xref: /linux/arch/x86/mm/init.c (revision 23c48a124b469cee2eb0c75e6d22d366d1caa118)
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 #include <asm/memtype.h>
29 
30 /*
31  * We need to define the tracepoints somewhere, and tlb.c
32  * is only compiled when SMP=y.
33  */
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 /*
81  * Check that the write-protect PAT entry is set for write-protect.
82  * To do this without making assumptions how PAT has been set up (Xen has
83  * another layout than the kernel), translate the _PAGE_CACHE_MODE_WP cache
84  * mode via the __cachemode2pte_tbl[] into protection bits (those protection
85  * bits will select a cache mode of WP or better), and then translate the
86  * protection bits back into the cache mode using __pte2cm_idx() and the
87  * __pte2cachemode_tbl[] array. This will return the really used cache mode.
88  */
89 bool x86_has_pat_wp(void)
90 {
91 	uint16_t prot = __cachemode2pte_tbl[_PAGE_CACHE_MODE_WP];
92 
93 	return __pte2cachemode_tbl[__pte2cm_idx(prot)] == _PAGE_CACHE_MODE_WP;
94 }
95 
96 enum page_cache_mode pgprot2cachemode(pgprot_t pgprot)
97 {
98 	unsigned long masked;
99 
100 	masked = pgprot_val(pgprot) & _PAGE_CACHE_MASK;
101 	if (likely(masked == 0))
102 		return 0;
103 	return __pte2cachemode_tbl[__pte2cm_idx(masked)];
104 }
105 
106 static unsigned long __initdata pgt_buf_start;
107 static unsigned long __initdata pgt_buf_end;
108 static unsigned long __initdata pgt_buf_top;
109 
110 static unsigned long min_pfn_mapped;
111 
112 static bool __initdata can_use_brk_pgt = true;
113 
114 /*
115  * Pages returned are already directly mapped.
116  *
117  * Changing that is likely to break Xen, see commit:
118  *
119  *    279b706 x86,xen: introduce x86_init.mapping.pagetable_reserve
120  *
121  * for detailed information.
122  */
123 __ref void *alloc_low_pages(unsigned int num)
124 {
125 	unsigned long pfn;
126 	int i;
127 
128 	if (after_bootmem) {
129 		unsigned int order;
130 
131 		order = get_order((unsigned long)num << PAGE_SHIFT);
132 		return (void *)__get_free_pages(GFP_ATOMIC | __GFP_ZERO, order);
133 	}
134 
135 	if ((pgt_buf_end + num) > pgt_buf_top || !can_use_brk_pgt) {
136 		unsigned long ret = 0;
137 
138 		if (min_pfn_mapped < max_pfn_mapped) {
139 			ret = memblock_phys_alloc_range(
140 					PAGE_SIZE * num, PAGE_SIZE,
141 					min_pfn_mapped << PAGE_SHIFT,
142 					max_pfn_mapped << PAGE_SHIFT);
143 		}
144 		if (!ret && can_use_brk_pgt)
145 			ret = __pa(extend_brk(PAGE_SIZE * num, PAGE_SIZE));
146 
147 		if (!ret)
148 			panic("alloc_low_pages: can not alloc memory");
149 
150 		pfn = ret >> PAGE_SHIFT;
151 	} else {
152 		pfn = pgt_buf_end;
153 		pgt_buf_end += num;
154 	}
155 
156 	for (i = 0; i < num; i++) {
157 		void *adr;
158 
159 		adr = __va((pfn + i) << PAGE_SHIFT);
160 		clear_page(adr);
161 	}
162 
163 	return __va(pfn << PAGE_SHIFT);
164 }
165 
166 /*
167  * By default need to be able to allocate page tables below PGD firstly for
168  * the 0-ISA_END_ADDRESS range and secondly for the initial PMD_SIZE mapping.
169  * With KASLR memory randomization, depending on the machine e820 memory and the
170  * PUD alignment, twice that many pages may be needed when KASLR memory
171  * randomization is enabled.
172  */
173 
174 #ifndef CONFIG_X86_5LEVEL
175 #define INIT_PGD_PAGE_TABLES    3
176 #else
177 #define INIT_PGD_PAGE_TABLES    4
178 #endif
179 
180 #ifndef CONFIG_RANDOMIZE_MEMORY
181 #define INIT_PGD_PAGE_COUNT      (2 * INIT_PGD_PAGE_TABLES)
182 #else
183 #define INIT_PGD_PAGE_COUNT      (4 * INIT_PGD_PAGE_TABLES)
184 #endif
185 
186 #define INIT_PGT_BUF_SIZE	(INIT_PGD_PAGE_COUNT * PAGE_SIZE)
187 RESERVE_BRK(early_pgt_alloc, INIT_PGT_BUF_SIZE);
188 void  __init early_alloc_pgt_buf(void)
189 {
190 	unsigned long tables = INIT_PGT_BUF_SIZE;
191 	phys_addr_t base;
192 
193 	base = __pa(extend_brk(tables, PAGE_SIZE));
194 
195 	pgt_buf_start = base >> PAGE_SHIFT;
196 	pgt_buf_end = pgt_buf_start;
197 	pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
198 }
199 
200 int after_bootmem;
201 
202 early_param_on_off("gbpages", "nogbpages", direct_gbpages, CONFIG_X86_DIRECT_GBPAGES);
203 
204 struct map_range {
205 	unsigned long start;
206 	unsigned long end;
207 	unsigned page_size_mask;
208 };
209 
210 static int page_size_mask;
211 
212 /*
213  * Save some of cr4 feature set we're using (e.g.  Pentium 4MB
214  * enable and PPro Global page enable), so that any CPU's that boot
215  * up after us can get the correct flags. Invoked on the boot CPU.
