?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
drm/radeon: remove r600_blit_suspend
[linux-2.6.git] / drivers / gpu / drm / radeon / ni.c
blob4004376362eec07d7c5646ca80157db156f94c36
1 /*
2 * Copyright 2010 Advanced Micro Devices, 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: Alex Deucher
23 */
24 #include <linux/firmware.h>
25 #include <linux/platform_device.h>
26 #include <linux/slab.h>
27 #include <linux/module.h>
28 #include "drmP.h"
29 #include "radeon.h"
30 #include "radeon_asic.h"
31 #include "radeon_drm.h"
32 #include "nid.h"
33 #include "atom.h"
34 #include "ni_reg.h"
35 #include "cayman_blit_shaders.h"
36
37 extern void evergreen_mc_stop(struct radeon_device *rdev, struct evergreen_mc_save *save);
38 extern void evergreen_mc_resume(struct radeon_device *rdev, struct evergreen_mc_save *save);
39 extern int evergreen_mc_wait_for_idle(struct radeon_device *rdev);
40 extern void evergreen_mc_program(struct radeon_device *rdev);
41 extern void evergreen_irq_suspend(struct radeon_device *rdev);
42 extern int evergreen_mc_init(struct radeon_device *rdev);
43 extern void evergreen_fix_pci_max_read_req_size(struct radeon_device *rdev);
44 extern void evergreen_pcie_gen2_enable(struct radeon_device *rdev);
45 extern void si_rlc_fini(struct radeon_device *rdev);
46 extern int si_rlc_init(struct radeon_device *rdev);
47
48 #define EVERGREEN_PFP_UCODE_SIZE 1120
49 #define EVERGREEN_PM4_UCODE_SIZE 1376
50 #define EVERGREEN_RLC_UCODE_SIZE 768
51 #define BTC_MC_UCODE_SIZE 6024
52
53 #define CAYMAN_PFP_UCODE_SIZE 2176
54 #define CAYMAN_PM4_UCODE_SIZE 2176
55 #define CAYMAN_RLC_UCODE_SIZE 1024
56 #define CAYMAN_MC_UCODE_SIZE 6037
57
58 #define ARUBA_RLC_UCODE_SIZE 1536
59
60 /* Firmware Names */
61 MODULE_FIRMWARE("radeon/BARTS_pfp.bin");
62 MODULE_FIRMWARE("radeon/BARTS_me.bin");
63 MODULE_FIRMWARE("radeon/BARTS_mc.bin");
64 MODULE_FIRMWARE("radeon/BTC_rlc.bin");
65 MODULE_FIRMWARE("radeon/TURKS_pfp.bin");
66 MODULE_FIRMWARE("radeon/TURKS_me.bin");
67 MODULE_FIRMWARE("radeon/TURKS_mc.bin");
68 MODULE_FIRMWARE("radeon/CAICOS_pfp.bin");
69 MODULE_FIRMWARE("radeon/CAICOS_me.bin");
70 MODULE_FIRMWARE("radeon/CAICOS_mc.bin");
71 MODULE_FIRMWARE("radeon/CAYMAN_pfp.bin");
72 MODULE_FIRMWARE("radeon/CAYMAN_me.bin");
73 MODULE_FIRMWARE("radeon/CAYMAN_mc.bin");
74 MODULE_FIRMWARE("radeon/CAYMAN_rlc.bin");
75 MODULE_FIRMWARE("radeon/ARUBA_pfp.bin");
76 MODULE_FIRMWARE("radeon/ARUBA_me.bin");
77 MODULE_FIRMWARE("radeon/ARUBA_rlc.bin");
78
79 #define BTC_IO_MC_REGS_SIZE 29
80
81 static const u32 barts_io_mc_regs[BTC_IO_MC_REGS_SIZE][2] = {
82 {0x00000077, 0xff010100},
83 {0x00000078, 0x00000000},
84 {0x00000079, 0x00001434},
85 {0x0000007a, 0xcc08ec08},
86 {0x0000007b, 0x00040000},
87 {0x0000007c, 0x000080c0},
88 {0x0000007d, 0x09000000},
89 {0x0000007e, 0x00210404},
90 {0x00000081, 0x08a8e800},
91 {0x00000082, 0x00030444},
92 {0x00000083, 0x00000000},
93 {0x00000085, 0x00000001},
94 {0x00000086, 0x00000002},
95 {0x00000087, 0x48490000},
96 {0x00000088, 0x20244647},
97 {0x00000089, 0x00000005},
98 {0x0000008b, 0x66030000},
99 {0x0000008c, 0x00006603},
100 {0x0000008d, 0x00000100},
101 {0x0000008f, 0x00001c0a},
102 {0x00000090, 0xff000001},
103 {0x00000094, 0x00101101},
104 {0x00000095, 0x00000fff},
105 {0x00000096, 0x00116fff},
106 {0x00000097, 0x60010000},
107 {0x00000098, 0x10010000},
108 {0x00000099, 0x00006000},
109 {0x0000009a, 0x00001000},
110 {0x0000009f, 0x00946a00}
111 };
112
113 static const u32 turks_io_mc_regs[BTC_IO_MC_REGS_SIZE][2] = {
114 {0x00000077, 0xff010100},
115 {0x00000078, 0x00000000},
116 {0x00000079, 0x00001434},
117 {0x0000007a, 0xcc08ec08},
118 {0x0000007b, 0x00040000},
119 {0x0000007c, 0x000080c0},
120 {0x0000007d, 0x09000000},
121 {0x0000007e, 0x00210404},
122 {0x00000081, 0x08a8e800},
123 {0x00000082, 0x00030444},
124 {0x00000083, 0x00000000},
125 {0x00000085, 0x00000001},
126 {0x00000086, 0x00000002},
127 {0x00000087, 0x48490000},
128 {0x00000088, 0x20244647},
129 {0x00000089, 0x00000005},
130 {0x0000008b, 0x66030000},
131 {0x0000008c, 0x00006603},
132 {0x0000008d, 0x00000100},
133 {0x0000008f, 0x00001c0a},
134 {0x00000090, 0xff000001},
135 {0x00000094, 0x00101101},
136 {0x00000095, 0x00000fff},
137 {0x00000096, 0x00116fff},
138 {0x00000097, 0x60010000},
139 {0x00000098, 0x10010000},
140 {0x00000099, 0x00006000},
141 {0x0000009a, 0x00001000},
142 {0x0000009f, 0x00936a00}
143 };
144
145 static const u32 caicos_io_mc_regs[BTC_IO_MC_REGS_SIZE][2] = {
146 {0x00000077, 0xff010100},
147 {0x00000078, 0x00000000},
148 {0x00000079, 0x00001434},
149 {0x0000007a, 0xcc08ec08},
150 {0x0000007b, 0x00040000},
151 {0x0000007c, 0x000080c0},
152 {0x0000007d, 0x09000000},
153 {0x0000007e, 0x00210404},
154 {0x00000081, 0x08a8e800},
155 {0x00000082, 0x00030444},
156 {0x00000083, 0x00000000},
157 {0x00000085, 0x00000001},
158 {0x00000086, 0x00000002},
159 {0x00000087, 0x48490000},
160 {0x00000088, 0x20244647},
161 {0x00000089, 0x00000005},
162 {0x0000008b, 0x66030000},
163 {0x0000008c, 0x00006603},
164 {0x0000008d, 0x00000100},
165 {0x0000008f, 0x00001c0a},
166 {0x00000090, 0xff000001},
167 {0x00000094, 0x00101101},
168 {0x00000095, 0x00000fff},
169 {0x00000096, 0x00116fff},
170 {0x00000097, 0x60010000},
171 {0x00000098, 0x10010000},
172 {0x00000099, 0x00006000},
173 {0x0000009a, 0x00001000},
174 {0x0000009f, 0x00916a00}
175 };
176
177 static const u32 cayman_io_mc_regs[BTC_IO_MC_REGS_SIZE][2] = {
178 {0x00000077, 0xff010100},
179 {0x00000078, 0x00000000},
180 {0x00000079, 0x00001434},
181 {0x0000007a, 0xcc08ec08},
182 {0x0000007b, 0x00040000},
183 {0x0000007c, 0x000080c0},
184 {0x0000007d, 0x09000000},
185 {0x0000007e, 0x00210404},
186 {0x00000081, 0x08a8e800},
187 {0x00000082, 0x00030444},
188 {0x00000083, 0x00000000},
189 {0x00000085, 0x00000001},
190 {0x00000086, 0x00000002},
191 {0x00000087, 0x48490000},
192 {0x00000088, 0x20244647},
193 {0x00000089, 0x00000005},
