xref: /kernel/linux/linux-5.10/arch/x86/mm/init.c (revision 8c2ecf20)
1#include <linux/gfp.h>
2#include <linux/initrd.h>
3#include <linux/ioport.h>
4#include <linux/swap.h>
5#include <linux/memblock.h>
6#include <linux/swapfile.h>
7#include <linux/swapops.h>
8#include <linux/kmemleak.h>
9#include <linux/sched/task.h>
10#include <linux/sched/mm.h>
11
12#include <asm/set_memory.h>
13#include <asm/cpu_device_id.h>
14#include <asm/e820/api.h>
15#include <asm/init.h>
16#include <asm/page.h>
17#include <asm/page_types.h>
18#include <asm/sections.h>
19#include <asm/setup.h>
20#include <asm/tlbflush.h>
21#include <asm/tlb.h>
22#include <asm/proto.h>
23#include <asm/dma.h>		/* for MAX_DMA_PFN */
24#include <asm/microcode.h>
25#include <asm/kaslr.h>
26#include <asm/hypervisor.h>
27#include <asm/cpufeature.h>
28#include <asm/pti.h>
29#include <asm/text-patching.h>
30#include <asm/memtype.h>
31#include <asm/paravirt.h>
32
33/*
34 * We need to define the tracepoints somewhere, and tlb.c
35 * is only compied when SMP=y.
36 */
37#define CREATE_TRACE_POINTS
38#include <trace/events/tlb.h>
39
40#include "mm_internal.h"
41
42/*
43 * Tables translating between page_cache_type_t and pte encoding.
44 *
45 * The default values are defined statically as minimal supported mode;
46 * WC and WT fall back to UC-.  pat_init() updates these values to support
47 * more cache modes, WC and WT, when it is safe to do so.  See pat_init()
48 * for the details.  Note, __early_ioremap() used during early boot-time
49 * takes pgprot_t (pte encoding) and does not use these tables.
50 *
51 *   Index into __cachemode2pte_tbl[] is the cachemode.
52 *
53 *   Index into __pte2cachemode_tbl[] are the caching attribute bits of the pte
54 *   (_PAGE_PWT, _PAGE_PCD, _PAGE_PAT) at index bit positions 0, 1, 2.
55 */
56static uint16_t __cachemode2pte_tbl[_PAGE_CACHE_MODE_NUM] = {
57	[_PAGE_CACHE_MODE_WB      ]	= 0         | 0        ,
58	[_PAGE_CACHE_MODE_WC      ]	= 0         | _PAGE_PCD,
59	[_PAGE_CACHE_MODE_UC_MINUS]	= 0         | _PAGE_PCD,
60	[_PAGE_CACHE_MODE_UC      ]	= _PAGE_PWT | _PAGE_PCD,
61	[_PAGE_CACHE_MODE_WT      ]	= 0         | _PAGE_PCD,
62	[_PAGE_CACHE_MODE_WP      ]	= 0         | _PAGE_PCD,
63};
64
65unsigned long cachemode2protval(enum page_cache_mode pcm)
66{
67	if (likely(pcm == 0))
68		return 0;
69	return __cachemode2pte_tbl[pcm];
70}
71EXPORT_SYMBOL(cachemode2protval);
72
73static uint8_t __pte2cachemode_tbl[8] = {
74	[__pte2cm_idx( 0        | 0         | 0        )] = _PAGE_CACHE_MODE_WB,
75	[__pte2cm_idx(_PAGE_PWT | 0         | 0        )] = _PAGE_CACHE_MODE_UC_MINUS,
76	[__pte2cm_idx( 0        | _PAGE_PCD | 0        )] = _PAGE_CACHE_MODE_UC_MINUS,
77	[__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | 0        )] = _PAGE_CACHE_MODE_UC,
78	[__pte2cm_idx( 0        | 0         | _PAGE_PAT)] = _PAGE_CACHE_MODE_WB,
79	[__pte2cm_idx(_PAGE_PWT | 0         | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
80	[__pte2cm_idx(0         | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
81	[__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC,
82};
83
84/*
85 * Check that the write-protect PAT entry is set for write-protect.
86 * To do this without making assumptions how PAT has been set up (Xen has
87 * another layout than the kernel), translate the _PAGE_CACHE_MODE_WP cache
88 * mode via the __cachemode2pte_tbl[] into protection bits (those protection
89 * bits will select a cache mode of WP or better), and then translate the
90 * protection bits back into the cache mode using __pte2cm_idx() and the
91 * __pte2cachemode_tbl[] array. This will return the really used cache mode.
