xref: /openbmc/linux/arch/x86/mm/init.c (revision 0545810f)
1 #include <linux/gfp.h>
2 #include <linux/initrd.h>
3 #include <linux/ioport.h>
4 #include <linux/swap.h>
5 #include <linux/memblock.h>
6 #include <linux/swapfile.h>
7 #include <linux/swapops.h>
8 #include <linux/kmemleak.h>
9 #include <linux/sched/task.h>
10 
11 #include <asm/set_memory.h>
12 #include <asm/e820/api.h>
13 #include <asm/init.h>
14 #include <asm/page.h>
15 #include <asm/page_types.h>
16 #include <asm/sections.h>
17 #include <asm/setup.h>
18 #include <asm/tlbflush.h>
19 #include <asm/tlb.h>
20 #include <asm/proto.h>
21 #include <asm/dma.h>		/* for MAX_DMA_PFN */
22 #include <asm/microcode.h>
23 #include <asm/kaslr.h>
24 #include <asm/hypervisor.h>
25 #include <asm/cpufeature.h>
26 #include <asm/pti.h>
27 #include <asm/text-patching.h>
28 #include <asm/memtype.h>
29 #include <asm/paravirt.h>
30 
31 /*
32  * We need to define the tracepoints somewhere, and tlb.c
33  * is only compiled when SMP=y.
34  */
35 #include <trace/events/tlb.h>
36 
37 #include "mm_internal.h"
38 
39 /*
40  * Tables translating between page_cache_type_t and pte encoding.
41  *
42  * The default values are defined statically as minimal supported mode;
43  * WC and WT fall back to UC-.  pat_init() updates these values to support
44  * more cache modes, WC and WT, when it is safe to do so.  See pat_init()
45  * for the details.  Note, __early_ioremap() used during early boot-time
46  * takes pgprot_t (pte encoding) and does not use these tables.
47  *
48  *   Index into __cachemode2pte_tbl[] is the cachemode.
49  *
50  *   Index into __pte2cachemode_tbl[] are the caching attribute bits of the pte
51  *   (_PAGE_PWT, _PAGE_PCD, _PAGE_PAT) at index bit positions 0, 1, 2.
52  */
53 static uint16_t __cachemode2pte_tbl[_PAGE_CACHE_MODE_NUM] = {
54 	[_PAGE_CACHE_MODE_WB      ]	= 0         | 0        ,
55 	[_PAGE_CACHE_MODE_WC      ]	= 0         | _PAGE_PCD,
56 	[_PAGE_CACHE_MODE_UC_MINUS]	= 0         | _PAGE_PCD,
57 	[_PAGE_CACHE_MODE_UC      ]	= _PAGE_PWT | _PAGE_PCD,
58 	[_PAGE_CACHE_MODE_WT      ]	= 0         | _PAGE_PCD,
59 	[_PAGE_CACHE_MODE_WP      ]	= 0         | _PAGE_PCD,
60 };
61 
62 unsigned long cachemode2protval(enum page_cache_mode pcm)
63 {
64 	if (likely(pcm == 0))
65 		return 0;
66 	return __cachemode2pte_tbl[pcm];
67 }
68 EXPORT_SYMBOL(cachemode2protval);
69 
70 static uint8_t __pte2cachemode_tbl[8] = {
71 	[__pte2cm_idx( 0        | 0         | 0        )] = _PAGE_CACHE_MODE_WB,
72 	[__pte2cm_idx(_PAGE_PWT | 0         | 0        )] = _PAGE_CACHE_MODE_UC_MINUS,
73 	[__pte2cm_idx( 0        | _PAGE_PCD | 0        )] = _PAGE_CACHE_MODE_UC_MINUS,
74 	[__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | 0        )] = _PAGE_CACHE_MODE_UC,
75 	[__pte2cm_idx( 0        | 0         | _PAGE_PAT)] = _PAGE_CACHE_MODE_WB,
76 	[__pte2cm_idx(_PAGE_PWT | 0         | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
77 	[__pte2cm_idx(0         | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
78 	[__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC,
79 };
80 
81 /*
82  * Check that the write-protect PAT entry is set for write-protect.
83  * To do this without making assumptions how PAT has been set up (Xen has
84  * another layout than the kernel), translate the _PAGE_CACHE_MODE_WP cache
85  * mode via the __cachemode2pte_tbl[] into protection bits (those protection
86  * bits will select a cache mode of WP or better), and then translate the
87  * protection bits back into the cache mode using __pte2cm_idx() and the
88  * __pte2cachemode_tbl[] array. This will return the really used cache mode.
89  */
90 bool x86_has_pat_wp(void)
91 {
92 	uint16_t prot = __cachemode2pte_tbl[_PAGE_CACHE_MODE_WP];
93 
94 	return __pte2cachemode_tbl[__pte2cm_idx(prot)] == _PAGE_CACHE_MODE_WP;
95 }
96 
97 enum page_cache_mode pgprot2cachemode(pgprot_t pgprot)
98 {
99 	unsigned long masked;
100 
101 	masked = pgprot_val(pgprot) & _PAGE_CACHE_MASK;
102 	if (likely(masked == 0))
103 		return 0;
104 	return __pte2cachemode_tbl[__pte2cm_idx(masked)];
105 }
106 
107 static unsigned long __initdata pgt_buf_start;
108 static unsigned long __initdata pgt_buf_end;
109 static unsigned long __initdata pgt_buf_top;
110 
111 static unsigned long min_pfn_mapped;
112 
113 static bool __initdata can_use_brk_pgt = true;
114 
115 /*
116  * Pages returned are already directly mapped.
