1 /*
2 * Copyright (C) 2006 Ben Skeggs.
3 *
4 * All Rights Reserved.
5 *
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:
13 *
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.
17 *
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.
25 *
26 */
27
28 /*
29 * Authors:
30 * Ben Skeggs <darktama@iinet.net.au>
31 */
32
33 #include "drmP.h"
34 #include "drm.h"
35 #include "nouveau_drv.h"
36 #include "nouveau_drm.h"
37 #include "nouveau_ramht.h"
38 #include "nouveau_vm.h"
39 #include "nv50_display.h"
40
41 struct nouveau_gpuobj_method {
42 struct list_head head;
43 u32 mthd;
44 int (*exec)(struct nouveau_channel *, u32 class, u32 mthd, u32 data);
45 };
46
47 struct nouveau_gpuobj_class {
48 struct list_head head;
49 struct list_head methods;
50 u32 id;
51 u32 engine;
52 };
53
54 int
nouveau_gpuobj_class_new(struct drm_device * dev,u32 class,u32 engine)55 nouveau_gpuobj_class_new(struct drm_device *dev, u32 class, u32 engine)
56 {
57 struct drm_nouveau_private *dev_priv = dev->dev_private;
58 struct nouveau_gpuobj_class *oc;
59
60 oc = kzalloc(sizeof(*oc), GFP_KERNEL);
61 if (!oc)
62 return -ENOMEM;
63
64 INIT_LIST_HEAD(&oc->methods);
65 oc->id = class;
66 oc->engine = engine;
67 list_add(&oc->head, &dev_priv->classes);
68 return 0;
69 }
70
71 int
nouveau_gpuobj_mthd_new(struct drm_device * dev,u32 class,u32 mthd,int (* exec)(struct nouveau_channel *,u32,u32,u32))72 nouveau_gpuobj_mthd_new(struct drm_device *dev, u32 class, u32 mthd,
73 int (*exec)(struct nouveau_channel *, u32, u32, u32))
74 {
75 struct drm_nouveau_private *dev_priv = dev->dev_private;
76 struct nouveau_gpuobj_method *om;
77 struct nouveau_gpuobj_class *oc;
78
79 list_for_each_entry(oc, &dev_priv->classes, head) {
80 if (oc->id == class)
81 goto found;
82 }
83
84 return -EINVAL;
85
86 found:
87 om = kzalloc(sizeof(*om), GFP_KERNEL);
88 if (!om)
89 return -ENOMEM;
90
91 om->mthd = mthd;
92 om->exec = exec;
93 list_add(&om->head, &oc->methods);
94 return 0;
95 }
96
97 int
nouveau_gpuobj_mthd_call(struct nouveau_channel * chan,u32 class,u32 mthd,u32 data)98 nouveau_gpuobj_mthd_call(struct nouveau_channel *chan,
99 u32 class, u32 mthd, u32 data)
100 {
101 struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
102 struct nouveau_gpuobj_method *om;
103 struct nouveau_gpuobj_class *oc;
104
105 list_for_each_entry(oc, &dev_priv->classes, head) {
106 if (oc->id != class)
107 continue;
108
109 list_for_each_entry(om, &oc->methods, head) {
110 if (om->mthd == mthd)
111 return om->exec(chan, class, mthd, data);
112 }
113 }
114
115 return -ENOENT;
116 }
117
118 int
nouveau_gpuobj_mthd_call2(struct drm_device * dev,int chid,u32 class,u32 mthd,u32 data)119 nouveau_gpuobj_mthd_call2(struct drm_device *dev, int chid,
120 u32 class, u32 mthd, u32 data)
121 {
122 struct drm_nouveau_private *dev_priv = dev->dev_private;
123 struct nouveau_channel *chan = NULL;
124 unsigned long flags;
125 int ret = -EINVAL;
126
127 spin_lock_irqsave(&dev_priv->channels.lock, flags);
128 if (chid >= 0 && chid < dev_priv->engine.fifo.channels)
129 chan = dev_priv->channels.ptr[chid];
130 if (chan)
131 ret = nouveau_gpuobj_mthd_call(chan, class, mthd, data);
132 spin_unlock_irqrestore(&dev_priv->channels.lock, flags);
133 return ret;
134 }
135
136 /* NVidia uses context objects to drive drawing operations.
137
138 Context objects can be selected into 8 subchannels in the FIFO,
139 and then used via DMA command buffers.
140
141 A context object is referenced by a user defined handle (CARD32). The HW
142 looks up graphics objects in a hash table in the instance RAM.
143
144 An entry in the hash table consists of 2 CARD32. The first CARD32 contains
145 the handle, the second one a bitfield, that contains the address of the
146 object in instance RAM.
