xref: /openbmc/linux/arch/x86/mm/init_64.c (revision 174cd4b1)
1 /*
2  *  linux/arch/x86_64/mm/init.c
3  *
4  *  Copyright (C) 1995  Linus Torvalds
5  *  Copyright (C) 2000  Pavel Machek <pavel@ucw.cz>
6  *  Copyright (C) 2002,2003 Andi Kleen <ak@suse.de>
7  */
8 
9 #include <linux/signal.h>
10 #include <linux/sched.h>
11 #include <linux/kernel.h>
12 #include <linux/errno.h>
13 #include <linux/string.h>
14 #include <linux/types.h>
15 #include <linux/ptrace.h>
16 #include <linux/mman.h>
17 #include <linux/mm.h>
18 #include <linux/swap.h>
19 #include <linux/smp.h>
20 #include <linux/init.h>
21 #include <linux/initrd.h>
22 #include <linux/pagemap.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/proc_fs.h>
26 #include <linux/pci.h>
27 #include <linux/pfn.h>
28 #include <linux/poison.h>
29 #include <linux/dma-mapping.h>
30 #include <linux/memory.h>
31 #include <linux/memory_hotplug.h>
32 #include <linux/memremap.h>
33 #include <linux/nmi.h>
34 #include <linux/gfp.h>
35 #include <linux/kcore.h>
36 
37 #include <asm/processor.h>
38 #include <asm/bios_ebda.h>
39 #include <linux/uaccess.h>
40 #include <asm/pgtable.h>
41 #include <asm/pgalloc.h>
42 #include <asm/dma.h>
43 #include <asm/fixmap.h>
44 #include <asm/e820.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 
60 #include "ident_map.c"
61 
62 /*
63  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
64  * physical space so we can cache the place of the first one and move
65  * around without checking the pgd every time.
66  */
67 
68 pteval_t __supported_pte_mask __read_mostly = ~0;
69 EXPORT_SYMBOL_GPL(__supported_pte_mask);
70 
71 int force_personality32;
72 
73 /*
74  * noexec32=on|off
75  * Control non executable heap for 32bit processes.
76  * To control the stack too use noexec=off
77  *
78  * on	PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
79  * off	PROT_READ implies PROT_EXEC
80  */
81 static int __init nonx32_setup(char *str)
82 {
83 	if (!strcmp(str, "on"))
84 		force_personality32 &= ~READ_IMPLIES_EXEC;
85 	else if (!strcmp(str, "off"))
86 		force_personality32 |= READ_IMPLIES_EXEC;
87 	return 1;
88 }
89 __setup("noexec32=", nonx32_setup);
90 
91 /*
92  * When memory was added make sure all the processes MM have
93  * suitable PGD entries in the local PGD level page.
94  */
95 void sync_global_pgds(unsigned long start, unsigned long end)
96 {
97 	unsigned long address;
98 
99 	for (address = start; address <= end; address += PGDIR_SIZE) {
100 		const pgd_t *pgd_ref = pgd_offset_k(address);
101 		struct page *page;
102 
103 		if (pgd_none(*pgd_ref))
104 			continue;
105 
106 		spin_lock(&pgd_lock);
107 		list_for_each_entry(page, &pgd_list, lru) {
108 			pgd_t *pgd;
109 			spinlock_t *pgt_lock;
110 
111 			pgd = (pgd_t *)page_address(page) + pgd_index(address);
112 			/* the pgt_lock only for Xen */
113 			pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
114 			spin_lock(pgt_lock);
115 
116 			if (!pgd_none(*pgd_ref) && !pgd_none(*pgd))
117 				BUG_ON(pgd_page_vaddr(*pgd)
118 				       != pgd_page_vaddr(*pgd_ref));
119 
120 			if (pgd_none(*pgd))
121 				set_pgd(pgd, *pgd_ref);
