drm/radeon: fix VGT_GS_INSTANCE_CNT register
[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);
811e4d58
AD
81static void si_enable_gui_idle_interrupt(struct radeon_device *rdev,
82 bool enable);
a6f4ae8d
AD
83static void si_fini_pg(struct radeon_device *rdev);
84static void si_fini_cg(struct radeon_device *rdev);
85static void si_rlc_stop(struct radeon_device *rdev);
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;
d8367112 1679 err = 0;
8a53fa23 1680 } else if (rdev->smc_fw->size != smc_req_size) {
a9e61410
AD
1681 printk(KERN_ERR
1682 "si_smc: Bogus length %zu in firmware \"%s\"\n",
1683 rdev->smc_fw->size, fw_name);
1684 err = -EINVAL;
1685 }
1686
0f0de06c 1687out:
0f0de06c
AD
1688 if (err) {
1689 if (err != -EINVAL)
1690 printk(KERN_ERR
1691 "si_cp: Failed to load firmware \"%s\"\n",
1692 fw_name);
1693 release_firmware(rdev->pfp_fw);
1694 rdev->pfp_fw = NULL;
1695 release_firmware(rdev->me_fw);
1696 rdev->me_fw = NULL;
1697 release_firmware(rdev->ce_fw);
1698 rdev->ce_fw = NULL;
1699 release_firmware(rdev->rlc_fw);
1700 rdev->rlc_fw = NULL;
1701 release_firmware(rdev->mc_fw);
1702 rdev->mc_fw = NULL;
a9e61410
AD
1703 release_firmware(rdev->smc_fw);
1704 rdev->smc_fw = NULL;
0f0de06c
AD
1705 }
1706 return err;
1707}
1708
43b3cd99
AD
1709/* watermark setup */
1710static u32 dce6_line_buffer_adjust(struct radeon_device *rdev,
1711 struct radeon_crtc *radeon_crtc,
1712 struct drm_display_mode *mode,
1713 struct drm_display_mode *other_mode)
1714{
290d2457
AD
1715 u32 tmp, buffer_alloc, i;
1716 u32 pipe_offset = radeon_crtc->crtc_id * 0x20;
43b3cd99
AD
1717 /*
1718 * Line Buffer Setup
1719 * There are 3 line buffers, each one shared by 2 display controllers.
1720 * DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
1721 * the display controllers. The paritioning is done via one of four
1722 * preset allocations specified in bits 21:20:
1723 * 0 - half lb
1724 * 2 - whole lb, other crtc must be disabled
1725 */
1726 /* this can get tricky if we have two large displays on a paired group
1727 * of crtcs. Ideally for multiple large displays we'd assign them to
1728 * non-linked crtcs for maximum line buffer allocation.
1729 */
1730 if (radeon_crtc->base.enabled && mode) {
290d2457 1731 if (other_mode) {
43b3cd99 1732 tmp = 0; /* 1/2 */
290d2457
AD
1733 buffer_alloc = 1;
1734 } else {
43b3cd99 1735 tmp = 2; /* whole */
290d2457
AD
1736 buffer_alloc = 2;
1737 }
1738 } else {
43b3cd99 1739 tmp = 0;
290d2457
AD
1740 buffer_alloc = 0;
1741 }
43b3cd99
AD
1742
1743 WREG32(DC_LB_MEMORY_SPLIT + radeon_crtc->crtc_offset,
1744 DC_LB_MEMORY_CONFIG(tmp));
1745
290d2457
AD
1746 WREG32(PIPE0_DMIF_BUFFER_CONTROL + pipe_offset,
1747 DMIF_BUFFERS_ALLOCATED(buffer_alloc));
1748 for (i = 0; i < rdev->usec_timeout; i++) {
1749 if (RREG32(PIPE0_DMIF_BUFFER_CONTROL + pipe_offset) &
1750 DMIF_BUFFERS_ALLOCATED_COMPLETED)
1751 break;
1752 udelay(1);
1753 }
1754
43b3cd99
AD
1755 if (radeon_crtc->base.enabled && mode) {
1756 switch (tmp) {
1757 case 0:
1758 default:
1759 return 4096 * 2;
1760 case 2:
1761 return 8192 * 2;
1762 }
1763 }
1764
1765 /* controller not enabled, so no lb used */
1766 return 0;
1767}
1768
ca7db22b 1769static u32 si_get_number_of_dram_channels(struct radeon_device *rdev)
43b3cd99
AD
1770{
1771 u32 tmp = RREG32(MC_SHARED_CHMAP);
1772
1773 switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
1774 case 0:
1775 default:
1776 return 1;
1777 case 1:
1778 return 2;
1779 case 2:
1780 return 4;
1781 case 3:
1782 return 8;
1783 case 4:
1784 return 3;
1785 case 5:
1786 return 6;
1787 case 6:
1788 return 10;
1789 case 7:
1790 return 12;
1791 case 8:
1792 return 16;
1793 }
1794}
1795
1796struct dce6_wm_params {
1797 u32 dram_channels; /* number of dram channels */
1798 u32 yclk; /* bandwidth per dram data pin in kHz */
1799 u32 sclk; /* engine clock in kHz */
1800 u32 disp_clk; /* display clock in kHz */
1801 u32 src_width; /* viewport width */
1802 u32 active_time; /* active display time in ns */
1803 u32 blank_time; /* blank time in ns */
1804 bool interlaced; /* mode is interlaced */
1805 fixed20_12 vsc; /* vertical scale ratio */
1806 u32 num_heads; /* number of active crtcs */
1807 u32 bytes_per_pixel; /* bytes per pixel display + overlay */
1808 u32 lb_size; /* line buffer allocated to pipe */
1809 u32 vtaps; /* vertical scaler taps */
1810};
1811
1812static u32 dce6_dram_bandwidth(struct dce6_wm_params *wm)
1813{
1814 /* Calculate raw DRAM Bandwidth */
1815 fixed20_12 dram_efficiency; /* 0.7 */
1816 fixed20_12 yclk, dram_channels, bandwidth;
1817 fixed20_12 a;
1818
1819 a.full = dfixed_const(1000);
1820 yclk.full = dfixed_const(wm->yclk);
1821 yclk.full = dfixed_div(yclk, a);
1822 dram_channels.full = dfixed_const(wm->dram_channels * 4);
1823 a.full = dfixed_const(10);
1824 dram_efficiency.full = dfixed_const(7);
1825 dram_efficiency.full = dfixed_div(dram_efficiency, a);
1826 bandwidth.full = dfixed_mul(dram_channels, yclk);
1827 bandwidth.full = dfixed_mul(bandwidth, dram_efficiency);
1828
1829 return dfixed_trunc(bandwidth);
1830}
1831
1832static u32 dce6_dram_bandwidth_for_display(struct dce6_wm_params *wm)
1833{
1834 /* Calculate DRAM Bandwidth and the part allocated to display. */
1835 fixed20_12 disp_dram_allocation; /* 0.3 to 0.7 */
1836 fixed20_12 yclk, dram_channels, bandwidth;
1837 fixed20_12 a;
1838
1839 a.full = dfixed_const(1000);
1840 yclk.full = dfixed_const(wm->yclk);
1841 yclk.full = dfixed_div(yclk, a);
1842 dram_channels.full = dfixed_const(wm->dram_channels * 4);
1843 a.full = dfixed_const(10);
1844 disp_dram_allocation.full = dfixed_const(3); /* XXX worse case value 0.3 */
1845 disp_dram_allocation.full = dfixed_div(disp_dram_allocation, a);
1846 bandwidth.full = dfixed_mul(dram_channels, yclk);
1847 bandwidth.full = dfixed_mul(bandwidth, disp_dram_allocation);
1848
1849 return dfixed_trunc(bandwidth);
1850}
1851
1852static u32 dce6_data_return_bandwidth(struct dce6_wm_params *wm)
1853{
1854 /* Calculate the display Data return Bandwidth */
1855 fixed20_12 return_efficiency; /* 0.8 */
1856 fixed20_12 sclk, bandwidth;
1857 fixed20_12 a;
1858
1859 a.full = dfixed_const(1000);
1860 sclk.full = dfixed_const(wm->sclk);
1861 sclk.full = dfixed_div(sclk, a);
1862 a.full = dfixed_const(10);
1863 return_efficiency.full = dfixed_const(8);
1864 return_efficiency.full = dfixed_div(return_efficiency, a);
1865 a.full = dfixed_const(32);
1866 bandwidth.full = dfixed_mul(a, sclk);
1867 bandwidth.full = dfixed_mul(bandwidth, return_efficiency);
1868
1869 return dfixed_trunc(bandwidth);
1870}
1871
1872static u32 dce6_get_dmif_bytes_per_request(struct dce6_wm_params *wm)
1873{
1874 return 32;
1875}
1876
1877static u32 dce6_dmif_request_bandwidth(struct dce6_wm_params *wm)
1878{
1879 /* Calculate the DMIF Request Bandwidth */
1880 fixed20_12 disp_clk_request_efficiency; /* 0.8 */
1881 fixed20_12 disp_clk, sclk, bandwidth;
1882 fixed20_12 a, b1, b2;
1883 u32 min_bandwidth;
1884
1885 a.full = dfixed_const(1000);
1886 disp_clk.full = dfixed_const(wm->disp_clk);
1887 disp_clk.full = dfixed_div(disp_clk, a);
1888 a.full = dfixed_const(dce6_get_dmif_bytes_per_request(wm) / 2);
1889 b1.full = dfixed_mul(a, disp_clk);
1890
1891 a.full = dfixed_const(1000);
1892 sclk.full = dfixed_const(wm->sclk);
1893 sclk.full = dfixed_div(sclk, a);
1894 a.full = dfixed_const(dce6_get_dmif_bytes_per_request(wm));
1895 b2.full = dfixed_mul(a, sclk);
1896
1897 a.full = dfixed_const(10);
1898 disp_clk_request_efficiency.full = dfixed_const(8);
1899 disp_clk_request_efficiency.full = dfixed_div(disp_clk_request_efficiency, a);
1900
1901 min_bandwidth = min(dfixed_trunc(b1), dfixed_trunc(b2));
1902
1903 a.full = dfixed_const(min_bandwidth);
1904 bandwidth.full = dfixed_mul(a, disp_clk_request_efficiency);
1905
1906 return dfixed_trunc(bandwidth);
1907}
1908
1909static u32 dce6_available_bandwidth(struct dce6_wm_params *wm)
1910{
1911 /* Calculate the Available bandwidth. Display can use this temporarily but not in average. */
1912 u32 dram_bandwidth = dce6_dram_bandwidth(wm);
1913 u32 data_return_bandwidth = dce6_data_return_bandwidth(wm);
1914 u32 dmif_req_bandwidth = dce6_dmif_request_bandwidth(wm);
1915
1916 return min(dram_bandwidth, min(data_return_bandwidth, dmif_req_bandwidth));
1917}
1918
1919static u32 dce6_average_bandwidth(struct dce6_wm_params *wm)
1920{
1921 /* Calculate the display mode Average Bandwidth
1922 * DisplayMode should contain the source and destination dimensions,
1923 * timing, etc.
1924 */
1925 fixed20_12 bpp;
1926 fixed20_12 line_time;
1927 fixed20_12 src_width;
1928 fixed20_12 bandwidth;
1929 fixed20_12 a;
1930
1931 a.full = dfixed_const(1000);
1932 line_time.full = dfixed_const(wm->active_time + wm->blank_time);
1933 line_time.full = dfixed_div(line_time, a);
1934 bpp.full = dfixed_const(wm->bytes_per_pixel);
1935 src_width.full = dfixed_const(wm->src_width);
1936 bandwidth.full = dfixed_mul(src_width, bpp);
1937 bandwidth.full = dfixed_mul(bandwidth, wm->vsc);
1938 bandwidth.full = dfixed_div(bandwidth, line_time);
1939
1940 return dfixed_trunc(bandwidth);
1941}
1942
1943static u32 dce6_latency_watermark(struct dce6_wm_params *wm)
1944{
1945 /* First calcualte the latency in ns */
1946 u32 mc_latency = 2000; /* 2000 ns. */
1947 u32 available_bandwidth = dce6_available_bandwidth(wm);
1948 u32 worst_chunk_return_time = (512 * 8 * 1000) / available_bandwidth;
1949 u32 cursor_line_pair_return_time = (128 * 4 * 1000) / available_bandwidth;
1950 u32 dc_latency = 40000000 / wm->disp_clk; /* dc pipe latency */
1951 u32 other_heads_data_return_time = ((wm->num_heads + 1) * worst_chunk_return_time) +
1952 (wm->num_heads * cursor_line_pair_return_time);
1953 u32 latency = mc_latency + other_heads_data_return_time + dc_latency;
1954 u32 max_src_lines_per_dst_line, lb_fill_bw, line_fill_time;
1955 u32 tmp, dmif_size = 12288;
1956 fixed20_12 a, b, c;
1957
1958 if (wm->num_heads == 0)
1959 return 0;
1960
1961 a.full = dfixed_const(2);
1962 b.full = dfixed_const(1);
1963 if ((wm->vsc.full > a.full) ||
1964 ((wm->vsc.full > b.full) && (wm->vtaps >= 3)) ||
1965 (wm->vtaps >= 5) ||
1966 ((wm->vsc.full >= a.full) && wm->interlaced))
1967 max_src_lines_per_dst_line = 4;
1968 else
1969 max_src_lines_per_dst_line = 2;
1970
1971 a.full = dfixed_const(available_bandwidth);
1972 b.full = dfixed_const(wm->num_heads);
1973 a.full = dfixed_div(a, b);
1974
1975 b.full = dfixed_const(mc_latency + 512);
1976 c.full = dfixed_const(wm->disp_clk);
1977 b.full = dfixed_div(b, c);
1978
1979 c.full = dfixed_const(dmif_size);
1980 b.full = dfixed_div(c, b);
1981
1982 tmp = min(dfixed_trunc(a), dfixed_trunc(b));
1983
1984 b.full = dfixed_const(1000);
1985 c.full = dfixed_const(wm->disp_clk);
1986 b.full = dfixed_div(c, b);
1987 c.full = dfixed_const(wm->bytes_per_pixel);
1988 b.full = dfixed_mul(b, c);
1989
1990 lb_fill_bw = min(tmp, dfixed_trunc(b));
1991
1992 a.full = dfixed_const(max_src_lines_per_dst_line * wm->src_width * wm->bytes_per_pixel);
1993 b.full = dfixed_const(1000);
1994 c.full = dfixed_const(lb_fill_bw);
1995 b.full = dfixed_div(c, b);
1996 a.full = dfixed_div(a, b);
1997 line_fill_time = dfixed_trunc(a);
1998
1999 if (line_fill_time < wm->active_time)
2000 return latency;
2001 else
2002 return latency + (line_fill_time - wm->active_time);
2003
2004}
2005
2006static bool dce6_average_bandwidth_vs_dram_bandwidth_for_display(struct dce6_wm_params *wm)
2007{
2008 if (dce6_average_bandwidth(wm) <=
2009 (dce6_dram_bandwidth_for_display(wm) / wm->num_heads))
2010 return true;
2011 else
2012 return false;
2013};
2014
2015static bool dce6_average_bandwidth_vs_available_bandwidth(struct dce6_wm_params *wm)
2016{
2017 if (dce6_average_bandwidth(wm) <=
2018 (dce6_available_bandwidth(wm) / wm->num_heads))
2019 return true;
2020 else
2021 return false;
2022};
2023
2024static bool dce6_check_latency_hiding(struct dce6_wm_params *wm)
2025{
2026 u32 lb_partitions = wm->lb_size / wm->src_width;
2027 u32 line_time = wm->active_time + wm->blank_time;
2028 u32 latency_tolerant_lines;
2029 u32 latency_hiding;
2030 fixed20_12 a;
2031
2032 a.full = dfixed_const(1);
2033 if (wm->vsc.full > a.full)
2034 latency_tolerant_lines = 1;
2035 else {
2036 if (lb_partitions <= (wm->vtaps + 1))
2037 latency_tolerant_lines = 1;
2038 else
2039 latency_tolerant_lines = 2;
2040 }
2041
2042 latency_hiding = (latency_tolerant_lines * line_time + wm->blank_time);
2043
2044 if (dce6_latency_watermark(wm) <= latency_hiding)
2045 return true;
2046 else
2047 return false;
2048}
2049
2050static void dce6_program_watermarks(struct radeon_device *rdev,
2051 struct radeon_crtc *radeon_crtc,
2052 u32 lb_size, u32 num_heads)
2053{
2054 struct drm_display_mode *mode = &radeon_crtc->base.mode;
c696e53f
AD
2055 struct dce6_wm_params wm_low, wm_high;
2056 u32 dram_channels;
43b3cd99
AD
2057 u32 pixel_period;
2058 u32 line_time = 0;
2059 u32 latency_watermark_a = 0, latency_watermark_b = 0;
2060 u32 priority_a_mark = 0, priority_b_mark = 0;
2061 u32 priority_a_cnt = PRIORITY_OFF;
2062 u32 priority_b_cnt = PRIORITY_OFF;
2063 u32 tmp, arb_control3;
2064 fixed20_12 a, b, c;
2065
2066 if (radeon_crtc->base.enabled && num_heads && mode) {
2067 pixel_period = 1000000 / (u32)mode->clock;
2068 line_time = min((u32)mode->crtc_htotal * pixel_period, (u32)65535);
2069 priority_a_cnt = 0;
2070 priority_b_cnt = 0;
2071
ca7db22b 2072 if (rdev->family == CHIP_ARUBA)
c696e53f 2073 dram_channels = evergreen_get_number_of_dram_channels(rdev);
ca7db22b 2074 else
c696e53f
AD
2075 dram_channels = si_get_number_of_dram_channels(rdev);
2076
2077 /* watermark for high clocks */
2078 if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
2079 wm_high.yclk =
2080 radeon_dpm_get_mclk(rdev, false) * 10;
2081 wm_high.sclk =
2082 radeon_dpm_get_sclk(rdev, false) * 10;
2083 } else {
2084 wm_high.yclk = rdev->pm.current_mclk * 10;
2085 wm_high.sclk = rdev->pm.current_sclk * 10;
2086 }
2087
2088 wm_high.disp_clk = mode->clock;
2089 wm_high.src_width = mode->crtc_hdisplay;
2090 wm_high.active_time = mode->crtc_hdisplay * pixel_period;
2091 wm_high.blank_time = line_time - wm_high.active_time;
2092 wm_high.interlaced = false;
2093 if (mode->flags & DRM_MODE_FLAG_INTERLACE)
2094 wm_high.interlaced = true;
2095 wm_high.vsc = radeon_crtc->vsc;
2096 wm_high.vtaps = 1;
2097 if (radeon_crtc->rmx_type != RMX_OFF)
2098 wm_high.vtaps = 2;
2099 wm_high.bytes_per_pixel = 4; /* XXX: get this from fb config */
2100 wm_high.lb_size = lb_size;
2101 wm_high.dram_channels = dram_channels;
2102 wm_high.num_heads = num_heads;
2103
2104 /* watermark for low clocks */
2105 if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
2106 wm_low.yclk =
2107 radeon_dpm_get_mclk(rdev, true) * 10;
2108 wm_low.sclk =
2109 radeon_dpm_get_sclk(rdev, true) * 10;
2110 } else {
2111 wm_low.yclk = rdev->pm.current_mclk * 10;
2112 wm_low.sclk = rdev->pm.current_sclk * 10;
2113 }
2114
2115 wm_low.disp_clk = mode->clock;
2116 wm_low.src_width = mode->crtc_hdisplay;
2117 wm_low.active_time = mode->crtc_hdisplay * pixel_period;
2118 wm_low.blank_time = line_time - wm_low.active_time;
2119 wm_low.interlaced = false;
2120 if (mode->flags & DRM_MODE_FLAG_INTERLACE)
2121 wm_low.interlaced = true;
2122 wm_low.vsc = radeon_crtc->vsc;
2123 wm_low.vtaps = 1;
2124 if (radeon_crtc->rmx_type != RMX_OFF)
2125 wm_low.vtaps = 2;
2126 wm_low.bytes_per_pixel = 4; /* XXX: get this from fb config */
2127 wm_low.lb_size = lb_size;
2128 wm_low.dram_channels = dram_channels;
2129 wm_low.num_heads = num_heads;
43b3cd99
AD
2130
2131 /* set for high clocks */
c696e53f 2132 latency_watermark_a = min(dce6_latency_watermark(&wm_high), (u32)65535);
43b3cd99 2133 /* set for low clocks */
c696e53f 2134 latency_watermark_b = min(dce6_latency_watermark(&wm_low), (u32)65535);
43b3cd99
AD
2135
2136 /* possibly force display priority to high */
2137 /* should really do this at mode validation time... */
c696e53f
AD
2138 if (!dce6_average_bandwidth_vs_dram_bandwidth_for_display(&wm_high) ||
2139 !dce6_average_bandwidth_vs_available_bandwidth(&wm_high) ||
2140 !dce6_check_latency_hiding(&wm_high) ||
2141 (rdev->disp_priority == 2)) {
2142 DRM_DEBUG_KMS("force priority to high\n");
2143 priority_a_cnt |= PRIORITY_ALWAYS_ON;
2144 priority_b_cnt |= PRIORITY_ALWAYS_ON;
2145 }
2146 if (!dce6_average_bandwidth_vs_dram_bandwidth_for_display(&wm_low) ||
2147 !dce6_average_bandwidth_vs_available_bandwidth(&wm_low) ||
2148 !dce6_check_latency_hiding(&wm_low) ||
43b3cd99
AD
2149 (rdev->disp_priority == 2)) {
2150 DRM_DEBUG_KMS("force priority to high\n");
2151 priority_a_cnt |= PRIORITY_ALWAYS_ON;
2152 priority_b_cnt |= PRIORITY_ALWAYS_ON;
2153 }
2154
2155 a.full = dfixed_const(1000);
2156 b.full = dfixed_const(mode->clock);
2157 b.full = dfixed_div(b, a);
2158 c.full = dfixed_const(latency_watermark_a);
2159 c.full = dfixed_mul(c, b);
2160 c.full = dfixed_mul(c, radeon_crtc->hsc);
2161 c.full = dfixed_div(c, a);
2162 a.full = dfixed_const(16);
2163 c.full = dfixed_div(c, a);
2164 priority_a_mark = dfixed_trunc(c);
2165 priority_a_cnt |= priority_a_mark & PRIORITY_MARK_MASK;
2166
2167 a.full = dfixed_const(1000);
2168 b.full = dfixed_const(mode->clock);
2169 b.full = dfixed_div(b, a);
2170 c.full = dfixed_const(latency_watermark_b);
2171 c.full = dfixed_mul(c, b);
2172 c.full = dfixed_mul(c, radeon_crtc->hsc);
2173 c.full = dfixed_div(c, a);
2174 a.full = dfixed_const(16);
2175 c.full = dfixed_div(c, a);
2176 priority_b_mark = dfixed_trunc(c);
2177 priority_b_cnt |= priority_b_mark & PRIORITY_MARK_MASK;
2178 }
2179
2180 /* select wm A */
2181 arb_control3 = RREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset);
2182 tmp = arb_control3;
2183 tmp &= ~LATENCY_WATERMARK_MASK(3);
2184 tmp |= LATENCY_WATERMARK_MASK(1);
2185 WREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset, tmp);
2186 WREG32(DPG_PIPE_LATENCY_CONTROL + radeon_crtc->crtc_offset,
2187 (LATENCY_LOW_WATERMARK(latency_watermark_a) |
2188 LATENCY_HIGH_WATERMARK(line_time)));
2189 /* select wm B */
2190 tmp = RREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset);
2191 tmp &= ~LATENCY_WATERMARK_MASK(3);
2192 tmp |= LATENCY_WATERMARK_MASK(2);
2193 WREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset, tmp);
2194 WREG32(DPG_PIPE_LATENCY_CONTROL + radeon_crtc->crtc_offset,
2195 (LATENCY_LOW_WATERMARK(latency_watermark_b) |
2196 LATENCY_HIGH_WATERMARK(line_time)));
2197 /* restore original selection */
2198 WREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset, arb_control3);
2199
2200 /* write the priority marks */
2201 WREG32(PRIORITY_A_CNT + radeon_crtc->crtc_offset, priority_a_cnt);
2202 WREG32(PRIORITY_B_CNT + radeon_crtc->crtc_offset, priority_b_cnt);
2203
7178d2a6
AD
2204 /* save values for DPM */
2205 radeon_crtc->line_time = line_time;
2206 radeon_crtc->wm_high = latency_watermark_a;
2207 radeon_crtc->wm_low = latency_watermark_b;
43b3cd99
AD
2208}
2209
2210void dce6_bandwidth_update(struct radeon_device *rdev)
2211{
2212 struct drm_display_mode *mode0 = NULL;
2213 struct drm_display_mode *mode1 = NULL;
2214 u32 num_heads = 0, lb_size;
2215 int i;
2216
2217 radeon_update_display_priority(rdev);
2218
2219 for (i = 0; i < rdev->num_crtc; i++) {
2220 if (rdev->mode_info.crtcs[i]->base.enabled)
2221 num_heads++;
2222 }
2223 for (i = 0; i < rdev->num_crtc; i += 2) {
2224 mode0 = &rdev->mode_info.crtcs[i]->base.mode;
2225 mode1 = &rdev->mode_info.crtcs[i+1]->base.mode;
2226 lb_size = dce6_line_buffer_adjust(rdev, rdev->mode_info.crtcs[i], mode0, mode1);
2227 dce6_program_watermarks(rdev, rdev->mode_info.crtcs[i], lb_size, num_heads);
2228 lb_size = dce6_line_buffer_adjust(rdev, rdev->mode_info.crtcs[i+1], mode1, mode0);
2229 dce6_program_watermarks(rdev, rdev->mode_info.crtcs[i+1], lb_size, num_heads);
2230 }
2231}
2232
0a96d72b
AD
2233/*
2234 * Core functions
2235 */
0a96d72b
AD
2236static void si_tiling_mode_table_init(struct radeon_device *rdev)
2237{
2238 const u32 num_tile_mode_states = 32;
2239 u32 reg_offset, gb_tile_moden, split_equal_to_row_size;
2240
2241 switch (rdev->config.si.mem_row_size_in_kb) {
2242 case 1:
2243 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_1KB;
2244 break;
2245 case 2:
2246 default:
2247 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_2KB;
2248 break;
2249 case 4:
2250 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_4KB;
2251 break;
2252 }
2253
2254 if ((rdev->family == CHIP_TAHITI) ||
2255 (rdev->family == CHIP_PITCAIRN)) {
2256 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
2257 switch (reg_offset) {
2258 case 0: /* non-AA compressed depth or any compressed stencil */
2259 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2260 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2261 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2262 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2263 NUM_BANKS(ADDR_SURF_16_BANK) |
2264 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2265 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2266 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2267 break;
2268 case 1: /* 2xAA/4xAA compressed depth only */
2269 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2270 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2271 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2272 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2273 NUM_BANKS(ADDR_SURF_16_BANK) |
2274 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2275 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2276 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2277 break;
2278 case 2: /* 8xAA compressed depth only */
2279 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2280 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2281 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2282 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2283 NUM_BANKS(ADDR_SURF_16_BANK) |
2284 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2285 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2286 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2287 break;
2288 case 3: /* 2xAA/4xAA compressed depth with stencil (for depth buffer) */
2289 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2290 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2291 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2292 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2293 NUM_BANKS(ADDR_SURF_16_BANK) |
2294 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2295 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2296 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2297 break;
2298 case 4: /* Maps w/ a dimension less than the 2D macro-tile dimensions (for mipmapped depth textures) */
2299 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2300 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2301 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2302 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2303 NUM_BANKS(ADDR_SURF_16_BANK) |
2304 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2305 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2306 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2307 break;
2308 case 5: /* Uncompressed 16bpp depth - and stencil buffer allocated with it */
2309 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2310 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2311 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2312 TILE_SPLIT(split_equal_to_row_size) |
2313 NUM_BANKS(ADDR_SURF_16_BANK) |
2314 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2315 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2316 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2317 break;
2318 case 6: /* Uncompressed 32bpp depth - and stencil buffer allocated with it */
2319 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2320 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2321 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2322 TILE_SPLIT(split_equal_to_row_size) |
2323 NUM_BANKS(ADDR_SURF_16_BANK) |
2324 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2325 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2326 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2327 break;
2328 case 7: /* Uncompressed 8bpp stencil without depth (drivers typically do not use) */
2329 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2330 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2331 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2332 TILE_SPLIT(split_equal_to_row_size) |
2333 NUM_BANKS(ADDR_SURF_16_BANK) |
2334 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2335 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2336 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2337 break;
2338 case 8: /* 1D and 1D Array Surfaces */
2339 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
2340 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2341 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2342 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2343 NUM_BANKS(ADDR_SURF_16_BANK) |
2344 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2345 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2346 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2347 break;
2348 case 9: /* Displayable maps. */
2349 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2350 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2351 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2352 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2353 NUM_BANKS(ADDR_SURF_16_BANK) |
2354 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2355 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2356 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2357 break;
2358 case 10: /* Display 8bpp. */
2359 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2360 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2361 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2362 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2363 NUM_BANKS(ADDR_SURF_16_BANK) |
2364 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2365 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2366 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2367 break;
2368 case 11: /* Display 16bpp. */
2369 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2370 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2371 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2372 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2373 NUM_BANKS(ADDR_SURF_16_BANK) |
2374 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2375 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2376 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2377 break;
2378 case 12: /* Display 32bpp. */
2379 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2380 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2381 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2382 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2383 NUM_BANKS(ADDR_SURF_16_BANK) |
2384 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2385 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2386 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2387 break;
2388 case 13: /* Thin. */
2389 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2390 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2391 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2392 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2393 NUM_BANKS(ADDR_SURF_16_BANK) |
2394 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2395 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2396 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2397 break;
2398 case 14: /* Thin 8 bpp. */
2399 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2400 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2401 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2402 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2403 NUM_BANKS(ADDR_SURF_16_BANK) |
2404 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2405 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2406 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2407 break;
2408 case 15: /* Thin 16 bpp. */
2409 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2410 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2411 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2412 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2413 NUM_BANKS(ADDR_SURF_16_BANK) |
2414 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2415 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2416 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2417 break;
2418 case 16: /* Thin 32 bpp. */
2419 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2420 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2421 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2422 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2423 NUM_BANKS(ADDR_SURF_16_BANK) |