216  */
217 static inline void cr4_set_bits_and_update_boot(unsigned long mask)
218 {
219 	mmu_cr4_features |= mask;
220 	if (trampoline_cr4_features)
221 		*trampoline_cr4_features = mmu_cr4_features;
222 	cr4_set_bits(mask);
223 }
224 
225 static void __init probe_page_size_mask(void)
226 {
227 	/*
228 	 * For pagealloc debugging, identity mapping will use small pages.
229 	 * This will simplify cpa(), which otherwise needs to support splitting
230 	 * large pages into small in interrupt context, etc.
231 	 */
232 	if (boot_cpu_has(X86_FEATURE_PSE) && !debug_pagealloc_enabled())
233 		page_size_mask |= 1 << PG_LEVEL_2M;
234 	else
235 		direct_gbpages = 0;
236 
237 	/* Enable PSE if available */
238 	if (boot_cpu_has(X86_FEATURE_PSE))
239 		cr4_set_bits_and_update_boot(X86_CR4_PSE);
240 
241 	/* Enable PGE if available */
242 	__supported_pte_mask &= ~_PAGE_GLOBAL;
243 	if (boot_cpu_has(X86_FEATURE_PGE)) {
244 		cr4_set_bits_and_update_boot(X86_CR4_PGE);
245 		__supported_pte_mask |= _PAGE_GLOBAL;
246 	}
247 
248 	/* By the default is everything supported: */
249 	__default_kernel_pte_mask = __supported_pte_mask;
250 	/* Except when with PTI where the kernel is mostly non-Global: */
251 	if (cpu_feature_enabled(X86_FEATURE_PTI))
252 		__default_kernel_pte_mask &= ~_PAGE_GLOBAL;
253 
254 	/* Enable 1 GB linear kernel mappings if available: */
255 	if (direct_gbpages && boot_cpu_has(X86_FEATURE_GBPAGES)) {
256 		printk(KERN_INFO "Using GB pages for direct mapping\n");
257 		page_size_mask |= 1 << PG_LEVEL_1G;
258 	} else {
259 		direct_gbpages = 0;
260 	}
261 }
262 
263 static void setup_pcid(void)
264 {
265 	if (!IS_ENABLED(CONFIG_X86_64))
266 		return;
267 
268 	if (!boot_cpu_has(X86_FEATURE_PCID))
269 		return;
270 
271 	if (boot_cpu_has(X86_FEATURE_PGE)) {
272 		/*
273 		 * This can't be cr4_set_bits_and_update_boot() -- the
274 		 * trampoline code can't handle CR4.PCIDE and it wouldn't
275 		 * do any good anyway.  Despite the name,
276 		 * cr4_set_bits_and_update_boot() doesn't actually cause
277 		 * the bits in question to remain set all the way through
278 		 * the secondary boot asm.
279 		 *
280 		 * Instead, we brute-force it and set CR4.PCIDE manually in
281 		 * start_secondary().
282 		 */
283 		cr4_set_bits(X86_CR4_PCIDE);
284 
285 		/*
286 		 * INVPCID's single-context modes (2/3) only work if we set
287 		 * X86_CR4_PCIDE, *and* we INVPCID support.  It's unusable
288 		 * on systems that have X86_CR4_PCIDE clear, or that have
289 		 * no INVPCID support at all.
290 		 */
291 		if (boot_cpu_has(X86_FEATURE_INVPCID))
292 			setup_force_cpu_cap(X86_FEATURE_INVPCID_SINGLE);
293 	} else {
294 		/*
295 		 * flush_tlb_all(), as currently implemented, won't work if
296 		 * PCID is on but PGE is not.  Since that combination
297 		 * doesn't exist on real hardware, there's no reason to try
298 		 * to fully support it, but it's polite to avoid corrupting
299 		 * data if we're on an improperly configured VM.
