xref: /linux/arch/x86/mm/init.c (revision 0e2b2a76278153d1ac312b0691cb65dabb9aef3e)
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/cpu_device_id.h>
13 #include <asm/e820/api.h>
14 #include <asm/init.h>
15 #include <asm/page.h>
16 #include <asm/page_types.h>
17 #include <asm/sections.h>
18 #include <asm/setup.h>
19 #include <asm/tlbflush.h>
20 #include <asm/tlb.h>
21 #include <asm/proto.h>
22 #include <asm/dma.h>		/* for MAX_DMA_PFN */
23 #include <asm/microcode.h>
24 #include <asm/kaslr.h>
25 #include <asm/hypervisor.h>
26 #include <asm/cpufeature.h>
27 #include <asm/pti.h>
28 #include <asm/text-patching.h>
29 #include <asm/memtype.h>
30 #include <asm/paravirt.h>
31 
32 /*
33  * We need to define the tracepoints somewhere, and tlb.c
34  * is only compiled when SMP=y.
35  */
36 #include <trace/events/tlb.h>
37 
38 #include "mm_internal.h"
39 
40 /*
41  * Tables translating between page_cache_type_t and pte encoding.
42  *
43  * The default values are defined statically as minimal supported mode;
44  * WC and WT fall back to UC-.  pat_init() updates these values to support
45  * more cache modes, WC and WT, when it is safe to do so.  See pat_init()
46  * for the details.  Note, __early_ioremap() used during early boot-time
47  * takes pgprot_t (pte encoding) and does not use these tables.
48  *
49  *   Index into __cachemode2pte_tbl[] is the cachemode.
50  *
51  *   Index into __pte2cachemode_tbl[] are the caching attribute bits of the pte
52  *   (_PAGE_PWT, _PAGE_PCD, _PAGE_PAT) at index bit positions 0, 1, 2.
53  */
54 static uint16_t __cachemode2pte_tbl[_PAGE_CACHE_MODE_NUM] = {
55 	[_PAGE_CACHE_MODE_WB      ]	= 0         | 0        ,
56 	[_PAGE_CACHE_MODE_WC      ]	= 0         | _PAGE_PCD,
57 	[_PAGE_CACHE_MODE_UC_MINUS]	= 0         | _PAGE_PCD,
58 	[_PAGE_CACHE_MODE_UC      ]	= _PAGE_PWT | _PAGE_PCD,
59 	[_PAGE_CACHE_MODE_WT      ]	= 0         | _PAGE_PCD,
60 	[_PAGE_CACHE_MODE_WP      ]	= 0         | _PAGE_PCD,
61 };
62 
63 unsigned long cachemode2protval(enum page_cache_mode pcm)
64 {
65 	if (likely(pcm == 0))
66 		return 0;
67 	return __cachemode2pte_tbl[pcm];
68 }
69 EXPORT_SYMBOL(cachemode2protval);
70 
71 static uint8_t __pte2cachemode_tbl[8] = {
72 	[__pte2cm_idx( 0        | 0         | 0        )] = _PAGE_CACHE_MODE_WB,
73 	[__pte2cm_idx(_PAGE_PWT | 0         | 0        )] = _PAGE_CACHE_MODE_UC_MINUS,
74 	[__pte2cm_idx( 0        | _PAGE_PCD | 0        )] = _PAGE_CACHE_MODE_UC_MINUS,
75 	[__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | 0        )] = _PAGE_CACHE_MODE_UC,
76 	[__pte2cm_idx( 0        | 0         | _PAGE_PAT)] = _PAGE_CACHE_MODE_WB,
77 	[__pte2cm_idx(_PAGE_PWT | 0         | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
78 	[__pte2cm_idx(0         | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
79 	[__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC,
80 };
81 
82 /*
83  * Check that the write-protect PAT entry is set for write-protect.
84  * To do this without making assumptions how PAT has been set up (Xen has
85  * another layout than the kernel), translate the _PAGE_CACHE_MODE_WP cache
86  * mode via the __cachemode2pte_tbl[] into protection bits (those protection
87  * bits will select a cache mode of WP or better), and then translate the
88  * protection bits back into the cache mode using __pte2cm_idx() and the
89  * __pte2cachemode_tbl[] array. This will return the really used cache mode.
90  */
91 bool x86_has_pat_wp(void)
92 {
93 	uint16_t prot = __cachemode2pte_tbl[_PAGE_CACHE_MODE_WP];
94 
95 	return __pte2cachemode_tbl[__pte2cm_idx(prot)] == _PAGE_CACHE_MODE_WP;
96 }
97 
98 enum page_cache_mode pgprot2cachemode(pgprot_t pgprot)
99 {
100 	unsigned long masked;
101 
102 	masked = pgprot_val(pgprot) & _PAGE_CACHE_MASK;
103 	if (likely(masked == 0))
104 		return 0;
105 	return __pte2cachemode_tbl[__pte2cm_idx(masked)];
106 }
107 
108 static unsigned long __initdata pgt_buf_start;
109 static unsigned long __initdata pgt_buf_end;
110 static unsigned long __initdata pgt_buf_top;
111 
112 static unsigned long min_pfn_mapped;
113 
114 static bool __initdata can_use_brk_pgt = true;
115 
116 /*
117  * Pages returned are already directly mapped.
118  *
119  * Changing that is likely to break Xen, see commit:
120  *
121  *    279b706 x86,xen: introduce x86_init.mapping.pagetable_reserve
122  *
123  * for detailed information.