122 
123 			spin_unlock(pgt_lock);
124 		}
125 		spin_unlock(&pgd_lock);
126 	}
127 }
128 
129 /*
130  * NOTE: This function is marked __ref because it calls __init function
131  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
132  */
133 static __ref void *spp_getpage(void)
134 {
135 	void *ptr;
136 
137 	if (after_bootmem)
138 		ptr = (void *) get_zeroed_page(GFP_ATOMIC | __GFP_NOTRACK);
139 	else
140 		ptr = alloc_bootmem_pages(PAGE_SIZE);
141 
142 	if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
143 		panic("set_pte_phys: cannot allocate page data %s\n",
144 			after_bootmem ? "after bootmem" : "");
145 	}
146 
147 	pr_debug("spp_getpage %p\n", ptr);
148 
149 	return ptr;
150 }
151 
152 static pud_t *fill_pud(pgd_t *pgd, unsigned long vaddr)
153 {
154 	if (pgd_none(*pgd)) {
155 		pud_t *pud = (pud_t *)spp_getpage();
156 		pgd_populate(&init_mm, pgd, pud);
157 		if (pud != pud_offset(pgd, 0))
158 			printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
159 			       pud, pud_offset(pgd, 0));
160 	}
161 	return pud_offset(pgd, vaddr);
162 }
163 
164 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
165 {
166 	if (pud_none(*pud)) {
167 		pmd_t *pmd = (pmd_t *) spp_getpage();
168 		pud_populate(&init_mm, pud, pmd);
169 		if (pmd != pmd_offset(pud, 0))
170 			printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
171 			       pmd, pmd_offset(pud, 0));
172 	}
173 	return pmd_offset(pud, vaddr);
174 }
175 
176 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
177 {
178 	if (pmd_none(*pmd)) {
179 		pte_t *pte = (pte_t *) spp_getpage();
180 		pmd_populate_kernel(&init_mm, pmd, pte);
181 		if (pte != pte_offset_kernel(pmd, 0))
182 			printk(KERN_ERR "PAGETABLE BUG #02!\n");
183 	}
184 	return pte_offset_kernel(pmd, vaddr);
185 }
186 
187 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
188 {
189 	pud_t *pud;
190 	pmd_t *pmd;
191 	pte_t *pte;
192 
193 	pud = pud_page + pud_index(vaddr);
194 	pmd = fill_pmd(pud, vaddr);
195 	pte = fill_pte(pmd, vaddr);
196 
197 	set_pte(pte, new_pte);
198 
199 	/*
200 	 * It's enough to flush this one mapping.
201 	 * (PGE mappings get flushed as well)
202 	 */
203 	__flush_tlb_one(vaddr);
204 }
205 
206 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
207 {
208 	pgd_t *pgd;
209 	pud_t *pud_page;
210 
211 	pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
212 
213 	pgd = pgd_offset_k(vaddr);
214 	if (pgd_none(*pgd)) {
215 		printk(KERN_ERR
216 			"PGD FIXMAP MISSING, it should be setup in head.S!\n");
217 		return;
218 	}
219 	pud_page = (pud_t*)pgd_page_vaddr(*pgd);
220 	set_pte_vaddr_pud(pud_page, vaddr, pteval);
221 }
222 
223 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
224 {
225 	pgd_t *pgd;
226 	pud_t *pud;
227 
228 	pgd = pgd_offset_k(vaddr);
229 	pud = fill_pud(pgd, vaddr);
230 	return fill_pmd(pud, vaddr);
231 }
232 
233 pte_t * __init populate_extra_pte(unsigned long vaddr)
234 {
235 	pmd_t *pmd;
236 
237 	pmd = populate_extra_pmd(vaddr);
238 	return fill_pte(pmd, vaddr);
239 }
240 
241 /*
242  * Create large page table mappings for a range of physical addresses.
