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 
25 #include <nvif/os.h>
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/ioctl.h>
33 
34 /*XXX*/
35 #include <core/client.h>
36 
37 #include "nouveau_drv.h"
38 #include "nouveau_dma.h"
39 #include "nouveau_bo.h"
40 #include "nouveau_chan.h"
41 #include "nouveau_fence.h"
42 #include "nouveau_abi16.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 int
49 nouveau_channel_idle(struct nouveau_channel *chan)
50 {
51 	if (likely(chan && chan->fence)) {
52 		struct nouveau_cli *cli = (void *)chan->user.client;
53 		struct nouveau_fence *fence = NULL;
54 		int ret;
55 
56 		ret = nouveau_fence_new(chan, false, &fence);
57 		if (!ret) {
58 			ret = nouveau_fence_wait(fence, false, false);
59 			nouveau_fence_unref(&fence);
60 		}
61 
62 		if (ret) {
63 			NV_PRINTK(err, cli, "failed to idle channel %d [%s]\n",
64 				  chan->chid, nvxx_client(&cli->base)->name);
65 			return ret;
66 		}
67 	}
68 	return 0;
69 }
70 
71 void
72 nouveau_channel_del(struct nouveau_channel **pchan)
73 {
74 	struct nouveau_channel *chan = *pchan;
75 	if (chan) {
76 		if (chan->fence)
77 			nouveau_fence(chan->drm)->context_del(chan);
78 		nvif_object_fini(&chan->nvsw);
79 		nvif_object_fini(&chan->gart);
80 		nvif_object_fini(&chan->vram);
81 		nvif_object_fini(&chan->user);
82 		nvif_object_fini(&chan->push.ctxdma);
83 		nouveau_bo_vma_del(chan->push.buffer, &chan->push.vma);
84 		nouveau_bo_unmap(chan->push.buffer);
85 		if (chan->push.buffer && chan->push.buffer->pin_refcnt)
86 			nouveau_bo_unpin(chan->push.buffer);
87 		nouveau_bo_ref(NULL, &chan->push.buffer);
88 		kfree(chan);
89 	}
90 	*pchan = NULL;
91 }
92 
93 static int
94 nouveau_channel_prep(struct nouveau_drm *drm, struct nvif_device *device,
95 		     u32 size, struct nouveau_channel **pchan)
96 {
97 	struct nouveau_cli *cli = (void *)device->object.client;
98 	struct nvkm_mmu *mmu = nvxx_mmu(device);
99 	struct nv_dma_v0 args = {};
100 	struct nouveau_channel *chan;
101 	u32 target;
102 	int ret;
103 
104 	chan = *pchan = kzalloc(sizeof(*chan), GFP_KERNEL);
105 	if (!chan)
106 		return -ENOMEM;
107 
108 	chan->device = device;
109 	chan->drm = drm;
110 
111 	/* allocate memory for dma push buffer */
112 	target = TTM_PL_FLAG_TT | TTM_PL_FLAG_UNCACHED;
113 	if (nouveau_vram_pushbuf)
114 		target = TTM_PL_FLAG_VRAM;
115 
116 	ret = nouveau_bo_new(drm->dev, size, 0, target, 0, 0, NULL, NULL,
117 			    &chan->push.buffer);
118 	if (ret == 0) {
119 		ret = nouveau_bo_pin(chan->push.buffer, target, false);
120 		if (ret == 0)
121 			ret = nouveau_bo_map(chan->push.buffer);
122 	}
123 
124 	if (ret) {
125 		nouveau_channel_del(pchan);
126 		return ret;
127 	}
128 
129 	/* create dma object covering the *entire* memory space that the
130 	 * pushbuf lives in, this is because the GEM code requires that
131 	 * we be able to call out to other (indirect) push buffers
132 	 */
133 	chan->push.vma.offset = chan->push.buffer->bo.offset;
134 
135 	if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
136 		ret = nouveau_bo_vma_add(chan->push.buffer, cli->vm,
137 					&chan->push.vma);
138 		if (ret) {
139 			nouveau_channel_del(pchan);
140 			return ret;
141 		}
142 
143 		args.target = NV_DMA_V0_TARGET_VM;
144 		args.access = NV_DMA_V0_ACCESS_VM;
145 		args.start = 0;
146 		args.limit = cli->vm->mmu->limit - 1;
147 	} else
148 	if (chan->push.buffer->bo.mem.mem_type == TTM_PL_VRAM) {
149 		if (device->info.family == NV_DEVICE_INFO_V0_TNT) {
150 			/* nv04 vram pushbuf hack, retarget to its location in
151 			 * the framebuffer bar rather than direct vram access..