147
148 The format of the second CARD32 seems to be:
149
150 NV4 to NV30:
151
152 15: 0 instance_addr >> 4
153 17:16 engine (here uses 1 = graphics)
154 28:24 channel id (here uses 0)
155 31 valid (use 1)
156
157 NV40:
158
159 15: 0 instance_addr >> 4 (maybe 19-0)
160 21:20 engine (here uses 1 = graphics)
161 I'm unsure about the other bits, but using 0 seems to work.
162
163 The key into the hash table depends on the object handle and channel id and
164 is given as:
165 */
166
167 int
nouveau_gpuobj_new(struct drm_device * dev,struct nouveau_channel * chan,uint32_t size,int align,uint32_t flags,struct nouveau_gpuobj ** gpuobj_ret)168 nouveau_gpuobj_new(struct drm_device *dev, struct nouveau_channel *chan,
169 uint32_t size, int align, uint32_t flags,
170 struct nouveau_gpuobj **gpuobj_ret)
171 {
172 struct drm_nouveau_private *dev_priv = dev->dev_private;
173 struct nouveau_instmem_engine *instmem = &dev_priv->engine.instmem;
174 struct nouveau_gpuobj *gpuobj;
175 struct drm_mm_node *ramin = NULL;
176 int ret, i;
177
178 NV_DEBUG(dev, "ch%d size=%u align=%d flags=0x%08x\n",
179 chan ? chan->id : -1, size, align, flags);
180
181 gpuobj = kzalloc(sizeof(*gpuobj), GFP_KERNEL);
182 if (!gpuobj)
183 return -ENOMEM;
184 NV_DEBUG(dev, "gpuobj %p\n", gpuobj);
185 gpuobj->dev = dev;
186 gpuobj->flags = flags;
187 kref_init(&gpuobj->refcount);
188 gpuobj->size = size;
189
190 spin_lock(&dev_priv->ramin_lock);
191 list_add_tail(&gpuobj->list, &dev_priv->gpuobj_list);
192 spin_unlock(&dev_priv->ramin_lock);
193
194 if (!(flags & NVOBJ_FLAG_VM) && chan) {
195 ramin = drm_mm_search_free(&chan->ramin_heap, size, align, 0);
196 if (ramin)
197 ramin = drm_mm_get_block(ramin, size, align);
198 if (!ramin) {
199 nouveau_gpuobj_ref(NULL, &gpuobj);
200 return -ENOMEM;
201 }
202
203 gpuobj->pinst = chan->ramin->pinst;
204 if (gpuobj->pinst != ~0)
205 gpuobj->pinst += ramin->start;
206
207 gpuobj->cinst = ramin->start;
208 gpuobj->vinst = ramin->start + chan->ramin->vinst;
209 gpuobj->node = ramin;
210 } else {
211 ret = instmem->get(gpuobj, chan, size, align);
212 if (ret) {
213 nouveau_gpuobj_ref(NULL, &gpuobj);
214 return ret;
215 }
216
217 ret = -ENOSYS;
218 if (!(flags & NVOBJ_FLAG_DONT_MAP))
219 ret = instmem->map(gpuobj);
220 if (ret)
221 gpuobj->pinst = ~0;
222
223 gpuobj->cinst = NVOBJ_CINST_GLOBAL;
224 }
225
226 if (gpuobj->flags & NVOBJ_FLAG_ZERO_ALLOC) {
227 for (i = 0; i < gpuobj->size; i += 4)
228 nv_wo32(gpuobj, i, 0);
229 instmem->flush(dev);
230 }
231
232
233 *gpuobj_ret = gpuobj;
234 return 0;
235 }
236
237 int
nouveau_gpuobj_init(struct drm_device * dev)238 nouveau_gpuobj_init(struct drm_device *dev)
239 {
240 struct drm_nouveau_private *dev_priv = dev->dev_private;
241
242 NV_DEBUG(dev, "\n");
243
244 INIT_LIST_HEAD(&dev_priv->gpuobj_list);
245 INIT_LIST_HEAD(&dev_priv->classes);
246 spin_lock_init(&dev_priv->ramin_lock);
247 dev_priv->ramin_base = ~0;
248
249 return 0;
250 }
251
252 void
nouveau_gpuobj_takedown(struct drm_device * dev)253 nouveau_gpuobj_takedown(struct drm_device *dev)
254 {
255 struct drm_nouveau_private *dev_priv = dev->dev_private;
256 struct nouveau_gpuobj_method *om, *tm;
257 struct nouveau_gpuobj_class *oc, *tc;
258
259 NV_DEBUG(dev, "\n");
260
261 list_for_each_entry_safe(oc, tc, &dev_priv->classes, head) {
262 list_for_each_entry_safe(om, tm, &oc->methods, head) {
263 list_del(&om->head);
264 kfree(om);
265 }
266 list_del(&oc->head);
267 kfree(oc);
268 }
269
270 BUG_ON(!list_empty(&dev_priv->gpuobj_list));
271 }
272
273
274 static void
