drm/radeon: enable mgcg on SI
[linux-2.6-block.git] / drivers / gpu / drm / radeon / si.c
CommitLineData
43b3cd99
AD
1/*
2 * Copyright 2011 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 */
0f0de06c 24#include <linux/firmware.h>
0f0de06c
AD
25#include <linux/slab.h>
26#include <linux/module.h>
760285e7 27#include <drm/drmP.h>
43b3cd99
AD
28#include "radeon.h"
29#include "radeon_asic.h"
760285e7 30#include <drm/radeon_drm.h>
43b3cd99
AD
31#include "sid.h"
32#include "atom.h"
48c0c902 33#include "si_blit_shaders.h"
bd8cd539 34#include "clearstate_si.h"
a0ceada6 35#include "radeon_ucode.h"
43b3cd99 36
0f0de06c
AD
37
38MODULE_FIRMWARE("radeon/TAHITI_pfp.bin");
39MODULE_FIRMWARE("radeon/TAHITI_me.bin");
40MODULE_FIRMWARE("radeon/TAHITI_ce.bin");
41MODULE_FIRMWARE("radeon/TAHITI_mc.bin");
42MODULE_FIRMWARE("radeon/TAHITI_rlc.bin");
a9e61410 43MODULE_FIRMWARE("radeon/TAHITI_smc.bin");
0f0de06c
AD
44MODULE_FIRMWARE("radeon/PITCAIRN_pfp.bin");
45MODULE_FIRMWARE("radeon/PITCAIRN_me.bin");
46MODULE_FIRMWARE("radeon/PITCAIRN_ce.bin");
47MODULE_FIRMWARE("radeon/PITCAIRN_mc.bin");
48MODULE_FIRMWARE("radeon/PITCAIRN_rlc.bin");
a9e61410 49MODULE_FIRMWARE("radeon/PITCAIRN_smc.bin");
0f0de06c
AD
50MODULE_FIRMWARE("radeon/VERDE_pfp.bin");
51MODULE_FIRMWARE("radeon/VERDE_me.bin");
52MODULE_FIRMWARE("radeon/VERDE_ce.bin");
53MODULE_FIRMWARE("radeon/VERDE_mc.bin");
54MODULE_FIRMWARE("radeon/VERDE_rlc.bin");
a9e61410 55MODULE_FIRMWARE("radeon/VERDE_smc.bin");
bcc7f5d2
AD
56MODULE_FIRMWARE("radeon/OLAND_pfp.bin");
57MODULE_FIRMWARE("radeon/OLAND_me.bin");
58MODULE_FIRMWARE("radeon/OLAND_ce.bin");
59MODULE_FIRMWARE("radeon/OLAND_mc.bin");
60MODULE_FIRMWARE("radeon/OLAND_rlc.bin");
a9e61410 61MODULE_FIRMWARE("radeon/OLAND_smc.bin");
c04c00b4
AD
62MODULE_FIRMWARE("radeon/HAINAN_pfp.bin");
63MODULE_FIRMWARE("radeon/HAINAN_me.bin");
64MODULE_FIRMWARE("radeon/HAINAN_ce.bin");
65MODULE_FIRMWARE("radeon/HAINAN_mc.bin");
66MODULE_FIRMWARE("radeon/HAINAN_rlc.bin");
a9e61410 67MODULE_FIRMWARE("radeon/HAINAN_smc.bin");
0f0de06c 68
b9d305df 69static void si_pcie_gen3_enable(struct radeon_device *rdev);
e0bcf165 70static void si_program_aspm(struct radeon_device *rdev);
1fd11777
AD
71extern void sumo_rlc_fini(struct radeon_device *rdev);
72extern int sumo_rlc_init(struct radeon_device *rdev);
25a857fb
AD
73extern int r600_ih_ring_alloc(struct radeon_device *rdev);
74extern void r600_ih_ring_fini(struct radeon_device *rdev);
0a96d72b 75extern void evergreen_fix_pci_max_read_req_size(struct radeon_device *rdev);
c476dde2
AD
76extern void evergreen_mc_stop(struct radeon_device *rdev, struct evergreen_mc_save *save);
77extern void evergreen_mc_resume(struct radeon_device *rdev, struct evergreen_mc_save *save);
ca7db22b 78extern u32 evergreen_get_number_of_dram_channels(struct radeon_device *rdev);
1c534671 79extern void evergreen_print_gpu_status_regs(struct radeon_device *rdev);
014bb209 80extern bool evergreen_is_display_hung(struct radeon_device *rdev);
2483b4ea
CK
81extern void si_dma_vm_set_page(struct radeon_device *rdev,
82 struct radeon_ib *ib,
83 uint64_t pe,
84 uint64_t addr, unsigned count,
85 uint32_t incr, uint32_t flags);
0a96d72b 86
6d8cf000
AD
87static const u32 verde_rlc_save_restore_register_list[] =
88{
89 (0x8000 << 16) | (0x98f4 >> 2),
90 0x00000000,
91 (0x8040 << 16) | (0x98f4 >> 2),
92 0x00000000,
93 (0x8000 << 16) | (0xe80 >> 2),
94 0x00000000,
95 (0x8040 << 16) | (0xe80 >> 2),
96 0x00000000,
97 (0x8000 << 16) | (0x89bc >> 2),
98 0x00000000,
99 (0x8040 << 16) | (0x89bc >> 2),
100 0x00000000,
101 (0x8000 << 16) | (0x8c1c >> 2),
102 0x00000000,
103 (0x8040 << 16) | (0x8c1c >> 2),
104 0x00000000,
105 (0x9c00 << 16) | (0x98f0 >> 2),
106 0x00000000,
107 (0x9c00 << 16) | (0xe7c >> 2),
108 0x00000000,
109 (0x8000 << 16) | (0x9148 >> 2),
110 0x00000000,
111 (0x8040 << 16) | (0x9148 >> 2),
112 0x00000000,
113 (0x9c00 << 16) | (0x9150 >> 2),
114 0x00000000,
115 (0x9c00 << 16) | (0x897c >> 2),
116 0x00000000,
117 (0x9c00 << 16) | (0x8d8c >> 2),
118 0x00000000,
119 (0x9c00 << 16) | (0xac54 >> 2),
120 0X00000000,
121 0x3,
122 (0x9c00 << 16) | (0x98f8 >> 2),
123 0x00000000,
124 (0x9c00 << 16) | (0x9910 >> 2),
125 0x00000000,
126 (0x9c00 << 16) | (0x9914 >> 2),
127 0x00000000,
128 (0x9c00 << 16) | (0x9918 >> 2),
129 0x00000000,
130 (0x9c00 << 16) | (0x991c >> 2),
131 0x00000000,
132 (0x9c00 << 16) | (0x9920 >> 2),
133 0x00000000,
134 (0x9c00 << 16) | (0x9924 >> 2),
135 0x00000000,
136 (0x9c00 << 16) | (0x9928 >> 2),
137 0x00000000,
138 (0x9c00 << 16) | (0x992c >> 2),
139 0x00000000,
140 (0x9c00 << 16) | (0x9930 >> 2),
141 0x00000000,
142 (0x9c00 << 16) | (0x9934 >> 2),
143 0x00000000,
144 (0x9c00 << 16) | (0x9938 >> 2),
145 0x00000000,
146 (0x9c00 << 16) | (0x993c >> 2),
147 0x00000000,
148 (0x9c00 << 16) | (0x9940 >> 2),
149 0x00000000,
150 (0x9c00 << 16) | (0x9944 >> 2),
151 0x00000000,
152 (0x9c00 << 16) | (0x9948 >> 2),
153 0x00000000,
154 (0x9c00 << 16) | (0x994c >> 2),
155 0x00000000,
156 (0x9c00 << 16) | (0x9950 >> 2),
157 0x00000000,
158 (0x9c00 << 16) | (0x9954 >> 2),
159 0x00000000,
160 (0x9c00 << 16) | (0x9958 >> 2),
161 0x00000000,
162 (0x9c00 << 16) | (0x995c >> 2),
163 0x00000000,
164 (0x9c00 << 16) | (0x9960 >> 2),
165 0x00000000,
166 (0x9c00 << 16) | (0x9964 >> 2),
167 0x00000000,
168 (0x9c00 << 16) | (0x9968 >> 2),
169 0x00000000,
170 (0x9c00 << 16) | (0x996c >> 2),
171 0x00000000,
172 (0x9c00 << 16) | (0x9970 >> 2),
173 0x00000000,
174 (0x9c00 << 16) | (0x9974 >> 2),
175 0x00000000,
176 (0x9c00 << 16) | (0x9978 >> 2),
177 0x00000000,
178 (0x9c00 << 16) | (0x997c >> 2),
179 0x00000000,
180 (0x9c00 << 16) | (0x9980 >> 2),
181 0x00000000,
182 (0x9c00 << 16) | (0x9984 >> 2),
183 0x00000000,
184 (0x9c00 << 16) | (0x9988 >> 2),
185 0x00000000,
186 (0x9c00 << 16) | (0x998c >> 2),
187 0x00000000,
188 (0x9c00 << 16) | (0x8c00 >> 2),
189 0x00000000,
190 (0x9c00 << 16) | (0x8c14 >> 2),
191 0x00000000,
192 (0x9c00 << 16) | (0x8c04 >> 2),
193 0x00000000,
194 (0x9c00 << 16) | (0x8c08 >> 2),
195 0x00000000,
196 (0x8000 << 16) | (0x9b7c >> 2),
197 0x00000000,
198 (0x8040 << 16) | (0x9b7c >> 2),
199 0x00000000,
200 (0x8000 << 16) | (0xe84 >> 2),
201 0x00000000,
202 (0x8040 << 16) | (0xe84 >> 2),
203 0x00000000,
204 (0x8000 << 16) | (0x89c0 >> 2),
205 0x00000000,
206 (0x8040 << 16) | (0x89c0 >> 2),
207 0x00000000,
208 (0x8000 << 16) | (0x914c >> 2),
209 0x00000000,
210 (0x8040 << 16) | (0x914c >> 2),
211 0x00000000,
212 (0x8000 << 16) | (0x8c20 >> 2),
213 0x00000000,
214 (0x8040 << 16) | (0x8c20 >> 2),
215 0x00000000,
216 (0x8000 << 16) | (0x9354 >> 2),
217 0x00000000,
218 (0x8040 << 16) | (0x9354 >> 2),
219 0x00000000,
220 (0x9c00 << 16) | (0x9060 >> 2),
221 0x00000000,
222 (0x9c00 << 16) | (0x9364 >> 2),
223 0x00000000,
224 (0x9c00 << 16) | (0x9100 >> 2),
225 0x00000000,
226 (0x9c00 << 16) | (0x913c >> 2),
227 0x00000000,
228 (0x8000 << 16) | (0x90e0 >> 2),
229 0x00000000,
230 (0x8000 << 16) | (0x90e4 >> 2),
231 0x00000000,
232 (0x8000 << 16) | (0x90e8 >> 2),
233 0x00000000,
234 (0x8040 << 16) | (0x90e0 >> 2),
235 0x00000000,
236 (0x8040 << 16) | (0x90e4 >> 2),
237 0x00000000,
238 (0x8040 << 16) | (0x90e8 >> 2),
239 0x00000000,
240 (0x9c00 << 16) | (0x8bcc >> 2),
241 0x00000000,
242 (0x9c00 << 16) | (0x8b24 >> 2),
243 0x00000000,
244 (0x9c00 << 16) | (0x88c4 >> 2),
245 0x00000000,
246 (0x9c00 << 16) | (0x8e50 >> 2),
247 0x00000000,
248 (0x9c00 << 16) | (0x8c0c >> 2),
249 0x00000000,
250 (0x9c00 << 16) | (0x8e58 >> 2),
251 0x00000000,
252 (0x9c00 << 16) | (0x8e5c >> 2),
253 0x00000000,
254 (0x9c00 << 16) | (0x9508 >> 2),
255 0x00000000,
256 (0x9c00 << 16) | (0x950c >> 2),
257 0x00000000,
258 (0x9c00 << 16) | (0x9494 >> 2),
259 0x00000000,
260 (0x9c00 << 16) | (0xac0c >> 2),
261 0x00000000,
262 (0x9c00 << 16) | (0xac10 >> 2),
263 0x00000000,
264 (0x9c00 << 16) | (0xac14 >> 2),
265 0x00000000,
266 (0x9c00 << 16) | (0xae00 >> 2),
267 0x00000000,
268 (0x9c00 << 16) | (0xac08 >> 2),
269 0x00000000,
270 (0x9c00 << 16) | (0x88d4 >> 2),
271 0x00000000,
272 (0x9c00 << 16) | (0x88c8 >> 2),
273 0x00000000,
274 (0x9c00 << 16) | (0x88cc >> 2),
275 0x00000000,
276 (0x9c00 << 16) | (0x89b0 >> 2),
277 0x00000000,
278 (0x9c00 << 16) | (0x8b10 >> 2),
279 0x00000000,
280 (0x9c00 << 16) | (0x8a14 >> 2),
281 0x00000000,
282 (0x9c00 << 16) | (0x9830 >> 2),
283 0x00000000,
284 (0x9c00 << 16) | (0x9834 >> 2),
285 0x00000000,
286 (0x9c00 << 16) | (0x9838 >> 2),
287 0x00000000,
288 (0x9c00 << 16) | (0x9a10 >> 2),
289 0x00000000,
290 (0x8000 << 16) | (0x9870 >> 2),
291 0x00000000,
292 (0x8000 << 16) | (0x9874 >> 2),
293 0x00000000,
294 (0x8001 << 16) | (0x9870 >> 2),
295 0x00000000,
296 (0x8001 << 16) | (0x9874 >> 2),
297 0x00000000,
298 (0x8040 << 16) | (0x9870 >> 2),
299 0x00000000,
300 (0x8040 << 16) | (0x9874 >> 2),
301 0x00000000,
302 (0x8041 << 16) | (0x9870 >> 2),
303 0x00000000,
304 (0x8041 << 16) | (0x9874 >> 2),
305 0x00000000,
306 0x00000000
307};
308
205996c0
AD
309static const u32 tahiti_golden_rlc_registers[] =
310{
311 0xc424, 0xffffffff, 0x00601005,
312 0xc47c, 0xffffffff, 0x10104040,
313 0xc488, 0xffffffff, 0x0100000a,
314 0xc314, 0xffffffff, 0x00000800,
315 0xc30c, 0xffffffff, 0x800000f4,
316 0xf4a8, 0xffffffff, 0x00000000
317};
318
319static const u32 tahiti_golden_registers[] =
320{
321 0x9a10, 0x00010000, 0x00018208,
322 0x9830, 0xffffffff, 0x00000000,
323 0x9834, 0xf00fffff, 0x00000400,
324 0x9838, 0x0002021c, 0x00020200,
325 0xc78, 0x00000080, 0x00000000,
326 0xd030, 0x000300c0, 0x00800040,
327 0xd830, 0x000300c0, 0x00800040,
328 0x5bb0, 0x000000f0, 0x00000070,
329 0x5bc0, 0x00200000, 0x50100000,
330 0x7030, 0x31000311, 0x00000011,
331 0x277c, 0x00000003, 0x000007ff,
332 0x240c, 0x000007ff, 0x00000000,
333 0x8a14, 0xf000001f, 0x00000007,
334 0x8b24, 0xffffffff, 0x00ffffff,
335 0x8b10, 0x0000ff0f, 0x00000000,
336 0x28a4c, 0x07ffffff, 0x4e000000,
337 0x28350, 0x3f3f3fff, 0x2a00126a,
338 0x30, 0x000000ff, 0x0040,
339 0x34, 0x00000040, 0x00004040,
340 0x9100, 0x07ffffff, 0x03000000,
341 0x8e88, 0x01ff1f3f, 0x00000000,
342 0x8e84, 0x01ff1f3f, 0x00000000,
343 0x9060, 0x0000007f, 0x00000020,
344 0x9508, 0x00010000, 0x00010000,
345 0xac14, 0x00000200, 0x000002fb,
346 0xac10, 0xffffffff, 0x0000543b,
347 0xac0c, 0xffffffff, 0xa9210876,
348 0x88d0, 0xffffffff, 0x000fff40,
349 0x88d4, 0x0000001f, 0x00000010,
350 0x1410, 0x20000000, 0x20fffed8,
351 0x15c0, 0x000c0fc0, 0x000c0400
352};
353
354static const u32 tahiti_golden_registers2[] =
355{
356 0xc64, 0x00000001, 0x00000001
357};
358
359static const u32 pitcairn_golden_rlc_registers[] =
360{
361 0xc424, 0xffffffff, 0x00601004,
362 0xc47c, 0xffffffff, 0x10102020,
363 0xc488, 0xffffffff, 0x01000020,
364 0xc314, 0xffffffff, 0x00000800,
365 0xc30c, 0xffffffff, 0x800000a4
366};
367
368static const u32 pitcairn_golden_registers[] =
369{
370 0x9a10, 0x00010000, 0x00018208,
371 0x9830, 0xffffffff, 0x00000000,
372 0x9834, 0xf00fffff, 0x00000400,
373 0x9838, 0x0002021c, 0x00020200,
374 0xc78, 0x00000080, 0x00000000,
375 0xd030, 0x000300c0, 0x00800040,
376 0xd830, 0x000300c0, 0x00800040,
377 0x5bb0, 0x000000f0, 0x00000070,
378 0x5bc0, 0x00200000, 0x50100000,
379 0x7030, 0x31000311, 0x00000011,
380 0x2ae4, 0x00073ffe, 0x000022a2,
381 0x240c, 0x000007ff, 0x00000000,
382 0x8a14, 0xf000001f, 0x00000007,
383 0x8b24, 0xffffffff, 0x00ffffff,
384 0x8b10, 0x0000ff0f, 0x00000000,
385 0x28a4c, 0x07ffffff, 0x4e000000,
386 0x28350, 0x3f3f3fff, 0x2a00126a,
387 0x30, 0x000000ff, 0x0040,
388 0x34, 0x00000040, 0x00004040,
389 0x9100, 0x07ffffff, 0x03000000,
390 0x9060, 0x0000007f, 0x00000020,
391 0x9508, 0x00010000, 0x00010000,
392 0xac14, 0x000003ff, 0x000000f7,
393 0xac10, 0xffffffff, 0x00000000,
394 0xac0c, 0xffffffff, 0x32761054,
395 0x88d4, 0x0000001f, 0x00000010,
396 0x15c0, 0x000c0fc0, 0x000c0400
397};
398
399static const u32 verde_golden_rlc_registers[] =
400{
401 0xc424, 0xffffffff, 0x033f1005,
402 0xc47c, 0xffffffff, 0x10808020,
403 0xc488, 0xffffffff, 0x00800008,
404 0xc314, 0xffffffff, 0x00001000,
405 0xc30c, 0xffffffff, 0x80010014
406};
407
408static const u32 verde_golden_registers[] =
409{
410 0x9a10, 0x00010000, 0x00018208,
411 0x9830, 0xffffffff, 0x00000000,
412 0x9834, 0xf00fffff, 0x00000400,
413 0x9838, 0x0002021c, 0x00020200,
414 0xc78, 0x00000080, 0x00000000,
415 0xd030, 0x000300c0, 0x00800040,
416 0xd030, 0x000300c0, 0x00800040,
417 0xd830, 0x000300c0, 0x00800040,
418 0xd830, 0x000300c0, 0x00800040,
419 0x5bb0, 0x000000f0, 0x00000070,
420 0x5bc0, 0x00200000, 0x50100000,
421 0x7030, 0x31000311, 0x00000011,
422 0x2ae4, 0x00073ffe, 0x000022a2,
423 0x2ae4, 0x00073ffe, 0x000022a2,
424 0x2ae4, 0x00073ffe, 0x000022a2,
425 0x240c, 0x000007ff, 0x00000000,
426 0x240c, 0x000007ff, 0x00000000,
427 0x240c, 0x000007ff, 0x00000000,
428 0x8a14, 0xf000001f, 0x00000007,
429 0x8a14, 0xf000001f, 0x00000007,
430 0x8a14, 0xf000001f, 0x00000007,
431 0x8b24, 0xffffffff, 0x00ffffff,
432 0x8b10, 0x0000ff0f, 0x00000000,
433 0x28a4c, 0x07ffffff, 0x4e000000,
434 0x28350, 0x3f3f3fff, 0x0000124a,
435 0x28350, 0x3f3f3fff, 0x0000124a,
436 0x28350, 0x3f3f3fff, 0x0000124a,
437 0x30, 0x000000ff, 0x0040,
438 0x34, 0x00000040, 0x00004040,
439 0x9100, 0x07ffffff, 0x03000000,
440 0x9100, 0x07ffffff, 0x03000000,
441 0x8e88, 0x01ff1f3f, 0x00000000,
442 0x8e88, 0x01ff1f3f, 0x00000000,
443 0x8e88, 0x01ff1f3f, 0x00000000,
444 0x8e84, 0x01ff1f3f, 0x00000000,
445 0x8e84, 0x01ff1f3f, 0x00000000,
446 0x8e84, 0x01ff1f3f, 0x00000000,
447 0x9060, 0x0000007f, 0x00000020,
448 0x9508, 0x00010000, 0x00010000,
449 0xac14, 0x000003ff, 0x00000003,
450 0xac14, 0x000003ff, 0x00000003,
451 0xac14, 0x000003ff, 0x00000003,
452 0xac10, 0xffffffff, 0x00000000,
453 0xac10, 0xffffffff, 0x00000000,
454 0xac10, 0xffffffff, 0x00000000,
455 0xac0c, 0xffffffff, 0x00001032,
456 0xac0c, 0xffffffff, 0x00001032,
457 0xac0c, 0xffffffff, 0x00001032,
458 0x88d4, 0x0000001f, 0x00000010,
459 0x88d4, 0x0000001f, 0x00000010,
460 0x88d4, 0x0000001f, 0x00000010,
461 0x15c0, 0x000c0fc0, 0x000c0400
462};
463
464static const u32 oland_golden_rlc_registers[] =
465{
466 0xc424, 0xffffffff, 0x00601005,
467 0xc47c, 0xffffffff, 0x10104040,
468 0xc488, 0xffffffff, 0x0100000a,
469 0xc314, 0xffffffff, 0x00000800,
470 0xc30c, 0xffffffff, 0x800000f4
471};
472
473static const u32 oland_golden_registers[] =
474{
475 0x9a10, 0x00010000, 0x00018208,
476 0x9830, 0xffffffff, 0x00000000,
477 0x9834, 0xf00fffff, 0x00000400,
478 0x9838, 0x0002021c, 0x00020200,
479 0xc78, 0x00000080, 0x00000000,
480 0xd030, 0x000300c0, 0x00800040,
481 0xd830, 0x000300c0, 0x00800040,
482 0x5bb0, 0x000000f0, 0x00000070,
483 0x5bc0, 0x00200000, 0x50100000,
484 0x7030, 0x31000311, 0x00000011,
485 0x2ae4, 0x00073ffe, 0x000022a2,
486 0x240c, 0x000007ff, 0x00000000,
487 0x8a14, 0xf000001f, 0x00000007,
488 0x8b24, 0xffffffff, 0x00ffffff,
489 0x8b10, 0x0000ff0f, 0x00000000,
490 0x28a4c, 0x07ffffff, 0x4e000000,
491 0x28350, 0x3f3f3fff, 0x00000082,
492 0x30, 0x000000ff, 0x0040,
493 0x34, 0x00000040, 0x00004040,
494 0x9100, 0x07ffffff, 0x03000000,
495 0x9060, 0x0000007f, 0x00000020,
496 0x9508, 0x00010000, 0x00010000,
497 0xac14, 0x000003ff, 0x000000f3,
498 0xac10, 0xffffffff, 0x00000000,
499 0xac0c, 0xffffffff, 0x00003210,
500 0x88d4, 0x0000001f, 0x00000010,
501 0x15c0, 0x000c0fc0, 0x000c0400
502};
503
fffbdda4
AD
504static const u32 hainan_golden_registers[] =
505{
506 0x9a10, 0x00010000, 0x00018208,
507 0x9830, 0xffffffff, 0x00000000,
508 0x9834, 0xf00fffff, 0x00000400,
509 0x9838, 0x0002021c, 0x00020200,
510 0xd0c0, 0xff000fff, 0x00000100,
511 0xd030, 0x000300c0, 0x00800040,
512 0xd8c0, 0xff000fff, 0x00000100,
513 0xd830, 0x000300c0, 0x00800040,
514 0x2ae4, 0x00073ffe, 0x000022a2,
515 0x240c, 0x000007ff, 0x00000000,
516 0x8a14, 0xf000001f, 0x00000007,
517 0x8b24, 0xffffffff, 0x00ffffff,
518 0x8b10, 0x0000ff0f, 0x00000000,
519 0x28a4c, 0x07ffffff, 0x4e000000,
520 0x28350, 0x3f3f3fff, 0x00000000,
521 0x30, 0x000000ff, 0x0040,
522 0x34, 0x00000040, 0x00004040,
523 0x9100, 0x03e00000, 0x03600000,
524 0x9060, 0x0000007f, 0x00000020,
525 0x9508, 0x00010000, 0x00010000,
526 0xac14, 0x000003ff, 0x000000f1,
527 0xac10, 0xffffffff, 0x00000000,
528 0xac0c, 0xffffffff, 0x00003210,
529 0x88d4, 0x0000001f, 0x00000010,
530 0x15c0, 0x000c0fc0, 0x000c0400
531};
532
533static const u32 hainan_golden_registers2[] =
534{
535 0x98f8, 0xffffffff, 0x02010001
536};
537
205996c0
AD
538static const u32 tahiti_mgcg_cgcg_init[] =
539{
540 0xc400, 0xffffffff, 0xfffffffc,
541 0x802c, 0xffffffff, 0xe0000000,
542 0x9a60, 0xffffffff, 0x00000100,
543 0x92a4, 0xffffffff, 0x00000100,
544 0xc164, 0xffffffff, 0x00000100,
545 0x9774, 0xffffffff, 0x00000100,
546 0x8984, 0xffffffff, 0x06000100,
547 0x8a18, 0xffffffff, 0x00000100,
548 0x92a0, 0xffffffff, 0x00000100,
549 0xc380, 0xffffffff, 0x00000100,
550 0x8b28, 0xffffffff, 0x00000100,
551 0x9144, 0xffffffff, 0x00000100,
552 0x8d88, 0xffffffff, 0x00000100,
553 0x8d8c, 0xffffffff, 0x00000100,
554 0x9030, 0xffffffff, 0x00000100,
555 0x9034, 0xffffffff, 0x00000100,
556 0x9038, 0xffffffff, 0x00000100,
557 0x903c, 0xffffffff, 0x00000100,
558 0xad80, 0xffffffff, 0x00000100,
559 0xac54, 0xffffffff, 0x00000100,
560 0x897c, 0xffffffff, 0x06000100,
561 0x9868, 0xffffffff, 0x00000100,
562 0x9510, 0xffffffff, 0x00000100,
563 0xaf04, 0xffffffff, 0x00000100,
564 0xae04, 0xffffffff, 0x00000100,
565 0x949c, 0xffffffff, 0x00000100,
566 0x802c, 0xffffffff, 0xe0000000,
567 0x9160, 0xffffffff, 0x00010000,
568 0x9164, 0xffffffff, 0x00030002,
569 0x9168, 0xffffffff, 0x00040007,
570 0x916c, 0xffffffff, 0x00060005,
571 0x9170, 0xffffffff, 0x00090008,
572 0x9174, 0xffffffff, 0x00020001,
573 0x9178, 0xffffffff, 0x00040003,
574 0x917c, 0xffffffff, 0x00000007,
575 0x9180, 0xffffffff, 0x00060005,
576 0x9184, 0xffffffff, 0x00090008,
577 0x9188, 0xffffffff, 0x00030002,
578 0x918c, 0xffffffff, 0x00050004,
579 0x9190, 0xffffffff, 0x00000008,
580 0x9194, 0xffffffff, 0x00070006,
581 0x9198, 0xffffffff, 0x000a0009,
582 0x919c, 0xffffffff, 0x00040003,
583 0x91a0, 0xffffffff, 0x00060005,
584 0x91a4, 0xffffffff, 0x00000009,
585 0x91a8, 0xffffffff, 0x00080007,
586 0x91ac, 0xffffffff, 0x000b000a,
587 0x91b0, 0xffffffff, 0x00050004,
588 0x91b4, 0xffffffff, 0x00070006,
589 0x91b8, 0xffffffff, 0x0008000b,
590 0x91bc, 0xffffffff, 0x000a0009,
591 0x91c0, 0xffffffff, 0x000d000c,
592 0x91c4, 0xffffffff, 0x00060005,
593 0x91c8, 0xffffffff, 0x00080007,
594 0x91cc, 0xffffffff, 0x0000000b,
595 0x91d0, 0xffffffff, 0x000a0009,
596 0x91d4, 0xffffffff, 0x000d000c,
597 0x91d8, 0xffffffff, 0x00070006,
598 0x91dc, 0xffffffff, 0x00090008,
599 0x91e0, 0xffffffff, 0x0000000c,
600 0x91e4, 0xffffffff, 0x000b000a,
601 0x91e8, 0xffffffff, 0x000e000d,
602 0x91ec, 0xffffffff, 0x00080007,
603 0x91f0, 0xffffffff, 0x000a0009,
604 0x91f4, 0xffffffff, 0x0000000d,
605 0x91f8, 0xffffffff, 0x000c000b,
606 0x91fc, 0xffffffff, 0x000f000e,
607 0x9200, 0xffffffff, 0x00090008,
608 0x9204, 0xffffffff, 0x000b000a,
609 0x9208, 0xffffffff, 0x000c000f,
610 0x920c, 0xffffffff, 0x000e000d,
611 0x9210, 0xffffffff, 0x00110010,
612 0x9214, 0xffffffff, 0x000a0009,
613 0x9218, 0xffffffff, 0x000c000b,
614 0x921c, 0xffffffff, 0x0000000f,
615 0x9220, 0xffffffff, 0x000e000d,
616 0x9224, 0xffffffff, 0x00110010,
617 0x9228, 0xffffffff, 0x000b000a,
618 0x922c, 0xffffffff, 0x000d000c,
619 0x9230, 0xffffffff, 0x00000010,
620 0x9234, 0xffffffff, 0x000f000e,
621 0x9238, 0xffffffff, 0x00120011,
622 0x923c, 0xffffffff, 0x000c000b,
623 0x9240, 0xffffffff, 0x000e000d,
624 0x9244, 0xffffffff, 0x00000011,
625 0x9248, 0xffffffff, 0x0010000f,
626 0x924c, 0xffffffff, 0x00130012,
627 0x9250, 0xffffffff, 0x000d000c,
628 0x9254, 0xffffffff, 0x000f000e,
629 0x9258, 0xffffffff, 0x00100013,
630 0x925c, 0xffffffff, 0x00120011,
631 0x9260, 0xffffffff, 0x00150014,
632 0x9264, 0xffffffff, 0x000e000d,
633 0x9268, 0xffffffff, 0x0010000f,
634 0x926c, 0xffffffff, 0x00000013,
635 0x9270, 0xffffffff, 0x00120011,
636 0x9274, 0xffffffff, 0x00150014,
637 0x9278, 0xffffffff, 0x000f000e,
638 0x927c, 0xffffffff, 0x00110010,
639 0x9280, 0xffffffff, 0x00000014,
640 0x9284, 0xffffffff, 0x00130012,
641 0x9288, 0xffffffff, 0x00160015,
642 0x928c, 0xffffffff, 0x0010000f,
643 0x9290, 0xffffffff, 0x00120011,
644 0x9294, 0xffffffff, 0x00000015,
645 0x9298, 0xffffffff, 0x00140013,
646 0x929c, 0xffffffff, 0x00170016,
647 0x9150, 0xffffffff, 0x96940200,
648 0x8708, 0xffffffff, 0x00900100,
649 0xc478, 0xffffffff, 0x00000080,
650 0xc404, 0xffffffff, 0x0020003f,
651 0x30, 0xffffffff, 0x0000001c,
652 0x34, 0x000f0000, 0x000f0000,
653 0x160c, 0xffffffff, 0x00000100,
654 0x1024, 0xffffffff, 0x00000100,
655 0x102c, 0x00000101, 0x00000000,
656 0x20a8, 0xffffffff, 0x00000104,
657 0x264c, 0x000c0000, 0x000c0000,
658 0x2648, 0x000c0000, 0x000c0000,
659 0x55e4, 0xff000fff, 0x00000100,
660 0x55e8, 0x00000001, 0x00000001,
661 0x2f50, 0x00000001, 0x00000001,
662 0x30cc, 0xc0000fff, 0x00000104,
663 0xc1e4, 0x00000001, 0x00000001,
664 0xd0c0, 0xfffffff0, 0x00000100,
665 0xd8c0, 0xfffffff0, 0x00000100
666};
667
668static const u32 pitcairn_mgcg_cgcg_init[] =
669{
670 0xc400, 0xffffffff, 0xfffffffc,
671 0x802c, 0xffffffff, 0xe0000000,
672 0x9a60, 0xffffffff, 0x00000100,
673 0x92a4, 0xffffffff, 0x00000100,
674 0xc164, 0xffffffff, 0x00000100,
675 0x9774, 0xffffffff, 0x00000100,
676 0x8984, 0xffffffff, 0x06000100,
677 0x8a18, 0xffffffff, 0x00000100,
678 0x92a0, 0xffffffff, 0x00000100,
679 0xc380, 0xffffffff, 0x00000100,
680 0x8b28, 0xffffffff, 0x00000100,
681 0x9144, 0xffffffff, 0x00000100,
682 0x8d88, 0xffffffff, 0x00000100,
683 0x8d8c, 0xffffffff, 0x00000100,
684 0x9030, 0xffffffff, 0x00000100,
685 0x9034, 0xffffffff, 0x00000100,
686 0x9038, 0xffffffff, 0x00000100,
687 0x903c, 0xffffffff, 0x00000100,
688 0xad80, 0xffffffff, 0x00000100,
689 0xac54, 0xffffffff, 0x00000100,
690 0x897c, 0xffffffff, 0x06000100,
691 0x9868, 0xffffffff, 0x00000100,
692 0x9510, 0xffffffff, 0x00000100,
693 0xaf04, 0xffffffff, 0x00000100,
694 0xae04, 0xffffffff, 0x00000100,
695 0x949c, 0xffffffff, 0x00000100,
696 0x802c, 0xffffffff, 0xe0000000,
697 0x9160, 0xffffffff, 0x00010000,
698 0x9164, 0xffffffff, 0x00030002,
699 0x9168, 0xffffffff, 0x00040007,
700 0x916c, 0xffffffff, 0x00060005,
701 0x9170, 0xffffffff, 0x00090008,
702 0x9174, 0xffffffff, 0x00020001,
703 0x9178, 0xffffffff, 0x00040003,
704 0x917c, 0xffffffff, 0x00000007,
705 0x9180, 0xffffffff, 0x00060005,
706 0x9184, 0xffffffff, 0x00090008,
707 0x9188, 0xffffffff, 0x00030002,
708 0x918c, 0xffffffff, 0x00050004,
709 0x9190, 0xffffffff, 0x00000008,
710 0x9194, 0xffffffff, 0x00070006,
711 0x9198, 0xffffffff, 0x000a0009,
712 0x919c, 0xffffffff, 0x00040003,
713 0x91a0, 0xffffffff, 0x00060005,
714 0x91a4, 0xffffffff, 0x00000009,
715 0x91a8, 0xffffffff, 0x00080007,
716 0x91ac, 0xffffffff, 0x000b000a,
717 0x91b0, 0xffffffff, 0x00050004,
718 0x91b4, 0xffffffff, 0x00070006,
719 0x91b8, 0xffffffff, 0x0008000b,
720 0x91bc, 0xffffffff, 0x000a0009,
721 0x91c0, 0xffffffff, 0x000d000c,
722 0x9200, 0xffffffff, 0x00090008,
723 0x9204, 0xffffffff, 0x000b000a,
724 0x9208, 0xffffffff, 0x000c000f,
725 0x920c, 0xffffffff, 0x000e000d,
726 0x9210, 0xffffffff, 0x00110010,
727 0x9214, 0xffffffff, 0x000a0009,
728 0x9218, 0xffffffff, 0x000c000b,
729 0x921c, 0xffffffff, 0x0000000f,
730 0x9220, 0xffffffff, 0x000e000d,
731 0x9224, 0xffffffff, 0x00110010,
732 0x9228, 0xffffffff, 0x000b000a,
733 0x922c, 0xffffffff, 0x000d000c,
734 0x9230, 0xffffffff, 0x00000010,
735 0x9234, 0xffffffff, 0x000f000e,
736 0x9238, 0xffffffff, 0x00120011,
737 0x923c, 0xffffffff, 0x000c000b,
738 0x9240, 0xffffffff, 0x000e000d,
739 0x9244, 0xffffffff, 0x00000011,
740 0x9248, 0xffffffff, 0x0010000f,
741 0x924c, 0xffffffff, 0x00130012,
742 0x9250, 0xffffffff, 0x000d000c,
743 0x9254, 0xffffffff, 0x000f000e,
744 0x9258, 0xffffffff, 0x00100013,
745 0x925c, 0xffffffff, 0x00120011,
746 0x9260, 0xffffffff, 0x00150014,
747 0x9150, 0xffffffff, 0x96940200,
748 0x8708, 0xffffffff, 0x00900100,
749 0xc478, 0xffffffff, 0x00000080,
750 0xc404, 0xffffffff, 0x0020003f,
751 0x30, 0xffffffff, 0x0000001c,
752 0x34, 0x000f0000, 0x000f0000,
753 0x160c, 0xffffffff, 0x00000100,
754 0x1024, 0xffffffff, 0x00000100,
755 0x102c, 0x00000101, 0x00000000,
756 0x20a8, 0xffffffff, 0x00000104,
757 0x55e4, 0xff000fff, 0x00000100,
758 0x55e8, 0x00000001, 0x00000001,
759 0x2f50, 0x00000001, 0x00000001,
760 0x30cc, 0xc0000fff, 0x00000104,
761 0xc1e4, 0x00000001, 0x00000001,
762 0xd0c0, 0xfffffff0, 0x00000100,
763 0xd8c0, 0xfffffff0, 0x00000100
764};
765
766static const u32 verde_mgcg_cgcg_init[] =
767{
768 0xc400, 0xffffffff, 0xfffffffc,
769 0x802c, 0xffffffff, 0xe0000000,
770 0x9a60, 0xffffffff, 0x00000100,
771 0x92a4, 0xffffffff, 0x00000100,
772 0xc164, 0xffffffff, 0x00000100,
773 0x9774, 0xffffffff, 0x00000100,
774 0x8984, 0xffffffff, 0x06000100,
775 0x8a18, 0xffffffff, 0x00000100,
776 0x92a0, 0xffffffff, 0x00000100,
777 0xc380, 0xffffffff, 0x00000100,
778 0x8b28, 0xffffffff, 0x00000100,
779 0x9144, 0xffffffff, 0x00000100,
780 0x8d88, 0xffffffff, 0x00000100,
781 0x8d8c, 0xffffffff, 0x00000100,
782 0x9030, 0xffffffff, 0x00000100,
783 0x9034, 0xffffffff, 0x00000100,
784 0x9038, 0xffffffff, 0x00000100,
785 0x903c, 0xffffffff, 0x00000100,
786 0xad80, 0xffffffff, 0x00000100,
787 0xac54, 0xffffffff, 0x00000100,
788 0x897c, 0xffffffff, 0x06000100,
789 0x9868, 0xffffffff, 0x00000100,
790 0x9510, 0xffffffff, 0x00000100,
791 0xaf04, 0xffffffff, 0x00000100,
792 0xae04, 0xffffffff, 0x00000100,
793 0x949c, 0xffffffff, 0x00000100,
794 0x802c, 0xffffffff, 0xe0000000,
795 0x9160, 0xffffffff, 0x00010000,
796 0x9164, 0xffffffff, 0x00030002,
797 0x9168, 0xffffffff, 0x00040007,
798 0x916c, 0xffffffff, 0x00060005,
799 0x9170, 0xffffffff, 0x00090008,
800 0x9174, 0xffffffff, 0x00020001,
801 0x9178, 0xffffffff, 0x00040003,
802 0x917c, 0xffffffff, 0x00000007,
803 0x9180, 0xffffffff, 0x00060005,
804 0x9184, 0xffffffff, 0x00090008,
805 0x9188, 0xffffffff, 0x00030002,
806 0x918c, 0xffffffff, 0x00050004,
807 0x9190, 0xffffffff, 0x00000008,
808 0x9194, 0xffffffff, 0x00070006,
809 0x9198, 0xffffffff, 0x000a0009,
810 0x919c, 0xffffffff, 0x00040003,
811 0x91a0, 0xffffffff, 0x00060005,
812 0x91a4, 0xffffffff, 0x00000009,
813 0x91a8, 0xffffffff, 0x00080007,
814 0x91ac, 0xffffffff, 0x000b000a,
815 0x91b0, 0xffffffff, 0x00050004,
816 0x91b4, 0xffffffff, 0x00070006,
817 0x91b8, 0xffffffff, 0x0008000b,
818 0x91bc, 0xffffffff, 0x000a0009,
819 0x91c0, 0xffffffff, 0x000d000c,
820 0x9200, 0xffffffff, 0x00090008,
821 0x9204, 0xffffffff, 0x000b000a,
822 0x9208, 0xffffffff, 0x000c000f,
823 0x920c, 0xffffffff, 0x000e000d,
824 0x9210, 0xffffffff, 0x00110010,
825 0x9214, 0xffffffff, 0x000a0009,
826 0x9218, 0xffffffff, 0x000c000b,
827 0x921c, 0xffffffff, 0x0000000f,
828 0x9220, 0xffffffff, 0x000e000d,
829 0x9224, 0xffffffff, 0x00110010,
830 0x9228, 0xffffffff, 0x000b000a,
831 0x922c, 0xffffffff, 0x000d000c,
832 0x9230, 0xffffffff, 0x00000010,
833 0x9234, 0xffffffff, 0x000f000e,
834 0x9238, 0xffffffff, 0x00120011,
835 0x923c, 0xffffffff, 0x000c000b,
836 0x9240, 0xffffffff, 0x000e000d,
837 0x9244, 0xffffffff, 0x00000011,
838 0x9248, 0xffffffff, 0x0010000f,
839 0x924c, 0xffffffff, 0x00130012,
840 0x9250, 0xffffffff, 0x000d000c,
841 0x9254, 0xffffffff, 0x000f000e,
842 0x9258, 0xffffffff, 0x00100013,
843 0x925c, 0xffffffff, 0x00120011,
844 0x9260, 0xffffffff, 0x00150014,
845 0x9150, 0xffffffff, 0x96940200,
846 0x8708, 0xffffffff, 0x00900100,
847 0xc478, 0xffffffff, 0x00000080,
848 0xc404, 0xffffffff, 0x0020003f,
849 0x30, 0xffffffff, 0x0000001c,
850 0x34, 0x000f0000, 0x000f0000,
851 0x160c, 0xffffffff, 0x00000100,
852 0x1024, 0xffffffff, 0x00000100,
853 0x102c, 0x00000101, 0x00000000,
854 0x20a8, 0xffffffff, 0x00000104,
855 0x264c, 0x000c0000, 0x000c0000,
856 0x2648, 0x000c0000, 0x000c0000,
857 0x55e4, 0xff000fff, 0x00000100,
858 0x55e8, 0x00000001, 0x00000001,
859 0x2f50, 0x00000001, 0x00000001,
860 0x30cc, 0xc0000fff, 0x00000104,
861 0xc1e4, 0x00000001, 0x00000001,