243  */
244 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
245 					enum page_cache_mode cache)
246 {
247 	pgd_t *pgd;
248 	pud_t *pud;
249 	pmd_t *pmd;
250 	pgprot_t prot;
251 
252 	pgprot_val(prot) = pgprot_val(PAGE_KERNEL_LARGE) |
253 		pgprot_val(pgprot_4k_2_large(cachemode2pgprot(cache)));
254 	BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
255 	for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
256 		pgd = pgd_offset_k((unsigned long)__va(phys));
257 		if (pgd_none(*pgd)) {
258 			pud = (pud_t *) spp_getpage();
259 			set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
260 						_PAGE_USER));
261 		}
262 		pud = pud_offset(pgd, (unsigned long)__va(phys));
263 		if (pud_none(*pud)) {
264 			pmd = (pmd_t *) spp_getpage();
265 			set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
266 						_PAGE_USER));
267 		}
268 		pmd = pmd_offset(pud, phys);
269 		BUG_ON(!pmd_none(*pmd));
270 		set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
271 	}
272 }
273 
274 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
275 {
276 	__init_extra_mapping(phys, size, _PAGE_CACHE_MODE_WB);
277 }
278 
279 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
280 {
281 	__init_extra_mapping(phys, size, _PAGE_CACHE_MODE_UC);
282 }
283 
284 /*
285  * The head.S code sets up the kernel high mapping:
286  *
287  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
288  *
289  * phys_base holds the negative offset to the kernel, which is added
290  * to the compile time generated pmds. This results in invalid pmds up
291  * to the point where we hit the physaddr 0 mapping.
292  *
293  * We limit the mappings to the region from _text to _brk_end.  _brk_end
294  * is rounded up to the 2MB boundary. This catches the invalid pmds as
295  * well, as they are located before _text:
296  */
297 void __init cleanup_highmap(void)
298 {
299 	unsigned long vaddr = __START_KERNEL_map;
300 	unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
301 	unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
302 	pmd_t *pmd = level2_kernel_pgt;
303 
304 	/*
305 	 * Native path, max_pfn_mapped is not set yet.
306 	 * Xen has valid max_pfn_mapped set in
307 	 *	arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
308 	 */
309 	if (max_pfn_mapped)
310 		vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
311 
312 	for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
313 		if (pmd_none(*pmd))
314 			continue;
315 		if (vaddr < (unsigned long) _text || vaddr > end)
316 			set_pmd(pmd, __pmd(0));
317 	}
318 }
319 
320 /*
321  * Create PTE level page table mapping for physical addresses.
322  * It returns the last physical address mapped.
323  */
324 static unsigned long __meminit
325 phys_pte_init(pte_t *pte_page, unsigned long paddr, unsigned long paddr_end,
326 	      pgprot_t prot)
327 {
328 	unsigned long pages = 0, paddr_next;
329 	unsigned long paddr_last = paddr_end;
330 	pte_t *pte;
331 	int i;
332 
333 	pte = pte_page + pte_index(paddr);
334 	i = pte_index(paddr);
335 
336 	for (; i < PTRS_PER_PTE; i++, paddr = paddr_next, pte++) {
337 		paddr_next = (paddr & PAGE_MASK) + PAGE_SIZE;
338 		if (paddr >= paddr_end) {
339 			if (!after_bootmem &&
340 			    !e820_any_mapped(paddr & PAGE_MASK, paddr_next,
341 					     E820_RAM) &&
342 			    !e820_any_mapped(paddr & PAGE_MASK, paddr_next,
343 					     E820_RESERVED_KERN))
344 				set_pte(pte, __pte(0));
345 			continue;
346 		}
347 
348 		/*
349 		 * We will re-use the existing mapping.
350 		 * Xen for example has some special requirements, like mapping
351 		 * pagetable pages as RO. So assume someone who pre-setup
352 		 * these mappings are more intelligent.
353 		 */
354 		if (!pte_none(*pte)) {
355 			if (!after_bootmem)
356 				pages++;
357 			continue;
358 		}
359 
360 		if (0)
361 			pr_info("   pte=%p addr=%lx pte=%016lx\n", pte, paddr,
362 				pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL).pte);
363 		pages++;
364 		set_pte(pte, pfn_pte(paddr >> PAGE_SHIFT, prot));
365 		paddr_last = (paddr & PAGE_MASK) + PAGE_SIZE;
366 	}
367 
368 	update_page_count(PG_LEVEL_4K, pages);
369 
370 	return paddr_last;
371 }
372 
373 /*
374  * Create PMD level page table mapping for physical addresses. The virtual
375  * and physical address have to be aligned at this level.
