xref: /linux/arch/x86/mm/init.c (revision fd639726bf15fca8ee1a00dce8e0096d0ad9bd18)
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/set_memory.h>
9 #include <asm/e820/api.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 #include <asm/hypervisor.h>
22 #include <asm/cpufeature.h>
23 
24 /*
25  * We need to define the tracepoints somewhere, and tlb.c
26  * is only compied when SMP=y.
27  */
28 #define CREATE_TRACE_POINTS
29 #include <trace/events/tlb.h>
30 
31 #include "mm_internal.h"
32 
33 /*
34  * Tables translating between page_cache_type_t and pte encoding.
35  *
36  * The default values are defined statically as minimal supported mode;
37  * WC and WT fall back to UC-.  pat_init() updates these values to support
38  * more cache modes, WC and WT, when it is safe to do so.  See pat_init()
39  * for the details.  Note, __early_ioremap() used during early boot-time
40  * takes pgprot_t (pte encoding) and does not use these tables.
41  *
42  *   Index into __cachemode2pte_tbl[] is the cachemode.
43  *
44  *   Index into __pte2cachemode_tbl[] are the caching attribute bits of the pte
45  *   (_PAGE_PWT, _PAGE_PCD, _PAGE_PAT) at index bit positions 0, 1, 2.
46  */
47 uint16_t __cachemode2pte_tbl[_PAGE_CACHE_MODE_NUM] = {
48 	[_PAGE_CACHE_MODE_WB      ]	= 0         | 0        ,
49 	[_PAGE_CACHE_MODE_WC      ]	= 0         | _PAGE_PCD,
50 	[_PAGE_CACHE_MODE_UC_MINUS]	= 0         | _PAGE_PCD,
51 	[_PAGE_CACHE_MODE_UC      ]	= _PAGE_PWT | _PAGE_PCD,
52 	[_PAGE_CACHE_MODE_WT      ]	= 0         | _PAGE_PCD,
53 	[_PAGE_CACHE_MODE_WP      ]	= 0         | _PAGE_PCD,
54 };
55 EXPORT_SYMBOL(__cachemode2pte_tbl);
56 
57 uint8_t __pte2cachemode_tbl[8] = {
58 	[__pte2cm_idx( 0        | 0         | 0        )] = _PAGE_CACHE_MODE_WB,
59 	[__pte2cm_idx(_PAGE_PWT | 0         | 0        )] = _PAGE_CACHE_MODE_UC_MINUS,
60 	[__pte2cm_idx( 0        | _PAGE_PCD | 0        )] = _PAGE_CACHE_MODE_UC_MINUS,
61 	[__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | 0        )] = _PAGE_CACHE_MODE_UC,
62 	[__pte2cm_idx( 0        | 0         | _PAGE_PAT)] = _PAGE_CACHE_MODE_WB,
63 	[__pte2cm_idx(_PAGE_PWT | 0         | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
64 	[__pte2cm_idx(0         | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
65 	[__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC,
66 };
67 EXPORT_SYMBOL(__pte2cachemode_tbl);
68 
69 static unsigned long __initdata pgt_buf_start;
70 static unsigned long __initdata pgt_buf_end;
71 static unsigned long __initdata pgt_buf_top;
72 
73 static unsigned long min_pfn_mapped;
74 
75 static bool __initdata can_use_brk_pgt = true;
76 
77 /*
78  * Pages returned are already directly mapped.
79  *
80  * Changing that is likely to break Xen, see commit:
81  *
82  *    279b706 x86,xen: introduce x86_init.mapping.pagetable_reserve
83  *
84  * for detailed information.
