xref: /linux/arch/x86/mm/init.c (revision 968e3000680713f712bcf02c51c4d7bb7d4d7685)
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 #include <linux/execmem.h>
11 
12 #include <asm/set_memory.h>
13 #include <asm/cpu_device_id.h>
14 #include <asm/e820/api.h>
15 #include <asm/init.h>
16 #include <asm/page.h>
17 #include <asm/page_types.h>
18 #include <asm/sections.h>
19 #include <asm/setup.h>
20 #include <asm/tlbflush.h>
21 #include <asm/tlb.h>
22 #include <asm/proto.h>
23 #include <asm/dma.h>		/* for MAX_DMA_PFN */
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 #include <asm/mmu_context.h>
32 
33 /*
34  * We need to define the tracepoints somewhere, and tlb.c
35  * is only compiled when SMP=y.
36  */
37 #include <trace/events/tlb.h>
38 
39 #include "mm_internal.h"
40 
41 /*
42  * Tables translating between page_cache_type_t and pte encoding.
43  *
44  * The default values are defined statically as minimal supported mode;
45  * WC and WT fall back to UC-.  pat_init() updates these values to support
46  * more cache modes, WC and WT, when it is safe to do so.  See pat_init()
47  * for the details.  Note, __early_ioremap() used during early boot-time
48  * takes pgprot_t (pte encoding) and does not use these tables.
49  *
50  *   Index into __cachemode2pte_tbl[] is the cachemode.
51  *
52  *   Index into __pte2cachemode_tbl[] are the caching attribute bits of the pte
53  *   (_PAGE_PWT, _PAGE_PCD, _PAGE_PAT) at index bit positions 0, 1, 2.
54  */
55 static uint16_t __cachemode2pte_tbl[_PAGE_CACHE_MODE_NUM] = {
56 	[_PAGE_CACHE_MODE_WB      ]	= 0         | 0        ,
57 	[_PAGE_CACHE_MODE_WC      ]	= 0         | _PAGE_PCD,
58 	[_PAGE_CACHE_MODE_UC_MINUS]	= 0         | _PAGE_PCD,
59 	[_PAGE_CACHE_MODE_UC      ]	= _PAGE_PWT | _PAGE_PCD,
60 	[_PAGE_CACHE_MODE_WT      ]	= 0         | _PAGE_PCD,
61 	[_PAGE_CACHE_MODE_WP      ]	= 0         | _PAGE_PCD,
62 };
63 
64 unsigned long cachemode2protval(enum page_cache_mode pcm)
65 {
66 	if (likely(pcm == 0))
67 		return 0;
68 	return __cachemode2pte_tbl[pcm];
69 }
70 EXPORT_SYMBOL(cachemode2protval);
71 
72 static uint8_t __pte2cachemode_tbl[8] = {
73 	[__pte2cm_idx( 0        | 0         | 0        )] = _PAGE_CACHE_MODE_WB,
74 	[__pte2cm_idx(_PAGE_PWT | 0         | 0        )] = _PAGE_CACHE_MODE_UC_MINUS,
75 	[__pte2cm_idx( 0        | _PAGE_PCD | 0        )] = _PAGE_CACHE_MODE_UC_MINUS,
76 	[__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | 0        )] = _PAGE_CACHE_MODE_UC,
77 	[__pte2cm_idx( 0        | 0         | _PAGE_PAT)] = _PAGE_CACHE_MODE_WB,
78 	[__pte2cm_idx(_PAGE_PWT | 0         | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
79 	[__pte2cm_idx(0         | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
80 	[__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC,
81 };
82 
83 /*
84  * Check that the write-protect PAT entry is set for write-protect.
85  * To do this without making assumptions how PAT has been set up (Xen has
86  * another layout than the kernel), translate the _PAGE_CACHE_MODE_WP cache
87  * mode via the __cachemode2pte_tbl[] into protection bits (those protection
88  * bits will select a cache mode of WP or better), and then translate the
89  * protection bits back into the cache mode using __pte2cm_idx() and the
90  * __pte2cachemode_tbl[] array. This will return the really used cache mode.
91  */
92 bool x86_has_pat_wp(void)
93 {
94 	uint16_t prot = __cachemode2pte_tbl[_PAGE_CACHE_MODE_WP];
95 
96 	return __pte2cachemode_tbl[__pte2cm_idx(prot)] == _PAGE_CACHE_MODE_WP;
97 }
98 
99 enum page_cache_mode pgprot2cachemode(pgprot_t pgprot)
100 {
101 	unsigned long masked;
102 
103 	masked = pgprot_val(pgprot) & _PAGE_CACHE_MASK;
104 	if (likely(masked == 0))
105 		return 0;
106 	return __pte2cachemode_tbl[__pte2cm_idx(masked)];
107 }
108 
109 static unsigned long __initdata pgt_buf_start;
110 static unsigned long __initdata pgt_buf_end;
111 static unsigned long __initdata pgt_buf_top;
112 
113 static unsigned long min_pfn_mapped;
114 
115 static bool __initdata can_use_brk_pgt = true;
116 
117 /*
118  * Pages returned are already directly mapped.
