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 (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, 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 static uint32_t
nouveau_gpuobj_class_instmem_size(struct drm_device * dev,int class)366 nouveau_gpuobj_class_instmem_size(struct drm_device *dev, int class)
367 {
368 	struct drm_nouveau_private *dev_priv = dev->dev_private;
369 
370 	/*XXX: dodgy hack for now */
371 	if (dev_priv->card_type >= NV_50)
372 		return 24;
373 	if (dev_priv->card_type >= NV_40)
374 		return 32;
375 	return 16;
376 }
377 
378 /*
379    DMA objects are used to reference a piece of memory in the
380    framebuffer, PCI or AGP address space. Each object is 16 bytes big
381    and looks as follows:
382 
383    entry[0]
384    11:0  class (seems like I can always use 0 here)
385    12    page table present?
386    13    page entry linear?
387    15:14 access: 0 rw, 1 ro, 2 wo
388    17:16 target: 0 NV memory, 1 NV memory tiled, 2 PCI, 3 AGP
389    31:20 dma adjust (bits 0-11 of the address)
390    entry[1]
391    dma limit (size of transfer)
392    entry[X]
393    1     0 readonly, 1 readwrite
394    31:12 dma frame address of the page (bits 12-31 of the address)
395    entry[N]
396    page table terminator, same value as the first pte, as does nvidia
397    rivatv uses 0xffffffff
398 
399    Non linear page tables need a list of frame addresses afterwards,
400    the rivatv project has some info on this.
401 
402    The method below creates a DMA object in instance RAM and returns a handle
403    to it that can be used to set up context objects.
404 */
405 
406 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)407 nv50_gpuobj_dma_init(struct nouveau_gpuobj *obj, u32 offset, int class,
408 		     u64 base, u64 size, int target, int access,
409 		     u32 type, u32 comp)
410 {
411 	struct drm_nouveau_private *dev_priv = obj->dev->dev_private;
412 	struct nouveau_instmem_engine *pinstmem = &dev_priv->engine.instmem;
413 	u32 flags0;
414 
415 	flags0  = (comp << 29) | (type << 22) | class;
416 	flags0 |= 0x00100000;
417 
418 	switch (access) {
419 	case NV_MEM_ACCESS_RO: flags0 |= 0x00040000; break;
420 	case NV_MEM_ACCESS_RW:
421 	case NV_MEM_ACCESS_WO: flags0 |= 0x00080000; break;
422 	default:
423 		break;
424 	}
425 
426 	switch (target) {
427 	case NV_MEM_TARGET_VRAM:
428 		flags0 |= 0x00010000;
429 		break;
430 	case NV_MEM_TARGET_PCI:
431 		flags0 |= 0x00020000;
432 		break;
433 	case NV_MEM_TARGET_PCI_NOSNOOP:
434 		flags0 |= 0x00030000;
435 		break;
436 	case NV_MEM_TARGET_GART:
437 		base += dev_priv->gart_info.aper_base;
438 	default:
439 		flags0 &= ~0x00100000;
440 		break;
441 	}
442 
443 	/* convert to base + limit */
444 	size = (base + size) - 1;
445 
446 	nv_wo32(obj, offset + 0x00, flags0);
447 	nv_wo32(obj, offset + 0x04, lower_32_bits(size));
448 	nv_wo32(obj, offset + 0x08, lower_32_bits(base));
449 	nv_wo32(obj, offset + 0x0c, upper_32_bits(size) << 24 |
450 				    upper_32_bits(base));
451 	nv_wo32(obj, offset + 0x10, 0x00000000);
452 	nv_wo32(obj, offset + 0x14, 0x00000000);
453 
454 	pinstmem->flush(obj->dev);
455 }
456 
457 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)458 nv50_gpuobj_dma_new(struct nouveau_channel *chan, int class, u64 base, u64 size,
459 		    int target, int access, u32 type, u32 comp,
460 		    struct nouveau_gpuobj **pobj)
461 {
462 	struct drm_device *dev = chan->dev;
463 	int ret;
464 
465 	ret = nouveau_gpuobj_new(dev, chan, 24, 16, NVOBJ_FLAG_ZERO_FREE, pobj);
466 	if (ret)
467 		return ret;
468 
