drm/radeon: dump full IB if we hit a packet error
[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"
bfc1f97d 30#include "radeon_audio.h"
760285e7 31#include <drm/radeon_drm.h>
43b3cd99
AD
32#include "sid.h"
33#include "atom.h"
48c0c902 34#include "si_blit_shaders.h"
bd8cd539 35#include "clearstate_si.h"
a0ceada6 36#include "radeon_ucode.h"
43b3cd99 37
0f0de06c
AD
38
39MODULE_FIRMWARE("radeon/TAHITI_pfp.bin");
40MODULE_FIRMWARE("radeon/TAHITI_me.bin");
41MODULE_FIRMWARE("radeon/TAHITI_ce.bin");
42MODULE_FIRMWARE("radeon/TAHITI_mc.bin");
1ebe9280 43MODULE_FIRMWARE("radeon/TAHITI_mc2.bin");
0f0de06c 44MODULE_FIRMWARE("radeon/TAHITI_rlc.bin");
a9e61410 45MODULE_FIRMWARE("radeon/TAHITI_smc.bin");
629bd33c
AD
46
47MODULE_FIRMWARE("radeon/tahiti_pfp.bin");
48MODULE_FIRMWARE("radeon/tahiti_me.bin");
49MODULE_FIRMWARE("radeon/tahiti_ce.bin");
50MODULE_FIRMWARE("radeon/tahiti_mc.bin");
51MODULE_FIRMWARE("radeon/tahiti_rlc.bin");
52MODULE_FIRMWARE("radeon/tahiti_smc.bin");
53
0f0de06c
AD
54MODULE_FIRMWARE("radeon/PITCAIRN_pfp.bin");
55MODULE_FIRMWARE("radeon/PITCAIRN_me.bin");
56MODULE_FIRMWARE("radeon/PITCAIRN_ce.bin");
57MODULE_FIRMWARE("radeon/PITCAIRN_mc.bin");
1ebe9280 58MODULE_FIRMWARE("radeon/PITCAIRN_mc2.bin");
0f0de06c 59MODULE_FIRMWARE("radeon/PITCAIRN_rlc.bin");
a9e61410 60MODULE_FIRMWARE("radeon/PITCAIRN_smc.bin");
629bd33c
AD
61
62MODULE_FIRMWARE("radeon/pitcairn_pfp.bin");
63MODULE_FIRMWARE("radeon/pitcairn_me.bin");
64MODULE_FIRMWARE("radeon/pitcairn_ce.bin");
65MODULE_FIRMWARE("radeon/pitcairn_mc.bin");
66MODULE_FIRMWARE("radeon/pitcairn_rlc.bin");
67MODULE_FIRMWARE("radeon/pitcairn_smc.bin");
68
0f0de06c
AD
69MODULE_FIRMWARE("radeon/VERDE_pfp.bin");
70MODULE_FIRMWARE("radeon/VERDE_me.bin");
71MODULE_FIRMWARE("radeon/VERDE_ce.bin");
72MODULE_FIRMWARE("radeon/VERDE_mc.bin");
1ebe9280 73MODULE_FIRMWARE("radeon/VERDE_mc2.bin");
0f0de06c 74MODULE_FIRMWARE("radeon/VERDE_rlc.bin");
a9e61410 75MODULE_FIRMWARE("radeon/VERDE_smc.bin");
629bd33c
AD
76
77MODULE_FIRMWARE("radeon/verde_pfp.bin");
78MODULE_FIRMWARE("radeon/verde_me.bin");
79MODULE_FIRMWARE("radeon/verde_ce.bin");
80MODULE_FIRMWARE("radeon/verde_mc.bin");
81MODULE_FIRMWARE("radeon/verde_rlc.bin");
82MODULE_FIRMWARE("radeon/verde_smc.bin");
83
bcc7f5d2
AD
84MODULE_FIRMWARE("radeon/OLAND_pfp.bin");
85MODULE_FIRMWARE("radeon/OLAND_me.bin");
86MODULE_FIRMWARE("radeon/OLAND_ce.bin");
87MODULE_FIRMWARE("radeon/OLAND_mc.bin");
1ebe9280 88MODULE_FIRMWARE("radeon/OLAND_mc2.bin");
bcc7f5d2 89MODULE_FIRMWARE("radeon/OLAND_rlc.bin");
a9e61410 90MODULE_FIRMWARE("radeon/OLAND_smc.bin");
629bd33c
AD
91
92MODULE_FIRMWARE("radeon/oland_pfp.bin");
93MODULE_FIRMWARE("radeon/oland_me.bin");
94MODULE_FIRMWARE("radeon/oland_ce.bin");
95MODULE_FIRMWARE("radeon/oland_mc.bin");
96MODULE_FIRMWARE("radeon/oland_rlc.bin");
97MODULE_FIRMWARE("radeon/oland_smc.bin");
98
c04c00b4
AD
99MODULE_FIRMWARE("radeon/HAINAN_pfp.bin");
100MODULE_FIRMWARE("radeon/HAINAN_me.bin");
101MODULE_FIRMWARE("radeon/HAINAN_ce.bin");
102MODULE_FIRMWARE("radeon/HAINAN_mc.bin");
1ebe9280 103MODULE_FIRMWARE("radeon/HAINAN_mc2.bin");
c04c00b4 104MODULE_FIRMWARE("radeon/HAINAN_rlc.bin");
a9e61410 105MODULE_FIRMWARE("radeon/HAINAN_smc.bin");
0f0de06c 106
629bd33c
AD
107MODULE_FIRMWARE("radeon/hainan_pfp.bin");
108MODULE_FIRMWARE("radeon/hainan_me.bin");
109MODULE_FIRMWARE("radeon/hainan_ce.bin");
110MODULE_FIRMWARE("radeon/hainan_mc.bin");
111MODULE_FIRMWARE("radeon/hainan_rlc.bin");
112MODULE_FIRMWARE("radeon/hainan_smc.bin");
113
65fcf668 114static u32 si_get_cu_active_bitmap(struct radeon_device *rdev, u32 se, u32 sh);
b9d305df 115static void si_pcie_gen3_enable(struct radeon_device *rdev);
e0bcf165 116static void si_program_aspm(struct radeon_device *rdev);
1fd11777
AD
117extern void sumo_rlc_fini(struct radeon_device *rdev);
118extern int sumo_rlc_init(struct radeon_device *rdev);
25a857fb
AD
119extern int r600_ih_ring_alloc(struct radeon_device *rdev);
120extern void r600_ih_ring_fini(struct radeon_device *rdev);
0a96d72b 121extern void evergreen_fix_pci_max_read_req_size(struct radeon_device *rdev);
c476dde2
AD
122extern void evergreen_mc_stop(struct radeon_device *rdev, struct evergreen_mc_save *save);
123extern void evergreen_mc_resume(struct radeon_device *rdev, struct evergreen_mc_save *save);
ca7db22b 124extern u32 evergreen_get_number_of_dram_channels(struct radeon_device *rdev);
1c534671 125extern void evergreen_print_gpu_status_regs(struct radeon_device *rdev);
014bb209 126extern bool evergreen_is_display_hung(struct radeon_device *rdev);
811e4d58
AD
127static void si_enable_gui_idle_interrupt(struct radeon_device *rdev,
128 bool enable);
4a5c8ea5
AD
129static void si_init_pg(struct radeon_device *rdev);
130static void si_init_cg(struct radeon_device *rdev);
a6f4ae8d
AD
131static void si_fini_pg(struct radeon_device *rdev);
132static void si_fini_cg(struct radeon_device *rdev);
133static void si_rlc_stop(struct radeon_device *rdev);
0a96d72b 134
6d8cf000
AD
135static const u32 verde_rlc_save_restore_register_list[] =
136{
137 (0x8000 << 16) | (0x98f4 >> 2),
138 0x00000000,
139 (0x8040 << 16) | (0x98f4 >> 2),
140 0x00000000,
141 (0x8000 << 16) | (0xe80 >> 2),
142 0x00000000,
143 (0x8040 << 16) | (0xe80 >> 2),
144 0x00000000,
145 (0x8000 << 16) | (0x89bc >> 2),
146 0x00000000,
147 (0x8040 << 16) | (0x89bc >> 2),
148 0x00000000,
149 (0x8000 << 16) | (0x8c1c >> 2),
150 0x00000000,
151 (0x8040 << 16) | (0x8c1c >> 2),
152 0x00000000,
153 (0x9c00 << 16) | (0x98f0 >> 2),
154 0x00000000,
155 (0x9c00 << 16) | (0xe7c >> 2),
156 0x00000000,
157 (0x8000 << 16) | (0x9148 >> 2),
158 0x00000000,
159 (0x8040 << 16) | (0x9148 >> 2),
160 0x00000000,
161 (0x9c00 << 16) | (0x9150 >> 2),
162 0x00000000,
163 (0x9c00 << 16) | (0x897c >> 2),
164 0x00000000,
165 (0x9c00 << 16) | (0x8d8c >> 2),
166 0x00000000,
167 (0x9c00 << 16) | (0xac54 >> 2),
168 0X00000000,
169 0x3,
170 (0x9c00 << 16) | (0x98f8 >> 2),
171 0x00000000,
172 (0x9c00 << 16) | (0x9910 >> 2),
173 0x00000000,
174 (0x9c00 << 16) | (0x9914 >> 2),
175 0x00000000,
176 (0x9c00 << 16) | (0x9918 >> 2),
177 0x00000000,
178 (0x9c00 << 16) | (0x991c >> 2),
179 0x00000000,
180 (0x9c00 << 16) | (0x9920 >> 2),
181 0x00000000,
182 (0x9c00 << 16) | (0x9924 >> 2),
183 0x00000000,
184 (0x9c00 << 16) | (0x9928 >> 2),
185 0x00000000,
186 (0x9c00 << 16) | (0x992c >> 2),
187 0x00000000,
188 (0x9c00 << 16) | (0x9930 >> 2),
189 0x00000000,
190 (0x9c00 << 16) | (0x9934 >> 2),
191 0x00000000,
192 (0x9c00 << 16) | (0x9938 >> 2),
193 0x00000000,
194 (0x9c00 << 16) | (0x993c >> 2),
195 0x00000000,
196 (0x9c00 << 16) | (0x9940 >> 2),
197 0x00000000,
198 (0x9c00 << 16) | (0x9944 >> 2),
199 0x00000000,
200 (0x9c00 << 16) | (0x9948 >> 2),
201 0x00000000,
202 (0x9c00 << 16) | (0x994c >> 2),
203 0x00000000,
204 (0x9c00 << 16) | (0x9950 >> 2),
205 0x00000000,
206 (0x9c00 << 16) | (0x9954 >> 2),
207 0x00000000,
208 (0x9c00 << 16) | (0x9958 >> 2),
209 0x00000000,
210 (0x9c00 << 16) | (0x995c >> 2),
211 0x00000000,
212 (0x9c00 << 16) | (0x9960 >> 2),
213 0x00000000,
214 (0x9c00 << 16) | (0x9964 >> 2),
215 0x00000000,
216 (0x9c00 << 16) | (0x9968 >> 2),
217 0x00000000,
218 (0x9c00 << 16) | (0x996c >> 2),
219 0x00000000,
220 (0x9c00 << 16) | (0x9970 >> 2),
221 0x00000000,
222 (0x9c00 << 16) | (0x9974 >> 2),
223 0x00000000,
224 (0x9c00 << 16) | (0x9978 >> 2),
225 0x00000000,
226 (0x9c00 << 16) | (0x997c >> 2),
227 0x00000000,
228 (0x9c00 << 16) | (0x9980 >> 2),
229 0x00000000,
230 (0x9c00 << 16) | (0x9984 >> 2),
231 0x00000000,
232 (0x9c00 << 16) | (0x9988 >> 2),
233 0x00000000,
234 (0x9c00 << 16) | (0x998c >> 2),
235 0x00000000,
236 (0x9c00 << 16) | (0x8c00 >> 2),
237 0x00000000,
238 (0x9c00 << 16) | (0x8c14 >> 2),
239 0x00000000,
240 (0x9c00 << 16) | (0x8c04 >> 2),
241 0x00000000,
242 (0x9c00 << 16) | (0x8c08 >> 2),
243 0x00000000,
244 (0x8000 << 16) | (0x9b7c >> 2),
245 0x00000000,
246 (0x8040 << 16) | (0x9b7c >> 2),
247 0x00000000,
248 (0x8000 << 16) | (0xe84 >> 2),
249 0x00000000,
250 (0x8040 << 16) | (0xe84 >> 2),
251 0x00000000,
252 (0x8000 << 16) | (0x89c0 >> 2),
253 0x00000000,
254 (0x8040 << 16) | (0x89c0 >> 2),
255 0x00000000,
256 (0x8000 << 16) | (0x914c >> 2),
257 0x00000000,
258 (0x8040 << 16) | (0x914c >> 2),
259 0x00000000,
260 (0x8000 << 16) | (0x8c20 >> 2),
261 0x00000000,
262 (0x8040 << 16) | (0x8c20 >> 2),
263 0x00000000,
264 (0x8000 << 16) | (0x9354 >> 2),
265 0x00000000,
266 (0x8040 << 16) | (0x9354 >> 2),
267 0x00000000,
268 (0x9c00 << 16) | (0x9060 >> 2),
269 0x00000000,
270 (0x9c00 << 16) | (0x9364 >> 2),
271 0x00000000,
272 (0x9c00 << 16) | (0x9100 >> 2),
273 0x00000000,
274 (0x9c00 << 16) | (0x913c >> 2),
275 0x00000000,
276 (0x8000 << 16) | (0x90e0 >> 2),
277 0x00000000,
278 (0x8000 << 16) | (0x90e4 >> 2),
279 0x00000000,
280 (0x8000 << 16) | (0x90e8 >> 2),
281 0x00000000,
282 (0x8040 << 16) | (0x90e0 >> 2),
283 0x00000000,
284 (0x8040 << 16) | (0x90e4 >> 2),
285 0x00000000,
286 (0x8040 << 16) | (0x90e8 >> 2),
287 0x00000000,
288 (0x9c00 << 16) | (0x8bcc >> 2),
289 0x00000000,
290 (0x9c00 << 16) | (0x8b24 >> 2),
291 0x00000000,
292 (0x9c00 << 16) | (0x88c4 >> 2),
293 0x00000000,
294 (0x9c00 << 16) | (0x8e50 >> 2),
295 0x00000000,
296 (0x9c00 << 16) | (0x8c0c >> 2),
297 0x00000000,
298 (0x9c00 << 16) | (0x8e58 >> 2),
299 0x00000000,
300 (0x9c00 << 16) | (0x8e5c >> 2),
301 0x00000000,
302 (0x9c00 << 16) | (0x9508 >> 2),
303 0x00000000,
304 (0x9c00 << 16) | (0x950c >> 2),
305 0x00000000,
306 (0x9c00 << 16) | (0x9494 >> 2),
307 0x00000000,
308 (0x9c00 << 16) | (0xac0c >> 2),
309 0x00000000,
310 (0x9c00 << 16) | (0xac10 >> 2),
311 0x00000000,
312 (0x9c00 << 16) | (0xac14 >> 2),
313 0x00000000,
314 (0x9c00 << 16) | (0xae00 >> 2),
315 0x00000000,
316 (0x9c00 << 16) | (0xac08 >> 2),
317 0x00000000,
318 (0x9c00 << 16) | (0x88d4 >> 2),
319 0x00000000,
320 (0x9c00 << 16) | (0x88c8 >> 2),
321 0x00000000,
322 (0x9c00 << 16) | (0x88cc >> 2),
323 0x00000000,
324 (0x9c00 << 16) | (0x89b0 >> 2),
325 0x00000000,
326 (0x9c00 << 16) | (0x8b10 >> 2),
327 0x00000000,
328 (0x9c00 << 16) | (0x8a14 >> 2),
329 0x00000000,
330 (0x9c00 << 16) | (0x9830 >> 2),
331 0x00000000,
332 (0x9c00 << 16) | (0x9834 >> 2),
333 0x00000000,
334 (0x9c00 << 16) | (0x9838 >> 2),
335 0x00000000,
336 (0x9c00 << 16) | (0x9a10 >> 2),
337 0x00000000,
338 (0x8000 << 16) | (0x9870 >> 2),
339 0x00000000,
340 (0x8000 << 16) | (0x9874 >> 2),
341 0x00000000,
342 (0x8001 << 16) | (0x9870 >> 2),
343 0x00000000,
344 (0x8001 << 16) | (0x9874 >> 2),
345 0x00000000,
346 (0x8040 << 16) | (0x9870 >> 2),
347 0x00000000,
348 (0x8040 << 16) | (0x9874 >> 2),
349 0x00000000,
350 (0x8041 << 16) | (0x9870 >> 2),
351 0x00000000,
352 (0x8041 << 16) | (0x9874 >> 2),
353 0x00000000,
354 0x00000000
355};
356
205996c0
AD
357static const u32 tahiti_golden_rlc_registers[] =
358{
359 0xc424, 0xffffffff, 0x00601005,
360 0xc47c, 0xffffffff, 0x10104040,
361 0xc488, 0xffffffff, 0x0100000a,
362 0xc314, 0xffffffff, 0x00000800,
363 0xc30c, 0xffffffff, 0x800000f4,
364 0xf4a8, 0xffffffff, 0x00000000
365};
366
367static const u32 tahiti_golden_registers[] =
368{
369 0x9a10, 0x00010000, 0x00018208,
370 0x9830, 0xffffffff, 0x00000000,
371 0x9834, 0xf00fffff, 0x00000400,
372 0x9838, 0x0002021c, 0x00020200,
373 0xc78, 0x00000080, 0x00000000,
374 0xd030, 0x000300c0, 0x00800040,
375 0xd830, 0x000300c0, 0x00800040,
376 0x5bb0, 0x000000f0, 0x00000070,
377 0x5bc0, 0x00200000, 0x50100000,
378 0x7030, 0x31000311, 0x00000011,
379 0x277c, 0x00000003, 0x000007ff,
380 0x240c, 0x000007ff, 0x00000000,
381 0x8a14, 0xf000001f, 0x00000007,
382 0x8b24, 0xffffffff, 0x00ffffff,
383 0x8b10, 0x0000ff0f, 0x00000000,
384 0x28a4c, 0x07ffffff, 0x4e000000,
385 0x28350, 0x3f3f3fff, 0x2a00126a,
386 0x30, 0x000000ff, 0x0040,
387 0x34, 0x00000040, 0x00004040,
388 0x9100, 0x07ffffff, 0x03000000,
389 0x8e88, 0x01ff1f3f, 0x00000000,
390 0x8e84, 0x01ff1f3f, 0x00000000,
391 0x9060, 0x0000007f, 0x00000020,
392 0x9508, 0x00010000, 0x00010000,
393 0xac14, 0x00000200, 0x000002fb,
394 0xac10, 0xffffffff, 0x0000543b,
395 0xac0c, 0xffffffff, 0xa9210876,
396 0x88d0, 0xffffffff, 0x000fff40,
397 0x88d4, 0x0000001f, 0x00000010,
398 0x1410, 0x20000000, 0x20fffed8,
399 0x15c0, 0x000c0fc0, 0x000c0400
400};
401
402static const u32 tahiti_golden_registers2[] =
403{
404 0xc64, 0x00000001, 0x00000001
405};
406
407static const u32 pitcairn_golden_rlc_registers[] =
408{
409 0xc424, 0xffffffff, 0x00601004,
410 0xc47c, 0xffffffff, 0x10102020,
411 0xc488, 0xffffffff, 0x01000020,
412 0xc314, 0xffffffff, 0x00000800,
413 0xc30c, 0xffffffff, 0x800000a4
414};
415
416static const u32 pitcairn_golden_registers[] =
417{
418 0x9a10, 0x00010000, 0x00018208,
419 0x9830, 0xffffffff, 0x00000000,
420 0x9834, 0xf00fffff, 0x00000400,
421 0x9838, 0x0002021c, 0x00020200,
422 0xc78, 0x00000080, 0x00000000,
423 0xd030, 0x000300c0, 0x00800040,
424 0xd830, 0x000300c0, 0x00800040,
425 0x5bb0, 0x000000f0, 0x00000070,
426 0x5bc0, 0x00200000, 0x50100000,
427 0x7030, 0x31000311, 0x00000011,
428 0x2ae4, 0x00073ffe, 0x000022a2,
429 0x240c, 0x000007ff, 0x00000000,
430 0x8a14, 0xf000001f, 0x00000007,
431 0x8b24, 0xffffffff, 0x00ffffff,
432 0x8b10, 0x0000ff0f, 0x00000000,
433 0x28a4c, 0x07ffffff, 0x4e000000,
434 0x28350, 0x3f3f3fff, 0x2a00126a,
435 0x30, 0x000000ff, 0x0040,
436 0x34, 0x00000040, 0x00004040,
437 0x9100, 0x07ffffff, 0x03000000,
438 0x9060, 0x0000007f, 0x00000020,
439 0x9508, 0x00010000, 0x00010000,
440 0xac14, 0x000003ff, 0x000000f7,
441 0xac10, 0xffffffff, 0x00000000,
442 0xac0c, 0xffffffff, 0x32761054,
443 0x88d4, 0x0000001f, 0x00000010,
444 0x15c0, 0x000c0fc0, 0x000c0400
445};
446
447static const u32 verde_golden_rlc_registers[] =
448{
449 0xc424, 0xffffffff, 0x033f1005,
450 0xc47c, 0xffffffff, 0x10808020,
451 0xc488, 0xffffffff, 0x00800008,
452 0xc314, 0xffffffff, 0x00001000,
453 0xc30c, 0xffffffff, 0x80010014
454};
455
456static const u32 verde_golden_registers[] =
457{
458 0x9a10, 0x00010000, 0x00018208,
459 0x9830, 0xffffffff, 0x00000000,
460 0x9834, 0xf00fffff, 0x00000400,
461 0x9838, 0x0002021c, 0x00020200,
462 0xc78, 0x00000080, 0x00000000,
463 0xd030, 0x000300c0, 0x00800040,
464 0xd030, 0x000300c0, 0x00800040,
465 0xd830, 0x000300c0, 0x00800040,
466 0xd830, 0x000300c0, 0x00800040,
467 0x5bb0, 0x000000f0, 0x00000070,
468 0x5bc0, 0x00200000, 0x50100000,
469 0x7030, 0x31000311, 0x00000011,
470 0x2ae4, 0x00073ffe, 0x000022a2,
471 0x2ae4, 0x00073ffe, 0x000022a2,
472 0x2ae4, 0x00073ffe, 0x000022a2,
473 0x240c, 0x000007ff, 0x00000000,
474 0x240c, 0x000007ff, 0x00000000,
475 0x240c, 0x000007ff, 0x00000000,
476 0x8a14, 0xf000001f, 0x00000007,
477 0x8a14, 0xf000001f, 0x00000007,
478 0x8a14, 0xf000001f, 0x00000007,
479 0x8b24, 0xffffffff, 0x00ffffff,
480 0x8b10, 0x0000ff0f, 0x00000000,
481 0x28a4c, 0x07ffffff, 0x4e000000,
482 0x28350, 0x3f3f3fff, 0x0000124a,
483 0x28350, 0x3f3f3fff, 0x0000124a,
484 0x28350, 0x3f3f3fff, 0x0000124a,
485 0x30, 0x000000ff, 0x0040,
486 0x34, 0x00000040, 0x00004040,
487 0x9100, 0x07ffffff, 0x03000000,
488 0x9100, 0x07ffffff, 0x03000000,
489 0x8e88, 0x01ff1f3f, 0x00000000,
490 0x8e88, 0x01ff1f3f, 0x00000000,
491 0x8e88, 0x01ff1f3f, 0x00000000,
492 0x8e84, 0x01ff1f3f, 0x00000000,
493 0x8e84, 0x01ff1f3f, 0x00000000,
494 0x8e84, 0x01ff1f3f, 0x00000000,
495 0x9060, 0x0000007f, 0x00000020,
496 0x9508, 0x00010000, 0x00010000,
497 0xac14, 0x000003ff, 0x00000003,
498 0xac14, 0x000003ff, 0x00000003,
499 0xac14, 0x000003ff, 0x00000003,
500 0xac10, 0xffffffff, 0x00000000,
501 0xac10, 0xffffffff, 0x00000000,
502 0xac10, 0xffffffff, 0x00000000,
503 0xac0c, 0xffffffff, 0x00001032,
504 0xac0c, 0xffffffff, 0x00001032,
505 0xac0c, 0xffffffff, 0x00001032,
506 0x88d4, 0x0000001f, 0x00000010,
507 0x88d4, 0x0000001f, 0x00000010,
508 0x88d4, 0x0000001f, 0x00000010,
509 0x15c0, 0x000c0fc0, 0x000c0400
510};
511
512static const u32 oland_golden_rlc_registers[] =
513{
514 0xc424, 0xffffffff, 0x00601005,
515 0xc47c, 0xffffffff, 0x10104040,
516 0xc488, 0xffffffff, 0x0100000a,
517 0xc314, 0xffffffff, 0x00000800,
518 0xc30c, 0xffffffff, 0x800000f4
519};
520
521static const u32 oland_golden_registers[] =
522{
523 0x9a10, 0x00010000, 0x00018208,
524 0x9830, 0xffffffff, 0x00000000,
525 0x9834, 0xf00fffff, 0x00000400,
526 0x9838, 0x0002021c, 0x00020200,
527 0xc78, 0x00000080, 0x00000000,
528 0xd030, 0x000300c0, 0x00800040,
529 0xd830, 0x000300c0, 0x00800040,
530 0x5bb0, 0x000000f0, 0x00000070,
531 0x5bc0, 0x00200000, 0x50100000,
532 0x7030, 0x31000311, 0x00000011,
533 0x2ae4, 0x00073ffe, 0x000022a2,
534 0x240c, 0x000007ff, 0x00000000,
535 0x8a14, 0xf000001f, 0x00000007,
536 0x8b24, 0xffffffff, 0x00ffffff,
537 0x8b10, 0x0000ff0f, 0x00000000,
538 0x28a4c, 0x07ffffff, 0x4e000000,
539 0x28350, 0x3f3f3fff, 0x00000082,
540 0x30, 0x000000ff, 0x0040,
541 0x34, 0x00000040, 0x00004040,
542 0x9100, 0x07ffffff, 0x03000000,
543 0x9060, 0x0000007f, 0x00000020,
544 0x9508, 0x00010000, 0x00010000,
545 0xac14, 0x000003ff, 0x000000f3,
546 0xac10, 0xffffffff, 0x00000000,
547 0xac0c, 0xffffffff, 0x00003210,
548 0x88d4, 0x0000001f, 0x00000010,
549 0x15c0, 0x000c0fc0, 0x000c0400
550};
551
fffbdda4
AD
552static const u32 hainan_golden_registers[] =
553{
554 0x9a10, 0x00010000, 0x00018208,
555 0x9830, 0xffffffff, 0x00000000,
556 0x9834, 0xf00fffff, 0x00000400,
557 0x9838, 0x0002021c, 0x00020200,
558 0xd0c0, 0xff000fff, 0x00000100,
559 0xd030, 0x000300c0, 0x00800040,
560 0xd8c0, 0xff000fff, 0x00000100,
561 0xd830, 0x000300c0, 0x00800040,
562 0x2ae4, 0x00073ffe, 0x000022a2,
563 0x240c, 0x000007ff, 0x00000000,
564 0x8a14, 0xf000001f, 0x00000007,
565 0x8b24, 0xffffffff, 0x00ffffff,
566 0x8b10, 0x0000ff0f, 0x00000000,
567 0x28a4c, 0x07ffffff, 0x4e000000,
568 0x28350, 0x3f3f3fff, 0x00000000,
569 0x30, 0x000000ff, 0x0040,
570 0x34, 0x00000040, 0x00004040,
571 0x9100, 0x03e00000, 0x03600000,
572 0x9060, 0x0000007f, 0x00000020,
573 0x9508, 0x00010000, 0x00010000,
574 0xac14, 0x000003ff, 0x000000f1,
575 0xac10, 0xffffffff, 0x00000000,
576 0xac0c, 0xffffffff, 0x00003210,
577 0x88d4, 0x0000001f, 0x00000010,
578 0x15c0, 0x000c0fc0, 0x000c0400
579};
580
581static const u32 hainan_golden_registers2[] =
582{
583 0x98f8, 0xffffffff, 0x02010001
584};
585
205996c0
AD
586static const u32 tahiti_mgcg_cgcg_init[] =
587{
588 0xc400, 0xffffffff, 0xfffffffc,
589 0x802c, 0xffffffff, 0xe0000000,
590 0x9a60, 0xffffffff, 0x00000100,
591 0x92a4, 0xffffffff, 0x00000100,
592 0xc164, 0xffffffff, 0x00000100,
593 0x9774, 0xffffffff, 0x00000100,
594 0x8984, 0xffffffff, 0x06000100,
595 0x8a18, 0xffffffff, 0x00000100,
596 0x92a0, 0xffffffff, 0x00000100,
597 0xc380, 0xffffffff, 0x00000100,
598 0x8b28, 0xffffffff, 0x00000100,
599 0x9144, 0xffffffff, 0x00000100,
600 0x8d88, 0xffffffff, 0x00000100,
601 0x8d8c, 0xffffffff, 0x00000100,
602 0x9030, 0xffffffff, 0x00000100,
603 0x9034, 0xffffffff, 0x00000100,
604 0x9038, 0xffffffff, 0x00000100,
605 0x903c, 0xffffffff, 0x00000100,
606 0xad80, 0xffffffff, 0x00000100,
607 0xac54, 0xffffffff, 0x00000100,
608 0x897c, 0xffffffff, 0x06000100,
609 0x9868, 0xffffffff, 0x00000100,
610 0x9510, 0xffffffff, 0x00000100,
611 0xaf04, 0xffffffff, 0x00000100,
612 0xae04, 0xffffffff, 0x00000100,
613 0x949c, 0xffffffff, 0x00000100,
614 0x802c, 0xffffffff, 0xe0000000,
615 0x9160, 0xffffffff, 0x00010000,
616 0x9164, 0xffffffff, 0x00030002,
617 0x9168, 0xffffffff, 0x00040007,
618 0x916c, 0xffffffff, 0x00060005,
619 0x9170, 0xffffffff, 0x00090008,
620 0x9174, 0xffffffff, 0x00020001,
621 0x9178, 0xffffffff, 0x00040003,
622 0x917c, 0xffffffff, 0x00000007,
623 0x9180, 0xffffffff, 0x00060005,
624 0x9184, 0xffffffff, 0x00090008,
625 0x9188, 0xffffffff, 0x00030002,
626 0x918c, 0xffffffff, 0x00050004,
627 0x9190, 0xffffffff, 0x00000008,
628 0x9194, 0xffffffff, 0x00070006,
629 0x9198, 0xffffffff, 0x000a0009,
630 0x919c, 0xffffffff, 0x00040003,
631 0x91a0, 0xffffffff, 0x00060005,
632 0x91a4, 0xffffffff, 0x00000009,
633 0x91a8, 0xffffffff, 0x00080007,
634 0x91ac, 0xffffffff, 0x000b000a,
635 0x91b0, 0xffffffff, 0x00050004,
636 0x91b4, 0xffffffff, 0x00070006,
637 0x91b8, 0xffffffff, 0x0008000b,
638 0x91bc, 0xffffffff, 0x000a0009,
639 0x91c0, 0xffffffff, 0x000d000c,
640 0x91c4, 0xffffffff, 0x00060005,
641 0x91c8, 0xffffffff, 0x00080007,
642 0x91cc, 0xffffffff, 0x0000000b,
643 0x91d0, 0xffffffff, 0x000a0009,
644 0x91d4, 0xffffffff, 0x000d000c,
645 0x91d8, 0xffffffff, 0x00070006,
646 0x91dc, 0xffffffff, 0x00090008,
647 0x91e0, 0xffffffff, 0x0000000c,
648 0x91e4, 0xffffffff, 0x000b000a,
649 0x91e8, 0xffffffff, 0x000e000d,
650 0x91ec, 0xffffffff, 0x00080007,
651 0x91f0, 0xffffffff, 0x000a0009,
652 0x91f4, 0xffffffff, 0x0000000d,
653 0x91f8, 0xffffffff, 0x000c000b,
654 0x91fc, 0xffffffff, 0x000f000e,
655 0x9200, 0xffffffff, 0x00090008,
656 0x9204, 0xffffffff, 0x000b000a,
657 0x9208, 0xffffffff, 0x000c000f,
658 0x920c, 0xffffffff, 0x000e000d,
659 0x9210, 0xffffffff, 0x00110010,
660 0x9214, 0xffffffff, 0x000a0009,
661 0x9218, 0xffffffff, 0x000c000b,
662 0x921c, 0xffffffff, 0x0000000f,
663 0x9220, 0xffffffff, 0x000e000d,
664 0x9224, 0xffffffff, 0x00110010,
665 0x9228, 0xffffffff, 0x000b000a,
666 0x922c, 0xffffffff, 0x000d000c,
667 0x9230, 0xffffffff, 0x00000010,
668 0x9234, 0xffffffff, 0x000f000e,
669 0x9238, 0xffffffff, 0x00120011,
670 0x923c, 0xffffffff, 0x000c000b,
671 0x9240, 0xffffffff, 0x000e000d,
672 0x9244, 0xffffffff, 0x00000011,
673 0x9248, 0xffffffff, 0x0010000f,
674 0x924c, 0xffffffff, 0x00130012,
675 0x9250, 0xffffffff, 0x000d000c,
676 0x9254, 0xffffffff, 0x000f000e,
677 0x9258, 0xffffffff, 0x00100013,
678 0x925c, 0xffffffff, 0x00120011,
679 0x9260, 0xffffffff, 0x00150014,
680 0x9264, 0xffffffff, 0x000e000d,
681 0x9268, 0xffffffff, 0x0010000f,
682 0x926c, 0xffffffff, 0x00000013,
683 0x9270, 0xffffffff, 0x00120011,
684 0x9274, 0xffffffff, 0x00150014,
685 0x9278, 0xffffffff, 0x000f000e,
686 0x927c, 0xffffffff, 0x00110010,
687 0x9280, 0xffffffff, 0x00000014,
688 0x9284, 0xffffffff, 0x00130012,
689 0x9288, 0xffffffff, 0x00160015,
690 0x928c, 0xffffffff, 0x0010000f,
691 0x9290, 0xffffffff, 0x00120011,
692 0x9294, 0xffffffff, 0x00000015,
693 0x9298, 0xffffffff, 0x00140013,
694 0x929c, 0xffffffff, 0x00170016,
695 0x9150, 0xffffffff, 0x96940200,
696 0x8708, 0xffffffff, 0x00900100,
697 0xc478, 0xffffffff, 0x00000080,
698 0xc404, 0xffffffff, 0x0020003f,
699 0x30, 0xffffffff, 0x0000001c,
700 0x34, 0x000f0000, 0x000f0000,
701 0x160c, 0xffffffff, 0x00000100,
702 0x1024, 0xffffffff, 0x00000100,
703 0x102c, 0x00000101, 0x00000000,
704 0x20a8, 0xffffffff, 0x00000104,
705 0x264c, 0x000c0000, 0x000c0000,
706 0x2648, 0x000c0000, 0x000c0000,
707 0x55e4, 0xff000fff, 0x00000100,
708 0x55e8, 0x00000001, 0x00000001,
709 0x2f50, 0x00000001, 0x00000001,
710 0x30cc, 0xc0000fff, 0x00000104,
711 0xc1e4, 0x00000001, 0x00000001,
712 0xd0c0, 0xfffffff0, 0x00000100,
713 0xd8c0, 0xfffffff0, 0x00000100
714};
715
716static const u32 pitcairn_mgcg_cgcg_init[] =
717{
718 0xc400, 0xffffffff, 0xfffffffc,
719 0x802c, 0xffffffff, 0xe0000000,
720 0x9a60, 0xffffffff, 0x00000100,
721 0x92a4, 0xffffffff, 0x00000100,
722 0xc164, 0xffffffff, 0x00000100,
723 0x9774, 0xffffffff, 0x00000100,
724 0x8984, 0xffffffff, 0x06000100,
725 0x8a18, 0xffffffff, 0x00000100,
726 0x92a0, 0xffffffff, 0x00000100,
727 0xc380, 0xffffffff, 0x00000100,
728 0x8b28, 0xffffffff, 0x00000100,
729 0x9144, 0xffffffff, 0x00000100,
730 0x8d88, 0xffffffff, 0x00000100,
731 0x8d8c, 0xffffffff, 0x00000100,
732 0x9030, 0xffffffff, 0x00000100,
733 0x9034, 0xffffffff, 0x00000100,
734 0x9038, 0xffffffff, 0x00000100,
735 0x903c, 0xffffffff, 0x00000100,
736 0xad80, 0xffffffff, 0x00000100,
737 0xac54, 0xffffffff, 0x00000100,
738 0x897c, 0xffffffff, 0x06000100,
739 0x9868, 0xffffffff, 0x00000100,
740 0x9510, 0xffffffff, 0x00000100,
741 0xaf04, 0xffffffff, 0x00000100,
742 0xae04, 0xffffffff, 0x00000100,
743 0x949c, 0xffffffff, 0x00000100,
744 0x802c, 0xffffffff, 0xe0000000,
745 0x9160, 0xffffffff, 0x00010000,
746 0x9164, 0xffffffff, 0x00030002,
747 0x9168, 0xffffffff, 0x00040007,
748 0x916c, 0xffffffff, 0x00060005,
749 0x9170, 0xffffffff, 0x00090008,
750 0x9174, 0xffffffff, 0x00020001,
751 0x9178, 0xffffffff, 0x00040003,
752 0x917c, 0xffffffff, 0x00000007,
753 0x9180, 0xffffffff, 0x00060005,
754 0x9184, 0xffffffff, 0x00090008,
755 0x9188, 0xffffffff, 0x00030002,
756 0x918c, 0xffffffff, 0x00050004,
757 0x9190, 0xffffffff, 0x00000008,
758 0x9194, 0xffffffff, 0x00070006,
759 0x9198, 0xffffffff, 0x000a0009,
760 0x919c, 0xffffffff, 0x00040003,
761 0x91a0, 0xffffffff, 0x00060005,
762 0x91a4, 0xffffffff, 0x00000009,
763 0x91a8, 0xffffffff, 0x00080007,
764 0x91ac, 0xffffffff, 0x000b000a,
765 0x91b0, 0xffffffff, 0x00050004,
766 0x91b4, 0xffffffff, 0x00070006,
767 0x91b8, 0xffffffff, 0x0008000b,
768 0x91bc, 0xffffffff, 0x000a0009,
769 0x91c0, 0xffffffff, 0x000d000c,
770 0x9200, 0xffffffff, 0x00090008,
771 0x9204, 0xffffffff, 0x000b000a,
772 0x9208, 0xffffffff, 0x000c000f,
773 0x920c, 0xffffffff, 0x000e000d,
774 0x9210, 0xffffffff, 0x00110010,
775 0x9214, 0xffffffff, 0x000a0009,
776 0x9218, 0xffffffff, 0x000c000b,
777 0x921c, 0xffffffff, 0x0000000f,
778 0x9220, 0xffffffff, 0x000e000d,
779 0x9224, 0xffffffff, 0x00110010,
780 0x9228, 0xffffffff, 0x000b000a,
781 0x922c, 0xffffffff, 0x000d000c,
782 0x9230, 0xffffffff, 0x00000010,
783 0x9234, 0xffffffff, 0x000f000e,
784 0x9238, 0xffffffff, 0x00120011,
785 0x923c, 0xffffffff, 0x000c000b,
786 0x9240, 0xffffffff, 0x000e000d,
787 0x9244, 0xffffffff, 0x00000011,
788 0x9248, 0xffffffff, 0x0010000f,
789 0x924c, 0xffffffff, 0x00130012,
790 0x9250, 0xffffffff, 0x000d000c,
791 0x9254, 0xffffffff, 0x000f000e,
792 0x9258, 0xffffffff, 0x00100013,
793 0x925c, 0xffffffff, 0x00120011,
794 0x9260, 0xffffffff, 0x00150014,
795 0x9150, 0xffffffff, 0x96940200,
796 0x8708, 0xffffffff, 0x00900100,
797 0xc478, 0xffffffff, 0x00000080,
798 0xc404, 0xffffffff, 0x0020003f,
799 0x30, 0xffffffff, 0x0000001c,
800 0x34, 0x000f0000, 0x000f0000,
801 0x160c, 0xffffffff, 0x00000100,
802 0x1024, 0xffffffff, 0x00000100,
803 0x102c, 0x00000101, 0x00000000,
804 0x20a8, 0xffffffff, 0x00000104,
805 0x55e4, 0xff000fff, 0x00000100,
806 0x55e8, 0x00000001, 0x00000001,
807 0x2f50, 0x00000001, 0x00000001,
808 0x30cc, 0xc0000fff, 0x00000104,
809 0xc1e4, 0x00000001, 0x00000001,
810 0xd0c0, 0xfffffff0, 0x00000100,
811 0xd8c0, 0xfffffff0, 0x00000100
812};
813
814static const u32 verde_mgcg_cgcg_init[] =
815{
816 0xc400, 0xffffffff, 0xfffffffc,
817 0x802c, 0xffffffff, 0xe0000000,
818 0x9a60, 0xffffffff, 0x00000100,
819 0x92a4, 0xffffffff, 0x00000100,
820 0xc164, 0xffffffff, 0x00000100,
821 0x9774, 0xffffffff, 0x00000100,
822 0x8984, 0xffffffff, 0x06000100,
823 0x8a18, 0xffffffff, 0x00000100,
824 0x92a0, 0xffffffff, 0x00000100,
825 0xc380, 0xffffffff, 0x00000100,
826 0x8b28, 0xffffffff, 0x00000100,
827 0x9144, 0xffffffff, 0x00000100,
828 0x8d88, 0xffffffff, 0x00000100,
829 0x8d8c, 0xffffffff, 0x00000100,
830 0x9030, 0xffffffff, 0x00000100,
831 0x9034, 0xffffffff, 0x00000100,
832 0x9038, 0xffffffff, 0x00000100,
833 0x903c, 0xffffffff, 0x00000100,
834 0xad80, 0xffffffff, 0x00000100,
835 0xac54, 0xffffffff, 0x00000100,
836 0x897c, 0xffffffff, 0x06000100,
837 0x9868, 0xffffffff, 0x00000100,
838 0x9510, 0xffffffff, 0x00000100,
839 0xaf04, 0xffffffff, 0x00000100,
840 0xae04, 0xffffffff, 0x00000100,
841 0x949c, 0xffffffff, 0x00000100,
842 0x802c, 0xffffffff, 0xe0000000,
843 0x9160, 0xffffffff, 0x00010000,
844 0x9164, 0xffffffff, 0x00030002,
845 0x9168, 0xffffffff, 0x00040007,
846 0x916c, 0xffffffff, 0x00060005,
847 0x9170, 0xffffffff, 0x00090008,
848 0x9174, 0xffffffff, 0x00020001,
849 0x9178, 0xffffffff, 0x00040003,
850 0x917c, 0xffffffff, 0x00000007,
851 0x9180, 0xffffffff, 0x00060005,
852 0x9184, 0xffffffff, 0x00090008,
853 0x9188, 0xffffffff, 0x00030002,
854 0x918c, 0xffffffff, 0x00050004,
855 0x9190, 0xffffffff, 0x00000008,
856 0x9194, 0xffffffff, 0x00070006,
857 0x9198, 0xffffffff, 0x000a0009,
858 0x919c, 0xffffffff, 0x00040003,
859 0x91a0, 0xffffffff, 0x00060005,
860 0x91a4, 0xffffffff, 0x00000009,
861 0x91a8, 0xffffffff, 0x00080007,
862 0x91ac, 0xffffffff, 0x000b000a,
863 0x91b0, 0xffffffff, 0x00050004,
864 0x91b4, 0xffffffff, 0x00070006,
865 0x91b8, 0xffffffff, 0x0008000b,
866 0x91bc, 0xffffffff, 0x000a0009,
867 0x91c0, 0xffffffff, 0x000d000c,
868 0x9200, 0xffffffff, 0x00090008,
869 0x9204, 0xffffffff, 0x000b000a,
870 0x9208, 0xffffffff, 0x000c000f,
871 0x920c, 0xffffffff, 0x000e000d,
872 0x9210, 0xffffffff, 0x00110010,
873 0x9214, 0xffffffff, 0x000a0009,
874 0x9218, 0xffffffff, 0x000c000b,
875 0x921c, 0xffffffff, 0x0000000f,
876 0x9220, 0xffffffff, 0x000e000d,
877 0x9224, 0xffffffff, 0x00110010,
878 0x9228, 0xffffffff, 0x000b000a,
879 0x922c, 0xffffffff, 0x000d000c,
880 0x9230, 0xffffffff, 0x00000010,
881 0x9234, 0xffffffff, 0x000f000e,
882 0x9238, 0xffffffff, 0x00120011,
883 0x923c, 0xffffffff, 0x000c000b,
884 0x9240, 0xffffffff, 0x000e000d,
885 0x9244, 0xffffffff, 0x00000011,
886 0x9248, 0xffffffff, 0x0010000f,
887 0x924c, 0xffffffff, 0x00130012,
888 0x9250, 0xffffffff, 0x000d000c,
889 0x9254, 0xffffffff, 0x000f000e,
890 0x9258, 0xffffffff, 0x00100013,
891 0x925c, 0xffffffff, 0x00120011,
892 0x9260, 0xffffffff, 0x00150014,
893 0x9150, 0xffffffff, 0x96940200,
894 0x8708, 0xffffffff, 0x00900100,
895 0xc478, 0xffffffff, 0x00000080,
896 0xc404, 0xffffffff, 0x0020003f,
897 0x30, 0xffffffff, 0x0000001c,
898 0x34, 0x000f0000, 0x000f0000,
899 0x160c, 0xffffffff, 0x00000100,
900 0x1024, 0xffffffff, 0x00000100,
901 0x102c, 0x00000101, 0x00000000,
902 0x20a8, 0xffffffff, 0x00000104,
903 0x264c, 0x000c0000, 0x000c0000,
904 0x2648, 0x000c0000, 0x000c0000,
905 0x55e4, 0xff000fff, 0x00000100,
906 0x55e8, 0x00000001, 0x00000001,
907 0x2f50, 0x00000001, 0x00000001,
908 0x30cc, 0xc0000fff, 0x00000104,
909 0xc1e4, 0x00000001, 0x00000001,
910 0xd0c0, 0xfffffff0, 0x00000100,
911 0xd8c0, 0xfffffff0, 0x00000100
912};
913
914static const u32 oland_mgcg_cgcg_init[] =
915{
916 0xc400, 0xffffffff, 0xfffffffc,
917 0x802c, 0xffffffff, 0xe0000000,
918 0x9a60, 0xffffffff, 0x00000100,
919 0x92a4, 0xffffffff, 0x00000100,
920 0xc164, 0xffffffff, 0x00000100,
921 0x9774, 0xffffffff, 0x00000100,
922 0x8984, 0xffffffff, 0x06000100,
923 0x8a18, 0xffffffff, 0x00000100,
924 0x92a0, 0xffffffff, 0x00000100,
925 0xc380, 0xffffffff, 0x00000100,
926 0x8b28, 0xffffffff, 0x00000100,
927 0x9144, 0xffffffff, 0x00000100,
928 0x8d88, 0xffffffff, 0x00000100,
929 0x8d8c, 0xffffffff, 0x00000100,
930 0x9030, 0xffffffff, 0x00000100,
931 0x9034, 0xffffffff, 0x00000100,
932 0x9038, 0xffffffff, 0x00000100,
933 0x903c, 0xffffffff, 0x00000100,
934 0xad80, 0xffffffff, 0x00000100,
935 0xac54, 0xffffffff, 0x00000100,
936 0x897c, 0xffffffff, 0x06000100,
937 0x9868, 0xffffffff, 0x00000100,
938 0x9510, 0xffffffff, 0x00000100,
939 0xaf04, 0xffffffff, 0x00000100,
940 0xae04, 0xffffffff, 0x00000100,
941 0x949c, 0xffffffff, 0x00000100,
942 0x802c, 0xffffffff, 0xe0000000,
943 0x9160, 0xffffffff, 0x00010000,
944 0x9164, 0xffffffff, 0x00030002,
945 0x9168, 0xffffffff, 0x00040007,
946 0x916c, 0xffffffff, 0x00060005,
947 0x9170, 0xffffffff, 0x00090008,
948 0x9174, 0xffffffff, 0x00020001,
