouveau_fence_unref(&fault->fence);
+	} else {
+		/*
+		 * FIXME wait for channel to be IDLE before calling finalizing
+		 * the hmem object below (nouveau_migrate_hmem_fini()).
+		 */
+	}
+
+	while (fault->npages--) {
+		dma_unmap_page(drm->dev->dev, fault->dma[fault->npages],
+			       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
+	}
+	kfree(fault->dma);
+}
+
+static const struct migrate_vma_ops nouveau_dmem_fault_migrate_ops = {
+	.alloc_and_copy		= nouveau_dmem_fault_alloc_and_copy,
+	.finalize_and_map	= nouveau_dmem_fault_finalize_and_map,
+};
+
+static int
+nouveau_dmem_fault(struct hmm_devmem *devmem,
+		   struct vm_area_struct *vma,
+		   unsigned long addr,
+		   const struct page *page,
+		   unsigned int flags,
+		   pmd_t *pmdp)
+{
+	struct drm_device *drm_dev = dev_get_drvdata(devmem->device);
+	unsigned long src[1] = {0}, dst[1] = {0};
+	struct nouveau_dmem_fault fault = {0};
+	int ret;
+
+
+
+	/*
+	 * FIXME what we really want is to find some heuristic to migrate more
+	 * than just one page on CPU fault. When such fault happens it is very
+	 * likely that more surrounding page will CPU fault too.
+	 */
+	fault.drm = nouveau_drm(drm_dev);
+	ret = migrate_vma(&nouveau_dmem_fault_migrate_ops, vma, addr,
+			  addr + PAGE_SIZE, src, dst, &fault);
+	if (ret)
+		return VM_FAULT_SIGBUS;
+
+	if (dst[0] == MIGRATE_PFN_ERROR)
+		return VM_FAULT_SIGBUS;
+
+	return 0;
+}
+
+static const struct hmm_devmem_ops
+nouveau_dmem_devmem_ops = {
+	.free = nouveau_dmem_free,
+	.fault = nouveau_dmem_fault,
+};
+
+static int
+nouveau_dmem_chunk_alloc(struct nouveau_drm *drm)
+{
+	struct nouveau_dmem_chunk *chunk;
+	int ret;
+
+	if (drm->dmem == NULL)
+		return -EINVAL;
+
+	mutex_lock(&drm->dmem->mutex);
+	chunk = list_first_entry_or_null(&drm->dmem->chunk_empty,
+					 struct nouveau_dmem_chunk,
+					 list);
+	if (chunk == NULL) {
+		mutex_unlock(&drm->dmem->mutex);
+		return -ENOMEM;
+	}
+
+	list_del(&chunk->list);
+	mutex_unlock(&drm->dmem->mutex);
+
+	ret = nouveau_bo_new(&drm->client, DMEM_CHUNK_SIZE, 0,
+			     TTM_PL_FLAG_VRAM, 0, 0, NULL, NULL,
+			     &chunk->bo);
+	if (ret)
+		goto out;
+
+	ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false);
+	if (ret) {
+		nouveau_bo_ref(NULL, &chunk->bo);
+		goto out;
+	}
+
+	bitmap_zero(chunk->bitmap, DMEM_CHUNK_NPAGES);
+	spin_lock_init(&chunk->lock);
+
+out:
+	mutex_lock(&drm->dmem->mutex);
+	if (chunk->bo)
+		list_add(&chunk->list, &drm->dmem->chunk_empty);
+	else
+		list_add_tail(&chunk->list, &drm->dmem->chunk_empty);
+	mutex_unlock(&drm->dmem->mutex);
+
+	return ret;
+}
+
+static struct nouveau_dmem_chunk *
+nouveau_dmem_chunk_first_free_locked(struct nouveau_drm *drm)
+{
+	struct nouveau_dmem_chunk *chunk;
+
+	chunk = list_first_entry_or_null(&drm->dmem->chunk_free,
+					 struct nouveau_dmem_chunk,
+					 list);
+	if (chunk)
+		return chunk;
+
+	chunk = list_first_entry_or_null(&drm->dmem->chunk_empty,
+					 struct nouveau_dmem_chunk,
+					 list);
+	if (chunk->bo)
+		return chunk;
+
+	return NULL;
+}
+
+static int
+nouveau_dmem_pages_alloc(struct nouveau_drm *drm,
+			 unsigned long npages,
+			 unsigned long *pages)
+{
+	struct nouveau_dmem_chunk *chunk;
+	unsigned long c;
+	int ret;
+
+	memset(pages, 0xff, npages * sizeof(*pages));
+
+	mutex_lock(&drm->dmem->mutex);
+	for (c = 0; c < npages;) {
+		unsigned long i;
+
+		chunk = nouveau_dmem_chunk_first_free_locked(drm);
+		if (chunk == NULL) {
+			mutex_unlock(&drm->dmem->mutex);
+			ret = nouveau_dmem_chunk_alloc(drm);
+			if (ret) {
+				if (c)
+					break;
+				return ret;
+			}
+			continue;
+		}
+
+		spin_lock(&chunk->lock);
+		i = find_first_zero_bit(chunk->bitmap, DMEM_CHUNK_NPAGES);
+		while (i < DMEM_CHUNK_NPAGES && c < npages) {
+			pages[c] = chunk->pfn_first + i;
+			set_bit(i, chunk->bitmap);
+			chunk->callocated++;
+			c++;
+
+			i = find_next_zero_bit(chunk->bitmap,
+					DMEM_CHUNK_NPAGES, i);
+		}
+		spin_unlock(&chunk->lock);
+	}
+	mutex_unlock(&drm->dmem->mutex);
+
+	return 0;
+}
+
+static struct page *
+nouveau_dmem_page_alloc_locked(struct nouveau_drm *drm)
+{
+	unsigned long pfns[1];
+	struct page *page;
+	int ret;
+
+	/* FIXME stop all the miss-match API ... */
+	ret = nouveau_dmem_pages_alloc(drm, 1, pfns);
+	if (ret)
+		return NULL;
+
+	page = pfn_to_page(pfns[0]);
+	get_page(page);
+	lock_page(page);
+	return page;
+}
+
+static void
+nouveau_dmem_page_free_locked(struct nouveau_drm *drm, struct page *page)
+{
+	unlock_page(page);
+	put_page(page);
+}
+
+void
+nouveau_dmem_resume(struct nouveau_drm *drm)
+{
+	struct nouveau_dmem_chunk *chunk;
+	int ret;
+
+	if (drm->dmem == NULL)
+		return;
+
+	mutex_lock(&drm->dmem->mutex);
+	list_for_each_entry (chunk, &drm->dmem->chunk_free, list) {
+		ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false);
+		/* FIXME handle pin failure */
+		WARN_ON(ret);
+	}
+	list_for_each_entry (chunk, &drm->dmem->chunk_full, list) {
+		ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false);
+		/* FIXME handle pin failure */
+		WARN_ON(ret);
+	}
+	list_for_each_entry (chunk, &drm->dmem->chunk_empty, list) {
+		ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false);
+		/* FIXME handle pin failure */
+		WARN_ON(ret);
+	}
+	mutex_unlock(&drm->dmem->mutex);
+}
+
+void
+nouveau_dmem_suspend(struct nouveau_drm *drm)
+{
+	struct nouveau_dmem_chunk *chunk;
+
+	if (drm->dmem == NULL)
+		return;
+
+	mutex_lock(&drm->dmem->mutex);
+	list_for_each_entry (chunk, &drm->dmem->chunk_free, list) {
+		nouveau_bo_unpin(chunk->bo);
+	}
+	list_for_each_entry (chunk, &drm->dmem->chunk_full, list) {
+		nouveau_bo_unpin(chunk->bo);
+	}
+	list_for_each_entry (chunk, &drm->dmem->chunk_empty, list) {
+		nouveau_bo_unpin(chunk->bo);
+	}
+	mutex_unlock(&drm->dmem->mutex);
+}
+
+void
+nouveau_dmem_fini(struct nouveau_drm *drm)
+{
+	struct nouveau_dmem_chunk *chunk, *tmp;
+
+	if (drm->dmem == NULL)
+		return;
+
+	mutex_lock(&drm->dmem->mutex);
+
+	WARN_ON(!list_empty(&drm->dmem->chunk_free));
+	WARN_ON(!list_empty(&drm->dmem->chunk_full));
+
+	list_for_each_entry_safe (chunk, tmp, &drm->dmem->chunk_empty, list) {
+		if (chunk->bo) {
+			nouveau_bo_unpin(chunk->bo);
+			nouveau_bo_ref(NULL, &chunk->bo);
+		}
+		list_del(&chunk->list);
+		kfree(chunk);
+	}
+
+	mutex_unlock(&drm->dmem->mutex);
+}
+
+static int
+nvc0b5_migrate_copy(struct nouveau_drm *drm, u64 npages,
+		    enum nouveau_aper dst_aper, u64 dst_addr,
+		    enum nouveau_aper src_aper, u64 src_addr)
+{
+	struct nouveau_channel *chan = drm->dmem->migrate.chan;
+	u32 launch_dma = (1 << 9) /* MULTI_LINE_ENABLE. */ |
+			 (1 << 8) /* DST_MEMORY_LAYOUT_PITCH. */ |
+			 (1 << 7) /* SRC_MEMORY_LAYOUT_PITCH. */ |
+			 (1 << 2) /* FLUSH_ENABLE_TRUE. */ |
+			 (2 << 0) /* DATA_TRANSFER_TYPE_NON_PIPELINED. */;
+	int ret;
+
+	ret = RING_SPACE(chan, 13);
+	if (ret)
+		return ret;
+
+	if (src_aper != NOUVEAU_APER_VIRT) {
+		switch (src_aper) {
+		case NOUVEAU_APER_VRAM:
+			BEGIN_IMC0(chan, NvSubCopy, 0x0260, 0);
+			break;
+		case NOUVEAU_APER_HOST:
+			BEGIN_IMC0(chan, NvSubCopy, 0x0260, 1);
+			break;
+		default:
+			return -EINVAL;
+		}
+		launch_dma |= 0x00001000; /* SRC_TYPE_PHYSICAL. */
+	}
+
+	if (dst_aper != NOUVEAU_APER_VIRT) {
+		switch (dst_aper) {
+		case NOUVEAU_APER_VRAM:
+			BEGIN_IMC0(chan, NvSubCopy, 0x0264, 0);
+			break;
+		case NOUVEAU_APER_HOST:
+			BEGIN_IMC0(chan, NvSubCopy, 0x0264, 1);
+			break;
+		default:
+			return -EINVAL;
+		}
+		launch_dma |= 0x00002000; /* DST_TYPE_PHYSICAL. */
+	}
+
+	BEGIN_NVC0(chan, NvSubCopy, 0x0400, 8);
+	OUT_RING  (chan, upper_32_bits(src_addr));
+	OUT_RING  (chan, lower_32_bits(src_addr));
+	OUT_RING  (chan, upper_32_bits(dst_addr));
+	OUT_RING  (chan, lower_32_bits(dst_addr));
+	OUT_RING  (chan, PAGE_SIZE);
+	OUT_RING  (chan, PAGE_SIZE);
+	OUT_RING  (chan, PAGE_SIZE);
+	OUT_RING  (chan, npages);
+	BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
+	OUT_RING  (chan, launch_dma);
+	return 0;
+}
+
+static int
+nouveau_dmem_migrate_init(struct nouveau_drm *drm)
+{
+	switch (drm->ttm.copy.oclass) {
+	case PASCAL_DMA_COPY_A:
+	case PASCAL_DMA_COPY_B:
+	case  VOLTA_DMA_COPY_A:
+	case TURING_DMA_COPY_A:
+		drm->dmem->migrate.copy_func = nvc0b5_migrate_copy;
+		drm->dmem->migrate.chan = drm->ttm.chan;
+		return 0;
+	default:
+		break;
+	}
+	return -ENODEV;
+}
+
+void
+nouveau_dmem_init(struct nouveau_drm *drm)
+{
+	struct device *device = drm->dev->dev;
+	unsigned long i, size;
+	int ret;
+
+	/* This only make sense on PASCAL or newer */
+	if (drm->client.device.info.family < NV_DEVICE_INFO_V0_PASCAL)
+		return;
+
+	if (!(drm->dmem = kzalloc(sizeof(*drm->dmem), GFP_KERNEL)))
+		return;
+
+	mutex_init(&drm->dmem->mutex);
+	INIT_LIST_HEAD(&drm->dmem->chunk_free);
+	INIT_LIST_HEAD(&drm->dmem->chunk_full);
+	INIT_LIST_HEAD(&drm->dmem->chunk_empty);
+
+	size = ALIGN(drm->client.device.info.ram_user, DMEM_CHUNK_SIZE);
+
+	/* Initialize migration dma helpers before registering memory */
+	ret = nouveau_dmem_migrate_init(drm);
+	if (ret) {
+		kfree(drm->dmem);
+		drm->dmem = NULL;
+		return;
+	}
+
+	/*
+	 * FIXME we need some kind of policy to decide how much VRAM we
+	 * want to register with HMM. For now just register everything
+	 * and latter if we want to do thing like over commit then we
+	 * could revisit this.
+	 */
+	drm->dmem->devmem = hmm_devmem_add(&nouveau_dmem_devmem_ops,
+					   device, size);
+	if (drm->dmem->devmem == NULL) {
+		kfree(drm->dmem);
+		drm->dmem = NULL;
+		return;
+	}
+
+	for (i = 0; i < (size / DMEM_CHUNK_SIZE); ++i) {
+		struct nouveau_dmem_chunk *chunk;
+		struct page *page;
+		unsigned long j;
+
+		chunk = kzalloc(sizeof(*chunk), GFP_KERNEL);
+		if (chunk == NULL) {
+			nouveau_dmem_fini(drm);
+			return;
+		}
+
+		chunk->drm = drm;
+		chunk->pfn_first = drm->dmem->devmem->pfn_first;
+		chunk->pfn_first += (i * DMEM_CHUNK_NPAGES);
+		list_add_tail(&chunk->list, &drm->dmem->chunk_empty);
+
+		page = pfn_to_page(chunk->pfn_first);
+		for (j = 0; j < DMEM_CHUNK_NPAGES; ++j, ++page) {
+			hmm_devmem_page_set_drvdata(page, (long)chunk);
+		}
+	}
+
+	NV_INFO(drm, "DMEM: registered %ldMB of device memory\n", size >> 20);
+}
+
+static void
+nouveau_dmem_migrate_alloc_and_copy(struct vm_area_struct *vma,
+				    const unsigned long *src_pfns,
+				    unsigned long *dst_pfns,
+				    unsigned long start,
+				    unsigned long end,
+				    void *private)
+{
+	struct nouveau_migrate *migrate = private;
+	struct nouveau_drm *drm = migrate->drm;
+	struct device *dev = drm->dev->dev;
+	unsigned long addr, i, npages = 0;
+	nouveau_migrate_copy_t copy;
+	int ret;
+
+	/* First allocate new memory */
+	for (addr = start, i = 0; addr < end; addr += PAGE_SIZE, i++) {
+		struct page *dpage, *spage;
+
+		dst_pfns[i] = 0;
+		spage = migrate_pfn_to_page(src_pfns[i]);
+		if (!spage || !(src_pfns[i] & MIGRATE_PFN_MIGRATE))
+			continue;
+
+		dpage = nouveau_dmem_page_alloc_locked(drm);
+		if (!dpage)
+			continue;
+
+		dst_pfns[i] = migrate_pfn(page_to_pfn(dpage)) |
+			      MIGRATE_PFN_LOCKED |
+			      MIGRATE_PFN_DEVICE;
+		npages++;
+	}
+
+	if (!npages)
+		return;
+
+	/* Allocate storage for DMA addresses, so we can unmap later. */
+	migrate->dma = kmalloc(sizeof(*migrate->dma) * npages, GFP_KERNEL);
+	if (!migrate->dma)
+		goto error;
+
+	/* Copy things over */
+	copy = drm->dmem->migrate.copy_func;
+	for (addr = start, i = 0; addr < end; addr += PAGE_SIZE, i++) {
+		struct nouveau_dmem_chunk *chunk;
+		struct page *spage, *dpage;
+		u64 src_addr, dst_addr;
+
+		dpage = migrate_pfn_to_page(dst_pfns[i]);
+		if (!dpage || dst_pfns[i] == MIGRATE_PFN_ERROR)
+			continue;
+
+		chunk = (void *)hmm_devmem_page_get_drvdata(dpage);
+		dst_addr = page_to_pfn(dpage) - chunk->pfn_first;
+		dst_addr = (dst_addr << PAGE_SHIFT) + chunk->bo->bo.offset;
+
+		spage = migrate_pfn_to_page(src_pfns[i]);
+		if (!spage || !(src_pfns[i] & MIGRATE_PFN_MIGRATE)) {
+			nouveau_dmem_page_free_locked(drm, dpage);
+			dst_pfns[i] = 0;
+			continue;
+		}
+
+		migrate->dma[migrate->dma_nr] =
+			dma_map_page_attrs(dev, spage, 0, PAGE_SIZE,
+					   PCI_DMA_BIDIRECTIONAL,
+					   DMA_ATTR_SKIP_CPU_SYNC);
+		if (dma_mapping_error(dev, migrate->dma[migrate->dma_nr])) {
+			nouveau_dmem_page_free_locked(drm, dpage);
+			dst_pfns[i] = 0;
+			continue;
+		}
+
+		src_addr = migrate->dma[migrate->dma_nr++];
+
+		ret = copy(drm, 1, NOUVEAU_APER_VRAM, dst_addr,
+				   NOUVEAU_APER_HOST, src_addr);
+		if (ret) {
+			nouveau_dmem_page_free_locked(drm, dpage);
+			dst_pfns[i] = 0;
+			continue;
+		}
+	}
+
+	nouveau_fence_new(drm->dmem->migrate.chan, false, &migrate->fence);
+
+	return;
+
+error:
+	for (addr = start, i = 0; addr < end; addr += PAGE_SIZE, ++i) {
+		struct page *page;
+
+		if (!dst_pfns[i] || dst_pfns[i] == MIGRATE_PFN_ERROR)
+			continue;
+
+		page = migrate_pfn_to_page(dst_pfns[i]);
+		dst_pfns[i] = MIGRATE_PFN_ERROR;
+		if (page == NULL)
+			continue;
+
+		__free_page(page);
+	}
+}
+
+void nouveau_dmem_migrate_finalize_and_map(struct vm_area_struct *vma,
+					   const unsigned long *src_pfns,
+					   const unsigned long *dst_pfns,
+					   unsigned long start,
+					   unsigned long end,
+					   void *private)
+{
+	struct nouveau_migrate *migrate = private;
+	struct nouveau_drm *drm = migrate->drm;
+
+	if (migrate->fence) {
+		nouveau_fence_wait(migrate->fence, true, false);
+		nouveau_fence_unref(&migrate->fence);
+	} else {
+		/*
+		 * FIXME wait for channel to be IDLE before finalizing
+		 * the hmem object below (nouveau_migrate_hmem_fini()) ?
+		 */
+	}
+
+	while (migrate->dma_nr--) {
+		dma_unmap_page(drm->dev->dev, migrate->dma[migrate->dma_nr],
+			       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
+	}
+	kfree(migrate->dma);
+
+	/*
+	 * FIXME optimization: update GPU page table to point to newly
+	 * migrated memory.
+	 */
+}
+
+static const struct migrate_vma_ops nouveau_dmem_migrate_ops = {
+	.alloc_and_copy		= nouveau_dmem_migrate_alloc_and_copy,
+	.finalize_and_map	= nouveau_dmem_migrate_finalize_and_map,
+};
+
+int
+nouveau_dmem_migrate_vma(struct nouveau_drm *drm,
+			 struct vm_area_struct *vma,
+			 unsigned long start,
+			 unsigned long end)
+{
+	unsigned long *src_pfns, *dst_pfns, npages;
+	struct nouveau_migrate migrate = {0};
+	unsigned long i, c, max;
+	int ret = 0;
+
+	npages = (end - start) >> PAGE_SHIFT;
+	max = min(SG_MAX_SINGLE_ALLOC, npages);
+	src_pfns = kzalloc(sizeof(long) * max, GFP_KERNEL);
+	if (src_pfns == NULL)
+		return -ENOMEM;
+	dst_pfns = kzalloc(sizeof(long) * max, GFP_KERNEL);
+	if (dst_pfns == NULL) {
+		kfree(src_pfns);
+		return -ENOMEM;
+	}
+
+	migrate.drm = drm;
+	migrate.vma = vma;
+	migrate.npages = npages;
+	for (i = 0; i < npages; i += c) {
+		unsigned long next;
+
+		c = min(SG_MAX_SINGLE_ALLOC, npages);
+		next = start + (c << PAGE_SHIFT);
+		ret = migrate_vma(&nouveau_dmem_migrate_ops, vma, start,
+				  next, src_pfns, dst_pfns, &migrate);
+		if (ret)
+			goto out;
+		start = next;
+	}
+
+out:
+	kfree(dst_pfns);
+	kfree(src_pfns);
+	return ret;
+}
+
+static inline bool
+nouveau_dmem_page(struct nouveau_drm *drm, struct page *page)
+{
+	if (!is_device_private_page(page))
+		return false;
+
+	if (drm->dmem->devmem != page->pgmap->data)
+		return false;
+
+	return true;
+}
+
+void
+nouveau_dmem_convert_pfn(struct nouveau_drm *drm,
+			 struct hmm_range *range)
+{
+	unsigned long i, npages;
+
+	npages = (range->end - range->start) >> PAGE_SHIFT;
+	for (i = 0; i < npages; ++i) {
+		struct nouveau_dmem_chunk *chunk;
+		struct page *page;
+		uint64_t addr;
+
+		page = hmm_pfn_to_page(range, range->pfns[i]);
+		if (page == NULL)
+			continue;
+
+		if (!(range->pfns[i] & range->flags[HMM_PFN_DEVICE_PRIVATE])) {
+			continue;
+		}
+
+		if (!nouveau_dmem_page(drm, page)) {
+			WARN(1, "Some unknown device memory !\n");
+			range->pfns[i] = 0;
+			continue;
+		}
+
+		chunk = (void *)hmm_devmem_page_get_drvdata(page);
+		addr = page_to_pfn(page) - chunk->pfn_first;
+		addr = (addr + chunk->bo->bo.mem.start) << PAGE_SHIFT;
+
+		range->pfns[i] &= ((1UL << range->pfn_shift) - 1);
+		range->pfns[i] |= (addr >> PAGE_SHIFT) << range->pfn_shift;
+	}
+}

drivers/gpu/drm/nouveau/nouveau_dmem.h

···
+/*
+ * Copyright 2018 Red Hat Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+#ifndef __NOUVEAU_DMEM_H__
+#define __NOUVEAU_DMEM_H__
+#include <nvif/os.h>
+struct drm_device;
+struct drm_file;
+struct nouveau_drm;
+struct hmm_range;
+
+#if IS_ENABLED(CONFIG_DRM_NOUVEAU_SVM)
+void nouveau_dmem_init(struct nouveau_drm *);
+void nouveau_dmem_fini(struct nouveau_drm *);
+void nouveau_dmem_suspend(struct nouveau_drm *);
+void nouveau_dmem_resume(struct nouveau_drm *);
+
+int nouveau_dmem_migrate_vma(struct nouveau_drm *drm,
+			     struct vm_area_struct *vma,
+			     unsigned long start,
+			     unsigned long end);
+
+void nouveau_dmem_convert_pfn(struct nouveau_drm *drm,
+			      struct hmm_range *range);
+#else /* IS_ENABLED(CONFIG_DRM_NOUVEAU_SVM) */
+static inline void nouveau_dmem_init(struct nouveau_drm *drm) {}
+static inline void nouveau_dmem_fini(struct nouveau_drm *drm) {}
+static inline void nouveau_dmem_suspend(struct nouveau_drm *drm) {}
+static inline void nouveau_dmem_resume(struct nouveau_drm *drm) {}
+
+static inline int nouveau_dmem_migrate_vma(struct nouveau_drm *drm,
+					   struct vm_area_struct *vma,
+					   unsigned long start,
+					   unsigned long end)
+{
+	return 0;
+}
+
+static inline void nouveau_dmem_convert_pfn(struct nouveau_drm *drm,
+					    struct hmm_range *range) {}
+#endif /* IS_ENABLED(CONFIG_DRM_NOUVEAU_SVM) */
+#endif

drivers/gpu/drm/nouveau/nouveau_drm.c

···
 #include <nvif/class.h>
 #include <nvif/cl0002.h>
 #include <nvif/cla06f.h>
-#include <nvif/if0004.h>
 
 #include "nouveau_drv.h"
 #include "nouveau_dma.h"
···
 #include "nouveau_usif.h"
 #include "nouveau_connector.h"
 #include "nouveau_platform.h"
+#include "nouveau_svm.h"
+#include "nouveau_dmem.h"
 
 MODULE_PARM_DESC(config, "option string to pass to driver core");
 static char *nouveau_config;
···
 	WARN_ON(!list_empty(&cli->worker));
 
 	usif_client_fini(cli);
+	nouveau_vmm_fini(&cli->svm);
 	nouveau_vmm_fini(&cli->vmm);
 	nvif_mmu_fini(&cli->mmu);
 	nvif_device_fini(&cli->device);
···
 }
 
 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_notify_fini(&drm->flip);
 	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);
 }
···
 {
 	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.
 	 */
···
 		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);
+	/* 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 = 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);
-		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);
-			}
-
-			ret = nvif_notify_init(&drm->nvsw,
-					       nouveau_flip_complete,
-					       false, NV04_NVSW_NTFY_UEVENT,
-					       NULL, 0, 0, &drm->flip);
-			if (ret == 0)
-				ret = nvif_notify_get(&drm->flip);
-			if (ret) {
-				nouveau_accel_fini(drm);
-				return;
-			}
-		}
-
-		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;
-		}
+	/* Allocate channels we need to support various functions. */
+	nouveau_accel_gr_init(drm);
+	nouveau_accel_ce_init(drm);
 
-		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;
-		}
-	}
-
-
+	/* Initialise accelerated TTM buffer moves. */
 	nouveau_bo_move_init(drm);
 }
 
···
 	if (ret)
 		goto fail_bios;
 
+	nouveau_accel_init(drm);
+
 	ret = nouveau_display_create(dev);
 	if (ret)
 		goto fail_dispctor;
 
 	if (dev->mode_config.num_crtc) {
-		ret = nouveau_display_init(dev);
+		ret = nouveau_display_init(dev, false, false);
 		if (ret)
 			goto fail_dispinit;
 	}
 
 	nouveau_debugfs_init(drm);
 	nouveau_hwmon_init(dev);
-	nouveau_accel_init(drm);
+	nouveau_svm_init(drm);
+	nouveau_dmem_init(drm);
 	nouveau_fbcon_init(dev);
 	nouveau_led_init(dev);
 
···
 fail_dispinit:
 	nouveau_display_destroy(dev);
 fail_dispctor:
+	nouveau_accel_fini(drm);
 	nouveau_bios_takedown(dev);
 fail_bios:
 	nouveau_ttm_fini(drm);
···
 
 	nouveau_led_fini(dev);
 	nouveau_fbcon_fini(dev);
-	nouveau_accel_fini(drm);
+	nouveau_dmem_fini(drm);
+	nouveau_svm_fini(drm);
 	nouveau_hwmon_fini(dev);
 	nouveau_debugfs_fini(drm);
 
···
 		nouveau_display_fini(dev, false, false);
 	nouveau_display_destroy(dev);
 
+	nouveau_accel_fini(drm);
 	nouveau_bios_takedown(dev);
 
 	nouveau_ttm_fini(drm);
···
 	struct nouveau_drm *drm = nouveau_drm(dev);
 	int ret;
 
+	nouveau_svm_suspend(drm);
+	nouveau_dmem_suspend(drm);
 	nouveau_led_suspend(dev);
 
 	if (dev->mode_config.num_crtc) {
···
 	}
 
 	nouveau_led_resume(dev);
-
+	nouveau_dmem_resume(drm);
+	nouveau_svm_resume(drm);
 	return 0;
 }
 
···
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GROBJ_ALLOC, nouveau_abi16_ioctl_grobj_alloc, DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(NOUVEAU_NOTIFIEROBJ_ALLOC, nouveau_abi16_ioctl_notifierobj_alloc, DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GPUOBJ_FREE, nouveau_abi16_ioctl_gpuobj_free, DRM_AUTH|DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(NOUVEAU_SVM_INIT, nouveau_svmm_init, DRM_AUTH|DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(NOUVEAU_SVM_BIND, nouveau_svmm_bind, DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_NEW, nouveau_gem_ioctl_new, DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_PUSHBUF, nouveau_gem_ioctl_pushbuf, DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_CPU_PREP, nouveau_gem_ioctl_cpu_prep, DRM_AUTH|DRM_RENDER_ALLOW),

drivers/gpu/drm/nouveau/nouveau_drv.h

···
 	struct nvif_device device;
 	struct nvif_mmu mmu;
 	struct nouveau_vmm vmm;
+	struct nouveau_vmm svm;
 	const struct nvif_mclass *mem;
 
 	struct list_head head;
···
 	struct nouveau_fbdev *fbcon;
 	struct nvif_object nvsw;
 	struct nvif_object ntfy;
-	struct nvif_notify flip;
 
 	/* nv10-nv40 tiling regions */
 	struct {
···
 	bool have_disp_power_ref;
 
 	struct dev_pm_domain vga_pm_domain;
+
+	struct nouveau_svm *svm;
+
+	struct nouveau_dmem *dmem;
 };
 
 static inline struct nouveau_drm *

drivers/gpu/drm/nouveau/nouveau_fbcon.c

···
 
 	chan = nouveau_nofbaccel ? NULL : drm->channel;
 	if (chan && device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
-		ret = nouveau_vma_new(nvbo, &drm->client.vmm, &fb->vma);
+		ret = nouveau_vma_new(nvbo, chan->vmm, &fb->vma);
 		if (ret) {
 			NV_ERROR(drm, "failed to map fb into chan: %d\n", ret);
 			chan = NULL;

drivers/gpu/drm/nouveau/nouveau_fence.h

···
 int nv84_fence_create(struct nouveau_drm *);
 int nvc0_fence_create(struct nouveau_drm *);
 