92 */
93bool x86_has_pat_wp(void)
94{
95	uint16_t prot = __cachemode2pte_tbl[_PAGE_CACHE_MODE_WP];
96
97	return __pte2cachemode_tbl[__pte2cm_idx(prot)] == _PAGE_CACHE_MODE_WP;
98}
99
100enum page_cache_mode pgprot2cachemode(pgprot_t pgprot)
101{
102	unsigned long masked;
103
104	masked = pgprot_val(pgprot) & _PAGE_CACHE_MASK;
105	if (likely(masked == 0))
106		return 0;
107	return __pte2cachemode_tbl[__pte2cm_idx(masked)];
108}
109
110static unsigned long __initdata pgt_buf_start;
111static unsigned long __initdata pgt_buf_end;
112static unsigned long __initdata pgt_buf_top;
113
114static unsigned long min_pfn_mapped;
115
116static bool __initdata can_use_brk_pgt = true;
117
118/*
119 * Pages returned are already directly mapped.
120 *
121 * Changing that is likely to break Xen, see commit:
122 *
123 *    279b706 x86,xen: introduce x86_init.mapping.pagetable_reserve
124 *
125 * for detailed information.
126 */
127__ref void *alloc_low_pages(unsigned int num)
128{
129	unsigned long pfn;
130	int i;
131
132	if (after_bootmem) {
133		unsigned int order;
134
135		order = get_order((unsigned long)num << PAGE_SHIFT);
136		return (void *)__get_free_pages(GFP_ATOMIC | __GFP_ZERO, order);
137	}
138
139	if ((pgt_buf_end + num) > pgt_buf_top || !can_use_brk_pgt) {
140		unsigned long ret = 0;
141
142		if (min_pfn_mapped < max_pfn_mapped) {
143			ret = memblock_find_in_range(
144					min_pfn_mapped << PAGE_SHIFT,
145					max_pfn_mapped << PAGE_SHIFT,
146					PAGE_SIZE * num , PAGE_SIZE);
147		}
148		if (ret)
149			memblock_reserve(ret, PAGE_SIZE * num);
150		else if (can_use_brk_pgt)
151			ret = __pa(extend_brk(PAGE_SIZE * num, PAGE_SIZE));
152
153		if (!ret)
154			panic("alloc_low_pages: can not alloc memory");
155
156		pfn = ret >> PAGE_SHIFT;
157	} else {
158		pfn = pgt_buf_end;
159		pgt_buf_end += num;
160	}
161
162	for (i = 0; i < num; i++) {
163		void *adr;
164
165		adr = __va((pfn + i) << PAGE_SHIFT);
166		clear_page(adr);
167	}
168
169	return __va(pfn << PAGE_SHIFT);
170}
171
172/*
173 * By default need 3 4k for initial PMD_SIZE,  3 4k for 0-ISA_END_ADDRESS.
174 * With KASLR memory randomization, depending on the machine e820 memory
175 * and the PUD alignment. We may need twice more pages when KASLR memory
176 * randomization is enabled.
177 */
178#ifndef CONFIG_RANDOMIZE_MEMORY
179#define INIT_PGD_PAGE_COUNT      6
180#else
181#define INIT_PGD_PAGE_COUNT      12
182#endif
183#define INIT_PGT_BUF_SIZE	(INIT_PGD_PAGE_COUNT * PAGE_SIZE)
184RESERVE_BRK(early_pgt_alloc, INIT_PGT_BUF_SIZE);
185void  __init early_alloc_pgt_buf(void)
186{
187	unsigned long tables = INIT_PGT_BUF_SIZE;
188	phys_addr_t base;
189
190	base = __pa(extend_brk(tables, PAGE_SIZE));
191
192	pgt_buf_start = base >> PAGE_SHIFT;
193	pgt_buf_end = pgt_buf_start;
194	pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
195}
196
197int after_bootmem;
198
199early_param_on_off("gbpages", "nogbpages", direct_gbpages, CONFIG_X86_DIRECT_GBPAGES);
200
201struct map_range {
202	unsigned long start;
203	unsigned long end;
204	unsigned page_size_mask;
205};
206
207static int page_size_mask;
208
209/*
210 * Save some of cr4 feature set we're using (e.g.  Pentium 4MB
211 * enable and PPro Global page enable), so that any CPU's that boot
212 * up after us can get the correct flags. Invoked on the boot CPU.
213 */
214static inline void cr4_set_bits_and_update_boot(unsigned long mask)
215{
216	mmu_cr4_features |= mask;
217	if (trampoline_cr4_features)
218		*trampoline_cr4_features = mmu_cr4_features;
219	cr4_set_bits(mask);
220}
221
222static void __init probe_page_size_mask(void)
223{
224	/*
225	 * For pagealloc debugging, identity mapping will use small pages.
226	 * This will simplify cpa(), which otherwise needs to support splitting
227	 * large pages into small in interrupt context, etc.
