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/ioctl.h>
28 
29 /*XXX*/
30 #include <core/client.h>
31 
32 #include "nouveau_drm.h"
33 #include "nouveau_dma.h"
34 #include "nouveau_bo.h"
35 #include "nouveau_chan.h"
36 #include "nouveau_fence.h"
37 #include "nouveau_abi16.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 int
44 nouveau_channel_idle(struct nouveau_channel *chan)
45 {
46 	if (likely(chan && chan->fence)) {
47 		struct nouveau_cli *cli = (void *)chan->user.client;
48 		struct nouveau_fence *fence = NULL;
49 		int ret;
50 
51 		ret = nouveau_fence_new(chan, false, &fence);
52 		if (!ret) {
53 			ret = nouveau_fence_wait(fence, false, false);
54 			nouveau_fence_unref(&fence);
55 		}
56 
57 		if (ret) {
58 			NV_PRINTK(err, cli, "failed to idle channel %d [%s]\n",
59 				  chan->chid, nvxx_client(&cli->base)->name);
60 			return ret;
61 		}
62 	}
63 	return 0;
64 }
65 
66 void
67 nouveau_channel_del(struct nouveau_channel **pchan)
68 {
69 	struct nouveau_channel *chan = *pchan;
70 	if (chan) {
71 		if (chan->fence)
72 			nouveau_fence(chan->drm)->context_del(chan);
73 		nvif_object_fini(&chan->nvsw);
74 		nvif_object_fini(&chan->gart);
75 		nvif_object_fini(&chan->vram);
76 		nvif_object_fini(&chan->user);
77 		nvif_object_fini(&chan->push.ctxdma);
78 		nouveau_bo_vma_del(chan->push.buffer, &chan->push.vma);
79 		nouveau_bo_unmap(chan->push.buffer);
80 		if (chan->push.buffer && chan->push.buffer->pin_refcnt)
81 			nouveau_bo_unpin(chan->push.buffer);
82 		nouveau_bo_ref(NULL, &chan->push.buffer);
83 		kfree(chan);
84 	}
85 	*pchan = NULL;
86 }
87 
88 static int
89 nouveau_channel_prep(struct nouveau_drm *drm, struct nvif_device *device,
90 		     u32 size, struct nouveau_channel **pchan)
91 {
92 	struct nouveau_cli *cli = (void *)device->object.client;
93 	struct nvkm_mmu *mmu = nvxx_mmu(device);
94 	struct nv_dma_v0 args = {};
95 	struct nouveau_channel *chan;
96 	u32 target;
97 	int ret;
98 
99 	chan = *pchan = kzalloc(sizeof(*chan), GFP_KERNEL);
100 	if (!chan)
101 		return -ENOMEM;
102 
103 	chan->device = device;
104 	chan->drm = drm;
105 
106 	/* allocate memory for dma push buffer */
107 	target = TTM_PL_FLAG_TT | TTM_PL_FLAG_UNCACHED;
108 	if (nouveau_vram_pushbuf)
109 		target = TTM_PL_FLAG_VRAM;
110 
111 	ret = nouveau_bo_new(drm->dev, size, 0, target, 0, 0, NULL, NULL,
112 			    &chan->push.buffer);
113 	if (ret == 0) {
114 		ret = nouveau_bo_pin(chan->push.buffer, target, false);
115 		if (ret == 0)
116 			ret = nouveau_bo_map(chan->push.buffer);
117 	}
118 
119 	if (ret) {
120 		nouveau_channel_del(pchan);
121 		return ret;
122 	}
123 
124 	/* create dma object covering the *entire* memory space that the
125 	 * pushbuf lives in, this is because the GEM code requires that
126 	 * we be able to call out to other (indirect) push buffers
127 	 */
128 	chan->push.vma.offset = chan->push.buffer->bo.offset;
129 
130 	if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
131 		ret = nouveau_bo_vma_add(chan->push.buffer, cli->vm,
132 					&chan->push.vma);
133 		if (ret) {
134 			nouveau_channel_del(pchan);
135 			return ret;
136 		}
137 
138 		args.target = NV_DMA_V0_TARGET_VM;
139 		args.access = NV_DMA_V0_ACCESS_VM;
140 		args.start = 0;
141 		args.limit = cli->vm->mmu->limit - 1;
142 	} else
143 	if (chan->push.buffer->bo.mem.mem_type == TTM_PL_VRAM) {
144 		if (device->info.family == NV_DEVICE_INFO_V0_TNT) {
145 			/* nv04 vram pushbuf hack, retarget to its location in
146 			 * the framebuffer bar rather than direct vram access..
