xref: /linux/arch/x86/mm/init.c (revision cea0f76a483d1270ac6f6513964e3e75193dda48)
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/swapfile.h>
7 #include <linux/swapops.h>
8 #include <linux/kmemleak.h>
9 #include <linux/sched/task.h>
10 
11 #include <asm/set_memory.h>
12 #include <asm/e820/api.h>
13 #include <asm/init.h>
14 #include <asm/page.h>
15 #include <asm/page_types.h>
16 #include <asm/sections.h>
17 #include <asm/setup.h>
18 #include <asm/tlbflush.h>
19 #include <asm/tlb.h>
20 #include <asm/proto.h>
21 #include <asm/dma.h>		/* for MAX_DMA_PFN */
22 #include <asm/microcode.h>
23 #include <asm/kaslr.h>
24 #include <asm/hypervisor.h>
25 #include <asm/cpufeature.h>
26 #include <asm/pti.h>
27 #include <asm/text-patching.h>
28 
29 /*
30  * We need to define the tracepoints somewhere, and tlb.c
31  * is only compied when SMP=y.
32  */
33 #define CREATE_TRACE_POINTS
34 #include <trace/events/tlb.h>
35 
36 #include "mm_internal.h"
37 
38 /*
39  * Tables translating between page_cache_type_t and pte encoding.
40  *
41  * The default values are defined statically as minimal supported mode;
42  * WC and WT fall back to UC-.  pat_init() updates these values to support
43  * more cache modes, WC and WT, when it is safe to do so.  See pat_init()
44  * for the details.  Note, __early_ioremap() used during early boot-time
45  * takes pgprot_t (pte encoding) and does not use these tables.
46  *
47  *   Index into __cachemode2pte_tbl[] is the cachemode.
48  *
49  *   Index into __pte2cachemode_tbl[] are the caching attribute bits of the pte
50  *   (_PAGE_PWT, _PAGE_PCD, _PAGE_PAT) at index bit positions 0, 1, 2.
51  */
52 static uint16_t __cachemode2pte_tbl[_PAGE_CACHE_MODE_NUM] = {
53 	[_PAGE_CACHE_MODE_WB      ]	= 0         | 0        ,
54 	[_PAGE_CACHE_MODE_WC      ]	= 0         | _PAGE_PCD,
55 	[_PAGE_CACHE_MODE_UC_MINUS]	= 0         | _PAGE_PCD,
56 	[_PAGE_CACHE_MODE_UC      ]	= _PAGE_PWT | _PAGE_PCD,
57 	[_PAGE_CACHE_MODE_WT      ]	= 0         | _PAGE_PCD,
58 	[_PAGE_CACHE_MODE_WP      ]	= 0         | _PAGE_PCD,
59 };
60 
61 unsigned long cachemode2protval(enum page_cache_mode pcm)
62 {
63 	if (likely(pcm == 0))
64 		return 0;
65 	return __cachemode2pte_tbl[pcm];
66 }
67 EXPORT_SYMBOL(cachemode2protval);
68 
69 static uint8_t __pte2cachemode_tbl[8] = {
70 	[__pte2cm_idx( 0        | 0         | 0        )] = _PAGE_CACHE_MODE_WB,
71 	[__pte2cm_idx(_PAGE_PWT | 0         | 0        )] = _PAGE_CACHE_MODE_UC_MINUS,
72 	[__pte2cm_idx( 0        | _PAGE_PCD | 0        )] = _PAGE_CACHE_MODE_UC_MINUS,
73 	[__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | 0        )] = _PAGE_CACHE_MODE_UC,
74 	[__pte2cm_idx( 0        | 0         | _PAGE_PAT)] = _PAGE_CACHE_MODE_WB,
75 	[__pte2cm_idx(_PAGE_PWT | 0         | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
76 	[__pte2cm_idx(0         | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
77 	[__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC,
78 };
79 
80 /* Check that the write-protect PAT entry is set for write-protect */
81 bool x86_has_pat_wp(void)
82 {
83 	return __pte2cachemode_tbl[_PAGE_CACHE_MODE_WP] == _PAGE_CACHE_MODE_WP;
84 }
85 
86 enum page_cache_mode pgprot2cachemode(pgprot_t pgprot)
87 {
88 	unsigned long masked;
89 
90 	masked = pgprot_val(pgprot) & _PAGE_CACHE_MASK;
91 	if (likely(masked == 0))
92 		return 0;
93 	return __pte2cachemode_tbl[__pte2cm_idx(masked)];
94 }
95 
96 static unsigned long __initdata pgt_buf_start;
97 static unsigned long __initdata pgt_buf_end;
98 static unsigned long __initdata pgt_buf_top;
99 
100 static unsigned long min_pfn_mapped;
101 
102 static bool __initdata can_use_brk_pgt = true;
103 
104 /*
105  * Pages returned are already directly mapped.
106  *
107  * Changing that is likely to break Xen, see commit:
108  *
109  *    279b706 x86,xen: introduce x86_init.mapping.pagetable_reserve
110  *
111  * for detailed information.
