xref: /linux/drivers/gpu/drm/nouveau/nouveau_chan.c (revision a4871e6201c46c8e1d04308265b4b4c5753c8209)
1 /*
2  * Copyright 2012 Red Hat Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  * Authors: Ben Skeggs
23  */
24 #include <nvif/push006c.h>
25 
26 #include <nvif/class.h>
27 #include <nvif/cl0002.h>
28 #include <nvif/if0020.h>
29 
30 #include "nouveau_drv.h"
31 #include "nouveau_dma.h"
32 #include "nouveau_bo.h"
33 #include "nouveau_chan.h"
34 #include "nouveau_fence.h"
35 #include "nouveau_abi16.h"
36 #include "nouveau_vmm.h"
37 #include "nouveau_svm.h"
38 
39 MODULE_PARM_DESC(vram_pushbuf, "Create DMA push buffers in VRAM");
40 int nouveau_vram_pushbuf;
41 module_param_named(vram_pushbuf, nouveau_vram_pushbuf, int, 0400);
42 
43 void
44 nouveau_channel_kill(struct nouveau_channel *chan)
45 {
46 	atomic_set(&chan->killed, 1);
47 	if (chan->fence)
48 		nouveau_fence_context_kill(chan->fence, -ENODEV);
49 }
50 
51 static int
52 nouveau_channel_killed(struct nvif_event *event, void *repv, u32 repc)
53 {
54 	struct nouveau_channel *chan = container_of(event, typeof(*chan), kill);
55 	struct nouveau_cli *cli = chan->cli;
56 
57 	NV_PRINTK(warn, cli, "channel %d killed!\n", chan->chid);
58 
59 	if (unlikely(!atomic_read(&chan->killed)))
60 		nouveau_channel_kill(chan);
61 
62 	return NVIF_EVENT_DROP;
63 }
64 
65 int
66 nouveau_channel_idle(struct nouveau_channel *chan)
67 {
68 	if (likely(chan && chan->fence && !atomic_read(&chan->killed))) {
69 		struct nouveau_cli *cli = chan->cli;
70 		struct nouveau_fence *fence = NULL;
71 		int ret;
72 
73 		ret = nouveau_fence_new(&fence, chan);
74 		if (!ret) {
75 			ret = nouveau_fence_wait(fence, false, false);
76 			nouveau_fence_unref(&fence);
77 		}
78 
79 		if (ret) {
80 			NV_PRINTK(err, cli, "failed to idle channel %d [%s]\n",
81 				  chan->chid, cli->name);
82 			return ret;
83 		}
84 	}
85 	return 0;
86 }
87 
88 void
89 nouveau_channel_del(struct nouveau_channel **pchan)
90 {
91 	struct nouveau_channel *chan = *pchan;
92 	if (chan) {
93 		if (chan->fence)
94 			nouveau_fence(chan->cli->drm)->context_del(chan);
95 
96 		if (nvif_object_constructed(&chan->user))
97 			nouveau_svmm_part(chan->vmm->svmm, chan->inst);
98 
99 		nvif_object_dtor(&chan->blit);
100 		nvif_object_dtor(&chan->nvsw);
101 		nvif_object_dtor(&chan->gart);
102 		nvif_object_dtor(&chan->vram);
103 		nvif_event_dtor(&chan->kill);
104 		nvif_object_dtor(&chan->user);
105 		nvif_mem_dtor(&chan->mem_userd);
106 		nvif_object_dtor(&chan->push.ctxdma);
107 		nouveau_vma_del(&chan->push.vma);
108 		nouveau_bo_unmap(chan->push.buffer);
109 		if (chan->push.buffer && chan->push.buffer->bo.pin_count)
110 			nouveau_bo_unpin(chan->push.buffer);
111 		nouveau_bo_fini(chan->push.buffer);
112 		kfree(chan);
113 	}
114 	*pchan = NULL;
115 }
116 
117 static void
118 nouveau_channel_kick(struct nvif_push *push)
119 {
120 	struct nouveau_channel *chan = container_of(push, typeof(*chan), chan.push);
121 	chan->dma.cur = chan->dma.cur + (chan->chan.push.cur - chan->chan.push.bgn);
122 	FIRE_RING(chan);
