1 /*
2 * Copyright 2010 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 <linux/firmware.h>
26 #include <linux/module.h>
27
28 #include "drmP.h"
29
30 #include "nouveau_drv.h"
31 #include "nouveau_mm.h"
32
33 #include "nvc0_graph.h"
34 #include "nvc0_grhub.fuc.h"
35 #include "nvc0_grgpc.fuc.h"
36
37 static void
nvc0_graph_ctxctl_debug_unit(struct drm_device * dev,u32 base)38 nvc0_graph_ctxctl_debug_unit(struct drm_device *dev, u32 base)
39 {
40 NV_INFO(dev, "PGRAPH: %06x - done 0x%08x\n", base,
41 nv_rd32(dev, base + 0x400));
42 NV_INFO(dev, "PGRAPH: %06x - stat 0x%08x 0x%08x 0x%08x 0x%08x\n", base,
43 nv_rd32(dev, base + 0x800), nv_rd32(dev, base + 0x804),
44 nv_rd32(dev, base + 0x808), nv_rd32(dev, base + 0x80c));
45 NV_INFO(dev, "PGRAPH: %06x - stat 0x%08x 0x%08x 0x%08x 0x%08x\n", base,
46 nv_rd32(dev, base + 0x810), nv_rd32(dev, base + 0x814),
47 nv_rd32(dev, base + 0x818), nv_rd32(dev, base + 0x81c));
48 }
49
50 static void
nvc0_graph_ctxctl_debug(struct drm_device * dev)51 nvc0_graph_ctxctl_debug(struct drm_device *dev)
52 {
53 u32 gpcnr = nv_rd32(dev, 0x409604) & 0xffff;
54 u32 gpc;
55
56 nvc0_graph_ctxctl_debug_unit(dev, 0x409000);
57 for (gpc = 0; gpc < gpcnr; gpc++)
58 nvc0_graph_ctxctl_debug_unit(dev, 0x502000 + (gpc * 0x8000));
59 }
60
61 static int
nvc0_graph_load_context(struct nouveau_channel * chan)62 nvc0_graph_load_context(struct nouveau_channel *chan)
63 {
64 struct drm_device *dev = chan->dev;
65
66 nv_wr32(dev, 0x409840, 0x00000030);
67 nv_wr32(dev, 0x409500, 0x80000000 | chan->ramin->vinst >> 12);
68 nv_wr32(dev, 0x409504, 0x00000003);
69 if (!nv_wait(dev, 0x409800, 0x00000010, 0x00000010))
70 NV_ERROR(dev, "PGRAPH: load_ctx timeout\n");
71
72 return 0;
73 }
74
75 static int
nvc0_graph_unload_context_to(struct drm_device * dev,u64 chan)76 nvc0_graph_unload_context_to(struct drm_device *dev, u64 chan)
77 {
78 nv_wr32(dev, 0x409840, 0x00000003);
79 nv_wr32(dev, 0x409500, 0x80000000 | chan >> 12);
80 nv_wr32(dev, 0x409504, 0x00000009);
81 if (!nv_wait(dev, 0x409800, 0x00000001, 0x00000000)) {
82 NV_ERROR(dev, "PGRAPH: unload_ctx timeout\n");
83 return -EBUSY;
84 }
85
86 return 0;
87 }
88
89 static int
nvc0_graph_construct_context(struct nouveau_channel * chan)90 nvc0_graph_construct_context(struct nouveau_channel *chan)
91 {
92 struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
93 struct nvc0_graph_priv *priv = nv_engine(chan->dev, NVOBJ_ENGINE_GR);
94 struct nvc0_graph_chan *grch = chan->engctx[NVOBJ_ENGINE_GR];
95 struct drm_device *dev = chan->dev;
96 int ret, i;
97 u32 *ctx;
98
99 ctx = kmalloc(priv->grctx_size, GFP_KERNEL);
100 if (!ctx)
101 return -ENOMEM;
102
103 if (!nouveau_ctxfw) {
104 nv_wr32(dev, 0x409840, 0x80000000);
105 nv_wr32(dev, 0x409500, 0x80000000 | chan->ramin->vinst >> 12);
106 nv_wr32(dev, 0x409504, 0x00000001);
107 if (!nv_wait(dev, 0x409800, 0x80000000, 0x80000000)) {
108 NV_ERROR(dev, "PGRAPH: HUB_SET_CHAN timeout\n");
109 nvc0_graph_ctxctl_debug(dev);
110 ret = -EBUSY;
111 goto err;
112 }
113 } else {
114 nvc0_graph_load_context(chan);
115
116 nv_wo32(grch->grctx, 0x1c, 1);
117 nv_wo32(grch->grctx, 0x20, 0);
118 nv_wo32(grch->grctx, 0x28, 0);
119 nv_wo32(grch->grctx, 0x2c, 0);
120 dev_priv->engine.instmem.flush(dev);
121 }
122
123 ret = nvc0_grctx_generate(chan);
124 if (ret)
125 goto err;
126
127 if (!nouveau_ctxfw) {
128 nv_wr32(dev, 0x409840, 0x80000000);
129 nv_wr32(dev, 0x409500, 0x80000000 | chan->ramin->vinst >> 12);
130 nv_wr32(dev, 0x409504, 0x00000002);
131 if (!nv_wait(dev, 0x409800, 0x80000000, 0x80000000)) {
