Loading...
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/module.h>
31#include <linux/memory.h>
32#include <linux/memory_hotplug.h>
33#include <linux/memremap.h>
34#include <linux/nmi.h>
35#include <linux/gfp.h>
36#include <linux/kcore.h>
37
38#include <asm/processor.h>
39#include <asm/bios_ebda.h>
40#include <asm/uaccess.h>
41#include <asm/pgtable.h>
42#include <asm/pgalloc.h>
43#include <asm/dma.h>
44#include <asm/fixmap.h>
45#include <asm/e820.h>
46#include <asm/apic.h>
47#include <asm/tlb.h>
48#include <asm/mmu_context.h>
49#include <asm/proto.h>
50#include <asm/smp.h>
51#include <asm/sections.h>
52#include <asm/kdebug.h>
53#include <asm/numa.h>
54#include <asm/cacheflush.h>
55#include <asm/init.h>
56#include <asm/uv/uv.h>
57#include <asm/setup.h>
58
59#include "mm_internal.h"
60
61static void ident_pmd_init(unsigned long pmd_flag, pmd_t *pmd_page,
62 unsigned long addr, unsigned long end)
63{
64 addr &= PMD_MASK;
65 for (; addr < end; addr += PMD_SIZE) {
66 pmd_t *pmd = pmd_page + pmd_index(addr);
67
68 if (!pmd_present(*pmd))
69 set_pmd(pmd, __pmd(addr | pmd_flag));
70 }
71}
72static int ident_pud_init(struct x86_mapping_info *info, pud_t *pud_page,
73 unsigned long addr, unsigned long end)
74{
75 unsigned long next;
76
77 for (; addr < end; addr = next) {
78 pud_t *pud = pud_page + pud_index(addr);
79 pmd_t *pmd;
80
81 next = (addr & PUD_MASK) + PUD_SIZE;
82 if (next > end)
83 next = end;
84
85 if (pud_present(*pud)) {
86 pmd = pmd_offset(pud, 0);
87 ident_pmd_init(info->pmd_flag, pmd, addr, next);
88 continue;
89 }
90 pmd = (pmd_t *)info->alloc_pgt_page(info->context);
91 if (!pmd)
92 return -ENOMEM;
93 ident_pmd_init(info->pmd_flag, pmd, addr, next);
94 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
95 }
96
97 return 0;
98}
99
100int kernel_ident_mapping_init(struct x86_mapping_info *info, pgd_t *pgd_page,
101 unsigned long addr, unsigned long end)
102{
103 unsigned long next;
104 int result;
105 int off = info->kernel_mapping ? pgd_index(__PAGE_OFFSET) : 0;
106
107 for (; addr < end; addr = next) {
108 pgd_t *pgd = pgd_page + pgd_index(addr) + off;
109 pud_t *pud;
110
111 next = (addr & PGDIR_MASK) + PGDIR_SIZE;
112 if (next > end)
113 next = end;
114
115 if (pgd_present(*pgd)) {
116 pud = pud_offset(pgd, 0);
117 result = ident_pud_init(info, pud, addr, next);
118 if (result)
119 return result;
120 continue;
121 }
122
123 pud = (pud_t *)info->alloc_pgt_page(info->context);
124 if (!pud)
125 return -ENOMEM;
126 result = ident_pud_init(info, pud, addr, next);
127 if (result)
128 return result;
129 set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE));
130 }
131
132 return 0;
133}
134
135/*
136 * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
137 * physical space so we can cache the place of the first one and move
138 * around without checking the pgd every time.
139 */
140
141pteval_t __supported_pte_mask __read_mostly = ~0;
142EXPORT_SYMBOL_GPL(__supported_pte_mask);
143
144int force_personality32;
145
146/*
147 * noexec32=on|off
148 * Control non executable heap for 32bit processes.
149 * To control the stack too use noexec=off
150 *
151 * on PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
152 * off PROT_READ implies PROT_EXEC
153 */
154static int __init nonx32_setup(char *str)
155{
156 if (!strcmp(str, "on"))
157 force_personality32 &= ~READ_IMPLIES_EXEC;
158 else if (!strcmp(str, "off"))
159 force_personality32 |= READ_IMPLIES_EXEC;
160 return 1;
161}
162__setup("noexec32=", nonx32_setup);
163
164/*
165 * When memory was added/removed make sure all the processes MM have
166 * suitable PGD entries in the local PGD level page.
167 */
168void sync_global_pgds(unsigned long start, unsigned long end, int removed)
169{
170 unsigned long address;
171
172 for (address = start; address <= end; address += PGDIR_SIZE) {
173 const pgd_t *pgd_ref = pgd_offset_k(address);
174 struct page *page;
175
176 /*
177 * When it is called after memory hot remove, pgd_none()
178 * returns true. In this case (removed == 1), we must clear
179 * the PGD entries in the local PGD level page.
180 */
181 if (pgd_none(*pgd_ref) && !removed)
182 continue;
183
184 spin_lock(&pgd_lock);
185 list_for_each_entry(page, &pgd_list, lru) {
186 pgd_t *pgd;
187 spinlock_t *pgt_lock;
188
189 pgd = (pgd_t *)page_address(page) + pgd_index(address);
190 /* the pgt_lock only for Xen */
191 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
192 spin_lock(pgt_lock);
193
194 if (!pgd_none(*pgd_ref) && !pgd_none(*pgd))
195 BUG_ON(pgd_page_vaddr(*pgd)
196 != pgd_page_vaddr(*pgd_ref));
197
198 if (removed) {
199 if (pgd_none(*pgd_ref) && !pgd_none(*pgd))
200 pgd_clear(pgd);
201 } else {
202 if (pgd_none(*pgd))
203 set_pgd(pgd, *pgd_ref);
204 }
205
206 spin_unlock(pgt_lock);
207 }
208 spin_unlock(&pgd_lock);
209 }
210}
211
212/*
213 * NOTE: This function is marked __ref because it calls __init function
214 * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
215 */
216static __ref void *spp_getpage(void)
217{
218 void *ptr;
219
220 if (after_bootmem)
221 ptr = (void *) get_zeroed_page(GFP_ATOMIC | __GFP_NOTRACK);
222 else
223 ptr = alloc_bootmem_pages(PAGE_SIZE);
224
225 if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
226 panic("set_pte_phys: cannot allocate page data %s\n",
227 after_bootmem ? "after bootmem" : "");
228 }
229
230 pr_debug("spp_getpage %p\n", ptr);
231
232 return ptr;
233}
234
235static pud_t *fill_pud(pgd_t *pgd, unsigned long vaddr)
236{
237 if (pgd_none(*pgd)) {
238 pud_t *pud = (pud_t *)spp_getpage();
239 pgd_populate(&init_mm, pgd, pud);
240 if (pud != pud_offset(pgd, 0))
241 printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
242 pud, pud_offset(pgd, 0));
243 }
244 return pud_offset(pgd, vaddr);
245}
246
247static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
248{
249 if (pud_none(*pud)) {
250 pmd_t *pmd = (pmd_t *) spp_getpage();
251 pud_populate(&init_mm, pud, pmd);
252 if (pmd != pmd_offset(pud, 0))
253 printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
254 pmd, pmd_offset(pud, 0));
255 }
256 return pmd_offset(pud, vaddr);
257}
258
259static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
260{
261 if (pmd_none(*pmd)) {
262 pte_t *pte = (pte_t *) spp_getpage();
263 pmd_populate_kernel(&init_mm, pmd, pte);
264 if (pte != pte_offset_kernel(pmd, 0))
265 printk(KERN_ERR "PAGETABLE BUG #02!\n");
266 }
267 return pte_offset_kernel(pmd, vaddr);
268}
269
270void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
271{
272 pud_t *pud;
273 pmd_t *pmd;
274 pte_t *pte;
275
276 pud = pud_page + pud_index(vaddr);
277 pmd = fill_pmd(pud, vaddr);
278 pte = fill_pte(pmd, vaddr);
279
280 set_pte(pte, new_pte);
281
282 /*
283 * It's enough to flush this one mapping.
284 * (PGE mappings get flushed as well)
285 */
286 __flush_tlb_one(vaddr);
287}
288
289void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
290{
291 pgd_t *pgd;
292 pud_t *pud_page;
293
294 pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
295
296 pgd = pgd_offset_k(vaddr);
297 if (pgd_none(*pgd)) {
298 printk(KERN_ERR
299 "PGD FIXMAP MISSING, it should be setup in head.S!\n");
300 return;
301 }
302 pud_page = (pud_t*)pgd_page_vaddr(*pgd);
303 set_pte_vaddr_pud(pud_page, vaddr, pteval);
304}
305
306pmd_t * __init populate_extra_pmd(unsigned long vaddr)
307{
308 pgd_t *pgd;
309 pud_t *pud;
310
311 pgd = pgd_offset_k(vaddr);
312 pud = fill_pud(pgd, vaddr);
313 return fill_pmd(pud, vaddr);
314}
315
316pte_t * __init populate_extra_pte(unsigned long vaddr)
317{
318 pmd_t *pmd;
319
320 pmd = populate_extra_pmd(vaddr);
321 return fill_pte(pmd, vaddr);
322}
323
324/*
325 * Create large page table mappings for a range of physical addresses.
326 */
327static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
328 enum page_cache_mode cache)
329{
330 pgd_t *pgd;
331 pud_t *pud;
332 pmd_t *pmd;
333 pgprot_t prot;
334
335 pgprot_val(prot) = pgprot_val(PAGE_KERNEL_LARGE) |
336 pgprot_val(pgprot_4k_2_large(cachemode2pgprot(cache)));
337 BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
338 for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
339 pgd = pgd_offset_k((unsigned long)__va(phys));
340 if (pgd_none(*pgd)) {
341 pud = (pud_t *) spp_getpage();
342 set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
343 _PAGE_USER));
344 }
345 pud = pud_offset(pgd, (unsigned long)__va(phys));
346 if (pud_none(*pud)) {
347 pmd = (pmd_t *) spp_getpage();
348 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
349 _PAGE_USER));
350 }
351 pmd = pmd_offset(pud, phys);
352 BUG_ON(!pmd_none(*pmd));
353 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
354 }
355}
356
357void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
358{
359 __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_WB);
360}
361
362void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
363{
364 __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_UC);
365}
366
367/*
368 * The head.S code sets up the kernel high mapping:
369 *
370 * from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
371 *
372 * phys_base holds the negative offset to the kernel, which is added
373 * to the compile time generated pmds. This results in invalid pmds up
374 * to the point where we hit the physaddr 0 mapping.
