xref: /openbmc/linux/arch/x86/mm/init.c (revision 26b0a1cd)
1 #include <linux/gfp.h>
2 #include <linux/initrd.h>
3 #include <linux/ioport.h>
4 #include <linux/swap.h>
5 #include <linux/memblock.h>
6 #include <linux/swapfile.h>
7 #include <linux/swapops.h>
8 #include <linux/kmemleak.h>
9 #include <linux/sched/task.h>
10 
11 #include <asm/set_memory.h>
12 #include <asm/cpu_device_id.h>
13 #include <asm/e820/api.h>
14 #include <asm/init.h>
15 #include <asm/page.h>
16 #include <asm/page_types.h>
17 #include <asm/sections.h>
18 #include <asm/setup.h>
19 #include <asm/tlbflush.h>
20 #include <asm/tlb.h>
21 #include <asm/proto.h>
22 #include <asm/dma.h>		/* for MAX_DMA_PFN */
23 #include <asm/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 
cachemode2protval(enum page_cache_mode pcm)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  */
x86_has_pat_wp(void)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 
pgprot2cachemode(pgprot_t pgprot)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  */
alloc_low_pages(unsigned int num)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);
early_alloc_pgt_buf(void)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  */
cr4_set_bits_and_update_boot(unsigned long mask)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 
probe_page_size_mask(void)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 /*
265  * INVLPG may not properly flush Global entries
266  * on these CPUs when PCIDs are enabled.
267  */
268 static const struct x86_cpu_id invlpg_miss_ids[] = {
269 	X86_MATCH_INTEL_FAM6_MODEL(ALDERLAKE,      0),
270 	X86_MATCH_INTEL_FAM6_MODEL(ALDERLAKE_L,    0),
271 	X86_MATCH_INTEL_FAM6_MODEL(ATOM_GRACEMONT, 0),
272 	X86_MATCH_INTEL_FAM6_MODEL(RAPTORLAKE,     0),
273 	X86_MATCH_INTEL_FAM6_MODEL(RAPTORLAKE_P,   0),
274 	X86_MATCH_INTEL_FAM6_MODEL(RAPTORLAKE_S,   0),
275 	{}
276 };
277 
setup_pcid(void)278 static void setup_pcid(void)
279 {
280 	if (!IS_ENABLED(CONFIG_X86_64))
281 		return;
282 
283 	if (!boot_cpu_has(X86_FEATURE_PCID))
284 		return;
285 
286 	if (x86_match_cpu(invlpg_miss_ids)) {
287 		pr_info("Incomplete global flushes, disabling PCID");
288 		setup_clear_cpu_cap(X86_FEATURE_PCID);
289 		return;
290 	}
291 
292 	if (boot_cpu_has(X86_FEATURE_PGE)) {
293 		/*
294 		 * This can't be cr4_set_bits_and_update_boot() -- the
295 		 * trampoline code can't handle CR4.PCIDE and it wouldn't
296 		 * do any good anyway.  Despite the name,
297 		 * cr4_set_bits_and_update_boot() doesn't actually cause
298 		 * the bits in question to remain set all the way through
299 		 * the secondary boot asm.
300 		 *
301 		 * Instead, we brute-force it and set CR4.PCIDE manually in
302 		 * start_secondary().
303 		 */
304 		cr4_set_bits(X86_CR4_PCIDE);
305 	} else {
306 		/*
307 		 * flush_tlb_all(), as currently implemented, won't work if
308 		 * PCID is on but PGE is not.  Since that combination
309 		 * doesn't exist on real hardware, there's no reason to try
310 		 * to fully support it, but it's polite to avoid corrupting
311 		 * data if we're on an improperly configured VM.
312 		 */
313 		setup_clear_cpu_cap(X86_FEATURE_PCID);
314 	}
315 }
316 
317 #ifdef CONFIG_X86_32
318 #define NR_RANGE_MR 3
319 #else /* CONFIG_X86_64 */
320 #define NR_RANGE_MR 5
321 #endif
322 
save_mr(struct map_range * mr,int nr_range,unsigned long start_pfn,unsigned long end_pfn,unsigned long page_size_mask)323 static int __meminit save_mr(struct map_range *mr, int nr_range,
324 			     unsigned long start_pfn, unsigned long end_pfn,
325 			     unsigned long page_size_mask)
326 {
327 	if (start_pfn < end_pfn) {
328 		if (nr_range >= NR_RANGE_MR)
329 			panic("run out of range for init_memory_mapping\n");
330 		mr[nr_range].start = start_pfn<<PAGE_SHIFT;
331 		mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
332 		mr[nr_range].page_size_mask = page_size_mask;
333 		nr_range++;
334 	}
335 
336 	return nr_range;
337 }
338 
339 /*
340  * adjust the page_size_mask for small range to go with
341  *	big page size instead small one if nearby are ram too.