-int nouveau_flip_complete(struct nvif_notify *);
-
 struct nv84_fence_chan {
 	struct nouveau_fence_chan base;
 	struct nouveau_vma *vma;

drivers/gpu/drm/nouveau/nouveau_gem.c

···
 	struct nouveau_bo *nvbo = nouveau_gem_object(gem);
 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
 	struct device *dev = drm->dev->dev;
+	struct nouveau_vmm *vmm = cli->svm.cli ? &cli->svm : &cli->vmm;
 	struct nouveau_vma *vma;
 	int ret;
 
-	if (cli->vmm.vmm.object.oclass < NVIF_CLASS_VMM_NV50)
+	if (vmm->vmm.object.oclass < NVIF_CLASS_VMM_NV50)
 		return 0;
 
 	ret = ttm_bo_reserve(&nvbo->bo, false, false, NULL);
···
 	if (ret < 0 && ret != -EACCES)
 		goto out;
 
-	ret = nouveau_vma_new(nvbo, &cli->vmm, &vma);
+	ret = nouveau_vma_new(nvbo, vmm, &vma);
 	pm_runtime_mark_last_busy(dev);
 	pm_runtime_put_autosuspend(dev);
 out:
···
 	struct nouveau_bo *nvbo = nouveau_gem_object(gem);
 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
 	struct device *dev = drm->dev->dev;
+	struct nouveau_vmm *vmm = cli->svm.cli ? &cli->svm : & cli->vmm;
 	struct nouveau_vma *vma;
 	int ret;
 
-	if (cli->vmm.vmm.object.oclass < NVIF_CLASS_VMM_NV50)
+	if (vmm->vmm.object.oclass < NVIF_CLASS_VMM_NV50)
 		return;
 
 	ret = ttm_bo_reserve(&nvbo->bo, false, false, NULL);
 	if (ret)
 		return;
 
-	vma = nouveau_vma_find(nvbo, &cli->vmm);
+	vma = nouveau_vma_find(nvbo, vmm);
 	if (vma) {
 		if (--vma->refs == 0) {
 			ret = pm_runtime_get_sync(dev);
···
 {
 	struct nouveau_cli *cli = nouveau_cli(file_priv);
 	struct nouveau_bo *nvbo = nouveau_gem_object(gem);
+	struct nouveau_vmm *vmm = cli->svm.cli ? &cli->svm : &cli->vmm;
 	struct nouveau_vma *vma;
 
 	if (is_power_of_2(nvbo->valid_domains))
···
 	else
 		rep->domain = NOUVEAU_GEM_DOMAIN_VRAM;
 	rep->offset = nvbo->bo.offset;
-	if (cli->vmm.vmm.object.oclass >= NVIF_CLASS_VMM_NV50) {
-		vma = nouveau_vma_find(nvbo, &cli->vmm);
+	if (vmm->vmm.object.oclass >= NVIF_CLASS_VMM_NV50) {
+		vma = nouveau_vma_find(nvbo, vmm);
 		if (!vma)
 			return -EINVAL;
 
···
 };
 
 static void
-validate_fini_no_ticket(struct validate_op *op, struct nouveau_fence *fence,
+validate_fini_no_ticket(struct validate_op *op, struct nouveau_channel *chan,
+			struct nouveau_fence *fence,
 			struct drm_nouveau_gem_pushbuf_bo *pbbo)
 {
 	struct nouveau_bo *nvbo;
···
 		b = &pbbo[nvbo->pbbo_index];
 
 		if (likely(fence)) {
-			struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
-			struct nouveau_vma *vma;
-
 			nouveau_bo_fence(nvbo, fence, !!b->write_domains);
 
-			if (drm->client.vmm.vmm.object.oclass >= NVIF_CLASS_VMM_NV50) {
-				vma = (void *)(unsigned long)b->user_priv;
+			if (chan->vmm->vmm.object.oclass >= NVIF_CLASS_VMM_NV50) {
+				struct nouveau_vma *vma =
+					(void *)(unsigned long)b->user_priv;
 				nouveau_fence_unref(&vma->fence);
 				dma_fence_get(&fence->base);
 				vma->fence = fence;
···
 }
 
 static void
-validate_fini(struct validate_op *op, struct nouveau_fence *fence,
+validate_fini(struct validate_op *op, struct nouveau_channel *chan,
+	      struct nouveau_fence *fence,
 	      struct drm_nouveau_gem_pushbuf_bo *pbbo)
 {
-	validate_fini_no_ticket(op, fence, pbbo);
+	validate_fini_no_ticket(op, chan, fence, pbbo);
 	ww_acquire_fini(&op->ticket);
 }
 
···
 			list_splice_tail_init(&vram_list, &op->list);
 			list_splice_tail_init(&gart_list, &op->list);
 			list_splice_tail_init(&both_list, &op->list);
-			validate_fini_no_ticket(op, NULL, NULL);
+			validate_fini_no_ticket(op, chan, NULL, NULL);
 			if (unlikely(ret == -EDEADLK)) {
 				ret = ttm_bo_reserve_slowpath(&nvbo->bo, true,
 							      &op->ticket);
···
 			}
 		}
 
-		if (cli->vmm.vmm.object.oclass >= NVIF_CLASS_VMM_NV50) {
-			struct nouveau_vmm *vmm = &cli->vmm;
+		if (chan->vmm->vmm.object.oclass >= NVIF_CLASS_VMM_NV50) {
+			struct nouveau_vmm *vmm = chan->vmm;
 			struct nouveau_vma *vma = nouveau_vma_find(nvbo, vmm);
 			if (!vma) {
 				NV_PRINTK(err, cli, "vma not found!\n");
···
 	list_splice_tail(&gart_list, &op->list);
 	list_splice_tail(&both_list, &op->list);
 	if (ret)
-		validate_fini(op, NULL, NULL);
+		validate_fini(op, chan, NULL, NULL);
 	return ret;
 
 }
···
 	if (unlikely(ret < 0)) {
 		if (ret != -ERESTARTSYS)
 			NV_PRINTK(err, cli, "validating bo list\n");
-		validate_fini(op, NULL, NULL);
+		validate_fini(op, chan, NULL, NULL);
 		return ret;
 	}
 	*apply_relocs = ret;
···
 	}
 
 out:
-	validate_fini(&op, fence, bo);
+	validate_fini(&op, chan, fence, bo);
 	nouveau_fence_unref(&fence);
 
 out_prevalid:

drivers/gpu/drm/nouveau/nouveau_svm.c

···
+/*
+ * Copyright 2018 Red Hat Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+#include "nouveau_svm.h"
+#include "nouveau_drv.h"
+#include "nouveau_chan.h"
+#include "nouveau_dmem.h"
+
+#include <nvif/notify.h>
+#include <nvif/object.h>
+#include <nvif/vmm.h>
+
+#include <nvif/class.h>
+#include <nvif/clb069.h>
+#include <nvif/ifc00d.h>
+
+#include <linux/sched/mm.h>
+#include <linux/sort.h>
+#include <linux/hmm.h>
+
+struct nouveau_svm {
+	struct nouveau_drm *drm;
+	struct mutex mutex;
+	struct list_head inst;
+
+	struct nouveau_svm_fault_buffer {
+		int id;
+		struct nvif_object object;
+		u32 entries;
+		u32 getaddr;
+		u32 putaddr;
+		u32 get;
+		u32 put;
+		struct nvif_notify notify;
+
+		struct nouveau_svm_fault {
+			u64 inst;
+			u64 addr;
+			u64 time;
+			u32 engine;
+			u8  gpc;
+			u8  hub;
+			u8  access;
+			u8  client;
+			u8  fault;
+			struct nouveau_svmm *svmm;
+		} **fault;
+		int fault_nr;
+	} buffer[1];
+};
+
+#define SVM_DBG(s,f,a...) NV_DEBUG((s)->drm, "svm: "f"\n", ##a)
+#define SVM_ERR(s,f,a...) NV_WARN((s)->drm, "svm: "f"\n", ##a)
+
+struct nouveau_ivmm {
+	struct nouveau_svmm *svmm;
+	u64 inst;
+	struct list_head head;
+};
+
+static struct nouveau_ivmm *
+nouveau_ivmm_find(struct nouveau_svm *svm, u64 inst)
+{
+	struct nouveau_ivmm *ivmm;
+	list_for_each_entry(ivmm, &svm->inst, head) {
+		if (ivmm->inst == inst)
+			return ivmm;
+	}
+	return NULL;
+}
+
+struct nouveau_svmm {
+	struct nouveau_vmm *vmm;
+	struct {
+		unsigned long start;
+		unsigned long limit;
+	} unmanaged;
+
+	struct mutex mutex;
+
+	struct mm_struct *mm;
+	struct hmm_mirror mirror;
+};
+
+#define SVMM_DBG(s,f,a...)                                                     \
+	NV_DEBUG((s)->vmm->cli->drm, "svm-%p: "f"\n", (s), ##a)
+#define SVMM_ERR(s,f,a...)                                                     \
+	NV_WARN((s)->vmm->cli->drm, "svm-%p: "f"\n", (s), ##a)
+
+int
+nouveau_svmm_bind(struct drm_device *dev, void *data,
+		  struct drm_file *file_priv)
+{
+	struct nouveau_cli *cli = nouveau_cli(file_priv);
+	struct drm_nouveau_svm_bind *args = data;
+	unsigned target, cmd, priority;
+	unsigned long addr, end, size;
+	struct mm_struct *mm;
+
+	args->va_start &= PAGE_MASK;
+	args->va_end &= PAGE_MASK;
+
+	/* Sanity check arguments */
+	if (args->reserved0 || args->reserved1)
+		return -EINVAL;
+	if (args->header & (~NOUVEAU_SVM_BIND_VALID_MASK))
+		return -EINVAL;
+	if (args->va_start >= args->va_end)
+		return -EINVAL;
+	if (!args->npages)
+		return -EINVAL;
+
+	cmd = args->header >> NOUVEAU_SVM_BIND_COMMAND_SHIFT;
+	cmd &= NOUVEAU_SVM_BIND_COMMAND_MASK;
+	switch (cmd) {
+	case NOUVEAU_SVM_BIND_COMMAND__MIGRATE:
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	priority = args->header >> NOUVEAU_SVM_BIND_PRIORITY_SHIFT;
+	priority &= NOUVEAU_SVM_BIND_PRIORITY_MASK;
+
+	/* FIXME support CPU target ie all target value < GPU_VRAM */
+	target = args->header >> NOUVEAU_SVM_BIND_TARGET_SHIFT;
+	target &= NOUVEAU_SVM_BIND_TARGET_MASK;
+	switch (target) {
+	case NOUVEAU_SVM_BIND_TARGET__GPU_VRAM:
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	/*
+	 * FIXME: For now refuse non 0 stride, we need to change the migrate
+	 * kernel function to handle stride to avoid to create a mess within
+	 * each device driver.
+	 */
+	if (args->stride)
+		return -EINVAL;
+
+	size = ((unsigned long)args->npages) << PAGE_SHIFT;
+	if ((args->va_start + size) <= args->va_start)
+		return -EINVAL;
+	if ((args->va_start + size) > args->va_end)
+		return -EINVAL;
+
+	/*
+	 * Ok we are ask to do something sane, for now we only support migrate
+	 * commands but we will add things like memory policy (what to do on
+	 * page fault) and maybe some other commands.
+	 */
+
+	mm = get_task_mm(current);
+	down_read(&mm->mmap_sem);
+
+	for (addr = args->va_start, end = args->va_start + size; addr < end;) {
+		struct vm_area_struct *vma;
+		unsigned long next;
+
+		vma = find_vma_intersection(mm, addr, end);
+		if (!vma)
+			break;
+
+		next = min(vma->vm_end, end);
+		/* This is a best effort so we ignore errors */
+		nouveau_dmem_migrate_vma(cli->drm, vma, addr, next);
+		addr = next;
+	}
+
+	/*
+	 * FIXME Return the number of page we have migrated, again we need to
+	 * update the migrate API to return that information so that we can
+	 * report it to user space.
+	 */
+	args->result = 0;
+
+	up_read(&mm->mmap_sem);
+	mmput(mm);
+
+	return 0;
+}
+
+/* Unlink channel instance from SVMM. */
+void
+nouveau_svmm_part(struct nouveau_svmm *svmm, u64 inst)
+{
+	struct nouveau_ivmm *ivmm;
+	if (svmm) {
+		mutex_lock(&svmm->vmm->cli->drm->svm->mutex);
+		ivmm = nouveau_ivmm_find(svmm->vmm->cli->drm->svm, inst);
+		if (ivmm) {
+			list_del(&ivmm->head);
+			kfree(ivmm);
+		}
+		mutex_unlock(&svmm->vmm->cli->drm->svm->mutex);
+	}
+}
+
+/* Link channel instance to SVMM. */
+int
+nouveau_svmm_join(struct nouveau_svmm *svmm, u64 inst)
+{
+	struct nouveau_ivmm *ivmm;
+	if (svmm) {
+		if (!(ivmm = kmalloc(sizeof(*ivmm), GFP_KERNEL)))
+			return -ENOMEM;
+		ivmm->svmm = svmm;
+		ivmm->inst = inst;
+
+		mutex_lock(&svmm->vmm->cli->drm->svm->mutex);
+		list_add(&ivmm->head, &svmm->vmm->cli->drm->svm->inst);
+		mutex_unlock(&svmm->vmm->cli->drm->svm->mutex);
+	}
+	return 0;
+}
+
+/* Invalidate SVMM address-range on GPU. */
+static void
+nouveau_svmm_invalidate(struct nouveau_svmm *svmm, u64 start, u64 limit)
+{
+	if (limit > start) {
+		bool super = svmm->vmm->vmm.object.client->super;
+		svmm->vmm->vmm.object.client->super = true;
+		nvif_object_mthd(&svmm->vmm->vmm.object, NVIF_VMM_V0_PFNCLR,
+				 &(struct nvif_vmm_pfnclr_v0) {
+					.addr = start,
+					.size = limit - start,
+				 }, sizeof(struct nvif_vmm_pfnclr_v0));
+		svmm->vmm->vmm.object.client->super = super;
+	}
+}
+
+static int
+nouveau_svmm_sync_cpu_device_pagetables(struct hmm_mirror *mirror,
+					const struct hmm_update *update)
+{
+	struct nouveau_svmm *svmm = container_of(mirror, typeof(*svmm), mirror);
+	unsigned long start = update->start;
+	unsigned long limit = update->end;
+
+	if (!update->blockable)
+		return -EAGAIN;
+
+	SVMM_DBG(svmm, "invalidate %016lx-%016lx", start, limit);
+
+	mutex_lock(&svmm->mutex);
+	if (limit > svmm->unmanaged.start && start < svmm->unmanaged.limit) {
+		if (start < svmm->unmanaged.start) {
+			nouveau_svmm_invalidate(svmm, start,
+						svmm->unmanaged.limit);
+		}
+		start = svmm->unmanaged.limit;
+	}
+
+	nouveau_svmm_invalidate(svmm, start, limit);
+	mutex_unlock(&svmm->mutex);
+	return 0;
+}
+
+static void
+nouveau_svmm_release(struct hmm_mirror *mirror)
+{
+}
+
+static const struct hmm_mirror_ops
+nouveau_svmm = {
+	.sync_cpu_device_pagetables = nouveau_svmm_sync_cpu_device_pagetables,
+	.release = nouveau_svmm_release,
+};
+
+void
+nouveau_svmm_fini(struct nouveau_svmm **psvmm)
+{
+	struct nouveau_svmm *svmm = *psvmm;
+	if (svmm) {
+		hmm_mirror_unregister(&svmm->mirror);
+		kfree(*psvmm);
+		*psvmm = NULL;
+	}
+}
+
+int
+nouveau_svmm_init(struct drm_device *dev, void *data,
+		  struct drm_file *file_priv)
+{
+	struct nouveau_cli *cli = nouveau_cli(file_priv);
+	struct nouveau_svmm *svmm;
+	struct drm_nouveau_svm_init *args = data;
+	int ret;
+
+	/* Allocate tracking for SVM-enabled VMM. */
+	if (!(svmm = kzalloc(sizeof(*svmm), GFP_KERNEL)))
+		return -ENOMEM;
+	svmm->vmm = &cli->svm;
+	svmm->unmanaged.start = args->unmanaged_addr;
+	svmm->unmanaged.limit = args->unmanaged_addr + args->unmanaged_size;
+	mutex_init(&svmm->mutex);
+
+	/* Check that SVM isn't already enabled for the client. */
+	mutex_lock(&cli->mutex);
+	if (cli->svm.cli) {
+		ret = -EBUSY;
+		goto done;
+	}
+
+	/* Allocate a new GPU VMM that can support SVM (managed by the
+	 * client, with replayable faults enabled).
+	 *
+	 * All future channel/memory allocations will make use of this
+	 * VMM instead of the standard one.
+	 */
+	ret = nvif_vmm_init(&cli->mmu, cli->vmm.vmm.object.oclass, true,
+			    args->unmanaged_addr, args->unmanaged_size,
+			    &(struct gp100_vmm_v0) {
+				.fault_replay = true,
+			    }, sizeof(struct gp100_vmm_v0), &cli->svm.vmm);
+	if (ret)
+		goto done;
+
+	/* Enable HMM mirroring of CPU address-space to VMM. */
+	svmm->mm = get_task_mm(current);
+	down_write(&svmm->mm->mmap_sem);
+	svmm->mirror.ops = &nouveau_svmm;
+	ret = hmm_mirror_register(&svmm->mirror, svmm->mm);
+	if (ret == 0) {
+		cli->svm.svmm = svmm;
+		cli->svm.cli = cli;
+	}
+	up_write(&svmm->mm->mmap_sem);
+	mmput(svmm->mm);
+
+done:
+	if (ret)
+		nouveau_svmm_fini(&svmm);
+	mutex_unlock(&cli->mutex);
+	return ret;
+}
+
+static const u64
+nouveau_svm_pfn_flags[HMM_PFN_FLAG_MAX] = {
+	[HMM_PFN_VALID         ] = NVIF_VMM_PFNMAP_V0_V,
+	[HMM_PFN_WRITE         ] = NVIF_VMM_PFNMAP_V0_W,
+	[HMM_PFN_DEVICE_PRIVATE] = NVIF_VMM_PFNMAP_V0_VRAM,
+};
+
+static const u64
+nouveau_svm_pfn_values[HMM_PFN_VALUE_MAX] = {
+	[HMM_PFN_ERROR  ] = ~NVIF_VMM_PFNMAP_V0_V,
+	[HMM_PFN_NONE   ] =  NVIF_VMM_PFNMAP_V0_NONE,
+	[HMM_PFN_SPECIAL] = ~NVIF_VMM_PFNMAP_V0_V,
+};
+
+/* Issue fault replay for GPU to retry accesses that faulted previously. */
+static void
+nouveau_svm_fault_replay(struct nouveau_svm *svm)
+{
+	SVM_DBG(svm, "replay");
+	WARN_ON(nvif_object_mthd(&svm->drm->client.vmm.vmm.object,
+				 GP100_VMM_VN_FAULT_REPLAY,
+				 &(struct gp100_vmm_fault_replay_vn) {},
+				 sizeof(struct gp100_vmm_fault_replay_vn)));
+}
+
+/* Cancel a replayable fault that could not be handled.
+ *
+ * Cancelling the fault will trigger recovery to reset the engine
+ * and kill the offending channel (ie. GPU SIGSEGV).
+ */
+static void
+nouveau_svm_fault_cancel(struct nouveau_svm *svm,
+			 u64 inst, u8 hub, u8 gpc, u8 client)
+{
+	SVM_DBG(svm, "cancel %016llx %d %02x %02x", inst, hub, gpc, client);
+	WARN_ON(nvif_object_mthd(&svm->drm->client.vmm.vmm.object,
+				 GP100_VMM_VN_FAULT_CANCEL,
+				 &(struct gp100_vmm_fault_cancel_v0) {
+					.hub = hub,
+					.gpc = gpc,
+					.client = client,
+					.inst = inst,
+				 }, sizeof(struct gp100_vmm_fault_cancel_v0)));
+}
+
+static void
+nouveau_svm_fault_cancel_fault(struct nouveau_svm *svm,
+			       struct nouveau_svm_fault *fault)
+{
+	nouveau_svm_fault_cancel(svm, fault->inst,
+				      fault->hub,
+				      fault->gpc,
+				      fault->client);
+}
+
+static int
+nouveau_svm_fault_cmp(const void *a, const void *b)
+{
+	const struct nouveau_svm_fault *fa = *(struct nouveau_svm_fault **)a;
+	const struct nouveau_svm_fault *fb = *(struct nouveau_svm_fault **)b;
+	int ret;
+	if ((ret = (s64)fa->inst - fb->inst))
+		return ret;
+	if ((ret = (s64)fa->addr - fb->addr))
+		return ret;
+	/*XXX: atomic? */
+	return (fa->access == 0 || fa->access == 3) -
+	       (fb->access == 0 || fb->access == 3);
+}
+
+static void
+nouveau_svm_fault_cache(struct nouveau_svm *svm,
+			struct nouveau_svm_fault_buffer *buffer, u32 offset)
+{
+	struct nvif_object *memory = &buffer->object;
+	const u32 instlo = nvif_rd32(memory, offset + 0x00);
+	const u32 insthi = nvif_rd32(memory, offset + 0x04);
+	const u32 addrlo = nvif_rd32(memory, offset + 0x08);
+	const u32 addrhi = nvif_rd32(memory, offset + 0x0c);
+	const u32 timelo = nvif_rd32(memory, offset + 0x10);
+	const u32 timehi = nvif_rd32(memory, offset + 0x14);
+	const u32 engine = nvif_rd32(memory, offset + 0x18);
+	const u32   info = nvif_rd32(memory, offset + 0x1c);
+	const u64   inst = (u64)insthi << 32 | instlo;
+	const u8     gpc = (info & 0x1f000000) >> 24;
+	const u8     hub = (info & 0x00100000) >> 20;
+	const u8  client = (info & 0x00007f00) >> 8;
+	struct nouveau_svm_fault *fault;
+
+	//XXX: i think we're supposed to spin waiting */
+	if (WARN_ON(!(info & 0x80000000)))
+		return;
+
+	nvif_mask(memory, offset + 0x1c, 0x80000000, 0x00000000);
+
+	if (!buffer->fault[buffer->fault_nr]) {
+		fault = kmalloc(sizeof(*fault), GFP_KERNEL);
+		if (WARN_ON(!fault)) {
+			nouveau_svm_fault_cancel(svm, inst, hub, gpc, client);
+			return;
+		}
+		buffer->fault[buffer->fault_nr] = fault;
+	}
+
+	fault = buffer->fault[buffer->fault_nr++];
+	fault->inst   = inst;
+	fault->addr   = (u64)addrhi << 32 | addrlo;
+	fault->time   = (u64)timehi << 32 | timelo;
+	fault->engine = engine;
+	fault->gpc    = gpc;
+	fault->hub    = hub;
+	fault->access = (info & 0x000f0000) >> 16;
+	fault->client = client;
+	fault->fault  = (info & 0x0000001f);
+
+	SVM_DBG(svm, "fault %016llx %016llx %02x",
+		fault->inst, fault->addr, fault->access);
+}
+
+static int
+nouveau_svm_fault(struct nvif_notify *notify)
+{
+	struct nouveau_svm_fault_buffer *buffer =
+		container_of(notify, typeof(*buffer), notify);
+	struct nouveau_svm *svm =
+		container_of(buffer, typeof(*svm), buffer[buffer->id]);
+	struct nvif_object *device = &svm->drm->client.device.object;
+	struct nouveau_svmm *svmm;
+	struct {
+		struct {
+			struct nvif_ioctl_v0 i;
+			struct nvif_ioctl_mthd_v0 m;
+			struct nvif_vmm_pfnmap_v0 p;
+		} i;
+		u64 phys[16];
+	} args;
+	struct hmm_range range;
+	struct vm_area_struct *vma;
+	u64 inst, start, limit;
+	int fi, fn, pi, fill;
+	int replay = 0, ret;
+
+	/* Parse available fault buffer entries into a cache, and update
+	 * the GET pointer so HW can reuse the entries.
+	 */
+	SVM_DBG(svm, "fault handler");
+	if (buffer->get == buffer->put) {
+		buffer->put = nvif_rd32(device, buffer->putaddr);
+		buffer->get = nvif_rd32(device, buffer->getaddr);
+		if (buffer->get == buffer->put)
+			return NVIF_NOTIFY_KEEP;
+	}
+	buffer->fault_nr = 0;
+
+	SVM_DBG(svm, "get %08x put %08x", buffer->get, buffer->put);
+	while (buffer->get != buffer->put) {
+		nouveau_svm_fault_cache(svm, buffer, buffer->get * 0x20);
+		if (++buffer->get == buffer->entries)
+			buffer->get = 0;
+	}
+	nvif_wr32(device, buffer->getaddr, buffer->get);
+	SVM_DBG(svm, "%d fault(s) pending", buffer->fault_nr);
+
+	/* Sort parsed faults by instance pointer to prevent unnecessary
+	 * instance to SVMM translations, followed by address and access
+	 * type to reduce the amount of work when handling the faults.
+	 */
+	sort(buffer->fault, buffer->fault_nr, sizeof(*buffer->fault),
+	     nouveau_svm_fault_cmp, NULL);
+
+	/* Lookup SVMM structure for each unique instance pointer. */
+	mutex_lock(&svm->mutex);
+	for (fi = 0, svmm = NULL; fi < buffer->fault_nr; fi++) {
+		if (!svmm || buffer->fault[fi]->inst != inst) {
+			struct nouveau_ivmm *ivmm =
+				nouveau_ivmm_find(svm, buffer->fault[fi]->inst);
+			svmm = ivmm ? ivmm->svmm : NULL;
+			inst = buffer->fault[fi]->inst;
+			SVM_DBG(svm, "inst %016llx -> svm-%p", inst, svmm);
+		}
+		buffer->fault[fi]->svmm = svmm;
+	}
+	mutex_unlock(&svm->mutex);
+
+	/* Process list of faults. */
+	args.i.i.version = 0;
+	args.i.i.type = NVIF_IOCTL_V0_MTHD;
+	args.i.m.version = 0;
+	args.i.m.method = NVIF_VMM_V0_PFNMAP;
+	args.i.p.version = 0;
+
+	for (fi = 0; fn = fi + 1, fi < buffer->fault_nr; fi = fn) {
+		/* Cancel any faults from non-SVM channels. */
+		if (!(svmm = buffer->fault[fi]->svmm)) {
+			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
+			continue;
+		}
+		SVMM_DBG(svmm, "addr %016llx", buffer->fault[fi]->addr);
+
+		/* We try and group handling of faults within a small
+		 * window into a single update.
+		 */
+		start = buffer->fault[fi]->addr;
+		limit = start + (ARRAY_SIZE(args.phys) << PAGE_SHIFT);
+		if (start < svmm->unmanaged.limit)
+			limit = min_t(u64, limit, svmm->unmanaged.start);
+		else
+		if (limit > svmm->unmanaged.start)
+			start = max_t(u64, start, svmm->unmanaged.limit);
+		SVMM_DBG(svmm, "wndw %016llx-%016llx", start, limit);
+
+		/* Intersect fault window with the CPU VMA, cancelling
+		 * the fault if the address is invalid.
+		 */
+		down_read(&svmm->mm->mmap_sem);
+		vma = find_vma_intersection(svmm->mm, start, limit);
+		if (!vma) {
+			SVMM_ERR(svmm, "wndw %016llx-%016llx", start, limit);
+			up_read(&svmm->mm->mmap_sem);
+			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
+			continue;
+		}
+		start = max_t(u64, start, vma->vm_start);
+		limit = min_t(u64, limit, vma->vm_end);
+		SVMM_DBG(svmm, "wndw %016llx-%016llx", start, limit);
+
+		if (buffer->fault[fi]->addr != start) {
+			SVMM_ERR(svmm, "addr %016llx", buffer->fault[fi]->addr);
+			up_read(&svmm->mm->mmap_sem);
+			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
+			continue;
+		}
+
+		/* Prepare the GPU-side update of all pages within the
+		 * fault window, determining required pages and access
+		 * permissions based on pending faults.
+		 */
+		args.i.p.page = PAGE_SHIFT;
+		args.i.p.addr = start;
+		for (fn = fi, pi = 0;;) {
+			/* Determine required permissions based on GPU fault
+			 * access flags.
+			 *XXX: atomic?
+			 */
+			if (buffer->fault[fn]->access != 0 /* READ. */ &&
+			    buffer->fault[fn]->access != 3 /* PREFETCH. */) {
+				args.phys[pi++] = NVIF_VMM_PFNMAP_V0_V |
+						  NVIF_VMM_PFNMAP_V0_W;
+			} else {
+				args.phys[pi++] = NVIF_VMM_PFNMAP_V0_V;
+			}
+			args.i.p.size = pi << PAGE_SHIFT;
+
+			/* It's okay to skip over duplicate addresses from the
+			 * same SVMM as faults are ordered by access type such
+			 * that only the first one needs to be handled.
+			 *
+			 * ie. WRITE faults appear first, thus any handling of
+			 * pending READ faults will already be satisfied.
+			 */
+			while (++fn < buffer->fault_nr &&
+			       buffer->fault[fn]->svmm == svmm &&
+			       buffer->fault[fn    ]->addr ==
+			       buffer->fault[fn - 1]->addr);
+
+			/* If the next fault is outside the window, or all GPU
+			 * faults have been dealt with, we're done here.
+			 */
+			if (fn >= buffer->fault_nr ||
+			    buffer->fault[fn]->svmm != svmm ||
+			    buffer->fault[fn]->addr >= limit)
+				break;
+
+			/* Fill in the gap between this fault and the next. */
+			fill = (buffer->fault[fn    ]->addr -
+				buffer->fault[fn - 1]->addr) >> PAGE_SHIFT;
+			while (--fill)
+				args.phys[pi++] = NVIF_VMM_PFNMAP_V0_NONE;
+		}
+
+		SVMM_DBG(svmm, "wndw %016llx-%016llx covering %d fault(s)",
+			 args.i.p.addr,
+			 args.i.p.addr + args.i.p.size, fn - fi);
+
+		/* Have HMM fault pages within the fault window to the GPU. */
+		range.vma = vma;
+		range.start = args.i.p.addr;
+		range.end = args.i.p.addr + args.i.p.size;
+		range.pfns = args.phys;
+		range.flags = nouveau_svm_pfn_flags;
+		range.values = nouveau_svm_pfn_values;
+		range.pfn_shift = NVIF_VMM_PFNMAP_V0_ADDR_SHIFT;
+again:
+		ret = hmm_vma_fault(&range, true);
+		if (ret == 0) {
+			mutex_lock(&svmm->mutex);
+			if (!hmm_vma_range_done(&range)) {
+				mutex_unlock(&svmm->mutex);
+				goto again;
+			}
+
+			nouveau_dmem_convert_pfn(svm->drm, &range);
+
+			svmm->vmm->vmm.object.client->super = true;
+			ret = nvif_object_ioctl(&svmm->vmm->vmm.object,
+						&args, sizeof(args.i) +
+						pi * sizeof(args.phys[0]),
+						NULL);
+			svmm->vmm->vmm.object.client->super = false;
+			mutex_unlock(&svmm->mutex);
+		}
+		up_read(&svmm->mm->mmap_sem);
+
+		/* Cancel any faults in the window whose pages didn't manage
+		 * to keep their valid bit, or stay writeable when required.
+		 *
+		 * If handling failed completely, cancel all faults.
+		 */
+		while (fi < fn) {
+			struct nouveau_svm_fault *fault = buffer->fault[fi++];
+			pi = (fault->addr - range.start) >> PAGE_SHIFT;
+			if (ret ||
+			     !(range.pfns[pi] & NVIF_VMM_PFNMAP_V0_V) ||
+			    (!(range.pfns[pi] & NVIF_VMM_PFNMAP_V0_W) &&
+			     fault->access != 0 && fault->access != 3)) {
+				nouveau_svm_fault_cancel_fault(svm, fault);
+				continue;
+			}
+			replay++;
+		}
+	}
+
+	/* Issue fault replay to the GPU. */
+	if (replay)
+		nouveau_svm_fault_replay(svm);
+	return NVIF_NOTIFY_KEEP;
+}
+
+static void
+nouveau_svm_fault_buffer_fini(struct nouveau_svm *svm, int id)
+{
+	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
+	nvif_notify_put(&buffer->notify);
+}
+
+static int
+nouveau_svm_fault_buffer_init(struct nouveau_svm *svm, int id)
+{
+	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
+	struct nvif_object *device = &svm->drm->client.device.object;
+	buffer->get = nvif_rd32(device, buffer->getaddr);
+	buffer->put = nvif_rd32(device, buffer->putaddr);
+	SVM_DBG(svm, "get %08x put %08x (init)", buffer->get, buffer->put);
+	return nvif_notify_get(&buffer->notify);
+}
+
+static void
+nouveau_svm_fault_buffer_dtor(struct nouveau_svm *svm, int id)
+{
+	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
+	int i;
+
+	if (buffer->fault) {
+		for (i = 0; buffer->fault[i] && i < buffer->entries; i++)
+			kfree(buffer->fault[i]);
+		kvfree(buffer->fault);
+	}
+
+	nouveau_svm_fault_buffer_fini(svm, id);
+
+	nvif_notify_fini(&buffer->notify);
+	nvif_object_fini(&buffer->object);
+}
+
+static int
+nouveau_svm_fault_buffer_ctor(struct nouveau_svm *svm, s32 oclass, int id)
+{
+	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
+	struct nouveau_drm *drm = svm->drm;
+	struct nvif_object *device = &drm->client.device.object;
+	struct nvif_clb069_v0 args = {};
+	int ret;
+
+	buffer->id = id;
+
+	ret = nvif_object_init(device, 0, oclass, &args, sizeof(args),
+			       &buffer->object);
+	if (ret < 0) {
+		SVM_ERR(svm, "Fault buffer allocation failed: %d", ret);
+		return ret;
+	}
+
+	nvif_object_map(&buffer->object, NULL, 0);
+	buffer->entries = args.entries;
+	buffer->getaddr = args.get;
+	buffer->putaddr = args.put;
+
+	ret = nvif_notify_init(&buffer->object, nouveau_svm_fault, true,
+			       NVB069_V0_NTFY_FAULT, NULL, 0, 0,
+			       &buffer->notify);
+	if (ret)
+		return ret;
+
+	buffer->fault = kvzalloc(sizeof(*buffer->fault) * buffer->entries, GFP_KERNEL);
+	if (!buffer->fault)
+		return -ENOMEM;
+
+	return nouveau_svm_fault_buffer_init(svm, id);
+}
+
+void
+nouveau_svm_resume(struct nouveau_drm *drm)
+{
+	struct nouveau_svm *svm = drm->svm;
+	if (svm)
+		nouveau_svm_fault_buffer_init(svm, 0);
+}
+
+void
+nouveau_svm_suspend(struct nouveau_drm *drm)
+{
+	struct nouveau_svm *svm = drm->svm;
+	if (svm)
+		nouveau_svm_fault_buffer_fini(svm, 0);
+}
+
+void
+nouveau_svm_fini(struct nouveau_drm *drm)
+{
+	struct nouveau_svm *svm = drm->svm;
+	if (svm) {
+		nouveau_svm_fault_buffer_dtor(svm, 0);
+		kfree(drm->svm);
+		drm->svm = NULL;
+	}
+}
+
+void
+nouveau_svm_init(struct nouveau_drm *drm)
+{
+	static const struct nvif_mclass buffers[] = {
+		{   VOLTA_FAULT_BUFFER_A, 0 },
+		{ MAXWELL_FAULT_BUFFER_A, 0 },
+		{}
+	};
+	struct nouveau_svm *svm;
+	int ret;
+
+	/* Disable on Volta and newer until channel recovery is fixed,
+	 * otherwise clients will have a trivial way to trash the GPU
+	 * for everyone.
+	 */
+	if (drm->client.device.info.family > NV_DEVICE_INFO_V0_PASCAL)
+		return;
+
+	if (!(drm->svm = svm = kzalloc(sizeof(*drm->svm), GFP_KERNEL)))
+		return;
+
+	drm->svm->drm = drm;
+	mutex_init(&drm->svm->mutex);
+	INIT_LIST_HEAD(&drm->svm->inst);
+
+	ret = nvif_mclass(&drm->client.device.object, buffers);
+	if (ret < 0) {
+		SVM_DBG(svm, "No supported fault buffer class");
+		nouveau_svm_fini(drm);
+		return;
+	}
+
+	ret = nouveau_svm_fault_buffer_ctor(svm, buffers[ret].oclass, 0);
+	if (ret) {
+		nouveau_svm_fini(drm);
+		return;
+	}
+
+	SVM_DBG(svm, "Initialised");
+}

