1 /*
2 * linux/arch/x86_64/mm/init.c
3 *
4 * Copyright (C) 1995 Linus Torvalds
5 * Copyright (C) 2000 Pavel Machek <pavel@suse.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/proc_fs.h>
25 #include <linux/pci.h>
26 #include <linux/pfn.h>
27 #include <linux/poison.h>
28 #include <linux/dma-mapping.h>
29 #include <linux/module.h>
30 #include <linux/memory_hotplug.h>
31 #include <linux/nmi.h>
32
33 #include <asm/processor.h>
34 #include <asm/bios_ebda.h>
35 #include <asm/system.h>
36 #include <asm/uaccess.h>
37 #include <asm/pgtable.h>
38 #include <asm/pgalloc.h>
39 #include <asm/dma.h>
40 #include <asm/fixmap.h>
41 #include <asm/e820.h>
42 #include <asm/apic.h>
43 #include <asm/tlb.h>
44 #include <asm/mmu_context.h>
45 #include <asm/proto.h>
46 #include <asm/smp.h>
47 #include <asm/sections.h>
48 #include <asm/kdebug.h>
49 #include <asm/numa.h>
50 #include <asm/cacheflush.h>
51
52 /*
53 * end_pfn only includes RAM, while max_pfn_mapped includes all e820 entries.
54 * The direct mapping extends to max_pfn_mapped, so that we can directly access
55 * apertures, ACPI and other tables without having to play with fixmaps.
56 */
57 unsigned long max_low_pfn_mapped;
58 unsigned long max_pfn_mapped;
59
60 static unsigned long dma_reserve __initdata;
61
62 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
63
64 int direct_gbpages
65 #ifdef CONFIG_DIRECT_GBPAGES
66 = 1
67 #endif
68 ;
69
parse_direct_gbpages_off(char * arg)70 static int __init parse_direct_gbpages_off(char *arg)
71 {
72 direct_gbpages = 0;
73 return 0;
74 }
75 early_param("nogbpages", parse_direct_gbpages_off);
76
parse_direct_gbpages_on(char * arg)77 static int __init parse_direct_gbpages_on(char *arg)
78 {
79 direct_gbpages = 1;
80 return 0;
81 }
82 early_param("gbpages", parse_direct_gbpages_on);
83
84 /*
85 * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
86 * physical space so we can cache the place of the first one and move
87 * around without checking the pgd every time.
88 */
89
90 int after_bootmem;
91
92 pteval_t __supported_pte_mask __read_mostly = ~_PAGE_IOMAP;
93 EXPORT_SYMBOL_GPL(__supported_pte_mask);
94
95 static int do_not_nx __cpuinitdata;
96
97 /*
98 * noexec=on|off
99 * Control non-executable mappings for 64-bit processes.
100 *
101 * on Enable (default)
102 * off Disable
103 */
nonx_setup(char * str)104 static int __init nonx_setup(char *str)
105 {
106 if (!str)
107 return -EINVAL;
108 if (!strncmp(str, "on", 2)) {
109 __supported_pte_mask |= _PAGE_NX;
110 do_not_nx = 0;
111 } else if (!strncmp(str, "off", 3)) {
112 do_not_nx = 1;
113 __supported_pte_mask &= ~_PAGE_NX;
114 }
115 return 0;
116 }
117 early_param("noexec", nonx_setup);
118
check_efer(void)119 void __cpuinit check_efer(void)
120 {
121 unsigned long efer;
122
123 rdmsrl(MSR_EFER, efer);
124 if (!(efer & EFER_NX) || do_not_nx)
125 __supported_pte_mask &= ~_PAGE_NX;
126 }
127
128 int force_personality32;
129
130 /*
131 * noexec32=on|off
132 * Control non executable heap for 32bit processes.
133 * To control the stack too use noexec=off
134 *
135 * on PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
136 * off PROT_READ implies PROT_EXEC
137 */
nonx32_setup(char * str)138 static int __init nonx32_setup(char *str)
139 {
140 if (!strcmp(str, "on"))
141 force_personality32 &= ~READ_IMPLIES_EXEC;
142 else if (!strcmp(str, "off"))
143 force_personality32 |= READ_IMPLIES_EXEC;
144 return 1;
145 }
146 __setup("noexec32=", nonx32_setup);
147
148 /*
149 * NOTE: This function is marked __ref because it calls __init function
150 * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
151 */
spp_getpage(void)152 static __ref void *spp_getpage(void)
153 {
154 void *ptr;
155
156 if (after_bootmem)
157 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
158 else
159 ptr = alloc_bootmem_pages(PAGE_SIZE);
160
161 if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
162 panic("set_pte_phys: cannot allocate page data %s\n",
163 after_bootmem ? "after bootmem" : "");
164 }
165
166 pr_debug("spp_getpage %p\n", ptr);
167
168 return ptr;
169 }
170
171 void
set_pte_vaddr_pud(pud_t * pud_page,unsigned long vaddr,pte_t new_pte)172 set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
173 {
174 pud_t *pud;
175 pmd_t *pmd;
176 pte_t *pte;
177
178 pud = pud_page + pud_index(vaddr);
179 if (pud_none(*pud)) {
180 pmd = (pmd_t *) spp_getpage();
181 pud_populate(&init_mm, pud, pmd);
182 if (pmd != pmd_offset(pud, 0)) {
183 printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
184 pmd, pmd_offset(pud, 0));
185 return;
186 }
187 }
188 pmd = pmd_offset(pud, vaddr);
189 if (pmd_none(*pmd)) {
190 pte = (pte_t *) spp_getpage();
191 pmd_populate_kernel(&init_mm, pmd, pte);
192 if (pte != pte_offset_kernel(pmd, 0)) {
193 printk(KERN_ERR "PAGETABLE BUG #02!\n");
194 return;
195 }
196 }
197
198 pte = pte_offset_kernel(pmd, vaddr);
199 set_pte(pte, new_pte);
200
201 /*
202 * It's enough to flush this one mapping.