123 	chan->chan.push.bgn = chan->chan.push.cur;
124 }
125 
126 static int
127 nouveau_channel_wait(struct nvif_push *push, u32 size)
128 {
129 	struct nouveau_channel *chan = container_of(push, typeof(*chan), chan.push);
130 	int ret;
131 	chan->dma.cur = chan->dma.cur + (chan->chan.push.cur - chan->chan.push.bgn);
132 	ret = RING_SPACE(chan, size);
133 	if (ret == 0) {
134 		chan->chan.push.bgn = chan->chan.push.mem.object.map.ptr;
135 		chan->chan.push.bgn = chan->chan.push.bgn + chan->dma.cur;
136 		chan->chan.push.cur = chan->chan.push.bgn;
137 		chan->chan.push.end = chan->chan.push.bgn + size;
138 	}
139 	return ret;
140 }
141 
142 static int
143 nouveau_channel_prep(struct nouveau_cli *cli,
144 		     u32 size, struct nouveau_channel **pchan)
145 {
146 	struct nouveau_drm *drm = cli->drm;
147 	struct nvif_device *device = &cli->device;
148 	struct nv_dma_v0 args = {};
149 	struct nouveau_channel *chan;
150 	u32 target;
151 	int ret;
152 
153 	chan = *pchan = kzalloc(sizeof(*chan), GFP_KERNEL);
154 	if (!chan)
155 		return -ENOMEM;
156 
157 	chan->cli = cli;
158 	chan->vmm = nouveau_cli_vmm(cli);
159 	atomic_set(&chan->killed, 0);
160 
161 	/* allocate memory for dma push buffer */
162 	target = NOUVEAU_GEM_DOMAIN_GART | NOUVEAU_GEM_DOMAIN_COHERENT;
163 	if (nouveau_vram_pushbuf)
164 		target = NOUVEAU_GEM_DOMAIN_VRAM;
165 
166 	ret = nouveau_bo_new(cli, size, 0, target, 0, 0, NULL, NULL,
167 			    &chan->push.buffer);
168 	if (ret == 0) {
169 		ret = nouveau_bo_pin(chan->push.buffer, target, false);
170 		if (ret == 0)
171 			ret = nouveau_bo_map(chan->push.buffer);
172 	}
173 
174 	if (ret) {
175 		nouveau_channel_del(pchan);
176 		return ret;
177 	}
178 
179 	chan->chan.push.mem.object.parent = cli->base.object.parent;
180 	chan->chan.push.mem.object.client = &cli->base;
181 	chan->chan.push.mem.object.name = "chanPush";
182 	chan->chan.push.mem.object.map.ptr = chan->push.buffer->kmap.virtual;
183 	chan->chan.push.wait = nouveau_channel_wait;
184 	chan->chan.push.kick = nouveau_channel_kick;
185 
186 	/* create dma object covering the *entire* memory space that the
187 	 * pushbuf lives in, this is because the GEM code requires that
188 	 * we be able to call out to other (indirect) push buffers
189 	 */
190 	chan->push.addr = chan->push.buffer->offset;
191 
192 	if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
193 		ret = nouveau_vma_new(chan->push.buffer, chan->vmm,
194 				      &chan->push.vma);
195 		if (ret) {
196 			nouveau_channel_del(pchan);
197 			return ret;
198 		}
199 
200 		chan->push.addr = chan->push.vma->addr;
201 
202 		if (device->info.family >= NV_DEVICE_INFO_V0_FERMI)
203 			return 0;
204 
205 		args.target = NV_DMA_V0_TARGET_VM;
206 		args.access = NV_DMA_V0_ACCESS_VM;
207 		args.start = 0;
208 		args.limit = chan->vmm->vmm.limit - 1;
209 	} else
210 	if (chan->push.buffer->bo.resource->mem_type == TTM_PL_VRAM) {
211 		if (device->info.family == NV_DEVICE_INFO_V0_TNT) {
212 			/* nv04 vram pushbuf hack, retarget to its location in
213 			 * the framebuffer bar rather than direct vram access..
