drm/nouveau: Add DRM driver for NVIDIA GPUs
[linux-2.6/linux-2.6-openrd.git] / drivers / gpu / drm / nouveau / nouveau_object.c
blob93379bb81beae845a0a91a769537cd90b790d87c
1 /*
2 * Copyright (C) 2006 Ben Skeggs.
4 * All Rights Reserved.
6 * Permission is hereby granted, free of charge, to any person obtaining
7 * a copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sublicense, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial
16 * portions of the Software.
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
19 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
21 * IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
22 * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
23 * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
24 * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
29 * Authors:
30 * Ben Skeggs <darktama@iinet.net.au>
33 #include "drmP.h"
34 #include "drm.h"
35 #include "nouveau_drv.h"
36 #include "nouveau_drm.h"
38 /* NVidia uses context objects to drive drawing operations.
40 Context objects can be selected into 8 subchannels in the FIFO,
41 and then used via DMA command buffers.
43 A context object is referenced by a user defined handle (CARD32). The HW
44 looks up graphics objects in a hash table in the instance RAM.
46 An entry in the hash table consists of 2 CARD32. The first CARD32 contains
47 the handle, the second one a bitfield, that contains the address of the
48 object in instance RAM.
50 The format of the second CARD32 seems to be:
52 NV4 to NV30:
54 15: 0 instance_addr >> 4
55 17:16 engine (here uses 1 = graphics)
56 28:24 channel id (here uses 0)
57 31 valid (use 1)
59 NV40:
61 15: 0 instance_addr >> 4 (maybe 19-0)
62 21:20 engine (here uses 1 = graphics)
63 I'm unsure about the other bits, but using 0 seems to work.
65 The key into the hash table depends on the object handle and channel id and
66 is given as:
68 static uint32_t
69 nouveau_ramht_hash_handle(struct drm_device *dev, int channel, uint32_t handle)
71 struct drm_nouveau_private *dev_priv = dev->dev_private;
72 uint32_t hash = 0;
73 int i;
75 NV_DEBUG(dev, "ch%d handle=0x%08x\n", channel, handle);
77 for (i = 32; i > 0; i -= dev_priv->ramht_bits) {
78 hash ^= (handle & ((1 << dev_priv->ramht_bits) - 1));
79 handle >>= dev_priv->ramht_bits;
82 if (dev_priv->card_type < NV_50)
83 hash ^= channel << (dev_priv->ramht_bits - 4);
84 hash <<= 3;
86 NV_DEBUG(dev, "hash=0x%08x\n", hash);
87 return hash;
90 static int
91 nouveau_ramht_entry_valid(struct drm_device *dev, struct nouveau_gpuobj *ramht,
92 uint32_t offset)
94 struct drm_nouveau_private *dev_priv = dev->dev_private;
95 uint32_t ctx = nv_ro32(dev, ramht, (offset + 4)/4);
97 if (dev_priv->card_type < NV_40)
98 return ((ctx & NV_RAMHT_CONTEXT_VALID) != 0);
99 return (ctx != 0);
102 static int
103 nouveau_ramht_insert(struct drm_device *dev, struct nouveau_gpuobj_ref *ref)
105 struct drm_nouveau_private *dev_priv = dev->dev_private;
106 struct nouveau_instmem_engine *instmem = &dev_priv->engine.instmem;
107 struct nouveau_channel *chan = ref->channel;
108 struct nouveau_gpuobj *ramht = chan->ramht ? chan->ramht->gpuobj : NULL;
109 uint32_t ctx, co, ho;
111 if (!ramht) {
112 NV_ERROR(dev, "No hash table!\n");
113 return -EINVAL;
116 if (dev_priv->card_type < NV_40) {
117 ctx = NV_RAMHT_CONTEXT_VALID | (ref->instance >> 4) |
118 (chan->id << NV_RAMHT_CONTEXT_CHANNEL_SHIFT) |
119 (ref->gpuobj->engine << NV_RAMHT_CONTEXT_ENGINE_SHIFT);
120 } else
121 if (dev_priv->card_type < NV_50) {
122 ctx = (ref->instance >> 4) |
123 (chan->id << NV40_RAMHT_CONTEXT_CHANNEL_SHIFT) |
124 (ref->gpuobj->engine << NV40_RAMHT_CONTEXT_ENGINE_SHIFT);
125 } else {
126 if (ref->gpuobj->engine == NVOBJ_ENGINE_DISPLAY) {
127 ctx = (ref->instance << 10) | 2;
128 } else {
129 ctx = (ref->instance >> 4) |
130 ((ref->gpuobj->engine <<
131 NV40_RAMHT_CONTEXT_ENGINE_SHIFT));
135 instmem->prepare_access(dev, true);
136 co = ho = nouveau_ramht_hash_handle(dev, chan->id, ref->handle);
137 do {
138 if (!nouveau_ramht_entry_valid(dev, ramht, co)) {
139 NV_DEBUG(dev,
140 "insert ch%d 0x%08x: h=0x%08x, c=0x%08x\n",
141 chan->id, co, ref->handle, ctx);
142 nv_wo32(dev, ramht, (co + 0)/4, ref->handle);
143 nv_wo32(dev, ramht, (co + 4)/4, ctx);
145 list_add_tail(&ref->list, &chan->ramht_refs);