85  */
86 __ref void *alloc_low_pages(unsigned int num)
87 {
88 	unsigned long pfn;
89 	int i;
90 
91 	if (after_bootmem) {
92 		unsigned int order;
93 
94 		order = get_order((unsigned long)num << PAGE_SHIFT);
95 		return (void *)__get_free_pages(GFP_ATOMIC | __GFP_ZERO, order);
96 	}
97 
98 	if ((pgt_buf_end + num) > pgt_buf_top || !can_use_brk_pgt) {
99 		unsigned long ret;
100 		if (min_pfn_mapped >= max_pfn_mapped)
101 			panic("alloc_low_pages: ran out of memory");
102 		ret = memblock_find_in_range(min_pfn_mapped << PAGE_SHIFT,
103 					max_pfn_mapped << PAGE_SHIFT,
104 					PAGE_SIZE * num , PAGE_SIZE);
105 		if (!ret)
106 			panic("alloc_low_pages: can not alloc memory");
107 		memblock_reserve(ret, PAGE_SIZE * num);
108 		pfn = ret >> PAGE_SHIFT;
109 	} else {
110 		pfn = pgt_buf_end;
111 		pgt_buf_end += num;
112 		printk(KERN_DEBUG "BRK [%#010lx, %#010lx] PGTABLE\n",
113 			pfn << PAGE_SHIFT, (pgt_buf_end << PAGE_SHIFT) - 1);
114 	}
115 
116 	for (i = 0; i < num; i++) {
117 		void *adr;
118 
119 		adr = __va((pfn + i) << PAGE_SHIFT);
120 		clear_page(adr);
121 	}
122 
123 	return __va(pfn << PAGE_SHIFT);
124 }
125 
126 /*
127  * By default need 3 4k for initial PMD_SIZE,  3 4k for 0-ISA_END_ADDRESS.
128  * With KASLR memory randomization, depending on the machine e820 memory
129  * and the PUD alignment. We may need twice more pages when KASLR memory
130  * randomization is enabled.
131  */
132 #ifndef CONFIG_RANDOMIZE_MEMORY
133 #define INIT_PGD_PAGE_COUNT      6
134 #else
135 #define INIT_PGD_PAGE_COUNT      12
136 #endif
137 #define INIT_PGT_BUF_SIZE	(INIT_PGD_PAGE_COUNT * PAGE_SIZE)
138 RESERVE_BRK(early_pgt_alloc, INIT_PGT_BUF_SIZE);
139 void  __init early_alloc_pgt_buf(void)
140 {
141 	unsigned long tables = INIT_PGT_BUF_SIZE;
142 	phys_addr_t base;
143 
144 	base = __pa(extend_brk(tables, PAGE_SIZE));
145 
146 	pgt_buf_start = base >> PAGE_SHIFT;
147 	pgt_buf_end = pgt_buf_start;
148 	pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
149 }
150 
151 int after_bootmem;
152 
153 early_param_on_off("gbpages", "nogbpages", direct_gbpages, CONFIG_X86_DIRECT_GBPAGES);
154 
155 struct map_range {
156 	unsigned long start;
157 	unsigned long end;
158 	unsigned page_size_mask;
159 };
160 
161 static int page_size_mask;
162 
163 static void __init probe_page_size_mask(void)
164 {
165 	/*
166 	 * For pagealloc debugging, identity mapping will use small pages.
167 	 * This will simplify cpa(), which otherwise needs to support splitting
168 	 * large pages into small in interrupt context, etc.
169 	 */
170 	if (boot_cpu_has(X86_FEATURE_PSE) && !debug_pagealloc_enabled())
171 		page_size_mask |= 1 << PG_LEVEL_2M;
172 	else
173 		direct_gbpages = 0;
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 static void setup_pcid(void)
196 {
197 #ifdef CONFIG_X86_64
198 	if (boot_cpu_has(X86_FEATURE_PCID)) {
199 		if (boot_cpu_has(X86_FEATURE_PGE)) {
200 			/*
201 			 * This can't be cr4_set_bits_and_update_boot() --
202 			 * the trampoline code can't handle CR4.PCIDE and
203 			 * it wouldn't do any good anyway.  Despite the name,
204 			 * cr4_set_bits_and_update_boot() doesn't actually
205 			 * cause the bits in question to remain set all the
206 			 * way through the secondary boot asm.