862 0xd0c0, 0xfffffff0, 0x00000100,
863 0xd8c0, 0xfffffff0, 0x00000100
864};
865
866static const u32 oland_mgcg_cgcg_init[] =
867{
868 0xc400, 0xffffffff, 0xfffffffc,
869 0x802c, 0xffffffff, 0xe0000000,
870 0x9a60, 0xffffffff, 0x00000100,
871 0x92a4, 0xffffffff, 0x00000100,
872 0xc164, 0xffffffff, 0x00000100,
873 0x9774, 0xffffffff, 0x00000100,
874 0x8984, 0xffffffff, 0x06000100,
875 0x8a18, 0xffffffff, 0x00000100,
876 0x92a0, 0xffffffff, 0x00000100,
877 0xc380, 0xffffffff, 0x00000100,
878 0x8b28, 0xffffffff, 0x00000100,
879 0x9144, 0xffffffff, 0x00000100,
880 0x8d88, 0xffffffff, 0x00000100,
881 0x8d8c, 0xffffffff, 0x00000100,
882 0x9030, 0xffffffff, 0x00000100,
883 0x9034, 0xffffffff, 0x00000100,
884 0x9038, 0xffffffff, 0x00000100,
885 0x903c, 0xffffffff, 0x00000100,
886 0xad80, 0xffffffff, 0x00000100,
887 0xac54, 0xffffffff, 0x00000100,
888 0x897c, 0xffffffff, 0x06000100,
889 0x9868, 0xffffffff, 0x00000100,
890 0x9510, 0xffffffff, 0x00000100,
891 0xaf04, 0xffffffff, 0x00000100,
892 0xae04, 0xffffffff, 0x00000100,
893 0x949c, 0xffffffff, 0x00000100,
894 0x802c, 0xffffffff, 0xe0000000,
895 0x9160, 0xffffffff, 0x00010000,
896 0x9164, 0xffffffff, 0x00030002,
897 0x9168, 0xffffffff, 0x00040007,
898 0x916c, 0xffffffff, 0x00060005,
899 0x9170, 0xffffffff, 0x00090008,
900 0x9174, 0xffffffff, 0x00020001,
901 0x9178, 0xffffffff, 0x00040003,
902 0x917c, 0xffffffff, 0x00000007,
903 0x9180, 0xffffffff, 0x00060005,
904 0x9184, 0xffffffff, 0x00090008,
905 0x9188, 0xffffffff, 0x00030002,
906 0x918c, 0xffffffff, 0x00050004,
907 0x9190, 0xffffffff, 0x00000008,
908 0x9194, 0xffffffff, 0x00070006,
909 0x9198, 0xffffffff, 0x000a0009,
910 0x919c, 0xffffffff, 0x00040003,
911 0x91a0, 0xffffffff, 0x00060005,
912 0x91a4, 0xffffffff, 0x00000009,
913 0x91a8, 0xffffffff, 0x00080007,
914 0x91ac, 0xffffffff, 0x000b000a,
915 0x91b0, 0xffffffff, 0x00050004,
916 0x91b4, 0xffffffff, 0x00070006,
917 0x91b8, 0xffffffff, 0x0008000b,
918 0x91bc, 0xffffffff, 0x000a0009,
919 0x91c0, 0xffffffff, 0x000d000c,
920 0x91c4, 0xffffffff, 0x00060005,
921 0x91c8, 0xffffffff, 0x00080007,
922 0x91cc, 0xffffffff, 0x0000000b,
923 0x91d0, 0xffffffff, 0x000a0009,
924 0x91d4, 0xffffffff, 0x000d000c,
925 0x9150, 0xffffffff, 0x96940200,
926 0x8708, 0xffffffff, 0x00900100,
927 0xc478, 0xffffffff, 0x00000080,
928 0xc404, 0xffffffff, 0x0020003f,
929 0x30, 0xffffffff, 0x0000001c,
930 0x34, 0x000f0000, 0x000f0000,
931 0x160c, 0xffffffff, 0x00000100,
932 0x1024, 0xffffffff, 0x00000100,
933 0x102c, 0x00000101, 0x00000000,
934 0x20a8, 0xffffffff, 0x00000104,
935 0x264c, 0x000c0000, 0x000c0000,
936 0x2648, 0x000c0000, 0x000c0000,
937 0x55e4, 0xff000fff, 0x00000100,
938 0x55e8, 0x00000001, 0x00000001,
939 0x2f50, 0x00000001, 0x00000001,
940 0x30cc, 0xc0000fff, 0x00000104,
941 0xc1e4, 0x00000001, 0x00000001,
942 0xd0c0, 0xfffffff0, 0x00000100,
943 0xd8c0, 0xfffffff0, 0x00000100
944};
945
fffbdda4
AD
946static const u32 hainan_mgcg_cgcg_init[] =
947{
948 0xc400, 0xffffffff, 0xfffffffc,
949 0x802c, 0xffffffff, 0xe0000000,
950 0x9a60, 0xffffffff, 0x00000100,
951 0x92a4, 0xffffffff, 0x00000100,
952 0xc164, 0xffffffff, 0x00000100,
953 0x9774, 0xffffffff, 0x00000100,
954 0x8984, 0xffffffff, 0x06000100,
955 0x8a18, 0xffffffff, 0x00000100,
956 0x92a0, 0xffffffff, 0x00000100,
957 0xc380, 0xffffffff, 0x00000100,
958 0x8b28, 0xffffffff, 0x00000100,
959 0x9144, 0xffffffff, 0x00000100,
960 0x8d88, 0xffffffff, 0x00000100,
961 0x8d8c, 0xffffffff, 0x00000100,
962 0x9030, 0xffffffff, 0x00000100,
963 0x9034, 0xffffffff, 0x00000100,
964 0x9038, 0xffffffff, 0x00000100,
965 0x903c, 0xffffffff, 0x00000100,
966 0xad80, 0xffffffff, 0x00000100,
967 0xac54, 0xffffffff, 0x00000100,
968 0x897c, 0xffffffff, 0x06000100,
969 0x9868, 0xffffffff, 0x00000100,
970 0x9510, 0xffffffff, 0x00000100,
971 0xaf04, 0xffffffff, 0x00000100,
972 0xae04, 0xffffffff, 0x00000100,
973 0x949c, 0xffffffff, 0x00000100,
974 0x802c, 0xffffffff, 0xe0000000,
975 0x9160, 0xffffffff, 0x00010000,
976 0x9164, 0xffffffff, 0x00030002,
977 0x9168, 0xffffffff, 0x00040007,
978 0x916c, 0xffffffff, 0x00060005,
979 0x9170, 0xffffffff, 0x00090008,
980 0x9174, 0xffffffff, 0x00020001,
981 0x9178, 0xffffffff, 0x00040003,
982 0x917c, 0xffffffff, 0x00000007,
983 0x9180, 0xffffffff, 0x00060005,
984 0x9184, 0xffffffff, 0x00090008,
985 0x9188, 0xffffffff, 0x00030002,
986 0x918c, 0xffffffff, 0x00050004,
987 0x9190, 0xffffffff, 0x00000008,
988 0x9194, 0xffffffff, 0x00070006,
989 0x9198, 0xffffffff, 0x000a0009,
990 0x919c, 0xffffffff, 0x00040003,
991 0x91a0, 0xffffffff, 0x00060005,
992 0x91a4, 0xffffffff, 0x00000009,
993 0x91a8, 0xffffffff, 0x00080007,
994 0x91ac, 0xffffffff, 0x000b000a,
995 0x91b0, 0xffffffff, 0x00050004,
996 0x91b4, 0xffffffff, 0x00070006,
997 0x91b8, 0xffffffff, 0x0008000b,
998 0x91bc, 0xffffffff, 0x000a0009,
999 0x91c0, 0xffffffff, 0x000d000c,
1000 0x91c4, 0xffffffff, 0x00060005,
1001 0x91c8, 0xffffffff, 0x00080007,
1002 0x91cc, 0xffffffff, 0x0000000b,
1003 0x91d0, 0xffffffff, 0x000a0009,
1004 0x91d4, 0xffffffff, 0x000d000c,
1005 0x9150, 0xffffffff, 0x96940200,
1006 0x8708, 0xffffffff, 0x00900100,
1007 0xc478, 0xffffffff, 0x00000080,
1008 0xc404, 0xffffffff, 0x0020003f,
1009 0x30, 0xffffffff, 0x0000001c,
1010 0x34, 0x000f0000, 0x000f0000,
1011 0x160c, 0xffffffff, 0x00000100,
1012 0x1024, 0xffffffff, 0x00000100,
1013 0x20a8, 0xffffffff, 0x00000104,
1014 0x264c, 0x000c0000, 0x000c0000,
1015 0x2648, 0x000c0000, 0x000c0000,
1016 0x2f50, 0x00000001, 0x00000001,
1017 0x30cc, 0xc0000fff, 0x00000104,
1018 0xc1e4, 0x00000001, 0x00000001,
1019 0xd0c0, 0xfffffff0, 0x00000100,
1020 0xd8c0, 0xfffffff0, 0x00000100
1021};
1022
205996c0
AD
1023static u32 verde_pg_init[] =
1024{
1025 0x353c, 0xffffffff, 0x40000,
1026 0x3538, 0xffffffff, 0x200010ff,
1027 0x353c, 0xffffffff, 0x0,
1028 0x353c, 0xffffffff, 0x0,
1029 0x353c, 0xffffffff, 0x0,
1030 0x353c, 0xffffffff, 0x0,
1031 0x353c, 0xffffffff, 0x0,
1032 0x353c, 0xffffffff, 0x7007,
1033 0x3538, 0xffffffff, 0x300010ff,
1034 0x353c, 0xffffffff, 0x0,
1035 0x353c, 0xffffffff, 0x0,
1036 0x353c, 0xffffffff, 0x0,
1037 0x353c, 0xffffffff, 0x0,
1038 0x353c, 0xffffffff, 0x0,
1039 0x353c, 0xffffffff, 0x400000,
1040 0x3538, 0xffffffff, 0x100010ff,
1041 0x353c, 0xffffffff, 0x0,
1042 0x353c, 0xffffffff, 0x0,
1043 0x353c, 0xffffffff, 0x0,
1044 0x353c, 0xffffffff, 0x0,
1045 0x353c, 0xffffffff, 0x0,
1046 0x353c, 0xffffffff, 0x120200,
1047 0x3538, 0xffffffff, 0x500010ff,
1048 0x353c, 0xffffffff, 0x0,
1049 0x353c, 0xffffffff, 0x0,
1050 0x353c, 0xffffffff, 0x0,
1051 0x353c, 0xffffffff, 0x0,
1052 0x353c, 0xffffffff, 0x0,
1053 0x353c, 0xffffffff, 0x1e1e16,
1054 0x3538, 0xffffffff, 0x600010ff,
1055 0x353c, 0xffffffff, 0x0,
1056 0x353c, 0xffffffff, 0x0,
1057 0x353c, 0xffffffff, 0x0,
1058 0x353c, 0xffffffff, 0x0,
1059 0x353c, 0xffffffff, 0x0,
1060 0x353c, 0xffffffff, 0x171f1e,
1061 0x3538, 0xffffffff, 0x700010ff,
1062 0x353c, 0xffffffff, 0x0,
1063 0x353c, 0xffffffff, 0x0,
1064 0x353c, 0xffffffff, 0x0,
1065 0x353c, 0xffffffff, 0x0,
1066 0x353c, 0xffffffff, 0x0,
1067 0x353c, 0xffffffff, 0x0,
1068 0x3538, 0xffffffff, 0x9ff,
1069 0x3500, 0xffffffff, 0x0,
1070 0x3504, 0xffffffff, 0x10000800,
1071 0x3504, 0xffffffff, 0xf,
1072 0x3504, 0xffffffff, 0xf,
1073 0x3500, 0xffffffff, 0x4,
1074 0x3504, 0xffffffff, 0x1000051e,
1075 0x3504, 0xffffffff, 0xffff,
1076 0x3504, 0xffffffff, 0xffff,
1077 0x3500, 0xffffffff, 0x8,
1078 0x3504, 0xffffffff, 0x80500,
1079 0x3500, 0xffffffff, 0x12,
1080 0x3504, 0xffffffff, 0x9050c,
1081 0x3500, 0xffffffff, 0x1d,
1082 0x3504, 0xffffffff, 0xb052c,
1083 0x3500, 0xffffffff, 0x2a,
1084 0x3504, 0xffffffff, 0x1053e,
1085 0x3500, 0xffffffff, 0x2d,
1086 0x3504, 0xffffffff, 0x10546,
1087 0x3500, 0xffffffff, 0x30,
1088 0x3504, 0xffffffff, 0xa054e,
1089 0x3500, 0xffffffff, 0x3c,
1090 0x3504, 0xffffffff, 0x1055f,
1091 0x3500, 0xffffffff, 0x3f,
1092 0x3504, 0xffffffff, 0x10567,
1093 0x3500, 0xffffffff, 0x42,
1094 0x3504, 0xffffffff, 0x1056f,
1095 0x3500, 0xffffffff, 0x45,
1096 0x3504, 0xffffffff, 0x10572,
1097 0x3500, 0xffffffff, 0x48,
1098 0x3504, 0xffffffff, 0x20575,
1099 0x3500, 0xffffffff, 0x4c,
1100 0x3504, 0xffffffff, 0x190801,
1101 0x3500, 0xffffffff, 0x67,
1102 0x3504, 0xffffffff, 0x1082a,
1103 0x3500, 0xffffffff, 0x6a,
1104 0x3504, 0xffffffff, 0x1b082d,
1105 0x3500, 0xffffffff, 0x87,
1106 0x3504, 0xffffffff, 0x310851,
1107 0x3500, 0xffffffff, 0xba,
1108 0x3504, 0xffffffff, 0x891,
1109 0x3500, 0xffffffff, 0xbc,
1110 0x3504, 0xffffffff, 0x893,
1111 0x3500, 0xffffffff, 0xbe,
1112 0x3504, 0xffffffff, 0x20895,
1113 0x3500, 0xffffffff, 0xc2,
1114 0x3504, 0xffffffff, 0x20899,
1115 0x3500, 0xffffffff, 0xc6,
1116 0x3504, 0xffffffff, 0x2089d,
1117 0x3500, 0xffffffff, 0xca,
1118 0x3504, 0xffffffff, 0x8a1,
1119 0x3500, 0xffffffff, 0xcc,
1120 0x3504, 0xffffffff, 0x8a3,
1121 0x3500, 0xffffffff, 0xce,
1122 0x3504, 0xffffffff, 0x308a5,
1123 0x3500, 0xffffffff, 0xd3,
1124 0x3504, 0xffffffff, 0x6d08cd,
1125 0x3500, 0xffffffff, 0x142,
1126 0x3504, 0xffffffff, 0x2000095a,
1127 0x3504, 0xffffffff, 0x1,
1128 0x3500, 0xffffffff, 0x144,
1129 0x3504, 0xffffffff, 0x301f095b,
1130 0x3500, 0xffffffff, 0x165,
1131 0x3504, 0xffffffff, 0xc094d,
1132 0x3500, 0xffffffff, 0x173,
1133 0x3504, 0xffffffff, 0xf096d,
1134 0x3500, 0xffffffff, 0x184,
1135 0x3504, 0xffffffff, 0x15097f,
1136 0x3500, 0xffffffff, 0x19b,
1137 0x3504, 0xffffffff, 0xc0998,
1138 0x3500, 0xffffffff, 0x1a9,
1139 0x3504, 0xffffffff, 0x409a7,
1140 0x3500, 0xffffffff, 0x1af,
1141 0x3504, 0xffffffff, 0xcdc,
1142 0x3500, 0xffffffff, 0x1b1,
1143 0x3504, 0xffffffff, 0x800,
1144 0x3508, 0xffffffff, 0x6c9b2000,
1145 0x3510, 0xfc00, 0x2000,
1146 0x3544, 0xffffffff, 0xfc0,
1147 0x28d4, 0x00000100, 0x100
1148};
1149
1150static void si_init_golden_registers(struct radeon_device *rdev)
1151{
1152 switch (rdev->family) {
1153 case CHIP_TAHITI:
1154 radeon_program_register_sequence(rdev,
1155 tahiti_golden_registers,
1156 (const u32)ARRAY_SIZE(tahiti_golden_registers));
1157 radeon_program_register_sequence(rdev,
1158 tahiti_golden_rlc_registers,
1159 (const u32)ARRAY_SIZE(tahiti_golden_rlc_registers));
1160 radeon_program_register_sequence(rdev,
1161 tahiti_mgcg_cgcg_init,
1162 (const u32)ARRAY_SIZE(tahiti_mgcg_cgcg_init));
1163 radeon_program_register_sequence(rdev,
1164 tahiti_golden_registers2,
1165 (const u32)ARRAY_SIZE(tahiti_golden_registers2));
1166 break;
1167 case CHIP_PITCAIRN:
1168 radeon_program_register_sequence(rdev,
1169 pitcairn_golden_registers,
1170 (const u32)ARRAY_SIZE(pitcairn_golden_registers));
1171 radeon_program_register_sequence(rdev,
1172 pitcairn_golden_rlc_registers,
1173 (const u32)ARRAY_SIZE(pitcairn_golden_rlc_registers));
1174 radeon_program_register_sequence(rdev,
1175 pitcairn_mgcg_cgcg_init,
1176 (const u32)ARRAY_SIZE(pitcairn_mgcg_cgcg_init));
1177 break;
1178 case CHIP_VERDE:
1179 radeon_program_register_sequence(rdev,
1180 verde_golden_registers,
1181 (const u32)ARRAY_SIZE(verde_golden_registers));
1182 radeon_program_register_sequence(rdev,
1183 verde_golden_rlc_registers,
1184 (const u32)ARRAY_SIZE(verde_golden_rlc_registers));
1185 radeon_program_register_sequence(rdev,
1186 verde_mgcg_cgcg_init,
1187 (const u32)ARRAY_SIZE(verde_mgcg_cgcg_init));
1188 radeon_program_register_sequence(rdev,
1189 verde_pg_init,
1190 (const u32)ARRAY_SIZE(verde_pg_init));
1191 break;
1192 case CHIP_OLAND:
1193 radeon_program_register_sequence(rdev,
1194 oland_golden_registers,
1195 (const u32)ARRAY_SIZE(oland_golden_registers));
1196 radeon_program_register_sequence(rdev,
1197 oland_golden_rlc_registers,
1198 (const u32)ARRAY_SIZE(oland_golden_rlc_registers));
1199 radeon_program_register_sequence(rdev,
1200 oland_mgcg_cgcg_init,
1201 (const u32)ARRAY_SIZE(oland_mgcg_cgcg_init));
1202 break;
fffbdda4
AD
1203 case CHIP_HAINAN:
1204 radeon_program_register_sequence(rdev,
1205 hainan_golden_registers,
1206 (const u32)ARRAY_SIZE(hainan_golden_registers));
1207 radeon_program_register_sequence(rdev,
1208 hainan_golden_registers2,
1209 (const u32)ARRAY_SIZE(hainan_golden_registers2));
1210 radeon_program_register_sequence(rdev,
1211 hainan_mgcg_cgcg_init,
1212 (const u32)ARRAY_SIZE(hainan_mgcg_cgcg_init));
1213 break;
205996c0
AD
1214 default:
1215 break;
1216 }
1217}
1218
454d2e2a
AD
1219#define PCIE_BUS_CLK 10000
1220#define TCLK (PCIE_BUS_CLK / 10)
1221
1222/**
1223 * si_get_xclk - get the xclk
1224 *
1225 * @rdev: radeon_device pointer
1226 *
1227 * Returns the reference clock used by the gfx engine
1228 * (SI).
1229 */
1230u32 si_get_xclk(struct radeon_device *rdev)
1231{
1232 u32 reference_clock = rdev->clock.spll.reference_freq;
1233 u32 tmp;
1234
1235 tmp = RREG32(CG_CLKPIN_CNTL_2);
1236 if (tmp & MUX_TCLK_TO_XCLK)
1237 return TCLK;
1238
1239 tmp = RREG32(CG_CLKPIN_CNTL);
1240 if (tmp & XTALIN_DIVIDE)
1241 return reference_clock / 4;
1242
1243 return reference_clock;
1244}
1245
1bd47d2e
AD
1246/* get temperature in millidegrees */
1247int si_get_temp(struct radeon_device *rdev)
1248{
1249 u32 temp;
1250 int actual_temp = 0;
1251
1252 temp = (RREG32(CG_MULT_THERMAL_STATUS) & CTF_TEMP_MASK) >>
1253 CTF_TEMP_SHIFT;
1254
1255 if (temp & 0x200)
1256 actual_temp = 255;
1257 else
1258 actual_temp = temp & 0x1ff;
1259
1260 actual_temp = (actual_temp * 1000);
1261
1262 return actual_temp;
1263}
1264
8b074dd6
AD
1265#define TAHITI_IO_MC_REGS_SIZE 36
1266
1267static const u32 tahiti_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
1268 {0x0000006f, 0x03044000},
1269 {0x00000070, 0x0480c018},
1270 {0x00000071, 0x00000040},
1271 {0x00000072, 0x01000000},
1272 {0x00000074, 0x000000ff},
1273 {0x00000075, 0x00143400},
1274 {0x00000076, 0x08ec0800},
1275 {0x00000077, 0x040000cc},
1276 {0x00000079, 0x00000000},
1277 {0x0000007a, 0x21000409},
1278 {0x0000007c, 0x00000000},
1279 {0x0000007d, 0xe8000000},
1280 {0x0000007e, 0x044408a8},
1281 {0x0000007f, 0x00000003},
1282 {0x00000080, 0x00000000},
1283 {0x00000081, 0x01000000},
1284 {0x00000082, 0x02000000},
1285 {0x00000083, 0x00000000},
1286 {0x00000084, 0xe3f3e4f4},
1287 {0x00000085, 0x00052024},
1288 {0x00000087, 0x00000000},
1289 {0x00000088, 0x66036603},
1290 {0x00000089, 0x01000000},
1291 {0x0000008b, 0x1c0a0000},
1292 {0x0000008c, 0xff010000},
1293 {0x0000008e, 0xffffefff},
1294 {0x0000008f, 0xfff3efff},
1295 {0x00000090, 0xfff3efbf},
1296 {0x00000094, 0x00101101},
1297 {0x00000095, 0x00000fff},
1298 {0x00000096, 0x00116fff},
1299 {0x00000097, 0x60010000},
1300 {0x00000098, 0x10010000},
1301 {0x00000099, 0x00006000},
1302 {0x0000009a, 0x00001000},
1303 {0x0000009f, 0x00a77400}
1304};
1305
1306static const u32 pitcairn_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
1307 {0x0000006f, 0x03044000},
1308 {0x00000070, 0x0480c018},
1309 {0x00000071, 0x00000040},
1310 {0x00000072, 0x01000000},
1311 {0x00000074, 0x000000ff},
1312 {0x00000075, 0x00143400},
1313 {0x00000076, 0x08ec0800},
1314 {0x00000077, 0x040000cc},
1315 {0x00000079, 0x00000000},
1316 {0x0000007a, 0x21000409},
1317 {0x0000007c, 0x00000000},
1318 {0x0000007d, 0xe8000000},
1319 {0x0000007e, 0x044408a8},
1320 {0x0000007f, 0x00000003},
1321 {0x00000080, 0x00000000},
1322 {0x00000081, 0x01000000},
1323 {0x00000082, 0x02000000},
1324 {0x00000083, 0x00000000},
1325 {0x00000084, 0xe3f3e4f4},
1326 {0x00000085, 0x00052024},
1327 {0x00000087, 0x00000000},
1328 {0x00000088, 0x66036603},
1329 {0x00000089, 0x01000000},
1330 {0x0000008b, 0x1c0a0000},
1331 {0x0000008c, 0xff010000},
1332 {0x0000008e, 0xffffefff},
1333 {0x0000008f, 0xfff3efff},
1334 {0x00000090, 0xfff3efbf},
1335 {0x00000094, 0x00101101},
1336 {0x00000095, 0x00000fff},
1337 {0x00000096, 0x00116fff},
1338 {0x00000097, 0x60010000},
1339 {0x00000098, 0x10010000},
1340 {0x00000099, 0x00006000},
1341 {0x0000009a, 0x00001000},
1342 {0x0000009f, 0x00a47400}
1343};
1344
1345static const u32 verde_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
1346 {0x0000006f, 0x03044000},
1347 {0x00000070, 0x0480c018},
1348 {0x00000071, 0x00000040},
1349 {0x00000072, 0x01000000},
1350 {0x00000074, 0x000000ff},
1351 {0x00000075, 0x00143400},
1352 {0x00000076, 0x08ec0800},
1353 {0x00000077, 0x040000cc},
1354 {0x00000079, 0x00000000},
1355 {0x0000007a, 0x21000409},
1356 {0x0000007c, 0x00000000},
1357 {0x0000007d, 0xe8000000},
1358 {0x0000007e, 0x044408a8},
1359 {0x0000007f, 0x00000003},
1360 {0x00000080, 0x00000000},
1361 {0x00000081, 0x01000000},
1362 {0x00000082, 0x02000000},
1363 {0x00000083, 0x00000000},
1364 {0x00000084, 0xe3f3e4f4},
1365 {0x00000085, 0x00052024},
1366 {0x00000087, 0x00000000},
1367 {0x00000088, 0x66036603},
1368 {0x00000089, 0x01000000},
1369 {0x0000008b, 0x1c0a0000},
1370 {0x0000008c, 0xff010000},
1371 {0x0000008e, 0xffffefff},
1372 {0x0000008f, 0xfff3efff},
1373 {0x00000090, 0xfff3efbf},
1374 {0x00000094, 0x00101101},
1375 {0x00000095, 0x00000fff},
1376 {0x00000096, 0x00116fff},
1377 {0x00000097, 0x60010000},
1378 {0x00000098, 0x10010000},
1379 {0x00000099, 0x00006000},
1380 {0x0000009a, 0x00001000},
1381 {0x0000009f, 0x00a37400}
1382};
1383
bcc7f5d2
AD
1384static const u32 oland_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
1385 {0x0000006f, 0x03044000},
1386 {0x00000070, 0x0480c018},
1387 {0x00000071, 0x00000040},
1388 {0x00000072, 0x01000000},
1389 {0x00000074, 0x000000ff},
1390 {0x00000075, 0x00143400},
1391 {0x00000076, 0x08ec0800},
1392 {0x00000077, 0x040000cc},
1393 {0x00000079, 0x00000000},
1394 {0x0000007a, 0x21000409},
1395 {0x0000007c, 0x00000000},
1396 {0x0000007d, 0xe8000000},
1397 {0x0000007e, 0x044408a8},
1398 {0x0000007f, 0x00000003},
1399 {0x00000080, 0x00000000},
1400 {0x00000081, 0x01000000},
1401 {0x00000082, 0x02000000},
1402 {0x00000083, 0x00000000},
1403 {0x00000084, 0xe3f3e4f4},
1404 {0x00000085, 0x00052024},
1405 {0x00000087, 0x00000000},
1406 {0x00000088, 0x66036603},
1407 {0x00000089, 0x01000000},
1408 {0x0000008b, 0x1c0a0000},
1409 {0x0000008c, 0xff010000},
1410 {0x0000008e, 0xffffefff},
1411 {0x0000008f, 0xfff3efff},
1412 {0x00000090, 0xfff3efbf},
1413 {0x00000094, 0x00101101},
1414 {0x00000095, 0x00000fff},
1415 {0x00000096, 0x00116fff},
1416 {0x00000097, 0x60010000},
1417 {0x00000098, 0x10010000},
1418 {0x00000099, 0x00006000},
1419 {0x0000009a, 0x00001000},
1420 {0x0000009f, 0x00a17730}
1421};
1422
c04c00b4
AD
1423static const u32 hainan_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
1424 {0x0000006f, 0x03044000},
1425 {0x00000070, 0x0480c018},
1426 {0x00000071, 0x00000040},
1427 {0x00000072, 0x01000000},
1428 {0x00000074, 0x000000ff},
1429 {0x00000075, 0x00143400},
1430 {0x00000076, 0x08ec0800},
1431 {0x00000077, 0x040000cc},
1432 {0x00000079, 0x00000000},
1433 {0x0000007a, 0x21000409},
1434 {0x0000007c, 0x00000000},
1435 {0x0000007d, 0xe8000000},
1436 {0x0000007e, 0x044408a8},
1437 {0x0000007f, 0x00000003},
1438 {0x00000080, 0x00000000},
1439 {0x00000081, 0x01000000},
1440 {0x00000082, 0x02000000},
1441 {0x00000083, 0x00000000},
1442 {0x00000084, 0xe3f3e4f4},
1443 {0x00000085, 0x00052024},
1444 {0x00000087, 0x00000000},
1445 {0x00000088, 0x66036603},
1446 {0x00000089, 0x01000000},
1447 {0x0000008b, 0x1c0a0000},
1448 {0x0000008c, 0xff010000},
1449 {0x0000008e, 0xffffefff},
1450 {0x0000008f, 0xfff3efff},
1451 {0x00000090, 0xfff3efbf},
1452 {0x00000094, 0x00101101},
1453 {0x00000095, 0x00000fff},
1454 {0x00000096, 0x00116fff},
1455 {0x00000097, 0x60010000},
1456 {0x00000098, 0x10010000},
1457 {0x00000099, 0x00006000},
1458 {0x0000009a, 0x00001000},
1459 {0x0000009f, 0x00a07730}
1460};
1461
8b074dd6
AD
1462/* ucode loading */
1463static int si_mc_load_microcode(struct radeon_device *rdev)
1464{
1465 const __be32 *fw_data;
1466 u32 running, blackout = 0;
1467 u32 *io_mc_regs;
1468 int i, ucode_size, regs_size;
1469
1470 if (!rdev->mc_fw)
1471 return -EINVAL;
1472
1473 switch (rdev->family) {
1474 case CHIP_TAHITI:
1475 io_mc_regs = (u32 *)&tahiti_io_mc_regs;
1476 ucode_size = SI_MC_UCODE_SIZE;
1477 regs_size = TAHITI_IO_MC_REGS_SIZE;
1478 break;
1479 case CHIP_PITCAIRN:
1480 io_mc_regs = (u32 *)&pitcairn_io_mc_regs;
1481 ucode_size = SI_MC_UCODE_SIZE;
1482 regs_size = TAHITI_IO_MC_REGS_SIZE;
1483 break;
1484 case CHIP_VERDE:
1485 default:
1486 io_mc_regs = (u32 *)&verde_io_mc_regs;
1487 ucode_size = SI_MC_UCODE_SIZE;
1488 regs_size = TAHITI_IO_MC_REGS_SIZE;
1489 break;
bcc7f5d2
AD
1490 case CHIP_OLAND:
1491 io_mc_regs = (u32 *)&oland_io_mc_regs;
1492 ucode_size = OLAND_MC_UCODE_SIZE;
1493 regs_size = TAHITI_IO_MC_REGS_SIZE;
1494 break;
c04c00b4
AD
1495 case CHIP_HAINAN:
1496 io_mc_regs = (u32 *)&hainan_io_mc_regs;
1497 ucode_size = OLAND_MC_UCODE_SIZE;
1498 regs_size = TAHITI_IO_MC_REGS_SIZE;
1499 break;
8b074dd6
AD
1500 }
1501
1502 running = RREG32(MC_SEQ_SUP_CNTL) & RUN_MASK;
1503
1504 if (running == 0) {
1505 if (running) {
1506 blackout = RREG32(MC_SHARED_BLACKOUT_CNTL);
1507 WREG32(MC_SHARED_BLACKOUT_CNTL, blackout | 1);
1508 }
1509
1510 /* reset the engine and set to writable */
1511 WREG32(MC_SEQ_SUP_CNTL, 0x00000008);
1512 WREG32(MC_SEQ_SUP_CNTL, 0x00000010);
1513
1514 /* load mc io regs */
1515 for (i = 0; i < regs_size; i++) {
1516 WREG32(MC_SEQ_IO_DEBUG_INDEX, io_mc_regs[(i << 1)]);
1517 WREG32(MC_SEQ_IO_DEBUG_DATA, io_mc_regs[(i << 1) + 1]);
1518 }
1519 /* load the MC ucode */
1520 fw_data = (const __be32 *)rdev->mc_fw->data;
1521 for (i = 0; i < ucode_size; i++)
1522 WREG32(MC_SEQ_SUP_PGM, be32_to_cpup(fw_data++));
1523
1524 /* put the engine back into the active state */
1525 WREG32(MC_SEQ_SUP_CNTL, 0x00000008);
1526 WREG32(MC_SEQ_SUP_CNTL, 0x00000004);
1527 WREG32(MC_SEQ_SUP_CNTL, 0x00000001);
1528
1529 /* wait for training to complete */
1530 for (i = 0; i < rdev->usec_timeout; i++) {
1531 if (RREG32(MC_SEQ_TRAIN_WAKEUP_CNTL) & TRAIN_DONE_D0)
1532 break;
1533 udelay(1);
1534 }
1535 for (i = 0; i < rdev->usec_timeout; i++) {
1536 if (RREG32(MC_SEQ_TRAIN_WAKEUP_CNTL) & TRAIN_DONE_D1)
1537 break;
1538 udelay(1);
1539 }
1540
1541 if (running)
1542 WREG32(MC_SHARED_BLACKOUT_CNTL, blackout);
1543 }
1544
1545 return 0;
1546}
1547
0f0de06c
AD
1548static int si_init_microcode(struct radeon_device *rdev)
1549{
0f0de06c
AD
1550 const char *chip_name;
1551 const char *rlc_chip_name;
1552 size_t pfp_req_size, me_req_size, ce_req_size, rlc_req_size, mc_req_size;
a9e61410 1553 size_t smc_req_size;
0f0de06c
AD
1554 char fw_name[30];
1555 int err;
1556
1557 DRM_DEBUG("\n");
1558
0f0de06c
AD
1559 switch (rdev->family) {
1560 case CHIP_TAHITI:
1561 chip_name = "TAHITI";
1562 rlc_chip_name = "TAHITI";
1563 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
1564 me_req_size = SI_PM4_UCODE_SIZE * 4;
1565 ce_req_size = SI_CE_UCODE_SIZE * 4;
1566 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
1567 mc_req_size = SI_MC_UCODE_SIZE * 4;
a9e61410 1568 smc_req_size = ALIGN(TAHITI_SMC_UCODE_SIZE, 4);
0f0de06c
AD
1569 break;
1570 case CHIP_PITCAIRN:
1571 chip_name = "PITCAIRN";
1572 rlc_chip_name = "PITCAIRN";
1573 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
1574 me_req_size = SI_PM4_UCODE_SIZE * 4;
1575 ce_req_size = SI_CE_UCODE_SIZE * 4;
1576 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
1577 mc_req_size = SI_MC_UCODE_SIZE * 4;
a9e61410 1578 smc_req_size = ALIGN(PITCAIRN_SMC_UCODE_SIZE, 4);
0f0de06c
AD
1579 break;
1580 case CHIP_VERDE:
1581 chip_name = "VERDE";
1582 rlc_chip_name = "VERDE";
1583 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
1584 me_req_size = SI_PM4_UCODE_SIZE * 4;
1585 ce_req_size = SI_CE_UCODE_SIZE * 4;
1586 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
1587 mc_req_size = SI_MC_UCODE_SIZE * 4;
a9e61410 1588 smc_req_size = ALIGN(VERDE_SMC_UCODE_SIZE, 4);
0f0de06c 1589 break;
bcc7f5d2
AD
1590 case CHIP_OLAND:
1591 chip_name = "OLAND";
1592 rlc_chip_name = "OLAND";
1593 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
1594 me_req_size = SI_PM4_UCODE_SIZE * 4;
1595 ce_req_size = SI_CE_UCODE_SIZE * 4;
1596 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
1597 mc_req_size = OLAND_MC_UCODE_SIZE * 4;
a9e61410 1598 smc_req_size = ALIGN(OLAND_SMC_UCODE_SIZE, 4);
bcc7f5d2 1599 break;
c04c00b4
AD
1600 case CHIP_HAINAN:
1601 chip_name = "HAINAN";
1602 rlc_chip_name = "HAINAN";
1603 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
1604 me_req_size = SI_PM4_UCODE_SIZE * 4;
1605 ce_req_size = SI_CE_UCODE_SIZE * 4;
1606 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
1607 mc_req_size = OLAND_MC_UCODE_SIZE * 4;
a9e61410 1608 smc_req_size = ALIGN(HAINAN_SMC_UCODE_SIZE, 4);
c04c00b4 1609 break;
0f0de06c
AD
1610 default: BUG();
1611 }
1612
1613 DRM_INFO("Loading %s Microcode\n", chip_name);
1614
1615 snprintf(fw_name, sizeof(fw_name), "radeon/%s_pfp.bin", chip_name);
0a168933 1616 err = request_firmware(&rdev->pfp_fw, fw_name, rdev->dev);
0f0de06c
AD
1617 if (err)
1618 goto out;
1619 if (rdev->pfp_fw->size != pfp_req_size) {
1620 printk(KERN_ERR
1621 "si_cp: Bogus length %zu in firmware \"%s\"\n",
1622 rdev->pfp_fw->size, fw_name);
1623 err = -EINVAL;
1624 goto out;
1625 }
1626
1627 snprintf(fw_name, sizeof(fw_name), "radeon/%s_me.bin", chip_name);
0a168933 1628 err = request_firmware(&rdev->me_fw, fw_name, rdev->dev);
0f0de06c
AD
1629 if (err)
1630 goto out;
1631 if (rdev->me_fw->size != me_req_size) {
1632 printk(KERN_ERR
1633 "si_cp: Bogus length %zu in firmware \"%s\"\n",
1634 rdev->me_fw->size, fw_name);
1635 err = -EINVAL;
1636 }
1637
1638 snprintf(fw_name, sizeof(fw_name), "radeon/%s_ce.bin", chip_name);
0a168933 1639 err = request_firmware(&rdev->ce_fw, fw_name, rdev->dev);
0f0de06c
AD
1640 if (err)
1641 goto out;
1642 if (rdev->ce_fw->size != ce_req_size) {
1643 printk(KERN_ERR
1644 "si_cp: Bogus length %zu in firmware \"%s\"\n",
1645 rdev->ce_fw->size, fw_name);
1646 err = -EINVAL;
1647 }
1648
1649 snprintf(fw_name, sizeof(fw_name), "radeon/%s_rlc.bin", rlc_chip_name);
0a168933 1650 err = request_firmware(&rdev->rlc_fw, fw_name, rdev->dev);
0f0de06c
AD
1651 if (err)
1652 goto out;
1653 if (rdev->rlc_fw->size != rlc_req_size) {
1654 printk(KERN_ERR
1655 "si_rlc: Bogus length %zu in firmware \"%s\"\n",
1656 rdev->rlc_fw->size, fw_name);
1657 err = -EINVAL;
1658 }
1659
1660 snprintf(fw_name, sizeof(fw_name), "radeon/%s_mc.bin", chip_name);
0a168933 1661 err = request_firmware(&rdev->mc_fw, fw_name, rdev->dev);
0f0de06c
AD
1662 if (err)
1663 goto out;
1664 if (rdev->mc_fw->size != mc_req_size) {
1665 printk(KERN_ERR
1666 "si_mc: Bogus length %zu in firmware \"%s\"\n",
1667 rdev->mc_fw->size, fw_name);
1668 err = -EINVAL;
1669 }
1670
a9e61410 1671 snprintf(fw_name, sizeof(fw_name), "radeon/%s_smc.bin", chip_name);
0a168933 1672 err = request_firmware(&rdev->smc_fw, fw_name, rdev->dev);
8a53fa23
AD
1673 if (err) {
1674 printk(KERN_ERR
1675 "smc: error loading firmware \"%s\"\n",
1676 fw_name);
1677 release_firmware(rdev->smc_fw);
1678 rdev->smc_fw = NULL;
1679 } else if (rdev->smc_fw->size != smc_req_size) {
a9e61410
AD
1680 printk(KERN_ERR
1681 "si_smc: Bogus length %zu in firmware \"%s\"\n",
1682 rdev->smc_fw->size, fw_name);
1683 err = -EINVAL;
1684 }
1685
0f0de06c 1686out:
0f0de06c
AD
1687 if (err) {
1688 if (err != -EINVAL)
1689 printk(KERN_ERR
1690 "si_cp: Failed to load firmware \"%s\"\n",
1691 fw_name);
1692 release_firmware(rdev->pfp_fw);
1693 rdev->pfp_fw = NULL;
1694 release_firmware(rdev->me_fw);
1695 rdev->me_fw = NULL;
1696 release_firmware(rdev->ce_fw);
1697 rdev->ce_fw = NULL;
1698 release_firmware(rdev->rlc_fw);
1699 rdev->rlc_fw = NULL;
1700 release_firmware(rdev->mc_fw);
1701 rdev->mc_fw = NULL;
a9e61410
AD
1702 release_firmware(rdev->smc_fw);
1703 rdev->smc_fw = NULL;
0f0de06c
AD
1704 }
1705 return err;
1706}
1707
43b3cd99
AD
1708/* watermark setup */
1709static u32 dce6_line_buffer_adjust(struct radeon_device *rdev,
1710 struct radeon_crtc *radeon_crtc,
1711 struct drm_display_mode *mode,
1712 struct drm_display_mode *other_mode)
1713{
1714 u32 tmp;
1715 /*
1716 * Line Buffer Setup
1717 * There are 3 line buffers, each one shared by 2 display controllers.
