Loading...
1/*
2 * MTRR (Memory Type Range Register) cleanup
3 *
4 * Copyright (C) 2009 Yinghai Lu
5 *
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Library General Public
8 * License as published by the Free Software Foundation; either
9 * version 2 of the License, or (at your option) any later version.
10 *
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Library General Public License for more details.
15 *
16 * You should have received a copy of the GNU Library General Public
17 * License along with this library; if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
20#include <linux/module.h>
21#include <linux/init.h>
22#include <linux/pci.h>
23#include <linux/smp.h>
24#include <linux/cpu.h>
25#include <linux/mutex.h>
26#include <linux/uaccess.h>
27#include <linux/kvm_para.h>
28#include <linux/range.h>
29
30#include <asm/processor.h>
31#include <asm/e820.h>
32#include <asm/mtrr.h>
33#include <asm/msr.h>
34
35#include "mtrr.h"
36
37struct var_mtrr_range_state {
38 unsigned long base_pfn;
39 unsigned long size_pfn;
40 mtrr_type type;
41};
42
43struct var_mtrr_state {
44 unsigned long range_startk;
45 unsigned long range_sizek;
46 unsigned long chunk_sizek;
47 unsigned long gran_sizek;
48 unsigned int reg;
49};
50
51/* Should be related to MTRR_VAR_RANGES nums */
52#define RANGE_NUM 256
53
54static struct range __initdata range[RANGE_NUM];
55static int __initdata nr_range;
56
57static struct var_mtrr_range_state __initdata range_state[RANGE_NUM];
58
59static int __initdata debug_print;
60#define Dprintk(x...) do { if (debug_print) pr_debug(x); } while (0)
61
62#define BIOS_BUG_MSG \
63 "WARNING: BIOS bug: VAR MTRR %d contains strange UC entry under 1M, check with your system vendor!\n"
64
65static int __init
66x86_get_mtrr_mem_range(struct range *range, int nr_range,
67 unsigned long extra_remove_base,
68 unsigned long extra_remove_size)
69{
70 unsigned long base, size;
71 mtrr_type type;
72 int i;
73
74 for (i = 0; i < num_var_ranges; i++) {
75 type = range_state[i].type;
76 if (type != MTRR_TYPE_WRBACK)
77 continue;
78 base = range_state[i].base_pfn;
79 size = range_state[i].size_pfn;
80 nr_range = add_range_with_merge(range, RANGE_NUM, nr_range,
81 base, base + size);
82 }
83 if (debug_print) {
84 pr_debug("After WB checking\n");
85 for (i = 0; i < nr_range; i++)
86 pr_debug("MTRR MAP PFN: %016llx - %016llx\n",
87 range[i].start, range[i].end);
88 }
89
90 /* Take out UC ranges: */
91 for (i = 0; i < num_var_ranges; i++) {
92 type = range_state[i].type;
93 if (type != MTRR_TYPE_UNCACHABLE &&
94 type != MTRR_TYPE_WRPROT)
95 continue;
96 size = range_state[i].size_pfn;
97 if (!size)
98 continue;
99 base = range_state[i].base_pfn;
100 if (base < (1<<(20-PAGE_SHIFT)) && mtrr_state.have_fixed &&
101 (mtrr_state.enabled & MTRR_STATE_MTRR_ENABLED) &&
102 (mtrr_state.enabled & MTRR_STATE_MTRR_FIXED_ENABLED)) {
103 /* Var MTRR contains UC entry below 1M? Skip it: */
104 pr_warn(BIOS_BUG_MSG, i);
105 if (base + size <= (1<<(20-PAGE_SHIFT)))
106 continue;
107 size -= (1<<(20-PAGE_SHIFT)) - base;
108 base = 1<<(20-PAGE_SHIFT);
109 }
110 subtract_range(range, RANGE_NUM, base, base + size);
111 }
112 if (extra_remove_size)
113 subtract_range(range, RANGE_NUM, extra_remove_base,
114 extra_remove_base + extra_remove_size);
115
116 if (debug_print) {
117 pr_debug("After UC checking\n");
118 for (i = 0; i < RANGE_NUM; i++) {
119 if (!range[i].end)
120 continue;
121 pr_debug("MTRR MAP PFN: %016llx - %016llx\n",
122 range[i].start, range[i].end);
123 }
124 }
125
126 /* sort the ranges */
127 nr_range = clean_sort_range(range, RANGE_NUM);
128 if (debug_print) {
129 pr_debug("After sorting\n");
130 for (i = 0; i < nr_range; i++)
131 pr_debug("MTRR MAP PFN: %016llx - %016llx\n",
132 range[i].start, range[i].end);
133 }
134
135 return nr_range;
136}
137
138#ifdef CONFIG_MTRR_SANITIZER
139
140static unsigned long __init sum_ranges(struct range *range, int nr_range)
141{
142 unsigned long sum = 0;
143 int i;
144
145 for (i = 0; i < nr_range; i++)
146 sum += range[i].end - range[i].start;
147
148 return sum;
149}
150
151static int enable_mtrr_cleanup __initdata =
152 CONFIG_MTRR_SANITIZER_ENABLE_DEFAULT;
153
154static int __init disable_mtrr_cleanup_setup(char *str)
155{
156 enable_mtrr_cleanup = 0;
157 return 0;
158}
159early_param("disable_mtrr_cleanup", disable_mtrr_cleanup_setup);
160
161static int __init enable_mtrr_cleanup_setup(char *str)
162{
163 enable_mtrr_cleanup = 1;
164 return 0;
165}
166early_param("enable_mtrr_cleanup", enable_mtrr_cleanup_setup);
167
168static int __init mtrr_cleanup_debug_setup(char *str)
169{
170 debug_print = 1;
171 return 0;
172}
173early_param("mtrr_cleanup_debug", mtrr_cleanup_debug_setup);
174
175static void __init
176set_var_mtrr(unsigned int reg, unsigned long basek, unsigned long sizek,
177 unsigned char type, unsigned int address_bits)
178{
179 u32 base_lo, base_hi, mask_lo, mask_hi;
180 u64 base, mask;
181
182 if (!sizek) {
183 fill_mtrr_var_range(reg, 0, 0, 0, 0);
184 return;
185 }
186
187 mask = (1ULL << address_bits) - 1;
188 mask &= ~((((u64)sizek) << 10) - 1);
189
190 base = ((u64)basek) << 10;
191
192 base |= type;
193 mask |= 0x800;
194
195 base_lo = base & ((1ULL<<32) - 1);
196 base_hi = base >> 32;
197
198 mask_lo = mask & ((1ULL<<32) - 1);
199 mask_hi = mask >> 32;
200
201 fill_mtrr_var_range(reg, base_lo, base_hi, mask_lo, mask_hi);
202}
203
204static void __init
205save_var_mtrr(unsigned int reg, unsigned long basek, unsigned long sizek,
206 unsigned char type)
207{
208 range_state[reg].base_pfn = basek >> (PAGE_SHIFT - 10);
209 range_state[reg].size_pfn = sizek >> (PAGE_SHIFT - 10);
210 range_state[reg].type = type;
211}
212
213static void __init set_var_mtrr_all(unsigned int address_bits)
214{
215 unsigned long basek, sizek;
216 unsigned char type;
217 unsigned int reg;
218
219 for (reg = 0; reg < num_var_ranges; reg++) {
220 basek = range_state[reg].base_pfn << (PAGE_SHIFT - 10);
221 sizek = range_state[reg].size_pfn << (PAGE_SHIFT - 10);
222 type = range_state[reg].type;
223
224 set_var_mtrr(reg, basek, sizek, type, address_bits);
225 }
226}
227
228static unsigned long to_size_factor(unsigned long sizek, char *factorp)
229{
230 unsigned long base = sizek;
231 char factor;
232
233 if (base & ((1<<10) - 1)) {
234 /* Not MB-aligned: */
235 factor = 'K';
236 } else if (base & ((1<<20) - 1)) {
237 factor = 'M';
238 base >>= 10;
239 } else {
240 factor = 'G';
241 base >>= 20;
242 }
243
244 *factorp = factor;
245
246 return base;
247}
248
249static unsigned int __init
250range_to_mtrr(unsigned int reg, unsigned long range_startk,
251 unsigned long range_sizek, unsigned char type)
252{
253 if (!range_sizek || (reg >= num_var_ranges))
254 return reg;
255
256 while (range_sizek) {
257 unsigned long max_align, align;
258 unsigned long sizek;
259
260 /* Compute the maximum size with which we can make a range: */
261 if (range_startk)
262 max_align = __ffs(range_startk);
263 else
