xref: /linux/arch/x86/mm/init.c (revision e78f70bad29c5ae1e1076698b690b15794e9b81e)
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 #define INIT_PGD_PAGE_TABLES    4
178 
179 #ifndef CONFIG_RANDOMIZE_MEMORY
180 #define INIT_PGD_PAGE_COUNT      (2 * INIT_PGD_PAGE_TABLES)
181 #else
182 #define INIT_PGD_PAGE_COUNT      (4 * INIT_PGD_PAGE_TABLES)
183 #endif
184 
185 #define INIT_PGT_BUF_SIZE	(INIT_PGD_PAGE_COUNT * PAGE_SIZE)
186 RESERVE_BRK(early_pgt_alloc, INIT_PGT_BUF_SIZE);
187 void  __init early_alloc_pgt_buf(void)
188 {
189 	unsigned long tables = INIT_PGT_BUF_SIZE;
190 	phys_addr_t base;
191 
192 	base = __pa(extend_brk(tables, PAGE_SIZE));
193 
194 	pgt_buf_start = base >> PAGE_SHIFT;
195 	pgt_buf_end = pgt_buf_start;
196 	pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
197 }
198 
199 int after_bootmem;
200 
201 early_param_on_off("gbpages", "nogbpages", direct_gbpages, CONFIG_X86_DIRECT_GBPAGES);
202 
203 struct map_range {
204 	unsigned long start;
205 	unsigned long end;
206 	unsigned page_size_mask;
207 };
208 
209 static int page_size_mask;
210 
211 /*
212  * Save some of cr4 feature set we're using (e.g.  Pentium 4MB
213  * enable and PPro Global page enable), so that any CPU's that boot
214  * up after us can get the correct flags. Invoked on the boot CPU.
215  */
216 static inline void cr4_set_bits_and_update_boot(unsigned long mask)
217 {
218 	mmu_cr4_features |= mask;
219 	if (trampoline_cr4_features)
220 		*trampoline_cr4_features = mmu_cr4_features;
221 	cr4_set_bits(mask);
222 }
223 
224 static void __init probe_page_size_mask(void)
225 {
226 	/*
227 	 * For pagealloc debugging, identity mapping will use small pages.
228 	 * This will simplify cpa(), which otherwise needs to support splitting
229 	 * large pages into small in interrupt context, etc.
230 	 */
231 	if (boot_cpu_has(X86_FEATURE_PSE) && !debug_pagealloc_enabled())
232 		page_size_mask |= 1 << PG_LEVEL_2M;
233 	else
234 		direct_gbpages = 0;
235 
236 	/* Enable PSE if available */
237 	if (boot_cpu_has(X86_FEATURE_PSE))
238 		cr4_set_bits_and_update_boot(X86_CR4_PSE);
239 
240 	/* Enable PGE if available */
241 	__supported_pte_mask &= ~_PAGE_GLOBAL;
242 	if (boot_cpu_has(X86_FEATURE_PGE)) {
243 		cr4_set_bits_and_update_boot(X86_CR4_PGE);
244 		__supported_pte_mask |= _PAGE_GLOBAL;
245 	}
246 
247 	/* By the default is everything supported: */
248 	__default_kernel_pte_mask = __supported_pte_mask;
249 	/* Except when with PTI where the kernel is mostly non-Global: */
250 	if (cpu_feature_enabled(X86_FEATURE_PTI))
251 		__default_kernel_pte_mask &= ~_PAGE_GLOBAL;
252 
253 	/* Enable 1 GB linear kernel mappings if available: */
254 	if (direct_gbpages && boot_cpu_has(X86_FEATURE_GBPAGES)) {
255 		printk(KERN_INFO "Using GB pages for direct mapping\n");
256 		page_size_mask |= 1 << PG_LEVEL_1G;
257 	} else {
258 		direct_gbpages = 0;
259 	}
260 }
261 
262 /*
263  * INVLPG may not properly flush Global entries on
264  * these CPUs.  New microcode fixes the issue.
