1 /* SPDX-License-Identifier: GPL-2.0-only */
2
3 #include <commonlib/helpers.h>
4 #include <stdint.h>
5 #include <arch/cpu.h>
6 #include <device/mmio.h>
7 #include <device/pci_ops.h>
8 #include <device/pci_def.h>
9 #include <device/device.h>
10 #include <device/smbus_host.h>
11 #include <spd.h>
12 #include <console/console.h>
13 #include <lib.h>
14 #include <delay.h>
15 #include <timestamp.h>
16 #include "gm45.h"
17 #include "chip.h"
18
19 static const gmch_gfx_t gmch_gfx_types[][5] = {
20 /* MAX_667MHz MAX_533MHz MAX_400MHz MAX_333MHz MAX_800MHz */
21 { GMCH_UNKNOWN, GMCH_UNKNOWN, GMCH_UNKNOWN, GMCH_UNKNOWN, GMCH_UNKNOWN },
22 { GMCH_GM47, GMCH_GM45, GMCH_UNKNOWN, GMCH_UNKNOWN, GMCH_GM49 },
23 { GMCH_GE45, GMCH_GE45, GMCH_GE45, GMCH_GE45, GMCH_GE45 },
24 { GMCH_UNKNOWN, GMCH_GL43, GMCH_GL40, GMCH_UNKNOWN, GMCH_UNKNOWN },
25 { GMCH_UNKNOWN, GMCH_GS45, GMCH_GS40, GMCH_UNKNOWN, GMCH_UNKNOWN },
26 { GMCH_UNKNOWN, GMCH_UNKNOWN, GMCH_UNKNOWN, GMCH_UNKNOWN, GMCH_UNKNOWN },
27 { GMCH_UNKNOWN, GMCH_UNKNOWN, GMCH_UNKNOWN, GMCH_UNKNOWN, GMCH_UNKNOWN },
28 { GMCH_PM45, GMCH_PM45, GMCH_PM45, GMCH_PM45, GMCH_PM45 },
29 };
30
get_gmch_info(sysinfo_t * sysinfo)31 void get_gmch_info(sysinfo_t *sysinfo)
32 {
33 sysinfo->stepping = pci_read_config8(PCI_DEV(0, 0, 0), PCI_CLASS_REVISION);
34 if ((sysinfo->stepping > STEPPING_B3) &&
35 (sysinfo->stepping != STEPPING_CONVERSION_A1))
36 die("Unknown stepping.\n");
37 if (sysinfo->stepping <= STEPPING_B3)
38 printk(BIOS_DEBUG, "Stepping %c%d\n", 'A' + sysinfo->stepping / 4, sysinfo->stepping % 4);
39 else
40 printk(BIOS_DEBUG, "Conversion stepping A1\n");
41
42 const u32 eax = cpuid_ext(0x04, 0).eax;
43 sysinfo->cores = ((eax >> 26) & 0x3f) + 1;
44 printk(BIOS_SPEW, "%d CPU cores\n", sysinfo->cores);
45
46 u32 capid = pci_read_config16(PCI_DEV(0, 0, 0), D0F0_CAPID0+8);
47 if (!(capid & (1<<(79-64)))) {
48 printk(BIOS_SPEW, "iTPM enabled\n");
49 }
50
51 capid = pci_read_config32(PCI_DEV(0, 0, 0), D0F0_CAPID0+4);
52 if (!(capid & (1<<(57-32)))) {
53 printk(BIOS_SPEW, "ME enabled\n");
54 }
55
56 if (!(capid & (1<<(56-32)))) {
57 printk(BIOS_SPEW, "AMT enabled\n");
58 }
59
60 sysinfo->max_ddr2_mt = (capid & (1<<(53-32)))?667:800;
61 printk(BIOS_SPEW, "capable of DDR2 of %d MHz or lower\n", sysinfo->max_ddr2_mt);
62
63 if (!(capid & (1<<(48-32)))) {
64 printk(BIOS_SPEW, "VT-d enabled\n");
65 }
66
67 const u32 gfx_variant = (capid>>(42-32)) & 0x7;
68 const u32 render_freq = ((capid>>(50-32) & 0x1) << 2) | ((capid>>(35-32)) & 0x3);
69 if (render_freq <= 4)
70 sysinfo->gfx_type = gmch_gfx_types[gfx_variant][render_freq];
71 else
72 sysinfo->gfx_type = GMCH_UNKNOWN;
73 switch (sysinfo->gfx_type) {
74 case GMCH_GM45:
75 printk(BIOS_SPEW, "GMCH: GM45\n");
76 break;
77 case GMCH_GM47:
78 printk(BIOS_SPEW, "GMCH: GM47\n");
79 break;
80 case GMCH_GM49:
81 printk(BIOS_SPEW, "GMCH: GM49\n");
82 break;
83 case GMCH_GE45:
84 printk(BIOS_SPEW, "GMCH: GE45\n");
85 break;
86 case GMCH_GL40:
87 printk(BIOS_SPEW, "GMCH: GL40\n");
88 break;
89 case GMCH_GL43:
90 printk(BIOS_SPEW, "GMCH: GL43\n");
91 break;
92 case GMCH_GS40:
93 printk(BIOS_SPEW, "GMCH: GS40\n");
94 break;
95 case GMCH_GS45:
96 printk(BIOS_SPEW, "GMCH: GS45, using %s-power mode\n",
97 sysinfo->gs45_low_power_mode ? "low" : "high");
98 break;
99 case GMCH_PM45:
100 printk(BIOS_SPEW, "GMCH: PM45\n");
101 break;
102 case GMCH_UNKNOWN:
103 printk(BIOS_SPEW, "unknown GMCH\n");
104 break;
105 }
106
107 sysinfo->txt_enabled = !(capid & (1 << (37-32)));
108 if (sysinfo->txt_enabled) {
109 printk(BIOS_SPEW, "TXT enabled\n");
110 }
111
112 switch (render_freq) {
113 case 4:
114 sysinfo->max_render_mhz = 800;
115 break;
116 case 0:
117 sysinfo->max_render_mhz = 667;
118 break;
119 case 1:
120 sysinfo->max_render_mhz = 533;
121 break;
122 case 2:
123 sysinfo->max_render_mhz = 400;
124 break;
125 case 3:
126 sysinfo->max_render_mhz = 333;
127 break;
128 default:
129 printk(BIOS_SPEW, "Unknown render frequency\n");
130 sysinfo->max_render_mhz = 0;
131 break;
132 }
133 if (sysinfo->max_render_mhz != 0) {
134 printk(BIOS_SPEW, "Render frequency: %d MHz\n", sysinfo->max_render_mhz);
135 }
136
137 if (!(capid & (1<<(33-32)))) {
138 printk(BIOS_SPEW, "IGD enabled\n");
139 }
140
141 if (!(capid & (1<<(32-32)))) {
142 printk(BIOS_SPEW, "PCIe-to-GMCH enabled\n");
143 }
144
145 capid = pci_read_config32(PCI_DEV(0, 0, 0), D0F0_CAPID0);
146
147 u32 ddr_cap = capid>>30 & 0x3;
148 switch (ddr_cap) {
149 case 0:
150 sysinfo->max_ddr3_mt = 1067;
151 break;
152 case 1:
153 sysinfo->max_ddr3_mt = 800;
154 break;
155 case 2:
156 case 3:
157 printk(BIOS_SPEW, "GMCH not DDR3 capable\n");
158 sysinfo->max_ddr3_mt = 0;
159 break;
160 }
161 if (sysinfo->max_ddr3_mt != 0) {
162 printk(BIOS_SPEW, "GMCH supports DDR3 with %d MT or less\n", sysinfo->max_ddr3_mt);
163 }
164
165 const unsigned int max_fsb = (capid >> 28) & 0x3;
166 switch (max_fsb) {
167 case 1:
168 sysinfo->max_fsb_mhz = 1067;
169 break;
170 case 2:
171 sysinfo->max_fsb_mhz = 800;
172 break;
173 case 3:
174 sysinfo->max_fsb_mhz = 667;
175 break;
176 default:
177 die("unknown FSB capability\n");
178 break;
179 }
180 if (sysinfo->max_fsb_mhz != 0) {
181 printk(BIOS_SPEW, "GMCH supports FSB with up to %d MHz\n", sysinfo->max_fsb_mhz);
182 }
183 sysinfo->max_fsb = max_fsb - 1;
184 }
185
186 /*
187 * Detect if the system went through an interrupted RAM init or is incon-
188 * sistent. If so, initiate a cold reboot. Otherwise mark the system to be
189 * in RAM init, so this function would detect it on an erroneous reboot.
190 */
enter_raminit_or_reset(void)191 void enter_raminit_or_reset(void)
192 {
193 /* Interrupted RAM init or inconsistent system? */
194 u8 reg8 = pci_read_config8(PCI_DEV(0, 0x1f, 0), 0xa2);
195
196 if (reg8 & (1 << 2)) { /* S4-assertion-width violation */
197 /* Ignore S4-assertion-width violation like original BIOS. */
198 printk(BIOS_WARNING, "Ignoring S4-assertion-width violation.\n");
199 /* Bit2 is R/WC, so it will clear itself below. */
200 }
201
202 if (reg8 & (1 << 7)) { /* interrupted RAM init */
203 /* Don't enable S4-assertion stretch. Makes trouble on roda/rk9.
204 reg8 = pci_read_config8(PCI_DEV(0, 0x1f, 0), 0xa4);
205 pci_write_config8(PCI_DEV(0, 0x1f, 0), 0xa4, reg8 | 0x08);
206 */
207
208 /* Clear bit7. */
209 pci_write_config8(PCI_DEV(0, 0x1f, 0), 0xa2, reg8 & ~(1 << 7));
210
211 printk(BIOS_INFO, "Interrupted RAM init, reset required.\n");
212 gm45_early_reset();
213 }
214 /* Mark system to be in RAM init. */
215 pci_write_config8(PCI_DEV(0, 0x1f, 0), 0xa2, reg8 | (1 << 7));
216 }
217
218 /* For a detected DIMM, test the value of an SPD byte to
219 match the expected value after masking some bits. */
test_dimm(sysinfo_t * const sysinfo,int dimm,int addr,int bitmask,int expected)220 static int test_dimm(sysinfo_t *const sysinfo,
221 int dimm, int addr, int bitmask, int expected)
222 {
223 return (smbus_read_byte(sysinfo->spd_map[dimm], addr) & bitmask) == expected;
224 }
225
226 /* This function dies if dimm is unsuitable for the chipset. */
verify_ddr2_dimm(sysinfo_t * const sysinfo,int dimm)227 static void verify_ddr2_dimm(sysinfo_t *const sysinfo, int dimm)
228 {
229 if (!test_dimm(sysinfo, dimm, 20, 0x04, 0x04))
230 die("Chipset only supports SO-DIMM\n");
231
232 if (!test_dimm(sysinfo, dimm, 6, 0xff, 0x40) ||
233 !test_dimm(sysinfo, dimm, 11, 0xff, 0x00))
234 die("Chipset doesn't support ECC RAM\n");
235
236 if (!test_dimm(sysinfo, dimm, 5, 0x07, 0) &&
237 !test_dimm(sysinfo, dimm, 5, 0x07, 1))
238 die("Chipset wants single or dual ranked DIMMs\n");
239
240 /*
241 * Generally supports:
242 * x8/x16
243 * 4 or 8 banks
244 * 10 column address bits
245 * 13, 14 or 15 (x8 only) row address bits
246 *
247 * FIXME: There seems to be an exception for 256Gb x16 chips. Not
248 * covered by the numbers above (9 column address bits?).
249 */
250 if (!test_dimm(sysinfo, dimm, 13, 0xff, 8) &&
251 !test_dimm(sysinfo, dimm, 13, 0xff, 16))
252 die("Chipset requires x8 or x16 width\n");
253
254 if (!test_dimm(sysinfo, dimm, 17, 0xff, 4) &&
255 !test_dimm(sysinfo, dimm, 17, 0xff, 8))
256 die("Chipset requires 4 or 8 banks\n");
257
258 if (!test_dimm(sysinfo, dimm, 4, 0xff, 10))
259 die("Chipset requires 10 column address bits\n");
260
261 if (!test_dimm(sysinfo, dimm, 3, 0xff, 13) &&
262 !test_dimm(sysinfo, dimm, 3, 0xff, 14) &&
263 !(test_dimm(sysinfo, dimm, 3, 0xff, 15) &&
264 test_dimm(sysinfo, dimm, 13, 0xff, 8)))
265 die("Chipset requires 13, 14 or 15 (with x8) row address bits");
266 }
267
268 /* For every detected DIMM, test if it's suitable for the chipset. */
verify_ddr2(sysinfo_t * const sysinfo,int mask)269 static void verify_ddr2(sysinfo_t *const sysinfo, int mask)
270 {
271 int cur;
272 for (cur = 0; mask; mask >>= 1, ++cur) {
273 if (mask & 1)
274 verify_ddr2_dimm(sysinfo, cur);
275 }
276 }
277
278 /* This function dies if dimm is unsuitable for the chipset. */
verify_ddr3_dimm(sysinfo_t * const sysinfo,int dimm)279 static void verify_ddr3_dimm(sysinfo_t *const sysinfo, int dimm)
280 {
281 if (!test_dimm(sysinfo, dimm, 3, 15, 3))
282 die("Chipset only supports SO-DIMM\n");
283
284 if (!test_dimm(sysinfo, dimm, 8, 0x18, 0))
285 die("Chipset doesn't support ECC RAM\n");
286
287 if (!test_dimm(sysinfo, dimm, 7, 0x38, 0) &&
288 !test_dimm(sysinfo, dimm, 7, 0x38, 8))
289 die("Chipset wants single or double sided DIMMs\n");
290
291 if (!test_dimm(sysinfo, dimm, 7, 7, 1) &&
292 !test_dimm(sysinfo, dimm, 7, 7, 2))
293 die("Chipset requires x8 or x16 width\n");
294
295 if (!test_dimm(sysinfo, dimm, 4, 0x0f, 0) &&
296 !test_dimm(sysinfo, dimm, 4, 0x0f, 1) &&
297 !test_dimm(sysinfo, dimm, 4, 0x0f, 2) &&
298 !test_dimm(sysinfo, dimm, 4, 0x0f, 3))
299 die("Chipset requires 256Mb, 512Mb, 1Gb or 2Gb chips.");
300
301 if (!test_dimm(sysinfo, dimm, 4, 0x70, 0))
302 die("Chipset requires 8 banks on DDR3\n");
303
304 /* How to check if burst length is 8?
