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 
28 /*XXX*/
29 #include <core/client.h>
30 
31 #include "nouveau_drm.h"
32 #include "nouveau_dma.h"
33 #include "nouveau_bo.h"
34 #include "nouveau_chan.h"
35 #include "nouveau_fence.h"
36 #include "nouveau_abi16.h"
37 
38 MODULE_PARM_DESC(vram_pushbuf, "Create DMA push buffers in VRAM");
39 int nouveau_vram_pushbuf;
40 module_param_named(vram_pushbuf, nouveau_vram_pushbuf, int, 0400);
41 
42 int
43 nouveau_channel_idle(struct nouveau_channel *chan)
44 {
45 	struct nouveau_cli *cli = (void *)nvif_client(chan->object);
46 	struct nouveau_fence *fence = NULL;
47 	int ret;
48 
49 	ret = nouveau_fence_new(chan, false, &fence);
50 	if (!ret) {
51 		ret = nouveau_fence_wait(fence, false, false);
52 		nouveau_fence_unref(&fence);
53 	}
54 
55 	if (ret)
56 		NV_PRINTK(error, cli, "failed to idle channel 0x%08x [%s]\n",
57 			  chan->object->handle, nvkm_client(&cli->base)->name);
58 	return ret;
59 }
60 
61 void
62 nouveau_channel_del(struct nouveau_channel **pchan)
63 {
64 	struct nouveau_channel *chan = *pchan;
65 	if (chan) {
66 		if (chan->fence) {
67 			nouveau_channel_idle(chan);
68 			nouveau_fence(chan->drm)->context_del(chan);
69 		}
70 		nvif_object_fini(&chan->nvsw);
71 		nvif_object_fini(&chan->gart);
72 		nvif_object_fini(&chan->vram);
73 		nvif_object_ref(NULL, &chan->object);
74 		nvif_object_fini(&chan->push.ctxdma);
75 		nouveau_bo_vma_del(chan->push.buffer, &chan->push.vma);
76 		nouveau_bo_unmap(chan->push.buffer);
77 		if (chan->push.buffer && chan->push.buffer->pin_refcnt)
78 			nouveau_bo_unpin(chan->push.buffer);
79 		nouveau_bo_ref(NULL, &chan->push.buffer);
80 		nvif_device_ref(NULL, &chan->device);
81 		kfree(chan);
82 	}
83 	*pchan = NULL;
84 }
85 
86 static int
87 nouveau_channel_prep(struct nouveau_drm *drm, struct nvif_device *device,
88 		     u32 handle, u32 size, struct nouveau_channel **pchan)
89 {
90 	struct nouveau_cli *cli = (void *)nvif_client(&device->base);
91 	struct nouveau_vmmgr *vmm = nvkm_vmmgr(device);
92 	struct nv_dma_v0 args = {};
93 	struct nouveau_channel *chan;
94 	u32 target;
95 	int ret;
96 
97 	chan = *pchan = kzalloc(sizeof(*chan), GFP_KERNEL);
98 	if (!chan)
99 		return -ENOMEM;
100 
101 	nvif_device_ref(device, &chan->device);
102 	chan->drm = drm;
103 
104 	/* allocate memory for dma push buffer */
105 	target = TTM_PL_FLAG_TT | TTM_PL_FLAG_UNCACHED;
106 	if (nouveau_vram_pushbuf)
107 		target = TTM_PL_FLAG_VRAM;
108 
109 	ret = nouveau_bo_new(drm->dev, size, 0, target, 0, 0, NULL, NULL,
110 			    &chan->push.buffer);
111 	if (ret == 0) {
112 		ret = nouveau_bo_pin(chan->push.buffer, target, false);
113 		if (ret == 0)
114 			ret = nouveau_bo_map(chan->push.buffer);
115 	}
116 
117 	if (ret) {
118 		nouveau_channel_del(pchan);
119 		return ret;
120 	}
121 
122 	/* create dma object covering the *entire* memory space that the
123 	 * pushbuf lives in, this is because the GEM code requires that
124 	 * we be able to call out to other (indirect) push buffers
125 	 */
126 	chan->push.vma.offset = chan->push.buffer->bo.offset;
127 
128 	if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
129 		ret = nouveau_bo_vma_add(chan->push.buffer, cli->vm,
130 					&chan->push.vma);
131 		if (ret) {
132 			nouveau_channel_del(pchan);
133 			return ret;
134 		}
135 
136 		args.target = NV_DMA_V0_TARGET_VM;
137 		args.access = NV_DMA_V0_ACCESS_VM;
138 		args.start = 0;
139 		args.limit = cli->vm->vmm->limit - 1;
140 	} else
141 	if (chan->push.buffer->bo.mem.mem_type == TTM_PL_VRAM) {
142 		if (device->info.family == NV_DEVICE_INFO_V0_TNT) {
143 			/* nv04 vram pushbuf hack, retarget to its location in
144 			 * the framebuffer bar rather than direct vram access..