194 {0x0000008b, 0x66030000},
195 {0x0000008c, 0x00006603},
196 {0x0000008d, 0x00000100},
197 {0x0000008f, 0x00001c0a},
198 {0x00000090, 0xff000001},
199 {0x00000094, 0x00101101},
200 {0x00000095, 0x00000fff},
201 {0x00000096, 0x00116fff},
202 {0x00000097, 0x60010000},
203 {0x00000098, 0x10010000},
204 {0x00000099, 0x00006000},
205 {0x0000009a, 0x00001000},
206 {0x0000009f, 0x00976b00}
207 };
208
209 int ni_mc_load_microcode(struct radeon_device *rdev)
210 {
211 const __be32 *fw_data;
212 u32 mem_type, running, blackout = 0;
213 u32 *io_mc_regs;
214 int i, ucode_size, regs_size;
215
216 if (!rdev->mc_fw)
217 return -EINVAL;
218
219 switch (rdev->family) {
220 case CHIP_BARTS:
221 io_mc_regs = (u32 *)&barts_io_mc_regs;
222 ucode_size = BTC_MC_UCODE_SIZE;
223 regs_size = BTC_IO_MC_REGS_SIZE;
224 break;
225 case CHIP_TURKS:
226 io_mc_regs = (u32 *)&turks_io_mc_regs;
227 ucode_size = BTC_MC_UCODE_SIZE;
228 regs_size = BTC_IO_MC_REGS_SIZE;
229 break;
230 case CHIP_CAICOS:
231 default:
232 io_mc_regs = (u32 *)&caicos_io_mc_regs;
233 ucode_size = BTC_MC_UCODE_SIZE;
234 regs_size = BTC_IO_MC_REGS_SIZE;
235 break;
236 case CHIP_CAYMAN:
237 io_mc_regs = (u32 *)&cayman_io_mc_regs;
238 ucode_size = CAYMAN_MC_UCODE_SIZE;
239 regs_size = BTC_IO_MC_REGS_SIZE;
240 break;
241 }
242
243 mem_type = (RREG32(MC_SEQ_MISC0) & MC_SEQ_MISC0_GDDR5_MASK) >> MC_SEQ_MISC0_GDDR5_SHIFT;
244 running = RREG32(MC_SEQ_SUP_CNTL) & RUN_MASK;
245
246 if ((mem_type == MC_SEQ_MISC0_GDDR5_VALUE) && (running == 0)) {
247 if (running) {
248 blackout = RREG32(MC_SHARED_BLACKOUT_CNTL);
249 WREG32(MC_SHARED_BLACKOUT_CNTL, 1);
250 }
251
252 /* reset the engine and set to writable */
253 WREG32(MC_SEQ_SUP_CNTL, 0x00000008);
254 WREG32(MC_SEQ_SUP_CNTL, 0x00000010);
255
256 /* load mc io regs */
257 for (i = 0; i < regs_size; i++) {
258 WREG32(MC_SEQ_IO_DEBUG_INDEX, io_mc_regs[(i << 1)]);
259 WREG32(MC_SEQ_IO_DEBUG_DATA, io_mc_regs[(i << 1) + 1]);
260 }
261 /* load the MC ucode */
262 fw_data = (const __be32 *)rdev->mc_fw->data;
263 for (i = 0; i < ucode_size; i++)
264 WREG32(MC_SEQ_SUP_PGM, be32_to_cpup(fw_data++));
265
266 /* put the engine back into the active state */
267 WREG32(MC_SEQ_SUP_CNTL, 0x00000008);
268 WREG32(MC_SEQ_SUP_CNTL, 0x00000004);
269 WREG32(MC_SEQ_SUP_CNTL, 0x00000001);
270
271 /* wait for training to complete */
272 for (i = 0; i < rdev->usec_timeout; i++) {
273 if (RREG32(MC_IO_PAD_CNTL_D0) & MEM_FALL_OUT_CMD)
274 break;
275 udelay(1);
276 }
277
278 if (running)
279 WREG32(MC_SHARED_BLACKOUT_CNTL, blackout);
280 }
281
282 return 0;
283 }
284
285 int ni_init_microcode(struct radeon_device *rdev)
286 {
287 struct platform_device *pdev;
288 const char *chip_name;
289 const char *rlc_chip_name;
290 size_t pfp_req_size, me_req_size, rlc_req_size, mc_req_size;
291 char fw_name[30];
292 int err;
293
294 DRM_DEBUG("\n");
295
296 pdev = platform_device_register_simple("radeon_cp", 0, NULL, 0);
297 err = IS_ERR(pdev);
298 if (err) {
299 printk(KERN_ERR "radeon_cp: Failed to register firmware\n");
300 return -EINVAL;
301 }
302
303 switch (rdev->family) {
304 case CHIP_BARTS:
305 chip_name = "BARTS";
306 rlc_chip_name = "BTC";
307 pfp_req_size = EVERGREEN_PFP_UCODE_SIZE * 4;
308 me_req_size = EVERGREEN_PM4_UCODE_SIZE * 4;
309 rlc_req_size = EVERGREEN_RLC_UCODE_SIZE * 4;
310 mc_req_size = BTC_MC_UCODE_SIZE * 4;
311 break;
312 case CHIP_TURKS:
313 chip_name = "TURKS";
314 rlc_chip_name = "BTC";
315 pfp_req_size = EVERGREEN_PFP_UCODE_SIZE * 4;
316 me_req_size = EVERGREEN_PM4_UCODE_SIZE * 4;
317 rlc_req_size = EVERGREEN_RLC_UCODE_SIZE * 4;
318 mc_req_size = BTC_MC_UCODE_SIZE * 4;
319 break;
320 case CHIP_CAICOS:
321 chip_name = "CAICOS";
322 rlc_chip_name = "BTC";
323 pfp_req_size = EVERGREEN_PFP_UCODE_SIZE * 4;
324 me_req_size = EVERGREEN_PM4_UCODE_SIZE * 4;
325 rlc_req_size = EVERGREEN_RLC_UCODE_SIZE * 4;
326 mc_req_size = BTC_MC_UCODE_SIZE * 4;
327 break;
328 case CHIP_CAYMAN:
329 chip_name = "CAYMAN";
330 rlc_chip_name = "CAYMAN";
331 pfp_req_size = CAYMAN_PFP_UCODE_SIZE * 4;
332 me_req_size = CAYMAN_PM4_UCODE_SIZE * 4;
333 rlc_req_size = CAYMAN_RLC_UCODE_SIZE * 4;
334 mc_req_size = CAYMAN_MC_UCODE_SIZE * 4;
335 break;
336 case CHIP_ARUBA:
337 chip_name = "ARUBA";
338 rlc_chip_name = "ARUBA";
339 /* pfp/me same size as CAYMAN */
340 pfp_req_size = CAYMAN_PFP_UCODE_SIZE * 4;
341 me_req_size = CAYMAN_PM4_UCODE_SIZE * 4;
342 rlc_req_size = ARUBA_RLC_UCODE_SIZE * 4;
343 mc_req_size = 0;
344 break;
345 default: BUG();
346 }
347
348 DRM_INFO("Loading %s Microcode\n", chip_name);
349
350 snprintf(fw_name, sizeof(fw_name), "radeon/%s_pfp.bin", chip_name);
351 err = request_firmware(&rdev->pfp_fw, fw_name, &pdev->dev);
352 if (err)
353 goto out;
354 if (rdev->pfp_fw->size != pfp_req_size) {
355 printk(KERN_ERR
356 "ni_cp: Bogus length %zu in firmware \"%s\"\n",
357 rdev->pfp_fw->size, fw_name);
358 err = -EINVAL;
359 goto out;
360 }
361
362 snprintf(fw_name, sizeof(fw_name), "radeon/%s_me.bin", chip_name);
363 err = request_firmware(&rdev->me_fw, fw_name, &pdev->dev);
364 if (err)
365 goto out;
366 if (rdev->me_fw->size != me_req_size) {
367 printk(KERN_ERR
368 "ni_cp: Bogus length %zu in firmware \"%s\"\n",
369 rdev->me_fw->size, fw_name);
370 err = -EINVAL;
371 }
372
373 snprintf(fw_name, sizeof(fw_name), "radeon/%s_rlc.bin", rlc_chip_name);
374 err = request_firmware(&rdev->rlc_fw, fw_name, &pdev->dev);
375 if (err)
376 goto out;
377 if (rdev->rlc_fw->size != rlc_req_size) {
378 printk(KERN_ERR
379 "ni_rlc: Bogus length %zu in firmware \"%s\"\n",
380 rdev->rlc_fw->size, fw_name);
381 err = -EINVAL;
382 }
383
384 /* no MC ucode on TN */
385 if (!(rdev->flags & RADEON_IS_IGP)) {
386 snprintf(fw_name, sizeof(fw_name), "radeon/%s_mc.bin", chip_name);
387 err = request_firmware(&rdev->mc_fw, fw_name, &pdev->dev);
388 if (err)
389 goto out;
390 if (rdev->mc_fw->size != mc_req_size) {
391 printk(KERN_ERR
