xref: /openbmc/linux/arch/x86/mm/init_64.c (revision 97a532c3)
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 				set_pte_init(pte, __pte(0), init);
474 			continue;
475 		}
476 
477 		/*
478 		 * We will re-use the existing mapping.
479 		 * Xen for example has some special requirements, like mapping
480 		 * pagetable pages as RO. So assume someone who pre-setup
481 		 * these mappings are more intelligent.
482 		 */
483 		if (!pte_none(*pte)) {
484 			if (!after_bootmem)
485 				pages++;
486 			continue;
487 		}
488 
489 		if (0)
490 			pr_info("   pte=%p addr=%lx pte=%016lx\n", pte, paddr,
491 				pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL).pte);
492 		pages++;
493 		set_pte_init(pte, pfn_pte(paddr >> PAGE_SHIFT, prot), init);
494 		paddr_last = (paddr & PAGE_MASK) + PAGE_SIZE;
495 	}
496 
497 	update_page_count(PG_LEVEL_4K, pages);
498 
499 	return paddr_last;
500 }
501 
502 /*
503  * Create PMD level page table mapping for physical addresses. The virtual
504  * and physical address have to be aligned at this level.
505  * It returns the last physical address mapped.
506  */
507 static unsigned long __meminit
508 phys_pmd_init(pmd_t *pmd_page, unsigned long paddr, unsigned long paddr_end,
509 	      unsigned long page_size_mask, pgprot_t prot, bool init)
510 {
511 	unsigned long pages = 0, paddr_next;
512 	unsigned long paddr_last = paddr_end;
513 
514 	int i = pmd_index(paddr);
515 
516 	for (; i < PTRS_PER_PMD; i++, paddr = paddr_next) {
517 		pmd_t *pmd = pmd_page + pmd_index(paddr);
518 		pte_t *pte;
519 		pgprot_t new_prot = prot;
520 
521 		paddr_next = (paddr & PMD_MASK) + PMD_SIZE;
522 		if (paddr >= paddr_end) {
523 			if (!after_bootmem &&
524 			    !e820__mapped_any(paddr & PMD_MASK, paddr_next,
525 					     E820_TYPE_RAM) &&
526 			    !e820__mapped_any(paddr & PMD_MASK, paddr_next,
527 					     E820_TYPE_RESERVED_KERN))
528 				set_pmd_init(pmd, __pmd(0), init);
529 			continue;
530 		}
531 
532 		if (!pmd_none(*pmd)) {
533 			if (!pmd_large(*pmd)) {
534 				spin_lock(&init_mm.page_table_lock);
535 				pte = (pte_t *)pmd_page_vaddr(*pmd);
536 				paddr_last = phys_pte_init(pte, paddr,
537 							   paddr_end, prot,
538 							   init);
539 				spin_unlock(&init_mm.page_table_lock);
540 				continue;
541 			}
542 			/*
543 			 * If we are ok with PG_LEVEL_2M mapping, then we will
544 			 * use the existing mapping,
545 			 *
546 			 * Otherwise, we will split the large page mapping but
547 			 * use the same existing protection bits except for
548 			 * large page, so that we don't violate Intel's TLB
549 			 * Application note (317080) which says, while changing
550 			 * the page sizes, new and old translations should
551 			 * not differ with respect to page frame and
552 			 * attributes.
553 			 */
554 			if (page_size_mask & (1 << PG_LEVEL_2M)) {
555 				if (!after_bootmem)
556 					pages++;
557 				paddr_last = paddr_next;
558 				continue;
559 			}
560 			new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
561 		}
562 
563 		if (page_size_mask & (1<<PG_LEVEL_2M)) {
564 			pages++;
565 			spin_lock(&init_mm.page_table_lock);
566 			set_pmd_init(pmd,
567 				     pfn_pmd(paddr >> PAGE_SHIFT, prot_sethuge(prot)),
568 				     init);
569 			spin_unlock(&init_mm.page_table_lock);
570 			paddr_last = paddr_next;
571 			continue;
572 		}
573 
574 		pte = alloc_low_page();
575 		paddr_last = phys_pte_init(pte, paddr, paddr_end, new_prot, init);
576 
577 		spin_lock(&init_mm.page_table_lock);
578 		pmd_populate_kernel_init(&init_mm, pmd, pte, init);
579 		spin_unlock(&init_mm.page_table_lock);
580 	}
581 	update_page_count(PG_LEVEL_2M, pages);
582 	return paddr_last;
583 }
584 
585 /*
586  * Create PUD level page table mapping for physical addresses. The virtual
587  * and physical address do not have to be aligned at this level. KASLR can
588  * randomize virtual addresses up to this level.
589  * It returns the last physical address mapped.