2424 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2425 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2426 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2427 break;
2428 case 17: /* Thin 64 bpp. */
2429 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2430 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2431 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2432 TILE_SPLIT(split_equal_to_row_size) |
2433 NUM_BANKS(ADDR_SURF_16_BANK) |
2434 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2435 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2436 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2437 break;
2438 case 21: /* 8 bpp PRT. */
2439 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2440 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2441 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2442 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2443 NUM_BANKS(ADDR_SURF_16_BANK) |
2444 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
2445 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2446 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2447 break;
2448 case 22: /* 16 bpp PRT */
2449 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2450 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2451 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2452 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2453 NUM_BANKS(ADDR_SURF_16_BANK) |
2454 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2455 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2456 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2457 break;
2458 case 23: /* 32 bpp PRT */
2459 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2460 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2461 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2462 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2463 NUM_BANKS(ADDR_SURF_16_BANK) |
2464 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2465 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2466 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2467 break;
2468 case 24: /* 64 bpp PRT */
2469 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2470 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2471 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2472 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2473 NUM_BANKS(ADDR_SURF_16_BANK) |
2474 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2475 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2476 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2477 break;
2478 case 25: /* 128 bpp PRT */
2479 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2480 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2481 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2482 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
2483 NUM_BANKS(ADDR_SURF_8_BANK) |
2484 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2485 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2486 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2487 break;
2488 default:
2489 gb_tile_moden = 0;
2490 break;
2491 }
64d7b8be 2492 rdev->config.si.tile_mode_array[reg_offset] = gb_tile_moden;
0a96d72b
AD
2493 WREG32(GB_TILE_MODE0 + (reg_offset * 4), gb_tile_moden);
2494 }
d0ae7fcc 2495 } else if ((rdev->family == CHIP_VERDE) ||
8b02859d
AD
2496 (rdev->family == CHIP_OLAND) ||
2497 (rdev->family == CHIP_HAINAN)) {
0a96d72b
AD
2498 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
2499 switch (reg_offset) {
2500 case 0: /* non-AA compressed depth or any compressed stencil */
2501 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2502 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2503 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2504 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2505 NUM_BANKS(ADDR_SURF_16_BANK) |
2506 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2507 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2508 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2509 break;
2510 case 1: /* 2xAA/4xAA compressed depth only */
2511 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2512 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2513 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2514 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2515 NUM_BANKS(ADDR_SURF_16_BANK) |
2516 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2517 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2518 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2519 break;
2520 case 2: /* 8xAA compressed depth only */
2521 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2522 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2523 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2524 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2525 NUM_BANKS(ADDR_SURF_16_BANK) |
2526 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2527 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2528 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2529 break;
2530 case 3: /* 2xAA/4xAA compressed depth with stencil (for depth buffer) */
2531 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2532 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2533 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2534 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2535 NUM_BANKS(ADDR_SURF_16_BANK) |
2536 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2537 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2538 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2539 break;
2540 case 4: /* Maps w/ a dimension less than the 2D macro-tile dimensions (for mipmapped depth textures) */
2541 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2542 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2543 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2544 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2545 NUM_BANKS(ADDR_SURF_16_BANK) |
2546 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2547 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2548 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2549 break;
2550 case 5: /* Uncompressed 16bpp depth - and stencil buffer allocated with it */
2551 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2552 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2553 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2554 TILE_SPLIT(split_equal_to_row_size) |
2555 NUM_BANKS(ADDR_SURF_16_BANK) |
2556 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2557 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2558 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2559 break;
2560 case 6: /* Uncompressed 32bpp depth - and stencil buffer allocated with it */
2561 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2562 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2563 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2564 TILE_SPLIT(split_equal_to_row_size) |
2565 NUM_BANKS(ADDR_SURF_16_BANK) |
2566 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2567 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2568 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2569 break;
2570 case 7: /* Uncompressed 8bpp stencil without depth (drivers typically do not use) */
2571 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2572 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2573 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2574 TILE_SPLIT(split_equal_to_row_size) |
2575 NUM_BANKS(ADDR_SURF_16_BANK) |
2576 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2577 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2578 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2579 break;
2580 case 8: /* 1D and 1D Array Surfaces */
2581 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
2582 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2583 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2584 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2585 NUM_BANKS(ADDR_SURF_16_BANK) |
2586 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2587 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2588 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2589 break;
2590 case 9: /* Displayable maps. */
2591 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2592 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2593 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2594 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2595 NUM_BANKS(ADDR_SURF_16_BANK) |
2596 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2597 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2598 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2599 break;
2600 case 10: /* Display 8bpp. */
2601 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2602 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2603 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2604 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2605 NUM_BANKS(ADDR_SURF_16_BANK) |
2606 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2607 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2608 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2609 break;
2610 case 11: /* Display 16bpp. */
2611 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2612 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2613 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2614 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2615 NUM_BANKS(ADDR_SURF_16_BANK) |
2616 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2617 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2618 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2619 break;
2620 case 12: /* Display 32bpp. */
2621 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2622 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2623 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2624 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2625 NUM_BANKS(ADDR_SURF_16_BANK) |
2626 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2627 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2628 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2629 break;
2630 case 13: /* Thin. */
2631 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2632 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2633 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2634 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2635 NUM_BANKS(ADDR_SURF_16_BANK) |
2636 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2637 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2638 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2639 break;
2640 case 14: /* Thin 8 bpp. */
2641 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2642 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2643 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2644 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2645 NUM_BANKS(ADDR_SURF_16_BANK) |
2646 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2647 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2648 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2649 break;
2650 case 15: /* Thin 16 bpp. */
2651 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2652 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2653 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2654 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2655 NUM_BANKS(ADDR_SURF_16_BANK) |
2656 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2657 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2658 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2659 break;
2660 case 16: /* Thin 32 bpp. */
2661 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2662 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2663 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2664 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2665 NUM_BANKS(ADDR_SURF_16_BANK) |
2666 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2667 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2668 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2669 break;
2670 case 17: /* Thin 64 bpp. */
2671 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2672 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2673 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2674 TILE_SPLIT(split_equal_to_row_size) |
2675 NUM_BANKS(ADDR_SURF_16_BANK) |
2676 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2677 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2678 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2679 break;
2680 case 21: /* 8 bpp PRT. */
2681 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2682 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2683 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2684 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2685 NUM_BANKS(ADDR_SURF_16_BANK) |
2686 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
2687 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2688 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2689 break;
2690 case 22: /* 16 bpp PRT */
2691 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2692 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2693 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2694 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2695 NUM_BANKS(ADDR_SURF_16_BANK) |
2696 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2697 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2698 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2699 break;
2700 case 23: /* 32 bpp PRT */
2701 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2702 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2703 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2704 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2705 NUM_BANKS(ADDR_SURF_16_BANK) |
2706 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2707 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2708 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2709 break;
2710 case 24: /* 64 bpp PRT */
2711 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2712 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2713 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2714 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2715 NUM_BANKS(ADDR_SURF_16_BANK) |
2716 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2717 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2718 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2719 break;
2720 case 25: /* 128 bpp PRT */
2721 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2722 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2723 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2724 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
2725 NUM_BANKS(ADDR_SURF_8_BANK) |
2726 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2727 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2728 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2729 break;
2730 default:
2731 gb_tile_moden = 0;
2732 break;
2733 }
64d7b8be 2734 rdev->config.si.tile_mode_array[reg_offset] = gb_tile_moden;
0a96d72b
AD
2735 WREG32(GB_TILE_MODE0 + (reg_offset * 4), gb_tile_moden);
2736 }
2737 } else
2738 DRM_ERROR("unknown asic: 0x%x\n", rdev->family);
2739}
2740
1a8ca750
AD
2741static void si_select_se_sh(struct radeon_device *rdev,
2742 u32 se_num, u32 sh_num)
2743{
2744 u32 data = INSTANCE_BROADCAST_WRITES;
2745
2746 if ((se_num == 0xffffffff) && (sh_num == 0xffffffff))
79b52d6a 2747 data |= SH_BROADCAST_WRITES | SE_BROADCAST_WRITES;
1a8ca750
AD
2748 else if (se_num == 0xffffffff)
2749 data |= SE_BROADCAST_WRITES | SH_INDEX(sh_num);
2750 else if (sh_num == 0xffffffff)
2751 data |= SH_BROADCAST_WRITES | SE_INDEX(se_num);
2752 else
2753 data |= SH_INDEX(sh_num) | SE_INDEX(se_num);
2754 WREG32(GRBM_GFX_INDEX, data);
2755}
2756
2757static u32 si_create_bitmask(u32 bit_width)
2758{
2759 u32 i, mask = 0;
2760
2761 for (i = 0; i < bit_width; i++) {
2762 mask <<= 1;
2763 mask |= 1;
2764 }
2765 return mask;
2766}
2767
2768static u32 si_get_cu_enabled(struct radeon_device *rdev, u32 cu_per_sh)
2769{
2770 u32 data, mask;
2771
2772 data = RREG32(CC_GC_SHADER_ARRAY_CONFIG);
2773 if (data & 1)
2774 data &= INACTIVE_CUS_MASK;
2775 else
2776 data = 0;
2777 data |= RREG32(GC_USER_SHADER_ARRAY_CONFIG);
2778
2779 data >>= INACTIVE_CUS_SHIFT;
2780
2781 mask = si_create_bitmask(cu_per_sh);
2782
2783 return ~data & mask;
2784}
2785
2786static void si_setup_spi(struct radeon_device *rdev,
2787 u32 se_num, u32 sh_per_se,
2788 u32 cu_per_sh)
2789{
2790 int i, j, k;
2791 u32 data, mask, active_cu;
2792
2793 for (i = 0; i < se_num; i++) {
2794 for (j = 0; j < sh_per_se; j++) {
2795 si_select_se_sh(rdev, i, j);
2796 data = RREG32(SPI_STATIC_THREAD_MGMT_3);
2797 active_cu = si_get_cu_enabled(rdev, cu_per_sh);
2798
2799 mask = 1;
2800 for (k = 0; k < 16; k++) {
2801 mask <<= k;
2802 if (active_cu & mask) {
2803 data &= ~mask;
2804 WREG32(SPI_STATIC_THREAD_MGMT_3, data);
2805 break;
2806 }
2807 }
2808 }
2809 }
2810 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
2811}
2812
2813static u32 si_get_rb_disabled(struct radeon_device *rdev,
2814 u32 max_rb_num, u32 se_num,
2815 u32 sh_per_se)
2816{
2817 u32 data, mask;
2818
2819 data = RREG32(CC_RB_BACKEND_DISABLE);
2820 if (data & 1)
2821 data &= BACKEND_DISABLE_MASK;
2822 else
2823 data = 0;
2824 data |= RREG32(GC_USER_RB_BACKEND_DISABLE);
2825
2826 data >>= BACKEND_DISABLE_SHIFT;
2827
2828 mask = si_create_bitmask(max_rb_num / se_num / sh_per_se);
2829
2830 return data & mask;
2831}
2832
2833static void si_setup_rb(struct radeon_device *rdev,
2834 u32 se_num, u32 sh_per_se,
2835 u32 max_rb_num)
2836{
2837 int i, j;
2838 u32 data, mask;
2839 u32 disabled_rbs = 0;
2840 u32 enabled_rbs = 0;
2841
2842 for (i = 0; i < se_num; i++) {
2843 for (j = 0; j < sh_per_se; j++) {
2844 si_select_se_sh(rdev, i, j);
2845 data = si_get_rb_disabled(rdev, max_rb_num, se_num, sh_per_se);
2846 disabled_rbs |= data << ((i * sh_per_se + j) * TAHITI_RB_BITMAP_WIDTH_PER_SH);
2847 }
2848 }
2849 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
2850
2851 mask = 1;
2852 for (i = 0; i < max_rb_num; i++) {
2853 if (!(disabled_rbs & mask))
2854 enabled_rbs |= mask;
2855 mask <<= 1;
2856 }
2857
2858 for (i = 0; i < se_num; i++) {
2859 si_select_se_sh(rdev, i, 0xffffffff);
2860 data = 0;
2861 for (j = 0; j < sh_per_se; j++) {
2862 switch (enabled_rbs & 3) {
2863 case 1:
2864 data |= (RASTER_CONFIG_RB_MAP_0 << (i * sh_per_se + j) * 2);
2865 break;
2866 case 2:
2867 data |= (RASTER_CONFIG_RB_MAP_3 << (i * sh_per_se + j) * 2);
2868 break;
2869 case 3:
2870 default:
2871 data |= (RASTER_CONFIG_RB_MAP_2 << (i * sh_per_se + j) * 2);
2872 break;
2873 }
2874 enabled_rbs >>= 2;
2875 }
2876 WREG32(PA_SC_RASTER_CONFIG, data);
2877 }
2878 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
2879}
2880
0a96d72b
AD
2881static void si_gpu_init(struct radeon_device *rdev)
2882{
0a96d72b
AD
2883 u32 gb_addr_config = 0;
2884 u32 mc_shared_chmap, mc_arb_ramcfg;
0a96d72b 2885 u32 sx_debug_1;
0a96d72b
AD
2886 u32 hdp_host_path_cntl;
2887 u32 tmp;
2888 int i, j;
2889
2890 switch (rdev->family) {
2891 case CHIP_TAHITI:
2892 rdev->config.si.max_shader_engines = 2;
0a96d72b 2893 rdev->config.si.max_tile_pipes = 12;
1a8ca750
AD
2894 rdev->config.si.max_cu_per_sh = 8;
2895 rdev->config.si.max_sh_per_se = 2;
0a96d72b
AD
2896 rdev->config.si.max_backends_per_se = 4;
2897 rdev->config.si.max_texture_channel_caches = 12;
2898 rdev->config.si.max_gprs = 256;
2899 rdev->config.si.max_gs_threads = 32;
2900 rdev->config.si.max_hw_contexts = 8;
2901
2902 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
2903 rdev->config.si.sc_prim_fifo_size_backend = 0x100;
2904 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
2905 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
1a8ca750 2906 gb_addr_config = TAHITI_GB_ADDR_CONFIG_GOLDEN;
0a96d72b
AD
2907 break;
2908 case CHIP_PITCAIRN:
2909 rdev->config.si.max_shader_engines = 2;
0a96d72b 2910 rdev->config.si.max_tile_pipes = 8;
1a8ca750
AD
2911 rdev->config.si.max_cu_per_sh = 5;
2912 rdev->config.si.max_sh_per_se = 2;
0a96d72b
AD
2913 rdev->config.si.max_backends_per_se = 4;
2914 rdev->config.si.max_texture_channel_caches = 8;
2915 rdev->config.si.max_gprs = 256;
2916 rdev->config.si.max_gs_threads = 32;
2917 rdev->config.si.max_hw_contexts = 8;
2918
2919 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
2920 rdev->config.si.sc_prim_fifo_size_backend = 0x100;
2921 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
2922 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
1a8ca750 2923 gb_addr_config = TAHITI_GB_ADDR_CONFIG_GOLDEN;
0a96d72b
AD
2924 break;
2925 case CHIP_VERDE:
2926 default:
2927 rdev->config.si.max_shader_engines = 1;
0a96d72b 2928 rdev->config.si.max_tile_pipes = 4;
468ef1a5 2929 rdev->config.si.max_cu_per_sh = 5;
1a8ca750 2930 rdev->config.si.max_sh_per_se = 2;
0a96d72b
AD
2931 rdev->config.si.max_backends_per_se = 4;
2932 rdev->config.si.max_texture_channel_caches = 4;
2933 rdev->config.si.max_gprs = 256;
2934 rdev->config.si.max_gs_threads = 32;
2935 rdev->config.si.max_hw_contexts = 8;
2936
d0ae7fcc
AD
2937 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
2938 rdev->config.si.sc_prim_fifo_size_backend = 0x40;
2939 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
2940 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
2941 gb_addr_config = VERDE_GB_ADDR_CONFIG_GOLDEN;
2942 break;
2943 case CHIP_OLAND:
2944 rdev->config.si.max_shader_engines = 1;
2945 rdev->config.si.max_tile_pipes = 4;
2946 rdev->config.si.max_cu_per_sh = 6;
2947 rdev->config.si.max_sh_per_se = 1;
2948 rdev->config.si.max_backends_per_se = 2;
2949 rdev->config.si.max_texture_channel_caches = 4;
2950 rdev->config.si.max_gprs = 256;
2951 rdev->config.si.max_gs_threads = 16;
2952 rdev->config.si.max_hw_contexts = 8;
2953
0a96d72b
AD
2954 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
2955 rdev->config.si.sc_prim_fifo_size_backend = 0x40;
2956 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
2957 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
1a8ca750 2958 gb_addr_config = VERDE_GB_ADDR_CONFIG_GOLDEN;
0a96d72b 2959 break;
8b02859d
AD
2960 case CHIP_HAINAN:
2961 rdev->config.si.max_shader_engines = 1;
2962 rdev->config.si.max_tile_pipes = 4;
2963 rdev->config.si.max_cu_per_sh = 5;
2964 rdev->config.si.max_sh_per_se = 1;
2965 rdev->config.si.max_backends_per_se = 1;
2966 rdev->config.si.max_texture_channel_caches = 2;
2967 rdev->config.si.max_gprs = 256;
2968 rdev->config.si.max_gs_threads = 16;
2969 rdev->config.si.max_hw_contexts = 8;
2970
2971 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
2972 rdev->config.si.sc_prim_fifo_size_backend = 0x40;
2973 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
2974 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
2975 gb_addr_config = HAINAN_GB_ADDR_CONFIG_GOLDEN;
2976 break;
0a96d72b
AD
2977 }
2978
2979 /* Initialize HDP */
2980 for (i = 0, j = 0; i < 32; i++, j += 0x18) {
2981 WREG32((0x2c14 + j), 0x00000000);
2982 WREG32((0x2c18 + j), 0x00000000);
2983 WREG32((0x2c1c + j), 0x00000000);
2984 WREG32((0x2c20 + j), 0x00000000);
2985 WREG32((0x2c24 + j), 0x00000000);
2986 }
2987
2988 WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff));
2989
2990 evergreen_fix_pci_max_read_req_size(rdev);
2991
2992 WREG32(BIF_FB_EN, FB_READ_EN | FB_WRITE_EN);
2993
2994 mc_shared_chmap = RREG32(MC_SHARED_CHMAP);
2995 mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG);
2996
0a96d72b 2997 rdev->config.si.num_tile_pipes = rdev->config.si.max_tile_pipes;
0a96d72b
AD
2998 rdev->config.si.mem_max_burst_length_bytes = 256;
2999 tmp = (mc_arb_ramcfg & NOOFCOLS_MASK) >> NOOFCOLS_SHIFT;
3000 rdev->config.si.mem_row_size_in_kb = (4 * (1 << (8 + tmp))) / 1024;
3001 if (rdev->config.si.mem_row_size_in_kb > 4)
3002 rdev->config.si.mem_row_size_in_kb = 4;
3003 /* XXX use MC settings? */
3004 rdev->config.si.shader_engine_tile_size = 32;
3005 rdev->config.si.num_gpus = 1;
3006 rdev->config.si.multi_gpu_tile_size = 64;
3007
1a8ca750
AD
3008 /* fix up row size */
3009 gb_addr_config &= ~ROW_SIZE_MASK;
0a96d72b
AD
3010 switch (rdev->config.si.mem_row_size_in_kb) {
3011 case 1:
3012 default:
3013 gb_addr_config |= ROW_SIZE(0);
3014 break;
3015 case 2:
3016 gb_addr_config |= ROW_SIZE(1);
3017 break;
3018 case 4:
3019 gb_addr_config |= ROW_SIZE(2);
3020 break;
3021 }
3022
0a96d72b
AD
3023 /* setup tiling info dword. gb_addr_config is not adequate since it does
3024 * not have bank info, so create a custom tiling dword.
3025 * bits 3:0 num_pipes
3026 * bits 7:4 num_banks
3027 * bits 11:8 group_size
3028 * bits 15:12 row_size
3029 */
3030 rdev->config.si.tile_config = 0;
3031 switch (rdev->config.si.num_tile_pipes) {
3032 case 1:
3033 rdev->config.si.tile_config |= (0 << 0);
3034 break;
3035 case 2:
3036 rdev->config.si.tile_config |= (1 << 0);
3037 break;
3038 case 4:
3039 rdev->config.si.tile_config |= (2 << 0);
3040 break;
3041 case 8:
3042 default:
3043 /* XXX what about 12? */
3044 rdev->config.si.tile_config |= (3 << 0);
3045 break;
dca571a6
CK
3046 }
3047 switch ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT) {
3048 case 0: /* four banks */
1a8ca750 3049 rdev->config.si.tile_config |= 0 << 4;
dca571a6
CK
3050 break;
3051 case 1: /* eight banks */
3052 rdev->config.si.tile_config |= 1 << 4;
3053 break;
3054 case 2: /* sixteen banks */
3055 default:
3056 rdev->config.si.tile_config |= 2 << 4;
3057 break;
3058 }
0a96d72b
AD
3059 rdev->config.si.tile_config |=
3060 ((gb_addr_config & PIPE_INTERLEAVE_SIZE_MASK) >> PIPE_INTERLEAVE_SIZE_SHIFT) << 8;
3061 rdev->config.si.tile_config |=
3062 ((gb_addr_config & ROW_SIZE_MASK) >> ROW_SIZE_SHIFT) << 12;
3063
0a96d72b
AD
3064 WREG32(GB_ADDR_CONFIG, gb_addr_config);
3065 WREG32(DMIF_ADDR_CONFIG, gb_addr_config);
7c1c7c18 3066 WREG32(DMIF_ADDR_CALC, gb_addr_config);
0a96d72b 3067 WREG32(HDP_ADDR_CONFIG, gb_addr_config);
8c5fd7ef
AD
3068 WREG32(DMA_TILING_CONFIG + DMA0_REGISTER_OFFSET, gb_addr_config);
3069 WREG32(DMA_TILING_CONFIG + DMA1_REGISTER_OFFSET, gb_addr_config);
1df0d523
AD
3070 if (rdev->has_uvd) {
3071 WREG32(UVD_UDEC_ADDR_CONFIG, gb_addr_config);
3072 WREG32(UVD_UDEC_DB_ADDR_CONFIG, gb_addr_config);
3073 WREG32(UVD_UDEC_DBW_ADDR_CONFIG, gb_addr_config);
3074 }
0a96d72b 3075
1a8ca750 3076 si_tiling_mode_table_init(rdev);
0a96d72b 3077
1a8ca750
AD
3078 si_setup_rb(rdev, rdev->config.si.max_shader_engines,
3079 rdev->config.si.max_sh_per_se,
3080 rdev->config.si.max_backends_per_se);
0a96d72b 3081
1a8ca750
AD
3082 si_setup_spi(rdev, rdev->config.si.max_shader_engines,
3083 rdev->config.si.max_sh_per_se,
3084 rdev->config.si.max_cu_per_sh);
0a96d72b 3085
0a96d72b
AD
3086
3087 /* set HW defaults for 3D engine */
3088 WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) |
3089 ROQ_IB2_START(0x2b)));
3090 WREG32(CP_MEQ_THRESHOLDS, MEQ1_START(0x30) | MEQ2_START(0x60));
3091
3092 sx_debug_1 = RREG32(SX_DEBUG_1);
3093 WREG32(SX_DEBUG_1, sx_debug_1);
3094
3095 WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4));
3096
3097 WREG32(PA_SC_FIFO_SIZE, (SC_FRONTEND_PRIM_FIFO_SIZE(rdev->config.si.sc_prim_fifo_size_frontend) |
3098 SC_BACKEND_PRIM_FIFO_SIZE(rdev->config.si.sc_prim_fifo_size_backend) |
3099 SC_HIZ_TILE_FIFO_SIZE(rdev->config.si.sc_hiz_tile_fifo_size) |
3100 SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.si.sc_earlyz_tile_fifo_size)));
3101
3102 WREG32(VGT_NUM_INSTANCES, 1);
3103
3104 WREG32(CP_PERFMON_CNTL, 0);
3105
3106 WREG32(SQ_CONFIG, 0);
3107
3108 WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) |
3109 FORCE_EOV_MAX_REZ_CNT(255)));
3110
3111 WREG32(VGT_CACHE_INVALIDATION, CACHE_INVALIDATION(VC_AND_TC) |
3112 AUTO_INVLD_EN(ES_AND_GS_AUTO));
3113
3114 WREG32(VGT_GS_VERTEX_REUSE, 16);
3115 WREG32(PA_SC_LINE_STIPPLE_STATE, 0);
3116
3117 WREG32(CB_PERFCOUNTER0_SELECT0, 0);
3118 WREG32(CB_PERFCOUNTER0_SELECT1, 0);
3119 WREG32(CB_PERFCOUNTER1_SELECT0, 0);
3120 WREG32(CB_PERFCOUNTER1_SELECT1, 0);
3121 WREG32(CB_PERFCOUNTER2_SELECT0, 0);
3122 WREG32(CB_PERFCOUNTER2_SELECT1, 0);
3123 WREG32(CB_PERFCOUNTER3_SELECT0, 0);
3124 WREG32(CB_PERFCOUNTER3_SELECT1, 0);
3125
3126 tmp = RREG32(HDP_MISC_CNTL);
3127 tmp |= HDP_FLUSH_INVALIDATE_CACHE;
3128 WREG32(HDP_MISC_CNTL, tmp);
3129
3130 hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL);
3131 WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
3132
3133 WREG32(PA_CL_ENHANCE, CLIP_VTX_REORDER_ENA | NUM_CLIP_SEQ(3));
3134
3135 udelay(50);
3136}
c476dde2 3137
2ece2e8b
AD
3138/*
3139 * GPU scratch registers helpers function.