203 * (PGE mappings get flushed as well)
204 */
205 __flush_tlb_one(vaddr);
206 }
207
208 void
set_pte_vaddr(unsigned long vaddr,pte_t pteval)209 set_pte_vaddr(unsigned long vaddr, pte_t pteval)
210 {
211 pgd_t *pgd;
212 pud_t *pud_page;
213
214 pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
215
216 pgd = pgd_offset_k(vaddr);
217 if (pgd_none(*pgd)) {
218 printk(KERN_ERR
219 "PGD FIXMAP MISSING, it should be setup in head.S!\n");
220 return;
221 }
222 pud_page = (pud_t*)pgd_page_vaddr(*pgd);
223 set_pte_vaddr_pud(pud_page, vaddr, pteval);
224 }
225
226 /*
227 * Create large page table mappings for a range of physical addresses.
228 */
__init_extra_mapping(unsigned long phys,unsigned long size,pgprot_t prot)229 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
230 pgprot_t prot)
231 {
232 pgd_t *pgd;
233 pud_t *pud;
234 pmd_t *pmd;
235
236 BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
237 for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
238 pgd = pgd_offset_k((unsigned long)__va(phys));
239 if (pgd_none(*pgd)) {
240 pud = (pud_t *) spp_getpage();
241 set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
242 _PAGE_USER));
243 }
244 pud = pud_offset(pgd, (unsigned long)__va(phys));
245 if (pud_none(*pud)) {
246 pmd = (pmd_t *) spp_getpage();
247 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
248 _PAGE_USER));
249 }
250 pmd = pmd_offset(pud, phys);
251 BUG_ON(!pmd_none(*pmd));
252 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
253 }
254 }
255
init_extra_mapping_wb(unsigned long phys,unsigned long size)256 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
257 {
258 __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
259 }
260
init_extra_mapping_uc(unsigned long phys,unsigned long size)261 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
262 {
263 __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
264 }
265
266 /*
267 * The head.S code sets up the kernel high mapping:
268 *
269 * from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
270 *
271 * phys_addr holds the negative offset to the kernel, which is added
272 * to the compile time generated pmds. This results in invalid pmds up
273 * to the point where we hit the physaddr 0 mapping.
274 *
275 * We limit the mappings to the region from _text to _end. _end is
276 * rounded up to the 2MB boundary. This catches the invalid pmds as
277 * well, as they are located before _text:
278 */
cleanup_highmap(void)279 void __init cleanup_highmap(void)
280 {
281 unsigned long vaddr = __START_KERNEL_map;
282 unsigned long end = roundup((unsigned long)_end, PMD_SIZE) - 1;
283 pmd_t *pmd = level2_kernel_pgt;
284 pmd_t *last_pmd = pmd + PTRS_PER_PMD;
285
286 for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
287 if (pmd_none(*pmd))
288 continue;
289 if (vaddr < (unsigned long) _text || vaddr > end)
290 set_pmd(pmd, __pmd(0));
291 }
292 }
293
294 static unsigned long __initdata table_start;
295 static unsigned long __meminitdata table_end;
296 static unsigned long __meminitdata table_top;
297
alloc_low_page(unsigned long * phys)298 static __ref void *alloc_low_page(unsigned long *phys)
299 {
300 unsigned long pfn = table_end++;
301 void *adr;
302
303 if (after_bootmem) {
304 adr = (void *)get_zeroed_page(GFP_ATOMIC);
305 *phys = __pa(adr);
306
307 return adr;
308 }
309
310 if (pfn >= table_top)
311 panic("alloc_low_page: ran out of memory");
312
313 adr = early_memremap(pfn * PAGE_SIZE, PAGE_SIZE);
314 memset(adr, 0, PAGE_SIZE);
315 *phys = pfn * PAGE_SIZE;
316 return adr;
317 }
318
unmap_low_page(void * adr)319 static __ref void unmap_low_page(void *adr)
320 {
321 if (after_bootmem)
322 return;
323
324 early_iounmap(adr, PAGE_SIZE);
325 }
326
327 static unsigned long __meminit
phys_pte_init(pte_t * pte_page,unsigned long addr,unsigned long end,pgprot_t prot)328 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
329 pgprot_t prot)
330 {
331 unsigned pages = 0;
332 unsigned long last_map_addr = end;
333 int i;
334
335 pte_t *pte = pte_page + pte_index(addr);
336
337 for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
338
339 if (addr >= end) {
340 if (!after_bootmem) {
341 for(; i < PTRS_PER_PTE; i++, pte++)
342 set_pte(pte, __pte(0));
343 }
344 break;
345 }
346
347 /*
348 * We will re-use the existing mapping.