112  */
113 __ref void *alloc_low_pages(unsigned int num)
114 {
115 	unsigned long pfn;
116 	int i;
117 
118 	if (after_bootmem) {
119 		unsigned int order;
120 
121 		order = get_order((unsigned long)num << PAGE_SHIFT);
122 		return (void *)__get_free_pages(GFP_ATOMIC | __GFP_ZERO, order);
123 	}
124 
125 	if ((pgt_buf_end + num) > pgt_buf_top || !can_use_brk_pgt) {
126 		unsigned long ret = 0;
127 
128 		if (min_pfn_mapped < max_pfn_mapped) {
129 			ret = memblock_find_in_range(
130 					min_pfn_mapped << PAGE_SHIFT,
131 					max_pfn_mapped << PAGE_SHIFT,
132 					PAGE_SIZE * num , PAGE_SIZE);
133 		}
134 		if (ret)
135 			memblock_reserve(ret, PAGE_SIZE * num);
136 		else if (can_use_brk_pgt)
137 			ret = __pa(extend_brk(PAGE_SIZE * num, PAGE_SIZE));
138 
139 		if (!ret)
140 			panic("alloc_low_pages: can not alloc memory");
141 
142 		pfn = ret >> PAGE_SHIFT;
143 	} else {
144 		pfn = pgt_buf_end;
145 		pgt_buf_end += num;
146 	}
147 
148 	for (i = 0; i < num; i++) {
149 		void *adr;
150 
151 		adr = __va((pfn + i) << PAGE_SHIFT);
152 		clear_page(adr);
153 	}
154 
155 	return __va(pfn << PAGE_SHIFT);
156 }
157 
158 /*
159  * By default need 3 4k for initial PMD_SIZE,  3 4k for 0-ISA_END_ADDRESS.
160  * With KASLR memory randomization, depending on the machine e820 memory
161  * and the PUD alignment. We may need twice more pages when KASLR memory
162  * randomization is enabled.
163  */
164 #ifndef CONFIG_RANDOMIZE_MEMORY
165 #define INIT_PGD_PAGE_COUNT      6
166 #else
167 #define INIT_PGD_PAGE_COUNT      12
168 #endif
169 #define INIT_PGT_BUF_SIZE	(INIT_PGD_PAGE_COUNT * PAGE_SIZE)
170 RESERVE_BRK(early_pgt_alloc, INIT_PGT_BUF_SIZE);
171 void  __init early_alloc_pgt_buf(void)
172 {
173 	unsigned long tables = INIT_PGT_BUF_SIZE;
174 	phys_addr_t base;
175 
176 	base = __pa(extend_brk(tables, PAGE_SIZE));
177 
178 	pgt_buf_start = base >> PAGE_SHIFT;
179 	pgt_buf_end = pgt_buf_start;
180 	pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
181 }
182 
183 int after_bootmem;
184 
185 early_param_on_off("gbpages", "nogbpages", direct_gbpages, CONFIG_X86_DIRECT_GBPAGES);
186 
187 struct map_range {
188 	unsigned long start;
189 	unsigned long end;
190 	unsigned page_size_mask;
191 };
192 
193 static int page_size_mask;
194 
195 /*
196  * Save some of cr4 feature set we're using (e.g.  Pentium 4MB
197  * enable and PPro Global page enable), so that any CPU's that boot
198  * up after us can get the correct flags. Invoked on the boot CPU.
199  */
200 static inline void cr4_set_bits_and_update_boot(unsigned long mask)
201 {
202 	mmu_cr4_features |= mask;
203 	if (trampoline_cr4_features)
204 		*trampoline_cr4_features = mmu_cr4_features;
205 	cr4_set_bits(mask);
206 }
207 
208 static void __init probe_page_size_mask(void)
209 {
210 	/*
211 	 * For pagealloc debugging, identity mapping will use small pages.
212 	 * This will simplify cpa(), which otherwise needs to support splitting
213 	 * large pages into small in interrupt context, etc.
214 	 */
215 	if (boot_cpu_has(X86_FEATURE_PSE) && !debug_pagealloc_enabled())
216 		page_size_mask |= 1 << PG_LEVEL_2M;
217 	else
218 		direct_gbpages = 0;
219 
220 	/* Enable PSE if available */
221 	if (boot_cpu_has(X86_FEATURE_PSE))
222 		cr4_set_bits_and_update_boot(X86_CR4_PSE);
223 
224 	/* Enable PGE if available */
225 	__supported_pte_mask &= ~_PAGE_GLOBAL;
226 	if (boot_cpu_has(X86_FEATURE_PGE)) {
227 		cr4_set_bits_and_update_boot(X86_CR4_PGE);
228 		__supported_pte_mask |= _PAGE_GLOBAL;
229 	}
230 
231 	/* By the default is everything supported: */
232 	__default_kernel_pte_mask = __supported_pte_mask;
233 	/* Except when with PTI where the kernel is mostly non-Global: */
234 	if (cpu_feature_enabled(X86_FEATURE_PTI))
235 		__default_kernel_pte_mask &= ~_PAGE_GLOBAL;
236 
237 	/* Enable 1 GB linear kernel mappings if available: */
238 	if (direct_gbpages && boot_cpu_has(X86_FEATURE_GBPAGES)) {
239 		printk(KERN_INFO "Using GB pages for direct mapping\n");
240 		page_size_mask |= 1 << PG_LEVEL_1G;
241 	} else {
242 		direct_gbpages = 0;
243 	}
244 }
245 
246 static void setup_pcid(void)
247 {
248 	if (!IS_ENABLED(CONFIG_X86_64))
249 		return;
250 
251 	if (!boot_cpu_has(X86_FEATURE_PCID))
252 		return;
253 
254 	if (boot_cpu_has(X86_FEATURE_PGE)) {
255 		/*
256 		 * This can't be cr4_set_bits_and_update_boot() -- the
257 		 * trampoline code can't handle CR4.PCIDE and it wouldn't
258 		 * do any good anyway.  Despite the name,
259 		 * cr4_set_bits_and_update_boot() doesn't actually cause
260 		 * the bits in question to remain set all the way through
261 		 * the secondary boot asm.