469 	nv50_gpuobj_dma_init(*pobj, 0, class, base, size, target,
470 			     access, type, comp);
471 	return 0;
472 }
473 
474 int
nouveau_gpuobj_dma_new(struct nouveau_channel * chan,int class,u64 base,u64 size,int access,int target,struct nouveau_gpuobj ** pobj)475 nouveau_gpuobj_dma_new(struct nouveau_channel *chan, int class, u64 base,
476 		       u64 size, int access, int target,
477 		       struct nouveau_gpuobj **pobj)
478 {
479 	struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
480 	struct drm_device *dev = chan->dev;
481 	struct nouveau_gpuobj *obj;
482 	u32 flags0, flags2;
483 	int ret;
484 
485 	if (dev_priv->card_type >= NV_50) {
486 		u32 comp = (target == NV_MEM_TARGET_VM) ? NV_MEM_COMP_VM : 0;
487 		u32 type = (target == NV_MEM_TARGET_VM) ? NV_MEM_TYPE_VM : 0;
488 
489 		return nv50_gpuobj_dma_new(chan, class, base, size,
490 					   target, access, type, comp, pobj);
491 	}
492 
493 	if (target == NV_MEM_TARGET_GART) {
494 		struct nouveau_gpuobj *gart = dev_priv->gart_info.sg_ctxdma;
495 
496 		if (dev_priv->gart_info.type == NOUVEAU_GART_PDMA) {
497 			if (base == 0) {
498 				nouveau_gpuobj_ref(gart, pobj);
499 				return 0;
500 			}
501 
502 			base   = nouveau_sgdma_get_physical(dev, base);
503 			target = NV_MEM_TARGET_PCI;
504 		} else {
505 			base += dev_priv->gart_info.aper_base;
506 			if (dev_priv->gart_info.type == NOUVEAU_GART_AGP)
507 				target = NV_MEM_TARGET_PCI_NOSNOOP;
508 			else
509 				target = NV_MEM_TARGET_PCI;
510 		}
511 	}
512 
513 	flags0  = class;
514 	flags0 |= 0x00003000; /* PT present, PT linear */
515 	flags2  = 0;
516 
517 	switch (target) {
518 	case NV_MEM_TARGET_PCI:
519 		flags0 |= 0x00020000;
520 		break;
521 	case NV_MEM_TARGET_PCI_NOSNOOP:
522 		flags0 |= 0x00030000;
523 		break;
524 	default:
525 		break;
526 	}
527 
528 	switch (access) {
529 	case NV_MEM_ACCESS_RO:
530 		flags0 |= 0x00004000;
531 		break;
532 	case NV_MEM_ACCESS_WO:
533 		flags0 |= 0x00008000;
534 	default:
535 		flags2 |= 0x00000002;
536 		break;
537 	}
538 
539 	flags0 |= (base & 0x00000fff) << 20;
540 	flags2 |= (base & 0xfffff000);
541 
542 	ret = nouveau_gpuobj_new(dev, chan, 16, 16, NVOBJ_FLAG_ZERO_FREE, &obj);
543 	if (ret)
544 		return ret;
545 
546 	nv_wo32(obj, 0x00, flags0);
547 	nv_wo32(obj, 0x04, size - 1);
548 	nv_wo32(obj, 0x08, flags2);
549 	nv_wo32(obj, 0x0c, flags2);
550 
551 	obj->engine = NVOBJ_ENGINE_SW;
552 	obj->class  = class;
553 	*pobj = obj;
554 	return 0;
555 }
556 
557 /* Context objects in the instance RAM have the following structure.
558  * On NV40 they are 32 byte long, on NV30 and smaller 16 bytes.
559 
560    NV4 - NV30:
561 
562    entry[0]
563    11:0 class
564    12   chroma key enable
565    13   user clip enable
566    14   swizzle enable
567    17:15 patch config:
568        scrcopy_and, rop_and, blend_and, scrcopy, srccopy_pre, blend_pre
569    18   synchronize enable
570    19   endian: 1 big, 0 little
571    21:20 dither mode
572    23    single step enable
573    24    patch status: 0 invalid, 1 valid
574    25    context_surface 0: 1 valid
575    26    context surface 1: 1 valid
576    27    context pattern: 1 valid
577    28    context rop: 1 valid
578    29,30 context beta, beta4
579    entry[1]
580    7:0   mono format
581    15:8  color format
582    31:16 notify instance address
583    entry[2]
584    15:0  dma 0 instance address
585    31:16 dma 1 instance address
586    entry[3]
587    dma method traps
588 
589    NV40:
590    No idea what the exact format is. Here's what can be deducted:
591 
592    entry[0]:
593    11:0  class  (maybe uses more bits here?)
