projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
drm/radeon: record what is next valid wptr for each ring v4
[linux-2.6-block.git] / drivers / gpu / drm / radeon / ni.c
CommitLineData
0af62b01
AD		 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>
e0cd3608 27#include <linux/module.h>
0af62b01
AD		 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"
0c88a02e 35#include "cayman_blit_shaders.h"
0af62b01 36
b9952a8a
AD		 37extern void evergreen_mc_stop(struct radeon_device *rdev, struct evergreen_mc_save *save);
38extern void evergreen_mc_resume(struct radeon_device *rdev, struct evergreen_mc_save *save);
39extern int evergreen_mc_wait_for_idle(struct radeon_device *rdev);
755d819e
AD		 40extern void evergreen_mc_program(struct radeon_device *rdev);
41extern void evergreen_irq_suspend(struct radeon_device *rdev);
42extern int evergreen_mc_init(struct radeon_device *rdev);
d054ac16 43extern void evergreen_fix_pci_max_read_req_size(struct radeon_device *rdev);
b07759bf 44extern void evergreen_pcie_gen2_enable(struct radeon_device *rdev);
c420c745
AD		 45extern void si_rlc_fini(struct radeon_device *rdev);
46extern int si_rlc_init(struct radeon_device *rdev);
b9952a8a 47
0af62b01
AD		 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
9b8253ce
AD		 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
c420c745
AD		 58#define ARUBA_RLC_UCODE_SIZE 1536
59
0af62b01
AD		 60/* Firmware Names */
61MODULE_FIRMWARE("radeon/BARTS_pfp.bin");
62MODULE_FIRMWARE("radeon/BARTS_me.bin");
63MODULE_FIRMWARE("radeon/BARTS_mc.bin");
64MODULE_FIRMWARE("radeon/BTC_rlc.bin");
65MODULE_FIRMWARE("radeon/TURKS_pfp.bin");
66MODULE_FIRMWARE("radeon/TURKS_me.bin");
67MODULE_FIRMWARE("radeon/TURKS_mc.bin");
68MODULE_FIRMWARE("radeon/CAICOS_pfp.bin");
69MODULE_FIRMWARE("radeon/CAICOS_me.bin");
70MODULE_FIRMWARE("radeon/CAICOS_mc.bin");
9b8253ce
AD		 71MODULE_FIRMWARE("radeon/CAYMAN_pfp.bin");
72MODULE_FIRMWARE("radeon/CAYMAN_me.bin");
73MODULE_FIRMWARE("radeon/CAYMAN_mc.bin");
74MODULE_FIRMWARE("radeon/CAYMAN_rlc.bin");
c420c745
AD		 75MODULE_FIRMWARE("radeon/ARUBA_pfp.bin");
76MODULE_FIRMWARE("radeon/ARUBA_me.bin");
77MODULE_FIRMWARE("radeon/ARUBA_rlc.bin");
0af62b01
AD		 78
79#define BTC_IO_MC_REGS_SIZE 29
80
81static 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
113static 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
145static 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
9b8253ce
AD		 177static 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
755d819e 209int ni_mc_load_microcode(struct radeon_device *rdev)
0af62b01
AD		 210{
211 const __be32 *fw_data;
212 u32 mem_type, running, blackout = 0;
213 u32 *io_mc_regs;
9b8253ce 214 int i, ucode_size, regs_size;
0af62b01
AD		 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;
9b8253ce
AD		 222 ucode_size = BTC_MC_UCODE_SIZE;
223 regs_size = BTC_IO_MC_REGS_SIZE;
0af62b01
AD		 224 break;
225 case CHIP_TURKS:
226 io_mc_regs = (u32 *)&turks_io_mc_regs;
9b8253ce
AD		 227 ucode_size = BTC_MC_UCODE_SIZE;
228 regs_size = BTC_IO_MC_REGS_SIZE;
0af62b01
AD		 229 break;
230 case CHIP_CAICOS:
231 default:
232 io_mc_regs = (u32 *)&caicos_io_mc_regs;
9b8253ce
AD		 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;
0af62b01
AD		 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 */
9b8253ce 257 for (i = 0; i < regs_size; i++) {
0af62b01
AD		 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;
9b8253ce 263 for (i = 0; i < ucode_size; i++)
0af62b01
AD		 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 */
0e2c978e
AD		 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 }
0af62b01
AD		 277
278 if (running)
279 WREG32(MC_SHARED_BLACKOUT_CNTL, blackout);
280 }
281
282 return 0;
283}
284
285int 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";
9b8253ce
AD		 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;
0af62b01
AD		 311 break;
312 case CHIP_TURKS:
313 chip_name = "TURKS";
314 rlc_chip_name = "BTC";
9b8253ce
AD		 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;
0af62b01
AD		 319 break;
320 case CHIP_CAICOS:
321 chip_name = "CAICOS";
322 rlc_chip_name = "BTC";
9b8253ce
AD		 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;
0af62b01 335 break;
c420c745
AD		 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;
0af62b01
AD		 345 default: BUG();
346 }
347
0af62b01
AD		 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
c420c745
AD		 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 }
0af62b01
AD		 396 }
397out:
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
fecf1d07
AD		 417/*
418 * Core functions
419 */
fecf1d07
AD		 420static void cayman_gpu_init(struct radeon_device *rdev)
421{
fecf1d07
AD		 422 u32 gb_addr_config = 0;
423 u32 mc_shared_chmap, mc_arb_ramcfg;
fecf1d07
AD		 424 u32 cgts_tcc_disable;
425 u32 sx_debug_1;