262 		 *
263 		 * Instead, we brute-force it and set CR4.PCIDE manually in
264 		 * start_secondary().
265 		 */
266 		cr4_set_bits(X86_CR4_PCIDE);
267 
268 		/*
269 		 * INVPCID's single-context modes (2/3) only work if we set
270 		 * X86_CR4_PCIDE, *and* we INVPCID support.  It's unusable
271 		 * on systems that have X86_CR4_PCIDE clear, or that have
272 		 * no INVPCID support at all.
273 		 */
274 		if (boot_cpu_has(X86_FEATURE_INVPCID))
275 			setup_force_cpu_cap(X86_FEATURE_INVPCID_SINGLE);
276 	} else {
277 		/*
278 		 * flush_tlb_all(), as currently implemented, won't work if
279 		 * PCID is on but PGE is not.  Since that combination
280 		 * doesn't exist on real hardware, there's no reason to try
281 		 * to fully support it, but it's polite to avoid corrupting
282 		 * data if we're on an improperly configured VM.
283 		 */
284 		setup_clear_cpu_cap(X86_FEATURE_PCID);
285 	}
286 }
287 
288 #ifdef CONFIG_X86_32
289 #define NR_RANGE_MR 3
290 #else /* CONFIG_X86_64 */
291 #define NR_RANGE_MR 5
292 #endif
293 
294 static int __meminit save_mr(struct map_range *mr, int nr_range,
295 			     unsigned long start_pfn, unsigned long end_pfn,
296 			     unsigned long page_size_mask)
297 {
298 	if (start_pfn < end_pfn) {
299 		if (nr_range >= NR_RANGE_MR)
300 			panic("run out of range for init_memory_mapping\n");
301 		mr[nr_range].start = start_pfn<<PAGE_SHIFT;
302 		mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
303 		mr[nr_range].page_size_mask = page_size_mask;
304 		nr_range++;
305 	}
306 
307 	return nr_range;
308 }
309 
310 /*
311  * adjust the page_size_mask for small range to go with
312  *	big page size instead small one if nearby are ram too.
313  */
314 static void __ref adjust_range_page_size_mask(struct map_range *mr,
315 							 int nr_range)
316 {
317 	int i;
318 
319 	for (i = 0; i < nr_range; i++) {
320 		if ((page_size_mask & (1<<PG_LEVEL_2M)) &&
321 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) {
322 			unsigned long start = round_down(mr[i].start, PMD_SIZE);
323 			unsigned long end = round_up(mr[i].end, PMD_SIZE);
324 
325 #ifdef CONFIG_X86_32
326 			if ((end >> PAGE_SHIFT) > max_low_pfn)
327 				continue;
328 #endif
329 
330 			if (memblock_is_region_memory(start, end - start))
331 				mr[i].page_size_mask |= 1<<PG_LEVEL_2M;
332 		}
333 		if ((page_size_mask & (1<<PG_LEVEL_1G)) &&
334 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) {
335 			unsigned long start = round_down(mr[i].start, PUD_SIZE);
336 			unsigned long end = round_up(mr[i].end, PUD_SIZE);
337 
338 			if (memblock_is_region_memory(start, end - start))
339 				mr[i].page_size_mask |= 1<<PG_LEVEL_1G;
340 		}
341 	}
342 }
343 
344 static const char *page_size_string(struct map_range *mr)
345 {
346 	static const char str_1g[] = "1G";
347 	static const char str_2m[] = "2M";
348 	static const char str_4m[] = "4M";
349 	static const char str_4k[] = "4k";
350 
351 	if (mr->page_size_mask & (1<<PG_LEVEL_1G))
352 		return str_1g;
353 	/*
354 	 * 32-bit without PAE has a 4M large page size.
355 	 * PG_LEVEL_2M is misnamed, but we can at least
356 	 * print out the right size in the string.