375 *
376 * We limit the mappings to the region from _text to _brk_end. _brk_end
377 * is rounded up to the 2MB boundary. This catches the invalid pmds as
378 * well, as they are located before _text:
379 */
380void __init cleanup_highmap(void)
381{
382 unsigned long vaddr = __START_KERNEL_map;
383 unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
384 unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
385 pmd_t *pmd = level2_kernel_pgt;
386
387 /*
388 * Native path, max_pfn_mapped is not set yet.
389 * Xen has valid max_pfn_mapped set in
390 * arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
391 */
392 if (max_pfn_mapped)
393 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
394
395 for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
396 if (pmd_none(*pmd))
397 continue;
398 if (vaddr < (unsigned long) _text || vaddr > end)
399 set_pmd(pmd, __pmd(0));
400 }
401}
402
403static unsigned long __meminit
404phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
405 pgprot_t prot)
406{
407 unsigned long pages = 0, next;
408 unsigned long last_map_addr = end;
409 int i;
410
411 pte_t *pte = pte_page + pte_index(addr);
412
413 for (i = pte_index(addr); i < PTRS_PER_PTE; i++, addr = next, pte++) {
414 next = (addr & PAGE_MASK) + PAGE_SIZE;
415 if (addr >= end) {
416 if (!after_bootmem &&
417 !e820_any_mapped(addr & PAGE_MASK, next, E820_RAM) &&
418 !e820_any_mapped(addr & PAGE_MASK, next, E820_RESERVED_KERN))
419 set_pte(pte, __pte(0));
420 continue;
421 }
422
423 /*
424 * We will re-use the existing mapping.
425 * Xen for example has some special requirements, like mapping
426 * pagetable pages as RO. So assume someone who pre-setup
427 * these mappings are more intelligent.
428 */
429 if (pte_val(*pte)) {
430 if (!after_bootmem)
431 pages++;
432 continue;
433 }
434
435 if (0)
436 printk(" pte=%p addr=%lx pte=%016lx\n",
437 pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
438 pages++;
439 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
440 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
441 }
442
443 update_page_count(PG_LEVEL_4K, pages);
444
445 return last_map_addr;
446}
447
448static unsigned long __meminit
449phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
450 unsigned long page_size_mask, pgprot_t prot)
451{
452 unsigned long pages = 0, next;
453 unsigned long last_map_addr = end;
454
455 int i = pmd_index(address);
456
457 for (; i < PTRS_PER_PMD; i++, address = next) {
458 pmd_t *pmd = pmd_page + pmd_index(address);
459 pte_t *pte;
460 pgprot_t new_prot = prot;
461
462 next = (address & PMD_MASK) + PMD_SIZE;
463 if (address >= end) {
464 if (!after_bootmem &&
465 !e820_any_mapped(address & PMD_MASK, next, E820_RAM) &&
466 !e820_any_mapped(address & PMD_MASK, next, E820_RESERVED_KERN))
467 set_pmd(pmd, __pmd(0));
468 continue;
469 }
470
471 if (pmd_val(*pmd)) {
472 if (!pmd_large(*pmd)) {
473 spin_lock(&init_mm.page_table_lock);
474 pte = (pte_t *)pmd_page_vaddr(*pmd);
475 last_map_addr = phys_pte_init(pte, address,
476 end, prot);
477 spin_unlock(&init_mm.page_table_lock);
478 continue;
479 }
480 /*
481 * If we are ok with PG_LEVEL_2M mapping, then we will
482 * use the existing mapping,
483 *
484 * Otherwise, we will split the large page mapping but
485 * use the same existing protection bits except for
486 * large page, so that we don't violate Intel's TLB
487 * Application note (317080) which says, while changing
488 * the page sizes, new and old translations should
489 * not differ with respect to page frame and
490 * attributes.
491 */
492 if (page_size_mask & (1 << PG_LEVEL_2M)) {
493 if (!after_bootmem)
494 pages++;
495 last_map_addr = next;
496 continue;
497 }
498 new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
499 }
500
501 if (page_size_mask & (1<<PG_LEVEL_2M)) {
502 pages++;
503 spin_lock(&init_mm.page_table_lock);
504 set_pte((pte_t *)pmd,
505 pfn_pte((address & PMD_MASK) >> PAGE_SHIFT,
506 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
507 spin_unlock(&init_mm.page_table_lock);
508 last_map_addr = next;
509 continue;
510 }
511
512 pte = alloc_low_page();
513 last_map_addr = phys_pte_init(pte, address, end, new_prot);
514
515 spin_lock(&init_mm.page_table_lock);
516 pmd_populate_kernel(&init_mm, pmd, pte);
517 spin_unlock(&init_mm.page_table_lock);
518 }
519 update_page_count(PG_LEVEL_2M, pages);
520 return last_map_addr;
521}
522
523static unsigned long __meminit
524phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
525 unsigned long page_size_mask)
526{
527 unsigned long pages = 0, next;
528 unsigned long last_map_addr = end;
529 int i = pud_index(addr);
530
531 for (; i < PTRS_PER_PUD; i++, addr = next) {
532 pud_t *pud = pud_page + pud_index(addr);
533 pmd_t *pmd;
534 pgprot_t prot = PAGE_KERNEL;
535
536 next = (addr & PUD_MASK) + PUD_SIZE;
537 if (addr >= end) {
538 if (!after_bootmem &&
539 !e820_any_mapped(addr & PUD_MASK, next, E820_RAM) &&
540 !e820_any_mapped(addr & PUD_MASK, next, E820_RESERVED_KERN))
541 set_pud(pud, __pud(0));
542 continue;
543 }
544
545 if (pud_val(*pud)) {
546 if (!pud_large(*pud)) {
547 pmd = pmd_offset(pud, 0);
548 last_map_addr = phys_pmd_init(pmd, addr, end,
549 page_size_mask, prot);
550 __flush_tlb_all();
551 continue;
552 }
553 /*
554 * If we are ok with PG_LEVEL_1G mapping, then we will
555 * use the existing mapping.
556 *
557 * Otherwise, we will split the gbpage mapping but use
558 * the same existing protection bits except for large
559 * page, so that we don't violate Intel's TLB
560 * Application note (317080) which says, while changing
561 * the page sizes, new and old translations should
562 * not differ with respect to page frame and
563 * attributes.
564 */
565 if (page_size_mask & (1 << PG_LEVEL_1G)) {
566 if (!after_bootmem)
567 pages++;
568 last_map_addr = next;
569 continue;
570 }
571 prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
572 }
573
574 if (page_size_mask & (1<<PG_LEVEL_1G)) {
575 pages++;
576 spin_lock(&init_mm.page_table_lock);
577 set_pte((pte_t *)pud,
578 pfn_pte((addr & PUD_MASK) >> PAGE_SHIFT,
579 PAGE_KERNEL_LARGE));
580 spin_unlock(&init_mm.page_table_lock);
581 last_map_addr = next;
582 continue;
583 }
584
585 pmd = alloc_low_page();
586 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
587 prot);
588
589 spin_lock(&init_mm.page_table_lock);
590 pud_populate(&init_mm, pud, pmd);
591 spin_unlock(&init_mm.page_table_lock);
592 }
593 __flush_tlb_all();
594
595 update_page_count(PG_LEVEL_1G, pages);
596
597 return last_map_addr;
598}
599
600unsigned long __meminit
601kernel_physical_mapping_init(unsigned long start,
602 unsigned long end,
603 unsigned long page_size_mask)
604{
605 bool pgd_changed = false;
606 unsigned long next, last_map_addr = end;
607 unsigned long addr;
608
609 start = (unsigned long)__va(start);
610 end = (unsigned long)__va(end);
611 addr = start;
612
613 for (; start < end; start = next) {
614 pgd_t *pgd = pgd_offset_k(start);
615 pud_t *pud;
616
617 next = (start & PGDIR_MASK) + PGDIR_SIZE;
618
619 if (pgd_val(*pgd)) {
620 pud = (pud_t *)pgd_page_vaddr(*pgd);
621 last_map_addr = phys_pud_init(pud, __pa(start),
622 __pa(end), page_size_mask);
623 continue;
624 }
625
626 pud = alloc_low_page();
627 last_map_addr = phys_pud_init(pud, __pa(start), __pa(end),
628 page_size_mask);
629
630 spin_lock(&init_mm.page_table_lock);
631 pgd_populate(&init_mm, pgd, pud);
632 spin_unlock(&init_mm.page_table_lock);
633 pgd_changed = true;
634 }
635
636 if (pgd_changed)
637 sync_global_pgds(addr, end - 1, 0);
638
639 __flush_tlb_all();
640
641 return last_map_addr;
642}
643
644#ifndef CONFIG_NUMA
645void __init initmem_init(void)
646{
647 memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
648}
649#endif
650
651void __init paging_init(void)
652{
653 sparse_memory_present_with_active_regions(MAX_NUMNODES);
654 sparse_init();
655
656 /*
657 * clear the default setting with node 0
658 * note: don't use nodes_clear here, that is really clearing when
659 * numa support is not compiled in, and later node_set_state
660 * will not set it back.
661 */
662 node_clear_state(0, N_MEMORY);
663 if (N_MEMORY != N_NORMAL_MEMORY)
664 node_clear_state(0, N_NORMAL_MEMORY);
665
666 zone_sizes_init();
667}
668
669/*
670 * Memory hotplug specific functions
671 */
672#ifdef CONFIG_MEMORY_HOTPLUG
673/*
674 * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
675 * updating.
676 */
677static void update_end_of_memory_vars(u64 start, u64 size)
678{
679 unsigned long end_pfn = PFN_UP(start + size);
680
681 if (end_pfn > max_pfn) {
682 max_pfn = end_pfn;
683 max_low_pfn = end_pfn;
684 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
685 }
686}
687
688/*
689 * Memory is added always to NORMAL zone. This means you will never get
690 * additional DMA/DMA32 memory.