949 0x9178, 0xffffffff, 0x00040003,
950 0x917c, 0xffffffff, 0x00000007,
951 0x9180, 0xffffffff, 0x00060005,
952 0x9184, 0xffffffff, 0x00090008,
953 0x9188, 0xffffffff, 0x00030002,
954 0x918c, 0xffffffff, 0x00050004,
955 0x9190, 0xffffffff, 0x00000008,
956 0x9194, 0xffffffff, 0x00070006,
957 0x9198, 0xffffffff, 0x000a0009,
958 0x919c, 0xffffffff, 0x00040003,
959 0x91a0, 0xffffffff, 0x00060005,
960 0x91a4, 0xffffffff, 0x00000009,
961 0x91a8, 0xffffffff, 0x00080007,
962 0x91ac, 0xffffffff, 0x000b000a,
963 0x91b0, 0xffffffff, 0x00050004,
964 0x91b4, 0xffffffff, 0x00070006,
965 0x91b8, 0xffffffff, 0x0008000b,
966 0x91bc, 0xffffffff, 0x000a0009,
967 0x91c0, 0xffffffff, 0x000d000c,
968 0x91c4, 0xffffffff, 0x00060005,
969 0x91c8, 0xffffffff, 0x00080007,
970 0x91cc, 0xffffffff, 0x0000000b,
971 0x91d0, 0xffffffff, 0x000a0009,
972 0x91d4, 0xffffffff, 0x000d000c,
973 0x9150, 0xffffffff, 0x96940200,
974 0x8708, 0xffffffff, 0x00900100,
975 0xc478, 0xffffffff, 0x00000080,
976 0xc404, 0xffffffff, 0x0020003f,
977 0x30, 0xffffffff, 0x0000001c,
978 0x34, 0x000f0000, 0x000f0000,
979 0x160c, 0xffffffff, 0x00000100,
980 0x1024, 0xffffffff, 0x00000100,
981 0x102c, 0x00000101, 0x00000000,
982 0x20a8, 0xffffffff, 0x00000104,
983 0x264c, 0x000c0000, 0x000c0000,
984 0x2648, 0x000c0000, 0x000c0000,
985 0x55e4, 0xff000fff, 0x00000100,
986 0x55e8, 0x00000001, 0x00000001,
987 0x2f50, 0x00000001, 0x00000001,
988 0x30cc, 0xc0000fff, 0x00000104,
989 0xc1e4, 0x00000001, 0x00000001,
990 0xd0c0, 0xfffffff0, 0x00000100,
991 0xd8c0, 0xfffffff0, 0x00000100
992};
993
fffbdda4
AD
994static const u32 hainan_mgcg_cgcg_init[] =
995{
996 0xc400, 0xffffffff, 0xfffffffc,
997 0x802c, 0xffffffff, 0xe0000000,
998 0x9a60, 0xffffffff, 0x00000100,
999 0x92a4, 0xffffffff, 0x00000100,
1000 0xc164, 0xffffffff, 0x00000100,
1001 0x9774, 0xffffffff, 0x00000100,
1002 0x8984, 0xffffffff, 0x06000100,
1003 0x8a18, 0xffffffff, 0x00000100,
1004 0x92a0, 0xffffffff, 0x00000100,
1005 0xc380, 0xffffffff, 0x00000100,
1006 0x8b28, 0xffffffff, 0x00000100,
1007 0x9144, 0xffffffff, 0x00000100,
1008 0x8d88, 0xffffffff, 0x00000100,
1009 0x8d8c, 0xffffffff, 0x00000100,
1010 0x9030, 0xffffffff, 0x00000100,
1011 0x9034, 0xffffffff, 0x00000100,
1012 0x9038, 0xffffffff, 0x00000100,
1013 0x903c, 0xffffffff, 0x00000100,
1014 0xad80, 0xffffffff, 0x00000100,
1015 0xac54, 0xffffffff, 0x00000100,
1016 0x897c, 0xffffffff, 0x06000100,
1017 0x9868, 0xffffffff, 0x00000100,
1018 0x9510, 0xffffffff, 0x00000100,
1019 0xaf04, 0xffffffff, 0x00000100,
1020 0xae04, 0xffffffff, 0x00000100,
1021 0x949c, 0xffffffff, 0x00000100,
1022 0x802c, 0xffffffff, 0xe0000000,
1023 0x9160, 0xffffffff, 0x00010000,
1024 0x9164, 0xffffffff, 0x00030002,
1025 0x9168, 0xffffffff, 0x00040007,
1026 0x916c, 0xffffffff, 0x00060005,
1027 0x9170, 0xffffffff, 0x00090008,
1028 0x9174, 0xffffffff, 0x00020001,
1029 0x9178, 0xffffffff, 0x00040003,
1030 0x917c, 0xffffffff, 0x00000007,
1031 0x9180, 0xffffffff, 0x00060005,
1032 0x9184, 0xffffffff, 0x00090008,
1033 0x9188, 0xffffffff, 0x00030002,
1034 0x918c, 0xffffffff, 0x00050004,
1035 0x9190, 0xffffffff, 0x00000008,
1036 0x9194, 0xffffffff, 0x00070006,
1037 0x9198, 0xffffffff, 0x000a0009,
1038 0x919c, 0xffffffff, 0x00040003,
1039 0x91a0, 0xffffffff, 0x00060005,
1040 0x91a4, 0xffffffff, 0x00000009,
1041 0x91a8, 0xffffffff, 0x00080007,
1042 0x91ac, 0xffffffff, 0x000b000a,
1043 0x91b0, 0xffffffff, 0x00050004,
1044 0x91b4, 0xffffffff, 0x00070006,
1045 0x91b8, 0xffffffff, 0x0008000b,
1046 0x91bc, 0xffffffff, 0x000a0009,
1047 0x91c0, 0xffffffff, 0x000d000c,
1048 0x91c4, 0xffffffff, 0x00060005,
1049 0x91c8, 0xffffffff, 0x00080007,
1050 0x91cc, 0xffffffff, 0x0000000b,
1051 0x91d0, 0xffffffff, 0x000a0009,
1052 0x91d4, 0xffffffff, 0x000d000c,
1053 0x9150, 0xffffffff, 0x96940200,
1054 0x8708, 0xffffffff, 0x00900100,
1055 0xc478, 0xffffffff, 0x00000080,
1056 0xc404, 0xffffffff, 0x0020003f,
1057 0x30, 0xffffffff, 0x0000001c,
1058 0x34, 0x000f0000, 0x000f0000,
1059 0x160c, 0xffffffff, 0x00000100,
1060 0x1024, 0xffffffff, 0x00000100,
1061 0x20a8, 0xffffffff, 0x00000104,
1062 0x264c, 0x000c0000, 0x000c0000,
1063 0x2648, 0x000c0000, 0x000c0000,
1064 0x2f50, 0x00000001, 0x00000001,
1065 0x30cc, 0xc0000fff, 0x00000104,
1066 0xc1e4, 0x00000001, 0x00000001,
1067 0xd0c0, 0xfffffff0, 0x00000100,
1068 0xd8c0, 0xfffffff0, 0x00000100
1069};
1070
205996c0
AD
1071static u32 verde_pg_init[] =
1072{
1073 0x353c, 0xffffffff, 0x40000,
1074 0x3538, 0xffffffff, 0x200010ff,
1075 0x353c, 0xffffffff, 0x0,
1076 0x353c, 0xffffffff, 0x0,
1077 0x353c, 0xffffffff, 0x0,
1078 0x353c, 0xffffffff, 0x0,
1079 0x353c, 0xffffffff, 0x0,
1080 0x353c, 0xffffffff, 0x7007,
1081 0x3538, 0xffffffff, 0x300010ff,
1082 0x353c, 0xffffffff, 0x0,
1083 0x353c, 0xffffffff, 0x0,
1084 0x353c, 0xffffffff, 0x0,
1085 0x353c, 0xffffffff, 0x0,
1086 0x353c, 0xffffffff, 0x0,
1087 0x353c, 0xffffffff, 0x400000,
1088 0x3538, 0xffffffff, 0x100010ff,
1089 0x353c, 0xffffffff, 0x0,
1090 0x353c, 0xffffffff, 0x0,
1091 0x353c, 0xffffffff, 0x0,
1092 0x353c, 0xffffffff, 0x0,
1093 0x353c, 0xffffffff, 0x0,
1094 0x353c, 0xffffffff, 0x120200,
1095 0x3538, 0xffffffff, 0x500010ff,
1096 0x353c, 0xffffffff, 0x0,
1097 0x353c, 0xffffffff, 0x0,
1098 0x353c, 0xffffffff, 0x0,
1099 0x353c, 0xffffffff, 0x0,
1100 0x353c, 0xffffffff, 0x0,
1101 0x353c, 0xffffffff, 0x1e1e16,
1102 0x3538, 0xffffffff, 0x600010ff,
1103 0x353c, 0xffffffff, 0x0,
1104 0x353c, 0xffffffff, 0x0,
1105 0x353c, 0xffffffff, 0x0,
1106 0x353c, 0xffffffff, 0x0,
1107 0x353c, 0xffffffff, 0x0,
1108 0x353c, 0xffffffff, 0x171f1e,
1109 0x3538, 0xffffffff, 0x700010ff,
1110 0x353c, 0xffffffff, 0x0,
1111 0x353c, 0xffffffff, 0x0,
1112 0x353c, 0xffffffff, 0x0,
1113 0x353c, 0xffffffff, 0x0,
1114 0x353c, 0xffffffff, 0x0,
1115 0x353c, 0xffffffff, 0x0,
1116 0x3538, 0xffffffff, 0x9ff,
1117 0x3500, 0xffffffff, 0x0,
1118 0x3504, 0xffffffff, 0x10000800,
1119 0x3504, 0xffffffff, 0xf,
1120 0x3504, 0xffffffff, 0xf,
1121 0x3500, 0xffffffff, 0x4,
1122 0x3504, 0xffffffff, 0x1000051e,
1123 0x3504, 0xffffffff, 0xffff,
1124 0x3504, 0xffffffff, 0xffff,
1125 0x3500, 0xffffffff, 0x8,
1126 0x3504, 0xffffffff, 0x80500,
1127 0x3500, 0xffffffff, 0x12,
1128 0x3504, 0xffffffff, 0x9050c,
1129 0x3500, 0xffffffff, 0x1d,
1130 0x3504, 0xffffffff, 0xb052c,
1131 0x3500, 0xffffffff, 0x2a,
1132 0x3504, 0xffffffff, 0x1053e,
1133 0x3500, 0xffffffff, 0x2d,
1134 0x3504, 0xffffffff, 0x10546,
1135 0x3500, 0xffffffff, 0x30,
1136 0x3504, 0xffffffff, 0xa054e,
1137 0x3500, 0xffffffff, 0x3c,
1138 0x3504, 0xffffffff, 0x1055f,
1139 0x3500, 0xffffffff, 0x3f,
1140 0x3504, 0xffffffff, 0x10567,
1141 0x3500, 0xffffffff, 0x42,
1142 0x3504, 0xffffffff, 0x1056f,
1143 0x3500, 0xffffffff, 0x45,
1144 0x3504, 0xffffffff, 0x10572,
1145 0x3500, 0xffffffff, 0x48,
1146 0x3504, 0xffffffff, 0x20575,
1147 0x3500, 0xffffffff, 0x4c,
1148 0x3504, 0xffffffff, 0x190801,
1149 0x3500, 0xffffffff, 0x67,
1150 0x3504, 0xffffffff, 0x1082a,
1151 0x3500, 0xffffffff, 0x6a,
1152 0x3504, 0xffffffff, 0x1b082d,
1153 0x3500, 0xffffffff, 0x87,
1154 0x3504, 0xffffffff, 0x310851,
1155 0x3500, 0xffffffff, 0xba,
1156 0x3504, 0xffffffff, 0x891,
1157 0x3500, 0xffffffff, 0xbc,
1158 0x3504, 0xffffffff, 0x893,
1159 0x3500, 0xffffffff, 0xbe,
1160 0x3504, 0xffffffff, 0x20895,
1161 0x3500, 0xffffffff, 0xc2,
1162 0x3504, 0xffffffff, 0x20899,
1163 0x3500, 0xffffffff, 0xc6,
1164 0x3504, 0xffffffff, 0x2089d,
1165 0x3500, 0xffffffff, 0xca,
1166 0x3504, 0xffffffff, 0x8a1,
1167 0x3500, 0xffffffff, 0xcc,
1168 0x3504, 0xffffffff, 0x8a3,
1169 0x3500, 0xffffffff, 0xce,
1170 0x3504, 0xffffffff, 0x308a5,
1171 0x3500, 0xffffffff, 0xd3,
1172 0x3504, 0xffffffff, 0x6d08cd,
1173 0x3500, 0xffffffff, 0x142,
1174 0x3504, 0xffffffff, 0x2000095a,
1175 0x3504, 0xffffffff, 0x1,
1176 0x3500, 0xffffffff, 0x144,
1177 0x3504, 0xffffffff, 0x301f095b,
1178 0x3500, 0xffffffff, 0x165,
1179 0x3504, 0xffffffff, 0xc094d,
1180 0x3500, 0xffffffff, 0x173,
1181 0x3504, 0xffffffff, 0xf096d,
1182 0x3500, 0xffffffff, 0x184,
1183 0x3504, 0xffffffff, 0x15097f,
1184 0x3500, 0xffffffff, 0x19b,
1185 0x3504, 0xffffffff, 0xc0998,
1186 0x3500, 0xffffffff, 0x1a9,
1187 0x3504, 0xffffffff, 0x409a7,
1188 0x3500, 0xffffffff, 0x1af,
1189 0x3504, 0xffffffff, 0xcdc,
1190 0x3500, 0xffffffff, 0x1b1,
1191 0x3504, 0xffffffff, 0x800,
1192 0x3508, 0xffffffff, 0x6c9b2000,
1193 0x3510, 0xfc00, 0x2000,
1194 0x3544, 0xffffffff, 0xfc0,
1195 0x28d4, 0x00000100, 0x100
1196};
1197
1198static void si_init_golden_registers(struct radeon_device *rdev)
1199{
1200 switch (rdev->family) {
1201 case CHIP_TAHITI:
1202 radeon_program_register_sequence(rdev,
1203 tahiti_golden_registers,
1204 (const u32)ARRAY_SIZE(tahiti_golden_registers));
1205 radeon_program_register_sequence(rdev,
1206 tahiti_golden_rlc_registers,
1207 (const u32)ARRAY_SIZE(tahiti_golden_rlc_registers));
1208 radeon_program_register_sequence(rdev,
1209 tahiti_mgcg_cgcg_init,
1210 (const u32)ARRAY_SIZE(tahiti_mgcg_cgcg_init));
1211 radeon_program_register_sequence(rdev,
1212 tahiti_golden_registers2,
1213 (const u32)ARRAY_SIZE(tahiti_golden_registers2));
1214 break;
1215 case CHIP_PITCAIRN:
1216 radeon_program_register_sequence(rdev,
1217 pitcairn_golden_registers,
1218 (const u32)ARRAY_SIZE(pitcairn_golden_registers));
1219 radeon_program_register_sequence(rdev,
1220 pitcairn_golden_rlc_registers,
1221 (const u32)ARRAY_SIZE(pitcairn_golden_rlc_registers));
1222 radeon_program_register_sequence(rdev,
1223 pitcairn_mgcg_cgcg_init,
1224 (const u32)ARRAY_SIZE(pitcairn_mgcg_cgcg_init));
1225 break;
1226 case CHIP_VERDE:
1227 radeon_program_register_sequence(rdev,
1228 verde_golden_registers,
1229 (const u32)ARRAY_SIZE(verde_golden_registers));
1230 radeon_program_register_sequence(rdev,
1231 verde_golden_rlc_registers,
1232 (const u32)ARRAY_SIZE(verde_golden_rlc_registers));
1233 radeon_program_register_sequence(rdev,
1234 verde_mgcg_cgcg_init,
1235 (const u32)ARRAY_SIZE(verde_mgcg_cgcg_init));
1236 radeon_program_register_sequence(rdev,
1237 verde_pg_init,
1238 (const u32)ARRAY_SIZE(verde_pg_init));
1239 break;
1240 case CHIP_OLAND:
1241 radeon_program_register_sequence(rdev,
1242 oland_golden_registers,
1243 (const u32)ARRAY_SIZE(oland_golden_registers));
1244 radeon_program_register_sequence(rdev,
1245 oland_golden_rlc_registers,
1246 (const u32)ARRAY_SIZE(oland_golden_rlc_registers));
1247 radeon_program_register_sequence(rdev,
1248 oland_mgcg_cgcg_init,
1249 (const u32)ARRAY_SIZE(oland_mgcg_cgcg_init));
1250 break;
fffbdda4
AD
1251 case CHIP_HAINAN:
1252 radeon_program_register_sequence(rdev,
1253 hainan_golden_registers,
1254 (const u32)ARRAY_SIZE(hainan_golden_registers));
1255 radeon_program_register_sequence(rdev,
1256 hainan_golden_registers2,
1257 (const u32)ARRAY_SIZE(hainan_golden_registers2));
1258 radeon_program_register_sequence(rdev,
1259 hainan_mgcg_cgcg_init,
1260 (const u32)ARRAY_SIZE(hainan_mgcg_cgcg_init));
1261 break;
205996c0
AD
1262 default:
1263 break;
1264 }
1265}
1266
454d2e2a
AD
1267#define PCIE_BUS_CLK 10000
1268#define TCLK (PCIE_BUS_CLK / 10)
1269
1270/**
1271 * si_get_xclk - get the xclk
1272 *
1273 * @rdev: radeon_device pointer
1274 *
1275 * Returns the reference clock used by the gfx engine
1276 * (SI).
1277 */
1278u32 si_get_xclk(struct radeon_device *rdev)
1279{
1280 u32 reference_clock = rdev->clock.spll.reference_freq;
1281 u32 tmp;
1282
1283 tmp = RREG32(CG_CLKPIN_CNTL_2);
1284 if (tmp & MUX_TCLK_TO_XCLK)
1285 return TCLK;
1286
1287 tmp = RREG32(CG_CLKPIN_CNTL);
1288 if (tmp & XTALIN_DIVIDE)
1289 return reference_clock / 4;
1290
1291 return reference_clock;
1292}
1293
1bd47d2e
AD
1294/* get temperature in millidegrees */
1295int si_get_temp(struct radeon_device *rdev)
1296{
1297 u32 temp;
1298 int actual_temp = 0;
1299
1300 temp = (RREG32(CG_MULT_THERMAL_STATUS) & CTF_TEMP_MASK) >>
1301 CTF_TEMP_SHIFT;
1302
1303 if (temp & 0x200)
1304 actual_temp = 255;
1305 else
1306 actual_temp = temp & 0x1ff;
1307
1308 actual_temp = (actual_temp * 1000);
1309
1310 return actual_temp;
1311}
1312
8b074dd6
AD
1313#define TAHITI_IO_MC_REGS_SIZE 36
1314
1315static const u32 tahiti_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
1316 {0x0000006f, 0x03044000},
1317 {0x00000070, 0x0480c018},
1318 {0x00000071, 0x00000040},
1319 {0x00000072, 0x01000000},
1320 {0x00000074, 0x000000ff},
1321 {0x00000075, 0x00143400},
1322 {0x00000076, 0x08ec0800},
1323 {0x00000077, 0x040000cc},
1324 {0x00000079, 0x00000000},
1325 {0x0000007a, 0x21000409},
1326 {0x0000007c, 0x00000000},
1327 {0x0000007d, 0xe8000000},
1328 {0x0000007e, 0x044408a8},
1329 {0x0000007f, 0x00000003},
1330 {0x00000080, 0x00000000},
1331 {0x00000081, 0x01000000},
1332 {0x00000082, 0x02000000},
1333 {0x00000083, 0x00000000},
1334 {0x00000084, 0xe3f3e4f4},
1335 {0x00000085, 0x00052024},
1336 {0x00000087, 0x00000000},
1337 {0x00000088, 0x66036603},
1338 {0x00000089, 0x01000000},
1339 {0x0000008b, 0x1c0a0000},
1340 {0x0000008c, 0xff010000},
1341 {0x0000008e, 0xffffefff},
1342 {0x0000008f, 0xfff3efff},
1343 {0x00000090, 0xfff3efbf},
1344 {0x00000094, 0x00101101},
1345 {0x00000095, 0x00000fff},
1346 {0x00000096, 0x00116fff},
1347 {0x00000097, 0x60010000},
1348 {0x00000098, 0x10010000},
1349 {0x00000099, 0x00006000},
1350 {0x0000009a, 0x00001000},
1351 {0x0000009f, 0x00a77400}
1352};
1353
1354static const u32 pitcairn_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
1355 {0x0000006f, 0x03044000},
1356 {0x00000070, 0x0480c018},
1357 {0x00000071, 0x00000040},
1358 {0x00000072, 0x01000000},
1359 {0x00000074, 0x000000ff},
1360 {0x00000075, 0x00143400},
1361 {0x00000076, 0x08ec0800},
1362 {0x00000077, 0x040000cc},
1363 {0x00000079, 0x00000000},
1364 {0x0000007a, 0x21000409},
1365 {0x0000007c, 0x00000000},
1366 {0x0000007d, 0xe8000000},
1367 {0x0000007e, 0x044408a8},
1368 {0x0000007f, 0x00000003},
1369 {0x00000080, 0x00000000},
1370 {0x00000081, 0x01000000},
1371 {0x00000082, 0x02000000},
1372 {0x00000083, 0x00000000},
1373 {0x00000084, 0xe3f3e4f4},
1374 {0x00000085, 0x00052024},
1375 {0x00000087, 0x00000000},
1376 {0x00000088, 0x66036603},
1377 {0x00000089, 0x01000000},
1378 {0x0000008b, 0x1c0a0000},
1379 {0x0000008c, 0xff010000},
1380 {0x0000008e, 0xffffefff},
1381 {0x0000008f, 0xfff3efff},
1382 {0x00000090, 0xfff3efbf},
1383 {0x00000094, 0x00101101},
1384 {0x00000095, 0x00000fff},
1385 {0x00000096, 0x00116fff},
1386 {0x00000097, 0x60010000},
1387 {0x00000098, 0x10010000},
1388 {0x00000099, 0x00006000},
1389 {0x0000009a, 0x00001000},
1390 {0x0000009f, 0x00a47400}
1391};
1392
1393static const u32 verde_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
1394 {0x0000006f, 0x03044000},
1395 {0x00000070, 0x0480c018},
1396 {0x00000071, 0x00000040},
1397 {0x00000072, 0x01000000},
1398 {0x00000074, 0x000000ff},
1399 {0x00000075, 0x00143400},
1400 {0x00000076, 0x08ec0800},
1401 {0x00000077, 0x040000cc},
1402 {0x00000079, 0x00000000},
1403 {0x0000007a, 0x21000409},
1404 {0x0000007c, 0x00000000},
1405 {0x0000007d, 0xe8000000},
1406 {0x0000007e, 0x044408a8},
1407 {0x0000007f, 0x00000003},
1408 {0x00000080, 0x00000000},
1409 {0x00000081, 0x01000000},
1410 {0x00000082, 0x02000000},
1411 {0x00000083, 0x00000000},
1412 {0x00000084, 0xe3f3e4f4},
1413 {0x00000085, 0x00052024},
1414 {0x00000087, 0x00000000},
1415 {0x00000088, 0x66036603},
1416 {0x00000089, 0x01000000},
1417 {0x0000008b, 0x1c0a0000},
1418 {0x0000008c, 0xff010000},
1419 {0x0000008e, 0xffffefff},
1420 {0x0000008f, 0xfff3efff},
1421 {0x00000090, 0xfff3efbf},
1422 {0x00000094, 0x00101101},
1423 {0x00000095, 0x00000fff},
1424 {0x00000096, 0x00116fff},
1425 {0x00000097, 0x60010000},
1426 {0x00000098, 0x10010000},
1427 {0x00000099, 0x00006000},
1428 {0x0000009a, 0x00001000},
1429 {0x0000009f, 0x00a37400}
1430};
1431
bcc7f5d2
AD
1432static const u32 oland_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
1433 {0x0000006f, 0x03044000},
1434 {0x00000070, 0x0480c018},
1435 {0x00000071, 0x00000040},
1436 {0x00000072, 0x01000000},
1437 {0x00000074, 0x000000ff},
1438 {0x00000075, 0x00143400},
1439 {0x00000076, 0x08ec0800},
1440 {0x00000077, 0x040000cc},
1441 {0x00000079, 0x00000000},
1442 {0x0000007a, 0x21000409},
1443 {0x0000007c, 0x00000000},
1444 {0x0000007d, 0xe8000000},
1445 {0x0000007e, 0x044408a8},
1446 {0x0000007f, 0x00000003},
1447 {0x00000080, 0x00000000},
1448 {0x00000081, 0x01000000},
1449 {0x00000082, 0x02000000},
1450 {0x00000083, 0x00000000},
1451 {0x00000084, 0xe3f3e4f4},
1452 {0x00000085, 0x00052024},
1453 {0x00000087, 0x00000000},
1454 {0x00000088, 0x66036603},
1455 {0x00000089, 0x01000000},
1456 {0x0000008b, 0x1c0a0000},
1457 {0x0000008c, 0xff010000},
1458 {0x0000008e, 0xffffefff},
1459 {0x0000008f, 0xfff3efff},
1460 {0x00000090, 0xfff3efbf},
1461 {0x00000094, 0x00101101},
1462 {0x00000095, 0x00000fff},
1463 {0x00000096, 0x00116fff},
1464 {0x00000097, 0x60010000},
1465 {0x00000098, 0x10010000},
1466 {0x00000099, 0x00006000},
1467 {0x0000009a, 0x00001000},
1468 {0x0000009f, 0x00a17730}
1469};
1470
c04c00b4
AD
1471static const u32 hainan_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
1472 {0x0000006f, 0x03044000},
1473 {0x00000070, 0x0480c018},
1474 {0x00000071, 0x00000040},
1475 {0x00000072, 0x01000000},
1476 {0x00000074, 0x000000ff},
1477 {0x00000075, 0x00143400},
1478 {0x00000076, 0x08ec0800},
1479 {0x00000077, 0x040000cc},
1480 {0x00000079, 0x00000000},
1481 {0x0000007a, 0x21000409},
1482 {0x0000007c, 0x00000000},
1483 {0x0000007d, 0xe8000000},
1484 {0x0000007e, 0x044408a8},
1485 {0x0000007f, 0x00000003},
1486 {0x00000080, 0x00000000},
1487 {0x00000081, 0x01000000},
1488 {0x00000082, 0x02000000},
1489 {0x00000083, 0x00000000},
1490 {0x00000084, 0xe3f3e4f4},
1491 {0x00000085, 0x00052024},
1492 {0x00000087, 0x00000000},
1493 {0x00000088, 0x66036603},
1494 {0x00000089, 0x01000000},
1495 {0x0000008b, 0x1c0a0000},
1496 {0x0000008c, 0xff010000},
1497 {0x0000008e, 0xffffefff},
1498 {0x0000008f, 0xfff3efff},
1499 {0x00000090, 0xfff3efbf},
1500 {0x00000094, 0x00101101},
1501 {0x00000095, 0x00000fff},
1502 {0x00000096, 0x00116fff},
1503 {0x00000097, 0x60010000},
1504 {0x00000098, 0x10010000},
1505 {0x00000099, 0x00006000},
1506 {0x0000009a, 0x00001000},
1507 {0x0000009f, 0x00a07730}
1508};
1509
8b074dd6 1510/* ucode loading */
6c7bccea 1511int si_mc_load_microcode(struct radeon_device *rdev)
8b074dd6 1512{
629bd33c
AD
1513 const __be32 *fw_data = NULL;
1514 const __le32 *new_fw_data = NULL;
8b074dd6 1515 u32 running, blackout = 0;
629bd33c
AD
1516 u32 *io_mc_regs = NULL;
1517 const __le32 *new_io_mc_regs = NULL;
8c79bae6 1518 int i, regs_size, ucode_size;
8b074dd6
AD
1519
1520 if (!rdev->mc_fw)
1521 return -EINVAL;
1522
629bd33c
AD
1523 if (rdev->new_fw) {
1524 const struct mc_firmware_header_v1_0 *hdr =
1525 (const struct mc_firmware_header_v1_0 *)rdev->mc_fw->data;
1526
1527 radeon_ucode_print_mc_hdr(&hdr->header);
1528 regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
1529 new_io_mc_regs = (const __le32 *)
1530 (rdev->mc_fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
1531 ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
1532 new_fw_data = (const __le32 *)
1533 (rdev->mc_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1534 } else {
1535 ucode_size = rdev->mc_fw->size / 4;
8c79bae6 1536
629bd33c
AD
1537 switch (rdev->family) {
1538 case CHIP_TAHITI:
1539 io_mc_regs = (u32 *)&tahiti_io_mc_regs;
1540 regs_size = TAHITI_IO_MC_REGS_SIZE;
1541 break;
1542 case CHIP_PITCAIRN:
1543 io_mc_regs = (u32 *)&pitcairn_io_mc_regs;
1544 regs_size = TAHITI_IO_MC_REGS_SIZE;
1545 break;
1546 case CHIP_VERDE:
1547 default:
1548 io_mc_regs = (u32 *)&verde_io_mc_regs;
1549 regs_size = TAHITI_IO_MC_REGS_SIZE;
1550 break;
1551 case CHIP_OLAND:
1552 io_mc_regs = (u32 *)&oland_io_mc_regs;
1553 regs_size = TAHITI_IO_MC_REGS_SIZE;
1554 break;
1555 case CHIP_HAINAN:
1556 io_mc_regs = (u32 *)&hainan_io_mc_regs;
1557 regs_size = TAHITI_IO_MC_REGS_SIZE;
1558 break;
1559 }
1560 fw_data = (const __be32 *)rdev->mc_fw->data;
8b074dd6
AD
1561 }
1562
1563 running = RREG32(MC_SEQ_SUP_CNTL) & RUN_MASK;
1564
1565 if (running == 0) {
1566 if (running) {
1567 blackout = RREG32(MC_SHARED_BLACKOUT_CNTL);
1568 WREG32(MC_SHARED_BLACKOUT_CNTL, blackout | 1);
1569 }
1570
1571 /* reset the engine and set to writable */
1572 WREG32(MC_SEQ_SUP_CNTL, 0x00000008);
1573 WREG32(MC_SEQ_SUP_CNTL, 0x00000010);
1574
1575 /* load mc io regs */
1576 for (i = 0; i < regs_size; i++) {
629bd33c
AD
1577 if (rdev->new_fw) {
1578 WREG32(MC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(new_io_mc_regs++));
1579 WREG32(MC_SEQ_IO_DEBUG_DATA, le32_to_cpup(new_io_mc_regs++));
1580 } else {
1581 WREG32(MC_SEQ_IO_DEBUG_INDEX, io_mc_regs[(i << 1)]);
1582 WREG32(MC_SEQ_IO_DEBUG_DATA, io_mc_regs[(i << 1) + 1]);
1583 }
8b074dd6
AD
1584 }
1585 /* load the MC ucode */
629bd33c
AD
1586 for (i = 0; i < ucode_size; i++) {
1587 if (rdev->new_fw)
1588 WREG32(MC_SEQ_SUP_PGM, le32_to_cpup(new_fw_data++));
1589 else
1590 WREG32(MC_SEQ_SUP_PGM, be32_to_cpup(fw_data++));
1591 }
8b074dd6
AD
1592
1593 /* put the engine back into the active state */
1594 WREG32(MC_SEQ_SUP_CNTL, 0x00000008);
1595 WREG32(MC_SEQ_SUP_CNTL, 0x00000004);
1596 WREG32(MC_SEQ_SUP_CNTL, 0x00000001);
1597
1598 /* wait for training to complete */
1599 for (i = 0; i < rdev->usec_timeout; i++) {
1600 if (RREG32(MC_SEQ_TRAIN_WAKEUP_CNTL) & TRAIN_DONE_D0)
1601 break;
1602 udelay(1);
1603 }
1604 for (i = 0; i < rdev->usec_timeout; i++) {
1605 if (RREG32(MC_SEQ_TRAIN_WAKEUP_CNTL) & TRAIN_DONE_D1)
1606 break;
1607 udelay(1);
1608 }
1609
1610 if (running)
1611 WREG32(MC_SHARED_BLACKOUT_CNTL, blackout);
1612 }
1613
1614 return 0;
1615}
1616
0f0de06c
AD
1617static int si_init_microcode(struct radeon_device *rdev)
1618{
0f0de06c 1619 const char *chip_name;
629bd33c 1620 const char *new_chip_name;
0f0de06c 1621 size_t pfp_req_size, me_req_size, ce_req_size, rlc_req_size, mc_req_size;
1ebe9280 1622 size_t smc_req_size, mc2_req_size;
0f0de06c
AD
1623 char fw_name[30];
1624 int err;
629bd33c 1625 int new_fw = 0;
0f0de06c
AD
1626
1627 DRM_DEBUG("\n");
1628
0f0de06c
AD
1629 switch (rdev->family) {
1630 case CHIP_TAHITI:
1631 chip_name = "TAHITI";
629bd33c 1632 new_chip_name = "tahiti";
0f0de06c
AD
1633 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
1634 me_req_size = SI_PM4_UCODE_SIZE * 4;
1635 ce_req_size = SI_CE_UCODE_SIZE * 4;
1636 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
1637 mc_req_size = SI_MC_UCODE_SIZE * 4;
1ebe9280 1638 mc2_req_size = TAHITI_MC_UCODE_SIZE * 4;
a9e61410 1639 smc_req_size = ALIGN(TAHITI_SMC_UCODE_SIZE, 4);
0f0de06c
AD
1640 break;
1641 case CHIP_PITCAIRN:
1642 chip_name = "PITCAIRN";
629bd33c 1643 new_chip_name = "pitcairn";
0f0de06c
AD
1644 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
1645 me_req_size = SI_PM4_UCODE_SIZE * 4;
1646 ce_req_size = SI_CE_UCODE_SIZE * 4;
1647 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
1648 mc_req_size = SI_MC_UCODE_SIZE * 4;
1ebe9280 1649 mc2_req_size = PITCAIRN_MC_UCODE_SIZE * 4;
a9e61410 1650 smc_req_size = ALIGN(PITCAIRN_SMC_UCODE_SIZE, 4);
0f0de06c
AD
1651 break;
1652 case CHIP_VERDE:
1653 chip_name = "VERDE";
629bd33c 1654 new_chip_name = "verde";
0f0de06c
AD
1655 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
1656 me_req_size = SI_PM4_UCODE_SIZE * 4;
1657 ce_req_size = SI_CE_UCODE_SIZE * 4;
1658 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
1659 mc_req_size = SI_MC_UCODE_SIZE * 4;
1ebe9280 1660 mc2_req_size = VERDE_MC_UCODE_SIZE * 4;
a9e61410 1661 smc_req_size = ALIGN(VERDE_SMC_UCODE_SIZE, 4);
0f0de06c 1662 break;
bcc7f5d2
AD
1663 case CHIP_OLAND:
1664 chip_name = "OLAND";
629bd33c 1665 new_chip_name = "oland";
bcc7f5d2
AD
1666 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
1667 me_req_size = SI_PM4_UCODE_SIZE * 4;
1668 ce_req_size = SI_CE_UCODE_SIZE * 4;
1669 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
1ebe9280 1670 mc_req_size = mc2_req_size = OLAND_MC_UCODE_SIZE * 4;
a9e61410 1671 smc_req_size = ALIGN(OLAND_SMC_UCODE_SIZE, 4);
bcc7f5d2 1672 break;
c04c00b4
AD
1673 case CHIP_HAINAN:
1674 chip_name = "HAINAN";
629bd33c 1675 new_chip_name = "hainan";
c04c00b4
AD
1676 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
1677 me_req_size = SI_PM4_UCODE_SIZE * 4;
1678 ce_req_size = SI_CE_UCODE_SIZE * 4;
1679 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
1ebe9280 1680 mc_req_size = mc2_req_size = OLAND_MC_UCODE_SIZE * 4;
a9e61410 1681 smc_req_size = ALIGN(HAINAN_SMC_UCODE_SIZE, 4);
c04c00b4 1682 break;
0f0de06c
AD
1683 default: BUG();
1684 }
1685
629bd33c 1686 DRM_INFO("Loading %s Microcode\n", new_chip_name);
0f0de06c 1687
629bd33c 1688 snprintf(fw_name, sizeof(fw_name), "radeon/%s_pfp.bin", new_chip_name);
0a168933 1689 err = request_firmware(&rdev->pfp_fw, fw_name, rdev->dev);
629bd33c
AD
1690 if (err) {
1691 snprintf(fw_name, sizeof(fw_name), "radeon/%s_pfp.bin", chip_name);
1692 err = request_firmware(&rdev->pfp_fw, fw_name, rdev->dev);
1693 if (err)
1694 goto out;
1695 if (rdev->pfp_fw->size != pfp_req_size) {
1696 printk(KERN_ERR
1697 "si_cp: Bogus length %zu in firmware \"%s\"\n",
1698 rdev->pfp_fw->size, fw_name);
1699 err = -EINVAL;
1700 goto out;
1701 }
1702 } else {
1703 err = radeon_ucode_validate(rdev->pfp_fw);
1704 if (err) {
1705 printk(KERN_ERR
1706 "si_cp: validation failed for firmware \"%s\"\n",
1707 fw_name);
1708 goto out;
1709 } else {
1710 new_fw++;
1711 }
0f0de06c
AD
1712 }
1713
629bd33c 1714 snprintf(fw_name, sizeof(fw_name), "radeon/%s_me.bin", new_chip_name);
0a168933 1715 err = request_firmware(&rdev->me_fw, fw_name, rdev->dev);
629bd33c
AD
1716 if (err) {
1717 snprintf(fw_name, sizeof(fw_name), "radeon/%s_me.bin", chip_name);
1718 err = request_firmware(&rdev->me_fw, fw_name, rdev->dev);
1719 if (err)
1720 goto out;
1721 if (rdev->me_fw->size != me_req_size) {
1722 printk(KERN_ERR
1723 "si_cp: Bogus length %zu in firmware \"%s\"\n",
1724 rdev->me_fw->size, fw_name);
1725 err = -EINVAL;
1726 }
1727 } else {
1728 err = radeon_ucode_validate(rdev->me_fw);
1729 if (err) {
1730 printk(KERN_ERR
1731 "si_cp: validation failed for firmware \"%s\"\n",
1732 fw_name);
1733 goto out;
1734 } else {
1735 new_fw++;
1736 }
0f0de06c
AD
1737 }
1738
629bd33c 1739 snprintf(fw_name, sizeof(fw_name), "radeon/%s_ce.bin", new_chip_name);
0a168933 1740 err = request_firmware(&rdev->ce_fw, fw_name, rdev->dev);
629bd33c
AD
1741 if (err) {
1742 snprintf(fw_name, sizeof(fw_name), "radeon/%s_ce.bin", chip_name);
1743 err = request_firmware(&rdev->ce_fw, fw_name, rdev->dev);
1744 if (err)
1745 goto out;
1746 if (rdev->ce_fw->size != ce_req_size) {
1747 printk(KERN_ERR
1748 "si_cp: Bogus length %zu in firmware \"%s\"\n",
1749 rdev->ce_fw->size, fw_name);
1750 err = -EINVAL;
1751 }
1752 } else {
1753 err = radeon_ucode_validate(rdev->ce_fw);
1754 if (err) {
1755 printk(KERN_ERR
1756 "si_cp: validation failed for firmware \"%s\"\n",
1757 fw_name);
1758 goto out;
1759 } else {
1760 new_fw++;
1761 }
0f0de06c
AD
1762 }
1763
629bd33c 1764 snprintf(fw_name, sizeof(fw_name), "radeon/%s_rlc.bin", new_chip_name);
0a168933 1765 err = request_firmware(&rdev->rlc_fw, fw_name, rdev->dev);
629bd33c
AD
1766 if (err) {
1767 snprintf(fw_name, sizeof(fw_name), "radeon/%s_rlc.bin", chip_name);
1768 err = request_firmware(&rdev->rlc_fw, fw_name, rdev->dev);
1769 if (err)
1770 goto out;
1771 if (rdev->rlc_fw->size != rlc_req_size) {
1772 printk(KERN_ERR
1773 "si_rlc: Bogus length %zu in firmware \"%s\"\n",
1774 rdev->rlc_fw->size, fw_name);
1775 err = -EINVAL;
1776 }
1777 } else {
1778 err = radeon_ucode_validate(rdev->rlc_fw);
1779 if (err) {
1780 printk(KERN_ERR
1781 "si_cp: validation failed for firmware \"%s\"\n",
1782 fw_name);
1783 goto out;
1784 } else {
1785 new_fw++;
1786 }
0f0de06c
AD
1787 }
1788
629bd33c 1789 snprintf(fw_name, sizeof(fw_name), "radeon/%s_mc.bin", new_chip_name);
0a168933 1790 err = request_firmware(&rdev->mc_fw, fw_name, rdev->dev);
1ebe9280 1791 if (err) {
629bd33c 1792 snprintf(fw_name, sizeof(fw_name), "radeon/%s_mc2.bin", chip_name);
1ebe9280 1793 err = request_firmware(&rdev->mc_fw, fw_name, rdev->dev);
629bd33c
AD
1794 if (err) {
1795 snprintf(fw_name, sizeof(fw_name), "radeon/%s_mc.bin", chip_name);
1796 err = request_firmware(&rdev->mc_fw, fw_name, rdev->dev);
1797 if (err)
1798 goto out;
1799 }
1800 if ((rdev->mc_fw->size != mc_req_size) &&
1801 (rdev->mc_fw->size != mc2_req_size)) {
1802 printk(KERN_ERR
1803 "si_mc: Bogus length %zu in firmware \"%s\"\n",
1804 rdev->mc_fw->size, fw_name);
1805 err = -EINVAL;
1806 }
1807 DRM_INFO("%s: %zu bytes\n", fw_name, rdev->mc_fw->size);
1808 } else {
1809 err = radeon_ucode_validate(rdev->mc_fw);
1810 if (err) {
1811 printk(KERN_ERR
1812 "si_cp: validation failed for firmware \"%s\"\n",
1813 fw_name);
1ebe9280 1814 goto out;
629bd33c
AD
1815 } else {
1816 new_fw++;
1817 }
1ebe9280 1818 }
0f0de06c 1819
629bd33c 1820 snprintf(fw_name, sizeof(fw_name), "radeon/%s_smc.bin", new_chip_name);
0a168933 1821 err = request_firmware(&rdev->smc_fw, fw_name, rdev->dev);
8a53fa23 1822 if (err) {
629bd33c
AD
1823 snprintf(fw_name, sizeof(fw_name), "radeon/%s_smc.bin", chip_name);
1824 err = request_firmware(&rdev->smc_fw, fw_name, rdev->dev);
1825 if (err) {
1826 printk(KERN_ERR
1827 "smc: error loading firmware \"%s\"\n",
1828 fw_name);
1829 release_firmware(rdev->smc_fw);
1830 rdev->smc_fw = NULL;
1831 err = 0;
1832 } else if (rdev->smc_fw->size != smc_req_size) {
1833 printk(KERN_ERR
1834 "si_smc: Bogus length %zu in firmware \"%s\"\n",
1835 rdev->smc_fw->size, fw_name);
1836 err = -EINVAL;
1837 }
1838 } else {
1839 err = radeon_ucode_validate(rdev->smc_fw);
1840 if (err) {
1841 printk(KERN_ERR
1842 "si_cp: validation failed for firmware \"%s\"\n",
1843 fw_name);
1844 goto out;
1845 } else {
1846 new_fw++;
1847 }
a9e61410
AD
1848 }
1849
629bd33c
AD
1850 if (new_fw == 0) {
1851 rdev->new_fw = false;
1852 } else if (new_fw < 6) {
1853 printk(KERN_ERR "si_fw: mixing new and old firmware!\n");
1854 err = -EINVAL;
1855 } else {
1856 rdev->new_fw = true;
1857 }
0f0de06c 1858out:
0f0de06c
AD
1859 if (err) {
1860 if (err != -EINVAL)
1861 printk(KERN_ERR
1862 "si_cp: Failed to load firmware \"%s\"\n",
1863 fw_name);
1864 release_firmware(rdev->pfp_fw);
1865 rdev->pfp_fw = NULL;
1866 release_firmware(rdev->me_fw);
1867 rdev->me_fw = NULL;
1868 release_firmware(rdev->ce_fw);
1869 rdev->ce_fw = NULL;
1870 release_firmware(rdev->rlc_fw);
1871 rdev->rlc_fw = NULL;
1872 release_firmware(rdev->mc_fw);
1873 rdev->mc_fw = NULL;
a9e61410
AD
1874 release_firmware(rdev->smc_fw);
1875 rdev->smc_fw = NULL;
0f0de06c
AD
1876 }
1877 return err;
1878}
1879
43b3cd99
AD
1880/* watermark setup */
1881static u32 dce6_line_buffer_adjust(struct radeon_device *rdev,
1882 struct radeon_crtc *radeon_crtc,
1883 struct drm_display_mode *mode,
1884 struct drm_display_mode *other_mode)
1885{
290d2457
AD
1886 u32 tmp, buffer_alloc, i;
1887 u32 pipe_offset = radeon_crtc->crtc_id * 0x20;
43b3cd99
AD
1888 /*
1889 * Line Buffer Setup
1890 * There are 3 line buffers, each one shared by 2 display controllers.