132 NV_ERROR(dev, "PGRAPH: HUB_CTX_SAVE timeout\n");
133 nvc0_graph_ctxctl_debug(dev);
134 ret = -EBUSY;
135 goto err;
136 }
137 } else {
138 ret = nvc0_graph_unload_context_to(dev, chan->ramin->vinst);
139 if (ret)
140 goto err;
141 }
142
143 for (i = 0; i < priv->grctx_size; i += 4)
144 ctx[i / 4] = nv_ro32(grch->grctx, i);
145
146 priv->grctx_vals = ctx;
147 return 0;
148
149 err:
150 kfree(ctx);
151 return ret;
152 }
153
154 static int
nvc0_graph_create_context_mmio_list(struct nouveau_channel * chan)155 nvc0_graph_create_context_mmio_list(struct nouveau_channel *chan)
156 {
157 struct nvc0_graph_priv *priv = nv_engine(chan->dev, NVOBJ_ENGINE_GR);
158 struct nvc0_graph_chan *grch = chan->engctx[NVOBJ_ENGINE_GR];
159 struct drm_device *dev = chan->dev;
160 struct drm_nouveau_private *dev_priv = dev->dev_private;
161 int i = 0, gpc, tp, ret;
162
163 ret = nouveau_gpuobj_new(dev, chan, 0x2000, 256, NVOBJ_FLAG_VM,
164 &grch->unk408004);
165 if (ret)
166 return ret;
167
168 ret = nouveau_gpuobj_new(dev, chan, 0x8000, 256, NVOBJ_FLAG_VM,
169 &grch->unk40800c);
170 if (ret)
171 return ret;
172
173 ret = nouveau_gpuobj_new(dev, chan, 384 * 1024, 4096,
174 NVOBJ_FLAG_VM | NVOBJ_FLAG_VM_USER,
175 &grch->unk418810);
176 if (ret)
177 return ret;
178
179 ret = nouveau_gpuobj_new(dev, chan, 0x1000, 0, NVOBJ_FLAG_VM,
180 &grch->mmio);
181 if (ret)
182 return ret;
183
184
185 nv_wo32(grch->mmio, i++ * 4, 0x00408004);
186 nv_wo32(grch->mmio, i++ * 4, grch->unk408004->linst >> 8);
187 nv_wo32(grch->mmio, i++ * 4, 0x00408008);
188 nv_wo32(grch->mmio, i++ * 4, 0x80000018);
189
190 nv_wo32(grch->mmio, i++ * 4, 0x0040800c);
191 nv_wo32(grch->mmio, i++ * 4, grch->unk40800c->linst >> 8);
192 nv_wo32(grch->mmio, i++ * 4, 0x00408010);
193 nv_wo32(grch->mmio, i++ * 4, 0x80000000);
194
195 nv_wo32(grch->mmio, i++ * 4, 0x00418810);
196 nv_wo32(grch->mmio, i++ * 4, 0x80000000 | grch->unk418810->linst >> 12);
197 nv_wo32(grch->mmio, i++ * 4, 0x00419848);
198 nv_wo32(grch->mmio, i++ * 4, 0x10000000 | grch->unk418810->linst >> 12);
199
200 nv_wo32(grch->mmio, i++ * 4, 0x00419004);
201 nv_wo32(grch->mmio, i++ * 4, grch->unk40800c->linst >> 8);
202 nv_wo32(grch->mmio, i++ * 4, 0x00419008);
203 nv_wo32(grch->mmio, i++ * 4, 0x00000000);
204
205 nv_wo32(grch->mmio, i++ * 4, 0x00418808);
206 nv_wo32(grch->mmio, i++ * 4, grch->unk408004->linst >> 8);
207 nv_wo32(grch->mmio, i++ * 4, 0x0041880c);
208 nv_wo32(grch->mmio, i++ * 4, 0x80000018);
209
210 if (dev_priv->chipset != 0xc1) {
211 u32 magic = 0x02180000;
212 nv_wo32(grch->mmio, i++ * 4, 0x00405830);
213 nv_wo32(grch->mmio, i++ * 4, magic);
214 for (gpc = 0; gpc < priv->gpc_nr; gpc++) {
215 for (tp = 0; tp < priv->tp_nr[gpc]; tp++) {
216 u32 reg = TP_UNIT(gpc, tp, 0x520);
217 nv_wo32(grch->mmio, i++ * 4, reg);
218 nv_wo32(grch->mmio, i++ * 4, magic);
219 magic += 0x0324;
220 }
221 }
222 } else {
223 u32 magic = 0x02180000;
224 nv_wo32(grch->mmio, i++ * 4, 0x00405830);
225 nv_wo32(grch->mmio, i++ * 4, magic | 0x0000218);
226 nv_wo32(grch->mmio, i++ * 4, 0x004064c4);
227 nv_wo32(grch->mmio, i++ * 4, 0x0086ffff);
228 for (gpc = 0; gpc < priv->gpc_nr; gpc++) {
229 for (tp = 0; tp < priv->tp_nr[gpc]; tp++) {
230 u32 reg = TP_UNIT(gpc, tp, 0x520);
231 nv_wo32(grch->mmio, i++ * 4, reg);
232 nv_wo32(grch->mmio, i++ * 4, (1 << 28) | magic);
233 magic += 0x0324;
234 }
235 for (tp = 0; tp < priv->tp_nr[gpc]; tp++) {
236 u32 reg = TP_UNIT(gpc, tp, 0x544);
237 nv_wo32(grch->mmio, i++ * 4, reg);
238 nv_wo32(grch->mmio, i++ * 4, magic);