594    17    user clip enable
595    21:19 patch config
596    25    patch status valid ?
597    entry[1]:
598    15:0  DMA notifier  (maybe 20:0)
599    entry[2]:
600    15:0  DMA 0 instance (maybe 20:0)
601    24    big endian
602    entry[3]:
603    15:0  DMA 1 instance (maybe 20:0)
604    entry[4]:
605    entry[5]:
606    set to 0?
607 */
608 static int
nouveau_gpuobj_sw_new(struct nouveau_channel * chan,int class,struct nouveau_gpuobj ** gpuobj_ret)609 nouveau_gpuobj_sw_new(struct nouveau_channel *chan, int class,
610 		      struct nouveau_gpuobj **gpuobj_ret)
611 {
612 	struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
613 	struct nouveau_gpuobj *gpuobj;
614 
615 	gpuobj = kzalloc(sizeof(*gpuobj), GFP_KERNEL);
616 	if (!gpuobj)
617 		return -ENOMEM;
618 	gpuobj->dev = chan->dev;
619 	gpuobj->engine = NVOBJ_ENGINE_SW;
620 	gpuobj->class = class;
621 	kref_init(&gpuobj->refcount);
622 	gpuobj->cinst = 0x40;
623 
624 	spin_lock(&dev_priv->ramin_lock);
625 	list_add_tail(&gpuobj->list, &dev_priv->gpuobj_list);
626 	spin_unlock(&dev_priv->ramin_lock);
627 	*gpuobj_ret = gpuobj;
628 	return 0;
629 }
630 
631 int
nouveau_gpuobj_gr_new(struct nouveau_channel * chan,u32 handle,int class)632 nouveau_gpuobj_gr_new(struct nouveau_channel *chan, u32 handle, int class)
633 {
634 	struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
635 	struct drm_device *dev = chan->dev;
636 	struct nouveau_gpuobj_class *oc;
637 	struct nouveau_gpuobj *gpuobj;
638 	int ret;
639 
640 	NV_DEBUG(dev, "ch%d class=0x%04x\n", chan->id, class);
641 
642 	list_for_each_entry(oc, &dev_priv->classes, head) {
643 		if (oc->id == class)
644 			goto found;
645 	}
646 
647 	NV_ERROR(dev, "illegal object class: 0x%x\n", class);
648 	return -EINVAL;
649 
650 found:
651 	switch (oc->engine) {
652 	case NVOBJ_ENGINE_SW:
653 		if (dev_priv->card_type < NV_C0) {
654 			ret = nouveau_gpuobj_sw_new(chan, class, &gpuobj);
655 			if (ret)
656 				return ret;
657 			goto insert;
658 		}
659 		break;
660 	case NVOBJ_ENGINE_GR:
661 		if ((dev_priv->card_type >= NV_20 && !chan->ramin_grctx) ||
662 		    (dev_priv->card_type  < NV_20 && !chan->pgraph_ctx)) {
663 			struct nouveau_pgraph_engine *pgraph =
664 				&dev_priv->engine.graph;
665 
666 			ret = pgraph->create_context(chan);
667 			if (ret)
668 				return ret;
669 		}
670 		break;
671 	case NVOBJ_ENGINE_CRYPT:
672 		if (!chan->crypt_ctx) {
673 			struct nouveau_crypt_engine *pcrypt =
674 				&dev_priv->engine.crypt;
675 
676 			ret = pcrypt->create_context(chan);
677 			if (ret)
678 				return ret;
679 		}
680 		break;
681 	}
682 
683 	/* we're done if this is fermi */
684 	if (dev_priv->card_type >= NV_C0)
685 		return 0;
686 
687 	ret = nouveau_gpuobj_new(dev, chan,
688 				 nouveau_gpuobj_class_instmem_size(dev, class),
689 				 16,
690 				 NVOBJ_FLAG_ZERO_ALLOC | NVOBJ_FLAG_ZERO_FREE,
691 				 &gpuobj);
692 	if (ret) {
693 		NV_ERROR(dev, "error creating gpuobj: %d\n", ret);