117  *
118  * Changing that is likely to break Xen, see commit:
119  *
120  *    279b706 x86,xen: introduce x86_init.mapping.pagetable_reserve
121  *
122  * for detailed information.
123  */
124 __ref void *alloc_low_pages(unsigned int num)
125 {
126 	unsigned long pfn;
127 	int i;
128 
129 	if (after_bootmem) {
130 		unsigned int order;
131 
132 		order = get_order((unsigned long)num << PAGE_SHIFT);
133 		return (void *)__get_free_pages(GFP_ATOMIC | __GFP_ZERO, order);
134 	}
135 
136 	if ((pgt_buf_end + num) > pgt_buf_top || !can_use_brk_pgt) {
137 		unsigned long ret = 0;
138 
139 		if (min_pfn_mapped < max_pfn_mapped) {
140 			ret = memblock_phys_alloc_range(
141 					PAGE_SIZE * num, PAGE_SIZE,
142 					min_pfn_mapped << PAGE_SHIFT,
143 					max_pfn_mapped << PAGE_SHIFT);
144 		}
145 		if (!ret && can_use_brk_pgt)
146 			ret = __pa(extend_brk(PAGE_SIZE * num, PAGE_SIZE));
147 
148 		if (!ret)
149 			panic("alloc_low_pages: can not alloc memory");
150 
151 		pfn = ret >> PAGE_SHIFT;
152 	} else {
153 		pfn = pgt_buf_end;
154 		pgt_buf_end += num;
155 	}
156 
157 	for (i = 0; i < num; i++) {
158 		void *adr;
159 
160 		adr = __va((pfn + i) << PAGE_SHIFT);
161 		clear_page(adr);
162 	}
163 
164 	return __va(pfn << PAGE_SHIFT);
165 }
166 
167 /*
168  * By default need to be able to allocate page tables below PGD firstly for
169  * the 0-ISA_END_ADDRESS range and secondly for the initial PMD_SIZE mapping.
170  * With KASLR memory randomization, depending on the machine e820 memory and the
171  * PUD alignment, twice that many pages may be needed when KASLR memory
172  * randomization is enabled.
173  */
174 
175 #ifndef CONFIG_X86_5LEVEL
176 #define INIT_PGD_PAGE_TABLES    3
177 #else
178 #define INIT_PGD_PAGE_TABLES    4
179 #endif
180 
181 #ifndef CONFIG_RANDOMIZE_MEMORY
182 #define INIT_PGD_PAGE_COUNT      (2 * INIT_PGD_PAGE_TABLES)
183 #else
184 #define INIT_PGD_PAGE_COUNT      (4 * INIT_PGD_PAGE_TABLES)
185 #endif
186 
187 #define INIT_PGT_BUF_SIZE	(INIT_PGD_PAGE_COUNT * PAGE_SIZE)
188 RESERVE_BRK(early_pgt_alloc, INIT_PGT_BUF_SIZE);
189 void  __init early_alloc_pgt_buf(void)
190 {
191 	unsigned long tables = INIT_PGT_BUF_SIZE;
192 	phys_addr_t base;
193 
194 	base = __pa(extend_brk(tables, PAGE_SIZE));
195 
196 	pgt_buf_start = base >> PAGE_SHIFT;
197 	pgt_buf_end = pgt_buf_start;
198 	pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
199 }
200 
201 int after_bootmem;
202 
203 early_param_on_off("gbpages", "nogbpages", direct_gbpages, CONFIG_X86_DIRECT_GBPAGES);
204 
205 struct map_range {
206 	unsigned long start;
207 	unsigned long end;
208 	unsigned page_size_mask;
209 };
210 
211 static int page_size_mask;
212 
213 /*
214  * Save some of cr4 feature set we're using (e.g.  Pentium 4MB
215  * enable and PPro Global page enable), so that any CPU's that boot
216  * up after us can get the correct flags. Invoked on the boot CPU.
217  */
218 static inline void cr4_set_bits_and_update_boot(unsigned long mask)
219 {
220 	mmu_cr4_features |= mask;
221 	if (trampoline_cr4_features)
222 		*trampoline_cr4_features = mmu_cr4_features;
223 	cr4_set_bits(mask);
224 }
225 
226 static void __init probe_page_size_mask(void)
227 {
228 	/*
229 	 * For pagealloc debugging, identity mapping will use small pages.
230 	 * This will simplify cpa(), which otherwise needs to support splitting
231 	 * large pages into small in interrupt context, etc.