264 max_align = BITS_PER_LONG - 1;
265
266 align = __fls(range_sizek);
267 if (align > max_align)
268 align = max_align;
269
270 sizek = 1UL << align;
271 if (debug_print) {
272 char start_factor = 'K', size_factor = 'K';
273 unsigned long start_base, size_base;
274
275 start_base = to_size_factor(range_startk, &start_factor);
276 size_base = to_size_factor(sizek, &size_factor);
277
278 Dprintk("Setting variable MTRR %d, "
279 "base: %ld%cB, range: %ld%cB, type %s\n",
280 reg, start_base, start_factor,
281 size_base, size_factor,
282 (type == MTRR_TYPE_UNCACHABLE) ? "UC" :
283 ((type == MTRR_TYPE_WRBACK) ? "WB" : "Other")
284 );
285 }
286 save_var_mtrr(reg++, range_startk, sizek, type);
287 range_startk += sizek;
288 range_sizek -= sizek;
289 if (reg >= num_var_ranges)
290 break;
291 }
292 return reg;
293}
294
295static unsigned __init
296range_to_mtrr_with_hole(struct var_mtrr_state *state, unsigned long basek,
297 unsigned long sizek)
298{
299 unsigned long hole_basek, hole_sizek;
300 unsigned long second_basek, second_sizek;
301 unsigned long range0_basek, range0_sizek;
302 unsigned long range_basek, range_sizek;
303 unsigned long chunk_sizek;
304 unsigned long gran_sizek;
305
306 hole_basek = 0;
307 hole_sizek = 0;
308 second_basek = 0;
309 second_sizek = 0;
310 chunk_sizek = state->chunk_sizek;
311 gran_sizek = state->gran_sizek;
312
313 /* Align with gran size, prevent small block used up MTRRs: */
314 range_basek = ALIGN(state->range_startk, gran_sizek);
315 if ((range_basek > basek) && basek)
316 return second_sizek;
317
318 state->range_sizek -= (range_basek - state->range_startk);
319 range_sizek = ALIGN(state->range_sizek, gran_sizek);
320
321 while (range_sizek > state->range_sizek) {
322 range_sizek -= gran_sizek;
323 if (!range_sizek)
324 return 0;
325 }
326 state->range_sizek = range_sizek;
327
328 /* Try to append some small hole: */
329 range0_basek = state->range_startk;
330 range0_sizek = ALIGN(state->range_sizek, chunk_sizek);
331
332 /* No increase: */
333 if (range0_sizek == state->range_sizek) {
334 Dprintk("rangeX: %016lx - %016lx\n",
335 range0_basek<<10,
336 (range0_basek + state->range_sizek)<<10);
337 state->reg = range_to_mtrr(state->reg, range0_basek,
338 state->range_sizek, MTRR_TYPE_WRBACK);
339 return 0;
340 }
341
342 /* Only cut back when it is not the last: */
343 if (sizek) {
344 while (range0_basek + range0_sizek > (basek + sizek)) {
345 if (range0_sizek >= chunk_sizek)
346 range0_sizek -= chunk_sizek;
347 else
348 range0_sizek = 0;
349
350 if (!range0_sizek)
351 break;
352 }
353 }
354
355second_try:
356 range_basek = range0_basek + range0_sizek;
357
358 /* One hole in the middle: */
359 if (range_basek > basek && range_basek <= (basek + sizek))
360 second_sizek = range_basek - basek;
361
362 if (range0_sizek > state->range_sizek) {
363
364 /* One hole in middle or at the end: */
365 hole_sizek = range0_sizek - state->range_sizek - second_sizek;
366
367 /* Hole size should be less than half of range0 size: */
368 if (hole_sizek >= (range0_sizek >> 1) &&
369 range0_sizek >= chunk_sizek) {
370 range0_sizek -= chunk_sizek;
371 second_sizek = 0;
372 hole_sizek = 0;
373
374 goto second_try;
375 }
376 }
377
378 if (range0_sizek) {
379 Dprintk("range0: %016lx - %016lx\n",
380 range0_basek<<10,
381 (range0_basek + range0_sizek)<<10);
382 state->reg = range_to_mtrr(state->reg, range0_basek,
383 range0_sizek, MTRR_TYPE_WRBACK);
384 }
385
386 if (range0_sizek < state->range_sizek) {
387 /* Need to handle left over range: */
388 range_sizek = state->range_sizek - range0_sizek;
389
390 Dprintk("range: %016lx - %016lx\n",
391 range_basek<<10,
392 (range_basek + range_sizek)<<10);
393
394 state->reg = range_to_mtrr(state->reg, range_basek,
395 range_sizek, MTRR_TYPE_WRBACK);
396 }
397
398 if (hole_sizek) {
399 hole_basek = range_basek - hole_sizek - second_sizek;
400 Dprintk("hole: %016lx - %016lx\n",
401 hole_basek<<10,
402 (hole_basek + hole_sizek)<<10);
403 state->reg = range_to_mtrr(state->reg, hole_basek,
404 hole_sizek, MTRR_TYPE_UNCACHABLE);
405 }
406
407 return second_sizek;
408}
409
410static void __init
411set_var_mtrr_range(struct var_mtrr_state *state, unsigned long base_pfn,
412 unsigned long size_pfn)
413{
414 unsigned long basek, sizek;
415 unsigned long second_sizek = 0;
416
417 if (state->reg >= num_var_ranges)
418 return;
419
420 basek = base_pfn << (PAGE_SHIFT - 10);
421 sizek = size_pfn << (PAGE_SHIFT - 10);
422
423 /* See if I can merge with the last range: */
424 if ((basek <= 1024) ||
425 (state->range_startk + state->range_sizek == basek)) {
426 unsigned long endk = basek + sizek;
427 state->range_sizek = endk - state->range_startk;
428 return;
429 }
430 /* Write the range mtrrs: */
431 if (state->range_sizek != 0)
432 second_sizek = range_to_mtrr_with_hole(state, basek, sizek);
433
434 /* Allocate an msr: */
435 state->range_startk = basek + second_sizek;
436 state->range_sizek = sizek - second_sizek;
437}
438
439/* Mininum size of mtrr block that can take hole: */
440static u64 mtrr_chunk_size __initdata = (256ULL<<20);
441
442static int __init parse_mtrr_chunk_size_opt(char *p)
443{
444 if (!p)
445 return -EINVAL;
446 mtrr_chunk_size = memparse(p, &p);
447 return 0;
448}
449early_param("mtrr_chunk_size", parse_mtrr_chunk_size_opt);
450
451/* Granularity of mtrr of block: */
452static u64 mtrr_gran_size __initdata;
453
454static int __init parse_mtrr_gran_size_opt(char *p)
455{
456 if (!p)
457 return -EINVAL;
458 mtrr_gran_size = memparse(p, &p);
459 return 0;
460}
461early_param("mtrr_gran_size", parse_mtrr_gran_size_opt);
462
463static unsigned long nr_mtrr_spare_reg __initdata =
464 CONFIG_MTRR_SANITIZER_SPARE_REG_NR_DEFAULT;
465
466static int __init parse_mtrr_spare_reg(char *arg)
467{
468 if (arg)
469 nr_mtrr_spare_reg = simple_strtoul(arg, NULL, 0);
470 return 0;
471}
472early_param("mtrr_spare_reg_nr", parse_mtrr_spare_reg);
473
474static int __init
475x86_setup_var_mtrrs(struct range *range, int nr_range,
476 u64 chunk_size, u64 gran_size)
477{
478 struct var_mtrr_state var_state;
479 int num_reg;
480 int i;
481
482 var_state.range_startk = 0;
483 var_state.range_sizek = 0;
484 var_state.reg = 0;
485 var_state.chunk_sizek = chunk_size >> 10;
486 var_state.gran_sizek = gran_size >> 10;
487
488 memset(range_state, 0, sizeof(range_state));
489
490 /* Write the range: */
491 for (i = 0; i < nr_range; i++) {
492 set_var_mtrr_range(&var_state, range[i].start,
493 range[i].end - range[i].start);
494 }
495
496 /* Write the last range: */
497 if (var_state.range_sizek != 0)
498 range_to_mtrr_with_hole(&var_state, 0, 0);
499
500 num_reg = var_state.reg;
501 /* Clear out the extra MTRR's: */
502 while (var_state.reg < num_var_ranges) {
503 save_var_mtrr(var_state.reg, 0, 0, 0);
504 var_state.reg++;
505 }
506
507 return num_reg;
508}
509
510struct mtrr_cleanup_result {
511 unsigned long gran_sizek;
512 unsigned long chunk_sizek;
513 unsigned long lose_cover_sizek;
514 unsigned int num_reg;
