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 	struct nouveau_cli *cli = (void *)chan->user.client;
47 	struct nouveau_fence *fence = NULL;
48 	int ret;
49 
50 	ret = nouveau_fence_new(chan, false, &fence);
51 	if (!ret) {
52 		ret = nouveau_fence_wait(fence, false, false);
53 		nouveau_fence_unref(&fence);
54 	}
55 
56 	if (ret)
57 		NV_PRINTK(err, cli, "failed to idle channel 0x%08x [%s]\n",
58 			  chan->user.handle, nvxx_client(&cli->base)->name);
59 	return ret;
60 }
61 
62 void
63 nouveau_channel_del(struct nouveau_channel **pchan)
64 {
65 	struct nouveau_channel *chan = *pchan;
66 	if (chan) {
67 		if (chan->fence) {
68 			nouveau_channel_idle(chan);
69 			nouveau_fence(chan->drm)->context_del(chan);
70 		}
71 		nvif_object_fini(&chan->nvsw);
72 		nvif_object_fini(&chan->gart);
73 		nvif_object_fini(&chan->vram);
74 		nvif_object_fini(&chan->user);
75 		nvif_object_fini(&chan->push.ctxdma);
76 		nouveau_bo_vma_del(chan->push.buffer, &chan->push.vma);
77 		nouveau_bo_unmap(chan->push.buffer);
78 		if (chan->push.buffer && chan->push.buffer->pin_refcnt)
79 			nouveau_bo_unpin(chan->push.buffer);
80 		nouveau_bo_ref(NULL, &chan->push.buffer);
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 *)device->object.client;
91 	struct nvkm_mmu *mmu = nvxx_mmu(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 	chan->device = 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->mmu->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(nvxx_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 = mmu->limit - 1;
169 		}
170 	}
171 
172 	ret = nvif_object_init(&device->object, 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[] = { MAXWELL_CHANNEL_GPFIFO_A,
188 					KEPLER_CHANNEL_GPFIFO_A,
189 					FERMI_CHANNEL_GPFIFO,
190 					G82_CHANNEL_GPFIFO,
191 					NV50_CHANNEL_GPFIFO,
192 					0 };
193 	const u16 *oclass = oclasses;
194 	union {
195 		struct nv50_channel_gpfifo_v0 nv50;
196 		struct fermi_channel_gpfifo_v0 fermi;
197 		struct kepler_channel_gpfifo_a_v0 kepler;
198 	} args;
199 	struct nouveau_channel *chan;
200 	u32 size;
201 	int ret;
202 
203 	/* allocate dma push buffer */
204 	ret = nouveau_channel_prep(drm, device, handle, 0x12000, &chan);
205 	*pchan = chan;
206 	if (ret)
207 		return ret;
208 
209 	/* create channel object */
210 	do {
211 		if (oclass[0] >= KEPLER_CHANNEL_GPFIFO_A) {
212 			args.kepler.version = 0;
213 			args.kepler.engine  = engine;
214 			args.kepler.ilength = 0x02000;
215 			args.kepler.ioffset = 0x10000 + chan->push.vma.offset;
216 			args.kepler.vm = 0;
217 			size = sizeof(args.kepler);
218 		} else
219 		if (oclass[0] >= FERMI_CHANNEL_GPFIFO) {
220 			args.fermi.version = 0;
221 			args.fermi.ilength = 0x02000;
222 			args.fermi.ioffset = 0x10000 + chan->push.vma.offset;
223 			args.fermi.vm = 0;
224 			size = sizeof(args.fermi);
225 		} else {
226 			args.nv50.version = 0;
227 			args.nv50.ilength = 0x02000;
228 			args.nv50.ioffset = 0x10000 + chan->push.vma.offset;
