Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/arch/x86_64/mm/init.c
4 *
5 * Copyright (C) 1995 Linus Torvalds
6 * Copyright (C) 2000 Pavel Machek <pavel@ucw.cz>
7 * Copyright (C) 2002,2003 Andi Kleen <ak@suse.de>
8 */
9
10#include <linux/signal.h>
11#include <linux/sched.h>
12#include <linux/kernel.h>
13#include <linux/errno.h>
14#include <linux/string.h>
15#include <linux/types.h>
16#include <linux/ptrace.h>
17#include <linux/mman.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/smp.h>
21#include <linux/init.h>
22#include <linux/initrd.h>
23#include <linux/pagemap.h>
24#include <linux/memblock.h>
25#include <linux/proc_fs.h>
26#include <linux/pci.h>
27#include <linux/pfn.h>
28#include <linux/poison.h>
29#include <linux/dma-mapping.h>
30#include <linux/memory.h>
31#include <linux/memory_hotplug.h>
32#include <linux/memremap.h>
33#include <linux/nmi.h>
34#include <linux/gfp.h>
35#include <linux/kcore.h>
36
37#include <asm/processor.h>
38#include <asm/bios_ebda.h>
39#include <linux/uaccess.h>
40#include <asm/pgtable.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
58#include "mm_internal.h"
59
60#include "ident_map.c"
61
62#define DEFINE_POPULATE(fname, type1, type2, init) \
63static inline void fname##_init(struct mm_struct *mm, \
64 type1##_t *arg1, type2##_t *arg2, bool init) \
65{ \
66 if (init) \
67 fname##_safe(mm, arg1, arg2); \
68 else \
69 fname(mm, arg1, arg2); \
70}
71
72DEFINE_POPULATE(p4d_populate, p4d, pud, init)
73DEFINE_POPULATE(pgd_populate, pgd, p4d, init)
74DEFINE_POPULATE(pud_populate, pud, pmd, init)
75DEFINE_POPULATE(pmd_populate_kernel, pmd, pte, init)
76
77#define DEFINE_ENTRY(type1, type2, init) \
78static inline void set_##type1##_init(type1##_t *arg1, \
79 type2##_t arg2, bool init) \
80{ \
81 if (init) \
82 set_##type1##_safe(arg1, arg2); \
83 else \
84 set_##type1(arg1, arg2); \
85}
86
87DEFINE_ENTRY(p4d, p4d, init)
88DEFINE_ENTRY(pud, pud, init)
89DEFINE_ENTRY(pmd, pmd, init)
90DEFINE_ENTRY(pte, pte, init)
91
92
93/*
94 * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
95 * physical space so we can cache the place of the first one and move
96 * around without checking the pgd every time.
97 */
98
99/* Bits supported by the hardware: */
100pteval_t __supported_pte_mask __read_mostly = ~0;
101/* Bits allowed in normal kernel mappings: */
102pteval_t __default_kernel_pte_mask __read_mostly = ~0;
103EXPORT_SYMBOL_GPL(__supported_pte_mask);
104/* Used in PAGE_KERNEL_* macros which are reasonably used out-of-tree: */
105EXPORT_SYMBOL(__default_kernel_pte_mask);
106
107int force_personality32;
108
109/*
110 * noexec32=on|off
111 * Control non executable heap for 32bit processes.
112 * To control the stack too use noexec=off
113 *
114 * on PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
115 * off PROT_READ implies PROT_EXEC
116 */
117static int __init nonx32_setup(char *str)
118{
119 if (!strcmp(str, "on"))
120 force_personality32 &= ~READ_IMPLIES_EXEC;
121 else if (!strcmp(str, "off"))
122 force_personality32 |= READ_IMPLIES_EXEC;
123 return 1;
124}
125__setup("noexec32=", nonx32_setup);
126
127static void sync_global_pgds_l5(unsigned long start, unsigned long end)
128{
129 unsigned long addr;
130
131 for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
132 const pgd_t *pgd_ref = pgd_offset_k(addr);
133 struct page *page;
134
135 /* Check for overflow */
136 if (addr < start)
137 break;
138
139 if (pgd_none(*pgd_ref))
140 continue;
141
142 spin_lock(&pgd_lock);
143 list_for_each_entry(page, &pgd_list, lru) {
144 pgd_t *pgd;
145 spinlock_t *pgt_lock;
146
147 pgd = (pgd_t *)page_address(page) + pgd_index(addr);
148 /* the pgt_lock only for Xen */
149 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
150 spin_lock(pgt_lock);
151
152 if (!pgd_none(*pgd_ref) && !pgd_none(*pgd))
153 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
154
155 if (pgd_none(*pgd))
156 set_pgd(pgd, *pgd_ref);
157
158 spin_unlock(pgt_lock);
159 }
160 spin_unlock(&pgd_lock);
161 }
162}
163
164static void sync_global_pgds_l4(unsigned long start, unsigned long end)
165{
166 unsigned long addr;
167
168 for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
169 pgd_t *pgd_ref = pgd_offset_k(addr);
170 const p4d_t *p4d_ref;
171 struct page *page;
172
173 /*
174 * With folded p4d, pgd_none() is always false, we need to
175 * handle synchonization on p4d level.
176 */
177 MAYBE_BUILD_BUG_ON(pgd_none(*pgd_ref));
178 p4d_ref = p4d_offset(pgd_ref, addr);
179
180 if (p4d_none(*p4d_ref))
181 continue;
182
183 spin_lock(&pgd_lock);
184 list_for_each_entry(page, &pgd_list, lru) {
185 pgd_t *pgd;
186 p4d_t *p4d;
187 spinlock_t *pgt_lock;
188
189 pgd = (pgd_t *)page_address(page) + pgd_index(addr);
190 p4d = p4d_offset(pgd, addr);
191 /* the pgt_lock only for Xen */
192 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
193 spin_lock(pgt_lock);
194
195 if (!p4d_none(*p4d_ref) && !p4d_none(*p4d))
196 BUG_ON(p4d_page_vaddr(*p4d)
197 != p4d_page_vaddr(*p4d_ref));
198
199 if (p4d_none(*p4d))
200 set_p4d(p4d, *p4d_ref);
201
202 spin_unlock(pgt_lock);
203 }
204 spin_unlock(&pgd_lock);
205 }
206}
207
208/*
209 * When memory was added make sure all the processes MM have
210 * suitable PGD entries in the local PGD level page.
211 */
212void sync_global_pgds(unsigned long start, unsigned long end)
213{
214 if (pgtable_l5_enabled())
215 sync_global_pgds_l5(start, end);
216 else
217 sync_global_pgds_l4(start, end);
218}
219
220/*
221 * NOTE: This function is marked __ref because it calls __init function
222 * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
223 */
224static __ref void *spp_getpage(void)
225{
226 void *ptr;
227
228 if (after_bootmem)
229 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
230 else
231 ptr = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
232
233 if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
234 panic("set_pte_phys: cannot allocate page data %s\n",
235 after_bootmem ? "after bootmem" : "");
236 }
237
238 pr_debug("spp_getpage %p\n", ptr);
239
240 return ptr;
241}
242
243static p4d_t *fill_p4d(pgd_t *pgd, unsigned long vaddr)
244{
245 if (pgd_none(*pgd)) {
246 p4d_t *p4d = (p4d_t *)spp_getpage();
247 pgd_populate(&init_mm, pgd, p4d);
248 if (p4d != p4d_offset(pgd, 0))
249 printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
250 p4d, p4d_offset(pgd, 0));
251 }
252 return p4d_offset(pgd, vaddr);
253}
254
255static pud_t *fill_pud(p4d_t *p4d, unsigned long vaddr)
256{
257 if (p4d_none(*p4d)) {
258 pud_t *pud = (pud_t *)spp_getpage();
259 p4d_populate(&init_mm, p4d, pud);
260 if (pud != pud_offset(p4d, 0))
261 printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
262 pud, pud_offset(p4d, 0));
263 }
264 return pud_offset(p4d, vaddr);
265}
266
267static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
268{
269 if (pud_none(*pud)) {
270 pmd_t *pmd = (pmd_t *) spp_getpage();
271 pud_populate(&init_mm, pud, pmd);
272 if (pmd != pmd_offset(pud, 0))
273 printk(KERN_ERR "PAGETABLE BUG #02! %p <-> %p\n",
274 pmd, pmd_offset(pud, 0));
275 }
276 return pmd_offset(pud, vaddr);
277}
278
279static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
280{
281 if (pmd_none(*pmd)) {
282 pte_t *pte = (pte_t *) spp_getpage();
283 pmd_populate_kernel(&init_mm, pmd, pte);
284 if (pte != pte_offset_kernel(pmd, 0))
285 printk(KERN_ERR "PAGETABLE BUG #03!\n");
286 }
287 return pte_offset_kernel(pmd, vaddr);
288}
289
290static void __set_pte_vaddr(pud_t *pud, unsigned long vaddr, pte_t new_pte)
291{
292 pmd_t *pmd = fill_pmd(pud, vaddr);
293 pte_t *pte = fill_pte(pmd, vaddr);
294
295 set_pte(pte, new_pte);
296
297 /*
298 * It's enough to flush this one mapping.
299 * (PGE mappings get flushed as well)
300 */
301 __flush_tlb_one_kernel(vaddr);
302}
303
304void set_pte_vaddr_p4d(p4d_t *p4d_page, unsigned long vaddr, pte_t new_pte)
305{
306 p4d_t *p4d = p4d_page + p4d_index(vaddr);
307 pud_t *pud = fill_pud(p4d, vaddr);
308
309 __set_pte_vaddr(pud, vaddr, new_pte);
310}
311
312void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
313{
314 pud_t *pud = pud_page + pud_index(vaddr);
315
316 __set_pte_vaddr(pud, vaddr, new_pte);
317}
318
319void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
320{
321 pgd_t *pgd;
322 p4d_t *p4d_page;
323
324 pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
325
326 pgd = pgd_offset_k(vaddr);
327 if (pgd_none(*pgd)) {
328 printk(KERN_ERR
329 "PGD FIXMAP MISSING, it should be setup in head.S!\n");
330 return;
331 }
332
333 p4d_page = p4d_offset(pgd, 0);
334 set_pte_vaddr_p4d(p4d_page, vaddr, pteval);
335}
336
337pmd_t * __init populate_extra_pmd(unsigned long vaddr)
338{
339 pgd_t *pgd;
340 p4d_t *p4d;
341 pud_t *pud;
342
343 pgd = pgd_offset_k(vaddr);
344 p4d = fill_p4d(pgd, vaddr);
345 pud = fill_pud(p4d, vaddr);
346 return fill_pmd(pud, vaddr);
347}
348
349pte_t * __init populate_extra_pte(unsigned long vaddr)
350{
351 pmd_t *pmd;
352
353 pmd = populate_extra_pmd(vaddr);
354 return fill_pte(pmd, vaddr);
355}
356
357/*
358 * Create large page table mappings for a range of physical addresses.
359 */
360static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
361 enum page_cache_mode cache)
362{
363 pgd_t *pgd;
364 p4d_t *p4d;
365 pud_t *pud;
366 pmd_t *pmd;
367 pgprot_t prot;
368
369 pgprot_val(prot) = pgprot_val(PAGE_KERNEL_LARGE) |
370 pgprot_val(pgprot_4k_2_large(cachemode2pgprot(cache)));
371 BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
372 for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
373 pgd = pgd_offset_k((unsigned long)__va(phys));
374 if (pgd_none(*pgd)) {
375 p4d = (p4d_t *) spp_getpage();
376 set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE |
377 _PAGE_USER));
378 }
379 p4d = p4d_offset(pgd, (unsigned long)__va(phys));
380 if (p4d_none(*p4d)) {
381 pud = (pud_t *) spp_getpage();
382 set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE |
383 _PAGE_USER));
384 }
385 pud = pud_offset(p4d, (unsigned long)__va(phys));
386 if (pud_none(*pud)) {
387 pmd = (pmd_t *) spp_getpage();
388 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
389 _PAGE_USER));
390 }
391 pmd = pmd_offset(pud, phys);
392 BUG_ON(!pmd_none(*pmd));
393 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
394 }
395}
396
397void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
398{
399 __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_WB);
400}
401
402void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
403{
404 __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_UC);
405}
406
407/*
408 * The head.S code sets up the kernel high mapping:
409 *
410 * from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
411 *
412 * phys_base holds the negative offset to the kernel, which is added
413 * to the compile time generated pmds. This results in invalid pmds up
414 * to the point where we hit the physaddr 0 mapping.
415 *
416 * We limit the mappings to the region from _text to _brk_end. _brk_end
417 * is rounded up to the 2MB boundary. This catches the invalid pmds as
418 * well, as they are located before _text:
419 */
420void __init cleanup_highmap(void)
421{
422 unsigned long vaddr = __START_KERNEL_map;
423 unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
424 unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
425 pmd_t *pmd = level2_kernel_pgt;
426
427 /*
428 * Native path, max_pfn_mapped is not set yet.
