xref: /linux/arch/x86/mm/init_64.c (revision 987b741c52c7c6c68d46fbaeb95b8d1087f10b7f)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  linux/arch/x86_64/mm/init.c
4  *
5  *  Copyright (C) 1995  Linus Torvalds
6  *  Copyright (C) 2000  Pavel Machek <pavel@ucw.cz>
7  *  Copyright (C) 2002,2003 Andi Kleen <ak@suse.de>
8  */
9 
10 #include <linux/signal.h>
11 #include <linux/sched.h>
12 #include <linux/kernel.h>
13 #include <linux/errno.h>
14 #include <linux/string.h>
15 #include <linux/types.h>
16 #include <linux/ptrace.h>
17 #include <linux/mman.h>
18 #include <linux/mm.h>
19 #include <linux/swap.h>
20 #include <linux/smp.h>
21 #include <linux/init.h>
22 #include <linux/initrd.h>
23 #include <linux/pagemap.h>
24 #include <linux/memblock.h>
25 #include <linux/proc_fs.h>
26 #include <linux/pci.h>
27 #include <linux/pfn.h>
28 #include <linux/poison.h>
29 #include <linux/dma-mapping.h>
30 #include <linux/memory.h>
31 #include <linux/memory_hotplug.h>
32 #include <linux/memremap.h>
33 #include <linux/nmi.h>
34 #include <linux/gfp.h>
35 #include <linux/kcore.h>
36 
37 #include <asm/processor.h>
38 #include <asm/bios_ebda.h>
39 #include <linux/uaccess.h>
40 #include <asm/pgalloc.h>
41 #include <asm/dma.h>
42 #include <asm/fixmap.h>
43 #include <asm/e820/api.h>
44 #include <asm/apic.h>
45 #include <asm/tlb.h>
46 #include <asm/mmu_context.h>
47 #include <asm/proto.h>
48 #include <asm/smp.h>
49 #include <asm/sections.h>
50 #include <asm/kdebug.h>
51 #include <asm/numa.h>
52 #include <asm/set_memory.h>
53 #include <asm/init.h>
54 #include <asm/uv/uv.h>
55 #include <asm/setup.h>
56 #include <asm/ftrace.h>
57 
58 #include "mm_internal.h"
59 
60 #include "ident_map.c"
61 
62 #define DEFINE_POPULATE(fname, type1, type2, init)		\
63 static inline void fname##_init(struct mm_struct *mm,		\
64 		type1##_t *arg1, type2##_t *arg2, bool init)	\
65 {								\
66 	if (init)						\
67 		fname##_safe(mm, arg1, arg2);			\
68 	else							\
69 		fname(mm, arg1, arg2);				\
70 }
71 
72 DEFINE_POPULATE(p4d_populate, p4d, pud, init)
73 DEFINE_POPULATE(pgd_populate, pgd, p4d, init)
74 DEFINE_POPULATE(pud_populate, pud, pmd, init)
75 DEFINE_POPULATE(pmd_populate_kernel, pmd, pte, init)
76 
77 #define DEFINE_ENTRY(type1, type2, init)			\
78 static inline void set_##type1##_init(type1##_t *arg1,		\
79 			type2##_t arg2, bool init)		\
80 {								\
81 	if (init)						\
82 		set_##type1##_safe(arg1, arg2);			\
83 	else							\
84 		set_##type1(arg1, arg2);			\
85 }
86 
87 DEFINE_ENTRY(p4d, p4d, init)
88 DEFINE_ENTRY(pud, pud, init)
89 DEFINE_ENTRY(pmd, pmd, init)
90 DEFINE_ENTRY(pte, pte, init)
91 
92 
93 /*
94  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
95  * physical space so we can cache the place of the first one and move
96  * around without checking the pgd every time.
97  */
98 
99 /* Bits supported by the hardware: */
100 pteval_t __supported_pte_mask __read_mostly = ~0;
101 /* Bits allowed in normal kernel mappings: */
102 pteval_t __default_kernel_pte_mask __read_mostly = ~0;
103 EXPORT_SYMBOL_GPL(__supported_pte_mask);
104 /* Used in PAGE_KERNEL_* macros which are reasonably used out-of-tree: */
105 EXPORT_SYMBOL(__default_kernel_pte_mask);
106 
107 int force_personality32;
108 
109 /*
110  * noexec32=on|off
111  * Control non executable heap for 32bit processes.
112  * To control the stack too use noexec=off
113  *
114  * on	PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
115  * off	PROT_READ implies PROT_EXEC
116  */
117 static int __init nonx32_setup(char *str)
118 {
119 	if (!strcmp(str, "on"))
120 		force_personality32 &= ~READ_IMPLIES_EXEC;
121 	else if (!strcmp(str, "off"))
122 		force_personality32 |= READ_IMPLIES_EXEC;
123 	return 1;
124 }
125 __setup("noexec32=", nonx32_setup);
126 
127 static void sync_global_pgds_l5(unsigned long start, unsigned long end)
128 {
129 	unsigned long addr;
130 
131 	for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
132 		const pgd_t *pgd_ref = pgd_offset_k(addr);
133 		struct page *page;
134 
135 		/* Check for overflow */
136 		if (addr < start)
137 			break;
138 
139 		if (pgd_none(*pgd_ref))
140 			continue;
141 
142 		spin_lock(&pgd_lock);
143 		list_for_each_entry(page, &pgd_list, lru) {
144 			pgd_t *pgd;
145 			spinlock_t *pgt_lock;
146 
147 			pgd = (pgd_t *)page_address(page) + pgd_index(addr);
148 			/* the pgt_lock only for Xen */
149 			pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
150 			spin_lock(pgt_lock);
151 
152 			if (!pgd_none(*pgd_ref) && !pgd_none(*pgd))
153 				BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
154 
155 			if (pgd_none(*pgd))
156 				set_pgd(pgd, *pgd_ref);
157 
158 			spin_unlock(pgt_lock);
159 		}
160 		spin_unlock(&pgd_lock);
161 	}
162 }
163 
164 static void sync_global_pgds_l4(unsigned long start, unsigned long end)
165 {
166 	unsigned long addr;
167 
168 	for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
169 		pgd_t *pgd_ref = pgd_offset_k(addr);
170 		const p4d_t *p4d_ref;
171 		struct page *page;
172 
173 		/*
174 		 * With folded p4d, pgd_none() is always false, we need to
175 		 * handle synchronization on p4d level.
176 		 */
177 		MAYBE_BUILD_BUG_ON(pgd_none(*pgd_ref));
178 		p4d_ref = p4d_offset(pgd_ref, addr);
179 
180 		if (p4d_none(*p4d_ref))
181 			continue;
182 
183 		spin_lock(&pgd_lock);
184 		list_for_each_entry(page, &pgd_list, lru) {
185 			pgd_t *pgd;
186 			p4d_t *p4d;
187 			spinlock_t *pgt_lock;
188 
189 			pgd = (pgd_t *)page_address(page) + pgd_index(addr);
190 			p4d = p4d_offset(pgd, addr);
191 			/* the pgt_lock only for Xen */
192 			pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
193 			spin_lock(pgt_lock);
194 
195 			if (!p4d_none(*p4d_ref) && !p4d_none(*p4d))
196 				BUG_ON(p4d_page_vaddr(*p4d)
197 				       != p4d_page_vaddr(*p4d_ref));
198 
199 			if (p4d_none(*p4d))
200 				set_p4d(p4d, *p4d_ref);
201 
202 			spin_unlock(pgt_lock);
203 		}
204 		spin_unlock(&pgd_lock);
205 	}
206 }
207 
208 /*
209  * When memory was added make sure all the processes MM have
210  * suitable PGD entries in the local PGD level page.
