xref: /linux/arch/x86/mm/init_64.c (revision 367b8112fe2ea5c39a7bb4d263dcdd9b612fae18)
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 
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 
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  */
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 
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  */
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  */
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
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
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  */
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 
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 
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  */
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 
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 
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
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
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
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
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
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
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 
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 
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 
599 static unsigned long __init 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 
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  */
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 (start_pfn < end_pfn) {
718 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
719 		pos = end_pfn << PAGE_SHIFT;
720 	}
721 
722 	/* big page (2M) range*/
723 	start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
724 			 << (PMD_SHIFT - PAGE_SHIFT);
725 	end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
726 			 << (PUD_SHIFT - PAGE_SHIFT);
727 	if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
728 		end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
729 	if (start_pfn < end_pfn) {
730 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
731 				page_size_mask & (1<<PG_LEVEL_2M));
732 		pos = end_pfn << PAGE_SHIFT;
733 	}
734 
735 	/* big page (1G) range */
736 	start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
737 			 << (PUD_SHIFT - PAGE_SHIFT);
738 	end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
739 	if (start_pfn < end_pfn) {
740 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
741 				page_size_mask &
742 				 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
743 		pos = end_pfn << PAGE_SHIFT;
744 	}
745 
746 	/* tail is not big page (1G) alignment */
747 	start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
748 			 << (PMD_SHIFT - PAGE_SHIFT);
749 	end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
750 	if (start_pfn < end_pfn) {
751 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
752 				page_size_mask & (1<<PG_LEVEL_2M));
753 		pos = end_pfn << PAGE_SHIFT;
754 	}
755 
756 	/* tail is not big page (2M) alignment */
757 	start_pfn = pos>>PAGE_SHIFT;
758 	end_pfn = end>>PAGE_SHIFT;
759 	nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
760 
761 	/* try to merge same page size and continuous */
762 	for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
763 		unsigned long old_start;
764 		if (mr[i].end != mr[i+1].start ||
765 		    mr[i].page_size_mask != mr[i+1].page_size_mask)
766 			continue;
767 		/* move it */
768 		old_start = mr[i].start;
769 		memmove(&mr[i], &mr[i+1],
770 			 (nr_range - 1 - i) * sizeof (struct map_range));
771 		mr[i--].start = old_start;
772 		nr_range--;
773 	}
774 
775 	for (i = 0; i < nr_range; i++)
776 		printk(KERN_DEBUG " %010lx - %010lx page %s\n",
777 				mr[i].start, mr[i].end,
778 			(mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
779 			 (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
780 
781 	if (!after_bootmem)
782 		find_early_table_space(end, use_pse, use_gbpages);
783 
784 	for (i = 0; i < nr_range; i++)
785 		last_map_addr = kernel_physical_mapping_init(
786 					mr[i].start, mr[i].end,
787 					mr[i].page_size_mask);
788 
789 	if (!after_bootmem)
790 		mmu_cr4_features = read_cr4();
791 	__flush_tlb_all();
792 
793 	if (!after_bootmem && table_end > table_start)
794 		reserve_early(table_start << PAGE_SHIFT,
795 				 table_end << PAGE_SHIFT, "PGTABLE");
796 
797 	printk(KERN_INFO "last_map_addr: %lx end: %lx\n",
798 			 last_map_addr, end);
799 
800 	if (!after_bootmem)
801 		early_memtest(start, end);
802 
803 	return last_map_addr >> PAGE_SHIFT;
804 }
805 
806 #ifndef CONFIG_NUMA
807 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
808 {
809 	unsigned long bootmap_size, bootmap;
810 
811 	bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
812 	bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
813 				 PAGE_SIZE);
814 	if (bootmap == -1L)
815 		panic("Cannot find bootmem map of size %ld\n", bootmap_size);
816 	/* don't touch min_low_pfn */
817 	bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
818 					 0, end_pfn);
819 	e820_register_active_regions(0, start_pfn, end_pfn);
820 	free_bootmem_with_active_regions(0, end_pfn);
821 	early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
822 	reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
823 }
824 
825 void __init paging_init(void)
826 {
827 	unsigned long max_zone_pfns[MAX_NR_ZONES];
828 
829 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
830 	max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
831 	max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
832 	max_zone_pfns[ZONE_NORMAL] = max_pfn;
833 
834 	memory_present(0, 0, max_pfn);
835 	sparse_init();
836 	free_area_init_nodes(max_zone_pfns);
837 }
838 #endif
839 
840 /*
841  * Memory hotplug specific functions
842  */
843 #ifdef CONFIG_MEMORY_HOTPLUG
844 /*
845  * Memory is added always to NORMAL zone. This means you will never get
846  * additional DMA/DMA32 memory.