207 			 *
208 			 * Instead, we brute-force it and set CR4.PCIDE
209 			 * manually in start_secondary().
210 			 */
211 			cr4_set_bits(X86_CR4_PCIDE);
212 		} else {
213 			/*
214 			 * flush_tlb_all(), as currently implemented, won't
215 			 * work if PCID is on but PGE is not.  Since that
216 			 * combination doesn't exist on real hardware, there's
217 			 * no reason to try to fully support it, but it's
218 			 * polite to avoid corrupting data if we're on
219 			 * an improperly configured VM.
220 			 */
221 			setup_clear_cpu_cap(X86_FEATURE_PCID);
222 		}
223 	}
224 #endif
225 }
226 
227 #ifdef CONFIG_X86_32
228 #define NR_RANGE_MR 3
229 #else /* CONFIG_X86_64 */
230 #define NR_RANGE_MR 5
231 #endif
232 
233 static int __meminit save_mr(struct map_range *mr, int nr_range,
234 			     unsigned long start_pfn, unsigned long end_pfn,
235 			     unsigned long page_size_mask)
236 {
237 	if (start_pfn < end_pfn) {
238 		if (nr_range >= NR_RANGE_MR)
239 			panic("run out of range for init_memory_mapping\n");
240 		mr[nr_range].start = start_pfn<<PAGE_SHIFT;
241 		mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
242 		mr[nr_range].page_size_mask = page_size_mask;
243 		nr_range++;
244 	}
245 
246 	return nr_range;
247 }
248 
249 /*
250  * adjust the page_size_mask for small range to go with
251  *	big page size instead small one if nearby are ram too.
252  */
253 static void __ref adjust_range_page_size_mask(struct map_range *mr,
254 							 int nr_range)
255 {
256 	int i;
257 
258 	for (i = 0; i < nr_range; i++) {
259 		if ((page_size_mask & (1<<PG_LEVEL_2M)) &&
260 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) {
261 			unsigned long start = round_down(mr[i].start, PMD_SIZE);
262 			unsigned long end = round_up(mr[i].end, PMD_SIZE);
263 
264 #ifdef CONFIG_X86_32
265 			if ((end >> PAGE_SHIFT) > max_low_pfn)
266 				continue;
267 #endif
268 
269 			if (memblock_is_region_memory(start, end - start))
270 				mr[i].page_size_mask |= 1<<PG_LEVEL_2M;
271 		}
272 		if ((page_size_mask & (1<<PG_LEVEL_1G)) &&
273 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) {
274 			unsigned long start = round_down(mr[i].start, PUD_SIZE);
275 			unsigned long end = round_up(mr[i].end, PUD_SIZE);
276 
277 			if (memblock_is_region_memory(start, end - start))
278 				mr[i].page_size_mask |= 1<<PG_LEVEL_1G;
279 		}
280 	}
281 }
282 
283 static const char *page_size_string(struct map_range *mr)
284 {
285 	static const char str_1g[] = "1G";
286 	static const char str_2m[] = "2M";
287 	static const char str_4m[] = "4M";
288 	static const char str_4k[] = "4k";
289 
290 	if (mr->page_size_mask & (1<<PG_LEVEL_1G))
291 		return str_1g;
292 	/*
293 	 * 32-bit without PAE has a 4M large page size.
294 	 * PG_LEVEL_2M is misnamed, but we can at least
295 	 * print out the right size in the string.
296 	 */
297 	if (IS_ENABLED(CONFIG_X86_32) &&
298 	    !IS_ENABLED(CONFIG_X86_PAE) &&
299 	    mr->page_size_mask & (1<<PG_LEVEL_2M))
300 		return str_4m;
301 
302 	if (mr->page_size_mask & (1<<PG_LEVEL_2M))
303 		return str_2m;
304 
305 	return str_4k;
306 }
307 
308 static int __meminit split_mem_range(struct map_range *mr, int nr_range,
309 				     unsigned long start,
310 				     unsigned long end)
311 {
312 	unsigned long start_pfn, end_pfn, limit_pfn;
313 	unsigned long pfn;
314 	int i;
315 
316 	limit_pfn = PFN_DOWN(end);
317 
318 	/* head if not big page alignment ? */
319 	pfn = start_pfn = PFN_DOWN(start);
320 #ifdef CONFIG_X86_32
321 	/*
322 	 * Don't use a large page for the first 2/4MB of memory
323 	 * because there are often fixed size MTRRs in there
324 	 * and overlapping MTRRs into large pages can cause
325 	 * slowdowns.