342  */
adjust_range_page_size_mask(struct map_range * mr,int nr_range)343 static void __ref adjust_range_page_size_mask(struct map_range *mr,
344 							 int nr_range)
345 {
346 	int i;
347 
348 	for (i = 0; i < nr_range; i++) {
349 		if ((page_size_mask & (1<<PG_LEVEL_2M)) &&
350 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) {
351 			unsigned long start = round_down(mr[i].start, PMD_SIZE);
352 			unsigned long end = round_up(mr[i].end, PMD_SIZE);
353 
354 #ifdef CONFIG_X86_32
355 			if ((end >> PAGE_SHIFT) > max_low_pfn)
356 				continue;
357 #endif
358 
359 			if (memblock_is_region_memory(start, end - start))
360 				mr[i].page_size_mask |= 1<<PG_LEVEL_2M;
361 		}
362 		if ((page_size_mask & (1<<PG_LEVEL_1G)) &&
363 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) {
364 			unsigned long start = round_down(mr[i].start, PUD_SIZE);
365 			unsigned long end = round_up(mr[i].end, PUD_SIZE);
366 
367 			if (memblock_is_region_memory(start, end - start))
368 				mr[i].page_size_mask |= 1<<PG_LEVEL_1G;
369 		}
370 	}
371 }
372 
page_size_string(struct map_range * mr)373 static const char *page_size_string(struct map_range *mr)
374 {
375 	static const char str_1g[] = "1G";
376 	static const char str_2m[] = "2M";
377 	static const char str_4m[] = "4M";
378 	static const char str_4k[] = "4k";
379 
380 	if (mr->page_size_mask & (1<<PG_LEVEL_1G))
381 		return str_1g;
382 	/*
383 	 * 32-bit without PAE has a 4M large page size.
384 	 * PG_LEVEL_2M is misnamed, but we can at least
385 	 * print out the right size in the string.
386 	 */
387 	if (IS_ENABLED(CONFIG_X86_32) &&
388 	    !IS_ENABLED(CONFIG_X86_PAE) &&
389 	    mr->page_size_mask & (1<<PG_LEVEL_2M))
390 		return str_4m;
391 
392 	if (mr->page_size_mask & (1<<PG_LEVEL_2M))
393 		return str_2m;
394 
395 	return str_4k;
396 }
397 
split_mem_range(struct map_range * mr,int nr_range,unsigned long start,unsigned long end)398 static int __meminit split_mem_range(struct map_range *mr, int nr_range,
399 				     unsigned long start,
400 				     unsigned long end)
401 {
402 	unsigned long start_pfn, end_pfn, limit_pfn;
403 	unsigned long pfn;
404 	int i;
405 
406 	limit_pfn = PFN_DOWN(end);
407 
408 	/* head if not big page alignment ? */
409 	pfn = start_pfn = PFN_DOWN(start);
410 #ifdef CONFIG_X86_32
411 	/*
412 	 * Don't use a large page for the first 2/4MB of memory
413 	 * because there are often fixed size MTRRs in there
414 	 * and overlapping MTRRs into large pages can cause
415 	 * slowdowns.