392 "ni_mc: Bogus length %zu in firmware \"%s\"\n",
393 rdev->mc_fw->size, fw_name);
394 err = -EINVAL;
395 }
396 }
397 out:
398 platform_device_unregister(pdev);
399
400 if (err) {
401 if (err != -EINVAL)
402 printk(KERN_ERR
403 "ni_cp: Failed to load firmware \"%s\"\n",
404 fw_name);
405 release_firmware(rdev->pfp_fw);
406 rdev->pfp_fw = NULL;
407 release_firmware(rdev->me_fw);
408 rdev->me_fw = NULL;
409 release_firmware(rdev->rlc_fw);
410 rdev->rlc_fw = NULL;
411 release_firmware(rdev->mc_fw);
412 rdev->mc_fw = NULL;
413 }
414 return err;
415 }
416
417 /*
418 * Core functions
419 */
420 static void cayman_gpu_init(struct radeon_device *rdev)
421 {
422 u32 gb_addr_config = 0;
423 u32 mc_shared_chmap, mc_arb_ramcfg;
424 u32 cgts_tcc_disable;
425 u32 sx_debug_1;
426 u32 smx_dc_ctl0;
427 u32 cgts_sm_ctrl_reg;
428 u32 hdp_host_path_cntl;
429 u32 tmp;
430 u32 disabled_rb_mask;
431 int i, j;
432
433 switch (rdev->family) {
434 case CHIP_CAYMAN:
435 rdev->config.cayman.max_shader_engines = 2;
436 rdev->config.cayman.max_pipes_per_simd = 4;
437 rdev->config.cayman.max_tile_pipes = 8;
438 rdev->config.cayman.max_simds_per_se = 12;
439 rdev->config.cayman.max_backends_per_se = 4;
440 rdev->config.cayman.max_texture_channel_caches = 8;
441 rdev->config.cayman.max_gprs = 256;
442 rdev->config.cayman.max_threads = 256;
443 rdev->config.cayman.max_gs_threads = 32;
444 rdev->config.cayman.max_stack_entries = 512;
445 rdev->config.cayman.sx_num_of_sets = 8;
446 rdev->config.cayman.sx_max_export_size = 256;
447 rdev->config.cayman.sx_max_export_pos_size = 64;
448 rdev->config.cayman.sx_max_export_smx_size = 192;
449 rdev->config.cayman.max_hw_contexts = 8;
450 rdev->config.cayman.sq_num_cf_insts = 2;
451
452 rdev->config.cayman.sc_prim_fifo_size = 0x100;
453 rdev->config.cayman.sc_hiz_tile_fifo_size = 0x30;
454 rdev->config.cayman.sc_earlyz_tile_fifo_size = 0x130;
455 gb_addr_config = CAYMAN_GB_ADDR_CONFIG_GOLDEN;
456 break;
457 case CHIP_ARUBA:
458 default:
459 rdev->config.cayman.max_shader_engines = 1;
460 rdev->config.cayman.max_pipes_per_simd = 4;
461 rdev->config.cayman.max_tile_pipes = 2;
462 if ((rdev->pdev->device == 0x9900) ||
463 (rdev->pdev->device == 0x9901) ||
464 (rdev->pdev->device == 0x9905) ||
465 (rdev->pdev->device == 0x9906) ||
466 (rdev->pdev->device == 0x9907) ||
467 (rdev->pdev->device == 0x9908) ||
468 (rdev->pdev->device == 0x9909) ||
469 (rdev->pdev->device == 0x9910) ||
470 (rdev->pdev->device == 0x9917)) {
471 rdev->config.cayman.max_simds_per_se = 6;
472 rdev->config.cayman.max_backends_per_se = 2;
473 } else if ((rdev->pdev->device == 0x9903) ||
474 (rdev->pdev->device == 0x9904) ||
475 (rdev->pdev->device == 0x990A) ||
476 (rdev->pdev->device == 0x9913) ||
477 (rdev->pdev->device == 0x9918)) {
478 rdev->config.cayman.max_simds_per_se = 4;
479 rdev->config.cayman.max_backends_per_se = 2;
480 } else if ((rdev->pdev->device == 0x9919) ||
481 (rdev->pdev->device == 0x9990) ||
482 (rdev->pdev->device == 0x9991) ||
483 (rdev->pdev->device == 0x9994) ||
484 (rdev->pdev->device == 0x99A0)) {
485 rdev->config.cayman.max_simds_per_se = 3;
486 rdev->config.cayman.max_backends_per_se = 1;
487 } else {
488 rdev->config.cayman.max_simds_per_se = 2;
489 rdev->config.cayman.max_backends_per_se = 1;
490 }
491 rdev->config.cayman.max_texture_channel_caches = 2;
492 rdev->config.cayman.max_gprs = 256;
493 rdev->config.cayman.max_threads = 256;
494 rdev->config.cayman.max_gs_threads = 32;
495 rdev->config.cayman.max_stack_entries = 512;
496 rdev->config.cayman.sx_num_of_sets = 8;
497 rdev->config.cayman.sx_max_export_size = 256;
498 rdev->config.cayman.sx_max_export_pos_size = 64;
499 rdev->config.cayman.sx_max_export_smx_size = 192;
500 rdev->config.cayman.max_hw_contexts = 8;
501 rdev->config.cayman.sq_num_cf_insts = 2;
502
503 rdev->config.cayman.sc_prim_fifo_size = 0x40;
504 rdev->config.cayman.sc_hiz_tile_fifo_size = 0x30;
505 rdev->config.cayman.sc_earlyz_tile_fifo_size = 0x130;
506 gb_addr_config = ARUBA_GB_ADDR_CONFIG_GOLDEN;
507 break;
508 }
509
510 /* Initialize HDP */
511 for (i = 0, j = 0; i < 32; i++, j += 0x18) {
512 WREG32((0x2c14 + j), 0x00000000);
513 WREG32((0x2c18 + j), 0x00000000);
514 WREG32((0x2c1c + j), 0x00000000);
515 WREG32((0x2c20 + j), 0x00000000);
516 WREG32((0x2c24 + j), 0x00000000);
517 }
518
519 WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff));
520
521 evergreen_fix_pci_max_read_req_size(rdev);
522
523 mc_shared_chmap = RREG32(MC_SHARED_CHMAP);
524 mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG);
525
526 tmp = (mc_arb_ramcfg & NOOFCOLS_MASK) >> NOOFCOLS_SHIFT;
527 rdev->config.cayman.mem_row_size_in_kb = (4 * (1 << (8 + tmp))) / 1024;
528 if (rdev->config.cayman.mem_row_size_in_kb > 4)
529 rdev->config.cayman.mem_row_size_in_kb = 4;
530 /* XXX use MC settings? */
531 rdev->config.cayman.shader_engine_tile_size = 32;
532 rdev->config.cayman.num_gpus = 1;
533 rdev->config.cayman.multi_gpu_tile_size = 64;
534
535 tmp = (gb_addr_config & NUM_PIPES_MASK) >> NUM_PIPES_SHIFT;
536 rdev->config.cayman.num_tile_pipes = (1 << tmp);
537 tmp = (gb_addr_config & PIPE_INTERLEAVE_SIZE_MASK) >> PIPE_INTERLEAVE_SIZE_SHIFT;
538 rdev->config.cayman.mem_max_burst_length_bytes = (tmp + 1) * 256;
539 tmp = (gb_addr_config & NUM_SHADER_ENGINES_MASK) >> NUM_SHADER_ENGINES_SHIFT;
540 rdev->config.cayman.num_shader_engines = tmp + 1;
541 tmp = (gb_addr_config & NUM_GPUS_MASK) >> NUM_GPUS_SHIFT;
542 rdev->config.cayman.num_gpus = tmp + 1;
543 tmp = (gb_addr_config & MULTI_GPU_TILE_SIZE_MASK) >> MULTI_GPU_TILE_SIZE_SHIFT;
544 rdev->config.cayman.multi_gpu_tile_size = 1 << tmp;
545 tmp = (gb_addr_config & ROW_SIZE_MASK) >> ROW_SIZE_SHIFT;
546 rdev->config.cayman.mem_row_size_in_kb = 1 << tmp;
547
548
549 /* setup tiling info dword. gb_addr_config is not adequate since it does
550 * not have bank info, so create a custom tiling dword.
