xref: /linux/arch/x86/mm/init_64.c (revision 34e3c4500cdc06094b37a41b622598098308ba8f)
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 #include <linux/bootmem_info.h>
37 
38 #include <asm/processor.h>
39 #include <asm/bios_ebda.h>
40 #include <linux/uaccess.h>
41 #include <asm/pgalloc.h>
42 #include <asm/dma.h>
43 #include <asm/fixmap.h>
44 #include <asm/e820/api.h>
45 #include <asm/apic.h>
46 #include <asm/tlb.h>
47 #include <asm/mmu_context.h>
48 #include <asm/proto.h>
49 #include <asm/smp.h>
50 #include <asm/sections.h>
51 #include <asm/kdebug.h>
52 #include <asm/numa.h>
53 #include <asm/set_memory.h>
54 #include <asm/init.h>
55 #include <asm/uv/uv.h>
56 #include <asm/setup.h>
57 #include <asm/ftrace.h>
58 
59 #include "mm_internal.h"
60 
61 #include "ident_map.c"
62 
63 #define DEFINE_POPULATE(fname, type1, type2, init)		\
64 static inline void fname##_init(struct mm_struct *mm,		\
65 		type1##_t *arg1, type2##_t *arg2, bool init)	\
66 {								\
67 	if (init)						\
68 		fname##_safe(mm, arg1, arg2);			\
69 	else							\
70 		fname(mm, arg1, arg2);				\
71 }
72 
73 DEFINE_POPULATE(p4d_populate, p4d, pud, init)
74 DEFINE_POPULATE(pgd_populate, pgd, p4d, init)
75 DEFINE_POPULATE(pud_populate, pud, pmd, init)
76 DEFINE_POPULATE(pmd_populate_kernel, pmd, pte, init)
77 
78 #define DEFINE_ENTRY(type1, type2, init)			\
79 static inline void set_##type1##_init(type1##_t *arg1,		\
80 			type2##_t arg2, bool init)		\
81 {								\
82 	if (init)						\
83 		set_##type1##_safe(arg1, arg2);			\
84 	else							\
85 		set_##type1(arg1, arg2);			\
86 }
87 
88 DEFINE_ENTRY(p4d, p4d, init)
89 DEFINE_ENTRY(pud, pud, init)
90 DEFINE_ENTRY(pmd, pmd, init)
91 DEFINE_ENTRY(pte, pte, init)
92 
93 static inline pgprot_t prot_sethuge(pgprot_t prot)
94 {
95 	WARN_ON_ONCE(pgprot_val(prot) & _PAGE_PAT);
96 
97 	return __pgprot(pgprot_val(prot) | _PAGE_PSE);
98 }
99 
100 /*
101  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
102  * physical space so we can cache the place of the first one and move
103  * around without checking the pgd every time.
104  */
105 
106 /* Bits supported by the hardware: */
107 pteval_t __supported_pte_mask __read_mostly = ~0;
108 /* Bits allowed in normal kernel mappings: */
109 pteval_t __default_kernel_pte_mask __read_mostly = ~0;
110 EXPORT_SYMBOL_GPL(__supported_pte_mask);
111 /* Used in PAGE_KERNEL_* macros which are reasonably used out-of-tree: */
112 EXPORT_SYMBOL(__default_kernel_pte_mask);
113 
114 int force_personality32;
115 
116 /*
117  * noexec32=on|off
118  * Control non executable heap for 32bit processes.
119  *
120  * on	PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
121  * off	PROT_READ implies PROT_EXEC
122  */
123 static int __init nonx32_setup(char *str)
124 {
125 	if (!strcmp(str, "on"))
126 		force_personality32 &= ~READ_IMPLIES_EXEC;
127 	else if (!strcmp(str, "off"))
128 		force_personality32 |= READ_IMPLIES_EXEC;
129 	return 1;
130 }
131 __setup("noexec32=", nonx32_setup);
132 
133 static void sync_global_pgds_l5(unsigned long start, unsigned long end)
134 {
135 	unsigned long addr;
136 
137 	for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
138 		const pgd_t *pgd_ref = pgd_offset_k(addr);
139 		struct page *page;
140 
141 		/* Check for overflow */
142 		if (addr < start)
143 			break;
144 
145 		if (pgd_none(*pgd_ref))
146 			continue;
147 
148 		spin_lock(&pgd_lock);
149 		list_for_each_entry(page, &pgd_list, lru) {
150 			pgd_t *pgd;
151 			spinlock_t *pgt_lock;
152 
153 			pgd = (pgd_t *)page_address(page) + pgd_index(addr);
154 			/* the pgt_lock only for Xen */
155 			pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
156 			spin_lock(pgt_lock);
157 
158 			if (!pgd_none(*pgd_ref) && !pgd_none(*pgd))
159 				BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
160 
161 			if (pgd_none(*pgd))
162 				set_pgd(pgd, *pgd_ref);
163 
164 			spin_unlock(pgt_lock);
165 		}
166 		spin_unlock(&pgd_lock);
167 	}
168 }
169 
170 static void sync_global_pgds_l4(unsigned long start, unsigned long end)
171 {
172 	unsigned long addr;
173 
174 	for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
175 		pgd_t *pgd_ref = pgd_offset_k(addr);
176 		const p4d_t *p4d_ref;
177 		struct page *page;
178 
179 		/*
180 		 * With folded p4d, pgd_none() is always false, we need to
181 		 * handle synchronization on p4d level.
182 		 */
183 		MAYBE_BUILD_BUG_ON(pgd_none(*pgd_ref));
184 		p4d_ref = p4d_offset(pgd_ref, addr);
185 
186 		if (p4d_none(*p4d_ref))
187 			continue;
188 
189 		spin_lock(&pgd_lock);
190 		list_for_each_entry(page, &pgd_list, lru) {
191 			pgd_t *pgd;
192 			p4d_t *p4d;
193 			spinlock_t *pgt_lock;
194 
195 			pgd = (pgd_t *)page_address(page) + pgd_index(addr);
196 			p4d = p4d_offset(pgd, addr);
197 			/* the pgt_lock only for Xen */
198 			pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
199 			spin_lock(pgt_lock);
200 
201 			if (!p4d_none(*p4d_ref) && !p4d_none(*p4d))
202 				BUG_ON(p4d_pgtable(*p4d)
203 				       != p4d_pgtable(*p4d_ref));
204 
205 			if (p4d_none(*p4d))
206 				set_p4d(p4d, *p4d_ref);
207 
208 			spin_unlock(pgt_lock);
209 		}
210 		spin_unlock(&pgd_lock);
211 	}
212 }
213 
214 /*
215  * When memory was added make sure all the processes MM have
216  * suitable PGD entries in the local PGD level page.