691 */
692int arch_add_memory(int nid, u64 start, u64 size, bool for_device)
693{
694 struct pglist_data *pgdat = NODE_DATA(nid);
695 struct zone *zone = pgdat->node_zones +
696 zone_for_memory(nid, start, size, ZONE_NORMAL, for_device);
697 unsigned long start_pfn = start >> PAGE_SHIFT;
698 unsigned long nr_pages = size >> PAGE_SHIFT;
699 int ret;
700
701 init_memory_mapping(start, start + size);
702
703 ret = __add_pages(nid, zone, start_pfn, nr_pages);
704 WARN_ON_ONCE(ret);
705
706 /* update max_pfn, max_low_pfn and high_memory */
707 update_end_of_memory_vars(start, size);
708
709 return ret;
710}
711EXPORT_SYMBOL_GPL(arch_add_memory);
712
713#define PAGE_INUSE 0xFD
714
715static void __meminit free_pagetable(struct page *page, int order)
716{
717 unsigned long magic;
718 unsigned int nr_pages = 1 << order;
719 struct vmem_altmap *altmap = to_vmem_altmap((unsigned long) page);
720
721 if (altmap) {
722 vmem_altmap_free(altmap, nr_pages);
723 return;
724 }
725
726 /* bootmem page has reserved flag */
727 if (PageReserved(page)) {
728 __ClearPageReserved(page);
729
730 magic = (unsigned long)page->lru.next;
731 if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
732 while (nr_pages--)
733 put_page_bootmem(page++);
734 } else
735 while (nr_pages--)
736 free_reserved_page(page++);
737 } else
738 free_pages((unsigned long)page_address(page), order);
739}
740
741static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
742{
743 pte_t *pte;
744 int i;
745
746 for (i = 0; i < PTRS_PER_PTE; i++) {
747 pte = pte_start + i;
748 if (pte_val(*pte))
749 return;
750 }
751
752 /* free a pte talbe */
753 free_pagetable(pmd_page(*pmd), 0);
754 spin_lock(&init_mm.page_table_lock);
755 pmd_clear(pmd);
756 spin_unlock(&init_mm.page_table_lock);
757}
758
759static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
760{
761 pmd_t *pmd;
762 int i;
763
764 for (i = 0; i < PTRS_PER_PMD; i++) {
765 pmd = pmd_start + i;
766 if (pmd_val(*pmd))
767 return;
768 }
769
770 /* free a pmd talbe */
771 free_pagetable(pud_page(*pud), 0);
772 spin_lock(&init_mm.page_table_lock);
773 pud_clear(pud);
774 spin_unlock(&init_mm.page_table_lock);
775}
776
777/* Return true if pgd is changed, otherwise return false. */
778static bool __meminit free_pud_table(pud_t *pud_start, pgd_t *pgd)
779{
780 pud_t *pud;
781 int i;
782
783 for (i = 0; i < PTRS_PER_PUD; i++) {
784 pud = pud_start + i;
785 if (pud_val(*pud))
786 return false;
787 }
788
789 /* free a pud table */
790 free_pagetable(pgd_page(*pgd), 0);
791 spin_lock(&init_mm.page_table_lock);
792 pgd_clear(pgd);
793 spin_unlock(&init_mm.page_table_lock);
794
795 return true;
796}
797
798static void __meminit
799remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
800 bool direct)
801{
802 unsigned long next, pages = 0;
803 pte_t *pte;
804 void *page_addr;
805 phys_addr_t phys_addr;
806
807 pte = pte_start + pte_index(addr);
808 for (; addr < end; addr = next, pte++) {
809 next = (addr + PAGE_SIZE) & PAGE_MASK;
810 if (next > end)
811 next = end;
812
813 if (!pte_present(*pte))
814 continue;
815
816 /*
817 * We mapped [0,1G) memory as identity mapping when
818 * initializing, in arch/x86/kernel/head_64.S. These
819 * pagetables cannot be removed.
820 */
821 phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
822 if (phys_addr < (phys_addr_t)0x40000000)
823 return;
824
825 if (PAGE_ALIGNED(addr) && PAGE_ALIGNED(next)) {
826 /*
827 * Do not free direct mapping pages since they were
828 * freed when offlining, or simplely not in use.
829 */
830 if (!direct)
831 free_pagetable(pte_page(*pte), 0);
832
833 spin_lock(&init_mm.page_table_lock);
834 pte_clear(&init_mm, addr, pte);
835 spin_unlock(&init_mm.page_table_lock);
836
837 /* For non-direct mapping, pages means nothing. */
838 pages++;
839 } else {
840 /*
841 * If we are here, we are freeing vmemmap pages since
842 * direct mapped memory ranges to be freed are aligned.
843 *
844 * If we are not removing the whole page, it means
845 * other page structs in this page are being used and
846 * we canot remove them. So fill the unused page_structs
847 * with 0xFD, and remove the page when it is wholly
848 * filled with 0xFD.
849 */
850 memset((void *)addr, PAGE_INUSE, next - addr);
851
852 page_addr = page_address(pte_page(*pte));
853 if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) {
854 free_pagetable(pte_page(*pte), 0);
855
856 spin_lock(&init_mm.page_table_lock);
857 pte_clear(&init_mm, addr, pte);
858 spin_unlock(&init_mm.page_table_lock);
859 }
860 }
861 }
862
863 /* Call free_pte_table() in remove_pmd_table(). */
864 flush_tlb_all();
865 if (direct)
866 update_page_count(PG_LEVEL_4K, -pages);
867}
868
869static void __meminit
870remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
871 bool direct)
872{
873 unsigned long next, pages = 0;
874 pte_t *pte_base;
875 pmd_t *pmd;
876 void *page_addr;
877
878 pmd = pmd_start + pmd_index(addr);
879 for (; addr < end; addr = next, pmd++) {
880 next = pmd_addr_end(addr, end);
881
882 if (!pmd_present(*pmd))
883 continue;
884
885 if (pmd_large(*pmd)) {
886 if (IS_ALIGNED(addr, PMD_SIZE) &&
887 IS_ALIGNED(next, PMD_SIZE)) {
888 if (!direct)
889 free_pagetable(pmd_page(*pmd),
890 get_order(PMD_SIZE));
891
892 spin_lock(&init_mm.page_table_lock);
893 pmd_clear(pmd);
894 spin_unlock(&init_mm.page_table_lock);
895 pages++;
896 } else {
897 /* If here, we are freeing vmemmap pages. */
898 memset((void *)addr, PAGE_INUSE, next - addr);
899
900 page_addr = page_address(pmd_page(*pmd));
901 if (!memchr_inv(page_addr, PAGE_INUSE,
902 PMD_SIZE)) {
903 free_pagetable(pmd_page(*pmd),
904 get_order(PMD_SIZE));
905
906 spin_lock(&init_mm.page_table_lock);
907 pmd_clear(pmd);
908 spin_unlock(&init_mm.page_table_lock);
909 }
910 }
911
912 continue;
913 }
914
915 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
916 remove_pte_table(pte_base, addr, next, direct);
917 free_pte_table(pte_base, pmd);
918 }
919
920 /* Call free_pmd_table() in remove_pud_table(). */
921 if (direct)
922 update_page_count(PG_LEVEL_2M, -pages);
923}
924
925static void __meminit
926remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
927 bool direct)
928{
929 unsigned long next, pages = 0;
930 pmd_t *pmd_base;
931 pud_t *pud;
932 void *page_addr;
933
934 pud = pud_start + pud_index(addr);
935 for (; addr < end; addr = next, pud++) {
936 next = pud_addr_end(addr, end);
937
938 if (!pud_present(*pud))
939 continue;
940
941 if (pud_large(*pud)) {
942 if (IS_ALIGNED(addr, PUD_SIZE) &&
943 IS_ALIGNED(next, PUD_SIZE)) {
944 if (!direct)
945 free_pagetable(pud_page(*pud),
946 get_order(PUD_SIZE));
947
948 spin_lock(&init_mm.page_table_lock);
949 pud_clear(pud);
950 spin_unlock(&init_mm.page_table_lock);
951 pages++;
952 } else {
953 /* If here, we are freeing vmemmap pages. */
954 memset((void *)addr, PAGE_INUSE, next - addr);
955
956 page_addr = page_address(pud_page(*pud));
957 if (!memchr_inv(page_addr, PAGE_INUSE,
958 PUD_SIZE)) {
959 free_pagetable(pud_page(*pud),
960 get_order(PUD_SIZE));
961
962 spin_lock(&init_mm.page_table_lock);
963 pud_clear(pud);
964 spin_unlock(&init_mm.page_table_lock);
965 }
966 }
967
968 continue;
969 }
970
971 pmd_base = (pmd_t *)pud_page_vaddr(*pud);
972 remove_pmd_table(pmd_base, addr, next, direct);
973 free_pmd_table(pmd_base, pud);
974 }
975
976 if (direct)
977 update_page_count(PG_LEVEL_1G, -pages);
978}
979
980/* start and end are both virtual address. */
981static void __meminit
982remove_pagetable(unsigned long start, unsigned long end, bool direct)
983{
984 unsigned long next;
985 unsigned long addr;
986 pgd_t *pgd;
987 pud_t *pud;
988 bool pgd_changed = false;
989
990 for (addr = start; addr < end; addr = next) {
991 next = pgd_addr_end(addr, end);
992
993 pgd = pgd_offset_k(addr);
994 if (!pgd_present(*pgd))
995 continue;
996
997 pud = (pud_t *)pgd_page_vaddr(*pgd);
998 remove_pud_table(pud, addr, next, direct);
999 if (free_pud_table(pud, pgd))
1000 pgd_changed = true;
1001 }
1002
1003 if (pgd_changed)
1004 sync_global_pgds(start, end - 1, 1);
1005
1006 flush_tlb_all();
1007}
1008
1009void __ref vmemmap_free(unsigned long start, unsigned long end)
1010{
1011 remove_pagetable(start, end, false);
1012}
1013
1014#ifdef CONFIG_MEMORY_HOTREMOVE
1015static void __meminit
1016kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1017{
1018 start = (unsigned long)__va(start);
1019 end = (unsigned long)__va(end);
1020
1021 remove_pagetable(start, end, true);
1022}
1023
1024int __ref arch_remove_memory(u64 start, u64 size)
1025{
1026 unsigned long start_pfn = start >> PAGE_SHIFT;
1027 unsigned long nr_pages = size >> PAGE_SHIFT;
1028 struct page *page = pfn_to_page(start_pfn);
1029 struct vmem_altmap *altmap;
1030 struct zone *zone;
1031 int ret;
1032
1033 /* With altmap the first mapped page is offset from @start */
1034 altmap = to_vmem_altmap((unsigned long) page);
1035 if (altmap)
1036 page += vmem_altmap_offset(altmap);
1037 zone = page_zone(page);
1038 ret = __remove_pages(zone, start_pfn, nr_pages);
1039 WARN_ON_ONCE(ret);
1040 kernel_physical_mapping_remove(start, start + size);
1041
1042 return ret;
1043}
1044#endif
1045#endif /* CONFIG_MEMORY_HOTPLUG */
1046
1047static struct kcore_list kcore_vsyscall;
1048
1049static void __init register_page_bootmem_info(void)
1050{
1051#ifdef CONFIG_NUMA
1052 int i;
1053
1054 for_each_online_node(i)
1055 register_page_bootmem_info_node(NODE_DATA(i));
1056#endif
1057}
1058
1059void __init mem_init(void)
1060{
1061 pci_iommu_alloc();
1062
1063 /* clear_bss() already clear the empty_zero_page */
1064
1065 register_page_bootmem_info();
1066
1067 /* this will put all memory onto the freelists */
1068 free_all_bootmem();
1069 after_bootmem = 1;
1070
1071 /* Register memory areas for /proc/kcore */
1072 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR,
1073 PAGE_SIZE, KCORE_OTHER);
1074
1075 mem_init_print_info(NULL);
1076}
1077
1078const int rodata_test_data = 0xC3;
1079EXPORT_SYMBOL_GPL(rodata_test_data);
1080
1081int kernel_set_to_readonly;
1082
1083void set_kernel_text_rw(void)
1084{
1085 unsigned long start = PFN_ALIGN(_text);
1086 unsigned long end = PFN_ALIGN(__stop___ex_table);
1087
1088 if (!kernel_set_to_readonly)
1089 return;
1090
1091 pr_debug("Set kernel text: %lx - %lx for read write\n",
1092 start, end);
1093
1094 /*
1095 * Make the kernel identity mapping for text RW. Kernel text
1096 * mapping will always be RO. Refer to the comment in
1097 * static_protections() in pageattr.c
1098 */
1099 set_memory_rw(start, (end - start) >> PAGE_SHIFT);
1100}
1101
1102void set_kernel_text_ro(void)
1103{
1104 unsigned long start = PFN_ALIGN(_text);
1105 unsigned long end = PFN_ALIGN(__stop___ex_table);
1106
1107 if (!kernel_set_to_readonly)
1108 return;
1109
1110 pr_debug("Set kernel text: %lx - %lx for read only\n",
1111 start, end);
1112
1113 /*
1114 * Set the kernel identity mapping for text RO.