124  */
125 __ref void *alloc_low_pages(unsigned int num)
126 {
127 	unsigned long pfn;
128 	int i;
129 
130 	if (after_bootmem) {
131 		unsigned int order;
132 
133 		order = get_order((unsigned long)num << PAGE_SHIFT);
134 		return (void *)__get_free_pages(GFP_ATOMIC | __GFP_ZERO, order);
135 	}
136 
137 	if ((pgt_buf_end + num) > pgt_buf_top || !can_use_brk_pgt) {
138 		unsigned long ret = 0;
139 
140 		if (min_pfn_mapped < max_pfn_mapped) {
141 			ret = memblock_phys_alloc_range(
142 					PAGE_SIZE * num, PAGE_SIZE,
143 					min_pfn_mapped << PAGE_SHIFT,
144 					max_pfn_mapped << PAGE_SHIFT);
145 		}
146 		if (!ret && can_use_brk_pgt)
147 			ret = __pa(extend_brk(PAGE_SIZE * num, PAGE_SIZE));
148 
149 		if (!ret)
150 			panic("alloc_low_pages: can not alloc memory");
151 
152 		pfn = ret >> PAGE_SHIFT;
153 	} else {
154 		pfn = pgt_buf_end;
155 		pgt_buf_end += num;
156 	}
157 
158 	for (i = 0; i < num; i++) {
159 		void *adr;
160 
161 		adr = __va((pfn + i) << PAGE_SHIFT);
162 		clear_page(adr);
163 	}
164 
165 	return __va(pfn << PAGE_SHIFT);
166 }
167 
168 /*
169  * By default need to be able to allocate page tables below PGD firstly for
170  * the 0-ISA_END_ADDRESS range and secondly for the initial PMD_SIZE mapping.
171  * With KASLR memory randomization, depending on the machine e820 memory and the
172  * PUD alignment, twice that many pages may be needed when KASLR memory
173  * randomization is enabled.
174  */
175 
176 #ifndef CONFIG_X86_5LEVEL
177 #define INIT_PGD_PAGE_TABLES    3
178 #else
179 #define INIT_PGD_PAGE_TABLES    4
180 #endif
181 
182 #ifndef CONFIG_RANDOMIZE_MEMORY
183 #define INIT_PGD_PAGE_COUNT      (2 * INIT_PGD_PAGE_TABLES)
184 #else
185 #define INIT_PGD_PAGE_COUNT      (4 * INIT_PGD_PAGE_TABLES)
186 #endif
187 
188 #define INIT_PGT_BUF_SIZE	(INIT_PGD_PAGE_COUNT * PAGE_SIZE)
189 RESERVE_BRK(early_pgt_alloc, INIT_PGT_BUF_SIZE);
190 void  __init early_alloc_pgt_buf(void)
191 {
192 	unsigned long tables = INIT_PGT_BUF_SIZE;
193 	phys_addr_t base;
194 
195 	base = __pa(extend_brk(tables, PAGE_SIZE));
196 
197 	pgt_buf_start = base >> PAGE_SHIFT;
198 	pgt_buf_end = pgt_buf_start;
199 	pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
200 }
201 
202 int after_bootmem;
203 
204 early_param_on_off("gbpages", "nogbpages", direct_gbpages, CONFIG_X86_DIRECT_GBPAGES);
205 
206 struct map_range {
207 	unsigned long start;
208 	unsigned long end;
209 	unsigned page_size_mask;
210 };
211 
212 static int page_size_mask;
213 
214 /*
215  * Save some of cr4 feature set we're using (e.g.  Pentium 4MB
216  * enable and PPro Global page enable), so that any CPU's that boot
217  * up after us can get the correct flags. Invoked on the boot CPU.
218  */
219 static inline void cr4_set_bits_and_update_boot(unsigned long mask)
220 {
221 	mmu_cr4_features |= mask;
222 	if (trampoline_cr4_features)
223 		*trampoline_cr4_features = mmu_cr4_features;
224 	cr4_set_bits(mask);
225 }
226 
227 static void __init probe_page_size_mask(void)
228 {
229 	/*
230 	 * For pagealloc debugging, identity mapping will use small pages.
231 	 * This will simplify cpa(), which otherwise needs to support splitting
232 	 * large pages into small in interrupt context, etc.
233 	 */
234 	if (boot_cpu_has(X86_FEATURE_PSE) && !debug_pagealloc_enabled())
235 		page_size_mask |= 1 << PG_LEVEL_2M;
236 	else
237 		direct_gbpages = 0;
238 
239 	/* Enable PSE if available */
240 	if (boot_cpu_has(X86_FEATURE_PSE))
241 		cr4_set_bits_and_update_boot(X86_CR4_PSE);
242 
243 	/* Enable PGE if available */
244 	__supported_pte_mask &= ~_PAGE_GLOBAL;
245 	if (boot_cpu_has(X86_FEATURE_PGE)) {
246 		cr4_set_bits_and_update_boot(X86_CR4_PGE);
247 		__supported_pte_mask |= _PAGE_GLOBAL;
248 	}
249 
250 	/* By the default is everything supported: */
251 	__default_kernel_pte_mask = __supported_pte_mask;
252 	/* Except when with PTI where the kernel is mostly non-Global: */
253 	if (cpu_feature_enabled(X86_FEATURE_PTI))
254 		__default_kernel_pte_mask &= ~_PAGE_GLOBAL;
255 
256 	/* Enable 1 GB linear kernel mappings if available: */
257 	if (direct_gbpages && boot_cpu_has(X86_FEATURE_GBPAGES)) {
258 		printk(KERN_INFO "Using GB pages for direct mapping\n");
259 		page_size_mask |= 1 << PG_LEVEL_1G;
260 	} else {
261 		direct_gbpages = 0;
262 	}
263 }
264 
265 #define INTEL_MATCH(_model) { .vendor  = X86_VENDOR_INTEL,	\
266 			      .family  = 6,			\
267 			      .model = _model,			\
268 			    }
269 /*
270  * INVLPG may not properly flush Global entries
271  * on these CPUs when PCIDs are enabled.