265  */
266 static const struct x86_cpu_id invlpg_miss_ids[] = {
267 	X86_MATCH_VFM(INTEL_ALDERLAKE,	    0x2e),
268 	X86_MATCH_VFM(INTEL_ALDERLAKE_L,    0x42c),
269 	X86_MATCH_VFM(INTEL_ATOM_GRACEMONT, 0x11),
270 	X86_MATCH_VFM(INTEL_RAPTORLAKE,	    0x118),
271 	X86_MATCH_VFM(INTEL_RAPTORLAKE_P,   0x4117),
272 	X86_MATCH_VFM(INTEL_RAPTORLAKE_S,   0x2e),
273 	{}
274 };
275 
276 static void setup_pcid(void)
277 {
278 	const struct x86_cpu_id *invlpg_miss_match;
279 
280 	if (!IS_ENABLED(CONFIG_X86_64))
281 		return;
282 
283 	if (!boot_cpu_has(X86_FEATURE_PCID))
284 		return;
285 
286 	invlpg_miss_match = x86_match_cpu(invlpg_miss_ids);
287 
288 	if (invlpg_miss_match &&
289 	    boot_cpu_data.microcode < invlpg_miss_match->driver_data) {
290 		pr_info("Incomplete global flushes, disabling PCID");
291 		setup_clear_cpu_cap(X86_FEATURE_PCID);
292 		return;
293 	}
294 
295 	if (boot_cpu_has(X86_FEATURE_PGE)) {
296 		/*
297 		 * This can't be cr4_set_bits_and_update_boot() -- the
298 		 * trampoline code can't handle CR4.PCIDE and it wouldn't
299 		 * do any good anyway.  Despite the name,
300 		 * cr4_set_bits_and_update_boot() doesn't actually cause
301 		 * the bits in question to remain set all the way through
302 		 * the secondary boot asm.
303 		 *
304 		 * Instead, we brute-force it and set CR4.PCIDE manually in
305 		 * start_secondary().
306 		 */
307 		cr4_set_bits(X86_CR4_PCIDE);
308 	} else {
309 		/*
310 		 * flush_tlb_all(), as currently implemented, won't work if
311 		 * PCID is on but PGE is not.  Since that combination
312 		 * doesn't exist on real hardware, there's no reason to try
313 		 * to fully support it, but it's polite to avoid corrupting
314 		 * data if we're on an improperly configured VM.
315 		 */
316 		setup_clear_cpu_cap(X86_FEATURE_PCID);
317 	}
318 }
319 
320 #ifdef CONFIG_X86_32
321 #define NR_RANGE_MR 3
322 #else /* CONFIG_X86_64 */
323 #define NR_RANGE_MR 5
324 #endif
325 
326 static int __meminit save_mr(struct map_range *mr, int nr_range,
327 			     unsigned long start_pfn, unsigned long end_pfn,
328 			     unsigned long page_size_mask)
329 {
330 	if (start_pfn < end_pfn) {
331 		if (nr_range >= NR_RANGE_MR)
332 			panic("run out of range for init_memory_mapping\n");
333 		mr[nr_range].start = start_pfn<<PAGE_SHIFT;
334 		mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
335 		mr[nr_range].page_size_mask = page_size_mask;
336 		nr_range++;
337 	}
338 
339 	return nr_range;
340 }
341 
342 /*
343  * adjust the page_size_mask for small range to go with
344  *	big page size instead small one if nearby are ram too.
345  */
346 static void __ref adjust_range_page_size_mask(struct map_range *mr,
347 							 int nr_range)
348 {
349 	int i;
350 
351 	for (i = 0; i < nr_range; i++) {
352 		if ((page_size_mask & (1<<PG_LEVEL_2M)) &&
353 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) {
354 			unsigned long start = round_down(mr[i].start, PMD_SIZE);
355 			unsigned long end = round_up(mr[i].end, PMD_SIZE);
356 
357 #ifdef CONFIG_X86_32
358 			if ((end >> PAGE_SHIFT) > max_low_pfn)
359 				continue;
360 #endif
361 
362 			if (memblock_is_region_memory(start, end - start))
363 				mr[i].page_size_mask |= 1<<PG_LEVEL_2M;
364 		}
365 		if ((page_size_mask & (1<<PG_LEVEL_1G)) &&
366 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) {
367 			unsigned long start = round_down(mr[i].start, PUD_SIZE);
368 			unsigned long end = round_up(mr[i].end, PUD_SIZE);
369 
370 			if (memblock_is_region_memory(start, end - start))
371 				mr[i].page_size_mask |= 1<<PG_LEVEL_1G;
372 		}
373 	}
374 }
375 
376 static const char *page_size_string(struct map_range *mr)
377 {
378 	static const char str_1g[] = "1G";
379 	static const char str_2m[] = "2M";
380 	static const char str_4m[] = "4M";
381 	static const char str_4k[] = "4k";
382 
383 	if (mr->page_size_mask & (1<<PG_LEVEL_1G))
384 		return str_1g;
385 	/*
386 	 * 32-bit without PAE has a 4M large page size.
387 	 * PG_LEVEL_2M is misnamed, but we can at least
388 	 * print out the right size in the string.