217  */
218 static void sync_global_pgds(unsigned long start, unsigned long end)
219 {
220 	if (pgtable_l5_enabled())
221 		sync_global_pgds_l5(start, end);
222 	else
223 		sync_global_pgds_l4(start, end);
224 }
225 
226 /*
227  * NOTE: This function is marked __ref because it calls __init function
228  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
229  */
230 static __ref void *spp_getpage(void)
231 {
232 	void *ptr;
233 
234 	if (after_bootmem)
235 		ptr = (void *) get_zeroed_page(GFP_ATOMIC);
236 	else
237 		ptr = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
238 
239 	if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
240 		panic("set_pte_phys: cannot allocate page data %s\n",
241 			after_bootmem ? "after bootmem" : "");
242 	}
243 
244 	pr_debug("spp_getpage %p\n", ptr);
245 
246 	return ptr;
247 }
248 
249 static p4d_t *fill_p4d(pgd_t *pgd, unsigned long vaddr)
250 {
251 	if (pgd_none(*pgd)) {
252 		p4d_t *p4d = (p4d_t *)spp_getpage();
253 		pgd_populate(&init_mm, pgd, p4d);
254 		if (p4d != p4d_offset(pgd, 0))
255 			printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
256 			       p4d, p4d_offset(pgd, 0));
257 	}
258 	return p4d_offset(pgd, vaddr);
259 }
260 
261 static pud_t *fill_pud(p4d_t *p4d, unsigned long vaddr)
262 {
263 	if (p4d_none(*p4d)) {
264 		pud_t *pud = (pud_t *)spp_getpage();
265 		p4d_populate(&init_mm, p4d, pud);
266 		if (pud != pud_offset(p4d, 0))
267 			printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
268 			       pud, pud_offset(p4d, 0));
269 	}
270 	return pud_offset(p4d, vaddr);
271 }
272 
273 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
274 {
275 	if (pud_none(*pud)) {
276 		pmd_t *pmd = (pmd_t *) spp_getpage();
277 		pud_populate(&init_mm, pud, pmd);
278 		if (pmd != pmd_offset(pud, 0))
279 			printk(KERN_ERR "PAGETABLE BUG #02! %p <-> %p\n",
280 			       pmd, pmd_offset(pud, 0));
281 	}
282 	return pmd_offset(pud, vaddr);
283 }
284 
285 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
286 {
287 	if (pmd_none(*pmd)) {
288 		pte_t *pte = (pte_t *) spp_getpage();
289 		pmd_populate_kernel(&init_mm, pmd, pte);
290 		if (pte != pte_offset_kernel(pmd, 0))
291 			printk(KERN_ERR "PAGETABLE BUG #03!\n");
292 	}
293 	return pte_offset_kernel(pmd, vaddr);
294 }
295 
296 static void __set_pte_vaddr(pud_t *pud, unsigned long vaddr, pte_t new_pte)
297 {
298 	pmd_t *pmd = fill_pmd(pud, vaddr);
299 	pte_t *pte = fill_pte(pmd, vaddr);
300 
301 	set_pte(pte, new_pte);
302 
303 	/*
304 	 * It's enough to flush this one mapping.
305 	 * (PGE mappings get flushed as well)
306 	 */
307 	flush_tlb_one_kernel(vaddr);
308 }
309 
310 void set_pte_vaddr_p4d(p4d_t *p4d_page, unsigned long vaddr, pte_t new_pte)
311 {
312 	p4d_t *p4d = p4d_page + p4d_index(vaddr);
313 	pud_t *pud = fill_pud(p4d, vaddr);
314 
315 	__set_pte_vaddr(pud, vaddr, new_pte);
316 }
317 
318 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
319 {
320 	pud_t *pud = pud_page + pud_index(vaddr);
321 
322 	__set_pte_vaddr(pud, vaddr, new_pte);
323 }
324 
325 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
326 {
327 	pgd_t *pgd;
328 	p4d_t *p4d_page;
329 
330 	pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
331 
332 	pgd = pgd_offset_k(vaddr);
333 	if (pgd_none(*pgd)) {
334 		printk(KERN_ERR
335 			"PGD FIXMAP MISSING, it should be setup in head.S!\n");
336 		return;
337 	}
338 
339 	p4d_page = p4d_offset(pgd, 0);
340 	set_pte_vaddr_p4d(p4d_page, vaddr, pteval);
341 }
342 
343 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
344 {
345 	pgd_t *pgd;
346 	p4d_t *p4d;
347 	pud_t *pud;
348 
349 	pgd = pgd_offset_k(vaddr);
350 	p4d = fill_p4d(pgd, vaddr);
351 	pud = fill_pud(p4d, vaddr);
352 	return fill_pmd(pud, vaddr);
353 }
354 
355 pte_t * __init populate_extra_pte(unsigned long vaddr)
356 {
357 	pmd_t *pmd;
358 
359 	pmd = populate_extra_pmd(vaddr);
360 	return fill_pte(pmd, vaddr);
361 }
362 
363 /*
364  * Create large page table mappings for a range of physical addresses.
365  */
366 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
367 					enum page_cache_mode cache)
368 {
369 	pgd_t *pgd;
370 	p4d_t *p4d;
371 	pud_t *pud;
372 	pmd_t *pmd;
373 	pgprot_t prot;
374 
375 	pgprot_val(prot) = pgprot_val(PAGE_KERNEL_LARGE) |
376 		protval_4k_2_large(cachemode2protval(cache));
377 	BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
378 	for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
379 		pgd = pgd_offset_k((unsigned long)__va(phys));
380 		if (pgd_none(*pgd)) {
381 			p4d = (p4d_t *) spp_getpage();
382 			set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE |
383 						_PAGE_USER));
384 		}
385 		p4d = p4d_offset(pgd, (unsigned long)__va(phys));
386 		if (p4d_none(*p4d)) {
387 			pud = (pud_t *) spp_getpage();
388 			set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE |
389 						_PAGE_USER));
390 		}
391 		pud = pud_offset(p4d, (unsigned long)__va(phys));
392 		if (pud_none(*pud)) {
393 			pmd = (pmd_t *) spp_getpage();
394 			set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
395 						_PAGE_USER));
396 		}
397 		pmd = pmd_offset(pud, phys);
398 		BUG_ON(!pmd_none(*pmd));
399 		set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
400 	}
401 }
402 
403 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
404 {
405 	__init_extra_mapping(phys, size, _PAGE_CACHE_MODE_WB);
406 }
407 
408 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
409 {
410 	__init_extra_mapping(phys, size, _PAGE_CACHE_MODE_UC);
411 }
412 
413 /*
414  * The head.S code sets up the kernel high mapping:
415  *
416  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
417  *
418  * phys_base holds the negative offset to the kernel, which is added
419  * to the compile time generated pmds. This results in invalid pmds up
420  * to the point where we hit the physaddr 0 mapping.
421  *
422  * We limit the mappings to the region from _text to _brk_end.  _brk_end
423  * is rounded up to the 2MB boundary. This catches the invalid pmds as
424  * well, as they are located before _text:
425  */
426 void __init cleanup_highmap(void)
427 {
428 	unsigned long vaddr = __START_KERNEL_map;
429 	unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
430 	unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
431 	pmd_t *pmd = level2_kernel_pgt;
432 
433 	/*
434 	 * Native path, max_pfn_mapped is not set yet.
435 	 * Xen has valid max_pfn_mapped set in
436 	 *	arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
437 	 */
438 	if (max_pfn_mapped)
439 		vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
440 
441 	for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
442 		if (pmd_none(*pmd))
443 			continue;
444 		if (vaddr < (unsigned long) _text || vaddr > end)
445 			set_pmd(pmd, __pmd(0));
446 	}
447 }
448 
449 /*
450  * Create PTE level page table mapping for physical addresses.
451  * It returns the last physical address mapped.
