Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / drivers / gpu / drm / nouveau / nvkm / engine / fifo / nv04.c
diff --git a/kernel/drivers/gpu/drm/nouveau/nvkm/engine/fifo/nv04.c b/kernel/drivers/gpu/drm/nouveau/nvkm/engine/fifo/nv04.c
new file mode 100644 (file)
index 0000000..043e429
--- /dev/null
@@ -0,0 +1,635 @@
+/*
+ * Copyright 2012 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.
+ *
+ * Authors: Ben Skeggs
+ */
+#include "nv04.h"
+
+#include <core/client.h>
+#include <core/device.h>
+#include <core/engctx.h>
+#include <core/handle.h>
+#include <core/ramht.h>
+#include <subdev/instmem/nv04.h>
+#include <subdev/timer.h>
+
+#include <nvif/class.h>
+#include <nvif/unpack.h>
+
+static struct ramfc_desc
+nv04_ramfc[] = {
+       { 32,  0, 0x00,  0, NV04_PFIFO_CACHE1_DMA_PUT },
+       { 32,  0, 0x04,  0, NV04_PFIFO_CACHE1_DMA_GET },
+       { 16,  0, 0x08,  0, NV04_PFIFO_CACHE1_DMA_INSTANCE },
+       { 16, 16, 0x08,  0, NV04_PFIFO_CACHE1_DMA_DCOUNT },
+       { 32,  0, 0x0c,  0, NV04_PFIFO_CACHE1_DMA_STATE },
+       { 32,  0, 0x10,  0, NV04_PFIFO_CACHE1_DMA_FETCH },
+       { 32,  0, 0x14,  0, NV04_PFIFO_CACHE1_ENGINE },
+       { 32,  0, 0x18,  0, NV04_PFIFO_CACHE1_PULL1 },
+       {}
+};
+
+/*******************************************************************************
+ * FIFO channel objects
+ ******************************************************************************/
+
+int
+nv04_fifo_object_attach(struct nvkm_object *parent,
+                       struct nvkm_object *object, u32 handle)
+{
+       struct nv04_fifo_priv *priv = (void *)parent->engine;
+       struct nv04_fifo_chan *chan = (void *)parent;
+       u32 context, chid = chan->base.chid;
+       int ret;
+
+       if (nv_iclass(object, NV_GPUOBJ_CLASS))
+               context = nv_gpuobj(object)->addr >> 4;
+       else
+               context = 0x00000004; /* just non-zero */
+
+       switch (nv_engidx(object->engine)) {
+       case NVDEV_ENGINE_DMAOBJ:
+       case NVDEV_ENGINE_SW:
+               context |= 0x00000000;
+               break;
+       case NVDEV_ENGINE_GR:
+               context |= 0x00010000;
+               break;
+       case NVDEV_ENGINE_MPEG:
+               context |= 0x00020000;
+               break;
+       default:
+               return -EINVAL;
+       }
+
+       context |= 0x80000000; /* valid */
+       context |= chid << 24;
+
+       mutex_lock(&nv_subdev(priv)->mutex);
+       ret = nvkm_ramht_insert(priv->ramht, chid, handle, context);
+       mutex_unlock(&nv_subdev(priv)->mutex);
+       return ret;
+}
+
+void
+nv04_fifo_object_detach(struct nvkm_object *parent, int cookie)
+{
+       struct nv04_fifo_priv *priv = (void *)parent->engine;
+       mutex_lock(&nv_subdev(priv)->mutex);
+       nvkm_ramht_remove(priv->ramht, cookie);
+       mutex_unlock(&nv_subdev(priv)->mutex);
+}
+
+int
+nv04_fifo_context_attach(struct nvkm_object *parent,
+                        struct nvkm_object *object)
+{
+       nv_engctx(object)->addr = nvkm_fifo_chan(parent)->chid;
+       return 0;
+}
+
+static int
+nv04_fifo_chan_ctor(struct nvkm_object *parent,
+                   struct nvkm_object *engine,
+                   struct nvkm_oclass *oclass, void *data, u32 size,
+                   struct nvkm_object **pobject)
+{
+       union {
+               struct nv03_channel_dma_v0 v0;
+       } *args = data;
+       struct nv04_fifo_priv *priv = (void *)engine;
+       struct nv04_fifo_chan *chan;
+       int ret;
+
+       nv_ioctl(parent, "create channel dma size %d\n", size);
