Loading...
1#include <linux/gfp.h>
2#include <linux/initrd.h>
3#include <linux/ioport.h>
4#include <linux/swap.h>
5#include <linux/memblock.h>
6#include <linux/bootmem.h> /* for max_low_pfn */
7
8#include <asm/cacheflush.h>
9#include <asm/e820.h>
10#include <asm/init.h>
11#include <asm/page.h>
12#include <asm/page_types.h>
13#include <asm/sections.h>
14#include <asm/setup.h>
15#include <asm/tlbflush.h>
16#include <asm/tlb.h>
17#include <asm/proto.h>
18#include <asm/dma.h> /* for MAX_DMA_PFN */
19#include <asm/microcode.h>
20#include <asm/kaslr.h>
21
22/*
23 * We need to define the tracepoints somewhere, and tlb.c
24 * is only compied when SMP=y.
25 */
26#define CREATE_TRACE_POINTS
27#include <trace/events/tlb.h>
28
29#include "mm_internal.h"
30
31/*
32 * Tables translating between page_cache_type_t and pte encoding.
33 *
34 * The default values are defined statically as minimal supported mode;
35 * WC and WT fall back to UC-. pat_init() updates these values to support
36 * more cache modes, WC and WT, when it is safe to do so. See pat_init()
37 * for the details. Note, __early_ioremap() used during early boot-time
38 * takes pgprot_t (pte encoding) and does not use these tables.
39 *
40 * Index into __cachemode2pte_tbl[] is the cachemode.
41 *
42 * Index into __pte2cachemode_tbl[] are the caching attribute bits of the pte
43 * (_PAGE_PWT, _PAGE_PCD, _PAGE_PAT) at index bit positions 0, 1, 2.
44 */
45uint16_t __cachemode2pte_tbl[_PAGE_CACHE_MODE_NUM] = {
46 [_PAGE_CACHE_MODE_WB ] = 0 | 0 ,
47 [_PAGE_CACHE_MODE_WC ] = 0 | _PAGE_PCD,
48 [_PAGE_CACHE_MODE_UC_MINUS] = 0 | _PAGE_PCD,
49 [_PAGE_CACHE_MODE_UC ] = _PAGE_PWT | _PAGE_PCD,
50 [_PAGE_CACHE_MODE_WT ] = 0 | _PAGE_PCD,
51 [_PAGE_CACHE_MODE_WP ] = 0 | _PAGE_PCD,
52};
53EXPORT_SYMBOL(__cachemode2pte_tbl);
54
55uint8_t __pte2cachemode_tbl[8] = {
56 [__pte2cm_idx( 0 | 0 | 0 )] = _PAGE_CACHE_MODE_WB,
57 [__pte2cm_idx(_PAGE_PWT | 0 | 0 )] = _PAGE_CACHE_MODE_UC_MINUS,
58 [__pte2cm_idx( 0 | _PAGE_PCD | 0 )] = _PAGE_CACHE_MODE_UC_MINUS,
59 [__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | 0 )] = _PAGE_CACHE_MODE_UC,
60 [__pte2cm_idx( 0 | 0 | _PAGE_PAT)] = _PAGE_CACHE_MODE_WB,
61 [__pte2cm_idx(_PAGE_PWT | 0 | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
62 [__pte2cm_idx(0 | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
63 [__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC,
64};
65EXPORT_SYMBOL(__pte2cachemode_tbl);
66
67static unsigned long __initdata pgt_buf_start;
68static unsigned long __initdata pgt_buf_end;
69static unsigned long __initdata pgt_buf_top;
70
71static unsigned long min_pfn_mapped;
72
73static bool __initdata can_use_brk_pgt = true;
74
75/*
76 * Pages returned are already directly mapped.
77 *
78 * Changing that is likely to break Xen, see commit:
79 *
80 * 279b706 x86,xen: introduce x86_init.mapping.pagetable_reserve
81 *
82 * for detailed information.
83 */
84__ref void *alloc_low_pages(unsigned int num)
85{
86 unsigned long pfn;
87 int i;
88
89 if (after_bootmem) {
90 unsigned int order;
91
92 order = get_order((unsigned long)num << PAGE_SHIFT);
93 return (void *)__get_free_pages(GFP_ATOMIC | __GFP_NOTRACK |
94 __GFP_ZERO, order);
95 }
96
97 if ((pgt_buf_end + num) > pgt_buf_top || !can_use_brk_pgt) {
98 unsigned long ret;
99 if (min_pfn_mapped >= max_pfn_mapped)
100 panic("alloc_low_pages: ran out of memory");
101 ret = memblock_find_in_range(min_pfn_mapped << PAGE_SHIFT,
102 max_pfn_mapped << PAGE_SHIFT,
103 PAGE_SIZE * num , PAGE_SIZE);
104 if (!ret)
105 panic("alloc_low_pages: can not alloc memory");
106 memblock_reserve(ret, PAGE_SIZE * num);
107 pfn = ret >> PAGE_SHIFT;
108 } else {
109 pfn = pgt_buf_end;
110 pgt_buf_end += num;
111 printk(KERN_DEBUG "BRK [%#010lx, %#010lx] PGTABLE\n",
112 pfn << PAGE_SHIFT, (pgt_buf_end << PAGE_SHIFT) - 1);
113 }
114
115 for (i = 0; i < num; i++) {
116 void *adr;
117
118 adr = __va((pfn + i) << PAGE_SHIFT);
119 clear_page(adr);
120 }
121
122 return __va(pfn << PAGE_SHIFT);
123}
124
125/*
126 * By default need 3 4k for initial PMD_SIZE, 3 4k for 0-ISA_END_ADDRESS.