3140 */
3141static void si_scratch_init(struct radeon_device *rdev)
3142{
3143 int i;
3144
3145 rdev->scratch.num_reg = 7;
3146 rdev->scratch.reg_base = SCRATCH_REG0;
3147 for (i = 0; i < rdev->scratch.num_reg; i++) {
3148 rdev->scratch.free[i] = true;
3149 rdev->scratch.reg[i] = rdev->scratch.reg_base + (i * 4);
3150 }
3151}
3152
3153void si_fence_ring_emit(struct radeon_device *rdev,
3154 struct radeon_fence *fence)
3155{
3156 struct radeon_ring *ring = &rdev->ring[fence->ring];
3157 u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
3158
3159 /* flush read cache over gart */
3160 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
3161 radeon_ring_write(ring, (CP_COHER_CNTL2 - PACKET3_SET_CONFIG_REG_START) >> 2);
3162 radeon_ring_write(ring, 0);
3163 radeon_ring_write(ring, PACKET3(PACKET3_SURFACE_SYNC, 3));
3164 radeon_ring_write(ring, PACKET3_TCL1_ACTION_ENA |
3165 PACKET3_TC_ACTION_ENA |
3166 PACKET3_SH_KCACHE_ACTION_ENA |
3167 PACKET3_SH_ICACHE_ACTION_ENA);
3168 radeon_ring_write(ring, 0xFFFFFFFF);
3169 radeon_ring_write(ring, 0);
3170 radeon_ring_write(ring, 10); /* poll interval */
3171 /* EVENT_WRITE_EOP - flush caches, send int */
3172 radeon_ring_write(ring, PACKET3(PACKET3_EVENT_WRITE_EOP, 4));
3173 radeon_ring_write(ring, EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) | EVENT_INDEX(5));
3174 radeon_ring_write(ring, addr & 0xffffffff);
3175 radeon_ring_write(ring, (upper_32_bits(addr) & 0xff) | DATA_SEL(1) | INT_SEL(2));
3176 radeon_ring_write(ring, fence->seq);
3177 radeon_ring_write(ring, 0);
3178}
3179
3180/*
3181 * IB stuff
3182 */
3183void si_ring_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib)
3184{
876dc9f3 3185 struct radeon_ring *ring = &rdev->ring[ib->ring];
2ece2e8b
AD
3186 u32 header;
3187
a85a7da4
AD
3188 if (ib->is_const_ib) {
3189 /* set switch buffer packet before const IB */
3190 radeon_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
3191 radeon_ring_write(ring, 0);
45df6803 3192
2ece2e8b 3193 header = PACKET3(PACKET3_INDIRECT_BUFFER_CONST, 2);
a85a7da4 3194 } else {
89d35807 3195 u32 next_rptr;
a85a7da4 3196 if (ring->rptr_save_reg) {
89d35807 3197 next_rptr = ring->wptr + 3 + 4 + 8;
a85a7da4
AD
3198 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
3199 radeon_ring_write(ring, ((ring->rptr_save_reg -
3200 PACKET3_SET_CONFIG_REG_START) >> 2));
3201 radeon_ring_write(ring, next_rptr);
89d35807
AD
3202 } else if (rdev->wb.enabled) {
3203 next_rptr = ring->wptr + 5 + 4 + 8;
3204 radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
3205 radeon_ring_write(ring, (1 << 8));
3206 radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
3207 radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr) & 0xffffffff);
3208 radeon_ring_write(ring, next_rptr);
a85a7da4
AD
3209 }
3210
2ece2e8b 3211 header = PACKET3(PACKET3_INDIRECT_BUFFER, 2);
a85a7da4 3212 }
2ece2e8b
AD
3213
3214 radeon_ring_write(ring, header);
3215 radeon_ring_write(ring,
3216#ifdef __BIG_ENDIAN
3217 (2 << 0) |
3218#endif
3219 (ib->gpu_addr & 0xFFFFFFFC));
3220 radeon_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xFFFF);
4bf3dd92
CK
3221 radeon_ring_write(ring, ib->length_dw |
3222 (ib->vm ? (ib->vm->id << 24) : 0));
2ece2e8b 3223
a85a7da4
AD
3224 if (!ib->is_const_ib) {
3225 /* flush read cache over gart for this vmid */
3226 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
3227 radeon_ring_write(ring, (CP_COHER_CNTL2 - PACKET3_SET_CONFIG_REG_START) >> 2);
4bf3dd92 3228 radeon_ring_write(ring, ib->vm ? ib->vm->id : 0);
a85a7da4
AD
3229 radeon_ring_write(ring, PACKET3(PACKET3_SURFACE_SYNC, 3));
3230 radeon_ring_write(ring, PACKET3_TCL1_ACTION_ENA |
3231 PACKET3_TC_ACTION_ENA |
3232 PACKET3_SH_KCACHE_ACTION_ENA |
3233 PACKET3_SH_ICACHE_ACTION_ENA);
3234 radeon_ring_write(ring, 0xFFFFFFFF);
3235 radeon_ring_write(ring, 0);
3236 radeon_ring_write(ring, 10); /* poll interval */
3237 }
2ece2e8b
AD
3238}
3239
48c0c902
AD
3240/*
3241 * CP.
3242 */
3243static void si_cp_enable(struct radeon_device *rdev, bool enable)
3244{
3245 if (enable)
3246 WREG32(CP_ME_CNTL, 0);
3247 else {
3248 radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
3249 WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT | CP_CE_HALT));
3250 WREG32(SCRATCH_UMSK, 0);
8c5fd7ef
AD
3251 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
3252 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
3253 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
48c0c902
AD
3254 }
3255 udelay(50);
3256}
3257
3258static int si_cp_load_microcode(struct radeon_device *rdev)
3259{
3260 const __be32 *fw_data;
3261 int i;
3262
3263 if (!rdev->me_fw || !rdev->pfp_fw)
3264 return -EINVAL;
3265
3266 si_cp_enable(rdev, false);
3267
3268 /* PFP */
3269 fw_data = (const __be32 *)rdev->pfp_fw->data;
3270 WREG32(CP_PFP_UCODE_ADDR, 0);
3271 for (i = 0; i < SI_PFP_UCODE_SIZE; i++)
3272 WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++));
3273 WREG32(CP_PFP_UCODE_ADDR, 0);
3274
3275 /* CE */
3276 fw_data = (const __be32 *)rdev->ce_fw->data;
3277 WREG32(CP_CE_UCODE_ADDR, 0);
3278 for (i = 0; i < SI_CE_UCODE_SIZE; i++)
3279 WREG32(CP_CE_UCODE_DATA, be32_to_cpup(fw_data++));
3280 WREG32(CP_CE_UCODE_ADDR, 0);
3281
3282 /* ME */
3283 fw_data = (const __be32 *)rdev->me_fw->data;
3284 WREG32(CP_ME_RAM_WADDR, 0);
3285 for (i = 0; i < SI_PM4_UCODE_SIZE; i++)
3286 WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++));
3287 WREG32(CP_ME_RAM_WADDR, 0);
3288
3289 WREG32(CP_PFP_UCODE_ADDR, 0);
3290 WREG32(CP_CE_UCODE_ADDR, 0);
3291 WREG32(CP_ME_RAM_WADDR, 0);
3292 WREG32(CP_ME_RAM_RADDR, 0);
3293 return 0;
3294}
3295
3296static int si_cp_start(struct radeon_device *rdev)
3297{
3298 struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
3299 int r, i;
3300
3301 r = radeon_ring_lock(rdev, ring, 7 + 4);
3302 if (r) {
3303 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
3304 return r;
3305 }
3306 /* init the CP */
3307 radeon_ring_write(ring, PACKET3(PACKET3_ME_INITIALIZE, 5));
3308 radeon_ring_write(ring, 0x1);
3309 radeon_ring_write(ring, 0x0);
3310 radeon_ring_write(ring, rdev->config.si.max_hw_contexts - 1);
3311 radeon_ring_write(ring, PACKET3_ME_INITIALIZE_DEVICE_ID(1));
3312 radeon_ring_write(ring, 0);
3313 radeon_ring_write(ring, 0);
3314
3315 /* init the CE partitions */
3316 radeon_ring_write(ring, PACKET3(PACKET3_SET_BASE, 2));
3317 radeon_ring_write(ring, PACKET3_BASE_INDEX(CE_PARTITION_BASE));
3318 radeon_ring_write(ring, 0xc000);
3319 radeon_ring_write(ring, 0xe000);
3320 radeon_ring_unlock_commit(rdev, ring);
3321
3322 si_cp_enable(rdev, true);
3323
3324 r = radeon_ring_lock(rdev, ring, si_default_size + 10);
3325 if (r) {
3326 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
3327 return r;
3328 }
3329
3330 /* setup clear context state */
3331 radeon_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
3332 radeon_ring_write(ring, PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
3333
3334 for (i = 0; i < si_default_size; i++)
3335 radeon_ring_write(ring, si_default_state[i]);
3336
3337 radeon_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
3338 radeon_ring_write(ring, PACKET3_PREAMBLE_END_CLEAR_STATE);
3339
3340 /* set clear context state */
3341 radeon_ring_write(ring, PACKET3(PACKET3_CLEAR_STATE, 0));
3342 radeon_ring_write(ring, 0);
3343
3344 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONTEXT_REG, 2));
3345 radeon_ring_write(ring, 0x00000316);
3346 radeon_ring_write(ring, 0x0000000e); /* VGT_VERTEX_REUSE_BLOCK_CNTL */
3347 radeon_ring_write(ring, 0x00000010); /* VGT_OUT_DEALLOC_CNTL */
3348
3349 radeon_ring_unlock_commit(rdev, ring);
3350
3351 for (i = RADEON_RING_TYPE_GFX_INDEX; i <= CAYMAN_RING_TYPE_CP2_INDEX; ++i) {
3352 ring = &rdev->ring[i];
3353 r = radeon_ring_lock(rdev, ring, 2);
3354
3355 /* clear the compute context state */
3356 radeon_ring_write(ring, PACKET3_COMPUTE(PACKET3_CLEAR_STATE, 0));
3357 radeon_ring_write(ring, 0);
3358
3359 radeon_ring_unlock_commit(rdev, ring);
3360 }
3361
3362 return 0;
3363}
3364
3365static void si_cp_fini(struct radeon_device *rdev)
3366{
45df6803 3367 struct radeon_ring *ring;
48c0c902 3368 si_cp_enable(rdev, false);
45df6803
CK
3369
3370 ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
3371 radeon_ring_fini(rdev, ring);
3372 radeon_scratch_free(rdev, ring->rptr_save_reg);
3373
3374 ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
3375 radeon_ring_fini(rdev, ring);
3376 radeon_scratch_free(rdev, ring->rptr_save_reg);
3377
3378 ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
3379 radeon_ring_fini(rdev, ring);
3380 radeon_scratch_free(rdev, ring->rptr_save_reg);
48c0c902
AD
3381}
3382
3383static int si_cp_resume(struct radeon_device *rdev)
3384{
3385 struct radeon_ring *ring;
3386 u32 tmp;
3387 u32 rb_bufsz;
3388 int r;
3389
811e4d58
AD
3390 si_enable_gui_idle_interrupt(rdev, false);
3391
48c0c902
AD
3392 WREG32(CP_SEM_WAIT_TIMER, 0x0);
3393 WREG32(CP_SEM_INCOMPLETE_TIMER_CNTL, 0x0);
3394
3395 /* Set the write pointer delay */
3396 WREG32(CP_RB_WPTR_DELAY, 0);
3397
3398 WREG32(CP_DEBUG, 0);
3399 WREG32(SCRATCH_ADDR, ((rdev->wb.gpu_addr + RADEON_WB_SCRATCH_OFFSET) >> 8) & 0xFFFFFFFF);
3400
3401 /* ring 0 - compute and gfx */
3402 /* Set ring buffer size */
3403 ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
b72a8925
DV
3404 rb_bufsz = order_base_2(ring->ring_size / 8);
3405 tmp = (order_base_2(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
48c0c902
AD
3406#ifdef __BIG_ENDIAN
3407 tmp |= BUF_SWAP_32BIT;
3408#endif
3409 WREG32(CP_RB0_CNTL, tmp);
3410
3411 /* Initialize the ring buffer's read and write pointers */
3412 WREG32(CP_RB0_CNTL, tmp | RB_RPTR_WR_ENA);
3413 ring->wptr = 0;
3414 WREG32(CP_RB0_WPTR, ring->wptr);
3415
48fc7f7e 3416 /* set the wb address whether it's enabled or not */
48c0c902
AD
3417 WREG32(CP_RB0_RPTR_ADDR, (rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET) & 0xFFFFFFFC);
3418 WREG32(CP_RB0_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET) & 0xFF);
3419
3420 if (rdev->wb.enabled)
3421 WREG32(SCRATCH_UMSK, 0xff);
3422 else {
3423 tmp |= RB_NO_UPDATE;
3424 WREG32(SCRATCH_UMSK, 0);
3425 }
3426
3427 mdelay(1);
3428 WREG32(CP_RB0_CNTL, tmp);
3429
3430 WREG32(CP_RB0_BASE, ring->gpu_addr >> 8);
3431
3432 ring->rptr = RREG32(CP_RB0_RPTR);
3433
3434 /* ring1 - compute only */
3435 /* Set ring buffer size */
3436 ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
b72a8925
DV
3437 rb_bufsz = order_base_2(ring->ring_size / 8);
3438 tmp = (order_base_2(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
48c0c902
AD
3439#ifdef __BIG_ENDIAN
3440 tmp |= BUF_SWAP_32BIT;
3441#endif
3442 WREG32(CP_RB1_CNTL, tmp);
3443
3444 /* Initialize the ring buffer's read and write pointers */
3445 WREG32(CP_RB1_CNTL, tmp | RB_RPTR_WR_ENA);
3446 ring->wptr = 0;
3447 WREG32(CP_RB1_WPTR, ring->wptr);
3448
48fc7f7e 3449 /* set the wb address whether it's enabled or not */
48c0c902
AD
3450 WREG32(CP_RB1_RPTR_ADDR, (rdev->wb.gpu_addr + RADEON_WB_CP1_RPTR_OFFSET) & 0xFFFFFFFC);
3451 WREG32(CP_RB1_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP1_RPTR_OFFSET) & 0xFF);
3452
3453 mdelay(1);
3454 WREG32(CP_RB1_CNTL, tmp);
3455
3456 WREG32(CP_RB1_BASE, ring->gpu_addr >> 8);
3457
3458 ring->rptr = RREG32(CP_RB1_RPTR);
3459
3460 /* ring2 - compute only */
3461 /* Set ring buffer size */
3462 ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
b72a8925
DV
3463 rb_bufsz = order_base_2(ring->ring_size / 8);
3464 tmp = (order_base_2(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
48c0c902
AD
3465#ifdef __BIG_ENDIAN
3466 tmp |= BUF_SWAP_32BIT;
3467#endif
3468 WREG32(CP_RB2_CNTL, tmp);
3469
3470 /* Initialize the ring buffer's read and write pointers */
3471 WREG32(CP_RB2_CNTL, tmp | RB_RPTR_WR_ENA);
3472 ring->wptr = 0;
3473 WREG32(CP_RB2_WPTR, ring->wptr);
3474
48fc7f7e 3475 /* set the wb address whether it's enabled or not */
48c0c902
AD
3476 WREG32(CP_RB2_RPTR_ADDR, (rdev->wb.gpu_addr + RADEON_WB_CP2_RPTR_OFFSET) & 0xFFFFFFFC);
3477 WREG32(CP_RB2_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP2_RPTR_OFFSET) & 0xFF);
3478
3479 mdelay(1);
3480 WREG32(CP_RB2_CNTL, tmp);
3481
3482 WREG32(CP_RB2_BASE, ring->gpu_addr >> 8);
3483
3484 ring->rptr = RREG32(CP_RB2_RPTR);
3485
3486 /* start the rings */
3487 si_cp_start(rdev);
3488 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = true;
3489 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = true;
3490 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = true;
3491 r = radeon_ring_test(rdev, RADEON_RING_TYPE_GFX_INDEX, &rdev->ring[RADEON_RING_TYPE_GFX_INDEX]);
3492 if (r) {
3493 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
3494 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
3495 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
3496 return r;
3497 }
3498 r = radeon_ring_test(rdev, CAYMAN_RING_TYPE_CP1_INDEX, &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX]);
3499 if (r) {
3500 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
3501 }
3502 r = radeon_ring_test(rdev, CAYMAN_RING_TYPE_CP2_INDEX, &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX]);
3503 if (r) {
3504 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
3505 }
3506
811e4d58
AD
3507 si_enable_gui_idle_interrupt(rdev, true);
3508
48c0c902
AD
3509 return 0;
3510}
3511
2483b4ea 3512u32 si_gpu_check_soft_reset(struct radeon_device *rdev)
06bc6df0 3513{
014bb209 3514 u32 reset_mask = 0;
1c534671 3515 u32 tmp;
06bc6df0 3516
014bb209
AD
3517 /* GRBM_STATUS */
3518 tmp = RREG32(GRBM_STATUS);
3519 if (tmp & (PA_BUSY | SC_BUSY |
3520 BCI_BUSY | SX_BUSY |
3521 TA_BUSY | VGT_BUSY |
3522 DB_BUSY | CB_BUSY |
3523 GDS_BUSY | SPI_BUSY |
3524 IA_BUSY | IA_BUSY_NO_DMA))
3525 reset_mask |= RADEON_RESET_GFX;
3526
3527 if (tmp & (CF_RQ_PENDING | PF_RQ_PENDING |
3528 CP_BUSY | CP_COHERENCY_BUSY))
3529 reset_mask |= RADEON_RESET_CP;
3530
3531 if (tmp & GRBM_EE_BUSY)
3532 reset_mask |= RADEON_RESET_GRBM | RADEON_RESET_GFX | RADEON_RESET_CP;
3533
3534 /* GRBM_STATUS2 */
3535 tmp = RREG32(GRBM_STATUS2);
3536 if (tmp & (RLC_RQ_PENDING | RLC_BUSY))
3537 reset_mask |= RADEON_RESET_RLC;
3538
3539 /* DMA_STATUS_REG 0 */
3540 tmp = RREG32(DMA_STATUS_REG + DMA0_REGISTER_OFFSET);
3541 if (!(tmp & DMA_IDLE))
3542 reset_mask |= RADEON_RESET_DMA;
3543
3544 /* DMA_STATUS_REG 1 */
3545 tmp = RREG32(DMA_STATUS_REG + DMA1_REGISTER_OFFSET);
3546 if (!(tmp & DMA_IDLE))
3547 reset_mask |= RADEON_RESET_DMA1;
3548
3549 /* SRBM_STATUS2 */
3550 tmp = RREG32(SRBM_STATUS2);
3551 if (tmp & DMA_BUSY)
3552 reset_mask |= RADEON_RESET_DMA;
3553
3554 if (tmp & DMA1_BUSY)
3555 reset_mask |= RADEON_RESET_DMA1;
3556
3557 /* SRBM_STATUS */
3558 tmp = RREG32(SRBM_STATUS);
3559
3560 if (tmp & IH_BUSY)
3561 reset_mask |= RADEON_RESET_IH;
3562
3563 if (tmp & SEM_BUSY)
3564 reset_mask |= RADEON_RESET_SEM;
3565
3566 if (tmp & GRBM_RQ_PENDING)
3567 reset_mask |= RADEON_RESET_GRBM;
3568
3569 if (tmp & VMC_BUSY)
3570 reset_mask |= RADEON_RESET_VMC;
19fc42ed 3571
014bb209
AD
3572 if (tmp & (MCB_BUSY | MCB_NON_DISPLAY_BUSY |
3573 MCC_BUSY | MCD_BUSY))
3574 reset_mask |= RADEON_RESET_MC;
3575
3576 if (evergreen_is_display_hung(rdev))
3577 reset_mask |= RADEON_RESET_DISPLAY;
3578
3579 /* VM_L2_STATUS */
3580 tmp = RREG32(VM_L2_STATUS);
3581 if (tmp & L2_BUSY)
3582 reset_mask |= RADEON_RESET_VMC;
3583
d808fc88
AD
3584 /* Skip MC reset as it's mostly likely not hung, just busy */
3585 if (reset_mask & RADEON_RESET_MC) {
3586 DRM_DEBUG("MC busy: 0x%08X, clearing.\n", reset_mask);
3587 reset_mask &= ~RADEON_RESET_MC;
3588 }
3589
014bb209
AD
3590 return reset_mask;
3591}
3592
3593static void si_gpu_soft_reset(struct radeon_device *rdev, u32 reset_mask)
3594{
3595 struct evergreen_mc_save save;
3596 u32 grbm_soft_reset = 0, srbm_soft_reset = 0;
3597 u32 tmp;
19fc42ed 3598
06bc6df0 3599 if (reset_mask == 0)
014bb209 3600 return;
06bc6df0
AD
3601
3602 dev_info(rdev->dev, "GPU softreset: 0x%08X\n", reset_mask);
3603
1c534671 3604 evergreen_print_gpu_status_regs(rdev);
06bc6df0
AD
3605 dev_info(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n",
3606 RREG32(VM_CONTEXT1_PROTECTION_FAULT_ADDR));
3607 dev_info(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
3608 RREG32(VM_CONTEXT1_PROTECTION_FAULT_STATUS));
3609
a6f4ae8d
AD
3610 /* disable PG/CG */
3611 si_fini_pg(rdev);
3612 si_fini_cg(rdev);
3613
3614 /* stop the rlc */
3615 si_rlc_stop(rdev);
3616
1c534671
AD
3617 /* Disable CP parsing/prefetching */
3618 WREG32(CP_ME_CNTL, CP_ME_HALT | CP_PFP_HALT | CP_CE_HALT);
3619
3620 if (reset_mask & RADEON_RESET_DMA) {
3621 /* dma0 */
3622 tmp = RREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET);
3623 tmp &= ~DMA_RB_ENABLE;
3624 WREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET, tmp);
014bb209
AD
3625 }
3626 if (reset_mask & RADEON_RESET_DMA1) {
1c534671
AD
3627 /* dma1 */
3628 tmp = RREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET);
3629 tmp &= ~DMA_RB_ENABLE;
3630 WREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET, tmp);
3631 }
3632
f770d78a
AD
3633 udelay(50);
3634
3635 evergreen_mc_stop(rdev, &save);
3636 if (evergreen_mc_wait_for_idle(rdev)) {
3637 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
3638 }
3639
1c534671
AD
3640 if (reset_mask & (RADEON_RESET_GFX | RADEON_RESET_COMPUTE | RADEON_RESET_CP)) {
3641 grbm_soft_reset = SOFT_RESET_CB |
3642 SOFT_RESET_DB |
3643 SOFT_RESET_GDS |
3644 SOFT_RESET_PA |
3645 SOFT_RESET_SC |
3646 SOFT_RESET_BCI |
3647 SOFT_RESET_SPI |
3648 SOFT_RESET_SX |
3649 SOFT_RESET_TC |
3650 SOFT_RESET_TA |
3651 SOFT_RESET_VGT |
3652 SOFT_RESET_IA;
3653 }
3654
3655 if (reset_mask & RADEON_RESET_CP) {
3656 grbm_soft_reset |= SOFT_RESET_CP | SOFT_RESET_VGT;
3657
3658 srbm_soft_reset |= SOFT_RESET_GRBM;
3659 }
06bc6df0
AD
3660
3661 if (reset_mask & RADEON_RESET_DMA)
014bb209
AD
3662 srbm_soft_reset |= SOFT_RESET_DMA;
3663
3664 if (reset_mask & RADEON_RESET_DMA1)
3665 srbm_soft_reset |= SOFT_RESET_DMA1;
3666
3667 if (reset_mask & RADEON_RESET_DISPLAY)
3668 srbm_soft_reset |= SOFT_RESET_DC;
3669
3670 if (reset_mask & RADEON_RESET_RLC)
3671 grbm_soft_reset |= SOFT_RESET_RLC;
3672
3673 if (reset_mask & RADEON_RESET_SEM)
3674 srbm_soft_reset |= SOFT_RESET_SEM;
3675
3676 if (reset_mask & RADEON_RESET_IH)
3677 srbm_soft_reset |= SOFT_RESET_IH;
3678
3679 if (reset_mask & RADEON_RESET_GRBM)
3680 srbm_soft_reset |= SOFT_RESET_GRBM;
3681
3682 if (reset_mask & RADEON_RESET_VMC)
3683 srbm_soft_reset |= SOFT_RESET_VMC;
3684
3685 if (reset_mask & RADEON_RESET_MC)
3686 srbm_soft_reset |= SOFT_RESET_MC;
1c534671
AD
3687
3688 if (grbm_soft_reset) {
3689 tmp = RREG32(GRBM_SOFT_RESET);
3690 tmp |= grbm_soft_reset;
3691 dev_info(rdev->dev, "GRBM_SOFT_RESET=0x%08X\n", tmp);
3692 WREG32(GRBM_SOFT_RESET, tmp);
3693 tmp = RREG32(GRBM_SOFT_RESET);
3694
3695 udelay(50);
3696
3697 tmp &= ~grbm_soft_reset;
3698 WREG32(GRBM_SOFT_RESET, tmp);
3699 tmp = RREG32(GRBM_SOFT_RESET);
3700 }
3701
3702 if (srbm_soft_reset) {
3703 tmp = RREG32(SRBM_SOFT_RESET);
3704 tmp |= srbm_soft_reset;
3705 dev_info(rdev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
3706 WREG32(SRBM_SOFT_RESET, tmp);
3707 tmp = RREG32(SRBM_SOFT_RESET);
3708
3709 udelay(50);
3710
3711 tmp &= ~srbm_soft_reset;
3712 WREG32(SRBM_SOFT_RESET, tmp);
3713 tmp = RREG32(SRBM_SOFT_RESET);
3714 }
06bc6df0
AD
3715
3716 /* Wait a little for things to settle down */
3717 udelay(50);
3718
c476dde2 3719 evergreen_mc_resume(rdev, &save);
1c534671
AD
3720 udelay(50);
3721
1c534671 3722 evergreen_print_gpu_status_regs(rdev);
c476dde2
AD
3723}
3724
3725int si_asic_reset(struct radeon_device *rdev)
3726{
014bb209
AD
3727 u32 reset_mask;
3728
3729 reset_mask = si_gpu_check_soft_reset(rdev);
3730
3731 if (reset_mask)
3732 r600_set_bios_scratch_engine_hung(rdev, true);
3733
3734 si_gpu_soft_reset(rdev, reset_mask);
3735
3736 reset_mask = si_gpu_check_soft_reset(rdev);
3737
3738 if (!reset_mask)
3739 r600_set_bios_scratch_engine_hung(rdev, false);
3740
3741 return 0;
c476dde2
AD
3742}
3743
123bc183
AD
3744/**
3745 * si_gfx_is_lockup - Check if the GFX engine is locked up
3746 *
3747 * @rdev: radeon_device pointer
3748 * @ring: radeon_ring structure holding ring information
3749 *
3750 * Check if the GFX engine is locked up.
3751 * Returns true if the engine appears to be locked up, false if not.