+       if (nvif_unpack(args->v0, 0, 0, false)) {
+               nv_ioctl(parent, "create channel dma vers %d pushbuf %08x "
+                                "offset %016llx\n", args->v0.version,
+                        args->v0.pushbuf, args->v0.offset);
+       } else
+               return ret;
+
+       ret = nvkm_fifo_channel_create(parent, engine, oclass, 0, 0x800000,
+                                      0x10000, args->v0.pushbuf,
+                                      (1ULL << NVDEV_ENGINE_DMAOBJ) |
+                                      (1ULL << NVDEV_ENGINE_SW) |
+                                      (1ULL << NVDEV_ENGINE_GR), &chan);
+       *pobject = nv_object(chan);
+       if (ret)
+               return ret;
+
+       args->v0.chid = chan->base.chid;
+
+       nv_parent(chan)->object_attach = nv04_fifo_object_attach;
+       nv_parent(chan)->object_detach = nv04_fifo_object_detach;
+       nv_parent(chan)->context_attach = nv04_fifo_context_attach;
+       chan->ramfc = chan->base.chid * 32;
+
+       nv_wo32(priv->ramfc, chan->ramfc + 0x00, args->v0.offset);
+       nv_wo32(priv->ramfc, chan->ramfc + 0x04, args->v0.offset);
+       nv_wo32(priv->ramfc, chan->ramfc + 0x08, chan->base.pushgpu->addr >> 4);
+       nv_wo32(priv->ramfc, chan->ramfc + 0x10,
+                            NV_PFIFO_CACHE1_DMA_FETCH_TRIG_128_BYTES |
+                            NV_PFIFO_CACHE1_DMA_FETCH_SIZE_128_BYTES |
+#ifdef __BIG_ENDIAN
+                            NV_PFIFO_CACHE1_BIG_ENDIAN |
+#endif
+                            NV_PFIFO_CACHE1_DMA_FETCH_MAX_REQS_8);
+       return 0;
+}
+
+void
+nv04_fifo_chan_dtor(struct nvkm_object *object)
+{
+       struct nv04_fifo_priv *priv = (void *)object->engine;
+       struct nv04_fifo_chan *chan = (void *)object;
+       struct ramfc_desc *c = priv->ramfc_desc;
+
+       do {
+               nv_wo32(priv->ramfc, chan->ramfc + c->ctxp, 0x00000000);
+       } while ((++c)->bits);
+
+       nvkm_fifo_channel_destroy(&chan->base);
+}
+
+int
+nv04_fifo_chan_init(struct nvkm_object *object)
+{
+       struct nv04_fifo_priv *priv = (void *)object->engine;
+       struct nv04_fifo_chan *chan = (void *)object;
+       u32 mask = 1 << chan->base.chid;
+       unsigned long flags;
+       int ret;
+
+       ret = nvkm_fifo_channel_init(&chan->base);
+       if (ret)
+               return ret;
+
+       spin_lock_irqsave(&priv->base.lock, flags);
+       nv_mask(priv, NV04_PFIFO_MODE, mask, mask);
+       spin_unlock_irqrestore(&priv->base.lock, flags);
+       return 0;
+}
+
+int
+nv04_fifo_chan_fini(struct nvkm_object *object, bool suspend)
+{
+       struct nv04_fifo_priv *priv = (void *)object->engine;
+       struct nv04_fifo_chan *chan = (void *)object;
+       struct nvkm_gpuobj *fctx = priv->ramfc;
+       struct ramfc_desc *c;
+       unsigned long flags;
+       u32 data = chan->ramfc;
+       u32 chid;
+
+       /* prevent fifo context switches */
+       spin_lock_irqsave(&priv->base.lock, flags);
+       nv_wr32(priv, NV03_PFIFO_CACHES, 0);
+
+       /* if this channel is active, replace it with a null context */
+       chid = nv_rd32(priv, NV03_PFIFO_CACHE1_PUSH1) & priv->base.max;
+       if (chid == chan->base.chid) {
+               nv_mask(priv, NV04_PFIFO_CACHE1_DMA_PUSH, 0x00000001, 0);
+               nv_wr32(priv, NV03_PFIFO_CACHE1_PUSH0, 0);
+               nv_mask(priv, NV04_PFIFO_CACHE1_PULL0, 0x00000001, 0);
+
+               c = priv->ramfc_desc;
+               do {
+                       u32 rm = ((1ULL << c->bits) - 1) << c->regs;