228	 */
229	if (boot_cpu_has(X86_FEATURE_PSE) && !debug_pagealloc_enabled())
230		page_size_mask |= 1 << PG_LEVEL_2M;
231	else
232		direct_gbpages = 0;
233
234	/* Enable PSE if available */
235	if (boot_cpu_has(X86_FEATURE_PSE))
236		cr4_set_bits_and_update_boot(X86_CR4_PSE);
237
238	/* Enable PGE if available */
239	__supported_pte_mask &= ~_PAGE_GLOBAL;
240	if (boot_cpu_has(X86_FEATURE_PGE)) {
241		cr4_set_bits_and_update_boot(X86_CR4_PGE);
242		__supported_pte_mask |= _PAGE_GLOBAL;
243	}
244
245	/* By the default is everything supported: */
246	__default_kernel_pte_mask = __supported_pte_mask;
247	/* Except when with PTI where the kernel is mostly non-Global: */
248	if (cpu_feature_enabled(X86_FEATURE_PTI))
249		__default_kernel_pte_mask &= ~_PAGE_GLOBAL;
250
251	/* Enable 1 GB linear kernel mappings if available: */
252	if (direct_gbpages && boot_cpu_has(X86_FEATURE_GBPAGES)) {
253		printk(KERN_INFO "Using GB pages for direct mapping\n");
254		page_size_mask |= 1 << PG_LEVEL_1G;
255	} else {
256		direct_gbpages = 0;
257	}
258}
259
260#define INTEL_MATCH(_model) { .vendor  = X86_VENDOR_INTEL,	\
261			      .family  = 6,			\
262			      .model = _model,			\
263			    }
264/*
265 * INVLPG may not properly flush Global entries
266 * on these CPUs when PCIDs are enabled.
267 */
268static const struct x86_cpu_id invlpg_miss_ids[] = {
269	INTEL_MATCH(INTEL_FAM6_ALDERLAKE   ),
270	INTEL_MATCH(INTEL_FAM6_ALDERLAKE_L ),
271	INTEL_MATCH(INTEL_FAM6_ALDERLAKE_N ),
272	INTEL_MATCH(INTEL_FAM6_RAPTORLAKE  ),
273	INTEL_MATCH(INTEL_FAM6_RAPTORLAKE_P),
274	INTEL_MATCH(INTEL_FAM6_RAPTORLAKE_S),
275	{}
276};
277
278static 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
306		/*
307		 * INVPCID's single-context modes (2/3) only work if we set
308		 * X86_CR4_PCIDE, *and* we INVPCID support.  It's unusable
309		 * on systems that have X86_CR4_PCIDE clear, or that have
310		 * no INVPCID support at all.
311		 */
312		if (boot_cpu_has(X86_FEATURE_INVPCID))
313			setup_force_cpu_cap(X86_FEATURE_INVPCID_SINGLE);
314	} else {
315		/*
316		 * flush_tlb_all(), as currently implemented, won't work if
317		 * PCID is on but PGE is not.  Since that combination
318		 * doesn't exist on real hardware, there's no reason to try
319		 * to fully support it, but it's polite to avoid corrupting
320		 * data if we're on an improperly configured VM.
321		 */
322		setup_clear_cpu_cap(X86_FEATURE_PCID);
323	}
324}
325
326#ifdef CONFIG_X86_32
327#define NR_RANGE_MR 3
328#else /* CONFIG_X86_64 */
329#define NR_RANGE_MR 5
330#endif
331
332static int __meminit save_mr(struct map_range *mr, int nr_range,
333			     unsigned long start_pfn, unsigned long end_pfn,
334			     unsigned long page_size_mask)
335{
336	if (start_pfn < end_pfn) {
337		if (nr_range >= NR_RANGE_MR)
338			panic("run out of range for init_memory_mapping\n");
339		mr[nr_range].start = start_pfn<<PAGE_SHIFT;
340		mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
341		mr[nr_range].page_size_mask = page_size_mask;
342		nr_range++;
343	}
344
345	return nr_range;
346}
347
348/*
349 * adjust the page_size_mask for small range to go with
350 *	big page size instead small one if nearby are ram too.
351 */
352static void __ref adjust_range_page_size_mask(struct map_range *mr,
353							 int nr_range)
354{
355	int i;
356
357	for (i = 0; i < nr_range; i++) {
358		if ((page_size_mask & (1<<PG_LEVEL_2M)) &&
359		    !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) {
360			unsigned long start = round_down(mr[i].start, PMD_SIZE);
361			unsigned long end = round_up(mr[i].end, PMD_SIZE);
362
363#ifdef CONFIG_X86_32
364			if ((end >> PAGE_SHIFT) > max_low_pfn)
365				continue;
366#endif
367
368			if (memblock_is_region_memory(start, end - start))
369				mr[i].page_size_mask |= 1<<PG_LEVEL_2M;
370		}
371		if ((page_size_mask & (1<<PG_LEVEL_1G)) &&
372		    !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) {
373			unsigned long start = round_down(mr[i].start, PUD_SIZE);
374			unsigned long end = round_up(mr[i].end, PUD_SIZE);
375
376			if (memblock_is_region_memory(start, end - start))
377				mr[i].page_size_mask |= 1<<PG_LEVEL_1G;
378		}
379	}
380}
381
382static const char *page_size_string(struct map_range *mr)
383{
384	static const char str_1g[] = "1G";
385	static const char str_2m[] = "2M";
386	static const char str_4m[] = "4M";
387	static const char str_4k[] = "4k";
388
389	if (mr->page_size_mask & (1<<PG_LEVEL_1G))
390		return str_1g;
391	/*
392	 * 32-bit without PAE has a 4M large page size.
