drm/nouveau: ltc/gf100-: fix cbc issues on certain boards
[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");
1ebe9280 42MODULE_FIRMWARE("radeon/TAHITI_mc2.bin");
0f0de06c 43MODULE_FIRMWARE("radeon/TAHITI_rlc.bin");
a9e61410 44MODULE_FIRMWARE("radeon/TAHITI_smc.bin");
629bd33c
AD
45
46MODULE_FIRMWARE("radeon/tahiti_pfp.bin");
47MODULE_FIRMWARE("radeon/tahiti_me.bin");
48MODULE_FIRMWARE("radeon/tahiti_ce.bin");
49MODULE_FIRMWARE("radeon/tahiti_mc.bin");
50MODULE_FIRMWARE("radeon/tahiti_rlc.bin");
51MODULE_FIRMWARE("radeon/tahiti_smc.bin");
52
0f0de06c
AD
53MODULE_FIRMWARE("radeon/PITCAIRN_pfp.bin");
54MODULE_FIRMWARE("radeon/PITCAIRN_me.bin");
55MODULE_FIRMWARE("radeon/PITCAIRN_ce.bin");
56MODULE_FIRMWARE("radeon/PITCAIRN_mc.bin");
1ebe9280 57MODULE_FIRMWARE("radeon/PITCAIRN_mc2.bin");
0f0de06c 58MODULE_FIRMWARE("radeon/PITCAIRN_rlc.bin");
a9e61410 59MODULE_FIRMWARE("radeon/PITCAIRN_smc.bin");
629bd33c
AD
60
61MODULE_FIRMWARE("radeon/pitcairn_pfp.bin");
62MODULE_FIRMWARE("radeon/pitcairn_me.bin");
63MODULE_FIRMWARE("radeon/pitcairn_ce.bin");
64MODULE_FIRMWARE("radeon/pitcairn_mc.bin");
65MODULE_FIRMWARE("radeon/pitcairn_rlc.bin");
66MODULE_FIRMWARE("radeon/pitcairn_smc.bin");
67
0f0de06c
AD
68MODULE_FIRMWARE("radeon/VERDE_pfp.bin");
69MODULE_FIRMWARE("radeon/VERDE_me.bin");
70MODULE_FIRMWARE("radeon/VERDE_ce.bin");
71MODULE_FIRMWARE("radeon/VERDE_mc.bin");
1ebe9280 72MODULE_FIRMWARE("radeon/VERDE_mc2.bin");
0f0de06c 73MODULE_FIRMWARE("radeon/VERDE_rlc.bin");
a9e61410 74MODULE_FIRMWARE("radeon/VERDE_smc.bin");
629bd33c
AD
75
76MODULE_FIRMWARE("radeon/verde_pfp.bin");
77MODULE_FIRMWARE("radeon/verde_me.bin");
78MODULE_FIRMWARE("radeon/verde_ce.bin");
79MODULE_FIRMWARE("radeon/verde_mc.bin");
80MODULE_FIRMWARE("radeon/verde_rlc.bin");
81MODULE_FIRMWARE("radeon/verde_smc.bin");
82
bcc7f5d2
AD
83MODULE_FIRMWARE("radeon/OLAND_pfp.bin");
84MODULE_FIRMWARE("radeon/OLAND_me.bin");
85MODULE_FIRMWARE("radeon/OLAND_ce.bin");
86MODULE_FIRMWARE("radeon/OLAND_mc.bin");
1ebe9280 87MODULE_FIRMWARE("radeon/OLAND_mc2.bin");
bcc7f5d2 88MODULE_FIRMWARE("radeon/OLAND_rlc.bin");
a9e61410 89MODULE_FIRMWARE("radeon/OLAND_smc.bin");
629bd33c
AD
90
91MODULE_FIRMWARE("radeon/oland_pfp.bin");
92MODULE_FIRMWARE("radeon/oland_me.bin");
93MODULE_FIRMWARE("radeon/oland_ce.bin");
94MODULE_FIRMWARE("radeon/oland_mc.bin");
95MODULE_FIRMWARE("radeon/oland_rlc.bin");
96MODULE_FIRMWARE("radeon/oland_smc.bin");
97
c04c00b4
AD
98MODULE_FIRMWARE("radeon/HAINAN_pfp.bin");
99MODULE_FIRMWARE("radeon/HAINAN_me.bin");
100MODULE_FIRMWARE("radeon/HAINAN_ce.bin");
101MODULE_FIRMWARE("radeon/HAINAN_mc.bin");
1ebe9280 102MODULE_FIRMWARE("radeon/HAINAN_mc2.bin");
c04c00b4 103MODULE_FIRMWARE("radeon/HAINAN_rlc.bin");
a9e61410 104MODULE_FIRMWARE("radeon/HAINAN_smc.bin");
0f0de06c 105
629bd33c
AD
106MODULE_FIRMWARE("radeon/hainan_pfp.bin");
107MODULE_FIRMWARE("radeon/hainan_me.bin");
108MODULE_FIRMWARE("radeon/hainan_ce.bin");
109MODULE_FIRMWARE("radeon/hainan_mc.bin");
110MODULE_FIRMWARE("radeon/hainan_rlc.bin");
111MODULE_FIRMWARE("radeon/hainan_smc.bin");
112
65fcf668 113static u32 si_get_cu_active_bitmap(struct radeon_device *rdev, u32 se, u32 sh);
b9d305df 114static void si_pcie_gen3_enable(struct radeon_device *rdev);
e0bcf165 115static void si_program_aspm(struct radeon_device *rdev);
1fd11777
AD
116extern void sumo_rlc_fini(struct radeon_device *rdev);
117extern int sumo_rlc_init(struct radeon_device *rdev);
25a857fb
AD
118extern int r600_ih_ring_alloc(struct radeon_device *rdev);
119extern void r600_ih_ring_fini(struct radeon_device *rdev);
0a96d72b 120extern void evergreen_fix_pci_max_read_req_size(struct radeon_device *rdev);
c476dde2
AD
121extern void evergreen_mc_stop(struct radeon_device *rdev, struct evergreen_mc_save *save);
122extern void evergreen_mc_resume(struct radeon_device *rdev, struct evergreen_mc_save *save);
ca7db22b 123extern u32 evergreen_get_number_of_dram_channels(struct radeon_device *rdev);
1c534671 124extern void evergreen_print_gpu_status_regs(struct radeon_device *rdev);
014bb209 125extern bool evergreen_is_display_hung(struct radeon_device *rdev);
811e4d58
AD
126static void si_enable_gui_idle_interrupt(struct radeon_device *rdev,
127 bool enable);
4a5c8ea5
AD
128static void si_init_pg(struct radeon_device *rdev);
129static void si_init_cg(struct radeon_device *rdev);
a6f4ae8d
AD
130static void si_fini_pg(struct radeon_device *rdev);
131static void si_fini_cg(struct radeon_device *rdev);
132static void si_rlc_stop(struct radeon_device *rdev);
0a96d72b 133
6d8cf000
AD
134static const u32 verde_rlc_save_restore_register_list[] =
135{
136 (0x8000 << 16) | (0x98f4 >> 2),
137 0x00000000,
138 (0x8040 << 16) | (0x98f4 >> 2),
139 0x00000000,
140 (0x8000 << 16) | (0xe80 >> 2),
141 0x00000000,
142 (0x8040 << 16) | (0xe80 >> 2),
143 0x00000000,
144 (0x8000 << 16) | (0x89bc >> 2),
145 0x00000000,
146 (0x8040 << 16) | (0x89bc >> 2),
147 0x00000000,
148 (0x8000 << 16) | (0x8c1c >> 2),
149 0x00000000,
150 (0x8040 << 16) | (0x8c1c >> 2),
151 0x00000000,
152 (0x9c00 << 16) | (0x98f0 >> 2),
153 0x00000000,
154 (0x9c00 << 16) | (0xe7c >> 2),
155 0x00000000,
156 (0x8000 << 16) | (0x9148 >> 2),
157 0x00000000,
158 (0x8040 << 16) | (0x9148 >> 2),
159 0x00000000,
160 (0x9c00 << 16) | (0x9150 >> 2),
161 0x00000000,
162 (0x9c00 << 16) | (0x897c >> 2),
163 0x00000000,
164 (0x9c00 << 16) | (0x8d8c >> 2),
165 0x00000000,
166 (0x9c00 << 16) | (0xac54 >> 2),
167 0X00000000,
168 0x3,
169 (0x9c00 << 16) | (0x98f8 >> 2),
170 0x00000000,
171 (0x9c00 << 16) | (0x9910 >> 2),
172 0x00000000,
173 (0x9c00 << 16) | (0x9914 >> 2),
174 0x00000000,
175 (0x9c00 << 16) | (0x9918 >> 2),
176 0x00000000,
177 (0x9c00 << 16) | (0x991c >> 2),
178 0x00000000,
179 (0x9c00 << 16) | (0x9920 >> 2),
180 0x00000000,
181 (0x9c00 << 16) | (0x9924 >> 2),
182 0x00000000,
183 (0x9c00 << 16) | (0x9928 >> 2),
184 0x00000000,
185 (0x9c00 << 16) | (0x992c >> 2),
186 0x00000000,
187 (0x9c00 << 16) | (0x9930 >> 2),
188 0x00000000,
189 (0x9c00 << 16) | (0x9934 >> 2),
190 0x00000000,
191 (0x9c00 << 16) | (0x9938 >> 2),
192 0x00000000,
193 (0x9c00 << 16) | (0x993c >> 2),
194 0x00000000,
195 (0x9c00 << 16) | (0x9940 >> 2),
196 0x00000000,
197 (0x9c00 << 16) | (0x9944 >> 2),
198 0x00000000,
199 (0x9c00 << 16) | (0x9948 >> 2),
200 0x00000000,
201 (0x9c00 << 16) | (0x994c >> 2),
202 0x00000000,
203 (0x9c00 << 16) | (0x9950 >> 2),
204 0x00000000,
205 (0x9c00 << 16) | (0x9954 >> 2),
206 0x00000000,
207 (0x9c00 << 16) | (0x9958 >> 2),
208 0x00000000,
209 (0x9c00 << 16) | (0x995c >> 2),
210 0x00000000,
211 (0x9c00 << 16) | (0x9960 >> 2),
212 0x00000000,
213 (0x9c00 << 16) | (0x9964 >> 2),
214 0x00000000,
215 (0x9c00 << 16) | (0x9968 >> 2),
216 0x00000000,
217 (0x9c00 << 16) | (0x996c >> 2),
218 0x00000000,
219 (0x9c00 << 16) | (0x9970 >> 2),
220 0x00000000,
221 (0x9c00 << 16) | (0x9974 >> 2),
222 0x00000000,
223 (0x9c00 << 16) | (0x9978 >> 2),
224 0x00000000,
225 (0x9c00 << 16) | (0x997c >> 2),
226 0x00000000,
227 (0x9c00 << 16) | (0x9980 >> 2),
228 0x00000000,
229 (0x9c00 << 16) | (0x9984 >> 2),
230 0x00000000,
231 (0x9c00 << 16) | (0x9988 >> 2),
232 0x00000000,
233 (0x9c00 << 16) | (0x998c >> 2),
234 0x00000000,
235 (0x9c00 << 16) | (0x8c00 >> 2),
236 0x00000000,
237 (0x9c00 << 16) | (0x8c14 >> 2),
238 0x00000000,
239 (0x9c00 << 16) | (0x8c04 >> 2),
240 0x00000000,
241 (0x9c00 << 16) | (0x8c08 >> 2),
242 0x00000000,
243 (0x8000 << 16) | (0x9b7c >> 2),
244 0x00000000,
245 (0x8040 << 16) | (0x9b7c >> 2),
246 0x00000000,
247 (0x8000 << 16) | (0xe84 >> 2),
248 0x00000000,
249 (0x8040 << 16) | (0xe84 >> 2),
250 0x00000000,
251 (0x8000 << 16) | (0x89c0 >> 2),
252 0x00000000,
253 (0x8040 << 16) | (0x89c0 >> 2),
254 0x00000000,
255 (0x8000 << 16) | (0x914c >> 2),
256 0x00000000,
257 (0x8040 << 16) | (0x914c >> 2),
258 0x00000000,
259 (0x8000 << 16) | (0x8c20 >> 2),
260 0x00000000,
261 (0x8040 << 16) | (0x8c20 >> 2),
262 0x00000000,
263 (0x8000 << 16) | (0x9354 >> 2),
264 0x00000000,
265 (0x8040 << 16) | (0x9354 >> 2),
266 0x00000000,
267 (0x9c00 << 16) | (0x9060 >> 2),
268 0x00000000,
269 (0x9c00 << 16) | (0x9364 >> 2),
270 0x00000000,
271 (0x9c00 << 16) | (0x9100 >> 2),
272 0x00000000,
273 (0x9c00 << 16) | (0x913c >> 2),
274 0x00000000,
275 (0x8000 << 16) | (0x90e0 >> 2),
276 0x00000000,
277 (0x8000 << 16) | (0x90e4 >> 2),
278 0x00000000,
279 (0x8000 << 16) | (0x90e8 >> 2),
280 0x00000000,
281 (0x8040 << 16) | (0x90e0 >> 2),
282 0x00000000,
283 (0x8040 << 16) | (0x90e4 >> 2),
284 0x00000000,
285 (0x8040 << 16) | (0x90e8 >> 2),
286 0x00000000,
287 (0x9c00 << 16) | (0x8bcc >> 2),
288 0x00000000,
289 (0x9c00 << 16) | (0x8b24 >> 2),
290 0x00000000,
291 (0x9c00 << 16) | (0x88c4 >> 2),
292 0x00000000,
293 (0x9c00 << 16) | (0x8e50 >> 2),
294 0x00000000,
295 (0x9c00 << 16) | (0x8c0c >> 2),
296 0x00000000,
297 (0x9c00 << 16) | (0x8e58 >> 2),
298 0x00000000,
299 (0x9c00 << 16) | (0x8e5c >> 2),
300 0x00000000,
301 (0x9c00 << 16) | (0x9508 >> 2),
302 0x00000000,
303 (0x9c00 << 16) | (0x950c >> 2),
304 0x00000000,
305 (0x9c00 << 16) | (0x9494 >> 2),
306 0x00000000,
307 (0x9c00 << 16) | (0xac0c >> 2),
308 0x00000000,
309 (0x9c00 << 16) | (0xac10 >> 2),
310 0x00000000,
311 (0x9c00 << 16) | (0xac14 >> 2),
312 0x00000000,
313 (0x9c00 << 16) | (0xae00 >> 2),
314 0x00000000,
315 (0x9c00 << 16) | (0xac08 >> 2),
316 0x00000000,
317 (0x9c00 << 16) | (0x88d4 >> 2),
318 0x00000000,
319 (0x9c00 << 16) | (0x88c8 >> 2),
320 0x00000000,
321 (0x9c00 << 16) | (0x88cc >> 2),
322 0x00000000,
323 (0x9c00 << 16) | (0x89b0 >> 2),
324 0x00000000,
325 (0x9c00 << 16) | (0x8b10 >> 2),
326 0x00000000,
327 (0x9c00 << 16) | (0x8a14 >> 2),
328 0x00000000,
329 (0x9c00 << 16) | (0x9830 >> 2),
330 0x00000000,
331 (0x9c00 << 16) | (0x9834 >> 2),
332 0x00000000,
333 (0x9c00 << 16) | (0x9838 >> 2),
334 0x00000000,
335 (0x9c00 << 16) | (0x9a10 >> 2),
336 0x00000000,
337 (0x8000 << 16) | (0x9870 >> 2),
338 0x00000000,
339 (0x8000 << 16) | (0x9874 >> 2),
340 0x00000000,
341 (0x8001 << 16) | (0x9870 >> 2),
342 0x00000000,
343 (0x8001 << 16) | (0x9874 >> 2),
344 0x00000000,
345 (0x8040 << 16) | (0x9870 >> 2),
346 0x00000000,
347 (0x8040 << 16) | (0x9874 >> 2),
348 0x00000000,
349 (0x8041 << 16) | (0x9870 >> 2),
350 0x00000000,
351 (0x8041 << 16) | (0x9874 >> 2),
352 0x00000000,
353 0x00000000
354};
355
205996c0
AD
356static const u32 tahiti_golden_rlc_registers[] =
357{
358 0xc424, 0xffffffff, 0x00601005,
359 0xc47c, 0xffffffff, 0x10104040,
360 0xc488, 0xffffffff, 0x0100000a,
361 0xc314, 0xffffffff, 0x00000800,
362 0xc30c, 0xffffffff, 0x800000f4,
363 0xf4a8, 0xffffffff, 0x00000000
364};
365
366static const u32 tahiti_golden_registers[] =
367{
368 0x9a10, 0x00010000, 0x00018208,
369 0x9830, 0xffffffff, 0x00000000,
370 0x9834, 0xf00fffff, 0x00000400,
371 0x9838, 0x0002021c, 0x00020200,
372 0xc78, 0x00000080, 0x00000000,
373 0xd030, 0x000300c0, 0x00800040,
374 0xd830, 0x000300c0, 0x00800040,
375 0x5bb0, 0x000000f0, 0x00000070,
376 0x5bc0, 0x00200000, 0x50100000,
377 0x7030, 0x31000311, 0x00000011,
378 0x277c, 0x00000003, 0x000007ff,
379 0x240c, 0x000007ff, 0x00000000,
380 0x8a14, 0xf000001f, 0x00000007,
381 0x8b24, 0xffffffff, 0x00ffffff,
382 0x8b10, 0x0000ff0f, 0x00000000,
383 0x28a4c, 0x07ffffff, 0x4e000000,
384 0x28350, 0x3f3f3fff, 0x2a00126a,
385 0x30, 0x000000ff, 0x0040,
386 0x34, 0x00000040, 0x00004040,
387 0x9100, 0x07ffffff, 0x03000000,
388 0x8e88, 0x01ff1f3f, 0x00000000,
389 0x8e84, 0x01ff1f3f, 0x00000000,
390 0x9060, 0x0000007f, 0x00000020,
391 0x9508, 0x00010000, 0x00010000,
392 0xac14, 0x00000200, 0x000002fb,
393 0xac10, 0xffffffff, 0x0000543b,
394 0xac0c, 0xffffffff, 0xa9210876,
395 0x88d0, 0xffffffff, 0x000fff40,
396 0x88d4, 0x0000001f, 0x00000010,
397 0x1410, 0x20000000, 0x20fffed8,
398 0x15c0, 0x000c0fc0, 0x000c0400
399};
400
401static const u32 tahiti_golden_registers2[] =
402{
403 0xc64, 0x00000001, 0x00000001
404};
405
406static const u32 pitcairn_golden_rlc_registers[] =
407{
408 0xc424, 0xffffffff, 0x00601004,
409 0xc47c, 0xffffffff, 0x10102020,
410 0xc488, 0xffffffff, 0x01000020,
411 0xc314, 0xffffffff, 0x00000800,
412 0xc30c, 0xffffffff, 0x800000a4
413};
414
415static const u32 pitcairn_golden_registers[] =
416{
417 0x9a10, 0x00010000, 0x00018208,
418 0x9830, 0xffffffff, 0x00000000,
419 0x9834, 0xf00fffff, 0x00000400,
420 0x9838, 0x0002021c, 0x00020200,
421 0xc78, 0x00000080, 0x00000000,
422 0xd030, 0x000300c0, 0x00800040,
423 0xd830, 0x000300c0, 0x00800040,
424 0x5bb0, 0x000000f0, 0x00000070,
425 0x5bc0, 0x00200000, 0x50100000,
426 0x7030, 0x31000311, 0x00000011,
427 0x2ae4, 0x00073ffe, 0x000022a2,
428 0x240c, 0x000007ff, 0x00000000,
429 0x8a14, 0xf000001f, 0x00000007,
430 0x8b24, 0xffffffff, 0x00ffffff,
431 0x8b10, 0x0000ff0f, 0x00000000,
432 0x28a4c, 0x07ffffff, 0x4e000000,
433 0x28350, 0x3f3f3fff, 0x2a00126a,
434 0x30, 0x000000ff, 0x0040,
435 0x34, 0x00000040, 0x00004040,
436 0x9100, 0x07ffffff, 0x03000000,
437 0x9060, 0x0000007f, 0x00000020,
438 0x9508, 0x00010000, 0x00010000,
439 0xac14, 0x000003ff, 0x000000f7,
440 0xac10, 0xffffffff, 0x00000000,
441 0xac0c, 0xffffffff, 0x32761054,
442 0x88d4, 0x0000001f, 0x00000010,
443 0x15c0, 0x000c0fc0, 0x000c0400
444};
445
446static const u32 verde_golden_rlc_registers[] =
447{
448 0xc424, 0xffffffff, 0x033f1005,
449 0xc47c, 0xffffffff, 0x10808020,
450 0xc488, 0xffffffff, 0x00800008,
451 0xc314, 0xffffffff, 0x00001000,
452 0xc30c, 0xffffffff, 0x80010014
453};
454
455static const u32 verde_golden_registers[] =
456{
457 0x9a10, 0x00010000, 0x00018208,
458 0x9830, 0xffffffff, 0x00000000,
459 0x9834, 0xf00fffff, 0x00000400,
460 0x9838, 0x0002021c, 0x00020200,
461 0xc78, 0x00000080, 0x00000000,
462 0xd030, 0x000300c0, 0x00800040,
463 0xd030, 0x000300c0, 0x00800040,
464 0xd830, 0x000300c0, 0x00800040,
465 0xd830, 0x000300c0, 0x00800040,
466 0x5bb0, 0x000000f0, 0x00000070,
467 0x5bc0, 0x00200000, 0x50100000,
468 0x7030, 0x31000311, 0x00000011,
469 0x2ae4, 0x00073ffe, 0x000022a2,
470 0x2ae4, 0x00073ffe, 0x000022a2,
471 0x2ae4, 0x00073ffe, 0x000022a2,
472 0x240c, 0x000007ff, 0x00000000,
473 0x240c, 0x000007ff, 0x00000000,
474 0x240c, 0x000007ff, 0x00000000,
475 0x8a14, 0xf000001f, 0x00000007,
476 0x8a14, 0xf000001f, 0x00000007,
477 0x8a14, 0xf000001f, 0x00000007,
478 0x8b24, 0xffffffff, 0x00ffffff,
479 0x8b10, 0x0000ff0f, 0x00000000,
480 0x28a4c, 0x07ffffff, 0x4e000000,
481 0x28350, 0x3f3f3fff, 0x0000124a,
482 0x28350, 0x3f3f3fff, 0x0000124a,
483 0x28350, 0x3f3f3fff, 0x0000124a,
484 0x30, 0x000000ff, 0x0040,
485 0x34, 0x00000040, 0x00004040,
486 0x9100, 0x07ffffff, 0x03000000,
487 0x9100, 0x07ffffff, 0x03000000,
488 0x8e88, 0x01ff1f3f, 0x00000000,
489 0x8e88, 0x01ff1f3f, 0x00000000,
490 0x8e88, 0x01ff1f3f, 0x00000000,
491 0x8e84, 0x01ff1f3f, 0x00000000,
492 0x8e84, 0x01ff1f3f, 0x00000000,
493 0x8e84, 0x01ff1f3f, 0x00000000,
494 0x9060, 0x0000007f, 0x00000020,
495 0x9508, 0x00010000, 0x00010000,
496 0xac14, 0x000003ff, 0x00000003,
497 0xac14, 0x000003ff, 0x00000003,
498 0xac14, 0x000003ff, 0x00000003,
499 0xac10, 0xffffffff, 0x00000000,
500 0xac10, 0xffffffff, 0x00000000,
501 0xac10, 0xffffffff, 0x00000000,
502 0xac0c, 0xffffffff, 0x00001032,
503 0xac0c, 0xffffffff, 0x00001032,
504 0xac0c, 0xffffffff, 0x00001032,
505 0x88d4, 0x0000001f, 0x00000010,
506 0x88d4, 0x0000001f, 0x00000010,
507 0x88d4, 0x0000001f, 0x00000010,
508 0x15c0, 0x000c0fc0, 0x000c0400
509};
510
511static const u32 oland_golden_rlc_registers[] =
512{
513 0xc424, 0xffffffff, 0x00601005,
514 0xc47c, 0xffffffff, 0x10104040,
515 0xc488, 0xffffffff, 0x0100000a,
516 0xc314, 0xffffffff, 0x00000800,
517 0xc30c, 0xffffffff, 0x800000f4
518};
519
520static const u32 oland_golden_registers[] =
521{
522 0x9a10, 0x00010000, 0x00018208,
523 0x9830, 0xffffffff, 0x00000000,
524 0x9834, 0xf00fffff, 0x00000400,
525 0x9838, 0x0002021c, 0x00020200,
526 0xc78, 0x00000080, 0x00000000,
527 0xd030, 0x000300c0, 0x00800040,
528 0xd830, 0x000300c0, 0x00800040,
529 0x5bb0, 0x000000f0, 0x00000070,
530 0x5bc0, 0x00200000, 0x50100000,
531 0x7030, 0x31000311, 0x00000011,
532 0x2ae4, 0x00073ffe, 0x000022a2,
533 0x240c, 0x000007ff, 0x00000000,
534 0x8a14, 0xf000001f, 0x00000007,
535 0x8b24, 0xffffffff, 0x00ffffff,
536 0x8b10, 0x0000ff0f, 0x00000000,
537 0x28a4c, 0x07ffffff, 0x4e000000,
538 0x28350, 0x3f3f3fff, 0x00000082,
539 0x30, 0x000000ff, 0x0040,
540 0x34, 0x00000040, 0x00004040,
541 0x9100, 0x07ffffff, 0x03000000,
542 0x9060, 0x0000007f, 0x00000020,
543 0x9508, 0x00010000, 0x00010000,
544 0xac14, 0x000003ff, 0x000000f3,
545 0xac10, 0xffffffff, 0x00000000,
546 0xac0c, 0xffffffff, 0x00003210,
547 0x88d4, 0x0000001f, 0x00000010,
548 0x15c0, 0x000c0fc0, 0x000c0400
549};
550
fffbdda4
AD
551static const u32 hainan_golden_registers[] =
552{
553 0x9a10, 0x00010000, 0x00018208,
554 0x9830, 0xffffffff, 0x00000000,
555 0x9834, 0xf00fffff, 0x00000400,
556 0x9838, 0x0002021c, 0x00020200,
557 0xd0c0, 0xff000fff, 0x00000100,
558 0xd030, 0x000300c0, 0x00800040,
559 0xd8c0, 0xff000fff, 0x00000100,
560 0xd830, 0x000300c0, 0x00800040,
561 0x2ae4, 0x00073ffe, 0x000022a2,
562 0x240c, 0x000007ff, 0x00000000,
563 0x8a14, 0xf000001f, 0x00000007,
564 0x8b24, 0xffffffff, 0x00ffffff,
565 0x8b10, 0x0000ff0f, 0x00000000,
566 0x28a4c, 0x07ffffff, 0x4e000000,
567 0x28350, 0x3f3f3fff, 0x00000000,
568 0x30, 0x000000ff, 0x0040,
569 0x34, 0x00000040, 0x00004040,
570 0x9100, 0x03e00000, 0x03600000,
571 0x9060, 0x0000007f, 0x00000020,
572 0x9508, 0x00010000, 0x00010000,
573 0xac14, 0x000003ff, 0x000000f1,
574 0xac10, 0xffffffff, 0x00000000,
575 0xac0c, 0xffffffff, 0x00003210,
576 0x88d4, 0x0000001f, 0x00000010,
577 0x15c0, 0x000c0fc0, 0x000c0400
578};
579
580static const u32 hainan_golden_registers2[] =
581{
582 0x98f8, 0xffffffff, 0x02010001
583};
584
205996c0
AD
585static const u32 tahiti_mgcg_cgcg_init[] =
586{
587 0xc400, 0xffffffff, 0xfffffffc,
588 0x802c, 0xffffffff, 0xe0000000,
589 0x9a60, 0xffffffff, 0x00000100,
590 0x92a4, 0xffffffff, 0x00000100,
591 0xc164, 0xffffffff, 0x00000100,
592 0x9774, 0xffffffff, 0x00000100,
593 0x8984, 0xffffffff, 0x06000100,
594 0x8a18, 0xffffffff, 0x00000100,
595 0x92a0, 0xffffffff, 0x00000100,
596 0xc380, 0xffffffff, 0x00000100,
597 0x8b28, 0xffffffff, 0x00000100,
598 0x9144, 0xffffffff, 0x00000100,
599 0x8d88, 0xffffffff, 0x00000100,
600 0x8d8c, 0xffffffff, 0x00000100,
601 0x9030, 0xffffffff, 0x00000100,
602 0x9034, 0xffffffff, 0x00000100,
603 0x9038, 0xffffffff, 0x00000100,
604 0x903c, 0xffffffff, 0x00000100,
605 0xad80, 0xffffffff, 0x00000100,
606 0xac54, 0xffffffff, 0x00000100,
607 0x897c, 0xffffffff, 0x06000100,
608 0x9868, 0xffffffff, 0x00000100,
609 0x9510, 0xffffffff, 0x00000100,
610 0xaf04, 0xffffffff, 0x00000100,
611 0xae04, 0xffffffff, 0x00000100,
612 0x949c, 0xffffffff, 0x00000100,
613 0x802c, 0xffffffff, 0xe0000000,
614 0x9160, 0xffffffff, 0x00010000,
615 0x9164, 0xffffffff, 0x00030002,
616 0x9168, 0xffffffff, 0x00040007,
617 0x916c, 0xffffffff, 0x00060005,
618 0x9170, 0xffffffff, 0x00090008,
619 0x9174, 0xffffffff, 0x00020001,
620 0x9178, 0xffffffff, 0x00040003,
621 0x917c, 0xffffffff, 0x00000007,
622 0x9180, 0xffffffff, 0x00060005,
623 0x9184, 0xffffffff, 0x00090008,
624 0x9188, 0xffffffff, 0x00030002,
625 0x918c, 0xffffffff, 0x00050004,
626 0x9190, 0xffffffff, 0x00000008,
627 0x9194, 0xffffffff, 0x00070006,
628 0x9198, 0xffffffff, 0x000a0009,
629 0x919c, 0xffffffff, 0x00040003,
630 0x91a0, 0xffffffff, 0x00060005,
631 0x91a4, 0xffffffff, 0x00000009,
632 0x91a8, 0xffffffff, 0x00080007,
633 0x91ac, 0xffffffff, 0x000b000a,
634 0x91b0, 0xffffffff, 0x00050004,
635 0x91b4, 0xffffffff, 0x00070006,
636 0x91b8, 0xffffffff, 0x0008000b,
637 0x91bc, 0xffffffff, 0x000a0009,
638 0x91c0, 0xffffffff, 0x000d000c,
639 0x91c4, 0xffffffff, 0x00060005,
640 0x91c8, 0xffffffff, 0x00080007,
641 0x91cc, 0xffffffff, 0x0000000b,
642 0x91d0, 0xffffffff, 0x000a0009,
643 0x91d4, 0xffffffff, 0x000d000c,
644 0x91d8, 0xffffffff, 0x00070006,
645 0x91dc, 0xffffffff, 0x00090008,
646 0x91e0, 0xffffffff, 0x0000000c,
647 0x91e4, 0xffffffff, 0x000b000a,
648 0x91e8, 0xffffffff, 0x000e000d,
649 0x91ec, 0xffffffff, 0x00080007,
650 0x91f0, 0xffffffff, 0x000a0009,
651 0x91f4, 0xffffffff, 0x0000000d,
652 0x91f8, 0xffffffff, 0x000c000b,
653 0x91fc, 0xffffffff, 0x000f000e,
654 0x9200, 0xffffffff, 0x00090008,
655 0x9204, 0xffffffff, 0x000b000a,
656 0x9208, 0xffffffff, 0x000c000f,
657 0x920c, 0xffffffff, 0x000e000d,
658 0x9210, 0xffffffff, 0x00110010,
659 0x9214, 0xffffffff, 0x000a0009,
660 0x9218, 0xffffffff, 0x000c000b,
661 0x921c, 0xffffffff, 0x0000000f,
662 0x9220, 0xffffffff, 0x000e000d,
663 0x9224, 0xffffffff, 0x00110010,
664 0x9228, 0xffffffff, 0x000b000a,
665 0x922c, 0xffffffff, 0x000d000c,
666 0x9230, 0xffffffff, 0x00000010,
667 0x9234, 0xffffffff, 0x000f000e,
668 0x9238, 0xffffffff, 0x00120011,
669 0x923c, 0xffffffff, 0x000c000b,
670 0x9240, 0xffffffff, 0x000e000d,
671 0x9244, 0xffffffff, 0x00000011,
672 0x9248, 0xffffffff, 0x0010000f,
673 0x924c, 0xffffffff, 0x00130012,
674 0x9250, 0xffffffff, 0x000d000c,
675 0x9254, 0xffffffff, 0x000f000e,
676 0x9258, 0xffffffff, 0x00100013,
677 0x925c, 0xffffffff, 0x00120011,
678 0x9260, 0xffffffff, 0x00150014,
679 0x9264, 0xffffffff, 0x000e000d,
680 0x9268, 0xffffffff, 0x0010000f,
681 0x926c, 0xffffffff, 0x00000013,
682 0x9270, 0xffffffff, 0x00120011,
683 0x9274, 0xffffffff, 0x00150014,
684 0x9278, 0xffffffff, 0x000f000e,
685 0x927c, 0xffffffff, 0x00110010,
686 0x9280, 0xffffffff, 0x00000014,
687 0x9284, 0xffffffff, 0x00130012,
688 0x9288, 0xffffffff, 0x00160015,
689 0x928c, 0xffffffff, 0x0010000f,
690 0x9290, 0xffffffff, 0x00120011,
691 0x9294, 0xffffffff, 0x00000015,
692 0x9298, 0xffffffff, 0x00140013,
693 0x929c, 0xffffffff, 0x00170016,
694 0x9150, 0xffffffff, 0x96940200,
695 0x8708, 0xffffffff, 0x00900100,
696 0xc478, 0xffffffff, 0x00000080,
697 0xc404, 0xffffffff, 0x0020003f,
698 0x30, 0xffffffff, 0x0000001c,
699 0x34, 0x000f0000, 0x000f0000,
700 0x160c, 0xffffffff, 0x00000100,
701 0x1024, 0xffffffff, 0x00000100,
702 0x102c, 0x00000101, 0x00000000,
703 0x20a8, 0xffffffff, 0x00000104,
704 0x264c, 0x000c0000, 0x000c0000,
705 0x2648, 0x000c0000, 0x000c0000,
706 0x55e4, 0xff000fff, 0x00000100,
707 0x55e8, 0x00000001, 0x00000001,
708 0x2f50, 0x00000001, 0x00000001,
709 0x30cc, 0xc0000fff, 0x00000104,
710 0xc1e4, 0x00000001, 0x00000001,
711 0xd0c0, 0xfffffff0, 0x00000100,
712 0xd8c0, 0xfffffff0, 0x00000100
713};
714
715static const u32 pitcairn_mgcg_cgcg_init[] =
716{
717 0xc400, 0xffffffff, 0xfffffffc,
718 0x802c, 0xffffffff, 0xe0000000,
719 0x9a60, 0xffffffff, 0x00000100,
720 0x92a4, 0xffffffff, 0x00000100,
721 0xc164, 0xffffffff, 0x00000100,
722 0x9774, 0xffffffff, 0x00000100,
723 0x8984, 0xffffffff, 0x06000100,
724 0x8a18, 0xffffffff, 0x00000100,
725 0x92a0, 0xffffffff, 0x00000100,
726 0xc380, 0xffffffff, 0x00000100,
727 0x8b28, 0xffffffff, 0x00000100,
728 0x9144, 0xffffffff, 0x00000100,
729 0x8d88, 0xffffffff, 0x00000100,
730 0x8d8c, 0xffffffff, 0x00000100,
731 0x9030, 0xffffffff, 0x00000100,
732 0x9034, 0xffffffff, 0x00000100,
733 0x9038, 0xffffffff, 0x00000100,
734 0x903c, 0xffffffff, 0x00000100,
735 0xad80, 0xffffffff, 0x00000100,
736 0xac54, 0xffffffff, 0x00000100,
737 0x897c, 0xffffffff, 0x06000100,
738 0x9868, 0xffffffff, 0x00000100,
739 0x9510, 0xffffffff, 0x00000100,
740 0xaf04, 0xffffffff, 0x00000100,
741 0xae04, 0xffffffff, 0x00000100,
742 0x949c, 0xffffffff, 0x00000100,
743 0x802c, 0xffffffff, 0xe0000000,
744 0x9160, 0xffffffff, 0x00010000,
745 0x9164, 0xffffffff, 0x00030002,
746 0x9168, 0xffffffff, 0x00040007,
747 0x916c, 0xffffffff, 0x00060005,
748 0x9170, 0xffffffff, 0x00090008,
749 0x9174, 0xffffffff, 0x00020001,
750 0x9178, 0xffffffff, 0x00040003,
751 0x917c, 0xffffffff, 0x00000007,
752 0x9180, 0xffffffff, 0x00060005,
753 0x9184, 0xffffffff, 0x00090008,
754 0x9188, 0xffffffff, 0x00030002,
755 0x918c, 0xffffffff, 0x00050004,
756 0x9190, 0xffffffff, 0x00000008,
757 0x9194, 0xffffffff, 0x00070006,
758 0x9198, 0xffffffff, 0x000a0009,
759 0x919c, 0xffffffff, 0x00040003,
760 0x91a0, 0xffffffff, 0x00060005,
761 0x91a4, 0xffffffff, 0x00000009,
762 0x91a8, 0xffffffff, 0x00080007,
763 0x91ac, 0xffffffff, 0x000b000a,
764 0x91b0, 0xffffffff, 0x00050004,
765 0x91b4, 0xffffffff, 0x00070006,
766 0x91b8, 0xffffffff, 0x0008000b,
767 0x91bc, 0xffffffff, 0x000a0009,
768 0x91c0, 0xffffffff, 0x000d000c,
769 0x9200, 0xffffffff, 0x00090008,
770 0x9204, 0xffffffff, 0x000b000a,
771 0x9208, 0xffffffff, 0x000c000f,
772 0x920c, 0xffffffff, 0x000e000d,
773 0x9210, 0xffffffff, 0x00110010,
774 0x9214, 0xffffffff, 0x000a0009,
775 0x9218, 0xffffffff, 0x000c000b,
776 0x921c, 0xffffffff, 0x0000000f,
777 0x9220, 0xffffffff, 0x000e000d,
778 0x9224, 0xffffffff, 0x00110010,
779 0x9228, 0xffffffff, 0x000b000a,
780 0x922c, 0xffffffff, 0x000d000c,
781 0x9230, 0xffffffff, 0x00000010,
782 0x9234, 0xffffffff, 0x000f000e,
783 0x9238, 0xffffffff, 0x00120011,
784 0x923c, 0xffffffff, 0x000c000b,
785 0x9240, 0xffffffff, 0x000e000d,
786 0x9244, 0xffffffff, 0x00000011,
787 0x9248, 0xffffffff, 0x0010000f,
788 0x924c, 0xffffffff, 0x00130012,
789 0x9250, 0xffffffff, 0x000d000c,
790 0x9254, 0xffffffff, 0x000f000e,
791 0x9258, 0xffffffff, 0x00100013,
792 0x925c, 0xffffffff, 0x00120011,
793 0x9260, 0xffffffff, 0x00150014,
794 0x9150, 0xffffffff, 0x96940200,
795 0x8708, 0xffffffff, 0x00900100,
796 0xc478, 0xffffffff, 0x00000080,
797 0xc404, 0xffffffff, 0x0020003f,
798 0x30, 0xffffffff, 0x0000001c,
799 0x34, 0x000f0000, 0x000f0000,
800 0x160c, 0xffffffff, 0x00000100,
801 0x1024, 0xffffffff, 0x00000100,
802 0x102c, 0x00000101, 0x00000000,
803 0x20a8, 0xffffffff, 0x00000104,
804 0x55e4, 0xff000fff, 0x00000100,
805 0x55e8, 0x00000001, 0x00000001,
806 0x2f50, 0x00000001, 0x00000001,
807 0x30cc, 0xc0000fff, 0x00000104,
808 0xc1e4, 0x00000001, 0x00000001,
809 0xd0c0, 0xfffffff0, 0x00000100,
810 0xd8c0, 0xfffffff0, 0x00000100
811};
812
813static const u32 verde_mgcg_cgcg_init[] =
814{
815 0xc400, 0xffffffff, 0xfffffffc,
816 0x802c, 0xffffffff, 0xe0000000,
817 0x9a60, 0xffffffff, 0x00000100,
818 0x92a4, 0xffffffff, 0x00000100,
819 0xc164, 0xffffffff, 0x00000100,
820 0x9774, 0xffffffff, 0x00000100,
821 0x8984, 0xffffffff, 0x06000100,
822 0x8a18, 0xffffffff, 0x00000100,
823 0x92a0, 0xffffffff, 0x00000100,
824 0xc380, 0xffffffff, 0x00000100,
825 0x8b28, 0xffffffff, 0x00000100,
826 0x9144, 0xffffffff, 0x00000100,
827 0x8d88, 0xffffffff, 0x00000100,
828 0x8d8c, 0xffffffff, 0x00000100,
829 0x9030, 0xffffffff, 0x00000100,
830 0x9034, 0xffffffff, 0x00000100,
831 0x9038, 0xffffffff, 0x00000100,
832 0x903c, 0xffffffff, 0x00000100,
833 0xad80, 0xffffffff, 0x00000100,
834 0xac54, 0xffffffff, 0x00000100,
835 0x897c, 0xffffffff, 0x06000100,
836 0x9868, 0xffffffff, 0x00000100,
837 0x9510, 0xffffffff, 0x00000100,
838 0xaf04, 0xffffffff, 0x00000100,
839 0xae04, 0xffffffff, 0x00000100,
840 0x949c, 0xffffffff, 0x00000100,
841 0x802c, 0xffffffff, 0xe0000000,
842 0x9160, 0xffffffff, 0x00010000,
843 0x9164, 0xffffffff, 0x00030002,
844 0x9168, 0xffffffff, 0x00040007,
845 0x916c, 0xffffffff, 0x00060005,
846 0x9170, 0xffffffff, 0x00090008,
847 0x9174, 0xffffffff, 0x00020001,
848 0x9178, 0xffffffff, 0x00040003,
849 0x917c, 0xffffffff, 0x00000007,
850 0x9180, 0xffffffff, 0x00060005,
851 0x9184, 0xffffffff, 0x00090008,
852 0x9188, 0xffffffff, 0x00030002,
853 0x918c, 0xffffffff, 0x00050004,
854 0x9190, 0xffffffff, 0x00000008,
855 0x9194, 0xffffffff, 0x00070006,
856 0x9198, 0xffffffff, 0x000a0009,
857 0x919c, 0xffffffff, 0x00040003,
858 0x91a0, 0xffffffff, 0x00060005,
859 0x91a4, 0xffffffff, 0x00000009,
860 0x91a8, 0xffffffff, 0x00080007,
861 0x91ac, 0xffffffff, 0x000b000a,
862 0x91b0, 0xffffffff, 0x00050004,
863 0x91b4, 0xffffffff, 0x00070006,
864 0x91b8, 0xffffffff, 0x0008000b,
865 0x91bc, 0xffffffff, 0x000a0009,
866 0x91c0, 0xffffffff, 0x000d000c,
867 0x9200, 0xffffffff, 0x00090008,
868 0x9204, 0xffffffff, 0x000b000a,
869 0x9208, 0xffffffff, 0x000c000f,
870 0x920c, 0xffffffff, 0x000e000d,
871 0x9210, 0xffffffff, 0x00110010,
872 0x9214, 0xffffffff, 0x000a0009,
873 0x9218, 0xffffffff, 0x000c000b,
874 0x921c, 0xffffffff, 0x0000000f,
875 0x9220, 0xffffffff, 0x000e000d,
876 0x9224, 0xffffffff, 0x00110010,
877 0x9228, 0xffffffff, 0x000b000a,
878 0x922c, 0xffffffff, 0x000d000c,
879 0x9230, 0xffffffff, 0x00000010,
880 0x9234, 0xffffffff, 0x000f000e,
881 0x9238, 0xffffffff, 0x00120011,
882 0x923c, 0xffffffff, 0x000c000b,
883 0x9240, 0xffffffff, 0x000e000d,
884 0x9244, 0xffffffff, 0x00000011,
885 0x9248, 0xffffffff, 0x0010000f,
886 0x924c, 0xffffffff, 0x00130012,
887 0x9250, 0xffffffff, 0x000d000c,
888 0x9254, 0xffffffff, 0x000f000e,
889 0x9258, 0xffffffff, 0x00100013,
890 0x925c, 0xffffffff, 0x00120011,
891 0x9260, 0xffffffff, 0x00150014,
892 0x9150, 0xffffffff, 0x96940200,
893 0x8708, 0xffffffff, 0x00900100,
894 0xc478, 0xffffffff, 0x00000080,
895 0xc404, 0xffffffff, 0x0020003f,
896 0x30, 0xffffffff, 0x0000001c,
897 0x34, 0x000f0000, 0x000f0000,
898 0x160c, 0xffffffff, 0x00000100,
899 0x1024, 0xffffffff, 0x00000100,
900 0x102c, 0x00000101, 0x00000000,
901 0x20a8, 0xffffffff, 0x00000104,
902 0x264c, 0x000c0000, 0x000c0000,
903 0x2648, 0x000c0000, 0x000c0000,
904 0x55e4, 0xff000fff, 0x00000100,
905 0x55e8, 0x00000001, 0x00000001,
906 0x2f50, 0x00000001, 0x00000001,
907 0x30cc, 0xc0000fff, 0x00000104,
908 0xc1e4, 0x00000001, 0x00000001,
909 0xd0c0, 0xfffffff0, 0x00000100,
910 0xd8c0, 0xfffffff0, 0x00000100
911};
912
913static const u32 oland_mgcg_cgcg_init[] =
914{
915 0xc400, 0xffffffff, 0xfffffffc,
916 0x802c, 0xffffffff, 0xe0000000,
917 0x9a60, 0xffffffff, 0x00000100,
918 0x92a4, 0xffffffff, 0x00000100,
919 0xc164, 0xffffffff, 0x00000100,
920 0x9774, 0xffffffff, 0x00000100,
921 0x8984, 0xffffffff, 0x06000100,
922 0x8a18, 0xffffffff, 0x00000100,
923 0x92a0, 0xffffffff, 0x00000100,
924 0xc380, 0xffffffff, 0x00000100,
925 0x8b28, 0xffffffff, 0x00000100,
926 0x9144, 0xffffffff, 0x00000100,
927 0x8d88, 0xffffffff, 0x00000100,
928 0x8d8c, 0xffffffff, 0x00000100,
929 0x9030, 0xffffffff, 0x00000100,
930 0x9034, 0xffffffff, 0x00000100,
931 0x9038, 0xffffffff, 0x00000100,
932 0x903c, 0xffffffff, 0x00000100,
933 0xad80, 0xffffffff, 0x00000100,
934 0xac54, 0xffffffff, 0x00000100,
935 0x897c, 0xffffffff, 0x06000100,
936 0x9868, 0xffffffff, 0x00000100,
937 0x9510, 0xffffffff, 0x00000100,
938 0xaf04, 0xffffffff, 0x00000100,
939 0xae04, 0xffffffff, 0x00000100,
940 0x949c, 0xffffffff, 0x00000100,
941 0x802c, 0xffffffff, 0xe0000000,
942 0x9160, 0xffffffff, 0x00010000,
943 0x9164, 0xffffffff, 0x00030002,
944 0x9168, 0xffffffff, 0x00040007,
945 0x916c, 0xffffffff, 0x00060005,
946 0x9170, 0xffffffff, 0x00090008,
947 0x9174, 0xffffffff, 0x00020001,
948 0x9178, 0xffffffff, 0x00040003,
949 0x917c, 0xffffffff, 0x00000007,
950 0x9180, 0xffffffff, 0x00060005,
951 0x9184, 0xffffffff, 0x00090008,
952 0x9188, 0xffffffff, 0x00030002,
953 0x918c, 0xffffffff, 0x00050004,