211  */
212 static void sync_global_pgds(unsigned long start, unsigned long end)
213 {
214 	if (pgtable_l5_enabled())
215 		sync_global_pgds_l5(start, end);
216 	else
217 		sync_global_pgds_l4(start, end);
218 }
219 
220 /*
221  * NOTE: This function is marked __ref because it calls __init function
222  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
223  */
224 static __ref void *spp_getpage(void)
225 {
226 	void *ptr;
227 
228 	if (after_bootmem)
229 		ptr = (void *) get_zeroed_page(GFP_ATOMIC);
230 	else
231 		ptr = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
232 
233 	if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
234 		panic("set_pte_phys: cannot allocate page data %s\n",
235 			after_bootmem ? "after bootmem" : "");
236 	}
237 
238 	pr_debug("spp_getpage %p\n", ptr);
239 
240 	return ptr;
241 }
242 
243 static p4d_t *fill_p4d(pgd_t *pgd, unsigned long vaddr)
244 {
245 	if (pgd_none(*pgd)) {
246 		p4d_t *p4d = (p4d_t *)spp_getpage();
247 		pgd_populate(&init_mm, pgd, p4d);
248 		if (p4d != p4d_offset(pgd, 0))
249 			printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
250 			       p4d, p4d_offset(pgd, 0));
251 	}
252 	return p4d_offset(pgd, vaddr);
253 }
254 
255 static pud_t *fill_pud(p4d_t *p4d, unsigned long vaddr)
256 {
257 	if (p4d_none(*p4d)) {
258 		pud_t *pud = (pud_t *)spp_getpage();
259 		p4d_populate(&init_mm, p4d, pud);
260 		if (pud != pud_offset(p4d, 0))
261 			printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
262 			       pud, pud_offset(p4d, 0));
263 	}
264 	return pud_offset(p4d, vaddr);
265 }
266 
267 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
268 {
269 	if (pud_none(*pud)) {
270 		pmd_t *pmd = (pmd_t *) spp_getpage();
271 		pud_populate(&init_mm, pud, pmd);
272 		if (pmd != pmd_offset(pud, 0))
273 			printk(KERN_ERR "PAGETABLE BUG #02! %p <-> %p\n",
274 			       pmd, pmd_offset(pud, 0));
275 	}
276 	return pmd_offset(pud, vaddr);
277 }
278 
279 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
280 {
281 	if (pmd_none(*pmd)) {
282 		pte_t *pte = (pte_t *) spp_getpage();
283 		pmd_populate_kernel(&init_mm, pmd, pte);
284 		if (pte != pte_offset_kernel(pmd, 0))
285 			printk(KERN_ERR "PAGETABLE BUG #03!\n");
286 	}
287 	return pte_offset_kernel(pmd, vaddr);
288 }
289 
290 static void __set_pte_vaddr(pud_t *pud, unsigned long vaddr, pte_t new_pte)
291 {
292 	pmd_t *pmd = fill_pmd(pud, vaddr);
293 	pte_t *pte = fill_pte(pmd, vaddr);
294 
295 	set_pte(pte, new_pte);
296 
297 	/*
298 	 * It's enough to flush this one mapping.
299 	 * (PGE mappings get flushed as well)
300 	 */
301 	flush_tlb_one_kernel(vaddr);
302 }
303 
304 void set_pte_vaddr_p4d(p4d_t *p4d_page, unsigned long vaddr, pte_t new_pte)
305 {
306 	p4d_t *p4d = p4d_page + p4d_index(vaddr);
307 	pud_t *pud = fill_pud(p4d, vaddr);
308 
309 	__set_pte_vaddr(pud, vaddr, new_pte);
310 }
311 
312 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
313 {
314 	pud_t *pud = pud_page + pud_index(vaddr);
315 
316 	__set_pte_vaddr(pud, vaddr, new_pte);
317 }
318 
319 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
320 {
321 	pgd_t *pgd;
322 	p4d_t *p4d_page;
323 
324 	pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
325 
326 	pgd = pgd_offset_k(vaddr);
327 	if (pgd_none(*pgd)) {
328 		printk(KERN_ERR
329 			"PGD FIXMAP MISSING, it should be setup in head.S!\n");
330 		return;
331 	}
332 
333 	p4d_page = p4d_offset(pgd, 0);
334 	set_pte_vaddr_p4d(p4d_page, vaddr, pteval);
335 }
336 
337 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
338 {
339 	pgd_t *pgd;
340 	p4d_t *p4d;
341 	pud_t *pud;
342 
343 	pgd = pgd_offset_k(vaddr);
344 	p4d = fill_p4d(pgd, vaddr);
345 	pud = fill_pud(p4d, vaddr);
346 	return fill_pmd(pud, vaddr);
347 }
348 
349 pte_t * __init populate_extra_pte(unsigned long vaddr)
350 {
351 	pmd_t *pmd;
352 
353 	pmd = populate_extra_pmd(vaddr);
354 	return fill_pte(pmd, vaddr);
355 }
356 
357 /*
358  * Create large page table mappings for a range of physical addresses.
359  */
360 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
361 					enum page_cache_mode cache)
362 {
363 	pgd_t *pgd;
364 	p4d_t *p4d;
365 	pud_t *pud;
366 	pmd_t *pmd;
367 	pgprot_t prot;
368 
369 	pgprot_val(prot) = pgprot_val(PAGE_KERNEL_LARGE) |
370 		protval_4k_2_large(cachemode2protval(cache));
371 	BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
372 	for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
373 		pgd = pgd_offset_k((unsigned long)__va(phys));
374 		if (pgd_none(*pgd)) {
375 			p4d = (p4d_t *) spp_getpage();
376 			set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE |
377 						_PAGE_USER));
378 		}
379 		p4d = p4d_offset(pgd, (unsigned long)__va(phys));
380 		if (p4d_none(*p4d)) {
381 			pud = (pud_t *) spp_getpage();
382 			set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE |
383 						_PAGE_USER));
384 		}
385 		pud = pud_offset(p4d, (unsigned long)__va(phys));
386 		if (pud_none(*pud)) {
387 			pmd = (pmd_t *) spp_getpage();
388 			set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
389 						_PAGE_USER));
390 		}
391 		pmd = pmd_offset(pud, phys);
392 		BUG_ON(!pmd_none(*pmd));
393 		set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
394 	}
395 }
396 
397 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
398 {
399 	__init_extra_mapping(phys, size, _PAGE_CACHE_MODE_WB);
400 }
401 
402 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
403 {
404 	__init_extra_mapping(phys, size, _PAGE_CACHE_MODE_UC);
405 }
406 
407 /*
408  * The head.S code sets up the kernel high mapping:
409  *
410  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
411  *
412  * phys_base holds the negative offset to the kernel, which is added
413  * to the compile time generated pmds. This results in invalid pmds up
414  * to the point where we hit the physaddr 0 mapping.
415  *
416  * We limit the mappings to the region from _text to _brk_end.  _brk_end
417  * is rounded up to the 2MB boundary. This catches the invalid pmds as
418  * well, as they are located before _text:
419  */
420 void __init cleanup_highmap(void)
421 {
422 	unsigned long vaddr = __START_KERNEL_map;
423 	unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
424 	unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
425 	pmd_t *pmd = level2_kernel_pgt;
426 
427 	/*
428 	 * Native path, max_pfn_mapped is not set yet.
429 	 * Xen has valid max_pfn_mapped set in
430 	 *	arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
431 	 */
432 	if (max_pfn_mapped)
433 		vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
434 
435 	for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
436 		if (pmd_none(*pmd))
437 			continue;
438 		if (vaddr < (unsigned long) _text || vaddr > end)
439 			set_pmd(pmd, __pmd(0));
440 	}
441 }
442 
443 /*
444  * Create PTE level page table mapping for physical addresses.