393	 * PG_LEVEL_2M is misnamed, but we can at least
394	 * print out the right size in the string.
395	 */
396	if (IS_ENABLED(CONFIG_X86_32) &&
397	    !IS_ENABLED(CONFIG_X86_PAE) &&
398	    mr->page_size_mask & (1<<PG_LEVEL_2M))
399		return str_4m;
400
401	if (mr->page_size_mask & (1<<PG_LEVEL_2M))
402		return str_2m;
403
404	return str_4k;
405}
406
407static int __meminit split_mem_range(struct map_range *mr, int nr_range,
408				     unsigned long start,
409				     unsigned long end)
410{
411	unsigned long start_pfn, end_pfn, limit_pfn;
412	unsigned long pfn;
413	int i;
414
415	limit_pfn = PFN_DOWN(end);
416
417	/* head if not big page alignment ? */
418	pfn = start_pfn = PFN_DOWN(start);
419#ifdef CONFIG_X86_32
420	/*
421	 * Don't use a large page for the first 2/4MB of memory
422	 * because there are often fixed size MTRRs in there
423	 * and overlapping MTRRs into large pages can cause
424	 * slowdowns.
425	 */
426	if (pfn == 0)
427		end_pfn = PFN_DOWN(PMD_SIZE);
428	else
429		end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
430#else /* CONFIG_X86_64 */
431	end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
432#endif
433	if (end_pfn > limit_pfn)
434		end_pfn = limit_pfn;
435	if (start_pfn < end_pfn) {
436		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
437		pfn = end_pfn;
438	}
439
440	/* big page (2M) range */
441	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
442#ifdef CONFIG_X86_32
443	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
444#else /* CONFIG_X86_64 */
445	end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
446	if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE)))
447		end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
448#endif
449
450	if (start_pfn < end_pfn) {
451		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
452				page_size_mask & (1<<PG_LEVEL_2M));
453		pfn = end_pfn;
454	}
455
456#ifdef CONFIG_X86_64
457	/* big page (1G) range */
458	start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
459	end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE));
460	if (start_pfn < end_pfn) {
461		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
462				page_size_mask &
463				 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
464		pfn = end_pfn;
465	}
466
467	/* tail is not big page (1G) alignment */
468	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
469	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
470	if (start_pfn < end_pfn) {
471		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
472				page_size_mask & (1<<PG_LEVEL_2M));
473		pfn = end_pfn;
474	}
475#endif
476
477	/* tail is not big page (2M) alignment */
478	start_pfn = pfn;
479	end_pfn = limit_pfn;
480	nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
481
482	if (!after_bootmem)
483		adjust_range_page_size_mask(mr, nr_range);
484
485	/* try to merge same page size and continuous */
486	for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
487		unsigned long old_start;
488		if (mr[i].end != mr[i+1].start ||
489		    mr[i].page_size_mask != mr[i+1].page_size_mask)
490			continue;
491		/* move it */
492		old_start = mr[i].start;
493		memmove(&mr[i], &mr[i+1],
494			(nr_range - 1 - i) * sizeof(struct map_range));
495		mr[i--].start = old_start;
496		nr_range--;
497	}
498
499	for (i = 0; i < nr_range; i++)
500		pr_debug(" [mem %#010lx-%#010lx] page %s\n",
501				mr[i].start, mr[i].end - 1,
502				page_size_string(&mr[i]));
503
504	return nr_range;
505}
506
507struct range pfn_mapped[E820_MAX_ENTRIES];
508int nr_pfn_mapped;
509
510static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn)
511{
512	nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_MAX_ENTRIES,
513					     nr_pfn_mapped, start_pfn, end_pfn);
514	nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_MAX_ENTRIES);
515
516	max_pfn_mapped = max(max_pfn_mapped, end_pfn);
517
518	if (start_pfn < (1UL<<(32-PAGE_SHIFT)))
519		max_low_pfn_mapped = max(max_low_pfn_mapped,
520					 min(end_pfn, 1UL<<(32-PAGE_SHIFT)));
521}
522
523bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn)
524{
525	int i;
526
527	for (i = 0; i < nr_pfn_mapped; i++)
528		if ((start_pfn >= pfn_mapped[i].start) &&
529		    (end_pfn <= pfn_mapped[i].end))
530			return true;
531
532	return false;
533}
534
535/*
536 * Setup the direct mapping of the physical memory at PAGE_OFFSET.
537 * This runs before bootmem is initialized and gets pages directly from
538 * the physical memory. To access them they are temporarily mapped.
