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