nouveau_gpuobj_del(struct kref * ref)275 nouveau_gpuobj_del(struct kref *ref)
276 {
277 struct nouveau_gpuobj *gpuobj =
278 container_of(ref, struct nouveau_gpuobj, refcount);
279 struct drm_device *dev = gpuobj->dev;
280 struct drm_nouveau_private *dev_priv = dev->dev_private;
281 struct nouveau_instmem_engine *instmem = &dev_priv->engine.instmem;
282 int i;
283
284 NV_DEBUG(dev, "gpuobj %p\n", gpuobj);
285
286 if (gpuobj->node && (gpuobj->flags & NVOBJ_FLAG_ZERO_FREE)) {
287 for (i = 0; i < gpuobj->size; i += 4)
288 nv_wo32(gpuobj, i, 0);
289 instmem->flush(dev);
290 }
291
292 if (gpuobj->dtor)
293 gpuobj->dtor(dev, gpuobj);
294
295 if (gpuobj->cinst == NVOBJ_CINST_GLOBAL) {
296 if (gpuobj->node) {
297 instmem->unmap(gpuobj);
298 instmem->put(gpuobj);
299 }
300 } else {
301 if (gpuobj->node) {
302 spin_lock(&dev_priv->ramin_lock);
303 drm_mm_put_block(gpuobj->node);
304 spin_unlock(&dev_priv->ramin_lock);
305 }
306 }
307
308 spin_lock(&dev_priv->ramin_lock);
309 list_del(&gpuobj->list);
310 spin_unlock(&dev_priv->ramin_lock);
311
312 kfree(gpuobj);
313 }
314
315 void
nouveau_gpuobj_ref(struct nouveau_gpuobj * ref,struct nouveau_gpuobj ** ptr)316 nouveau_gpuobj_ref(struct nouveau_gpuobj *ref, struct nouveau_gpuobj **ptr)
317 {
318 if (ref)
319 kref_get(&ref->refcount);
320
321 if (*ptr)
322 kref_put(&(*ptr)->refcount, nouveau_gpuobj_del);
323
324 *ptr = ref;
325 }
326
327 int
nouveau_gpuobj_new_fake(struct drm_device * dev,u32 pinst,u64 vinst,u32 size,u32 flags,struct nouveau_gpuobj ** pgpuobj)328 nouveau_gpuobj_new_fake(struct drm_device *dev, u32 pinst, u64 vinst,
329 u32 size, u32 flags, struct nouveau_gpuobj **pgpuobj)
330 {
331 struct drm_nouveau_private *dev_priv = dev->dev_private;
332 struct nouveau_gpuobj *gpuobj = NULL;
333 int i;
334
335 NV_DEBUG(dev,
336 "pinst=0x%08x vinst=0x%010llx size=0x%08x flags=0x%08x\n",
337 pinst, vinst, size, flags);
338
339 gpuobj = kzalloc(sizeof(*gpuobj), GFP_KERNEL);
340 if (!gpuobj)
341 return -ENOMEM;
342 NV_DEBUG(dev, "gpuobj %p\n", gpuobj);
343 gpuobj->dev = dev;
344 gpuobj->flags = flags;
345 kref_init(&gpuobj->refcount);
346 gpuobj->size = size;
347 gpuobj->pinst = pinst;
348 gpuobj->cinst = NVOBJ_CINST_GLOBAL;
349 gpuobj->vinst = vinst;
350
351 if (gpuobj->flags & NVOBJ_FLAG_ZERO_ALLOC) {
352 for (i = 0; i < gpuobj->size; i += 4)
353 nv_wo32(gpuobj, i, 0);
354 dev_priv->engine.instmem.flush(dev);
355 }
356
357 spin_lock(&dev_priv->ramin_lock);
358 list_add_tail(&gpuobj->list, &dev_priv->gpuobj_list);
359 spin_unlock(&dev_priv->ramin_lock);
360 *pgpuobj = gpuobj;
361 return 0;
362 }
363
364 /*
365 DMA objects are used to reference a piece of memory in the
366 framebuffer, PCI or AGP address space. Each object is 16 bytes big
367 and looks as follows:
368
369 entry[0]
370 11:0 class (seems like I can always use 0 here)
371 12 page table present?
372 13 page entry linear?
373 15:14 access: 0 rw, 1 ro, 2 wo
374 17:16 target: 0 NV memory, 1 NV memory tiled, 2 PCI, 3 AGP
375 31:20 dma adjust (bits 0-11 of the address)
376 entry[1]
377 dma limit (size of transfer)
378 entry[X]
379 1 0 readonly, 1 readwrite
380 31:12 dma frame address of the page (bits 12-31 of the address)
381 entry[N]
382 page table terminator, same value as the first pte, as does nvidia
383 rivatv uses 0xffffffff
384
385 Non linear page tables need a list of frame addresses afterwards,
386 the rivatv project has some info on this.
387
388 The method below creates a DMA object in instance RAM and returns a handle
389 to it that can be used to set up context objects.