429 * Xen has valid max_pfn_mapped set in
430 * arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
431 */
432 if (max_pfn_mapped)
433 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
434
435 for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
436 if (pmd_none(*pmd))
437 continue;
438 if (vaddr < (unsigned long) _text || vaddr > end)
439 set_pmd(pmd, __pmd(0));
440 }
441}
442
443/*
444 * Create PTE level page table mapping for physical addresses.
445 * It returns the last physical address mapped.
446 */
447static unsigned long __meminit
448phys_pte_init(pte_t *pte_page, unsigned long paddr, unsigned long paddr_end,
449 pgprot_t prot, bool init)
450{
451 unsigned long pages = 0, paddr_next;
452 unsigned long paddr_last = paddr_end;
453 pte_t *pte;
454 int i;
455
456 pte = pte_page + pte_index(paddr);
457 i = pte_index(paddr);
458
459 for (; i < PTRS_PER_PTE; i++, paddr = paddr_next, pte++) {
460 paddr_next = (paddr & PAGE_MASK) + PAGE_SIZE;
461 if (paddr >= paddr_end) {
462 if (!after_bootmem &&
463 !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
464 E820_TYPE_RAM) &&
465 !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
466 E820_TYPE_RESERVED_KERN))
467 set_pte_init(pte, __pte(0), init);
468 continue;
469 }
470
471 /*
472 * We will re-use the existing mapping.
473 * Xen for example has some special requirements, like mapping
474 * pagetable pages as RO. So assume someone who pre-setup
475 * these mappings are more intelligent.
476 */
477 if (!pte_none(*pte)) {
478 if (!after_bootmem)
479 pages++;
480 continue;
481 }
482
483 if (0)
484 pr_info(" pte=%p addr=%lx pte=%016lx\n", pte, paddr,
485 pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL).pte);
486 pages++;
487 set_pte_init(pte, pfn_pte(paddr >> PAGE_SHIFT, prot), init);
488 paddr_last = (paddr & PAGE_MASK) + PAGE_SIZE;
489 }
490
491 update_page_count(PG_LEVEL_4K, pages);
492
493 return paddr_last;
494}
495
496/*
497 * Create PMD level page table mapping for physical addresses. The virtual
498 * and physical address have to be aligned at this level.
499 * It returns the last physical address mapped.
500 */
501static unsigned long __meminit
502phys_pmd_init(pmd_t *pmd_page, unsigned long paddr, unsigned long paddr_end,
503 unsigned long page_size_mask, pgprot_t prot, bool init)
504{
505 unsigned long pages = 0, paddr_next;
506 unsigned long paddr_last = paddr_end;
507
508 int i = pmd_index(paddr);
509
510 for (; i < PTRS_PER_PMD; i++, paddr = paddr_next) {
511 pmd_t *pmd = pmd_page + pmd_index(paddr);
512 pte_t *pte;
513 pgprot_t new_prot = prot;
514
515 paddr_next = (paddr & PMD_MASK) + PMD_SIZE;
516 if (paddr >= paddr_end) {
517 if (!after_bootmem &&
518 !e820__mapped_any(paddr & PMD_MASK, paddr_next,
519 E820_TYPE_RAM) &&
520 !e820__mapped_any(paddr & PMD_MASK, paddr_next,
521 E820_TYPE_RESERVED_KERN))
522 set_pmd_init(pmd, __pmd(0), init);
523 continue;
524 }
525
526 if (!pmd_none(*pmd)) {
527 if (!pmd_large(*pmd)) {
528 spin_lock(&init_mm.page_table_lock);
529 pte = (pte_t *)pmd_page_vaddr(*pmd);
530 paddr_last = phys_pte_init(pte, paddr,
531 paddr_end, prot,
532 init);
533 spin_unlock(&init_mm.page_table_lock);
534 continue;
535 }
536 /*
537 * If we are ok with PG_LEVEL_2M mapping, then we will
538 * use the existing mapping,
539 *
540 * Otherwise, we will split the large page mapping but
541 * use the same existing protection bits except for
542 * large page, so that we don't violate Intel's TLB
543 * Application note (317080) which says, while changing
544 * the page sizes, new and old translations should
545 * not differ with respect to page frame and
546 * attributes.
547 */
548 if (page_size_mask & (1 << PG_LEVEL_2M)) {
549 if (!after_bootmem)
550 pages++;
551 paddr_last = paddr_next;
552 continue;
553 }
554 new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
555 }
556
557 if (page_size_mask & (1<<PG_LEVEL_2M)) {
558 pages++;
559 spin_lock(&init_mm.page_table_lock);
560 set_pte_init((pte_t *)pmd,
561 pfn_pte((paddr & PMD_MASK) >> PAGE_SHIFT,
562 __pgprot(pgprot_val(prot) | _PAGE_PSE)),
563 init);
564 spin_unlock(&init_mm.page_table_lock);
565 paddr_last = paddr_next;
566 continue;
567 }
568
569 pte = alloc_low_page();
570 paddr_last = phys_pte_init(pte, paddr, paddr_end, new_prot, init);
571
572 spin_lock(&init_mm.page_table_lock);
573 pmd_populate_kernel_init(&init_mm, pmd, pte, init);
574 spin_unlock(&init_mm.page_table_lock);
575 }
576 update_page_count(PG_LEVEL_2M, pages);
577 return paddr_last;
578}
579
580/*
581 * Create PUD level page table mapping for physical addresses. The virtual
582 * and physical address do not have to be aligned at this level. KASLR can
583 * randomize virtual addresses up to this level.
584 * It returns the last physical address mapped.
585 */
586static unsigned long __meminit
587phys_pud_init(pud_t *pud_page, unsigned long paddr, unsigned long paddr_end,
588 unsigned long page_size_mask, bool init)
589{
590 unsigned long pages = 0, paddr_next;
591 unsigned long paddr_last = paddr_end;
592 unsigned long vaddr = (unsigned long)__va(paddr);
593 int i = pud_index(vaddr);
594
595 for (; i < PTRS_PER_PUD; i++, paddr = paddr_next) {
596 pud_t *pud;
597 pmd_t *pmd;
598 pgprot_t prot = PAGE_KERNEL;
599
600 vaddr = (unsigned long)__va(paddr);
601 pud = pud_page + pud_index(vaddr);
602 paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
603
604 if (paddr >= paddr_end) {
605 if (!after_bootmem &&
606 !e820__mapped_any(paddr & PUD_MASK, paddr_next,
607 E820_TYPE_RAM) &&
608 !e820__mapped_any(paddr & PUD_MASK, paddr_next,
609 E820_TYPE_RESERVED_KERN))
610 set_pud_init(pud, __pud(0), init);
611 continue;
612 }
613
614 if (!pud_none(*pud)) {
615 if (!pud_large(*pud)) {
616 pmd = pmd_offset(pud, 0);
617 paddr_last = phys_pmd_init(pmd, paddr,
618 paddr_end,
619 page_size_mask,
620 prot, init);
621 continue;
622 }
623 /*
624 * If we are ok with PG_LEVEL_1G mapping, then we will
625 * use the existing mapping.
626 *
627 * Otherwise, we will split the gbpage mapping but use
628 * the same existing protection bits except for large
629 * page, so that we don't violate Intel's TLB
630 * Application note (317080) which says, while changing
631 * the page sizes, new and old translations should
632 * not differ with respect to page frame and
633 * attributes.
634 */
635 if (page_size_mask & (1 << PG_LEVEL_1G)) {
636 if (!after_bootmem)
637 pages++;
638 paddr_last = paddr_next;
639 continue;
640 }
641 prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
642 }
643
644 if (page_size_mask & (1<<PG_LEVEL_1G)) {
645 pages++;
646 spin_lock(&init_mm.page_table_lock);
647 set_pte_init((pte_t *)pud,
648 pfn_pte((paddr & PUD_MASK) >> PAGE_SHIFT,
649 PAGE_KERNEL_LARGE),
650 init);
651 spin_unlock(&init_mm.page_table_lock);
652 paddr_last = paddr_next;
653 continue;
654 }
655
656 pmd = alloc_low_page();
657 paddr_last = phys_pmd_init(pmd, paddr, paddr_end,
658 page_size_mask, prot, init);
659
660 spin_lock(&init_mm.page_table_lock);
661 pud_populate_init(&init_mm, pud, pmd, init);
662 spin_unlock(&init_mm.page_table_lock);
663 }
664
665 update_page_count(PG_LEVEL_1G, pages);
666
667 return paddr_last;
668}
669
670static unsigned long __meminit
671phys_p4d_init(p4d_t *p4d_page, unsigned long paddr, unsigned long paddr_end,
672 unsigned long page_size_mask, bool init)
673{
674 unsigned long vaddr, vaddr_end, vaddr_next, paddr_next, paddr_last;
675
676 paddr_last = paddr_end;
677 vaddr = (unsigned long)__va(paddr);
678 vaddr_end = (unsigned long)__va(paddr_end);
679
680 if (!pgtable_l5_enabled())
681 return phys_pud_init((pud_t *) p4d_page, paddr, paddr_end,
682 page_size_mask, init);
683
684 for (; vaddr < vaddr_end; vaddr = vaddr_next) {
685 p4d_t *p4d = p4d_page + p4d_index(vaddr);
686 pud_t *pud;
687
688 vaddr_next = (vaddr & P4D_MASK) + P4D_SIZE;
689 paddr = __pa(vaddr);
690
691 if (paddr >= paddr_end) {
692 paddr_next = __pa(vaddr_next);
693 if (!after_bootmem &&
694 !e820__mapped_any(paddr & P4D_MASK, paddr_next,
695 E820_TYPE_RAM) &&
696 !e820__mapped_any(paddr & P4D_MASK, paddr_next,
697 E820_TYPE_RESERVED_KERN))
698 set_p4d_init(p4d, __p4d(0), init);
699 continue;
700 }
701
702 if (!p4d_none(*p4d)) {
703 pud = pud_offset(p4d, 0);
704 paddr_last = phys_pud_init(pud, paddr, __pa(vaddr_end),
705 page_size_mask, init);
706 continue;
707 }
708
709 pud = alloc_low_page();
710 paddr_last = phys_pud_init(pud, paddr, __pa(vaddr_end),
711 page_size_mask, init);
712
713 spin_lock(&init_mm.page_table_lock);
714 p4d_populate_init(&init_mm, p4d, pud, init);
715 spin_unlock(&init_mm.page_table_lock);
716 }
717
718 return paddr_last;
719}
720
721static unsigned long __meminit
722__kernel_physical_mapping_init(unsigned long paddr_start,
723 unsigned long paddr_end,
724 unsigned long page_size_mask,
725 bool init)
726{
727 bool pgd_changed = false;
728 unsigned long vaddr, vaddr_start, vaddr_end, vaddr_next, paddr_last;
729
730 paddr_last = paddr_end;
731 vaddr = (unsigned long)__va(paddr_start);
732 vaddr_end = (unsigned long)__va(paddr_end);
733 vaddr_start = vaddr;
734
735 for (; vaddr < vaddr_end; vaddr = vaddr_next) {
736 pgd_t *pgd = pgd_offset_k(vaddr);
737 p4d_t *p4d;
738
739 vaddr_next = (vaddr & PGDIR_MASK) + PGDIR_SIZE;
740
741 if (pgd_val(*pgd)) {
742 p4d = (p4d_t *)pgd_page_vaddr(*pgd);
743 paddr_last = phys_p4d_init(p4d, __pa(vaddr),
744 __pa(vaddr_end),
745 page_size_mask,
746 init);
747 continue;
748 }
749
750 p4d = alloc_low_page();
751 paddr_last = phys_p4d_init(p4d, __pa(vaddr), __pa(vaddr_end),
752 page_size_mask, init);
753
754 spin_lock(&init_mm.page_table_lock);
755 if (pgtable_l5_enabled())
756 pgd_populate_init(&init_mm, pgd, p4d, init);
757 else
758 p4d_populate_init(&init_mm, p4d_offset(pgd, vaddr),
759 (pud_t *) p4d, init);
760
761 spin_unlock(&init_mm.page_table_lock);
762 pgd_changed = true;
763 }
764
765 if (pgd_changed)
766 sync_global_pgds(vaddr_start, vaddr_end - 1);
767
768 return paddr_last;
769}
770
771
772/*
773 * Create page table mapping for the physical memory for specific physical
774 * addresses. Note that it can only be used to populate non-present entries.
775 * The virtual and physical addresses have to be aligned on PMD level
776 * down. It returns the last physical address mapped.