1891 * DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
1892 * the display controllers. The paritioning is done via one of four
1893 * preset allocations specified in bits 21:20:
1894 * 0 - half lb
1895 * 2 - whole lb, other crtc must be disabled
1896 */
1897 /* this can get tricky if we have two large displays on a paired group
1898 * of crtcs. Ideally for multiple large displays we'd assign them to
1899 * non-linked crtcs for maximum line buffer allocation.
1900 */
1901 if (radeon_crtc->base.enabled && mode) {
290d2457 1902 if (other_mode) {
43b3cd99 1903 tmp = 0; /* 1/2 */
290d2457
AD
1904 buffer_alloc = 1;
1905 } else {
43b3cd99 1906 tmp = 2; /* whole */
290d2457
AD
1907 buffer_alloc = 2;
1908 }
1909 } else {
43b3cd99 1910 tmp = 0;
290d2457
AD
1911 buffer_alloc = 0;
1912 }
43b3cd99
AD
1913
1914 WREG32(DC_LB_MEMORY_SPLIT + radeon_crtc->crtc_offset,
1915 DC_LB_MEMORY_CONFIG(tmp));
1916
290d2457
AD
1917 WREG32(PIPE0_DMIF_BUFFER_CONTROL + pipe_offset,
1918 DMIF_BUFFERS_ALLOCATED(buffer_alloc));
1919 for (i = 0; i < rdev->usec_timeout; i++) {
1920 if (RREG32(PIPE0_DMIF_BUFFER_CONTROL + pipe_offset) &
1921 DMIF_BUFFERS_ALLOCATED_COMPLETED)
1922 break;
1923 udelay(1);
1924 }
1925
43b3cd99
AD
1926 if (radeon_crtc->base.enabled && mode) {
1927 switch (tmp) {
1928 case 0:
1929 default:
1930 return 4096 * 2;
1931 case 2:
1932 return 8192 * 2;
1933 }
1934 }
1935
1936 /* controller not enabled, so no lb used */
1937 return 0;
1938}
1939
ca7db22b 1940static u32 si_get_number_of_dram_channels(struct radeon_device *rdev)
43b3cd99
AD
1941{
1942 u32 tmp = RREG32(MC_SHARED_CHMAP);
1943
1944 switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
1945 case 0:
1946 default:
1947 return 1;
1948 case 1:
1949 return 2;
1950 case 2:
1951 return 4;
1952 case 3:
1953 return 8;
1954 case 4:
1955 return 3;
1956 case 5:
1957 return 6;
1958 case 6:
1959 return 10;
1960 case 7:
1961 return 12;
1962 case 8:
1963 return 16;
1964 }
1965}
1966
1967struct dce6_wm_params {
1968 u32 dram_channels; /* number of dram channels */
1969 u32 yclk; /* bandwidth per dram data pin in kHz */
1970 u32 sclk; /* engine clock in kHz */
1971 u32 disp_clk; /* display clock in kHz */
1972 u32 src_width; /* viewport width */
1973 u32 active_time; /* active display time in ns */
1974 u32 blank_time; /* blank time in ns */
1975 bool interlaced; /* mode is interlaced */
1976 fixed20_12 vsc; /* vertical scale ratio */
1977 u32 num_heads; /* number of active crtcs */
1978 u32 bytes_per_pixel; /* bytes per pixel display + overlay */
1979 u32 lb_size; /* line buffer allocated to pipe */
1980 u32 vtaps; /* vertical scaler taps */
1981};
1982
1983static u32 dce6_dram_bandwidth(struct dce6_wm_params *wm)
1984{
1985 /* Calculate raw DRAM Bandwidth */
1986 fixed20_12 dram_efficiency; /* 0.7 */
1987 fixed20_12 yclk, dram_channels, bandwidth;
1988 fixed20_12 a;
1989
1990 a.full = dfixed_const(1000);
1991 yclk.full = dfixed_const(wm->yclk);
1992 yclk.full = dfixed_div(yclk, a);
1993 dram_channels.full = dfixed_const(wm->dram_channels * 4);
1994 a.full = dfixed_const(10);
1995 dram_efficiency.full = dfixed_const(7);
1996 dram_efficiency.full = dfixed_div(dram_efficiency, a);
1997 bandwidth.full = dfixed_mul(dram_channels, yclk);
1998 bandwidth.full = dfixed_mul(bandwidth, dram_efficiency);
1999
2000 return dfixed_trunc(bandwidth);
2001}
2002
2003static u32 dce6_dram_bandwidth_for_display(struct dce6_wm_params *wm)
2004{
2005 /* Calculate DRAM Bandwidth and the part allocated to display. */
2006 fixed20_12 disp_dram_allocation; /* 0.3 to 0.7 */
2007 fixed20_12 yclk, dram_channels, bandwidth;
2008 fixed20_12 a;
2009
2010 a.full = dfixed_const(1000);
2011 yclk.full = dfixed_const(wm->yclk);
2012 yclk.full = dfixed_div(yclk, a);
2013 dram_channels.full = dfixed_const(wm->dram_channels * 4);
2014 a.full = dfixed_const(10);
2015 disp_dram_allocation.full = dfixed_const(3); /* XXX worse case value 0.3 */
2016 disp_dram_allocation.full = dfixed_div(disp_dram_allocation, a);
2017 bandwidth.full = dfixed_mul(dram_channels, yclk);
2018 bandwidth.full = dfixed_mul(bandwidth, disp_dram_allocation);
2019
2020 return dfixed_trunc(bandwidth);
2021}
2022
2023static u32 dce6_data_return_bandwidth(struct dce6_wm_params *wm)
2024{
2025 /* Calculate the display Data return Bandwidth */
2026 fixed20_12 return_efficiency; /* 0.8 */
2027 fixed20_12 sclk, bandwidth;
2028 fixed20_12 a;
2029
2030 a.full = dfixed_const(1000);
2031 sclk.full = dfixed_const(wm->sclk);
2032 sclk.full = dfixed_div(sclk, a);
2033 a.full = dfixed_const(10);
2034 return_efficiency.full = dfixed_const(8);
2035 return_efficiency.full = dfixed_div(return_efficiency, a);
2036 a.full = dfixed_const(32);
2037 bandwidth.full = dfixed_mul(a, sclk);
2038 bandwidth.full = dfixed_mul(bandwidth, return_efficiency);
2039
2040 return dfixed_trunc(bandwidth);
2041}
2042
2043static u32 dce6_get_dmif_bytes_per_request(struct dce6_wm_params *wm)
2044{
2045 return 32;
2046}
2047
2048static u32 dce6_dmif_request_bandwidth(struct dce6_wm_params *wm)
2049{
2050 /* Calculate the DMIF Request Bandwidth */
2051 fixed20_12 disp_clk_request_efficiency; /* 0.8 */
2052 fixed20_12 disp_clk, sclk, bandwidth;
2053 fixed20_12 a, b1, b2;
2054 u32 min_bandwidth;
2055
2056 a.full = dfixed_const(1000);
2057 disp_clk.full = dfixed_const(wm->disp_clk);
2058 disp_clk.full = dfixed_div(disp_clk, a);
2059 a.full = dfixed_const(dce6_get_dmif_bytes_per_request(wm) / 2);
2060 b1.full = dfixed_mul(a, disp_clk);
2061
2062 a.full = dfixed_const(1000);
2063 sclk.full = dfixed_const(wm->sclk);
2064 sclk.full = dfixed_div(sclk, a);
2065 a.full = dfixed_const(dce6_get_dmif_bytes_per_request(wm));
2066 b2.full = dfixed_mul(a, sclk);
2067
2068 a.full = dfixed_const(10);
2069 disp_clk_request_efficiency.full = dfixed_const(8);
2070 disp_clk_request_efficiency.full = dfixed_div(disp_clk_request_efficiency, a);
2071
2072 min_bandwidth = min(dfixed_trunc(b1), dfixed_trunc(b2));
2073
2074 a.full = dfixed_const(min_bandwidth);
2075 bandwidth.full = dfixed_mul(a, disp_clk_request_efficiency);
2076
2077 return dfixed_trunc(bandwidth);
2078}
2079
2080static u32 dce6_available_bandwidth(struct dce6_wm_params *wm)
2081{
2082 /* Calculate the Available bandwidth. Display can use this temporarily but not in average. */
2083 u32 dram_bandwidth = dce6_dram_bandwidth(wm);
2084 u32 data_return_bandwidth = dce6_data_return_bandwidth(wm);
2085 u32 dmif_req_bandwidth = dce6_dmif_request_bandwidth(wm);
2086
2087 return min(dram_bandwidth, min(data_return_bandwidth, dmif_req_bandwidth));
2088}
2089
2090static u32 dce6_average_bandwidth(struct dce6_wm_params *wm)
2091{
2092 /* Calculate the display mode Average Bandwidth
2093 * DisplayMode should contain the source and destination dimensions,
2094 * timing, etc.
2095 */
2096 fixed20_12 bpp;
2097 fixed20_12 line_time;
2098 fixed20_12 src_width;
2099 fixed20_12 bandwidth;
2100 fixed20_12 a;
2101
2102 a.full = dfixed_const(1000);
2103 line_time.full = dfixed_const(wm->active_time + wm->blank_time);
2104 line_time.full = dfixed_div(line_time, a);
2105 bpp.full = dfixed_const(wm->bytes_per_pixel);
2106 src_width.full = dfixed_const(wm->src_width);
2107 bandwidth.full = dfixed_mul(src_width, bpp);
2108 bandwidth.full = dfixed_mul(bandwidth, wm->vsc);
2109 bandwidth.full = dfixed_div(bandwidth, line_time);
2110
2111 return dfixed_trunc(bandwidth);
2112}
2113
2114static u32 dce6_latency_watermark(struct dce6_wm_params *wm)
2115{
2116 /* First calcualte the latency in ns */
2117 u32 mc_latency = 2000; /* 2000 ns. */
2118 u32 available_bandwidth = dce6_available_bandwidth(wm);
2119 u32 worst_chunk_return_time = (512 * 8 * 1000) / available_bandwidth;
2120 u32 cursor_line_pair_return_time = (128 * 4 * 1000) / available_bandwidth;
2121 u32 dc_latency = 40000000 / wm->disp_clk; /* dc pipe latency */
2122 u32 other_heads_data_return_time = ((wm->num_heads + 1) * worst_chunk_return_time) +
2123 (wm->num_heads * cursor_line_pair_return_time);
2124 u32 latency = mc_latency + other_heads_data_return_time + dc_latency;
2125 u32 max_src_lines_per_dst_line, lb_fill_bw, line_fill_time;
2126 u32 tmp, dmif_size = 12288;
2127 fixed20_12 a, b, c;
2128
2129 if (wm->num_heads == 0)
2130 return 0;
2131
2132 a.full = dfixed_const(2);
2133 b.full = dfixed_const(1);
2134 if ((wm->vsc.full > a.full) ||
2135 ((wm->vsc.full > b.full) && (wm->vtaps >= 3)) ||
2136 (wm->vtaps >= 5) ||
2137 ((wm->vsc.full >= a.full) && wm->interlaced))
2138 max_src_lines_per_dst_line = 4;
2139 else
2140 max_src_lines_per_dst_line = 2;
2141
2142 a.full = dfixed_const(available_bandwidth);
2143 b.full = dfixed_const(wm->num_heads);
2144 a.full = dfixed_div(a, b);
2145
2146 b.full = dfixed_const(mc_latency + 512);
2147 c.full = dfixed_const(wm->disp_clk);
2148 b.full = dfixed_div(b, c);
2149
2150 c.full = dfixed_const(dmif_size);
2151 b.full = dfixed_div(c, b);
2152
2153 tmp = min(dfixed_trunc(a), dfixed_trunc(b));
2154
2155 b.full = dfixed_const(1000);
2156 c.full = dfixed_const(wm->disp_clk);
2157 b.full = dfixed_div(c, b);
2158 c.full = dfixed_const(wm->bytes_per_pixel);
2159 b.full = dfixed_mul(b, c);
2160
2161 lb_fill_bw = min(tmp, dfixed_trunc(b));
2162
2163 a.full = dfixed_const(max_src_lines_per_dst_line * wm->src_width * wm->bytes_per_pixel);
2164 b.full = dfixed_const(1000);
2165 c.full = dfixed_const(lb_fill_bw);
2166 b.full = dfixed_div(c, b);
2167 a.full = dfixed_div(a, b);
2168 line_fill_time = dfixed_trunc(a);
2169
2170 if (line_fill_time < wm->active_time)
2171 return latency;
2172 else
2173 return latency + (line_fill_time - wm->active_time);
2174
2175}
2176
2177static bool dce6_average_bandwidth_vs_dram_bandwidth_for_display(struct dce6_wm_params *wm)
2178{
2179 if (dce6_average_bandwidth(wm) <=
2180 (dce6_dram_bandwidth_for_display(wm) / wm->num_heads))
2181 return true;
2182 else
2183 return false;
2184};
2185
2186static bool dce6_average_bandwidth_vs_available_bandwidth(struct dce6_wm_params *wm)
2187{
2188 if (dce6_average_bandwidth(wm) <=
2189 (dce6_available_bandwidth(wm) / wm->num_heads))
2190 return true;
2191 else
2192 return false;
2193};
2194
2195static bool dce6_check_latency_hiding(struct dce6_wm_params *wm)
2196{
2197 u32 lb_partitions = wm->lb_size / wm->src_width;
2198 u32 line_time = wm->active_time + wm->blank_time;
2199 u32 latency_tolerant_lines;
2200 u32 latency_hiding;
2201 fixed20_12 a;
2202
2203 a.full = dfixed_const(1);
2204 if (wm->vsc.full > a.full)
2205 latency_tolerant_lines = 1;
2206 else {
2207 if (lb_partitions <= (wm->vtaps + 1))
2208 latency_tolerant_lines = 1;
2209 else
2210 latency_tolerant_lines = 2;
2211 }
2212
2213 latency_hiding = (latency_tolerant_lines * line_time + wm->blank_time);
2214
2215 if (dce6_latency_watermark(wm) <= latency_hiding)
2216 return true;
2217 else
2218 return false;
2219}
2220
2221static void dce6_program_watermarks(struct radeon_device *rdev,
2222 struct radeon_crtc *radeon_crtc,
2223 u32 lb_size, u32 num_heads)
2224{
2225 struct drm_display_mode *mode = &radeon_crtc->base.mode;
c696e53f
AD
2226 struct dce6_wm_params wm_low, wm_high;
2227 u32 dram_channels;
43b3cd99
AD
2228 u32 pixel_period;
2229 u32 line_time = 0;
2230 u32 latency_watermark_a = 0, latency_watermark_b = 0;
2231 u32 priority_a_mark = 0, priority_b_mark = 0;
2232 u32 priority_a_cnt = PRIORITY_OFF;
2233 u32 priority_b_cnt = PRIORITY_OFF;
2234 u32 tmp, arb_control3;
2235 fixed20_12 a, b, c;
2236
2237 if (radeon_crtc->base.enabled && num_heads && mode) {
2238 pixel_period = 1000000 / (u32)mode->clock;
2239 line_time = min((u32)mode->crtc_htotal * pixel_period, (u32)65535);
2240 priority_a_cnt = 0;
2241 priority_b_cnt = 0;
2242
ca7db22b 2243 if (rdev->family == CHIP_ARUBA)
c696e53f 2244 dram_channels = evergreen_get_number_of_dram_channels(rdev);
ca7db22b 2245 else
c696e53f
AD
2246 dram_channels = si_get_number_of_dram_channels(rdev);
2247
2248 /* watermark for high clocks */
2249 if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
2250 wm_high.yclk =
2251 radeon_dpm_get_mclk(rdev, false) * 10;
2252 wm_high.sclk =
2253 radeon_dpm_get_sclk(rdev, false) * 10;
2254 } else {
2255 wm_high.yclk = rdev->pm.current_mclk * 10;
2256 wm_high.sclk = rdev->pm.current_sclk * 10;
2257 }
2258
2259 wm_high.disp_clk = mode->clock;
2260 wm_high.src_width = mode->crtc_hdisplay;
2261 wm_high.active_time = mode->crtc_hdisplay * pixel_period;
2262 wm_high.blank_time = line_time - wm_high.active_time;
2263 wm_high.interlaced = false;
2264 if (mode->flags & DRM_MODE_FLAG_INTERLACE)
2265 wm_high.interlaced = true;
2266 wm_high.vsc = radeon_crtc->vsc;
2267 wm_high.vtaps = 1;
2268 if (radeon_crtc->rmx_type != RMX_OFF)
2269 wm_high.vtaps = 2;
2270 wm_high.bytes_per_pixel = 4; /* XXX: get this from fb config */
2271 wm_high.lb_size = lb_size;
2272 wm_high.dram_channels = dram_channels;
2273 wm_high.num_heads = num_heads;
2274
2275 /* watermark for low clocks */
2276 if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
2277 wm_low.yclk =
2278 radeon_dpm_get_mclk(rdev, true) * 10;
2279 wm_low.sclk =
2280 radeon_dpm_get_sclk(rdev, true) * 10;
2281 } else {
2282 wm_low.yclk = rdev->pm.current_mclk * 10;
2283 wm_low.sclk = rdev->pm.current_sclk * 10;
2284 }
2285
2286 wm_low.disp_clk = mode->clock;
2287 wm_low.src_width = mode->crtc_hdisplay;
2288 wm_low.active_time = mode->crtc_hdisplay * pixel_period;
2289 wm_low.blank_time = line_time - wm_low.active_time;
2290 wm_low.interlaced = false;
2291 if (mode->flags & DRM_MODE_FLAG_INTERLACE)
2292 wm_low.interlaced = true;
2293 wm_low.vsc = radeon_crtc->vsc;
2294 wm_low.vtaps = 1;
2295 if (radeon_crtc->rmx_type != RMX_OFF)
2296 wm_low.vtaps = 2;
2297 wm_low.bytes_per_pixel = 4; /* XXX: get this from fb config */
2298 wm_low.lb_size = lb_size;
2299 wm_low.dram_channels = dram_channels;
2300 wm_low.num_heads = num_heads;
43b3cd99
AD
2301
2302 /* set for high clocks */
c696e53f 2303 latency_watermark_a = min(dce6_latency_watermark(&wm_high), (u32)65535);
43b3cd99 2304 /* set for low clocks */
c696e53f 2305 latency_watermark_b = min(dce6_latency_watermark(&wm_low), (u32)65535);
43b3cd99
AD
2306
2307 /* possibly force display priority to high */
2308 /* should really do this at mode validation time... */
c696e53f
AD
2309 if (!dce6_average_bandwidth_vs_dram_bandwidth_for_display(&wm_high) ||
2310 !dce6_average_bandwidth_vs_available_bandwidth(&wm_high) ||
2311 !dce6_check_latency_hiding(&wm_high) ||
2312 (rdev->disp_priority == 2)) {
2313 DRM_DEBUG_KMS("force priority to high\n");
2314 priority_a_cnt |= PRIORITY_ALWAYS_ON;
2315 priority_b_cnt |= PRIORITY_ALWAYS_ON;
2316 }
2317 if (!dce6_average_bandwidth_vs_dram_bandwidth_for_display(&wm_low) ||
2318 !dce6_average_bandwidth_vs_available_bandwidth(&wm_low) ||
2319 !dce6_check_latency_hiding(&wm_low) ||
43b3cd99
AD
2320 (rdev->disp_priority == 2)) {
2321 DRM_DEBUG_KMS("force priority to high\n");
2322 priority_a_cnt |= PRIORITY_ALWAYS_ON;
2323 priority_b_cnt |= PRIORITY_ALWAYS_ON;
2324 }
2325
2326 a.full = dfixed_const(1000);
2327 b.full = dfixed_const(mode->clock);
2328 b.full = dfixed_div(b, a);
2329 c.full = dfixed_const(latency_watermark_a);
2330 c.full = dfixed_mul(c, b);
2331 c.full = dfixed_mul(c, radeon_crtc->hsc);
2332 c.full = dfixed_div(c, a);
2333 a.full = dfixed_const(16);
2334 c.full = dfixed_div(c, a);
2335 priority_a_mark = dfixed_trunc(c);
2336 priority_a_cnt |= priority_a_mark & PRIORITY_MARK_MASK;
2337
2338 a.full = dfixed_const(1000);
2339 b.full = dfixed_const(mode->clock);
2340 b.full = dfixed_div(b, a);
2341 c.full = dfixed_const(latency_watermark_b);
2342 c.full = dfixed_mul(c, b);
2343 c.full = dfixed_mul(c, radeon_crtc->hsc);
2344 c.full = dfixed_div(c, a);
2345 a.full = dfixed_const(16);
2346 c.full = dfixed_div(c, a);
2347 priority_b_mark = dfixed_trunc(c);
2348 priority_b_cnt |= priority_b_mark & PRIORITY_MARK_MASK;
2349 }
2350
2351 /* select wm A */
2352 arb_control3 = RREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset);
2353 tmp = arb_control3;
2354 tmp &= ~LATENCY_WATERMARK_MASK(3);
2355 tmp |= LATENCY_WATERMARK_MASK(1);
2356 WREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset, tmp);
2357 WREG32(DPG_PIPE_LATENCY_CONTROL + radeon_crtc->crtc_offset,
2358 (LATENCY_LOW_WATERMARK(latency_watermark_a) |
2359 LATENCY_HIGH_WATERMARK(line_time)));
2360 /* select wm B */
2361 tmp = RREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset);
2362 tmp &= ~LATENCY_WATERMARK_MASK(3);
2363 tmp |= LATENCY_WATERMARK_MASK(2);
2364 WREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset, tmp);
2365 WREG32(DPG_PIPE_LATENCY_CONTROL + radeon_crtc->crtc_offset,
2366 (LATENCY_LOW_WATERMARK(latency_watermark_b) |
2367 LATENCY_HIGH_WATERMARK(line_time)));
2368 /* restore original selection */
2369 WREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset, arb_control3);
2370
2371 /* write the priority marks */
2372 WREG32(PRIORITY_A_CNT + radeon_crtc->crtc_offset, priority_a_cnt);
2373 WREG32(PRIORITY_B_CNT + radeon_crtc->crtc_offset, priority_b_cnt);
2374
7178d2a6
AD
2375 /* save values for DPM */
2376 radeon_crtc->line_time = line_time;
2377 radeon_crtc->wm_high = latency_watermark_a;
2378 radeon_crtc->wm_low = latency_watermark_b;
43b3cd99
AD
2379}
2380
2381void dce6_bandwidth_update(struct radeon_device *rdev)
2382{
2383 struct drm_display_mode *mode0 = NULL;
2384 struct drm_display_mode *mode1 = NULL;
2385 u32 num_heads = 0, lb_size;
2386 int i;
2387
8efe82ca
AD
2388 if (!rdev->mode_info.mode_config_initialized)
2389 return;
2390
43b3cd99
AD
2391 radeon_update_display_priority(rdev);
2392
2393 for (i = 0; i < rdev->num_crtc; i++) {
2394 if (rdev->mode_info.crtcs[i]->base.enabled)
2395 num_heads++;
2396 }
2397 for (i = 0; i < rdev->num_crtc; i += 2) {
2398 mode0 = &rdev->mode_info.crtcs[i]->base.mode;
2399 mode1 = &rdev->mode_info.crtcs[i+1]->base.mode;
2400 lb_size = dce6_line_buffer_adjust(rdev, rdev->mode_info.crtcs[i], mode0, mode1);
2401 dce6_program_watermarks(rdev, rdev->mode_info.crtcs[i], lb_size, num_heads);
2402 lb_size = dce6_line_buffer_adjust(rdev, rdev->mode_info.crtcs[i+1], mode1, mode0);
2403 dce6_program_watermarks(rdev, rdev->mode_info.crtcs[i+1], lb_size, num_heads);
2404 }
2405}
2406
0a96d72b
AD
2407/*
2408 * Core functions
2409 */
0a96d72b
AD
2410static void si_tiling_mode_table_init(struct radeon_device *rdev)
2411{
2412 const u32 num_tile_mode_states = 32;
2413 u32 reg_offset, gb_tile_moden, split_equal_to_row_size;
2414
2415 switch (rdev->config.si.mem_row_size_in_kb) {
2416 case 1:
2417 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_1KB;
2418 break;
2419 case 2:
2420 default:
2421 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_2KB;
2422 break;
2423 case 4:
2424 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_4KB;
2425 break;
2426 }
2427
2428 if ((rdev->family == CHIP_TAHITI) ||
2429 (rdev->family == CHIP_PITCAIRN)) {
2430 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
2431 switch (reg_offset) {
2432 case 0: /* non-AA compressed depth or any compressed stencil */
2433 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2434 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2435 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2436 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2437 NUM_BANKS(ADDR_SURF_16_BANK) |
2438 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2439 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2440 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2441 break;
2442 case 1: /* 2xAA/4xAA compressed depth only */
2443 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2444 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2445 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2446 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2447 NUM_BANKS(ADDR_SURF_16_BANK) |
2448 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2449 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2450 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2451 break;
2452 case 2: /* 8xAA compressed depth only */
2453 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2454 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2455 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2456 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2457 NUM_BANKS(ADDR_SURF_16_BANK) |
2458 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2459 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2460 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2461 break;
2462 case 3: /* 2xAA/4xAA compressed depth with stencil (for depth buffer) */
2463 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2464 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2465 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2466 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2467 NUM_BANKS(ADDR_SURF_16_BANK) |
2468 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2469 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2470 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2471 break;
2472 case 4: /* Maps w/ a dimension less than the 2D macro-tile dimensions (for mipmapped depth textures) */
2473 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2474 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2475 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2476 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2477 NUM_BANKS(ADDR_SURF_16_BANK) |
2478 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2479 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2480 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2481 break;
2482 case 5: /* Uncompressed 16bpp depth - and stencil buffer allocated with it */
2483 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2484 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2485 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2486 TILE_SPLIT(split_equal_to_row_size) |
2487 NUM_BANKS(ADDR_SURF_16_BANK) |
2488 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2489 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2490 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2491 break;
2492 case 6: /* Uncompressed 32bpp depth - and stencil buffer allocated with it */
2493 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2494 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2495 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2496 TILE_SPLIT(split_equal_to_row_size) |
2497 NUM_BANKS(ADDR_SURF_16_BANK) |
2498 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2499 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2500 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2501 break;
2502 case 7: /* Uncompressed 8bpp stencil without depth (drivers typically do not use) */
2503 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2504 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2505 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2506 TILE_SPLIT(split_equal_to_row_size) |
2507 NUM_BANKS(ADDR_SURF_16_BANK) |
2508 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2509 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2510 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2511 break;
2512 case 8: /* 1D and 1D Array Surfaces */
2513 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
2514 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2515 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2516 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2517 NUM_BANKS(ADDR_SURF_16_BANK) |
2518 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2519 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2520 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2521 break;
2522 case 9: /* Displayable maps. */
2523 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2524 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2525 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2526 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2527 NUM_BANKS(ADDR_SURF_16_BANK) |
2528 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2529 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2530 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2531 break;
2532 case 10: /* Display 8bpp. */
2533 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2534 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2535 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2536 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2537 NUM_BANKS(ADDR_SURF_16_BANK) |
2538 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2539 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2540 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2541 break;
2542 case 11: /* Display 16bpp. */
2543 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2544 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2545 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2546 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2547 NUM_BANKS(ADDR_SURF_16_BANK) |
2548 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2549 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2550 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2551 break;
2552 case 12: /* Display 32bpp. */
2553 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2554 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2555 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2556 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2557 NUM_BANKS(ADDR_SURF_16_BANK) |
2558 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2559 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2560 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2561 break;
2562 case 13: /* Thin. */
2563 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2564 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2565 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2566 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2567 NUM_BANKS(ADDR_SURF_16_BANK) |
2568 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2569 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2570 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2571 break;
2572 case 14: /* Thin 8 bpp. */
2573 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2574 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2575 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2576 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2577 NUM_BANKS(ADDR_SURF_16_BANK) |
2578 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2579 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2580 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2581 break;
2582 case 15: /* Thin 16 bpp. */
2583 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2584 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2585 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2586 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2587 NUM_BANKS(ADDR_SURF_16_BANK) |
2588 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2589 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2590 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2591 break;
2592 case 16: /* Thin 32 bpp. */
2593 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2594 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2595 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2596 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2597 NUM_BANKS(ADDR_SURF_16_BANK) |
2598 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2599 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2600 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2601 break;
2602 case 17: /* Thin 64 bpp. */
2603 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2604 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2605 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2606 TILE_SPLIT(split_equal_to_row_size) |
2607 NUM_BANKS(ADDR_SURF_16_BANK) |
2608 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2609 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2610 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2611 break;
2612 case 21: /* 8 bpp PRT. */
2613 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2614 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2615 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2616 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2617 NUM_BANKS(ADDR_SURF_16_BANK) |
2618 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
2619 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2620 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2621 break;
2622 case 22: /* 16 bpp PRT */
2623 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2624 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2625 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2626 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2627 NUM_BANKS(ADDR_SURF_16_BANK) |
2628 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2629 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2630 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2631 break;
2632 case 23: /* 32 bpp PRT */
2633 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2634 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2635 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2636 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2637 NUM_BANKS(ADDR_SURF_16_BANK) |
2638 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2639 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2640 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2641 break;
2642 case 24: /* 64 bpp PRT */
2643 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2644 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2645 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2646 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2647 NUM_BANKS(ADDR_SURF_16_BANK) |
2648 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2649 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2650 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2651 break;
2652 case 25: /* 128 bpp PRT */
2653 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2654 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2655 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2656 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
2657 NUM_BANKS(ADDR_SURF_8_BANK) |