145 			 * nfi why this exists, it came from the -nv ddx.
146 			 */
147 			args.target = NV_DMA_V0_TARGET_PCI;
148 			args.access = NV_DMA_V0_ACCESS_RDWR;
149 			args.start = nv_device_resource_start(nvkm_device(device), 1);
150 			args.limit = args.start + device->info.ram_user - 1;
151 		} else {
152 			args.target = NV_DMA_V0_TARGET_VRAM;
153 			args.access = NV_DMA_V0_ACCESS_RDWR;
154 			args.start = 0;
155 			args.limit = device->info.ram_user - 1;
156 		}
157 	} else {
158 		if (chan->drm->agp.stat == ENABLED) {
159 			args.target = NV_DMA_V0_TARGET_AGP;
160 			args.access = NV_DMA_V0_ACCESS_RDWR;
161 			args.start = chan->drm->agp.base;
162 			args.limit = chan->drm->agp.base +
163 				     chan->drm->agp.size - 1;
164 		} else {
165 			args.target = NV_DMA_V0_TARGET_VM;
166 			args.access = NV_DMA_V0_ACCESS_RDWR;
167 			args.start = 0;
168 			args.limit = vmm->limit - 1;
169 		}
170 	}
171 
172 	ret = nvif_object_init(nvif_object(device), NULL, NVDRM_PUSH |
173 			       (handle & 0xffff), NV_DMA_FROM_MEMORY,
174 			       &args, sizeof(args), &chan->push.ctxdma);
175 	if (ret) {
176 		nouveau_channel_del(pchan);
177 		return ret;
178 	}
179 
180 	return 0;
181 }
182 
183 static int
184 nouveau_channel_ind(struct nouveau_drm *drm, struct nvif_device *device,
185 		    u32 handle, u32 engine, struct nouveau_channel **pchan)
186 {
187 	static const u16 oclasses[] = { KEPLER_CHANNEL_GPFIFO_A,
188 					FERMI_CHANNEL_GPFIFO,
189 					G82_CHANNEL_GPFIFO,
190 					NV50_CHANNEL_GPFIFO,
191 					0 };
192 	const u16 *oclass = oclasses;
193 	union {
194 		struct nv50_channel_gpfifo_v0 nv50;
195 		struct kepler_channel_gpfifo_a_v0 kepler;
196 	} args, *retn;
197 	struct nouveau_channel *chan;
198 	u32 size;
199 	int ret;
200 
201 	/* allocate dma push buffer */
202 	ret = nouveau_channel_prep(drm, device, handle, 0x12000, &chan);
203 	*pchan = chan;
204 	if (ret)
205 		return ret;
206 
207 	/* create channel object */
208 	do {
209 		if (oclass[0] >= KEPLER_CHANNEL_GPFIFO_A) {
210 			args.kepler.version = 0;
211 			args.kepler.engine  = engine;
212 			args.kepler.pushbuf = chan->push.ctxdma.handle;
213 			args.kepler.ilength = 0x02000;
214 			args.kepler.ioffset = 0x10000 + chan->push.vma.offset;
215 			size = sizeof(args.kepler);
216 		} else {
217 			args.nv50.version = 0;
218 			args.nv50.pushbuf = chan->push.ctxdma.handle;
219 			args.nv50.ilength = 0x02000;
220 			args.nv50.ioffset = 0x10000 + chan->push.vma.offset;
221 			size = sizeof(args.nv50);
222 		}
223 
224 		ret = nvif_object_new(nvif_object(device), handle, *oclass++,
225 				      &args, size, &chan->object);
226 		if (ret == 0) {
227 			retn = chan->object->data;
228 			if (chan->object->oclass >= KEPLER_CHANNEL_GPFIFO_A)
229 				chan->chid = retn->kepler.chid;
230 			else
231 				chan->chid = retn->nv50.chid;
232 			return ret;