3752 */
3753bool si_gfx_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
3754{
3755 u32 reset_mask = si_gpu_check_soft_reset(rdev);
3756
3757 if (!(reset_mask & (RADEON_RESET_GFX |
3758 RADEON_RESET_COMPUTE |
3759 RADEON_RESET_CP))) {
3760 radeon_ring_lockup_update(ring);
3761 return false;
3762 }
3763 /* force CP activities */
3764 radeon_ring_force_activity(rdev, ring);
3765 return radeon_ring_test_lockup(rdev, ring);
3766}
3767
d2800ee5
AD
3768/* MC */
3769static void si_mc_program(struct radeon_device *rdev)
3770{
3771 struct evergreen_mc_save save;
3772 u32 tmp;
3773 int i, j;
3774
3775 /* Initialize HDP */
3776 for (i = 0, j = 0; i < 32; i++, j += 0x18) {
3777 WREG32((0x2c14 + j), 0x00000000);
3778 WREG32((0x2c18 + j), 0x00000000);
3779 WREG32((0x2c1c + j), 0x00000000);
3780 WREG32((0x2c20 + j), 0x00000000);
3781 WREG32((0x2c24 + j), 0x00000000);
3782 }
3783 WREG32(HDP_REG_COHERENCY_FLUSH_CNTL, 0);
3784
3785 evergreen_mc_stop(rdev, &save);
3786 if (radeon_mc_wait_for_idle(rdev)) {
3787 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
3788 }
5153550a
AD
3789 if (!ASIC_IS_NODCE(rdev))
3790 /* Lockout access through VGA aperture*/
3791 WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE);
d2800ee5
AD
3792 /* Update configuration */
3793 WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
3794 rdev->mc.vram_start >> 12);
3795 WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
3796 rdev->mc.vram_end >> 12);
3797 WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR,
3798 rdev->vram_scratch.gpu_addr >> 12);
3799 tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16;
3800 tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF);
3801 WREG32(MC_VM_FB_LOCATION, tmp);
3802 /* XXX double check these! */
3803 WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8));
3804 WREG32(HDP_NONSURFACE_INFO, (2 << 7) | (1 << 30));
3805 WREG32(HDP_NONSURFACE_SIZE, 0x3FFFFFFF);
3806 WREG32(MC_VM_AGP_BASE, 0);
3807 WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF);
3808 WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF);
3809 if (radeon_mc_wait_for_idle(rdev)) {
3810 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
3811 }
3812 evergreen_mc_resume(rdev, &save);
5153550a
AD
3813 if (!ASIC_IS_NODCE(rdev)) {
3814 /* we need to own VRAM, so turn off the VGA renderer here
3815 * to stop it overwriting our objects */
3816 rv515_vga_render_disable(rdev);
3817 }
d2800ee5
AD
3818}
3819
1c49165d
AD
3820void si_vram_gtt_location(struct radeon_device *rdev,
3821 struct radeon_mc *mc)
d2800ee5
AD
3822{
3823 if (mc->mc_vram_size > 0xFFC0000000ULL) {
3824 /* leave room for at least 1024M GTT */
3825 dev_warn(rdev->dev, "limiting VRAM\n");
3826 mc->real_vram_size = 0xFFC0000000ULL;
3827 mc->mc_vram_size = 0xFFC0000000ULL;
3828 }
9ed8b1f9 3829 radeon_vram_location(rdev, &rdev->mc, 0);
d2800ee5 3830 rdev->mc.gtt_base_align = 0;
9ed8b1f9 3831 radeon_gtt_location(rdev, mc);
d2800ee5
AD
3832}
3833
3834static int si_mc_init(struct radeon_device *rdev)
3835{
3836 u32 tmp;
3837 int chansize, numchan;
3838
3839 /* Get VRAM informations */
3840 rdev->mc.vram_is_ddr = true;
3841 tmp = RREG32(MC_ARB_RAMCFG);
3842 if (tmp & CHANSIZE_OVERRIDE) {
3843 chansize = 16;
3844 } else if (tmp & CHANSIZE_MASK) {
3845 chansize = 64;
3846 } else {
3847 chansize = 32;
3848 }
3849 tmp = RREG32(MC_SHARED_CHMAP);
3850 switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
3851 case 0:
3852 default:
3853 numchan = 1;
3854 break;
3855 case 1:
3856 numchan = 2;
3857 break;
3858 case 2:
3859 numchan = 4;
3860 break;
3861 case 3:
3862 numchan = 8;
3863 break;
3864 case 4:
3865 numchan = 3;
3866 break;
3867 case 5:
3868 numchan = 6;
3869 break;
3870 case 6:
3871 numchan = 10;
3872 break;
3873 case 7:
3874 numchan = 12;
3875 break;
3876 case 8:
3877 numchan = 16;
3878 break;
3879 }
3880 rdev->mc.vram_width = numchan * chansize;
3881 /* Could aper size report 0 ? */
3882 rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
3883 rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
3884 /* size in MB on si */
fc986034
NOS
3885 rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE) * 1024ULL * 1024ULL;
3886 rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE) * 1024ULL * 1024ULL;
d2800ee5
AD
3887 rdev->mc.visible_vram_size = rdev->mc.aper_size;
3888 si_vram_gtt_location(rdev, &rdev->mc);
3889 radeon_update_bandwidth_info(rdev);
3890
3891 return 0;
3892}
3893
3894/*
3895 * GART
3896 */
3897void si_pcie_gart_tlb_flush(struct radeon_device *rdev)
3898{
3899 /* flush hdp cache */
3900 WREG32(HDP_MEM_COHERENCY_FLUSH_CNTL, 0x1);
3901
3902 /* bits 0-15 are the VM contexts0-15 */
3903 WREG32(VM_INVALIDATE_REQUEST, 1);
3904}
3905
1109ca09 3906static int si_pcie_gart_enable(struct radeon_device *rdev)
d2800ee5
AD
3907{
3908 int r, i;
3909
3910 if (rdev->gart.robj == NULL) {
3911 dev_err(rdev->dev, "No VRAM object for PCIE GART.\n");
3912 return -EINVAL;
3913 }
3914 r = radeon_gart_table_vram_pin(rdev);
3915 if (r)
3916 return r;
3917 radeon_gart_restore(rdev);
3918 /* Setup TLB control */
3919 WREG32(MC_VM_MX_L1_TLB_CNTL,
3920 (0xA << 7) |
3921 ENABLE_L1_TLB |
3922 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
3923 ENABLE_ADVANCED_DRIVER_MODEL |
3924 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
3925 /* Setup L2 cache */
3926 WREG32(VM_L2_CNTL, ENABLE_L2_CACHE |
3927 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
3928 ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
3929 EFFECTIVE_L2_QUEUE_SIZE(7) |
3930 CONTEXT1_IDENTITY_ACCESS_MODE(1));
3931 WREG32(VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS | INVALIDATE_L2_CACHE);
3932 WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
3933 L2_CACHE_BIGK_FRAGMENT_SIZE(0));
3934 /* setup context0 */
3935 WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);
3936 WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);
3937 WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12);
3938 WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
3939 (u32)(rdev->dummy_page.addr >> 12));
3940 WREG32(VM_CONTEXT0_CNTL2, 0);
3941 WREG32(VM_CONTEXT0_CNTL, (ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) |
3942 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT));
3943
3944 WREG32(0x15D4, 0);
3945 WREG32(0x15D8, 0);
3946 WREG32(0x15DC, 0);
3947
3948 /* empty context1-15 */
d2800ee5
AD
3949 /* set vm size, must be a multiple of 4 */
3950 WREG32(VM_CONTEXT1_PAGE_TABLE_START_ADDR, 0);
c21b328e 3951 WREG32(VM_CONTEXT1_PAGE_TABLE_END_ADDR, rdev->vm_manager.max_pfn);
23d4f1f2
AD
3952 /* Assign the pt base to something valid for now; the pts used for
3953 * the VMs are determined by the application and setup and assigned
3954 * on the fly in the vm part of radeon_gart.c
3955 */
d2800ee5
AD
3956 for (i = 1; i < 16; i++) {
3957 if (i < 8)
3958 WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (i << 2),
3959 rdev->gart.table_addr >> 12);
3960 else
3961 WREG32(VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((i - 8) << 2),
3962 rdev->gart.table_addr >> 12);
3963 }
3964
3965 /* enable context1-15 */
3966 WREG32(VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR,
3967 (u32)(rdev->dummy_page.addr >> 12));
ae133a11 3968 WREG32(VM_CONTEXT1_CNTL2, 4);
fa87e62d 3969 WREG32(VM_CONTEXT1_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(1) |
ae133a11
CK
3970 RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT |
3971 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT |
3972 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT |
3973 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT |
3974 PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT |
3975 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT |
3976 VALID_PROTECTION_FAULT_ENABLE_INTERRUPT |
3977 VALID_PROTECTION_FAULT_ENABLE_DEFAULT |
3978 READ_PROTECTION_FAULT_ENABLE_INTERRUPT |
3979 READ_PROTECTION_FAULT_ENABLE_DEFAULT |
3980 WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT |
3981 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT);
d2800ee5
AD
3982
3983 si_pcie_gart_tlb_flush(rdev);
3984 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
3985 (unsigned)(rdev->mc.gtt_size >> 20),
3986 (unsigned long long)rdev->gart.table_addr);
3987 rdev->gart.ready = true;
3988 return 0;
3989}
3990
1109ca09 3991static void si_pcie_gart_disable(struct radeon_device *rdev)
d2800ee5
AD
3992{
3993 /* Disable all tables */
3994 WREG32(VM_CONTEXT0_CNTL, 0);
3995 WREG32(VM_CONTEXT1_CNTL, 0);
3996 /* Setup TLB control */
3997 WREG32(MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE_NOT_IN_SYS |
3998 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
3999 /* Setup L2 cache */
4000 WREG32(VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
4001 ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
4002 EFFECTIVE_L2_QUEUE_SIZE(7) |
4003 CONTEXT1_IDENTITY_ACCESS_MODE(1));
4004 WREG32(VM_L2_CNTL2, 0);
4005 WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
4006 L2_CACHE_BIGK_FRAGMENT_SIZE(0));
4007 radeon_gart_table_vram_unpin(rdev);
4008}
4009
1109ca09 4010static void si_pcie_gart_fini(struct radeon_device *rdev)
d2800ee5
AD
4011{
4012 si_pcie_gart_disable(rdev);
4013 radeon_gart_table_vram_free(rdev);
4014 radeon_gart_fini(rdev);
4015}
4016
498dd8b3
AD
4017/* vm parser */
4018static bool si_vm_reg_valid(u32 reg)
4019{
4020 /* context regs are fine */
4021 if (reg >= 0x28000)
4022 return true;
4023
4024 /* check config regs */
4025 switch (reg) {
4026 case GRBM_GFX_INDEX:
f418b88a 4027 case CP_STRMOUT_CNTL:
498dd8b3
AD
4028 case VGT_VTX_VECT_EJECT_REG:
4029 case VGT_CACHE_INVALIDATION:
4030 case VGT_ESGS_RING_SIZE:
4031 case VGT_GSVS_RING_SIZE:
4032 case VGT_GS_VERTEX_REUSE:
4033 case VGT_PRIMITIVE_TYPE:
4034 case VGT_INDEX_TYPE:
4035 case VGT_NUM_INDICES:
4036 case VGT_NUM_INSTANCES:
4037 case VGT_TF_RING_SIZE:
4038 case VGT_HS_OFFCHIP_PARAM:
4039 case VGT_TF_MEMORY_BASE:
4040 case PA_CL_ENHANCE:
4041 case PA_SU_LINE_STIPPLE_VALUE:
4042 case PA_SC_LINE_STIPPLE_STATE:
4043 case PA_SC_ENHANCE:
4044 case SQC_CACHES:
4045 case SPI_STATIC_THREAD_MGMT_1:
4046 case SPI_STATIC_THREAD_MGMT_2:
4047 case SPI_STATIC_THREAD_MGMT_3:
4048 case SPI_PS_MAX_WAVE_ID:
4049 case SPI_CONFIG_CNTL:
4050 case SPI_CONFIG_CNTL_1:
4051 case TA_CNTL_AUX:
4052 return true;
4053 default:
4054 DRM_ERROR("Invalid register 0x%x in CS\n", reg);
4055 return false;
4056 }
4057}
4058
4059static int si_vm_packet3_ce_check(struct radeon_device *rdev,
4060 u32 *ib, struct radeon_cs_packet *pkt)
4061{
4062 switch (pkt->opcode) {
4063 case PACKET3_NOP:
4064 case PACKET3_SET_BASE:
4065 case PACKET3_SET_CE_DE_COUNTERS:
4066 case PACKET3_LOAD_CONST_RAM:
4067 case PACKET3_WRITE_CONST_RAM:
4068 case PACKET3_WRITE_CONST_RAM_OFFSET:
4069 case PACKET3_DUMP_CONST_RAM:
4070 case PACKET3_INCREMENT_CE_COUNTER:
4071 case PACKET3_WAIT_ON_DE_COUNTER:
4072 case PACKET3_CE_WRITE:
4073 break;
4074 default:
4075 DRM_ERROR("Invalid CE packet3: 0x%x\n", pkt->opcode);
4076 return -EINVAL;
4077 }
4078 return 0;
4079}
4080
e5b9e750
TS
4081static int si_vm_packet3_cp_dma_check(u32 *ib, u32 idx)
4082{
4083 u32 start_reg, reg, i;
4084 u32 command = ib[idx + 4];
4085 u32 info = ib[idx + 1];
4086 u32 idx_value = ib[idx];
4087 if (command & PACKET3_CP_DMA_CMD_SAS) {
4088 /* src address space is register */
4089 if (((info & 0x60000000) >> 29) == 0) {
4090 start_reg = idx_value << 2;
4091 if (command & PACKET3_CP_DMA_CMD_SAIC) {
4092 reg = start_reg;
4093 if (!si_vm_reg_valid(reg)) {
4094 DRM_ERROR("CP DMA Bad SRC register\n");
4095 return -EINVAL;
4096 }
4097 } else {
4098 for (i = 0; i < (command & 0x1fffff); i++) {
4099 reg = start_reg + (4 * i);
4100 if (!si_vm_reg_valid(reg)) {
4101 DRM_ERROR("CP DMA Bad SRC register\n");
4102 return -EINVAL;
4103 }
4104 }
4105 }
4106 }
4107 }
4108 if (command & PACKET3_CP_DMA_CMD_DAS) {
4109 /* dst address space is register */
4110 if (((info & 0x00300000) >> 20) == 0) {
4111 start_reg = ib[idx + 2];
4112 if (command & PACKET3_CP_DMA_CMD_DAIC) {
4113 reg = start_reg;
4114 if (!si_vm_reg_valid(reg)) {
4115 DRM_ERROR("CP DMA Bad DST register\n");
4116 return -EINVAL;
4117 }
4118 } else {
4119 for (i = 0; i < (command & 0x1fffff); i++) {
4120 reg = start_reg + (4 * i);
4121 if (!si_vm_reg_valid(reg)) {
4122 DRM_ERROR("CP DMA Bad DST register\n");
4123 return -EINVAL;
4124 }
4125 }
4126 }
4127 }
4128 }
4129 return 0;
4130}
4131
498dd8b3
AD
4132static int si_vm_packet3_gfx_check(struct radeon_device *rdev,
4133 u32 *ib, struct radeon_cs_packet *pkt)
4134{
e5b9e750 4135 int r;
498dd8b3
AD
4136 u32 idx = pkt->idx + 1;
4137 u32 idx_value = ib[idx];
4138 u32 start_reg, end_reg, reg, i;
4139
4140 switch (pkt->opcode) {
4141 case PACKET3_NOP:
4142 case PACKET3_SET_BASE:
4143 case PACKET3_CLEAR_STATE:
4144 case PACKET3_INDEX_BUFFER_SIZE:
4145 case PACKET3_DISPATCH_DIRECT:
4146 case PACKET3_DISPATCH_INDIRECT:
4147 case PACKET3_ALLOC_GDS:
4148 case PACKET3_WRITE_GDS_RAM:
4149 case PACKET3_ATOMIC_GDS:
4150 case PACKET3_ATOMIC:
4151 case PACKET3_OCCLUSION_QUERY:
4152 case PACKET3_SET_PREDICATION:
4153 case PACKET3_COND_EXEC:
4154 case PACKET3_PRED_EXEC:
4155 case PACKET3_DRAW_INDIRECT:
4156 case PACKET3_DRAW_INDEX_INDIRECT:
4157 case PACKET3_INDEX_BASE:
4158 case PACKET3_DRAW_INDEX_2:
4159 case PACKET3_CONTEXT_CONTROL:
4160 case PACKET3_INDEX_TYPE:
4161 case PACKET3_DRAW_INDIRECT_MULTI:
4162 case PACKET3_DRAW_INDEX_AUTO:
4163 case PACKET3_DRAW_INDEX_IMMD:
4164 case PACKET3_NUM_INSTANCES:
4165 case PACKET3_DRAW_INDEX_MULTI_AUTO:
4166 case PACKET3_STRMOUT_BUFFER_UPDATE:
4167 case PACKET3_DRAW_INDEX_OFFSET_2:
4168 case PACKET3_DRAW_INDEX_MULTI_ELEMENT:
4169 case PACKET3_DRAW_INDEX_INDIRECT_MULTI:
4170 case PACKET3_MPEG_INDEX:
4171 case PACKET3_WAIT_REG_MEM:
4172 case PACKET3_MEM_WRITE:
4173 case PACKET3_PFP_SYNC_ME:
4174 case PACKET3_SURFACE_SYNC:
4175 case PACKET3_EVENT_WRITE:
4176 case PACKET3_EVENT_WRITE_EOP:
4177 case PACKET3_EVENT_WRITE_EOS:
4178 case PACKET3_SET_CONTEXT_REG:
4179 case PACKET3_SET_CONTEXT_REG_INDIRECT:
4180 case PACKET3_SET_SH_REG:
4181 case PACKET3_SET_SH_REG_OFFSET:
4182 case PACKET3_INCREMENT_DE_COUNTER:
4183 case PACKET3_WAIT_ON_CE_COUNTER:
4184 case PACKET3_WAIT_ON_AVAIL_BUFFER:
4185 case PACKET3_ME_WRITE:
4186 break;
4187 case PACKET3_COPY_DATA:
4188 if ((idx_value & 0xf00) == 0) {
4189 reg = ib[idx + 3] * 4;
4190 if (!si_vm_reg_valid(reg))
4191 return -EINVAL;
4192 }
4193 break;
4194 case PACKET3_WRITE_DATA:
4195 if ((idx_value & 0xf00) == 0) {
4196 start_reg = ib[idx + 1] * 4;
4197 if (idx_value & 0x10000) {
4198 if (!si_vm_reg_valid(start_reg))
4199 return -EINVAL;
4200 } else {
4201 for (i = 0; i < (pkt->count - 2); i++) {
4202 reg = start_reg + (4 * i);
4203 if (!si_vm_reg_valid(reg))
4204 return -EINVAL;
4205 }
4206 }
4207 }
4208 break;
4209 case PACKET3_COND_WRITE:
4210 if (idx_value & 0x100) {
4211 reg = ib[idx + 5] * 4;
4212 if (!si_vm_reg_valid(reg))
4213 return -EINVAL;
4214 }
4215 break;
4216 case PACKET3_COPY_DW:
4217 if (idx_value & 0x2) {
4218 reg = ib[idx + 3] * 4;
4219 if (!si_vm_reg_valid(reg))
4220 return -EINVAL;
4221 }
4222 break;
4223 case PACKET3_SET_CONFIG_REG:
4224 start_reg = (idx_value << 2) + PACKET3_SET_CONFIG_REG_START;
4225 end_reg = 4 * pkt->count + start_reg - 4;
4226 if ((start_reg < PACKET3_SET_CONFIG_REG_START) ||
4227 (start_reg >= PACKET3_SET_CONFIG_REG_END) ||
4228 (end_reg >= PACKET3_SET_CONFIG_REG_END)) {
4229 DRM_ERROR("bad PACKET3_SET_CONFIG_REG\n");
4230 return -EINVAL;
4231 }
4232 for (i = 0; i < pkt->count; i++) {
4233 reg = start_reg + (4 * i);
4234 if (!si_vm_reg_valid(reg))
4235 return -EINVAL;
4236 }
4237 break;
5aa709be 4238 case PACKET3_CP_DMA:
e5b9e750
TS
4239 r = si_vm_packet3_cp_dma_check(ib, idx);
4240 if (r)
4241 return r;
5aa709be 4242 break;
498dd8b3
AD
4243 default:
4244 DRM_ERROR("Invalid GFX packet3: 0x%x\n", pkt->opcode);
4245 return -EINVAL;
4246 }
4247 return 0;
4248}
4249
4250static int si_vm_packet3_compute_check(struct radeon_device *rdev,
4251 u32 *ib, struct radeon_cs_packet *pkt)
4252{
e5b9e750 4253 int r;
498dd8b3
AD
4254 u32 idx = pkt->idx + 1;
4255 u32 idx_value = ib[idx];
4256 u32 start_reg, reg, i;
4257
4258 switch (pkt->opcode) {
4259 case PACKET3_NOP:
4260 case PACKET3_SET_BASE:
4261 case PACKET3_CLEAR_STATE:
4262 case PACKET3_DISPATCH_DIRECT:
4263 case PACKET3_DISPATCH_INDIRECT:
4264 case PACKET3_ALLOC_GDS:
4265 case PACKET3_WRITE_GDS_RAM:
4266 case PACKET3_ATOMIC_GDS:
4267 case PACKET3_ATOMIC:
4268 case PACKET3_OCCLUSION_QUERY:
4269 case PACKET3_SET_PREDICATION:
4270 case PACKET3_COND_EXEC:
4271 case PACKET3_PRED_EXEC:
4272 case PACKET3_CONTEXT_CONTROL:
4273 case PACKET3_STRMOUT_BUFFER_UPDATE:
4274 case PACKET3_WAIT_REG_MEM:
4275 case PACKET3_MEM_WRITE:
4276 case PACKET3_PFP_SYNC_ME:
4277 case PACKET3_SURFACE_SYNC:
4278 case PACKET3_EVENT_WRITE:
4279 case PACKET3_EVENT_WRITE_EOP:
4280 case PACKET3_EVENT_WRITE_EOS:
4281 case PACKET3_SET_CONTEXT_REG:
4282 case PACKET3_SET_CONTEXT_REG_INDIRECT:
4283 case PACKET3_SET_SH_REG:
4284 case PACKET3_SET_SH_REG_OFFSET:
4285 case PACKET3_INCREMENT_DE_COUNTER:
4286 case PACKET3_WAIT_ON_CE_COUNTER:
4287 case PACKET3_WAIT_ON_AVAIL_BUFFER:
4288 case PACKET3_ME_WRITE:
4289 break;
4290 case PACKET3_COPY_DATA:
4291 if ((idx_value & 0xf00) == 0) {
4292 reg = ib[idx + 3] * 4;
4293 if (!si_vm_reg_valid(reg))
4294 return -EINVAL;
4295 }
4296 break;
4297 case PACKET3_WRITE_DATA:
4298 if ((idx_value & 0xf00) == 0) {
4299 start_reg = ib[idx + 1] * 4;
4300 if (idx_value & 0x10000) {
4301 if (!si_vm_reg_valid(start_reg))
4302 return -EINVAL;
4303 } else {
4304 for (i = 0; i < (pkt->count - 2); i++) {
4305 reg = start_reg + (4 * i);
4306 if (!si_vm_reg_valid(reg))
4307 return -EINVAL;
4308 }
4309 }
4310 }
4311 break;
4312 case PACKET3_COND_WRITE:
4313 if (idx_value & 0x100) {
4314 reg = ib[idx + 5] * 4;
4315 if (!si_vm_reg_valid(reg))
4316 return -EINVAL;
4317 }
4318 break;
4319 case PACKET3_COPY_DW:
4320 if (idx_value & 0x2) {
4321 reg = ib[idx + 3] * 4;
4322 if (!si_vm_reg_valid(reg))
4323 return -EINVAL;
4324 }
4325 break;
e5b9e750
TS
4326 case PACKET3_CP_DMA:
4327 r = si_vm_packet3_cp_dma_check(ib, idx);
4328 if (r)
4329 return r;
4330 break;
498dd8b3
AD
4331 default:
4332 DRM_ERROR("Invalid Compute packet3: 0x%x\n", pkt->opcode);
4333 return -EINVAL;
4334 }
4335 return 0;
4336}
4337
4338int si_ib_parse(struct radeon_device *rdev, struct radeon_ib *ib)
4339{
4340 int ret = 0;
4341 u32 idx = 0;
4342 struct radeon_cs_packet pkt;
4343
4344 do {
4345 pkt.idx = idx;
4e872ae2
IH
4346 pkt.type = RADEON_CP_PACKET_GET_TYPE(ib->ptr[idx]);
4347 pkt.count = RADEON_CP_PACKET_GET_COUNT(ib->ptr[idx]);
498dd8b3
AD
4348 pkt.one_reg_wr = 0;
4349 switch (pkt.type) {
4e872ae2 4350 case RADEON_PACKET_TYPE0:
498dd8b3
AD
4351 dev_err(rdev->dev, "Packet0 not allowed!\n");
4352 ret = -EINVAL;
4353 break;
4e872ae2 4354 case RADEON_PACKET_TYPE2:
498dd8b3
AD
4355 idx += 1;
4356 break;
4e872ae2
IH
4357 case RADEON_PACKET_TYPE3:
4358 pkt.opcode = RADEON_CP_PACKET3_GET_OPCODE(ib->ptr[idx]);
498dd8b3
AD
4359 if (ib->is_const_ib)
4360 ret = si_vm_packet3_ce_check(rdev, ib->ptr, &pkt);
4361 else {
876dc9f3 4362 switch (ib->ring) {
498dd8b3
AD
4363 case RADEON_RING_TYPE_GFX_INDEX:
4364 ret = si_vm_packet3_gfx_check(rdev, ib->ptr, &pkt);
4365 break;
4366 case CAYMAN_RING_TYPE_CP1_INDEX:
4367 case CAYMAN_RING_TYPE_CP2_INDEX:
4368 ret = si_vm_packet3_compute_check(rdev, ib->ptr, &pkt);
4369 break;
4370 default:
876dc9f3 4371 dev_err(rdev->dev, "Non-PM4 ring %d !\n", ib->ring);
498dd8b3
AD
4372 ret = -EINVAL;
4373 break;
4374 }
4375 }
4376 idx += pkt.count + 2;
4377 break;
4378 default:
4379 dev_err(rdev->dev, "Unknown packet type %d !\n", pkt.type);
4380 ret = -EINVAL;
4381 break;
4382 }
4383 if (ret)
4384 break;
4385 } while (idx < ib->length_dw);
4386
4387 return ret;
4388}
4389
d2800ee5
AD
4390/*
4391 * vm
4392 */
4393int si_vm_init(struct radeon_device *rdev)
4394{
4395 /* number of VMs */
4396 rdev->vm_manager.nvm = 16;
4397 /* base offset of vram pages */
4398 rdev->vm_manager.vram_base_offset = 0;
4399
4400 return 0;
4401}
4402
4403void si_vm_fini(struct radeon_device *rdev)
4404{
4405}
4406
fbf6dc7a
AD
4407/**
4408 * si_vm_decode_fault - print human readable fault info
4409 *
4410 * @rdev: radeon_device pointer
4411 * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value
4412 * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value
4413 *
4414 * Print human readable fault information (SI).