452  */
453 static unsigned long __meminit
454 phys_pte_init(pte_t *pte_page, unsigned long paddr, unsigned long paddr_end,
455 	      pgprot_t prot, bool init)
456 {
457 	unsigned long pages = 0, paddr_next;
458 	unsigned long paddr_last = paddr_end;
459 	pte_t *pte;
460 	int i;
461 
462 	pte = pte_page + pte_index(paddr);
463 	i = pte_index(paddr);
464 
465 	for (; i < PTRS_PER_PTE; i++, paddr = paddr_next, pte++) {
466 		paddr_next = (paddr & PAGE_MASK) + PAGE_SIZE;
467 		if (paddr >= paddr_end) {
468 			if (!after_bootmem &&
469 			    !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
470 					     E820_TYPE_RAM) &&
471 			    !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
472 					     E820_TYPE_RESERVED_KERN) &&
473 			    !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
474 					     E820_TYPE_ACPI))
475 				set_pte_init(pte, __pte(0), init);
476 			continue;
477 		}
478 
479 		/*
480 		 * We will re-use the existing mapping.
481 		 * Xen for example has some special requirements, like mapping
482 		 * pagetable pages as RO. So assume someone who pre-setup
483 		 * these mappings are more intelligent.
484 		 */
485 		if (!pte_none(*pte)) {
486 			if (!after_bootmem)
487 				pages++;
488 			continue;
489 		}
490 
491 		if (0)
492 			pr_info("   pte=%p addr=%lx pte=%016lx\n", pte, paddr,
493 				pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL).pte);
494 		pages++;
495 		set_pte_init(pte, pfn_pte(paddr >> PAGE_SHIFT, prot), init);
496 		paddr_last = (paddr & PAGE_MASK) + PAGE_SIZE;
497 	}
498 
499 	update_page_count(PG_LEVEL_4K, pages);
500 
501 	return paddr_last;
502 }
503 
504 /*
505  * Create PMD level page table mapping for physical addresses. The virtual
506  * and physical address have to be aligned at this level.
507  * It returns the last physical address mapped.
508  */
509 static unsigned long __meminit
510 phys_pmd_init(pmd_t *pmd_page, unsigned long paddr, unsigned long paddr_end,
511 	      unsigned long page_size_mask, pgprot_t prot, bool init)
512 {
513 	unsigned long pages = 0, paddr_next;
514 	unsigned long paddr_last = paddr_end;
515 
516 	int i = pmd_index(paddr);
517 
518 	for (; i < PTRS_PER_PMD; i++, paddr = paddr_next) {
519 		pmd_t *pmd = pmd_page + pmd_index(paddr);
520 		pte_t *pte;
521 		pgprot_t new_prot = prot;
522 
523 		paddr_next = (paddr & PMD_MASK) + PMD_SIZE;
524 		if (paddr >= paddr_end) {
525 			if (!after_bootmem &&
526 			    !e820__mapped_any(paddr & PMD_MASK, paddr_next,
527 					     E820_TYPE_RAM) &&
528 			    !e820__mapped_any(paddr & PMD_MASK, paddr_next,
529 					     E820_TYPE_RESERVED_KERN) &&
530 			    !e820__mapped_any(paddr & PMD_MASK, paddr_next,
531 					     E820_TYPE_ACPI))
532 				set_pmd_init(pmd, __pmd(0), init);
533 			continue;
534 		}
535 
536 		if (!pmd_none(*pmd)) {
537 			if (!pmd_leaf(*pmd)) {
538 				spin_lock(&init_mm.page_table_lock);
539 				pte = (pte_t *)pmd_page_vaddr(*pmd);
540 				paddr_last = phys_pte_init(pte, paddr,
541 							   paddr_end, prot,
542 							   init);
543 				spin_unlock(&init_mm.page_table_lock);
544 				continue;
545 			}
546 			/*
547 			 * If we are ok with PG_LEVEL_2M mapping, then we will
548 			 * use the existing mapping,
549 			 *
550 			 * Otherwise, we will split the large page mapping but
551 			 * use the same existing protection bits except for
552 			 * large page, so that we don't violate Intel's TLB
553 			 * Application note (317080) which says, while changing
554 			 * the page sizes, new and old translations should
555 			 * not differ with respect to page frame and
556 			 * attributes.
557 			 */
558 			if (page_size_mask & (1 << PG_LEVEL_2M)) {
559 				if (!after_bootmem)
560 					pages++;
561 				paddr_last = paddr_next;
562 				continue;
563 			}
564 			new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
565 		}
566 
567 		if (page_size_mask & (1<<PG_LEVEL_2M)) {
568 			pages++;
569 			spin_lock(&init_mm.page_table_lock);
570 			set_pmd_init(pmd,
571 				     pfn_pmd(paddr >> PAGE_SHIFT, prot_sethuge(prot)),
572 				     init);
573 			spin_unlock(&init_mm.page_table_lock);
574 			paddr_last = paddr_next;
575 			continue;
576 		}
577 
578 		pte = alloc_low_page();
579 		paddr_last = phys_pte_init(pte, paddr, paddr_end, new_prot, init);
580 
581 		spin_lock(&init_mm.page_table_lock);
582 		pmd_populate_kernel_init(&init_mm, pmd, pte, init);
583 		spin_unlock(&init_mm.page_table_lock);
584 	}
585 	update_page_count(PG_LEVEL_2M, pages);
586 	return paddr_last;
587 }
588 
589 /*
590  * Create PUD level page table mapping for physical addresses. The virtual
591  * and physical address do not have to be aligned at this level. KASLR can
592  * randomize virtual addresses up to this level.
593  * It returns the last physical address mapped.
594  */
595 static unsigned long __meminit
596 phys_pud_init(pud_t *pud_page, unsigned long paddr, unsigned long paddr_end,
597 	      unsigned long page_size_mask, pgprot_t _prot, bool init)
598 {
599 	unsigned long pages = 0, paddr_next;
600 	unsigned long paddr_last = paddr_end;
601 	unsigned long vaddr = (unsigned long)__va(paddr);
602 	int i = pud_index(vaddr);
603 
604 	for (; i < PTRS_PER_PUD; i++, paddr = paddr_next) {
605 		pud_t *pud;
606 		pmd_t *pmd;
607 		pgprot_t prot = _prot;
608 
609 		vaddr = (unsigned long)__va(paddr);
610 		pud = pud_page + pud_index(vaddr);
611 		paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
612 
613 		if (paddr >= paddr_end) {
614 			if (!after_bootmem &&
615 			    !e820__mapped_any(paddr & PUD_MASK, paddr_next,
616 					     E820_TYPE_RAM) &&
617 			    !e820__mapped_any(paddr & PUD_MASK, paddr_next,
618 					     E820_TYPE_RESERVED_KERN) &&
619 			    !e820__mapped_any(paddr & PUD_MASK, paddr_next,
620 					     E820_TYPE_ACPI))
621 				set_pud_init(pud, __pud(0), init);
622 			continue;
623 		}
624 
625 		if (!pud_none(*pud)) {
626 			if (!pud_leaf(*pud)) {
627 				pmd = pmd_offset(pud, 0);
628 				paddr_last = phys_pmd_init(pmd, paddr,
629 							   paddr_end,
630 							   page_size_mask,
631 							   prot, init);
632 				continue;
633 			}
634 			/*
635 			 * If we are ok with PG_LEVEL_1G mapping, then we will
636 			 * use the existing mapping.