239 magic += 0x0324;
240 }
241 }
242 }
243
244 grch->mmio_nr = i / 2;
245 return 0;
246 }
247
248 static int
nvc0_graph_context_new(struct nouveau_channel * chan,int engine)249 nvc0_graph_context_new(struct nouveau_channel *chan, int engine)
250 {
251 struct drm_device *dev = chan->dev;
252 struct drm_nouveau_private *dev_priv = dev->dev_private;
253 struct nouveau_instmem_engine *pinstmem = &dev_priv->engine.instmem;
254 struct nvc0_graph_priv *priv = nv_engine(dev, engine);
255 struct nvc0_graph_chan *grch;
256 struct nouveau_gpuobj *grctx;
257 int ret, i;
258
259 grch = kzalloc(sizeof(*grch), GFP_KERNEL);
260 if (!grch)
261 return -ENOMEM;
262 chan->engctx[NVOBJ_ENGINE_GR] = grch;
263
264 ret = nouveau_gpuobj_new(dev, chan, priv->grctx_size, 256,
265 NVOBJ_FLAG_VM | NVOBJ_FLAG_ZERO_ALLOC,
266 &grch->grctx);
267 if (ret)
268 goto error;
269 grctx = grch->grctx;
270
271 ret = nvc0_graph_create_context_mmio_list(chan);
272 if (ret)
273 goto error;
274
275 nv_wo32(chan->ramin, 0x0210, lower_32_bits(grctx->linst) | 4);
276 nv_wo32(chan->ramin, 0x0214, upper_32_bits(grctx->linst));
277 pinstmem->flush(dev);
278
279 if (!priv->grctx_vals) {
280 ret = nvc0_graph_construct_context(chan);
281 if (ret)
282 goto error;
283 }
284
285 for (i = 0; i < priv->grctx_size; i += 4)
286 nv_wo32(grctx, i, priv->grctx_vals[i / 4]);
287
288 if (!nouveau_ctxfw) {
289 nv_wo32(grctx, 0x00, grch->mmio_nr);
290 nv_wo32(grctx, 0x04, grch->mmio->linst >> 8);
291 } else {
292 nv_wo32(grctx, 0xf4, 0);
293 nv_wo32(grctx, 0xf8, 0);
294 nv_wo32(grctx, 0x10, grch->mmio_nr);
295 nv_wo32(grctx, 0x14, lower_32_bits(grch->mmio->linst));
296 nv_wo32(grctx, 0x18, upper_32_bits(grch->mmio->linst));
297 nv_wo32(grctx, 0x1c, 1);
298 nv_wo32(grctx, 0x20, 0);
299 nv_wo32(grctx, 0x28, 0);
300 nv_wo32(grctx, 0x2c, 0);
301 }
302 pinstmem->flush(dev);
303 return 0;
304
305 error:
306 priv->base.context_del(chan, engine);
307 return ret;
308 }
309
310 static void
nvc0_graph_context_del(struct nouveau_channel * chan,int engine)311 nvc0_graph_context_del(struct nouveau_channel *chan, int engine)
312 {
313 struct nvc0_graph_chan *grch = chan->engctx[engine];
314
315 nouveau_gpuobj_ref(NULL, &grch->mmio);
316 nouveau_gpuobj_ref(NULL, &grch->unk418810);
317 nouveau_gpuobj_ref(NULL, &grch->unk40800c);
318 nouveau_gpuobj_ref(NULL, &grch->unk408004);
319 nouveau_gpuobj_ref(NULL, &grch->grctx);
320 chan->engctx[engine] = NULL;
321 }
322
323 static int
nvc0_graph_object_new(struct nouveau_channel * chan,int engine,u32 handle,u16 class)324 nvc0_graph_object_new(struct nouveau_channel *chan, int engine,
325 u32 handle, u16 class)
326 {
327 return 0;
328 }
329
330 static int
nvc0_graph_fini(struct drm_device * dev,int engine,bool suspend)331 nvc0_graph_fini(struct drm_device *dev, int engine, bool suspend)
332 {
333 return 0;
334 }
335
336 static int
nvc0_graph_mthd_page_flip(struct nouveau_channel * chan,u32 class,u32 mthd,u32 data)337 nvc0_graph_mthd_page_flip(struct nouveau_channel *chan,
338 u32 class, u32 mthd, u32 data)
339 {
340 nouveau_finish_page_flip(chan, NULL);
341 return 0;
342 }
343
344 static void
nvc0_graph_init_obj418880(struct drm_device * dev)345 nvc0_graph_init_obj418880(struct drm_device *dev)
346 {
347 struct nvc0_graph_priv *priv = nv_engine(dev, NVOBJ_ENGINE_GR);
348 int i;
349
350 nv_wr32(dev, GPC_BCAST(0x0880), 0x00000000);
351 nv_wr32(dev, GPC_BCAST(0x08a4), 0x00000000);
352 for (i = 0; i < 4; i++)
353 nv_wr32(dev, GPC_BCAST(0x0888) + (i * 4), 0x00000000);