416 	 */
417 	if (pfn == 0)
418 		end_pfn = PFN_DOWN(PMD_SIZE);
419 	else
420 		end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
421 #else /* CONFIG_X86_64 */
422 	end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
423 #endif
424 	if (end_pfn > limit_pfn)
425 		end_pfn = limit_pfn;
426 	if (start_pfn < end_pfn) {
427 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
428 		pfn = end_pfn;
429 	}
430 
431 	/* big page (2M) range */
432 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
433 #ifdef CONFIG_X86_32
434 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
435 #else /* CONFIG_X86_64 */
436 	end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
437 	if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE)))
438 		end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
439 #endif
440 
441 	if (start_pfn < end_pfn) {
442 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
443 				page_size_mask & (1<<PG_LEVEL_2M));
444 		pfn = end_pfn;
445 	}
446 
447 #ifdef CONFIG_X86_64
448 	/* big page (1G) range */
449 	start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
450 	end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE));
451 	if (start_pfn < end_pfn) {
452 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
453 				page_size_mask &
454 				 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
455 		pfn = end_pfn;
456 	}
457 
458 	/* tail is not big page (1G) alignment */
459 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
460 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
461 	if (start_pfn < end_pfn) {
462 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
463 				page_size_mask & (1<<PG_LEVEL_2M));
464 		pfn = end_pfn;
465 	}
466 #endif
467 
468 	/* tail is not big page (2M) alignment */
469 	start_pfn = pfn;
470 	end_pfn = limit_pfn;
471 	nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
472 
473 	if (!after_bootmem)
474 		adjust_range_page_size_mask(mr, nr_range);
475 
476 	/* try to merge same page size and continuous */
477 	for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
478 		unsigned long old_start;
479 		if (mr[i].end != mr[i+1].start ||
480 		    mr[i].page_size_mask != mr[i+1].page_size_mask)
481 			continue;
482 		/* move it */
483 		old_start = mr[i].start;
484 		memmove(&mr[i], &mr[i+1],
485 			(nr_range - 1 - i) * sizeof(struct map_range));
486 		mr[i--].start = old_start;
487 		nr_range--;
488 	}
489 
490 	for (i = 0; i < nr_range; i++)
491 		pr_debug(" [mem %#010lx-%#010lx] page %s\n",
492 				mr[i].start, mr[i].end - 1,
493 				page_size_string(&mr[i]));
494 
495 	return nr_range;
496 }
497 
498 struct range pfn_mapped[E820_MAX_ENTRIES];
499 int nr_pfn_mapped;
500 
add_pfn_range_mapped(unsigned long start_pfn,unsigned long end_pfn)501 static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn)
502 {
503 	nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_MAX_ENTRIES,
504 					     nr_pfn_mapped, start_pfn, end_pfn);
505 	nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_MAX_ENTRIES);
506 
507 	max_pfn_mapped = max(max_pfn_mapped, end_pfn);
508 
509 	if (start_pfn < (1UL<<(32-PAGE_SHIFT)))
510 		max_low_pfn_mapped = max(max_low_pfn_mapped,
511 					 min(end_pfn, 1UL<<(32-PAGE_SHIFT)));
512 }
513 
pfn_range_is_mapped(unsigned long start_pfn,unsigned long end_pfn)514 bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn)
515 {
516 	int i;
517 
518 	for (i = 0; i < nr_pfn_mapped; i++)
519 		if ((start_pfn >= pfn_mapped[i].start) &&
520 		    (end_pfn <= pfn_mapped[i].end))
521 			return true;
522 
523 	return false;
524 }
525 
526 /*
527  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
528  * This runs before bootmem is initialized and gets pages directly from
529  * the physical memory. To access them they are temporarily mapped.
530  */
init_memory_mapping(unsigned long start,unsigned long end,pgprot_t prot)531 unsigned long __ref init_memory_mapping(unsigned long start,
532 					unsigned long end, pgprot_t prot)
533 {
534 	struct map_range mr[NR_RANGE_MR];
535 	unsigned long ret = 0;
536 	int nr_range, i;
537 
538 	pr_debug("init_memory_mapping: [mem %#010lx-%#010lx]\n",
539 	       start, end - 1);
540 
541 	memset(mr, 0, sizeof(mr));
542 	nr_range = split_mem_range(mr, 0, start, end);
543 
544 	for (i = 0; i < nr_range; i++)
545 		ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
546 						   mr[i].page_size_mask,
547 						   prot);
548 
549 	add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT);
550 
551 	return ret >> PAGE_SHIFT;
552 }
553 
554 /*
555  * We need to iterate through the E820 memory map and create direct mappings
556  * for only E820_TYPE_RAM and E820_KERN_RESERVED regions. We cannot simply
557  * create direct mappings for all pfns from [0 to max_low_pfn) and
558  * [4GB to max_pfn) because of possible memory holes in high addresses
559  * that cannot be marked as UC by fixed/variable range MTRRs.
560  * Depending on the alignment of E820 ranges, this may possibly result
561  * in using smaller size (i.e. 4K instead of 2M or 1G) page tables.
562  *
563  * init_mem_mapping() calls init_range_memory_mapping() with big range.