2658 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2659 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2660 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2661 break;
2662 default:
2663 gb_tile_moden = 0;
2664 break;
2665 }
64d7b8be 2666 rdev->config.si.tile_mode_array[reg_offset] = gb_tile_moden;
0a96d72b
AD
2667 WREG32(GB_TILE_MODE0 + (reg_offset * 4), gb_tile_moden);
2668 }
d0ae7fcc 2669 } else if ((rdev->family == CHIP_VERDE) ||
8b02859d
AD
2670 (rdev->family == CHIP_OLAND) ||
2671 (rdev->family == CHIP_HAINAN)) {
0a96d72b
AD
2672 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
2673 switch (reg_offset) {
2674 case 0: /* non-AA compressed depth or any compressed stencil */
2675 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2676 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2677 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2678 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2679 NUM_BANKS(ADDR_SURF_16_BANK) |
2680 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2681 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2682 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2683 break;
2684 case 1: /* 2xAA/4xAA compressed depth only */
2685 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2686 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2687 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2688 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2689 NUM_BANKS(ADDR_SURF_16_BANK) |
2690 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2691 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2692 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2693 break;
2694 case 2: /* 8xAA compressed depth only */
2695 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2696 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2697 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2698 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2699 NUM_BANKS(ADDR_SURF_16_BANK) |
2700 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2701 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2702 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2703 break;
2704 case 3: /* 2xAA/4xAA compressed depth with stencil (for depth buffer) */
2705 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2706 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2707 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2708 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2709 NUM_BANKS(ADDR_SURF_16_BANK) |
2710 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2711 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2712 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2713 break;
2714 case 4: /* Maps w/ a dimension less than the 2D macro-tile dimensions (for mipmapped depth textures) */
2715 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2716 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2717 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2718 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2719 NUM_BANKS(ADDR_SURF_16_BANK) |
2720 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2721 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2722 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2723 break;
2724 case 5: /* Uncompressed 16bpp depth - and stencil buffer allocated with it */
2725 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2726 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2727 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2728 TILE_SPLIT(split_equal_to_row_size) |
2729 NUM_BANKS(ADDR_SURF_16_BANK) |
2730 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2731 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2732 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2733 break;
2734 case 6: /* Uncompressed 32bpp depth - and stencil buffer allocated with it */
2735 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2736 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2737 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2738 TILE_SPLIT(split_equal_to_row_size) |
2739 NUM_BANKS(ADDR_SURF_16_BANK) |
2740 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2741 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2742 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2743 break;
2744 case 7: /* Uncompressed 8bpp stencil without depth (drivers typically do not use) */
2745 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2746 MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
2747 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2748 TILE_SPLIT(split_equal_to_row_size) |
2749 NUM_BANKS(ADDR_SURF_16_BANK) |
2750 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2751 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2752 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2753 break;
2754 case 8: /* 1D and 1D Array Surfaces */
2755 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
2756 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2757 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2758 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2759 NUM_BANKS(ADDR_SURF_16_BANK) |
2760 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2761 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2762 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2763 break;
2764 case 9: /* Displayable maps. */
2765 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2766 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2767 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2768 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2769 NUM_BANKS(ADDR_SURF_16_BANK) |
2770 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2771 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2772 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2773 break;
2774 case 10: /* Display 8bpp. */
2775 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2776 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2777 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2778 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2779 NUM_BANKS(ADDR_SURF_16_BANK) |
2780 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2781 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2782 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2783 break;
2784 case 11: /* Display 16bpp. */
2785 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2786 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2787 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2788 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2789 NUM_BANKS(ADDR_SURF_16_BANK) |
2790 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2791 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2792 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2793 break;
2794 case 12: /* Display 32bpp. */
2795 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2796 MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
2797 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2798 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2799 NUM_BANKS(ADDR_SURF_16_BANK) |
2800 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2801 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2802 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2803 break;
2804 case 13: /* Thin. */
2805 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2806 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2807 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2808 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2809 NUM_BANKS(ADDR_SURF_16_BANK) |
2810 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2811 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2812 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2813 break;
2814 case 14: /* Thin 8 bpp. */
2815 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2816 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2817 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2818 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2819 NUM_BANKS(ADDR_SURF_16_BANK) |
2820 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2821 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2822 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2823 break;
2824 case 15: /* Thin 16 bpp. */
2825 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2826 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2827 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2828 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2829 NUM_BANKS(ADDR_SURF_16_BANK) |
2830 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2831 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2832 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2833 break;
2834 case 16: /* Thin 32 bpp. */
2835 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2836 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2837 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2838 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2839 NUM_BANKS(ADDR_SURF_16_BANK) |
2840 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2841 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2842 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2843 break;
2844 case 17: /* Thin 64 bpp. */
2845 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2846 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2847 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2848 TILE_SPLIT(split_equal_to_row_size) |
2849 NUM_BANKS(ADDR_SURF_16_BANK) |
2850 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2851 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2852 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2853 break;
2854 case 21: /* 8 bpp PRT. */
2855 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2856 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2857 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2858 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2859 NUM_BANKS(ADDR_SURF_16_BANK) |
2860 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
2861 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2862 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2863 break;
2864 case 22: /* 16 bpp PRT */
2865 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2866 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2867 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2868 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2869 NUM_BANKS(ADDR_SURF_16_BANK) |
2870 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2871 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2872 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
2873 break;
2874 case 23: /* 32 bpp PRT */
2875 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2876 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2877 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2878 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2879 NUM_BANKS(ADDR_SURF_16_BANK) |
2880 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2881 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2882 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2883 break;
2884 case 24: /* 64 bpp PRT */
2885 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2886 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2887 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2888 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2889 NUM_BANKS(ADDR_SURF_16_BANK) |
2890 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2891 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2892 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
2893 break;
2894 case 25: /* 128 bpp PRT */
2895 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2896 MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
2897 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2898 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
2899 NUM_BANKS(ADDR_SURF_8_BANK) |
2900 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2901 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2902 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
2903 break;
2904 default:
2905 gb_tile_moden = 0;
2906 break;
2907 }
64d7b8be 2908 rdev->config.si.tile_mode_array[reg_offset] = gb_tile_moden;
0a96d72b
AD
2909 WREG32(GB_TILE_MODE0 + (reg_offset * 4), gb_tile_moden);
2910 }
2911 } else
2912 DRM_ERROR("unknown asic: 0x%x\n", rdev->family);
2913}
2914
1a8ca750
AD
2915static void si_select_se_sh(struct radeon_device *rdev,
2916 u32 se_num, u32 sh_num)
2917{
2918 u32 data = INSTANCE_BROADCAST_WRITES;
2919
2920 if ((se_num == 0xffffffff) && (sh_num == 0xffffffff))
79b52d6a 2921 data |= SH_BROADCAST_WRITES | SE_BROADCAST_WRITES;
1a8ca750
AD
2922 else if (se_num == 0xffffffff)
2923 data |= SE_BROADCAST_WRITES | SH_INDEX(sh_num);
2924 else if (sh_num == 0xffffffff)
2925 data |= SH_BROADCAST_WRITES | SE_INDEX(se_num);
2926 else
2927 data |= SH_INDEX(sh_num) | SE_INDEX(se_num);
2928 WREG32(GRBM_GFX_INDEX, data);
2929}
2930
2931static u32 si_create_bitmask(u32 bit_width)
2932{
2933 u32 i, mask = 0;
2934
2935 for (i = 0; i < bit_width; i++) {
2936 mask <<= 1;
2937 mask |= 1;
2938 }
2939 return mask;
2940}
2941
2942static u32 si_get_cu_enabled(struct radeon_device *rdev, u32 cu_per_sh)
2943{
2944 u32 data, mask;
2945
2946 data = RREG32(CC_GC_SHADER_ARRAY_CONFIG);
2947 if (data & 1)
2948 data &= INACTIVE_CUS_MASK;
2949 else
2950 data = 0;
2951 data |= RREG32(GC_USER_SHADER_ARRAY_CONFIG);
2952
2953 data >>= INACTIVE_CUS_SHIFT;
2954
2955 mask = si_create_bitmask(cu_per_sh);
2956
2957 return ~data & mask;
2958}
2959
2960static void si_setup_spi(struct radeon_device *rdev,
2961 u32 se_num, u32 sh_per_se,
2962 u32 cu_per_sh)
2963{
2964 int i, j, k;
2965 u32 data, mask, active_cu;
2966
2967 for (i = 0; i < se_num; i++) {
2968 for (j = 0; j < sh_per_se; j++) {
2969 si_select_se_sh(rdev, i, j);
2970 data = RREG32(SPI_STATIC_THREAD_MGMT_3);
2971 active_cu = si_get_cu_enabled(rdev, cu_per_sh);
2972
2973 mask = 1;
2974 for (k = 0; k < 16; k++) {
2975 mask <<= k;
2976 if (active_cu & mask) {
2977 data &= ~mask;
2978 WREG32(SPI_STATIC_THREAD_MGMT_3, data);
2979 break;
2980 }
2981 }
2982 }
2983 }
2984 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
2985}
2986
2987static u32 si_get_rb_disabled(struct radeon_device *rdev,
9fadb352 2988 u32 max_rb_num_per_se,
1a8ca750
AD
2989 u32 sh_per_se)
2990{
2991 u32 data, mask;
2992
2993 data = RREG32(CC_RB_BACKEND_DISABLE);
2994 if (data & 1)
2995 data &= BACKEND_DISABLE_MASK;
2996 else
2997 data = 0;
2998 data |= RREG32(GC_USER_RB_BACKEND_DISABLE);
2999
3000 data >>= BACKEND_DISABLE_SHIFT;
3001
9fadb352 3002 mask = si_create_bitmask(max_rb_num_per_se / sh_per_se);
1a8ca750
AD
3003
3004 return data & mask;
3005}
3006
3007static void si_setup_rb(struct radeon_device *rdev,
3008 u32 se_num, u32 sh_per_se,
9fadb352 3009 u32 max_rb_num_per_se)
1a8ca750
AD
3010{
3011 int i, j;
3012 u32 data, mask;
3013 u32 disabled_rbs = 0;
3014 u32 enabled_rbs = 0;
3015
3016 for (i = 0; i < se_num; i++) {
3017 for (j = 0; j < sh_per_se; j++) {
3018 si_select_se_sh(rdev, i, j);
9fadb352 3019 data = si_get_rb_disabled(rdev, max_rb_num_per_se, sh_per_se);
1a8ca750
AD
3020 disabled_rbs |= data << ((i * sh_per_se + j) * TAHITI_RB_BITMAP_WIDTH_PER_SH);
3021 }
3022 }
3023 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
3024
3025 mask = 1;
9fadb352 3026 for (i = 0; i < max_rb_num_per_se * se_num; i++) {
1a8ca750
AD
3027 if (!(disabled_rbs & mask))
3028 enabled_rbs |= mask;
3029 mask <<= 1;
3030 }
3031
439a1cff
MO
3032 rdev->config.si.backend_enable_mask = enabled_rbs;
3033
1a8ca750
AD
3034 for (i = 0; i < se_num; i++) {
3035 si_select_se_sh(rdev, i, 0xffffffff);
3036 data = 0;
3037 for (j = 0; j < sh_per_se; j++) {
3038 switch (enabled_rbs & 3) {
3039 case 1:
3040 data |= (RASTER_CONFIG_RB_MAP_0 << (i * sh_per_se + j) * 2);
3041 break;
3042 case 2:
3043 data |= (RASTER_CONFIG_RB_MAP_3 << (i * sh_per_se + j) * 2);
3044 break;
3045 case 3:
3046 default:
3047 data |= (RASTER_CONFIG_RB_MAP_2 << (i * sh_per_se + j) * 2);
3048 break;
3049 }
3050 enabled_rbs >>= 2;
3051 }
3052 WREG32(PA_SC_RASTER_CONFIG, data);
3053 }
3054 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
3055}
3056
0a96d72b
AD
3057static void si_gpu_init(struct radeon_device *rdev)
3058{
0a96d72b
AD
3059 u32 gb_addr_config = 0;
3060 u32 mc_shared_chmap, mc_arb_ramcfg;
0a96d72b 3061 u32 sx_debug_1;
0a96d72b
AD
3062 u32 hdp_host_path_cntl;
3063 u32 tmp;
6101b3ae 3064 int i, j;
0a96d72b
AD
3065
3066 switch (rdev->family) {
3067 case CHIP_TAHITI:
3068 rdev->config.si.max_shader_engines = 2;
0a96d72b 3069 rdev->config.si.max_tile_pipes = 12;
1a8ca750
AD
3070 rdev->config.si.max_cu_per_sh = 8;
3071 rdev->config.si.max_sh_per_se = 2;
0a96d72b
AD
3072 rdev->config.si.max_backends_per_se = 4;
3073 rdev->config.si.max_texture_channel_caches = 12;
3074 rdev->config.si.max_gprs = 256;
3075 rdev->config.si.max_gs_threads = 32;
3076 rdev->config.si.max_hw_contexts = 8;
3077
3078 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
3079 rdev->config.si.sc_prim_fifo_size_backend = 0x100;
3080 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
3081 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
1a8ca750 3082 gb_addr_config = TAHITI_GB_ADDR_CONFIG_GOLDEN;
0a96d72b
AD
3083 break;
3084 case CHIP_PITCAIRN:
3085 rdev->config.si.max_shader_engines = 2;
0a96d72b 3086 rdev->config.si.max_tile_pipes = 8;
1a8ca750
AD
3087 rdev->config.si.max_cu_per_sh = 5;
3088 rdev->config.si.max_sh_per_se = 2;
0a96d72b
AD
3089 rdev->config.si.max_backends_per_se = 4;
3090 rdev->config.si.max_texture_channel_caches = 8;
3091 rdev->config.si.max_gprs = 256;
3092 rdev->config.si.max_gs_threads = 32;
3093 rdev->config.si.max_hw_contexts = 8;
3094
3095 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
3096 rdev->config.si.sc_prim_fifo_size_backend = 0x100;
3097 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
3098 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
1a8ca750 3099 gb_addr_config = TAHITI_GB_ADDR_CONFIG_GOLDEN;
0a96d72b
AD
3100 break;
3101 case CHIP_VERDE:
3102 default:
3103 rdev->config.si.max_shader_engines = 1;
0a96d72b 3104 rdev->config.si.max_tile_pipes = 4;
468ef1a5 3105 rdev->config.si.max_cu_per_sh = 5;
1a8ca750 3106 rdev->config.si.max_sh_per_se = 2;
0a96d72b
AD
3107 rdev->config.si.max_backends_per_se = 4;
3108 rdev->config.si.max_texture_channel_caches = 4;
3109 rdev->config.si.max_gprs = 256;
3110 rdev->config.si.max_gs_threads = 32;
3111 rdev->config.si.max_hw_contexts = 8;
3112
d0ae7fcc
AD
3113 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
3114 rdev->config.si.sc_prim_fifo_size_backend = 0x40;
3115 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
3116 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
3117 gb_addr_config = VERDE_GB_ADDR_CONFIG_GOLDEN;
3118 break;
3119 case CHIP_OLAND:
3120 rdev->config.si.max_shader_engines = 1;
3121 rdev->config.si.max_tile_pipes = 4;
3122 rdev->config.si.max_cu_per_sh = 6;
3123 rdev->config.si.max_sh_per_se = 1;
3124 rdev->config.si.max_backends_per_se = 2;
3125 rdev->config.si.max_texture_channel_caches = 4;
3126 rdev->config.si.max_gprs = 256;
3127 rdev->config.si.max_gs_threads = 16;
3128 rdev->config.si.max_hw_contexts = 8;
3129
0a96d72b
AD
3130 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
3131 rdev->config.si.sc_prim_fifo_size_backend = 0x40;
3132 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
3133 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
1a8ca750 3134 gb_addr_config = VERDE_GB_ADDR_CONFIG_GOLDEN;
0a96d72b 3135 break;
8b02859d
AD
3136 case CHIP_HAINAN:
3137 rdev->config.si.max_shader_engines = 1;
3138 rdev->config.si.max_tile_pipes = 4;
3139 rdev->config.si.max_cu_per_sh = 5;
3140 rdev->config.si.max_sh_per_se = 1;
3141 rdev->config.si.max_backends_per_se = 1;
3142 rdev->config.si.max_texture_channel_caches = 2;
3143 rdev->config.si.max_gprs = 256;
3144 rdev->config.si.max_gs_threads = 16;
3145 rdev->config.si.max_hw_contexts = 8;
3146
3147 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
3148 rdev->config.si.sc_prim_fifo_size_backend = 0x40;
3149 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
3150 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
3151 gb_addr_config = HAINAN_GB_ADDR_CONFIG_GOLDEN;
3152 break;
0a96d72b
AD
3153 }
3154
3155 /* Initialize HDP */
3156 for (i = 0, j = 0; i < 32; i++, j += 0x18) {
3157 WREG32((0x2c14 + j), 0x00000000);
3158 WREG32((0x2c18 + j), 0x00000000);
3159 WREG32((0x2c1c + j), 0x00000000);
3160 WREG32((0x2c20 + j), 0x00000000);
3161 WREG32((0x2c24 + j), 0x00000000);
3162 }
3163
3164 WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff));
3165
3166 evergreen_fix_pci_max_read_req_size(rdev);
3167
3168 WREG32(BIF_FB_EN, FB_READ_EN | FB_WRITE_EN);
3169
3170 mc_shared_chmap = RREG32(MC_SHARED_CHMAP);
3171 mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG);
3172
0a96d72b 3173 rdev->config.si.num_tile_pipes = rdev->config.si.max_tile_pipes;
0a96d72b
AD
3174 rdev->config.si.mem_max_burst_length_bytes = 256;
3175 tmp = (mc_arb_ramcfg & NOOFCOLS_MASK) >> NOOFCOLS_SHIFT;
3176 rdev->config.si.mem_row_size_in_kb = (4 * (1 << (8 + tmp))) / 1024;
3177 if (rdev->config.si.mem_row_size_in_kb > 4)
3178 rdev->config.si.mem_row_size_in_kb = 4;
3179 /* XXX use MC settings? */
3180 rdev->config.si.shader_engine_tile_size = 32;
3181 rdev->config.si.num_gpus = 1;
3182 rdev->config.si.multi_gpu_tile_size = 64;
3183
1a8ca750
AD
3184 /* fix up row size */
3185 gb_addr_config &= ~ROW_SIZE_MASK;
0a96d72b
AD
3186 switch (rdev->config.si.mem_row_size_in_kb) {
3187 case 1:
3188 default:
3189 gb_addr_config |= ROW_SIZE(0);
3190 break;
3191 case 2:
3192 gb_addr_config |= ROW_SIZE(1);
3193 break;
3194 case 4:
3195 gb_addr_config |= ROW_SIZE(2);
3196 break;
3197 }
3198
0a96d72b
AD
3199 /* setup tiling info dword. gb_addr_config is not adequate since it does
3200 * not have bank info, so create a custom tiling dword.
3201 * bits 3:0 num_pipes
3202 * bits 7:4 num_banks
3203 * bits 11:8 group_size
3204 * bits 15:12 row_size
3205 */
3206 rdev->config.si.tile_config = 0;
3207 switch (rdev->config.si.num_tile_pipes) {
3208 case 1:
3209 rdev->config.si.tile_config |= (0 << 0);
3210 break;
3211 case 2:
3212 rdev->config.si.tile_config |= (1 << 0);
3213 break;
3214 case 4:
3215 rdev->config.si.tile_config |= (2 << 0);
3216 break;
3217 case 8:
3218 default:
3219 /* XXX what about 12? */
3220 rdev->config.si.tile_config |= (3 << 0);
3221 break;
dca571a6
CK
3222 }
3223 switch ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT) {
3224 case 0: /* four banks */
1a8ca750 3225 rdev->config.si.tile_config |= 0 << 4;
dca571a6
CK
3226 break;
3227 case 1: /* eight banks */
3228 rdev->config.si.tile_config |= 1 << 4;
3229 break;
3230 case 2: /* sixteen banks */
3231 default:
3232 rdev->config.si.tile_config |= 2 << 4;
3233 break;
3234 }
0a96d72b
AD
3235 rdev->config.si.tile_config |=
3236 ((gb_addr_config & PIPE_INTERLEAVE_SIZE_MASK) >> PIPE_INTERLEAVE_SIZE_SHIFT) << 8;
3237 rdev->config.si.tile_config |=
3238 ((gb_addr_config & ROW_SIZE_MASK) >> ROW_SIZE_SHIFT) << 12;
3239
0a96d72b
AD
3240 WREG32(GB_ADDR_CONFIG, gb_addr_config);
3241 WREG32(DMIF_ADDR_CONFIG, gb_addr_config);
7c1c7c18 3242 WREG32(DMIF_ADDR_CALC, gb_addr_config);
0a96d72b 3243 WREG32(HDP_ADDR_CONFIG, gb_addr_config);
8c5fd7ef
AD
3244 WREG32(DMA_TILING_CONFIG + DMA0_REGISTER_OFFSET, gb_addr_config);
3245 WREG32(DMA_TILING_CONFIG + DMA1_REGISTER_OFFSET, gb_addr_config);
1df0d523
AD
3246 if (rdev->has_uvd) {
3247 WREG32(UVD_UDEC_ADDR_CONFIG, gb_addr_config);
3248 WREG32(UVD_UDEC_DB_ADDR_CONFIG, gb_addr_config);
3249 WREG32(UVD_UDEC_DBW_ADDR_CONFIG, gb_addr_config);
3250 }
0a96d72b 3251
1a8ca750 3252 si_tiling_mode_table_init(rdev);
0a96d72b 3253
1a8ca750
AD
3254 si_setup_rb(rdev, rdev->config.si.max_shader_engines,
3255 rdev->config.si.max_sh_per_se,
3256 rdev->config.si.max_backends_per_se);
0a96d72b 3257
1a8ca750
AD
3258 si_setup_spi(rdev, rdev->config.si.max_shader_engines,
3259 rdev->config.si.max_sh_per_se,
3260 rdev->config.si.max_cu_per_sh);
0a96d72b 3261
52da51f0 3262 rdev->config.si.active_cus = 0;
65fcf668
AD
3263 for (i = 0; i < rdev->config.si.max_shader_engines; i++) {
3264 for (j = 0; j < rdev->config.si.max_sh_per_se; j++) {
6101b3ae
AD
3265 rdev->config.si.active_cus +=
3266 hweight32(si_get_cu_active_bitmap(rdev, i, j));
65fcf668
AD
3267 }
3268 }
0a96d72b
AD
3269
3270 /* set HW defaults for 3D engine */
3271 WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) |
3272 ROQ_IB2_START(0x2b)));
3273 WREG32(CP_MEQ_THRESHOLDS, MEQ1_START(0x30) | MEQ2_START(0x60));
3274
3275 sx_debug_1 = RREG32(SX_DEBUG_1);
3276 WREG32(SX_DEBUG_1, sx_debug_1);
3277
3278 WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4));
3279
3280 WREG32(PA_SC_FIFO_SIZE, (SC_FRONTEND_PRIM_FIFO_SIZE(rdev->config.si.sc_prim_fifo_size_frontend) |
3281 SC_BACKEND_PRIM_FIFO_SIZE(rdev->config.si.sc_prim_fifo_size_backend) |
3282 SC_HIZ_TILE_FIFO_SIZE(rdev->config.si.sc_hiz_tile_fifo_size) |
3283 SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.si.sc_earlyz_tile_fifo_size)));
3284
3285 WREG32(VGT_NUM_INSTANCES, 1);
3286
3287 WREG32(CP_PERFMON_CNTL, 0);
3288
3289 WREG32(SQ_CONFIG, 0);
3290
3291 WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) |
3292 FORCE_EOV_MAX_REZ_CNT(255)));
3293
3294 WREG32(VGT_CACHE_INVALIDATION, CACHE_INVALIDATION(VC_AND_TC) |
3295 AUTO_INVLD_EN(ES_AND_GS_AUTO));
3296
3297 WREG32(VGT_GS_VERTEX_REUSE, 16);
3298 WREG32(PA_SC_LINE_STIPPLE_STATE, 0);
3299
3300 WREG32(CB_PERFCOUNTER0_SELECT0, 0);
3301 WREG32(CB_PERFCOUNTER0_SELECT1, 0);
3302 WREG32(CB_PERFCOUNTER1_SELECT0, 0);
3303 WREG32(CB_PERFCOUNTER1_SELECT1, 0);
3304 WREG32(CB_PERFCOUNTER2_SELECT0, 0);
3305 WREG32(CB_PERFCOUNTER2_SELECT1, 0);
3306 WREG32(CB_PERFCOUNTER3_SELECT0, 0);
3307 WREG32(CB_PERFCOUNTER3_SELECT1, 0);
3308
3309 tmp = RREG32(HDP_MISC_CNTL);
3310 tmp |= HDP_FLUSH_INVALIDATE_CACHE;
3311 WREG32(HDP_MISC_CNTL, tmp);
3312
3313 hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL);
3314 WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
3315
3316 WREG32(PA_CL_ENHANCE, CLIP_VTX_REORDER_ENA | NUM_CLIP_SEQ(3));
3317
3318 udelay(50);
3319}
c476dde2 3320
2ece2e8b
AD
3321/*
3322 * GPU scratch registers helpers function.
3323 */
3324static void si_scratch_init(struct radeon_device *rdev)
3325{
3326 int i;
3327
3328 rdev->scratch.num_reg = 7;
3329 rdev->scratch.reg_base = SCRATCH_REG0;
3330 for (i = 0; i < rdev->scratch.num_reg; i++) {
3331 rdev->scratch.free[i] = true;
3332 rdev->scratch.reg[i] = rdev->scratch.reg_base + (i * 4);
3333 }
3334}
3335
3336void si_fence_ring_emit(struct radeon_device *rdev,
3337 struct radeon_fence *fence)
3338{
3339 struct radeon_ring *ring = &rdev->ring[fence->ring];
3340 u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
3341
3342 /* flush read cache over gart */
3343 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
3344 radeon_ring_write(ring, (CP_COHER_CNTL2 - PACKET3_SET_CONFIG_REG_START) >> 2);
3345 radeon_ring_write(ring, 0);
3346 radeon_ring_write(ring, PACKET3(PACKET3_SURFACE_SYNC, 3));
3347 radeon_ring_write(ring, PACKET3_TCL1_ACTION_ENA |
3348 PACKET3_TC_ACTION_ENA |
3349 PACKET3_SH_KCACHE_ACTION_ENA |
3350 PACKET3_SH_ICACHE_ACTION_ENA);
3351 radeon_ring_write(ring, 0xFFFFFFFF);
3352 radeon_ring_write(ring, 0);
3353 radeon_ring_write(ring, 10); /* poll interval */
3354 /* EVENT_WRITE_EOP - flush caches, send int */
3355 radeon_ring_write(ring, PACKET3(PACKET3_EVENT_WRITE_EOP, 4));
3356 radeon_ring_write(ring, EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) | EVENT_INDEX(5));
5e167cdb 3357 radeon_ring_write(ring, lower_32_bits(addr));
2ece2e8b
AD
3358 radeon_ring_write(ring, (upper_32_bits(addr) & 0xff) | DATA_SEL(1) | INT_SEL(2));
3359 radeon_ring_write(ring, fence->seq);
3360 radeon_ring_write(ring, 0);
3361}
3362
3363/*
3364 * IB stuff
3365 */
3366void si_ring_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib)
3367{
876dc9f3 3368 struct radeon_ring *ring = &rdev->ring[ib->ring];
7c42bc1a 3369 unsigned vm_id = ib->vm ? ib->vm->ids[ib->ring].id : 0;
2ece2e8b
AD
3370 u32 header;
3371
a85a7da4
AD
3372 if (ib->is_const_ib) {
3373 /* set switch buffer packet before const IB */
3374 radeon_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
3375 radeon_ring_write(ring, 0);
45df6803 3376
2ece2e8b 3377 header = PACKET3(PACKET3_INDIRECT_BUFFER_CONST, 2);
a85a7da4 3378 } else {
89d35807 3379 u32 next_rptr;
a85a7da4 3380 if (ring->rptr_save_reg) {
89d35807 3381 next_rptr = ring->wptr + 3 + 4 + 8;
a85a7da4
AD
3382 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
3383 radeon_ring_write(ring, ((ring->rptr_save_reg -
3384 PACKET3_SET_CONFIG_REG_START) >> 2));
3385 radeon_ring_write(ring, next_rptr);
89d35807
AD
3386 } else if (rdev->wb.enabled) {
3387 next_rptr = ring->wptr + 5 + 4 + 8;
3388 radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
3389 radeon_ring_write(ring, (1 << 8));
3390 radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
5e167cdb 3391 radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr));
89d35807 3392 radeon_ring_write(ring, next_rptr);
a85a7da4
AD
3393 }
3394
2ece2e8b 3395 header = PACKET3(PACKET3_INDIRECT_BUFFER, 2);
a85a7da4 3396 }
2ece2e8b
AD
3397
3398 radeon_ring_write(ring, header);
3399 radeon_ring_write(ring,
3400#ifdef __BIG_ENDIAN
3401 (2 << 0) |
3402#endif
3403 (ib->gpu_addr & 0xFFFFFFFC));
3404 radeon_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xFFFF);
7c42bc1a 3405 radeon_ring_write(ring, ib->length_dw | (vm_id << 24));
2ece2e8b 3406
a85a7da4
AD
3407 if (!ib->is_const_ib) {
3408 /* flush read cache over gart for this vmid */
3409 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
3410 radeon_ring_write(ring, (CP_COHER_CNTL2 - PACKET3_SET_CONFIG_REG_START) >> 2);
7c42bc1a 3411 radeon_ring_write(ring, vm_id);
a85a7da4
AD
3412 radeon_ring_write(ring, PACKET3(PACKET3_SURFACE_SYNC, 3));
3413 radeon_ring_write(ring, PACKET3_TCL1_ACTION_ENA |
3414 PACKET3_TC_ACTION_ENA |
3415 PACKET3_SH_KCACHE_ACTION_ENA |
3416 PACKET3_SH_ICACHE_ACTION_ENA);
3417 radeon_ring_write(ring, 0xFFFFFFFF);
3418 radeon_ring_write(ring, 0);
3419 radeon_ring_write(ring, 10); /* poll interval */
3420 }
2ece2e8b
AD
3421}
3422
48c0c902
AD
3423/*
3424 * CP.