376  * It returns the last physical address mapped.
377  */
378 static unsigned long __meminit
379 phys_pmd_init(pmd_t *pmd_page, unsigned long paddr, unsigned long paddr_end,
380 	      unsigned long page_size_mask, pgprot_t prot)
381 {
382 	unsigned long pages = 0, paddr_next;
383 	unsigned long paddr_last = paddr_end;
384 
385 	int i = pmd_index(paddr);
386 
387 	for (; i < PTRS_PER_PMD; i++, paddr = paddr_next) {
388 		pmd_t *pmd = pmd_page + pmd_index(paddr);
389 		pte_t *pte;
390 		pgprot_t new_prot = prot;
391 
392 		paddr_next = (paddr & PMD_MASK) + PMD_SIZE;
393 		if (paddr >= paddr_end) {
394 			if (!after_bootmem &&
395 			    !e820_any_mapped(paddr & PMD_MASK, paddr_next,
396 					     E820_RAM) &&
397 			    !e820_any_mapped(paddr & PMD_MASK, paddr_next,
398 					     E820_RESERVED_KERN))
399 				set_pmd(pmd, __pmd(0));
400 			continue;
401 		}
402 
403 		if (!pmd_none(*pmd)) {
404 			if (!pmd_large(*pmd)) {
405 				spin_lock(&init_mm.page_table_lock);
406 				pte = (pte_t *)pmd_page_vaddr(*pmd);
407 				paddr_last = phys_pte_init(pte, paddr,
408 							   paddr_end, prot);
409 				spin_unlock(&init_mm.page_table_lock);
410 				continue;
411 			}
412 			/*
413 			 * If we are ok with PG_LEVEL_2M mapping, then we will
414 			 * use the existing mapping,
415 			 *
416 			 * Otherwise, we will split the large page mapping but
417 			 * use the same existing protection bits except for
418 			 * large page, so that we don't violate Intel's TLB
419 			 * Application note (317080) which says, while changing
420 			 * the page sizes, new and old translations should
421 			 * not differ with respect to page frame and
422 			 * attributes.
423 			 */
424 			if (page_size_mask & (1 << PG_LEVEL_2M)) {
425 				if (!after_bootmem)
426 					pages++;
427 				paddr_last = paddr_next;
428 				continue;
429 			}
430 			new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
431 		}
432 
433 		if (page_size_mask & (1<<PG_LEVEL_2M)) {
434 			pages++;
435 			spin_lock(&init_mm.page_table_lock);
436 			set_pte((pte_t *)pmd,
437 				pfn_pte((paddr & PMD_MASK) >> PAGE_SHIFT,
438 					__pgprot(pgprot_val(prot) | _PAGE_PSE)));
439 			spin_unlock(&init_mm.page_table_lock);
440 			paddr_last = paddr_next;
441 			continue;
442 		}
443 
444 		pte = alloc_low_page();
445 		paddr_last = phys_pte_init(pte, paddr, paddr_end, new_prot);
446 
447 		spin_lock(&init_mm.page_table_lock);
448 		pmd_populate_kernel(&init_mm, pmd, pte);
449 		spin_unlock(&init_mm.page_table_lock);
450 	}
451 	update_page_count(PG_LEVEL_2M, pages);
452 	return paddr_last;
453 }
454 
455 /*
456  * Create PUD level page table mapping for physical addresses. The virtual
457  * and physical address do not have to be aligned at this level. KASLR can
458  * randomize virtual addresses up to this level.
459  * It returns the last physical address mapped.
