1 /*	$NetBSD: nouveau_chan.c,v 1.7 2021/12/18 23:45:32 riastradh Exp $	*/
2 
3 /*
4  * Copyright 2012 Red Hat Inc.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the "Software"),
8  * to deal in the Software without restriction, including without limitation
9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10  * and/or sell copies of the Software, and to permit persons to whom the
11  * Software is furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22  * OTHER DEALINGS IN THE SOFTWARE.
23  *
24  * Authors: Ben Skeggs
25  */
26 
27 #include <sys/cdefs.h>
28 __KERNEL_RCSID(0, "$NetBSD: nouveau_chan.c,v 1.7 2021/12/18 23:45:32 riastradh Exp $");
29 
30 #include <nvif/os.h>
31 #include <nvif/class.h>
32 #include <nvif/cl0002.h>
33 #include <nvif/cl006b.h>
34 #include <nvif/cl506f.h>
35 #include <nvif/cl906f.h>
36 #include <nvif/cla06f.h>
37 #include <nvif/clc36f.h>
38 #include <nvif/ioctl.h>
39 
40 /*XXX*/
41 #include <core/client.h>
42 
43 #include "nouveau_drv.h"
44 #include "nouveau_dma.h"
45 #include "nouveau_bo.h"
46 #include "nouveau_chan.h"
47 #include "nouveau_fence.h"
48 #include "nouveau_abi16.h"
49 #include "nouveau_vmm.h"
50 #include "nouveau_svm.h"
51 
52 MODULE_PARM_DESC(vram_pushbuf, "Create DMA push buffers in VRAM");
53 int nouveau_vram_pushbuf;
54 module_param_named(vram_pushbuf, nouveau_vram_pushbuf, int, 0400);
55 
56 static int
nouveau_channel_killed(struct nvif_notify * ntfy)57 nouveau_channel_killed(struct nvif_notify *ntfy)
58 {
59 	struct nouveau_channel *chan = container_of(ntfy, typeof(*chan), kill);
60 	struct nouveau_cli *cli = (void *)chan->user.client;
61 	NV_PRINTK(warn, cli, "channel %d killed!\n", chan->chid);
62 	atomic_set(&chan->killed, 1);
63 	if (chan->fence)
64 		nouveau_fence_context_kill(chan->fence, -ENODEV);
65 	return NVIF_NOTIFY_DROP;
66 }
67 
68 int
nouveau_channel_idle(struct nouveau_channel * chan)69 nouveau_channel_idle(struct nouveau_channel *chan)
70 {
71 	if (likely(chan && chan->fence && !atomic_read(&chan->killed))) {
72 		struct nouveau_cli *cli = (void *)chan->user.client;
73 		struct nouveau_fence *fence = NULL;
74 		int ret;
75 
76 		ret = nouveau_fence_new(chan, false, &fence);
77 		if (!ret) {
78 			ret = nouveau_fence_wait(fence, false, false);
79 			nouveau_fence_unref(&fence);
80 		}
81 
82 		if (ret) {
83 			NV_PRINTK(err, cli, "failed to idle channel %d [%s]\n",
84 				  chan->chid, nvxx_client(&cli->base)->name);
85 			return ret;
86 		}
87 	}
88 	return 0;
89 }
90 
91 void
nouveau_channel_del(struct nouveau_channel ** pchan)92 nouveau_channel_del(struct nouveau_channel **pchan)
93 {
94 	struct nouveau_channel *chan = *pchan;
95 	if (chan) {
96 		struct nouveau_cli *cli = (void *)chan->user.client;
97 		bool super;
98 
99 		if (cli) {
100 			super = cli->base.super;
101 			cli->base.super = true;
102 		}
103 
104 		if (chan->fence)
105 			nouveau_fence(chan->drm)->context_del(chan);
106 
107 		if (cli)
108 			nouveau_svmm_part(chan->vmm->svmm, chan->inst);
109 
110 		nvif_object_fini(&chan->nvsw);
111 		nvif_object_fini(&chan->gart);
112 		nvif_object_fini(&chan->vram);
113 		nvif_notify_fini(&chan->kill);
114 		nvif_object_fini(&chan->user);
115 		nvif_object_fini(&chan->push.ctxdma);
116 		nouveau_vma_del(&chan->push.vma);
117 		nouveau_bo_unmap(chan->push.buffer);
118 		if (chan->push.buffer && chan->push.buffer->pin_refcnt)
119 			nouveau_bo_unpin(chan->push.buffer);