drivers/gpu/drm/nouveau/nouveau_svm.h

···
+#ifndef __NOUVEAU_SVM_H__
+#define __NOUVEAU_SVM_H__
+#include <nvif/os.h>
+struct drm_device;
+struct drm_file;
+struct nouveau_drm;
+
+struct nouveau_svmm;
+
+#if IS_ENABLED(CONFIG_DRM_NOUVEAU_SVM)
+void nouveau_svm_init(struct nouveau_drm *);
+void nouveau_svm_fini(struct nouveau_drm *);
+void nouveau_svm_suspend(struct nouveau_drm *);
+void nouveau_svm_resume(struct nouveau_drm *);
+
+int nouveau_svmm_init(struct drm_device *, void *, struct drm_file *);
+void nouveau_svmm_fini(struct nouveau_svmm **);
+int nouveau_svmm_join(struct nouveau_svmm *, u64 inst);
+void nouveau_svmm_part(struct nouveau_svmm *, u64 inst);
+int nouveau_svmm_bind(struct drm_device *, void *, struct drm_file *);
+#else /* IS_ENABLED(CONFIG_DRM_NOUVEAU_SVM) */
+static inline void nouveau_svm_init(struct nouveau_drm *drm) {}
+static inline void nouveau_svm_fini(struct nouveau_drm *drm) {}
+static inline void nouveau_svm_suspend(struct nouveau_drm *drm) {}
+static inline void nouveau_svm_resume(struct nouveau_drm *drm) {}
+
+static inline int nouveau_svmm_init(struct drm_device *device, void *p,
+				    struct drm_file *file)
+{
+	return -ENOSYS;
+}
+
+static inline void nouveau_svmm_fini(struct nouveau_svmm **svmmp) {}
+
+static inline int nouveau_svmm_join(struct nouveau_svmm *svmm, u64 inst)
+{
+	return 0;
+}
+
+static inline void nouveau_svmm_part(struct nouveau_svmm *svmm, u64 inst) {}
+
+static inline int nouveau_svmm_bind(struct drm_device *device, void *p,
+				    struct drm_file *file)
+{
+	return -ENOSYS;
+}
+#endif /* IS_ENABLED(CONFIG_DRM_NOUVEAU_SVM) */
+#endif

drivers/gpu/drm/nouveau/nouveau_vmm.c

···
 #include "nouveau_vmm.h"
 #include "nouveau_drv.h"
 #include "nouveau_bo.h"
+#include "nouveau_svm.h"
 #include "nouveau_mem.h"
 
 void
···
 void
 nouveau_vmm_fini(struct nouveau_vmm *vmm)
 {
+	nouveau_svmm_fini(&vmm->svmm);
 	nvif_vmm_fini(&vmm->vmm);
 	vmm->cli = NULL;
 }
···
 int
 nouveau_vmm_init(struct nouveau_cli *cli, s32 oclass, struct nouveau_vmm *vmm)
 {
-	int ret = nvif_vmm_init(&cli->mmu, oclass, PAGE_SIZE, 0, NULL, 0,
+	int ret = nvif_vmm_init(&cli->mmu, oclass, false, PAGE_SIZE, 0, NULL, 0,
 				&vmm->vmm);
 	if (ret)
 		return ret;

drivers/gpu/drm/nouveau/nouveau_vmm.h

···
 struct nouveau_vmm {
 	struct nouveau_cli *cli;
 	struct nvif_vmm vmm;
+	struct nouveau_svmm *svmm;
 };
 
 int nouveau_vmm_init(struct nouveau_cli *, s32 oclass, struct nouveau_vmm *);

drivers/gpu/drm/nouveau/nv84_fence.c

···
 int
 nv84_fence_context_new(struct nouveau_channel *chan)
 {
-	struct nouveau_cli *cli = (void *)chan->user.client;
 	struct nv84_fence_priv *priv = chan->drm->fence;
 	struct nv84_fence_chan *fctx;
 	int ret;
···
 	fctx->base.sequence = nv84_fence_read(chan);
 
 	mutex_lock(&priv->mutex);
-	ret = nouveau_vma_new(priv->bo, &cli->vmm, &fctx->vma);
+	ret = nouveau_vma_new(priv->bo, chan->vmm, &fctx->vma);
 	mutex_unlock(&priv->mutex);
 
 	if (ret)

drivers/gpu/drm/nouveau/nvif/disp.c

···
 nvif_disp_ctor(struct nvif_device *device, s32 oclass, struct nvif_disp *disp)
 {
 	static const struct nvif_mclass disps[] = {
-		{ TU104_DISP, -1 },
+		{ TU102_DISP, -1 },
 		{ GV100_DISP, -1 },
 		{ GP102_DISP, -1 },
 		{ GP100_DISP, -1 },

drivers/gpu/drm/nouveau/nvif/vmm.c

···
 }
 
 int
-nvif_vmm_init(struct nvif_mmu *mmu, s32 oclass, u64 addr, u64 size,
-	      void *argv, u32 argc, struct nvif_vmm *vmm)
+nvif_vmm_init(struct nvif_mmu *mmu, s32 oclass, bool managed, u64 addr,
+	      u64 size, void *argv, u32 argc, struct nvif_vmm *vmm)
 {
 	struct nvif_vmm_v0 *args;
 	u32 argn = sizeof(*args) + argc;
···
 	if (!(args = kmalloc(argn, GFP_KERNEL)))
 		return -ENOMEM;
 	args->version = 0;
+	args->managed = managed;
 	args->addr = addr;
 	args->size = size;
 	memcpy(args->data, argv, argc);

drivers/gpu/drm/nouveau/nvkm/core/subdev.c

···
 	[NVKM_SUBDEV_FB      ] = "fb",
 	[NVKM_SUBDEV_FUSE    ] = "fuse",
 	[NVKM_SUBDEV_GPIO    ] = "gpio",
+	[NVKM_SUBDEV_GSP     ] = "gsp",
 	[NVKM_SUBDEV_I2C     ] = "i2c",
 	[NVKM_SUBDEV_IBUS    ] = "priv",
 	[NVKM_SUBDEV_ICCSENSE] = "iccsense",

drivers/gpu/drm/nouveau/nvkm/engine/ce/Kbuild

···
 nvkm-y += nvkm/engine/ce/gp100.o
 nvkm-y += nvkm/engine/ce/gp102.o
 nvkm-y += nvkm/engine/ce/gv100.o
-nvkm-y += nvkm/engine/ce/tu104.o
+nvkm-y += nvkm/engine/ce/tu102.o

drivers/gpu/drm/nouveau/nvkm/engine/ce/tu104.c drivers/gpu/drm/nouveau/nvkm/engine/fifo/usertu102.c

···
  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  * OTHER DEALINGS IN THE SOFTWARE.
  */
-#include "priv.h"
+#include "user.h"
 
-#include <nvif/class.h>
+static int
+tu102_fifo_user_map(struct nvkm_object *object, void *argv, u32 argc,
+		    enum nvkm_object_map *type, u64 *addr, u64 *size)
+{
+	struct nvkm_device *device = object->engine->subdev.device;
+	*addr = 0xbb0000 + device->func->resource_addr(device, 0);
+	*size = 0x010000;
+	*type = NVKM_OBJECT_MAP_IO;
+	return 0;
+}
 
-static const struct nvkm_engine_func
-tu104_ce = {
-	.intr = gp100_ce_intr,
-	.sclass = {
-		{ -1, -1, TURING_DMA_COPY_A },
-		{}
-	}
+static const struct nvkm_object_func
+tu102_fifo_user = {
+	.map = tu102_fifo_user_map,
 };
 
 int
-tu104_ce_new(struct nvkm_device *device, int index,
-	     struct nvkm_engine **pengine)
+tu102_fifo_user_new(const struct nvkm_oclass *oclass, void *argv, u32 argc,
+		    struct nvkm_object **pobject)
 {
-	return nvkm_engine_new_(&tu104_ce, device, index, true, pengine);
+	return nvkm_object_new_(&tu102_fifo_user, oclass, argv, argc, pobject);
 }

drivers/gpu/drm/nouveau/nvkm/engine/device/base.c

···
 	.pci = gf106_pci_new,
 	.therm = gf119_therm_new,
 	.timer = nv41_timer_new,
-	.volt = gf100_volt_new,
+	.volt = gf117_volt_new,
 	.ce[0] = gf100_ce_new,
 	.disp = gf119_disp_new,
 	.dma = gf119_dma_new,
···
 	.fb = gv100_fb_new,
 	.fuse = gm107_fuse_new,
 	.gpio = gk104_gpio_new,
+	.gsp = gv100_gsp_new,
 	.i2c = gm200_i2c_new,
 	.ibus = gm200_ibus_new,
 	.imem = nv50_instmem_new,
···
 static const struct nvkm_device_chip
 nv162_chipset = {
 	.name = "TU102",
-	.bar = tu104_bar_new,
+	.bar = tu102_bar_new,
 	.bios = nvkm_bios_new,
 	.bus = gf100_bus_new,
-	.devinit = tu104_devinit_new,
-	.fault = tu104_fault_new,
+	.devinit = tu102_devinit_new,
+	.fault = tu102_fault_new,
 	.fb = gv100_fb_new,
 	.fuse = gm107_fuse_new,
 	.gpio = gk104_gpio_new,
+	.gsp = gv100_gsp_new,
 	.i2c = gm200_i2c_new,
 	.ibus = gm200_ibus_new,
 	.imem = nv50_instmem_new,
 	.ltc = gp102_ltc_new,
-	.mc = tu104_mc_new,
-	.mmu = tu104_mmu_new,
+	.mc = tu102_mc_new,
+	.mmu = tu102_mmu_new,
 	.pci = gp100_pci_new,
 	.pmu = gp102_pmu_new,
 	.therm = gp100_therm_new,
 	.timer = gk20a_timer_new,
 	.top = gk104_top_new,
-	.ce[0] = tu104_ce_new,
-	.ce[1] = tu104_ce_new,
-	.ce[2] = tu104_ce_new,
-	.ce[3] = tu104_ce_new,
-	.ce[4] = tu104_ce_new,
-	.disp = tu104_disp_new,
+	.ce[0] = tu102_ce_new,
+	.ce[1] = tu102_ce_new,
+	.ce[2] = tu102_ce_new,
+	.ce[3] = tu102_ce_new,
+	.ce[4] = tu102_ce_new,
+	.disp = tu102_disp_new,
 	.dma = gv100_dma_new,
-	.fifo = tu104_fifo_new,
+	.fifo = tu102_fifo_new,
+	.nvdec[0] = gp102_nvdec_new,
+	.sec2 = tu102_sec2_new,
 };
 
 static const struct nvkm_device_chip
 nv164_chipset = {
 	.name = "TU104",
-	.bar = tu104_bar_new,
+	.bar = tu102_bar_new,
 	.bios = nvkm_bios_new,
 	.bus = gf100_bus_new,
-	.devinit = tu104_devinit_new,
-	.fault = tu104_fault_new,
+	.devinit = tu102_devinit_new,
+	.fault = tu102_fault_new,
 	.fb = gv100_fb_new,
 	.fuse = gm107_fuse_new,
 	.gpio = gk104_gpio_new,
+	.gsp = gv100_gsp_new,
 	.i2c = gm200_i2c_new,
 	.ibus = gm200_ibus_new,
 	.imem = nv50_instmem_new,
 	.ltc = gp102_ltc_new,
-	.mc = tu104_mc_new,
-	.mmu = tu104_mmu_new,
+	.mc = tu102_mc_new,
+	.mmu = tu102_mmu_new,
 	.pci = gp100_pci_new,
 	.pmu = gp102_pmu_new,
 	.therm = gp100_therm_new,
 	.timer = gk20a_timer_new,
 	.top = gk104_top_new,
-	.ce[0] = tu104_ce_new,
-	.ce[1] = tu104_ce_new,
-	.ce[2] = tu104_ce_new,
-	.ce[3] = tu104_ce_new,
-	.ce[4] = tu104_ce_new,
-	.disp = tu104_disp_new,
+	.ce[0] = tu102_ce_new,
+	.ce[1] = tu102_ce_new,
+	.ce[2] = tu102_ce_new,
+	.ce[3] = tu102_ce_new,
+	.ce[4] = tu102_ce_new,
+	.disp = tu102_disp_new,
 	.dma = gv100_dma_new,
-	.fifo = tu104_fifo_new,
+	.fifo = tu102_fifo_new,
+	.nvdec[0] = gp102_nvdec_new,
+	.sec2 = tu102_sec2_new,
 };
 
 static const struct nvkm_device_chip
 nv166_chipset = {
 	.name = "TU106",
-	.bar = tu104_bar_new,
+	.bar = tu102_bar_new,
 	.bios = nvkm_bios_new,
 	.bus = gf100_bus_new,
-	.devinit = tu104_devinit_new,
-	.fault = tu104_fault_new,
+	.devinit = tu102_devinit_new,
+	.fault = tu102_fault_new,
 	.fb = gv100_fb_new,
 	.fuse = gm107_fuse_new,
 	.gpio = gk104_gpio_new,
+	.gsp = gv100_gsp_new,
 	.i2c = gm200_i2c_new,
 	.ibus = gm200_ibus_new,
 	.imem = nv50_instmem_new,
 	.ltc = gp102_ltc_new,
-	.mc = tu104_mc_new,
-	.mmu = tu104_mmu_new,
+	.mc = tu102_mc_new,
+	.mmu = tu102_mmu_new,
 	.pci = gp100_pci_new,
 	.pmu = gp102_pmu_new,
 	.therm = gp100_therm_new,
 	.timer = gk20a_timer_new,
 	.top = gk104_top_new,
-	.ce[0] = tu104_ce_new,
-	.ce[1] = tu104_ce_new,
-	.ce[2] = tu104_ce_new,
-	.ce[3] = tu104_ce_new,
-	.ce[4] = tu104_ce_new,
-	.disp = tu104_disp_new,
+	.ce[0] = tu102_ce_new,
+	.ce[1] = tu102_ce_new,
+	.ce[2] = tu102_ce_new,
+	.ce[3] = tu102_ce_new,
+	.ce[4] = tu102_ce_new,
+	.disp = tu102_disp_new,
 	.dma = gv100_dma_new,
-	.fifo = tu104_fifo_new,
+	.fifo = tu102_fifo_new,
+	.nvdec[0] = gp102_nvdec_new,
+	.sec2 = tu102_sec2_new,
 };
 
 static int
···
 	_(FB      , device->fb      , &device->fb->subdev);
 	_(FUSE    , device->fuse    , &device->fuse->subdev);
 	_(GPIO    , device->gpio    , &device->gpio->subdev);
+	_(GSP     , device->gsp     , &device->gsp->subdev);
 	_(I2C     , device->i2c     , &device->i2c->subdev);
 	_(IBUS    , device->ibus    ,  device->ibus);
 	_(ICCSENSE, device->iccsense, &device->iccsense->subdev);
···
 		_(NVKM_SUBDEV_FB      ,       fb);
 		_(NVKM_SUBDEV_FUSE    ,     fuse);
 		_(NVKM_SUBDEV_GPIO    ,     gpio);
+		_(NVKM_SUBDEV_GSP     ,      gsp);
 		_(NVKM_SUBDEV_I2C     ,      i2c);
 		_(NVKM_SUBDEV_IBUS    ,     ibus);
 		_(NVKM_SUBDEV_ICCSENSE, iccsense);

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

···
 #include <subdev/fb.h>
 #include <subdev/fuse.h>
 #include <subdev/gpio.h>
+#include <subdev/gsp.h>
 #include <subdev/i2c.h>
 #include <subdev/ibus.h>
 #include <subdev/iccsense.h>

drivers/gpu/drm/nouveau/nvkm/engine/device/user.c

···
 	}
 
 	if (!sclass) {
-		switch (index) {
-		case 0: sclass = &nvkm_control_oclass; break;
-		case 1:
-			if (!device->mmu)
-				return -EINVAL;
+		if (index-- == 0)
+			sclass = &nvkm_control_oclass;
+		else if (device->mmu && index-- == 0)
 			sclass = &device->mmu->user;
-			break;
-		default:
+		else if (device->fault && index-- == 0)
+			sclass = &device->fault->user;
+		else
 			return -EINVAL;
-		}
+
 		oclass->base = sclass->base;
 	}
 

drivers/gpu/drm/nouveau/nvkm/engine/disp/Kbuild

···
 nvkm-y += nvkm/engine/disp/gp100.o
 nvkm-y += nvkm/engine/disp/gp102.o
 nvkm-y += nvkm/engine/disp/gv100.o
-nvkm-y += nvkm/engine/disp/tu104.o
+nvkm-y += nvkm/engine/disp/tu102.o
 nvkm-y += nvkm/engine/disp/vga.o
 
 nvkm-y += nvkm/engine/disp/head.o
···
 nvkm-y += nvkm/engine/disp/sorgm107.o
 nvkm-y += nvkm/engine/disp/sorgm200.o
 nvkm-y += nvkm/engine/disp/sorgv100.o
-nvkm-y += nvkm/engine/disp/sortu104.o
+nvkm-y += nvkm/engine/disp/sortu102.o
 
 nvkm-y += nvkm/engine/disp/outp.o
 nvkm-y += nvkm/engine/disp/dp.o
···
 nvkm-y += nvkm/engine/disp/rootgp100.o
 nvkm-y += nvkm/engine/disp/rootgp102.o
 nvkm-y += nvkm/engine/disp/rootgv100.o
-nvkm-y += nvkm/engine/disp/roottu104.o
+nvkm-y += nvkm/engine/disp/roottu102.o
 
 nvkm-y += nvkm/engine/disp/channv50.o
 nvkm-y += nvkm/engine/disp/changf119.o

drivers/gpu/drm/nouveau/nvkm/engine/disp/gf119.c

···
 {
 	struct nvkm_subdev *subdev = &disp->base.engine.subdev;
 	struct nvkm_device *device = subdev->device;
-	u32 mthd = nvkm_rd32(device, 0x6101f0 + (chid * 12));
+	u32 stat = nvkm_rd32(device, 0x6101f0 + (chid * 12));
+	u32 type = (stat & 0x00007000) >> 12;
+	u32 mthd = (stat & 0x00000ffc);
 	u32 data = nvkm_rd32(device, 0x6101f4 + (chid * 12));
-	u32 unkn = nvkm_rd32(device, 0x6101f8 + (chid * 12));
+	u32 code = nvkm_rd32(device, 0x6101f8 + (chid * 12));
+	const struct nvkm_enum *reason =
+		nvkm_enum_find(nv50_disp_intr_error_type, type);
 
-	nvkm_error(subdev, "chid %d mthd %04x data %08x %08x %08x\n",
-		   chid, (mthd & 0x0000ffc), data, mthd, unkn);
+	nvkm_error(subdev, "chid %d stat %08x reason %d [%s] mthd %04x "
+			   "data %08x code %08x\n",
+		   chid, stat, type, reason ? reason->name : "",
+		   mthd, data, code);
 
 	if (chid < ARRAY_SIZE(disp->chan)) {
-		switch (mthd & 0xffc) {
+		switch (mthd) {
 		case 0x0080:
 			nv50_disp_chan_mthd(disp->chan[chid], NV_DBG_ERROR);
 			break;

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

···
 	u32 mthd = (stat & 0x00000fff) << 2;
 	u32 data = nvkm_rd32(device, 0x611024 + (chid * 12));
 	u32 code = nvkm_rd32(device, 0x611028 + (chid * 12));
+	const struct nvkm_enum *reason =
+		nvkm_enum_find(nv50_disp_intr_error_type, type);
 
-	nvkm_error(subdev, "chid %d %08x [type %d mthd %04x] "
+	nvkm_error(subdev, "chid %d stat %08x reason %d [%s] mthd %04x "
 			   "data %08x code %08x\n",
-		   chid, stat, type, mthd, data, code);
+		   chid, stat, type, reason ? reason->name : "",
+		   mthd, data, code);
 
 	if (chid < ARRAY_SIZE(disp->chan) && disp->chan[chid]) {
 		switch (mthd) {

drivers/gpu/drm/nouveau/nvkm/engine/disp/ior.h

···
 int gv100_sor_cnt(struct nvkm_disp *, unsigned long *);
 int gv100_sor_new(struct nvkm_disp *, int);
 
-int tu104_sor_new(struct nvkm_disp *, int);
+int tu102_sor_new(struct nvkm_disp *, int);
 #endif

drivers/gpu/drm/nouveau/nvkm/engine/disp/nv50.c

···
 #include "rootnv50.h"
 
 #include <core/client.h>
-#include <core/enum.h>
 #include <core/ramht.h>
 #include <subdev/bios.h>
 #include <subdev/bios/disp.h>
···
 	nvkm_wr32(device, 0x610030, 0x80000000);
 }
 
-static const struct nvkm_enum
+const struct nvkm_enum
 nv50_disp_intr_error_type[] = {
-	{ 3, "ILLEGAL_MTHD" },
-	{ 4, "INVALID_VALUE" },
+	{ 0, "NONE" },
+	{ 1, "PUSHBUFFER_ERR" },
+	{ 2, "TRAP" },
+	{ 3, "RESERVED_METHOD" },
+	{ 4, "INVALID_ARG" },
 	{ 5, "INVALID_STATE" },
-	{ 7, "INVALID_HANDLE" },
+	{ 7, "UNRESOLVABLE_HANDLE" },
 	{}
 };
 

drivers/gpu/drm/nouveau/nvkm/engine/disp/nv50.h

···
 #include "priv.h"
 struct nvkm_head;
 
+#include <core/enum.h>
+
 struct nv50_disp {
 	const struct nv50_disp_func *func;
 	struct nvkm_disp base;
···
 void nv50_disp_fini(struct nv50_disp *);
 void nv50_disp_intr(struct nv50_disp *);
 void nv50_disp_super(struct work_struct *);
+extern const struct nvkm_enum nv50_disp_intr_error_type[];
 
 int gf119_disp_init(struct nv50_disp *);
 void gf119_disp_fini(struct nv50_disp *);

drivers/gpu/drm/nouveau/nvkm/engine/disp/rootnv50.h

···
 extern const struct nvkm_disp_oclass gp100_disp_root_oclass;
 extern const struct nvkm_disp_oclass gp102_disp_root_oclass;
 extern const struct nvkm_disp_oclass gv100_disp_root_oclass;
-extern const struct nvkm_disp_oclass tu104_disp_root_oclass;
+extern const struct nvkm_disp_oclass tu102_disp_root_oclass;
 #endif

drivers/gpu/drm/nouveau/nvkm/engine/disp/roottu104.c drivers/gpu/drm/nouveau/nvkm/engine/ce/tu102.c