127 * With KASLR memory randomization, depending on the machine e820 memory
128 * and the PUD alignment. We may need twice more pages when KASLR memory
129 * randomization is enabled.
130 */
131#ifndef CONFIG_RANDOMIZE_MEMORY
132#define INIT_PGD_PAGE_COUNT 6
133#else
134#define INIT_PGD_PAGE_COUNT 12
135#endif
136#define INIT_PGT_BUF_SIZE (INIT_PGD_PAGE_COUNT * PAGE_SIZE)
137RESERVE_BRK(early_pgt_alloc, INIT_PGT_BUF_SIZE);
138void __init early_alloc_pgt_buf(void)
139{
140 unsigned long tables = INIT_PGT_BUF_SIZE;
141 phys_addr_t base;
142
143 base = __pa(extend_brk(tables, PAGE_SIZE));
144
145 pgt_buf_start = base >> PAGE_SHIFT;
146 pgt_buf_end = pgt_buf_start;
147 pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
148}
149
150int after_bootmem;
151
152early_param_on_off("gbpages", "nogbpages", direct_gbpages, CONFIG_X86_DIRECT_GBPAGES);
153
154struct map_range {
155 unsigned long start;
156 unsigned long end;
157 unsigned page_size_mask;
158};
159
160static int page_size_mask;
161
162static void __init probe_page_size_mask(void)
163{
164#if !defined(CONFIG_KMEMCHECK)
165 /*
166 * For CONFIG_KMEMCHECK or pagealloc debugging, identity mapping will
167 * use small pages.
168 * This will simplify cpa(), which otherwise needs to support splitting
169 * large pages into small in interrupt context, etc.
170 */
171 if (boot_cpu_has(X86_FEATURE_PSE) && !debug_pagealloc_enabled())
172 page_size_mask |= 1 << PG_LEVEL_2M;
173#endif
174
175 /* Enable PSE if available */
176 if (boot_cpu_has(X86_FEATURE_PSE))
177 cr4_set_bits_and_update_boot(X86_CR4_PSE);
178
179 /* Enable PGE if available */
180 if (boot_cpu_has(X86_FEATURE_PGE)) {
181 cr4_set_bits_and_update_boot(X86_CR4_PGE);
182 __supported_pte_mask |= _PAGE_GLOBAL;
183 } else
184 __supported_pte_mask &= ~_PAGE_GLOBAL;
185
186 /* Enable 1 GB linear kernel mappings if available: */
187 if (direct_gbpages && boot_cpu_has(X86_FEATURE_GBPAGES)) {
188 printk(KERN_INFO "Using GB pages for direct mapping\n");
189 page_size_mask |= 1 << PG_LEVEL_1G;
190 } else {
191 direct_gbpages = 0;
192 }
193}
194
195#ifdef CONFIG_X86_32
196#define NR_RANGE_MR 3
197#else /* CONFIG_X86_64 */
198#define NR_RANGE_MR 5
199#endif
200
201static int __meminit save_mr(struct map_range *mr, int nr_range,
202 unsigned long start_pfn, unsigned long end_pfn,
203 unsigned long page_size_mask)
204{
205 if (start_pfn < end_pfn) {
206 if (nr_range >= NR_RANGE_MR)
207 panic("run out of range for init_memory_mapping\n");
208 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
209 mr[nr_range].end = end_pfn<<PAGE_SHIFT;
210 mr[nr_range].page_size_mask = page_size_mask;
211 nr_range++;
212 }
213
214 return nr_range;
215}
216
217/*
218 * adjust the page_size_mask for small range to go with
219 * big page size instead small one if nearby are ram too.
