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/cl006b.h>
29 #include <nvif/cl506f.h>
30 #include <nvif/cl906f.h>
31 #include <nvif/cla06f.h>
32 #include <nvif/clc36f.h>
33 #include <nvif/ioctl.h>
34 
35 #include "nouveau_drv.h"
36 #include "nouveau_dma.h"
37 #include "nouveau_bo.h"
38 #include "nouveau_chan.h"
39 #include "nouveau_fence.h"
40 #include "nouveau_abi16.h"
41 #include "nouveau_vmm.h"
42 #include "nouveau_svm.h"
43 
44 MODULE_PARM_DESC(vram_pushbuf, "Create DMA push buffers in VRAM");
45 int nouveau_vram_pushbuf;
46 module_param_named(vram_pushbuf, nouveau_vram_pushbuf, int, 0400);
47 
48 static int
49 nouveau_channel_killed(struct nvif_notify *ntfy)
50 {
51 	struct nouveau_channel *chan = container_of(ntfy, typeof(*chan), kill);
52 	struct nouveau_cli *cli = (void *)chan->user.client;
53 	NV_PRINTK(warn, cli, "channel %d killed!\n", chan->chid);
54 	atomic_set(&chan->killed, 1);
55 	if (chan->fence)
56 		nouveau_fence_context_kill(chan->fence, -ENODEV);
57 	return NVIF_NOTIFY_DROP;
58 }
59 
60 int
61 nouveau_channel_idle(struct nouveau_channel *chan)
62 {
63 	if (likely(chan && chan->fence && !atomic_read(&chan->killed))) {
64 		struct nouveau_cli *cli = (void *)chan->user.client;
65 		struct nouveau_fence *fence = NULL;
66 		int ret;
67 
68 		ret = nouveau_fence_new(chan, false, &fence);
69 		if (!ret) {
70 			ret = nouveau_fence_wait(fence, false, false);
71 			nouveau_fence_unref(&fence);
72 		}
73 
74 		if (ret) {
75 			NV_PRINTK(err, cli, "failed to idle channel %d [%s]\n",
76 				  chan->chid, nvxx_client(&cli->base)->name);
77 			return ret;
78 		}
79 	}
80 	return 0;
81 }
82 
83 void
84 nouveau_channel_del(struct nouveau_channel **pchan)
85 {
86 	struct nouveau_channel *chan = *pchan;
87 	if (chan) {
88 		struct nouveau_cli *cli = (void *)chan->user.client;
89 		bool super;
90 
91 		if (cli) {
92 			super = cli->base.super;
93 			cli->base.super = true;
94 		}
95 
96 		if (chan->fence)
97 			nouveau_fence(chan->drm)->context_del(chan);
98 
99 		if (cli)
100 			nouveau_svmm_part(chan->vmm->svmm, chan->inst);
101 
102 		nvif_object_dtor(&chan->nvsw);
103 		nvif_object_dtor(&chan->gart);
104 		nvif_object_dtor(&chan->vram);
105 		nvif_notify_dtor(&chan->kill);
106 		nvif_object_dtor(&chan->user);
107 		nvif_object_dtor(&chan->push.ctxdma);
108 		nouveau_vma_del(&chan->push.vma);
109 		nouveau_bo_unmap(chan->push.buffer);
110 		if (chan->push.buffer && chan->push.buffer->pin_refcnt)
111 			nouveau_bo_unpin(chan->push.buffer);
112 		nouveau_bo_ref(NULL, &chan->push.buffer);
113 		kfree(chan);
114 
115 		if (cli)
116 			cli->base.super = super;
117 	}
118 	*pchan = NULL;
119 }
120 
121 static void
122 nouveau_channel_kick(struct nvif_push *push)
123 {
124 	struct nouveau_channel *chan = container_of(push, typeof(*chan), chan._push);
125 	chan->dma.cur = chan->dma.cur + (chan->chan._push.cur - chan->chan._push.bgn);
126 	FIRE_RING(chan);
127 	chan->chan._push.bgn = chan->chan._push.cur;
128 }
129 
130 static int
131 nouveau_channel_wait(struct nvif_push *push, u32 size)
132 {
