xref: /linux/drivers/gpu/drm/nouveau/nouveau_chan.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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 	struct {
274 		struct nvif_chan_v0 chan;
275 		char name[TASK_COMM_LEN+16];
276 	} args;
277 	struct nvif_device *device = &cli->device;
278 	struct nouveau_channel *chan;
279 	const u64 plength = 0x10000;
280 	const u64 ioffset = plength;
281 	const u64 ilength = 0x02000;
282 	char name[TASK_COMM_LEN];
283 	int cid, ret;
284 	u64 size;
285 
286 	cid = nvif_mclass(&device->object, hosts);
287 	if (cid < 0)
288 		return cid;
289 
290 	if (hosts[cid].oclass < NV50_CHANNEL_GPFIFO)
291 		size = plength;
292 	else
293 		size = ioffset + ilength;
294 
295 	/* allocate dma push buffer */
296 	ret = nouveau_channel_prep(cli, size, &chan);
297 	*pchan = chan;
298 	if (ret)
299 		return ret;
300 
301 	/* create channel object */
302 	args.chan.version = 0;
303 	args.chan.namelen = sizeof(args.name);
304 	args.chan.runlist = __ffs64(runm);
305 	args.chan.runq = 0;
306 	args.chan.priv = priv;
307 	args.chan.devm = BIT(0);
308 	if (hosts[cid].oclass < NV50_CHANNEL_GPFIFO) {
309 		args.chan.vmm = 0;
310 		args.chan.ctxdma = nvif_handle(&chan->push.ctxdma);
311 		args.chan.offset = chan->push.addr;
312 		args.chan.length = 0;
313 	} else {
314 		args.chan.vmm = nvif_handle(&chan->vmm->vmm.object);
315 		if (hosts[cid].oclass < FERMI_CHANNEL_GPFIFO)
316 			args.chan.ctxdma = nvif_handle(&chan->push.ctxdma);
317 		else
318 			args.chan.ctxdma = 0;
319 		args.chan.offset = ioffset + chan->push.addr;
320 		args.chan.length = ilength;
321 	}
322 	args.chan.huserd = 0;
323 	args.chan.ouserd = 0;
324 
325 	/* allocate userd */
326 	if (hosts[cid].oclass >= VOLTA_CHANNEL_GPFIFO_A) {
327 		ret = nvif_mem_ctor(&cli->mmu, "abi16ChanUSERD", NVIF_CLASS_MEM_GF100,
328 				    NVIF_MEM_VRAM | NVIF_MEM_COHERENT | NVIF_MEM_MAPPABLE,
329 				    0, PAGE_SIZE, NULL, 0, &chan->mem_userd);
330 		if (ret)
331 			return ret;
332 
333 		args.chan.huserd = nvif_handle(&chan->mem_userd.object);
334 		args.chan.ouserd = 0;
335 
336 		chan->userd = &chan->mem_userd.object;
337 	} else {
338 		chan->userd = &chan->user;
339 	}
340 
341 	get_task_comm(name, current);
342 	snprintf(args.name, sizeof(args.name), "%s[%d]", name, task_pid_nr(current));
343 
344 	ret = nvif_object_ctor(&device->object, "abi16ChanUser", 0, hosts[cid].oclass,
345 			       &args, sizeof(args), &chan->user);
346 	if (ret) {
347 		nouveau_channel_del(pchan);
348 		return ret;
349 	}
350 
351 	chan->runlist = args.chan.runlist;
352 	chan->chid = args.chan.chid;
353 	chan->inst = args.chan.inst;
354 	chan->token = args.chan.token;
355 	return 0;
356 }
357 
358 static int
359 nouveau_channel_init(struct nouveau_channel *chan, u32 vram, u32 gart)
360 {
361 	struct nouveau_cli *cli = chan->cli;
362 	struct nouveau_drm *drm = cli->drm;
363 	struct nvif_device *device = &cli->device;
364 	struct nv_dma_v0 args = {};
365 	int ret, i;
366 
367 	ret = nvif_object_map(chan->userd, NULL, 0);
368 	if (ret)
369 		return ret;
370 
371 	if (chan->user.oclass >= FERMI_CHANNEL_GPFIFO) {
372 		struct {
373 			struct nvif_event_v0 base;