146 instmem->finish_access(dev);
147 return 0;
149 NV_DEBUG(dev, "collision ch%d 0x%08x: h=0x%08x\n",
150 chan->id, co, nv_ro32(dev, ramht, co/4));
152 co += 8;
153 if (co >= dev_priv->ramht_size)
154 co = 0;
155 } while (co != ho);
156 instmem->finish_access(dev);
158 NV_ERROR(dev, "RAMHT space exhausted. ch=%d\n", chan->id);
159 return -ENOMEM;
162 static void
163 nouveau_ramht_remove(struct drm_device *dev, struct nouveau_gpuobj_ref *ref)
165 struct drm_nouveau_private *dev_priv = dev->dev_private;
166 struct nouveau_instmem_engine *instmem = &dev_priv->engine.instmem;
167 struct nouveau_channel *chan = ref->channel;
168 struct nouveau_gpuobj *ramht = chan->ramht ? chan->ramht->gpuobj : NULL;
169 uint32_t co, ho;
171 if (!ramht) {
172 NV_ERROR(dev, "No hash table!\n");
173 return;
176 instmem->prepare_access(dev, true);
177 co = ho = nouveau_ramht_hash_handle(dev, chan->id, ref->handle);
178 do {
179 if (nouveau_ramht_entry_valid(dev, ramht, co) &&
180 (ref->handle == nv_ro32(dev, ramht, (co/4)))) {
181 NV_DEBUG(dev,
182 "remove ch%d 0x%08x: h=0x%08x, c=0x%08x\n",
183 chan->id, co, ref->handle,
184 nv_ro32(dev, ramht, (co + 4)));
185 nv_wo32(dev, ramht, (co + 0)/4, 0x00000000);
186 nv_wo32(dev, ramht, (co + 4)/4, 0x00000000);
188 list_del(&ref->list);
189 instmem->finish_access(dev);
190 return;
193 co += 8;
194 if (co >= dev_priv->ramht_size)
195 co = 0;
196 } while (co != ho);
197 list_del(&ref->list);
198 instmem->finish_access(dev);
200 NV_ERROR(dev, "RAMHT entry not found. ch=%d, handle=0x%08x\n",
201 chan->id, ref->handle);
205 nouveau_gpuobj_new(struct drm_device *dev, struct nouveau_channel *chan,
206 uint32_t size, int align, uint32_t flags,
207 struct nouveau_gpuobj **gpuobj_ret)
209 struct drm_nouveau_private *dev_priv = dev->dev_private;
210 struct nouveau_engine *engine = &dev_priv->engine;
211 struct nouveau_gpuobj *gpuobj;
212 struct mem_block *pramin = NULL;
213 int ret;
215 NV_DEBUG(dev, "ch%d size=%u align=%d flags=0x%08x\n",
216 chan ? chan->id : -1, size, align, flags);
218 if (!dev_priv || !gpuobj_ret || *gpuobj_ret != NULL)
219 return -EINVAL;
221 gpuobj = kzalloc(sizeof(*gpuobj), GFP_KERNEL);
222 if (!gpuobj)
223 return -ENOMEM;
224 NV_DEBUG(dev, "gpuobj %p\n", gpuobj);
225 gpuobj->flags = flags;
226 gpuobj->im_channel = chan;
228 list_add_tail(&gpuobj->list, &dev_priv->gpuobj_list);
230 /* Choose between global instmem heap, and per-channel private
231 * instmem heap. On <NV50 allow requests for private instmem
232 * to be satisfied from global heap if no per-channel area
233 * available.
235 if (chan) {
236 if (chan->ramin_heap) {
237 NV_DEBUG(dev, "private heap\n");
238 pramin = chan->ramin_heap;
239 } else
240 if (dev_priv->card_type < NV_50) {
241 NV_DEBUG(dev, "global heap fallback\n");
242 pramin = dev_priv->ramin_heap;
244 } else {
245 NV_DEBUG(dev, "global heap\n");
246 pramin = dev_priv->ramin_heap;
249 if (!pramin) {
250 NV_ERROR(dev, "No PRAMIN heap!\n");
251 return -EINVAL;
254 if (!chan) {
255 ret = engine->instmem.populate(dev, gpuobj, &size);
256 if (ret) {
257 nouveau_gpuobj_del(dev, &gpuobj);
258 return ret;
262 /* Allocate a chunk of the PRAMIN aperture */
263 gpuobj->im_pramin = nouveau_mem_alloc_block(pramin, size,
264 drm_order(align),
265 (struct drm_file *)-2, 0);
266 if (!gpuobj->im_pramin) {
267 nouveau_gpuobj_del(dev, &gpuobj);
268 return -ENOMEM;
271 if (!chan) {
272 ret = engine->instmem.bind(dev, gpuobj);
273 if (ret) {
274 nouveau_gpuobj_del(dev, &gpuobj);
275 return ret;
279 if (gpuobj->flags & NVOBJ_FLAG_ZERO_ALLOC) {
280 int i;
282 engine->instmem.prepare_access(dev, true);
283 for (i = 0; i < gpuobj->im_pramin->size; i += 4)
284 nv_wo32(dev, gpuobj, i/4, 0);
285 engine->instmem.finish_access(dev);
288 *gpuobj_ret = gpuobj;
289 return 0;
293 nouveau_gpuobj_early_init(struct drm_device *dev)
295 struct drm_nouveau_private *dev_priv = dev->dev_private;
297 NV_DEBUG(dev, "\n");
299 INIT_LIST_HEAD(&dev_priv->gpuobj_list);
301 return 0;
305 nouveau_gpuobj_init(struct drm_device *dev)
307 struct drm_nouveau_private *dev_priv = dev->dev_private;
308 int ret;
310 NV_DEBUG(dev, "\n");
312 if (dev_priv->card_type < NV_50) {
313 ret = nouveau_gpuobj_new_fake(dev,
314 dev_priv->ramht_offset, ~0, dev_priv->ramht_size,
315 NVOBJ_FLAG_ZERO_ALLOC | NVOBJ_FLAG_ALLOW_NO_REFS,