232 	 */
233 	if (boot_cpu_has(X86_FEATURE_PSE) && !debug_pagealloc_enabled())
234 		page_size_mask |= 1 << PG_LEVEL_2M;
235 	else
236 		direct_gbpages = 0;
237 
238 	/* Enable PSE if available */
239 	if (boot_cpu_has(X86_FEATURE_PSE))
240 		cr4_set_bits_and_update_boot(X86_CR4_PSE);
241 
242 	/* Enable PGE if available */
243 	__supported_pte_mask &= ~_PAGE_GLOBAL;
244 	if (boot_cpu_has(X86_FEATURE_PGE)) {
245 		cr4_set_bits_and_update_boot(X86_CR4_PGE);
246 		__supported_pte_mask |= _PAGE_GLOBAL;
247 	}
248 
249 	/* By the default is everything supported: */
250 	__default_kernel_pte_mask = __supported_pte_mask;
251 	/* Except when with PTI where the kernel is mostly non-Global: */
252 	if (cpu_feature_enabled(X86_FEATURE_PTI))
253 		__default_kernel_pte_mask &= ~_PAGE_GLOBAL;
254 
255 	/* Enable 1 GB linear kernel mappings if available: */
256 	if (direct_gbpages && boot_cpu_has(X86_FEATURE_GBPAGES)) {
257 		printk(KERN_INFO "Using GB pages for direct mapping\n");
258 		page_size_mask |= 1 << PG_LEVEL_1G;
259 	} else {
260 		direct_gbpages = 0;
261 	}
262 }
263 
264 static void setup_pcid(void)
265 {
266 	if (!IS_ENABLED(CONFIG_X86_64))
267 		return;
268 
269 	if (!boot_cpu_has(X86_FEATURE_PCID))
270 		return;
271 
272 	if (boot_cpu_has(X86_FEATURE_PGE)) {
273 		/*
274 		 * This can't be cr4_set_bits_and_update_boot() -- the
275 		 * trampoline code can't handle CR4.PCIDE and it wouldn't
276 		 * do any good anyway.  Despite the name,
277 		 * cr4_set_bits_and_update_boot() doesn't actually cause
278 		 * the bits in question to remain set all the way through
279 		 * the secondary boot asm.
280 		 *
281 		 * Instead, we brute-force it and set CR4.PCIDE manually in
282 		 * start_secondary().
283 		 */
284 		cr4_set_bits(X86_CR4_PCIDE);
285 
286 		/*
287 		 * INVPCID's single-context modes (2/3) only work if we set
288 		 * X86_CR4_PCIDE, *and* we INVPCID support.  It's unusable
289 		 * on systems that have X86_CR4_PCIDE clear, or that have
290 		 * no INVPCID support at all.
291 		 */
292 		if (boot_cpu_has(X86_FEATURE_INVPCID))
293 			setup_force_cpu_cap(X86_FEATURE_INVPCID_SINGLE);
294 	} else {
295 		/*
296 		 * flush_tlb_all(), as currently implemented, won't work if
297 		 * PCID is on but PGE is not.  Since that combination
298 		 * doesn't exist on real hardware, there's no reason to try
299 		 * to fully support it, but it's polite to avoid corrupting
300 		 * data if we're on an improperly configured VM.
301 		 */
302 		setup_clear_cpu_cap(X86_FEATURE_PCID);
303 	}
304 }
305 
306 #ifdef CONFIG_X86_32
307 #define NR_RANGE_MR 3
308 #else /* CONFIG_X86_64 */
309 #define NR_RANGE_MR 5
310 #endif
311 
312 static int __meminit save_mr(struct map_range *mr, int nr_range,
313 			     unsigned long start_pfn, unsigned long end_pfn,
314 			     unsigned long page_size_mask)
315 {
316 	if (start_pfn < end_pfn) {
317 		if (nr_range >= NR_RANGE_MR)
318 			panic("run out of range for init_memory_mapping\n");
319 		mr[nr_range].start = start_pfn<<PAGE_SHIFT;
320 		mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
321 		mr[nr_range].page_size_mask = page_size_mask;
322 		nr_range++;
323 	}
324 
325 	return nr_range;
326 }
327 
328 /*
329  * adjust the page_size_mask for small range to go with
330  *	big page size instead small one if nearby are ram too.
331  */
332 static void __ref adjust_range_page_size_mask(struct map_range *mr,
333 							 int nr_range)
334 {
335 	int i;
336 
337 	for (i = 0; i < nr_range; i++) {
338 		if ((page_size_mask & (1<<PG_LEVEL_2M)) &&
339 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) {
340 			unsigned long start = round_down(mr[i].start, PMD_SIZE);
341 			unsigned long end = round_up(mr[i].end, PMD_SIZE);
342 
343 #ifdef CONFIG_X86_32
344 			if ((end >> PAGE_SHIFT) > max_low_pfn)
345 				continue;
346 #endif
347 
348 			if (memblock_is_region_memory(start, end - start))
349 				mr[i].page_size_mask |= 1<<PG_LEVEL_2M;
350 		}
351 		if ((page_size_mask & (1<<PG_LEVEL_1G)) &&
352 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) {
353 			unsigned long start = round_down(mr[i].start, PUD_SIZE);
354 			unsigned long end = round_up(mr[i].end, PUD_SIZE);
355 
356 			if (memblock_is_region_memory(start, end - start))
357 				mr[i].page_size_mask |= 1<<PG_LEVEL_1G;
358 		}
359 	}
360 }
361 
362 static const char *page_size_string(struct map_range *mr)
363 {
364 	static const char str_1g[] = "1G";
365 	static const char str_2m[] = "2M";
366 	static const char str_4m[] = "4M";
367 	static const char str_4k[] = "4k";
368 
369 	if (mr->page_size_mask & (1<<PG_LEVEL_1G))
370 		return str_1g;
371 	/*
372 	 * 32-bit without PAE has a 4M large page size.
373 	 * PG_LEVEL_2M is misnamed, but we can at least
374 	 * print out the right size in the string.