120 		nouveau_bo_ref(NULL, &chan->push.buffer);
121 		kfree(chan);
122 
123 		if (cli)
124 			cli->base.super = super;
125 	}
126 	*pchan = NULL;
127 }
128 
129 static int
nouveau_channel_prep(struct nouveau_drm * drm,struct nvif_device * device,u32 size,struct nouveau_channel ** pchan)130 nouveau_channel_prep(struct nouveau_drm *drm, struct nvif_device *device,
131 		     u32 size, struct nouveau_channel **pchan)
132 {
133 	struct nouveau_cli *cli = (void *)device->object.client;
134 	struct nv_dma_v0 args = {};
135 	struct nouveau_channel *chan;
136 	u32 target;
137 	int ret;
138 
139 	chan = *pchan = kzalloc(sizeof(*chan), GFP_KERNEL);
140 	if (!chan)
141 		return -ENOMEM;
142 
143 	chan->device = device;
144 	chan->drm = drm;
145 	chan->vmm = cli->svm.cli ? &cli->svm : &cli->vmm;
146 	atomic_set(&chan->killed, 0);
147 
148 	/* allocate memory for dma push buffer */
149 	target = TTM_PL_FLAG_TT | TTM_PL_FLAG_UNCACHED;
150 	if (nouveau_vram_pushbuf)
151 		target = TTM_PL_FLAG_VRAM;
152 
153 	ret = nouveau_bo_new(cli, size, 0, target, 0, 0, NULL, NULL,
154 			    &chan->push.buffer);
155 	if (ret == 0) {
156 		ret = nouveau_bo_pin(chan->push.buffer, target, false);
157 		if (ret == 0)
158 			ret = nouveau_bo_map(chan->push.buffer);
159 	}
160 
161 	if (ret) {
162 		nouveau_channel_del(pchan);
163 		return ret;
164 	}
165 
166 	/* create dma object covering the *entire* memory space that the
167 	 * pushbuf lives in, this is because the GEM code requires that
168 	 * we be able to call out to other (indirect) push buffers
169 	 */
170 	chan->push.addr = chan->push.buffer->bo.offset;
171 
172 	if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
173 		ret = nouveau_vma_new(chan->push.buffer, chan->vmm,
174 				      &chan->push.vma);
175 		if (ret) {
176 			nouveau_channel_del(pchan);
177 			return ret;
178 		}
179 
180 		chan->push.addr = chan->push.vma->addr;
181 
182 		if (device->info.family >= NV_DEVICE_INFO_V0_FERMI)
183 			return 0;
184 
185 		args.target = NV_DMA_V0_TARGET_VM;
186 		args.access = NV_DMA_V0_ACCESS_VM;
187 		args.start = 0;
188 		args.limit = chan->vmm->vmm.limit - 1;
189 	} else
190 	if (chan->push.buffer->bo.mem.mem_type == TTM_PL_VRAM) {
191 		if (device->info.family == NV_DEVICE_INFO_V0_TNT) {
192 			/* nv04 vram pushbuf hack, retarget to its location in
193 			 * the framebuffer bar rather than direct vram access..
194 			 * nfi why this exists, it came from the -nv ddx.
195 			 */
196 			args.target = NV_DMA_V0_TARGET_PCI;
197 			args.access = NV_DMA_V0_ACCESS_RDWR;
198 			args.start = nvxx_device(device)->func->
199 				resource_addr(nvxx_device(device), 1);
200 			args.limit = args.start + device->info.ram_user - 1;
201 		} else {
202 			args.target = NV_DMA_V0_TARGET_VRAM;
203 			args.access = NV_DMA_V0_ACCESS_RDWR;
204 			args.start = 0;
205 			args.limit = device->info.ram_user - 1;
206 		}
207 	} else {
208 		if (chan->drm->agp.bridge) {
209 			args.target = NV_DMA_V0_TARGET_AGP;
210 			args.access = NV_DMA_V0_ACCESS_RDWR;
211 			args.start = chan->drm->agp.base;
212 			args.limit = chan->drm->agp.base +
213 				     chan->drm->agp.size - 1;
214 		} else {
215 			args.target = NV_DMA_V0_TARGET_VM;
216 			args.access = NV_DMA_V0_ACCESS_RDWR;
217 			args.start = 0;
218 			args.limit = chan->vmm->vmm.limit - 1;
219 		}
220 	}
221 
222 	ret = nvif_object_init(&device->object, 0, NV_DMA_FROM_MEMORY,