539 */
540unsigned long __ref init_memory_mapping(unsigned long start,
541					unsigned long end, pgprot_t prot)
542{
543	struct map_range mr[NR_RANGE_MR];
544	unsigned long ret = 0;
545	int nr_range, i;
546
547	pr_debug("init_memory_mapping: [mem %#010lx-%#010lx]\n",
548	       start, end - 1);
549
550	memset(mr, 0, sizeof(mr));
551	nr_range = split_mem_range(mr, 0, start, end);
552
553	for (i = 0; i < nr_range; i++)
554		ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
555						   mr[i].page_size_mask,
556						   prot);
557
558	add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT);
559
560	return ret >> PAGE_SHIFT;
561}
562
563/*
564 * We need to iterate through the E820 memory map and create direct mappings
565 * for only E820_TYPE_RAM and E820_KERN_RESERVED regions. We cannot simply
566 * create direct mappings for all pfns from [0 to max_low_pfn) and
567 * [4GB to max_pfn) because of possible memory holes in high addresses
568 * that cannot be marked as UC by fixed/variable range MTRRs.
569 * Depending on the alignment of E820 ranges, this may possibly result
570 * in using smaller size (i.e. 4K instead of 2M or 1G) page tables.
571 *
572 * init_mem_mapping() calls init_range_memory_mapping() with big range.
573 * That range would have hole in the middle or ends, and only ram parts
574 * will be mapped in init_range_memory_mapping().
575 */
576static unsigned long __init init_range_memory_mapping(
577					   unsigned long r_start,
578					   unsigned long r_end)
579{
580	unsigned long start_pfn, end_pfn;
581	unsigned long mapped_ram_size = 0;
582	int i;
583
584	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
585		u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end);
586		u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end);
587		if (start >= end)
588			continue;
589
590		/*
591		 * if it is overlapping with brk pgt, we need to
592		 * alloc pgt buf from memblock instead.
593		 */
594		can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >=
595				    min(end, (u64)pgt_buf_top<<PAGE_SHIFT);
596		init_memory_mapping(start, end, PAGE_KERNEL);
597		mapped_ram_size += end - start;
598		can_use_brk_pgt = true;
599	}
600
601	return mapped_ram_size;
602}
603
604static unsigned long __init get_new_step_size(unsigned long step_size)
605{
606	/*
607	 * Initial mapped size is PMD_SIZE (2M).
608	 * We can not set step_size to be PUD_SIZE (1G) yet.
609	 * In worse case, when we cross the 1G boundary, and
610	 * PG_LEVEL_2M is not set, we will need 1+1+512 pages (2M + 8k)
611	 * to map 1G range with PTE. Hence we use one less than the
612	 * difference of page table level shifts.
613	 *
614	 * Don't need to worry about overflow in the top-down case, on 32bit,
615	 * when step_size is 0, round_down() returns 0 for start, and that
616	 * turns it into 0x100000000ULL.
617	 * In the bottom-up case, round_up(x, 0) returns 0 though too, which
618	 * needs to be taken into consideration by the code below.
619	 */
620	return step_size << (PMD_SHIFT - PAGE_SHIFT - 1);
621}
622
623/**
624 * memory_map_top_down - Map [map_start, map_end) top down
625 * @map_start: start address of the target memory range
626 * @map_end: end address of the target memory range
627 *
628 * This function will setup direct mapping for memory range
629 * [map_start, map_end) in top-down. That said, the page tables
630 * will be allocated at the end of the memory, and we map the
631 * memory in top-down.
632 */
633static void __init memory_map_top_down(unsigned long map_start,
634				       unsigned long map_end)
635{
636	unsigned long real_end, start, last_start;
637	unsigned long step_size;
638	unsigned long addr;
639	unsigned long mapped_ram_size = 0;
640
641	/* xen has big range in reserved near end of ram, skip it at first.*/
642	addr = memblock_find_in_range(map_start, map_end, PMD_SIZE, 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 = 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		if (last_start > step_size) {
659			start = round_down(last_start - 1, step_size);
660			if (start < map_start)
661				start = map_start;
662		} else
663			start = map_start;
664		mapped_ram_size += init_range_memory_mapping(start,
665							last_start);
666		last_start = start;
667		min_pfn_mapped = last_start >> PAGE_SHIFT;
668		if (mapped_ram_size >= step_size)
669			step_size = get_new_step_size(step_size);
670	}
671
672	if (real_end < map_end)
673		init_range_memory_mapping(real_end, map_end);
674}
675
676/**
677 * memory_map_bottom_up - Map [map_start, map_end) bottom up
678 * @map_start: start address of the target memory range
679 * @map_end: end address of the target memory range
680 *
681 * This function will setup direct mapping for memory range
682 * [map_start, map_end) in bottom-up. Since we have limited the
683 * bottom-up allocation above the kernel, the page tables will
684 * be allocated just above the kernel and we map the memory
685 * in [map_start, map_end) in bottom-up.