426 u32 smx_dc_ctl0;
fecf1d07
AD		 427 u32 cgts_sm_ctrl_reg;
428 u32 hdp_host_path_cntl;
429 u32 tmp;
416a2bd2 430 u32 disabled_rb_mask;
fecf1d07
AD		 431 int i, j;
432
433 switch (rdev->family) {
434 case CHIP_CAYMAN:
fecf1d07
AD		 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;
416a2bd2 455 gb_addr_config = CAYMAN_GB_ADDR_CONFIG_GOLDEN;
fecf1d07 456 break;
7b76e479
AD		 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) ||
d430f7db
AD		 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)) {
7b76e479
AD		 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) ||
d430f7db
AD		 474 (rdev->pdev->device == 0x9904) ||
475 (rdev->pdev->device == 0x990A) ||
476 (rdev->pdev->device == 0x9913) ||
477 (rdev->pdev->device == 0x9918)) {
7b76e479
AD		 478 rdev->config.cayman.max_simds_per_se = 4;
479 rdev->config.cayman.max_backends_per_se = 2;
d430f7db
AD		 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)) {
7b76e479
AD		 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;
416a2bd2 506 gb_addr_config = ARUBA_GB_ADDR_CONFIG_GOLDEN;
7b76e479 507 break;
fecf1d07
AD		 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
d054ac16
AD		 521 evergreen_fix_pci_max_read_req_size(rdev);
522
fecf1d07
AD		 523 mc_shared_chmap = RREG32(MC_SHARED_CHMAP);
524 mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG);
525
fecf1d07
AD		 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
fecf1d07
AD		 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
416a2bd2 548
fecf1d07
AD		 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 }
7b76e479
AD		 572
573 /* num banks is 8 on all fusion asics. 0 = 4, 1 = 8, 2 = 16 */
574 if (rdev->flags & RADEON_IS_IGP)
1f73cca7 575 rdev->config.cayman.tile_config |= 1 << 4;
29d65406
AD		 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 }
fecf1d07 582 rdev->config.cayman.tile_config |=
cde5083b 583 ((gb_addr_config & PIPE_INTERLEAVE_SIZE_MASK) >> PIPE_INTERLEAVE_SIZE_SHIFT) << 8;
fecf1d07
AD		 584 rdev->config.cayman.tile_config |=
585 ((gb_addr_config & ROW_SIZE_MASK) >> ROW_SIZE_SHIFT) << 12;
586
416a2bd2
AD		 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
fecf1d07
AD		 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
416a2bd2
AD		 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);
fecf1d07 613
416a2bd2
AD		 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));
fecf1d07
AD		 617 WREG32(CGTS_TCC_DISABLE, cgts_tcc_disable);
618 WREG32(CGTS_SYS_TCC_DISABLE, cgts_tcc_disable);
fecf1d07
AD		 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);
285e042d 637 smx_dc_ctl0 |= NUMBER_OF_SETS(rdev->config.cayman.sx_num_of_sets);
fecf1d07
AD		 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
285e042d
DA		 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)));
fecf1d07 656
285e042d
DA		 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)));
fecf1d07
AD		 660
661
662 WREG32(VGT_NUM_INSTANCES, 1);
663
664 WREG32(CP_PERFMON_CNTL, 0);
665
285e042d 666 WREG32(SQ_MS_FIFO_SIZES, (CACHE_FIFO_SIZE(16 * rdev->config.cayman.sq_num_cf_insts) |
fecf1d07
AD		 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
0b65f83f
DA		 697 tmp = RREG32(HDP_MISC_CNTL);
698 tmp |= HDP_FLUSH_INVALIDATE_CACHE;
699 WREG32(HDP_MISC_CNTL, tmp);
700
fecf1d07
AD		 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
fa8198ea
AD		 709/*
710 * GART
711 */
712void 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
721int cayman_pcie_gart_enable(struct radeon_device *rdev)
722{
721604a1 723 int i, r;
fa8198ea 724
c9a1be96 725 if (rdev->gart.robj == NULL) {
fa8198ea
AD		 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 */
721604a1
JG		 734 WREG32(MC_VM_MX_L1_TLB_CNTL,
735 (0xA << 7) |
736 ENABLE_L1_TLB |
fa8198ea
AD		 737 ENABLE_L1_FRAGMENT_PROCESSING |
738 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
721604a1 739 ENABLE_ADVANCED_DRIVER_MODEL |
fa8198ea
AD		 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);
721604a1
JG		 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));
fa8198ea
AD		 775 WREG32(VM_CONTEXT1_CNTL2, 0);
776 WREG32(VM_CONTEXT1_CNTL, 0);
721604a1
JG		 777 WREG32(VM_CONTEXT1_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) |
778 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT);
fa8198ea
AD		 779
780 cayman_pcie_gart_tlb_flush(rdev);
fcf4de5a
TV		 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);
fa8198ea
AD		 784 rdev->gart.ready = true;
785 return 0;
786}
787
788void cayman_pcie_gart_disable(struct radeon_device *rdev)
789{
fa8198ea
AD		 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));
c9a1be96 805 radeon_gart_table_vram_unpin(rdev);
fa8198ea
AD		 806}
807
808void 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
1b37078b
AD		 815void 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
0c88a02e
AD		 824/*
825 * CP.