220 */
221static void __ref adjust_range_page_size_mask(struct map_range *mr,
222 int nr_range)
223{
224 int i;
225
226 for (i = 0; i < nr_range; i++) {
227 if ((page_size_mask & (1<<PG_LEVEL_2M)) &&
228 !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) {
229 unsigned long start = round_down(mr[i].start, PMD_SIZE);
230 unsigned long end = round_up(mr[i].end, PMD_SIZE);
231
232#ifdef CONFIG_X86_32
233 if ((end >> PAGE_SHIFT) > max_low_pfn)
234 continue;
235#endif
236
237 if (memblock_is_region_memory(start, end - start))
238 mr[i].page_size_mask |= 1<<PG_LEVEL_2M;
239 }
240 if ((page_size_mask & (1<<PG_LEVEL_1G)) &&
241 !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) {
242 unsigned long start = round_down(mr[i].start, PUD_SIZE);
243 unsigned long end = round_up(mr[i].end, PUD_SIZE);
244
245 if (memblock_is_region_memory(start, end - start))
246 mr[i].page_size_mask |= 1<<PG_LEVEL_1G;
247 }
248 }
249}
250
251static const char *page_size_string(struct map_range *mr)
252{
253 static const char str_1g[] = "1G";
254 static const char str_2m[] = "2M";
255 static const char str_4m[] = "4M";
256 static const char str_4k[] = "4k";
257
258 if (mr->page_size_mask & (1<<PG_LEVEL_1G))
259 return str_1g;
260 /*
261 * 32-bit without PAE has a 4M large page size.
262 * PG_LEVEL_2M is misnamed, but we can at least
263 * print out the right size in the string.
264 */
265 if (IS_ENABLED(CONFIG_X86_32) &&
266 !IS_ENABLED(CONFIG_X86_PAE) &&
267 mr->page_size_mask & (1<<PG_LEVEL_2M))
268 return str_4m;
269
270 if (mr->page_size_mask & (1<<PG_LEVEL_2M))
271 return str_2m;
272
273 return str_4k;
274}
275
276static int __meminit split_mem_range(struct map_range *mr, int nr_range,
277 unsigned long start,
278 unsigned long end)
279{
280 unsigned long start_pfn, end_pfn, limit_pfn;
281 unsigned long pfn;
282 int i;
283
284 limit_pfn = PFN_DOWN(end);
285
286 /* head if not big page alignment ? */
287 pfn = start_pfn = PFN_DOWN(start);
288#ifdef CONFIG_X86_32
289 /*
290 * Don't use a large page for the first 2/4MB of memory
291 * because there are often fixed size MTRRs in there
292 * and overlapping MTRRs into large pages can cause
293 * slowdowns.
294 */
295 if (pfn == 0)
296 end_pfn = PFN_DOWN(PMD_SIZE);
297 else
298 end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
299#else /* CONFIG_X86_64 */
300 end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
301#endif
302 if (end_pfn > limit_pfn)
303 end_pfn = limit_pfn;
304 if (start_pfn < end_pfn) {
305 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
306 pfn = end_pfn;
307 }
308
309 /* big page (2M) range */
310 start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
311#ifdef CONFIG_X86_32
312 end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
313#else /* CONFIG_X86_64 */
314 end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
315 if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE)))
316 end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
317#endif
318
319 if (start_pfn < end_pfn) {
320 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
321 page_size_mask & (1<<PG_LEVEL_2M));
322 pfn = end_pfn;
323 }
324
325#ifdef CONFIG_X86_64
326 /* big page (1G) range */
327 start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
328 end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE));
329 if (start_pfn < end_pfn) {
330 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
331 page_size_mask &
332 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
333 pfn = end_pfn;
334 }
335
336 /* tail is not big page (1G) alignment */
337 start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
338 end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
339 if (start_pfn < end_pfn) {
340 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
341 page_size_mask & (1<<PG_LEVEL_2M));
342 pfn = end_pfn;
343 }
344#endif
345
346 /* tail is not big page (2M) alignment */
347 start_pfn = pfn;
348 end_pfn = limit_pfn;
349 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
350
351 if (!after_bootmem)
352 adjust_range_page_size_mask(mr, nr_range);
353
354 /* try to merge same page size and continuous */
355 for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
356 unsigned long old_start;
357 if (mr[i].end != mr[i+1].start ||
358 mr[i].page_size_mask != mr[i+1].page_size_mask)
359 continue;
360 /* move it */
361 old_start = mr[i].start;
362 memmove(&mr[i], &mr[i+1],
363 (nr_range - 1 - i) * sizeof(struct map_range));
364 mr[i--].start = old_start;
365 nr_range--;
366 }
367
368 for (i = 0; i < nr_range; i++)
369 pr_debug(" [mem %#010lx-%#010lx] page %s\n",
370 mr[i].start, mr[i].end - 1,
371 page_size_string(&mr[i]));
372
373 return nr_range;
374}
375
376struct range pfn_mapped[E820_X_MAX];
377int nr_pfn_mapped;
378
379static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn)
380{
381 nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_X_MAX,
382 nr_pfn_mapped, start_pfn, end_pfn);
383 nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_X_MAX);
384
385 max_pfn_mapped = max(max_pfn_mapped, end_pfn);
386
387 if (start_pfn < (1UL<<(32-PAGE_SHIFT)))
388 max_low_pfn_mapped = max(max_low_pfn_mapped,
389 min(end_pfn, 1UL<<(32-PAGE_SHIFT)));
390}
391
392bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn)
393{
394 int i;
395
396 for (i = 0; i < nr_pfn_mapped; i++)
397 if ((start_pfn >= pfn_mapped[i].start) &&
398 (end_pfn <= pfn_mapped[i].end))
399 return true;
400
401 return false;
402}
403
404/*
405 * Setup the direct mapping of the physical memory at PAGE_OFFSET.