305 Other values are not supported, are they even possible? */
306
307 if (!test_dimm(sysinfo, dimm, 10, 0xff, 1))
308 die("Code assumes 1/8ns MTB\n");
309
310 if (!test_dimm(sysinfo, dimm, 11, 0xff, 8))
311 die("Code assumes 1/8ns MTB\n");
312
313 if (!test_dimm(sysinfo, dimm, 62, 0x9f, 0) &&
314 !test_dimm(sysinfo, dimm, 62, 0x9f, 1) &&
315 !test_dimm(sysinfo, dimm, 62, 0x9f, 2) &&
316 !test_dimm(sysinfo, dimm, 62, 0x9f, 3) &&
317 !test_dimm(sysinfo, dimm, 62, 0x9f, 5))
318 die("Only raw card types A, B, C, D and F are supported.\n");
319 }
320
321 /* For every detected DIMM, test if it's suitable for the chipset. */
verify_ddr3(sysinfo_t * const sysinfo,int mask)322 static void verify_ddr3(sysinfo_t *const sysinfo, int mask)
323 {
324 int cur = 0;
325 while (mask) {
326 if (mask & 1) {
327 verify_ddr3_dimm(sysinfo, cur);
328 }
329 mask >>= 1;
330 cur++;
331 }
332 }
333
334 typedef struct {
335 int dimm_mask;
336 struct spd_dimminfo {
337 unsigned int rows;
338 unsigned int cols;
339 unsigned int chip_capacity;
340 unsigned int banks;
341 unsigned int ranks;
342 unsigned int cas_latencies;
343 unsigned int tAAmin;
344 unsigned int tCKmin;
345 unsigned int width;
346 unsigned int tRAS;
347 unsigned int tRP;
348 unsigned int tRCD;
349 unsigned int tWR;
350 unsigned int page_size;
351 unsigned int raw_card;
352 unsigned int refresh;
353 } channel[2];
354 } spdinfo_t;
355 /**
356 * \brief Decode SPD tck cycle time
357 *
358 * Decodes a raw SPD data from a DDR2 DIMM.
359 * Returns cycle time in 1/256th ns.
360 */
spd_decode_tck_time(u8 c)361 static unsigned int spd_decode_tck_time(u8 c)
362 {
363 u8 high, low;
364
365 high = c >> 4;
366
367 switch (c & 0xf) {
368 case 0xa:
369 low = 25;
370 break;
371 case 0xb:
372 low = 33;
373 break;
374 case 0xc:
375 low = 66;
376 break;
377 case 0xd:
378 low = 75;
379 break;
380 case 0xe:
381 case 0xf:
382 die("Invalid tck setting. lower nibble is 0x%x\n", c & 0xf);
383 default:
384 low = (c & 0xf) * 10;
385 }
386
387 return ((high * 100 + low) << 8) / 100;
388 }
collect_ddr2_dimm(struct spd_dimminfo * const di,const int smb_addr)389 static void collect_ddr2_dimm(struct spd_dimminfo *const di, const int smb_addr)
390 {
391 static const int tCK_offsets[] = { 9, 23, 25 };
392
393 di->rows = smbus_read_byte(smb_addr, 3);
394 di->cols = smbus_read_byte(smb_addr, 4);
395 di->banks = smbus_read_byte(smb_addr, 17);
396 di->width = smbus_read_byte(smb_addr, 13) / 8; /* in bytes */
397
398 /* 0: 256Mb .. 3: 2Gb */
399 di->chip_capacity =
400 di->rows + di->cols
401 + (di->width == 1 ? 3 : 4) /* 1B: 2^3 bits, 2B: 2^4 bits */
402 + (di->banks == 4 ? 2 : 3) /* 4 banks: 2^2, 8 banks: 2^3 */
403 - 28;
404
405 di->page_size = di->width * (1 << di->cols); /* in bytes */
406
407 di->ranks = (smbus_read_byte(smb_addr, 5) & 7) + 1;
408
409 di->cas_latencies = smbus_read_byte(smb_addr, 18);
410 /* assuming tCKmin for the highest CAS is the absolute minimum */
411 di->tCKmin = spd_decode_tck_time(smbus_read_byte(smb_addr, 9));
412
413 /* try to reconstruct tAAmin from available data (I hate DDR2 SPDs) */
414 unsigned int i;
415 unsigned int cas = 7;
416 di->tAAmin = UINT32_MAX; /* we don't have UINT_MAX? */
417 for (i = 0; i < ARRAY_SIZE(tCK_offsets); ++i, --cas) {
418 for (; cas > 1; --cas)
419 if (di->cas_latencies & (1 << cas))
420 break;
421 if (cas <= 1)
422 break;
423
424 const unsigned int tCK_enc =
425 smbus_read_byte(smb_addr, tCK_offsets[i]);
426 const unsigned int tAA = spd_decode_tck_time(tCK_enc) * cas;
427 if (tAA < di->tAAmin)
428 di->tAAmin = tAA;
429 }
430
431 /* convert to 1/256ns */
432 di->tRAS = smbus_read_byte(smb_addr, 30) << 8; /* given in ns */
433 di->tRP = smbus_read_byte(smb_addr, 27) << 6; /* given in 1/4ns */
434 di->tRCD = smbus_read_byte(smb_addr, 29) << 6; /* given in 1/4ns */
435 di->tWR = smbus_read_byte(smb_addr, 36) << 6; /* given in 1/4ns */
436
437 di->raw_card = 0; /* Use same path as for DDR3 type A. */
438 di->refresh = smbus_read_byte(smb_addr, 12);
439 }
440 /*
441 * This function collects RAM characteristics from SPD, assuming that RAM
442 * is generally within chipset's requirements, since verify_ddr2() passed.
443 */
collect_ddr2(sysinfo_t * const sysinfo,spdinfo_t * const config)444 static void collect_ddr2(sysinfo_t *const sysinfo, spdinfo_t *const config)
445 {
446 int cur;
447 for (cur = 0; cur < 2; ++cur) {
448 if (config->dimm_mask & (1 << (2 * cur))) {
449 collect_ddr2_dimm(&config->channel[cur],
450 sysinfo->spd_map[2 * cur]);
451 }
452 }
453 }
454
455 /*
456 * This function collects RAM characteristics from SPD, assuming that RAM
457 * is generally within chipset's requirements, since verify_ddr3() passed.
458 */
collect_ddr3(sysinfo_t * const sysinfo,spdinfo_t * const config)459 static void collect_ddr3(sysinfo_t *const sysinfo, spdinfo_t *const config)
460 {
461 int mask = config->dimm_mask;
462 int cur = 0;
463 while (mask != 0) {
464 /* FIXME: support several dimms on same channel. */
465 if ((mask & 1) && sysinfo->spd_map[2 * cur]) {
466 int tmp;
467 const int smb_addr = sysinfo->spd_map[2 * cur];
468
469 config->channel[cur].rows = ((smbus_read_byte(smb_addr, 5) >> 3) & 7) + 12;
470 config->channel[cur].cols = (smbus_read_byte(smb_addr, 5) & 7) + 9;
471
472 config->channel[cur].chip_capacity = smbus_read_byte(smb_addr, 4) & 0xf;
473
474 config->channel[cur].banks = 8; /* GM45 only accepts this for DDR3.
475 verify_ddr3() fails for other values. */
476 config->channel[cur].ranks = ((smbus_read_byte(smb_addr, 7) >> 3) & 7) + 1;
477
478 config->channel[cur].cas_latencies =
479 ((smbus_read_byte(smb_addr, 15) << 8) | smbus_read_byte(smb_addr, 14))
480 << 4; /* so bit x is CAS x */
481 config->channel[cur].tAAmin = smbus_read_byte(smb_addr, 16) * 32; /* convert from MTB to 1/256 ns */
482 config->channel[cur].tCKmin = smbus_read_byte(smb_addr, 12) * 32; /* convert from MTB to 1/256 ns */
483
484 config->channel[cur].width = smbus_read_byte(smb_addr, 7) & 7;
485 config->channel[cur].page_size = config->channel[cur].width *
486 (1 << config->channel[cur].cols); /* in Bytes */
487
488 tmp = smbus_read_byte(smb_addr, 21);
489 config->channel[cur].tRAS = (smbus_read_byte(smb_addr, 22) | ((tmp & 0xf) << 8)) * 32;
490 config->channel[cur].tRP = smbus_read_byte(smb_addr, 20) * 32;
491 config->channel[cur].tRCD = smbus_read_byte(smb_addr, 18) * 32;
492 config->channel[cur].tWR = smbus_read_byte(smb_addr, 17) * 32;
493
494 config->channel[cur].raw_card = smbus_read_byte(smb_addr, 62) & 0x1f;
495 config->channel[cur].refresh = REFRESH_7_8;
496 }
497 cur++;
498 mask >>= 2;
499 }
500 }
501
read_fsb_clock(void)502 static fsb_clock_t read_fsb_clock(void)
503 {
504 switch (mchbar_read32(CLKCFG_MCHBAR) & CLKCFG_FSBCLK_MASK) {
505 case 6:
506 return FSB_CLOCK_1067MHz;
507 case 2:
508 return FSB_CLOCK_800MHz;
509 case 3:
510 return FSB_CLOCK_667MHz;
511 default:
512 die("Unsupported FSB clock.\n");
513 }
514 }
clock_index(const unsigned int clock)515 static mem_clock_t clock_index(const unsigned int clock)
516 {
517 switch (clock) {
518 case 533: return MEM_CLOCK_533MHz;
519 case 400: return MEM_CLOCK_400MHz;
520 case 333: return MEM_CLOCK_333MHz;
521 default: die("Unknown clock value.\n");
522 }
523 return -1; /* Won't be reached. */
524 }
normalize_clock(unsigned int * const clock)525 static void normalize_clock(unsigned int *const clock)
526 {
527 if (*clock >= 533)
528 *clock = 533;
529 else if (*clock >= 400)
530 *clock = 400;
531 else if (*clock >= 333)
532 *clock = 333;
533 else
534 *clock = 0;
535 }
lower_clock(unsigned int * const clock)536 static void lower_clock(unsigned int *const clock)
537 {
538 --*clock;
539 normalize_clock(clock);
540 }
find_common_clock_cas(sysinfo_t * const sysinfo,const spdinfo_t * const spdinfo)541 static unsigned int find_common_clock_cas(sysinfo_t *const sysinfo,
542 const spdinfo_t *const spdinfo)
543 {
544 /* various constraints must be fulfilled:
545 CAS * tCK < 20ns == 160MTB
546 tCK_max >= tCK >= tCK_min
547 CAS >= roundup(tAA_min/tCK)
548 CAS supported
549 Clock(MHz) = 1000 / tCK(ns)
550 Clock(MHz) = 8000 / tCK(MTB)
551 AND BTW: Clock(MT) = 2000 / tCK(ns) - intel uses MTs but calls them MHz
552 */
553 int i;
554
555 /* Calculate common cas_latencies mask, tCKmin and tAAmin. */
556 unsigned int cas_latencies = (unsigned int)-1;
557 unsigned int tCKmin = 0, tAAmin = 0;
558 FOR_EACH_POPULATED_CHANNEL(sysinfo->dimms, i) {
559 cas_latencies &= spdinfo->channel[i].cas_latencies;
560 if (spdinfo->channel[i].tCKmin > tCKmin)
561 tCKmin = spdinfo->channel[i].tCKmin;
562 if (spdinfo->channel[i].tAAmin > tAAmin)
563 tAAmin = spdinfo->channel[i].tAAmin;
564 }
565
566 /* Get actual value of fsb clock. */
567 sysinfo->selected_timings.fsb_clock = read_fsb_clock();
568 unsigned int fsb_mhz = 0;
569 switch (sysinfo->selected_timings.fsb_clock) {
570 case FSB_CLOCK_1067MHz: fsb_mhz = 1067; break;
571 case FSB_CLOCK_800MHz: fsb_mhz = 800; break;
572 case FSB_CLOCK_667MHz: fsb_mhz = 667; break;
573 }
574
575 unsigned int clock = 256000 / tCKmin;
576 const unsigned int max_ddr_clock = (sysinfo->spd_type == DDR2)
577 ? sysinfo->max_ddr2_mt / 2
578 : sysinfo->max_ddr3_mt / 2;
579 if ((clock > max_ddr_clock) || (clock > fsb_mhz / 2)) {
580 int new_clock = MIN(max_ddr_clock, fsb_mhz / 2);
581 printk(BIOS_INFO, "DIMMs support %d MHz, but chipset only runs at up to %d. Limiting...\n",
582 clock, new_clock);
583 clock = new_clock;
584 }
585 normalize_clock(&clock);
586
587 /* Find compatible clock / CAS pair. */
588 unsigned int tCKproposed;
589 unsigned int CAS;
590 while (1) {
591 if (!clock)
592 die("Couldn't find compatible clock / CAS settings.\n");
593 tCKproposed = 256000 / clock;
594 CAS = DIV_ROUND_UP(tAAmin, tCKproposed);
595 printk(BIOS_SPEW, "Trying CAS %u, tCK %u.\n", CAS, tCKproposed);
596 for (; CAS <= DDR3_MAX_CAS; ++CAS)
597 if (cas_latencies & (1 << CAS))
598 break;
