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