152 			 * nfi why this exists, it came from the -nv ddx.
153 			 */
154 			args.target = NV_DMA_V0_TARGET_PCI;
155 			args.access = NV_DMA_V0_ACCESS_RDWR;
156 			args.start = nvxx_device(device)->func->
157 				resource_addr(nvxx_device(device), 1);
158 			args.limit = args.start + device->info.ram_user - 1;
159 		} else {
160 			args.target = NV_DMA_V0_TARGET_VRAM;
161 			args.access = NV_DMA_V0_ACCESS_RDWR;
162 			args.start = 0;
163 			args.limit = device->info.ram_user - 1;
164 		}
165 	} else {
166 		if (chan->drm->agp.bridge) {
167 			args.target = NV_DMA_V0_TARGET_AGP;
168 			args.access = NV_DMA_V0_ACCESS_RDWR;
169 			args.start = chan->drm->agp.base;
170 			args.limit = chan->drm->agp.base +
171 				     chan->drm->agp.size - 1;
172 		} else {
173 			args.target = NV_DMA_V0_TARGET_VM;
174 			args.access = NV_DMA_V0_ACCESS_RDWR;
175 			args.start = 0;
176 			args.limit = mmu->limit - 1;
177 		}
178 	}
179 
180 	ret = nvif_object_init(&device->object, 0, NV_DMA_FROM_MEMORY,
181 			       &args, sizeof(args), &chan->push.ctxdma);
182 	if (ret) {
183 		nouveau_channel_del(pchan);
184 		return ret;
185 	}
186 
187 	return 0;
188 }
189 
190 static int
191 nouveau_channel_ind(struct nouveau_drm *drm, struct nvif_device *device,
192 		    u32 engine, struct nouveau_channel **pchan)
193 {
194 	static const u16 oclasses[] = { PASCAL_CHANNEL_GPFIFO_A,
195 					MAXWELL_CHANNEL_GPFIFO_A,
196 					KEPLER_CHANNEL_GPFIFO_B,
197 					KEPLER_CHANNEL_GPFIFO_A,
198 					FERMI_CHANNEL_GPFIFO,
199 					G82_CHANNEL_GPFIFO,
200 					NV50_CHANNEL_GPFIFO,
201 					0 };
202 	const u16 *oclass = oclasses;
203 	union {
204 		struct nv50_channel_gpfifo_v0 nv50;
205 		struct fermi_channel_gpfifo_v0 fermi;
206 		struct kepler_channel_gpfifo_a_v0 kepler;
207 	} args;
208 	struct nouveau_channel *chan;
209 	u32 size;
210 	int ret;
211 
212 	/* allocate dma push buffer */
213 	ret = nouveau_channel_prep(drm, device, 0x12000, &chan);
214 	*pchan = chan;
215 	if (ret)
216 		return ret;
217 
218 	/* create channel object */
219 	do {
220 		if (oclass[0] >= KEPLER_CHANNEL_GPFIFO_A) {
221 			args.kepler.version = 0;
222 			args.kepler.engines = engine;
223 			args.kepler.ilength = 0x02000;
224 			args.kepler.ioffset = 0x10000 + chan->push.vma.offset;
225 			args.kepler.vm = 0;
226 			size = sizeof(args.kepler);
227 		} else
228 		if (oclass[0] >= FERMI_CHANNEL_GPFIFO) {
229 			args.fermi.version = 0;
230 			args.fermi.ilength = 0x02000;
231 			args.fermi.ioffset = 0x10000 + chan->push.vma.offset;
232 			args.fermi.vm = 0;
233 			size = sizeof(args.fermi);
234 		} else {
235 			args.nv50.version = 0;
236 			args.nv50.ilength = 0x02000;
237 			args.nv50.ioffset = 0x10000 + chan->push.vma.offset;
238 			args.nv50.pushbuf = nvif_handle(&chan->push.ctxdma);
239 			args.nv50.vm = 0;
240 			size = sizeof(args.nv50);
241 		}
242 