316 &dev_priv->ramht, NULL);
317 if (ret)
318 return ret;
321 return 0;
324 void
325 nouveau_gpuobj_takedown(struct drm_device *dev)
327 struct drm_nouveau_private *dev_priv = dev->dev_private;
329 NV_DEBUG(dev, "\n");
331 nouveau_gpuobj_del(dev, &dev_priv->ramht);
334 void
335 nouveau_gpuobj_late_takedown(struct drm_device *dev)
337 struct drm_nouveau_private *dev_priv = dev->dev_private;
338 struct nouveau_gpuobj *gpuobj = NULL;
339 struct list_head *entry, *tmp;
341 NV_DEBUG(dev, "\n");
343 list_for_each_safe(entry, tmp, &dev_priv->gpuobj_list) {
344 gpuobj = list_entry(entry, struct nouveau_gpuobj, list);
346 NV_ERROR(dev, "gpuobj %p still exists at takedown, refs=%d\n",
347 gpuobj, gpuobj->refcount);
348 gpuobj->refcount = 0;
349 nouveau_gpuobj_del(dev, &gpuobj);
354 nouveau_gpuobj_del(struct drm_device *dev, struct nouveau_gpuobj **pgpuobj)
356 struct drm_nouveau_private *dev_priv = dev->dev_private;
357 struct nouveau_engine *engine = &dev_priv->engine;
358 struct nouveau_gpuobj *gpuobj;
359 int i;
361 NV_DEBUG(dev, "gpuobj %p\n", pgpuobj ? *pgpuobj : NULL);
363 if (!dev_priv || !pgpuobj || !(*pgpuobj))
364 return -EINVAL;
365 gpuobj = *pgpuobj;
367 if (gpuobj->refcount != 0) {
368 NV_ERROR(dev, "gpuobj refcount is %d\n", gpuobj->refcount);
369 return -EINVAL;
372 if (gpuobj->im_pramin && (gpuobj->flags & NVOBJ_FLAG_ZERO_FREE)) {
373 engine->instmem.prepare_access(dev, true);
374 for (i = 0; i < gpuobj->im_pramin->size; i += 4)
375 nv_wo32(dev, gpuobj, i/4, 0);
376 engine->instmem.finish_access(dev);
379 if (gpuobj->dtor)
380 gpuobj->dtor(dev, gpuobj);
382 if (gpuobj->im_backing && !(gpuobj->flags & NVOBJ_FLAG_FAKE))
383 engine->instmem.clear(dev, gpuobj);
385 if (gpuobj->im_pramin) {
386 if (gpuobj->flags & NVOBJ_FLAG_FAKE)
387 kfree(gpuobj->im_pramin);
388 else
389 nouveau_mem_free_block(gpuobj->im_pramin);
392 list_del(&gpuobj->list);
394 *pgpuobj = NULL;
395 kfree(gpuobj);
396 return 0;
399 static int
400 nouveau_gpuobj_instance_get(struct drm_device *dev,
401 struct nouveau_channel *chan,
402 struct nouveau_gpuobj *gpuobj, uint32_t *inst)
404 struct drm_nouveau_private *dev_priv = dev->dev_private;
405 struct nouveau_gpuobj *cpramin;
407 /* <NV50 use PRAMIN address everywhere */
408 if (dev_priv->card_type < NV_50) {
409 *inst = gpuobj->im_pramin->start;
410 return 0;
413 if (chan && gpuobj->im_channel != chan) {
414 NV_ERROR(dev, "Channel mismatch: obj %d, ref %d\n",
415 gpuobj->im_channel->id, chan->id);
416 return -EINVAL;
419 /* NV50 channel-local instance */
420 if (chan) {
421 cpramin = chan->ramin->gpuobj;
422 *inst = gpuobj->im_pramin->start - cpramin->im_pramin->start;
423 return 0;
426 /* NV50 global (VRAM) instance */
427 if (!gpuobj->im_channel) {
428 /* ...from global heap */
429 if (!gpuobj->im_backing) {
430 NV_ERROR(dev, "AII, no VRAM backing gpuobj\n");
431 return -EINVAL;
433 *inst = gpuobj->im_backing_start;
434 return 0;
435 } else {
436 /* ...from local heap */
437 cpramin = gpuobj->im_channel->ramin->gpuobj;
438 *inst = cpramin->im_backing_start +
439 (gpuobj->im_pramin->start - cpramin->im_pramin->start);
440 return 0;
443 return -EINVAL;
447 nouveau_gpuobj_ref_add(struct drm_device *dev, struct nouveau_channel *chan,
448 uint32_t handle, struct nouveau_gpuobj *gpuobj,
449 struct nouveau_gpuobj_ref **ref_ret)
451 struct drm_nouveau_private *dev_priv = dev->dev_private;
452 struct nouveau_gpuobj_ref *ref;
453 uint32_t instance;
454 int ret;
456 NV_DEBUG(dev, "ch%d h=0x%08x gpuobj=%p\n",
457 chan ? chan->id : -1, handle, gpuobj);
459 if (!dev_priv || !gpuobj || (ref_ret && *ref_ret != NULL))
460 return -EINVAL;
462 if (!chan && !ref_ret)
463 return -EINVAL;
465 if (gpuobj->engine == NVOBJ_ENGINE_SW && !gpuobj->im_pramin) {
466 /* sw object */
467 instance = 0x40;
468 } else {
469 ret = nouveau_gpuobj_instance_get(dev, chan, gpuobj, &instance);
470 if (ret)
471 return ret;
474 ref = kzalloc(sizeof(*ref), GFP_KERNEL);
475 if (!ref)
476 return -ENOMEM;
477 INIT_LIST_HEAD(&ref->list);
478 ref->gpuobj = gpuobj;
479 ref->channel = chan;
480 ref->instance = instance;
482 if (!ref_ret) {
483 ref->handle = handle;
485 ret = nouveau_ramht_insert(dev, ref);
486 if (ret) {
487 kfree(ref);
488 return ret;
490 } else {
491 ref->handle = ~0;
492 *ref_ret = ref;