551 * bits 3:0 num_pipes
552 * bits 7:4 num_banks
553 * bits 11:8 group_size
554 * bits 15:12 row_size
555 */
556 rdev->config.cayman.tile_config = 0;
557 switch (rdev->config.cayman.num_tile_pipes) {
558 case 1:
559 default:
560 rdev->config.cayman.tile_config |= (0 << 0);
561 break;
562 case 2:
563 rdev->config.cayman.tile_config |= (1 << 0);
564 break;
565 case 4:
566 rdev->config.cayman.tile_config |= (2 << 0);
567 break;
568 case 8:
569 rdev->config.cayman.tile_config |= (3 << 0);
570 break;
571 }
572
573 /* num banks is 8 on all fusion asics. 0 = 4, 1 = 8, 2 = 16 */
574 if (rdev->flags & RADEON_IS_IGP)
575 rdev->config.cayman.tile_config |= 1 << 4;
576 else {
577 if ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT)
578 rdev->config.cayman.tile_config |= 1 << 4;
579 else
580 rdev->config.cayman.tile_config |= 0 << 4;
581 }
582 rdev->config.cayman.tile_config |=
583 ((gb_addr_config & PIPE_INTERLEAVE_SIZE_MASK) >> PIPE_INTERLEAVE_SIZE_SHIFT) << 8;
584 rdev->config.cayman.tile_config |=
585 ((gb_addr_config & ROW_SIZE_MASK) >> ROW_SIZE_SHIFT) << 12;
586
587 tmp = 0;
588 for (i = (rdev->config.cayman.max_shader_engines - 1); i >= 0; i--) {
589 u32 rb_disable_bitmap;
590
591 WREG32(GRBM_GFX_INDEX, INSTANCE_BROADCAST_WRITES | SE_INDEX(i));
592 WREG32(RLC_GFX_INDEX, INSTANCE_BROADCAST_WRITES | SE_INDEX(i));
593 rb_disable_bitmap = (RREG32(CC_RB_BACKEND_DISABLE) & 0x00ff0000) >> 16;
594 tmp <<= 4;
595 tmp |= rb_disable_bitmap;
596 }
597 /* enabled rb are just the one not disabled :) */
598 disabled_rb_mask = tmp;
599
600 WREG32(GRBM_GFX_INDEX, INSTANCE_BROADCAST_WRITES | SE_BROADCAST_WRITES);
601 WREG32(RLC_GFX_INDEX, INSTANCE_BROADCAST_WRITES | SE_BROADCAST_WRITES);
602
603 WREG32(GB_ADDR_CONFIG, gb_addr_config);
604 WREG32(DMIF_ADDR_CONFIG, gb_addr_config);
605 WREG32(HDP_ADDR_CONFIG, gb_addr_config);
606
607 tmp = gb_addr_config & NUM_PIPES_MASK;
608 tmp = r6xx_remap_render_backend(rdev, tmp,
609 rdev->config.cayman.max_backends_per_se *
610 rdev->config.cayman.max_shader_engines,
611 CAYMAN_MAX_BACKENDS, disabled_rb_mask);
612 WREG32(GB_BACKEND_MAP, tmp);
613
614 cgts_tcc_disable = 0xffff0000;
615 for (i = 0; i < rdev->config.cayman.max_texture_channel_caches; i++)
616 cgts_tcc_disable &= ~(1 << (16 + i));
617 WREG32(CGTS_TCC_DISABLE, cgts_tcc_disable);
618 WREG32(CGTS_SYS_TCC_DISABLE, cgts_tcc_disable);
619 WREG32(CGTS_USER_SYS_TCC_DISABLE, cgts_tcc_disable);
620 WREG32(CGTS_USER_TCC_DISABLE, cgts_tcc_disable);
621
622 /* reprogram the shader complex */
623 cgts_sm_ctrl_reg = RREG32(CGTS_SM_CTRL_REG);
624 for (i = 0; i < 16; i++)
625 WREG32(CGTS_SM_CTRL_REG, OVERRIDE);
626 WREG32(CGTS_SM_CTRL_REG, cgts_sm_ctrl_reg);
627
628 /* set HW defaults for 3D engine */
629 WREG32(CP_MEQ_THRESHOLDS, MEQ1_START(0x30) | MEQ2_START(0x60));
630
631 sx_debug_1 = RREG32(SX_DEBUG_1);
632 sx_debug_1 |= ENABLE_NEW_SMX_ADDRESS;
633 WREG32(SX_DEBUG_1, sx_debug_1);
634
635 smx_dc_ctl0 = RREG32(SMX_DC_CTL0);
636 smx_dc_ctl0 &= ~NUMBER_OF_SETS(0x1ff);
637 smx_dc_ctl0 |= NUMBER_OF_SETS(rdev->config.cayman.sx_num_of_sets);
638 WREG32(SMX_DC_CTL0, smx_dc_ctl0);
639
640 WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4) | CRC_SIMD_ID_WADDR_DISABLE);
641
642 /* need to be explicitly zero-ed */
643 WREG32(VGT_OFFCHIP_LDS_BASE, 0);
644 WREG32(SQ_LSTMP_RING_BASE, 0);
645 WREG32(SQ_HSTMP_RING_BASE, 0);
646 WREG32(SQ_ESTMP_RING_BASE, 0);
647 WREG32(SQ_GSTMP_RING_BASE, 0);
648 WREG32(SQ_VSTMP_RING_BASE, 0);
649 WREG32(SQ_PSTMP_RING_BASE, 0);
650
651 WREG32(TA_CNTL_AUX, DISABLE_CUBE_ANISO);
652
653 WREG32(SX_EXPORT_BUFFER_SIZES, (COLOR_BUFFER_SIZE((rdev->config.cayman.sx_max_export_size / 4) - 1) |
654 POSITION_BUFFER_SIZE((rdev->config.cayman.sx_max_export_pos_size / 4) - 1) |
655 SMX_BUFFER_SIZE((rdev->config.cayman.sx_max_export_smx_size / 4) - 1)));
656
657 WREG32(PA_SC_FIFO_SIZE, (SC_PRIM_FIFO_SIZE(rdev->config.cayman.sc_prim_fifo_size) |
658 SC_HIZ_TILE_FIFO_SIZE(rdev->config.cayman.sc_hiz_tile_fifo_size) |
659 SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.cayman.sc_earlyz_tile_fifo_size)));
660
661
662 WREG32(VGT_NUM_INSTANCES, 1);
663
664 WREG32(CP_PERFMON_CNTL, 0);
665
666 WREG32(SQ_MS_FIFO_SIZES, (CACHE_FIFO_SIZE(16 * rdev->config.cayman.sq_num_cf_insts) |
667 FETCH_FIFO_HIWATER(0x4) |
668 DONE_FIFO_HIWATER(0xe0) |
669 ALU_UPDATE_FIFO_HIWATER(0x8)));
670
671 WREG32(SQ_GPR_RESOURCE_MGMT_1, NUM_CLAUSE_TEMP_GPRS(4));
672 WREG32(SQ_CONFIG, (VC_ENABLE |
673 EXPORT_SRC_C |
674 GFX_PRIO(0) |
675 CS1_PRIO(0) |
676 CS2_PRIO(1)));
677 WREG32(SQ_DYN_GPR_CNTL_PS_FLUSH_REQ, DYN_GPR_ENABLE);
678
679 WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) |
680 FORCE_EOV_MAX_REZ_CNT(255)));
681
682 WREG32(VGT_CACHE_INVALIDATION, CACHE_INVALIDATION(VC_AND_TC) |
683 AUTO_INVLD_EN(ES_AND_GS_AUTO));
684
685 WREG32(VGT_GS_VERTEX_REUSE, 16);
686 WREG32(PA_SC_LINE_STIPPLE_STATE, 0);
687
688 WREG32(CB_PERF_CTR0_SEL_0, 0);
689 WREG32(CB_PERF_CTR0_SEL_1, 0);
690 WREG32(CB_PERF_CTR1_SEL_0, 0);
691 WREG32(CB_PERF_CTR1_SEL_1, 0);
692 WREG32(CB_PERF_CTR2_SEL_0, 0);
693 WREG32(CB_PERF_CTR2_SEL_1, 0);
694 WREG32(CB_PERF_CTR3_SEL_0, 0);
695 WREG32(CB_PERF_CTR3_SEL_1, 0);
696
697 tmp = RREG32(HDP_MISC_CNTL);
698 tmp |= HDP_FLUSH_INVALIDATE_CACHE;
699 WREG32(HDP_MISC_CNTL, tmp);
700
701 hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL);
702 WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
703
704 WREG32(PA_CL_ENHANCE, CLIP_VTX_REORDER_ENA | NUM_CLIP_SEQ(3));
705
706 udelay(50);
707 }
708
709 /*
710 * GART
711 */
712 void cayman_pcie_gart_tlb_flush(struct radeon_device *rdev)
713 {
714 /* flush hdp cache */
715 WREG32(HDP_MEM_COHERENCY_FLUSH_CNTL, 0x1);
716
717 /* bits 0-7 are the VM contexts0-7 */
718 WREG32(VM_INVALIDATE_REQUEST, 1);
719 }
720
721 int cayman_pcie_gart_enable(struct radeon_device *rdev)
722 {
723 int i, r;
724
725 if (rdev->gart.robj == NULL) {
726 dev_err(rdev->dev, "No VRAM object for PCIE GART.\n");
727 return -EINVAL;
728 }
729 r = radeon_gart_table_vram_pin(rdev);
730 if (r)
731 return r;
732 radeon_gart_restore(rdev);
733 /* Setup TLB control */
734 WREG32(MC_VM_MX_L1_TLB_CNTL,
735 (0xA << 7) |
736 ENABLE_L1_TLB |
737 ENABLE_L1_FRAGMENT_PROCESSING |