357 	 */
358 	if (IS_ENABLED(CONFIG_X86_32) &&
359 	    !IS_ENABLED(CONFIG_X86_PAE) &&
360 	    mr->page_size_mask & (1<<PG_LEVEL_2M))
361 		return str_4m;
362 
363 	if (mr->page_size_mask & (1<<PG_LEVEL_2M))
364 		return str_2m;
365 
366 	return str_4k;
367 }
368 
369 static int __meminit split_mem_range(struct map_range *mr, int nr_range,
370 				     unsigned long start,
371 				     unsigned long end)
372 {
373 	unsigned long start_pfn, end_pfn, limit_pfn;
374 	unsigned long pfn;
375 	int i;
376 
377 	limit_pfn = PFN_DOWN(end);
378 
379 	/* head if not big page alignment ? */
380 	pfn = start_pfn = PFN_DOWN(start);
381 #ifdef CONFIG_X86_32
382 	/*
383 	 * Don't use a large page for the first 2/4MB of memory
384 	 * because there are often fixed size MTRRs in there
385 	 * and overlapping MTRRs into large pages can cause
386 	 * slowdowns.
387 	 */
388 	if (pfn == 0)
389 		end_pfn = PFN_DOWN(PMD_SIZE);
390 	else
391 		end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
392 #else /* CONFIG_X86_64 */
393 	end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
394 #endif
395 	if (end_pfn > limit_pfn)
396 		end_pfn = limit_pfn;
397 	if (start_pfn < end_pfn) {
398 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
399 		pfn = end_pfn;
400 	}
401 
402 	/* big page (2M) range */
403 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
404 #ifdef CONFIG_X86_32
405 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
406 #else /* CONFIG_X86_64 */
407 	end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
408 	if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE)))
409 		end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
410 #endif
411 
412 	if (start_pfn < end_pfn) {
413 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
414 				page_size_mask & (1<<PG_LEVEL_2M));
415 		pfn = end_pfn;
416 	}
417 
418 #ifdef CONFIG_X86_64
419 	/* big page (1G) range */
420 	start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
421 	end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE));
422 	if (start_pfn < end_pfn) {
423 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
424 				page_size_mask &
425 				 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
426 		pfn = end_pfn;
427 	}
428 
429 	/* tail is not big page (1G) alignment */
430 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
431 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
432 	if (start_pfn < end_pfn) {
433 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
434 				page_size_mask & (1<<PG_LEVEL_2M));
435 		pfn = end_pfn;
436 	}
437 #endif
438 
439 	/* tail is not big page (2M) alignment */
440 	start_pfn = pfn;
441 	end_pfn = limit_pfn;
442 	nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
443 
444 	if (!after_bootmem)
445 		adjust_range_page_size_mask(mr, nr_range);
446 
447 	/* try to merge same page size and continuous */
448 	for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
449 		unsigned long old_start;
450 		if (mr[i].end != mr[i+1].start ||
451 		    mr[i].page_size_mask != mr[i+1].page_size_mask)
452 			continue;
453 		/* move it */
454 		old_start = mr[i].start;
455 		memmove(&mr[i], &mr[i+1],
456 			(nr_range - 1 - i) * sizeof(struct map_range));
457 		mr[i--].start = old_start;
458 		nr_range--;
459 	}
460 
461 	for (i = 0; i < nr_range; i++)
462 		pr_debug(" [mem %#010lx-%#010lx] page %s\n",
463 				mr[i].start, mr[i].end - 1,
464 				page_size_string(&mr[i]));
465 
466 	return nr_range;
467 }
468 
469 struct range pfn_mapped[E820_MAX_ENTRIES];
470 int nr_pfn_mapped;
471 
472 static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn)
473 {
474 	nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_MAX_ENTRIES,
475 					     nr_pfn_mapped, start_pfn, end_pfn);
476 	nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_MAX_ENTRIES);
477 
478 	max_pfn_mapped = max(max_pfn_mapped, end_pfn);
479 
480 	if (start_pfn < (1UL<<(32-PAGE_SHIFT)))
481 		max_low_pfn_mapped = max(max_low_pfn_mapped,
482 					 min(end_pfn, 1UL<<(32-PAGE_SHIFT)));
483 }
484 
485 bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn)
486 {
487 	int i;
488 
489 	for (i = 0; i < nr_pfn_mapped; i++)
490 		if ((start_pfn >= pfn_mapped[i].start) &&
491 		    (end_pfn <= pfn_mapped[i].end))
492 			return true;
493 
494 	return false;
495 }
496 
497 /*
498  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
499  * This runs before bootmem is initialized and gets pages directly from
500  * the physical memory. To access them they are temporarily mapped.
501  */
502 unsigned long __ref init_memory_mapping(unsigned long start,
503 					unsigned long end, pgprot_t prot)
504 {
505 	struct map_range mr[NR_RANGE_MR];
506 	unsigned long ret = 0;
507 	int nr_range, i;
508 
509 	pr_debug("init_memory_mapping: [mem %#010lx-%#010lx]\n",
510 	       start, end - 1);
511 
512 	memset(mr, 0, sizeof(mr));
513 	nr_range = split_mem_range(mr, 0, start, end);
514 
515 	for (i = 0; i < nr_range; i++)
516 		ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
517 						   mr[i].page_size_mask,
518 						   prot);
519 
520 	add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT);
521 
522 	return ret >> PAGE_SHIFT;
523 }
524 
525 /*
526  * We need to iterate through the E820 memory map and create direct mappings
527  * for only E820_TYPE_RAM and E820_KERN_RESERVED regions. We cannot simply
528  * create direct mappings for all pfns from [0 to max_low_pfn) and
529  * [4GB to max_pfn) because of possible memory holes in high addresses
530  * that cannot be marked as UC by fixed/variable range MTRRs.