460  */
461 static unsigned long __meminit
462 phys_pud_init(pud_t *pud_page, unsigned long paddr, unsigned long paddr_end,
463 	      unsigned long page_size_mask)
464 {
465 	unsigned long pages = 0, paddr_next;
466 	unsigned long paddr_last = paddr_end;
467 	unsigned long vaddr = (unsigned long)__va(paddr);
468 	int i = pud_index(vaddr);
469 
470 	for (; i < PTRS_PER_PUD; i++, paddr = paddr_next) {
471 		pud_t *pud;
472 		pmd_t *pmd;
473 		pgprot_t prot = PAGE_KERNEL;
474 
475 		vaddr = (unsigned long)__va(paddr);
476 		pud = pud_page + pud_index(vaddr);
477 		paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
478 
479 		if (paddr >= paddr_end) {
480 			if (!after_bootmem &&
481 			    !e820_any_mapped(paddr & PUD_MASK, paddr_next,
482 					     E820_RAM) &&
483 			    !e820_any_mapped(paddr & PUD_MASK, paddr_next,
484 					     E820_RESERVED_KERN))
485 				set_pud(pud, __pud(0));
486 			continue;
487 		}
488 
489 		if (!pud_none(*pud)) {
490 			if (!pud_large(*pud)) {
491 				pmd = pmd_offset(pud, 0);
492 				paddr_last = phys_pmd_init(pmd, paddr,
493 							   paddr_end,
494 							   page_size_mask,
495 							   prot);
496 				__flush_tlb_all();
497 				continue;
498 			}
499 			/*
500 			 * If we are ok with PG_LEVEL_1G mapping, then we will
501 			 * use the existing mapping.
502 			 *
503 			 * Otherwise, we will split the gbpage mapping but use
504 			 * the same existing protection  bits except for large
505 			 * page, so that we don't violate Intel's TLB
506 			 * Application note (317080) which says, while changing
507 			 * the page sizes, new and old translations should
508 			 * not differ with respect to page frame and
509 			 * attributes.
510 			 */
511 			if (page_size_mask & (1 << PG_LEVEL_1G)) {
512 				if (!after_bootmem)
513 					pages++;
514 				paddr_last = paddr_next;
515 				continue;
516 			}
517 			prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
518 		}
519 
520 		if (page_size_mask & (1<<PG_LEVEL_1G)) {
521 			pages++;
522 			spin_lock(&init_mm.page_table_lock);
523 			set_pte((pte_t *)pud,
524 				pfn_pte((paddr & PUD_MASK) >> PAGE_SHIFT,
525 					PAGE_KERNEL_LARGE));
526 			spin_unlock(&init_mm.page_table_lock);
527 			paddr_last = paddr_next;
528 			continue;
529 		}
530 
531 		pmd = alloc_low_page();
532 		paddr_last = phys_pmd_init(pmd, paddr, paddr_end,
533 					   page_size_mask, prot);
534 
535 		spin_lock(&init_mm.page_table_lock);
536 		pud_populate(&init_mm, pud, pmd);
537 		spin_unlock(&init_mm.page_table_lock);
538 	}
539 	__flush_tlb_all();
540 
541 	update_page_count(PG_LEVEL_1G, pages);
542 
543 	return paddr_last;
544 }
545 
546 /*
547  * Create page table mapping for the physical memory for specific physical
548  * addresses. The virtual and physical addresses have to be aligned on PMD level
549  * down. It returns the last physical address mapped.