777 */
778unsigned long __meminit
779kernel_physical_mapping_init(unsigned long paddr_start,
780 unsigned long paddr_end,
781 unsigned long page_size_mask)
782{
783 return __kernel_physical_mapping_init(paddr_start, paddr_end,
784 page_size_mask, true);
785}
786
787/*
788 * This function is similar to kernel_physical_mapping_init() above with the
789 * exception that it uses set_{pud,pmd}() instead of the set_{pud,pte}_safe()
790 * when updating the mapping. The caller is responsible to flush the TLBs after
791 * the function returns.
792 */
793unsigned long __meminit
794kernel_physical_mapping_change(unsigned long paddr_start,
795 unsigned long paddr_end,
796 unsigned long page_size_mask)
797{
798 return __kernel_physical_mapping_init(paddr_start, paddr_end,
799 page_size_mask, false);
800}
801
802#ifndef CONFIG_NUMA
803void __init initmem_init(void)
804{
805 memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
806}
807#endif
808
809void __init paging_init(void)
810{
811 sparse_memory_present_with_active_regions(MAX_NUMNODES);
812 sparse_init();
813
814 /*
815 * clear the default setting with node 0
816 * note: don't use nodes_clear here, that is really clearing when
817 * numa support is not compiled in, and later node_set_state
818 * will not set it back.
819 */
820 node_clear_state(0, N_MEMORY);
821 if (N_MEMORY != N_NORMAL_MEMORY)
822 node_clear_state(0, N_NORMAL_MEMORY);
823
824 zone_sizes_init();
825}
826
827/*
828 * Memory hotplug specific functions
829 */
830#ifdef CONFIG_MEMORY_HOTPLUG
831/*
832 * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
833 * updating.
834 */
835static void update_end_of_memory_vars(u64 start, u64 size)
836{
837 unsigned long end_pfn = PFN_UP(start + size);
838
839 if (end_pfn > max_pfn) {
840 max_pfn = end_pfn;
841 max_low_pfn = end_pfn;
842 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
843 }
844}
845
846int add_pages(int nid, unsigned long start_pfn, unsigned long nr_pages,
847 struct mhp_restrictions *restrictions)
848{
849 int ret;
850
851 ret = __add_pages(nid, start_pfn, nr_pages, restrictions);
852 WARN_ON_ONCE(ret);
853
854 /* update max_pfn, max_low_pfn and high_memory */
855 update_end_of_memory_vars(start_pfn << PAGE_SHIFT,
856 nr_pages << PAGE_SHIFT);
857
858 return ret;
859}
860
861int arch_add_memory(int nid, u64 start, u64 size,
862 struct mhp_restrictions *restrictions)
863{
864 unsigned long start_pfn = start >> PAGE_SHIFT;
865 unsigned long nr_pages = size >> PAGE_SHIFT;
866
867 init_memory_mapping(start, start + size);
868
869 return add_pages(nid, start_pfn, nr_pages, restrictions);
870}
871
872#define PAGE_INUSE 0xFD
873
874static void __meminit free_pagetable(struct page *page, int order)
875{
876 unsigned long magic;
877 unsigned int nr_pages = 1 << order;
878
879 /* bootmem page has reserved flag */
880 if (PageReserved(page)) {
881 __ClearPageReserved(page);
882
883 magic = (unsigned long)page->freelist;
884 if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
885 while (nr_pages--)
886 put_page_bootmem(page++);
887 } else
888 while (nr_pages--)
889 free_reserved_page(page++);
890 } else
891 free_pages((unsigned long)page_address(page), order);
892}
893
894static void __meminit free_hugepage_table(struct page *page,
895 struct vmem_altmap *altmap)
896{
897 if (altmap)
898 vmem_altmap_free(altmap, PMD_SIZE / PAGE_SIZE);
899 else
900 free_pagetable(page, get_order(PMD_SIZE));
901}
902
903static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
904{
905 pte_t *pte;
906 int i;
907
908 for (i = 0; i < PTRS_PER_PTE; i++) {
909 pte = pte_start + i;
910 if (!pte_none(*pte))
911 return;
912 }
913
914 /* free a pte talbe */
915 free_pagetable(pmd_page(*pmd), 0);
916 spin_lock(&init_mm.page_table_lock);
917 pmd_clear(pmd);
918 spin_unlock(&init_mm.page_table_lock);
919}
920
921static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
922{
923 pmd_t *pmd;
924 int i;
925
926 for (i = 0; i < PTRS_PER_PMD; i++) {
927 pmd = pmd_start + i;
928 if (!pmd_none(*pmd))
929 return;
930 }
931
932 /* free a pmd talbe */
933 free_pagetable(pud_page(*pud), 0);
934 spin_lock(&init_mm.page_table_lock);
935 pud_clear(pud);
936 spin_unlock(&init_mm.page_table_lock);
937}
938
939static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
940{
941 pud_t *pud;
942 int i;
943
944 for (i = 0; i < PTRS_PER_PUD; i++) {
945 pud = pud_start + i;
946 if (!pud_none(*pud))
947 return;
948 }
949
950 /* free a pud talbe */
951 free_pagetable(p4d_page(*p4d), 0);
952 spin_lock(&init_mm.page_table_lock);
953 p4d_clear(p4d);
954 spin_unlock(&init_mm.page_table_lock);
955}
956
957static void __meminit
958remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
959 bool direct)
960{
961 unsigned long next, pages = 0;
962 pte_t *pte;
963 void *page_addr;
964 phys_addr_t phys_addr;
965
966 pte = pte_start + pte_index(addr);
967 for (; addr < end; addr = next, pte++) {
968 next = (addr + PAGE_SIZE) & PAGE_MASK;
969 if (next > end)
970 next = end;
971
972 if (!pte_present(*pte))
973 continue;
974
975 /*
976 * We mapped [0,1G) memory as identity mapping when
977 * initializing, in arch/x86/kernel/head_64.S. These
978 * pagetables cannot be removed.
979 */
980 phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
981 if (phys_addr < (phys_addr_t)0x40000000)
982 return;
983
984 if (PAGE_ALIGNED(addr) && PAGE_ALIGNED(next)) {
985 /*
986 * Do not free direct mapping pages since they were
987 * freed when offlining, or simplely not in use.
988 */
989 if (!direct)
990 free_pagetable(pte_page(*pte), 0);
991
992 spin_lock(&init_mm.page_table_lock);
993 pte_clear(&init_mm, addr, pte);
994 spin_unlock(&init_mm.page_table_lock);
995
996 /* For non-direct mapping, pages means nothing. */
997 pages++;
998 } else {
999 /*
1000 * If we are here, we are freeing vmemmap pages since
1001 * direct mapped memory ranges to be freed are aligned.
1002 *
1003 * If we are not removing the whole page, it means
1004 * other page structs in this page are being used and
1005 * we canot remove them. So fill the unused page_structs
1006 * with 0xFD, and remove the page when it is wholly
1007 * filled with 0xFD.
1008 */
1009 memset((void *)addr, PAGE_INUSE, next - addr);
1010
1011 page_addr = page_address(pte_page(*pte));
1012 if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) {
1013 free_pagetable(pte_page(*pte), 0);
1014
1015 spin_lock(&init_mm.page_table_lock);
1016 pte_clear(&init_mm, addr, pte);
1017 spin_unlock(&init_mm.page_table_lock);
1018 }
1019 }
1020 }
1021
1022 /* Call free_pte_table() in remove_pmd_table(). */
1023 flush_tlb_all();
1024 if (direct)
1025 update_page_count(PG_LEVEL_4K, -pages);
1026}
1027
1028static void __meminit
1029remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
1030 bool direct, struct vmem_altmap *altmap)
1031{
1032 unsigned long next, pages = 0;
1033 pte_t *pte_base;
1034 pmd_t *pmd;
1035 void *page_addr;
1036
1037 pmd = pmd_start + pmd_index(addr);
1038 for (; addr < end; addr = next, pmd++) {
1039 next = pmd_addr_end(addr, end);
1040
1041 if (!pmd_present(*pmd))
1042 continue;
1043
1044 if (pmd_large(*pmd)) {
1045 if (IS_ALIGNED(addr, PMD_SIZE) &&
1046 IS_ALIGNED(next, PMD_SIZE)) {
1047 if (!direct)
1048 free_hugepage_table(pmd_page(*pmd),
1049 altmap);
1050
1051 spin_lock(&init_mm.page_table_lock);
1052 pmd_clear(pmd);
1053 spin_unlock(&init_mm.page_table_lock);
1054 pages++;
1055 } else {
1056 /* If here, we are freeing vmemmap pages. */
1057 memset((void *)addr, PAGE_INUSE, next - addr);
1058
1059 page_addr = page_address(pmd_page(*pmd));
1060 if (!memchr_inv(page_addr, PAGE_INUSE,
1061 PMD_SIZE)) {
1062 free_hugepage_table(pmd_page(*pmd),
1063 altmap);
1064
1065 spin_lock(&init_mm.page_table_lock);
1066 pmd_clear(pmd);
1067 spin_unlock(&init_mm.page_table_lock);
1068 }
1069 }
1070
1071 continue;
1072 }
1073
1074 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
1075 remove_pte_table(pte_base, addr, next, direct);
1076 free_pte_table(pte_base, pmd);
1077 }
1078
1079 /* Call free_pmd_table() in remove_pud_table(). */
1080 if (direct)
1081 update_page_count(PG_LEVEL_2M, -pages);
1082}
1083
1084static void __meminit
1085remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
1086 struct vmem_altmap *altmap, bool direct)
1087{
1088 unsigned long next, pages = 0;
1089 pmd_t *pmd_base;
1090 pud_t *pud;
1091 void *page_addr;
1092
1093 pud = pud_start + pud_index(addr);
1094 for (; addr < end; addr = next, pud++) {
1095 next = pud_addr_end(addr, end);
1096
1097 if (!pud_present(*pud))
1098 continue;
1099
1100 if (pud_large(*pud)) {
1101 if (IS_ALIGNED(addr, PUD_SIZE) &&
1102 IS_ALIGNED(next, PUD_SIZE)) {
1103 if (!direct)
1104 free_pagetable(pud_page(*pud),
1105 get_order(PUD_SIZE));
1106
1107 spin_lock(&init_mm.page_table_lock);
1108 pud_clear(pud);
1109 spin_unlock(&init_mm.page_table_lock);
1110 pages++;
1111 } else {
1112 /* If here, we are freeing vmemmap pages. */
1113 memset((void *)addr, PAGE_INUSE, next - addr);
1114
1115 page_addr = page_address(pud_page(*pud));
1116 if (!memchr_inv(page_addr, PAGE_INUSE,
1117 PUD_SIZE)) {
1118 free_pagetable(pud_page(*pud),
1119 get_order(PUD_SIZE));
1120
1121 spin_lock(&init_mm.page_table_lock);
1122 pud_clear(pud);
1123 spin_unlock(&init_mm.page_table_lock);
1124 }
1125 }
1126
1127 continue;
1128 }
1129
1130 pmd_base = pmd_offset(pud, 0);
1131 remove_pmd_table(pmd_base, addr, next, direct, altmap);
1132 free_pmd_table(pmd_base, pud);
1133 }
1134
1135 if (direct)
1136 update_page_count(PG_LEVEL_1G, -pages);
1137}
1138
1139static void __meminit
1140remove_p4d_table(p4d_t *p4d_start, unsigned long addr, unsigned long end,
1141 struct vmem_altmap *altmap, bool direct)
1142{
1143 unsigned long next, pages = 0;
1144 pud_t *pud_base;
1145 p4d_t *p4d;
1146
1147 p4d = p4d_start + p4d_index(addr);
1148 for (; addr < end; addr = next, p4d++) {
1149 next = p4d_addr_end(addr, end);
1150
1151 if (!p4d_present(*p4d))
1152 continue;
1153
1154 BUILD_BUG_ON(p4d_large(*p4d));
1155
1156 pud_base = pud_offset(p4d, 0);
1157 remove_pud_table(pud_base, addr, next, altmap, direct);
1158 /*
1159 * For 4-level page tables we do not want to free PUDs, but in the
1160 * 5-level case we should free them. This code will have to change
1161 * to adapt for boot-time switching between 4 and 5 level page tables.
