diff --git a/drivers/gpu/drm/nouveau/include/nvkm/engine/fifo.h b/drivers/gpu/drm/nouveau/include/nvkm/engine/fifo.h
index 1e12697beae2..62fe64dab11a 100644
--- a/drivers/gpu/drm/nouveau/include/nvkm/engine/fifo.h
+++ b/drivers/gpu/drm/nouveau/include/nvkm/engine/fifo.h
@@ -6,7 +6,6 @@
 #include <core/event.h>
 struct nvkm_fault_data;
 
-#define NVKM_FIFO_CHID_NR 4096
 #define NVKM_FIFO_ENGN_NR 16
 
 struct nvkm_fifo_engn {
@@ -17,13 +16,16 @@ struct nvkm_fifo_engn {
 
 struct nvkm_chan {
 	const struct nvkm_chan_func *func;
+	char name[64];
+	struct nvkm_cgrp *cgrp;
+
+	union { int id; int chid; }; /*FIXME: remove later */
 
 	struct nvkm_fifo *fifo;
 	u32 engm;
 	struct nvkm_object object;
 
 	struct list_head head;
-	u16 chid;
 	struct nvkm_gpuobj *inst;
 	struct nvkm_gpuobj *push;
 	struct nvkm_vmm *vmm;
@@ -43,7 +45,6 @@ struct nvkm_fifo {
 	struct list_head runqs;
 	struct list_head runls;
 
-	DECLARE_BITMAP(mask, NVKM_FIFO_CHID_NR);
 	int nr;
 	struct list_head chan;
 	spinlock_t lock;
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/Kbuild b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/Kbuild
index a92a88a72e59..eb82da7e7d87 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/Kbuild
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/Kbuild
@@ -1,5 +1,6 @@
 # SPDX-License-Identifier: MIT
 nvkm-y += nvkm/engine/fifo/base.o
+nvkm-y += nvkm/engine/fifo/cgrp.o
 nvkm-y += nvkm/engine/fifo/chan.o
 nvkm-y += nvkm/engine/fifo/chid.o
 nvkm-y += nvkm/engine/fifo/runl.o
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/base.c b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/base.c
index 0f603d9f00cb..3e62eacc48ad 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/base.c
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/base.c
@@ -389,10 +389,7 @@ nvkm_fifo_ctor(const struct nvkm_fifo_func *func, struct nvkm_device *device,
 	INIT_LIST_HEAD(&fifo->chan);
 
 	nr = func->chid_nr(fifo);
-	if (WARN_ON(fifo->nr > NVKM_FIFO_CHID_NR))
-		fifo->nr = NVKM_FIFO_CHID_NR;
-	else
-		fifo->nr = nr;
+	fifo->nr = nr;
 
 	if (func->uevent_init) {
 		ret = nvkm_event_init(&nvkm_fifo_uevent_func, &fifo->engine.subdev, 1, 1,
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/cgrp.c b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/cgrp.c
new file mode 100644
index 000000000000..e6ac178db8b5
--- /dev/null
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/cgrp.c
@@ -0,0 +1,94 @@
+/*
+ * Copyright 2021 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 "cgrp.h"
+#include "chan.h"
+#include "chid.h"
+#include "runl.h"
+#include "priv.h"
+
+#include <subdev/mmu.h>
+
+static void
+nvkm_cgrp_del(struct kref *kref)
+{
+	struct nvkm_cgrp *cgrp = container_of(kref, typeof(*cgrp), kref);
+	struct nvkm_runl *runl = cgrp->runl;
+
+	if (runl->cgid)
+		nvkm_chid_put(runl->cgid, cgrp->id, &cgrp->lock);
+
+	nvkm_vmm_unref(&cgrp->vmm);
+	kfree(cgrp);
+}
+
+void
+nvkm_cgrp_unref(struct nvkm_cgrp **pcgrp)
+{
+	struct nvkm_cgrp *cgrp = *pcgrp;
+
+	if (!cgrp)
+		return;
+
+	kref_put(&cgrp->kref, nvkm_cgrp_del);
+	*pcgrp = NULL;
+}
+
+struct nvkm_cgrp *
+nvkm_cgrp_ref(struct nvkm_cgrp *cgrp)
+{
+	if (cgrp)
+		kref_get(&cgrp->kref);
+
+	return cgrp;
+}
+
+int
+nvkm_cgrp_new(struct nvkm_runl *runl, const char *name, struct nvkm_vmm *vmm, bool hw,
+	      struct nvkm_cgrp **pcgrp)
+{
+	struct nvkm_cgrp *cgrp;
+
+	if (!(cgrp = *pcgrp = kmalloc(sizeof(*cgrp), GFP_KERNEL)))
+		return -ENOMEM;
+
+	cgrp->func = runl->fifo->func->cgrp.func;
+	strscpy(cgrp->name, name, sizeof(cgrp->name));
+	cgrp->runl = runl;
+	cgrp->vmm = nvkm_vmm_ref(vmm);
+	cgrp->hw = hw;
+	cgrp->id = -1;
+	kref_init(&cgrp->kref);
+	cgrp->chans = NULL;
+	cgrp->chan_nr = 0;
+	spin_lock_init(&cgrp->lock);
+
+	if (runl->cgid) {
+		cgrp->id = nvkm_chid_get(runl->cgid, cgrp);
+		if (cgrp->id < 0) {
+			RUNL_ERROR(runl, "!cgids");
+			nvkm_cgrp_unref(pcgrp);
+			return -ENOSPC;
+		}
+	}
+
+	return 0;
+}
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/cgrp.h b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/cgrp.h
index 9a677a1acf9b..d0510df9286a 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/cgrp.h
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/cgrp.h
@@ -2,13 +2,34 @@
 #ifndef __NVKM_CGRP_H__
 #define __NVKM_CGRP_H__
 #include <core/os.h>
+struct nvkm_chan;
 