738 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
739 ENABLE_ADVANCED_DRIVER_MODEL |
740 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
741 /* Setup L2 cache */
742 WREG32(VM_L2_CNTL, ENABLE_L2_CACHE |
743 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
744 ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
745 EFFECTIVE_L2_QUEUE_SIZE(7) |
746 CONTEXT1_IDENTITY_ACCESS_MODE(1));
747 WREG32(VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS | INVALIDATE_L2_CACHE);
748 WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
749 L2_CACHE_BIGK_FRAGMENT_SIZE(6));
750 /* setup context0 */
751 WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);
752 WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);
753 WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12);
754 WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
755 (u32)(rdev->dummy_page.addr >> 12));
756 WREG32(VM_CONTEXT0_CNTL2, 0);
757 WREG32(VM_CONTEXT0_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) |
758 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT);
759
760 WREG32(0x15D4, 0);
761 WREG32(0x15D8, 0);
762 WREG32(0x15DC, 0);
763
764 /* empty context1-7 */
765 for (i = 1; i < 8; i++) {
766 WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR + (i << 2), 0);
767 WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR + (i << 2), 0);
768 WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (i << 2),
769 rdev->gart.table_addr >> 12);
770 }
771
772 /* enable context1-7 */
773 WREG32(VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR,
774 (u32)(rdev->dummy_page.addr >> 12));
775 WREG32(VM_CONTEXT1_CNTL2, 0);
776 WREG32(VM_CONTEXT1_CNTL, 0);
777 WREG32(VM_CONTEXT1_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) |
778 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT);
779
780 cayman_pcie_gart_tlb_flush(rdev);
781 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
782 (unsigned)(rdev->mc.gtt_size >> 20),
783 (unsigned long long)rdev->gart.table_addr);
784 rdev->gart.ready = true;
785 return 0;
786 }
787
788 void cayman_pcie_gart_disable(struct radeon_device *rdev)
789 {
790 /* Disable all tables */
791 WREG32(VM_CONTEXT0_CNTL, 0);
792 WREG32(VM_CONTEXT1_CNTL, 0);
793 /* Setup TLB control */
794 WREG32(MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING |
795 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
796 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
797 /* Setup L2 cache */
798 WREG32(VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
799 ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
800 EFFECTIVE_L2_QUEUE_SIZE(7) |
801 CONTEXT1_IDENTITY_ACCESS_MODE(1));
802 WREG32(VM_L2_CNTL2, 0);
803 WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
804 L2_CACHE_BIGK_FRAGMENT_SIZE(6));
805 radeon_gart_table_vram_unpin(rdev);
806 }
807
808 void cayman_pcie_gart_fini(struct radeon_device *rdev)
809 {
810 cayman_pcie_gart_disable(rdev);
811 radeon_gart_table_vram_free(rdev);
812 radeon_gart_fini(rdev);
813 }
814
815 void cayman_cp_int_cntl_setup(struct radeon_device *rdev,
816 int ring, u32 cp_int_cntl)
817 {
818 u32 srbm_gfx_cntl = RREG32(SRBM_GFX_CNTL) & ~3;
819
820 WREG32(SRBM_GFX_CNTL, srbm_gfx_cntl | (ring & 3));
821 WREG32(CP_INT_CNTL, cp_int_cntl);
822 }
823
824 /*
825 * CP.
826 */
827 void cayman_fence_ring_emit(struct radeon_device *rdev,
828 struct radeon_fence *fence)
829 {
830 struct radeon_ring *ring = &rdev->ring[fence->ring];
831 u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
832
833 /* flush read cache over gart for this vmid */
834 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
835 radeon_ring_write(ring, (CP_COHER_CNTL2 - PACKET3_SET_CONFIG_REG_START) >> 2);
836 radeon_ring_write(ring, 0);
837 radeon_ring_write(ring, PACKET3(PACKET3_SURFACE_SYNC, 3));
838 radeon_ring_write(ring, PACKET3_TC_ACTION_ENA | PACKET3_SH_ACTION_ENA);
839 radeon_ring_write(ring, 0xFFFFFFFF);
840 radeon_ring_write(ring, 0);
841 radeon_ring_write(ring, 10); /* poll interval */
842 /* EVENT_WRITE_EOP - flush caches, send int */
843 radeon_ring_write(ring, PACKET3(PACKET3_EVENT_WRITE_EOP, 4));
844 radeon_ring_write(ring, EVENT_TYPE(CACHE_FLUSH_AND_INV_EVENT_TS) | EVENT_INDEX(5));
845 radeon_ring_write(ring, addr & 0xffffffff);
846 radeon_ring_write(ring, (upper_32_bits(addr) & 0xff) | DATA_SEL(1) | INT_SEL(2));
847 radeon_ring_write(ring, fence->seq);
848 radeon_ring_write(ring, 0);
849 }
850
851 void cayman_ring_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib)
852 {
853 struct radeon_ring *ring = &rdev->ring[ib->ring];
854
855 /* set to DX10/11 mode */
856 radeon_ring_write(ring, PACKET3(PACKET3_MODE_CONTROL, 0));
857 radeon_ring_write(ring, 1);
858 radeon_ring_write(ring, PACKET3(PACKET3_INDIRECT_BUFFER, 2));
859 radeon_ring_write(ring,
860 #ifdef __BIG_ENDIAN
861 (2 << 0) |
862 #endif
863 (ib->gpu_addr & 0xFFFFFFFC));
864 radeon_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xFF);
865 radeon_ring_write(ring, ib->length_dw | (ib->vm_id << 24));
866
867 /* flush read cache over gart for this vmid */
868 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
869 radeon_ring_write(ring, (CP_COHER_CNTL2 - PACKET3_SET_CONFIG_REG_START) >> 2);
870 radeon_ring_write(ring, ib->vm_id);
871 radeon_ring_write(ring, PACKET3(PACKET3_SURFACE_SYNC, 3));
872 radeon_ring_write(ring, PACKET3_TC_ACTION_ENA | PACKET3_SH_ACTION_ENA);
873 radeon_ring_write(ring, 0xFFFFFFFF);
874 radeon_ring_write(ring, 0);
875 radeon_ring_write(ring, 10); /* poll interval */
876 }
877
878 static void cayman_cp_enable(struct radeon_device *rdev, bool enable)
879 {
880 if (enable)
881 WREG32(CP_ME_CNTL, 0);
882 else {
883 radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
884 WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT));
885 WREG32(SCRATCH_UMSK, 0);
886 }
887 }
888
889 static int cayman_cp_load_microcode(struct radeon_device *rdev)
890 {
891 const __be32 *fw_data;
892 int i;
893
894 if (!rdev->me_fw || !rdev->pfp_fw)
895 return -EINVAL;
896
897 cayman_cp_enable(rdev, false);
898
899 fw_data = (const __be32 *)rdev->pfp_fw->data;
900 WREG32(CP_PFP_UCODE_ADDR, 0);
901 for (i = 0; i < CAYMAN_PFP_UCODE_SIZE; i++)