1162 */
1163 if (pgtable_l5_enabled())
1164 free_pud_table(pud_base, p4d);
1165 }
1166
1167 if (direct)
1168 update_page_count(PG_LEVEL_512G, -pages);
1169}
1170
1171/* start and end are both virtual address. */
1172static void __meminit
1173remove_pagetable(unsigned long start, unsigned long end, bool direct,
1174 struct vmem_altmap *altmap)
1175{
1176 unsigned long next;
1177 unsigned long addr;
1178 pgd_t *pgd;
1179 p4d_t *p4d;
1180
1181 for (addr = start; addr < end; addr = next) {
1182 next = pgd_addr_end(addr, end);
1183
1184 pgd = pgd_offset_k(addr);
1185 if (!pgd_present(*pgd))
1186 continue;
1187
1188 p4d = p4d_offset(pgd, 0);
1189 remove_p4d_table(p4d, addr, next, altmap, direct);
1190 }
1191
1192 flush_tlb_all();
1193}
1194
1195void __ref vmemmap_free(unsigned long start, unsigned long end,
1196 struct vmem_altmap *altmap)
1197{
1198 remove_pagetable(start, end, false, altmap);
1199}
1200
1201static void __meminit
1202kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1203{
1204 start = (unsigned long)__va(start);
1205 end = (unsigned long)__va(end);
1206
1207 remove_pagetable(start, end, true, NULL);
1208}
1209
1210void __ref arch_remove_memory(int nid, u64 start, u64 size,
1211 struct vmem_altmap *altmap)
1212{
1213 unsigned long start_pfn = start >> PAGE_SHIFT;
1214 unsigned long nr_pages = size >> PAGE_SHIFT;
1215 struct page *page = pfn_to_page(start_pfn) + vmem_altmap_offset(altmap);
1216 struct zone *zone = page_zone(page);
1217
1218 __remove_pages(zone, start_pfn, nr_pages, altmap);
1219 kernel_physical_mapping_remove(start, start + size);
1220}
1221#endif /* CONFIG_MEMORY_HOTPLUG */
1222
1223static struct kcore_list kcore_vsyscall;
1224
1225static void __init register_page_bootmem_info(void)
1226{
1227#ifdef CONFIG_NUMA
1228 int i;
1229
1230 for_each_online_node(i)
1231 register_page_bootmem_info_node(NODE_DATA(i));
1232#endif
1233}
1234
1235void __init mem_init(void)
1236{
1237 pci_iommu_alloc();
1238
1239 /* clear_bss() already clear the empty_zero_page */
1240
1241 /* this will put all memory onto the freelists */
1242 memblock_free_all();
1243 after_bootmem = 1;
1244 x86_init.hyper.init_after_bootmem();
1245
1246 /*
1247 * Must be done after boot memory is put on freelist, because here we
1248 * might set fields in deferred struct pages that have not yet been
1249 * initialized, and memblock_free_all() initializes all the reserved
1250 * deferred pages for us.
1251 */
1252 register_page_bootmem_info();
1253
1254 /* Register memory areas for /proc/kcore */
1255 if (get_gate_vma(&init_mm))
1256 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR, PAGE_SIZE, KCORE_USER);
1257
1258 mem_init_print_info(NULL);
1259}
1260
1261int kernel_set_to_readonly;
1262
1263void set_kernel_text_rw(void)
1264{
1265 unsigned long start = PFN_ALIGN(_text);
1266 unsigned long end = PFN_ALIGN(__stop___ex_table);
1267
1268 if (!kernel_set_to_readonly)
1269 return;
1270
1271 pr_debug("Set kernel text: %lx - %lx for read write\n",
1272 start, end);
1273
1274 /*
1275 * Make the kernel identity mapping for text RW. Kernel text
1276 * mapping will always be RO. Refer to the comment in
1277 * static_protections() in pageattr.c
1278 */
1279 set_memory_rw(start, (end - start) >> PAGE_SHIFT);
1280}
1281
1282void set_kernel_text_ro(void)
1283{
1284 unsigned long start = PFN_ALIGN(_text);
1285 unsigned long end = PFN_ALIGN(__stop___ex_table);
1286
1287 if (!kernel_set_to_readonly)
1288 return;
1289
1290 pr_debug("Set kernel text: %lx - %lx for read only\n",
1291 start, end);
1292
1293 /*
1294 * Set the kernel identity mapping for text RO.
1295 */
1296 set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1297}
1298
1299void mark_rodata_ro(void)
1300{
1301 unsigned long start = PFN_ALIGN(_text);
1302 unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1303 unsigned long end = (unsigned long) &__end_rodata_hpage_align;
1304 unsigned long text_end = PFN_ALIGN(&__stop___ex_table);
1305 unsigned long rodata_end = PFN_ALIGN(&__end_rodata);
1306 unsigned long all_end;
1307
1308 printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1309 (end - start) >> 10);
1310 set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1311
1312 kernel_set_to_readonly = 1;
1313
1314 /*
1315 * The rodata/data/bss/brk section (but not the kernel text!)
1316 * should also be not-executable.
1317 *
1318 * We align all_end to PMD_SIZE because the existing mapping
1319 * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1320 * split the PMD and the reminder between _brk_end and the end
1321 * of the PMD will remain mapped executable.
1322 *
1323 * Any PMD which was setup after the one which covers _brk_end
1324 * has been zapped already via cleanup_highmem().
1325 */
1326 all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1327 set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
1328
1329#ifdef CONFIG_CPA_DEBUG
1330 printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1331 set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1332
1333 printk(KERN_INFO "Testing CPA: again\n");
1334 set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1335#endif
1336
1337 free_kernel_image_pages((void *)text_end, (void *)rodata_start);
1338 free_kernel_image_pages((void *)rodata_end, (void *)_sdata);
1339
1340 debug_checkwx();
1341}
1342
1343int kern_addr_valid(unsigned long addr)
1344{
1345 unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1346 pgd_t *pgd;
1347 p4d_t *p4d;
1348 pud_t *pud;
1349 pmd_t *pmd;
1350 pte_t *pte;
1351
1352 if (above != 0 && above != -1UL)
1353 return 0;
1354
1355 pgd = pgd_offset_k(addr);
1356 if (pgd_none(*pgd))
1357 return 0;
1358
1359 p4d = p4d_offset(pgd, addr);
1360 if (p4d_none(*p4d))
1361 return 0;
1362
1363 pud = pud_offset(p4d, addr);
1364 if (pud_none(*pud))
1365 return 0;
1366
1367 if (pud_large(*pud))
1368 return pfn_valid(pud_pfn(*pud));
1369
1370 pmd = pmd_offset(pud, addr);
1371 if (pmd_none(*pmd))
1372 return 0;
1373
1374 if (pmd_large(*pmd))
1375 return pfn_valid(pmd_pfn(*pmd));
1376
1377 pte = pte_offset_kernel(pmd, addr);
1378 if (pte_none(*pte))
1379 return 0;
1380
1381 return pfn_valid(pte_pfn(*pte));
1382}
1383
1384/*
1385 * Block size is the minimum amount of memory which can be hotplugged or
1386 * hotremoved. It must be power of two and must be equal or larger than
1387 * MIN_MEMORY_BLOCK_SIZE.
1388 */
1389#define MAX_BLOCK_SIZE (2UL << 30)
1390
1391/* Amount of ram needed to start using large blocks */
1392#define MEM_SIZE_FOR_LARGE_BLOCK (64UL << 30)
1393
1394/* Adjustable memory block size */
1395static unsigned long set_memory_block_size;
1396int __init set_memory_block_size_order(unsigned int order)
1397{
1398 unsigned long size = 1UL << order;
1399
1400 if (size > MEM_SIZE_FOR_LARGE_BLOCK || size < MIN_MEMORY_BLOCK_SIZE)
1401 return -EINVAL;
1402
1403 set_memory_block_size = size;
1404 return 0;
1405}
1406
1407static unsigned long probe_memory_block_size(void)
1408{
1409 unsigned long boot_mem_end = max_pfn << PAGE_SHIFT;
1410 unsigned long bz;
1411
1412 /* If memory block size has been set, then use it */
1413 bz = set_memory_block_size;
1414 if (bz)
1415 goto done;
1416
1417 /* Use regular block if RAM is smaller than MEM_SIZE_FOR_LARGE_BLOCK */
1418 if (boot_mem_end < MEM_SIZE_FOR_LARGE_BLOCK) {
1419 bz = MIN_MEMORY_BLOCK_SIZE;
1420 goto done;
1421 }
1422
1423 /* Find the largest allowed block size that aligns to memory end */
1424 for (bz = MAX_BLOCK_SIZE; bz > MIN_MEMORY_BLOCK_SIZE; bz >>= 1) {
1425 if (IS_ALIGNED(boot_mem_end, bz))
1426 break;
1427 }
1428done:
1429 pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20);
1430
1431 return bz;
1432}
1433
1434static unsigned long memory_block_size_probed;
1435unsigned long memory_block_size_bytes(void)
1436{
1437 if (!memory_block_size_probed)
1438 memory_block_size_probed = probe_memory_block_size();
1439
1440 return memory_block_size_probed;
1441}
1442
1443#ifdef CONFIG_SPARSEMEM_VMEMMAP
1444/*
1445 * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1446 */
1447static long __meminitdata addr_start, addr_end;
1448static void __meminitdata *p_start, *p_end;
1449static int __meminitdata node_start;
1450
1451static int __meminit vmemmap_populate_hugepages(unsigned long start,
1452 unsigned long end, int node, struct vmem_altmap *altmap)
1453{
1454 unsigned long addr;
1455 unsigned long next;
1456 pgd_t *pgd;
1457 p4d_t *p4d;
1458 pud_t *pud;
1459 pmd_t *pmd;
1460
1461 for (addr = start; addr < end; addr = next) {
1462 next = pmd_addr_end(addr, end);
1463
1464 pgd = vmemmap_pgd_populate(addr, node);
1465 if (!pgd)
1466 return -ENOMEM;
1467
1468 p4d = vmemmap_p4d_populate(pgd, addr, node);
1469 if (!p4d)
1470 return -ENOMEM;
1471
1472 pud = vmemmap_pud_populate(p4d, addr, node);
1473 if (!pud)
1474 return -ENOMEM;
1475
1476 pmd = pmd_offset(pud, addr);
1477 if (pmd_none(*pmd)) {
1478 void *p;
1479
1480 if (altmap)
1481 p = altmap_alloc_block_buf(PMD_SIZE, altmap);
1482 else
1483 p = vmemmap_alloc_block_buf(PMD_SIZE, node);
1484 if (p) {
1485 pte_t entry;
1486
1487 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1488 PAGE_KERNEL_LARGE);
1489 set_pmd(pmd, __pmd(pte_val(entry)));
1490
1491 /* check to see if we have contiguous blocks */
1492 if (p_end != p || node_start != node) {
1493 if (p_start)
1494 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1495 addr_start, addr_end-1, p_start, p_end-1, node_start);
1496 addr_start = addr;
1497 node_start = node;
1498 p_start = p;
1499 }
1500
1501 addr_end = addr + PMD_SIZE;
1502 p_end = p + PMD_SIZE;
1503 continue;
1504 } else if (altmap)
1505 return -ENOMEM; /* no fallback */
1506 } else if (pmd_large(*pmd)) {
1507 vmemmap_verify((pte_t *)pmd, node, addr, next);
1508 continue;
1509 }
1510 if (vmemmap_populate_basepages(addr, next, node))
1511 return -ENOMEM;
1512 }
1513 return 0;
1514}
1515
1516int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1517 struct vmem_altmap *altmap)
1518{
1519 int err;
1520
1521 if (end - start < PAGES_PER_SECTION * sizeof(struct page))
1522 err = vmemmap_populate_basepages(start, end, node);
1523 else if (boot_cpu_has(X86_FEATURE_PSE))
1524 err = vmemmap_populate_hugepages(start, end, node, altmap);
1525 else if (altmap) {
1526 pr_err_once("%s: no cpu support for altmap allocations\n",
1527 __func__);
1528 err = -ENOMEM;
1529 } else
1530 err = vmemmap_populate_basepages(start, end, node);
1531 if (!err)
1532 sync_global_pgds(start, end - 1);
1533 return err;
1534}
1535
1536#if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1537void register_page_bootmem_memmap(unsigned long section_nr,
1538 struct page *start_page, unsigned long nr_pages)
1539{
1540 unsigned long addr = (unsigned long)start_page;
1541 unsigned long end = (unsigned long)(start_page + nr_pages);
1542 unsigned long next;
1543 pgd_t *pgd;
1544 p4d_t *p4d;
1545 pud_t *pud;
1546 pmd_t *pmd;
1547 unsigned int nr_pmd_pages;
1548 struct page *page;
1549
1550 for (; addr < end; addr = next) {
1551 pte_t *pte = NULL;
1552
1553 pgd = pgd_offset_k(addr);
1554 if (pgd_none(*pgd)) {
1555 next = (addr + PAGE_SIZE) & PAGE_MASK;
1556 continue;
1557 }
1558 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1559
1560 p4d = p4d_offset(pgd, addr);
1561 if (p4d_none(*p4d)) {
1562 next = (addr + PAGE_SIZE) & PAGE_MASK;
1563 continue;
1564 }
1565 get_page_bootmem(section_nr, p4d_page(*p4d), MIX_SECTION_INFO);
1566
1567 pud = pud_offset(p4d, addr);
1568 if (pud_none(*pud)) {
1569 next = (addr + PAGE_SIZE) & PAGE_MASK;
1570 continue;
1571 }
1572 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1573
1574 if (!boot_cpu_has(X86_FEATURE_PSE)) {
1575 next = (addr + PAGE_SIZE) & PAGE_MASK;
1576 pmd = pmd_offset(pud, addr);
1577 if (pmd_none(*pmd))
1578 continue;
1579 get_page_bootmem(section_nr, pmd_page(*pmd),
1580 MIX_SECTION_INFO);
1581
1582 pte = pte_offset_kernel(pmd, addr);
1583 if (pte_none(*pte))
1584 continue;
1585 get_page_bootmem(section_nr, pte_page(*pte),
1586 SECTION_INFO);
1587 } else {
1588 next = pmd_addr_end(addr, end);
1589
1590 pmd = pmd_offset(pud, addr);
1591 if (pmd_none(*pmd))
1592 continue;
1593
1594 nr_pmd_pages = 1 << get_order(PMD_SIZE);
1595 page = pmd_page(*pmd);
1596 while (nr_pmd_pages--)
1597 get_page_bootmem(section_nr, page++,
1598 SECTION_INFO);
1599 }
1600 }
1601}
1602#endif
1603
1604void __meminit vmemmap_populate_print_last(void)
1605{
1606 if (p_start) {
1607 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1608 addr_start, addr_end-1, p_start, p_end-1, node_start);
1609 p_start = NULL;
1610 p_end = NULL;
1611 node_start = 0;
1612 }
1613}
1614#endif
1/*
2 * linux/arch/x86_64/mm/init.c
3 *
4 * Copyright (C) 1995 Linus Torvalds
5 * Copyright (C) 2000 Pavel Machek <pavel@ucw.cz>
6 * Copyright (C) 2002,2003 Andi Kleen <ak@suse.de>
7 */
8
9#include <linux/signal.h>
10#include <linux/sched.h>
11#include <linux/kernel.h>
12#include <linux/errno.h>
13#include <linux/string.h>
14#include <linux/types.h>
15#include <linux/ptrace.h>
16#include <linux/mman.h>
17#include <linux/mm.h>
18#include <linux/swap.h>
19#include <linux/smp.h>
20#include <linux/init.h>
21#include <linux/initrd.h>
22#include <linux/pagemap.h>
23#include <linux/bootmem.h>
24#include <linux/memblock.h>
25#include <linux/proc_fs.h>
26#include <linux/pci.h>
27#include <linux/pfn.h>
28#include <linux/poison.h>
29#include <linux/dma-mapping.h>
30#include <linux/memory.h>
31#include <linux/memory_hotplug.h>
32#include <linux/memremap.h>
33#include <linux/nmi.h>
34#include <linux/gfp.h>
35#include <linux/kcore.h>
36
37#include <asm/processor.h>
38#include <asm/bios_ebda.h>
39#include <linux/uaccess.h>
40#include <asm/pgtable.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
58#include "mm_internal.h"
59
60#include "ident_map.c"
61
62/*
63 * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
64 * physical space so we can cache the place of the first one and move
65 * around without checking the pgd every time.