564  * That range would have hole in the middle or ends, and only ram parts
565  * will be mapped in init_range_memory_mapping().
566  */
init_range_memory_mapping(unsigned long r_start,unsigned long r_end)567 static unsigned long __init init_range_memory_mapping(
568 					   unsigned long r_start,
569 					   unsigned long r_end)
570 {
571 	unsigned long start_pfn, end_pfn;
572 	unsigned long mapped_ram_size = 0;
573 	int i;
574 
575 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
576 		u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end);
577 		u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end);
578 		if (start >= end)
579 			continue;
580 
581 		/*
582 		 * if it is overlapping with brk pgt, we need to
583 		 * alloc pgt buf from memblock instead.
584 		 */
585 		can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >=
586 				    min(end, (u64)pgt_buf_top<<PAGE_SHIFT);
587 		init_memory_mapping(start, end, PAGE_KERNEL);
588 		mapped_ram_size += end - start;
589 		can_use_brk_pgt = true;
590 	}
591 
592 	return mapped_ram_size;
593 }
594 
get_new_step_size(unsigned long step_size)595 static unsigned long __init get_new_step_size(unsigned long step_size)
596 {
597 	/*
598 	 * Initial mapped size is PMD_SIZE (2M).
599 	 * We can not set step_size to be PUD_SIZE (1G) yet.
600 	 * In worse case, when we cross the 1G boundary, and
601 	 * PG_LEVEL_2M is not set, we will need 1+1+512 pages (2M + 8k)
602 	 * to map 1G range with PTE. Hence we use one less than the
603 	 * difference of page table level shifts.
604 	 *
605 	 * Don't need to worry about overflow in the top-down case, on 32bit,
606 	 * when step_size is 0, round_down() returns 0 for start, and that
607 	 * turns it into 0x100000000ULL.
608 	 * In the bottom-up case, round_up(x, 0) returns 0 though too, which
609 	 * needs to be taken into consideration by the code below.
610 	 */
611 	return step_size << (PMD_SHIFT - PAGE_SHIFT - 1);
612 }
613 
614 /**
615  * memory_map_top_down - Map [map_start, map_end) top down
616  * @map_start: start address of the target memory range
617  * @map_end: end address of the target memory range
618  *
619  * This function will setup direct mapping for memory range
620  * [map_start, map_end) in top-down. That said, the page tables
621  * will be allocated at the end of the memory, and we map the
622  * memory in top-down.
623  */
memory_map_top_down(unsigned long map_start,unsigned long map_end)624 static void __init memory_map_top_down(unsigned long map_start,
625 				       unsigned long map_end)
626 {
627 	unsigned long real_end, last_start;
628 	unsigned long step_size;
629 	unsigned long addr;
630 	unsigned long mapped_ram_size = 0;
631 
632 	/*
633 	 * Systems that have many reserved areas near top of the memory,
634 	 * e.g. QEMU with less than 1G RAM and EFI enabled, or Xen, will
635 	 * require lots of 4K mappings which may exhaust pgt_buf.
636 	 * Start with top-most PMD_SIZE range aligned at PMD_SIZE to ensure
637 	 * there is enough mapped memory that can be allocated from
638 	 * memblock.
639 	 */
640 	addr = memblock_phys_alloc_range(PMD_SIZE, PMD_SIZE, map_start,
641 					 map_end);
642 	memblock_phys_free(addr, PMD_SIZE);
643 	real_end = addr + PMD_SIZE;
644 
645 	/* step_size need to be small so pgt_buf from BRK could cover it */
646 	step_size = PMD_SIZE;
647 	max_pfn_mapped = 0; /* will get exact value next */
648 	min_pfn_mapped = real_end >> PAGE_SHIFT;
649 	last_start = real_end;
650 
651 	/*
652 	 * We start from the top (end of memory) and go to the bottom.
653 	 * The memblock_find_in_range() gets us a block of RAM from the
654 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
655 	 * for page table.