3425 */
3426static void si_cp_enable(struct radeon_device *rdev, bool enable)
3427{
3428 if (enable)
3429 WREG32(CP_ME_CNTL, 0);
3430 else {
50efa51a
AD
3431 if (rdev->asic->copy.copy_ring_index == RADEON_RING_TYPE_GFX_INDEX)
3432 radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
48c0c902
AD
3433 WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT | CP_CE_HALT));
3434 WREG32(SCRATCH_UMSK, 0);
8c5fd7ef
AD
3435 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
3436 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
3437 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
48c0c902
AD
3438 }
3439 udelay(50);
3440}
3441
3442static int si_cp_load_microcode(struct radeon_device *rdev)
3443{
48c0c902
AD
3444 int i;
3445
629bd33c 3446 if (!rdev->me_fw || !rdev->pfp_fw || !rdev->ce_fw)
48c0c902
AD
3447 return -EINVAL;
3448
3449 si_cp_enable(rdev, false);
3450
629bd33c
AD
3451 if (rdev->new_fw) {
3452 const struct gfx_firmware_header_v1_0 *pfp_hdr =
3453 (const struct gfx_firmware_header_v1_0 *)rdev->pfp_fw->data;
3454 const struct gfx_firmware_header_v1_0 *ce_hdr =
3455 (const struct gfx_firmware_header_v1_0 *)rdev->ce_fw->data;
3456 const struct gfx_firmware_header_v1_0 *me_hdr =
3457 (const struct gfx_firmware_header_v1_0 *)rdev->me_fw->data;
3458 const __le32 *fw_data;
3459 u32 fw_size;
3460
3461 radeon_ucode_print_gfx_hdr(&pfp_hdr->header);
3462 radeon_ucode_print_gfx_hdr(&ce_hdr->header);
3463 radeon_ucode_print_gfx_hdr(&me_hdr->header);
3464
3465 /* PFP */
3466 fw_data = (const __le32 *)
3467 (rdev->pfp_fw->data + le32_to_cpu(pfp_hdr->header.ucode_array_offset_bytes));
3468 fw_size = le32_to_cpu(pfp_hdr->header.ucode_size_bytes) / 4;
3469 WREG32(CP_PFP_UCODE_ADDR, 0);
3470 for (i = 0; i < fw_size; i++)
3471 WREG32(CP_PFP_UCODE_DATA, le32_to_cpup(fw_data++));
3472 WREG32(CP_PFP_UCODE_ADDR, 0);
3473
3474 /* CE */
3475 fw_data = (const __le32 *)
3476 (rdev->ce_fw->data + le32_to_cpu(ce_hdr->header.ucode_array_offset_bytes));
3477 fw_size = le32_to_cpu(ce_hdr->header.ucode_size_bytes) / 4;
3478 WREG32(CP_CE_UCODE_ADDR, 0);
3479 for (i = 0; i < fw_size; i++)
3480 WREG32(CP_CE_UCODE_DATA, le32_to_cpup(fw_data++));
3481 WREG32(CP_CE_UCODE_ADDR, 0);
3482
3483 /* ME */
3484 fw_data = (const __be32 *)
3485 (rdev->me_fw->data + le32_to_cpu(me_hdr->header.ucode_array_offset_bytes));
3486 fw_size = le32_to_cpu(me_hdr->header.ucode_size_bytes) / 4;
3487 WREG32(CP_ME_RAM_WADDR, 0);
3488 for (i = 0; i < fw_size; i++)
3489 WREG32(CP_ME_RAM_DATA, le32_to_cpup(fw_data++));
3490 WREG32(CP_ME_RAM_WADDR, 0);
3491 } else {
3492 const __be32 *fw_data;
3493
3494 /* PFP */
3495 fw_data = (const __be32 *)rdev->pfp_fw->data;
3496 WREG32(CP_PFP_UCODE_ADDR, 0);
3497 for (i = 0; i < SI_PFP_UCODE_SIZE; i++)
3498 WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++));
3499 WREG32(CP_PFP_UCODE_ADDR, 0);
3500
3501 /* CE */
3502 fw_data = (const __be32 *)rdev->ce_fw->data;
3503 WREG32(CP_CE_UCODE_ADDR, 0);
3504 for (i = 0; i < SI_CE_UCODE_SIZE; i++)
3505 WREG32(CP_CE_UCODE_DATA, be32_to_cpup(fw_data++));
3506 WREG32(CP_CE_UCODE_ADDR, 0);
3507
3508 /* ME */
3509 fw_data = (const __be32 *)rdev->me_fw->data;
3510 WREG32(CP_ME_RAM_WADDR, 0);
3511 for (i = 0; i < SI_PM4_UCODE_SIZE; i++)
3512 WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++));
3513 WREG32(CP_ME_RAM_WADDR, 0);
3514 }
48c0c902
AD
3515
3516 WREG32(CP_PFP_UCODE_ADDR, 0);
3517 WREG32(CP_CE_UCODE_ADDR, 0);
3518 WREG32(CP_ME_RAM_WADDR, 0);
3519 WREG32(CP_ME_RAM_RADDR, 0);
3520 return 0;
3521}
3522
3523static int si_cp_start(struct radeon_device *rdev)
3524{
3525 struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
3526 int r, i;
3527
3528 r = radeon_ring_lock(rdev, ring, 7 + 4);
3529 if (r) {
3530 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
3531 return r;
3532 }
3533 /* init the CP */
3534 radeon_ring_write(ring, PACKET3(PACKET3_ME_INITIALIZE, 5));
3535 radeon_ring_write(ring, 0x1);
3536 radeon_ring_write(ring, 0x0);
3537 radeon_ring_write(ring, rdev->config.si.max_hw_contexts - 1);
3538 radeon_ring_write(ring, PACKET3_ME_INITIALIZE_DEVICE_ID(1));
3539 radeon_ring_write(ring, 0);
3540 radeon_ring_write(ring, 0);
3541
3542 /* init the CE partitions */
3543 radeon_ring_write(ring, PACKET3(PACKET3_SET_BASE, 2));
3544 radeon_ring_write(ring, PACKET3_BASE_INDEX(CE_PARTITION_BASE));
3545 radeon_ring_write(ring, 0xc000);
3546 radeon_ring_write(ring, 0xe000);
1538a9e0 3547 radeon_ring_unlock_commit(rdev, ring, false);
48c0c902
AD
3548
3549 si_cp_enable(rdev, true);
3550
3551 r = radeon_ring_lock(rdev, ring, si_default_size + 10);
3552 if (r) {
3553 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
3554 return r;
3555 }
3556
3557 /* setup clear context state */
3558 radeon_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
3559 radeon_ring_write(ring, PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
3560
3561 for (i = 0; i < si_default_size; i++)
3562 radeon_ring_write(ring, si_default_state[i]);
3563
3564 radeon_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
3565 radeon_ring_write(ring, PACKET3_PREAMBLE_END_CLEAR_STATE);
3566
3567 /* set clear context state */
3568 radeon_ring_write(ring, PACKET3(PACKET3_CLEAR_STATE, 0));
3569 radeon_ring_write(ring, 0);
3570
3571 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONTEXT_REG, 2));
3572 radeon_ring_write(ring, 0x00000316);
3573 radeon_ring_write(ring, 0x0000000e); /* VGT_VERTEX_REUSE_BLOCK_CNTL */
3574 radeon_ring_write(ring, 0x00000010); /* VGT_OUT_DEALLOC_CNTL */
3575
1538a9e0 3576 radeon_ring_unlock_commit(rdev, ring, false);
48c0c902
AD
3577
3578 for (i = RADEON_RING_TYPE_GFX_INDEX; i <= CAYMAN_RING_TYPE_CP2_INDEX; ++i) {
3579 ring = &rdev->ring[i];
3580 r = radeon_ring_lock(rdev, ring, 2);
3581
3582 /* clear the compute context state */
3583 radeon_ring_write(ring, PACKET3_COMPUTE(PACKET3_CLEAR_STATE, 0));
3584 radeon_ring_write(ring, 0);
3585
1538a9e0 3586 radeon_ring_unlock_commit(rdev, ring, false);
48c0c902
AD
3587 }
3588
3589 return 0;
3590}
3591
3592static void si_cp_fini(struct radeon_device *rdev)
3593{
45df6803 3594 struct radeon_ring *ring;
48c0c902 3595 si_cp_enable(rdev, false);
45df6803
CK
3596
3597 ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
3598 radeon_ring_fini(rdev, ring);
3599 radeon_scratch_free(rdev, ring->rptr_save_reg);
3600
3601 ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
3602 radeon_ring_fini(rdev, ring);
3603 radeon_scratch_free(rdev, ring->rptr_save_reg);
3604
3605 ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
3606 radeon_ring_fini(rdev, ring);
3607 radeon_scratch_free(rdev, ring->rptr_save_reg);
48c0c902
AD
3608}
3609
3610static int si_cp_resume(struct radeon_device *rdev)
3611{
3612 struct radeon_ring *ring;
3613 u32 tmp;
3614 u32 rb_bufsz;
3615 int r;
3616
811e4d58
AD
3617 si_enable_gui_idle_interrupt(rdev, false);
3618
48c0c902
AD
3619 WREG32(CP_SEM_WAIT_TIMER, 0x0);
3620 WREG32(CP_SEM_INCOMPLETE_TIMER_CNTL, 0x0);
3621
3622 /* Set the write pointer delay */
3623 WREG32(CP_RB_WPTR_DELAY, 0);
3624
3625 WREG32(CP_DEBUG, 0);
3626 WREG32(SCRATCH_ADDR, ((rdev->wb.gpu_addr + RADEON_WB_SCRATCH_OFFSET) >> 8) & 0xFFFFFFFF);
3627
3628 /* ring 0 - compute and gfx */
3629 /* Set ring buffer size */
3630 ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
b72a8925
DV
3631 rb_bufsz = order_base_2(ring->ring_size / 8);
3632 tmp = (order_base_2(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
48c0c902
AD
3633#ifdef __BIG_ENDIAN
3634 tmp |= BUF_SWAP_32BIT;
3635#endif
3636 WREG32(CP_RB0_CNTL, tmp);
3637
3638 /* Initialize the ring buffer's read and write pointers */
3639 WREG32(CP_RB0_CNTL, tmp | RB_RPTR_WR_ENA);
3640 ring->wptr = 0;
3641 WREG32(CP_RB0_WPTR, ring->wptr);
3642
48fc7f7e 3643 /* set the wb address whether it's enabled or not */
48c0c902
AD
3644 WREG32(CP_RB0_RPTR_ADDR, (rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET) & 0xFFFFFFFC);
3645 WREG32(CP_RB0_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET) & 0xFF);
3646
3647 if (rdev->wb.enabled)
3648 WREG32(SCRATCH_UMSK, 0xff);
3649 else {
3650 tmp |= RB_NO_UPDATE;
3651 WREG32(SCRATCH_UMSK, 0);
3652 }
3653
3654 mdelay(1);
3655 WREG32(CP_RB0_CNTL, tmp);
3656
3657 WREG32(CP_RB0_BASE, ring->gpu_addr >> 8);
3658
48c0c902
AD
3659 /* ring1 - compute only */
3660 /* Set ring buffer size */
3661 ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
b72a8925
DV
3662 rb_bufsz = order_base_2(ring->ring_size / 8);
3663 tmp = (order_base_2(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
48c0c902
AD
3664#ifdef __BIG_ENDIAN
3665 tmp |= BUF_SWAP_32BIT;
3666#endif
3667 WREG32(CP_RB1_CNTL, tmp);
3668
3669 /* Initialize the ring buffer's read and write pointers */
3670 WREG32(CP_RB1_CNTL, tmp | RB_RPTR_WR_ENA);
3671 ring->wptr = 0;
3672 WREG32(CP_RB1_WPTR, ring->wptr);
3673
48fc7f7e 3674 /* set the wb address whether it's enabled or not */
48c0c902
AD
3675 WREG32(CP_RB1_RPTR_ADDR, (rdev->wb.gpu_addr + RADEON_WB_CP1_RPTR_OFFSET) & 0xFFFFFFFC);
3676 WREG32(CP_RB1_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP1_RPTR_OFFSET) & 0xFF);
3677
3678 mdelay(1);
3679 WREG32(CP_RB1_CNTL, tmp);
3680
3681 WREG32(CP_RB1_BASE, ring->gpu_addr >> 8);
3682
48c0c902
AD
3683 /* ring2 - compute only */
3684 /* Set ring buffer size */
3685 ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
b72a8925
DV
3686 rb_bufsz = order_base_2(ring->ring_size / 8);
3687 tmp = (order_base_2(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
48c0c902
AD
3688#ifdef __BIG_ENDIAN
3689 tmp |= BUF_SWAP_32BIT;
3690#endif
3691 WREG32(CP_RB2_CNTL, tmp);
3692
3693 /* Initialize the ring buffer's read and write pointers */
3694 WREG32(CP_RB2_CNTL, tmp | RB_RPTR_WR_ENA);
3695 ring->wptr = 0;
3696 WREG32(CP_RB2_WPTR, ring->wptr);
3697
48fc7f7e 3698 /* set the wb address whether it's enabled or not */
48c0c902
AD
3699 WREG32(CP_RB2_RPTR_ADDR, (rdev->wb.gpu_addr + RADEON_WB_CP2_RPTR_OFFSET) & 0xFFFFFFFC);
3700 WREG32(CP_RB2_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP2_RPTR_OFFSET) & 0xFF);
3701
3702 mdelay(1);
3703 WREG32(CP_RB2_CNTL, tmp);
3704
3705 WREG32(CP_RB2_BASE, ring->gpu_addr >> 8);
3706
48c0c902
AD
3707 /* start the rings */
3708 si_cp_start(rdev);
3709 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = true;
3710 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = true;
3711 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = true;
3712 r = radeon_ring_test(rdev, RADEON_RING_TYPE_GFX_INDEX, &rdev->ring[RADEON_RING_TYPE_GFX_INDEX]);
3713 if (r) {
3714 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
3715 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
3716 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
3717 return r;
3718 }
3719 r = radeon_ring_test(rdev, CAYMAN_RING_TYPE_CP1_INDEX, &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX]);
3720 if (r) {
3721 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
3722 }
3723 r = radeon_ring_test(rdev, CAYMAN_RING_TYPE_CP2_INDEX, &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX]);
3724 if (r) {
3725 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
3726 }
3727
811e4d58
AD
3728 si_enable_gui_idle_interrupt(rdev, true);
3729
50efa51a
AD
3730 if (rdev->asic->copy.copy_ring_index == RADEON_RING_TYPE_GFX_INDEX)
3731 radeon_ttm_set_active_vram_size(rdev, rdev->mc.real_vram_size);
3732
48c0c902
AD
3733 return 0;
3734}
3735
2483b4ea 3736u32 si_gpu_check_soft_reset(struct radeon_device *rdev)
06bc6df0 3737{
014bb209 3738 u32 reset_mask = 0;
1c534671 3739 u32 tmp;
06bc6df0 3740
014bb209
AD
3741 /* GRBM_STATUS */
3742 tmp = RREG32(GRBM_STATUS);
3743 if (tmp & (PA_BUSY | SC_BUSY |
3744 BCI_BUSY | SX_BUSY |
3745 TA_BUSY | VGT_BUSY |
3746 DB_BUSY | CB_BUSY |
3747 GDS_BUSY | SPI_BUSY |
3748 IA_BUSY | IA_BUSY_NO_DMA))
3749 reset_mask |= RADEON_RESET_GFX;
3750
3751 if (tmp & (CF_RQ_PENDING | PF_RQ_PENDING |
3752 CP_BUSY | CP_COHERENCY_BUSY))
3753 reset_mask |= RADEON_RESET_CP;
3754
3755 if (tmp & GRBM_EE_BUSY)
3756 reset_mask |= RADEON_RESET_GRBM | RADEON_RESET_GFX | RADEON_RESET_CP;
3757
3758 /* GRBM_STATUS2 */
3759 tmp = RREG32(GRBM_STATUS2);
3760 if (tmp & (RLC_RQ_PENDING | RLC_BUSY))
3761 reset_mask |= RADEON_RESET_RLC;
3762
3763 /* DMA_STATUS_REG 0 */
3764 tmp = RREG32(DMA_STATUS_REG + DMA0_REGISTER_OFFSET);
3765 if (!(tmp & DMA_IDLE))
3766 reset_mask |= RADEON_RESET_DMA;
3767
3768 /* DMA_STATUS_REG 1 */
3769 tmp = RREG32(DMA_STATUS_REG + DMA1_REGISTER_OFFSET);
3770 if (!(tmp & DMA_IDLE))
3771 reset_mask |= RADEON_RESET_DMA1;
3772
3773 /* SRBM_STATUS2 */
3774 tmp = RREG32(SRBM_STATUS2);
3775 if (tmp & DMA_BUSY)
3776 reset_mask |= RADEON_RESET_DMA;
3777
3778 if (tmp & DMA1_BUSY)
3779 reset_mask |= RADEON_RESET_DMA1;
3780
3781 /* SRBM_STATUS */
3782 tmp = RREG32(SRBM_STATUS);
3783
3784 if (tmp & IH_BUSY)
3785 reset_mask |= RADEON_RESET_IH;
3786
3787 if (tmp & SEM_BUSY)
3788 reset_mask |= RADEON_RESET_SEM;
3789
3790 if (tmp & GRBM_RQ_PENDING)
3791 reset_mask |= RADEON_RESET_GRBM;
3792
3793 if (tmp & VMC_BUSY)
3794 reset_mask |= RADEON_RESET_VMC;
19fc42ed 3795
014bb209
AD
3796 if (tmp & (MCB_BUSY | MCB_NON_DISPLAY_BUSY |
3797 MCC_BUSY | MCD_BUSY))
3798 reset_mask |= RADEON_RESET_MC;
3799
3800 if (evergreen_is_display_hung(rdev))
3801 reset_mask |= RADEON_RESET_DISPLAY;
3802
3803 /* VM_L2_STATUS */
3804 tmp = RREG32(VM_L2_STATUS);
3805 if (tmp & L2_BUSY)
3806 reset_mask |= RADEON_RESET_VMC;
3807
d808fc88
AD
3808 /* Skip MC reset as it's mostly likely not hung, just busy */
3809 if (reset_mask & RADEON_RESET_MC) {
3810 DRM_DEBUG("MC busy: 0x%08X, clearing.\n", reset_mask);
3811 reset_mask &= ~RADEON_RESET_MC;
3812 }
3813
014bb209
AD
3814 return reset_mask;
3815}
3816
3817static void si_gpu_soft_reset(struct radeon_device *rdev, u32 reset_mask)
3818{
3819 struct evergreen_mc_save save;
3820 u32 grbm_soft_reset = 0, srbm_soft_reset = 0;
3821 u32 tmp;
19fc42ed 3822
06bc6df0 3823 if (reset_mask == 0)
014bb209 3824 return;
06bc6df0
AD
3825
3826 dev_info(rdev->dev, "GPU softreset: 0x%08X\n", reset_mask);
3827
1c534671 3828 evergreen_print_gpu_status_regs(rdev);
06bc6df0
AD
3829 dev_info(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n",
3830 RREG32(VM_CONTEXT1_PROTECTION_FAULT_ADDR));
3831 dev_info(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
3832 RREG32(VM_CONTEXT1_PROTECTION_FAULT_STATUS));
3833
a6f4ae8d
AD
3834 /* disable PG/CG */
3835 si_fini_pg(rdev);
3836 si_fini_cg(rdev);
3837
3838 /* stop the rlc */
3839 si_rlc_stop(rdev);
3840
1c534671
AD
3841 /* Disable CP parsing/prefetching */
3842 WREG32(CP_ME_CNTL, CP_ME_HALT | CP_PFP_HALT | CP_CE_HALT);
3843
3844 if (reset_mask & RADEON_RESET_DMA) {
3845 /* dma0 */
3846 tmp = RREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET);
3847 tmp &= ~DMA_RB_ENABLE;
3848 WREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET, tmp);
014bb209
AD
3849 }
3850 if (reset_mask & RADEON_RESET_DMA1) {
1c534671
AD
3851 /* dma1 */
3852 tmp = RREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET);
3853 tmp &= ~DMA_RB_ENABLE;
3854 WREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET, tmp);
3855 }
3856
f770d78a
AD
3857 udelay(50);
3858
3859 evergreen_mc_stop(rdev, &save);
3860 if (evergreen_mc_wait_for_idle(rdev)) {
3861 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
3862 }
3863
1c534671
AD
3864 if (reset_mask & (RADEON_RESET_GFX | RADEON_RESET_COMPUTE | RADEON_RESET_CP)) {
3865 grbm_soft_reset = SOFT_RESET_CB |
3866 SOFT_RESET_DB |
3867 SOFT_RESET_GDS |
3868 SOFT_RESET_PA |
3869 SOFT_RESET_SC |
3870 SOFT_RESET_BCI |
3871 SOFT_RESET_SPI |
3872 SOFT_RESET_SX |
3873 SOFT_RESET_TC |
3874 SOFT_RESET_TA |
3875 SOFT_RESET_VGT |
3876 SOFT_RESET_IA;
3877 }
3878
3879 if (reset_mask & RADEON_RESET_CP) {
3880 grbm_soft_reset |= SOFT_RESET_CP | SOFT_RESET_VGT;
3881
3882 srbm_soft_reset |= SOFT_RESET_GRBM;
3883 }
06bc6df0
AD
3884
3885 if (reset_mask & RADEON_RESET_DMA)
014bb209
AD
3886 srbm_soft_reset |= SOFT_RESET_DMA;
3887
3888 if (reset_mask & RADEON_RESET_DMA1)
3889 srbm_soft_reset |= SOFT_RESET_DMA1;
3890
3891 if (reset_mask & RADEON_RESET_DISPLAY)
3892 srbm_soft_reset |= SOFT_RESET_DC;
3893
3894 if (reset_mask & RADEON_RESET_RLC)
3895 grbm_soft_reset |= SOFT_RESET_RLC;
3896
3897 if (reset_mask & RADEON_RESET_SEM)
3898 srbm_soft_reset |= SOFT_RESET_SEM;
3899
3900 if (reset_mask & RADEON_RESET_IH)
3901 srbm_soft_reset |= SOFT_RESET_IH;
3902
3903 if (reset_mask & RADEON_RESET_GRBM)
3904 srbm_soft_reset |= SOFT_RESET_GRBM;
3905
3906 if (reset_mask & RADEON_RESET_VMC)
3907 srbm_soft_reset |= SOFT_RESET_VMC;
3908
3909 if (reset_mask & RADEON_RESET_MC)
3910 srbm_soft_reset |= SOFT_RESET_MC;
1c534671
AD
3911
3912 if (grbm_soft_reset) {
3913 tmp = RREG32(GRBM_SOFT_RESET);
3914 tmp |= grbm_soft_reset;
3915 dev_info(rdev->dev, "GRBM_SOFT_RESET=0x%08X\n", tmp);
3916 WREG32(GRBM_SOFT_RESET, tmp);
3917 tmp = RREG32(GRBM_SOFT_RESET);
3918
3919 udelay(50);
3920
3921 tmp &= ~grbm_soft_reset;
3922 WREG32(GRBM_SOFT_RESET, tmp);
3923 tmp = RREG32(GRBM_SOFT_RESET);
3924 }
3925
3926 if (srbm_soft_reset) {
3927 tmp = RREG32(SRBM_SOFT_RESET);
3928 tmp |= srbm_soft_reset;
3929 dev_info(rdev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
3930 WREG32(SRBM_SOFT_RESET, tmp);
3931 tmp = RREG32(SRBM_SOFT_RESET);
3932
3933 udelay(50);
3934
3935 tmp &= ~srbm_soft_reset;
3936 WREG32(SRBM_SOFT_RESET, tmp);
3937 tmp = RREG32(SRBM_SOFT_RESET);
3938 }
06bc6df0
AD
3939
3940 /* Wait a little for things to settle down */
3941 udelay(50);
3942
c476dde2 3943 evergreen_mc_resume(rdev, &save);
1c534671
AD
3944 udelay(50);
3945
1c534671 3946 evergreen_print_gpu_status_regs(rdev);
c476dde2
AD
3947}
3948
4a5c8ea5
AD
3949static void si_set_clk_bypass_mode(struct radeon_device *rdev)
3950{
3951 u32 tmp, i;
3952
3953 tmp = RREG32(CG_SPLL_FUNC_CNTL);
3954 tmp |= SPLL_BYPASS_EN;
3955 WREG32(CG_SPLL_FUNC_CNTL, tmp);
3956
3957 tmp = RREG32(CG_SPLL_FUNC_CNTL_2);
3958 tmp |= SPLL_CTLREQ_CHG;
3959 WREG32(CG_SPLL_FUNC_CNTL_2, tmp);
3960
3961 for (i = 0; i < rdev->usec_timeout; i++) {
3962 if (RREG32(SPLL_STATUS) & SPLL_CHG_STATUS)
3963 break;
3964 udelay(1);
3965 }
3966
3967 tmp = RREG32(CG_SPLL_FUNC_CNTL_2);
3968 tmp &= ~(SPLL_CTLREQ_CHG | SCLK_MUX_UPDATE);
3969 WREG32(CG_SPLL_FUNC_CNTL_2, tmp);
3970
3971 tmp = RREG32(MPLL_CNTL_MODE);
3972 tmp &= ~MPLL_MCLK_SEL;
3973 WREG32(MPLL_CNTL_MODE, tmp);
3974}
3975
3976static void si_spll_powerdown(struct radeon_device *rdev)
3977{
3978 u32 tmp;
3979
3980 tmp = RREG32(SPLL_CNTL_MODE);
3981 tmp |= SPLL_SW_DIR_CONTROL;
3982 WREG32(SPLL_CNTL_MODE, tmp);
3983
3984 tmp = RREG32(CG_SPLL_FUNC_CNTL);
3985 tmp |= SPLL_RESET;
3986 WREG32(CG_SPLL_FUNC_CNTL, tmp);
3987
3988 tmp = RREG32(CG_SPLL_FUNC_CNTL);
3989 tmp |= SPLL_SLEEP;
3990 WREG32(CG_SPLL_FUNC_CNTL, tmp);
3991
3992 tmp = RREG32(SPLL_CNTL_MODE);
3993 tmp &= ~SPLL_SW_DIR_CONTROL;
3994 WREG32(SPLL_CNTL_MODE, tmp);
3995}
3996
3997static void si_gpu_pci_config_reset(struct radeon_device *rdev)
3998{
3999 struct evergreen_mc_save save;
4000 u32 tmp, i;
4001
4002 dev_info(rdev->dev, "GPU pci config reset\n");
4003
4004 /* disable dpm? */
4005
4006 /* disable cg/pg */
4007 si_fini_pg(rdev);
4008 si_fini_cg(rdev);
4009
4010 /* Disable CP parsing/prefetching */
4011 WREG32(CP_ME_CNTL, CP_ME_HALT | CP_PFP_HALT | CP_CE_HALT);
4012 /* dma0 */
4013 tmp = RREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET);
4014 tmp &= ~DMA_RB_ENABLE;
4015 WREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET, tmp);
4016 /* dma1 */
4017 tmp = RREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET);
4018 tmp &= ~DMA_RB_ENABLE;
4019 WREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET, tmp);
4020 /* XXX other engines? */
4021
4022 /* halt the rlc, disable cp internal ints */
4023 si_rlc_stop(rdev);
4024
4025 udelay(50);
4026
4027 /* disable mem access */
4028 evergreen_mc_stop(rdev, &save);
4029 if (evergreen_mc_wait_for_idle(rdev)) {
4030 dev_warn(rdev->dev, "Wait for MC idle timed out !\n");
4031 }
4032
4033 /* set mclk/sclk to bypass */
4034 si_set_clk_bypass_mode(rdev);
4035 /* powerdown spll */
4036 si_spll_powerdown(rdev);
4037 /* disable BM */
4038 pci_clear_master(rdev->pdev);
4039 /* reset */
4040 radeon_pci_config_reset(rdev);
4041 /* wait for asic to come out of reset */
4042 for (i = 0; i < rdev->usec_timeout; i++) {
4043 if (RREG32(CONFIG_MEMSIZE) != 0xffffffff)
4044 break;
4045 udelay(1);
4046 }
4047}
4048
c476dde2
AD
4049int si_asic_reset(struct radeon_device *rdev)
4050{
014bb209
AD
4051 u32 reset_mask;
4052
4053 reset_mask = si_gpu_check_soft_reset(rdev);
4054
4055 if (reset_mask)
4056 r600_set_bios_scratch_engine_hung(rdev, true);
4057
4a5c8ea5 4058 /* try soft reset */
014bb209
AD
4059 si_gpu_soft_reset(rdev, reset_mask);
4060
4061 reset_mask = si_gpu_check_soft_reset(rdev);
4062
4a5c8ea5
AD
4063 /* try pci config reset */
4064 if (reset_mask && radeon_hard_reset)
4065 si_gpu_pci_config_reset(rdev);
4066
4067 reset_mask = si_gpu_check_soft_reset(rdev);
4068
014bb209
AD
4069 if (!reset_mask)
4070 r600_set_bios_scratch_engine_hung(rdev, false);
4071
4072 return 0;
c476dde2
AD
4073}
4074
123bc183
AD
4075/**
4076 * si_gfx_is_lockup - Check if the GFX engine is locked up
4077 *
4078 * @rdev: radeon_device pointer
4079 * @ring: radeon_ring structure holding ring information
4080 *
4081 * Check if the GFX engine is locked up.
4082 * Returns true if the engine appears to be locked up, false if not.
4083 */
4084bool si_gfx_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
4085{
4086 u32 reset_mask = si_gpu_check_soft_reset(rdev);
4087
4088 if (!(reset_mask & (RADEON_RESET_GFX |
4089 RADEON_RESET_COMPUTE |
4090 RADEON_RESET_CP))) {
ff212f25 4091 radeon_ring_lockup_update(rdev, ring);
123bc183
AD
4092 return false;
4093 }
123bc183
AD
4094 return radeon_ring_test_lockup(rdev, ring);
4095}
4096
d2800ee5
AD
4097/* MC */
4098static void si_mc_program(struct radeon_device *rdev)
4099{
4100 struct evergreen_mc_save save;
4101 u32 tmp;
4102 int i, j;
4103
4104 /* Initialize HDP */
4105 for (i = 0, j = 0; i < 32; i++, j += 0x18) {
4106 WREG32((0x2c14 + j), 0x00000000);
4107 WREG32((0x2c18 + j), 0x00000000);
4108 WREG32((0x2c1c + j), 0x00000000);
4109 WREG32((0x2c20 + j), 0x00000000);
4110 WREG32((0x2c24 + j), 0x00000000);
4111 }
4112 WREG32(HDP_REG_COHERENCY_FLUSH_CNTL, 0);
4113
4114 evergreen_mc_stop(rdev, &save);
4115 if (radeon_mc_wait_for_idle(rdev)) {
4116 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
4117 }
5153550a
AD
4118 if (!ASIC_IS_NODCE(rdev))
4119 /* Lockout access through VGA aperture*/
4120 WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE);
d2800ee5
AD
4121 /* Update configuration */
4122 WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
4123 rdev->mc.vram_start >> 12);
4124 WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
4125 rdev->mc.vram_end >> 12);
4126 WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR,
4127 rdev->vram_scratch.gpu_addr >> 12);
4128 tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16;
4129 tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF);
4130 WREG32(MC_VM_FB_LOCATION, tmp);
4131 /* XXX double check these! */
4132 WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8));
4133 WREG32(HDP_NONSURFACE_INFO, (2 << 7) | (1 << 30));
4134 WREG32(HDP_NONSURFACE_SIZE, 0x3FFFFFFF);
4135 WREG32(MC_VM_AGP_BASE, 0);
4136 WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF);
4137 WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF);
4138 if (radeon_mc_wait_for_idle(rdev)) {
4139 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
4140 }
4141 evergreen_mc_resume(rdev, &save);
5153550a
AD
4142 if (!ASIC_IS_NODCE(rdev)) {
4143 /* we need to own VRAM, so turn off the VGA renderer here
4144 * to stop it overwriting our objects */
4145 rv515_vga_render_disable(rdev);
4146 }
d2800ee5
AD
4147}
4148
1c49165d
AD
4149void si_vram_gtt_location(struct radeon_device *rdev,
4150 struct radeon_mc *mc)
d2800ee5
AD
4151{
4152 if (mc->mc_vram_size > 0xFFC0000000ULL) {
4153 /* leave room for at least 1024M GTT */
4154 dev_warn(rdev->dev, "limiting VRAM\n");
4155 mc->real_vram_size = 0xFFC0000000ULL;
4156 mc->mc_vram_size = 0xFFC0000000ULL;
4157 }
9ed8b1f9 4158 radeon_vram_location(rdev, &rdev->mc, 0);
d2800ee5 4159 rdev->mc.gtt_base_align = 0;
9ed8b1f9 4160 radeon_gtt_location(rdev, mc);
d2800ee5
AD
4161}
4162
4163static int si_mc_init(struct radeon_device *rdev)
4164{
4165 u32 tmp;
4166 int chansize, numchan;
4167
4168 /* Get VRAM informations */
4169 rdev->mc.vram_is_ddr = true;
4170 tmp = RREG32(MC_ARB_RAMCFG);
4171 if (tmp & CHANSIZE_OVERRIDE) {
4172 chansize = 16;
4173 } else if (tmp & CHANSIZE_MASK) {
4174 chansize = 64;
4175 } else {
4176 chansize = 32;
4177 }
4178 tmp = RREG32(MC_SHARED_CHMAP);
4179 switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
4180 case 0:
4181 default:
4182 numchan = 1;
4183 break;
4184 case 1:
4185 numchan = 2;
4186 break;
4187 case 2:
4188 numchan = 4;
4189 break;
4190 case 3:
4191 numchan = 8;
4192 break;
4193 case 4:
4194 numchan = 3;
4195 break;
4196 case 5:
4197 numchan = 6;
4198 break;
4199 case 6:
4200 numchan = 10;
4201 break;
4202 case 7:
4203 numchan = 12;
4204 break;
4205 case 8:
4206 numchan = 16;
4207 break;
4208 }
4209 rdev->mc.vram_width = numchan * chansize;
4210 /* Could aper size report 0 ? */
4211 rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
4212 rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
4213 /* size in MB on si */
0ca223b0
AD
4214 tmp = RREG32(CONFIG_MEMSIZE);
4215 /* some boards may have garbage in the upper 16 bits */
4216 if (tmp & 0xffff0000) {
4217 DRM_INFO("Probable bad vram size: 0x%08x\n", tmp);
4218 if (tmp & 0xffff)
4219 tmp &= 0xffff;
4220 }
4221 rdev->mc.mc_vram_size = tmp * 1024ULL * 1024ULL;
4222 rdev->mc.real_vram_size = rdev->mc.mc_vram_size;
d2800ee5
AD
4223 rdev->mc.visible_vram_size = rdev->mc.aper_size;
4224 si_vram_gtt_location(rdev, &rdev->mc);
4225 radeon_update_bandwidth_info(rdev);
4226
4227 return 0;
4228}
4229
4230/*
4231 * GART
4232 */
4233void si_pcie_gart_tlb_flush(struct radeon_device *rdev)
4234{
4235 /* flush hdp cache */
4236 WREG32(HDP_MEM_COHERENCY_FLUSH_CNTL, 0x1);
4237
4238 /* bits 0-15 are the VM contexts0-15 */
4239 WREG32(VM_INVALIDATE_REQUEST, 1);
4240}
4241
1109ca09 4242static int si_pcie_gart_enable(struct radeon_device *rdev)
d2800ee5
AD
4243{
4244 int r, i;
4245
4246 if (rdev->gart.robj == NULL) {
4247 dev_err(rdev->dev, "No VRAM object for PCIE GART.\n");
4248 return -EINVAL;
4249 }
4250 r = radeon_gart_table_vram_pin(rdev);
4251 if (r)
4252 return r;
d2800ee5
AD
4253 /* Setup TLB control */
4254 WREG32(MC_VM_MX_L1_TLB_CNTL,
4255 (0xA << 7) |
4256 ENABLE_L1_TLB |
ec3dbbcb 4257 ENABLE_L1_FRAGMENT_PROCESSING |
d2800ee5
AD
4258 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
4259 ENABLE_ADVANCED_DRIVER_MODEL |
4260 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
4261 /* Setup L2 cache */
4262 WREG32(VM_L2_CNTL, ENABLE_L2_CACHE |
ec3dbbcb 4263 ENABLE_L2_FRAGMENT_PROCESSING |
d2800ee5
AD
4264 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
4265 ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
4266 EFFECTIVE_L2_QUEUE_SIZE(7) |
4267 CONTEXT1_IDENTITY_ACCESS_MODE(1));
4268 WREG32(VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS | INVALIDATE_L2_CACHE);
4269 WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
ec3dbbcb
CK
4270 BANK_SELECT(4) |
4271 L2_CACHE_BIGK_FRAGMENT_SIZE(4));
d2800ee5
AD
4272 /* setup context0 */
4273 WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);
4274 WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);
4275 WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12);
4276 WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
4277 (u32)(rdev->dummy_page.addr >> 12));
4278 WREG32(VM_CONTEXT0_CNTL2, 0);
4279 WREG32(VM_CONTEXT0_CNTL, (ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) |
4280 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT));
4281
4282 WREG32(0x15D4, 0);
4283 WREG32(0x15D8, 0);
4284 WREG32(0x15DC, 0);
4285
4286 /* empty context1-15 */
d2800ee5
AD
4287 /* set vm size, must be a multiple of 4 */
4288 WREG32(VM_CONTEXT1_PAGE_TABLE_START_ADDR, 0);
c21b328e 4289 WREG32(VM_CONTEXT1_PAGE_TABLE_END_ADDR, rdev->vm_manager.max_pfn);
23d4f1f2
AD
4290 /* Assign the pt base to something valid for now; the pts used for
4291 * the VMs are determined by the application and setup and assigned
4292 * on the fly in the vm part of radeon_gart.c
4293 */
d2800ee5
AD
4294 for (i = 1; i < 16; i++) {
4295 if (i < 8)
4296 WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (i << 2),
054e01d6 4297 rdev->vm_manager.saved_table_addr[i]);
d2800ee5
AD
4298 else
4299 WREG32(VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((i - 8) << 2),
054e01d6 4300 rdev->vm_manager.saved_table_addr[i]);
d2800ee5
AD
4301 }
4302
4303 /* enable context1-15 */
4304 WREG32(VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR,
4305 (u32)(rdev->dummy_page.addr >> 12));
ae133a11 4306 WREG32(VM_CONTEXT1_CNTL2, 4);
fa87e62d 4307 WREG32(VM_CONTEXT1_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(1) |
4510fb98 4308 PAGE_TABLE_BLOCK_SIZE(radeon_vm_block_size - 9) |
ae133a11
CK
4309 RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT |
4310 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT |
4311 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT |
4312 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT |
4313 PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT |
4314 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT |
4315 VALID_PROTECTION_FAULT_ENABLE_INTERRUPT |
4316 VALID_PROTECTION_FAULT_ENABLE_DEFAULT |
4317 READ_PROTECTION_FAULT_ENABLE_INTERRUPT |
4318 READ_PROTECTION_FAULT_ENABLE_DEFAULT |
4319 WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT |
4320 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT);
d2800ee5
AD
4321
4322 si_pcie_gart_tlb_flush(rdev);
4323 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
4324 (unsigned)(rdev->mc.gtt_size >> 20),
4325 (unsigned long long)rdev->gart.table_addr);
4326 rdev->gart.ready = true;
4327 return 0;
4328}
4329
1109ca09 4330static void si_pcie_gart_disable(struct radeon_device *rdev)
d2800ee5 4331{
054e01d6
CK
4332 unsigned i;
4333
4334 for (i = 1; i < 16; ++i) {
4335 uint32_t reg;
4336 if (i < 8)
4337 reg = VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (i << 2);
4338 else
4339 reg = VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((i - 8) << 2);
4340 rdev->vm_manager.saved_table_addr[i] = RREG32(reg);
4341 }
4342
d2800ee5
AD
4343 /* Disable all tables */
4344 WREG32(VM_CONTEXT0_CNTL, 0);
4345 WREG32(VM_CONTEXT1_CNTL, 0);
4346 /* Setup TLB control */
4347 WREG32(MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE_NOT_IN_SYS |
4348 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
4349 /* Setup L2 cache */
4350 WREG32(VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
4351 ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
4352 EFFECTIVE_L2_QUEUE_SIZE(7) |
4353 CONTEXT1_IDENTITY_ACCESS_MODE(1));
4354 WREG32(VM_L2_CNTL2, 0);
4355 WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
4356 L2_CACHE_BIGK_FRAGMENT_SIZE(0));
4357 radeon_gart_table_vram_unpin(rdev);
4358}
4359
1109ca09 4360static void si_pcie_gart_fini(struct radeon_device *rdev)
d2800ee5
AD
4361{
4362 si_pcie_gart_disable(rdev);
4363 radeon_gart_table_vram_free(rdev);
4364 radeon_gart_fini(rdev);
4365}
4366
498dd8b3
AD
4367/* vm parser */
4368static bool si_vm_reg_valid(u32 reg)
4369{
4370 /* context regs are fine */
4371 if (reg >= 0x28000)
4372 return true;
4373
4374 /* check config regs */
4375 switch (reg) {
4376 case GRBM_GFX_INDEX:
f418b88a 4377 case CP_STRMOUT_CNTL:
498dd8b3
AD
4378 case VGT_VTX_VECT_EJECT_REG:
4379 case VGT_CACHE_INVALIDATION:
4380 case VGT_ESGS_RING_SIZE:
4381 case VGT_GSVS_RING_SIZE:
4382 case VGT_GS_VERTEX_REUSE:
4383 case VGT_PRIMITIVE_TYPE:
4384 case VGT_INDEX_TYPE:
4385 case VGT_NUM_INDICES:
4386 case VGT_NUM_INSTANCES:
4387 case VGT_TF_RING_SIZE:
4388 case VGT_HS_OFFCHIP_PARAM:
4389 case VGT_TF_MEMORY_BASE:
4390 case PA_CL_ENHANCE:
4391 case PA_SU_LINE_STIPPLE_VALUE:
4392 case PA_SC_LINE_STIPPLE_STATE:
4393 case PA_SC_ENHANCE:
4394 case SQC_CACHES:
4395 case SPI_STATIC_THREAD_MGMT_1:
4396 case SPI_STATIC_THREAD_MGMT_2:
4397 case SPI_STATIC_THREAD_MGMT_3:
4398 case SPI_PS_MAX_WAVE_ID:
4399 case SPI_CONFIG_CNTL:
4400 case SPI_CONFIG_CNTL_1:
4401 case TA_CNTL_AUX:
4402 return true;
4403 default:
4404 DRM_ERROR("Invalid register 0x%x in CS\n", reg);
4405 return false;
4406 }
4407}
4408
4409static int si_vm_packet3_ce_check(struct radeon_device *rdev,
4410 u32 *ib, struct radeon_cs_packet *pkt)
4411{
4412 switch (pkt->opcode) {
4413 case PACKET3_NOP:
4414 case PACKET3_SET_BASE:
4415 case PACKET3_SET_CE_DE_COUNTERS:
4416 case PACKET3_LOAD_CONST_RAM:
4417 case PACKET3_WRITE_CONST_RAM:
4418 case PACKET3_WRITE_CONST_RAM_OFFSET:
4419 case PACKET3_DUMP_CONST_RAM:
4420 case PACKET3_INCREMENT_CE_COUNTER:
4421 case PACKET3_WAIT_ON_DE_COUNTER:
4422 case PACKET3_CE_WRITE:
4423 break;
4424 default:
4425 DRM_ERROR("Invalid CE packet3: 0x%x\n", pkt->opcode);
4426 return -EINVAL;
4427 }
4428 return 0;
4429}
4430
e5b9e750
TS
4431static int si_vm_packet3_cp_dma_check(u32 *ib, u32 idx)
4432{
4433 u32 start_reg, reg, i;
4434 u32 command = ib[idx + 4];
4435 u32 info = ib[idx + 1];
4436 u32 idx_value = ib[idx];
4437 if (command & PACKET3_CP_DMA_CMD_SAS) {
4438 /* src address space is register */
4439 if (((info & 0x60000000) >> 29) == 0) {
4440 start_reg = idx_value << 2;
4441 if (command & PACKET3_CP_DMA_CMD_SAIC) {
4442 reg = start_reg;
4443 if (!si_vm_reg_valid(reg)) {
4444 DRM_ERROR("CP DMA Bad SRC register\n");
4445 return -EINVAL;
4446 }
4447 } else {
4448 for (i = 0; i < (command & 0x1fffff); i++) {
4449 reg = start_reg + (4 * i);
4450 if (!si_vm_reg_valid(reg)) {
4451 DRM_ERROR("CP DMA Bad SRC register\n");
4452 return -EINVAL;
4453 }
4454 }
4455 }
4456 }
4457 }
4458 if (command & PACKET3_CP_DMA_CMD_DAS) {
4459 /* dst address space is register */
4460 if (((info & 0x00300000) >> 20) == 0) {
4461 start_reg = ib[idx + 2];
4462 if (command & PACKET3_CP_DMA_CMD_DAIC) {
4463 reg = start_reg;
4464 if (!si_vm_reg_valid(reg)) {
4465 DRM_ERROR("CP DMA Bad DST register\n");
4466 return -EINVAL;
4467 }
4468 } else {
4469 for (i = 0; i < (command & 0x1fffff); i++) {
4470 reg = start_reg + (4 * i);
4471 if (!si_vm_reg_valid(reg)) {
4472 DRM_ERROR("CP DMA Bad DST register\n");
4473 return -EINVAL;
4474 }
4475 }
4476 }
4477 }
4478 }
4479 return 0;
4480}
4481
498dd8b3
AD
4482static int si_vm_packet3_gfx_check(struct radeon_device *rdev,
4483 u32 *ib, struct radeon_cs_packet *pkt)
4484{
e5b9e750 4485 int r;
498dd8b3
AD
4486 u32 idx = pkt->idx + 1;
4487 u32 idx_value = ib[idx];
4488 u32 start_reg, end_reg, reg, i;
4489
4490 switch (pkt->opcode) {
4491 case PACKET3_NOP:
4492 case PACKET3_SET_BASE:
4493 case PACKET3_CLEAR_STATE:
4494 case PACKET3_INDEX_BUFFER_SIZE:
4495 case PACKET3_DISPATCH_DIRECT:
4496 case PACKET3_DISPATCH_INDIRECT:
4497 case PACKET3_ALLOC_GDS:
4498 case PACKET3_WRITE_GDS_RAM:
4499 case PACKET3_ATOMIC_GDS:
4500 case PACKET3_ATOMIC:
4501 case PACKET3_OCCLUSION_QUERY:
4502 case PACKET3_SET_PREDICATION:
4503 case PACKET3_COND_EXEC:
4504 case PACKET3_PRED_EXEC:
4505 case PACKET3_DRAW_INDIRECT:
4506 case PACKET3_DRAW_INDEX_INDIRECT:
4507 case PACKET3_INDEX_BASE:
4508 case PACKET3_DRAW_INDEX_2:
4509 case PACKET3_CONTEXT_CONTROL:
4510 case PACKET3_INDEX_TYPE:
4511 case PACKET3_DRAW_INDIRECT_MULTI:
4512 case PACKET3_DRAW_INDEX_AUTO:
4513 case PACKET3_DRAW_INDEX_IMMD:
4514 case PACKET3_NUM_INSTANCES:
4515 case PACKET3_DRAW_INDEX_MULTI_AUTO:
4516 case PACKET3_STRMOUT_BUFFER_UPDATE:
4517 case PACKET3_DRAW_INDEX_OFFSET_2:
4518 case PACKET3_DRAW_INDEX_MULTI_ELEMENT:
4519 case PACKET3_DRAW_INDEX_INDIRECT_MULTI:
4520 case PACKET3_MPEG_INDEX:
4521 case PACKET3_WAIT_REG_MEM:
4522 case PACKET3_MEM_WRITE:
4523 case PACKET3_PFP_SYNC_ME:
4524 case PACKET3_SURFACE_SYNC:
4525 case PACKET3_EVENT_WRITE:
4526 case PACKET3_EVENT_WRITE_EOP:
4527 case PACKET3_EVENT_WRITE_EOS:
4528 case PACKET3_SET_CONTEXT_REG:
4529 case PACKET3_SET_CONTEXT_REG_INDIRECT:
4530 case PACKET3_SET_SH_REG:
4531 case PACKET3_SET_SH_REG_OFFSET:
4532 case PACKET3_INCREMENT_DE_COUNTER:
4533 case PACKET3_WAIT_ON_CE_COUNTER:
4534 case PACKET3_WAIT_ON_AVAIL_BUFFER:
4535 case PACKET3_ME_WRITE:
4536 break;
4537 case PACKET3_COPY_DATA:
4538 if ((idx_value & 0xf00) == 0) {
4539 reg = ib[idx + 3] * 4;
4540 if (!si_vm_reg_valid(reg))
4541 return -EINVAL;
4542 }
4543 break;
4544 case PACKET3_WRITE_DATA:
4545 if ((idx_value & 0xf00) == 0) {
4546 start_reg = ib[idx + 1] * 4;
4547 if (idx_value & 0x10000) {
4548 if (!si_vm_reg_valid(start_reg))
4549 return -EINVAL;
4550 } else {
4551 for (i = 0; i < (pkt->count - 2); i++) {
4552 reg = start_reg + (4 * i);
4553 if (!si_vm_reg_valid(reg))
4554 return -EINVAL;
4555 }
4556 }
4557 }
4558 break;
4559 case PACKET3_COND_WRITE:
4560 if (idx_value & 0x100) {
4561 reg = ib[idx + 5] * 4;
4562 if (!si_vm_reg_valid(reg))
4563 return -EINVAL;
4564 }
4565 break;
4566 case PACKET3_COPY_DW:
4567 if (idx_value & 0x2) {
4568 reg = ib[idx + 3] * 4;
4569 if (!si_vm_reg_valid(reg))
4570 return -EINVAL;
4571 }
4572 break;
4573 case PACKET3_SET_CONFIG_REG:
4574 start_reg = (idx_value << 2) + PACKET3_SET_CONFIG_REG_START;
4575 end_reg = 4 * pkt->count + start_reg - 4;
4576 if ((start_reg < PACKET3_SET_CONFIG_REG_START) ||
4577 (start_reg >= PACKET3_SET_CONFIG_REG_END) ||
4578 (end_reg >= PACKET3_SET_CONFIG_REG_END)) {
4579 DRM_ERROR("bad PACKET3_SET_CONFIG_REG\n");
4580 return -EINVAL;
4581 }
4582 for (i = 0; i < pkt->count; i++) {
4583 reg = start_reg + (4 * i);
4584 if (!si_vm_reg_valid(reg))
4585 return -EINVAL;
4586 }
4587 break;
5aa709be 4588 case PACKET3_CP_DMA:
e5b9e750
TS
4589 r = si_vm_packet3_cp_dma_check(ib, idx);
4590 if (r)
4591 return r;
5aa709be 4592 break;
498dd8b3
AD
4593 default:
4594 DRM_ERROR("Invalid GFX packet3: 0x%x\n", pkt->opcode);
4595 return -EINVAL;
4596 }
4597 return 0;
4598}
4599
4600static int si_vm_packet3_compute_check(struct radeon_device *rdev,
4601 u32 *ib, struct radeon_cs_packet *pkt)
4602{
e5b9e750 4603 int r;
498dd8b3
AD
4604 u32 idx = pkt->idx + 1;
4605 u32 idx_value = ib[idx];
4606 u32 start_reg, reg, i;
4607
4608 switch (pkt->opcode) {
4609 case PACKET3_NOP:
4610 case PACKET3_SET_BASE:
4611 case PACKET3_CLEAR_STATE:
4612 case PACKET3_DISPATCH_DIRECT:
4613 case PACKET3_DISPATCH_INDIRECT:
4614 case PACKET3_ALLOC_GDS:
4615 case PACKET3_WRITE_GDS_RAM:
4616 case PACKET3_ATOMIC_GDS:
4617 case PACKET3_ATOMIC:
4618 case PACKET3_OCCLUSION_QUERY:
4619 case PACKET3_SET_PREDICATION:
4620 case PACKET3_COND_EXEC:
4621 case PACKET3_PRED_EXEC:
4622 case PACKET3_CONTEXT_CONTROL:
4623 case PACKET3_STRMOUT_BUFFER_UPDATE:
4624 case PACKET3_WAIT_REG_MEM:
4625 case PACKET3_MEM_WRITE:
4626 case PACKET3_PFP_SYNC_ME:
4627 case PACKET3_SURFACE_SYNC:
4628 case PACKET3_EVENT_WRITE:
4629 case PACKET3_EVENT_WRITE_EOP:
4630 case PACKET3_EVENT_WRITE_EOS:
4631 case PACKET3_SET_CONTEXT_REG:
4632 case PACKET3_SET_CONTEXT_REG_INDIRECT:
4633 case PACKET3_SET_SH_REG:
4634 case PACKET3_SET_SH_REG_OFFSET:
4635 case PACKET3_INCREMENT_DE_COUNTER:
4636 case PACKET3_WAIT_ON_CE_COUNTER:
4637 case PACKET3_WAIT_ON_AVAIL_BUFFER:
4638 case PACKET3_ME_WRITE:
4639 break;
4640 case PACKET3_COPY_DATA:
4641 if ((idx_value & 0xf00) == 0) {
4642 reg = ib[idx + 3] * 4;
4643 if (!si_vm_reg_valid(reg))
4644 return -EINVAL;
4645 }
4646 break;
4647 case PACKET3_WRITE_DATA:
4648 if ((idx_value & 0xf00) == 0) {
4649 start_reg = ib[idx + 1] * 4;
4650 if (idx_value & 0x10000) {
4651 if (!si_vm_reg_valid(start_reg))
4652 return -EINVAL;
4653 } else {
4654 for (i = 0; i < (pkt->count - 2); i++) {
4655 reg = start_reg + (4 * i);
4656 if (!si_vm_reg_valid(reg))
4657 return -EINVAL;
4658 }
4659 }
4660 }
4661 break;
4662 case PACKET3_COND_WRITE:
4663 if (idx_value & 0x100) {
4664 reg = ib[idx + 5] * 4;
4665 if (!si_vm_reg_valid(reg))
4666 return -EINVAL;
4667 }
4668 break;
4669 case PACKET3_COPY_DW:
4670 if (idx_value & 0x2) {
4671 reg = ib[idx + 3] * 4;
4672 if (!si_vm_reg_valid(reg))
4673 return -EINVAL;
4674 }
4675 break;
e5b9e750
TS
4676 case PACKET3_CP_DMA:
4677 r = si_vm_packet3_cp_dma_check(ib, idx);
4678 if (r)
4679 return r;
4680 break;
498dd8b3
AD
4681 default:
4682 DRM_ERROR("Invalid Compute packet3: 0x%x\n", pkt->opcode);
4683 return -EINVAL;
4684 }
4685 return 0;
4686}
4687
4688int si_ib_parse(struct radeon_device *rdev, struct radeon_ib *ib)
4689{
4690 int ret = 0;
9d0223d5 4691 u32 idx = 0, i;
498dd8b3
AD
4692 struct radeon_cs_packet pkt;
4693
4694 do {
4695 pkt.idx = idx;
4e872ae2
IH
4696 pkt.type = RADEON_CP_PACKET_GET_TYPE(ib->ptr[idx]);
4697 pkt.count = RADEON_CP_PACKET_GET_COUNT(ib->ptr[idx]);
498dd8b3
AD
4698 pkt.one_reg_wr = 0;
4699 switch (pkt.type) {
4e872ae2 4700 case RADEON_PACKET_TYPE0:
498dd8b3
AD
4701 dev_err(rdev->dev, "Packet0 not allowed!\n");
4702 ret = -EINVAL;
4703 break;
4e872ae2 4704 case RADEON_PACKET_TYPE2:
498dd8b3
AD
4705 idx += 1;
4706 break;
4e872ae2
IH
4707 case RADEON_PACKET_TYPE3:
4708 pkt.opcode = RADEON_CP_PACKET3_GET_OPCODE(ib->ptr[idx]);
498dd8b3
AD
4709 if (ib->is_const_ib)
4710 ret = si_vm_packet3_ce_check(rdev, ib->ptr, &pkt);
4711 else {
876dc9f3 4712 switch (ib->ring) {
498dd8b3
AD
4713 case RADEON_RING_TYPE_GFX_INDEX:
4714 ret = si_vm_packet3_gfx_check(rdev, ib->ptr, &pkt);
4715 break;
4716 case CAYMAN_RING_TYPE_CP1_INDEX:
4717 case CAYMAN_RING_TYPE_CP2_INDEX:
4718 ret = si_vm_packet3_compute_check(rdev, ib->ptr, &pkt);
4719 break;
4720 default:
876dc9f3 4721 dev_err(rdev->dev, "Non-PM4 ring %d !\n", ib->ring);
498dd8b3
AD
4722 ret = -EINVAL;
4723 break;
4724 }
4725 }
4726 idx += pkt.count + 2;
4727 break;
4728 default:
4729 dev_err(rdev->dev, "Unknown packet type %d !\n", pkt.type);
4730 ret = -EINVAL;
4731 break;
4732 }
e1b4e722
AD
4733 if (ret) {
4734 for (i = 0; i < ib->length_dw; i++) {
4735 if (i == idx)
4736 printk("\t0x%08x <---\n", ib->ptr[i]);
4737 else
4738 printk("\t0x%08x\n", ib->ptr[i]);
4739 }
498dd8b3 4740 break;
e1b4e722 4741 }
498dd8b3
AD
4742 } while (idx < ib->length_dw);
4743
4744 return ret;
4745}
4746
d2800ee5
AD
4747/*
4748 * vm
4749 */
4750int si_vm_init(struct radeon_device *rdev)
4751{
4752 /* number of VMs */
4753 rdev->vm_manager.nvm = 16;
4754 /* base offset of vram pages */
4755 rdev->vm_manager.vram_base_offset = 0;
4756
4757 return 0;
4758}
4759
4760void si_vm_fini(struct radeon_device *rdev)
4761{
4762}
4763
fbf6dc7a
AD
4764/**
4765 * si_vm_decode_fault - print human readable fault info
4766 *
4767 * @rdev: radeon_device pointer
4768 * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value
4769 * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value
4770 *
4771 * Print human readable fault information (SI).