66 */
67
68/* Bits supported by the hardware: */
69pteval_t __supported_pte_mask __read_mostly = ~0;
70/* Bits allowed in normal kernel mappings: */
71pteval_t __default_kernel_pte_mask __read_mostly = ~0;
72EXPORT_SYMBOL_GPL(__supported_pte_mask);
73/* Used in PAGE_KERNEL_* macros which are reasonably used out-of-tree: */
74EXPORT_SYMBOL(__default_kernel_pte_mask);
75
76int force_personality32;
77
78/*
79 * noexec32=on|off
80 * Control non executable heap for 32bit processes.
81 * To control the stack too use noexec=off
82 *
83 * on PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
84 * off PROT_READ implies PROT_EXEC
85 */
86static int __init nonx32_setup(char *str)
87{
88 if (!strcmp(str, "on"))
89 force_personality32 &= ~READ_IMPLIES_EXEC;
90 else if (!strcmp(str, "off"))
91 force_personality32 |= READ_IMPLIES_EXEC;
92 return 1;
93}
94__setup("noexec32=", nonx32_setup);
95
96static void sync_global_pgds_l5(unsigned long start, unsigned long end)
97{
98 unsigned long addr;
99
100 for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
101 const pgd_t *pgd_ref = pgd_offset_k(addr);
102 struct page *page;
103
104 /* Check for overflow */
105 if (addr < start)
106 break;
107
108 if (pgd_none(*pgd_ref))
109 continue;
110
111 spin_lock(&pgd_lock);
112 list_for_each_entry(page, &pgd_list, lru) {
113 pgd_t *pgd;
114 spinlock_t *pgt_lock;
115
116 pgd = (pgd_t *)page_address(page) + pgd_index(addr);
117 /* the pgt_lock only for Xen */
118 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
119 spin_lock(pgt_lock);
120
121 if (!pgd_none(*pgd_ref) && !pgd_none(*pgd))
122 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
123
124 if (pgd_none(*pgd))
125 set_pgd(pgd, *pgd_ref);
126
127 spin_unlock(pgt_lock);
128 }
129 spin_unlock(&pgd_lock);
130 }
131}
132
133static void sync_global_pgds_l4(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 pgd_t *pgd_ref = pgd_offset_k(addr);
139 const p4d_t *p4d_ref;
140 struct page *page;
141
142 /*
143 * With folded p4d, pgd_none() is always false, we need to
144 * handle synchonization on p4d level.
145 */
146 MAYBE_BUILD_BUG_ON(pgd_none(*pgd_ref));
147 p4d_ref = p4d_offset(pgd_ref, addr);
148
149 if (p4d_none(*p4d_ref))
150 continue;
151
152 spin_lock(&pgd_lock);
153 list_for_each_entry(page, &pgd_list, lru) {
154 pgd_t *pgd;
155 p4d_t *p4d;
156 spinlock_t *pgt_lock;
157
158 pgd = (pgd_t *)page_address(page) + pgd_index(addr);
159 p4d = p4d_offset(pgd, addr);
160 /* the pgt_lock only for Xen */
161 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
162 spin_lock(pgt_lock);
163
164 if (!p4d_none(*p4d_ref) && !p4d_none(*p4d))
165 BUG_ON(p4d_page_vaddr(*p4d)
166 != p4d_page_vaddr(*p4d_ref));
167
168 if (p4d_none(*p4d))
169 set_p4d(p4d, *p4d_ref);
170
171 spin_unlock(pgt_lock);
172 }
173 spin_unlock(&pgd_lock);
174 }
175}
176
177/*
178 * When memory was added make sure all the processes MM have
179 * suitable PGD entries in the local PGD level page.
180 */
181void sync_global_pgds(unsigned long start, unsigned long end)
182{
183 if (pgtable_l5_enabled)
184 sync_global_pgds_l5(start, end);
185 else
186 sync_global_pgds_l4(start, end);
187}
188
189/*
190 * NOTE: This function is marked __ref because it calls __init function
191 * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
192 */
193static __ref void *spp_getpage(void)
194{
195 void *ptr;
196
197 if (after_bootmem)
198 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
199 else
200 ptr = alloc_bootmem_pages(PAGE_SIZE);
201
202 if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
203 panic("set_pte_phys: cannot allocate page data %s\n",
204 after_bootmem ? "after bootmem" : "");
205 }
206
207 pr_debug("spp_getpage %p\n", ptr);
208
209 return ptr;
210}
211
212static p4d_t *fill_p4d(pgd_t *pgd, unsigned long vaddr)
213{
214 if (pgd_none(*pgd)) {
215 p4d_t *p4d = (p4d_t *)spp_getpage();
216 pgd_populate(&init_mm, pgd, p4d);
217 if (p4d != p4d_offset(pgd, 0))
218 printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
219 p4d, p4d_offset(pgd, 0));
220 }
221 return p4d_offset(pgd, vaddr);
222}
223
224static pud_t *fill_pud(p4d_t *p4d, unsigned long vaddr)
225{
226 if (p4d_none(*p4d)) {
227 pud_t *pud = (pud_t *)spp_getpage();
228 p4d_populate(&init_mm, p4d, pud);
229 if (pud != pud_offset(p4d, 0))
230 printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
231 pud, pud_offset(p4d, 0));
232 }
233 return pud_offset(p4d, vaddr);
234}
235
236static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
237{
238 if (pud_none(*pud)) {
239 pmd_t *pmd = (pmd_t *) spp_getpage();
240 pud_populate(&init_mm, pud, pmd);
241 if (pmd != pmd_offset(pud, 0))
242 printk(KERN_ERR "PAGETABLE BUG #02! %p <-> %p\n",
243 pmd, pmd_offset(pud, 0));
244 }
245 return pmd_offset(pud, vaddr);
246}
247
248static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
249{
250 if (pmd_none(*pmd)) {
251 pte_t *pte = (pte_t *) spp_getpage();
252 pmd_populate_kernel(&init_mm, pmd, pte);
253 if (pte != pte_offset_kernel(pmd, 0))
254 printk(KERN_ERR "PAGETABLE BUG #03!\n");
255 }
256 return pte_offset_kernel(pmd, vaddr);
257}
258
259static void __set_pte_vaddr(pud_t *pud, unsigned long vaddr, pte_t new_pte)
260{
261 pmd_t *pmd = fill_pmd(pud, vaddr);
262 pte_t *pte = fill_pte(pmd, vaddr);
263
264 set_pte(pte, new_pte);
265
266 /*
267 * It's enough to flush this one mapping.
268 * (PGE mappings get flushed as well)
269 */
270 __flush_tlb_one_kernel(vaddr);
271}
272
273void set_pte_vaddr_p4d(p4d_t *p4d_page, unsigned long vaddr, pte_t new_pte)
274{
275 p4d_t *p4d = p4d_page + p4d_index(vaddr);
276 pud_t *pud = fill_pud(p4d, vaddr);
277
278 __set_pte_vaddr(pud, vaddr, new_pte);
279}
280
281void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
282{
283 pud_t *pud = pud_page + pud_index(vaddr);
284
285 __set_pte_vaddr(pud, vaddr, new_pte);
286}
287
288void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
289{
290 pgd_t *pgd;
291 p4d_t *p4d_page;
292
293 pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
294
295 pgd = pgd_offset_k(vaddr);
296 if (pgd_none(*pgd)) {
297 printk(KERN_ERR
298 "PGD FIXMAP MISSING, it should be setup in head.S!\n");
299 return;
300 }
301
302 p4d_page = p4d_offset(pgd, 0);
303 set_pte_vaddr_p4d(p4d_page, vaddr, pteval);
304}
305
306pmd_t * __init populate_extra_pmd(unsigned long vaddr)
307{
308 pgd_t *pgd;
309 p4d_t *p4d;
310 pud_t *pud;
311
312 pgd = pgd_offset_k(vaddr);
313 p4d = fill_p4d(pgd, vaddr);
314 pud = fill_pud(p4d, vaddr);
315 return fill_pmd(pud, vaddr);
316}
317
318pte_t * __init populate_extra_pte(unsigned long vaddr)
319{
320 pmd_t *pmd;
321
322 pmd = populate_extra_pmd(vaddr);
323 return fill_pte(pmd, vaddr);
324}
325
326/*
327 * Create large page table mappings for a range of physical addresses.
328 */
329static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
330 enum page_cache_mode cache)
331{
332 pgd_t *pgd;
333 p4d_t *p4d;
334 pud_t *pud;
335 pmd_t *pmd;
336 pgprot_t prot;
337
338 pgprot_val(prot) = pgprot_val(PAGE_KERNEL_LARGE) |
339 pgprot_val(pgprot_4k_2_large(cachemode2pgprot(cache)));
340 BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
341 for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
342 pgd = pgd_offset_k((unsigned long)__va(phys));
343 if (pgd_none(*pgd)) {
344 p4d = (p4d_t *) spp_getpage();
345 set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE |
346 _PAGE_USER));
347 }
348 p4d = p4d_offset(pgd, (unsigned long)__va(phys));
349 if (p4d_none(*p4d)) {
350 pud = (pud_t *) spp_getpage();
351 set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE |
352 _PAGE_USER));
353 }
354 pud = pud_offset(p4d, (unsigned long)__va(phys));
355 if (pud_none(*pud)) {
356 pmd = (pmd_t *) spp_getpage();
357 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
358 _PAGE_USER));
359 }
360 pmd = pmd_offset(pud, phys);
361 BUG_ON(!pmd_none(*pmd));
362 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
363 }
364}
365
366void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
367{
368 __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_WB);
369}
370
371void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
372{
373 __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_UC);
374}
375
376/*
377 * The head.S code sets up the kernel high mapping:
378 *
379 * from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
380 *
381 * phys_base holds the negative offset to the kernel, which is added
382 * to the compile time generated pmds. This results in invalid pmds up
383 * to the point where we hit the physaddr 0 mapping.
384 *
385 * We limit the mappings to the region from _text to _brk_end. _brk_end
386 * is rounded up to the 2MB boundary. This catches the invalid pmds as
387 * well, as they are located before _text:
388 */
389void __init cleanup_highmap(void)
390{
391 unsigned long vaddr = __START_KERNEL_map;
392 unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
393 unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
394 pmd_t *pmd = level2_kernel_pgt;
395
396 /*
397 * Native path, max_pfn_mapped is not set yet.