847  */
848 int arch_add_memory(int nid, u64 start, u64 size)
849 {
850 	struct pglist_data *pgdat = NODE_DATA(nid);
851 	struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
852 	unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
853 	unsigned long nr_pages = size >> PAGE_SHIFT;
854 	int ret;
855 
856 	last_mapped_pfn = init_memory_mapping(start, start + size);
857 	if (last_mapped_pfn > max_pfn_mapped)
858 		max_pfn_mapped = last_mapped_pfn;
859 
860 	ret = __add_pages(zone, start_pfn, nr_pages);
861 	WARN_ON_ONCE(ret);
862 
863 	return ret;
864 }
865 EXPORT_SYMBOL_GPL(arch_add_memory);
866 
867 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
868 int memory_add_physaddr_to_nid(u64 start)
869 {
870 	return 0;
871 }
872 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
873 #endif
874 
875 #endif /* CONFIG_MEMORY_HOTPLUG */
876 
877 /*
878  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
879  * is valid. The argument is a physical page number.
880  *
881  *
882  * On x86, access has to be given to the first megabyte of ram because that area
883  * contains bios code and data regions used by X and dosemu and similar apps.
884  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
885  * mmio resources as well as potential bios/acpi data regions.
886  */
887 int devmem_is_allowed(unsigned long pagenr)
888 {
889 	if (pagenr <= 256)
890 		return 1;
891 	if (!page_is_ram(pagenr))
892 		return 1;
893 	return 0;
894 }
895 
896 
897 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
898 			 kcore_modules, kcore_vsyscall;
899 
900 void __init mem_init(void)
901 {
902 	long codesize, reservedpages, datasize, initsize;
903 	unsigned long absent_pages;
904 
905 	start_periodic_check_for_corruption();
906 
907 	pci_iommu_alloc();
908 
909 	/* clear_bss() already clear the empty_zero_page */
910 
911 	reservedpages = 0;
912 
913 	/* this will put all low memory onto the freelists */
914 #ifdef CONFIG_NUMA
915 	totalram_pages = numa_free_all_bootmem();
916 #else
917 	totalram_pages = free_all_bootmem();
918 #endif
919 
920 	absent_pages = absent_pages_in_range(0, max_pfn);
921 	reservedpages = max_pfn - totalram_pages - absent_pages;
922 	after_bootmem = 1;
923 
924 	codesize =  (unsigned long) &_etext - (unsigned long) &_text;
925 	datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
926 	initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
927 
928 	/* Register memory areas for /proc/kcore */
929 	kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
930 	kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
931 		   VMALLOC_END-VMALLOC_START);
932 	kclist_add(&kcore_kernel, &_stext, _end - _stext);
933 	kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
934 	kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
935 				 VSYSCALL_END - VSYSCALL_START);
936 
937 	printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
938 			 "%ldk absent, %ldk reserved, %ldk data, %ldk init)\n",
939 		(unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
940 		max_pfn << (PAGE_SHIFT-10),
941 		codesize >> 10,
942 		absent_pages << (PAGE_SHIFT-10),
943 		reservedpages << (PAGE_SHIFT-10),
944 		datasize >> 10,
945 		initsize >> 10);
946 }
947 
948 void free_init_pages(char *what, unsigned long begin, unsigned long end)
949 {
950 	unsigned long addr = begin;
951 
952 	if (addr >= end)
953 		return;
954 
955 	/*
956 	 * If debugging page accesses then do not free this memory but
957 	 * mark them not present - any buggy init-section access will
958 	 * create a kernel page fault:
959 	 */
960 #ifdef CONFIG_DEBUG_PAGEALLOC
961 	printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
962 		begin, PAGE_ALIGN(end));
963 	set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
964 #else
965 	printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
966 
967 	for (; addr < end; addr += PAGE_SIZE) {
968 		ClearPageReserved(virt_to_page(addr));
969 		init_page_count(virt_to_page(addr));
970 		memset((void *)(addr & ~(PAGE_SIZE-1)),
971 			POISON_FREE_INITMEM, PAGE_SIZE);
972 		free_page(addr);
973 		totalram_pages++;
974 	}
975 #endif
976 }
977 
978 void free_initmem(void)
979 {
980 	free_init_pages("unused kernel memory",
981 			(unsigned long)(&__init_begin),
982 			(unsigned long)(&__init_end));
983 }
984 
985 #ifdef CONFIG_DEBUG_RODATA
986 const int rodata_test_data = 0xC3;
987 EXPORT_SYMBOL_GPL(rodata_test_data);
988 
989 void mark_rodata_ro(void)
990 {
991 	unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
992 	unsigned long rodata_start =
993 		((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
994 
995 #ifdef CONFIG_DYNAMIC_FTRACE
996 	/* Dynamic tracing modifies the kernel text section */
997 	start = rodata_start;
998 #endif
999 
1000 	printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1001 	       (end - start) >> 10);
1002 	set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1003 
1004 	/*
1005 	 * The rodata section (but not the kernel text!) should also be
1006 	 * not-executable.