826 */
b40e7e16
AD		 827void 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
721604a1
JG		 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);
b40e7e16
AD		 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
721604a1
JG		 851void cayman_ring_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib)
852{
876dc9f3 853 struct radeon_ring *ring = &rdev->ring[ib->ring];
721604a1
JG		 854
855 /* set to DX10/11 mode */
856 radeon_ring_write(ring, PACKET3(PACKET3_MODE_CONTROL, 0));
857 radeon_ring_write(ring, 1);
45df6803
CK		 858
859 if (ring->rptr_save_reg) {
860 uint32_t next_rptr = ring->wptr + 3 + 4 + 8;
861 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
862 radeon_ring_write(ring, ((ring->rptr_save_reg -
863 PACKET3_SET_CONFIG_REG_START) >> 2));
864 radeon_ring_write(ring, next_rptr);
865 }
866
721604a1
JG		 867 radeon_ring_write(ring, PACKET3(PACKET3_INDIRECT_BUFFER, 2));
868 radeon_ring_write(ring,
869#ifdef __BIG_ENDIAN
870 (2 << 0) |
871#endif
872 (ib->gpu_addr & 0xFFFFFFFC));
873 radeon_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xFF);
874 radeon_ring_write(ring, ib->length_dw | (ib->vm_id << 24));
875
876 /* flush read cache over gart for this vmid */
877 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
878 radeon_ring_write(ring, (CP_COHER_CNTL2 - PACKET3_SET_CONFIG_REG_START) >> 2);
879 radeon_ring_write(ring, ib->vm_id);
880 radeon_ring_write(ring, PACKET3(PACKET3_SURFACE_SYNC, 3));
881 radeon_ring_write(ring, PACKET3_TC_ACTION_ENA | PACKET3_SH_ACTION_ENA);
882 radeon_ring_write(ring, 0xFFFFFFFF);
883 radeon_ring_write(ring, 0);
884 radeon_ring_write(ring, 10); /* poll interval */
885}
886
0c88a02e
AD		 887static void cayman_cp_enable(struct radeon_device *rdev, bool enable)
888{
889 if (enable)
890 WREG32(CP_ME_CNTL, 0);
891 else {
38f1cff0 892 radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
0c88a02e
AD		 893 WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT));
894 WREG32(SCRATCH_UMSK, 0);
895 }
896}
897
898static int cayman_cp_load_microcode(struct radeon_device *rdev)
899{
900 const __be32 *fw_data;
901 int i;
902
903 if (!rdev->me_fw || !rdev->pfp_fw)
904 return -EINVAL;
905
906 cayman_cp_enable(rdev, false);
907
908 fw_data = (const __be32 *)rdev->pfp_fw->data;
909 WREG32(CP_PFP_UCODE_ADDR, 0);
910 for (i = 0; i < CAYMAN_PFP_UCODE_SIZE; i++)
911 WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++));
912 WREG32(CP_PFP_UCODE_ADDR, 0);
913
914 fw_data = (const __be32 *)rdev->me_fw->data;
915 WREG32(CP_ME_RAM_WADDR, 0);
916 for (i = 0; i < CAYMAN_PM4_UCODE_SIZE; i++)
917 WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++));
918
919 WREG32(CP_PFP_UCODE_ADDR, 0);
920 WREG32(CP_ME_RAM_WADDR, 0);
921 WREG32(CP_ME_RAM_RADDR, 0);
922 return 0;
923}
924
925static int cayman_cp_start(struct radeon_device *rdev)
926{
e32eb50d 927 struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
0c88a02e
AD		 928 int r, i;
929
e32eb50d 930 r = radeon_ring_lock(rdev, ring, 7);
0c88a02e
AD		 931 if (r) {
932 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
933 return r;
934 }
e32eb50d
CK		 935 radeon_ring_write(ring, PACKET3(PACKET3_ME_INITIALIZE, 5));
936 radeon_ring_write(ring, 0x1);
937 radeon_ring_write(ring, 0x0);
938 radeon_ring_write(ring, rdev->config.cayman.max_hw_contexts - 1);
939 radeon_ring_write(ring, PACKET3_ME_INITIALIZE_DEVICE_ID(1));
940 radeon_ring_write(ring, 0);
941 radeon_ring_write(ring, 0);
942 radeon_ring_unlock_commit(rdev, ring);
0c88a02e
AD		 943
944 cayman_cp_enable(rdev, true);
945
e32eb50d 946 r = radeon_ring_lock(rdev, ring, cayman_default_size + 19);
0c88a02e
AD		 947 if (r) {
948 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
949 return r;
950 }
951
952 /* setup clear context state */
e32eb50d