495 ref->gpuobj->refcount++;
496 return 0;
499 int nouveau_gpuobj_ref_del(struct drm_device *dev, struct nouveau_gpuobj_ref **pref)
501 struct nouveau_gpuobj_ref *ref;
503 NV_DEBUG(dev, "ref %p\n", pref ? *pref : NULL);
505 if (!dev || !pref || *pref == NULL)
506 return -EINVAL;
507 ref = *pref;
509 if (ref->handle != ~0)
510 nouveau_ramht_remove(dev, ref);
512 if (ref->gpuobj) {
513 ref->gpuobj->refcount--;
515 if (ref->gpuobj->refcount == 0) {
516 if (!(ref->gpuobj->flags & NVOBJ_FLAG_ALLOW_NO_REFS))
517 nouveau_gpuobj_del(dev, &ref->gpuobj);
521 *pref = NULL;
522 kfree(ref);
523 return 0;
527 nouveau_gpuobj_new_ref(struct drm_device *dev,
528 struct nouveau_channel *oc, struct nouveau_channel *rc,
529 uint32_t handle, uint32_t size, int align,
530 uint32_t flags, struct nouveau_gpuobj_ref **ref)
532 struct nouveau_gpuobj *gpuobj = NULL;
533 int ret;
535 ret = nouveau_gpuobj_new(dev, oc, size, align, flags, &gpuobj);
536 if (ret)
537 return ret;
539 ret = nouveau_gpuobj_ref_add(dev, rc, handle, gpuobj, ref);
540 if (ret) {
541 nouveau_gpuobj_del(dev, &gpuobj);
542 return ret;
545 return 0;
549 nouveau_gpuobj_ref_find(struct nouveau_channel *chan, uint32_t handle,
550 struct nouveau_gpuobj_ref **ref_ret)
552 struct nouveau_gpuobj_ref *ref;
553 struct list_head *entry, *tmp;
555 list_for_each_safe(entry, tmp, &chan->ramht_refs) {
556 ref = list_entry(entry, struct nouveau_gpuobj_ref, list);
558 if (ref->handle == handle) {
559 if (ref_ret)
560 *ref_ret = ref;
561 return 0;
565 return -EINVAL;
569 nouveau_gpuobj_new_fake(struct drm_device *dev, uint32_t p_offset,
570 uint32_t b_offset, uint32_t size,
571 uint32_t flags, struct nouveau_gpuobj **pgpuobj,
572 struct nouveau_gpuobj_ref **pref)
574 struct drm_nouveau_private *dev_priv = dev->dev_private;
575 struct nouveau_gpuobj *gpuobj = NULL;
576 int i;
578 NV_DEBUG(dev,
579 "p_offset=0x%08x b_offset=0x%08x size=0x%08x flags=0x%08x\n",
580 p_offset, b_offset, size, flags);
582 gpuobj = kzalloc(sizeof(*gpuobj), GFP_KERNEL);
583 if (!gpuobj)
584 return -ENOMEM;
585 NV_DEBUG(dev, "gpuobj %p\n", gpuobj);
586 gpuobj->im_channel = NULL;
587 gpuobj->flags = flags | NVOBJ_FLAG_FAKE;
589 list_add_tail(&gpuobj->list, &dev_priv->gpuobj_list);
591 if (p_offset != ~0) {
592 gpuobj->im_pramin = kzalloc(sizeof(struct mem_block),
593 GFP_KERNEL);
594 if (!gpuobj->im_pramin) {
595 nouveau_gpuobj_del(dev, &gpuobj);
596 return -ENOMEM;
598 gpuobj->im_pramin->start = p_offset;
599 gpuobj->im_pramin->size = size;
602 if (b_offset != ~0) {
603 gpuobj->im_backing = (struct nouveau_bo *)-1;
604 gpuobj->im_backing_start = b_offset;
607 if (gpuobj->flags & NVOBJ_FLAG_ZERO_ALLOC) {
608 dev_priv->engine.instmem.prepare_access(dev, true);
609 for (i = 0; i < gpuobj->im_pramin->size; i += 4)
610 nv_wo32(dev, gpuobj, i/4, 0);
611 dev_priv->engine.instmem.finish_access(dev);
614 if (pref) {
615 i = nouveau_gpuobj_ref_add(dev, NULL, 0, gpuobj, pref);
616 if (i) {
617 nouveau_gpuobj_del(dev, &gpuobj);
618 return i;
622 if (pgpuobj)
623 *pgpuobj = gpuobj;
624 return 0;
628 static uint32_t
629 nouveau_gpuobj_class_instmem_size(struct drm_device *dev, int class)
631 struct drm_nouveau_private *dev_priv = dev->dev_private;
633 /*XXX: dodgy hack for now */
634 if (dev_priv->card_type >= NV_50)
635 return 24;
636 if (dev_priv->card_type >= NV_40)
637 return 32;
638 return 16;
642 DMA objects are used to reference a piece of memory in the
643 framebuffer, PCI or AGP address space. Each object is 16 bytes big
644 and looks as follows:
646 entry[0]
647 11:0 class (seems like I can always use 0 here)
648 12 page table present?
649 13 page entry linear?
650 15:14 access: 0 rw, 1 ro, 2 wo
651 17:16 target: 0 NV memory, 1 NV memory tiled, 2 PCI, 3 AGP
652 31:20 dma adjust (bits 0-11 of the address)
653 entry[1]
654 dma limit (size of transfer)
655 entry[X]
656 1 0 readonly, 1 readwrite
657 31:12 dma frame address of the page (bits 12-31 of the address)
658 entry[N]
659 page table terminator, same value as the first pte, as does nvidia
660 rivatv uses 0xffffffff
662 Non linear page tables need a list of frame addresses afterwards,
663 the rivatv project has some info on this.
665 The method below creates a DMA object in instance RAM and returns a handle
666 to it that can be used to set up context objects.