445  * It returns the last physical address mapped.
446  */
447 static unsigned long __meminit
448 phys_pte_init(pte_t *pte_page, unsigned long paddr, unsigned long paddr_end,
449 	      pgprot_t prot, bool init)
450 {
451 	unsigned long pages = 0, paddr_next;
452 	unsigned long paddr_last = paddr_end;
453 	pte_t *pte;
454 	int i;
455 
456 	pte = pte_page + pte_index(paddr);
457 	i = pte_index(paddr);
458 
459 	for (; i < PTRS_PER_PTE; i++, paddr = paddr_next, pte++) {
460 		paddr_next = (paddr & PAGE_MASK) + PAGE_SIZE;
461 		if (paddr >= paddr_end) {
462 			if (!after_bootmem &&
463 			    !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
464 					     E820_TYPE_RAM) &&
465 			    !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
466 					     E820_TYPE_RESERVED_KERN))
467 				set_pte_init(pte, __pte(0), init);
468 			continue;
469 		}
470 
471 		/*
472 		 * We will re-use the existing mapping.
473 		 * Xen for example has some special requirements, like mapping
474 		 * pagetable pages as RO. So assume someone who pre-setup
475 		 * these mappings are more intelligent.
476 		 */
477 		if (!pte_none(*pte)) {
478 			if (!after_bootmem)
479 				pages++;
480 			continue;
481 		}
482 
483 		if (0)
484 			pr_info("   pte=%p addr=%lx pte=%016lx\n", pte, paddr,
485 				pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL).pte);
486 		pages++;
487 		set_pte_init(pte, pfn_pte(paddr >> PAGE_SHIFT, prot), init);
488 		paddr_last = (paddr & PAGE_MASK) + PAGE_SIZE;
489 	}
490 
491 	update_page_count(PG_LEVEL_4K, pages);
492 
493 	return paddr_last;
494 }
495 
496 /*
497  * Create PMD level page table mapping for physical addresses. The virtual
498  * and physical address have to be aligned at this level.
499  * It returns the last physical address mapped.
500  */
501 static unsigned long __meminit
502 phys_pmd_init(pmd_t *pmd_page, unsigned long paddr, unsigned long paddr_end,
503 	      unsigned long page_size_mask, pgprot_t prot, bool init)
504 {
505 	unsigned long pages = 0, paddr_next;
506 	unsigned long paddr_last = paddr_end;
507 
508 	int i = pmd_index(paddr);
509 
510 	for (; i < PTRS_PER_PMD; i++, paddr = paddr_next) {
511 		pmd_t *pmd = pmd_page + pmd_index(paddr);
512 		pte_t *pte;
513 		pgprot_t new_prot = prot;
514 
515 		paddr_next = (paddr & PMD_MASK) + PMD_SIZE;
516 		if (paddr >= paddr_end) {
517 			if (!after_bootmem &&
518 			    !e820__mapped_any(paddr & PMD_MASK, paddr_next,
519 					     E820_TYPE_RAM) &&
520 			    !e820__mapped_any(paddr & PMD_MASK, paddr_next,
521 					     E820_TYPE_RESERVED_KERN))
522 				set_pmd_init(pmd, __pmd(0), init);
523 			continue;
524 		}
525 
526 		if (!pmd_none(*pmd)) {
527 			if (!pmd_large(*pmd)) {
528 				spin_lock(&init_mm.page_table_lock);
529 				pte = (pte_t *)pmd_page_vaddr(*pmd);
530 				paddr_last = phys_pte_init(pte, paddr,
531 							   paddr_end, prot,
532 							   init);
533 				spin_unlock(&init_mm.page_table_lock);
534 				continue;
535 			}
536 			/*
537 			 * If we are ok with PG_LEVEL_2M mapping, then we will
538 			 * use the existing mapping,
539 			 *
540 			 * Otherwise, we will split the large page mapping but
541 			 * use the same existing protection bits except for
542 			 * large page, so that we don't violate Intel's TLB
543 			 * Application note (317080) which says, while changing
544 			 * the page sizes, new and old translations should
545 			 * not differ with respect to page frame and
546 			 * attributes.
547 			 */
548 			if (page_size_mask & (1 << PG_LEVEL_2M)) {
549 				if (!after_bootmem)
550 					pages++;
551 				paddr_last = paddr_next;
552 				continue;
553 			}
554 			new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
555 		}
556 
557 		if (page_size_mask & (1<<PG_LEVEL_2M)) {
558 			pages++;
559 			spin_lock(&init_mm.page_table_lock);
560 			set_pte_init((pte_t *)pmd,
561 				     pfn_pte((paddr & PMD_MASK) >> PAGE_SHIFT,
562 					     __pgprot(pgprot_val(prot) | _PAGE_PSE)),
563 				     init);
564 			spin_unlock(&init_mm.page_table_lock);
565 			paddr_last = paddr_next;
566 			continue;
567 		}
568 
569 		pte = alloc_low_page();
570 		paddr_last = phys_pte_init(pte, paddr, paddr_end, new_prot, init);
571 
572 		spin_lock(&init_mm.page_table_lock);
573 		pmd_populate_kernel_init(&init_mm, pmd, pte, init);
574 		spin_unlock(&init_mm.page_table_lock);
575 	}
576 	update_page_count(PG_LEVEL_2M, pages);
577 	return paddr_last;
578 }
579 
580 /*
581  * Create PUD level page table mapping for physical addresses. The virtual
582  * and physical address do not have to be aligned at this level. KASLR can
583  * randomize virtual addresses up to this level.
584  * It returns the last physical address mapped.
585  */
586 static unsigned long __meminit
587 phys_pud_init(pud_t *pud_page, unsigned long paddr, unsigned long paddr_end,
588 	      unsigned long page_size_mask, pgprot_t _prot, bool init)
589 {
590 	unsigned long pages = 0, paddr_next;
591 	unsigned long paddr_last = paddr_end;
592 	unsigned long vaddr = (unsigned long)__va(paddr);
593 	int i = pud_index(vaddr);
594 
595 	for (; i < PTRS_PER_PUD; i++, paddr = paddr_next) {
596 		pud_t *pud;
597 		pmd_t *pmd;
598 		pgprot_t prot = _prot;
599 
600 		vaddr = (unsigned long)__va(paddr);
601 		pud = pud_page + pud_index(vaddr);
602 		paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
603 
604 		if (paddr >= paddr_end) {
605 			if (!after_bootmem &&
606 			    !e820__mapped_any(paddr & PUD_MASK, paddr_next,
607 					     E820_TYPE_RAM) &&
608 			    !e820__mapped_any(paddr & PUD_MASK, paddr_next,
609 					     E820_TYPE_RESERVED_KERN))
610 				set_pud_init(pud, __pud(0), init);
611 			continue;
612 		}
613 
614 		if (!pud_none(*pud)) {
615 			if (!pud_large(*pud)) {
616 				pmd = pmd_offset(pud, 0);
617 				paddr_last = phys_pmd_init(pmd, paddr,
618 							   paddr_end,
619 							   page_size_mask,
620 							   prot, init);
621 				continue;
622 			}
623 			/*
624 			 * If we are ok with PG_LEVEL_1G mapping, then we will
625 			 * use the existing mapping.
626 			 *
627 			 * Otherwise, we will split the gbpage mapping but use
628 			 * the same existing protection  bits except for large
629 			 * page, so that we don't violate Intel's TLB
630 			 * Application note (317080) which says, while changing
631 			 * the page sizes, new and old translations should
632 			 * not differ with respect to page frame and
633 			 * attributes.