406 * This runs before bootmem is initialized and gets pages directly from
407 * the physical memory. To access them they are temporarily mapped.
408 */
409unsigned long __ref init_memory_mapping(unsigned long start,
410 unsigned long end)
411{
412 struct map_range mr[NR_RANGE_MR];
413 unsigned long ret = 0;
414 int nr_range, i;
415
416 pr_debug("init_memory_mapping: [mem %#010lx-%#010lx]\n",
417 start, end - 1);
418
419 memset(mr, 0, sizeof(mr));
420 nr_range = split_mem_range(mr, 0, start, end);
421
422 for (i = 0; i < nr_range; i++)
423 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
424 mr[i].page_size_mask);
425
426 add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT);
427
428 return ret >> PAGE_SHIFT;
429}
430
431/*
432 * We need to iterate through the E820 memory map and create direct mappings
433 * for only E820_RAM and E820_KERN_RESERVED regions. We cannot simply
434 * create direct mappings for all pfns from [0 to max_low_pfn) and
435 * [4GB to max_pfn) because of possible memory holes in high addresses
436 * that cannot be marked as UC by fixed/variable range MTRRs.
437 * Depending on the alignment of E820 ranges, this may possibly result
438 * in using smaller size (i.e. 4K instead of 2M or 1G) page tables.
439 *
440 * init_mem_mapping() calls init_range_memory_mapping() with big range.
441 * That range would have hole in the middle or ends, and only ram parts
442 * will be mapped in init_range_memory_mapping().
443 */
444static unsigned long __init init_range_memory_mapping(
445 unsigned long r_start,
446 unsigned long r_end)
447{
448 unsigned long start_pfn, end_pfn;
449 unsigned long mapped_ram_size = 0;
450 int i;
451
452 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
453 u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end);
454 u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end);
455 if (start >= end)
456 continue;
457
458 /*
459 * if it is overlapping with brk pgt, we need to
460 * alloc pgt buf from memblock instead.
461 */
462 can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >=
463 min(end, (u64)pgt_buf_top<<PAGE_SHIFT);
464 init_memory_mapping(start, end);
465 mapped_ram_size += end - start;
466 can_use_brk_pgt = true;
467 }
468
469 return mapped_ram_size;
470}
471
472static unsigned long __init get_new_step_size(unsigned long step_size)
473{
474 /*
475 * Initial mapped size is PMD_SIZE (2M).
476 * We can not set step_size to be PUD_SIZE (1G) yet.
477 * In worse case, when we cross the 1G boundary, and
478 * PG_LEVEL_2M is not set, we will need 1+1+512 pages (2M + 8k)
479 * to map 1G range with PTE. Hence we use one less than the
480 * difference of page table level shifts.
481 *
482 * Don't need to worry about overflow in the top-down case, on 32bit,
483 * when step_size is 0, round_down() returns 0 for start, and that
484 * turns it into 0x100000000ULL.
485 * In the bottom-up case, round_up(x, 0) returns 0 though too, which
486 * needs to be taken into consideration by the code below.
487 */
488 return step_size << (PMD_SHIFT - PAGE_SHIFT - 1);
489}
490
491/**
492 * memory_map_top_down - Map [map_start, map_end) top down
493 * @map_start: start address of the target memory range
494 * @map_end: end address of the target memory range
495 *
496 * This function will setup direct mapping for memory range
497 * [map_start, map_end) in top-down. That said, the page tables
498 * will be allocated at the end of the memory, and we map the
499 * memory in top-down.
500 */
501static void __init memory_map_top_down(unsigned long map_start,
502 unsigned long map_end)
503{
504 unsigned long real_end, start, last_start;
505 unsigned long step_size;
506 unsigned long addr;
507 unsigned long mapped_ram_size = 0;
508
509 /* xen has big range in reserved near end of ram, skip it at first.*/
510 addr = memblock_find_in_range(map_start, map_end, PMD_SIZE, PMD_SIZE);
511 real_end = addr + PMD_SIZE;
512
513 /* step_size need to be small so pgt_buf from BRK could cover it */
514 step_size = PMD_SIZE;
515 max_pfn_mapped = 0; /* will get exact value next */
516 min_pfn_mapped = real_end >> PAGE_SHIFT;
517 last_start = start = real_end;
518
519 /*
520 * We start from the top (end of memory) and go to the bottom.
521 * The memblock_find_in_range() gets us a block of RAM from the
522 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
523 * for page table.