954 0x9190, 0xffffffff, 0x00000008,
955 0x9194, 0xffffffff, 0x00070006,
956 0x9198, 0xffffffff, 0x000a0009,
957 0x919c, 0xffffffff, 0x00040003,
958 0x91a0, 0xffffffff, 0x00060005,
959 0x91a4, 0xffffffff, 0x00000009,
960 0x91a8, 0xffffffff, 0x00080007,
961 0x91ac, 0xffffffff, 0x000b000a,
962 0x91b0, 0xffffffff, 0x00050004,
963 0x91b4, 0xffffffff, 0x00070006,
964 0x91b8, 0xffffffff, 0x0008000b,
965 0x91bc, 0xffffffff, 0x000a0009,
966 0x91c0, 0xffffffff, 0x000d000c,
967 0x91c4, 0xffffffff, 0x00060005,
968 0x91c8, 0xffffffff, 0x00080007,
969 0x91cc, 0xffffffff, 0x0000000b,
970 0x91d0, 0xffffffff, 0x000a0009,
971 0x91d4, 0xffffffff, 0x000d000c,
972 0x9150, 0xffffffff, 0x96940200,
973 0x8708, 0xffffffff, 0x00900100,
974 0xc478, 0xffffffff, 0x00000080,
975 0xc404, 0xffffffff, 0x0020003f,
976 0x30, 0xffffffff, 0x0000001c,
977 0x34, 0x000f0000, 0x000f0000,
978 0x160c, 0xffffffff, 0x00000100,
979 0x1024, 0xffffffff, 0x00000100,
980 0x102c, 0x00000101, 0x00000000,
981 0x20a8, 0xffffffff, 0x00000104,
982 0x264c, 0x000c0000, 0x000c0000,
983 0x2648, 0x000c0000, 0x000c0000,
984 0x55e4, 0xff000fff, 0x00000100,
985 0x55e8, 0x00000001, 0x00000001,
986 0x2f50, 0x00000001, 0x00000001,
987 0x30cc, 0xc0000fff, 0x00000104,
988 0xc1e4, 0x00000001, 0x00000001,
989 0xd0c0, 0xfffffff0, 0x00000100,
990 0xd8c0, 0xfffffff0, 0x00000100
991};
992
fffbdda4
AD
993static const u32 hainan_mgcg_cgcg_init[] =
994{
995 0xc400, 0xffffffff, 0xfffffffc,
996 0x802c, 0xffffffff, 0xe0000000,
997 0x9a60, 0xffffffff, 0x00000100,
998 0x92a4, 0xffffffff, 0x00000100,
999 0xc164, 0xffffffff, 0x00000100,
1000 0x9774, 0xffffffff, 0x00000100,
1001 0x8984, 0xffffffff, 0x06000100,
1002 0x8a18, 0xffffffff, 0x00000100,
1003 0x92a0, 0xffffffff, 0x00000100,
1004 0xc380, 0xffffffff, 0x00000100,
1005 0x8b28, 0xffffffff, 0x00000100,
1006 0x9144, 0xffffffff, 0x00000100,
1007 0x8d88, 0xffffffff, 0x00000100,
1008 0x8d8c, 0xffffffff, 0x00000100,
1009 0x9030, 0xffffffff, 0x00000100,
1010 0x9034, 0xffffffff, 0x00000100,
1011 0x9038, 0xffffffff, 0x00000100,
1012 0x903c, 0xffffffff, 0x00000100,
1013 0xad80, 0xffffffff, 0x00000100,
1014 0xac54, 0xffffffff, 0x00000100,
1015 0x897c, 0xffffffff, 0x06000100,
1016 0x9868, 0xffffffff, 0x00000100,
1017 0x9510, 0xffffffff, 0x00000100,
1018 0xaf04, 0xffffffff, 0x00000100,
1019 0xae04, 0xffffffff, 0x00000100,
1020 0x949c, 0xffffffff, 0x00000100,
1021 0x802c, 0xffffffff, 0xe0000000,
1022 0x9160, 0xffffffff, 0x00010000,
1023 0x9164, 0xffffffff, 0x00030002,
1024 0x9168, 0xffffffff, 0x00040007,
1025 0x916c, 0xffffffff, 0x00060005,
1026 0x9170, 0xffffffff, 0x00090008,
1027 0x9174, 0xffffffff, 0x00020001,
1028 0x9178, 0xffffffff, 0x00040003,
1029 0x917c, 0xffffffff, 0x00000007,
1030 0x9180, 0xffffffff, 0x00060005,
1031 0x9184, 0xffffffff, 0x00090008,
1032 0x9188, 0xffffffff, 0x00030002,
1033 0x918c, 0xffffffff, 0x00050004,
1034 0x9190, 0xffffffff, 0x00000008,
1035 0x9194, 0xffffffff, 0x00070006,
1036 0x9198, 0xffffffff, 0x000a0009,
1037 0x919c, 0xffffffff, 0x00040003,
1038 0x91a0, 0xffffffff, 0x00060005,
1039 0x91a4, 0xffffffff, 0x00000009,
1040 0x91a8, 0xffffffff, 0x00080007,
1041 0x91ac, 0xffffffff, 0x000b000a,
1042 0x91b0, 0xffffffff, 0x00050004,
1043 0x91b4, 0xffffffff, 0x00070006,
1044 0x91b8, 0xffffffff, 0x0008000b,
1045 0x91bc, 0xffffffff, 0x000a0009,
1046 0x91c0, 0xffffffff, 0x000d000c,
1047 0x91c4, 0xffffffff, 0x00060005,
1048 0x91c8, 0xffffffff, 0x00080007,
1049 0x91cc, 0xffffffff, 0x0000000b,
1050 0x91d0, 0xffffffff, 0x000a0009,
1051 0x91d4, 0xffffffff, 0x000d000c,
1052 0x9150, 0xffffffff, 0x96940200,
1053 0x8708, 0xffffffff, 0x00900100,
1054 0xc478, 0xffffffff, 0x00000080,
1055 0xc404, 0xffffffff, 0x0020003f,
1056 0x30, 0xffffffff, 0x0000001c,
1057 0x34, 0x000f0000, 0x000f0000,
1058 0x160c, 0xffffffff, 0x00000100,
1059 0x1024, 0xffffffff, 0x00000100,
1060 0x20a8, 0xffffffff, 0x00000104,
1061 0x264c, 0x000c0000, 0x000c0000,
1062 0x2648, 0x000c0000, 0x000c0000,
1063 0x2f50, 0x00000001, 0x00000001,
1064 0x30cc, 0xc0000fff, 0x00000104,
1065 0xc1e4, 0x00000001, 0x00000001,
1066 0xd0c0, 0xfffffff0, 0x00000100,
1067 0xd8c0, 0xfffffff0, 0x00000100
1068};
1069
205996c0
AD
1070static u32 verde_pg_init[] =
1071{
1072 0x353c, 0xffffffff, 0x40000,
1073 0x3538, 0xffffffff, 0x200010ff,
1074 0x353c, 0xffffffff, 0x0,
1075 0x353c, 0xffffffff, 0x0,
1076 0x353c, 0xffffffff, 0x0,
1077 0x353c, 0xffffffff, 0x0,
1078 0x353c, 0xffffffff, 0x0,
1079 0x353c, 0xffffffff, 0x7007,
1080 0x3538, 0xffffffff, 0x300010ff,
1081 0x353c, 0xffffffff, 0x0,
1082 0x353c, 0xffffffff, 0x0,
1083 0x353c, 0xffffffff, 0x0,
1084 0x353c, 0xffffffff, 0x0,
1085 0x353c, 0xffffffff, 0x0,
1086 0x353c, 0xffffffff, 0x400000,
1087 0x3538, 0xffffffff, 0x100010ff,
1088 0x353c, 0xffffffff, 0x0,
1089 0x353c, 0xffffffff, 0x0,
1090 0x353c, 0xffffffff, 0x0,
1091 0x353c, 0xffffffff, 0x0,
1092 0x353c, 0xffffffff, 0x0,
1093 0x353c, 0xffffffff, 0x120200,
1094 0x3538, 0xffffffff, 0x500010ff,
1095 0x353c, 0xffffffff, 0x0,
1096 0x353c, 0xffffffff, 0x0,
1097 0x353c, 0xffffffff, 0x0,
1098 0x353c, 0xffffffff, 0x0,
1099 0x353c, 0xffffffff, 0x0,
1100 0x353c, 0xffffffff, 0x1e1e16,
1101 0x3538, 0xffffffff, 0x600010ff,
1102 0x353c, 0xffffffff, 0x0,
1103 0x353c, 0xffffffff, 0x0,
1104 0x353c, 0xffffffff, 0x0,
1105 0x353c, 0xffffffff, 0x0,
1106 0x353c, 0xffffffff, 0x0,
1107 0x353c, 0xffffffff, 0x171f1e,
1108 0x3538, 0xffffffff, 0x700010ff,
1109 0x353c, 0xffffffff, 0x0,
1110 0x353c, 0xffffffff, 0x0,
1111 0x353c, 0xffffffff, 0x0,
1112 0x353c, 0xffffffff, 0x0,
1113 0x353c, 0xffffffff, 0x0,
1114 0x353c, 0xffffffff, 0x0,
1115 0x3538, 0xffffffff, 0x9ff,
1116 0x3500, 0xffffffff, 0x0,
1117 0x3504, 0xffffffff, 0x10000800,
1118 0x3504, 0xffffffff, 0xf,
1119 0x3504, 0xffffffff, 0xf,
1120 0x3500, 0xffffffff, 0x4,
1121 0x3504, 0xffffffff, 0x1000051e,
1122 0x3504, 0xffffffff, 0xffff,
1123 0x3504, 0xffffffff, 0xffff,
1124 0x3500, 0xffffffff, 0x8,
1125 0x3504, 0xffffffff, 0x80500,
1126 0x3500, 0xffffffff, 0x12,
1127 0x3504, 0xffffffff, 0x9050c,
1128 0x3500, 0xffffffff, 0x1d,
1129 0x3504, 0xffffffff, 0xb052c,
1130 0x3500, 0xffffffff, 0x2a,
1131 0x3504, 0xffffffff, 0x1053e,
1132 0x3500, 0xffffffff, 0x2d,
1133 0x3504, 0xffffffff, 0x10546,
1134 0x3500, 0xffffffff, 0x30,
1135 0x3504, 0xffffffff, 0xa054e,
1136 0x3500, 0xffffffff, 0x3c,
1137 0x3504, 0xffffffff, 0x1055f,
1138 0x3500, 0xffffffff, 0x3f,
1139 0x3504, 0xffffffff, 0x10567,
1140 0x3500, 0xffffffff, 0x42,
1141 0x3504, 0xffffffff, 0x1056f,
1142 0x3500, 0xffffffff, 0x45,
1143 0x3504, 0xffffffff, 0x10572,
1144 0x3500, 0xffffffff, 0x48,
1145 0x3504, 0xffffffff, 0x20575,
1146 0x3500, 0xffffffff, 0x4c,
1147 0x3504, 0xffffffff, 0x190801,
1148 0x3500, 0xffffffff, 0x67,
1149 0x3504, 0xffffffff, 0x1082a,
1150 0x3500, 0xffffffff, 0x6a,
1151 0x3504, 0xffffffff, 0x1b082d,
1152 0x3500, 0xffffffff, 0x87,
1153 0x3504, 0xffffffff, 0x310851,
1154 0x3500, 0xffffffff, 0xba,
1155 0x3504, 0xffffffff, 0x891,
1156 0x3500, 0xffffffff, 0xbc,
1157 0x3504, 0xffffffff, 0x893,
1158 0x3500, 0xffffffff, 0xbe,
1159 0x3504, 0xffffffff, 0x20895,
1160 0x3500, 0xffffffff, 0xc2,
1161 0x3504, 0xffffffff, 0x20899,
1162 0x3500, 0xffffffff, 0xc6,
1163 0x3504, 0xffffffff, 0x2089d,
1164 0x3500, 0xffffffff, 0xca,
1165 0x3504, 0xffffffff, 0x8a1,
1166 0x3500, 0xffffffff, 0xcc,
1167 0x3504, 0xffffffff, 0x8a3,
1168 0x3500, 0xffffffff, 0xce,
1169 0x3504, 0xffffffff, 0x308a5,
1170 0x3500, 0xffffffff, 0xd3,
1171 0x3504, 0xffffffff, 0x6d08cd,
1172 0x3500, 0xffffffff, 0x142,
1173 0x3504, 0xffffffff, 0x2000095a,
1174 0x3504, 0xffffffff, 0x1,
1175 0x3500, 0xffffffff, 0x144,
1176 0x3504, 0xffffffff, 0x301f095b,
1177 0x3500, 0xffffffff, 0x165,
1178 0x3504, 0xffffffff, 0xc094d,
1179 0x3500, 0xffffffff, 0x173,
1180 0x3504, 0xffffffff, 0xf096d,
1181 0x3500, 0xffffffff, 0x184,
1182 0x3504, 0xffffffff, 0x15097f,
1183 0x3500, 0xffffffff, 0x19b,
1184 0x3504, 0xffffffff, 0xc0998,
1185 0x3500, 0xffffffff, 0x1a9,
1186 0x3504, 0xffffffff, 0x409a7,
1187 0x3500, 0xffffffff, 0x1af,
1188 0x3504, 0xffffffff, 0xcdc,
1189 0x3500, 0xffffffff, 0x1b1,
1190 0x3504, 0xffffffff, 0x800,
1191 0x3508, 0xffffffff, 0x6c9b2000,
1192 0x3510, 0xfc00, 0x2000,
1193 0x3544, 0xffffffff, 0xfc0,
1194 0x28d4, 0x00000100, 0x100
1195};
1196
1197static void si_init_golden_registers(struct radeon_device *rdev)
1198{
1199 switch (rdev->family) {
1200 case CHIP_TAHITI:
1201 radeon_program_register_sequence(rdev,
1202 tahiti_golden_registers,
1203 (const u32)ARRAY_SIZE(tahiti_golden_registers));
1204 radeon_program_register_sequence(rdev,
1205 tahiti_golden_rlc_registers,
1206 (const u32)ARRAY_SIZE(tahiti_golden_rlc_registers));
1207 radeon_program_register_sequence(rdev,
1208 tahiti_mgcg_cgcg_init,
1209 (const u32)ARRAY_SIZE(tahiti_mgcg_cgcg_init));
1210 radeon_program_register_sequence(rdev,
1211 tahiti_golden_registers2,
1212 (const u32)ARRAY_SIZE(tahiti_golden_registers2));
1213 break;
1214 case CHIP_PITCAIRN:
1215 radeon_program_register_sequence(rdev,
1216 pitcairn_golden_registers,
1217 (const u32)ARRAY_SIZE(pitcairn_golden_registers));
1218 radeon_program_register_sequence(rdev,
1219 pitcairn_golden_rlc_registers,
1220 (const u32)ARRAY_SIZE(pitcairn_golden_rlc_registers));
1221 radeon_program_register_sequence(rdev,
1222 pitcairn_mgcg_cgcg_init,
1223 (const u32)ARRAY_SIZE(pitcairn_mgcg_cgcg_init));
1224 break;
1225 case CHIP_VERDE:
1226 radeon_program_register_sequence(rdev,
1227 verde_golden_registers,
1228 (const u32)ARRAY_SIZE(verde_golden_registers));
1229 radeon_program_register_sequence(rdev,
1230 verde_golden_rlc_registers,
1231 (const u32)ARRAY_SIZE(verde_golden_rlc_registers));
1232 radeon_program_register_sequence(rdev,
1233 verde_mgcg_cgcg_init,
1234 (const u32)ARRAY_SIZE(verde_mgcg_cgcg_init));
1235 radeon_program_register_sequence(rdev,
1236 verde_pg_init,
1237 (const u32)ARRAY_SIZE(verde_pg_init));
1238 break;
1239 case CHIP_OLAND:
1240 radeon_program_register_sequence(rdev,
1241 oland_golden_registers,
1242 (const u32)ARRAY_SIZE(oland_golden_registers));
1243 radeon_program_register_sequence(rdev,
1244 oland_golden_rlc_registers,
1245 (const u32)ARRAY_SIZE(oland_golden_rlc_registers));
1246 radeon_program_register_sequence(rdev,
1247 oland_mgcg_cgcg_init,
1248 (const u32)ARRAY_SIZE(oland_mgcg_cgcg_init));
1249 break;
fffbdda4
AD
1250 case CHIP_HAINAN:
1251 radeon_program_register_sequence(rdev,
1252 hainan_golden_registers,
1253 (const u32)ARRAY_SIZE(hainan_golden_registers));
1254 radeon_program_register_sequence(rdev,
1255 hainan_golden_registers2,
1256 (const u32)ARRAY_SIZE(hainan_golden_registers2));
1257 radeon_program_register_sequence(rdev,
1258 hainan_mgcg_cgcg_init,
1259 (const u32)ARRAY_SIZE(hainan_mgcg_cgcg_init));
1260 break;
205996c0
AD
1261 default:
1262 break;
1263 }
1264}
1265
454d2e2a
AD
1266#define PCIE_BUS_CLK 10000
1267#define TCLK (PCIE_BUS_CLK / 10)
1268
1269/**
1270 * si_get_xclk - get the xclk
1271 *
1272 * @rdev: radeon_device pointer
1273 *
1274 * Returns the reference clock used by the gfx engine
1275 * (SI).
1276 */
1277u32 si_get_xclk(struct radeon_device *rdev)
1278{
1279 u32 reference_clock = rdev->clock.spll.reference_freq;
1280 u32 tmp;
1281
1282 tmp = RREG32(CG_CLKPIN_CNTL_2);
1283 if (tmp & MUX_TCLK_TO_XCLK)
1284 return TCLK;
1285
1286 tmp = RREG32(CG_CLKPIN_CNTL);
1287 if (tmp & XTALIN_DIVIDE)
1288 return reference_clock / 4;
1289
1290 return reference_clock;
1291}
1292
1bd47d2e
AD
1293/* get temperature in millidegrees */
1294int si_get_temp(struct radeon_device *rdev)
1295{
1296 u32 temp;
1297 int actual_temp = 0;
1298
1299 temp = (RREG32(CG_MULT_THERMAL_STATUS) & CTF_TEMP_MASK) >>
1300 CTF_TEMP_SHIFT;
1301
1302 if (temp & 0x200)
1303 actual_temp = 255;
1304 else
1305 actual_temp = temp & 0x1ff;
1306
1307 actual_temp = (actual_temp * 1000);
1308
1309 return actual_temp;
1310}
1311
8b074dd6
AD
1312#define TAHITI_IO_MC_REGS_SIZE 36
1313
1314static const u32 tahiti_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
1315 {0x0000006f, 0x03044000},
1316 {0x00000070, 0x0480c018},
1317 {0x00000071, 0x00000040},
1318 {0x00000072, 0x01000000},
1319 {0x00000074, 0x000000ff},
1320 {0x00000075, 0x00143400},
1321 {0x00000076, 0x08ec0800},
1322 {0x00000077, 0x040000cc},
1323 {0x00000079, 0x00000000},
1324 {0x0000007a, 0x21000409},
1325 {0x0000007c, 0x00000000},
1326 {0x0000007d, 0xe8000000},
1327 {0x0000007e, 0x044408a8},
1328 {0x0000007f, 0x00000003},
1329 {0x00000080, 0x00000000},
1330 {0x00000081, 0x01000000},
1331 {0x00000082, 0x02000000},
1332 {0x00000083, 0x00000000},
1333 {0x00000084, 0xe3f3e4f4},
1334 {0x00000085, 0x00052024},
1335 {0x00000087, 0x00000000},
1336 {0x00000088, 0x66036603},
1337 {0x00000089, 0x01000000},
1338 {0x0000008b, 0x1c0a0000},
1339 {0x0000008c, 0xff010000},
1340 {0x0000008e, 0xffffefff},
1341 {0x0000008f, 0xfff3efff},
1342 {0x00000090, 0xfff3efbf},
1343 {0x00000094, 0x00101101},
1344 {0x00000095, 0x00000fff},
1345 {0x00000096, 0x00116fff},
1346 {0x00000097, 0x60010000},
1347 {0x00000098, 0x10010000},
1348 {0x00000099, 0x00006000},
1349 {0x0000009a, 0x00001000},
1350 {0x0000009f, 0x00a77400}
1351};
1352
1353static const u32 pitcairn_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
1354 {0x0000006f, 0x03044000},
1355 {0x00000070, 0x0480c018},
1356 {0x00000071, 0x00000040},
1357 {0x00000072, 0x01000000},
1358 {0x00000074, 0x000000ff},
1359 {0x00000075, 0x00143400},
1360 {0x00000076, 0x08ec0800},
1361 {0x00000077, 0x040000cc},
1362 {0x00000079, 0x00000000},
1363 {0x0000007a, 0x21000409},
1364 {0x0000007c, 0x00000000},
1365 {0x0000007d, 0xe8000000},
1366 {0x0000007e, 0x044408a8},
1367 {0x0000007f, 0x00000003},
1368 {0x00000080, 0x00000000},
1369 {0x00000081, 0x01000000},
1370 {0x00000082, 0x02000000},
1371 {0x00000083, 0x00000000},
1372 {0x00000084, 0xe3f3e4f4},
1373 {0x00000085, 0x00052024},
1374 {0x00000087, 0x00000000},
1375 {0x00000088, 0x66036603},
1376 {0x00000089, 0x01000000},
1377 {0x0000008b, 0x1c0a0000},
1378 {0x0000008c, 0xff010000},
1379 {0x0000008e, 0xffffefff},
1380 {0x0000008f, 0xfff3efff},
1381 {0x00000090, 0xfff3efbf},
1382 {0x00000094, 0x00101101},
1383 {0x00000095, 0x00000fff},
1384 {0x00000096, 0x00116fff},
1385 {0x00000097, 0x60010000},
1386 {0x00000098, 0x10010000},
1387 {0x00000099, 0x00006000},
1388 {0x0000009a, 0x00001000},
1389 {0x0000009f, 0x00a47400}
1390};
1391
1392static const u32 verde_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
1393 {0x0000006f, 0x03044000},
1394 {0x00000070, 0x0480c018},
1395 {0x00000071, 0x00000040},
1396 {0x00000072, 0x01000000},
1397 {0x00000074, 0x000000ff},
1398 {0x00000075, 0x00143400},
1399 {0x00000076, 0x08ec0800},
1400 {0x00000077, 0x040000cc},
1401 {0x00000079, 0x00000000},
1402 {0x0000007a, 0x21000409},
1403 {0x0000007c, 0x00000000},
1404 {0x0000007d, 0xe8000000},
1405 {0x0000007e, 0x044408a8},
1406 {0x0000007f, 0x00000003},
1407 {0x00000080, 0x00000000},
1408 {0x00000081, 0x01000000},
1409 {0x00000082, 0x02000000},
1410 {0x00000083, 0x00000000},
1411 {0x00000084, 0xe3f3e4f4},
1412 {0x00000085, 0x00052024},
1413 {0x00000087, 0x00000000},
1414 {0x00000088, 0x66036603},
1415 {0x00000089, 0x01000000},
1416 {0x0000008b, 0x1c0a0000},
1417 {0x0000008c, 0xff010000},
1418 {0x0000008e, 0xffffefff},
1419 {0x0000008f, 0xfff3efff},
1420 {0x00000090, 0xfff3efbf},
1421 {0x00000094, 0x00101101},
1422 {0x00000095, 0x00000fff},
1423 {0x00000096, 0x00116fff},
1424 {0x00000097, 0x60010000},
1425 {0x00000098, 0x10010000},
1426 {0x00000099, 0x00006000},
1427 {0x0000009a, 0x00001000},
1428 {0x0000009f, 0x00a37400}
1429};
1430
bcc7f5d2
AD
1431static const u32 oland_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
1432 {0x0000006f, 0x03044000},
1433 {0x00000070, 0x0480c018},
1434 {0x00000071, 0x00000040},
1435 {0x00000072, 0x01000000},
1436 {0x00000074, 0x000000ff},
1437 {0x00000075, 0x00143400},
1438 {0x00000076, 0x08ec0800},
1439 {0x00000077, 0x040000cc},
1440 {0x00000079, 0x00000000},
1441 {0x0000007a, 0x21000409},
1442 {0x0000007c, 0x00000000},
1443 {0x0000007d, 0xe8000000},
1444 {0x0000007e, 0x044408a8},
1445 {0x0000007f, 0x00000003},
1446 {0x00000080, 0x00000000},
1447 {0x00000081, 0x01000000},
1448 {0x00000082, 0x02000000},
1449 {0x00000083, 0x00000000},
1450 {0x00000084, 0xe3f3e4f4},
1451 {0x00000085, 0x00052024},
1452 {0x00000087, 0x00000000},
1453 {0x00000088, 0x66036603},
1454 {0x00000089, 0x01000000},
1455 {0x0000008b, 0x1c0a0000},
1456 {0x0000008c, 0xff010000},
1457 {0x0000008e, 0xffffefff},
1458 {0x0000008f, 0xfff3efff},
1459 {0x00000090, 0xfff3efbf},
1460 {0x00000094, 0x00101101},
1461 {0x00000095, 0x00000fff},
1462 {0x00000096, 0x00116fff},
1463 {0x00000097, 0x60010000},
1464 {0x00000098, 0x10010000},
1465 {0x00000099, 0x00006000},
1466 {0x0000009a, 0x00001000},
1467 {0x0000009f, 0x00a17730}
1468};
1469
c04c00b4
AD
1470static const u32 hainan_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
1471 {0x0000006f, 0x03044000},
1472 {0x00000070, 0x0480c018},
1473 {0x00000071, 0x00000040},
1474 {0x00000072, 0x01000000},
1475 {0x00000074, 0x000000ff},
1476 {0x00000075, 0x00143400},
1477 {0x00000076, 0x08ec0800},
1478 {0x00000077, 0x040000cc},
1479 {0x00000079, 0x00000000},
1480 {0x0000007a, 0x21000409},
1481 {0x0000007c, 0x00000000},
1482 {0x0000007d, 0xe8000000},
1483 {0x0000007e, 0x044408a8},
1484 {0x0000007f, 0x00000003},
1485 {0x00000080, 0x00000000},
1486 {0x00000081, 0x01000000},
1487 {0x00000082, 0x02000000},
1488 {0x00000083, 0x00000000},
1489 {0x00000084, 0xe3f3e4f4},
1490 {0x00000085, 0x00052024},
1491 {0x00000087, 0x00000000},
1492 {0x00000088, 0x66036603},
1493 {0x00000089, 0x01000000},
1494 {0x0000008b, 0x1c0a0000},
1495 {0x0000008c, 0xff010000},
1496 {0x0000008e, 0xffffefff},
1497 {0x0000008f, 0xfff3efff},
1498 {0x00000090, 0xfff3efbf},
1499 {0x00000094, 0x00101101},
1500 {0x00000095, 0x00000fff},
1501 {0x00000096, 0x00116fff},
1502 {0x00000097, 0x60010000},
1503 {0x00000098, 0x10010000},
1504 {0x00000099, 0x00006000},
1505 {0x0000009a, 0x00001000},
1506 {0x0000009f, 0x00a07730}
1507};
1508
8b074dd6 1509/* ucode loading */
6c7bccea 1510int si_mc_load_microcode(struct radeon_device *rdev)
8b074dd6 1511{
629bd33c
AD
1512 const __be32 *fw_data = NULL;
1513 const __le32 *new_fw_data = NULL;
8b074dd6 1514 u32 running, blackout = 0;
629bd33c
AD
1515 u32 *io_mc_regs = NULL;
1516 const __le32 *new_io_mc_regs = NULL;
8c79bae6 1517 int i, regs_size, ucode_size;
8b074dd6
AD
1518
1519 if (!rdev->mc_fw)
1520 return -EINVAL;
1521
629bd33c
AD
1522 if (rdev->new_fw) {
1523 const struct mc_firmware_header_v1_0 *hdr =
1524 (const struct mc_firmware_header_v1_0 *)rdev->mc_fw->data;
1525
1526 radeon_ucode_print_mc_hdr(&hdr->header);
1527 regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
1528 new_io_mc_regs = (const __le32 *)
1529 (rdev->mc_fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
1530 ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
1531 new_fw_data = (const __le32 *)
1532 (rdev->mc_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1533 } else {
1534 ucode_size = rdev->mc_fw->size / 4;
8c79bae6 1535
629bd33c
AD
1536 switch (rdev->family) {
1537 case CHIP_TAHITI:
1538 io_mc_regs = (u32 *)&tahiti_io_mc_regs;
1539 regs_size = TAHITI_IO_MC_REGS_SIZE;
1540 break;
1541 case CHIP_PITCAIRN:
1542 io_mc_regs = (u32 *)&pitcairn_io_mc_regs;
1543 regs_size = TAHITI_IO_MC_REGS_SIZE;
1544 break;
1545 case CHIP_VERDE:
1546 default:
1547 io_mc_regs = (u32 *)&verde_io_mc_regs;
1548 regs_size = TAHITI_IO_MC_REGS_SIZE;
1549 break;
1550 case CHIP_OLAND:
1551 io_mc_regs = (u32 *)&oland_io_mc_regs;
1552 regs_size = TAHITI_IO_MC_REGS_SIZE;
1553 break;
1554 case CHIP_HAINAN:
1555 io_mc_regs = (u32 *)&hainan_io_mc_regs;
1556 regs_size = TAHITI_IO_MC_REGS_SIZE;
1557 break;
1558 }
1559 fw_data = (const __be32 *)rdev->mc_fw->data;
8b074dd6
AD
1560 }
1561
1562 running = RREG32(MC_SEQ_SUP_CNTL) & RUN_MASK;
1563
1564 if (running == 0) {
1565 if (running) {
1566 blackout = RREG32(MC_SHARED_BLACKOUT_CNTL);
1567 WREG32(MC_SHARED_BLACKOUT_CNTL, blackout | 1);
1568 }
1569
1570 /* reset the engine and set to writable */
1571 WREG32(MC_SEQ_SUP_CNTL, 0x00000008);
1572 WREG32(MC_SEQ_SUP_CNTL, 0x00000010);
1573
1574 /* load mc io regs */
1575 for (i = 0; i < regs_size; i++) {
629bd33c
AD
1576 if (rdev->new_fw) {
1577 WREG32(MC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(new_io_mc_regs++));
1578 WREG32(MC_SEQ_IO_DEBUG_DATA, le32_to_cpup(new_io_mc_regs++));
1579 } else {
1580 WREG32(MC_SEQ_IO_DEBUG_INDEX, io_mc_regs[(i << 1)]);
1581 WREG32(MC_SEQ_IO_DEBUG_DATA, io_mc_regs[(i << 1) + 1]);
1582 }
8b074dd6
AD
1583 }
1584 /* load the MC ucode */
629bd33c
AD
1585 for (i = 0; i < ucode_size; i++) {
1586 if (rdev->new_fw)
1587 WREG32(MC_SEQ_SUP_PGM, le32_to_cpup(new_fw_data++));
1588 else
1589 WREG32(MC_SEQ_SUP_PGM, be32_to_cpup(fw_data++));
1590 }
8b074dd6
AD
1591
1592 /* put the engine back into the active state */
1593 WREG32(MC_SEQ_SUP_CNTL, 0x00000008);
1594 WREG32(MC_SEQ_SUP_CNTL, 0x00000004);
1595 WREG32(MC_SEQ_SUP_CNTL, 0x00000001);
1596
1597 /* wait for training to complete */
1598 for (i = 0; i < rdev->usec_timeout; i++) {
1599 if (RREG32(MC_SEQ_TRAIN_WAKEUP_CNTL) & TRAIN_DONE_D0)
1600 break;
1601 udelay(1);
1602 }
1603 for (i = 0; i < rdev->usec_timeout; i++) {
1604 if (RREG32(MC_SEQ_TRAIN_WAKEUP_CNTL) & TRAIN_DONE_D1)
1605 break;
1606 udelay(1);
1607 }
1608
1609 if (running)
1610 WREG32(MC_SHARED_BLACKOUT_CNTL, blackout);
1611 }
1612
1613 return 0;
1614}
1615
0f0de06c
AD
1616static int si_init_microcode(struct radeon_device *rdev)
1617{
0f0de06c 1618 const char *chip_name;
629bd33c 1619 const char *new_chip_name;
0f0de06c 1620 size_t pfp_req_size, me_req_size, ce_req_size, rlc_req_size, mc_req_size;
1ebe9280 1621 size_t smc_req_size, mc2_req_size;
0f0de06c
AD
1622 char fw_name[30];
1623 int err;
629bd33c 1624 int new_fw = 0;
0f0de06c
AD
1625
1626 DRM_DEBUG("\n");
1627
0f0de06c
AD
1628 switch (rdev->family) {
1629 case CHIP_TAHITI:
1630 chip_name = "TAHITI";
629bd33c 1631 new_chip_name = "tahiti";
0f0de06c
AD
1632 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
1633 me_req_size = SI_PM4_UCODE_SIZE * 4;
1634 ce_req_size = SI_CE_UCODE_SIZE * 4;
1635 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
1636 mc_req_size = SI_MC_UCODE_SIZE * 4;
1ebe9280 1637 mc2_req_size = TAHITI_MC_UCODE_SIZE * 4;
a9e61410 1638 smc_req_size = ALIGN(TAHITI_SMC_UCODE_SIZE, 4);
0f0de06c
AD
1639 break;
1640 case CHIP_PITCAIRN:
1641 chip_name = "PITCAIRN";
629bd33c 1642 new_chip_name = "pitcairn";
0f0de06c
AD
1643 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
1644 me_req_size = SI_PM4_UCODE_SIZE * 4;
1645 ce_req_size = SI_CE_UCODE_SIZE * 4;
1646 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
1647 mc_req_size = SI_MC_UCODE_SIZE * 4;
1ebe9280 1648 mc2_req_size = PITCAIRN_MC_UCODE_SIZE * 4;
a9e61410 1649 smc_req_size = ALIGN(PITCAIRN_SMC_UCODE_SIZE, 4);
0f0de06c
AD
1650 break;
1651 case CHIP_VERDE:
1652 chip_name = "VERDE";
629bd33c 1653 new_chip_name = "verde";
0f0de06c
AD
1654 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
1655 me_req_size = SI_PM4_UCODE_SIZE * 4;
1656 ce_req_size = SI_CE_UCODE_SIZE * 4;
1657 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
1658 mc_req_size = SI_MC_UCODE_SIZE * 4;
1ebe9280 1659 mc2_req_size = VERDE_MC_UCODE_SIZE * 4;
a9e61410 1660 smc_req_size = ALIGN(VERDE_SMC_UCODE_SIZE, 4);
0f0de06c 1661 break;
bcc7f5d2
AD
1662 case CHIP_OLAND:
1663 chip_name = "OLAND";
629bd33c 1664 new_chip_name = "oland";
bcc7f5d2
AD
1665 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
1666 me_req_size = SI_PM4_UCODE_SIZE * 4;
1667 ce_req_size = SI_CE_UCODE_SIZE * 4;
1668 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
1ebe9280 1669 mc_req_size = mc2_req_size = OLAND_MC_UCODE_SIZE * 4;
a9e61410 1670 smc_req_size = ALIGN(OLAND_SMC_UCODE_SIZE, 4);
bcc7f5d2 1671 break;
c04c00b4
AD
1672 case CHIP_HAINAN:
1673 chip_name = "HAINAN";
629bd33c 1674 new_chip_name = "hainan";
c04c00b4
AD
1675 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
1676 me_req_size = SI_PM4_UCODE_SIZE * 4;
1677 ce_req_size = SI_CE_UCODE_SIZE * 4;
1678 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
1ebe9280 1679 mc_req_size = mc2_req_size = OLAND_MC_UCODE_SIZE * 4;
a9e61410 1680 smc_req_size = ALIGN(HAINAN_SMC_UCODE_SIZE, 4);
c04c00b4 1681 break;
0f0de06c
AD
1682 default: BUG();
1683 }
1684
629bd33c 1685 DRM_INFO("Loading %s Microcode\n", new_chip_name);
0f0de06c 1686
629bd33c 1687 snprintf(fw_name, sizeof(fw_name), "radeon/%s_pfp.bin", new_chip_name);
0a168933 1688 err = request_firmware(&rdev->pfp_fw, fw_name, rdev->dev);
629bd33c
AD
1689 if (err) {
1690 snprintf(fw_name, sizeof(fw_name), "radeon/%s_pfp.bin", chip_name);
1691 err = request_firmware(&rdev->pfp_fw, fw_name, rdev->dev);
1692 if (err)
1693 goto out;
1694 if (rdev->pfp_fw->size != pfp_req_size) {
1695 printk(KERN_ERR
1696 "si_cp: Bogus length %zu in firmware \"%s\"\n",
1697 rdev->pfp_fw->size, fw_name);
1698 err = -EINVAL;
1699 goto out;
1700 }
1701 } else {
1702 err = radeon_ucode_validate(rdev->pfp_fw);
1703 if (err) {
1704 printk(KERN_ERR
1705 "si_cp: validation failed for firmware \"%s\"\n",
1706 fw_name);
1707 goto out;
1708 } else {
1709 new_fw++;
1710 }
0f0de06c
AD
1711 }
1712
629bd33c 1713 snprintf(fw_name, sizeof(fw_name), "radeon/%s_me.bin", new_chip_name);
0a168933 1714 err = request_firmware(&rdev->me_fw, fw_name, rdev->dev);
629bd33c
AD
1715 if (err) {
1716 snprintf(fw_name, sizeof(fw_name), "radeon/%s_me.bin", chip_name);
1717 err = request_firmware(&rdev->me_fw, fw_name, rdev->dev);
1718 if (err)
1719 goto out;
1720 if (rdev->me_fw->size != me_req_size) {
1721 printk(KERN_ERR
1722 "si_cp: Bogus length %zu in firmware \"%s\"\n",
1723 rdev->me_fw->size, fw_name);
1724 err = -EINVAL;
1725 }
1726 } else {
1727 err = radeon_ucode_validate(rdev->me_fw);
1728 if (err) {
1729 printk(KERN_ERR
1730 "si_cp: validation failed for firmware \"%s\"\n",
1731 fw_name);
1732 goto out;
1733 } else {
1734 new_fw++;
1735 }
0f0de06c
AD
1736 }
1737
629bd33c 1738 snprintf(fw_name, sizeof(fw_name), "radeon/%s_ce.bin", new_chip_name);
0a168933 1739 err = request_firmware(&rdev->ce_fw, fw_name, rdev->dev);
629bd33c
AD
1740 if (err) {
1741 snprintf(fw_name, sizeof(fw_name), "radeon/%s_ce.bin", chip_name);
1742 err = request_firmware(&rdev->ce_fw, fw_name, rdev->dev);
1743 if (err)
1744 goto out;
1745 if (rdev->ce_fw->size != ce_req_size) {
1746 printk(KERN_ERR
1747 "si_cp: Bogus length %zu in firmware \"%s\"\n",
1748 rdev->ce_fw->size, fw_name);
1749 err = -EINVAL;
1750 }
1751 } else {
1752 err = radeon_ucode_validate(rdev->ce_fw);
1753 if (err) {
1754 printk(KERN_ERR
1755 "si_cp: validation failed for firmware \"%s\"\n",
1756 fw_name);
1757 goto out;
1758 } else {
1759 new_fw++;
1760 }
0f0de06c
AD
1761 }
1762
629bd33c 1763 snprintf(fw_name, sizeof(fw_name), "radeon/%s_rlc.bin", new_chip_name);
0a168933 1764 err = request_firmware(&rdev->rlc_fw, fw_name, rdev->dev);
629bd33c
AD
1765 if (err) {
1766 snprintf(fw_name, sizeof(fw_name), "radeon/%s_rlc.bin", chip_name);
1767 err = request_firmware(&rdev->rlc_fw, fw_name, rdev->dev);
1768 if (err)
1769 goto out;
1770 if (rdev->rlc_fw->size != rlc_req_size) {
1771 printk(KERN_ERR
1772 "si_rlc: Bogus length %zu in firmware \"%s\"\n",
1773 rdev->rlc_fw->size, fw_name);
1774 err = -EINVAL;
1775 }
1776 } else {
1777 err = radeon_ucode_validate(rdev->rlc_fw);
1778 if (err) {
1779 printk(KERN_ERR
1780 "si_cp: validation failed for firmware \"%s\"\n",
1781 fw_name);
1782 goto out;
1783 } else {
1784 new_fw++;
1785 }
0f0de06c
AD
1786 }
1787
629bd33c 1788 snprintf(fw_name, sizeof(fw_name), "radeon/%s_mc.bin", new_chip_name);
0a168933 1789 err = request_firmware(&rdev->mc_fw, fw_name, rdev->dev);
1ebe9280 1790 if (err) {
629bd33c 1791 snprintf(fw_name, sizeof(fw_name), "radeon/%s_mc2.bin", chip_name);
1ebe9280 1792 err = request_firmware(&rdev->mc_fw, fw_name, rdev->dev);
629bd33c
AD
1793 if (err) {
1794 snprintf(fw_name, sizeof(fw_name), "radeon/%s_mc.bin", chip_name);
1795 err = request_firmware(&rdev->mc_fw, fw_name, rdev->dev);
1796 if (err)
1797 goto out;
1798 }
1799 if ((rdev->mc_fw->size != mc_req_size) &&
1800 (rdev->mc_fw->size != mc2_req_size)) {
1801 printk(KERN_ERR
1802 "si_mc: Bogus length %zu in firmware \"%s\"\n",
1803 rdev->mc_fw->size, fw_name);
1804 err = -EINVAL;
1805 }
1806 DRM_INFO("%s: %zu bytes\n", fw_name, rdev->mc_fw->size);
1807 } else {
1808 err = radeon_ucode_validate(rdev->mc_fw);
1809 if (err) {
1810 printk(KERN_ERR
1811 "si_cp: validation failed for firmware \"%s\"\n",
1812 fw_name);
1ebe9280 1813 goto out;
629bd33c
AD
1814 } else {
1815 new_fw++;
1816 }
1ebe9280 1817 }
0f0de06c 1818
629bd33c 1819 snprintf(fw_name, sizeof(fw_name), "radeon/%s_smc.bin", new_chip_name);
0a168933 1820 err = request_firmware(&rdev->smc_fw, fw_name, rdev->dev);
8a53fa23 1821 if (err) {
629bd33c
AD
1822 snprintf(fw_name, sizeof(fw_name), "radeon/%s_smc.bin", chip_name);
1823 err = request_firmware(&rdev->smc_fw, fw_name, rdev->dev);
1824 if (err) {
1825 printk(KERN_ERR
1826 "smc: error loading firmware \"%s\"\n",
1827 fw_name);
1828 release_firmware(rdev->smc_fw);
1829 rdev->smc_fw = NULL;
1830 err = 0;
1831 } else if (rdev->smc_fw->size != smc_req_size) {
1832 printk(KERN_ERR
1833 "si_smc: Bogus length %zu in firmware \"%s\"\n",
1834 rdev->smc_fw->size, fw_name);
1835 err = -EINVAL;
1836 }
1837 } else {
1838 err = radeon_ucode_validate(rdev->smc_fw);
1839 if (err) {
1840 printk(KERN_ERR
1841 "si_cp: validation failed for firmware \"%s\"\n",
1842 fw_name);
1843 goto out;
1844 } else {
1845 new_fw++;
1846 }
a9e61410
AD
1847 }
1848
629bd33c
AD
1849 if (new_fw == 0) {
1850 rdev->new_fw = false;
1851 } else if (new_fw < 6) {
1852 printk(KERN_ERR "si_fw: mixing new and old firmware!\n");
1853 err = -EINVAL;
1854 } else {
1855 rdev->new_fw = true;
1856 }
0f0de06c 1857out:
0f0de06c
AD
1858 if (err) {
1859 if (err != -EINVAL)
1860 printk(KERN_ERR
1861 "si_cp: Failed to load firmware \"%s\"\n",
1862 fw_name);
1863 release_firmware(rdev->pfp_fw);
1864 rdev->pfp_fw = NULL;
1865 release_firmware(rdev->me_fw);
1866 rdev->me_fw = NULL;
1867 release_firmware(rdev->ce_fw);
1868 rdev->ce_fw = NULL;
1869 release_firmware(rdev->rlc_fw);
1870 rdev->rlc_fw = NULL;
1871 release_firmware(rdev->mc_fw);
1872 rdev->mc_fw = NULL;
a9e61410
AD
1873 release_firmware(rdev->smc_fw);
1874 rdev->smc_fw = NULL;
0f0de06c
AD
1875 }
1876 return err;
1877}
1878
43b3cd99
AD
1879/* watermark setup */
1880static u32 dce6_line_buffer_adjust(struct radeon_device *rdev,
1881 struct radeon_crtc *radeon_crtc,
1882 struct drm_display_mode *mode,
1883 struct drm_display_mode *other_mode)
1884{
290d2457
AD
1885 u32 tmp, buffer_alloc, i;
1886 u32 pipe_offset = radeon_crtc->crtc_id * 0x20;
43b3cd99
AD
1887 /*
1888 * Line Buffer Setup
1889 * There are 3 line buffers, each one shared by 2 display controllers.