374 			struct nvif_chan_event_v0 host;
375 		} args;
376 
377 		args.host.version = 0;
378 		args.host.type = NVIF_CHAN_EVENT_V0_KILLED;
379 
380 		ret = nvif_event_ctor(&chan->user, "abi16ChanKilled", chan->chid,
381 				      nouveau_channel_killed, false,
382 				      &args.base, sizeof(args), &chan->kill);
383 		if (ret == 0)
384 			ret = nvif_event_allow(&chan->kill);
385 		if (ret) {
386 			NV_ERROR(drm, "Failed to request channel kill "
387 				      "notification: %d\n", ret);
388 			return ret;
389 		}
390 	}
391 
392 	/* allocate dma objects to cover all allowed vram, and gart */
393 	if (device->info.family < NV_DEVICE_INFO_V0_FERMI) {
394 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
395 			args.target = NV_DMA_V0_TARGET_VM;
396 			args.access = NV_DMA_V0_ACCESS_VM;
397 			args.start = 0;
398 			args.limit = chan->vmm->vmm.limit - 1;
399 		} else {
400 			args.target = NV_DMA_V0_TARGET_VRAM;
401 			args.access = NV_DMA_V0_ACCESS_RDWR;
402 			args.start = 0;
403 			args.limit = device->info.ram_user - 1;
404 		}
405 
406 		ret = nvif_object_ctor(&chan->user, "abi16ChanVramCtxDma", vram,
407 				       NV_DMA_IN_MEMORY, &args, sizeof(args),
408 				       &chan->vram);
409 		if (ret)
410 			return ret;
411 
412 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
413 			args.target = NV_DMA_V0_TARGET_VM;
414 			args.access = NV_DMA_V0_ACCESS_VM;
415 			args.start = 0;
416 			args.limit = chan->vmm->vmm.limit - 1;
417 		} else
418 		if (drm->agp.bridge) {
419 			args.target = NV_DMA_V0_TARGET_AGP;
420 			args.access = NV_DMA_V0_ACCESS_RDWR;
421 			args.start = drm->agp.base;
422 			args.limit = drm->agp.base + drm->agp.size - 1;
423 		} else {
424 			args.target = NV_DMA_V0_TARGET_VM;
425 			args.access = NV_DMA_V0_ACCESS_RDWR;
426 			args.start = 0;
427 			args.limit = chan->vmm->vmm.limit - 1;
428 		}
429 
430 		ret = nvif_object_ctor(&chan->user, "abi16ChanGartCtxDma", gart,
431 				       NV_DMA_IN_MEMORY, &args, sizeof(args),
432 				       &chan->gart);
433 		if (ret)
434 			return ret;
435 	}
436 
437 	/* initialise dma tracking parameters */
438 	switch (chan->user.oclass) {
439 	case NV03_CHANNEL_DMA:
440 	case NV10_CHANNEL_DMA:
441 	case NV17_CHANNEL_DMA:
442 	case NV40_CHANNEL_DMA:
443 		chan->user_put = 0x40;
444 		chan->user_get = 0x44;
445 		chan->dma.max = (0x10000 / 4) - 2;
446 		break;
447 	default:
448 		chan->user_put = 0x40;
449 		chan->user_get = 0x44;
450 		chan->user_get_hi = 0x60;
451 		chan->dma.ib_base =  0x10000 / 4;
452 		chan->dma.ib_max  = NV50_DMA_IB_MAX;
453 		chan->dma.ib_put  = 0;
454 		chan->dma.ib_free = chan->dma.ib_max - chan->dma.ib_put;
455 		chan->dma.max = chan->dma.ib_base;
456 		break;
457 	}
458 
459 	chan->dma.put = 0;
460 	chan->dma.cur = chan->dma.put;
461 	chan->dma.free = chan->dma.max - chan->dma.cur;
462 
463 	ret = PUSH_WAIT(&chan->chan.push, NOUVEAU_DMA_SKIPS);
464 	if (ret)
465 		return ret;
466 
467 	for (i = 0; i < NOUVEAU_DMA_SKIPS; i++)
468 		PUSH_DATA(&chan->chan.push, 0x00000000);
469 
470 	/* allocate software object class (used for fences on <= nv05) */
471 	if (device->info.family < NV_DEVICE_INFO_V0_CELSIUS) {