599 if ((CAS <= DDR3_MAX_CAS) && (CAS * tCKproposed < 32 * 160)) {
600 /* Found good CAS. */
601 printk(BIOS_SPEW, "Found compatible clock / CAS pair: %u / %u.\n", clock, CAS);
602 break;
603 }
604 lower_clock(&clock);
605 }
606 sysinfo->selected_timings.CAS = CAS;
607 sysinfo->selected_timings.mem_clock = clock_index(clock);
608
609 return tCKproposed;
610 }
611
calculate_derived_timings(sysinfo_t * const sysinfo,const unsigned int tCLK,const spdinfo_t * const spdinfo)612 static void calculate_derived_timings(sysinfo_t *const sysinfo,
613 const unsigned int tCLK,
614 const spdinfo_t *const spdinfo)
615 {
616 int i;
617
618 /* Calculate common tRASmin, tRPmin, tRCDmin and tWRmin. */
619 unsigned int tRASmin = 0, tRPmin = 0, tRCDmin = 0, tWRmin = 0;
620 FOR_EACH_POPULATED_CHANNEL(sysinfo->dimms, i) {
621 if (spdinfo->channel[i].tRAS > tRASmin)
622 tRASmin = spdinfo->channel[i].tRAS;
623 if (spdinfo->channel[i].tRP > tRPmin)
624 tRPmin = spdinfo->channel[i].tRP;
625 if (spdinfo->channel[i].tRCD > tRCDmin)
626 tRCDmin = spdinfo->channel[i].tRCD;
627 if (spdinfo->channel[i].tWR > tWRmin)
628 tWRmin = spdinfo->channel[i].tWR;
629 }
630 tRASmin = DIV_ROUND_UP(tRASmin, tCLK);
631 tRPmin = DIV_ROUND_UP(tRPmin, tCLK);
632 tRCDmin = DIV_ROUND_UP(tRCDmin, tCLK);
633 tWRmin = DIV_ROUND_UP(tWRmin, tCLK);
634
635 /* Lookup tRFC and calculate common tRFCmin. */
636 const unsigned int tRFC_from_clock_and_cap[][4] = {
637 /* CAP_256M CAP_512M CAP_1G CAP_2G */
638 /* 533MHz */ { 40, 56, 68, 104 },
639 /* 400MHz */ { 30, 42, 51, 78 },
640 /* 333MHz */ { 25, 35, 43, 65 },
641 };
642 unsigned int tRFCmin = 0;
643 FOR_EACH_POPULATED_CHANNEL(sysinfo->dimms, i) {
644 const unsigned int tRFC = tRFC_from_clock_and_cap
645 [sysinfo->selected_timings.mem_clock][spdinfo->channel[i].chip_capacity];
646 if (tRFC > tRFCmin)
647 tRFCmin = tRFC;
648 }
649
650 /* Calculate common tRD from CAS and FSB and DRAM clocks. */
651 unsigned int tRDmin = sysinfo->selected_timings.CAS;
652 switch (sysinfo->selected_timings.fsb_clock) {
653 case FSB_CLOCK_667MHz:
654 tRDmin += 1;
655 break;
656 case FSB_CLOCK_800MHz:
657 tRDmin += 2;
658 break;
659 case FSB_CLOCK_1067MHz:
660 tRDmin += 3;
661 if (sysinfo->selected_timings.mem_clock == MEM_CLOCK_1067MT)
662 tRDmin += 1;
663 break;
664 }
665
666 /* Calculate common tRRDmin. */
667 unsigned int tRRDmin = 0;
668 FOR_EACH_POPULATED_CHANNEL(sysinfo->dimms, i) {
669 unsigned int tRRD = 2 + (spdinfo->channel[i].page_size / 1024);
670 if (sysinfo->selected_timings.mem_clock == MEM_CLOCK_1067MT)
671 tRRD += (spdinfo->channel[i].page_size / 1024);
672 if (tRRD > tRRDmin)
673 tRRDmin = tRRD;
674 }
675
676 /* Lookup and calculate common tFAWmin. */
677 unsigned int tFAW_from_pagesize_and_clock[][3] = {
678 /* 533MHz 400MHz 333MHz */
679 /* 1K */ { 20, 15, 13 },
680 /* 2K */ { 27, 20, 17 },
681 };
682 unsigned int tFAWmin = 0;
683 FOR_EACH_POPULATED_CHANNEL(sysinfo->dimms, i) {
684 const unsigned int tFAW = tFAW_from_pagesize_and_clock
685 [spdinfo->channel[i].page_size / 1024 - 1]
686 [sysinfo->selected_timings.mem_clock];
687 if (tFAW > tFAWmin)
688 tFAWmin = tFAW;
689 }
690
691 /* Refresh rate is fixed. */
692 unsigned int tWL;
693 if (sysinfo->spd_type == DDR2) {
694 tWL = sysinfo->selected_timings.CAS - 1;
695 } else if (sysinfo->selected_timings.mem_clock == MEM_CLOCK_1067MT) {
696 tWL = 6;
697 } else {
698 tWL = 5;
699 }
700
701 printk(BIOS_SPEW, "Timing values:\n"
702 " tCLK: %3u\n"
703 " tRAS: %3u\n"
704 " tRP: %3u\n"
705 " tRCD: %3u\n"
706 " tRFC: %3u\n"
707 " tWR: %3u\n"
708 " tRD: %3u\n"
709 " tRRD: %3u\n"
710 " tFAW: %3u\n"
711 " tWL: %3u\n",
712 tCLK, tRASmin, tRPmin, tRCDmin, tRFCmin, tWRmin, tRDmin, tRRDmin, tFAWmin, tWL);
713
714 sysinfo->selected_timings.tRAS = tRASmin;
715 sysinfo->selected_timings.tRP = tRPmin;
716 sysinfo->selected_timings.tRCD = tRCDmin;
717 sysinfo->selected_timings.tRFC = tRFCmin;
718 sysinfo->selected_timings.tWR = tWRmin;
719 sysinfo->selected_timings.tRD = tRDmin;
720 sysinfo->selected_timings.tRRD = tRRDmin;
721 sysinfo->selected_timings.tFAW = tFAWmin;
722 sysinfo->selected_timings.tWL = tWL;
723 }
724
collect_dimm_config(sysinfo_t * const sysinfo)725 static void collect_dimm_config(sysinfo_t *const sysinfo)
726 {
727 int i;
728 spdinfo_t spdinfo;
729
730 spdinfo.dimm_mask = 0;
731 sysinfo->spd_type = 0;
732
733 for (i = 0; i < 4; i++)
734 if (sysinfo->spd_map[i]) {
735 const u8 spd = smbus_read_byte(sysinfo->spd_map[i], 2);
736 printk(BIOS_DEBUG, "%x:%x:%x\n",
737 i, sysinfo->spd_map[i],
738 spd);
739 if ((spd == 7) || (spd == 8) || (spd == 0xb)) {
740 spdinfo.dimm_mask |= 1 << i;
741 if (sysinfo->spd_type && sysinfo->spd_type != spd) {
742 die("Multiple types of DIMM installed in the system, don't do that!\n");
743 }
744 sysinfo->spd_type = spd;
745 }
746 }
747 if (spdinfo.dimm_mask == 0) {
748 die("Could not find any DIMM.\n");
749 }
750
751 /* Normalize spd_type to 1, 2, 3. */
752 sysinfo->spd_type = (sysinfo->spd_type & 1) | ((sysinfo->spd_type & 8) >> 2);
753 printk(BIOS_SPEW, "DDR mask %x, DDR %d\n", spdinfo.dimm_mask, sysinfo->spd_type);
754
755 if (sysinfo->spd_type == DDR2) {
756 verify_ddr2(sysinfo, spdinfo.dimm_mask);
757 collect_ddr2(sysinfo, &spdinfo);
758 } else if (sysinfo->spd_type == DDR3) {
759 verify_ddr3(sysinfo, spdinfo.dimm_mask);
760 collect_ddr3(sysinfo, &spdinfo);
761 } else {
762 die("Will never support DDR1.\n");
763 }
764
765 for (i = 0; i < 2; i++) {
766 if ((spdinfo.dimm_mask >> (i*2)) & 1) {
767 printk(BIOS_SPEW, "Bank %d populated:\n"
768 " Raw card type: %4c\n"
769 " Row addr bits: %4u\n"
770 " Col addr bits: %4u\n"
771 " byte width: %4u\n"
772 " page size: %4u\n"
773 " banks: %4u\n"
774 " ranks: %4u\n"
775 " tAAmin: %3u\n"
776 " tCKmin: %3u\n"
777 " Max clock: %3u MHz\n"
778 " CAS: 0x%04x\n",
779 i, spdinfo.channel[i].raw_card + 'A',
780 spdinfo.channel[i].rows, spdinfo.channel[i].cols,
781 spdinfo.channel[i].width, spdinfo.channel[i].page_size,
782 spdinfo.channel[i].banks, spdinfo.channel[i].ranks,
783 spdinfo.channel[i].tAAmin, spdinfo.channel[i].tCKmin,
784 256000 / spdinfo.channel[i].tCKmin, spdinfo.channel[i].cas_latencies);
785 }
786 }
787
788 FOR_EACH_CHANNEL(i) {
789 sysinfo->dimms[i].card_type =
790 (spdinfo.dimm_mask & (1 << (i * 2))) ? spdinfo.channel[i].raw_card + 0xa : 0;
791 sysinfo->dimms[i].refresh = spdinfo.channel[i].refresh;
792 }
793
794 /* Find common memory clock and CAS. */
795 const unsigned int tCLK = find_common_clock_cas(sysinfo, &spdinfo);
796
797 /* Calculate other timings from clock and CAS. */
798 calculate_derived_timings(sysinfo, tCLK, &spdinfo);
799
800 /* Initialize DIMM infos. */
801 /* Always prefer interleaved over async channel mode. */
802 FOR_EACH_CHANNEL(i) {
803 IF_CHANNEL_POPULATED(sysinfo->dimms, i) {
804 sysinfo->dimms[i].banks = spdinfo.channel[i].banks;
805 sysinfo->dimms[i].ranks = spdinfo.channel[i].ranks;
806
807 /* .width is 1 for x8 or 2 for x16, bus width is 8 bytes. */
808 const unsigned int chips_per_rank = 8 / spdinfo.channel[i].width;
809
810 sysinfo->dimms[i].chip_width = spdinfo.channel[i].width;
811 sysinfo->dimms[i].chip_capacity = spdinfo.channel[i].chip_capacity;
812 sysinfo->dimms[i].page_size = spdinfo.channel[i].page_size * chips_per_rank;
813 sysinfo->dimms[i].rank_capacity_mb =
814 /* offset of chip_capacity is 8 (256M), therefore, add 8
815 chip_capacity is in Mbit, we want MByte, therefore, subtract 3 */
816 (1 << (spdinfo.channel[i].chip_capacity + 8 - 3)) * chips_per_rank;
817 }
818 }
819 if (CHANNEL_IS_POPULATED(sysinfo->dimms, 0) &&
820 CHANNEL_IS_POPULATED(sysinfo->dimms, 1))
821 sysinfo->selected_timings.channel_mode = CHANNEL_MODE_DUAL_INTERLEAVED;
822 else
823 sysinfo->selected_timings.channel_mode = CHANNEL_MODE_SINGLE;
824 }
825
reset_on_bad_warmboot(void)826 static void reset_on_bad_warmboot(void)
827 {
828 /* Check self refresh channel status. */
829 const u32 reg = mchbar_read32(PMSTS_MCHBAR);
830 /* Clear status bits. R/WC */
831 mchbar_write32(PMSTS_MCHBAR, reg);
832 if ((reg & PMSTS_WARM_RESET) && !(reg & PMSTS_BOTH_SELFREFRESH)) {
833 printk(BIOS_INFO, "DRAM was not in self refresh "
834 "during warm boot, reset required.\n");
835 gm45_early_reset();
836 }
837 }
838
set_system_memory_frequency(const timings_t * const timings)839 static void set_system_memory_frequency(const timings_t *const timings)
840 {
841 mchbar_clrbits16(CLKCFG_MCHBAR + 0x60, 1 << 15);
842 mchbar_clrbits16(CLKCFG_MCHBAR + 0x48, 1 << 15);
843
844 /* Calculate wanted frequency setting. */
845 const int want_freq = 6 - timings->mem_clock;
846
847 /* Read current memory frequency. */
848 const u32 clkcfg = mchbar_read32(CLKCFG_MCHBAR);
849 int cur_freq = (clkcfg & CLKCFG_MEMCLK_MASK) >> CLKCFG_MEMCLK_SHIFT;
850 if (0 == cur_freq) {
851 /* Try memory frequency from scratchpad. */
852 printk(BIOS_DEBUG, "Reading current memory frequency from scratchpad.\n");
853 cur_freq = (mchbar_read16(SSKPD_MCHBAR) & SSKPD_CLK_MASK) >> SSKPD_CLK_SHIFT;
854 }
855
856 if (cur_freq != want_freq) {
857 printk(BIOS_DEBUG, "Changing memory frequency: old %x, new %x.\n", cur_freq, want_freq);
858 /* When writing new frequency setting, reset, then set update bit. */
859 mchbar_clrsetbits32(CLKCFG_MCHBAR, CLKCFG_UPDATE | CLKCFG_MEMCLK_MASK,
860 want_freq << CLKCFG_MEMCLK_SHIFT);
861 mchbar_clrsetbits32(CLKCFG_MCHBAR, CLKCFG_MEMCLK_MASK,
862 want_freq << CLKCFG_MEMCLK_SHIFT | CLKCFG_UPDATE);
863 /* Reset update bit. */
864 mchbar_clrbits32(CLKCFG_MCHBAR, CLKCFG_UPDATE);
865 }
866
867 if ((timings->fsb_clock == FSB_CLOCK_1067MHz) && (timings->mem_clock == MEM_CLOCK_667MT)) {
868 mchbar_write32(CLKCFG_MCHBAR + 0x16, 0x000030f0);
869 mchbar_write32(CLKCFG_MCHBAR + 0x64, 0x000050c1);
870
871 mchbar_clrsetbits32(CLKCFG_MCHBAR, 1 << 12, 1 << 17);
872 mchbar_setbits32(CLKCFG_MCHBAR, 1 << 17 | 1 << 12);
873 mchbar_clrbits32(CLKCFG_MCHBAR, 1 << 12);
874