326 	 */
327 	if (pfn == 0)
328 		end_pfn = PFN_DOWN(PMD_SIZE);
329 	else
330 		end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
331 #else /* CONFIG_X86_64 */
332 	end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
333 #endif
334 	if (end_pfn > limit_pfn)
335 		end_pfn = limit_pfn;
336 	if (start_pfn < end_pfn) {
337 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
338 		pfn = end_pfn;
339 	}
340 
341 	/* big page (2M) range */
342 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
343 #ifdef CONFIG_X86_32
344 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
345 #else /* CONFIG_X86_64 */
346 	end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
347 	if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE)))
348 		end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
349 #endif
350 
351 	if (start_pfn < end_pfn) {
352 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
353 				page_size_mask & (1<<PG_LEVEL_2M));
354 		pfn = end_pfn;
355 	}
356 
357 #ifdef CONFIG_X86_64
358 	/* big page (1G) range */
359 	start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
360 	end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE));
361 	if (start_pfn < end_pfn) {
362 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
363 				page_size_mask &
364 				 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
365 		pfn = end_pfn;
366 	}
367 
368 	/* tail is not big page (1G) alignment */
369 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
370 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
371 	if (start_pfn < end_pfn) {
372 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
373 				page_size_mask & (1<<PG_LEVEL_2M));
374 		pfn = end_pfn;
375 	}
376 #endif
377 
378 	/* tail is not big page (2M) alignment */
379 	start_pfn = pfn;
380 	end_pfn = limit_pfn;
381 	nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
382 
383 	if (!after_bootmem)
384 		adjust_range_page_size_mask(mr, nr_range);
385 
386 	/* try to merge same page size and continuous */
387 	for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
388 		unsigned long old_start;
389 		if (mr[i].end != mr[i+1].start ||
390 		    mr[i].page_size_mask != mr[i+1].page_size_mask)
391 			continue;
392 		/* move it */
393 		old_start = mr[i].start;
394 		memmove(&mr[i], &mr[i+1],
395 			(nr_range - 1 - i) * sizeof(struct map_range));
396 		mr[i--].start = old_start;
397 		nr_range--;
398 	}
399 
400 	for (i = 0; i < nr_range; i++)
401 		pr_debug(" [mem %#010lx-%#010lx] page %s\n",
402 				mr[i].start, mr[i].end - 1,
403 				page_size_string(&mr[i]));
404 
405 	return nr_range;
406 }
407 
408 struct range pfn_mapped[E820_MAX_ENTRIES];
409 int nr_pfn_mapped;
410 
411 static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn)
412 {
413 	nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_MAX_ENTRIES,
414 					     nr_pfn_mapped, start_pfn, end_pfn);
415 	nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_MAX_ENTRIES);
416 
417 	max_pfn_mapped = max(max_pfn_mapped, end_pfn);
418 
419 	if (start_pfn < (1UL<<(32-PAGE_SHIFT)))
420 		max_low_pfn_mapped = max(max_low_pfn_mapped,
421 					 min(end_pfn, 1UL<<(32-PAGE_SHIFT)));
422 }
423 
424 bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn)
425 {
426 	int i;
427 
428 	for (i = 0; i < nr_pfn_mapped; i++)
429 		if ((start_pfn >= pfn_mapped[i].start) &&
430 		    (end_pfn <= pfn_mapped[i].end))
431 			return true;
432 
433 	return false;
434 }
435 
436 /*
437  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
438  * This runs before bootmem is initialized and gets pages directly from
439  * the physical memory. To access them they are temporarily mapped.