233 		}
234 	} while (*oclass);
235 
236 	nouveau_channel_del(pchan);
237 	return ret;
238 }
239 
240 static int
241 nouveau_channel_dma(struct nouveau_drm *drm, struct nvif_device *device,
242 		    u32 handle, struct nouveau_channel **pchan)
243 {
244 	static const u16 oclasses[] = { NV40_CHANNEL_DMA,
245 					NV17_CHANNEL_DMA,
246 					NV10_CHANNEL_DMA,
247 					NV03_CHANNEL_DMA,
248 					0 };
249 	const u16 *oclass = oclasses;
250 	struct nv03_channel_dma_v0 args, *retn;
251 	struct nouveau_channel *chan;
252 	int ret;
253 
254 	/* allocate dma push buffer */
255 	ret = nouveau_channel_prep(drm, device, handle, 0x10000, &chan);
256 	*pchan = chan;
257 	if (ret)
258 		return ret;
259 
260 	/* create channel object */
261 	args.version = 0;
262 	args.pushbuf = chan->push.ctxdma.handle;
263 	args.offset = chan->push.vma.offset;
264 
265 	do {
266 		ret = nvif_object_new(nvif_object(device), handle, *oclass++,
267 				      &args, sizeof(args), &chan->object);
268 		if (ret == 0) {
269 			retn = chan->object->data;
270 			chan->chid = retn->chid;
271 			return ret;
272 		}
273 	} while (ret && *oclass);
274 
275 	nouveau_channel_del(pchan);
276 	return ret;
277 }
278 
279 static int
280 nouveau_channel_init(struct nouveau_channel *chan, u32 vram, u32 gart)
281 {
282 	struct nvif_device *device = chan->device;
283 	struct nouveau_cli *cli = (void *)nvif_client(&device->base);
284 	struct nouveau_vmmgr *vmm = nvkm_vmmgr(device);
285 	struct nouveau_software_chan *swch;
286 	struct nv_dma_v0 args = {};
287 	int ret, i;
288 
289 	nvif_object_map(chan->object);
290 
291 	/* allocate dma objects to cover all allowed vram, and gart */
292 	if (device->info.family < NV_DEVICE_INFO_V0_FERMI) {
293 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
294 			args.target = NV_DMA_V0_TARGET_VM;
295 			args.access = NV_DMA_V0_ACCESS_VM;
296 			args.start = 0;
297 			args.limit = cli->vm->vmm->limit - 1;
298 		} else {
299 			args.target = NV_DMA_V0_TARGET_VRAM;
300 			args.access = NV_DMA_V0_ACCESS_RDWR;
301 			args.start = 0;
302 			args.limit = device->info.ram_user - 1;
303 		}
304 
305 		ret = nvif_object_init(chan->object, NULL, vram,
306 				       NV_DMA_IN_MEMORY, &args,
307 				       sizeof(args), &chan->vram);
308 		if (ret)
309 			return ret;
310 
311 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
312 			args.target = NV_DMA_V0_TARGET_VM;
313 			args.access = NV_DMA_V0_ACCESS_VM;
314 			args.start = 0;
315 			args.limit = cli->vm->vmm->limit - 1;
316 		} else
317 		if (chan->drm->agp.stat == ENABLED) {
318 			args.target = NV_DMA_V0_TARGET_AGP;
319 			args.access = NV_DMA_V0_ACCESS_RDWR;
320 			args.start = chan->drm->agp.base;
321 			args.limit = chan->drm->agp.base +
322 				     chan->drm->agp.size - 1;
323 		} else {
324 			args.target = NV_DMA_V0_TARGET_VM;