694 		return ret;
695 	}
696 
697 	if (dev_priv->card_type >= NV_50) {
698 		nv_wo32(gpuobj,  0, class);
699 		nv_wo32(gpuobj, 20, 0x00010000);
700 	} else {
701 		switch (class) {
702 		case NV_CLASS_NULL:
703 			nv_wo32(gpuobj, 0, 0x00001030);
704 			nv_wo32(gpuobj, 4, 0xFFFFFFFF);
705 			break;
706 		default:
707 			if (dev_priv->card_type >= NV_40) {
708 				nv_wo32(gpuobj, 0, class);
709 #ifdef __BIG_ENDIAN
710 				nv_wo32(gpuobj, 8, 0x01000000);
711 #endif
712 			} else {
713 #ifdef __BIG_ENDIAN
714 				nv_wo32(gpuobj, 0, class | 0x00080000);
715 #else
716 				nv_wo32(gpuobj, 0, class);
717 #endif
718 			}
719 		}
720 	}
721 	dev_priv->engine.instmem.flush(dev);
722 
723 	gpuobj->engine = oc->engine;
724 	gpuobj->class  = oc->id;
725 
726 insert:
727 	ret = nouveau_ramht_insert(chan, handle, gpuobj);
728 	if (ret)
729 		NV_ERROR(dev, "error adding gpuobj to RAMHT: %d\n", ret);
730 	nouveau_gpuobj_ref(NULL, &gpuobj);
731 	return ret;
732 }
733 
734 static int
nouveau_gpuobj_channel_init_pramin(struct nouveau_channel * chan)735 nouveau_gpuobj_channel_init_pramin(struct nouveau_channel *chan)
736 {
737 	struct drm_device *dev = chan->dev;
738 	struct drm_nouveau_private *dev_priv = dev->dev_private;
739 	uint32_t size;
740 	uint32_t base;
741 	int ret;
742 
743 	NV_DEBUG(dev, "ch%d\n", chan->id);
744 
745 	/* Base amount for object storage (4KiB enough?) */
746 	size = 0x2000;
747 	base = 0;
748 
749 	/* PGRAPH context */
750 	size += dev_priv->engine.graph.grctx_size;
751 
752 	if (dev_priv->card_type == NV_50) {
753 		/* Various fixed table thingos */
754 		size += 0x1400; /* mostly unknown stuff */
755 		size += 0x4000; /* vm pd */
756 		base  = 0x6000;
757 		/* RAMHT, not sure about setting size yet, 32KiB to be safe */
758 		size += 0x8000;
759 		/* RAMFC */
760 		size += 0x1000;
761 	}
762 
763 	ret = nouveau_gpuobj_new(dev, NULL, size, 0x1000, 0, &chan->ramin);
764 	if (ret) {
765 		NV_ERROR(dev, "Error allocating channel PRAMIN: %d\n", ret);
766 		return ret;
767 	}
768 
769 	ret = drm_mm_init(&chan->ramin_heap, base, size);
770 	if (ret) {
771 		NV_ERROR(dev, "Error creating PRAMIN heap: %d\n", ret);
772 		nouveau_gpuobj_ref(NULL, &chan->ramin);
773 		return ret;
774 	}
775 
776 	return 0;
777 }
778 
779 int
nouveau_gpuobj_channel_init(struct nouveau_channel * chan,uint32_t vram_h,uint32_t tt_h)780 nouveau_gpuobj_channel_init(struct nouveau_channel *chan,
781 			    uint32_t vram_h, uint32_t tt_h)
782 {
783 	struct drm_device *dev = chan->dev;
784 	struct drm_nouveau_private *dev_priv = dev->dev_private;
785 	struct nouveau_gpuobj *vram = NULL, *tt = NULL;
786 	int ret, i;
787 
788 	NV_DEBUG(dev, "ch%d vram=0x%08x tt=0x%08x\n", chan->id, vram_h, tt_h);
789 
790 	if (dev_priv->card_type == NV_C0) {
791 		struct nouveau_vm *vm = dev_priv->chan_vm;
792 		struct nouveau_vm_pgd *vpgd;
793 
794 		ret = nouveau_gpuobj_new(dev, NULL, 4096, 0x1000, 0,
795 					 &chan->ramin);