300 		 */
301 		setup_clear_cpu_cap(X86_FEATURE_PCID);
302 	}
303 }
304 
305 #ifdef CONFIG_X86_32
306 #define NR_RANGE_MR 3
307 #else /* CONFIG_X86_64 */
308 #define NR_RANGE_MR 5
309 #endif
310 
311 static int __meminit save_mr(struct map_range *mr, int nr_range,
312 			     unsigned long start_pfn, unsigned long end_pfn,
313 			     unsigned long page_size_mask)
314 {
315 	if (start_pfn < end_pfn) {
316 		if (nr_range >= NR_RANGE_MR)
317 			panic("run out of range for init_memory_mapping\n");
318 		mr[nr_range].start = start_pfn<<PAGE_SHIFT;
319 		mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
320 		mr[nr_range].page_size_mask = page_size_mask;
321 		nr_range++;
322 	}
323 
324 	return nr_range;
325 }
326 
327 /*
328  * adjust the page_size_mask for small range to go with
329  *	big page size instead small one if nearby are ram too.
330  */
331 static void __ref adjust_range_page_size_mask(struct map_range *mr,
332 							 int nr_range)
333 {
334 	int i;
335 
336 	for (i = 0; i < nr_range; i++) {
337 		if ((page_size_mask & (1<<PG_LEVEL_2M)) &&
338 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) {
339 			unsigned long start = round_down(mr[i].start, PMD_SIZE);
340 			unsigned long end = round_up(mr[i].end, PMD_SIZE);
341 
342 #ifdef CONFIG_X86_32
343 			if ((end >> PAGE_SHIFT) > max_low_pfn)
344 				continue;
345 #endif
346 
347 			if (memblock_is_region_memory(start, end - start))
348 				mr[i].page_size_mask |= 1<<PG_LEVEL_2M;
349 		}
350 		if ((page_size_mask & (1<<PG_LEVEL_1G)) &&
351 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) {
352 			unsigned long start = round_down(mr[i].start, PUD_SIZE);
353 			unsigned long end = round_up(mr[i].end, PUD_SIZE);
354 
355 			if (memblock_is_region_memory(start, end - start))
356 				mr[i].page_size_mask |= 1<<PG_LEVEL_1G;
357 		}
358 	}
359 }
360 
361 static const char *page_size_string(struct map_range *mr)
362 {
363 	static const char str_1g[] = "1G";
364 	static const char str_2m[] = "2M";
365 	static const char str_4m[] = "4M";
366 	static const char str_4k[] = "4k";
367 
368 	if (mr->page_size_mask & (1<<PG_LEVEL_1G))
369 		return str_1g;
370 	/*
371 	 * 32-bit without PAE has a 4M large page size.
372 	 * PG_LEVEL_2M is misnamed, but we can at least
373 	 * print out the right size in the string.
374 	 */
375 	if (IS_ENABLED(CONFIG_X86_32) &&
376 	    !IS_ENABLED(CONFIG_X86_PAE) &&
377 	    mr->page_size_mask & (1<<PG_LEVEL_2M))
378 		return str_4m;
379 
380 	if (mr->page_size_mask & (1<<PG_LEVEL_2M))
381 		return str_2m;
382 
383 	return str_4k;
384 }
385 
386 static int __meminit split_mem_range(struct map_range *mr, int nr_range,
387 				     unsigned long start,
388 				     unsigned long end)
389 {
390 	unsigned long start_pfn, end_pfn, limit_pfn;
391 	unsigned long pfn;
392 	int i;
393 
394 	limit_pfn = PFN_DOWN(end);
395 
396 	/* head if not big page alignment ? */
397 	pfn = start_pfn = PFN_DOWN(start);
398 #ifdef CONFIG_X86_32
399 	/*
400 	 * Don't use a large page for the first 2/4MB of memory
401 	 * because there are often fixed size MTRRs in there
402 	 * and overlapping MTRRs into large pages can cause
403 	 * slowdowns.