214 			 * nfi why this exists, it came from the -nv ddx.
215 			 */
216 			args.target = NV_DMA_V0_TARGET_PCI;
217 			args.access = NV_DMA_V0_ACCESS_RDWR;
218 			args.start = nvxx_device(drm)->func->resource_addr(nvxx_device(drm), 1);
219 			args.limit = args.start + device->info.ram_user - 1;
220 		} else {
221 			args.target = NV_DMA_V0_TARGET_VRAM;
222 			args.access = NV_DMA_V0_ACCESS_RDWR;
223 			args.start = 0;
224 			args.limit = device->info.ram_user - 1;
225 		}
226 	} else {
227 		if (drm->agp.bridge) {
228 			args.target = NV_DMA_V0_TARGET_AGP;
229 			args.access = NV_DMA_V0_ACCESS_RDWR;
230 			args.start = drm->agp.base;
231 			args.limit = drm->agp.base + drm->agp.size - 1;
232 		} else {
233 			args.target = NV_DMA_V0_TARGET_VM;
234 			args.access = NV_DMA_V0_ACCESS_RDWR;
235 			args.start = 0;
236 			args.limit = chan->vmm->vmm.limit - 1;
237 		}
238 	}
239 
240 	ret = nvif_object_ctor(&device->object, "abi16PushCtxDma", 0,
241 			       NV_DMA_FROM_MEMORY, &args, sizeof(args),
242 			       &chan->push.ctxdma);
243 	if (ret) {
244 		nouveau_channel_del(pchan);
245 		return ret;
246 	}
247 
248 	return 0;
249 }
250 
251 static int
252 nouveau_channel_ctor(struct nouveau_cli *cli, bool priv, u64 runm,
253 		     struct nouveau_channel **pchan)
254 {
255 	const struct nvif_mclass hosts[] = {
256 		{  AMPERE_CHANNEL_GPFIFO_B, 0 },
257 		{  AMPERE_CHANNEL_GPFIFO_A, 0 },
258 		{  TURING_CHANNEL_GPFIFO_A, 0 },
259 		{   VOLTA_CHANNEL_GPFIFO_A, 0 },
260 		{  PASCAL_CHANNEL_GPFIFO_A, 0 },
261 		{ MAXWELL_CHANNEL_GPFIFO_A, 0 },
262 		{  KEPLER_CHANNEL_GPFIFO_B, 0 },
263 		{  KEPLER_CHANNEL_GPFIFO_A, 0 },
264 		{   FERMI_CHANNEL_GPFIFO  , 0 },
265 		{     G82_CHANNEL_GPFIFO  , 0 },
266 		{    NV50_CHANNEL_GPFIFO  , 0 },
267 		{    NV40_CHANNEL_DMA     , 0 },
268 		{    NV17_CHANNEL_DMA     , 0 },
269 		{    NV10_CHANNEL_DMA     , 0 },
270 		{    NV03_CHANNEL_DMA     , 0 },
271 		{}
272 	};
273 	DEFINE_RAW_FLEX(struct nvif_chan_v0, args, name, TASK_COMM_LEN + 16);
274 	struct nvif_device *device = &cli->device;
275 	struct nouveau_channel *chan;
276 	const u64 plength = 0x10000;
277 	const u64 ioffset = plength;
278 	const u64 ilength = 0x02000;
279 	int cid, ret;
280 	u64 size;
281 
282 	cid = nvif_mclass(&device->object, hosts);
283 	if (cid < 0)
284 		return cid;
285 
286 	if (hosts[cid].oclass < NV50_CHANNEL_GPFIFO)
287 		size = plength;
288 	else
289 		size = ioffset + ilength;
290 
291 	/* allocate dma push buffer */
292 	ret = nouveau_channel_prep(cli, size, &chan);
293 	*pchan = chan;
294 	if (ret)
295 		return ret;
296 
297 	/* create channel object */
298 	args->version = 0;
299 	args->namelen = __member_size(args->name);
300 	args->runlist = __ffs64(runm);
301 	args->runq = 0;
302 	args->priv = priv;
303 	args->devm = BIT(0);