CK		 953 radeon_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
954 radeon_ring_write(ring, PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
0c88a02e
AD		 955
956 for (i = 0; i < cayman_default_size; i++)
e32eb50d 957 radeon_ring_write(ring, cayman_default_state[i]);
0c88a02e 958
e32eb50d
CK		 959 radeon_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
960 radeon_ring_write(ring, PACKET3_PREAMBLE_END_CLEAR_STATE);
0c88a02e
AD		 961
962 /* set clear context state */
e32eb50d
CK		 963 radeon_ring_write(ring, PACKET3(PACKET3_CLEAR_STATE, 0));
964 radeon_ring_write(ring, 0);
0c88a02e
AD		 965
966 /* SQ_VTX_BASE_VTX_LOC */
e32eb50d
CK		 967 radeon_ring_write(ring, 0xc0026f00);
968 radeon_ring_write(ring, 0x00000000);
969 radeon_ring_write(ring, 0x00000000);
970 radeon_ring_write(ring, 0x00000000);
0c88a02e
AD		 971
972 /* Clear consts */
e32eb50d
CK		 973 radeon_ring_write(ring, 0xc0036f00);
974 radeon_ring_write(ring, 0x00000bc4);
975 radeon_ring_write(ring, 0xffffffff);
976 radeon_ring_write(ring, 0xffffffff);
977 radeon_ring_write(ring, 0xffffffff);
0c88a02e 978
e32eb50d
CK		 979 radeon_ring_write(ring, 0xc0026900);
980 radeon_ring_write(ring, 0x00000316);
981 radeon_ring_write(ring, 0x0000000e); /* VGT_VERTEX_REUSE_BLOCK_CNTL */
982 radeon_ring_write(ring, 0x00000010); /* */
9b91d18d 983
e32eb50d 984 radeon_ring_unlock_commit(rdev, ring);
0c88a02e
AD		 985
986 /* XXX init other rings */
987
988 return 0;
989}
990
755d819e
AD		 991static void cayman_cp_fini(struct radeon_device *rdev)
992{
45df6803 993 struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
755d819e 994 cayman_cp_enable(rdev, false);
45df6803
CK		 995 radeon_ring_fini(rdev, ring);
996 radeon_scratch_free(rdev, ring->rptr_save_reg);
755d819e
AD		 997}
998
0c88a02e
AD		 999int cayman_cp_resume(struct radeon_device *rdev)
1000{
b90ca986
CK		 1001 static const int ridx[] = {
1002 RADEON_RING_TYPE_GFX_INDEX,
1003 CAYMAN_RING_TYPE_CP1_INDEX,
1004 CAYMAN_RING_TYPE_CP2_INDEX
1005 };
1006 static const unsigned cp_rb_cntl[] = {
1007 CP_RB0_CNTL,
1008 CP_RB1_CNTL,
1009 CP_RB2_CNTL,
1010 };
1011 static const unsigned cp_rb_rptr_addr[] = {
1012 CP_RB0_RPTR_ADDR,
1013 CP_RB1_RPTR_ADDR,
1014 CP_RB2_RPTR_ADDR
1015 };
1016 static const unsigned cp_rb_rptr_addr_hi[] = {
1017 CP_RB0_RPTR_ADDR_HI,
1018 CP_RB1_RPTR_ADDR_HI,
1019 CP_RB2_RPTR_ADDR_HI
1020 };
1021 static const unsigned cp_rb_base[] = {
1022 CP_RB0_BASE,
1023 CP_RB1_BASE,
1024 CP_RB2_BASE
1025 };
e32eb50d 1026 struct radeon_ring *ring;
b90ca986 1027 int i, r;
0c88a02e
AD		 1028
1029 /* Reset cp; if cp is reset, then PA, SH, VGT also need to be reset */
1030 WREG32(GRBM_SOFT_RESET, (SOFT_RESET_CP |
1031 SOFT_RESET_PA |
1032 SOFT_RESET_SH |
1033 SOFT_RESET_VGT |
a49a50da 1034 SOFT_RESET_SPI |
0c88a02e
AD		 1035 SOFT_RESET_SX));
1036 RREG32(GRBM_SOFT_RESET);
1037 mdelay(15);
1038 WREG32(GRBM_SOFT_RESET, 0);
1039 RREG32(GRBM_SOFT_RESET);
1040
15d3332f 1041 WREG32(CP_SEM_WAIT_TIMER, 0x0);
11ef3f1f 1042 WREG32(CP_SEM_INCOMPLETE_TIMER_CNTL, 0x0);
0c88a02e
AD		 1043
1044 /* Set the write pointer delay */
1045 WREG32(CP_RB_WPTR_DELAY, 0);
1046
1047 WREG32(CP_DEBUG, (1 << 27));
1048
0c88a02e 1049 /* set the wb address wether it's enabled or not */
0c88a02e 1050 WREG32(SCRATCH_ADDR, ((rdev->wb.gpu_addr + RADEON_WB_SCRATCH_OFFSET) >> 8) & 0xFFFFFFFF);
b90ca986 1051 WREG32(SCRATCH_UMSK, 0xff);
0c88a02e 1052
b90ca986
CK		 1053 for (i = 0; i < 3; ++i) {
1054 uint32_t rb_cntl;
1055 uint64_t addr;
0c88a02e 1056
b90ca986
CK		 1057 /* Set ring buffer size */
1058 ring = &rdev->ring[ridx[i]];
1059 rb_cntl = drm_order(ring->ring_size / 8);
1060 rb_cntl |= drm_order(RADEON_GPU_PAGE_SIZE/8) << 8;
0c88a02e 1061#ifdef __BIG_ENDIAN
b90ca986 1062 rb_cntl |= BUF_SWAP_32BIT;
0c88a02e 1063#endif
b90ca986 1064 WREG32(cp_rb_cntl[i], rb_cntl);