902 WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++));
903 WREG32(CP_PFP_UCODE_ADDR, 0);
904
905 fw_data = (const __be32 *)rdev->me_fw->data;
906 WREG32(CP_ME_RAM_WADDR, 0);
907 for (i = 0; i < CAYMAN_PM4_UCODE_SIZE; i++)
908 WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++));
909
910 WREG32(CP_PFP_UCODE_ADDR, 0);
911 WREG32(CP_ME_RAM_WADDR, 0);
912 WREG32(CP_ME_RAM_RADDR, 0);
913 return 0;
914 }
915
916 static int cayman_cp_start(struct radeon_device *rdev)
917 {
918 struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
919 int r, i;
920
921 r = radeon_ring_lock(rdev, ring, 7);
922 if (r) {
923 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
924 return r;
925 }
926 radeon_ring_write(ring, PACKET3(PACKET3_ME_INITIALIZE, 5));
927 radeon_ring_write(ring, 0x1);
928 radeon_ring_write(ring, 0x0);
929 radeon_ring_write(ring, rdev->config.cayman.max_hw_contexts - 1);
930 radeon_ring_write(ring, PACKET3_ME_INITIALIZE_DEVICE_ID(1));
931 radeon_ring_write(ring, 0);
932 radeon_ring_write(ring, 0);
933 radeon_ring_unlock_commit(rdev, ring);
934
935 cayman_cp_enable(rdev, true);
936
937 r = radeon_ring_lock(rdev, ring, cayman_default_size + 19);
938 if (r) {
939 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
940 return r;
941 }
942
943 /* setup clear context state */
944 radeon_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
945 radeon_ring_write(ring, PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
946
947 for (i = 0; i < cayman_default_size; i++)
948 radeon_ring_write(ring, cayman_default_state[i]);
949
950 radeon_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
951 radeon_ring_write(ring, PACKET3_PREAMBLE_END_CLEAR_STATE);
952
953 /* set clear context state */
954 radeon_ring_write(ring, PACKET3(PACKET3_CLEAR_STATE, 0));
955 radeon_ring_write(ring, 0);
956
957 /* SQ_VTX_BASE_VTX_LOC */
958 radeon_ring_write(ring, 0xc0026f00);
959 radeon_ring_write(ring, 0x00000000);
960 radeon_ring_write(ring, 0x00000000);
961 radeon_ring_write(ring, 0x00000000);
962
963 /* Clear consts */
964 radeon_ring_write(ring, 0xc0036f00);
965 radeon_ring_write(ring, 0x00000bc4);
966 radeon_ring_write(ring, 0xffffffff);
967 radeon_ring_write(ring, 0xffffffff);
968 radeon_ring_write(ring, 0xffffffff);
969
970 radeon_ring_write(ring, 0xc0026900);
971 radeon_ring_write(ring, 0x00000316);
972 radeon_ring_write(ring, 0x0000000e); /* VGT_VERTEX_REUSE_BLOCK_CNTL */
973 radeon_ring_write(ring, 0x00000010); /* */
974
975 radeon_ring_unlock_commit(rdev, ring);
976
977 /* XXX init other rings */
978
979 return 0;
980 }
981
982 static void cayman_cp_fini(struct radeon_device *rdev)
983 {
984 cayman_cp_enable(rdev, false);
985 radeon_ring_fini(rdev, &rdev->ring[RADEON_RING_TYPE_GFX_INDEX]);
986 }
987
988 int cayman_cp_resume(struct radeon_device *rdev)
989 {
990 static const int ridx[] = {
991 RADEON_RING_TYPE_GFX_INDEX,
992 CAYMAN_RING_TYPE_CP1_INDEX,
993 CAYMAN_RING_TYPE_CP2_INDEX
994 };
995 static const unsigned cp_rb_cntl[] = {
996 CP_RB0_CNTL,
997 CP_RB1_CNTL,
998 CP_RB2_CNTL,
999 };
1000 static const unsigned cp_rb_rptr_addr[] = {
1001 CP_RB0_RPTR_ADDR,
1002 CP_RB1_RPTR_ADDR,
1003 CP_RB2_RPTR_ADDR
1004 };
1005 static const unsigned cp_rb_rptr_addr_hi[] = {
1006 CP_RB0_RPTR_ADDR_HI,
1007 CP_RB1_RPTR_ADDR_HI,
1008 CP_RB2_RPTR_ADDR_HI
1009 };
1010 static const unsigned cp_rb_base[] = {
1011 CP_RB0_BASE,
1012 CP_RB1_BASE,
1013 CP_RB2_BASE
1014 };
1015 struct radeon_ring *ring;
1016 int i, r;
1017
1018 /* Reset cp; if cp is reset, then PA, SH, VGT also need to be reset */
1019 WREG32(GRBM_SOFT_RESET, (SOFT_RESET_CP |
1020 SOFT_RESET_PA |
1021 SOFT_RESET_SH |
1022 SOFT_RESET_VGT |
1023 SOFT_RESET_SPI |
1024 SOFT_RESET_SX));
1025 RREG32(GRBM_SOFT_RESET);
1026 mdelay(15);
1027 WREG32(GRBM_SOFT_RESET, 0);
1028 RREG32(GRBM_SOFT_RESET);
1029
1030 WREG32(CP_SEM_WAIT_TIMER, 0x0);
1031 WREG32(CP_SEM_INCOMPLETE_TIMER_CNTL, 0x0);
1032
1033 /* Set the write pointer delay */
1034 WREG32(CP_RB_WPTR_DELAY, 0);
1035
1036 WREG32(CP_DEBUG, (1 << 27));
1037
1038 /* set the wb address wether it's enabled or not */
1039 WREG32(SCRATCH_ADDR, ((rdev->wb.gpu_addr + RADEON_WB_SCRATCH_OFFSET) >> 8) & 0xFFFFFFFF);
1040 WREG32(SCRATCH_UMSK, 0xff);
1041
1042 for (i = 0; i < 3; ++i) {
1043 uint32_t rb_cntl;
1044 uint64_t addr;
1045
1046 /* Set ring buffer size */
1047 ring = &rdev->ring[ridx[i]];
1048 rb_cntl = drm_order(ring->ring_size / 8);
1049 rb_cntl |= drm_order(RADEON_GPU_PAGE_SIZE/8) << 8;
1050 #ifdef __BIG_ENDIAN
1051 rb_cntl |= BUF_SWAP_32BIT;
1052 #endif
1053 WREG32(cp_rb_cntl[i], rb_cntl);
1054
1055 /* set the wb address wether it's enabled or not */
1056 addr = rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET;
1057 WREG32(cp_rb_rptr_addr[i], addr & 0xFFFFFFFC);
1058 WREG32(cp_rb_rptr_addr_hi[i], upper_32_bits(addr) & 0xFF);
1059 }
1060
1061 /* set the rb base addr, this causes an internal reset of ALL rings */
1062 for (i = 0; i < 3; ++i) {
1063 ring = &rdev->ring[ridx[i]];
1064 WREG32(cp_rb_base[i], ring->gpu_addr >> 8);
1065 }
1066
1067 for (i = 0; i < 3; ++i) {
1068 /* Initialize the ring buffer's read and write pointers */
1069 ring = &rdev->ring[ridx[i]];
1070 WREG32_P(cp_rb_cntl[i], RB_RPTR_WR_ENA, ~RB_RPTR_WR_ENA);
1071
1072 ring->rptr = ring->wptr = 0;
1073 WREG32(ring->rptr_reg, ring->rptr);
1074 WREG32(ring->wptr_reg, ring->wptr);
1075
1076 mdelay(1);
1077 WREG32_P(cp_rb_cntl[i], 0, ~RB_RPTR_WR_ENA);
1078 }
1079
1080 /* start the rings */
1081 cayman_cp_start(rdev);
1082 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = true;
1083 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
1084 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
1085 /* this only test cp0 */
1086 r = radeon_ring_test(rdev, RADEON_RING_TYPE_GFX_INDEX, &rdev->ring[RADEON_RING_TYPE_GFX_INDEX]);
1087 if (r) {
1088 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
1089 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
1090 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
1091 return r;
1092 }
1093
1094 return 0;
1095 }
1096
1097 static int cayman_gpu_soft_reset(struct radeon_device *rdev)
1098 {
1099 struct evergreen_mc_save save;
1100 u32 grbm_reset = 0;
1101
1102 if (!(RREG32(GRBM_STATUS) & GUI_ACTIVE))