398 * Xen has valid max_pfn_mapped set in
399 * arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
400 */
401 if (max_pfn_mapped)
402 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
403
404 for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
405 if (pmd_none(*pmd))
406 continue;
407 if (vaddr < (unsigned long) _text || vaddr > end)
408 set_pmd(pmd, __pmd(0));
409 }
410}
411
412/*
413 * Create PTE level page table mapping for physical addresses.
414 * It returns the last physical address mapped.
415 */
416static unsigned long __meminit
417phys_pte_init(pte_t *pte_page, unsigned long paddr, unsigned long paddr_end,
418 pgprot_t prot)
419{
420 unsigned long pages = 0, paddr_next;
421 unsigned long paddr_last = paddr_end;
422 pte_t *pte;
423 int i;
424
425 pte = pte_page + pte_index(paddr);
426 i = pte_index(paddr);
427
428 for (; i < PTRS_PER_PTE; i++, paddr = paddr_next, pte++) {
429 paddr_next = (paddr & PAGE_MASK) + PAGE_SIZE;
430 if (paddr >= paddr_end) {
431 if (!after_bootmem &&
432 !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
433 E820_TYPE_RAM) &&
434 !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
435 E820_TYPE_RESERVED_KERN))
436 set_pte(pte, __pte(0));
437 continue;
438 }
439
440 /*
441 * We will re-use the existing mapping.
442 * Xen for example has some special requirements, like mapping
443 * pagetable pages as RO. So assume someone who pre-setup
444 * these mappings are more intelligent.
445 */
446 if (!pte_none(*pte)) {
447 if (!after_bootmem)
448 pages++;
449 continue;
450 }
451
452 if (0)
453 pr_info(" pte=%p addr=%lx pte=%016lx\n", pte, paddr,
454 pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL).pte);
455 pages++;
456 set_pte(pte, pfn_pte(paddr >> PAGE_SHIFT, prot));
457 paddr_last = (paddr & PAGE_MASK) + PAGE_SIZE;
458 }
459
460 update_page_count(PG_LEVEL_4K, pages);
461
462 return paddr_last;
463}
464
465/*
466 * Create PMD level page table mapping for physical addresses. The virtual
467 * and physical address have to be aligned at this level.
468 * It returns the last physical address mapped.
469 */
470static unsigned long __meminit
471phys_pmd_init(pmd_t *pmd_page, unsigned long paddr, unsigned long paddr_end,
472 unsigned long page_size_mask, pgprot_t prot)
473{
474 unsigned long pages = 0, paddr_next;
475 unsigned long paddr_last = paddr_end;
476
477 int i = pmd_index(paddr);
478
479 for (; i < PTRS_PER_PMD; i++, paddr = paddr_next) {
480 pmd_t *pmd = pmd_page + pmd_index(paddr);
481 pte_t *pte;
482 pgprot_t new_prot = prot;
483
484 paddr_next = (paddr & PMD_MASK) + PMD_SIZE;
485 if (paddr >= paddr_end) {
486 if (!after_bootmem &&
487 !e820__mapped_any(paddr & PMD_MASK, paddr_next,
488 E820_TYPE_RAM) &&
489 !e820__mapped_any(paddr & PMD_MASK, paddr_next,
490 E820_TYPE_RESERVED_KERN))
491 set_pmd(pmd, __pmd(0));
492 continue;
493 }
494
495 if (!pmd_none(*pmd)) {
496 if (!pmd_large(*pmd)) {
497 spin_lock(&init_mm.page_table_lock);
498 pte = (pte_t *)pmd_page_vaddr(*pmd);
499 paddr_last = phys_pte_init(pte, paddr,
500 paddr_end, prot);
501 spin_unlock(&init_mm.page_table_lock);
502 continue;
503 }
504 /*
505 * If we are ok with PG_LEVEL_2M mapping, then we will
506 * use the existing mapping,
507 *
508 * Otherwise, we will split the large page mapping but
509 * use the same existing protection bits except for
510 * large page, so that we don't violate Intel's TLB
511 * Application note (317080) which says, while changing
512 * the page sizes, new and old translations should
513 * not differ with respect to page frame and
514 * attributes.
515 */
516 if (page_size_mask & (1 << PG_LEVEL_2M)) {
517 if (!after_bootmem)
518 pages++;
519 paddr_last = paddr_next;
520 continue;
521 }
522 new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
523 }
524
525 if (page_size_mask & (1<<PG_LEVEL_2M)) {
526 pages++;
527 spin_lock(&init_mm.page_table_lock);
528 set_pte((pte_t *)pmd,
529 pfn_pte((paddr & PMD_MASK) >> PAGE_SHIFT,
530 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
531 spin_unlock(&init_mm.page_table_lock);
532 paddr_last = paddr_next;
533 continue;
534 }
535
536 pte = alloc_low_page();
537 paddr_last = phys_pte_init(pte, paddr, paddr_end, new_prot);
538
539 spin_lock(&init_mm.page_table_lock);
540 pmd_populate_kernel(&init_mm, pmd, pte);
541 spin_unlock(&init_mm.page_table_lock);
542 }
543 update_page_count(PG_LEVEL_2M, pages);
544 return paddr_last;
545}
546
547/*
548 * Create PUD level page table mapping for physical addresses. The virtual
549 * and physical address do not have to be aligned at this level. KASLR can
550 * randomize virtual addresses up to this level.
551 * It returns the last physical address mapped.
552 */
553static unsigned long __meminit
554phys_pud_init(pud_t *pud_page, unsigned long paddr, unsigned long paddr_end,
555 unsigned long page_size_mask)
556{
557 unsigned long pages = 0, paddr_next;
558 unsigned long paddr_last = paddr_end;
559 unsigned long vaddr = (unsigned long)__va(paddr);
560 int i = pud_index(vaddr);
561
562 for (; i < PTRS_PER_PUD; i++, paddr = paddr_next) {
563 pud_t *pud;
564 pmd_t *pmd;
565 pgprot_t prot = PAGE_KERNEL;
566
567 vaddr = (unsigned long)__va(paddr);
568 pud = pud_page + pud_index(vaddr);
569 paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
570
571 if (paddr >= paddr_end) {
572 if (!after_bootmem &&
573 !e820__mapped_any(paddr & PUD_MASK, paddr_next,
574 E820_TYPE_RAM) &&
575 !e820__mapped_any(paddr & PUD_MASK, paddr_next,
576 E820_TYPE_RESERVED_KERN))
577 set_pud(pud, __pud(0));
578 continue;
579 }
580
581 if (!pud_none(*pud)) {
582 if (!pud_large(*pud)) {
583 pmd = pmd_offset(pud, 0);
584 paddr_last = phys_pmd_init(pmd, paddr,
585 paddr_end,
586 page_size_mask,
587 prot);
588 __flush_tlb_all();
589 continue;
590 }
591 /*
592 * If we are ok with PG_LEVEL_1G mapping, then we will
593 * use the existing mapping.
594 *
595 * Otherwise, we will split the gbpage mapping but use
596 * the same existing protection bits except for large
597 * page, so that we don't violate Intel's TLB
598 * Application note (317080) which says, while changing
599 * the page sizes, new and old translations should
600 * not differ with respect to page frame and
601 * attributes.
602 */
603 if (page_size_mask & (1 << PG_LEVEL_1G)) {
604 if (!after_bootmem)
605 pages++;
606 paddr_last = paddr_next;
607 continue;
608 }
609 prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
610 }
611
612 if (page_size_mask & (1<<PG_LEVEL_1G)) {
613 pages++;
614 spin_lock(&init_mm.page_table_lock);
615 set_pte((pte_t *)pud,
616 pfn_pte((paddr & PUD_MASK) >> PAGE_SHIFT,
617 PAGE_KERNEL_LARGE));
618 spin_unlock(&init_mm.page_table_lock);
619 paddr_last = paddr_next;
620 continue;
621 }
622
623 pmd = alloc_low_page();
624 paddr_last = phys_pmd_init(pmd, paddr, paddr_end,
625 page_size_mask, prot);
626
627 spin_lock(&init_mm.page_table_lock);
628 pud_populate(&init_mm, pud, pmd);
629 spin_unlock(&init_mm.page_table_lock);
630 }
631 __flush_tlb_all();
632
633 update_page_count(PG_LEVEL_1G, pages);
634
635 return paddr_last;
636}
637
638static unsigned long __meminit
639phys_p4d_init(p4d_t *p4d_page, unsigned long paddr, unsigned long paddr_end,
640 unsigned long page_size_mask)
641{
642 unsigned long paddr_next, paddr_last = paddr_end;
643 unsigned long vaddr = (unsigned long)__va(paddr);
644 int i = p4d_index(vaddr);
645
646 if (!pgtable_l5_enabled)
647 return phys_pud_init((pud_t *) p4d_page, paddr, paddr_end, page_size_mask);
648
649 for (; i < PTRS_PER_P4D; i++, paddr = paddr_next) {
650 p4d_t *p4d;
651 pud_t *pud;
652
653 vaddr = (unsigned long)__va(paddr);
654 p4d = p4d_page + p4d_index(vaddr);
655 paddr_next = (paddr & P4D_MASK) + P4D_SIZE;
656
657 if (paddr >= paddr_end) {
658 if (!after_bootmem &&
659 !e820__mapped_any(paddr & P4D_MASK, paddr_next,
660 E820_TYPE_RAM) &&
661 !e820__mapped_any(paddr & P4D_MASK, paddr_next,
662 E820_TYPE_RESERVED_KERN))
663 set_p4d(p4d, __p4d(0));
664 continue;
665 }
666
667 if (!p4d_none(*p4d)) {
668 pud = pud_offset(p4d, 0);
669 paddr_last = phys_pud_init(pud, paddr,
670 paddr_end,
671 page_size_mask);
672 __flush_tlb_all();
673 continue;
674 }
675
676 pud = alloc_low_page();
677 paddr_last = phys_pud_init(pud, paddr, paddr_end,
678 page_size_mask);
679
680 spin_lock(&init_mm.page_table_lock);
681 p4d_populate(&init_mm, p4d, pud);
682 spin_unlock(&init_mm.page_table_lock);
683 }
684 __flush_tlb_all();
685
686 return paddr_last;
687}
688
689/*
690 * Create page table mapping for the physical memory for specific physical
691 * addresses. The virtual and physical addresses have to be aligned on PMD level
692 * down. It returns the last physical address mapped.
693 */
694unsigned long __meminit
695kernel_physical_mapping_init(unsigned long paddr_start,
696 unsigned long paddr_end,
697 unsigned long page_size_mask)
698{
699 bool pgd_changed = false;
700 unsigned long vaddr, vaddr_start, vaddr_end, vaddr_next, paddr_last;
701
702 paddr_last = paddr_end;
703 vaddr = (unsigned long)__va(paddr_start);
704 vaddr_end = (unsigned long)__va(paddr_end);
705 vaddr_start = vaddr;
706
707 for (; vaddr < vaddr_end; vaddr = vaddr_next) {
708 pgd_t *pgd = pgd_offset_k(vaddr);
709 p4d_t *p4d;
710
711 vaddr_next = (vaddr & PGDIR_MASK) + PGDIR_SIZE;
712
713 if (pgd_val(*pgd)) {
714 p4d = (p4d_t *)pgd_page_vaddr(*pgd);
715 paddr_last = phys_p4d_init(p4d, __pa(vaddr),
716 __pa(vaddr_end),
717 page_size_mask);
718 continue;
719 }
720
721 p4d = alloc_low_page();
722 paddr_last = phys_p4d_init(p4d, __pa(vaddr), __pa(vaddr_end),
723 page_size_mask);
724
725 spin_lock(&init_mm.page_table_lock);
726 if (pgtable_l5_enabled)
727 pgd_populate(&init_mm, pgd, p4d);
728 else
729 p4d_populate(&init_mm, p4d_offset(pgd, vaddr), (pud_t *) p4d);
730 spin_unlock(&init_mm.page_table_lock);
731 pgd_changed = true;
732 }
733
734 if (pgd_changed)
735 sync_global_pgds(vaddr_start, vaddr_end - 1);
736
737 __flush_tlb_all();
738
739 return paddr_last;
740}
741
742#ifndef CONFIG_NUMA
743void __init initmem_init(void)
744{
745 memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
746}
747#endif
748
749void __init paging_init(void)
750{
751 sparse_memory_present_with_active_regions(MAX_NUMNODES);
752 sparse_init();
753
754 /*
755 * clear the default setting with node 0
756 * note: don't use nodes_clear here, that is really clearing when
757 * numa support is not compiled in, and later node_set_state
758 * will not set it back.