354 nv_wr32(dev, GPC_BCAST(0x08b4), priv->unk4188b4->vinst >> 8);
355 nv_wr32(dev, GPC_BCAST(0x08b8), priv->unk4188b8->vinst >> 8);
356 }
357
358 static void
nvc0_graph_init_regs(struct drm_device * dev)359 nvc0_graph_init_regs(struct drm_device *dev)
360 {
361 nv_wr32(dev, 0x400080, 0x003083c2);
362 nv_wr32(dev, 0x400088, 0x00006fe7);
363 nv_wr32(dev, 0x40008c, 0x00000000);
364 nv_wr32(dev, 0x400090, 0x00000030);
365 nv_wr32(dev, 0x40013c, 0x013901f7);
366 nv_wr32(dev, 0x400140, 0x00000100);
367 nv_wr32(dev, 0x400144, 0x00000000);
368 nv_wr32(dev, 0x400148, 0x00000110);
369 nv_wr32(dev, 0x400138, 0x00000000);
370 nv_wr32(dev, 0x400130, 0x00000000);
371 nv_wr32(dev, 0x400134, 0x00000000);
372 nv_wr32(dev, 0x400124, 0x00000002);
373 }
374
375 static void
nvc0_graph_init_gpc_0(struct drm_device * dev)376 nvc0_graph_init_gpc_0(struct drm_device *dev)
377 {
378 struct nvc0_graph_priv *priv = nv_engine(dev, NVOBJ_ENGINE_GR);
379 const u32 magicgpc918 = DIV_ROUND_UP(0x00800000, priv->tp_total);
380 u32 data[TP_MAX / 8];
381 u8 tpnr[GPC_MAX];
382 int i, gpc, tpc;
383
384 nv_wr32(dev, TP_UNIT(0, 0, 0x5c), 1); /* affects TFB offset queries */
385
386 /*
387 * TP ROP UNKVAL(magic_not_rop_nr)
388 * 450: 4/0/0/0 2 3
389 * 460: 3/4/0/0 4 1
390 * 465: 3/4/4/0 4 7
391 * 470: 3/3/4/4 5 5
392 * 480: 3/4/4/4 6 6
393 */
394
395 memset(data, 0x00, sizeof(data));
396 memcpy(tpnr, priv->tp_nr, sizeof(priv->tp_nr));
397 for (i = 0, gpc = -1; i < priv->tp_total; i++) {
398 do {
399 gpc = (gpc + 1) % priv->gpc_nr;
400 } while (!tpnr[gpc]);
401 tpc = priv->tp_nr[gpc] - tpnr[gpc]--;
402
403 data[i / 8] |= tpc << ((i % 8) * 4);
404 }
405
406 nv_wr32(dev, GPC_BCAST(0x0980), data[0]);
407 nv_wr32(dev, GPC_BCAST(0x0984), data[1]);
408 nv_wr32(dev, GPC_BCAST(0x0988), data[2]);
409 nv_wr32(dev, GPC_BCAST(0x098c), data[3]);
410
411 for (gpc = 0; gpc < priv->gpc_nr; gpc++) {
412 nv_wr32(dev, GPC_UNIT(gpc, 0x0914), priv->magic_not_rop_nr << 8 |
413 priv->tp_nr[gpc]);
414 nv_wr32(dev, GPC_UNIT(gpc, 0x0910), 0x00040000 | priv->tp_total);
415 nv_wr32(dev, GPC_UNIT(gpc, 0x0918), magicgpc918);
416 }
417
418 nv_wr32(dev, GPC_BCAST(0x1bd4), magicgpc918);
419 nv_wr32(dev, GPC_BCAST(0x08ac), nv_rd32(dev, 0x100800));
420 }
421
422 static void
nvc0_graph_init_units(struct drm_device * dev)423 nvc0_graph_init_units(struct drm_device *dev)
424 {
425 nv_wr32(dev, 0x409c24, 0x000f0000);
426 nv_wr32(dev, 0x404000, 0xc0000000); /* DISPATCH */
427 nv_wr32(dev, 0x404600, 0xc0000000); /* M2MF */
428 nv_wr32(dev, 0x408030, 0xc0000000);
429 nv_wr32(dev, 0x40601c, 0xc0000000);
430 nv_wr32(dev, 0x404490, 0xc0000000); /* MACRO */
431 nv_wr32(dev, 0x406018, 0xc0000000);
432 nv_wr32(dev, 0x405840, 0xc0000000);
433 nv_wr32(dev, 0x405844, 0x00ffffff);
434 nv_mask(dev, 0x419cc0, 0x00000008, 0x00000008);
435 nv_mask(dev, 0x419eb4, 0x00001000, 0x00001000);
436 }
437
438 static void
nvc0_graph_init_gpc_1(struct drm_device * dev)439 nvc0_graph_init_gpc_1(struct drm_device *dev)
440 {
441 struct nvc0_graph_priv *priv = nv_engine(dev, NVOBJ_ENGINE_GR);
442 int gpc, tp;
443
444 for (gpc = 0; gpc < priv->gpc_nr; gpc++) {
445 nv_wr32(dev, GPC_UNIT(gpc, 0x0420), 0xc0000000);
446 nv_wr32(dev, GPC_UNIT(gpc, 0x0900), 0xc0000000);
447 nv_wr32(dev, GPC_UNIT(gpc, 0x1028), 0xc0000000);
448 nv_wr32(dev, GPC_UNIT(gpc, 0x0824), 0xc0000000);
449 for (tp = 0; tp < priv->tp_nr[gpc]; tp++) {
450 nv_wr32(dev, TP_UNIT(gpc, tp, 0x508), 0xffffffff);
451 nv_wr32(dev, TP_UNIT(gpc, tp, 0x50c), 0xffffffff);
452 nv_wr32(dev, TP_UNIT(gpc, tp, 0x224), 0xc0000000);