390 */
391
392 void
nv50_gpuobj_dma_init(struct nouveau_gpuobj * obj,u32 offset,int class,u64 base,u64 size,int target,int access,u32 type,u32 comp)393 nv50_gpuobj_dma_init(struct nouveau_gpuobj *obj, u32 offset, int class,
394 u64 base, u64 size, int target, int access,
395 u32 type, u32 comp)
396 {
397 struct drm_nouveau_private *dev_priv = obj->dev->dev_private;
398 struct nouveau_instmem_engine *pinstmem = &dev_priv->engine.instmem;
399 u32 flags0;
400
401 flags0 = (comp << 29) | (type << 22) | class;
402 flags0 |= 0x00100000;
403
404 switch (access) {
405 case NV_MEM_ACCESS_RO: flags0 |= 0x00040000; break;
406 case NV_MEM_ACCESS_RW:
407 case NV_MEM_ACCESS_WO: flags0 |= 0x00080000; break;
408 default:
409 break;
410 }
411
412 switch (target) {
413 case NV_MEM_TARGET_VRAM:
414 flags0 |= 0x00010000;
415 break;
416 case NV_MEM_TARGET_PCI:
417 flags0 |= 0x00020000;
418 break;
419 case NV_MEM_TARGET_PCI_NOSNOOP:
420 flags0 |= 0x00030000;
421 break;
422 case NV_MEM_TARGET_GART:
423 base += dev_priv->gart_info.aper_base;
424 default:
425 flags0 &= ~0x00100000;
426 break;
427 }
428
429 /* convert to base + limit */
430 size = (base + size) - 1;
431
432 nv_wo32(obj, offset + 0x00, flags0);
433 nv_wo32(obj, offset + 0x04, lower_32_bits(size));
434 nv_wo32(obj, offset + 0x08, lower_32_bits(base));
435 nv_wo32(obj, offset + 0x0c, upper_32_bits(size) << 24 |
436 upper_32_bits(base));
437 nv_wo32(obj, offset + 0x10, 0x00000000);
438 nv_wo32(obj, offset + 0x14, 0x00000000);
439
440 pinstmem->flush(obj->dev);
441 }
442
443 int
nv50_gpuobj_dma_new(struct nouveau_channel * chan,int class,u64 base,u64 size,int target,int access,u32 type,u32 comp,struct nouveau_gpuobj ** pobj)444 nv50_gpuobj_dma_new(struct nouveau_channel *chan, int class, u64 base, u64 size,
445 int target, int access, u32 type, u32 comp,
446 struct nouveau_gpuobj **pobj)
447 {
448 struct drm_device *dev = chan->dev;
449 int ret;
450
451 ret = nouveau_gpuobj_new(dev, chan, 24, 16, NVOBJ_FLAG_ZERO_FREE, pobj);
452 if (ret)
453 return ret;
454
455 nv50_gpuobj_dma_init(*pobj, 0, class, base, size, target,
456 access, type, comp);
457 return 0;
458 }
459
460 int
nouveau_gpuobj_dma_new(struct nouveau_channel * chan,int class,u64 base,u64 size,int access,int target,struct nouveau_gpuobj ** pobj)461 nouveau_gpuobj_dma_new(struct nouveau_channel *chan, int class, u64 base,
462 u64 size, int access, int target,
463 struct nouveau_gpuobj **pobj)
464 {
465 struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
466 struct drm_device *dev = chan->dev;
467 struct nouveau_gpuobj *obj;
468 u32 flags0, flags2;
469 int ret;
470
471 if (dev_priv->card_type >= NV_50) {
472 u32 comp = (target == NV_MEM_TARGET_VM) ? NV_MEM_COMP_VM : 0;
473 u32 type = (target == NV_MEM_TARGET_VM) ? NV_MEM_TYPE_VM : 0;
474
475 return nv50_gpuobj_dma_new(chan, class, base, size,
476 target, access, type, comp, pobj);
477 }
478
479 if (target == NV_MEM_TARGET_GART) {
480 struct nouveau_gpuobj *gart = dev_priv->gart_info.sg_ctxdma;
481
482 if (dev_priv->gart_info.type == NOUVEAU_GART_PDMA) {
483 if (base == 0) {
484 nouveau_gpuobj_ref(gart, pobj);
485 return 0;
486 }
487
488 base = nouveau_sgdma_get_physical(dev, base);
489 target = NV_MEM_TARGET_PCI;
490 } else {
491 base += dev_priv->gart_info.aper_base;
492 if (dev_priv->gart_info.type == NOUVEAU_GART_AGP)
493 target = NV_MEM_TARGET_PCI_NOSNOOP;
494 else
495 target = NV_MEM_TARGET_PCI;
496 }
497 }
498
499 flags0 = class;
500 flags0 |= 0x00003000; /* PT present, PT linear */
501 flags2 = 0;
502
503 switch (target) {
504 case NV_MEM_TARGET_PCI:
505 flags0 |= 0x00020000;
506 break;
507 case NV_MEM_TARGET_PCI_NOSNOOP:
508 flags0 |= 0x00030000;
509 break;
510 default:
511 break;
512 }
513
514 switch (access) {
515 case NV_MEM_ACCESS_RO:
516 flags0 |= 0x00004000;
517 break;
518 case NV_MEM_ACCESS_WO:
519 flags0 |= 0x00008000;
520 default:
521 flags2 |= 0x00000002;
522 break;
523 }
524
525 flags0 |= (base & 0x00000fff) << 20;
526 flags2 |= (base & 0xfffff000);
527
528 ret = nouveau_gpuobj_new(dev, chan, 16, 16, NVOBJ_FLAG_ZERO_FREE, &obj);
529 if (ret)
530 return ret;
531
532 nv_wo32(obj, 0x00, flags0);
533 nv_wo32(obj, 0x04, size - 1);
534 nv_wo32(obj, 0x08, flags2);
535 nv_wo32(obj, 0x0c, flags2);
536
537 obj->engine = NVOBJ_ENGINE_SW;
538 obj->class = class;
539 *pobj = obj;
540 return 0;
541 }
542
543 /* Context objects in the instance RAM have the following structure.