1718 * DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
1719 * the display controllers. The paritioning is done via one of four
1720 * preset allocations specified in bits 21:20:
1721 * 0 - half lb
1722 * 2 - whole lb, other crtc must be disabled
1723 */
1724 /* this can get tricky if we have two large displays on a paired group
1725 * of crtcs. Ideally for multiple large displays we'd assign them to
1726 * non-linked crtcs for maximum line buffer allocation.
1727 */
1728 if (radeon_crtc->base.enabled && mode) {
1729 if (other_mode)
1730 tmp = 0; /* 1/2 */
1731 else
1732 tmp = 2; /* whole */
1733 } else
1734 tmp = 0;
1735
1736 WREG32(DC_LB_MEMORY_SPLIT + radeon_crtc->crtc_offset,
1737 DC_LB_MEMORY_CONFIG(tmp));
1738
1739 if (radeon_crtc->base.enabled && mode) {
1740 switch (tmp) {
1741 case 0:
1742 default:
1743 return 4096 * 2;
1744 case 2:
1745 return 8192 * 2;
1746 }
1747 }
1748
1749 /* controller not enabled, so no lb used */
1750 return 0;
1751}
1752
ca7db22b 1753static u32 si_get_number_of_dram_channels(struct radeon_device *rdev)
43b3cd99
AD
1754{
1755 u32 tmp = RREG32(MC_SHARED_CHMAP);
1756
1757 switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
1758 case 0:
1759 default:
1760 return 1;
1761 case 1:
1762 return 2;
1763 case 2:
1764 return 4;
1765 case 3:
1766 return 8;
1767 case 4:
1768 return 3;
1769 case 5:
1770 return 6;
1771 case 6:
1772 return 10;
1773 case 7:
1774 return 12;
1775 case 8:
1776 return 16;
1777 }
1778}
1779
1780struct dce6_wm_params {
1781 u32 dram_channels; /* number of dram channels */
1782 u32 yclk; /* bandwidth per dram data pin in kHz */
1783 u32 sclk; /* engine clock in kHz */
1784 u32 disp_clk; /* display clock in kHz */
1785 u32 src_width; /* viewport width */
1786 u32 active_time; /* active display time in ns */
1787 u32 blank_time; /* blank time in ns */
1788 bool interlaced; /* mode is interlaced */
1789 fixed20_12 vsc; /* vertical scale ratio */
1790 u32 num_heads; /* number of active crtcs */
1791 u32 bytes_per_pixel; /* bytes per pixel display + overlay */
1792 u32 lb_size; /* line buffer allocated to pipe */
1793 u32 vtaps; /* vertical scaler taps */
1794};
1795
1796static u32 dce6_dram_bandwidth(struct dce6_wm_params *wm)
1797{
1798 /* Calculate raw DRAM Bandwidth */
1799 fixed20_12 dram_efficiency; /* 0.7 */
1800 fixed20_12 yclk, dram_channels, bandwidth;
1801 fixed20_12 a;
1802
1803 a.full = dfixed_const(1000);
1804 yclk.full = dfixed_const(wm->yclk);
1805 yclk.full = dfixed_div(yclk, a);
1806 dram_channels.full = dfixed_const(wm->dram_channels * 4);
1807 a.full = dfixed_const(10);
1808 dram_efficiency.full = dfixed_const(7);
1809 dram_efficiency.full = dfixed_div(dram_efficiency, a);
1810 bandwidth.full = dfixed_mul(dram_channels, yclk);
1811 bandwidth.full = dfixed_mul(bandwidth, dram_efficiency);
1812
1813 return dfixed_trunc(bandwidth);
1814}
1815
1816static u32 dce6_dram_bandwidth_for_display(struct dce6_wm_params *wm)
1817{
1818 /* Calculate DRAM Bandwidth and the part allocated to display. */
1819 fixed20_12 disp_dram_allocation; /* 0.3 to 0.7 */
1820 fixed20_12 yclk, dram_channels, bandwidth;
1821 fixed20_12 a;
1822
1823 a.full = dfixed_const(1000);
1824 yclk.full = dfixed_const(wm->yclk);
1825 yclk.full = dfixed_div(yclk, a);
1826 dram_channels.full = dfixed_const(wm->dram_channels * 4);
1827 a.full = dfixed_const(10);
1828 disp_dram_allocation.full = dfixed_const(3); /* XXX worse case value 0.3 */
1829 disp_dram_allocation.full = dfixed_div(disp_dram_allocation, a);
1830 bandwidth.full = dfixed_mul(dram_channels, yclk);
1831 bandwidth.full = dfixed_mul(bandwidth, disp_dram_allocation);
1832
1833 return dfixed_trunc(bandwidth);
1834}
1835
1836static u32 dce6_data_return_bandwidth(struct dce6_wm_params *wm)
1837{
1838 /* Calculate the display Data return Bandwidth */
1839 fixed20_12 return_efficiency; /* 0.8 */
1840 fixed20_12 sclk, bandwidth;
1841 fixed20_12 a;
1842
1843 a.full = dfixed_const(1000);
1844 sclk.full = dfixed_const(wm->sclk);
1845 sclk.full = dfixed_div(sclk, a);
1846 a.full = dfixed_const(10);
1847 return_efficiency.full = dfixed_const(8);
1848 return_efficiency.full = dfixed_div(return_efficiency, a);
1849 a.full = dfixed_const(32);
1850 bandwidth.full = dfixed_mul(a, sclk);
1851 bandwidth.full = dfixed_mul(bandwidth, return_efficiency);
1852
1853 return dfixed_trunc(bandwidth);
1854}
1855
1856static u32 dce6_get_dmif_bytes_per_request(struct dce6_wm_params *wm)
1857{
1858 return 32;
1859}
1860
1861static u32 dce6_dmif_request_bandwidth(struct dce6_wm_params *wm)
1862{
1863 /* Calculate the DMIF Request Bandwidth */
1864 fixed20_12 disp_clk_request_efficiency; /* 0.8 */
1865 fixed20_12 disp_clk, sclk, bandwidth;
1866 fixed20_12 a, b1, b2;
1867 u32 min_bandwidth;
1868
1869 a.full = dfixed_const(1000);
1870 disp_clk.full = dfixed_const(wm->disp_clk);
1871 disp_clk.full = dfixed_div(disp_clk, a);
1872 a.full = dfixed_const(dce6_get_dmif_bytes_per_request(wm) / 2);
1873 b1.full = dfixed_mul(a, disp_clk);
1874
1875 a.full = dfixed_const(1000);
1876 sclk.full = dfixed_const(wm->sclk);
1877 sclk.full = dfixed_div(sclk, a);
1878 a.full = dfixed_const(dce6_get_dmif_bytes_per_request(wm));
1879 b2.full = dfixed_mul(a, sclk);
1880
1881 a.full = dfixed_const(10);
1882 disp_clk_request_efficiency.full = dfixed_const(8);
1883 disp_clk_request_efficiency.full = dfixed_div(disp_clk_request_efficiency, a);
1884
1885 min_bandwidth = min(dfixed_trunc(b1), dfixed_trunc(b2));
1886
1887 a.full = dfixed_const(min_bandwidth);
1888 bandwidth.full = dfixed_mul(a, disp_clk_request_efficiency);
1889
1890 return dfixed_trunc(bandwidth);
1891}
1892
1893static u32 dce6_available_bandwidth(struct dce6_wm_params *wm)
1894{
1895 /* Calculate the Available bandwidth. Display can use this temporarily but not in average. */
1896 u32 dram_bandwidth = dce6_dram_bandwidth(wm);
1897 u32 data_return_bandwidth = dce6_data_return_bandwidth(wm);
1898 u32 dmif_req_bandwidth = dce6_dmif_request_bandwidth(wm);
1899
1900 return min(dram_bandwidth, min(data_return_bandwidth, dmif_req_bandwidth));
1901}
1902
1903static u32 dce6_average_bandwidth(struct dce6_wm_params *wm)
1904{
1905 /* Calculate the display mode Average Bandwidth
1906 * DisplayMode should contain the source and destination dimensions,
1907 * timing, etc.
1908 */
1909 fixed20_12 bpp;
1910 fixed20_12 line_time;
1911 fixed20_12 src_width;
1912 fixed20_12 bandwidth;
1913 fixed20_12 a;
1914
1915 a.full = dfixed_const(1000);
1916 line_time.full = dfixed_const(wm->active_time + wm->blank_time);
1917 line_time.full = dfixed_div(line_time, a);
1918 bpp.full = dfixed_const(wm->bytes_per_pixel);
1919 src_width.full = dfixed_const(wm->src_width);
1920 bandwidth.full = dfixed_mul(src_width, bpp);
1921 bandwidth.full = dfixed_mul(bandwidth, wm->vsc);
1922 bandwidth.full = dfixed_div(bandwidth, line_time);
1923
1924 return dfixed_trunc(bandwidth);
1925}
1926
1927static u32 dce6_latency_watermark(struct dce6_wm_params *wm)
1928{
1929 /* First calcualte the latency in ns */
1930 u32 mc_latency = 2000; /* 2000 ns. */
1931 u32 available_bandwidth = dce6_available_bandwidth(wm);
1932 u32 worst_chunk_return_time = (512 * 8 * 1000) / available_bandwidth;
1933 u32 cursor_line_pair_return_time = (128 * 4 * 1000) / available_bandwidth;
1934 u32 dc_latency = 40000000 / wm->disp_clk; /* dc pipe latency */
1935 u32 other_heads_data_return_time = ((wm->num_heads + 1) * worst_chunk_return_time) +
1936 (wm->num_heads * cursor_line_pair_return_time);
1937 u32 latency = mc_latency + other_heads_data_return_time + dc_latency;
1938 u32 max_src_lines_per_dst_line, lb_fill_bw, line_fill_time;
1939 u32 tmp, dmif_size = 12288;
1940 fixed20_12 a, b, c;
1941
1942 if (wm->num_heads == 0)
1943 return 0;
1944
1945 a.full = dfixed_const(2);
1946 b.full = dfixed_const(1);
1947 if ((wm->vsc.full > a.full) ||
1948 ((wm->vsc.full > b.full) && (wm->vtaps >= 3)) ||
1949 (wm->vtaps >= 5) ||
1950 ((wm->vsc.full >= a.full) && wm->interlaced))
1951 max_src_lines_per_dst_line = 4;
1952 else
1953 max_src_lines_per_dst_line = 2;
1954
1955 a.full = dfixed_const(available_bandwidth);
1956 b.full = dfixed_const(wm->num_heads);
1957 a.full = dfixed_div(a, b);
1958
1959 b.full = dfixed_const(mc_latency + 512);
1960 c.full = dfixed_const(wm->disp_clk);
1961 b.full = dfixed_div(b, c);
1962
1963 c.full = dfixed_const(dmif_size);
1964 b.full = dfixed_div(c, b);
1965
1966 tmp = min(dfixed_trunc(a), dfixed_trunc(b));
1967
1968 b.full = dfixed_const(1000);
1969 c.full = dfixed_const(wm->disp_clk);
1970 b.full = dfixed_div(c, b);
1971 c.full = dfixed_const(wm->bytes_per_pixel);
1972 b.full = dfixed_mul(b, c);
1973
1974 lb_fill_bw = min(tmp, dfixed_trunc(b));
1975
1976 a.full = dfixed_const(max_src_lines_per_dst_line * wm->src_width * wm->bytes_per_pixel);
1977 b.full = dfixed_const(1000);
1978 c.full = dfixed_const(lb_fill_bw);
1979 b.full = dfixed_div(c, b);
1980 a.full = dfixed_div(a, b);
1981 line_fill_time = dfixed_trunc(a);
1982
1983 if (line_fill_time < wm->active_time)
1984 return latency;
1985 else
1986 return latency + (line_fill_time - wm->active_time);
1987
1988}
1989
1990static bool dce6_average_bandwidth_vs_dram_bandwidth_for_display(struct dce6_wm_params *wm)
1991{
1992 if (dce6_average_bandwidth(wm) <=
1993 (dce6_dram_bandwidth_for_display(wm) / wm->num_heads))
1994 return true;
1995 else
1996 return false;
1997};
1998
1999static bool dce6_average_bandwidth_vs_available_bandwidth(struct dce6_wm_params *wm)
2000{
2001 if (dce6_average_bandwidth(wm) <=
2002 (dce6_available_bandwidth(wm) / wm->num_heads))
2003 return true;
2004 else
2005 return false;
2006};
2007
2008static bool dce6_check_latency_hiding(struct dce6_wm_params *wm)
2009{
2010 u32 lb_partitions = wm->lb_size / wm->src_width;
2011 u32 line_time = wm->active_time + wm->blank_time;
2012 u32 latency_tolerant_lines;
2013 u32 latency_hiding;
2014 fixed20_12 a;
2015
2016 a.full = dfixed_const(1);
2017 if (wm->vsc.full > a.full)
2018 latency_tolerant_lines = 1;
2019 else {
2020 if (lb_partitions <= (wm->vtaps + 1))
2021 latency_tolerant_lines = 1;
2022 else
2023 latency_tolerant_lines = 2;
2024 }
2025
2026 latency_hiding = (latency_tolerant_lines * line_time + wm->blank_time);
2027
2028 if (dce6_latency_watermark(wm) <= latency_hiding)
2029 return true;
2030 else
2031 return false;
2032}
2033
2034static void dce6_program_watermarks(struct radeon_device *rdev,
2035 struct radeon_crtc *radeon_crtc,
2036 u32 lb_size, u32 num_heads)
2037{
2038 struct drm_display_mode *mode = &radeon_crtc->base.mode;
c696e53f
AD
2039 struct dce6_wm_params wm_low, wm_high;
2040 u32 dram_channels;
43b3cd99
AD
2041 u32 pixel_period;
2042 u32 line_time = 0;
2043 u32 latency_watermark_a = 0, latency_watermark_b = 0;
2044 u32 priority_a_mark = 0, priority_b_mark = 0;
2045 u32 priority_a_cnt = PRIORITY_OFF;
2046 u32 priority_b_cnt = PRIORITY_OFF;
2047 u32 tmp, arb_control3;
2048 fixed20_12 a, b, c;
2049
2050 if (radeon_crtc->base.enabled && num_heads && mode) {
2051 pixel_period = 1000000 / (u32)mode->clock;
2052 line_time = min((u32)mode->crtc_htotal * pixel_period, (u32)65535);
2053 priority_a_cnt = 0;
2054 priority_b_cnt = 0;
2055
ca7db22b 2056 if (rdev->family == CHIP_ARUBA)
c696e53f 2057 dram_channels = evergreen_get_number_of_dram_channels(rdev);
ca7db22b 2058 else
c696e53f
AD
2059 dram_channels = si_get_number_of_dram_channels(rdev);
2060
2061 /* watermark for high clocks */
2062 if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
2063 wm_high.yclk =
2064 radeon_dpm_get_mclk(rdev, false) * 10;
2065 wm_high.sclk =
2066 radeon_dpm_get_sclk(rdev, false) * 10;
2067 } else {
2068 wm_high.yclk = rdev->pm.current_mclk * 10;
2069 wm_high.sclk = rdev->pm.current_sclk * 10;
2070 }
2071
2072 wm_high.disp_clk = mode->clock;
2073 wm_high.src_width = mode->crtc_hdisplay;
2074 wm_high.active_time = mode->crtc_hdisplay * pixel_period;
2075 wm_high.blank_time = line_time - wm_high.active_time;
2076 wm_high.interlaced = false;
2077 if (mode->flags & DRM_MODE_FLAG_INTERLACE)
2078 wm_high.interlaced = true;
2079 wm_high.vsc = radeon_crtc->vsc;
2080 wm_high.vtaps = 1;
2081 if (radeon_crtc->rmx_type != RMX_OFF)
2082 wm_high.vtaps = 2;
2083 wm_high.bytes_per_pixel = 4; /* XXX: get this from fb config */
2084 wm_high.lb_size = lb_size;
2085 wm_high.dram_channels = dram_channels;
2086 wm_high.num_heads = num_heads;
2087
2088 /* watermark for low clocks */
2089 if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
2090 wm_low.yclk =
2091 radeon_dpm_get_mclk(rdev, true) * 10;
2092 wm_low.sclk =
2093 radeon_dpm_get_sclk(rdev, true) * 10;
2094 } else {
2095 wm_low.yclk = rdev->pm.current_mclk * 10;
2096 wm_low.sclk = rdev->pm.current_sclk * 10;
2097 }
2098
2099 wm_low.disp_clk = mode->clock;
2100 wm_low.src_width = mode->crtc_hdisplay;
2101 wm_low.active_time = mode->crtc_hdisplay * pixel_period;
2102 wm_low.blank_time = line_time - wm_low.active_time;
2103 wm_low.interlaced = false;
2104 if (mode->flags & DRM_MODE_FLAG_INTERLACE)
2105 wm_low.interlaced = true;
2106 wm_low.vsc = radeon_crtc->vsc;
2107 wm_low.vtaps = 1;
2108 if (radeon_crtc->rmx_type != RMX_OFF)
2109 wm_low.vtaps = 2;
2110 wm_low.bytes_per_pixel = 4; /* XXX: get this from fb config */
2111 wm_low.lb_size = lb_size;
2112 wm_low.dram_channels = dram_channels;
2113 wm_low.num_heads = num_heads;
43b3cd99
AD
2114
2115 /* set for high clocks */
c696e53f 2116 latency_watermark_a = min(dce6_latency_watermark(&wm_high), (u32)65535);
43b3cd99 2117 /* set for low clocks */
c696e53f 2118 latency_watermark_b = min(dce6_latency_watermark(&wm_low), (u32)65535);
43b3cd99
AD
2119
2120 /* possibly force display priority to high */
2121 /* should really do this at mode validation time... */
c696e53f
AD
2122 if (!dce6_average_bandwidth_vs_dram_bandwidth_for_display(&wm_high) ||
2123 !dce6_average_bandwidth_vs_available_bandwidth(&wm_high) ||
2124 !dce6_check_latency_hiding(&wm_high) ||
2125 (rdev->disp_priority == 2)) {
2126 DRM_DEBUG_KMS("force priority to high\n");
2127 priority_a_cnt |= PRIORITY_ALWAYS_ON;
2128 priority_b_cnt |= PRIORITY_ALWAYS_ON;
2129 }
2130 if (!dce6_average_bandwidth_vs_dram_bandwidth_for_display(&wm_low) ||
2131 !dce6_average_bandwidth_vs_available_bandwidth(&wm_low) ||
2132 !dce6_check_latency_hiding(&wm_low) ||
43b3cd99
AD
2133 (rdev->disp_priority == 2)) {
2134 DRM_DEBUG_KMS("force priority to high\n");
2135 priority_a_cnt |= PRIORITY_ALWAYS_ON;
2136 priority_b_cnt |= PRIORITY_ALWAYS_ON;
2137 }
2138
2139 a.full = dfixed_const(1000);
2140 b.full = dfixed_const(mode->clock);
2141 b.full = dfixed_div(b, a);
2142 c.full = dfixed_const(latency_watermark_a);
2143 c.full = dfixed_mul(c, b);
2144 c.full = dfixed_mul(c, radeon_crtc->hsc);
2145 c.full = dfixed_div(c, a);
2146 a.full = dfixed_const(16);
2147 c.full = dfixed_div(c, a);
2148 priority_a_mark = dfixed_trunc(c);
2149 priority_a_cnt |= priority_a_mark & PRIORITY_MARK_MASK;
2150
2151 a.full = dfixed_const(1000);
2152 b.full = dfixed_const(mode->clock);
2153 b.full = dfixed_div(b, a);
2154 c.full = dfixed_const(latency_watermark_b);
2155 c.full = dfixed_mul(c, b);
2156 c.full = dfixed_mul(c, radeon_crtc->hsc);
2157 c.full = dfixed_div(c, a);
2158 a.full = dfixed_const(16);
2159 c.full = dfixed_div(c, a);
2160 priority_b_mark = dfixed_trunc(c);
2161 priority_b_cnt |= priority_b_mark & PRIORITY_MARK_MASK;
2162 }
2163
2164 /* select wm A */
2165 arb_control3 = RREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset);
2166 tmp = arb_control3;
2167 tmp &= ~LATENCY_WATERMARK_MASK(3);
2168 tmp |= LATENCY_WATERMARK_MASK(1);
2169 WREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset, tmp);
2170 WREG32(DPG_PIPE_LATENCY_CONTROL + radeon_crtc->crtc_offset,
2171 (LATENCY_LOW_WATERMARK(latency_watermark_a) |
2172 LATENCY_HIGH_WATERMARK(line_time)));
2173 /* select wm B */
2174 tmp = RREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset);
2175 tmp &= ~LATENCY_WATERMARK_MASK(3);
2176 tmp |= LATENCY_WATERMARK_MASK(2);
2177 WREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset, tmp);
2178 WREG32(DPG_PIPE_LATENCY_CONTROL + radeon_crtc->crtc_offset,
2179 (LATENCY_LOW_WATERMARK(latency_watermark_b) |
2180 LATENCY_HIGH_WATERMARK(line_time)));
2181 /* restore original selection */
2182 WREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset, arb_control3);
2183
2184 /* write the priority marks */
2185 WREG32(PRIORITY_A_CNT + radeon_crtc->crtc_offset, priority_a_cnt);
2186 WREG32(PRIORITY_B_CNT + radeon_crtc->crtc_offset, priority_b_cnt);
2187
7178d2a6
AD
2188 /* save values for DPM */
2189 radeon_crtc->line_time = line_time;
2190 radeon_crtc->wm_high = latency_watermark_a;
2191 radeon_crtc->wm_low = latency_watermark_b;
43b3cd99
AD
2192}
2193
2194void dce6_bandwidth_update(struct radeon_device *rdev)
2195{
2196 struct drm_display_mode *mode0 = NULL;
2197 struct drm_display_mode *mode1 = NULL;
2198 u32 num_heads = 0, lb_size;
2199 int i;
2200
2201 radeon_update_display_priority(rdev);
2202
2203 for (i = 0; i < rdev->num_crtc; i++) {
2204 if (rdev->mode_info.crtcs[i]->base.enabled)
2205 num_heads++;
2206 }
2207 for (i = 0; i < rdev->num_crtc; i += 2) {
2208 mode0 = &rdev->mode_info.crtcs[i]->base.mode;
2209 mode1 = &rdev->mode_info.crtcs[i+1]->base.mode;
2210 lb_size = dce6_line_buffer_adjust(rdev, rdev->mode_info.crtcs[i], mode0, mode1);
2211 dce6_program_watermarks(rdev, rdev->mode_info.crtcs[i], lb_size, num_heads);
2212 lb_size = dce6_line_buffer_adjust(rdev, rdev->mode_info.crtcs[i+1], mode1, mode0);
2213 dce6_program_watermarks(rdev, rdev->mode_info.crtcs[i+1], lb_size, num_heads);
2214 }
2215}
2216
0a96d72b
AD
2217/*
2218 * Core functions
2219 */
0a96d72b
AD
2220static void si_tiling_mode_table_init(struct radeon_device *rdev)
2221{
2222 const u32 num_tile_mode_states = 32;
2223 u32 reg_offset, gb_tile_moden, split_equal_to_row_size;
2224
2225 switch (rdev->config.si.mem_row_size_in_kb) {
2226 case 1:
2227 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_1KB;
2228 break;
2229 case 2:
2230 default:
2231 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_2KB;
2232 break;
2233 case 4:
2234 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_4KB;
2235 break;
2236 }
2237
2238 if ((rdev->family == CHIP_TAHITI) ||
2239 (rdev->family == CHIP_PITCAIRN)) {
2240 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
2241 switch (reg_offset) {
2242 case 0: /* non-AA compressed depth or any compressed stencil */
2243 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2244 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2245 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2246 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2247 NUM_BANKS(ADDR_SURF_16_BANK) |
2248 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2249 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2250 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2251 break;
2252 case 1: /* 2xAA/4xAA compressed depth only */
2253 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2254 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2255 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2256 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2257 NUM_BANKS(ADDR_SURF_16_BANK) |
2258 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2259 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2260 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2261 break;
2262 case 2: /* 8xAA compressed depth only */
2263 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2264 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2265 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2266 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2267 NUM_BANKS(ADDR_SURF_16_BANK) |
2268 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2269 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2270 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2271 break;
2272 case 3: /* 2xAA/4xAA compressed depth with stencil (for depth buffer) */
2273 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2274 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2275 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2276 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2277 NUM_BANKS(ADDR_SURF_16_BANK) |
2278 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2279 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2280 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2281 break;
2282 case 4: /* Maps w/ a dimension less than the 2D macro-tile dimensions (for mipmapped depth textures) */
2283 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2284 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2285 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2286 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2287 NUM_BANKS(ADDR_SURF_16_BANK) |
2288 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2289 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2290 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2291 break;
2292 case 5: /* Uncompressed 16bpp depth - and stencil buffer allocated with it */
2293 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2294 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2295 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2296 TILE_SPLIT(split_equal_to_row_size) |
2297 NUM_BANKS(ADDR_SURF_16_BANK) |
2298 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2299 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2300 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2301 break;
2302 case 6: /* Uncompressed 32bpp depth - and stencil buffer allocated with it */
2303 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2304 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2305 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2306 TILE_SPLIT(split_equal_to_row_size) |
2307 NUM_BANKS(ADDR_SURF_16_BANK) |
2308 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2309 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2310 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2311 break;
2312 case 7: /* Uncompressed 8bpp stencil without depth (drivers typically do not use) */
2313 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2314 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2315 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2316 TILE_SPLIT(split_equal_to_row_size) |
2317 NUM_BANKS(ADDR_SURF_16_BANK) |
2318 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2319 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2320 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2321 break;
2322 case 8: /* 1D and 1D Array Surfaces */
2323 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
2324 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2325 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2326 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2327 NUM_BANKS(ADDR_SURF_16_BANK) |
2328 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2329 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2330 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2331 break;
2332 case 9: /* Displayable maps. */
2333 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2334 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2335 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2336 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2337 NUM_BANKS(ADDR_SURF_16_BANK) |
2338 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2339 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2340 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2341 break;
2342 case 10: /* Display 8bpp. */
2343 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2344 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2345 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2346 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2347 NUM_BANKS(ADDR_SURF_16_BANK) |
2348 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2349 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2350 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2351 break;
2352 case 11: /* Display 16bpp. */
2353 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2354 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2355 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2356 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2357 NUM_BANKS(ADDR_SURF_16_BANK) |
2358 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2359 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2360 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2361 break;
2362 case 12: /* Display 32bpp. */
2363 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2364 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2365 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2366 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2367 NUM_BANKS(ADDR_SURF_16_BANK) |
2368 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2369 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2370 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2371 break;
2372 case 13: /* Thin. */
2373 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2374 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2375 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2376 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2377 NUM_BANKS(ADDR_SURF_16_BANK) |
2378 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2379 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2380 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2381 break;
2382 case 14: /* Thin 8 bpp. */
2383 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2384 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2385 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2386 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2387 NUM_BANKS(ADDR_SURF_16_BANK) |
2388 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2389 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2390 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2391 break;
2392 case 15: /* Thin 16 bpp. */
2393 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2394 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2395 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2396 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2397 NUM_BANKS(ADDR_SURF_16_BANK) |
2398 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2399 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2400 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2401 break;
2402 case 16: /* Thin 32 bpp. */
2403 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2404 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2405 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2406 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2407 NUM_BANKS(ADDR_SURF_16_BANK) |
2408 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2409 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2410 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2411 break;
2412 case 17: /* Thin 64 bpp. */
2413 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2414 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2415 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2416 TILE_SPLIT(split_equal_to_row_size) |
2417 NUM_BANKS(ADDR_SURF_16_BANK) |
2418 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2419 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2420 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2421 break;
2422 case 21: /* 8 bpp PRT. */
2423 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2424 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2425 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2426 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2427 NUM_BANKS(ADDR_SURF_16_BANK) |
2428 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
2429 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2430 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2431 break;
2432 case 22: /* 16 bpp PRT */
2433 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2434 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2435 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2436 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2437 NUM_BANKS(ADDR_SURF_16_BANK) |
2438 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2439 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2440 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2441 break;
2442 case 23: /* 32 bpp PRT */
2443 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2444 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2445 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2446 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2447 NUM_BANKS(ADDR_SURF_16_BANK) |
2448 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2449 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2450 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2451 break;
2452 case 24: /* 64 bpp PRT */
2453 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2454 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2455 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2456 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2457 NUM_BANKS(ADDR_SURF_16_BANK) |
2458 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2459 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2460 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2461 break;
2462 case 25: /* 128 bpp PRT */
2463 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2464 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2465 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2466 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
2467 NUM_BANKS(ADDR_SURF_8_BANK) |