515 int bad;
516};
517
518/*
519 * gran_size: 64K, 128K, 256K, 512K, 1M, 2M, ..., 2G
520 * chunk size: gran_size, ..., 2G
521 * so we need (1+16)*8
522 */
523#define NUM_RESULT 136
524#define PSHIFT (PAGE_SHIFT - 10)
525
526static struct mtrr_cleanup_result __initdata result[NUM_RESULT];
527static unsigned long __initdata min_loss_pfn[RANGE_NUM];
528
529static void __init print_out_mtrr_range_state(void)
530{
531 char start_factor = 'K', size_factor = 'K';
532 unsigned long start_base, size_base;
533 mtrr_type type;
534 int i;
535
536 for (i = 0; i < num_var_ranges; i++) {
537
538 size_base = range_state[i].size_pfn << (PAGE_SHIFT - 10);
539 if (!size_base)
540 continue;
541
542 size_base = to_size_factor(size_base, &size_factor),
543 start_base = range_state[i].base_pfn << (PAGE_SHIFT - 10);
544 start_base = to_size_factor(start_base, &start_factor),
545 type = range_state[i].type;
546
547 pr_debug("reg %d, base: %ld%cB, range: %ld%cB, type %s\n",
548 i, start_base, start_factor,
549 size_base, size_factor,
550 (type == MTRR_TYPE_UNCACHABLE) ? "UC" :
551 ((type == MTRR_TYPE_WRPROT) ? "WP" :
552 ((type == MTRR_TYPE_WRBACK) ? "WB" : "Other"))
553 );
554 }
555}
556
557static int __init mtrr_need_cleanup(void)
558{
559 int i;
560 mtrr_type type;
561 unsigned long size;
562 /* Extra one for all 0: */
563 int num[MTRR_NUM_TYPES + 1];
564
565 /* Check entries number: */
566 memset(num, 0, sizeof(num));
567 for (i = 0; i < num_var_ranges; i++) {
568 type = range_state[i].type;
569 size = range_state[i].size_pfn;
570 if (type >= MTRR_NUM_TYPES)
571 continue;
572 if (!size)
573 type = MTRR_NUM_TYPES;
574 num[type]++;
575 }
576
577 /* Check if we got UC entries: */
578 if (!num[MTRR_TYPE_UNCACHABLE])
579 return 0;
580
581 /* Check if we only had WB and UC */
582 if (num[MTRR_TYPE_WRBACK] + num[MTRR_TYPE_UNCACHABLE] !=
583 num_var_ranges - num[MTRR_NUM_TYPES])
584 return 0;
585
586 return 1;
587}
588
589static unsigned long __initdata range_sums;
590
591static void __init
592mtrr_calc_range_state(u64 chunk_size, u64 gran_size,
593 unsigned long x_remove_base,
594 unsigned long x_remove_size, int i)
595{
596 /*
597 * range_new should really be an automatic variable, but
598 * putting 4096 bytes on the stack is frowned upon, to put it
599 * mildly. It is safe to make it a static __initdata variable,
600 * since mtrr_calc_range_state is only called during init and
601 * there's no way it will call itself recursively.
602 */
603 static struct range range_new[RANGE_NUM] __initdata;
604 unsigned long range_sums_new;
605 int nr_range_new;
606 int num_reg;
607
608 /* Convert ranges to var ranges state: */
609 num_reg = x86_setup_var_mtrrs(range, nr_range, chunk_size, gran_size);
610
611 /* We got new setting in range_state, check it: */
612 memset(range_new, 0, sizeof(range_new));
613 nr_range_new = x86_get_mtrr_mem_range(range_new, 0,
614 x_remove_base, x_remove_size);
615 range_sums_new = sum_ranges(range_new, nr_range_new);
616
617 result[i].chunk_sizek = chunk_size >> 10;
618 result[i].gran_sizek = gran_size >> 10;
619 result[i].num_reg = num_reg;
620
621 if (range_sums < range_sums_new) {
622 result[i].lose_cover_sizek = (range_sums_new - range_sums) << PSHIFT;
623 result[i].bad = 1;
624 } else {
625 result[i].lose_cover_sizek = (range_sums - range_sums_new) << PSHIFT;
626 }
627
628 /* Double check it: */
629 if (!result[i].bad && !result[i].lose_cover_sizek) {
630 if (nr_range_new != nr_range || memcmp(range, range_new, sizeof(range)))
631 result[i].bad = 1;
632 }
633
634 if (!result[i].bad && (range_sums - range_sums_new < min_loss_pfn[num_reg]))
635 min_loss_pfn[num_reg] = range_sums - range_sums_new;
636}
637
638static void __init mtrr_print_out_one_result(int i)
639{
640 unsigned long gran_base, chunk_base, lose_base;
641 char gran_factor, chunk_factor, lose_factor;
642
643 gran_base = to_size_factor(result[i].gran_sizek, &gran_factor);
644 chunk_base = to_size_factor(result[i].chunk_sizek, &chunk_factor);
645 lose_base = to_size_factor(result[i].lose_cover_sizek, &lose_factor);
646
647 pr_info("%sgran_size: %ld%c \tchunk_size: %ld%c \t",
648 result[i].bad ? "*BAD*" : " ",
649 gran_base, gran_factor, chunk_base, chunk_factor);
650 pr_cont("num_reg: %d \tlose cover RAM: %s%ld%c\n",
651 result[i].num_reg, result[i].bad ? "-" : "",
652 lose_base, lose_factor);
653}
654
655static int __init mtrr_search_optimal_index(void)
656{
657 int num_reg_good;
658 int index_good;
659 int i;
660
661 if (nr_mtrr_spare_reg >= num_var_ranges)
662 nr_mtrr_spare_reg = num_var_ranges - 1;
663
664 num_reg_good = -1;
665 for (i = num_var_ranges - nr_mtrr_spare_reg; i > 0; i--) {
666 if (!min_loss_pfn[i])
667 num_reg_good = i;
668 }
669
670 index_good = -1;
671 if (num_reg_good != -1) {
672 for (i = 0; i < NUM_RESULT; i++) {
673 if (!result[i].bad &&
674 result[i].num_reg == num_reg_good &&
675 !result[i].lose_cover_sizek) {
676 index_good = i;
677 break;
678 }
679 }
680 }
681
682 return index_good;
683}
684
685int __init mtrr_cleanup(unsigned address_bits)
686{
687 unsigned long x_remove_base, x_remove_size;
688 unsigned long base, size, def, dummy;
689 u64 chunk_size, gran_size;
690 mtrr_type type;
691 int index_good;
692 int i;
693
694 if (!is_cpu(INTEL) || enable_mtrr_cleanup < 1)
695 return 0;
696
697 rdmsr(MSR_MTRRdefType, def, dummy);
698 def &= 0xff;
699 if (def != MTRR_TYPE_UNCACHABLE)
700 return 0;
701
702 /* Get it and store it aside: */
703 memset(range_state, 0, sizeof(range_state));
704 for (i = 0; i < num_var_ranges; i++) {
705 mtrr_if->get(i, &base, &size, &type);
706 range_state[i].base_pfn = base;
707 range_state[i].size_pfn = size;
708 range_state[i].type = type;
709 }
710
711 /* Check if we need handle it and can handle it: */
712 if (!mtrr_need_cleanup())
713 return 0;
714
715 /* Print original var MTRRs at first, for debugging: */
716 pr_debug("original variable MTRRs\n");
717 print_out_mtrr_range_state();
718
719 memset(range, 0, sizeof(range));
720 x_remove_size = 0;
721 x_remove_base = 1 << (32 - PAGE_SHIFT);
722 if (mtrr_tom2)
723 x_remove_size = (mtrr_tom2 >> PAGE_SHIFT) - x_remove_base;
724
725 /*
726 * [0, 1M) should always be covered by var mtrr with WB
727 * and fixed mtrrs should take effect before var mtrr for it:
728 */
729 nr_range = add_range_with_merge(range, RANGE_NUM, 0, 0,
730 1ULL<<(20 - PAGE_SHIFT));
731 /* add from var mtrr at last */
732 nr_range = x86_get_mtrr_mem_range(range, nr_range,
733 x_remove_base, x_remove_size);
734
735 range_sums = sum_ranges(range, nr_range);
736 pr_info("total RAM covered: %ldM\n",
737 range_sums >> (20 - PAGE_SHIFT));
738
739 if (mtrr_chunk_size && mtrr_gran_size) {
740 i = 0;
741 mtrr_calc_range_state(mtrr_chunk_size, mtrr_gran_size,
742 x_remove_base, x_remove_size, i);
743
744 mtrr_print_out_one_result(i);
745
746 if (!result[i].bad) {
747 set_var_mtrr_all(address_bits);
748 pr_debug("New variable MTRRs\n");
749 print_out_mtrr_range_state();
750 return 1;
751 }