1890 * DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
1891 * the display controllers. The paritioning is done via one of four
1892 * preset allocations specified in bits 21:20:
1893 * 0 - half lb
1894 * 2 - whole lb, other crtc must be disabled
1895 */
1896 /* this can get tricky if we have two large displays on a paired group
1897 * of crtcs. Ideally for multiple large displays we'd assign them to
1898 * non-linked crtcs for maximum line buffer allocation.
1899 */
1900 if (radeon_crtc->base.enabled && mode) {
290d2457 1901 if (other_mode) {
43b3cd99 1902 tmp = 0; /* 1/2 */
290d2457
AD
1903 buffer_alloc = 1;
1904 } else {
43b3cd99 1905 tmp = 2; /* whole */
290d2457
AD
1906 buffer_alloc = 2;
1907 }
1908 } else {
43b3cd99 1909 tmp = 0;
290d2457
AD
1910 buffer_alloc = 0;
1911 }
43b3cd99
AD
1912
1913 WREG32(DC_LB_MEMORY_SPLIT + radeon_crtc->crtc_offset,
1914 DC_LB_MEMORY_CONFIG(tmp));
1915
290d2457
AD
1916 WREG32(PIPE0_DMIF_BUFFER_CONTROL + pipe_offset,
1917 DMIF_BUFFERS_ALLOCATED(buffer_alloc));
1918 for (i = 0; i < rdev->usec_timeout; i++) {
1919 if (RREG32(PIPE0_DMIF_BUFFER_CONTROL + pipe_offset) &
1920 DMIF_BUFFERS_ALLOCATED_COMPLETED)
1921 break;
1922 udelay(1);
1923 }
1924
43b3cd99
AD
1925 if (radeon_crtc->base.enabled && mode) {
1926 switch (tmp) {
1927 case 0:
1928 default:
1929 return 4096 * 2;
1930 case 2:
1931 return 8192 * 2;
1932 }
1933 }
1934
1935 /* controller not enabled, so no lb used */
1936 return 0;
1937}
1938
ca7db22b 1939static u32 si_get_number_of_dram_channels(struct radeon_device *rdev)
43b3cd99
AD
1940{
1941 u32 tmp = RREG32(MC_SHARED_CHMAP);
1942
1943 switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
1944 case 0:
1945 default:
1946 return 1;
1947 case 1:
1948 return 2;
1949 case 2:
1950 return 4;
1951 case 3:
1952 return 8;
1953 case 4:
1954 return 3;
1955 case 5:
1956 return 6;
1957 case 6:
1958 return 10;
1959 case 7:
1960 return 12;
1961 case 8:
1962 return 16;
1963 }
1964}
1965
1966struct dce6_wm_params {
1967 u32 dram_channels; /* number of dram channels */
1968 u32 yclk; /* bandwidth per dram data pin in kHz */
1969 u32 sclk; /* engine clock in kHz */
1970 u32 disp_clk; /* display clock in kHz */
1971 u32 src_width; /* viewport width */
1972 u32 active_time; /* active display time in ns */
1973 u32 blank_time; /* blank time in ns */
1974 bool interlaced; /* mode is interlaced */
1975 fixed20_12 vsc; /* vertical scale ratio */
1976 u32 num_heads; /* number of active crtcs */
1977 u32 bytes_per_pixel; /* bytes per pixel display + overlay */
1978 u32 lb_size; /* line buffer allocated to pipe */
1979 u32 vtaps; /* vertical scaler taps */
1980};
1981
1982static u32 dce6_dram_bandwidth(struct dce6_wm_params *wm)
1983{
1984 /* Calculate raw DRAM Bandwidth */
1985 fixed20_12 dram_efficiency; /* 0.7 */
1986 fixed20_12 yclk, dram_channels, bandwidth;
1987 fixed20_12 a;
1988
1989 a.full = dfixed_const(1000);
1990 yclk.full = dfixed_const(wm->yclk);
1991 yclk.full = dfixed_div(yclk, a);
1992 dram_channels.full = dfixed_const(wm->dram_channels * 4);
1993 a.full = dfixed_const(10);
1994 dram_efficiency.full = dfixed_const(7);
1995 dram_efficiency.full = dfixed_div(dram_efficiency, a);
1996 bandwidth.full = dfixed_mul(dram_channels, yclk);
1997 bandwidth.full = dfixed_mul(bandwidth, dram_efficiency);
1998
1999 return dfixed_trunc(bandwidth);
2000}
2001
2002static u32 dce6_dram_bandwidth_for_display(struct dce6_wm_params *wm)
2003{
2004 /* Calculate DRAM Bandwidth and the part allocated to display. */
2005 fixed20_12 disp_dram_allocation; /* 0.3 to 0.7 */
2006 fixed20_12 yclk, dram_channels, bandwidth;
2007 fixed20_12 a;
2008
2009 a.full = dfixed_const(1000);
2010 yclk.full = dfixed_const(wm->yclk);
2011 yclk.full = dfixed_div(yclk, a);
2012 dram_channels.full = dfixed_const(wm->dram_channels * 4);
2013 a.full = dfixed_const(10);
2014 disp_dram_allocation.full = dfixed_const(3); /* XXX worse case value 0.3 */
2015 disp_dram_allocation.full = dfixed_div(disp_dram_allocation, a);
2016 bandwidth.full = dfixed_mul(dram_channels, yclk);
2017 bandwidth.full = dfixed_mul(bandwidth, disp_dram_allocation);
2018
2019 return dfixed_trunc(bandwidth);
2020}
2021
2022static u32 dce6_data_return_bandwidth(struct dce6_wm_params *wm)
2023{
2024 /* Calculate the display Data return Bandwidth */
2025 fixed20_12 return_efficiency; /* 0.8 */
2026 fixed20_12 sclk, bandwidth;
2027 fixed20_12 a;
2028
2029 a.full = dfixed_const(1000);
2030 sclk.full = dfixed_const(wm->sclk);
2031 sclk.full = dfixed_div(sclk, a);
2032 a.full = dfixed_const(10);
2033 return_efficiency.full = dfixed_const(8);
2034 return_efficiency.full = dfixed_div(return_efficiency, a);
2035 a.full = dfixed_const(32);
2036 bandwidth.full = dfixed_mul(a, sclk);
2037 bandwidth.full = dfixed_mul(bandwidth, return_efficiency);
2038
2039 return dfixed_trunc(bandwidth);
2040}
2041
2042static u32 dce6_get_dmif_bytes_per_request(struct dce6_wm_params *wm)
2043{
2044 return 32;
2045}
2046
2047static u32 dce6_dmif_request_bandwidth(struct dce6_wm_params *wm)
2048{
2049 /* Calculate the DMIF Request Bandwidth */
2050 fixed20_12 disp_clk_request_efficiency; /* 0.8 */
2051 fixed20_12 disp_clk, sclk, bandwidth;
2052 fixed20_12 a, b1, b2;
2053 u32 min_bandwidth;
2054
2055 a.full = dfixed_const(1000);
2056 disp_clk.full = dfixed_const(wm->disp_clk);
2057 disp_clk.full = dfixed_div(disp_clk, a);
2058 a.full = dfixed_const(dce6_get_dmif_bytes_per_request(wm) / 2);
2059 b1.full = dfixed_mul(a, disp_clk);
2060
2061 a.full = dfixed_const(1000);
2062 sclk.full = dfixed_const(wm->sclk);
2063 sclk.full = dfixed_div(sclk, a);
2064 a.full = dfixed_const(dce6_get_dmif_bytes_per_request(wm));
2065 b2.full = dfixed_mul(a, sclk);
2066
2067 a.full = dfixed_const(10);
2068 disp_clk_request_efficiency.full = dfixed_const(8);
2069 disp_clk_request_efficiency.full = dfixed_div(disp_clk_request_efficiency, a);
2070
2071 min_bandwidth = min(dfixed_trunc(b1), dfixed_trunc(b2));
2072
2073 a.full = dfixed_const(min_bandwidth);
2074 bandwidth.full = dfixed_mul(a, disp_clk_request_efficiency);
2075
2076 return dfixed_trunc(bandwidth);
2077}
2078
2079static u32 dce6_available_bandwidth(struct dce6_wm_params *wm)
2080{
2081 /* Calculate the Available bandwidth. Display can use this temporarily but not in average. */
2082 u32 dram_bandwidth = dce6_dram_bandwidth(wm);
2083 u32 data_return_bandwidth = dce6_data_return_bandwidth(wm);
2084 u32 dmif_req_bandwidth = dce6_dmif_request_bandwidth(wm);
2085
2086 return min(dram_bandwidth, min(data_return_bandwidth, dmif_req_bandwidth));
2087}
2088
2089static u32 dce6_average_bandwidth(struct dce6_wm_params *wm)
2090{
2091 /* Calculate the display mode Average Bandwidth
2092 * DisplayMode should contain the source and destination dimensions,
2093 * timing, etc.
2094 */
2095 fixed20_12 bpp;
2096 fixed20_12 line_time;
2097 fixed20_12 src_width;
2098 fixed20_12 bandwidth;
2099 fixed20_12 a;
2100
2101 a.full = dfixed_const(1000);
2102 line_time.full = dfixed_const(wm->active_time + wm->blank_time);
2103 line_time.full = dfixed_div(line_time, a);
2104 bpp.full = dfixed_const(wm->bytes_per_pixel);
2105 src_width.full = dfixed_const(wm->src_width);
2106 bandwidth.full = dfixed_mul(src_width, bpp);
2107 bandwidth.full = dfixed_mul(bandwidth, wm->vsc);
2108 bandwidth.full = dfixed_div(bandwidth, line_time);
2109
2110 return dfixed_trunc(bandwidth);
2111}
2112
2113static u32 dce6_latency_watermark(struct dce6_wm_params *wm)
2114{
2115 /* First calcualte the latency in ns */
2116 u32 mc_latency = 2000; /* 2000 ns. */
2117 u32 available_bandwidth = dce6_available_bandwidth(wm);
2118 u32 worst_chunk_return_time = (512 * 8 * 1000) / available_bandwidth;
2119 u32 cursor_line_pair_return_time = (128 * 4 * 1000) / available_bandwidth;
2120 u32 dc_latency = 40000000 / wm->disp_clk; /* dc pipe latency */
2121 u32 other_heads_data_return_time = ((wm->num_heads + 1) * worst_chunk_return_time) +
2122 (wm->num_heads * cursor_line_pair_return_time);
2123 u32 latency = mc_latency + other_heads_data_return_time + dc_latency;
2124 u32 max_src_lines_per_dst_line, lb_fill_bw, line_fill_time;
2125 u32 tmp, dmif_size = 12288;
2126 fixed20_12 a, b, c;
2127
2128 if (wm->num_heads == 0)
2129 return 0;
2130
2131 a.full = dfixed_const(2);
2132 b.full = dfixed_const(1);
2133 if ((wm->vsc.full > a.full) ||
2134 ((wm->vsc.full > b.full) && (wm->vtaps >= 3)) ||
2135 (wm->vtaps >= 5) ||
2136 ((wm->vsc.full >= a.full) && wm->interlaced))
2137 max_src_lines_per_dst_line = 4;
2138 else
2139 max_src_lines_per_dst_line = 2;
2140
2141 a.full = dfixed_const(available_bandwidth);
2142 b.full = dfixed_const(wm->num_heads);
2143 a.full = dfixed_div(a, b);
2144
2145 b.full = dfixed_const(mc_latency + 512);
2146 c.full = dfixed_const(wm->disp_clk);
2147 b.full = dfixed_div(b, c);
2148
2149 c.full = dfixed_const(dmif_size);
2150 b.full = dfixed_div(c, b);
2151
2152 tmp = min(dfixed_trunc(a), dfixed_trunc(b));
2153
2154 b.full = dfixed_const(1000);
2155 c.full = dfixed_const(wm->disp_clk);
2156 b.full = dfixed_div(c, b);
2157 c.full = dfixed_const(wm->bytes_per_pixel);
2158 b.full = dfixed_mul(b, c);
2159
2160 lb_fill_bw = min(tmp, dfixed_trunc(b));
2161
2162 a.full = dfixed_const(max_src_lines_per_dst_line * wm->src_width * wm->bytes_per_pixel);
2163 b.full = dfixed_const(1000);
2164 c.full = dfixed_const(lb_fill_bw);
2165 b.full = dfixed_div(c, b);
2166 a.full = dfixed_div(a, b);
2167 line_fill_time = dfixed_trunc(a);
2168
2169 if (line_fill_time < wm->active_time)
2170 return latency;
2171 else
2172 return latency + (line_fill_time - wm->active_time);
2173
2174}
2175
2176static bool dce6_average_bandwidth_vs_dram_bandwidth_for_display(struct dce6_wm_params *wm)
2177{
2178 if (dce6_average_bandwidth(wm) <=
2179 (dce6_dram_bandwidth_for_display(wm) / wm->num_heads))
2180 return true;
2181 else
2182 return false;
2183};
2184
2185static bool dce6_average_bandwidth_vs_available_bandwidth(struct dce6_wm_params *wm)
2186{
2187 if (dce6_average_bandwidth(wm) <=
2188 (dce6_available_bandwidth(wm) / wm->num_heads))
2189 return true;
2190 else
2191 return false;
2192};
2193
2194static bool dce6_check_latency_hiding(struct dce6_wm_params *wm)
2195{
2196 u32 lb_partitions = wm->lb_size / wm->src_width;
2197 u32 line_time = wm->active_time + wm->blank_time;
2198 u32 latency_tolerant_lines;
2199 u32 latency_hiding;
2200 fixed20_12 a;
2201
2202 a.full = dfixed_const(1);
2203 if (wm->vsc.full > a.full)
2204 latency_tolerant_lines = 1;
2205 else {
2206 if (lb_partitions <= (wm->vtaps + 1))
2207 latency_tolerant_lines = 1;
2208 else
2209 latency_tolerant_lines = 2;
2210 }
2211
2212 latency_hiding = (latency_tolerant_lines * line_time + wm->blank_time);
2213
2214 if (dce6_latency_watermark(wm) <= latency_hiding)
2215 return true;
2216 else
2217 return false;
2218}
2219
2220static void dce6_program_watermarks(struct radeon_device *rdev,
2221 struct radeon_crtc *radeon_crtc,
2222 u32 lb_size, u32 num_heads)
2223{
2224 struct drm_display_mode *mode = &radeon_crtc->base.mode;
c696e53f
AD
2225 struct dce6_wm_params wm_low, wm_high;
2226 u32 dram_channels;
43b3cd99
AD
2227 u32 pixel_period;
2228 u32 line_time = 0;
2229 u32 latency_watermark_a = 0, latency_watermark_b = 0;
2230 u32 priority_a_mark = 0, priority_b_mark = 0;
2231 u32 priority_a_cnt = PRIORITY_OFF;
2232 u32 priority_b_cnt = PRIORITY_OFF;
2233 u32 tmp, arb_control3;
2234 fixed20_12 a, b, c;
2235
2236 if (radeon_crtc->base.enabled && num_heads && mode) {
2237 pixel_period = 1000000 / (u32)mode->clock;
2238 line_time = min((u32)mode->crtc_htotal * pixel_period, (u32)65535);
2239 priority_a_cnt = 0;
2240 priority_b_cnt = 0;
2241
ca7db22b 2242 if (rdev->family == CHIP_ARUBA)
c696e53f 2243 dram_channels = evergreen_get_number_of_dram_channels(rdev);
ca7db22b 2244 else
c696e53f
AD
2245 dram_channels = si_get_number_of_dram_channels(rdev);
2246
2247 /* watermark for high clocks */
2248 if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
2249 wm_high.yclk =
2250 radeon_dpm_get_mclk(rdev, false) * 10;
2251 wm_high.sclk =
2252 radeon_dpm_get_sclk(rdev, false) * 10;
2253 } else {
2254 wm_high.yclk = rdev->pm.current_mclk * 10;
2255 wm_high.sclk = rdev->pm.current_sclk * 10;
2256 }
2257
2258 wm_high.disp_clk = mode->clock;
2259 wm_high.src_width = mode->crtc_hdisplay;
2260 wm_high.active_time = mode->crtc_hdisplay * pixel_period;
2261 wm_high.blank_time = line_time - wm_high.active_time;
2262 wm_high.interlaced = false;
2263 if (mode->flags & DRM_MODE_FLAG_INTERLACE)
2264 wm_high.interlaced = true;
2265 wm_high.vsc = radeon_crtc->vsc;
2266 wm_high.vtaps = 1;
2267 if (radeon_crtc->rmx_type != RMX_OFF)
2268 wm_high.vtaps = 2;
2269 wm_high.bytes_per_pixel = 4; /* XXX: get this from fb config */
2270 wm_high.lb_size = lb_size;
2271 wm_high.dram_channels = dram_channels;
2272 wm_high.num_heads = num_heads;
2273
2274 /* watermark for low clocks */
2275 if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
2276 wm_low.yclk =
2277 radeon_dpm_get_mclk(rdev, true) * 10;
2278 wm_low.sclk =
2279 radeon_dpm_get_sclk(rdev, true) * 10;
2280 } else {
2281 wm_low.yclk = rdev->pm.current_mclk * 10;
2282 wm_low.sclk = rdev->pm.current_sclk * 10;
2283 }
2284
2285 wm_low.disp_clk = mode->clock;
2286 wm_low.src_width = mode->crtc_hdisplay;
2287 wm_low.active_time = mode->crtc_hdisplay * pixel_period;
2288 wm_low.blank_time = line_time - wm_low.active_time;
2289 wm_low.interlaced = false;
2290 if (mode->flags & DRM_MODE_FLAG_INTERLACE)
2291 wm_low.interlaced = true;
2292 wm_low.vsc = radeon_crtc->vsc;
2293 wm_low.vtaps = 1;
2294 if (radeon_crtc->rmx_type != RMX_OFF)
2295 wm_low.vtaps = 2;
2296 wm_low.bytes_per_pixel = 4; /* XXX: get this from fb config */
2297 wm_low.lb_size = lb_size;
2298 wm_low.dram_channels = dram_channels;
2299 wm_low.num_heads = num_heads;
43b3cd99
AD
2300
2301 /* set for high clocks */
c696e53f 2302 latency_watermark_a = min(dce6_latency_watermark(&wm_high), (u32)65535);
43b3cd99 2303 /* set for low clocks */
c696e53f 2304 latency_watermark_b = min(dce6_latency_watermark(&wm_low), (u32)65535);
43b3cd99
AD
2305
2306 /* possibly force display priority to high */
2307 /* should really do this at mode validation time... */
c696e53f
AD
2308 if (!dce6_average_bandwidth_vs_dram_bandwidth_for_display(&wm_high) ||
2309 !dce6_average_bandwidth_vs_available_bandwidth(&wm_high) ||
2310 !dce6_check_latency_hiding(&wm_high) ||
2311 (rdev->disp_priority == 2)) {
2312 DRM_DEBUG_KMS("force priority to high\n");
2313 priority_a_cnt |= PRIORITY_ALWAYS_ON;
2314 priority_b_cnt |= PRIORITY_ALWAYS_ON;
2315 }
2316 if (!dce6_average_bandwidth_vs_dram_bandwidth_for_display(&wm_low) ||
2317 !dce6_average_bandwidth_vs_available_bandwidth(&wm_low) ||
2318 !dce6_check_latency_hiding(&wm_low) ||
43b3cd99
AD
2319 (rdev->disp_priority == 2)) {
2320 DRM_DEBUG_KMS("force priority to high\n");
2321 priority_a_cnt |= PRIORITY_ALWAYS_ON;
2322 priority_b_cnt |= PRIORITY_ALWAYS_ON;
2323 }
2324
2325 a.full = dfixed_const(1000);
2326 b.full = dfixed_const(mode->clock);
2327 b.full = dfixed_div(b, a);
2328 c.full = dfixed_const(latency_watermark_a);
2329 c.full = dfixed_mul(c, b);
2330 c.full = dfixed_mul(c, radeon_crtc->hsc);
2331 c.full = dfixed_div(c, a);
2332 a.full = dfixed_const(16);
2333 c.full = dfixed_div(c, a);
2334 priority_a_mark = dfixed_trunc(c);
2335 priority_a_cnt |= priority_a_mark & PRIORITY_MARK_MASK;
2336
2337 a.full = dfixed_const(1000);
2338 b.full = dfixed_const(mode->clock);
2339 b.full = dfixed_div(b, a);
2340 c.full = dfixed_const(latency_watermark_b);
2341 c.full = dfixed_mul(c, b);
2342 c.full = dfixed_mul(c, radeon_crtc->hsc);
2343 c.full = dfixed_div(c, a);
2344 a.full = dfixed_const(16);
2345 c.full = dfixed_div(c, a);
2346 priority_b_mark = dfixed_trunc(c);
2347 priority_b_cnt |= priority_b_mark & PRIORITY_MARK_MASK;
2348 }
2349
2350 /* select wm A */
2351 arb_control3 = RREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset);
2352 tmp = arb_control3;
2353 tmp &= ~LATENCY_WATERMARK_MASK(3);
2354 tmp |= LATENCY_WATERMARK_MASK(1);
2355 WREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset, tmp);
2356 WREG32(DPG_PIPE_LATENCY_CONTROL + radeon_crtc->crtc_offset,
2357 (LATENCY_LOW_WATERMARK(latency_watermark_a) |
2358 LATENCY_HIGH_WATERMARK(line_time)));
2359 /* select wm B */
2360 tmp = RREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset);
2361 tmp &= ~LATENCY_WATERMARK_MASK(3);
2362 tmp |= LATENCY_WATERMARK_MASK(2);
2363 WREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset, tmp);
2364 WREG32(DPG_PIPE_LATENCY_CONTROL + radeon_crtc->crtc_offset,
2365 (LATENCY_LOW_WATERMARK(latency_watermark_b) |
2366 LATENCY_HIGH_WATERMARK(line_time)));
2367 /* restore original selection */
2368 WREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset, arb_control3);
2369
2370 /* write the priority marks */
2371 WREG32(PRIORITY_A_CNT + radeon_crtc->crtc_offset, priority_a_cnt);
2372 WREG32(PRIORITY_B_CNT + radeon_crtc->crtc_offset, priority_b_cnt);
2373
7178d2a6
AD
2374 /* save values for DPM */
2375 radeon_crtc->line_time = line_time;
2376 radeon_crtc->wm_high = latency_watermark_a;
2377 radeon_crtc->wm_low = latency_watermark_b;
43b3cd99
AD
2378}
2379
2380void dce6_bandwidth_update(struct radeon_device *rdev)
2381{
2382 struct drm_display_mode *mode0 = NULL;
2383 struct drm_display_mode *mode1 = NULL;
2384 u32 num_heads = 0, lb_size;
2385 int i;
2386
2387 radeon_update_display_priority(rdev);
2388
2389 for (i = 0; i < rdev->num_crtc; i++) {
2390 if (rdev->mode_info.crtcs[i]->base.enabled)
2391 num_heads++;
2392 }
2393 for (i = 0; i < rdev->num_crtc; i += 2) {
2394 mode0 = &rdev->mode_info.crtcs[i]->base.mode;
2395 mode1 = &rdev->mode_info.crtcs[i+1]->base.mode;
2396 lb_size = dce6_line_buffer_adjust(rdev, rdev->mode_info.crtcs[i], mode0, mode1);
2397 dce6_program_watermarks(rdev, rdev->mode_info.crtcs[i], lb_size, num_heads);
2398 lb_size = dce6_line_buffer_adjust(rdev, rdev->mode_info.crtcs[i+1], mode1, mode0);
2399 dce6_program_watermarks(rdev, rdev->mode_info.crtcs[i+1], lb_size, num_heads);
2400 }
2401}
2402
0a96d72b
AD
2403/*
2404 * Core functions
2405 */
0a96d72b
AD
2406static void si_tiling_mode_table_init(struct radeon_device *rdev)
2407{
2408 const u32 num_tile_mode_states = 32;
2409 u32 reg_offset, gb_tile_moden, split_equal_to_row_size;
2410
2411 switch (rdev->config.si.mem_row_size_in_kb) {
2412 case 1:
2413 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_1KB;
2414 break;
2415 case 2:
2416 default:
2417 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_2KB;
2418 break;
2419 case 4:
2420 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_4KB;
2421 break;
2422 }
2423
2424 if ((rdev->family == CHIP_TAHITI) ||
2425 (rdev->family == CHIP_PITCAIRN)) {
2426 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
2427 switch (reg_offset) {
2428 case 0: /* non-AA compressed depth or any compressed stencil */
2429 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2430 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2431 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2432 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2433 NUM_BANKS(ADDR_SURF_16_BANK) |
2434 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2435 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2436 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2437 break;
2438 case 1: /* 2xAA/4xAA compressed depth only */
2439 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2440 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2441 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2442 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2443 NUM_BANKS(ADDR_SURF_16_BANK) |
2444 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2445 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2446 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2447 break;
2448 case 2: /* 8xAA compressed depth only */
2449 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2450 MICRO_TILE_MODE(ADDR_SURF_DEPTH_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_2));
2457 break;
2458 case 3: /* 2xAA/4xAA compressed depth with stencil (for depth buffer) */
2459 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2460 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2461 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2462 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2463 NUM_BANKS(ADDR_SURF_16_BANK) |
2464 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2465 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2466 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2467 break;
2468 case 4: /* Maps w/ a dimension less than the 2D macro-tile dimensions (for mipmapped depth textures) */
2469 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2470 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2471 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2472 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2473 NUM_BANKS(ADDR_SURF_16_BANK) |
2474 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2475 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2476 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2477 break;
2478 case 5: /* Uncompressed 16bpp depth - and stencil buffer allocated with it */
2479 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2480 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2481 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2482 TILE_SPLIT(split_equal_to_row_size) |
2483 NUM_BANKS(ADDR_SURF_16_BANK) |
2484 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2485 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2486 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2487 break;
2488 case 6: /* Uncompressed 32bpp depth - and stencil buffer allocated with it */
2489 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2490 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2491 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2492 TILE_SPLIT(split_equal_to_row_size) |
2493 NUM_BANKS(ADDR_SURF_16_BANK) |
2494 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2495 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2496 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2497 break;
2498 case 7: /* Uncompressed 8bpp stencil without depth (drivers typically do not use) */
2499 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2500 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2501 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2502 TILE_SPLIT(split_equal_to_row_size) |
2503 NUM_BANKS(ADDR_SURF_16_BANK) |
2504 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2505 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2506 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2507 break;
2508 case 8: /* 1D and 1D Array Surfaces */
2509 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
2510 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2511 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2512 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2513 NUM_BANKS(ADDR_SURF_16_BANK) |
2514 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2515 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2516 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2517 break;
2518 case 9: /* Displayable maps. */
2519 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2520 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2521 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2522 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2523 NUM_BANKS(ADDR_SURF_16_BANK) |
2524 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2525 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2526 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2527 break;
2528 case 10: /* Display 8bpp. */
2529 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2530 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2531 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2532 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2533 NUM_BANKS(ADDR_SURF_16_BANK) |
2534 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2535 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2536 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2537 break;
2538 case 11: /* Display 16bpp. */
2539 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2540 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2541 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2542 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2543 NUM_BANKS(ADDR_SURF_16_BANK) |
2544 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2545 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2546 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2547 break;
2548 case 12: /* Display 32bpp. */
2549 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2550 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2551 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2552 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2553 NUM_BANKS(ADDR_SURF_16_BANK) |
2554 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2555 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2556 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2557 break;
2558 case 13: /* Thin. */
2559 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2560 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2561 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2562 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2563 NUM_BANKS(ADDR_SURF_16_BANK) |
2564 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2565 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2566 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2567 break;
2568 case 14: /* Thin 8 bpp. */
2569 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2570 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2571 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2572 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2573 NUM_BANKS(ADDR_SURF_16_BANK) |
2574 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2575 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2576 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2577 break;
2578 case 15: /* Thin 16 bpp. */
2579 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2580 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2581 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2582 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2583 NUM_BANKS(ADDR_SURF_16_BANK) |
2584 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2585 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2586 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2587 break;
2588 case 16: /* Thin 32 bpp. */
2589 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2590 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2591 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2592 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2593 NUM_BANKS(ADDR_SURF_16_BANK) |
2594 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2595 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2596 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2597 break;
2598 case 17: /* Thin 64 bpp. */
2599 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2600 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2601 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2602 TILE_SPLIT(split_equal_to_row_size) |
2603 NUM_BANKS(ADDR_SURF_16_BANK) |
2604 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2605 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2606 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2607 break;
2608 case 21: /* 8 bpp PRT. */
2609 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2610 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2611 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2612 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2613 NUM_BANKS(ADDR_SURF_16_BANK) |
2614 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
2615 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2616 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2617 break;
2618 case 22: /* 16 bpp PRT */
2619 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2620 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2621 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2622 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2623 NUM_BANKS(ADDR_SURF_16_BANK) |
2624 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2625 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2626 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2627 break;
2628 case 23: /* 32 bpp PRT */
2629 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2630 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2631 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2632 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2633 NUM_BANKS(ADDR_SURF_16_BANK) |
2634 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2635 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2636 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2637 break;
2638 case 24: /* 64 bpp PRT */
2639 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2640 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2641 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2642 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2643 NUM_BANKS(ADDR_SURF_16_BANK) |
2644 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2645 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2646 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2647 break;
2648 case 25: /* 128 bpp PRT */
2649 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2650 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2651 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2652 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
2653 NUM_BANKS(ADDR_SURF_8_BANK) |
