xref: /openbmc/linux/arch/x86/mm/init.c (revision 78bb17f7)
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 
29 /*
30  * We need to define the tracepoints somewhere, and tlb.c
31  * is only compied when SMP=y.
32  */
33 #define CREATE_TRACE_POINTS
34 #include <trace/events/tlb.h>
35 
36 #include "mm_internal.h"
37 
38 /*
39  * Tables translating between page_cache_type_t and pte encoding.
40  *
41  * The default values are defined statically as minimal supported mode;
42  * WC and WT fall back to UC-.  pat_init() updates these values to support
43  * more cache modes, WC and WT, when it is safe to do so.  See pat_init()
44  * for the details.  Note, __early_ioremap() used during early boot-time
45  * takes pgprot_t (pte encoding) and does not use these tables.
46  *
47  *   Index into __cachemode2pte_tbl[] is the cachemode.
48  *
49  *   Index into __pte2cachemode_tbl[] are the caching attribute bits of the pte
50  *   (_PAGE_PWT, _PAGE_PCD, _PAGE_PAT) at index bit positions 0, 1, 2.
51  */
52 uint16_t __cachemode2pte_tbl[_PAGE_CACHE_MODE_NUM] = {
53 	[_PAGE_CACHE_MODE_WB      ]	= 0         | 0        ,
54 	[_PAGE_CACHE_MODE_WC      ]	= 0         | _PAGE_PCD,
55 	[_PAGE_CACHE_MODE_UC_MINUS]	= 0         | _PAGE_PCD,
56 	[_PAGE_CACHE_MODE_UC      ]	= _PAGE_PWT | _PAGE_PCD,
57 	[_PAGE_CACHE_MODE_WT      ]	= 0         | _PAGE_PCD,
58 	[_PAGE_CACHE_MODE_WP      ]	= 0         | _PAGE_PCD,
59 };
60 EXPORT_SYMBOL(__cachemode2pte_tbl);
61 
62 uint8_t __pte2cachemode_tbl[8] = {
63 	[__pte2cm_idx( 0        | 0         | 0        )] = _PAGE_CACHE_MODE_WB,
64 	[__pte2cm_idx(_PAGE_PWT | 0         | 0        )] = _PAGE_CACHE_MODE_UC_MINUS,
65 	[__pte2cm_idx( 0        | _PAGE_PCD | 0        )] = _PAGE_CACHE_MODE_UC_MINUS,
66 	[__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | 0        )] = _PAGE_CACHE_MODE_UC,
67 	[__pte2cm_idx( 0        | 0         | _PAGE_PAT)] = _PAGE_CACHE_MODE_WB,
68 	[__pte2cm_idx(_PAGE_PWT | 0         | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
69 	[__pte2cm_idx(0         | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
70 	[__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC,
71 };
72 EXPORT_SYMBOL(__pte2cachemode_tbl);
73 
74 static unsigned long __initdata pgt_buf_start;
75 static unsigned long __initdata pgt_buf_end;
76 static unsigned long __initdata pgt_buf_top;
77 
78 static unsigned long min_pfn_mapped;
79 
80 static bool __initdata can_use_brk_pgt = true;
81 
82 /*
83  * Pages returned are already directly mapped.
84  *
85  * Changing that is likely to break Xen, see commit:
86  *
87  *    279b706 x86,xen: introduce x86_init.mapping.pagetable_reserve
88  *
89  * for detailed information.
90  */
91 __ref void *alloc_low_pages(unsigned int num)
92 {
93 	unsigned long pfn;
94 	int i;
95 
96 	if (after_bootmem) {
97 		unsigned int order;
98 
99 		order = get_order((unsigned long)num << PAGE_SHIFT);
100 		return (void *)__get_free_pages(GFP_ATOMIC | __GFP_ZERO, order);
101 	}
102 
103 	if ((pgt_buf_end + num) > pgt_buf_top || !can_use_brk_pgt) {
104 		unsigned long ret = 0;
105 
106 		if (min_pfn_mapped < max_pfn_mapped) {
107 			ret = memblock_find_in_range(
108 					min_pfn_mapped << PAGE_SHIFT,
109 					max_pfn_mapped << PAGE_SHIFT,
110 					PAGE_SIZE * num , PAGE_SIZE);
111 		}
112 		if (ret)
113 			memblock_reserve(ret, PAGE_SIZE * num);
114 		else if (can_use_brk_pgt)
115 			ret = __pa(extend_brk(PAGE_SIZE * num, PAGE_SIZE));
116 
117 		if (!ret)
118 			panic("alloc_low_pages: can not alloc memory");
119 
120 		pfn = ret >> PAGE_SHIFT;
121 	} else {
122 		pfn = pgt_buf_end;
123 		pgt_buf_end += num;
124 		printk(KERN_DEBUG "BRK [%#010lx, %#010lx] PGTABLE\n",
125 			pfn << PAGE_SHIFT, (pgt_buf_end << PAGE_SHIFT) - 1);
126 	}
127 
128 	for (i = 0; i < num; i++) {
129 		void *adr;
130 
131 		adr = __va((pfn + i) << PAGE_SHIFT);
132 		clear_page(adr);
133 	}
134 
135 	return __va(pfn << PAGE_SHIFT);
136 }
137 
138 /*
139  * By default need 3 4k for initial PMD_SIZE,  3 4k for 0-ISA_END_ADDRESS.