272  */
273 static const struct x86_cpu_id invlpg_miss_ids[] = {
274 	INTEL_MATCH(INTEL_FAM6_ALDERLAKE   ),
275 	INTEL_MATCH(INTEL_FAM6_ALDERLAKE_L ),
276 	INTEL_MATCH(INTEL_FAM6_ALDERLAKE_N ),
277 	INTEL_MATCH(INTEL_FAM6_RAPTORLAKE  ),
278 	INTEL_MATCH(INTEL_FAM6_RAPTORLAKE_P),
279 	INTEL_MATCH(INTEL_FAM6_RAPTORLAKE_S),
280 	{}
281 };
282 
283 static void setup_pcid(void)
284 {
285 	if (!IS_ENABLED(CONFIG_X86_64))
286 		return;
287 
288 	if (!boot_cpu_has(X86_FEATURE_PCID))
289 		return;
290 
291 	if (x86_match_cpu(invlpg_miss_ids)) {
292 		pr_info("Incomplete global flushes, disabling PCID");
293 		setup_clear_cpu_cap(X86_FEATURE_PCID);
294 		return;
295 	}
296 
297 	if (boot_cpu_has(X86_FEATURE_PGE)) {
298 		/*
299 		 * This can't be cr4_set_bits_and_update_boot() -- the
300 		 * trampoline code can't handle CR4.PCIDE and it wouldn't
301 		 * do any good anyway.  Despite the name,
302 		 * cr4_set_bits_and_update_boot() doesn't actually cause
303 		 * the bits in question to remain set all the way through
304 		 * the secondary boot asm.
305 		 *
306 		 * Instead, we brute-force it and set CR4.PCIDE manually in
307 		 * start_secondary().
308 		 */
309 		cr4_set_bits(X86_CR4_PCIDE);
310 
311 		/*
312 		 * INVPCID's single-context modes (2/3) only work if we set
313 		 * X86_CR4_PCIDE, *and* we INVPCID support.  It's unusable
314 		 * on systems that have X86_CR4_PCIDE clear, or that have
315 		 * no INVPCID support at all.
316 		 */
317 		if (boot_cpu_has(X86_FEATURE_INVPCID))
318 			setup_force_cpu_cap(X86_FEATURE_INVPCID_SINGLE);
319 	} else {
320 		/*
321 		 * flush_tlb_all(), as currently implemented, won't work if
322 		 * PCID is on but PGE is not.  Since that combination
323 		 * doesn't exist on real hardware, there's no reason to try
324 		 * to fully support it, but it's polite to avoid corrupting
325 		 * data if we're on an improperly configured VM.
326 		 */
327 		setup_clear_cpu_cap(X86_FEATURE_PCID);
328 	}
329 }
330 
331 #ifdef CONFIG_X86_32
332 #define NR_RANGE_MR 3
333 #else /* CONFIG_X86_64 */
334 #define NR_RANGE_MR 5
335 #endif
336 
337 static int __meminit save_mr(struct map_range *mr, int nr_range,
338 			     unsigned long start_pfn, unsigned long end_pfn,
339 			     unsigned long page_size_mask)
340 {
341 	if (start_pfn < end_pfn) {
342 		if (nr_range >= NR_RANGE_MR)
343 			panic("run out of range for init_memory_mapping\n");
344 		mr[nr_range].start = start_pfn<<PAGE_SHIFT;
345 		mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
346 		mr[nr_range].page_size_mask = page_size_mask;
347 		nr_range++;
348 	}
349 
350 	return nr_range;
351 }
352 
353 /*
354  * adjust the page_size_mask for small range to go with
355  *	big page size instead small one if nearby are ram too.
356  */
357 static void __ref adjust_range_page_size_mask(struct map_range *mr,
358 							 int nr_range)
359 {
360 	int i;
361 
362 	for (i = 0; i < nr_range; i++) {
363 		if ((page_size_mask & (1<<PG_LEVEL_2M)) &&
364 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) {
365 			unsigned long start = round_down(mr[i].start, PMD_SIZE);
366 			unsigned long end = round_up(mr[i].end, PMD_SIZE);
367 
368 #ifdef CONFIG_X86_32
369 			if ((end >> PAGE_SHIFT) > max_low_pfn)
370 				continue;
371 #endif
372 
373 			if (memblock_is_region_memory(start, end - start))
374 				mr[i].page_size_mask |= 1<<PG_LEVEL_2M;
375 		}
376 		if ((page_size_mask & (1<<PG_LEVEL_1G)) &&
377 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) {
378 			unsigned long start = round_down(mr[i].start, PUD_SIZE);
379 			unsigned long end = round_up(mr[i].end, PUD_SIZE);
380 
381 			if (memblock_is_region_memory(start, end - start))
382 				mr[i].page_size_mask |= 1<<PG_LEVEL_1G;
383 		}
384 	}
385 }
386 
387 static const char *page_size_string(struct map_range *mr)
388 {
389 	static const char str_1g[] = "1G";
390 	static const char str_2m[] = "2M";
391 	static const char str_4m[] = "4M";
392 	static const char str_4k[] = "4k";
393 
394 	if (mr->page_size_mask & (1<<PG_LEVEL_1G))
395 		return str_1g;
396 	/*
397 	 * 32-bit without PAE has a 4M large page size.