759 */
760 node_clear_state(0, N_MEMORY);
761 if (N_MEMORY != N_NORMAL_MEMORY)
762 node_clear_state(0, N_NORMAL_MEMORY);
763
764 zone_sizes_init();
765}
766
767/*
768 * Memory hotplug specific functions
769 */
770#ifdef CONFIG_MEMORY_HOTPLUG
771/*
772 * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
773 * updating.
774 */
775static void update_end_of_memory_vars(u64 start, u64 size)
776{
777 unsigned long end_pfn = PFN_UP(start + size);
778
779 if (end_pfn > max_pfn) {
780 max_pfn = end_pfn;
781 max_low_pfn = end_pfn;
782 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
783 }
784}
785
786int add_pages(int nid, unsigned long start_pfn, unsigned long nr_pages,
787 struct vmem_altmap *altmap, bool want_memblock)
788{
789 int ret;
790
791 ret = __add_pages(nid, start_pfn, nr_pages, altmap, want_memblock);
792 WARN_ON_ONCE(ret);
793
794 /* update max_pfn, max_low_pfn and high_memory */
795 update_end_of_memory_vars(start_pfn << PAGE_SHIFT,
796 nr_pages << PAGE_SHIFT);
797
798 return ret;
799}
800
801int arch_add_memory(int nid, u64 start, u64 size, struct vmem_altmap *altmap,
802 bool want_memblock)
803{
804 unsigned long start_pfn = start >> PAGE_SHIFT;
805 unsigned long nr_pages = size >> PAGE_SHIFT;
806
807 init_memory_mapping(start, start + size);
808
809 return add_pages(nid, start_pfn, nr_pages, altmap, want_memblock);
810}
811
812#define PAGE_INUSE 0xFD
813
814static void __meminit free_pagetable(struct page *page, int order)
815{
816 unsigned long magic;
817 unsigned int nr_pages = 1 << order;
818
819 /* bootmem page has reserved flag */
820 if (PageReserved(page)) {
821 __ClearPageReserved(page);
822
823 magic = (unsigned long)page->freelist;
824 if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
825 while (nr_pages--)
826 put_page_bootmem(page++);
827 } else
828 while (nr_pages--)
829 free_reserved_page(page++);
830 } else
831 free_pages((unsigned long)page_address(page), order);
832}
833
834static void __meminit free_hugepage_table(struct page *page,
835 struct vmem_altmap *altmap)
836{
837 if (altmap)
838 vmem_altmap_free(altmap, PMD_SIZE / PAGE_SIZE);
839 else
840 free_pagetable(page, get_order(PMD_SIZE));
841}
842
843static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
844{
845 pte_t *pte;
846 int i;
847
848 for (i = 0; i < PTRS_PER_PTE; i++) {
849 pte = pte_start + i;
850 if (!pte_none(*pte))
851 return;
852 }
853
854 /* free a pte talbe */
855 free_pagetable(pmd_page(*pmd), 0);
856 spin_lock(&init_mm.page_table_lock);
857 pmd_clear(pmd);
858 spin_unlock(&init_mm.page_table_lock);
859}
860
861static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
862{
863 pmd_t *pmd;
864 int i;
865
866 for (i = 0; i < PTRS_PER_PMD; i++) {
867 pmd = pmd_start + i;
868 if (!pmd_none(*pmd))
869 return;
870 }
871
872 /* free a pmd talbe */
873 free_pagetable(pud_page(*pud), 0);
874 spin_lock(&init_mm.page_table_lock);
875 pud_clear(pud);
876 spin_unlock(&init_mm.page_table_lock);
877}
878
879static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
880{
881 pud_t *pud;
882 int i;
883
884 for (i = 0; i < PTRS_PER_PUD; i++) {
885 pud = pud_start + i;
886 if (!pud_none(*pud))
887 return;
888 }
889
890 /* free a pud talbe */
891 free_pagetable(p4d_page(*p4d), 0);
892 spin_lock(&init_mm.page_table_lock);
893 p4d_clear(p4d);
894 spin_unlock(&init_mm.page_table_lock);
895}
896
897static void __meminit
898remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
899 bool direct)
900{
901 unsigned long next, pages = 0;
902 pte_t *pte;
903 void *page_addr;
904 phys_addr_t phys_addr;
905
906 pte = pte_start + pte_index(addr);
907 for (; addr < end; addr = next, pte++) {
908 next = (addr + PAGE_SIZE) & PAGE_MASK;
909 if (next > end)
910 next = end;
911
912 if (!pte_present(*pte))
913 continue;
914
915 /*
916 * We mapped [0,1G) memory as identity mapping when
917 * initializing, in arch/x86/kernel/head_64.S. These
918 * pagetables cannot be removed.
919 */
920 phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
921 if (phys_addr < (phys_addr_t)0x40000000)
922 return;
923
924 if (PAGE_ALIGNED(addr) && PAGE_ALIGNED(next)) {
925 /*
926 * Do not free direct mapping pages since they were
927 * freed when offlining, or simplely not in use.
928 */
929 if (!direct)
930 free_pagetable(pte_page(*pte), 0);
931
932 spin_lock(&init_mm.page_table_lock);
933 pte_clear(&init_mm, addr, pte);
934 spin_unlock(&init_mm.page_table_lock);
935
936 /* For non-direct mapping, pages means nothing. */
937 pages++;
938 } else {
939 /*
940 * If we are here, we are freeing vmemmap pages since
941 * direct mapped memory ranges to be freed are aligned.
942 *
943 * If we are not removing the whole page, it means
944 * other page structs in this page are being used and
945 * we canot remove them. So fill the unused page_structs
946 * with 0xFD, and remove the page when it is wholly
947 * filled with 0xFD.
948 */
949 memset((void *)addr, PAGE_INUSE, next - addr);
950
951 page_addr = page_address(pte_page(*pte));
952 if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) {
953 free_pagetable(pte_page(*pte), 0);
954
955 spin_lock(&init_mm.page_table_lock);
956 pte_clear(&init_mm, addr, pte);
957 spin_unlock(&init_mm.page_table_lock);
958 }
959 }
960 }
961
962 /* Call free_pte_table() in remove_pmd_table(). */
963 flush_tlb_all();
964 if (direct)
965 update_page_count(PG_LEVEL_4K, -pages);
966}
967
968static void __meminit
969remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
970 bool direct, struct vmem_altmap *altmap)
971{
972 unsigned long next, pages = 0;
973 pte_t *pte_base;
974 pmd_t *pmd;
975 void *page_addr;
976
977 pmd = pmd_start + pmd_index(addr);
978 for (; addr < end; addr = next, pmd++) {
979 next = pmd_addr_end(addr, end);
980
981 if (!pmd_present(*pmd))
982 continue;
983
984 if (pmd_large(*pmd)) {
985 if (IS_ALIGNED(addr, PMD_SIZE) &&
986 IS_ALIGNED(next, PMD_SIZE)) {
987 if (!direct)
988 free_hugepage_table(pmd_page(*pmd),
989 altmap);
990
991 spin_lock(&init_mm.page_table_lock);
992 pmd_clear(pmd);
993 spin_unlock(&init_mm.page_table_lock);
994 pages++;
995 } else {
996 /* If here, we are freeing vmemmap pages. */
997 memset((void *)addr, PAGE_INUSE, next - addr);
998
999 page_addr = page_address(pmd_page(*pmd));
1000 if (!memchr_inv(page_addr, PAGE_INUSE,
1001 PMD_SIZE)) {
1002 free_hugepage_table(pmd_page(*pmd),
1003 altmap);
1004
1005 spin_lock(&init_mm.page_table_lock);
1006 pmd_clear(pmd);
1007 spin_unlock(&init_mm.page_table_lock);
1008 }
1009 }
1010
1011 continue;
1012 }
1013
1014 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
1015 remove_pte_table(pte_base, addr, next, direct);
1016 free_pte_table(pte_base, pmd);
1017 }
1018
1019 /* Call free_pmd_table() in remove_pud_table(). */
1020 if (direct)
1021 update_page_count(PG_LEVEL_2M, -pages);
1022}
1023
1024static void __meminit
1025remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
1026 struct vmem_altmap *altmap, bool direct)
1027{
1028 unsigned long next, pages = 0;
1029 pmd_t *pmd_base;
1030 pud_t *pud;
1031 void *page_addr;
1032
1033 pud = pud_start + pud_index(addr);
1034 for (; addr < end; addr = next, pud++) {
1035 next = pud_addr_end(addr, end);
1036
1037 if (!pud_present(*pud))
1038 continue;
1039
1040 if (pud_large(*pud)) {
1041 if (IS_ALIGNED(addr, PUD_SIZE) &&
1042 IS_ALIGNED(next, PUD_SIZE)) {
1043 if (!direct)
1044 free_pagetable(pud_page(*pud),
1045 get_order(PUD_SIZE));
1046
1047 spin_lock(&init_mm.page_table_lock);
1048 pud_clear(pud);
1049 spin_unlock(&init_mm.page_table_lock);
1050 pages++;
1051 } else {
1052 /* If here, we are freeing vmemmap pages. */
1053 memset((void *)addr, PAGE_INUSE, next - addr);
1054
1055 page_addr = page_address(pud_page(*pud));
1056 if (!memchr_inv(page_addr, PAGE_INUSE,
1057 PUD_SIZE)) {
1058 free_pagetable(pud_page(*pud),
1059 get_order(PUD_SIZE));
1060
1061 spin_lock(&init_mm.page_table_lock);
1062 pud_clear(pud);
1063 spin_unlock(&init_mm.page_table_lock);
1064 }
1065 }
1066
1067 continue;
1068 }
1069
1070 pmd_base = pmd_offset(pud, 0);
1071 remove_pmd_table(pmd_base, addr, next, direct, altmap);
1072 free_pmd_table(pmd_base, pud);
1073 }
1074
1075 if (direct)
1076 update_page_count(PG_LEVEL_1G, -pages);
1077}
1078
1079static void __meminit
1080remove_p4d_table(p4d_t *p4d_start, unsigned long addr, unsigned long end,
1081 struct vmem_altmap *altmap, bool direct)
1082{
1083 unsigned long next, pages = 0;
1084 pud_t *pud_base;
1085 p4d_t *p4d;
1086
1087 p4d = p4d_start + p4d_index(addr);
1088 for (; addr < end; addr = next, p4d++) {
1089 next = p4d_addr_end(addr, end);
1090
1091 if (!p4d_present(*p4d))
1092 continue;
1093
1094 BUILD_BUG_ON(p4d_large(*p4d));
1095
1096 pud_base = pud_offset(p4d, 0);
1097 remove_pud_table(pud_base, addr, next, altmap, direct);
1098 /*
1099 * For 4-level page tables we do not want to free PUDs, but in the
1100 * 5-level case we should free them. This code will have to change
1101 * to adapt for boot-time switching between 4 and 5 level page tables.
1102 */
1103 if (pgtable_l5_enabled)
1104 free_pud_table(pud_base, p4d);
1105 }
1106
1107 if (direct)
1108 update_page_count(PG_LEVEL_512G, -pages);
1109}
1110
1111/* start and end are both virtual address. */
1112static void __meminit
1113remove_pagetable(unsigned long start, unsigned long end, bool direct,
1114 struct vmem_altmap *altmap)
1115{
1116 unsigned long next;
1117 unsigned long addr;
1118 pgd_t *pgd;
1119 p4d_t *p4d;
1120
1121 for (addr = start; addr < end; addr = next) {
1122 next = pgd_addr_end(addr, end);
1123
1124 pgd = pgd_offset_k(addr);
1125 if (!pgd_present(*pgd))
1126 continue;
1127
1128 p4d = p4d_offset(pgd, 0);
1129 remove_p4d_table(p4d, addr, next, altmap, direct);
1130 }
1131
1132 flush_tlb_all();
1133}
1134
1135void __ref vmemmap_free(unsigned long start, unsigned long end,
1136 struct vmem_altmap *altmap)
1137{
1138 remove_pagetable(start, end, false, altmap);
1139}
1140
1141#ifdef CONFIG_MEMORY_HOTREMOVE
1142static void __meminit
1143kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1144{
1145 start = (unsigned long)__va(start);
1146 end = (unsigned long)__va(end);
1147
1148 remove_pagetable(start, end, true, NULL);
1149}
1150
1151int __ref arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap)
1152{
1153 unsigned long start_pfn = start >> PAGE_SHIFT;
1154 unsigned long nr_pages = size >> PAGE_SHIFT;
1155 struct page *page = pfn_to_page(start_pfn);
1156 struct zone *zone;
1157 int ret;
1158
1159 /* With altmap the first mapped page is offset from @start */
1160 if (altmap)
1161 page += vmem_altmap_offset(altmap);
1162 zone = page_zone(page);
1163 ret = __remove_pages(zone, start_pfn, nr_pages, altmap);
1164 WARN_ON_ONCE(ret);
1165 kernel_physical_mapping_remove(start, start + size);
1166
1167 return ret;
1168}
1169#endif
1170#endif /* CONFIG_MEMORY_HOTPLUG */
1171
1172static struct kcore_list kcore_vsyscall;
1173
1174static void __init register_page_bootmem_info(void)
1175{
1176#ifdef CONFIG_NUMA
1177 int i;
1178
1179 for_each_online_node(i)
1180 register_page_bootmem_info_node(NODE_DATA(i));
1181#endif
1182}
1183
1184void __init mem_init(void)
1185{
1186 pci_iommu_alloc();
1187
1188 /* clear_bss() already clear the empty_zero_page */
1189
1190 /* this will put all memory onto the freelists */
1191 free_all_bootmem();
1192 after_bootmem = 1;
1193 x86_init.hyper.init_after_bootmem();
1194
1195 /*
1196 * Must be done after boot memory is put on freelist, because here we
1197 * might set fields in deferred struct pages that have not yet been
1198 * initialized, and free_all_bootmem() initializes all the reserved
1199 * deferred pages for us.