404 	 */
405 	if (pfn == 0)
406 		end_pfn = PFN_DOWN(PMD_SIZE);
407 	else
408 		end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
409 #else /* CONFIG_X86_64 */
410 	end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
411 #endif
412 	if (end_pfn > limit_pfn)
413 		end_pfn = limit_pfn;
414 	if (start_pfn < end_pfn) {
415 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
416 		pfn = end_pfn;
417 	}
418 
419 	/* big page (2M) range */
420 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
421 #ifdef CONFIG_X86_32
422 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
423 #else /* CONFIG_X86_64 */
424 	end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
425 	if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE)))
426 		end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
427 #endif
428 
429 	if (start_pfn < end_pfn) {
430 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
431 				page_size_mask & (1<<PG_LEVEL_2M));
432 		pfn = end_pfn;
433 	}
434 
435 #ifdef CONFIG_X86_64
436 	/* big page (1G) range */
437 	start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
438 	end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE));
439 	if (start_pfn < end_pfn) {
440 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
441 				page_size_mask &
442 				 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
443 		pfn = end_pfn;
444 	}
445 
446 	/* tail is not big page (1G) alignment */
447 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
448 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
449 	if (start_pfn < end_pfn) {
450 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
451 				page_size_mask & (1<<PG_LEVEL_2M));
452 		pfn = end_pfn;
453 	}
454 #endif
455 
456 	/* tail is not big page (2M) alignment */
457 	start_pfn = pfn;
458 	end_pfn = limit_pfn;
459 	nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
460 
461 	if (!after_bootmem)
462 		adjust_range_page_size_mask(mr, nr_range);
463 
464 	/* try to merge same page size and continuous */
465 	for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
466 		unsigned long old_start;
467 		if (mr[i].end != mr[i+1].start ||
468 		    mr[i].page_size_mask != mr[i+1].page_size_mask)
469 			continue;
470 		/* move it */
471 		old_start = mr[i].start;
472 		memmove(&mr[i], &mr[i+1],
473 			(nr_range - 1 - i) * sizeof(struct map_range));
474 		mr[i--].start = old_start;
475 		nr_range--;
476 	}
477 
478 	for (i = 0; i < nr_range; i++)
479 		pr_debug(" [mem %#010lx-%#010lx] page %s\n",
480 				mr[i].start, mr[i].end - 1,
481 				page_size_string(&mr[i]));
482 
483 	return nr_range;
484 }
485 
486 struct range pfn_mapped[E820_MAX_ENTRIES];
487 int nr_pfn_mapped;
488 
489 static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn)
490 {
491 	nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_MAX_ENTRIES,
492 					     nr_pfn_mapped, start_pfn, end_pfn);
493 	nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_MAX_ENTRIES);
494 
495 	max_pfn_mapped = max(max_pfn_mapped, end_pfn);
496 
497 	if (start_pfn < (1UL<<(32-PAGE_SHIFT)))
498 		max_low_pfn_mapped = max(max_low_pfn_mapped,
499 					 min(end_pfn, 1UL<<(32-PAGE_SHIFT)));
500 }
501 
502 bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn)
503 {
504 	int i;
505 
506 	for (i = 0; i < nr_pfn_mapped; i++)
507 		if ((start_pfn >= pfn_mapped[i].start) &&
508 		    (end_pfn <= pfn_mapped[i].end))
509 			return true;
510 
511 	return false;
512 }
513 
514 /*
515  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
516  * This runs before bootmem is initialized and gets pages directly from
517  * the physical memory. To access them they are temporarily mapped.
518  */
519 unsigned long __ref init_memory_mapping(unsigned long start,
520 					unsigned long end, pgprot_t prot)
521 {
522 	struct map_range mr[NR_RANGE_MR];
523 	unsigned long ret = 0;
524 	int nr_range, i;
525 
526 	pr_debug("init_memory_mapping: [mem %#010lx-%#010lx]\n",
527 	       start, end - 1);
528 
529 	memset(mr, 0, sizeof(mr));
530 	nr_range = split_mem_range(mr, 0, start, end);
531 
532 	for (i = 0; i < nr_range; i++)
533 		ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
534 						   mr[i].page_size_mask,
535 						   prot);
536 
537 	add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT);
538 
539 	return ret >> PAGE_SHIFT;
540 }
541 
542 /*
543  * We need to iterate through the E820 memory map and create direct mappings
544  * for only E820_TYPE_RAM and E820_KERN_RESERVED regions. We cannot simply
545  * create direct mappings for all pfns from [0 to max_low_pfn) and
546  * [4GB to max_pfn) because of possible memory holes in high addresses
547  * that cannot be marked as UC by fixed/variable range MTRRs.
548  * Depending on the alignment of E820 ranges, this may possibly result
549  * in using smaller size (i.e. 4K instead of 2M or 1G) page tables.
550  *
551  * init_mem_mapping() calls init_range_memory_mapping() with big range.
552  * That range would have hole in the middle or ends, and only ram parts
553  * will be mapped in init_range_memory_mapping().
554  */
555 static unsigned long __init init_range_memory_mapping(
556 					   unsigned long r_start,
557 					   unsigned long r_end)
558 {
559 	unsigned long start_pfn, end_pfn;
560 	unsigned long mapped_ram_size = 0;
561 	int i;
562 
563 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
564 		u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end);
565 		u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end);
566 		if (start >= end)
567 			continue;
568 
569 		/*
570 		 * if it is overlapping with brk pgt, we need to
571 		 * alloc pgt buf from memblock instead.
572 		 */
573 		can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >=
574 				    min(end, (u64)pgt_buf_top<<PAGE_SHIFT);
575 		init_memory_mapping(start, end, PAGE_KERNEL);
576 		mapped_ram_size += end - start;
577 		can_use_brk_pgt = true;
578 	}
579 
580 	return mapped_ram_size;
581 }
582 
583 static unsigned long __init get_new_step_size(unsigned long step_size)
584 {
585 	/*
586 	 * Initial mapped size is PMD_SIZE (2M).