1103 return 0;
1104
1105 dev_info(rdev->dev, "GPU softreset \n");
1106 dev_info(rdev->dev, " GRBM_STATUS=0x%08X\n",
1107 RREG32(GRBM_STATUS));
1108 dev_info(rdev->dev, " GRBM_STATUS_SE0=0x%08X\n",
1109 RREG32(GRBM_STATUS_SE0));
1110 dev_info(rdev->dev, " GRBM_STATUS_SE1=0x%08X\n",
1111 RREG32(GRBM_STATUS_SE1));
1112 dev_info(rdev->dev, " SRBM_STATUS=0x%08X\n",
1113 RREG32(SRBM_STATUS));
1114 dev_info(rdev->dev, " VM_CONTEXT0_PROTECTION_FAULT_ADDR 0x%08X\n",
1115 RREG32(0x14F8));
1116 dev_info(rdev->dev, " VM_CONTEXT0_PROTECTION_FAULT_STATUS 0x%08X\n",
1117 RREG32(0x14D8));
1118 dev_info(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n",
1119 RREG32(0x14FC));
1120 dev_info(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
1121 RREG32(0x14DC));
1122
1123 evergreen_mc_stop(rdev, &save);
1124 if (evergreen_mc_wait_for_idle(rdev)) {
1125 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
1126 }
1127 /* Disable CP parsing/prefetching */
1128 WREG32(CP_ME_CNTL, CP_ME_HALT | CP_PFP_HALT);
1129
1130 /* reset all the gfx blocks */
1131 grbm_reset = (SOFT_RESET_CP |
1132 SOFT_RESET_CB |
1133 SOFT_RESET_DB |
1134 SOFT_RESET_GDS |
1135 SOFT_RESET_PA |
1136 SOFT_RESET_SC |
1137 SOFT_RESET_SPI |
1138 SOFT_RESET_SH |
1139 SOFT_RESET_SX |
1140 SOFT_RESET_TC |
1141 SOFT_RESET_TA |
1142 SOFT_RESET_VGT |
1143 SOFT_RESET_IA);
1144
1145 dev_info(rdev->dev, " GRBM_SOFT_RESET=0x%08X\n", grbm_reset);
1146 WREG32(GRBM_SOFT_RESET, grbm_reset);
1147 (void)RREG32(GRBM_SOFT_RESET);
1148 udelay(50);
1149 WREG32(GRBM_SOFT_RESET, 0);
1150 (void)RREG32(GRBM_SOFT_RESET);
1151 /* Wait a little for things to settle down */
1152 udelay(50);
1153
1154 dev_info(rdev->dev, " GRBM_STATUS=0x%08X\n",
1155 RREG32(GRBM_STATUS));
1156 dev_info(rdev->dev, " GRBM_STATUS_SE0=0x%08X\n",
1157 RREG32(GRBM_STATUS_SE0));
1158 dev_info(rdev->dev, " GRBM_STATUS_SE1=0x%08X\n",
1159 RREG32(GRBM_STATUS_SE1));
1160 dev_info(rdev->dev, " SRBM_STATUS=0x%08X\n",
1161 RREG32(SRBM_STATUS));
1162 evergreen_mc_resume(rdev, &save);
1163 return 0;
1164 }
1165
1166 int cayman_asic_reset(struct radeon_device *rdev)
1167 {
1168 return cayman_gpu_soft_reset(rdev);
1169 }
1170
1171 static int cayman_startup(struct radeon_device *rdev)
1172 {
1173 struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
1174 int r;
1175
1176 /* enable pcie gen2 link */
1177 evergreen_pcie_gen2_enable(rdev);
1178
1179 if (rdev->flags & RADEON_IS_IGP) {
1180 if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw) {
1181 r = ni_init_microcode(rdev);
1182 if (r) {
1183 DRM_ERROR("Failed to load firmware!\n");
1184 return r;
1185 }
1186 }
1187 } else {
1188 if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw || !rdev->mc_fw) {
1189 r = ni_init_microcode(rdev);
1190 if (r) {
1191 DRM_ERROR("Failed to load firmware!\n");
1192 return r;
1193 }
1194 }
1195
1196 r = ni_mc_load_microcode(rdev);
1197 if (r) {
1198 DRM_ERROR("Failed to load MC firmware!\n");
1199 return r;
1200 }
1201 }
1202
1203 r = r600_vram_scratch_init(rdev);
1204 if (r)
1205 return r;
1206
1207 evergreen_mc_program(rdev);
1208 r = cayman_pcie_gart_enable(rdev);
1209 if (r)
1210 return r;
1211 cayman_gpu_init(rdev);
1212
1213 r = evergreen_blit_init(rdev);
1214 if (r) {
1215 r600_blit_fini(rdev);
1216 rdev->asic->copy.copy = NULL;
1217 dev_warn(rdev->dev, "failed blitter (%d) falling back to memcpy\n", r);
1218 }
1219
1220 /* allocate rlc buffers */
1221 if (rdev->flags & RADEON_IS_IGP) {
1222 r = si_rlc_init(rdev);
1223 if (r) {
1224 DRM_ERROR("Failed to init rlc BOs!\n");
1225 return r;
1226 }
1227 }
1228
1229 /* allocate wb buffer */
1230 r = radeon_wb_init(rdev);
1231 if (r)
1232 return r;
1233
1234 r = radeon_fence_driver_start_ring(rdev, RADEON_RING_TYPE_GFX_INDEX);
1235 if (r) {
1236 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
1237 return r;
1238 }
1239
1240 r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
1241 if (r) {
1242 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
1243 return r;
1244 }
1245
1246 r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
1247 if (r) {
1248 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
1249 return r;
1250 }
1251
1252 /* Enable IRQ */
1253 r = r600_irq_init(rdev);
1254 if (r) {
1255 DRM_ERROR("radeon: IH init failed (%d).\n", r);
1256 radeon_irq_kms_fini(rdev);
1257 return r;
1258 }
1259 evergreen_irq_set(rdev);
1260
1261 r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP_RPTR_OFFSET,
1262 CP_RB0_RPTR, CP_RB0_WPTR,
1263 0, 0xfffff, RADEON_CP_PACKET2);
1264 if (r)
1265 return r;
1266 r = cayman_cp_load_microcode(rdev);
1267 if (r)
1268 return r;
1269 r = cayman_cp_resume(rdev);
1270 if (r)
1271 return r;
1272
1273 r = radeon_ib_pool_init(rdev);
1274 if (r) {
1275 dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
1276 return r;
1277 }
1278
1279 r = radeon_ib_ring_tests(rdev);
1280 if (r)
1281 return r;
1282
1283 r = radeon_vm_manager_start(rdev);
1284 if (r)
1285 return r;
1286
1287 r = r600_audio_init(rdev);
1288 if (r)
1289 return r;
1290
1291 return 0;
1292 }
1293
1294 int cayman_resume(struct radeon_device *rdev)
1295 {
1296 int r;
1297
1298 /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw,
1299 * posting will perform necessary task to bring back GPU into good
1300 * shape.
1301 */
1302 /* post card */
1303 atom_asic_init(rdev->mode_info.atom_context);
1304
1305 rdev->accel_working = true;
1306 r = cayman_startup(rdev);
1307 if (r) {
1308 DRM_ERROR("cayman startup failed on resume\n");
1309 rdev->accel_working = false;
1310 return r;
1311 }
1312 return r;
1313 }
1314
1315 int cayman_suspend(struct radeon_device *rdev)
1316 {
1317 r600_audio_fini(rdev);
1318 radeon_vm_manager_suspend(rdev);
1319 cayman_cp_enable(rdev, false);
1320 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
1321 evergreen_irq_suspend(rdev);
1322 radeon_wb_disable(rdev);
1323 cayman_pcie_gart_disable(rdev);
1324 return 0;
1325 }
1326
1327 /* Plan is to move initialization in that function and use
1328 * helper function so that radeon_device_init pretty much
1329 * do nothing more than calling asic specific function. This
1330 * should also allow to remove a bunch of callback function
1331 * like vram_info.