223 			       &args, sizeof(args), &chan->push.ctxdma);
224 	if (ret) {
225 		nouveau_channel_del(pchan);
226 		return ret;
227 	}
228 
229 	return 0;
230 }
231 
232 static int
nouveau_channel_ind(struct nouveau_drm * drm,struct nvif_device * device,u64 runlist,bool priv,struct nouveau_channel ** pchan)233 nouveau_channel_ind(struct nouveau_drm *drm, struct nvif_device *device,
234 		    u64 runlist, bool priv, struct nouveau_channel **pchan)
235 {
236 	static const u16 oclasses[] = { TURING_CHANNEL_GPFIFO_A,
237 					VOLTA_CHANNEL_GPFIFO_A,
238 					PASCAL_CHANNEL_GPFIFO_A,
239 					MAXWELL_CHANNEL_GPFIFO_A,
240 					KEPLER_CHANNEL_GPFIFO_B,
241 					KEPLER_CHANNEL_GPFIFO_A,
242 					FERMI_CHANNEL_GPFIFO,
243 					G82_CHANNEL_GPFIFO,
244 					NV50_CHANNEL_GPFIFO,
245 					0 };
246 	const u16 *oclass = oclasses;
247 	union {
248 		struct nv50_channel_gpfifo_v0 nv50;
249 		struct fermi_channel_gpfifo_v0 fermi;
250 		struct kepler_channel_gpfifo_a_v0 kepler;
251 		struct volta_channel_gpfifo_a_v0 volta;
252 	} args;
253 	struct nouveau_channel *chan;
254 	u32 size;
255 	int ret;
256 
257 	/* allocate dma push buffer */
258 	ret = nouveau_channel_prep(drm, device, 0x12000, &chan);
259 	*pchan = chan;
260 	if (ret)
261 		return ret;
262 
263 	/* create channel object */
264 	do {
265 		if (oclass[0] >= VOLTA_CHANNEL_GPFIFO_A) {
266 			args.volta.version = 0;
267 			args.volta.ilength = 0x02000;
268 			args.volta.ioffset = 0x10000 + chan->push.addr;
269 			args.volta.runlist = runlist;
270 			args.volta.vmm = nvif_handle(&chan->vmm->vmm.object);
271 			args.volta.priv = priv;
272 			size = sizeof(args.volta);
273 		} else
274 		if (oclass[0] >= KEPLER_CHANNEL_GPFIFO_A) {
275 			args.kepler.version = 0;
276 			args.kepler.ilength = 0x02000;
277 			args.kepler.ioffset = 0x10000 + chan->push.addr;
278 			args.kepler.runlist = runlist;
279 			args.kepler.vmm = nvif_handle(&chan->vmm->vmm.object);
280 			args.kepler.priv = priv;
281 			size = sizeof(args.kepler);
282 		} else
283 		if (oclass[0] >= FERMI_CHANNEL_GPFIFO) {
284 			args.fermi.version = 0;
285 			args.fermi.ilength = 0x02000;
286 			args.fermi.ioffset = 0x10000 + chan->push.addr;
287 			args.fermi.vmm = nvif_handle(&chan->vmm->vmm.object);
288 			size = sizeof(args.fermi);
289 		} else {
290 			args.nv50.version = 0;
291 			args.nv50.ilength = 0x02000;
292 			args.nv50.ioffset = 0x10000 + chan->push.addr;
293 			args.nv50.pushbuf = nvif_handle(&chan->push.ctxdma);
294 			args.nv50.vmm = nvif_handle(&chan->vmm->vmm.object);
295 			size = sizeof(args.nv50);
296 		}
297 
298 		ret = nvif_object_init(&device->object, 0, *oclass++,
299 				       &args, size, &chan->user);
300 		if (ret == 0) {
301 			if (chan->user.oclass >= VOLTA_CHANNEL_GPFIFO_A) {
302 				chan->chid = args.volta.chid;
303 				chan->inst = args.volta.inst;
304 				chan->token = args.volta.token;
305 			} else
306 			if (chan->user.oclass >= KEPLER_CHANNEL_GPFIFO_A) {
307 				chan->chid = args.kepler.chid;
308 				chan->inst = args.kepler.inst;
309 			} else
310 			if (chan->user.oclass >= FERMI_CHANNEL_GPFIFO) {
311 				chan->chid = args.fermi.chid;
312 			} else {
313 				chan->chid = args.nv50.chid;
314 			}
315 			return ret;
316 		}
317 	} while (*oclass);
318 
319 	nouveau_channel_del(pchan);
320 	return ret;
321 }
322 
323 static int