···
  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  * OTHER DEALINGS IN THE SOFTWARE.
  */
-#include "rootnv50.h"
-#include "channv50.h"
+#include "priv.h"
 
 #include <nvif/class.h>
 
-static const struct nv50_disp_root_func
-tu104_disp_root = {
-	.user = {
-		{{0,0,TU104_DISP_CURSOR                }, gv100_disp_curs_new },
-		{{0,0,TU104_DISP_WINDOW_IMM_CHANNEL_DMA}, gv100_disp_wimm_new },
-		{{0,0,TU104_DISP_CORE_CHANNEL_DMA      }, gv100_disp_core_new },
-		{{0,0,TU104_DISP_WINDOW_CHANNEL_DMA    }, gv100_disp_wndw_new },
+static const struct nvkm_engine_func
+tu102_ce = {
+	.intr = gp100_ce_intr,
+	.sclass = {
+		{ -1, -1, TURING_DMA_COPY_A },
 		{}
-	},
+	}
 };
 
-static int
-tu104_disp_root_new(struct nvkm_disp *disp, const struct nvkm_oclass *oclass,
-		    void *data, u32 size, struct nvkm_object **pobject)
+int
+tu102_ce_new(struct nvkm_device *device, int index,
+	     struct nvkm_engine **pengine)
 {
-	return nv50_disp_root_new_(&tu104_disp_root, disp, oclass,
-				   data, size, pobject);
+	return nvkm_engine_new_(&tu102_ce, device, index, true, pengine);
 }
-
-const struct nvkm_disp_oclass
-tu104_disp_root_oclass = {
-	.base.oclass = TU104_DISP,
-	.base.minver = -1,
-	.base.maxver = -1,
-	.ctor = tu104_disp_root_new,
-};

drivers/gpu/drm/nouveau/nvkm/engine/disp/sortu104.c drivers/gpu/drm/nouveau/nvkm/engine/disp/sortu102.c

···
 #include <subdev/timer.h>
 
 static void
-tu104_sor_dp_vcpi(struct nvkm_ior *sor, int head,
+tu102_sor_dp_vcpi(struct nvkm_ior *sor, int head,
 		  u8 slot, u8 slot_nr, u16 pbn, u16 aligned)
 {
 	struct nvkm_device *device = sor->disp->engine.subdev.device;
···
 }
 
 static int
-tu104_sor_dp_links(struct nvkm_ior *sor, struct nvkm_i2c_aux *aux)
+tu102_sor_dp_links(struct nvkm_ior *sor, struct nvkm_i2c_aux *aux)
 {
 	struct nvkm_device *device = sor->disp->engine.subdev.device;
 	const u32 soff = nv50_ior_base(sor);
···
 }
 
 static const struct nvkm_ior_func
-tu104_sor = {
+tu102_sor = {
 	.route = {
 		.get = gm200_sor_route_get,
 		.set = gm200_sor_route_set,
···
 	},
 	.dp = {
 		.lanes = { 0, 1, 2, 3 },
-		.links = tu104_sor_dp_links,
+		.links = tu102_sor_dp_links,
 		.power = g94_sor_dp_power,
 		.pattern = gm107_sor_dp_pattern,
 		.drive = gm200_sor_dp_drive,
-		.vcpi = tu104_sor_dp_vcpi,
+		.vcpi = tu102_sor_dp_vcpi,
 		.audio = gv100_sor_dp_audio,
 		.audio_sym = gv100_sor_dp_audio_sym,
 		.watermark = gv100_sor_dp_watermark,
···
 };
 
 int
-tu104_sor_new(struct nvkm_disp *disp, int id)
+tu102_sor_new(struct nvkm_disp *disp, int id)
 {
-	return nvkm_ior_new_(&tu104_sor, disp, SOR, id);
+	return nvkm_ior_new_(&tu102_sor, disp, SOR, id);
 }

drivers/gpu/drm/nouveau/nvkm/engine/disp/tu104.c drivers/gpu/drm/nouveau/nvkm/engine/disp/tu102.c

···
 #include <subdev/timer.h>
 
 static int
-tu104_disp_init(struct nv50_disp *disp)
+tu102_disp_init(struct nv50_disp *disp)
 {
 	struct nvkm_device *device = disp->base.engine.subdev.device;
 	struct nvkm_head *head;
···
 }
 
 static const struct nv50_disp_func
-tu104_disp = {
-	.init = tu104_disp_init,
+tu102_disp = {
+	.init = tu102_disp_init,
 	.fini = gv100_disp_fini,
 	.intr = gv100_disp_intr,
 	.uevent = &gv100_disp_chan_uevent,
 	.super = gv100_disp_super,
-	.root = &tu104_disp_root_oclass,
+	.root = &tu102_disp_root_oclass,
 	.wndw = { .cnt = gv100_disp_wndw_cnt },
 	.head = { .cnt = gv100_head_cnt, .new = gv100_head_new },
-	.sor = { .cnt = gv100_sor_cnt, .new = tu104_sor_new },
+	.sor = { .cnt = gv100_sor_cnt, .new = tu102_sor_new },
 	.ramht_size = 0x2000,
 };
 
 int
-tu104_disp_new(struct nvkm_device *device, int index, struct nvkm_disp **pdisp)
+tu102_disp_new(struct nvkm_device *device, int index, struct nvkm_disp **pdisp)
 {
-	return nv50_disp_new_(&tu104_disp, device, index, pdisp);
+	return nv50_disp_new_(&tu102_disp, device, index, pdisp);
 }

drivers/gpu/drm/nouveau/nvkm/engine/fifo/Kbuild

···
 nvkm-y += nvkm/engine/fifo/gp100.o
 nvkm-y += nvkm/engine/fifo/gp10b.o
 nvkm-y += nvkm/engine/fifo/gv100.o
-nvkm-y += nvkm/engine/fifo/tu104.o
+nvkm-y += nvkm/engine/fifo/tu102.o
 
 nvkm-y += nvkm/engine/fifo/chan.o
 nvkm-y += nvkm/engine/fifo/channv50.o
···
 nvkm-y += nvkm/engine/fifo/gpfifogf100.o
 nvkm-y += nvkm/engine/fifo/gpfifogk104.o
 nvkm-y += nvkm/engine/fifo/gpfifogv100.o
-nvkm-y += nvkm/engine/fifo/gpfifotu104.o
+nvkm-y += nvkm/engine/fifo/gpfifotu102.o
 
 nvkm-y += nvkm/engine/fifo/usergv100.o
-nvkm-y += nvkm/engine/fifo/usertu104.o
+nvkm-y += nvkm/engine/fifo/usertu102.o

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

···
 int gv100_fifo_gpfifo_engine_fini(struct nvkm_fifo_chan *,
 				  struct nvkm_engine *, bool);
 
-int tu104_fifo_gpfifo_new(struct gk104_fifo *, const struct nvkm_oclass *,
+int tu102_fifo_gpfifo_new(struct gk104_fifo *, const struct nvkm_oclass *,
 			  void *data, u32 size, struct nvkm_object **);
 #endif

drivers/gpu/drm/nouveau/nvkm/engine/fifo/gpfifotu104.c drivers/gpu/drm/nouveau/nvkm/engine/fifo/gpfifotu102.c

···
 #include <nvif/unpack.h>
 
 static u32
-tu104_fifo_gpfifo_submit_token(struct nvkm_fifo_chan *base)
+tu102_fifo_gpfifo_submit_token(struct nvkm_fifo_chan *base)
 {
 	struct gk104_fifo_chan *chan = gk104_fifo_chan(base);
 	return (chan->runl << 16) | chan->base.chid;
 }
 
 static const struct nvkm_fifo_chan_func
-tu104_fifo_gpfifo = {
+tu102_fifo_gpfifo = {
 	.dtor = gk104_fifo_gpfifo_dtor,
 	.init = gk104_fifo_gpfifo_init,
 	.fini = gk104_fifo_gpfifo_fini,
···
 	.engine_dtor = gk104_fifo_gpfifo_engine_dtor,
 	.engine_init = gv100_fifo_gpfifo_engine_init,
 	.engine_fini = gv100_fifo_gpfifo_engine_fini,
-	.submit_token = tu104_fifo_gpfifo_submit_token,
+	.submit_token = tu102_fifo_gpfifo_submit_token,
 };
 
 int
-tu104_fifo_gpfifo_new(struct gk104_fifo *fifo, const struct nvkm_oclass *oclass,
+tu102_fifo_gpfifo_new(struct gk104_fifo *fifo, const struct nvkm_oclass *oclass,
 		      void *data, u32 size, struct nvkm_object **pobject)
 {
 	struct nvkm_object *parent = oclass->parent;
···
 			   args->v0.ilength, args->v0.runlist, args->v0.priv);
 		if (args->v0.priv && !oclass->client->super)
 			return -EINVAL;
-		return gv100_fifo_gpfifo_new_(&tu104_fifo_gpfifo, fifo,
+		return gv100_fifo_gpfifo_new_(&tu102_fifo_gpfifo, fifo,
 					      &args->v0.runlist,
 					      &args->v0.chid,
 					       args->v0.vmm,

drivers/gpu/drm/nouveau/nvkm/engine/fifo/tu104.c drivers/gpu/drm/nouveau/nvkm/engine/fifo/tu102.c

···
 #include <nvif/class.h>
 
 static void
-tu104_fifo_runlist_commit(struct gk104_fifo *fifo, int runl,
+tu102_fifo_runlist_commit(struct gk104_fifo *fifo, int runl,
 			  struct nvkm_memory *mem, int nr)
 {
 	struct nvkm_device *device = fifo->base.engine.subdev.device;
···
 }
 
 const struct gk104_fifo_runlist_func
-tu104_fifo_runlist = {
+tu102_fifo_runlist = {
 	.size = 16,
 	.cgrp = gv100_fifo_runlist_cgrp,
 	.chan = gv100_fifo_runlist_chan,
-	.commit = tu104_fifo_runlist_commit,
+	.commit = tu102_fifo_runlist_commit,
 };
 
 static const struct nvkm_enum
-tu104_fifo_fault_engine[] = {
+tu102_fifo_fault_engine[] = {
 	{ 0x01, "DISPLAY" },
 	{ 0x03, "PTP" },
 	{ 0x06, "PWR_PMU" },
···
 };
 
 static void
-tu104_fifo_pbdma_init(struct gk104_fifo *fifo)
+tu102_fifo_pbdma_init(struct gk104_fifo *fifo)
 {
 	struct nvkm_device *device = fifo->base.engine.subdev.device;
 	const u32 mask = (1 << fifo->pbdma_nr) - 1;
···
 }
 
 static const struct gk104_fifo_pbdma_func
-tu104_fifo_pbdma = {
+tu102_fifo_pbdma = {
 	.nr = gm200_fifo_pbdma_nr,
-	.init = tu104_fifo_pbdma_init,
+	.init = tu102_fifo_pbdma_init,
 	.init_timeout = gk208_fifo_pbdma_init_timeout,
 };
 
 static const struct gk104_fifo_func
-tu104_fifo = {
-	.pbdma = &tu104_fifo_pbdma,
+tu102_fifo = {
+	.pbdma = &tu102_fifo_pbdma,
 	.fault.access = gv100_fifo_fault_access,
-	.fault.engine = tu104_fifo_fault_engine,
+	.fault.engine = tu102_fifo_fault_engine,
 	.fault.reason = gv100_fifo_fault_reason,
 	.fault.hubclient = gv100_fifo_fault_hubclient,
 	.fault.gpcclient = gv100_fifo_fault_gpcclient,
-	.runlist = &tu104_fifo_runlist,
-	.user = {{-1,-1,VOLTA_USERMODE_A       }, tu104_fifo_user_new   },
-	.chan = {{ 0, 0,TURING_CHANNEL_GPFIFO_A}, tu104_fifo_gpfifo_new },
+	.runlist = &tu102_fifo_runlist,
+	.user = {{-1,-1,VOLTA_USERMODE_A       }, tu102_fifo_user_new   },
+	.chan = {{ 0, 0,TURING_CHANNEL_GPFIFO_A}, tu102_fifo_gpfifo_new },
 	.cgrp_force = true,
 };
 
 int
-tu104_fifo_new(struct nvkm_device *device, int index, struct nvkm_fifo **pfifo)
+tu102_fifo_new(struct nvkm_device *device, int index, struct nvkm_fifo **pfifo)
 {
-	return gk104_fifo_new_(&tu104_fifo, device, index, 4096, pfifo);
+	return gk104_fifo_new_(&tu102_fifo, device, index, 4096, pfifo);
 }

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

···
 #include "priv.h"
 int gv100_fifo_user_new(const struct nvkm_oclass *, void *, u32,
 			struct nvkm_object **);
-int tu104_fifo_user_new(const struct nvkm_oclass *, void *, u32,
+int tu102_fifo_user_new(const struct nvkm_oclass *, void *, u32,
 			struct nvkm_object **);
 #endif

drivers/gpu/drm/nouveau/nvkm/engine/fifo/usertu104.c drivers/gpu/drm/nouveau/nvkm/engine/disp/roottu102.c

···
  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  * OTHER DEALINGS IN THE SOFTWARE.
  */
-#include "user.h"
+#include "rootnv50.h"
+#include "channv50.h"
 
-static int
-tu104_fifo_user_map(struct nvkm_object *object, void *argv, u32 argc,
-		    enum nvkm_object_map *type, u64 *addr, u64 *size)
-{
-	struct nvkm_device *device = object->engine->subdev.device;
-	*addr = 0xbb0000 + device->func->resource_addr(device, 0);
-	*size = 0x010000;
-	*type = NVKM_OBJECT_MAP_IO;
-	return 0;
-}
+#include <nvif/class.h>
 
-static const struct nvkm_object_func
-tu104_fifo_user = {
-	.map = tu104_fifo_user_map,
+static const struct nv50_disp_root_func
+tu102_disp_root = {
+	.user = {
+		{{0,0,TU102_DISP_CURSOR                }, gv100_disp_curs_new },
+		{{0,0,TU102_DISP_WINDOW_IMM_CHANNEL_DMA}, gv100_disp_wimm_new },
+		{{0,0,TU102_DISP_CORE_CHANNEL_DMA      }, gv100_disp_core_new },
+		{{0,0,TU102_DISP_WINDOW_CHANNEL_DMA    }, gv100_disp_wndw_new },
+		{}
+	},
 };
 
-int
-tu104_fifo_user_new(const struct nvkm_oclass *oclass, void *argv, u32 argc,
-		    struct nvkm_object **pobject)
+static int
+tu102_disp_root_new(struct nvkm_disp *disp, const struct nvkm_oclass *oclass,
+		    void *data, u32 size, struct nvkm_object **pobject)
 {
-	return nvkm_object_new_(&tu104_fifo_user, oclass, argv, argc, pobject);
+	return nv50_disp_root_new_(&tu102_disp_root, disp, oclass,
+				   data, size, pobject);
 }
+
+const struct nvkm_disp_oclass
+tu102_disp_root_oclass = {
+	.base.oclass = TU102_DISP,
+	.base.minver = -1,
+	.base.maxver = -1,
+	.ctor = tu102_disp_root_new,
+};

drivers/gpu/drm/nouveau/nvkm/engine/gr/base.c

···
 
 #include <engine/fifo.h>
 
+u32
+nvkm_gr_ctxsw_inst(struct nvkm_device *device)
+{
+	struct nvkm_gr *gr = device->gr;
+	if (gr && gr->func->ctxsw.inst)
+		return gr->func->ctxsw.inst(gr);
+	return 0;
+}
+
+int
+nvkm_gr_ctxsw_resume(struct nvkm_device *device)
+{
+	struct nvkm_gr *gr = device->gr;
+	if (gr && gr->func->ctxsw.resume)
+		return gr->func->ctxsw.resume(gr);
+	return 0;
+}
+
+int
+nvkm_gr_ctxsw_pause(struct nvkm_device *device)
+{
+	struct nvkm_gr *gr = device->gr;
+	if (gr && gr->func->ctxsw.pause)
+		return gr->func->ctxsw.pause(gr);
+	return 0;
+}
+
 static bool
 nvkm_gr_chsw_load(struct nvkm_engine *engine)
 {

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgf100.c

···
 	/* Make channel current. */
 	addr = nvkm_memory_addr(inst) >> 12;
 	if (gr->firmware) {
-		nvkm_wr32(device, 0x409840, 0x00000030);
-		nvkm_wr32(device, 0x409500, 0x80000000 | addr);
-		nvkm_wr32(device, 0x409504, 0x00000003);
-		nvkm_msec(device, 2000,
-			if (nvkm_rd32(device, 0x409800) & 0x00000010)
-				break;
-		);
+		ret = gf100_gr_fecs_bind_pointer(gr, 0x80000000 | addr);
+		if (ret)
+			goto done;
 
 		nvkm_kmap(data);
 		nvkm_wo32(data, 0x1c, 1);

drivers/gpu/drm/nouveau/nvkm/engine/gr/gf100.c

···
  * PGRAPH engine/subdev functions
  ******************************************************************************/
 
+static u32
+gf100_gr_ctxsw_inst(struct nvkm_gr *gr)
+{
+	return nvkm_rd32(gr->engine.subdev.device, 0x409b00);
+}
+
+static int
+gf100_gr_fecs_ctrl_ctxsw(struct gf100_gr *gr, u32 mthd)
+{
+	struct nvkm_device *device = gr->base.engine.subdev.device;
+
+	nvkm_wr32(device, 0x409804, 0xffffffff);
+	nvkm_wr32(device, 0x409840, 0xffffffff);
+	nvkm_wr32(device, 0x409500, 0xffffffff);
+	nvkm_wr32(device, 0x409504, mthd);
+	nvkm_msec(device, 2000,
+		u32 stat = nvkm_rd32(device, 0x409804);
+		if (stat == 0x00000002)
+			return -EIO;
+		if (stat == 0x00000001)
+			return 0;
+	);
+
+	return -ETIMEDOUT;
+}
+
+int
+gf100_gr_fecs_start_ctxsw(struct nvkm_gr *base)
+{
+	struct gf100_gr *gr = gf100_gr(base);
+	int ret = 0;
+
+	mutex_lock(&gr->fecs.mutex);
+	if (!--gr->fecs.disable) {
+		if (WARN_ON(ret = gf100_gr_fecs_ctrl_ctxsw(gr, 0x39)))
+			gr->fecs.disable++;
+	}
+	mutex_unlock(&gr->fecs.mutex);
+	return ret;
+}
+
+int
+gf100_gr_fecs_stop_ctxsw(struct nvkm_gr *base)
+{
+	struct gf100_gr *gr = gf100_gr(base);
+	int ret = 0;
+
+	mutex_lock(&gr->fecs.mutex);
+	if (!gr->fecs.disable++) {
+		if (WARN_ON(ret = gf100_gr_fecs_ctrl_ctxsw(gr, 0x38)))
+			gr->fecs.disable--;
+	}
+	mutex_unlock(&gr->fecs.mutex);
+	return ret;
+}
+
+int
+gf100_gr_fecs_bind_pointer(struct gf100_gr *gr, u32 inst)
+{
+	struct nvkm_device *device = gr->base.engine.subdev.device;
+
+	nvkm_wr32(device, 0x409840, 0x00000030);
+	nvkm_wr32(device, 0x409500, inst);
+	nvkm_wr32(device, 0x409504, 0x00000003);
+	nvkm_msec(device, 2000,
+		u32 stat = nvkm_rd32(device, 0x409800);
+		if (stat & 0x00000020)
+			return -EIO;
+		if (stat & 0x00000010)
+			return 0;
+	);
+
+	return -ETIMEDOUT;
+}
+
+static int
+gf100_gr_fecs_set_reglist_virtual_address(struct gf100_gr *gr, u64 addr)
+{
+	struct nvkm_device *device = gr->base.engine.subdev.device;
+
+	nvkm_wr32(device, 0x409810, addr >> 8);
+	nvkm_wr32(device, 0x409800, 0x00000000);
+	nvkm_wr32(device, 0x409500, 0x00000001);
+	nvkm_wr32(device, 0x409504, 0x00000032);
+	nvkm_msec(device, 2000,
+		if (nvkm_rd32(device, 0x409800) == 0x00000001)
+			return 0;
+	);
+
+	return -ETIMEDOUT;
+}
+
+static int
+gf100_gr_fecs_set_reglist_bind_instance(struct gf100_gr *gr, u32 inst)
+{
+	struct nvkm_device *device = gr->base.engine.subdev.device;
+
+	nvkm_wr32(device, 0x409810, inst);
+	nvkm_wr32(device, 0x409800, 0x00000000);
+	nvkm_wr32(device, 0x409500, 0x00000001);
+	nvkm_wr32(device, 0x409504, 0x00000031);
+	nvkm_msec(device, 2000,
+		if (nvkm_rd32(device, 0x409800) == 0x00000001)
+			return 0;
+	);
+
+	return -ETIMEDOUT;
+}
+
+static int
+gf100_gr_fecs_discover_reglist_image_size(struct gf100_gr *gr, u32 *psize)
+{
+	struct nvkm_device *device = gr->base.engine.subdev.device;
+
+	nvkm_wr32(device, 0x409800, 0x00000000);
+	nvkm_wr32(device, 0x409500, 0x00000001);
+	nvkm_wr32(device, 0x409504, 0x00000030);
+	nvkm_msec(device, 2000,
+		if ((*psize = nvkm_rd32(device, 0x409800)))
+			return 0;
+	);
+
+	return -ETIMEDOUT;
+}
+
+static int
+gf100_gr_fecs_elpg_bind(struct gf100_gr *gr)
+{
+	u32 size;
+	int ret;
+
+	ret = gf100_gr_fecs_discover_reglist_image_size(gr, &size);
+	if (ret)
+		return ret;
+
+	/*XXX: We need to allocate + map the above into PMU's inst block,
+	 *     which which means we probably need a proper PMU before we
+	 *     even bother.
+	 */
+
+	ret = gf100_gr_fecs_set_reglist_bind_instance(gr, 0);
+	if (ret)
+		return ret;
+
+	return gf100_gr_fecs_set_reglist_virtual_address(gr, 0);
+}
+
+static int
+gf100_gr_fecs_discover_pm_image_size(struct gf100_gr *gr, u32 *psize)
+{
+	struct nvkm_device *device = gr->base.engine.subdev.device;
+
+	nvkm_wr32(device, 0x409840, 0xffffffff);
+	nvkm_wr32(device, 0x409500, 0x00000000);
+	nvkm_wr32(device, 0x409504, 0x00000025);
+	nvkm_msec(device, 2000,
+		if ((*psize = nvkm_rd32(device, 0x409800)))
+			return 0;
+	);
+
+	return -ETIMEDOUT;
+}
+
+static int
+gf100_gr_fecs_discover_zcull_image_size(struct gf100_gr *gr, u32 *psize)
+{
+	struct nvkm_device *device = gr->base.engine.subdev.device;
+
+	nvkm_wr32(device, 0x409840, 0xffffffff);
+	nvkm_wr32(device, 0x409500, 0x00000000);
+	nvkm_wr32(device, 0x409504, 0x00000016);
+	nvkm_msec(device, 2000,
+		if ((*psize = nvkm_rd32(device, 0x409800)))
+			return 0;
+	);
+
+	return -ETIMEDOUT;
+}
+
+static int
+gf100_gr_fecs_discover_image_size(struct gf100_gr *gr, u32 *psize)
+{
+	struct nvkm_device *device = gr->base.engine.subdev.device;
+
+	nvkm_wr32(device, 0x409840, 0xffffffff);
+	nvkm_wr32(device, 0x409500, 0x00000000);
+	nvkm_wr32(device, 0x409504, 0x00000010);
+	nvkm_msec(device, 2000,
+		if ((*psize = nvkm_rd32(device, 0x409800)))
+			return 0;
+	);
+
+	return -ETIMEDOUT;
+}
+
+static void
+gf100_gr_fecs_set_watchdog_timeout(struct gf100_gr *gr, u32 timeout)
+{
+	struct nvkm_device *device = gr->base.engine.subdev.device;
+
+	nvkm_wr32(device, 0x409840, 0xffffffff);
+	nvkm_wr32(device, 0x409500, timeout);
+	nvkm_wr32(device, 0x409504, 0x00000021);
+}
+
 static bool
 gf100_gr_chsw_load(struct nvkm_gr *base)
 {
···
 	struct nvkm_device *device = subdev->device;
 	struct nvkm_secboot *sb = device->secboot;
 	u32 secboot_mask = 0;
+	int ret;
 
 	/* load fuc microcode */
 	nvkm_mc_unk260(device, 0);
···
 	if (nvkm_secboot_is_managed(sb, NVKM_SECBOOT_FALCON_FECS))
 		secboot_mask |= BIT(NVKM_SECBOOT_FALCON_FECS);
 	else
-		gf100_gr_init_fw(gr->fecs, &gr->fuc409c, &gr->fuc409d);
+		gf100_gr_init_fw(gr->fecs.falcon, &gr->fuc409c, &gr->fuc409d);
 
 	if (nvkm_secboot_is_managed(sb, NVKM_SECBOOT_FALCON_GPCCS))
 		secboot_mask |= BIT(NVKM_SECBOOT_FALCON_GPCCS);
 	else
-		gf100_gr_init_fw(gr->gpccs, &gr->fuc41ac, &gr->fuc41ad);
+		gf100_gr_init_fw(gr->gpccs.falcon, &gr->fuc41ac, &gr->fuc41ad);
 
 	if (secboot_mask != 0) {
 		int ret = nvkm_secboot_reset(sb, secboot_mask);
···
 	nvkm_wr32(device, 0x41a10c, 0x00000000);
 	nvkm_wr32(device, 0x40910c, 0x00000000);
 
-	nvkm_falcon_start(gr->gpccs);
-	nvkm_falcon_start(gr->fecs);
+	nvkm_falcon_start(gr->gpccs.falcon);
+	nvkm_falcon_start(gr->fecs.falcon);
 
 	if (nvkm_msec(device, 2000,
 		if (nvkm_rd32(device, 0x409800) & 0x00000001)
···
 	) < 0)
 		return -EBUSY;
 
-	nvkm_wr32(device, 0x409840, 0xffffffff);
-	nvkm_wr32(device, 0x409500, 0x7fffffff);
-	nvkm_wr32(device, 0x409504, 0x00000021);
+	gf100_gr_fecs_set_watchdog_timeout(gr, 0x7fffffff);
 
-	nvkm_wr32(device, 0x409840, 0xffffffff);
-	nvkm_wr32(device, 0x409500, 0x00000000);
-	nvkm_wr32(device, 0x409504, 0x00000010);
-	if (nvkm_msec(device, 2000,
-		if ((gr->size = nvkm_rd32(device, 0x409800)))
-			break;
-	) < 0)
-		return -EBUSY;
+	/* Determine how much memory is required to store main context image. */
+	ret = gf100_gr_fecs_discover_image_size(gr, &gr->size);
+	if (ret)
+		return ret;
 
-	nvkm_wr32(device, 0x409840, 0xffffffff);
-	nvkm_wr32(device, 0x409500, 0x00000000);
-	nvkm_wr32(device, 0x409504, 0x00000016);
-	if (nvkm_msec(device, 2000,
-		if (nvkm_rd32(device, 0x409800))
-			break;
-	) < 0)
-		return -EBUSY;
+	/* Determine how much memory is required to store ZCULL image. */
+	ret = gf100_gr_fecs_discover_zcull_image_size(gr, &gr->size_zcull);
+	if (ret)
+		return ret;
 
-	nvkm_wr32(device, 0x409840, 0xffffffff);
-	nvkm_wr32(device, 0x409500, 0x00000000);
-	nvkm_wr32(device, 0x409504, 0x00000025);
-	if (nvkm_msec(device, 2000,
-		if (nvkm_rd32(device, 0x409800))
-			break;
-	) < 0)
-		return -EBUSY;
+	/* Determine how much memory is required to store PerfMon image. */
+	ret = gf100_gr_fecs_discover_pm_image_size(gr, &gr->size_pm);
+	if (ret)
+		return ret;
 
-	if (device->chipset >= 0xe0) {
-		nvkm_wr32(device, 0x409800, 0x00000000);
-		nvkm_wr32(device, 0x409500, 0x00000001);
-		nvkm_wr32(device, 0x409504, 0x00000030);
-		if (nvkm_msec(device, 2000,
-			if (nvkm_rd32(device, 0x409800))
-				break;
-		) < 0)
-			return -EBUSY;
-
-		nvkm_wr32(device, 0x409810, 0xb00095c8);
-		nvkm_wr32(device, 0x409800, 0x00000000);
-		nvkm_wr32(device, 0x409500, 0x00000001);
-		nvkm_wr32(device, 0x409504, 0x00000031);
-		if (nvkm_msec(device, 2000,
-			if (nvkm_rd32(device, 0x409800))
-				break;
-		) < 0)
-			return -EBUSY;
-
-		nvkm_wr32(device, 0x409810, 0x00080420);
-		nvkm_wr32(device, 0x409800, 0x00000000);
-		nvkm_wr32(device, 0x409500, 0x00000001);
-		nvkm_wr32(device, 0x409504, 0x00000032);
-		if (nvkm_msec(device, 2000,
-			if (nvkm_rd32(device, 0x409800))
-				break;
-		) < 0)
-			return -EBUSY;
-
-		nvkm_wr32(device, 0x409614, 0x00000070);
-		nvkm_wr32(device, 0x409614, 0x00000770);
-		nvkm_wr32(device, 0x40802c, 0x00000001);
+	/*XXX: We (likely) require PMU support to even bother with this.
+	 *
+	 *     Also, it seems like not all GPUs support ELPG.  Traces I
+	 *     have here show RM enabling it on Kepler/Turing, but none
+	 *     of the GPUs between those.  NVGPU decides this by PCIID.
+	 */
+	if (0) {
+		ret = gf100_gr_fecs_elpg_bind(gr);
+		if (ret)
+			return ret;
 	}
 