+                       u32 cm = ((1ULL << c->bits) - 1) << c->ctxs;
+                       u32 rv = (nv_rd32(priv, c->regp) &  rm) >> c->regs;
+                       u32 cv = (nv_ro32(fctx, c->ctxp + data) & ~cm);
+                       nv_wo32(fctx, c->ctxp + data, cv | (rv << c->ctxs));
+               } while ((++c)->bits);
+
+               c = priv->ramfc_desc;
+               do {
+                       nv_wr32(priv, c->regp, 0x00000000);
+               } while ((++c)->bits);
+
+               nv_wr32(priv, NV03_PFIFO_CACHE1_GET, 0);
+               nv_wr32(priv, NV03_PFIFO_CACHE1_PUT, 0);
+               nv_wr32(priv, NV03_PFIFO_CACHE1_PUSH1, priv->base.max);
+               nv_wr32(priv, NV03_PFIFO_CACHE1_PUSH0, 1);
+               nv_wr32(priv, NV04_PFIFO_CACHE1_PULL0, 1);
+       }
+
+       /* restore normal operation, after disabling dma mode */
+       nv_mask(priv, NV04_PFIFO_MODE, 1 << chan->base.chid, 0);
+       nv_wr32(priv, NV03_PFIFO_CACHES, 1);
+       spin_unlock_irqrestore(&priv->base.lock, flags);
+
+       return nvkm_fifo_channel_fini(&chan->base, suspend);
+}
+
+static struct nvkm_ofuncs
+nv04_fifo_ofuncs = {
+       .ctor = nv04_fifo_chan_ctor,
+       .dtor = nv04_fifo_chan_dtor,
+       .init = nv04_fifo_chan_init,
+       .fini = nv04_fifo_chan_fini,
+       .map  = _nvkm_fifo_channel_map,
+       .rd32 = _nvkm_fifo_channel_rd32,
+       .wr32 = _nvkm_fifo_channel_wr32,
+       .ntfy = _nvkm_fifo_channel_ntfy
+};
+
+static struct nvkm_oclass
+nv04_fifo_sclass[] = {
+       { NV03_CHANNEL_DMA, &nv04_fifo_ofuncs },
+       {}
+};
+
+/*******************************************************************************
+ * FIFO context - basically just the instmem reserved for the channel
+ ******************************************************************************/
+
+int
+nv04_fifo_context_ctor(struct nvkm_object *parent,
+                      struct nvkm_object *engine,
+                      struct nvkm_oclass *oclass, void *data, u32 size,
+                      struct nvkm_object **pobject)
+{
+       struct nv04_fifo_base *base;
+       int ret;
+
+       ret = nvkm_fifo_context_create(parent, engine, oclass, NULL, 0x1000,
+                                      0x1000, NVOBJ_FLAG_HEAP, &base);
+       *pobject = nv_object(base);
+       if (ret)
+               return ret;
+
+       return 0;
+}
+
+static struct nvkm_oclass
+nv04_fifo_cclass = {
+       .handle = NV_ENGCTX(FIFO, 0x04),
+       .ofuncs = &(struct nvkm_ofuncs) {
+               .ctor = nv04_fifo_context_ctor,
+               .dtor = _nvkm_fifo_context_dtor,
+               .init = _nvkm_fifo_context_init,
+               .fini = _nvkm_fifo_context_fini,
+               .rd32 = _nvkm_fifo_context_rd32,
+               .wr32 = _nvkm_fifo_context_wr32,
+       },
+};
+
+/*******************************************************************************
+ * PFIFO engine
+ ******************************************************************************/
+
+void
+nv04_fifo_pause(struct nvkm_fifo *pfifo, unsigned long *pflags)
+__acquires(priv->base.lock)
+{
+       struct nv04_fifo_priv *priv = (void *)pfifo;
+       unsigned long flags;
+
+       spin_lock_irqsave(&priv->base.lock, flags);
+       *pflags = flags;
+
+       nv_wr32(priv, NV03_PFIFO_CACHES, 0x00000000);
+       nv_mask(priv, NV04_PFIFO_CACHE1_PULL0, 0x00000001, 0x00000000);
+
+       /* in some cases the puller may be left in an inconsistent state
+        * if you try to stop it while it's busy translating handles.