1115 */
1116 set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1117}
1118
1119void mark_rodata_ro(void)
1120{
1121 unsigned long start = PFN_ALIGN(_text);
1122 unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1123 unsigned long end = (unsigned long) &__end_rodata_hpage_align;
1124 unsigned long text_end = PFN_ALIGN(&__stop___ex_table);
1125 unsigned long rodata_end = PFN_ALIGN(&__end_rodata);
1126 unsigned long all_end;
1127
1128 printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1129 (end - start) >> 10);
1130 set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1131
1132 kernel_set_to_readonly = 1;
1133
1134 /*
1135 * The rodata/data/bss/brk section (but not the kernel text!)
1136 * should also be not-executable.
1137 *
1138 * We align all_end to PMD_SIZE because the existing mapping
1139 * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1140 * split the PMD and the reminder between _brk_end and the end
1141 * of the PMD will remain mapped executable.
1142 *
1143 * Any PMD which was setup after the one which covers _brk_end
1144 * has been zapped already via cleanup_highmem().
1145 */
1146 all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1147 set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
1148
1149 rodata_test();
1150
1151#ifdef CONFIG_CPA_DEBUG
1152 printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1153 set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1154
1155 printk(KERN_INFO "Testing CPA: again\n");
1156 set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1157#endif
1158
1159 free_init_pages("unused kernel",
1160 (unsigned long) __va(__pa_symbol(text_end)),
1161 (unsigned long) __va(__pa_symbol(rodata_start)));
1162 free_init_pages("unused kernel",
1163 (unsigned long) __va(__pa_symbol(rodata_end)),
1164 (unsigned long) __va(__pa_symbol(_sdata)));
1165
1166 debug_checkwx();
1167}
1168
1169int kern_addr_valid(unsigned long addr)
1170{
1171 unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1172 pgd_t *pgd;
1173 pud_t *pud;
1174 pmd_t *pmd;
1175 pte_t *pte;
1176
1177 if (above != 0 && above != -1UL)
1178 return 0;
1179
1180 pgd = pgd_offset_k(addr);
1181 if (pgd_none(*pgd))
1182 return 0;
1183
1184 pud = pud_offset(pgd, addr);
1185 if (pud_none(*pud))
1186 return 0;
1187
1188 if (pud_large(*pud))
1189 return pfn_valid(pud_pfn(*pud));
1190
1191 pmd = pmd_offset(pud, addr);
1192 if (pmd_none(*pmd))
1193 return 0;
1194
1195 if (pmd_large(*pmd))
1196 return pfn_valid(pmd_pfn(*pmd));
1197
1198 pte = pte_offset_kernel(pmd, addr);
1199 if (pte_none(*pte))
1200 return 0;
1201
1202 return pfn_valid(pte_pfn(*pte));
1203}
1204
1205static unsigned long probe_memory_block_size(void)
1206{
1207 unsigned long bz = MIN_MEMORY_BLOCK_SIZE;
1208
1209 /* if system is UV or has 64GB of RAM or more, use large blocks */
1210 if (is_uv_system() || ((max_pfn << PAGE_SHIFT) >= (64UL << 30)))
1211 bz = 2UL << 30; /* 2GB */
1212
1213 pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20);
1214
1215 return bz;
1216}
1217
1218static unsigned long memory_block_size_probed;
1219unsigned long memory_block_size_bytes(void)
1220{
1221 if (!memory_block_size_probed)
1222 memory_block_size_probed = probe_memory_block_size();
1223
1224 return memory_block_size_probed;
1225}
1226
1227#ifdef CONFIG_SPARSEMEM_VMEMMAP
1228/*
1229 * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1230 */
1231static long __meminitdata addr_start, addr_end;
1232static void __meminitdata *p_start, *p_end;
1233static int __meminitdata node_start;
1234
1235static int __meminit vmemmap_populate_hugepages(unsigned long start,
1236 unsigned long end, int node, struct vmem_altmap *altmap)
1237{
1238 unsigned long addr;
1239 unsigned long next;
1240 pgd_t *pgd;
1241 pud_t *pud;
1242 pmd_t *pmd;
1243
1244 for (addr = start; addr < end; addr = next) {
1245 next = pmd_addr_end(addr, end);
1246
1247 pgd = vmemmap_pgd_populate(addr, node);
1248 if (!pgd)
1249 return -ENOMEM;
1250
1251 pud = vmemmap_pud_populate(pgd, addr, node);
1252 if (!pud)
1253 return -ENOMEM;
1254
1255 pmd = pmd_offset(pud, addr);
1256 if (pmd_none(*pmd)) {
1257 void *p;
1258
1259 p = __vmemmap_alloc_block_buf(PMD_SIZE, node, altmap);
1260 if (p) {
1261 pte_t entry;
1262
1263 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1264 PAGE_KERNEL_LARGE);
1265 set_pmd(pmd, __pmd(pte_val(entry)));
1266
1267 /* check to see if we have contiguous blocks */
1268 if (p_end != p || node_start != node) {
1269 if (p_start)
1270 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1271 addr_start, addr_end-1, p_start, p_end-1, node_start);
1272 addr_start = addr;
1273 node_start = node;
1274 p_start = p;
1275 }
1276
1277 addr_end = addr + PMD_SIZE;
1278 p_end = p + PMD_SIZE;
1279 continue;
1280 } else if (altmap)
1281 return -ENOMEM; /* no fallback */
1282 } else if (pmd_large(*pmd)) {
1283 vmemmap_verify((pte_t *)pmd, node, addr, next);
1284 continue;
1285 }
1286 pr_warn_once("vmemmap: falling back to regular page backing\n");
1287 if (vmemmap_populate_basepages(addr, next, node))
1288 return -ENOMEM;
1289 }
1290 return 0;
1291}
1292
1293int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
1294{
1295 struct vmem_altmap *altmap = to_vmem_altmap(start);
1296 int err;
1297
1298 if (cpu_has_pse)
1299 err = vmemmap_populate_hugepages(start, end, node, altmap);
1300 else if (altmap) {
1301 pr_err_once("%s: no cpu support for altmap allocations\n",
1302 __func__);
1303 err = -ENOMEM;
1304 } else
1305 err = vmemmap_populate_basepages(start, end, node);
1306 if (!err)
1307 sync_global_pgds(start, end - 1, 0);
1308 return err;
1309}
1310
1311#if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1312void register_page_bootmem_memmap(unsigned long section_nr,
1313 struct page *start_page, unsigned long size)
1314{
1315 unsigned long addr = (unsigned long)start_page;
1316 unsigned long end = (unsigned long)(start_page + size);
1317 unsigned long next;
1318 pgd_t *pgd;
1319 pud_t *pud;
1320 pmd_t *pmd;
1321 unsigned int nr_pages;
1322 struct page *page;
1323
1324 for (; addr < end; addr = next) {
1325 pte_t *pte = NULL;
1326
1327 pgd = pgd_offset_k(addr);
1328 if (pgd_none(*pgd)) {
1329 next = (addr + PAGE_SIZE) & PAGE_MASK;
1330 continue;
1331 }
1332 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1333
1334 pud = pud_offset(pgd, addr);
1335 if (pud_none(*pud)) {
1336 next = (addr + PAGE_SIZE) & PAGE_MASK;
1337 continue;
1338 }
1339 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1340
1341 if (!cpu_has_pse) {
1342 next = (addr + PAGE_SIZE) & PAGE_MASK;
1343 pmd = pmd_offset(pud, addr);
1344 if (pmd_none(*pmd))
1345 continue;
1346 get_page_bootmem(section_nr, pmd_page(*pmd),
1347 MIX_SECTION_INFO);
1348
1349 pte = pte_offset_kernel(pmd, addr);
1350 if (pte_none(*pte))
1351 continue;
1352 get_page_bootmem(section_nr, pte_page(*pte),
1353 SECTION_INFO);
1354 } else {
1355 next = pmd_addr_end(addr, end);
1356
1357 pmd = pmd_offset(pud, addr);
1358 if (pmd_none(*pmd))
1359 continue;
1360
1361 nr_pages = 1 << (get_order(PMD_SIZE));
1362 page = pmd_page(*pmd);
1363 while (nr_pages--)
1364 get_page_bootmem(section_nr, page++,
1365 SECTION_INFO);
1366 }
1367 }
1368}
1369#endif
1370
1371void __meminit vmemmap_populate_print_last(void)
1372{
1373 if (p_start) {
1374 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1375 addr_start, addr_end-1, p_start, p_end-1, node_start);
1376 p_start = NULL;
1377 p_end = NULL;
1378 node_start = 0;
1379 }
1380}
1381#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/module.h>
31#include <linux/memory.h>
32#include <linux/memory_hotplug.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 <asm/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.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/cacheflush.h>
54#include <asm/init.h>
55#include <asm/uv/uv.h>
56#include <asm/setup.h>
57
58#include "mm_internal.h"
59
60static void ident_pmd_init(unsigned long pmd_flag, pmd_t *pmd_page,
61 unsigned long addr, unsigned long end)
62{
63 addr &= PMD_MASK;
64 for (; addr < end; addr += PMD_SIZE) {
65 pmd_t *pmd = pmd_page + pmd_index(addr);
66
67 if (!pmd_present(*pmd))