389 	 */
390 	if (IS_ENABLED(CONFIG_X86_32) &&
391 	    !IS_ENABLED(CONFIG_X86_PAE) &&
392 	    mr->page_size_mask & (1<<PG_LEVEL_2M))
393 		return str_4m;
394 
395 	if (mr->page_size_mask & (1<<PG_LEVEL_2M))
396 		return str_2m;
397 
398 	return str_4k;
399 }
400 
401 static int __meminit split_mem_range(struct map_range *mr, int nr_range,
402 				     unsigned long start,
403 				     unsigned long end)
404 {
405 	unsigned long start_pfn, end_pfn, limit_pfn;
406 	unsigned long pfn;
407 	int i;
408 
409 	limit_pfn = PFN_DOWN(end);
410 
411 	/* head if not big page alignment ? */
412 	pfn = start_pfn = PFN_DOWN(start);
413 #ifdef CONFIG_X86_32
414 	/*
415 	 * Don't use a large page for the first 2/4MB of memory
416 	 * because there are often fixed size MTRRs in there
417 	 * and overlapping MTRRs into large pages can cause
418 	 * slowdowns.
419 	 */
420 	if (pfn == 0)
421 		end_pfn = PFN_DOWN(PMD_SIZE);
422 	else
423 		end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
424 #else /* CONFIG_X86_64 */
425 	end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
426 #endif
427 	if (end_pfn > limit_pfn)
428 		end_pfn = limit_pfn;
429 	if (start_pfn < end_pfn) {
430 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
431 		pfn = end_pfn;
432 	}
433 
434 	/* big page (2M) range */
435 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
436 #ifdef CONFIG_X86_32
437 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
438 #else /* CONFIG_X86_64 */
439 	end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
440 	if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE)))
441 		end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
442 #endif
443 
444 	if (start_pfn < end_pfn) {
445 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
446 				page_size_mask & (1<<PG_LEVEL_2M));
447 		pfn = end_pfn;
448 	}
449 
450 #ifdef CONFIG_X86_64
451 	/* big page (1G) range */
452 	start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
453 	end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE));
454 	if (start_pfn < end_pfn) {
455 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
456 				page_size_mask &
457 				 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
458 		pfn = end_pfn;
459 	}
460 
461 	/* tail is not big page (1G) alignment */
462 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
463 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
464 	if (start_pfn < end_pfn) {
465 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
466 				page_size_mask & (1<<PG_LEVEL_2M));
467 		pfn = end_pfn;
468 	}
469 #endif
470 
471 	/* tail is not big page (2M) alignment */
472 	start_pfn = pfn;
473 	end_pfn = limit_pfn;
474 	nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
475 
476 	if (!after_bootmem)
477 		adjust_range_page_size_mask(mr, nr_range);
478 
479 	/* try to merge same page size and continuous */
480 	for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
481 		unsigned long old_start;
482 		if (mr[i].end != mr[i+1].start ||
483 		    mr[i].page_size_mask != mr[i+1].page_size_mask)
484 			continue;
485 		/* move it */
486 		old_start = mr[i].start;
487 		memmove(&mr[i], &mr[i+1],
488 			(nr_range - 1 - i) * sizeof(struct map_range));
489 		mr[i--].start = old_start;
490 		nr_range--;
491 	}
492 
493 	for (i = 0; i < nr_range; i++)
494 		pr_debug(" [mem %#010lx-%#010lx] page %s\n",
495 				mr[i].start, mr[i].end - 1,
496 				page_size_string(&mr[i]));
497 
498 	return nr_range;
499 }
500 
501 struct range pfn_mapped[E820_MAX_ENTRIES];
502 int nr_pfn_mapped;
503 
504 static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn)
505 {
506 	nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_MAX_ENTRIES,
507 					     nr_pfn_mapped, start_pfn, end_pfn);
508 	nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_MAX_ENTRIES);
509 
510 	max_pfn_mapped = max(max_pfn_mapped, end_pfn);
511 
512 	if (start_pfn < (1UL<<(32-PAGE_SHIFT)))
513 		max_low_pfn_mapped = max(max_low_pfn_mapped,
514 					 min(end_pfn, 1UL<<(32-PAGE_SHIFT)));
515 }
516 
517 bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn)
518 {
519 	int i;
520 
521 	for (i = 0; i < nr_pfn_mapped; i++)
522 		if ((start_pfn >= pfn_mapped[i].start) &&
523 		    (end_pfn <= pfn_mapped[i].end))
524 			return true;
525 
526 	return false;
527 }
528 
529 /*
530  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
531  * This runs before bootmem is initialized and gets pages directly from
532  * the physical memory. To access them they are temporarily mapped.