531  * Depending on the alignment of E820 ranges, this may possibly result
532  * in using smaller size (i.e. 4K instead of 2M or 1G) page tables.
533  *
534  * init_mem_mapping() calls init_range_memory_mapping() with big range.
535  * That range would have hole in the middle or ends, and only ram parts
536  * will be mapped in init_range_memory_mapping().
537  */
538 static unsigned long __init init_range_memory_mapping(
539 					   unsigned long r_start,
540 					   unsigned long r_end)
541 {
542 	unsigned long start_pfn, end_pfn;
543 	unsigned long mapped_ram_size = 0;
544 	int i;
545 
546 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
547 		u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end);
548 		u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end);
549 		if (start >= end)
550 			continue;
551 
552 		/*
553 		 * if it is overlapping with brk pgt, we need to
554 		 * alloc pgt buf from memblock instead.
555 		 */
556 		can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >=
557 				    min(end, (u64)pgt_buf_top<<PAGE_SHIFT);
558 		init_memory_mapping(start, end, PAGE_KERNEL);
559 		mapped_ram_size += end - start;
560 		can_use_brk_pgt = true;
561 	}
562 
563 	return mapped_ram_size;
564 }
565 
566 static unsigned long __init get_new_step_size(unsigned long step_size)
567 {
568 	/*
569 	 * Initial mapped size is PMD_SIZE (2M).
570 	 * We can not set step_size to be PUD_SIZE (1G) yet.
571 	 * In worse case, when we cross the 1G boundary, and
572 	 * PG_LEVEL_2M is not set, we will need 1+1+512 pages (2M + 8k)
573 	 * to map 1G range with PTE. Hence we use one less than the
574 	 * difference of page table level shifts.
575 	 *
576 	 * Don't need to worry about overflow in the top-down case, on 32bit,
577 	 * when step_size is 0, round_down() returns 0 for start, and that
578 	 * turns it into 0x100000000ULL.
579 	 * In the bottom-up case, round_up(x, 0) returns 0 though too, which
580 	 * needs to be taken into consideration by the code below.
581 	 */
582 	return step_size << (PMD_SHIFT - PAGE_SHIFT - 1);
583 }
584 
585 /**
586  * memory_map_top_down - Map [map_start, map_end) top down
587  * @map_start: start address of the target memory range
588  * @map_end: end address of the target memory range
589  *
590  * This function will setup direct mapping for memory range
591  * [map_start, map_end) in top-down. That said, the page tables
592  * will be allocated at the end of the memory, and we map the
593  * memory in top-down.
594  */
595 static void __init memory_map_top_down(unsigned long map_start,
596 				       unsigned long map_end)
597 {
598 	unsigned long real_end, start, last_start;
599 	unsigned long step_size;
600 	unsigned long addr;
601 	unsigned long mapped_ram_size = 0;
602 
603 	/* xen has big range in reserved near end of ram, skip it at first.*/
604 	addr = memblock_find_in_range(map_start, map_end, PMD_SIZE, PMD_SIZE);
605 	real_end = addr + PMD_SIZE;
606 
607 	/* step_size need to be small so pgt_buf from BRK could cover it */
608 	step_size = PMD_SIZE;
609 	max_pfn_mapped = 0; /* will get exact value next */
610 	min_pfn_mapped = real_end >> PAGE_SHIFT;
611 	last_start = start = real_end;
612 
613 	/*
614 	 * We start from the top (end of memory) and go to the bottom.
615 	 * The memblock_find_in_range() gets us a block of RAM from the
616 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
617 	 * for page table.
618 	 */
619 	while (last_start > map_start) {
620 		if (last_start > step_size) {
621 			start = round_down(last_start - 1, step_size);
622 			if (start < map_start)
623 				start = map_start;
624 		} else
625 			start = map_start;
626 		mapped_ram_size += init_range_memory_mapping(start,
627 							last_start);
628 		last_start = start;
629 		min_pfn_mapped = last_start >> PAGE_SHIFT;
630 		if (mapped_ram_size >= step_size)
631 			step_size = get_new_step_size(step_size);
632 	}
633 
634 	if (real_end < map_end)
635 		init_range_memory_mapping(real_end, map_end);
636 }
637 
638 /**
639  * memory_map_bottom_up - Map [map_start, map_end) bottom up
640  * @map_start: start address of the target memory range
641  * @map_end: end address of the target memory range
642  *
643  * This function will setup direct mapping for memory range
644  * [map_start, map_end) in bottom-up. Since we have limited the
645  * bottom-up allocation above the kernel, the page tables will
646  * be allocated just above the kernel and we map the memory
647  * in [map_start, map_end) in bottom-up.