304 	if (hosts[cid].oclass < NV50_CHANNEL_GPFIFO) {
305 		args->vmm = 0;
306 		args->ctxdma = nvif_handle(&chan->push.ctxdma);
307 		args->offset = chan->push.addr;
308 		args->length = 0;
309 	} else {
310 		args->vmm = nvif_handle(&chan->vmm->vmm.object);
311 		if (hosts[cid].oclass < FERMI_CHANNEL_GPFIFO)
312 			args->ctxdma = nvif_handle(&chan->push.ctxdma);
313 		else
314 			args->ctxdma = 0;
315 		args->offset = ioffset + chan->push.addr;
316 		args->length = ilength;
317 	}
318 	args->huserd = 0;
319 	args->ouserd = 0;
320 
321 	/* allocate userd */
322 	if (hosts[cid].oclass >= VOLTA_CHANNEL_GPFIFO_A) {
323 		ret = nvif_mem_ctor(&cli->mmu, "abi16ChanUSERD", NVIF_CLASS_MEM_GF100,
324 				    NVIF_MEM_VRAM | NVIF_MEM_COHERENT | NVIF_MEM_MAPPABLE,
325 				    0, PAGE_SIZE, NULL, 0, &chan->mem_userd);
326 		if (ret)
327 			return ret;
328 
329 		args->huserd = nvif_handle(&chan->mem_userd.object);
330 		args->ouserd = 0;
331 
332 		chan->userd = &chan->mem_userd.object;
333 	} else {
334 		chan->userd = &chan->user;
335 	}
336 
337 	snprintf(args->name, __member_size(args->name), "%s[%d]",
338 		 current->comm, task_pid_nr(current));
339 
340 	ret = nvif_object_ctor(&device->object, "abi16ChanUser", 0, hosts[cid].oclass,
341 			       args, __struct_size(args), &chan->user);
342 	if (ret) {
343 		nouveau_channel_del(pchan);
344 		return ret;
345 	}
346 
347 	chan->runlist = args->runlist;
348 	chan->chid = args->chid;
349 	chan->inst = args->inst;
350 	chan->token = args->token;
351 	return 0;
352 }
353 
354 static int
355 nouveau_channel_init(struct nouveau_channel *chan, u32 vram, u32 gart)
356 {
357 	struct nouveau_cli *cli = chan->cli;
358 	struct nouveau_drm *drm = cli->drm;
359 	struct nvif_device *device = &cli->device;
360 	struct nv_dma_v0 args = {};
361 	int ret, i;
362 
363 	ret = nvif_object_map(chan->userd, NULL, 0);
364 	if (ret)
365 		return ret;
366 
367 	if (chan->user.oclass >= FERMI_CHANNEL_GPFIFO) {
368 		DEFINE_RAW_FLEX(struct nvif_event_v0, args, data,
369 				sizeof(struct nvif_chan_event_v0));
370 		struct nvif_chan_event_v0 *host =
371 				(struct nvif_chan_event_v0 *)args->data;
372 
373 		host->version = 0;
374 		host->type = NVIF_CHAN_EVENT_V0_KILLED;
375 
376 		ret = nvif_event_ctor(&chan->user, "abi16ChanKilled", chan->chid,
377 				      nouveau_channel_killed, false,
378 				      args, __struct_size(args), &chan->kill);
379 		if (ret == 0)
380 			ret = nvif_event_allow(&chan->kill);
381 		if (ret) {
382 			NV_ERROR(drm, "Failed to request channel kill "
383 				      "notification: %d\n", ret);
384 			return ret;
385 		}
386 	}
387 
388 	/* allocate dma objects to cover all allowed vram, and gart */
389 	if (device->info.family < NV_DEVICE_INFO_V0_FERMI) {
390 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