634 			 */
635 			if (page_size_mask & (1 << PG_LEVEL_1G)) {
636 				if (!after_bootmem)
637 					pages++;
638 				paddr_last = paddr_next;
639 				continue;
640 			}
641 			prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
642 		}
643 
644 		if (page_size_mask & (1<<PG_LEVEL_1G)) {
645 			pages++;
646 			spin_lock(&init_mm.page_table_lock);
647 
648 			prot = __pgprot(pgprot_val(prot) | __PAGE_KERNEL_LARGE);
649 
650 			set_pte_init((pte_t *)pud,
651 				     pfn_pte((paddr & PUD_MASK) >> PAGE_SHIFT,
652 					     prot),
653 				     init);
654 			spin_unlock(&init_mm.page_table_lock);
655 			paddr_last = paddr_next;
656 			continue;
657 		}
658 
659 		pmd = alloc_low_page();
660 		paddr_last = phys_pmd_init(pmd, paddr, paddr_end,
661 					   page_size_mask, prot, init);
662 
663 		spin_lock(&init_mm.page_table_lock);
664 		pud_populate_init(&init_mm, pud, pmd, init);
665 		spin_unlock(&init_mm.page_table_lock);
666 	}
667 
668 	update_page_count(PG_LEVEL_1G, pages);
669 
670 	return paddr_last;
671 }
672 
673 static unsigned long __meminit
674 phys_p4d_init(p4d_t *p4d_page, unsigned long paddr, unsigned long paddr_end,
675 	      unsigned long page_size_mask, pgprot_t prot, bool init)
676 {
677 	unsigned long vaddr, vaddr_end, vaddr_next, paddr_next, paddr_last;
678 
679 	paddr_last = paddr_end;
680 	vaddr = (unsigned long)__va(paddr);
681 	vaddr_end = (unsigned long)__va(paddr_end);
682 
683 	if (!pgtable_l5_enabled())
684 		return phys_pud_init((pud_t *) p4d_page, paddr, paddr_end,
685 				     page_size_mask, prot, init);
686 
687 	for (; vaddr < vaddr_end; vaddr = vaddr_next) {
688 		p4d_t *p4d = p4d_page + p4d_index(vaddr);
689 		pud_t *pud;
690 
691 		vaddr_next = (vaddr & P4D_MASK) + P4D_SIZE;
692 		paddr = __pa(vaddr);
693 
694 		if (paddr >= paddr_end) {
695 			paddr_next = __pa(vaddr_next);
696 			if (!after_bootmem &&
697 			    !e820__mapped_any(paddr & P4D_MASK, paddr_next,
698 					     E820_TYPE_RAM) &&
699 			    !e820__mapped_any(paddr & P4D_MASK, paddr_next,
700 					     E820_TYPE_RESERVED_KERN))
701 				set_p4d_init(p4d, __p4d(0), init);
702 			continue;
703 		}
704 
705 		if (!p4d_none(*p4d)) {
706 			pud = pud_offset(p4d, 0);
707 			paddr_last = phys_pud_init(pud, paddr, __pa(vaddr_end),
708 					page_size_mask, prot, init);
709 			continue;
710 		}
711 
712 		pud = alloc_low_page();
713 		paddr_last = phys_pud_init(pud, paddr, __pa(vaddr_end),
714 					   page_size_mask, prot, init);
715 
716 		spin_lock(&init_mm.page_table_lock);
717 		p4d_populate_init(&init_mm, p4d, pud, init);
718 		spin_unlock(&init_mm.page_table_lock);
719 	}
720 
721 	return paddr_last;
722 }
723 
724 static unsigned long __meminit
725 __kernel_physical_mapping_init(unsigned long paddr_start,
726 			       unsigned long paddr_end,
727 			       unsigned long page_size_mask,
728 			       pgprot_t prot, bool init)
729 {
730 	bool pgd_changed = false;
731 	unsigned long vaddr, vaddr_start, vaddr_end, vaddr_next, paddr_last;
732 
733 	paddr_last = paddr_end;
734 	vaddr = (unsigned long)__va(paddr_start);
735 	vaddr_end = (unsigned long)__va(paddr_end);
736 	vaddr_start = vaddr;
737 
738 	for (; vaddr < vaddr_end; vaddr = vaddr_next) {
739 		pgd_t *pgd = pgd_offset_k(vaddr);
740 		p4d_t *p4d;
741 
742 		vaddr_next = (vaddr & PGDIR_MASK) + PGDIR_SIZE;
743 
744 		if (pgd_val(*pgd)) {
745 			p4d = (p4d_t *)pgd_page_vaddr(*pgd);
746 			paddr_last = phys_p4d_init(p4d, __pa(vaddr),
747 						   __pa(vaddr_end),
748 						   page_size_mask,
749 						   prot, init);
750 			continue;
751 		}
752 
753 		p4d = alloc_low_page();
754 		paddr_last = phys_p4d_init(p4d, __pa(vaddr), __pa(vaddr_end),
755 					   page_size_mask, prot, init);
756 
757 		spin_lock(&init_mm.page_table_lock);
758 		if (pgtable_l5_enabled())
759 			pgd_populate_init(&init_mm, pgd, p4d, init);
760 		else
761 			p4d_populate_init(&init_mm, p4d_offset(pgd, vaddr),
762 					  (pud_t *) p4d, init);
763 
764 		spin_unlock(&init_mm.page_table_lock);
765 		pgd_changed = true;
766 	}
767 
768 	if (pgd_changed)
769 		sync_global_pgds(vaddr_start, vaddr_end - 1);
770 
771 	return paddr_last;
772 }
773 
774 
775 /*
776  * Create page table mapping for the physical memory for specific physical
777  * addresses. Note that it can only be used to populate non-present entries.
778  * The virtual and physical addresses have to be aligned on PMD level
779  * down. It returns the last physical address mapped.
780  */
781 unsigned long __meminit
782 kernel_physical_mapping_init(unsigned long paddr_start,
783 			     unsigned long paddr_end,
784 			     unsigned long page_size_mask, pgprot_t prot)
785 {
786 	return __kernel_physical_mapping_init(paddr_start, paddr_end,
787 					      page_size_mask, prot, true);
788 }
789 
790 /*
791  * This function is similar to kernel_physical_mapping_init() above with the
792  * exception that it uses set_{pud,pmd}() instead of the set_{pud,pte}_safe()
793  * when updating the mapping. The caller is responsible to flush the TLBs after
794  * the function returns.
795  */
796 unsigned long __meminit
797 kernel_physical_mapping_change(unsigned long paddr_start,
798 			       unsigned long paddr_end,
799 			       unsigned long page_size_mask)
800 {
801 	return __kernel_physical_mapping_init(paddr_start, paddr_end,
802 					      page_size_mask, PAGE_KERNEL,
803 					      false);
804 }
805 
806 #ifndef CONFIG_NUMA
807 void __init initmem_init(void)
808 {
809 	memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
810 }
811 #endif
812 
813 void __init paging_init(void)
814 {
815 	sparse_init();
816 
817 	/*
818 	 * clear the default setting with node 0
819 	 * note: don't use nodes_clear here, that is really clearing when
820 	 *	 numa support is not compiled in, and later node_set_state
821 	 *	 will not set it back.
822 	 */
823 	node_clear_state(0, N_MEMORY);
824 	node_clear_state(0, N_NORMAL_MEMORY);
825 
826 	zone_sizes_init();
827 }
828 
829 #ifdef CONFIG_SPARSEMEM_VMEMMAP
830 #define PAGE_UNUSED 0xFD
831 
832 /*
833  * The unused vmemmap range, which was not yet memset(PAGE_UNUSED), ranges
834  * from unused_pmd_start to next PMD_SIZE boundary.