119  *
120  * Changing that is likely to break Xen, see commit:
121  *
122  *    279b706 x86,xen: introduce x86_init.mapping.pagetable_reserve
123  *
124  * for detailed information.
125  */
126 __ref void *alloc_low_pages(unsigned int num)
127 {
128 	unsigned long pfn;
129 	int i;
130 
131 	if (after_bootmem) {
132 		unsigned int order;
133 
134 		order = get_order((unsigned long)num << PAGE_SHIFT);
135 		return (void *)__get_free_pages(GFP_ATOMIC | __GFP_ZERO, order);
136 	}
137 
138 	if ((pgt_buf_end + num) > pgt_buf_top || !can_use_brk_pgt) {
139 		unsigned long ret = 0;
140 
141 		if (min_pfn_mapped < max_pfn_mapped) {
142 			ret = memblock_phys_alloc_range(
143 					PAGE_SIZE * num, PAGE_SIZE,
144 					min_pfn_mapped << PAGE_SHIFT,
145 					max_pfn_mapped << PAGE_SHIFT);
146 		}
147 		if (!ret && can_use_brk_pgt)
148 			ret = __pa(extend_brk(PAGE_SIZE * num, PAGE_SIZE));
149 
150 		if (!ret)
151 			panic("alloc_low_pages: can not alloc memory");
152 
153 		pfn = ret >> PAGE_SHIFT;
154 	} else {
155 		pfn = pgt_buf_end;
156 		pgt_buf_end += num;
157 	}
158 
159 	for (i = 0; i < num; i++) {
160 		void *adr;
161 
162 		adr = __va((pfn + i) << PAGE_SHIFT);
163 		clear_page(adr);
164 	}
165 
166 	return __va(pfn << PAGE_SHIFT);
167 }
168 
169 /*
170  * By default need to be able to allocate page tables below PGD firstly for
171  * the 0-ISA_END_ADDRESS range and secondly for the initial PMD_SIZE mapping.
172  * With KASLR memory randomization, depending on the machine e820 memory and the
173  * PUD alignment, twice that many pages may be needed when KASLR memory
174  * randomization is enabled.
175  */
176 
177 #ifndef CONFIG_X86_5LEVEL
178 #define INIT_PGD_PAGE_TABLES    3
179 #else
180 #define INIT_PGD_PAGE_TABLES    4
181 #endif
182 
183 #ifndef CONFIG_RANDOMIZE_MEMORY
184 #define INIT_PGD_PAGE_COUNT      (2 * INIT_PGD_PAGE_TABLES)
185 #else
186 #define INIT_PGD_PAGE_COUNT      (4 * INIT_PGD_PAGE_TABLES)
187 #endif
188 
189 #define INIT_PGT_BUF_SIZE	(INIT_PGD_PAGE_COUNT * PAGE_SIZE)
190 RESERVE_BRK(early_pgt_alloc, INIT_PGT_BUF_SIZE);
191 void  __init early_alloc_pgt_buf(void)
192 {
193 	unsigned long tables = INIT_PGT_BUF_SIZE;
194 	phys_addr_t base;
195 
196 	base = __pa(extend_brk(tables, PAGE_SIZE));
197 
198 	pgt_buf_start = base >> PAGE_SHIFT;
199 	pgt_buf_end = pgt_buf_start;
200 	pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
201 }
202 
203 int after_bootmem;
204 
205 early_param_on_off("gbpages", "nogbpages", direct_gbpages, CONFIG_X86_DIRECT_GBPAGES);
206 
207 struct map_range {
208 	unsigned long start;
209 	unsigned long end;
210 	unsigned page_size_mask;
211 };
212 
213 static int page_size_mask;
214 
215 /*
216  * Save some of cr4 feature set we're using (e.g.  Pentium 4MB
217  * enable and PPro Global page enable), so that any CPU's that boot
218  * up after us can get the correct flags. Invoked on the boot CPU.
219  */
220 static inline void cr4_set_bits_and_update_boot(unsigned long mask)
221 {
222 	mmu_cr4_features |= mask;
223 	if (trampoline_cr4_features)
224 		*trampoline_cr4_features = mmu_cr4_features;
225 	cr4_set_bits(mask);
226 }
227 
228 static void __init probe_page_size_mask(void)
229 {
230 	/*
231 	 * For pagealloc debugging, identity mapping will use small pages.
232 	 * This will simplify cpa(), which otherwise needs to support splitting
233 	 * large pages into small in interrupt context, etc.