375 	 */
376 	if (IS_ENABLED(CONFIG_X86_32) &&
377 	    !IS_ENABLED(CONFIG_X86_PAE) &&
378 	    mr->page_size_mask & (1<<PG_LEVEL_2M))
379 		return str_4m;
380 
381 	if (mr->page_size_mask & (1<<PG_LEVEL_2M))
382 		return str_2m;
383 
384 	return str_4k;
385 }
386 
387 static int __meminit split_mem_range(struct map_range *mr, int nr_range,
388 				     unsigned long start,
389 				     unsigned long end)
390 {
391 	unsigned long start_pfn, end_pfn, limit_pfn;
392 	unsigned long pfn;
393 	int i;
394 
395 	limit_pfn = PFN_DOWN(end);
396 
397 	/* head if not big page alignment ? */
398 	pfn = start_pfn = PFN_DOWN(start);
399 #ifdef CONFIG_X86_32
400 	/*
401 	 * Don't use a large page for the first 2/4MB of memory
402 	 * because there are often fixed size MTRRs in there
403 	 * and overlapping MTRRs into large pages can cause
404 	 * slowdowns.
405 	 */
406 	if (pfn == 0)
407 		end_pfn = PFN_DOWN(PMD_SIZE);
408 	else
409 		end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
410 #else /* CONFIG_X86_64 */
411 	end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
412 #endif
413 	if (end_pfn > limit_pfn)
414 		end_pfn = limit_pfn;
415 	if (start_pfn < end_pfn) {
416 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
417 		pfn = end_pfn;
418 	}
419 
420 	/* big page (2M) range */
421 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
422 #ifdef CONFIG_X86_32
423 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
424 #else /* CONFIG_X86_64 */
425 	end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
426 	if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE)))
427 		end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
428 #endif
429 
430 	if (start_pfn < end_pfn) {
431 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
432 				page_size_mask & (1<<PG_LEVEL_2M));
433 		pfn = end_pfn;
434 	}
435 
436 #ifdef CONFIG_X86_64
437 	/* big page (1G) range */
438 	start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
439 	end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE));
440 	if (start_pfn < end_pfn) {
441 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
442 				page_size_mask &
443 				 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
444 		pfn = end_pfn;
445 	}
446 
447 	/* tail is not big page (1G) alignment */
448 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
449 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
450 	if (start_pfn < end_pfn) {
451 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
452 				page_size_mask & (1<<PG_LEVEL_2M));
453 		pfn = end_pfn;
454 	}
455 #endif
456 
457 	/* tail is not big page (2M) alignment */
458 	start_pfn = pfn;
459 	end_pfn = limit_pfn;
460 	nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
461 
462 	if (!after_bootmem)
463 		adjust_range_page_size_mask(mr, nr_range);
464 
465 	/* try to merge same page size and continuous */
466 	for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
467 		unsigned long old_start;
468 		if (mr[i].end != mr[i+1].start ||
469 		    mr[i].page_size_mask != mr[i+1].page_size_mask)
470 			continue;
471 		/* move it */
472 		old_start = mr[i].start;
473 		memmove(&mr[i], &mr[i+1],
474 			(nr_range - 1 - i) * sizeof(struct map_range));
475 		mr[i--].start = old_start;
476 		nr_range--;
477 	}
478 
479 	for (i = 0; i < nr_range; i++)
480 		pr_debug(" [mem %#010lx-%#010lx] page %s\n",
481 				mr[i].start, mr[i].end - 1,
482 				page_size_string(&mr[i]));
483 
484 	return nr_range;
485 }
486 
487 struct range pfn_mapped[E820_MAX_ENTRIES];
488 int nr_pfn_mapped;
489 
490 static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn)
491 {
492 	nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_MAX_ENTRIES,
493 					     nr_pfn_mapped, start_pfn, end_pfn);
494 	nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_MAX_ENTRIES);
495 
496 	max_pfn_mapped = max(max_pfn_mapped, end_pfn);
497 
498 	if (start_pfn < (1UL<<(32-PAGE_SHIFT)))
499 		max_low_pfn_mapped = max(max_low_pfn_mapped,
500 					 min(end_pfn, 1UL<<(32-PAGE_SHIFT)));
501 }
502 
503 bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn)
504 {
505 	int i;
506 
507 	for (i = 0; i < nr_pfn_mapped; i++)
508 		if ((start_pfn >= pfn_mapped[i].start) &&
509 		    (end_pfn <= pfn_mapped[i].end))
510 			return true;
511 
512 	return false;
513 }
514 
515 /*
516  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
517  * This runs before bootmem is initialized and gets pages directly from
518  * the physical memory. To access them they are temporarily mapped.
519  */
520 unsigned long __ref init_memory_mapping(unsigned long start,
521 					unsigned long end, pgprot_t prot)
522 {
523 	struct map_range mr[NR_RANGE_MR];
524 	unsigned long ret = 0;
525 	int nr_range, i;
526 
527 	pr_debug("init_memory_mapping: [mem %#010lx-%#010lx]\n",
528 	       start, end - 1);
529 
530 	memset(mr, 0, sizeof(mr));
531 	nr_range = split_mem_range(mr, 0, start, end);
532 
533 	for (i = 0; i < nr_range; i++)
534 		ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
535 						   mr[i].page_size_mask,
536 						   prot);
537 
538 	add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT);
539 
540 	return ret >> PAGE_SHIFT;
541 }
542 
543 /*
544  * We need to iterate through the E820 memory map and create direct mappings
545  * for only E820_TYPE_RAM and E820_KERN_RESERVED regions. We cannot simply
546  * create direct mappings for all pfns from [0 to max_low_pfn) and
547  * [4GB to max_pfn) because of possible memory holes in high addresses
548  * that cannot be marked as UC by fixed/variable range MTRRs.