140  * With KASLR memory randomization, depending on the machine e820 memory
141  * and the PUD alignment. We may need twice more pages when KASLR memory
142  * randomization is enabled.
143  */
144 #ifndef CONFIG_RANDOMIZE_MEMORY
145 #define INIT_PGD_PAGE_COUNT      6
146 #else
147 #define INIT_PGD_PAGE_COUNT      12
148 #endif
149 #define INIT_PGT_BUF_SIZE	(INIT_PGD_PAGE_COUNT * PAGE_SIZE)
150 RESERVE_BRK(early_pgt_alloc, INIT_PGT_BUF_SIZE);
151 void  __init early_alloc_pgt_buf(void)
152 {
153 	unsigned long tables = INIT_PGT_BUF_SIZE;
154 	phys_addr_t base;
155 
156 	base = __pa(extend_brk(tables, PAGE_SIZE));
157 
158 	pgt_buf_start = base >> PAGE_SHIFT;
159 	pgt_buf_end = pgt_buf_start;
160 	pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
161 }
162 
163 int after_bootmem;
164 
165 early_param_on_off("gbpages", "nogbpages", direct_gbpages, CONFIG_X86_DIRECT_GBPAGES);
166 
167 struct map_range {
168 	unsigned long start;
169 	unsigned long end;
170 	unsigned page_size_mask;
171 };
172 
173 static int page_size_mask;
174 
175 static void __init probe_page_size_mask(void)
176 {
177 	/*
178 	 * For pagealloc debugging, identity mapping will use small pages.
179 	 * This will simplify cpa(), which otherwise needs to support splitting
180 	 * large pages into small in interrupt context, etc.
181 	 */
182 	if (boot_cpu_has(X86_FEATURE_PSE) && !debug_pagealloc_enabled())
183 		page_size_mask |= 1 << PG_LEVEL_2M;
184 	else
185 		direct_gbpages = 0;
186 
187 	/* Enable PSE if available */
188 	if (boot_cpu_has(X86_FEATURE_PSE))
189 		cr4_set_bits_and_update_boot(X86_CR4_PSE);
190 
191 	/* Enable PGE if available */
192 	__supported_pte_mask &= ~_PAGE_GLOBAL;
193 	if (boot_cpu_has(X86_FEATURE_PGE)) {
194 		cr4_set_bits_and_update_boot(X86_CR4_PGE);
195 		__supported_pte_mask |= _PAGE_GLOBAL;
196 	}
197 
198 	/* By the default is everything supported: */
199 	__default_kernel_pte_mask = __supported_pte_mask;
200 	/* Except when with PTI where the kernel is mostly non-Global: */
201 	if (cpu_feature_enabled(X86_FEATURE_PTI))
202 		__default_kernel_pte_mask &= ~_PAGE_GLOBAL;
203 
204 	/* Enable 1 GB linear kernel mappings if available: */
205 	if (direct_gbpages && boot_cpu_has(X86_FEATURE_GBPAGES)) {
206 		printk(KERN_INFO "Using GB pages for direct mapping\n");
207 		page_size_mask |= 1 << PG_LEVEL_1G;
208 	} else {
209 		direct_gbpages = 0;
210 	}
211 }
212 
213 static void setup_pcid(void)
214 {
215 	if (!IS_ENABLED(CONFIG_X86_64))
216 		return;
217 
218 	if (!boot_cpu_has(X86_FEATURE_PCID))
219 		return;
220 
221 	if (boot_cpu_has(X86_FEATURE_PGE)) {
222 		/*
223 		 * This can't be cr4_set_bits_and_update_boot() -- the
224 		 * trampoline code can't handle CR4.PCIDE and it wouldn't
225 		 * do any good anyway.  Despite the name,
226 		 * cr4_set_bits_and_update_boot() doesn't actually cause
227 		 * the bits in question to remain set all the way through
228 		 * the secondary boot asm.
229 		 *
230 		 * Instead, we brute-force it and set CR4.PCIDE manually in
231 		 * start_secondary().