440  */
441 unsigned long __ref init_memory_mapping(unsigned long start,
442 					       unsigned long end)
443 {
444 	struct map_range mr[NR_RANGE_MR];
445 	unsigned long ret = 0;
446 	int nr_range, i;
447 
448 	pr_debug("init_memory_mapping: [mem %#010lx-%#010lx]\n",
449 	       start, end - 1);
450 
451 	memset(mr, 0, sizeof(mr));
452 	nr_range = split_mem_range(mr, 0, start, end);
453 
454 	for (i = 0; i < nr_range; i++)
455 		ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
456 						   mr[i].page_size_mask);
457 
458 	add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT);
459 
460 	return ret >> PAGE_SHIFT;
461 }
462 
463 /*
464  * We need to iterate through the E820 memory map and create direct mappings
465  * for only E820_TYPE_RAM and E820_KERN_RESERVED regions. We cannot simply
466  * create direct mappings for all pfns from [0 to max_low_pfn) and
467  * [4GB to max_pfn) because of possible memory holes in high addresses
468  * that cannot be marked as UC by fixed/variable range MTRRs.
469  * Depending on the alignment of E820 ranges, this may possibly result
470  * in using smaller size (i.e. 4K instead of 2M or 1G) page tables.
471  *
472  * init_mem_mapping() calls init_range_memory_mapping() with big range.
473  * That range would have hole in the middle or ends, and only ram parts
474  * will be mapped in init_range_memory_mapping().
475  */
476 static unsigned long __init init_range_memory_mapping(
477 					   unsigned long r_start,
478 					   unsigned long r_end)
479 {
480 	unsigned long start_pfn, end_pfn;
481 	unsigned long mapped_ram_size = 0;
482 	int i;
483 
484 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
485 		u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end);
486 		u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end);
487 		if (start >= end)
488 			continue;
489 
490 		/*
491 		 * if it is overlapping with brk pgt, we need to
492 		 * alloc pgt buf from memblock instead.
493 		 */
494 		can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >=
495 				    min(end, (u64)pgt_buf_top<<PAGE_SHIFT);
496 		init_memory_mapping(start, end);
497 		mapped_ram_size += end - start;
498 		can_use_brk_pgt = true;
499 	}
500 
501 	return mapped_ram_size;
502 }
503 
504 static unsigned long __init get_new_step_size(unsigned long step_size)
505 {
506 	/*
507 	 * Initial mapped size is PMD_SIZE (2M).
508 	 * We can not set step_size to be PUD_SIZE (1G) yet.
509 	 * In worse case, when we cross the 1G boundary, and
510 	 * PG_LEVEL_2M is not set, we will need 1+1+512 pages (2M + 8k)
511 	 * to map 1G range with PTE. Hence we use one less than the
512 	 * difference of page table level shifts.
513 	 *
514 	 * Don't need to worry about overflow in the top-down case, on 32bit,
515 	 * when step_size is 0, round_down() returns 0 for start, and that
516 	 * turns it into 0x100000000ULL.
517 	 * In the bottom-up case, round_up(x, 0) returns 0 though too, which
518 	 * needs to be taken into consideration by the code below.
519 	 */
520 	return step_size << (PMD_SHIFT - PAGE_SHIFT - 1);
521 }
522 
523 /**
524  * memory_map_top_down - Map [map_start, map_end) top down
525  * @map_start: start address of the target memory range
526  * @map_end: end address of the target memory range
527  *
528  * This function will setup direct mapping for memory range
529  * [map_start, map_end) in top-down. That said, the page tables
530  * will be allocated at the end of the memory, and we map the
531  * memory in top-down.