590  */
591 static unsigned long __meminit
592 phys_pud_init(pud_t *pud_page, unsigned long paddr, unsigned long paddr_end,
593 	      unsigned long page_size_mask, pgprot_t _prot, bool init)
594 {
595 	unsigned long pages = 0, paddr_next;
596 	unsigned long paddr_last = paddr_end;
597 	unsigned long vaddr = (unsigned long)__va(paddr);
598 	int i = pud_index(vaddr);
599 
600 	for (; i < PTRS_PER_PUD; i++, paddr = paddr_next) {
601 		pud_t *pud;
602 		pmd_t *pmd;
603 		pgprot_t prot = _prot;
604 
605 		vaddr = (unsigned long)__va(paddr);
606 		pud = pud_page + pud_index(vaddr);
607 		paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
608 
609 		if (paddr >= paddr_end) {
610 			if (!after_bootmem &&
611 			    !e820__mapped_any(paddr & PUD_MASK, paddr_next,
612 					     E820_TYPE_RAM) &&
613 			    !e820__mapped_any(paddr & PUD_MASK, paddr_next,
614 					     E820_TYPE_RESERVED_KERN))
615 				set_pud_init(pud, __pud(0), init);
616 			continue;
617 		}
618 
619 		if (!pud_none(*pud)) {
620 			if (!pud_leaf(*pud)) {
621 				pmd = pmd_offset(pud, 0);
622 				paddr_last = phys_pmd_init(pmd, paddr,
623 							   paddr_end,
624 							   page_size_mask,
625 							   prot, init);
626 				continue;
627 			}
628 			/*
629 			 * If we are ok with PG_LEVEL_1G mapping, then we will
630 			 * use the existing mapping.
631 			 *
632 			 * Otherwise, we will split the gbpage mapping but use
633 			 * the same existing protection  bits except for large
634 			 * page, so that we don't violate Intel's TLB
635 			 * Application note (317080) which says, while changing
636 			 * the page sizes, new and old translations should
637 			 * not differ with respect to page frame and
638 			 * attributes.
639 			 */
640 			if (page_size_mask & (1 << PG_LEVEL_1G)) {
641 				if (!after_bootmem)
642 					pages++;
643 				paddr_last = paddr_next;
644 				continue;
645 			}
646 			prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
647 		}
648 
649 		if (page_size_mask & (1<<PG_LEVEL_1G)) {
650 			pages++;
651 			spin_lock(&init_mm.page_table_lock);
652 			set_pud_init(pud,
653 				     pfn_pud(paddr >> PAGE_SHIFT, prot_sethuge(prot)),
654 				     init);
655 			spin_unlock(&init_mm.page_table_lock);
656 			paddr_last = paddr_next;
657 			continue;
658 		}
659 
660 		pmd = alloc_low_page();
661 		paddr_last = phys_pmd_init(pmd, paddr, paddr_end,
662 					   page_size_mask, prot, init);
663 
664 		spin_lock(&init_mm.page_table_lock);
665 		pud_populate_init(&init_mm, pud, pmd, init);
666 		spin_unlock(&init_mm.page_table_lock);
667 	}
668 
669 	update_page_count(PG_LEVEL_1G, pages);
670 
671 	return paddr_last;
672 }
673 
674 static unsigned long __meminit
675 phys_p4d_init(p4d_t *p4d_page, unsigned long paddr, unsigned long paddr_end,
676 	      unsigned long page_size_mask, pgprot_t prot, bool init)
677 {
678 	unsigned long vaddr, vaddr_end, vaddr_next, paddr_next, paddr_last;
679 
680 	paddr_last = paddr_end;
681 	vaddr = (unsigned long)__va(paddr);
682 	vaddr_end = (unsigned long)__va(paddr_end);
683 
684 	if (!pgtable_l5_enabled())
685 		return phys_pud_init((pud_t *) p4d_page, paddr, paddr_end,
686 				     page_size_mask, prot, init);
687 
688 	for (; vaddr < vaddr_end; vaddr = vaddr_next) {
689 		p4d_t *p4d = p4d_page + p4d_index(vaddr);
690 		pud_t *pud;
691 
692 		vaddr_next = (vaddr & P4D_MASK) + P4D_SIZE;
693 		paddr = __pa(vaddr);
694 
695 		if (paddr >= paddr_end) {
696 			paddr_next = __pa(vaddr_next);
697 			if (!after_bootmem &&
698 			    !e820__mapped_any(paddr & P4D_MASK, paddr_next,
699 					     E820_TYPE_RAM) &&
700 			    !e820__mapped_any(paddr & P4D_MASK, paddr_next,
701 					     E820_TYPE_RESERVED_KERN))
702 				set_p4d_init(p4d, __p4d(0), init);
703 			continue;
704 		}
705 
706 		if (!p4d_none(*p4d)) {
707 			pud = pud_offset(p4d, 0);
708 			paddr_last = phys_pud_init(pud, paddr, __pa(vaddr_end),
709 					page_size_mask, prot, init);
710 			continue;
711 		}
712 
713 		pud = alloc_low_page();
714 		paddr_last = phys_pud_init(pud, paddr, __pa(vaddr_end),
715 					   page_size_mask, prot, init);
716 
717 		spin_lock(&init_mm.page_table_lock);
718 		p4d_populate_init(&init_mm, p4d, pud, init);
719 		spin_unlock(&init_mm.page_table_lock);
720 	}
721 
722 	return paddr_last;
723 }
724 
725 static unsigned long __meminit
726 __kernel_physical_mapping_init(unsigned long paddr_start,
727 			       unsigned long paddr_end,
728 			       unsigned long page_size_mask,
729 			       pgprot_t prot, bool init)
730 {
731 	bool pgd_changed = false;
732 	unsigned long vaddr, vaddr_start, vaddr_end, vaddr_next, paddr_last;
733 
734 	paddr_last = paddr_end;
735 	vaddr = (unsigned long)__va(paddr_start);
736 	vaddr_end = (unsigned long)__va(paddr_end);
737 	vaddr_start = vaddr;
738 
739 	for (; vaddr < vaddr_end; vaddr = vaddr_next) {
740 		pgd_t *pgd = pgd_offset_k(vaddr);
741 		p4d_t *p4d;
742 
743 		vaddr_next = (vaddr & PGDIR_MASK) + PGDIR_SIZE;
744 
745 		if (pgd_val(*pgd)) {
746 			p4d = (p4d_t *)pgd_page_vaddr(*pgd);
747 			paddr_last = phys_p4d_init(p4d, __pa(vaddr),
748 						   __pa(vaddr_end),
749 						   page_size_mask,
750 						   prot, init);
751 			continue;
752 		}
753 
754 		p4d = alloc_low_page();
755 		paddr_last = phys_p4d_init(p4d, __pa(vaddr), __pa(vaddr_end),
756 					   page_size_mask, prot, init);
757 
758 		spin_lock(&init_mm.page_table_lock);
759 		if (pgtable_l5_enabled())
760 			pgd_populate_init(&init_mm, pgd, p4d, init);
761 		else
762 			p4d_populate_init(&init_mm, p4d_offset(pgd, vaddr),
763 					  (pud_t *) p4d, init);
764 
765 		spin_unlock(&init_mm.page_table_lock);
766 		pgd_changed = true;
767 	}
768 
769 	if (pgd_changed)
770 		sync_global_pgds(vaddr_start, vaddr_end - 1);
771 
772 	return paddr_last;
773 }
774 
775 
776 /*
777  * Create page table mapping for the physical memory for specific physical
778  * addresses. Note that it can only be used to populate non-present entries.