2468 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2469 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2470 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2471 break;
2472 default:
2473 gb_tile_moden = 0;
2474 break;
2475 }
64d7b8be 2476 rdev->config.si.tile_mode_array[reg_offset] = gb_tile_moden;
0a96d72b
AD
2477 WREG32(GB_TILE_MODE0 + (reg_offset * 4), gb_tile_moden);
2478 }
d0ae7fcc 2479 } else if ((rdev->family == CHIP_VERDE) ||
8b02859d
AD
2480 (rdev->family == CHIP_OLAND) ||
2481 (rdev->family == CHIP_HAINAN)) {
0a96d72b
AD
2482 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
2483 switch (reg_offset) {
2484 case 0: /* non-AA compressed depth or any compressed stencil */
2485 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2486 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2487 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2488 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2489 NUM_BANKS(ADDR_SURF_16_BANK) |
2490 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2491 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2492 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2493 break;
2494 case 1: /* 2xAA/4xAA compressed depth only */
2495 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2496 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2497 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2498 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2499 NUM_BANKS(ADDR_SURF_16_BANK) |
2500 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2501 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2502 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2503 break;
2504 case 2: /* 8xAA compressed depth only */
2505 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2506 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2507 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2508 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2509 NUM_BANKS(ADDR_SURF_16_BANK) |
2510 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2511 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2512 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2513 break;
2514 case 3: /* 2xAA/4xAA compressed depth with stencil (for depth buffer) */
2515 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2516 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2517 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2518 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2519 NUM_BANKS(ADDR_SURF_16_BANK) |
2520 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2521 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2522 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2523 break;
2524 case 4: /* Maps w/ a dimension less than the 2D macro-tile dimensions (for mipmapped depth textures) */
2525 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2526 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2527 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2528 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2529 NUM_BANKS(ADDR_SURF_16_BANK) |
2530 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2531 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2532 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2533 break;
2534 case 5: /* Uncompressed 16bpp depth - and stencil buffer allocated with it */
2535 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2536 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2537 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2538 TILE_SPLIT(split_equal_to_row_size) |
2539 NUM_BANKS(ADDR_SURF_16_BANK) |
2540 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2541 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2542 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2543 break;
2544 case 6: /* Uncompressed 32bpp depth - and stencil buffer allocated with it */
2545 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2546 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2547 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2548 TILE_SPLIT(split_equal_to_row_size) |
2549 NUM_BANKS(ADDR_SURF_16_BANK) |
2550 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2551 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2552 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2553 break;
2554 case 7: /* Uncompressed 8bpp stencil without depth (drivers typically do not use) */
2555 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2556 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2557 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2558 TILE_SPLIT(split_equal_to_row_size) |
2559 NUM_BANKS(ADDR_SURF_16_BANK) |
2560 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2561 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2562 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2563 break;
2564 case 8: /* 1D and 1D Array Surfaces */
2565 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
2566 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2567 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2568 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2569 NUM_BANKS(ADDR_SURF_16_BANK) |
2570 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2571 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2572 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2573 break;
2574 case 9: /* Displayable maps. */
2575 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2576 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2577 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2578 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2579 NUM_BANKS(ADDR_SURF_16_BANK) |
2580 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2581 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2582 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2583 break;
2584 case 10: /* Display 8bpp. */
2585 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2586 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2587 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2588 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2589 NUM_BANKS(ADDR_SURF_16_BANK) |
2590 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2591 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2592 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2593 break;
2594 case 11: /* Display 16bpp. */
2595 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2596 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2597 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2598 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2599 NUM_BANKS(ADDR_SURF_16_BANK) |
2600 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2601 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2602 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2603 break;
2604 case 12: /* Display 32bpp. */
2605 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2606 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2607 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2608 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2609 NUM_BANKS(ADDR_SURF_16_BANK) |
2610 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2611 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2612 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2613 break;
2614 case 13: /* Thin. */
2615 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2616 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2617 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2618 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2619 NUM_BANKS(ADDR_SURF_16_BANK) |
2620 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2621 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2622 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2623 break;
2624 case 14: /* Thin 8 bpp. */
2625 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2626 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2627 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2628 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2629 NUM_BANKS(ADDR_SURF_16_BANK) |
2630 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2631 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2632 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2633 break;
2634 case 15: /* Thin 16 bpp. */
2635 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2636 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2637 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2638 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2639 NUM_BANKS(ADDR_SURF_16_BANK) |
2640 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2641 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2642 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2643 break;
2644 case 16: /* Thin 32 bpp. */
2645 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2646 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2647 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2648 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2649 NUM_BANKS(ADDR_SURF_16_BANK) |
2650 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2651 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2652 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2653 break;
2654 case 17: /* Thin 64 bpp. */
2655 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2656 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2657 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2658 TILE_SPLIT(split_equal_to_row_size) |
2659 NUM_BANKS(ADDR_SURF_16_BANK) |
2660 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2661 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2662 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2663 break;
2664 case 21: /* 8 bpp PRT. */
2665 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2666 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2667 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2668 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2669 NUM_BANKS(ADDR_SURF_16_BANK) |
2670 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
2671 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2672 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2673 break;
2674 case 22: /* 16 bpp PRT */
2675 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2676 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2677 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2678 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2679 NUM_BANKS(ADDR_SURF_16_BANK) |
2680 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2681 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2682 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2683 break;
2684 case 23: /* 32 bpp PRT */
2685 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2686 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2687 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2688 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2689 NUM_BANKS(ADDR_SURF_16_BANK) |
2690 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2691 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2692 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2693 break;
2694 case 24: /* 64 bpp PRT */
2695 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2696 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2697 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2698 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2699 NUM_BANKS(ADDR_SURF_16_BANK) |
2700 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2701 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2702 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2703 break;
2704 case 25: /* 128 bpp PRT */
2705 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2706 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2707 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2708 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
2709 NUM_BANKS(ADDR_SURF_8_BANK) |
2710 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2711 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2712 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2713 break;
2714 default:
2715 gb_tile_moden = 0;
2716 break;
2717 }
64d7b8be 2718 rdev->config.si.tile_mode_array[reg_offset] = gb_tile_moden;
0a96d72b
AD
2719 WREG32(GB_TILE_MODE0 + (reg_offset * 4), gb_tile_moden);
2720 }
2721 } else
2722 DRM_ERROR("unknown asic: 0x%x\n", rdev->family);
2723}
2724
1a8ca750
AD
2725static void si_select_se_sh(struct radeon_device *rdev,
2726 u32 se_num, u32 sh_num)
2727{
2728 u32 data = INSTANCE_BROADCAST_WRITES;
2729
2730 if ((se_num == 0xffffffff) && (sh_num == 0xffffffff))
79b52d6a 2731 data |= SH_BROADCAST_WRITES | SE_BROADCAST_WRITES;
1a8ca750
AD
2732 else if (se_num == 0xffffffff)
2733 data |= SE_BROADCAST_WRITES | SH_INDEX(sh_num);
2734 else if (sh_num == 0xffffffff)
2735 data |= SH_BROADCAST_WRITES | SE_INDEX(se_num);
2736 else
2737 data |= SH_INDEX(sh_num) | SE_INDEX(se_num);
2738 WREG32(GRBM_GFX_INDEX, data);
2739}
2740
2741static u32 si_create_bitmask(u32 bit_width)
2742{
2743 u32 i, mask = 0;
2744
2745 for (i = 0; i < bit_width; i++) {
2746 mask <<= 1;
2747 mask |= 1;
2748 }
2749 return mask;
2750}
2751
2752static u32 si_get_cu_enabled(struct radeon_device *rdev, u32 cu_per_sh)
2753{
2754 u32 data, mask;
2755
2756 data = RREG32(CC_GC_SHADER_ARRAY_CONFIG);
2757 if (data & 1)
2758 data &= INACTIVE_CUS_MASK;
2759 else
2760 data = 0;
2761 data |= RREG32(GC_USER_SHADER_ARRAY_CONFIG);
2762
2763 data >>= INACTIVE_CUS_SHIFT;
2764
2765 mask = si_create_bitmask(cu_per_sh);
2766
2767 return ~data & mask;
2768}
2769
2770static void si_setup_spi(struct radeon_device *rdev,
2771 u32 se_num, u32 sh_per_se,
2772 u32 cu_per_sh)
2773{
2774 int i, j, k;
2775 u32 data, mask, active_cu;
2776
2777 for (i = 0; i < se_num; i++) {
2778 for (j = 0; j < sh_per_se; j++) {
2779 si_select_se_sh(rdev, i, j);
2780 data = RREG32(SPI_STATIC_THREAD_MGMT_3);
2781 active_cu = si_get_cu_enabled(rdev, cu_per_sh);
2782
2783 mask = 1;
2784 for (k = 0; k < 16; k++) {
2785 mask <<= k;
2786 if (active_cu & mask) {
2787 data &= ~mask;
2788 WREG32(SPI_STATIC_THREAD_MGMT_3, data);
2789 break;
2790 }
2791 }
2792 }
2793 }
2794 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
2795}
2796
2797static u32 si_get_rb_disabled(struct radeon_device *rdev,
2798 u32 max_rb_num, u32 se_num,
2799 u32 sh_per_se)
2800{
2801 u32 data, mask;
2802
2803 data = RREG32(CC_RB_BACKEND_DISABLE);
2804 if (data & 1)
2805 data &= BACKEND_DISABLE_MASK;
2806 else
2807 data = 0;
2808 data |= RREG32(GC_USER_RB_BACKEND_DISABLE);
2809
2810 data >>= BACKEND_DISABLE_SHIFT;
2811
2812 mask = si_create_bitmask(max_rb_num / se_num / sh_per_se);
2813
2814 return data & mask;
2815}
2816
2817static void si_setup_rb(struct radeon_device *rdev,
2818 u32 se_num, u32 sh_per_se,
2819 u32 max_rb_num)
2820{
2821 int i, j;
2822 u32 data, mask;
2823 u32 disabled_rbs = 0;
2824 u32 enabled_rbs = 0;
2825
2826 for (i = 0; i < se_num; i++) {
2827 for (j = 0; j < sh_per_se; j++) {
2828 si_select_se_sh(rdev, i, j);
2829 data = si_get_rb_disabled(rdev, max_rb_num, se_num, sh_per_se);
2830 disabled_rbs |= data << ((i * sh_per_se + j) * TAHITI_RB_BITMAP_WIDTH_PER_SH);
2831 }
2832 }
2833 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
2834
2835 mask = 1;
2836 for (i = 0; i < max_rb_num; i++) {
2837 if (!(disabled_rbs & mask))
2838 enabled_rbs |= mask;
2839 mask <<= 1;
2840 }
2841
2842 for (i = 0; i < se_num; i++) {
2843 si_select_se_sh(rdev, i, 0xffffffff);
2844 data = 0;
2845 for (j = 0; j < sh_per_se; j++) {
2846 switch (enabled_rbs & 3) {
2847 case 1:
2848 data |= (RASTER_CONFIG_RB_MAP_0 << (i * sh_per_se + j) * 2);
2849 break;
2850 case 2:
2851 data |= (RASTER_CONFIG_RB_MAP_3 << (i * sh_per_se + j) * 2);
2852 break;
2853 case 3:
2854 default:
2855 data |= (RASTER_CONFIG_RB_MAP_2 << (i * sh_per_se + j) * 2);
2856 break;
2857 }
2858 enabled_rbs >>= 2;
2859 }
2860 WREG32(PA_SC_RASTER_CONFIG, data);
2861 }
2862 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
2863}
2864
0a96d72b
AD
2865static void si_gpu_init(struct radeon_device *rdev)
2866{
0a96d72b
AD
2867 u32 gb_addr_config = 0;
2868 u32 mc_shared_chmap, mc_arb_ramcfg;
0a96d72b 2869 u32 sx_debug_1;
0a96d72b
AD
2870 u32 hdp_host_path_cntl;
2871 u32 tmp;
2872 int i, j;
2873
2874 switch (rdev->family) {
2875 case CHIP_TAHITI:
2876 rdev->config.si.max_shader_engines = 2;
0a96d72b 2877 rdev->config.si.max_tile_pipes = 12;
1a8ca750
AD
2878 rdev->config.si.max_cu_per_sh = 8;
2879 rdev->config.si.max_sh_per_se = 2;
0a96d72b
AD
2880 rdev->config.si.max_backends_per_se = 4;
2881 rdev->config.si.max_texture_channel_caches = 12;
2882 rdev->config.si.max_gprs = 256;
2883 rdev->config.si.max_gs_threads = 32;
2884 rdev->config.si.max_hw_contexts = 8;
2885
2886 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
2887 rdev->config.si.sc_prim_fifo_size_backend = 0x100;
2888 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
2889 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
1a8ca750 2890 gb_addr_config = TAHITI_GB_ADDR_CONFIG_GOLDEN;
0a96d72b
AD
2891 break;
2892 case CHIP_PITCAIRN:
2893 rdev->config.si.max_shader_engines = 2;
0a96d72b 2894 rdev->config.si.max_tile_pipes = 8;
1a8ca750
AD
2895 rdev->config.si.max_cu_per_sh = 5;
2896 rdev->config.si.max_sh_per_se = 2;
0a96d72b
AD
2897 rdev->config.si.max_backends_per_se = 4;
2898 rdev->config.si.max_texture_channel_caches = 8;
2899 rdev->config.si.max_gprs = 256;
2900 rdev->config.si.max_gs_threads = 32;
2901 rdev->config.si.max_hw_contexts = 8;
2902
2903 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
2904 rdev->config.si.sc_prim_fifo_size_backend = 0x100;
2905 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
2906 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
1a8ca750 2907 gb_addr_config = TAHITI_GB_ADDR_CONFIG_GOLDEN;
0a96d72b
AD
2908 break;
2909 case CHIP_VERDE:
2910 default:
2911 rdev->config.si.max_shader_engines = 1;
0a96d72b 2912 rdev->config.si.max_tile_pipes = 4;
468ef1a5 2913 rdev->config.si.max_cu_per_sh = 5;
1a8ca750 2914 rdev->config.si.max_sh_per_se = 2;
0a96d72b
AD
2915 rdev->config.si.max_backends_per_se = 4;
2916 rdev->config.si.max_texture_channel_caches = 4;
2917 rdev->config.si.max_gprs = 256;
2918 rdev->config.si.max_gs_threads = 32;
2919 rdev->config.si.max_hw_contexts = 8;
2920
d0ae7fcc
AD
2921 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
2922 rdev->config.si.sc_prim_fifo_size_backend = 0x40;
2923 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
2924 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
2925 gb_addr_config = VERDE_GB_ADDR_CONFIG_GOLDEN;
2926 break;
2927 case CHIP_OLAND:
2928 rdev->config.si.max_shader_engines = 1;
2929 rdev->config.si.max_tile_pipes = 4;
2930 rdev->config.si.max_cu_per_sh = 6;
2931 rdev->config.si.max_sh_per_se = 1;
2932 rdev->config.si.max_backends_per_se = 2;
2933 rdev->config.si.max_texture_channel_caches = 4;
2934 rdev->config.si.max_gprs = 256;
2935 rdev->config.si.max_gs_threads = 16;
2936 rdev->config.si.max_hw_contexts = 8;
2937
0a96d72b
AD
2938 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
2939 rdev->config.si.sc_prim_fifo_size_backend = 0x40;
2940 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
2941 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
1a8ca750 2942 gb_addr_config = VERDE_GB_ADDR_CONFIG_GOLDEN;
0a96d72b 2943 break;
8b02859d
AD
2944 case CHIP_HAINAN:
2945 rdev->config.si.max_shader_engines = 1;
2946 rdev->config.si.max_tile_pipes = 4;
2947 rdev->config.si.max_cu_per_sh = 5;
2948 rdev->config.si.max_sh_per_se = 1;
2949 rdev->config.si.max_backends_per_se = 1;
2950 rdev->config.si.max_texture_channel_caches = 2;
2951 rdev->config.si.max_gprs = 256;
2952 rdev->config.si.max_gs_threads = 16;
2953 rdev->config.si.max_hw_contexts = 8;
2954
2955 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
2956 rdev->config.si.sc_prim_fifo_size_backend = 0x40;
2957 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
2958 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
2959 gb_addr_config = HAINAN_GB_ADDR_CONFIG_GOLDEN;
2960 break;
0a96d72b
AD
2961 }
2962
2963 /* Initialize HDP */
2964 for (i = 0, j = 0; i < 32; i++, j += 0x18) {
2965 WREG32((0x2c14 + j), 0x00000000);
2966 WREG32((0x2c18 + j), 0x00000000);
2967 WREG32((0x2c1c + j), 0x00000000);
2968 WREG32((0x2c20 + j), 0x00000000);
2969 WREG32((0x2c24 + j), 0x00000000);
2970 }
2971
2972 WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff));
2973
2974 evergreen_fix_pci_max_read_req_size(rdev);
2975
2976 WREG32(BIF_FB_EN, FB_READ_EN | FB_WRITE_EN);
2977
2978 mc_shared_chmap = RREG32(MC_SHARED_CHMAP);
2979 mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG);
2980
0a96d72b 2981 rdev->config.si.num_tile_pipes = rdev->config.si.max_tile_pipes;
0a96d72b
AD
2982 rdev->config.si.mem_max_burst_length_bytes = 256;
2983 tmp = (mc_arb_ramcfg & NOOFCOLS_MASK) >> NOOFCOLS_SHIFT;
2984 rdev->config.si.mem_row_size_in_kb = (4 * (1 << (8 + tmp))) / 1024;
2985 if (rdev->config.si.mem_row_size_in_kb > 4)
2986 rdev->config.si.mem_row_size_in_kb = 4;
2987 /* XXX use MC settings? */
2988 rdev->config.si.shader_engine_tile_size = 32;
2989 rdev->config.si.num_gpus = 1;
2990 rdev->config.si.multi_gpu_tile_size = 64;
2991
1a8ca750
AD
2992 /* fix up row size */
2993 gb_addr_config &= ~ROW_SIZE_MASK;
0a96d72b
AD
2994 switch (rdev->config.si.mem_row_size_in_kb) {
2995 case 1:
2996 default:
2997 gb_addr_config |= ROW_SIZE(0);
2998 break;
2999 case 2:
3000 gb_addr_config |= ROW_SIZE(1);
3001 break;
3002 case 4:
3003 gb_addr_config |= ROW_SIZE(2);
3004 break;
3005 }
3006
0a96d72b
AD
3007 /* setup tiling info dword. gb_addr_config is not adequate since it does
3008 * not have bank info, so create a custom tiling dword.
3009 * bits 3:0 num_pipes
3010 * bits 7:4 num_banks
3011 * bits 11:8 group_size
3012 * bits 15:12 row_size
3013 */
3014 rdev->config.si.tile_config = 0;
3015 switch (rdev->config.si.num_tile_pipes) {
3016 case 1:
3017 rdev->config.si.tile_config |= (0 << 0);
3018 break;
3019 case 2:
3020 rdev->config.si.tile_config |= (1 << 0);
3021 break;
3022 case 4:
3023 rdev->config.si.tile_config |= (2 << 0);
3024 break;
3025 case 8:
3026 default:
3027 /* XXX what about 12? */
3028 rdev->config.si.tile_config |= (3 << 0);
3029 break;
dca571a6
CK
3030 }
3031 switch ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT) {
3032 case 0: /* four banks */
1a8ca750 3033 rdev->config.si.tile_config |= 0 << 4;
dca571a6
CK
3034 break;
3035 case 1: /* eight banks */
3036 rdev->config.si.tile_config |= 1 << 4;
3037 break;
3038 case 2: /* sixteen banks */
3039 default:
3040 rdev->config.si.tile_config |= 2 << 4;
3041 break;
3042 }
0a96d72b
AD
3043 rdev->config.si.tile_config |=
3044 ((gb_addr_config & PIPE_INTERLEAVE_SIZE_MASK) >> PIPE_INTERLEAVE_SIZE_SHIFT) << 8;
3045 rdev->config.si.tile_config |=
3046 ((gb_addr_config & ROW_SIZE_MASK) >> ROW_SIZE_SHIFT) << 12;
3047
0a96d72b
AD
3048 WREG32(GB_ADDR_CONFIG, gb_addr_config);
3049 WREG32(DMIF_ADDR_CONFIG, gb_addr_config);
7c1c7c18 3050 WREG32(DMIF_ADDR_CALC, gb_addr_config);
0a96d72b 3051 WREG32(HDP_ADDR_CONFIG, gb_addr_config);
8c5fd7ef
AD
3052 WREG32(DMA_TILING_CONFIG + DMA0_REGISTER_OFFSET, gb_addr_config);
3053 WREG32(DMA_TILING_CONFIG + DMA1_REGISTER_OFFSET, gb_addr_config);
1df0d523
AD
3054 if (rdev->has_uvd) {
3055 WREG32(UVD_UDEC_ADDR_CONFIG, gb_addr_config);
3056 WREG32(UVD_UDEC_DB_ADDR_CONFIG, gb_addr_config);
3057 WREG32(UVD_UDEC_DBW_ADDR_CONFIG, gb_addr_config);
3058 }
0a96d72b 3059
1a8ca750 3060 si_tiling_mode_table_init(rdev);
0a96d72b 3061
1a8ca750
AD
3062 si_setup_rb(rdev, rdev->config.si.max_shader_engines,
3063 rdev->config.si.max_sh_per_se,
3064 rdev->config.si.max_backends_per_se);
0a96d72b 3065
1a8ca750
AD
3066 si_setup_spi(rdev, rdev->config.si.max_shader_engines,
3067 rdev->config.si.max_sh_per_se,
3068 rdev->config.si.max_cu_per_sh);
0a96d72b 3069
0a96d72b
AD
3070
3071 /* set HW defaults for 3D engine */
3072 WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) |
3073 ROQ_IB2_START(0x2b)));
3074 WREG32(CP_MEQ_THRESHOLDS, MEQ1_START(0x30) | MEQ2_START(0x60));
3075
3076 sx_debug_1 = RREG32(SX_DEBUG_1);
3077 WREG32(SX_DEBUG_1, sx_debug_1);
3078
3079 WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4));
3080
3081 WREG32(PA_SC_FIFO_SIZE, (SC_FRONTEND_PRIM_FIFO_SIZE(rdev->config.si.sc_prim_fifo_size_frontend) |
3082 SC_BACKEND_PRIM_FIFO_SIZE(rdev->config.si.sc_prim_fifo_size_backend) |
3083 SC_HIZ_TILE_FIFO_SIZE(rdev->config.si.sc_hiz_tile_fifo_size) |
3084 SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.si.sc_earlyz_tile_fifo_size)));
3085
3086 WREG32(VGT_NUM_INSTANCES, 1);
3087
3088 WREG32(CP_PERFMON_CNTL, 0);
3089
3090 WREG32(SQ_CONFIG, 0);
3091
3092 WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) |
3093 FORCE_EOV_MAX_REZ_CNT(255)));
3094
3095 WREG32(VGT_CACHE_INVALIDATION, CACHE_INVALIDATION(VC_AND_TC) |
3096 AUTO_INVLD_EN(ES_AND_GS_AUTO));
3097
3098 WREG32(VGT_GS_VERTEX_REUSE, 16);
3099 WREG32(PA_SC_LINE_STIPPLE_STATE, 0);
3100
3101 WREG32(CB_PERFCOUNTER0_SELECT0, 0);
3102 WREG32(CB_PERFCOUNTER0_SELECT1, 0);
3103 WREG32(CB_PERFCOUNTER1_SELECT0, 0);
3104 WREG32(CB_PERFCOUNTER1_SELECT1, 0);
3105 WREG32(CB_PERFCOUNTER2_SELECT0, 0);
3106 WREG32(CB_PERFCOUNTER2_SELECT1, 0);
3107 WREG32(CB_PERFCOUNTER3_SELECT0, 0);
3108 WREG32(CB_PERFCOUNTER3_SELECT1, 0);
3109
3110 tmp = RREG32(HDP_MISC_CNTL);
3111 tmp |= HDP_FLUSH_INVALIDATE_CACHE;
3112 WREG32(HDP_MISC_CNTL, tmp);
3113
3114 hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL);
3115 WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
3116
3117 WREG32(PA_CL_ENHANCE, CLIP_VTX_REORDER_ENA | NUM_CLIP_SEQ(3));
3118
3119 udelay(50);
3120}
c476dde2 3121
2ece2e8b
AD
3122/*
3123 * GPU scratch registers helpers function.
3124 */
3125static void si_scratch_init(struct radeon_device *rdev)
3126{
3127 int i;
3128
3129 rdev->scratch.num_reg = 7;
3130 rdev->scratch.reg_base = SCRATCH_REG0;
3131 for (i = 0; i < rdev->scratch.num_reg; i++) {
3132 rdev->scratch.free[i] = true;
3133 rdev->scratch.reg[i] = rdev->scratch.reg_base + (i * 4);
3134 }
3135}
3136
3137void si_fence_ring_emit(struct radeon_device *rdev,
3138 struct radeon_fence *fence)
3139{
3140 struct radeon_ring *ring = &rdev->ring[fence->ring];
3141 u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
3142
3143 /* flush read cache over gart */
3144 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
3145 radeon_ring_write(ring, (CP_COHER_CNTL2 - PACKET3_SET_CONFIG_REG_START) >> 2);
3146 radeon_ring_write(ring, 0);
3147 radeon_ring_write(ring, PACKET3(PACKET3_SURFACE_SYNC, 3));
3148 radeon_ring_write(ring, PACKET3_TCL1_ACTION_ENA |
3149 PACKET3_TC_ACTION_ENA |
3150 PACKET3_SH_KCACHE_ACTION_ENA |
3151 PACKET3_SH_ICACHE_ACTION_ENA);
3152 radeon_ring_write(ring, 0xFFFFFFFF);
3153 radeon_ring_write(ring, 0);
3154 radeon_ring_write(ring, 10); /* poll interval */
3155 /* EVENT_WRITE_EOP - flush caches, send int */
3156 radeon_ring_write(ring, PACKET3(PACKET3_EVENT_WRITE_EOP, 4));
3157 radeon_ring_write(ring, EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) | EVENT_INDEX(5));
3158 radeon_ring_write(ring, addr & 0xffffffff);
3159 radeon_ring_write(ring, (upper_32_bits(addr) & 0xff) | DATA_SEL(1) | INT_SEL(2));
3160 radeon_ring_write(ring, fence->seq);
3161 radeon_ring_write(ring, 0);
3162}
3163
3164/*
3165 * IB stuff
3166 */
3167void si_ring_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib)
3168{
876dc9f3 3169 struct radeon_ring *ring = &rdev->ring[ib->ring];
2ece2e8b
AD
3170 u32 header;
3171
a85a7da4
AD
3172 if (ib->is_const_ib) {
3173 /* set switch buffer packet before const IB */
3174 radeon_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
3175 radeon_ring_write(ring, 0);
45df6803 3176
2ece2e8b 3177 header = PACKET3(PACKET3_INDIRECT_BUFFER_CONST, 2);
a85a7da4 3178 } else {
89d35807 3179 u32 next_rptr;
a85a7da4 3180 if (ring->rptr_save_reg) {
89d35807 3181 next_rptr = ring->wptr + 3 + 4 + 8;
a85a7da4
AD
3182 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
3183 radeon_ring_write(ring, ((ring->rptr_save_reg -
3184 PACKET3_SET_CONFIG_REG_START) >> 2));
3185 radeon_ring_write(ring, next_rptr);
89d35807
AD
3186 } else if (rdev->wb.enabled) {
3187 next_rptr = ring->wptr + 5 + 4 + 8;
3188 radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
3189 radeon_ring_write(ring, (1 << 8));
3190 radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
3191 radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr) & 0xffffffff);
3192 radeon_ring_write(ring, next_rptr);
a85a7da4
AD
3193 }
3194
2ece2e8b 3195 header = PACKET3(PACKET3_INDIRECT_BUFFER, 2);
a85a7da4 3196 }
2ece2e8b
AD
3197
3198 radeon_ring_write(ring, header);
3199 radeon_ring_write(ring,
3200#ifdef __BIG_ENDIAN
3201 (2 << 0) |
3202#endif
3203 (ib->gpu_addr & 0xFFFFFFFC));
3204 radeon_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xFFFF);
4bf3dd92
CK
3205 radeon_ring_write(ring, ib->length_dw |
3206 (ib->vm ? (ib->vm->id << 24) : 0));
2ece2e8b 3207
a85a7da4
AD
3208 if (!ib->is_const_ib) {
3209 /* flush read cache over gart for this vmid */
3210 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
3211 radeon_ring_write(ring, (CP_COHER_CNTL2 - PACKET3_SET_CONFIG_REG_START) >> 2);
4bf3dd92 3212 radeon_ring_write(ring, ib->vm ? ib->vm->id : 0);
a85a7da4
AD
3213 radeon_ring_write(ring, PACKET3(PACKET3_SURFACE_SYNC, 3));
3214 radeon_ring_write(ring, PACKET3_TCL1_ACTION_ENA |
3215 PACKET3_TC_ACTION_ENA |
3216 PACKET3_SH_KCACHE_ACTION_ENA |
3217 PACKET3_SH_ICACHE_ACTION_ENA);
3218 radeon_ring_write(ring, 0xFFFFFFFF);
3219 radeon_ring_write(ring, 0);
3220 radeon_ring_write(ring, 10); /* poll interval */
3221 }
2ece2e8b
AD
3222}
3223
48c0c902
AD
3224/*
3225 * CP.