752 pr_info("invalid mtrr_gran_size or mtrr_chunk_size, will find optimal one\n");
753 }
754
755 i = 0;
756 memset(min_loss_pfn, 0xff, sizeof(min_loss_pfn));
757 memset(result, 0, sizeof(result));
758 for (gran_size = (1ULL<<16); gran_size < (1ULL<<32); gran_size <<= 1) {
759
760 for (chunk_size = gran_size; chunk_size < (1ULL<<32);
761 chunk_size <<= 1) {
762
763 if (i >= NUM_RESULT)
764 continue;
765
766 mtrr_calc_range_state(chunk_size, gran_size,
767 x_remove_base, x_remove_size, i);
768 if (debug_print) {
769 mtrr_print_out_one_result(i);
770 pr_info("\n");
771 }
772
773 i++;
774 }
775 }
776
777 /* Try to find the optimal index: */
778 index_good = mtrr_search_optimal_index();
779
780 if (index_good != -1) {
781 pr_info("Found optimal setting for mtrr clean up\n");
782 i = index_good;
783 mtrr_print_out_one_result(i);
784
785 /* Convert ranges to var ranges state: */
786 chunk_size = result[i].chunk_sizek;
787 chunk_size <<= 10;
788 gran_size = result[i].gran_sizek;
789 gran_size <<= 10;
790 x86_setup_var_mtrrs(range, nr_range, chunk_size, gran_size);
791 set_var_mtrr_all(address_bits);
792 pr_debug("New variable MTRRs\n");
793 print_out_mtrr_range_state();
794 return 1;
795 } else {
796 /* print out all */
797 for (i = 0; i < NUM_RESULT; i++)
798 mtrr_print_out_one_result(i);
799 }
800
801 pr_info("mtrr_cleanup: can not find optimal value\n");
802 pr_info("please specify mtrr_gran_size/mtrr_chunk_size\n");
803
804 return 0;
805}
806#else
807int __init mtrr_cleanup(unsigned address_bits)
808{
809 return 0;
810}
811#endif
812
813static int disable_mtrr_trim;
814
815static int __init disable_mtrr_trim_setup(char *str)
816{
817 disable_mtrr_trim = 1;
818 return 0;
819}
820early_param("disable_mtrr_trim", disable_mtrr_trim_setup);
821
822/*
823 * Newer AMD K8s and later CPUs have a special magic MSR way to force WB
824 * for memory >4GB. Check for that here.
825 * Note this won't check if the MTRRs < 4GB where the magic bit doesn't
826 * apply to are wrong, but so far we don't know of any such case in the wild.
827 */
828#define Tom2Enabled (1U << 21)
829#define Tom2ForceMemTypeWB (1U << 22)
830
831int __init amd_special_default_mtrr(void)
832{
833 u32 l, h;
834
835 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
836 return 0;
837 if (boot_cpu_data.x86 < 0xf)
838 return 0;
839 /* In case some hypervisor doesn't pass SYSCFG through: */
840 if (rdmsr_safe(MSR_K8_SYSCFG, &l, &h) < 0)
841 return 0;
842 /*
843 * Memory between 4GB and top of mem is forced WB by this magic bit.
844 * Reserved before K8RevF, but should be zero there.
845 */
846 if ((l & (Tom2Enabled | Tom2ForceMemTypeWB)) ==
847 (Tom2Enabled | Tom2ForceMemTypeWB))
848 return 1;
849 return 0;
850}
851
852static u64 __init
853real_trim_memory(unsigned long start_pfn, unsigned long limit_pfn)
854{
855 u64 trim_start, trim_size;
856
857 trim_start = start_pfn;
858 trim_start <<= PAGE_SHIFT;
859
860 trim_size = limit_pfn;
861 trim_size <<= PAGE_SHIFT;
862 trim_size -= trim_start;
863
864 return e820_update_range(trim_start, trim_size, E820_RAM, E820_RESERVED);
865}
866
867/**
868 * mtrr_trim_uncached_memory - trim RAM not covered by MTRRs
869 * @end_pfn: ending page frame number
870 *
871 * Some buggy BIOSes don't setup the MTRRs properly for systems with certain
872 * memory configurations. This routine checks that the highest MTRR matches
873 * the end of memory, to make sure the MTRRs having a write back type cover
874 * all of the memory the kernel is intending to use. If not, it'll trim any
875 * memory off the end by adjusting end_pfn, removing it from the kernel's
876 * allocation pools, warning the user with an obnoxious message.
877 */
878int __init mtrr_trim_uncached_memory(unsigned long end_pfn)
879{
880 unsigned long i, base, size, highest_pfn = 0, def, dummy;
881 mtrr_type type;
882 u64 total_trim_size;
883 /* extra one for all 0 */
884 int num[MTRR_NUM_TYPES + 1];
885
886 /*
887 * Make sure we only trim uncachable memory on machines that
888 * support the Intel MTRR architecture:
889 */
890 if (!is_cpu(INTEL) || disable_mtrr_trim)
891 return 0;
892
893 rdmsr(MSR_MTRRdefType, def, dummy);
894 def &= 0xff;
895 if (def != MTRR_TYPE_UNCACHABLE)
896 return 0;
897
898 /* Get it and store it aside: */
899 memset(range_state, 0, sizeof(range_state));
900 for (i = 0; i < num_var_ranges; i++) {
901 mtrr_if->get(i, &base, &size, &type);
902 range_state[i].base_pfn = base;
903 range_state[i].size_pfn = size;
904 range_state[i].type = type;
905 }
906
907 /* Find highest cached pfn: */
908 for (i = 0; i < num_var_ranges; i++) {
909 type = range_state[i].type;
910 if (type != MTRR_TYPE_WRBACK)
911 continue;
912 base = range_state[i].base_pfn;
913 size = range_state[i].size_pfn;
914 if (highest_pfn < base + size)
915 highest_pfn = base + size;
916 }
917
918 /* kvm/qemu doesn't have mtrr set right, don't trim them all: */
919 if (!highest_pfn) {
920 pr_info("CPU MTRRs all blank - virtualized system.\n");
921 return 0;
922 }
923
924 /* Check entries number: */
925 memset(num, 0, sizeof(num));
926 for (i = 0; i < num_var_ranges; i++) {
927 type = range_state[i].type;
928 if (type >= MTRR_NUM_TYPES)
929 continue;
930 size = range_state[i].size_pfn;
931 if (!size)
932 type = MTRR_NUM_TYPES;
933 num[type]++;
934 }
935
936 /* No entry for WB? */
937 if (!num[MTRR_TYPE_WRBACK])
938 return 0;
939
940 /* Check if we only had WB and UC: */
941 if (num[MTRR_TYPE_WRBACK] + num[MTRR_TYPE_UNCACHABLE] !=
942 num_var_ranges - num[MTRR_NUM_TYPES])
943 return 0;
944
945 memset(range, 0, sizeof(range));
946 nr_range = 0;
947 if (mtrr_tom2) {
948 range[nr_range].start = (1ULL<<(32 - PAGE_SHIFT));
949 range[nr_range].end = mtrr_tom2 >> PAGE_SHIFT;
950 if (highest_pfn < range[nr_range].end)
951 highest_pfn = range[nr_range].end;
952 nr_range++;
953 }
954 nr_range = x86_get_mtrr_mem_range(range, nr_range, 0, 0);
955
956 /* Check the head: */
957 total_trim_size = 0;
958 if (range[0].start)
959 total_trim_size += real_trim_memory(0, range[0].start);
960
961 /* Check the holes: */
962 for (i = 0; i < nr_range - 1; i++) {
963 if (range[i].end < range[i+1].start)
964 total_trim_size += real_trim_memory(range[i].end,
965 range[i+1].start);
966 }
967
968 /* Check the top: */
969 i = nr_range - 1;
970 if (range[i].end < end_pfn)
971 total_trim_size += real_trim_memory(range[i].end,
972 end_pfn);
973
974 if (total_trim_size) {
975 pr_warn("WARNING: BIOS bug: CPU MTRRs don't cover all of memory, losing %lluMB of RAM.\n",
976 total_trim_size >> 20);
977
978 if (!changed_by_mtrr_cleanup)
979 WARN_ON(1);
980
981 pr_info("update e820 for mtrr\n");
982 update_e820();
983
984 return 1;
985 }
986
987 return 0;
988}
1/*
2 * MTRR (Memory Type Range Register) cleanup
3 *
4 * Copyright (C) 2009 Yinghai Lu
5 *
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Library General Public
8 * License as published by the Free Software Foundation; either
9 * version 2 of the License, or (at your option) any later version.