147 			 * nfi why this exists, it came from the -nv ddx.
148 			 */
149 			args.target = NV_DMA_V0_TARGET_PCI;
150 			args.access = NV_DMA_V0_ACCESS_RDWR;
151 			args.start = nvxx_device(device)->func->
152 				resource_addr(nvxx_device(device), 1);
153 			args.limit = args.start + device->info.ram_user - 1;
154 		} else {
155 			args.target = NV_DMA_V0_TARGET_VRAM;
156 			args.access = NV_DMA_V0_ACCESS_RDWR;
157 			args.start = 0;
158 			args.limit = device->info.ram_user - 1;
159 		}
160 	} else {
161 		if (chan->drm->agp.bridge) {
162 			args.target = NV_DMA_V0_TARGET_AGP;
163 			args.access = NV_DMA_V0_ACCESS_RDWR;
164 			args.start = chan->drm->agp.base;
165 			args.limit = chan->drm->agp.base +
166 				     chan->drm->agp.size - 1;
167 		} else {
168 			args.target = NV_DMA_V0_TARGET_VM;
169 			args.access = NV_DMA_V0_ACCESS_RDWR;
170 			args.start = 0;
171 			args.limit = mmu->limit - 1;
172 		}
173 	}
174 
175 	ret = nvif_object_init(&device->object, 0, NV_DMA_FROM_MEMORY,
176 			       &args, sizeof(args), &chan->push.ctxdma);
177 	if (ret) {
178 		nouveau_channel_del(pchan);
179 		return ret;
180 	}
181 
182 	return 0;
183 }
184 
185 static int
186 nouveau_channel_ind(struct nouveau_drm *drm, struct nvif_device *device,
187 		    u32 engine, struct nouveau_channel **pchan)
188 {
189 	static const u16 oclasses[] = { MAXWELL_CHANNEL_GPFIFO_A,
190 					KEPLER_CHANNEL_GPFIFO_A,
191 					FERMI_CHANNEL_GPFIFO,
192 					G82_CHANNEL_GPFIFO,
193 					NV50_CHANNEL_GPFIFO,
194 					0 };
195 	const u16 *oclass = oclasses;
196 	union {
197 		struct nv50_channel_gpfifo_v0 nv50;
198 		struct fermi_channel_gpfifo_v0 fermi;
199 		struct kepler_channel_gpfifo_a_v0 kepler;
200 	} args;
201 	struct nouveau_channel *chan;
202 	u32 size;
203 	int ret;
204 
205 	/* allocate dma push buffer */
206 	ret = nouveau_channel_prep(drm, device, 0x12000, &chan);
207 	*pchan = chan;
208 	if (ret)
209 		return ret;
210 
211 	/* create channel object */
212 	do {
213 		if (oclass[0] >= KEPLER_CHANNEL_GPFIFO_A) {
214 			args.kepler.version = 0;
215 			args.kepler.engine  = engine;
216 			args.kepler.ilength = 0x02000;
217 			args.kepler.ioffset = 0x10000 + chan->push.vma.offset;
218 			args.kepler.vm = 0;
219 			size = sizeof(args.kepler);
220 		} else
221 		if (oclass[0] >= FERMI_CHANNEL_GPFIFO) {
222 			args.fermi.version = 0;
223 			args.fermi.ilength = 0x02000;
224 			args.fermi.ioffset = 0x10000 + chan->push.vma.offset;
225 			args.fermi.vm = 0;
226 			size = sizeof(args.fermi);
227 		} else {
228 			args.nv50.version = 0;
229 			args.nv50.ilength = 0x02000;
230 			args.nv50.ioffset = 0x10000 + chan->push.vma.offset;
231 			args.nv50.pushbuf = nvif_handle(&chan->push.ctxdma);
232 			args.nv50.vm = 0;
233 			size = sizeof(args.nv50);
234 		}
235 
236 		ret = nvif_object_init(&device->object, 0, *oclass++,