232 		 */
233 		cr4_set_bits(X86_CR4_PCIDE);
234 
235 		/*
236 		 * INVPCID's single-context modes (2/3) only work if we set
237 		 * X86_CR4_PCIDE, *and* we INVPCID support.  It's unusable
238 		 * on systems that have X86_CR4_PCIDE clear, or that have
239 		 * no INVPCID support at all.
240 		 */
241 		if (boot_cpu_has(X86_FEATURE_INVPCID))
242 			setup_force_cpu_cap(X86_FEATURE_INVPCID_SINGLE);
243 	} else {
244 		/*
245 		 * flush_tlb_all(), as currently implemented, won't work if
246 		 * PCID is on but PGE is not.  Since that combination
247 		 * doesn't exist on real hardware, there's no reason to try
248 		 * to fully support it, but it's polite to avoid corrupting
249 		 * data if we're on an improperly configured VM.
250 		 */
251 		setup_clear_cpu_cap(X86_FEATURE_PCID);
252 	}
253 }
254 
255 #ifdef CONFIG_X86_32
256 #define NR_RANGE_MR 3
257 #else /* CONFIG_X86_64 */
258 #define NR_RANGE_MR 5
259 #endif
260 
261 static int __meminit save_mr(struct map_range *mr, int nr_range,
262 			     unsigned long start_pfn, unsigned long end_pfn,
263 			     unsigned long page_size_mask)
264 {
265 	if (start_pfn < end_pfn) {
266 		if (nr_range >= NR_RANGE_MR)
267 			panic("run out of range for init_memory_mapping\n");
268 		mr[nr_range].start = start_pfn<<PAGE_SHIFT;
269 		mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
270 		mr[nr_range].page_size_mask = page_size_mask;
271 		nr_range++;
272 	}
273 
274 	return nr_range;
275 }
276 
277 /*
278  * adjust the page_size_mask for small range to go with
279  *	big page size instead small one if nearby are ram too.
280  */
281 static void __ref adjust_range_page_size_mask(struct map_range *mr,
282 							 int nr_range)
283 {
284 	int i;
285 
286 	for (i = 0; i < nr_range; i++) {
287 		if ((page_size_mask & (1<<PG_LEVEL_2M)) &&
288 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) {
289 			unsigned long start = round_down(mr[i].start, PMD_SIZE);
290 			unsigned long end = round_up(mr[i].end, PMD_SIZE);
291 
292 #ifdef CONFIG_X86_32
293 			if ((end >> PAGE_SHIFT) > max_low_pfn)
294 				continue;
295 #endif
296 
297 			if (memblock_is_region_memory(start, end - start))
298 				mr[i].page_size_mask |= 1<<PG_LEVEL_2M;
299 		}
300 		if ((page_size_mask & (1<<PG_LEVEL_1G)) &&
301 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) {
302 			unsigned long start = round_down(mr[i].start, PUD_SIZE);
303 			unsigned long end = round_up(mr[i].end, PUD_SIZE);
304 
305 			if (memblock_is_region_memory(start, end - start))
306 				mr[i].page_size_mask |= 1<<PG_LEVEL_1G;
307 		}
308 	}
309 }
310 
311 static const char *page_size_string(struct map_range *mr)
312 {
313 	static const char str_1g[] = "1G";
314 	static const char str_2m[] = "2M";
315 	static const char str_4m[] = "4M";
316 	static const char str_4k[] = "4k";
317 
318 	if (mr->page_size_mask & (1<<PG_LEVEL_1G))
319 		return str_1g;
320 	/*
321 	 * 32-bit without PAE has a 4M large page size.
322 	 * PG_LEVEL_2M is misnamed, but we can at least
323 	 * print out the right size in the string.
324 	 */
325 	if (IS_ENABLED(CONFIG_X86_32) &&
326 	    !IS_ENABLED(CONFIG_X86_PAE) &&
327 	    mr->page_size_mask & (1<<PG_LEVEL_2M))
328 		return str_4m;
329 
330 	if (mr->page_size_mask & (1<<PG_LEVEL_2M))
331 		return str_2m;
332 
333 	return str_4k;
334 }
335 
336 static int __meminit split_mem_range(struct map_range *mr, int nr_range,
337 				     unsigned long start,
338 				     unsigned long end)
339 {
340 	unsigned long start_pfn, end_pfn, limit_pfn;
341 	unsigned long pfn;
342 	int i;
343 
344 	limit_pfn = PFN_DOWN(end);
345 
346 	/* head if not big page alignment ? */
347 	pfn = start_pfn = PFN_DOWN(start);
348 #ifdef CONFIG_X86_32
349 	/*
350 	 * Don't use a large page for the first 2/4MB of memory
351 	 * because there are often fixed size MTRRs in there
352 	 * and overlapping MTRRs into large pages can cause
353 	 * slowdowns.