 struct nvkm_cgrp {
 	const struct nvkm_cgrp_func {
 	} *func;
+	char name[64];
+	struct nvkm_runl *runl;
+	struct nvkm_vmm *vmm;
+	bool hw;
 	int id;
+	struct kref kref;
+
+	struct nvkm_chan *chans;
+	int chan_nr;
+
+	spinlock_t lock; /* protects irq handler channel (group) lookup */
+
 	struct list_head head;
 	struct list_head chan;
-	int chan_nr;
 };
+
+int nvkm_cgrp_new(struct nvkm_runl *, const char *name, struct nvkm_vmm *, bool hw,
+		  struct nvkm_cgrp **);
+struct nvkm_cgrp *nvkm_cgrp_ref(struct nvkm_cgrp *);
+void nvkm_cgrp_unref(struct nvkm_cgrp **);
+
+#define CGRP_PRCLI(c,l,p,f,a...) RUNL_PRINT((c)->runl, l, p, "%04x:[%s]"f, (c)->id, (c)->name, ##a)
+#define CGRP_PRINT(c,l,p,f,a...) RUNL_PRINT((c)->runl, l, p, "%04x:"f, (c)->id, ##a)
+#define CGRP_ERROR(c,f,a...) CGRP_PRCLI((c), ERROR,    err, " "f"\n", ##a)
+#define CGRP_TRACE(c,f,a...) CGRP_PRINT((c), TRACE,   info, " "f"\n", ##a)
 #endif
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chan.c b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chan.c
index 464c956aaca9..e2178a8f78a7 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chan.c
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chan.c
@@ -22,6 +22,9 @@
  * Authors: Ben Skeggs
  */
 #include "chan.h"
+#include "chid.h"
+#include "cgrp.h"
+#include "runl.h"
 #include "priv.h"
 
 #include <core/client.h>
@@ -312,6 +315,11 @@ nvkm_chan_del(struct nvkm_chan **pchan)
 	if (!chan)
 		return;
 
+	if (chan->cgrp) {
+		nvkm_chid_put(chan->cgrp->runl->chid, chan->id, &chan->cgrp->lock);
+		nvkm_cgrp_unref(&chan->cgrp);
+	}
+
 	chan = nvkm_object_dtor(&chan->object);
 	kfree(chan);
 }
@@ -326,7 +334,6 @@ nvkm_fifo_chan_dtor(struct nvkm_object *object)
 
 	spin_lock_irqsave(&fifo->lock, flags);
 	if (!list_empty(&chan->head)) {
-		__clear_bit(chan->chid, fifo->mask);
 		list_del(&chan->head);
 	}
 	spin_unlock_irqrestore(&fifo->lock, flags);
@@ -363,9 +370,22 @@ nvkm_fifo_chan_ctor(const struct nvkm_fifo_chan_func *fn,
 	struct nvkm_client *client = oclass->client;
 	struct nvkm_device *device = fifo->engine.subdev.device;
 	struct nvkm_dmaobj *dmaobj;
+	struct nvkm_cgrp *cgrp = NULL;
+	struct nvkm_runl *runl;
+	struct nvkm_engn *engn = NULL;
+	struct nvkm_vmm *vmm = NULL;
 	unsigned long flags;
 	int ret;
 