+        * sometimes you get a CACHE_ERROR, sometimes it just fails
+        * silently; sending incorrect instance offsets to PGRAPH after
+        * it's started up again.
+        *
+        * to avoid this, we invalidate the most recently calculated
+        * instance.
+        */
+       if (!nv_wait(priv, NV04_PFIFO_CACHE1_PULL0,
+                          NV04_PFIFO_CACHE1_PULL0_HASH_BUSY, 0x00000000))
+               nv_warn(priv, "timeout idling puller\n");
+
+       if (nv_rd32(priv, NV04_PFIFO_CACHE1_PULL0) &
+                         NV04_PFIFO_CACHE1_PULL0_HASH_FAILED)
+               nv_wr32(priv, NV03_PFIFO_INTR_0, NV_PFIFO_INTR_CACHE_ERROR);
+
+       nv_wr32(priv, NV04_PFIFO_CACHE1_HASH, 0x00000000);
+}
+
+void
+nv04_fifo_start(struct nvkm_fifo *pfifo, unsigned long *pflags)
+__releases(priv->base.lock)
+{
+       struct nv04_fifo_priv *priv = (void *)pfifo;
+       unsigned long flags = *pflags;
+
+       nv_mask(priv, NV04_PFIFO_CACHE1_PULL0, 0x00000001, 0x00000001);
+       nv_wr32(priv, NV03_PFIFO_CACHES, 0x00000001);
+
+       spin_unlock_irqrestore(&priv->base.lock, flags);
+}
+
+static const char *
+nv_dma_state_err(u32 state)
+{
+       static const char * const desc[] = {
+               "NONE", "CALL_SUBR_ACTIVE", "INVALID_MTHD", "RET_SUBR_INACTIVE",
+               "INVALID_CMD", "IB_EMPTY"/* NV50+ */, "MEM_FAULT", "UNK"
+       };
+       return desc[(state >> 29) & 0x7];
+}
+
+static bool
+nv04_fifo_swmthd(struct nv04_fifo_priv *priv, u32 chid, u32 addr, u32 data)
+{
+       struct nv04_fifo_chan *chan = NULL;
+       struct nvkm_handle *bind;
+       const int subc = (addr >> 13) & 0x7;
+       const int mthd = addr & 0x1ffc;
+       bool handled = false;
+       unsigned long flags;
+       u32 engine;
+
+       spin_lock_irqsave(&priv->base.lock, flags);
+       if (likely(chid >= priv->base.min && chid <= priv->base.max))
+               chan = (void *)priv->base.channel[chid];
+       if (unlikely(!chan))
+               goto out;
+
+       switch (mthd) {
+       case 0x0000:
+               bind = nvkm_namedb_get(nv_namedb(chan), data);
+               if (unlikely(!bind))
+                       break;
+
+               if (nv_engidx(bind->object->engine) == NVDEV_ENGINE_SW) {
+                       engine = 0x0000000f << (subc * 4);
+                       chan->subc[subc] = data;
+                       handled = true;
+
+                       nv_mask(priv, NV04_PFIFO_CACHE1_ENGINE, engine, 0);
+               }
+
+               nvkm_namedb_put(bind);
+               break;
+       default:
+               engine = nv_rd32(priv, NV04_PFIFO_CACHE1_ENGINE);
+               if (unlikely(((engine >> (subc * 4)) & 0xf) != 0))
+                       break;
+
+               bind = nvkm_namedb_get(nv_namedb(chan), chan->subc[subc]);
+               if (likely(bind)) {
+                       if (!nv_call(bind->object, mthd, data))
+                               handled = true;
+                       nvkm_namedb_put(bind);
+               }
+               break;
+       }
+
+out:
+       spin_unlock_irqrestore(&priv->base.lock, flags);
+       return handled;
+}
+
+static void
+nv04_fifo_cache_error(struct nvkm_device *device,
+                     struct nv04_fifo_priv *priv, u32 chid, u32 get)
+{
+       u32 mthd, data;
+       int ptr;
+
+       /* NV_PFIFO_CACHE1_GET actually goes to 0xffc before wrapping on my
+        * G80 chips, but CACHE1 isn't big enough for this much data.. Tests
+        * show that it wraps around to the start at GET=0x800.. No clue as to
+        * why..