10 *
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Library General Public License for more details.
15 *
16 * You should have received a copy of the GNU Library General Public
17 * License along with this library; if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
20#include <linux/init.h>
21#include <linux/pci.h>
22#include <linux/smp.h>
23#include <linux/cpu.h>
24#include <linux/mutex.h>
25#include <linux/uaccess.h>
26#include <linux/kvm_para.h>
27#include <linux/range.h>
28
29#include <asm/processor.h>
30#include <asm/e820/api.h>
31#include <asm/mtrr.h>
32#include <asm/msr.h>
33
34#include "mtrr.h"
35
36struct var_mtrr_range_state {
37 unsigned long base_pfn;
38 unsigned long size_pfn;
39 mtrr_type type;
40};
41
42struct var_mtrr_state {
43 unsigned long range_startk;
44 unsigned long range_sizek;
45 unsigned long chunk_sizek;
46 unsigned long gran_sizek;
47 unsigned int reg;
48};
49
50/* Should be related to MTRR_VAR_RANGES nums */
51#define RANGE_NUM 256
52
53static struct range __initdata range[RANGE_NUM];
54static int __initdata nr_range;
55
56static struct var_mtrr_range_state __initdata range_state[RANGE_NUM];
57
58#define BIOS_BUG_MSG \
59 "WARNING: BIOS bug: VAR MTRR %d contains strange UC entry under 1M, check with your system vendor!\n"
60
61static int __init
62x86_get_mtrr_mem_range(struct range *range, int nr_range,
63 unsigned long extra_remove_base,
64 unsigned long extra_remove_size)
65{
66 unsigned long base, size;
67 mtrr_type type;
68 int i;
69
70 for (i = 0; i < num_var_ranges; i++) {
71 type = range_state[i].type;
72 if (type != MTRR_TYPE_WRBACK)
73 continue;
74 base = range_state[i].base_pfn;
75 size = range_state[i].size_pfn;
76 nr_range = add_range_with_merge(range, RANGE_NUM, nr_range,
77 base, base + size);
78 }
79
80 Dprintk("After WB checking\n");
81 for (i = 0; i < nr_range; i++)
82 Dprintk("MTRR MAP PFN: %016llx - %016llx\n",
83 range[i].start, range[i].end);
84
85 /* Take out UC ranges: */
86 for (i = 0; i < num_var_ranges; i++) {
87 type = range_state[i].type;
88 if (type != MTRR_TYPE_UNCACHABLE &&
89 type != MTRR_TYPE_WRPROT)
90 continue;
91 size = range_state[i].size_pfn;
92 if (!size)
93 continue;
94 base = range_state[i].base_pfn;
95 if (base < (1<<(20-PAGE_SHIFT)) && mtrr_state.have_fixed &&
96 (mtrr_state.enabled & MTRR_STATE_MTRR_ENABLED) &&
97 (mtrr_state.enabled & MTRR_STATE_MTRR_FIXED_ENABLED)) {
98 /* Var MTRR contains UC entry below 1M? Skip it: */
99 pr_warn(BIOS_BUG_MSG, i);
100 if (base + size <= (1<<(20-PAGE_SHIFT)))
101 continue;
102 size -= (1<<(20-PAGE_SHIFT)) - base;
103 base = 1<<(20-PAGE_SHIFT);
104 }
105 subtract_range(range, RANGE_NUM, base, base + size);
106 }
107 if (extra_remove_size)
108 subtract_range(range, RANGE_NUM, extra_remove_base,
109 extra_remove_base + extra_remove_size);
110
111 Dprintk("After UC checking\n");
112 for (i = 0; i < RANGE_NUM; i++) {
113 if (!range[i].end)
114 continue;
115
116 Dprintk("MTRR MAP PFN: %016llx - %016llx\n",
117 range[i].start, range[i].end);
118 }
119
120 /* sort the ranges */
121 nr_range = clean_sort_range(range, RANGE_NUM);
122
123 Dprintk("After sorting\n");
124 for (i = 0; i < nr_range; i++)
125 Dprintk("MTRR MAP PFN: %016llx - %016llx\n",
126 range[i].start, range[i].end);
127
128 return nr_range;
129}
130
131#ifdef CONFIG_MTRR_SANITIZER
132
133static unsigned long __init sum_ranges(struct range *range, int nr_range)
134{
135 unsigned long sum = 0;
136 int i;
137
138 for (i = 0; i < nr_range; i++)
139 sum += range[i].end - range[i].start;
140
141 return sum;
142}
143
144static int enable_mtrr_cleanup __initdata =
145 CONFIG_MTRR_SANITIZER_ENABLE_DEFAULT;
146
147static int __init disable_mtrr_cleanup_setup(char *str)
148{
149 enable_mtrr_cleanup = 0;
150 return 0;
151}
152early_param("disable_mtrr_cleanup", disable_mtrr_cleanup_setup);
153
154static int __init enable_mtrr_cleanup_setup(char *str)
155{
156 enable_mtrr_cleanup = 1;
157 return 0;
158}
159early_param("enable_mtrr_cleanup", enable_mtrr_cleanup_setup);
160
161static void __init
162set_var_mtrr(unsigned int reg, unsigned long basek, unsigned long sizek,
163 unsigned char type)
164{
165 u32 base_lo, base_hi, mask_lo, mask_hi;
166 u64 base, mask;
167
168 if (!sizek) {
169 fill_mtrr_var_range(reg, 0, 0, 0, 0);
170 return;
171 }
172
173 mask = (1ULL << boot_cpu_data.x86_phys_bits) - 1;
174 mask &= ~((((u64)sizek) << 10) - 1);
175
176 base = ((u64)basek) << 10;
177
178 base |= type;
179 mask |= 0x800;
180
181 base_lo = base & ((1ULL<<32) - 1);
182 base_hi = base >> 32;
183
184 mask_lo = mask & ((1ULL<<32) - 1);
185 mask_hi = mask >> 32;
186
187 fill_mtrr_var_range(reg, base_lo, base_hi, mask_lo, mask_hi);
188}
189
190static void __init
191save_var_mtrr(unsigned int reg, unsigned long basek, unsigned long sizek,
192 unsigned char type)
193{
194 range_state[reg].base_pfn = basek >> (PAGE_SHIFT - 10);
195 range_state[reg].size_pfn = sizek >> (PAGE_SHIFT - 10);
196 range_state[reg].type = type;
197}
198
199static void __init set_var_mtrr_all(void)
200{
201 unsigned long basek, sizek;
202 unsigned char type;
203 unsigned int reg;
204
205 for (reg = 0; reg < num_var_ranges; reg++) {
206 basek = range_state[reg].base_pfn << (PAGE_SHIFT - 10);
207 sizek = range_state[reg].size_pfn << (PAGE_SHIFT - 10);
208 type = range_state[reg].type;
209
210 set_var_mtrr(reg, basek, sizek, type);
211 }
212}
213
214static unsigned long to_size_factor(unsigned long sizek, char *factorp)
215{
216 unsigned long base = sizek;
217 char factor;
218
219 if (base & ((1<<10) - 1)) {
220 /* Not MB-aligned: */
221 factor = 'K';
222 } else if (base & ((1<<20) - 1)) {
223 factor = 'M';
224 base >>= 10;
225 } else {
226 factor = 'G';
227 base >>= 20;
228 }
229
230 *factorp = factor;
231
232 return base;
233}
234
235static unsigned int __init
236range_to_mtrr(unsigned int reg, unsigned long range_startk,
237 unsigned long range_sizek, unsigned char type)
238{
239 if (!range_sizek || (reg >= num_var_ranges))
240 return reg;
241
242 while (range_sizek) {
243 unsigned long max_align, align;
244 unsigned long sizek;
245
246 /* Compute the maximum size with which we can make a range: */
247 if (range_startk)
248 max_align = __ffs(range_startk);
249 else
250 max_align = BITS_PER_LONG - 1;
251
252 align = __fls(range_sizek);
253 if (align > max_align)
254 align = max_align;
255
256 sizek = 1UL << align;
257 if (mtrr_debug) {
258 char start_factor = 'K', size_factor = 'K';
259 unsigned long start_base, size_base;
260
261 start_base = to_size_factor(range_startk, &start_factor);
262 size_base = to_size_factor(sizek, &size_factor);
263
264 Dprintk("Setting variable MTRR %d, "
265 "base: %ld%cB, range: %ld%cB, type %s\n",
266 reg, start_base, start_factor,
267 size_base, size_factor,
268 (type == MTRR_TYPE_UNCACHABLE) ? "UC" :
269 ((type == MTRR_TYPE_WRBACK) ? "WB" : "Other")
270 );
271 }
272 save_var_mtrr(reg++, range_startk, sizek, type);
273 range_startk += sizek;
274 range_sizek -= sizek;
275 if (reg >= num_var_ranges)
276 break;
277 }
278 return reg;
279}
280
281static unsigned __init
282range_to_mtrr_with_hole(struct var_mtrr_state *state, unsigned long basek,
283 unsigned long sizek)
284{
285 unsigned long hole_basek, hole_sizek;
286 unsigned long second_sizek;
287 unsigned long range0_basek, range0_sizek;
288 unsigned long range_basek, range_sizek;