524 */
525 while (last_start > map_start) {
526 if (last_start > step_size) {
527 start = round_down(last_start - 1, step_size);
528 if (start < map_start)
529 start = map_start;
530 } else
531 start = map_start;
532 mapped_ram_size += init_range_memory_mapping(start,
533 last_start);
534 last_start = start;
535 min_pfn_mapped = last_start >> PAGE_SHIFT;
536 if (mapped_ram_size >= step_size)
537 step_size = get_new_step_size(step_size);
538 }
539
540 if (real_end < map_end)
541 init_range_memory_mapping(real_end, map_end);
542}
543
544/**
545 * memory_map_bottom_up - Map [map_start, map_end) bottom up
546 * @map_start: start address of the target memory range
547 * @map_end: end address of the target memory range
548 *
549 * This function will setup direct mapping for memory range
550 * [map_start, map_end) in bottom-up. Since we have limited the
551 * bottom-up allocation above the kernel, the page tables will
552 * be allocated just above the kernel and we map the memory
553 * in [map_start, map_end) in bottom-up.
554 */
555static void __init memory_map_bottom_up(unsigned long map_start,
556 unsigned long map_end)
557{
558 unsigned long next, start;
559 unsigned long mapped_ram_size = 0;
560 /* step_size need to be small so pgt_buf from BRK could cover it */
561 unsigned long step_size = PMD_SIZE;
562
563 start = map_start;
564 min_pfn_mapped = start >> PAGE_SHIFT;
565
566 /*
567 * We start from the bottom (@map_start) and go to the top (@map_end).
568 * The memblock_find_in_range() gets us a block of RAM from the
569 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
570 * for page table.
571 */
572 while (start < map_end) {
573 if (step_size && map_end - start > step_size) {
574 next = round_up(start + 1, step_size);
575 if (next > map_end)
576 next = map_end;
577 } else {
578 next = map_end;
579 }
580
581 mapped_ram_size += init_range_memory_mapping(start, next);
582 start = next;
583
584 if (mapped_ram_size >= step_size)
585 step_size = get_new_step_size(step_size);
586 }
587}
588
589void __init init_mem_mapping(void)
590{
591 unsigned long end;
592
593 probe_page_size_mask();
594
595#ifdef CONFIG_X86_64
596 end = max_pfn << PAGE_SHIFT;
597#else
598 end = max_low_pfn << PAGE_SHIFT;
599#endif
600
601 /* the ISA range is always mapped regardless of memory holes */
602 init_memory_mapping(0, ISA_END_ADDRESS);
603
604 /* Init the trampoline, possibly with KASLR memory offset */
605 init_trampoline();
606
607 /*
608 * If the allocation is in bottom-up direction, we setup direct mapping
609 * in bottom-up, otherwise we setup direct mapping in top-down.
610 */
611 if (memblock_bottom_up()) {
612 unsigned long kernel_end = __pa_symbol(_end);
613
614 /*
615 * we need two separate calls here. This is because we want to
616 * allocate page tables above the kernel. So we first map
617 * [kernel_end, end) to make memory above the kernel be mapped
618 * as soon as possible. And then use page tables allocated above
619 * the kernel to map [ISA_END_ADDRESS, kernel_end).
620 */
621 memory_map_bottom_up(kernel_end, end);
622 memory_map_bottom_up(ISA_END_ADDRESS, kernel_end);
623 } else {
624 memory_map_top_down(ISA_END_ADDRESS, end);
625 }
626
627#ifdef CONFIG_X86_64
628 if (max_pfn > max_low_pfn) {
629 /* can we preseve max_low_pfn ?*/
630 max_low_pfn = max_pfn;
631 }
632#else
633 early_ioremap_page_table_range_init();
634#endif
635
636 load_cr3(swapper_pg_dir);
637 __flush_tlb_all();
638
639 early_memtest(0, max_pfn_mapped << PAGE_SHIFT);
640}
641
642/*
643 * devmem_is_allowed() checks to see if /dev/mem access to a certain address
644 * is valid. The argument is a physical page number.
645 *
646 *
647 * On x86, access has to be given to the first megabyte of ram because that area
648 * contains BIOS code and data regions used by X and dosemu and similar apps.
649 * Access has to be given to non-kernel-ram areas as well, these contain the PCI
650 * mmio resources as well as potential bios/acpi data regions.
651 */
652int devmem_is_allowed(unsigned long pagenr)
653{
654 if (pagenr < 256)
655 return 1;
656 if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
657 return 0;
658 if (!page_is_ram(pagenr))
659 return 1;
660 return 0;
661}
662
663void free_init_pages(char *what, unsigned long begin, unsigned long end)
664{
665 unsigned long begin_aligned, end_aligned;
666
667 /* Make sure boundaries are page aligned */
668 begin_aligned = PAGE_ALIGN(begin);
669 end_aligned = end & PAGE_MASK;
670
671 if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
672 begin = begin_aligned;
673 end = end_aligned;
674 }
675
676 if (begin >= end)
677 return;
678
679 /*
680 * If debugging page accesses then do not free this memory but
681 * mark them not present - any buggy init-section access will
682 * create a kernel page fault:
683 */
684 if (debug_pagealloc_enabled()) {
685 pr_info("debug: unmapping init [mem %#010lx-%#010lx]\n",
686 begin, end - 1);
687 set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
688 } else {
689 /*
690 * We just marked the kernel text read only above, now that
691 * we are going to free part of that, we need to make that
692 * writeable and non-executable first.