349 * Xen for example has some special requirements, like mapping
350 * pagetable pages as RO. So assume someone who pre-setup
351 * these mappings are more intelligent.
352 */
353 if (pte_val(*pte)) {
354 pages++;
355 continue;
356 }
357
358 if (0)
359 printk(" pte=%p addr=%lx pte=%016lx\n",
360 pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
361 pages++;
362 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
363 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
364 }
365
366 update_page_count(PG_LEVEL_4K, pages);
367
368 return last_map_addr;
369 }
370
371 static unsigned long __meminit
phys_pte_update(pmd_t * pmd,unsigned long address,unsigned long end,pgprot_t prot)372 phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end,
373 pgprot_t prot)
374 {
375 pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd);
376
377 return phys_pte_init(pte, address, end, prot);
378 }
379
380 static unsigned long __meminit
phys_pmd_init(pmd_t * pmd_page,unsigned long address,unsigned long end,unsigned long page_size_mask,pgprot_t prot)381 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
382 unsigned long page_size_mask, pgprot_t prot)
383 {
384 unsigned long pages = 0;
385 unsigned long last_map_addr = end;
386
387 int i = pmd_index(address);
388
389 for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
390 unsigned long pte_phys;
391 pmd_t *pmd = pmd_page + pmd_index(address);
392 pte_t *pte;
393 pgprot_t new_prot = prot;
394
395 if (address >= end) {
396 if (!after_bootmem) {
397 for (; i < PTRS_PER_PMD; i++, pmd++)
398 set_pmd(pmd, __pmd(0));
399 }
400 break;
401 }
402
403 if (pmd_val(*pmd)) {
404 if (!pmd_large(*pmd)) {
405 spin_lock(&init_mm.page_table_lock);
406 last_map_addr = phys_pte_update(pmd, address,
407 end, prot);
408 spin_unlock(&init_mm.page_table_lock);
409 continue;
410 }
411 /*
412 * If we are ok with PG_LEVEL_2M mapping, then we will
413 * use the existing mapping,
414 *
415 * Otherwise, we will split the large page mapping but
416 * use the same existing protection bits except for
417 * large page, so that we don't violate Intel's TLB
418 * Application note (317080) which says, while changing
419 * the page sizes, new and old translations should
420 * not differ with respect to page frame and
421 * attributes.
422 */
423 if (page_size_mask & (1 << PG_LEVEL_2M)) {
424 pages++;
425 continue;
426 }
427 new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
428 }
429
430 if (page_size_mask & (1<<PG_LEVEL_2M)) {
431 pages++;
432 spin_lock(&init_mm.page_table_lock);
433 set_pte((pte_t *)pmd,
434 pfn_pte(address >> PAGE_SHIFT,
435 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
436 spin_unlock(&init_mm.page_table_lock);
437 last_map_addr = (address & PMD_MASK) + PMD_SIZE;
438 continue;
439 }
440
441 pte = alloc_low_page(&pte_phys);
442 last_map_addr = phys_pte_init(pte, address, end, new_prot);
443 unmap_low_page(pte);
444
445 spin_lock(&init_mm.page_table_lock);
446 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
447 spin_unlock(&init_mm.page_table_lock);
448 }
449 update_page_count(PG_LEVEL_2M, pages);
450 return last_map_addr;
451 }
452
453 static unsigned long __meminit
phys_pmd_update(pud_t * pud,unsigned long address,unsigned long end,unsigned long page_size_mask,pgprot_t prot)454 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end,
455 unsigned long page_size_mask, pgprot_t prot)
456 {
457 pmd_t *pmd = pmd_offset(pud, 0);
458 unsigned long last_map_addr;
459
460 last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask, prot);
461 __flush_tlb_all();
462 return last_map_addr;
463 }
464
465 static unsigned long __meminit
phys_pud_init(pud_t * pud_page,unsigned long addr,unsigned long end,unsigned long page_size_mask)466 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
467 unsigned long page_size_mask)
468 {
469 unsigned long pages = 0;
470 unsigned long last_map_addr = end;
471 int i = pud_index(addr);
472
473 for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
474 unsigned long pmd_phys;
475 pud_t *pud = pud_page + pud_index(addr);
476 pmd_t *pmd;
477 pgprot_t prot = PAGE_KERNEL;
478
479 if (addr >= end)
480 break;
481
482 if (!after_bootmem &&
483 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
484 set_pud(pud, __pud(0));
485 continue;
486 }
487
488 if (pud_val(*pud)) {
489 if (!pud_large(*pud)) {
490 last_map_addr = phys_pmd_update(pud, addr, end,
491 page_size_mask, prot);
492 continue;
493 }
494 /*
495 * If we are ok with PG_LEVEL_1G mapping, then we will
496 * use the existing mapping.