289 unsigned long chunk_sizek;
290 unsigned long gran_sizek;
291
292 hole_basek = 0;
293 hole_sizek = 0;
294 second_sizek = 0;
295 chunk_sizek = state->chunk_sizek;
296 gran_sizek = state->gran_sizek;
297
298 /* Align with gran size, prevent small block used up MTRRs: */
299 range_basek = ALIGN(state->range_startk, gran_sizek);
300 if ((range_basek > basek) && basek)
301 return second_sizek;
302
303 state->range_sizek -= (range_basek - state->range_startk);
304 range_sizek = ALIGN(state->range_sizek, gran_sizek);
305
306 while (range_sizek > state->range_sizek) {
307 range_sizek -= gran_sizek;
308 if (!range_sizek)
309 return 0;
310 }
311 state->range_sizek = range_sizek;
312
313 /* Try to append some small hole: */
314 range0_basek = state->range_startk;
315 range0_sizek = ALIGN(state->range_sizek, chunk_sizek);
316
317 /* No increase: */
318 if (range0_sizek == state->range_sizek) {
319 Dprintk("rangeX: %016lx - %016lx\n",
320 range0_basek<<10,
321 (range0_basek + state->range_sizek)<<10);
322 state->reg = range_to_mtrr(state->reg, range0_basek,
323 state->range_sizek, MTRR_TYPE_WRBACK);
324 return 0;
325 }
326
327 /* Only cut back when it is not the last: */
328 if (sizek) {
329 while (range0_basek + range0_sizek > (basek + sizek)) {
330 if (range0_sizek >= chunk_sizek)
331 range0_sizek -= chunk_sizek;
332 else
333 range0_sizek = 0;
334
335 if (!range0_sizek)
336 break;
337 }
338 }
339
340second_try:
341 range_basek = range0_basek + range0_sizek;
342
343 /* One hole in the middle: */
344 if (range_basek > basek && range_basek <= (basek + sizek))
345 second_sizek = range_basek - basek;
346
347 if (range0_sizek > state->range_sizek) {
348
349 /* One hole in middle or at the end: */
350 hole_sizek = range0_sizek - state->range_sizek - second_sizek;
351
352 /* Hole size should be less than half of range0 size: */
353 if (hole_sizek >= (range0_sizek >> 1) &&
354 range0_sizek >= chunk_sizek) {
355 range0_sizek -= chunk_sizek;
356 second_sizek = 0;
357 hole_sizek = 0;
358
359 goto second_try;
360 }
361 }
362
363 if (range0_sizek) {
364 Dprintk("range0: %016lx - %016lx\n",
365 range0_basek<<10,
366 (range0_basek + range0_sizek)<<10);
367 state->reg = range_to_mtrr(state->reg, range0_basek,
368 range0_sizek, MTRR_TYPE_WRBACK);
369 }
370
371 if (range0_sizek < state->range_sizek) {
372 /* Need to handle left over range: */
373 range_sizek = state->range_sizek - range0_sizek;
374
375 Dprintk("range: %016lx - %016lx\n",
376 range_basek<<10,
377 (range_basek + range_sizek)<<10);
378
379 state->reg = range_to_mtrr(state->reg, range_basek,
380 range_sizek, MTRR_TYPE_WRBACK);
381 }
382
383 if (hole_sizek) {
384 hole_basek = range_basek - hole_sizek - second_sizek;
385 Dprintk("hole: %016lx - %016lx\n",
386 hole_basek<<10,
387 (hole_basek + hole_sizek)<<10);
388 state->reg = range_to_mtrr(state->reg, hole_basek,
389 hole_sizek, MTRR_TYPE_UNCACHABLE);
390 }
391
392 return second_sizek;
393}
394
395static void __init
396set_var_mtrr_range(struct var_mtrr_state *state, unsigned long base_pfn,
397 unsigned long size_pfn)
398{
399 unsigned long basek, sizek;
400 unsigned long second_sizek = 0;
401
402 if (state->reg >= num_var_ranges)
403 return;
404
405 basek = base_pfn << (PAGE_SHIFT - 10);
406 sizek = size_pfn << (PAGE_SHIFT - 10);
407
408 /* See if I can merge with the last range: */
409 if ((basek <= 1024) ||
410 (state->range_startk + state->range_sizek == basek)) {
411 unsigned long endk = basek + sizek;
412 state->range_sizek = endk - state->range_startk;
413 return;
414 }
415 /* Write the range mtrrs: */
416 if (state->range_sizek != 0)
417 second_sizek = range_to_mtrr_with_hole(state, basek, sizek);
418
419 /* Allocate an msr: */
420 state->range_startk = basek + second_sizek;
421 state->range_sizek = sizek - second_sizek;
422}
423
424/* Minimum size of mtrr block that can take hole: */
425static u64 mtrr_chunk_size __initdata = (256ULL<<20);
426
427static int __init parse_mtrr_chunk_size_opt(char *p)
428{
429 if (!p)
430 return -EINVAL;
431 mtrr_chunk_size = memparse(p, &p);
432 return 0;
433}
434early_param("mtrr_chunk_size", parse_mtrr_chunk_size_opt);
435
436/* Granularity of mtrr of block: */
437static u64 mtrr_gran_size __initdata;
438
439static int __init parse_mtrr_gran_size_opt(char *p)
440{
441 if (!p)
442 return -EINVAL;
443 mtrr_gran_size = memparse(p, &p);
444 return 0;
445}
446early_param("mtrr_gran_size", parse_mtrr_gran_size_opt);
447
448static unsigned long nr_mtrr_spare_reg __initdata =
449 CONFIG_MTRR_SANITIZER_SPARE_REG_NR_DEFAULT;
450
451static int __init parse_mtrr_spare_reg(char *arg)
452{
453 if (arg)
454 nr_mtrr_spare_reg = simple_strtoul(arg, NULL, 0);
455 return 0;
456}
457early_param("mtrr_spare_reg_nr", parse_mtrr_spare_reg);
458
459static int __init
460x86_setup_var_mtrrs(struct range *range, int nr_range,
461 u64 chunk_size, u64 gran_size)
462{
463 struct var_mtrr_state var_state;
464 int num_reg;
465 int i;
466
467 var_state.range_startk = 0;
468 var_state.range_sizek = 0;
469 var_state.reg = 0;
470 var_state.chunk_sizek = chunk_size >> 10;
471 var_state.gran_sizek = gran_size >> 10;
472
473 memset(range_state, 0, sizeof(range_state));
474
475 /* Write the range: */
476 for (i = 0; i < nr_range; i++) {
477 set_var_mtrr_range(&var_state, range[i].start,
478 range[i].end - range[i].start);
479 }
480
481 /* Write the last range: */
482 if (var_state.range_sizek != 0)
483 range_to_mtrr_with_hole(&var_state, 0, 0);
484
485 num_reg = var_state.reg;
486 /* Clear out the extra MTRR's: */
487 while (var_state.reg < num_var_ranges) {
488 save_var_mtrr(var_state.reg, 0, 0, 0);
489 var_state.reg++;
490 }
491
492 return num_reg;
493}
494
495struct mtrr_cleanup_result {
496 unsigned long gran_sizek;
497 unsigned long chunk_sizek;
498 unsigned long lose_cover_sizek;
499 unsigned int num_reg;
500 int bad;
501};
502
503/*
504 * gran_size: 64K, 128K, 256K, 512K, 1M, 2M, ..., 2G
505 * chunk size: gran_size, ..., 2G
506 * so we need (1+16)*8
507 */
508#define NUM_RESULT 136
509#define PSHIFT (PAGE_SHIFT - 10)
510
511static struct mtrr_cleanup_result __initdata result[NUM_RESULT];
512static unsigned long __initdata min_loss_pfn[RANGE_NUM];
513
514static void __init print_out_mtrr_range_state(void)
515{
516 char start_factor = 'K', size_factor = 'K';
517 unsigned long start_base, size_base;
518 mtrr_type type;
519 int i;
520
521 for (i = 0; i < num_var_ranges; i++) {
522
523 size_base = range_state[i].size_pfn << (PAGE_SHIFT - 10);
524 if (!size_base)
525 continue;
526
527 size_base = to_size_factor(size_base, &size_factor);
528 start_base = range_state[i].base_pfn << (PAGE_SHIFT - 10);
529 start_base = to_size_factor(start_base, &start_factor);
530 type = range_state[i].type;
531
532 Dprintk("reg %d, base: %ld%cB, range: %ld%cB, type %s\n",
533 i, start_base, start_factor,
534 size_base, size_factor,
535 (type == MTRR_TYPE_UNCACHABLE) ? "UC" :
536 ((type == MTRR_TYPE_WRPROT) ? "WP" :
537 ((type == MTRR_TYPE_WRBACK) ? "WB" : "Other"))
538 );
539 }
540}
541
542static int __init mtrr_need_cleanup(void)
543{
544 int i;
545 mtrr_type type;
546 unsigned long size;
547 /* Extra one for all 0: */
548 int num[MTRR_NUM_TYPES + 1];
549
550 /* Check entries number: */
551 memset(num, 0, sizeof(num));
552 for (i = 0; i < num_var_ranges; i++) {
553 type = range_state[i].type;