453 nv_wr32(dev, TP_UNIT(gpc, tp, 0x48c), 0xc0000000);
454 nv_wr32(dev, TP_UNIT(gpc, tp, 0x084), 0xc0000000);
455 nv_wr32(dev, TP_UNIT(gpc, tp, 0x644), 0x001ffffe);
456 nv_wr32(dev, TP_UNIT(gpc, tp, 0x64c), 0x0000000f);
457 }
458 nv_wr32(dev, GPC_UNIT(gpc, 0x2c90), 0xffffffff);
459 nv_wr32(dev, GPC_UNIT(gpc, 0x2c94), 0xffffffff);
460 }
461 }
462
463 static void
nvc0_graph_init_rop(struct drm_device * dev)464 nvc0_graph_init_rop(struct drm_device *dev)
465 {
466 struct nvc0_graph_priv *priv = nv_engine(dev, NVOBJ_ENGINE_GR);
467 int rop;
468
469 for (rop = 0; rop < priv->rop_nr; rop++) {
470 nv_wr32(dev, ROP_UNIT(rop, 0x144), 0xc0000000);
471 nv_wr32(dev, ROP_UNIT(rop, 0x070), 0xc0000000);
472 nv_wr32(dev, ROP_UNIT(rop, 0x204), 0xffffffff);
473 nv_wr32(dev, ROP_UNIT(rop, 0x208), 0xffffffff);
474 }
475 }
476
477 static void
nvc0_graph_init_fuc(struct drm_device * dev,u32 fuc_base,struct nvc0_graph_fuc * code,struct nvc0_graph_fuc * data)478 nvc0_graph_init_fuc(struct drm_device *dev, u32 fuc_base,
479 struct nvc0_graph_fuc *code, struct nvc0_graph_fuc *data)
480 {
481 int i;
482
483 nv_wr32(dev, fuc_base + 0x01c0, 0x01000000);
484 for (i = 0; i < data->size / 4; i++)
485 nv_wr32(dev, fuc_base + 0x01c4, data->data[i]);
486
487 nv_wr32(dev, fuc_base + 0x0180, 0x01000000);
488 for (i = 0; i < code->size / 4; i++) {
489 if ((i & 0x3f) == 0)
490 nv_wr32(dev, fuc_base + 0x0188, i >> 6);
491 nv_wr32(dev, fuc_base + 0x0184, code->data[i]);
492 }
493 }
494
495 static int
nvc0_graph_init_ctxctl(struct drm_device * dev)496 nvc0_graph_init_ctxctl(struct drm_device *dev)
497 {
498 struct drm_nouveau_private *dev_priv = dev->dev_private;
499 struct nvc0_graph_priv *priv = nv_engine(dev, NVOBJ_ENGINE_GR);
500 u32 r000260;
501 int i;
502
503 if (!nouveau_ctxfw) {
504 /* load HUB microcode */
505 r000260 = nv_mask(dev, 0x000260, 0x00000001, 0x00000000);
506 nv_wr32(dev, 0x4091c0, 0x01000000);
507 for (i = 0; i < sizeof(nvc0_grhub_data) / 4; i++)
508 nv_wr32(dev, 0x4091c4, nvc0_grhub_data[i]);
509
510 nv_wr32(dev, 0x409180, 0x01000000);
511 for (i = 0; i < sizeof(nvc0_grhub_code) / 4; i++) {
512 if ((i & 0x3f) == 0)
513 nv_wr32(dev, 0x409188, i >> 6);
514 nv_wr32(dev, 0x409184, nvc0_grhub_code[i]);
515 }
516
517 /* load GPC microcode */
518 nv_wr32(dev, 0x41a1c0, 0x01000000);
519 for (i = 0; i < sizeof(nvc0_grgpc_data) / 4; i++)
520 nv_wr32(dev, 0x41a1c4, nvc0_grgpc_data[i]);
521
522 nv_wr32(dev, 0x41a180, 0x01000000);
523 for (i = 0; i < sizeof(nvc0_grgpc_code) / 4; i++) {
524 if ((i & 0x3f) == 0)
525 nv_wr32(dev, 0x41a188, i >> 6);
526 nv_wr32(dev, 0x41a184, nvc0_grgpc_code[i]);
527 }
528 nv_wr32(dev, 0x000260, r000260);
529
530 /* start HUB ucode running, it'll init the GPCs */
531 nv_wr32(dev, 0x409800, dev_priv->chipset);
532 nv_wr32(dev, 0x40910c, 0x00000000);
533 nv_wr32(dev, 0x409100, 0x00000002);
534 if (!nv_wait(dev, 0x409800, 0x80000000, 0x80000000)) {
535 NV_ERROR(dev, "PGRAPH: HUB_INIT timed out\n");
536 nvc0_graph_ctxctl_debug(dev);
537 return -EBUSY;
538 }
539
540 priv->grctx_size = nv_rd32(dev, 0x409804);
541 return 0;
542 }
543
544 /* load fuc microcode */
545 r000260 = nv_mask(dev, 0x000260, 0x00000001, 0x00000000);
546 nvc0_graph_init_fuc(dev, 0x409000, &priv->fuc409c, &priv->fuc409d);
547 nvc0_graph_init_fuc(dev, 0x41a000, &priv->fuc41ac, &priv->fuc41ad);
548 nv_wr32(dev, 0x000260, r000260);
549
550 /* start both of them running */
551 nv_wr32(dev, 0x409840, 0xffffffff);
552 nv_wr32(dev, 0x41a10c, 0x00000000);
553 nv_wr32(dev, 0x40910c, 0x00000000);