587 	 * We can not set step_size to be PUD_SIZE (1G) yet.
588 	 * In worse case, when we cross the 1G boundary, and
589 	 * PG_LEVEL_2M is not set, we will need 1+1+512 pages (2M + 8k)
590 	 * to map 1G range with PTE. Hence we use one less than the
591 	 * difference of page table level shifts.
592 	 *
593 	 * Don't need to worry about overflow in the top-down case, on 32bit,
594 	 * when step_size is 0, round_down() returns 0 for start, and that
595 	 * turns it into 0x100000000ULL.
596 	 * In the bottom-up case, round_up(x, 0) returns 0 though too, which
597 	 * needs to be taken into consideration by the code below.
598 	 */
599 	return step_size << (PMD_SHIFT - PAGE_SHIFT - 1);
600 }
601 
602 /**
603  * memory_map_top_down - Map [map_start, map_end) top down
604  * @map_start: start address of the target memory range
605  * @map_end: end address of the target memory range
606  *
607  * This function will setup direct mapping for memory range
608  * [map_start, map_end) in top-down. That said, the page tables
609  * will be allocated at the end of the memory, and we map the
610  * memory in top-down.
611  */
612 static void __init memory_map_top_down(unsigned long map_start,
613 				       unsigned long map_end)
614 {
615 	unsigned long real_end, last_start;
616 	unsigned long step_size;
617 	unsigned long addr;
618 	unsigned long mapped_ram_size = 0;
619 
620 	/*
621 	 * Systems that have many reserved areas near top of the memory,
622 	 * e.g. QEMU with less than 1G RAM and EFI enabled, or Xen, will
623 	 * require lots of 4K mappings which may exhaust pgt_buf.
624 	 * Start with top-most PMD_SIZE range aligned at PMD_SIZE to ensure
625 	 * there is enough mapped memory that can be allocated from
626 	 * memblock.
627 	 */
628 	addr = memblock_phys_alloc_range(PMD_SIZE, PMD_SIZE, map_start,
629 					 map_end);
630 	memblock_phys_free(addr, PMD_SIZE);
631 	real_end = addr + PMD_SIZE;
632 
633 	/* step_size need to be small so pgt_buf from BRK could cover it */
634 	step_size = PMD_SIZE;
635 	max_pfn_mapped = 0; /* will get exact value next */
636 	min_pfn_mapped = real_end >> PAGE_SHIFT;
637 	last_start = real_end;
638 
639 	/*
640 	 * We start from the top (end of memory) and go to the bottom.
641 	 * The memblock_find_in_range() gets us a block of RAM from the
642 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
643 	 * for page table.
644 	 */
645 	while (last_start > map_start) {
646 		unsigned long start;
647 
648 		if (last_start > step_size) {
649 			start = round_down(last_start - 1, step_size);
650 			if (start < map_start)
651 				start = map_start;
652 		} else
653 			start = map_start;
654 		mapped_ram_size += init_range_memory_mapping(start,
655 							last_start);
656 		last_start = start;
657 		min_pfn_mapped = last_start >> PAGE_SHIFT;
658 		if (mapped_ram_size >= step_size)
659 			step_size = get_new_step_size(step_size);
660 	}
661 
662 	if (real_end < map_end)
663 		init_range_memory_mapping(real_end, map_end);
664 }
665 
666 /**
667  * memory_map_bottom_up - Map [map_start, map_end) bottom up
668  * @map_start: start address of the target memory range
669  * @map_end: end address of the target memory range
670  *
671  * This function will setup direct mapping for memory range
672  * [map_start, map_end) in bottom-up. Since we have limited the
673  * bottom-up allocation above the kernel, the page tables will
674  * be allocated just above the kernel and we map the memory
675  * in [map_start, map_end) in bottom-up.
676  */
677 static void __init memory_map_bottom_up(unsigned long map_start,
678 					unsigned long map_end)
679 {
680 	unsigned long next, start;
681 	unsigned long mapped_ram_size = 0;
682 	/* step_size need to be small so pgt_buf from BRK could cover it */
683 	unsigned long step_size = PMD_SIZE;
684 
685 	start = map_start;
686 	min_pfn_mapped = start >> PAGE_SHIFT;
687 
688 	/*
689 	 * We start from the bottom (@map_start) and go to the top (@map_end).
690 	 * The memblock_find_in_range() gets us a block of RAM from the
691 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
692 	 * for page table.
693 	 */
694 	while (start < map_end) {
695 		if (step_size && map_end - start > step_size) {
696 			next = round_up(start + 1, step_size);
697 			if (next > map_end)
698 				next = map_end;
699 		} else {
700 			next = map_end;
701 		}
702 
703 		mapped_ram_size += init_range_memory_mapping(start, next);
704 		start = next;
705 
706 		if (mapped_ram_size >= step_size)
707 			step_size = get_new_step_size(step_size);
708 	}
709 }
710 
711 /*
712  * The real mode trampoline, which is required for bootstrapping CPUs
713  * occupies only a small area under the low 1MB.  See reserve_real_mode()
714  * for details.