550  */
551 unsigned long __meminit
552 kernel_physical_mapping_init(unsigned long paddr_start,
553 			     unsigned long paddr_end,
554 			     unsigned long page_size_mask)
555 {
556 	bool pgd_changed = false;
557 	unsigned long vaddr, vaddr_start, vaddr_end, vaddr_next, paddr_last;
558 
559 	paddr_last = paddr_end;
560 	vaddr = (unsigned long)__va(paddr_start);
561 	vaddr_end = (unsigned long)__va(paddr_end);
562 	vaddr_start = vaddr;
563 
564 	for (; vaddr < vaddr_end; vaddr = vaddr_next) {
565 		pgd_t *pgd = pgd_offset_k(vaddr);
566 		pud_t *pud;
567 
568 		vaddr_next = (vaddr & PGDIR_MASK) + PGDIR_SIZE;
569 
570 		if (pgd_val(*pgd)) {
571 			pud = (pud_t *)pgd_page_vaddr(*pgd);
572 			paddr_last = phys_pud_init(pud, __pa(vaddr),
573 						   __pa(vaddr_end),
574 						   page_size_mask);
575 			continue;
576 		}
577 
578 		pud = alloc_low_page();
579 		paddr_last = phys_pud_init(pud, __pa(vaddr), __pa(vaddr_end),
580 					   page_size_mask);
581 
582 		spin_lock(&init_mm.page_table_lock);
583 		pgd_populate(&init_mm, pgd, pud);
584 		spin_unlock(&init_mm.page_table_lock);
585 		pgd_changed = true;
586 	}
587 
588 	if (pgd_changed)
589 		sync_global_pgds(vaddr_start, vaddr_end - 1);
590 
591 	__flush_tlb_all();
592 
593 	return paddr_last;
594 }
595 
596 #ifndef CONFIG_NUMA
597 void __init initmem_init(void)
598 {
599 	memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
600 }
601 #endif
602 
603 void __init paging_init(void)
604 {
605 	sparse_memory_present_with_active_regions(MAX_NUMNODES);
606 	sparse_init();
607 
608 	/*
609 	 * clear the default setting with node 0
610 	 * note: don't use nodes_clear here, that is really clearing when
611 	 *	 numa support is not compiled in, and later node_set_state
612 	 *	 will not set it back.
613 	 */
614 	node_clear_state(0, N_MEMORY);
615 	if (N_MEMORY != N_NORMAL_MEMORY)
616 		node_clear_state(0, N_NORMAL_MEMORY);
617 
618 	zone_sizes_init();
619 }
620 
621 /*
622  * Memory hotplug specific functions
623  */
624 #ifdef CONFIG_MEMORY_HOTPLUG
625 /*
626  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
627  * updating.
628  */
629 static void  update_end_of_memory_vars(u64 start, u64 size)
630 {
631 	unsigned long end_pfn = PFN_UP(start + size);
632 
633 	if (end_pfn > max_pfn) {
634 		max_pfn = end_pfn;
635 		max_low_pfn = end_pfn;
636 		high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
637 	}
638 }
639 
640 /*
641  * Memory is added always to NORMAL zone. This means you will never get
642  * additional DMA/DMA32 memory.
643  */
644 int arch_add_memory(int nid, u64 start, u64 size, bool for_device)
645 {
646 	struct pglist_data *pgdat = NODE_DATA(nid);
647 	struct zone *zone = pgdat->node_zones +
648 		zone_for_memory(nid, start, size, ZONE_NORMAL, for_device);
649 	unsigned long start_pfn = start >> PAGE_SHIFT;
650 	unsigned long nr_pages = size >> PAGE_SHIFT;
651 	int ret;
652 
653 	init_memory_mapping(start, start + size);
654 
655 	ret = __add_pages(nid, zone, start_pfn, nr_pages);
656 	WARN_ON_ONCE(ret);
657 
658 	/* update max_pfn, max_low_pfn and high_memory */
659 	update_end_of_memory_vars(start, size);
660 
661 	return ret;
662 }
663 EXPORT_SYMBOL_GPL(arch_add_memory);
664 