549  * Depending on the alignment of E820 ranges, this may possibly result
550  * in using smaller size (i.e. 4K instead of 2M or 1G) page tables.
551  *
552  * init_mem_mapping() calls init_range_memory_mapping() with big range.
553  * That range would have hole in the middle or ends, and only ram parts
554  * will be mapped in init_range_memory_mapping().
555  */
556 static unsigned long __init init_range_memory_mapping(
557 					   unsigned long r_start,
558 					   unsigned long r_end)
559 {
560 	unsigned long start_pfn, end_pfn;
561 	unsigned long mapped_ram_size = 0;
562 	int i;
563 
564 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
565 		u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end);
566 		u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end);
567 		if (start >= end)
568 			continue;
569 
570 		/*
571 		 * if it is overlapping with brk pgt, we need to
572 		 * alloc pgt buf from memblock instead.
573 		 */
574 		can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >=
575 				    min(end, (u64)pgt_buf_top<<PAGE_SHIFT);
576 		init_memory_mapping(start, end, PAGE_KERNEL);
577 		mapped_ram_size += end - start;
578 		can_use_brk_pgt = true;
579 	}
580 
581 	return mapped_ram_size;
582 }
583 
584 static unsigned long __init get_new_step_size(unsigned long step_size)
585 {
586 	/*
587 	 * Initial mapped size is PMD_SIZE (2M).
588 	 * We can not set step_size to be PUD_SIZE (1G) yet.
589 	 * In worse case, when we cross the 1G boundary, and
590 	 * PG_LEVEL_2M is not set, we will need 1+1+512 pages (2M + 8k)
591 	 * to map 1G range with PTE. Hence we use one less than the
592 	 * difference of page table level shifts.
593 	 *
594 	 * Don't need to worry about overflow in the top-down case, on 32bit,
595 	 * when step_size is 0, round_down() returns 0 for start, and that
596 	 * turns it into 0x100000000ULL.
597 	 * In the bottom-up case, round_up(x, 0) returns 0 though too, which
598 	 * needs to be taken into consideration by the code below.
599 	 */
600 	return step_size << (PMD_SHIFT - PAGE_SHIFT - 1);
601 }
602 
603 /**
604  * memory_map_top_down - Map [map_start, map_end) top down
605  * @map_start: start address of the target memory range
606  * @map_end: end address of the target memory range
607  *
608  * This function will setup direct mapping for memory range
609  * [map_start, map_end) in top-down. That said, the page tables
610  * will be allocated at the end of the memory, and we map the
611  * memory in top-down.
612  */
613 static void __init memory_map_top_down(unsigned long map_start,
614 				       unsigned long map_end)
615 {
616 	unsigned long real_end, last_start;
617 	unsigned long step_size;
618 	unsigned long addr;
619 	unsigned long mapped_ram_size = 0;
620 
621 	/*
622 	 * Systems that have many reserved areas near top of the memory,
623 	 * e.g. QEMU with less than 1G RAM and EFI enabled, or Xen, will
624 	 * require lots of 4K mappings which may exhaust pgt_buf.
625 	 * Start with top-most PMD_SIZE range aligned at PMD_SIZE to ensure
626 	 * there is enough mapped memory that can be allocated from
627 	 * memblock.
628 	 */
629 	addr = memblock_phys_alloc_range(PMD_SIZE, PMD_SIZE, map_start,
630 					 map_end);
631 	memblock_phys_free(addr, PMD_SIZE);
632 	real_end = addr + PMD_SIZE;
633 
634 	/* step_size need to be small so pgt_buf from BRK could cover it */
635 	step_size = PMD_SIZE;
636 	max_pfn_mapped = 0; /* will get exact value next */
637 	min_pfn_mapped = real_end >> PAGE_SHIFT;
638 	last_start = real_end;
639 
640 	/*
641 	 * We start from the top (end of memory) and go to the bottom.
642 	 * The memblock_find_in_range() gets us a block of RAM from the
643 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
644 	 * for page table.
645 	 */
646 	while (last_start > map_start) {
647 		unsigned long start;
648 
649 		if (last_start > step_size) {
650 			start = round_down(last_start - 1, step_size);
651 			if (start < map_start)
652 				start = map_start;
653 		} else
654 			start = map_start;
655 		mapped_ram_size += init_range_memory_mapping(start,
656 							last_start);
657 		last_start = start;
658 		min_pfn_mapped = last_start >> PAGE_SHIFT;
659 		if (mapped_ram_size >= step_size)
660 			step_size = get_new_step_size(step_size);
661 	}
662 
663 	if (real_end < map_end)
664 		init_range_memory_mapping(real_end, map_end);
665 }
666 
667 /**
668  * memory_map_bottom_up - Map [map_start, map_end) bottom up
669  * @map_start: start address of the target memory range
670  * @map_end: end address of the target memory range
671  *
672  * This function will setup direct mapping for memory range
673  * [map_start, map_end) in bottom-up. Since we have limited the
674  * bottom-up allocation above the kernel, the page tables will
675  * be allocated just above the kernel and we map the memory
676  * in [map_start, map_end) in bottom-up.