875 mchbar_write32(CLKCFG_MCHBAR + 0x04, 0x9bad1f1f);
876 mchbar_write8(CLKCFG_MCHBAR + 0x08, 0xf4);
877 mchbar_write8(CLKCFG_MCHBAR + 0x0a, 0x43);
878 mchbar_write8(CLKCFG_MCHBAR + 0x0c, 0x10);
879 mchbar_write8(CLKCFG_MCHBAR + 0x0d, 0x80);
880 mchbar_write32(CLKCFG_MCHBAR + 0x50, 0x0b0e151b);
881 mchbar_write8(CLKCFG_MCHBAR + 0x54, 0xb4);
882 mchbar_write8(CLKCFG_MCHBAR + 0x55, 0x10);
883 mchbar_write8(CLKCFG_MCHBAR + 0x56, 0x08);
884
885 mchbar_setbits32(CLKCFG_MCHBAR, 1 << 10);
886 mchbar_setbits32(CLKCFG_MCHBAR, 1 << 11);
887 mchbar_clrbits32(CLKCFG_MCHBAR, 1 << 10);
888 mchbar_clrbits32(CLKCFG_MCHBAR, 1 << 11);
889 }
890
891 mchbar_setbits32(CLKCFG_MCHBAR + 0x48, 0x3f << 24);
892 }
893
raminit_read_vco_index(void)894 int raminit_read_vco_index(void)
895 {
896 switch (mchbar_read8(HPLLVCO_MCHBAR) & 0x7) {
897 case VCO_2666:
898 return 0;
899 case VCO_3200:
900 return 1;
901 case VCO_4000:
902 return 2;
903 case VCO_5333:
904 return 3;
905 default:
906 die("Unknown VCO frequency.\n");
907 return 0;
908 }
909 }
set_igd_memory_frequencies(const sysinfo_t * const sysinfo)910 static void set_igd_memory_frequencies(const sysinfo_t *const sysinfo)
911 {
912 const int gfx_idx = ((sysinfo->gfx_type == GMCH_GS45) &&
913 !sysinfo->gs45_low_power_mode)
914 ? (GMCH_GS45 + 1) : sysinfo->gfx_type;
915
916 /* Render and sampler frequency values seem to be some kind of factor. */
917 const u16 render_freq_from_vco_and_gfxtype[][10] = {
918 /* GM45 GM47 GM49 GE45 GL40 GL43 GS40 GS45 (perf) */
919 /* VCO 2666 */ { 0xd, 0xd, 0xe, 0xd, 0xb, 0xd, 0xb, 0xa, 0xd },
920 /* VCO 3200 */ { 0xd, 0xe, 0xf, 0xd, 0xb, 0xd, 0xb, 0x9, 0xd },
921 /* VCO 4000 */ { 0xc, 0xd, 0xf, 0xc, 0xa, 0xc, 0xa, 0x9, 0xc },
922 /* VCO 5333 */ { 0xb, 0xc, 0xe, 0xb, 0x9, 0xb, 0x9, 0x8, 0xb },
923 };
924 const u16 sampler_freq_from_vco_and_gfxtype[][10] = {
925 /* GM45 GM47 GM49 GE45 GL40 GL43 GS40 GS45 (perf) */
926 /* VCO 2666 */ { 0xc, 0xc, 0xd, 0xc, 0x9, 0xc, 0x9, 0x8, 0xc },
927 /* VCO 3200 */ { 0xc, 0xd, 0xe, 0xc, 0x9, 0xc, 0x9, 0x8, 0xc },
928 /* VCO 4000 */ { 0xa, 0xc, 0xd, 0xa, 0x8, 0xa, 0x8, 0x8, 0xa },
929 /* VCO 5333 */ { 0xa, 0xa, 0xc, 0xa, 0x7, 0xa, 0x7, 0x6, 0xa },
930 };
931 const u16 display_clock_select_from_gfxtype[] = {
932 /* GM45 GM47 GM49 GE45 GL40 GL43 GS40 GS45 (perf) */
933 1, 1, 1, 1, 1, 1, 1, 0, 1
934 };
935
936 if (pci_read_config16(GCFGC_PCIDEV, 0) != 0x8086) {
937 printk(BIOS_DEBUG, "Skipping IGD memory frequency setting.\n");
938 return;
939 }
940
941 mchbar_write16(0x119e, 0xa800);
942 mchbar_clrsetbits16(0x11c0, 0xff << 8, 0x01 << 8);
943 mchbar_write16(0x119e, 0xb800);
944 mchbar_setbits8(0x0f10, 1 << 7);
945
946 /* Read VCO. */
947 const int vco_idx = raminit_read_vco_index();
948 printk(BIOS_DEBUG, "Setting IGD memory frequencies for VCO #%d.\n", vco_idx);
949
950 const u32 freqcfg =
951 ((render_freq_from_vco_and_gfxtype[vco_idx][gfx_idx]
952 << GCFGC_CR_SHIFT) & GCFGC_CR_MASK) |
953 ((sampler_freq_from_vco_and_gfxtype[vco_idx][gfx_idx]
954 << GCFGC_CS_SHIFT) & GCFGC_CS_MASK);
955
956 /* Set frequencies, clear update bit. */
957 u32 gcfgc = pci_read_config16(GCFGC_PCIDEV, GCFGC_OFFSET);
958 gcfgc &= ~(GCFGC_CS_MASK | GCFGC_UPDATE | GCFGC_CR_MASK);
959 gcfgc |= freqcfg;
960 pci_write_config16(GCFGC_PCIDEV, GCFGC_OFFSET, gcfgc);
961
962 /* Set frequencies, set update bit. */
963 gcfgc = pci_read_config16(GCFGC_PCIDEV, GCFGC_OFFSET);
964 gcfgc &= ~(GCFGC_CS_MASK | GCFGC_CR_MASK);
965 gcfgc |= freqcfg | GCFGC_UPDATE;
966 pci_write_config16(GCFGC_PCIDEV, GCFGC_OFFSET, gcfgc);
967
968 /* Clear update bit. */
969 pci_and_config16(GCFGC_PCIDEV, GCFGC_OFFSET, ~GCFGC_UPDATE);
970
971 /* Set display clock select bit. */
972 pci_write_config16(GCFGC_PCIDEV, GCFGC_OFFSET,
973 (pci_read_config16(GCFGC_PCIDEV, GCFGC_OFFSET) & ~GCFGC_CD_MASK) |
974 (display_clock_select_from_gfxtype[gfx_idx] << GCFGC_CD_SHIFT));
975 }
976
configure_dram_control_mode(const timings_t * const timings,const dimminfo_t * const dimms)977 static void configure_dram_control_mode(const timings_t *const timings, const dimminfo_t *const dimms)
978 {
979 int ch, r;
980
981 FOR_EACH_CHANNEL(ch) {
982 unsigned int mchbar = CxDRC0_MCHBAR(ch);
983 u32 cxdrc = mchbar_read32(mchbar);
984 cxdrc &= ~CxDRC0_RANKEN_MASK;
985 FOR_EACH_POPULATED_RANK_IN_CHANNEL(dimms, ch, r)
986 cxdrc |= CxDRC0_RANKEN(r);
987 if (dimms[ch].refresh == REFRESH_3_9)
988 cxdrc = (cxdrc & ~CxDRC0_RMS_MASK) | CxDRC0_RMS_39US;
989 else
990 cxdrc = (cxdrc & ~CxDRC0_RMS_MASK) | CxDRC0_RMS_78US;
991 mchbar_write32(mchbar, cxdrc);
992
993 mchbar = CxDRC1_MCHBAR(ch);
994 cxdrc = mchbar_read32(mchbar);
995 cxdrc |= CxDRC1_NOTPOP_MASK;
996 FOR_EACH_POPULATED_RANK_IN_CHANNEL(dimms, ch, r)
997 cxdrc &= ~CxDRC1_NOTPOP(r);
998 cxdrc |= CxDRC1_MUSTWR;
999 mchbar_write32(mchbar, cxdrc);
1000
1001 mchbar = CxDRC2_MCHBAR(ch);
1002 cxdrc = mchbar_read32(mchbar);
1003 cxdrc |= CxDRC2_NOTPOP_MASK;
1004 FOR_EACH_POPULATED_RANK_IN_CHANNEL(dimms, ch, r)
1005 cxdrc &= ~CxDRC2_NOTPOP(r);
1006 cxdrc |= CxDRC2_MUSTWR;
1007 if (timings->mem_clock == MEM_CLOCK_1067MT)
1008 cxdrc |= CxDRC2_CLK1067MT;
1009 mchbar_write32(mchbar, cxdrc);
1010 }
1011 }
1012
rcomp_initialization(const int spd_type,const stepping_t stepping,const int sff)1013 static void rcomp_initialization(const int spd_type, const stepping_t stepping, const int sff)
1014 {
1015 /* Program RCOMP codes. */
1016 if (sff)
1017 die("SFF platform unsupported in RCOMP initialization.\n");
1018
1019 if (spd_type == DDR2) {
1020 unsigned int o;
1021 for (o = 0; o <= 0x200; o += 0x40) {
1022 mchbar_clrsetbits8(0x6ac + o, 0x0f, 0x0a);
1023 mchbar_write8(0x6b0 + o, 0x55);
1024 }
1025 /* ODT multiplier bits. */
1026 mchbar_clrsetbits32(0x04d0, 7 << 3 | 7 << 0, 1 << 3 | 1 << 0);
1027 } else {
1028 /* Values are for DDR3. */
1029 mchbar_clrbits8(0x6ac, 0x0f);
1030 mchbar_write8(0x6b0, 0x55);
1031 mchbar_clrbits8(0x6ec, 0x0f);
1032 mchbar_write8(0x6f0, 0x66);
1033 mchbar_clrbits8(0x72c, 0x0f);
1034 mchbar_write8(0x730, 0x66);
1035 mchbar_clrbits8(0x76c, 0x0f);
1036 mchbar_write8(0x770, 0x66);
1037 mchbar_clrbits8(0x7ac, 0x0f);
1038 mchbar_write8(0x7b0, 0x66);
1039 mchbar_clrbits8(0x7ec, 0x0f);
1040 mchbar_write8(0x7f0, 0x66);
1041 mchbar_clrbits8(0x86c, 0x0f);
1042 mchbar_write8(0x870, 0x55);
1043 mchbar_clrbits8(0x8ac, 0x0f);
1044 mchbar_write8(0x8b0, 0x66);
1045 /* ODT multiplier bits. */
1046 mchbar_clrsetbits32(0x04d0, 7 << 3 | 7 << 0, 2 << 3 | 2 << 0);
1047 }
1048
1049 /* Perform RCOMP calibration for DDR3. */
1050 raminit_rcomp_calibration(stepping);
1051
1052 /* Run initial RCOMP. */
1053 mchbar_setbits32(0x418, 1 << 17);
1054 mchbar_clrbits32(0x40c, 1 << 23);
1055 mchbar_clrbits32(0x41c, 1 << 7 | 1 << 3);
1056 mchbar_setbits32(0x400, 1);
1057 while (mchbar_read32(0x400) & 1) {}
1058
1059 /* Run second RCOMP. */
1060 mchbar_setbits32(0x40c, 1 << 19);
1061 mchbar_setbits32(0x400, 1);
1062 while (mchbar_read32(0x400) & 1) {}
1063
1064 /* Cleanup and start periodic RCOMP. */
1065 mchbar_clrbits32(0x40c, 1 << 19);
1066 mchbar_setbits32(0x40c, 1 << 23);
1067 mchbar_clrbits32(0x418, 1 << 17);
1068 mchbar_setbits32(0x41c, 1 << 7 | 1 << 3);
1069 mchbar_setbits32(0x400, 1 << 1);
1070 }
1071
dram_powerup(const int spd_type,const int stepping,const int resume)1072 static void dram_powerup(const int spd_type, const int stepping, const int resume)
1073 {
1074 u32 tmp;
1075
1076 udelay(200);
1077 tmp = mchbar_read32(CLKCFG_MCHBAR);
1078 tmp &= ~(3 << 21 | 1 << 3);
1079 if (spd_type == DDR2 && stepping < STEPPING_B0)
1080 tmp |= 2 << 21 | 1 << 3;
1081 else
1082 tmp |= 3 << 21;
1083 mchbar_write32(CLKCFG_MCHBAR, tmp);
1084
1085 if (spd_type == DDR3 && !resume) {
1086 mchbar_setbits32(0x1434, 1 << 10);
1087 udelay(1);
1088 }
1089 mchbar_setbits32(0x1434, 1 << 6);
1090 if (spd_type == DDR3 && !resume) {
1091 udelay(1);
1092 mchbar_setbits32(0x1434, 1 << 9);
1093 mchbar_clrbits32(0x1434, 1 << 10);
1094 udelay(500);
1095 }
1096 }
1097
dram_program_timings(const int spd_type,const timings_t * const timings)1098 static void dram_program_timings(const int spd_type, const timings_t *const timings)
1099 {
1100 /* Values are for DDR3. */
1101 const int burst_length = 8;
1102 const int tWTR = (spd_type == DDR2) ? 3 : 4, tRTP = 1;
1103 int i;
1104
1105 FOR_EACH_CHANNEL(i) {
1106 u32 reg = mchbar_read32(CxDRT0_MCHBAR(i));
1107 const int btb_wtp = timings->tWL + burst_length/2 + timings->tWR;
1108 const int btb_wtr =
1109 ((spd_type == DDR2) ? timings->CAS - 1 : timings->tWL)
1110 + burst_length/2 + tWTR;
1111 reg = (reg & ~(CxDRT0_BtB_WtP_MASK | CxDRT0_BtB_WtR_MASK)) |
1112 ((btb_wtp << CxDRT0_BtB_WtP_SHIFT) & CxDRT0_BtB_WtP_MASK) |
1113 ((btb_wtr << CxDRT0_BtB_WtR_SHIFT) & CxDRT0_BtB_WtR_MASK);
1114 if (spd_type == DDR2) {
1115 reg = (reg & ~(0x7 << 15)) | (2 << 15);
1116 if (timings->mem_clock == MEM_CLOCK_667MT)
1117 reg = (reg & ~(0xf << 10)) | (2 << 10);
1118 else
1119 reg = (reg & ~(0xf << 10)) | (3 << 10);
1120 reg = (reg & ~(0x7 << 5)) | (3 << 5);
1121 } else if (timings->mem_clock != MEM_CLOCK_1067MT) {
1122 reg = (reg & ~(0x7 << 15)) | ((9 - timings->CAS) << 15);
1123 reg = (reg & ~(0xf << 10)) | ((timings->CAS - 3) << 10);
1124 reg = (reg & ~(0x7 << 5)) | (3 << 5);
1125 } else {
1126 reg = (reg & ~(0x7 << 15)) | ((10 - timings->CAS) << 15);
1127 reg = (reg & ~(0xf << 10)) | ((timings->CAS - 4) << 10);
1128 reg = (reg & ~(0x7 << 5)) | (3 << 5);
1129 }
1130 reg = (reg & ~(0x7 << 0)) | (1 << 0);
1131 mchbar_write32(CxDRT0_MCHBAR(i), reg);
1132
1133 reg = mchbar_read32(CxDRT1_MCHBAR(i));
1134 reg = (reg & ~(0x03 << 28)) | ((tRTP & 0x03) << 28);
1135 reg = (reg & ~(0x1f << 21)) | ((timings->tRAS & 0x1f) << 21);
1136 reg = (reg & ~(0x07 << 10)) | (((timings->tRRD - 2) & 0x07) << 10);
1137 reg = (reg & ~(0x07 << 5)) | (((timings->tRCD - 2) & 0x07) << 5);
1138 reg = (reg & ~(0x07 << 0)) | (((timings->tRP - 2) & 0x07) << 0);
1139 mchbar_write32(CxDRT1_MCHBAR(i), reg);