497 *
498 * Otherwise, we will split the gbpage mapping but use
499 * the same existing protection bits except for large
500 * page, so that we don't violate Intel's TLB
501 * Application note (317080) which says, while changing
502 * the page sizes, new and old translations should
503 * not differ with respect to page frame and
504 * attributes.
505 */
506 if (page_size_mask & (1 << PG_LEVEL_1G)) {
507 pages++;
508 continue;
509 }
510 prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
511 }
512
513 if (page_size_mask & (1<<PG_LEVEL_1G)) {
514 pages++;
515 spin_lock(&init_mm.page_table_lock);
516 set_pte((pte_t *)pud,
517 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
518 spin_unlock(&init_mm.page_table_lock);
519 last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
520 continue;
521 }
522
523 pmd = alloc_low_page(&pmd_phys);
524 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
525 prot);
526 unmap_low_page(pmd);
527
528 spin_lock(&init_mm.page_table_lock);
529 pud_populate(&init_mm, pud, __va(pmd_phys));
530 spin_unlock(&init_mm.page_table_lock);
531 }
532 __flush_tlb_all();
533
534 update_page_count(PG_LEVEL_1G, pages);
535
536 return last_map_addr;
537 }
538
539 static unsigned long __meminit
phys_pud_update(pgd_t * pgd,unsigned long addr,unsigned long end,unsigned long page_size_mask)540 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end,
541 unsigned long page_size_mask)
542 {
543 pud_t *pud;
544
545 pud = (pud_t *)pgd_page_vaddr(*pgd);
546
547 return phys_pud_init(pud, addr, end, page_size_mask);
548 }
549
find_early_table_space(unsigned long end,int use_pse,int use_gbpages)550 static void __init find_early_table_space(unsigned long end, int use_pse,
551 int use_gbpages)
552 {
553 unsigned long puds, pmds, ptes, tables, start;
554
555 puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
556 tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
557 if (use_gbpages) {
558 unsigned long extra;
559 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
560 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
561 } else
562 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
563 tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
564
565 if (use_pse) {
566 unsigned long extra;
567 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
568 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
569 } else
570 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
571 tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
572
573 /*
574 * RED-PEN putting page tables only on node 0 could
575 * cause a hotspot and fill up ZONE_DMA. The page tables
576 * need roughly 0.5KB per GB.
577 */
578 start = 0x8000;
579 table_start = find_e820_area(start, end, tables, PAGE_SIZE);
580 if (table_start == -1UL)
581 panic("Cannot find space for the kernel page tables");
582
583 table_start >>= PAGE_SHIFT;
584 table_end = table_start;
585 table_top = table_start + (tables >> PAGE_SHIFT);
586
587 printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
588 end, table_start << PAGE_SHIFT, table_top << PAGE_SHIFT);
589 }
590
init_gbpages(void)591 static void __init init_gbpages(void)
592 {
593 if (direct_gbpages && cpu_has_gbpages)
594 printk(KERN_INFO "Using GB pages for direct mapping\n");
595 else
596 direct_gbpages = 0;
597 }
598
kernel_physical_mapping_init(unsigned long start,unsigned long end,unsigned long page_size_mask)599 static unsigned long __meminit kernel_physical_mapping_init(unsigned long start,
600 unsigned long end,
601 unsigned long page_size_mask)
602 {
603
604 unsigned long next, last_map_addr = end;
605
606 start = (unsigned long)__va(start);
607 end = (unsigned long)__va(end);
608
609 for (; start < end; start = next) {
610 pgd_t *pgd = pgd_offset_k(start);
611 unsigned long pud_phys;
612 pud_t *pud;
613
614 next = (start + PGDIR_SIZE) & PGDIR_MASK;
615 if (next > end)
616 next = end;
617
618 if (pgd_val(*pgd)) {
619 last_map_addr = phys_pud_update(pgd, __pa(start),
620 __pa(end), page_size_mask);
621 continue;
622 }
623
624 pud = alloc_low_page(&pud_phys);
625 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
626 page_size_mask);
627 unmap_low_page(pud);
628
629 spin_lock(&init_mm.page_table_lock);
630 pgd_populate(&init_mm, pgd, __va(pud_phys));
631 spin_unlock(&init_mm.page_table_lock);
632 }
633 __flush_tlb_all();
634
635 return last_map_addr;
636 }
637
638 struct map_range {
639 unsigned long start;
640 unsigned long end;
641 unsigned page_size_mask;
642 };
643
644 #define NR_RANGE_MR 5
645
save_mr(struct map_range * mr,int nr_range,unsigned long start_pfn,unsigned long end_pfn,unsigned long page_size_mask)646 static int save_mr(struct map_range *mr, int nr_range,
647 unsigned long start_pfn, unsigned long end_pfn,
648 unsigned long page_size_mask)
649 {
650
651 if (start_pfn < end_pfn) {
652 if (nr_range >= NR_RANGE_MR)
653 panic("run out of range for init_memory_mapping\n");
654 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
655 mr[nr_range].end = end_pfn<<PAGE_SHIFT;
656 mr[nr_range].page_size_mask = page_size_mask;
657 nr_range++;
658 }
659
660 return nr_range;
661 }
662
663 /*
664 * Setup the direct mapping of the physical memory at PAGE_OFFSET.