532  */
533 static void __init memory_map_top_down(unsigned long map_start,
534 				       unsigned long map_end)
535 {
536 	unsigned long real_end, start, last_start;
537 	unsigned long step_size;
538 	unsigned long addr;
539 	unsigned long mapped_ram_size = 0;
540 
541 	/* xen has big range in reserved near end of ram, skip it at first.*/
542 	addr = memblock_find_in_range(map_start, map_end, PMD_SIZE, PMD_SIZE);
543 	real_end = addr + PMD_SIZE;
544 
545 	/* step_size need to be small so pgt_buf from BRK could cover it */
546 	step_size = PMD_SIZE;
547 	max_pfn_mapped = 0; /* will get exact value next */
548 	min_pfn_mapped = real_end >> PAGE_SHIFT;
549 	last_start = start = real_end;
550 
551 	/*
552 	 * We start from the top (end of memory) and go to the bottom.
553 	 * The memblock_find_in_range() gets us a block of RAM from the
554 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
555 	 * for page table.
556 	 */
557 	while (last_start > map_start) {
558 		if (last_start > step_size) {
559 			start = round_down(last_start - 1, step_size);
560 			if (start < map_start)
561 				start = map_start;
562 		} else
563 			start = map_start;
564 		mapped_ram_size += init_range_memory_mapping(start,
565 							last_start);
566 		last_start = start;
567 		min_pfn_mapped = last_start >> PAGE_SHIFT;
568 		if (mapped_ram_size >= step_size)
569 			step_size = get_new_step_size(step_size);
570 	}
571 
572 	if (real_end < map_end)
573 		init_range_memory_mapping(real_end, map_end);
574 }
575 
576 /**
577  * memory_map_bottom_up - Map [map_start, map_end) bottom up
578  * @map_start: start address of the target memory range
579  * @map_end: end address of the target memory range
580  *
581  * This function will setup direct mapping for memory range
582  * [map_start, map_end) in bottom-up. Since we have limited the
583  * bottom-up allocation above the kernel, the page tables will
584  * be allocated just above the kernel and we map the memory
585  * in [map_start, map_end) in bottom-up.
586  */
587 static void __init memory_map_bottom_up(unsigned long map_start,
588 					unsigned long map_end)
589 {
590 	unsigned long next, start;
591 	unsigned long mapped_ram_size = 0;
592 	/* step_size need to be small so pgt_buf from BRK could cover it */
593 	unsigned long step_size = PMD_SIZE;
594 
595 	start = map_start;
596 	min_pfn_mapped = start >> PAGE_SHIFT;
597 
598 	/*
599 	 * We start from the bottom (@map_start) and go to the top (@map_end).
600 	 * The memblock_find_in_range() gets us a block of RAM from the
601 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
602 	 * for page table.
603 	 */
604 	while (start < map_end) {
605 		if (step_size && map_end - start > step_size) {
606 			next = round_up(start + 1, step_size);
607 			if (next > map_end)
608 				next = map_end;
609 		} else {
610 			next = map_end;
611 		}
612 
613 		mapped_ram_size += init_range_memory_mapping(start, next);
614 		start = next;
615 
616 		if (mapped_ram_size >= step_size)
617 			step_size = get_new_step_size(step_size);
618 	}
619 }
620 
621 void __init init_mem_mapping(void)
622 {
623 	unsigned long end;
624 
625 	probe_page_size_mask();
626 	setup_pcid();
627 
628 #ifdef CONFIG_X86_64
629 	end = max_pfn << PAGE_SHIFT;
630 #else
631 	end = max_low_pfn << PAGE_SHIFT;
632 #endif
633 
634 	/* the ISA range is always mapped regardless of memory holes */
635 	init_memory_mapping(0, ISA_END_ADDRESS);
636 
637 	/* Init the trampoline, possibly with KASLR memory offset */
638 	init_trampoline();
639 
640 	/*
641 	 * If the allocation is in bottom-up direction, we setup direct mapping
642 	 * in bottom-up, otherwise we setup direct mapping in top-down.
643 	 */
644 	if (memblock_bottom_up()) {
645 		unsigned long kernel_end = __pa_symbol(_end);
646 
647 		/*
648 		 * we need two separate calls here. This is because we want to
649 		 * allocate page tables above the kernel. So we first map
650 		 * [kernel_end, end) to make memory above the kernel be mapped
651 		 * as soon as possible. And then use page tables allocated above
652 		 * the kernel to map [ISA_END_ADDRESS, kernel_end).