656 	 */
657 	while (last_start > map_start) {
658 		unsigned long start;
659 
660 		if (last_start > step_size) {
661 			start = round_down(last_start - 1, step_size);
662 			if (start < map_start)
663 				start = map_start;
664 		} else
665 			start = map_start;
666 		mapped_ram_size += init_range_memory_mapping(start,
667 							last_start);
668 		last_start = start;
669 		min_pfn_mapped = last_start >> PAGE_SHIFT;
670 		if (mapped_ram_size >= step_size)
671 			step_size = get_new_step_size(step_size);
672 	}
673 
674 	if (real_end < map_end)
675 		init_range_memory_mapping(real_end, map_end);
676 }
677 
678 /**
679  * memory_map_bottom_up - Map [map_start, map_end) bottom up
680  * @map_start: start address of the target memory range
681  * @map_end: end address of the target memory range
682  *
683  * This function will setup direct mapping for memory range
684  * [map_start, map_end) in bottom-up. Since we have limited the
685  * bottom-up allocation above the kernel, the page tables will
686  * be allocated just above the kernel and we map the memory
687  * in [map_start, map_end) in bottom-up.
688  */
memory_map_bottom_up(unsigned long map_start,unsigned long map_end)689 static void __init memory_map_bottom_up(unsigned long map_start,
690 					unsigned long map_end)
691 {
692 	unsigned long next, start;
693 	unsigned long mapped_ram_size = 0;
694 	/* step_size need to be small so pgt_buf from BRK could cover it */
695 	unsigned long step_size = PMD_SIZE;
696 
697 	start = map_start;
698 	min_pfn_mapped = start >> PAGE_SHIFT;
699 
700 	/*
701 	 * We start from the bottom (@map_start) and go to the top (@map_end).
702 	 * The memblock_find_in_range() gets us a block of RAM from the
703 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
704 	 * for page table.
705 	 */
706 	while (start < map_end) {
707 		if (step_size && map_end - start > step_size) {
708 			next = round_up(start + 1, step_size);
709 			if (next > map_end)
710 				next = map_end;
711 		} else {
712 			next = map_end;
713 		}
714 
715 		mapped_ram_size += init_range_memory_mapping(start, next);
716 		start = next;
717 
718 		if (mapped_ram_size >= step_size)
719 			step_size = get_new_step_size(step_size);
720 	}
721 }
722 
723 /*
724  * The real mode trampoline, which is required for bootstrapping CPUs
725  * occupies only a small area under the low 1MB.  See reserve_real_mode()
726  * for details.
727  *
728  * If KASLR is disabled the first PGD entry of the direct mapping is copied
729  * to map the real mode trampoline.
730  *
731  * If KASLR is enabled, copy only the PUD which covers the low 1MB
732  * area. This limits the randomization granularity to 1GB for both 4-level
733  * and 5-level paging.
734  */
init_trampoline(void)735 static void __init init_trampoline(void)
736 {
737 #ifdef CONFIG_X86_64
738 	/*
739 	 * The code below will alias kernel page-tables in the user-range of the
740 	 * address space, including the Global bit. So global TLB entries will
741 	 * be created when using the trampoline page-table.
742 	 */
743 	if (!kaslr_memory_enabled())
744 		trampoline_pgd_entry = init_top_pgt[pgd_index(__PAGE_OFFSET)];
745 	else
746 		init_trampoline_kaslr();
747 #endif
748 }
749 
init_mem_mapping(void)750 void __init init_mem_mapping(void)
751 {
752 	unsigned long end;
753 
754 	pti_check_boottime_disable();
755 	probe_page_size_mask();
756 	setup_pcid();
757 
758 #ifdef CONFIG_X86_64
759 	end = max_pfn << PAGE_SHIFT;
760 #else
761 	end = max_low_pfn << PAGE_SHIFT;
762 #endif
763 
764 	/* the ISA range is always mapped regardless of memory holes */
765 	init_memory_mapping(0, ISA_END_ADDRESS, PAGE_KERNEL);
766 
767 	/* Init the trampoline, possibly with KASLR memory offset */
768 	init_trampoline();
769 
770 	/*
771 	 * If the allocation is in bottom-up direction, we setup direct mapping
772 	 * in bottom-up, otherwise we setup direct mapping in top-down.
773 	 */
774 	if (memblock_bottom_up()) {
775 		unsigned long kernel_end = __pa_symbol(_end);
776 
777 		/*
778 		 * we need two separate calls here. This is because we want to
779 		 * allocate page tables above the kernel. So we first map
780 		 * [kernel_end, end) to make memory above the kernel be mapped
781 		 * as soon as possible. And then use page tables allocated above
782 		 * the kernel to map [ISA_END_ADDRESS, kernel_end).