4415 */
4416static void si_vm_decode_fault(struct radeon_device *rdev,
4417 u32 status, u32 addr)
4418{
4419 u32 mc_id = (status & MEMORY_CLIENT_ID_MASK) >> MEMORY_CLIENT_ID_SHIFT;
4420 u32 vmid = (status & FAULT_VMID_MASK) >> FAULT_VMID_SHIFT;
4421 u32 protections = (status & PROTECTIONS_MASK) >> PROTECTIONS_SHIFT;
4422 char *block;
4423
4424 if (rdev->family == CHIP_TAHITI) {
4425 switch (mc_id) {
4426 case 160:
4427 case 144:
4428 case 96:
4429 case 80:
4430 case 224:
4431 case 208:
4432 case 32:
4433 case 16:
4434 block = "CB";
4435 break;
4436 case 161:
4437 case 145:
4438 case 97:
4439 case 81:
4440 case 225:
4441 case 209:
4442 case 33:
4443 case 17:
4444 block = "CB_FMASK";
4445 break;
4446 case 162:
4447 case 146:
4448 case 98:
4449 case 82:
4450 case 226:
4451 case 210:
4452 case 34:
4453 case 18:
4454 block = "CB_CMASK";
4455 break;
4456 case 163:
4457 case 147:
4458 case 99:
4459 case 83:
4460 case 227:
4461 case 211:
4462 case 35:
4463 case 19:
4464 block = "CB_IMMED";
4465 break;
4466 case 164:
4467 case 148:
4468 case 100:
4469 case 84:
4470 case 228:
4471 case 212:
4472 case 36:
4473 case 20:
4474 block = "DB";
4475 break;
4476 case 165:
4477 case 149:
4478 case 101:
4479 case 85:
4480 case 229:
4481 case 213:
4482 case 37:
4483 case 21:
4484 block = "DB_HTILE";
4485 break;
4486 case 167:
4487 case 151:
4488 case 103:
4489 case 87:
4490 case 231:
4491 case 215:
4492 case 39:
4493 case 23:
4494 block = "DB_STEN";
4495 break;
4496 case 72:
4497 case 68:
4498 case 64:
4499 case 8:
4500 case 4:
4501 case 0:
4502 case 136:
4503 case 132:
4504 case 128:
4505 case 200:
4506 case 196:
4507 case 192:
4508 block = "TC";
4509 break;
4510 case 112:
4511 case 48:
4512 block = "CP";
4513 break;
4514 case 49:
4515 case 177:
4516 case 50:
4517 case 178:
4518 block = "SH";
4519 break;
4520 case 53:
4521 case 190:
4522 block = "VGT";
4523 break;
4524 case 117:
4525 block = "IH";
4526 break;
4527 case 51:
4528 case 115:
4529 block = "RLC";
4530 break;
4531 case 119:
4532 case 183:
4533 block = "DMA0";
4534 break;
4535 case 61:
4536 block = "DMA1";
4537 break;
4538 case 248:
4539 case 120:
4540 block = "HDP";
4541 break;
4542 default:
4543 block = "unknown";
4544 break;
4545 }
4546 } else {
4547 switch (mc_id) {
4548 case 32:
4549 case 16:
4550 case 96:
4551 case 80:
4552 case 160:
4553 case 144:
4554 case 224:
4555 case 208:
4556 block = "CB";
4557 break;
4558 case 33:
4559 case 17:
4560 case 97:
4561 case 81:
4562 case 161:
4563 case 145:
4564 case 225:
4565 case 209:
4566 block = "CB_FMASK";
4567 break;
4568 case 34:
4569 case 18:
4570 case 98:
4571 case 82:
4572 case 162:
4573 case 146:
4574 case 226:
4575 case 210:
4576 block = "CB_CMASK";
4577 break;
4578 case 35:
4579 case 19:
4580 case 99:
4581 case 83:
4582 case 163:
4583 case 147:
4584 case 227:
4585 case 211:
4586 block = "CB_IMMED";
4587 break;
4588 case 36:
4589 case 20:
4590 case 100:
4591 case 84:
4592 case 164:
4593 case 148:
4594 case 228:
4595 case 212:
4596 block = "DB";
4597 break;
4598 case 37:
4599 case 21:
4600 case 101:
4601 case 85:
4602 case 165:
4603 case 149:
4604 case 229:
4605 case 213:
4606 block = "DB_HTILE";
4607 break;
4608 case 39:
4609 case 23:
4610 case 103:
4611 case 87:
4612 case 167:
4613 case 151:
4614 case 231:
4615 case 215:
4616 block = "DB_STEN";
4617 break;
4618 case 72:
4619 case 68:
4620 case 8:
4621 case 4:
4622 case 136:
4623 case 132:
4624 case 200:
4625 case 196:
4626 block = "TC";
4627 break;
4628 case 112:
4629 case 48:
4630 block = "CP";
4631 break;
4632 case 49:
4633 case 177:
4634 case 50:
4635 case 178:
4636 block = "SH";
4637 break;
4638 case 53:
4639 block = "VGT";
4640 break;
4641 case 117:
4642 block = "IH";
4643 break;
4644 case 51:
4645 case 115:
4646 block = "RLC";
4647 break;
4648 case 119:
4649 case 183:
4650 block = "DMA0";
4651 break;
4652 case 61:
4653 block = "DMA1";
4654 break;
4655 case 248:
4656 case 120:
4657 block = "HDP";
4658 break;
4659 default:
4660 block = "unknown";
4661 break;
4662 }
4663 }
4664
4665 printk("VM fault (0x%02x, vmid %d) at page %u, %s from %s (%d)\n",
4666 protections, vmid, addr,
4667 (status & MEMORY_CLIENT_RW_MASK) ? "write" : "read",
4668 block, mc_id);
4669}
4670
498522b4 4671void si_vm_flush(struct radeon_device *rdev, int ridx, struct radeon_vm *vm)
d2800ee5 4672{
498522b4 4673 struct radeon_ring *ring = &rdev->ring[ridx];
d2800ee5 4674
ee60e29f 4675 if (vm == NULL)
d2800ee5
AD
4676 return;
4677
76c44f2c
AD
4678 /* write new base address */
4679 radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4680 radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
4681 WRITE_DATA_DST_SEL(0)));
4682
ee60e29f 4683 if (vm->id < 8) {
76c44f2c
AD
4684 radeon_ring_write(ring,
4685 (VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm->id << 2)) >> 2);
ee60e29f 4686 } else {
76c44f2c
AD
4687 radeon_ring_write(ring,
4688 (VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((vm->id - 8) << 2)) >> 2);
ee60e29f 4689 }
76c44f2c 4690 radeon_ring_write(ring, 0);
fa87e62d 4691 radeon_ring_write(ring, vm->pd_gpu_addr >> 12);
ee60e29f 4692
d2800ee5 4693 /* flush hdp cache */
76c44f2c
AD
4694 radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4695 radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
4696 WRITE_DATA_DST_SEL(0)));
4697 radeon_ring_write(ring, HDP_MEM_COHERENCY_FLUSH_CNTL >> 2);
4698 radeon_ring_write(ring, 0);
ee60e29f
CK
4699 radeon_ring_write(ring, 0x1);
4700
d2800ee5 4701 /* bits 0-15 are the VM contexts0-15 */
76c44f2c
AD
4702 radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4703 radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
4704 WRITE_DATA_DST_SEL(0)));
4705 radeon_ring_write(ring, VM_INVALIDATE_REQUEST >> 2);
4706 radeon_ring_write(ring, 0);
498522b4 4707 radeon_ring_write(ring, 1 << vm->id);
58f8cf56
CK
4708
4709 /* sync PFP to ME, otherwise we might get invalid PFP reads */
4710 radeon_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
4711 radeon_ring_write(ring, 0x0);
d2800ee5
AD
4712}
4713
f8f84ac5
AD
4714/*
4715 * Power and clock gating
4716 */
4717static void si_wait_for_rlc_serdes(struct radeon_device *rdev)
4718{
4719 int i;
4720
4721 for (i = 0; i < rdev->usec_timeout; i++) {
4722 if (RREG32(RLC_SERDES_MASTER_BUSY_0) == 0)
4723 break;
4724 udelay(1);
4725 }
4726
4727 for (i = 0; i < rdev->usec_timeout; i++) {
4728 if (RREG32(RLC_SERDES_MASTER_BUSY_1) == 0)
4729 break;
4730 udelay(1);
4731 }
4732}
4733
4734static void si_enable_gui_idle_interrupt(struct radeon_device *rdev,
4735 bool enable)
4736{
4737 u32 tmp = RREG32(CP_INT_CNTL_RING0);
4738 u32 mask;
4739 int i;
4740
4741 if (enable)
4742 tmp |= (CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
4743 else
4744 tmp &= ~(CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
4745 WREG32(CP_INT_CNTL_RING0, tmp);
4746
4747 if (!enable) {
4748 /* read a gfx register */
4749 tmp = RREG32(DB_DEPTH_INFO);
4750
4751 mask = RLC_BUSY_STATUS | GFX_POWER_STATUS | GFX_CLOCK_STATUS | GFX_LS_STATUS;
4752 for (i = 0; i < rdev->usec_timeout; i++) {
4753 if ((RREG32(RLC_STAT) & mask) == (GFX_CLOCK_STATUS | GFX_POWER_STATUS))
4754 break;
4755 udelay(1);
4756 }
4757 }
4758}
4759
4760static void si_set_uvd_dcm(struct radeon_device *rdev,
4761 bool sw_mode)
4762{
4763 u32 tmp, tmp2;
4764
4765 tmp = RREG32(UVD_CGC_CTRL);
4766 tmp &= ~(CLK_OD_MASK | CG_DT_MASK);
4767 tmp |= DCM | CG_DT(1) | CLK_OD(4);
4768
4769 if (sw_mode) {
4770 tmp &= ~0x7ffff800;
4771 tmp2 = DYN_OR_EN | DYN_RR_EN | G_DIV_ID(7);
4772 } else {
4773 tmp |= 0x7ffff800;
4774 tmp2 = 0;
4775 }
4776
4777 WREG32(UVD_CGC_CTRL, tmp);
4778 WREG32_UVD_CTX(UVD_CGC_CTRL2, tmp2);
4779}
4780
22c775ce 4781void si_init_uvd_internal_cg(struct radeon_device *rdev)
f8f84ac5
AD
4782{
4783 bool hw_mode = true;
4784
4785 if (hw_mode) {
4786 si_set_uvd_dcm(rdev, false);
4787 } else {
4788 u32 tmp = RREG32(UVD_CGC_CTRL);
4789 tmp &= ~DCM;
4790 WREG32(UVD_CGC_CTRL, tmp);
4791 }
4792}
4793
4794static u32 si_halt_rlc(struct radeon_device *rdev)
4795{
4796 u32 data, orig;
4797
4798 orig = data = RREG32(RLC_CNTL);
4799
4800 if (data & RLC_ENABLE) {
4801 data &= ~RLC_ENABLE;
4802 WREG32(RLC_CNTL, data);
4803
4804 si_wait_for_rlc_serdes(rdev);
4805 }
4806
4807 return orig;
4808}
4809
4810static void si_update_rlc(struct radeon_device *rdev, u32 rlc)
4811{
4812 u32 tmp;
4813
4814 tmp = RREG32(RLC_CNTL);
4815 if (tmp != rlc)
4816 WREG32(RLC_CNTL, rlc);
4817}
4818
4819static void si_enable_dma_pg(struct radeon_device *rdev, bool enable)
4820{
4821 u32 data, orig;
4822
4823 orig = data = RREG32(DMA_PG);
e16866ec 4824 if (enable && (rdev->pg_flags & RADEON_PG_SUPPORT_SDMA))
f8f84ac5
AD
4825 data |= PG_CNTL_ENABLE;
4826 else
4827 data &= ~PG_CNTL_ENABLE;
4828 if (orig != data)
4829 WREG32(DMA_PG, data);
4830}
4831
4832static void si_init_dma_pg(struct radeon_device *rdev)
4833{
4834 u32 tmp;
4835
4836 WREG32(DMA_PGFSM_WRITE, 0x00002000);
4837 WREG32(DMA_PGFSM_CONFIG, 0x100010ff);
4838
4839 for (tmp = 0; tmp < 5; tmp++)
4840 WREG32(DMA_PGFSM_WRITE, 0);
4841}
4842
4843static void si_enable_gfx_cgpg(struct radeon_device *rdev,
4844 bool enable)
4845{
4846 u32 tmp;
4847
2b19d17f 4848 if (enable && (rdev->pg_flags & RADEON_PG_SUPPORT_GFX_PG)) {
f8f84ac5
AD
4849 tmp = RLC_PUD(0x10) | RLC_PDD(0x10) | RLC_TTPD(0x10) | RLC_MSD(0x10);
4850 WREG32(RLC_TTOP_D, tmp);
4851
4852 tmp = RREG32(RLC_PG_CNTL);
4853 tmp |= GFX_PG_ENABLE;
4854 WREG32(RLC_PG_CNTL, tmp);
4855
4856 tmp = RREG32(RLC_AUTO_PG_CTRL);
4857 tmp |= AUTO_PG_EN;
4858 WREG32(RLC_AUTO_PG_CTRL, tmp);
4859 } else {
4860 tmp = RREG32(RLC_AUTO_PG_CTRL);
4861 tmp &= ~AUTO_PG_EN;
4862 WREG32(RLC_AUTO_PG_CTRL, tmp);
4863
4864 tmp = RREG32(DB_RENDER_CONTROL);
4865 }
4866}
4867
4868static void si_init_gfx_cgpg(struct radeon_device *rdev)
4869{
4870 u32 tmp;
4871
4872 WREG32(RLC_SAVE_AND_RESTORE_BASE, rdev->rlc.save_restore_gpu_addr >> 8);
4873
4874 tmp = RREG32(RLC_PG_CNTL);
4875 tmp |= GFX_PG_SRC;
4876 WREG32(RLC_PG_CNTL, tmp);
4877
4878 WREG32(RLC_CLEAR_STATE_RESTORE_BASE, rdev->rlc.clear_state_gpu_addr >> 8);
4879
4880 tmp = RREG32(RLC_AUTO_PG_CTRL);
4881
4882 tmp &= ~GRBM_REG_SGIT_MASK;
4883 tmp |= GRBM_REG_SGIT(0x700);
4884 tmp &= ~PG_AFTER_GRBM_REG_ST_MASK;
4885 WREG32(RLC_AUTO_PG_CTRL, tmp);
4886}
4887
ba19031a 4888static u32 si_get_cu_active_bitmap(struct radeon_device *rdev, u32 se, u32 sh)
f8f84ac5
AD
4889{
4890 u32 mask = 0, tmp, tmp1;
4891 int i;
4892
4893 si_select_se_sh(rdev, se, sh);
4894 tmp = RREG32(CC_GC_SHADER_ARRAY_CONFIG);
4895 tmp1 = RREG32(GC_USER_SHADER_ARRAY_CONFIG);
4896 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
4897
4898 tmp &= 0xffff0000;
4899
4900 tmp |= tmp1;
4901 tmp >>= 16;
4902
4903 for (i = 0; i < rdev->config.si.max_cu_per_sh; i ++) {
4904 mask <<= 1;
4905 mask |= 1;
4906 }
4907
4908 return (~tmp) & mask;
4909}
4910
4911static void si_init_ao_cu_mask(struct radeon_device *rdev)
4912{
4913 u32 i, j, k, active_cu_number = 0;
4914 u32 mask, counter, cu_bitmap;
4915 u32 tmp = 0;
4916
4917 for (i = 0; i < rdev->config.si.max_shader_engines; i++) {
4918 for (j = 0; j < rdev->config.si.max_sh_per_se; j++) {
4919 mask = 1;
4920 cu_bitmap = 0;
4921 counter = 0;
4922 for (k = 0; k < rdev->config.si.max_cu_per_sh; k++) {
ba19031a 4923 if (si_get_cu_active_bitmap(rdev, i, j) & mask) {
f8f84ac5
AD
4924 if (counter < 2)
4925 cu_bitmap |= mask;
4926 counter++;
4927 }
4928 mask <<= 1;
4929 }
4930
4931 active_cu_number += counter;
4932 tmp |= (cu_bitmap << (i * 16 + j * 8));
4933 }
4934 }
4935
4936 WREG32(RLC_PG_AO_CU_MASK, tmp);
4937
4938 tmp = RREG32(RLC_MAX_PG_CU);
4939 tmp &= ~MAX_PU_CU_MASK;
4940 tmp |= MAX_PU_CU(active_cu_number);
4941 WREG32(RLC_MAX_PG_CU, tmp);
4942}
4943
4944static void si_enable_cgcg(struct radeon_device *rdev,
4945 bool enable)
4946{
4947 u32 data, orig, tmp;
4948
4949 orig = data = RREG32(RLC_CGCG_CGLS_CTRL);
4950
e16866ec 4951 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_CGCG)) {
5594a558 4952 si_enable_gui_idle_interrupt(rdev, true);
f8f84ac5 4953
f8f84ac5
AD
4954 WREG32(RLC_GCPM_GENERAL_3, 0x00000080);
4955
4956 tmp = si_halt_rlc(rdev);
4957
4958 WREG32(RLC_SERDES_WR_MASTER_MASK_0, 0xffffffff);
4959 WREG32(RLC_SERDES_WR_MASTER_MASK_1, 0xffffffff);
4960 WREG32(RLC_SERDES_WR_CTRL, 0x00b000ff);
4961
4962 si_wait_for_rlc_serdes(rdev);
4963
4964 si_update_rlc(rdev, tmp);
4965
4966 WREG32(RLC_SERDES_WR_CTRL, 0x007000ff);
4967
4968 data |= CGCG_EN | CGLS_EN;
4969 } else {
5594a558
AD
4970 si_enable_gui_idle_interrupt(rdev, false);
4971
f8f84ac5
AD
4972 RREG32(CB_CGTT_SCLK_CTRL);
4973 RREG32(CB_CGTT_SCLK_CTRL);
4974 RREG32(CB_CGTT_SCLK_CTRL);
4975 RREG32(CB_CGTT_SCLK_CTRL);
4976
4977 data &= ~(CGCG_EN | CGLS_EN);
4978 }
4979
4980 if (orig != data)
4981 WREG32(RLC_CGCG_CGLS_CTRL, data);
4982}
4983
4984static void si_enable_mgcg(struct radeon_device *rdev,
4985 bool enable)
4986{
4987 u32 data, orig, tmp = 0;
4988
e16866ec 4989 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_MGCG)) {
f8f84ac5
AD
4990 orig = data = RREG32(CGTS_SM_CTRL_REG);
4991 data = 0x96940200;
4992 if (orig != data)
4993 WREG32(CGTS_SM_CTRL_REG, data);
4994
e16866ec
AD
4995 if (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_CP_LS) {
4996 orig = data = RREG32(CP_MEM_SLP_CNTL);
4997 data |= CP_MEM_LS_EN;
4998 if (orig != data)
4999 WREG32(CP_MEM_SLP_CNTL, data);
5000 }
f8f84ac5
AD
5001
5002 orig = data = RREG32(RLC_CGTT_MGCG_OVERRIDE);
5003 data &= 0xffffffc0;
5004 if (orig != data)
5005 WREG32(RLC_CGTT_MGCG_OVERRIDE, data);
5006
5007 tmp = si_halt_rlc(rdev);
5008
5009 WREG32(RLC_SERDES_WR_MASTER_MASK_0, 0xffffffff);
5010 WREG32(RLC_SERDES_WR_MASTER_MASK_1, 0xffffffff);
5011 WREG32(RLC_SERDES_WR_CTRL, 0x00d000ff);
5012
5013 si_update_rlc(rdev, tmp);
5014 } else {
5015 orig = data = RREG32(RLC_CGTT_MGCG_OVERRIDE);
5016 data |= 0x00000003;
5017 if (orig != data)
5018 WREG32(RLC_CGTT_MGCG_OVERRIDE, data);
5019
5020 data = RREG32(CP_MEM_SLP_CNTL);
5021 if (data & CP_MEM_LS_EN) {
5022 data &= ~CP_MEM_LS_EN;
5023 WREG32(CP_MEM_SLP_CNTL, data);
5024 }
5025 orig = data = RREG32(CGTS_SM_CTRL_REG);
5026 data |= LS_OVERRIDE | OVERRIDE;
5027 if (orig != data)
5028 WREG32(CGTS_SM_CTRL_REG, data);
5029
5030 tmp = si_halt_rlc(rdev);
5031
5032 WREG32(RLC_SERDES_WR_MASTER_MASK_0, 0xffffffff);
5033 WREG32(RLC_SERDES_WR_MASTER_MASK_1, 0xffffffff);
5034 WREG32(RLC_SERDES_WR_CTRL, 0x00e000ff);
5035
5036 si_update_rlc(rdev, tmp);
5037 }
5038}
5039
5040static void si_enable_uvd_mgcg(struct radeon_device *rdev,
5041 bool enable)
5042{
5043 u32 orig, data, tmp;
5044
e16866ec 5045 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_UVD_MGCG)) {
f8f84ac5
AD
5046 tmp = RREG32_UVD_CTX(UVD_CGC_MEM_CTRL);
5047 tmp |= 0x3fff;
5048 WREG32_UVD_CTX(UVD_CGC_MEM_CTRL, tmp);
5049
5050 orig = data = RREG32(UVD_CGC_CTRL);
5051 data |= DCM;
5052 if (orig != data)
5053 WREG32(UVD_CGC_CTRL, data);
5054
5055 WREG32_SMC(SMC_CG_IND_START + CG_CGTT_LOCAL_0, 0);
5056 WREG32_SMC(SMC_CG_IND_START + CG_CGTT_LOCAL_1, 0);
5057 } else {
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, 0xffffffff);
5068 WREG32_SMC(SMC_CG_IND_START + CG_CGTT_LOCAL_1, 0xffffffff);
5069 }
5070}
5071
5072static const u32 mc_cg_registers[] =
5073{
5074 MC_HUB_MISC_HUB_CG,
5075 MC_HUB_MISC_SIP_CG,
5076 MC_HUB_MISC_VM_CG,
5077 MC_XPB_CLK_GAT,
5078 ATC_MISC_CG,
5079 MC_CITF_MISC_WR_CG,
5080 MC_CITF_MISC_RD_CG,
5081 MC_CITF_MISC_VM_CG,
5082 VM_L2_CG,
5083};
5084
5085static void si_enable_mc_ls(struct radeon_device *rdev,
5086 bool enable)
5087{
5088 int i;
5089 u32 orig, data;
5090
5091 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) {
5092 orig = data = RREG32(mc_cg_registers[i]);
e16866ec 5093 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_MC_LS))
f8f84ac5
AD
5094 data |= MC_LS_ENABLE;
5095 else
5096 data &= ~MC_LS_ENABLE;
5097 if (data != orig)
5098 WREG32(mc_cg_registers[i], data);
5099 }
5100}
5101
e16866ec
AD
5102static void si_enable_mc_mgcg(struct radeon_device *rdev,
5103 bool enable)
f8f84ac5 5104{
e16866ec
AD
5105 int i;
5106 u32 orig, data;
5107
5108 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) {
5109 orig = data = RREG32(mc_cg_registers[i]);
5110 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_MC_MGCG))
5111 data |= MC_CG_ENABLE;
5112 else
5113 data &= ~MC_CG_ENABLE;
5114 if (data != orig)
5115 WREG32(mc_cg_registers[i], data);
f8f84ac5
AD
5116 }
5117}
5118
e16866ec
AD
5119static void si_enable_dma_mgcg(struct radeon_device *rdev,
5120 bool enable)
f8f84ac5 5121{
e16866ec
AD
5122 u32 orig, data, offset;
5123 int i;
f8f84ac5 5124
e16866ec
AD
5125 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_SDMA_MGCG)) {
5126 for (i = 0; i < 2; i++) {
5127 if (i == 0)
5128 offset = DMA0_REGISTER_OFFSET;
5129 else
5130 offset = DMA1_REGISTER_OFFSET;
5131 orig = data = RREG32(DMA_POWER_CNTL + offset);
5132 data &= ~MEM_POWER_OVERRIDE;
5133 if (data != orig)
5134 WREG32(DMA_POWER_CNTL + offset, data);
5135 WREG32(DMA_CLK_CTRL + offset, 0x00000100);
5136 }
5137 } else {
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
5148 orig = data = RREG32(DMA_CLK_CTRL + offset);
5149 data = 0xff000000;
5150 if (data != orig)
5151 WREG32(DMA_CLK_CTRL + offset, data);
5152 }
5153 }
f8f84ac5
AD
5154}
5155
e16866ec
AD
5156static void si_enable_bif_mgls(struct radeon_device *rdev,
5157 bool enable)
f8f84ac5 5158{
e16866ec 5159 u32 orig, data;
f8f84ac5 5160
e16866ec 5161 orig = data = RREG32_PCIE(PCIE_CNTL2);
f8f84ac5 5162
e16866ec
AD
5163 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_BIF_LS))
5164 data |= SLV_MEM_LS_EN | MST_MEM_LS_EN |
5165 REPLAY_MEM_LS_EN | SLV_MEM_AGGRESSIVE_LS_EN;
5166 else
5167 data &= ~(SLV_MEM_LS_EN | MST_MEM_LS_EN |
5168 REPLAY_MEM_LS_EN | SLV_MEM_AGGRESSIVE_LS_EN);
5169
5170 if (orig != data)
5171 WREG32_PCIE(PCIE_CNTL2, data);
f8f84ac5
AD
5172}
5173
e16866ec
AD
5174static void si_enable_hdp_mgcg(struct radeon_device *rdev,
5175 bool enable)
f8f84ac5 5176{
e16866ec 5177 u32 orig, data;
f8f84ac5 5178
e16866ec 5179 orig = data = RREG32(HDP_HOST_PATH_CNTL);
f8f84ac5 5180
e16866ec
AD
5181 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_HDP_MGCG))
5182 data &= ~CLOCK_GATING_DIS;
5183 else
5184 data |= CLOCK_GATING_DIS;
5185
5186 if (orig != data)
5187 WREG32(HDP_HOST_PATH_CNTL, data);
f8f84ac5
AD
5188}
5189
e16866ec
AD
5190static void si_enable_hdp_ls(struct radeon_device *rdev,
5191 bool enable)
347e7592 5192{
e16866ec 5193 u32 orig, data;
347e7592 5194
e16866ec 5195 orig = data = RREG32(HDP_MEM_POWER_LS);
347e7592 5196
e16866ec
AD
5197 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_HDP_LS))
5198 data |= HDP_LS_ENABLE;
5199 else
5200 data &= ~HDP_LS_ENABLE;
347e7592 5201
e16866ec
AD
5202 if (orig != data)
5203 WREG32(HDP_MEM_POWER_LS, data);
347e7592
AD
5204}
5205
e16866ec
AD
5206void si_update_cg(struct radeon_device *rdev,
5207 u32 block, bool enable)
347e7592 5208{
e16866ec 5209 if (block & RADEON_CG_BLOCK_GFX) {
811e4d58 5210 si_enable_gui_idle_interrupt(rdev, false);
e16866ec
AD
5211 /* order matters! */
5212 if (enable) {
5213 si_enable_mgcg(rdev, true);
5214 si_enable_cgcg(rdev, true);
5215 } else {
5216 si_enable_cgcg(rdev, false);
5217 si_enable_mgcg(rdev, false);
347e7592 5218 }
811e4d58 5219 si_enable_gui_idle_interrupt(rdev, true);
347e7592
AD
5220 }
5221
e16866ec
AD
5222 if (block & RADEON_CG_BLOCK_MC) {
5223 si_enable_mc_mgcg(rdev, enable);
5224 si_enable_mc_ls(rdev, enable);
347e7592 5225 }
e16866ec
AD
5226
5227 if (block & RADEON_CG_BLOCK_SDMA) {
5228 si_enable_dma_mgcg(rdev, enable);
347e7592
AD
5229 }
5230
e16866ec
AD
5231 if (block & RADEON_CG_BLOCK_BIF) {
5232 si_enable_bif_mgls(rdev, enable);
6d8cf000 5233 }
6d8cf000 5234
e16866ec
AD
5235 if (block & RADEON_CG_BLOCK_UVD) {
5236 if (rdev->has_uvd) {
5237 si_enable_uvd_mgcg(rdev, enable);
bd8cd539
AD
5238 }
5239 }
bd8cd539 5240
e16866ec
AD
5241 if (block & RADEON_CG_BLOCK_HDP) {
5242 si_enable_hdp_mgcg(rdev, enable);
5243 si_enable_hdp_ls(rdev, enable);
347e7592 5244 }
e16866ec 5245}
f8f84ac5
AD
5246
5247static void si_init_cg(struct radeon_device *rdev)
5248{
e16866ec
AD
5249 si_update_cg(rdev, (RADEON_CG_BLOCK_GFX |
5250 RADEON_CG_BLOCK_MC |
5251 RADEON_CG_BLOCK_SDMA |
5252 RADEON_CG_BLOCK_BIF |
5253 RADEON_CG_BLOCK_HDP), true);
b2d70917 5254 if (rdev->has_uvd) {
e16866ec 5255 si_update_cg(rdev, RADEON_CG_BLOCK_UVD, true);
f8f84ac5 5256 si_init_uvd_internal_cg(rdev);
347e7592 5257 }
f8f84ac5 5258}
bd8cd539 5259
f8f84ac5
AD
5260static void si_fini_cg(struct radeon_device *rdev)
5261{
0116e1ef 5262 if (rdev->has_uvd) {
e16866ec 5263 si_update_cg(rdev, RADEON_CG_BLOCK_UVD, false);
347e7592 5264 }
e16866ec
AD
5265 si_update_cg(rdev, (RADEON_CG_BLOCK_GFX |
5266 RADEON_CG_BLOCK_MC |
5267 RADEON_CG_BLOCK_SDMA |
5268 RADEON_CG_BLOCK_BIF |
5269 RADEON_CG_BLOCK_HDP), false);
5270}
5271
59a82d0e
AD
5272u32 si_get_csb_size(struct radeon_device *rdev)
5273{
5274 u32 count = 0;
5275 const struct cs_section_def *sect = NULL;
5276 const struct cs_extent_def *ext = NULL;
5277
5278 if (rdev->rlc.cs_data == NULL)
5279 return 0;
5280
5281 /* begin clear state */
5282 count += 2;
5283 /* context control state */
5284 count += 3;
5285
5286 for (sect = rdev->rlc.cs_data; sect->section != NULL; ++sect) {
5287 for (ext = sect->section; ext->extent != NULL; ++ext) {
5288 if (sect->id == SECT_CONTEXT)
5289 count += 2 + ext->reg_count;
5290 else
5291 return 0;
5292 }
bd8cd539 5293 }
59a82d0e
AD
5294 /* pa_sc_raster_config */
5295 count += 3;
5296 /* end clear state */
5297 count += 2;
5298 /* clear state */
5299 count += 2;
5300
5301 return count;
5302}
5303
5304void si_get_csb_buffer(struct radeon_device *rdev, volatile u32 *buffer)
5305{
5306 u32 count = 0, i;
5307 const struct cs_section_def *sect = NULL;
5308 const struct cs_extent_def *ext = NULL;
5309
5310 if (rdev->rlc.cs_data == NULL)
5311 return;
5312 if (buffer == NULL)
5313 return;
5314
6ba81e53
AD
5315 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0));
5316 buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
59a82d0e 5317
6ba81e53
AD
5318 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CONTEXT_CONTROL, 1));
5319 buffer[count++] = cpu_to_le32(0x80000000);
5320 buffer[count++] = cpu_to_le32(0x80000000);
59a82d0e
AD
5321
5322 for (sect = rdev->rlc.cs_data; sect->section != NULL; ++sect) {
5323 for (ext = sect->section; ext->extent != NULL; ++ext) {
5324 if (sect->id == SECT_CONTEXT) {
6ba81e53
AD
5325 buffer[count++] =
5326 cpu_to_le32(PACKET3(PACKET3_SET_CONTEXT_REG, ext->reg_count));
5327 buffer[count++] = cpu_to_le32(ext->reg_index - 0xa000);
59a82d0e 5328 for (i = 0; i < ext->reg_count; i++)
6ba81e53 5329 buffer[count++] = cpu_to_le32(ext->extent[i]);
59a82d0e
AD
5330 } else {
5331 return;
bd8cd539 5332 }
bd8cd539
AD
5333 }
5334 }
bd8cd539 5335
6ba81e53
AD
5336 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_SET_CONTEXT_REG, 1));