229 			args.nv50.pushbuf = nvif_handle(&chan->push.ctxdma);
230 			args.nv50.vm = 0;
231 			size = sizeof(args.nv50);
232 		}
233 
234 		ret = nvif_object_init(&device->object, handle, *oclass++,
235 				       &args, size, &chan->user);
236 		if (ret == 0) {
237 			if (chan->user.oclass >= KEPLER_CHANNEL_GPFIFO_A)
238 				chan->chid = args.kepler.chid;
239 			else
240 			if (chan->user.oclass >= FERMI_CHANNEL_GPFIFO)
241 				chan->chid = args.fermi.chid;
242 			else
243 				chan->chid = args.nv50.chid;
244 			return ret;
245 		}
246 	} while (*oclass);
247 
248 	nouveau_channel_del(pchan);
249 	return ret;
250 }
251 
252 static int
253 nouveau_channel_dma(struct nouveau_drm *drm, struct nvif_device *device,
254 		    u32 handle, struct nouveau_channel **pchan)
255 {
256 	static const u16 oclasses[] = { NV40_CHANNEL_DMA,
257 					NV17_CHANNEL_DMA,
258 					NV10_CHANNEL_DMA,
259 					NV03_CHANNEL_DMA,
260 					0 };
261 	const u16 *oclass = oclasses;
262 	struct nv03_channel_dma_v0 args;
263 	struct nouveau_channel *chan;
264 	int ret;
265 
266 	/* allocate dma push buffer */
267 	ret = nouveau_channel_prep(drm, device, handle, 0x10000, &chan);
268 	*pchan = chan;
269 	if (ret)
270 		return ret;
271 
272 	/* create channel object */
273 	args.version = 0;
274 	args.pushbuf = nvif_handle(&chan->push.ctxdma);
275 	args.offset = chan->push.vma.offset;
276 
277 	do {
278 		ret = nvif_object_init(&device->object, handle, *oclass++,
279 				       &args, sizeof(args), &chan->user);
280 		if (ret == 0) {
281 			chan->chid = args.chid;
282 			return ret;
283 		}
284 	} while (ret && *oclass);
285 
286 	nouveau_channel_del(pchan);
287 	return ret;
288 }
289 
290 static int
291 nouveau_channel_init(struct nouveau_channel *chan, u32 vram, u32 gart)
292 {
293 	struct nvif_device *device = chan->device;
294 	struct nouveau_cli *cli = (void *)chan->user.client;
295 	struct nvkm_mmu *mmu = nvxx_mmu(device);
296 	struct nv_dma_v0 args = {};
297 	int ret, i;
298 
299 	nvif_object_map(&chan->user);
300 
301 	/* allocate dma objects to cover all allowed vram, and gart */
302 	if (device->info.family < NV_DEVICE_INFO_V0_FERMI) {
303 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
304 			args.target = NV_DMA_V0_TARGET_VM;
305 			args.access = NV_DMA_V0_ACCESS_VM;
306 			args.start = 0;
307 			args.limit = cli->vm->mmu->limit - 1;
308 		} else {
309 			args.target = NV_DMA_V0_TARGET_VRAM;
310 			args.access = NV_DMA_V0_ACCESS_RDWR;
311 			args.start = 0;
312 			args.limit = device->info.ram_user - 1;
313 		}
314 
315 		ret = nvif_object_init(&chan->user, vram, NV_DMA_IN_MEMORY,
316 				       &args, sizeof(args), &chan->vram);
317 		if (ret)
318 			return ret;
319 
320 		if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
321 			args.target = NV_DMA_V0_TARGET_VM;
322 			args.access = NV_DMA_V0_ACCESS_VM;
323 			args.start = 0;
324 			args.limit = cli->vm->mmu->limit - 1;
325 		} else
326 		if (chan->drm->agp.stat == ENABLED) {