544 * On NV40 they are 32 byte long, on NV30 and smaller 16 bytes.
545
546 NV4 - NV30:
547
548 entry[0]
549 11:0 class
550 12 chroma key enable
551 13 user clip enable
552 14 swizzle enable
553 17:15 patch config:
554 scrcopy_and, rop_and, blend_and, scrcopy, srccopy_pre, blend_pre
555 18 synchronize enable
556 19 endian: 1 big, 0 little
557 21:20 dither mode
558 23 single step enable
559 24 patch status: 0 invalid, 1 valid
560 25 context_surface 0: 1 valid
561 26 context surface 1: 1 valid
562 27 context pattern: 1 valid
563 28 context rop: 1 valid
564 29,30 context beta, beta4
565 entry[1]
566 7:0 mono format
567 15:8 color format
568 31:16 notify instance address
569 entry[2]
570 15:0 dma 0 instance address
571 31:16 dma 1 instance address
572 entry[3]
573 dma method traps
574
575 NV40:
576 No idea what the exact format is. Here's what can be deducted:
577
578 entry[0]:
579 11:0 class (maybe uses more bits here?)
580 17 user clip enable
581 21:19 patch config
582 25 patch status valid ?
583 entry[1]:
584 15:0 DMA notifier (maybe 20:0)
585 entry[2]:
586 15:0 DMA 0 instance (maybe 20:0)
587 24 big endian
588 entry[3]:
589 15:0 DMA 1 instance (maybe 20:0)
590 entry[4]:
591 entry[5]:
592 set to 0?
593 */
594 static int
nouveau_gpuobj_sw_new(struct nouveau_channel * chan,u32 handle,u16 class)595 nouveau_gpuobj_sw_new(struct nouveau_channel *chan, u32 handle, u16 class)
596 {
597 struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
598 struct nouveau_gpuobj *gpuobj;
599 int ret;
600
601 gpuobj = kzalloc(sizeof(*gpuobj), GFP_KERNEL);
602 if (!gpuobj)
603 return -ENOMEM;
604 gpuobj->dev = chan->dev;
605 gpuobj->engine = NVOBJ_ENGINE_SW;
606 gpuobj->class = class;
607 kref_init(&gpuobj->refcount);
608 gpuobj->cinst = 0x40;
609
610 spin_lock(&dev_priv->ramin_lock);
611 list_add_tail(&gpuobj->list, &dev_priv->gpuobj_list);
612 spin_unlock(&dev_priv->ramin_lock);
613
614 ret = nouveau_ramht_insert(chan, handle, gpuobj);
615 nouveau_gpuobj_ref(NULL, &gpuobj);
616 return ret;
617 }
618
619 int
nouveau_gpuobj_gr_new(struct nouveau_channel * chan,u32 handle,int class)620 nouveau_gpuobj_gr_new(struct nouveau_channel *chan, u32 handle, int class)
621 {
622 struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
623 struct drm_device *dev = chan->dev;
624 struct nouveau_gpuobj_class *oc;
625 int ret;
626
627 NV_DEBUG(dev, "ch%d class=0x%04x\n", chan->id, class);
628
629 list_for_each_entry(oc, &dev_priv->classes, head) {
630 struct nouveau_exec_engine *eng = dev_priv->eng[oc->engine];
631
632 if (oc->id != class)
633 continue;
634
635 if (oc->engine == NVOBJ_ENGINE_SW)
636 return nouveau_gpuobj_sw_new(chan, handle, class);
637
638 if (!chan->engctx[oc->engine]) {
639 ret = eng->context_new(chan, oc->engine);
640 if (ret)
641 return ret;
642 }
643
644 return eng->object_new(chan, oc->engine, handle, class);
645 }
646
647 NV_ERROR(dev, "illegal object class: 0x%x\n", class);
648 return -EINVAL;
649 }
650
651 static int
nouveau_gpuobj_channel_init_pramin(struct nouveau_channel * chan)652 nouveau_gpuobj_channel_init_pramin(struct nouveau_channel *chan)
653 {
654 struct drm_device *dev = chan->dev;
655 struct drm_nouveau_private *dev_priv = dev->dev_private;
656 uint32_t size;
657 uint32_t base;
658 int ret;
659
660 NV_DEBUG(dev, "ch%d\n", chan->id);
661
662 /* Base amount for object storage (4KiB enough?) */
663 size = 0x2000;
664 base = 0;
665
666 if (dev_priv->card_type == NV_50) {
667 /* Various fixed table thingos */
668 size += 0x1400; /* mostly unknown stuff */
669 size += 0x4000; /* vm pd */
670 base = 0x6000;
671 /* RAMHT, not sure about setting size yet, 32KiB to be safe */
672 size += 0x8000;
673 /* RAMFC */