533  */
534 unsigned long __ref init_memory_mapping(unsigned long start,
535 					unsigned long end, pgprot_t prot)
536 {
537 	struct map_range mr[NR_RANGE_MR];
538 	unsigned long ret = 0;
539 	int nr_range, i;
540 
541 	pr_debug("init_memory_mapping: [mem %#010lx-%#010lx]\n",
542 	       start, end - 1);
543 
544 	memset(mr, 0, sizeof(mr));
545 	nr_range = split_mem_range(mr, 0, start, end);
546 
547 	for (i = 0; i < nr_range; i++)
548 		ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
549 						   mr[i].page_size_mask,
550 						   prot);
551 
552 	add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT);
553 
554 	return ret >> PAGE_SHIFT;
555 }
556 
557 /*
558  * We need to iterate through the E820 memory map and create direct mappings
559  * for only E820_TYPE_RAM and E820_KERN_RESERVED regions. We cannot simply
560  * create direct mappings for all pfns from [0 to max_low_pfn) and
561  * [4GB to max_pfn) because of possible memory holes in high addresses
562  * that cannot be marked as UC by fixed/variable range MTRRs.
563  * Depending on the alignment of E820 ranges, this may possibly result
564  * in using smaller size (i.e. 4K instead of 2M or 1G) page tables.
565  *
566  * init_mem_mapping() calls init_range_memory_mapping() with big range.
567  * That range would have hole in the middle or ends, and only ram parts
568  * will be mapped in init_range_memory_mapping().
569  */
570 static unsigned long __init init_range_memory_mapping(
571 					   unsigned long r_start,
572 					   unsigned long r_end)
573 {
574 	unsigned long start_pfn, end_pfn;
575 	unsigned long mapped_ram_size = 0;
576 	int i;
577 
578 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
579 		u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end);
580 		u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end);
581 		if (start >= end)
582 			continue;
583 
584 		/*
585 		 * if it is overlapping with brk pgt, we need to
586 		 * alloc pgt buf from memblock instead.
587 		 */
588 		can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >=
589 				    min(end, (u64)pgt_buf_top<<PAGE_SHIFT);
590 		init_memory_mapping(start, end, PAGE_KERNEL);
591 		mapped_ram_size += end - start;
592 		can_use_brk_pgt = true;
593 	}
594 
595 	return mapped_ram_size;
596 }
597 
598 static unsigned long __init get_new_step_size(unsigned long step_size)
599 {
600 	/*
601 	 * Initial mapped size is PMD_SIZE (2M).
602 	 * We can not set step_size to be PUD_SIZE (1G) yet.
603 	 * In worse case, when we cross the 1G boundary, and
604 	 * PG_LEVEL_2M is not set, we will need 1+1+512 pages (2M + 8k)
605 	 * to map 1G range with PTE. Hence we use one less than the
606 	 * difference of page table level shifts.
607 	 *
608 	 * Don't need to worry about overflow in the top-down case, on 32bit,
609 	 * when step_size is 0, round_down() returns 0 for start, and that
610 	 * turns it into 0x100000000ULL.
611 	 * In the bottom-up case, round_up(x, 0) returns 0 though too, which
612 	 * needs to be taken into consideration by the code below.
613 	 */
614 	return step_size << (PMD_SHIFT - PAGE_SHIFT - 1);
615 }
616 
617 /**
618  * memory_map_top_down - Map [map_start, map_end) top down
619  * @map_start: start address of the target memory range
620  * @map_end: end address of the target memory range
621  *
622  * This function will setup direct mapping for memory range
623  * [map_start, map_end) in top-down. That said, the page tables
624  * will be allocated at the end of the memory, and we map the
625  * memory in top-down.
626  */
627 static void __init memory_map_top_down(unsigned long map_start,
628 				       unsigned long map_end)
629 {
630 	unsigned long real_end, last_start;
631 	unsigned long step_size;
632 	unsigned long addr;
633 	unsigned long mapped_ram_size = 0;
634 
635 	/*
636 	 * Systems that have many reserved areas near top of the memory,
637 	 * e.g. QEMU with less than 1G RAM and EFI enabled, or Xen, will
638 	 * require lots of 4K mappings which may exhaust pgt_buf.
639 	 * Start with top-most PMD_SIZE range aligned at PMD_SIZE to ensure
640 	 * there is enough mapped memory that can be allocated from
641 	 * memblock.
642 	 */
643 	addr = memblock_phys_alloc_range(PMD_SIZE, PMD_SIZE, map_start,
644 					 map_end);
645 	if (!addr) {
646 		pr_warn("Failed to release memory for alloc_low_pages()");
647 		real_end = max(map_start, ALIGN_DOWN(map_end, PMD_SIZE));
648 	} else {
649 		memblock_phys_free(addr, PMD_SIZE);
650 		real_end = addr + PMD_SIZE;
651 	}
652 
653 	/* step_size need to be small so pgt_buf from BRK could cover it */
654 	step_size = PMD_SIZE;
655 	max_pfn_mapped = 0; /* will get exact value next */
656 	min_pfn_mapped = real_end >> PAGE_SHIFT;
657 	last_start = real_end;
658 
659 	/*
660 	 * We start from the top (end of memory) and go to the bottom.