693 */
694 set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
695 set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
696
697 free_reserved_area((void *)begin, (void *)end,
698 POISON_FREE_INITMEM, what);
699 }
700}
701
702void __ref free_initmem(void)
703{
704 e820_reallocate_tables();
705
706 free_init_pages("unused kernel",
707 (unsigned long)(&__init_begin),
708 (unsigned long)(&__init_end));
709}
710
711#ifdef CONFIG_BLK_DEV_INITRD
712void __init free_initrd_mem(unsigned long start, unsigned long end)
713{
714 /*
715 * end could be not aligned, and We can not align that,
716 * decompresser could be confused by aligned initrd_end
717 * We already reserve the end partial page before in
718 * - i386_start_kernel()
719 * - x86_64_start_kernel()
720 * - relocate_initrd()
721 * So here We can do PAGE_ALIGN() safely to get partial page to be freed
722 */
723 free_init_pages("initrd", start, PAGE_ALIGN(end));
724}
725#endif
726
727void __init zone_sizes_init(void)
728{
729 unsigned long max_zone_pfns[MAX_NR_ZONES];
730
731 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
732
733#ifdef CONFIG_ZONE_DMA
734 max_zone_pfns[ZONE_DMA] = min(MAX_DMA_PFN, max_low_pfn);
735#endif
736#ifdef CONFIG_ZONE_DMA32
737 max_zone_pfns[ZONE_DMA32] = min(MAX_DMA32_PFN, max_low_pfn);
738#endif
739 max_zone_pfns[ZONE_NORMAL] = max_low_pfn;
740#ifdef CONFIG_HIGHMEM
741 max_zone_pfns[ZONE_HIGHMEM] = max_pfn;
742#endif
743
744 free_area_init_nodes(max_zone_pfns);
745}
746
747DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate) = {
748#ifdef CONFIG_SMP
749 .active_mm = &init_mm,
750 .state = 0,
751#endif
752 .cr4 = ~0UL, /* fail hard if we screw up cr4 shadow initialization */
753};
754EXPORT_SYMBOL_GPL(cpu_tlbstate);
755
756void update_cache_mode_entry(unsigned entry, enum page_cache_mode cache)
757{
758 /* entry 0 MUST be WB (hardwired to speed up translations) */
759 BUG_ON(!entry && cache != _PAGE_CACHE_MODE_WB);
760
761 __cachemode2pte_tbl[cache] = __cm_idx2pte(entry);
762 __pte2cachemode_tbl[entry] = cache;
763}
1#include <linux/gfp.h>
2#include <linux/initrd.h>
3#include <linux/ioport.h>
4#include <linux/swap.h>
5#include <linux/memblock.h>
6
7#include <asm/cacheflush.h>
8#include <asm/e820.h>
9#include <asm/init.h>
10#include <asm/page.h>
11#include <asm/page_types.h>
12#include <asm/sections.h>
13#include <asm/setup.h>
14#include <asm/system.h>
15#include <asm/tlbflush.h>
16#include <asm/tlb.h>
17#include <asm/proto.h>
18
19unsigned long __initdata pgt_buf_start;
20unsigned long __meminitdata pgt_buf_end;
21unsigned long __meminitdata pgt_buf_top;
22
23int after_bootmem;
24
25int direct_gbpages
26#ifdef CONFIG_DIRECT_GBPAGES
27 = 1
28#endif
29;
30
31static void __init find_early_table_space(unsigned long end, int use_pse,
32 int use_gbpages)
33{
34 unsigned long puds, pmds, ptes, tables, start = 0, good_end = end;
35 phys_addr_t base;
36
37 puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
38 tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
39
40 if (use_gbpages) {
41 unsigned long extra;
42
43 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
44 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
45 } else
46 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
47
48 tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
49
50 if (use_pse) {
51 unsigned long extra;
52
53 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
54#ifdef CONFIG_X86_32
55 extra += PMD_SIZE;
56#endif
57 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
58 } else
59 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
60
61 tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
62
63#ifdef CONFIG_X86_32
64 /* for fixmap */
65 tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
66#endif
67 good_end = max_pfn_mapped << PAGE_SHIFT;
68
69 base = memblock_find_in_range(start, good_end, tables, PAGE_SIZE);
70 if (base == MEMBLOCK_ERROR)
71 panic("Cannot find space for the kernel page tables");
72
73 pgt_buf_start = base >> PAGE_SHIFT;
74 pgt_buf_end = pgt_buf_start;
75 pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
76
77 printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
78 end, pgt_buf_start << PAGE_SHIFT, pgt_buf_top << PAGE_SHIFT);
79}
80
81void __init native_pagetable_reserve(u64 start, u64 end)
82{
83 memblock_x86_reserve_range(start, end, "PGTABLE");
84}
85
86struct map_range {
87 unsigned long start;
88 unsigned long end;
89 unsigned page_size_mask;
90};
91
92#ifdef CONFIG_X86_32
93#define NR_RANGE_MR 3
94#else /* CONFIG_X86_64 */
95#define NR_RANGE_MR 5
96#endif
97
98static int __meminit save_mr(struct map_range *mr, int nr_range,
99 unsigned long start_pfn, unsigned long end_pfn,
100 unsigned long page_size_mask)
101{
102 if (start_pfn < end_pfn) {
103 if (nr_range >= NR_RANGE_MR)
104 panic("run out of range for init_memory_mapping\n");
105 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
106 mr[nr_range].end = end_pfn<<PAGE_SHIFT;
107 mr[nr_range].page_size_mask = page_size_mask;
108 nr_range++;
109 }
110
111 return nr_range;
112}
113
114/*
115 * Setup the direct mapping of the physical memory at PAGE_OFFSET.