674 size += 0x1000;
675 }
676
677 ret = nouveau_gpuobj_new(dev, NULL, size, 0x1000, 0, &chan->ramin);
678 if (ret) {
679 NV_ERROR(dev, "Error allocating channel PRAMIN: %d\n", ret);
680 return ret;
681 }
682
683 ret = drm_mm_init(&chan->ramin_heap, base, size - base);
684 if (ret) {
685 NV_ERROR(dev, "Error creating PRAMIN heap: %d\n", ret);
686 nouveau_gpuobj_ref(NULL, &chan->ramin);
687 return ret;
688 }
689
690 return 0;
691 }
692
693 static int
nvc0_gpuobj_channel_init(struct nouveau_channel * chan,struct nouveau_vm * vm)694 nvc0_gpuobj_channel_init(struct nouveau_channel *chan, struct nouveau_vm *vm)
695 {
696 struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
697 struct drm_device *dev = chan->dev;
698 struct nouveau_gpuobj *pgd = NULL;
699 struct nouveau_vm_pgd *vpgd;
700 int ret, i;
701
702 ret = nouveau_gpuobj_new(dev, NULL, 4096, 0x1000, 0, &chan->ramin);
703 if (ret)
704 return ret;
705
706 /* create page directory for this vm if none currently exists,
707 * will be destroyed automagically when last reference to the
708 * vm is removed
709 */
710 if (list_empty(&vm->pgd_list)) {
711 ret = nouveau_gpuobj_new(dev, NULL, 65536, 0x1000, 0, &pgd);
712 if (ret)
713 return ret;
714 }
715 nouveau_vm_ref(vm, &chan->vm, pgd);
716 nouveau_gpuobj_ref(NULL, &pgd);
717
718 /* point channel at vm's page directory */
719 vpgd = list_first_entry(&vm->pgd_list, struct nouveau_vm_pgd, head);
720 nv_wo32(chan->ramin, 0x0200, lower_32_bits(vpgd->obj->vinst));
721 nv_wo32(chan->ramin, 0x0204, upper_32_bits(vpgd->obj->vinst));
722 nv_wo32(chan->ramin, 0x0208, 0xffffffff);
723 nv_wo32(chan->ramin, 0x020c, 0x000000ff);
724
725 /* map display semaphore buffers into channel's vm */
726 for (i = 0; i < dev->mode_config.num_crtc; i++) {
727 struct nouveau_bo *bo;
728 if (dev_priv->card_type >= NV_D0)
729 bo = nvd0_display_crtc_sema(dev, i);
730 else
731 bo = nv50_display(dev)->crtc[i].sem.bo;
732
733 ret = nouveau_bo_vma_add(bo, chan->vm, &chan->dispc_vma[i]);
734 if (ret)
735 return ret;
736 }
737
738 return 0;
739 }
740
741 int
nouveau_gpuobj_channel_init(struct nouveau_channel * chan,uint32_t vram_h,uint32_t tt_h)742 nouveau_gpuobj_channel_init(struct nouveau_channel *chan,
743 uint32_t vram_h, uint32_t tt_h)
744 {
745 struct drm_device *dev = chan->dev;
746 struct drm_nouveau_private *dev_priv = dev->dev_private;
747 struct nouveau_fpriv *fpriv = nouveau_fpriv(chan->file_priv);
748 struct nouveau_vm *vm = fpriv ? fpriv->vm : dev_priv->chan_vm;
749 struct nouveau_gpuobj *vram = NULL, *tt = NULL;
750 int ret, i;
751
752 NV_DEBUG(dev, "ch%d vram=0x%08x tt=0x%08x\n", chan->id, vram_h, tt_h);
753 if (dev_priv->card_type >= NV_C0)
754 return nvc0_gpuobj_channel_init(chan, vm);
755
756 /* Allocate a chunk of memory for per-channel object storage */
757 ret = nouveau_gpuobj_channel_init_pramin(chan);
758 if (ret) {
759 NV_ERROR(dev, "init pramin\n");
760 return ret;
761 }
762
763 /* NV50 VM
764 * - Allocate per-channel page-directory
765 * - Link with shared channel VM
766 */
767 if (vm) {
768 u32 pgd_offs = (dev_priv->chipset == 0x50) ? 0x1400 : 0x0200;
769 u64 vm_vinst = chan->ramin->vinst + pgd_offs;
770 u32 vm_pinst = chan->ramin->pinst;
771
772 if (vm_pinst != ~0)
773 vm_pinst += pgd_offs;
774
775 ret = nouveau_gpuobj_new_fake(dev, vm_pinst, vm_vinst, 0x4000,
776 0, &chan->vm_pd);
777 if (ret)
778 return ret;
779
780 nouveau_vm_ref(vm, &chan->vm, chan->vm_pd);
781 }
782
783 /* RAMHT */
784 if (dev_priv->card_type < NV_50) {
785 nouveau_ramht_ref(dev_priv->ramht, &chan->ramht, NULL);
786 } else {
787 struct nouveau_gpuobj *ramht = NULL;
788
789 ret = nouveau_gpuobj_new(dev, chan, 0x8000, 16,
790 NVOBJ_FLAG_ZERO_ALLOC, &ramht);