+	nvkm_runl_foreach(runl, fifo) {
+		engn = nvkm_runl_find_engn(engn, runl, engm & BIT(engn->id));
+		if (engn)
+			break;
+	}
+
+	if (!engn)
+		return -EINVAL;
+
 	/*FIXME: temp kludge to ease transition, remove later */
 	if (!(func = kmalloc(sizeof(*func), GFP_KERNEL)))
 		return -ENOMEM;
@@ -383,12 +403,38 @@ nvkm_fifo_chan_ctor(const struct nvkm_fifo_chan_func *fn,
 	func->submit_token = fn->submit_token;
 
 	chan->func = func;
+	chan->id = -1;
 
 	nvkm_object_ctor(&nvkm_fifo_chan_func, oclass, &chan->object);
 	chan->fifo = fifo;
 	chan->engm = engm;
 	INIT_LIST_HEAD(&chan->head);
 
+	/* Join channel group.
+	 *
+	 * GK110 and newer support channel groups (aka TSGs), where individual channels
+	 * share a timeslice, and, engine context(s).
+	 *
+	 * As such, engine contexts are tracked in nvkm_cgrp and we need them even when
+	 * channels aren't in an API channel group, and on HW that doesn't support TSGs.
+	 */
+	if (!cgrp) {
+		ret = nvkm_cgrp_new(runl, chan->name, vmm, fifo->func->cgrp.force, &chan->cgrp);
+		if (ret) {
+			RUNL_DEBUG(runl, "cgrp %d", ret);
+			return ret;
+		}
+
+		cgrp = chan->cgrp;
+	} else {
+		if (cgrp->runl != runl || cgrp->vmm != vmm) {
+			RUNL_DEBUG(runl, "cgrp %d %d", cgrp->runl != runl, cgrp->vmm != vmm);
+			return -EINVAL;
+		}
+
+		chan->cgrp = nvkm_cgrp_ref(cgrp);
+	}
+
 	/* instance memory */
 	ret = nvkm_gpuobj_new(device, size, align, zero, NULL, &chan->inst);
 	if (ret)
@@ -422,15 +468,23 @@ nvkm_fifo_chan_ctor(const struct nvkm_fifo_chan_func *fn,
 		chan->vmm = nvkm_vmm_ref(vmm);
 	}
 
-	/* allocate channel id */
-	spin_lock_irqsave(&fifo->lock, flags);
-	chan->chid = find_first_zero_bit(fifo->mask, NVKM_FIFO_CHID_NR);
-	if (chan->chid >= NVKM_FIFO_CHID_NR) {
-		spin_unlock_irqrestore(&fifo->lock, flags);
+	/* Allocate channel ID. */
+	if (runl->cgid) {
+		chan->id = chan->cgrp->id;
+		runl->chid->data[chan->id] = chan;
+		set_bit(chan->id, runl->chid->used);
+		goto temp_hack_until_no_chid_eq_cgid_req;
+	}
+
+	chan->id = nvkm_chid_get(runl->chid, chan);
+	if (chan->id < 0) {
+		RUNL_ERROR(runl, "!chids");
 		return -ENOSPC;
 	}
+
+temp_hack_until_no_chid_eq_cgid_req:
+	spin_lock_irqsave(&fifo->lock, flags);
 	list_add(&chan->head, &fifo->chan);
-	__set_bit(chan->chid, fifo->mask);
 	spin_unlock_irqrestore(&fifo->lock, flags);
 
 	/* determine address of this channel's user registers */
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chan.h b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chan.h
index a9c0e02b6f9f..0f1added5c24 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chan.h
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chan.h
@@ -27,5 +27,10 @@ int nvkm_fifo_chan_ctor(const struct nvkm_fifo_chan_func *, struct nvkm_fifo *,
 			const struct nvkm_oclass *, struct nvkm_fifo_chan *);
 void nvkm_chan_del(struct nvkm_chan **);
 
+#define CHAN_PRCLI(c,l,p,f,a...) CGRP_PRINT((c)->cgrp, l, p, "%04x:[%s]"f, (c)->id, (c)->name, ##a)
+#define CHAN_PRINT(c,l,p,f,a...) CGRP_PRINT((c)->cgrp, l, p, "%04x:"f, (c)->id, ##a)
+#define CHAN_ERROR(c,f,a...) CHAN_PRCLI((c), ERROR,    err, " "f"\n", ##a)
+#define CHAN_TRACE(c,f,a...) CHAN_PRINT((c), TRACE,   info, " "f"\n", ##a)
+
 int nvkm_fifo_chan_child_new(const struct nvkm_oclass *, void *, u32, struct nvkm_object **);
 #endif
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chid.c b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chid.c
index 7c582bce7e24..23944d95efd5 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chid.c
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chid.c
@@ -21,6 +21,35 @@
  */
 #include "chid.h"
 