669 nouveau_gpuobj_dma_new(struct nouveau_channel *chan, int class,
670 uint64_t offset, uint64_t size, int access,
671 int target, struct nouveau_gpuobj **gpuobj)
673 struct drm_device *dev = chan->dev;
674 struct drm_nouveau_private *dev_priv = dev->dev_private;
675 struct nouveau_instmem_engine *instmem = &dev_priv->engine.instmem;
676 int ret;
678 NV_DEBUG(dev, "ch%d class=0x%04x offset=0x%llx size=0x%llx\n",
679 chan->id, class, offset, size);
680 NV_DEBUG(dev, "access=%d target=%d\n", access, target);
682 switch (target) {
683 case NV_DMA_TARGET_AGP:
684 offset += dev_priv->gart_info.aper_base;
685 break;
686 default:
687 break;
690 ret = nouveau_gpuobj_new(dev, chan,
691 nouveau_gpuobj_class_instmem_size(dev, class),
692 16, NVOBJ_FLAG_ZERO_ALLOC |
693 NVOBJ_FLAG_ZERO_FREE, gpuobj);
694 if (ret) {
695 NV_ERROR(dev, "Error creating gpuobj: %d\n", ret);
696 return ret;
699 instmem->prepare_access(dev, true);
701 if (dev_priv->card_type < NV_50) {
702 uint32_t frame, adjust, pte_flags = 0;
704 if (access != NV_DMA_ACCESS_RO)
705 pte_flags |= (1<<1);
706 adjust = offset & 0x00000fff;
707 frame = offset & ~0x00000fff;
709 nv_wo32(dev, *gpuobj, 0, ((1<<12) | (1<<13) |
710 (adjust << 20) |
711 (access << 14) |
712 (target << 16) |
713 class));
714 nv_wo32(dev, *gpuobj, 1, size - 1);
715 nv_wo32(dev, *gpuobj, 2, frame | pte_flags);
716 nv_wo32(dev, *gpuobj, 3, frame | pte_flags);
717 } else {
718 uint64_t limit = offset + size - 1;
719 uint32_t flags0, flags5;
721 if (target == NV_DMA_TARGET_VIDMEM) {
722 flags0 = 0x00190000;
723 flags5 = 0x00010000;
724 } else {
725 flags0 = 0x7fc00000;
726 flags5 = 0x00080000;
729 nv_wo32(dev, *gpuobj, 0, flags0 | class);
730 nv_wo32(dev, *gpuobj, 1, lower_32_bits(limit));
731 nv_wo32(dev, *gpuobj, 2, lower_32_bits(offset));
732 nv_wo32(dev, *gpuobj, 3, ((upper_32_bits(limit) & 0xff) << 24) |
733 (upper_32_bits(offset) & 0xff));
734 nv_wo32(dev, *gpuobj, 5, flags5);
737 instmem->finish_access(dev);
739 (*gpuobj)->engine = NVOBJ_ENGINE_SW;
740 (*gpuobj)->class = class;
741 return 0;
745 nouveau_gpuobj_gart_dma_new(struct nouveau_channel *chan,
746 uint64_t offset, uint64_t size, int access,
747 struct nouveau_gpuobj **gpuobj,
748 uint32_t *o_ret)
750 struct drm_device *dev = chan->dev;
751 struct drm_nouveau_private *dev_priv = dev->dev_private;
752 int ret;
754 if (dev_priv->gart_info.type == NOUVEAU_GART_AGP ||
755 (dev_priv->card_type >= NV_50 &&
756 dev_priv->gart_info.type == NOUVEAU_GART_SGDMA)) {
757 ret = nouveau_gpuobj_dma_new(chan, NV_CLASS_DMA_IN_MEMORY,
758 offset + dev_priv->vm_gart_base,
759 size, access, NV_DMA_TARGET_AGP,
760 gpuobj);
761 if (o_ret)
762 *o_ret = 0;
763 } else
764 if (dev_priv->gart_info.type == NOUVEAU_GART_SGDMA) {
765 *gpuobj = dev_priv->gart_info.sg_ctxdma;
766 if (offset & ~0xffffffffULL) {
767 NV_ERROR(dev, "obj offset exceeds 32-bits\n");
768 return -EINVAL;
770 if (o_ret)
771 *o_ret = (uint32_t)offset;
772 ret = (*gpuobj != NULL) ? 0 : -EINVAL;
773 } else {
774 NV_ERROR(dev, "Invalid GART type %d\n", dev_priv->gart_info.type);
775 return -EINVAL;
778 return ret;
781 /* Context objects in the instance RAM have the following structure.
782 * On NV40 they are 32 byte long, on NV30 and smaller 16 bytes.
784 NV4 - NV30:
786 entry[0]
787 11:0 class
788 12 chroma key enable
789 13 user clip enable
790 14 swizzle enable
791 17:15 patch config:
792 scrcopy_and, rop_and, blend_and, scrcopy, srccopy_pre, blend_pre
793 18 synchronize enable
794 19 endian: 1 big, 0 little
795 21:20 dither mode
796 23 single step enable
797 24 patch status: 0 invalid, 1 valid
798 25 context_surface 0: 1 valid
799 26 context surface 1: 1 valid
800 27 context pattern: 1 valid
801 28 context rop: 1 valid
802 29,30 context beta, beta4
803 entry[1]
804 7:0 mono format
805 15:8 color format
806 31:16 notify instance address
807 entry[2]
808 15:0 dma 0 instance address
809 31:16 dma 1 instance address
810 entry[3]
811 dma method traps
813 NV40:
814 No idea what the exact format is. Here's what can be deducted:
816 entry[0]:
817 11:0 class (maybe uses more bits here?)
818 17 user clip enable
819 21:19 patch config
820 25 patch status valid ?
821 entry[1]:
822 15:0 DMA notifier (maybe 20:0)
823 entry[2]:
824 15:0 DMA 0 instance (maybe 20:0)
825 24 big endian
826 entry[3]:
827 15:0 DMA 1 instance (maybe 20:0)
828 entry[4]:
829 entry[5]:
830 set to 0?