133 	struct nouveau_channel *chan = container_of(push, typeof(*chan), chan._push);
134 	int ret;
135 	chan->dma.cur = chan->dma.cur + (chan->chan._push.cur - chan->chan._push.bgn);
136 	ret = RING_SPACE(chan, size);
137 	if (ret == 0) {
138 		chan->chan._push.bgn = chan->chan._push.mem.object.map.ptr;
139 		chan->chan._push.bgn = chan->chan._push.bgn + chan->dma.cur;
140 		chan->chan._push.cur = chan->chan._push.bgn;
141 		chan->chan._push.end = chan->chan._push.bgn + size;
142 	}
143 	return ret;
144 }
145 
146 static int
147 nouveau_channel_prep(struct nouveau_drm *drm, struct nvif_device *device,
148 		     u32 size, struct nouveau_channel **pchan)
149 {
150 	struct nouveau_cli *cli = (void *)device->object.client;
151 	struct nv_dma_v0 args = {};
152 	struct nouveau_channel *chan;
153 	u32 target;
154 	int ret;
155 
156 	chan = *pchan = kzalloc(sizeof(*chan), GFP_KERNEL);
157 	if (!chan)
158 		return -ENOMEM;
159 
160 	chan->device = device;
161 	chan->drm = drm;
162 	chan->vmm = cli->svm.cli ? &cli->svm : &cli->vmm;
163 	atomic_set(&chan->killed, 0);
164 
165 	/* allocate memory for dma push buffer */
166 	target = TTM_PL_FLAG_TT | TTM_PL_FLAG_UNCACHED;
167 	if (nouveau_vram_pushbuf)
168 		target = TTM_PL_FLAG_VRAM;
169 
170 	ret = nouveau_bo_new(cli, size, 0, target, 0, 0, NULL, NULL,
171 			    &chan->push.buffer);
172 	if (ret == 0) {
173 		ret = nouveau_bo_pin(chan->push.buffer, target, false);
174 		if (ret == 0)
175 			ret = nouveau_bo_map(chan->push.buffer);
176 	}
177 
178 	if (ret) {
179 		nouveau_channel_del(pchan);
180 		return ret;
181 	}
182 
183 	chan->chan._push.mem.object.parent = cli->base.object.parent;
184 	chan->chan._push.mem.object.client = &cli->base;
185 	chan->chan._push.mem.object.name = "chanPush";
186 	chan->chan._push.mem.object.map.ptr = chan->push.buffer->kmap.virtual;
187 	chan->chan._push.wait = nouveau_channel_wait;
188 	chan->chan._push.kick = nouveau_channel_kick;
189 	chan->chan.push = &chan->chan._push;
190 
191 	/* create dma object covering the *entire* memory space that the
192 	 * pushbuf lives in, this is because the GEM code requires that
193 	 * we be able to call out to other (indirect) push buffers
194 	 */
195 	chan->push.addr = chan->push.buffer->offset;
196 
197 	if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
198 		ret = nouveau_vma_new(chan->push.buffer, chan->vmm,
199 				      &chan->push.vma);
200 		if (ret) {
201 			nouveau_channel_del(pchan);
202 			return ret;
203 		}
204 
205 		chan->push.addr = chan->push.vma->addr;
206 
207 		if (device->info.family >= NV_DEVICE_INFO_V0_FERMI)
208 			return 0;
209 
210 		args.target = NV_DMA_V0_TARGET_VM;
211 		args.access = NV_DMA_V0_ACCESS_VM;
212 		args.start = 0;
213 		args.limit = chan->vmm->vmm.limit - 1;
214 	} else
215 	if (chan->push.buffer->bo.mem.mem_type == TTM_PL_VRAM) {
216 		if (device->info.family == NV_DEVICE_INFO_V0_TNT) {
217 			/* nv04 vram pushbuf hack, retarget to its location in
218 			 * the framebuffer bar rather than direct vram access..
219 			 * nfi why this exists, it came from the -nv ddx.