835  */
836 static unsigned long unused_pmd_start __meminitdata;
837 
838 static void __meminit vmemmap_flush_unused_pmd(void)
839 {
840 	if (!unused_pmd_start)
841 		return;
842 	/*
843 	 * Clears (unused_pmd_start, PMD_END]
844 	 */
845 	memset((void *)unused_pmd_start, PAGE_UNUSED,
846 	       ALIGN(unused_pmd_start, PMD_SIZE) - unused_pmd_start);
847 	unused_pmd_start = 0;
848 }
849 
850 #ifdef CONFIG_MEMORY_HOTPLUG
851 /* Returns true if the PMD is completely unused and thus it can be freed */
852 static bool __meminit vmemmap_pmd_is_unused(unsigned long addr, unsigned long end)
853 {
854 	unsigned long start = ALIGN_DOWN(addr, PMD_SIZE);
855 
856 	/*
857 	 * Flush the unused range cache to ensure that memchr_inv() will work
858 	 * for the whole range.
859 	 */
860 	vmemmap_flush_unused_pmd();
861 	memset((void *)addr, PAGE_UNUSED, end - addr);
862 
863 	return !memchr_inv((void *)start, PAGE_UNUSED, PMD_SIZE);
864 }
865 #endif
866 
867 static void __meminit __vmemmap_use_sub_pmd(unsigned long start)
868 {
869 	/*
870 	 * As we expect to add in the same granularity as we remove, it's
871 	 * sufficient to mark only some piece used to block the memmap page from
872 	 * getting removed when removing some other adjacent memmap (just in
873 	 * case the first memmap never gets initialized e.g., because the memory
874 	 * block never gets onlined).
875 	 */
876 	memset((void *)start, 0, sizeof(struct page));
877 }
878 
879 static void __meminit vmemmap_use_sub_pmd(unsigned long start, unsigned long end)
880 {
881 	/*
882 	 * We only optimize if the new used range directly follows the
883 	 * previously unused range (esp., when populating consecutive sections).
884 	 */
885 	if (unused_pmd_start == start) {
886 		if (likely(IS_ALIGNED(end, PMD_SIZE)))
887 			unused_pmd_start = 0;
888 		else
889 			unused_pmd_start = end;
890 		return;
891 	}
892 
893 	/*
894 	 * If the range does not contiguously follows previous one, make sure
895 	 * to mark the unused range of the previous one so it can be removed.
896 	 */
897 	vmemmap_flush_unused_pmd();
898 	__vmemmap_use_sub_pmd(start);
899 }
900 
901 
902 static void __meminit vmemmap_use_new_sub_pmd(unsigned long start, unsigned long end)
903 {
904 	vmemmap_flush_unused_pmd();
905 
906 	/*
907 	 * Could be our memmap page is filled with PAGE_UNUSED already from a
908 	 * previous remove. Make sure to reset it.
909 	 */
910 	__vmemmap_use_sub_pmd(start);
911 
912 	/*
913 	 * Mark with PAGE_UNUSED the unused parts of the new memmap range
914 	 */
915 	if (!IS_ALIGNED(start, PMD_SIZE))
916 		memset((void *)start, PAGE_UNUSED,
917 			start - ALIGN_DOWN(start, PMD_SIZE));
918 
919 	/*
920 	 * We want to avoid memset(PAGE_UNUSED) when populating the vmemmap of
921 	 * consecutive sections. Remember for the last added PMD where the
922 	 * unused range begins.
923 	 */
924 	if (!IS_ALIGNED(end, PMD_SIZE))
925 		unused_pmd_start = end;
926 }
927 #endif
928 
929 /*
930  * Memory hotplug specific functions
931  */
932 #ifdef CONFIG_MEMORY_HOTPLUG
933 /*
934  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
935  * updating.
936  */
937 static void update_end_of_memory_vars(u64 start, u64 size)
938 {
939 	unsigned long end_pfn = PFN_UP(start + size);
940 
941 	if (end_pfn > max_pfn) {
942 		max_pfn = end_pfn;
943 		max_low_pfn = end_pfn;
944 		high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
945 	}
946 }
947 
948 int add_pages(int nid, unsigned long start_pfn, unsigned long nr_pages,
949 	      struct mhp_params *params)
950 {
951 	int ret;
952 
953 	ret = __add_pages(nid, start_pfn, nr_pages, params);
954 	WARN_ON_ONCE(ret);
955 
956 	/* update max_pfn, max_low_pfn and high_memory */
957 	update_end_of_memory_vars(start_pfn << PAGE_SHIFT,
958 				  nr_pages << PAGE_SHIFT);
959 
960 	return ret;
961 }
962 
963 int arch_add_memory(int nid, u64 start, u64 size,
964 		    struct mhp_params *params)
965 {
966 	unsigned long start_pfn = start >> PAGE_SHIFT;
967 	unsigned long nr_pages = size >> PAGE_SHIFT;
968 
969 	init_memory_mapping(start, start + size, params->pgprot);
970 
971 	return add_pages(nid, start_pfn, nr_pages, params);
972 }
973 
974 static void __meminit free_pagetable(struct page *page, int order)
975 {
976 	unsigned long magic;
977 	unsigned int nr_pages = 1 << order;
978 
979 	/* bootmem page has reserved flag */
980 	if (PageReserved(page)) {
981 		__ClearPageReserved(page);
982 
983 		magic = (unsigned long)page->freelist;
984 		if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
985 			while (nr_pages--)
986 				put_page_bootmem(page++);
987 		} else
988 			while (nr_pages--)
989 				free_reserved_page(page++);
990 	} else
991 		free_pages((unsigned long)page_address(page), order);
992 }
993 
994 static void __meminit free_hugepage_table(struct page *page,
995 		struct vmem_altmap *altmap)
996 {
997 	if (altmap)
998 		vmem_altmap_free(altmap, PMD_SIZE / PAGE_SIZE);
999 	else
1000 		free_pagetable(page, get_order(PMD_SIZE));
1001 }
1002 
1003 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
1004 {
1005 	pte_t *pte;
1006 	int i;
1007 
1008 	for (i = 0; i < PTRS_PER_PTE; i++) {
1009 		pte = pte_start + i;
1010 		if (!pte_none(*pte))
1011 			return;
1012 	}
1013 
1014 	/* free a pte talbe */
1015 	free_pagetable(pmd_page(*pmd), 0);
1016 	spin_lock(&init_mm.page_table_lock);
1017 	pmd_clear(pmd);
1018 	spin_unlock(&init_mm.page_table_lock);
1019 }
1020 
1021 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
1022 {
1023 	pmd_t *pmd;
1024 	int i;
1025 
1026 	for (i = 0; i < PTRS_PER_PMD; i++) {
1027 		pmd = pmd_start + i;
1028 		if (!pmd_none(*pmd))
1029 			return;
1030 	}
1031 
1032 	/* free a pmd talbe */
1033 	free_pagetable(pud_page(*pud), 0);
1034 	spin_lock(&init_mm.page_table_lock);
1035 	pud_clear(pud);
1036 	spin_unlock(&init_mm.page_table_lock);
1037 }
1038 
1039 static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
1040 {
1041 	pud_t *pud;
1042 	int i;
1043 
1044 	for (i = 0; i < PTRS_PER_PUD; i++) {
1045 		pud = pud_start + i;
1046 		if (!pud_none(*pud))
1047 			return;
1048 	}
1049 
1050 	/* free a pud talbe */
1051 	free_pagetable(p4d_page(*p4d), 0);
1052 	spin_lock(&init_mm.page_table_lock);
1053 	p4d_clear(p4d);
1054 	spin_unlock(&init_mm.page_table_lock);
1055 }
1056 
1057 static void __meminit
1058 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
1059 		 bool direct)
1060 {
1061 	unsigned long next, pages = 0;
1062 	pte_t *pte;
1063 	phys_addr_t phys_addr;
1064 
1065 	pte = pte_start + pte_index(addr);
1066 	for (; addr < end; addr = next, pte++) {
1067 		next = (addr + PAGE_SIZE) & PAGE_MASK;
1068 		if (next > end)
1069 			next = end;
1070 
1071 		if (!pte_present(*pte))
1072 			continue;
1073 
1074 		/*
1075 		 * We mapped [0,1G) memory as identity mapping when
1076 		 * initializing, in arch/x86/kernel/head_64.S. These
1077 		 * pagetables cannot be removed.