783 		 */
784 		memory_map_bottom_up(kernel_end, end);
785 		memory_map_bottom_up(ISA_END_ADDRESS, kernel_end);
786 	} else {
787 		memory_map_top_down(ISA_END_ADDRESS, end);
788 	}
789 
790 #ifdef CONFIG_X86_64
791 	if (max_pfn > max_low_pfn) {
792 		/* can we preserve max_low_pfn ?*/
793 		max_low_pfn = max_pfn;
794 	}
795 #else
796 	early_ioremap_page_table_range_init();
797 #endif
798 
799 	load_cr3(swapper_pg_dir);
800 	__flush_tlb_all();
801 
802 	x86_init.hyper.init_mem_mapping();
803 
804 	early_memtest(0, max_pfn_mapped << PAGE_SHIFT);
805 }
806 
807 /*
808  * Initialize an mm_struct to be used during poking and a pointer to be used
809  * during patching.
810  */
poking_init(void)811 void __init poking_init(void)
812 {
813 	spinlock_t *ptl;
814 	pte_t *ptep;
815 
816 	poking_mm = mm_alloc();
817 	BUG_ON(!poking_mm);
818 
819 	/* Xen PV guests need the PGD to be pinned. */
820 	paravirt_enter_mmap(poking_mm);
821 
822 	/*
823 	 * Randomize the poking address, but make sure that the following page
824 	 * will be mapped at the same PMD. We need 2 pages, so find space for 3,
825 	 * and adjust the address if the PMD ends after the first one.
826 	 */
827 	poking_addr = TASK_UNMAPPED_BASE;
828 	if (IS_ENABLED(CONFIG_RANDOMIZE_BASE))
829 		poking_addr += (kaslr_get_random_long("Poking") & PAGE_MASK) %
830 			(TASK_SIZE - TASK_UNMAPPED_BASE - 3 * PAGE_SIZE);
831 
832 	if (((poking_addr + PAGE_SIZE) & ~PMD_MASK) == 0)
833 		poking_addr += PAGE_SIZE;
834 
835 	/*
836 	 * We need to trigger the allocation of the page-tables that will be
837 	 * needed for poking now. Later, poking may be performed in an atomic
838 	 * section, which might cause allocation to fail.
839 	 */
840 	ptep = get_locked_pte(poking_mm, poking_addr, &ptl);
841 	BUG_ON(!ptep);
842 	pte_unmap_unlock(ptep, ptl);
843 }
844 
845 /*
846  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
847  * is valid. The argument is a physical page number.
848  *
849  * On x86, access has to be given to the first megabyte of RAM because that
850  * area traditionally contains BIOS code and data regions used by X, dosemu,
851  * and similar apps. Since they map the entire memory range, the whole range
852  * must be allowed (for mapping), but any areas that would otherwise be
853  * disallowed are flagged as being "zero filled" instead of rejected.
854  * Access has to be given to non-kernel-ram areas as well, these contain the
855  * PCI mmio resources as well as potential bios/acpi data regions.
856  */
devmem_is_allowed(unsigned long pagenr)857 int devmem_is_allowed(unsigned long pagenr)
858 {
859 	if (region_intersects(PFN_PHYS(pagenr), PAGE_SIZE,
860 				IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE)
861 			!= REGION_DISJOINT) {
862 		/*
863 		 * For disallowed memory regions in the low 1MB range,
864 		 * request that the page be shown as all zeros.
865 		 */
866 		if (pagenr < 256)
867 			return 2;
868 
869 		return 0;
870 	}
871 
872 	/*
873 	 * This must follow RAM test, since System RAM is considered a
874 	 * restricted resource under CONFIG_STRICT_DEVMEM.
875 	 */
876 	if (iomem_is_exclusive(pagenr << PAGE_SHIFT)) {
877 		/* Low 1MB bypasses iomem restrictions. */
878 		if (pagenr < 256)
879 			return 1;
880 
881 		return 0;
882 	}
883 
884 	return 1;
885 }
886 
free_init_pages(const char * what,unsigned long begin,unsigned long end)887 void free_init_pages(const char *what, unsigned long begin, unsigned long end)
888 {
889 	unsigned long begin_aligned, end_aligned;
890 
891 	/* Make sure boundaries are page aligned */
892 	begin_aligned = PAGE_ALIGN(begin);
893 	end_aligned   = end & PAGE_MASK;
894 
895 	if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
896 		begin = begin_aligned;
897 		end   = end_aligned;
898 	}
899 
900 	if (begin >= end)
901 		return;
902 
903 	/*
904 	 * If debugging page accesses then do not free this memory but
905 	 * mark them not present - any buggy init-section access will
906 	 * create a kernel page fault:
907 	 */
908 	if (debug_pagealloc_enabled()) {
909 		pr_info("debug: unmapping init [mem %#010lx-%#010lx]\n",
910 			begin, end - 1);
911 		/*
912 		 * Inform kmemleak about the hole in the memory since the
913 		 * corresponding pages will be unmapped.