779  * The virtual and physical addresses have to be aligned on PMD level
780  * down. It returns the last physical address mapped.
781  */
782 unsigned long __meminit
783 kernel_physical_mapping_init(unsigned long paddr_start,
784 			     unsigned long paddr_end,
785 			     unsigned long page_size_mask, pgprot_t prot)
786 {
787 	return __kernel_physical_mapping_init(paddr_start, paddr_end,
788 					      page_size_mask, prot, true);
789 }
790 
791 /*
792  * This function is similar to kernel_physical_mapping_init() above with the
793  * exception that it uses set_{pud,pmd}() instead of the set_{pud,pte}_safe()
794  * when updating the mapping. The caller is responsible to flush the TLBs after
795  * the function returns.
796  */
797 unsigned long __meminit
798 kernel_physical_mapping_change(unsigned long paddr_start,
799 			       unsigned long paddr_end,
800 			       unsigned long page_size_mask)
801 {
802 	return __kernel_physical_mapping_init(paddr_start, paddr_end,
803 					      page_size_mask, PAGE_KERNEL,
804 					      false);
805 }
806 
807 #ifndef CONFIG_NUMA
808 void __init initmem_init(void)
809 {
810 	memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
811 }
812 #endif
813 
814 void __init paging_init(void)
815 {
816 	sparse_init();
817 
818 	/*
819 	 * clear the default setting with node 0
820 	 * note: don't use nodes_clear here, that is really clearing when
821 	 *	 numa support is not compiled in, and later node_set_state
822 	 *	 will not set it back.
823 	 */
824 	node_clear_state(0, N_MEMORY);
825 	node_clear_state(0, N_NORMAL_MEMORY);
826 
827 	zone_sizes_init();
828 }
829 
830 #ifdef CONFIG_SPARSEMEM_VMEMMAP
831 #define PAGE_UNUSED 0xFD
832 
833 /*
834  * The unused vmemmap range, which was not yet memset(PAGE_UNUSED), ranges
835  * from unused_pmd_start to next PMD_SIZE boundary.
836  */
837 static unsigned long unused_pmd_start __meminitdata;
838 
839 static void __meminit vmemmap_flush_unused_pmd(void)
840 {
841 	if (!unused_pmd_start)
842 		return;
843 	/*
844 	 * Clears (unused_pmd_start, PMD_END]
845 	 */
846 	memset((void *)unused_pmd_start, PAGE_UNUSED,
847 	       ALIGN(unused_pmd_start, PMD_SIZE) - unused_pmd_start);
848 	unused_pmd_start = 0;
849 }
850 
851 #ifdef CONFIG_MEMORY_HOTPLUG
852 /* Returns true if the PMD is completely unused and thus it can be freed */
853 static bool __meminit vmemmap_pmd_is_unused(unsigned long addr, unsigned long end)
854 {
855 	unsigned long start = ALIGN_DOWN(addr, PMD_SIZE);
856 
857 	/*
858 	 * Flush the unused range cache to ensure that memchr_inv() will work
859 	 * for the whole range.
860 	 */
861 	vmemmap_flush_unused_pmd();
862 	memset((void *)addr, PAGE_UNUSED, end - addr);
863 
864 	return !memchr_inv((void *)start, PAGE_UNUSED, PMD_SIZE);
865 }
866 #endif
867 
868 static void __meminit __vmemmap_use_sub_pmd(unsigned long start)
869 {
870 	/*
871 	 * As we expect to add in the same granularity as we remove, it's
872 	 * sufficient to mark only some piece used to block the memmap page from
873 	 * getting removed when removing some other adjacent memmap (just in
874 	 * case the first memmap never gets initialized e.g., because the memory
875 	 * block never gets onlined).
876 	 */
877 	memset((void *)start, 0, sizeof(struct page));
878 }
879 
880 static void __meminit vmemmap_use_sub_pmd(unsigned long start, unsigned long end)
881 {
882 	/*
883 	 * We only optimize if the new used range directly follows the
884 	 * previously unused range (esp., when populating consecutive sections).