1078 		 */
1079 		phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
1080 		if (phys_addr < (phys_addr_t)0x40000000)
1081 			return;
1082 
1083 		if (!direct)
1084 			free_pagetable(pte_page(*pte), 0);
1085 
1086 		spin_lock(&init_mm.page_table_lock);
1087 		pte_clear(&init_mm, addr, pte);
1088 		spin_unlock(&init_mm.page_table_lock);
1089 
1090 		/* For non-direct mapping, pages means nothing. */
1091 		pages++;
1092 	}
1093 
1094 	/* Call free_pte_table() in remove_pmd_table(). */
1095 	flush_tlb_all();
1096 	if (direct)
1097 		update_page_count(PG_LEVEL_4K, -pages);
1098 }
1099 
1100 static void __meminit
1101 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
1102 		 bool direct, struct vmem_altmap *altmap)
1103 {
1104 	unsigned long next, pages = 0;
1105 	pte_t *pte_base;
1106 	pmd_t *pmd;
1107 
1108 	pmd = pmd_start + pmd_index(addr);
1109 	for (; addr < end; addr = next, pmd++) {
1110 		next = pmd_addr_end(addr, end);
1111 
1112 		if (!pmd_present(*pmd))
1113 			continue;
1114 
1115 		if (pmd_large(*pmd)) {
1116 			if (IS_ALIGNED(addr, PMD_SIZE) &&
1117 			    IS_ALIGNED(next, PMD_SIZE)) {
1118 				if (!direct)
1119 					free_hugepage_table(pmd_page(*pmd),
1120 							    altmap);
1121 
1122 				spin_lock(&init_mm.page_table_lock);
1123 				pmd_clear(pmd);
1124 				spin_unlock(&init_mm.page_table_lock);
1125 				pages++;
1126 			}
1127 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1128 			else if (vmemmap_pmd_is_unused(addr, next)) {
1129 					free_hugepage_table(pmd_page(*pmd),
1130 							    altmap);
1131 					spin_lock(&init_mm.page_table_lock);
1132 					pmd_clear(pmd);
1133 					spin_unlock(&init_mm.page_table_lock);
1134 			}
1135 #endif
1136 			continue;
1137 		}
1138 
1139 		pte_base = (pte_t *)pmd_page_vaddr(*pmd);
1140 		remove_pte_table(pte_base, addr, next, direct);
1141 		free_pte_table(pte_base, pmd);
1142 	}
1143 
1144 	/* Call free_pmd_table() in remove_pud_table(). */
1145 	if (direct)
1146 		update_page_count(PG_LEVEL_2M, -pages);
1147 }
1148 
1149 static void __meminit
1150 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
1151 		 struct vmem_altmap *altmap, bool direct)
1152 {
1153 	unsigned long next, pages = 0;
1154 	pmd_t *pmd_base;
1155 	pud_t *pud;
1156 
1157 	pud = pud_start + pud_index(addr);
1158 	for (; addr < end; addr = next, pud++) {
1159 		next = pud_addr_end(addr, end);
1160 
1161 		if (!pud_present(*pud))
1162 			continue;
1163 
1164 		if (pud_large(*pud) &&
1165 		    IS_ALIGNED(addr, PUD_SIZE) &&
1166 		    IS_ALIGNED(next, PUD_SIZE)) {
1167 			spin_lock(&init_mm.page_table_lock);
1168 			pud_clear(pud);
1169 			spin_unlock(&init_mm.page_table_lock);
1170 			pages++;
1171 			continue;
1172 		}
1173 
1174 		pmd_base = pmd_offset(pud, 0);
1175 		remove_pmd_table(pmd_base, addr, next, direct, altmap);
1176 		free_pmd_table(pmd_base, pud);
1177 	}
1178 
1179 	if (direct)
1180 		update_page_count(PG_LEVEL_1G, -pages);
1181 }
1182 
1183 static void __meminit
1184 remove_p4d_table(p4d_t *p4d_start, unsigned long addr, unsigned long end,
1185 		 struct vmem_altmap *altmap, bool direct)
1186 {
1187 	unsigned long next, pages = 0;
1188 	pud_t *pud_base;
1189 	p4d_t *p4d;
1190 
1191 	p4d = p4d_start + p4d_index(addr);
1192 	for (; addr < end; addr = next, p4d++) {
1193 		next = p4d_addr_end(addr, end);
1194 
1195 		if (!p4d_present(*p4d))
1196 			continue;
1197 
1198 		BUILD_BUG_ON(p4d_large(*p4d));
1199 
1200 		pud_base = pud_offset(p4d, 0);
1201 		remove_pud_table(pud_base, addr, next, altmap, direct);
1202 		/*
1203 		 * For 4-level page tables we do not want to free PUDs, but in the
1204 		 * 5-level case we should free them. This code will have to change
1205 		 * to adapt for boot-time switching between 4 and 5 level page tables.
1206 		 */
1207 		if (pgtable_l5_enabled())
1208 			free_pud_table(pud_base, p4d);
1209 	}
1210 
1211 	if (direct)
1212 		update_page_count(PG_LEVEL_512G, -pages);
1213 }
1214 
1215 /* start and end are both virtual address. */
1216 static void __meminit
1217 remove_pagetable(unsigned long start, unsigned long end, bool direct,
1218 		struct vmem_altmap *altmap)
1219 {
1220 	unsigned long next;
1221 	unsigned long addr;
1222 	pgd_t *pgd;
1223 	p4d_t *p4d;
1224 
1225 	for (addr = start; addr < end; addr = next) {
1226 		next = pgd_addr_end(addr, end);
1227 
1228 		pgd = pgd_offset_k(addr);
1229 		if (!pgd_present(*pgd))
1230 			continue;
1231 
1232 		p4d = p4d_offset(pgd, 0);
1233 		remove_p4d_table(p4d, addr, next, altmap, direct);
1234 	}
1235 
1236 	flush_tlb_all();
1237 }
1238 
1239 void __ref vmemmap_free(unsigned long start, unsigned long end,
1240 		struct vmem_altmap *altmap)
1241 {
1242 	VM_BUG_ON(!IS_ALIGNED(start, PAGE_SIZE));
1243 	VM_BUG_ON(!IS_ALIGNED(end, PAGE_SIZE));
1244 
1245 	remove_pagetable(start, end, false, altmap);
1246 }
1247 
1248 static void __meminit
1249 kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1250 {
1251 	start = (unsigned long)__va(start);
1252 	end = (unsigned long)__va(end);
1253 
1254 	remove_pagetable(start, end, true, NULL);
1255 }
1256 
1257 void __ref arch_remove_memory(int nid, u64 start, u64 size,
1258 			      struct vmem_altmap *altmap)
1259 {
1260 	unsigned long start_pfn = start >> PAGE_SHIFT;
1261 	unsigned long nr_pages = size >> PAGE_SHIFT;
1262 
1263 	__remove_pages(start_pfn, nr_pages, altmap);
1264 	kernel_physical_mapping_remove(start, start + size);
1265 }
1266 #endif /* CONFIG_MEMORY_HOTPLUG */
1267 
1268 static struct kcore_list kcore_vsyscall;
1269 
1270 static void __init register_page_bootmem_info(void)
1271 {
1272 #ifdef CONFIG_NUMA
1273 	int i;
1274 
1275 	for_each_online_node(i)
1276 		register_page_bootmem_info_node(NODE_DATA(i));
1277 #endif
1278 }
1279 
1280 /*
1281  * Pre-allocates page-table pages for the vmalloc area in the kernel page-table.