677  */
678 static void __init memory_map_bottom_up(unsigned long map_start,
679 					unsigned long map_end)
680 {
681 	unsigned long next, start;
682 	unsigned long mapped_ram_size = 0;
683 	/* step_size need to be small so pgt_buf from BRK could cover it */
684 	unsigned long step_size = PMD_SIZE;
685 
686 	start = map_start;
687 	min_pfn_mapped = start >> PAGE_SHIFT;
688 
689 	/*
690 	 * We start from the bottom (@map_start) and go to the top (@map_end).
691 	 * The memblock_find_in_range() gets us a block of RAM from the
692 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
693 	 * for page table.
694 	 */
695 	while (start < map_end) {
696 		if (step_size && map_end - start > step_size) {
697 			next = round_up(start + 1, step_size);
698 			if (next > map_end)
699 				next = map_end;
700 		} else {
701 			next = map_end;
702 		}
703 
704 		mapped_ram_size += init_range_memory_mapping(start, next);
705 		start = next;
706 
707 		if (mapped_ram_size >= step_size)
708 			step_size = get_new_step_size(step_size);
709 	}
710 }
711 
712 /*
713  * The real mode trampoline, which is required for bootstrapping CPUs
714  * occupies only a small area under the low 1MB.  See reserve_real_mode()
715  * for details.
716  *
717  * If KASLR is disabled the first PGD entry of the direct mapping is copied
718  * to map the real mode trampoline.
719  *
720  * If KASLR is enabled, copy only the PUD which covers the low 1MB
721  * area. This limits the randomization granularity to 1GB for both 4-level
722  * and 5-level paging.
723  */
724 static void __init init_trampoline(void)
725 {
726 #ifdef CONFIG_X86_64
727 	/*
728 	 * The code below will alias kernel page-tables in the user-range of the
729 	 * address space, including the Global bit. So global TLB entries will
730 	 * be created when using the trampoline page-table.
731 	 */
732 	if (!kaslr_memory_enabled())
733 		trampoline_pgd_entry = init_top_pgt[pgd_index(__PAGE_OFFSET)];
734 	else
735 		init_trampoline_kaslr();
736 #endif
737 }
738 
739 void __init init_mem_mapping(void)
740 {
741 	unsigned long end;
742 
743 	pti_check_boottime_disable();
744 	probe_page_size_mask();
745 	setup_pcid();
746 
747 #ifdef CONFIG_X86_64
748 	end = max_pfn << PAGE_SHIFT;
749 #else
750 	end = max_low_pfn << PAGE_SHIFT;
751 #endif
752 
753 	/* the ISA range is always mapped regardless of memory holes */
754 	init_memory_mapping(0, ISA_END_ADDRESS, PAGE_KERNEL);
755 
756 	/* Init the trampoline, possibly with KASLR memory offset */
757 	init_trampoline();
758 
759 	/*
760 	 * If the allocation is in bottom-up direction, we setup direct mapping
761 	 * in bottom-up, otherwise we setup direct mapping in top-down.
762 	 */
763 	if (memblock_bottom_up()) {
764 		unsigned long kernel_end = __pa_symbol(_end);
765 
766 		/*
767 		 * we need two separate calls here. This is because we want to
768 		 * allocate page tables above the kernel. So we first map
769 		 * [kernel_end, end) to make memory above the kernel be mapped
770 		 * as soon as possible. And then use page tables allocated above
771 		 * the kernel to map [ISA_END_ADDRESS, kernel_end).
772 		 */
773 		memory_map_bottom_up(kernel_end, end);
774 		memory_map_bottom_up(ISA_END_ADDRESS, kernel_end);
775 	} else {
776 		memory_map_top_down(ISA_END_ADDRESS, end);
777 	}
778 
779 #ifdef CONFIG_X86_64
780 	if (max_pfn > max_low_pfn) {
781 		/* can we preserve max_low_pfn ?*/
782 		max_low_pfn = max_pfn;
783 	}
784 #else
785 	early_ioremap_page_table_range_init();
786 #endif
787 
788 	load_cr3(swapper_pg_dir);
789 	__flush_tlb_all();
790 
791 	x86_init.hyper.init_mem_mapping();
792 
793 	early_memtest(0, max_pfn_mapped << PAGE_SHIFT);
794 }
795 
796 /*
797  * Initialize an mm_struct to be used during poking and a pointer to be used
798  * during patching.
799  */
800 void __init poking_init(void)
801 {
802 	spinlock_t *ptl;
803 	pte_t *ptep;
804 
805 	poking_mm = mm_alloc();
806 	BUG_ON(!poking_mm);
807 
808 	/* Xen PV guests need the PGD to be pinned. */
809 	paravirt_enter_mmap(poking_mm);
810 
811 	/*
812 	 * Randomize the poking address, but make sure that the following page
813 	 * will be mapped at the same PMD. We need 2 pages, so find space for 3,
814 	 * and adjust the address if the PMD ends after the first one.
815 	 */
816 	poking_addr = TASK_UNMAPPED_BASE;
817 	if (IS_ENABLED(CONFIG_RANDOMIZE_BASE))
818 		poking_addr += (kaslr_get_random_long("Poking") & PAGE_MASK) %
819 			(TASK_SIZE - TASK_UNMAPPED_BASE - 3 * PAGE_SIZE);
820 
821 	if (((poking_addr + PAGE_SIZE) & ~PMD_MASK) == 0)
822 		poking_addr += PAGE_SIZE;
823 
824 	/*
825 	 * We need to trigger the allocation of the page-tables that will be
826 	 * needed for poking now. Later, poking may be performed in an atomic
827 	 * section, which might cause allocation to fail.