4772 */
4773static void si_vm_decode_fault(struct radeon_device *rdev,
4774 u32 status, u32 addr)
4775{
4776 u32 mc_id = (status & MEMORY_CLIENT_ID_MASK) >> MEMORY_CLIENT_ID_SHIFT;
4777 u32 vmid = (status & FAULT_VMID_MASK) >> FAULT_VMID_SHIFT;
4778 u32 protections = (status & PROTECTIONS_MASK) >> PROTECTIONS_SHIFT;
4779 char *block;
4780
4781 if (rdev->family == CHIP_TAHITI) {
4782 switch (mc_id) {
4783 case 160:
4784 case 144:
4785 case 96:
4786 case 80:
4787 case 224:
4788 case 208:
4789 case 32:
4790 case 16:
4791 block = "CB";
4792 break;
4793 case 161:
4794 case 145:
4795 case 97:
4796 case 81:
4797 case 225:
4798 case 209:
4799 case 33:
4800 case 17:
4801 block = "CB_FMASK";
4802 break;
4803 case 162:
4804 case 146:
4805 case 98:
4806 case 82:
4807 case 226:
4808 case 210:
4809 case 34:
4810 case 18:
4811 block = "CB_CMASK";
4812 break;
4813 case 163:
4814 case 147:
4815 case 99:
4816 case 83:
4817 case 227:
4818 case 211:
4819 case 35:
4820 case 19:
4821 block = "CB_IMMED";
4822 break;
4823 case 164:
4824 case 148:
4825 case 100:
4826 case 84:
4827 case 228:
4828 case 212:
4829 case 36:
4830 case 20:
4831 block = "DB";
4832 break;
4833 case 165:
4834 case 149:
4835 case 101:
4836 case 85:
4837 case 229:
4838 case 213:
4839 case 37:
4840 case 21:
4841 block = "DB_HTILE";
4842 break;
4843 case 167:
4844 case 151:
4845 case 103:
4846 case 87:
4847 case 231:
4848 case 215:
4849 case 39:
4850 case 23:
4851 block = "DB_STEN";
4852 break;
4853 case 72:
4854 case 68:
4855 case 64:
4856 case 8:
4857 case 4:
4858 case 0:
4859 case 136:
4860 case 132:
4861 case 128:
4862 case 200:
4863 case 196:
4864 case 192:
4865 block = "TC";
4866 break;
4867 case 112:
4868 case 48:
4869 block = "CP";
4870 break;
4871 case 49:
4872 case 177:
4873 case 50:
4874 case 178:
4875 block = "SH";
4876 break;
4877 case 53:
4878 case 190:
4879 block = "VGT";
4880 break;
4881 case 117:
4882 block = "IH";
4883 break;
4884 case 51:
4885 case 115:
4886 block = "RLC";
4887 break;
4888 case 119:
4889 case 183:
4890 block = "DMA0";
4891 break;
4892 case 61:
4893 block = "DMA1";
4894 break;
4895 case 248:
4896 case 120:
4897 block = "HDP";
4898 break;
4899 default:
4900 block = "unknown";
4901 break;
4902 }
4903 } else {
4904 switch (mc_id) {
4905 case 32:
4906 case 16:
4907 case 96:
4908 case 80:
4909 case 160:
4910 case 144:
4911 case 224:
4912 case 208:
4913 block = "CB";
4914 break;
4915 case 33:
4916 case 17:
4917 case 97:
4918 case 81:
4919 case 161:
4920 case 145:
4921 case 225:
4922 case 209:
4923 block = "CB_FMASK";
4924 break;
4925 case 34:
4926 case 18:
4927 case 98:
4928 case 82:
4929 case 162:
4930 case 146:
4931 case 226:
4932 case 210:
4933 block = "CB_CMASK";
4934 break;
4935 case 35:
4936 case 19:
4937 case 99:
4938 case 83:
4939 case 163:
4940 case 147:
4941 case 227:
4942 case 211:
4943 block = "CB_IMMED";
4944 break;
4945 case 36:
4946 case 20:
4947 case 100:
4948 case 84:
4949 case 164:
4950 case 148:
4951 case 228:
4952 case 212:
4953 block = "DB";
4954 break;
4955 case 37:
4956 case 21:
4957 case 101:
4958 case 85:
4959 case 165:
4960 case 149:
4961 case 229:
4962 case 213:
4963 block = "DB_HTILE";
4964 break;
4965 case 39:
4966 case 23:
4967 case 103:
4968 case 87:
4969 case 167:
4970 case 151:
4971 case 231:
4972 case 215:
4973 block = "DB_STEN";
4974 break;
4975 case 72:
4976 case 68:
4977 case 8:
4978 case 4:
4979 case 136:
4980 case 132:
4981 case 200:
4982 case 196:
4983 block = "TC";
4984 break;
4985 case 112:
4986 case 48:
4987 block = "CP";
4988 break;
4989 case 49:
4990 case 177:
4991 case 50:
4992 case 178:
4993 block = "SH";
4994 break;
4995 case 53:
4996 block = "VGT";
4997 break;
4998 case 117:
4999 block = "IH";
5000 break;
5001 case 51:
5002 case 115:
5003 block = "RLC";
5004 break;
5005 case 119:
5006 case 183:
5007 block = "DMA0";
5008 break;
5009 case 61:
5010 block = "DMA1";
5011 break;
5012 case 248:
5013 case 120:
5014 block = "HDP";
5015 break;
5016 default:
5017 block = "unknown";
5018 break;
5019 }
5020 }
5021
5022 printk("VM fault (0x%02x, vmid %d) at page %u, %s from %s (%d)\n",
5023 protections, vmid, addr,
5024 (status & MEMORY_CLIENT_RW_MASK) ? "write" : "read",
5025 block, mc_id);
5026}
5027
faffaf62
CK
5028void si_vm_flush(struct radeon_device *rdev, struct radeon_ring *ring,
5029 unsigned vm_id, uint64_t pd_addr)
d2800ee5 5030{
76c44f2c
AD
5031 /* write new base address */
5032 radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
f1d2a26b 5033 radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(1) |
76c44f2c
AD
5034 WRITE_DATA_DST_SEL(0)));
5035
faffaf62 5036 if (vm_id < 8) {
76c44f2c 5037 radeon_ring_write(ring,
faffaf62 5038 (VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm_id << 2)) >> 2);
ee60e29f 5039 } else {
76c44f2c 5040 radeon_ring_write(ring,
faffaf62 5041 (VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((vm_id - 8) << 2)) >> 2);
ee60e29f 5042 }
76c44f2c 5043 radeon_ring_write(ring, 0);
faffaf62 5044 radeon_ring_write(ring, pd_addr >> 12);
ee60e29f 5045
d2800ee5 5046 /* flush hdp cache */
76c44f2c 5047 radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4fb0bbd5 5048 radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(1) |
76c44f2c
AD
5049 WRITE_DATA_DST_SEL(0)));
5050 radeon_ring_write(ring, HDP_MEM_COHERENCY_FLUSH_CNTL >> 2);
5051 radeon_ring_write(ring, 0);
ee60e29f
CK
5052 radeon_ring_write(ring, 0x1);
5053
d2800ee5 5054 /* bits 0-15 are the VM contexts0-15 */
76c44f2c 5055 radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4fb0bbd5 5056 radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(1) |
76c44f2c
AD
5057 WRITE_DATA_DST_SEL(0)));
5058 radeon_ring_write(ring, VM_INVALIDATE_REQUEST >> 2);
5059 radeon_ring_write(ring, 0);
faffaf62 5060 radeon_ring_write(ring, 1 << vm_id);
58f8cf56 5061
d474ea7e
AD
5062 /* wait for the invalidate to complete */
5063 radeon_ring_write(ring, PACKET3(PACKET3_WAIT_REG_MEM, 5));
5064 radeon_ring_write(ring, (WAIT_REG_MEM_FUNCTION(0) | /* always */
5065 WAIT_REG_MEM_ENGINE(0))); /* me */
5066 radeon_ring_write(ring, VM_INVALIDATE_REQUEST >> 2);
5067 radeon_ring_write(ring, 0);
5068 radeon_ring_write(ring, 0); /* ref */
5069 radeon_ring_write(ring, 0); /* mask */
5070 radeon_ring_write(ring, 0x20); /* poll interval */
5071
58f8cf56
CK
5072 /* sync PFP to ME, otherwise we might get invalid PFP reads */
5073 radeon_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
5074 radeon_ring_write(ring, 0x0);
d2800ee5
AD
5075}
5076
f8f84ac5
AD
5077/*
5078 * Power and clock gating
5079 */
5080static void si_wait_for_rlc_serdes(struct radeon_device *rdev)
5081{
5082 int i;
5083
5084 for (i = 0; i < rdev->usec_timeout; i++) {
5085 if (RREG32(RLC_SERDES_MASTER_BUSY_0) == 0)
5086 break;
5087 udelay(1);
5088 }
5089
5090 for (i = 0; i < rdev->usec_timeout; i++) {
5091 if (RREG32(RLC_SERDES_MASTER_BUSY_1) == 0)
5092 break;
5093 udelay(1);
5094 }
5095}
5096
5097static void si_enable_gui_idle_interrupt(struct radeon_device *rdev,
5098 bool enable)
5099{
5100 u32 tmp = RREG32(CP_INT_CNTL_RING0);
5101 u32 mask;
5102 int i;
5103
5104 if (enable)
5105 tmp |= (CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
5106 else
5107 tmp &= ~(CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
5108 WREG32(CP_INT_CNTL_RING0, tmp);
5109
5110 if (!enable) {
5111 /* read a gfx register */
5112 tmp = RREG32(DB_DEPTH_INFO);
5113
5114 mask = RLC_BUSY_STATUS | GFX_POWER_STATUS | GFX_CLOCK_STATUS | GFX_LS_STATUS;
5115 for (i = 0; i < rdev->usec_timeout; i++) {
5116 if ((RREG32(RLC_STAT) & mask) == (GFX_CLOCK_STATUS | GFX_POWER_STATUS))
5117 break;
5118 udelay(1);
5119 }
5120 }
5121}
5122
5123static void si_set_uvd_dcm(struct radeon_device *rdev,
5124 bool sw_mode)
5125{
5126 u32 tmp, tmp2;
5127
5128 tmp = RREG32(UVD_CGC_CTRL);
5129 tmp &= ~(CLK_OD_MASK | CG_DT_MASK);
5130 tmp |= DCM | CG_DT(1) | CLK_OD(4);
5131
5132 if (sw_mode) {
5133 tmp &= ~0x7ffff800;
5134 tmp2 = DYN_OR_EN | DYN_RR_EN | G_DIV_ID(7);
5135 } else {
5136 tmp |= 0x7ffff800;
5137 tmp2 = 0;
5138 }
5139
5140 WREG32(UVD_CGC_CTRL, tmp);
5141 WREG32_UVD_CTX(UVD_CGC_CTRL2, tmp2);
5142}
5143
22c775ce 5144void si_init_uvd_internal_cg(struct radeon_device *rdev)
f8f84ac5
AD
5145{
5146 bool hw_mode = true;
5147
5148 if (hw_mode) {
5149 si_set_uvd_dcm(rdev, false);
5150 } else {
5151 u32 tmp = RREG32(UVD_CGC_CTRL);
5152 tmp &= ~DCM;
5153 WREG32(UVD_CGC_CTRL, tmp);
5154 }
5155}
5156
5157static u32 si_halt_rlc(struct radeon_device *rdev)
5158{
5159 u32 data, orig;
5160
5161 orig = data = RREG32(RLC_CNTL);
5162
5163 if (data & RLC_ENABLE) {
5164 data &= ~RLC_ENABLE;
5165 WREG32(RLC_CNTL, data);
5166
5167 si_wait_for_rlc_serdes(rdev);
5168 }
5169
5170 return orig;
5171}
5172
5173static void si_update_rlc(struct radeon_device *rdev, u32 rlc)
5174{
5175 u32 tmp;
5176
5177 tmp = RREG32(RLC_CNTL);
5178 if (tmp != rlc)
5179 WREG32(RLC_CNTL, rlc);
5180}
5181
5182static void si_enable_dma_pg(struct radeon_device *rdev, bool enable)
5183{
5184 u32 data, orig;
5185
5186 orig = data = RREG32(DMA_PG);
e16866ec 5187 if (enable && (rdev->pg_flags & RADEON_PG_SUPPORT_SDMA))
f8f84ac5
AD
5188 data |= PG_CNTL_ENABLE;
5189 else
5190 data &= ~PG_CNTL_ENABLE;
5191 if (orig != data)
5192 WREG32(DMA_PG, data);
5193}
5194
5195static void si_init_dma_pg(struct radeon_device *rdev)
5196{
5197 u32 tmp;
5198
5199 WREG32(DMA_PGFSM_WRITE, 0x00002000);
5200 WREG32(DMA_PGFSM_CONFIG, 0x100010ff);
5201
5202 for (tmp = 0; tmp < 5; tmp++)
5203 WREG32(DMA_PGFSM_WRITE, 0);
5204}
5205
5206static void si_enable_gfx_cgpg(struct radeon_device *rdev,
5207 bool enable)
5208{
5209 u32 tmp;
5210
2b19d17f 5211 if (enable && (rdev->pg_flags & RADEON_PG_SUPPORT_GFX_PG)) {
f8f84ac5
AD
5212 tmp = RLC_PUD(0x10) | RLC_PDD(0x10) | RLC_TTPD(0x10) | RLC_MSD(0x10);
5213 WREG32(RLC_TTOP_D, tmp);
5214
5215 tmp = RREG32(RLC_PG_CNTL);
5216 tmp |= GFX_PG_ENABLE;
5217 WREG32(RLC_PG_CNTL, tmp);
5218
5219 tmp = RREG32(RLC_AUTO_PG_CTRL);
5220 tmp |= AUTO_PG_EN;
5221 WREG32(RLC_AUTO_PG_CTRL, tmp);
5222 } else {
5223 tmp = RREG32(RLC_AUTO_PG_CTRL);
5224 tmp &= ~AUTO_PG_EN;
5225 WREG32(RLC_AUTO_PG_CTRL, tmp);
5226
5227 tmp = RREG32(DB_RENDER_CONTROL);
5228 }
5229}
5230
5231static void si_init_gfx_cgpg(struct radeon_device *rdev)
5232{
5233 u32 tmp;
5234
5235 WREG32(RLC_SAVE_AND_RESTORE_BASE, rdev->rlc.save_restore_gpu_addr >> 8);
5236
5237 tmp = RREG32(RLC_PG_CNTL);
5238 tmp |= GFX_PG_SRC;
5239 WREG32(RLC_PG_CNTL, tmp);
5240
5241 WREG32(RLC_CLEAR_STATE_RESTORE_BASE, rdev->rlc.clear_state_gpu_addr >> 8);
5242
5243 tmp = RREG32(RLC_AUTO_PG_CTRL);
5244
5245 tmp &= ~GRBM_REG_SGIT_MASK;
5246 tmp |= GRBM_REG_SGIT(0x700);
5247 tmp &= ~PG_AFTER_GRBM_REG_ST_MASK;
5248 WREG32(RLC_AUTO_PG_CTRL, tmp);
5249}
5250
ba19031a 5251static u32 si_get_cu_active_bitmap(struct radeon_device *rdev, u32 se, u32 sh)
f8f84ac5
AD
5252{
5253 u32 mask = 0, tmp, tmp1;
5254 int i;
5255
5256 si_select_se_sh(rdev, se, sh);
5257 tmp = RREG32(CC_GC_SHADER_ARRAY_CONFIG);
5258 tmp1 = RREG32(GC_USER_SHADER_ARRAY_CONFIG);
5259 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
5260
5261 tmp &= 0xffff0000;
5262
5263 tmp |= tmp1;
5264 tmp >>= 16;
5265
5266 for (i = 0; i < rdev->config.si.max_cu_per_sh; i ++) {
5267 mask <<= 1;
5268 mask |= 1;
5269 }
5270
5271 return (~tmp) & mask;
5272}
5273
5274static void si_init_ao_cu_mask(struct radeon_device *rdev)
5275{
5276 u32 i, j, k, active_cu_number = 0;
5277 u32 mask, counter, cu_bitmap;
5278 u32 tmp = 0;
5279
5280 for (i = 0; i < rdev->config.si.max_shader_engines; i++) {
5281 for (j = 0; j < rdev->config.si.max_sh_per_se; j++) {
5282 mask = 1;
5283 cu_bitmap = 0;
5284 counter = 0;
5285 for (k = 0; k < rdev->config.si.max_cu_per_sh; k++) {
ba19031a 5286 if (si_get_cu_active_bitmap(rdev, i, j) & mask) {
f8f84ac5
AD
5287 if (counter < 2)
5288 cu_bitmap |= mask;
5289 counter++;
5290 }
5291 mask <<= 1;
5292 }
5293
5294 active_cu_number += counter;
5295 tmp |= (cu_bitmap << (i * 16 + j * 8));
5296 }
5297 }
5298
5299 WREG32(RLC_PG_AO_CU_MASK, tmp);
5300
5301 tmp = RREG32(RLC_MAX_PG_CU);
5302 tmp &= ~MAX_PU_CU_MASK;
5303 tmp |= MAX_PU_CU(active_cu_number);
5304 WREG32(RLC_MAX_PG_CU, tmp);
5305}
5306
5307static void si_enable_cgcg(struct radeon_device *rdev,
5308 bool enable)
5309{
5310 u32 data, orig, tmp;
5311
5312 orig = data = RREG32(RLC_CGCG_CGLS_CTRL);
5313
e16866ec 5314 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_CGCG)) {
5594a558 5315 si_enable_gui_idle_interrupt(rdev, true);
f8f84ac5 5316
f8f84ac5
AD
5317 WREG32(RLC_GCPM_GENERAL_3, 0x00000080);
5318
5319 tmp = si_halt_rlc(rdev);
5320
5321 WREG32(RLC_SERDES_WR_MASTER_MASK_0, 0xffffffff);
5322 WREG32(RLC_SERDES_WR_MASTER_MASK_1, 0xffffffff);
5323 WREG32(RLC_SERDES_WR_CTRL, 0x00b000ff);
5324
5325 si_wait_for_rlc_serdes(rdev);
5326
5327 si_update_rlc(rdev, tmp);
5328
5329 WREG32(RLC_SERDES_WR_CTRL, 0x007000ff);
5330
5331 data |= CGCG_EN | CGLS_EN;
5332 } else {
5594a558
AD
5333 si_enable_gui_idle_interrupt(rdev, false);
5334
f8f84ac5
AD
5335 RREG32(CB_CGTT_SCLK_CTRL);
5336 RREG32(CB_CGTT_SCLK_CTRL);
5337 RREG32(CB_CGTT_SCLK_CTRL);
5338 RREG32(CB_CGTT_SCLK_CTRL);
5339
5340 data &= ~(CGCG_EN | CGLS_EN);
5341 }
5342
5343 if (orig != data)
5344 WREG32(RLC_CGCG_CGLS_CTRL, data);
5345}
5346
5347static void si_enable_mgcg(struct radeon_device *rdev,
5348 bool enable)
5349{
5350 u32 data, orig, tmp = 0;
5351
e16866ec 5352 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_MGCG)) {
f8f84ac5
AD
5353 orig = data = RREG32(CGTS_SM_CTRL_REG);
5354 data = 0x96940200;
5355 if (orig != data)
5356 WREG32(CGTS_SM_CTRL_REG, data);
5357
e16866ec
AD
5358 if (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_CP_LS) {
5359 orig = data = RREG32(CP_MEM_SLP_CNTL);
5360 data |= CP_MEM_LS_EN;
5361 if (orig != data)
5362 WREG32(CP_MEM_SLP_CNTL, data);
5363 }
f8f84ac5
AD
5364
5365 orig = data = RREG32(RLC_CGTT_MGCG_OVERRIDE);
5366 data &= 0xffffffc0;
5367 if (orig != data)
5368 WREG32(RLC_CGTT_MGCG_OVERRIDE, data);
5369
5370 tmp = si_halt_rlc(rdev);
5371
5372 WREG32(RLC_SERDES_WR_MASTER_MASK_0, 0xffffffff);
5373 WREG32(RLC_SERDES_WR_MASTER_MASK_1, 0xffffffff);
5374 WREG32(RLC_SERDES_WR_CTRL, 0x00d000ff);
5375
5376 si_update_rlc(rdev, tmp);
5377 } else {
5378 orig = data = RREG32(RLC_CGTT_MGCG_OVERRIDE);
5379 data |= 0x00000003;
5380 if (orig != data)
5381 WREG32(RLC_CGTT_MGCG_OVERRIDE, data);
5382
5383 data = RREG32(CP_MEM_SLP_CNTL);
5384 if (data & CP_MEM_LS_EN) {
5385 data &= ~CP_MEM_LS_EN;
5386 WREG32(CP_MEM_SLP_CNTL, data);
5387 }
5388 orig = data = RREG32(CGTS_SM_CTRL_REG);
5389 data |= LS_OVERRIDE | OVERRIDE;
5390 if (orig != data)
5391 WREG32(CGTS_SM_CTRL_REG, data);
5392
5393 tmp = si_halt_rlc(rdev);
5394
5395 WREG32(RLC_SERDES_WR_MASTER_MASK_0, 0xffffffff);
5396 WREG32(RLC_SERDES_WR_MASTER_MASK_1, 0xffffffff);
5397 WREG32(RLC_SERDES_WR_CTRL, 0x00e000ff);
5398
5399 si_update_rlc(rdev, tmp);
5400 }
5401}
5402
5403static void si_enable_uvd_mgcg(struct radeon_device *rdev,
5404 bool enable)
5405{
5406 u32 orig, data, tmp;
5407
e16866ec 5408 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_UVD_MGCG)) {
f8f84ac5
AD
5409 tmp = RREG32_UVD_CTX(UVD_CGC_MEM_CTRL);
5410 tmp |= 0x3fff;
5411 WREG32_UVD_CTX(UVD_CGC_MEM_CTRL, tmp);
5412
5413 orig = data = RREG32(UVD_CGC_CTRL);
5414 data |= DCM;
5415 if (orig != data)
5416 WREG32(UVD_CGC_CTRL, data);
5417
5418 WREG32_SMC(SMC_CG_IND_START + CG_CGTT_LOCAL_0, 0);
5419 WREG32_SMC(SMC_CG_IND_START + CG_CGTT_LOCAL_1, 0);
5420 } else {
5421 tmp = RREG32_UVD_CTX(UVD_CGC_MEM_CTRL);
5422 tmp &= ~0x3fff;
5423 WREG32_UVD_CTX(UVD_CGC_MEM_CTRL, tmp);
5424
5425 orig = data = RREG32(UVD_CGC_CTRL);
5426 data &= ~DCM;
5427 if (orig != data)
5428 WREG32(UVD_CGC_CTRL, data);
5429
5430 WREG32_SMC(SMC_CG_IND_START + CG_CGTT_LOCAL_0, 0xffffffff);
5431 WREG32_SMC(SMC_CG_IND_START + CG_CGTT_LOCAL_1, 0xffffffff);
5432 }
5433}
5434
5435static const u32 mc_cg_registers[] =
5436{
5437 MC_HUB_MISC_HUB_CG,
5438 MC_HUB_MISC_SIP_CG,
5439 MC_HUB_MISC_VM_CG,
5440 MC_XPB_CLK_GAT,
5441 ATC_MISC_CG,
5442 MC_CITF_MISC_WR_CG,
5443 MC_CITF_MISC_RD_CG,
5444 MC_CITF_MISC_VM_CG,
5445 VM_L2_CG,
5446};
5447
5448static void si_enable_mc_ls(struct radeon_device *rdev,
5449 bool enable)
5450{
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]);
e16866ec 5456 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_MC_LS))
f8f84ac5
AD
5457 data |= MC_LS_ENABLE;
5458 else
5459 data &= ~MC_LS_ENABLE;
5460 if (data != orig)
5461 WREG32(mc_cg_registers[i], data);
5462 }
5463}
5464
e16866ec
AD
5465static void si_enable_mc_mgcg(struct radeon_device *rdev,
5466 bool enable)
f8f84ac5 5467{
e16866ec
AD
5468 int i;
5469 u32 orig, data;
5470
5471 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) {
5472 orig = data = RREG32(mc_cg_registers[i]);
5473 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_MC_MGCG))
5474 data |= MC_CG_ENABLE;
5475 else
5476 data &= ~MC_CG_ENABLE;
5477 if (data != orig)
5478 WREG32(mc_cg_registers[i], data);
f8f84ac5
AD
5479 }
5480}
5481
e16866ec
AD
5482static void si_enable_dma_mgcg(struct radeon_device *rdev,
5483 bool enable)
f8f84ac5 5484{
e16866ec
AD
5485 u32 orig, data, offset;
5486 int i;
f8f84ac5 5487
e16866ec
AD
5488 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_SDMA_MGCG)) {
5489 for (i = 0; i < 2; i++) {
5490 if (i == 0)
5491 offset = DMA0_REGISTER_OFFSET;
5492 else
5493 offset = DMA1_REGISTER_OFFSET;
5494 orig = data = RREG32(DMA_POWER_CNTL + offset);
5495 data &= ~MEM_POWER_OVERRIDE;
5496 if (data != orig)
5497 WREG32(DMA_POWER_CNTL + offset, data);
5498 WREG32(DMA_CLK_CTRL + offset, 0x00000100);
5499 }
5500 } else {
5501 for (i = 0; i < 2; i++) {
5502 if (i == 0)
5503 offset = DMA0_REGISTER_OFFSET;
5504 else
5505 offset = DMA1_REGISTER_OFFSET;
5506 orig = data = RREG32(DMA_POWER_CNTL + offset);
5507 data |= MEM_POWER_OVERRIDE;
5508 if (data != orig)
5509 WREG32(DMA_POWER_CNTL + offset, data);
5510
5511 orig = data = RREG32(DMA_CLK_CTRL + offset);
5512 data = 0xff000000;
5513 if (data != orig)
5514 WREG32(DMA_CLK_CTRL + offset, data);
5515 }
5516 }
f8f84ac5
AD
5517}
5518
e16866ec
AD
5519static void si_enable_bif_mgls(struct radeon_device *rdev,
5520 bool enable)
f8f84ac5 5521{
e16866ec 5522 u32 orig, data;
f8f84ac5 5523
e16866ec 5524 orig = data = RREG32_PCIE(PCIE_CNTL2);
f8f84ac5 5525
e16866ec
AD
5526 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_BIF_LS))
5527 data |= SLV_MEM_LS_EN | MST_MEM_LS_EN |
5528 REPLAY_MEM_LS_EN | SLV_MEM_AGGRESSIVE_LS_EN;
5529 else
5530 data &= ~(SLV_MEM_LS_EN | MST_MEM_LS_EN |
5531 REPLAY_MEM_LS_EN | SLV_MEM_AGGRESSIVE_LS_EN);
5532
5533 if (orig != data)
5534 WREG32_PCIE(PCIE_CNTL2, data);
f8f84ac5
AD
5535}
5536
e16866ec
AD
5537static void si_enable_hdp_mgcg(struct radeon_device *rdev,
5538 bool enable)
f8f84ac5 5539{
e16866ec 5540 u32 orig, data;
f8f84ac5 5541
e16866ec 5542 orig = data = RREG32(HDP_HOST_PATH_CNTL);
f8f84ac5 5543
e16866ec
AD
5544 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_HDP_MGCG))
5545 data &= ~CLOCK_GATING_DIS;
5546 else
5547 data |= CLOCK_GATING_DIS;
5548
5549 if (orig != data)
5550 WREG32(HDP_HOST_PATH_CNTL, data);
f8f84ac5
AD
5551}
5552
e16866ec
AD
5553static void si_enable_hdp_ls(struct radeon_device *rdev,
5554 bool enable)
347e7592 5555{
e16866ec 5556 u32 orig, data;
347e7592 5557
e16866ec 5558 orig = data = RREG32(HDP_MEM_POWER_LS);
347e7592 5559
e16866ec
AD
5560 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_HDP_LS))
5561 data |= HDP_LS_ENABLE;
5562 else
5563 data &= ~HDP_LS_ENABLE;
347e7592 5564
e16866ec
AD
5565 if (orig != data)
5566 WREG32(HDP_MEM_POWER_LS, data);
347e7592
AD
5567}
5568
68e3a092
AD
5569static void si_update_cg(struct radeon_device *rdev,
5570 u32 block, bool enable)
347e7592 5571{
e16866ec 5572 if (block & RADEON_CG_BLOCK_GFX) {
811e4d58 5573 si_enable_gui_idle_interrupt(rdev, false);
e16866ec
AD
5574 /* order matters! */
5575 if (enable) {
5576 si_enable_mgcg(rdev, true);
5577 si_enable_cgcg(rdev, true);
5578 } else {
5579 si_enable_cgcg(rdev, false);
5580 si_enable_mgcg(rdev, false);
347e7592 5581 }
811e4d58 5582 si_enable_gui_idle_interrupt(rdev, true);
347e7592
AD
5583 }
5584
e16866ec
AD
5585 if (block & RADEON_CG_BLOCK_MC) {
5586 si_enable_mc_mgcg(rdev, enable);
5587 si_enable_mc_ls(rdev, enable);
347e7592 5588 }
e16866ec
AD
5589
5590 if (block & RADEON_CG_BLOCK_SDMA) {
5591 si_enable_dma_mgcg(rdev, enable);
347e7592
AD
5592 }
5593
e16866ec
AD
5594 if (block & RADEON_CG_BLOCK_BIF) {
5595 si_enable_bif_mgls(rdev, enable);
6d8cf000 5596 }
6d8cf000 5597
e16866ec
AD
5598 if (block & RADEON_CG_BLOCK_UVD) {
5599 if (rdev->has_uvd) {
5600 si_enable_uvd_mgcg(rdev, enable);
bd8cd539
AD
5601 }
5602 }
bd8cd539 5603
e16866ec
AD
5604 if (block & RADEON_CG_BLOCK_HDP) {
5605 si_enable_hdp_mgcg(rdev, enable);
5606 si_enable_hdp_ls(rdev, enable);
347e7592 5607 }
e16866ec 5608}
f8f84ac5
AD
5609
5610static void si_init_cg(struct radeon_device *rdev)
5611{
e16866ec
AD
5612 si_update_cg(rdev, (RADEON_CG_BLOCK_GFX |
5613 RADEON_CG_BLOCK_MC |
5614 RADEON_CG_BLOCK_SDMA |
5615 RADEON_CG_BLOCK_BIF |
5616 RADEON_CG_BLOCK_HDP), true);
b2d70917 5617 if (rdev->has_uvd) {
e16866ec 5618 si_update_cg(rdev, RADEON_CG_BLOCK_UVD, true);
f8f84ac5 5619 si_init_uvd_internal_cg(rdev);
347e7592 5620 }
f8f84ac5 5621}
bd8cd539 5622
f8f84ac5
AD
5623static void si_fini_cg(struct radeon_device *rdev)
5624{
0116e1ef 5625 if (rdev->has_uvd) {
e16866ec 5626 si_update_cg(rdev, RADEON_CG_BLOCK_UVD, false);
347e7592 5627 }
e16866ec
AD
5628 si_update_cg(rdev, (RADEON_CG_BLOCK_GFX |
5629 RADEON_CG_BLOCK_MC |
5630 RADEON_CG_BLOCK_SDMA |
5631 RADEON_CG_BLOCK_BIF |
5632 RADEON_CG_BLOCK_HDP), false);
5633}
5634
59a82d0e
AD
5635u32 si_get_csb_size(struct radeon_device *rdev)
5636{
5637 u32 count = 0;
5638 const struct cs_section_def *sect = NULL;
5639 const struct cs_extent_def *ext = NULL;
5640
5641 if (rdev->rlc.cs_data == NULL)
5642 return 0;
5643
5644 /* begin clear state */
5645 count += 2;
5646 /* context control state */
5647 count += 3;
5648
5649 for (sect = rdev->rlc.cs_data; sect->section != NULL; ++sect) {
5650 for (ext = sect->section; ext->extent != NULL; ++ext) {
5651 if (sect->id == SECT_CONTEXT)
5652 count += 2 + ext->reg_count;
5653 else
5654 return 0;
5655 }
bd8cd539 5656 }
59a82d0e
AD
5657 /* pa_sc_raster_config */
5658 count += 3;
5659 /* end clear state */
5660 count += 2;
5661 /* clear state */
5662 count += 2;
5663
5664 return count;
5665}
5666
5667void si_get_csb_buffer(struct radeon_device *rdev, volatile u32 *buffer)
5668{
5669 u32 count = 0, i;
5670 const struct cs_section_def *sect = NULL;
5671 const struct cs_extent_def *ext = NULL;
5672
5673 if (rdev->rlc.cs_data == NULL)
5674 return;
5675 if (buffer == NULL)
5676 return;
5677
6ba81e53
AD
5678 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0));
5679 buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
59a82d0e 5680
6ba81e53
AD
5681 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CONTEXT_CONTROL, 1));
5682 buffer[count++] = cpu_to_le32(0x80000000);
5683 buffer[count++] = cpu_to_le32(0x80000000);
59a82d0e
AD
5684
5685 for (sect = rdev->rlc.cs_data; sect->section != NULL; ++sect) {
5686 for (ext = sect->section; ext->extent != NULL; ++ext) {
5687 if (sect->id == SECT_CONTEXT) {
6ba81e53
AD
5688 buffer[count++] =
5689 cpu_to_le32(PACKET3(PACKET3_SET_CONTEXT_REG, ext->reg_count));
5690 buffer[count++] = cpu_to_le32(ext->reg_index - 0xa000);
59a82d0e 5691 for (i = 0; i < ext->reg_count; i++)
6ba81e53 5692 buffer[count++] = cpu_to_le32(ext->extent[i]);
59a82d0e
AD
5693 } else {
5694 return;
bd8cd539 5695 }
bd8cd539
AD
5696 }
5697 }
bd8cd539 5698
6ba81e53
AD
5699 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_SET_CONTEXT_REG, 1));
5700 buffer[count++] = cpu_to_le32(PA_SC_RASTER_CONFIG - PACKET3_SET_CONTEXT_REG_START);