554 nv_wr32(dev, 0x41a100, 0x00000002);
555 nv_wr32(dev, 0x409100, 0x00000002);
556 if (!nv_wait(dev, 0x409800, 0x00000001, 0x00000001))
557 NV_INFO(dev, "0x409800 wait failed\n");
558
559 nv_wr32(dev, 0x409840, 0xffffffff);
560 nv_wr32(dev, 0x409500, 0x7fffffff);
561 nv_wr32(dev, 0x409504, 0x00000021);
562
563 nv_wr32(dev, 0x409840, 0xffffffff);
564 nv_wr32(dev, 0x409500, 0x00000000);
565 nv_wr32(dev, 0x409504, 0x00000010);
566 if (!nv_wait_ne(dev, 0x409800, 0xffffffff, 0x00000000)) {
567 NV_ERROR(dev, "fuc09 req 0x10 timeout\n");
568 return -EBUSY;
569 }
570 priv->grctx_size = nv_rd32(dev, 0x409800);
571
572 nv_wr32(dev, 0x409840, 0xffffffff);
573 nv_wr32(dev, 0x409500, 0x00000000);
574 nv_wr32(dev, 0x409504, 0x00000016);
575 if (!nv_wait_ne(dev, 0x409800, 0xffffffff, 0x00000000)) {
576 NV_ERROR(dev, "fuc09 req 0x16 timeout\n");
577 return -EBUSY;
578 }
579
580 nv_wr32(dev, 0x409840, 0xffffffff);
581 nv_wr32(dev, 0x409500, 0x00000000);
582 nv_wr32(dev, 0x409504, 0x00000025);
583 if (!nv_wait_ne(dev, 0x409800, 0xffffffff, 0x00000000)) {
584 NV_ERROR(dev, "fuc09 req 0x25 timeout\n");
585 return -EBUSY;
586 }
587
588 return 0;
589 }
590
591 static int
nvc0_graph_init(struct drm_device * dev,int engine)592 nvc0_graph_init(struct drm_device *dev, int engine)
593 {
594 int ret;
595
596 nv_mask(dev, 0x000200, 0x18001000, 0x00000000);
597 nv_mask(dev, 0x000200, 0x18001000, 0x18001000);
598
599 nvc0_graph_init_obj418880(dev);
600 nvc0_graph_init_regs(dev);
601 /*nvc0_graph_init_unitplemented_magics(dev);*/
602 nvc0_graph_init_gpc_0(dev);
603 /*nvc0_graph_init_unitplemented_c242(dev);*/
604
605 nv_wr32(dev, 0x400500, 0x00010001);
606 nv_wr32(dev, 0x400100, 0xffffffff);
607 nv_wr32(dev, 0x40013c, 0xffffffff);
608
609 nvc0_graph_init_units(dev);
610 nvc0_graph_init_gpc_1(dev);
611 nvc0_graph_init_rop(dev);
612
613 nv_wr32(dev, 0x400108, 0xffffffff);
614 nv_wr32(dev, 0x400138, 0xffffffff);
615 nv_wr32(dev, 0x400118, 0xffffffff);
616 nv_wr32(dev, 0x400130, 0xffffffff);
617 nv_wr32(dev, 0x40011c, 0xffffffff);
618 nv_wr32(dev, 0x400134, 0xffffffff);
619 nv_wr32(dev, 0x400054, 0x34ce3464);
620
621 ret = nvc0_graph_init_ctxctl(dev);
622 if (ret)
623 return ret;
624
625 return 0;
626 }
627
628 int
nvc0_graph_isr_chid(struct drm_device * dev,u64 inst)629 nvc0_graph_isr_chid(struct drm_device *dev, u64 inst)
630 {
631 struct drm_nouveau_private *dev_priv = dev->dev_private;
632 struct nouveau_channel *chan;
633 unsigned long flags;
634 int i;
635
636 spin_lock_irqsave(&dev_priv->channels.lock, flags);
637 for (i = 0; i < dev_priv->engine.fifo.channels; i++) {
638 chan = dev_priv->channels.ptr[i];
639 if (!chan || !chan->ramin)
640 continue;
641
642 if (inst == chan->ramin->vinst)
643 break;
644 }
645 spin_unlock_irqrestore(&dev_priv->channels.lock, flags);
646 return i;
647 }
648
649 static void
nvc0_graph_ctxctl_isr(struct drm_device * dev)650 nvc0_graph_ctxctl_isr(struct drm_device *dev)
651 {
652 u32 ustat = nv_rd32(dev, 0x409c18);
653
654 if (ustat & 0x00000001)
655 NV_INFO(dev, "PGRAPH: CTXCTRL ucode error\n");
656 if (ustat & 0x00080000)
657 NV_INFO(dev, "PGRAPH: CTXCTRL watchdog timeout\n");
658 if (ustat & ~0x00080001)
659 NV_INFO(dev, "PGRAPH: CTXCTRL 0x%08x\n", ustat);
660
661 nvc0_graph_ctxctl_debug(dev);
662 nv_wr32(dev, 0x409c20, ustat);
663 }
664
665 static void
nvc0_graph_isr(struct drm_device * dev)666 nvc0_graph_isr(struct drm_device *dev)
667 {
668 u64 inst = (u64)(nv_rd32(dev, 0x409b00) & 0x0fffffff) << 12;
669 u32 chid = nvc0_graph_isr_chid(dev, inst);
