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_drm.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[] = { MAXWELL_CHANNEL_GPFIFO_A,
195 					KEPLER_CHANNEL_GPFIFO_A,
196 					FERMI_CHANNEL_GPFIFO,
197 					G82_CHANNEL_GPFIFO,
198 					NV50_CHANNEL_GPFIFO,
199 					0 };
200 	const u16 *oclass = oclasses;
201 	union {
202 		struct nv50_channel_gpfifo_v0 nv50;
203 		struct fermi_channel_gpfifo_v0 fermi;
204 		struct kepler_channel_gpfifo_a_v0 kepler;
205 	} args;
206 	struct nouveau_channel *chan;
207 	u32 size;
208 	int ret;
209 
210 	/* allocate dma push buffer */
211 	ret = nouveau_channel_prep(drm, device, 0x12000, &chan);
212 	*pchan = chan;
213 	if (ret)
214 		return ret;
215 
216 	/* create channel object */
217 	do {
218 		if (oclass[0] >= KEPLER_CHANNEL_GPFIFO_A) {
219 			args.kepler.version = 0;
220 			args.kepler.engine  = engine;
221 			args.kepler.ilength = 0x02000;
222 			args.kepler.ioffset = 0x10000 + chan->push.vma.offset;
223 			args.kepler.vm = 0;
224 			size = sizeof(args.kepler);
225 		} else
226 		if (oclass[0] >= FERMI_CHANNEL_GPFIFO) {
227 			args.fermi.version = 0;
228 			args.fermi.ilength = 0x02000;
229 			args.fermi.ioffset = 0x10000 + chan->push.vma.offset;
230 			args.fermi.vm = 0;
231 			size = sizeof(args.fermi);
232 		} else {
233 			args.nv50.version = 0;
234 			args.nv50.ilength = 0x02000;
235 			args.nv50.ioffset = 0x10000 + chan->push.vma.offset;
236 			args.nv50.pushbuf = nvif_handle(&chan->push.ctxdma);
237 			args.nv50.vm = 0;
238 			size = sizeof(args.nv50);
239 		}
240 
241 		ret = nvif_object_init(&device->object, 0, *oclass++,
242 				       &args, size, &chan->user);
243 		if (ret == 0) {
244 			if (chan->user.oclass >= KEPLER_CHANNEL_GPFIFO_A)
245 				chan->chid = args.kepler.chid;
246 			else
247 			if (chan->user.oclass >= FERMI_CHANNEL_GPFIFO)
248 				chan->chid = args.fermi.chid;
249 			else
250 				chan->chid = args.nv50.chid;
251 			return ret;
252 		}
253 	} while (*oclass);
254 
255 	nouveau_channel_del(pchan);
256 	return ret;
257 }
258 
259 static int
260 nouveau_channel_dma(struct nouveau_drm *drm, struct nvif_device *device,
261 		    struct nouveau_channel **pchan)
262 {
263 	static const u16 oclasses[] = { NV40_CHANNEL_DMA,
264 					NV17_CHANNEL_DMA,
265 					NV10_CHANNEL_DMA,
266 					NV03_CHANNEL_DMA,
267 					0 };
268 	const u16 *oclass = oclasses;
269 	struct nv03_channel_dma_v0 args;
270 	struct nouveau_channel *chan;
271 	int ret;
272 
273 	/* allocate dma push buffer */
274 	ret = nouveau_channel_prep(drm, device, 0x10000, &chan);
275 	*pchan = chan;
276 	if (ret)
277 		return ret;
278 
279 	/* create channel object */
280 	args.version = 0;
281 	args.pushbuf = nvif_handle(&chan->push.ctxdma);
282 	args.offset = chan->push.vma.offset;
283 
284 	do {
285 		ret = nvif_object_init(&device->object, 0, *oclass++,
286 				       &args, sizeof(args), &chan->user);
287 		if (ret == 0) {
288 			chan->chid = args.chid;
289 			return ret;
290 		}
291 	} while (ret && *oclass);
292 
293 	nouveau_channel_del(pchan);
294 	return ret;
295 }
296 
297 static int
298 nouveau_channel_init(struct nouveau_channel *chan, u32 vram, u32 gart)
299 {
300 	struct nvif_device *device = chan->device;
301 	struct nouveau_cli *cli = (void *)chan->user.client;
302 	struct nvkm_mmu *mmu = nvxx_mmu(device);
303 	struct nv_dma_v0 args = {};
304 	int ret, i;
305 
306 	nvif_object_map(&chan->user);
307 
308 	/* allocate dma objects to cover all allowed vram, and gart */
309 	if (device->info.family < NV_DEVICE_INFO_V0_FERMI) {
310 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
311 			args.target = NV_DMA_V0_TARGET_VM;