2654 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2655 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2656 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2657 break;
2658 default:
2659 gb_tile_moden = 0;
2660 break;
2661 }
64d7b8be 2662 rdev->config.si.tile_mode_array[reg_offset] = gb_tile_moden;
0a96d72b
AD
2663 WREG32(GB_TILE_MODE0 + (reg_offset * 4), gb_tile_moden);
2664 }
d0ae7fcc 2665 } else if ((rdev->family == CHIP_VERDE) ||
8b02859d
AD
2666 (rdev->family == CHIP_OLAND) ||
2667 (rdev->family == CHIP_HAINAN)) {
0a96d72b
AD
2668 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
2669 switch (reg_offset) {
2670 case 0: /* non-AA compressed depth or any compressed stencil */
2671 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2672 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2673 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2674 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2675 NUM_BANKS(ADDR_SURF_16_BANK) |
2676 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2677 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2678 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2679 break;
2680 case 1: /* 2xAA/4xAA compressed depth only */
2681 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2682 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2683 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2684 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2685 NUM_BANKS(ADDR_SURF_16_BANK) |
2686 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2687 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2688 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2689 break;
2690 case 2: /* 8xAA compressed depth only */
2691 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2692 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2693 PIPE_CONFIG(ADDR_SURF_P4_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 3: /* 2xAA/4xAA compressed depth with stencil (for depth buffer) */
2701 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2702 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2703 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2704 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2705 NUM_BANKS(ADDR_SURF_16_BANK) |
2706 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2707 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2708 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2709 break;
2710 case 4: /* Maps w/ a dimension less than the 2D macro-tile dimensions (for mipmapped depth textures) */
2711 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2712 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2713 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2714 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2715 NUM_BANKS(ADDR_SURF_16_BANK) |
2716 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2717 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2718 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2719 break;
2720 case 5: /* Uncompressed 16bpp depth - and stencil buffer allocated with it */
2721 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2722 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2723 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2724 TILE_SPLIT(split_equal_to_row_size) |
2725 NUM_BANKS(ADDR_SURF_16_BANK) |
2726 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2727 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2728 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2729 break;
2730 case 6: /* Uncompressed 32bpp depth - and stencil buffer allocated with it */
2731 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2732 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2733 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2734 TILE_SPLIT(split_equal_to_row_size) |
2735 NUM_BANKS(ADDR_SURF_16_BANK) |
2736 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2737 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2738 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2739 break;
2740 case 7: /* Uncompressed 8bpp stencil without depth (drivers typically do not use) */
2741 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2742 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2743 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2744 TILE_SPLIT(split_equal_to_row_size) |
2745 NUM_BANKS(ADDR_SURF_16_BANK) |
2746 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2747 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2748 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2749 break;
2750 case 8: /* 1D and 1D Array Surfaces */
2751 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
2752 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2753 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2754 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2755 NUM_BANKS(ADDR_SURF_16_BANK) |
2756 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2757 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2758 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2759 break;
2760 case 9: /* Displayable maps. */
2761 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2762 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2763 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2764 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2765 NUM_BANKS(ADDR_SURF_16_BANK) |
2766 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2767 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2768 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2769 break;
2770 case 10: /* Display 8bpp. */
2771 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2772 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2773 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2774 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2775 NUM_BANKS(ADDR_SURF_16_BANK) |
2776 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2777 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2778 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2779 break;
2780 case 11: /* Display 16bpp. */
2781 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2782 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2783 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2784 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2785 NUM_BANKS(ADDR_SURF_16_BANK) |
2786 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2787 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2788 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2789 break;
2790 case 12: /* Display 32bpp. */
2791 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2792 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2793 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2794 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2795 NUM_BANKS(ADDR_SURF_16_BANK) |
2796 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2797 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2798 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2799 break;
2800 case 13: /* Thin. */
2801 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2802 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2803 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2804 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2805 NUM_BANKS(ADDR_SURF_16_BANK) |
2806 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2807 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2808 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2809 break;
2810 case 14: /* Thin 8 bpp. */
2811 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2812 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2813 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2814 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2815 NUM_BANKS(ADDR_SURF_16_BANK) |
2816 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2817 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2818 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2819 break;
2820 case 15: /* Thin 16 bpp. */
2821 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2822 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2823 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2824 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2825 NUM_BANKS(ADDR_SURF_16_BANK) |
2826 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2827 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2828 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2829 break;
2830 case 16: /* Thin 32 bpp. */
2831 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2832 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2833 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2834 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2835 NUM_BANKS(ADDR_SURF_16_BANK) |
2836 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2837 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2838 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2839 break;
2840 case 17: /* Thin 64 bpp. */
2841 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2842 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2843 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2844 TILE_SPLIT(split_equal_to_row_size) |
2845 NUM_BANKS(ADDR_SURF_16_BANK) |
2846 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2847 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2848 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2849 break;
2850 case 21: /* 8 bpp PRT. */
2851 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2852 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2853 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2854 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2855 NUM_BANKS(ADDR_SURF_16_BANK) |
2856 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
2857 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2858 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2859 break;
2860 case 22: /* 16 bpp PRT */
2861 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2862 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2863 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2864 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2865 NUM_BANKS(ADDR_SURF_16_BANK) |
2866 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2867 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2868 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2869 break;
2870 case 23: /* 32 bpp PRT */
2871 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2872 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2873 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2874 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2875 NUM_BANKS(ADDR_SURF_16_BANK) |
2876 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2877 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2878 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2879 break;
2880 case 24: /* 64 bpp PRT */
2881 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2882 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2883 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2884 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2885 NUM_BANKS(ADDR_SURF_16_BANK) |
2886 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2887 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2888 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2889 break;
2890 case 25: /* 128 bpp PRT */
2891 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2892 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2893 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2894 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
2895 NUM_BANKS(ADDR_SURF_8_BANK) |
2896 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2897 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2898 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2899 break;
2900 default:
2901 gb_tile_moden = 0;
2902 break;
2903 }
64d7b8be 2904 rdev->config.si.tile_mode_array[reg_offset] = gb_tile_moden;
0a96d72b
AD
2905 WREG32(GB_TILE_MODE0 + (reg_offset * 4), gb_tile_moden);
2906 }
2907 } else
2908 DRM_ERROR("unknown asic: 0x%x\n", rdev->family);
2909}
2910
1a8ca750
AD
2911static void si_select_se_sh(struct radeon_device *rdev,
2912 u32 se_num, u32 sh_num)
2913{
2914 u32 data = INSTANCE_BROADCAST_WRITES;
2915
2916 if ((se_num == 0xffffffff) && (sh_num == 0xffffffff))
79b52d6a 2917 data |= SH_BROADCAST_WRITES | SE_BROADCAST_WRITES;
1a8ca750
AD
2918 else if (se_num == 0xffffffff)
2919 data |= SE_BROADCAST_WRITES | SH_INDEX(sh_num);
2920 else if (sh_num == 0xffffffff)
2921 data |= SH_BROADCAST_WRITES | SE_INDEX(se_num);
2922 else
2923 data |= SH_INDEX(sh_num) | SE_INDEX(se_num);
2924 WREG32(GRBM_GFX_INDEX, data);
2925}
2926
2927static u32 si_create_bitmask(u32 bit_width)
2928{
2929 u32 i, mask = 0;
2930
2931 for (i = 0; i < bit_width; i++) {
2932 mask <<= 1;
2933 mask |= 1;
2934 }
2935 return mask;
2936}
2937
2938static u32 si_get_cu_enabled(struct radeon_device *rdev, u32 cu_per_sh)
2939{
2940 u32 data, mask;
2941
2942 data = RREG32(CC_GC_SHADER_ARRAY_CONFIG);
2943 if (data & 1)
2944 data &= INACTIVE_CUS_MASK;
2945 else
2946 data = 0;
2947 data |= RREG32(GC_USER_SHADER_ARRAY_CONFIG);
2948
2949 data >>= INACTIVE_CUS_SHIFT;
2950
2951 mask = si_create_bitmask(cu_per_sh);
2952
2953 return ~data & mask;
2954}
2955
2956static void si_setup_spi(struct radeon_device *rdev,
2957 u32 se_num, u32 sh_per_se,
2958 u32 cu_per_sh)
2959{
2960 int i, j, k;
2961 u32 data, mask, active_cu;
2962
2963 for (i = 0; i < se_num; i++) {
2964 for (j = 0; j < sh_per_se; j++) {
2965 si_select_se_sh(rdev, i, j);
2966 data = RREG32(SPI_STATIC_THREAD_MGMT_3);
2967 active_cu = si_get_cu_enabled(rdev, cu_per_sh);
2968
2969 mask = 1;
2970 for (k = 0; k < 16; k++) {
2971 mask <<= k;
2972 if (active_cu & mask) {
2973 data &= ~mask;
2974 WREG32(SPI_STATIC_THREAD_MGMT_3, data);
2975 break;
2976 }
2977 }
2978 }
2979 }
2980 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
2981}
2982
2983static u32 si_get_rb_disabled(struct radeon_device *rdev,
9fadb352 2984 u32 max_rb_num_per_se,
1a8ca750
AD
2985 u32 sh_per_se)
2986{
2987 u32 data, mask;
2988
2989 data = RREG32(CC_RB_BACKEND_DISABLE);
2990 if (data & 1)
2991 data &= BACKEND_DISABLE_MASK;
2992 else
2993 data = 0;
2994 data |= RREG32(GC_USER_RB_BACKEND_DISABLE);
2995
2996 data >>= BACKEND_DISABLE_SHIFT;
2997
9fadb352 2998 mask = si_create_bitmask(max_rb_num_per_se / sh_per_se);
1a8ca750
AD
2999
3000 return data & mask;
3001}
3002
3003static void si_setup_rb(struct radeon_device *rdev,
3004 u32 se_num, u32 sh_per_se,
9fadb352 3005 u32 max_rb_num_per_se)
1a8ca750
AD
3006{
3007 int i, j;
3008 u32 data, mask;
3009 u32 disabled_rbs = 0;
3010 u32 enabled_rbs = 0;
3011
3012 for (i = 0; i < se_num; i++) {
3013 for (j = 0; j < sh_per_se; j++) {
3014 si_select_se_sh(rdev, i, j);
9fadb352 3015 data = si_get_rb_disabled(rdev, max_rb_num_per_se, sh_per_se);
1a8ca750
AD
3016 disabled_rbs |= data << ((i * sh_per_se + j) * TAHITI_RB_BITMAP_WIDTH_PER_SH);
3017 }
3018 }
3019 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
3020
3021 mask = 1;
9fadb352 3022 for (i = 0; i < max_rb_num_per_se * se_num; i++) {
1a8ca750
AD
3023 if (!(disabled_rbs & mask))
3024 enabled_rbs |= mask;
3025 mask <<= 1;
3026 }
3027
439a1cff
MO
3028 rdev->config.si.backend_enable_mask = enabled_rbs;
3029
1a8ca750
AD
3030 for (i = 0; i < se_num; i++) {
3031 si_select_se_sh(rdev, i, 0xffffffff);
3032 data = 0;
3033 for (j = 0; j < sh_per_se; j++) {
3034 switch (enabled_rbs & 3) {
3035 case 1:
3036 data |= (RASTER_CONFIG_RB_MAP_0 << (i * sh_per_se + j) * 2);
3037 break;
3038 case 2:
3039 data |= (RASTER_CONFIG_RB_MAP_3 << (i * sh_per_se + j) * 2);
3040 break;
3041 case 3:
3042 default:
3043 data |= (RASTER_CONFIG_RB_MAP_2 << (i * sh_per_se + j) * 2);
3044 break;
3045 }
3046 enabled_rbs >>= 2;
3047 }
3048 WREG32(PA_SC_RASTER_CONFIG, data);
3049 }
3050 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
3051}
3052
0a96d72b
AD
3053static void si_gpu_init(struct radeon_device *rdev)
3054{
0a96d72b
AD
3055 u32 gb_addr_config = 0;
3056 u32 mc_shared_chmap, mc_arb_ramcfg;
0a96d72b 3057 u32 sx_debug_1;
0a96d72b
AD
3058 u32 hdp_host_path_cntl;
3059 u32 tmp;
6101b3ae 3060 int i, j;
0a96d72b
AD
3061
3062 switch (rdev->family) {
3063 case CHIP_TAHITI:
3064 rdev->config.si.max_shader_engines = 2;
0a96d72b 3065 rdev->config.si.max_tile_pipes = 12;
1a8ca750
AD
3066 rdev->config.si.max_cu_per_sh = 8;
3067 rdev->config.si.max_sh_per_se = 2;
0a96d72b
AD
3068 rdev->config.si.max_backends_per_se = 4;
3069 rdev->config.si.max_texture_channel_caches = 12;
3070 rdev->config.si.max_gprs = 256;
3071 rdev->config.si.max_gs_threads = 32;
3072 rdev->config.si.max_hw_contexts = 8;
3073
3074 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
3075 rdev->config.si.sc_prim_fifo_size_backend = 0x100;
3076 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
3077 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
1a8ca750 3078 gb_addr_config = TAHITI_GB_ADDR_CONFIG_GOLDEN;
0a96d72b
AD
3079 break;
3080 case CHIP_PITCAIRN:
3081 rdev->config.si.max_shader_engines = 2;
0a96d72b 3082 rdev->config.si.max_tile_pipes = 8;
1a8ca750
AD
3083 rdev->config.si.max_cu_per_sh = 5;
3084 rdev->config.si.max_sh_per_se = 2;
0a96d72b
AD
3085 rdev->config.si.max_backends_per_se = 4;
3086 rdev->config.si.max_texture_channel_caches = 8;
3087 rdev->config.si.max_gprs = 256;
3088 rdev->config.si.max_gs_threads = 32;
3089 rdev->config.si.max_hw_contexts = 8;
3090
3091 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
3092 rdev->config.si.sc_prim_fifo_size_backend = 0x100;
3093 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
3094 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
1a8ca750 3095 gb_addr_config = TAHITI_GB_ADDR_CONFIG_GOLDEN;
0a96d72b
AD
3096 break;
3097 case CHIP_VERDE:
3098 default:
3099 rdev->config.si.max_shader_engines = 1;
0a96d72b 3100 rdev->config.si.max_tile_pipes = 4;
468ef1a5 3101 rdev->config.si.max_cu_per_sh = 5;
1a8ca750 3102 rdev->config.si.max_sh_per_se = 2;
0a96d72b
AD
3103 rdev->config.si.max_backends_per_se = 4;
3104 rdev->config.si.max_texture_channel_caches = 4;
3105 rdev->config.si.max_gprs = 256;
3106 rdev->config.si.max_gs_threads = 32;
3107 rdev->config.si.max_hw_contexts = 8;
3108
d0ae7fcc
AD
3109 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
3110 rdev->config.si.sc_prim_fifo_size_backend = 0x40;
3111 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
3112 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
3113 gb_addr_config = VERDE_GB_ADDR_CONFIG_GOLDEN;
3114 break;
3115 case CHIP_OLAND:
3116 rdev->config.si.max_shader_engines = 1;
3117 rdev->config.si.max_tile_pipes = 4;
3118 rdev->config.si.max_cu_per_sh = 6;
3119 rdev->config.si.max_sh_per_se = 1;
3120 rdev->config.si.max_backends_per_se = 2;
3121 rdev->config.si.max_texture_channel_caches = 4;
3122 rdev->config.si.max_gprs = 256;
3123 rdev->config.si.max_gs_threads = 16;
3124 rdev->config.si.max_hw_contexts = 8;
3125
0a96d72b
AD
3126 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
3127 rdev->config.si.sc_prim_fifo_size_backend = 0x40;
3128 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
3129 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
1a8ca750 3130 gb_addr_config = VERDE_GB_ADDR_CONFIG_GOLDEN;
0a96d72b 3131 break;
8b02859d
AD
3132 case CHIP_HAINAN:
3133 rdev->config.si.max_shader_engines = 1;
3134 rdev->config.si.max_tile_pipes = 4;
3135 rdev->config.si.max_cu_per_sh = 5;
3136 rdev->config.si.max_sh_per_se = 1;
3137 rdev->config.si.max_backends_per_se = 1;
3138 rdev->config.si.max_texture_channel_caches = 2;
3139 rdev->config.si.max_gprs = 256;
3140 rdev->config.si.max_gs_threads = 16;
3141 rdev->config.si.max_hw_contexts = 8;
3142
3143 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
3144 rdev->config.si.sc_prim_fifo_size_backend = 0x40;
3145 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
3146 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
3147 gb_addr_config = HAINAN_GB_ADDR_CONFIG_GOLDEN;
3148 break;
0a96d72b
AD
3149 }
3150
3151 /* Initialize HDP */
3152 for (i = 0, j = 0; i < 32; i++, j += 0x18) {
3153 WREG32((0x2c14 + j), 0x00000000);
3154 WREG32((0x2c18 + j), 0x00000000);
3155 WREG32((0x2c1c + j), 0x00000000);
3156 WREG32((0x2c20 + j), 0x00000000);
3157 WREG32((0x2c24 + j), 0x00000000);
3158 }
3159
3160 WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff));
3161
3162 evergreen_fix_pci_max_read_req_size(rdev);
3163
3164 WREG32(BIF_FB_EN, FB_READ_EN | FB_WRITE_EN);
3165
3166 mc_shared_chmap = RREG32(MC_SHARED_CHMAP);
3167 mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG);
3168
0a96d72b 3169 rdev->config.si.num_tile_pipes = rdev->config.si.max_tile_pipes;
0a96d72b
AD
3170 rdev->config.si.mem_max_burst_length_bytes = 256;
3171 tmp = (mc_arb_ramcfg & NOOFCOLS_MASK) >> NOOFCOLS_SHIFT;
3172 rdev->config.si.mem_row_size_in_kb = (4 * (1 << (8 + tmp))) / 1024;
3173 if (rdev->config.si.mem_row_size_in_kb > 4)
3174 rdev->config.si.mem_row_size_in_kb = 4;
3175 /* XXX use MC settings? */
3176 rdev->config.si.shader_engine_tile_size = 32;
3177 rdev->config.si.num_gpus = 1;
3178 rdev->config.si.multi_gpu_tile_size = 64;
3179
1a8ca750
AD
3180 /* fix up row size */
3181 gb_addr_config &= ~ROW_SIZE_MASK;
0a96d72b
AD
3182 switch (rdev->config.si.mem_row_size_in_kb) {
3183 case 1:
3184 default:
3185 gb_addr_config |= ROW_SIZE(0);
3186 break;
3187 case 2:
3188 gb_addr_config |= ROW_SIZE(1);
3189 break;
3190 case 4:
3191 gb_addr_config |= ROW_SIZE(2);
3192 break;
3193 }
3194
0a96d72b
AD
3195 /* setup tiling info dword. gb_addr_config is not adequate since it does
3196 * not have bank info, so create a custom tiling dword.
3197 * bits 3:0 num_pipes
3198 * bits 7:4 num_banks
3199 * bits 11:8 group_size
3200 * bits 15:12 row_size
3201 */
3202 rdev->config.si.tile_config = 0;
3203 switch (rdev->config.si.num_tile_pipes) {
3204 case 1:
3205 rdev->config.si.tile_config |= (0 << 0);
3206 break;
3207 case 2:
3208 rdev->config.si.tile_config |= (1 << 0);
3209 break;
3210 case 4:
3211 rdev->config.si.tile_config |= (2 << 0);
3212 break;
3213 case 8:
3214 default:
3215 /* XXX what about 12? */
3216 rdev->config.si.tile_config |= (3 << 0);
3217 break;
dca571a6
CK
3218 }
3219 switch ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT) {
3220 case 0: /* four banks */
1a8ca750 3221 rdev->config.si.tile_config |= 0 << 4;
dca571a6
CK
3222 break;
3223 case 1: /* eight banks */
3224 rdev->config.si.tile_config |= 1 << 4;
3225 break;
3226 case 2: /* sixteen banks */
3227 default:
3228 rdev->config.si.tile_config |= 2 << 4;
3229 break;
3230 }
0a96d72b
AD
3231 rdev->config.si.tile_config |=
3232 ((gb_addr_config & PIPE_INTERLEAVE_SIZE_MASK) >> PIPE_INTERLEAVE_SIZE_SHIFT) << 8;
3233 rdev->config.si.tile_config |=
3234 ((gb_addr_config & ROW_SIZE_MASK) >> ROW_SIZE_SHIFT) << 12;
3235
0a96d72b
AD
3236 WREG32(GB_ADDR_CONFIG, gb_addr_config);
3237 WREG32(DMIF_ADDR_CONFIG, gb_addr_config);
7c1c7c18 3238 WREG32(DMIF_ADDR_CALC, gb_addr_config);
0a96d72b 3239 WREG32(HDP_ADDR_CONFIG, gb_addr_config);
8c5fd7ef
AD
3240 WREG32(DMA_TILING_CONFIG + DMA0_REGISTER_OFFSET, gb_addr_config);
3241 WREG32(DMA_TILING_CONFIG + DMA1_REGISTER_OFFSET, gb_addr_config);
1df0d523
AD
3242 if (rdev->has_uvd) {
3243 WREG32(UVD_UDEC_ADDR_CONFIG, gb_addr_config);
3244 WREG32(UVD_UDEC_DB_ADDR_CONFIG, gb_addr_config);
3245 WREG32(UVD_UDEC_DBW_ADDR_CONFIG, gb_addr_config);
3246 }
0a96d72b 3247
1a8ca750 3248 si_tiling_mode_table_init(rdev);
0a96d72b 3249
1a8ca750
AD
3250 si_setup_rb(rdev, rdev->config.si.max_shader_engines,
3251 rdev->config.si.max_sh_per_se,
3252 rdev->config.si.max_backends_per_se);
0a96d72b 3253
1a8ca750
AD
3254 si_setup_spi(rdev, rdev->config.si.max_shader_engines,
3255 rdev->config.si.max_sh_per_se,
3256 rdev->config.si.max_cu_per_sh);
0a96d72b 3257
52da51f0 3258 rdev->config.si.active_cus = 0;
65fcf668
AD
3259 for (i = 0; i < rdev->config.si.max_shader_engines; i++) {
3260 for (j = 0; j < rdev->config.si.max_sh_per_se; j++) {
6101b3ae
AD
3261 rdev->config.si.active_cus +=
3262 hweight32(si_get_cu_active_bitmap(rdev, i, j));
65fcf668
AD
3263 }
3264 }
0a96d72b
AD
3265
3266 /* set HW defaults for 3D engine */
3267 WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) |
3268 ROQ_IB2_START(0x2b)));
3269 WREG32(CP_MEQ_THRESHOLDS, MEQ1_START(0x30) | MEQ2_START(0x60));
3270
3271 sx_debug_1 = RREG32(SX_DEBUG_1);
3272 WREG32(SX_DEBUG_1, sx_debug_1);
3273
3274 WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4));
3275
3276 WREG32(PA_SC_FIFO_SIZE, (SC_FRONTEND_PRIM_FIFO_SIZE(rdev->config.si.sc_prim_fifo_size_frontend) |
3277 SC_BACKEND_PRIM_FIFO_SIZE(rdev->config.si.sc_prim_fifo_size_backend) |
3278 SC_HIZ_TILE_FIFO_SIZE(rdev->config.si.sc_hiz_tile_fifo_size) |
3279 SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.si.sc_earlyz_tile_fifo_size)));
3280
3281 WREG32(VGT_NUM_INSTANCES, 1);
3282
3283 WREG32(CP_PERFMON_CNTL, 0);
3284
3285 WREG32(SQ_CONFIG, 0);
3286
3287 WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) |
3288 FORCE_EOV_MAX_REZ_CNT(255)));
3289
3290 WREG32(VGT_CACHE_INVALIDATION, CACHE_INVALIDATION(VC_AND_TC) |
3291 AUTO_INVLD_EN(ES_AND_GS_AUTO));
3292
3293 WREG32(VGT_GS_VERTEX_REUSE, 16);
3294 WREG32(PA_SC_LINE_STIPPLE_STATE, 0);
3295
3296 WREG32(CB_PERFCOUNTER0_SELECT0, 0);
3297 WREG32(CB_PERFCOUNTER0_SELECT1, 0);
3298 WREG32(CB_PERFCOUNTER1_SELECT0, 0);
3299 WREG32(CB_PERFCOUNTER1_SELECT1, 0);
3300 WREG32(CB_PERFCOUNTER2_SELECT0, 0);
3301 WREG32(CB_PERFCOUNTER2_SELECT1, 0);
3302 WREG32(CB_PERFCOUNTER3_SELECT0, 0);
3303 WREG32(CB_PERFCOUNTER3_SELECT1, 0);
3304
3305 tmp = RREG32(HDP_MISC_CNTL);
3306 tmp |= HDP_FLUSH_INVALIDATE_CACHE;
3307 WREG32(HDP_MISC_CNTL, tmp);
3308
3309 hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL);
3310 WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
3311
3312 WREG32(PA_CL_ENHANCE, CLIP_VTX_REORDER_ENA | NUM_CLIP_SEQ(3));
3313
3314 udelay(50);
3315}
c476dde2 3316
2ece2e8b
AD
3317/*
3318 * GPU scratch registers helpers function.
3319 */
3320static void si_scratch_init(struct radeon_device *rdev)
3321{
3322 int i;
3323
3324 rdev->scratch.num_reg = 7;
3325 rdev->scratch.reg_base = SCRATCH_REG0;
3326 for (i = 0; i < rdev->scratch.num_reg; i++) {
3327 rdev->scratch.free[i] = true;
3328 rdev->scratch.reg[i] = rdev->scratch.reg_base + (i * 4);
3329 }
3330}
3331
3332void si_fence_ring_emit(struct radeon_device *rdev,
3333 struct radeon_fence *fence)
3334{
3335 struct radeon_ring *ring = &rdev->ring[fence->ring];
3336 u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
3337
3338 /* flush read cache over gart */
3339 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
3340 radeon_ring_write(ring, (CP_COHER_CNTL2 - PACKET3_SET_CONFIG_REG_START) >> 2);
3341 radeon_ring_write(ring, 0);
3342 radeon_ring_write(ring, PACKET3(PACKET3_SURFACE_SYNC, 3));
3343 radeon_ring_write(ring, PACKET3_TCL1_ACTION_ENA |
3344 PACKET3_TC_ACTION_ENA |
3345 PACKET3_SH_KCACHE_ACTION_ENA |
3346 PACKET3_SH_ICACHE_ACTION_ENA);
3347 radeon_ring_write(ring, 0xFFFFFFFF);
3348 radeon_ring_write(ring, 0);
3349 radeon_ring_write(ring, 10); /* poll interval */
3350 /* EVENT_WRITE_EOP - flush caches, send int */
3351 radeon_ring_write(ring, PACKET3(PACKET3_EVENT_WRITE_EOP, 4));
3352 radeon_ring_write(ring, EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) | EVENT_INDEX(5));
5e167cdb 3353 radeon_ring_write(ring, lower_32_bits(addr));
2ece2e8b
AD
3354 radeon_ring_write(ring, (upper_32_bits(addr) & 0xff) | DATA_SEL(1) | INT_SEL(2));
3355 radeon_ring_write(ring, fence->seq);
3356 radeon_ring_write(ring, 0);
3357}
3358
3359/*
3360 * IB stuff
3361 */
3362void si_ring_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib)
3363{
876dc9f3 3364 struct radeon_ring *ring = &rdev->ring[ib->ring];
2ece2e8b
AD
3365 u32 header;
3366
a85a7da4
AD
3367 if (ib->is_const_ib) {
3368 /* set switch buffer packet before const IB */
3369 radeon_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
3370 radeon_ring_write(ring, 0);
45df6803 3371
2ece2e8b 3372 header = PACKET3(PACKET3_INDIRECT_BUFFER_CONST, 2);
a85a7da4 3373 } else {
89d35807 3374 u32 next_rptr;
a85a7da4 3375 if (ring->rptr_save_reg) {
89d35807 3376 next_rptr = ring->wptr + 3 + 4 + 8;
a85a7da4
AD
3377 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
3378 radeon_ring_write(ring, ((ring->rptr_save_reg -
3379 PACKET3_SET_CONFIG_REG_START) >> 2));
3380 radeon_ring_write(ring, next_rptr);
89d35807
AD
3381 } else if (rdev->wb.enabled) {
3382 next_rptr = ring->wptr + 5 + 4 + 8;
3383 radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
3384 radeon_ring_write(ring, (1 << 8));
3385 radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
5e167cdb 3386 radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr));
89d35807 3387 radeon_ring_write(ring, next_rptr);
a85a7da4
AD
3388 }
3389
2ece2e8b 3390 header = PACKET3(PACKET3_INDIRECT_BUFFER, 2);
a85a7da4 3391 }
2ece2e8b
AD
3392
3393 radeon_ring_write(ring, header);
3394 radeon_ring_write(ring,
3395#ifdef __BIG_ENDIAN
3396 (2 << 0) |
3397#endif
3398 (ib->gpu_addr & 0xFFFFFFFC));
3399 radeon_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xFFFF);
4bf3dd92
CK
3400 radeon_ring_write(ring, ib->length_dw |
3401 (ib->vm ? (ib->vm->id << 24) : 0));
2ece2e8b 3402
a85a7da4
AD
3403 if (!ib->is_const_ib) {
3404 /* flush read cache over gart for this vmid */
3405 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
3406 radeon_ring_write(ring, (CP_COHER_CNTL2 - PACKET3_SET_CONFIG_REG_START) >> 2);
4bf3dd92 3407 radeon_ring_write(ring, ib->vm ? ib->vm->id : 0);
a85a7da4
AD
3408 radeon_ring_write(ring, PACKET3(PACKET3_SURFACE_SYNC, 3));
3409 radeon_ring_write(ring, PACKET3_TCL1_ACTION_ENA |
3410 PACKET3_TC_ACTION_ENA |
3411 PACKET3_SH_KCACHE_ACTION_ENA |
3412 PACKET3_SH_ICACHE_ACTION_ENA);
3413 radeon_ring_write(ring, 0xFFFFFFFF);
3414 radeon_ring_write(ring, 0);
3415 radeon_ring_write(ring, 10); /* poll interval */
3416 }
2ece2e8b
AD
3417}
3418
48c0c902
AD
3419/*
3420 * CP.