398 	 * PG_LEVEL_2M is misnamed, but we can at least
399 	 * print out the right size in the string.
400 	 */
401 	if (IS_ENABLED(CONFIG_X86_32) &&
402 	    !IS_ENABLED(CONFIG_X86_PAE) &&
403 	    mr->page_size_mask & (1<<PG_LEVEL_2M))
404 		return str_4m;
405 
406 	if (mr->page_size_mask & (1<<PG_LEVEL_2M))
407 		return str_2m;
408 
409 	return str_4k;
410 }
411 
412 static int __meminit split_mem_range(struct map_range *mr, int nr_range,
413 				     unsigned long start,
414 				     unsigned long end)
415 {
416 	unsigned long start_pfn, end_pfn, limit_pfn;
417 	unsigned long pfn;
418 	int i;
419 
420 	limit_pfn = PFN_DOWN(end);
421 
422 	/* head if not big page alignment ? */
423 	pfn = start_pfn = PFN_DOWN(start);
424 #ifdef CONFIG_X86_32
425 	/*
426 	 * Don't use a large page for the first 2/4MB of memory
427 	 * because there are often fixed size MTRRs in there
428 	 * and overlapping MTRRs into large pages can cause
429 	 * slowdowns.
430 	 */
431 	if (pfn == 0)
432 		end_pfn = PFN_DOWN(PMD_SIZE);
433 	else
434 		end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
435 #else /* CONFIG_X86_64 */
436 	end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
437 #endif
438 	if (end_pfn > limit_pfn)
439 		end_pfn = limit_pfn;
440 	if (start_pfn < end_pfn) {
441 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
442 		pfn = end_pfn;
443 	}
444 
445 	/* big page (2M) range */
446 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
447 #ifdef CONFIG_X86_32
448 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
449 #else /* CONFIG_X86_64 */
450 	end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
451 	if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE)))
452 		end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
453 #endif
454 
455 	if (start_pfn < end_pfn) {
456 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
457 				page_size_mask & (1<<PG_LEVEL_2M));
458 		pfn = end_pfn;
459 	}
460 
461 #ifdef CONFIG_X86_64
462 	/* big page (1G) range */
463 	start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
464 	end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE));
465 	if (start_pfn < end_pfn) {
466 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
467 				page_size_mask &
468 				 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
469 		pfn = end_pfn;
470 	}
471 
472 	/* tail is not big page (1G) alignment */
473 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
474 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
475 	if (start_pfn < end_pfn) {
476 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
477 				page_size_mask & (1<<PG_LEVEL_2M));
478 		pfn = end_pfn;
479 	}
480 #endif
481 
482 	/* tail is not big page (2M) alignment */
483 	start_pfn = pfn;
484 	end_pfn = limit_pfn;
485 	nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
486 
487 	if (!after_bootmem)
488 		adjust_range_page_size_mask(mr, nr_range);
489 
490 	/* try to merge same page size and continuous */
491 	for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
492 		unsigned long old_start;
493 		if (mr[i].end != mr[i+1].start ||
494 		    mr[i].page_size_mask != mr[i+1].page_size_mask)
495 			continue;
496 		/* move it */
497 		old_start = mr[i].start;
498 		memmove(&mr[i], &mr[i+1],
499 			(nr_range - 1 - i) * sizeof(struct map_range));
500 		mr[i--].start = old_start;
501 		nr_range--;
502 	}
503 
504 	for (i = 0; i < nr_range; i++)
505 		pr_debug(" [mem %#010lx-%#010lx] page %s\n",
506 				mr[i].start, mr[i].end - 1,
507 				page_size_string(&mr[i]));
508 
509 	return nr_range;
510 }
511 
512 struct range pfn_mapped[E820_MAX_ENTRIES];
513 int nr_pfn_mapped;
514 
515 static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn)
516 {
517 	nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_MAX_ENTRIES,
518 					     nr_pfn_mapped, start_pfn, end_pfn);
519 	nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_MAX_ENTRIES);
520 
521 	max_pfn_mapped = max(max_pfn_mapped, end_pfn);
522 
523 	if (start_pfn < (1UL<<(32-PAGE_SHIFT)))
524 		max_low_pfn_mapped = max(max_low_pfn_mapped,
525 					 min(end_pfn, 1UL<<(32-PAGE_SHIFT)));
526 }
527 
528 bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn)
529 {
530 	int i;
531 
532 	for (i = 0; i < nr_pfn_mapped; i++)
533 		if ((start_pfn >= pfn_mapped[i].start) &&
534 		    (end_pfn <= pfn_mapped[i].end))
535 			return true;
536 
537 	return false;
538 }
539 
540 /*
541  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
542  * This runs before bootmem is initialized and gets pages directly from
543  * the physical memory. To access them they are temporarily mapped.