354 	 */
355 	if (pfn == 0)
356 		end_pfn = PFN_DOWN(PMD_SIZE);
357 	else
358 		end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
359 #else /* CONFIG_X86_64 */
360 	end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
361 #endif
362 	if (end_pfn > limit_pfn)
363 		end_pfn = limit_pfn;
364 	if (start_pfn < end_pfn) {
365 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
366 		pfn = end_pfn;
367 	}
368 
369 	/* big page (2M) range */
370 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
371 #ifdef CONFIG_X86_32
372 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
373 #else /* CONFIG_X86_64 */
374 	end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
375 	if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE)))
376 		end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
377 #endif
378 
379 	if (start_pfn < end_pfn) {
380 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
381 				page_size_mask & (1<<PG_LEVEL_2M));
382 		pfn = end_pfn;
383 	}
384 
385 #ifdef CONFIG_X86_64
386 	/* big page (1G) range */
387 	start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
388 	end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE));
389 	if (start_pfn < end_pfn) {
390 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
391 				page_size_mask &
392 				 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
393 		pfn = end_pfn;
394 	}
395 
396 	/* tail is not big page (1G) alignment */
397 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
398 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
399 	if (start_pfn < end_pfn) {
400 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
401 				page_size_mask & (1<<PG_LEVEL_2M));
402 		pfn = end_pfn;
403 	}
404 #endif
405 
406 	/* tail is not big page (2M) alignment */
407 	start_pfn = pfn;
408 	end_pfn = limit_pfn;
409 	nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
410 
411 	if (!after_bootmem)
412 		adjust_range_page_size_mask(mr, nr_range);
413 
414 	/* try to merge same page size and continuous */
415 	for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
416 		unsigned long old_start;
417 		if (mr[i].end != mr[i+1].start ||
418 		    mr[i].page_size_mask != mr[i+1].page_size_mask)
419 			continue;
420 		/* move it */
421 		old_start = mr[i].start;
422 		memmove(&mr[i], &mr[i+1],
423 			(nr_range - 1 - i) * sizeof(struct map_range));
424 		mr[i--].start = old_start;
425 		nr_range--;
426 	}
427 
428 	for (i = 0; i < nr_range; i++)
429 		pr_debug(" [mem %#010lx-%#010lx] page %s\n",
430 				mr[i].start, mr[i].end - 1,
431 				page_size_string(&mr[i]));
432 
433 	return nr_range;
434 }
435 
436 struct range pfn_mapped[E820_MAX_ENTRIES];
437 int nr_pfn_mapped;
438 
439 static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn)
440 {
441 	nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_MAX_ENTRIES,
442 					     nr_pfn_mapped, start_pfn, end_pfn);
443 	nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_MAX_ENTRIES);
444 
445 	max_pfn_mapped = max(max_pfn_mapped, end_pfn);
446 
447 	if (start_pfn < (1UL<<(32-PAGE_SHIFT)))
448 		max_low_pfn_mapped = max(max_low_pfn_mapped,
449 					 min(end_pfn, 1UL<<(32-PAGE_SHIFT)));
450 }
451 
452 bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn)
453 {
454 	int i;
455 
456 	for (i = 0; i < nr_pfn_mapped; i++)
457 		if ((start_pfn >= pfn_mapped[i].start) &&
458 		    (end_pfn <= pfn_mapped[i].end))
459 			return true;
460 
461 	return false;
462 }
463 
464 /*
465  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
466  * This runs before bootmem is initialized and gets pages directly from
467  * the physical memory. To access them they are temporarily mapped.
468  */
469 unsigned long __ref init_memory_mapping(unsigned long start,
470 					unsigned long end, pgprot_t prot)
471 {
472 	struct map_range mr[NR_RANGE_MR];
473 	unsigned long ret = 0;
474 	int nr_range, i;
475 
476 	pr_debug("init_memory_mapping: [mem %#010lx-%#010lx]\n",
477 	       start, end - 1);
478 
479 	memset(mr, 0, sizeof(mr));
480 	nr_range = split_mem_range(mr, 0, start, end);
481 
482 	for (i = 0; i < nr_range; i++)
483 		ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
484 						   mr[i].page_size_mask,
485 						   prot);
486 
487 	add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT);
488 
489 	return ret >> PAGE_SHIFT;
490 }
491 
492 /*
493  * We need to iterate through the E820 memory map and create direct mappings
494  * for only E820_TYPE_RAM and E820_KERN_RESERVED regions. We cannot simply
495  * create direct mappings for all pfns from [0 to max_low_pfn) and
496  * [4GB to max_pfn) because of possible memory holes in high addresses
497  * that cannot be marked as UC by fixed/variable range MTRRs.
498  * Depending on the alignment of E820 ranges, this may possibly result
499  * in using smaller size (i.e. 4K instead of 2M or 1G) page tables.
500  *
501  * init_mem_mapping() calls init_range_memory_mapping() with big range.