220 			 */
221 			args.target = NV_DMA_V0_TARGET_PCI;
222 			args.access = NV_DMA_V0_ACCESS_RDWR;
223 			args.start = nvxx_device(device)->func->
224 				resource_addr(nvxx_device(device), 1);
225 			args.limit = args.start + device->info.ram_user - 1;
226 		} else {
227 			args.target = NV_DMA_V0_TARGET_VRAM;
228 			args.access = NV_DMA_V0_ACCESS_RDWR;
229 			args.start = 0;
230 			args.limit = device->info.ram_user - 1;
231 		}
232 	} else {
233 		if (chan->drm->agp.bridge) {
234 			args.target = NV_DMA_V0_TARGET_AGP;
235 			args.access = NV_DMA_V0_ACCESS_RDWR;
236 			args.start = chan->drm->agp.base;
237 			args.limit = chan->drm->agp.base +
238 				     chan->drm->agp.size - 1;
239 		} else {
240 			args.target = NV_DMA_V0_TARGET_VM;
241 			args.access = NV_DMA_V0_ACCESS_RDWR;
242 			args.start = 0;
243 			args.limit = chan->vmm->vmm.limit - 1;
244 		}
245 	}
246 
247 	ret = nvif_object_ctor(&device->object, "abi16PushCtxDma", 0,
248 			       NV_DMA_FROM_MEMORY, &args, sizeof(args),
249 			       &chan->push.ctxdma);
250 	if (ret) {
251 		nouveau_channel_del(pchan);
252 		return ret;
253 	}
254 
255 	return 0;
256 }
257 
258 static int
259 nouveau_channel_ind(struct nouveau_drm *drm, struct nvif_device *device,
260 		    u64 runlist, bool priv, struct nouveau_channel **pchan)
261 {
262 	static const u16 oclasses[] = { TURING_CHANNEL_GPFIFO_A,
263 					VOLTA_CHANNEL_GPFIFO_A,
264 					PASCAL_CHANNEL_GPFIFO_A,
265 					MAXWELL_CHANNEL_GPFIFO_A,
266 					KEPLER_CHANNEL_GPFIFO_B,
267 					KEPLER_CHANNEL_GPFIFO_A,
268 					FERMI_CHANNEL_GPFIFO,
269 					G82_CHANNEL_GPFIFO,
270 					NV50_CHANNEL_GPFIFO,
271 					0 };
272 	const u16 *oclass = oclasses;
273 	union {
274 		struct nv50_channel_gpfifo_v0 nv50;
275 		struct fermi_channel_gpfifo_v0 fermi;
276 		struct kepler_channel_gpfifo_a_v0 kepler;
277 		struct volta_channel_gpfifo_a_v0 volta;
278 	} args;
279 	struct nouveau_channel *chan;
280 	u32 size;
281 	int ret;
282 
283 	/* allocate dma push buffer */
284 	ret = nouveau_channel_prep(drm, device, 0x12000, &chan);
285 	*pchan = chan;
286 	if (ret)
287 		return ret;
288 
289 	/* create channel object */
290 	do {
291 		if (oclass[0] >= VOLTA_CHANNEL_GPFIFO_A) {
292 			args.volta.version = 0;
293 			args.volta.ilength = 0x02000;
294 			args.volta.ioffset = 0x10000 + chan->push.addr;
295 			args.volta.runlist = runlist;
296 			args.volta.vmm = nvif_handle(&chan->vmm->vmm.object);
297 			args.volta.priv = priv;
298 			size = sizeof(args.volta);
299 		} else
300 		if (oclass[0] >= KEPLER_CHANNEL_GPFIFO_A) {
301 			args.kepler.version = 0;
302 			args.kepler.ilength = 0x02000;
303 			args.kepler.ioffset = 0x10000 + chan->push.addr;
304 			args.kepler.runlist = runlist;
305 			args.kepler.vmm = nvif_handle(&chan->vmm->vmm.object);
306 			args.kepler.priv = priv;
307 			size = sizeof(args.kepler);
308 		} else
309 		if (oclass[0] >= FERMI_CHANNEL_GPFIFO) {
310 			args.fermi.version = 0;
311 			args.fermi.ilength = 0x02000;
312 			args.fermi.ioffset = 0x10000 + chan->push.addr;
313 			args.fermi.vmm = nvif_handle(&chan->vmm->vmm.object);
314 			size = sizeof(args.fermi);
315 		} else {
316 			args.nv50.version = 0;
317 			args.nv50.ilength = 0x02000;
318 			args.nv50.ioffset = 0x10000 + chan->push.addr;
319 			args.nv50.pushbuf = nvif_handle(&chan->push.ctxdma);
320 			args.nv50.vmm = nvif_handle(&chan->vmm->vmm.object);
321 			size = sizeof(args.nv50);
322 		}
323 
324 		ret = nvif_object_ctor(&device->object, "abi16ChanUser", 0,
325 				       *oclass++, &args, size, &chan->user);
326 		if (ret == 0) {
327 			if (chan->user.oclass >= VOLTA_CHANNEL_GPFIFO_A) {