472 		ret = nvif_object_ctor(&chan->user, "abi16NvswFence", 0x006e,
473 				       NVIF_CLASS_SW_NV04,
474 				       NULL, 0, &chan->nvsw);
475 		if (ret)
476 			return ret;
477 
478 		ret = PUSH_WAIT(&chan->chan.push, 2);
479 		if (ret)
480 			return ret;
481 
482 		PUSH_NVSQ(&chan->chan.push, NV_SW, 0x0000, chan->nvsw.handle);
483 		PUSH_KICK(&chan->chan.push);
484 	}
485 
486 	/* initialise synchronisation */
487 	return nouveau_fence(drm)->context_new(chan);
488 }
489 
490 int
491 nouveau_channel_new(struct nouveau_cli *cli,
492 		    bool priv, u64 runm, u32 vram, u32 gart, struct nouveau_channel **pchan)
493 {
494 	int ret;
495 
496 	ret = nouveau_channel_ctor(cli, priv, runm, pchan);
497 	if (ret) {
498 		NV_PRINTK(dbg, cli, "channel create, %d\n", ret);
499 		return ret;
500 	}
501 
502 	ret = nouveau_channel_init(*pchan, vram, gart);
503 	if (ret) {
504 		NV_PRINTK(err, cli, "channel failed to initialise, %d\n", ret);
505 		nouveau_channel_del(pchan);
506 		return ret;
507 	}
508 
509 	ret = nouveau_svmm_join((*pchan)->vmm->svmm, (*pchan)->inst);
510 	if (ret)
511 		nouveau_channel_del(pchan);
512 
513 	return ret;
514 }
515 
516 void
517 nouveau_channels_fini(struct nouveau_drm *drm)
518 {
519 	kfree(drm->runl);
520 }
521 
522 int
523 nouveau_channels_init(struct nouveau_drm *drm)
524 {
525 	struct {
526 		struct nv_device_info_v1 m;
527 		struct {
528 			struct nv_device_info_v1_data channels;
529 			struct nv_device_info_v1_data runlists;
530 		} v;
531 	} args = {
532 		.m.version = 1,
533 		.m.count = sizeof(args.v) / sizeof(args.v.channels),
534 		.v.channels.mthd = NV_DEVICE_HOST_CHANNELS,
535 		.v.runlists.mthd = NV_DEVICE_HOST_RUNLISTS,
536 	};
537 	struct nvif_object *device = &drm->client.device.object;
538 	int ret, i;
539 
540 	ret = nvif_object_mthd(device, NV_DEVICE_V0_INFO, &args, sizeof(args));
541 	if (ret ||
542 	    args.v.runlists.mthd == NV_DEVICE_INFO_INVALID || !args.v.runlists.data ||
543 	    args.v.channels.mthd == NV_DEVICE_INFO_INVALID)
544 		return -ENODEV;
545 
546 	drm->chan_nr = drm->chan_total = args.v.channels.data;
547 	drm->runl_nr = fls64(args.v.runlists.data);
548 	drm->runl = kcalloc(drm->runl_nr, sizeof(*drm->runl), GFP_KERNEL);
549 	if (!drm->runl)
550 		return -ENOMEM;
551 
552 	if (drm->chan_nr == 0) {
553 		for (i = 0; i < drm->runl_nr; i++) {
554 			if (!(args.v.runlists.data & BIT(i)))
555 				continue;
556 
557 			args.v.channels.mthd = NV_DEVICE_HOST_RUNLIST_CHANNELS;
558 			args.v.channels.data = i;
559 
560 			ret = nvif_object_mthd(device, NV_DEVICE_V0_INFO, &args, sizeof(args));
561 			if (ret || args.v.channels.mthd == NV_DEVICE_INFO_INVALID)
562 				return -ENODEV;
563 
564 			drm->runl[i].chan_nr = args.v.channels.data;
565 			drm->runl[i].chan_id_base = drm->chan_total;
566 			drm->runl[i].context_base = dma_fence_context_alloc(drm->runl[i].chan_nr);
567 
568 			drm->chan_total += drm->runl[i].chan_nr;
569 		}
570 	} else {
571 		drm->runl[0].context_base = dma_fence_context_alloc(drm->chan_nr);
572 		for (i = 1; i < drm->runl_nr; i++)
573 			drm->runl[i].context_base = drm->runl[0].context_base;
574 
575 	}
576 
577 	return 0;
578 }
579