243 		ret = nvif_object_init(&device->object, 0, *oclass++,
244 				       &args, size, &chan->user);
245 		if (ret == 0) {
246 			if (chan->user.oclass >= KEPLER_CHANNEL_GPFIFO_A)
247 				chan->chid = args.kepler.chid;
248 			else
249 			if (chan->user.oclass >= FERMI_CHANNEL_GPFIFO)
250 				chan->chid = args.fermi.chid;
251 			else
252 				chan->chid = args.nv50.chid;
253 			return ret;
254 		}
255 	} while (*oclass);
256 
257 	nouveau_channel_del(pchan);
258 	return ret;
259 }
260 
261 static int
262 nouveau_channel_dma(struct nouveau_drm *drm, struct nvif_device *device,
263 		    struct nouveau_channel **pchan)
264 {
265 	static const u16 oclasses[] = { NV40_CHANNEL_DMA,
266 					NV17_CHANNEL_DMA,
267 					NV10_CHANNEL_DMA,
268 					NV03_CHANNEL_DMA,
269 					0 };
270 	const u16 *oclass = oclasses;
271 	struct nv03_channel_dma_v0 args;
272 	struct nouveau_channel *chan;
273 	int ret;
274 
275 	/* allocate dma push buffer */
276 	ret = nouveau_channel_prep(drm, device, 0x10000, &chan);
277 	*pchan = chan;
278 	if (ret)
279 		return ret;
280 
281 	/* create channel object */
282 	args.version = 0;
283 	args.pushbuf = nvif_handle(&chan->push.ctxdma);
284 	args.offset = chan->push.vma.offset;
285 
286 	do {
287 		ret = nvif_object_init(&device->object, 0, *oclass++,
288 				       &args, sizeof(args), &chan->user);
289 		if (ret == 0) {
290 			chan->chid = args.chid;
291 			return ret;
292 		}
293 	} while (ret && *oclass);
294 
295 	nouveau_channel_del(pchan);
296 	return ret;
297 }
298 
299 static int
300 nouveau_channel_init(struct nouveau_channel *chan, u32 vram, u32 gart)
301 {
302 	struct nvif_device *device = chan->device;
303 	struct nouveau_cli *cli = (void *)chan->user.client;
304 	struct nvkm_mmu *mmu = nvxx_mmu(device);
305 	struct nv_dma_v0 args = {};
306 	int ret, i;
307 
308 	nvif_object_map(&chan->user);
309 
310 	/* allocate dma objects to cover all allowed vram, and gart */
311 	if (device->info.family < NV_DEVICE_INFO_V0_FERMI) {
312 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
313 			args.target = NV_DMA_V0_TARGET_VM;
314 			args.access = NV_DMA_V0_ACCESS_VM;
315 			args.start = 0;
316 			args.limit = cli->vm->mmu->limit - 1;
317 		} else {
318 			args.target = NV_DMA_V0_TARGET_VRAM;
319 			args.access = NV_DMA_V0_ACCESS_RDWR;
320 			args.start = 0;
321 			args.limit = device->info.ram_user - 1;
322 		}
323 
324 		ret = nvif_object_init(&chan->user, vram, NV_DMA_IN_MEMORY,
325 				       &args, sizeof(args), &chan->vram);
326 		if (ret)
327 			return ret;
328 
329 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
330 			args.target = NV_DMA_V0_TARGET_VM;
331 			args.access = NV_DMA_V0_ACCESS_VM;
332 			args.start = 0;
333 			args.limit = cli->vm->mmu->limit - 1;
334 		} else
335 		if (chan->drm->agp.bridge) {
336 			args.target = NV_DMA_V0_TARGET_AGP;
337 			args.access = NV_DMA_V0_ACCESS_RDWR;