1140
1141 reg = mchbar_read32(CxDRT2_MCHBAR(i));
1142 reg = (reg & ~(0x1f << 17)) | ((timings->tFAW & 0x1f) << 17);
1143 if (spd_type == DDR2) {
1144 reg = (reg & ~(0x7 << 12)) | (0x1 << 12);
1145 reg = (reg & ~(0xf << 6)) | (0x1 << 6);
1146 } else if (timings->mem_clock != MEM_CLOCK_1067MT) {
1147 reg = (reg & ~(0x7 << 12)) | (0x2 << 12);
1148 reg = (reg & ~(0xf << 6)) | (0x9 << 6);
1149 } else {
1150 reg = (reg & ~(0x7 << 12)) | (0x3 << 12);
1151 reg = (reg & ~(0xf << 6)) | (0xc << 6);
1152 }
1153 reg = (reg & ~(0x1f << 0)) | (0x13 << 0);
1154 mchbar_write32(CxDRT2_MCHBAR(i), reg);
1155
1156 reg = mchbar_read32(CxDRT3_MCHBAR(i));
1157 if (spd_type == DDR2)
1158 reg &= ~(0x3 << 28);
1159 else
1160 reg |= (0x3 << 28);
1161 reg = (reg & ~(0x03 << 26));
1162 reg = (reg & ~(0x07 << 23)) | (((timings->CAS - 3) & 0x07) << 23);
1163 reg = (reg & ~(0xff << 13)) | ((timings->tRFC & 0xff) << 13);
1164 reg = (reg & ~(0x07 << 0)) | (((timings->tWL - 2) & 0x07) << 0);
1165 mchbar_write32(CxDRT3_MCHBAR(i), reg);
1166
1167 reg = mchbar_read32(CxDRT4_MCHBAR(i));
1168 static const u8 timings_by_clock[4][3] = {
1169 /* 333MHz 400MHz 533MHz
1170 667MT 800MT 1067MT */
1171 { 0x07, 0x0a, 0x0d },
1172 { 0x3a, 0x46, 0x5d },
1173 { 0x0c, 0x0e, 0x18 },
1174 { 0x21, 0x28, 0x35 },
1175 };
1176 const int clk_idx = 2 - timings->mem_clock;
1177 reg = (reg & ~(0x01f << 27)) | (timings_by_clock[0][clk_idx] << 27);
1178 reg = (reg & ~(0x3ff << 17)) | (timings_by_clock[1][clk_idx] << 17);
1179 reg = (reg & ~(0x03f << 10)) | (timings_by_clock[2][clk_idx] << 10);
1180 reg = (reg & ~(0x1ff << 0)) | (timings_by_clock[3][clk_idx] << 0);
1181 mchbar_write32(CxDRT4_MCHBAR(i), reg);
1182
1183 reg = mchbar_read32(CxDRT5_MCHBAR(i));
1184 if (timings->mem_clock == MEM_CLOCK_1067MT)
1185 reg = (reg & ~(0xf << 28)) | (0x8 << 28);
1186 reg = (reg & ~(0x00f << 22)) | ((burst_length/2 + timings->CAS + 2) << 22);
1187 if (spd_type == DDR2) {
1188 if (timings->mem_clock == MEM_CLOCK_667MT)
1189 reg = (reg & ~(0x1ff << 12)) | (0x21 << 12);
1190 else
1191 reg = (reg & ~(0x1ff << 12)) | (0x28 << 12);
1192 } else {
1193 reg = (reg & ~(0x1ff << 12)) | (0x190 << 12);
1194 }
1195 reg = (reg & ~(0x00f << 4)) | ((timings->CAS - 2) << 4);
1196 reg = (reg & ~(0x003 << 2)) | (0x001 << 2);
1197 reg = (reg & ~(0x003 << 0));
1198 mchbar_write32(CxDRT5_MCHBAR(i), reg);
1199
1200 reg = mchbar_read32(CxDRT6_MCHBAR(i));
1201 if (spd_type == DDR2) {
1202 reg &= ~(1 << 2);
1203 } else {
1204 reg = (reg & ~(0xffff << 16)) | (0x066a << 16); /* always 7.8us refresh rate for DDR3 */
1205 reg |= (1 << 2);
1206 }
1207 mchbar_write32(CxDRT6_MCHBAR(i), reg);
1208 }
1209 }
1210
dram_program_banks(const dimminfo_t * const dimms)1211 static void dram_program_banks(const dimminfo_t *const dimms)
1212 {
1213 int ch, r;
1214
1215 FOR_EACH_CHANNEL(ch) {
1216 const int tRPALL = dimms[ch].banks == 8;
1217
1218 u32 reg = mchbar_read32(CxDRT1_MCHBAR(ch)) & ~(0x01 << 15);
1219 IF_CHANNEL_POPULATED(dimms, ch)
1220 reg |= tRPALL << 15;
1221 mchbar_write32(CxDRT1_MCHBAR(ch), reg);
1222
1223 reg = mchbar_read32(CxDRA_MCHBAR(ch)) & ~CxDRA_BANKS_MASK;
1224 FOR_EACH_POPULATED_RANK_IN_CHANNEL(dimms, ch, r) {
1225 reg |= CxDRA_BANKS(r, dimms[ch].banks);
1226 }
1227 mchbar_write32(CxDRA_MCHBAR(ch), reg);
1228 }
1229 }
1230
ddr3_odt_setup(const timings_t * const timings,const int sff)1231 static void ddr3_odt_setup(const timings_t *const timings, const int sff)
1232 {
1233 int ch;
1234
1235 FOR_EACH_CHANNEL(ch) {
1236 u32 reg = mchbar_read32(CxODT_HIGH(ch));
1237 if (sff && (timings->mem_clock != MEM_CLOCK_1067MT))
1238 reg &= ~(0x3 << (61 - 32));
1239 else
1240 reg |= 0x3 << (61 - 32);
1241 reg = (reg & ~(0x3 << (52 - 32))) | (0x2 << (52 - 32));
1242 reg = (reg & ~(0x7 << (48 - 32))) | ((timings->CAS - 3) << (48 - 32));
1243 reg = (reg & ~(0xf << (44 - 32))) | (0x7 << (44 - 32));
1244 if (timings->mem_clock != MEM_CLOCK_1067MT) {
1245 reg = (reg & ~(0xf << (40 - 32))) | ((12 - timings->CAS) << (40 - 32));
1246 reg = (reg & ~(0xf << (36 - 32))) | (( 2 + timings->CAS) << (36 - 32));
1247 } else {
1248 reg = (reg & ~(0xf << (40 - 32))) | ((13 - timings->CAS) << (40 - 32));
1249 reg = (reg & ~(0xf << (36 - 32))) | (( 1 + timings->CAS) << (36 - 32));
1250 }
1251 reg = (reg & ~(0xf << (32 - 32))) | (0x7 << (32 - 32));
1252 mchbar_write32(CxODT_HIGH(ch), reg);
1253
1254 reg = mchbar_read32(CxODT_LOW(ch));
1255 reg = (reg & ~(0x7 << 28)) | (0x2 << 28);
1256 reg = (reg & ~(0x3 << 22)) | (0x2 << 22);
1257 reg = (reg & ~(0x7 << 12)) | (0x2 << 12);
1258 reg = (reg & ~(0x7 << 4)) | (0x2 << 4);
1259 switch (timings->mem_clock) {
1260 case MEM_CLOCK_667MT:
1261 reg = (reg & ~0x7);
1262 break;
1263 case MEM_CLOCK_800MT:
1264 reg = (reg & ~0x7) | 0x2;
1265 break;
1266 case MEM_CLOCK_1067MT:
1267 reg = (reg & ~0x7) | 0x5;
1268 break;
1269 }
1270 mchbar_write32(CxODT_LOW(ch), reg);
1271 }
1272 }
1273
ddr2_odt_setup(const timings_t * const timings,const int sff)1274 static void ddr2_odt_setup(const timings_t *const timings, const int sff)
1275 {
1276 int ch;
1277
1278 FOR_EACH_CHANNEL(ch) {
1279 u32 reg = mchbar_read32(CxODT_HIGH(ch));
1280 if (sff && (timings->mem_clock == MEM_CLOCK_667MT))
1281 reg &= ~(0x3 << (61 - 32));
1282 else
1283 reg |= 0x3 << (61 - 32);
1284 reg = (reg & ~(0x3 << (52 - 32))) | (1 << (52 - 32));
1285 reg = (reg & ~(0x7 << (48 - 32))) | ((timings->CAS - 2) << (48 - 32));
1286 reg = (reg & ~(0xf << (44 - 32))) | (8 << (44 - 32));
1287 reg = (reg & ~(0xf << (40 - 32))) | (7 << (40 - 32));
1288 if (timings->mem_clock == MEM_CLOCK_667MT) {
1289 reg = (reg & ~(0xf << (36 - 32))) | (4 << (36 - 32));
1290 reg = (reg & ~(0xf << (32 - 32))) | (4 << (32 - 32));
1291 } else {
1292 reg = (reg & ~(0xf << (36 - 32))) | (5 << (36 - 32));
1293 reg = (reg & ~(0xf << (32 - 32))) | (5 << (32 - 32));
1294 }
1295 mchbar_write32(CxODT_HIGH(ch), reg);
1296
1297 reg = mchbar_read32(CxODT_LOW(ch));
1298 if (timings->mem_clock == MEM_CLOCK_667MT)
1299 reg = (reg & ~(0x7 << 28)) | (2 << 28);
1300 else
1301 reg = (reg & ~(0x7 << 28)) | (3 << 28);
1302 reg = (reg & ~(0x3 << 22)) | (1 << 22);
1303 if (timings->mem_clock == MEM_CLOCK_667MT)
1304 reg = (reg & ~(0x7 << 12)) | ((timings->tWL - 1) << 12);
1305 else
1306 reg = (reg & ~(0x7 << 12)) | ((timings->tWL - 2) << 12);
1307 reg = (reg & ~(0x7 << 4)) | ((timings->tWL - 1) << 4);
1308 reg = (reg & ~(0x7 << 0));
1309 mchbar_write32(CxODT_LOW(ch), reg);
1310 }
1311 }
1312
misc_settings(const timings_t * const timings,const stepping_t stepping)1313 static void misc_settings(const timings_t *const timings,
1314 const stepping_t stepping)
1315 {
1316 mchbar_clrsetbits32(0x1260, 1 << 24 | 0x1f, timings->tRD);
1317 mchbar_clrsetbits32(0x1360, 1 << 24 | 0x1f, timings->tRD);
1318
1319 mchbar_clrsetbits8(0x1268, 0xf, timings->tWL);
1320 mchbar_clrsetbits8(0x1368, 0xf, timings->tWL);
1321 mchbar_clrsetbits8(0x12a0, 0xf, 0xa);
1322 mchbar_clrsetbits8(0x13a0, 0xf, 0xa);
1323
1324 mchbar_clrsetbits32(0x218, 7 << 29 | 7 << 25 | 3 << 22 | 3 << 10,
1325 4 << 29 | 3 << 25 | 0 << 22 | 1 << 10);
1326 mchbar_clrsetbits32(0x220, 7 << 16, 1 << 21 | 1 << 16);
1327 mchbar_clrsetbits32(0x224, 7 << 8, 3 << 8);
1328 if (stepping >= STEPPING_B1)
1329 mchbar_setbits8(0x234, 1 << 3);
1330 }
1331
clock_crossing_setup(const fsb_clock_t fsb,const mem_clock_t ddr3clock,const dimminfo_t * const dimms)1332 static void clock_crossing_setup(const fsb_clock_t fsb,
1333 const mem_clock_t ddr3clock,
1334 const dimminfo_t *const dimms)
1335 {
1336 int ch;
1337
1338 static const u32 values_from_fsb_and_mem[][3][4] = {
1339 /* FSB 1067MHz */{
1340 /* DDR3-1067 */ { 0x00000000, 0x00000000, 0x00180006, 0x00810060 },
1341 /* DDR3-800 */ { 0x00000000, 0x00000000, 0x0000001c, 0x000300e0 },
1342 /* DDR3-667 */ { 0x00000000, 0x00001c00, 0x03c00038, 0x0007e000 },
1343 },
1344 /* FSB 800MHz */{
1345 /* DDR3-1067 */ { 0, 0, 0, 0 },
1346 /* DDR3-800 */ { 0x00000000, 0x00000000, 0x0030000c, 0x000300c0 },
1347 /* DDR3-667 */ { 0x00000000, 0x00000380, 0x0060001c, 0x00030c00 },
1348 },
1349 /* FSB 667MHz */{
1350 /* DDR3-1067 */ { 0, 0, 0, 0 },
1351 /* DDR3-800 */ { 0, 0, 0, 0 },
1352 /* DDR3-667 */ { 0x00000000, 0x00000000, 0x0030000c, 0x000300c0 },
1353 },
1354 };
1355
1356 const u32 *data = values_from_fsb_and_mem[fsb][ddr3clock];
1357 mchbar_write32(0x0208, data[3]);
1358 mchbar_write32(0x020c, data[2]);
1359 if (((fsb == FSB_CLOCK_1067MHz) || (fsb == FSB_CLOCK_800MHz)) && (ddr3clock == MEM_CLOCK_667MT))
1360 mchbar_write32(0x0210, data[1]);
1361
1362 static const u32 from_fsb_and_mem[][3] = {
1363 /* DDR3-1067 DDR3-800 DDR3-667 */
1364 /* FSB 1067MHz */{ 0x40100401, 0x10040220, 0x08040110, },
1365 /* FSB 800MHz */{ 0x00000000, 0x40100401, 0x00080201, },
1366 /* FSB 667MHz */{ 0x00000000, 0x00000000, 0x40100401, },
1367 };
1368 FOR_EACH_CHANNEL(ch) {
1369 const unsigned int mchbar = 0x1258 + (ch * 0x0100);
1370 if ((fsb == FSB_CLOCK_1067MHz) && (ddr3clock == MEM_CLOCK_800MT) && CHANNEL_IS_CARDF(dimms, ch))
1371 mchbar_write32(mchbar, 0x08040120);
1372 else
1373 mchbar_write32(mchbar, from_fsb_and_mem[fsb][ddr3clock]);
1374 mchbar_write32(mchbar + 4, 0);
1375 }
1376 }
1377
1378 /* Program egress VC1 isoch timings. */
vc1_program_timings(const fsb_clock_t fsb)1379 static void vc1_program_timings(const fsb_clock_t fsb)
1380 {
1381 const u32 timings_by_fsb[][2] = {
1382 /* FSB 1067MHz */ { 0x1a, 0x01380138 },
1383 /* FSB 800MHz */ { 0x14, 0x00f000f0 },
1384 /* FSB 667MHz */ { 0x10, 0x00c000c0 },
1385 };
1386 epbar_write8(EPVC1ITC, timings_by_fsb[fsb][0]);
1387 epbar_write32(EPVC1IST + 0, timings_by_fsb[fsb][1]);
1388 epbar_write32(EPVC1IST + 4, timings_by_fsb[fsb][1]);
1389 }
1390
1391 #define DEFAULT_PCI_MMIO_SIZE 2048
1392 #define HOST_BRIDGE PCI_DEVFN(0, 0)
1393
get_mmio_size(void)1394 static unsigned int get_mmio_size(void)
1395 {
1396 const struct device *dev;