1200 */
1201 register_page_bootmem_info();
1202
1203 /* Register memory areas for /proc/kcore */
1204 if (get_gate_vma(&init_mm))
1205 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR, PAGE_SIZE, KCORE_USER);
1206
1207 mem_init_print_info(NULL);
1208}
1209
1210int kernel_set_to_readonly;
1211
1212void set_kernel_text_rw(void)
1213{
1214 unsigned long start = PFN_ALIGN(_text);
1215 unsigned long end = PFN_ALIGN(__stop___ex_table);
1216
1217 if (!kernel_set_to_readonly)
1218 return;
1219
1220 pr_debug("Set kernel text: %lx - %lx for read write\n",
1221 start, end);
1222
1223 /*
1224 * Make the kernel identity mapping for text RW. Kernel text
1225 * mapping will always be RO. Refer to the comment in
1226 * static_protections() in pageattr.c
1227 */
1228 set_memory_rw(start, (end - start) >> PAGE_SHIFT);
1229}
1230
1231void set_kernel_text_ro(void)
1232{
1233 unsigned long start = PFN_ALIGN(_text);
1234 unsigned long end = PFN_ALIGN(__stop___ex_table);
1235
1236 if (!kernel_set_to_readonly)
1237 return;
1238
1239 pr_debug("Set kernel text: %lx - %lx for read only\n",
1240 start, end);
1241
1242 /*
1243 * Set the kernel identity mapping for text RO.
1244 */
1245 set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1246}
1247
1248void mark_rodata_ro(void)
1249{
1250 unsigned long start = PFN_ALIGN(_text);
1251 unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1252 unsigned long end = (unsigned long) &__end_rodata_hpage_align;
1253 unsigned long text_end = PFN_ALIGN(&__stop___ex_table);
1254 unsigned long rodata_end = PFN_ALIGN(&__end_rodata);
1255 unsigned long all_end;
1256
1257 printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1258 (end - start) >> 10);
1259 set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1260
1261 kernel_set_to_readonly = 1;
1262
1263 /*
1264 * The rodata/data/bss/brk section (but not the kernel text!)
1265 * should also be not-executable.
1266 *
1267 * We align all_end to PMD_SIZE because the existing mapping
1268 * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1269 * split the PMD and the reminder between _brk_end and the end
1270 * of the PMD will remain mapped executable.
1271 *
1272 * Any PMD which was setup after the one which covers _brk_end
1273 * has been zapped already via cleanup_highmem().
1274 */
1275 all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1276 set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
1277
1278#ifdef CONFIG_CPA_DEBUG
1279 printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1280 set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1281
1282 printk(KERN_INFO "Testing CPA: again\n");
1283 set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1284#endif
1285
1286 free_init_pages("unused kernel",
1287 (unsigned long) __va(__pa_symbol(text_end)),
1288 (unsigned long) __va(__pa_symbol(rodata_start)));
1289 free_init_pages("unused kernel",
1290 (unsigned long) __va(__pa_symbol(rodata_end)),
1291 (unsigned long) __va(__pa_symbol(_sdata)));
1292
1293 debug_checkwx();
1294
1295 /*
1296 * Do this after all of the manipulation of the
1297 * kernel text page tables are complete.
1298 */
1299 pti_clone_kernel_text();
1300}
1301
1302int kern_addr_valid(unsigned long addr)
1303{
1304 unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1305 pgd_t *pgd;
1306 p4d_t *p4d;
1307 pud_t *pud;
1308 pmd_t *pmd;
1309 pte_t *pte;
1310
1311 if (above != 0 && above != -1UL)
1312 return 0;
1313
1314 pgd = pgd_offset_k(addr);
1315 if (pgd_none(*pgd))
1316 return 0;
1317
1318 p4d = p4d_offset(pgd, addr);
1319 if (p4d_none(*p4d))
1320 return 0;
1321
1322 pud = pud_offset(p4d, addr);
1323 if (pud_none(*pud))
1324 return 0;
1325
1326 if (pud_large(*pud))
1327 return pfn_valid(pud_pfn(*pud));
1328
1329 pmd = pmd_offset(pud, addr);
1330 if (pmd_none(*pmd))
1331 return 0;
1332
1333 if (pmd_large(*pmd))
1334 return pfn_valid(pmd_pfn(*pmd));
1335
1336 pte = pte_offset_kernel(pmd, addr);
1337 if (pte_none(*pte))
1338 return 0;
1339
1340 return pfn_valid(pte_pfn(*pte));
1341}
1342
1343/*
1344 * Block size is the minimum amount of memory which can be hotplugged or
1345 * hotremoved. It must be power of two and must be equal or larger than
1346 * MIN_MEMORY_BLOCK_SIZE.
1347 */
1348#define MAX_BLOCK_SIZE (2UL << 30)
1349
1350/* Amount of ram needed to start using large blocks */
1351#define MEM_SIZE_FOR_LARGE_BLOCK (64UL << 30)
1352
1353static unsigned long probe_memory_block_size(void)
1354{
1355 unsigned long boot_mem_end = max_pfn << PAGE_SHIFT;
1356 unsigned long bz;
1357
1358 /* If this is UV system, always set 2G block size */
1359 if (is_uv_system()) {
1360 bz = MAX_BLOCK_SIZE;
1361 goto done;
1362 }
1363
1364 /* Use regular block if RAM is smaller than MEM_SIZE_FOR_LARGE_BLOCK */
1365 if (boot_mem_end < MEM_SIZE_FOR_LARGE_BLOCK) {
1366 bz = MIN_MEMORY_BLOCK_SIZE;
1367 goto done;
1368 }
1369
1370 /* Find the largest allowed block size that aligns to memory end */
1371 for (bz = MAX_BLOCK_SIZE; bz > MIN_MEMORY_BLOCK_SIZE; bz >>= 1) {
1372 if (IS_ALIGNED(boot_mem_end, bz))
1373 break;
1374 }
1375done:
1376 pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20);
1377
1378 return bz;
1379}
1380
1381static unsigned long memory_block_size_probed;
1382unsigned long memory_block_size_bytes(void)
1383{
1384 if (!memory_block_size_probed)
1385 memory_block_size_probed = probe_memory_block_size();
1386
1387 return memory_block_size_probed;
1388}
1389
1390#ifdef CONFIG_SPARSEMEM_VMEMMAP
1391/*
1392 * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1393 */
1394static long __meminitdata addr_start, addr_end;
1395static void __meminitdata *p_start, *p_end;
1396static int __meminitdata node_start;
1397
1398static int __meminit vmemmap_populate_hugepages(unsigned long start,
1399 unsigned long end, int node, struct vmem_altmap *altmap)
1400{
1401 unsigned long addr;
1402 unsigned long next;
1403 pgd_t *pgd;
1404 p4d_t *p4d;
1405 pud_t *pud;
1406 pmd_t *pmd;
1407
1408 for (addr = start; addr < end; addr = next) {
1409 next = pmd_addr_end(addr, end);
1410
1411 pgd = vmemmap_pgd_populate(addr, node);
1412 if (!pgd)
1413 return -ENOMEM;
1414
1415 p4d = vmemmap_p4d_populate(pgd, addr, node);
1416 if (!p4d)
1417 return -ENOMEM;
1418
1419 pud = vmemmap_pud_populate(p4d, addr, node);
1420 if (!pud)
1421 return -ENOMEM;
1422
1423 pmd = pmd_offset(pud, addr);
1424 if (pmd_none(*pmd)) {
1425 void *p;
1426
1427 if (altmap)
1428 p = altmap_alloc_block_buf(PMD_SIZE, altmap);
1429 else
1430 p = vmemmap_alloc_block_buf(PMD_SIZE, node);
1431 if (p) {
1432 pte_t entry;
1433
1434 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1435 PAGE_KERNEL_LARGE);
1436 set_pmd(pmd, __pmd(pte_val(entry)));
1437
1438 /* check to see if we have contiguous blocks */
1439 if (p_end != p || node_start != node) {
1440 if (p_start)
1441 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1442 addr_start, addr_end-1, p_start, p_end-1, node_start);
1443 addr_start = addr;
1444 node_start = node;
1445 p_start = p;
1446 }
1447
1448 addr_end = addr + PMD_SIZE;
1449 p_end = p + PMD_SIZE;
1450 continue;
1451 } else if (altmap)
1452 return -ENOMEM; /* no fallback */
1453 } else if (pmd_large(*pmd)) {
1454 vmemmap_verify((pte_t *)pmd, node, addr, next);
1455 continue;
1456 }
1457 if (vmemmap_populate_basepages(addr, next, node))
1458 return -ENOMEM;
1459 }
1460 return 0;
1461}
1462
1463int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1464 struct vmem_altmap *altmap)
1465{
1466 int err;
1467
1468 if (boot_cpu_has(X86_FEATURE_PSE))
1469 err = vmemmap_populate_hugepages(start, end, node, altmap);
1470 else if (altmap) {
1471 pr_err_once("%s: no cpu support for altmap allocations\n",
1472 __func__);
1473 err = -ENOMEM;
1474 } else
1475 err = vmemmap_populate_basepages(start, end, node);
1476 if (!err)
1477 sync_global_pgds(start, end - 1);
1478 return err;
1479}
1480
1481#if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1482void register_page_bootmem_memmap(unsigned long section_nr,
1483 struct page *start_page, unsigned long nr_pages)
1484{
1485 unsigned long addr = (unsigned long)start_page;
1486 unsigned long end = (unsigned long)(start_page + nr_pages);
1487 unsigned long next;
1488 pgd_t *pgd;
1489 p4d_t *p4d;
1490 pud_t *pud;
1491 pmd_t *pmd;
1492 unsigned int nr_pmd_pages;
1493 struct page *page;
1494
1495 for (; addr < end; addr = next) {
1496 pte_t *pte = NULL;
1497
1498 pgd = pgd_offset_k(addr);
1499 if (pgd_none(*pgd)) {
1500 next = (addr + PAGE_SIZE) & PAGE_MASK;
1501 continue;
1502 }
1503 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1504
1505 p4d = p4d_offset(pgd, addr);
1506 if (p4d_none(*p4d)) {
1507 next = (addr + PAGE_SIZE) & PAGE_MASK;
1508 continue;
1509 }
1510 get_page_bootmem(section_nr, p4d_page(*p4d), MIX_SECTION_INFO);
1511
1512 pud = pud_offset(p4d, addr);
1513 if (pud_none(*pud)) {
1514 next = (addr + PAGE_SIZE) & PAGE_MASK;
1515 continue;
1516 }
1517 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1518
1519 if (!boot_cpu_has(X86_FEATURE_PSE)) {
1520 next = (addr + PAGE_SIZE) & PAGE_MASK;
1521 pmd = pmd_offset(pud, addr);
1522 if (pmd_none(*pmd))
1523 continue;
1524 get_page_bootmem(section_nr, pmd_page(*pmd),
1525 MIX_SECTION_INFO);
1526
1527 pte = pte_offset_kernel(pmd, addr);
1528 if (pte_none(*pte))
1529 continue;
1530 get_page_bootmem(section_nr, pte_page(*pte),
1531 SECTION_INFO);
1532 } else {
1533 next = pmd_addr_end(addr, end);
1534
1535 pmd = pmd_offset(pud, addr);
1536 if (pmd_none(*pmd))
1537 continue;
1538
1539 nr_pmd_pages = 1 << get_order(PMD_SIZE);
1540 page = pmd_page(*pmd);
1541 while (nr_pmd_pages--)
1542 get_page_bootmem(section_nr, page++,
1543 SECTION_INFO);
1544 }
1545 }
1546}
1547#endif
1548
1549void __meminit vmemmap_populate_print_last(void)
1550{
1551 if (p_start) {
1552 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1553 addr_start, addr_end-1, p_start, p_end-1, node_start);
1554 p_start = NULL;
1555 p_end = NULL;
1556 node_start = 0;
1557 }
1558}
1559#endif