nouveau_channel_dma(struct nouveau_drm * drm,struct nvif_device * device,struct nouveau_channel ** pchan)324 nouveau_channel_dma(struct nouveau_drm *drm, struct nvif_device *device,
325 		    struct nouveau_channel **pchan)
326 {
327 	static const u16 oclasses[] = { NV40_CHANNEL_DMA,
328 					NV17_CHANNEL_DMA,
329 					NV10_CHANNEL_DMA,
330 					NV03_CHANNEL_DMA,
331 					0 };
332 	const u16 *oclass = oclasses;
333 	struct nv03_channel_dma_v0 args;
334 	struct nouveau_channel *chan;
335 	int ret;
336 
337 	/* allocate dma push buffer */
338 	ret = nouveau_channel_prep(drm, device, 0x10000, &chan);
339 	*pchan = chan;
340 	if (ret)
341 		return ret;
342 
343 	/* create channel object */
344 	args.version = 0;
345 	args.pushbuf = nvif_handle(&chan->push.ctxdma);
346 	args.offset = chan->push.addr;
347 
348 	do {
349 		ret = nvif_object_init(&device->object, 0, *oclass++,
350 				       &args, sizeof(args), &chan->user);
351 		if (ret == 0) {
352 			chan->chid = args.chid;
353 			return ret;
354 		}
355 	} while (ret && *oclass);
356 
357 	nouveau_channel_del(pchan);
358 	return ret;
359 }
360 
361 static int
nouveau_channel_init(struct nouveau_channel * chan,u32 vram,u32 gart)362 nouveau_channel_init(struct nouveau_channel *chan, u32 vram, u32 gart)
363 {
364 	struct nvif_device *device = chan->device;
365 	struct nouveau_drm *drm = chan->drm;
366 	struct nv_dma_v0 args = {};
367 	int ret, i;
368 
369 	nvif_object_map(&chan->user, NULL, 0);
370 
371 	if (chan->user.oclass >= FERMI_CHANNEL_GPFIFO) {
372 		ret = nvif_notify_init(&chan->user, nouveau_channel_killed,
373 				       true, NV906F_V0_NTFY_KILLED,
374 				       NULL, 0, 0, &chan->kill);
375 		if (ret == 0)
376 			ret = nvif_notify_get(&chan->kill);
377 		if (ret) {
378 			NV_ERROR(drm, "Failed to request channel kill "
379 				      "notification: %d\n", ret);
380 			return ret;
381 		}
382 	}
383 
384 	/* allocate dma objects to cover all allowed vram, and gart */
385 	if (device->info.family < NV_DEVICE_INFO_V0_FERMI) {
386 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
387 			args.target = NV_DMA_V0_TARGET_VM;
388 			args.access = NV_DMA_V0_ACCESS_VM;
389 			args.start = 0;
390 			args.limit = chan->vmm->vmm.limit - 1;
391 		} else {
392 			args.target = NV_DMA_V0_TARGET_VRAM;
393 			args.access = NV_DMA_V0_ACCESS_RDWR;
394 			args.start = 0;
395 			args.limit = device->info.ram_user - 1;
396 		}
397 
398 		ret = nvif_object_init(&chan->user, vram, NV_DMA_IN_MEMORY,
399 				       &args, sizeof(args), &chan->vram);
400 		if (ret)
401 			return ret;
402 
403 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
404 			args.target = NV_DMA_V0_TARGET_VM;
405 			args.access = NV_DMA_V0_ACCESS_VM;
406 			args.start = 0;
407 			args.limit = chan->vmm->vmm.limit - 1;
408 		} else
409 		if (chan->drm->agp.bridge) {
410 			args.target = NV_DMA_V0_TARGET_AGP;
411 			args.access = NV_DMA_V0_ACCESS_RDWR;
412 			args.start = chan->drm->agp.base;
413 			args.limit = chan->drm->agp.base +
414 				     chan->drm->agp.size - 1;
415 		} else {
416 			args.target = NV_DMA_V0_TARGET_VM;
417 			args.access = NV_DMA_V0_ACCESS_RDWR;
418 			args.start = 0;
419 			args.limit = chan->vmm->vmm.limit - 1;
420 		}
421 
422 		ret = nvif_object_init(&chan->user, gart, NV_DMA_IN_MEMORY,
423 				       &args, sizeof(args), &chan->gart);
424 		if (ret)
425 			return ret;
426 	}
427 
428 	/* initialise dma tracking parameters */