3421 */
3422static void si_cp_enable(struct radeon_device *rdev, bool enable)
3423{
3424 if (enable)
3425 WREG32(CP_ME_CNTL, 0);
3426 else {
50efa51a
AD
3427 if (rdev->asic->copy.copy_ring_index == RADEON_RING_TYPE_GFX_INDEX)
3428 radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
48c0c902
AD
3429 WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT | CP_CE_HALT));
3430 WREG32(SCRATCH_UMSK, 0);
8c5fd7ef
AD
3431 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
3432 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
3433 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
48c0c902
AD
3434 }
3435 udelay(50);
3436}
3437
3438static int si_cp_load_microcode(struct radeon_device *rdev)
3439{
48c0c902
AD
3440 int i;
3441
629bd33c 3442 if (!rdev->me_fw || !rdev->pfp_fw || !rdev->ce_fw)
48c0c902
AD
3443 return -EINVAL;
3444
3445 si_cp_enable(rdev, false);
3446
629bd33c
AD
3447 if (rdev->new_fw) {
3448 const struct gfx_firmware_header_v1_0 *pfp_hdr =
3449 (const struct gfx_firmware_header_v1_0 *)rdev->pfp_fw->data;
3450 const struct gfx_firmware_header_v1_0 *ce_hdr =
3451 (const struct gfx_firmware_header_v1_0 *)rdev->ce_fw->data;
3452 const struct gfx_firmware_header_v1_0 *me_hdr =
3453 (const struct gfx_firmware_header_v1_0 *)rdev->me_fw->data;
3454 const __le32 *fw_data;
3455 u32 fw_size;
3456
3457 radeon_ucode_print_gfx_hdr(&pfp_hdr->header);
3458 radeon_ucode_print_gfx_hdr(&ce_hdr->header);
3459 radeon_ucode_print_gfx_hdr(&me_hdr->header);
3460
3461 /* PFP */
3462 fw_data = (const __le32 *)
3463 (rdev->pfp_fw->data + le32_to_cpu(pfp_hdr->header.ucode_array_offset_bytes));
3464 fw_size = le32_to_cpu(pfp_hdr->header.ucode_size_bytes) / 4;
3465 WREG32(CP_PFP_UCODE_ADDR, 0);
3466 for (i = 0; i < fw_size; i++)
3467 WREG32(CP_PFP_UCODE_DATA, le32_to_cpup(fw_data++));
3468 WREG32(CP_PFP_UCODE_ADDR, 0);
3469
3470 /* CE */
3471 fw_data = (const __le32 *)
3472 (rdev->ce_fw->data + le32_to_cpu(ce_hdr->header.ucode_array_offset_bytes));
3473 fw_size = le32_to_cpu(ce_hdr->header.ucode_size_bytes) / 4;
3474 WREG32(CP_CE_UCODE_ADDR, 0);
3475 for (i = 0; i < fw_size; i++)
3476 WREG32(CP_CE_UCODE_DATA, le32_to_cpup(fw_data++));
3477 WREG32(CP_CE_UCODE_ADDR, 0);
3478
3479 /* ME */
3480 fw_data = (const __be32 *)
3481 (rdev->me_fw->data + le32_to_cpu(me_hdr->header.ucode_array_offset_bytes));
3482 fw_size = le32_to_cpu(me_hdr->header.ucode_size_bytes) / 4;
3483 WREG32(CP_ME_RAM_WADDR, 0);
3484 for (i = 0; i < fw_size; i++)
3485 WREG32(CP_ME_RAM_DATA, le32_to_cpup(fw_data++));
3486 WREG32(CP_ME_RAM_WADDR, 0);
3487 } else {
3488 const __be32 *fw_data;
3489
3490 /* PFP */
3491 fw_data = (const __be32 *)rdev->pfp_fw->data;
3492 WREG32(CP_PFP_UCODE_ADDR, 0);
3493 for (i = 0; i < SI_PFP_UCODE_SIZE; i++)
3494 WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++));
3495 WREG32(CP_PFP_UCODE_ADDR, 0);
3496
3497 /* CE */
3498 fw_data = (const __be32 *)rdev->ce_fw->data;
3499 WREG32(CP_CE_UCODE_ADDR, 0);
3500 for (i = 0; i < SI_CE_UCODE_SIZE; i++)
3501 WREG32(CP_CE_UCODE_DATA, be32_to_cpup(fw_data++));
3502 WREG32(CP_CE_UCODE_ADDR, 0);
3503
3504 /* ME */
3505 fw_data = (const __be32 *)rdev->me_fw->data;
3506 WREG32(CP_ME_RAM_WADDR, 0);
3507 for (i = 0; i < SI_PM4_UCODE_SIZE; i++)
3508 WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++));
3509 WREG32(CP_ME_RAM_WADDR, 0);
3510 }
48c0c902
AD
3511
3512 WREG32(CP_PFP_UCODE_ADDR, 0);
3513 WREG32(CP_CE_UCODE_ADDR, 0);
3514 WREG32(CP_ME_RAM_WADDR, 0);
3515 WREG32(CP_ME_RAM_RADDR, 0);
3516 return 0;
3517}
3518
3519static int si_cp_start(struct radeon_device *rdev)
3520{
3521 struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
3522 int r, i;
3523
3524 r = radeon_ring_lock(rdev, ring, 7 + 4);
3525 if (r) {
3526 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
3527 return r;
3528 }
3529 /* init the CP */
3530 radeon_ring_write(ring, PACKET3(PACKET3_ME_INITIALIZE, 5));
3531 radeon_ring_write(ring, 0x1);
3532 radeon_ring_write(ring, 0x0);
3533 radeon_ring_write(ring, rdev->config.si.max_hw_contexts - 1);
3534 radeon_ring_write(ring, PACKET3_ME_INITIALIZE_DEVICE_ID(1));
3535 radeon_ring_write(ring, 0);
3536 radeon_ring_write(ring, 0);
3537
3538 /* init the CE partitions */
3539 radeon_ring_write(ring, PACKET3(PACKET3_SET_BASE, 2));
3540 radeon_ring_write(ring, PACKET3_BASE_INDEX(CE_PARTITION_BASE));
3541 radeon_ring_write(ring, 0xc000);
3542 radeon_ring_write(ring, 0xe000);
1538a9e0 3543 radeon_ring_unlock_commit(rdev, ring, false);
48c0c902
AD
3544
3545 si_cp_enable(rdev, true);
3546
3547 r = radeon_ring_lock(rdev, ring, si_default_size + 10);
3548 if (r) {
3549 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
3550 return r;
3551 }
3552
3553 /* setup clear context state */
3554 radeon_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
3555 radeon_ring_write(ring, PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
3556
3557 for (i = 0; i < si_default_size; i++)
3558 radeon_ring_write(ring, si_default_state[i]);
3559
3560 radeon_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
3561 radeon_ring_write(ring, PACKET3_PREAMBLE_END_CLEAR_STATE);
3562
3563 /* set clear context state */
3564 radeon_ring_write(ring, PACKET3(PACKET3_CLEAR_STATE, 0));
3565 radeon_ring_write(ring, 0);
3566
3567 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONTEXT_REG, 2));
3568 radeon_ring_write(ring, 0x00000316);
3569 radeon_ring_write(ring, 0x0000000e); /* VGT_VERTEX_REUSE_BLOCK_CNTL */
3570 radeon_ring_write(ring, 0x00000010); /* VGT_OUT_DEALLOC_CNTL */
3571
1538a9e0 3572 radeon_ring_unlock_commit(rdev, ring, false);
48c0c902
AD
3573
3574 for (i = RADEON_RING_TYPE_GFX_INDEX; i <= CAYMAN_RING_TYPE_CP2_INDEX; ++i) {
3575 ring = &rdev->ring[i];
3576 r = radeon_ring_lock(rdev, ring, 2);
3577
3578 /* clear the compute context state */
3579 radeon_ring_write(ring, PACKET3_COMPUTE(PACKET3_CLEAR_STATE, 0));
3580 radeon_ring_write(ring, 0);
3581
1538a9e0 3582 radeon_ring_unlock_commit(rdev, ring, false);
48c0c902
AD
3583 }
3584
3585 return 0;
3586}
3587
3588static void si_cp_fini(struct radeon_device *rdev)
3589{
45df6803 3590 struct radeon_ring *ring;
48c0c902 3591 si_cp_enable(rdev, false);
45df6803
CK
3592
3593 ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
3594 radeon_ring_fini(rdev, ring);
3595 radeon_scratch_free(rdev, ring->rptr_save_reg);
3596
3597 ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
3598 radeon_ring_fini(rdev, ring);
3599 radeon_scratch_free(rdev, ring->rptr_save_reg);
3600
3601 ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
3602 radeon_ring_fini(rdev, ring);
3603 radeon_scratch_free(rdev, ring->rptr_save_reg);
48c0c902
AD
3604}
3605
3606static int si_cp_resume(struct radeon_device *rdev)
3607{
3608 struct radeon_ring *ring;
3609 u32 tmp;
3610 u32 rb_bufsz;
3611 int r;
3612
811e4d58
AD
3613 si_enable_gui_idle_interrupt(rdev, false);
3614
48c0c902
AD
3615 WREG32(CP_SEM_WAIT_TIMER, 0x0);
3616 WREG32(CP_SEM_INCOMPLETE_TIMER_CNTL, 0x0);
3617
3618 /* Set the write pointer delay */
3619 WREG32(CP_RB_WPTR_DELAY, 0);
3620
3621 WREG32(CP_DEBUG, 0);
3622 WREG32(SCRATCH_ADDR, ((rdev->wb.gpu_addr + RADEON_WB_SCRATCH_OFFSET) >> 8) & 0xFFFFFFFF);
3623
3624 /* ring 0 - compute and gfx */
3625 /* Set ring buffer size */
3626 ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
b72a8925
DV
3627 rb_bufsz = order_base_2(ring->ring_size / 8);
3628 tmp = (order_base_2(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
48c0c902
AD
3629#ifdef __BIG_ENDIAN
3630 tmp |= BUF_SWAP_32BIT;
3631#endif
3632 WREG32(CP_RB0_CNTL, tmp);
3633
3634 /* Initialize the ring buffer's read and write pointers */
3635 WREG32(CP_RB0_CNTL, tmp | RB_RPTR_WR_ENA);
3636 ring->wptr = 0;
3637 WREG32(CP_RB0_WPTR, ring->wptr);
3638
48fc7f7e 3639 /* set the wb address whether it's enabled or not */
48c0c902
AD
3640 WREG32(CP_RB0_RPTR_ADDR, (rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET) & 0xFFFFFFFC);
3641 WREG32(CP_RB0_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET) & 0xFF);
3642
3643 if (rdev->wb.enabled)
3644 WREG32(SCRATCH_UMSK, 0xff);
3645 else {
3646 tmp |= RB_NO_UPDATE;
3647 WREG32(SCRATCH_UMSK, 0);
3648 }
3649
3650 mdelay(1);
3651 WREG32(CP_RB0_CNTL, tmp);
3652
3653 WREG32(CP_RB0_BASE, ring->gpu_addr >> 8);
3654
48c0c902
AD
3655 /* ring1 - compute only */
3656 /* Set ring buffer size */
3657 ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
b72a8925
DV
3658 rb_bufsz = order_base_2(ring->ring_size / 8);
3659 tmp = (order_base_2(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
48c0c902
AD
3660#ifdef __BIG_ENDIAN
3661 tmp |= BUF_SWAP_32BIT;
3662#endif
3663 WREG32(CP_RB1_CNTL, tmp);
3664
3665 /* Initialize the ring buffer's read and write pointers */
3666 WREG32(CP_RB1_CNTL, tmp | RB_RPTR_WR_ENA);
3667 ring->wptr = 0;
3668 WREG32(CP_RB1_WPTR, ring->wptr);
3669
48fc7f7e 3670 /* set the wb address whether it's enabled or not */
48c0c902
AD
3671 WREG32(CP_RB1_RPTR_ADDR, (rdev->wb.gpu_addr + RADEON_WB_CP1_RPTR_OFFSET) & 0xFFFFFFFC);
3672 WREG32(CP_RB1_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP1_RPTR_OFFSET) & 0xFF);
3673
3674 mdelay(1);
3675 WREG32(CP_RB1_CNTL, tmp);
3676
3677 WREG32(CP_RB1_BASE, ring->gpu_addr >> 8);
3678
48c0c902
AD
3679 /* ring2 - compute only */
3680 /* Set ring buffer size */
3681 ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
b72a8925
DV
3682 rb_bufsz = order_base_2(ring->ring_size / 8);
3683 tmp = (order_base_2(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
48c0c902
AD
3684#ifdef __BIG_ENDIAN
3685 tmp |= BUF_SWAP_32BIT;
3686#endif
3687 WREG32(CP_RB2_CNTL, tmp);
3688
3689 /* Initialize the ring buffer's read and write pointers */
3690 WREG32(CP_RB2_CNTL, tmp | RB_RPTR_WR_ENA);
3691 ring->wptr = 0;
3692 WREG32(CP_RB2_WPTR, ring->wptr);
3693
48fc7f7e 3694 /* set the wb address whether it's enabled or not */
48c0c902
AD
3695 WREG32(CP_RB2_RPTR_ADDR, (rdev->wb.gpu_addr + RADEON_WB_CP2_RPTR_OFFSET) & 0xFFFFFFFC);
3696 WREG32(CP_RB2_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP2_RPTR_OFFSET) & 0xFF);
3697
3698 mdelay(1);
3699 WREG32(CP_RB2_CNTL, tmp);
3700
3701 WREG32(CP_RB2_BASE, ring->gpu_addr >> 8);
3702
48c0c902
AD
3703 /* start the rings */
3704 si_cp_start(rdev);
3705 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = true;
3706 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = true;
3707 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = true;
3708 r = radeon_ring_test(rdev, RADEON_RING_TYPE_GFX_INDEX, &rdev->ring[RADEON_RING_TYPE_GFX_INDEX]);
3709 if (r) {
3710 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
3711 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
3712 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
3713 return r;
3714 }
3715 r = radeon_ring_test(rdev, CAYMAN_RING_TYPE_CP1_INDEX, &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX]);
3716 if (r) {
3717 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
3718 }
3719 r = radeon_ring_test(rdev, CAYMAN_RING_TYPE_CP2_INDEX, &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX]);
3720 if (r) {
3721 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
3722 }
3723
811e4d58
AD
3724 si_enable_gui_idle_interrupt(rdev, true);
3725
50efa51a
AD
3726 if (rdev->asic->copy.copy_ring_index == RADEON_RING_TYPE_GFX_INDEX)
3727 radeon_ttm_set_active_vram_size(rdev, rdev->mc.real_vram_size);
3728
48c0c902
AD
3729 return 0;
3730}
3731
2483b4ea 3732u32 si_gpu_check_soft_reset(struct radeon_device *rdev)
06bc6df0 3733{
014bb209 3734 u32 reset_mask = 0;
1c534671 3735 u32 tmp;
06bc6df0 3736
014bb209
AD
3737 /* GRBM_STATUS */
3738 tmp = RREG32(GRBM_STATUS);
3739 if (tmp & (PA_BUSY | SC_BUSY |
3740 BCI_BUSY | SX_BUSY |
3741 TA_BUSY | VGT_BUSY |
3742 DB_BUSY | CB_BUSY |
3743 GDS_BUSY | SPI_BUSY |
3744 IA_BUSY | IA_BUSY_NO_DMA))
3745 reset_mask |= RADEON_RESET_GFX;
3746
3747 if (tmp & (CF_RQ_PENDING | PF_RQ_PENDING |
3748 CP_BUSY | CP_COHERENCY_BUSY))
3749 reset_mask |= RADEON_RESET_CP;
3750
3751 if (tmp & GRBM_EE_BUSY)
3752 reset_mask |= RADEON_RESET_GRBM | RADEON_RESET_GFX | RADEON_RESET_CP;
3753
3754 /* GRBM_STATUS2 */
3755 tmp = RREG32(GRBM_STATUS2);
3756 if (tmp & (RLC_RQ_PENDING | RLC_BUSY))
3757 reset_mask |= RADEON_RESET_RLC;
3758
3759 /* DMA_STATUS_REG 0 */
3760 tmp = RREG32(DMA_STATUS_REG + DMA0_REGISTER_OFFSET);
3761 if (!(tmp & DMA_IDLE))
3762 reset_mask |= RADEON_RESET_DMA;
3763
3764 /* DMA_STATUS_REG 1 */
3765 tmp = RREG32(DMA_STATUS_REG + DMA1_REGISTER_OFFSET);
3766 if (!(tmp & DMA_IDLE))
3767 reset_mask |= RADEON_RESET_DMA1;
3768
3769 /* SRBM_STATUS2 */
3770 tmp = RREG32(SRBM_STATUS2);
3771 if (tmp & DMA_BUSY)
3772 reset_mask |= RADEON_RESET_DMA;
3773
3774 if (tmp & DMA1_BUSY)
3775 reset_mask |= RADEON_RESET_DMA1;
3776
3777 /* SRBM_STATUS */
3778 tmp = RREG32(SRBM_STATUS);
3779
3780 if (tmp & IH_BUSY)
3781 reset_mask |= RADEON_RESET_IH;
3782
3783 if (tmp & SEM_BUSY)
3784 reset_mask |= RADEON_RESET_SEM;
3785
3786 if (tmp & GRBM_RQ_PENDING)
3787 reset_mask |= RADEON_RESET_GRBM;
3788
3789 if (tmp & VMC_BUSY)
3790 reset_mask |= RADEON_RESET_VMC;
19fc42ed 3791
014bb209
AD
3792 if (tmp & (MCB_BUSY | MCB_NON_DISPLAY_BUSY |
3793 MCC_BUSY | MCD_BUSY))
3794 reset_mask |= RADEON_RESET_MC;
3795
3796 if (evergreen_is_display_hung(rdev))
3797 reset_mask |= RADEON_RESET_DISPLAY;
3798
3799 /* VM_L2_STATUS */
3800 tmp = RREG32(VM_L2_STATUS);
3801 if (tmp & L2_BUSY)
3802 reset_mask |= RADEON_RESET_VMC;
3803
d808fc88
AD
3804 /* Skip MC reset as it's mostly likely not hung, just busy */
3805 if (reset_mask & RADEON_RESET_MC) {
3806 DRM_DEBUG("MC busy: 0x%08X, clearing.\n", reset_mask);
3807 reset_mask &= ~RADEON_RESET_MC;
3808 }
3809
014bb209
AD
3810 return reset_mask;
3811}
3812
3813static void si_gpu_soft_reset(struct radeon_device *rdev, u32 reset_mask)
3814{
3815 struct evergreen_mc_save save;
3816 u32 grbm_soft_reset = 0, srbm_soft_reset = 0;
3817 u32 tmp;
19fc42ed 3818
06bc6df0 3819 if (reset_mask == 0)
014bb209 3820 return;
06bc6df0
AD
3821
3822 dev_info(rdev->dev, "GPU softreset: 0x%08X\n", reset_mask);
3823
1c534671 3824 evergreen_print_gpu_status_regs(rdev);
06bc6df0
AD
3825 dev_info(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n",
3826 RREG32(VM_CONTEXT1_PROTECTION_FAULT_ADDR));
3827 dev_info(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
3828 RREG32(VM_CONTEXT1_PROTECTION_FAULT_STATUS));
3829
a6f4ae8d
AD
3830 /* disable PG/CG */
3831 si_fini_pg(rdev);
3832 si_fini_cg(rdev);
3833
3834 /* stop the rlc */
3835 si_rlc_stop(rdev);
3836
1c534671
AD
3837 /* Disable CP parsing/prefetching */
3838 WREG32(CP_ME_CNTL, CP_ME_HALT | CP_PFP_HALT | CP_CE_HALT);
3839
3840 if (reset_mask & RADEON_RESET_DMA) {
3841 /* dma0 */
3842 tmp = RREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET);
3843 tmp &= ~DMA_RB_ENABLE;
3844 WREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET, tmp);
014bb209
AD
3845 }
3846 if (reset_mask & RADEON_RESET_DMA1) {
1c534671
AD
3847 /* dma1 */
3848 tmp = RREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET);
3849 tmp &= ~DMA_RB_ENABLE;
3850 WREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET, tmp);
3851 }
3852
f770d78a
AD
3853 udelay(50);
3854
3855 evergreen_mc_stop(rdev, &save);
3856 if (evergreen_mc_wait_for_idle(rdev)) {
3857 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
3858 }
3859
1c534671
AD
3860 if (reset_mask & (RADEON_RESET_GFX | RADEON_RESET_COMPUTE | RADEON_RESET_CP)) {
3861 grbm_soft_reset = SOFT_RESET_CB |
3862 SOFT_RESET_DB |
3863 SOFT_RESET_GDS |
3864 SOFT_RESET_PA |
3865 SOFT_RESET_SC |
3866 SOFT_RESET_BCI |
3867 SOFT_RESET_SPI |
3868 SOFT_RESET_SX |
3869 SOFT_RESET_TC |
3870 SOFT_RESET_TA |
3871 SOFT_RESET_VGT |
3872 SOFT_RESET_IA;
3873 }
3874
3875 if (reset_mask & RADEON_RESET_CP) {
3876 grbm_soft_reset |= SOFT_RESET_CP | SOFT_RESET_VGT;
3877
3878 srbm_soft_reset |= SOFT_RESET_GRBM;
3879 }
06bc6df0
AD
3880
3881 if (reset_mask & RADEON_RESET_DMA)
014bb209
AD
3882 srbm_soft_reset |= SOFT_RESET_DMA;
3883
3884 if (reset_mask & RADEON_RESET_DMA1)
3885 srbm_soft_reset |= SOFT_RESET_DMA1;
3886
3887 if (reset_mask & RADEON_RESET_DISPLAY)
3888 srbm_soft_reset |= SOFT_RESET_DC;
3889
3890 if (reset_mask & RADEON_RESET_RLC)
3891 grbm_soft_reset |= SOFT_RESET_RLC;
3892
3893 if (reset_mask & RADEON_RESET_SEM)
3894 srbm_soft_reset |= SOFT_RESET_SEM;
3895
3896 if (reset_mask & RADEON_RESET_IH)
3897 srbm_soft_reset |= SOFT_RESET_IH;
3898
3899 if (reset_mask & RADEON_RESET_GRBM)
3900 srbm_soft_reset |= SOFT_RESET_GRBM;
3901
3902 if (reset_mask & RADEON_RESET_VMC)
3903 srbm_soft_reset |= SOFT_RESET_VMC;
3904
3905 if (reset_mask & RADEON_RESET_MC)
3906 srbm_soft_reset |= SOFT_RESET_MC;
1c534671
AD
3907
3908 if (grbm_soft_reset) {
3909 tmp = RREG32(GRBM_SOFT_RESET);
3910 tmp |= grbm_soft_reset;
3911 dev_info(rdev->dev, "GRBM_SOFT_RESET=0x%08X\n", tmp);
3912 WREG32(GRBM_SOFT_RESET, tmp);
3913 tmp = RREG32(GRBM_SOFT_RESET);
3914
3915 udelay(50);
3916
3917 tmp &= ~grbm_soft_reset;
3918 WREG32(GRBM_SOFT_RESET, tmp);
3919 tmp = RREG32(GRBM_SOFT_RESET);
3920 }
3921
3922 if (srbm_soft_reset) {
3923 tmp = RREG32(SRBM_SOFT_RESET);
3924 tmp |= srbm_soft_reset;
3925 dev_info(rdev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
3926 WREG32(SRBM_SOFT_RESET, tmp);
3927 tmp = RREG32(SRBM_SOFT_RESET);
3928
3929 udelay(50);
3930
3931 tmp &= ~srbm_soft_reset;
3932 WREG32(SRBM_SOFT_RESET, tmp);
3933 tmp = RREG32(SRBM_SOFT_RESET);
3934 }
06bc6df0
AD
3935
3936 /* Wait a little for things to settle down */
3937 udelay(50);
3938
c476dde2 3939 evergreen_mc_resume(rdev, &save);
1c534671
AD
3940 udelay(50);
3941
1c534671 3942 evergreen_print_gpu_status_regs(rdev);
c476dde2
AD
3943}
3944
4a5c8ea5
AD
3945static void si_set_clk_bypass_mode(struct radeon_device *rdev)
3946{
3947 u32 tmp, i;
3948
3949 tmp = RREG32(CG_SPLL_FUNC_CNTL);
3950 tmp |= SPLL_BYPASS_EN;
3951 WREG32(CG_SPLL_FUNC_CNTL, tmp);
3952
3953 tmp = RREG32(CG_SPLL_FUNC_CNTL_2);
3954 tmp |= SPLL_CTLREQ_CHG;
3955 WREG32(CG_SPLL_FUNC_CNTL_2, tmp);
3956
3957 for (i = 0; i < rdev->usec_timeout; i++) {
3958 if (RREG32(SPLL_STATUS) & SPLL_CHG_STATUS)
3959 break;
3960 udelay(1);
3961 }
3962
3963 tmp = RREG32(CG_SPLL_FUNC_CNTL_2);
3964 tmp &= ~(SPLL_CTLREQ_CHG | SCLK_MUX_UPDATE);
3965 WREG32(CG_SPLL_FUNC_CNTL_2, tmp);
3966
3967 tmp = RREG32(MPLL_CNTL_MODE);
3968 tmp &= ~MPLL_MCLK_SEL;
3969 WREG32(MPLL_CNTL_MODE, tmp);
3970}
3971
3972static void si_spll_powerdown(struct radeon_device *rdev)
3973{
3974 u32 tmp;
3975
3976 tmp = RREG32(SPLL_CNTL_MODE);
3977 tmp |= SPLL_SW_DIR_CONTROL;
3978 WREG32(SPLL_CNTL_MODE, tmp);
3979
3980 tmp = RREG32(CG_SPLL_FUNC_CNTL);
3981 tmp |= SPLL_RESET;
3982 WREG32(CG_SPLL_FUNC_CNTL, tmp);
3983
3984 tmp = RREG32(CG_SPLL_FUNC_CNTL);
3985 tmp |= SPLL_SLEEP;
3986 WREG32(CG_SPLL_FUNC_CNTL, tmp);
3987
3988 tmp = RREG32(SPLL_CNTL_MODE);
3989 tmp &= ~SPLL_SW_DIR_CONTROL;
3990 WREG32(SPLL_CNTL_MODE, tmp);
3991}
3992
3993static void si_gpu_pci_config_reset(struct radeon_device *rdev)
3994{
3995 struct evergreen_mc_save save;
3996 u32 tmp, i;
3997
3998 dev_info(rdev->dev, "GPU pci config reset\n");
3999
4000 /* disable dpm? */
4001
4002 /* disable cg/pg */
4003 si_fini_pg(rdev);
4004 si_fini_cg(rdev);
4005
4006 /* Disable CP parsing/prefetching */
4007 WREG32(CP_ME_CNTL, CP_ME_HALT | CP_PFP_HALT | CP_CE_HALT);
4008 /* dma0 */
4009 tmp = RREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET);
4010 tmp &= ~DMA_RB_ENABLE;
4011 WREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET, tmp);
4012 /* dma1 */
4013 tmp = RREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET);
4014 tmp &= ~DMA_RB_ENABLE;
4015 WREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET, tmp);
4016 /* XXX other engines? */
4017
4018 /* halt the rlc, disable cp internal ints */
4019 si_rlc_stop(rdev);
4020
4021 udelay(50);
4022
4023 /* disable mem access */
4024 evergreen_mc_stop(rdev, &save);
4025 if (evergreen_mc_wait_for_idle(rdev)) {
4026 dev_warn(rdev->dev, "Wait for MC idle timed out !\n");
4027 }
4028
4029 /* set mclk/sclk to bypass */
4030 si_set_clk_bypass_mode(rdev);
4031 /* powerdown spll */
4032 si_spll_powerdown(rdev);
4033 /* disable BM */
4034 pci_clear_master(rdev->pdev);
4035 /* reset */
4036 radeon_pci_config_reset(rdev);
4037 /* wait for asic to come out of reset */
4038 for (i = 0; i < rdev->usec_timeout; i++) {
4039 if (RREG32(CONFIG_MEMSIZE) != 0xffffffff)
4040 break;
4041 udelay(1);
4042 }
4043}
4044
c476dde2
AD
4045int si_asic_reset(struct radeon_device *rdev)
4046{
014bb209
AD
4047 u32 reset_mask;
4048
4049 reset_mask = si_gpu_check_soft_reset(rdev);
4050
4051 if (reset_mask)
4052 r600_set_bios_scratch_engine_hung(rdev, true);
4053
4a5c8ea5 4054 /* try soft reset */
014bb209
AD
4055 si_gpu_soft_reset(rdev, reset_mask);
4056
4057 reset_mask = si_gpu_check_soft_reset(rdev);
4058
4a5c8ea5
AD
4059 /* try pci config reset */
4060 if (reset_mask && radeon_hard_reset)
4061 si_gpu_pci_config_reset(rdev);
4062
4063 reset_mask = si_gpu_check_soft_reset(rdev);
4064
014bb209
AD
4065 if (!reset_mask)
4066 r600_set_bios_scratch_engine_hung(rdev, false);
4067
4068 return 0;
c476dde2
AD
4069}
4070
123bc183
AD
4071/**
4072 * si_gfx_is_lockup - Check if the GFX engine is locked up
4073 *
4074 * @rdev: radeon_device pointer
4075 * @ring: radeon_ring structure holding ring information
4076 *
4077 * Check if the GFX engine is locked up.
4078 * Returns true if the engine appears to be locked up, false if not.
4079 */
4080bool si_gfx_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
4081{
4082 u32 reset_mask = si_gpu_check_soft_reset(rdev);
4083
4084 if (!(reset_mask & (RADEON_RESET_GFX |
4085 RADEON_RESET_COMPUTE |
4086 RADEON_RESET_CP))) {
ff212f25 4087 radeon_ring_lockup_update(rdev, ring);
123bc183
AD
4088 return false;
4089 }
123bc183
AD
4090 return radeon_ring_test_lockup(rdev, ring);
4091}
4092
d2800ee5
AD
4093/* MC */
4094static void si_mc_program(struct radeon_device *rdev)
4095{
4096 struct evergreen_mc_save save;
4097 u32 tmp;
4098 int i, j;
4099
4100 /* Initialize HDP */
4101 for (i = 0, j = 0; i < 32; i++, j += 0x18) {
4102 WREG32((0x2c14 + j), 0x00000000);
4103 WREG32((0x2c18 + j), 0x00000000);
4104 WREG32((0x2c1c + j), 0x00000000);
4105 WREG32((0x2c20 + j), 0x00000000);
4106 WREG32((0x2c24 + j), 0x00000000);
4107 }
4108 WREG32(HDP_REG_COHERENCY_FLUSH_CNTL, 0);
4109
4110 evergreen_mc_stop(rdev, &save);
4111 if (radeon_mc_wait_for_idle(rdev)) {
4112 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
4113 }
5153550a
AD
4114 if (!ASIC_IS_NODCE(rdev))
4115 /* Lockout access through VGA aperture*/
4116 WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE);
d2800ee5
AD
4117 /* Update configuration */
4118 WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
4119 rdev->mc.vram_start >> 12);
4120 WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
4121 rdev->mc.vram_end >> 12);
4122 WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR,
4123 rdev->vram_scratch.gpu_addr >> 12);
4124 tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16;
4125 tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF);
4126 WREG32(MC_VM_FB_LOCATION, tmp);
4127 /* XXX double check these! */
4128 WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8));
4129 WREG32(HDP_NONSURFACE_INFO, (2 << 7) | (1 << 30));
4130 WREG32(HDP_NONSURFACE_SIZE, 0x3FFFFFFF);
4131 WREG32(MC_VM_AGP_BASE, 0);
4132 WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF);
4133 WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF);
4134 if (radeon_mc_wait_for_idle(rdev)) {
4135 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
4136 }
4137 evergreen_mc_resume(rdev, &save);
5153550a
AD
4138 if (!ASIC_IS_NODCE(rdev)) {
4139 /* we need to own VRAM, so turn off the VGA renderer here
4140 * to stop it overwriting our objects */
4141 rv515_vga_render_disable(rdev);
4142 }
d2800ee5
AD
4143}
4144
1c49165d
AD
4145void si_vram_gtt_location(struct radeon_device *rdev,
4146 struct radeon_mc *mc)
d2800ee5
AD
4147{
4148 if (mc->mc_vram_size > 0xFFC0000000ULL) {
4149 /* leave room for at least 1024M GTT */
4150 dev_warn(rdev->dev, "limiting VRAM\n");
4151 mc->real_vram_size = 0xFFC0000000ULL;
4152 mc->mc_vram_size = 0xFFC0000000ULL;
4153 }
9ed8b1f9 4154 radeon_vram_location(rdev, &rdev->mc, 0);
d2800ee5 4155 rdev->mc.gtt_base_align = 0;
9ed8b1f9 4156 radeon_gtt_location(rdev, mc);
d2800ee5
AD
4157}
4158
4159static int si_mc_init(struct radeon_device *rdev)
4160{
4161 u32 tmp;
4162 int chansize, numchan;
4163
4164 /* Get VRAM informations */
4165 rdev->mc.vram_is_ddr = true;
4166 tmp = RREG32(MC_ARB_RAMCFG);
4167 if (tmp & CHANSIZE_OVERRIDE) {
4168 chansize = 16;
4169 } else if (tmp & CHANSIZE_MASK) {
4170 chansize = 64;
4171 } else {
4172 chansize = 32;
4173 }
4174 tmp = RREG32(MC_SHARED_CHMAP);
4175 switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
4176 case 0:
4177 default:
4178 numchan = 1;
4179 break;
4180 case 1:
4181 numchan = 2;
4182 break;
4183 case 2:
4184 numchan = 4;
4185 break;
4186 case 3:
4187 numchan = 8;
4188 break;
4189 case 4:
4190 numchan = 3;
4191 break;
4192 case 5:
4193 numchan = 6;
4194 break;
4195 case 6:
4196 numchan = 10;
4197 break;
4198 case 7:
4199 numchan = 12;
4200 break;
4201 case 8:
4202 numchan = 16;
4203 break;
4204 }
4205 rdev->mc.vram_width = numchan * chansize;
4206 /* Could aper size report 0 ? */
4207 rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
4208 rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
4209 /* size in MB on si */
0ca223b0
AD
4210 tmp = RREG32(CONFIG_MEMSIZE);
4211 /* some boards may have garbage in the upper 16 bits */
4212 if (tmp & 0xffff0000) {
4213 DRM_INFO("Probable bad vram size: 0x%08x\n", tmp);
4214 if (tmp & 0xffff)
4215 tmp &= 0xffff;
4216 }
4217 rdev->mc.mc_vram_size = tmp * 1024ULL * 1024ULL;
4218 rdev->mc.real_vram_size = rdev->mc.mc_vram_size;
d2800ee5
AD
4219 rdev->mc.visible_vram_size = rdev->mc.aper_size;
4220 si_vram_gtt_location(rdev, &rdev->mc);
4221 radeon_update_bandwidth_info(rdev);
4222
4223 return 0;
4224}
4225
4226/*
4227 * GART
4228 */
4229void si_pcie_gart_tlb_flush(struct radeon_device *rdev)
4230{
4231 /* flush hdp cache */
4232 WREG32(HDP_MEM_COHERENCY_FLUSH_CNTL, 0x1);
4233
4234 /* bits 0-15 are the VM contexts0-15 */
4235 WREG32(VM_INVALIDATE_REQUEST, 1);
4236}
4237
1109ca09 4238static int si_pcie_gart_enable(struct radeon_device *rdev)
d2800ee5
AD
4239{
4240 int r, i;
4241
4242 if (rdev->gart.robj == NULL) {
4243 dev_err(rdev->dev, "No VRAM object for PCIE GART.\n");
4244 return -EINVAL;
4245 }
4246 r = radeon_gart_table_vram_pin(rdev);
4247 if (r)
4248 return r;
d2800ee5
AD
4249 /* Setup TLB control */
4250 WREG32(MC_VM_MX_L1_TLB_CNTL,
4251 (0xA << 7) |
4252 ENABLE_L1_TLB |
ec3dbbcb 4253 ENABLE_L1_FRAGMENT_PROCESSING |
d2800ee5
AD
4254 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
4255 ENABLE_ADVANCED_DRIVER_MODEL |
4256 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
4257 /* Setup L2 cache */
4258 WREG32(VM_L2_CNTL, ENABLE_L2_CACHE |
ec3dbbcb 4259 ENABLE_L2_FRAGMENT_PROCESSING |
d2800ee5
AD
4260 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
4261 ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
4262 EFFECTIVE_L2_QUEUE_SIZE(7) |
4263 CONTEXT1_IDENTITY_ACCESS_MODE(1));
4264 WREG32(VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS | INVALIDATE_L2_CACHE);
4265 WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
ec3dbbcb
CK
4266 BANK_SELECT(4) |
4267 L2_CACHE_BIGK_FRAGMENT_SIZE(4));
d2800ee5
AD
4268 /* setup context0 */
4269 WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);
4270 WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);
4271 WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12);
4272 WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
4273 (u32)(rdev->dummy_page.addr >> 12));
4274 WREG32(VM_CONTEXT0_CNTL2, 0);
4275 WREG32(VM_CONTEXT0_CNTL, (ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) |
4276 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT));
4277
4278 WREG32(0x15D4, 0);
4279 WREG32(0x15D8, 0);
4280 WREG32(0x15DC, 0);
4281
4282 /* empty context1-15 */
d2800ee5
AD
4283 /* set vm size, must be a multiple of 4 */
4284 WREG32(VM_CONTEXT1_PAGE_TABLE_START_ADDR, 0);
c21b328e 4285 WREG32(VM_CONTEXT1_PAGE_TABLE_END_ADDR, rdev->vm_manager.max_pfn);
23d4f1f2
AD
4286 /* Assign the pt base to something valid for now; the pts used for
4287 * the VMs are determined by the application and setup and assigned
4288 * on the fly in the vm part of radeon_gart.c
4289 */
d2800ee5
AD
4290 for (i = 1; i < 16; i++) {
4291 if (i < 8)
4292 WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (i << 2),
054e01d6 4293 rdev->vm_manager.saved_table_addr[i]);
d2800ee5
AD
4294 else
4295 WREG32(VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((i - 8) << 2),
054e01d6 4296 rdev->vm_manager.saved_table_addr[i]);
d2800ee5
AD
4297 }
4298
4299 /* enable context1-15 */
4300 WREG32(VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR,
4301 (u32)(rdev->dummy_page.addr >> 12));
ae133a11 4302 WREG32(VM_CONTEXT1_CNTL2, 4);
fa87e62d 4303 WREG32(VM_CONTEXT1_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(1) |
4510fb98 4304 PAGE_TABLE_BLOCK_SIZE(radeon_vm_block_size - 9) |
ae133a11
CK
4305 RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT |
4306 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT |
4307 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT |
4308 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT |
4309 PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT |
4310 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT |
4311 VALID_PROTECTION_FAULT_ENABLE_INTERRUPT |
4312 VALID_PROTECTION_FAULT_ENABLE_DEFAULT |
4313 READ_PROTECTION_FAULT_ENABLE_INTERRUPT |
4314 READ_PROTECTION_FAULT_ENABLE_DEFAULT |
4315 WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT |
4316 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT);
d2800ee5
AD
4317
4318 si_pcie_gart_tlb_flush(rdev);
4319 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
4320 (unsigned)(rdev->mc.gtt_size >> 20),
4321 (unsigned long long)rdev->gart.table_addr);
4322 rdev->gart.ready = true;
4323 return 0;
4324}
4325
1109ca09 4326static void si_pcie_gart_disable(struct radeon_device *rdev)
d2800ee5 4327{
054e01d6
CK
4328 unsigned i;
4329
4330 for (i = 1; i < 16; ++i) {
4331 uint32_t reg;
4332 if (i < 8)
4333 reg = VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (i << 2);
4334 else
4335 reg = VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((i - 8) << 2);
4336 rdev->vm_manager.saved_table_addr[i] = RREG32(reg);
4337 }
4338
d2800ee5
AD
4339 /* Disable all tables */
4340 WREG32(VM_CONTEXT0_CNTL, 0);
4341 WREG32(VM_CONTEXT1_CNTL, 0);
4342 /* Setup TLB control */
4343 WREG32(MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE_NOT_IN_SYS |
4344 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
4345 /* Setup L2 cache */
4346 WREG32(VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
4347 ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
4348 EFFECTIVE_L2_QUEUE_SIZE(7) |
4349 CONTEXT1_IDENTITY_ACCESS_MODE(1));
4350 WREG32(VM_L2_CNTL2, 0);
4351 WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
4352 L2_CACHE_BIGK_FRAGMENT_SIZE(0));
4353 radeon_gart_table_vram_unpin(rdev);
4354}
4355
1109ca09 4356static void si_pcie_gart_fini(struct radeon_device *rdev)
d2800ee5
AD
4357{
4358 si_pcie_gart_disable(rdev);
4359 radeon_gart_table_vram_free(rdev);
4360 radeon_gart_fini(rdev);
4361}
4362
498dd8b3
AD
4363/* vm parser */
4364static bool si_vm_reg_valid(u32 reg)
4365{
4366 /* context regs are fine */
4367 if (reg >= 0x28000)
4368 return true;
4369
4370 /* check config regs */
4371 switch (reg) {
4372 case GRBM_GFX_INDEX:
f418b88a 4373 case CP_STRMOUT_CNTL:
498dd8b3
AD
4374 case VGT_VTX_VECT_EJECT_REG:
4375 case VGT_CACHE_INVALIDATION:
4376 case VGT_ESGS_RING_SIZE:
4377 case VGT_GSVS_RING_SIZE:
4378 case VGT_GS_VERTEX_REUSE:
4379 case VGT_PRIMITIVE_TYPE:
4380 case VGT_INDEX_TYPE:
4381 case VGT_NUM_INDICES:
4382 case VGT_NUM_INSTANCES:
4383 case VGT_TF_RING_SIZE:
4384 case VGT_HS_OFFCHIP_PARAM:
4385 case VGT_TF_MEMORY_BASE:
4386 case PA_CL_ENHANCE:
4387 case PA_SU_LINE_STIPPLE_VALUE:
4388 case PA_SC_LINE_STIPPLE_STATE:
4389 case PA_SC_ENHANCE:
4390 case SQC_CACHES:
4391 case SPI_STATIC_THREAD_MGMT_1:
4392 case SPI_STATIC_THREAD_MGMT_2:
4393 case SPI_STATIC_THREAD_MGMT_3:
4394 case SPI_PS_MAX_WAVE_ID:
4395 case SPI_CONFIG_CNTL:
4396 case SPI_CONFIG_CNTL_1:
4397 case TA_CNTL_AUX:
4398 return true;
4399 default:
4400 DRM_ERROR("Invalid register 0x%x in CS\n", reg);
4401 return false;
4402 }
4403}
4404
4405static int si_vm_packet3_ce_check(struct radeon_device *rdev,
4406 u32 *ib, struct radeon_cs_packet *pkt)
4407{
4408 switch (pkt->opcode) {
4409 case PACKET3_NOP:
4410 case PACKET3_SET_BASE:
4411 case PACKET3_SET_CE_DE_COUNTERS:
4412 case PACKET3_LOAD_CONST_RAM:
4413 case PACKET3_WRITE_CONST_RAM:
4414 case PACKET3_WRITE_CONST_RAM_OFFSET:
4415 case PACKET3_DUMP_CONST_RAM:
4416 case PACKET3_INCREMENT_CE_COUNTER:
4417 case PACKET3_WAIT_ON_DE_COUNTER:
4418 case PACKET3_CE_WRITE:
4419 break;
4420 default:
4421 DRM_ERROR("Invalid CE packet3: 0x%x\n", pkt->opcode);
4422 return -EINVAL;
4423 }
4424 return 0;
4425}
4426
e5b9e750
TS
4427static int si_vm_packet3_cp_dma_check(u32 *ib, u32 idx)
4428{
4429 u32 start_reg, reg, i;
4430 u32 command = ib[idx + 4];
4431 u32 info = ib[idx + 1];
4432 u32 idx_value = ib[idx];
4433 if (command & PACKET3_CP_DMA_CMD_SAS) {
4434 /* src address space is register */
4435 if (((info & 0x60000000) >> 29) == 0) {
4436 start_reg = idx_value << 2;
4437 if (command & PACKET3_CP_DMA_CMD_SAIC) {
4438 reg = start_reg;
4439 if (!si_vm_reg_valid(reg)) {
4440 DRM_ERROR("CP DMA Bad SRC register\n");
4441 return -EINVAL;
4442 }
4443 } else {
4444 for (i = 0; i < (command & 0x1fffff); i++) {
4445 reg = start_reg + (4 * i);
4446 if (!si_vm_reg_valid(reg)) {
4447 DRM_ERROR("CP DMA Bad SRC register\n");
4448 return -EINVAL;
4449 }
4450 }
4451 }
4452 }
4453 }
4454 if (command & PACKET3_CP_DMA_CMD_DAS) {
4455 /* dst address space is register */
4456 if (((info & 0x00300000) >> 20) == 0) {
4457 start_reg = ib[idx + 2];
4458 if (command & PACKET3_CP_DMA_CMD_DAIC) {
4459 reg = start_reg;
4460 if (!si_vm_reg_valid(reg)) {
4461 DRM_ERROR("CP DMA Bad DST register\n");
4462 return -EINVAL;
4463 }
4464 } else {
4465 for (i = 0; i < (command & 0x1fffff); i++) {
4466 reg = start_reg + (4 * i);
4467 if (!si_vm_reg_valid(reg)) {
4468 DRM_ERROR("CP DMA Bad DST register\n");
4469 return -EINVAL;
4470 }
4471 }
4472 }
4473 }
4474 }
4475 return 0;
4476}
4477
498dd8b3
AD
4478static int si_vm_packet3_gfx_check(struct radeon_device *rdev,
4479 u32 *ib, struct radeon_cs_packet *pkt)
4480{
e5b9e750 4481 int r;
498dd8b3
AD
4482 u32 idx = pkt->idx + 1;
4483 u32 idx_value = ib[idx];
4484 u32 start_reg, end_reg, reg, i;
4485
4486 switch (pkt->opcode) {
4487 case PACKET3_NOP:
4488 case PACKET3_SET_BASE:
4489 case PACKET3_CLEAR_STATE:
4490 case PACKET3_INDEX_BUFFER_SIZE:
4491 case PACKET3_DISPATCH_DIRECT:
4492 case PACKET3_DISPATCH_INDIRECT:
4493 case PACKET3_ALLOC_GDS:
4494 case PACKET3_WRITE_GDS_RAM:
4495 case PACKET3_ATOMIC_GDS:
4496 case PACKET3_ATOMIC:
4497 case PACKET3_OCCLUSION_QUERY:
4498 case PACKET3_SET_PREDICATION:
4499 case PACKET3_COND_EXEC:
4500 case PACKET3_PRED_EXEC:
4501 case PACKET3_DRAW_INDIRECT:
4502 case PACKET3_DRAW_INDEX_INDIRECT:
4503 case PACKET3_INDEX_BASE:
4504 case PACKET3_DRAW_INDEX_2:
4505 case PACKET3_CONTEXT_CONTROL:
4506 case PACKET3_INDEX_TYPE:
4507 case PACKET3_DRAW_INDIRECT_MULTI:
4508 case PACKET3_DRAW_INDEX_AUTO:
4509 case PACKET3_DRAW_INDEX_IMMD:
4510 case PACKET3_NUM_INSTANCES:
4511 case PACKET3_DRAW_INDEX_MULTI_AUTO:
4512 case PACKET3_STRMOUT_BUFFER_UPDATE:
4513 case PACKET3_DRAW_INDEX_OFFSET_2:
4514 case PACKET3_DRAW_INDEX_MULTI_ELEMENT:
4515 case PACKET3_DRAW_INDEX_INDIRECT_MULTI:
4516 case PACKET3_MPEG_INDEX:
4517 case PACKET3_WAIT_REG_MEM:
4518 case PACKET3_MEM_WRITE:
4519 case PACKET3_PFP_SYNC_ME:
4520 case PACKET3_SURFACE_SYNC:
4521 case PACKET3_EVENT_WRITE:
4522 case PACKET3_EVENT_WRITE_EOP:
4523 case PACKET3_EVENT_WRITE_EOS:
4524 case PACKET3_SET_CONTEXT_REG:
4525 case PACKET3_SET_CONTEXT_REG_INDIRECT:
4526 case PACKET3_SET_SH_REG:
4527 case PACKET3_SET_SH_REG_OFFSET:
4528 case PACKET3_INCREMENT_DE_COUNTER:
4529 case PACKET3_WAIT_ON_CE_COUNTER:
4530 case PACKET3_WAIT_ON_AVAIL_BUFFER:
4531 case PACKET3_ME_WRITE:
4532 break;
4533 case PACKET3_COPY_DATA:
4534 if ((idx_value & 0xf00) == 0) {
4535 reg = ib[idx + 3] * 4;
4536 if (!si_vm_reg_valid(reg))
4537 return -EINVAL;
4538 }
4539 break;
4540 case PACKET3_WRITE_DATA:
4541 if ((idx_value & 0xf00) == 0) {
4542 start_reg = ib[idx + 1] * 4;
4543 if (idx_value & 0x10000) {
4544 if (!si_vm_reg_valid(start_reg))
4545 return -EINVAL;
4546 } else {
4547 for (i = 0; i < (pkt->count - 2); i++) {
4548 reg = start_reg + (4 * i);
4549 if (!si_vm_reg_valid(reg))
4550 return -EINVAL;
4551 }
4552 }
4553 }
4554 break;
4555 case PACKET3_COND_WRITE:
4556 if (idx_value & 0x100) {
4557 reg = ib[idx + 5] * 4;
4558 if (!si_vm_reg_valid(reg))
4559 return -EINVAL;
4560 }
4561 break;
4562 case PACKET3_COPY_DW:
4563 if (idx_value & 0x2) {
4564 reg = ib[idx + 3] * 4;
4565 if (!si_vm_reg_valid(reg))
4566 return -EINVAL;
4567 }
4568 break;
4569 case PACKET3_SET_CONFIG_REG:
4570 start_reg = (idx_value << 2) + PACKET3_SET_CONFIG_REG_START;
4571 end_reg = 4 * pkt->count + start_reg - 4;
4572 if ((start_reg < PACKET3_SET_CONFIG_REG_START) ||
4573 (start_reg >= PACKET3_SET_CONFIG_REG_END) ||
4574 (end_reg >= PACKET3_SET_CONFIG_REG_END)) {
4575 DRM_ERROR("bad PACKET3_SET_CONFIG_REG\n");
4576 return -EINVAL;
4577 }
4578 for (i = 0; i < pkt->count; i++) {
4579 reg = start_reg + (4 * i);
4580 if (!si_vm_reg_valid(reg))
4581 return -EINVAL;
4582 }
4583 break;
5aa709be 4584 case PACKET3_CP_DMA:
e5b9e750
TS
4585 r = si_vm_packet3_cp_dma_check(ib, idx);
4586 if (r)
4587 return r;
5aa709be 4588 break;
498dd8b3
AD
4589 default:
4590 DRM_ERROR("Invalid GFX packet3: 0x%x\n", pkt->opcode);
4591 return -EINVAL;
4592 }
4593 return 0;
4594}
4595
4596static int si_vm_packet3_compute_check(struct radeon_device *rdev,
4597 u32 *ib, struct radeon_cs_packet *pkt)
4598{
e5b9e750 4599 int r;
498dd8b3
AD
4600 u32 idx = pkt->idx + 1;
4601 u32 idx_value = ib[idx];
4602 u32 start_reg, reg, i;
4603
4604 switch (pkt->opcode) {
4605 case PACKET3_NOP:
4606 case PACKET3_SET_BASE:
4607 case PACKET3_CLEAR_STATE:
4608 case PACKET3_DISPATCH_DIRECT:
4609 case PACKET3_DISPATCH_INDIRECT:
4610 case PACKET3_ALLOC_GDS:
4611 case PACKET3_WRITE_GDS_RAM:
4612 case PACKET3_ATOMIC_GDS:
4613 case PACKET3_ATOMIC:
4614 case PACKET3_OCCLUSION_QUERY:
4615 case PACKET3_SET_PREDICATION:
4616 case PACKET3_COND_EXEC:
4617 case PACKET3_PRED_EXEC:
4618 case PACKET3_CONTEXT_CONTROL:
4619 case PACKET3_STRMOUT_BUFFER_UPDATE:
4620 case PACKET3_WAIT_REG_MEM:
4621 case PACKET3_MEM_WRITE:
4622 case PACKET3_PFP_SYNC_ME:
4623 case PACKET3_SURFACE_SYNC:
4624 case PACKET3_EVENT_WRITE:
4625 case PACKET3_EVENT_WRITE_EOP:
4626 case PACKET3_EVENT_WRITE_EOS:
4627 case PACKET3_SET_CONTEXT_REG:
4628 case PACKET3_SET_CONTEXT_REG_INDIRECT:
4629 case PACKET3_SET_SH_REG:
4630 case PACKET3_SET_SH_REG_OFFSET:
4631 case PACKET3_INCREMENT_DE_COUNTER:
4632 case PACKET3_WAIT_ON_CE_COUNTER:
4633 case PACKET3_WAIT_ON_AVAIL_BUFFER:
4634 case PACKET3_ME_WRITE:
4635 break;
4636 case PACKET3_COPY_DATA:
4637 if ((idx_value & 0xf00) == 0) {
4638 reg = ib[idx + 3] * 4;
4639 if (!si_vm_reg_valid(reg))
4640 return -EINVAL;
4641 }
4642 break;
4643 case PACKET3_WRITE_DATA:
4644 if ((idx_value & 0xf00) == 0) {
4645 start_reg = ib[idx + 1] * 4;
4646 if (idx_value & 0x10000) {
4647 if (!si_vm_reg_valid(start_reg))
4648 return -EINVAL;
4649 } else {
4650 for (i = 0; i < (pkt->count - 2); i++) {
4651 reg = start_reg + (4 * i);
4652 if (!si_vm_reg_valid(reg))
4653 return -EINVAL;
4654 }
4655 }
4656 }
4657 break;
4658 case PACKET3_COND_WRITE:
4659 if (idx_value & 0x100) {
4660 reg = ib[idx + 5] * 4;
4661 if (!si_vm_reg_valid(reg))
4662 return -EINVAL;
4663 }
4664 break;
4665 case PACKET3_COPY_DW:
4666 if (idx_value & 0x2) {
4667 reg = ib[idx + 3] * 4;
4668 if (!si_vm_reg_valid(reg))
4669 return -EINVAL;
4670 }
4671 break;
e5b9e750
TS
4672 case PACKET3_CP_DMA:
4673 r = si_vm_packet3_cp_dma_check(ib, idx);
4674 if (r)
4675 return r;
4676 break;
498dd8b3
AD
4677 default:
4678 DRM_ERROR("Invalid Compute packet3: 0x%x\n", pkt->opcode);
4679 return -EINVAL;
4680 }
4681 return 0;
4682}
4683
4684int si_ib_parse(struct radeon_device *rdev, struct radeon_ib *ib)
4685{
4686 int ret = 0;
4687 u32 idx = 0;
4688 struct radeon_cs_packet pkt;
4689
4690 do {
4691 pkt.idx = idx;
4e872ae2
IH
4692 pkt.type = RADEON_CP_PACKET_GET_TYPE(ib->ptr[idx]);
4693 pkt.count = RADEON_CP_PACKET_GET_COUNT(ib->ptr[idx]);
498dd8b3
AD
4694 pkt.one_reg_wr = 0;
4695 switch (pkt.type) {
4e872ae2 4696 case RADEON_PACKET_TYPE0:
498dd8b3
AD
4697 dev_err(rdev->dev, "Packet0 not allowed!\n");
4698 ret = -EINVAL;
4699 break;
4e872ae2 4700 case RADEON_PACKET_TYPE2:
498dd8b3
AD
4701 idx += 1;
4702 break;
4e872ae2
IH
4703 case RADEON_PACKET_TYPE3:
4704 pkt.opcode = RADEON_CP_PACKET3_GET_OPCODE(ib->ptr[idx]);
498dd8b3
AD
4705 if (ib->is_const_ib)
4706 ret = si_vm_packet3_ce_check(rdev, ib->ptr, &pkt);
4707 else {
876dc9f3 4708 switch (ib->ring) {
498dd8b3
AD
4709 case RADEON_RING_TYPE_GFX_INDEX:
4710 ret = si_vm_packet3_gfx_check(rdev, ib->ptr, &pkt);
4711 break;
4712 case CAYMAN_RING_TYPE_CP1_INDEX:
4713 case CAYMAN_RING_TYPE_CP2_INDEX:
4714 ret = si_vm_packet3_compute_check(rdev, ib->ptr, &pkt);
4715 break;
4716 default:
876dc9f3 4717 dev_err(rdev->dev, "Non-PM4 ring %d !\n", ib->ring);
498dd8b3
AD
4718 ret = -EINVAL;
4719 break;
4720 }
4721 }
4722 idx += pkt.count + 2;
4723 break;
4724 default:
4725 dev_err(rdev->dev, "Unknown packet type %d !\n", pkt.type);
4726 ret = -EINVAL;
4727 break;
4728 }
4729 if (ret)
4730 break;
4731 } while (idx < ib->length_dw);
4732
4733 return ret;
4734}
4735
d2800ee5
AD
4736/*
4737 * vm
4738 */
4739int si_vm_init(struct radeon_device *rdev)
4740{
4741 /* number of VMs */
4742 rdev->vm_manager.nvm = 16;
4743 /* base offset of vram pages */
4744 rdev->vm_manager.vram_base_offset = 0;
4745
4746 return 0;
4747}
4748
4749void si_vm_fini(struct radeon_device *rdev)
4750{
4751}
4752
fbf6dc7a
AD
4753/**
4754 * si_vm_decode_fault - print human readable fault info
4755 *
4756 * @rdev: radeon_device pointer
4757 * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value
4758 * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value
4759 *
4760 * Print human readable fault information (SI).