686 */
687static void __init memory_map_bottom_up(unsigned long map_start,
688					unsigned long map_end)
689{
690	unsigned long next, start;
691	unsigned long mapped_ram_size = 0;
692	/* step_size need to be small so pgt_buf from BRK could cover it */
693	unsigned long step_size = PMD_SIZE;
694
695	start = map_start;
696	min_pfn_mapped = start >> PAGE_SHIFT;
697
698	/*
699	 * We start from the bottom (@map_start) and go to the top (@map_end).
700	 * The memblock_find_in_range() gets us a block of RAM from the
701	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
702	 * for page table.
703	 */
704	while (start < map_end) {
705		if (step_size && map_end - start > step_size) {
706			next = round_up(start + 1, step_size);
707			if (next > map_end)
708				next = map_end;
709		} else {
710			next = map_end;
711		}
712
713		mapped_ram_size += init_range_memory_mapping(start, next);
714		start = next;
715
716		if (mapped_ram_size >= step_size)
717			step_size = get_new_step_size(step_size);
718	}
719}
720
721/*
722 * The real mode trampoline, which is required for bootstrapping CPUs
723 * occupies only a small area under the low 1MB.  See reserve_real_mode()
724 * for details.
725 *
726 * If KASLR is disabled the first PGD entry of the direct mapping is copied
727 * to map the real mode trampoline.
728 *
729 * If KASLR is enabled, copy only the PUD which covers the low 1MB
730 * area. This limits the randomization granularity to 1GB for both 4-level
731 * and 5-level paging.
732 */
733static void __init init_trampoline(void)
734{
735#ifdef CONFIG_X86_64
736	if (!kaslr_memory_enabled())
737		trampoline_pgd_entry = init_top_pgt[pgd_index(__PAGE_OFFSET)];
738	else
739		init_trampoline_kaslr();
740#endif
741}
742
743void __init init_mem_mapping(void)
744{
745	unsigned long end;
746
747	pti_check_boottime_disable();
748	probe_page_size_mask();
749	setup_pcid();
750
751#ifdef CONFIG_X86_64
752	end = max_pfn << PAGE_SHIFT;
753#else
754	end = max_low_pfn << PAGE_SHIFT;
755#endif
756
757	/* the ISA range is always mapped regardless of memory holes */
758	init_memory_mapping(0, ISA_END_ADDRESS, PAGE_KERNEL);
759
760	/* Init the trampoline, possibly with KASLR memory offset */
761	init_trampoline();
762
763	/*
764	 * If the allocation is in bottom-up direction, we setup direct mapping
765	 * in bottom-up, otherwise we setup direct mapping in top-down.
766	 */
767	if (memblock_bottom_up()) {
768		unsigned long kernel_end = __pa_symbol(_end);
769
770		/*
771		 * we need two separate calls here. This is because we want to
772		 * allocate page tables above the kernel. So we first map
773		 * [kernel_end, end) to make memory above the kernel be mapped
774		 * as soon as possible. And then use page tables allocated above
775		 * the kernel to map [ISA_END_ADDRESS, kernel_end).
776		 */
777		memory_map_bottom_up(kernel_end, end);
778		memory_map_bottom_up(ISA_END_ADDRESS, kernel_end);
779	} else {
780		memory_map_top_down(ISA_END_ADDRESS, end);
781	}
782
783#ifdef CONFIG_X86_64
784	if (max_pfn > max_low_pfn) {
785		/* can we preseve max_low_pfn ?*/
786		max_low_pfn = max_pfn;
787	}
788#else
789	early_ioremap_page_table_range_init();
790#endif
791
792	load_cr3(swapper_pg_dir);
793	__flush_tlb_all();
794
795	x86_init.hyper.init_mem_mapping();
796
797	early_memtest(0, max_pfn_mapped << PAGE_SHIFT);
798}
799
800/*
801 * Initialize an mm_struct to be used during poking and a pointer to be used
802 * during patching.
803 */
804void __init poking_init(void)
805{
806	spinlock_t *ptl;
807	pte_t *ptep;
808
809	poking_mm = mm_alloc();
810	BUG_ON(!poking_mm);
811
812	/* Xen PV guests need the PGD to be pinned. */
813	paravirt_arch_dup_mmap(NULL, poking_mm);
814
815	/*
816	 * Randomize the poking address, but make sure that the following page
817	 * will be mapped at the same PMD. We need 2 pages, so find space for 3,
818	 * and adjust the address if the PMD ends after the first one.
819	 */
820	poking_addr = TASK_UNMAPPED_BASE;
821	if (IS_ENABLED(CONFIG_RANDOMIZE_BASE))
822		poking_addr += (kaslr_get_random_long("Poking") & PAGE_MASK) %
823			(TASK_SIZE - TASK_UNMAPPED_BASE - 3 * PAGE_SIZE);
824
825	if (((poking_addr + PAGE_SIZE) & ~PMD_MASK) == 0)
826		poking_addr += PAGE_SIZE;
827
828	/*
829	 * We need to trigger the allocation of the page-tables that will be
830	 * needed for poking now. Later, poking may be performed in an atomic
831	 * section, which might cause allocation to fail.