+	/* Generate golden context image. */
 	if (gr->data == NULL) {
 		int ret = gf100_grctx_generate(gr);
 		if (ret) {
···
 
 	/* load HUB microcode */
 	nvkm_mc_unk260(device, 0);
-	nvkm_falcon_load_dmem(gr->fecs, gr->func->fecs.ucode->data.data, 0x0,
+	nvkm_falcon_load_dmem(gr->fecs.falcon,
+			      gr->func->fecs.ucode->data.data, 0x0,
 			      gr->func->fecs.ucode->data.size, 0);
-	nvkm_falcon_load_imem(gr->fecs, gr->func->fecs.ucode->code.data, 0x0,
+	nvkm_falcon_load_imem(gr->fecs.falcon,
+			      gr->func->fecs.ucode->code.data, 0x0,
 			      gr->func->fecs.ucode->code.size, 0, 0, false);
 
 	/* load GPC microcode */
-	nvkm_falcon_load_dmem(gr->gpccs, gr->func->gpccs.ucode->data.data, 0x0,
+	nvkm_falcon_load_dmem(gr->gpccs.falcon,
+			      gr->func->gpccs.ucode->data.data, 0x0,
 			      gr->func->gpccs.ucode->data.size, 0);
-	nvkm_falcon_load_imem(gr->gpccs, gr->func->gpccs.ucode->code.data, 0x0,
+	nvkm_falcon_load_imem(gr->gpccs.falcon,
+			      gr->func->gpccs.ucode->code.data, 0x0,
 			      gr->func->gpccs.ucode->code.size, 0, 0, false);
 	nvkm_mc_unk260(device, 1);
 
···
 	int i, j;
 	int ret;
 
-	ret = nvkm_falcon_v1_new(subdev, "FECS", 0x409000, &gr->fecs);
+	ret = nvkm_falcon_v1_new(subdev, "FECS", 0x409000, &gr->fecs.falcon);
 	if (ret)
 		return ret;
 
-	ret = nvkm_falcon_v1_new(subdev, "GPCCS", 0x41a000, &gr->gpccs);
+	mutex_init(&gr->fecs.mutex);
+
+	ret = nvkm_falcon_v1_new(subdev, "GPCCS", 0x41a000, &gr->gpccs.falcon);
 	if (ret)
 		return ret;
 
···
 
 	nvkm_pmu_pgob(gr->base.engine.subdev.device->pmu, false);
 
-	ret = nvkm_falcon_get(gr->fecs, subdev);
+	ret = nvkm_falcon_get(gr->fecs.falcon, subdev);
 	if (ret)
 		return ret;
 
-	ret = nvkm_falcon_get(gr->gpccs, subdev);
+	ret = nvkm_falcon_get(gr->gpccs.falcon, subdev);
 	if (ret)
 		return ret;
 
···
 {
 	struct gf100_gr *gr = gf100_gr(base);
 	struct nvkm_subdev *subdev = &gr->base.engine.subdev;
-	nvkm_falcon_put(gr->gpccs, subdev);
-	nvkm_falcon_put(gr->fecs, subdev);
+	nvkm_falcon_put(gr->gpccs.falcon, subdev);
+	nvkm_falcon_put(gr->fecs.falcon, subdev);
 	return 0;
 }
 
···
 		gr->func->dtor(gr);
 	kfree(gr->data);
 
-	nvkm_falcon_del(&gr->gpccs);
-	nvkm_falcon_del(&gr->fecs);
+	nvkm_falcon_del(&gr->gpccs.falcon);
+	nvkm_falcon_del(&gr->fecs.falcon);
 
 	gf100_gr_dtor_fw(&gr->fuc409c);
 	gf100_gr_dtor_fw(&gr->fuc409d);
···
 	.chan_new = gf100_gr_chan_new,
 	.object_get = gf100_gr_object_get,
 	.chsw_load = gf100_gr_chsw_load,
+	.ctxsw.pause = gf100_gr_fecs_stop_ctxsw,
+	.ctxsw.resume = gf100_gr_fecs_start_ctxsw,
+	.ctxsw.inst = gf100_gr_ctxsw_inst,
 };
 
 int

drivers/gpu/drm/nouveau/nvkm/engine/gr/gf100.h

···
 	const struct gf100_gr_func *func;
 	struct nvkm_gr base;
 
-	struct nvkm_falcon *fecs;
-	struct nvkm_falcon *gpccs;
+	struct {
+		struct nvkm_falcon *falcon;
+		struct mutex mutex;
+		u32 disable;
+	} fecs;
+
+	struct {
+		struct nvkm_falcon *falcon;
+	} gpccs;
+
 	struct gf100_gr_fuc fuc409c;
 	struct gf100_gr_fuc fuc409d;
 	struct gf100_gr_fuc fuc41ac;
···
 	struct gf100_gr_mmio mmio_list[4096/8];
 	u32  size;
 	u32 *data;
+	u32 size_zcull;
+	u32 size_pm;
 };
 
 int gf100_gr_ctor(const struct gf100_gr_func *, struct nvkm_device *,
···
 int gf100_gr_new_(const struct gf100_gr_func *, struct nvkm_device *,
 		  int, struct nvkm_gr **);
 void *gf100_gr_dtor(struct nvkm_gr *);
+
+int gf100_gr_fecs_bind_pointer(struct gf100_gr *, u32 inst);
 
 struct gf100_gr_func_zbc {
 	void (*clear_color)(struct gf100_gr *, int zbc);

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

···
 	 */
 	u64 (*units)(struct nvkm_gr *);
 	bool (*chsw_load)(struct nvkm_gr *);
+	struct {
+		int (*pause)(struct nvkm_gr *);
+		int (*resume)(struct nvkm_gr *);
+		u32 (*inst)(struct nvkm_gr *);
+	} ctxsw;
 	struct nvkm_sclass sclass[];
 };
 

drivers/gpu/drm/nouveau/nvkm/engine/nvdec/base.c

···
  */
 #include "priv.h"
 
+#include <subdev/top.h>
 #include <engine/falcon.h>
 
 static int
 nvkm_nvdec_oneinit(struct nvkm_engine *engine)
 {
 	struct nvkm_nvdec *nvdec = nvkm_nvdec(engine);
-	return nvkm_falcon_v1_new(&nvdec->engine.subdev, "NVDEC", 0x84000,
+	struct nvkm_subdev *subdev = &nvdec->engine.subdev;
+
+	nvdec->addr = nvkm_top_addr(subdev->device, subdev->index);
+	if (!nvdec->addr)
+		return -EINVAL;
+
+	/*XXX: fix naming of this when adding support for multiple-NVDEC */
+	return nvkm_falcon_v1_new(subdev, "NVDEC", nvdec->addr,
 				  &nvdec->falcon);
 }
 

drivers/gpu/drm/nouveau/nvkm/engine/sec2/Kbuild

···
 nvkm-y += nvkm/engine/sec2/base.o
 nvkm-y += nvkm/engine/sec2/gp102.o
+nvkm-y += nvkm/engine/sec2/tu102.o

drivers/gpu/drm/nouveau/nvkm/engine/sec2/base.c

···
 #include "priv.h"
 
 #include <core/msgqueue.h>
+#include <subdev/top.h>
 #include <engine/falcon.h>
 
 static void *
···
 	struct nvkm_sec2 *sec2 = nvkm_sec2(engine);
 	struct nvkm_subdev *subdev = &engine->subdev;
 	struct nvkm_device *device = subdev->device;
-	u32 disp = nvkm_rd32(device, 0x8701c);
-	u32 intr = nvkm_rd32(device, 0x87008) & disp & ~(disp >> 16);
+	u32 disp = nvkm_rd32(device, sec2->addr + 0x01c);
+	u32 intr = nvkm_rd32(device, sec2->addr + 0x008) & disp & ~(disp >> 16);
 
 	if (intr & 0x00000040) {
 		schedule_work(&sec2->work);
-		nvkm_wr32(device, 0x87004, 0x00000040);
+		nvkm_wr32(device, sec2->addr + 0x004, 0x00000040);
 		intr &= ~0x00000040;
 	}
 
 	if (intr) {
 		nvkm_error(subdev, "unhandled intr %08x\n", intr);
-		nvkm_wr32(device, 0x87004, intr);
+		nvkm_wr32(device, sec2->addr + 0x004, intr);
 
 	}
 }
···
 nvkm_sec2_oneinit(struct nvkm_engine *engine)
 {
 	struct nvkm_sec2 *sec2 = nvkm_sec2(engine);
-	return nvkm_falcon_v1_new(&sec2->engine.subdev, "SEC2", 0x87000,
-				  &sec2->falcon);
+	struct nvkm_subdev *subdev = &sec2->engine.subdev;
+
+	if (!sec2->addr) {
+		sec2->addr = nvkm_top_addr(subdev->device, subdev->index);
+		if (WARN_ON(!sec2->addr))
+			return -EINVAL;
+	}
+
+	return nvkm_falcon_v1_new(subdev, "SEC2", sec2->addr, &sec2->falcon);
 }
 
 static int
···
 };
 
 int
-nvkm_sec2_new_(struct nvkm_device *device, int index,
+nvkm_sec2_new_(struct nvkm_device *device, int index, u32 addr,
 	       struct nvkm_sec2 **psec2)
 {
 	struct nvkm_sec2 *sec2;
 
 	if (!(sec2 = *psec2 = kzalloc(sizeof(*sec2), GFP_KERNEL)))
 		return -ENOMEM;
+	sec2->addr = addr;
 	INIT_WORK(&sec2->work, nvkm_sec2_recv);
 
 	return nvkm_engine_ctor(&nvkm_sec2, device, index, true, &sec2->engine);

drivers/gpu/drm/nouveau/nvkm/engine/sec2/gp102.c

···
 gp102_sec2_new(struct nvkm_device *device, int index,
 	       struct nvkm_sec2 **psec2)
 {
-	return nvkm_sec2_new_(device, index, psec2);
+	return nvkm_sec2_new_(device, index, 0, psec2);
 }

drivers/gpu/drm/nouveau/nvkm/engine/sec2/priv.h

···
 
 #define nvkm_sec2(p) container_of((p), struct nvkm_sec2, engine)
 
-int nvkm_sec2_new_(struct nvkm_device *, int, struct nvkm_sec2 **);
-
+int nvkm_sec2_new_(struct nvkm_device *, int, u32 addr, struct nvkm_sec2 **);
 #endif

drivers/gpu/drm/nouveau/nvkm/engine/sec2/tu102.c

···
+/*
+ * Copyright 2019 Red Hat Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#include "priv.h"
+
+int
+tu102_sec2_new(struct nvkm_device *device, int index,
+	       struct nvkm_sec2 **psec2)
+{
+	/* TOP info wasn't updated on Turing to reflect the PRI
+	 * address change for some reason.  We override it here.
+	 */
+	return nvkm_sec2_new_(device, index, 0x840000, psec2);
+}

drivers/gpu/drm/nouveau/nvkm/falcon/base.c

···
 		debug_reg = 0x408;
 		falcon->has_emem = true;
 		break;
+	case NVKM_SUBDEV_GSP:
+		debug_reg = 0x0; /*XXX*/
+		break;
 	default:
 		nvkm_warn(subdev, "unsupported falcon %s!\n",
 			  nvkm_subdev_name[subdev->index]);

drivers/gpu/drm/nouveau/nvkm/falcon/msgqueue.c

···
 		commit = false;
 	}
 
-	   cmd_queue_close(priv, queue, commit);
+	cmd_queue_close(priv, queue, commit);
 
 	return ret;
 }
···
 	ret = cmd_write(priv, cmd, queue);
 	if (ret) {
 		seq->state = SEQ_STATE_PENDING;
-		      msgqueue_seq_release(priv, seq);
+		msgqueue_seq_release(priv, seq);
 	}
 
 	return ret;
···
 	if (seq->completion)
 		complete(seq->completion);
 
-	   msgqueue_seq_release(priv, seq);
+	msgqueue_seq_release(priv, seq);
 
 	return 0;
 }

drivers/gpu/drm/nouveau/nvkm/subdev/Kbuild

···
 include $(src)/nvkm/subdev/fb/Kbuild
 include $(src)/nvkm/subdev/fuse/Kbuild
 include $(src)/nvkm/subdev/gpio/Kbuild
+include $(src)/nvkm/subdev/gsp/Kbuild
 include $(src)/nvkm/subdev/i2c/Kbuild
 include $(src)/nvkm/subdev/ibus/Kbuild
 include $(src)/nvkm/subdev/iccsense/Kbuild

drivers/gpu/drm/nouveau/nvkm/subdev/bar/Kbuild

···
 nvkm-y += nvkm/subdev/bar/gk20a.o
 nvkm-y += nvkm/subdev/bar/gm107.o
 nvkm-y += nvkm/subdev/bar/gm20b.o
-nvkm-y += nvkm/subdev/bar/tu104.o
+nvkm-y += nvkm/subdev/bar/tu102.o

drivers/gpu/drm/nouveau/nvkm/subdev/bar/tu104.c drivers/gpu/drm/nouveau/nvkm/subdev/bar/tu102.c

···
 #include <subdev/timer.h>
 
 static void
-tu104_bar_bar2_wait(struct nvkm_bar *bar)
+tu102_bar_bar2_wait(struct nvkm_bar *bar)
 {
 	struct nvkm_device *device = bar->subdev.device;
 	nvkm_msec(device, 2000,
···
 }
 
 static void
-tu104_bar_bar2_fini(struct nvkm_bar *bar)
+tu102_bar_bar2_fini(struct nvkm_bar *bar)
 {
 	nvkm_mask(bar->subdev.device, 0xb80f48, 0x80000000, 0x00000000);
 }
 
 static void
-tu104_bar_bar2_init(struct nvkm_bar *base)
+tu102_bar_bar2_init(struct nvkm_bar *base)
 {
 	struct nvkm_device *device = base->subdev.device;
 	struct gf100_bar *bar = gf100_bar(base);
···
 }
 
 static void
-tu104_bar_bar1_wait(struct nvkm_bar *bar)
+tu102_bar_bar1_wait(struct nvkm_bar *bar)
 {
 	struct nvkm_device *device = bar->subdev.device;
 	nvkm_msec(device, 2000,
···
 }
 
 static void
-tu104_bar_bar1_fini(struct nvkm_bar *bar)
+tu102_bar_bar1_fini(struct nvkm_bar *bar)
 {
 	nvkm_mask(bar->subdev.device, 0xb80f40, 0x80000000, 0x00000000);
 }
 
 static void
-tu104_bar_bar1_init(struct nvkm_bar *base)
+tu102_bar_bar1_init(struct nvkm_bar *base)
 {
 	struct nvkm_device *device = base->subdev.device;
 	struct gf100_bar *bar = gf100_bar(base);
···
 }
 
 static const struct nvkm_bar_func
-tu104_bar = {
+tu102_bar = {
 	.dtor = gf100_bar_dtor,
 	.oneinit = gf100_bar_oneinit,
-	.bar1.init = tu104_bar_bar1_init,
-	.bar1.fini = tu104_bar_bar1_fini,
-	.bar1.wait = tu104_bar_bar1_wait,
+	.bar1.init = tu102_bar_bar1_init,
+	.bar1.fini = tu102_bar_bar1_fini,
+	.bar1.wait = tu102_bar_bar1_wait,
 	.bar1.vmm = gf100_bar_bar1_vmm,
-	.bar2.init = tu104_bar_bar2_init,
-	.bar2.fini = tu104_bar_bar2_fini,
-	.bar2.wait = tu104_bar_bar2_wait,
+	.bar2.init = tu102_bar_bar2_init,
+	.bar2.fini = tu102_bar_bar2_fini,
+	.bar2.wait = tu102_bar_bar2_wait,
 	.bar2.vmm = gf100_bar_bar2_vmm,
 	.flush = g84_bar_flush,
 };
 
 int
-tu104_bar_new(struct nvkm_device *device, int index, struct nvkm_bar **pbar)
+tu102_bar_new(struct nvkm_device *device, int index, struct nvkm_bar **pbar)
 {
-	return gf100_bar_new_(&tu104_bar, device, index, pbar);
+	return gf100_bar_new_(&tu102_bar, device, index, pbar);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/bios/dp.c

···
 	u16 data;
 
 	if (*ver >= 0x30) {
-		const u8 vsoff[] = { 0, 4, 7, 9 };
+		static const u8 vsoff[] = { 0, 4, 7, 9 };
 		idx = (pc * 10) + vsoff[vs] + pe;
 		if (*ver >= 0x40 && *ver <= 0x41 && *hdr >= 0x12)
 			idx += nvbios_rd08(bios, outp + 0x11) * 40;

drivers/gpu/drm/nouveau/nvkm/subdev/bios/init.c

···
 	init->offset += 3;
 
 	switch (cond) {
-	case 0:
+	case 0: /* CONDITION_ID_INT_DP. */
 		if (init_conn(init) != DCB_CONNECTOR_eDP)
 			init_exec_set(init, false);
 		break;
-	case 1:
-	case 2:
+	case 1: /* CONDITION_ID_USE_SPPLL0. */
+	case 2: /* CONDITION_ID_USE_SPPLL1. */
 		if ( init->outp &&
 		    (data = nvbios_dpout_match(bios, DCB_OUTPUT_DP,
 					       (init->outp->or << 0) |
···
 		if (init_exec(init))
 			warn("script needs dp output table data\n");
 		break;
-	case 5:
+	case 5: /* CONDITION_ID_ASSR_SUPPORT. */
 		if (!(init_rdauxr(init, 0x0d) & 1))
 			init_exec_set(init, false);
+		break;
+	case 7: /* CONDITION_ID_NO_PANEL_SEQ_DELAYS. */
+		init_exec_set(init, false);
 		break;
 	default:
 		warn("INIT_GENERIC_CONDITON: unknown 0x%02x\n", cond);

drivers/gpu/drm/nouveau/nvkm/subdev/devinit/Kbuild

···
 nvkm-y += nvkm/subdev/devinit/gm107.o
 nvkm-y += nvkm/subdev/devinit/gm200.o
 nvkm-y += nvkm/subdev/devinit/gv100.o
-nvkm-y += nvkm/subdev/devinit/tu104.o
+nvkm-y += nvkm/subdev/devinit/tu102.o

drivers/gpu/drm/nouveau/nvkm/subdev/devinit/tu104.c drivers/gpu/drm/nouveau/nvkm/subdev/devinit/tu102.c

···
 #include <subdev/clk/pll.h>
 
 static int
-tu104_devinit_pll_set(struct nvkm_devinit *init, u32 type, u32 freq)
+tu102_devinit_pll_set(struct nvkm_devinit *init, u32 type, u32 freq)
 {
 	struct nvkm_subdev *subdev = &init->subdev;
 	struct nvkm_device *device = subdev->device;
···
 }
 
 static int
-tu104_devinit_post(struct nvkm_devinit *base, bool post)
+tu102_devinit_post(struct nvkm_devinit *base, bool post)
 {
 	struct nv50_devinit *init = nv50_devinit(base);
 	gm200_devinit_preos(init, post);
···
 }
 
 static const struct nvkm_devinit_func
-tu104_devinit = {
+tu102_devinit = {
 	.init = nv50_devinit_init,
-	.post = tu104_devinit_post,
-	.pll_set = tu104_devinit_pll_set,
+	.post = tu102_devinit_post,
+	.pll_set = tu102_devinit_pll_set,
 	.disable = gm107_devinit_disable,
 };
 
 int
-tu104_devinit_new(struct nvkm_device *device, int index,
+tu102_devinit_new(struct nvkm_device *device, int index,
 		struct nvkm_devinit **pinit)
 {
-	return nv50_devinit_new_(&tu104_devinit, device, index, pinit);
+	return nv50_devinit_new_(&tu102_devinit, device, index, pinit);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/fault/Kbuild

···
 nvkm-y += nvkm/subdev/fault/base.o
+nvkm-y += nvkm/subdev/fault/user.o
 nvkm-y += nvkm/subdev/fault/gp100.o
 nvkm-y += nvkm/subdev/fault/gv100.o
-nvkm-y += nvkm/subdev/fault/tu104.o
+nvkm-y += nvkm/subdev/fault/tu102.o

drivers/gpu/drm/nouveau/nvkm/subdev/fault/base.c

···
 		return -ENOMEM;
 	nvkm_subdev_ctor(&nvkm_fault, device, index, &fault->subdev);
 	fault->func = func;
+	fault->user.ctor = nvkm_ufault_new;
+	fault->user.base = func->user.base;
 	return 0;
 }

drivers/gpu/drm/nouveau/nvkm/subdev/fault/gp100.c

···
 
 #include <subdev/mc.h>
 
+#include <nvif/class.h>
+
 static void
 gp100_fault_buffer_intr(struct nvkm_fault_buffer *buffer, bool enable)
 {
···
 	.buffer.init = gp100_fault_buffer_init,
 	.buffer.fini = gp100_fault_buffer_fini,
 	.buffer.intr = gp100_fault_buffer_intr,
+	.user = { { 0, 0, MAXWELL_FAULT_BUFFER_A }, 0 },
 };
 
 int

drivers/gpu/drm/nouveau/nvkm/subdev/fault/gv100.c

···
 #include <subdev/mmu.h>
 #include <engine/fifo.h>
 
+#include <nvif/class.h>
+
 static void
 gv100_fault_buffer_process(struct nvkm_fault_buffer *buffer)
 {
···
 		}
 	}
 
+	if (stat & 0x08000000) {
+		if (fault->buffer[1]) {
+			nvkm_event_send(&fault->event, 1, 1, NULL, 0);
+			stat &= ~0x08000000;
+		}
+	}
+
 	if (stat) {
 		nvkm_debug(subdev, "intr %08x\n", stat);
 	}
···
 	.buffer.init = gv100_fault_buffer_init,
 	.buffer.fini = gv100_fault_buffer_fini,
 	.buffer.intr = gv100_fault_buffer_intr,
+	/*TODO: Figure out how to expose non-replayable fault buffer, which,
+	 *      for some reason, is where recoverable CE faults appear...
+	 *
+	 * 	It's a bit tricky, as both NVKM and SVM will need access to
+	 * 	the non-replayable fault buffer.
+	 */
+	.user = { { 0, 0, VOLTA_FAULT_BUFFER_A }, 1 },
 };
 
 int

drivers/gpu/drm/nouveau/nvkm/subdev/fault/priv.h

···
 		void (*fini)(struct nvkm_fault_buffer *);
 		void (*intr)(struct nvkm_fault_buffer *, bool enable);
 	} buffer;
+	struct {
+		struct nvkm_sclass base;
+		int rp;
+	} user;
 };
 
 int gv100_fault_oneinit(struct nvkm_fault *);
+
+int nvkm_ufault_new(struct nvkm_device *, const struct nvkm_oclass *,
+		    void *, u32, struct nvkm_object **);
 #endif

drivers/gpu/drm/nouveau/nvkm/subdev/fault/tu104.c drivers/gpu/drm/nouveau/nvkm/subdev/fault/tu102.c

···
 #include <nvif/class.h>
 
 static void
-tu104_fault_buffer_intr(struct nvkm_fault_buffer *buffer, bool enable)
+tu102_fault_buffer_intr(struct nvkm_fault_buffer *buffer, bool enable)
 {
 	/*XXX: Earlier versions of RM touched the old regs on Turing,
 	 *     which don't appear to actually work anymore, but newer
···
 }
 
 static void
-tu104_fault_buffer_fini(struct nvkm_fault_buffer *buffer)
+tu102_fault_buffer_fini(struct nvkm_fault_buffer *buffer)
 {
 	struct nvkm_device *device = buffer->fault->subdev.device;
 	const u32 foff = buffer->id * 0x20;
···
 }
 
 static void
-tu104_fault_buffer_init(struct nvkm_fault_buffer *buffer)
+tu102_fault_buffer_init(struct nvkm_fault_buffer *buffer)
 {
 	struct nvkm_device *device = buffer->fault->subdev.device;
 	const u32 foff = buffer->id * 0x20;
···
 }
 
 static void
-tu104_fault_buffer_info(struct nvkm_fault_buffer *buffer)
+tu102_fault_buffer_info(struct nvkm_fault_buffer *buffer)
 {
 	struct nvkm_device *device = buffer->fault->subdev.device;
 	const u32 foff = buffer->id * 0x20;
···
 }
 
 static void
-tu104_fault_intr_fault(struct nvkm_fault *fault)
+tu102_fault_intr_fault(struct nvkm_fault *fault)
 {
 	struct nvkm_subdev *subdev = &fault->subdev;
 	struct nvkm_device *device = subdev->device;
···
 }
 
 static void
-tu104_fault_intr(struct nvkm_fault *fault)
+tu102_fault_intr(struct nvkm_fault *fault)
 {
 	struct nvkm_subdev *subdev = &fault->subdev;
 	struct nvkm_device *device = subdev->device;
 	u32 stat = nvkm_rd32(device, 0xb83094);
 
 	if (stat & 0x80000000) {
-		tu104_fault_intr_fault(fault);
+		tu102_fault_intr_fault(fault);
 		nvkm_wr32(device, 0xb83094, 0x80000000);
 		stat &= ~0x80000000;
 	}
···
 }
 
 static void
-tu104_fault_fini(struct nvkm_fault *fault)
+tu102_fault_fini(struct nvkm_fault *fault)
 {
 	nvkm_notify_put(&fault->nrpfb);
 	if (fault->buffer[0])
···
 }
 
 static void
-tu104_fault_init(struct nvkm_fault *fault)
+tu102_fault_init(struct nvkm_fault *fault)
 {
 	/*XXX: enable priv faults */
 	fault->func->buffer.init(fault->buffer[0]);
···
 }
 
 static const struct nvkm_fault_func
-tu104_fault = {
+tu102_fault = {
 	.oneinit = gv100_fault_oneinit,
-	.init = tu104_fault_init,
-	.fini = tu104_fault_fini,
-	.intr = tu104_fault_intr,
+	.init = tu102_fault_init,
+	.fini = tu102_fault_fini,
+	.intr = tu102_fault_intr,
 	.buffer.nr = 2,
 	.buffer.entry_size = 32,
-	.buffer.info = tu104_fault_buffer_info,
-	.buffer.init = tu104_fault_buffer_init,
-	.buffer.fini = tu104_fault_buffer_fini,
-	.buffer.intr = tu104_fault_buffer_intr,
+	.buffer.info = tu102_fault_buffer_info,
+	.buffer.init = tu102_fault_buffer_init,
+	.buffer.fini = tu102_fault_buffer_fini,
+	.buffer.intr = tu102_fault_buffer_intr,
+	.user = { { 0, 0, VOLTA_FAULT_BUFFER_A }, 1 },
 };
 
 int
-tu104_fault_new(struct nvkm_device *device, int index,
+tu102_fault_new(struct nvkm_device *device, int index,
 		struct nvkm_fault **pfault)
 {
-	return nvkm_fault_new_(&tu104_fault, device, index, pfault);
+	return nvkm_fault_new_(&tu102_fault, device, index, pfault);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/fault/user.c