3226 */
3227static void si_cp_enable(struct radeon_device *rdev, bool enable)
3228{
3229 if (enable)
3230 WREG32(CP_ME_CNTL, 0);
3231 else {
3232 radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
3233 WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT | CP_CE_HALT));
3234 WREG32(SCRATCH_UMSK, 0);
8c5fd7ef
AD
3235 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
3236 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
3237 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
48c0c902
AD
3238 }
3239 udelay(50);
3240}
3241
3242static int si_cp_load_microcode(struct radeon_device *rdev)
3243{
3244 const __be32 *fw_data;
3245 int i;
3246
3247 if (!rdev->me_fw || !rdev->pfp_fw)
3248 return -EINVAL;
3249
3250 si_cp_enable(rdev, false);
3251
3252 /* PFP */
3253 fw_data = (const __be32 *)rdev->pfp_fw->data;
3254 WREG32(CP_PFP_UCODE_ADDR, 0);
3255 for (i = 0; i < SI_PFP_UCODE_SIZE; i++)
3256 WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++));
3257 WREG32(CP_PFP_UCODE_ADDR, 0);
3258
3259 /* CE */
3260 fw_data = (const __be32 *)rdev->ce_fw->data;
3261 WREG32(CP_CE_UCODE_ADDR, 0);
3262 for (i = 0; i < SI_CE_UCODE_SIZE; i++)
3263 WREG32(CP_CE_UCODE_DATA, be32_to_cpup(fw_data++));
3264 WREG32(CP_CE_UCODE_ADDR, 0);
3265
3266 /* ME */
3267 fw_data = (const __be32 *)rdev->me_fw->data;
3268 WREG32(CP_ME_RAM_WADDR, 0);
3269 for (i = 0; i < SI_PM4_UCODE_SIZE; i++)
3270 WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++));
3271 WREG32(CP_ME_RAM_WADDR, 0);
3272
3273 WREG32(CP_PFP_UCODE_ADDR, 0);
3274 WREG32(CP_CE_UCODE_ADDR, 0);
3275 WREG32(CP_ME_RAM_WADDR, 0);
3276 WREG32(CP_ME_RAM_RADDR, 0);
3277 return 0;
3278}
3279
3280static int si_cp_start(struct radeon_device *rdev)
3281{
3282 struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
3283 int r, i;
3284
3285 r = radeon_ring_lock(rdev, ring, 7 + 4);
3286 if (r) {
3287 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
3288 return r;
3289 }
3290 /* init the CP */
3291 radeon_ring_write(ring, PACKET3(PACKET3_ME_INITIALIZE, 5));
3292 radeon_ring_write(ring, 0x1);
3293 radeon_ring_write(ring, 0x0);
3294 radeon_ring_write(ring, rdev->config.si.max_hw_contexts - 1);
3295 radeon_ring_write(ring, PACKET3_ME_INITIALIZE_DEVICE_ID(1));
3296 radeon_ring_write(ring, 0);
3297 radeon_ring_write(ring, 0);
3298
3299 /* init the CE partitions */
3300 radeon_ring_write(ring, PACKET3(PACKET3_SET_BASE, 2));
3301 radeon_ring_write(ring, PACKET3_BASE_INDEX(CE_PARTITION_BASE));
3302 radeon_ring_write(ring, 0xc000);
3303 radeon_ring_write(ring, 0xe000);
3304 radeon_ring_unlock_commit(rdev, ring);
3305
3306 si_cp_enable(rdev, true);
3307
3308 r = radeon_ring_lock(rdev, ring, si_default_size + 10);
3309 if (r) {
3310 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
3311 return r;
3312 }
3313
3314 /* setup clear context state */
3315 radeon_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
3316 radeon_ring_write(ring, PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
3317
3318 for (i = 0; i < si_default_size; i++)
3319 radeon_ring_write(ring, si_default_state[i]);
3320
3321 radeon_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
3322 radeon_ring_write(ring, PACKET3_PREAMBLE_END_CLEAR_STATE);
3323
3324 /* set clear context state */
3325 radeon_ring_write(ring, PACKET3(PACKET3_CLEAR_STATE, 0));
3326 radeon_ring_write(ring, 0);
3327
3328 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONTEXT_REG, 2));
3329 radeon_ring_write(ring, 0x00000316);
3330 radeon_ring_write(ring, 0x0000000e); /* VGT_VERTEX_REUSE_BLOCK_CNTL */
3331 radeon_ring_write(ring, 0x00000010); /* VGT_OUT_DEALLOC_CNTL */
3332
3333 radeon_ring_unlock_commit(rdev, ring);
3334
3335 for (i = RADEON_RING_TYPE_GFX_INDEX; i <= CAYMAN_RING_TYPE_CP2_INDEX; ++i) {
3336 ring = &rdev->ring[i];
3337 r = radeon_ring_lock(rdev, ring, 2);
3338
3339 /* clear the compute context state */
3340 radeon_ring_write(ring, PACKET3_COMPUTE(PACKET3_CLEAR_STATE, 0));
3341 radeon_ring_write(ring, 0);
3342
3343 radeon_ring_unlock_commit(rdev, ring);
3344 }
3345
3346 return 0;
3347}
3348
3349static void si_cp_fini(struct radeon_device *rdev)
3350{
45df6803 3351 struct radeon_ring *ring;
48c0c902 3352 si_cp_enable(rdev, false);
45df6803
CK
3353
3354 ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
3355 radeon_ring_fini(rdev, ring);
3356 radeon_scratch_free(rdev, ring->rptr_save_reg);
3357
3358 ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
3359 radeon_ring_fini(rdev, ring);
3360 radeon_scratch_free(rdev, ring->rptr_save_reg);
3361
3362 ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
3363 radeon_ring_fini(rdev, ring);
3364 radeon_scratch_free(rdev, ring->rptr_save_reg);
48c0c902
AD
3365}
3366
3367static int si_cp_resume(struct radeon_device *rdev)
3368{
3369 struct radeon_ring *ring;
3370 u32 tmp;
3371 u32 rb_bufsz;
3372 int r;
3373
48c0c902
AD
3374 WREG32(CP_SEM_WAIT_TIMER, 0x0);
3375 WREG32(CP_SEM_INCOMPLETE_TIMER_CNTL, 0x0);
3376
3377 /* Set the write pointer delay */
3378 WREG32(CP_RB_WPTR_DELAY, 0);
3379
3380 WREG32(CP_DEBUG, 0);
3381 WREG32(SCRATCH_ADDR, ((rdev->wb.gpu_addr + RADEON_WB_SCRATCH_OFFSET) >> 8) & 0xFFFFFFFF);
3382
3383 /* ring 0 - compute and gfx */
3384 /* Set ring buffer size */
3385 ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
3386 rb_bufsz = drm_order(ring->ring_size / 8);
3387 tmp = (drm_order(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
3388#ifdef __BIG_ENDIAN
3389 tmp |= BUF_SWAP_32BIT;
3390#endif
3391 WREG32(CP_RB0_CNTL, tmp);
3392
3393 /* Initialize the ring buffer's read and write pointers */
3394 WREG32(CP_RB0_CNTL, tmp | RB_RPTR_WR_ENA);
3395 ring->wptr = 0;
3396 WREG32(CP_RB0_WPTR, ring->wptr);
3397
48fc7f7e 3398 /* set the wb address whether it's enabled or not */
48c0c902
AD
3399 WREG32(CP_RB0_RPTR_ADDR, (rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET) & 0xFFFFFFFC);
3400 WREG32(CP_RB0_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET) & 0xFF);
3401
3402 if (rdev->wb.enabled)
3403 WREG32(SCRATCH_UMSK, 0xff);
3404 else {
3405 tmp |= RB_NO_UPDATE;
3406 WREG32(SCRATCH_UMSK, 0);
3407 }
3408
3409 mdelay(1);
3410 WREG32(CP_RB0_CNTL, tmp);
3411
3412 WREG32(CP_RB0_BASE, ring->gpu_addr >> 8);
3413
3414 ring->rptr = RREG32(CP_RB0_RPTR);
3415
3416 /* ring1 - compute only */
3417 /* Set ring buffer size */
3418 ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
3419 rb_bufsz = drm_order(ring->ring_size / 8);
3420 tmp = (drm_order(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
3421#ifdef __BIG_ENDIAN
3422 tmp |= BUF_SWAP_32BIT;
3423#endif
3424 WREG32(CP_RB1_CNTL, tmp);
3425
3426 /* Initialize the ring buffer's read and write pointers */
3427 WREG32(CP_RB1_CNTL, tmp | RB_RPTR_WR_ENA);
3428 ring->wptr = 0;
3429 WREG32(CP_RB1_WPTR, ring->wptr);
3430
48fc7f7e 3431 /* set the wb address whether it's enabled or not */
48c0c902
AD
3432 WREG32(CP_RB1_RPTR_ADDR, (rdev->wb.gpu_addr + RADEON_WB_CP1_RPTR_OFFSET) & 0xFFFFFFFC);
3433 WREG32(CP_RB1_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP1_RPTR_OFFSET) & 0xFF);
3434
3435 mdelay(1);
3436 WREG32(CP_RB1_CNTL, tmp);
3437
3438 WREG32(CP_RB1_BASE, ring->gpu_addr >> 8);
3439
3440 ring->rptr = RREG32(CP_RB1_RPTR);
3441
3442 /* ring2 - compute only */
3443 /* Set ring buffer size */
3444 ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
3445 rb_bufsz = drm_order(ring->ring_size / 8);
3446 tmp = (drm_order(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
3447#ifdef __BIG_ENDIAN
3448 tmp |= BUF_SWAP_32BIT;
3449#endif
3450 WREG32(CP_RB2_CNTL, tmp);
3451
3452 /* Initialize the ring buffer's read and write pointers */
3453 WREG32(CP_RB2_CNTL, tmp | RB_RPTR_WR_ENA);
3454 ring->wptr = 0;
3455 WREG32(CP_RB2_WPTR, ring->wptr);
3456
48fc7f7e 3457 /* set the wb address whether it's enabled or not */
48c0c902
AD
3458 WREG32(CP_RB2_RPTR_ADDR, (rdev->wb.gpu_addr + RADEON_WB_CP2_RPTR_OFFSET) & 0xFFFFFFFC);
3459 WREG32(CP_RB2_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP2_RPTR_OFFSET) & 0xFF);
3460
3461 mdelay(1);
3462 WREG32(CP_RB2_CNTL, tmp);
3463
3464 WREG32(CP_RB2_BASE, ring->gpu_addr >> 8);
3465
3466 ring->rptr = RREG32(CP_RB2_RPTR);
3467
3468 /* start the rings */
3469 si_cp_start(rdev);
3470 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = true;
3471 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = true;
3472 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = true;
3473 r = radeon_ring_test(rdev, RADEON_RING_TYPE_GFX_INDEX, &rdev->ring[RADEON_RING_TYPE_GFX_INDEX]);
3474 if (r) {
3475 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
3476 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
3477 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
3478 return r;
3479 }
3480 r = radeon_ring_test(rdev, CAYMAN_RING_TYPE_CP1_INDEX, &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX]);
3481 if (r) {
3482 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
3483 }
3484 r = radeon_ring_test(rdev, CAYMAN_RING_TYPE_CP2_INDEX, &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX]);
3485 if (r) {
3486 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
3487 }
3488
3489 return 0;
3490}
3491
2483b4ea 3492u32 si_gpu_check_soft_reset(struct radeon_device *rdev)
06bc6df0 3493{
014bb209 3494 u32 reset_mask = 0;
1c534671 3495 u32 tmp;
06bc6df0 3496
014bb209
AD
3497 /* GRBM_STATUS */
3498 tmp = RREG32(GRBM_STATUS);
3499 if (tmp & (PA_BUSY | SC_BUSY |
3500 BCI_BUSY | SX_BUSY |
3501 TA_BUSY | VGT_BUSY |
3502 DB_BUSY | CB_BUSY |
3503 GDS_BUSY | SPI_BUSY |
3504 IA_BUSY | IA_BUSY_NO_DMA))
3505 reset_mask |= RADEON_RESET_GFX;
3506
3507 if (tmp & (CF_RQ_PENDING | PF_RQ_PENDING |
3508 CP_BUSY | CP_COHERENCY_BUSY))
3509 reset_mask |= RADEON_RESET_CP;
3510
3511 if (tmp & GRBM_EE_BUSY)
3512 reset_mask |= RADEON_RESET_GRBM | RADEON_RESET_GFX | RADEON_RESET_CP;
3513
3514 /* GRBM_STATUS2 */
3515 tmp = RREG32(GRBM_STATUS2);
3516 if (tmp & (RLC_RQ_PENDING | RLC_BUSY))
3517 reset_mask |= RADEON_RESET_RLC;
3518
3519 /* DMA_STATUS_REG 0 */
3520 tmp = RREG32(DMA_STATUS_REG + DMA0_REGISTER_OFFSET);
3521 if (!(tmp & DMA_IDLE))
3522 reset_mask |= RADEON_RESET_DMA;
3523
3524 /* DMA_STATUS_REG 1 */
3525 tmp = RREG32(DMA_STATUS_REG + DMA1_REGISTER_OFFSET);
3526 if (!(tmp & DMA_IDLE))
3527 reset_mask |= RADEON_RESET_DMA1;
3528
3529 /* SRBM_STATUS2 */
3530 tmp = RREG32(SRBM_STATUS2);
3531 if (tmp & DMA_BUSY)
3532 reset_mask |= RADEON_RESET_DMA;
3533
3534 if (tmp & DMA1_BUSY)
3535 reset_mask |= RADEON_RESET_DMA1;
3536
3537 /* SRBM_STATUS */
3538 tmp = RREG32(SRBM_STATUS);
3539
3540 if (tmp & IH_BUSY)
3541 reset_mask |= RADEON_RESET_IH;
3542
3543 if (tmp & SEM_BUSY)
3544 reset_mask |= RADEON_RESET_SEM;
3545
3546 if (tmp & GRBM_RQ_PENDING)
3547 reset_mask |= RADEON_RESET_GRBM;
3548
3549 if (tmp & VMC_BUSY)
3550 reset_mask |= RADEON_RESET_VMC;
19fc42ed 3551
014bb209
AD
3552 if (tmp & (MCB_BUSY | MCB_NON_DISPLAY_BUSY |
3553 MCC_BUSY | MCD_BUSY))
3554 reset_mask |= RADEON_RESET_MC;
3555
3556 if (evergreen_is_display_hung(rdev))
3557 reset_mask |= RADEON_RESET_DISPLAY;
3558
3559 /* VM_L2_STATUS */
3560 tmp = RREG32(VM_L2_STATUS);
3561 if (tmp & L2_BUSY)
3562 reset_mask |= RADEON_RESET_VMC;
3563
d808fc88
AD
3564 /* Skip MC reset as it's mostly likely not hung, just busy */
3565 if (reset_mask & RADEON_RESET_MC) {
3566 DRM_DEBUG("MC busy: 0x%08X, clearing.\n", reset_mask);
3567 reset_mask &= ~RADEON_RESET_MC;
3568 }
3569
014bb209
AD
3570 return reset_mask;
3571}
3572
3573static void si_gpu_soft_reset(struct radeon_device *rdev, u32 reset_mask)
3574{
3575 struct evergreen_mc_save save;
3576 u32 grbm_soft_reset = 0, srbm_soft_reset = 0;
3577 u32 tmp;
19fc42ed 3578
06bc6df0 3579 if (reset_mask == 0)
014bb209 3580 return;
06bc6df0
AD
3581
3582 dev_info(rdev->dev, "GPU softreset: 0x%08X\n", reset_mask);
3583
1c534671 3584 evergreen_print_gpu_status_regs(rdev);
06bc6df0
AD
3585 dev_info(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n",
3586 RREG32(VM_CONTEXT1_PROTECTION_FAULT_ADDR));
3587 dev_info(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
3588 RREG32(VM_CONTEXT1_PROTECTION_FAULT_STATUS));
3589
1c534671
AD
3590 /* Disable CP parsing/prefetching */
3591 WREG32(CP_ME_CNTL, CP_ME_HALT | CP_PFP_HALT | CP_CE_HALT);
3592
3593 if (reset_mask & RADEON_RESET_DMA) {
3594 /* dma0 */
3595 tmp = RREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET);
3596 tmp &= ~DMA_RB_ENABLE;
3597 WREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET, tmp);
014bb209
AD
3598 }
3599 if (reset_mask & RADEON_RESET_DMA1) {
1c534671
AD
3600 /* dma1 */
3601 tmp = RREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET);
3602 tmp &= ~DMA_RB_ENABLE;
3603 WREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET, tmp);
3604 }
3605
f770d78a
AD
3606 udelay(50);
3607
3608 evergreen_mc_stop(rdev, &save);
3609 if (evergreen_mc_wait_for_idle(rdev)) {
3610 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
3611 }
3612
1c534671
AD
3613 if (reset_mask & (RADEON_RESET_GFX | RADEON_RESET_COMPUTE | RADEON_RESET_CP)) {
3614 grbm_soft_reset = SOFT_RESET_CB |
3615 SOFT_RESET_DB |
3616 SOFT_RESET_GDS |
3617 SOFT_RESET_PA |
3618 SOFT_RESET_SC |
3619 SOFT_RESET_BCI |
3620 SOFT_RESET_SPI |
3621 SOFT_RESET_SX |
3622 SOFT_RESET_TC |
3623 SOFT_RESET_TA |
3624 SOFT_RESET_VGT |
3625 SOFT_RESET_IA;
3626 }
3627
3628 if (reset_mask & RADEON_RESET_CP) {
3629 grbm_soft_reset |= SOFT_RESET_CP | SOFT_RESET_VGT;
3630
3631 srbm_soft_reset |= SOFT_RESET_GRBM;
3632 }
06bc6df0
AD
3633
3634 if (reset_mask & RADEON_RESET_DMA)
014bb209
AD
3635 srbm_soft_reset |= SOFT_RESET_DMA;
3636
3637 if (reset_mask & RADEON_RESET_DMA1)
3638 srbm_soft_reset |= SOFT_RESET_DMA1;
3639
3640 if (reset_mask & RADEON_RESET_DISPLAY)
3641 srbm_soft_reset |= SOFT_RESET_DC;
3642
3643 if (reset_mask & RADEON_RESET_RLC)
3644 grbm_soft_reset |= SOFT_RESET_RLC;
3645
3646 if (reset_mask & RADEON_RESET_SEM)
3647 srbm_soft_reset |= SOFT_RESET_SEM;
3648
3649 if (reset_mask & RADEON_RESET_IH)
3650 srbm_soft_reset |= SOFT_RESET_IH;
3651
3652 if (reset_mask & RADEON_RESET_GRBM)
3653 srbm_soft_reset |= SOFT_RESET_GRBM;
3654
3655 if (reset_mask & RADEON_RESET_VMC)
3656 srbm_soft_reset |= SOFT_RESET_VMC;
3657
3658 if (reset_mask & RADEON_RESET_MC)
3659 srbm_soft_reset |= SOFT_RESET_MC;
1c534671
AD
3660
3661 if (grbm_soft_reset) {
3662 tmp = RREG32(GRBM_SOFT_RESET);
3663 tmp |= grbm_soft_reset;
3664 dev_info(rdev->dev, "GRBM_SOFT_RESET=0x%08X\n", tmp);
3665 WREG32(GRBM_SOFT_RESET, tmp);
3666 tmp = RREG32(GRBM_SOFT_RESET);
3667
3668 udelay(50);
3669
3670 tmp &= ~grbm_soft_reset;
3671 WREG32(GRBM_SOFT_RESET, tmp);
3672 tmp = RREG32(GRBM_SOFT_RESET);
3673 }
3674
3675 if (srbm_soft_reset) {
3676 tmp = RREG32(SRBM_SOFT_RESET);
3677 tmp |= srbm_soft_reset;
3678 dev_info(rdev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
3679 WREG32(SRBM_SOFT_RESET, tmp);
3680 tmp = RREG32(SRBM_SOFT_RESET);
3681
3682 udelay(50);
3683
3684 tmp &= ~srbm_soft_reset;
3685 WREG32(SRBM_SOFT_RESET, tmp);
3686 tmp = RREG32(SRBM_SOFT_RESET);
3687 }
06bc6df0
AD
3688
3689 /* Wait a little for things to settle down */
3690 udelay(50);
3691
c476dde2 3692 evergreen_mc_resume(rdev, &save);
1c534671
AD
3693 udelay(50);
3694
1c534671 3695 evergreen_print_gpu_status_regs(rdev);
c476dde2
AD
3696}
3697
3698int si_asic_reset(struct radeon_device *rdev)
3699{
014bb209
AD
3700 u32 reset_mask;
3701
3702 reset_mask = si_gpu_check_soft_reset(rdev);
3703
3704 if (reset_mask)
3705 r600_set_bios_scratch_engine_hung(rdev, true);
3706
3707 si_gpu_soft_reset(rdev, reset_mask);
3708
3709 reset_mask = si_gpu_check_soft_reset(rdev);
3710
3711 if (!reset_mask)
3712 r600_set_bios_scratch_engine_hung(rdev, false);
3713
3714 return 0;
c476dde2
AD
3715}
3716
123bc183
AD
3717/**
3718 * si_gfx_is_lockup - Check if the GFX engine is locked up
3719 *
3720 * @rdev: radeon_device pointer
3721 * @ring: radeon_ring structure holding ring information
3722 *
3723 * Check if the GFX engine is locked up.
3724 * Returns true if the engine appears to be locked up, false if not.
3725 */
3726bool si_gfx_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
3727{
3728 u32 reset_mask = si_gpu_check_soft_reset(rdev);
3729
3730 if (!(reset_mask & (RADEON_RESET_GFX |
3731 RADEON_RESET_COMPUTE |
3732 RADEON_RESET_CP))) {
3733 radeon_ring_lockup_update(ring);
3734 return false;
3735 }
3736 /* force CP activities */
3737 radeon_ring_force_activity(rdev, ring);
3738 return radeon_ring_test_lockup(rdev, ring);
3739}
3740
d2800ee5
AD
3741/* MC */
3742static void si_mc_program(struct radeon_device *rdev)
3743{
3744 struct evergreen_mc_save save;
3745 u32 tmp;
3746 int i, j;
3747
3748 /* Initialize HDP */
3749 for (i = 0, j = 0; i < 32; i++, j += 0x18) {
3750 WREG32((0x2c14 + j), 0x00000000);
3751 WREG32((0x2c18 + j), 0x00000000);
3752 WREG32((0x2c1c + j), 0x00000000);
3753 WREG32((0x2c20 + j), 0x00000000);
3754 WREG32((0x2c24 + j), 0x00000000);
3755 }
3756 WREG32(HDP_REG_COHERENCY_FLUSH_CNTL, 0);
3757
3758 evergreen_mc_stop(rdev, &save);
3759 if (radeon_mc_wait_for_idle(rdev)) {
3760 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
3761 }
5153550a
AD
3762 if (!ASIC_IS_NODCE(rdev))
3763 /* Lockout access through VGA aperture*/
3764 WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE);
d2800ee5
AD
3765 /* Update configuration */
3766 WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
3767 rdev->mc.vram_start >> 12);
3768 WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
3769 rdev->mc.vram_end >> 12);
3770 WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR,
3771 rdev->vram_scratch.gpu_addr >> 12);
3772 tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16;
3773 tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF);
3774 WREG32(MC_VM_FB_LOCATION, tmp);
3775 /* XXX double check these! */
3776 WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8));
3777 WREG32(HDP_NONSURFACE_INFO, (2 << 7) | (1 << 30));
3778 WREG32(HDP_NONSURFACE_SIZE, 0x3FFFFFFF);
3779 WREG32(MC_VM_AGP_BASE, 0);
3780 WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF);
3781 WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF);
3782 if (radeon_mc_wait_for_idle(rdev)) {
3783 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
3784 }
3785 evergreen_mc_resume(rdev, &save);
5153550a
AD
3786 if (!ASIC_IS_NODCE(rdev)) {
3787 /* we need to own VRAM, so turn off the VGA renderer here
3788 * to stop it overwriting our objects */
3789 rv515_vga_render_disable(rdev);
3790 }
d2800ee5
AD
3791}
3792
1c49165d
AD
3793void si_vram_gtt_location(struct radeon_device *rdev,
3794 struct radeon_mc *mc)
d2800ee5
AD
3795{
3796 if (mc->mc_vram_size > 0xFFC0000000ULL) {
3797 /* leave room for at least 1024M GTT */
3798 dev_warn(rdev->dev, "limiting VRAM\n");
3799 mc->real_vram_size = 0xFFC0000000ULL;
3800 mc->mc_vram_size = 0xFFC0000000ULL;
3801 }
9ed8b1f9 3802 radeon_vram_location(rdev, &rdev->mc, 0);
d2800ee5 3803 rdev->mc.gtt_base_align = 0;
9ed8b1f9 3804 radeon_gtt_location(rdev, mc);
d2800ee5
AD
3805}
3806
3807static int si_mc_init(struct radeon_device *rdev)
3808{
3809 u32 tmp;
3810 int chansize, numchan;
3811
3812 /* Get VRAM informations */
3813 rdev->mc.vram_is_ddr = true;
3814 tmp = RREG32(MC_ARB_RAMCFG);
3815 if (tmp & CHANSIZE_OVERRIDE) {
3816 chansize = 16;
3817 } else if (tmp & CHANSIZE_MASK) {
3818 chansize = 64;
3819 } else {
3820 chansize = 32;
3821 }
3822 tmp = RREG32(MC_SHARED_CHMAP);
3823 switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
3824 case 0:
3825 default:
3826 numchan = 1;
3827 break;
3828 case 1:
3829 numchan = 2;
3830 break;
3831 case 2:
3832 numchan = 4;
3833 break;
3834 case 3:
3835 numchan = 8;
3836 break;
3837 case 4:
3838 numchan = 3;
3839 break;
3840 case 5:
3841 numchan = 6;
3842 break;
3843 case 6:
3844 numchan = 10;
3845 break;
3846 case 7:
3847 numchan = 12;
3848 break;
3849 case 8:
3850 numchan = 16;
3851 break;
3852 }
3853 rdev->mc.vram_width = numchan * chansize;
3854 /* Could aper size report 0 ? */
3855 rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
3856 rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
3857 /* size in MB on si */
fc986034
NOS
3858 rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE) * 1024ULL * 1024ULL;
3859 rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE) * 1024ULL * 1024ULL;
d2800ee5
AD
3860 rdev->mc.visible_vram_size = rdev->mc.aper_size;
3861 si_vram_gtt_location(rdev, &rdev->mc);
3862 radeon_update_bandwidth_info(rdev);
3863
3864 return 0;
3865}
3866
3867/*
3868 * GART
3869 */
3870void si_pcie_gart_tlb_flush(struct radeon_device *rdev)
3871{
3872 /* flush hdp cache */
3873 WREG32(HDP_MEM_COHERENCY_FLUSH_CNTL, 0x1);
3874
3875 /* bits 0-15 are the VM contexts0-15 */
3876 WREG32(VM_INVALIDATE_REQUEST, 1);
3877}
3878
1109ca09 3879static int si_pcie_gart_enable(struct radeon_device *rdev)
d2800ee5
AD
3880{
3881 int r, i;
3882
3883 if (rdev->gart.robj == NULL) {
3884 dev_err(rdev->dev, "No VRAM object for PCIE GART.\n");
3885 return -EINVAL;
3886 }
3887 r = radeon_gart_table_vram_pin(rdev);
3888 if (r)
3889 return r;
3890 radeon_gart_restore(rdev);
3891 /* Setup TLB control */
3892 WREG32(MC_VM_MX_L1_TLB_CNTL,
3893 (0xA << 7) |
3894 ENABLE_L1_TLB |
3895 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
3896 ENABLE_ADVANCED_DRIVER_MODEL |
3897 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
3898 /* Setup L2 cache */
3899 WREG32(VM_L2_CNTL, ENABLE_L2_CACHE |
3900 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
3901 ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
3902 EFFECTIVE_L2_QUEUE_SIZE(7) |
3903 CONTEXT1_IDENTITY_ACCESS_MODE(1));
3904 WREG32(VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS | INVALIDATE_L2_CACHE);
3905 WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
3906 L2_CACHE_BIGK_FRAGMENT_SIZE(0));
3907 /* setup context0 */
3908 WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);
3909 WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);
3910 WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12);
3911 WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
3912 (u32)(rdev->dummy_page.addr >> 12));
3913 WREG32(VM_CONTEXT0_CNTL2, 0);
3914 WREG32(VM_CONTEXT0_CNTL, (ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) |
3915 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT));
3916
3917 WREG32(0x15D4, 0);
3918 WREG32(0x15D8, 0);
3919 WREG32(0x15DC, 0);
3920
3921 /* empty context1-15 */
d2800ee5
AD
3922 /* set vm size, must be a multiple of 4 */
3923 WREG32(VM_CONTEXT1_PAGE_TABLE_START_ADDR, 0);
c21b328e 3924 WREG32(VM_CONTEXT1_PAGE_TABLE_END_ADDR, rdev->vm_manager.max_pfn);
23d4f1f2
AD
3925 /* Assign the pt base to something valid for now; the pts used for
3926 * the VMs are determined by the application and setup and assigned
3927 * on the fly in the vm part of radeon_gart.c
3928 */
d2800ee5
AD
3929 for (i = 1; i < 16; i++) {
3930 if (i < 8)
3931 WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (i << 2),
3932 rdev->gart.table_addr >> 12);
3933 else
3934 WREG32(VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((i - 8) << 2),
3935 rdev->gart.table_addr >> 12);
3936 }
3937
3938 /* enable context1-15 */
3939 WREG32(VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR,
3940 (u32)(rdev->dummy_page.addr >> 12));
ae133a11 3941 WREG32(VM_CONTEXT1_CNTL2, 4);
fa87e62d 3942 WREG32(VM_CONTEXT1_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(1) |
ae133a11
CK
3943 RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT |
3944 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT |
3945 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT |
3946 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT |
3947 PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT |
3948 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT |
3949 VALID_PROTECTION_FAULT_ENABLE_INTERRUPT |
3950 VALID_PROTECTION_FAULT_ENABLE_DEFAULT |
3951 READ_PROTECTION_FAULT_ENABLE_INTERRUPT |
3952 READ_PROTECTION_FAULT_ENABLE_DEFAULT |
3953 WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT |
3954 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT);
d2800ee5
AD
3955
3956 si_pcie_gart_tlb_flush(rdev);
3957 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
3958 (unsigned)(rdev->mc.gtt_size >> 20),
3959 (unsigned long long)rdev->gart.table_addr);
3960 rdev->gart.ready = true;
3961 return 0;
3962}
3963
1109ca09 3964static void si_pcie_gart_disable(struct radeon_device *rdev)
d2800ee5
AD
3965{
3966 /* Disable all tables */
3967 WREG32(VM_CONTEXT0_CNTL, 0);
3968 WREG32(VM_CONTEXT1_CNTL, 0);
3969 /* Setup TLB control */
3970 WREG32(MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE_NOT_IN_SYS |
3971 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
3972 /* Setup L2 cache */
3973 WREG32(VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
3974 ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
3975 EFFECTIVE_L2_QUEUE_SIZE(7) |
3976 CONTEXT1_IDENTITY_ACCESS_MODE(1));
3977 WREG32(VM_L2_CNTL2, 0);
3978 WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
3979 L2_CACHE_BIGK_FRAGMENT_SIZE(0));
3980 radeon_gart_table_vram_unpin(rdev);
3981}
3982
1109ca09 3983static void si_pcie_gart_fini(struct radeon_device *rdev)
d2800ee5
AD
3984{
3985 si_pcie_gart_disable(rdev);
3986 radeon_gart_table_vram_free(rdev);
3987 radeon_gart_fini(rdev);
3988}
3989
498dd8b3
AD
3990/* vm parser */
3991static bool si_vm_reg_valid(u32 reg)
3992{
3993 /* context regs are fine */
3994 if (reg >= 0x28000)
3995 return true;
3996
3997 /* check config regs */
3998 switch (reg) {
3999 case GRBM_GFX_INDEX:
f418b88a 4000 case CP_STRMOUT_CNTL:
498dd8b3
AD
4001 case VGT_VTX_VECT_EJECT_REG:
4002 case VGT_CACHE_INVALIDATION:
4003 case VGT_ESGS_RING_SIZE:
4004 case VGT_GSVS_RING_SIZE:
4005 case VGT_GS_VERTEX_REUSE:
4006 case VGT_PRIMITIVE_TYPE:
4007 case VGT_INDEX_TYPE:
4008 case VGT_NUM_INDICES:
4009 case VGT_NUM_INSTANCES:
4010 case VGT_TF_RING_SIZE:
4011 case VGT_HS_OFFCHIP_PARAM:
4012 case VGT_TF_MEMORY_BASE:
4013 case PA_CL_ENHANCE:
4014 case PA_SU_LINE_STIPPLE_VALUE:
4015 case PA_SC_LINE_STIPPLE_STATE:
4016 case PA_SC_ENHANCE:
4017 case SQC_CACHES:
4018 case SPI_STATIC_THREAD_MGMT_1:
4019 case SPI_STATIC_THREAD_MGMT_2:
4020 case SPI_STATIC_THREAD_MGMT_3:
4021 case SPI_PS_MAX_WAVE_ID:
4022 case SPI_CONFIG_CNTL:
4023 case SPI_CONFIG_CNTL_1:
4024 case TA_CNTL_AUX:
4025 return true;
4026 default:
4027 DRM_ERROR("Invalid register 0x%x in CS\n", reg);
4028 return false;
4029 }
4030}
4031
4032static int si_vm_packet3_ce_check(struct radeon_device *rdev,
4033 u32 *ib, struct radeon_cs_packet *pkt)
4034{
4035 switch (pkt->opcode) {
4036 case PACKET3_NOP:
4037 case PACKET3_SET_BASE:
4038 case PACKET3_SET_CE_DE_COUNTERS:
4039 case PACKET3_LOAD_CONST_RAM:
4040 case PACKET3_WRITE_CONST_RAM:
4041 case PACKET3_WRITE_CONST_RAM_OFFSET:
4042 case PACKET3_DUMP_CONST_RAM:
4043 case PACKET3_INCREMENT_CE_COUNTER:
4044 case PACKET3_WAIT_ON_DE_COUNTER:
4045 case PACKET3_CE_WRITE:
4046 break;
4047 default:
4048 DRM_ERROR("Invalid CE packet3: 0x%x\n", pkt->opcode);
4049 return -EINVAL;
4050 }
4051 return 0;
4052}
4053
4054static int si_vm_packet3_gfx_check(struct radeon_device *rdev,
4055 u32 *ib, struct radeon_cs_packet *pkt)
4056{
4057 u32 idx = pkt->idx + 1;
4058 u32 idx_value = ib[idx];
4059 u32 start_reg, end_reg, reg, i;
5aa709be 4060 u32 command, info;
498dd8b3
AD
4061
4062 switch (pkt->opcode) {
4063 case PACKET3_NOP:
4064 case PACKET3_SET_BASE:
4065 case PACKET3_CLEAR_STATE:
4066 case PACKET3_INDEX_BUFFER_SIZE:
4067 case PACKET3_DISPATCH_DIRECT:
4068 case PACKET3_DISPATCH_INDIRECT:
4069 case PACKET3_ALLOC_GDS:
4070 case PACKET3_WRITE_GDS_RAM:
4071 case PACKET3_ATOMIC_GDS:
4072 case PACKET3_ATOMIC:
4073 case PACKET3_OCCLUSION_QUERY:
4074 case PACKET3_SET_PREDICATION:
4075 case PACKET3_COND_EXEC:
4076 case PACKET3_PRED_EXEC:
4077 case PACKET3_DRAW_INDIRECT:
4078 case PACKET3_DRAW_INDEX_INDIRECT:
4079 case PACKET3_INDEX_BASE:
4080 case PACKET3_DRAW_INDEX_2:
4081 case PACKET3_CONTEXT_CONTROL:
4082 case PACKET3_INDEX_TYPE:
4083 case PACKET3_DRAW_INDIRECT_MULTI:
4084 case PACKET3_DRAW_INDEX_AUTO:
4085 case PACKET3_DRAW_INDEX_IMMD:
4086 case PACKET3_NUM_INSTANCES:
4087 case PACKET3_DRAW_INDEX_MULTI_AUTO:
4088 case PACKET3_STRMOUT_BUFFER_UPDATE:
4089 case PACKET3_DRAW_INDEX_OFFSET_2:
4090 case PACKET3_DRAW_INDEX_MULTI_ELEMENT:
4091 case PACKET3_DRAW_INDEX_INDIRECT_MULTI:
4092 case PACKET3_MPEG_INDEX:
4093 case PACKET3_WAIT_REG_MEM:
4094 case PACKET3_MEM_WRITE:
4095 case PACKET3_PFP_SYNC_ME:
4096 case PACKET3_SURFACE_SYNC:
4097 case PACKET3_EVENT_WRITE:
4098 case PACKET3_EVENT_WRITE_EOP:
4099 case PACKET3_EVENT_WRITE_EOS:
4100 case PACKET3_SET_CONTEXT_REG:
4101 case PACKET3_SET_CONTEXT_REG_INDIRECT:
4102 case PACKET3_SET_SH_REG:
4103 case PACKET3_SET_SH_REG_OFFSET:
4104 case PACKET3_INCREMENT_DE_COUNTER:
4105 case PACKET3_WAIT_ON_CE_COUNTER:
4106 case PACKET3_WAIT_ON_AVAIL_BUFFER:
4107 case PACKET3_ME_WRITE:
4108 break;
4109 case PACKET3_COPY_DATA:
4110 if ((idx_value & 0xf00) == 0) {
4111 reg = ib[idx + 3] * 4;
4112 if (!si_vm_reg_valid(reg))
4113 return -EINVAL;
4114 }
4115 break;
4116 case PACKET3_WRITE_DATA:
4117 if ((idx_value & 0xf00) == 0) {
4118 start_reg = ib[idx + 1] * 4;
4119 if (idx_value & 0x10000) {
4120 if (!si_vm_reg_valid(start_reg))
4121 return -EINVAL;
4122 } else {
4123 for (i = 0; i < (pkt->count - 2); i++) {
4124 reg = start_reg + (4 * i);
4125 if (!si_vm_reg_valid(reg))
4126 return -EINVAL;
4127 }
4128 }
4129 }
4130 break;
4131 case PACKET3_COND_WRITE:
4132 if (idx_value & 0x100) {
4133 reg = ib[idx + 5] * 4;
4134 if (!si_vm_reg_valid(reg))
4135 return -EINVAL;
4136 }
4137 break;
4138 case PACKET3_COPY_DW:
4139 if (idx_value & 0x2) {
4140 reg = ib[idx + 3] * 4;
4141 if (!si_vm_reg_valid(reg))
4142 return -EINVAL;
4143 }
4144 break;
4145 case PACKET3_SET_CONFIG_REG:
4146 start_reg = (idx_value << 2) + PACKET3_SET_CONFIG_REG_START;
4147 end_reg = 4 * pkt->count + start_reg - 4;
4148 if ((start_reg < PACKET3_SET_CONFIG_REG_START) ||
4149 (start_reg >= PACKET3_SET_CONFIG_REG_END) ||
4150 (end_reg >= PACKET3_SET_CONFIG_REG_END)) {
4151 DRM_ERROR("bad PACKET3_SET_CONFIG_REG\n");
4152 return -EINVAL;
4153 }
4154 for (i = 0; i < pkt->count; i++) {
4155 reg = start_reg + (4 * i);
4156 if (!si_vm_reg_valid(reg))
4157 return -EINVAL;
4158 }
4159 break;
5aa709be
AD
4160 case PACKET3_CP_DMA:
4161 command = ib[idx + 4];
4162 info = ib[idx + 1];
4163 if (command & PACKET3_CP_DMA_CMD_SAS) {
4164 /* src address space is register */
4165 if (((info & 0x60000000) >> 29) == 0) {
4166 start_reg = idx_value << 2;
4167 if (command & PACKET3_CP_DMA_CMD_SAIC) {
4168 reg = start_reg;
4169 if (!si_vm_reg_valid(reg)) {
4170 DRM_ERROR("CP DMA Bad SRC register\n");
4171 return -EINVAL;
4172 }
4173 } else {
4174 for (i = 0; i < (command & 0x1fffff); i++) {
4175 reg = start_reg + (4 * i);
4176 if (!si_vm_reg_valid(reg)) {
4177 DRM_ERROR("CP DMA Bad SRC register\n");
4178 return -EINVAL;
4179 }
4180 }
4181 }
4182 }
4183 }
4184 if (command & PACKET3_CP_DMA_CMD_DAS) {
4185 /* dst address space is register */
4186 if (((info & 0x00300000) >> 20) == 0) {
4187 start_reg = ib[idx + 2];
4188 if (command & PACKET3_CP_DMA_CMD_DAIC) {
4189 reg = start_reg;
4190 if (!si_vm_reg_valid(reg)) {
4191 DRM_ERROR("CP DMA Bad DST register\n");
4192 return -EINVAL;
4193 }
4194 } else {
4195 for (i = 0; i < (command & 0x1fffff); i++) {
4196 reg = start_reg + (4 * i);
4197 if (!si_vm_reg_valid(reg)) {
4198 DRM_ERROR("CP DMA Bad DST register\n");
4199 return -EINVAL;
4200 }
4201 }
4202 }
4203 }
4204 }
4205 break;
498dd8b3
AD
4206 default:
4207 DRM_ERROR("Invalid GFX packet3: 0x%x\n", pkt->opcode);
4208 return -EINVAL;
4209 }
4210 return 0;
4211}
4212
4213static int si_vm_packet3_compute_check(struct radeon_device *rdev,
4214 u32 *ib, struct radeon_cs_packet *pkt)
4215{
4216 u32 idx = pkt->idx + 1;
4217 u32 idx_value = ib[idx];
4218 u32 start_reg, reg, i;
4219
4220 switch (pkt->opcode) {
4221 case PACKET3_NOP:
4222 case PACKET3_SET_BASE:
4223 case PACKET3_CLEAR_STATE:
4224 case PACKET3_DISPATCH_DIRECT:
4225 case PACKET3_DISPATCH_INDIRECT:
4226 case PACKET3_ALLOC_GDS:
4227 case PACKET3_WRITE_GDS_RAM:
4228 case PACKET3_ATOMIC_GDS:
4229 case PACKET3_ATOMIC:
4230 case PACKET3_OCCLUSION_QUERY:
4231 case PACKET3_SET_PREDICATION:
4232 case PACKET3_COND_EXEC:
4233 case PACKET3_PRED_EXEC:
4234 case PACKET3_CONTEXT_CONTROL:
4235 case PACKET3_STRMOUT_BUFFER_UPDATE:
4236 case PACKET3_WAIT_REG_MEM:
4237 case PACKET3_MEM_WRITE:
4238 case PACKET3_PFP_SYNC_ME:
4239 case PACKET3_SURFACE_SYNC:
4240 case PACKET3_EVENT_WRITE:
4241 case PACKET3_EVENT_WRITE_EOP:
4242 case PACKET3_EVENT_WRITE_EOS:
4243 case PACKET3_SET_CONTEXT_REG:
4244 case PACKET3_SET_CONTEXT_REG_INDIRECT:
4245 case PACKET3_SET_SH_REG:
4246 case PACKET3_SET_SH_REG_OFFSET:
4247 case PACKET3_INCREMENT_DE_COUNTER:
4248 case PACKET3_WAIT_ON_CE_COUNTER:
4249 case PACKET3_WAIT_ON_AVAIL_BUFFER:
4250 case PACKET3_ME_WRITE:
4251 break;
4252 case PACKET3_COPY_DATA:
4253 if ((idx_value & 0xf00) == 0) {
4254 reg = ib[idx + 3] * 4;
4255 if (!si_vm_reg_valid(reg))
4256 return -EINVAL;
4257 }
4258 break;
4259 case PACKET3_WRITE_DATA:
4260 if ((idx_value & 0xf00) == 0) {
4261 start_reg = ib[idx + 1] * 4;
4262 if (idx_value & 0x10000) {
4263 if (!si_vm_reg_valid(start_reg))
4264 return -EINVAL;
4265 } else {
4266 for (i = 0; i < (pkt->count - 2); i++) {
4267 reg = start_reg + (4 * i);
4268 if (!si_vm_reg_valid(reg))
4269 return -EINVAL;
4270 }
4271 }
4272 }
4273 break;
4274 case PACKET3_COND_WRITE:
4275 if (idx_value & 0x100) {
4276 reg = ib[idx + 5] * 4;
4277 if (!si_vm_reg_valid(reg))
4278 return -EINVAL;
4279 }
4280 break;
4281 case PACKET3_COPY_DW:
4282 if (idx_value & 0x2) {
4283 reg = ib[idx + 3] * 4;
4284 if (!si_vm_reg_valid(reg))
4285 return -EINVAL;
4286 }
4287 break;
4288 default:
4289 DRM_ERROR("Invalid Compute packet3: 0x%x\n", pkt->opcode);
4290 return -EINVAL;
4291 }
4292 return 0;
4293}
4294
4295int si_ib_parse(struct radeon_device *rdev, struct radeon_ib *ib)
4296{
4297 int ret = 0;
4298 u32 idx = 0;
4299 struct radeon_cs_packet pkt;
4300
4301 do {
4302 pkt.idx = idx;
4e872ae2
IH
4303 pkt.type = RADEON_CP_PACKET_GET_TYPE(ib->ptr[idx]);
4304 pkt.count = RADEON_CP_PACKET_GET_COUNT(ib->ptr[idx]);
498dd8b3
AD
4305 pkt.one_reg_wr = 0;
4306 switch (pkt.type) {
4e872ae2 4307 case RADEON_PACKET_TYPE0:
498dd8b3
AD
4308 dev_err(rdev->dev, "Packet0 not allowed!\n");
4309 ret = -EINVAL;
4310 break;
4e872ae2 4311 case RADEON_PACKET_TYPE2:
498dd8b3
AD
4312 idx += 1;
4313 break;
4e872ae2
IH
4314 case RADEON_PACKET_TYPE3:
4315 pkt.opcode = RADEON_CP_PACKET3_GET_OPCODE(ib->ptr[idx]);
498dd8b3
AD
4316 if (ib->is_const_ib)
4317 ret = si_vm_packet3_ce_check(rdev, ib->ptr, &pkt);
4318 else {
876dc9f3 4319 switch (ib->ring) {
498dd8b3
AD
4320 case RADEON_RING_TYPE_GFX_INDEX:
4321 ret = si_vm_packet3_gfx_check(rdev, ib->ptr, &pkt);
4322 break;
4323 case CAYMAN_RING_TYPE_CP1_INDEX:
4324 case CAYMAN_RING_TYPE_CP2_INDEX:
4325 ret = si_vm_packet3_compute_check(rdev, ib->ptr, &pkt);
4326 break;
4327 default:
876dc9f3 4328 dev_err(rdev->dev, "Non-PM4 ring %d !\n", ib->ring);
498dd8b3
AD
4329 ret = -EINVAL;
4330 break;
4331 }
4332 }
4333 idx += pkt.count + 2;
4334 break;
4335 default:
4336 dev_err(rdev->dev, "Unknown packet type %d !\n", pkt.type);
4337 ret = -EINVAL;
4338 break;
4339 }
4340 if (ret)
4341 break;
4342 } while (idx < ib->length_dw);
4343
4344 return ret;
4345}
4346
d2800ee5
AD
4347/*
4348 * vm
4349 */
4350int si_vm_init(struct radeon_device *rdev)
4351{
4352 /* number of VMs */
4353 rdev->vm_manager.nvm = 16;
4354 /* base offset of vram pages */
4355 rdev->vm_manager.vram_base_offset = 0;
4356
4357 return 0;
4358}
4359
4360void si_vm_fini(struct radeon_device *rdev)
4361{
4362}
4363
fbf6dc7a
AD
4364/**
4365 * si_vm_decode_fault - print human readable fault info
4366 *
4367 * @rdev: radeon_device pointer
4368 * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value
4369 * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value
4370 *
4371 * Print human readable fault information (SI).