327 			args.target = NV_DMA_V0_TARGET_AGP;
328 			args.access = NV_DMA_V0_ACCESS_RDWR;
329 			args.start = chan->drm->agp.base;
330 			args.limit = chan->drm->agp.base +
331 				     chan->drm->agp.size - 1;
332 		} else {
333 			args.target = NV_DMA_V0_TARGET_VM;
334 			args.access = NV_DMA_V0_ACCESS_RDWR;
335 			args.start = 0;
336 			args.limit = mmu->limit - 1;
337 		}
338 
339 		ret = nvif_object_init(&chan->user, gart, NV_DMA_IN_MEMORY,
340 				       &args, sizeof(args), &chan->gart);
341 		if (ret)
342 			return ret;
343 	}
344 
345 	/* initialise dma tracking parameters */
346 	switch (chan->user.oclass & 0x00ff) {
347 	case 0x006b:
348 	case 0x006e:
349 		chan->user_put = 0x40;
350 		chan->user_get = 0x44;
351 		chan->dma.max = (0x10000 / 4) - 2;
352 		break;
353 	default:
354 		chan->user_put = 0x40;
355 		chan->user_get = 0x44;
356 		chan->user_get_hi = 0x60;
357 		chan->dma.ib_base =  0x10000 / 4;
358 		chan->dma.ib_max  = (0x02000 / 8) - 1;
359 		chan->dma.ib_put  = 0;
360 		chan->dma.ib_free = chan->dma.ib_max - chan->dma.ib_put;
361 		chan->dma.max = chan->dma.ib_base;
362 		break;
363 	}
364 
365 	chan->dma.put = 0;
366 	chan->dma.cur = chan->dma.put;
367 	chan->dma.free = chan->dma.max - chan->dma.cur;
368 
369 	ret = RING_SPACE(chan, NOUVEAU_DMA_SKIPS);
370 	if (ret)
371 		return ret;
372 
373 	for (i = 0; i < NOUVEAU_DMA_SKIPS; i++)
374 		OUT_RING(chan, 0x00000000);
375 
376 	/* allocate software object class (used for fences on <= nv05) */
377 	if (device->info.family < NV_DEVICE_INFO_V0_CELSIUS) {
378 		ret = nvif_object_init(&chan->user, 0x006e,
379 				       NVIF_IOCTL_NEW_V0_SW_NV04,
380 				       NULL, 0, &chan->nvsw);
381 		if (ret)
382 			return ret;
383 
384 		ret = RING_SPACE(chan, 2);
385 		if (ret)
386 			return ret;
387 
388 		BEGIN_NV04(chan, NvSubSw, 0x0000, 1);
389 		OUT_RING  (chan, chan->nvsw.handle);
390 		FIRE_RING (chan);
391 	}
392 
393 	/* initialise synchronisation */
394 	return nouveau_fence(chan->drm)->context_new(chan);
395 }
396 
397 int
398 nouveau_channel_new(struct nouveau_drm *drm, struct nvif_device *device,
399 		    u32 handle, u32 arg0, u32 arg1,
400 		    struct nouveau_channel **pchan)
401 {
402 	struct nouveau_cli *cli = (void *)device->object.client;
403 	bool super;
404 	int ret;
405 
406 	/* hack until fencenv50 is fixed, and agp access relaxed */
407 	super = cli->base.super;
408 	cli->base.super = true;
409 
410 	ret = nouveau_channel_ind(drm, device, handle, arg0, pchan);
411 	if (ret) {
412 		NV_PRINTK(dbg, cli, "ib channel create, %d\n", ret);
413 		ret = nouveau_channel_dma(drm, device, handle, pchan);
414 		if (ret) {
415 			NV_PRINTK(dbg, cli, "dma channel create, %d\n", ret);
416 			goto done;
417 		}
418 	}
419 
420 	ret = nouveau_channel_init(*pchan, arg0, arg1);
421 	if (ret) {
422 		NV_PRINTK(err, cli, "channel failed to initialise, %d\n", ret);
423 		nouveau_channel_del(pchan);
424 	}
425 
426 done:
427 	cli->base.super = super;
428 	return ret;
429 }
430