1397 const struct northbridge_intel_gm45_config *cfg = NULL;
1398
1399 dev = pcidev_path_on_root(HOST_BRIDGE);
1400 if (dev)
1401 cfg = dev->chip_info;
1402
1403 /* If this is zero, it just means devicetree.cb didn't set it */
1404 if (!cfg || cfg->pci_mmio_size == 0)
1405 return DEFAULT_PCI_MMIO_SIZE;
1406 else
1407 return cfg->pci_mmio_size;
1408 }
1409
1410 /* @prejedec if not zero, set rank size to 128MB and page size to 4KB. */
program_memory_map(const dimminfo_t * const dimms,const channel_mode_t mode,const int prejedec,u16 ggc)1411 static void program_memory_map(const dimminfo_t *const dimms, const channel_mode_t mode, const int prejedec, u16 ggc)
1412 {
1413 int ch, r;
1414
1415 /* Program rank boundaries (CxDRBy). */
1416 unsigned int base = 0; /* start of next rank in MB */
1417 unsigned int total_mb[2] = { 0, 0 }; /* total memory per channel in MB */
1418 FOR_EACH_CHANNEL(ch) {
1419 if (mode == CHANNEL_MODE_DUAL_INTERLEAVED)
1420 /* In interleaved mode, start every channel from 0. */
1421 base = 0;
1422 for (r = 0; r < RANKS_PER_CHANNEL; r += 2) {
1423 /* Fixed capacity for pre-jedec config. */
1424 const unsigned int rank_capacity_mb =
1425 prejedec ? 128 : dimms[ch].rank_capacity_mb;
1426 u32 reg = 0;
1427
1428 /* Program bounds in CxDRBy. */
1429 IF_RANK_POPULATED(dimms, ch, r) {
1430 base += rank_capacity_mb;
1431 total_mb[ch] += rank_capacity_mb;
1432 }
1433 reg |= CxDRBy_BOUND_MB(r, base);
1434 IF_RANK_POPULATED(dimms, ch, r+1) {
1435 base += rank_capacity_mb;
1436 total_mb[ch] += rank_capacity_mb;
1437 }
1438 reg |= CxDRBy_BOUND_MB(r+1, base);
1439
1440 mchbar_write32(CxDRBy_MCHBAR(ch, r), reg);
1441 }
1442 }
1443
1444 /* Program page size (CxDRA). */
1445 FOR_EACH_CHANNEL(ch) {
1446 u32 reg = mchbar_read32(CxDRA_MCHBAR(ch)) & ~CxDRA_PAGESIZE_MASK;
1447 FOR_EACH_POPULATED_RANK_IN_CHANNEL(dimms, ch, r) {
1448 /* Fixed page size for pre-jedec config. */
1449 const unsigned int page_size = /* dimm page size in bytes */
1450 prejedec ? 4096 : dimms[ch].page_size;
1451 reg |= CxDRA_PAGESIZE(r, log2(page_size));
1452 /* deferred to f5_27: reg |= CxDRA_BANKS(r, dimms[ch].banks); */
1453 }
1454 mchbar_write32(CxDRA_MCHBAR(ch), reg);
1455 }
1456
1457 /* Calculate memory mapping, all values in MB. */
1458
1459 u32 uma_sizem = 0;
1460 if (!prejedec) {
1461 if (!(ggc & 2)) {
1462 printk(BIOS_DEBUG, "IGD decoded, subtracting ");
1463
1464 /* Graphics memory */
1465 const u32 gms_sizek = decode_igd_memory_size((ggc >> 4) & 0xf);
1466 printk(BIOS_DEBUG, "%uM UMA", gms_sizek >> 10);
1467
1468 /* GTT Graphics Stolen Memory Size (GGMS) */
1469 const u32 gsm_sizek = decode_igd_gtt_size((ggc >> 8) & 0xf);
1470 printk(BIOS_DEBUG, " and %uM GTT\n", gsm_sizek >> 10);
1471
1472 uma_sizem = (gms_sizek + gsm_sizek) >> 10;
1473 }
1474 /* TSEG 2M, This amount can easily be covered by SMRR MTRR's,
1475 which requires to have TSEG_BASE aligned to TSEG_SIZE. */
1476 pci_update_config8(PCI_DEV(0, 0, 0), D0F0_ESMRAMC, ~0x07, (1 << 1) | (1 << 0));
1477 uma_sizem += 2;
1478 }
1479
1480 const unsigned int mmio_size = get_mmio_size();
1481 const unsigned int MMIOstart = 4096 - mmio_size + uma_sizem;
1482 const int me_active = pci_read_config8(PCI_DEV(0, 3, 0), PCI_CLASS_REVISION) != 0xff;
1483 const unsigned int ME_SIZE = prejedec || !me_active ? 0 : 32;
1484 const unsigned int usedMEsize = (total_mb[0] != total_mb[1]) ? ME_SIZE : 2 * ME_SIZE;
1485 const unsigned int claimCapable =
1486 !(pci_read_config32(PCI_DEV(0, 0, 0), D0F0_CAPID0 + 4) & (1 << (47 - 32)));
1487
1488 const unsigned int TOM = total_mb[0] + total_mb[1];
1489 unsigned int TOMminusME = TOM - usedMEsize;
1490 unsigned int TOLUD = (TOMminusME < MMIOstart) ? TOMminusME : MMIOstart;
1491 unsigned int TOUUD = TOMminusME;
1492 unsigned int REMAPbase = 0xffff, REMAPlimit = 0;
1493
1494 if (claimCapable && (TOMminusME >= (MMIOstart + 64))) {
1495 /* 64MB alignment: We'll lose some MBs here, if ME is on. */
1496 TOMminusME &= ~(64 - 1);
1497 /* 64MB alignment: Loss will be reclaimed. */
1498 TOLUD &= ~(64 - 1);
1499 if (TOMminusME > 4096) {
1500 REMAPbase = TOMminusME;
1501 REMAPlimit = REMAPbase + (4096 - TOLUD);
1502 } else {
1503 REMAPbase = 4096;
1504 REMAPlimit = REMAPbase + (TOMminusME - TOLUD);
1505 }
1506 TOUUD = REMAPlimit;
1507 /* REMAPlimit is an inclusive bound, all others exclusive. */
1508 REMAPlimit -= 64;
1509 }
1510
1511 pci_write_config16(PCI_DEV(0, 0, 0), D0F0_TOM, (TOM >> 7) & 0x1ff);
1512 pci_write_config16(PCI_DEV(0, 0, 0), D0F0_TOLUD, TOLUD << 4);
1513 pci_write_config16(PCI_DEV(0, 0, 0), D0F0_TOUUD, TOUUD);
1514 pci_write_config16(PCI_DEV(0, 0, 0), D0F0_REMAPBASE, (REMAPbase >> 6) & 0x03ff);
1515 pci_write_config16(PCI_DEV(0, 0, 0), D0F0_REMAPLIMIT, (REMAPlimit >> 6) & 0x03ff);
1516
1517 /* Program channel mode. */
1518 switch (mode) {
1519 case CHANNEL_MODE_SINGLE:
1520 printk(BIOS_DEBUG, "Memory configured in single-channel mode.\n");
1521 mchbar_clrbits32(DCC_MCHBAR, DCC_INTERLEAVED);
1522 break;
1523 case CHANNEL_MODE_DUAL_ASYNC:
1524 printk(BIOS_DEBUG, "Memory configured in dual-channel asymmetric mode.\n");
1525 mchbar_clrbits32(DCC_MCHBAR, DCC_INTERLEAVED);
1526 break;
1527 case CHANNEL_MODE_DUAL_INTERLEAVED:
1528 printk(BIOS_DEBUG, "Memory configured in dual-channel interleaved mode.\n");
1529 mchbar_clrbits32(DCC_MCHBAR, DCC_NO_CHANXOR | 1 << 9);
1530 mchbar_setbits32(DCC_MCHBAR, DCC_INTERLEAVED);
1531 break;
1532 }
1533
1534 printk(BIOS_SPEW, "Memory map:\n"
1535 "TOM = %5uMB\n"
1536 "TOLUD = %5uMB\n"
1537 "TOUUD = %5uMB\n"
1538 "REMAP:\t base = %5uMB\n"
1539 "\t limit = %5uMB\n"
1540 "usedMEsize: %dMB\n",
1541 TOM, TOLUD, TOUUD, REMAPbase, REMAPlimit, usedMEsize);
1542 }
prejedec_memory_map(const dimminfo_t * const dimms,channel_mode_t mode)1543 static void prejedec_memory_map(const dimminfo_t *const dimms, channel_mode_t mode)
1544 {
1545 /* Never use dual-interleaved mode in pre-jedec config. */
1546 if (CHANNEL_MODE_DUAL_INTERLEAVED == mode)
1547 mode = CHANNEL_MODE_DUAL_ASYNC;
1548
1549 program_memory_map(dimms, mode, 1, 0);
1550 mchbar_setbits32(DCC_MCHBAR, DCC_NO_CHANXOR);
1551 }
1552
ddr3_select_clock_mux(const mem_clock_t ddr3clock,const dimminfo_t * const dimms,const stepping_t stepping)1553 static void ddr3_select_clock_mux(const mem_clock_t ddr3clock,
1554 const dimminfo_t *const dimms,
1555 const stepping_t stepping)
1556 {
1557 const int clk1067 = (ddr3clock == MEM_CLOCK_1067MT);
1558 const int cardF[] = { CHANNEL_IS_CARDF(dimms, 0), CHANNEL_IS_CARDF(dimms, 1) };
1559
1560 int ch;
1561
1562 if (stepping < STEPPING_B1)
1563 die("Stepping <B1 unsupported in clock-multiplexer selection.\n");
1564
1565 FOR_EACH_POPULATED_CHANNEL(dimms, ch) {
1566 int mixed = 0;
1567 if ((1 == ch) && (!CHANNEL_IS_POPULATED(dimms, 0) || (cardF[0] != cardF[1])))
1568 mixed = 4 << 11;
1569 const unsigned int b = 0x14b0 + (ch * 0x0100);
1570 mchbar_write32(b + 0x1c, (mchbar_read32(b + 0x1c) & ~(7 << 11)) |
1571 ((( cardF[ch])?1:0) << 11) | mixed);
1572 mchbar_write32(b + 0x18, (mchbar_read32(b + 0x18) & ~(7 << 11)) | mixed);
1573 mchbar_write32(b + 0x14, (mchbar_read32(b + 0x14) & ~(7 << 11)) |
1574 (((!clk1067 && !cardF[ch])?0:1) << 11) | mixed);
1575 mchbar_write32(b + 0x10, (mchbar_read32(b + 0x10) & ~(7 << 11)) |
1576 ((( clk1067 && !cardF[ch])?1:0) << 11) | mixed);
1577 mchbar_write32(b + 0x0c, (mchbar_read32(b + 0x0c) & ~(7 << 11)) |
1578 ((( cardF[ch])?3:2) << 11) | mixed);
1579 mchbar_write32(b + 0x08, (mchbar_read32(b + 0x08) & ~(7 << 11)) |
1580 (2 << 11) | mixed);
1581 mchbar_write32(b + 0x04, (mchbar_read32(b + 0x04) & ~(7 << 11)) |
1582 (((!clk1067 && !cardF[ch])?2:3) << 11) | mixed);
1583 mchbar_write32(b + 0x00, (mchbar_read32(b + 0x00) & ~(7 << 11)) |
1584 ((( clk1067 && !cardF[ch])?3:2) << 11) | mixed);
1585 }
1586 }
1587
ddr3_write_io_init(const mem_clock_t ddr3clock,const dimminfo_t * const dimms,const stepping_t stepping,const int sff)1588 static void ddr3_write_io_init(const mem_clock_t ddr3clock,
1589 const dimminfo_t *const dimms,
1590 const stepping_t stepping,
1591 const int sff)
1592 {
1593 const int a1step = stepping >= STEPPING_CONVERSION_A1;
1594 const int cardF[] = { CHANNEL_IS_CARDF(dimms, 0), CHANNEL_IS_CARDF(dimms, 1) };
1595
1596 int ch;
1597
1598 if (stepping < STEPPING_B1)
1599 die("Stepping <B1 unsupported in write i/o initialization.\n");
1600 if (sff)
1601 die("SFF platform unsupported in write i/o initialization.\n");
1602
1603 static const u32 ddr3_667_800_by_stepping_ddr3_and_card[][2][2][4] = {
1604 { /* Stepping B3 and below */
1605 { /* 667 MHz */
1606 { 0xa3255008, 0x26888209, 0x26288208, 0x6188040f },
1607 { 0x7524240b, 0xa5255608, 0x232b8508, 0x5528040f },
1608 },
1609 { /* 800 MHz */
1610 { 0xa6255308, 0x26888209, 0x212b7508, 0x6188040f },
1611 { 0x7524240b, 0xa6255708, 0x132b7508, 0x5528040f },
1612 },
1613 },
1614 { /* Conversion stepping A1 and above */
1615 { /* 667 MHz */
1616 { 0xc5257208, 0x26888209, 0x26288208, 0x6188040f },
1617 { 0x7524240b, 0xc5257608, 0x232b8508, 0x5528040f },
1618 },
1619 { /* 800 MHz */
1620 { 0xb6256308, 0x26888209, 0x212b7508, 0x6188040f },
1621 { 0x7524240b, 0xb6256708, 0x132b7508, 0x5528040f },
1622 }
1623 }};
1624
1625 static const u32 ddr3_1067_by_channel_and_card[][2][4] = {
1626 { /* Channel A */
1627 { 0xb2254708, 0x002b7408, 0x132b8008, 0x7228060f },
1628 { 0xb0255008, 0xa4254108, 0x4528b409, 0x9428230f },
1629 },
1630 { /* Channel B */
1631 { 0xa4254208, 0x022b6108, 0x132b8208, 0x9228210f },
1632 { 0x6024140b, 0x92244408, 0x252ba409, 0x9328360c },
1633 },
1634 };
1635
1636 FOR_EACH_POPULATED_CHANNEL(dimms, ch) {
1637 if ((1 == ch) && CHANNEL_IS_POPULATED(dimms, 0) && (cardF[0] == cardF[1]))
1638 /* Only write if second channel population differs. */
1639 continue;
1640 const u32 *const data = (ddr3clock != MEM_CLOCK_1067MT)
1641 ? ddr3_667_800_by_stepping_ddr3_and_card[a1step][2 - ddr3clock][cardF[ch]]
1642 : ddr3_1067_by_channel_and_card[ch][cardF[ch]];