0c88a02e 1065
b90ca986
CK		 1066 /* set the wb address wether it's enabled or not */
1067 addr = rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET;
1068 WREG32(cp_rb_rptr_addr[i], addr & 0xFFFFFFFC);
1069 WREG32(cp_rb_rptr_addr_hi[i], upper_32_bits(addr) & 0xFF);
1070 }
0c88a02e 1071
b90ca986
CK		 1072 /* set the rb base addr, this causes an internal reset of ALL rings */
1073 for (i = 0; i < 3; ++i) {
1074 ring = &rdev->ring[ridx[i]];
1075 WREG32(cp_rb_base[i], ring->gpu_addr >> 8);
1076 }
0c88a02e 1077
b90ca986
CK		 1078 for (i = 0; i < 3; ++i) {
1079 /* Initialize the ring buffer's read and write pointers */
1080 ring = &rdev->ring[ridx[i]];
1081 WREG32_P(cp_rb_cntl[i], RB_RPTR_WR_ENA, ~RB_RPTR_WR_ENA);
0c88a02e 1082
b90ca986
CK		 1083 ring->rptr = ring->wptr = 0;
1084 WREG32(ring->rptr_reg, ring->rptr);
1085 WREG32(ring->wptr_reg, ring->wptr);
0c88a02e 1086
b90ca986
CK		 1087 mdelay(1);
1088 WREG32_P(cp_rb_cntl[i], 0, ~RB_RPTR_WR_ENA);
1089 }
0c88a02e
AD		 1090
1091 /* start the rings */
1092 cayman_cp_start(rdev);
e32eb50d
CK		 1093 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = true;
1094 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
1095 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
0c88a02e 1096 /* this only test cp0 */
f712812e 1097 r = radeon_ring_test(rdev, RADEON_RING_TYPE_GFX_INDEX, &rdev->ring[RADEON_RING_TYPE_GFX_INDEX]);
0c88a02e 1098 if (r) {
e32eb50d
CK		 1099 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
1100 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
1101 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
0c88a02e
AD		 1102 return r;
1103 }
1104
1105 return 0;
1106}
1107
b9952a8a
AD		 1108static int cayman_gpu_soft_reset(struct radeon_device *rdev)
1109{
1110 struct evergreen_mc_save save;
1111 u32 grbm_reset = 0;
1112
1113 if (!(RREG32(GRBM_STATUS) & GUI_ACTIVE))
1114 return 0;
1115
1116 dev_info(rdev->dev, "GPU softreset \n");
1117 dev_info(rdev->dev, " GRBM_STATUS=0x%08X\n",
1118 RREG32(GRBM_STATUS));
1119 dev_info(rdev->dev, " GRBM_STATUS_SE0=0x%08X\n",
1120 RREG32(GRBM_STATUS_SE0));
1121 dev_info(rdev->dev, " GRBM_STATUS_SE1=0x%08X\n",
1122 RREG32(GRBM_STATUS_SE1));
1123 dev_info(rdev->dev, " SRBM_STATUS=0x%08X\n",
1124 RREG32(SRBM_STATUS));
721604a1
JG		 1125 dev_info(rdev->dev, " VM_CONTEXT0_PROTECTION_FAULT_ADDR 0x%08X\n",
1126 RREG32(0x14F8));
1127 dev_info(rdev->dev, " VM_CONTEXT0_PROTECTION_FAULT_STATUS 0x%08X\n",
1128 RREG32(0x14D8));
1129 dev_info(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n",
1130 RREG32(0x14FC));
1131 dev_info(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
1132 RREG32(0x14DC));
1133
b9952a8a
AD		 1134 evergreen_mc_stop(rdev, &save);
1135 if (evergreen_mc_wait_for_idle(rdev)) {
1136 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
1137 }
1138 /* Disable CP parsing/prefetching */
1139 WREG32(CP_ME_CNTL, CP_ME_HALT | CP_PFP_HALT);
1140
1141 /* reset all the gfx blocks */
1142 grbm_reset = (SOFT_RESET_CP |
1143 SOFT_RESET_CB |
1144 SOFT_RESET_DB |
1145 SOFT_RESET_GDS |
1146 SOFT_RESET_PA |
1147 SOFT_RESET_SC |
1148 SOFT_RESET_SPI |
1149 SOFT_RESET_SH |
1150 SOFT_RESET_SX |
1151 SOFT_RESET_TC |
1152 SOFT_RESET_TA |
1153 SOFT_RESET_VGT |
1154 SOFT_RESET_IA);
1155
1156 dev_info(rdev->dev, " GRBM_SOFT_RESET=0x%08X\n", grbm_reset);
1157 WREG32(GRBM_SOFT_RESET, grbm_reset);
1158 (void)RREG32(GRBM_SOFT_RESET);
1159 udelay(50);
1160 WREG32(GRBM_SOFT_RESET, 0);
1161 (void)RREG32(GRBM_SOFT_RESET);
1162 /* Wait a little for things to settle down */
1163 udelay(50);
721604a1 1164
b9952a8a
AD		 1165 dev_info(rdev->dev, " GRBM_STATUS=0x%08X\n",
1166 RREG32(GRBM_STATUS));
1167 dev_info(rdev->dev, " GRBM_STATUS_SE0=0x%08X\n",
1168 RREG32(GRBM_STATUS_SE0));
1169 dev_info(rdev->dev, " GRBM_STATUS_SE1=0x%08X\n",