715  *
716  * If KASLR is disabled the first PGD entry of the direct mapping is copied
717  * to map the real mode trampoline.
718  *
719  * If KASLR is enabled, copy only the PUD which covers the low 1MB
720  * area. This limits the randomization granularity to 1GB for both 4-level
721  * and 5-level paging.
722  */
723 static void __init init_trampoline(void)
724 {
725 #ifdef CONFIG_X86_64
726 	/*
727 	 * The code below will alias kernel page-tables in the user-range of the
728 	 * address space, including the Global bit. So global TLB entries will
729 	 * be created when using the trampoline page-table.
730 	 */
731 	if (!kaslr_memory_enabled())
732 		trampoline_pgd_entry = init_top_pgt[pgd_index(__PAGE_OFFSET)];
733 	else
734 		init_trampoline_kaslr();
735 #endif
736 }
737 
738 void __init init_mem_mapping(void)
739 {
740 	unsigned long end;
741 
742 	pti_check_boottime_disable();
743 	probe_page_size_mask();
744 	setup_pcid();
745 
746 #ifdef CONFIG_X86_64
747 	end = max_pfn << PAGE_SHIFT;
748 #else
749 	end = max_low_pfn << PAGE_SHIFT;
750 #endif
751 
752 	/* the ISA range is always mapped regardless of memory holes */
753 	init_memory_mapping(0, ISA_END_ADDRESS, PAGE_KERNEL);
754 
755 	/* Init the trampoline, possibly with KASLR memory offset */
756 	init_trampoline();
757 
758 	/*
759 	 * If the allocation is in bottom-up direction, we setup direct mapping
760 	 * in bottom-up, otherwise we setup direct mapping in top-down.
761 	 */
762 	if (memblock_bottom_up()) {
763 		unsigned long kernel_end = __pa_symbol(_end);
764 
765 		/*
766 		 * we need two separate calls here. This is because we want to
767 		 * allocate page tables above the kernel. So we first map
768 		 * [kernel_end, end) to make memory above the kernel be mapped
769 		 * as soon as possible. And then use page tables allocated above
770 		 * the kernel to map [ISA_END_ADDRESS, kernel_end).
771 		 */
772 		memory_map_bottom_up(kernel_end, end);
773 		memory_map_bottom_up(ISA_END_ADDRESS, kernel_end);
774 	} else {
775 		memory_map_top_down(ISA_END_ADDRESS, end);
776 	}
777 
778 #ifdef CONFIG_X86_64
779 	if (max_pfn > max_low_pfn) {
780 		/* can we preserve max_low_pfn ?*/
781 		max_low_pfn = max_pfn;
782 	}
783 #else
784 	early_ioremap_page_table_range_init();
785 #endif
786 
787 	load_cr3(swapper_pg_dir);
788 	__flush_tlb_all();
789 
790 	x86_init.hyper.init_mem_mapping();
791 
792 	early_memtest(0, max_pfn_mapped << PAGE_SHIFT);
793 }
794 
795 /*
796  * Initialize an mm_struct to be used during poking and a pointer to be used
797  * during patching.
798  */
799 void __init poking_init(void)
800 {
801 	spinlock_t *ptl;
802 	pte_t *ptep;
803 
804 	poking_mm = copy_init_mm();
805 	BUG_ON(!poking_mm);
806 
807 	/*
808 	 * Randomize the poking address, but make sure that the following page
809 	 * will be mapped at the same PMD. We need 2 pages, so find space for 3,
810 	 * and adjust the address if the PMD ends after the first one.
811 	 */
812 	poking_addr = TASK_UNMAPPED_BASE;
813 	if (IS_ENABLED(CONFIG_RANDOMIZE_BASE))
814 		poking_addr += (kaslr_get_random_long("Poking") & PAGE_MASK) %
815 			(TASK_SIZE - TASK_UNMAPPED_BASE - 3 * PAGE_SIZE);
816 
817 	if (((poking_addr + PAGE_SIZE) & ~PMD_MASK) == 0)
818 		poking_addr += PAGE_SIZE;
819 
820 	/*
821 	 * We need to trigger the allocation of the page-tables that will be
822 	 * needed for poking now. Later, poking may be performed in an atomic
823 	 * section, which might cause allocation to fail.
824 	 */
825 	ptep = get_locked_pte(poking_mm, poking_addr, &ptl);
826 	BUG_ON(!ptep);
827 	pte_unmap_unlock(ptep, ptl);
828 }
829 
830 /*
831  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
832  * is valid. The argument is a physical page number.
833  *
834  * On x86, access has to be given to the first megabyte of RAM because that
835  * area traditionally contains BIOS code and data regions used by X, dosemu,
836  * and similar apps. Since they map the entire memory range, the whole range
837  * must be allowed (for mapping), but any areas that would otherwise be
838  * disallowed are flagged as being "zero filled" instead of rejected.