637 			 *
638 			 * Otherwise, we will split the gbpage mapping but use
639 			 * the same existing protection  bits except for large
640 			 * page, so that we don't violate Intel's TLB
641 			 * Application note (317080) which says, while changing
642 			 * the page sizes, new and old translations should
643 			 * not differ with respect to page frame and
644 			 * attributes.
645 			 */
646 			if (page_size_mask & (1 << PG_LEVEL_1G)) {
647 				if (!after_bootmem)
648 					pages++;
649 				paddr_last = paddr_next;
650 				continue;
651 			}
652 			prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
653 		}
654 
655 		if (page_size_mask & (1<<PG_LEVEL_1G)) {
656 			pages++;
657 			spin_lock(&init_mm.page_table_lock);
658 			set_pud_init(pud,
659 				     pfn_pud(paddr >> PAGE_SHIFT, prot_sethuge(prot)),
660 				     init);
661 			spin_unlock(&init_mm.page_table_lock);
662 			paddr_last = paddr_next;
663 			continue;
664 		}
665 
666 		pmd = alloc_low_page();
667 		paddr_last = phys_pmd_init(pmd, paddr, paddr_end,
668 					   page_size_mask, prot, init);
669 
670 		spin_lock(&init_mm.page_table_lock);
671 		pud_populate_init(&init_mm, pud, pmd, init);
672 		spin_unlock(&init_mm.page_table_lock);
673 	}
674 
675 	update_page_count(PG_LEVEL_1G, pages);
676 
677 	return paddr_last;
678 }
679 
680 static unsigned long __meminit
681 phys_p4d_init(p4d_t *p4d_page, unsigned long paddr, unsigned long paddr_end,
682 	      unsigned long page_size_mask, pgprot_t prot, bool init)
683 {
684 	unsigned long vaddr, vaddr_end, vaddr_next, paddr_next, paddr_last;
685 
686 	paddr_last = paddr_end;
687 	vaddr = (unsigned long)__va(paddr);
688 	vaddr_end = (unsigned long)__va(paddr_end);
689 
690 	if (!pgtable_l5_enabled())
691 		return phys_pud_init((pud_t *) p4d_page, paddr, paddr_end,
692 				     page_size_mask, prot, init);
693 
694 	for (; vaddr < vaddr_end; vaddr = vaddr_next) {
695 		p4d_t *p4d = p4d_page + p4d_index(vaddr);
696 		pud_t *pud;
697 
698 		vaddr_next = (vaddr & P4D_MASK) + P4D_SIZE;
699 		paddr = __pa(vaddr);
700 
701 		if (paddr >= paddr_end) {
702 			paddr_next = __pa(vaddr_next);
703 			if (!after_bootmem &&
704 			    !e820__mapped_any(paddr & P4D_MASK, paddr_next,
705 					     E820_TYPE_RAM) &&
706 			    !e820__mapped_any(paddr & P4D_MASK, paddr_next,
707 					     E820_TYPE_RESERVED_KERN) &&
708 			    !e820__mapped_any(paddr & P4D_MASK, paddr_next,
709 					     E820_TYPE_ACPI))
710 				set_p4d_init(p4d, __p4d(0), init);
711 			continue;
712 		}
713 
714 		if (!p4d_none(*p4d)) {
715 			pud = pud_offset(p4d, 0);
716 			paddr_last = phys_pud_init(pud, paddr, __pa(vaddr_end),
717 					page_size_mask, prot, init);
718 			continue;
719 		}
720 
721 		pud = alloc_low_page();
722 		paddr_last = phys_pud_init(pud, paddr, __pa(vaddr_end),
723 					   page_size_mask, prot, init);
724 
725 		spin_lock(&init_mm.page_table_lock);
726 		p4d_populate_init(&init_mm, p4d, pud, init);
727 		spin_unlock(&init_mm.page_table_lock);
728 	}
729 
730 	return paddr_last;
731 }
732 
733 static unsigned long __meminit
734 __kernel_physical_mapping_init(unsigned long paddr_start,
735 			       unsigned long paddr_end,
736 			       unsigned long page_size_mask,
737 			       pgprot_t prot, bool init)
738 {
739 	bool pgd_changed = false;
740 	unsigned long vaddr, vaddr_start, vaddr_end, vaddr_next, paddr_last;
741 
742 	paddr_last = paddr_end;
743 	vaddr = (unsigned long)__va(paddr_start);
744 	vaddr_end = (unsigned long)__va(paddr_end);
745 	vaddr_start = vaddr;
746 
747 	for (; vaddr < vaddr_end; vaddr = vaddr_next) {
748 		pgd_t *pgd = pgd_offset_k(vaddr);
749 		p4d_t *p4d;
750 
751 		vaddr_next = (vaddr & PGDIR_MASK) + PGDIR_SIZE;
752 
753 		if (pgd_val(*pgd)) {
754 			p4d = (p4d_t *)pgd_page_vaddr(*pgd);
755 			paddr_last = phys_p4d_init(p4d, __pa(vaddr),
756 						   __pa(vaddr_end),
757 						   page_size_mask,
758 						   prot, init);
759 			continue;
760 		}
761 
762 		p4d = alloc_low_page();
763 		paddr_last = phys_p4d_init(p4d, __pa(vaddr), __pa(vaddr_end),
764 					   page_size_mask, prot, init);
765 
766 		spin_lock(&init_mm.page_table_lock);
767 		if (pgtable_l5_enabled())
768 			pgd_populate_init(&init_mm, pgd, p4d, init);
769 		else
770 			p4d_populate_init(&init_mm, p4d_offset(pgd, vaddr),
771 					  (pud_t *) p4d, init);
772 
773 		spin_unlock(&init_mm.page_table_lock);
774 		pgd_changed = true;
775 	}
776 
777 	if (pgd_changed)
778 		sync_global_pgds(vaddr_start, vaddr_end - 1);
779 
780 	return paddr_last;
781 }
782 
783 
784 /*
785  * Create page table mapping for the physical memory for specific physical
786  * addresses. Note that it can only be used to populate non-present entries.
787  * The virtual and physical addresses have to be aligned on PMD level
788  * down. It returns the last physical address mapped.
789  */
790 unsigned long __meminit
791 kernel_physical_mapping_init(unsigned long paddr_start,
792 			     unsigned long paddr_end,
793 			     unsigned long page_size_mask, pgprot_t prot)
794 {
795 	return __kernel_physical_mapping_init(paddr_start, paddr_end,
796 					      page_size_mask, prot, true);
797 }
798 
799 /*
800  * This function is similar to kernel_physical_mapping_init() above with the
801  * exception that it uses set_{pud,pmd}() instead of the set_{pud,pte}_safe()
802  * when updating the mapping. The caller is responsible to flush the TLBs after
803  * the function returns.
804  */
805 unsigned long __meminit
806 kernel_physical_mapping_change(unsigned long paddr_start,
807 			       unsigned long paddr_end,
808 			       unsigned long page_size_mask)
809 {
810 	return __kernel_physical_mapping_init(paddr_start, paddr_end,
811 					      page_size_mask, PAGE_KERNEL,
812 					      false);
813 }
814 
815 #ifndef CONFIG_NUMA
816 void __init initmem_init(void)
817 {
818 	memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
819 }
820 #endif
821 
822 void __init paging_init(void)
823 {
824 	sparse_init();
825 
826 	/*
827 	 * clear the default setting with node 0
828 	 * note: don't use nodes_clear here, that is really clearing when
829 	 *	 numa support is not compiled in, and later node_set_state
830 	 *	 will not set it back.