1007 	 */
1008 	set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
1009 
1010 	rodata_test();
1011 
1012 #ifdef CONFIG_CPA_DEBUG
1013 	printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1014 	set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1015 
1016 	printk(KERN_INFO "Testing CPA: again\n");
1017 	set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1018 #endif
1019 }
1020 
1021 #endif
1022 
1023 #ifdef CONFIG_BLK_DEV_INITRD
1024 void free_initrd_mem(unsigned long start, unsigned long end)
1025 {
1026 	free_init_pages("initrd memory", start, end);
1027 }
1028 #endif
1029 
1030 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
1031 				   int flags)
1032 {
1033 #ifdef CONFIG_NUMA
1034 	int nid, next_nid;
1035 	int ret;
1036 #endif
1037 	unsigned long pfn = phys >> PAGE_SHIFT;
1038 
1039 	if (pfn >= max_pfn) {
1040 		/*
1041 		 * This can happen with kdump kernels when accessing
1042 		 * firmware tables:
1043 		 */
1044 		if (pfn < max_pfn_mapped)
1045 			return -EFAULT;
1046 
1047 		printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
1048 				phys, len);
1049 		return -EFAULT;
1050 	}
1051 
1052 	/* Should check here against the e820 map to avoid double free */
1053 #ifdef CONFIG_NUMA
1054 	nid = phys_to_nid(phys);
1055 	next_nid = phys_to_nid(phys + len - 1);
1056 	if (nid == next_nid)
1057 		ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
1058 	else
1059 		ret = reserve_bootmem(phys, len, flags);
1060 
1061 	if (ret != 0)
1062 		return ret;
1063 
1064 #else
1065 	reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
1066 #endif
1067 
1068 	if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
1069 		dma_reserve += len / PAGE_SIZE;
1070 		set_dma_reserve(dma_reserve);
1071 	}
1072 
1073 	return 0;
1074 }
1075 
1076 int kern_addr_valid(unsigned long addr)
1077 {
1078 	unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1079 	pgd_t *pgd;
1080 	pud_t *pud;
1081 	pmd_t *pmd;
1082 	pte_t *pte;
1083 
1084 	if (above != 0 && above != -1UL)
1085 		return 0;
1086 
1087 	pgd = pgd_offset_k(addr);
1088 	if (pgd_none(*pgd))
1089 		return 0;
1090 
1091 	pud = pud_offset(pgd, addr);
1092 	if (pud_none(*pud))
1093 		return 0;
1094 
1095 	pmd = pmd_offset(pud, addr);
1096 	if (pmd_none(*pmd))
1097 		return 0;
1098 
1099 	if (pmd_large(*pmd))
1100 		return pfn_valid(pmd_pfn(*pmd));
1101 
1102 	pte = pte_offset_kernel(pmd, addr);
1103 	if (pte_none(*pte))
1104 		return 0;
1105 
1106 	return pfn_valid(pte_pfn(*pte));
1107 }
1108 
1109 /*
1110  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
1111  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1112  * not need special handling anymore:
1113  */
1114 static struct vm_area_struct gate_vma = {
1115 	.vm_start	= VSYSCALL_START,
1116 	.vm_end		= VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
1117 	.vm_page_prot	= PAGE_READONLY_EXEC,
1118 	.vm_flags	= VM_READ | VM_EXEC
1119 };
1120 
1121 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1122 {
1123 #ifdef CONFIG_IA32_EMULATION
1124 	if (test_tsk_thread_flag(tsk, TIF_IA32))
1125 		return NULL;
1126 #endif
1127 	return &gate_vma;
1128 }
1129 
1130 int in_gate_area(struct task_struct *task, unsigned long addr)
1131 {