828 	 */
829 	ptep = get_locked_pte(poking_mm, poking_addr, &ptl);
830 	BUG_ON(!ptep);
831 	pte_unmap_unlock(ptep, ptl);
832 }
833 
834 /*
835  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
836  * is valid. The argument is a physical page number.
837  *
838  * On x86, access has to be given to the first megabyte of RAM because that
839  * area traditionally contains BIOS code and data regions used by X, dosemu,
840  * and similar apps. Since they map the entire memory range, the whole range
841  * must be allowed (for mapping), but any areas that would otherwise be
842  * disallowed are flagged as being "zero filled" instead of rejected.
843  * Access has to be given to non-kernel-ram areas as well, these contain the
844  * PCI mmio resources as well as potential bios/acpi data regions.
845  */
846 int devmem_is_allowed(unsigned long pagenr)
847 {
848 	if (region_intersects(PFN_PHYS(pagenr), PAGE_SIZE,
849 				IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE)
850 			!= REGION_DISJOINT) {
851 		/*
852 		 * For disallowed memory regions in the low 1MB range,
853 		 * request that the page be shown as all zeros.
854 		 */
855 		if (pagenr < 256)
856 			return 2;
857 
858 		return 0;
859 	}
860 
861 	/*
862 	 * This must follow RAM test, since System RAM is considered a
863 	 * restricted resource under CONFIG_STRICT_DEVMEM.
864 	 */
865 	if (iomem_is_exclusive(pagenr << PAGE_SHIFT)) {
866 		/* Low 1MB bypasses iomem restrictions. */
867 		if (pagenr < 256)
868 			return 1;
869 
870 		return 0;
871 	}
872 
873 	return 1;
874 }
875 
876 void free_init_pages(const char *what, unsigned long begin, unsigned long end)
877 {
878 	unsigned long begin_aligned, end_aligned;
879 
880 	/* Make sure boundaries are page aligned */
881 	begin_aligned = PAGE_ALIGN(begin);
882 	end_aligned   = end & PAGE_MASK;
883 
884 	if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
885 		begin = begin_aligned;
886 		end   = end_aligned;
887 	}
888 
889 	if (begin >= end)
890 		return;
891 
892 	/*
893 	 * If debugging page accesses then do not free this memory but
894 	 * mark them not present - any buggy init-section access will
895 	 * create a kernel page fault:
896 	 */
897 	if (debug_pagealloc_enabled()) {
898 		pr_info("debug: unmapping init [mem %#010lx-%#010lx]\n",
899 			begin, end - 1);
900 		/*
901 		 * Inform kmemleak about the hole in the memory since the
902 		 * corresponding pages will be unmapped.
903 		 */
904 		kmemleak_free_part((void *)begin, end - begin);
905 		set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
906 	} else {
907 		/*
908 		 * We just marked the kernel text read only above, now that
909 		 * we are going to free part of that, we need to make that
910 		 * writeable and non-executable first.
911 		 */
912 		set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
913 		set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
914 
915 		free_reserved_area((void *)begin, (void *)end,
916 				   POISON_FREE_INITMEM, what);
917 	}
918 }
919 
920 /*
921  * begin/end can be in the direct map or the "high kernel mapping"
922  * used for the kernel image only.  free_init_pages() will do the
923  * right thing for either kind of address.
924  */
925 void free_kernel_image_pages(const char *what, void *begin, void *end)
926 {
927 	unsigned long begin_ul = (unsigned long)begin;
928 	unsigned long end_ul = (unsigned long)end;
929 	unsigned long len_pages = (end_ul - begin_ul) >> PAGE_SHIFT;
930 
931 	free_init_pages(what, begin_ul, end_ul);
932 
933 	/*
934 	 * PTI maps some of the kernel into userspace.  For performance,
935 	 * this includes some kernel areas that do not contain secrets.
936 	 * Those areas might be adjacent to the parts of the kernel image
937 	 * being freed, which may contain secrets.  Remove the "high kernel
938 	 * image mapping" for these freed areas, ensuring they are not even
939 	 * potentially vulnerable to Meltdown regardless of the specific
940 	 * optimizations PTI is currently using.
941 	 *
942 	 * The "noalias" prevents unmapping the direct map alias which is
943 	 * needed to access the freed pages.
944 	 *
945 	 * This is only valid for 64bit kernels. 32bit has only one mapping
946 	 * which can't be treated in this way for obvious reasons.
947 	 */
948 	if (IS_ENABLED(CONFIG_X86_64) && cpu_feature_enabled(X86_FEATURE_PTI))
949 		set_memory_np_noalias(begin_ul, len_pages);
950 }
951 
952 void __ref free_initmem(void)
953 {
954 	e820__reallocate_tables();
955 
956 	mem_encrypt_free_decrypted_mem();
957 
958 	free_kernel_image_pages("unused kernel image (initmem)",
959 				&__init_begin, &__init_end);
960 }
961 
962 #ifdef CONFIG_BLK_DEV_INITRD
963 void __init free_initrd_mem(unsigned long start, unsigned long end)
964 {
965 	/*
966 	 * end could be not aligned, and We can not align that,
967 	 * decompressor could be confused by aligned initrd_end
968 	 * We already reserve the end partial page before in
969 	 *   - i386_start_kernel()
970 	 *   - x86_64_start_kernel()
971 	 *   - relocate_initrd()
972 	 * So here We can do PAGE_ALIGN() safely to get partial page to be freed
973 	 */
974 	free_init_pages("initrd", start, PAGE_ALIGN(end));
975 }
976 #endif
977 
978 /*
979  * Calculate the precise size of the DMA zone (first 16 MB of RAM),
980  * and pass it to the MM layer - to help it set zone watermarks more
981  * accurately.