59a82d0e
AD
5701 switch (rdev->family) {
5702 case CHIP_TAHITI:
5703 case CHIP_PITCAIRN:
6ba81e53 5704 buffer[count++] = cpu_to_le32(0x2a00126a);
59a82d0e
AD
5705 break;
5706 case CHIP_VERDE:
6ba81e53 5707 buffer[count++] = cpu_to_le32(0x0000124a);
59a82d0e
AD
5708 break;
5709 case CHIP_OLAND:
6ba81e53 5710 buffer[count++] = cpu_to_le32(0x00000082);
59a82d0e
AD
5711 break;
5712 case CHIP_HAINAN:
6ba81e53 5713 buffer[count++] = cpu_to_le32(0x00000000);
59a82d0e
AD
5714 break;
5715 default:
6ba81e53 5716 buffer[count++] = cpu_to_le32(0x00000000);
59a82d0e
AD
5717 break;
5718 }
5719
6ba81e53
AD
5720 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0));
5721 buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_END_CLEAR_STATE);
347e7592 5722
6ba81e53
AD
5723 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CLEAR_STATE, 0));
5724 buffer[count++] = cpu_to_le32(0);
59a82d0e
AD
5725}
5726
f8f84ac5
AD
5727static void si_init_pg(struct radeon_device *rdev)
5728{
0116e1ef
AD
5729 if (rdev->pg_flags) {
5730 if (rdev->pg_flags & RADEON_PG_SUPPORT_SDMA) {
5731 si_init_dma_pg(rdev);
0116e1ef 5732 }
f8f84ac5 5733 si_init_ao_cu_mask(rdev);
2b19d17f 5734 if (rdev->pg_flags & RADEON_PG_SUPPORT_GFX_PG) {
0116e1ef 5735 si_init_gfx_cgpg(rdev);
aa34dba8
AD
5736 } else {
5737 WREG32(RLC_SAVE_AND_RESTORE_BASE, rdev->rlc.save_restore_gpu_addr >> 8);
5738 WREG32(RLC_CLEAR_STATE_RESTORE_BASE, rdev->rlc.clear_state_gpu_addr >> 8);
0116e1ef 5739 }
ca6ebb39
AD
5740 si_enable_dma_pg(rdev, true);
5741 si_enable_gfx_cgpg(rdev, true);
f8f84ac5
AD
5742 } else {
5743 WREG32(RLC_SAVE_AND_RESTORE_BASE, rdev->rlc.save_restore_gpu_addr >> 8);
5744 WREG32(RLC_CLEAR_STATE_RESTORE_BASE, rdev->rlc.clear_state_gpu_addr >> 8);
5745 }
5746}
5747
5748static void si_fini_pg(struct radeon_device *rdev)
5749{
0116e1ef 5750 if (rdev->pg_flags) {
ca6ebb39
AD
5751 si_enable_dma_pg(rdev, false);
5752 si_enable_gfx_cgpg(rdev, false);
f8f84ac5 5753 }
347e7592
AD
5754}
5755
347e7592
AD
5756/*
5757 * RLC
5758 */
866d83de 5759void si_rlc_reset(struct radeon_device *rdev)
d719cef3 5760{
f8f84ac5 5761 u32 tmp = RREG32(GRBM_SOFT_RESET);
d719cef3 5762
f8f84ac5
AD
5763 tmp |= SOFT_RESET_RLC;
5764 WREG32(GRBM_SOFT_RESET, tmp);
5765 udelay(50);
5766 tmp &= ~SOFT_RESET_RLC;
5767 WREG32(GRBM_SOFT_RESET, tmp);
5768 udelay(50);
d719cef3
AD
5769}
5770
347e7592
AD
5771static void si_rlc_stop(struct radeon_device *rdev)
5772{
5773 WREG32(RLC_CNTL, 0);
d719cef3
AD
5774
5775 si_enable_gui_idle_interrupt(rdev, false);
5776
5777 si_wait_for_rlc_serdes(rdev);
347e7592
AD
5778}
5779
5780static void si_rlc_start(struct radeon_device *rdev)
5781{
5782 WREG32(RLC_CNTL, RLC_ENABLE);
d719cef3
AD
5783
5784 si_enable_gui_idle_interrupt(rdev, true);
5785
5786 udelay(50);
5787}
5788
5789static bool si_lbpw_supported(struct radeon_device *rdev)
5790{
5791 u32 tmp;
5792
5793 /* Enable LBPW only for DDR3 */
5794 tmp = RREG32(MC_SEQ_MISC0);
5795 if ((tmp & 0xF0000000) == 0xB0000000)
5796 return true;
5797 return false;
5798}
5799
5800static void si_enable_lbpw(struct radeon_device *rdev, bool enable)
5801{
5802 u32 tmp;
5803
5804 tmp = RREG32(RLC_LB_CNTL);
5805 if (enable)
5806 tmp |= LOAD_BALANCE_ENABLE;
5807 else
5808 tmp &= ~LOAD_BALANCE_ENABLE;
5809 WREG32(RLC_LB_CNTL, tmp);
5810
5811 if (!enable) {
5812 si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
5813 WREG32(SPI_LB_CU_MASK, 0x00ff);
5814 }
347e7592
AD
5815}
5816
5817static int si_rlc_resume(struct radeon_device *rdev)
5818{
5819 u32 i;
347e7592
AD
5820
5821 if (!rdev->rlc_fw)
5822 return -EINVAL;
5823
5824 si_rlc_stop(rdev);
5825
f8f84ac5
AD
5826 si_rlc_reset(rdev);
5827
5828 si_init_pg(rdev);
5829
5830 si_init_cg(rdev);
5831
347e7592
AD
5832 WREG32(RLC_RL_BASE, 0);
5833 WREG32(RLC_RL_SIZE, 0);
5834 WREG32(RLC_LB_CNTL, 0);
5835 WREG32(RLC_LB_CNTR_MAX, 0xffffffff);
5836 WREG32(RLC_LB_CNTR_INIT, 0);
d719cef3 5837 WREG32(RLC_LB_INIT_CU_MASK, 0xffffffff);
347e7592 5838
347e7592
AD
5839 WREG32(RLC_MC_CNTL, 0);
5840 WREG32(RLC_UCODE_CNTL, 0);
5841
629bd33c
AD
5842 if (rdev->new_fw) {
5843 const struct rlc_firmware_header_v1_0 *hdr =
5844 (const struct rlc_firmware_header_v1_0 *)rdev->rlc_fw->data;
5845 u32 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
5846 const __le32 *fw_data = (const __le32 *)
5847 (rdev->rlc_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
5848
5849 radeon_ucode_print_rlc_hdr(&hdr->header);
5850
5851 for (i = 0; i < fw_size; i++) {
5852 WREG32(RLC_UCODE_ADDR, i);
5853 WREG32(RLC_UCODE_DATA, le32_to_cpup(fw_data++));
5854 }
5855 } else {
5856 const __be32 *fw_data =
5857 (const __be32 *)rdev->rlc_fw->data;
5858 for (i = 0; i < SI_RLC_UCODE_SIZE; i++) {
5859 WREG32(RLC_UCODE_ADDR, i);
5860 WREG32(RLC_UCODE_DATA, be32_to_cpup(fw_data++));
5861 }
347e7592
AD
5862 }
5863 WREG32(RLC_UCODE_ADDR, 0);
5864
d719cef3
AD
5865 si_enable_lbpw(rdev, si_lbpw_supported(rdev));
5866
347e7592
AD
5867 si_rlc_start(rdev);
5868
5869 return 0;
5870}
5871
25a857fb
AD
5872static void si_enable_interrupts(struct radeon_device *rdev)
5873{
5874 u32 ih_cntl = RREG32(IH_CNTL);
5875 u32 ih_rb_cntl = RREG32(IH_RB_CNTL);
5876
5877 ih_cntl |= ENABLE_INTR;
5878 ih_rb_cntl |= IH_RB_ENABLE;
5879 WREG32(IH_CNTL, ih_cntl);
5880 WREG32(IH_RB_CNTL, ih_rb_cntl);
5881 rdev->ih.enabled = true;
5882}
5883
5884static void si_disable_interrupts(struct radeon_device *rdev)
5885{
5886 u32 ih_rb_cntl = RREG32(IH_RB_CNTL);
5887 u32 ih_cntl = RREG32(IH_CNTL);
5888
5889 ih_rb_cntl &= ~IH_RB_ENABLE;
5890 ih_cntl &= ~ENABLE_INTR;
5891 WREG32(IH_RB_CNTL, ih_rb_cntl);
5892 WREG32(IH_CNTL, ih_cntl);
5893 /* set rptr, wptr to 0 */
5894 WREG32(IH_RB_RPTR, 0);
5895 WREG32(IH_RB_WPTR, 0);
5896 rdev->ih.enabled = false;
25a857fb
AD
5897 rdev->ih.rptr = 0;
5898}
5899
5900static void si_disable_interrupt_state(struct radeon_device *rdev)
5901{
5902 u32 tmp;
5903
811e4d58
AD
5904 tmp = RREG32(CP_INT_CNTL_RING0) &
5905 (CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
5906 WREG32(CP_INT_CNTL_RING0, tmp);
25a857fb
AD
5907 WREG32(CP_INT_CNTL_RING1, 0);
5908 WREG32(CP_INT_CNTL_RING2, 0);
8c5fd7ef
AD
5909 tmp = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET) & ~TRAP_ENABLE;
5910 WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, tmp);
5911 tmp = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET) & ~TRAP_ENABLE;
5912 WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, tmp);
25a857fb 5913 WREG32(GRBM_INT_CNTL, 0);
5153550a
AD
5914 if (rdev->num_crtc >= 2) {
5915 WREG32(INT_MASK + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
5916 WREG32(INT_MASK + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
5917 }
25a857fb
AD
5918 if (rdev->num_crtc >= 4) {
5919 WREG32(INT_MASK + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
5920 WREG32(INT_MASK + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
5921 }
5922 if (rdev->num_crtc >= 6) {
5923 WREG32(INT_MASK + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
5924 WREG32(INT_MASK + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
5925 }
5926
5153550a
AD
5927 if (rdev->num_crtc >= 2) {
5928 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
5929 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
5930 }
25a857fb
AD
5931 if (rdev->num_crtc >= 4) {
5932 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
5933 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
5934 }
5935 if (rdev->num_crtc >= 6) {
5936 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
5937 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
5938 }
5939
5153550a 5940 if (!ASIC_IS_NODCE(rdev)) {
e9a321c6 5941 WREG32(DAC_AUTODETECT_INT_CONTROL, 0);
5153550a
AD
5942
5943 tmp = RREG32(DC_HPD1_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5944 WREG32(DC_HPD1_INT_CONTROL, tmp);
5945 tmp = RREG32(DC_HPD2_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5946 WREG32(DC_HPD2_INT_CONTROL, tmp);
5947 tmp = RREG32(DC_HPD3_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5948 WREG32(DC_HPD3_INT_CONTROL, tmp);
5949 tmp = RREG32(DC_HPD4_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5950 WREG32(DC_HPD4_INT_CONTROL, tmp);
5951 tmp = RREG32(DC_HPD5_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5952 WREG32(DC_HPD5_INT_CONTROL, tmp);
5953 tmp = RREG32(DC_HPD6_INT_CONTROL) & DC_HPDx_INT_POLARITY;
5954 WREG32(DC_HPD6_INT_CONTROL, tmp);
5955 }
25a857fb
AD
5956}
5957
5958static int si_irq_init(struct radeon_device *rdev)
5959{
5960 int ret = 0;
5961 int rb_bufsz;
5962 u32 interrupt_cntl, ih_cntl, ih_rb_cntl;
5963
5964 /* allocate ring */
5965 ret = r600_ih_ring_alloc(rdev);
5966 if (ret)
5967 return ret;
5968
5969 /* disable irqs */
5970 si_disable_interrupts(rdev);
5971
5972 /* init rlc */
5973 ret = si_rlc_resume(rdev);
5974 if (ret) {
5975 r600_ih_ring_fini(rdev);
5976 return ret;
5977 }
5978
5979 /* setup interrupt control */
5980 /* set dummy read address to ring address */
5981 WREG32(INTERRUPT_CNTL2, rdev->ih.gpu_addr >> 8);
5982 interrupt_cntl = RREG32(INTERRUPT_CNTL);
5983 /* IH_DUMMY_RD_OVERRIDE=0 - dummy read disabled with msi, enabled without msi
5984 * IH_DUMMY_RD_OVERRIDE=1 - dummy read controlled by IH_DUMMY_RD_EN
5985 */
5986 interrupt_cntl &= ~IH_DUMMY_RD_OVERRIDE;
5987 /* IH_REQ_NONSNOOP_EN=1 if ring is in non-cacheable memory, e.g., vram */
5988 interrupt_cntl &= ~IH_REQ_NONSNOOP_EN;
5989 WREG32(INTERRUPT_CNTL, interrupt_cntl);
5990
5991 WREG32(IH_RB_BASE, rdev->ih.gpu_addr >> 8);
b72a8925 5992 rb_bufsz = order_base_2(rdev->ih.ring_size / 4);
25a857fb
AD
5993
5994 ih_rb_cntl = (IH_WPTR_OVERFLOW_ENABLE |
5995 IH_WPTR_OVERFLOW_CLEAR |
5996 (rb_bufsz << 1));
5997
5998 if (rdev->wb.enabled)
5999 ih_rb_cntl |= IH_WPTR_WRITEBACK_ENABLE;
6000
6001 /* set the writeback address whether it's enabled or not */
6002 WREG32(IH_RB_WPTR_ADDR_LO, (rdev->wb.gpu_addr + R600_WB_IH_WPTR_OFFSET) & 0xFFFFFFFC);
6003 WREG32(IH_RB_WPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + R600_WB_IH_WPTR_OFFSET) & 0xFF);
6004
6005 WREG32(IH_RB_CNTL, ih_rb_cntl);
6006
6007 /* set rptr, wptr to 0 */
6008 WREG32(IH_RB_RPTR, 0);
6009 WREG32(IH_RB_WPTR, 0);
6010
6011 /* Default settings for IH_CNTL (disabled at first) */
6012 ih_cntl = MC_WRREQ_CREDIT(0x10) | MC_WR_CLEAN_CNT(0x10) | MC_VMID(0);
6013 /* RPTR_REARM only works if msi's are enabled */
6014 if (rdev->msi_enabled)
6015 ih_cntl |= RPTR_REARM;
6016 WREG32(IH_CNTL, ih_cntl);
6017
6018 /* force the active interrupt state to all disabled */
6019 si_disable_interrupt_state(rdev);
6020
2099810f
DA
6021 pci_set_master(rdev->pdev);
6022
25a857fb
AD
6023 /* enable irqs */
6024 si_enable_interrupts(rdev);
6025
6026 return ret;
6027}
6028
6029int si_irq_set(struct radeon_device *rdev)
6030{
811e4d58 6031 u32 cp_int_cntl;
25a857fb
AD
6032 u32 cp_int_cntl1 = 0, cp_int_cntl2 = 0;
6033 u32 crtc1 = 0, crtc2 = 0, crtc3 = 0, crtc4 = 0, crtc5 = 0, crtc6 = 0;
5153550a 6034 u32 hpd1 = 0, hpd2 = 0, hpd3 = 0, hpd4 = 0, hpd5 = 0, hpd6 = 0;
25a857fb 6035 u32 grbm_int_cntl = 0;
8c5fd7ef 6036 u32 dma_cntl, dma_cntl1;
a9e61410 6037 u32 thermal_int = 0;
25a857fb
AD
6038
6039 if (!rdev->irq.installed) {
6040 WARN(1, "Can't enable IRQ/MSI because no handler is installed\n");
6041 return -EINVAL;
6042 }
6043 /* don't enable anything if the ih is disabled */
6044 if (!rdev->ih.enabled) {
6045 si_disable_interrupts(rdev);
6046 /* force the active interrupt state to all disabled */
6047 si_disable_interrupt_state(rdev);
6048 return 0;
6049 }
6050
811e4d58
AD
6051 cp_int_cntl = RREG32(CP_INT_CNTL_RING0) &
6052 (CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
6053
5153550a
AD
6054 if (!ASIC_IS_NODCE(rdev)) {
6055 hpd1 = RREG32(DC_HPD1_INT_CONTROL) & ~DC_HPDx_INT_EN;
6056 hpd2 = RREG32(DC_HPD2_INT_CONTROL) & ~DC_HPDx_INT_EN;
6057 hpd3 = RREG32(DC_HPD3_INT_CONTROL) & ~DC_HPDx_INT_EN;
6058 hpd4 = RREG32(DC_HPD4_INT_CONTROL) & ~DC_HPDx_INT_EN;
6059 hpd5 = RREG32(DC_HPD5_INT_CONTROL) & ~DC_HPDx_INT_EN;
6060 hpd6 = RREG32(DC_HPD6_INT_CONTROL) & ~DC_HPDx_INT_EN;
6061 }
25a857fb 6062
8c5fd7ef
AD
6063 dma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET) & ~TRAP_ENABLE;
6064 dma_cntl1 = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET) & ~TRAP_ENABLE;
6065
a9e61410
AD
6066 thermal_int = RREG32(CG_THERMAL_INT) &
6067 ~(THERM_INT_MASK_HIGH | THERM_INT_MASK_LOW);
6068
25a857fb 6069 /* enable CP interrupts on all rings */
736fc37f 6070 if (atomic_read(&rdev->irq.ring_int[RADEON_RING_TYPE_GFX_INDEX])) {
25a857fb
AD
6071 DRM_DEBUG("si_irq_set: sw int gfx\n");
6072 cp_int_cntl |= TIME_STAMP_INT_ENABLE;
6073 }
736fc37f 6074 if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_CP1_INDEX])) {
25a857fb
AD
6075 DRM_DEBUG("si_irq_set: sw int cp1\n");
6076 cp_int_cntl1 |= TIME_STAMP_INT_ENABLE;
6077 }
736fc37f 6078 if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_CP2_INDEX])) {
25a857fb
AD
6079 DRM_DEBUG("si_irq_set: sw int cp2\n");
6080 cp_int_cntl2 |= TIME_STAMP_INT_ENABLE;
6081 }
8c5fd7ef
AD
6082 if (atomic_read(&rdev->irq.ring_int[R600_RING_TYPE_DMA_INDEX])) {
6083 DRM_DEBUG("si_irq_set: sw int dma\n");
6084 dma_cntl |= TRAP_ENABLE;
6085 }
6086
6087 if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_DMA1_INDEX])) {
6088 DRM_DEBUG("si_irq_set: sw int dma1\n");
6089 dma_cntl1 |= TRAP_ENABLE;
6090 }
25a857fb 6091 if (rdev->irq.crtc_vblank_int[0] ||
736fc37f 6092 atomic_read(&rdev->irq.pflip[0])) {
25a857fb
AD
6093 DRM_DEBUG("si_irq_set: vblank 0\n");
6094 crtc1 |= VBLANK_INT_MASK;
6095 }
6096 if (rdev->irq.crtc_vblank_int[1] ||
736fc37f 6097 atomic_read(&rdev->irq.pflip[1])) {
25a857fb
AD
6098 DRM_DEBUG("si_irq_set: vblank 1\n");
6099 crtc2 |= VBLANK_INT_MASK;
6100 }
6101 if (rdev->irq.crtc_vblank_int[2] ||
736fc37f 6102 atomic_read(&rdev->irq.pflip[2])) {
25a857fb
AD
6103 DRM_DEBUG("si_irq_set: vblank 2\n");
6104 crtc3 |= VBLANK_INT_MASK;
6105 }
6106 if (rdev->irq.crtc_vblank_int[3] ||
736fc37f 6107 atomic_read(&rdev->irq.pflip[3])) {
25a857fb
AD
6108 DRM_DEBUG("si_irq_set: vblank 3\n");
6109 crtc4 |= VBLANK_INT_MASK;
6110 }
6111 if (rdev->irq.crtc_vblank_int[4] ||
736fc37f 6112 atomic_read(&rdev->irq.pflip[4])) {
25a857fb
AD
6113 DRM_DEBUG("si_irq_set: vblank 4\n");
6114 crtc5 |= VBLANK_INT_MASK;
6115 }
6116 if (rdev->irq.crtc_vblank_int[5] ||
736fc37f 6117 atomic_read(&rdev->irq.pflip[5])) {
25a857fb
AD
6118 DRM_DEBUG("si_irq_set: vblank 5\n");
6119 crtc6 |= VBLANK_INT_MASK;
6120 }
6121 if (rdev->irq.hpd[0]) {
6122 DRM_DEBUG("si_irq_set: hpd 1\n");
6123 hpd1 |= DC_HPDx_INT_EN;
6124 }
6125 if (rdev->irq.hpd[1]) {
6126 DRM_DEBUG("si_irq_set: hpd 2\n");
6127 hpd2 |= DC_HPDx_INT_EN;
6128 }
6129 if (rdev->irq.hpd[2]) {
6130 DRM_DEBUG("si_irq_set: hpd 3\n");
6131 hpd3 |= DC_HPDx_INT_EN;
6132 }
6133 if (rdev->irq.hpd[3]) {
6134 DRM_DEBUG("si_irq_set: hpd 4\n");
6135 hpd4 |= DC_HPDx_INT_EN;
6136 }
6137 if (rdev->irq.hpd[4]) {
6138 DRM_DEBUG("si_irq_set: hpd 5\n");
6139 hpd5 |= DC_HPDx_INT_EN;
6140 }
6141 if (rdev->irq.hpd[5]) {
6142 DRM_DEBUG("si_irq_set: hpd 6\n");
6143 hpd6 |= DC_HPDx_INT_EN;
6144 }
25a857fb
AD
6145
6146 WREG32(CP_INT_CNTL_RING0, cp_int_cntl);
6147 WREG32(CP_INT_CNTL_RING1, cp_int_cntl1);
6148 WREG32(CP_INT_CNTL_RING2, cp_int_cntl2);
6149
8c5fd7ef
AD
6150 WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, dma_cntl);
6151 WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, dma_cntl1);
6152
25a857fb
AD
6153 WREG32(GRBM_INT_CNTL, grbm_int_cntl);
6154
a9e61410
AD
6155 if (rdev->irq.dpm_thermal) {
6156 DRM_DEBUG("dpm thermal\n");
6157 thermal_int |= THERM_INT_MASK_HIGH | THERM_INT_MASK_LOW;
6158 }
6159
5153550a
AD
6160 if (rdev->num_crtc >= 2) {
6161 WREG32(INT_MASK + EVERGREEN_CRTC0_REGISTER_OFFSET, crtc1);
6162 WREG32(INT_MASK + EVERGREEN_CRTC1_REGISTER_OFFSET, crtc2);
6163 }
25a857fb
AD
6164 if (rdev->num_crtc >= 4) {
6165 WREG32(INT_MASK + EVERGREEN_CRTC2_REGISTER_OFFSET, crtc3);
6166 WREG32(INT_MASK + EVERGREEN_CRTC3_REGISTER_OFFSET, crtc4);
6167 }
6168 if (rdev->num_crtc >= 6) {
6169 WREG32(INT_MASK + EVERGREEN_CRTC4_REGISTER_OFFSET, crtc5);
6170 WREG32(INT_MASK + EVERGREEN_CRTC5_REGISTER_OFFSET, crtc6);
6171 }
6172
5153550a 6173 if (rdev->num_crtc >= 2) {
f5d636d2
CK
6174 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET,
6175 GRPH_PFLIP_INT_MASK);
6176 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET,
6177 GRPH_PFLIP_INT_MASK);
5153550a 6178 }
25a857fb 6179 if (rdev->num_crtc >= 4) {
f5d636d2
CK
6180 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET,
6181 GRPH_PFLIP_INT_MASK);
6182 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET,
6183 GRPH_PFLIP_INT_MASK);
25a857fb
AD
6184 }
6185 if (rdev->num_crtc >= 6) {
f5d636d2
CK
6186 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET,
6187 GRPH_PFLIP_INT_MASK);
6188 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET,
6189 GRPH_PFLIP_INT_MASK);
25a857fb
AD
6190 }
6191
5153550a
AD
6192 if (!ASIC_IS_NODCE(rdev)) {
6193 WREG32(DC_HPD1_INT_CONTROL, hpd1);
6194 WREG32(DC_HPD2_INT_CONTROL, hpd2);
6195 WREG32(DC_HPD3_INT_CONTROL, hpd3);
6196 WREG32(DC_HPD4_INT_CONTROL, hpd4);
6197 WREG32(DC_HPD5_INT_CONTROL, hpd5);
6198 WREG32(DC_HPD6_INT_CONTROL, hpd6);
6199 }
25a857fb 6200
a9e61410
AD
6201 WREG32(CG_THERMAL_INT, thermal_int);
6202
25a857fb
AD
6203 return 0;
6204}
6205
6206static inline void si_irq_ack(struct radeon_device *rdev)
6207{
6208 u32 tmp;
6209
5153550a
AD
6210 if (ASIC_IS_NODCE(rdev))
6211 return;
6212
25a857fb
AD
6213 rdev->irq.stat_regs.evergreen.disp_int = RREG32(DISP_INTERRUPT_STATUS);
6214 rdev->irq.stat_regs.evergreen.disp_int_cont = RREG32(DISP_INTERRUPT_STATUS_CONTINUE);
6215 rdev->irq.stat_regs.evergreen.disp_int_cont2 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE2);
6216 rdev->irq.stat_regs.evergreen.disp_int_cont3 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE3);
6217 rdev->irq.stat_regs.evergreen.disp_int_cont4 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE4);
6218 rdev->irq.stat_regs.evergreen.disp_int_cont5 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE5);
6219 rdev->irq.stat_regs.evergreen.d1grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET);
6220 rdev->irq.stat_regs.evergreen.d2grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET);
6221 if (rdev->num_crtc >= 4) {
6222 rdev->irq.stat_regs.evergreen.d3grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET);
6223 rdev->irq.stat_regs.evergreen.d4grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET);
6224 }
6225 if (rdev->num_crtc >= 6) {
6226 rdev->irq.stat_regs.evergreen.d5grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET);
6227 rdev->irq.stat_regs.evergreen.d6grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET);
6228 }
6229
6230 if (rdev->irq.stat_regs.evergreen.d1grph_int & GRPH_PFLIP_INT_OCCURRED)
6231 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
6232 if (rdev->irq.stat_regs.evergreen.d2grph_int & GRPH_PFLIP_INT_OCCURRED)
6233 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
6234 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VBLANK_INTERRUPT)
6235 WREG32(VBLANK_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, VBLANK_ACK);
6236 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VLINE_INTERRUPT)
6237 WREG32(VLINE_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, VLINE_ACK);
6238 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VBLANK_INTERRUPT)
6239 WREG32(VBLANK_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, VBLANK_ACK);
6240 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VLINE_INTERRUPT)
6241 WREG32(VLINE_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, VLINE_ACK);
6242
6243 if (rdev->num_crtc >= 4) {
6244 if (rdev->irq.stat_regs.evergreen.d3grph_int & GRPH_PFLIP_INT_OCCURRED)
6245 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
6246 if (rdev->irq.stat_regs.evergreen.d4grph_int & GRPH_PFLIP_INT_OCCURRED)
6247 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
6248 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VBLANK_INTERRUPT)
6249 WREG32(VBLANK_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, VBLANK_ACK);
6250 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VLINE_INTERRUPT)
6251 WREG32(VLINE_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, VLINE_ACK);
6252 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VBLANK_INTERRUPT)
6253 WREG32(VBLANK_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, VBLANK_ACK);
6254 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VLINE_INTERRUPT)
6255 WREG32(VLINE_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, VLINE_ACK);
6256 }
6257
6258 if (rdev->num_crtc >= 6) {
6259 if (rdev->irq.stat_regs.evergreen.d5grph_int & GRPH_PFLIP_INT_OCCURRED)
6260 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
6261 if (rdev->irq.stat_regs.evergreen.d6grph_int & GRPH_PFLIP_INT_OCCURRED)
6262 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
6263 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VBLANK_INTERRUPT)
6264 WREG32(VBLANK_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, VBLANK_ACK);
6265 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VLINE_INTERRUPT)
6266 WREG32(VLINE_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, VLINE_ACK);
6267 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VBLANK_INTERRUPT)
6268 WREG32(VBLANK_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, VBLANK_ACK);
6269 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VLINE_INTERRUPT)
6270 WREG32(VLINE_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, VLINE_ACK);
6271 }
6272
6273 if (rdev->irq.stat_regs.evergreen.disp_int & DC_HPD1_INTERRUPT) {
6274 tmp = RREG32(DC_HPD1_INT_CONTROL);
6275 tmp |= DC_HPDx_INT_ACK;
6276 WREG32(DC_HPD1_INT_CONTROL, tmp);
6277 }
6278 if (rdev->irq.stat_regs.evergreen.disp_int_cont & DC_HPD2_INTERRUPT) {
6279 tmp = RREG32(DC_HPD2_INT_CONTROL);
6280 tmp |= DC_HPDx_INT_ACK;
6281 WREG32(DC_HPD2_INT_CONTROL, tmp);
6282 }
6283 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & DC_HPD3_INTERRUPT) {
6284 tmp = RREG32(DC_HPD3_INT_CONTROL);
6285 tmp |= DC_HPDx_INT_ACK;
6286 WREG32(DC_HPD3_INT_CONTROL, tmp);
6287 }
6288 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & DC_HPD4_INTERRUPT) {
6289 tmp = RREG32(DC_HPD4_INT_CONTROL);
6290 tmp |= DC_HPDx_INT_ACK;
6291 WREG32(DC_HPD4_INT_CONTROL, tmp);
6292 }
6293 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & DC_HPD5_INTERRUPT) {
6294 tmp = RREG32(DC_HPD5_INT_CONTROL);
6295 tmp |= DC_HPDx_INT_ACK;
6296 WREG32(DC_HPD5_INT_CONTROL, tmp);
6297 }
6298 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & DC_HPD6_INTERRUPT) {
6299 tmp = RREG32(DC_HPD5_INT_CONTROL);
6300 tmp |= DC_HPDx_INT_ACK;
6301 WREG32(DC_HPD6_INT_CONTROL, tmp);
6302 }
6303}
6304
6305static void si_irq_disable(struct radeon_device *rdev)
6306{
6307 si_disable_interrupts(rdev);
6308 /* Wait and acknowledge irq */
6309 mdelay(1);
6310 si_irq_ack(rdev);
6311 si_disable_interrupt_state(rdev);
6312}
6313
6314static void si_irq_suspend(struct radeon_device *rdev)
6315{
6316 si_irq_disable(rdev);
6317 si_rlc_stop(rdev);
6318}
6319
9b136d51
AD
6320static void si_irq_fini(struct radeon_device *rdev)
6321{
6322 si_irq_suspend(rdev);
6323 r600_ih_ring_fini(rdev);
6324}
6325
25a857fb
AD
6326static inline u32 si_get_ih_wptr(struct radeon_device *rdev)
6327{
6328 u32 wptr, tmp;
6329
6330 if (rdev->wb.enabled)
6331 wptr = le32_to_cpu(rdev->wb.wb[R600_WB_IH_WPTR_OFFSET/4]);
6332 else
6333 wptr = RREG32(IH_RB_WPTR);
6334
6335 if (wptr & RB_OVERFLOW) {
11bab0ae 6336 wptr &= ~RB_OVERFLOW;
25a857fb
AD
6337 /* When a ring buffer overflow happen start parsing interrupt
6338 * from the last not overwritten vector (wptr + 16). Hopefully
6339 * this should allow us to catchup.