661 	 * The memblock_find_in_range() gets us a block of RAM from the
662 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
663 	 * for page table.
664 	 */
665 	while (last_start > map_start) {
666 		unsigned long start;
667 
668 		if (last_start > step_size) {
669 			start = round_down(last_start - 1, step_size);
670 			if (start < map_start)
671 				start = map_start;
672 		} else
673 			start = map_start;
674 		mapped_ram_size += init_range_memory_mapping(start,
675 							last_start);
676 		last_start = start;
677 		min_pfn_mapped = last_start >> PAGE_SHIFT;
678 		if (mapped_ram_size >= step_size)
679 			step_size = get_new_step_size(step_size);
680 	}
681 
682 	if (real_end < map_end)
683 		init_range_memory_mapping(real_end, map_end);
684 }
685 
686 /**
687  * memory_map_bottom_up - Map [map_start, map_end) bottom up
688  * @map_start: start address of the target memory range
689  * @map_end: end address of the target memory range
690  *
691  * This function will setup direct mapping for memory range
692  * [map_start, map_end) in bottom-up. Since we have limited the
693  * bottom-up allocation above the kernel, the page tables will
694  * be allocated just above the kernel and we map the memory
695  * in [map_start, map_end) in bottom-up.
696  */
697 static void __init memory_map_bottom_up(unsigned long map_start,
698 					unsigned long map_end)
699 {
700 	unsigned long next, start;
701 	unsigned long mapped_ram_size = 0;
702 	/* step_size need to be small so pgt_buf from BRK could cover it */
703 	unsigned long step_size = PMD_SIZE;
704 
705 	start = map_start;
706 	min_pfn_mapped = start >> PAGE_SHIFT;
707 
708 	/*
709 	 * We start from the bottom (@map_start) and go to the top (@map_end).
710 	 * The memblock_find_in_range() gets us a block of RAM from the
711 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
712 	 * for page table.
713 	 */
714 	while (start < map_end) {
715 		if (step_size && map_end - start > step_size) {
716 			next = round_up(start + 1, step_size);
717 			if (next > map_end)
718 				next = map_end;
719 		} else {
720 			next = map_end;
721 		}
722 
723 		mapped_ram_size += init_range_memory_mapping(start, next);
724 		start = next;
725 
726 		if (mapped_ram_size >= step_size)
727 			step_size = get_new_step_size(step_size);
728 	}
729 }
730 
731 /*
732  * The real mode trampoline, which is required for bootstrapping CPUs
733  * occupies only a small area under the low 1MB.  See reserve_real_mode()
734  * for details.
735  *
736  * If KASLR is disabled the first PGD entry of the direct mapping is copied
737  * to map the real mode trampoline.
738  *
739  * If KASLR is enabled, copy only the PUD which covers the low 1MB
740  * area. This limits the randomization granularity to 1GB for both 4-level
741  * and 5-level paging.
742  */
743 static void __init init_trampoline(void)
744 {
745 #ifdef CONFIG_X86_64
746 	/*
747 	 * The code below will alias kernel page-tables in the user-range of the
748 	 * address space, including the Global bit. So global TLB entries will
749 	 * be created when using the trampoline page-table.
750 	 */
751 	if (!kaslr_memory_enabled())
752 		trampoline_pgd_entry = init_top_pgt[pgd_index(__PAGE_OFFSET)];
753 	else
754 		init_trampoline_kaslr();
755 #endif
756 }
757 
758 void __init init_mem_mapping(void)
759 {
760 	unsigned long end;
761 
762 	pti_check_boottime_disable();
763 	probe_page_size_mask();
764 	setup_pcid();
765 
766 #ifdef CONFIG_X86_64
767 	end = max_pfn << PAGE_SHIFT;
768 #else
769 	end = max_low_pfn << PAGE_SHIFT;
770 #endif
771 
772 	/* the ISA range is always mapped regardless of memory holes */
773 	init_memory_mapping(0, ISA_END_ADDRESS, PAGE_KERNEL);
774 
775 	/* Init the trampoline, possibly with KASLR memory offset */
776 	init_trampoline();
777 
778 	/*
779 	 * If the allocation is in bottom-up direction, we setup direct mapping
780 	 * in bottom-up, otherwise we setup direct mapping in top-down.