832	 */
833	ptep = get_locked_pte(poking_mm, poking_addr, &ptl);
834	BUG_ON(!ptep);
835	pte_unmap_unlock(ptep, ptl);
836}
837
838/*
839 * devmem_is_allowed() checks to see if /dev/mem access to a certain address
840 * is valid. The argument is a physical page number.
841 *
842 * On x86, access has to be given to the first megabyte of RAM because that
843 * area traditionally contains BIOS code and data regions used by X, dosemu,
844 * and similar apps. Since they map the entire memory range, the whole range
845 * must be allowed (for mapping), but any areas that would otherwise be
846 * disallowed are flagged as being "zero filled" instead of rejected.
847 * Access has to be given to non-kernel-ram areas as well, these contain the
848 * PCI mmio resources as well as potential bios/acpi data regions.
849 */
850int devmem_is_allowed(unsigned long pagenr)
851{
852	if (region_intersects(PFN_PHYS(pagenr), PAGE_SIZE,
853				IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE)
854			!= REGION_DISJOINT) {
855		/*
856		 * For disallowed memory regions in the low 1MB range,
857		 * request that the page be shown as all zeros.
858		 */
859		if (pagenr < 256)
860			return 2;
861
862		return 0;
863	}
864
865	/*
866	 * This must follow RAM test, since System RAM is considered a
867	 * restricted resource under CONFIG_STRICT_IOMEM.
868	 */
869	if (iomem_is_exclusive(pagenr << PAGE_SHIFT)) {
870		/* Low 1MB bypasses iomem restrictions. */
871		if (pagenr < 256)
872			return 1;
873
874		return 0;
875	}
876
877	return 1;
878}
879
880void free_init_pages(const char *what, unsigned long begin, unsigned long end)
881{
882	unsigned long begin_aligned, end_aligned;
883
884	/* Make sure boundaries are page aligned */
885	begin_aligned = PAGE_ALIGN(begin);
886	end_aligned   = end & PAGE_MASK;
887
888	if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
889		begin = begin_aligned;
890		end   = end_aligned;
891	}
892
893	if (begin >= end)
894		return;
895
896	/*
897	 * If debugging page accesses then do not free this memory but
898	 * mark them not present - any buggy init-section access will
899	 * create a kernel page fault:
900	 */
901	if (debug_pagealloc_enabled()) {
902		pr_info("debug: unmapping init [mem %#010lx-%#010lx]\n",
903			begin, end - 1);
904		/*
905		 * Inform kmemleak about the hole in the memory since the
906		 * corresponding pages will be unmapped.
907		 */
908		kmemleak_free_part((void *)begin, end - begin);
909		set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
910	} else {
911		/*
912		 * We just marked the kernel text read only above, now that
913		 * we are going to free part of that, we need to make that
914		 * writeable and non-executable first.
915		 */
916		set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
917		set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
918
919		free_reserved_area((void *)begin, (void *)end,
920				   POISON_FREE_INITMEM, what);
921	}
922}
923
924/*
925 * begin/end can be in the direct map or the "high kernel mapping"
926 * used for the kernel image only.  free_init_pages() will do the
927 * right thing for either kind of address.
928 */
929void free_kernel_image_pages(const char *what, void *begin, void *end)
930{
931	unsigned long begin_ul = (unsigned long)begin;
932	unsigned long end_ul = (unsigned long)end;
933	unsigned long len_pages = (end_ul - begin_ul) >> PAGE_SHIFT;
934
935	free_init_pages(what, begin_ul, end_ul);
936
937	/*
938	 * PTI maps some of the kernel into userspace.  For performance,
939	 * this includes some kernel areas that do not contain secrets.
940	 * Those areas might be adjacent to the parts of the kernel image
941	 * being freed, which may contain secrets.  Remove the "high kernel
942	 * image mapping" for these freed areas, ensuring they are not even
943	 * potentially vulnerable to Meltdown regardless of the specific
944	 * optimizations PTI is currently using.
945	 *
946	 * The "noalias" prevents unmapping the direct map alias which is
947	 * needed to access the freed pages.
948	 *
949	 * This is only valid for 64bit kernels. 32bit has only one mapping
950	 * which can't be treated in this way for obvious reasons.