982  *
983  * Done on 64-bit systems only for the time being, although 32-bit systems
984  * might benefit from this as well.
985  */
986 void __init memblock_find_dma_reserve(void)
987 {
988 #ifdef CONFIG_X86_64
989 	u64 nr_pages = 0, nr_free_pages = 0;
990 	unsigned long start_pfn, end_pfn;
991 	phys_addr_t start_addr, end_addr;
992 	int i;
993 	u64 u;
994 
995 	/*
996 	 * Iterate over all memory ranges (free and reserved ones alike),
997 	 * to calculate the total number of pages in the first 16 MB of RAM:
998 	 */
999 	nr_pages = 0;
1000 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
1001 		start_pfn = min(start_pfn, MAX_DMA_PFN);
1002 		end_pfn   = min(end_pfn,   MAX_DMA_PFN);
1003 
1004 		nr_pages += end_pfn - start_pfn;
1005 	}
1006 
1007 	/*
1008 	 * Iterate over free memory ranges to calculate the number of free
1009 	 * pages in the DMA zone, while not counting potential partial
1010 	 * pages at the beginning or the end of the range:
1011 	 */
1012 	nr_free_pages = 0;
1013 	for_each_free_mem_range(u, NUMA_NO_NODE, MEMBLOCK_NONE, &start_addr, &end_addr, NULL) {
1014 		start_pfn = min_t(unsigned long, PFN_UP(start_addr), MAX_DMA_PFN);
1015 		end_pfn   = min_t(unsigned long, PFN_DOWN(end_addr), MAX_DMA_PFN);
1016 
1017 		if (start_pfn < end_pfn)
1018 			nr_free_pages += end_pfn - start_pfn;
1019 	}
1020 
1021 	set_dma_reserve(nr_pages - nr_free_pages);
1022 #endif
1023 }
1024 
1025 void __init zone_sizes_init(void)
1026 {
1027 	unsigned long max_zone_pfns[MAX_NR_ZONES];
1028 
1029 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1030 
1031 #ifdef CONFIG_ZONE_DMA
1032 	max_zone_pfns[ZONE_DMA]		= min(MAX_DMA_PFN, max_low_pfn);
1033 #endif
1034 #ifdef CONFIG_ZONE_DMA32
1035 	max_zone_pfns[ZONE_DMA32]	= min(MAX_DMA32_PFN, max_low_pfn);
1036 #endif
1037 	max_zone_pfns[ZONE_NORMAL]	= max_low_pfn;
1038 #ifdef CONFIG_HIGHMEM
1039 	max_zone_pfns[ZONE_HIGHMEM]	= max_pfn;
1040 #endif
1041 
1042 	free_area_init(max_zone_pfns);
1043 }
1044 
1045 __visible DEFINE_PER_CPU_ALIGNED(struct tlb_state, cpu_tlbstate) = {
1046 	.loaded_mm = &init_mm,
1047 	.next_asid = 1,
1048 	.cr4 = ~0UL,	/* fail hard if we screw up cr4 shadow initialization */
1049 };
1050 
1051 #ifdef CONFIG_ADDRESS_MASKING
1052 DEFINE_PER_CPU(u64, tlbstate_untag_mask);
1053 EXPORT_PER_CPU_SYMBOL(tlbstate_untag_mask);
1054 #endif
1055 
1056 void update_cache_mode_entry(unsigned entry, enum page_cache_mode cache)
1057 {
1058 	/* entry 0 MUST be WB (hardwired to speed up translations) */
1059 	BUG_ON(!entry && cache != _PAGE_CACHE_MODE_WB);
1060 
1061 	__cachemode2pte_tbl[cache] = __cm_idx2pte(entry);
1062 	__pte2cachemode_tbl[entry] = cache;
1063 }
1064 
1065 #ifdef CONFIG_SWAP
1066 unsigned long arch_max_swapfile_size(void)
1067 {
1068 	unsigned long pages;
1069 
1070 	pages = generic_max_swapfile_size();
1071 
1072 	if (boot_cpu_has_bug(X86_BUG_L1TF) && l1tf_mitigation != L1TF_MITIGATION_OFF) {
1073 		/* Limit the swap file size to MAX_PA/2 for L1TF workaround */
1074 		unsigned long long l1tf_limit = l1tf_pfn_limit();
1075 		/*
1076 		 * We encode swap offsets also with 3 bits below those for pfn
1077 		 * which makes the usable limit higher.
1078 		 */
1079 #if CONFIG_PGTABLE_LEVELS > 2
1080 		l1tf_limit <<= PAGE_SHIFT - SWP_OFFSET_FIRST_BIT;
1081 #endif
1082 		pages = min_t(unsigned long long, l1tf_limit, pages);
1083 	}
1084 	return pages;
1085 }
1086 #endif
1087