885 	 */
886 	if (unused_pmd_start == start) {
887 		if (likely(IS_ALIGNED(end, PMD_SIZE)))
888 			unused_pmd_start = 0;
889 		else
890 			unused_pmd_start = end;
891 		return;
892 	}
893 
894 	/*
895 	 * If the range does not contiguously follows previous one, make sure
896 	 * to mark the unused range of the previous one so it can be removed.
897 	 */
898 	vmemmap_flush_unused_pmd();
899 	__vmemmap_use_sub_pmd(start);
900 }
901 
902 
903 static void __meminit vmemmap_use_new_sub_pmd(unsigned long start, unsigned long end)
904 {
905 	const unsigned long page = ALIGN_DOWN(start, PMD_SIZE);
906 
907 	vmemmap_flush_unused_pmd();
908 
909 	/*
910 	 * Could be our memmap page is filled with PAGE_UNUSED already from a
911 	 * previous remove. Make sure to reset it.
912 	 */
913 	__vmemmap_use_sub_pmd(start);
914 
915 	/*
916 	 * Mark with PAGE_UNUSED the unused parts of the new memmap range
917 	 */
918 	if (!IS_ALIGNED(start, PMD_SIZE))
919 		memset((void *)page, PAGE_UNUSED, start - page);
920 
921 	/*
922 	 * We want to avoid memset(PAGE_UNUSED) when populating the vmemmap of
923 	 * consecutive sections. Remember for the last added PMD where the
924 	 * unused range begins.
925 	 */
926 	if (!IS_ALIGNED(end, PMD_SIZE))
927 		unused_pmd_start = end;
928 }
929 #endif
930 
931 /*
932  * Memory hotplug specific functions
933  */
934 #ifdef CONFIG_MEMORY_HOTPLUG
935 /*
936  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
937  * updating.
938  */
939 static void update_end_of_memory_vars(u64 start, u64 size)
940 {
941 	unsigned long end_pfn = PFN_UP(start + size);
942 
943 	if (end_pfn > max_pfn) {
944 		max_pfn = end_pfn;
945 		max_low_pfn = end_pfn;
946 		high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
947 	}
948 }
949 
950 int add_pages(int nid, unsigned long start_pfn, unsigned long nr_pages,
951 	      struct mhp_params *params)
952 {
953 	int ret;
954 
955 	ret = __add_pages(nid, start_pfn, nr_pages, params);
956 	WARN_ON_ONCE(ret);
957 
958 	/* update max_pfn, max_low_pfn and high_memory */
959 	update_end_of_memory_vars(start_pfn << PAGE_SHIFT,
960 				  nr_pages << PAGE_SHIFT);
961 
962 	return ret;
963 }
964 
965 int arch_add_memory(int nid, u64 start, u64 size,
966 		    struct mhp_params *params)
967 {
968 	unsigned long start_pfn = start >> PAGE_SHIFT;
969 	unsigned long nr_pages = size >> PAGE_SHIFT;
970 
971 	init_memory_mapping(start, start + size, params->pgprot);
972 
973 	return add_pages(nid, start_pfn, nr_pages, params);
974 }
975 
976 static void __meminit free_pagetable(struct page *page, int order)
977 {
978 	unsigned long magic;
979 	unsigned int nr_pages = 1 << order;
980 
981 	/* bootmem page has reserved flag */
982 	if (PageReserved(page)) {
983 		__ClearPageReserved(page);
984 
985 		magic = page->index;
986 		if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
987 			while (nr_pages--)
988 				put_page_bootmem(page++);
989 		} else
990 			while (nr_pages--)
991 				free_reserved_page(page++);
992 	} else
993 		free_pages((unsigned long)page_address(page), order);
994 }
995 
996 static void __meminit free_hugepage_table(struct page *page,
997 		struct vmem_altmap *altmap)
998 {
999 	if (altmap)
1000 		vmem_altmap_free(altmap, PMD_SIZE / PAGE_SIZE);
1001 	else
1002 		free_pagetable(page, get_order(PMD_SIZE));
1003 }
1004 
1005 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
1006 {
1007 	pte_t *pte;
1008 	int i;
1009 
1010 	for (i = 0; i < PTRS_PER_PTE; i++) {
1011 		pte = pte_start + i;
1012 		if (!pte_none(*pte))
1013 			return;
1014 	}
1015 
1016 	/* free a pte talbe */
1017 	free_pagetable(pmd_page(*pmd), 0);
1018 	spin_lock(&init_mm.page_table_lock);
1019 	pmd_clear(pmd);
1020 	spin_unlock(&init_mm.page_table_lock);
1021 }
1022 
1023 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
1024 {
1025 	pmd_t *pmd;
1026 	int i;
1027 
1028 	for (i = 0; i < PTRS_PER_PMD; i++) {
1029 		pmd = pmd_start + i;
1030 		if (!pmd_none(*pmd))
1031 			return;
1032 	}
1033 
1034 	/* free a pmd talbe */
1035 	free_pagetable(pud_page(*pud), 0);
1036 	spin_lock(&init_mm.page_table_lock);
1037 	pud_clear(pud);
1038 	spin_unlock(&init_mm.page_table_lock);
1039 }
1040 
1041 static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
1042 {
1043 	pud_t *pud;
1044 	int i;
1045 
1046 	for (i = 0; i < PTRS_PER_PUD; i++) {
1047 		pud = pud_start + i;
1048 		if (!pud_none(*pud))
1049 			return;
1050 	}
1051 
1052 	/* free a pud talbe */
1053 	free_pagetable(p4d_page(*p4d), 0);
1054 	spin_lock(&init_mm.page_table_lock);
1055 	p4d_clear(p4d);
1056 	spin_unlock(&init_mm.page_table_lock);
1057 }
1058 
1059 static void __meminit
1060 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
1061 		 bool direct)
1062 {
1063 	unsigned long next, pages = 0;
1064 	pte_t *pte;
1065 	phys_addr_t phys_addr;
1066 
1067 	pte = pte_start + pte_index(addr);
1068 	for (; addr < end; addr = next, pte++) {
1069 		next = (addr + PAGE_SIZE) & PAGE_MASK;
1070 		if (next > end)
1071 			next = end;
1072 
1073 		if (!pte_present(*pte))
1074 			continue;
1075 
1076 		/*
1077 		 * We mapped [0,1G) memory as identity mapping when
1078 		 * initializing, in arch/x86/kernel/head_64.S. These
1079 		 * pagetables cannot be removed.