796 		if (ret)
797 			return ret;
798 
799 		nouveau_vm_ref(vm, &chan->vm, NULL);
800 
801 		vpgd = list_first_entry(&vm->pgd_list, struct nouveau_vm_pgd, head);
802 		nv_wo32(chan->ramin, 0x0200, lower_32_bits(vpgd->obj->vinst));
803 		nv_wo32(chan->ramin, 0x0204, upper_32_bits(vpgd->obj->vinst));
804 		nv_wo32(chan->ramin, 0x0208, 0xffffffff);
805 		nv_wo32(chan->ramin, 0x020c, 0x000000ff);
806 		return 0;
807 	}
808 
809 	/* Allocate a chunk of memory for per-channel object storage */
810 	ret = nouveau_gpuobj_channel_init_pramin(chan);
811 	if (ret) {
812 		NV_ERROR(dev, "init pramin\n");
813 		return ret;
814 	}
815 
816 	/* NV50 VM
817 	 *  - Allocate per-channel page-directory
818 	 *  - Link with shared channel VM
819 	 */
820 	if (dev_priv->chan_vm) {
821 		u32 pgd_offs = (dev_priv->chipset == 0x50) ? 0x1400 : 0x0200;
822 		u64 vm_vinst = chan->ramin->vinst + pgd_offs;
823 		u32 vm_pinst = chan->ramin->pinst;
824 
825 		if (vm_pinst != ~0)
826 			vm_pinst += pgd_offs;
827 
828 		ret = nouveau_gpuobj_new_fake(dev, vm_pinst, vm_vinst, 0x4000,
829 					      0, &chan->vm_pd);
830 		if (ret)
831 			return ret;
832 
833 		nouveau_vm_ref(dev_priv->chan_vm, &chan->vm, chan->vm_pd);
834 	}
835 
836 	/* RAMHT */
837 	if (dev_priv->card_type < NV_50) {
838 		nouveau_ramht_ref(dev_priv->ramht, &chan->ramht, NULL);
839 	} else {
840 		struct nouveau_gpuobj *ramht = NULL;
841 
842 		ret = nouveau_gpuobj_new(dev, chan, 0x8000, 16,
843 					 NVOBJ_FLAG_ZERO_ALLOC, &ramht);
844 		if (ret)
845 			return ret;
846 
847 		ret = nouveau_ramht_new(dev, ramht, &chan->ramht);
848 		nouveau_gpuobj_ref(NULL, &ramht);
849 		if (ret)
850 			return ret;
851 
852 		/* dma objects for display sync channel semaphore blocks */
853 		for (i = 0; i < 2; i++) {
854 			struct nouveau_gpuobj *sem = NULL;
855 			struct nv50_display_crtc *dispc =
856 				&nv50_display(dev)->crtc[i];
857 			u64 offset = dispc->sem.bo->bo.mem.start << PAGE_SHIFT;
858 
859 			ret = nouveau_gpuobj_dma_new(chan, 0x3d, offset, 0xfff,
860 						     NV_MEM_ACCESS_RW,
861 						     NV_MEM_TARGET_VRAM, &sem);
862 			if (ret)
863 				return ret;
864 
865 			ret = nouveau_ramht_insert(chan, NvEvoSema0 + i, sem);
866 			nouveau_gpuobj_ref(NULL, &sem);
867 			if (ret)
868 				return ret;
869 		}
870 	}
871 
872 	/* VRAM ctxdma */
873 	if (dev_priv->card_type >= NV_50) {
874 		ret = nouveau_gpuobj_dma_new(chan, NV_CLASS_DMA_IN_MEMORY,
875 					     0, (1ULL << 40), NV_MEM_ACCESS_RW,
876 					     NV_MEM_TARGET_VM, &vram);
877 		if (ret) {
878 			NV_ERROR(dev, "Error creating VRAM ctxdma: %d\n", ret);
879 			return ret;
880 		}
881 	} else {
882 		ret = nouveau_gpuobj_dma_new(chan, NV_CLASS_DMA_IN_MEMORY,
883 					     0, dev_priv->fb_available_size,
884 					     NV_MEM_ACCESS_RW,
885 					     NV_MEM_TARGET_VRAM, &vram);
886 		if (ret) {
887 			NV_ERROR(dev, "Error creating VRAM ctxdma: %d\n", ret);
888 			return ret;
889 		}
890 	}
891 