237 				       &args, size, &chan->user);
238 		if (ret == 0) {
239 			if (chan->user.oclass >= KEPLER_CHANNEL_GPFIFO_A)
240 				chan->chid = args.kepler.chid;
241 			else
242 			if (chan->user.oclass >= FERMI_CHANNEL_GPFIFO)
243 				chan->chid = args.fermi.chid;
244 			else
245 				chan->chid = args.nv50.chid;
246 			return ret;
247 		}
248 	} while (*oclass);
249 
250 	nouveau_channel_del(pchan);
251 	return ret;
252 }
253 
254 static int
255 nouveau_channel_dma(struct nouveau_drm *drm, struct nvif_device *device,
256 		    struct nouveau_channel **pchan)
257 {
258 	static const u16 oclasses[] = { NV40_CHANNEL_DMA,
259 					NV17_CHANNEL_DMA,
260 					NV10_CHANNEL_DMA,
261 					NV03_CHANNEL_DMA,
262 					0 };
263 	const u16 *oclass = oclasses;
264 	struct nv03_channel_dma_v0 args;
265 	struct nouveau_channel *chan;
266 	int ret;
267 
268 	/* allocate dma push buffer */
269 	ret = nouveau_channel_prep(drm, device, 0x10000, &chan);
270 	*pchan = chan;
271 	if (ret)
272 		return ret;
273 
274 	/* create channel object */
275 	args.version = 0;
276 	args.pushbuf = nvif_handle(&chan->push.ctxdma);
277 	args.offset = chan->push.vma.offset;
278 
279 	do {
280 		ret = nvif_object_init(&device->object, 0, *oclass++,
281 				       &args, sizeof(args), &chan->user);
282 		if (ret == 0) {
283 			chan->chid = args.chid;
284 			return ret;
285 		}
286 	} while (ret && *oclass);
287 
288 	nouveau_channel_del(pchan);
289 	return ret;
290 }
291 
292 static int
293 nouveau_channel_init(struct nouveau_channel *chan, u32 vram, u32 gart)
294 {
295 	struct nvif_device *device = chan->device;
296 	struct nouveau_cli *cli = (void *)chan->user.client;
297 	struct nvkm_mmu *mmu = nvxx_mmu(device);
298 	struct nv_dma_v0 args = {};
299 	int ret, i;
300 
301 	nvif_object_map(&chan->user);
302 
303 	/* allocate dma objects to cover all allowed vram, and gart */
304 	if (device->info.family < NV_DEVICE_INFO_V0_FERMI) {
305 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
306 			args.target = NV_DMA_V0_TARGET_VM;
307 			args.access = NV_DMA_V0_ACCESS_VM;
308 			args.start = 0;
309 			args.limit = cli->vm->mmu->limit - 1;
310 		} else {
311 			args.target = NV_DMA_V0_TARGET_VRAM;
312 			args.access = NV_DMA_V0_ACCESS_RDWR;
313 			args.start = 0;
314 			args.limit = device->info.ram_user - 1;
315 		}
316 
317 		ret = nvif_object_init(&chan->user, vram, NV_DMA_IN_MEMORY,
318 				       &args, sizeof(args), &chan->vram);
319 		if (ret)
320 			return ret;
321 
322 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
323 			args.target = NV_DMA_V0_TARGET_VM;
324 			args.access = NV_DMA_V0_ACCESS_VM;
325 			args.start = 0;
326 			args.limit = cli->vm->mmu->limit - 1;
327 		} else
328 		if (chan->drm->agp.bridge) {
329 			args.target = NV_DMA_V0_TARGET_AGP;
330 			args.access = NV_DMA_V0_ACCESS_RDWR;
331 			args.start = chan->drm->agp.base;