429 	switch (chan->user.oclass & 0x00ff) {
430 	case 0x006b:
431 	case 0x006e:
432 		chan->user_put = 0x40;
433 		chan->user_get = 0x44;
434 		chan->dma.max = (0x10000 / 4) - 2;
435 		break;
436 	default:
437 		chan->user_put = 0x40;
438 		chan->user_get = 0x44;
439 		chan->user_get_hi = 0x60;
440 		chan->dma.ib_base =  0x10000 / 4;
441 		chan->dma.ib_max  = (0x02000 / 8) - 1;
442 		chan->dma.ib_put  = 0;
443 		chan->dma.ib_free = chan->dma.ib_max - chan->dma.ib_put;
444 		chan->dma.max = chan->dma.ib_base;
445 		break;
446 	}
447 
448 	chan->dma.put = 0;
449 	chan->dma.cur = chan->dma.put;
450 	chan->dma.free = chan->dma.max - chan->dma.cur;
451 
452 	ret = RING_SPACE(chan, NOUVEAU_DMA_SKIPS);
453 	if (ret)
454 		return ret;
455 
456 	for (i = 0; i < NOUVEAU_DMA_SKIPS; i++)
457 		OUT_RING(chan, 0x00000000);
458 
459 	/* allocate software object class (used for fences on <= nv05) */
460 	if (device->info.family < NV_DEVICE_INFO_V0_CELSIUS) {
461 		ret = nvif_object_init(&chan->user, 0x006e,
462 				       NVIF_CLASS_SW_NV04,
463 				       NULL, 0, &chan->nvsw);
464 		if (ret)
465 			return ret;
466 
467 		ret = RING_SPACE(chan, 2);
468 		if (ret)
469 			return ret;
470 
471 		BEGIN_NV04(chan, NvSubSw, 0x0000, 1);
472 		OUT_RING  (chan, chan->nvsw.handle);
473 		FIRE_RING (chan);
474 	}
475 
476 	/* initialise synchronisation */
477 	return nouveau_fence(chan->drm)->context_new(chan);
478 }
479 
480 int
nouveau_channel_new(struct nouveau_drm * drm,struct nvif_device * device,u32 arg0,u32 arg1,bool priv,struct nouveau_channel ** pchan)481 nouveau_channel_new(struct nouveau_drm *drm, struct nvif_device *device,
482 		    u32 arg0, u32 arg1, bool priv,
483 		    struct nouveau_channel **pchan)
484 {
485 	struct nouveau_cli *cli = (void *)device->object.client;
486 	bool super;
487 	int ret;
488 
489 	/* hack until fencenv50 is fixed, and agp access relaxed */
490 	super = cli->base.super;
491 	cli->base.super = true;
492 
493 	ret = nouveau_channel_ind(drm, device, arg0, priv, pchan);
494 	if (ret) {
495 		NV_PRINTK(dbg, cli, "ib channel create, %d\n", ret);
496 		ret = nouveau_channel_dma(drm, device, pchan);
497 		if (ret) {
498 			NV_PRINTK(dbg, cli, "dma channel create, %d\n", ret);
499 			goto done;
500 		}
501 	}
502 
503 	ret = nouveau_channel_init(*pchan, arg0, arg1);
504 	if (ret) {
505 		NV_PRINTK(err, cli, "channel failed to initialise, %d\n", ret);
506 		nouveau_channel_del(pchan);
507 	}
508 
509 	ret = nouveau_svmm_join((*pchan)->vmm->svmm, (*pchan)->inst);
510 	if (ret)
511 		nouveau_channel_del(pchan);
512 
513 done:
514 	cli->base.super = super;
515 	return ret;
516 }
517 
518 int
nouveau_channels_init(struct nouveau_drm * drm)519 nouveau_channels_init(struct nouveau_drm *drm)
520 {
521 	struct {
522 		struct nv_device_info_v1 m;
523 		struct {
524 			struct nv_device_info_v1_data channels;
525 		} v;
526 	} args = {
527 		.m.version = 1,
528 		.m.count = sizeof(args.v) / sizeof(args.v.channels),
529 		.v.channels.mthd = NV_DEVICE_FIFO_CHANNELS,
530 	};
531 	struct nvif_object *device = &drm->client.device.object;
532 	int ret;
533 
534 	ret = nvif_object_mthd(device, NV_DEVICE_V0_INFO, &args, sizeof(args));
535 	if (ret || args.v.channels.mthd == NV_DEVICE_INFO_INVALID)
536 		return -ENODEV;
537 
538 	drm->chan.nr = args.v.channels.data;
539 	drm->chan.context_base = dma_fence_context_alloc(drm->chan.nr);
540 	return 0;
541 }
542