502  * That range would have hole in the middle or ends, and only ram parts
503  * will be mapped in init_range_memory_mapping().
504  */
505 static unsigned long __init init_range_memory_mapping(
506 					   unsigned long r_start,
507 					   unsigned long r_end)
508 {
509 	unsigned long start_pfn, end_pfn;
510 	unsigned long mapped_ram_size = 0;
511 	int i;
512 
513 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
514 		u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end);
515 		u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end);
516 		if (start >= end)
517 			continue;
518 
519 		/*
520 		 * if it is overlapping with brk pgt, we need to
521 		 * alloc pgt buf from memblock instead.
522 		 */
523 		can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >=
524 				    min(end, (u64)pgt_buf_top<<PAGE_SHIFT);
525 		init_memory_mapping(start, end, PAGE_KERNEL);
526 		mapped_ram_size += end - start;
527 		can_use_brk_pgt = true;
528 	}
529 
530 	return mapped_ram_size;
531 }
532 
533 static unsigned long __init get_new_step_size(unsigned long step_size)
534 {
535 	/*
536 	 * Initial mapped size is PMD_SIZE (2M).
537 	 * We can not set step_size to be PUD_SIZE (1G) yet.
538 	 * In worse case, when we cross the 1G boundary, and
539 	 * PG_LEVEL_2M is not set, we will need 1+1+512 pages (2M + 8k)
540 	 * to map 1G range with PTE. Hence we use one less than the
541 	 * difference of page table level shifts.
542 	 *
543 	 * Don't need to worry about overflow in the top-down case, on 32bit,
544 	 * when step_size is 0, round_down() returns 0 for start, and that
545 	 * turns it into 0x100000000ULL.
546 	 * In the bottom-up case, round_up(x, 0) returns 0 though too, which
547 	 * needs to be taken into consideration by the code below.
548 	 */
549 	return step_size << (PMD_SHIFT - PAGE_SHIFT - 1);
550 }
551 
552 /**
553  * memory_map_top_down - Map [map_start, map_end) top down
554  * @map_start: start address of the target memory range
555  * @map_end: end address of the target memory range
556  *
557  * This function will setup direct mapping for memory range
558  * [map_start, map_end) in top-down. That said, the page tables
559  * will be allocated at the end of the memory, and we map the
560  * memory in top-down.
561  */
562 static void __init memory_map_top_down(unsigned long map_start,
563 				       unsigned long map_end)
564 {
565 	unsigned long real_end, start, last_start;
566 	unsigned long step_size;
567 	unsigned long addr;
568 	unsigned long mapped_ram_size = 0;
569 
570 	/* xen has big range in reserved near end of ram, skip it at first.*/
571 	addr = memblock_find_in_range(map_start, map_end, PMD_SIZE, PMD_SIZE);
572 	real_end = addr + PMD_SIZE;
573 
574 	/* step_size need to be small so pgt_buf from BRK could cover it */
575 	step_size = PMD_SIZE;
576 	max_pfn_mapped = 0; /* will get exact value next */
577 	min_pfn_mapped = real_end >> PAGE_SHIFT;
578 	last_start = start = real_end;
579 
580 	/*
581 	 * We start from the top (end of memory) and go to the bottom.
582 	 * The memblock_find_in_range() gets us a block of RAM from the
583 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
584 	 * for page table.
585 	 */
586 	while (last_start > map_start) {
587 		if (last_start > step_size) {
588 			start = round_down(last_start - 1, step_size);
589 			if (start < map_start)
590 				start = map_start;
591 		} else
592 			start = map_start;
593 		mapped_ram_size += init_range_memory_mapping(start,
594 							last_start);
595 		last_start = start;
596 		min_pfn_mapped = last_start >> PAGE_SHIFT;
597 		if (mapped_ram_size >= step_size)
598 			step_size = get_new_step_size(step_size);
599 	}
600 
601 	if (real_end < map_end)
602 		init_range_memory_mapping(real_end, map_end);
603 }
604 
605 /**
606  * memory_map_bottom_up - Map [map_start, map_end) bottom up
607  * @map_start: start address of the target memory range
608  * @map_end: end address of the target memory range
609  *
610  * This function will setup direct mapping for memory range
611  * [map_start, map_end) in bottom-up. Since we have limited the
612  * bottom-up allocation above the kernel, the page tables will
613  * be allocated just above the kernel and we map the memory
614  * in [map_start, map_end) in bottom-up.
615  */
616 static void __init memory_map_bottom_up(unsigned long map_start,
617 					unsigned long map_end)
618 {
619 	unsigned long next, start;
620 	unsigned long mapped_ram_size = 0;
621 	/* step_size need to be small so pgt_buf from BRK could cover it */
622 	unsigned long step_size = PMD_SIZE;
623 
624 	start = map_start;
625 	min_pfn_mapped = start >> PAGE_SHIFT;
626 
627 	/*
628 	 * We start from the bottom (@map_start) and go to the top (@map_end).