665 #define PAGE_INUSE 0xFD
666 
667 static void __meminit free_pagetable(struct page *page, int order)
668 {
669 	unsigned long magic;
670 	unsigned int nr_pages = 1 << order;
671 	struct vmem_altmap *altmap = to_vmem_altmap((unsigned long) page);
672 
673 	if (altmap) {
674 		vmem_altmap_free(altmap, nr_pages);
675 		return;
676 	}
677 
678 	/* bootmem page has reserved flag */
679 	if (PageReserved(page)) {
680 		__ClearPageReserved(page);
681 
682 		magic = (unsigned long)page->freelist;
683 		if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
684 			while (nr_pages--)
685 				put_page_bootmem(page++);
686 		} else
687 			while (nr_pages--)
688 				free_reserved_page(page++);
689 	} else
690 		free_pages((unsigned long)page_address(page), order);
691 }
692 
693 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
694 {
695 	pte_t *pte;
696 	int i;
697 
698 	for (i = 0; i < PTRS_PER_PTE; i++) {
699 		pte = pte_start + i;
700 		if (!pte_none(*pte))
701 			return;
702 	}
703 
704 	/* free a pte talbe */
705 	free_pagetable(pmd_page(*pmd), 0);
706 	spin_lock(&init_mm.page_table_lock);
707 	pmd_clear(pmd);
708 	spin_unlock(&init_mm.page_table_lock);
709 }
710 
711 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
712 {
713 	pmd_t *pmd;
714 	int i;
715 
716 	for (i = 0; i < PTRS_PER_PMD; i++) {
717 		pmd = pmd_start + i;
718 		if (!pmd_none(*pmd))
719 			return;
720 	}
721 
722 	/* free a pmd talbe */
723 	free_pagetable(pud_page(*pud), 0);
724 	spin_lock(&init_mm.page_table_lock);
725 	pud_clear(pud);
726 	spin_unlock(&init_mm.page_table_lock);
727 }
728 
729 static void __meminit
730 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
731 		 bool direct)
732 {
733 	unsigned long next, pages = 0;
734 	pte_t *pte;
735 	void *page_addr;
736 	phys_addr_t phys_addr;
737 
738 	pte = pte_start + pte_index(addr);
739 	for (; addr < end; addr = next, pte++) {
740 		next = (addr + PAGE_SIZE) & PAGE_MASK;
741 		if (next > end)
742 			next = end;
743 
744 		if (!pte_present(*pte))
745 			continue;
746 
747 		/*
748 		 * We mapped [0,1G) memory as identity mapping when
749 		 * initializing, in arch/x86/kernel/head_64.S. These
750 		 * pagetables cannot be removed.
751 		 */
752 		phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
753 		if (phys_addr < (phys_addr_t)0x40000000)
754 			return;
755 
756 		if (PAGE_ALIGNED(addr) && PAGE_ALIGNED(next)) {
757 			/*
758 			 * Do not free direct mapping pages since they were
759 			 * freed when offlining, or simplely not in use.
760 			 */
761 			if (!direct)
762 				free_pagetable(pte_page(*pte), 0);
763 
764 			spin_lock(&init_mm.page_table_lock);
765 			pte_clear(&init_mm, addr, pte);
766 			spin_unlock(&init_mm.page_table_lock);
767 
768 			/* For non-direct mapping, pages means nothing. */
769 			pages++;
770 		} else {
771 			/*
772 			 * If we are here, we are freeing vmemmap pages since
773 			 * direct mapped memory ranges to be freed are aligned.
774 			 *
775 			 * If we are not removing the whole page, it means
776 			 * other page structs in this page are being used and
777 			 * we canot remove them. So fill the unused page_structs
778 			 * with 0xFD, and remove the page when it is wholly
779 			 * filled with 0xFD.