234 	 */
235 	if (boot_cpu_has(X86_FEATURE_PSE) && !debug_pagealloc_enabled())
236 		page_size_mask |= 1 << PG_LEVEL_2M;
237 	else
238 		direct_gbpages = 0;
239 
240 	/* Enable PSE if available */
241 	if (boot_cpu_has(X86_FEATURE_PSE))
242 		cr4_set_bits_and_update_boot(X86_CR4_PSE);
243 
244 	/* Enable PGE if available */
245 	__supported_pte_mask &= ~_PAGE_GLOBAL;
246 	if (boot_cpu_has(X86_FEATURE_PGE)) {
247 		cr4_set_bits_and_update_boot(X86_CR4_PGE);
248 		__supported_pte_mask |= _PAGE_GLOBAL;
249 	}
250 
251 	/* By the default is everything supported: */
252 	__default_kernel_pte_mask = __supported_pte_mask;
253 	/* Except when with PTI where the kernel is mostly non-Global: */
254 	if (cpu_feature_enabled(X86_FEATURE_PTI))
255 		__default_kernel_pte_mask &= ~_PAGE_GLOBAL;
256 
257 	/* Enable 1 GB linear kernel mappings if available: */
258 	if (direct_gbpages && boot_cpu_has(X86_FEATURE_GBPAGES)) {
259 		printk(KERN_INFO "Using GB pages for direct mapping\n");
260 		page_size_mask |= 1 << PG_LEVEL_1G;
261 	} else {
262 		direct_gbpages = 0;
263 	}
264 }
265 
266 /*
267  * INVLPG may not properly flush Global entries on
268  * these CPUs.  New microcode fixes the issue.
269  */
270 static const struct x86_cpu_id invlpg_miss_ids[] = {
271 	X86_MATCH_VFM(INTEL_ALDERLAKE,	    0x2e),
272 	X86_MATCH_VFM(INTEL_ALDERLAKE_L,    0x42c),
273 	X86_MATCH_VFM(INTEL_ATOM_GRACEMONT, 0x11),
274 	X86_MATCH_VFM(INTEL_RAPTORLAKE,	    0x118),
275 	X86_MATCH_VFM(INTEL_RAPTORLAKE_P,   0x4117),
276 	X86_MATCH_VFM(INTEL_RAPTORLAKE_S,   0x2e),
277 	{}
278 };
279 
280 static void setup_pcid(void)
281 {
282 	const struct x86_cpu_id *invlpg_miss_match;
283 
284 	if (!IS_ENABLED(CONFIG_X86_64))
285 		return;
286 
287 	if (!boot_cpu_has(X86_FEATURE_PCID))
288 		return;
289 
290 	invlpg_miss_match = x86_match_cpu(invlpg_miss_ids);
291 
292 	if (invlpg_miss_match &&
293 	    boot_cpu_data.microcode < invlpg_miss_match->driver_data) {
294 		pr_info("Incomplete global flushes, disabling PCID");
295 		setup_clear_cpu_cap(X86_FEATURE_PCID);
296 		return;
297 	}
298 
299 	if (boot_cpu_has(X86_FEATURE_PGE)) {
300 		/*
301 		 * This can't be cr4_set_bits_and_update_boot() -- the
302 		 * trampoline code can't handle CR4.PCIDE and it wouldn't
303 		 * do any good anyway.  Despite the name,
304 		 * cr4_set_bits_and_update_boot() doesn't actually cause
305 		 * the bits in question to remain set all the way through
306 		 * the secondary boot asm.
307 		 *
308 		 * Instead, we brute-force it and set CR4.PCIDE manually in
309 		 * start_secondary().
310 		 */
311 		cr4_set_bits(X86_CR4_PCIDE);
312 	} else {
313 		/*
314 		 * flush_tlb_all(), as currently implemented, won't work if
315 		 * PCID is on but PGE is not.  Since that combination
316 		 * doesn't exist on real hardware, there's no reason to try
317 		 * to fully support it, but it's polite to avoid corrupting
318 		 * data if we're on an improperly configured VM.
319 		 */
320 		setup_clear_cpu_cap(X86_FEATURE_PCID);
321 	}
322 }
323 
324 #ifdef CONFIG_X86_32
325 #define NR_RANGE_MR 3
326 #else /* CONFIG_X86_64 */
327 #define NR_RANGE_MR 5
328 #endif
329 
330 static int __meminit save_mr(struct map_range *mr, int nr_range,
331 			     unsigned long start_pfn, unsigned long end_pfn,
332 			     unsigned long page_size_mask)
333 {
334 	if (start_pfn < end_pfn) {
335 		if (nr_range >= NR_RANGE_MR)
336 			panic("run out of range for init_memory_mapping\n");
337 		mr[nr_range].start = start_pfn<<PAGE_SHIFT;
338 		mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
339 		mr[nr_range].page_size_mask = page_size_mask;
340 		nr_range++;
341 	}
342 
343 	return nr_range;
344 }
345 
346 /*
347  * adjust the page_size_mask for small range to go with
348  *	big page size instead small one if nearby are ram too.