1170 RREG32(GRBM_STATUS_SE1));
1171 dev_info(rdev->dev, " SRBM_STATUS=0x%08X\n",
1172 RREG32(SRBM_STATUS));
1173 evergreen_mc_resume(rdev, &save);
1174 return 0;
1175}
1176
1177int cayman_asic_reset(struct radeon_device *rdev)
1178{
1179 return cayman_gpu_soft_reset(rdev);
1180}
1181
755d819e
AD		 1182static int cayman_startup(struct radeon_device *rdev)
1183{
e32eb50d 1184 struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
755d819e
AD		 1185 int r;
1186
b07759bf
IH		 1187 /* enable pcie gen2 link */
1188 evergreen_pcie_gen2_enable(rdev);
1189
c420c745
AD		 1190 if (rdev->flags & RADEON_IS_IGP) {
1191 if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw) {
1192 r = ni_init_microcode(rdev);
1193 if (r) {
1194 DRM_ERROR("Failed to load firmware!\n");
1195 return r;
1196 }
1197 }
1198 } else {
1199 if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw || !rdev->mc_fw) {
1200 r = ni_init_microcode(rdev);
1201 if (r) {
1202 DRM_ERROR("Failed to load firmware!\n");
1203 return r;
1204 }
1205 }
1206
1207 r = ni_mc_load_microcode(rdev);
755d819e 1208 if (r) {
c420c745 1209 DRM_ERROR("Failed to load MC firmware!\n");
755d819e
AD		 1210 return r;
1211 }
1212 }
755d819e 1213
16cdf04d
AD		 1214 r = r600_vram_scratch_init(rdev);
1215 if (r)
1216 return r;
1217
755d819e
AD		 1218 evergreen_mc_program(rdev);
1219 r = cayman_pcie_gart_enable(rdev);
1220 if (r)
1221 return r;
1222 cayman_gpu_init(rdev);
1223
cb92d452 1224 r = evergreen_blit_init(rdev);
755d819e 1225 if (r) {
fb3d9e97 1226 r600_blit_fini(rdev);
27cd7769 1227 rdev->asic->copy.copy = NULL;
755d819e
AD		 1228 dev_warn(rdev->dev, "failed blitter (%d) falling back to memcpy\n", r);
1229 }
755d819e 1230
c420c745
AD		 1231 /* allocate rlc buffers */
1232 if (rdev->flags & RADEON_IS_IGP) {
1233 r = si_rlc_init(rdev);
1234 if (r) {
1235 DRM_ERROR("Failed to init rlc BOs!\n");
1236 return r;
1237 }
1238 }
1239
755d819e
AD		 1240 /* allocate wb buffer */
1241 r = radeon_wb_init(rdev);
1242 if (r)
1243 return r;
1244
30eb77f4
JG		 1245 r = radeon_fence_driver_start_ring(rdev, RADEON_RING_TYPE_GFX_INDEX);
1246 if (r) {
1247 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
1248 return r;
1249 }
1250
1251 r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
1252 if (r) {
1253 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
1254 return r;
1255 }
1256
1257 r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
1258 if (r) {
1259 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
1260 return r;
1261 }
1262
755d819e
AD		 1263 /* Enable IRQ */
1264 r = r600_irq_init(rdev);
1265 if (r) {
1266 DRM_ERROR("radeon: IH init failed (%d).\n", r);
1267 radeon_irq_kms_fini(rdev);
1268 return r;
1269 }
1270 evergreen_irq_set(rdev);
1271
e32eb50d 1272 r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP_RPTR_OFFSET,
78c5560a
AD		 1273 CP_RB0_RPTR, CP_RB0_WPTR,
1274 0, 0xfffff, RADEON_CP_PACKET2);
755d819e
AD		 1275 if (r)
1276 return r;
1277 r = cayman_cp_load_microcode(rdev);
1278 if (r)
1279 return r;
1280 r = cayman_cp_resume(rdev);
1281 if (r)
1282 return r;
1283
2898c348
CK		 1284 r = radeon_ib_pool_init(rdev);
1285 if (r) {
1286 dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
b15ba512 1287 return r;
2898c348 1288 }
b15ba512 1289
c6105f24
CK		 1290 r = radeon_vm_manager_init(rdev);
1291 if (r) {
1292 dev_err(rdev->dev, "vm manager initialization failed (%d).\n", r);
721604a1 1293 return r;
c6105f24 1294 }
721604a1 1295
6b53a050
RM		 1296 r = r600_audio_init(rdev);
1297 if (r)
1298 return r;
1299
755d819e
AD		 1300 return 0;
1301}
1302
1303int cayman_resume(struct radeon_device *rdev)
1304{
1305 int r;
1306
1307 /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw,
1308 * posting will perform necessary task to bring back GPU into good
1309 * shape.