312 			args.access = NV_DMA_V0_ACCESS_VM;
313 			args.start = 0;
314 			args.limit = cli->vm->mmu->limit - 1;
315 		} else {
316 			args.target = NV_DMA_V0_TARGET_VRAM;
317 			args.access = NV_DMA_V0_ACCESS_RDWR;
318 			args.start = 0;
319 			args.limit = device->info.ram_user - 1;
320 		}
321 
322 		ret = nvif_object_init(&chan->user, vram, NV_DMA_IN_MEMORY,
323 				       &args, sizeof(args), &chan->vram);
324 		if (ret)
325 			return ret;
326 
327 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
328 			args.target = NV_DMA_V0_TARGET_VM;
329 			args.access = NV_DMA_V0_ACCESS_VM;
330 			args.start = 0;
331 			args.limit = cli->vm->mmu->limit - 1;
332 		} else
333 		if (chan->drm->agp.bridge) {
334 			args.target = NV_DMA_V0_TARGET_AGP;
335 			args.access = NV_DMA_V0_ACCESS_RDWR;
336 			args.start = chan->drm->agp.base;
337 			args.limit = chan->drm->agp.base +
338 				     chan->drm->agp.size - 1;
339 		} else {
340 			args.target = NV_DMA_V0_TARGET_VM;
341 			args.access = NV_DMA_V0_ACCESS_RDWR;
342 			args.start = 0;
343 			args.limit = mmu->limit - 1;
344 		}
345 
346 		ret = nvif_object_init(&chan->user, gart, NV_DMA_IN_MEMORY,
347 				       &args, sizeof(args), &chan->gart);
348 		if (ret)
349 			return ret;
350 	}
351 
352 	/* initialise dma tracking parameters */
353 	switch (chan->user.oclass & 0x00ff) {
354 	case 0x006b:
355 	case 0x006e:
356 		chan->user_put = 0x40;
357 		chan->user_get = 0x44;
358 		chan->dma.max = (0x10000 / 4) - 2;
359 		break;
360 	default:
361 		chan->user_put = 0x40;
362 		chan->user_get = 0x44;
363 		chan->user_get_hi = 0x60;
364 		chan->dma.ib_base =  0x10000 / 4;
365 		chan->dma.ib_max  = (0x02000 / 8) - 1;
366 		chan->dma.ib_put  = 0;
367 		chan->dma.ib_free = chan->dma.ib_max - chan->dma.ib_put;
368 		chan->dma.max = chan->dma.ib_base;
369 		break;
370 	}
371 
372 	chan->dma.put = 0;
373 	chan->dma.cur = chan->dma.put;
374 	chan->dma.free = chan->dma.max - chan->dma.cur;
375 
376 	ret = RING_SPACE(chan, NOUVEAU_DMA_SKIPS);
377 	if (ret)
378 		return ret;
379 
380 	for (i = 0; i < NOUVEAU_DMA_SKIPS; i++)
381 		OUT_RING(chan, 0x00000000);
382 
383 	/* allocate software object class (used for fences on <= nv05) */
384 	if (device->info.family < NV_DEVICE_INFO_V0_CELSIUS) {
385 		ret = nvif_object_init(&chan->user, 0x006e,
386 				       NVIF_CLASS_SW_NV04,
387 				       NULL, 0, &chan->nvsw);
388 		if (ret)
389 			return ret;
390 
391 		ret = RING_SPACE(chan, 2);
392 		if (ret)
393 			return ret;
394 
395 		BEGIN_NV04(chan, NvSubSw, 0x0000, 1);
396 		OUT_RING  (chan, chan->nvsw.handle);
397 		FIRE_RING (chan);
398 	}
399 
400 	/* initialise synchronisation */
401 	return nouveau_fence(chan->drm)->context_new(chan);
402 }
403 
404 int
405 nouveau_channel_new(struct nouveau_drm *drm, struct nvif_device *device,
406 		    u32 arg0, u32 arg1, struct nouveau_channel **pchan)
407 {
408 	struct nouveau_cli *cli = (void *)device->object.client;
409 	bool super;
410 	int ret;
411 
412 	/* hack until fencenv50 is fixed, and agp access relaxed */
413 	super = cli->base.super;
414 	cli->base.super = true;
415 
416 	ret = nouveau_channel_ind(drm, device, arg0, pchan);
417 	if (ret) {
418 		NV_PRINTK(dbg, cli, "ib channel create, %d\n", ret);
419 		ret = nouveau_channel_dma(drm, device, pchan);
420 		if (ret) {
421 			NV_PRINTK(dbg, cli, "dma channel create, %d\n", ret);
422 			goto done;
423 		}
424 	}
425 
426 	ret = nouveau_channel_init(*pchan, arg0, arg1);
427 	if (ret) {
428 		NV_PRINTK(err, cli, "channel failed to initialise, %d\n", ret);
429 		nouveau_channel_del(pchan);
430 	}
431 
432 done:
433 	cli->base.super = super;
434 	return ret;
435 }
436