1080 		 */
1081 		phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
1082 		if (phys_addr < (phys_addr_t)0x40000000)
1083 			return;
1084 
1085 		if (!direct)
1086 			free_pagetable(pte_page(*pte), 0);
1087 
1088 		spin_lock(&init_mm.page_table_lock);
1089 		pte_clear(&init_mm, addr, pte);
1090 		spin_unlock(&init_mm.page_table_lock);
1091 
1092 		/* For non-direct mapping, pages means nothing. */
1093 		pages++;
1094 	}
1095 
1096 	/* Call free_pte_table() in remove_pmd_table(). */
1097 	flush_tlb_all();
1098 	if (direct)
1099 		update_page_count(PG_LEVEL_4K, -pages);
1100 }
1101 
1102 static void __meminit
1103 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
1104 		 bool direct, struct vmem_altmap *altmap)
1105 {
1106 	unsigned long next, pages = 0;
1107 	pte_t *pte_base;
1108 	pmd_t *pmd;
1109 
1110 	pmd = pmd_start + pmd_index(addr);
1111 	for (; addr < end; addr = next, pmd++) {
1112 		next = pmd_addr_end(addr, end);
1113 
1114 		if (!pmd_present(*pmd))
1115 			continue;
1116 
1117 		if (pmd_large(*pmd)) {
1118 			if (IS_ALIGNED(addr, PMD_SIZE) &&
1119 			    IS_ALIGNED(next, PMD_SIZE)) {
1120 				if (!direct)
1121 					free_hugepage_table(pmd_page(*pmd),
1122 							    altmap);
1123 
1124 				spin_lock(&init_mm.page_table_lock);
1125 				pmd_clear(pmd);
1126 				spin_unlock(&init_mm.page_table_lock);
1127 				pages++;
1128 			}
1129 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1130 			else if (vmemmap_pmd_is_unused(addr, next)) {
1131 					free_hugepage_table(pmd_page(*pmd),
1132 							    altmap);
1133 					spin_lock(&init_mm.page_table_lock);
1134 					pmd_clear(pmd);
1135 					spin_unlock(&init_mm.page_table_lock);
1136 			}
1137 #endif
1138 			continue;
1139 		}
1140 
1141 		pte_base = (pte_t *)pmd_page_vaddr(*pmd);
1142 		remove_pte_table(pte_base, addr, next, direct);
1143 		free_pte_table(pte_base, pmd);
1144 	}
1145 
1146 	/* Call free_pmd_table() in remove_pud_table(). */
1147 	if (direct)
1148 		update_page_count(PG_LEVEL_2M, -pages);
1149 }
1150 
1151 static void __meminit
1152 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
1153 		 struct vmem_altmap *altmap, bool direct)
1154 {
1155 	unsigned long next, pages = 0;
1156 	pmd_t *pmd_base;
1157 	pud_t *pud;
1158 
1159 	pud = pud_start + pud_index(addr);
1160 	for (; addr < end; addr = next, pud++) {
1161 		next = pud_addr_end(addr, end);
1162 
1163 		if (!pud_present(*pud))
1164 			continue;
1165 
1166 		if (pud_leaf(*pud) &&
1167 		    IS_ALIGNED(addr, PUD_SIZE) &&
1168 		    IS_ALIGNED(next, PUD_SIZE)) {
1169 			spin_lock(&init_mm.page_table_lock);
1170 			pud_clear(pud);
1171 			spin_unlock(&init_mm.page_table_lock);
1172 			pages++;
1173 			continue;
1174 		}
1175 
1176 		pmd_base = pmd_offset(pud, 0);
1177 		remove_pmd_table(pmd_base, addr, next, direct, altmap);
1178 		free_pmd_table(pmd_base, pud);
1179 	}
1180 
1181 	if (direct)
1182 		update_page_count(PG_LEVEL_1G, -pages);
1183 }
1184 
1185 static void __meminit
1186 remove_p4d_table(p4d_t *p4d_start, unsigned long addr, unsigned long end,
1187 		 struct vmem_altmap *altmap, bool direct)
1188 {
1189 	unsigned long next, pages = 0;
1190 	pud_t *pud_base;
1191 	p4d_t *p4d;
1192 
1193 	p4d = p4d_start + p4d_index(addr);
1194 	for (; addr < end; addr = next, p4d++) {
1195 		next = p4d_addr_end(addr, end);
1196 
1197 		if (!p4d_present(*p4d))
1198 			continue;
1199 
1200 		BUILD_BUG_ON(p4d_large(*p4d));
1201 
1202 		pud_base = pud_offset(p4d, 0);
1203 		remove_pud_table(pud_base, addr, next, altmap, direct);
1204 		/*
1205 		 * For 4-level page tables we do not want to free PUDs, but in the
1206 		 * 5-level case we should free them. This code will have to change
1207 		 * to adapt for boot-time switching between 4 and 5 level page tables.