···
+/*
+ * Copyright 2018 Red Hat Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+#include "priv.h"
+
+#include <core/memory.h>
+#include <subdev/mmu.h>
+
+#include <nvif/clb069.h>
+#include <nvif/unpack.h>
+
+static int
+nvkm_ufault_map(struct nvkm_object *object, void *argv, u32 argc,
+		enum nvkm_object_map *type, u64 *addr, u64 *size)
+{
+	struct nvkm_fault_buffer *buffer = nvkm_fault_buffer(object);
+	struct nvkm_device *device = buffer->fault->subdev.device;
+	*type = NVKM_OBJECT_MAP_IO;
+	*addr = device->func->resource_addr(device, 3) + buffer->addr;
+	*size = nvkm_memory_size(buffer->mem);
+	return 0;
+}
+
+static int
+nvkm_ufault_ntfy(struct nvkm_object *object, u32 type,
+		 struct nvkm_event **pevent)
+{
+	struct nvkm_fault_buffer *buffer = nvkm_fault_buffer(object);
+	if (type == NVB069_V0_NTFY_FAULT) {
+		*pevent = &buffer->fault->event;
+		return 0;
+	}
+	return -EINVAL;
+}
+
+static int
+nvkm_ufault_fini(struct nvkm_object *object, bool suspend)
+{
+	struct nvkm_fault_buffer *buffer = nvkm_fault_buffer(object);
+	buffer->fault->func->buffer.fini(buffer);
+	return 0;
+}
+
+static int
+nvkm_ufault_init(struct nvkm_object *object)
+{
+	struct nvkm_fault_buffer *buffer = nvkm_fault_buffer(object);
+	buffer->fault->func->buffer.init(buffer);
+	return 0;
+}
+
+static void *
+nvkm_ufault_dtor(struct nvkm_object *object)
+{
+	return NULL;
+}
+
+static const struct nvkm_object_func
+nvkm_ufault = {
+	.dtor = nvkm_ufault_dtor,
+	.init = nvkm_ufault_init,
+	.fini = nvkm_ufault_fini,
+	.ntfy = nvkm_ufault_ntfy,
+	.map = nvkm_ufault_map,
+};
+
+int
+nvkm_ufault_new(struct nvkm_device *device, const struct nvkm_oclass *oclass,
+		void *argv, u32 argc, struct nvkm_object **pobject)
+{
+	union {
+		struct nvif_clb069_v0 v0;
+	} *args = argv;
+	struct nvkm_fault *fault = device->fault;
+	struct nvkm_fault_buffer *buffer = fault->buffer[fault->func->user.rp];
+	int ret = -ENOSYS;
+
+	if (!(ret = nvif_unpack(ret, &argv, &argc, args->v0, 0, 0, false))) {
+		args->v0.entries = buffer->entries;
+		args->v0.get = buffer->get;
+		args->v0.put = buffer->put;
+	} else
+		return ret;
+
+	nvkm_object_ctor(&nvkm_ufault, oclass, &buffer->object);
+	*pobject = &buffer->object;
+	return 0;
+}

drivers/gpu/drm/nouveau/nvkm/subdev/fb/gddr3.c

···
 		WR  = (ram->next->bios.timing[2] & 0x007f0000) >> 16;
 		/* XXX: Get these values from the VBIOS instead */
 		DLL = !(ram->mr[1] & 0x1);
-		RON = !(ram->mr[1] & 0x300) >> 8;
+		RON = !((ram->mr[1] & 0x300) >> 8);
 		break;
 	default:
 		return -ENOSYS;

drivers/gpu/drm/nouveau/nvkm/subdev/gsp/Kbuild

···
+nvkm-y += nvkm/subdev/gsp/gv100.o

drivers/gpu/drm/nouveau/nvkm/subdev/gsp/gv100.c

···
+/*
+ * Copyright 2019 Red Hat Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+#include <subdev/gsp.h>
+#include <subdev/top.h>
+#include <engine/falcon.h>
+
+static int
+gv100_gsp_oneinit(struct nvkm_subdev *subdev)
+{
+	struct nvkm_gsp *gsp = nvkm_gsp(subdev);
+
+	gsp->addr = nvkm_top_addr(subdev->device, subdev->index);
+	if (!gsp->addr)
+		return -EINVAL;
+
+	return nvkm_falcon_v1_new(subdev, "GSP", gsp->addr, &gsp->falcon);
+}
+
+static void *
+gv100_gsp_dtor(struct nvkm_subdev *subdev)
+{
+	struct nvkm_gsp *gsp = nvkm_gsp(subdev);
+	nvkm_falcon_del(&gsp->falcon);
+	return gsp;
+}
+
+static const struct nvkm_subdev_func
+gv100_gsp = {
+	.dtor = gv100_gsp_dtor,
+	.oneinit = gv100_gsp_oneinit,
+};
+
+int
+gv100_gsp_new(struct nvkm_device *device, int index, struct nvkm_gsp **pgsp)
+{
+	struct nvkm_gsp *gsp;
+
+	if (!(gsp = *pgsp = kzalloc(sizeof(*gsp), GFP_KERNEL)))
+		return -ENOMEM;
+
+	nvkm_subdev_ctor(&gv100_gsp, device, index, &gsp->subdev);
+	return 0;
+}

drivers/gpu/drm/nouveau/nvkm/subdev/mc/Kbuild

···
 nvkm-y += nvkm/subdev/mc/gk20a.o
 nvkm-y += nvkm/subdev/mc/gp100.o
 nvkm-y += nvkm/subdev/mc/gp10b.o
-nvkm-y += nvkm/subdev/mc/tu104.o
+nvkm-y += nvkm/subdev/mc/tu102.o

drivers/gpu/drm/nouveau/nvkm/subdev/mc/tu102.c

···
+/*
+ * Copyright 2018 Red Hat Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+#include "priv.h"
+
+static void
+tu102_mc_intr_hack(struct nvkm_mc *mc, bool *handled)
+{
+	struct nvkm_device *device = mc->subdev.device;
+	u32 stat = nvkm_rd32(device, 0xb81010);
+	if (stat & 0x00000050) {
+		struct nvkm_subdev *subdev =
+			nvkm_device_subdev(device, NVKM_SUBDEV_FAULT);
+		nvkm_wr32(device, 0xb81010, stat & 0x00000050);
+		if (subdev)
+			nvkm_subdev_intr(subdev);
+		*handled = true;
+	}
+}
+
+static const struct nvkm_mc_func
+tu102_mc = {
+	.init = nv50_mc_init,
+	.intr = gp100_mc_intr,
+	.intr_unarm = gp100_mc_intr_unarm,
+	.intr_rearm = gp100_mc_intr_rearm,
+	.intr_mask = gp100_mc_intr_mask,
+	.intr_stat = gf100_mc_intr_stat,
+	.intr_hack = tu102_mc_intr_hack,
+	.reset = gk104_mc_reset,
+};
+
+int
+tu102_mc_new(struct nvkm_device *device, int index, struct nvkm_mc **pmc)
+{
+	return gp100_mc_new_(&tu102_mc, device, index, pmc);
+}

drivers/gpu/drm/nouveau/nvkm/subdev/mc/tu104.c drivers/gpu/drm/nouveau/nvkm/subdev/mmu/tu102.c

···
  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  * OTHER DEALINGS IN THE SOFTWARE.
  */
-#include "priv.h"
+#include "mem.h"
+#include "vmm.h"
 
-static void
-tu104_mc_intr_hack(struct nvkm_mc *mc, bool *handled)
-{
-	struct nvkm_device *device = mc->subdev.device;
-	u32 stat = nvkm_rd32(device, 0xb81010);
-	if (stat & 0x00000050) {
-		struct nvkm_subdev *subdev =
-			nvkm_device_subdev(device, NVKM_SUBDEV_FAULT);
-		nvkm_wr32(device, 0xb81010, stat & 0x00000050);
-		if (subdev)
-			nvkm_subdev_intr(subdev);
-		*handled = true;
-	}
-}
+#include <core/option.h>
 
-static const struct nvkm_mc_func
-tu104_mc = {
-	.init = nv50_mc_init,
-	.intr = gp100_mc_intr,
-	.intr_unarm = gp100_mc_intr_unarm,
-	.intr_rearm = gp100_mc_intr_rearm,
-	.intr_mask = gp100_mc_intr_mask,
-	.intr_stat = gf100_mc_intr_stat,
-	.intr_hack = tu104_mc_intr_hack,
-	.reset = gk104_mc_reset,
+#include <nvif/class.h>
+
+static const struct nvkm_mmu_func
+tu102_mmu = {
+	.dma_bits = 47,
+	.mmu = {{ -1, -1, NVIF_CLASS_MMU_GF100}},
+	.mem = {{ -1,  0, NVIF_CLASS_MEM_GF100}, gf100_mem_new, gf100_mem_map },
+	.vmm = {{ -1,  0, NVIF_CLASS_VMM_GP100}, tu102_vmm_new },
+	.kind = gm200_mmu_kind,
+	.kind_sys = true,
 };
 
 int
-tu104_mc_new(struct nvkm_device *device, int index, struct nvkm_mc **pmc)
+tu102_mmu_new(struct nvkm_device *device, int index, struct nvkm_mmu **pmmu)
 {
-	return gp100_mc_new_(&tu104_mc, device, index, pmc);
+	return nvkm_mmu_new_(&tu102_mmu, device, index, pmmu);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/Kbuild

···
 nvkm-y += nvkm/subdev/mmu/gp100.o
 nvkm-y += nvkm/subdev/mmu/gp10b.o
 nvkm-y += nvkm/subdev/mmu/gv100.o
-nvkm-y += nvkm/subdev/mmu/tu104.o
+nvkm-y += nvkm/subdev/mmu/tu102.o
 
 nvkm-y += nvkm/subdev/mmu/mem.o
 nvkm-y += nvkm/subdev/mmu/memnv04.o
···
 nvkm-y += nvkm/subdev/mmu/vmmgp100.o
 nvkm-y += nvkm/subdev/mmu/vmmgp10b.o
 nvkm-y += nvkm/subdev/mmu/vmmgv100.o
-nvkm-y += nvkm/subdev/mmu/vmmtu104.o
+nvkm-y += nvkm/subdev/mmu/vmmtu102.o
 
 nvkm-y += nvkm/subdev/mmu/umem.o
 nvkm-y += nvkm/subdev/mmu/ummu.o

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/gp100.c

···
 	.dma_bits = 47,
 	.mmu = {{ -1, -1, NVIF_CLASS_MMU_GF100}},
 	.mem = {{ -1,  0, NVIF_CLASS_MEM_GF100}, gf100_mem_new, gf100_mem_map },
-	.vmm = {{ -1, -1, NVIF_CLASS_VMM_GP100}, gp100_vmm_new },
+	.vmm = {{ -1,  0, NVIF_CLASS_VMM_GP100}, gp100_vmm_new },
 	.kind = gm200_mmu_kind,
 	.kind_sys = true,
 };

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/gp10b.c

···
 	.dma_bits = 47,
 	.mmu = {{ -1, -1, NVIF_CLASS_MMU_GF100}},
 	.mem = {{ -1, -1, NVIF_CLASS_MEM_GF100}, .umap = gf100_mem_map },
-	.vmm = {{ -1, -1, NVIF_CLASS_VMM_GP100}, gp10b_vmm_new },
+	.vmm = {{ -1,  0, NVIF_CLASS_VMM_GP100}, gp10b_vmm_new },
 	.kind = gm200_mmu_kind,
 	.kind_sys = true,
 };

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/gv100.c

···
 	.dma_bits = 47,
 	.mmu = {{ -1, -1, NVIF_CLASS_MMU_GF100}},
 	.mem = {{ -1,  0, NVIF_CLASS_MEM_GF100}, gf100_mem_new, gf100_mem_map },
-	.vmm = {{ -1, -1, NVIF_CLASS_VMM_GP100}, gv100_vmm_new },
+	.vmm = {{ -1,  0, NVIF_CLASS_VMM_GP100}, gv100_vmm_new },
 	.kind = gm200_mmu_kind,
 	.kind_sys = true,
 };

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h

···
 
 	struct {
 		struct nvkm_sclass user;
-		int (*ctor)(struct nvkm_mmu *, u64 addr, u64 size,
+		int (*ctor)(struct nvkm_mmu *, bool managed, u64 addr, u64 size,
 			    void *argv, u32 argc, struct lock_class_key *,
 			    const char *name, struct nvkm_vmm **);
 		bool global;

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/tu104.c drivers/gpu/drm/nouveau/nvkm/subdev/volt/gf117.c

···
 /*
- * Copyright 2018 Red Hat Inc.
+ * Copyright 2019 Ilia Mirkin
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
  * copy of this software and associated documentation files (the "Software"),
···
  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  * OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * Authors: Ilia Mirkin
  */
-#include "mem.h"
-#include "vmm.h"
+#include "priv.h"
 
-#include <core/option.h>
+#include <subdev/fuse.h>
 
-#include <nvif/class.h>
+static int
+gf117_volt_speedo_read(struct nvkm_volt *volt)
+{
+	struct nvkm_device *device = volt->subdev.device;
+	struct nvkm_fuse *fuse = device->fuse;
 
-static const struct nvkm_mmu_func
-tu104_mmu = {
-	.dma_bits = 47,
-	.mmu = {{ -1, -1, NVIF_CLASS_MMU_GF100}},
-	.mem = {{ -1,  0, NVIF_CLASS_MEM_GF100}, gf100_mem_new, gf100_mem_map },
-	.vmm = {{ -1,  0, NVIF_CLASS_VMM_GP100}, tu104_vmm_new },
-	.kind = gm200_mmu_kind,
-	.kind_sys = true,
+	if (!fuse)
+		return -EINVAL;
+
+	return nvkm_fuse_read(fuse, 0x3a8);
+}
+
+static const struct nvkm_volt_func
+gf117_volt = {
+	.oneinit = gf100_volt_oneinit,
+	.vid_get = nvkm_voltgpio_get,
+	.vid_set = nvkm_voltgpio_set,
+	.speedo_read = gf117_volt_speedo_read,
 };
 
 int
-tu104_mmu_new(struct nvkm_device *device, int index, struct nvkm_mmu **pmmu)
+gf117_volt_new(struct nvkm_device *device, int index, struct nvkm_volt **pvolt)
 {
-	return nvkm_mmu_new_(&tu104_mmu, device, index, pmmu);
+	struct nvkm_volt *volt;
+	int ret;
+
+	ret = nvkm_volt_new_(&gf117_volt, device, index, &volt);
+	*pvolt = volt;
+	if (ret)
+		return ret;
+
+	return nvkm_voltgpio_init(volt);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/uvmm.c

···
 }
 
 static int
+nvkm_uvmm_mthd_pfnclr(struct nvkm_uvmm *uvmm, void *argv, u32 argc)
+{
+	struct nvkm_client *client = uvmm->object.client;
+	union {
+		struct nvif_vmm_pfnclr_v0 v0;
+	} *args = argv;
+	struct nvkm_vmm *vmm = uvmm->vmm;
+	int ret = -ENOSYS;
+	u64 addr, size;
+
+	if (!(ret = nvif_unpack(ret, &argv, &argc, args->v0, 0, 0, false))) {
+		addr = args->v0.addr;
+		size = args->v0.size;
+	} else
+		return ret;
+
+	if (!client->super)
+		return -ENOENT;
+
+	if (size) {
+		mutex_lock(&vmm->mutex);
+		ret = nvkm_vmm_pfn_unmap(vmm, addr, size);
+		mutex_unlock(&vmm->mutex);
+	}
+
+	return ret;
+}
+
+static int
+nvkm_uvmm_mthd_pfnmap(struct nvkm_uvmm *uvmm, void *argv, u32 argc)
+{
+	struct nvkm_client *client = uvmm->object.client;
+	union {
+		struct nvif_vmm_pfnmap_v0 v0;
+	} *args = argv;
+	struct nvkm_vmm *vmm = uvmm->vmm;
+	int ret = -ENOSYS;
+	u64 addr, size, *phys;
+	u8  page;
+
+	if (!(ret = nvif_unpack(ret, &argv, &argc, args->v0, 0, 0, true))) {
+		page = args->v0.page;
+		addr = args->v0.addr;
+		size = args->v0.size;
+		phys = args->v0.phys;
+		if (argc != (size >> page) * sizeof(args->v0.phys[0]))
+			return -EINVAL;
+	} else
+		return ret;
+
+	if (!client->super)
+		return -ENOENT;
+
+	if (size) {
+		mutex_lock(&vmm->mutex);
+		ret = nvkm_vmm_pfn_map(vmm, page, addr, size, phys);
+		mutex_unlock(&vmm->mutex);
+	}
+
+	return ret;
+}
+
+static int
 nvkm_uvmm_mthd_unmap(struct nvkm_uvmm *uvmm, void *argv, u32 argc)
 {
 	struct nvkm_client *client = uvmm->object.client;
···
 		goto done;
 	}
 
-	nvkm_vmm_unmap_locked(vmm, vma);
+	nvkm_vmm_unmap_locked(vmm, vma, false);
 	ret = 0;
 done:
 	mutex_unlock(&vmm->mutex);
···
 	if (ret = -ENOENT, (!vma->user && !client->super) || vma->busy) {
 		VMM_DEBUG(vmm, "denied %016llx: %d %d %d", addr,
 			  vma->user, !client->super, vma->busy);
+		goto fail;
+	}
+
+	if (ret = -EINVAL, vma->mapped && !vma->memory) {
+		VMM_DEBUG(vmm, "pfnmap %016llx", addr);
 		goto fail;
 	}
 
···
 	case NVIF_VMM_V0_PUT   : return nvkm_uvmm_mthd_put   (uvmm, argv, argc);
 	case NVIF_VMM_V0_MAP   : return nvkm_uvmm_mthd_map   (uvmm, argv, argc);
 	case NVIF_VMM_V0_UNMAP : return nvkm_uvmm_mthd_unmap (uvmm, argv, argc);
+	case NVIF_VMM_V0_PFNMAP: return nvkm_uvmm_mthd_pfnmap(uvmm, argv, argc);
+	case NVIF_VMM_V0_PFNCLR: return nvkm_uvmm_mthd_pfnclr(uvmm, argv, argc);
+	case NVIF_VMM_V0_MTHD(0x00) ... NVIF_VMM_V0_MTHD(0x7f):
+		if (uvmm->vmm->func->mthd) {
+			return uvmm->vmm->func->mthd(uvmm->vmm,
+						     uvmm->object.client,
+						     mthd, argv, argc);
+		}
+		break;
 	default:
 		break;
 	}
···
 	struct nvkm_uvmm *uvmm;
 	int ret = -ENOSYS;
 	u64 addr, size;
+	bool managed;
 
 	if (!(ret = nvif_unpack(ret, &argv, &argc, args->v0, 0, 0, more))) {
+		managed = args->v0.managed != 0;
 		addr = args->v0.addr;
 		size = args->v0.size;
 	} else
···
 	*pobject = &uvmm->object;
 
 	if (!mmu->vmm) {
-		ret = mmu->func->vmm.ctor(mmu, addr, size, argv, argc,
+		ret = mmu->func->vmm.ctor(mmu, managed, addr, size, argv, argc,
 					  NULL, "user", &uvmm->vmm);
 		if (ret)
 			return ret;

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c

···
 }
 
 static bool
-nvkm_vmm_unref_ptes(struct nvkm_vmm_iter *it, u32 ptei, u32 ptes)
+nvkm_vmm_unref_ptes(struct nvkm_vmm_iter *it, bool pfn, u32 ptei, u32 ptes)
 {
 	const struct nvkm_vmm_desc *desc = it->desc;
 	const int type = desc->type == SPT;
 	struct nvkm_vmm_pt *pgt = it->pt[0];
+	bool dma;
+
+	if (pfn) {
+		/* Need to clear PTE valid bits before we dma_unmap_page(). */
+		dma = desc->func->pfn_clear(it->vmm, pgt->pt[type], ptei, ptes);
+		if (dma) {
+			/* GPU may have cached the PT, flush before unmap. */
+			nvkm_vmm_flush_mark(it);
+			nvkm_vmm_flush(it);
+			desc->func->pfn_unmap(it->vmm, pgt->pt[type], ptei, ptes);
+		}
+	}
 
 	/* Drop PTE references. */
 	pgt->refs[type] -= ptes;
···
 }
 
 static bool
-nvkm_vmm_ref_ptes(struct nvkm_vmm_iter *it, u32 ptei, u32 ptes)
+nvkm_vmm_ref_ptes(struct nvkm_vmm_iter *it, bool pfn, u32 ptei, u32 ptes)
 {
 	const struct nvkm_vmm_desc *desc = it->desc;
 	const int type = desc->type == SPT;
···
 }
 
 static bool
-nvkm_vmm_sparse_unref_ptes(struct nvkm_vmm_iter *it, u32 ptei, u32 ptes)
+nvkm_vmm_sparse_unref_ptes(struct nvkm_vmm_iter *it, bool pfn, u32 ptei, u32 ptes)
 {
 	struct nvkm_vmm_pt *pt = it->pt[0];
 	if (it->desc->type == PGD)
···
 	else
 	if (it->desc->type == LPT)
 		memset(&pt->pte[ptei], 0x00, sizeof(pt->pte[0]) * ptes);
-	return nvkm_vmm_unref_ptes(it, ptei, ptes);
+	return nvkm_vmm_unref_ptes(it, pfn, ptei, ptes);
 }
 
 static bool
-nvkm_vmm_sparse_ref_ptes(struct nvkm_vmm_iter *it, u32 ptei, u32 ptes)
+nvkm_vmm_sparse_ref_ptes(struct nvkm_vmm_iter *it, bool pfn, u32 ptei, u32 ptes)
 {
 	nvkm_vmm_sparse_ptes(it->desc, it->pt[0], ptei, ptes);
-	return nvkm_vmm_ref_ptes(it, ptei, ptes);
+	return nvkm_vmm_ref_ptes(it, pfn, ptei, ptes);
 }
 
 static bool
···
 
 static inline u64
 nvkm_vmm_iter(struct nvkm_vmm *vmm, const struct nvkm_vmm_page *page,
-	      u64 addr, u64 size, const char *name, bool ref,
-	      bool (*REF_PTES)(struct nvkm_vmm_iter *, u32, u32),
+	      u64 addr, u64 size, const char *name, bool ref, bool pfn,
+	      bool (*REF_PTES)(struct nvkm_vmm_iter *, bool pfn, u32, u32),
 	      nvkm_vmm_pte_func MAP_PTES, struct nvkm_vmm_map *map,
 	      nvkm_vmm_pxe_func CLR_PTES)
 {
···
 		}
 
 		/* Handle PTE updates. */
-		if (!REF_PTES || REF_PTES(&it, ptei, ptes)) {
+		if (!REF_PTES || REF_PTES(&it, pfn, ptei, ptes)) {
 			struct nvkm_mmu_pt *pt = pgt->pt[type];
 			if (MAP_PTES || CLR_PTES) {
 				if (MAP_PTES)
···
 nvkm_vmm_ptes_sparse_put(struct nvkm_vmm *vmm, const struct nvkm_vmm_page *page,
 			 u64 addr, u64 size)
 {
-	nvkm_vmm_iter(vmm, page, addr, size, "sparse unref", false,
+	nvkm_vmm_iter(vmm, page, addr, size, "sparse unref", false, false,
 		      nvkm_vmm_sparse_unref_ptes, NULL, NULL,
 		      page->desc->func->invalid ?
 		      page->desc->func->invalid : page->desc->func->unmap);
···
 {
 	if ((page->type & NVKM_VMM_PAGE_SPARSE)) {
 		u64 fail = nvkm_vmm_iter(vmm, page, addr, size, "sparse ref",
-					 true, nvkm_vmm_sparse_ref_ptes, NULL,
-					 NULL, page->desc->func->sparse);
+					 true, false, nvkm_vmm_sparse_ref_ptes,
+					 NULL, NULL, page->desc->func->sparse);
 		if (fail != ~0ULL) {
 			if ((size = fail - addr))
 				nvkm_vmm_ptes_sparse_put(vmm, page, addr, size);
···
 
 static void
 nvkm_vmm_ptes_unmap_put(struct nvkm_vmm *vmm, const struct nvkm_vmm_page *page,
-			u64 addr, u64 size, bool sparse)
+			u64 addr, u64 size, bool sparse, bool pfn)
 {
 	const struct nvkm_vmm_desc_func *func = page->desc->func;
 	nvkm_vmm_iter(vmm, page, addr, size, "unmap + unref",
-		      false, nvkm_vmm_unref_ptes, NULL, NULL,
+		      false, pfn, nvkm_vmm_unref_ptes, NULL, NULL,
 		      sparse ? func->sparse : func->invalid ? func->invalid :
 							      func->unmap);
 }
···
 		      nvkm_vmm_pte_func func)
 {
 	u64 fail = nvkm_vmm_iter(vmm, page, addr, size, "ref + map", true,
-				 nvkm_vmm_ref_ptes, func, map, NULL);
+				 false, nvkm_vmm_ref_ptes, func, map, NULL);
 	if (fail != ~0ULL) {
 		if ((size = fail - addr))
-			nvkm_vmm_ptes_unmap_put(vmm, page, addr, size, false);
+			nvkm_vmm_ptes_unmap_put(vmm, page, addr, size, false, false);
 		return -ENOMEM;
 	}
 	return 0;
···
 
 static void
 nvkm_vmm_ptes_unmap(struct nvkm_vmm *vmm, const struct nvkm_vmm_page *page,
-		    u64 addr, u64 size, bool sparse)
+		    u64 addr, u64 size, bool sparse, bool pfn)
 {
 	const struct nvkm_vmm_desc_func *func = page->desc->func;
-	nvkm_vmm_iter(vmm, page, addr, size, "unmap", false, NULL, NULL, NULL,
+	nvkm_vmm_iter(vmm, page, addr, size, "unmap", false, pfn,
+		      NULL, NULL, NULL,
 		      sparse ? func->sparse : func->invalid ? func->invalid :
 							      func->unmap);
 }
···
 		  u64 addr, u64 size, struct nvkm_vmm_map *map,
 		  nvkm_vmm_pte_func func)
 {
-	nvkm_vmm_iter(vmm, page, addr, size, "map", false,
+	nvkm_vmm_iter(vmm, page, addr, size, "map", false, false,
 		      NULL, func, map, NULL);
 }
 
···
 nvkm_vmm_ptes_put(struct nvkm_vmm *vmm, const struct nvkm_vmm_page *page,
 		  u64 addr, u64 size)
 {
-	nvkm_vmm_iter(vmm, page, addr, size, "unref", false,
+	nvkm_vmm_iter(vmm, page, addr, size, "unref", false, false,
 		      nvkm_vmm_unref_ptes, NULL, NULL, NULL);
 }
 
···
 nvkm_vmm_ptes_get(struct nvkm_vmm *vmm, const struct nvkm_vmm_page *page,
 		  u64 addr, u64 size)
 {
-	u64 fail = nvkm_vmm_iter(vmm, page, addr, size, "ref", true,
+	u64 fail = nvkm_vmm_iter(vmm, page, addr, size, "ref", true, false,
 				 nvkm_vmm_ref_ptes, NULL, NULL, NULL);
 	if (fail != ~0ULL) {
 		if (fail != addr)
···
 	new->part = vma->part;
 	new->user = vma->user;
 	new->busy = vma->busy;
+	new->mapped = vma->mapped;
 	list_add(&new->head, &vma->head);
 	return new;
 }
···
 }
 
 static void
+nvkm_vma_dump(struct nvkm_vma *vma)
+{
+	printk(KERN_ERR "%016llx %016llx %c%c%c%c%c%c%c%c%c %p\n",
+	       vma->addr, (u64)vma->size,
+	       vma->used ? '-' : 'F',
+	       vma->mapref ? 'R' : '-',
+	       vma->sparse ? 'S' : '-',
+	       vma->page != NVKM_VMA_PAGE_NONE ? '0' + vma->page : '-',
+	       vma->refd != NVKM_VMA_PAGE_NONE ? '0' + vma->refd : '-',
+	       vma->part ? 'P' : '-',
+	       vma->user ? 'U' : '-',
+	       vma->busy ? 'B' : '-',
+	       vma->mapped ? 'M' : '-',
+	       vma->memory);
+}
+
+static void
+nvkm_vmm_dump(struct nvkm_vmm *vmm)
+{
+	struct nvkm_vma *vma;
+	list_for_each_entry(vma, &vmm->list, head) {
+		nvkm_vma_dump(vma);
+	}
+}
+
+static void
 nvkm_vmm_dtor(struct nvkm_vmm *vmm)
 {
 	struct nvkm_vma *vma;
 	struct rb_node *node;
+
+	if (0)
+		nvkm_vmm_dump(vmm);
 
 	while ((node = rb_first(&vmm->root))) {
 		struct nvkm_vma *vma = rb_entry(node, typeof(*vma), tree);
···
 	}
 }
 
+static int
+nvkm_vmm_ctor_managed(struct nvkm_vmm *vmm, u64 addr, u64 size)
+{
+	struct nvkm_vma *vma;
+	if (!(vma = nvkm_vma_new(addr, size)))
+		return -ENOMEM;
+	vma->mapref = true;
+	vma->sparse = false;
+	vma->used = true;
+	vma->user = true;
+	nvkm_vmm_node_insert(vmm, vma);
+	list_add_tail(&vma->head, &vmm->list);
+	return 0;
+}
+
 int
 nvkm_vmm_ctor(const struct nvkm_vmm_func *func, struct nvkm_mmu *mmu,
-	      u32 pd_header, u64 addr, u64 size, struct lock_class_key *key,
-	      const char *name, struct nvkm_vmm *vmm)
+	      u32 pd_header, bool managed, u64 addr, u64 size,
+	      struct lock_class_key *key, const char *name,
+	      struct nvkm_vmm *vmm)
 {
 	static struct lock_class_key _key;
 	const struct nvkm_vmm_page *page = func->page;
 	const struct nvkm_vmm_desc *desc;
 	struct nvkm_vma *vma;
-	int levels, bits = 0;
+	int levels, bits = 0, ret;
 
 	vmm->func = func;
 	vmm->mmu = mmu;
···
 	if (WARN_ON(levels > NVKM_VMM_LEVELS_MAX))
 		return -EINVAL;
 
-	vmm->start = addr;
-	vmm->limit = size ? (addr + size) : (1ULL << bits);
-	if (vmm->start > vmm->limit || vmm->limit > (1ULL << bits))
-		return -EINVAL;
-
 	/* Allocate top-level page table. */
 	vmm->pd = nvkm_vmm_pt_new(desc, false, NULL);
 	if (!vmm->pd)
···
 	vmm->free = RB_ROOT;
 	vmm->root = RB_ROOT;
 
-	if (!(vma = nvkm_vma_new(vmm->start, vmm->limit - vmm->start)))
-		return -ENOMEM;
+	if (managed) {
+		/* Address-space will be managed by the client for the most
+		 * part, except for a specified area where NVKM allocations
+		 * are allowed to be placed.
+		 */
+		vmm->start = 0;
+		vmm->limit = 1ULL << bits;
+		if (addr + size < addr || addr + size > vmm->limit)
+			return -EINVAL;
 