839  * Access has to be given to non-kernel-ram areas as well, these contain the
840  * PCI mmio resources as well as potential bios/acpi data regions.
841  */
842 int devmem_is_allowed(unsigned long pagenr)
843 {
844 	if (region_intersects(PFN_PHYS(pagenr), PAGE_SIZE,
845 				IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE)
846 			!= REGION_DISJOINT) {
847 		/*
848 		 * For disallowed memory regions in the low 1MB range,
849 		 * request that the page be shown as all zeros.
850 		 */
851 		if (pagenr < 256)
852 			return 2;
853 
854 		return 0;
855 	}
856 
857 	/*
858 	 * This must follow RAM test, since System RAM is considered a
859 	 * restricted resource under CONFIG_STRICT_DEVMEM.
860 	 */
861 	if (iomem_is_exclusive(pagenr << PAGE_SHIFT)) {
862 		/* Low 1MB bypasses iomem restrictions. */
863 		if (pagenr < 256)
864 			return 1;
865 
866 		return 0;
867 	}
868 
869 	return 1;
870 }
871 
872 void free_init_pages(const char *what, unsigned long begin, unsigned long end)
873 {
874 	unsigned long begin_aligned, end_aligned;
875 
876 	/* Make sure boundaries are page aligned */
877 	begin_aligned = PAGE_ALIGN(begin);
878 	end_aligned   = end & PAGE_MASK;
879 
880 	if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
881 		begin = begin_aligned;
882 		end   = end_aligned;
883 	}
884 
885 	if (begin >= end)
886 		return;
887 
888 	/*
889 	 * If debugging page accesses then do not free this memory but
890 	 * mark them not present - any buggy init-section access will
891 	 * create a kernel page fault:
892 	 */
893 	if (debug_pagealloc_enabled()) {
894 		pr_info("debug: unmapping init [mem %#010lx-%#010lx]\n",
895 			begin, end - 1);
896 		/*
897 		 * Inform kmemleak about the hole in the memory since the
898 		 * corresponding pages will be unmapped.
899 		 */
900 		kmemleak_free_part((void *)begin, end - begin);
901 		set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
902 	} else {
903 		/*
904 		 * We just marked the kernel text read only above, now that
905 		 * we are going to free part of that, we need to make that
906 		 * writeable and non-executable first.
907 		 */
908 		set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
909 		set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
910 
911 		free_reserved_area((void *)begin, (void *)end,
912 				   POISON_FREE_INITMEM, what);
913 	}
914 }
915 
916 /*
917  * begin/end can be in the direct map or the "high kernel mapping"
918  * used for the kernel image only.  free_init_pages() will do the
919  * right thing for either kind of address.
920  */
921 void free_kernel_image_pages(const char *what, void *begin, void *end)
922 {
923 	unsigned long begin_ul = (unsigned long)begin;
924 	unsigned long end_ul = (unsigned long)end;
925 	unsigned long len_pages = (end_ul - begin_ul) >> PAGE_SHIFT;
926 
927 	free_init_pages(what, begin_ul, end_ul);
928 
929 	/*
930 	 * PTI maps some of the kernel into userspace.  For performance,
931 	 * this includes some kernel areas that do not contain secrets.
932 	 * Those areas might be adjacent to the parts of the kernel image
933 	 * being freed, which may contain secrets.  Remove the "high kernel
934 	 * image mapping" for these freed areas, ensuring they are not even
935 	 * potentially vulnerable to Meltdown regardless of the specific
936 	 * optimizations PTI is currently using.
937 	 *
938 	 * The "noalias" prevents unmapping the direct map alias which is
939 	 * needed to access the freed pages.
940 	 *
941 	 * This is only valid for 64bit kernels. 32bit has only one mapping
942 	 * which can't be treated in this way for obvious reasons.
943 	 */
944 	if (IS_ENABLED(CONFIG_X86_64) && cpu_feature_enabled(X86_FEATURE_PTI))
945 		set_memory_np_noalias(begin_ul, len_pages);
946 }
947 
948 void __ref free_initmem(void)
949 {
950 	e820__reallocate_tables();
951 
952 	mem_encrypt_free_decrypted_mem();
953 
954 	free_kernel_image_pages("unused kernel image (initmem)",
955 				&__init_begin, &__init_end);
956 }
957 
958 #ifdef CONFIG_BLK_DEV_INITRD
959 void __init free_initrd_mem(unsigned long start, unsigned long end)
960 {
961 	/*
962 	 * end could be not aligned, and We can not align that,
963 	 * decompressor could be confused by aligned initrd_end
964 	 * We already reserve the end partial page before in
965 	 *   - i386_start_kernel()
966 	 *   - x86_64_start_kernel()
967 	 *   - relocate_initrd()
968 	 * So here We can do PAGE_ALIGN() safely to get partial page to be freed
969 	 */
970 	free_init_pages("initrd", start, PAGE_ALIGN(end));
971 }
972 #endif
973 
974 /*
975  * Calculate the precise size of the DMA zone (first 16 MB of RAM),
976  * and pass it to the MM layer - to help it set zone watermarks more
977  * accurately.