554 size = range_state[i].size_pfn;
555 if (type >= MTRR_NUM_TYPES)
556 continue;
557 if (!size)
558 type = MTRR_NUM_TYPES;
559 num[type]++;
560 }
561
562 /* Check if we got UC entries: */
563 if (!num[MTRR_TYPE_UNCACHABLE])
564 return 0;
565
566 /* Check if we only had WB and UC */
567 if (num[MTRR_TYPE_WRBACK] + num[MTRR_TYPE_UNCACHABLE] !=
568 num_var_ranges - num[MTRR_NUM_TYPES])
569 return 0;
570
571 return 1;
572}
573
574static unsigned long __initdata range_sums;
575
576static void __init
577mtrr_calc_range_state(u64 chunk_size, u64 gran_size,
578 unsigned long x_remove_base,
579 unsigned long x_remove_size, int i)
580{
581 /*
582 * range_new should really be an automatic variable, but
583 * putting 4096 bytes on the stack is frowned upon, to put it
584 * mildly. It is safe to make it a static __initdata variable,
585 * since mtrr_calc_range_state is only called during init and
586 * there's no way it will call itself recursively.
587 */
588 static struct range range_new[RANGE_NUM] __initdata;
589 unsigned long range_sums_new;
590 int nr_range_new;
591 int num_reg;
592
593 /* Convert ranges to var ranges state: */
594 num_reg = x86_setup_var_mtrrs(range, nr_range, chunk_size, gran_size);
595
596 /* We got new setting in range_state, check it: */
597 memset(range_new, 0, sizeof(range_new));
598 nr_range_new = x86_get_mtrr_mem_range(range_new, 0,
599 x_remove_base, x_remove_size);
600 range_sums_new = sum_ranges(range_new, nr_range_new);
601
602 result[i].chunk_sizek = chunk_size >> 10;
603 result[i].gran_sizek = gran_size >> 10;
604 result[i].num_reg = num_reg;
605
606 if (range_sums < range_sums_new) {
607 result[i].lose_cover_sizek = (range_sums_new - range_sums) << PSHIFT;
608 result[i].bad = 1;
609 } else {
610 result[i].lose_cover_sizek = (range_sums - range_sums_new) << PSHIFT;
611 }
612
613 /* Double check it: */
614 if (!result[i].bad && !result[i].lose_cover_sizek) {
615 if (nr_range_new != nr_range || memcmp(range, range_new, sizeof(range)))
616 result[i].bad = 1;
617 }
618
619 if (!result[i].bad && (range_sums - range_sums_new < min_loss_pfn[num_reg]))
620 min_loss_pfn[num_reg] = range_sums - range_sums_new;
621}
622
623static void __init mtrr_print_out_one_result(int i)
624{
625 unsigned long gran_base, chunk_base, lose_base;
626 char gran_factor, chunk_factor, lose_factor;
627
628 gran_base = to_size_factor(result[i].gran_sizek, &gran_factor);
629 chunk_base = to_size_factor(result[i].chunk_sizek, &chunk_factor);
630 lose_base = to_size_factor(result[i].lose_cover_sizek, &lose_factor);
631
632 pr_info("%sgran_size: %ld%c \tchunk_size: %ld%c \t",
633 result[i].bad ? "*BAD*" : " ",
634 gran_base, gran_factor, chunk_base, chunk_factor);
635 pr_cont("num_reg: %d \tlose cover RAM: %s%ld%c\n",
636 result[i].num_reg, result[i].bad ? "-" : "",
637 lose_base, lose_factor);
638}
639
640static int __init mtrr_search_optimal_index(void)
641{
642 int num_reg_good;
643 int index_good;
644 int i;
645
646 if (nr_mtrr_spare_reg >= num_var_ranges)
647 nr_mtrr_spare_reg = num_var_ranges - 1;
648
649 num_reg_good = -1;
650 for (i = num_var_ranges - nr_mtrr_spare_reg; i > 0; i--) {
651 if (!min_loss_pfn[i])
652 num_reg_good = i;
653 }
654
655 index_good = -1;
656 if (num_reg_good != -1) {
657 for (i = 0; i < NUM_RESULT; i++) {
658 if (!result[i].bad &&
659 result[i].num_reg == num_reg_good &&
660 !result[i].lose_cover_sizek) {
661 index_good = i;
662 break;
663 }
664 }
665 }
666
667 return index_good;
668}
669
670int __init mtrr_cleanup(void)
671{
672 unsigned long x_remove_base, x_remove_size;
673 unsigned long base, size, def, dummy;
674 u64 chunk_size, gran_size;
675 mtrr_type type;
676 int index_good;
677 int i;
678
679 if (!mtrr_enabled())
680 return 0;
681
682 if (!cpu_feature_enabled(X86_FEATURE_MTRR) || enable_mtrr_cleanup < 1)
683 return 0;
684
685 rdmsr(MSR_MTRRdefType, def, dummy);
686 def &= 0xff;
687 if (def != MTRR_TYPE_UNCACHABLE)
688 return 0;
689
690 /* Get it and store it aside: */
691 memset(range_state, 0, sizeof(range_state));
692 for (i = 0; i < num_var_ranges; i++) {
693 mtrr_if->get(i, &base, &size, &type);
694 range_state[i].base_pfn = base;
695 range_state[i].size_pfn = size;
696 range_state[i].type = type;
697 }
698
699 /* Check if we need handle it and can handle it: */
700 if (!mtrr_need_cleanup())
701 return 0;
702
703 /* Print original var MTRRs at first, for debugging: */
704 Dprintk("original variable MTRRs\n");
705 print_out_mtrr_range_state();
706
707 memset(range, 0, sizeof(range));
708 x_remove_size = 0;
709 x_remove_base = 1 << (32 - PAGE_SHIFT);
710 if (mtrr_tom2)
711 x_remove_size = (mtrr_tom2 >> PAGE_SHIFT) - x_remove_base;
712
713 /*
714 * [0, 1M) should always be covered by var mtrr with WB
715 * and fixed mtrrs should take effect before var mtrr for it:
716 */
717 nr_range = add_range_with_merge(range, RANGE_NUM, 0, 0,
718 1ULL<<(20 - PAGE_SHIFT));
719 /* add from var mtrr at last */
720 nr_range = x86_get_mtrr_mem_range(range, nr_range,
721 x_remove_base, x_remove_size);
722
723 range_sums = sum_ranges(range, nr_range);
724 pr_info("total RAM covered: %ldM\n",
725 range_sums >> (20 - PAGE_SHIFT));
726
727 if (mtrr_chunk_size && mtrr_gran_size) {
728 i = 0;
729 mtrr_calc_range_state(mtrr_chunk_size, mtrr_gran_size,
730 x_remove_base, x_remove_size, i);
731
732 mtrr_print_out_one_result(i);
733
734 if (!result[i].bad) {
735 set_var_mtrr_all();
736 Dprintk("New variable MTRRs\n");
737 print_out_mtrr_range_state();
738 return 1;
739 }
740 pr_info("invalid mtrr_gran_size or mtrr_chunk_size, will find optimal one\n");
741 }
742
743 i = 0;
744 memset(min_loss_pfn, 0xff, sizeof(min_loss_pfn));
745 memset(result, 0, sizeof(result));
746 for (gran_size = (1ULL<<16); gran_size < (1ULL<<32); gran_size <<= 1) {
747
748 for (chunk_size = gran_size; chunk_size < (1ULL<<32);
749 chunk_size <<= 1) {
750
751 if (i >= NUM_RESULT)
752 continue;
753
754 mtrr_calc_range_state(chunk_size, gran_size,
755 x_remove_base, x_remove_size, i);
756 if (mtrr_debug) {
757 mtrr_print_out_one_result(i);
758 pr_info("\n");
759 }
760
761 i++;
762 }
763 }
764
765 /* Try to find the optimal index: */
766 index_good = mtrr_search_optimal_index();
767
768 if (index_good != -1) {
769 pr_info("Found optimal setting for mtrr clean up\n");
770 i = index_good;
771 mtrr_print_out_one_result(i);
772
773 /* Convert ranges to var ranges state: */
774 chunk_size = result[i].chunk_sizek;
775 chunk_size <<= 10;
776 gran_size = result[i].gran_sizek;
777 gran_size <<= 10;
778 x86_setup_var_mtrrs(range, nr_range, chunk_size, gran_size);
779 set_var_mtrr_all();
780 Dprintk("New variable MTRRs\n");
781 print_out_mtrr_range_state();
782 return 1;
783 } else {
784 /* print out all */
785 for (i = 0; i < NUM_RESULT; i++)
786 mtrr_print_out_one_result(i);
787 }
788
789 pr_info("mtrr_cleanup: can not find optimal value\n");
790 pr_info("please specify mtrr_gran_size/mtrr_chunk_size\n");
791
792 return 0;
793}
794#else
795int __init mtrr_cleanup(void)
796{
797 return 0;
798}
799#endif
800
801static int disable_mtrr_trim;
802
803static int __init disable_mtrr_trim_setup(char *str)
804{
805 disable_mtrr_trim = 1;
806 return 0;
807}
808early_param("disable_mtrr_trim", disable_mtrr_trim_setup);
809
810/*
811 * Newer AMD K8s and later CPUs have a special magic MSR way to force WB
812 * for memory >4GB. Check for that here.