+        */
+       ptr = (get & 0x7ff) >> 2;
+
+       if (device->card_type < NV_40) {
+               mthd = nv_rd32(priv, NV04_PFIFO_CACHE1_METHOD(ptr));
+               data = nv_rd32(priv, NV04_PFIFO_CACHE1_DATA(ptr));
+       } else {
+               mthd = nv_rd32(priv, NV40_PFIFO_CACHE1_METHOD(ptr));
+               data = nv_rd32(priv, NV40_PFIFO_CACHE1_DATA(ptr));
+       }
+
+       if (!nv04_fifo_swmthd(priv, chid, mthd, data)) {
+               const char *client_name =
+                       nvkm_client_name_for_fifo_chid(&priv->base, chid);
+               nv_error(priv,
+                        "CACHE_ERROR - ch %d [%s] subc %d mthd 0x%04x data 0x%08x\n",
+                        chid, client_name, (mthd >> 13) & 7, mthd & 0x1ffc,
+                        data);
+       }
+
+       nv_wr32(priv, NV04_PFIFO_CACHE1_DMA_PUSH, 0);
+       nv_wr32(priv, NV03_PFIFO_INTR_0, NV_PFIFO_INTR_CACHE_ERROR);
+
+       nv_wr32(priv, NV03_PFIFO_CACHE1_PUSH0,
+               nv_rd32(priv, NV03_PFIFO_CACHE1_PUSH0) & ~1);
+       nv_wr32(priv, NV03_PFIFO_CACHE1_GET, get + 4);
+       nv_wr32(priv, NV03_PFIFO_CACHE1_PUSH0,
+               nv_rd32(priv, NV03_PFIFO_CACHE1_PUSH0) | 1);
+       nv_wr32(priv, NV04_PFIFO_CACHE1_HASH, 0);
+
+       nv_wr32(priv, NV04_PFIFO_CACHE1_DMA_PUSH,
+               nv_rd32(priv, NV04_PFIFO_CACHE1_DMA_PUSH) | 1);
+       nv_wr32(priv, NV04_PFIFO_CACHE1_PULL0, 1);
+}
+
+static void
+nv04_fifo_dma_pusher(struct nvkm_device *device,
+                    struct nv04_fifo_priv *priv, u32 chid)
+{
+       const char *client_name;
+       u32 dma_get = nv_rd32(priv, 0x003244);
+       u32 dma_put = nv_rd32(priv, 0x003240);
+       u32 push = nv_rd32(priv, 0x003220);
+       u32 state = nv_rd32(priv, 0x003228);
+
+       client_name = nvkm_client_name_for_fifo_chid(&priv->base, chid);
+
+       if (device->card_type == NV_50) {
+               u32 ho_get = nv_rd32(priv, 0x003328);
+               u32 ho_put = nv_rd32(priv, 0x003320);
+               u32 ib_get = nv_rd32(priv, 0x003334);
+               u32 ib_put = nv_rd32(priv, 0x003330);
+
+               nv_error(priv,
+                        "DMA_PUSHER - ch %d [%s] get 0x%02x%08x put 0x%02x%08x ib_get 0x%08x ib_put 0x%08x state 0x%08x (err: %s) push 0x%08x\n",
+                        chid, client_name, ho_get, dma_get, ho_put, dma_put,
+                        ib_get, ib_put, state, nv_dma_state_err(state), push);
+
+               /* METHOD_COUNT, in DMA_STATE on earlier chipsets */
+               nv_wr32(priv, 0x003364, 0x00000000);
+               if (dma_get != dma_put || ho_get != ho_put) {
+                       nv_wr32(priv, 0x003244, dma_put);