914 		 */
915 		kmemleak_free_part((void *)begin, end - begin);
916 		set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
917 	} else {
918 		/*
919 		 * We just marked the kernel text read only above, now that
920 		 * we are going to free part of that, we need to make that
921 		 * writeable and non-executable first.
922 		 */
923 		set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
924 		set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
925 
926 		free_reserved_area((void *)begin, (void *)end,
927 				   POISON_FREE_INITMEM, what);
928 	}
929 }
930 
931 /*
932  * begin/end can be in the direct map or the "high kernel mapping"
933  * used for the kernel image only.  free_init_pages() will do the
934  * right thing for either kind of address.
935  */
free_kernel_image_pages(const char * what,void * begin,void * end)936 void free_kernel_image_pages(const char *what, void *begin, void *end)
937 {
938 	unsigned long begin_ul = (unsigned long)begin;
939 	unsigned long end_ul = (unsigned long)end;
940 	unsigned long len_pages = (end_ul - begin_ul) >> PAGE_SHIFT;
941 
942 	free_init_pages(what, begin_ul, end_ul);
943 
944 	/*
945 	 * PTI maps some of the kernel into userspace.  For performance,
946 	 * this includes some kernel areas that do not contain secrets.
947 	 * Those areas might be adjacent to the parts of the kernel image
948 	 * being freed, which may contain secrets.  Remove the "high kernel
949 	 * image mapping" for these freed areas, ensuring they are not even
950 	 * potentially vulnerable to Meltdown regardless of the specific
951 	 * optimizations PTI is currently using.
952 	 *
953 	 * The "noalias" prevents unmapping the direct map alias which is
954 	 * needed to access the freed pages.
955 	 *
956 	 * This is only valid for 64bit kernels. 32bit has only one mapping
957 	 * which can't be treated in this way for obvious reasons.
958 	 */
959 	if (IS_ENABLED(CONFIG_X86_64) && cpu_feature_enabled(X86_FEATURE_PTI))
960 		set_memory_np_noalias(begin_ul, len_pages);
961 }
962 
free_initmem(void)963 void __ref free_initmem(void)
964 {
965 	e820__reallocate_tables();
966 
967 	mem_encrypt_free_decrypted_mem();
968 
969 	free_kernel_image_pages("unused kernel image (initmem)",
970 				&__init_begin, &__init_end);
971 }
972 
973 #ifdef CONFIG_BLK_DEV_INITRD
free_initrd_mem(unsigned long start,unsigned long end)974 void __init free_initrd_mem(unsigned long start, unsigned long end)
975 {
976 	/*
977 	 * end could be not aligned, and We can not align that,
978 	 * decompressor could be confused by aligned initrd_end
979 	 * We already reserve the end partial page before in
980 	 *   - i386_start_kernel()
981 	 *   - x86_64_start_kernel()
982 	 *   - relocate_initrd()
983 	 * So here We can do PAGE_ALIGN() safely to get partial page to be freed
984 	 */
985 	free_init_pages("initrd", start, PAGE_ALIGN(end));
986 }
987 #endif
988 
989 /*
990  * Calculate the precise size of the DMA zone (first 16 MB of RAM),
991  * and pass it to the MM layer - to help it set zone watermarks more
992  * accurately.
993  *
994  * Done on 64-bit systems only for the time being, although 32-bit systems
995  * might benefit from this as well.