1310 */
1311 /* post card */
1312 atom_asic_init(rdev->mode_info.atom_context);
1313
b15ba512 1314 rdev->accel_working = true;
755d819e
AD		 1315 r = cayman_startup(rdev);
1316 if (r) {
1317 DRM_ERROR("cayman startup failed on resume\n");
6b7746e8 1318 rdev->accel_working = false;
755d819e
AD		 1319 return r;
1320 }
755d819e 1321 return r;
755d819e
AD		 1322}
1323
1324int cayman_suspend(struct radeon_device *rdev)
1325{
6b53a050 1326 r600_audio_fini(rdev);
755d819e 1327 cayman_cp_enable(rdev, false);
e32eb50d 1328 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
755d819e
AD		 1329 evergreen_irq_suspend(rdev);
1330 radeon_wb_disable(rdev);
1331 cayman_pcie_gart_disable(rdev);
755d819e
AD		 1332 return 0;
1333}
1334
1335/* Plan is to move initialization in that function and use
1336 * helper function so that radeon_device_init pretty much
1337 * do nothing more than calling asic specific function. This
1338 * should also allow to remove a bunch of callback function
1339 * like vram_info.
1340 */
1341int cayman_init(struct radeon_device *rdev)
1342{
e32eb50d 1343 struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
755d819e
AD		 1344 int r;
1345
755d819e
AD		 1346 /* Read BIOS */
1347 if (!radeon_get_bios(rdev)) {
1348 if (ASIC_IS_AVIVO(rdev))
1349 return -EINVAL;
1350 }
1351 /* Must be an ATOMBIOS */
1352 if (!rdev->is_atom_bios) {
1353 dev_err(rdev->dev, "Expecting atombios for cayman GPU\n");
1354 return -EINVAL;
1355 }
1356 r = radeon_atombios_init(rdev);
1357 if (r)
1358 return r;
1359
1360 /* Post card if necessary */
1361 if (!radeon_card_posted(rdev)) {
1362 if (!rdev->bios) {
1363 dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n");
1364 return -EINVAL;
1365 }
1366 DRM_INFO("GPU not posted. posting now...\n");
1367 atom_asic_init(rdev->mode_info.atom_context);
1368 }
1369 /* Initialize scratch registers */
1370 r600_scratch_init(rdev);
1371 /* Initialize surface registers */
1372 radeon_surface_init(rdev);
1373 /* Initialize clocks */
1374 radeon_get_clock_info(rdev->ddev);
1375 /* Fence driver */
30eb77f4 1376 r = radeon_fence_driver_init(rdev);
755d819e
AD		 1377 if (r)
1378 return r;
1379 /* initialize memory controller */
1380 r = evergreen_mc_init(rdev);
1381 if (r)
1382 return r;
1383 /* Memory manager */
1384 r = radeon_bo_init(rdev);
1385 if (r)
1386 return r;
1387
1388 r = radeon_irq_kms_init(rdev);
1389 if (r)
1390 return r;
1391
e32eb50d
CK		 1392 ring->ring_obj = NULL;
1393 r600_ring_init(rdev, ring, 1024 * 1024);
755d819e
AD		 1394
1395 rdev->ih.ring_obj = NULL;
1396 r600_ih_ring_init(rdev, 64 * 1024);
1397
1398 r = r600_pcie_gart_init(rdev);
1399 if (r)
1400 return r;
1401
1402 rdev->accel_working = true;
1403 r = cayman_startup(rdev);
1404 if (r) {
1405 dev_err(rdev->dev, "disabling GPU acceleration\n");
1406 cayman_cp_fini(rdev);
1407 r600_irq_fini(rdev);
c420c745
AD		 1408 if (rdev->flags & RADEON_IS_IGP)
1409 si_rlc_fini(rdev);
755d819e 1410 radeon_wb_fini(rdev);
2898c348 1411 radeon_ib_pool_fini(rdev);
721604a1 1412 radeon_vm_manager_fini(rdev);
755d819e
AD		 1413 radeon_irq_kms_fini(rdev);
1414 cayman_pcie_gart_fini(rdev);
1415 rdev->accel_working = false;
1416 }
755d819e
AD		 1417
1418 /* Don't start up if the MC ucode is missing.
1419 * The default clocks and voltages before the MC ucode
1420 * is loaded are not suffient for advanced operations.
c420c745
AD		 1421 *
1422 * We can skip this check for TN, because there is no MC
1423 * ucode.