5337 buffer[count++] = cpu_to_le32(PA_SC_RASTER_CONFIG - PACKET3_SET_CONTEXT_REG_START);
59a82d0e
AD
5338 switch (rdev->family) {
5339 case CHIP_TAHITI:
5340 case CHIP_PITCAIRN:
6ba81e53 5341 buffer[count++] = cpu_to_le32(0x2a00126a);
59a82d0e
AD
5342 break;
5343 case CHIP_VERDE:
6ba81e53 5344 buffer[count++] = cpu_to_le32(0x0000124a);
59a82d0e
AD
5345 break;
5346 case CHIP_OLAND:
6ba81e53 5347 buffer[count++] = cpu_to_le32(0x00000082);
59a82d0e
AD
5348 break;
5349 case CHIP_HAINAN:
6ba81e53 5350 buffer[count++] = cpu_to_le32(0x00000000);
59a82d0e
AD
5351 break;
5352 default:
6ba81e53 5353 buffer[count++] = cpu_to_le32(0x00000000);
59a82d0e
AD
5354 break;
5355 }
5356
6ba81e53
AD
5357 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0));
5358 buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_END_CLEAR_STATE);
347e7592 5359
6ba81e53
AD
5360 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CLEAR_STATE, 0));
5361 buffer[count++] = cpu_to_le32(0);
59a82d0e
AD
5362}
5363
f8f84ac5
AD
5364static void si_init_pg(struct radeon_device *rdev)
5365{
0116e1ef
AD
5366 if (rdev->pg_flags) {
5367 if (rdev->pg_flags & RADEON_PG_SUPPORT_SDMA) {
5368 si_init_dma_pg(rdev);
0116e1ef 5369 }
f8f84ac5 5370 si_init_ao_cu_mask(rdev);
2b19d17f 5371 if (rdev->pg_flags & RADEON_PG_SUPPORT_GFX_PG) {
0116e1ef 5372 si_init_gfx_cgpg(rdev);
0116e1ef 5373 }
ca6ebb39
AD
5374 si_enable_dma_pg(rdev, true);
5375 si_enable_gfx_cgpg(rdev, true);
f8f84ac5
AD
5376 } else {
5377 WREG32(RLC_SAVE_AND_RESTORE_BASE, rdev->rlc.save_restore_gpu_addr >> 8);
5378 WREG32(RLC_CLEAR_STATE_RESTORE_BASE, rdev->rlc.clear_state_gpu_addr >> 8);
5379 }
5380}
5381
5382static void si_fini_pg(struct radeon_device *rdev)
5383{
0116e1ef 5384 if (rdev->pg_flags) {
ca6ebb39
AD
5385 si_enable_dma_pg(rdev, false);
5386 si_enable_gfx_cgpg(rdev, false);
f8f84ac5 5387 }
347e7592
AD
5388}
5389
347e7592
AD
5390/*
5391 * RLC
5392 */
866d83de 5393void si_rlc_reset(struct radeon_device *rdev)
d719cef3 5394{
f8f84ac5 5395 u32 tmp = RREG32(GRBM_SOFT_RESET);
d719cef3 5396
f8f84ac5
AD
5397 tmp |= SOFT_RESET_RLC;
5398 WREG32(GRBM_SOFT_RESET, tmp);
5399 udelay(50);
5400 tmp &= ~SOFT_RESET_RLC;
5401 WREG32(GRBM_SOFT_RESET, tmp);
5402 udelay(50);
d719cef3
AD
5403}
5404
347e7592
AD
5405static void si_rlc_stop(struct radeon_device *rdev)
5406{
5407 WREG32(RLC_CNTL, 0);
d719cef3
AD
5408
5409 si_enable_gui_idle_interrupt(rdev, false);
5410
5411 si_wait_for_rlc_serdes(rdev);
347e7592
AD
5412}
5413
5414static void si_rlc_start(struct radeon_device *rdev)
5415{
5416 WREG32(RLC_CNTL, RLC_ENABLE);
d719cef3
AD
5417
5418 si_enable_gui_idle_interrupt(rdev, true);
5419
5420 udelay(50);
5421}
5422
5423static bool si_lbpw_supported(struct radeon_device *rdev)
5424{
5425 u32 tmp;
5426
5427 /* Enable LBPW only for DDR3 */
5428 tmp = RREG32(MC_SEQ_MISC0);
5429 if ((tmp & 0xF0000000) == 0xB0000000)
5430 return true;
5431 return false;
5432}
5433
5434static void si_enable_lbpw(struct radeon_device *rdev, bool enable)
5435{
5436 u32 tmp;
5437
5438 tmp = RREG32(RLC_LB_CNTL);
5439 if (enable)
5440 tmp |= LOAD_BALANCE_ENABLE;
5441 else
5442 tmp &= ~LOAD_BALANCE_ENABLE;
5443 WREG32(RLC_LB_CNTL, tmp);
5444
5445 if (!enable) {
5446 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
5447 WREG32(SPI_LB_CU_MASK, 0x00ff);
5448 }
347e7592
AD
5449}
5450
5451static int si_rlc_resume(struct radeon_device *rdev)
5452{
5453 u32 i;
5454 const __be32 *fw_data;
5455
5456 if (!rdev->rlc_fw)
5457 return -EINVAL;
5458
5459 si_rlc_stop(rdev);
5460
f8f84ac5
AD
5461 si_rlc_reset(rdev);
5462
5463 si_init_pg(rdev);
5464
5465 si_init_cg(rdev);
5466
347e7592
AD
5467 WREG32(RLC_RL_BASE, 0);
5468 WREG32(RLC_RL_SIZE, 0);
5469 WREG32(RLC_LB_CNTL, 0);
5470 WREG32(RLC_LB_CNTR_MAX, 0xffffffff);
5471 WREG32(RLC_LB_CNTR_INIT, 0);
d719cef3 5472 WREG32(RLC_LB_INIT_CU_MASK, 0xffffffff);
347e7592 5473
347e7592
AD
5474 WREG32(RLC_MC_CNTL, 0);
5475 WREG32(RLC_UCODE_CNTL, 0);
5476
5477 fw_data = (const __be32 *)rdev->rlc_fw->data;
5478 for (i = 0; i < SI_RLC_UCODE_SIZE; i++) {
5479 WREG32(RLC_UCODE_ADDR, i);
5480 WREG32(RLC_UCODE_DATA, be32_to_cpup(fw_data++));
5481 }
5482 WREG32(RLC_UCODE_ADDR, 0);
5483
d719cef3
AD
5484 si_enable_lbpw(rdev, si_lbpw_supported(rdev));
5485
347e7592
AD
5486 si_rlc_start(rdev);
5487
5488 return 0;
5489}
5490
25a857fb
AD
5491static void si_enable_interrupts(struct radeon_device *rdev)
5492{
5493 u32 ih_cntl = RREG32(IH_CNTL);
5494 u32 ih_rb_cntl = RREG32(IH_RB_CNTL);
5495
5496 ih_cntl |= ENABLE_INTR;
5497 ih_rb_cntl |= IH_RB_ENABLE;
5498 WREG32(IH_CNTL, ih_cntl);
5499 WREG32(IH_RB_CNTL, ih_rb_cntl);
5500 rdev->ih.enabled = true;
5501}
5502
5503static void si_disable_interrupts(struct radeon_device *rdev)
5504{
5505 u32 ih_rb_cntl = RREG32(IH_RB_CNTL);
5506 u32 ih_cntl = RREG32(IH_CNTL);
5507
5508 ih_rb_cntl &= ~IH_RB_ENABLE;
5509 ih_cntl &= ~ENABLE_INTR;
5510 WREG32(IH_RB_CNTL, ih_rb_cntl);
5511 WREG32(IH_CNTL, ih_cntl);
5512 /* set rptr, wptr to 0 */
5513 WREG32(IH_RB_RPTR, 0);
5514 WREG32(IH_RB_WPTR, 0);
5515 rdev->ih.enabled = false;
25a857fb
AD
5516 rdev->ih.rptr = 0;
5517}
5518
5519static void si_disable_interrupt_state(struct radeon_device *rdev)
5520{
5521 u32 tmp;
5522
811e4d58
AD
5523 tmp = RREG32(CP_INT_CNTL_RING0) &
5524 (CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
5525 WREG32(CP_INT_CNTL_RING0, tmp);
25a857fb
AD
5526 WREG32(CP_INT_CNTL_RING1, 0);
5527 WREG32(CP_INT_CNTL_RING2, 0);
8c5fd7ef
AD
5528 tmp = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET) & ~TRAP_ENABLE;
5529 WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, tmp);
5530 tmp = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET) & ~TRAP_ENABLE;
5531 WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, tmp);
25a857fb 5532 WREG32(GRBM_INT_CNTL, 0);
5153550a
AD
5533 if (rdev->num_crtc >= 2) {
5534 WREG32(INT_MASK + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
5535 WREG32(INT_MASK + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
5536 }
25a857fb
AD
5537 if (rdev->num_crtc >= 4) {
5538 WREG32(INT_MASK + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
5539 WREG32(INT_MASK + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
5540 }
5541 if (rdev->num_crtc >= 6) {
5542 WREG32(INT_MASK + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
5543 WREG32(INT_MASK + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
5544 }
5545
5153550a
AD
5546 if (rdev->num_crtc >= 2) {
5547 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
5548 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
5549 }
25a857fb
AD
5550 if (rdev->num_crtc >= 4) {
5551 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
5552 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
5553 }
5554 if (rdev->num_crtc >= 6) {
5555 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
5556 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
5557 }
5558
5153550a
AD
5559 if (!ASIC_IS_NODCE(rdev)) {
5560 WREG32(DACA_AUTODETECT_INT_CONTROL, 0);
5561
5562 tmp = RREG32(DC_HPD1_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5563 WREG32(DC_HPD1_INT_CONTROL, tmp);
5564 tmp = RREG32(DC_HPD2_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5565 WREG32(DC_HPD2_INT_CONTROL, tmp);
5566 tmp = RREG32(DC_HPD3_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5567 WREG32(DC_HPD3_INT_CONTROL, tmp);
5568 tmp = RREG32(DC_HPD4_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5569 WREG32(DC_HPD4_INT_CONTROL, tmp);
5570 tmp = RREG32(DC_HPD5_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5571 WREG32(DC_HPD5_INT_CONTROL, tmp);
5572 tmp = RREG32(DC_HPD6_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5573 WREG32(DC_HPD6_INT_CONTROL, tmp);
5574 }
25a857fb
AD
5575}
5576
5577static int si_irq_init(struct radeon_device *rdev)
5578{
5579 int ret = 0;
5580 int rb_bufsz;
5581 u32 interrupt_cntl, ih_cntl, ih_rb_cntl;
5582
5583 /* allocate ring */
5584 ret = r600_ih_ring_alloc(rdev);
5585 if (ret)
5586 return ret;
5587
5588 /* disable irqs */
5589 si_disable_interrupts(rdev);
5590
5591 /* init rlc */
5592 ret = si_rlc_resume(rdev);
5593 if (ret) {
5594 r600_ih_ring_fini(rdev);
5595 return ret;
5596 }
5597
5598 /* setup interrupt control */
5599 /* set dummy read address to ring address */
5600 WREG32(INTERRUPT_CNTL2, rdev->ih.gpu_addr >> 8);
5601 interrupt_cntl = RREG32(INTERRUPT_CNTL);
5602 /* IH_DUMMY_RD_OVERRIDE=0 - dummy read disabled with msi, enabled without msi
5603 * IH_DUMMY_RD_OVERRIDE=1 - dummy read controlled by IH_DUMMY_RD_EN
5604 */
5605 interrupt_cntl &= ~IH_DUMMY_RD_OVERRIDE;
5606 /* IH_REQ_NONSNOOP_EN=1 if ring is in non-cacheable memory, e.g., vram */
5607 interrupt_cntl &= ~IH_REQ_NONSNOOP_EN;
5608 WREG32(INTERRUPT_CNTL, interrupt_cntl);
5609
5610 WREG32(IH_RB_BASE, rdev->ih.gpu_addr >> 8);
b72a8925 5611 rb_bufsz = order_base_2(rdev->ih.ring_size / 4);
25a857fb
AD
5612
5613 ih_rb_cntl = (IH_WPTR_OVERFLOW_ENABLE |
5614 IH_WPTR_OVERFLOW_CLEAR |
5615 (rb_bufsz << 1));
5616
5617 if (rdev->wb.enabled)
5618 ih_rb_cntl |= IH_WPTR_WRITEBACK_ENABLE;
5619
5620 /* set the writeback address whether it's enabled or not */
5621 WREG32(IH_RB_WPTR_ADDR_LO, (rdev->wb.gpu_addr + R600_WB_IH_WPTR_OFFSET) & 0xFFFFFFFC);
5622 WREG32(IH_RB_WPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + R600_WB_IH_WPTR_OFFSET) & 0xFF);
5623
5624 WREG32(IH_RB_CNTL, ih_rb_cntl);
5625
5626 /* set rptr, wptr to 0 */
5627 WREG32(IH_RB_RPTR, 0);
5628 WREG32(IH_RB_WPTR, 0);
5629
5630 /* Default settings for IH_CNTL (disabled at first) */
5631 ih_cntl = MC_WRREQ_CREDIT(0x10) | MC_WR_CLEAN_CNT(0x10) | MC_VMID(0);
5632 /* RPTR_REARM only works if msi's are enabled */
5633 if (rdev->msi_enabled)
5634 ih_cntl |= RPTR_REARM;
5635 WREG32(IH_CNTL, ih_cntl);
5636
5637 /* force the active interrupt state to all disabled */
5638 si_disable_interrupt_state(rdev);
5639
2099810f
DA
5640 pci_set_master(rdev->pdev);
5641
25a857fb
AD
5642 /* enable irqs */
5643 si_enable_interrupts(rdev);
5644
5645 return ret;
5646}
5647
5648int si_irq_set(struct radeon_device *rdev)
5649{
811e4d58 5650 u32 cp_int_cntl;
25a857fb
AD
5651 u32 cp_int_cntl1 = 0, cp_int_cntl2 = 0;
5652 u32 crtc1 = 0, crtc2 = 0, crtc3 = 0, crtc4 = 0, crtc5 = 0, crtc6 = 0;
5153550a 5653 u32 hpd1 = 0, hpd2 = 0, hpd3 = 0, hpd4 = 0, hpd5 = 0, hpd6 = 0;
25a857fb
AD
5654 u32 grbm_int_cntl = 0;
5655 u32 grph1 = 0, grph2 = 0, grph3 = 0, grph4 = 0, grph5 = 0, grph6 = 0;
8c5fd7ef 5656 u32 dma_cntl, dma_cntl1;
a9e61410 5657 u32 thermal_int = 0;
25a857fb
AD
5658
5659 if (!rdev->irq.installed) {
5660 WARN(1, "Can't enable IRQ/MSI because no handler is installed\n");
5661 return -EINVAL;
5662 }
5663 /* don't enable anything if the ih is disabled */
5664 if (!rdev->ih.enabled) {
5665 si_disable_interrupts(rdev);
5666 /* force the active interrupt state to all disabled */
5667 si_disable_interrupt_state(rdev);
5668 return 0;
5669 }
5670
811e4d58
AD
5671 cp_int_cntl = RREG32(CP_INT_CNTL_RING0) &
5672 (CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
5673
5153550a
AD
5674 if (!ASIC_IS_NODCE(rdev)) {
5675 hpd1 = RREG32(DC_HPD1_INT_CONTROL) & ~DC_HPDx_INT_EN;
5676 hpd2 = RREG32(DC_HPD2_INT_CONTROL) & ~DC_HPDx_INT_EN;
5677 hpd3 = RREG32(DC_HPD3_INT_CONTROL) & ~DC_HPDx_INT_EN;
5678 hpd4 = RREG32(DC_HPD4_INT_CONTROL) & ~DC_HPDx_INT_EN;
5679 hpd5 = RREG32(DC_HPD5_INT_CONTROL) & ~DC_HPDx_INT_EN;
5680 hpd6 = RREG32(DC_HPD6_INT_CONTROL) & ~DC_HPDx_INT_EN;
5681 }
25a857fb 5682
8c5fd7ef
AD
5683 dma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET) & ~TRAP_ENABLE;
5684 dma_cntl1 = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET) & ~TRAP_ENABLE;
5685
a9e61410
AD
5686 thermal_int = RREG32(CG_THERMAL_INT) &
5687 ~(THERM_INT_MASK_HIGH | THERM_INT_MASK_LOW);
5688
25a857fb 5689 /* enable CP interrupts on all rings */
736fc37f 5690 if (atomic_read(&rdev->irq.ring_int[RADEON_RING_TYPE_GFX_INDEX])) {
25a857fb
AD
5691 DRM_DEBUG("si_irq_set: sw int gfx\n");
5692 cp_int_cntl |= TIME_STAMP_INT_ENABLE;
5693 }
736fc37f 5694 if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_CP1_INDEX])) {
25a857fb
AD
5695 DRM_DEBUG("si_irq_set: sw int cp1\n");
5696 cp_int_cntl1 |= TIME_STAMP_INT_ENABLE;
5697 }
736fc37f 5698 if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_CP2_INDEX])) {
25a857fb
AD
5699 DRM_DEBUG("si_irq_set: sw int cp2\n");
5700 cp_int_cntl2 |= TIME_STAMP_INT_ENABLE;
5701 }
8c5fd7ef
AD
5702 if (atomic_read(&rdev->irq.ring_int[R600_RING_TYPE_DMA_INDEX])) {
5703 DRM_DEBUG("si_irq_set: sw int dma\n");
5704 dma_cntl |= TRAP_ENABLE;
5705 }
5706
5707 if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_DMA1_INDEX])) {
5708 DRM_DEBUG("si_irq_set: sw int dma1\n");
5709 dma_cntl1 |= TRAP_ENABLE;
5710 }
25a857fb 5711 if (rdev->irq.crtc_vblank_int[0] ||
736fc37f 5712 atomic_read(&rdev->irq.pflip[0])) {
25a857fb
AD
5713 DRM_DEBUG("si_irq_set: vblank 0\n");
5714 crtc1 |= VBLANK_INT_MASK;
5715 }
5716 if (rdev->irq.crtc_vblank_int[1] ||
736fc37f 5717 atomic_read(&rdev->irq.pflip[1])) {
25a857fb
AD
5718 DRM_DEBUG("si_irq_set: vblank 1\n");
5719 crtc2 |= VBLANK_INT_MASK;
5720 }
5721 if (rdev->irq.crtc_vblank_int[2] ||
736fc37f 5722 atomic_read(&rdev->irq.pflip[2])) {
25a857fb
AD
5723 DRM_DEBUG("si_irq_set: vblank 2\n");
5724 crtc3 |= VBLANK_INT_MASK;
5725 }
5726 if (rdev->irq.crtc_vblank_int[3] ||
736fc37f 5727 atomic_read(&rdev->irq.pflip[3])) {
25a857fb
AD
5728 DRM_DEBUG("si_irq_set: vblank 3\n");
5729 crtc4 |= VBLANK_INT_MASK;
5730 }
5731 if (rdev->irq.crtc_vblank_int[4] ||
736fc37f 5732 atomic_read(&rdev->irq.pflip[4])) {
25a857fb
AD
5733 DRM_DEBUG("si_irq_set: vblank 4\n");
5734 crtc5 |= VBLANK_INT_MASK;
5735 }
5736 if (rdev->irq.crtc_vblank_int[5] ||
736fc37f 5737 atomic_read(&rdev->irq.pflip[5])) {
25a857fb
AD
5738 DRM_DEBUG("si_irq_set: vblank 5\n");
5739 crtc6 |= VBLANK_INT_MASK;
5740 }
5741 if (rdev->irq.hpd[0]) {
5742 DRM_DEBUG("si_irq_set: hpd 1\n");
5743 hpd1 |= DC_HPDx_INT_EN;
5744 }
5745 if (rdev->irq.hpd[1]) {
5746 DRM_DEBUG("si_irq_set: hpd 2\n");
5747 hpd2 |= DC_HPDx_INT_EN;
5748 }
5749 if (rdev->irq.hpd[2]) {
5750 DRM_DEBUG("si_irq_set: hpd 3\n");
5751 hpd3 |= DC_HPDx_INT_EN;
5752 }
5753 if (rdev->irq.hpd[3]) {
5754 DRM_DEBUG("si_irq_set: hpd 4\n");
5755 hpd4 |= DC_HPDx_INT_EN;
5756 }
5757 if (rdev->irq.hpd[4]) {
5758 DRM_DEBUG("si_irq_set: hpd 5\n");
5759 hpd5 |= DC_HPDx_INT_EN;
5760 }
5761 if (rdev->irq.hpd[5]) {
5762 DRM_DEBUG("si_irq_set: hpd 6\n");
5763 hpd6 |= DC_HPDx_INT_EN;
5764 }
25a857fb
AD
5765
5766 WREG32(CP_INT_CNTL_RING0, cp_int_cntl);
5767 WREG32(CP_INT_CNTL_RING1, cp_int_cntl1);
5768 WREG32(CP_INT_CNTL_RING2, cp_int_cntl2);
5769
8c5fd7ef
AD
5770 WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, dma_cntl);
5771 WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, dma_cntl1);
5772
25a857fb
AD
5773 WREG32(GRBM_INT_CNTL, grbm_int_cntl);
5774
a9e61410
AD
5775 if (rdev->irq.dpm_thermal) {
5776 DRM_DEBUG("dpm thermal\n");
5777 thermal_int |= THERM_INT_MASK_HIGH | THERM_INT_MASK_LOW;
5778 }
5779
5153550a
AD
5780 if (rdev->num_crtc >= 2) {
5781 WREG32(INT_MASK + EVERGREEN_CRTC0_REGISTER_OFFSET, crtc1);
5782 WREG32(INT_MASK + EVERGREEN_CRTC1_REGISTER_OFFSET, crtc2);
5783 }
25a857fb
AD
5784 if (rdev->num_crtc >= 4) {
5785 WREG32(INT_MASK + EVERGREEN_CRTC2_REGISTER_OFFSET, crtc3);
5786 WREG32(INT_MASK + EVERGREEN_CRTC3_REGISTER_OFFSET, crtc4);
5787 }
5788 if (rdev->num_crtc >= 6) {
5789 WREG32(INT_MASK + EVERGREEN_CRTC4_REGISTER_OFFSET, crtc5);
5790 WREG32(INT_MASK + EVERGREEN_CRTC5_REGISTER_OFFSET, crtc6);
5791 }
5792
5153550a
AD
5793 if (rdev->num_crtc >= 2) {
5794 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET, grph1);
5795 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET, grph2);
5796 }
25a857fb
AD
5797 if (rdev->num_crtc >= 4) {
5798 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET, grph3);
5799 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET, grph4);
5800 }
5801 if (rdev->num_crtc >= 6) {
5802 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET, grph5);
5803 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET, grph6);
5804 }
5805
5153550a
AD
5806 if (!ASIC_IS_NODCE(rdev)) {
5807 WREG32(DC_HPD1_INT_CONTROL, hpd1);
5808 WREG32(DC_HPD2_INT_CONTROL, hpd2);
5809 WREG32(DC_HPD3_INT_CONTROL, hpd3);
5810 WREG32(DC_HPD4_INT_CONTROL, hpd4);
5811 WREG32(DC_HPD5_INT_CONTROL, hpd5);
5812 WREG32(DC_HPD6_INT_CONTROL, hpd6);
5813 }
25a857fb 5814
a9e61410
AD
5815 WREG32(CG_THERMAL_INT, thermal_int);
5816
25a857fb
AD
5817 return 0;
5818}
5819
5820static inline void si_irq_ack(struct radeon_device *rdev)
5821{
5822 u32 tmp;
5823
5153550a
AD
5824 if (ASIC_IS_NODCE(rdev))
5825 return;
5826
25a857fb
AD
5827 rdev->irq.stat_regs.evergreen.disp_int = RREG32(DISP_INTERRUPT_STATUS);
5828 rdev->irq.stat_regs.evergreen.disp_int_cont = RREG32(DISP_INTERRUPT_STATUS_CONTINUE);
5829 rdev->irq.stat_regs.evergreen.disp_int_cont2 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE2);
5830 rdev->irq.stat_regs.evergreen.disp_int_cont3 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE3);
5831 rdev->irq.stat_regs.evergreen.disp_int_cont4 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE4);
5832 rdev->irq.stat_regs.evergreen.disp_int_cont5 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE5);
5833 rdev->irq.stat_regs.evergreen.d1grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET);
5834 rdev->irq.stat_regs.evergreen.d2grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET);
5835 if (rdev->num_crtc >= 4) {
5836 rdev->irq.stat_regs.evergreen.d3grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET);
5837 rdev->irq.stat_regs.evergreen.d4grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET);
5838 }
5839 if (rdev->num_crtc >= 6) {
5840 rdev->irq.stat_regs.evergreen.d5grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET);
5841 rdev->irq.stat_regs.evergreen.d6grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET);
5842 }
5843
5844 if (rdev->irq.stat_regs.evergreen.d1grph_int & GRPH_PFLIP_INT_OCCURRED)
5845 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
5846 if (rdev->irq.stat_regs.evergreen.d2grph_int & GRPH_PFLIP_INT_OCCURRED)
5847 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
5848 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VBLANK_INTERRUPT)
5849 WREG32(VBLANK_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, VBLANK_ACK);
5850 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VLINE_INTERRUPT)
5851 WREG32(VLINE_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, VLINE_ACK);
5852 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VBLANK_INTERRUPT)
5853 WREG32(VBLANK_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, VBLANK_ACK);
5854 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VLINE_INTERRUPT)
5855 WREG32(VLINE_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, VLINE_ACK);
5856
5857 if (rdev->num_crtc >= 4) {
5858 if (rdev->irq.stat_regs.evergreen.d3grph_int & GRPH_PFLIP_INT_OCCURRED)
5859 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
5860 if (rdev->irq.stat_regs.evergreen.d4grph_int & GRPH_PFLIP_INT_OCCURRED)
5861 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
5862 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VBLANK_INTERRUPT)
5863 WREG32(VBLANK_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, VBLANK_ACK);
5864 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VLINE_INTERRUPT)
5865 WREG32(VLINE_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, VLINE_ACK);
5866 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VBLANK_INTERRUPT)
5867 WREG32(VBLANK_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, VBLANK_ACK);
5868 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VLINE_INTERRUPT)
5869 WREG32(VLINE_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, VLINE_ACK);
5870 }
5871
5872 if (rdev->num_crtc >= 6) {
5873 if (rdev->irq.stat_regs.evergreen.d5grph_int & GRPH_PFLIP_INT_OCCURRED)
5874 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
5875 if (rdev->irq.stat_regs.evergreen.d6grph_int & GRPH_PFLIP_INT_OCCURRED)
5876 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
5877 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VBLANK_INTERRUPT)
5878 WREG32(VBLANK_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, VBLANK_ACK);
5879 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VLINE_INTERRUPT)
5880 WREG32(VLINE_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, VLINE_ACK);
5881 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VBLANK_INTERRUPT)
5882 WREG32(VBLANK_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, VBLANK_ACK);
5883 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VLINE_INTERRUPT)
5884 WREG32(VLINE_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, VLINE_ACK);
5885 }
5886
5887 if (rdev->irq.stat_regs.evergreen.disp_int & DC_HPD1_INTERRUPT) {
5888 tmp = RREG32(DC_HPD1_INT_CONTROL);
5889 tmp |= DC_HPDx_INT_ACK;
5890 WREG32(DC_HPD1_INT_CONTROL, tmp);
5891 }
5892 if (rdev->irq.stat_regs.evergreen.disp_int_cont & DC_HPD2_INTERRUPT) {
5893 tmp = RREG32(DC_HPD2_INT_CONTROL);
5894 tmp |= DC_HPDx_INT_ACK;
5895 WREG32(DC_HPD2_INT_CONTROL, tmp);
5896 }
5897 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & DC_HPD3_INTERRUPT) {
5898 tmp = RREG32(DC_HPD3_INT_CONTROL);
5899 tmp |= DC_HPDx_INT_ACK;
5900 WREG32(DC_HPD3_INT_CONTROL, tmp);
5901 }
5902 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & DC_HPD4_INTERRUPT) {
5903 tmp = RREG32(DC_HPD4_INT_CONTROL);
5904 tmp |= DC_HPDx_INT_ACK;
5905 WREG32(DC_HPD4_INT_CONTROL, tmp);
5906 }
5907 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & DC_HPD5_INTERRUPT) {
5908 tmp = RREG32(DC_HPD5_INT_CONTROL);
5909 tmp |= DC_HPDx_INT_ACK;
5910 WREG32(DC_HPD5_INT_CONTROL, tmp);
5911 }
5912 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & DC_HPD6_INTERRUPT) {
5913 tmp = RREG32(DC_HPD5_INT_CONTROL);
5914 tmp |= DC_HPDx_INT_ACK;
5915 WREG32(DC_HPD6_INT_CONTROL, tmp);
5916 }
5917}
5918
5919static void si_irq_disable(struct radeon_device *rdev)
5920{
5921 si_disable_interrupts(rdev);
5922 /* Wait and acknowledge irq */
5923 mdelay(1);
5924 si_irq_ack(rdev);
5925 si_disable_interrupt_state(rdev);
5926}
5927
5928static void si_irq_suspend(struct radeon_device *rdev)
5929{
5930 si_irq_disable(rdev);
5931 si_rlc_stop(rdev);
5932}
5933
9b136d51
AD
5934static void si_irq_fini(struct radeon_device *rdev)
5935{
5936 si_irq_suspend(rdev);
5937 r600_ih_ring_fini(rdev);
5938}
5939
25a857fb
AD
5940static inline u32 si_get_ih_wptr(struct radeon_device *rdev)
5941{
5942 u32 wptr, tmp;
5943
5944 if (rdev->wb.enabled)
5945 wptr = le32_to_cpu(rdev->wb.wb[R600_WB_IH_WPTR_OFFSET/4]);
5946 else
5947 wptr = RREG32(IH_RB_WPTR);
5948
5949 if (wptr & RB_OVERFLOW) {
5950 /* When a ring buffer overflow happen start parsing interrupt
5951 * from the last not overwritten vector (wptr + 16). Hopefully
5952 * this should allow us to catchup.