+void
+nvkm_chid_put(struct nvkm_chid *chid, int id, spinlock_t *data_lock)
+{
+	if (id >= 0) {
+		spin_lock_irq(&chid->lock);
+		spin_lock(data_lock);
+		chid->data[id] = NULL;
+		spin_unlock(data_lock);
+		clear_bit(id, chid->used);
+		spin_unlock_irq(&chid->lock);
+	}
+}
+
+int
+nvkm_chid_get(struct nvkm_chid *chid, void *data)
+{
+	int id = -1, cid;
+
+	spin_lock_irq(&chid->lock);
+	cid = find_first_zero_bit(chid->used, chid->nr);
+	if (cid < chid->nr) {
+		set_bit(cid, chid->used);
+		chid->data[cid] = data;
+		id = cid;
+	}
+	spin_unlock_irq(&chid->lock);
+	return id;
+}
+
 static void
 nvkm_chid_del(struct kref *kref)
 {
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chid.h b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chid.h
index d359828dde80..2a42efb18401 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chid.h
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chid.h
@@ -20,4 +20,6 @@ int nvkm_chid_new(const struct nvkm_event_func *, struct nvkm_subdev *,
 		  int nr, int first, int count, struct nvkm_chid **pchid);
 struct nvkm_chid *nvkm_chid_ref(struct nvkm_chid *);
 void nvkm_chid_unref(struct nvkm_chid **);
+int nvkm_chid_get(struct nvkm_chid *, void *data);
+void nvkm_chid_put(struct nvkm_chid *, int id, spinlock_t *data_lock);
 #endif
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/nv04.c b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/nv04.c
index ce965d6aa874..52fe9d5a5a4a 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/nv04.c
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/nv04.c
@@ -489,7 +489,6 @@ nv04_fifo_new_(const struct nvkm_fifo_func *func, struct nvkm_device *device,
 	if (ret)
 		return ret;
 
-	set_bit(nr - 1, fifo->base.mask); /* inactive channel */
 	return 0;
 }
 
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/nv50.c b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/nv50.c
index e4a7ef5aa461..4fcf49dbfd90 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/nv50.c
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/nv50.c
@@ -155,8 +155,6 @@ nv50_fifo_new_(const struct nvkm_fifo_func *func, struct nvkm_device *device,
 	if (ret)
 		return ret;
 
-	set_bit(0, fifo->base.mask); /* PIO channel */
-	set_bit(127, fifo->base.mask); /* inactive channel */
 	return 0;
 }
 
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/runl.c b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/runl.c
index 4d7319674128..1a15b0a9ab7c 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/runl.c
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/runl.c
@@ -38,6 +38,7 @@ nvkm_runl_del(struct nvkm_runl *runl)
 	nvkm_chid_unref(&runl->cgid);
 
 	list_del(&runl->head);
+	mutex_destroy(&runl->mutex);
 	kfree(runl);
 }
 
@@ -94,6 +95,8 @@ nvkm_runl_new(struct nvkm_fifo *fifo, int runi, u32 addr, int id_nr)
 	runl->id = runi;
 	runl->addr = addr;
 	INIT_LIST_HEAD(&runl->engns);
+	INIT_LIST_HEAD(&runl->cgrps);
+	mutex_init(&runl->mutex);
 	list_add_tail(&runl->head, &fifo->runls);
 
 	if (!fifo->chid) {
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/runl.h b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/runl.h
index 9f27babc8caf..cb0de870a118 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/runl.h
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/runl.h
@@ -35,6 +35,11 @@ struct nvkm_runl {
 	struct nvkm_runq *runq[2];
 	int runq_nr;
 
+	struct list_head cgrps;
+	int cgrp_nr;
+	int chan_nr;
+	struct mutex mutex;
+
 	struct list_head head;
 };
 
@@ -44,8 +49,12 @@ struct nvkm_engn *nvkm_runl_add(struct nvkm_runl *, int engi, const struct nvkm_
 				enum nvkm_subdev_type, int inst);
 void nvkm_runl_del(struct nvkm_runl *);
 
+#define nvkm_runl_find_engn(engn,runl,cond) nvkm_list_find(engn, &(runl)->engns, head, (cond))
+
 #define nvkm_runl_foreach(runl,fifo) list_for_each_entry((runl), &(fifo)->runls, head)
 #define nvkm_runl_foreach_engn(engn,runl) list_for_each_entry((engn), &(runl)->engns, head)
+#define nvkm_runl_foreach_engn_cond(engn,runl,cond) \
+	nvkm_list_foreach(engn, &(runl)->engns, head, (cond))
 
 #define RUNL_PRINT(r,l,p,f,a...)                                                          \
 	nvkm_printk__(&(r)->fifo->engine.subdev, NV_DBG_##l, p, "%06x:"f, (r)->addr, ##a)