1282  * Only the level which needs to be synchronized between all page-tables is
1283  * allocated because the synchronization can be expensive.
1284  */
1285 static void __init preallocate_vmalloc_pages(void)
1286 {
1287 	unsigned long addr;
1288 	const char *lvl;
1289 
1290 	for (addr = VMALLOC_START; addr <= VMALLOC_END; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
1291 		pgd_t *pgd = pgd_offset_k(addr);
1292 		p4d_t *p4d;
1293 		pud_t *pud;
1294 
1295 		lvl = "p4d";
1296 		p4d = p4d_alloc(&init_mm, pgd, addr);
1297 		if (!p4d)
1298 			goto failed;
1299 
1300 		if (pgtable_l5_enabled())
1301 			continue;
1302 
1303 		/*
1304 		 * The goal here is to allocate all possibly required
1305 		 * hardware page tables pointed to by the top hardware
1306 		 * level.
1307 		 *
1308 		 * On 4-level systems, the P4D layer is folded away and
1309 		 * the above code does no preallocation.  Below, go down
1310 		 * to the pud _software_ level to ensure the second
1311 		 * hardware level is allocated on 4-level systems too.
1312 		 */
1313 		lvl = "pud";
1314 		pud = pud_alloc(&init_mm, p4d, addr);
1315 		if (!pud)
1316 			goto failed;
1317 	}
1318 
1319 	return;
1320 
1321 failed:
1322 
1323 	/*
1324 	 * The pages have to be there now or they will be missing in
1325 	 * process page-tables later.
1326 	 */
1327 	panic("Failed to pre-allocate %s pages for vmalloc area\n", lvl);
1328 }
1329 
1330 void __init mem_init(void)
1331 {
1332 	pci_iommu_alloc();
1333 
1334 	/* clear_bss() already clear the empty_zero_page */
1335 
1336 	/* this will put all memory onto the freelists */
1337 	memblock_free_all();
1338 	after_bootmem = 1;
1339 	x86_init.hyper.init_after_bootmem();
1340 
1341 	/*
1342 	 * Must be done after boot memory is put on freelist, because here we
1343 	 * might set fields in deferred struct pages that have not yet been
1344 	 * initialized, and memblock_free_all() initializes all the reserved
1345 	 * deferred pages for us.
1346 	 */
1347 	register_page_bootmem_info();
1348 
1349 	/* Register memory areas for /proc/kcore */
1350 	if (get_gate_vma(&init_mm))
1351 		kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR, PAGE_SIZE, KCORE_USER);
1352 
1353 	preallocate_vmalloc_pages();
1354 }
1355 
1356 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1357 int __init deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1358 {
1359 	/*
1360 	 * More CPUs always led to greater speedups on tested systems, up to
1361 	 * all the nodes' CPUs.  Use all since the system is otherwise idle
1362 	 * now.
1363 	 */
1364 	return max_t(int, cpumask_weight(node_cpumask), 1);
1365 }
1366 #endif
1367 
1368 int kernel_set_to_readonly;
1369 
1370 void mark_rodata_ro(void)
1371 {
1372 	unsigned long start = PFN_ALIGN(_text);
1373 	unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1374 	unsigned long end = (unsigned long)__end_rodata_hpage_align;
1375 	unsigned long text_end = PFN_ALIGN(_etext);
1376 	unsigned long rodata_end = PFN_ALIGN(__end_rodata);
1377 	unsigned long all_end;
1378 
1379 	printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1380 	       (end - start) >> 10);
1381 	set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1382 
1383 	kernel_set_to_readonly = 1;
1384 
1385 	/*
1386 	 * The rodata/data/bss/brk section (but not the kernel text!)
1387 	 * should also be not-executable.
1388 	 *
1389 	 * We align all_end to PMD_SIZE because the existing mapping
1390 	 * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1391 	 * split the PMD and the reminder between _brk_end and the end
1392 	 * of the PMD will remain mapped executable.
1393 	 *
1394 	 * Any PMD which was setup after the one which covers _brk_end
1395 	 * has been zapped already via cleanup_highmem().
1396 	 */
1397 	all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1398 	set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
1399 
1400 	set_ftrace_ops_ro();
1401 
1402 #ifdef CONFIG_CPA_DEBUG
1403 	printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1404 	set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1405 
1406 	printk(KERN_INFO "Testing CPA: again\n");
1407 	set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1408 #endif
1409 
1410 	free_kernel_image_pages("unused kernel image (text/rodata gap)",
1411 				(void *)text_end, (void *)rodata_start);
1412 	free_kernel_image_pages("unused kernel image (rodata/data gap)",
1413 				(void *)rodata_end, (void *)_sdata);
1414 
1415 	debug_checkwx();
1416 }
1417 
1418 int kern_addr_valid(unsigned long addr)
1419 {
1420 	unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1421 	pgd_t *pgd;
1422 	p4d_t *p4d;
1423 	pud_t *pud;
1424 	pmd_t *pmd;
1425 	pte_t *pte;
1426 
1427 	if (above != 0 && above != -1UL)
1428 		return 0;
1429 
1430 	pgd = pgd_offset_k(addr);
1431 	if (pgd_none(*pgd))
1432 		return 0;
1433 
1434 	p4d = p4d_offset(pgd, addr);
1435 	if (p4d_none(*p4d))
1436 		return 0;
1437 
1438 	pud = pud_offset(p4d, addr);
1439 	if (pud_none(*pud))
1440 		return 0;
1441 
1442 	if (pud_large(*pud))
1443 		return pfn_valid(pud_pfn(*pud));
1444 
1445 	pmd = pmd_offset(pud, addr);
1446 	if (pmd_none(*pmd))
1447 		return 0;
1448 
1449 	if (pmd_large(*pmd))
1450 		return pfn_valid(pmd_pfn(*pmd));
1451 
1452 	pte = pte_offset_kernel(pmd, addr);
1453 	if (pte_none(*pte))
1454 		return 0;
1455 
1456 	return pfn_valid(pte_pfn(*pte));
1457 }
1458 
1459 /*
1460  * Block size is the minimum amount of memory which can be hotplugged or
1461  * hotremoved. It must be power of two and must be equal or larger than
1462  * MIN_MEMORY_BLOCK_SIZE.
1463  */
1464 #define MAX_BLOCK_SIZE (2UL << 30)
1465 
1466 /* Amount of ram needed to start using large blocks */
1467 #define MEM_SIZE_FOR_LARGE_BLOCK (64UL << 30)
1468 
1469 /* Adjustable memory block size */
1470 static unsigned long set_memory_block_size;
1471 int __init set_memory_block_size_order(unsigned int order)
1472 {
1473 	unsigned long size = 1UL << order;
1474 
1475 	if (size > MEM_SIZE_FOR_LARGE_BLOCK || size < MIN_MEMORY_BLOCK_SIZE)
1476 		return -EINVAL;
1477 
1478 	set_memory_block_size = size;
1479 	return 0;
1480 }
1481 
1482 static unsigned long probe_memory_block_size(void)
1483 {
1484 	unsigned long boot_mem_end = max_pfn << PAGE_SHIFT;
1485 	unsigned long bz;
1486 
1487 	/* If memory block size has been set, then use it */
1488 	bz = set_memory_block_size;
1489 	if (bz)
1490 		goto done;
1491 
1492 	/* Use regular block if RAM is smaller than MEM_SIZE_FOR_LARGE_BLOCK */
1493 	if (boot_mem_end < MEM_SIZE_FOR_LARGE_BLOCK) {
1494 		bz = MIN_MEMORY_BLOCK_SIZE;
1495 		goto done;
1496 	}
1497 
1498 	/*
1499 	 * Use max block size to minimize overhead on bare metal, where
1500 	 * alignment for memory hotplug isn't a concern.