6340 */
6cc2fda2
MD
6341 dev_warn(rdev->dev, "IH ring buffer overflow (0x%08X, 0x%08X, 0x%08X)\n",
6342 wptr, rdev->ih.rptr, (wptr + 16) & rdev->ih.ptr_mask);
25a857fb
AD
6343 rdev->ih.rptr = (wptr + 16) & rdev->ih.ptr_mask;
6344 tmp = RREG32(IH_RB_CNTL);
6345 tmp |= IH_WPTR_OVERFLOW_CLEAR;
6346 WREG32(IH_RB_CNTL, tmp);
6347 }
6348 return (wptr & rdev->ih.ptr_mask);
6349}
6350
6351/* SI IV Ring
6352 * Each IV ring entry is 128 bits:
6353 * [7:0] - interrupt source id
6354 * [31:8] - reserved
6355 * [59:32] - interrupt source data
6356 * [63:60] - reserved
6357 * [71:64] - RINGID
6358 * [79:72] - VMID
6359 * [127:80] - reserved
6360 */
6361int si_irq_process(struct radeon_device *rdev)
6362{
6363 u32 wptr;
6364 u32 rptr;
6365 u32 src_id, src_data, ring_id;
6366 u32 ring_index;
25a857fb 6367 bool queue_hotplug = false;
a9e61410 6368 bool queue_thermal = false;
fbf6dc7a 6369 u32 status, addr;
25a857fb
AD
6370
6371 if (!rdev->ih.enabled || rdev->shutdown)
6372 return IRQ_NONE;
6373
6374 wptr = si_get_ih_wptr(rdev);
c20dc369
CK
6375
6376restart_ih:
6377 /* is somebody else already processing irqs? */
6378 if (atomic_xchg(&rdev->ih.lock, 1))
6379 return IRQ_NONE;
6380
25a857fb
AD
6381 rptr = rdev->ih.rptr;
6382 DRM_DEBUG("si_irq_process start: rptr %d, wptr %d\n", rptr, wptr);
6383
25a857fb
AD
6384 /* Order reading of wptr vs. reading of IH ring data */
6385 rmb();
6386
6387 /* display interrupts */
6388 si_irq_ack(rdev);
6389
25a857fb
AD
6390 while (rptr != wptr) {
6391 /* wptr/rptr are in bytes! */
6392 ring_index = rptr / 4;
6393 src_id = le32_to_cpu(rdev->ih.ring[ring_index]) & 0xff;
6394 src_data = le32_to_cpu(rdev->ih.ring[ring_index + 1]) & 0xfffffff;
6395 ring_id = le32_to_cpu(rdev->ih.ring[ring_index + 2]) & 0xff;
6396
6397 switch (src_id) {
6398 case 1: /* D1 vblank/vline */
6399 switch (src_data) {
6400 case 0: /* D1 vblank */
6401 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VBLANK_INTERRUPT) {
6402 if (rdev->irq.crtc_vblank_int[0]) {
6403 drm_handle_vblank(rdev->ddev, 0);
6404 rdev->pm.vblank_sync = true;
6405 wake_up(&rdev->irq.vblank_queue);
6406 }
736fc37f 6407 if (atomic_read(&rdev->irq.pflip[0]))
1a0e7918 6408 radeon_crtc_handle_vblank(rdev, 0);
25a857fb
AD
6409 rdev->irq.stat_regs.evergreen.disp_int &= ~LB_D1_VBLANK_INTERRUPT;
6410 DRM_DEBUG("IH: D1 vblank\n");
6411 }
6412 break;
6413 case 1: /* D1 vline */
6414 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VLINE_INTERRUPT) {
6415 rdev->irq.stat_regs.evergreen.disp_int &= ~LB_D1_VLINE_INTERRUPT;
6416 DRM_DEBUG("IH: D1 vline\n");
6417 }
6418 break;
6419 default:
6420 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6421 break;
6422 }
6423 break;
6424 case 2: /* D2 vblank/vline */
6425 switch (src_data) {
6426 case 0: /* D2 vblank */
6427 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VBLANK_INTERRUPT) {
6428 if (rdev->irq.crtc_vblank_int[1]) {
6429 drm_handle_vblank(rdev->ddev, 1);
6430 rdev->pm.vblank_sync = true;
6431 wake_up(&rdev->irq.vblank_queue);
6432 }
736fc37f 6433 if (atomic_read(&rdev->irq.pflip[1]))
1a0e7918 6434 radeon_crtc_handle_vblank(rdev, 1);
25a857fb
AD
6435 rdev->irq.stat_regs.evergreen.disp_int_cont &= ~LB_D2_VBLANK_INTERRUPT;
6436 DRM_DEBUG("IH: D2 vblank\n");
6437 }
6438 break;
6439 case 1: /* D2 vline */
6440 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VLINE_INTERRUPT) {
6441 rdev->irq.stat_regs.evergreen.disp_int_cont &= ~LB_D2_VLINE_INTERRUPT;
6442 DRM_DEBUG("IH: D2 vline\n");
6443 }
6444 break;
6445 default:
6446 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6447 break;
6448 }
6449 break;
6450 case 3: /* D3 vblank/vline */
6451 switch (src_data) {
6452 case 0: /* D3 vblank */
6453 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VBLANK_INTERRUPT) {
6454 if (rdev->irq.crtc_vblank_int[2]) {
6455 drm_handle_vblank(rdev->ddev, 2);
6456 rdev->pm.vblank_sync = true;
6457 wake_up(&rdev->irq.vblank_queue);
6458 }
736fc37f 6459 if (atomic_read(&rdev->irq.pflip[2]))
1a0e7918 6460 radeon_crtc_handle_vblank(rdev, 2);
25a857fb
AD
6461 rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~LB_D3_VBLANK_INTERRUPT;
6462 DRM_DEBUG("IH: D3 vblank\n");
6463 }
6464 break;
6465 case 1: /* D3 vline */
6466 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VLINE_INTERRUPT) {
6467 rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~LB_D3_VLINE_INTERRUPT;
6468 DRM_DEBUG("IH: D3 vline\n");
6469 }
6470 break;
6471 default:
6472 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6473 break;
6474 }
6475 break;
6476 case 4: /* D4 vblank/vline */
6477 switch (src_data) {
6478 case 0: /* D4 vblank */
6479 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VBLANK_INTERRUPT) {
6480 if (rdev->irq.crtc_vblank_int[3]) {
6481 drm_handle_vblank(rdev->ddev, 3);
6482 rdev->pm.vblank_sync = true;
6483 wake_up(&rdev->irq.vblank_queue);
6484 }
736fc37f 6485 if (atomic_read(&rdev->irq.pflip[3]))
1a0e7918 6486 radeon_crtc_handle_vblank(rdev, 3);
25a857fb
AD
6487 rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~LB_D4_VBLANK_INTERRUPT;
6488 DRM_DEBUG("IH: D4 vblank\n");
6489 }
6490 break;
6491 case 1: /* D4 vline */
6492 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VLINE_INTERRUPT) {
6493 rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~LB_D4_VLINE_INTERRUPT;
6494 DRM_DEBUG("IH: D4 vline\n");
6495 }
6496 break;
6497 default:
6498 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6499 break;
6500 }
6501 break;
6502 case 5: /* D5 vblank/vline */
6503 switch (src_data) {
6504 case 0: /* D5 vblank */
6505 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VBLANK_INTERRUPT) {
6506 if (rdev->irq.crtc_vblank_int[4]) {
6507 drm_handle_vblank(rdev->ddev, 4);
6508 rdev->pm.vblank_sync = true;
6509 wake_up(&rdev->irq.vblank_queue);
6510 }
736fc37f 6511 if (atomic_read(&rdev->irq.pflip[4]))
1a0e7918 6512 radeon_crtc_handle_vblank(rdev, 4);
25a857fb
AD
6513 rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~LB_D5_VBLANK_INTERRUPT;
6514 DRM_DEBUG("IH: D5 vblank\n");
6515 }
6516 break;
6517 case 1: /* D5 vline */
6518 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VLINE_INTERRUPT) {
6519 rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~LB_D5_VLINE_INTERRUPT;
6520 DRM_DEBUG("IH: D5 vline\n");
6521 }
6522 break;
6523 default:
6524 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6525 break;
6526 }
6527 break;
6528 case 6: /* D6 vblank/vline */
6529 switch (src_data) {
6530 case 0: /* D6 vblank */
6531 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VBLANK_INTERRUPT) {
6532 if (rdev->irq.crtc_vblank_int[5]) {
6533 drm_handle_vblank(rdev->ddev, 5);
6534 rdev->pm.vblank_sync = true;
6535 wake_up(&rdev->irq.vblank_queue);
6536 }
736fc37f 6537 if (atomic_read(&rdev->irq.pflip[5]))
1a0e7918 6538 radeon_crtc_handle_vblank(rdev, 5);
25a857fb
AD
6539 rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~LB_D6_VBLANK_INTERRUPT;
6540 DRM_DEBUG("IH: D6 vblank\n");
6541 }
6542 break;
6543 case 1: /* D6 vline */
6544 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VLINE_INTERRUPT) {
6545 rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~LB_D6_VLINE_INTERRUPT;
6546 DRM_DEBUG("IH: D6 vline\n");
6547 }
6548 break;
6549 default:
6550 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6551 break;
6552 }
6553 break;
f5d636d2
CK
6554 case 8: /* D1 page flip */
6555 case 10: /* D2 page flip */
6556 case 12: /* D3 page flip */
6557 case 14: /* D4 page flip */
6558 case 16: /* D5 page flip */
6559 case 18: /* D6 page flip */
6560 DRM_DEBUG("IH: D%d flip\n", ((src_id - 8) >> 1) + 1);
39dc5454
MK
6561 if (radeon_use_pflipirq > 0)
6562 radeon_crtc_handle_flip(rdev, (src_id - 8) >> 1);
f5d636d2 6563 break;
25a857fb
AD
6564 case 42: /* HPD hotplug */
6565 switch (src_data) {
6566 case 0:
6567 if (rdev->irq.stat_regs.evergreen.disp_int & DC_HPD1_INTERRUPT) {
6568 rdev->irq.stat_regs.evergreen.disp_int &= ~DC_HPD1_INTERRUPT;
6569 queue_hotplug = true;
6570 DRM_DEBUG("IH: HPD1\n");
6571 }
6572 break;
6573 case 1:
6574 if (rdev->irq.stat_regs.evergreen.disp_int_cont & DC_HPD2_INTERRUPT) {
6575 rdev->irq.stat_regs.evergreen.disp_int_cont &= ~DC_HPD2_INTERRUPT;
6576 queue_hotplug = true;
6577 DRM_DEBUG("IH: HPD2\n");
6578 }
6579 break;
6580 case 2:
6581 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & DC_HPD3_INTERRUPT) {
6582 rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~DC_HPD3_INTERRUPT;
6583 queue_hotplug = true;
6584 DRM_DEBUG("IH: HPD3\n");
6585 }
6586 break;
6587 case 3:
6588 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & DC_HPD4_INTERRUPT) {
6589 rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~DC_HPD4_INTERRUPT;
6590 queue_hotplug = true;
6591 DRM_DEBUG("IH: HPD4\n");
6592 }
6593 break;
6594 case 4:
6595 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & DC_HPD5_INTERRUPT) {
6596 rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~DC_HPD5_INTERRUPT;
6597 queue_hotplug = true;
6598 DRM_DEBUG("IH: HPD5\n");
6599 }
6600 break;
6601 case 5:
6602 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & DC_HPD6_INTERRUPT) {
6603 rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~DC_HPD6_INTERRUPT;
6604 queue_hotplug = true;
6605 DRM_DEBUG("IH: HPD6\n");
6606 }
6607 break;
6608 default:
6609 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6610 break;
6611 }
6612 break;
b927e1c2
CK
6613 case 124: /* UVD */
6614 DRM_DEBUG("IH: UVD int: 0x%08x\n", src_data);
6615 radeon_fence_process(rdev, R600_RING_TYPE_UVD_INDEX);
6616 break;
ae133a11
CK
6617 case 146:
6618 case 147:
fbf6dc7a
AD
6619 addr = RREG32(VM_CONTEXT1_PROTECTION_FAULT_ADDR);
6620 status = RREG32(VM_CONTEXT1_PROTECTION_FAULT_STATUS);
9b7d786b
CK
6621 /* reset addr and status */
6622 WREG32_P(VM_CONTEXT1_CNTL2, 1, ~1);
6623 if (addr == 0x0 && status == 0x0)
6624 break;
ae133a11
CK
6625 dev_err(rdev->dev, "GPU fault detected: %d 0x%08x\n", src_id, src_data);
6626 dev_err(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n",
fbf6dc7a 6627 addr);
ae133a11 6628 dev_err(rdev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
fbf6dc7a
AD
6629 status);
6630 si_vm_decode_fault(rdev, status, addr);
ae133a11 6631 break;
25a857fb
AD
6632 case 176: /* RINGID0 CP_INT */
6633 radeon_fence_process(rdev, RADEON_RING_TYPE_GFX_INDEX);
6634 break;
6635 case 177: /* RINGID1 CP_INT */
6636 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
6637 break;
6638 case 178: /* RINGID2 CP_INT */
6639 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
6640 break;
6641 case 181: /* CP EOP event */
6642 DRM_DEBUG("IH: CP EOP\n");
6643 switch (ring_id) {
6644 case 0:
6645 radeon_fence_process(rdev, RADEON_RING_TYPE_GFX_INDEX);
6646 break;
6647 case 1:
6648 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
6649 break;
6650 case 2:
6651 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
6652 break;
6653 }
6654 break;
8c5fd7ef
AD
6655 case 224: /* DMA trap event */
6656 DRM_DEBUG("IH: DMA trap\n");
6657 radeon_fence_process(rdev, R600_RING_TYPE_DMA_INDEX);
6658 break;
a9e61410
AD
6659 case 230: /* thermal low to high */
6660 DRM_DEBUG("IH: thermal low to high\n");
6661 rdev->pm.dpm.thermal.high_to_low = false;
6662 queue_thermal = true;
6663 break;
6664 case 231: /* thermal high to low */
6665 DRM_DEBUG("IH: thermal high to low\n");
6666 rdev->pm.dpm.thermal.high_to_low = true;
6667 queue_thermal = true;
6668 break;
25a857fb
AD
6669 case 233: /* GUI IDLE */
6670 DRM_DEBUG("IH: GUI idle\n");
25a857fb 6671 break;
8c5fd7ef
AD
6672 case 244: /* DMA trap event */
6673 DRM_DEBUG("IH: DMA1 trap\n");
6674 radeon_fence_process(rdev, CAYMAN_RING_TYPE_DMA1_INDEX);
6675 break;
25a857fb
AD
6676 default:
6677 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
6678 break;
6679 }
6680
6681 /* wptr/rptr are in bytes! */
6682 rptr += 16;
6683 rptr &= rdev->ih.ptr_mask;
f55e03b9 6684 WREG32(IH_RB_RPTR, rptr);
25a857fb 6685 }
25a857fb
AD
6686 if (queue_hotplug)
6687 schedule_work(&rdev->hotplug_work);
a9e61410
AD
6688 if (queue_thermal && rdev->pm.dpm_enabled)
6689 schedule_work(&rdev->pm.dpm.thermal.work);
25a857fb 6690 rdev->ih.rptr = rptr;
c20dc369
CK
6691 atomic_set(&rdev->ih.lock, 0);
6692
6693 /* make sure wptr hasn't changed while processing */
6694 wptr = si_get_ih_wptr(rdev);
6695 if (wptr != rptr)
6696 goto restart_ih;
6697
25a857fb
AD
6698 return IRQ_HANDLED;
6699}
6700
9b136d51
AD
6701/*
6702 * startup/shutdown callbacks
6703 */
6704static int si_startup(struct radeon_device *rdev)
6705{
6706 struct radeon_ring *ring;
6707 int r;
6708
b9d305df
AD
6709 /* enable pcie gen2/3 link */
6710 si_pcie_gen3_enable(rdev);
e0bcf165
AD
6711 /* enable aspm */
6712 si_program_aspm(rdev);
b9d305df 6713
e5903d39
AD
6714 /* scratch needs to be initialized before MC */
6715 r = r600_vram_scratch_init(rdev);
6716 if (r)
6717 return r;
6718
6fab3feb
AD
6719 si_mc_program(rdev);
6720
6c7bccea
AD
6721 if (!rdev->pm.dpm_enabled) {
6722 r = si_mc_load_microcode(rdev);
6723 if (r) {
6724 DRM_ERROR("Failed to load MC firmware!\n");
6725 return r;
6726 }
9b136d51
AD
6727 }
6728
9b136d51
AD
6729 r = si_pcie_gart_enable(rdev);
6730 if (r)
6731 return r;
6732 si_gpu_init(rdev);
6733
9b136d51 6734 /* allocate rlc buffers */
1fd11777
AD
6735 if (rdev->family == CHIP_VERDE) {
6736 rdev->rlc.reg_list = verde_rlc_save_restore_register_list;
6737 rdev->rlc.reg_list_size =
6738 (u32)ARRAY_SIZE(verde_rlc_save_restore_register_list);
6739 }
6740 rdev->rlc.cs_data = si_cs_data;
6741 r = sumo_rlc_init(rdev);
9b136d51
AD
6742 if (r) {
6743 DRM_ERROR("Failed to init rlc BOs!\n");
6744 return r;
6745 }
6746
6747 /* allocate wb buffer */
6748 r = radeon_wb_init(rdev);
6749 if (r)
6750 return r;
6751
6752 r = radeon_fence_driver_start_ring(rdev, RADEON_RING_TYPE_GFX_INDEX);
6753 if (r) {
6754 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
6755 return r;
6756 }
6757
6758 r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
6759 if (r) {
6760 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
6761 return r;
6762 }
6763
6764 r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
6765 if (r) {
6766 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
6767 return r;
6768 }
6769
8c5fd7ef
AD
6770 r = radeon_fence_driver_start_ring(rdev, R600_RING_TYPE_DMA_INDEX);
6771 if (r) {
6772 dev_err(rdev->dev, "failed initializing DMA fences (%d).\n", r);
6773 return r;
6774 }
6775
6776 r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_DMA1_INDEX);
6777 if (r) {
6778 dev_err(rdev->dev, "failed initializing DMA fences (%d).\n", r);
6779 return r;
6780 }
6781
1df0d523 6782 if (rdev->has_uvd) {
e409b128 6783 r = uvd_v2_2_resume(rdev);
1df0d523
AD
6784 if (!r) {
6785 r = radeon_fence_driver_start_ring(rdev,
6786 R600_RING_TYPE_UVD_INDEX);
6787 if (r)
6788 dev_err(rdev->dev, "UVD fences init error (%d).\n", r);
6789 }
f2ba57b5 6790 if (r)
1df0d523 6791 rdev->ring[R600_RING_TYPE_UVD_INDEX].ring_size = 0;
f2ba57b5 6792 }
f2ba57b5 6793
9b136d51 6794 /* Enable IRQ */
e49f3959
AH
6795 if (!rdev->irq.installed) {
6796 r = radeon_irq_kms_init(rdev);
6797 if (r)
6798 return r;
6799 }
6800
9b136d51
AD
6801 r = si_irq_init(rdev);
6802 if (r) {
6803 DRM_ERROR("radeon: IH init failed (%d).\n", r);
6804 radeon_irq_kms_fini(rdev);
6805 return r;
6806 }
6807 si_irq_set(rdev);
6808
6809 ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
6810 r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP_RPTR_OFFSET,
2e1e6dad 6811 RADEON_CP_PACKET2);
9b136d51
AD
6812 if (r)
6813 return r;
6814
6815 ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
6816 r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP1_RPTR_OFFSET,
2e1e6dad 6817 RADEON_CP_PACKET2);
9b136d51
AD
6818 if (r)
6819 return r;
6820
6821 ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
6822 r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP2_RPTR_OFFSET,
2e1e6dad 6823 RADEON_CP_PACKET2);
9b136d51
AD
6824 if (r)
6825 return r;
6826
8c5fd7ef
AD
6827 ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
6828 r = radeon_ring_init(rdev, ring, ring->ring_size, R600_WB_DMA_RPTR_OFFSET,
2e1e6dad 6829 DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0));
8c5fd7ef
AD
6830 if (r)
6831 return r;
6832
6833 ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
6834 r = radeon_ring_init(rdev, ring, ring->ring_size, CAYMAN_WB_DMA1_RPTR_OFFSET,
2e1e6dad 6835 DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0));
8c5fd7ef
AD
6836 if (r)
6837 return r;
6838
9b136d51
AD
6839 r = si_cp_load_microcode(rdev);
6840 if (r)
6841 return r;
6842 r = si_cp_resume(rdev);
6843 if (r)
6844 return r;
6845
8c5fd7ef
AD
6846 r = cayman_dma_resume(rdev);
6847 if (r)
6848 return r;
6849
1df0d523
AD
6850 if (rdev->has_uvd) {
6851 ring = &rdev->ring[R600_RING_TYPE_UVD_INDEX];
6852 if (ring->ring_size) {
02c9f7fa 6853 r = radeon_ring_init(rdev, ring, ring->ring_size, 0,
2e1e6dad 6854 RADEON_CP_PACKET2);
1df0d523 6855 if (!r)
e409b128 6856 r = uvd_v1_0_init(rdev);
1df0d523
AD
6857 if (r)
6858 DRM_ERROR("radeon: failed initializing UVD (%d).\n", r);
6859 }
f2ba57b5
CK
6860 }
6861
2898c348
CK
6862 r = radeon_ib_pool_init(rdev);
6863 if (r) {
6864 dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
9b136d51 6865 return r;
2898c348 6866 }
9b136d51 6867
c6105f24
CK
6868 r = radeon_vm_manager_init(rdev);
6869 if (r) {
6870 dev_err(rdev->dev, "vm manager initialization failed (%d).\n", r);
9b136d51 6871 return r;
c6105f24 6872 }
9b136d51 6873
bfc1f97d 6874 r = radeon_audio_init(rdev);
b530602f
AD
6875 if (r)
6876 return r;
6877
9b136d51
AD
6878 return 0;
6879}
6880
6881int si_resume(struct radeon_device *rdev)
6882{
6883 int r;
6884
6885 /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw,
6886 * posting will perform necessary task to bring back GPU into good
6887 * shape.
6888 */
6889 /* post card */
6890 atom_asic_init(rdev->mode_info.atom_context);
6891
205996c0
AD
6892 /* init golden registers */
6893 si_init_golden_registers(rdev);
6894
bc6a6295
AD
6895 if (rdev->pm.pm_method == PM_METHOD_DPM)
6896 radeon_pm_resume(rdev);
6c7bccea 6897
9b136d51
AD
6898 rdev->accel_working = true;
6899 r = si_startup(rdev);
6900 if (r) {
6901 DRM_ERROR("si startup failed on resume\n");
6902 rdev->accel_working = false;
6903 return r;
6904 }
6905
6906 return r;
6907
6908}
6909
6910int si_suspend(struct radeon_device *rdev)
6911{
6c7bccea 6912 radeon_pm_suspend(rdev);
7991d665 6913 radeon_audio_fini(rdev);
fa3daf9a 6914 radeon_vm_manager_fini(rdev);
9b136d51 6915 si_cp_enable(rdev, false);
8c5fd7ef 6916 cayman_dma_stop(rdev);
1df0d523 6917 if (rdev->has_uvd) {
e409b128 6918 uvd_v1_0_fini(rdev);
1df0d523
AD
6919 radeon_uvd_suspend(rdev);
6920 }
e16866ec
AD
6921 si_fini_pg(rdev);
6922 si_fini_cg(rdev);
9b136d51
AD
6923 si_irq_suspend(rdev);
6924 radeon_wb_disable(rdev);
6925 si_pcie_gart_disable(rdev);
6926 return 0;
6927}
6928
6929/* Plan is to move initialization in that function and use
6930 * helper function so that radeon_device_init pretty much
6931 * do nothing more than calling asic specific function. This
6932 * should also allow to remove a bunch of callback function
6933 * like vram_info.
6934 */
6935int si_init(struct radeon_device *rdev)
6936{
6937 struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
6938 int r;
6939
9b136d51
AD
6940 /* Read BIOS */
6941 if (!radeon_get_bios(rdev)) {
6942 if (ASIC_IS_AVIVO(rdev))
6943 return -EINVAL;
6944 }
6945 /* Must be an ATOMBIOS */
6946 if (!rdev->is_atom_bios) {
6947 dev_err(rdev->dev, "Expecting atombios for cayman GPU\n");
6948 return -EINVAL;
6949 }
6950 r = radeon_atombios_init(rdev);
6951 if (r)
6952 return r;
6953
6954 /* Post card if necessary */
6955 if (!radeon_card_posted(rdev)) {
6956 if (!rdev->bios) {
6957 dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n");
6958 return -EINVAL;
6959 }
6960 DRM_INFO("GPU not posted. posting now...\n");
6961 atom_asic_init(rdev->mode_info.atom_context);
6962 }
205996c0
AD
6963 /* init golden registers */
6964 si_init_golden_registers(rdev);
9b136d51
AD
6965 /* Initialize scratch registers */
6966 si_scratch_init(rdev);
6967 /* Initialize surface registers */
6968 radeon_surface_init(rdev);
6969 /* Initialize clocks */
6970 radeon_get_clock_info(rdev->ddev);
6971
6972 /* Fence driver */
6973 r = radeon_fence_driver_init(rdev);
6974 if (r)
6975 return r;
6976
6977 /* initialize memory controller */
6978 r = si_mc_init(rdev);
6979 if (r)
6980 return r;
6981 /* Memory manager */
6982 r = radeon_bo_init(rdev);
6983 if (r)
6984 return r;
6985
01ac8794
AD
6986 if (!rdev->me_fw || !rdev->pfp_fw || !rdev->ce_fw ||
6987 !rdev->rlc_fw || !rdev->mc_fw) {
6988 r = si_init_microcode(rdev);
6989 if (r) {
6990 DRM_ERROR("Failed to load firmware!\n");
6991 return r;
6992 }
6993 }
6994
6c7bccea
AD
6995 /* Initialize power management */
6996 radeon_pm_init(rdev);
6997
9b136d51
AD
6998 ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
6999 ring->ring_obj = NULL;
7000 r600_ring_init(rdev, ring, 1024 * 1024);
7001
7002 ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
7003 ring->ring_obj = NULL;
7004 r600_ring_init(rdev, ring, 1024 * 1024);
7005
7006 ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
7007 ring->ring_obj = NULL;
7008 r600_ring_init(rdev, ring, 1024 * 1024);
7009
8c5fd7ef
AD
7010 ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
7011 ring->ring_obj = NULL;
7012 r600_ring_init(rdev, ring, 64 * 1024);
7013
7014 ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
7015 ring->ring_obj = NULL;
7016 r600_ring_init(rdev, ring, 64 * 1024);
7017
1df0d523
AD
7018 if (rdev->has_uvd) {
7019 r = radeon_uvd_init(rdev);
7020 if (!r) {
7021 ring = &rdev->ring[R600_RING_TYPE_UVD_INDEX];
7022 ring->ring_obj = NULL;
7023 r600_ring_init(rdev, ring, 4096);
7024 }
f2ba57b5
CK
7025 }
7026
9b136d51
AD
7027 rdev->ih.ring_obj = NULL;
7028 r600_ih_ring_init(rdev, 64 * 1024);
7029
7030 r = r600_pcie_gart_init(rdev);
7031 if (r)
7032 return r;
7033
9b136d51 7034 rdev->accel_working = true;
9b136d51
AD
7035 r = si_startup(rdev);
7036 if (r) {
7037 dev_err(rdev->dev, "disabling GPU acceleration\n");
7038 si_cp_fini(rdev);
8c5fd7ef 7039 cayman_dma_fini(rdev);
9b136d51 7040 si_irq_fini(rdev);
1fd11777 7041 sumo_rlc_fini(rdev);
9b136d51 7042 radeon_wb_fini(rdev);
2898c348 7043 radeon_ib_pool_fini(rdev);
9b136d51
AD
7044 radeon_vm_manager_fini(rdev);
7045 radeon_irq_kms_fini(rdev);
7046 si_pcie_gart_fini(rdev);
7047 rdev->accel_working = false;
7048 }
7049
7050 /* Don't start up if the MC ucode is missing.
7051 * The default clocks and voltages before the MC ucode
7052 * is loaded are not suffient for advanced operations.
7053 */
7054 if (!rdev->mc_fw) {
7055 DRM_ERROR("radeon: MC ucode required for NI+.\n");
7056 return -EINVAL;
7057 }
7058
7059 return 0;
7060}
7061
7062void si_fini(struct radeon_device *rdev)
7063{
6c7bccea 7064 radeon_pm_fini(rdev);
9b136d51 7065 si_cp_fini(rdev);
8c5fd7ef 7066 cayman_dma_fini(rdev);
f8f84ac5 7067 si_fini_pg(rdev);
e16866ec 7068 si_fini_cg(rdev);
9b136d51 7069 si_irq_fini(rdev);
1fd11777 7070 sumo_rlc_fini(rdev);
9b136d51
AD
7071 radeon_wb_fini(rdev);
7072 radeon_vm_manager_fini(rdev);
2898c348 7073 radeon_ib_pool_fini(rdev);
9b136d51 7074 radeon_irq_kms_fini(rdev);
2858c00d 7075 if (rdev->has_uvd) {
e409b128 7076 uvd_v1_0_fini(rdev);
1df0d523 7077 radeon_uvd_fini(rdev);
2858c00d 7078 }
9b136d51
AD
7079 si_pcie_gart_fini(rdev);
7080 r600_vram_scratch_fini(rdev);
7081 radeon_gem_fini(rdev);
9b136d51
AD
7082 radeon_fence_driver_fini(rdev);
7083 radeon_bo_fini(rdev);
7084 radeon_atombios_fini(rdev);
7085 kfree(rdev->bios);
7086 rdev->bios = NULL;
7087}
7088
6759a0a7 7089/**
d0418894 7090 * si_get_gpu_clock_counter - return GPU clock counter snapshot
6759a0a7
MO
7091 *
7092 * @rdev: radeon_device pointer
7093 *
7094 * Fetches a GPU clock counter snapshot (SI).
7095 * Returns the 64 bit clock counter snapshot.
7096 */
d0418894 7097uint64_t si_get_gpu_clock_counter(struct radeon_device *rdev)
6759a0a7
MO
7098{
7099 uint64_t clock;
7100
7101 mutex_lock(&rdev->gpu_clock_mutex);
7102 WREG32(RLC_CAPTURE_GPU_CLOCK_COUNT, 1);
7103 clock = (uint64_t)RREG32(RLC_GPU_CLOCK_COUNT_LSB) |
7104 ((uint64_t)RREG32(RLC_GPU_CLOCK_COUNT_MSB) << 32ULL);
7105 mutex_unlock(&rdev->gpu_clock_mutex);
7106 return clock;
7107}
2539eb02 7108
2539eb02
CK
7109int si_set_uvd_clocks(struct radeon_device *rdev, u32 vclk, u32 dclk)
7110{
facd112d 7111 unsigned fb_div = 0, vclk_div = 0, dclk_div = 0;
2539eb02
CK
7112 int r;
7113
4ed10835
CK
7114 /* bypass vclk and dclk with bclk */
7115 WREG32_P(CG_UPLL_FUNC_CNTL_2,
7116 VCLK_SRC_SEL(1) | DCLK_SRC_SEL(1),
7117 ~(VCLK_SRC_SEL_MASK | DCLK_SRC_SEL_MASK));
7118
7119 /* put PLL in bypass mode */
7120 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_BYPASS_EN_MASK, ~UPLL_BYPASS_EN_MASK);
7121
7122 if (!vclk || !dclk) {
7123 /* keep the Bypass mode, put PLL to sleep */
7124 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_SLEEP_MASK, ~UPLL_SLEEP_MASK);
7125 return 0;
7126 }
7127
facd112d
CK
7128 r = radeon_uvd_calc_upll_dividers(rdev, vclk, dclk, 125000, 250000,
7129 16384, 0x03FFFFFF, 0, 128, 5,
7130 &fb_div, &vclk_div, &dclk_div);
7131 if (r)
7132 return r;
2539eb02
CK
7133
7134 /* set RESET_ANTI_MUX to 0 */
7135 WREG32_P(CG_UPLL_FUNC_CNTL_5, 0, ~RESET_ANTI_MUX_MASK);
7136
7137 /* set VCO_MODE to 1 */
7138 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_VCO_MODE_MASK, ~UPLL_VCO_MODE_MASK);
7139
7140 /* toggle UPLL_SLEEP to 1 then back to 0 */
7141 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_SLEEP_MASK, ~UPLL_SLEEP_MASK);
7142 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_SLEEP_MASK);
7143
7144 /* deassert UPLL_RESET */
7145 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_RESET_MASK);
7146
7147 mdelay(1);
7148
facd112d 7149 r = radeon_uvd_send_upll_ctlreq(rdev, CG_UPLL_FUNC_CNTL);
2539eb02
CK
7150 if (r)
7151 return r;
7152
7153 /* assert UPLL_RESET again */
7154 WREG32_P(CG_UPLL_FUNC_CNTL, UPLL_RESET_MASK, ~UPLL_RESET_MASK);
7155
7156 /* disable spread spectrum. */
7157 WREG32_P(CG_UPLL_SPREAD_SPECTRUM, 0, ~SSEN_MASK);
7158
7159 /* set feedback divider */
facd112d 7160 WREG32_P(CG_UPLL_FUNC_CNTL_3, UPLL_FB_DIV(fb_div), ~UPLL_FB_DIV_MASK);
2539eb02
CK
7161
7162 /* set ref divider to 0 */
7163 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_REF_DIV_MASK);
7164
facd112d 7165 if (fb_div < 307200)
2539eb02
CK
7166 WREG32_P(CG_UPLL_FUNC_CNTL_4, 0, ~UPLL_SPARE_ISPARE9);
7167 else
7168 WREG32_P(CG_UPLL_FUNC_CNTL_4, UPLL_SPARE_ISPARE9, ~UPLL_SPARE_ISPARE9);
7169
7170 /* set PDIV_A and PDIV_B */
7171 WREG32_P(CG_UPLL_FUNC_CNTL_2,
facd112d 7172 UPLL_PDIV_A(vclk_div) | UPLL_PDIV_B(dclk_div),
2539eb02
CK
7173 ~(UPLL_PDIV_A_MASK | UPLL_PDIV_B_MASK));
7174
7175 /* give the PLL some time to settle */
7176 mdelay(15);
7177
7178 /* deassert PLL_RESET */
7179 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_RESET_MASK);
7180
7181 mdelay(15);
7182
7183 /* switch from bypass mode to normal mode */
7184 WREG32_P(CG_UPLL_FUNC_CNTL, 0, ~UPLL_BYPASS_EN_MASK);
7185
facd112d 7186 r = radeon_uvd_send_upll_ctlreq(rdev, CG_UPLL_FUNC_CNTL);
2539eb02
CK
7187 if (r)
7188 return r;
7189
7190 /* switch VCLK and DCLK selection */
7191 WREG32_P(CG_UPLL_FUNC_CNTL_2,
7192 VCLK_SRC_SEL(2) | DCLK_SRC_SEL(2),
7193 ~(VCLK_SRC_SEL_MASK | DCLK_SRC_SEL_MASK));
7194
7195 mdelay(100);
7196
7197 return 0;
7198}
b9d305df
AD
7199
7200static void si_pcie_gen3_enable(struct radeon_device *rdev)
7201{
7202 struct pci_dev *root = rdev->pdev->bus->self;
7203 int bridge_pos, gpu_pos;
7204 u32 speed_cntl, mask, current_data_rate;
7205 int ret, i;
7206 u16 tmp16;
7207
0bd252de
AW
7208 if (pci_is_root_bus(rdev->pdev->bus))
7209 return;
7210
b9d305df
AD
7211 if (radeon_pcie_gen2 == 0)
7212 return;
7213
7214 if (rdev->flags & RADEON_IS_IGP)
7215 return;
7216
7217 if (!(rdev->flags & RADEON_IS_PCIE))
7218 return;
7219
7220 ret = drm_pcie_get_speed_cap_mask(rdev->ddev, &mask);
7221 if (ret != 0)
7222 return;
7223
7224 if (!(mask & (DRM_PCIE_SPEED_50 | DRM_PCIE_SPEED_80)))
7225 return;
7226
7227 speed_cntl = RREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL);
7228 current_data_rate = (speed_cntl & LC_CURRENT_DATA_RATE_MASK) >>
7229 LC_CURRENT_DATA_RATE_SHIFT;
7230 if (mask & DRM_PCIE_SPEED_80) {
7231 if (current_data_rate == 2) {
7232 DRM_INFO("PCIE gen 3 link speeds already enabled\n");
7233 return;
7234 }
7235 DRM_INFO("enabling PCIE gen 3 link speeds, disable with radeon.pcie_gen2=0\n");
7236 } else if (mask & DRM_PCIE_SPEED_50) {
7237 if (current_data_rate == 1) {
7238 DRM_INFO("PCIE gen 2 link speeds already enabled\n");
7239 return;
7240 }
7241 DRM_INFO("enabling PCIE gen 2 link speeds, disable with radeon.pcie_gen2=0\n");
7242 }
7243
7244 bridge_pos = pci_pcie_cap(root);
7245 if (!bridge_pos)
7246 return;
7247
7248 gpu_pos = pci_pcie_cap(rdev->pdev);
7249 if (!gpu_pos)
7250 return;
7251
7252 if (mask & DRM_PCIE_SPEED_80) {
7253 /* re-try equalization if gen3 is not already enabled */
7254 if (current_data_rate != 2) {
7255 u16 bridge_cfg, gpu_cfg;
7256 u16 bridge_cfg2, gpu_cfg2;
7257 u32 max_lw, current_lw, tmp;
7258
7259 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL, &bridge_cfg);
7260 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, &gpu_cfg);
7261
7262 tmp16 = bridge_cfg | PCI_EXP_LNKCTL_HAWD;
7263 pci_write_config_word(root, bridge_pos + PCI_EXP_LNKCTL, tmp16);
7264
7265 tmp16 = gpu_cfg | PCI_EXP_LNKCTL_HAWD;
7266 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, tmp16);
7267
7268 tmp = RREG32_PCIE(PCIE_LC_STATUS1);
7269 max_lw = (tmp & LC_DETECTED_LINK_WIDTH_MASK) >> LC_DETECTED_LINK_WIDTH_SHIFT;
7270 current_lw = (tmp & LC_OPERATING_LINK_WIDTH_MASK) >> LC_OPERATING_LINK_WIDTH_SHIFT;
7271
7272 if (current_lw < max_lw) {
7273 tmp = RREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL);
7274 if (tmp & LC_RENEGOTIATION_SUPPORT) {
7275 tmp &= ~(LC_LINK_WIDTH_MASK | LC_UPCONFIGURE_DIS);
7276 tmp |= (max_lw << LC_LINK_WIDTH_SHIFT);
7277 tmp |= LC_UPCONFIGURE_SUPPORT | LC_RENEGOTIATE_EN | LC_RECONFIG_NOW;
7278 WREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL, tmp);
7279 }
7280 }
7281
7282 for (i = 0; i < 10; i++) {
7283 /* check status */
7284 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_DEVSTA, &tmp16);
7285 if (tmp16 & PCI_EXP_DEVSTA_TRPND)
7286 break;
7287
7288 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL, &bridge_cfg);
7289 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, &gpu_cfg);
7290
7291 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL2, &bridge_cfg2);
7292 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, &gpu_cfg2);
7293
7294 tmp = RREG32_PCIE_PORT(PCIE_LC_CNTL4);
7295 tmp |= LC_SET_QUIESCE;
7296 WREG32_PCIE_PORT(PCIE_LC_CNTL4, tmp);
7297
7298 tmp = RREG32_PCIE_PORT(PCIE_LC_CNTL4);
7299 tmp |= LC_REDO_EQ;
7300 WREG32_PCIE_PORT(PCIE_LC_CNTL4, tmp);
7301
7302 mdelay(100);
7303
7304 /* linkctl */
7305 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL, &tmp16);
7306 tmp16 &= ~PCI_EXP_LNKCTL_HAWD;
7307 tmp16 |= (bridge_cfg & PCI_EXP_LNKCTL_HAWD);
7308 pci_write_config_word(root, bridge_pos + PCI_EXP_LNKCTL, tmp16);
7309
7310 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, &tmp16);
7311 tmp16 &= ~PCI_EXP_LNKCTL_HAWD;
7312 tmp16 |= (gpu_cfg & PCI_EXP_LNKCTL_HAWD);
7313 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, tmp16);
7314
7315 /* linkctl2 */
7316 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL2, &tmp16);
7317 tmp16 &= ~((1 << 4) | (7 << 9));
7318 tmp16 |= (bridge_cfg2 & ((1 << 4) | (7 << 9)));
7319 pci_write_config_word(root, bridge_pos + PCI_EXP_LNKCTL2, tmp16);
7320
7321 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, &tmp16);
7322 tmp16 &= ~((1 << 4) | (7 << 9));
7323 tmp16 |= (gpu_cfg2 & ((1 << 4) | (7 << 9)));
7324 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, tmp16);
7325
7326 tmp = RREG32_PCIE_PORT(PCIE_LC_CNTL4);
7327 tmp &= ~LC_SET_QUIESCE;
7328 WREG32_PCIE_PORT(PCIE_LC_CNTL4, tmp);
7329 }
7330 }
7331 }
7332
7333 /* set the link speed */
7334 speed_cntl |= LC_FORCE_EN_SW_SPEED_CHANGE | LC_FORCE_DIS_HW_SPEED_CHANGE;
7335 speed_cntl &= ~LC_FORCE_DIS_SW_SPEED_CHANGE;
7336 WREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL, speed_cntl);
7337
7338 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, &tmp16);
7339 tmp16 &= ~0xf;
7340 if (mask & DRM_PCIE_SPEED_80)
7341 tmp16 |= 3; /* gen3 */
7342 else if (mask & DRM_PCIE_SPEED_50)
7343 tmp16 |= 2; /* gen2 */
7344 else
7345 tmp16 |= 1; /* gen1 */
7346 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, tmp16);
7347
7348 speed_cntl = RREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL);
7349 speed_cntl |= LC_INITIATE_LINK_SPEED_CHANGE;
7350 WREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL, speed_cntl);
7351
7352 for (i = 0; i < rdev->usec_timeout; i++) {
7353 speed_cntl = RREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL);
7354 if ((speed_cntl & LC_INITIATE_LINK_SPEED_CHANGE) == 0)
7355 break;
7356 udelay(1);
7357 }
7358}
7359
e0bcf165
AD
7360static void si_program_aspm(struct radeon_device *rdev)
7361{
7362 u32 data, orig;
7363 bool disable_l0s = false, disable_l1 = false, disable_plloff_in_l1 = false;
7364 bool disable_clkreq = false;
7365
1294d4a3
AD
7366 if (radeon_aspm == 0)
7367 return;
7368
e0bcf165
AD
7369 if (!(rdev->flags & RADEON_IS_PCIE))
7370 return;
7371
7372 orig = data = RREG32_PCIE_PORT(PCIE_LC_N_FTS_CNTL);
7373 data &= ~LC_XMIT_N_FTS_MASK;
7374 data |= LC_XMIT_N_FTS(0x24) | LC_XMIT_N_FTS_OVERRIDE_EN;
7375 if (orig != data)
7376 WREG32_PCIE_PORT(PCIE_LC_N_FTS_CNTL, data);
7377
7378 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL3);
7379 data |= LC_GO_TO_RECOVERY;
7380 if (orig != data)
7381 WREG32_PCIE_PORT(PCIE_LC_CNTL3, data);
7382
7383 orig = data = RREG32_PCIE(PCIE_P_CNTL);
7384 data |= P_IGNORE_EDB_ERR;
7385 if (orig != data)
7386 WREG32_PCIE(PCIE_P_CNTL, data);
7387
7388 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL);
7389 data &= ~(LC_L0S_INACTIVITY_MASK | LC_L1_INACTIVITY_MASK);
7390 data |= LC_PMI_TO_L1_DIS;
7391 if (!disable_l0s)
7392 data |= LC_L0S_INACTIVITY(7);
7393
7394 if (!disable_l1) {
7395 data |= LC_L1_INACTIVITY(7);
7396 data &= ~LC_PMI_TO_L1_DIS;
7397 if (orig != data)
7398 WREG32_PCIE_PORT(PCIE_LC_CNTL, data);
7399
7400 if (!disable_plloff_in_l1) {
7401 bool clk_req_support;
7402
7403 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_0);
7404 data &= ~(PLL_POWER_STATE_IN_OFF_0_MASK | PLL_POWER_STATE_IN_TXS2_0_MASK);
7405 data |= PLL_POWER_STATE_IN_OFF_0(7) | PLL_POWER_STATE_IN_TXS2_0(7);
7406 if (orig != data)
7407 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_0, data);
7408
7409 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_1);
7410 data &= ~(PLL_POWER_STATE_IN_OFF_1_MASK | PLL_POWER_STATE_IN_TXS2_1_MASK);
7411 data |= PLL_POWER_STATE_IN_OFF_1(7) | PLL_POWER_STATE_IN_TXS2_1(7);
7412 if (orig != data)
7413 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_1, data);
7414
7415 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_0);
7416 data &= ~(PLL_POWER_STATE_IN_OFF_0_MASK | PLL_POWER_STATE_IN_TXS2_0_MASK);
7417 data |= PLL_POWER_STATE_IN_OFF_0(7) | PLL_POWER_STATE_IN_TXS2_0(7);
7418 if (orig != data)
7419 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_0, data);
7420
7421 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_1);
7422 data &= ~(PLL_POWER_STATE_IN_OFF_1_MASK | PLL_POWER_STATE_IN_TXS2_1_MASK);
7423 data |= PLL_POWER_STATE_IN_OFF_1(7) | PLL_POWER_STATE_IN_TXS2_1(7);
7424 if (orig != data)
7425 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_1, data);
7426
7427 if ((rdev->family != CHIP_OLAND) && (rdev->family != CHIP_HAINAN)) {
7428 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_0);
7429 data &= ~PLL_RAMP_UP_TIME_0_MASK;
7430 if (orig != data)
7431 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_0, data);
7432
7433 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_1);
7434 data &= ~PLL_RAMP_UP_TIME_1_MASK;
7435 if (orig != data)
7436 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_1, data);
7437
7438 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_2);
7439 data &= ~PLL_RAMP_UP_TIME_2_MASK;
7440 if (orig != data)
7441 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_2, data);
7442
7443 orig = data = RREG32_PIF_PHY0(PB0_PIF_PWRDOWN_3);
7444 data &= ~PLL_RAMP_UP_TIME_3_MASK;
7445 if (orig != data)
7446 WREG32_PIF_PHY0(PB0_PIF_PWRDOWN_3, data);
7447
7448 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_0);
7449 data &= ~PLL_RAMP_UP_TIME_0_MASK;
7450 if (orig != data)
7451 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_0, data);
7452
7453 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_1);
7454 data &= ~PLL_RAMP_UP_TIME_1_MASK;
7455 if (orig != data)
7456 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_1, data);
7457
7458 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_2);
7459 data &= ~PLL_RAMP_UP_TIME_2_MASK;
7460 if (orig != data)
7461 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_2, data);
7462
7463 orig = data = RREG32_PIF_PHY1(PB1_PIF_PWRDOWN_3);
7464 data &= ~PLL_RAMP_UP_TIME_3_MASK;
7465 if (orig != data)
7466 WREG32_PIF_PHY1(PB1_PIF_PWRDOWN_3, data);
7467 }
7468 orig = data = RREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL);
7469 data &= ~LC_DYN_LANES_PWR_STATE_MASK;
7470 data |= LC_DYN_LANES_PWR_STATE(3);
7471 if (orig != data)
7472 WREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL, data);
7473
7474 orig = data = RREG32_PIF_PHY0(PB0_PIF_CNTL);
7475 data &= ~LS2_EXIT_TIME_MASK;
7476 if ((rdev->family == CHIP_OLAND) || (rdev->family == CHIP_HAINAN))
7477 data |= LS2_EXIT_TIME(5);
7478 if (orig != data)
7479 WREG32_PIF_PHY0(PB0_PIF_CNTL, data);
7480
7481 orig = data = RREG32_PIF_PHY1(PB1_PIF_CNTL);
7482 data &= ~LS2_EXIT_TIME_MASK;
7483 if ((rdev->family == CHIP_OLAND) || (rdev->family == CHIP_HAINAN))
7484 data |= LS2_EXIT_TIME(5);
7485 if (orig != data)
7486 WREG32_PIF_PHY1(PB1_PIF_CNTL, data);
7487
0bd252de
AW
7488 if (!disable_clkreq &&
7489 !pci_is_root_bus(rdev->pdev->bus)) {
e0bcf165
AD
7490 struct pci_dev *root = rdev->pdev->bus->self;
7491 u32 lnkcap;
7492
7493 clk_req_support = false;
7494 pcie_capability_read_dword(root, PCI_EXP_LNKCAP, &lnkcap);
7495 if (lnkcap & PCI_EXP_LNKCAP_CLKPM)
7496 clk_req_support = true;
7497 } else {
7498 clk_req_support = false;
7499 }
7500
7501 if (clk_req_support) {
7502 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL2);
7503 data |= LC_ALLOW_PDWN_IN_L1 | LC_ALLOW_PDWN_IN_L23;
7504 if (orig != data)
7505 WREG32_PCIE_PORT(PCIE_LC_CNTL2, data);
7506
7507 orig = data = RREG32(THM_CLK_CNTL);
7508 data &= ~(CMON_CLK_SEL_MASK | TMON_CLK_SEL_MASK);
7509 data |= CMON_CLK_SEL(1) | TMON_CLK_SEL(1);
7510 if (orig != data)
7511 WREG32(THM_CLK_CNTL, data);
7512
7513 orig = data = RREG32(MISC_CLK_CNTL);
7514 data &= ~(DEEP_SLEEP_CLK_SEL_MASK | ZCLK_SEL_MASK);
7515 data |= DEEP_SLEEP_CLK_SEL(1) | ZCLK_SEL(1);
7516 if (orig != data)
7517 WREG32(MISC_CLK_CNTL, data);
7518
7519 orig = data = RREG32(CG_CLKPIN_CNTL);
7520 data &= ~BCLK_AS_XCLK;
7521 if (orig != data)
7522 WREG32(CG_CLKPIN_CNTL, data);
7523
7524 orig = data = RREG32(CG_CLKPIN_CNTL_2);
7525 data &= ~FORCE_BIF_REFCLK_EN;
7526 if (orig != data)
7527 WREG32(CG_CLKPIN_CNTL_2, data);
7528
7529 orig = data = RREG32(MPLL_BYPASSCLK_SEL);
7530 data &= ~MPLL_CLKOUT_SEL_MASK;
7531 data |= MPLL_CLKOUT_SEL(4);
7532 if (orig != data)
7533 WREG32(MPLL_BYPASSCLK_SEL, data);
7534
7535 orig = data = RREG32(SPLL_CNTL_MODE);
7536 data &= ~SPLL_REFCLK_SEL_MASK;
7537 if (orig != data)
7538 WREG32(SPLL_CNTL_MODE, data);
7539 }
7540 }
7541 } else {
7542 if (orig != data)
7543 WREG32_PCIE_PORT(PCIE_LC_CNTL, data);
7544 }
7545
7546 orig = data = RREG32_PCIE(PCIE_CNTL2);
7547 data |= SLV_MEM_LS_EN | MST_MEM_LS_EN | REPLAY_MEM_LS_EN;
7548 if (orig != data)
7549 WREG32_PCIE(PCIE_CNTL2, data);
7550
7551 if (!disable_l0s) {
7552 data = RREG32_PCIE_PORT(PCIE_LC_N_FTS_CNTL);
7553 if((data & LC_N_FTS_MASK) == LC_N_FTS_MASK) {
7554 data = RREG32_PCIE(PCIE_LC_STATUS1);
7555 if ((data & LC_REVERSE_XMIT) && (data & LC_REVERSE_RCVR)) {
7556 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL);
7557 data &= ~LC_L0S_INACTIVITY_MASK;
7558 if (orig != data)
7559 WREG32_PCIE_PORT(PCIE_LC_CNTL, data);
7560 }
7561 }
7562 }
7563}