4372 */
4373static void si_vm_decode_fault(struct radeon_device *rdev,
4374 u32 status, u32 addr)
4375{
4376 u32 mc_id = (status & MEMORY_CLIENT_ID_MASK) >> MEMORY_CLIENT_ID_SHIFT;
4377 u32 vmid = (status & FAULT_VMID_MASK) >> FAULT_VMID_SHIFT;
4378 u32 protections = (status & PROTECTIONS_MASK) >> PROTECTIONS_SHIFT;
4379 char *block;
4380
4381 if (rdev->family == CHIP_TAHITI) {
4382 switch (mc_id) {
4383 case 160:
4384 case 144:
4385 case 96:
4386 case 80:
4387 case 224:
4388 case 208:
4389 case 32:
4390 case 16:
4391 block = "CB";
4392 break;
4393 case 161:
4394 case 145:
4395 case 97:
4396 case 81:
4397 case 225:
4398 case 209:
4399 case 33:
4400 case 17:
4401 block = "CB_FMASK";
4402 break;
4403 case 162:
4404 case 146:
4405 case 98:
4406 case 82:
4407 case 226:
4408 case 210:
4409 case 34:
4410 case 18:
4411 block = "CB_CMASK";
4412 break;
4413 case 163:
4414 case 147:
4415 case 99:
4416 case 83:
4417 case 227:
4418 case 211:
4419 case 35:
4420 case 19:
4421 block = "CB_IMMED";
4422 break;
4423 case 164:
4424 case 148:
4425 case 100:
4426 case 84:
4427 case 228:
4428 case 212:
4429 case 36:
4430 case 20:
4431 block = "DB";
4432 break;
4433 case 165:
4434 case 149:
4435 case 101:
4436 case 85:
4437 case 229:
4438 case 213:
4439 case 37:
4440 case 21:
4441 block = "DB_HTILE";
4442 break;
4443 case 167:
4444 case 151:
4445 case 103:
4446 case 87:
4447 case 231:
4448 case 215:
4449 case 39:
4450 case 23:
4451 block = "DB_STEN";
4452 break;
4453 case 72:
4454 case 68:
4455 case 64:
4456 case 8:
4457 case 4:
4458 case 0:
4459 case 136:
4460 case 132:
4461 case 128:
4462 case 200:
4463 case 196:
4464 case 192:
4465 block = "TC";
4466 break;
4467 case 112:
4468 case 48:
4469 block = "CP";
4470 break;
4471 case 49:
4472 case 177:
4473 case 50:
4474 case 178:
4475 block = "SH";
4476 break;
4477 case 53:
4478 case 190:
4479 block = "VGT";
4480 break;
4481 case 117:
4482 block = "IH";
4483 break;
4484 case 51:
4485 case 115:
4486 block = "RLC";
4487 break;
4488 case 119:
4489 case 183:
4490 block = "DMA0";
4491 break;
4492 case 61:
4493 block = "DMA1";
4494 break;
4495 case 248:
4496 case 120:
4497 block = "HDP";
4498 break;
4499 default:
4500 block = "unknown";
4501 break;
4502 }
4503 } else {
4504 switch (mc_id) {
4505 case 32:
4506 case 16:
4507 case 96:
4508 case 80:
4509 case 160:
4510 case 144:
4511 case 224:
4512 case 208:
4513 block = "CB";
4514 break;
4515 case 33:
4516 case 17:
4517 case 97:
4518 case 81:
4519 case 161:
4520 case 145:
4521 case 225:
4522 case 209:
4523 block = "CB_FMASK";
4524 break;
4525 case 34:
4526 case 18:
4527 case 98:
4528 case 82:
4529 case 162:
4530 case 146:
4531 case 226:
4532 case 210:
4533 block = "CB_CMASK";
4534 break;
4535 case 35:
4536 case 19:
4537 case 99:
4538 case 83:
4539 case 163:
4540 case 147:
4541 case 227:
4542 case 211:
4543 block = "CB_IMMED";
4544 break;
4545 case 36:
4546 case 20:
4547 case 100:
4548 case 84:
4549 case 164:
4550 case 148:
4551 case 228:
4552 case 212:
4553 block = "DB";
4554 break;
4555 case 37:
4556 case 21:
4557 case 101:
4558 case 85:
4559 case 165:
4560 case 149:
4561 case 229:
4562 case 213:
4563 block = "DB_HTILE";
4564 break;
4565 case 39:
4566 case 23:
4567 case 103:
4568 case 87:
4569 case 167:
4570 case 151:
4571 case 231:
4572 case 215:
4573 block = "DB_STEN";
4574 break;
4575 case 72:
4576 case 68:
4577 case 8:
4578 case 4:
4579 case 136:
4580 case 132:
4581 case 200:
4582 case 196:
4583 block = "TC";
4584 break;
4585 case 112:
4586 case 48:
4587 block = "CP";
4588 break;
4589 case 49:
4590 case 177:
4591 case 50:
4592 case 178:
4593 block = "SH";
4594 break;
4595 case 53:
4596 block = "VGT";
4597 break;
4598 case 117:
4599 block = "IH";
4600 break;
4601 case 51:
4602 case 115:
4603 block = "RLC";
4604 break;
4605 case 119:
4606 case 183:
4607 block = "DMA0";
4608 break;
4609 case 61:
4610 block = "DMA1";
4611 break;
4612 case 248:
4613 case 120:
4614 block = "HDP";
4615 break;
4616 default:
4617 block = "unknown";
4618 break;
4619 }
4620 }
4621
4622 printk("VM fault (0x%02x, vmid %d) at page %u, %s from %s (%d)\n",
4623 protections, vmid, addr,
4624 (status & MEMORY_CLIENT_RW_MASK) ? "write" : "read",
4625 block, mc_id);
4626}
4627
82ffd92b
AD
4628/**
4629 * si_vm_set_page - update the page tables using the CP
4630 *
4631 * @rdev: radeon_device pointer
43f1214a 4632 * @ib: indirect buffer to fill with commands
82ffd92b
AD
4633 * @pe: addr of the page entry
4634 * @addr: dst addr to write into pe
4635 * @count: number of page entries to update
4636 * @incr: increase next addr by incr bytes
4637 * @flags: access flags
4638 *
43f1214a 4639 * Update the page tables using the CP (SI).
82ffd92b 4640 */
43f1214a
AD
4641void si_vm_set_page(struct radeon_device *rdev,
4642 struct radeon_ib *ib,
4643 uint64_t pe,
82ffd92b
AD
4644 uint64_t addr, unsigned count,
4645 uint32_t incr, uint32_t flags)
d2800ee5 4646{
82ffd92b 4647 uint32_t r600_flags = cayman_vm_page_flags(rdev, flags);
deab48f1
AD
4648 uint64_t value;
4649 unsigned ndw;
4650
4651 if (rdev->asic->vm.pt_ring_index == RADEON_RING_TYPE_GFX_INDEX) {
4652 while (count) {
4653 ndw = 2 + count * 2;
4654 if (ndw > 0x3FFE)
4655 ndw = 0x3FFE;
4656
43f1214a
AD
4657 ib->ptr[ib->length_dw++] = PACKET3(PACKET3_WRITE_DATA, ndw);
4658 ib->ptr[ib->length_dw++] = (WRITE_DATA_ENGINE_SEL(0) |
4659 WRITE_DATA_DST_SEL(1));
4660 ib->ptr[ib->length_dw++] = pe;
4661 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
deab48f1
AD
4662 for (; ndw > 2; ndw -= 2, --count, pe += 8) {
4663 if (flags & RADEON_VM_PAGE_SYSTEM) {
4664 value = radeon_vm_map_gart(rdev, addr);
4665 value &= 0xFFFFFFFFFFFFF000ULL;
4666 } else if (flags & RADEON_VM_PAGE_VALID) {
4667 value = addr;
4668 } else {
4669 value = 0;
4670 }
4671 addr += incr;
4672 value |= r600_flags;
43f1214a
AD
4673 ib->ptr[ib->length_dw++] = value;
4674 ib->ptr[ib->length_dw++] = upper_32_bits(value);
deab48f1
AD
4675 }
4676 }
4677 } else {
4678 /* DMA */
2483b4ea 4679 si_dma_vm_set_page(rdev, ib, pe, addr, count, incr, flags);
82ffd92b 4680 }
d2800ee5
AD
4681}
4682
498522b4 4683void si_vm_flush(struct radeon_device *rdev, int ridx, struct radeon_vm *vm)
d2800ee5 4684{
498522b4 4685 struct radeon_ring *ring = &rdev->ring[ridx];
d2800ee5 4686
ee60e29f 4687 if (vm == NULL)
d2800ee5
AD
4688 return;
4689
76c44f2c
AD
4690 /* write new base address */
4691 radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4692 radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
4693 WRITE_DATA_DST_SEL(0)));
4694
ee60e29f 4695 if (vm->id < 8) {
76c44f2c
AD
4696 radeon_ring_write(ring,
4697 (VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm->id << 2)) >> 2);
ee60e29f 4698 } else {
76c44f2c
AD
4699 radeon_ring_write(ring,
4700 (VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((vm->id - 8) << 2)) >> 2);
ee60e29f 4701 }
76c44f2c 4702 radeon_ring_write(ring, 0);
fa87e62d 4703 radeon_ring_write(ring, vm->pd_gpu_addr >> 12);
ee60e29f 4704
d2800ee5 4705 /* flush hdp cache */
76c44f2c
AD
4706 radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4707 radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
4708 WRITE_DATA_DST_SEL(0)));
4709 radeon_ring_write(ring, HDP_MEM_COHERENCY_FLUSH_CNTL >> 2);
4710 radeon_ring_write(ring, 0);
ee60e29f
CK
4711 radeon_ring_write(ring, 0x1);
4712
d2800ee5 4713 /* bits 0-15 are the VM contexts0-15 */
76c44f2c
AD
4714 radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4715 radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
4716 WRITE_DATA_DST_SEL(0)));
4717 radeon_ring_write(ring, VM_INVALIDATE_REQUEST >> 2);
4718 radeon_ring_write(ring, 0);
498522b4 4719 radeon_ring_write(ring, 1 << vm->id);
58f8cf56
CK
4720
4721 /* sync PFP to ME, otherwise we might get invalid PFP reads */
4722 radeon_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
4723 radeon_ring_write(ring, 0x0);
d2800ee5
AD
4724}
4725
f8f84ac5
AD
4726/*
4727 * Power and clock gating
4728 */
4729static void si_wait_for_rlc_serdes(struct radeon_device *rdev)
4730{
4731 int i;
4732
4733 for (i = 0; i < rdev->usec_timeout; i++) {
4734 if (RREG32(RLC_SERDES_MASTER_BUSY_0) == 0)
4735 break;
4736 udelay(1);
4737 }
4738
4739 for (i = 0; i < rdev->usec_timeout; i++) {
4740 if (RREG32(RLC_SERDES_MASTER_BUSY_1) == 0)
4741 break;
4742 udelay(1);
4743 }
4744}
4745
4746static void si_enable_gui_idle_interrupt(struct radeon_device *rdev,
4747 bool enable)
4748{
4749 u32 tmp = RREG32(CP_INT_CNTL_RING0);
4750 u32 mask;
4751 int i;
4752
4753 if (enable)
4754 tmp |= (CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
4755 else
4756 tmp &= ~(CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
4757 WREG32(CP_INT_CNTL_RING0, tmp);
4758
4759 if (!enable) {
4760 /* read a gfx register */
4761 tmp = RREG32(DB_DEPTH_INFO);
4762
4763 mask = RLC_BUSY_STATUS | GFX_POWER_STATUS | GFX_CLOCK_STATUS | GFX_LS_STATUS;
4764 for (i = 0; i < rdev->usec_timeout; i++) {
4765 if ((RREG32(RLC_STAT) & mask) == (GFX_CLOCK_STATUS | GFX_POWER_STATUS))
4766 break;
4767 udelay(1);
4768 }
4769 }
4770}
4771
4772static void si_set_uvd_dcm(struct radeon_device *rdev,
4773 bool sw_mode)
4774{
4775 u32 tmp, tmp2;
4776
4777 tmp = RREG32(UVD_CGC_CTRL);
4778 tmp &= ~(CLK_OD_MASK | CG_DT_MASK);
4779 tmp |= DCM | CG_DT(1) | CLK_OD(4);
4780
4781 if (sw_mode) {
4782 tmp &= ~0x7ffff800;
4783 tmp2 = DYN_OR_EN | DYN_RR_EN | G_DIV_ID(7);
4784 } else {
4785 tmp |= 0x7ffff800;
4786 tmp2 = 0;
4787 }
4788
4789 WREG32(UVD_CGC_CTRL, tmp);
4790 WREG32_UVD_CTX(UVD_CGC_CTRL2, tmp2);
4791}
4792
22c775ce 4793void si_init_uvd_internal_cg(struct radeon_device *rdev)
f8f84ac5
AD
4794{
4795 bool hw_mode = true;
4796
4797 if (hw_mode) {
4798 si_set_uvd_dcm(rdev, false);
4799 } else {
4800 u32 tmp = RREG32(UVD_CGC_CTRL);
4801 tmp &= ~DCM;
4802 WREG32(UVD_CGC_CTRL, tmp);
4803 }
4804}
4805
4806static u32 si_halt_rlc(struct radeon_device *rdev)
4807{
4808 u32 data, orig;
4809
4810 orig = data = RREG32(RLC_CNTL);
4811
4812 if (data & RLC_ENABLE) {
4813 data &= ~RLC_ENABLE;
4814 WREG32(RLC_CNTL, data);
4815
4816 si_wait_for_rlc_serdes(rdev);
4817 }
4818
4819 return orig;
4820}
4821
4822static void si_update_rlc(struct radeon_device *rdev, u32 rlc)
4823{
4824 u32 tmp;
4825
4826 tmp = RREG32(RLC_CNTL);
4827 if (tmp != rlc)
4828 WREG32(RLC_CNTL, rlc);
4829}
4830
4831static void si_enable_dma_pg(struct radeon_device *rdev, bool enable)
4832{
4833 u32 data, orig;
4834
4835 orig = data = RREG32(DMA_PG);
e16866ec 4836 if (enable && (rdev->pg_flags & RADEON_PG_SUPPORT_SDMA))
f8f84ac5
AD
4837 data |= PG_CNTL_ENABLE;
4838 else
4839 data &= ~PG_CNTL_ENABLE;
4840 if (orig != data)
4841 WREG32(DMA_PG, data);
4842}
4843
4844static void si_init_dma_pg(struct radeon_device *rdev)
4845{
4846 u32 tmp;
4847
4848 WREG32(DMA_PGFSM_WRITE, 0x00002000);
4849 WREG32(DMA_PGFSM_CONFIG, 0x100010ff);
4850
4851 for (tmp = 0; tmp < 5; tmp++)
4852 WREG32(DMA_PGFSM_WRITE, 0);
4853}
4854
4855static void si_enable_gfx_cgpg(struct radeon_device *rdev,
4856 bool enable)
4857{
4858 u32 tmp;
4859
e16866ec 4860 if (enable && (rdev->pg_flags & RADEON_PG_SUPPORT_GFX_CG)) {
f8f84ac5
AD
4861 tmp = RLC_PUD(0x10) | RLC_PDD(0x10) | RLC_TTPD(0x10) | RLC_MSD(0x10);
4862 WREG32(RLC_TTOP_D, tmp);
4863
4864 tmp = RREG32(RLC_PG_CNTL);
4865 tmp |= GFX_PG_ENABLE;
4866 WREG32(RLC_PG_CNTL, tmp);
4867
4868 tmp = RREG32(RLC_AUTO_PG_CTRL);
4869 tmp |= AUTO_PG_EN;
4870 WREG32(RLC_AUTO_PG_CTRL, tmp);
4871 } else {
4872 tmp = RREG32(RLC_AUTO_PG_CTRL);
4873 tmp &= ~AUTO_PG_EN;
4874 WREG32(RLC_AUTO_PG_CTRL, tmp);
4875
4876 tmp = RREG32(DB_RENDER_CONTROL);
4877 }
4878}
4879
4880static void si_init_gfx_cgpg(struct radeon_device *rdev)
4881{
4882 u32 tmp;
4883
4884 WREG32(RLC_SAVE_AND_RESTORE_BASE, rdev->rlc.save_restore_gpu_addr >> 8);
4885
4886 tmp = RREG32(RLC_PG_CNTL);
4887 tmp |= GFX_PG_SRC;
4888 WREG32(RLC_PG_CNTL, tmp);
4889
4890 WREG32(RLC_CLEAR_STATE_RESTORE_BASE, rdev->rlc.clear_state_gpu_addr >> 8);
4891
4892 tmp = RREG32(RLC_AUTO_PG_CTRL);
4893
4894 tmp &= ~GRBM_REG_SGIT_MASK;
4895 tmp |= GRBM_REG_SGIT(0x700);
4896 tmp &= ~PG_AFTER_GRBM_REG_ST_MASK;
4897 WREG32(RLC_AUTO_PG_CTRL, tmp);
4898}
4899
ba19031a 4900static u32 si_get_cu_active_bitmap(struct radeon_device *rdev, u32 se, u32 sh)
f8f84ac5
AD
4901{
4902 u32 mask = 0, tmp, tmp1;
4903 int i;
4904
4905 si_select_se_sh(rdev, se, sh);
4906 tmp = RREG32(CC_GC_SHADER_ARRAY_CONFIG);
4907 tmp1 = RREG32(GC_USER_SHADER_ARRAY_CONFIG);
4908 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
4909
4910 tmp &= 0xffff0000;
4911
4912 tmp |= tmp1;
4913 tmp >>= 16;
4914
4915 for (i = 0; i < rdev->config.si.max_cu_per_sh; i ++) {
4916 mask <<= 1;
4917 mask |= 1;
4918 }
4919
4920 return (~tmp) & mask;
4921}
4922
4923static void si_init_ao_cu_mask(struct radeon_device *rdev)
4924{
4925 u32 i, j, k, active_cu_number = 0;
4926 u32 mask, counter, cu_bitmap;
4927 u32 tmp = 0;
4928
4929 for (i = 0; i < rdev->config.si.max_shader_engines; i++) {
4930 for (j = 0; j < rdev->config.si.max_sh_per_se; j++) {
4931 mask = 1;
4932 cu_bitmap = 0;
4933 counter = 0;
4934 for (k = 0; k < rdev->config.si.max_cu_per_sh; k++) {
ba19031a 4935 if (si_get_cu_active_bitmap(rdev, i, j) & mask) {
f8f84ac5
AD
4936 if (counter < 2)
4937 cu_bitmap |= mask;
4938 counter++;
4939 }
4940 mask <<= 1;
4941 }
4942
4943 active_cu_number += counter;
4944 tmp |= (cu_bitmap << (i * 16 + j * 8));
4945 }
4946 }
4947
4948 WREG32(RLC_PG_AO_CU_MASK, tmp);
4949
4950 tmp = RREG32(RLC_MAX_PG_CU);
4951 tmp &= ~MAX_PU_CU_MASK;
4952 tmp |= MAX_PU_CU(active_cu_number);
4953 WREG32(RLC_MAX_PG_CU, tmp);
4954}
4955
4956static void si_enable_cgcg(struct radeon_device *rdev,
4957 bool enable)
4958{
4959 u32 data, orig, tmp;
4960
4961 orig = data = RREG32(RLC_CGCG_CGLS_CTRL);
4962
e16866ec 4963 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_CGCG)) {
5594a558
AD
4964 si_enable_gui_idle_interrupt(rdev, true);
4965
f8f84ac5
AD
4966 WREG32(RLC_GCPM_GENERAL_3, 0x00000080);
4967
4968 tmp = si_halt_rlc(rdev);
4969
4970 WREG32(RLC_SERDES_WR_MASTER_MASK_0, 0xffffffff);
4971 WREG32(RLC_SERDES_WR_MASTER_MASK_1, 0xffffffff);
4972 WREG32(RLC_SERDES_WR_CTRL, 0x00b000ff);
4973
4974 si_wait_for_rlc_serdes(rdev);
4975
4976 si_update_rlc(rdev, tmp);
4977
4978 WREG32(RLC_SERDES_WR_CTRL, 0x007000ff);
4979
4980 data |= CGCG_EN | CGLS_EN;
4981 } else {
5594a558
AD
4982 si_enable_gui_idle_interrupt(rdev, false);
4983
f8f84ac5
AD
4984 RREG32(CB_CGTT_SCLK_CTRL);
4985 RREG32(CB_CGTT_SCLK_CTRL);
4986 RREG32(CB_CGTT_SCLK_CTRL);
4987 RREG32(CB_CGTT_SCLK_CTRL);
4988
4989 data &= ~(CGCG_EN | CGLS_EN);
4990 }
4991
4992 if (orig != data)
4993 WREG32(RLC_CGCG_CGLS_CTRL, data);
4994}
4995
4996static void si_enable_mgcg(struct radeon_device *rdev,
4997 bool enable)
4998{
4999 u32 data, orig, tmp = 0;
5000
e16866ec 5001 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_MGCG)) {
f8f84ac5
AD
5002 orig = data = RREG32(CGTS_SM_CTRL_REG);
5003 data = 0x96940200;
5004 if (orig != data)
5005 WREG32(CGTS_SM_CTRL_REG, data);
5006
e16866ec
AD
5007 if (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_CP_LS) {
5008 orig = data = RREG32(CP_MEM_SLP_CNTL);
5009 data |= CP_MEM_LS_EN;
5010 if (orig != data)
5011 WREG32(CP_MEM_SLP_CNTL, data);
5012 }
f8f84ac5
AD
5013
5014 orig = data = RREG32(RLC_CGTT_MGCG_OVERRIDE);
5015 data &= 0xffffffc0;
5016 if (orig != data)
5017 WREG32(RLC_CGTT_MGCG_OVERRIDE, data);
5018
5019 tmp = si_halt_rlc(rdev);
5020
5021 WREG32(RLC_SERDES_WR_MASTER_MASK_0, 0xffffffff);
5022 WREG32(RLC_SERDES_WR_MASTER_MASK_1, 0xffffffff);
5023 WREG32(RLC_SERDES_WR_CTRL, 0x00d000ff);
5024
5025 si_update_rlc(rdev, tmp);
5026 } else {
5027 orig = data = RREG32(RLC_CGTT_MGCG_OVERRIDE);
5028 data |= 0x00000003;
5029 if (orig != data)
5030 WREG32(RLC_CGTT_MGCG_OVERRIDE, data);
5031
5032 data = RREG32(CP_MEM_SLP_CNTL);
5033 if (data & CP_MEM_LS_EN) {
5034 data &= ~CP_MEM_LS_EN;
5035 WREG32(CP_MEM_SLP_CNTL, data);
5036 }
5037 orig = data = RREG32(CGTS_SM_CTRL_REG);
5038 data |= LS_OVERRIDE | OVERRIDE;
5039 if (orig != data)
5040 WREG32(CGTS_SM_CTRL_REG, data);
5041
5042 tmp = si_halt_rlc(rdev);
5043
5044 WREG32(RLC_SERDES_WR_MASTER_MASK_0, 0xffffffff);
5045 WREG32(RLC_SERDES_WR_MASTER_MASK_1, 0xffffffff);
5046 WREG32(RLC_SERDES_WR_CTRL, 0x00e000ff);
5047
5048 si_update_rlc(rdev, tmp);
5049 }
5050}
5051
5052static void si_enable_uvd_mgcg(struct radeon_device *rdev,
5053 bool enable)
5054{
5055 u32 orig, data, tmp;
5056
e16866ec 5057 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_UVD_MGCG)) {
f8f84ac5
AD
5058 tmp = RREG32_UVD_CTX(UVD_CGC_MEM_CTRL);
5059 tmp |= 0x3fff;
5060 WREG32_UVD_CTX(UVD_CGC_MEM_CTRL, tmp);
5061
5062 orig = data = RREG32(UVD_CGC_CTRL);
5063 data |= DCM;
5064 if (orig != data)
5065 WREG32(UVD_CGC_CTRL, data);
5066
5067 WREG32_SMC(SMC_CG_IND_START + CG_CGTT_LOCAL_0, 0);
5068 WREG32_SMC(SMC_CG_IND_START + CG_CGTT_LOCAL_1, 0);
5069 } else {
5070 tmp = RREG32_UVD_CTX(UVD_CGC_MEM_CTRL);
5071 tmp &= ~0x3fff;
5072 WREG32_UVD_CTX(UVD_CGC_MEM_CTRL, tmp);
5073
5074 orig = data = RREG32(UVD_CGC_CTRL);
5075 data &= ~DCM;
5076 if (orig != data)
5077 WREG32(UVD_CGC_CTRL, data);
5078
5079 WREG32_SMC(SMC_CG_IND_START + CG_CGTT_LOCAL_0, 0xffffffff);
5080 WREG32_SMC(SMC_CG_IND_START + CG_CGTT_LOCAL_1, 0xffffffff);
5081 }
5082}
5083
5084static const u32 mc_cg_registers[] =
5085{
5086 MC_HUB_MISC_HUB_CG,
5087 MC_HUB_MISC_SIP_CG,
5088 MC_HUB_MISC_VM_CG,
5089 MC_XPB_CLK_GAT,
5090 ATC_MISC_CG,
5091 MC_CITF_MISC_WR_CG,
5092 MC_CITF_MISC_RD_CG,
5093 MC_CITF_MISC_VM_CG,
5094 VM_L2_CG,
5095};
5096
5097static void si_enable_mc_ls(struct radeon_device *rdev,
5098 bool enable)
5099{
5100 int i;
5101 u32 orig, data;
5102
5103 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) {
5104 orig = data = RREG32(mc_cg_registers[i]);
e16866ec 5105 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_MC_LS))
f8f84ac5
AD
5106 data |= MC_LS_ENABLE;
5107 else
5108 data &= ~MC_LS_ENABLE;
5109 if (data != orig)
5110 WREG32(mc_cg_registers[i], data);
5111 }
5112}
5113
e16866ec
AD
5114static void si_enable_mc_mgcg(struct radeon_device *rdev,
5115 bool enable)
5116{
5117 int i;
5118 u32 orig, data;
5119
5120 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) {
5121 orig = data = RREG32(mc_cg_registers[i]);
5122 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_MC_MGCG))
5123 data |= MC_CG_ENABLE;
5124 else
5125 data &= ~MC_CG_ENABLE;
5126 if (data != orig)
5127 WREG32(mc_cg_registers[i], data);
5128 }
5129}
5130
5131static void si_enable_dma_mgcg(struct radeon_device *rdev,
5132 bool enable)
5133{
5134 u32 orig, data, offset;
5135 int i;
5136
5137 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_SDMA_MGCG)) {
5138 for (i = 0; i < 2; i++) {
5139 if (i == 0)
5140 offset = DMA0_REGISTER_OFFSET;
5141 else
5142 offset = DMA1_REGISTER_OFFSET;
5143 orig = data = RREG32(DMA_POWER_CNTL + offset);
5144 data &= ~MEM_POWER_OVERRIDE;
5145 if (data != orig)
5146 WREG32(DMA_POWER_CNTL + offset, data);
5147 WREG32(DMA_CLK_CTRL + offset, 0x00000100);
5148 }
5149 } else {
5150 for (i = 0; i < 2; i++) {
5151 if (i == 0)
5152 offset = DMA0_REGISTER_OFFSET;
5153 else
5154 offset = DMA1_REGISTER_OFFSET;
5155 orig = data = RREG32(DMA_POWER_CNTL + offset);
5156 data |= MEM_POWER_OVERRIDE;
5157 if (data != orig)
5158 WREG32(DMA_POWER_CNTL + offset, data);
5159
5160 orig = data = RREG32(DMA_CLK_CTRL + offset);
5161 data = 0xff000000;
5162 if (data != orig)
5163 WREG32(DMA_CLK_CTRL + offset, data);
5164 }
5165 }
5166}
5167
5168static void si_enable_bif_mgls(struct radeon_device *rdev,
5169 bool enable)
5170{
5171 u32 orig, data;
5172
5173 orig = data = RREG32_PCIE(PCIE_CNTL2);
5174
5175 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_BIF_LS))
5176 data |= SLV_MEM_LS_EN | MST_MEM_LS_EN |
5177 REPLAY_MEM_LS_EN | SLV_MEM_AGGRESSIVE_LS_EN;
5178 else
5179 data &= ~(SLV_MEM_LS_EN | MST_MEM_LS_EN |
5180 REPLAY_MEM_LS_EN | SLV_MEM_AGGRESSIVE_LS_EN);
5181
5182 if (orig != data)
5183 WREG32_PCIE(PCIE_CNTL2, data);
5184}
5185
5186static void si_enable_hdp_mgcg(struct radeon_device *rdev,
5187 bool enable)
5188{
5189 u32 orig, data;
5190
5191 orig = data = RREG32(HDP_HOST_PATH_CNTL);
5192
5193 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_HDP_MGCG))
5194 data &= ~CLOCK_GATING_DIS;
5195 else
5196 data |= CLOCK_GATING_DIS;
5197
5198 if (orig != data)
5199 WREG32(HDP_HOST_PATH_CNTL, data);
5200}
5201
5202static void si_enable_hdp_ls(struct radeon_device *rdev,
5203 bool enable)
5204{
5205 u32 orig, data;
5206
5207 orig = data = RREG32(HDP_MEM_POWER_LS);
5208
5209 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_HDP_LS))
5210 data |= HDP_LS_ENABLE;
5211 else
5212 data &= ~HDP_LS_ENABLE;
5213
5214 if (orig != data)
5215 WREG32(HDP_MEM_POWER_LS, data);
5216}
5217
5218void si_update_cg(struct radeon_device *rdev,
5219 u32 block, bool enable)
5220{
5221 if (block & RADEON_CG_BLOCK_GFX) {
5222 /* order matters! */
5223 if (enable) {
5224 si_enable_mgcg(rdev, true);
5225 si_enable_cgcg(rdev, true);
5226 } else {
5227 si_enable_cgcg(rdev, false);
5228 si_enable_mgcg(rdev, false);
5229 }
5230 }
5231
5232 if (block & RADEON_CG_BLOCK_MC) {
5233 si_enable_mc_mgcg(rdev, enable);
5234 si_enable_mc_ls(rdev, enable);
5235 }
5236
5237 if (block & RADEON_CG_BLOCK_SDMA) {
5238 si_enable_dma_mgcg(rdev, enable);
5239 }
5240
5241 if (block & RADEON_CG_BLOCK_BIF) {
5242 si_enable_bif_mgls(rdev, enable);
5243 }
5244
5245 if (block & RADEON_CG_BLOCK_UVD) {
5246 if (rdev->has_uvd) {
5247 si_enable_uvd_mgcg(rdev, enable);
5248 }
5249 }
5250
5251 if (block & RADEON_CG_BLOCK_HDP) {
5252 si_enable_hdp_mgcg(rdev, enable);
5253 si_enable_hdp_ls(rdev, enable);
5254 }
5255}
f8f84ac5
AD
5256
5257static void si_init_cg(struct radeon_device *rdev)
5258{
e16866ec
AD
5259 si_update_cg(rdev, (RADEON_CG_BLOCK_GFX |
5260 RADEON_CG_BLOCK_MC |
5261 RADEON_CG_BLOCK_SDMA |
5262 RADEON_CG_BLOCK_BIF |
5263 RADEON_CG_BLOCK_HDP), true);
b2d70917 5264 if (rdev->has_uvd) {
e16866ec 5265 si_update_cg(rdev, RADEON_CG_BLOCK_UVD, true);
f8f84ac5
AD
5266 si_init_uvd_internal_cg(rdev);
5267 }
5268}
5269
5270static void si_fini_cg(struct radeon_device *rdev)
5271{
0116e1ef 5272 if (rdev->has_uvd) {
e16866ec 5273 si_update_cg(rdev, RADEON_CG_BLOCK_UVD, false);
0116e1ef 5274 }
e16866ec
AD
5275 si_update_cg(rdev, (RADEON_CG_BLOCK_GFX |
5276 RADEON_CG_BLOCK_MC |
5277 RADEON_CG_BLOCK_SDMA |
5278 RADEON_CG_BLOCK_BIF |
5279 RADEON_CG_BLOCK_HDP), false);
5280}
5281
5282void si_update_pg(struct radeon_device *rdev,
5283 bool enable)
5284{
5285 si_enable_dma_pg(rdev, enable);
5286 si_enable_gfx_cgpg(rdev, enable);
f8f84ac5
AD
5287}
5288
5289static void si_init_pg(struct radeon_device *rdev)
5290{
0116e1ef
AD
5291 if (rdev->pg_flags) {
5292 if (rdev->pg_flags & RADEON_PG_SUPPORT_SDMA) {
5293 si_init_dma_pg(rdev);
0116e1ef 5294 }
f8f84ac5 5295 si_init_ao_cu_mask(rdev);
0116e1ef
AD
5296 if (rdev->pg_flags & RADEON_PG_SUPPORT_GFX_CG) {
5297 si_init_gfx_cgpg(rdev);
0116e1ef 5298 }
e16866ec 5299 si_update_pg(rdev, false);
f8f84ac5
AD
5300 } else {
5301 WREG32(RLC_SAVE_AND_RESTORE_BASE, rdev->rlc.save_restore_gpu_addr >> 8);
5302 WREG32(RLC_CLEAR_STATE_RESTORE_BASE, rdev->rlc.clear_state_gpu_addr >> 8);
5303 }
5304}
5305
5306static void si_fini_pg(struct radeon_device *rdev)
5307{
0116e1ef
AD
5308 if (rdev->pg_flags) {
5309 if (rdev->pg_flags & RADEON_PG_SUPPORT_SDMA)
5310 si_enable_dma_pg(rdev, false);
5311 if (rdev->pg_flags & RADEON_PG_SUPPORT_GFX_CG)
5312 si_enable_gfx_cgpg(rdev, false);
f8f84ac5
AD
5313 }
5314}
5315
347e7592
AD
5316/*
5317 * RLC
5318 */
866d83de 5319void si_rlc_reset(struct radeon_device *rdev)
d719cef3 5320{
f8f84ac5 5321 u32 tmp = RREG32(GRBM_SOFT_RESET);
d719cef3 5322
f8f84ac5
AD
5323 tmp |= SOFT_RESET_RLC;
5324 WREG32(GRBM_SOFT_RESET, tmp);
5325 udelay(50);
5326 tmp &= ~SOFT_RESET_RLC;
5327 WREG32(GRBM_SOFT_RESET, tmp);
5328 udelay(50);
d719cef3
AD
5329}
5330
347e7592
AD
5331static void si_rlc_stop(struct radeon_device *rdev)
5332{
5333 WREG32(RLC_CNTL, 0);
d719cef3
AD
5334
5335 si_enable_gui_idle_interrupt(rdev, false);
5336
5337 si_wait_for_rlc_serdes(rdev);
347e7592
AD
5338}
5339
5340static void si_rlc_start(struct radeon_device *rdev)
5341{
5342 WREG32(RLC_CNTL, RLC_ENABLE);
d719cef3
AD
5343
5344 si_enable_gui_idle_interrupt(rdev, true);
5345
5346 udelay(50);
5347}
5348
5349static bool si_lbpw_supported(struct radeon_device *rdev)
5350{
5351 u32 tmp;
5352
5353 /* Enable LBPW only for DDR3 */
5354 tmp = RREG32(MC_SEQ_MISC0);
5355 if ((tmp & 0xF0000000) == 0xB0000000)
5356 return true;
5357 return false;
5358}
5359
5360static void si_enable_lbpw(struct radeon_device *rdev, bool enable)
5361{
5362 u32 tmp;
5363
5364 tmp = RREG32(RLC_LB_CNTL);
5365 if (enable)
5366 tmp |= LOAD_BALANCE_ENABLE;
5367 else
5368 tmp &= ~LOAD_BALANCE_ENABLE;
5369 WREG32(RLC_LB_CNTL, tmp);
5370
5371 if (!enable) {
5372 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
5373 WREG32(SPI_LB_CU_MASK, 0x00ff);
5374 }
347e7592
AD
5375}
5376
5377static int si_rlc_resume(struct radeon_device *rdev)
5378{
5379 u32 i;
5380 const __be32 *fw_data;
5381
5382 if (!rdev->rlc_fw)
5383 return -EINVAL;
5384
5385 si_rlc_stop(rdev);
5386
f8f84ac5
AD
5387 si_rlc_reset(rdev);
5388
5389 si_init_pg(rdev);
5390
5391 si_init_cg(rdev);
5392
347e7592
AD
5393 WREG32(RLC_RL_BASE, 0);
5394 WREG32(RLC_RL_SIZE, 0);
5395 WREG32(RLC_LB_CNTL, 0);
5396 WREG32(RLC_LB_CNTR_MAX, 0xffffffff);
5397 WREG32(RLC_LB_CNTR_INIT, 0);
d719cef3 5398 WREG32(RLC_LB_INIT_CU_MASK, 0xffffffff);
347e7592 5399
347e7592
AD
5400 WREG32(RLC_MC_CNTL, 0);
5401 WREG32(RLC_UCODE_CNTL, 0);
5402
5403 fw_data = (const __be32 *)rdev->rlc_fw->data;
5404 for (i = 0; i < SI_RLC_UCODE_SIZE; i++) {
5405 WREG32(RLC_UCODE_ADDR, i);
5406 WREG32(RLC_UCODE_DATA, be32_to_cpup(fw_data++));
5407 }
5408 WREG32(RLC_UCODE_ADDR, 0);
5409
d719cef3
AD
5410 si_enable_lbpw(rdev, si_lbpw_supported(rdev));
5411
347e7592
AD
5412 si_rlc_start(rdev);
5413
5414 return 0;
5415}
5416
25a857fb
AD
5417static void si_enable_interrupts(struct radeon_device *rdev)
5418{
5419 u32 ih_cntl = RREG32(IH_CNTL);
5420 u32 ih_rb_cntl = RREG32(IH_RB_CNTL);
5421
5422 ih_cntl |= ENABLE_INTR;
5423 ih_rb_cntl |= IH_RB_ENABLE;
5424 WREG32(IH_CNTL, ih_cntl);
5425 WREG32(IH_RB_CNTL, ih_rb_cntl);
5426 rdev->ih.enabled = true;
5427}
5428
5429static void si_disable_interrupts(struct radeon_device *rdev)
5430{
5431 u32 ih_rb_cntl = RREG32(IH_RB_CNTL);
5432 u32 ih_cntl = RREG32(IH_CNTL);
5433
5434 ih_rb_cntl &= ~IH_RB_ENABLE;
5435 ih_cntl &= ~ENABLE_INTR;
5436 WREG32(IH_RB_CNTL, ih_rb_cntl);
5437 WREG32(IH_CNTL, ih_cntl);
5438 /* set rptr, wptr to 0 */
5439 WREG32(IH_RB_RPTR, 0);
5440 WREG32(IH_RB_WPTR, 0);
5441 rdev->ih.enabled = false;
25a857fb
AD
5442 rdev->ih.rptr = 0;
5443}
5444
5445static void si_disable_interrupt_state(struct radeon_device *rdev)
5446{
5447 u32 tmp;
5448
5449 WREG32(CP_INT_CNTL_RING0, CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
5450 WREG32(CP_INT_CNTL_RING1, 0);
5451 WREG32(CP_INT_CNTL_RING2, 0);
8c5fd7ef
AD
5452 tmp = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET) & ~TRAP_ENABLE;
5453 WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, tmp);
5454 tmp = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET) & ~TRAP_ENABLE;
5455 WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, tmp);
25a857fb 5456 WREG32(GRBM_INT_CNTL, 0);
5153550a
AD
5457 if (rdev->num_crtc >= 2) {
5458 WREG32(INT_MASK + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
5459 WREG32(INT_MASK + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
5460 }
25a857fb
AD
5461 if (rdev->num_crtc >= 4) {
5462 WREG32(INT_MASK + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
5463 WREG32(INT_MASK + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
5464 }
5465 if (rdev->num_crtc >= 6) {
5466 WREG32(INT_MASK + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
5467 WREG32(INT_MASK + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
5468 }
5469
5153550a
AD
5470 if (rdev->num_crtc >= 2) {
5471 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
5472 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
5473 }
25a857fb
AD
5474 if (rdev->num_crtc >= 4) {
5475 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
5476 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
5477 }
5478 if (rdev->num_crtc >= 6) {
5479 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
5480 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
5481 }
5482
5153550a
AD
5483 if (!ASIC_IS_NODCE(rdev)) {
5484 WREG32(DACA_AUTODETECT_INT_CONTROL, 0);
5485
5486 tmp = RREG32(DC_HPD1_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5487 WREG32(DC_HPD1_INT_CONTROL, tmp);
5488 tmp = RREG32(DC_HPD2_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5489 WREG32(DC_HPD2_INT_CONTROL, tmp);
5490 tmp = RREG32(DC_HPD3_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5491 WREG32(DC_HPD3_INT_CONTROL, tmp);
5492 tmp = RREG32(DC_HPD4_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5493 WREG32(DC_HPD4_INT_CONTROL, tmp);
5494 tmp = RREG32(DC_HPD5_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5495 WREG32(DC_HPD5_INT_CONTROL, tmp);
5496 tmp = RREG32(DC_HPD6_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5497 WREG32(DC_HPD6_INT_CONTROL, tmp);
5498 }
25a857fb
AD
5499}
5500
5501static int si_irq_init(struct radeon_device *rdev)
5502{
5503 int ret = 0;
5504 int rb_bufsz;
5505 u32 interrupt_cntl, ih_cntl, ih_rb_cntl;
5506
5507 /* allocate ring */
5508 ret = r600_ih_ring_alloc(rdev);
5509 if (ret)
5510 return ret;
5511
5512 /* disable irqs */
5513 si_disable_interrupts(rdev);
5514
5515 /* init rlc */
5516 ret = si_rlc_resume(rdev);
5517 if (ret) {
5518 r600_ih_ring_fini(rdev);
5519 return ret;
5520 }
5521
5522 /* setup interrupt control */
5523 /* set dummy read address to ring address */
5524 WREG32(INTERRUPT_CNTL2, rdev->ih.gpu_addr >> 8);
5525 interrupt_cntl = RREG32(INTERRUPT_CNTL);
5526 /* IH_DUMMY_RD_OVERRIDE=0 - dummy read disabled with msi, enabled without msi
5527 * IH_DUMMY_RD_OVERRIDE=1 - dummy read controlled by IH_DUMMY_RD_EN
5528 */
5529 interrupt_cntl &= ~IH_DUMMY_RD_OVERRIDE;
5530 /* IH_REQ_NONSNOOP_EN=1 if ring is in non-cacheable memory, e.g., vram */
5531 interrupt_cntl &= ~IH_REQ_NONSNOOP_EN;
5532 WREG32(INTERRUPT_CNTL, interrupt_cntl);
5533
5534 WREG32(IH_RB_BASE, rdev->ih.gpu_addr >> 8);
5535 rb_bufsz = drm_order(rdev->ih.ring_size / 4);
5536
5537 ih_rb_cntl = (IH_WPTR_OVERFLOW_ENABLE |
5538 IH_WPTR_OVERFLOW_CLEAR |
5539 (rb_bufsz << 1));
5540
5541 if (rdev->wb.enabled)
5542 ih_rb_cntl |= IH_WPTR_WRITEBACK_ENABLE;
5543
5544 /* set the writeback address whether it's enabled or not */
5545 WREG32(IH_RB_WPTR_ADDR_LO, (rdev->wb.gpu_addr + R600_WB_IH_WPTR_OFFSET) & 0xFFFFFFFC);
5546 WREG32(IH_RB_WPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + R600_WB_IH_WPTR_OFFSET) & 0xFF);
5547
5548 WREG32(IH_RB_CNTL, ih_rb_cntl);
5549