665 * This runs before bootmem is initialized and gets pages directly from
666 * the physical memory. To access them they are temporarily mapped.
667 */
init_memory_mapping(unsigned long start,unsigned long end)668 unsigned long __init_refok init_memory_mapping(unsigned long start,
669 unsigned long end)
670 {
671 unsigned long last_map_addr = 0;
672 unsigned long page_size_mask = 0;
673 unsigned long start_pfn, end_pfn;
674 unsigned long pos;
675
676 struct map_range mr[NR_RANGE_MR];
677 int nr_range, i;
678 int use_pse, use_gbpages;
679
680 printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
681
682 /*
683 * Find space for the kernel direct mapping tables.
684 *
685 * Later we should allocate these tables in the local node of the
686 * memory mapped. Unfortunately this is done currently before the
687 * nodes are discovered.
688 */
689 if (!after_bootmem)
690 init_gbpages();
691
692 #ifdef CONFIG_DEBUG_PAGEALLOC
693 /*
694 * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
695 * This will simplify cpa(), which otherwise needs to support splitting
696 * large pages into small in interrupt context, etc.
697 */
698 use_pse = use_gbpages = 0;
699 #else
700 use_pse = cpu_has_pse;
701 use_gbpages = direct_gbpages;
702 #endif
703
704 if (use_gbpages)
705 page_size_mask |= 1 << PG_LEVEL_1G;
706 if (use_pse)
707 page_size_mask |= 1 << PG_LEVEL_2M;
708
709 memset(mr, 0, sizeof(mr));
710 nr_range = 0;
711
712 /* head if not big page alignment ?*/
713 start_pfn = start >> PAGE_SHIFT;
714 pos = start_pfn << PAGE_SHIFT;
715 end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
716 << (PMD_SHIFT - PAGE_SHIFT);
717 if (end_pfn > (end >> PAGE_SHIFT))
718 end_pfn = end >> PAGE_SHIFT;
719 if (start_pfn < end_pfn) {
720 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
721 pos = end_pfn << PAGE_SHIFT;
722 }
723
724 /* big page (2M) range*/
725 start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
726 << (PMD_SHIFT - PAGE_SHIFT);
727 end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
728 << (PUD_SHIFT - PAGE_SHIFT);
729 if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
730 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
731 if (start_pfn < end_pfn) {
732 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
733 page_size_mask & (1<<PG_LEVEL_2M));
734 pos = end_pfn << PAGE_SHIFT;
735 }
736
737 /* big page (1G) range */
738 start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
739 << (PUD_SHIFT - PAGE_SHIFT);
740 end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
741 if (start_pfn < end_pfn) {
742 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
743 page_size_mask &
744 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
745 pos = end_pfn << PAGE_SHIFT;
746 }
747
748 /* tail is not big page (1G) alignment */
749 start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
750 << (PMD_SHIFT - PAGE_SHIFT);
751 end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
752 if (start_pfn < end_pfn) {
753 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
754 page_size_mask & (1<<PG_LEVEL_2M));
755 pos = end_pfn << PAGE_SHIFT;
756 }
757
758 /* tail is not big page (2M) alignment */
759 start_pfn = pos>>PAGE_SHIFT;
760 end_pfn = end>>PAGE_SHIFT;
761 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
762
763 /* try to merge same page size and continuous */
764 for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
765 unsigned long old_start;
766 if (mr[i].end != mr[i+1].start ||
767 mr[i].page_size_mask != mr[i+1].page_size_mask)
768 continue;
769 /* move it */
770 old_start = mr[i].start;
771 memmove(&mr[i], &mr[i+1],
772 (nr_range - 1 - i) * sizeof (struct map_range));
773 mr[i--].start = old_start;
774 nr_range--;
775 }
776
777 for (i = 0; i < nr_range; i++)
778 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
779 mr[i].start, mr[i].end,
780 (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
781 (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
782
783 if (!after_bootmem)
784 find_early_table_space(end, use_pse, use_gbpages);
785
786 for (i = 0; i < nr_range; i++)
787 last_map_addr = kernel_physical_mapping_init(
788 mr[i].start, mr[i].end,
789 mr[i].page_size_mask);
790
791 if (!after_bootmem)
792 mmu_cr4_features = read_cr4();
793 __flush_tlb_all();
794
795 if (!after_bootmem && table_end > table_start)
796 reserve_early(table_start << PAGE_SHIFT,
797 table_end << PAGE_SHIFT, "PGTABLE");
798