5953 */
5954 dev_warn(rdev->dev, "IH ring buffer overflow (0x%08X, %d, %d)\n",
5955 wptr, rdev->ih.rptr, (wptr + 16) + rdev->ih.ptr_mask);
5956 rdev->ih.rptr = (wptr + 16) & rdev->ih.ptr_mask;
5957 tmp = RREG32(IH_RB_CNTL);
5958 tmp |= IH_WPTR_OVERFLOW_CLEAR;
5959 WREG32(IH_RB_CNTL, tmp);
5960 }
5961 return (wptr & rdev->ih.ptr_mask);
5962}
5963
5964/* SI IV Ring
5965 * Each IV ring entry is 128 bits:
5966 * [7:0] - interrupt source id
5967 * [31:8] - reserved
5968 * [59:32] - interrupt source data
5969 * [63:60] - reserved
5970 * [71:64] - RINGID
5971 * [79:72] - VMID
5972 * [127:80] - reserved
5973 */
5974int si_irq_process(struct radeon_device *rdev)
5975{
5976 u32 wptr;
5977 u32 rptr;
5978 u32 src_id, src_data, ring_id;
5979 u32 ring_index;
25a857fb 5980 bool queue_hotplug = false;
a9e61410 5981 bool queue_thermal = false;
fbf6dc7a 5982 u32 status, addr;
25a857fb
AD
5983
5984 if (!rdev->ih.enabled || rdev->shutdown)
5985 return IRQ_NONE;
5986
5987 wptr = si_get_ih_wptr(rdev);
c20dc369
CK
5988
5989restart_ih:
5990 /* is somebody else already processing irqs? */
5991 if (atomic_xchg(&rdev->ih.lock, 1))
5992 return IRQ_NONE;
5993
25a857fb
AD
5994 rptr = rdev->ih.rptr;
5995 DRM_DEBUG("si_irq_process start: rptr %d, wptr %d\n", rptr, wptr);
5996
25a857fb
AD
5997 /* Order reading of wptr vs. reading of IH ring data */
5998 rmb();
5999
6000 /* display interrupts */
6001 si_irq_ack(rdev);
6002
25a857fb
AD
6003 while (rptr != wptr) {
6004 /* wptr/rptr are in bytes! */
6005 ring_index = rptr / 4;
6006 src_id = le32_to_cpu(rdev->ih.ring[ring_index]) & 0xff;
6007 src_data = le32_to_cpu(rdev->ih.ring[ring_index + 1]) & 0xfffffff;
6008 ring_id = le32_to_cpu(rdev->ih.ring[ring_index + 2]) & 0xff;
6009
6010 switch (src_id) {
6011 case 1: /* D1 vblank/vline */
6012 switch (src_data) {
6013 case 0: /* D1 vblank */
6014 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VBLANK_INTERRUPT) {
6015 if (rdev->irq.crtc_vblank_int[0]) {
6016 drm_handle_vblank(rdev->ddev, 0);
6017 rdev->pm.vblank_sync = true;
6018 wake_up(&rdev->irq.vblank_queue);
6019 }
736fc37f 6020 if (atomic_read(&rdev->irq.pflip[0]))
25a857fb
AD
6021 radeon_crtc_handle_flip(rdev, 0);
6022 rdev->irq.stat_regs.evergreen.disp_int &= ~LB_D1_VBLANK_INTERRUPT;
6023 DRM_DEBUG("IH: D1 vblank\n");
6024 }
6025 break;
6026 case 1: /* D1 vline */
6027 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VLINE_INTERRUPT) {
6028 rdev->irq.stat_regs.evergreen.disp_int &= ~LB_D1_VLINE_INTERRUPT;
6029 DRM_DEBUG("IH: D1 vline\n");
6030 }
6031 break;
6032 default:
6033 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6034 break;
6035 }
6036 break;
6037 case 2: /* D2 vblank/vline */
6038 switch (src_data) {
6039 case 0: /* D2 vblank */
6040 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VBLANK_INTERRUPT) {
6041 if (rdev->irq.crtc_vblank_int[1]) {
6042 drm_handle_vblank(rdev->ddev, 1);
6043 rdev->pm.vblank_sync = true;
6044 wake_up(&rdev->irq.vblank_queue);
6045 }
736fc37f 6046 if (atomic_read(&rdev->irq.pflip[1]))
25a857fb
AD
6047 radeon_crtc_handle_flip(rdev, 1);
6048 rdev->irq.stat_regs.evergreen.disp_int_cont &= ~LB_D2_VBLANK_INTERRUPT;
6049 DRM_DEBUG("IH: D2 vblank\n");
6050 }
6051 break;
6052 case 1: /* D2 vline */
6053 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VLINE_INTERRUPT) {
6054 rdev->irq.stat_regs.evergreen.disp_int_cont &= ~LB_D2_VLINE_INTERRUPT;
6055 DRM_DEBUG("IH: D2 vline\n");
6056 }
6057 break;
6058 default:
6059 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6060 break;
6061 }
6062 break;
6063 case 3: /* D3 vblank/vline */
6064 switch (src_data) {
6065 case 0: /* D3 vblank */
6066 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VBLANK_INTERRUPT) {
6067 if (rdev->irq.crtc_vblank_int[2]) {
6068 drm_handle_vblank(rdev->ddev, 2);
6069 rdev->pm.vblank_sync = true;
6070 wake_up(&rdev->irq.vblank_queue);
6071 }
736fc37f 6072 if (atomic_read(&rdev->irq.pflip[2]))
25a857fb
AD
6073 radeon_crtc_handle_flip(rdev, 2);
6074 rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~LB_D3_VBLANK_INTERRUPT;
6075 DRM_DEBUG("IH: D3 vblank\n");
6076 }
6077 break;
6078 case 1: /* D3 vline */
6079 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VLINE_INTERRUPT) {
6080 rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~LB_D3_VLINE_INTERRUPT;
6081 DRM_DEBUG("IH: D3 vline\n");
6082 }
6083 break;
6084 default:
6085 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6086 break;
6087 }
6088 break;
6089 case 4: /* D4 vblank/vline */
6090 switch (src_data) {
6091 case 0: /* D4 vblank */
6092 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VBLANK_INTERRUPT) {
6093 if (rdev->irq.crtc_vblank_int[3]) {
6094 drm_handle_vblank(rdev->ddev, 3);
6095 rdev->pm.vblank_sync = true;
6096 wake_up(&rdev->irq.vblank_queue);
6097 }
736fc37f 6098 if (atomic_read(&rdev->irq.pflip[3]))
25a857fb
AD
6099 radeon_crtc_handle_flip(rdev, 3);
6100 rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~LB_D4_VBLANK_INTERRUPT;
6101 DRM_DEBUG("IH: D4 vblank\n");
6102 }
6103 break;
6104 case 1: /* D4 vline */
6105 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VLINE_INTERRUPT) {
6106 rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~LB_D4_VLINE_INTERRUPT;
6107 DRM_DEBUG("IH: D4 vline\n");
6108 }
6109 break;
6110 default:
6111 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6112 break;
6113 }
6114 break;
6115 case 5: /* D5 vblank/vline */
6116 switch (src_data) {
6117 case 0: /* D5 vblank */
6118 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VBLANK_INTERRUPT) {
6119 if (rdev->irq.crtc_vblank_int[4]) {
6120 drm_handle_vblank(rdev->ddev, 4);
6121 rdev->pm.vblank_sync = true;
6122 wake_up(&rdev->irq.vblank_queue);
6123 }
736fc37f 6124 if (atomic_read(&rdev->irq.pflip[4]))
25a857fb
AD
6125 radeon_crtc_handle_flip(rdev, 4);
6126 rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~LB_D5_VBLANK_INTERRUPT;
6127 DRM_DEBUG("IH: D5 vblank\n");
6128 }
6129 break;
6130 case 1: /* D5 vline */
6131 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VLINE_INTERRUPT) {
6132 rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~LB_D5_VLINE_INTERRUPT;
6133 DRM_DEBUG("IH: D5 vline\n");
6134 }
6135 break;
6136 default:
6137 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6138 break;
6139 }
6140 break;
6141 case 6: /* D6 vblank/vline */
6142 switch (src_data) {
6143 case 0: /* D6 vblank */
6144 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VBLANK_INTERRUPT) {
6145 if (rdev->irq.crtc_vblank_int[5]) {
6146 drm_handle_vblank(rdev->ddev, 5);
6147 rdev->pm.vblank_sync = true;
6148 wake_up(&rdev->irq.vblank_queue);
6149 }
736fc37f 6150 if (atomic_read(&rdev->irq.pflip[5]))
25a857fb
AD
6151 radeon_crtc_handle_flip(rdev, 5);
6152 rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~LB_D6_VBLANK_INTERRUPT;
6153 DRM_DEBUG("IH: D6 vblank\n");
6154 }
6155 break;
6156 case 1: /* D6 vline */
6157 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VLINE_INTERRUPT) {
6158 rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~LB_D6_VLINE_INTERRUPT;
6159 DRM_DEBUG("IH: D6 vline\n");
6160 }
6161 break;
6162 default:
6163 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6164 break;
6165 }
6166 break;
6167 case 42: /* HPD hotplug */
6168 switch (src_data) {
6169 case 0:
6170 if (rdev->irq.stat_regs.evergreen.disp_int & DC_HPD1_INTERRUPT) {
6171 rdev->irq.stat_regs.evergreen.disp_int &= ~DC_HPD1_INTERRUPT;
6172 queue_hotplug = true;
6173 DRM_DEBUG("IH: HPD1\n");
6174 }
6175 break;
6176 case 1:
6177 if (rdev->irq.stat_regs.evergreen.disp_int_cont & DC_HPD2_INTERRUPT) {
6178 rdev->irq.stat_regs.evergreen.disp_int_cont &= ~DC_HPD2_INTERRUPT;
6179 queue_hotplug = true;
6180 DRM_DEBUG("IH: HPD2\n");
6181 }
6182 break;
6183 case 2:
6184 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & DC_HPD3_INTERRUPT) {
6185 rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~DC_HPD3_INTERRUPT;
6186 queue_hotplug = true;
6187 DRM_DEBUG("IH: HPD3\n");
6188 }
6189 break;
6190 case 3:
6191 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & DC_HPD4_INTERRUPT) {
6192 rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~DC_HPD4_INTERRUPT;
6193 queue_hotplug = true;
6194 DRM_DEBUG("IH: HPD4\n");
6195 }
6196 break;
6197 case 4:
6198 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & DC_HPD5_INTERRUPT) {
6199 rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~DC_HPD5_INTERRUPT;
6200 queue_hotplug = true;
6201 DRM_DEBUG("IH: HPD5\n");
6202 }
6203 break;
6204 case 5:
6205 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & DC_HPD6_INTERRUPT) {
6206 rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~DC_HPD6_INTERRUPT;
6207 queue_hotplug = true;
6208 DRM_DEBUG("IH: HPD6\n");
6209 }
6210 break;
6211 default:
6212 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6213 break;
6214 }
6215 break;
ae133a11
CK
6216 case 146:
6217 case 147:
fbf6dc7a
AD
6218 addr = RREG32(VM_CONTEXT1_PROTECTION_FAULT_ADDR);
6219 status = RREG32(VM_CONTEXT1_PROTECTION_FAULT_STATUS);
ae133a11
CK
6220 dev_err(rdev->dev, "GPU fault detected: %d 0x%08x\n", src_id, src_data);
6221 dev_err(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n",
fbf6dc7a 6222 addr);
ae133a11 6223 dev_err(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
fbf6dc7a
AD
6224 status);
6225 si_vm_decode_fault(rdev, status, addr);
ae133a11
CK
6226 /* reset addr and status */
6227 WREG32_P(VM_CONTEXT1_CNTL2, 1, ~1);
6228 break;
25a857fb
AD
6229 case 176: /* RINGID0 CP_INT */
6230 radeon_fence_process(rdev, RADEON_RING_TYPE_GFX_INDEX);
6231 break;
6232 case 177: /* RINGID1 CP_INT */
6233 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
6234 break;
6235 case 178: /* RINGID2 CP_INT */
6236 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
6237 break;
6238 case 181: /* CP EOP event */
6239 DRM_DEBUG("IH: CP EOP\n");
6240 switch (ring_id) {
6241 case 0:
6242 radeon_fence_process(rdev, RADEON_RING_TYPE_GFX_INDEX);
6243 break;
6244 case 1:
6245 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
6246 break;
6247 case 2:
6248 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
6249 break;
6250 }
6251 break;
8c5fd7ef
AD
6252 case 224: /* DMA trap event */
6253 DRM_DEBUG("IH: DMA trap\n");
6254 radeon_fence_process(rdev, R600_RING_TYPE_DMA_INDEX);
6255 break;
a9e61410
AD
6256 case 230: /* thermal low to high */
6257 DRM_DEBUG("IH: thermal low to high\n");
6258 rdev->pm.dpm.thermal.high_to_low = false;
6259 queue_thermal = true;
6260 break;
6261 case 231: /* thermal high to low */
6262 DRM_DEBUG("IH: thermal high to low\n");
6263 rdev->pm.dpm.thermal.high_to_low = true;
6264 queue_thermal = true;
6265 break;
25a857fb
AD
6266 case 233: /* GUI IDLE */
6267 DRM_DEBUG("IH: GUI idle\n");
25a857fb 6268 break;
8c5fd7ef
AD
6269 case 244: /* DMA trap event */
6270 DRM_DEBUG("IH: DMA1 trap\n");
6271 radeon_fence_process(rdev, CAYMAN_RING_TYPE_DMA1_INDEX);
6272 break;
25a857fb
AD
6273 default:
6274 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6275 break;
6276 }
6277
6278 /* wptr/rptr are in bytes! */
6279 rptr += 16;
6280 rptr &= rdev->ih.ptr_mask;
6281 }
25a857fb
AD
6282 if (queue_hotplug)
6283 schedule_work(&rdev->hotplug_work);
a9e61410
AD
6284 if (queue_thermal && rdev->pm.dpm_enabled)
6285 schedule_work(&rdev->pm.dpm.thermal.work);
25a857fb
AD
6286 rdev->ih.rptr = rptr;
6287 WREG32(IH_RB_RPTR, rdev->ih.rptr);
c20dc369
CK
6288 atomic_set(&rdev->ih.lock, 0);
6289
6290 /* make sure wptr hasn't changed while processing */
6291 wptr = si_get_ih_wptr(rdev);
6292 if (wptr != rptr)
6293 goto restart_ih;
6294
25a857fb
AD
6295 return IRQ_HANDLED;
6296}
6297
9b136d51
AD
6298/*
6299 * startup/shutdown callbacks
6300 */
6301static int si_startup(struct radeon_device *rdev)
6302{
6303 struct radeon_ring *ring;
6304 int r;
6305
b9d305df
AD
6306 /* enable pcie gen2/3 link */
6307 si_pcie_gen3_enable(rdev);
e0bcf165
AD
6308 /* enable aspm */
6309 si_program_aspm(rdev);
b9d305df 6310
e5903d39
AD
6311 /* scratch needs to be initialized before MC */
6312 r = r600_vram_scratch_init(rdev);
6313 if (r)
6314 return r;
6315
6fab3feb
AD
6316 si_mc_program(rdev);
6317
9b136d51
AD
6318 if (!rdev->me_fw || !rdev->pfp_fw || !rdev->ce_fw ||
6319 !rdev->rlc_fw || !rdev->mc_fw) {
6320 r = si_init_microcode(rdev);
6321 if (r) {
6322 DRM_ERROR("Failed to load firmware!\n");
6323 return r;
6324 }
6325 }
6326
6327 r = si_mc_load_microcode(rdev);
6328 if (r) {
6329 DRM_ERROR("Failed to load MC firmware!\n");
6330 return r;
6331 }
6332
9b136d51
AD
6333 r = si_pcie_gart_enable(rdev);
6334 if (r)
6335 return r;
6336 si_gpu_init(rdev);
6337
9b136d51 6338 /* allocate rlc buffers */
1fd11777
AD
6339 if (rdev->family == CHIP_VERDE) {
6340 rdev->rlc.reg_list = verde_rlc_save_restore_register_list;
6341 rdev->rlc.reg_list_size =
6342 (u32)ARRAY_SIZE(verde_rlc_save_restore_register_list);
6343 }
6344 rdev->rlc.cs_data = si_cs_data;
6345 r = sumo_rlc_init(rdev);
9b136d51
AD
6346 if (r) {
6347 DRM_ERROR("Failed to init rlc BOs!\n");
6348 return r;
6349 }
6350
6351 /* allocate wb buffer */
6352 r = radeon_wb_init(rdev);
6353 if (r)
6354 return r;
6355
6356 r = radeon_fence_driver_start_ring(rdev, RADEON_RING_TYPE_GFX_INDEX);
6357 if (r) {
6358 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
6359 return r;
6360 }
6361
6362 r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
6363 if (r) {
6364 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
6365 return r;
6366 }
6367
6368 r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
6369 if (r) {
6370 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
6371 return r;
6372 }
6373
8c5fd7ef
AD
6374 r = radeon_fence_driver_start_ring(rdev, R600_RING_TYPE_DMA_INDEX);
6375 if (r) {
6376 dev_err(rdev->dev, "failed initializing DMA fences (%d).\n", r);
6377 return r;
6378 }
6379
6380 r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_DMA1_INDEX);
6381 if (r) {
6382 dev_err(rdev->dev, "failed initializing DMA fences (%d).\n", r);
6383 return r;
6384 }
6385
1df0d523 6386 if (rdev->has_uvd) {
e409b128 6387 r = uvd_v2_2_resume(rdev);
1df0d523
AD
6388 if (!r) {
6389 r = radeon_fence_driver_start_ring(rdev,
6390 R600_RING_TYPE_UVD_INDEX);
6391 if (r)
6392 dev_err(rdev->dev, "UVD fences init error (%d).\n", r);
6393 }
f2ba57b5 6394 if (r)
1df0d523 6395 rdev->ring[R600_RING_TYPE_UVD_INDEX].ring_size = 0;
f2ba57b5 6396 }
f2ba57b5 6397
9b136d51 6398 /* Enable IRQ */
e49f3959
AH
6399 if (!rdev->irq.installed) {
6400 r = radeon_irq_kms_init(rdev);
6401 if (r)
6402 return r;
6403 }
6404
9b136d51
AD
6405 r = si_irq_init(rdev);
6406 if (r) {
6407 DRM_ERROR("radeon: IH init failed (%d).\n", r);
6408 radeon_irq_kms_fini(rdev);
6409 return r;
6410 }
6411 si_irq_set(rdev);
6412
6413 ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
6414 r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP_RPTR_OFFSET,
6415 CP_RB0_RPTR, CP_RB0_WPTR,
2e1e6dad 6416 RADEON_CP_PACKET2);
9b136d51
AD
6417 if (r)
6418 return r;
6419
6420 ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
6421 r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP1_RPTR_OFFSET,
6422 CP_RB1_RPTR, CP_RB1_WPTR,
2e1e6dad 6423 RADEON_CP_PACKET2);
9b136d51
AD
6424 if (r)
6425 return r;
6426
6427 ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
6428 r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP2_RPTR_OFFSET,
6429 CP_RB2_RPTR, CP_RB2_WPTR,
2e1e6dad 6430 RADEON_CP_PACKET2);
9b136d51
AD
6431 if (r)
6432 return r;
6433
8c5fd7ef
AD
6434 ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
6435 r = radeon_ring_init(rdev, ring, ring->ring_size, R600_WB_DMA_RPTR_OFFSET,
6436 DMA_RB_RPTR + DMA0_REGISTER_OFFSET,
6437 DMA_RB_WPTR + DMA0_REGISTER_OFFSET,
2e1e6dad 6438 DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0));
8c5fd7ef
AD
6439 if (r)
6440 return r;
6441
6442 ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
6443 r = radeon_ring_init(rdev, ring, ring->ring_size, CAYMAN_WB_DMA1_RPTR_OFFSET,
6444 DMA_RB_RPTR + DMA1_REGISTER_OFFSET,
6445 DMA_RB_WPTR + DMA1_REGISTER_OFFSET,
2e1e6dad 6446 DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0));
8c5fd7ef
AD
6447 if (r)
6448 return r;
6449
9b136d51
AD
6450 r = si_cp_load_microcode(rdev);
6451 if (r)
6452 return r;
6453 r = si_cp_resume(rdev);
6454 if (r)
6455 return r;
6456
8c5fd7ef
AD
6457 r = cayman_dma_resume(rdev);
6458 if (r)
6459 return r;
6460
1df0d523
AD
6461 if (rdev->has_uvd) {
6462 ring = &rdev->ring[R600_RING_TYPE_UVD_INDEX];
6463 if (ring->ring_size) {
02c9f7fa 6464 r = radeon_ring_init(rdev, ring, ring->ring_size, 0,
1df0d523 6465 UVD_RBC_RB_RPTR, UVD_RBC_RB_WPTR,
2e1e6dad 6466 RADEON_CP_PACKET2);
1df0d523 6467 if (!r)
e409b128 6468 r = uvd_v1_0_init(rdev);
1df0d523
AD
6469 if (r)
6470 DRM_ERROR("radeon: failed initializing UVD (%d).\n", r);
6471 }
f2ba57b5
CK
6472 }
6473
2898c348
CK
6474 r = radeon_ib_pool_init(rdev);
6475 if (r) {
6476 dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
9b136d51 6477 return r;
2898c348 6478 }
9b136d51 6479
c6105f24
CK
6480 r = radeon_vm_manager_init(rdev);
6481 if (r) {
6482 dev_err(rdev->dev, "vm manager initialization failed (%d).\n", r);
9b136d51 6483 return r;
c6105f24 6484 }
9b136d51 6485
b530602f
AD
6486 r = dce6_audio_init(rdev);
6487 if (r)
6488 return r;
6489
9b136d51
AD
6490 return 0;
6491}
6492
6493int si_resume(struct radeon_device *rdev)
6494{
6495 int r;
6496
6497 /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw,
6498 * posting will perform necessary task to bring back GPU into good
6499 * shape.
6500 */
6501 /* post card */
6502 atom_asic_init(rdev->mode_info.atom_context);
6503
205996c0
AD
6504 /* init golden registers */
6505 si_init_golden_registers(rdev);
6506
9b136d51
AD
6507 rdev->accel_working = true;
6508 r = si_startup(rdev);
6509 if (r) {
6510 DRM_ERROR("si startup failed on resume\n");
6511 rdev->accel_working = false;
6512 return r;
6513 }
6514
6515 return r;
6516
6517}
6518
6519int si_suspend(struct radeon_device *rdev)
6520{
b530602f 6521 dce6_audio_fini(rdev);
fa3daf9a 6522 radeon_vm_manager_fini(rdev);
9b136d51 6523 si_cp_enable(rdev, false);
8c5fd7ef 6524 cayman_dma_stop(rdev);
1df0d523 6525 if (rdev->has_uvd) {
e409b128 6526 uvd_v1_0_fini(rdev);
1df0d523
AD
6527 radeon_uvd_suspend(rdev);
6528 }
e16866ec
AD
6529 si_fini_pg(rdev);
6530 si_fini_cg(rdev);
9b136d51
AD
6531 si_irq_suspend(rdev);
6532 radeon_wb_disable(rdev);
6533 si_pcie_gart_disable(rdev);
6534 return 0;
6535}
6536
6537/* Plan is to move initialization in that function and use
6538 * helper function so that radeon_device_init pretty much
6539 * do nothing more than calling asic specific function. This
6540 * should also allow to remove a bunch of callback function
6541 * like vram_info.