1208 		 */
1209 		if (pgtable_l5_enabled())
1210 			free_pud_table(pud_base, p4d);
1211 	}
1212 
1213 	if (direct)
1214 		update_page_count(PG_LEVEL_512G, -pages);
1215 }
1216 
1217 /* start and end are both virtual address. */
1218 static void __meminit
1219 remove_pagetable(unsigned long start, unsigned long end, bool direct,
1220 		struct vmem_altmap *altmap)
1221 {
1222 	unsigned long next;
1223 	unsigned long addr;
1224 	pgd_t *pgd;
1225 	p4d_t *p4d;
1226 
1227 	for (addr = start; addr < end; addr = next) {
1228 		next = pgd_addr_end(addr, end);
1229 
1230 		pgd = pgd_offset_k(addr);
1231 		if (!pgd_present(*pgd))
1232 			continue;
1233 
1234 		p4d = p4d_offset(pgd, 0);
1235 		remove_p4d_table(p4d, addr, next, altmap, direct);
1236 	}
1237 
1238 	flush_tlb_all();
1239 }
1240 
1241 void __ref vmemmap_free(unsigned long start, unsigned long end,
1242 		struct vmem_altmap *altmap)
1243 {
1244 	VM_BUG_ON(!PAGE_ALIGNED(start));
1245 	VM_BUG_ON(!PAGE_ALIGNED(end));
1246 
1247 	remove_pagetable(start, end, false, altmap);
1248 }
1249 
1250 static void __meminit
1251 kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1252 {
1253 	start = (unsigned long)__va(start);
1254 	end = (unsigned long)__va(end);
1255 
1256 	remove_pagetable(start, end, true, NULL);
1257 }
1258 
1259 void __ref arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap)
1260 {
1261 	unsigned long start_pfn = start >> PAGE_SHIFT;
1262 	unsigned long nr_pages = size >> PAGE_SHIFT;
1263 
1264 	__remove_pages(start_pfn, nr_pages, altmap);
1265 	kernel_physical_mapping_remove(start, start + size);
1266 }
1267 #endif /* CONFIG_MEMORY_HOTPLUG */
1268 
1269 static struct kcore_list kcore_vsyscall;
1270 
1271 static void __init register_page_bootmem_info(void)
1272 {
1273 #if defined(CONFIG_NUMA) || defined(CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP)
1274 	int i;
1275 
1276 	for_each_online_node(i)
1277 		register_page_bootmem_info_node(NODE_DATA(i));
1278 #endif
1279 }
1280 
1281 /*
1282  * Pre-allocates page-table pages for the vmalloc area in the kernel page-table.
1283  * Only the level which needs to be synchronized between all page-tables is
1284  * allocated because the synchronization can be expensive.
1285  */
1286 static void __init preallocate_vmalloc_pages(void)
1287 {
1288 	unsigned long addr;
1289 	const char *lvl;
1290 
1291 	for (addr = VMALLOC_START; addr <= VMEMORY_END; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
1292 		pgd_t *pgd = pgd_offset_k(addr);
1293 		p4d_t *p4d;
1294 		pud_t *pud;
1295 
1296 		lvl = "p4d";
1297 		p4d = p4d_alloc(&init_mm, pgd, addr);
1298 		if (!p4d)
1299 			goto failed;
1300 
1301 		if (pgtable_l5_enabled())
1302 			continue;
1303 
1304 		/*
1305 		 * The goal here is to allocate all possibly required
1306 		 * hardware page tables pointed to by the top hardware
1307 		 * level.
1308 		 *
1309 		 * On 4-level systems, the P4D layer is folded away and
1310 		 * the above code does no preallocation.  Below, go down
1311 		 * to the pud _software_ level to ensure the second
1312 		 * hardware level is allocated on 4-level systems too.
1313 		 */
1314 		lvl = "pud";
1315 		pud = pud_alloc(&init_mm, p4d, addr);
1316 		if (!pud)
1317 			goto failed;
1318 	}
1319 
1320 	return;
1321 
1322 failed:
1323 
1324 	/*
1325 	 * The pages have to be there now or they will be missing in
1326 	 * process page-tables later.
1327 	 */
1328 	panic("Failed to pre-allocate %s pages for vmalloc area\n", lvl);
1329 }
1330 
1331 void __init mem_init(void)
1332 {
1333 	pci_iommu_alloc();
1334 
1335 	/* clear_bss() already clear the empty_zero_page */
1336 
1337 	/* this will put all memory onto the freelists */
1338 	memblock_free_all();
1339 	after_bootmem = 1;
1340 	x86_init.hyper.init_after_bootmem();
1341 
1342 	/*
1343 	 * Must be done after boot memory is put on freelist, because here we
1344 	 * might set fields in deferred struct pages that have not yet been
1345 	 * initialized, and memblock_free_all() initializes all the reserved
1346 	 * deferred pages for us.