-	nvkm_vmm_free_insert(vmm, vma);
-	list_add(&vma->head, &vmm->list);
+		/* Client-managed area before the NVKM-managed area. */
+		if (addr && (ret = nvkm_vmm_ctor_managed(vmm, 0, addr)))
+			return ret;
+
+		/* NVKM-managed area. */
+		if (size) {
+			if (!(vma = nvkm_vma_new(addr, size)))
+				return -ENOMEM;
+			nvkm_vmm_free_insert(vmm, vma);
+			list_add_tail(&vma->head, &vmm->list);
+		}
+
+		/* Client-managed area after the NVKM-managed area. */
+		addr = addr + size;
+		size = vmm->limit - addr;
+		if (size && (ret = nvkm_vmm_ctor_managed(vmm, addr, size)))
+			return ret;
+	} else {
+		/* Address-space fully managed by NVKM, requiring calls to
+		 * nvkm_vmm_get()/nvkm_vmm_put() to allocate address-space.
+		 */
+		vmm->start = addr;
+		vmm->limit = size ? (addr + size) : (1ULL << bits);
+		if (vmm->start > vmm->limit || vmm->limit > (1ULL << bits))
+			return -EINVAL;
+
+		if (!(vma = nvkm_vma_new(vmm->start, vmm->limit - vmm->start)))
+			return -ENOMEM;
+
+		nvkm_vmm_free_insert(vmm, vma);
+		list_add(&vma->head, &vmm->list);
+	}
+
 	return 0;
 }
 
 int
 nvkm_vmm_new_(const struct nvkm_vmm_func *func, struct nvkm_mmu *mmu,
-	      u32 hdr, u64 addr, u64 size, struct lock_class_key *key,
-	      const char *name, struct nvkm_vmm **pvmm)
+	      u32 hdr, bool managed, u64 addr, u64 size,
+	      struct lock_class_key *key, const char *name,
+	      struct nvkm_vmm **pvmm)
 {
 	if (!(*pvmm = kzalloc(sizeof(**pvmm), GFP_KERNEL)))
 		return -ENOMEM;
-	return nvkm_vmm_ctor(func, mmu, hdr, addr, size, key, name, *pvmm);
+	return nvkm_vmm_ctor(func, mmu, hdr, managed, addr, size, key, name, *pvmm);
+}
+
+static struct nvkm_vma *
+nvkm_vmm_pfn_split_merge(struct nvkm_vmm *vmm, struct nvkm_vma *vma,
+			 u64 addr, u64 size, u8 page, bool map)
+{
+	struct nvkm_vma *prev = NULL;
+	struct nvkm_vma *next = NULL;
+
+	if (vma->addr == addr && vma->part && (prev = node(vma, prev))) {
+		if (prev->memory || prev->mapped != map)
+			prev = NULL;
+	}
+
+	if (vma->addr + vma->size == addr + size && (next = node(vma, next))) {
+		if (!next->part ||
+		    next->memory || next->mapped != map)
+			next = NULL;
+	}
+
+	if (prev || next)
+		return nvkm_vmm_node_merge(vmm, prev, vma, next, size);
+	return nvkm_vmm_node_split(vmm, vma, addr, size);
+}
+
+int
+nvkm_vmm_pfn_unmap(struct nvkm_vmm *vmm, u64 addr, u64 size)
+{
+	struct nvkm_vma *vma = nvkm_vmm_node_search(vmm, addr);
+	struct nvkm_vma *next;
+	u64 limit = addr + size;
+	u64 start = addr;
+
+	if (!vma)
+		return -EINVAL;
+
+	do {
+		if (!vma->mapped || vma->memory)
+			continue;
+
+		size = min(limit - start, vma->size - (start - vma->addr));
+
+		nvkm_vmm_ptes_unmap_put(vmm, &vmm->func->page[vma->refd],
+					start, size, false, true);
+
+		next = nvkm_vmm_pfn_split_merge(vmm, vma, start, size, 0, false);
+		if (!WARN_ON(!next)) {
+			vma = next;
+			vma->refd = NVKM_VMA_PAGE_NONE;
+			vma->mapped = false;
+		}
+	} while ((vma = node(vma, next)) && (start = vma->addr) < limit);
+
+	return 0;
+}
+
+/*TODO:
+ * - Avoid PT readback (for dma_unmap etc), this might end up being dealt
+ *   with inside HMM, which would be a lot nicer for us to deal with.
+ * - Multiple page sizes (particularly for huge page support).
+ * - Support for systems without a 4KiB page size.
+ */
+int
+nvkm_vmm_pfn_map(struct nvkm_vmm *vmm, u8 shift, u64 addr, u64 size, u64 *pfn)
+{
+	const struct nvkm_vmm_page *page = vmm->func->page;
+	struct nvkm_vma *vma, *tmp;
+	u64 limit = addr + size;
+	u64 start = addr;
+	int pm = size >> shift;
+	int pi = 0;
+
+	/* Only support mapping where the page size of the incoming page
+	 * array matches a page size available for direct mapping.
+	 */
+	while (page->shift && page->shift != shift &&
+	       page->desc->func->pfn == NULL)
+		page++;
+
+	if (!page->shift || !IS_ALIGNED(addr, 1ULL << shift) ||
+			    !IS_ALIGNED(size, 1ULL << shift) ||
+	    addr + size < addr || addr + size > vmm->limit) {
+		VMM_DEBUG(vmm, "paged map %d %d %016llx %016llx\n",
+			  shift, page->shift, addr, size);
+		return -EINVAL;
+	}
+
+	if (!(vma = nvkm_vmm_node_search(vmm, addr)))
+		return -ENOENT;
+
+	do {
+		bool map = !!(pfn[pi] & NVKM_VMM_PFN_V);
+		bool mapped = vma->mapped;
+		u64 size = limit - start;
+		u64 addr = start;
+		int pn, ret = 0;
+
+		/* Narrow the operation window to cover a single action (page
+		 * should be mapped or not) within a single VMA.
+		 */
+		for (pn = 0; pi + pn < pm; pn++) {
+			if (map != !!(pfn[pi + pn] & NVKM_VMM_PFN_V))
+				break;
+		}
+		size = min_t(u64, size, pn << page->shift);
+		size = min_t(u64, size, vma->size + vma->addr - addr);
+
+		/* Reject any operation to unmanaged regions, and areas that
+		 * have nvkm_memory objects mapped in them already.
+		 */
+		if (!vma->mapref || vma->memory) {
+			ret = -EINVAL;
+			goto next;
+		}
+
+		/* In order to both properly refcount GPU page tables, and
+		 * prevent "normal" mappings and these direct mappings from
+		 * interfering with each other, we need to track contiguous
+		 * ranges that have been mapped with this interface.
+		 *
+		 * Here we attempt to either split an existing VMA so we're
+		 * able to flag the region as either unmapped/mapped, or to
+		 * merge with adjacent VMAs that are already compatible.
+		 *
+		 * If the region is already compatible, nothing is required.
+		 */
+		if (map != mapped) {
+			tmp = nvkm_vmm_pfn_split_merge(vmm, vma, addr, size,
+						       page -
+						       vmm->func->page, map);
+			if (WARN_ON(!tmp)) {
+				ret = -ENOMEM;
+				goto next;
+			}
+
+			if ((tmp->mapped = map))
+				tmp->refd = page - vmm->func->page;
+			else
+				tmp->refd = NVKM_VMA_PAGE_NONE;
+			vma = tmp;
+		}
+
+		/* Update HW page tables. */
+		if (map) {
+			struct nvkm_vmm_map args;
+			args.page = page;
+			args.pfn = &pfn[pi];
+
+			if (!mapped) {
+				ret = nvkm_vmm_ptes_get_map(vmm, page, addr,
+							    size, &args, page->
+							    desc->func->pfn);
+			} else {
+				nvkm_vmm_ptes_map(vmm, page, addr, size, &args,
+						  page->desc->func->pfn);
+			}
+		} else {
+			if (mapped) {
+				nvkm_vmm_ptes_unmap_put(vmm, page, addr, size,
+							false, true);
+			}
+		}
+
+next:
+		/* Iterate to next operation. */
+		if (vma->addr + vma->size == addr + size)
+			vma = node(vma, next);
+		start += size;
+
+		if (ret) {
+			/* Failure is signalled by clearing the valid bit on
+			 * any PFN that couldn't be modified as requested.
+			 */
+			while (size) {
+				pfn[pi++] = NVKM_VMM_PFN_NONE;
+				size -= 1 << page->shift;
+			}
+		} else {
+			pi += size >> page->shift;
+		}
+	} while (vma && start < limit);
+
+	return 0;
 }
 
 void
 nvkm_vmm_unmap_region(struct nvkm_vmm *vmm, struct nvkm_vma *vma)
 {
-	struct nvkm_vma *next = node(vma, next);
 	struct nvkm_vma *prev = NULL;
+	struct nvkm_vma *next;
 
 	nvkm_memory_tags_put(vma->memory, vmm->mmu->subdev.device, &vma->tags);
 	nvkm_memory_unref(&vma->memory);
+	vma->mapped = false;
 
-	if (!vma->part || ((prev = node(vma, prev)), prev->memory))
+	if (vma->part && (prev = node(vma, prev)) && prev->mapped)
 		prev = NULL;
-	if (!next->part || next->memory)
+	if ((next = node(vma, next)) && (!next->part || next->mapped))
 		next = NULL;
 	nvkm_vmm_node_merge(vmm, prev, vma, next, vma->size);
 }
 
 void
-nvkm_vmm_unmap_locked(struct nvkm_vmm *vmm, struct nvkm_vma *vma)
+nvkm_vmm_unmap_locked(struct nvkm_vmm *vmm, struct nvkm_vma *vma, bool pfn)
 {
 	const struct nvkm_vmm_page *page = &vmm->func->page[vma->refd];
 
 	if (vma->mapref) {
-		nvkm_vmm_ptes_unmap_put(vmm, page, vma->addr, vma->size, vma->sparse);
+		nvkm_vmm_ptes_unmap_put(vmm, page, vma->addr, vma->size, vma->sparse, pfn);
 		vma->refd = NVKM_VMA_PAGE_NONE;
 	} else {
-		nvkm_vmm_ptes_unmap(vmm, page, vma->addr, vma->size, vma->sparse);
+		nvkm_vmm_ptes_unmap(vmm, page, vma->addr, vma->size, vma->sparse, pfn);
 	}
 
 	nvkm_vmm_unmap_region(vmm, vma);
···
 {
 	if (vma->memory) {
 		mutex_lock(&vmm->mutex);
-		nvkm_vmm_unmap_locked(vmm, vma);
+		nvkm_vmm_unmap_locked(vmm, vma, false);
 		mutex_unlock(&vmm->mutex);
 	}
 }
···
 	nvkm_memory_tags_put(vma->memory, vmm->mmu->subdev.device, &vma->tags);
 	nvkm_memory_unref(&vma->memory);
 	vma->memory = nvkm_memory_ref(map->memory);
+	vma->mapped = true;
 	vma->tags = map->tags;
 	return 0;
 }
···
 
 	if (vma->mapref || !vma->sparse) {
 		do {
-			const bool map = next->memory != NULL;
+			const bool mem = next->memory != NULL;
+			const bool map = next->mapped;
 			const u8  refd = next->refd;
 			const u64 addr = next->addr;
 			u64 size = next->size;
 
 			/* Merge regions that are in the same state. */
 			while ((next = node(next, next)) && next->part &&
-			       (next->memory != NULL) == map &&
+			       (next->mapped == map) &&
+			       (next->memory != NULL) == mem &&
 			       (next->refd == refd))
 				size += next->size;
 
···
 				 * the page tree.
 				 */
 				nvkm_vmm_ptes_unmap_put(vmm, &page[refd], addr,
-							size, vma->sparse);
+							size, vma->sparse,
+							!mem);
 			} else
 			if (refd != NVKM_VMA_PAGE_NONE) {
 				/* Drop allocation-time PTE references. */
···
 	 */
 	next = vma;
 	do {
-		if (next->memory)
+		if (next->mapped)
 			nvkm_vmm_unmap_region(vmm, next);
 	} while ((next = node(vma, next)) && next->part);
 
···
 }
 
 static bool
-nvkm_vmm_boot_ptes(struct nvkm_vmm_iter *it, u32 ptei, u32 ptes)
+nvkm_vmm_boot_ptes(struct nvkm_vmm_iter *it, bool pfn, u32 ptei, u32 ptes)
 {
 	const struct nvkm_vmm_desc *desc = it->desc;
 	const int type = desc->type == SPT;
···
 	if (ret)
 		return ret;
 
-	nvkm_vmm_iter(vmm, page, vmm->start, limit, "bootstrap", false,
+	nvkm_vmm_iter(vmm, page, vmm->start, limit, "bootstrap", false, false,
 		      nvkm_vmm_boot_ptes, NULL, NULL, NULL);
 	vmm->bootstrapped = true;
 	return 0;
···
 	struct nvkm_mmu *mmu = device->mmu;
 	struct nvkm_vmm *vmm = NULL;
 	int ret;
-	ret = mmu->func->vmm.ctor(mmu, addr, size, argv, argc, key, name, &vmm);
+	ret = mmu->func->vmm.ctor(mmu, false, addr, size, argv, argc,
+				  key, name, &vmm);
 	if (ret)
 		nvkm_vmm_unref(&vmm);
 	*pvmm = vmm;

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h

···
 	nvkm_vmm_pte_func mem;
 	nvkm_vmm_pte_func dma;
 	nvkm_vmm_pte_func sgl;
+
+	nvkm_vmm_pte_func pfn;
+	bool (*pfn_clear)(struct nvkm_vmm *, struct nvkm_mmu_pt *, u32 ptei, u32 ptes);
+	nvkm_vmm_pxe_func pfn_unmap;
 };
 
 extern const struct nvkm_vmm_desc_func gf100_vmm_pgd;
···
 		     struct nvkm_vmm_map *);
 	void (*flush)(struct nvkm_vmm *, int depth);
 
+	int (*mthd)(struct nvkm_vmm *, struct nvkm_client *,
+		    u32 mthd, void *argv, u32 argc);
+
+	void (*invalidate_pdb)(struct nvkm_vmm *, u64 addr);
+
 	u64 page_block;
 	const struct nvkm_vmm_page page[];
 };
···
 };
 
 int nvkm_vmm_new_(const struct nvkm_vmm_func *, struct nvkm_mmu *,
-		  u32 pd_header, u64 addr, u64 size, struct lock_class_key *,
-		  const char *name, struct nvkm_vmm **);
+		  u32 pd_header, bool managed, u64 addr, u64 size,
+		  struct lock_class_key *, const char *name,
+		  struct nvkm_vmm **);
 int nvkm_vmm_ctor(const struct nvkm_vmm_func *, struct nvkm_mmu *,
-		  u32 pd_header, u64 addr, u64 size, struct lock_class_key *,
-		  const char *name, struct nvkm_vmm *);
+		  u32 pd_header, bool managed, u64 addr, u64 size,
+		  struct lock_class_key *, const char *name, struct nvkm_vmm *);
 struct nvkm_vma *nvkm_vmm_node_search(struct nvkm_vmm *, u64 addr);
 struct nvkm_vma *nvkm_vmm_node_split(struct nvkm_vmm *, struct nvkm_vma *,
 				     u64 addr, u64 size);
···
 			bool sparse, u8 page, u8 align, u64 size,
 			struct nvkm_vma **pvma);
 void nvkm_vmm_put_locked(struct nvkm_vmm *, struct nvkm_vma *);
-void nvkm_vmm_unmap_locked(struct nvkm_vmm *, struct nvkm_vma *);
-void nvkm_vmm_unmap_region(struct nvkm_vmm *vmm, struct nvkm_vma *vma);
+void nvkm_vmm_unmap_locked(struct nvkm_vmm *, struct nvkm_vma *, bool pfn);
+void nvkm_vmm_unmap_region(struct nvkm_vmm *, struct nvkm_vma *);
+
+#define NVKM_VMM_PFN_ADDR                                 0xfffffffffffff000ULL
+#define NVKM_VMM_PFN_ADDR_SHIFT                                              12
+#define NVKM_VMM_PFN_APER                                 0x00000000000000f0ULL
+#define NVKM_VMM_PFN_HOST                                 0x0000000000000000ULL
+#define NVKM_VMM_PFN_VRAM                                 0x0000000000000010ULL
+#define NVKM_VMM_PFN_W                                    0x0000000000000002ULL
+#define NVKM_VMM_PFN_V                                    0x0000000000000001ULL
+#define NVKM_VMM_PFN_NONE                                 0x0000000000000000ULL
+
+int nvkm_vmm_pfn_map(struct nvkm_vmm *, u8 page, u64 addr, u64 size, u64 *pfn);
+int nvkm_vmm_pfn_unmap(struct nvkm_vmm *, u64 addr, u64 size);
 
 struct nvkm_vma *nvkm_vma_tail(struct nvkm_vma *, u64 tail);
 
 int nv04_vmm_new_(const struct nvkm_vmm_func *, struct nvkm_mmu *, u32,
-		  u64, u64, void *, u32, struct lock_class_key *,
+		  bool, u64, u64, void *, u32, struct lock_class_key *,
 		  const char *, struct nvkm_vmm **);
 int nv04_vmm_valid(struct nvkm_vmm *, void *, u32, struct nvkm_vmm_map *);
 
···
 void nv50_vmm_flush(struct nvkm_vmm *, int);
 
 int gf100_vmm_new_(const struct nvkm_vmm_func *, const struct nvkm_vmm_func *,
-		   struct nvkm_mmu *, u64, u64, void *, u32,
+		   struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		   struct lock_class_key *, const char *, struct nvkm_vmm **);
 int gf100_vmm_join_(struct nvkm_vmm *, struct nvkm_memory *, u64 base);
 int gf100_vmm_join(struct nvkm_vmm *, struct nvkm_memory *);
 void gf100_vmm_part(struct nvkm_vmm *, struct nvkm_memory *);
 int gf100_vmm_aper(enum nvkm_memory_target);
 int gf100_vmm_valid(struct nvkm_vmm *, void *, u32, struct nvkm_vmm_map *);
-void gf100_vmm_flush_(struct nvkm_vmm *, int);
 void gf100_vmm_flush(struct nvkm_vmm *, int);
+void gf100_vmm_invalidate(struct nvkm_vmm *, u32 type);
+void gf100_vmm_invalidate_pdb(struct nvkm_vmm *, u64 addr);
 
 int gk20a_vmm_aper(enum nvkm_memory_target);
 
 int gm200_vmm_new_(const struct nvkm_vmm_func *, const struct nvkm_vmm_func *,
-		   struct nvkm_mmu *, u64, u64, void *, u32,
+		   struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		   struct lock_class_key *, const char *, struct nvkm_vmm **);
 int gm200_vmm_join_(struct nvkm_vmm *, struct nvkm_memory *, u64 base);
 int gm200_vmm_join(struct nvkm_vmm *, struct nvkm_memory *);
 
+int gp100_vmm_new_(const struct nvkm_vmm_func *,
+		   struct nvkm_mmu *, bool, u64, u64, void *, u32,
+		   struct lock_class_key *, const char *, struct nvkm_vmm **);
 int gp100_vmm_join(struct nvkm_vmm *, struct nvkm_memory *);
 int gp100_vmm_valid(struct nvkm_vmm *, void *, u32, struct nvkm_vmm_map *);
 void gp100_vmm_flush(struct nvkm_vmm *, int);
+int gp100_vmm_mthd(struct nvkm_vmm *, struct nvkm_client *, u32, void *, u32);
+void gp100_vmm_invalidate_pdb(struct nvkm_vmm *, u64 addr);
 
 int gv100_vmm_join(struct nvkm_vmm *, struct nvkm_memory *);
 
-int nv04_vmm_new(struct nvkm_mmu *, u64, u64, void *, u32,
+int nv04_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		 struct lock_class_key *, const char *, struct nvkm_vmm **);
-int nv41_vmm_new(struct nvkm_mmu *, u64, u64, void *, u32,
+int nv41_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		 struct lock_class_key *, const char *, struct nvkm_vmm **);
-int nv44_vmm_new(struct nvkm_mmu *, u64, u64, void *, u32,
+int nv44_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		 struct lock_class_key *, const char *, struct nvkm_vmm **);
-int nv50_vmm_new(struct nvkm_mmu *, u64, u64, void *, u32,
+int nv50_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		 struct lock_class_key *, const char *, struct nvkm_vmm **);
-int mcp77_vmm_new(struct nvkm_mmu *, u64, u64, void *, u32,
+int mcp77_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		  struct lock_class_key *, const char *, struct nvkm_vmm **);
-int g84_vmm_new(struct nvkm_mmu *, u64, u64, void *, u32,
+int g84_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		struct lock_class_key *, const char *, struct nvkm_vmm **);
-int gf100_vmm_new(struct nvkm_mmu *, u64, u64, void *, u32,
+int gf100_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		  struct lock_class_key *, const char *, struct nvkm_vmm **);
-int gk104_vmm_new(struct nvkm_mmu *, u64, u64, void *, u32,
+int gk104_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		  struct lock_class_key *, const char *, struct nvkm_vmm **);
-int gk20a_vmm_new(struct nvkm_mmu *, u64, u64, void *, u32,
+int gk20a_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		  struct lock_class_key *, const char *, struct nvkm_vmm **);
-int gm200_vmm_new_fixed(struct nvkm_mmu *, u64, u64, void *, u32,
+int gm200_vmm_new_fixed(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 			struct lock_class_key *, const char *,
 			struct nvkm_vmm **);
-int gm200_vmm_new(struct nvkm_mmu *, u64, u64, void *, u32,
+int gm200_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		  struct lock_class_key *, const char *,
 		  struct nvkm_vmm **);
-int gm20b_vmm_new_fixed(struct nvkm_mmu *, u64, u64, void *, u32,
+int gm20b_vmm_new_fixed(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 			struct lock_class_key *, const char *,
 			struct nvkm_vmm **);
-int gm20b_vmm_new(struct nvkm_mmu *, u64, u64, void *, u32,
+int gm20b_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		  struct lock_class_key *, const char *,
 		  struct nvkm_vmm **);
-int gp100_vmm_new(struct nvkm_mmu *, u64, u64, void *, u32,
+int gp100_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		  struct lock_class_key *, const char *,
 		  struct nvkm_vmm **);
-int gp10b_vmm_new(struct nvkm_mmu *, u64, u64, void *, u32,
+int gp10b_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		  struct lock_class_key *, const char *,
 		  struct nvkm_vmm **);
-int gv100_vmm_new(struct nvkm_mmu *, u64, u64, void *, u32,
+int gv100_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		  struct lock_class_key *, const char *,
 		  struct nvkm_vmm **);
-int tu104_vmm_new(struct nvkm_mmu *, u64, u64, void *, u32,
+int tu102_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		  struct lock_class_key *, const char *,
 		  struct nvkm_vmm **);
 

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgf100.c

···
 };
 
 void
-gf100_vmm_flush_(struct nvkm_vmm *vmm, int depth)
+gf100_vmm_invalidate_pdb(struct nvkm_vmm *vmm, u64 addr)
+{
+	struct nvkm_device *device = vmm->mmu->subdev.device;
+	nvkm_wr32(device, 0x100cb8, addr);
+}
+
+void
+gf100_vmm_invalidate(struct nvkm_vmm *vmm, u32 type)
 {
 	struct nvkm_subdev *subdev = &vmm->mmu->subdev;
 	struct nvkm_device *device = subdev->device;
-	u32 type = depth << 24;
-
-	type = 0x00000001; /* PAGE_ALL */
-	if (atomic_read(&vmm->engref[NVKM_SUBDEV_BAR]))
-		type |= 0x00000004; /* HUB_ONLY */
+	struct nvkm_mmu_pt *pd = vmm->pd->pt[0];
+	u64 addr = 0;
 
 	mutex_lock(&subdev->mutex);
 	/* Looks like maybe a "free flush slots" counter, the
···
 			break;
 	);
 
-	nvkm_wr32(device, 0x100cb8, vmm->pd->pt[0]->addr >> 8);
+	if (!(type & 0x00000002) /* ALL_PDB. */) {
+		switch (nvkm_memory_target(pd->memory)) {
+		case NVKM_MEM_TARGET_VRAM: addr |= 0x00000000; break;
+		case NVKM_MEM_TARGET_HOST: addr |= 0x00000002; break;
+		case NVKM_MEM_TARGET_NCOH: addr |= 0x00000003; break;
+		default:
+			WARN_ON(1);
+			break;
+		}
+		addr |= (vmm->pd->pt[0]->addr >> 12) << 4;
+
+		vmm->func->invalidate_pdb(vmm, addr);
+	}
+
 	nvkm_wr32(device, 0x100cbc, 0x80000000 | type);
 
 	/* Wait for flush to be queued? */
···
 void
 gf100_vmm_flush(struct nvkm_vmm *vmm, int depth)
 {
-	gf100_vmm_flush_(vmm, 0);
+	u32 type = 0x00000001; /* PAGE_ALL */
+	if (atomic_read(&vmm->engref[NVKM_SUBDEV_BAR]))
+		type |= 0x00000004; /* HUB_ONLY */
+	gf100_vmm_invalidate(vmm, type);
 }
 
 int
···
 	.aper = gf100_vmm_aper,
 	.valid = gf100_vmm_valid,
 	.flush = gf100_vmm_flush,
+	.invalidate_pdb = gf100_vmm_invalidate_pdb,
 	.page = {
 		{ 17, &gf100_vmm_desc_17_17[0], NVKM_VMM_PAGE_xVxC },
 		{ 12, &gf100_vmm_desc_17_12[0], NVKM_VMM_PAGE_xVHx },
···
 	.aper = gf100_vmm_aper,
 	.valid = gf100_vmm_valid,
 	.flush = gf100_vmm_flush,
+	.invalidate_pdb = gf100_vmm_invalidate_pdb,
 	.page = {
 		{ 16, &gf100_vmm_desc_16_16[0], NVKM_VMM_PAGE_xVxC },
 		{ 12, &gf100_vmm_desc_16_12[0], NVKM_VMM_PAGE_xVHx },
···
 int
 gf100_vmm_new_(const struct nvkm_vmm_func *func_16,
 	       const struct nvkm_vmm_func *func_17,
-	       struct nvkm_mmu *mmu, u64 addr, u64 size, void *argv, u32 argc,
-	       struct lock_class_key *key, const char *name,
-	       struct nvkm_vmm **pvmm)
+	       struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	       void *argv, u32 argc, struct lock_class_key *key,
+	       const char *name, struct nvkm_vmm **pvmm)
 {
 	switch (mmu->subdev.device->fb->page) {
-	case 16: return nv04_vmm_new_(func_16, mmu, 0, addr, size,
+	case 16: return nv04_vmm_new_(func_16, mmu, 0, managed, addr, size,
 				      argv, argc, key, name, pvmm);
-	case 17: return nv04_vmm_new_(func_17, mmu, 0, addr, size,
+	case 17: return nv04_vmm_new_(func_17, mmu, 0, managed, addr, size,
 				      argv, argc, key, name, pvmm);
 	default:
 		WARN_ON(1);
···
 }
 
 int
-gf100_vmm_new(struct nvkm_mmu *mmu, u64 addr, u64 size, void *argv, u32 argc,
-	      struct lock_class_key *key, const char *name,
-	      struct nvkm_vmm **pvmm)
+gf100_vmm_new(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	      void *argv, u32 argc, struct lock_class_key *key,
+	      const char *name, struct nvkm_vmm **pvmm)
 {
-	return gf100_vmm_new_(&gf100_vmm_16, &gf100_vmm_17, mmu, addr,
+	return gf100_vmm_new_(&gf100_vmm_16, &gf100_vmm_17, mmu, managed, addr,
 			      size, argv, argc, key, name, pvmm);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgk104.c

···
 	.aper = gf100_vmm_aper,
 	.valid = gf100_vmm_valid,
 	.flush = gf100_vmm_flush,
+	.invalidate_pdb = gf100_vmm_invalidate_pdb,
 	.page = {
 		{ 17, &gk104_vmm_desc_17_17[0], NVKM_VMM_PAGE_xVxC },
 		{ 12, &gk104_vmm_desc_17_12[0], NVKM_VMM_PAGE_xVHx },
···
 	.aper = gf100_vmm_aper,
 	.valid = gf100_vmm_valid,
 	.flush = gf100_vmm_flush,
+	.invalidate_pdb = gf100_vmm_invalidate_pdb,
 	.page = {
 		{ 16, &gk104_vmm_desc_16_16[0], NVKM_VMM_PAGE_xVxC },
 		{ 12, &gk104_vmm_desc_16_12[0], NVKM_VMM_PAGE_xVHx },
···
 };
 
 int
-gk104_vmm_new(struct nvkm_mmu *mmu, u64 addr, u64 size, void *argv, u32 argc,
-	      struct lock_class_key *key, const char *name,
-	      struct nvkm_vmm **pvmm)
+gk104_vmm_new(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	      void *argv, u32 argc, struct lock_class_key *key,
+	      const char *name, struct nvkm_vmm **pvmm)
 {
-	return gf100_vmm_new_(&gk104_vmm_16, &gk104_vmm_17, mmu, addr,
+	return gf100_vmm_new_(&gk104_vmm_16, &gk104_vmm_17, mmu, managed, addr,
 			      size, argv, argc, key, name, pvmm);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgk20a.c

···
 	.aper = gf100_vmm_aper,
 	.valid = gf100_vmm_valid,
 	.flush = gf100_vmm_flush,
+	.invalidate_pdb = gf100_vmm_invalidate_pdb,
 	.page = {
 		{ 17, &gk104_vmm_desc_17_17[0], NVKM_VMM_PAGE_xxHC },
 		{ 12, &gk104_vmm_desc_17_12[0], NVKM_VMM_PAGE_xxHx },
···
 	.aper = gf100_vmm_aper,
 	.valid = gf100_vmm_valid,
 	.flush = gf100_vmm_flush,
+	.invalidate_pdb = gf100_vmm_invalidate_pdb,
 	.page = {
 		{ 16, &gk104_vmm_desc_16_16[0], NVKM_VMM_PAGE_xxHC },
 		{ 12, &gk104_vmm_desc_16_12[0], NVKM_VMM_PAGE_xxHx },
···
 };
 
 int
-gk20a_vmm_new(struct nvkm_mmu *mmu, u64 addr, u64 size, void *argv, u32 argc,
-	      struct lock_class_key *key, const char *name,
-	      struct nvkm_vmm **pvmm)
+gk20a_vmm_new(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	      void *argv, u32 argc, struct lock_class_key *key,
+	      const char *name, struct nvkm_vmm **pvmm)
 {
-	return gf100_vmm_new_(&gk20a_vmm_16, &gk20a_vmm_17, mmu, addr,
+	return gf100_vmm_new_(&gk20a_vmm_16, &gk20a_vmm_17, mmu, managed, addr,
 			      size, argv, argc, key, name, pvmm);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgm200.c