648  */
649 static void __init memory_map_bottom_up(unsigned long map_start,
650 					unsigned long map_end)
651 {
652 	unsigned long next, start;
653 	unsigned long mapped_ram_size = 0;
654 	/* step_size need to be small so pgt_buf from BRK could cover it */
655 	unsigned long step_size = PMD_SIZE;
656 
657 	start = map_start;
658 	min_pfn_mapped = start >> PAGE_SHIFT;
659 
660 	/*
661 	 * We start from the bottom (@map_start) and go to the top (@map_end).
662 	 * The memblock_find_in_range() gets us a block of RAM from the
663 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
664 	 * for page table.
665 	 */
666 	while (start < map_end) {
667 		if (step_size && map_end - start > step_size) {
668 			next = round_up(start + 1, step_size);
669 			if (next > map_end)
670 				next = map_end;
671 		} else {
672 			next = map_end;
673 		}
674 
675 		mapped_ram_size += init_range_memory_mapping(start, next);
676 		start = next;
677 
678 		if (mapped_ram_size >= step_size)
679 			step_size = get_new_step_size(step_size);
680 	}
681 }
682 
683 /*
684  * The real mode trampoline, which is required for bootstrapping CPUs
685  * occupies only a small area under the low 1MB.  See reserve_real_mode()
686  * for details.
687  *
688  * If KASLR is disabled the first PGD entry of the direct mapping is copied
689  * to map the real mode trampoline.
690  *
691  * If KASLR is enabled, copy only the PUD which covers the low 1MB
692  * area. This limits the randomization granularity to 1GB for both 4-level
693  * and 5-level paging.
694  */
695 static void __init init_trampoline(void)
696 {
697 #ifdef CONFIG_X86_64
698 	if (!kaslr_memory_enabled())
699 		trampoline_pgd_entry = init_top_pgt[pgd_index(__PAGE_OFFSET)];
700 	else
701 		init_trampoline_kaslr();
702 #endif
703 }
704 
705 void __init init_mem_mapping(void)
706 {
707 	unsigned long end;
708 
709 	pti_check_boottime_disable();
710 	probe_page_size_mask();
711 	setup_pcid();
712 
713 #ifdef CONFIG_X86_64
714 	end = max_pfn << PAGE_SHIFT;
715 #else
716 	end = max_low_pfn << PAGE_SHIFT;
717 #endif
718 
719 	/* the ISA range is always mapped regardless of memory holes */
720 	init_memory_mapping(0, ISA_END_ADDRESS, PAGE_KERNEL);
721 
722 	/* Init the trampoline, possibly with KASLR memory offset */
723 	init_trampoline();
724 
725 	/*
726 	 * If the allocation is in bottom-up direction, we setup direct mapping
727 	 * in bottom-up, otherwise we setup direct mapping in top-down.
728 	 */
729 	if (memblock_bottom_up()) {
730 		unsigned long kernel_end = __pa_symbol(_end);
731 
732 		/*
733 		 * we need two separate calls here. This is because we want to
734 		 * allocate page tables above the kernel. So we first map
735 		 * [kernel_end, end) to make memory above the kernel be mapped
736 		 * as soon as possible. And then use page tables allocated above
737 		 * the kernel to map [ISA_END_ADDRESS, kernel_end).
738 		 */
739 		memory_map_bottom_up(kernel_end, end);
740 		memory_map_bottom_up(ISA_END_ADDRESS, kernel_end);
741 	} else {
742 		memory_map_top_down(ISA_END_ADDRESS, end);
743 	}
744 
745 #ifdef CONFIG_X86_64
746 	if (max_pfn > max_low_pfn) {
747 		/* can we preseve max_low_pfn ?*/
748 		max_low_pfn = max_pfn;
749 	}
750 #else
751 	early_ioremap_page_table_range_init();
752 #endif
753 
754 	load_cr3(swapper_pg_dir);
755 	__flush_tlb_all();
756 
757 	x86_init.hyper.init_mem_mapping();
758 
759 	early_memtest(0, max_pfn_mapped << PAGE_SHIFT);
760 }
761 
762 /*
763  * Initialize an mm_struct to be used during poking and a pointer to be used
764  * during patching.
765  */
766 void __init poking_init(void)
767 {
768 	spinlock_t *ptl;
769 	pte_t *ptep;
770 
771 	poking_mm = copy_init_mm();
772 	BUG_ON(!poking_mm);
773 
774 	/*
775 	 * Randomize the poking address, but make sure that the following page
776 	 * will be mapped at the same PMD. We need 2 pages, so find space for 3,
777 	 * and adjust the address if the PMD ends after the first one.
778 	 */
779 	poking_addr = TASK_UNMAPPED_BASE;
780 	if (IS_ENABLED(CONFIG_RANDOMIZE_BASE))
781 		poking_addr += (kaslr_get_random_long("Poking") & PAGE_MASK) %
782 			(TASK_SIZE - TASK_UNMAPPED_BASE - 3 * PAGE_SIZE);
783 
784 	if (((poking_addr + PAGE_SIZE) & ~PMD_MASK) == 0)
785 		poking_addr += PAGE_SIZE;
786 
787 	/*
788 	 * We need to trigger the allocation of the page-tables that will be
789 	 * needed for poking now. Later, poking may be performed in an atomic
790 	 * section, which might cause allocation to fail.