544  */
545 unsigned long __ref init_memory_mapping(unsigned long start,
546 					unsigned long end, pgprot_t prot)
547 {
548 	struct map_range mr[NR_RANGE_MR];
549 	unsigned long ret = 0;
550 	int nr_range, i;
551 
552 	pr_debug("init_memory_mapping: [mem %#010lx-%#010lx]\n",
553 	       start, end - 1);
554 
555 	memset(mr, 0, sizeof(mr));
556 	nr_range = split_mem_range(mr, 0, start, end);
557 
558 	for (i = 0; i < nr_range; i++)
559 		ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
560 						   mr[i].page_size_mask,
561 						   prot);
562 
563 	add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT);
564 
565 	return ret >> PAGE_SHIFT;
566 }
567 
568 /*
569  * We need to iterate through the E820 memory map and create direct mappings
570  * for only E820_TYPE_RAM and E820_KERN_RESERVED regions. We cannot simply
571  * create direct mappings for all pfns from [0 to max_low_pfn) and
572  * [4GB to max_pfn) because of possible memory holes in high addresses
573  * that cannot be marked as UC by fixed/variable range MTRRs.
574  * Depending on the alignment of E820 ranges, this may possibly result
575  * in using smaller size (i.e. 4K instead of 2M or 1G) page tables.
576  *
577  * init_mem_mapping() calls init_range_memory_mapping() with big range.
578  * That range would have hole in the middle or ends, and only ram parts
579  * will be mapped in init_range_memory_mapping().
580  */
581 static unsigned long __init init_range_memory_mapping(
582 					   unsigned long r_start,
583 					   unsigned long r_end)
584 {
585 	unsigned long start_pfn, end_pfn;
586 	unsigned long mapped_ram_size = 0;
587 	int i;
588 
589 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
590 		u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end);
591 		u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end);
592 		if (start >= end)
593 			continue;
594 
595 		/*
596 		 * if it is overlapping with brk pgt, we need to
597 		 * alloc pgt buf from memblock instead.
598 		 */
599 		can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >=
600 				    min(end, (u64)pgt_buf_top<<PAGE_SHIFT);
601 		init_memory_mapping(start, end, PAGE_KERNEL);
602 		mapped_ram_size += end - start;
603 		can_use_brk_pgt = true;
604 	}
605 
606 	return mapped_ram_size;
607 }
608 
609 static unsigned long __init get_new_step_size(unsigned long step_size)
610 {
611 	/*
612 	 * Initial mapped size is PMD_SIZE (2M).
613 	 * We can not set step_size to be PUD_SIZE (1G) yet.
614 	 * In worse case, when we cross the 1G boundary, and
615 	 * PG_LEVEL_2M is not set, we will need 1+1+512 pages (2M + 8k)
616 	 * to map 1G range with PTE. Hence we use one less than the
617 	 * difference of page table level shifts.
618 	 *
619 	 * Don't need to worry about overflow in the top-down case, on 32bit,
620 	 * when step_size is 0, round_down() returns 0 for start, and that
621 	 * turns it into 0x100000000ULL.
622 	 * In the bottom-up case, round_up(x, 0) returns 0 though too, which
623 	 * needs to be taken into consideration by the code below.
624 	 */
625 	return step_size << (PMD_SHIFT - PAGE_SHIFT - 1);
626 }
627 
628 /**
629  * memory_map_top_down - Map [map_start, map_end) top down
630  * @map_start: start address of the target memory range
631  * @map_end: end address of the target memory range
632  *
633  * This function will setup direct mapping for memory range
634  * [map_start, map_end) in top-down. That said, the page tables
635  * will be allocated at the end of the memory, and we map the
636  * memory in top-down.
637  */
638 static void __init memory_map_top_down(unsigned long map_start,
639 				       unsigned long map_end)
640 {
641 	unsigned long real_end, last_start;
642 	unsigned long step_size;
643 	unsigned long addr;
644 	unsigned long mapped_ram_size = 0;
645 
646 	/*
647 	 * Systems that have many reserved areas near top of the memory,
648 	 * e.g. QEMU with less than 1G RAM and EFI enabled, or Xen, will
649 	 * require lots of 4K mappings which may exhaust pgt_buf.
650 	 * Start with top-most PMD_SIZE range aligned at PMD_SIZE to ensure
651 	 * there is enough mapped memory that can be allocated from
652 	 * memblock.
653 	 */
654 	addr = memblock_phys_alloc_range(PMD_SIZE, PMD_SIZE, map_start,
655 					 map_end);
656 	memblock_phys_free(addr, PMD_SIZE);
657 	real_end = addr + PMD_SIZE;
658 
659 	/* step_size need to be small so pgt_buf from BRK could cover it */
660 	step_size = PMD_SIZE;
661 	max_pfn_mapped = 0; /* will get exact value next */
662 	min_pfn_mapped = real_end >> PAGE_SHIFT;
663 	last_start = real_end;
664 
665 	/*
666 	 * We start from the top (end of memory) and go to the bottom.
667 	 * The memblock_find_in_range() gets us a block of RAM from the
668 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
669 	 * for page table.
670 	 */
671 	while (last_start > map_start) {
672 		unsigned long start;
673 
674 		if (last_start > step_size) {
675 			start = round_down(last_start - 1, step_size);
676 			if (start < map_start)
677 				start = map_start;
678 		} else
679 			start = map_start;
680 		mapped_ram_size += init_range_memory_mapping(start,
681 							last_start);
682 		last_start = start;
683 		min_pfn_mapped = last_start >> PAGE_SHIFT;
684 		if (mapped_ram_size >= step_size)
685 			step_size = get_new_step_size(step_size);
686 	}
687 
688 	if (real_end < map_end)
689 		init_range_memory_mapping(real_end, map_end);
690 }
691 
692 /**
693  * memory_map_bottom_up - Map [map_start, map_end) bottom up
694  * @map_start: start address of the target memory range
695  * @map_end: end address of the target memory range
696  *
697  * This function will setup direct mapping for memory range
698  * [map_start, map_end) in bottom-up. Since we have limited the
699  * bottom-up allocation above the kernel, the page tables will
700  * be allocated just above the kernel and we map the memory
701  * in [map_start, map_end) in bottom-up.