1347 	 */
1348 	register_page_bootmem_info();
1349 
1350 	/* Register memory areas for /proc/kcore */
1351 	if (get_gate_vma(&init_mm))
1352 		kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR, PAGE_SIZE, KCORE_USER);
1353 
1354 	preallocate_vmalloc_pages();
1355 }
1356 
1357 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1358 int __init deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1359 {
1360 	/*
1361 	 * More CPUs always led to greater speedups on tested systems, up to
1362 	 * all the nodes' CPUs.  Use all since the system is otherwise idle
1363 	 * now.
1364 	 */
1365 	return max_t(int, cpumask_weight(node_cpumask), 1);
1366 }
1367 #endif
1368 
1369 int kernel_set_to_readonly;
1370 
1371 void mark_rodata_ro(void)
1372 {
1373 	unsigned long start = PFN_ALIGN(_text);
1374 	unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1375 	unsigned long end = (unsigned long)__end_rodata_hpage_align;
1376 	unsigned long text_end = PFN_ALIGN(_etext);
1377 	unsigned long rodata_end = PFN_ALIGN(__end_rodata);
1378 	unsigned long all_end;
1379 
1380 	printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1381 	       (end - start) >> 10);
1382 	set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1383 
1384 	kernel_set_to_readonly = 1;
1385 
1386 	/*
1387 	 * The rodata/data/bss/brk section (but not the kernel text!)
1388 	 * should also be not-executable.
1389 	 *
1390 	 * We align all_end to PMD_SIZE because the existing mapping
1391 	 * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1392 	 * split the PMD and the reminder between _brk_end and the end
1393 	 * of the PMD will remain mapped executable.
1394 	 *
1395 	 * Any PMD which was setup after the one which covers _brk_end
1396 	 * has been zapped already via cleanup_highmem().
1397 	 */
1398 	all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1399 	set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
1400 
1401 	set_ftrace_ops_ro();
1402 
1403 #ifdef CONFIG_CPA_DEBUG
1404 	printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1405 	set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1406 
1407 	printk(KERN_INFO "Testing CPA: again\n");
1408 	set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1409 #endif
1410 
1411 	free_kernel_image_pages("unused kernel image (text/rodata gap)",
1412 				(void *)text_end, (void *)rodata_start);
1413 	free_kernel_image_pages("unused kernel image (rodata/data gap)",
1414 				(void *)rodata_end, (void *)_sdata);
1415 
1416 	debug_checkwx();
1417 }
1418 
1419 /*
1420  * Block size is the minimum amount of memory which can be hotplugged or
1421  * hotremoved. It must be power of two and must be equal or larger than
1422  * MIN_MEMORY_BLOCK_SIZE.
1423  */
1424 #define MAX_BLOCK_SIZE (2UL << 30)
1425 
1426 /* Amount of ram needed to start using large blocks */
1427 #define MEM_SIZE_FOR_LARGE_BLOCK (64UL << 30)
1428 
1429 /* Adjustable memory block size */
1430 static unsigned long set_memory_block_size;
1431 int __init set_memory_block_size_order(unsigned int order)
1432 {
1433 	unsigned long size = 1UL << order;
1434 
1435 	if (size > MEM_SIZE_FOR_LARGE_BLOCK || size < MIN_MEMORY_BLOCK_SIZE)
1436 		return -EINVAL;
1437 
1438 	set_memory_block_size = size;
1439 	return 0;
1440 }
1441 
1442 static unsigned long probe_memory_block_size(void)
1443 {
1444 	unsigned long boot_mem_end = max_pfn << PAGE_SHIFT;
1445 	unsigned long bz;
1446 
1447 	/* If memory block size has been set, then use it */
1448 	bz = set_memory_block_size;
1449 	if (bz)
1450 		goto done;
1451 
1452 	/* Use regular block if RAM is smaller than MEM_SIZE_FOR_LARGE_BLOCK */
1453 	if (boot_mem_end < MEM_SIZE_FOR_LARGE_BLOCK) {
1454 		bz = MIN_MEMORY_BLOCK_SIZE;
1455 		goto done;
1456 	}
1457 
1458 	/*
1459 	 * Use max block size to minimize overhead on bare metal, where
1460 	 * alignment for memory hotplug isn't a concern.