338 			args.start = chan->drm->agp.base;
339 			args.limit = chan->drm->agp.base +
340 				     chan->drm->agp.size - 1;
341 		} else {
342 			args.target = NV_DMA_V0_TARGET_VM;
343 			args.access = NV_DMA_V0_ACCESS_RDWR;
344 			args.start = 0;
345 			args.limit = mmu->limit - 1;
346 		}
347 
348 		ret = nvif_object_init(&chan->user, gart, NV_DMA_IN_MEMORY,
349 				       &args, sizeof(args), &chan->gart);
350 		if (ret)
351 			return ret;
352 	}
353 
354 	/* initialise dma tracking parameters */
355 	switch (chan->user.oclass & 0x00ff) {
356 	case 0x006b:
357 	case 0x006e:
358 		chan->user_put = 0x40;
359 		chan->user_get = 0x44;
360 		chan->dma.max = (0x10000 / 4) - 2;
361 		break;
362 	default:
363 		chan->user_put = 0x40;
364 		chan->user_get = 0x44;
365 		chan->user_get_hi = 0x60;
366 		chan->dma.ib_base =  0x10000 / 4;
367 		chan->dma.ib_max  = (0x02000 / 8) - 1;
368 		chan->dma.ib_put  = 0;
369 		chan->dma.ib_free = chan->dma.ib_max - chan->dma.ib_put;
370 		chan->dma.max = chan->dma.ib_base;
371 		break;
372 	}
373 
374 	chan->dma.put = 0;
375 	chan->dma.cur = chan->dma.put;
376 	chan->dma.free = chan->dma.max - chan->dma.cur;
377 
378 	ret = RING_SPACE(chan, NOUVEAU_DMA_SKIPS);
379 	if (ret)
380 		return ret;
381 
382 	for (i = 0; i < NOUVEAU_DMA_SKIPS; i++)
383 		OUT_RING(chan, 0x00000000);
384 
385 	/* allocate software object class (used for fences on <= nv05) */
386 	if (device->info.family < NV_DEVICE_INFO_V0_CELSIUS) {
387 		ret = nvif_object_init(&chan->user, 0x006e,
388 				       NVIF_CLASS_SW_NV04,
389 				       NULL, 0, &chan->nvsw);
390 		if (ret)
391 			return ret;
392 
393 		ret = RING_SPACE(chan, 2);
394 		if (ret)
395 			return ret;
396 
397 		BEGIN_NV04(chan, NvSubSw, 0x0000, 1);
398 		OUT_RING  (chan, chan->nvsw.handle);
399 		FIRE_RING (chan);
400 	}
401 
402 	/* initialise synchronisation */
403 	return nouveau_fence(chan->drm)->context_new(chan);
404 }
405 
406 int
407 nouveau_channel_new(struct nouveau_drm *drm, struct nvif_device *device,
408 		    u32 arg0, u32 arg1, struct nouveau_channel **pchan)
409 {
410 	struct nouveau_cli *cli = (void *)device->object.client;
411 	bool super;
412 	int ret;
413 
414 	/* hack until fencenv50 is fixed, and agp access relaxed */
415 	super = cli->base.super;
416 	cli->base.super = true;
417 
418 	ret = nouveau_channel_ind(drm, device, arg0, pchan);
419 	if (ret) {
420 		NV_PRINTK(dbg, cli, "ib channel create, %d\n", ret);
421 		ret = nouveau_channel_dma(drm, device, pchan);
422 		if (ret) {
423 			NV_PRINTK(dbg, cli, "dma channel create, %d\n", ret);
424 			goto done;
425 		}
426 	}
427 
428 	ret = nouveau_channel_init(*pchan, arg0, arg1);
429 	if (ret) {
430 		NV_PRINTK(err, cli, "channel failed to initialise, %d\n", ret);
431 		nouveau_channel_del(pchan);
432 	}
433 
434 done:
435 	cli->base.super = super;
436 	return ret;
437 }
438