702  */
703 static void __init memory_map_bottom_up(unsigned long map_start,
704 					unsigned long map_end)
705 {
706 	unsigned long next, start;
707 	unsigned long mapped_ram_size = 0;
708 	/* step_size need to be small so pgt_buf from BRK could cover it */
709 	unsigned long step_size = PMD_SIZE;
710 
711 	start = map_start;
712 	min_pfn_mapped = start >> PAGE_SHIFT;
713 
714 	/*
715 	 * We start from the bottom (@map_start) and go to the top (@map_end).
716 	 * The memblock_find_in_range() gets us a block of RAM from the
717 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
718 	 * for page table.
719 	 */
720 	while (start < map_end) {
721 		if (step_size && map_end - start > step_size) {
722 			next = round_up(start + 1, step_size);
723 			if (next > map_end)
724 				next = map_end;
725 		} else {
726 			next = map_end;
727 		}
728 
729 		mapped_ram_size += init_range_memory_mapping(start, next);
730 		start = next;
731 
732 		if (mapped_ram_size >= step_size)
733 			step_size = get_new_step_size(step_size);
734 	}
735 }
736 
737 /*
738  * The real mode trampoline, which is required for bootstrapping CPUs
739  * occupies only a small area under the low 1MB.  See reserve_real_mode()
740  * for details.
741  *
742  * If KASLR is disabled the first PGD entry of the direct mapping is copied
743  * to map the real mode trampoline.
744  *
745  * If KASLR is enabled, copy only the PUD which covers the low 1MB
746  * area. This limits the randomization granularity to 1GB for both 4-level
747  * and 5-level paging.
748  */
749 static void __init init_trampoline(void)
750 {
751 #ifdef CONFIG_X86_64
752 	/*
753 	 * The code below will alias kernel page-tables in the user-range of the
754 	 * address space, including the Global bit. So global TLB entries will
755 	 * be created when using the trampoline page-table.
756 	 */
757 	if (!kaslr_memory_enabled())
758 		trampoline_pgd_entry = init_top_pgt[pgd_index(__PAGE_OFFSET)];
759 	else
760 		init_trampoline_kaslr();
761 #endif
762 }
763 
764 void __init init_mem_mapping(void)
765 {
766 	unsigned long end;
767 
768 	pti_check_boottime_disable();
769 	probe_page_size_mask();
770 	setup_pcid();
771 
772 #ifdef CONFIG_X86_64
773 	end = max_pfn << PAGE_SHIFT;
774 #else
775 	end = max_low_pfn << PAGE_SHIFT;
776 #endif
777 
778 	/* the ISA range is always mapped regardless of memory holes */
779 	init_memory_mapping(0, ISA_END_ADDRESS, PAGE_KERNEL);
780 
781 	/* Init the trampoline, possibly with KASLR memory offset */
782 	init_trampoline();
783 
784 	/*
785 	 * If the allocation is in bottom-up direction, we setup direct mapping
786 	 * in bottom-up, otherwise we setup direct mapping in top-down.
787 	 */
788 	if (memblock_bottom_up()) {
789 		unsigned long kernel_end = __pa_symbol(_end);
790 
791 		/*
792 		 * we need two separate calls here. This is because we want to
793 		 * allocate page tables above the kernel. So we first map
794 		 * [kernel_end, end) to make memory above the kernel be mapped
795 		 * as soon as possible. And then use page tables allocated above
796 		 * the kernel to map [ISA_END_ADDRESS, kernel_end).
797 		 */
798 		memory_map_bottom_up(kernel_end, end);
799 		memory_map_bottom_up(ISA_END_ADDRESS, kernel_end);
800 	} else {
801 		memory_map_top_down(ISA_END_ADDRESS, end);
802 	}
803 
804 #ifdef CONFIG_X86_64
805 	if (max_pfn > max_low_pfn) {
806 		/* can we preserve max_low_pfn ?*/
807 		max_low_pfn = max_pfn;
808 	}
809 #else
810 	early_ioremap_page_table_range_init();
811 #endif
812 
813 	load_cr3(swapper_pg_dir);
814 	__flush_tlb_all();
815 
816 	x86_init.hyper.init_mem_mapping();
817 
818 	early_memtest(0, max_pfn_mapped << PAGE_SHIFT);
819 }
820 
821 /*
822  * Initialize an mm_struct to be used during poking and a pointer to be used
823  * during patching.
824  */
825 void __init poking_init(void)
826 {
827 	spinlock_t *ptl;
828 	pte_t *ptep;
829 
830 	poking_mm = mm_alloc();
831 	BUG_ON(!poking_mm);
832 
833 	/* Xen PV guests need the PGD to be pinned. */
834 	paravirt_enter_mmap(poking_mm);
835 
836 	/*
837 	 * Randomize the poking address, but make sure that the following page
838 	 * will be mapped at the same PMD. We need 2 pages, so find space for 3,
839 	 * and adjust the address if the PMD ends after the first one.