653 		 */
654 		memory_map_bottom_up(kernel_end, end);
655 		memory_map_bottom_up(ISA_END_ADDRESS, kernel_end);
656 	} else {
657 		memory_map_top_down(ISA_END_ADDRESS, end);
658 	}
659 
660 #ifdef CONFIG_X86_64
661 	if (max_pfn > max_low_pfn) {
662 		/* can we preseve max_low_pfn ?*/
663 		max_low_pfn = max_pfn;
664 	}
665 #else
666 	early_ioremap_page_table_range_init();
667 #endif
668 
669 	load_cr3(swapper_pg_dir);
670 	__flush_tlb_all();
671 
672 	x86_init.hyper.init_mem_mapping();
673 
674 	early_memtest(0, max_pfn_mapped << PAGE_SHIFT);
675 }
676 
677 /*
678  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
679  * is valid. The argument is a physical page number.
680  *
681  * On x86, access has to be given to the first megabyte of RAM because that
682  * area traditionally contains BIOS code and data regions used by X, dosemu,
683  * and similar apps. Since they map the entire memory range, the whole range
684  * must be allowed (for mapping), but any areas that would otherwise be
685  * disallowed are flagged as being "zero filled" instead of rejected.
686  * Access has to be given to non-kernel-ram areas as well, these contain the
687  * PCI mmio resources as well as potential bios/acpi data regions.
688  */
689 int devmem_is_allowed(unsigned long pagenr)
690 {
691 	if (page_is_ram(pagenr)) {
692 		/*
693 		 * For disallowed memory regions in the low 1MB range,
694 		 * request that the page be shown as all zeros.
695 		 */
696 		if (pagenr < 256)
697 			return 2;
698 
699 		return 0;
700 	}
701 
702 	/*
703 	 * This must follow RAM test, since System RAM is considered a
704 	 * restricted resource under CONFIG_STRICT_IOMEM.
705 	 */
706 	if (iomem_is_exclusive(pagenr << PAGE_SHIFT)) {
707 		/* Low 1MB bypasses iomem restrictions. */
708 		if (pagenr < 256)
709 			return 1;
710 
711 		return 0;
712 	}
713 
714 	return 1;
715 }
716 
717 void free_init_pages(char *what, unsigned long begin, unsigned long end)
718 {
719 	unsigned long begin_aligned, end_aligned;
720 
721 	/* Make sure boundaries are page aligned */
722 	begin_aligned = PAGE_ALIGN(begin);
723 	end_aligned   = end & PAGE_MASK;
724 
725 	if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
726 		begin = begin_aligned;
727 		end   = end_aligned;
728 	}
729 
730 	if (begin >= end)
731 		return;
732 
733 	/*
734 	 * If debugging page accesses then do not free this memory but
735 	 * mark them not present - any buggy init-section access will
736 	 * create a kernel page fault:
737 	 */
738 	if (debug_pagealloc_enabled()) {
739 		pr_info("debug: unmapping init [mem %#010lx-%#010lx]\n",
740 			begin, end - 1);
741 		set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
742 	} else {
743 		/*
744 		 * We just marked the kernel text read only above, now that
745 		 * we are going to free part of that, we need to make that
746 		 * writeable and non-executable first.