791 if (ret)
792 return ret;
793
794 ret = nouveau_ramht_new(dev, ramht, &chan->ramht);
795 nouveau_gpuobj_ref(NULL, &ramht);
796 if (ret)
797 return ret;
798
799 /* dma objects for display sync channel semaphore blocks */
800 for (i = 0; i < dev->mode_config.num_crtc; i++) {
801 struct nouveau_gpuobj *sem = NULL;
802 struct nv50_display_crtc *dispc =
803 &nv50_display(dev)->crtc[i];
804 u64 offset = dispc->sem.bo->bo.offset;
805
806 ret = nouveau_gpuobj_dma_new(chan, 0x3d, offset, 0xfff,
807 NV_MEM_ACCESS_RW,
808 NV_MEM_TARGET_VRAM, &sem);
809 if (ret)
810 return ret;
811
812 ret = nouveau_ramht_insert(chan, NvEvoSema0 + i, sem);
813 nouveau_gpuobj_ref(NULL, &sem);
814 if (ret)
815 return ret;
816 }
817 }
818
819 /* VRAM ctxdma */
820 if (dev_priv->card_type >= NV_50) {
821 ret = nouveau_gpuobj_dma_new(chan, NV_CLASS_DMA_IN_MEMORY,
822 0, (1ULL << 40), NV_MEM_ACCESS_RW,
823 NV_MEM_TARGET_VM, &vram);
824 if (ret) {
825 NV_ERROR(dev, "Error creating VRAM ctxdma: %d\n", ret);
826 return ret;
827 }
828 } else {
829 ret = nouveau_gpuobj_dma_new(chan, NV_CLASS_DMA_IN_MEMORY,
830 0, dev_priv->fb_available_size,
831 NV_MEM_ACCESS_RW,
832 NV_MEM_TARGET_VRAM, &vram);
833 if (ret) {
834 NV_ERROR(dev, "Error creating VRAM ctxdma: %d\n", ret);
835 return ret;
836 }
837 }
838
839 ret = nouveau_ramht_insert(chan, vram_h, vram);
840 nouveau_gpuobj_ref(NULL, &vram);
841 if (ret) {
842 NV_ERROR(dev, "Error adding VRAM ctxdma to RAMHT: %d\n", ret);
843 return ret;
844 }
845
846 /* TT memory ctxdma */
847 if (dev_priv->card_type >= NV_50) {
848 ret = nouveau_gpuobj_dma_new(chan, NV_CLASS_DMA_IN_MEMORY,
849 0, (1ULL << 40), NV_MEM_ACCESS_RW,
850 NV_MEM_TARGET_VM, &tt);
851 } else {
852 ret = nouveau_gpuobj_dma_new(chan, NV_CLASS_DMA_IN_MEMORY,
853 0, dev_priv->gart_info.aper_size,
854 NV_MEM_ACCESS_RW,
855 NV_MEM_TARGET_GART, &tt);
856 }
857
858 if (ret) {
859 NV_ERROR(dev, "Error creating TT ctxdma: %d\n", ret);
860 return ret;
861 }
862
863 ret = nouveau_ramht_insert(chan, tt_h, tt);
864 nouveau_gpuobj_ref(NULL, &tt);
865 if (ret) {
866 NV_ERROR(dev, "Error adding TT ctxdma to RAMHT: %d\n", ret);
867 return ret;
868 }
869
870 return 0;
871 }
872
873 void
nouveau_gpuobj_channel_takedown(struct nouveau_channel * chan)874 nouveau_gpuobj_channel_takedown(struct nouveau_channel *chan)
875 {
876 struct drm_device *dev = chan->dev;
877 struct drm_nouveau_private *dev_priv = dev->dev_private;
878 int i;
879
880 NV_DEBUG(dev, "ch%d\n", chan->id);
881
882 if (dev_priv->card_type >= NV_D0) {
883 for (i = 0; i < dev->mode_config.num_crtc; i++) {
884 struct nouveau_bo *bo = nvd0_display_crtc_sema(dev, i);
885 nouveau_bo_vma_del(bo, &chan->dispc_vma[i]);
886 }
887 } else
888 if (dev_priv->card_type >= NV_50) {
889 struct nv50_display *disp = nv50_display(dev);
890 for (i = 0; i < dev->mode_config.num_crtc; i++) {
891 struct nv50_display_crtc *dispc = &disp->crtc[i];
892 nouveau_bo_vma_del(dispc->sem.bo, &chan->dispc_vma[i]);
893 }
894 }
895
896 nouveau_vm_ref(NULL, &chan->vm, chan->vm_pd);
897 nouveau_gpuobj_ref(NULL, &chan->vm_pd);
898
899 if (drm_mm_initialized(&chan->ramin_heap))
900 drm_mm_takedown(&chan->ramin_heap);
901 nouveau_gpuobj_ref(NULL, &chan->ramin);
902 }
903
904 int
nouveau_gpuobj_suspend(struct drm_device * dev)905 nouveau_gpuobj_suspend(struct drm_device *dev)
906 {
907 struct drm_nouveau_private *dev_priv = dev->dev_private;
908 struct nouveau_gpuobj *gpuobj;
909 int i;
910
911 list_for_each_entry(gpuobj, &dev_priv->gpuobj_list, list) {
912 if (gpuobj->cinst != NVOBJ_CINST_GLOBAL)
913 continue;
914
915 gpuobj->suspend = vmalloc(gpuobj->size);
916 if (!gpuobj->suspend) {
917 nouveau_gpuobj_resume(dev);
918 return -ENOMEM;