833 nouveau_gpuobj_gr_new(struct nouveau_channel *chan, int class,
834 struct nouveau_gpuobj **gpuobj)
836 struct drm_device *dev = chan->dev;
837 struct drm_nouveau_private *dev_priv = dev->dev_private;
838 int ret;
840 NV_DEBUG(dev, "ch%d class=0x%04x\n", chan->id, class);
842 ret = nouveau_gpuobj_new(dev, chan,
843 nouveau_gpuobj_class_instmem_size(dev, class),
845 NVOBJ_FLAG_ZERO_ALLOC | NVOBJ_FLAG_ZERO_FREE,
846 gpuobj);
847 if (ret) {
848 NV_ERROR(dev, "Error creating gpuobj: %d\n", ret);
849 return ret;
852 dev_priv->engine.instmem.prepare_access(dev, true);
853 if (dev_priv->card_type >= NV_50) {
854 nv_wo32(dev, *gpuobj, 0, class);
855 nv_wo32(dev, *gpuobj, 5, 0x00010000);
856 } else {
857 switch (class) {
858 case NV_CLASS_NULL:
859 nv_wo32(dev, *gpuobj, 0, 0x00001030);
860 nv_wo32(dev, *gpuobj, 1, 0xFFFFFFFF);
861 break;
862 default:
863 if (dev_priv->card_type >= NV_40) {
864 nv_wo32(dev, *gpuobj, 0, class);
865 #ifdef __BIG_ENDIAN
866 nv_wo32(dev, *gpuobj, 2, 0x01000000);
867 #endif
868 } else {
869 #ifdef __BIG_ENDIAN
870 nv_wo32(dev, *gpuobj, 0, class | 0x00080000);
871 #else
872 nv_wo32(dev, *gpuobj, 0, class);
873 #endif
877 dev_priv->engine.instmem.finish_access(dev);
879 (*gpuobj)->engine = NVOBJ_ENGINE_GR;
880 (*gpuobj)->class = class;
881 return 0;
884 static int
885 nouveau_gpuobj_sw_new(struct nouveau_channel *chan, int class,
886 struct nouveau_gpuobj **gpuobj_ret)
888 struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
889 struct nouveau_gpuobj *gpuobj;
891 if (!chan || !gpuobj_ret || *gpuobj_ret != NULL)
892 return -EINVAL;
894 gpuobj = kzalloc(sizeof(*gpuobj), GFP_KERNEL);
895 if (!gpuobj)
896 return -ENOMEM;
897 gpuobj->engine = NVOBJ_ENGINE_SW;
898 gpuobj->class = class;
900 list_add_tail(&gpuobj->list, &dev_priv->gpuobj_list);
901 *gpuobj_ret = gpuobj;
902 return 0;
905 static int
906 nouveau_gpuobj_channel_init_pramin(struct nouveau_channel *chan)
908 struct drm_device *dev = chan->dev;
909 struct drm_nouveau_private *dev_priv = dev->dev_private;
910 struct nouveau_gpuobj *pramin = NULL;
911 uint32_t size;
912 uint32_t base;
913 int ret;
915 NV_DEBUG(dev, "ch%d\n", chan->id);
917 /* Base amount for object storage (4KiB enough?) */
918 size = 0x1000;
919 base = 0;
921 /* PGRAPH context */
923 if (dev_priv->card_type == NV_50) {
924 /* Various fixed table thingos */
925 size += 0x1400; /* mostly unknown stuff */
926 size += 0x4000; /* vm pd */
927 base = 0x6000;
928 /* RAMHT, not sure about setting size yet, 32KiB to be safe */
929 size += 0x8000;
930 /* RAMFC */
931 size += 0x1000;
932 /* PGRAPH context */
933 size += 0x70000;
936 NV_DEBUG(dev, "ch%d PRAMIN size: 0x%08x bytes, base alloc=0x%08x\n",
937 chan->id, size, base);
938 ret = nouveau_gpuobj_new_ref(dev, NULL, NULL, 0, size, 0x1000, 0,
939 &chan->ramin);
940 if (ret) {
941 NV_ERROR(dev, "Error allocating channel PRAMIN: %d\n", ret);
942 return ret;
944 pramin = chan->ramin->gpuobj;
946 ret = nouveau_mem_init_heap(&chan->ramin_heap,
947 pramin->im_pramin->start + base, size);
948 if (ret) {
949 NV_ERROR(dev, "Error creating PRAMIN heap: %d\n", ret);
950 nouveau_gpuobj_ref_del(dev, &chan->ramin);
951 return ret;
954 return 0;
958 nouveau_gpuobj_channel_init(struct nouveau_channel *chan,
959 uint32_t vram_h, uint32_t tt_h)
961 struct drm_device *dev = chan->dev;
962 struct drm_nouveau_private *dev_priv = dev->dev_private;
963 struct nouveau_instmem_engine *instmem = &dev_priv->engine.instmem;
964 struct nouveau_gpuobj *vram = NULL, *tt = NULL;
965 int ret, i;
967 INIT_LIST_HEAD(&chan->ramht_refs);
969 NV_DEBUG(dev, "ch%d vram=0x%08x tt=0x%08x\n", chan->id, vram_h, tt_h);
971 /* Reserve a block of PRAMIN for the channel
972 *XXX: maybe on <NV50 too at some point
974 if (0 || dev_priv->card_type == NV_50) {
975 ret = nouveau_gpuobj_channel_init_pramin(chan);
976 if (ret) {
977 NV_ERROR(dev, "init pramin\n");
978 return ret;
982 /* NV50 VM
983 * - Allocate per-channel page-directory
984 * - Map GART and VRAM into the channel's address space at the
985 * locations determined during init.