629 	 * The memblock_find_in_range() gets us a block of RAM from the
630 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
631 	 * for page table.
632 	 */
633 	while (start < map_end) {
634 		if (step_size && map_end - start > step_size) {
635 			next = round_up(start + 1, step_size);
636 			if (next > map_end)
637 				next = map_end;
638 		} else {
639 			next = map_end;
640 		}
641 
642 		mapped_ram_size += init_range_memory_mapping(start, next);
643 		start = next;
644 
645 		if (mapped_ram_size >= step_size)
646 			step_size = get_new_step_size(step_size);
647 	}
648 }
649 
650 void __init init_mem_mapping(void)
651 {
652 	unsigned long end;
653 
654 	pti_check_boottime_disable();
655 	probe_page_size_mask();
656 	setup_pcid();
657 
658 #ifdef CONFIG_X86_64
659 	end = max_pfn << PAGE_SHIFT;
660 #else
661 	end = max_low_pfn << PAGE_SHIFT;
662 #endif
663 
664 	/* the ISA range is always mapped regardless of memory holes */
665 	init_memory_mapping(0, ISA_END_ADDRESS, PAGE_KERNEL);
666 
667 	/* Init the trampoline, possibly with KASLR memory offset */
668 	init_trampoline();
669 
670 	/*
671 	 * If the allocation is in bottom-up direction, we setup direct mapping
672 	 * in bottom-up, otherwise we setup direct mapping in top-down.
673 	 */
674 	if (memblock_bottom_up()) {
675 		unsigned long kernel_end = __pa_symbol(_end);
676 
677 		/*
678 		 * we need two separate calls here. This is because we want to
679 		 * allocate page tables above the kernel. So we first map
680 		 * [kernel_end, end) to make memory above the kernel be mapped
681 		 * as soon as possible. And then use page tables allocated above
682 		 * the kernel to map [ISA_END_ADDRESS, kernel_end).
683 		 */
684 		memory_map_bottom_up(kernel_end, end);
685 		memory_map_bottom_up(ISA_END_ADDRESS, kernel_end);
686 	} else {
687 		memory_map_top_down(ISA_END_ADDRESS, end);
688 	}
689 
690 #ifdef CONFIG_X86_64
691 	if (max_pfn > max_low_pfn) {
692 		/* can we preseve max_low_pfn ?*/
693 		max_low_pfn = max_pfn;
694 	}
695 #else
696 	early_ioremap_page_table_range_init();
697 #endif
698 
699 	load_cr3(swapper_pg_dir);
700 	__flush_tlb_all();
701 
702 	x86_init.hyper.init_mem_mapping();
703 
704 	early_memtest(0, max_pfn_mapped << PAGE_SHIFT);
705 }
706 
707 /*
708  * Initialize an mm_struct to be used during poking and a pointer to be used
709  * during patching.
710  */
711 void __init poking_init(void)
712 {
713 	spinlock_t *ptl;
714 	pte_t *ptep;
715 
716 	poking_mm = copy_init_mm();
717 	BUG_ON(!poking_mm);
718 
719 	/*
720 	 * Randomize the poking address, but make sure that the following page
721 	 * will be mapped at the same PMD. We need 2 pages, so find space for 3,
722 	 * and adjust the address if the PMD ends after the first one.
723 	 */
724 	poking_addr = TASK_UNMAPPED_BASE;
725 	if (IS_ENABLED(CONFIG_RANDOMIZE_BASE))
726 		poking_addr += (kaslr_get_random_long("Poking") & PAGE_MASK) %
727 			(TASK_SIZE - TASK_UNMAPPED_BASE - 3 * PAGE_SIZE);
728 
729 	if (((poking_addr + PAGE_SIZE) & ~PMD_MASK) == 0)
730 		poking_addr += PAGE_SIZE;
731 
732 	/*
733 	 * We need to trigger the allocation of the page-tables that will be
734 	 * needed for poking now. Later, poking may be performed in an atomic
735 	 * section, which might cause allocation to fail.
736 	 */
737 	ptep = get_locked_pte(poking_mm, poking_addr, &ptl);
738 	BUG_ON(!ptep);
739 	pte_unmap_unlock(ptep, ptl);
740 }
741 
742 /*
743  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
744  * is valid. The argument is a physical page number.
745  *
746  * On x86, access has to be given to the first megabyte of RAM because that
747  * area traditionally contains BIOS code and data regions used by X, dosemu,
748  * and similar apps. Since they map the entire memory range, the whole range
749  * must be allowed (for mapping), but any areas that would otherwise be
750  * disallowed are flagged as being "zero filled" instead of rejected.