978  *
979  * Done on 64-bit systems only for the time being, although 32-bit systems
980  * might benefit from this as well.
981  */
982 void __init memblock_find_dma_reserve(void)
983 {
984 #ifdef CONFIG_X86_64
985 	u64 nr_pages = 0, nr_free_pages = 0;
986 	unsigned long start_pfn, end_pfn;
987 	phys_addr_t start_addr, end_addr;
988 	int i;
989 	u64 u;
990 
991 	/*
992 	 * Iterate over all memory ranges (free and reserved ones alike),
993 	 * to calculate the total number of pages in the first 16 MB of RAM:
994 	 */
995 	nr_pages = 0;
996 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
997 		start_pfn = min(start_pfn, MAX_DMA_PFN);
998 		end_pfn   = min(end_pfn,   MAX_DMA_PFN);
999 
1000 		nr_pages += end_pfn - start_pfn;
1001 	}
1002 
1003 	/*
1004 	 * Iterate over free memory ranges to calculate the number of free
1005 	 * pages in the DMA zone, while not counting potential partial
1006 	 * pages at the beginning or the end of the range:
1007 	 */
1008 	nr_free_pages = 0;
1009 	for_each_free_mem_range(u, NUMA_NO_NODE, MEMBLOCK_NONE, &start_addr, &end_addr, NULL) {
1010 		start_pfn = min_t(unsigned long, PFN_UP(start_addr), MAX_DMA_PFN);
1011 		end_pfn   = min_t(unsigned long, PFN_DOWN(end_addr), MAX_DMA_PFN);
1012 
1013 		if (start_pfn < end_pfn)
1014 			nr_free_pages += end_pfn - start_pfn;
1015 	}
1016 
1017 	set_dma_reserve(nr_pages - nr_free_pages);
1018 #endif
1019 }
1020 
1021 void __init zone_sizes_init(void)
1022 {
1023 	unsigned long max_zone_pfns[MAX_NR_ZONES];
1024 
1025 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1026 
1027 #ifdef CONFIG_ZONE_DMA
1028 	max_zone_pfns[ZONE_DMA]		= min(MAX_DMA_PFN, max_low_pfn);
1029 #endif
1030 #ifdef CONFIG_ZONE_DMA32
1031 	max_zone_pfns[ZONE_DMA32]	= min(MAX_DMA32_PFN, max_low_pfn);
1032 #endif
1033 	max_zone_pfns[ZONE_NORMAL]	= max_low_pfn;
1034 #ifdef CONFIG_HIGHMEM
1035 	max_zone_pfns[ZONE_HIGHMEM]	= max_pfn;
1036 #endif
1037 
1038 	free_area_init(max_zone_pfns);
1039 }
1040 
1041 __visible DEFINE_PER_CPU_ALIGNED(struct tlb_state, cpu_tlbstate) = {
1042 	.loaded_mm = &init_mm,
1043 	.next_asid = 1,
1044 	.cr4 = ~0UL,	/* fail hard if we screw up cr4 shadow initialization */
1045 };
1046 
1047 void update_cache_mode_entry(unsigned entry, enum page_cache_mode cache)
1048 {
1049 	/* entry 0 MUST be WB (hardwired to speed up translations) */
1050 	BUG_ON(!entry && cache != _PAGE_CACHE_MODE_WB);
1051 
1052 	__cachemode2pte_tbl[cache] = __cm_idx2pte(entry);
1053 	__pte2cachemode_tbl[entry] = cache;
1054 }
1055 
1056 #ifdef CONFIG_SWAP
1057 unsigned long max_swapfile_size(void)
1058 {
1059 	unsigned long pages;
1060 
1061 	pages = generic_max_swapfile_size();
1062 
1063 	if (boot_cpu_has_bug(X86_BUG_L1TF) && l1tf_mitigation != L1TF_MITIGATION_OFF) {
1064 		/* Limit the swap file size to MAX_PA/2 for L1TF workaround */
1065 		unsigned long long l1tf_limit = l1tf_pfn_limit();
1066 		/*
1067 		 * We encode swap offsets also with 3 bits below those for pfn
1068 		 * which makes the usable limit higher.
1069 		 */
1070 #if CONFIG_PGTABLE_LEVELS > 2
1071 		l1tf_limit <<= PAGE_SHIFT - SWP_OFFSET_FIRST_BIT;
1072 #endif
1073 		pages = min_t(unsigned long long, l1tf_limit, pages);
1074 	}
1075 	return pages;
1076 }
1077 #endif
1078