349  */
350 static void __ref adjust_range_page_size_mask(struct map_range *mr,
351 							 int nr_range)
352 {
353 	int i;
354 
355 	for (i = 0; i < nr_range; i++) {
356 		if ((page_size_mask & (1<<PG_LEVEL_2M)) &&
357 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) {
358 			unsigned long start = round_down(mr[i].start, PMD_SIZE);
359 			unsigned long end = round_up(mr[i].end, PMD_SIZE);
360 
361 #ifdef CONFIG_X86_32
362 			if ((end >> PAGE_SHIFT) > max_low_pfn)
363 				continue;
364 #endif
365 
366 			if (memblock_is_region_memory(start, end - start))
367 				mr[i].page_size_mask |= 1<<PG_LEVEL_2M;
368 		}
369 		if ((page_size_mask & (1<<PG_LEVEL_1G)) &&
370 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) {
371 			unsigned long start = round_down(mr[i].start, PUD_SIZE);
372 			unsigned long end = round_up(mr[i].end, PUD_SIZE);
373 
374 			if (memblock_is_region_memory(start, end - start))
375 				mr[i].page_size_mask |= 1<<PG_LEVEL_1G;
376 		}
377 	}
378 }
379 
380 static const char *page_size_string(struct map_range *mr)
381 {
382 	static const char str_1g[] = "1G";
383 	static const char str_2m[] = "2M";
384 	static const char str_4m[] = "4M";
385 	static const char str_4k[] = "4k";
386 
387 	if (mr->page_size_mask & (1<<PG_LEVEL_1G))
388 		return str_1g;
389 	/*
390 	 * 32-bit without PAE has a 4M large page size.
391 	 * PG_LEVEL_2M is misnamed, but we can at least
392 	 * print out the right size in the string.
393 	 */
394 	if (IS_ENABLED(CONFIG_X86_32) &&
395 	    !IS_ENABLED(CONFIG_X86_PAE) &&
396 	    mr->page_size_mask & (1<<PG_LEVEL_2M))
397 		return str_4m;
398 
399 	if (mr->page_size_mask & (1<<PG_LEVEL_2M))
400 		return str_2m;
401 
402 	return str_4k;
403 }
404 
405 static int __meminit split_mem_range(struct map_range *mr, int nr_range,
406 				     unsigned long start,
407 				     unsigned long end)
408 {
409 	unsigned long start_pfn, end_pfn, limit_pfn;
410 	unsigned long pfn;
411 	int i;
412 
413 	limit_pfn = PFN_DOWN(end);
414 
415 	/* head if not big page alignment ? */
416 	pfn = start_pfn = PFN_DOWN(start);
417 #ifdef CONFIG_X86_32
418 	/*
419 	 * Don't use a large page for the first 2/4MB of memory
420 	 * because there are often fixed size MTRRs in there
421 	 * and overlapping MTRRs into large pages can cause
422 	 * slowdowns.
423 	 */
424 	if (pfn == 0)
425 		end_pfn = PFN_DOWN(PMD_SIZE);
426 	else
427 		end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
428 #else /* CONFIG_X86_64 */
429 	end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
430 #endif
431 	if (end_pfn > limit_pfn)
432 		end_pfn = limit_pfn;
433 	if (start_pfn < end_pfn) {
434 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
435 		pfn = end_pfn;
436 	}
437 
438 	/* big page (2M) range */
439 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
440 #ifdef CONFIG_X86_32
441 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
442 #else /* CONFIG_X86_64 */
443 	end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
444 	if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE)))
445 		end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
446 #endif
447 
448 	if (start_pfn < end_pfn) {
449 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
450 				page_size_mask & (1<<PG_LEVEL_2M));
451 		pfn = end_pfn;
452 	}
453 
454 #ifdef CONFIG_X86_64
455 	/* big page (1G) range */
456 	start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
457 	end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE));
458 	if (start_pfn < end_pfn) {
459 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
460 				page_size_mask &
461 				 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
462 		pfn = end_pfn;
463 	}
464 
465 	/* tail is not big page (1G) alignment */
466 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
467 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
468 	if (start_pfn < end_pfn) {
469 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
470 				page_size_mask & (1<<PG_LEVEL_2M));
471 		pfn = end_pfn;
472 	}
473 #endif
474 
475 	/* tail is not big page (2M) alignment */
476 	start_pfn = pfn;
477 	end_pfn = limit_pfn;
478 	nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
479 
480 	if (!after_bootmem)
481 		adjust_range_page_size_mask(mr, nr_range);
482 
483 	/* try to merge same page size and continuous */
484 	for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
485 		unsigned long old_start;
486 		if (mr[i].end != mr[i+1].start ||