751  * Access has to be given to non-kernel-ram areas as well, these contain the
752  * PCI mmio resources as well as potential bios/acpi data regions.
753  */
754 int devmem_is_allowed(unsigned long pagenr)
755 {
756 	if (region_intersects(PFN_PHYS(pagenr), PAGE_SIZE,
757 				IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE)
758 			!= REGION_DISJOINT) {
759 		/*
760 		 * For disallowed memory regions in the low 1MB range,
761 		 * request that the page be shown as all zeros.
762 		 */
763 		if (pagenr < 256)
764 			return 2;
765 
766 		return 0;
767 	}
768 
769 	/*
770 	 * This must follow RAM test, since System RAM is considered a
771 	 * restricted resource under CONFIG_STRICT_IOMEM.
772 	 */
773 	if (iomem_is_exclusive(pagenr << PAGE_SHIFT)) {
774 		/* Low 1MB bypasses iomem restrictions. */
775 		if (pagenr < 256)
776 			return 1;
777 
778 		return 0;
779 	}
780 
781 	return 1;
782 }
783 
784 void free_init_pages(const char *what, unsigned long begin, unsigned long end)
785 {
786 	unsigned long begin_aligned, end_aligned;
787 
788 	/* Make sure boundaries are page aligned */
789 	begin_aligned = PAGE_ALIGN(begin);
790 	end_aligned   = end & PAGE_MASK;
791 
792 	if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
793 		begin = begin_aligned;
794 		end   = end_aligned;
795 	}
796 
797 	if (begin >= end)
798 		return;
799 
800 	/*
801 	 * If debugging page accesses then do not free this memory but
802 	 * mark them not present - any buggy init-section access will
803 	 * create a kernel page fault:
804 	 */
805 	if (debug_pagealloc_enabled()) {
806 		pr_info("debug: unmapping init [mem %#010lx-%#010lx]\n",
807 			begin, end - 1);
808 		/*
809 		 * Inform kmemleak about the hole in the memory since the
810 		 * corresponding pages will be unmapped.
811 		 */
812 		kmemleak_free_part((void *)begin, end - begin);
813 		set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
814 	} else {
815 		/*
816 		 * We just marked the kernel text read only above, now that
817 		 * we are going to free part of that, we need to make that
818 		 * writeable and non-executable first.
819 		 */
820 		set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
821 		set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
822 
823 		free_reserved_area((void *)begin, (void *)end,
824 				   POISON_FREE_INITMEM, what);
825 	}
826 }
827 
828 /*
829  * begin/end can be in the direct map or the "high kernel mapping"
830  * used for the kernel image only.  free_init_pages() will do the
831  * right thing for either kind of address.
832  */
833 void free_kernel_image_pages(const char *what, void *begin, void *end)
834 {
835 	unsigned long begin_ul = (unsigned long)begin;
836 	unsigned long end_ul = (unsigned long)end;
837 	unsigned long len_pages = (end_ul - begin_ul) >> PAGE_SHIFT;
838 
839 	free_init_pages(what, begin_ul, end_ul);
840 
841 	/*
842 	 * PTI maps some of the kernel into userspace.  For performance,
843 	 * this includes some kernel areas that do not contain secrets.
844 	 * Those areas might be adjacent to the parts of the kernel image
845 	 * being freed, which may contain secrets.  Remove the "high kernel
846 	 * image mapping" for these freed areas, ensuring they are not even
847 	 * potentially vulnerable to Meltdown regardless of the specific
848 	 * optimizations PTI is currently using.
849 	 *
850 	 * The "noalias" prevents unmapping the direct map alias which is
851 	 * needed to access the freed pages.
852 	 *
853 	 * This is only valid for 64bit kernels. 32bit has only one mapping
854 	 * which can't be treated in this way for obvious reasons.
855 	 */
856 	if (IS_ENABLED(CONFIG_X86_64) && cpu_feature_enabled(X86_FEATURE_PTI))
857 		set_memory_np_noalias(begin_ul, len_pages);
858 }
859 
860 void __weak mem_encrypt_free_decrypted_mem(void) { }
861 
862 void __ref free_initmem(void)
863 {
864 	e820__reallocate_tables();
865 
866 	mem_encrypt_free_decrypted_mem();
867 
868 	free_kernel_image_pages("unused kernel image (initmem)",
869 				&__init_begin, &__init_end);
870 }
871 
872 #ifdef CONFIG_BLK_DEV_INITRD
873 void __init free_initrd_mem(unsigned long start, unsigned long end)
874 {
875 	/*
876 	 * end could be not aligned, and We can not align that,
877 	 * decompresser could be confused by aligned initrd_end
878 	 * We already reserve the end partial page before in
879 	 *   - i386_start_kernel()
880 	 *   - x86_64_start_kernel()
881 	 *   - relocate_initrd()
882 	 * So here We can do PAGE_ALIGN() safely to get partial page to be freed
883 	 */
884 	free_init_pages("initrd", start, PAGE_ALIGN(end));
885 }
886 #endif
887 
888 /*
889  * Calculate the precise size of the DMA zone (first 16 MB of RAM),
890  * and pass it to the MM layer - to help it set zone watermarks more
891  * accurately.