1461 	 */
1462 	if (!boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
1463 		bz = MAX_BLOCK_SIZE;
1464 		goto done;
1465 	}
1466 
1467 	/* Find the largest allowed block size that aligns to memory end */
1468 	for (bz = MAX_BLOCK_SIZE; bz > MIN_MEMORY_BLOCK_SIZE; bz >>= 1) {
1469 		if (IS_ALIGNED(boot_mem_end, bz))
1470 			break;
1471 	}
1472 done:
1473 	pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20);
1474 
1475 	return bz;
1476 }
1477 
1478 static unsigned long memory_block_size_probed;
1479 unsigned long memory_block_size_bytes(void)
1480 {
1481 	if (!memory_block_size_probed)
1482 		memory_block_size_probed = probe_memory_block_size();
1483 
1484 	return memory_block_size_probed;
1485 }
1486 
1487 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1488 /*
1489  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1490  */
1491 static long __meminitdata addr_start, addr_end;
1492 static void __meminitdata *p_start, *p_end;
1493 static int __meminitdata node_start;
1494 
1495 void __meminit vmemmap_set_pmd(pmd_t *pmd, void *p, int node,
1496 			       unsigned long addr, unsigned long next)
1497 {
1498 	pte_t entry;
1499 
1500 	entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1501 			PAGE_KERNEL_LARGE);
1502 	set_pmd(pmd, __pmd(pte_val(entry)));
1503 
1504 	/* check to see if we have contiguous blocks */
1505 	if (p_end != p || node_start != node) {
1506 		if (p_start)
1507 			pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1508 				addr_start, addr_end-1, p_start, p_end-1, node_start);
1509 		addr_start = addr;
1510 		node_start = node;
1511 		p_start = p;
1512 	}
1513 
1514 	addr_end = addr + PMD_SIZE;
1515 	p_end = p + PMD_SIZE;
1516 
1517 	if (!IS_ALIGNED(addr, PMD_SIZE) ||
1518 		!IS_ALIGNED(next, PMD_SIZE))
1519 		vmemmap_use_new_sub_pmd(addr, next);
1520 }
1521 
1522 int __meminit vmemmap_check_pmd(pmd_t *pmd, int node,
1523 				unsigned long addr, unsigned long next)
1524 {
1525 	int large = pmd_large(*pmd);
1526 
1527 	if (pmd_large(*pmd)) {
1528 		vmemmap_verify((pte_t *)pmd, node, addr, next);
1529 		vmemmap_use_sub_pmd(addr, next);
1530 	}
1531 
1532 	return large;
1533 }
1534 
1535 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1536 		struct vmem_altmap *altmap)
1537 {
1538 	int err;
1539 
1540 	VM_BUG_ON(!PAGE_ALIGNED(start));
1541 	VM_BUG_ON(!PAGE_ALIGNED(end));
1542 
1543 	if (end - start < PAGES_PER_SECTION * sizeof(struct page))
1544 		err = vmemmap_populate_basepages(start, end, node, NULL);
1545 	else if (boot_cpu_has(X86_FEATURE_PSE))
1546 		err = vmemmap_populate_hugepages(start, end, node, altmap);
1547 	else if (altmap) {
1548 		pr_err_once("%s: no cpu support for altmap allocations\n",
1549 				__func__);
1550 		err = -ENOMEM;
1551 	} else
1552 		err = vmemmap_populate_basepages(start, end, node, NULL);
1553 	if (!err)
1554 		sync_global_pgds(start, end - 1);
1555 	return err;
1556 }
1557 
1558 #ifdef CONFIG_HAVE_BOOTMEM_INFO_NODE
1559 void register_page_bootmem_memmap(unsigned long section_nr,
1560 				  struct page *start_page, unsigned long nr_pages)
1561 {
1562 	unsigned long addr = (unsigned long)start_page;
1563 	unsigned long end = (unsigned long)(start_page + nr_pages);
1564 	unsigned long next;
1565 	pgd_t *pgd;
1566 	p4d_t *p4d;
1567 	pud_t *pud;
1568 	pmd_t *pmd;
1569 	unsigned int nr_pmd_pages;
1570 	struct page *page;
1571 
1572 	for (; addr < end; addr = next) {
1573 		pte_t *pte = NULL;
1574 
1575 		pgd = pgd_offset_k(addr);
1576 		if (pgd_none(*pgd)) {
1577 			next = (addr + PAGE_SIZE) & PAGE_MASK;
1578 			continue;
1579 		}
1580 		get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1581 
1582 		p4d = p4d_offset(pgd, addr);
1583 		if (p4d_none(*p4d)) {
1584 			next = (addr + PAGE_SIZE) & PAGE_MASK;
1585 			continue;
1586 		}
1587 		get_page_bootmem(section_nr, p4d_page(*p4d), MIX_SECTION_INFO);
1588 
1589 		pud = pud_offset(p4d, addr);
1590 		if (pud_none(*pud)) {
1591 			next = (addr + PAGE_SIZE) & PAGE_MASK;
1592 			continue;
1593 		}
1594 		get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1595 
1596 		if (!boot_cpu_has(X86_FEATURE_PSE)) {
1597 			next = (addr + PAGE_SIZE) & PAGE_MASK;
1598 			pmd = pmd_offset(pud, addr);
1599 			if (pmd_none(*pmd))
1600 				continue;
1601 			get_page_bootmem(section_nr, pmd_page(*pmd),
1602 					 MIX_SECTION_INFO);
1603 
1604 			pte = pte_offset_kernel(pmd, addr);
1605 			if (pte_none(*pte))
1606 				continue;
1607 			get_page_bootmem(section_nr, pte_page(*pte),
1608 					 SECTION_INFO);
1609 		} else {
1610 			next = pmd_addr_end(addr, end);
1611 
1612 			pmd = pmd_offset(pud, addr);
1613 			if (pmd_none(*pmd))
1614 				continue;
1615 
1616 			nr_pmd_pages = 1 << get_order(PMD_SIZE);
1617 			page = pmd_page(*pmd);
1618 			while (nr_pmd_pages--)
1619 				get_page_bootmem(section_nr, page++,
1620 						 SECTION_INFO);
1621 		}
1622 	}
1623 }
1624 #endif
1625 
1626 void __meminit vmemmap_populate_print_last(void)
1627 {
1628 	if (p_start) {
1629 		pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1630 			addr_start, addr_end-1, p_start, p_end-1, node_start);
1631 		p_start = NULL;
1632 		p_end = NULL;
1633 		node_start = 0;
1634 	}
1635 }
1636 #endif
1637