813 * Note this won't check if the MTRRs < 4GB where the magic bit doesn't
814 * apply to are wrong, but so far we don't know of any such case in the wild.
815 */
816#define Tom2Enabled (1U << 21)
817#define Tom2ForceMemTypeWB (1U << 22)
818
819int __init amd_special_default_mtrr(void)
820{
821 u32 l, h;
822
823 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD &&
824 boot_cpu_data.x86_vendor != X86_VENDOR_HYGON)
825 return 0;
826 if (boot_cpu_data.x86 < 0xf)
827 return 0;
828 /* In case some hypervisor doesn't pass SYSCFG through: */
829 if (rdmsr_safe(MSR_AMD64_SYSCFG, &l, &h) < 0)
830 return 0;
831 /*
832 * Memory between 4GB and top of mem is forced WB by this magic bit.
833 * Reserved before K8RevF, but should be zero there.
834 */
835 if ((l & (Tom2Enabled | Tom2ForceMemTypeWB)) ==
836 (Tom2Enabled | Tom2ForceMemTypeWB))
837 return 1;
838 return 0;
839}
840
841static u64 __init
842real_trim_memory(unsigned long start_pfn, unsigned long limit_pfn)
843{
844 u64 trim_start, trim_size;
845
846 trim_start = start_pfn;
847 trim_start <<= PAGE_SHIFT;
848
849 trim_size = limit_pfn;
850 trim_size <<= PAGE_SHIFT;
851 trim_size -= trim_start;
852
853 return e820__range_update(trim_start, trim_size, E820_TYPE_RAM, E820_TYPE_RESERVED);
854}
855
856/**
857 * mtrr_trim_uncached_memory - trim RAM not covered by MTRRs
858 * @end_pfn: ending page frame number
859 *
860 * Some buggy BIOSes don't setup the MTRRs properly for systems with certain
861 * memory configurations. This routine checks that the highest MTRR matches
862 * the end of memory, to make sure the MTRRs having a write back type cover
863 * all of the memory the kernel is intending to use. If not, it'll trim any
864 * memory off the end by adjusting end_pfn, removing it from the kernel's
865 * allocation pools, warning the user with an obnoxious message.
866 */
867int __init mtrr_trim_uncached_memory(unsigned long end_pfn)
868{
869 unsigned long i, base, size, highest_pfn = 0, def, dummy;
870 mtrr_type type;
871 u64 total_trim_size;
872 /* extra one for all 0 */
873 int num[MTRR_NUM_TYPES + 1];
874
875 if (!mtrr_enabled())
876 return 0;
877
878 /*
879 * Make sure we only trim uncachable memory on machines that
880 * support the Intel MTRR architecture:
881 */
882 if (!cpu_feature_enabled(X86_FEATURE_MTRR) || disable_mtrr_trim)
883 return 0;
884
885 rdmsr(MSR_MTRRdefType, def, dummy);
886 def &= MTRR_DEF_TYPE_TYPE;
887 if (def != MTRR_TYPE_UNCACHABLE)
888 return 0;
889
890 /* Get it and store it aside: */
891 memset(range_state, 0, sizeof(range_state));
892 for (i = 0; i < num_var_ranges; i++) {
893 mtrr_if->get(i, &base, &size, &type);
894 range_state[i].base_pfn = base;
895 range_state[i].size_pfn = size;
896 range_state[i].type = type;
897 }
898
899 /* Find highest cached pfn: */
900 for (i = 0; i < num_var_ranges; i++) {
901 type = range_state[i].type;
902 if (type != MTRR_TYPE_WRBACK)
903 continue;
904 base = range_state[i].base_pfn;
905 size = range_state[i].size_pfn;
906 if (highest_pfn < base + size)
907 highest_pfn = base + size;
908 }
909
910 /* kvm/qemu doesn't have mtrr set right, don't trim them all: */
911 if (!highest_pfn) {
912 pr_info("CPU MTRRs all blank - virtualized system.\n");
913 return 0;
914 }
915
916 /* Check entries number: */
917 memset(num, 0, sizeof(num));
918 for (i = 0; i < num_var_ranges; i++) {
919 type = range_state[i].type;
920 if (type >= MTRR_NUM_TYPES)
921 continue;
922 size = range_state[i].size_pfn;
923 if (!size)
924 type = MTRR_NUM_TYPES;
925 num[type]++;
926 }
927
928 /* No entry for WB? */
929 if (!num[MTRR_TYPE_WRBACK])
930 return 0;
931
932 /* Check if we only had WB and UC: */
933 if (num[MTRR_TYPE_WRBACK] + num[MTRR_TYPE_UNCACHABLE] !=
934 num_var_ranges - num[MTRR_NUM_TYPES])
935 return 0;
936
937 memset(range, 0, sizeof(range));
938 nr_range = 0;
939 if (mtrr_tom2) {
940 range[nr_range].start = (1ULL<<(32 - PAGE_SHIFT));
941 range[nr_range].end = mtrr_tom2 >> PAGE_SHIFT;
942 if (highest_pfn < range[nr_range].end)
943 highest_pfn = range[nr_range].end;
944 nr_range++;
945 }
946 nr_range = x86_get_mtrr_mem_range(range, nr_range, 0, 0);
947
948 /* Check the head: */
949 total_trim_size = 0;
950 if (range[0].start)
951 total_trim_size += real_trim_memory(0, range[0].start);
952
953 /* Check the holes: */
954 for (i = 0; i < nr_range - 1; i++) {
955 if (range[i].end < range[i+1].start)
956 total_trim_size += real_trim_memory(range[i].end,
957 range[i+1].start);
958 }
959
960 /* Check the top: */
961 i = nr_range - 1;
962 if (range[i].end < end_pfn)
963 total_trim_size += real_trim_memory(range[i].end,
964 end_pfn);
965
966 if (total_trim_size) {
967 pr_warn("WARNING: BIOS bug: CPU MTRRs don't cover all of memory, losing %lluMB of RAM.\n",
968 total_trim_size >> 20);
969
970 if (!changed_by_mtrr_cleanup)
971 WARN_ON(1);
972
973 pr_info("update e820 for mtrr\n");
974 e820__update_table_print();
975
976 return 1;
977 }
978
979 return 0;
980}