4761 */
4762static void si_vm_decode_fault(struct radeon_device *rdev,
4763 u32 status, u32 addr)
4764{
4765 u32 mc_id = (status & MEMORY_CLIENT_ID_MASK) >> MEMORY_CLIENT_ID_SHIFT;
4766 u32 vmid = (status & FAULT_VMID_MASK) >> FAULT_VMID_SHIFT;
4767 u32 protections = (status & PROTECTIONS_MASK) >> PROTECTIONS_SHIFT;
4768 char *block;
4769
4770 if (rdev->family == CHIP_TAHITI) {
4771 switch (mc_id) {
4772 case 160:
4773 case 144:
4774 case 96:
4775 case 80:
4776 case 224:
4777 case 208:
4778 case 32:
4779 case 16:
4780 block = "CB";
4781 break;
4782 case 161:
4783 case 145:
4784 case 97:
4785 case 81:
4786 case 225:
4787 case 209:
4788 case 33:
4789 case 17:
4790 block = "CB_FMASK";
4791 break;
4792 case 162:
4793 case 146:
4794 case 98:
4795 case 82:
4796 case 226:
4797 case 210:
4798 case 34:
4799 case 18:
4800 block = "CB_CMASK";
4801 break;
4802 case 163:
4803 case 147:
4804 case 99:
4805 case 83:
4806 case 227:
4807 case 211:
4808 case 35:
4809 case 19:
4810 block = "CB_IMMED";
4811 break;
4812 case 164:
4813 case 148:
4814 case 100:
4815 case 84:
4816 case 228:
4817 case 212:
4818 case 36:
4819 case 20:
4820 block = "DB";
4821 break;
4822 case 165:
4823 case 149:
4824 case 101:
4825 case 85:
4826 case 229:
4827 case 213:
4828 case 37:
4829 case 21:
4830 block = "DB_HTILE";
4831 break;
4832 case 167:
4833 case 151:
4834 case 103:
4835 case 87:
4836 case 231:
4837 case 215:
4838 case 39:
4839 case 23:
4840 block = "DB_STEN";
4841 break;
4842 case 72:
4843 case 68:
4844 case 64:
4845 case 8:
4846 case 4:
4847 case 0:
4848 case 136:
4849 case 132:
4850 case 128:
4851 case 200:
4852 case 196:
4853 case 192:
4854 block = "TC";
4855 break;
4856 case 112:
4857 case 48:
4858 block = "CP";
4859 break;
4860 case 49:
4861 case 177:
4862 case 50:
4863 case 178:
4864 block = "SH";
4865 break;
4866 case 53:
4867 case 190:
4868 block = "VGT";
4869 break;
4870 case 117:
4871 block = "IH";
4872 break;
4873 case 51:
4874 case 115:
4875 block = "RLC";
4876 break;
4877 case 119:
4878 case 183:
4879 block = "DMA0";
4880 break;
4881 case 61:
4882 block = "DMA1";
4883 break;
4884 case 248:
4885 case 120:
4886 block = "HDP";
4887 break;
4888 default:
4889 block = "unknown";
4890 break;
4891 }
4892 } else {
4893 switch (mc_id) {
4894 case 32:
4895 case 16:
4896 case 96:
4897 case 80:
4898 case 160:
4899 case 144:
4900 case 224:
4901 case 208:
4902 block = "CB";
4903 break;
4904 case 33:
4905 case 17:
4906 case 97:
4907 case 81:
4908 case 161:
4909 case 145:
4910 case 225:
4911 case 209:
4912 block = "CB_FMASK";
4913 break;
4914 case 34:
4915 case 18:
4916 case 98:
4917 case 82:
4918 case 162:
4919 case 146:
4920 case 226:
4921 case 210:
4922 block = "CB_CMASK";
4923 break;
4924 case 35:
4925 case 19:
4926 case 99:
4927 case 83:
4928 case 163:
4929 case 147:
4930 case 227:
4931 case 211:
4932 block = "CB_IMMED";
4933 break;
4934 case 36:
4935 case 20:
4936 case 100:
4937 case 84:
4938 case 164:
4939 case 148:
4940 case 228:
4941 case 212:
4942 block = "DB";
4943 break;
4944 case 37:
4945 case 21:
4946 case 101:
4947 case 85:
4948 case 165:
4949 case 149:
4950 case 229:
4951 case 213:
4952 block = "DB_HTILE";
4953 break;
4954 case 39:
4955 case 23:
4956 case 103:
4957 case 87:
4958 case 167:
4959 case 151:
4960 case 231:
4961 case 215:
4962 block = "DB_STEN";
4963 break;
4964 case 72:
4965 case 68:
4966 case 8:
4967 case 4:
4968 case 136:
4969 case 132:
4970 case 200:
4971 case 196:
4972 block = "TC";
4973 break;
4974 case 112:
4975 case 48:
4976 block = "CP";
4977 break;
4978 case 49:
4979 case 177:
4980 case 50:
4981 case 178:
4982 block = "SH";
4983 break;
4984 case 53:
4985 block = "VGT";
4986 break;
4987 case 117:
4988 block = "IH";
4989 break;
4990 case 51:
4991 case 115:
4992 block = "RLC";
4993 break;
4994 case 119:
4995 case 183:
4996 block = "DMA0";
4997 break;
4998 case 61:
4999 block = "DMA1";
5000 break;
5001 case 248:
5002 case 120:
5003 block = "HDP";
5004 break;
5005 default:
5006 block = "unknown";
5007 break;
5008 }
5009 }
5010
5011 printk("VM fault (0x%02x, vmid %d) at page %u, %s from %s (%d)\n",
5012 protections, vmid, addr,
5013 (status & MEMORY_CLIENT_RW_MASK) ? "write" : "read",
5014 block, mc_id);
5015}
5016
498522b4 5017void si_vm_flush(struct radeon_device *rdev, int ridx, struct radeon_vm *vm)
d2800ee5 5018{
498522b4 5019 struct radeon_ring *ring = &rdev->ring[ridx];
d2800ee5 5020
ee60e29f 5021 if (vm == NULL)
d2800ee5
AD
5022 return;
5023
76c44f2c
AD
5024 /* write new base address */
5025 radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
f1d2a26b 5026 radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(1) |
76c44f2c
AD
5027 WRITE_DATA_DST_SEL(0)));
5028
ee60e29f 5029 if (vm->id < 8) {
76c44f2c
AD
5030 radeon_ring_write(ring,
5031 (VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm->id << 2)) >> 2);
ee60e29f 5032 } else {
76c44f2c
AD
5033 radeon_ring_write(ring,
5034 (VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((vm->id - 8) << 2)) >> 2);
ee60e29f 5035 }
76c44f2c 5036 radeon_ring_write(ring, 0);
fa87e62d 5037 radeon_ring_write(ring, vm->pd_gpu_addr >> 12);
ee60e29f 5038
d2800ee5 5039 /* flush hdp cache */
76c44f2c 5040 radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4fb0bbd5 5041 radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(1) |
76c44f2c
AD
5042 WRITE_DATA_DST_SEL(0)));
5043 radeon_ring_write(ring, HDP_MEM_COHERENCY_FLUSH_CNTL >> 2);
5044 radeon_ring_write(ring, 0);
ee60e29f
CK
5045 radeon_ring_write(ring, 0x1);
5046
d2800ee5 5047 /* bits 0-15 are the VM contexts0-15 */
76c44f2c 5048 radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4fb0bbd5 5049 radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(1) |
76c44f2c
AD
5050 WRITE_DATA_DST_SEL(0)));
5051 radeon_ring_write(ring, VM_INVALIDATE_REQUEST >> 2);
5052 radeon_ring_write(ring, 0);
498522b4 5053 radeon_ring_write(ring, 1 << vm->id);
58f8cf56
CK
5054
5055 /* sync PFP to ME, otherwise we might get invalid PFP reads */
5056 radeon_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
5057 radeon_ring_write(ring, 0x0);
d2800ee5
AD
5058}
5059
f8f84ac5
AD
5060/*
5061 * Power and clock gating
5062 */
5063static void si_wait_for_rlc_serdes(struct radeon_device *rdev)
5064{
5065 int i;
5066
5067 for (i = 0; i < rdev->usec_timeout; i++) {
5068 if (RREG32(RLC_SERDES_MASTER_BUSY_0) == 0)
5069 break;
5070 udelay(1);
5071 }
5072
5073 for (i = 0; i < rdev->usec_timeout; i++) {
5074 if (RREG32(RLC_SERDES_MASTER_BUSY_1) == 0)
5075 break;
5076 udelay(1);
5077 }
5078}
5079
5080static void si_enable_gui_idle_interrupt(struct radeon_device *rdev,
5081 bool enable)
5082{
5083 u32 tmp = RREG32(CP_INT_CNTL_RING0);
5084 u32 mask;
5085 int i;
5086
5087 if (enable)
5088 tmp |= (CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
5089 else
5090 tmp &= ~(CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
5091 WREG32(CP_INT_CNTL_RING0, tmp);
5092
5093 if (!enable) {
5094 /* read a gfx register */
5095 tmp = RREG32(DB_DEPTH_INFO);
5096
5097 mask = RLC_BUSY_STATUS | GFX_POWER_STATUS | GFX_CLOCK_STATUS | GFX_LS_STATUS;
5098 for (i = 0; i < rdev->usec_timeout; i++) {
5099 if ((RREG32(RLC_STAT) & mask) == (GFX_CLOCK_STATUS | GFX_POWER_STATUS))
5100 break;
5101 udelay(1);
5102 }
5103 }
5104}
5105
5106static void si_set_uvd_dcm(struct radeon_device *rdev,
5107 bool sw_mode)
5108{
5109 u32 tmp, tmp2;
5110
5111 tmp = RREG32(UVD_CGC_CTRL);
5112 tmp &= ~(CLK_OD_MASK | CG_DT_MASK);
5113 tmp |= DCM | CG_DT(1) | CLK_OD(4);
5114
5115 if (sw_mode) {
5116 tmp &= ~0x7ffff800;
5117 tmp2 = DYN_OR_EN | DYN_RR_EN | G_DIV_ID(7);
5118 } else {
5119 tmp |= 0x7ffff800;
5120 tmp2 = 0;
5121 }
5122
5123 WREG32(UVD_CGC_CTRL, tmp);
5124 WREG32_UVD_CTX(UVD_CGC_CTRL2, tmp2);
5125}
5126
22c775ce 5127void si_init_uvd_internal_cg(struct radeon_device *rdev)
f8f84ac5
AD
5128{
5129 bool hw_mode = true;
5130
5131 if (hw_mode) {
5132 si_set_uvd_dcm(rdev, false);
5133 } else {
5134 u32 tmp = RREG32(UVD_CGC_CTRL);
5135 tmp &= ~DCM;
5136 WREG32(UVD_CGC_CTRL, tmp);
5137 }
5138}
5139
5140static u32 si_halt_rlc(struct radeon_device *rdev)
5141{
5142 u32 data, orig;
5143
5144 orig = data = RREG32(RLC_CNTL);
5145
5146 if (data & RLC_ENABLE) {
5147 data &= ~RLC_ENABLE;
5148 WREG32(RLC_CNTL, data);
5149
5150 si_wait_for_rlc_serdes(rdev);
5151 }
5152
5153 return orig;
5154}
5155
5156static void si_update_rlc(struct radeon_device *rdev, u32 rlc)
5157{
5158 u32 tmp;
5159
5160 tmp = RREG32(RLC_CNTL);
5161 if (tmp != rlc)
5162 WREG32(RLC_CNTL, rlc);
5163}
5164
5165static void si_enable_dma_pg(struct radeon_device *rdev, bool enable)
5166{
5167 u32 data, orig;
5168
5169 orig = data = RREG32(DMA_PG);
e16866ec 5170 if (enable && (rdev->pg_flags & RADEON_PG_SUPPORT_SDMA))
f8f84ac5
AD
5171 data |= PG_CNTL_ENABLE;
5172 else
5173 data &= ~PG_CNTL_ENABLE;
5174 if (orig != data)
5175 WREG32(DMA_PG, data);
5176}
5177
5178static void si_init_dma_pg(struct radeon_device *rdev)
5179{
5180 u32 tmp;
5181
5182 WREG32(DMA_PGFSM_WRITE, 0x00002000);
5183 WREG32(DMA_PGFSM_CONFIG, 0x100010ff);
5184
5185 for (tmp = 0; tmp < 5; tmp++)
5186 WREG32(DMA_PGFSM_WRITE, 0);
5187}
5188
5189static void si_enable_gfx_cgpg(struct radeon_device *rdev,
5190 bool enable)
5191{
5192 u32 tmp;
5193
2b19d17f 5194 if (enable && (rdev->pg_flags & RADEON_PG_SUPPORT_GFX_PG)) {
f8f84ac5
AD
5195 tmp = RLC_PUD(0x10) | RLC_PDD(0x10) | RLC_TTPD(0x10) | RLC_MSD(0x10);
5196 WREG32(RLC_TTOP_D, tmp);
5197
5198 tmp = RREG32(RLC_PG_CNTL);
5199 tmp |= GFX_PG_ENABLE;
5200 WREG32(RLC_PG_CNTL, tmp);
5201
5202 tmp = RREG32(RLC_AUTO_PG_CTRL);
5203 tmp |= AUTO_PG_EN;
5204 WREG32(RLC_AUTO_PG_CTRL, tmp);
5205 } else {
5206 tmp = RREG32(RLC_AUTO_PG_CTRL);
5207 tmp &= ~AUTO_PG_EN;
5208 WREG32(RLC_AUTO_PG_CTRL, tmp);
5209
5210 tmp = RREG32(DB_RENDER_CONTROL);
5211 }
5212}
5213
5214static void si_init_gfx_cgpg(struct radeon_device *rdev)
5215{
5216 u32 tmp;
5217
5218 WREG32(RLC_SAVE_AND_RESTORE_BASE, rdev->rlc.save_restore_gpu_addr >> 8);
5219
5220 tmp = RREG32(RLC_PG_CNTL);
5221 tmp |= GFX_PG_SRC;
5222 WREG32(RLC_PG_CNTL, tmp);
5223
5224 WREG32(RLC_CLEAR_STATE_RESTORE_BASE, rdev->rlc.clear_state_gpu_addr >> 8);
5225
5226 tmp = RREG32(RLC_AUTO_PG_CTRL);
5227
5228 tmp &= ~GRBM_REG_SGIT_MASK;
5229 tmp |= GRBM_REG_SGIT(0x700);
5230 tmp &= ~PG_AFTER_GRBM_REG_ST_MASK;
5231 WREG32(RLC_AUTO_PG_CTRL, tmp);
5232}
5233
ba19031a 5234static u32 si_get_cu_active_bitmap(struct radeon_device *rdev, u32 se, u32 sh)
f8f84ac5
AD
5235{
5236 u32 mask = 0, tmp, tmp1;
5237 int i;
5238
5239 si_select_se_sh(rdev, se, sh);
5240 tmp = RREG32(CC_GC_SHADER_ARRAY_CONFIG);
5241 tmp1 = RREG32(GC_USER_SHADER_ARRAY_CONFIG);
5242 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
5243
5244 tmp &= 0xffff0000;
5245
5246 tmp |= tmp1;
5247 tmp >>= 16;
5248
5249 for (i = 0; i < rdev->config.si.max_cu_per_sh; i ++) {
5250 mask <<= 1;
5251 mask |= 1;
5252 }
5253
5254 return (~tmp) & mask;
5255}
5256
5257static void si_init_ao_cu_mask(struct radeon_device *rdev)
5258{
5259 u32 i, j, k, active_cu_number = 0;
5260 u32 mask, counter, cu_bitmap;
5261 u32 tmp = 0;
5262
5263 for (i = 0; i < rdev->config.si.max_shader_engines; i++) {
5264 for (j = 0; j < rdev->config.si.max_sh_per_se; j++) {
5265 mask = 1;
5266 cu_bitmap = 0;
5267 counter = 0;
5268 for (k = 0; k < rdev->config.si.max_cu_per_sh; k++) {
ba19031a 5269 if (si_get_cu_active_bitmap(rdev, i, j) & mask) {
f8f84ac5
AD
5270 if (counter < 2)
5271 cu_bitmap |= mask;
5272 counter++;
5273 }
5274 mask <<= 1;
5275 }
5276
5277 active_cu_number += counter;
5278 tmp |= (cu_bitmap << (i * 16 + j * 8));
5279 }
5280 }
5281
5282 WREG32(RLC_PG_AO_CU_MASK, tmp);
5283
5284 tmp = RREG32(RLC_MAX_PG_CU);
5285 tmp &= ~MAX_PU_CU_MASK;
5286 tmp |= MAX_PU_CU(active_cu_number);
5287 WREG32(RLC_MAX_PG_CU, tmp);
5288}
5289
5290static void si_enable_cgcg(struct radeon_device *rdev,
5291 bool enable)
5292{
5293 u32 data, orig, tmp;
5294
5295 orig = data = RREG32(RLC_CGCG_CGLS_CTRL);
5296
e16866ec 5297 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_CGCG)) {
5594a558 5298 si_enable_gui_idle_interrupt(rdev, true);
f8f84ac5 5299
f8f84ac5
AD
5300 WREG32(RLC_GCPM_GENERAL_3, 0x00000080);
5301
5302 tmp = si_halt_rlc(rdev);
5303
5304 WREG32(RLC_SERDES_WR_MASTER_MASK_0, 0xffffffff);
5305 WREG32(RLC_SERDES_WR_MASTER_MASK_1, 0xffffffff);
5306 WREG32(RLC_SERDES_WR_CTRL, 0x00b000ff);
5307
5308 si_wait_for_rlc_serdes(rdev);
5309
5310 si_update_rlc(rdev, tmp);
5311
5312 WREG32(RLC_SERDES_WR_CTRL, 0x007000ff);
5313
5314 data |= CGCG_EN | CGLS_EN;
5315 } else {
5594a558
AD
5316 si_enable_gui_idle_interrupt(rdev, false);
5317
f8f84ac5
AD
5318 RREG32(CB_CGTT_SCLK_CTRL);
5319 RREG32(CB_CGTT_SCLK_CTRL);
5320 RREG32(CB_CGTT_SCLK_CTRL);
5321 RREG32(CB_CGTT_SCLK_CTRL);
5322
5323 data &= ~(CGCG_EN | CGLS_EN);
5324 }
5325
5326 if (orig != data)
5327 WREG32(RLC_CGCG_CGLS_CTRL, data);
5328}
5329
5330static void si_enable_mgcg(struct radeon_device *rdev,
5331 bool enable)
5332{
5333 u32 data, orig, tmp = 0;
5334
e16866ec 5335 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_MGCG)) {
f8f84ac5
AD
5336 orig = data = RREG32(CGTS_SM_CTRL_REG);
5337 data = 0x96940200;
5338 if (orig != data)
5339 WREG32(CGTS_SM_CTRL_REG, data);
5340
e16866ec
AD
5341 if (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_CP_LS) {
5342 orig = data = RREG32(CP_MEM_SLP_CNTL);
5343 data |= CP_MEM_LS_EN;
5344 if (orig != data)
5345 WREG32(CP_MEM_SLP_CNTL, data);
5346 }
f8f84ac5
AD
5347
5348 orig = data = RREG32(RLC_CGTT_MGCG_OVERRIDE);
5349 data &= 0xffffffc0;
5350 if (orig != data)
5351 WREG32(RLC_CGTT_MGCG_OVERRIDE, data);
5352
5353 tmp = si_halt_rlc(rdev);
5354
5355 WREG32(RLC_SERDES_WR_MASTER_MASK_0, 0xffffffff);
5356 WREG32(RLC_SERDES_WR_MASTER_MASK_1, 0xffffffff);
5357 WREG32(RLC_SERDES_WR_CTRL, 0x00d000ff);
5358
5359 si_update_rlc(rdev, tmp);
5360 } else {
5361 orig = data = RREG32(RLC_CGTT_MGCG_OVERRIDE);
5362 data |= 0x00000003;
5363 if (orig != data)
5364 WREG32(RLC_CGTT_MGCG_OVERRIDE, data);
5365
5366 data = RREG32(CP_MEM_SLP_CNTL);
5367 if (data & CP_MEM_LS_EN) {
5368 data &= ~CP_MEM_LS_EN;
5369 WREG32(CP_MEM_SLP_CNTL, data);
5370 }
5371 orig = data = RREG32(CGTS_SM_CTRL_REG);
5372 data |= LS_OVERRIDE | OVERRIDE;
5373 if (orig != data)
5374 WREG32(CGTS_SM_CTRL_REG, data);
5375
5376 tmp = si_halt_rlc(rdev);
5377
5378 WREG32(RLC_SERDES_WR_MASTER_MASK_0, 0xffffffff);
5379 WREG32(RLC_SERDES_WR_MASTER_MASK_1, 0xffffffff);
5380 WREG32(RLC_SERDES_WR_CTRL, 0x00e000ff);
5381
5382 si_update_rlc(rdev, tmp);
5383 }
5384}
5385
5386static void si_enable_uvd_mgcg(struct radeon_device *rdev,
5387 bool enable)
5388{
5389 u32 orig, data, tmp;
5390
e16866ec 5391 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_UVD_MGCG)) {
f8f84ac5
AD
5392 tmp = RREG32_UVD_CTX(UVD_CGC_MEM_CTRL);
5393 tmp |= 0x3fff;
5394 WREG32_UVD_CTX(UVD_CGC_MEM_CTRL, tmp);
5395
5396 orig = data = RREG32(UVD_CGC_CTRL);
5397 data |= DCM;
5398 if (orig != data)
5399 WREG32(UVD_CGC_CTRL, data);
5400
5401 WREG32_SMC(SMC_CG_IND_START + CG_CGTT_LOCAL_0, 0);
5402 WREG32_SMC(SMC_CG_IND_START + CG_CGTT_LOCAL_1, 0);
5403 } else {
5404 tmp = RREG32_UVD_CTX(UVD_CGC_MEM_CTRL);
5405 tmp &= ~0x3fff;
5406 WREG32_UVD_CTX(UVD_CGC_MEM_CTRL, tmp);
5407
5408 orig = data = RREG32(UVD_CGC_CTRL);
5409 data &= ~DCM;
5410 if (orig != data)
5411 WREG32(UVD_CGC_CTRL, data);
5412
5413 WREG32_SMC(SMC_CG_IND_START + CG_CGTT_LOCAL_0, 0xffffffff);
5414 WREG32_SMC(SMC_CG_IND_START + CG_CGTT_LOCAL_1, 0xffffffff);
5415 }
5416}
5417
5418static const u32 mc_cg_registers[] =
5419{
5420 MC_HUB_MISC_HUB_CG,
5421 MC_HUB_MISC_SIP_CG,
5422 MC_HUB_MISC_VM_CG,
5423 MC_XPB_CLK_GAT,
5424 ATC_MISC_CG,
5425 MC_CITF_MISC_WR_CG,
5426 MC_CITF_MISC_RD_CG,
5427 MC_CITF_MISC_VM_CG,
5428 VM_L2_CG,
5429};
5430
5431static void si_enable_mc_ls(struct radeon_device *rdev,
5432 bool enable)
5433{
5434 int i;
5435 u32 orig, data;
5436
5437 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) {
5438 orig = data = RREG32(mc_cg_registers[i]);
e16866ec 5439 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_MC_LS))
f8f84ac5
AD
5440 data |= MC_LS_ENABLE;
5441 else
5442 data &= ~MC_LS_ENABLE;
5443 if (data != orig)
5444 WREG32(mc_cg_registers[i], data);
5445 }
5446}
5447
e16866ec
AD
5448static void si_enable_mc_mgcg(struct radeon_device *rdev,
5449 bool enable)
f8f84ac5 5450{
e16866ec
AD
5451 int i;
5452 u32 orig, data;
5453
5454 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) {
5455 orig = data = RREG32(mc_cg_registers[i]);
5456 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_MC_MGCG))
5457 data |= MC_CG_ENABLE;
5458 else
5459 data &= ~MC_CG_ENABLE;
5460 if (data != orig)
5461 WREG32(mc_cg_registers[i], data);
f8f84ac5
AD
5462 }
5463}
5464
e16866ec
AD
5465static void si_enable_dma_mgcg(struct radeon_device *rdev,
5466 bool enable)
f8f84ac5 5467{
e16866ec
AD
5468 u32 orig, data, offset;
5469 int i;
f8f84ac5 5470
e16866ec
AD
5471 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_SDMA_MGCG)) {
5472 for (i = 0; i < 2; i++) {
5473 if (i == 0)
5474 offset = DMA0_REGISTER_OFFSET;
5475 else
5476 offset = DMA1_REGISTER_OFFSET;
5477 orig = data = RREG32(DMA_POWER_CNTL + offset);
5478 data &= ~MEM_POWER_OVERRIDE;
5479 if (data != orig)
5480 WREG32(DMA_POWER_CNTL + offset, data);
5481 WREG32(DMA_CLK_CTRL + offset, 0x00000100);
5482 }
5483 } else {
5484 for (i = 0; i < 2; i++) {
5485 if (i == 0)
5486 offset = DMA0_REGISTER_OFFSET;
5487 else
5488 offset = DMA1_REGISTER_OFFSET;
5489 orig = data = RREG32(DMA_POWER_CNTL + offset);
5490 data |= MEM_POWER_OVERRIDE;
5491 if (data != orig)
5492 WREG32(DMA_POWER_CNTL + offset, data);
5493
5494 orig = data = RREG32(DMA_CLK_CTRL + offset);
5495 data = 0xff000000;
5496 if (data != orig)
5497 WREG32(DMA_CLK_CTRL + offset, data);
5498 }
5499 }
f8f84ac5
AD
5500}
5501
e16866ec
AD
5502static void si_enable_bif_mgls(struct radeon_device *rdev,
5503 bool enable)
f8f84ac5 5504{
e16866ec 5505 u32 orig, data;
f8f84ac5 5506
e16866ec 5507 orig = data = RREG32_PCIE(PCIE_CNTL2);
f8f84ac5 5508
e16866ec
AD
5509 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_BIF_LS))
5510 data |= SLV_MEM_LS_EN | MST_MEM_LS_EN |
5511 REPLAY_MEM_LS_EN | SLV_MEM_AGGRESSIVE_LS_EN;
5512 else
5513 data &= ~(SLV_MEM_LS_EN | MST_MEM_LS_EN |
5514 REPLAY_MEM_LS_EN | SLV_MEM_AGGRESSIVE_LS_EN);
5515
5516 if (orig != data)
5517 WREG32_PCIE(PCIE_CNTL2, data);
f8f84ac5
AD
5518}
5519
e16866ec
AD
5520static void si_enable_hdp_mgcg(struct radeon_device *rdev,
5521 bool enable)
f8f84ac5 5522{
e16866ec 5523 u32 orig, data;
f8f84ac5 5524
e16866ec 5525 orig = data = RREG32(HDP_HOST_PATH_CNTL);
f8f84ac5 5526
e16866ec
AD
5527 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_HDP_MGCG))
5528 data &= ~CLOCK_GATING_DIS;
5529 else
5530 data |= CLOCK_GATING_DIS;
5531
5532 if (orig != data)
5533 WREG32(HDP_HOST_PATH_CNTL, data);
f8f84ac5
AD
5534}
5535
e16866ec
AD
5536static void si_enable_hdp_ls(struct radeon_device *rdev,
5537 bool enable)
347e7592 5538{
e16866ec 5539 u32 orig, data;
347e7592 5540
e16866ec 5541 orig = data = RREG32(HDP_MEM_POWER_LS);
347e7592 5542
e16866ec
AD
5543 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_HDP_LS))
5544 data |= HDP_LS_ENABLE;
5545 else
5546 data &= ~HDP_LS_ENABLE;
347e7592 5547
e16866ec
AD
5548 if (orig != data)
5549 WREG32(HDP_MEM_POWER_LS, data);
347e7592
AD
5550}
5551
68e3a092
AD
5552static void si_update_cg(struct radeon_device *rdev,
5553 u32 block, bool enable)
347e7592 5554{
e16866ec 5555 if (block & RADEON_CG_BLOCK_GFX) {
811e4d58 5556 si_enable_gui_idle_interrupt(rdev, false);
e16866ec
AD
5557 /* order matters! */
5558 if (enable) {
5559 si_enable_mgcg(rdev, true);
5560 si_enable_cgcg(rdev, true);
5561 } else {
5562 si_enable_cgcg(rdev, false);
5563 si_enable_mgcg(rdev, false);
347e7592 5564 }
811e4d58 5565 si_enable_gui_idle_interrupt(rdev, true);
347e7592
AD
5566 }
5567
e16866ec
AD
5568 if (block & RADEON_CG_BLOCK_MC) {
5569 si_enable_mc_mgcg(rdev, enable);
5570 si_enable_mc_ls(rdev, enable);
347e7592 5571 }
e16866ec
AD
5572
5573 if (block & RADEON_CG_BLOCK_SDMA) {
5574 si_enable_dma_mgcg(rdev, enable);
347e7592
AD
5575 }
5576
e16866ec
AD
5577 if (block & RADEON_CG_BLOCK_BIF) {
5578 si_enable_bif_mgls(rdev, enable);
6d8cf000 5579 }
6d8cf000 5580
e16866ec
AD
5581 if (block & RADEON_CG_BLOCK_UVD) {
5582 if (rdev->has_uvd) {
5583 si_enable_uvd_mgcg(rdev, enable);
bd8cd539
AD
5584 }
5585 }
bd8cd539 5586
e16866ec
AD
5587 if (block & RADEON_CG_BLOCK_HDP) {
5588 si_enable_hdp_mgcg(rdev, enable);
5589 si_enable_hdp_ls(rdev, enable);
347e7592 5590 }
e16866ec 5591}
f8f84ac5
AD
5592
5593static void si_init_cg(struct radeon_device *rdev)
5594{
e16866ec
AD
5595 si_update_cg(rdev, (RADEON_CG_BLOCK_GFX |
5596 RADEON_CG_BLOCK_MC |
5597 RADEON_CG_BLOCK_SDMA |
5598 RADEON_CG_BLOCK_BIF |
5599 RADEON_CG_BLOCK_HDP), true);
b2d70917 5600 if (rdev->has_uvd) {
e16866ec 5601 si_update_cg(rdev, RADEON_CG_BLOCK_UVD, true);
f8f84ac5 5602 si_init_uvd_internal_cg(rdev);
347e7592 5603 }
f8f84ac5 5604}
bd8cd539 5605
f8f84ac5
AD
5606static void si_fini_cg(struct radeon_device *rdev)
5607{
0116e1ef 5608 if (rdev->has_uvd) {
e16866ec 5609 si_update_cg(rdev, RADEON_CG_BLOCK_UVD, false);
347e7592 5610 }
e16866ec
AD
5611 si_update_cg(rdev, (RADEON_CG_BLOCK_GFX |
5612 RADEON_CG_BLOCK_MC |
5613 RADEON_CG_BLOCK_SDMA |
5614 RADEON_CG_BLOCK_BIF |
5615 RADEON_CG_BLOCK_HDP), false);
5616}
5617
59a82d0e
AD
5618u32 si_get_csb_size(struct radeon_device *rdev)
5619{
5620 u32 count = 0;
5621 const struct cs_section_def *sect = NULL;
5622 const struct cs_extent_def *ext = NULL;
5623
5624 if (rdev->rlc.cs_data == NULL)
5625 return 0;
5626
5627 /* begin clear state */
5628 count += 2;
5629 /* context control state */
5630 count += 3;
5631
5632 for (sect = rdev->rlc.cs_data; sect->section != NULL; ++sect) {
5633 for (ext = sect->section; ext->extent != NULL; ++ext) {
5634 if (sect->id == SECT_CONTEXT)
5635 count += 2 + ext->reg_count;
5636 else
5637 return 0;
5638 }
bd8cd539 5639 }
59a82d0e
AD
5640 /* pa_sc_raster_config */
5641 count += 3;
5642 /* end clear state */
5643 count += 2;
5644 /* clear state */
5645 count += 2;
5646
5647 return count;
5648}
5649
5650void si_get_csb_buffer(struct radeon_device *rdev, volatile u32 *buffer)
5651{
5652 u32 count = 0, i;
5653 const struct cs_section_def *sect = NULL;
5654 const struct cs_extent_def *ext = NULL;
5655
5656 if (rdev->rlc.cs_data == NULL)
5657 return;
5658 if (buffer == NULL)
5659 return;
5660
6ba81e53
AD
5661 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0));
5662 buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
59a82d0e 5663
6ba81e53
AD
5664 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CONTEXT_CONTROL, 1));
5665 buffer[count++] = cpu_to_le32(0x80000000);
5666 buffer[count++] = cpu_to_le32(0x80000000);
59a82d0e
AD
5667
5668 for (sect = rdev->rlc.cs_data; sect->section != NULL; ++sect) {
5669 for (ext = sect->section; ext->extent != NULL; ++ext) {
5670 if (sect->id == SECT_CONTEXT) {
6ba81e53
AD
5671 buffer[count++] =
5672 cpu_to_le32(PACKET3(PACKET3_SET_CONTEXT_REG, ext->reg_count));
5673 buffer[count++] = cpu_to_le32(ext->reg_index - 0xa000);
59a82d0e 5674 for (i = 0; i < ext->reg_count; i++)
6ba81e53 5675 buffer[count++] = cpu_to_le32(ext->extent[i]);
59a82d0e
AD
5676 } else {
5677 return;
bd8cd539 5678 }
bd8cd539
AD
5679 }
5680 }
bd8cd539 5681
6ba81e53
AD
5682 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_SET_CONTEXT_REG, 1));
5683 buffer[count++] = cpu_to_le32(PA_SC_RASTER_CONFIG - PACKET3_SET_CONTEXT_REG_START);
59a82d0e
AD
5684 switch (rdev->family) {
5685 case CHIP_TAHITI:
5686 case CHIP_PITCAIRN:
6ba81e53 5687 buffer[count++] = cpu_to_le32(0x2a00126a);
59a82d0e
AD
5688 break;
5689 case CHIP_VERDE:
6ba81e53 5690 buffer[count++] = cpu_to_le32(0x0000124a);
59a82d0e
AD
5691 break;
5692 case CHIP_OLAND:
6ba81e53 5693 buffer[count++] = cpu_to_le32(0x00000082);
59a82d0e
AD
5694 break;
5695 case CHIP_HAINAN:
6ba81e53 5696 buffer[count++] = cpu_to_le32(0x00000000);
59a82d0e
AD
5697 break;
5698 default:
6ba81e53 5699 buffer[count++] = cpu_to_le32(0x00000000);
59a82d0e
AD
5700 break;
5701 }
5702
6ba81e53
AD
5703 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0));
5704 buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_END_CLEAR_STATE);
347e7592 5705
6ba81e53
AD
5706 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CLEAR_STATE, 0));
5707 buffer[count++] = cpu_to_le32(0);
59a82d0e
AD
5708}
5709
f8f84ac5
AD
5710static void si_init_pg(struct radeon_device *rdev)
5711{
0116e1ef
AD
5712 if (rdev->pg_flags) {
5713 if (rdev->pg_flags & RADEON_PG_SUPPORT_SDMA) {
5714 si_init_dma_pg(rdev);
0116e1ef 5715 }
f8f84ac5 5716 si_init_ao_cu_mask(rdev);
2b19d17f 5717 if (rdev->pg_flags & RADEON_PG_SUPPORT_GFX_PG) {
0116e1ef 5718 si_init_gfx_cgpg(rdev);
aa34dba8
AD
5719 } else {
5720 WREG32(RLC_SAVE_AND_RESTORE_BASE, rdev->rlc.save_restore_gpu_addr >> 8);
5721 WREG32(RLC_CLEAR_STATE_RESTORE_BASE, rdev->rlc.clear_state_gpu_addr >> 8);
0116e1ef 5722 }
ca6ebb39
AD
5723 si_enable_dma_pg(rdev, true);
5724 si_enable_gfx_cgpg(rdev, true);
f8f84ac5
AD
5725 } else {
5726 WREG32(RLC_SAVE_AND_RESTORE_BASE, rdev->rlc.save_restore_gpu_addr >> 8);
5727 WREG32(RLC_CLEAR_STATE_RESTORE_BASE, rdev->rlc.clear_state_gpu_addr >> 8);
5728 }
5729}
5730
5731static void si_fini_pg(struct radeon_device *rdev)
5732{
0116e1ef 5733 if (rdev->pg_flags) {
ca6ebb39
AD
5734 si_enable_dma_pg(rdev, false);
5735 si_enable_gfx_cgpg(rdev, false);
f8f84ac5 5736 }
347e7592
AD
5737}
5738
347e7592
AD
5739/*
5740 * RLC
5741 */
866d83de 5742void si_rlc_reset(struct radeon_device *rdev)
d719cef3 5743{
f8f84ac5 5744 u32 tmp = RREG32(GRBM_SOFT_RESET);
d719cef3 5745
f8f84ac5
AD
5746 tmp |= SOFT_RESET_RLC;
5747 WREG32(GRBM_SOFT_RESET, tmp);
5748 udelay(50);
5749 tmp &= ~SOFT_RESET_RLC;
5750 WREG32(GRBM_SOFT_RESET, tmp);
5751 udelay(50);
d719cef3
AD
5752}
5753
347e7592
AD
5754static void si_rlc_stop(struct radeon_device *rdev)
5755{
5756 WREG32(RLC_CNTL, 0);
d719cef3
AD
5757
5758 si_enable_gui_idle_interrupt(rdev, false);
5759
5760 si_wait_for_rlc_serdes(rdev);
347e7592
AD
5761}
5762
5763static void si_rlc_start(struct radeon_device *rdev)
5764{
5765 WREG32(RLC_CNTL, RLC_ENABLE);
d719cef3
AD
5766
5767 si_enable_gui_idle_interrupt(rdev, true);
5768
5769 udelay(50);
5770}
5771
5772static bool si_lbpw_supported(struct radeon_device *rdev)
5773{
5774 u32 tmp;
5775
5776 /* Enable LBPW only for DDR3 */
5777 tmp = RREG32(MC_SEQ_MISC0);
5778 if ((tmp & 0xF0000000) == 0xB0000000)
5779 return true;
5780 return false;
5781}
5782
5783static void si_enable_lbpw(struct radeon_device *rdev, bool enable)
5784{
5785 u32 tmp;
5786
5787 tmp = RREG32(RLC_LB_CNTL);
5788 if (enable)
5789 tmp |= LOAD_BALANCE_ENABLE;
5790 else
5791 tmp &= ~LOAD_BALANCE_ENABLE;
5792 WREG32(RLC_LB_CNTL, tmp);
5793
5794 if (!enable) {
5795 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
5796 WREG32(SPI_LB_CU_MASK, 0x00ff);
5797 }
347e7592
AD
5798}
5799
5800static int si_rlc_resume(struct radeon_device *rdev)
5801{
5802 u32 i;
347e7592
AD
5803
5804 if (!rdev->rlc_fw)
5805 return -EINVAL;
5806
5807 si_rlc_stop(rdev);
5808
f8f84ac5
AD
5809 si_rlc_reset(rdev);
5810
5811 si_init_pg(rdev);
5812
5813 si_init_cg(rdev);
5814
347e7592
AD
5815 WREG32(RLC_RL_BASE, 0);
5816 WREG32(RLC_RL_SIZE, 0);
5817 WREG32(RLC_LB_CNTL, 0);
5818 WREG32(RLC_LB_CNTR_MAX, 0xffffffff);
5819 WREG32(RLC_LB_CNTR_INIT, 0);
d719cef3 5820 WREG32(RLC_LB_INIT_CU_MASK, 0xffffffff);
347e7592 5821
347e7592
AD
5822 WREG32(RLC_MC_CNTL, 0);
5823 WREG32(RLC_UCODE_CNTL, 0);
5824
629bd33c
AD
5825 if (rdev->new_fw) {
5826 const struct rlc_firmware_header_v1_0 *hdr =
5827 (const struct rlc_firmware_header_v1_0 *)rdev->rlc_fw->data;
5828 u32 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
5829 const __le32 *fw_data = (const __le32 *)
5830 (rdev->rlc_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
5831
5832 radeon_ucode_print_rlc_hdr(&hdr->header);
5833
5834 for (i = 0; i < fw_size; i++) {
5835 WREG32(RLC_UCODE_ADDR, i);
5836 WREG32(RLC_UCODE_DATA, le32_to_cpup(fw_data++));
5837 }
5838 } else {
5839 const __be32 *fw_data =
5840 (const __be32 *)rdev->rlc_fw->data;
5841 for (i = 0; i < SI_RLC_UCODE_SIZE; i++) {
5842 WREG32(RLC_UCODE_ADDR, i);
5843 WREG32(RLC_UCODE_DATA, be32_to_cpup(fw_data++));
5844 }
347e7592
AD
5845 }
5846 WREG32(RLC_UCODE_ADDR, 0);
5847
d719cef3
AD
5848 si_enable_lbpw(rdev, si_lbpw_supported(rdev));
5849
347e7592
AD
5850 si_rlc_start(rdev);
5851
5852 return 0;
5853}
5854
25a857fb
AD
5855static void si_enable_interrupts(struct radeon_device *rdev)
5856{
5857 u32 ih_cntl = RREG32(IH_CNTL);
5858 u32 ih_rb_cntl = RREG32(IH_RB_CNTL);
5859
5860 ih_cntl |= ENABLE_INTR;
5861 ih_rb_cntl |= IH_RB_ENABLE;
5862 WREG32(IH_CNTL, ih_cntl);
5863 WREG32(IH_RB_CNTL, ih_rb_cntl);
5864 rdev->ih.enabled = true;
5865}
5866
5867static void si_disable_interrupts(struct radeon_device *rdev)
5868{
5869 u32 ih_rb_cntl = RREG32(IH_RB_CNTL);
5870 u32 ih_cntl = RREG32(IH_CNTL);
5871
5872 ih_rb_cntl &= ~IH_RB_ENABLE;
5873 ih_cntl &= ~ENABLE_INTR;
5874 WREG32(IH_RB_CNTL, ih_rb_cntl);
5875 WREG32(IH_CNTL, ih_cntl);
5876 /* set rptr, wptr to 0 */
5877 WREG32(IH_RB_RPTR, 0);
5878 WREG32(IH_RB_WPTR, 0);
5879 rdev->ih.enabled = false;
25a857fb
AD
5880 rdev->ih.rptr = 0;
5881}
5882
5883static void si_disable_interrupt_state(struct radeon_device *rdev)
5884{
5885 u32 tmp;
5886
811e4d58
AD
5887 tmp = RREG32(CP_INT_CNTL_RING0) &
5888 (CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
5889 WREG32(CP_INT_CNTL_RING0, tmp);
25a857fb
AD
5890 WREG32(CP_INT_CNTL_RING1, 0);
5891 WREG32(CP_INT_CNTL_RING2, 0);
8c5fd7ef
AD
5892 tmp = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET) & ~TRAP_ENABLE;
5893 WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, tmp);
5894 tmp = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET) & ~TRAP_ENABLE;
5895 WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, tmp);
25a857fb 5896 WREG32(GRBM_INT_CNTL, 0);
5153550a
AD
5897 if (rdev->num_crtc >= 2) {
5898 WREG32(INT_MASK + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
5899 WREG32(INT_MASK + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
5900 }
25a857fb
AD
5901 if (rdev->num_crtc >= 4) {
5902 WREG32(INT_MASK + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
5903 WREG32(INT_MASK + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
5904 }
5905 if (rdev->num_crtc >= 6) {
5906 WREG32(INT_MASK + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
5907 WREG32(INT_MASK + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
5908 }
5909
5153550a
AD
5910 if (rdev->num_crtc >= 2) {
5911 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
5912 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
5913 }
25a857fb
AD
5914 if (rdev->num_crtc >= 4) {
5915 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
5916 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
5917 }
5918 if (rdev->num_crtc >= 6) {
5919 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
5920 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
5921 }
5922
5153550a 5923 if (!ASIC_IS_NODCE(rdev)) {
e9a321c6 5924 WREG32(DAC_AUTODETECT_INT_CONTROL, 0);
5153550a
AD
5925
5926 tmp = RREG32(DC_HPD1_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5927 WREG32(DC_HPD1_INT_CONTROL, tmp);
5928 tmp = RREG32(DC_HPD2_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5929 WREG32(DC_HPD2_INT_CONTROL, tmp);
5930 tmp = RREG32(DC_HPD3_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5931 WREG32(DC_HPD3_INT_CONTROL, tmp);
5932 tmp = RREG32(DC_HPD4_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5933 WREG32(DC_HPD4_INT_CONTROL, tmp);
5934 tmp = RREG32(DC_HPD5_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5935 WREG32(DC_HPD5_INT_CONTROL, tmp);
5936 tmp = RREG32(DC_HPD6_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5937 WREG32(DC_HPD6_INT_CONTROL, tmp);
5938 }
25a857fb
AD
5939}
5940
5941static int si_irq_init(struct radeon_device *rdev)
5942{
5943 int ret = 0;
5944 int rb_bufsz;
5945 u32 interrupt_cntl, ih_cntl, ih_rb_cntl;
5946
5947 /* allocate ring */
5948 ret = r600_ih_ring_alloc(rdev);
5949 if (ret)
5950 return ret;
5951
5952 /* disable irqs */
5953 si_disable_interrupts(rdev);
5954
5955 /* init rlc */
5956 ret = si_rlc_resume(rdev);
5957 if (ret) {
5958 r600_ih_ring_fini(rdev);
5959 return ret;
5960 }
5961
5962 /* setup interrupt control */
5963 /* set dummy read address to ring address */
5964 WREG32(INTERRUPT_CNTL2, rdev->ih.gpu_addr >> 8);
5965 interrupt_cntl = RREG32(INTERRUPT_CNTL);
5966 /* IH_DUMMY_RD_OVERRIDE=0 - dummy read disabled with msi, enabled without msi
5967 * IH_DUMMY_RD_OVERRIDE=1 - dummy read controlled by IH_DUMMY_RD_EN
5968 */
5969 interrupt_cntl &= ~IH_DUMMY_RD_OVERRIDE;
5970 /* IH_REQ_NONSNOOP_EN=1 if ring is in non-cacheable memory, e.g., vram */
5971 interrupt_cntl &= ~IH_REQ_NONSNOOP_EN;
5972 WREG32(INTERRUPT_CNTL, interrupt_cntl);
5973
5974 WREG32(IH_RB_BASE, rdev->ih.gpu_addr >> 8);
b72a8925 5975 rb_bufsz = order_base_2(rdev->ih.ring_size / 4);
25a857fb
AD
5976
5977 ih_rb_cntl = (IH_WPTR_OVERFLOW_ENABLE |
5978 IH_WPTR_OVERFLOW_CLEAR |
5979 (rb_bufsz << 1));
5980
5981 if (rdev->wb.enabled)
5982 ih_rb_cntl |= IH_WPTR_WRITEBACK_ENABLE;
5983
5984 /* set the writeback address whether it's enabled or not */
5985 WREG32(IH_RB_WPTR_ADDR_LO, (rdev->wb.gpu_addr + R600_WB_IH_WPTR_OFFSET) & 0xFFFFFFFC);
5986 WREG32(IH_RB_WPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + R600_WB_IH_WPTR_OFFSET) & 0xFF);
5987
5988 WREG32(IH_RB_CNTL, ih_rb_cntl);
5989
5990 /* set rptr, wptr to 0 */
5991 WREG32(IH_RB_RPTR, 0);
5992 WREG32(IH_RB_WPTR, 0);
5993
5994 /* Default settings for IH_CNTL (disabled at first) */
5995 ih_cntl = MC_WRREQ_CREDIT(0x10) | MC_WR_CLEAN_CNT(0x10) | MC_VMID(0);
5996 /* RPTR_REARM only works if msi's are enabled */
5997 if (rdev->msi_enabled)
5998 ih_cntl |= RPTR_REARM;
5999 WREG32(IH_CNTL, ih_cntl);
6000
6001 /* force the active interrupt state to all disabled */
6002 si_disable_interrupt_state(rdev);
6003
2099810f
DA
6004 pci_set_master(rdev->pdev);
6005
25a857fb
AD
6006 /* enable irqs */
6007 si_enable_interrupts(rdev);
6008
6009 return ret;
6010}
6011
6012int si_irq_set(struct radeon_device *rdev)
6013{
811e4d58 6014 u32 cp_int_cntl;
25a857fb
AD
6015 u32 cp_int_cntl1 = 0, cp_int_cntl2 = 0;
6016 u32 crtc1 = 0, crtc2 = 0, crtc3 = 0, crtc4 = 0, crtc5 = 0, crtc6 = 0;
5153550a 6017 u32 hpd1 = 0, hpd2 = 0, hpd3 = 0, hpd4 = 0, hpd5 = 0, hpd6 = 0;
25a857fb 6018 u32 grbm_int_cntl = 0;
8c5fd7ef 6019 u32 dma_cntl, dma_cntl1;
a9e61410 6020 u32 thermal_int = 0;
25a857fb
AD
6021
6022 if (!rdev->irq.installed) {
6023 WARN(1, "Can't enable IRQ/MSI because no handler is installed\n");
6024 return -EINVAL;
6025 }
6026 /* don't enable anything if the ih is disabled */
6027 if (!rdev->ih.enabled) {
6028 si_disable_interrupts(rdev);
6029 /* force the active interrupt state to all disabled */
6030 si_disable_interrupt_state(rdev);
6031 return 0;
6032 }
6033
811e4d58
AD
6034 cp_int_cntl = RREG32(CP_INT_CNTL_RING0) &
6035 (CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
6036
5153550a
AD
6037 if (!ASIC_IS_NODCE(rdev)) {
6038 hpd1 = RREG32(DC_HPD1_INT_CONTROL) & ~DC_HPDx_INT_EN;
6039 hpd2 = RREG32(DC_HPD2_INT_CONTROL) & ~DC_HPDx_INT_EN;
6040 hpd3 = RREG32(DC_HPD3_INT_CONTROL) & ~DC_HPDx_INT_EN;
6041 hpd4 = RREG32(DC_HPD4_INT_CONTROL) & ~DC_HPDx_INT_EN;
6042 hpd5 = RREG32(DC_HPD5_INT_CONTROL) & ~DC_HPDx_INT_EN;
6043 hpd6 = RREG32(DC_HPD6_INT_CONTROL) & ~DC_HPDx_INT_EN;
6044 }
25a857fb 6045
8c5fd7ef
AD
6046 dma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET) & ~TRAP_ENABLE;
6047 dma_cntl1 = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET) & ~TRAP_ENABLE;
6048
a9e61410
AD
6049 thermal_int = RREG32(CG_THERMAL_INT) &
6050 ~(THERM_INT_MASK_HIGH | THERM_INT_MASK_LOW);
6051
25a857fb 6052 /* enable CP interrupts on all rings */
736fc37f 6053 if (atomic_read(&rdev->irq.ring_int[RADEON_RING_TYPE_GFX_INDEX])) {
25a857fb
AD
6054 DRM_DEBUG("si_irq_set: sw int gfx\n");
6055 cp_int_cntl |= TIME_STAMP_INT_ENABLE;
6056 }
736fc37f 6057 if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_CP1_INDEX])) {
25a857fb
AD
6058 DRM_DEBUG("si_irq_set: sw int cp1\n");
6059 cp_int_cntl1 |= TIME_STAMP_INT_ENABLE;
6060 }
736fc37f 6061 if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_CP2_INDEX])) {
25a857fb
AD
6062 DRM_DEBUG("si_irq_set: sw int cp2\n");
6063 cp_int_cntl2 |= TIME_STAMP_INT_ENABLE;
6064 }
8c5fd7ef
AD
6065 if (atomic_read(&rdev->irq.ring_int[R600_RING_TYPE_DMA_INDEX])) {
6066 DRM_DEBUG("si_irq_set: sw int dma\n");
6067 dma_cntl |= TRAP_ENABLE;
6068 }
6069
6070 if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_DMA1_INDEX])) {
6071 DRM_DEBUG("si_irq_set: sw int dma1\n");
6072 dma_cntl1 |= TRAP_ENABLE;
6073 }
25a857fb 6074 if (rdev->irq.crtc_vblank_int[0] ||
736fc37f 6075 atomic_read(&rdev->irq.pflip[0])) {
25a857fb
AD
6076 DRM_DEBUG("si_irq_set: vblank 0\n");
6077 crtc1 |= VBLANK_INT_MASK;
6078 }
6079 if (rdev->irq.crtc_vblank_int[1] ||
736fc37f 6080 atomic_read(&rdev->irq.pflip[1])) {
25a857fb
AD
6081 DRM_DEBUG("si_irq_set: vblank 1\n");
6082 crtc2 |= VBLANK_INT_MASK;
6083 }
6084 if (rdev->irq.crtc_vblank_int[2] ||
736fc37f 6085 atomic_read(&rdev->irq.pflip[2])) {
25a857fb
AD
6086 DRM_DEBUG("si_irq_set: vblank 2\n");
6087 crtc3 |= VBLANK_INT_MASK;
6088 }
6089 if (rdev->irq.crtc_vblank_int[3] ||
736fc37f 6090 atomic_read(&rdev->irq.pflip[3])) {
25a857fb
AD
6091 DRM_DEBUG("si_irq_set: vblank 3\n");
6092 crtc4 |= VBLANK_INT_MASK;
6093 }
6094 if (rdev->irq.crtc_vblank_int[4] ||
736fc37f 6095 atomic_read(&rdev->irq.pflip[4])) {
25a857fb
AD
6096 DRM_DEBUG("si_irq_set: vblank 4\n");
6097 crtc5 |= VBLANK_INT_MASK;
6098 }
6099 if (rdev->irq.crtc_vblank_int[5] ||
736fc37f 6100 atomic_read(&rdev->irq.pflip[5])) {
25a857fb
AD
6101 DRM_DEBUG("si_irq_set: vblank 5\n");
6102 crtc6 |= VBLANK_INT_MASK;
6103 }
6104 if (rdev->irq.hpd[0]) {
6105 DRM_DEBUG("si_irq_set: hpd 1\n");
6106 hpd1 |= DC_HPDx_INT_EN;
6107 }
6108 if (rdev->irq.hpd[1]) {
6109 DRM_DEBUG("si_irq_set: hpd 2\n");
6110 hpd2 |= DC_HPDx_INT_EN;
6111 }
6112 if (rdev->irq.hpd[2]) {
6113 DRM_DEBUG("si_irq_set: hpd 3\n");
6114 hpd3 |= DC_HPDx_INT_EN;
6115 }
6116 if (rdev->irq.hpd[3]) {
6117 DRM_DEBUG("si_irq_set: hpd 4\n");
6118 hpd4 |= DC_HPDx_INT_EN;
6119 }
6120 if (rdev->irq.hpd[4]) {
6121 DRM_DEBUG("si_irq_set: hpd 5\n");
6122 hpd5 |= DC_HPDx_INT_EN;
6123 }
6124 if (rdev->irq.hpd[5]) {
6125 DRM_DEBUG("si_irq_set: hpd 6\n");
6126 hpd6 |= DC_HPDx_INT_EN;
6127 }
25a857fb
AD
6128
6129 WREG32(CP_INT_CNTL_RING0, cp_int_cntl);
6130 WREG32(CP_INT_CNTL_RING1, cp_int_cntl1);
6131 WREG32(CP_INT_CNTL_RING2, cp_int_cntl2);
6132
8c5fd7ef
AD
6133 WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, dma_cntl);
6134 WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, dma_cntl1);
6135
25a857fb
AD
6136 WREG32(GRBM_INT_CNTL, grbm_int_cntl);
6137
a9e61410
AD
6138 if (rdev->irq.dpm_thermal) {
6139 DRM_DEBUG("dpm thermal\n");
6140 thermal_int |= THERM_INT_MASK_HIGH | THERM_INT_MASK_LOW;
6141 }
6142
5153550a
AD
6143 if (rdev->num_crtc >= 2) {
6144 WREG32(INT_MASK + EVERGREEN_CRTC0_REGISTER_OFFSET, crtc1);
6145 WREG32(INT_MASK + EVERGREEN_CRTC1_REGISTER_OFFSET, crtc2);
6146 }
25a857fb
AD
6147 if (rdev->num_crtc >= 4) {
6148 WREG32(INT_MASK + EVERGREEN_CRTC2_REGISTER_OFFSET, crtc3);
6149 WREG32(INT_MASK + EVERGREEN_CRTC3_REGISTER_OFFSET, crtc4);
6150 }
6151 if (rdev->num_crtc >= 6) {
6152 WREG32(INT_MASK + EVERGREEN_CRTC4_REGISTER_OFFSET, crtc5);
6153 WREG32(INT_MASK + EVERGREEN_CRTC5_REGISTER_OFFSET, crtc6);
6154 }
6155
5153550a 6156 if (rdev->num_crtc >= 2) {
f5d636d2
CK
6157 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET,
6158 GRPH_PFLIP_INT_MASK);
6159 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET,
6160 GRPH_PFLIP_INT_MASK);
5153550a 6161 }
25a857fb 6162 if (rdev->num_crtc >= 4) {
f5d636d2
CK
6163 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET,
6164 GRPH_PFLIP_INT_MASK);
6165 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET,
6166 GRPH_PFLIP_INT_MASK);
25a857fb
AD
6167 }
6168 if (rdev->num_crtc >= 6) {
f5d636d2
CK
6169 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET,
6170 GRPH_PFLIP_INT_MASK);
6171 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET,
6172 GRPH_PFLIP_INT_MASK);
25a857fb
AD
6173 }
6174
5153550a
AD
6175 if (!ASIC_IS_NODCE(rdev)) {
6176 WREG32(DC_HPD1_INT_CONTROL, hpd1);
6177 WREG32(DC_HPD2_INT_CONTROL, hpd2);
6178 WREG32(DC_HPD3_INT_CONTROL, hpd3);
6179 WREG32(DC_HPD4_INT_CONTROL, hpd4);
6180 WREG32(DC_HPD5_INT_CONTROL, hpd5);
6181 WREG32(DC_HPD6_INT_CONTROL, hpd6);
6182 }
25a857fb 6183
a9e61410
AD
6184 WREG32(CG_THERMAL_INT, thermal_int);
6185
25a857fb
AD
6186 return 0;
6187}
6188
6189static inline void si_irq_ack(struct radeon_device *rdev)
6190{
6191 u32 tmp;
6192
5153550a
AD
6193 if (ASIC_IS_NODCE(rdev))
6194 return;
6195
25a857fb
AD
6196 rdev->irq.stat_regs.evergreen.disp_int = RREG32(DISP_INTERRUPT_STATUS);
6197 rdev->irq.stat_regs.evergreen.disp_int_cont = RREG32(DISP_INTERRUPT_STATUS_CONTINUE);
6198 rdev->irq.stat_regs.evergreen.disp_int_cont2 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE2);
6199 rdev->irq.stat_regs.evergreen.disp_int_cont3 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE3);
6200 rdev->irq.stat_regs.evergreen.disp_int_cont4 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE4);
6201 rdev->irq.stat_regs.evergreen.disp_int_cont5 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE5);
6202 rdev->irq.stat_regs.evergreen.d1grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET);
6203 rdev->irq.stat_regs.evergreen.d2grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET);
6204 if (rdev->num_crtc >= 4) {
6205 rdev->irq.stat_regs.evergreen.d3grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET);
6206 rdev->irq.stat_regs.evergreen.d4grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET);
6207 }
6208 if (rdev->num_crtc >= 6) {
6209 rdev->irq.stat_regs.evergreen.d5grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET);
6210 rdev->irq.stat_regs.evergreen.d6grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET);
6211 }
6212
6213 if (rdev->irq.stat_regs.evergreen.d1grph_int & GRPH_PFLIP_INT_OCCURRED)
6214 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
6215 if (rdev->irq.stat_regs.evergreen.d2grph_int & GRPH_PFLIP_INT_OCCURRED)
6216 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
6217 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VBLANK_INTERRUPT)
6218 WREG32(VBLANK_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, VBLANK_ACK);
6219 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VLINE_INTERRUPT)
6220 WREG32(VLINE_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, VLINE_ACK);
6221 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VBLANK_INTERRUPT)
6222 WREG32(VBLANK_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, VBLANK_ACK);
6223 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VLINE_INTERRUPT)
6224 WREG32(VLINE_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, VLINE_ACK);
6225
6226 if (rdev->num_crtc >= 4) {
6227 if (rdev->irq.stat_regs.evergreen.d3grph_int & GRPH_PFLIP_INT_OCCURRED)
6228 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
6229 if (rdev->irq.stat_regs.evergreen.d4grph_int & GRPH_PFLIP_INT_OCCURRED)
6230 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
6231 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VBLANK_INTERRUPT)
6232 WREG32(VBLANK_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, VBLANK_ACK);
6233 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VLINE_INTERRUPT)
6234 WREG32(VLINE_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, VLINE_ACK);
6235 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VBLANK_INTERRUPT)
6236 WREG32(VBLANK_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, VBLANK_ACK);
6237 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VLINE_INTERRUPT)
6238 WREG32(VLINE_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, VLINE_ACK);
6239 }
6240
6241 if (rdev->num_crtc >= 6) {
6242 if (rdev->irq.stat_regs.evergreen.d5grph_int & GRPH_PFLIP_INT_OCCURRED)
6243 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
6244 if (rdev->irq.stat_regs.evergreen.d6grph_int & GRPH_PFLIP_INT_OCCURRED)
6245 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
6246 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VBLANK_INTERRUPT)
6247 WREG32(VBLANK_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, VBLANK_ACK);
6248 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VLINE_INTERRUPT)
6249 WREG32(VLINE_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, VLINE_ACK);
6250 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VBLANK_INTERRUPT)
6251 WREG32(VBLANK_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, VBLANK_ACK);
6252 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VLINE_INTERRUPT)
6253 WREG32(VLINE_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, VLINE_ACK);
6254 }
6255
6256 if (rdev->irq.stat_regs.evergreen.disp_int & DC_HPD1_INTERRUPT) {
6257 tmp = RREG32(DC_HPD1_INT_CONTROL);
6258 tmp |= DC_HPDx_INT_ACK;
6259 WREG32(DC_HPD1_INT_CONTROL, tmp);
6260 }
6261 if (rdev->irq.stat_regs.evergreen.disp_int_cont & DC_HPD2_INTERRUPT) {
6262 tmp = RREG32(DC_HPD2_INT_CONTROL);
6263 tmp |= DC_HPDx_INT_ACK;
6264 WREG32(DC_HPD2_INT_CONTROL, tmp);
6265 }
6266 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & DC_HPD3_INTERRUPT) {
6267 tmp = RREG32(DC_HPD3_INT_CONTROL);
6268 tmp |= DC_HPDx_INT_ACK;
6269 WREG32(DC_HPD3_INT_CONTROL, tmp);
6270 }
6271 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & DC_HPD4_INTERRUPT) {
6272 tmp = RREG32(DC_HPD4_INT_CONTROL);
6273 tmp |= DC_HPDx_INT_ACK;
6274 WREG32(DC_HPD4_INT_CONTROL, tmp);
6275 }
6276 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & DC_HPD5_INTERRUPT) {
6277 tmp = RREG32(DC_HPD5_INT_CONTROL);
6278 tmp |= DC_HPDx_INT_ACK;
6279 WREG32(DC_HPD5_INT_CONTROL, tmp);
6280 }
6281 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & DC_HPD6_INTERRUPT) {
6282 tmp = RREG32(DC_HPD5_INT_CONTROL);
6283 tmp |= DC_HPDx_INT_ACK;
6284 WREG32(DC_HPD6_INT_CONTROL, tmp);
6285 }
6286}
6287
6288static void si_irq_disable(struct radeon_device *rdev)
6289{
6290 si_disable_interrupts(rdev);
6291 /* Wait and acknowledge irq */
6292 mdelay(1);
6293 si_irq_ack(rdev);
6294 si_disable_interrupt_state(rdev);
6295}
6296
6297static void si_irq_suspend(struct radeon_device *rdev)
6298{
6299 si_irq_disable(rdev);
6300 si_rlc_stop(rdev);
6301}
6302
9b136d51
AD
6303static void si_irq_fini(struct radeon_device *rdev)
6304{
6305 si_irq_suspend(rdev);
6306 r600_ih_ring_fini(rdev);
6307}
6308
25a857fb
AD
6309static inline u32 si_get_ih_wptr(struct radeon_device *rdev)
6310{
6311 u32 wptr, tmp;
6312
6313 if (rdev->wb.enabled)
6314 wptr = le32_to_cpu(rdev->wb.wb[R600_WB_IH_WPTR_OFFSET/4]);
6315 else
6316 wptr = RREG32(IH_RB_WPTR);
6317
6318 if (wptr & RB_OVERFLOW) {
6319 /* When a ring buffer overflow happen start parsing interrupt
6320 * from the last not overwritten vector (wptr + 16). Hopefully
6321 * this should allow us to catchup.