831 	 */
832 	node_clear_state(0, N_MEMORY);
833 	node_clear_state(0, N_NORMAL_MEMORY);
834 
835 	zone_sizes_init();
836 }
837 
838 #ifdef CONFIG_SPARSEMEM_VMEMMAP
839 #define PAGE_UNUSED 0xFD
840 
841 /*
842  * The unused vmemmap range, which was not yet memset(PAGE_UNUSED), ranges
843  * from unused_pmd_start to next PMD_SIZE boundary.
844  */
845 static unsigned long unused_pmd_start __meminitdata;
846 
847 static void __meminit vmemmap_flush_unused_pmd(void)
848 {
849 	if (!unused_pmd_start)
850 		return;
851 	/*
852 	 * Clears (unused_pmd_start, PMD_END]
853 	 */
854 	memset((void *)unused_pmd_start, PAGE_UNUSED,
855 	       ALIGN(unused_pmd_start, PMD_SIZE) - unused_pmd_start);
856 	unused_pmd_start = 0;
857 }
858 
859 #ifdef CONFIG_MEMORY_HOTPLUG
860 /* Returns true if the PMD is completely unused and thus it can be freed */
861 static bool __meminit vmemmap_pmd_is_unused(unsigned long addr, unsigned long end)
862 {
863 	unsigned long start = ALIGN_DOWN(addr, PMD_SIZE);
864 
865 	/*
866 	 * Flush the unused range cache to ensure that memchr_inv() will work
867 	 * for the whole range.
868 	 */
869 	vmemmap_flush_unused_pmd();
870 	memset((void *)addr, PAGE_UNUSED, end - addr);
871 
872 	return !memchr_inv((void *)start, PAGE_UNUSED, PMD_SIZE);
873 }
874 #endif
875 
876 static void __meminit __vmemmap_use_sub_pmd(unsigned long start)
877 {
878 	/*
879 	 * As we expect to add in the same granularity as we remove, it's
880 	 * sufficient to mark only some piece used to block the memmap page from
881 	 * getting removed when removing some other adjacent memmap (just in
882 	 * case the first memmap never gets initialized e.g., because the memory
883 	 * block never gets onlined).
884 	 */
885 	memset((void *)start, 0, sizeof(struct page));
886 }
887 
888 static void __meminit vmemmap_use_sub_pmd(unsigned long start, unsigned long end)
889 {
890 	/*
891 	 * We only optimize if the new used range directly follows the
892 	 * previously unused range (esp., when populating consecutive sections).
893 	 */
894 	if (unused_pmd_start == start) {
895 		if (likely(IS_ALIGNED(end, PMD_SIZE)))
896 			unused_pmd_start = 0;
897 		else
898 			unused_pmd_start = end;
899 		return;
900 	}
901 
902 	/*
903 	 * If the range does not contiguously follows previous one, make sure
904 	 * to mark the unused range of the previous one so it can be removed.
905 	 */
906 	vmemmap_flush_unused_pmd();
907 	__vmemmap_use_sub_pmd(start);
908 }
909 
910 
911 static void __meminit vmemmap_use_new_sub_pmd(unsigned long start, unsigned long end)
912 {
913 	const unsigned long page = ALIGN_DOWN(start, PMD_SIZE);
914 
915 	vmemmap_flush_unused_pmd();
916 
917 	/*
918 	 * Could be our memmap page is filled with PAGE_UNUSED already from a
919 	 * previous remove. Make sure to reset it.
920 	 */
921 	__vmemmap_use_sub_pmd(start);
922 
923 	/*
924 	 * Mark with PAGE_UNUSED the unused parts of the new memmap range
925 	 */
926 	if (!IS_ALIGNED(start, PMD_SIZE))
927 		memset((void *)page, PAGE_UNUSED, start - page);
928 
929 	/*
930 	 * We want to avoid memset(PAGE_UNUSED) when populating the vmemmap of
931 	 * consecutive sections. Remember for the last added PMD where the
932 	 * unused range begins.
933 	 */
934 	if (!IS_ALIGNED(end, PMD_SIZE))
935 		unused_pmd_start = end;
936 }
937 #endif
938 
939 /*
940  * Memory hotplug specific functions
941  */
942 #ifdef CONFIG_MEMORY_HOTPLUG
943 /*
944  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
945  * updating.
946  */
947 static void update_end_of_memory_vars(u64 start, u64 size)
948 {
949 	unsigned long end_pfn = PFN_UP(start + size);
950 
951 	if (end_pfn > max_pfn) {
952 		max_pfn = end_pfn;
953 		max_low_pfn = end_pfn;
954 		high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
955 	}
956 }
957 
958 int add_pages(int nid, unsigned long start_pfn, unsigned long nr_pages,
959 	      struct mhp_params *params)
960 {
961 	unsigned long end = ((start_pfn + nr_pages) << PAGE_SHIFT) - 1;
962 	int ret;
963 
964 	if (WARN_ON_ONCE(end > PHYSMEM_END))
965 		return -ERANGE;
966 
967 	ret = __add_pages(nid, start_pfn, nr_pages, params);
968 	WARN_ON_ONCE(ret);
969 
970 	/* update max_pfn, max_low_pfn and high_memory */
971 	update_end_of_memory_vars(start_pfn << PAGE_SHIFT,
972 				  nr_pages << PAGE_SHIFT);
973 
974 	return ret;
975 }
976 
977 int arch_add_memory(int nid, u64 start, u64 size,
978 		    struct mhp_params *params)
979 {
980 	unsigned long start_pfn = start >> PAGE_SHIFT;
981 	unsigned long nr_pages = size >> PAGE_SHIFT;
982 
983 	init_memory_mapping(start, start + size, params->pgprot);
984 
985 	return add_pages(nid, start_pfn, nr_pages, params);
986 }
987 
988 static void __meminit free_pagetable(struct page *page, int order)
989 {
990 	unsigned long magic;
991 	unsigned int nr_pages = 1 << order;
992 
993 	/* bootmem page has reserved flag */
994 	if (PageReserved(page)) {
995 		magic = page->index;
996 		if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
997 			while (nr_pages--)
998 				put_page_bootmem(page++);
999 		} else
1000 			while (nr_pages--)
1001 				free_reserved_page(page++);
1002 	} else
1003 		free_pages((unsigned long)page_address(page), order);
1004 }
1005 
1006 static void __meminit free_hugepage_table(struct page *page,
1007 		struct vmem_altmap *altmap)
1008 {
1009 	if (altmap)
1010 		vmem_altmap_free(altmap, PMD_SIZE / PAGE_SIZE);
1011 	else
1012 		free_pagetable(page, get_order(PMD_SIZE));
1013 }
1014 
1015 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
1016 {
1017 	pte_t *pte;
1018 	int i;
1019 
1020 	for (i = 0; i < PTRS_PER_PTE; i++) {
1021 		pte = pte_start + i;
1022 		if (!pte_none(*pte))
1023 			return;
1024 	}
1025 
1026 	/* free a pte table */
1027 	free_pagetable(pmd_page(*pmd), 0);
1028 	spin_lock(&init_mm.page_table_lock);
1029 	pmd_clear(pmd);
1030 	spin_unlock(&init_mm.page_table_lock);
1031 }
1032 
1033 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
1034 {
1035 	pmd_t *pmd;
1036 	int i;
1037 
1038 	for (i = 0; i < PTRS_PER_PMD; i++) {
1039 		pmd = pmd_start + i;
1040 		if (!pmd_none(*pmd))
1041 			return;
1042 	}
1043 
1044 	/* free a pmd table */
1045 	free_pagetable(pud_page(*pud), 0);
1046 	spin_lock(&init_mm.page_table_lock);
1047 	pud_clear(pud);
1048 	spin_unlock(&init_mm.page_table_lock);
1049 }
1050 
1051 static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
1052 {
1053 	pud_t *pud;
1054 	int i;
1055 
1056 	for (i = 0; i < PTRS_PER_PUD; i++) {
1057 		pud = pud_start + i;
1058 		if (!pud_none(*pud))
1059 			return;
1060 	}
1061 
1062 	/* free a pud table */
1063 	free_pagetable(p4d_page(*p4d), 0);
1064 	spin_lock(&init_mm.page_table_lock);
1065 	p4d_clear(p4d);
1066 	spin_unlock(&init_mm.page_table_lock);
1067 }
1068 
1069 static void __meminit
1070 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
1071 		 bool direct)
1072 {
1073 	unsigned long next, pages = 0;
1074 	pte_t *pte;
1075 	phys_addr_t phys_addr;
1076 
1077 	pte = pte_start + pte_index(addr);
1078 	for (; addr < end; addr = next, pte++) {
1079 		next = (addr + PAGE_SIZE) & PAGE_MASK;
1080 		if (next > end)
1081 			next = end;
1082 
1083 		if (!pte_present(*pte))
1084 			continue;
1085 
1086 		/*
1087 		 * We mapped [0,1G) memory as identity mapping when
1088 		 * initializing, in arch/x86/kernel/head_64.S. These
1089 		 * pagetables cannot be removed.