892  *
893  * Done on 64-bit systems only for the time being, although 32-bit systems
894  * might benefit from this as well.
895  */
896 void __init memblock_find_dma_reserve(void)
897 {
898 #ifdef CONFIG_X86_64
899 	u64 nr_pages = 0, nr_free_pages = 0;
900 	unsigned long start_pfn, end_pfn;
901 	phys_addr_t start_addr, end_addr;
902 	int i;
903 	u64 u;
904 
905 	/*
906 	 * Iterate over all memory ranges (free and reserved ones alike),
907 	 * to calculate the total number of pages in the first 16 MB of RAM:
908 	 */
909 	nr_pages = 0;
910 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
911 		start_pfn = min(start_pfn, MAX_DMA_PFN);
912 		end_pfn   = min(end_pfn,   MAX_DMA_PFN);
913 
914 		nr_pages += end_pfn - start_pfn;
915 	}
916 
917 	/*
918 	 * Iterate over free memory ranges to calculate the number of free
919 	 * pages in the DMA zone, while not counting potential partial
920 	 * pages at the beginning or the end of the range:
921 	 */
922 	nr_free_pages = 0;
923 	for_each_free_mem_range(u, NUMA_NO_NODE, MEMBLOCK_NONE, &start_addr, &end_addr, NULL) {
924 		start_pfn = min_t(unsigned long, PFN_UP(start_addr), MAX_DMA_PFN);
925 		end_pfn   = min_t(unsigned long, PFN_DOWN(end_addr), MAX_DMA_PFN);
926 
927 		if (start_pfn < end_pfn)
928 			nr_free_pages += end_pfn - start_pfn;
929 	}
930 
931 	set_dma_reserve(nr_pages - nr_free_pages);
932 #endif
933 }
934 
935 void __init zone_sizes_init(void)
936 {
937 	unsigned long max_zone_pfns[MAX_NR_ZONES];
938 
939 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
940 
941 #ifdef CONFIG_ZONE_DMA
942 	max_zone_pfns[ZONE_DMA]		= min(MAX_DMA_PFN, max_low_pfn);
943 #endif
944 #ifdef CONFIG_ZONE_DMA32
945 	max_zone_pfns[ZONE_DMA32]	= min(MAX_DMA32_PFN, max_low_pfn);
946 #endif
947 	max_zone_pfns[ZONE_NORMAL]	= max_low_pfn;
948 #ifdef CONFIG_HIGHMEM
949 	max_zone_pfns[ZONE_HIGHMEM]	= max_pfn;
950 #endif
951 
952 	free_area_init_nodes(max_zone_pfns);
953 }
954 
955 __visible DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate) = {
956 	.loaded_mm = &init_mm,
957 	.next_asid = 1,
958 	.cr4 = ~0UL,	/* fail hard if we screw up cr4 shadow initialization */
959 };
960 EXPORT_PER_CPU_SYMBOL(cpu_tlbstate);
961 
962 void update_cache_mode_entry(unsigned entry, enum page_cache_mode cache)
963 {
964 	/* entry 0 MUST be WB (hardwired to speed up translations) */
965 	BUG_ON(!entry && cache != _PAGE_CACHE_MODE_WB);
966 
967 	__cachemode2pte_tbl[cache] = __cm_idx2pte(entry);
968 	__pte2cachemode_tbl[entry] = cache;
969 }
970 
971 #ifdef CONFIG_SWAP
972 unsigned long max_swapfile_size(void)
973 {
974 	unsigned long pages;
975 
976 	pages = generic_max_swapfile_size();
977 
978 	if (boot_cpu_has_bug(X86_BUG_L1TF) && l1tf_mitigation != L1TF_MITIGATION_OFF) {
979 		/* Limit the swap file size to MAX_PA/2 for L1TF workaround */
980 		unsigned long long l1tf_limit = l1tf_pfn_limit();
981 		/*
982 		 * We encode swap offsets also with 3 bits below those for pfn
983 		 * which makes the usable limit higher.
984 		 */
985 #if CONFIG_PGTABLE_LEVELS > 2
986 		l1tf_limit <<= PAGE_SHIFT - SWP_OFFSET_FIRST_BIT;
987 #endif
988 		pages = min_t(unsigned long long, l1tf_limit, pages);
989 	}
990 	return pages;
991 }
992 #endif
993