670 u32 stat = nv_rd32(dev, 0x400100);
671 u32 addr = nv_rd32(dev, 0x400704);
672 u32 mthd = (addr & 0x00003ffc);
673 u32 subc = (addr & 0x00070000) >> 16;
674 u32 data = nv_rd32(dev, 0x400708);
675 u32 code = nv_rd32(dev, 0x400110);
676 u32 class = nv_rd32(dev, 0x404200 + (subc * 4));
677
678 if (stat & 0x00000010) {
679 if (nouveau_gpuobj_mthd_call2(dev, chid, class, mthd, data)) {
680 NV_INFO(dev, "PGRAPH: ILLEGAL_MTHD ch %d [0x%010llx] "
681 "subc %d class 0x%04x mthd 0x%04x "
682 "data 0x%08x\n",
683 chid, inst, subc, class, mthd, data);
684 }
685 nv_wr32(dev, 0x400100, 0x00000010);
686 stat &= ~0x00000010;
687 }
688
689 if (stat & 0x00000020) {
690 NV_INFO(dev, "PGRAPH: ILLEGAL_CLASS ch %d [0x%010llx] subc %d "
691 "class 0x%04x mthd 0x%04x data 0x%08x\n",
692 chid, inst, subc, class, mthd, data);
693 nv_wr32(dev, 0x400100, 0x00000020);
694 stat &= ~0x00000020;
695 }
696
697 if (stat & 0x00100000) {
698 NV_INFO(dev, "PGRAPH: DATA_ERROR [");
699 nouveau_enum_print(nv50_data_error_names, code);
700 printk("] ch %d [0x%010llx] subc %d class 0x%04x "
701 "mthd 0x%04x data 0x%08x\n",
702 chid, inst, subc, class, mthd, data);
703 nv_wr32(dev, 0x400100, 0x00100000);
704 stat &= ~0x00100000;
705 }
706
707 if (stat & 0x00200000) {
708 u32 trap = nv_rd32(dev, 0x400108);
709 NV_INFO(dev, "PGRAPH: TRAP ch %d status 0x%08x\n", chid, trap);
710 nv_wr32(dev, 0x400108, trap);
711 nv_wr32(dev, 0x400100, 0x00200000);
712 stat &= ~0x00200000;
713 }
714
715 if (stat & 0x00080000) {
716 nvc0_graph_ctxctl_isr(dev);
717 nv_wr32(dev, 0x400100, 0x00080000);
718 stat &= ~0x00080000;
719 }
720
721 if (stat) {
722 NV_INFO(dev, "PGRAPH: unknown stat 0x%08x\n", stat);
723 nv_wr32(dev, 0x400100, stat);
724 }
725
726 nv_wr32(dev, 0x400500, 0x00010001);
727 }
728
729 static int
nvc0_graph_create_fw(struct drm_device * dev,const char * fwname,struct nvc0_graph_fuc * fuc)730 nvc0_graph_create_fw(struct drm_device *dev, const char *fwname,
731 struct nvc0_graph_fuc *fuc)
732 {
733 struct drm_nouveau_private *dev_priv = dev->dev_private;
734 const struct firmware *fw;
735 char f[32];
736 int ret;
737
738 snprintf(f, sizeof(f), "nouveau/nv%02x_%s", dev_priv->chipset, fwname);
739 ret = request_firmware(&fw, f, &dev->pdev->dev);
740 if (ret) {
741 snprintf(f, sizeof(f), "nouveau/%s", fwname);
742 ret = request_firmware(&fw, f, &dev->pdev->dev);
743 if (ret) {
744 NV_ERROR(dev, "failed to load %s\n", fwname);
745 return ret;
746 }
747 }
748
749 fuc->size = fw->size;
750 fuc->data = kmemdup(fw->data, fuc->size, GFP_KERNEL);
751 release_firmware(fw);
752 return (fuc->data != NULL) ? 0 : -ENOMEM;
753 }
754
755 static void
nvc0_graph_destroy_fw(struct nvc0_graph_fuc * fuc)756 nvc0_graph_destroy_fw(struct nvc0_graph_fuc *fuc)
757 {
758 if (fuc->data) {
759 kfree(fuc->data);
760 fuc->data = NULL;
761 }
762 }
763
764 static void
nvc0_graph_destroy(struct drm_device * dev,int engine)765 nvc0_graph_destroy(struct drm_device *dev, int engine)
766 {
767 struct nvc0_graph_priv *priv = nv_engine(dev, engine);
768
769 if (nouveau_ctxfw) {
770 nvc0_graph_destroy_fw(&priv->fuc409c);
771 nvc0_graph_destroy_fw(&priv->fuc409d);
772 nvc0_graph_destroy_fw(&priv->fuc41ac);
773 nvc0_graph_destroy_fw(&priv->fuc41ad);
774 }
775
776 nouveau_irq_unregister(dev, 12);
777
778 nouveau_gpuobj_ref(NULL, &priv->unk4188b8);
779 nouveau_gpuobj_ref(NULL, &priv->unk4188b4);
780
781 if (priv->grctx_vals)
782 kfree(priv->grctx_vals);
783
784 NVOBJ_ENGINE_DEL(dev, GR);
785 kfree(priv);
786 }
787
788 int