892 	ret = nouveau_ramht_insert(chan, vram_h, vram);
893 	nouveau_gpuobj_ref(NULL, &vram);
894 	if (ret) {
895 		NV_ERROR(dev, "Error adding VRAM ctxdma to RAMHT: %d\n", ret);
896 		return ret;
897 	}
898 
899 	/* TT memory ctxdma */
900 	if (dev_priv->card_type >= NV_50) {
901 		ret = nouveau_gpuobj_dma_new(chan, NV_CLASS_DMA_IN_MEMORY,
902 					     0, (1ULL << 40), NV_MEM_ACCESS_RW,
903 					     NV_MEM_TARGET_VM, &tt);
904 	} else {
905 		ret = nouveau_gpuobj_dma_new(chan, NV_CLASS_DMA_IN_MEMORY,
906 					     0, dev_priv->gart_info.aper_size,
907 					     NV_MEM_ACCESS_RW,
908 					     NV_MEM_TARGET_GART, &tt);
909 	}
910 
911 	if (ret) {
912 		NV_ERROR(dev, "Error creating TT ctxdma: %d\n", ret);
913 		return ret;
914 	}
915 
916 	ret = nouveau_ramht_insert(chan, tt_h, tt);
917 	nouveau_gpuobj_ref(NULL, &tt);
918 	if (ret) {
919 		NV_ERROR(dev, "Error adding TT ctxdma to RAMHT: %d\n", ret);
920 		return ret;
921 	}
922 
923 	return 0;
924 }
925 
926 void
nouveau_gpuobj_channel_takedown(struct nouveau_channel * chan)927 nouveau_gpuobj_channel_takedown(struct nouveau_channel *chan)
928 {
929 	struct drm_device *dev = chan->dev;
930 
931 	NV_DEBUG(dev, "ch%d\n", chan->id);
932 
933 	nouveau_ramht_ref(NULL, &chan->ramht, chan);
934 
935 	nouveau_vm_ref(NULL, &chan->vm, chan->vm_pd);
936 	nouveau_gpuobj_ref(NULL, &chan->vm_pd);
937 
938 	if (drm_mm_initialized(&chan->ramin_heap))
939 		drm_mm_takedown(&chan->ramin_heap);
940 	nouveau_gpuobj_ref(NULL, &chan->ramin);
941 }
942 
943 int
nouveau_gpuobj_suspend(struct drm_device * dev)944 nouveau_gpuobj_suspend(struct drm_device *dev)
945 {
946 	struct drm_nouveau_private *dev_priv = dev->dev_private;
947 	struct nouveau_gpuobj *gpuobj;
948 	int i;
949 
950 	list_for_each_entry(gpuobj, &dev_priv->gpuobj_list, list) {
951 		if (gpuobj->cinst != NVOBJ_CINST_GLOBAL)
952 			continue;
953 
954 		gpuobj->suspend = vmalloc(gpuobj->size);
955 		if (!gpuobj->suspend) {
956 			nouveau_gpuobj_resume(dev);
957 			return -ENOMEM;
958 		}
959 
960 		for (i = 0; i < gpuobj->size; i += 4)
961 			gpuobj->suspend[i/4] = nv_ro32(gpuobj, i);
962 	}
963 
964 	return 0;
965 }
966 
967 void
nouveau_gpuobj_resume(struct drm_device * dev)968 nouveau_gpuobj_resume(struct drm_device *dev)
969 {
970 	struct drm_nouveau_private *dev_priv = dev->dev_private;
971 	struct nouveau_gpuobj *gpuobj;
972 	int i;
973 
974 	list_for_each_entry(gpuobj, &dev_priv->gpuobj_list, list) {
975 		if (!gpuobj->suspend)
976 			continue;
977 
978 		for (i = 0; i < gpuobj->size; i += 4)
979 			nv_wo32(gpuobj, i, gpuobj->suspend[i/4]);
980 
981 		vfree(gpuobj->suspend);
982 		gpuobj->suspend = NULL;
983 	}
984 
985 	dev_priv->engine.instmem.flush(dev);
986 }
987 
nouveau_ioctl_grobj_alloc(struct drm_device * dev,void * data,struct drm_file * file_priv)988 int nouveau_ioctl_grobj_alloc(struct drm_device *dev, void *data,
989 			      struct drm_file *file_priv)
990 {
991 	struct drm_nouveau_grobj_alloc *init = data;