799 printk(KERN_INFO "last_map_addr: %lx end: %lx\n",
800 last_map_addr, end);
801
802 if (!after_bootmem)
803 early_memtest(start, end);
804
805 return last_map_addr >> PAGE_SHIFT;
806 }
807
808 #ifndef CONFIG_NUMA
initmem_init(unsigned long start_pfn,unsigned long end_pfn)809 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
810 {
811 unsigned long bootmap_size, bootmap;
812
813 bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
814 bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
815 PAGE_SIZE);
816 if (bootmap == -1L)
817 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
818 /* don't touch min_low_pfn */
819 bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
820 0, end_pfn);
821 e820_register_active_regions(0, start_pfn, end_pfn);
822 free_bootmem_with_active_regions(0, end_pfn);
823 early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
824 reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
825 }
826
paging_init(void)827 void __init paging_init(void)
828 {
829 unsigned long max_zone_pfns[MAX_NR_ZONES];
830
831 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
832 max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
833 max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
834 max_zone_pfns[ZONE_NORMAL] = max_pfn;
835
836 memory_present(0, 0, max_pfn);
837 sparse_init();
838 free_area_init_nodes(max_zone_pfns);
839 }
840 #endif
841
842 /*
843 * Memory hotplug specific functions
844 */
845 #ifdef CONFIG_MEMORY_HOTPLUG
846 /*
847 * Memory is added always to NORMAL zone. This means you will never get
848 * additional DMA/DMA32 memory.
849 */
arch_add_memory(int nid,u64 start,u64 size)850 int arch_add_memory(int nid, u64 start, u64 size)
851 {
852 struct pglist_data *pgdat = NODE_DATA(nid);
853 struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
854 unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
855 unsigned long nr_pages = size >> PAGE_SHIFT;
856 int ret;
857
858 last_mapped_pfn = init_memory_mapping(start, start + size);
859 if (last_mapped_pfn > max_pfn_mapped)
860 max_pfn_mapped = last_mapped_pfn;
861
862 ret = __add_pages(nid, zone, start_pfn, nr_pages);
863 WARN_ON_ONCE(ret);
864
865 return ret;
866 }
867 EXPORT_SYMBOL_GPL(arch_add_memory);
868
869 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
memory_add_physaddr_to_nid(u64 start)870 int memory_add_physaddr_to_nid(u64 start)
871 {
872 return 0;
873 }
874 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
875 #endif
876
877 #endif /* CONFIG_MEMORY_HOTPLUG */
878
879 /*
880 * devmem_is_allowed() checks to see if /dev/mem access to a certain address
881 * is valid. The argument is a physical page number.
882 *
883 *
884 * On x86, access has to be given to the first megabyte of ram because that area
885 * contains bios code and data regions used by X and dosemu and similar apps.
886 * Access has to be given to non-kernel-ram areas as well, these contain the PCI
887 * mmio resources as well as potential bios/acpi data regions.
888 */
devmem_is_allowed(unsigned long pagenr)889 int devmem_is_allowed(unsigned long pagenr)
890 {
891 if (pagenr <= 256)
892 return 1;
893 if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
894 return 0;
895 if (!page_is_ram(pagenr))
896 return 1;
897 return 0;
898 }
899
900
901 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
902 kcore_modules, kcore_vsyscall;
903
mem_init(void)904 void __init mem_init(void)
905 {
906 long codesize, reservedpages, datasize, initsize;
907 unsigned long absent_pages;
908
909 pci_iommu_alloc();
910
911 /* clear_bss() already clear the empty_zero_page */
912
913 reservedpages = 0;
914
915 /* this will put all low memory onto the freelists */
916 #ifdef CONFIG_NUMA
917 totalram_pages = numa_free_all_bootmem();
918 #else
919 totalram_pages = free_all_bootmem();
920 #endif
921
922 absent_pages = absent_pages_in_range(0, max_pfn);
923 reservedpages = max_pfn - totalram_pages - absent_pages;
924 after_bootmem = 1;
925
926 codesize = (unsigned long) &_etext - (unsigned long) &_text;
927 datasize = (unsigned long) &_edata - (unsigned long) &_etext;
928 initsize = (unsigned long) &__init_end - (unsigned long) &__init_begin;
929
930 /* Register memory areas for /proc/kcore */
931 kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
932 kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
933 VMALLOC_END-VMALLOC_START);
934 kclist_add(&kcore_kernel, &_stext, _end - _stext);