487 		    mr[i].page_size_mask != mr[i+1].page_size_mask)
488 			continue;
489 		/* move it */
490 		old_start = mr[i].start;
491 		memmove(&mr[i], &mr[i+1],
492 			(nr_range - 1 - i) * sizeof(struct map_range));
493 		mr[i--].start = old_start;
494 		nr_range--;
495 	}
496 
497 	for (i = 0; i < nr_range; i++)
498 		pr_debug(" [mem %#010lx-%#010lx] page %s\n",
499 				mr[i].start, mr[i].end - 1,
500 				page_size_string(&mr[i]));
501 
502 	return nr_range;
503 }
504 
505 struct range pfn_mapped[E820_MAX_ENTRIES];
506 int nr_pfn_mapped;
507 
508 static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn)
509 {
510 	nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_MAX_ENTRIES,
511 					     nr_pfn_mapped, start_pfn, end_pfn);
512 	nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_MAX_ENTRIES);
513 
514 	max_pfn_mapped = max(max_pfn_mapped, end_pfn);
515 
516 	if (start_pfn < (1UL<<(32-PAGE_SHIFT)))
517 		max_low_pfn_mapped = max(max_low_pfn_mapped,
518 					 min(end_pfn, 1UL<<(32-PAGE_SHIFT)));
519 }
520 
521 bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn)
522 {
523 	int i;
524 
525 	for (i = 0; i < nr_pfn_mapped; i++)
526 		if ((start_pfn >= pfn_mapped[i].start) &&
527 		    (end_pfn <= pfn_mapped[i].end))
528 			return true;
529 
530 	return false;
531 }
532 
533 /*
534  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
535  * This runs before bootmem is initialized and gets pages directly from
536  * the physical memory. To access them they are temporarily mapped.
537  */
538 unsigned long __ref init_memory_mapping(unsigned long start,
539 					unsigned long end, pgprot_t prot)
540 {
541 	struct map_range mr[NR_RANGE_MR];
542 	unsigned long ret = 0;
543 	int nr_range, i;
544 
545 	pr_debug("init_memory_mapping: [mem %#010lx-%#010lx]\n",
546 	       start, end - 1);
547 
548 	memset(mr, 0, sizeof(mr));
549 	nr_range = split_mem_range(mr, 0, start, end);
550 
551 	for (i = 0; i < nr_range; i++)
552 		ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
553 						   mr[i].page_size_mask,
554 						   prot);
555 
556 	add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT);
557 
558 	return ret >> PAGE_SHIFT;
559 }
560 
561 /*
562  * We need to iterate through the E820 memory map and create direct mappings
563  * for only E820_TYPE_RAM and E820_KERN_RESERVED regions. We cannot simply
564  * create direct mappings for all pfns from [0 to max_low_pfn) and
565  * [4GB to max_pfn) because of possible memory holes in high addresses
566  * that cannot be marked as UC by fixed/variable range MTRRs.
567  * Depending on the alignment of E820 ranges, this may possibly result
568  * in using smaller size (i.e. 4K instead of 2M or 1G) page tables.
569  *
570  * init_mem_mapping() calls init_range_memory_mapping() with big range.
571  * That range would have hole in the middle or ends, and only ram parts
572  * will be mapped in init_range_memory_mapping().
573  */
574 static unsigned long __init init_range_memory_mapping(
575 					   unsigned long r_start,
576 					   unsigned long r_end)
577 {
578 	unsigned long start_pfn, end_pfn;
579 	unsigned long mapped_ram_size = 0;
580 	int i;
581 
582 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
583 		u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end);
584 		u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end);
585 		if (start >= end)
586 			continue;
587 
588 		/*
589 		 * if it is overlapping with brk pgt, we need to
590 		 * alloc pgt buf from memblock instead.
591 		 */
592 		can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >=
593 				    min(end, (u64)pgt_buf_top<<PAGE_SHIFT);
594 		init_memory_mapping(start, end, PAGE_KERNEL);
595 		mapped_ram_size += end - start;
596 		can_use_brk_pgt = true;
597 	}
598 
599 	return mapped_ram_size;
600 }
601 
602 static unsigned long __init get_new_step_size(unsigned long step_size)
603 {
604 	/*
605 	 * Initial mapped size is PMD_SIZE (2M).
606 	 * We can not set step_size to be PUD_SIZE (1G) yet.
607 	 * In worse case, when we cross the 1G boundary, and
608 	 * PG_LEVEL_2M is not set, we will need 1+1+512 pages (2M + 8k)
609 	 * to map 1G range with PTE. Hence we use one less than the
610 	 * difference of page table level shifts.
611 	 *
612 	 * Don't need to worry about overflow in the top-down case, on 32bit,
613 	 * when step_size is 0, round_down() returns 0 for start, and that
614 	 * turns it into 0x100000000ULL.