6542 */
6543int si_init(struct radeon_device *rdev)
6544{
6545 struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
6546 int r;
6547
9b136d51
AD
6548 /* Read BIOS */
6549 if (!radeon_get_bios(rdev)) {
6550 if (ASIC_IS_AVIVO(rdev))
6551 return -EINVAL;
6552 }
6553 /* Must be an ATOMBIOS */
6554 if (!rdev->is_atom_bios) {
6555 dev_err(rdev->dev, "Expecting atombios for cayman GPU\n");
6556 return -EINVAL;
6557 }
6558 r = radeon_atombios_init(rdev);
6559 if (r)
6560 return r;
6561
6562 /* Post card if necessary */
6563 if (!radeon_card_posted(rdev)) {
6564 if (!rdev->bios) {
6565 dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n");
6566 return -EINVAL;
6567 }
6568 DRM_INFO("GPU not posted. posting now...\n");
6569 atom_asic_init(rdev->mode_info.atom_context);
6570 }
205996c0
AD
6571 /* init golden registers */
6572 si_init_golden_registers(rdev);
9b136d51
AD
6573 /* Initialize scratch registers */
6574 si_scratch_init(rdev);
6575 /* Initialize surface registers */
6576 radeon_surface_init(rdev);
6577 /* Initialize clocks */
6578 radeon_get_clock_info(rdev->ddev);
6579
6580 /* Fence driver */
6581 r = radeon_fence_driver_init(rdev);
6582 if (r)
6583 return r;
6584
6585 /* initialize memory controller */
6586 r = si_mc_init(rdev);
6587 if (r)
6588 return r;
6589 /* Memory manager */
6590 r = radeon_bo_init(rdev);
6591 if (r)
6592 return r;
6593
9b136d51
AD
6594 ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
6595 ring->ring_obj = NULL;
6596 r600_ring_init(rdev, ring, 1024 * 1024);
6597
6598 ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
6599 ring->ring_obj = NULL;
6600 r600_ring_init(rdev, ring, 1024 * 1024);
6601
6602 ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
6603 ring->ring_obj = NULL;
6604 r600_ring_init(rdev, ring, 1024 * 1024);
6605
8c5fd7ef
AD
6606 ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
6607 ring->ring_obj = NULL;
6608 r600_ring_init(rdev, ring, 64 * 1024);
6609
6610 ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
6611 ring->ring_obj = NULL;
6612 r600_ring_init(rdev, ring, 64 * 1024);
6613
1df0d523
AD
6614 if (rdev->has_uvd) {
6615 r = radeon_uvd_init(rdev);
6616 if (!r) {
6617 ring = &rdev->ring[R600_RING_TYPE_UVD_INDEX];
6618 ring->ring_obj = NULL;
6619 r600_ring_init(rdev, ring, 4096);
6620 }
f2ba57b5
CK
6621 }
6622
9b136d51
AD
6623 rdev->ih.ring_obj = NULL;
6624 r600_ih_ring_init(rdev, 64 * 1024);
6625
6626 r = r600_pcie_gart_init(rdev);
6627 if (r)
6628 return r;
6629
9b136d51 6630 rdev->accel_working = true;
9b136d51
AD
6631 r = si_startup(rdev);
6632 if (r) {
6633 dev_err(rdev->dev, "disabling GPU acceleration\n");
6634 si_cp_fini(rdev);
8c5fd7ef 6635 cayman_dma_fini(rdev);
9b136d51 6636 si_irq_fini(rdev);
1fd11777 6637 sumo_rlc_fini(rdev);
9b136d51 6638 radeon_wb_fini(rdev);
2898c348 6639 radeon_ib_pool_fini(rdev);
9b136d51
AD
6640 radeon_vm_manager_fini(rdev);
6641 radeon_irq_kms_fini(rdev);
6642 si_pcie_gart_fini(rdev);
6643 rdev->accel_working = false;
6644 }
6645
6646 /* Don't start up if the MC ucode is missing.
6647 * The default clocks and voltages before the MC ucode
6648 * is loaded are not suffient for advanced operations.
6649 */
6650 if (!rdev->mc_fw) {
6651 DRM_ERROR("radeon: MC ucode required for NI+.\n");
6652 return -EINVAL;
6653 }
6654
6655 return 0;
6656}
6657
6658void si_fini(struct radeon_device *rdev)
6659{
9b136d51 6660 si_cp_fini(rdev);
8c5fd7ef 6661 cayman_dma_fini(rdev);
f8f84ac5 6662 si_fini_pg(rdev);
e16866ec 6663 si_fini_cg(rdev);
9b136d51 6664 si_irq_fini(rdev);
1fd11777 6665 sumo_rlc_fini(rdev);
9b136d51
AD
6666 radeon_wb_fini(rdev);
6667 radeon_vm_manager_fini(rdev);
2898c348 6668 radeon_ib_pool_fini(rdev);
9b136d51 6669 radeon_irq_kms_fini(rdev);
2858c00d 6670 if (rdev->has_uvd) {
e409b128 6671 uvd_v1_0_fini(rdev);
1df0d523 6672 radeon_uvd_fini(rdev);
2858c00d 6673 }
9b136d51
AD
6674 si_pcie_gart_fini(rdev);
6675 r600_vram_scratch_fini(rdev);
6676 radeon_gem_fini(rdev);
9b136d51
AD
6677 radeon_fence_driver_fini(rdev);
6678 radeon_bo_fini(rdev);
6679 radeon_atombios_fini(rdev);
6680 kfree(rdev->bios);
6681 rdev->bios = NULL;
6682}
6683
6759a0a7 6684/**
d0418894 6685 * si_get_gpu_clock_counter - return GPU clock counter snapshot
6759a0a7
MO
6686 *
6687 * @rdev: radeon_device pointer
6688 *
6689 * Fetches a GPU clock counter snapshot (SI).
6690 * Returns the 64 bit clock counter snapshot.
6691 */
d0418894 6692uint64_t si_get_gpu_clock_counter(struct radeon_device *rdev)
6759a0a7
MO
6693{
6694 uint64_t clock;
6695
6696 mutex_lock(&rdev->gpu_clock_mutex);
6697 WREG32(RLC_CAPTURE_GPU_CLOCK_COUNT, 1);
6698 clock = (uint64_t)RREG32(RLC_GPU_CLOCK_COUNT_LSB) |
6699 ((uint64_t)RREG32(RLC_GPU_CLOCK_COUNT_MSB) << 32ULL);
6700 mutex_unlock(&rdev->gpu_clock_mutex);
6701 return clock;
6702}
2539eb02 6703
2539eb02
CK
6704int si_set_uvd_clocks(struct radeon_device *rdev, u32 vclk, u32 dclk)
6705{
facd112d 6706 unsigned fb_div = 0, vclk_div = 0, dclk_div = 0;
2539eb02
CK
6707 int r;
6708
4ed10835
CK
6709 /* bypass vclk and dclk with bclk */
6710 WREG32_P(CG_UPLL_FUNC_CNTL_2,
6711 VCLK_SRC_SEL(1) | DCLK_SRC_SEL(1),
6712 ~(VCLK_SRC_SEL_MASK | DCLK_SRC_SEL_MASK));
6713
6714 /* put PLL in bypass mode */
6715 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_BYPASS_EN_MASK, ~UPLL_BYPASS_EN_MASK);
6716
6717 if (!vclk || !dclk) {
6718 /* keep the Bypass mode, put PLL to sleep */
6719 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_SLEEP_MASK, ~UPLL_SLEEP_MASK);
6720 return 0;
6721 }
6722
facd112d
CK
6723 r = radeon_uvd_calc_upll_dividers(rdev, vclk, dclk, 125000, 250000,
6724 16384, 0x03FFFFFF, 0, 128, 5,
6725 &fb_div, &vclk_div, &dclk_div);
6726 if (r)
6727 return r;
2539eb02
CK
6728
6729 /* set RESET_ANTI_MUX to 0 */
6730 WREG32_P(CG_UPLL_FUNC_CNTL_5, 0, ~RESET_ANTI_MUX_MASK);
6731
6732 /* set VCO_MODE to 1 */
6733 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_VCO_MODE_MASK, ~UPLL_VCO_MODE_MASK);
6734
6735 /* toggle UPLL_SLEEP to 1 then back to 0 */
6736 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_SLEEP_MASK, ~UPLL_SLEEP_MASK);
6737 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_SLEEP_MASK);
6738
6739 /* deassert UPLL_RESET */
6740 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_RESET_MASK);
6741
6742 mdelay(1);
6743
facd112d 6744 r = radeon_uvd_send_upll_ctlreq(rdev, CG_UPLL_FUNC_CNTL);
2539eb02
CK
6745 if (r)
6746 return r;
6747
6748 /* assert UPLL_RESET again */
6749 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_RESET_MASK, ~UPLL_RESET_MASK);
6750
6751 /* disable spread spectrum. */
6752 WREG32_P(CG_UPLL_SPREAD_SPECTRUM, 0, ~SSEN_MASK);
6753
6754 /* set feedback divider */
facd112d 6755 WREG32_P(CG_UPLL_FUNC_CNTL_3, UPLL_FB_DIV(fb_div), ~UPLL_FB_DIV_MASK);
2539eb02
CK
6756
6757 /* set ref divider to 0 */
6758 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_REF_DIV_MASK);
6759
facd112d 6760 if (fb_div < 307200)
2539eb02
CK
6761 WREG32_P(CG_UPLL_FUNC_CNTL_4, 0, ~UPLL_SPARE_ISPARE9);
6762 else
6763 WREG32_P(CG_UPLL_FUNC_CNTL_4, UPLL_SPARE_ISPARE9, ~UPLL_SPARE_ISPARE9);
6764
6765 /* set PDIV_A and PDIV_B */
6766 WREG32_P(CG_UPLL_FUNC_CNTL_2,
facd112d 6767 UPLL_PDIV_A(vclk_div) | UPLL_PDIV_B(dclk_div),
2539eb02
CK
6768 ~(UPLL_PDIV_A_MASK | UPLL_PDIV_B_MASK));
6769
6770 /* give the PLL some time to settle */
6771 mdelay(15);
6772
6773 /* deassert PLL_RESET */
6774 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_RESET_MASK);
6775
6776 mdelay(15);
6777
6778 /* switch from bypass mode to normal mode */
6779 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_BYPASS_EN_MASK);
6780
facd112d 6781 r = radeon_uvd_send_upll_ctlreq(rdev, CG_UPLL_FUNC_CNTL);
2539eb02
CK
6782 if (r)
6783 return r;
6784
6785 /* switch VCLK and DCLK selection */
6786 WREG32_P(CG_UPLL_FUNC_CNTL_2,
6787 VCLK_SRC_SEL(2) | DCLK_SRC_SEL(2),
6788 ~(VCLK_SRC_SEL_MASK | DCLK_SRC_SEL_MASK));
6789
6790 mdelay(100);
6791
6792 return 0;
6793}
b9d305df
AD
6794
6795static void si_pcie_gen3_enable(struct radeon_device *rdev)
6796{
6797 struct pci_dev *root = rdev->pdev->bus->self;
6798 int bridge_pos, gpu_pos;
6799 u32 speed_cntl, mask, current_data_rate;
6800 int ret, i;
6801 u16 tmp16;
6802
6803 if (radeon_pcie_gen2 == 0)
6804 return;
6805
6806 if (rdev->flags & RADEON_IS_IGP)
6807 return;
6808
6809 if (!(rdev->flags & RADEON_IS_PCIE))
6810 return;
6811
6812 ret = drm_pcie_get_speed_cap_mask(rdev->ddev, &mask);
6813 if (ret != 0)
6814 return;
6815
6816 if (!(mask & (DRM_PCIE_SPEED_50 | DRM_PCIE_SPEED_80)))
6817 return;
6818
6819 speed_cntl = RREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL);
6820 current_data_rate = (speed_cntl & LC_CURRENT_DATA_RATE_MASK) >>
6821 LC_CURRENT_DATA_RATE_SHIFT;
6822 if (mask & DRM_PCIE_SPEED_80) {
6823 if (current_data_rate == 2) {
6824 DRM_INFO("PCIE gen 3 link speeds already enabled\n");
6825 return;
6826 }
6827 DRM_INFO("enabling PCIE gen 3 link speeds, disable with radeon.pcie_gen2=0\n");
6828 } else if (mask & DRM_PCIE_SPEED_50) {
6829 if (current_data_rate == 1) {
6830 DRM_INFO("PCIE gen 2 link speeds already enabled\n");
6831 return;
6832 }
6833 DRM_INFO("enabling PCIE gen 2 link speeds, disable with radeon.pcie_gen2=0\n");
6834 }
6835
6836 bridge_pos = pci_pcie_cap(root);
6837 if (!bridge_pos)
6838 return;
6839
6840 gpu_pos = pci_pcie_cap(rdev->pdev);
6841 if (!gpu_pos)
6842 return;
6843
6844 if (mask & DRM_PCIE_SPEED_80) {
6845 /* re-try equalization if gen3 is not already enabled */
6846 if (current_data_rate != 2) {
6847 u16 bridge_cfg, gpu_cfg;
6848 u16 bridge_cfg2, gpu_cfg2;
6849 u32 max_lw, current_lw, tmp;
6850
6851 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL, &bridge_cfg);
6852 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, &gpu_cfg);
6853
6854 tmp16 = bridge_cfg | PCI_EXP_LNKCTL_HAWD;
6855 pci_write_config_word(root, bridge_pos + PCI_EXP_LNKCTL, tmp16);
6856
6857 tmp16 = gpu_cfg | PCI_EXP_LNKCTL_HAWD;
6858 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, tmp16);
6859
6860 tmp = RREG32_PCIE(PCIE_LC_STATUS1);
6861 max_lw = (tmp & LC_DETECTED_LINK_WIDTH_MASK) >> LC_DETECTED_LINK_WIDTH_SHIFT;
6862 current_lw = (tmp & LC_OPERATING_LINK_WIDTH_MASK) >> LC_OPERATING_LINK_WIDTH_SHIFT;
6863
6864 if (current_lw < max_lw) {
6865 tmp = RREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL);
6866 if (tmp & LC_RENEGOTIATION_SUPPORT) {
6867 tmp &= ~(LC_LINK_WIDTH_MASK | LC_UPCONFIGURE_DIS);
6868 tmp |= (max_lw << LC_LINK_WIDTH_SHIFT);
6869 tmp |= LC_UPCONFIGURE_SUPPORT | LC_RENEGOTIATE_EN | LC_RECONFIG_NOW;
6870 WREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL, tmp);
6871 }
6872 }
6873
6874 for (i = 0; i < 10; i++) {
6875 /* check status */
6876 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_DEVSTA, &tmp16);
6877 if (tmp16 & PCI_EXP_DEVSTA_TRPND)
6878 break;
6879
6880 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL, &bridge_cfg);
6881 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, &gpu_cfg);
6882
6883 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL2, &bridge_cfg2);
6884 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, &gpu_cfg2);
6885
6886 tmp = RREG32_PCIE_PORT(PCIE_LC_CNTL4);
6887 tmp |= LC_SET_QUIESCE;
6888 WREG32_PCIE_PORT(PCIE_LC_CNTL4, tmp);
6889
6890 tmp = RREG32_PCIE_PORT(PCIE_LC_CNTL4);
6891 tmp |= LC_REDO_EQ;
6892 WREG32_PCIE_PORT(PCIE_LC_CNTL4, tmp);
6893
6894 mdelay(100);
6895
6896 /* linkctl */
6897 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL, &tmp16);
6898 tmp16 &= ~PCI_EXP_LNKCTL_HAWD;
6899 tmp16 |= (bridge_cfg & PCI_EXP_LNKCTL_HAWD);
6900 pci_write_config_word(root, bridge_pos + PCI_EXP_LNKCTL, tmp16);
6901
6902 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, &tmp16);
6903 tmp16 &= ~PCI_EXP_LNKCTL_HAWD;
6904 tmp16 |= (gpu_cfg & PCI_EXP_LNKCTL_HAWD);
6905 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, tmp16);
6906
6907 /* linkctl2 */
6908 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL2, &tmp16);
6909 tmp16 &= ~((1 << 4) | (7 << 9));
6910 tmp16 |= (bridge_cfg2 & ((1 << 4) | (7 << 9)));
6911 pci_write_config_word(root, bridge_pos + PCI_EXP_LNKCTL2, tmp16);
6912
6913 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, &tmp16);
6914 tmp16 &= ~((1 << 4) | (7 << 9));
6915 tmp16 |= (gpu_cfg2 & ((1 << 4) | (7 << 9)));
6916 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, tmp16);
6917
6918 tmp = RREG32_PCIE_PORT(PCIE_LC_CNTL4);
6919 tmp &= ~LC_SET_QUIESCE;
6920 WREG32_PCIE_PORT(PCIE_LC_CNTL4, tmp);
6921 }
6922 }
6923 }
6924
6925 /* set the link speed */
6926 speed_cntl |= LC_FORCE_EN_SW_SPEED_CHANGE | LC_FORCE_DIS_HW_SPEED_CHANGE;
6927 speed_cntl &= ~LC_FORCE_DIS_SW_SPEED_CHANGE;
6928 WREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL, speed_cntl);
6929
6930 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, &tmp16);
6931 tmp16 &= ~0xf;
6932 if (mask & DRM_PCIE_SPEED_80)
6933 tmp16 |= 3; /* gen3 */
6934 else if (mask & DRM_PCIE_SPEED_50)
6935 tmp16 |= 2; /* gen2 */
6936 else
6937 tmp16 |= 1; /* gen1 */
6938 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, tmp16);
6939
6940 speed_cntl = RREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL);
6941 speed_cntl |= LC_INITIATE_LINK_SPEED_CHANGE;
6942 WREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL, speed_cntl);
6943
6944 for (i = 0; i < rdev->usec_timeout; i++) {
6945 speed_cntl = RREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL);
6946 if ((speed_cntl & LC_INITIATE_LINK_SPEED_CHANGE) == 0)
6947 break;
6948 udelay(1);
6949 }
6950}
6951
e0bcf165
AD
6952static void si_program_aspm(struct radeon_device *rdev)
6953{
6954 u32 data, orig;
6955 bool disable_l0s = false, disable_l1 = false, disable_plloff_in_l1 = false;
6956 bool disable_clkreq = false;
6957
1294d4a3
AD
6958 if (radeon_aspm == 0)
6959 return;
6960
e0bcf165
AD
6961 if (!(rdev->flags & RADEON_IS_PCIE))
6962 return;
6963
6964 orig = data = RREG32_PCIE_PORT(PCIE_LC_N_FTS_CNTL);
6965 data &= ~LC_XMIT_N_FTS_MASK;
6966 data |= LC_XMIT_N_FTS(0x24) | LC_XMIT_N_FTS_OVERRIDE_EN;
6967 if (orig != data)
6968 WREG32_PCIE_PORT(PCIE_LC_N_FTS_CNTL, data);
6969
6970 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL3);
6971 data |= LC_GO_TO_RECOVERY;
6972 if (orig != data)
6973 WREG32_PCIE_PORT(PCIE_LC_CNTL3, data);
6974
6975 orig = data = RREG32_PCIE(PCIE_P_CNTL);
6976 data |= P_IGNORE_EDB_ERR;
6977 if (orig != data)
6978 WREG32_PCIE(PCIE_P_CNTL, data);
6979
6980 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL);
6981 data &= ~(LC_L0S_INACTIVITY_MASK | LC_L1_INACTIVITY_MASK);
6982 data |= LC_PMI_TO_L1_DIS;
6983 if (!disable_l0s)
6984 data |= LC_L0S_INACTIVITY(7);
6985
6986 if (!disable_l1) {
6987 data |= LC_L1_INACTIVITY(7);
6988 data &= ~LC_PMI_TO_L1_DIS;
6989 if (orig != data)
6990 WREG32_PCIE_PORT(PCIE_LC_CNTL, data);
6991
6992 if (!disable_plloff_in_l1) {
6993 bool clk_req_support;
6994
6995 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_0);
6996 data &= ~(PLL_POWER_STATE_IN_OFF_0_MASK | PLL_POWER_STATE_IN_TXS2_0_MASK);
6997 data |= PLL_POWER_STATE_IN_OFF_0(7) | PLL_POWER_STATE_IN_TXS2_0(7);
6998 if (orig != data)
6999 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_0, data);
7000
7001 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_1);
7002 data &= ~(PLL_POWER_STATE_IN_OFF_1_MASK | PLL_POWER_STATE_IN_TXS2_1_MASK);
7003 data |= PLL_POWER_STATE_IN_OFF_1(7) | PLL_POWER_STATE_IN_TXS2_1(7);
7004 if (orig != data)
7005 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_1, data);
7006
7007 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_0);
7008 data &= ~(PLL_POWER_STATE_IN_OFF_0_MASK | PLL_POWER_STATE_IN_TXS2_0_MASK);
7009 data |= PLL_POWER_STATE_IN_OFF_0(7) | PLL_POWER_STATE_IN_TXS2_0(7);
7010 if (orig != data)
7011 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_0, data);
7012
7013 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_1);
7014 data &= ~(PLL_POWER_STATE_IN_OFF_1_MASK | PLL_POWER_STATE_IN_TXS2_1_MASK);
7015 data |= PLL_POWER_STATE_IN_OFF_1(7) | PLL_POWER_STATE_IN_TXS2_1(7);
7016 if (orig != data)
7017 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_1, data);
7018
7019 if ((rdev->family != CHIP_OLAND) && (rdev->family != CHIP_HAINAN)) {
7020 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_0);
7021 data &= ~PLL_RAMP_UP_TIME_0_MASK;
7022 if (orig != data)
7023 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_0, data);
7024
7025 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_1);
7026 data &= ~PLL_RAMP_UP_TIME_1_MASK;
7027 if (orig != data)
7028 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_1, data);
7029
7030 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_2);
7031 data &= ~PLL_RAMP_UP_TIME_2_MASK;
7032 if (orig != data)
7033 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_2, data);
7034
7035 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_3);
7036 data &= ~PLL_RAMP_UP_TIME_3_MASK;
7037 if (orig != data)
7038 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_3, data);
7039
7040 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_0);
7041 data &= ~PLL_RAMP_UP_TIME_0_MASK;
7042 if (orig != data)
7043 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_0, data);
7044
7045 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_1);
7046 data &= ~PLL_RAMP_UP_TIME_1_MASK;
7047 if (orig != data)
7048 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_1, data);
7049
7050 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_2);
7051 data &= ~PLL_RAMP_UP_TIME_2_MASK;
7052 if (orig != data)
7053 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_2, data);
7054
7055 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_3);
7056 data &= ~PLL_RAMP_UP_TIME_3_MASK;
7057 if (orig != data)
7058 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_3, data);
7059 }
7060 orig = data = RREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL);
7061 data &= ~LC_DYN_LANES_PWR_STATE_MASK;
7062 data |= LC_DYN_LANES_PWR_STATE(3);
7063 if (orig != data)
7064 WREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL, data);
7065
7066 orig = data = RREG32_PIF_PHY0(PB0_PIF_CNTL);
7067 data &= ~LS2_EXIT_TIME_MASK;
7068 if ((rdev->family == CHIP_OLAND) || (rdev->family == CHIP_HAINAN))
7069 data |= LS2_EXIT_TIME(5);
7070 if (orig != data)
7071 WREG32_PIF_PHY0(PB0_PIF_CNTL, data);
7072
7073 orig = data = RREG32_PIF_PHY1(PB1_PIF_CNTL);
7074 data &= ~LS2_EXIT_TIME_MASK;
7075 if ((rdev->family == CHIP_OLAND) || (rdev->family == CHIP_HAINAN))
7076 data |= LS2_EXIT_TIME(5);
7077 if (orig != data)
7078 WREG32_PIF_PHY1(PB1_PIF_CNTL, data);
7079
7080 if (!disable_clkreq) {
7081 struct pci_dev *root = rdev->pdev->bus->self;
7082 u32 lnkcap;
7083
7084 clk_req_support = false;
7085 pcie_capability_read_dword(root, PCI_EXP_LNKCAP, &lnkcap);
7086 if (lnkcap & PCI_EXP_LNKCAP_CLKPM)
7087 clk_req_support = true;
7088 } else {
7089 clk_req_support = false;
7090 }
7091
7092 if (clk_req_support) {
7093 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL2);
7094 data |= LC_ALLOW_PDWN_IN_L1 | LC_ALLOW_PDWN_IN_L23;
7095 if (orig != data)
7096 WREG32_PCIE_PORT(PCIE_LC_CNTL2, data);
7097
7098 orig = data = RREG32(THM_CLK_CNTL);
7099 data &= ~(CMON_CLK_SEL_MASK | TMON_CLK_SEL_MASK);
7100 data |= CMON_CLK_SEL(1) | TMON_CLK_SEL(1);
7101 if (orig != data)
7102 WREG32(THM_CLK_CNTL, data);
7103
7104 orig = data = RREG32(MISC_CLK_CNTL);
7105 data &= ~(DEEP_SLEEP_CLK_SEL_MASK | ZCLK_SEL_MASK);
7106 data |= DEEP_SLEEP_CLK_SEL(1) | ZCLK_SEL(1);
7107 if (orig != data)
7108 WREG32(MISC_CLK_CNTL, data);
7109
7110 orig = data = RREG32(CG_CLKPIN_CNTL);
7111 data &= ~BCLK_AS_XCLK;
7112 if (orig != data)
7113 WREG32(CG_CLKPIN_CNTL, data);
7114
7115 orig = data = RREG32(CG_CLKPIN_CNTL_2);
7116 data &= ~FORCE_BIF_REFCLK_EN;
7117 if (orig != data)
7118 WREG32(CG_CLKPIN_CNTL_2, data);
7119
7120 orig = data = RREG32(MPLL_BYPASSCLK_SEL);
7121 data &= ~MPLL_CLKOUT_SEL_MASK;
7122 data |= MPLL_CLKOUT_SEL(4);
7123 if (orig != data)
7124 WREG32(MPLL_BYPASSCLK_SEL, data);
7125
7126 orig = data = RREG32(SPLL_CNTL_MODE);
7127 data &= ~SPLL_REFCLK_SEL_MASK;
7128 if (orig != data)
7129 WREG32(SPLL_CNTL_MODE, data);
7130 }
7131 }
7132 } else {
7133 if (orig != data)
7134 WREG32_PCIE_PORT(PCIE_LC_CNTL, data);
7135 }
7136
7137 orig = data = RREG32_PCIE(PCIE_CNTL2);
7138 data |= SLV_MEM_LS_EN | MST_MEM_LS_EN | REPLAY_MEM_LS_EN;
7139 if (orig != data)
7140 WREG32_PCIE(PCIE_CNTL2, data);
7141
7142 if (!disable_l0s) {
7143 data = RREG32_PCIE_PORT(PCIE_LC_N_FTS_CNTL);
7144 if((data & LC_N_FTS_MASK) == LC_N_FTS_MASK) {
7145 data = RREG32_PCIE(PCIE_LC_STATUS1);
7146 if ((data & LC_REVERSE_XMIT) && (data & LC_REVERSE_RCVR)) {
7147 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL);
7148 data &= ~LC_L0S_INACTIVITY_MASK;
7149 if (orig != data)
7150 WREG32_PCIE_PORT(PCIE_LC_CNTL, data);
7151 }
7152 }
7153 }
7154}