935 kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
936 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
937 VSYSCALL_END - VSYSCALL_START);
938
939 printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
940 "%ldk absent, %ldk reserved, %ldk data, %ldk init)\n",
941 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
942 max_pfn << (PAGE_SHIFT-10),
943 codesize >> 10,
944 absent_pages << (PAGE_SHIFT-10),
945 reservedpages << (PAGE_SHIFT-10),
946 datasize >> 10,
947 initsize >> 10);
948 }
949
free_init_pages(char * what,unsigned long begin,unsigned long end)950 void free_init_pages(char *what, unsigned long begin, unsigned long end)
951 {
952 unsigned long addr = begin;
953
954 if (addr >= end)
955 return;
956
957 /*
958 * If debugging page accesses then do not free this memory but
959 * mark them not present - any buggy init-section access will
960 * create a kernel page fault:
961 */
962 #ifdef CONFIG_DEBUG_PAGEALLOC
963 printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
964 begin, PAGE_ALIGN(end));
965 set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
966 #else
967 printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
968
969 for (; addr < end; addr += PAGE_SIZE) {
970 ClearPageReserved(virt_to_page(addr));
971 init_page_count(virt_to_page(addr));
972 memset((void *)(addr & ~(PAGE_SIZE-1)),
973 POISON_FREE_INITMEM, PAGE_SIZE);
974 free_page(addr);
975 totalram_pages++;
976 }
977 #endif
978 }
979
free_initmem(void)980 void free_initmem(void)
981 {
982 free_init_pages("unused kernel memory",
983 (unsigned long)(&__init_begin),
984 (unsigned long)(&__init_end));
985 }
986
987 #ifdef CONFIG_DEBUG_RODATA
988 const int rodata_test_data = 0xC3;
989 EXPORT_SYMBOL_GPL(rodata_test_data);
990
mark_rodata_ro(void)991 void mark_rodata_ro(void)
992 {
993 unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
994 unsigned long rodata_start =
995 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
996
997 #ifdef CONFIG_DYNAMIC_FTRACE
998 /* Dynamic tracing modifies the kernel text section */
999 start = rodata_start;
1000 #endif
1001
1002 printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1003 (end - start) >> 10);
1004 set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1005
1006 /*
1007 * The rodata section (but not the kernel text!) should also be
1008 * not-executable.
1009 */
1010 set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
1011
1012 rodata_test();
1013
1014 #ifdef CONFIG_CPA_DEBUG
1015 printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1016 set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1017
1018 printk(KERN_INFO "Testing CPA: again\n");
1019 set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1020 #endif
1021 }
1022
1023 #endif
1024
1025 #ifdef CONFIG_BLK_DEV_INITRD
free_initrd_mem(unsigned long start,unsigned long end)1026 void free_initrd_mem(unsigned long start, unsigned long end)
1027 {
1028 free_init_pages("initrd memory", start, end);
1029 }
1030 #endif
1031
reserve_bootmem_generic(unsigned long phys,unsigned long len,int flags)1032 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
1033 int flags)
1034 {
1035 #ifdef CONFIG_NUMA
1036 int nid, next_nid;
1037 int ret;
1038 #endif
1039 unsigned long pfn = phys >> PAGE_SHIFT;
1040
1041 if (pfn >= max_pfn) {
1042 /*
1043 * This can happen with kdump kernels when accessing
1044 * firmware tables:
1045 */
1046 if (pfn < max_pfn_mapped)
1047 return -EFAULT;
1048
1049 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
1050 phys, len);
1051 return -EFAULT;
1052 }
1053
1054 /* Should check here against the e820 map to avoid double free */
1055 #ifdef CONFIG_NUMA
1056 nid = phys_to_nid(phys);
1057 next_nid = phys_to_nid(phys + len - 1);
1058 if (nid == next_nid)
1059 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
1060 else
1061 ret = reserve_bootmem(phys, len, flags);
1062
1063 if (ret != 0)
1064 return ret;
1065
1066 #else
1067 reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
1068 #endif
1069
1070 if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
1071 dma_reserve += len / PAGE_SIZE;
1072 set_dma_reserve(dma_reserve);
1073 }
1074
1075 return 0;
1076 }
1077
kern_addr_valid(unsigned long addr)1078 int kern_addr_valid(unsigned long addr)
1079 {
1080 unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1081 pgd_t *pgd;
1082 pud_t *pud;
1083 pmd_t *pmd;
1084 pte_t *pte;
1085
1086 if (above != 0 && above != -1UL)
1087 return 0;