68 set_pmd(pmd, __pmd(addr | pmd_flag));
69 }
70}
71static int ident_pud_init(struct x86_mapping_info *info, pud_t *pud_page,
72 unsigned long addr, unsigned long end)
73{
74 unsigned long next;
75
76 for (; addr < end; addr = next) {
77 pud_t *pud = pud_page + pud_index(addr);
78 pmd_t *pmd;
79
80 next = (addr & PUD_MASK) + PUD_SIZE;
81 if (next > end)
82 next = end;
83
84 if (pud_present(*pud)) {
85 pmd = pmd_offset(pud, 0);
86 ident_pmd_init(info->pmd_flag, pmd, addr, next);
87 continue;
88 }
89 pmd = (pmd_t *)info->alloc_pgt_page(info->context);
90 if (!pmd)
91 return -ENOMEM;
92 ident_pmd_init(info->pmd_flag, pmd, addr, next);
93 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
94 }
95
96 return 0;
97}
98
99int kernel_ident_mapping_init(struct x86_mapping_info *info, pgd_t *pgd_page,
100 unsigned long addr, unsigned long end)
101{
102 unsigned long next;
103 int result;
104 int off = info->kernel_mapping ? pgd_index(__PAGE_OFFSET) : 0;
105
106 for (; addr < end; addr = next) {
107 pgd_t *pgd = pgd_page + pgd_index(addr) + off;
108 pud_t *pud;
109
110 next = (addr & PGDIR_MASK) + PGDIR_SIZE;
111 if (next > end)
112 next = end;
113
114 if (pgd_present(*pgd)) {
115 pud = pud_offset(pgd, 0);
116 result = ident_pud_init(info, pud, addr, next);
117 if (result)
118 return result;
119 continue;
120 }
121
122 pud = (pud_t *)info->alloc_pgt_page(info->context);
123 if (!pud)
124 return -ENOMEM;
125 result = ident_pud_init(info, pud, addr, next);
126 if (result)
127 return result;
128 set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE));
129 }
130
131 return 0;
132}
133
134static int __init parse_direct_gbpages_off(char *arg)
135{
136 direct_gbpages = 0;
137 return 0;
138}
139early_param("nogbpages", parse_direct_gbpages_off);
140
141static int __init parse_direct_gbpages_on(char *arg)
142{
143 direct_gbpages = 1;
144 return 0;
145}
146early_param("gbpages", parse_direct_gbpages_on);
147
148/*
149 * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
150 * physical space so we can cache the place of the first one and move
151 * around without checking the pgd every time.
152 */
153
154pteval_t __supported_pte_mask __read_mostly = ~_PAGE_IOMAP;
155EXPORT_SYMBOL_GPL(__supported_pte_mask);
156
157int force_personality32;
158
159/*
160 * noexec32=on|off
161 * Control non executable heap for 32bit processes.
162 * To control the stack too use noexec=off
163 *
164 * on PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
165 * off PROT_READ implies PROT_EXEC
166 */
167static int __init nonx32_setup(char *str)
168{
169 if (!strcmp(str, "on"))
170 force_personality32 &= ~READ_IMPLIES_EXEC;
171 else if (!strcmp(str, "off"))
172 force_personality32 |= READ_IMPLIES_EXEC;
173 return 1;
174}
175__setup("noexec32=", nonx32_setup);
176
177/*
178 * When memory was added/removed 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 unsigned long address;
184
185 for (address = start; address <= end; address += PGDIR_SIZE) {
186 const pgd_t *pgd_ref = pgd_offset_k(address);
187 struct page *page;
188
189 if (pgd_none(*pgd_ref))
190 continue;
191
192 spin_lock(&pgd_lock);
193 list_for_each_entry(page, &pgd_list, lru) {
194 pgd_t *pgd;
195 spinlock_t *pgt_lock;
196
197 pgd = (pgd_t *)page_address(page) + pgd_index(address);
198 /* the pgt_lock only for Xen */
199 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
200 spin_lock(pgt_lock);
201
202 if (pgd_none(*pgd))
203 set_pgd(pgd, *pgd_ref);
204 else
205 BUG_ON(pgd_page_vaddr(*pgd)
206 != pgd_page_vaddr(*pgd_ref));
207
208 spin_unlock(pgt_lock);
209 }
210 spin_unlock(&pgd_lock);
211 }
212}
213
214/*
215 * NOTE: This function is marked __ref because it calls __init function
216 * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
217 */
218static __ref void *spp_getpage(void)
219{
220 void *ptr;
221
222 if (after_bootmem)
223 ptr = (void *) get_zeroed_page(GFP_ATOMIC | __GFP_NOTRACK);
224 else
225 ptr = alloc_bootmem_pages(PAGE_SIZE);
226
227 if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
228 panic("set_pte_phys: cannot allocate page data %s\n",
229 after_bootmem ? "after bootmem" : "");
230 }
231
232 pr_debug("spp_getpage %p\n", ptr);
233
234 return ptr;
235}
236
237static pud_t *fill_pud(pgd_t *pgd, unsigned long vaddr)
238{
239 if (pgd_none(*pgd)) {
240 pud_t *pud = (pud_t *)spp_getpage();
241 pgd_populate(&init_mm, pgd, pud);
242 if (pud != pud_offset(pgd, 0))
243 printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
244 pud, pud_offset(pgd, 0));
245 }
246 return pud_offset(pgd, vaddr);
247}
248
249static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
250{
251 if (pud_none(*pud)) {
252 pmd_t *pmd = (pmd_t *) spp_getpage();
253 pud_populate(&init_mm, pud, pmd);
254 if (pmd != pmd_offset(pud, 0))
255 printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
256 pmd, pmd_offset(pud, 0));
257 }
258 return pmd_offset(pud, vaddr);
259}
260
261static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
262{
263 if (pmd_none(*pmd)) {
264 pte_t *pte = (pte_t *) spp_getpage();
265 pmd_populate_kernel(&init_mm, pmd, pte);
266 if (pte != pte_offset_kernel(pmd, 0))
267 printk(KERN_ERR "PAGETABLE BUG #02!\n");
268 }
269 return pte_offset_kernel(pmd, vaddr);
270}
271
272void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
273{
274 pud_t *pud;
275 pmd_t *pmd;
276 pte_t *pte;
277
278 pud = pud_page + pud_index(vaddr);
279 pmd = fill_pmd(pud, vaddr);
280 pte = fill_pte(pmd, vaddr);
281
282 set_pte(pte, new_pte);
283
284 /*
285 * It's enough to flush this one mapping.
286 * (PGE mappings get flushed as well)
287 */
288 __flush_tlb_one(vaddr);
289}
290
291void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
292{
293 pgd_t *pgd;
294 pud_t *pud_page;
295
296 pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
297
298 pgd = pgd_offset_k(vaddr);
299 if (pgd_none(*pgd)) {
300 printk(KERN_ERR
301 "PGD FIXMAP MISSING, it should be setup in head.S!\n");
302 return;
303 }
304 pud_page = (pud_t*)pgd_page_vaddr(*pgd);
305 set_pte_vaddr_pud(pud_page, vaddr, pteval);
306}
307
308pmd_t * __init populate_extra_pmd(unsigned long vaddr)
309{
310 pgd_t *pgd;
311 pud_t *pud;
312
313 pgd = pgd_offset_k(vaddr);
314 pud = fill_pud(pgd, 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 pgprot_t prot)
331{
332 pgd_t *pgd;
333 pud_t *pud;
334 pmd_t *pmd;
335
336 BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
337 for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
338 pgd = pgd_offset_k((unsigned long)__va(phys));
339 if (pgd_none(*pgd)) {
340 pud = (pud_t *) spp_getpage();
341 set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
342 _PAGE_USER));
343 }
344 pud = pud_offset(pgd, (unsigned long)__va(phys));
345 if (pud_none(*pud)) {
346 pmd = (pmd_t *) spp_getpage();
347 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
348 _PAGE_USER));
349 }
350 pmd = pmd_offset(pud, phys);
351 BUG_ON(!pmd_none(*pmd));
352 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
353 }
354}
355
356void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
357{
358 __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
359}
360
361void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
362{
363 __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
364}
365
366/*
367 * The head.S code sets up the kernel high mapping:
368 *
369 * from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
370 *
371 * phys_base holds the negative offset to the kernel, which is added
372 * to the compile time generated pmds. This results in invalid pmds up
373 * to the point where we hit the physaddr 0 mapping.
374 *
375 * We limit the mappings to the region from _text to _brk_end. _brk_end
376 * is rounded up to the 2MB boundary. This catches the invalid pmds as
377 * well, as they are located before _text:
378 */
379void __init cleanup_highmap(void)
380{
381 unsigned long vaddr = __START_KERNEL_map;
382 unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
383 unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
384 pmd_t *pmd = level2_kernel_pgt;
385
386 /*
387 * Native path, max_pfn_mapped is not set yet.