6322 */
6323 dev_warn(rdev->dev, "IH ring buffer overflow (0x%08X, %d, %d)\n",
6324 wptr, rdev->ih.rptr, (wptr + 16) + rdev->ih.ptr_mask);
6325 rdev->ih.rptr = (wptr + 16) & rdev->ih.ptr_mask;
6326 tmp = RREG32(IH_RB_CNTL);
6327 tmp |= IH_WPTR_OVERFLOW_CLEAR;
6328 WREG32(IH_RB_CNTL, tmp);
e8c214d2 6329 wptr &= ~RB_OVERFLOW;
25a857fb
AD
6330 }
6331 return (wptr & rdev->ih.ptr_mask);
6332}
6333
6334/* SI IV Ring
6335 * Each IV ring entry is 128 bits:
6336 * [7:0] - interrupt source id
6337 * [31:8] - reserved
6338 * [59:32] - interrupt source data
6339 * [63:60] - reserved
6340 * [71:64] - RINGID
6341 * [79:72] - VMID
6342 * [127:80] - reserved
6343 */
6344int si_irq_process(struct radeon_device *rdev)
6345{
6346 u32 wptr;
6347 u32 rptr;
6348 u32 src_id, src_data, ring_id;
6349 u32 ring_index;
25a857fb 6350 bool queue_hotplug = false;
a9e61410 6351 bool queue_thermal = false;
fbf6dc7a 6352 u32 status, addr;
25a857fb
AD
6353
6354 if (!rdev->ih.enabled || rdev->shutdown)
6355 return IRQ_NONE;
6356
6357 wptr = si_get_ih_wptr(rdev);
c20dc369
CK
6358
6359restart_ih:
6360 /* is somebody else already processing irqs? */
6361 if (atomic_xchg(&rdev->ih.lock, 1))
6362 return IRQ_NONE;
6363
25a857fb
AD
6364 rptr = rdev->ih.rptr;
6365 DRM_DEBUG("si_irq_process start: rptr %d, wptr %d\n", rptr, wptr);
6366
25a857fb
AD
6367 /* Order reading of wptr vs. reading of IH ring data */
6368 rmb();
6369
6370 /* display interrupts */
6371 si_irq_ack(rdev);
6372
25a857fb
AD
6373 while (rptr != wptr) {
6374 /* wptr/rptr are in bytes! */
6375 ring_index = rptr / 4;
6376 src_id = le32_to_cpu(rdev->ih.ring[ring_index]) & 0xff;
6377 src_data = le32_to_cpu(rdev->ih.ring[ring_index + 1]) & 0xfffffff;
6378 ring_id = le32_to_cpu(rdev->ih.ring[ring_index + 2]) & 0xff;
6379
6380 switch (src_id) {
6381 case 1: /* D1 vblank/vline */
6382 switch (src_data) {
6383 case 0: /* D1 vblank */
6384 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VBLANK_INTERRUPT) {
6385 if (rdev->irq.crtc_vblank_int[0]) {
6386 drm_handle_vblank(rdev->ddev, 0);
6387 rdev->pm.vblank_sync = true;
6388 wake_up(&rdev->irq.vblank_queue);
6389 }
736fc37f 6390 if (atomic_read(&rdev->irq.pflip[0]))
1a0e7918 6391 radeon_crtc_handle_vblank(rdev, 0);
25a857fb
AD
6392 rdev->irq.stat_regs.evergreen.disp_int &= ~LB_D1_VBLANK_INTERRUPT;
6393 DRM_DEBUG("IH: D1 vblank\n");
6394 }
6395 break;
6396 case 1: /* D1 vline */
6397 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VLINE_INTERRUPT) {
6398 rdev->irq.stat_regs.evergreen.disp_int &= ~LB_D1_VLINE_INTERRUPT;
6399 DRM_DEBUG("IH: D1 vline\n");
6400 }
6401 break;
6402 default:
6403 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6404 break;
6405 }
6406 break;
6407 case 2: /* D2 vblank/vline */
6408 switch (src_data) {
6409 case 0: /* D2 vblank */
6410 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VBLANK_INTERRUPT) {
6411 if (rdev->irq.crtc_vblank_int[1]) {
6412 drm_handle_vblank(rdev->ddev, 1);
6413 rdev->pm.vblank_sync = true;
6414 wake_up(&rdev->irq.vblank_queue);
6415 }
736fc37f 6416 if (atomic_read(&rdev->irq.pflip[1]))
1a0e7918 6417 radeon_crtc_handle_vblank(rdev, 1);
25a857fb
AD
6418 rdev->irq.stat_regs.evergreen.disp_int_cont &= ~LB_D2_VBLANK_INTERRUPT;
6419 DRM_DEBUG("IH: D2 vblank\n");
6420 }
6421 break;
6422 case 1: /* D2 vline */
6423 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VLINE_INTERRUPT) {
6424 rdev->irq.stat_regs.evergreen.disp_int_cont &= ~LB_D2_VLINE_INTERRUPT;
6425 DRM_DEBUG("IH: D2 vline\n");
6426 }
6427 break;
6428 default:
6429 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6430 break;
6431 }
6432 break;
6433 case 3: /* D3 vblank/vline */
6434 switch (src_data) {
6435 case 0: /* D3 vblank */
6436 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VBLANK_INTERRUPT) {
6437 if (rdev->irq.crtc_vblank_int[2]) {
6438 drm_handle_vblank(rdev->ddev, 2);
6439 rdev->pm.vblank_sync = true;
6440 wake_up(&rdev->irq.vblank_queue);
6441 }
736fc37f 6442 if (atomic_read(&rdev->irq.pflip[2]))
1a0e7918 6443 radeon_crtc_handle_vblank(rdev, 2);
25a857fb
AD
6444 rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~LB_D3_VBLANK_INTERRUPT;
6445 DRM_DEBUG("IH: D3 vblank\n");
6446 }
6447 break;
6448 case 1: /* D3 vline */
6449 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VLINE_INTERRUPT) {
6450 rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~LB_D3_VLINE_INTERRUPT;
6451 DRM_DEBUG("IH: D3 vline\n");
6452 }
6453 break;
6454 default:
6455 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6456 break;
6457 }
6458 break;
6459 case 4: /* D4 vblank/vline */
6460 switch (src_data) {
6461 case 0: /* D4 vblank */
6462 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VBLANK_INTERRUPT) {
6463 if (rdev->irq.crtc_vblank_int[3]) {
6464 drm_handle_vblank(rdev->ddev, 3);
6465 rdev->pm.vblank_sync = true;
6466 wake_up(&rdev->irq.vblank_queue);
6467 }
736fc37f 6468 if (atomic_read(&rdev->irq.pflip[3]))
1a0e7918 6469 radeon_crtc_handle_vblank(rdev, 3);
25a857fb
AD
6470 rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~LB_D4_VBLANK_INTERRUPT;
6471 DRM_DEBUG("IH: D4 vblank\n");
6472 }
6473 break;
6474 case 1: /* D4 vline */
6475 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VLINE_INTERRUPT) {
6476 rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~LB_D4_VLINE_INTERRUPT;
6477 DRM_DEBUG("IH: D4 vline\n");
6478 }
6479 break;
6480 default:
6481 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6482 break;
6483 }
6484 break;
6485 case 5: /* D5 vblank/vline */
6486 switch (src_data) {
6487 case 0: /* D5 vblank */
6488 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VBLANK_INTERRUPT) {
6489 if (rdev->irq.crtc_vblank_int[4]) {
6490 drm_handle_vblank(rdev->ddev, 4);
6491 rdev->pm.vblank_sync = true;
6492 wake_up(&rdev->irq.vblank_queue);
6493 }
736fc37f 6494 if (atomic_read(&rdev->irq.pflip[4]))
1a0e7918 6495 radeon_crtc_handle_vblank(rdev, 4);
25a857fb
AD
6496 rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~LB_D5_VBLANK_INTERRUPT;
6497 DRM_DEBUG("IH: D5 vblank\n");
6498 }
6499 break;
6500 case 1: /* D5 vline */
6501 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VLINE_INTERRUPT) {
6502 rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~LB_D5_VLINE_INTERRUPT;
6503 DRM_DEBUG("IH: D5 vline\n");
6504 }
6505 break;
6506 default:
6507 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6508 break;
6509 }
6510 break;
6511 case 6: /* D6 vblank/vline */
6512 switch (src_data) {
6513 case 0: /* D6 vblank */
6514 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VBLANK_INTERRUPT) {
6515 if (rdev->irq.crtc_vblank_int[5]) {
6516 drm_handle_vblank(rdev->ddev, 5);
6517 rdev->pm.vblank_sync = true;
6518 wake_up(&rdev->irq.vblank_queue);
6519 }
736fc37f 6520 if (atomic_read(&rdev->irq.pflip[5]))
1a0e7918 6521 radeon_crtc_handle_vblank(rdev, 5);
25a857fb
AD
6522 rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~LB_D6_VBLANK_INTERRUPT;
6523 DRM_DEBUG("IH: D6 vblank\n");
6524 }
6525 break;
6526 case 1: /* D6 vline */
6527 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VLINE_INTERRUPT) {
6528 rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~LB_D6_VLINE_INTERRUPT;
6529 DRM_DEBUG("IH: D6 vline\n");
6530 }
6531 break;
6532 default:
6533 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6534 break;
6535 }
6536 break;
f5d636d2
CK
6537 case 8: /* D1 page flip */
6538 case 10: /* D2 page flip */
6539 case 12: /* D3 page flip */
6540 case 14: /* D4 page flip */
6541 case 16: /* D5 page flip */
6542 case 18: /* D6 page flip */
6543 DRM_DEBUG("IH: D%d flip\n", ((src_id - 8) >> 1) + 1);
39dc5454
MK
6544 if (radeon_use_pflipirq > 0)
6545 radeon_crtc_handle_flip(rdev, (src_id - 8) >> 1);
f5d636d2 6546 break;
25a857fb
AD
6547 case 42: /* HPD hotplug */
6548 switch (src_data) {
6549 case 0:
6550 if (rdev->irq.stat_regs.evergreen.disp_int & DC_HPD1_INTERRUPT) {
6551 rdev->irq.stat_regs.evergreen.disp_int &= ~DC_HPD1_INTERRUPT;
6552 queue_hotplug = true;
6553 DRM_DEBUG("IH: HPD1\n");
6554 }
6555 break;
6556 case 1:
6557 if (rdev->irq.stat_regs.evergreen.disp_int_cont & DC_HPD2_INTERRUPT) {
6558 rdev->irq.stat_regs.evergreen.disp_int_cont &= ~DC_HPD2_INTERRUPT;
6559 queue_hotplug = true;
6560 DRM_DEBUG("IH: HPD2\n");
6561 }
6562 break;
6563 case 2:
6564 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & DC_HPD3_INTERRUPT) {
6565 rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~DC_HPD3_INTERRUPT;
6566 queue_hotplug = true;
6567 DRM_DEBUG("IH: HPD3\n");
6568 }
6569 break;
6570 case 3:
6571 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & DC_HPD4_INTERRUPT) {
6572 rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~DC_HPD4_INTERRUPT;
6573 queue_hotplug = true;
6574 DRM_DEBUG("IH: HPD4\n");
6575 }
6576 break;
6577 case 4:
6578 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & DC_HPD5_INTERRUPT) {
6579 rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~DC_HPD5_INTERRUPT;
6580 queue_hotplug = true;
6581 DRM_DEBUG("IH: HPD5\n");
6582 }
6583 break;
6584 case 5:
6585 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & DC_HPD6_INTERRUPT) {
6586 rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~DC_HPD6_INTERRUPT;
6587 queue_hotplug = true;
6588 DRM_DEBUG("IH: HPD6\n");
6589 }
6590 break;
6591 default:
6592 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6593 break;
6594 }
6595 break;
b927e1c2
CK
6596 case 124: /* UVD */
6597 DRM_DEBUG("IH: UVD int: 0x%08x\n", src_data);
6598 radeon_fence_process(rdev, R600_RING_TYPE_UVD_INDEX);
6599 break;
ae133a11
CK
6600 case 146:
6601 case 147:
fbf6dc7a
AD
6602 addr = RREG32(VM_CONTEXT1_PROTECTION_FAULT_ADDR);
6603 status = RREG32(VM_CONTEXT1_PROTECTION_FAULT_STATUS);
9b7d786b
CK
6604 /* reset addr and status */
6605 WREG32_P(VM_CONTEXT1_CNTL2, 1, ~1);
6606 if (addr == 0x0 && status == 0x0)
6607 break;
ae133a11
CK
6608 dev_err(rdev->dev, "GPU fault detected: %d 0x%08x\n", src_id, src_data);
6609 dev_err(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n",
fbf6dc7a 6610 addr);
ae133a11 6611 dev_err(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
fbf6dc7a
AD
6612 status);
6613 si_vm_decode_fault(rdev, status, addr);
ae133a11 6614 break;
25a857fb
AD
6615 case 176: /* RINGID0 CP_INT */
6616 radeon_fence_process(rdev, RADEON_RING_TYPE_GFX_INDEX);
6617 break;
6618 case 177: /* RINGID1 CP_INT */
6619 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
6620 break;
6621 case 178: /* RINGID2 CP_INT */
6622 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
6623 break;
6624 case 181: /* CP EOP event */
6625 DRM_DEBUG("IH: CP EOP\n");
6626 switch (ring_id) {
6627 case 0:
6628 radeon_fence_process(rdev, RADEON_RING_TYPE_GFX_INDEX);
6629 break;
6630 case 1:
6631 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
6632 break;
6633 case 2:
6634 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
6635 break;
6636 }
6637 break;
8c5fd7ef
AD
6638 case 224: /* DMA trap event */
6639 DRM_DEBUG("IH: DMA trap\n");
6640 radeon_fence_process(rdev, R600_RING_TYPE_DMA_INDEX);
6641 break;
a9e61410
AD
6642 case 230: /* thermal low to high */
6643 DRM_DEBUG("IH: thermal low to high\n");
6644 rdev->pm.dpm.thermal.high_to_low = false;
6645 queue_thermal = true;
6646 break;
6647 case 231: /* thermal high to low */
6648 DRM_DEBUG("IH: thermal high to low\n");
6649 rdev->pm.dpm.thermal.high_to_low = true;
6650 queue_thermal = true;
6651 break;
25a857fb
AD
6652 case 233: /* GUI IDLE */
6653 DRM_DEBUG("IH: GUI idle\n");
25a857fb 6654 break;
8c5fd7ef
AD
6655 case 244: /* DMA trap event */
6656 DRM_DEBUG("IH: DMA1 trap\n");
6657 radeon_fence_process(rdev, CAYMAN_RING_TYPE_DMA1_INDEX);
6658 break;
25a857fb
AD
6659 default:
6660 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6661 break;
6662 }
6663
6664 /* wptr/rptr are in bytes! */
6665 rptr += 16;
6666 rptr &= rdev->ih.ptr_mask;
6667 }
25a857fb
AD
6668 if (queue_hotplug)
6669 schedule_work(&rdev->hotplug_work);
a9e61410
AD
6670 if (queue_thermal && rdev->pm.dpm_enabled)
6671 schedule_work(&rdev->pm.dpm.thermal.work);
25a857fb
AD
6672 rdev->ih.rptr = rptr;
6673 WREG32(IH_RB_RPTR, rdev->ih.rptr);
c20dc369
CK
6674 atomic_set(&rdev->ih.lock, 0);
6675
6676 /* make sure wptr hasn't changed while processing */
6677 wptr = si_get_ih_wptr(rdev);
6678 if (wptr != rptr)
6679 goto restart_ih;
6680
25a857fb
AD
6681 return IRQ_HANDLED;
6682}
6683
9b136d51
AD
6684/*
6685 * startup/shutdown callbacks
6686 */
6687static int si_startup(struct radeon_device *rdev)
6688{
6689 struct radeon_ring *ring;
6690 int r;
6691
b9d305df
AD
6692 /* enable pcie gen2/3 link */
6693 si_pcie_gen3_enable(rdev);
e0bcf165
AD
6694 /* enable aspm */
6695 si_program_aspm(rdev);
b9d305df 6696
e5903d39
AD
6697 /* scratch needs to be initialized before MC */
6698 r = r600_vram_scratch_init(rdev);
6699 if (r)
6700 return r;
6701
6fab3feb
AD
6702 si_mc_program(rdev);
6703
6c7bccea
AD
6704 if (!rdev->pm.dpm_enabled) {
6705 r = si_mc_load_microcode(rdev);
6706 if (r) {
6707 DRM_ERROR("Failed to load MC firmware!\n");
6708 return r;
6709 }
9b136d51
AD
6710 }
6711
9b136d51
AD
6712 r = si_pcie_gart_enable(rdev);
6713 if (r)
6714 return r;
6715 si_gpu_init(rdev);
6716
9b136d51 6717 /* allocate rlc buffers */
1fd11777
AD
6718 if (rdev->family == CHIP_VERDE) {
6719 rdev->rlc.reg_list = verde_rlc_save_restore_register_list;
6720 rdev->rlc.reg_list_size =
6721 (u32)ARRAY_SIZE(verde_rlc_save_restore_register_list);
6722 }
6723 rdev->rlc.cs_data = si_cs_data;
6724 r = sumo_rlc_init(rdev);
9b136d51
AD
6725 if (r) {
6726 DRM_ERROR("Failed to init rlc BOs!\n");
6727 return r;
6728 }
6729
6730 /* allocate wb buffer */
6731 r = radeon_wb_init(rdev);
6732 if (r)
6733 return r;
6734
6735 r = radeon_fence_driver_start_ring(rdev, RADEON_RING_TYPE_GFX_INDEX);
6736 if (r) {
6737 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
6738 return r;
6739 }
6740
6741 r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
6742 if (r) {
6743 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
6744 return r;
6745 }
6746
6747 r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
6748 if (r) {
6749 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
6750 return r;
6751 }
6752
8c5fd7ef
AD
6753 r = radeon_fence_driver_start_ring(rdev, R600_RING_TYPE_DMA_INDEX);
6754 if (r) {
6755 dev_err(rdev->dev, "failed initializing DMA fences (%d).\n", r);
6756 return r;
6757 }
6758
6759 r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_DMA1_INDEX);
6760 if (r) {
6761 dev_err(rdev->dev, "failed initializing DMA fences (%d).\n", r);
6762 return r;
6763 }
6764
1df0d523 6765 if (rdev->has_uvd) {
e409b128 6766 r = uvd_v2_2_resume(rdev);
1df0d523
AD
6767 if (!r) {
6768 r = radeon_fence_driver_start_ring(rdev,
6769 R600_RING_TYPE_UVD_INDEX);
6770 if (r)
6771 dev_err(rdev->dev, "UVD fences init error (%d).\n", r);
6772 }
f2ba57b5 6773 if (r)
1df0d523 6774 rdev->ring[R600_RING_TYPE_UVD_INDEX].ring_size = 0;
f2ba57b5 6775 }
f2ba57b5 6776
9b136d51 6777 /* Enable IRQ */
e49f3959
AH
6778 if (!rdev->irq.installed) {
6779 r = radeon_irq_kms_init(rdev);
6780 if (r)
6781 return r;
6782 }
6783
9b136d51
AD
6784 r = si_irq_init(rdev);
6785 if (r) {
6786 DRM_ERROR("radeon: IH init failed (%d).\n", r);
6787 radeon_irq_kms_fini(rdev);
6788 return r;
6789 }
6790 si_irq_set(rdev);
6791
6792 ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
6793 r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP_RPTR_OFFSET,
2e1e6dad 6794 RADEON_CP_PACKET2);
9b136d51
AD
6795 if (r)
6796 return r;
6797
6798 ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
6799 r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP1_RPTR_OFFSET,
2e1e6dad 6800 RADEON_CP_PACKET2);
9b136d51
AD
6801 if (r)
6802 return r;
6803
6804 ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
6805 r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP2_RPTR_OFFSET,
2e1e6dad 6806 RADEON_CP_PACKET2);
9b136d51
AD
6807 if (r)
6808 return r;
6809
8c5fd7ef
AD
6810 ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
6811 r = radeon_ring_init(rdev, ring, ring->ring_size, R600_WB_DMA_RPTR_OFFSET,
2e1e6dad 6812 DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0));
8c5fd7ef
AD
6813 if (r)
6814 return r;
6815
6816 ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
6817 r = radeon_ring_init(rdev, ring, ring->ring_size, CAYMAN_WB_DMA1_RPTR_OFFSET,
2e1e6dad 6818 DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0));
8c5fd7ef
AD
6819 if (r)
6820 return r;
6821
9b136d51
AD
6822 r = si_cp_load_microcode(rdev);
6823 if (r)
6824 return r;
6825 r = si_cp_resume(rdev);
6826 if (r)
6827 return r;
6828
8c5fd7ef
AD
6829 r = cayman_dma_resume(rdev);
6830 if (r)
6831 return r;
6832
1df0d523
AD
6833 if (rdev->has_uvd) {
6834 ring = &rdev->ring[R600_RING_TYPE_UVD_INDEX];
6835 if (ring->ring_size) {
02c9f7fa 6836 r = radeon_ring_init(rdev, ring, ring->ring_size, 0,
2e1e6dad 6837 RADEON_CP_PACKET2);
1df0d523 6838 if (!r)
e409b128 6839 r = uvd_v1_0_init(rdev);
1df0d523
AD
6840 if (r)
6841 DRM_ERROR("radeon: failed initializing UVD (%d).\n", r);
6842 }
f2ba57b5
CK
6843 }
6844
2898c348
CK
6845 r = radeon_ib_pool_init(rdev);
6846 if (r) {
6847 dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
9b136d51 6848 return r;
2898c348 6849 }
9b136d51 6850
c6105f24
CK
6851 r = radeon_vm_manager_init(rdev);
6852 if (r) {
6853 dev_err(rdev->dev, "vm manager initialization failed (%d).\n", r);
9b136d51 6854 return r;
c6105f24 6855 }
9b136d51 6856
b530602f
AD
6857 r = dce6_audio_init(rdev);
6858 if (r)
6859 return r;
6860
9b136d51
AD
6861 return 0;
6862}
6863
6864int si_resume(struct radeon_device *rdev)
6865{
6866 int r;
6867
6868 /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw,
6869 * posting will perform necessary task to bring back GPU into good
6870 * shape.
6871 */
6872 /* post card */
6873 atom_asic_init(rdev->mode_info.atom_context);
6874
205996c0
AD
6875 /* init golden registers */
6876 si_init_golden_registers(rdev);
6877
bc6a6295
AD
6878 if (rdev->pm.pm_method == PM_METHOD_DPM)
6879 radeon_pm_resume(rdev);
6c7bccea 6880
9b136d51
AD
6881 rdev->accel_working = true;
6882 r = si_startup(rdev);
6883 if (r) {
6884 DRM_ERROR("si startup failed on resume\n");
6885 rdev->accel_working = false;
6886 return r;
6887 }
6888
6889 return r;
6890
6891}
6892
6893int si_suspend(struct radeon_device *rdev)
6894{
6c7bccea 6895 radeon_pm_suspend(rdev);
b530602f 6896 dce6_audio_fini(rdev);
fa3daf9a 6897 radeon_vm_manager_fini(rdev);
9b136d51 6898 si_cp_enable(rdev, false);
8c5fd7ef 6899 cayman_dma_stop(rdev);
1df0d523 6900 if (rdev->has_uvd) {
e409b128 6901 uvd_v1_0_fini(rdev);
1df0d523
AD
6902 radeon_uvd_suspend(rdev);
6903 }
e16866ec
AD
6904 si_fini_pg(rdev);
6905 si_fini_cg(rdev);
9b136d51
AD
6906 si_irq_suspend(rdev);
6907 radeon_wb_disable(rdev);
6908 si_pcie_gart_disable(rdev);
6909 return 0;
6910}
6911
6912/* Plan is to move initialization in that function and use
6913 * helper function so that radeon_device_init pretty much
6914 * do nothing more than calling asic specific function. This
6915 * should also allow to remove a bunch of callback function
6916 * like vram_info.
6917 */
6918int si_init(struct radeon_device *rdev)
6919{
6920 struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
6921 int r;
6922
9b136d51
AD
6923 /* Read BIOS */
6924 if (!radeon_get_bios(rdev)) {
6925 if (ASIC_IS_AVIVO(rdev))
6926 return -EINVAL;
6927 }
6928 /* Must be an ATOMBIOS */
6929 if (!rdev->is_atom_bios) {
6930 dev_err(rdev->dev, "Expecting atombios for cayman GPU\n");
6931 return -EINVAL;
6932 }
6933 r = radeon_atombios_init(rdev);
6934 if (r)
6935 return r;
6936
6937 /* Post card if necessary */
6938 if (!radeon_card_posted(rdev)) {
6939 if (!rdev->bios) {
6940 dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n");
6941 return -EINVAL;
6942 }
6943 DRM_INFO("GPU not posted. posting now...\n");
6944 atom_asic_init(rdev->mode_info.atom_context);
6945 }
205996c0
AD
6946 /* init golden registers */
6947 si_init_golden_registers(rdev);
9b136d51
AD
6948 /* Initialize scratch registers */
6949 si_scratch_init(rdev);
6950 /* Initialize surface registers */
6951 radeon_surface_init(rdev);
6952 /* Initialize clocks */
6953 radeon_get_clock_info(rdev->ddev);
6954
6955 /* Fence driver */
6956 r = radeon_fence_driver_init(rdev);
6957 if (r)
6958 return r;
6959
6960 /* initialize memory controller */
6961 r = si_mc_init(rdev);
6962 if (r)
6963 return r;
6964 /* Memory manager */
6965 r = radeon_bo_init(rdev);
6966 if (r)
6967 return r;
6968
01ac8794
AD
6969 if (!rdev->me_fw || !rdev->pfp_fw || !rdev->ce_fw ||
6970 !rdev->rlc_fw || !rdev->mc_fw) {
6971 r = si_init_microcode(rdev);
6972 if (r) {
6973 DRM_ERROR("Failed to load firmware!\n");
6974 return r;
6975 }
6976 }
6977
6c7bccea
AD
6978 /* Initialize power management */
6979 radeon_pm_init(rdev);
6980
9b136d51
AD
6981 ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
6982 ring->ring_obj = NULL;
6983 r600_ring_init(rdev, ring, 1024 * 1024);
6984
6985 ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
6986 ring->ring_obj = NULL;
6987 r600_ring_init(rdev, ring, 1024 * 1024);
6988
6989 ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
6990 ring->ring_obj = NULL;
6991 r600_ring_init(rdev, ring, 1024 * 1024);
6992
8c5fd7ef
AD
6993 ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
6994 ring->ring_obj = NULL;
6995 r600_ring_init(rdev, ring, 64 * 1024);
6996
6997 ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
6998 ring->ring_obj = NULL;
6999 r600_ring_init(rdev, ring, 64 * 1024);
7000
1df0d523
AD
7001 if (rdev->has_uvd) {
7002 r = radeon_uvd_init(rdev);
7003 if (!r) {
7004 ring = &rdev->ring[R600_RING_TYPE_UVD_INDEX];
7005 ring->ring_obj = NULL;
7006 r600_ring_init(rdev, ring, 4096);
7007 }
f2ba57b5
CK
7008 }
7009
9b136d51
AD
7010 rdev->ih.ring_obj = NULL;
7011 r600_ih_ring_init(rdev, 64 * 1024);
7012
7013 r = r600_pcie_gart_init(rdev);
7014 if (r)
7015 return r;
7016
9b136d51 7017 rdev->accel_working = true;
9b136d51
AD
7018 r = si_startup(rdev);
7019 if (r) {
7020 dev_err(rdev->dev, "disabling GPU acceleration\n");
7021 si_cp_fini(rdev);
8c5fd7ef 7022 cayman_dma_fini(rdev);
9b136d51 7023 si_irq_fini(rdev);
1fd11777 7024 sumo_rlc_fini(rdev);
9b136d51 7025 radeon_wb_fini(rdev);
2898c348 7026 radeon_ib_pool_fini(rdev);
9b136d51
AD
7027 radeon_vm_manager_fini(rdev);
7028 radeon_irq_kms_fini(rdev);
7029 si_pcie_gart_fini(rdev);
7030 rdev->accel_working = false;
7031 }
7032
7033 /* Don't start up if the MC ucode is missing.
7034 * The default clocks and voltages before the MC ucode
7035 * is loaded are not suffient for advanced operations.
7036 */
7037 if (!rdev->mc_fw) {
7038 DRM_ERROR("radeon: MC ucode required for NI+.\n");
7039 return -EINVAL;
7040 }
7041
7042 return 0;
7043}
7044
7045void si_fini(struct radeon_device *rdev)
7046{
6c7bccea 7047 radeon_pm_fini(rdev);
9b136d51 7048 si_cp_fini(rdev);
8c5fd7ef 7049 cayman_dma_fini(rdev);
f8f84ac5 7050 si_fini_pg(rdev);
e16866ec 7051 si_fini_cg(rdev);
9b136d51 7052 si_irq_fini(rdev);
1fd11777 7053 sumo_rlc_fini(rdev);
9b136d51
AD
7054 radeon_wb_fini(rdev);
7055 radeon_vm_manager_fini(rdev);
2898c348 7056 radeon_ib_pool_fini(rdev);
9b136d51 7057 radeon_irq_kms_fini(rdev);
2858c00d 7058 if (rdev->has_uvd) {
e409b128 7059 uvd_v1_0_fini(rdev);
1df0d523 7060 radeon_uvd_fini(rdev);
2858c00d 7061 }
9b136d51
AD
7062 si_pcie_gart_fini(rdev);
7063 r600_vram_scratch_fini(rdev);
7064 radeon_gem_fini(rdev);
9b136d51
AD
7065 radeon_fence_driver_fini(rdev);
7066 radeon_bo_fini(rdev);
7067 radeon_atombios_fini(rdev);
7068 kfree(rdev->bios);
7069 rdev->bios = NULL;
7070}
7071
6759a0a7 7072/**
d0418894 7073 * si_get_gpu_clock_counter - return GPU clock counter snapshot
6759a0a7
MO
7074 *
7075 * @rdev: radeon_device pointer
7076 *
7077 * Fetches a GPU clock counter snapshot (SI).
7078 * Returns the 64 bit clock counter snapshot.
7079 */
d0418894 7080uint64_t si_get_gpu_clock_counter(struct radeon_device *rdev)
6759a0a7
MO
7081{
7082 uint64_t clock;
7083
7084 mutex_lock(&rdev->gpu_clock_mutex);
7085 WREG32(RLC_CAPTURE_GPU_CLOCK_COUNT, 1);
7086 clock = (uint64_t)RREG32(RLC_GPU_CLOCK_COUNT_LSB) |
7087 ((uint64_t)RREG32(RLC_GPU_CLOCK_COUNT_MSB) << 32ULL);
7088 mutex_unlock(&rdev->gpu_clock_mutex);
7089 return clock;
7090}
2539eb02 7091
2539eb02
CK
7092int si_set_uvd_clocks(struct radeon_device *rdev, u32 vclk, u32 dclk)
7093{
facd112d 7094 unsigned fb_div = 0, vclk_div = 0, dclk_div = 0;
2539eb02
CK
7095 int r;
7096
4ed10835
CK
7097 /* bypass vclk and dclk with bclk */
7098 WREG32_P(CG_UPLL_FUNC_CNTL_2,
7099 VCLK_SRC_SEL(1) | DCLK_SRC_SEL(1),
7100 ~(VCLK_SRC_SEL_MASK | DCLK_SRC_SEL_MASK));
7101
7102 /* put PLL in bypass mode */
7103 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_BYPASS_EN_MASK, ~UPLL_BYPASS_EN_MASK);
7104
7105 if (!vclk || !dclk) {
7106 /* keep the Bypass mode, put PLL to sleep */
7107 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_SLEEP_MASK, ~UPLL_SLEEP_MASK);
7108 return 0;
7109 }
7110
facd112d
CK
7111 r = radeon_uvd_calc_upll_dividers(rdev, vclk, dclk, 125000, 250000,
7112 16384, 0x03FFFFFF, 0, 128, 5,
7113 &fb_div, &vclk_div, &dclk_div);
7114 if (r)
7115 return r;
2539eb02
CK
7116
7117 /* set RESET_ANTI_MUX to 0 */
7118 WREG32_P(CG_UPLL_FUNC_CNTL_5, 0, ~RESET_ANTI_MUX_MASK);
7119
7120 /* set VCO_MODE to 1 */
7121 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_VCO_MODE_MASK, ~UPLL_VCO_MODE_MASK);
7122
7123 /* toggle UPLL_SLEEP to 1 then back to 0 */
7124 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_SLEEP_MASK, ~UPLL_SLEEP_MASK);
7125 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_SLEEP_MASK);
7126
7127 /* deassert UPLL_RESET */
7128 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_RESET_MASK);
7129
7130 mdelay(1);
7131
facd112d 7132 r = radeon_uvd_send_upll_ctlreq(rdev, CG_UPLL_FUNC_CNTL);
2539eb02
CK
7133 if (r)
7134 return r;
7135
7136 /* assert UPLL_RESET again */
7137 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_RESET_MASK, ~UPLL_RESET_MASK);
7138
7139 /* disable spread spectrum. */
7140 WREG32_P(CG_UPLL_SPREAD_SPECTRUM, 0, ~SSEN_MASK);
7141
7142 /* set feedback divider */
facd112d 7143 WREG32_P(CG_UPLL_FUNC_CNTL_3, UPLL_FB_DIV(fb_div), ~UPLL_FB_DIV_MASK);
2539eb02
CK
7144
7145 /* set ref divider to 0 */
7146 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_REF_DIV_MASK);
7147
facd112d 7148 if (fb_div < 307200)
2539eb02
CK
7149 WREG32_P(CG_UPLL_FUNC_CNTL_4, 0, ~UPLL_SPARE_ISPARE9);
7150 else
7151 WREG32_P(CG_UPLL_FUNC_CNTL_4, UPLL_SPARE_ISPARE9, ~UPLL_SPARE_ISPARE9);
7152
7153 /* set PDIV_A and PDIV_B */
7154 WREG32_P(CG_UPLL_FUNC_CNTL_2,
facd112d 7155 UPLL_PDIV_A(vclk_div) | UPLL_PDIV_B(dclk_div),
2539eb02
CK
7156 ~(UPLL_PDIV_A_MASK | UPLL_PDIV_B_MASK));
7157
7158 /* give the PLL some time to settle */
7159 mdelay(15);
7160
7161 /* deassert PLL_RESET */
7162 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_RESET_MASK);
7163
7164 mdelay(15);
7165
7166 /* switch from bypass mode to normal mode */
7167 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_BYPASS_EN_MASK);
7168
facd112d 7169 r = radeon_uvd_send_upll_ctlreq(rdev, CG_UPLL_FUNC_CNTL);
2539eb02
CK
7170 if (r)
7171 return r;
7172
7173 /* switch VCLK and DCLK selection */
7174 WREG32_P(CG_UPLL_FUNC_CNTL_2,
7175 VCLK_SRC_SEL(2) | DCLK_SRC_SEL(2),
7176 ~(VCLK_SRC_SEL_MASK | DCLK_SRC_SEL_MASK));
7177
7178 mdelay(100);
7179
7180 return 0;
7181}
b9d305df
AD
7182
7183static void si_pcie_gen3_enable(struct radeon_device *rdev)
7184{
7185 struct pci_dev *root = rdev->pdev->bus->self;
7186 int bridge_pos, gpu_pos;
7187 u32 speed_cntl, mask, current_data_rate;
7188 int ret, i;
7189 u16 tmp16;
7190
0bd252de
AW
7191 if (pci_is_root_bus(rdev->pdev->bus))
7192 return;
7193
b9d305df
AD
7194 if (radeon_pcie_gen2 == 0)
7195 return;
7196
7197 if (rdev->flags & RADEON_IS_IGP)
7198 return;
7199
7200 if (!(rdev->flags & RADEON_IS_PCIE))
7201 return;
7202
7203 ret = drm_pcie_get_speed_cap_mask(rdev->ddev, &mask);
7204 if (ret != 0)
7205 return;
7206
7207 if (!(mask & (DRM_PCIE_SPEED_50 | DRM_PCIE_SPEED_80)))
7208 return;
7209
7210 speed_cntl = RREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL);
7211 current_data_rate = (speed_cntl & LC_CURRENT_DATA_RATE_MASK) >>
7212 LC_CURRENT_DATA_RATE_SHIFT;
7213 if (mask & DRM_PCIE_SPEED_80) {
7214 if (current_data_rate == 2) {
7215 DRM_INFO("PCIE gen 3 link speeds already enabled\n");
7216 return;
7217 }
7218 DRM_INFO("enabling PCIE gen 3 link speeds, disable with radeon.pcie_gen2=0\n");
7219 } else if (mask & DRM_PCIE_SPEED_50) {
7220 if (current_data_rate == 1) {
7221 DRM_INFO("PCIE gen 2 link speeds already enabled\n");
7222 return;
7223 }
7224 DRM_INFO("enabling PCIE gen 2 link speeds, disable with radeon.pcie_gen2=0\n");
7225 }
7226
7227 bridge_pos = pci_pcie_cap(root);
7228 if (!bridge_pos)
7229 return;
7230
7231 gpu_pos = pci_pcie_cap(rdev->pdev);
7232 if (!gpu_pos)
7233 return;
7234
7235 if (mask & DRM_PCIE_SPEED_80) {
7236 /* re-try equalization if gen3 is not already enabled */
7237 if (current_data_rate != 2) {
7238 u16 bridge_cfg, gpu_cfg;
7239 u16 bridge_cfg2, gpu_cfg2;
7240 u32 max_lw, current_lw, tmp;
7241
7242 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL, &bridge_cfg);
7243 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, &gpu_cfg);
7244
7245 tmp16 = bridge_cfg | PCI_EXP_LNKCTL_HAWD;
7246 pci_write_config_word(root, bridge_pos + PCI_EXP_LNKCTL, tmp16);
7247
7248 tmp16 = gpu_cfg | PCI_EXP_LNKCTL_HAWD;
7249 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, tmp16);
7250
7251 tmp = RREG32_PCIE(PCIE_LC_STATUS1);
7252 max_lw = (tmp & LC_DETECTED_LINK_WIDTH_MASK) >> LC_DETECTED_LINK_WIDTH_SHIFT;
7253 current_lw = (tmp & LC_OPERATING_LINK_WIDTH_MASK) >> LC_OPERATING_LINK_WIDTH_SHIFT;
7254
7255 if (current_lw < max_lw) {
7256 tmp = RREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL);
7257 if (tmp & LC_RENEGOTIATION_SUPPORT) {
7258 tmp &= ~(LC_LINK_WIDTH_MASK | LC_UPCONFIGURE_DIS);
7259 tmp |= (max_lw << LC_LINK_WIDTH_SHIFT);
7260 tmp |= LC_UPCONFIGURE_SUPPORT | LC_RENEGOTIATE_EN | LC_RECONFIG_NOW;
7261 WREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL, tmp);
7262 }
7263 }
7264
7265 for (i = 0; i < 10; i++) {
7266 /* check status */
7267 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_DEVSTA, &tmp16);
7268 if (tmp16 & PCI_EXP_DEVSTA_TRPND)
7269 break;
7270
7271 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL, &bridge_cfg);
7272 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, &gpu_cfg);
7273
7274 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL2, &bridge_cfg2);
7275 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, &gpu_cfg2);
7276
7277 tmp = RREG32_PCIE_PORT(PCIE_LC_CNTL4);
7278 tmp |= LC_SET_QUIESCE;
7279 WREG32_PCIE_PORT(PCIE_LC_CNTL4, tmp);
7280
7281 tmp = RREG32_PCIE_PORT(PCIE_LC_CNTL4);
7282 tmp |= LC_REDO_EQ;
7283 WREG32_PCIE_PORT(PCIE_LC_CNTL4, tmp);
7284
7285 mdelay(100);
7286
7287 /* linkctl */
7288 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL, &tmp16);
7289 tmp16 &= ~PCI_EXP_LNKCTL_HAWD;
7290 tmp16 |= (bridge_cfg & PCI_EXP_LNKCTL_HAWD);
7291 pci_write_config_word(root, bridge_pos + PCI_EXP_LNKCTL, tmp16);
7292
7293 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, &tmp16);
7294 tmp16 &= ~PCI_EXP_LNKCTL_HAWD;
7295 tmp16 |= (gpu_cfg & PCI_EXP_LNKCTL_HAWD);
7296 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, tmp16);
7297
7298 /* linkctl2 */
7299 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL2, &tmp16);
7300 tmp16 &= ~((1 << 4) | (7 << 9));
7301 tmp16 |= (bridge_cfg2 & ((1 << 4) | (7 << 9)));
7302 pci_write_config_word(root, bridge_pos + PCI_EXP_LNKCTL2, tmp16);
7303
7304 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, &tmp16);
7305 tmp16 &= ~((1 << 4) | (7 << 9));
7306 tmp16 |= (gpu_cfg2 & ((1 << 4) | (7 << 9)));
7307 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, tmp16);
7308
7309 tmp = RREG32_PCIE_PORT(PCIE_LC_CNTL4);
7310 tmp &= ~LC_SET_QUIESCE;
7311 WREG32_PCIE_PORT(PCIE_LC_CNTL4, tmp);
7312 }
7313 }
7314 }
7315
7316 /* set the link speed */
7317 speed_cntl |= LC_FORCE_EN_SW_SPEED_CHANGE | LC_FORCE_DIS_HW_SPEED_CHANGE;
7318 speed_cntl &= ~LC_FORCE_DIS_SW_SPEED_CHANGE;
7319 WREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL, speed_cntl);
7320
7321 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, &tmp16);
7322 tmp16 &= ~0xf;
7323 if (mask & DRM_PCIE_SPEED_80)
7324 tmp16 |= 3; /* gen3 */
7325 else if (mask & DRM_PCIE_SPEED_50)
7326 tmp16 |= 2; /* gen2 */
7327 else
7328 tmp16 |= 1; /* gen1 */
7329 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, tmp16);
7330
7331 speed_cntl = RREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL);
7332 speed_cntl |= LC_INITIATE_LINK_SPEED_CHANGE;
7333 WREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL, speed_cntl);
7334
7335 for (i = 0; i < rdev->usec_timeout; i++) {
7336 speed_cntl = RREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL);
7337 if ((speed_cntl & LC_INITIATE_LINK_SPEED_CHANGE) == 0)
7338 break;
7339 udelay(1);
7340 }
7341}
7342
e0bcf165
AD
7343static void si_program_aspm(struct radeon_device *rdev)
7344{
7345 u32 data, orig;
7346 bool disable_l0s = false, disable_l1 = false, disable_plloff_in_l1 = false;
7347 bool disable_clkreq = false;
7348
1294d4a3
AD
7349 if (radeon_aspm == 0)
7350 return;
7351
e0bcf165
AD
7352 if (!(rdev->flags & RADEON_IS_PCIE))
7353 return;
7354
7355 orig = data = RREG32_PCIE_PORT(PCIE_LC_N_FTS_CNTL);
7356 data &= ~LC_XMIT_N_FTS_MASK;
7357 data |= LC_XMIT_N_FTS(0x24) | LC_XMIT_N_FTS_OVERRIDE_EN;
7358 if (orig != data)
7359 WREG32_PCIE_PORT(PCIE_LC_N_FTS_CNTL, data);
7360
7361 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL3);
7362 data |= LC_GO_TO_RECOVERY;
7363 if (orig != data)
7364 WREG32_PCIE_PORT(PCIE_LC_CNTL3, data);
7365
7366 orig = data = RREG32_PCIE(PCIE_P_CNTL);
7367 data |= P_IGNORE_EDB_ERR;
7368 if (orig != data)
7369 WREG32_PCIE(PCIE_P_CNTL, data);
7370
7371 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL);
7372 data &= ~(LC_L0S_INACTIVITY_MASK | LC_L1_INACTIVITY_MASK);
7373 data |= LC_PMI_TO_L1_DIS;
7374 if (!disable_l0s)
7375 data |= LC_L0S_INACTIVITY(7);
7376
7377 if (!disable_l1) {
7378 data |= LC_L1_INACTIVITY(7);
7379 data &= ~LC_PMI_TO_L1_DIS;
7380 if (orig != data)
7381 WREG32_PCIE_PORT(PCIE_LC_CNTL, data);
7382
7383 if (!disable_plloff_in_l1) {
7384 bool clk_req_support;
7385
7386 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_0);
7387 data &= ~(PLL_POWER_STATE_IN_OFF_0_MASK | PLL_POWER_STATE_IN_TXS2_0_MASK);
7388 data |= PLL_POWER_STATE_IN_OFF_0(7) | PLL_POWER_STATE_IN_TXS2_0(7);
7389 if (orig != data)
7390 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_0, data);
7391
7392 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_1);
7393 data &= ~(PLL_POWER_STATE_IN_OFF_1_MASK | PLL_POWER_STATE_IN_TXS2_1_MASK);
7394 data |= PLL_POWER_STATE_IN_OFF_1(7) | PLL_POWER_STATE_IN_TXS2_1(7);
7395 if (orig != data)
7396 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_1, data);
7397
7398 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_0);
7399 data &= ~(PLL_POWER_STATE_IN_OFF_0_MASK | PLL_POWER_STATE_IN_TXS2_0_MASK);
7400 data |= PLL_POWER_STATE_IN_OFF_0(7) | PLL_POWER_STATE_IN_TXS2_0(7);
7401 if (orig != data)
7402 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_0, data);
7403
7404 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_1);
7405 data &= ~(PLL_POWER_STATE_IN_OFF_1_MASK | PLL_POWER_STATE_IN_TXS2_1_MASK);
7406 data |= PLL_POWER_STATE_IN_OFF_1(7) | PLL_POWER_STATE_IN_TXS2_1(7);
7407 if (orig != data)
7408 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_1, data);
7409
7410 if ((rdev->family != CHIP_OLAND) && (rdev->family != CHIP_HAINAN)) {
7411 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_0);
7412 data &= ~PLL_RAMP_UP_TIME_0_MASK;
7413 if (orig != data)
7414 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_0, data);
7415
7416 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_1);
7417 data &= ~PLL_RAMP_UP_TIME_1_MASK;
7418 if (orig != data)
7419 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_1, data);
7420
7421 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_2);
7422 data &= ~PLL_RAMP_UP_TIME_2_MASK;
7423 if (orig != data)
7424 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_2, data);
7425
7426 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_3);
7427 data &= ~PLL_RAMP_UP_TIME_3_MASK;
7428 if (orig != data)
7429 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_3, data);
7430
7431 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_0);
7432 data &= ~PLL_RAMP_UP_TIME_0_MASK;
7433 if (orig != data)
7434 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_0, data);
7435
7436 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_1);
7437 data &= ~PLL_RAMP_UP_TIME_1_MASK;
7438 if (orig != data)
7439 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_1, data);
7440
7441 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_2);
7442 data &= ~PLL_RAMP_UP_TIME_2_MASK;
7443 if (orig != data)
7444 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_2, data);
7445
7446 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_3);
7447 data &= ~PLL_RAMP_UP_TIME_3_MASK;
7448 if (orig != data)
7449 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_3, data);
7450 }
7451 orig = data = RREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL);
7452 data &= ~LC_DYN_LANES_PWR_STATE_MASK;
7453 data |= LC_DYN_LANES_PWR_STATE(3);
7454 if (orig != data)
7455 WREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL, data);
7456
7457 orig = data = RREG32_PIF_PHY0(PB0_PIF_CNTL);
7458 data &= ~LS2_EXIT_TIME_MASK;
7459 if ((rdev->family == CHIP_OLAND) || (rdev->family == CHIP_HAINAN))
7460 data |= LS2_EXIT_TIME(5);
7461 if (orig != data)
7462 WREG32_PIF_PHY0(PB0_PIF_CNTL, data);
7463
7464 orig = data = RREG32_PIF_PHY1(PB1_PIF_CNTL);
7465 data &= ~LS2_EXIT_TIME_MASK;
7466 if ((rdev->family == CHIP_OLAND) || (rdev->family == CHIP_HAINAN))
7467 data |= LS2_EXIT_TIME(5);
7468 if (orig != data)
7469 WREG32_PIF_PHY1(PB1_PIF_CNTL, data);
7470
0bd252de
AW
7471 if (!disable_clkreq &&
7472 !pci_is_root_bus(rdev->pdev->bus)) {
e0bcf165
AD
7473 struct pci_dev *root = rdev->pdev->bus->self;
7474 u32 lnkcap;
7475
7476 clk_req_support = false;
7477 pcie_capability_read_dword(root, PCI_EXP_LNKCAP, &lnkcap);
7478 if (lnkcap & PCI_EXP_LNKCAP_CLKPM)
7479 clk_req_support = true;
7480 } else {
7481 clk_req_support = false;
7482 }
7483
7484 if (clk_req_support) {
7485 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL2);
7486 data |= LC_ALLOW_PDWN_IN_L1 | LC_ALLOW_PDWN_IN_L23;
7487 if (orig != data)
7488 WREG32_PCIE_PORT(PCIE_LC_CNTL2, data);
7489
7490 orig = data = RREG32(THM_CLK_CNTL);
7491 data &= ~(CMON_CLK_SEL_MASK | TMON_CLK_SEL_MASK);
7492 data |= CMON_CLK_SEL(1) | TMON_CLK_SEL(1);
7493 if (orig != data)
7494 WREG32(THM_CLK_CNTL, data);
7495
7496 orig = data = RREG32(MISC_CLK_CNTL);
7497 data &= ~(DEEP_SLEEP_CLK_SEL_MASK | ZCLK_SEL_MASK);
7498 data |= DEEP_SLEEP_CLK_SEL(1) | ZCLK_SEL(1);
7499 if (orig != data)
7500 WREG32(MISC_CLK_CNTL, data);
7501
7502 orig = data = RREG32(CG_CLKPIN_CNTL);
7503 data &= ~BCLK_AS_XCLK;
7504 if (orig != data)
7505 WREG32(CG_CLKPIN_CNTL, data);
7506
7507 orig = data = RREG32(CG_CLKPIN_CNTL_2);
7508 data &= ~FORCE_BIF_REFCLK_EN;
7509 if (orig != data)
7510 WREG32(CG_CLKPIN_CNTL_2, data);
7511
7512 orig = data = RREG32(MPLL_BYPASSCLK_SEL);
7513 data &= ~MPLL_CLKOUT_SEL_MASK;
7514 data |= MPLL_CLKOUT_SEL(4);
7515 if (orig != data)
7516 WREG32(MPLL_BYPASSCLK_SEL, data);
7517
7518 orig = data = RREG32(SPLL_CNTL_MODE);
7519 data &= ~SPLL_REFCLK_SEL_MASK;
7520 if (orig != data)
7521 WREG32(SPLL_CNTL_MODE, data);
7522 }
7523 }
7524 } else {
7525 if (orig != data)
7526 WREG32_PCIE_PORT(PCIE_LC_CNTL, data);
7527 }
7528
7529 orig = data = RREG32_PCIE(PCIE_CNTL2);
7530 data |= SLV_MEM_LS_EN | MST_MEM_LS_EN | REPLAY_MEM_LS_EN;
7531 if (orig != data)
7532 WREG32_PCIE(PCIE_CNTL2, data);
7533
7534 if (!disable_l0s) {
7535 data = RREG32_PCIE_PORT(PCIE_LC_N_FTS_CNTL);
7536 if((data & LC_N_FTS_MASK) == LC_N_FTS_MASK) {
7537 data = RREG32_PCIE(PCIE_LC_STATUS1);
7538 if ((data & LC_REVERSE_XMIT) && (data & LC_REVERSE_RCVR)) {
7539 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL);
7540 data &= ~LC_L0S_INACTIVITY_MASK;
7541 if (orig != data)
7542 WREG32_PCIE_PORT(PCIE_LC_CNTL, data);
7543 }
7544 }
7545 }
7546}