747 		 */
748 		set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
749 		set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
750 
751 		free_reserved_area((void *)begin, (void *)end,
752 				   POISON_FREE_INITMEM, what);
753 	}
754 }
755 
756 void __ref free_initmem(void)
757 {
758 	e820__reallocate_tables();
759 
760 	free_init_pages("unused kernel",
761 			(unsigned long)(&__init_begin),
762 			(unsigned long)(&__init_end));
763 }
764 
765 #ifdef CONFIG_BLK_DEV_INITRD
766 void __init free_initrd_mem(unsigned long start, unsigned long end)
767 {
768 	/*
769 	 * end could be not aligned, and We can not align that,
770 	 * decompresser could be confused by aligned initrd_end
771 	 * We already reserve the end partial page before in
772 	 *   - i386_start_kernel()
773 	 *   - x86_64_start_kernel()
774 	 *   - relocate_initrd()
775 	 * So here We can do PAGE_ALIGN() safely to get partial page to be freed
776 	 */
777 	free_init_pages("initrd", start, PAGE_ALIGN(end));
778 }
779 #endif
780 
781 /*
782  * Calculate the precise size of the DMA zone (first 16 MB of RAM),
783  * and pass it to the MM layer - to help it set zone watermarks more
784  * accurately.
785  *
786  * Done on 64-bit systems only for the time being, although 32-bit systems
787  * might benefit from this as well.
788  */
789 void __init memblock_find_dma_reserve(void)
790 {
791 #ifdef CONFIG_X86_64
792 	u64 nr_pages = 0, nr_free_pages = 0;
793 	unsigned long start_pfn, end_pfn;
794 	phys_addr_t start_addr, end_addr;
795 	int i;
796 	u64 u;
797 
798 	/*
799 	 * Iterate over all memory ranges (free and reserved ones alike),
800 	 * to calculate the total number of pages in the first 16 MB of RAM:
801 	 */
802 	nr_pages = 0;
803 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
804 		start_pfn = min(start_pfn, MAX_DMA_PFN);
805 		end_pfn   = min(end_pfn,   MAX_DMA_PFN);
806 
807 		nr_pages += end_pfn - start_pfn;
808 	}
809 
810 	/*
811 	 * Iterate over free memory ranges to calculate the number of free
812 	 * pages in the DMA zone, while not counting potential partial
813 	 * pages at the beginning or the end of the range:
814 	 */
815 	nr_free_pages = 0;
816 	for_each_free_mem_range(u, NUMA_NO_NODE, MEMBLOCK_NONE, &start_addr, &end_addr, NULL) {
817 		start_pfn = min_t(unsigned long, PFN_UP(start_addr), MAX_DMA_PFN);
818 		end_pfn   = min_t(unsigned long, PFN_DOWN(end_addr), MAX_DMA_PFN);
819 
820 		if (start_pfn < end_pfn)
821 			nr_free_pages += end_pfn - start_pfn;
822 	}
823 
824 	set_dma_reserve(nr_pages - nr_free_pages);
825 #endif
826 }
827 
828 void __init zone_sizes_init(void)
829 {
830 	unsigned long max_zone_pfns[MAX_NR_ZONES];
831 
832 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
833 
834 #ifdef CONFIG_ZONE_DMA
835 	max_zone_pfns[ZONE_DMA]		= min(MAX_DMA_PFN, max_low_pfn);
836 #endif
837 #ifdef CONFIG_ZONE_DMA32
838 	max_zone_pfns[ZONE_DMA32]	= min(MAX_DMA32_PFN, max_low_pfn);
839 #endif
840 	max_zone_pfns[ZONE_NORMAL]	= max_low_pfn;
841 #ifdef CONFIG_HIGHMEM
842 	max_zone_pfns[ZONE_HIGHMEM]	= max_pfn;
843 #endif
844 
845 	free_area_init_nodes(max_zone_pfns);
846 }
847 
848 DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate) = {
849 	.loaded_mm = &init_mm,
850 	.next_asid = 1,
851 	.cr4 = ~0UL,	/* fail hard if we screw up cr4 shadow initialization */
852 };
853 EXPORT_SYMBOL_GPL(cpu_tlbstate);
854 
855 void update_cache_mode_entry(unsigned entry, enum page_cache_mode cache)
856 {
857 	/* entry 0 MUST be WB (hardwired to speed up translations) */
858 	BUG_ON(!entry && cache != _PAGE_CACHE_MODE_WB);
859 
860 	__cachemode2pte_tbl[cache] = __cm_idx2pte(entry);
861 	__pte2cachemode_tbl[entry] = cache;
862 }
863