116 * This runs before bootmem is initialized and gets pages directly from
117 * the physical memory. To access them they are temporarily mapped.
118 */
119unsigned long __init_refok init_memory_mapping(unsigned long start,
120 unsigned long end)
121{
122 unsigned long page_size_mask = 0;
123 unsigned long start_pfn, end_pfn;
124 unsigned long ret = 0;
125 unsigned long pos;
126
127 struct map_range mr[NR_RANGE_MR];
128 int nr_range, i;
129 int use_pse, use_gbpages;
130
131 printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
132
133#if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_KMEMCHECK)
134 /*
135 * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
136 * This will simplify cpa(), which otherwise needs to support splitting
137 * large pages into small in interrupt context, etc.
138 */
139 use_pse = use_gbpages = 0;
140#else
141 use_pse = cpu_has_pse;
142 use_gbpages = direct_gbpages;
143#endif
144
145 /* Enable PSE if available */
146 if (cpu_has_pse)
147 set_in_cr4(X86_CR4_PSE);
148
149 /* Enable PGE if available */
150 if (cpu_has_pge) {
151 set_in_cr4(X86_CR4_PGE);
152 __supported_pte_mask |= _PAGE_GLOBAL;
153 }
154
155 if (use_gbpages)
156 page_size_mask |= 1 << PG_LEVEL_1G;
157 if (use_pse)
158 page_size_mask |= 1 << PG_LEVEL_2M;
159
160 memset(mr, 0, sizeof(mr));
161 nr_range = 0;
162
163 /* head if not big page alignment ? */
164 start_pfn = start >> PAGE_SHIFT;
165 pos = start_pfn << PAGE_SHIFT;
166#ifdef CONFIG_X86_32
167 /*
168 * Don't use a large page for the first 2/4MB of memory
169 * because there are often fixed size MTRRs in there
170 * and overlapping MTRRs into large pages can cause
171 * slowdowns.
172 */
173 if (pos == 0)
174 end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
175 else
176 end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
177 << (PMD_SHIFT - PAGE_SHIFT);
178#else /* CONFIG_X86_64 */
179 end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
180 << (PMD_SHIFT - PAGE_SHIFT);
181#endif
182 if (end_pfn > (end >> PAGE_SHIFT))
183 end_pfn = end >> PAGE_SHIFT;
184 if (start_pfn < end_pfn) {
185 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
186 pos = end_pfn << PAGE_SHIFT;
187 }
188
189 /* big page (2M) range */
190 start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
191 << (PMD_SHIFT - PAGE_SHIFT);
192#ifdef CONFIG_X86_32
193 end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
194#else /* CONFIG_X86_64 */
195 end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
196 << (PUD_SHIFT - PAGE_SHIFT);
197 if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
198 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
199#endif
200
201 if (start_pfn < end_pfn) {
202 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
203 page_size_mask & (1<<PG_LEVEL_2M));
204 pos = end_pfn << PAGE_SHIFT;
205 }
206
207#ifdef CONFIG_X86_64
208 /* big page (1G) range */
209 start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
210 << (PUD_SHIFT - PAGE_SHIFT);
211 end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
212 if (start_pfn < end_pfn) {
213 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
214 page_size_mask &
215 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
216 pos = end_pfn << PAGE_SHIFT;
217 }
218
219 /* tail is not big page (1G) alignment */
220 start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
221 << (PMD_SHIFT - PAGE_SHIFT);
222 end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
223 if (start_pfn < end_pfn) {
224 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
225 page_size_mask & (1<<PG_LEVEL_2M));
226 pos = end_pfn << PAGE_SHIFT;
227 }
228#endif
229
230 /* tail is not big page (2M) alignment */
231 start_pfn = pos>>PAGE_SHIFT;
232 end_pfn = end>>PAGE_SHIFT;
233 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
234
235 /* try to merge same page size and continuous */
236 for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
237 unsigned long old_start;
238 if (mr[i].end != mr[i+1].start ||
239 mr[i].page_size_mask != mr[i+1].page_size_mask)
240 continue;
241 /* move it */
242 old_start = mr[i].start;
243 memmove(&mr[i], &mr[i+1],
244 (nr_range - 1 - i) * sizeof(struct map_range));
245 mr[i--].start = old_start;
246 nr_range--;
247 }
248
249 for (i = 0; i < nr_range; i++)
250 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
251 mr[i].start, mr[i].end,
252 (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
253 (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
254
255 /*
256 * Find space for the kernel direct mapping tables.