987 if (dev_priv->card_type >= NV_50) {
988 uint32_t vm_offset, pde;
990 instmem->prepare_access(dev, true);
992 vm_offset = (dev_priv->chipset & 0xf0) == 0x50 ? 0x1400 : 0x200;
993 vm_offset += chan->ramin->gpuobj->im_pramin->start;
995 ret = nouveau_gpuobj_new_fake(dev, vm_offset, ~0, 0x4000,
996 0, &chan->vm_pd, NULL);
997 if (ret) {
998 instmem->finish_access(dev);
999 return ret;
1001 for (i = 0; i < 0x4000; i += 8) {
1002 nv_wo32(dev, chan->vm_pd, (i+0)/4, 0x00000000);
1003 nv_wo32(dev, chan->vm_pd, (i+4)/4, 0xdeadcafe);
1006 pde = (dev_priv->vm_gart_base / (512*1024*1024)) * 2;
1007 ret = nouveau_gpuobj_ref_add(dev, NULL, 0,
1008 dev_priv->gart_info.sg_ctxdma,
1009 &chan->vm_gart_pt);
1010 if (ret) {
1011 instmem->finish_access(dev);
1012 return ret;
1014 nv_wo32(dev, chan->vm_pd, pde++,
1015 chan->vm_gart_pt->instance | 0x03);
1016 nv_wo32(dev, chan->vm_pd, pde++, 0x00000000);
1018 pde = (dev_priv->vm_vram_base / (512*1024*1024)) * 2;
1019 for (i = 0; i < dev_priv->vm_vram_pt_nr; i++) {
1020 ret = nouveau_gpuobj_ref_add(dev, NULL, 0,
1021 dev_priv->vm_vram_pt[i],
1022 &chan->vm_vram_pt[i]);
1023 if (ret) {
1024 instmem->finish_access(dev);
1025 return ret;
1028 nv_wo32(dev, chan->vm_pd, pde++,
1029 chan->vm_vram_pt[i]->instance | 0x61);
1030 nv_wo32(dev, chan->vm_pd, pde++, 0x00000000);
1033 instmem->finish_access(dev);
1036 /* RAMHT */
1037 if (dev_priv->card_type < NV_50) {
1038 ret = nouveau_gpuobj_ref_add(dev, NULL, 0, dev_priv->ramht,
1039 &chan->ramht);
1040 if (ret)
1041 return ret;
1042 } else {
1043 ret = nouveau_gpuobj_new_ref(dev, chan, chan, 0,
1044 0x8000, 16,
1045 NVOBJ_FLAG_ZERO_ALLOC,
1046 &chan->ramht);
1047 if (ret)
1048 return ret;
1051 /* VRAM ctxdma */
1052 if (dev_priv->card_type >= NV_50) {
1053 ret = nouveau_gpuobj_dma_new(chan, NV_CLASS_DMA_IN_MEMORY,
1054 0, dev_priv->vm_end,
1055 NV_DMA_ACCESS_RW,
1056 NV_DMA_TARGET_AGP, &vram);
1057 if (ret) {
1058 NV_ERROR(dev, "Error creating VRAM ctxdma: %d\n", ret);
1059 return ret;
1061 } else {
1062 ret = nouveau_gpuobj_dma_new(chan, NV_CLASS_DMA_IN_MEMORY,
1063 0, dev_priv->fb_available_size,
1064 NV_DMA_ACCESS_RW,
1065 NV_DMA_TARGET_VIDMEM, &vram);
1066 if (ret) {
1067 NV_ERROR(dev, "Error creating VRAM ctxdma: %d\n", ret);
1068 return ret;
1072 ret = nouveau_gpuobj_ref_add(dev, chan, vram_h, vram, NULL);
1073 if (ret) {
1074 NV_ERROR(dev, "Error referencing VRAM ctxdma: %d\n", ret);
1075 return ret;
1078 /* TT memory ctxdma */
1079 if (dev_priv->card_type >= NV_50) {
1080 tt = vram;
1081 } else
1082 if (dev_priv->gart_info.type != NOUVEAU_GART_NONE) {
1083 ret = nouveau_gpuobj_gart_dma_new(chan, 0,
1084 dev_priv->gart_info.aper_size,
1085 NV_DMA_ACCESS_RW, &tt, NULL);
1086 } else {
1087 NV_ERROR(dev, "Invalid GART type %d\n", dev_priv->gart_info.type);
1088 ret = -EINVAL;
1091 if (ret) {
1092 NV_ERROR(dev, "Error creating TT ctxdma: %d\n", ret);
1093 return ret;
1096 ret = nouveau_gpuobj_ref_add(dev, chan, tt_h, tt, NULL);
1097 if (ret) {
1098 NV_ERROR(dev, "Error referencing TT ctxdma: %d\n", ret);
1099 return ret;
1102 return 0;
1105 void
1106 nouveau_gpuobj_channel_takedown(struct nouveau_channel *chan)
1108 struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
1109 struct drm_device *dev = chan->dev;
1110 struct list_head *entry, *tmp;
1111 struct nouveau_gpuobj_ref *ref;
1112 int i;
1114 NV_DEBUG(dev, "ch%d\n", chan->id);
1116 if (!chan->ramht_refs.next)
1117 return;
1119 list_for_each_safe(entry, tmp, &chan->ramht_refs) {
1120 ref = list_entry(entry, struct nouveau_gpuobj_ref, list);
1122 nouveau_gpuobj_ref_del(dev, &ref);
1125 nouveau_gpuobj_ref_del(dev, &chan->ramht);
1127 nouveau_gpuobj_del(dev, &chan->vm_pd);
1128 nouveau_gpuobj_ref_del(dev, &chan->vm_gart_pt);
1129 for (i = 0; i < dev_priv->vm_vram_pt_nr; i++)
1130 nouveau_gpuobj_ref_del(dev, &chan->vm_vram_pt[i]);
1132 if (chan->ramin_heap)
1133 nouveau_mem_takedown(&chan->ramin_heap);
1134 if (chan->ramin)
1135 nouveau_gpuobj_ref_del(dev, &chan->ramin);
1140 nouveau_gpuobj_suspend(struct drm_device *dev)
1142 struct drm_nouveau_private *dev_priv = dev->dev_private;