1501 	 */
1502 	if (!boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
1503 		bz = MAX_BLOCK_SIZE;
1504 		goto done;
1505 	}
1506 
1507 	/* Find the largest allowed block size that aligns to memory end */
1508 	for (bz = MAX_BLOCK_SIZE; bz > MIN_MEMORY_BLOCK_SIZE; bz >>= 1) {
1509 		if (IS_ALIGNED(boot_mem_end, bz))
1510 			break;
1511 	}
1512 done:
1513 	pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20);
1514 
1515 	return bz;
1516 }
1517 
1518 static unsigned long memory_block_size_probed;
1519 unsigned long memory_block_size_bytes(void)
1520 {
1521 	if (!memory_block_size_probed)
1522 		memory_block_size_probed = probe_memory_block_size();
1523 
1524 	return memory_block_size_probed;
1525 }
1526 
1527 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1528 /*
1529  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1530  */
1531 static long __meminitdata addr_start, addr_end;
1532 static void __meminitdata *p_start, *p_end;
1533 static int __meminitdata node_start;
1534 
1535 static int __meminit vmemmap_populate_hugepages(unsigned long start,
1536 		unsigned long end, int node, struct vmem_altmap *altmap)
1537 {
1538 	unsigned long addr;
1539 	unsigned long next;
1540 	pgd_t *pgd;
1541 	p4d_t *p4d;
1542 	pud_t *pud;
1543 	pmd_t *pmd;
1544 
1545 	for (addr = start; addr < end; addr = next) {
1546 		next = pmd_addr_end(addr, end);
1547 
1548 		pgd = vmemmap_pgd_populate(addr, node);
1549 		if (!pgd)
1550 			return -ENOMEM;
1551 
1552 		p4d = vmemmap_p4d_populate(pgd, addr, node);
1553 		if (!p4d)
1554 			return -ENOMEM;
1555 
1556 		pud = vmemmap_pud_populate(p4d, addr, node);
1557 		if (!pud)
1558 			return -ENOMEM;
1559 
1560 		pmd = pmd_offset(pud, addr);
1561 		if (pmd_none(*pmd)) {
1562 			void *p;
1563 
1564 			p = vmemmap_alloc_block_buf(PMD_SIZE, node, altmap);
1565 			if (p) {
1566 				pte_t entry;
1567 
1568 				entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1569 						PAGE_KERNEL_LARGE);
1570 				set_pmd(pmd, __pmd(pte_val(entry)));
1571 
1572 				/* check to see if we have contiguous blocks */
1573 				if (p_end != p || node_start != node) {
1574 					if (p_start)
1575 						pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1576 						       addr_start, addr_end-1, p_start, p_end-1, node_start);
1577 					addr_start = addr;
1578 					node_start = node;
1579 					p_start = p;
1580 				}
1581 
1582 				addr_end = addr + PMD_SIZE;
1583 				p_end = p + PMD_SIZE;
1584 
1585 				if (!IS_ALIGNED(addr, PMD_SIZE) ||
1586 				    !IS_ALIGNED(next, PMD_SIZE))
1587 					vmemmap_use_new_sub_pmd(addr, next);
1588 
1589 				continue;
1590 			} else if (altmap)
1591 				return -ENOMEM; /* no fallback */
1592 		} else if (pmd_large(*pmd)) {
1593 			vmemmap_verify((pte_t *)pmd, node, addr, next);
1594 			vmemmap_use_sub_pmd(addr, next);
1595 			continue;
1596 		}
1597 		if (vmemmap_populate_basepages(addr, next, node, NULL))
1598 			return -ENOMEM;
1599 	}
1600 	return 0;
1601 }
1602 
1603 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1604 		struct vmem_altmap *altmap)
1605 {
1606 	int err;
1607 
1608 	VM_BUG_ON(!IS_ALIGNED(start, PAGE_SIZE));
1609 	VM_BUG_ON(!IS_ALIGNED(end, PAGE_SIZE));
1610 
1611 	if (end - start < PAGES_PER_SECTION * sizeof(struct page))
1612 		err = vmemmap_populate_basepages(start, end, node, NULL);
1613 	else if (boot_cpu_has(X86_FEATURE_PSE))
1614 		err = vmemmap_populate_hugepages(start, end, node, altmap);
1615 	else if (altmap) {
1616 		pr_err_once("%s: no cpu support for altmap allocations\n",
1617 				__func__);
1618 		err = -ENOMEM;
1619 	} else
1620 		err = vmemmap_populate_basepages(start, end, node, NULL);
1621 	if (!err)
1622 		sync_global_pgds(start, end - 1);
1623 	return err;
1624 }
1625 
1626 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1627 void register_page_bootmem_memmap(unsigned long section_nr,
1628 				  struct page *start_page, unsigned long nr_pages)
1629 {
1630 	unsigned long addr = (unsigned long)start_page;
1631 	unsigned long end = (unsigned long)(start_page + nr_pages);
1632 	unsigned long next;
1633 	pgd_t *pgd;
1634 	p4d_t *p4d;
1635 	pud_t *pud;
1636 	pmd_t *pmd;
1637 	unsigned int nr_pmd_pages;
1638 	struct page *page;
1639 
1640 	for (; addr < end; addr = next) {
1641 		pte_t *pte = NULL;
1642 
1643 		pgd = pgd_offset_k(addr);
1644 		if (pgd_none(*pgd)) {
1645 			next = (addr + PAGE_SIZE) & PAGE_MASK;
1646 			continue;
1647 		}
1648 		get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1649 
1650 		p4d = p4d_offset(pgd, addr);
1651 		if (p4d_none(*p4d)) {
1652 			next = (addr + PAGE_SIZE) & PAGE_MASK;
1653 			continue;
1654 		}
1655 		get_page_bootmem(section_nr, p4d_page(*p4d), MIX_SECTION_INFO);
1656 
1657 		pud = pud_offset(p4d, addr);
1658 		if (pud_none(*pud)) {
1659 			next = (addr + PAGE_SIZE) & PAGE_MASK;
1660 			continue;
1661 		}
1662 		get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1663 
1664 		if (!boot_cpu_has(X86_FEATURE_PSE)) {
1665 			next = (addr + PAGE_SIZE) & PAGE_MASK;
1666 			pmd = pmd_offset(pud, addr);
1667 			if (pmd_none(*pmd))
1668 				continue;
1669 			get_page_bootmem(section_nr, pmd_page(*pmd),
1670 					 MIX_SECTION_INFO);
1671 
1672 			pte = pte_offset_kernel(pmd, addr);
1673 			if (pte_none(*pte))
1674 				continue;
1675 			get_page_bootmem(section_nr, pte_page(*pte),
1676 					 SECTION_INFO);
1677 		} else {
1678 			next = pmd_addr_end(addr, end);
1679 
1680 			pmd = pmd_offset(pud, addr);
1681 			if (pmd_none(*pmd))
1682 				continue;
1683 
1684 			nr_pmd_pages = 1 << get_order(PMD_SIZE);
1685 			page = pmd_page(*pmd);
1686 			while (nr_pmd_pages--)
1687 				get_page_bootmem(section_nr, page++,
1688 						 SECTION_INFO);
1689 		}
1690 	}
1691 }
1692 #endif
1693 
1694 void __meminit vmemmap_populate_print_last(void)
1695 {
1696 	if (p_start) {
1697 		pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1698 			addr_start, addr_end-1, p_start, p_end-1, node_start);
1699 		p_start = NULL;
1700 		p_end = NULL;
1701 		node_start = 0;
1702 	}
1703 }
1704 #endif
1705