nvc0_graph_create(struct drm_device * dev)789 nvc0_graph_create(struct drm_device *dev)
790 {
791 struct drm_nouveau_private *dev_priv = dev->dev_private;
792 struct nvc0_graph_priv *priv;
793 int ret, gpc, i;
794 u32 fermi;
795
796 fermi = nvc0_graph_class(dev);
797 if (!fermi) {
798 NV_ERROR(dev, "PGRAPH: unsupported chipset, please report!\n");
799 return 0;
800 }
801
802 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
803 if (!priv)
804 return -ENOMEM;
805
806 priv->base.destroy = nvc0_graph_destroy;
807 priv->base.init = nvc0_graph_init;
808 priv->base.fini = nvc0_graph_fini;
809 priv->base.context_new = nvc0_graph_context_new;
810 priv->base.context_del = nvc0_graph_context_del;
811 priv->base.object_new = nvc0_graph_object_new;
812
813 NVOBJ_ENGINE_ADD(dev, GR, &priv->base);
814 nouveau_irq_register(dev, 12, nvc0_graph_isr);
815
816 if (nouveau_ctxfw) {
817 NV_INFO(dev, "PGRAPH: using external firmware\n");
818 if (nvc0_graph_create_fw(dev, "fuc409c", &priv->fuc409c) ||
819 nvc0_graph_create_fw(dev, "fuc409d", &priv->fuc409d) ||
820 nvc0_graph_create_fw(dev, "fuc41ac", &priv->fuc41ac) ||
821 nvc0_graph_create_fw(dev, "fuc41ad", &priv->fuc41ad)) {
822 ret = 0;
823 goto error;
824 }
825 }
826
827 ret = nouveau_gpuobj_new(dev, NULL, 0x1000, 256, 0, &priv->unk4188b4);
828 if (ret)
829 goto error;
830
831 ret = nouveau_gpuobj_new(dev, NULL, 0x1000, 256, 0, &priv->unk4188b8);
832 if (ret)
833 goto error;
834
835 for (i = 0; i < 0x1000; i += 4) {
836 nv_wo32(priv->unk4188b4, i, 0x00000010);
837 nv_wo32(priv->unk4188b8, i, 0x00000010);
838 }
839
840 priv->gpc_nr = nv_rd32(dev, 0x409604) & 0x0000001f;
841 priv->rop_nr = (nv_rd32(dev, 0x409604) & 0x001f0000) >> 16;
842 for (gpc = 0; gpc < priv->gpc_nr; gpc++) {
843 priv->tp_nr[gpc] = nv_rd32(dev, GPC_UNIT(gpc, 0x2608));
844 priv->tp_total += priv->tp_nr[gpc];
845 }
846
847 /*XXX: these need figuring out... */
848 switch (dev_priv->chipset) {
849 case 0xc0:
850 if (priv->tp_total == 11) { /* 465, 3/4/4/0, 4 */
851 priv->magic_not_rop_nr = 0x07;
852 } else
853 if (priv->tp_total == 14) { /* 470, 3/3/4/4, 5 */
854 priv->magic_not_rop_nr = 0x05;
855 } else
856 if (priv->tp_total == 15) { /* 480, 3/4/4/4, 6 */
857 priv->magic_not_rop_nr = 0x06;
858 }
859 break;
860 case 0xc3: /* 450, 4/0/0/0, 2 */
861 priv->magic_not_rop_nr = 0x03;
862 break;
863 case 0xc4: /* 460, 3/4/0/0, 4 */
864 priv->magic_not_rop_nr = 0x01;
865 break;
866 case 0xc1: /* 2/0/0/0, 1 */
867 priv->magic_not_rop_nr = 0x01;
868 break;
869 case 0xc8: /* 4/4/3/4, 5 */
870 priv->magic_not_rop_nr = 0x06;
871 break;
872 case 0xce: /* 4/4/0/0, 4 */
873 priv->magic_not_rop_nr = 0x03;
874 break;
875 case 0xcf: /* 4/0/0/0, 3 */
876 priv->magic_not_rop_nr = 0x03;
877 break;
878 case 0xd9: /* 1/0/0/0, 1 */
879 priv->magic_not_rop_nr = 0x01;
880 break;
881 }
882
883 if (!priv->magic_not_rop_nr) {
884 NV_ERROR(dev, "PGRAPH: unknown config: %d/%d/%d/%d, %d\n",
885 priv->tp_nr[0], priv->tp_nr[1], priv->tp_nr[2],
886 priv->tp_nr[3], priv->rop_nr);
887 priv->magic_not_rop_nr = 0x00;
888 }
889
890 NVOBJ_CLASS(dev, 0x902d, GR); /* 2D */
891 NVOBJ_CLASS(dev, 0x9039, GR); /* M2MF */
892 NVOBJ_MTHD (dev, 0x9039, 0x0500, nvc0_graph_mthd_page_flip);
893 NVOBJ_CLASS(dev, 0x9097, GR); /* 3D */
894 if (fermi >= 0x9197)
895 NVOBJ_CLASS(dev, 0x9197, GR); /* 3D (NVC1-) */
896 if (fermi >= 0x9297)
897 NVOBJ_CLASS(dev, 0x9297, GR); /* 3D (NVC8-) */
898 NVOBJ_CLASS(dev, 0x90c0, GR); /* COMPUTE */
899 return 0;
900
901 error:
902 nvc0_graph_destroy(dev, NVOBJ_ENGINE_GR);
903 return ret;
904 }
905