328 				chan->chid = args.volta.chid;
329 				chan->inst = args.volta.inst;
330 				chan->token = args.volta.token;
331 			} else
332 			if (chan->user.oclass >= KEPLER_CHANNEL_GPFIFO_A) {
333 				chan->chid = args.kepler.chid;
334 				chan->inst = args.kepler.inst;
335 			} else
336 			if (chan->user.oclass >= FERMI_CHANNEL_GPFIFO) {
337 				chan->chid = args.fermi.chid;
338 			} else {
339 				chan->chid = args.nv50.chid;
340 			}
341 			return ret;
342 		}
343 	} while (*oclass);
344 
345 	nouveau_channel_del(pchan);
346 	return ret;
347 }
348 
349 static int
350 nouveau_channel_dma(struct nouveau_drm *drm, struct nvif_device *device,
351 		    struct nouveau_channel **pchan)
352 {
353 	static const u16 oclasses[] = { NV40_CHANNEL_DMA,
354 					NV17_CHANNEL_DMA,
355 					NV10_CHANNEL_DMA,
356 					NV03_CHANNEL_DMA,
357 					0 };
358 	const u16 *oclass = oclasses;
359 	struct nv03_channel_dma_v0 args;
360 	struct nouveau_channel *chan;
361 	int ret;
362 
363 	/* allocate dma push buffer */
364 	ret = nouveau_channel_prep(drm, device, 0x10000, &chan);
365 	*pchan = chan;
366 	if (ret)
367 		return ret;
368 
369 	/* create channel object */
370 	args.version = 0;
371 	args.pushbuf = nvif_handle(&chan->push.ctxdma);
372 	args.offset = chan->push.addr;
373 
374 	do {
375 		ret = nvif_object_ctor(&device->object, "abi16ChanUser", 0,
376 				       *oclass++, &args, sizeof(args),
377 				       &chan->user);
378 		if (ret == 0) {
379 			chan->chid = args.chid;
380 			return ret;
381 		}
382 	} while (ret && *oclass);
383 
384 	nouveau_channel_del(pchan);
385 	return ret;
386 }
387 
388 static int
389 nouveau_channel_init(struct nouveau_channel *chan, u32 vram, u32 gart)
390 {
391 	struct nvif_device *device = chan->device;
392 	struct nouveau_drm *drm = chan->drm;
393 	struct nv_dma_v0 args = {};
394 	int ret, i;
395 
396 	nvif_object_map(&chan->user, NULL, 0);
397 
398 	if (chan->user.oclass >= FERMI_CHANNEL_GPFIFO) {
399 		ret = nvif_notify_ctor(&chan->user, "abi16ChanKilled",
400 				       nouveau_channel_killed,
401 				       true, NV906F_V0_NTFY_KILLED,
402 				       NULL, 0, 0, &chan->kill);
403 		if (ret == 0)
404 			ret = nvif_notify_get(&chan->kill);
405 		if (ret) {
406 			NV_ERROR(drm, "Failed to request channel kill "
407 				      "notification: %d\n", ret);
408 			return ret;
409 		}
410 	}
411 
412 	/* allocate dma objects to cover all allowed vram, and gart */
413 	if (device->info.family < NV_DEVICE_INFO_V0_FERMI) {
414 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
415 			args.target = NV_DMA_V0_TARGET_VM;
416 			args.access = NV_DMA_V0_ACCESS_VM;
417 			args.start = 0;
418 			args.limit = chan->vmm->vmm.limit - 1;
419 		} else {
420 			args.target = NV_DMA_V0_TARGET_VRAM;
421 			args.access = NV_DMA_V0_ACCESS_RDWR;
422 			args.start = 0;
423 			args.limit = device->info.ram_user - 1;
424 		}
425 
426 		ret = nvif_object_ctor(&chan->user, "abi16ChanVramCtxDma", vram,
427 				       NV_DMA_IN_MEMORY, &args, sizeof(args),
428 				       &chan->vram);
429 		if (ret)
430 			return ret;
431 
432 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
433 			args.target = NV_DMA_V0_TARGET_VM;
434 			args.access = NV_DMA_V0_ACCESS_VM;
435 			args.start = 0;
436 			args.limit = chan->vmm->vmm.limit - 1;
437 		} else
438 		if (chan->drm->agp.bridge) {
439 			args.target = NV_DMA_V0_TARGET_AGP;
440 			args.access = NV_DMA_V0_ACCESS_RDWR;
441 			args.start = chan->drm->agp.base;
442 			args.limit = chan->drm->agp.base +
443 				     chan->drm->agp.size - 1;
444 		} else {
445 			args.target = NV_DMA_V0_TARGET_VM;
446 			args.access = NV_DMA_V0_ACCESS_RDWR;