1143 struct nouveau_gpuobj *gpuobj;
1144 int i;
1146 if (dev_priv->card_type < NV_50) {
1147 dev_priv->susres.ramin_copy = vmalloc(dev_priv->ramin_rsvd_vram);
1148 if (!dev_priv->susres.ramin_copy)
1149 return -ENOMEM;
1151 for (i = 0; i < dev_priv->ramin_rsvd_vram; i += 4)
1152 dev_priv->susres.ramin_copy[i/4] = nv_ri32(dev, i);
1153 return 0;
1156 list_for_each_entry(gpuobj, &dev_priv->gpuobj_list, list) {
1157 if (!gpuobj->im_backing || (gpuobj->flags & NVOBJ_FLAG_FAKE))
1158 continue;
1160 gpuobj->im_backing_suspend = vmalloc(gpuobj->im_pramin->size);
1161 if (!gpuobj->im_backing_suspend) {
1162 nouveau_gpuobj_resume(dev);
1163 return -ENOMEM;
1166 dev_priv->engine.instmem.prepare_access(dev, false);
1167 for (i = 0; i < gpuobj->im_pramin->size / 4; i++)
1168 gpuobj->im_backing_suspend[i] = nv_ro32(dev, gpuobj, i);
1169 dev_priv->engine.instmem.finish_access(dev);
1172 return 0;
1175 void
1176 nouveau_gpuobj_suspend_cleanup(struct drm_device *dev)
1178 struct drm_nouveau_private *dev_priv = dev->dev_private;
1179 struct nouveau_gpuobj *gpuobj;
1181 if (dev_priv->card_type < NV_50) {
1182 vfree(dev_priv->susres.ramin_copy);
1183 dev_priv->susres.ramin_copy = NULL;
1184 return;
1187 list_for_each_entry(gpuobj, &dev_priv->gpuobj_list, list) {
1188 if (!gpuobj->im_backing_suspend)
1189 continue;
1191 vfree(gpuobj->im_backing_suspend);
1192 gpuobj->im_backing_suspend = NULL;
1196 void
1197 nouveau_gpuobj_resume(struct drm_device *dev)
1199 struct drm_nouveau_private *dev_priv = dev->dev_private;
1200 struct nouveau_gpuobj *gpuobj;
1201 int i;
1203 if (dev_priv->card_type < NV_50) {
1204 for (i = 0; i < dev_priv->ramin_rsvd_vram; i += 4)
1205 nv_wi32(dev, i, dev_priv->susres.ramin_copy[i/4]);
1206 nouveau_gpuobj_suspend_cleanup(dev);
1207 return;
1210 list_for_each_entry(gpuobj, &dev_priv->gpuobj_list, list) {
1211 if (!gpuobj->im_backing_suspend)
1212 continue;
1214 dev_priv->engine.instmem.prepare_access(dev, true);
1215 for (i = 0; i < gpuobj->im_pramin->size / 4; i++)
1216 nv_wo32(dev, gpuobj, i, gpuobj->im_backing_suspend[i]);
1217 dev_priv->engine.instmem.finish_access(dev);
1220 nouveau_gpuobj_suspend_cleanup(dev);
1223 int nouveau_ioctl_grobj_alloc(struct drm_device *dev, void *data,
1224 struct drm_file *file_priv)
1226 struct drm_nouveau_private *dev_priv = dev->dev_private;
1227 struct drm_nouveau_grobj_alloc *init = data;
1228 struct nouveau_pgraph_engine *pgraph = &dev_priv->engine.graph;
1229 struct nouveau_pgraph_object_class *grc;
1230 struct nouveau_gpuobj *gr = NULL;
1231 struct nouveau_channel *chan;
1232 int ret;
1234 NOUVEAU_CHECK_INITIALISED_WITH_RETURN;
1235 NOUVEAU_GET_USER_CHANNEL_WITH_RETURN(init->channel, file_priv, chan);
1237 if (init->handle == ~0)
1238 return -EINVAL;
1240 grc = pgraph->grclass;
1241 while (grc->id) {
1242 if (grc->id == init->class)
1243 break;
1244 grc++;
1247 if (!grc->id) {
1248 NV_ERROR(dev, "Illegal object class: 0x%x\n", init->class);
1249 return -EPERM;
1252 if (nouveau_gpuobj_ref_find(chan, init->handle, NULL) == 0)
1253 return -EEXIST;
1255 if (!grc->software)
1256 ret = nouveau_gpuobj_gr_new(chan, grc->id, &gr);
1257 else
1258 ret = nouveau_gpuobj_sw_new(chan, grc->id, &gr);
1260 if (ret) {
1261 NV_ERROR(dev, "Error creating object: %d (%d/0x%08x)\n",
1262 ret, init->channel, init->handle);
1263 return ret;
1266 ret = nouveau_gpuobj_ref_add(dev, chan, init->handle, gr, NULL);
1267 if (ret) {
1268 NV_ERROR(dev, "Error referencing object: %d (%d/0x%08x)\n",
1269 ret, init->channel, init->handle);
1270 nouveau_gpuobj_del(dev, &gr);
1271 return ret;
1274 return 0;
1277 int nouveau_ioctl_gpuobj_free(struct drm_device *dev, void *data,
1278 struct drm_file *file_priv)
1280 struct drm_nouveau_gpuobj_free *objfree = data;
1281 struct nouveau_gpuobj_ref *ref;
1282 struct nouveau_channel *chan;
1283 int ret;
1285 NOUVEAU_CHECK_INITIALISED_WITH_RETURN;
1286 NOUVEAU_GET_USER_CHANNEL_WITH_RETURN(objfree->channel, file_priv, chan);
1288 ret = nouveau_gpuobj_ref_find(chan, objfree->handle, &ref);
1289 if (ret)
1290 return ret;
1291 nouveau_gpuobj_ref_del(dev, &ref);
1293 return 0;