781 	 */
782 	if (memblock_bottom_up()) {
783 		unsigned long kernel_end = __pa_symbol(_end);
784 
785 		/*
786 		 * we need two separate calls here. This is because we want to
787 		 * allocate page tables above the kernel. So we first map
788 		 * [kernel_end, end) to make memory above the kernel be mapped
789 		 * as soon as possible. And then use page tables allocated above
790 		 * the kernel to map [ISA_END_ADDRESS, kernel_end).
791 		 */
792 		memory_map_bottom_up(kernel_end, end);
793 		memory_map_bottom_up(ISA_END_ADDRESS, kernel_end);
794 	} else {
795 		memory_map_top_down(ISA_END_ADDRESS, end);
796 	}
797 
798 #ifdef CONFIG_X86_64
799 	if (max_pfn > max_low_pfn) {
800 		/* can we preserve max_low_pfn ?*/
801 		max_low_pfn = max_pfn;
802 	}
803 #else
804 	early_ioremap_page_table_range_init();
805 #endif
806 
807 	load_cr3(swapper_pg_dir);
808 	__flush_tlb_all();
809 
810 	x86_init.hyper.init_mem_mapping();
811 
812 	early_memtest(0, max_pfn_mapped << PAGE_SHIFT);
813 }
814 
815 /*
816  * Initialize an mm_struct to be used during poking and a pointer to be used
817  * during patching.
818  */
819 void __init poking_init(void)
820 {
821 	spinlock_t *ptl;
822 	pte_t *ptep;
823 
824 	text_poke_mm = mm_alloc();
825 	BUG_ON(!text_poke_mm);
826 
827 	/* Xen PV guests need the PGD to be pinned. */
828 	paravirt_enter_mmap(text_poke_mm);
829 
830 	set_notrack_mm(text_poke_mm);
831 
832 	/*
833 	 * Randomize the poking address, but make sure that the following page
834 	 * will be mapped at the same PMD. We need 2 pages, so find space for 3,
835 	 * and adjust the address if the PMD ends after the first one.
836 	 */
837 	text_poke_mm_addr = TASK_UNMAPPED_BASE;
838 	if (IS_ENABLED(CONFIG_RANDOMIZE_BASE))
839 		text_poke_mm_addr += (kaslr_get_random_long("Poking") & PAGE_MASK) %
840 			(TASK_SIZE - TASK_UNMAPPED_BASE - 3 * PAGE_SIZE);
841 
842 	if (((text_poke_mm_addr + PAGE_SIZE) & ~PMD_MASK) == 0)
843 		text_poke_mm_addr += PAGE_SIZE;
844 
845 	/*
846 	 * We need to trigger the allocation of the page-tables that will be
847 	 * needed for poking now. Later, poking may be performed in an atomic
848 	 * section, which might cause allocation to fail.
849 	 */
850 	ptep = get_locked_pte(text_poke_mm, text_poke_mm_addr, &ptl);
851 	BUG_ON(!ptep);
852 	pte_unmap_unlock(ptep, ptl);
853 }
854 
855 /*
856  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
857  * is valid. The argument is a physical page number.
858  *
859  * On x86, access has to be given to the first megabyte of RAM because that
860  * area traditionally contains BIOS code and data regions used by X, dosemu,
861  * and similar apps. Since they map the entire memory range, the whole range
862  * must be allowed (for mapping), but any areas that would otherwise be
863  * disallowed are flagged as being "zero filled" instead of rejected.
864  * Access has to be given to non-kernel-ram areas as well, these contain the
865  * PCI mmio resources as well as potential bios/acpi data regions.
866  */
867 int devmem_is_allowed(unsigned long pagenr)
868 {
869 	if (region_intersects(PFN_PHYS(pagenr), PAGE_SIZE,
870 				IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE)
871 			!= REGION_DISJOINT) {
872 		/*
873 		 * For disallowed memory regions in the low 1MB range,
874 		 * request that the page be shown as all zeros.
875 		 */
876 		if (pagenr < 256)
877 			return 2;
878 
879 		return 0;
880 	}
881 
882 	/*
883 	 * This must follow RAM test, since System RAM is considered a
884 	 * restricted resource under CONFIG_STRICT_DEVMEM.