1090 		 */
1091 		phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
1092 		if (phys_addr < (phys_addr_t)0x40000000)
1093 			return;
1094 
1095 		if (!direct)
1096 			free_pagetable(pte_page(*pte), 0);
1097 
1098 		spin_lock(&init_mm.page_table_lock);
1099 		pte_clear(&init_mm, addr, pte);
1100 		spin_unlock(&init_mm.page_table_lock);
1101 
1102 		/* For non-direct mapping, pages means nothing. */
1103 		pages++;
1104 	}
1105 
1106 	/* Call free_pte_table() in remove_pmd_table(). */
1107 	flush_tlb_all();
1108 	if (direct)
1109 		update_page_count(PG_LEVEL_4K, -pages);
1110 }
1111 
1112 static void __meminit
1113 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
1114 		 bool direct, struct vmem_altmap *altmap)
1115 {
1116 	unsigned long next, pages = 0;
1117 	pte_t *pte_base;
1118 	pmd_t *pmd;
1119 
1120 	pmd = pmd_start + pmd_index(addr);
1121 	for (; addr < end; addr = next, pmd++) {
1122 		next = pmd_addr_end(addr, end);
1123 
1124 		if (!pmd_present(*pmd))
1125 			continue;
1126 
1127 		if (pmd_leaf(*pmd)) {
1128 			if (IS_ALIGNED(addr, PMD_SIZE) &&
1129 			    IS_ALIGNED(next, PMD_SIZE)) {
1130 				if (!direct)
1131 					free_hugepage_table(pmd_page(*pmd),
1132 							    altmap);
1133 
1134 				spin_lock(&init_mm.page_table_lock);
1135 				pmd_clear(pmd);
1136 				spin_unlock(&init_mm.page_table_lock);
1137 				pages++;
1138 			}
1139 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1140 			else if (vmemmap_pmd_is_unused(addr, next)) {
1141 					free_hugepage_table(pmd_page(*pmd),
1142 							    altmap);
1143 					spin_lock(&init_mm.page_table_lock);
1144 					pmd_clear(pmd);
1145 					spin_unlock(&init_mm.page_table_lock);
1146 			}
1147 #endif
1148 			continue;
1149 		}
1150 
1151 		pte_base = (pte_t *)pmd_page_vaddr(*pmd);
1152 		remove_pte_table(pte_base, addr, next, direct);
1153 		free_pte_table(pte_base, pmd);
1154 	}
1155 
1156 	/* Call free_pmd_table() in remove_pud_table(). */
1157 	if (direct)
1158 		update_page_count(PG_LEVEL_2M, -pages);
1159 }
1160 
1161 static void __meminit
1162 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
1163 		 struct vmem_altmap *altmap, bool direct)
1164 {
1165 	unsigned long next, pages = 0;
1166 	pmd_t *pmd_base;
1167 	pud_t *pud;
1168 
1169 	pud = pud_start + pud_index(addr);
1170 	for (; addr < end; addr = next, pud++) {
1171 		next = pud_addr_end(addr, end);
1172 
1173 		if (!pud_present(*pud))
1174 			continue;
1175 
1176 		if (pud_leaf(*pud) &&
1177 		    IS_ALIGNED(addr, PUD_SIZE) &&
1178 		    IS_ALIGNED(next, PUD_SIZE)) {
1179 			spin_lock(&init_mm.page_table_lock);
1180 			pud_clear(pud);
1181 			spin_unlock(&init_mm.page_table_lock);
1182 			pages++;
1183 			continue;
1184 		}
1185 
1186 		pmd_base = pmd_offset(pud, 0);
1187 		remove_pmd_table(pmd_base, addr, next, direct, altmap);
1188 		free_pmd_table(pmd_base, pud);
1189 	}
1190 
1191 	if (direct)
1192 		update_page_count(PG_LEVEL_1G, -pages);
1193 }
1194 
1195 static void __meminit
1196 remove_p4d_table(p4d_t *p4d_start, unsigned long addr, unsigned long end,
1197 		 struct vmem_altmap *altmap, bool direct)
1198 {
1199 	unsigned long next, pages = 0;
1200 	pud_t *pud_base;
1201 	p4d_t *p4d;
1202 
1203 	p4d = p4d_start + p4d_index(addr);
1204 	for (; addr < end; addr = next, p4d++) {
1205 		next = p4d_addr_end(addr, end);
1206 
1207 		if (!p4d_present(*p4d))
1208 			continue;
1209 
1210 		BUILD_BUG_ON(p4d_leaf(*p4d));
1211 
1212 		pud_base = pud_offset(p4d, 0);
1213 		remove_pud_table(pud_base, addr, next, altmap, direct);
1214 		/*
1215 		 * For 4-level page tables we do not want to free PUDs, but in the
1216 		 * 5-level case we should free them. This code will have to change
1217 		 * to adapt for boot-time switching between 4 and 5 level page tables.