615 	 * In the bottom-up case, round_up(x, 0) returns 0 though too, which
616 	 * needs to be taken into consideration by the code below.
617 	 */
618 	return step_size << (PMD_SHIFT - PAGE_SHIFT - 1);
619 }
620 
621 /**
622  * memory_map_top_down - Map [map_start, map_end) top down
623  * @map_start: start address of the target memory range
624  * @map_end: end address of the target memory range
625  *
626  * This function will setup direct mapping for memory range
627  * [map_start, map_end) in top-down. That said, the page tables
628  * will be allocated at the end of the memory, and we map the
629  * memory in top-down.
630  */
631 static void __init memory_map_top_down(unsigned long map_start,
632 				       unsigned long map_end)
633 {
634 	unsigned long real_end, last_start;
635 	unsigned long step_size;
636 	unsigned long addr;
637 	unsigned long mapped_ram_size = 0;
638 
639 	/*
640 	 * Systems that have many reserved areas near top of the memory,
641 	 * e.g. QEMU with less than 1G RAM and EFI enabled, or Xen, will
642 	 * require lots of 4K mappings which may exhaust pgt_buf.
643 	 * Start with top-most PMD_SIZE range aligned at PMD_SIZE to ensure
644 	 * there is enough mapped memory that can be allocated from
645 	 * memblock.
646 	 */
647 	addr = memblock_phys_alloc_range(PMD_SIZE, PMD_SIZE, map_start,
648 					 map_end);
649 	if (!addr) {
650 		pr_warn("Failed to release memory for alloc_low_pages()");
651 		real_end = max(map_start, ALIGN_DOWN(map_end, PMD_SIZE));
652 	} else {
653 		memblock_phys_free(addr, PMD_SIZE);
654 		real_end = addr + PMD_SIZE;
655 	}
656 
657 	/* step_size need to be small so pgt_buf from BRK could cover it */
658 	step_size = PMD_SIZE;
659 	max_pfn_mapped = 0; /* will get exact value next */
660 	min_pfn_mapped = real_end >> PAGE_SHIFT;
661 	last_start = real_end;
662 
663 	/*
664 	 * We start from the top (end of memory) and go to the bottom.
665 	 * The memblock_find_in_range() gets us a block of RAM from the
666 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
667 	 * for page table.
668 	 */
669 	while (last_start > map_start) {
670 		unsigned long start;
671 
672 		if (last_start > step_size) {
673 			start = round_down(last_start - 1, step_size);
674 			if (start < map_start)
675 				start = map_start;
676 		} else
677 			start = map_start;
678 		mapped_ram_size += init_range_memory_mapping(start,
679 							last_start);
680 		last_start = start;
681 		min_pfn_mapped = last_start >> PAGE_SHIFT;
682 		if (mapped_ram_size >= step_size)
683 			step_size = get_new_step_size(step_size);
684 	}
685 
686 	if (real_end < map_end)
687 		init_range_memory_mapping(real_end, map_end);
688 }
689 
690 /**
691  * memory_map_bottom_up - Map [map_start, map_end) bottom up
692  * @map_start: start address of the target memory range
693  * @map_end: end address of the target memory range
694  *
695  * This function will setup direct mapping for memory range
696  * [map_start, map_end) in bottom-up. Since we have limited the
697  * bottom-up allocation above the kernel, the page tables will
698  * be allocated just above the kernel and we map the memory
699  * in [map_start, map_end) in bottom-up.
700  */
701 static void __init memory_map_bottom_up(unsigned long map_start,
702 					unsigned long map_end)
703 {
704 	unsigned long next, start;
705 	unsigned long mapped_ram_size = 0;
706 	/* step_size need to be small so pgt_buf from BRK could cover it */
707 	unsigned long step_size = PMD_SIZE;
708 
709 	start = map_start;
710 	min_pfn_mapped = start >> PAGE_SHIFT;
711 
712 	/*
713 	 * We start from the bottom (@map_start) and go to the top (@map_end).
714 	 * The memblock_find_in_range() gets us a block of RAM from the
715 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
716 	 * for page table.
717 	 */
718 	while (start < map_end) {
719 		if (step_size && map_end - start > step_size) {
720 			next = round_up(start + 1, step_size);
721 			if (next > map_end)
722 				next = map_end;
723 		} else {
724 			next = map_end;
725 		}
726 
727 		mapped_ram_size += init_range_memory_mapping(start, next);
728 		start = next;
729 
730 		if (mapped_ram_size >= step_size)
731 			step_size = get_new_step_size(step_size);
732 	}
733 }
734 
735 /*
736  * The real mode trampoline, which is required for bootstrapping CPUs
737  * occupies only a small area under the low 1MB.  See reserve_real_mode()
738  * for details.
739  *
740  * If KASLR is disabled the first PGD entry of the direct mapping is copied
741  * to map the real mode trampoline.