5550 /* set rptr, wptr to 0 */
5551 WREG32(IH_RB_RPTR, 0);
5552 WREG32(IH_RB_WPTR, 0);
5553
5554 /* Default settings for IH_CNTL (disabled at first) */
5555 ih_cntl = MC_WRREQ_CREDIT(0x10) | MC_WR_CLEAN_CNT(0x10) | MC_VMID(0);
5556 /* RPTR_REARM only works if msi's are enabled */
5557 if (rdev->msi_enabled)
5558 ih_cntl |= RPTR_REARM;
5559 WREG32(IH_CNTL, ih_cntl);
5560
5561 /* force the active interrupt state to all disabled */
5562 si_disable_interrupt_state(rdev);
5563
2099810f
DA
5564 pci_set_master(rdev->pdev);
5565
25a857fb
AD
5566 /* enable irqs */
5567 si_enable_interrupts(rdev);
5568
5569 return ret;
5570}
5571
5572int si_irq_set(struct radeon_device *rdev)
5573{
5574 u32 cp_int_cntl = CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE;
5575 u32 cp_int_cntl1 = 0, cp_int_cntl2 = 0;
5576 u32 crtc1 = 0, crtc2 = 0, crtc3 = 0, crtc4 = 0, crtc5 = 0, crtc6 = 0;
5153550a 5577 u32 hpd1 = 0, hpd2 = 0, hpd3 = 0, hpd4 = 0, hpd5 = 0, hpd6 = 0;
25a857fb
AD
5578 u32 grbm_int_cntl = 0;
5579 u32 grph1 = 0, grph2 = 0, grph3 = 0, grph4 = 0, grph5 = 0, grph6 = 0;
8c5fd7ef 5580 u32 dma_cntl, dma_cntl1;
a9e61410 5581 u32 thermal_int = 0;
25a857fb
AD
5582
5583 if (!rdev->irq.installed) {
5584 WARN(1, "Can't enable IRQ/MSI because no handler is installed\n");
5585 return -EINVAL;
5586 }
5587 /* don't enable anything if the ih is disabled */
5588 if (!rdev->ih.enabled) {
5589 si_disable_interrupts(rdev);
5590 /* force the active interrupt state to all disabled */
5591 si_disable_interrupt_state(rdev);
5592 return 0;
5593 }
5594
5153550a
AD
5595 if (!ASIC_IS_NODCE(rdev)) {
5596 hpd1 = RREG32(DC_HPD1_INT_CONTROL) & ~DC_HPDx_INT_EN;
5597 hpd2 = RREG32(DC_HPD2_INT_CONTROL) & ~DC_HPDx_INT_EN;
5598 hpd3 = RREG32(DC_HPD3_INT_CONTROL) & ~DC_HPDx_INT_EN;
5599 hpd4 = RREG32(DC_HPD4_INT_CONTROL) & ~DC_HPDx_INT_EN;
5600 hpd5 = RREG32(DC_HPD5_INT_CONTROL) & ~DC_HPDx_INT_EN;
5601 hpd6 = RREG32(DC_HPD6_INT_CONTROL) & ~DC_HPDx_INT_EN;
5602 }
25a857fb 5603
8c5fd7ef
AD
5604 dma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET) & ~TRAP_ENABLE;
5605 dma_cntl1 = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET) & ~TRAP_ENABLE;
5606
a9e61410
AD
5607 thermal_int = RREG32(CG_THERMAL_INT) &
5608 ~(THERM_INT_MASK_HIGH | THERM_INT_MASK_LOW);
5609
25a857fb 5610 /* enable CP interrupts on all rings */
736fc37f 5611 if (atomic_read(&rdev->irq.ring_int[RADEON_RING_TYPE_GFX_INDEX])) {
25a857fb
AD
5612 DRM_DEBUG("si_irq_set: sw int gfx\n");
5613 cp_int_cntl |= TIME_STAMP_INT_ENABLE;
5614 }
736fc37f 5615 if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_CP1_INDEX])) {
25a857fb
AD
5616 DRM_DEBUG("si_irq_set: sw int cp1\n");
5617 cp_int_cntl1 |= TIME_STAMP_INT_ENABLE;
5618 }
736fc37f 5619 if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_CP2_INDEX])) {
25a857fb
AD
5620 DRM_DEBUG("si_irq_set: sw int cp2\n");
5621 cp_int_cntl2 |= TIME_STAMP_INT_ENABLE;
5622 }
8c5fd7ef
AD
5623 if (atomic_read(&rdev->irq.ring_int[R600_RING_TYPE_DMA_INDEX])) {
5624 DRM_DEBUG("si_irq_set: sw int dma\n");
5625 dma_cntl |= TRAP_ENABLE;
5626 }
5627
5628 if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_DMA1_INDEX])) {
5629 DRM_DEBUG("si_irq_set: sw int dma1\n");
5630 dma_cntl1 |= TRAP_ENABLE;
5631 }
25a857fb 5632 if (rdev->irq.crtc_vblank_int[0] ||
736fc37f 5633 atomic_read(&rdev->irq.pflip[0])) {
25a857fb
AD
5634 DRM_DEBUG("si_irq_set: vblank 0\n");
5635 crtc1 |= VBLANK_INT_MASK;
5636 }
5637 if (rdev->irq.crtc_vblank_int[1] ||
736fc37f 5638 atomic_read(&rdev->irq.pflip[1])) {
25a857fb
AD
5639 DRM_DEBUG("si_irq_set: vblank 1\n");
5640 crtc2 |= VBLANK_INT_MASK;
5641 }
5642 if (rdev->irq.crtc_vblank_int[2] ||
736fc37f 5643 atomic_read(&rdev->irq.pflip[2])) {
25a857fb
AD
5644 DRM_DEBUG("si_irq_set: vblank 2\n");
5645 crtc3 |= VBLANK_INT_MASK;
5646 }
5647 if (rdev->irq.crtc_vblank_int[3] ||
736fc37f 5648 atomic_read(&rdev->irq.pflip[3])) {
25a857fb
AD
5649 DRM_DEBUG("si_irq_set: vblank 3\n");
5650 crtc4 |= VBLANK_INT_MASK;
5651 }
5652 if (rdev->irq.crtc_vblank_int[4] ||
736fc37f 5653 atomic_read(&rdev->irq.pflip[4])) {
25a857fb
AD
5654 DRM_DEBUG("si_irq_set: vblank 4\n");
5655 crtc5 |= VBLANK_INT_MASK;
5656 }
5657 if (rdev->irq.crtc_vblank_int[5] ||
736fc37f 5658 atomic_read(&rdev->irq.pflip[5])) {
25a857fb
AD
5659 DRM_DEBUG("si_irq_set: vblank 5\n");
5660 crtc6 |= VBLANK_INT_MASK;
5661 }
5662 if (rdev->irq.hpd[0]) {
5663 DRM_DEBUG("si_irq_set: hpd 1\n");
5664 hpd1 |= DC_HPDx_INT_EN;
5665 }
5666 if (rdev->irq.hpd[1]) {
5667 DRM_DEBUG("si_irq_set: hpd 2\n");
5668 hpd2 |= DC_HPDx_INT_EN;
5669 }
5670 if (rdev->irq.hpd[2]) {
5671 DRM_DEBUG("si_irq_set: hpd 3\n");
5672 hpd3 |= DC_HPDx_INT_EN;
5673 }
5674 if (rdev->irq.hpd[3]) {
5675 DRM_DEBUG("si_irq_set: hpd 4\n");
5676 hpd4 |= DC_HPDx_INT_EN;
5677 }
5678 if (rdev->irq.hpd[4]) {
5679 DRM_DEBUG("si_irq_set: hpd 5\n");
5680 hpd5 |= DC_HPDx_INT_EN;
5681 }
5682 if (rdev->irq.hpd[5]) {
5683 DRM_DEBUG("si_irq_set: hpd 6\n");
5684 hpd6 |= DC_HPDx_INT_EN;
5685 }
25a857fb
AD
5686
5687 WREG32(CP_INT_CNTL_RING0, cp_int_cntl);
5688 WREG32(CP_INT_CNTL_RING1, cp_int_cntl1);
5689 WREG32(CP_INT_CNTL_RING2, cp_int_cntl2);
5690
8c5fd7ef
AD
5691 WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, dma_cntl);
5692 WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, dma_cntl1);
5693
25a857fb
AD
5694 WREG32(GRBM_INT_CNTL, grbm_int_cntl);
5695
a9e61410
AD
5696 if (rdev->irq.dpm_thermal) {
5697 DRM_DEBUG("dpm thermal\n");
5698 thermal_int |= THERM_INT_MASK_HIGH | THERM_INT_MASK_LOW;
5699 }
5700
5153550a
AD
5701 if (rdev->num_crtc >= 2) {
5702 WREG32(INT_MASK + EVERGREEN_CRTC0_REGISTER_OFFSET, crtc1);
5703 WREG32(INT_MASK + EVERGREEN_CRTC1_REGISTER_OFFSET, crtc2);
5704 }
25a857fb
AD
5705 if (rdev->num_crtc >= 4) {
5706 WREG32(INT_MASK + EVERGREEN_CRTC2_REGISTER_OFFSET, crtc3);
5707 WREG32(INT_MASK + EVERGREEN_CRTC3_REGISTER_OFFSET, crtc4);
5708 }
5709 if (rdev->num_crtc >= 6) {
5710 WREG32(INT_MASK + EVERGREEN_CRTC4_REGISTER_OFFSET, crtc5);
5711 WREG32(INT_MASK + EVERGREEN_CRTC5_REGISTER_OFFSET, crtc6);
5712 }
5713
5153550a
AD
5714 if (rdev->num_crtc >= 2) {
5715 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET, grph1);
5716 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET, grph2);
5717 }
25a857fb
AD
5718 if (rdev->num_crtc >= 4) {
5719 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET, grph3);
5720 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET, grph4);
5721 }
5722 if (rdev->num_crtc >= 6) {
5723 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET, grph5);
5724 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET, grph6);
5725 }
5726
5153550a
AD
5727 if (!ASIC_IS_NODCE(rdev)) {
5728 WREG32(DC_HPD1_INT_CONTROL, hpd1);
5729 WREG32(DC_HPD2_INT_CONTROL, hpd2);
5730 WREG32(DC_HPD3_INT_CONTROL, hpd3);
5731 WREG32(DC_HPD4_INT_CONTROL, hpd4);
5732 WREG32(DC_HPD5_INT_CONTROL, hpd5);
5733 WREG32(DC_HPD6_INT_CONTROL, hpd6);
5734 }
25a857fb 5735
a9e61410
AD
5736 WREG32(CG_THERMAL_INT, thermal_int);
5737
25a857fb
AD
5738 return 0;
5739}
5740
5741static inline void si_irq_ack(struct radeon_device *rdev)
5742{
5743 u32 tmp;
5744
5153550a
AD
5745 if (ASIC_IS_NODCE(rdev))
5746 return;
5747
25a857fb
AD
5748 rdev->irq.stat_regs.evergreen.disp_int = RREG32(DISP_INTERRUPT_STATUS);
5749 rdev->irq.stat_regs.evergreen.disp_int_cont = RREG32(DISP_INTERRUPT_STATUS_CONTINUE);
5750 rdev->irq.stat_regs.evergreen.disp_int_cont2 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE2);
5751 rdev->irq.stat_regs.evergreen.disp_int_cont3 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE3);
5752 rdev->irq.stat_regs.evergreen.disp_int_cont4 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE4);
5753 rdev->irq.stat_regs.evergreen.disp_int_cont5 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE5);
5754 rdev->irq.stat_regs.evergreen.d1grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET);
5755 rdev->irq.stat_regs.evergreen.d2grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET);
5756 if (rdev->num_crtc >= 4) {
5757 rdev->irq.stat_regs.evergreen.d3grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET);
5758 rdev->irq.stat_regs.evergreen.d4grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET);
5759 }
5760 if (rdev->num_crtc >= 6) {
5761 rdev->irq.stat_regs.evergreen.d5grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET);
5762 rdev->irq.stat_regs.evergreen.d6grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET);
5763 }
5764
5765 if (rdev->irq.stat_regs.evergreen.d1grph_int & GRPH_PFLIP_INT_OCCURRED)
5766 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
5767 if (rdev->irq.stat_regs.evergreen.d2grph_int & GRPH_PFLIP_INT_OCCURRED)
5768 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
5769 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VBLANK_INTERRUPT)
5770 WREG32(VBLANK_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, VBLANK_ACK);
5771 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VLINE_INTERRUPT)
5772 WREG32(VLINE_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, VLINE_ACK);
5773 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VBLANK_INTERRUPT)
5774 WREG32(VBLANK_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, VBLANK_ACK);
5775 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VLINE_INTERRUPT)
5776 WREG32(VLINE_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, VLINE_ACK);
5777
5778 if (rdev->num_crtc >= 4) {
5779 if (rdev->irq.stat_regs.evergreen.d3grph_int & GRPH_PFLIP_INT_OCCURRED)
5780 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
5781 if (rdev->irq.stat_regs.evergreen.d4grph_int & GRPH_PFLIP_INT_OCCURRED)
5782 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
5783 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VBLANK_INTERRUPT)
5784 WREG32(VBLANK_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, VBLANK_ACK);
5785 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VLINE_INTERRUPT)
5786 WREG32(VLINE_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, VLINE_ACK);
5787 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VBLANK_INTERRUPT)
5788 WREG32(VBLANK_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, VBLANK_ACK);
5789 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VLINE_INTERRUPT)
5790 WREG32(VLINE_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, VLINE_ACK);
5791 }
5792
5793 if (rdev->num_crtc >= 6) {
5794 if (rdev->irq.stat_regs.evergreen.d5grph_int & GRPH_PFLIP_INT_OCCURRED)
5795 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
5796 if (rdev->irq.stat_regs.evergreen.d6grph_int & GRPH_PFLIP_INT_OCCURRED)
5797 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
5798 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VBLANK_INTERRUPT)
5799 WREG32(VBLANK_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, VBLANK_ACK);
5800 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VLINE_INTERRUPT)
5801 WREG32(VLINE_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, VLINE_ACK);
5802 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VBLANK_INTERRUPT)
5803 WREG32(VBLANK_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, VBLANK_ACK);
5804 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VLINE_INTERRUPT)
5805 WREG32(VLINE_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, VLINE_ACK);
5806 }
5807
5808 if (rdev->irq.stat_regs.evergreen.disp_int & DC_HPD1_INTERRUPT) {
5809 tmp = RREG32(DC_HPD1_INT_CONTROL);
5810 tmp |= DC_HPDx_INT_ACK;
5811 WREG32(DC_HPD1_INT_CONTROL, tmp);
5812 }
5813 if (rdev->irq.stat_regs.evergreen.disp_int_cont & DC_HPD2_INTERRUPT) {
5814 tmp = RREG32(DC_HPD2_INT_CONTROL);
5815 tmp |= DC_HPDx_INT_ACK;
5816 WREG32(DC_HPD2_INT_CONTROL, tmp);
5817 }
5818 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & DC_HPD3_INTERRUPT) {
5819 tmp = RREG32(DC_HPD3_INT_CONTROL);
5820 tmp |= DC_HPDx_INT_ACK;
5821 WREG32(DC_HPD3_INT_CONTROL, tmp);
5822 }
5823 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & DC_HPD4_INTERRUPT) {
5824 tmp = RREG32(DC_HPD4_INT_CONTROL);
5825 tmp |= DC_HPDx_INT_ACK;
5826 WREG32(DC_HPD4_INT_CONTROL, tmp);
5827 }
5828 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & DC_HPD5_INTERRUPT) {
5829 tmp = RREG32(DC_HPD5_INT_CONTROL);
5830 tmp |= DC_HPDx_INT_ACK;
5831 WREG32(DC_HPD5_INT_CONTROL, tmp);
5832 }
5833 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & DC_HPD6_INTERRUPT) {
5834 tmp = RREG32(DC_HPD5_INT_CONTROL);
5835 tmp |= DC_HPDx_INT_ACK;
5836 WREG32(DC_HPD6_INT_CONTROL, tmp);
5837 }
5838}
5839
5840static void si_irq_disable(struct radeon_device *rdev)
5841{
5842 si_disable_interrupts(rdev);
5843 /* Wait and acknowledge irq */
5844 mdelay(1);
5845 si_irq_ack(rdev);
5846 si_disable_interrupt_state(rdev);
5847}
5848
5849static void si_irq_suspend(struct radeon_device *rdev)
5850{
5851 si_irq_disable(rdev);
5852 si_rlc_stop(rdev);
5853}
5854
9b136d51
AD
5855static void si_irq_fini(struct radeon_device *rdev)
5856{
5857 si_irq_suspend(rdev);
5858 r600_ih_ring_fini(rdev);
5859}
5860
25a857fb
AD
5861static inline u32 si_get_ih_wptr(struct radeon_device *rdev)
5862{
5863 u32 wptr, tmp;
5864
5865 if (rdev->wb.enabled)
5866 wptr = le32_to_cpu(rdev->wb.wb[R600_WB_IH_WPTR_OFFSET/4]);
5867 else
5868 wptr = RREG32(IH_RB_WPTR);
5869
5870 if (wptr & RB_OVERFLOW) {
5871 /* When a ring buffer overflow happen start parsing interrupt
5872 * from the last not overwritten vector (wptr + 16). Hopefully
5873 * this should allow us to catchup.
5874 */
5875 dev_warn(rdev->dev, "IH ring buffer overflow (0x%08X, %d, %d)\n",
5876 wptr, rdev->ih.rptr, (wptr + 16) + rdev->ih.ptr_mask);
5877 rdev->ih.rptr = (wptr + 16) & rdev->ih.ptr_mask;
5878 tmp = RREG32(IH_RB_CNTL);
5879 tmp |= IH_WPTR_OVERFLOW_CLEAR;
5880 WREG32(IH_RB_CNTL, tmp);
5881 }
5882 return (wptr & rdev->ih.ptr_mask);
5883}
5884
5885/* SI IV Ring
5886 * Each IV ring entry is 128 bits:
5887 * [7:0] - interrupt source id
5888 * [31:8] - reserved
5889 * [59:32] - interrupt source data
5890 * [63:60] - reserved
5891 * [71:64] - RINGID
5892 * [79:72] - VMID
5893 * [127:80] - reserved
5894 */
5895int si_irq_process(struct radeon_device *rdev)
5896{
5897 u32 wptr;
5898 u32 rptr;
5899 u32 src_id, src_data, ring_id;
5900 u32 ring_index;
25a857fb 5901 bool queue_hotplug = false;
a9e61410 5902 bool queue_thermal = false;
fbf6dc7a 5903 u32 status, addr;
25a857fb
AD
5904
5905 if (!rdev->ih.enabled || rdev->shutdown)
5906 return IRQ_NONE;
5907
5908 wptr = si_get_ih_wptr(rdev);
c20dc369
CK
5909
5910restart_ih:
5911 /* is somebody else already processing irqs? */
5912 if (atomic_xchg(&rdev->ih.lock, 1))
5913 return IRQ_NONE;
5914
25a857fb
AD
5915 rptr = rdev->ih.rptr;
5916 DRM_DEBUG("si_irq_process start: rptr %d, wptr %d\n", rptr, wptr);
5917
25a857fb
AD
5918 /* Order reading of wptr vs. reading of IH ring data */
5919 rmb();
5920
5921 /* display interrupts */
5922 si_irq_ack(rdev);
5923
25a857fb
AD
5924 while (rptr != wptr) {
5925 /* wptr/rptr are in bytes! */
5926 ring_index = rptr / 4;
5927 src_id = le32_to_cpu(rdev->ih.ring[ring_index]) & 0xff;
5928 src_data = le32_to_cpu(rdev->ih.ring[ring_index + 1]) & 0xfffffff;
5929 ring_id = le32_to_cpu(rdev->ih.ring[ring_index + 2]) & 0xff;
5930
5931 switch (src_id) {
5932 case 1: /* D1 vblank/vline */
5933 switch (src_data) {
5934 case 0: /* D1 vblank */
5935 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VBLANK_INTERRUPT) {
5936 if (rdev->irq.crtc_vblank_int[0]) {
5937 drm_handle_vblank(rdev->ddev, 0);
5938 rdev->pm.vblank_sync = true;
5939 wake_up(&rdev->irq.vblank_queue);
5940 }
736fc37f 5941 if (atomic_read(&rdev->irq.pflip[0]))
25a857fb
AD
5942 radeon_crtc_handle_flip(rdev, 0);
5943 rdev->irq.stat_regs.evergreen.disp_int &= ~LB_D1_VBLANK_INTERRUPT;
5944 DRM_DEBUG("IH: D1 vblank\n");
5945 }
5946 break;
5947 case 1: /* D1 vline */
5948 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VLINE_INTERRUPT) {
5949 rdev->irq.stat_regs.evergreen.disp_int &= ~LB_D1_VLINE_INTERRUPT;
5950 DRM_DEBUG("IH: D1 vline\n");
5951 }
5952 break;
5953 default:
5954 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
5955 break;
5956 }
5957 break;
5958 case 2: /* D2 vblank/vline */
5959 switch (src_data) {
5960 case 0: /* D2 vblank */
5961 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VBLANK_INTERRUPT) {
5962 if (rdev->irq.crtc_vblank_int[1]) {
5963 drm_handle_vblank(rdev->ddev, 1);
5964 rdev->pm.vblank_sync = true;
5965 wake_up(&rdev->irq.vblank_queue);
5966 }
736fc37f 5967 if (atomic_read(&rdev->irq.pflip[1]))
25a857fb
AD
5968 radeon_crtc_handle_flip(rdev, 1);
5969 rdev->irq.stat_regs.evergreen.disp_int_cont &= ~LB_D2_VBLANK_INTERRUPT;
5970 DRM_DEBUG("IH: D2 vblank\n");
5971 }
5972 break;
5973 case 1: /* D2 vline */
5974 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VLINE_INTERRUPT) {
5975 rdev->irq.stat_regs.evergreen.disp_int_cont &= ~LB_D2_VLINE_INTERRUPT;
5976 DRM_DEBUG("IH: D2 vline\n");
5977 }
5978 break;
5979 default:
5980 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
5981 break;
5982 }
5983 break;
5984 case 3: /* D3 vblank/vline */
5985 switch (src_data) {
5986 case 0: /* D3 vblank */
5987 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VBLANK_INTERRUPT) {
5988 if (rdev->irq.crtc_vblank_int[2]) {
5989 drm_handle_vblank(rdev->ddev, 2);
5990 rdev->pm.vblank_sync = true;
5991 wake_up(&rdev->irq.vblank_queue);
5992 }
736fc37f 5993 if (atomic_read(&rdev->irq.pflip[2]))
25a857fb
AD
5994 radeon_crtc_handle_flip(rdev, 2);
5995 rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~LB_D3_VBLANK_INTERRUPT;
5996 DRM_DEBUG("IH: D3 vblank\n");
5997 }
5998 break;
5999 case 1: /* D3 vline */
6000 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VLINE_INTERRUPT) {
6001 rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~LB_D3_VLINE_INTERRUPT;
6002 DRM_DEBUG("IH: D3 vline\n");
6003 }
6004 break;
6005 default:
6006 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6007 break;
6008 }
6009 break;
6010 case 4: /* D4 vblank/vline */
6011 switch (src_data) {
6012 case 0: /* D4 vblank */
6013 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VBLANK_INTERRUPT) {
6014 if (rdev->irq.crtc_vblank_int[3]) {
6015 drm_handle_vblank(rdev->ddev, 3);
6016 rdev->pm.vblank_sync = true;
6017 wake_up(&rdev->irq.vblank_queue);
6018 }
736fc37f 6019 if (atomic_read(&rdev->irq.pflip[3]))
25a857fb
AD
6020 radeon_crtc_handle_flip(rdev, 3);
6021 rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~LB_D4_VBLANK_INTERRUPT;
6022 DRM_DEBUG("IH: D4 vblank\n");
6023 }
6024 break;
6025 case 1: /* D4 vline */
6026 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VLINE_INTERRUPT) {
6027 rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~LB_D4_VLINE_INTERRUPT;
6028 DRM_DEBUG("IH: D4 vline\n");
6029 }
6030 break;
6031 default:
6032 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6033 break;
6034 }
6035 break;
6036 case 5: /* D5 vblank/vline */
6037 switch (src_data) {
6038 case 0: /* D5 vblank */
6039 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VBLANK_INTERRUPT) {
6040 if (rdev->irq.crtc_vblank_int[4]) {
6041 drm_handle_vblank(rdev->ddev, 4);
6042 rdev->pm.vblank_sync = true;
6043 wake_up(&rdev->irq.vblank_queue);
6044 }
736fc37f 6045 if (atomic_read(&rdev->irq.pflip[4]))
25a857fb
AD
6046 radeon_crtc_handle_flip(rdev, 4);
6047 rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~LB_D5_VBLANK_INTERRUPT;
6048 DRM_DEBUG("IH: D5 vblank\n");
6049 }
6050 break;
6051 case 1: /* D5 vline */
6052 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VLINE_INTERRUPT) {
6053 rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~LB_D5_VLINE_INTERRUPT;
6054 DRM_DEBUG("IH: D5 vline\n");
6055 }
6056 break;
6057 default:
6058 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6059 break;
6060 }
6061 break;
6062 case 6: /* D6 vblank/vline */
6063 switch (src_data) {
6064 case 0: /* D6 vblank */
6065 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VBLANK_INTERRUPT) {
6066 if (rdev->irq.crtc_vblank_int[5]) {
6067 drm_handle_vblank(rdev->ddev, 5);
6068 rdev->pm.vblank_sync = true;
6069 wake_up(&rdev->irq.vblank_queue);
6070 }
736fc37f 6071 if (atomic_read(&rdev->irq.pflip[5]))
25a857fb
AD
6072 radeon_crtc_handle_flip(rdev, 5);
6073 rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~LB_D6_VBLANK_INTERRUPT;
6074 DRM_DEBUG("IH: D6 vblank\n");
6075 }
6076 break;
6077 case 1: /* D6 vline */
6078 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VLINE_INTERRUPT) {
6079 rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~LB_D6_VLINE_INTERRUPT;
6080 DRM_DEBUG("IH: D6 vline\n");
6081 }
6082 break;
6083 default:
6084 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6085 break;
6086 }
6087 break;
6088 case 42: /* HPD hotplug */
6089 switch (src_data) {
6090 case 0:
6091 if (rdev->irq.stat_regs.evergreen.disp_int & DC_HPD1_INTERRUPT) {
6092 rdev->irq.stat_regs.evergreen.disp_int &= ~DC_HPD1_INTERRUPT;
6093 queue_hotplug = true;
6094 DRM_DEBUG("IH: HPD1\n");
6095 }
6096 break;
6097 case 1:
6098 if (rdev->irq.stat_regs.evergreen.disp_int_cont & DC_HPD2_INTERRUPT) {
6099 rdev->irq.stat_regs.evergreen.disp_int_cont &= ~DC_HPD2_INTERRUPT;
6100 queue_hotplug = true;
6101 DRM_DEBUG("IH: HPD2\n");
6102 }
6103 break;
6104 case 2:
6105 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & DC_HPD3_INTERRUPT) {
6106 rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~DC_HPD3_INTERRUPT;
6107 queue_hotplug = true;
6108 DRM_DEBUG("IH: HPD3\n");
6109 }
6110 break;
6111 case 3:
6112 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & DC_HPD4_INTERRUPT) {
6113 rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~DC_HPD4_INTERRUPT;
6114 queue_hotplug = true;
6115 DRM_DEBUG("IH: HPD4\n");
6116 }
6117 break;
6118 case 4:
6119 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & DC_HPD5_INTERRUPT) {
6120 rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~DC_HPD5_INTERRUPT;
6121 queue_hotplug = true;
6122 DRM_DEBUG("IH: HPD5\n");
6123 }
6124 break;
6125 case 5:
6126 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & DC_HPD6_INTERRUPT) {
6127 rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~DC_HPD6_INTERRUPT;
6128 queue_hotplug = true;
6129 DRM_DEBUG("IH: HPD6\n");
6130 }
6131 break;
6132 default:
6133 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6134 break;
6135 }
6136 break;
ae133a11
CK
6137 case 146:
6138 case 147:
fbf6dc7a
AD
6139 addr = RREG32(VM_CONTEXT1_PROTECTION_FAULT_ADDR);
6140 status = RREG32(VM_CONTEXT1_PROTECTION_FAULT_STATUS);
ae133a11
CK
6141 dev_err(rdev->dev, "GPU fault detected: %d 0x%08x\n", src_id, src_data);
6142 dev_err(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n",
fbf6dc7a 6143 addr);
ae133a11 6144 dev_err(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
fbf6dc7a
AD
6145 status);
6146 si_vm_decode_fault(rdev, status, addr);
ae133a11
CK
6147 /* reset addr and status */
6148 WREG32_P(VM_CONTEXT1_CNTL2, 1, ~1);
6149 break;
25a857fb
AD
6150 case 176: /* RINGID0 CP_INT */
6151 radeon_fence_process(rdev, RADEON_RING_TYPE_GFX_INDEX);
6152 break;
6153 case 177: /* RINGID1 CP_INT */
6154 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
6155 break;
6156 case 178: /* RINGID2 CP_INT */
6157 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
6158 break;
6159 case 181: /* CP EOP event */
6160 DRM_DEBUG("IH: CP EOP\n");
6161 switch (ring_id) {
6162 case 0:
6163 radeon_fence_process(rdev, RADEON_RING_TYPE_GFX_INDEX);
6164 break;
6165 case 1:
6166 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
6167 break;
6168 case 2:
6169 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
6170 break;
6171 }
6172 break;
8c5fd7ef
AD
6173 case 224: /* DMA trap event */
6174 DRM_DEBUG("IH: DMA trap\n");
6175 radeon_fence_process(rdev, R600_RING_TYPE_DMA_INDEX);
6176 break;
a9e61410
AD
6177 case 230: /* thermal low to high */
6178 DRM_DEBUG("IH: thermal low to high\n");
6179 rdev->pm.dpm.thermal.high_to_low = false;
6180 queue_thermal = true;
6181 break;
6182 case 231: /* thermal high to low */
6183 DRM_DEBUG("IH: thermal high to low\n");
6184 rdev->pm.dpm.thermal.high_to_low = true;
6185 queue_thermal = true;
6186 break;
25a857fb
AD
6187 case 233: /* GUI IDLE */
6188 DRM_DEBUG("IH: GUI idle\n");
25a857fb 6189 break;
8c5fd7ef
AD
6190 case 244: /* DMA trap event */
6191 DRM_DEBUG("IH: DMA1 trap\n");
6192 radeon_fence_process(rdev, CAYMAN_RING_TYPE_DMA1_INDEX);
6193 break;
25a857fb
AD
6194 default:
6195 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6196 break;
6197 }
6198
6199 /* wptr/rptr are in bytes! */
6200 rptr += 16;
6201 rptr &= rdev->ih.ptr_mask;
6202 }
25a857fb
AD
6203 if (queue_hotplug)
6204 schedule_work(&rdev->hotplug_work);
a9e61410
AD
6205 if (queue_thermal && rdev->pm.dpm_enabled)
6206 schedule_work(&rdev->pm.dpm.thermal.work);
25a857fb
AD
6207 rdev->ih.rptr = rptr;
6208 WREG32(IH_RB_RPTR, rdev->ih.rptr);
c20dc369
CK
6209 atomic_set(&rdev->ih.lock, 0);
6210
6211 /* make sure wptr hasn't changed while processing */
6212 wptr = si_get_ih_wptr(rdev);
6213 if (wptr != rptr)
6214 goto restart_ih;
6215
25a857fb
AD
6216 return IRQ_HANDLED;
6217}
6218
9b136d51
AD
6219/*
6220 * startup/shutdown callbacks
6221 */
6222static int si_startup(struct radeon_device *rdev)
6223{
6224 struct radeon_ring *ring;
6225 int r;
6226
b9d305df
AD
6227 /* enable pcie gen2/3 link */
6228 si_pcie_gen3_enable(rdev);
e0bcf165
AD
6229 /* enable aspm */
6230 si_program_aspm(rdev);
b9d305df 6231
6fab3feb
AD
6232 si_mc_program(rdev);
6233
9b136d51
AD
6234 if (!rdev->me_fw || !rdev->pfp_fw || !rdev->ce_fw ||
6235 !rdev->rlc_fw || !rdev->mc_fw) {
6236 r = si_init_microcode(rdev);
6237 if (r) {
6238 DRM_ERROR("Failed to load firmware!\n");
6239 return r;
6240 }
6241 }
6242
6243 r = si_mc_load_microcode(rdev);
6244 if (r) {
6245 DRM_ERROR("Failed to load MC firmware!\n");
6246 return r;
6247 }
6248
6249 r = r600_vram_scratch_init(rdev);
6250 if (r)
6251 return r;
6252
9b136d51
AD
6253 r = si_pcie_gart_enable(rdev);
6254 if (r)
6255 return r;
6256 si_gpu_init(rdev);
6257
9b136d51 6258 /* allocate rlc buffers */
1fd11777
AD
6259 if (rdev->family == CHIP_VERDE) {
6260 rdev->rlc.reg_list = verde_rlc_save_restore_register_list;
6261 rdev->rlc.reg_list_size =
6262 (u32)ARRAY_SIZE(verde_rlc_save_restore_register_list);
6263 }
6264 rdev->rlc.cs_data = si_cs_data;
6265 r = sumo_rlc_init(rdev);
9b136d51
AD
6266 if (r) {
6267 DRM_ERROR("Failed to init rlc BOs!\n");
6268 return r;
6269 }
6270
6271 /* allocate wb buffer */
6272 r = radeon_wb_init(rdev);
6273 if (r)
6274 return r;
6275
6276 r = radeon_fence_driver_start_ring(rdev, RADEON_RING_TYPE_GFX_INDEX);
6277 if (r) {
6278 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
6279 return r;
6280 }
6281
6282 r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
6283 if (r) {
6284 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
6285 return r;
6286 }
6287
6288 r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
6289 if (r) {
6290 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
6291 return r;
6292 }
6293
8c5fd7ef
AD
6294 r = radeon_fence_driver_start_ring(rdev, R600_RING_TYPE_DMA_INDEX);
6295 if (r) {
6296 dev_err(rdev->dev, "failed initializing DMA fences (%d).\n", r);
6297 return r;
6298 }
6299
6300 r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_DMA1_INDEX);
6301 if (r) {
6302 dev_err(rdev->dev, "failed initializing DMA fences (%d).\n", r);
6303 return r;
6304 }
6305
1df0d523 6306 if (rdev->has_uvd) {
e409b128 6307 r = uvd_v2_2_resume(rdev);
1df0d523
AD
6308 if (!r) {
6309 r = radeon_fence_driver_start_ring(rdev,
6310 R600_RING_TYPE_UVD_INDEX);
6311 if (r)
6312 dev_err(rdev->dev, "UVD fences init error (%d).\n", r);
6313 }
f2ba57b5 6314 if (r)
1df0d523 6315 rdev->ring[R600_RING_TYPE_UVD_INDEX].ring_size = 0;
f2ba57b5 6316 }
f2ba57b5 6317
9b136d51 6318 /* Enable IRQ */
e49f3959
AH
6319 if (!rdev->irq.installed) {
6320 r = radeon_irq_kms_init(rdev);
6321 if (r)
6322 return r;
6323 }
6324
9b136d51
AD
6325 r = si_irq_init(rdev);
6326 if (r) {
6327 DRM_ERROR("radeon: IH init failed (%d).\n", r);
6328 radeon_irq_kms_fini(rdev);
6329 return r;
6330 }
6331 si_irq_set(rdev);
6332
6333 ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
6334 r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP_RPTR_OFFSET,
6335 CP_RB0_RPTR, CP_RB0_WPTR,
2e1e6dad 6336 RADEON_CP_PACKET2);
9b136d51
AD
6337 if (r)
6338 return r;
6339
6340 ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
6341 r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP1_RPTR_OFFSET,
6342 CP_RB1_RPTR, CP_RB1_WPTR,
2e1e6dad 6343 RADEON_CP_PACKET2);
9b136d51
AD
6344 if (r)
6345 return r;
6346
6347 ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
6348 r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP2_RPTR_OFFSET,
6349 CP_RB2_RPTR, CP_RB2_WPTR,
2e1e6dad 6350 RADEON_CP_PACKET2);
9b136d51
AD
6351 if (r)
6352 return r;
6353
8c5fd7ef
AD
6354 ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
6355 r = radeon_ring_init(rdev, ring, ring->ring_size, R600_WB_DMA_RPTR_OFFSET,
6356 DMA_RB_RPTR + DMA0_REGISTER_OFFSET,
6357 DMA_RB_WPTR + DMA0_REGISTER_OFFSET,
2e1e6dad 6358 DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0));
8c5fd7ef
AD
6359 if (r)
6360 return r;
6361
6362 ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
6363 r = radeon_ring_init(rdev, ring, ring->ring_size, CAYMAN_WB_DMA1_RPTR_OFFSET,
6364 DMA_RB_RPTR + DMA1_REGISTER_OFFSET,
6365 DMA_RB_WPTR + DMA1_REGISTER_OFFSET,
2e1e6dad 6366 DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0));
8c5fd7ef
AD
6367 if (r)
6368 return r;
6369
9b136d51
AD
6370 r = si_cp_load_microcode(rdev);
6371 if (r)
6372 return r;
6373 r = si_cp_resume(rdev);
6374 if (r)
6375 return r;
6376
8c5fd7ef
AD
6377 r = cayman_dma_resume(rdev);
6378 if (r)
6379 return r;
6380
1df0d523
AD
6381 if (rdev->has_uvd) {
6382 ring = &rdev->ring[R600_RING_TYPE_UVD_INDEX];
6383 if (ring->ring_size) {
02c9f7fa 6384 r = radeon_ring_init(rdev, ring, ring->ring_size, 0,
1df0d523 6385 UVD_RBC_RB_RPTR, UVD_RBC_RB_WPTR,
2e1e6dad 6386 RADEON_CP_PACKET2);
1df0d523 6387 if (!r)
e409b128 6388 r = uvd_v1_0_init(rdev);
1df0d523
AD
6389 if (r)
6390 DRM_ERROR("radeon: failed initializing UVD (%d).\n", r);
6391 }
f2ba57b5
CK
6392 }
6393
2898c348
CK
6394 r = radeon_ib_pool_init(rdev);
6395 if (r) {
6396 dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
9b136d51 6397 return r;
2898c348 6398 }
9b136d51 6399
c6105f24
CK
6400 r = radeon_vm_manager_init(rdev);
6401 if (r) {
6402 dev_err(rdev->dev, "vm manager initialization failed (%d).\n", r);
9b136d51 6403 return r;
c6105f24 6404 }
9b136d51 6405
b530602f
AD
6406 r = dce6_audio_init(rdev);
6407 if (r)
6408 return r;
6409
9b136d51
AD
6410 return 0;
6411}
6412
6413int si_resume(struct radeon_device *rdev)
6414{
6415 int r;
6416
6417 /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw,
6418 * posting will perform necessary task to bring back GPU into good
6419 * shape.