996  */
memblock_find_dma_reserve(void)997 void __init memblock_find_dma_reserve(void)
998 {
999 #ifdef CONFIG_X86_64
1000 	u64 nr_pages = 0, nr_free_pages = 0;
1001 	unsigned long start_pfn, end_pfn;
1002 	phys_addr_t start_addr, end_addr;
1003 	int i;
1004 	u64 u;
1005 
1006 	/*
1007 	 * Iterate over all memory ranges (free and reserved ones alike),
1008 	 * to calculate the total number of pages in the first 16 MB of RAM:
1009 	 */
1010 	nr_pages = 0;
1011 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
1012 		start_pfn = min(start_pfn, MAX_DMA_PFN);
1013 		end_pfn   = min(end_pfn,   MAX_DMA_PFN);
1014 
1015 		nr_pages += end_pfn - start_pfn;
1016 	}
1017 
1018 	/*
1019 	 * Iterate over free memory ranges to calculate the number of free
1020 	 * pages in the DMA zone, while not counting potential partial
1021 	 * pages at the beginning or the end of the range:
1022 	 */
1023 	nr_free_pages = 0;
1024 	for_each_free_mem_range(u, NUMA_NO_NODE, MEMBLOCK_NONE, &start_addr, &end_addr, NULL) {
1025 		start_pfn = min_t(unsigned long, PFN_UP(start_addr), MAX_DMA_PFN);
1026 		end_pfn   = min_t(unsigned long, PFN_DOWN(end_addr), MAX_DMA_PFN);
1027 
1028 		if (start_pfn < end_pfn)
1029 			nr_free_pages += end_pfn - start_pfn;
1030 	}
1031 
1032 	set_dma_reserve(nr_pages - nr_free_pages);
1033 #endif
1034 }
1035 
zone_sizes_init(void)1036 void __init zone_sizes_init(void)
1037 {
1038 	unsigned long max_zone_pfns[MAX_NR_ZONES];
1039 
1040 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1041 
1042 #ifdef CONFIG_ZONE_DMA
1043 	max_zone_pfns[ZONE_DMA]		= min(MAX_DMA_PFN, max_low_pfn);
1044 #endif
1045 #ifdef CONFIG_ZONE_DMA32
1046 	max_zone_pfns[ZONE_DMA32]	= min(MAX_DMA32_PFN, max_low_pfn);
1047 #endif
1048 	max_zone_pfns[ZONE_NORMAL]	= max_low_pfn;
1049 #ifdef CONFIG_HIGHMEM
1050 	max_zone_pfns[ZONE_HIGHMEM]	= max_pfn;
1051 #endif
1052 
1053 	free_area_init(max_zone_pfns);
1054 }
1055 
1056 __visible DEFINE_PER_CPU_ALIGNED(struct tlb_state, cpu_tlbstate) = {
1057 	.loaded_mm = &init_mm,
1058 	.next_asid = 1,
1059 	.cr4 = ~0UL,	/* fail hard if we screw up cr4 shadow initialization */
1060 };
1061 
1062 #ifdef CONFIG_ADDRESS_MASKING
1063 DEFINE_PER_CPU(u64, tlbstate_untag_mask);
1064 EXPORT_PER_CPU_SYMBOL(tlbstate_untag_mask);
1065 #endif
1066 
update_cache_mode_entry(unsigned entry,enum page_cache_mode cache)1067 void update_cache_mode_entry(unsigned entry, enum page_cache_mode cache)
1068 {
1069 	/* entry 0 MUST be WB (hardwired to speed up translations) */
1070 	BUG_ON(!entry && cache != _PAGE_CACHE_MODE_WB);
1071 
1072 	__cachemode2pte_tbl[cache] = __cm_idx2pte(entry);
1073 	__pte2cachemode_tbl[entry] = cache;
1074 }
1075 
1076 #ifdef CONFIG_SWAP
arch_max_swapfile_size(void)1077 unsigned long arch_max_swapfile_size(void)
1078 {
1079 	unsigned long pages;
1080 
1081 	pages = generic_max_swapfile_size();
1082 
1083 	if (boot_cpu_has_bug(X86_BUG_L1TF) && l1tf_mitigation != L1TF_MITIGATION_OFF) {
1084 		/* Limit the swap file size to MAX_PA/2 for L1TF workaround */
1085 		unsigned long long l1tf_limit = l1tf_pfn_limit();
1086 		/*
1087 		 * We encode swap offsets also with 3 bits below those for pfn
1088 		 * which makes the usable limit higher.
1089 		 */
1090 #if CONFIG_PGTABLE_LEVELS > 2
1091 		l1tf_limit <<= PAGE_SHIFT - SWP_OFFSET_FIRST_BIT;
1092 #endif
1093 		pages = min_t(unsigned long long, l1tf_limit, pages);
1094 	}
1095 	return pages;
1096 }
1097 #endif
1098