391 			args.target = NV_DMA_V0_TARGET_VM;
392 			args.access = NV_DMA_V0_ACCESS_VM;
393 			args.start = 0;
394 			args.limit = chan->vmm->vmm.limit - 1;
395 		} else {
396 			args.target = NV_DMA_V0_TARGET_VRAM;
397 			args.access = NV_DMA_V0_ACCESS_RDWR;
398 			args.start = 0;
399 			args.limit = device->info.ram_user - 1;
400 		}
401 
402 		ret = nvif_object_ctor(&chan->user, "abi16ChanVramCtxDma", vram,
403 				       NV_DMA_IN_MEMORY, &args, sizeof(args),
404 				       &chan->vram);
405 		if (ret)
406 			return ret;
407 
408 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
409 			args.target = NV_DMA_V0_TARGET_VM;
410 			args.access = NV_DMA_V0_ACCESS_VM;
411 			args.start = 0;
412 			args.limit = chan->vmm->vmm.limit - 1;
413 		} else
414 		if (drm->agp.bridge) {
415 			args.target = NV_DMA_V0_TARGET_AGP;
416 			args.access = NV_DMA_V0_ACCESS_RDWR;
417 			args.start = drm->agp.base;
418 			args.limit = drm->agp.base + drm->agp.size - 1;
419 		} else {
420 			args.target = NV_DMA_V0_TARGET_VM;
421 			args.access = NV_DMA_V0_ACCESS_RDWR;
422 			args.start = 0;
423 			args.limit = chan->vmm->vmm.limit - 1;
424 		}
425 
426 		ret = nvif_object_ctor(&chan->user, "abi16ChanGartCtxDma", gart,
427 				       NV_DMA_IN_MEMORY, &args, sizeof(args),
428 				       &chan->gart);
429 		if (ret)
430 			return ret;
431 	}
432 
433 	/* initialise dma tracking parameters */
434 	switch (chan->user.oclass) {
435 	case NV03_CHANNEL_DMA:
436 	case NV10_CHANNEL_DMA:
437 	case NV17_CHANNEL_DMA:
438 	case NV40_CHANNEL_DMA:
439 		chan->user_put = 0x40;
440 		chan->user_get = 0x44;
441 		chan->dma.max = (0x10000 / 4) - 2;
442 		break;
443 	default:
444 		chan->user_put = 0x40;
445 		chan->user_get = 0x44;
446 		chan->user_get_hi = 0x60;
447 		chan->dma.ib_base =  0x10000 / 4;
448 		chan->dma.ib_max  = NV50_DMA_IB_MAX;
449 		chan->dma.ib_put  = 0;
450 		chan->dma.ib_free = chan->dma.ib_max - chan->dma.ib_put;
451 		chan->dma.max = chan->dma.ib_base;
452 		break;
453 	}
454 
455 	chan->dma.put = 0;
456 	chan->dma.cur = chan->dma.put;
457 	chan->dma.free = chan->dma.max - chan->dma.cur;
458 
459 	ret = PUSH_WAIT(&chan->chan.push, NOUVEAU_DMA_SKIPS);
460 	if (ret)
461 		return ret;
462 
463 	for (i = 0; i < NOUVEAU_DMA_SKIPS; i++)
464 		PUSH_DATA(&chan->chan.push, 0x00000000);
465 
466 	/* allocate software object class (used for fences on <= nv05) */
467 	if (device->info.family < NV_DEVICE_INFO_V0_CELSIUS) {
468 		ret = nvif_object_ctor(&chan->user, "abi16NvswFence", 0x006e,
469 				       NVIF_CLASS_SW_NV04,
470 				       NULL, 0, &chan->nvsw);
471 		if (ret)
472 			return ret;
473 
474 		ret = PUSH_WAIT(&chan->chan.push, 2);
475 		if (ret)
476 			return ret;
477 
478 		PUSH_NVSQ(&chan->chan.push, NV_SW, 0x0000, chan->nvsw.handle);
479 		PUSH_KICK(&chan->chan.push);