6420 */
6421 /* post card */
6422 atom_asic_init(rdev->mode_info.atom_context);
6423
205996c0
AD
6424 /* init golden registers */
6425 si_init_golden_registers(rdev);
6426
9b136d51
AD
6427 rdev->accel_working = true;
6428 r = si_startup(rdev);
6429 if (r) {
6430 DRM_ERROR("si startup failed on resume\n");
6431 rdev->accel_working = false;
6432 return r;
6433 }
6434
6435 return r;
6436
6437}
6438
6439int si_suspend(struct radeon_device *rdev)
6440{
b530602f 6441 dce6_audio_fini(rdev);
fa3daf9a 6442 radeon_vm_manager_fini(rdev);
9b136d51 6443 si_cp_enable(rdev, false);
8c5fd7ef 6444 cayman_dma_stop(rdev);
1df0d523 6445 if (rdev->has_uvd) {
e409b128 6446 uvd_v1_0_fini(rdev);
1df0d523
AD
6447 radeon_uvd_suspend(rdev);
6448 }
e16866ec
AD
6449 si_fini_pg(rdev);
6450 si_fini_cg(rdev);
9b136d51
AD
6451 si_irq_suspend(rdev);
6452 radeon_wb_disable(rdev);
6453 si_pcie_gart_disable(rdev);
6454 return 0;
6455}
6456
6457/* Plan is to move initialization in that function and use
6458 * helper function so that radeon_device_init pretty much
6459 * do nothing more than calling asic specific function. This
6460 * should also allow to remove a bunch of callback function
6461 * like vram_info.
6462 */
6463int si_init(struct radeon_device *rdev)
6464{
6465 struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
6466 int r;
6467
9b136d51
AD
6468 /* Read BIOS */
6469 if (!radeon_get_bios(rdev)) {
6470 if (ASIC_IS_AVIVO(rdev))
6471 return -EINVAL;
6472 }
6473 /* Must be an ATOMBIOS */
6474 if (!rdev->is_atom_bios) {
6475 dev_err(rdev->dev, "Expecting atombios for cayman GPU\n");
6476 return -EINVAL;
6477 }
6478 r = radeon_atombios_init(rdev);
6479 if (r)
6480 return r;
6481
6482 /* Post card if necessary */
6483 if (!radeon_card_posted(rdev)) {
6484 if (!rdev->bios) {
6485 dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n");
6486 return -EINVAL;
6487 }
6488 DRM_INFO("GPU not posted. posting now...\n");
6489 atom_asic_init(rdev->mode_info.atom_context);
6490 }
205996c0
AD
6491 /* init golden registers */
6492 si_init_golden_registers(rdev);
9b136d51
AD
6493 /* Initialize scratch registers */
6494 si_scratch_init(rdev);
6495 /* Initialize surface registers */
6496 radeon_surface_init(rdev);
6497 /* Initialize clocks */
6498 radeon_get_clock_info(rdev->ddev);
6499
6500 /* Fence driver */
6501 r = radeon_fence_driver_init(rdev);
6502 if (r)
6503 return r;
6504
6505 /* initialize memory controller */
6506 r = si_mc_init(rdev);
6507 if (r)
6508 return r;
6509 /* Memory manager */
6510 r = radeon_bo_init(rdev);
6511 if (r)
6512 return r;
6513
9b136d51
AD
6514 ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
6515 ring->ring_obj = NULL;
6516 r600_ring_init(rdev, ring, 1024 * 1024);
6517
6518 ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
6519 ring->ring_obj = NULL;
6520 r600_ring_init(rdev, ring, 1024 * 1024);
6521
6522 ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
6523 ring->ring_obj = NULL;
6524 r600_ring_init(rdev, ring, 1024 * 1024);
6525
8c5fd7ef
AD
6526 ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
6527 ring->ring_obj = NULL;
6528 r600_ring_init(rdev, ring, 64 * 1024);
6529
6530 ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
6531 ring->ring_obj = NULL;
6532 r600_ring_init(rdev, ring, 64 * 1024);
6533
1df0d523
AD
6534 if (rdev->has_uvd) {
6535 r = radeon_uvd_init(rdev);
6536 if (!r) {
6537 ring = &rdev->ring[R600_RING_TYPE_UVD_INDEX];
6538 ring->ring_obj = NULL;
6539 r600_ring_init(rdev, ring, 4096);
6540 }
f2ba57b5
CK
6541 }
6542
9b136d51
AD
6543 rdev->ih.ring_obj = NULL;
6544 r600_ih_ring_init(rdev, 64 * 1024);
6545
6546 r = r600_pcie_gart_init(rdev);
6547 if (r)
6548 return r;
6549
9b136d51 6550 rdev->accel_working = true;
9b136d51
AD
6551 r = si_startup(rdev);
6552 if (r) {
6553 dev_err(rdev->dev, "disabling GPU acceleration\n");
6554 si_cp_fini(rdev);
8c5fd7ef 6555 cayman_dma_fini(rdev);
9b136d51 6556 si_irq_fini(rdev);
1fd11777 6557 sumo_rlc_fini(rdev);
9b136d51 6558 radeon_wb_fini(rdev);
2898c348 6559 radeon_ib_pool_fini(rdev);
9b136d51
AD
6560 radeon_vm_manager_fini(rdev);
6561 radeon_irq_kms_fini(rdev);
6562 si_pcie_gart_fini(rdev);
6563 rdev->accel_working = false;
6564 }
6565
6566 /* Don't start up if the MC ucode is missing.
6567 * The default clocks and voltages before the MC ucode
6568 * is loaded are not suffient for advanced operations.
6569 */
6570 if (!rdev->mc_fw) {
6571 DRM_ERROR("radeon: MC ucode required for NI+.\n");
6572 return -EINVAL;
6573 }
6574
6575 return 0;
6576}
6577
6578void si_fini(struct radeon_device *rdev)
6579{
9b136d51 6580 si_cp_fini(rdev);
8c5fd7ef 6581 cayman_dma_fini(rdev);
e16866ec
AD
6582 si_fini_pg(rdev);
6583 si_fini_cg(rdev);
9b136d51 6584 si_irq_fini(rdev);
1fd11777 6585 sumo_rlc_fini(rdev);
9b136d51
AD
6586 radeon_wb_fini(rdev);
6587 radeon_vm_manager_fini(rdev);
2898c348 6588 radeon_ib_pool_fini(rdev);
9b136d51 6589 radeon_irq_kms_fini(rdev);
2858c00d 6590 if (rdev->has_uvd) {
e409b128 6591 uvd_v1_0_fini(rdev);
1df0d523 6592 radeon_uvd_fini(rdev);
2858c00d 6593 }
9b136d51
AD
6594 si_pcie_gart_fini(rdev);
6595 r600_vram_scratch_fini(rdev);
6596 radeon_gem_fini(rdev);
9b136d51
AD
6597 radeon_fence_driver_fini(rdev);
6598 radeon_bo_fini(rdev);
6599 radeon_atombios_fini(rdev);
6600 kfree(rdev->bios);
6601 rdev->bios = NULL;
6602}
6603
6759a0a7 6604/**
d0418894 6605 * si_get_gpu_clock_counter - return GPU clock counter snapshot
6759a0a7
MO
6606 *
6607 * @rdev: radeon_device pointer
6608 *
6609 * Fetches a GPU clock counter snapshot (SI).
6610 * Returns the 64 bit clock counter snapshot.
6611 */
d0418894 6612uint64_t si_get_gpu_clock_counter(struct radeon_device *rdev)
6759a0a7
MO
6613{
6614 uint64_t clock;
6615
6616 mutex_lock(&rdev->gpu_clock_mutex);
6617 WREG32(RLC_CAPTURE_GPU_CLOCK_COUNT, 1);
6618 clock = (uint64_t)RREG32(RLC_GPU_CLOCK_COUNT_LSB) |
6619 ((uint64_t)RREG32(RLC_GPU_CLOCK_COUNT_MSB) << 32ULL);
6620 mutex_unlock(&rdev->gpu_clock_mutex);
6621 return clock;
6622}
2539eb02 6623
2539eb02
CK
6624int si_set_uvd_clocks(struct radeon_device *rdev, u32 vclk, u32 dclk)
6625{
facd112d 6626 unsigned fb_div = 0, vclk_div = 0, dclk_div = 0;
2539eb02
CK
6627 int r;
6628
4ed10835
CK
6629 /* bypass vclk and dclk with bclk */
6630 WREG32_P(CG_UPLL_FUNC_CNTL_2,
6631 VCLK_SRC_SEL(1) | DCLK_SRC_SEL(1),
6632 ~(VCLK_SRC_SEL_MASK | DCLK_SRC_SEL_MASK));
6633
6634 /* put PLL in bypass mode */
6635 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_BYPASS_EN_MASK, ~UPLL_BYPASS_EN_MASK);
6636
6637 if (!vclk || !dclk) {
6638 /* keep the Bypass mode, put PLL to sleep */
6639 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_SLEEP_MASK, ~UPLL_SLEEP_MASK);
6640 return 0;
6641 }
6642
facd112d
CK
6643 r = radeon_uvd_calc_upll_dividers(rdev, vclk, dclk, 125000, 250000,
6644 16384, 0x03FFFFFF, 0, 128, 5,
6645 &fb_div, &vclk_div, &dclk_div);
6646 if (r)
6647 return r;
2539eb02
CK
6648
6649 /* set RESET_ANTI_MUX to 0 */
6650 WREG32_P(CG_UPLL_FUNC_CNTL_5, 0, ~RESET_ANTI_MUX_MASK);
6651
6652 /* set VCO_MODE to 1 */
6653 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_VCO_MODE_MASK, ~UPLL_VCO_MODE_MASK);
6654
6655 /* toggle UPLL_SLEEP to 1 then back to 0 */
6656 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_SLEEP_MASK, ~UPLL_SLEEP_MASK);
6657 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_SLEEP_MASK);
6658
6659 /* deassert UPLL_RESET */
6660 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_RESET_MASK);
6661
6662 mdelay(1);
6663
facd112d 6664 r = radeon_uvd_send_upll_ctlreq(rdev, CG_UPLL_FUNC_CNTL);
2539eb02
CK
6665 if (r)
6666 return r;
6667
6668 /* assert UPLL_RESET again */
6669 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_RESET_MASK, ~UPLL_RESET_MASK);
6670
6671 /* disable spread spectrum. */
6672 WREG32_P(CG_UPLL_SPREAD_SPECTRUM, 0, ~SSEN_MASK);
6673
6674 /* set feedback divider */
facd112d 6675 WREG32_P(CG_UPLL_FUNC_CNTL_3, UPLL_FB_DIV(fb_div), ~UPLL_FB_DIV_MASK);
2539eb02
CK
6676
6677 /* set ref divider to 0 */
6678 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_REF_DIV_MASK);
6679
facd112d 6680 if (fb_div < 307200)
2539eb02
CK
6681 WREG32_P(CG_UPLL_FUNC_CNTL_4, 0, ~UPLL_SPARE_ISPARE9);
6682 else
6683 WREG32_P(CG_UPLL_FUNC_CNTL_4, UPLL_SPARE_ISPARE9, ~UPLL_SPARE_ISPARE9);
6684
6685 /* set PDIV_A and PDIV_B */
6686 WREG32_P(CG_UPLL_FUNC_CNTL_2,
facd112d 6687 UPLL_PDIV_A(vclk_div) | UPLL_PDIV_B(dclk_div),
2539eb02
CK
6688 ~(UPLL_PDIV_A_MASK | UPLL_PDIV_B_MASK));
6689
6690 /* give the PLL some time to settle */
6691 mdelay(15);
6692
6693 /* deassert PLL_RESET */
6694 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_RESET_MASK);
6695
6696 mdelay(15);
6697
6698 /* switch from bypass mode to normal mode */
6699 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_BYPASS_EN_MASK);
6700
facd112d 6701 r = radeon_uvd_send_upll_ctlreq(rdev, CG_UPLL_FUNC_CNTL);
2539eb02
CK
6702 if (r)
6703 return r;
6704
6705 /* switch VCLK and DCLK selection */
6706 WREG32_P(CG_UPLL_FUNC_CNTL_2,
6707 VCLK_SRC_SEL(2) | DCLK_SRC_SEL(2),
6708 ~(VCLK_SRC_SEL_MASK | DCLK_SRC_SEL_MASK));
6709
6710 mdelay(100);
6711
6712 return 0;
6713}
b9d305df
AD
6714
6715static void si_pcie_gen3_enable(struct radeon_device *rdev)
6716{
6717 struct pci_dev *root = rdev->pdev->bus->self;
6718 int bridge_pos, gpu_pos;
6719 u32 speed_cntl, mask, current_data_rate;
6720 int ret, i;
6721 u16 tmp16;
6722
6723 if (radeon_pcie_gen2 == 0)
6724 return;
6725
6726 if (rdev->flags & RADEON_IS_IGP)
6727 return;
6728
6729 if (!(rdev->flags & RADEON_IS_PCIE))
6730 return;
6731
6732 ret = drm_pcie_get_speed_cap_mask(rdev->ddev, &mask);
6733 if (ret != 0)
6734 return;
6735
6736 if (!(mask & (DRM_PCIE_SPEED_50 | DRM_PCIE_SPEED_80)))
6737 return;
6738
6739 speed_cntl = RREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL);
6740 current_data_rate = (speed_cntl & LC_CURRENT_DATA_RATE_MASK) >>
6741 LC_CURRENT_DATA_RATE_SHIFT;
6742 if (mask & DRM_PCIE_SPEED_80) {
6743 if (current_data_rate == 2) {
6744 DRM_INFO("PCIE gen 3 link speeds already enabled\n");
6745 return;
6746 }
6747 DRM_INFO("enabling PCIE gen 3 link speeds, disable with radeon.pcie_gen2=0\n");
6748 } else if (mask & DRM_PCIE_SPEED_50) {
6749 if (current_data_rate == 1) {
6750 DRM_INFO("PCIE gen 2 link speeds already enabled\n");
6751 return;
6752 }
6753 DRM_INFO("enabling PCIE gen 2 link speeds, disable with radeon.pcie_gen2=0\n");
6754 }
6755
6756 bridge_pos = pci_pcie_cap(root);
6757 if (!bridge_pos)
6758 return;
6759
6760 gpu_pos = pci_pcie_cap(rdev->pdev);
6761 if (!gpu_pos)
6762 return;
6763
6764 if (mask & DRM_PCIE_SPEED_80) {
6765 /* re-try equalization if gen3 is not already enabled */
6766 if (current_data_rate != 2) {
6767 u16 bridge_cfg, gpu_cfg;
6768 u16 bridge_cfg2, gpu_cfg2;
6769 u32 max_lw, current_lw, tmp;
6770
6771 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL, &bridge_cfg);
6772 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, &gpu_cfg);
6773
6774 tmp16 = bridge_cfg | PCI_EXP_LNKCTL_HAWD;
6775 pci_write_config_word(root, bridge_pos + PCI_EXP_LNKCTL, tmp16);
6776
6777 tmp16 = gpu_cfg | PCI_EXP_LNKCTL_HAWD;
6778 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, tmp16);
6779
6780 tmp = RREG32_PCIE(PCIE_LC_STATUS1);
6781 max_lw = (tmp & LC_DETECTED_LINK_WIDTH_MASK) >> LC_DETECTED_LINK_WIDTH_SHIFT;
6782 current_lw = (tmp & LC_OPERATING_LINK_WIDTH_MASK) >> LC_OPERATING_LINK_WIDTH_SHIFT;
6783
6784 if (current_lw < max_lw) {
6785 tmp = RREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL);
6786 if (tmp & LC_RENEGOTIATION_SUPPORT) {
6787 tmp &= ~(LC_LINK_WIDTH_MASK | LC_UPCONFIGURE_DIS);
6788 tmp |= (max_lw << LC_LINK_WIDTH_SHIFT);
6789 tmp |= LC_UPCONFIGURE_SUPPORT | LC_RENEGOTIATE_EN | LC_RECONFIG_NOW;
6790 WREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL, tmp);
6791 }
6792 }
6793
6794 for (i = 0; i < 10; i++) {
6795 /* check status */
6796 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_DEVSTA, &tmp16);
6797 if (tmp16 & PCI_EXP_DEVSTA_TRPND)
6798 break;
6799
6800 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL, &bridge_cfg);
6801 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, &gpu_cfg);
6802
6803 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL2, &bridge_cfg2);
6804 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, &gpu_cfg2);
6805
6806 tmp = RREG32_PCIE_PORT(PCIE_LC_CNTL4);
6807 tmp |= LC_SET_QUIESCE;
6808 WREG32_PCIE_PORT(PCIE_LC_CNTL4, tmp);
6809
6810 tmp = RREG32_PCIE_PORT(PCIE_LC_CNTL4);
6811 tmp |= LC_REDO_EQ;
6812 WREG32_PCIE_PORT(PCIE_LC_CNTL4, tmp);
6813
6814 mdelay(100);
6815
6816 /* linkctl */
6817 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL, &tmp16);
6818 tmp16 &= ~PCI_EXP_LNKCTL_HAWD;
6819 tmp16 |= (bridge_cfg & PCI_EXP_LNKCTL_HAWD);
6820 pci_write_config_word(root, bridge_pos + PCI_EXP_LNKCTL, tmp16);
6821
6822 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, &tmp16);
6823 tmp16 &= ~PCI_EXP_LNKCTL_HAWD;
6824 tmp16 |= (gpu_cfg & PCI_EXP_LNKCTL_HAWD);
6825 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, tmp16);
6826
6827 /* linkctl2 */
6828 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL2, &tmp16);
6829 tmp16 &= ~((1 << 4) | (7 << 9));
6830 tmp16 |= (bridge_cfg2 & ((1 << 4) | (7 << 9)));
6831 pci_write_config_word(root, bridge_pos + PCI_EXP_LNKCTL2, tmp16);
6832
6833 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, &tmp16);
6834 tmp16 &= ~((1 << 4) | (7 << 9));
6835 tmp16 |= (gpu_cfg2 & ((1 << 4) | (7 << 9)));
6836 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, tmp16);
6837
6838 tmp = RREG32_PCIE_PORT(PCIE_LC_CNTL4);
6839 tmp &= ~LC_SET_QUIESCE;
6840 WREG32_PCIE_PORT(PCIE_LC_CNTL4, tmp);
6841 }
6842 }
6843 }
6844
6845 /* set the link speed */
6846 speed_cntl |= LC_FORCE_EN_SW_SPEED_CHANGE | LC_FORCE_DIS_HW_SPEED_CHANGE;
6847 speed_cntl &= ~LC_FORCE_DIS_SW_SPEED_CHANGE;
6848 WREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL, speed_cntl);
6849
6850 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, &tmp16);
6851 tmp16 &= ~0xf;
6852 if (mask & DRM_PCIE_SPEED_80)
6853 tmp16 |= 3; /* gen3 */
6854 else if (mask & DRM_PCIE_SPEED_50)
6855 tmp16 |= 2; /* gen2 */
6856 else
6857 tmp16 |= 1; /* gen1 */
6858 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, tmp16);
6859
6860 speed_cntl = RREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL);
6861 speed_cntl |= LC_INITIATE_LINK_SPEED_CHANGE;
6862 WREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL, speed_cntl);
6863
6864 for (i = 0; i < rdev->usec_timeout; i++) {
6865 speed_cntl = RREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL);
6866 if ((speed_cntl & LC_INITIATE_LINK_SPEED_CHANGE) == 0)
6867 break;
6868 udelay(1);
6869 }
6870}
6871
e0bcf165
AD
6872static void si_program_aspm(struct radeon_device *rdev)
6873{
6874 u32 data, orig;
6875 bool disable_l0s = false, disable_l1 = false, disable_plloff_in_l1 = false;
6876 bool disable_clkreq = false;
6877
1294d4a3
AD
6878 if (radeon_aspm == 0)
6879 return;
6880
e0bcf165
AD
6881 if (!(rdev->flags & RADEON_IS_PCIE))
6882 return;
6883
6884 orig = data = RREG32_PCIE_PORT(PCIE_LC_N_FTS_CNTL);
6885 data &= ~LC_XMIT_N_FTS_MASK;
6886 data |= LC_XMIT_N_FTS(0x24) | LC_XMIT_N_FTS_OVERRIDE_EN;
6887 if (orig != data)
6888 WREG32_PCIE_PORT(PCIE_LC_N_FTS_CNTL, data);
6889
6890 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL3);
6891 data |= LC_GO_TO_RECOVERY;
6892 if (orig != data)
6893 WREG32_PCIE_PORT(PCIE_LC_CNTL3, data);
6894
6895 orig = data = RREG32_PCIE(PCIE_P_CNTL);
6896 data |= P_IGNORE_EDB_ERR;
6897 if (orig != data)
6898 WREG32_PCIE(PCIE_P_CNTL, data);
6899
6900 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL);
6901 data &= ~(LC_L0S_INACTIVITY_MASK | LC_L1_INACTIVITY_MASK);
6902 data |= LC_PMI_TO_L1_DIS;
6903 if (!disable_l0s)
6904 data |= LC_L0S_INACTIVITY(7);
6905
6906 if (!disable_l1) {
6907 data |= LC_L1_INACTIVITY(7);
6908 data &= ~LC_PMI_TO_L1_DIS;
6909 if (orig != data)
6910 WREG32_PCIE_PORT(PCIE_LC_CNTL, data);
6911
6912 if (!disable_plloff_in_l1) {
6913 bool clk_req_support;
6914
6915 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_0);
6916 data &= ~(PLL_POWER_STATE_IN_OFF_0_MASK | PLL_POWER_STATE_IN_TXS2_0_MASK);
6917 data |= PLL_POWER_STATE_IN_OFF_0(7) | PLL_POWER_STATE_IN_TXS2_0(7);
6918 if (orig != data)
6919 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_0, data);
6920
6921 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_1);
6922 data &= ~(PLL_POWER_STATE_IN_OFF_1_MASK | PLL_POWER_STATE_IN_TXS2_1_MASK);
6923 data |= PLL_POWER_STATE_IN_OFF_1(7) | PLL_POWER_STATE_IN_TXS2_1(7);
6924 if (orig != data)
6925 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_1, data);
6926
6927 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_0);
6928 data &= ~(PLL_POWER_STATE_IN_OFF_0_MASK | PLL_POWER_STATE_IN_TXS2_0_MASK);
6929 data |= PLL_POWER_STATE_IN_OFF_0(7) | PLL_POWER_STATE_IN_TXS2_0(7);
6930 if (orig != data)
6931 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_0, data);
6932
6933 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_1);
6934 data &= ~(PLL_POWER_STATE_IN_OFF_1_MASK | PLL_POWER_STATE_IN_TXS2_1_MASK);
6935 data |= PLL_POWER_STATE_IN_OFF_1(7) | PLL_POWER_STATE_IN_TXS2_1(7);
6936 if (orig != data)
6937 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_1, data);
6938
6939 if ((rdev->family != CHIP_OLAND) && (rdev->family != CHIP_HAINAN)) {
6940 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_0);
6941 data &= ~PLL_RAMP_UP_TIME_0_MASK;
6942 if (orig != data)
6943 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_0, data);
6944
6945 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_1);
6946 data &= ~PLL_RAMP_UP_TIME_1_MASK;
6947 if (orig != data)
6948 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_1, data);
6949
6950 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_2);
6951 data &= ~PLL_RAMP_UP_TIME_2_MASK;
6952 if (orig != data)
6953 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_2, data);
6954
6955 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_3);
6956 data &= ~PLL_RAMP_UP_TIME_3_MASK;
6957 if (orig != data)
6958 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_3, data);
6959
6960 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_0);
6961 data &= ~PLL_RAMP_UP_TIME_0_MASK;
6962 if (orig != data)
6963 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_0, data);
6964
6965 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_1);
6966 data &= ~PLL_RAMP_UP_TIME_1_MASK;
6967 if (orig != data)
6968 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_1, data);
6969
6970 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_2);
6971 data &= ~PLL_RAMP_UP_TIME_2_MASK;
6972 if (orig != data)
6973 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_2, data);
6974
6975 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_3);
6976 data &= ~PLL_RAMP_UP_TIME_3_MASK;
6977 if (orig != data)
6978 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_3, data);
6979 }
6980 orig = data = RREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL);
6981 data &= ~LC_DYN_LANES_PWR_STATE_MASK;
6982 data |= LC_DYN_LANES_PWR_STATE(3);
6983 if (orig != data)
6984 WREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL, data);
6985
6986 orig = data = RREG32_PIF_PHY0(PB0_PIF_CNTL);
6987 data &= ~LS2_EXIT_TIME_MASK;
6988 if ((rdev->family == CHIP_OLAND) || (rdev->family == CHIP_HAINAN))
6989 data |= LS2_EXIT_TIME(5);
6990 if (orig != data)
6991 WREG32_PIF_PHY0(PB0_PIF_CNTL, data);
6992
6993 orig = data = RREG32_PIF_PHY1(PB1_PIF_CNTL);
6994 data &= ~LS2_EXIT_TIME_MASK;
6995 if ((rdev->family == CHIP_OLAND) || (rdev->family == CHIP_HAINAN))
6996 data |= LS2_EXIT_TIME(5);
6997 if (orig != data)
6998 WREG32_PIF_PHY1(PB1_PIF_CNTL, data);
6999
7000 if (!disable_clkreq) {
7001 struct pci_dev *root = rdev->pdev->bus->self;
7002 u32 lnkcap;
7003
7004 clk_req_support = false;
7005 pcie_capability_read_dword(root, PCI_EXP_LNKCAP, &lnkcap);
7006 if (lnkcap & PCI_EXP_LNKCAP_CLKPM)
7007 clk_req_support = true;
7008 } else {
7009 clk_req_support = false;
7010 }
7011
7012 if (clk_req_support) {
7013 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL2);
7014 data |= LC_ALLOW_PDWN_IN_L1 | LC_ALLOW_PDWN_IN_L23;
7015 if (orig != data)
7016 WREG32_PCIE_PORT(PCIE_LC_CNTL2, data);
7017
7018 orig = data = RREG32(THM_CLK_CNTL);
7019 data &= ~(CMON_CLK_SEL_MASK | TMON_CLK_SEL_MASK);
7020 data |= CMON_CLK_SEL(1) | TMON_CLK_SEL(1);
7021 if (orig != data)
7022 WREG32(THM_CLK_CNTL, data);
7023
7024 orig = data = RREG32(MISC_CLK_CNTL);
7025 data &= ~(DEEP_SLEEP_CLK_SEL_MASK | ZCLK_SEL_MASK);
7026 data |= DEEP_SLEEP_CLK_SEL(1) | ZCLK_SEL(1);
7027 if (orig != data)
7028 WREG32(MISC_CLK_CNTL, data);
7029
7030 orig = data = RREG32(CG_CLKPIN_CNTL);
7031 data &= ~BCLK_AS_XCLK;
7032 if (orig != data)
7033 WREG32(CG_CLKPIN_CNTL, data);
7034
7035 orig = data = RREG32(CG_CLKPIN_CNTL_2);
7036 data &= ~FORCE_BIF_REFCLK_EN;
7037 if (orig != data)
7038 WREG32(CG_CLKPIN_CNTL_2, data);
7039
7040 orig = data = RREG32(MPLL_BYPASSCLK_SEL);
7041 data &= ~MPLL_CLKOUT_SEL_MASK;
7042 data |= MPLL_CLKOUT_SEL(4);
7043 if (orig != data)
7044 WREG32(MPLL_BYPASSCLK_SEL, data);
7045
7046 orig = data = RREG32(SPLL_CNTL_MODE);
7047 data &= ~SPLL_REFCLK_SEL_MASK;
7048 if (orig != data)
7049 WREG32(SPLL_CNTL_MODE, data);
7050 }
7051 }
7052 } else {
7053 if (orig != data)
7054 WREG32_PCIE_PORT(PCIE_LC_CNTL, data);
7055 }
7056
7057 orig = data = RREG32_PCIE(PCIE_CNTL2);
7058 data |= SLV_MEM_LS_EN | MST_MEM_LS_EN | REPLAY_MEM_LS_EN;
7059 if (orig != data)
7060 WREG32_PCIE(PCIE_CNTL2, data);
7061
7062 if (!disable_l0s) {
7063 data = RREG32_PCIE_PORT(PCIE_LC_N_FTS_CNTL);
7064 if((data & LC_N_FTS_MASK) == LC_N_FTS_MASK) {
7065 data = RREG32_PCIE(PCIE_LC_STATUS1);
7066 if ((data & LC_REVERSE_XMIT) && (data & LC_REVERSE_RCVR)) {
7067 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL);
7068 data &= ~LC_L0S_INACTIVITY_MASK;
7069 if (orig != data)
7070 WREG32_PCIE_PORT(PCIE_LC_CNTL, data);
7071 }
7072 }
7073 }
7074}