755d819e 1424 */
c420c745 1425 if (!rdev->mc_fw && !(rdev->flags & RADEON_IS_IGP)) {
755d819e
AD		 1426 DRM_ERROR("radeon: MC ucode required for NI+.\n");
1427 return -EINVAL;
1428 }
1429
1430 return 0;
1431}
1432
1433void cayman_fini(struct radeon_device *rdev)
1434{
fb3d9e97 1435 r600_blit_fini(rdev);
755d819e
AD		 1436 cayman_cp_fini(rdev);
1437 r600_irq_fini(rdev);
c420c745
AD		 1438 if (rdev->flags & RADEON_IS_IGP)
1439 si_rlc_fini(rdev);
755d819e 1440 radeon_wb_fini(rdev);
721604a1 1441 radeon_vm_manager_fini(rdev);
2898c348 1442 radeon_ib_pool_fini(rdev);
755d819e
AD		 1443 radeon_irq_kms_fini(rdev);
1444 cayman_pcie_gart_fini(rdev);
16cdf04d 1445 r600_vram_scratch_fini(rdev);
755d819e
AD		 1446 radeon_gem_fini(rdev);
1447 radeon_fence_driver_fini(rdev);
1448 radeon_bo_fini(rdev);
1449 radeon_atombios_fini(rdev);
1450 kfree(rdev->bios);
1451 rdev->bios = NULL;
1452}
1453
721604a1
JG		 1454/*
1455 * vm
1456 */
1457int cayman_vm_init(struct radeon_device *rdev)
1458{
1459 /* number of VMs */
1460 rdev->vm_manager.nvm = 8;
1461 /* base offset of vram pages */
e71270fd
AD		 1462 if (rdev->flags & RADEON_IS_IGP) {
1463 u64 tmp = RREG32(FUS_MC_VM_FB_OFFSET);
1464 tmp <<= 22;
1465 rdev->vm_manager.vram_base_offset = tmp;
1466 } else
1467 rdev->vm_manager.vram_base_offset = 0;
721604a1
JG		 1468 return 0;
1469}
1470
1471void cayman_vm_fini(struct radeon_device *rdev)
1472{
1473}
1474
1475int cayman_vm_bind(struct radeon_device *rdev, struct radeon_vm *vm, int id)
1476{
1477 WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR + (id << 2), 0);
1478 WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR + (id << 2), vm->last_pfn);
1479 WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (id << 2), vm->pt_gpu_addr >> 12);
1480 /* flush hdp cache */
1481 WREG32(HDP_MEM_COHERENCY_FLUSH_CNTL, 0x1);
1482 /* bits 0-7 are the VM contexts0-7 */
1483 WREG32(VM_INVALIDATE_REQUEST, 1 << id);
1484 return 0;
1485}
1486
1487void cayman_vm_unbind(struct radeon_device *rdev, struct radeon_vm *vm)
1488{
1489 WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR + (vm->id << 2), 0);
1490 WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR + (vm->id << 2), 0);
1491 WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm->id << 2), 0);
1492 /* flush hdp cache */
1493 WREG32(HDP_MEM_COHERENCY_FLUSH_CNTL, 0x1);
1494 /* bits 0-7 are the VM contexts0-7 */
1495 WREG32(VM_INVALIDATE_REQUEST, 1 << vm->id);
1496}
1497
1498void cayman_vm_tlb_flush(struct radeon_device *rdev, struct radeon_vm *vm)
1499{
1500 if (vm->id == -1)
1501 return;
1502
1503 /* flush hdp cache */
1504 WREG32(HDP_MEM_COHERENCY_FLUSH_CNTL, 0x1);
1505 /* bits 0-7 are the VM contexts0-7 */
1506 WREG32(VM_INVALIDATE_REQUEST, 1 << vm->id);
1507}
1508
1509#define R600_PTE_VALID (1 << 0)
1510#define R600_PTE_SYSTEM (1 << 1)
1511#define R600_PTE_SNOOPED (1 << 2)
1512#define R600_PTE_READABLE (1 << 5)
1513#define R600_PTE_WRITEABLE (1 << 6)
1514
1515uint32_t cayman_vm_page_flags(struct radeon_device *rdev,
1516 struct radeon_vm *vm,
1517 uint32_t flags)
1518{
1519 uint32_t r600_flags = 0;
1520
1521 r600_flags |= (flags & RADEON_VM_PAGE_VALID) ? R600_PTE_VALID : 0;
1522 r600_flags |= (flags & RADEON_VM_PAGE_READABLE) ? R600_PTE_READABLE : 0;
1523 r600_flags |= (flags & RADEON_VM_PAGE_WRITEABLE) ? R600_PTE_WRITEABLE : 0;
1524 if (flags & RADEON_VM_PAGE_SYSTEM) {
1525 r600_flags |= R600_PTE_SYSTEM;
1526 r600_flags |= (flags & RADEON_VM_PAGE_SNOOPED) ? R600_PTE_SNOOPED : 0;
1527 }
1528 return r600_flags;
1529}
1530
1531void cayman_vm_set_page(struct radeon_device *rdev, struct radeon_vm *vm,
1532 unsigned pfn, uint64_t addr, uint32_t flags)
1533{
1534 void __iomem *ptr = (void *)vm->pt;
1535
1536 addr = addr & 0xFFFFFFFFFFFFF000ULL;
1537 addr |= flags;
1538 writeq(addr, ptr + (pfn * 8));
1539}
Linux 6.x block layer and io_uring tree(s)