885 	 */
886 	if (iomem_is_exclusive(pagenr << PAGE_SHIFT)) {
887 		/* Low 1MB bypasses iomem restrictions. */
888 		if (pagenr < 256)
889 			return 1;
890 
891 		return 0;
892 	}
893 
894 	return 1;
895 }
896 
897 void free_init_pages(const char *what, unsigned long begin, unsigned long end)
898 {
899 	unsigned long begin_aligned, end_aligned;
900 
901 	/* Make sure boundaries are page aligned */
902 	begin_aligned = PAGE_ALIGN(begin);
903 	end_aligned   = end & PAGE_MASK;
904 
905 	if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
906 		begin = begin_aligned;
907 		end   = end_aligned;
908 	}
909 
910 	if (begin >= end)
911 		return;
912 
913 	/*
914 	 * If debugging page accesses then do not free this memory but
915 	 * mark them not present - any buggy init-section access will
916 	 * create a kernel page fault:
917 	 */
918 	if (debug_pagealloc_enabled()) {
919 		pr_info("debug: unmapping init [mem %#010lx-%#010lx]\n",
920 			begin, end - 1);
921 		/*
922 		 * Inform kmemleak about the hole in the memory since the
923 		 * corresponding pages will be unmapped.
924 		 */
925 		kmemleak_free_part((void *)begin, end - begin);
926 		set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
927 	} else {
928 		/*
929 		 * We just marked the kernel text read only above, now that
930 		 * we are going to free part of that, we need to make that
931 		 * writeable and non-executable first.
932 		 */
933 		set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
934 		set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
935 
936 		free_reserved_area((void *)begin, (void *)end,
937 				   POISON_FREE_INITMEM, what);
938 	}
939 }
940 
941 /*
942  * begin/end can be in the direct map or the "high kernel mapping"
943  * used for the kernel image only.  free_init_pages() will do the
944  * right thing for either kind of address.
945  */
946 void free_kernel_image_pages(const char *what, void *begin, void *end)
947 {
948 	unsigned long begin_ul = (unsigned long)begin;
949 	unsigned long end_ul = (unsigned long)end;
950 	unsigned long len_pages = (end_ul - begin_ul) >> PAGE_SHIFT;
951 
952 	free_init_pages(what, begin_ul, end_ul);
953 
954 	/*
955 	 * PTI maps some of the kernel into userspace.  For performance,
956 	 * this includes some kernel areas that do not contain secrets.
957 	 * Those areas might be adjacent to the parts of the kernel image
958 	 * being freed, which may contain secrets.  Remove the "high kernel
959 	 * image mapping" for these freed areas, ensuring they are not even
960 	 * potentially vulnerable to Meltdown regardless of the specific
961 	 * optimizations PTI is currently using.
962 	 *
963 	 * The "noalias" prevents unmapping the direct map alias which is
964 	 * needed to access the freed pages.
965 	 *
966 	 * This is only valid for 64bit kernels. 32bit has only one mapping
967 	 * which can't be treated in this way for obvious reasons.
968 	 */
969 	if (IS_ENABLED(CONFIG_X86_64) && cpu_feature_enabled(X86_FEATURE_PTI))
970 		set_memory_np_noalias(begin_ul, len_pages);
971 }
972 
973 void __ref free_initmem(void)
974 {
975 	e820__reallocate_tables();
976 
977 	mem_encrypt_free_decrypted_mem();
978 
979 	free_kernel_image_pages("unused kernel image (initmem)",
980 				&__init_begin, &__init_end);
981 }
982 
983 #ifdef CONFIG_BLK_DEV_INITRD
984 void __init free_initrd_mem(unsigned long start, unsigned long end)
985 {
986 	/*
987 	 * end could be not aligned, and We can not align that,
988 	 * decompressor could be confused by aligned initrd_end
989 	 * We already reserve the end partial page before in
990 	 *   - i386_start_kernel()
991 	 *   - x86_64_start_kernel()
992 	 *   - relocate_initrd()
993 	 * So here We can do PAGE_ALIGN() safely to get partial page to be freed
994 	 */
995 	free_init_pages("initrd", start, PAGE_ALIGN(end));