840 	 */
841 	poking_addr = TASK_UNMAPPED_BASE;
842 	if (IS_ENABLED(CONFIG_RANDOMIZE_BASE))
843 		poking_addr += (kaslr_get_random_long("Poking") & PAGE_MASK) %
844 			(TASK_SIZE - TASK_UNMAPPED_BASE - 3 * PAGE_SIZE);
845 
846 	if (((poking_addr + PAGE_SIZE) & ~PMD_MASK) == 0)
847 		poking_addr += PAGE_SIZE;
848 
849 	/*
850 	 * We need to trigger the allocation of the page-tables that will be
851 	 * needed for poking now. Later, poking may be performed in an atomic
852 	 * section, which might cause allocation to fail.
853 	 */
854 	ptep = get_locked_pte(poking_mm, poking_addr, &ptl);
855 	BUG_ON(!ptep);
856 	pte_unmap_unlock(ptep, ptl);
857 }
858 
859 /*
860  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
861  * is valid. The argument is a physical page number.
862  *
863  * On x86, access has to be given to the first megabyte of RAM because that
864  * area traditionally contains BIOS code and data regions used by X, dosemu,
865  * and similar apps. Since they map the entire memory range, the whole range
866  * must be allowed (for mapping), but any areas that would otherwise be
867  * disallowed are flagged as being "zero filled" instead of rejected.
868  * Access has to be given to non-kernel-ram areas as well, these contain the
869  * PCI mmio resources as well as potential bios/acpi data regions.
870  */
871 int devmem_is_allowed(unsigned long pagenr)
872 {
873 	if (region_intersects(PFN_PHYS(pagenr), PAGE_SIZE,
874 				IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE)
875 			!= REGION_DISJOINT) {
876 		/*
877 		 * For disallowed memory regions in the low 1MB range,
878 		 * request that the page be shown as all zeros.
879 		 */
880 		if (pagenr < 256)
881 			return 2;
882 
883 		return 0;
884 	}
885 
886 	/*
887 	 * This must follow RAM test, since System RAM is considered a
888 	 * restricted resource under CONFIG_STRICT_DEVMEM.
889 	 */
890 	if (iomem_is_exclusive(pagenr << PAGE_SHIFT)) {
891 		/* Low 1MB bypasses iomem restrictions. */
892 		if (pagenr < 256)
893 			return 1;
894 
895 		return 0;
896 	}
897 
898 	return 1;
899 }
900 
901 void free_init_pages(const char *what, unsigned long begin, unsigned long end)
902 {
903 	unsigned long begin_aligned, end_aligned;
904 
905 	/* Make sure boundaries are page aligned */
906 	begin_aligned = PAGE_ALIGN(begin);
907 	end_aligned   = end & PAGE_MASK;
908 
909 	if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
910 		begin = begin_aligned;
911 		end   = end_aligned;
912 	}
913 
914 	if (begin >= end)
915 		return;
916 
917 	/*
918 	 * If debugging page accesses then do not free this memory but
919 	 * mark them not present - any buggy init-section access will
920 	 * create a kernel page fault:
921 	 */
922 	if (debug_pagealloc_enabled()) {
923 		pr_info("debug: unmapping init [mem %#010lx-%#010lx]\n",
924 			begin, end - 1);
925 		/*
926 		 * Inform kmemleak about the hole in the memory since the
927 		 * corresponding pages will be unmapped.
928 		 */
929 		kmemleak_free_part((void *)begin, end - begin);
930 		set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
931 	} else {
932 		/*
933 		 * We just marked the kernel text read only above, now that
934 		 * we are going to free part of that, we need to make that
935 		 * writeable and non-executable first.
936 		 */
937 		set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
938 		set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
939 
940 		free_reserved_area((void *)begin, (void *)end,
941 				   POISON_FREE_INITMEM, what);
942 	}
943 }
944 
945 /*
946  * begin/end can be in the direct map or the "high kernel mapping"
947  * used for the kernel image only.  free_init_pages() will do the
948  * right thing for either kind of address.
949  */
950 void free_kernel_image_pages(const char *what, void *begin, void *end)
951 {
952 	unsigned long begin_ul = (unsigned long)begin;
953 	unsigned long end_ul = (unsigned long)end;
954 	unsigned long len_pages = (end_ul - begin_ul) >> PAGE_SHIFT;
955 
956 	free_init_pages(what, begin_ul, end_ul);
957 
958 	/*
959 	 * PTI maps some of the kernel into userspace.  For performance,
960 	 * this includes some kernel areas that do not contain secrets.
961 	 * Those areas might be adjacent to the parts of the kernel image
962 	 * being freed, which may contain secrets.  Remove the "high kernel
963 	 * image mapping" for these freed areas, ensuring they are not even
964 	 * potentially vulnerable to Meltdown regardless of the specific
965 	 * optimizations PTI is currently using.
966 	 *
967 	 * The "noalias" prevents unmapping the direct map alias which is
968 	 * needed to access the freed pages.
969 	 *
970 	 * This is only valid for 64bit kernels. 32bit has only one mapping
971 	 * which can't be treated in this way for obvious reasons.