257 *
258 * Later we should allocate these tables in the local node of the
259 * memory mapped. Unfortunately this is done currently before the
260 * nodes are discovered.
261 */
262 if (!after_bootmem)
263 find_early_table_space(end, use_pse, use_gbpages);
264
265 for (i = 0; i < nr_range; i++)
266 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
267 mr[i].page_size_mask);
268
269#ifdef CONFIG_X86_32
270 early_ioremap_page_table_range_init();
271
272 load_cr3(swapper_pg_dir);
273#endif
274
275 __flush_tlb_all();
276
277 /*
278 * Reserve the kernel pagetable pages we used (pgt_buf_start -
279 * pgt_buf_end) and free the other ones (pgt_buf_end - pgt_buf_top)
280 * so that they can be reused for other purposes.
281 *
282 * On native it just means calling memblock_x86_reserve_range, on Xen it
283 * also means marking RW the pagetable pages that we allocated before
284 * but that haven't been used.
285 *
286 * In fact on xen we mark RO the whole range pgt_buf_start -
287 * pgt_buf_top, because we have to make sure that when
288 * init_memory_mapping reaches the pagetable pages area, it maps
289 * RO all the pagetable pages, including the ones that are beyond
290 * pgt_buf_end at that time.
291 */
292 if (!after_bootmem && pgt_buf_end > pgt_buf_start)
293 x86_init.mapping.pagetable_reserve(PFN_PHYS(pgt_buf_start),
294 PFN_PHYS(pgt_buf_end));
295
296 if (!after_bootmem)
297 early_memtest(start, end);
298
299 return ret >> PAGE_SHIFT;
300}
301
302
303/*
304 * devmem_is_allowed() checks to see if /dev/mem access to a certain address
305 * is valid. The argument is a physical page number.
306 *
307 *
308 * On x86, access has to be given to the first megabyte of ram because that area
309 * contains bios code and data regions used by X and dosemu and similar apps.
310 * Access has to be given to non-kernel-ram areas as well, these contain the PCI
311 * mmio resources as well as potential bios/acpi data regions.
312 */
313int devmem_is_allowed(unsigned long pagenr)
314{
315 if (pagenr <= 256)
316 return 1;
317 if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
318 return 0;
319 if (!page_is_ram(pagenr))
320 return 1;
321 return 0;
322}
323
324void free_init_pages(char *what, unsigned long begin, unsigned long end)
325{
326 unsigned long addr;
327 unsigned long begin_aligned, end_aligned;
328
329 /* Make sure boundaries are page aligned */
330 begin_aligned = PAGE_ALIGN(begin);
331 end_aligned = end & PAGE_MASK;
332
333 if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
334 begin = begin_aligned;
335 end = end_aligned;
336 }
337
338 if (begin >= end)
339 return;
340
341 addr = begin;
342
343 /*
344 * If debugging page accesses then do not free this memory but
345 * mark them not present - any buggy init-section access will
346 * create a kernel page fault:
347 */
348#ifdef CONFIG_DEBUG_PAGEALLOC
349 printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
350 begin, end);
351 set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
352#else
353 /*
354 * We just marked the kernel text read only above, now that
355 * we are going to free part of that, we need to make that
356 * writeable and non-executable first.
357 */
358 set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
359 set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
360
361 printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
362
363 for (; addr < end; addr += PAGE_SIZE) {
364 ClearPageReserved(virt_to_page(addr));
365 init_page_count(virt_to_page(addr));
366 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
367 free_page(addr);
368 totalram_pages++;
369 }
370#endif
371}
372
373void free_initmem(void)
374{
375 free_init_pages("unused kernel memory",
376 (unsigned long)(&__init_begin),
377 (unsigned long)(&__init_end));
378}
379
380#ifdef CONFIG_BLK_DEV_INITRD
381void free_initrd_mem(unsigned long start, unsigned long end)
382{
383 /*
384 * end could be not aligned, and We can not align that,
385 * decompresser could be confused by aligned initrd_end
386 * We already reserve the end partial page before in
387 * - i386_start_kernel()
388 * - x86_64_start_kernel()
389 * - relocate_initrd()
390 * So here We can do PAGE_ALIGN() safely to get partial page to be freed
391 */
392 free_init_pages("initrd memory", start, PAGE_ALIGN(end));
393}
394#endif