1332 */
1333 int cayman_init(struct radeon_device *rdev)
1334 {
1335 struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
1336 int r;
1337
1338 /* Read BIOS */
1339 if (!radeon_get_bios(rdev)) {
1340 if (ASIC_IS_AVIVO(rdev))
1341 return -EINVAL;
1342 }
1343 /* Must be an ATOMBIOS */
1344 if (!rdev->is_atom_bios) {
1345 dev_err(rdev->dev, "Expecting atombios for cayman GPU\n");
1346 return -EINVAL;
1347 }
1348 r = radeon_atombios_init(rdev);
1349 if (r)
1350 return r;
1351
1352 /* Post card if necessary */
1353 if (!radeon_card_posted(rdev)) {
1354 if (!rdev->bios) {
1355 dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n");
1356 return -EINVAL;
1357 }
1358 DRM_INFO("GPU not posted. posting now...\n");
1359 atom_asic_init(rdev->mode_info.atom_context);
1360 }
1361 /* Initialize scratch registers */
1362 r600_scratch_init(rdev);
1363 /* Initialize surface registers */
1364 radeon_surface_init(rdev);
1365 /* Initialize clocks */
1366 radeon_get_clock_info(rdev->ddev);
1367 /* Fence driver */
1368 r = radeon_fence_driver_init(rdev);
1369 if (r)
1370 return r;
1371 /* initialize memory controller */
1372 r = evergreen_mc_init(rdev);
1373 if (r)
1374 return r;
1375 /* Memory manager */
1376 r = radeon_bo_init(rdev);
1377 if (r)
1378 return r;
1379
1380 r = radeon_irq_kms_init(rdev);
1381 if (r)
1382 return r;
1383
1384 ring->ring_obj = NULL;
1385 r600_ring_init(rdev, ring, 1024 * 1024);
1386
1387 rdev->ih.ring_obj = NULL;
1388 r600_ih_ring_init(rdev, 64 * 1024);
1389
1390 r = r600_pcie_gart_init(rdev);
1391 if (r)
1392 return r;
1393
1394 rdev->accel_working = true;
1395 r = radeon_vm_manager_init(rdev);
1396 if (r) {
1397 dev_err(rdev->dev, "vm manager initialization failed (%d).\n", r);
1398 }
1399
1400 r = cayman_startup(rdev);
1401 if (r) {
1402 dev_err(rdev->dev, "disabling GPU acceleration\n");
1403 cayman_cp_fini(rdev);
1404 r600_irq_fini(rdev);
1405 if (rdev->flags & RADEON_IS_IGP)
1406 si_rlc_fini(rdev);
1407 radeon_wb_fini(rdev);
1408 radeon_ib_pool_fini(rdev);
1409 radeon_vm_manager_fini(rdev);
1410 radeon_irq_kms_fini(rdev);
1411 cayman_pcie_gart_fini(rdev);
1412 rdev->accel_working = false;
1413 }
1414
1415 /* Don't start up if the MC ucode is missing.
1416 * The default clocks and voltages before the MC ucode
1417 * is loaded are not suffient for advanced operations.
1418 *
1419 * We can skip this check for TN, because there is no MC
1420 * ucode.
1421 */
1422 if (!rdev->mc_fw && !(rdev->flags & RADEON_IS_IGP)) {
1423 DRM_ERROR("radeon: MC ucode required for NI+.\n");
1424 return -EINVAL;
1425 }
1426
1427 return 0;
1428 }
1429
1430 void cayman_fini(struct radeon_device *rdev)
1431 {
1432 r600_blit_fini(rdev);
1433 cayman_cp_fini(rdev);
1434 r600_irq_fini(rdev);
1435 if (rdev->flags & RADEON_IS_IGP)
1436 si_rlc_fini(rdev);
1437 radeon_wb_fini(rdev);
1438 radeon_vm_manager_fini(rdev);
1439 radeon_ib_pool_fini(rdev);
1440 radeon_irq_kms_fini(rdev);
1441 cayman_pcie_gart_fini(rdev);
1442 r600_vram_scratch_fini(rdev);
1443 radeon_gem_fini(rdev);
1444 radeon_fence_driver_fini(rdev);
1445 radeon_bo_fini(rdev);
1446 radeon_atombios_fini(rdev);
1447 kfree(rdev->bios);
1448 rdev->bios = NULL;
1449 }
1450
1451 /*
1452 * vm
1453 */
1454 int cayman_vm_init(struct radeon_device *rdev)
1455 {
1456 /* number of VMs */
1457 rdev->vm_manager.nvm = 8;
1458 /* base offset of vram pages */
1459 if (rdev->flags & RADEON_IS_IGP) {
1460 u64 tmp = RREG32(FUS_MC_VM_FB_OFFSET);
1461 tmp <<= 22;
1462 rdev->vm_manager.vram_base_offset = tmp;
1463 } else
1464 rdev->vm_manager.vram_base_offset = 0;
1465 return 0;
1466 }
1467
1468 void cayman_vm_fini(struct radeon_device *rdev)
1469 {
1470 }
1471
1472 int cayman_vm_bind(struct radeon_device *rdev, struct radeon_vm *vm, int id)
1473 {
1474 WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR + (id << 2), 0);
1475 WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR + (id << 2), vm->last_pfn);
1476 WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (id << 2), vm->pt_gpu_addr >> 12);
1477 /* flush hdp cache */
1478 WREG32(HDP_MEM_COHERENCY_FLUSH_CNTL, 0x1);
1479 /* bits 0-7 are the VM contexts0-7 */
1480 WREG32(VM_INVALIDATE_REQUEST, 1 << id);
1481 return 0;
1482 }
1483
1484 void cayman_vm_unbind(struct radeon_device *rdev, struct radeon_vm *vm)
1485 {
1486 WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR + (vm->id << 2), 0);
1487 WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR + (vm->id << 2), 0);
1488 WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm->id << 2), 0);
1489 /* flush hdp cache */
1490 WREG32(HDP_MEM_COHERENCY_FLUSH_CNTL, 0x1);
1491 /* bits 0-7 are the VM contexts0-7 */
1492 WREG32(VM_INVALIDATE_REQUEST, 1 << vm->id);
1493 }
1494
1495 void cayman_vm_tlb_flush(struct radeon_device *rdev, struct radeon_vm *vm)
1496 {
1497 if (vm->id == -1)
1498 return;
1499
1500 /* flush hdp cache */
1501 WREG32(HDP_MEM_COHERENCY_FLUSH_CNTL, 0x1);
1502 /* bits 0-7 are the VM contexts0-7 */
1503 WREG32(VM_INVALIDATE_REQUEST, 1 << vm->id);
1504 }
1505
1506 #define R600_PTE_VALID (1 << 0)
1507 #define R600_PTE_SYSTEM (1 << 1)
1508 #define R600_PTE_SNOOPED (1 << 2)
1509 #define R600_PTE_READABLE (1 << 5)
1510 #define R600_PTE_WRITEABLE (1 << 6)
1511
1512 uint32_t cayman_vm_page_flags(struct radeon_device *rdev,
1513 struct radeon_vm *vm,
1514 uint32_t flags)
1515 {
1516 uint32_t r600_flags = 0;
1517
1518 r600_flags |= (flags & RADEON_VM_PAGE_VALID) ? R600_PTE_VALID : 0;
1519 r600_flags |= (flags & RADEON_VM_PAGE_READABLE) ? R600_PTE_READABLE : 0;
1520 r600_flags |= (flags & RADEON_VM_PAGE_WRITEABLE) ? R600_PTE_WRITEABLE : 0;
1521 if (flags & RADEON_VM_PAGE_SYSTEM) {
1522 r600_flags |= R600_PTE_SYSTEM;
1523 r600_flags |= (flags & RADEON_VM_PAGE_SNOOPED) ? R600_PTE_SNOOPED : 0;
1524 }
1525 return r600_flags;
1526 }
1527
1528 void cayman_vm_set_page(struct radeon_device *rdev, struct radeon_vm *vm,
1529 unsigned pfn, uint64_t addr, uint32_t flags)
1530 {
1531 void __iomem *ptr = (void *)vm->pt;
1532
1533 addr = addr & 0xFFFFFFFFFFFFF000ULL;
1534 addr |= flags;
1535 writeq(addr, ptr + (pfn * 8));
1536 }
Linus' kernel tree