1643 mchbar_write32(CxWRTy_MCHBAR(ch, 0), data[0]);
1644 mchbar_write32(CxWRTy_MCHBAR(ch, 1), data[1]);
1645 mchbar_write32(CxWRTy_MCHBAR(ch, 2), data[2]);
1646 mchbar_write32(CxWRTy_MCHBAR(ch, 3), data[3]);
1647 }
1648
1649 mchbar_write32(0x1490, 0x00e70067);
1650 mchbar_write32(0x1494, 0x000d8000);
1651 mchbar_write32(0x1590, 0x00e70067);
1652 mchbar_write32(0x1594, 0x000d8000);
1653 }
1654
ddr_read_io_init(const mem_clock_t ddr_clock,const dimminfo_t * const dimms,const int sff)1655 static void ddr_read_io_init(const mem_clock_t ddr_clock,
1656 const dimminfo_t *const dimms,
1657 const int sff)
1658 {
1659 int ch;
1660
1661 FOR_EACH_POPULATED_CHANNEL(dimms, ch) {
1662 u32 addr, tmp;
1663 const unsigned int base = 0x14b0 + (ch * 0x0100);
1664 for (addr = base + 0x1c; addr >= base; addr -= 4) {
1665 tmp = mchbar_read32(addr);
1666 tmp &= ~((3 << 25) | (1 << 8) | (7 << 16) | (0xf << 20) | (1 << 27));
1667 tmp |= (1 << 27);
1668 switch (ddr_clock) {
1669 case MEM_CLOCK_667MT:
1670 tmp |= (1 << 16) | (4 << 20);
1671 break;
1672 case MEM_CLOCK_800MT:
1673 tmp |= (2 << 16) | (3 << 20);
1674 break;
1675 case MEM_CLOCK_1067MT:
1676 if (!sff)
1677 tmp |= (2 << 16) | (1 << 20);
1678 else
1679 tmp |= (2 << 16) | (2 << 20);
1680 break;
1681 default:
1682 die("Wrong clock");
1683 }
1684 mchbar_write32(addr, tmp);
1685 }
1686 }
1687 }
1688
ddr3_memory_io_init(const mem_clock_t ddr3clock,const dimminfo_t * const dimms,const stepping_t stepping,const int sff)1689 static void ddr3_memory_io_init(const mem_clock_t ddr3clock,
1690 const dimminfo_t *const dimms,
1691 const stepping_t stepping,
1692 const int sff)
1693 {
1694 u32 tmp;
1695
1696 if (stepping < STEPPING_B1)
1697 die("Stepping <B1 unsupported in "
1698 "system-memory i/o initialization.\n");
1699
1700 tmp = mchbar_read32(0x1400);
1701 tmp &= ~(3<<13);
1702 tmp |= (1<<9) | (1<<13);
1703 mchbar_write32(0x1400, tmp);
1704
1705 tmp = mchbar_read32(0x140c);
1706 tmp &= ~(0xff | (1<<11) | (1<<12) |
1707 (1<<16) | (1<<18) | (1<<27) | (0xf<<28));
1708 tmp |= (1<<7) | (1<<11) | (1<<16);
1709 switch (ddr3clock) {
1710 case MEM_CLOCK_667MT:
1711 tmp |= 9 << 28;
1712 break;
1713 case MEM_CLOCK_800MT:
1714 tmp |= 7 << 28;
1715 break;
1716 case MEM_CLOCK_1067MT:
1717 tmp |= 8 << 28;
1718 break;
1719 }
1720 mchbar_write32(0x140c, tmp);
1721
1722 mchbar_clrbits32(0x1440, 1);
1723
1724 tmp = mchbar_read32(0x1414);
1725 tmp &= ~((1<<20) | (7<<11) | (0xf << 24) | (0xf << 16));
1726 tmp |= (3<<11);
1727 switch (ddr3clock) {
1728 case MEM_CLOCK_667MT:
1729 tmp |= (2 << 24) | (10 << 16);
1730 break;
1731 case MEM_CLOCK_800MT:
1732 tmp |= (3 << 24) | (7 << 16);
1733 break;
1734 case MEM_CLOCK_1067MT:
1735 tmp |= (4 << 24) | (4 << 16);
1736 break;
1737 }
1738 mchbar_write32(0x1414, tmp);
1739
1740 mchbar_clrbits32(0x1418, 1 << 3 | 1 << 11 | 1 << 19 | 1 << 27);
1741
1742 mchbar_clrbits32(0x141c, 1 << 3 | 1 << 11 | 1 << 19 | 1 << 27);
1743
1744 mchbar_setbits32(0x1428, 1 << 14);
1745
1746 tmp = mchbar_read32(0x142c);
1747 tmp &= ~((0xf << 8) | (0x7 << 20) | 0xf | (0xf << 24));
1748 tmp |= (0x3 << 20) | (5 << 24);
1749 switch (ddr3clock) {
1750 case MEM_CLOCK_667MT:
1751 tmp |= (2 << 8) | 0xc;
1752 break;
1753 case MEM_CLOCK_800MT:
1754 tmp |= (3 << 8) | 0xa;
1755 break;
1756 case MEM_CLOCK_1067MT:
1757 tmp |= (4 << 8) | 0x7;
1758 break;
1759 }
1760 mchbar_write32(0x142c, tmp);
1761
1762 tmp = mchbar_read32(0x400);
1763 tmp &= ~((3 << 4) | (3 << 16) | (3 << 30));
1764 tmp |= (2 << 4) | (2 << 16);
1765 mchbar_write32(0x400, tmp);
1766
1767 mchbar_clrbits32(0x404, 0xf << 20);
1768
1769 mchbar_clrbits32(0x40c, 1 << 6);
1770
1771 tmp = mchbar_read32(0x410);
1772 tmp &= ~(7 << 28);
1773 tmp |= 2 << 28;
1774 mchbar_write32(0x410, tmp);
1775
1776 tmp = mchbar_read32(0x41c);
1777 tmp &= ~0x77;
1778 tmp |= 0x11;
1779 mchbar_write32(0x41c, tmp);
1780
1781 ddr3_select_clock_mux(ddr3clock, dimms, stepping);
1782
1783 ddr3_write_io_init(ddr3clock, dimms, stepping, sff);
1784
1785 ddr_read_io_init(ddr3clock, dimms, sff);
1786 }
1787
ddr2_select_clock_mux(const dimminfo_t * const dimms)1788 static void ddr2_select_clock_mux(const dimminfo_t *const dimms)
1789 {
1790 int ch;
1791 unsigned int o;
1792 FOR_EACH_POPULATED_CHANNEL(dimms, ch) {
1793 const unsigned int b = 0x14b0 + (ch * 0x0100);
1794 for (o = 0; o < 0x20; o += 4)
1795 mchbar_clrbits32(b + o, 7 << 11);
1796 }
1797 }
1798
ddr2_write_io_init(const dimminfo_t * const dimms)1799 static void ddr2_write_io_init(const dimminfo_t *const dimms)
1800 {
1801 int s;
1802
1803 mchbar_clrsetbits32(CxWRTy_MCHBAR(0, 0), 0xf7bff71f, 0x008b0008);
1804
1805 for (s = 1; s < 4; ++s) {
1806 mchbar_clrsetbits32(CxWRTy_MCHBAR(0, s), 0xf7bff71f, 0x00800000);
1807 }
1808
1809 mchbar_clrsetbits32(0x1490, 0xf7fff77f, 0x00800000);
1810 mchbar_clrsetbits32(0x1494, 0xf71f8000, 0x00040000);
1811
1812 mchbar_clrsetbits32(CxWRTy_MCHBAR(1, 0), 0xf7bff71f, 0x00890008);
1813
1814 for (s = 1; s < 4; ++s) {
1815 mchbar_clrsetbits32(CxWRTy_MCHBAR(1, s), 0xf7bff71f, 0x00890000);
1816 }
1817
1818 mchbar_clrsetbits32(0x1590, 0xf7fff77f, 0x00800000);
1819 mchbar_clrsetbits32(0x1594, 0xf71f8000, 0x00040000);
1820 }
1821
ddr2_memory_io_init(const mem_clock_t ddr2clock,const dimminfo_t * const dimms,const stepping_t stepping,const int sff)1822 static void ddr2_memory_io_init(const mem_clock_t ddr2clock,
1823 const dimminfo_t *const dimms,
1824 const stepping_t stepping,
1825 const int sff)
1826 {
1827 u32 tmp;
1828 u32 tmp2;
1829
1830 if (stepping < STEPPING_B1)
1831 die("Stepping <B1 unsupported in DDR2 memory i/o initialization.\n");
1832 if (sff)
1833 die("SFF platform unsupported in DDR2 memory i/o initialization.\n");
1834
1835 tmp = mchbar_read32(0x140c);
1836 tmp &= ~(0xff | (1<<11) | (0xf<<28));
1837 tmp |= (1<<0) | (1<<12) | (1<<16) | (1<<18) | (1<<27);
1838 mchbar_write32(0x140c, tmp);
1839
1840 tmp = mchbar_read32(0x1440);
1841 tmp &= ~(1<<5);
1842 tmp |= (1<<0) | (1<<2) | (1<<3) | (1<<4) | (1<<6);
1843 mchbar_write32(0x1440, tmp);
1844
1845 tmp = mchbar_read32(0x1414);
1846 tmp &= ~((1<<20) | (7<<11) | (0xf << 24) | (0xf << 16));
1847 tmp |= (3<<11);
1848 tmp2 = mchbar_read32(0x142c);
1849 tmp2 &= ~((0xf << 8) | (0x7 << 20) | 0xf);
1850 tmp2 |= (0x3 << 20);
1851 switch (ddr2clock) {
1852 case MEM_CLOCK_667MT:
1853 tmp |= (2 << 24) | (10 << 16);
1854 tmp2 |= (2 << 8) | 0xc;
1855 break;
1856 case MEM_CLOCK_800MT:
1857 tmp |= (3 << 24) | (7 << 16);
1858 tmp2 |= (3 << 8) | 0xa;
1859 break;
1860 default:
1861 die("Wrong clock");
1862 }
1863 mchbar_write32(0x1414, tmp);
1864 mchbar_write32(0x142c, tmp2);
1865
1866 mchbar_clrbits32(0x1418, (1<<3) | (1<<11) | (1<<19) | (1<<27));
1867 mchbar_clrbits32(0x141c, (1<<3) | (1<<11) | (1<<19) | (1<<27));
1868
1869 tmp = mchbar_read32(0x400);
1870 tmp &= ~((3 << 4) | (3 << 16) | (3 << 30));
1871 tmp |= (2 << 4) | (2 << 16);
1872 mchbar_write32(0x400, tmp);
1873
1874 mchbar_clrbits32(0x404, 0xf << 20);
1875
1876 mchbar_clrbits32(0x40c, 1 << 6);
1877
1878 tmp = mchbar_read32(0x410);
1879 tmp &= ~(0xf << 28);
1880 tmp |= 2 << 28;
1881 mchbar_write32(0x410, tmp);
1882
1883 tmp = mchbar_read32(0x41c);
1884 tmp &= ~((7<<0) | (7<<4));
1885 tmp |= (1<<0) | (1<<3) | (1<<4) | (1<<7);
1886 mchbar_write32(0x41c, tmp);
1887
1888 ddr2_select_clock_mux(dimms);
1889
1890 ddr2_write_io_init(dimms);
1891
1892 ddr_read_io_init(ddr2clock, dimms, sff);
1893 }
1894
jedec_command(const uintptr_t rankaddr,const u32 cmd,const u32 val)1895 static void jedec_command(const uintptr_t rankaddr, const u32 cmd, const u32 val)
1896 {
1897 mchbar_clrsetbits32(DCC_MCHBAR, DCC_SET_EREG_MASK, cmd);
1898 read32p(rankaddr | val);
1899 }
1900
jedec_init_ddr3(const timings_t * const timings,const dimminfo_t * const dimms)1901 static void jedec_init_ddr3(const timings_t *const timings,
1902 const dimminfo_t *const dimms)
1903 {
1904 if ((timings->tWR < 5) || (timings->tWR > 12))
1905 die("tWR value unsupported in Jedec initialization.\n");
1906
1907
1908
1909 /* 5 6 7 8 9 10 11 12 */
1910 static const u8 wr_lut[] = { 1, 2, 3, 4, 5, 5, 6, 6 };
1911
1912 const int WL = ((timings->tWL - 5) & 7) << 6;
1913 const int ODT_120OHMS = (1 << 9);
1914 const int ODS_34OHMS = (1 << 4);
1915 const int WR = (wr_lut[timings->tWR - 5] & 7) << 12;
1916 const int DLL1 = 1 << 11;
1917 const int CAS = ((timings->CAS - 4) & 7) << 7;
1918 const int INTERLEAVED = 1 << 6;/* This is READ Burst Type == interleaved. */
1919
1920 int ch, r;
1921 FOR_EACH_POPULATED_RANK(dimms, ch, r) {
1922 /* We won't do this in dual-interleaved mode,
1923 so don't care about the offset.
1924 Mirrored ranks aren't taken into account here. */
1925 const uintptr_t rankaddr = raminit_get_rank_addr(ch, r);
1926 printk(BIOS_DEBUG, "JEDEC init @0x%08x\n", (u32)rankaddr);
1927
1928 jedec_command(rankaddr, DCC_SET_EREGx(2), WL);
1929 jedec_command(rankaddr, DCC_SET_EREGx(3), 0);
1930 jedec_command(rankaddr, DCC_SET_EREGx(1), ODT_120OHMS | ODS_34OHMS);
1931 jedec_command(rankaddr, DCC_SET_MREG, WR | DLL1 | CAS | INTERLEAVED);
1932 jedec_command(rankaddr, DCC_SET_MREG, WR | CAS | INTERLEAVED);
1933 }
1934 }
1935
jedec_init_ddr2(const timings_t * const timings,const dimminfo_t * const dimms)1936 static void jedec_init_ddr2(const timings_t *const timings,
1937 const dimminfo_t *const dimms)
1938 {
1939 /* All bit offsets are off by 3 (2^3 bytes bus width). */
1940
1941 /* Mode Register (MR) settings */
1942 const int WR = ((timings->tWR - 1) & 7) << 12;
1943 const int DLLreset = 1 << 11;
1944 const int CAS = (timings->CAS & 7) << 7;
1945 const int BTinterleaved = 1 << 6;
1946 const int BL8 = 3 << 3;
1947
1948 /* Extended Mode Register 1 (EMR1) */
1949 const int OCDdefault = 7 << 10;
1950 const int ODT_150OHMS = 1 << 9 | 0 << 5;
1951
1952 int ch, r;
1953 FOR_EACH_POPULATED_RANK(dimms, ch, r) {
1954 /* We won't do this in dual-interleaved mode,
1955 so don't care about the offset.