992 	struct nouveau_channel *chan;
993 	int ret;
994 
995 	if (init->handle == ~0)
996 		return -EINVAL;
997 
998 	chan = nouveau_channel_get(dev, file_priv, init->channel);
999 	if (IS_ERR(chan))
1000 		return PTR_ERR(chan);
1001 
1002 	if (nouveau_ramht_find(chan, init->handle)) {
1003 		ret = -EEXIST;
1004 		goto out;
1005 	}
1006 
1007 	ret = nouveau_gpuobj_gr_new(chan, init->handle, init->class);
1008 	if (ret) {
1009 		NV_ERROR(dev, "Error creating object: %d (%d/0x%08x)\n",
1010 			 ret, init->channel, init->handle);
1011 	}
1012 
1013 out:
1014 	nouveau_channel_put(&chan);
1015 	return ret;
1016 }
1017 
nouveau_ioctl_gpuobj_free(struct drm_device * dev,void * data,struct drm_file * file_priv)1018 int nouveau_ioctl_gpuobj_free(struct drm_device *dev, void *data,
1019 			      struct drm_file *file_priv)
1020 {
1021 	struct drm_nouveau_gpuobj_free *objfree = data;
1022 	struct nouveau_channel *chan;
1023 	int ret;
1024 
1025 	chan = nouveau_channel_get(dev, file_priv, objfree->channel);
1026 	if (IS_ERR(chan))
1027 		return PTR_ERR(chan);
1028 
1029 	/* Synchronize with the user channel */
1030 	nouveau_channel_idle(chan);
1031 
1032 	ret = nouveau_ramht_remove(chan, objfree->handle);
1033 	nouveau_channel_put(&chan);
1034 	return ret;
1035 }
1036 
1037 u32
nv_ro32(struct nouveau_gpuobj * gpuobj,u32 offset)1038 nv_ro32(struct nouveau_gpuobj *gpuobj, u32 offset)
1039 {
1040 	struct drm_nouveau_private *dev_priv = gpuobj->dev->dev_private;
1041 	struct drm_device *dev = gpuobj->dev;
1042 	unsigned long flags;
1043 
1044 	if (gpuobj->pinst == ~0 || !dev_priv->ramin_available) {
1045 		u64  ptr = gpuobj->vinst + offset;
1046 		u32 base = ptr >> 16;
1047 		u32  val;
1048 
1049 		spin_lock_irqsave(&dev_priv->vm_lock, flags);
1050 		if (dev_priv->ramin_base != base) {
1051 			dev_priv->ramin_base = base;
1052 			nv_wr32(dev, 0x001700, dev_priv->ramin_base);
1053 		}
1054 		val = nv_rd32(dev, 0x700000 + (ptr & 0xffff));
1055 		spin_unlock_irqrestore(&dev_priv->vm_lock, flags);
1056 		return val;
1057 	}
1058 
1059 	return nv_ri32(dev, gpuobj->pinst + offset);
1060 }
1061 
1062 void
nv_wo32(struct nouveau_gpuobj * gpuobj,u32 offset,u32 val)1063 nv_wo32(struct nouveau_gpuobj *gpuobj, u32 offset, u32 val)
1064 {
1065 	struct drm_nouveau_private *dev_priv = gpuobj->dev->dev_private;
1066 	struct drm_device *dev = gpuobj->dev;
1067 	unsigned long flags;
1068 
1069 	if (gpuobj->pinst == ~0 || !dev_priv->ramin_available) {
1070 		u64  ptr = gpuobj->vinst + offset;
1071 		u32 base = ptr >> 16;
1072 
1073 		spin_lock_irqsave(&dev_priv->vm_lock, flags);
1074 		if (dev_priv->ramin_base != base) {
1075 			dev_priv->ramin_base = base;
1076 			nv_wr32(dev, 0x001700, dev_priv->ramin_base);
1077 		}
1078 		nv_wr32(dev, 0x700000 + (ptr & 0xffff), val);
1079 		spin_unlock_irqrestore(&dev_priv->vm_lock, flags);
1080 		return;
1081 	}
1082 
1083 	nv_wi32(dev, gpuobj->pinst + offset, val);
1084 }
1085