791 	 */
792 	ptep = get_locked_pte(poking_mm, poking_addr, &ptl);
793 	BUG_ON(!ptep);
794 	pte_unmap_unlock(ptep, ptl);
795 }
796 
797 /*
798  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
799  * is valid. The argument is a physical page number.
800  *
801  * On x86, access has to be given to the first megabyte of RAM because that
802  * area traditionally contains BIOS code and data regions used by X, dosemu,
803  * and similar apps. Since they map the entire memory range, the whole range
804  * must be allowed (for mapping), but any areas that would otherwise be
805  * disallowed are flagged as being "zero filled" instead of rejected.
806  * Access has to be given to non-kernel-ram areas as well, these contain the
807  * PCI mmio resources as well as potential bios/acpi data regions.
808  */
809 int devmem_is_allowed(unsigned long pagenr)
810 {
811 	if (region_intersects(PFN_PHYS(pagenr), PAGE_SIZE,
812 				IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE)
813 			!= REGION_DISJOINT) {
814 		/*
815 		 * For disallowed memory regions in the low 1MB range,
816 		 * request that the page be shown as all zeros.
817 		 */
818 		if (pagenr < 256)
819 			return 2;
820 
821 		return 0;
822 	}
823 
824 	/*
825 	 * This must follow RAM test, since System RAM is considered a
826 	 * restricted resource under CONFIG_STRICT_IOMEM.
827 	 */
828 	if (iomem_is_exclusive(pagenr << PAGE_SHIFT)) {
829 		/* Low 1MB bypasses iomem restrictions. */
830 		if (pagenr < 256)
831 			return 1;
832 
833 		return 0;
834 	}
835 
836 	return 1;
837 }
838 
839 void free_init_pages(const char *what, unsigned long begin, unsigned long end)
840 {
841 	unsigned long begin_aligned, end_aligned;
842 
843 	/* Make sure boundaries are page aligned */
844 	begin_aligned = PAGE_ALIGN(begin);
845 	end_aligned   = end & PAGE_MASK;
846 
847 	if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
848 		begin = begin_aligned;
849 		end   = end_aligned;
850 	}
851 
852 	if (begin >= end)
853 		return;
854 
855 	/*
856 	 * If debugging page accesses then do not free this memory but
857 	 * mark them not present - any buggy init-section access will
858 	 * create a kernel page fault:
859 	 */
860 	if (debug_pagealloc_enabled()) {
861 		pr_info("debug: unmapping init [mem %#010lx-%#010lx]\n",
862 			begin, end - 1);
863 		/*
864 		 * Inform kmemleak about the hole in the memory since the
865 		 * corresponding pages will be unmapped.
866 		 */
867 		kmemleak_free_part((void *)begin, end - begin);
868 		set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
869 	} else {
870 		/*
871 		 * We just marked the kernel text read only above, now that
872 		 * we are going to free part of that, we need to make that
873 		 * writeable and non-executable first.
874 		 */
875 		set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
876 		set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
877 
878 		free_reserved_area((void *)begin, (void *)end,
879 				   POISON_FREE_INITMEM, what);
880 	}
881 }
882 
883 /*
884  * begin/end can be in the direct map or the "high kernel mapping"
885  * used for the kernel image only.  free_init_pages() will do the
886  * right thing for either kind of address.
887  */
888 void free_kernel_image_pages(const char *what, void *begin, void *end)
889 {
890 	unsigned long begin_ul = (unsigned long)begin;
891 	unsigned long end_ul = (unsigned long)end;
892 	unsigned long len_pages = (end_ul - begin_ul) >> PAGE_SHIFT;
893 
894 	free_init_pages(what, begin_ul, end_ul);
895 
896 	/*
897 	 * PTI maps some of the kernel into userspace.  For performance,
898 	 * this includes some kernel areas that do not contain secrets.
899 	 * Those areas might be adjacent to the parts of the kernel image
900 	 * being freed, which may contain secrets.  Remove the "high kernel
901 	 * image mapping" for these freed areas, ensuring they are not even
902 	 * potentially vulnerable to Meltdown regardless of the specific
903 	 * optimizations PTI is currently using.
904 	 *
905 	 * The "noalias" prevents unmapping the direct map alias which is
906 	 * needed to access the freed pages.
907 	 *
908 	 * This is only valid for 64bit kernels. 32bit has only one mapping
909 	 * which can't be treated in this way for obvious reasons.