447 			args.start = 0;
448 			args.limit = chan->vmm->vmm.limit - 1;
449 		}
450 
451 		ret = nvif_object_ctor(&chan->user, "abi16ChanGartCtxDma", gart,
452 				       NV_DMA_IN_MEMORY, &args, sizeof(args),
453 				       &chan->gart);
454 		if (ret)
455 			return ret;
456 	}
457 
458 	/* initialise dma tracking parameters */
459 	switch (chan->user.oclass & 0x00ff) {
460 	case 0x006b:
461 	case 0x006e:
462 		chan->user_put = 0x40;
463 		chan->user_get = 0x44;
464 		chan->dma.max = (0x10000 / 4) - 2;
465 		break;
466 	default:
467 		chan->user_put = 0x40;
468 		chan->user_get = 0x44;
469 		chan->user_get_hi = 0x60;
470 		chan->dma.ib_base =  0x10000 / 4;
471 		chan->dma.ib_max  = (0x02000 / 8) - 1;
472 		chan->dma.ib_put  = 0;
473 		chan->dma.ib_free = chan->dma.ib_max - chan->dma.ib_put;
474 		chan->dma.max = chan->dma.ib_base;
475 		break;
476 	}
477 
478 	chan->dma.put = 0;
479 	chan->dma.cur = chan->dma.put;
480 	chan->dma.free = chan->dma.max - chan->dma.cur;
481 
482 	ret = PUSH_WAIT(chan->chan.push, NOUVEAU_DMA_SKIPS);
483 	if (ret)
484 		return ret;
485 
486 	for (i = 0; i < NOUVEAU_DMA_SKIPS; i++)
487 		PUSH_DATA(chan->chan.push, 0x00000000);
488 
489 	/* allocate software object class (used for fences on <= nv05) */
490 	if (device->info.family < NV_DEVICE_INFO_V0_CELSIUS) {
491 		ret = nvif_object_ctor(&chan->user, "abi16NvswFence", 0x006e,
492 				       NVIF_CLASS_SW_NV04,
493 				       NULL, 0, &chan->nvsw);
494 		if (ret)
495 			return ret;
496 
497 		ret = PUSH_WAIT(chan->chan.push, 2);
498 		if (ret)
499 			return ret;
500 
501 		PUSH_NVSQ(chan->chan.push, NV_SW, 0x0000, chan->nvsw.handle);
502 		PUSH_KICK(chan->chan.push);
503 	}
504 
505 	/* initialise synchronisation */
506 	return nouveau_fence(chan->drm)->context_new(chan);
507 }
508 
509 int
510 nouveau_channel_new(struct nouveau_drm *drm, struct nvif_device *device,
511 		    u32 arg0, u32 arg1, bool priv,
512 		    struct nouveau_channel **pchan)
513 {
514 	struct nouveau_cli *cli = (void *)device->object.client;
515 	bool super;
516 	int ret;
517 
518 	/* hack until fencenv50 is fixed, and agp access relaxed */
519 	super = cli->base.super;
520 	cli->base.super = true;
521 
522 	ret = nouveau_channel_ind(drm, device, arg0, priv, pchan);
523 	if (ret) {
524 		NV_PRINTK(dbg, cli, "ib channel create, %d\n", ret);
525 		ret = nouveau_channel_dma(drm, device, pchan);
526 		if (ret) {
527 			NV_PRINTK(dbg, cli, "dma channel create, %d\n", ret);
528 			goto done;
529 		}
530 	}
531 
532 	ret = nouveau_channel_init(*pchan, arg0, arg1);
533 	if (ret) {
534 		NV_PRINTK(err, cli, "channel failed to initialise, %d\n", ret);
535 		nouveau_channel_del(pchan);
536 	}
537 
538 	ret = nouveau_svmm_join((*pchan)->vmm->svmm, (*pchan)->inst);
539 	if (ret)
540 		nouveau_channel_del(pchan);
541 
542 done:
543 	cli->base.super = super;
544 	return ret;
545 }
546 
547 int
548 nouveau_channels_init(struct nouveau_drm *drm)
549 {
550 	struct {
551 		struct nv_device_info_v1 m;
552 		struct {
553 			struct nv_device_info_v1_data channels;
554 		} v;
555 	} args = {
556 		.m.version = 1,
557 		.m.count = sizeof(args.v) / sizeof(args.v.channels),
558 		.v.channels.mthd = NV_DEVICE_FIFO_CHANNELS,
559 	};
560 	struct nvif_object *device = &drm->client.device.object;
561 	int ret;
562 
563 	ret = nvif_object_mthd(device, NV_DEVICE_V0_INFO, &args, sizeof(args));
564 	if (ret || args.v.channels.mthd == NV_DEVICE_INFO_INVALID)
565 		return -ENODEV;
566 
567 	drm->chan.nr = args.v.channels.data;
568 	drm->chan.context_base = dma_fence_context_alloc(drm->chan.nr);
569 	return 0;
570 }
571