1218 		 */
1219 		if (pgtable_l5_enabled())
1220 			free_pud_table(pud_base, p4d);
1221 	}
1222 
1223 	if (direct)
1224 		update_page_count(PG_LEVEL_512G, -pages);
1225 }
1226 
1227 /* start and end are both virtual address. */
1228 static void __meminit
1229 remove_pagetable(unsigned long start, unsigned long end, bool direct,
1230 		struct vmem_altmap *altmap)
1231 {
1232 	unsigned long next;
1233 	unsigned long addr;
1234 	pgd_t *pgd;
1235 	p4d_t *p4d;
1236 
1237 	for (addr = start; addr < end; addr = next) {
1238 		next = pgd_addr_end(addr, end);
1239 
1240 		pgd = pgd_offset_k(addr);
1241 		if (!pgd_present(*pgd))
1242 			continue;
1243 
1244 		p4d = p4d_offset(pgd, 0);
1245 		remove_p4d_table(p4d, addr, next, altmap, direct);
1246 	}
1247 
1248 	flush_tlb_all();
1249 }
1250 
1251 void __ref vmemmap_free(unsigned long start, unsigned long end,
1252 		struct vmem_altmap *altmap)
1253 {
1254 	VM_BUG_ON(!PAGE_ALIGNED(start));
1255 	VM_BUG_ON(!PAGE_ALIGNED(end));
1256 
1257 	remove_pagetable(start, end, false, altmap);
1258 }
1259 
1260 static void __meminit
1261 kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1262 {
1263 	start = (unsigned long)__va(start);
1264 	end = (unsigned long)__va(end);
1265 
1266 	remove_pagetable(start, end, true, NULL);
1267 }
1268 
1269 void __ref arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap)
1270 {
1271 	unsigned long start_pfn = start >> PAGE_SHIFT;
1272 	unsigned long nr_pages = size >> PAGE_SHIFT;
1273 
1274 	__remove_pages(start_pfn, nr_pages, altmap);
1275 	kernel_physical_mapping_remove(start, start + size);
1276 }
1277 #endif /* CONFIG_MEMORY_HOTPLUG */
1278 
1279 static struct kcore_list kcore_vsyscall;
1280 
1281 static void __init register_page_bootmem_info(void)
1282 {
1283 #if defined(CONFIG_NUMA) || defined(CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP)
1284 	int i;
1285 
1286 	for_each_online_node(i)
1287 		register_page_bootmem_info_node(NODE_DATA(i));
1288 #endif
1289 }
1290 
1291 /*
1292  * Pre-allocates page-table pages for the vmalloc area in the kernel page-table.
1293  * Only the level which needs to be synchronized between all page-tables is
1294  * allocated because the synchronization can be expensive.
1295  */
1296 static void __init preallocate_vmalloc_pages(void)
1297 {
1298 	unsigned long addr;
1299 	const char *lvl;
1300 
1301 	for (addr = VMALLOC_START; addr <= VMEMORY_END; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
1302 		pgd_t *pgd = pgd_offset_k(addr);
1303 		p4d_t *p4d;
1304 		pud_t *pud;
1305 
1306 		lvl = "p4d";
1307 		p4d = p4d_alloc(&init_mm, pgd, addr);
1308 		if (!p4d)
1309 			goto failed;
1310 
1311 		if (pgtable_l5_enabled())
1312 			continue;
1313 
1314 		/*
1315 		 * The goal here is to allocate all possibly required
1316 		 * hardware page tables pointed to by the top hardware
1317 		 * level.
1318 		 *
1319 		 * On 4-level systems, the P4D layer is folded away and
1320 		 * the above code does no preallocation.  Below, go down
1321 		 * to the pud _software_ level to ensure the second
1322 		 * hardware level is allocated on 4-level systems too.
1323 		 */
1324 		lvl = "pud";
1325 		pud = pud_alloc(&init_mm, p4d, addr);
1326 		if (!pud)
1327 			goto failed;
1328 	}
1329 
1330 	return;
1331 
1332 failed:
1333 
1334 	/*
1335 	 * The pages have to be there now or they will be missing in
1336 	 * process page-tables later.
1337 	 */
1338 	panic("Failed to pre-allocate %s pages for vmalloc area\n", lvl);
1339 }
1340 
1341 void __init mem_init(void)
1342 {
1343 	pci_iommu_alloc();
1344 
1345 	/* clear_bss() already clear the empty_zero_page */
1346 
1347 	/* this will put all memory onto the freelists */
1348 	memblock_free_all();
1349 	after_bootmem = 1;
1350 	x86_init.hyper.init_after_bootmem();
1351 
1352 	/*
1353 	 * Must be done after boot memory is put on freelist, because here we
1354 	 * might set fields in deferred struct pages that have not yet been
1355 	 * initialized, and memblock_free_all() initializes all the reserved
1356 	 * deferred pages for us.
1357 	 */
1358 	register_page_bootmem_info();
1359 
1360 	/* Register memory areas for /proc/kcore */
1361 	if (get_gate_vma(&init_mm))
1362 		kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR, PAGE_SIZE, KCORE_USER);
1363 
1364 	preallocate_vmalloc_pages();
1365 }
1366 
1367 int kernel_set_to_readonly;
1368 
1369 void mark_rodata_ro(void)
1370 {
1371 	unsigned long start = PFN_ALIGN(_text);
1372 	unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1373 	unsigned long end = (unsigned long)__end_rodata_hpage_align;
1374 	unsigned long text_end = PFN_ALIGN(_etext);
1375 	unsigned long rodata_end = PFN_ALIGN(__end_rodata);
1376 	unsigned long all_end;
1377 
1378 	printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1379 	       (end - start) >> 10);
1380 	set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1381 
1382 	kernel_set_to_readonly = 1;
1383 
1384 	/*
1385 	 * The rodata/data/bss/brk section (but not the kernel text!)
1386 	 * should also be not-executable.
1387 	 *
1388 	 * We align all_end to PMD_SIZE because the existing mapping
1389 	 * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1390 	 * split the PMD and the reminder between _brk_end and the end
1391 	 * of the PMD will remain mapped executable.
1392 	 *
1393 	 * Any PMD which was setup after the one which covers _brk_end
1394 	 * has been zapped already via cleanup_highmem().
1395 	 */
1396 	all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1397 	set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
1398 
1399 	set_ftrace_ops_ro();
1400 
1401 #ifdef CONFIG_CPA_DEBUG
1402 	printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1403 	set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1404 
1405 	printk(KERN_INFO "Testing CPA: again\n");
1406 	set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1407 #endif
1408 
1409 	free_kernel_image_pages("unused kernel image (text/rodata gap)",
1410 				(void *)text_end, (void *)rodata_start);
1411 	free_kernel_image_pages("unused kernel image (rodata/data gap)",
1412 				(void *)rodata_end, (void *)_sdata);
1413 }
1414 
1415 /*
1416  * Block size is the minimum amount of memory which can be hotplugged or
1417  * hotremoved. It must be power of two and must be equal or larger than
1418  * MIN_MEMORY_BLOCK_SIZE.