919 }
920
921 for (i = 0; i < gpuobj->size; i += 4)
922 gpuobj->suspend[i/4] = nv_ro32(gpuobj, i);
923 }
924
925 return 0;
926 }
927
928 void
nouveau_gpuobj_resume(struct drm_device * dev)929 nouveau_gpuobj_resume(struct drm_device *dev)
930 {
931 struct drm_nouveau_private *dev_priv = dev->dev_private;
932 struct nouveau_gpuobj *gpuobj;
933 int i;
934
935 list_for_each_entry(gpuobj, &dev_priv->gpuobj_list, list) {
936 if (!gpuobj->suspend)
937 continue;
938
939 for (i = 0; i < gpuobj->size; i += 4)
940 nv_wo32(gpuobj, i, gpuobj->suspend[i/4]);
941
942 vfree(gpuobj->suspend);
943 gpuobj->suspend = NULL;
944 }
945
946 dev_priv->engine.instmem.flush(dev);
947 }
948
nouveau_ioctl_grobj_alloc(struct drm_device * dev,void * data,struct drm_file * file_priv)949 int nouveau_ioctl_grobj_alloc(struct drm_device *dev, void *data,
950 struct drm_file *file_priv)
951 {
952 struct drm_nouveau_grobj_alloc *init = data;
953 struct nouveau_channel *chan;
954 int ret;
955
956 if (init->handle == ~0)
957 return -EINVAL;
958
959 chan = nouveau_channel_get(file_priv, init->channel);
960 if (IS_ERR(chan))
961 return PTR_ERR(chan);
962
963 if (nouveau_ramht_find(chan, init->handle)) {
964 ret = -EEXIST;
965 goto out;
966 }
967
968 ret = nouveau_gpuobj_gr_new(chan, init->handle, init->class);
969 if (ret) {
970 NV_ERROR(dev, "Error creating object: %d (%d/0x%08x)\n",
971 ret, init->channel, init->handle);
972 }
973
974 out:
975 nouveau_channel_put(&chan);
976 return ret;
977 }
978
nouveau_ioctl_gpuobj_free(struct drm_device * dev,void * data,struct drm_file * file_priv)979 int nouveau_ioctl_gpuobj_free(struct drm_device *dev, void *data,
980 struct drm_file *file_priv)
981 {
982 struct drm_nouveau_gpuobj_free *objfree = data;
983 struct nouveau_channel *chan;
984 int ret;
985
986 chan = nouveau_channel_get(file_priv, objfree->channel);
987 if (IS_ERR(chan))
988 return PTR_ERR(chan);
989
990 /* Synchronize with the user channel */
991 nouveau_channel_idle(chan);
992
993 ret = nouveau_ramht_remove(chan, objfree->handle);
994 nouveau_channel_put(&chan);
995 return ret;
996 }
997
998 u32
nv_ro32(struct nouveau_gpuobj * gpuobj,u32 offset)999 nv_ro32(struct nouveau_gpuobj *gpuobj, u32 offset)
1000 {
1001 struct drm_nouveau_private *dev_priv = gpuobj->dev->dev_private;
1002 struct drm_device *dev = gpuobj->dev;
1003 unsigned long flags;
1004
1005 if (gpuobj->pinst == ~0 || !dev_priv->ramin_available) {
1006 u64 ptr = gpuobj->vinst + offset;
1007 u32 base = ptr >> 16;
1008 u32 val;
1009
1010 spin_lock_irqsave(&dev_priv->vm_lock, flags);
1011 if (dev_priv->ramin_base != base) {
1012 dev_priv->ramin_base = base;
1013 nv_wr32(dev, 0x001700, dev_priv->ramin_base);
1014 }
1015 val = nv_rd32(dev, 0x700000 + (ptr & 0xffff));
1016 spin_unlock_irqrestore(&dev_priv->vm_lock, flags);
1017 return val;
1018 }
1019
1020 return nv_ri32(dev, gpuobj->pinst + offset);
1021 }
1022
1023 void
nv_wo32(struct nouveau_gpuobj * gpuobj,u32 offset,u32 val)1024 nv_wo32(struct nouveau_gpuobj *gpuobj, u32 offset, u32 val)
1025 {
1026 struct drm_nouveau_private *dev_priv = gpuobj->dev->dev_private;
1027 struct drm_device *dev = gpuobj->dev;
1028 unsigned long flags;
1029
1030 if (gpuobj->pinst == ~0 || !dev_priv->ramin_available) {
1031 u64 ptr = gpuobj->vinst + offset;
1032 u32 base = ptr >> 16;
1033
1034 spin_lock_irqsave(&dev_priv->vm_lock, flags);
1035 if (dev_priv->ramin_base != base) {
1036 dev_priv->ramin_base = base;
1037 nv_wr32(dev, 0x001700, dev_priv->ramin_base);
1038 }
1039 nv_wr32(dev, 0x700000 + (ptr & 0xffff), val);
1040 spin_unlock_irqrestore(&dev_priv->vm_lock, flags);
1041 return;
1042 }
1043
1044 nv_wi32(dev, gpuobj->pinst + offset, val);
1045 }
1046