···
 	.aper = gf100_vmm_aper,
 	.valid = gf100_vmm_valid,
 	.flush = gf100_vmm_flush,
+	.invalidate_pdb = gf100_vmm_invalidate_pdb,
 	.page = {
 		{ 27, &gm200_vmm_desc_17_17[1], NVKM_VMM_PAGE_Sxxx },
 		{ 17, &gm200_vmm_desc_17_17[0], NVKM_VMM_PAGE_SVxC },
···
 	.aper = gf100_vmm_aper,
 	.valid = gf100_vmm_valid,
 	.flush = gf100_vmm_flush,
+	.invalidate_pdb = gf100_vmm_invalidate_pdb,
 	.page = {
 		{ 27, &gm200_vmm_desc_16_16[1], NVKM_VMM_PAGE_Sxxx },
 		{ 16, &gm200_vmm_desc_16_16[0], NVKM_VMM_PAGE_SVxC },
···
 int
 gm200_vmm_new_(const struct nvkm_vmm_func *func_16,
 	       const struct nvkm_vmm_func *func_17,
-	       struct nvkm_mmu *mmu, u64 addr, u64 size, void *argv, u32 argc,
-	       struct lock_class_key *key, const char *name,
-	       struct nvkm_vmm **pvmm)
+	       struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	       void *argv, u32 argc, struct lock_class_key *key,
+	       const char *name, struct nvkm_vmm **pvmm)
 {
 	const struct nvkm_vmm_func *func;
 	union {
···
 	} else
 		return ret;
 
-	return nvkm_vmm_new_(func, mmu, 0, addr, size, key, name, pvmm);
+	return nvkm_vmm_new_(func, mmu, 0, managed, addr, size, key, name, pvmm);
 }
 
 int
-gm200_vmm_new(struct nvkm_mmu *mmu, u64 addr, u64 size, void *argv, u32 argc,
-	      struct lock_class_key *key, const char *name,
-	      struct nvkm_vmm **pvmm)
+gm200_vmm_new(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	      void *argv, u32 argc, struct lock_class_key *key,
+	      const char *name, struct nvkm_vmm **pvmm)
 {
-	return gm200_vmm_new_(&gm200_vmm_16, &gm200_vmm_17, mmu, addr,
+	return gm200_vmm_new_(&gm200_vmm_16, &gm200_vmm_17, mmu, managed, addr,
 			      size, argv, argc, key, name, pvmm);
 }
 
 int
-gm200_vmm_new_fixed(struct nvkm_mmu *mmu, u64 addr, u64 size,
+gm200_vmm_new_fixed(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
 		    void *argv, u32 argc, struct lock_class_key *key,
 		    const char *name, struct nvkm_vmm **pvmm)
 {
-	return gf100_vmm_new_(&gm200_vmm_16, &gm200_vmm_17, mmu, addr,
+	return gf100_vmm_new_(&gm200_vmm_16, &gm200_vmm_17, mmu, managed, addr,
 			      size, argv, argc, key, name, pvmm);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgm20b.c

···
 	.aper = gk20a_vmm_aper,
 	.valid = gf100_vmm_valid,
 	.flush = gf100_vmm_flush,
+	.invalidate_pdb = gf100_vmm_invalidate_pdb,
 	.page = {
 		{ 27, &gm200_vmm_desc_17_17[1], NVKM_VMM_PAGE_Sxxx },
 		{ 17, &gm200_vmm_desc_17_17[0], NVKM_VMM_PAGE_SxHC },
···
 	.aper = gk20a_vmm_aper,
 	.valid = gf100_vmm_valid,
 	.flush = gf100_vmm_flush,
+	.invalidate_pdb = gf100_vmm_invalidate_pdb,
 	.page = {
 		{ 27, &gm200_vmm_desc_16_16[1], NVKM_VMM_PAGE_Sxxx },
 		{ 16, &gm200_vmm_desc_16_16[0], NVKM_VMM_PAGE_SxHC },
···
 };
 
 int
-gm20b_vmm_new(struct nvkm_mmu *mmu, u64 addr, u64 size, void *argv, u32 argc,
-	      struct lock_class_key *key, const char *name,
-	      struct nvkm_vmm **pvmm)
+gm20b_vmm_new(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	      void *argv, u32 argc, struct lock_class_key *key,
+	      const char *name, struct nvkm_vmm **pvmm)
 {
-	return gm200_vmm_new_(&gm20b_vmm_16, &gm20b_vmm_17, mmu, addr,
+	return gm200_vmm_new_(&gm20b_vmm_16, &gm20b_vmm_17, mmu, managed, addr,
 			      size, argv, argc, key, name, pvmm);
 }
 
 int
-gm20b_vmm_new_fixed(struct nvkm_mmu *mmu, u64 addr, u64 size,
+gm20b_vmm_new_fixed(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
 		    void *argv, u32 argc, struct lock_class_key *key,
 		    const char *name, struct nvkm_vmm **pvmm)
 {
-	return gf100_vmm_new_(&gm20b_vmm_16, &gm20b_vmm_17, mmu, addr,
+	return gf100_vmm_new_(&gm20b_vmm_16, &gm20b_vmm_17, mmu, managed, addr,
 			      size, argv, argc, key, name, pvmm);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgp100.c

···
  */
 #include "vmm.h"
 
+#include <core/client.h>
 #include <subdev/fb.h>
 #include <subdev/ltc.h>
+#include <subdev/timer.h>
+#include <engine/gr.h>
 
 #include <nvif/ifc00d.h>
 #include <nvif/unpack.h>
+
+static void
+gp100_vmm_pfn_unmap(struct nvkm_vmm *vmm,
+		    struct nvkm_mmu_pt *pt, u32 ptei, u32 ptes)
+{
+	struct device *dev = vmm->mmu->subdev.device->dev;
+	dma_addr_t addr;
+
+	nvkm_kmap(pt->memory);
+	while (ptes--) {
+		u32 datalo = nvkm_ro32(pt->memory, pt->base + ptei * 8 + 0);
+		u32 datahi = nvkm_ro32(pt->memory, pt->base + ptei * 8 + 4);
+		u64 data   = (u64)datahi << 32 | datalo;
+		if ((data & (3ULL << 1)) != 0) {
+			addr = (data >> 8) << 12;
+			dma_unmap_page(dev, addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
+		}
+		ptei++;
+	}
+	nvkm_done(pt->memory);
+}
+
+static bool
+gp100_vmm_pfn_clear(struct nvkm_vmm *vmm,
+		    struct nvkm_mmu_pt *pt, u32 ptei, u32 ptes)
+{
+	bool dma = false;
+	nvkm_kmap(pt->memory);
+	while (ptes--) {
+		u32 datalo = nvkm_ro32(pt->memory, pt->base + ptei * 8 + 0);
+		u32 datahi = nvkm_ro32(pt->memory, pt->base + ptei * 8 + 4);
+		u64 data   = (u64)datahi << 32 | datalo;
+		if ((data & BIT_ULL(0)) && (data & (3ULL << 1)) != 0) {
+			VMM_WO064(pt, vmm, ptei * 8, data & ~BIT_ULL(0));
+			dma = true;
+		}
+		ptei++;
+	}
+	nvkm_done(pt->memory);
+	return dma;
+}
+
+static void
+gp100_vmm_pgt_pfn(struct nvkm_vmm *vmm, struct nvkm_mmu_pt *pt,
+		  u32 ptei, u32 ptes, struct nvkm_vmm_map *map)
+{
+	struct device *dev = vmm->mmu->subdev.device->dev;
+	dma_addr_t addr;
+
+	nvkm_kmap(pt->memory);
+	while (ptes--) {
+		u64 data = 0;
+		if (!(*map->pfn & NVKM_VMM_PFN_W))
+			data |= BIT_ULL(6); /* RO. */
+
+		if (!(*map->pfn & NVKM_VMM_PFN_VRAM)) {
+			addr = *map->pfn >> NVKM_VMM_PFN_ADDR_SHIFT;
+			addr = dma_map_page(dev, pfn_to_page(addr), 0,
+					    PAGE_SIZE, DMA_BIDIRECTIONAL);
+			if (!WARN_ON(dma_mapping_error(dev, addr))) {
+				data |= addr >> 4;
+				data |= 2ULL << 1; /* SYSTEM_COHERENT_MEMORY. */
+				data |= BIT_ULL(3); /* VOL. */
+				data |= BIT_ULL(0); /* VALID. */
+			}
+		} else {
+			data |= (*map->pfn & NVKM_VMM_PFN_ADDR) >> 4;
+			data |= BIT_ULL(0); /* VALID. */
+		}
+
+		VMM_WO064(pt, vmm, ptei++ * 8, data);
+		map->pfn++;
+	}
+	nvkm_done(pt->memory);
+}
 
 static inline void
 gp100_vmm_pgt_pte(struct nvkm_vmm *vmm, struct nvkm_mmu_pt *pt,
···
 	.mem = gp100_vmm_pgt_mem,
 	.dma = gp100_vmm_pgt_dma,
 	.sgl = gp100_vmm_pgt_sgl,
+	.pfn = gp100_vmm_pgt_pfn,
+	.pfn_clear = gp100_vmm_pfn_clear,
+	.pfn_unmap = gp100_vmm_pfn_unmap,
 };
 
 static void
···
 	return 0;
 }
 
+static int
+gp100_vmm_fault_cancel(struct nvkm_vmm *vmm, void *argv, u32 argc)
+{
+	struct nvkm_device *device = vmm->mmu->subdev.device;
+	union {
+		struct gp100_vmm_fault_cancel_v0 v0;
+	} *args = argv;
+	int ret = -ENOSYS;
+	u32 inst, aper;
+
+	if ((ret = nvif_unpack(ret, &argv, &argc, args->v0, 0, 0, false)))
+		return ret;
+
+	/* Translate MaxwellFaultBufferA instance pointer to the same
+	 * format as the NV_GR_FECS_CURRENT_CTX register.
+	 */
+	aper = (args->v0.inst >> 8) & 3;
+	args->v0.inst >>= 12;
+	args->v0.inst |= aper << 28;
+	args->v0.inst |= 0x80000000;
+
+	if (!WARN_ON(nvkm_gr_ctxsw_pause(device))) {
+		if ((inst = nvkm_gr_ctxsw_inst(device)) == args->v0.inst) {
+			gf100_vmm_invalidate(vmm, 0x0000001b
+					     /* CANCEL_TARGETED. */ |
+					     (args->v0.hub    << 20) |
+					     (args->v0.gpc    << 15) |
+					     (args->v0.client << 9));
+		}
+		WARN_ON(nvkm_gr_ctxsw_resume(device));
+	}
+
+	return 0;
+}
+
+static int
+gp100_vmm_fault_replay(struct nvkm_vmm *vmm, void *argv, u32 argc)
+{
+	union {
+		struct gp100_vmm_fault_replay_vn vn;
+	} *args = argv;
+	int ret = -ENOSYS;
+
+	if (!(ret = nvif_unvers(ret, &argv, &argc, args->vn))) {
+		gf100_vmm_invalidate(vmm, 0x0000000b); /* REPLAY_GLOBAL. */
+	}
+
+	return ret;
+}
+
+int
+gp100_vmm_mthd(struct nvkm_vmm *vmm,
+	       struct nvkm_client *client, u32 mthd, void *argv, u32 argc)
+{
+	if (client->super) {
+		switch (mthd) {
+		case GP100_VMM_VN_FAULT_REPLAY:
+			return gp100_vmm_fault_replay(vmm, argv, argc);
+		case GP100_VMM_VN_FAULT_CANCEL:
+			return gp100_vmm_fault_cancel(vmm, argv, argc);
+		default:
+			break;
+		}
+	}
+	return -EINVAL;
+}
+
+void
+gp100_vmm_invalidate_pdb(struct nvkm_vmm *vmm, u64 addr)
+{
+	struct nvkm_device *device = vmm->mmu->subdev.device;
+	nvkm_wr32(device, 0x100cb8, lower_32_bits(addr));
+	nvkm_wr32(device, 0x100cec, upper_32_bits(addr));
+}
+
 void
 gp100_vmm_flush(struct nvkm_vmm *vmm, int depth)
 {
-	gf100_vmm_flush_(vmm, 5 /* CACHE_LEVEL_UP_TO_PDE3 */ - depth);
+	u32 type = (5 /* CACHE_LEVEL_UP_TO_PDE3 */ - depth) << 24;
+	type = 0; /*XXX: need to confirm stuff works with depth enabled... */
+	if (atomic_read(&vmm->engref[NVKM_SUBDEV_BAR]))
+		type |= 0x00000004; /* HUB_ONLY */
+	type |= 0x00000001; /* PAGE_ALL */
+	gf100_vmm_invalidate(vmm, type);
 }
 
 int
 gp100_vmm_join(struct nvkm_vmm *vmm, struct nvkm_memory *inst)
 {
-	const u64 base = BIT_ULL(10) /* VER2 */ | BIT_ULL(11); /* 64KiB */
+	u64 base = BIT_ULL(10) /* VER2 */ | BIT_ULL(11) /* 64KiB */;
+	if (vmm->replay) {
+		base |= BIT_ULL(4); /* FAULT_REPLAY_TEX */
+		base |= BIT_ULL(5); /* FAULT_REPLAY_GCC */
+	}
 	return gf100_vmm_join_(vmm, inst, base);
 }
 
···
 	.aper = gf100_vmm_aper,
 	.valid = gp100_vmm_valid,
 	.flush = gp100_vmm_flush,
+	.mthd = gp100_vmm_mthd,
+	.invalidate_pdb = gp100_vmm_invalidate_pdb,
 	.page = {
 		{ 47, &gp100_vmm_desc_16[4], NVKM_VMM_PAGE_Sxxx },
 		{ 38, &gp100_vmm_desc_16[3], NVKM_VMM_PAGE_Sxxx },
···
 };
 
 int
-gp100_vmm_new(struct nvkm_mmu *mmu, u64 addr, u64 size, void *argv, u32 argc,
-	      struct lock_class_key *key, const char *name,
-	      struct nvkm_vmm **pvmm)
+gp100_vmm_new_(const struct nvkm_vmm_func *func,
+	       struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	       void *argv, u32 argc, struct lock_class_key *key,
+	       const char *name, struct nvkm_vmm **pvmm)
 {
-	return nv04_vmm_new_(&gp100_vmm, mmu, 0, addr, size,
-			     argv, argc, key, name, pvmm);
+	union {
+		struct gp100_vmm_vn vn;
+		struct gp100_vmm_v0 v0;
+	} *args = argv;
+	int ret = -ENOSYS;
+	bool replay;
+
+	if (!(ret = nvif_unpack(ret, &argv, &argc, args->v0, 0, 0, false))) {
+		replay = args->v0.fault_replay != 0;
+	} else
+	if (!(ret = nvif_unvers(ret, &argv, &argc, args->vn))) {
+		replay = false;
+	} else
+		return ret;
+
+	ret = nvkm_vmm_new_(func, mmu, 0, managed, addr, size, key, name, pvmm);
+	if (ret)
+		return ret;
+
+	(*pvmm)->replay = replay;
+	return 0;
+}
+
+int
+gp100_vmm_new(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	      void *argv, u32 argc, struct lock_class_key *key,
+	      const char *name, struct nvkm_vmm **pvmm)
+{
+	return gp100_vmm_new_(&gp100_vmm, mmu, managed, addr, size,
+			      argv, argc, key, name, pvmm);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgp10b.c

···
 	.aper = gk20a_vmm_aper,
 	.valid = gp100_vmm_valid,
 	.flush = gp100_vmm_flush,
+	.mthd = gp100_vmm_mthd,
+	.invalidate_pdb = gp100_vmm_invalidate_pdb,
 	.page = {
 		{ 47, &gp100_vmm_desc_16[4], NVKM_VMM_PAGE_Sxxx },
 		{ 38, &gp100_vmm_desc_16[3], NVKM_VMM_PAGE_Sxxx },
···
 };
 
 int
-gp10b_vmm_new(struct nvkm_mmu *mmu, u64 addr, u64 size, void *argv, u32 argc,
-	      struct lock_class_key *key, const char *name,
-	      struct nvkm_vmm **pvmm)
+gp10b_vmm_new(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	      void *argv, u32 argc, struct lock_class_key *key,
+	      const char *name, struct nvkm_vmm **pvmm)
 {
-	return nv04_vmm_new_(&gp10b_vmm, mmu, 0, addr, size,
-			     argv, argc, key, name, pvmm);
+	return gp100_vmm_new_(&gp10b_vmm, mmu, managed, addr, size,
+			      argv, argc, key, name, pvmm);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgv100.c

···
 	.aper = gf100_vmm_aper,
 	.valid = gp100_vmm_valid,
 	.flush = gp100_vmm_flush,
+	.mthd = gp100_vmm_mthd,
+	.invalidate_pdb = gp100_vmm_invalidate_pdb,
 	.page = {
 		{ 47, &gp100_vmm_desc_16[4], NVKM_VMM_PAGE_Sxxx },
 		{ 38, &gp100_vmm_desc_16[3], NVKM_VMM_PAGE_Sxxx },
···
 };
 
 int
-gv100_vmm_new(struct nvkm_mmu *mmu, u64 addr, u64 size, void *argv, u32 argc,
-	      struct lock_class_key *key, const char *name,
-	      struct nvkm_vmm **pvmm)
+gv100_vmm_new(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	      void *argv, u32 argc, struct lock_class_key *key,
+	      const char *name, struct nvkm_vmm **pvmm)
 {
-	return nv04_vmm_new_(&gv100_vmm, mmu, 0, addr, size,
-			     argv, argc, key, name, pvmm);
+	return gp100_vmm_new_(&gv100_vmm, mmu, managed, addr, size,
+			      argv, argc, key, name, pvmm);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmmcp77.c

···
 };
 
 int
-mcp77_vmm_new(struct nvkm_mmu *mmu, u64 addr, u64 size, void *argv, u32 argc,
-	      struct lock_class_key *key, const char *name,
-	      struct nvkm_vmm **pvmm)
+mcp77_vmm_new(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	      void *argv, u32 argc, struct lock_class_key *key,
+	      const char *name, struct nvkm_vmm **pvmm)
 {
-	return nv04_vmm_new_(&mcp77_vmm, mmu, 0, addr, size,
+	return nv04_vmm_new_(&mcp77_vmm, mmu, 0, managed, addr, size,
 			     argv, argc, key, name, pvmm);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmnv04.c

···
 
 int
 nv04_vmm_new_(const struct nvkm_vmm_func *func, struct nvkm_mmu *mmu,
-	      u32 pd_header, u64 addr, u64 size, void *argv, u32 argc,
-	      struct lock_class_key *key, const char *name,
-	      struct nvkm_vmm **pvmm)
+	      u32 pd_header, bool managed, u64 addr, u64 size,
+	      void *argv, u32 argc, struct lock_class_key *key,
+	      const char *name, struct nvkm_vmm **pvmm)
 {
 	union {
 		struct nv04_vmm_vn vn;
 	} *args = argv;
 	int ret;
 
-	ret = nvkm_vmm_new_(func, mmu, pd_header, addr, size, key, name, pvmm);
+	ret = nvkm_vmm_new_(func, mmu, pd_header, managed, addr, size,
+			    key, name, pvmm);
 	if (ret)
 		return ret;
 
···
 }
 
 int
-nv04_vmm_new(struct nvkm_mmu *mmu, u64 addr, u64 size, void *argv, u32 argc,
-	     struct lock_class_key *key, const char *name,
+nv04_vmm_new(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	     void *argv, u32 argc, struct lock_class_key *key, const char *name,
 	     struct nvkm_vmm **pvmm)
 {
 	struct nvkm_memory *mem;
 	struct nvkm_vmm *vmm;
 	int ret;
 
-	ret = nv04_vmm_new_(&nv04_vmm, mmu, 8, addr, size,
+	ret = nv04_vmm_new_(&nv04_vmm, mmu, 8, managed, addr, size,
 			    argv, argc, key, name, &vmm);
 	*pvmm = vmm;
 	if (ret)

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmnv41.c

···
 };
 
 int
-nv41_vmm_new(struct nvkm_mmu *mmu, u64 addr, u64 size, void *argv, u32 argc,
-	     struct lock_class_key *key, const char *name,
+nv41_vmm_new(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	     void *argv, u32 argc, struct lock_class_key *key, const char *name,
 	     struct nvkm_vmm **pvmm)
 {
-	return nv04_vmm_new_(&nv41_vmm, mmu, 0, addr, size,
+	return nv04_vmm_new_(&nv41_vmm, mmu, 0, managed, addr, size,
 			     argv, argc, key, name, pvmm);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmnv44.c

···
 };
 
 int
-nv44_vmm_new(struct nvkm_mmu *mmu, u64 addr, u64 size, void *argv, u32 argc,
-	     struct lock_class_key *key, const char *name,
+nv44_vmm_new(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	     void *argv, u32 argc, struct lock_class_key *key, const char *name,
 	     struct nvkm_vmm **pvmm)
 {
 	struct nvkm_subdev *subdev = &mmu->subdev;
 	struct nvkm_vmm *vmm;
 	int ret;
 
-	ret = nv04_vmm_new_(&nv44_vmm, mmu, 0, addr, size,
+	ret = nv04_vmm_new_(&nv44_vmm, mmu, 0, managed, addr, size,
 			    argv, argc, key, name, &vmm);
 	*pvmm = vmm;
 	if (ret)

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmnv50.c

···
 };
 
 int
-nv50_vmm_new(struct nvkm_mmu *mmu, u64 addr, u64 size, void *argv, u32 argc,
-	     struct lock_class_key *key, const char *name,
+nv50_vmm_new(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	     void *argv, u32 argc, struct lock_class_key *key, const char *name,
 	     struct nvkm_vmm **pvmm)
 {
-	return nv04_vmm_new_(&nv50_vmm, mmu, 0, addr, size,
+	return nv04_vmm_new_(&nv50_vmm, mmu, 0, managed, addr, size,
 			     argv, argc, key, name, pvmm);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmtu104.c drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmtu102.c

···
 #include <subdev/timer.h>
 
 static void
-tu104_vmm_flush(struct nvkm_vmm *vmm, int depth)
+tu102_vmm_flush(struct nvkm_vmm *vmm, int depth)
 {
 	struct nvkm_subdev *subdev = &vmm->mmu->subdev;
 	struct nvkm_device *device = subdev->device;
···
 }
 
 static const struct nvkm_vmm_func
-tu104_vmm = {
+tu102_vmm = {
 	.join = gv100_vmm_join,
 	.part = gf100_vmm_part,
 	.aper = gf100_vmm_aper,
 	.valid = gp100_vmm_valid,
-	.flush = tu104_vmm_flush,
+	.flush = tu102_vmm_flush,
+	.mthd = gp100_vmm_mthd,
 	.page = {
 		{ 47, &gp100_vmm_desc_16[4], NVKM_VMM_PAGE_Sxxx },
 		{ 38, &gp100_vmm_desc_16[3], NVKM_VMM_PAGE_Sxxx },
···
 };
 
 int
-tu104_vmm_new(struct nvkm_mmu *mmu, u64 addr, u64 size,
+tu102_vmm_new(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
 	      void *argv, u32 argc, struct lock_class_key *key,
 	      const char *name, struct nvkm_vmm **pvmm)
 {
-	return nv04_vmm_new_(&tu104_vmm, mmu, 0, addr, size,
-			     argv, argc, key, name, pvmm);
+	return gp100_vmm_new_(&tu102_vmm, mmu, managed, addr, size,
+			      argv, argc, key, name, pvmm);
 }

drivers/gpu/drm/nouveau/nvkm/subdev/pmu/memx.c

···
 	if (exec) {
 		nvkm_pmu_send(pmu, reply, PROC_MEMX, MEMX_MSG_EXEC,
 			      memx->base, finish);
+		nvkm_debug(subdev, "Exec took %uns, PMU_IN %08x\n",
+			   reply[0], reply[1]);
 	}
 
-	nvkm_debug(subdev, "Exec took %uns, PMU_IN %08x\n",
-		   reply[0], reply[1]);
 	kfree(memx);
 	return 0;
 }

drivers/gpu/drm/nouveau/nvkm/subdev/secboot/acr_r352.c

···
 		 * and the expected behavior on RM as well
 		 */
 		if (ret && ret != 0x1d) {
-			nvkm_error(subdev, "HS unload failed, ret 0x%08x", ret);
+			nvkm_error(subdev, "HS unload failed, ret 0x%08x\n", ret);
 			return -EINVAL;
 		}
 		nvkm_debug(subdev, "HS unload blob completed\n");
···
 	if (ret < 0) {
 		return ret;
 	} else if (ret > 0) {
-		nvkm_error(subdev, "HS load failed, ret 0x%08x", ret);
+		nvkm_error(subdev, "HS load failed, ret 0x%08x\n", ret);
 		return -EINVAL;
 	}
 	nvkm_debug(subdev, "HS load blob completed\n");

drivers/gpu/drm/nouveau/nvkm/subdev/top/base.c

···
 }
 
 u32
+nvkm_top_addr(struct nvkm_device *device, enum nvkm_devidx index)
+{
+	struct nvkm_top *top = device->top;
+	struct nvkm_top_device *info;
+
+	if (top) {
+		list_for_each_entry(info, &top->device, head) {
+			if (info->index == index)
+				return info->addr;
+		}
+	}
+
+	return 0;
+}
+
+u32
 nvkm_top_reset(struct nvkm_device *device, enum nvkm_devidx index)
 {
 	struct nvkm_top *top = device->top;

drivers/gpu/drm/nouveau/nvkm/subdev/top/gk104.c

···
 #define A_(A) if (inst == 0) info->index = NVKM_ENGINE_##A
 #define B_(A) if (inst + NVKM_ENGINE_##A##0 < NVKM_ENGINE_##A##_LAST + 1)      \
 		info->index = NVKM_ENGINE_##A##0 + inst
+#define C_(A) if (inst == 0) info->index = NVKM_SUBDEV_##A
 		switch (type) {
 		case 0x00000000: A_(GR    ); break;
 		case 0x00000001: A_(CE0   ); break;
···
 		case 0x0000000f: A_(NVENC1); break;
 		case 0x00000010: B_(NVDEC ); break;
 		case 0x00000013: B_(CE    ); break;
+		case 0x00000014: C_(GSP   ); break;
 			break;
 		default:
 			break;

drivers/gpu/drm/nouveau/nvkm/subdev/volt/Kbuild

···
 nvkm-y += nvkm/subdev/volt/gpio.o
 nvkm-y += nvkm/subdev/volt/nv40.o
 nvkm-y += nvkm/subdev/volt/gf100.o
+nvkm-y += nvkm/subdev/volt/gf117.o
 nvkm-y += nvkm/subdev/volt/gk104.o
 nvkm-y += nvkm/subdev/volt/gk20a.o
 nvkm-y += nvkm/subdev/volt/gm20b.o

include/uapi/drm/nouveau_drm.h

···
 #define DRM_NOUVEAU_NOTIFIEROBJ_ALLOC  0x05 /* deprecated */
 #define DRM_NOUVEAU_GPUOBJ_FREE        0x06 /* deprecated */
 #define DRM_NOUVEAU_NVIF               0x07
+#define DRM_NOUVEAU_SVM_INIT           0x08
+#define DRM_NOUVEAU_SVM_BIND           0x09
 #define DRM_NOUVEAU_GEM_NEW            0x40
 #define DRM_NOUVEAU_GEM_PUSHBUF        0x41
 #define DRM_NOUVEAU_GEM_CPU_PREP       0x42
 #define DRM_NOUVEAU_GEM_CPU_FINI       0x43
 #define DRM_NOUVEAU_GEM_INFO           0x44
+
+struct drm_nouveau_svm_init {
+	__u64 unmanaged_addr;
+	__u64 unmanaged_size;
+};
+
+struct drm_nouveau_svm_bind {
+	__u64 header;
+	__u64 va_start;
+	__u64 va_end;
+	__u64 npages;
+	__u64 stride;
+	__u64 result;
+	__u64 reserved0;
+	__u64 reserved1;
+};
+
+#define NOUVEAU_SVM_BIND_COMMAND_SHIFT          0
+#define NOUVEAU_SVM_BIND_COMMAND_BITS           8
+#define NOUVEAU_SVM_BIND_COMMAND_MASK           ((1 << 8) - 1)
+#define NOUVEAU_SVM_BIND_PRIORITY_SHIFT         8
+#define NOUVEAU_SVM_BIND_PRIORITY_BITS          8
+#define NOUVEAU_SVM_BIND_PRIORITY_MASK          ((1 << 8) - 1)
+#define NOUVEAU_SVM_BIND_TARGET_SHIFT           16
+#define NOUVEAU_SVM_BIND_TARGET_BITS            32
+#define NOUVEAU_SVM_BIND_TARGET_MASK            0xffffffff
+
+/*
+ * Below is use to validate ioctl argument, userspace can also use it to make
+ * sure that no bit are set beyond known fields for a given kernel version.
+ */
+#define NOUVEAU_SVM_BIND_VALID_BITS     48
+#define NOUVEAU_SVM_BIND_VALID_MASK     ((1ULL << NOUVEAU_SVM_BIND_VALID_BITS) - 1)
+
+
+/*
+ * NOUVEAU_BIND_COMMAND__MIGRATE: synchronous migrate to target memory.
+ * result: number of page successfuly migrate to the target memory.
+ */
+#define NOUVEAU_SVM_BIND_COMMAND__MIGRATE               0
+
+/*
+ * NOUVEAU_SVM_BIND_HEADER_TARGET__GPU_VRAM: target the GPU VRAM memory.
+ */
+#define NOUVEAU_SVM_BIND_TARGET__GPU_VRAM               (1UL << 31)
+
+
+#define DRM_IOCTL_NOUVEAU_SVM_INIT           DRM_IOWR(DRM_COMMAND_BASE + DRM_NOUVEAU_SVM_INIT, struct drm_nouveau_svm_init)
+#define DRM_IOCTL_NOUVEAU_SVM_BIND           DRM_IOWR(DRM_COMMAND_BASE + DRM_NOUVEAU_SVM_BIND, struct drm_nouveau_svm_bind)
 
 #define DRM_IOCTL_NOUVEAU_GEM_NEW            DRM_IOWR(DRM_COMMAND_BASE + DRM_NOUVEAU_GEM_NEW, struct drm_nouveau_gem_new)
 #define DRM_IOCTL_NOUVEAU_GEM_PUSHBUF        DRM_IOWR(DRM_COMMAND_BASE + DRM_NOUVEAU_GEM_PUSHBUF, struct drm_nouveau_gem_pushbuf)