1132 	struct vm_area_struct *vma = get_gate_vma(task);
1133 
1134 	if (!vma)
1135 		return 0;
1136 
1137 	return (addr >= vma->vm_start) && (addr < vma->vm_end);
1138 }
1139 
1140 /*
1141  * Use this when you have no reliable task/vma, typically from interrupt
1142  * context. It is less reliable than using the task's vma and may give
1143  * false positives:
1144  */
1145 int in_gate_area_no_task(unsigned long addr)
1146 {
1147 	return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1148 }
1149 
1150 const char *arch_vma_name(struct vm_area_struct *vma)
1151 {
1152 	if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1153 		return "[vdso]";
1154 	if (vma == &gate_vma)
1155 		return "[vsyscall]";
1156 	return NULL;
1157 }
1158 
1159 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1160 /*
1161  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1162  */
1163 static long __meminitdata addr_start, addr_end;
1164 static void __meminitdata *p_start, *p_end;
1165 static int __meminitdata node_start;
1166 
1167 int __meminit
1168 vmemmap_populate(struct page *start_page, unsigned long size, int node)
1169 {
1170 	unsigned long addr = (unsigned long)start_page;
1171 	unsigned long end = (unsigned long)(start_page + size);
1172 	unsigned long next;
1173 	pgd_t *pgd;
1174 	pud_t *pud;
1175 	pmd_t *pmd;
1176 
1177 	for (; addr < end; addr = next) {
1178 		void *p = NULL;
1179 
1180 		pgd = vmemmap_pgd_populate(addr, node);
1181 		if (!pgd)
1182 			return -ENOMEM;
1183 
1184 		pud = vmemmap_pud_populate(pgd, addr, node);
1185 		if (!pud)
1186 			return -ENOMEM;
1187 
1188 		if (!cpu_has_pse) {
1189 			next = (addr + PAGE_SIZE) & PAGE_MASK;
1190 			pmd = vmemmap_pmd_populate(pud, addr, node);
1191 
1192 			if (!pmd)
1193 				return -ENOMEM;
1194 
1195 			p = vmemmap_pte_populate(pmd, addr, node);
1196 
1197 			if (!p)
1198 				return -ENOMEM;
1199 
1200 			addr_end = addr + PAGE_SIZE;
1201 			p_end = p + PAGE_SIZE;
1202 		} else {
1203 			next = pmd_addr_end(addr, end);
1204 
1205 			pmd = pmd_offset(pud, addr);
1206 			if (pmd_none(*pmd)) {
1207 				pte_t entry;
1208 
1209 				p = vmemmap_alloc_block(PMD_SIZE, node);
1210 				if (!p)
1211 					return -ENOMEM;
1212 
1213 				entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1214 						PAGE_KERNEL_LARGE);
1215 				set_pmd(pmd, __pmd(pte_val(entry)));
1216 
1217 				/* check to see if we have contiguous blocks */
1218 				if (p_end != p || node_start != node) {
1219 					if (p_start)
1220 						printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1221 						       addr_start, addr_end-1, p_start, p_end-1, node_start);
1222 					addr_start = addr;
1223 					node_start = node;
1224 					p_start = p;
1225 				}
1226 
1227 				addr_end = addr + PMD_SIZE;
1228 				p_end = p + PMD_SIZE;
1229 			} else
1230 				vmemmap_verify((pte_t *)pmd, node, addr, next);
1231 		}
1232 
1233 	}
1234 	return 0;
1235 }
1236 
1237 void __meminit vmemmap_populate_print_last(void)
1238 {
1239 	if (p_start) {
1240 		printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1241 			addr_start, addr_end-1, p_start, p_end-1, node_start);
1242 		p_start = NULL;
1243 		p_end = NULL;
1244 		node_start = 0;
1245 	}
1246 }
1247 #endif
1248