325 			args.access = NV_DMA_V0_ACCESS_RDWR;
326 			args.start = 0;
327 			args.limit = vmm->limit - 1;
328 		}
329 
330 		ret = nvif_object_init(chan->object, NULL, gart,
331 				       NV_DMA_IN_MEMORY, &args,
332 				       sizeof(args), &chan->gart);
333 		if (ret)
334 			return ret;
335 	}
336 
337 	/* initialise dma tracking parameters */
338 	switch (chan->object->oclass & 0x00ff) {
339 	case 0x006b:
340 	case 0x006e:
341 		chan->user_put = 0x40;
342 		chan->user_get = 0x44;
343 		chan->dma.max = (0x10000 / 4) - 2;
344 		break;
345 	default:
346 		chan->user_put = 0x40;
347 		chan->user_get = 0x44;
348 		chan->user_get_hi = 0x60;
349 		chan->dma.ib_base =  0x10000 / 4;
350 		chan->dma.ib_max  = (0x02000 / 8) - 1;
351 		chan->dma.ib_put  = 0;
352 		chan->dma.ib_free = chan->dma.ib_max - chan->dma.ib_put;
353 		chan->dma.max = chan->dma.ib_base;
354 		break;
355 	}
356 
357 	chan->dma.put = 0;
358 	chan->dma.cur = chan->dma.put;
359 	chan->dma.free = chan->dma.max - chan->dma.cur;
360 
361 	ret = RING_SPACE(chan, NOUVEAU_DMA_SKIPS);
362 	if (ret)
363 		return ret;
364 
365 	for (i = 0; i < NOUVEAU_DMA_SKIPS; i++)
366 		OUT_RING(chan, 0x00000000);
367 
368 	/* allocate software object class (used for fences on <= nv05) */
369 	if (device->info.family < NV_DEVICE_INFO_V0_CELSIUS) {
370 		ret = nvif_object_init(chan->object, NULL, 0x006e, 0x006e,
371 				       NULL, 0, &chan->nvsw);
372 		if (ret)
373 			return ret;
374 
375 		swch = (void *)nvkm_object(&chan->nvsw)->parent;
376 		swch->flip = nouveau_flip_complete;
377 		swch->flip_data = chan;
378 
379 		ret = RING_SPACE(chan, 2);
380 		if (ret)
381 			return ret;
382 
383 		BEGIN_NV04(chan, NvSubSw, 0x0000, 1);
384 		OUT_RING  (chan, chan->nvsw.handle);
385 		FIRE_RING (chan);
386 	}
387 
388 	/* initialise synchronisation */
389 	return nouveau_fence(chan->drm)->context_new(chan);
390 }
391 
392 int
393 nouveau_channel_new(struct nouveau_drm *drm, struct nvif_device *device,
394 		    u32 handle, u32 arg0, u32 arg1,
395 		    struct nouveau_channel **pchan)
396 {
397 	struct nouveau_cli *cli = (void *)nvif_client(&device->base);
398 	bool super;
399 	int ret;
400 
401 	/* hack until fencenv50 is fixed, and agp access relaxed */
402 	super = cli->base.super;
403 	cli->base.super = true;
404 
405 	ret = nouveau_channel_ind(drm, device, handle, arg0, pchan);
406 	if (ret) {
407 		NV_PRINTK(debug, cli, "ib channel create, %d\n", ret);
408 		ret = nouveau_channel_dma(drm, device, handle, pchan);
409 		if (ret) {
410 			NV_PRINTK(debug, cli, "dma channel create, %d\n", ret);
411 			goto done;
412 		}
413 	}
414 
415 	ret = nouveau_channel_init(*pchan, arg0, arg1);
416 	if (ret) {
417 		NV_PRINTK(error, cli, "channel failed to initialise, %d\n", ret);
418 		nouveau_channel_del(pchan);
419 	}
420 
421 done:
422 	cli->base.super = super;
423 	return ret;
424 }
425