388 * Xen has valid max_pfn_mapped set in
389 * arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
390 */
391 if (max_pfn_mapped)
392 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
393
394 for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
395 if (pmd_none(*pmd))
396 continue;
397 if (vaddr < (unsigned long) _text || vaddr > end)
398 set_pmd(pmd, __pmd(0));
399 }
400}
401
402static unsigned long __meminit
403phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
404 pgprot_t prot)
405{
406 unsigned long pages = 0, next;
407 unsigned long last_map_addr = end;
408 int i;
409
410 pte_t *pte = pte_page + pte_index(addr);
411
412 for (i = pte_index(addr); i < PTRS_PER_PTE; i++, addr = next, pte++) {
413 next = (addr & PAGE_MASK) + PAGE_SIZE;
414 if (addr >= end) {
415 if (!after_bootmem &&
416 !e820_any_mapped(addr & PAGE_MASK, next, E820_RAM) &&
417 !e820_any_mapped(addr & PAGE_MASK, next, E820_RESERVED_KERN))
418 set_pte(pte, __pte(0));
419 continue;
420 }
421
422 /*
423 * We will re-use the existing mapping.
424 * Xen for example has some special requirements, like mapping
425 * pagetable pages as RO. So assume someone who pre-setup
426 * these mappings are more intelligent.
427 */
428 if (pte_val(*pte)) {
429 if (!after_bootmem)
430 pages++;
431 continue;
432 }
433
434 if (0)
435 printk(" pte=%p addr=%lx pte=%016lx\n",
436 pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
437 pages++;
438 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
439 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
440 }
441
442 update_page_count(PG_LEVEL_4K, pages);
443
444 return last_map_addr;
445}
446
447static unsigned long __meminit
448phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
449 unsigned long page_size_mask, pgprot_t prot)
450{
451 unsigned long pages = 0, next;
452 unsigned long last_map_addr = end;
453
454 int i = pmd_index(address);
455
456 for (; i < PTRS_PER_PMD; i++, address = next) {
457 pmd_t *pmd = pmd_page + pmd_index(address);
458 pte_t *pte;
459 pgprot_t new_prot = prot;
460
461 next = (address & PMD_MASK) + PMD_SIZE;
462 if (address >= end) {
463 if (!after_bootmem &&
464 !e820_any_mapped(address & PMD_MASK, next, E820_RAM) &&
465 !e820_any_mapped(address & PMD_MASK, next, E820_RESERVED_KERN))
466 set_pmd(pmd, __pmd(0));
467 continue;
468 }
469
470 if (pmd_val(*pmd)) {
471 if (!pmd_large(*pmd)) {
472 spin_lock(&init_mm.page_table_lock);
473 pte = (pte_t *)pmd_page_vaddr(*pmd);
474 last_map_addr = phys_pte_init(pte, address,
475 end, prot);
476 spin_unlock(&init_mm.page_table_lock);
477 continue;
478 }
479 /*
480 * If we are ok with PG_LEVEL_2M mapping, then we will
481 * use the existing mapping,
482 *
483 * Otherwise, we will split the large page mapping but
484 * use the same existing protection bits except for
485 * large page, so that we don't violate Intel's TLB
486 * Application note (317080) which says, while changing
487 * the page sizes, new and old translations should
488 * not differ with respect to page frame and
489 * attributes.
490 */
491 if (page_size_mask & (1 << PG_LEVEL_2M)) {
492 if (!after_bootmem)
493 pages++;
494 last_map_addr = next;
495 continue;
496 }
497 new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
498 }
499
500 if (page_size_mask & (1<<PG_LEVEL_2M)) {
501 pages++;
502 spin_lock(&init_mm.page_table_lock);
503 set_pte((pte_t *)pmd,
504 pfn_pte((address & PMD_MASK) >> PAGE_SHIFT,
505 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
506 spin_unlock(&init_mm.page_table_lock);
507 last_map_addr = next;
508 continue;
509 }
510
511 pte = alloc_low_page();
512 last_map_addr = phys_pte_init(pte, address, end, new_prot);
513
514 spin_lock(&init_mm.page_table_lock);
515 pmd_populate_kernel(&init_mm, pmd, pte);
516 spin_unlock(&init_mm.page_table_lock);
517 }
518 update_page_count(PG_LEVEL_2M, pages);
519 return last_map_addr;
520}
521
522static unsigned long __meminit
523phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
524 unsigned long page_size_mask)
525{
526 unsigned long pages = 0, next;
527 unsigned long last_map_addr = end;
528 int i = pud_index(addr);
529
530 for (; i < PTRS_PER_PUD; i++, addr = next) {
531 pud_t *pud = pud_page + pud_index(addr);
532 pmd_t *pmd;
533 pgprot_t prot = PAGE_KERNEL;
534
535 next = (addr & PUD_MASK) + PUD_SIZE;
536 if (addr >= end) {
537 if (!after_bootmem &&
538 !e820_any_mapped(addr & PUD_MASK, next, E820_RAM) &&
539 !e820_any_mapped(addr & PUD_MASK, next, E820_RESERVED_KERN))
540 set_pud(pud, __pud(0));
541 continue;
542 }
543
544 if (pud_val(*pud)) {
545 if (!pud_large(*pud)) {
546 pmd = pmd_offset(pud, 0);
547 last_map_addr = phys_pmd_init(pmd, addr, end,
548 page_size_mask, prot);
549 __flush_tlb_all();
550 continue;
551 }
552 /*
553 * If we are ok with PG_LEVEL_1G mapping, then we will
554 * use the existing mapping.
555 *
556 * Otherwise, we will split the gbpage mapping but use
557 * the same existing protection bits except for large
558 * page, so that we don't violate Intel's TLB
559 * Application note (317080) which says, while changing
560 * the page sizes, new and old translations should
561 * not differ with respect to page frame and
562 * attributes.
563 */
564 if (page_size_mask & (1 << PG_LEVEL_1G)) {
565 if (!after_bootmem)
566 pages++;
567 last_map_addr = next;
568 continue;
569 }
570 prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
571 }
572
573 if (page_size_mask & (1<<PG_LEVEL_1G)) {
574 pages++;
575 spin_lock(&init_mm.page_table_lock);
576 set_pte((pte_t *)pud,
577 pfn_pte((addr & PUD_MASK) >> PAGE_SHIFT,
578 PAGE_KERNEL_LARGE));
579 spin_unlock(&init_mm.page_table_lock);
580 last_map_addr = next;
581 continue;
582 }
583
584 pmd = alloc_low_page();
585 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
586 prot);
587
588 spin_lock(&init_mm.page_table_lock);
589 pud_populate(&init_mm, pud, pmd);
590 spin_unlock(&init_mm.page_table_lock);
591 }
592 __flush_tlb_all();
593
594 update_page_count(PG_LEVEL_1G, pages);
595
596 return last_map_addr;
597}
598
599unsigned long __meminit
600kernel_physical_mapping_init(unsigned long start,
601 unsigned long end,
602 unsigned long page_size_mask)
603{
604 bool pgd_changed = false;
605 unsigned long next, last_map_addr = end;
606 unsigned long addr;
607
608 start = (unsigned long)__va(start);
609 end = (unsigned long)__va(end);
610 addr = start;
611
612 for (; start < end; start = next) {
613 pgd_t *pgd = pgd_offset_k(start);
614 pud_t *pud;
615
616 next = (start & PGDIR_MASK) + PGDIR_SIZE;
617
618 if (pgd_val(*pgd)) {
619 pud = (pud_t *)pgd_page_vaddr(*pgd);
620 last_map_addr = phys_pud_init(pud, __pa(start),
621 __pa(end), page_size_mask);
622 continue;
623 }
624
625 pud = alloc_low_page();
626 last_map_addr = phys_pud_init(pud, __pa(start), __pa(end),
627 page_size_mask);
628
629 spin_lock(&init_mm.page_table_lock);
630 pgd_populate(&init_mm, pgd, pud);
631 spin_unlock(&init_mm.page_table_lock);
632 pgd_changed = true;
633 }
634
635 if (pgd_changed)
636 sync_global_pgds(addr, end - 1);
637
638 __flush_tlb_all();
639
640 return last_map_addr;
641}
642
643#ifndef CONFIG_NUMA
644void __init initmem_init(void)
645{
646 memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
647}
648#endif
649
650void __init paging_init(void)
651{
652 sparse_memory_present_with_active_regions(MAX_NUMNODES);
653 sparse_init();
654
655 /*
656 * clear the default setting with node 0
657 * note: don't use nodes_clear here, that is really clearing when
658 * numa support is not compiled in, and later node_set_state
659 * will not set it back.
660 */
661 node_clear_state(0, N_MEMORY);
662 if (N_MEMORY != N_NORMAL_MEMORY)
663 node_clear_state(0, N_NORMAL_MEMORY);
664
665 zone_sizes_init();
666}
667
668/*
669 * Memory hotplug specific functions
670 */
671#ifdef CONFIG_MEMORY_HOTPLUG
672/*
673 * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
674 * updating.
675 */
676static void update_end_of_memory_vars(u64 start, u64 size)
677{
678 unsigned long end_pfn = PFN_UP(start + size);
679
680 if (end_pfn > max_pfn) {
681 max_pfn = end_pfn;
682 max_low_pfn = end_pfn;
683 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
684 }
685}
686
687/*
688 * Memory is added always to NORMAL zone. This means you will never get
689 * additional DMA/DMA32 memory.