1419  */
1420 #define MAX_BLOCK_SIZE (2UL << 30)
1421 
1422 /* Amount of ram needed to start using large blocks */
1423 #define MEM_SIZE_FOR_LARGE_BLOCK (64UL << 30)
1424 
1425 /* Adjustable memory block size */
1426 static unsigned long set_memory_block_size;
1427 int __init set_memory_block_size_order(unsigned int order)
1428 {
1429 	unsigned long size = 1UL << order;
1430 
1431 	if (size > MEM_SIZE_FOR_LARGE_BLOCK || size < MIN_MEMORY_BLOCK_SIZE)
1432 		return -EINVAL;
1433 
1434 	set_memory_block_size = size;
1435 	return 0;
1436 }
1437 
1438 static unsigned long probe_memory_block_size(void)
1439 {
1440 	unsigned long boot_mem_end = max_pfn << PAGE_SHIFT;
1441 	unsigned long bz;
1442 
1443 	/* If memory block size has been set, then use it */
1444 	bz = set_memory_block_size;
1445 	if (bz)
1446 		goto done;
1447 
1448 	/* Use regular block if RAM is smaller than MEM_SIZE_FOR_LARGE_BLOCK */
1449 	if (boot_mem_end < MEM_SIZE_FOR_LARGE_BLOCK) {
1450 		bz = MIN_MEMORY_BLOCK_SIZE;
1451 		goto done;
1452 	}
1453 
1454 	/*
1455 	 * Use max block size to minimize overhead on bare metal, where
1456 	 * alignment for memory hotplug isn't a concern.
1457 	 */
1458 	if (!boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
1459 		bz = MAX_BLOCK_SIZE;
1460 		goto done;
1461 	}
1462 
1463 	/* Find the largest allowed block size that aligns to memory end */
1464 	for (bz = MAX_BLOCK_SIZE; bz > MIN_MEMORY_BLOCK_SIZE; bz >>= 1) {
1465 		if (IS_ALIGNED(boot_mem_end, bz))
1466 			break;
1467 	}
1468 done:
1469 	pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20);
1470 
1471 	return bz;
1472 }
1473 
1474 static unsigned long memory_block_size_probed;
1475 unsigned long memory_block_size_bytes(void)
1476 {
1477 	if (!memory_block_size_probed)
1478 		memory_block_size_probed = probe_memory_block_size();
1479 
1480 	return memory_block_size_probed;
1481 }
1482 
1483 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1484 /*
1485  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1486  */
1487 static long __meminitdata addr_start, addr_end;
1488 static void __meminitdata *p_start, *p_end;
1489 static int __meminitdata node_start;
1490 
1491 void __meminit vmemmap_set_pmd(pmd_t *pmd, void *p, int node,
1492 			       unsigned long addr, unsigned long next)
1493 {
1494 	pte_t entry;
1495 
1496 	entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1497 			PAGE_KERNEL_LARGE);
1498 	set_pmd(pmd, __pmd(pte_val(entry)));
1499 
1500 	/* check to see if we have contiguous blocks */
1501 	if (p_end != p || node_start != node) {
1502 		if (p_start)
1503 			pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1504 				addr_start, addr_end-1, p_start, p_end-1, node_start);
1505 		addr_start = addr;
1506 		node_start = node;
1507 		p_start = p;
1508 	}
1509 
1510 	addr_end = addr + PMD_SIZE;
1511 	p_end = p + PMD_SIZE;
1512 
1513 	if (!IS_ALIGNED(addr, PMD_SIZE) ||
1514 		!IS_ALIGNED(next, PMD_SIZE))
1515 		vmemmap_use_new_sub_pmd(addr, next);
1516 }
1517 
1518 int __meminit vmemmap_check_pmd(pmd_t *pmd, int node,
1519 				unsigned long addr, unsigned long next)
1520 {
1521 	int large = pmd_leaf(*pmd);
1522 
1523 	if (pmd_leaf(*pmd)) {
1524 		vmemmap_verify((pte_t *)pmd, node, addr, next);
1525 		vmemmap_use_sub_pmd(addr, next);
1526 	}
1527 
1528 	return large;
1529 }
1530 
1531 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1532 		struct vmem_altmap *altmap)
1533 {
1534 	int err;
1535 
1536 	VM_BUG_ON(!PAGE_ALIGNED(start));
1537 	VM_BUG_ON(!PAGE_ALIGNED(end));
1538 
1539 	if (end - start < PAGES_PER_SECTION * sizeof(struct page))
1540 		err = vmemmap_populate_basepages(start, end, node, NULL);
1541 	else if (boot_cpu_has(X86_FEATURE_PSE))
1542 		err = vmemmap_populate_hugepages(start, end, node, altmap);
1543 	else if (altmap) {
1544 		pr_err_once("%s: no cpu support for altmap allocations\n",
1545 				__func__);
1546 		err = -ENOMEM;
1547 	} else
1548 		err = vmemmap_populate_basepages(start, end, node, NULL);
1549 	if (!err)
1550 		sync_global_pgds(start, end - 1);
1551 	return err;
1552 }
1553 
1554 #ifdef CONFIG_HAVE_BOOTMEM_INFO_NODE
1555 void register_page_bootmem_memmap(unsigned long section_nr,
1556 				  struct page *start_page, unsigned long nr_pages)
1557 {
1558 	unsigned long addr = (unsigned long)start_page;
1559 	unsigned long end = (unsigned long)(start_page + nr_pages);
1560 	unsigned long next;
1561 	pgd_t *pgd;
1562 	p4d_t *p4d;
1563 	pud_t *pud;
1564 	pmd_t *pmd;
1565 	unsigned int nr_pmd_pages;
1566 	struct page *page;
1567 
1568 	for (; addr < end; addr = next) {
1569 		pte_t *pte = NULL;
1570 
1571 		pgd = pgd_offset_k(addr);
1572 		if (pgd_none(*pgd)) {
1573 			next = (addr + PAGE_SIZE) & PAGE_MASK;
1574 			continue;
1575 		}
1576 		get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1577 
1578 		p4d = p4d_offset(pgd, addr);
1579 		if (p4d_none(*p4d)) {
1580 			next = (addr + PAGE_SIZE) & PAGE_MASK;
1581 			continue;
1582 		}
1583 		get_page_bootmem(section_nr, p4d_page(*p4d), MIX_SECTION_INFO);
1584 
1585 		pud = pud_offset(p4d, addr);
1586 		if (pud_none(*pud)) {
1587 			next = (addr + PAGE_SIZE) & PAGE_MASK;
1588 			continue;
1589 		}
1590 		get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1591 
1592 		if (!boot_cpu_has(X86_FEATURE_PSE)) {
1593 			next = (addr + PAGE_SIZE) & PAGE_MASK;
1594 			pmd = pmd_offset(pud, addr);
1595 			if (pmd_none(*pmd))
1596 				continue;
1597 			get_page_bootmem(section_nr, pmd_page(*pmd),
1598 					 MIX_SECTION_INFO);
1599 
1600 			pte = pte_offset_kernel(pmd, addr);
1601 			if (pte_none(*pte))
1602 				continue;
1603 			get_page_bootmem(section_nr, pte_page(*pte),
1604 					 SECTION_INFO);
1605 		} else {
1606 			next = pmd_addr_end(addr, end);
1607 
1608 			pmd = pmd_offset(pud, addr);
1609 			if (pmd_none(*pmd))
1610 				continue;
1611 
1612 			nr_pmd_pages = 1 << get_order(PMD_SIZE);
1613 			page = pmd_page(*pmd);
1614 			while (nr_pmd_pages--)
1615 				get_page_bootmem(section_nr, page++,
1616 						 SECTION_INFO);
1617 		}
1618 	}
1619 }
1620 #endif
1621 
1622 void __meminit vmemmap_populate_print_last(void)
1623 {
1624 	if (p_start) {
1625 		pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1626 			addr_start, addr_end-1, p_start, p_end-1, node_start);
1627 		p_start = NULL;
1628 		p_end = NULL;
1629 		node_start = 0;
1630 	}
1631 }
1632 #endif
1633