+                       nv_wr32(priv, 0x003328, ho_put);
+               } else
+               if (ib_get != ib_put)
+                       nv_wr32(priv, 0x003334, ib_put);
+       } else {
+               nv_error(priv,
+                        "DMA_PUSHER - ch %d [%s] get 0x%08x put 0x%08x state 0x%08x (err: %s) push 0x%08x\n",
+                        chid, client_name, dma_get, dma_put, state,
+                        nv_dma_state_err(state), push);
+
+               if (dma_get != dma_put)
+                       nv_wr32(priv, 0x003244, dma_put);
+       }
+
+       nv_wr32(priv, 0x003228, 0x00000000);
+       nv_wr32(priv, 0x003220, 0x00000001);
+       nv_wr32(priv, 0x002100, NV_PFIFO_INTR_DMA_PUSHER);
+}
+
+void
+nv04_fifo_intr(struct nvkm_subdev *subdev)
+{
+       struct nvkm_device *device = nv_device(subdev);
+       struct nv04_fifo_priv *priv = (void *)subdev;
+       u32 mask = nv_rd32(priv, NV03_PFIFO_INTR_EN_0);
+       u32 stat = nv_rd32(priv, NV03_PFIFO_INTR_0) & mask;
+       u32 reassign, chid, get, sem;
+
+       reassign = nv_rd32(priv, NV03_PFIFO_CACHES) & 1;
+       nv_wr32(priv, NV03_PFIFO_CACHES, 0);
+
+       chid = nv_rd32(priv, NV03_PFIFO_CACHE1_PUSH1) & priv->base.max;
+       get  = nv_rd32(priv, NV03_PFIFO_CACHE1_GET);
+
+       if (stat & NV_PFIFO_INTR_CACHE_ERROR) {
+               nv04_fifo_cache_error(device, priv, chid, get);
+               stat &= ~NV_PFIFO_INTR_CACHE_ERROR;
+       }
+
+       if (stat & NV_PFIFO_INTR_DMA_PUSHER) {
+               nv04_fifo_dma_pusher(device, priv, chid);
+               stat &= ~NV_PFIFO_INTR_DMA_PUSHER;
+       }
+
+       if (stat & NV_PFIFO_INTR_SEMAPHORE) {
+               stat &= ~NV_PFIFO_INTR_SEMAPHORE;
+               nv_wr32(priv, NV03_PFIFO_INTR_0, NV_PFIFO_INTR_SEMAPHORE);
+
+               sem = nv_rd32(priv, NV10_PFIFO_CACHE1_SEMAPHORE);
+               nv_wr32(priv, NV10_PFIFO_CACHE1_SEMAPHORE, sem | 0x1);
+
+               nv_wr32(priv, NV03_PFIFO_CACHE1_GET, get + 4);
+               nv_wr32(priv, NV04_PFIFO_CACHE1_PULL0, 1);
+       }
+
+       if (device->card_type == NV_50) {
+               if (stat & 0x00000010) {
+                       stat &= ~0x00000010;
+                       nv_wr32(priv, 0x002100, 0x00000010);
+               }
+
+               if (stat & 0x40000000) {
+                       nv_wr32(priv, 0x002100, 0x40000000);
+                       nvkm_fifo_uevent(&priv->base);
+                       stat &= ~0x40000000;
+               }
+       }
+
+       if (stat) {
+               nv_warn(priv, "unknown intr 0x%08x\n", stat);
+               nv_mask(priv, NV03_PFIFO_INTR_EN_0, stat, 0x00000000);
+               nv_wr32(priv, NV03_PFIFO_INTR_0, stat);