910 	 */
911 	if (IS_ENABLED(CONFIG_X86_64) && cpu_feature_enabled(X86_FEATURE_PTI))
912 		set_memory_np_noalias(begin_ul, len_pages);
913 }
914 
915 void __weak mem_encrypt_free_decrypted_mem(void) { }
916 
917 void __ref free_initmem(void)
918 {
919 	e820__reallocate_tables();
920 
921 	mem_encrypt_free_decrypted_mem();
922 
923 	free_kernel_image_pages("unused kernel image (initmem)",
924 				&__init_begin, &__init_end);
925 }
926 
927 #ifdef CONFIG_BLK_DEV_INITRD
928 void __init free_initrd_mem(unsigned long start, unsigned long end)
929 {
930 	/*
931 	 * end could be not aligned, and We can not align that,
932 	 * decompresser could be confused by aligned initrd_end
933 	 * We already reserve the end partial page before in
934 	 *   - i386_start_kernel()
935 	 *   - x86_64_start_kernel()
936 	 *   - relocate_initrd()
937 	 * So here We can do PAGE_ALIGN() safely to get partial page to be freed
938 	 */
939 	free_init_pages("initrd", start, PAGE_ALIGN(end));
940 }
941 #endif
942 
943 /*
944  * Calculate the precise size of the DMA zone (first 16 MB of RAM),
945  * and pass it to the MM layer - to help it set zone watermarks more
946  * accurately.
947  *
948  * Done on 64-bit systems only for the time being, although 32-bit systems
949  * might benefit from this as well.
950  */
951 void __init memblock_find_dma_reserve(void)
952 {
953 #ifdef CONFIG_X86_64
954 	u64 nr_pages = 0, nr_free_pages = 0;
955 	unsigned long start_pfn, end_pfn;
956 	phys_addr_t start_addr, end_addr;
957 	int i;
958 	u64 u;
959 
960 	/*
961 	 * Iterate over all memory ranges (free and reserved ones alike),
962 	 * to calculate the total number of pages in the first 16 MB of RAM:
963 	 */
964 	nr_pages = 0;
965 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
966 		start_pfn = min(start_pfn, MAX_DMA_PFN);
967 		end_pfn   = min(end_pfn,   MAX_DMA_PFN);
968 
969 		nr_pages += end_pfn - start_pfn;
970 	}
971 
972 	/*
973 	 * Iterate over free memory ranges to calculate the number of free
974 	 * pages in the DMA zone, while not counting potential partial
975 	 * pages at the beginning or the end of the range:
976 	 */
977 	nr_free_pages = 0;
978 	for_each_free_mem_range(u, NUMA_NO_NODE, MEMBLOCK_NONE, &start_addr, &end_addr, NULL) {
979 		start_pfn = min_t(unsigned long, PFN_UP(start_addr), MAX_DMA_PFN);
980 		end_pfn   = min_t(unsigned long, PFN_DOWN(end_addr), MAX_DMA_PFN);
981 
982 		if (start_pfn < end_pfn)
983 			nr_free_pages += end_pfn - start_pfn;
984 	}
985 
986 	set_dma_reserve(nr_pages - nr_free_pages);
987 #endif
988 }
989 
990 void __init zone_sizes_init(void)
991 {
992 	unsigned long max_zone_pfns[MAX_NR_ZONES];
993 
994 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
995 
996 #ifdef CONFIG_ZONE_DMA
997 	max_zone_pfns[ZONE_DMA]		= min(MAX_DMA_PFN, max_low_pfn);
998 #endif
999 #ifdef CONFIG_ZONE_DMA32
1000 	max_zone_pfns[ZONE_DMA32]	= min(MAX_DMA32_PFN, max_low_pfn);
1001 #endif
1002 	max_zone_pfns[ZONE_NORMAL]	= max_low_pfn;
1003 #ifdef CONFIG_HIGHMEM
1004 	max_zone_pfns[ZONE_HIGHMEM]	= max_pfn;
1005 #endif
1006 
1007 	free_area_init(max_zone_pfns);
1008 }
1009 
1010 __visible DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate) = {
1011 	.loaded_mm = &init_mm,
1012 	.next_asid = 1,
1013 	.cr4 = ~0UL,	/* fail hard if we screw up cr4 shadow initialization */
1014 };
1015 
1016 void update_cache_mode_entry(unsigned entry, enum page_cache_mode cache)
1017 {
1018 	/* entry 0 MUST be WB (hardwired to speed up translations) */
1019 	BUG_ON(!entry && cache != _PAGE_CACHE_MODE_WB);
1020 
1021 	__cachemode2pte_tbl[cache] = __cm_idx2pte(entry);
1022 	__pte2cachemode_tbl[entry] = cache;
1023 }
1024 
1025 #ifdef CONFIG_SWAP
1026 unsigned long max_swapfile_size(void)
1027 {
1028 	unsigned long pages;
1029 
1030 	pages = generic_max_swapfile_size();
1031 
1032 	if (boot_cpu_has_bug(X86_BUG_L1TF) && l1tf_mitigation != L1TF_MITIGATION_OFF) {
1033 		/* Limit the swap file size to MAX_PA/2 for L1TF workaround */
1034 		unsigned long long l1tf_limit = l1tf_pfn_limit();
1035 		/*
1036 		 * We encode swap offsets also with 3 bits below those for pfn
1037 		 * which makes the usable limit higher.
1038 		 */
1039 #if CONFIG_PGTABLE_LEVELS > 2
1040 		l1tf_limit <<= PAGE_SHIFT - SWP_OFFSET_FIRST_BIT;
1041 #endif
1042 		pages = min_t(unsigned long long, l1tf_limit, pages);
1043 	}
1044 	return pages;
1045 }
1046 #endif
1047