690 */
691int arch_add_memory(int nid, u64 start, u64 size)
692{
693 struct pglist_data *pgdat = NODE_DATA(nid);
694 struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
695 unsigned long start_pfn = start >> PAGE_SHIFT;
696 unsigned long nr_pages = size >> PAGE_SHIFT;
697 int ret;
698
699 init_memory_mapping(start, start + size);
700
701 ret = __add_pages(nid, zone, start_pfn, nr_pages);
702 WARN_ON_ONCE(ret);
703
704 /* update max_pfn, max_low_pfn and high_memory */
705 update_end_of_memory_vars(start, size);
706
707 return ret;
708}
709EXPORT_SYMBOL_GPL(arch_add_memory);
710
711#define PAGE_INUSE 0xFD
712
713static void __meminit free_pagetable(struct page *page, int order)
714{
715 unsigned long magic;
716 unsigned int nr_pages = 1 << order;
717
718 /* bootmem page has reserved flag */
719 if (PageReserved(page)) {
720 __ClearPageReserved(page);
721
722 magic = (unsigned long)page->lru.next;
723 if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
724 while (nr_pages--)
725 put_page_bootmem(page++);
726 } else
727 while (nr_pages--)
728 free_reserved_page(page++);
729 } else
730 free_pages((unsigned long)page_address(page), order);
731}
732
733static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
734{
735 pte_t *pte;
736 int i;
737
738 for (i = 0; i < PTRS_PER_PTE; i++) {
739 pte = pte_start + i;
740 if (pte_val(*pte))
741 return;
742 }
743
744 /* free a pte talbe */
745 free_pagetable(pmd_page(*pmd), 0);
746 spin_lock(&init_mm.page_table_lock);
747 pmd_clear(pmd);
748 spin_unlock(&init_mm.page_table_lock);
749}
750
751static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
752{
753 pmd_t *pmd;
754 int i;
755
756 for (i = 0; i < PTRS_PER_PMD; i++) {
757 pmd = pmd_start + i;
758 if (pmd_val(*pmd))
759 return;
760 }
761
762 /* free a pmd talbe */
763 free_pagetable(pud_page(*pud), 0);
764 spin_lock(&init_mm.page_table_lock);
765 pud_clear(pud);
766 spin_unlock(&init_mm.page_table_lock);
767}
768
769/* Return true if pgd is changed, otherwise return false. */
770static bool __meminit free_pud_table(pud_t *pud_start, pgd_t *pgd)
771{
772 pud_t *pud;
773 int i;
774
775 for (i = 0; i < PTRS_PER_PUD; i++) {
776 pud = pud_start + i;
777 if (pud_val(*pud))
778 return false;
779 }
780
781 /* free a pud table */
782 free_pagetable(pgd_page(*pgd), 0);
783 spin_lock(&init_mm.page_table_lock);
784 pgd_clear(pgd);
785 spin_unlock(&init_mm.page_table_lock);
786
787 return true;
788}
789
790static void __meminit
791remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
792 bool direct)
793{
794 unsigned long next, pages = 0;
795 pte_t *pte;
796 void *page_addr;
797 phys_addr_t phys_addr;
798
799 pte = pte_start + pte_index(addr);
800 for (; addr < end; addr = next, pte++) {
801 next = (addr + PAGE_SIZE) & PAGE_MASK;
802 if (next > end)
803 next = end;
804
805 if (!pte_present(*pte))
806 continue;
807
808 /*
809 * We mapped [0,1G) memory as identity mapping when
810 * initializing, in arch/x86/kernel/head_64.S. These
811 * pagetables cannot be removed.
812 */
813 phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
814 if (phys_addr < (phys_addr_t)0x40000000)
815 return;
816
817 if (IS_ALIGNED(addr, PAGE_SIZE) &&
818 IS_ALIGNED(next, PAGE_SIZE)) {
819 /*
820 * Do not free direct mapping pages since they were
821 * freed when offlining, or simplely not in use.
822 */
823 if (!direct)
824 free_pagetable(pte_page(*pte), 0);
825
826 spin_lock(&init_mm.page_table_lock);
827 pte_clear(&init_mm, addr, pte);
828 spin_unlock(&init_mm.page_table_lock);
829
830 /* For non-direct mapping, pages means nothing. */
831 pages++;
832 } else {
833 /*
834 * If we are here, we are freeing vmemmap pages since
835 * direct mapped memory ranges to be freed are aligned.
836 *
837 * If we are not removing the whole page, it means
838 * other page structs in this page are being used and
839 * we canot remove them. So fill the unused page_structs
840 * with 0xFD, and remove the page when it is wholly
841 * filled with 0xFD.
842 */
843 memset((void *)addr, PAGE_INUSE, next - addr);
844
845 page_addr = page_address(pte_page(*pte));
846 if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) {
847 free_pagetable(pte_page(*pte), 0);
848
849 spin_lock(&init_mm.page_table_lock);
850 pte_clear(&init_mm, addr, pte);
851 spin_unlock(&init_mm.page_table_lock);
852 }
853 }
854 }
855
856 /* Call free_pte_table() in remove_pmd_table(). */
857 flush_tlb_all();
858 if (direct)
859 update_page_count(PG_LEVEL_4K, -pages);
860}
861
862static void __meminit
863remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
864 bool direct)
865{
866 unsigned long next, pages = 0;
867 pte_t *pte_base;
868 pmd_t *pmd;
869 void *page_addr;
870
871 pmd = pmd_start + pmd_index(addr);
872 for (; addr < end; addr = next, pmd++) {
873 next = pmd_addr_end(addr, end);
874
875 if (!pmd_present(*pmd))
876 continue;
877
878 if (pmd_large(*pmd)) {
879 if (IS_ALIGNED(addr, PMD_SIZE) &&
880 IS_ALIGNED(next, PMD_SIZE)) {
881 if (!direct)
882 free_pagetable(pmd_page(*pmd),
883 get_order(PMD_SIZE));
884
885 spin_lock(&init_mm.page_table_lock);
886 pmd_clear(pmd);
887 spin_unlock(&init_mm.page_table_lock);
888 pages++;
889 } else {
890 /* If here, we are freeing vmemmap pages. */
891 memset((void *)addr, PAGE_INUSE, next - addr);
892
893 page_addr = page_address(pmd_page(*pmd));
894 if (!memchr_inv(page_addr, PAGE_INUSE,
895 PMD_SIZE)) {
896 free_pagetable(pmd_page(*pmd),
897 get_order(PMD_SIZE));
898
899 spin_lock(&init_mm.page_table_lock);
900 pmd_clear(pmd);
901 spin_unlock(&init_mm.page_table_lock);
902 }
903 }
904
905 continue;
906 }
907
908 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
909 remove_pte_table(pte_base, addr, next, direct);
910 free_pte_table(pte_base, pmd);
911 }
912
913 /* Call free_pmd_table() in remove_pud_table(). */
914 if (direct)
915 update_page_count(PG_LEVEL_2M, -pages);
916}
917
918static void __meminit
919remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
920 bool direct)
921{
922 unsigned long next, pages = 0;
923 pmd_t *pmd_base;
924 pud_t *pud;
925 void *page_addr;
926
927 pud = pud_start + pud_index(addr);
928 for (; addr < end; addr = next, pud++) {
929 next = pud_addr_end(addr, end);
930
931 if (!pud_present(*pud))
932 continue;
933
934 if (pud_large(*pud)) {
935 if (IS_ALIGNED(addr, PUD_SIZE) &&
936 IS_ALIGNED(next, PUD_SIZE)) {
937 if (!direct)
938 free_pagetable(pud_page(*pud),
939 get_order(PUD_SIZE));
940
941 spin_lock(&init_mm.page_table_lock);
942 pud_clear(pud);
943 spin_unlock(&init_mm.page_table_lock);
944 pages++;
945 } else {
946 /* If here, we are freeing vmemmap pages. */
947 memset((void *)addr, PAGE_INUSE, next - addr);
948
949 page_addr = page_address(pud_page(*pud));
950 if (!memchr_inv(page_addr, PAGE_INUSE,
951 PUD_SIZE)) {
952 free_pagetable(pud_page(*pud),
953 get_order(PUD_SIZE));
954
955 spin_lock(&init_mm.page_table_lock);
956 pud_clear(pud);
957 spin_unlock(&init_mm.page_table_lock);
958 }
959 }
960
961 continue;
962 }
963
964 pmd_base = (pmd_t *)pud_page_vaddr(*pud);
965 remove_pmd_table(pmd_base, addr, next, direct);
966 free_pmd_table(pmd_base, pud);
967 }
968
969 if (direct)
970 update_page_count(PG_LEVEL_1G, -pages);
971}
972
973/* start and end are both virtual address. */
974static void __meminit
975remove_pagetable(unsigned long start, unsigned long end, bool direct)
976{
977 unsigned long next;
978 pgd_t *pgd;
979 pud_t *pud;
980 bool pgd_changed = false;
981
982 for (; start < end; start = next) {
983 next = pgd_addr_end(start, end);
984
985 pgd = pgd_offset_k(start);
986 if (!pgd_present(*pgd))
987 continue;
988
989 pud = (pud_t *)pgd_page_vaddr(*pgd);
990 remove_pud_table(pud, start, next, direct);
991 if (free_pud_table(pud, pgd))
992 pgd_changed = true;
993 }
994
995 if (pgd_changed)
996 sync_global_pgds(start, end - 1);
997
998 flush_tlb_all();
999}
1000
1001void __ref vmemmap_free(unsigned long start, unsigned long end)
1002{
1003 remove_pagetable(start, end, false);
1004}
1005
1006#ifdef CONFIG_MEMORY_HOTREMOVE
1007static void __meminit
1008kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1009{
1010 start = (unsigned long)__va(start);
1011 end = (unsigned long)__va(end);
1012
1013 remove_pagetable(start, end, true);
1014}
1015
1016int __ref arch_remove_memory(u64 start, u64 size)
1017{
1018 unsigned long start_pfn = start >> PAGE_SHIFT;
1019 unsigned long nr_pages = size >> PAGE_SHIFT;
1020 struct zone *zone;
1021 int ret;
1022
1023 zone = page_zone(pfn_to_page(start_pfn));
1024 kernel_physical_mapping_remove(start, start + size);
1025 ret = __remove_pages(zone, start_pfn, nr_pages);
1026 WARN_ON_ONCE(ret);
1027
1028 return ret;
1029}
1030#endif
1031#endif /* CONFIG_MEMORY_HOTPLUG */
1032
1033static struct kcore_list kcore_vsyscall;
1034
1035static void __init register_page_bootmem_info(void)
1036{
1037#ifdef CONFIG_NUMA
1038 int i;
1039
1040 for_each_online_node(i)
1041 register_page_bootmem_info_node(NODE_DATA(i));
1042#endif
1043}
1044
1045void __init mem_init(void)
1046{
1047 pci_iommu_alloc();
1048
1049 /* clear_bss() already clear the empty_zero_page */
1050
1051 register_page_bootmem_info();
1052
1053 /* this will put all memory onto the freelists */
1054 free_all_bootmem();
1055 after_bootmem = 1;
1056
1057 /* Register memory areas for /proc/kcore */
1058 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
1059 VSYSCALL_END - VSYSCALL_START, KCORE_OTHER);
1060
1061 mem_init_print_info(NULL);
1062}
1063
1064#ifdef CONFIG_DEBUG_RODATA
1065const int rodata_test_data = 0xC3;
1066EXPORT_SYMBOL_GPL(rodata_test_data);
1067
1068int kernel_set_to_readonly;
1069
1070void set_kernel_text_rw(void)
1071{
1072 unsigned long start = PFN_ALIGN(_text);
1073 unsigned long end = PFN_ALIGN(__stop___ex_table);
1074
1075 if (!kernel_set_to_readonly)
1076 return;
1077
1078 pr_debug("Set kernel text: %lx - %lx for read write\n",
1079 start, end);
1080
1081 /*
1082 * Make the kernel identity mapping for text RW. Kernel text
1083 * mapping will always be RO. Refer to the comment in
1084 * static_protections() in pageattr.c
1085 */
1086 set_memory_rw(start, (end - start) >> PAGE_SHIFT);
1087}
1088
1089void set_kernel_text_ro(void)
1090{
1091 unsigned long start = PFN_ALIGN(_text);
1092 unsigned long end = PFN_ALIGN(__stop___ex_table);
1093
1094 if (!kernel_set_to_readonly)
1095 return;
1096
1097 pr_debug("Set kernel text: %lx - %lx for read only\n",
1098 start, end);
1099
1100 /*
1101 * Set the kernel identity mapping for text RO.