+       }
+
+       nv_wr32(priv, NV03_PFIFO_CACHES, reassign);
+}
+
+static int
+nv04_fifo_ctor(struct nvkm_object *parent, struct nvkm_object *engine,
+              struct nvkm_oclass *oclass, void *data, u32 size,
+              struct nvkm_object **pobject)
+{
+       struct nv04_instmem_priv *imem = nv04_instmem(parent);
+       struct nv04_fifo_priv *priv;
+       int ret;
+
+       ret = nvkm_fifo_create(parent, engine, oclass, 0, 15, &priv);
+       *pobject = nv_object(priv);
+       if (ret)
+               return ret;
+
+       nvkm_ramht_ref(imem->ramht, &priv->ramht);
+       nvkm_gpuobj_ref(imem->ramro, &priv->ramro);
+       nvkm_gpuobj_ref(imem->ramfc, &priv->ramfc);
+
+       nv_subdev(priv)->unit = 0x00000100;
+       nv_subdev(priv)->intr = nv04_fifo_intr;
+       nv_engine(priv)->cclass = &nv04_fifo_cclass;
+       nv_engine(priv)->sclass = nv04_fifo_sclass;
+       priv->base.pause = nv04_fifo_pause;
+       priv->base.start = nv04_fifo_start;
+       priv->ramfc_desc = nv04_ramfc;
+       return 0;
+}
+
+void
+nv04_fifo_dtor(struct nvkm_object *object)
+{
+       struct nv04_fifo_priv *priv = (void *)object;
+       nvkm_gpuobj_ref(NULL, &priv->ramfc);
+       nvkm_gpuobj_ref(NULL, &priv->ramro);
+       nvkm_ramht_ref(NULL, &priv->ramht);
+       nvkm_fifo_destroy(&priv->base);
+}
+
+int
+nv04_fifo_init(struct nvkm_object *object)
+{
+       struct nv04_fifo_priv *priv = (void *)object;
+       int ret;
+
+       ret = nvkm_fifo_init(&priv->base);
+       if (ret)
+               return ret;
+
+       nv_wr32(priv, NV04_PFIFO_DELAY_0, 0x000000ff);
+       nv_wr32(priv, NV04_PFIFO_DMA_TIMESLICE, 0x0101ffff);
+
+       nv_wr32(priv, NV03_PFIFO_RAMHT, (0x03 << 24) /* search 128 */ |
+                                      ((priv->ramht->bits - 9) << 16) |
+                                       (priv->ramht->gpuobj.addr >> 8));
+       nv_wr32(priv, NV03_PFIFO_RAMRO, priv->ramro->addr >> 8);
+       nv_wr32(priv, NV03_PFIFO_RAMFC, priv->ramfc->addr >> 8);
+
+       nv_wr32(priv, NV03_PFIFO_CACHE1_PUSH1, priv->base.max);
+
+       nv_wr32(priv, NV03_PFIFO_INTR_0, 0xffffffff);
+       nv_wr32(priv, NV03_PFIFO_INTR_EN_0, 0xffffffff);
+
+       nv_wr32(priv, NV03_PFIFO_CACHE1_PUSH0, 1);
+       nv_wr32(priv, NV04_PFIFO_CACHE1_PULL0, 1);
+       nv_wr32(priv, NV03_PFIFO_CACHES, 1);
+       return 0;
+}
+
+struct nvkm_oclass *
+nv04_fifo_oclass = &(struct nvkm_oclass) {
+       .handle = NV_ENGINE(FIFO, 0x04),
+       .ofuncs = &(struct nvkm_ofuncs) {
+               .ctor = nv04_fifo_ctor,
+               .dtor = nv04_fifo_dtor,
+               .init = nv04_fifo_init,
+               .fini = _nvkm_fifo_fini,
+       },
+};