996 }
997 #endif
998 
999 void __init zone_sizes_init(void)
1000 {
1001 	unsigned long max_zone_pfns[MAX_NR_ZONES];
1002 
1003 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1004 
1005 #ifdef CONFIG_ZONE_DMA
1006 	max_zone_pfns[ZONE_DMA]		= min(MAX_DMA_PFN, max_low_pfn);
1007 #endif
1008 #ifdef CONFIG_ZONE_DMA32
1009 	max_zone_pfns[ZONE_DMA32]	= min(MAX_DMA32_PFN, max_low_pfn);
1010 #endif
1011 	max_zone_pfns[ZONE_NORMAL]	= max_low_pfn;
1012 #ifdef CONFIG_HIGHMEM
1013 	max_zone_pfns[ZONE_HIGHMEM]	= max_pfn;
1014 #endif
1015 
1016 	free_area_init(max_zone_pfns);
1017 }
1018 
1019 __visible DEFINE_PER_CPU_ALIGNED(struct tlb_state, cpu_tlbstate) = {
1020 	.loaded_mm = &init_mm,
1021 	.next_asid = 1,
1022 	.cr4 = ~0UL,	/* fail hard if we screw up cr4 shadow initialization */
1023 };
1024 
1025 #ifdef CONFIG_ADDRESS_MASKING
1026 DEFINE_PER_CPU(u64, tlbstate_untag_mask);
1027 EXPORT_PER_CPU_SYMBOL(tlbstate_untag_mask);
1028 #endif
1029 
1030 void update_cache_mode_entry(unsigned entry, enum page_cache_mode cache)
1031 {
1032 	/* entry 0 MUST be WB (hardwired to speed up translations) */
1033 	BUG_ON(!entry && cache != _PAGE_CACHE_MODE_WB);
1034 
1035 	__cachemode2pte_tbl[cache] = __cm_idx2pte(entry);
1036 	__pte2cachemode_tbl[entry] = cache;
1037 }
1038 
1039 #ifdef CONFIG_SWAP
1040 unsigned long arch_max_swapfile_size(void)
1041 {
1042 	unsigned long pages;
1043 
1044 	pages = generic_max_swapfile_size();
1045 
1046 	if (boot_cpu_has_bug(X86_BUG_L1TF) && l1tf_mitigation != L1TF_MITIGATION_OFF) {
1047 		/* Limit the swap file size to MAX_PA/2 for L1TF workaround */
1048 		unsigned long long l1tf_limit = l1tf_pfn_limit();
1049 		/*
1050 		 * We encode swap offsets also with 3 bits below those for pfn
1051 		 * which makes the usable limit higher.
1052 		 */
1053 #if CONFIG_PGTABLE_LEVELS > 2
1054 		l1tf_limit <<= PAGE_SHIFT - SWP_OFFSET_FIRST_BIT;
1055 #endif
1056 		pages = min_t(unsigned long long, l1tf_limit, pages);
1057 	}
1058 	return pages;
1059 }
1060 #endif
1061 
1062 #ifdef CONFIG_EXECMEM
1063 static struct execmem_info execmem_info __ro_after_init;
1064 
1065 #ifdef CONFIG_ARCH_HAS_EXECMEM_ROX
1066 void execmem_fill_trapping_insns(void *ptr, size_t size, bool writeable)
1067 {
1068 	/* fill memory with INT3 instructions */
1069 	if (writeable)
1070 		memset(ptr, INT3_INSN_OPCODE, size);
1071 	else
1072 		text_poke_set(ptr, INT3_INSN_OPCODE, size);
1073 }
1074 #endif
1075 
1076 struct execmem_info __init *execmem_arch_setup(void)
1077 {
1078 	unsigned long start, offset = 0;
1079 	enum execmem_range_flags flags;
1080 	pgprot_t pgprot;
1081 
1082 	if (kaslr_enabled())
1083 		offset = get_random_u32_inclusive(1, 1024) * PAGE_SIZE;
1084 
1085 	start = MODULES_VADDR + offset;
1086 
1087 	if (IS_ENABLED(CONFIG_ARCH_HAS_EXECMEM_ROX) &&
1088 	    cpu_feature_enabled(X86_FEATURE_PSE)) {
1089 		pgprot = PAGE_KERNEL_ROX;
1090 		flags = EXECMEM_KASAN_SHADOW | EXECMEM_ROX_CACHE;
1091 	} else {
1092 		pgprot = PAGE_KERNEL;
1093 		flags = EXECMEM_KASAN_SHADOW;
1094 	}
1095 
1096 	execmem_info = (struct execmem_info){
1097 		.ranges = {
1098 			[EXECMEM_MODULE_TEXT] = {
1099 				.flags	= flags,
1100 				.start	= start,
1101 				.end	= MODULES_END,
1102 				.pgprot	= pgprot,
1103 				.alignment = MODULE_ALIGN,
1104 			},
1105 			[EXECMEM_KPROBES ... EXECMEM_BPF] = {
1106 				.flags	= EXECMEM_KASAN_SHADOW,
1107 				.start	= start,
1108 				.end	= MODULES_END,
1109 				.pgprot	= PAGE_KERNEL,
1110 				.alignment = MODULE_ALIGN,
1111 			},
1112 			[EXECMEM_MODULE_DATA] = {
1113 				.flags	= EXECMEM_KASAN_SHADOW,
1114 				.start	= start,
1115 				.end	= MODULES_END,
1116 				.pgprot	= PAGE_KERNEL,
1117 				.alignment = MODULE_ALIGN,
1118 			},
1119 		},
1120 	};
1121 
1122 	return &execmem_info;
1123 }
1124 #endif /* CONFIG_EXECMEM */
1125