972 	 */
973 	if (IS_ENABLED(CONFIG_X86_64) && cpu_feature_enabled(X86_FEATURE_PTI))
974 		set_memory_np_noalias(begin_ul, len_pages);
975 }
976 
977 void __ref free_initmem(void)
978 {
979 	e820__reallocate_tables();
980 
981 	mem_encrypt_free_decrypted_mem();
982 
983 	free_kernel_image_pages("unused kernel image (initmem)",
984 				&__init_begin, &__init_end);
985 }
986 
987 #ifdef CONFIG_BLK_DEV_INITRD
988 void __init free_initrd_mem(unsigned long start, unsigned long end)
989 {
990 	/*
991 	 * end could be not aligned, and We can not align that,
992 	 * decompressor could be confused by aligned initrd_end
993 	 * We already reserve the end partial page before in
994 	 *   - i386_start_kernel()
995 	 *   - x86_64_start_kernel()
996 	 *   - relocate_initrd()
997 	 * So here We can do PAGE_ALIGN() safely to get partial page to be freed
998 	 */
999 	free_init_pages("initrd", start, PAGE_ALIGN(end));
1000 }
1001 #endif
1002 
1003 /*
1004  * Calculate the precise size of the DMA zone (first 16 MB of RAM),
1005  * and pass it to the MM layer - to help it set zone watermarks more
1006  * accurately.
1007  *
1008  * Done on 64-bit systems only for the time being, although 32-bit systems
1009  * might benefit from this as well.
1010  */
1011 void __init memblock_find_dma_reserve(void)
1012 {
1013 #ifdef CONFIG_X86_64
1014 	u64 nr_pages = 0, nr_free_pages = 0;
1015 	unsigned long start_pfn, end_pfn;
1016 	phys_addr_t start_addr, end_addr;
1017 	int i;
1018 	u64 u;
1019 
1020 	/*
1021 	 * Iterate over all memory ranges (free and reserved ones alike),
1022 	 * to calculate the total number of pages in the first 16 MB of RAM:
1023 	 */
1024 	nr_pages = 0;
1025 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
1026 		start_pfn = min(start_pfn, MAX_DMA_PFN);
1027 		end_pfn   = min(end_pfn,   MAX_DMA_PFN);
1028 
1029 		nr_pages += end_pfn - start_pfn;
1030 	}
1031 
1032 	/*
1033 	 * Iterate over free memory ranges to calculate the number of free
1034 	 * pages in the DMA zone, while not counting potential partial
1035 	 * pages at the beginning or the end of the range:
1036 	 */
1037 	nr_free_pages = 0;
1038 	for_each_free_mem_range(u, NUMA_NO_NODE, MEMBLOCK_NONE, &start_addr, &end_addr, NULL) {
1039 		start_pfn = min_t(unsigned long, PFN_UP(start_addr), MAX_DMA_PFN);
1040 		end_pfn   = min_t(unsigned long, PFN_DOWN(end_addr), MAX_DMA_PFN);
1041 
1042 		if (start_pfn < end_pfn)
1043 			nr_free_pages += end_pfn - start_pfn;
1044 	}
1045 
1046 	set_dma_reserve(nr_pages - nr_free_pages);
1047 #endif
1048 }
1049 
1050 void __init zone_sizes_init(void)
1051 {
1052 	unsigned long max_zone_pfns[MAX_NR_ZONES];
1053 
1054 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1055 
1056 #ifdef CONFIG_ZONE_DMA
1057 	max_zone_pfns[ZONE_DMA]		= min(MAX_DMA_PFN, max_low_pfn);
1058 #endif
1059 #ifdef CONFIG_ZONE_DMA32
1060 	max_zone_pfns[ZONE_DMA32]	= min(MAX_DMA32_PFN, max_low_pfn);
1061 #endif
1062 	max_zone_pfns[ZONE_NORMAL]	= max_low_pfn;
1063 #ifdef CONFIG_HIGHMEM
1064 	max_zone_pfns[ZONE_HIGHMEM]	= max_pfn;
1065 #endif
1066 
1067 	free_area_init(max_zone_pfns);
1068 }
1069 
1070 __visible DEFINE_PER_CPU_ALIGNED(struct tlb_state, cpu_tlbstate) = {
1071 	.loaded_mm = &init_mm,
1072 	.next_asid = 1,
1073 	.cr4 = ~0UL,	/* fail hard if we screw up cr4 shadow initialization */
1074 };
1075 
1076 #ifdef CONFIG_ADDRESS_MASKING
1077 DEFINE_PER_CPU(u64, tlbstate_untag_mask);
1078 EXPORT_PER_CPU_SYMBOL(tlbstate_untag_mask);
1079 #endif
1080 
1081 void update_cache_mode_entry(unsigned entry, enum page_cache_mode cache)
1082 {
1083 	/* entry 0 MUST be WB (hardwired to speed up translations) */
1084 	BUG_ON(!entry && cache != _PAGE_CACHE_MODE_WB);
1085 
1086 	__cachemode2pte_tbl[cache] = __cm_idx2pte(entry);
1087 	__pte2cachemode_tbl[entry] = cache;
1088 }
1089 
1090 #ifdef CONFIG_SWAP
1091 unsigned long arch_max_swapfile_size(void)
1092 {
1093 	unsigned long pages;
1094 
1095 	pages = generic_max_swapfile_size();
1096 
1097 	if (boot_cpu_has_bug(X86_BUG_L1TF) && l1tf_mitigation != L1TF_MITIGATION_OFF) {
1098 		/* Limit the swap file size to MAX_PA/2 for L1TF workaround */
1099 		unsigned long long l1tf_limit = l1tf_pfn_limit();
1100 		/*
1101 		 * We encode swap offsets also with 3 bits below those for pfn
1102 		 * which makes the usable limit higher.
1103 		 */
1104 #if CONFIG_PGTABLE_LEVELS > 2
1105 		l1tf_limit <<= PAGE_SHIFT - SWP_OFFSET_FIRST_BIT;
1106 #endif
1107 		pages = min_t(unsigned long long, l1tf_limit, pages);
1108 	}
1109 	return pages;
1110 }
1111 #endif
1112