742  *
743  * If KASLR is enabled, copy only the PUD which covers the low 1MB
744  * area. This limits the randomization granularity to 1GB for both 4-level
745  * and 5-level paging.
746  */
747 static void __init init_trampoline(void)
748 {
749 #ifdef CONFIG_X86_64
750 	/*
751 	 * The code below will alias kernel page-tables in the user-range of the
752 	 * address space, including the Global bit. So global TLB entries will
753 	 * be created when using the trampoline page-table.
754 	 */
755 	if (!kaslr_memory_enabled())
756 		trampoline_pgd_entry = init_top_pgt[pgd_index(__PAGE_OFFSET)];
757 	else
758 		init_trampoline_kaslr();
759 #endif
760 }
761 
762 void __init init_mem_mapping(void)
763 {
764 	unsigned long end;
765 
766 	pti_check_boottime_disable();
767 	probe_page_size_mask();
768 	setup_pcid();
769 
770 #ifdef CONFIG_X86_64
771 	end = max_pfn << PAGE_SHIFT;
772 #else
773 	end = max_low_pfn << PAGE_SHIFT;
774 #endif
775 
776 	/* the ISA range is always mapped regardless of memory holes */
777 	init_memory_mapping(0, ISA_END_ADDRESS, PAGE_KERNEL);
778 
779 	/* Init the trampoline, possibly with KASLR memory offset */
780 	init_trampoline();
781 
782 	/*
783 	 * If the allocation is in bottom-up direction, we setup direct mapping
784 	 * in bottom-up, otherwise we setup direct mapping in top-down.
785 	 */
786 	if (memblock_bottom_up()) {
787 		unsigned long kernel_end = __pa_symbol(_end);
788 
789 		/*
790 		 * we need two separate calls here. This is because we want to
791 		 * allocate page tables above the kernel. So we first map
792 		 * [kernel_end, end) to make memory above the kernel be mapped
793 		 * as soon as possible. And then use page tables allocated above
794 		 * the kernel to map [ISA_END_ADDRESS, kernel_end).
795 		 */
796 		memory_map_bottom_up(kernel_end, end);
797 		memory_map_bottom_up(ISA_END_ADDRESS, kernel_end);
798 	} else {
799 		memory_map_top_down(ISA_END_ADDRESS, end);
800 	}
801 
802 #ifdef CONFIG_X86_64
803 	if (max_pfn > max_low_pfn) {
804 		/* can we preserve max_low_pfn ?*/
805 		max_low_pfn = max_pfn;
806 	}
807 #else
808 	early_ioremap_page_table_range_init();
809 #endif
810 
811 	load_cr3(swapper_pg_dir);
812 	__flush_tlb_all();
813 
814 	x86_init.hyper.init_mem_mapping();
815 
816 	early_memtest(0, max_pfn_mapped << PAGE_SHIFT);
817 }
818 
819 /*
820  * Initialize an mm_struct to be used during poking and a pointer to be used
821  * during patching.
822  */
823 void __init poking_init(void)
824 {
825 	spinlock_t *ptl;
826 	pte_t *ptep;
827 
828 	text_poke_mm = mm_alloc();
829 	BUG_ON(!text_poke_mm);
830 
831 	/* Xen PV guests need the PGD to be pinned. */
832 	paravirt_enter_mmap(text_poke_mm);
833 
834 	set_notrack_mm(text_poke_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 	text_poke_mm_addr = TASK_UNMAPPED_BASE;
842 	if (IS_ENABLED(CONFIG_RANDOMIZE_BASE))
843 		text_poke_mm_addr += (kaslr_get_random_long("Poking") & PAGE_MASK) %
844 			(TASK_SIZE - TASK_UNMAPPED_BASE - 3 * PAGE_SIZE);
845 
846 	if (((text_poke_mm_addr + PAGE_SIZE) & ~PMD_MASK) == 0)
847 		text_poke_mm_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(text_poke_mm, text_poke_mm_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 void __init zone_sizes_init(void)
1004 {
1005 	unsigned long max_zone_pfns[MAX_NR_ZONES];
1006 
1007 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1008 
1009 #ifdef CONFIG_ZONE_DMA
1010 	max_zone_pfns[ZONE_DMA]		= min(MAX_DMA_PFN, max_low_pfn);
1011 #endif