1102 */
1103 set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1104}
1105
1106void mark_rodata_ro(void)
1107{
1108 unsigned long start = PFN_ALIGN(_text);
1109 unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1110 unsigned long end = (unsigned long) &__end_rodata_hpage_align;
1111 unsigned long text_end = PFN_ALIGN(&__stop___ex_table);
1112 unsigned long rodata_end = PFN_ALIGN(&__end_rodata);
1113 unsigned long all_end = PFN_ALIGN(&_end);
1114
1115 printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1116 (end - start) >> 10);
1117 set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1118
1119 kernel_set_to_readonly = 1;
1120
1121 /*
1122 * The rodata/data/bss/brk section (but not the kernel text!)
1123 * should also be not-executable.
1124 */
1125 set_memory_nx(rodata_start, (all_end - rodata_start) >> PAGE_SHIFT);
1126
1127 rodata_test();
1128
1129#ifdef CONFIG_CPA_DEBUG
1130 printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1131 set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1132
1133 printk(KERN_INFO "Testing CPA: again\n");
1134 set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1135#endif
1136
1137 free_init_pages("unused kernel",
1138 (unsigned long) __va(__pa_symbol(text_end)),
1139 (unsigned long) __va(__pa_symbol(rodata_start)));
1140 free_init_pages("unused kernel",
1141 (unsigned long) __va(__pa_symbol(rodata_end)),
1142 (unsigned long) __va(__pa_symbol(_sdata)));
1143}
1144
1145#endif
1146
1147int kern_addr_valid(unsigned long addr)
1148{
1149 unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1150 pgd_t *pgd;
1151 pud_t *pud;
1152 pmd_t *pmd;
1153 pte_t *pte;
1154
1155 if (above != 0 && above != -1UL)
1156 return 0;
1157
1158 pgd = pgd_offset_k(addr);
1159 if (pgd_none(*pgd))
1160 return 0;
1161
1162 pud = pud_offset(pgd, addr);
1163 if (pud_none(*pud))
1164 return 0;
1165
1166 if (pud_large(*pud))
1167 return pfn_valid(pud_pfn(*pud));
1168
1169 pmd = pmd_offset(pud, addr);
1170 if (pmd_none(*pmd))
1171 return 0;
1172
1173 if (pmd_large(*pmd))
1174 return pfn_valid(pmd_pfn(*pmd));
1175
1176 pte = pte_offset_kernel(pmd, addr);
1177 if (pte_none(*pte))
1178 return 0;
1179
1180 return pfn_valid(pte_pfn(*pte));
1181}
1182
1183/*
1184 * A pseudo VMA to allow ptrace access for the vsyscall page. This only
1185 * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1186 * not need special handling anymore:
1187 */
1188static struct vm_area_struct gate_vma = {
1189 .vm_start = VSYSCALL_START,
1190 .vm_end = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
1191 .vm_page_prot = PAGE_READONLY_EXEC,
1192 .vm_flags = VM_READ | VM_EXEC
1193};
1194
1195struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
1196{
1197#ifdef CONFIG_IA32_EMULATION
1198 if (!mm || mm->context.ia32_compat)
1199 return NULL;
1200#endif
1201 return &gate_vma;
1202}
1203
1204int in_gate_area(struct mm_struct *mm, unsigned long addr)
1205{
1206 struct vm_area_struct *vma = get_gate_vma(mm);
1207
1208 if (!vma)
1209 return 0;
1210
1211 return (addr >= vma->vm_start) && (addr < vma->vm_end);
1212}
1213
1214/*
1215 * Use this when you have no reliable mm, typically from interrupt
1216 * context. It is less reliable than using a task's mm and may give
1217 * false positives.
1218 */
1219int in_gate_area_no_mm(unsigned long addr)
1220{
1221 return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1222}
1223
1224const char *arch_vma_name(struct vm_area_struct *vma)
1225{
1226 if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1227 return "[vdso]";
1228 if (vma == &gate_vma)
1229 return "[vsyscall]";
1230 return NULL;
1231}
1232
1233#ifdef CONFIG_X86_UV
1234unsigned long memory_block_size_bytes(void)
1235{
1236 if (is_uv_system()) {
1237 printk(KERN_INFO "UV: memory block size 2GB\n");
1238 return 2UL * 1024 * 1024 * 1024;
1239 }
1240 return MIN_MEMORY_BLOCK_SIZE;
1241}
1242#endif
1243
1244#ifdef CONFIG_SPARSEMEM_VMEMMAP
1245/*
1246 * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1247 */
1248static long __meminitdata addr_start, addr_end;
1249static void __meminitdata *p_start, *p_end;
1250static int __meminitdata node_start;
1251
1252static int __meminit vmemmap_populate_hugepages(unsigned long start,
1253 unsigned long end, int node)
1254{
1255 unsigned long addr;
1256 unsigned long next;
1257 pgd_t *pgd;
1258 pud_t *pud;
1259 pmd_t *pmd;
1260
1261 for (addr = start; addr < end; addr = next) {
1262 next = pmd_addr_end(addr, end);
1263
1264 pgd = vmemmap_pgd_populate(addr, node);
1265 if (!pgd)
1266 return -ENOMEM;
1267
1268 pud = vmemmap_pud_populate(pgd, addr, node);
1269 if (!pud)
1270 return -ENOMEM;
1271
1272 pmd = pmd_offset(pud, addr);
1273 if (pmd_none(*pmd)) {
1274 void *p;
1275
1276 p = vmemmap_alloc_block_buf(PMD_SIZE, node);
1277 if (p) {
1278 pte_t entry;
1279
1280 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1281 PAGE_KERNEL_LARGE);
1282 set_pmd(pmd, __pmd(pte_val(entry)));
1283
1284 /* check to see if we have contiguous blocks */
1285 if (p_end != p || node_start != node) {
1286 if (p_start)
1287 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1288 addr_start, addr_end-1, p_start, p_end-1, node_start);
1289 addr_start = addr;
1290 node_start = node;
1291 p_start = p;
1292 }
1293
1294 addr_end = addr + PMD_SIZE;
1295 p_end = p + PMD_SIZE;
1296 continue;
1297 }
1298 } else if (pmd_large(*pmd)) {
1299 vmemmap_verify((pte_t *)pmd, node, addr, next);
1300 continue;
1301 }
1302 pr_warn_once("vmemmap: falling back to regular page backing\n");
1303 if (vmemmap_populate_basepages(addr, next, node))
1304 return -ENOMEM;
1305 }
1306 return 0;
1307}
1308
1309int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
1310{
1311 int err;
1312
1313 if (cpu_has_pse)
1314 err = vmemmap_populate_hugepages(start, end, node);
1315 else
1316 err = vmemmap_populate_basepages(start, end, node);
1317 if (!err)
1318 sync_global_pgds(start, end - 1);
1319 return err;
1320}
1321
1322#if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1323void register_page_bootmem_memmap(unsigned long section_nr,
1324 struct page *start_page, unsigned long size)
1325{
1326 unsigned long addr = (unsigned long)start_page;
1327 unsigned long end = (unsigned long)(start_page + size);
1328 unsigned long next;
1329 pgd_t *pgd;
1330 pud_t *pud;
1331 pmd_t *pmd;
1332 unsigned int nr_pages;
1333 struct page *page;
1334
1335 for (; addr < end; addr = next) {
1336 pte_t *pte = NULL;
1337
1338 pgd = pgd_offset_k(addr);
1339 if (pgd_none(*pgd)) {
1340 next = (addr + PAGE_SIZE) & PAGE_MASK;
1341 continue;
1342 }
1343 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1344
1345 pud = pud_offset(pgd, addr);
1346 if (pud_none(*pud)) {
1347 next = (addr + PAGE_SIZE) & PAGE_MASK;
1348 continue;
1349 }
1350 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1351
1352 if (!cpu_has_pse) {
1353 next = (addr + PAGE_SIZE) & PAGE_MASK;
1354 pmd = pmd_offset(pud, addr);
1355 if (pmd_none(*pmd))
1356 continue;
1357 get_page_bootmem(section_nr, pmd_page(*pmd),
1358 MIX_SECTION_INFO);
1359
1360 pte = pte_offset_kernel(pmd, addr);
1361 if (pte_none(*pte))
1362 continue;
1363 get_page_bootmem(section_nr, pte_page(*pte),
1364 SECTION_INFO);
1365 } else {
1366 next = pmd_addr_end(addr, end);
1367
1368 pmd = pmd_offset(pud, addr);
1369 if (pmd_none(*pmd))
1370 continue;
1371
1372 nr_pages = 1 << (get_order(PMD_SIZE));
1373 page = pmd_page(*pmd);
1374 while (nr_pages--)
1375 get_page_bootmem(section_nr, page++,
1376 SECTION_INFO);
1377 }
1378 }
1379}
1380#endif
1381
1382void __meminit vmemmap_populate_print_last(void)
1383{
1384 if (p_start) {
1385 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1386 addr_start, addr_end-1, p_start, p_end-1, node_start);
1387 p_start = NULL;
1388 p_end = NULL;
1389 node_start = 0;
1390 }
1391}
1392#endif