1956 Mirrored ranks aren't taken into account here. */
1957 const uintptr_t rankaddr = raminit_get_rank_addr(ch, r);
1958 printk(BIOS_DEBUG, "JEDEC init @0x%08x\n", (u32)rankaddr);
1959
1960 jedec_command(rankaddr, DCC_CMD_ABP, 0);
1961 jedec_command(rankaddr, DCC_SET_EREGx(2), 0);
1962 jedec_command(rankaddr, DCC_SET_EREGx(3), 0);
1963 jedec_command(rankaddr, DCC_SET_EREGx(1), ODT_150OHMS);
1964 jedec_command(rankaddr, DCC_SET_MREG, WR | DLLreset | CAS | BTinterleaved | BL8);
1965 jedec_command(rankaddr, DCC_CMD_ABP, 0);
1966 jedec_command(rankaddr, DCC_CMD_CBR, 0);
1967 udelay(1);
1968 read32p(rankaddr);
1969
1970 jedec_command(rankaddr, DCC_SET_MREG, WR | CAS | BTinterleaved | BL8);
1971 jedec_command(rankaddr, DCC_SET_EREGx(1), OCDdefault | ODT_150OHMS);
1972 jedec_command(rankaddr, DCC_SET_EREGx(1), ODT_150OHMS);
1973 }
1974 }
1975
jedec_init(const int spd_type,const timings_t * const timings,const dimminfo_t * const dimms)1976 static void jedec_init(const int spd_type,
1977 const timings_t *const timings,
1978 const dimminfo_t *const dimms)
1979 {
1980 /* Pre-jedec settings */
1981 mchbar_setbits32(0x40, 1 << 1);
1982 mchbar_setbits32(0x230, 3 << 1);
1983 mchbar_setbits32(0x238, 3 << 24);
1984 mchbar_setbits32(0x23c, 3 << 24);
1985
1986 /* Normal write pointer operation */
1987 mchbar_setbits32(0x14f0, 1 << 9);
1988 mchbar_setbits32(0x15f0, 1 << 9);
1989
1990 mchbar_clrsetbits32(DCC_MCHBAR, DCC_CMD_MASK, DCC_CMD_NOP);
1991
1992 pci_and_config8(PCI_DEV(0, 0, 0), 0xf0, ~(1 << 2));
1993
1994 pci_or_config8(PCI_DEV(0, 0, 0), 0xf0, 1 << 2);
1995 udelay(2);
1996
1997 if (spd_type == DDR2) {
1998 jedec_init_ddr2(timings, dimms);
1999 } else if (spd_type == DDR3) {
2000 jedec_init_ddr3(timings, dimms);
2001 }
2002 }
2003
ddr3_calibrate_zq(void)2004 static void ddr3_calibrate_zq(void) {
2005 udelay(2);
2006
2007 u32 tmp = mchbar_read32(DCC_MCHBAR);
2008 tmp &= ~(7 << 16);
2009 tmp |= (5 << 16); /* ZQ calibration mode */
2010 mchbar_write32(DCC_MCHBAR, tmp);
2011
2012 mchbar_setbits32(CxDRT6_MCHBAR(0), 1 << 3);
2013 mchbar_setbits32(CxDRT6_MCHBAR(1), 1 << 3);
2014
2015 udelay(1);
2016
2017 mchbar_clrbits32(CxDRT6_MCHBAR(0), 1 << 3);
2018 mchbar_clrbits32(CxDRT6_MCHBAR(1), 1 << 3);
2019
2020 mchbar_setbits32(DCC_MCHBAR, 7 << 16); /* Normal operation */
2021 }
2022
post_jedec_sequence(const int cores)2023 static void post_jedec_sequence(const int cores) {
2024 const int quadcore = cores == 4;
2025
2026 mchbar_clrbits32(0x0040, 1 << 1);
2027 mchbar_clrbits32(0x0230, 3 << 1);
2028 mchbar_setbits32(0x0230, 1 << 15);
2029 mchbar_clrbits32(0x0230, 1 << 19);
2030 mchbar_write32(0x1250, 0x6c4);
2031 mchbar_write32(0x1350, 0x6c4);
2032 mchbar_write32(0x1254, 0x871a066d);
2033 mchbar_write32(0x1354, 0x871a066d);
2034 mchbar_setbits32(0x0238, 1 << 26);
2035 mchbar_clrbits32(0x0238, 3 << 24);
2036 mchbar_setbits32(0x0238, 1 << 23);
2037 mchbar_clrsetbits32(0x0238, 7 << 20, 3 << 20);
2038 mchbar_clrsetbits32(0x0238, 7 << 17, 6 << 17);
2039 mchbar_clrsetbits32(0x0238, 7 << 14, 6 << 14);
2040 mchbar_clrsetbits32(0x0238, 7 << 11, 6 << 11);
2041 mchbar_clrsetbits32(0x0238, 7 << 8, 6 << 8);
2042 mchbar_clrbits32(0x023c, 3 << 24);
2043 mchbar_clrbits32(0x023c, 1 << 23);
2044 mchbar_clrsetbits32(0x023c, 7 << 20, 3 << 20);
2045 mchbar_clrsetbits32(0x023c, 7 << 17, 6 << 17);
2046 mchbar_clrsetbits32(0x023c, 7 << 14, 6 << 14);
2047 mchbar_clrsetbits32(0x023c, 7 << 11, 6 << 11);
2048 mchbar_clrsetbits32(0x023c, 7 << 8, 6 << 8);
2049
2050 if (quadcore) {
2051 mchbar_setbits32(0xb14, 0xbfbf << 16);
2052 }
2053 }
2054
dram_optimizations(const timings_t * const timings,const dimminfo_t * const dimms)2055 static void dram_optimizations(const timings_t *const timings,
2056 const dimminfo_t *const dimms)
2057 {
2058 int ch;
2059
2060 FOR_EACH_POPULATED_CHANNEL(dimms, ch) {
2061 const unsigned int mchbar = CxDRC1_MCHBAR(ch);
2062 u32 cxdrc1 = mchbar_read32(mchbar);
2063 cxdrc1 &= ~CxDRC1_SSDS_MASK;
2064 if (dimms[ch].ranks == 1)
2065 cxdrc1 |= CxDRC1_SS;
2066 else
2067 cxdrc1 |= CxDRC1_DS;
2068 mchbar_write32(mchbar, cxdrc1);
2069 }
2070 }
2071
raminit_get_rank_addr(unsigned int channel,unsigned int rank)2072 u32 raminit_get_rank_addr(unsigned int channel, unsigned int rank)
2073 {
2074 if (!channel && !rank)
2075 return 0; /* Address of first rank */
2076
2077 /* Read the bound of the previous rank. */
2078 if (rank > 0) {
2079 rank--;
2080 } else {
2081 rank = 3; /* Highest rank per channel */
2082 channel--;
2083 }
2084 const u32 reg = mchbar_read32(CxDRBy_MCHBAR(channel, rank));
2085 /* Bound is in 32MB. */
2086 return ((reg & CxDRBy_BOUND_MASK(rank)) >> CxDRBy_BOUND_SHIFT(rank)) << 25;
2087 }
2088
raminit_reset_readwrite_pointers(void)2089 void raminit_reset_readwrite_pointers(void)
2090 {
2091 mchbar_setbits32(0x1234, 1 << 6);
2092 mchbar_clrbits32(0x1234, 1 << 6);
2093 mchbar_setbits32(0x1334, 1 << 6);
2094 mchbar_clrbits32(0x1334, 1 << 6);
2095 mchbar_clrbits32(0x14f0, 1 << 9);
2096 mchbar_setbits32(0x14f0, 1 << 9);
2097 mchbar_setbits32(0x14f0, 1 << 10);
2098 mchbar_clrbits32(0x15f0, 1 << 9);
2099 mchbar_setbits32(0x15f0, 1 << 9);
2100 mchbar_setbits32(0x15f0, 1 << 10);
2101 }
2102
raminit(sysinfo_t * const sysinfo,const int s3resume)2103 void raminit(sysinfo_t *const sysinfo, const int s3resume)
2104 {
2105 const dimminfo_t *const dimms = sysinfo->dimms;
2106 const timings_t *const timings = &sysinfo->selected_timings;
2107
2108 int ch;
2109
2110 timestamp_add_now(TS_INITRAM_START);
2111
2112 /* Wait for some bit, maybe TXT clear. */
2113 if (sysinfo->txt_enabled) {
2114 while (!(read8((u8 *)0xfed40000) & (1 << 7))) {}
2115 }
2116
2117 /* Collect information about DIMMs and find common settings. */
2118 collect_dimm_config(sysinfo);
2119
2120 /* Check for bad warm boot. */
2121 reset_on_bad_warmboot();
2122
2123 /***** From now on, program according to collected infos: *****/
2124
2125 /* Program DRAM type. */
2126 switch (sysinfo->spd_type) {
2127 case DDR2:
2128 mchbar_setbits8(0x1434, 1 << 7);
2129 break;
2130 case DDR3:
2131 mchbar_setbits8(0x1434, 3 << 0);
2132 break;
2133 }
2134
2135 /* Program system memory frequency. */
2136 set_system_memory_frequency(timings);
2137 /* Program IGD memory frequency. */
2138 set_igd_memory_frequencies(sysinfo);
2139
2140 /* Configure DRAM control mode for populated channels. */
2141 configure_dram_control_mode(timings, dimms);
2142
2143 /* Initialize RCOMP. */
2144 rcomp_initialization(sysinfo->spd_type, sysinfo->stepping, sysinfo->sff);
2145
2146 /* Power-up DRAM. */
2147 dram_powerup(sysinfo->spd_type, sysinfo->stepping, s3resume);
2148 /* Program DRAM timings. */
2149 dram_program_timings(sysinfo->spd_type, timings);
2150 /* Program number of banks. */
2151 dram_program_banks(dimms);
2152 /* Enable DRAM clock pairs for populated DIMMs. */
2153 FOR_EACH_POPULATED_CHANNEL(dimms, ch)
2154 mchbar_setbits32(CxDCLKDIS_MCHBAR(ch), CxDCLKDIS_ENABLE);
2155
2156 /* Enable On-Die Termination. */
2157 if (sysinfo->spd_type == DDR2)
2158 ddr2_odt_setup(timings, sysinfo->sff);
2159 else
2160 ddr3_odt_setup(timings, sysinfo->sff);
2161 /* Miscellaneous settings. */
2162 misc_settings(timings, sysinfo->stepping);
2163 /* Program clock crossing registers. */
2164 clock_crossing_setup(timings->fsb_clock, timings->mem_clock, dimms);
2165 /* Program egress VC1 timings. */
2166 vc1_program_timings(timings->fsb_clock);
2167 /* Perform system-memory i/o initialization. */
2168 if (sysinfo->spd_type == DDR2) {
2169 ddr2_memory_io_init(timings->mem_clock, dimms,
2170 sysinfo->stepping, sysinfo->sff);
2171 } else {
2172 ddr3_memory_io_init(timings->mem_clock, dimms,
2173 sysinfo->stepping, sysinfo->sff);
2174 }
2175
2176 /* Initialize memory map with dummy values of 128MB per rank with a
2177 page size of 4KB. This makes the JEDEC initialization code easier. */
2178 prejedec_memory_map(dimms, timings->channel_mode);
2179 if (!s3resume)
2180 /* Perform JEDEC initialization of DIMMS. */
2181 jedec_init(sysinfo->spd_type, timings, dimms);
2182 /* Some programming steps after JEDEC initialization. */
2183 post_jedec_sequence(sysinfo->cores);
2184
2185 /* Announce normal operation, initialization completed. */
2186 mchbar_setbits32(DCC_MCHBAR, 0x7 << 16 | 0x1 << 19);
2187
2188 pci_or_config8(PCI_DEV(0, 0, 0), 0xf0, 1 << 2);
2189
2190 pci_and_config8(PCI_DEV(0, 0, 0), 0xf0, ~(1 << 2));
2191
2192 /* Take a breath (the reader). */
2193
2194 /* Perform ZQ calibration for DDR3. */
2195 if (sysinfo->spd_type == DDR3)
2196 ddr3_calibrate_zq();
2197
2198 /* Perform receive-enable calibration. */
2199 raminit_receive_enable_calibration(sysinfo->spd_type, timings, dimms);
2200 /* Lend clock values from receive-enable calibration. */
2201 mchbar_clrsetbits32(CxDRT5_MCHBAR(0), 0xf0,
2202 (((mchbar_read32(CxDRT3_MCHBAR(0)) >> 7) - 1) & 0xf) << 4);
2203 mchbar_clrsetbits32(CxDRT5_MCHBAR(1), 0xf0,
2204 (((mchbar_read32(CxDRT3_MCHBAR(1)) >> 7) - 1) & 0xf) << 4);
2205
2206 /* Perform read/write training for high clock rate. */
2207 if (timings->mem_clock == MEM_CLOCK_1067MT) {
2208 raminit_read_training(dimms, s3resume);
2209 raminit_write_training(timings->mem_clock, dimms, s3resume);
2210 }
2211
2212 igd_compute_ggc(sysinfo);
2213
2214 /* Program final memory map (with real values). */
2215 program_memory_map(dimms, timings->channel_mode, 0, sysinfo->ggc);
2216
2217 /* Some last optimizations. */
2218 dram_optimizations(timings, dimms);
2219
2220 /* Mark raminit being finished. :-) */
2221 pci_and_config8(PCI_DEV(0, 0x1f, 0), 0xa2, (u8)~(1 << 7));
2222
2223 raminit_thermal(sysinfo);
2224 init_igd(sysinfo);
2225
2226 timestamp_add_now(TS_INITRAM_END);
2227 }
2228