1012 #ifdef CONFIG_ZONE_DMA32
1013 	max_zone_pfns[ZONE_DMA32]	= min(MAX_DMA32_PFN, max_low_pfn);
1014 #endif
1015 	max_zone_pfns[ZONE_NORMAL]	= max_low_pfn;
1016 #ifdef CONFIG_HIGHMEM
1017 	max_zone_pfns[ZONE_HIGHMEM]	= max_pfn;
1018 #endif
1019 
1020 	free_area_init(max_zone_pfns);
1021 }
1022 
1023 __visible DEFINE_PER_CPU_ALIGNED(struct tlb_state, cpu_tlbstate) = {
1024 	.loaded_mm = &init_mm,
1025 	.next_asid = 1,
1026 	.cr4 = ~0UL,	/* fail hard if we screw up cr4 shadow initialization */
1027 };
1028 
1029 #ifdef CONFIG_ADDRESS_MASKING
1030 DEFINE_PER_CPU(u64, tlbstate_untag_mask);
1031 EXPORT_PER_CPU_SYMBOL(tlbstate_untag_mask);
1032 #endif
1033 
1034 void update_cache_mode_entry(unsigned entry, enum page_cache_mode cache)
1035 {
1036 	/* entry 0 MUST be WB (hardwired to speed up translations) */
1037 	BUG_ON(!entry && cache != _PAGE_CACHE_MODE_WB);
1038 
1039 	__cachemode2pte_tbl[cache] = __cm_idx2pte(entry);
1040 	__pte2cachemode_tbl[entry] = cache;
1041 }
1042 
1043 #ifdef CONFIG_SWAP
1044 unsigned long arch_max_swapfile_size(void)
1045 {
1046 	unsigned long pages;
1047 
1048 	pages = generic_max_swapfile_size();
1049 
1050 	if (boot_cpu_has_bug(X86_BUG_L1TF) && l1tf_mitigation != L1TF_MITIGATION_OFF) {
1051 		/* Limit the swap file size to MAX_PA/2 for L1TF workaround */
1052 		unsigned long long l1tf_limit = l1tf_pfn_limit();
1053 		/*
1054 		 * We encode swap offsets also with 3 bits below those for pfn
1055 		 * which makes the usable limit higher.
1056 		 */
1057 #if CONFIG_PGTABLE_LEVELS > 2
1058 		l1tf_limit <<= PAGE_SHIFT - SWP_OFFSET_FIRST_BIT;
1059 #endif
1060 		pages = min_t(unsigned long long, l1tf_limit, pages);
1061 	}
1062 	return pages;
1063 }
1064 #endif
1065 
1066 #ifdef CONFIG_EXECMEM
1067 static struct execmem_info execmem_info __ro_after_init;
1068 
1069 #ifdef CONFIG_ARCH_HAS_EXECMEM_ROX
1070 void execmem_fill_trapping_insns(void *ptr, size_t size, bool writeable)
1071 {
1072 	/* fill memory with INT3 instructions */
1073 	if (writeable)
1074 		memset(ptr, INT3_INSN_OPCODE, size);
1075 	else
1076 		text_poke_set(ptr, INT3_INSN_OPCODE, size);
1077 }
1078 #endif
1079 
1080 struct execmem_info __init *execmem_arch_setup(void)
1081 {
1082 	unsigned long start, offset = 0;
1083 	enum execmem_range_flags flags;
1084 	pgprot_t pgprot;
1085 
1086 	if (kaslr_enabled())
1087 		offset = get_random_u32_inclusive(1, 1024) * PAGE_SIZE;
1088 
1089 	start = MODULES_VADDR + offset;
1090 
1091 	if (IS_ENABLED(CONFIG_ARCH_HAS_EXECMEM_ROX) &&
1092 	    cpu_feature_enabled(X86_FEATURE_PSE)) {
1093 		pgprot = PAGE_KERNEL_ROX;
1094 		flags = EXECMEM_KASAN_SHADOW | EXECMEM_ROX_CACHE;
1095 	} else {
1096 		pgprot = PAGE_KERNEL;
1097 		flags = EXECMEM_KASAN_SHADOW;
1098 	}
1099 
1100 	execmem_info = (struct execmem_info){
1101 		.ranges = {
1102 			[EXECMEM_MODULE_TEXT] = {
1103 				.flags	= flags,
1104 				.start	= start,
1105 				.end	= MODULES_END,
1106 				.pgprot	= pgprot,
1107 				.alignment = MODULE_ALIGN,
1108 			},
1109 			[EXECMEM_KPROBES ... EXECMEM_BPF] = {
1110 				.flags	= EXECMEM_KASAN_SHADOW,
1111 				.start	= start,
1112 				.end	= MODULES_END,
1113 				.pgprot	= PAGE_KERNEL,
1114 				.alignment = MODULE_ALIGN,
1115 			},
1116 			[EXECMEM_MODULE_DATA] = {
1117 				.flags	= EXECMEM_KASAN_SHADOW,
1118 				.start	= start,
1119 				.end	= MODULES_END,
1120 				.pgprot	= PAGE_KERNEL,
1121 				.alignment = MODULE_ALIGN,
1122 			},
1123 		},
1124 	};
1125 
1126 	return &execmem_info;
1127 }
1128 #endif /* CONFIG_EXECMEM */
1129