480 	}
481 
482 	/* initialise synchronisation */
483 	return nouveau_fence(drm)->context_new(chan);
484 }
485 
486 int
487 nouveau_channel_new(struct nouveau_cli *cli,
488 		    bool priv, u64 runm, u32 vram, u32 gart, struct nouveau_channel **pchan)
489 {
490 	int ret;
491 
492 	ret = nouveau_channel_ctor(cli, priv, runm, pchan);
493 	if (ret) {
494 		NV_PRINTK(dbg, cli, "channel create, %d\n", ret);
495 		return ret;
496 	}
497 
498 	ret = nouveau_channel_init(*pchan, vram, gart);
499 	if (ret) {
500 		NV_PRINTK(err, cli, "channel failed to initialise, %d\n", ret);
501 		nouveau_channel_del(pchan);
502 		return ret;
503 	}
504 
505 	ret = nouveau_svmm_join((*pchan)->vmm->svmm, (*pchan)->inst);
506 	if (ret)
507 		nouveau_channel_del(pchan);
508 
509 	return ret;
510 }
511 
512 void
513 nouveau_channels_fini(struct nouveau_drm *drm)
514 {
515 	kfree(drm->runl);
516 }
517 
518 int
519 nouveau_channels_init(struct nouveau_drm *drm)
520 {
521 	DEFINE_RAW_FLEX(struct nv_device_info_v1, args, data, 2);
522 	struct nv_device_info_v1_data *channels = &args->data[0];
523 	struct nv_device_info_v1_data *runlists = &args->data[1];
524 	struct nvif_object *device = &drm->client.device.object;
525 	int ret, i;
526 
527 	args->version = 1;
528 	args->count = __member_size(args->data) / sizeof(*args->data);
529 	channels->mthd = NV_DEVICE_HOST_CHANNELS;
530 	runlists->mthd = NV_DEVICE_HOST_RUNLISTS;
531 
532 	ret = nvif_object_mthd(device, NV_DEVICE_V0_INFO, args,
533 			       __struct_size(args));
534 	if (ret ||
535 	    runlists->mthd == NV_DEVICE_INFO_INVALID || !runlists->data ||
536 	    channels->mthd == NV_DEVICE_INFO_INVALID)
537 		return -ENODEV;
538 
539 	drm->chan_nr = drm->chan_total = channels->data;
540 	drm->runl_nr = fls64(runlists->data);
541 	drm->runl = kcalloc(drm->runl_nr, sizeof(*drm->runl), GFP_KERNEL);
542 	if (!drm->runl)
543 		return -ENOMEM;
544 
545 	if (drm->chan_nr == 0) {
546 		for (i = 0; i < drm->runl_nr; i++) {
547 			if (!(runlists->data & BIT(i)))
548 				continue;
549 
550 			channels->mthd = NV_DEVICE_HOST_RUNLIST_CHANNELS;
551 			channels->data = i;
552 
553 			ret = nvif_object_mthd(device, NV_DEVICE_V0_INFO, args,
554 					       __struct_size(args));
555 			if (ret || channels->mthd == NV_DEVICE_INFO_INVALID)
556 				return -ENODEV;
557 
558 			drm->runl[i].chan_nr = channels->data;
559 			drm->runl[i].chan_id_base = drm->chan_total;
560 			drm->runl[i].context_base = dma_fence_context_alloc(drm->runl[i].chan_nr);
561 
562 			drm->chan_total += drm->runl[i].chan_nr;
563 		}
564 	} else {
565 		drm->runl[0].context_base = dma_fence_context_alloc(drm->chan_nr);
566 		for (i = 1; i < drm->runl_nr; i++)
567 			drm->runl[i].context_base = drm->runl[0].context_base;
568 
569 	}
570 
571 	return 0;
572 }
573