951	 */
952	if (IS_ENABLED(CONFIG_X86_64) && cpu_feature_enabled(X86_FEATURE_PTI))
953		set_memory_np_noalias(begin_ul, len_pages);
954}
955
956void __ref free_initmem(void)
957{
958	e820__reallocate_tables();
959
960	mem_encrypt_free_decrypted_mem();
961
962	free_kernel_image_pages("unused kernel image (initmem)",
963				&__init_begin, &__init_end);
964}
965
966#ifdef CONFIG_BLK_DEV_INITRD
967void __init free_initrd_mem(unsigned long start, unsigned long end)
968{
969	/*
970	 * end could be not aligned, and We can not align that,
971	 * decompresser could be confused by aligned initrd_end
972	 * We already reserve the end partial page before in
973	 *   - i386_start_kernel()
974	 *   - x86_64_start_kernel()
975	 *   - relocate_initrd()
976	 * So here We can do PAGE_ALIGN() safely to get partial page to be freed
977	 */
978	free_init_pages("initrd", start, PAGE_ALIGN(end));
979}
980#endif
981
982/*
983 * Calculate the precise size of the DMA zone (first 16 MB of RAM),
984 * and pass it to the MM layer - to help it set zone watermarks more
985 * accurately.
986 *
987 * Done on 64-bit systems only for the time being, although 32-bit systems
988 * might benefit from this as well.
989 */
990void __init memblock_find_dma_reserve(void)
991{
992#ifdef CONFIG_X86_64
993	u64 nr_pages = 0, nr_free_pages = 0;
994	unsigned long start_pfn, end_pfn;
995	phys_addr_t start_addr, end_addr;
996	int i;
997	u64 u;
998
999	/*
1000	 * Iterate over all memory ranges (free and reserved ones alike),
1001	 * to calculate the total number of pages in the first 16 MB of RAM:
1002	 */
1003	nr_pages = 0;
1004	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
1005		start_pfn = min(start_pfn, MAX_DMA_PFN);
1006		end_pfn   = min(end_pfn,   MAX_DMA_PFN);
1007
1008		nr_pages += end_pfn - start_pfn;
1009	}
1010
1011	/*
1012	 * Iterate over free memory ranges to calculate the number of free
1013	 * pages in the DMA zone, while not counting potential partial
1014	 * pages at the beginning or the end of the range:
1015	 */
1016	nr_free_pages = 0;
1017	for_each_free_mem_range(u, NUMA_NO_NODE, MEMBLOCK_NONE, &start_addr, &end_addr, NULL) {
1018		start_pfn = min_t(unsigned long, PFN_UP(start_addr), MAX_DMA_PFN);
1019		end_pfn   = min_t(unsigned long, PFN_DOWN(end_addr), MAX_DMA_PFN);
1020
1021		if (start_pfn < end_pfn)
1022			nr_free_pages += end_pfn - start_pfn;
1023	}
1024
1025	set_dma_reserve(nr_pages - nr_free_pages);
1026#endif
1027}
1028
1029void __init zone_sizes_init(void)
1030{
1031	unsigned long max_zone_pfns[MAX_NR_ZONES];
1032
1033	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1034
1035#ifdef CONFIG_ZONE_DMA
1036	max_zone_pfns[ZONE_DMA]		= min(MAX_DMA_PFN, max_low_pfn);
1037#endif
1038#ifdef CONFIG_ZONE_DMA32
1039	max_zone_pfns[ZONE_DMA32]	= min(MAX_DMA32_PFN, max_low_pfn);
1040#endif
1041	max_zone_pfns[ZONE_NORMAL]	= max_low_pfn;
1042#ifdef CONFIG_HIGHMEM
1043	max_zone_pfns[ZONE_HIGHMEM]	= max_pfn;
1044#endif
1045
1046	free_area_init(max_zone_pfns);
1047}
1048
1049__visible DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate) = {
1050	.loaded_mm = &init_mm,
1051	.next_asid = 1,
1052	.cr4 = ~0UL,	/* fail hard if we screw up cr4 shadow initialization */
1053};
1054
1055void update_cache_mode_entry(unsigned entry, enum page_cache_mode cache)
1056{
1057	/* entry 0 MUST be WB (hardwired to speed up translations) */
1058	BUG_ON(!entry && cache != _PAGE_CACHE_MODE_WB);
1059
1060	__cachemode2pte_tbl[cache] = __cm_idx2pte(entry);
1061	__pte2cachemode_tbl[entry] = cache;
1062}
1063
1064#ifdef CONFIG_SWAP
1065unsigned long max_swapfile_size(void)
1066{
1067	unsigned long pages;
1068
1069	pages = generic_max_swapfile_size();
1070
1071	if (boot_cpu_has_bug(X86_BUG_L1TF) && l1tf_mitigation != L1TF_MITIGATION_OFF) {
1072		/* Limit the swap file size to MAX_PA/2 for L1TF workaround */
1073		unsigned long long l1tf_limit = l1tf_pfn_limit();
1074		/*
1075		 * We encode swap offsets also with 3 bits below those for pfn
1076		 * which makes the usable limit higher.
1077		 */
1078#if CONFIG_PGTABLE_LEVELS > 2
1079		l1tf_limit <<= PAGE_SHIFT - SWP_OFFSET_FIRST_BIT;
1080#endif
1081		pages = min_t(unsigned long long, l1tf_limit, pages);
1082	}
1083	return pages;
1084}
1085#endif
1086