332 			args.limit = chan->drm->agp.base +
333 				     chan->drm->agp.size - 1;
334 		} else {
335 			args.target = NV_DMA_V0_TARGET_VM;
336 			args.access = NV_DMA_V0_ACCESS_RDWR;
337 			args.start = 0;
338 			args.limit = mmu->limit - 1;
339 		}
340 
341 		ret = nvif_object_init(&chan->user, gart, NV_DMA_IN_MEMORY,
342 				       &args, sizeof(args), &chan->gart);
343 		if (ret)
344 			return ret;
345 	}
346 
347 	/* initialise dma tracking parameters */
348 	switch (chan->user.oclass & 0x00ff) {
349 	case 0x006b:
350 	case 0x006e:
351 		chan->user_put = 0x40;
352 		chan->user_get = 0x44;
353 		chan->dma.max = (0x10000 / 4) - 2;
354 		break;
355 	default:
356 		chan->user_put = 0x40;
357 		chan->user_get = 0x44;
358 		chan->user_get_hi = 0x60;
359 		chan->dma.ib_base =  0x10000 / 4;
360 		chan->dma.ib_max  = (0x02000 / 8) - 1;
361 		chan->dma.ib_put  = 0;
362 		chan->dma.ib_free = chan->dma.ib_max - chan->dma.ib_put;
363 		chan->dma.max = chan->dma.ib_base;
364 		break;
365 	}
366 
367 	chan->dma.put = 0;
368 	chan->dma.cur = chan->dma.put;
369 	chan->dma.free = chan->dma.max - chan->dma.cur;
370 
371 	ret = RING_SPACE(chan, NOUVEAU_DMA_SKIPS);
372 	if (ret)
373 		return ret;
374 
375 	for (i = 0; i < NOUVEAU_DMA_SKIPS; i++)
376 		OUT_RING(chan, 0x00000000);
377 
378 	/* allocate software object class (used for fences on <= nv05) */
379 	if (device->info.family < NV_DEVICE_INFO_V0_CELSIUS) {
380 		ret = nvif_object_init(&chan->user, 0x006e,
381 				       NVIF_IOCTL_NEW_V0_SW_NV04,
382 				       NULL, 0, &chan->nvsw);
383 		if (ret)
384 			return ret;
385 
386 		ret = RING_SPACE(chan, 2);
387 		if (ret)
388 			return ret;
389 
390 		BEGIN_NV04(chan, NvSubSw, 0x0000, 1);
391 		OUT_RING  (chan, chan->nvsw.handle);
392 		FIRE_RING (chan);
393 	}
394 
395 	/* initialise synchronisation */
396 	return nouveau_fence(chan->drm)->context_new(chan);
397 }
398 
399 int
400 nouveau_channel_new(struct nouveau_drm *drm, struct nvif_device *device,
401 		    u32 arg0, u32 arg1, struct nouveau_channel **pchan)
402 {
403 	struct nouveau_cli *cli = (void *)device->object.client;
404 	bool super;
405 	int ret;
406 
407 	/* hack until fencenv50 is fixed, and agp access relaxed */
408 	super = cli->base.super;
409 	cli->base.super = true;
410 
411 	ret = nouveau_channel_ind(drm, device, arg0, pchan);
412 	if (ret) {
413 		NV_PRINTK(dbg, cli, "ib channel create, %d\n", ret);
414 		ret = nouveau_channel_dma(drm, device, pchan);
415 		if (ret) {
416 			NV_PRINTK(dbg, cli, "dma channel create, %d\n", ret);
417 			goto done;
418 		}
419 	}
420 
421 	ret = nouveau_channel_init(*pchan, arg0, arg1);
422 	if (ret) {
423 		NV_PRINTK(err, cli, "channel failed to initialise, %d\n", ret);
424 		nouveau_channel_del(pchan);
425 	}
426 
427 done:
428 	cli->base.super = super;
429 	return ret;
430 }
431