1088
1089 pgd = pgd_offset_k(addr);
1090 if (pgd_none(*pgd))
1091 return 0;
1092
1093 pud = pud_offset(pgd, addr);
1094 if (pud_none(*pud))
1095 return 0;
1096
1097 pmd = pmd_offset(pud, addr);
1098 if (pmd_none(*pmd))
1099 return 0;
1100
1101 if (pmd_large(*pmd))
1102 return pfn_valid(pmd_pfn(*pmd));
1103
1104 pte = pte_offset_kernel(pmd, addr);
1105 if (pte_none(*pte))
1106 return 0;
1107
1108 return pfn_valid(pte_pfn(*pte));
1109 }
1110
1111 /*
1112 * A pseudo VMA to allow ptrace access for the vsyscall page. This only
1113 * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1114 * not need special handling anymore:
1115 */
1116 static struct vm_area_struct gate_vma = {
1117 .vm_start = VSYSCALL_START,
1118 .vm_end = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
1119 .vm_page_prot = PAGE_READONLY_EXEC,
1120 .vm_flags = VM_READ | VM_EXEC
1121 };
1122
get_gate_vma(struct task_struct * tsk)1123 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1124 {
1125 #ifdef CONFIG_IA32_EMULATION
1126 if (test_tsk_thread_flag(tsk, TIF_IA32))
1127 return NULL;
1128 #endif
1129 return &gate_vma;
1130 }
1131
in_gate_area(struct task_struct * task,unsigned long addr)1132 int in_gate_area(struct task_struct *task, unsigned long addr)
1133 {
1134 struct vm_area_struct *vma = get_gate_vma(task);
1135
1136 if (!vma)
1137 return 0;
1138
1139 return (addr >= vma->vm_start) && (addr < vma->vm_end);
1140 }
1141
1142 /*
1143 * Use this when you have no reliable task/vma, typically from interrupt
1144 * context. It is less reliable than using the task's vma and may give
1145 * false positives:
1146 */
in_gate_area_no_task(unsigned long addr)1147 int in_gate_area_no_task(unsigned long addr)
1148 {
1149 return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1150 }
1151
arch_vma_name(struct vm_area_struct * vma)1152 const char *arch_vma_name(struct vm_area_struct *vma)
1153 {
1154 if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1155 return "[vdso]";
1156 if (vma == &gate_vma)
1157 return "[vsyscall]";
1158 return NULL;
1159 }
1160
1161 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1162 /*
1163 * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1164 */
1165 static long __meminitdata addr_start, addr_end;
1166 static void __meminitdata *p_start, *p_end;
1167 static int __meminitdata node_start;
1168
1169 int __meminit
vmemmap_populate(struct page * start_page,unsigned long size,int node)1170 vmemmap_populate(struct page *start_page, unsigned long size, int node)
1171 {
1172 unsigned long addr = (unsigned long)start_page;
1173 unsigned long end = (unsigned long)(start_page + size);
1174 unsigned long next;
1175 pgd_t *pgd;
1176 pud_t *pud;
1177 pmd_t *pmd;
1178
1179 for (; addr < end; addr = next) {
1180 void *p = NULL;
1181
1182 pgd = vmemmap_pgd_populate(addr, node);
1183 if (!pgd)
1184 return -ENOMEM;
1185
1186 pud = vmemmap_pud_populate(pgd, addr, node);
1187 if (!pud)
1188 return -ENOMEM;
1189
1190 if (!cpu_has_pse) {
1191 next = (addr + PAGE_SIZE) & PAGE_MASK;
1192 pmd = vmemmap_pmd_populate(pud, addr, node);
1193
1194 if (!pmd)
1195 return -ENOMEM;
1196
1197 p = vmemmap_pte_populate(pmd, addr, node);
1198
1199 if (!p)
1200 return -ENOMEM;
1201
1202 addr_end = addr + PAGE_SIZE;
1203 p_end = p + PAGE_SIZE;
1204 } else {
1205 next = pmd_addr_end(addr, end);
1206
1207 pmd = pmd_offset(pud, addr);
1208 if (pmd_none(*pmd)) {
1209 pte_t entry;
1210
1211 p = vmemmap_alloc_block(PMD_SIZE, node);
1212 if (!p)
1213 return -ENOMEM;
1214
1215 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1216 PAGE_KERNEL_LARGE);
1217 set_pmd(pmd, __pmd(pte_val(entry)));
1218
1219 /* check to see if we have contiguous blocks */
1220 if (p_end != p || node_start != node) {
1221 if (p_start)
1222 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1223 addr_start, addr_end-1, p_start, p_end-1, node_start);
1224 addr_start = addr;
1225 node_start = node;
1226 p_start = p;
1227 }
1228
1229 addr_end = addr + PMD_SIZE;
1230 p_end = p + PMD_SIZE;
1231 } else
1232 vmemmap_verify((pte_t *)pmd, node, addr, next);
1233 }
1234
1235 }
1236 return 0;
1237 }
1238
vmemmap_populate_print_last(void)1239 void __meminit vmemmap_populate_print_last(void)
1240 {
1241 if (p_start) {
1242 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1243 addr_start, addr_end-1, p_start, p_end-1, node_start);
1244 p_start = NULL;
1245 p_end = NULL;
1246 node_start = 0;
1247 }
1248 }
1249 #endif
1250