780 			 */
781 			memset((void *)addr, PAGE_INUSE, next - addr);
782 
783 			page_addr = page_address(pte_page(*pte));
784 			if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) {
785 				free_pagetable(pte_page(*pte), 0);
786 
787 				spin_lock(&init_mm.page_table_lock);
788 				pte_clear(&init_mm, addr, pte);
789 				spin_unlock(&init_mm.page_table_lock);
790 			}
791 		}
792 	}
793 
794 	/* Call free_pte_table() in remove_pmd_table(). */
795 	flush_tlb_all();
796 	if (direct)
797 		update_page_count(PG_LEVEL_4K, -pages);
798 }
799 
800 static void __meminit
801 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
802 		 bool direct)
803 {
804 	unsigned long next, pages = 0;
805 	pte_t *pte_base;
806 	pmd_t *pmd;
807 	void *page_addr;
808 
809 	pmd = pmd_start + pmd_index(addr);
810 	for (; addr < end; addr = next, pmd++) {
811 		next = pmd_addr_end(addr, end);
812 
813 		if (!pmd_present(*pmd))
814 			continue;
815 
816 		if (pmd_large(*pmd)) {
817 			if (IS_ALIGNED(addr, PMD_SIZE) &&
818 			    IS_ALIGNED(next, PMD_SIZE)) {
819 				if (!direct)
820 					free_pagetable(pmd_page(*pmd),
821 						       get_order(PMD_SIZE));
822 
823 				spin_lock(&init_mm.page_table_lock);
824 				pmd_clear(pmd);
825 				spin_unlock(&init_mm.page_table_lock);
826 				pages++;
827 			} else {
828 				/* If here, we are freeing vmemmap pages. */
829 				memset((void *)addr, PAGE_INUSE, next - addr);
830 
831 				page_addr = page_address(pmd_page(*pmd));
832 				if (!memchr_inv(page_addr, PAGE_INUSE,
833 						PMD_SIZE)) {
834 					free_pagetable(pmd_page(*pmd),
835 						       get_order(PMD_SIZE));
836 
837 					spin_lock(&init_mm.page_table_lock);
838 					pmd_clear(pmd);
839 					spin_unlock(&init_mm.page_table_lock);
840 				}
841 			}
842 
843 			continue;
844 		}
845 
846 		pte_base = (pte_t *)pmd_page_vaddr(*pmd);
847 		remove_pte_table(pte_base, addr, next, direct);
848 		free_pte_table(pte_base, pmd);
849 	}
850 
851 	/* Call free_pmd_table() in remove_pud_table(). */
852 	if (direct)
853 		update_page_count(PG_LEVEL_2M, -pages);
854 }
855 
856 static void __meminit
857 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
858 		 bool direct)
859 {
860 	unsigned long next, pages = 0;
861 	pmd_t *pmd_base;
862 	pud_t *pud;
863 	void *page_addr;
864 
865 	pud = pud_start + pud_index(addr);
866 	for (; addr < end; addr = next, pud++) {
867 		next = pud_addr_end(addr, end);
868 
869 		if (!pud_present(*pud))
870 			continue;
871 
872 		if (pud_large(*pud)) {
873 			if (IS_ALIGNED(addr, PUD_SIZE) &&
874 			    IS_ALIGNED(next, PUD_SIZE)) {
875 				if (!direct)
876 					free_pagetable(pud_page(*pud),
877 						       get_order(PUD_SIZE));
878 
879 				spin_lock(&init_mm.page_table_lock);
880 				pud_clear(pud);
881 				spin_unlock(&init_mm.page_table_lock);
882 				pages++;
883 			} else {
884 				/* If here, we are freeing vmemmap pages. */
885 				memset((void *)addr, PAGE_INUSE, next - addr);
886 
887 				page_addr = page_address(pud_page(*pud));
888 				if (!memchr_inv(page_addr, PAGE_INUSE,
889 						PUD_SIZE)) {
890 					free_pagetable(pud_page(*pud),
891 						       get_order(PUD_SIZE));
892 
893 					spin_lock(&init_mm.page_table_lock);
894 					pud_clear(pud);
895 					spin_unlock(&init_mm.page_table_lock);
896 				}
897 			}
898 
899 			continue;
900 		}
901 
902 		pmd_base = (pmd_t *)pud_page_vaddr(*pud);
903 		remove_pmd_table(pmd_base, addr, next, direct);
904 		free_pmd_table(pmd_base, pud);
905 	}
906 
907 	if (direct)
908 		update_page_count(PG_LEVEL_1G, -pages);
909 }
910 
911 /* start and end are both virtual address. */
912 static void __meminit
913 remove_pagetable(unsigned long start, unsigned long end, bool direct)
914 {
915 	unsigned long next;
916 	unsigned long addr;
917 	pgd_t *pgd;
918 	pud_t *pud;
919 
920 	for (addr = start; addr < end; addr = next) {
921 		next = pgd_addr_end(addr, end);
922 
923 		pgd = pgd_offset_k(addr);
924 		if (!pgd_present(*pgd))
925 			continue;
926 
927 		pud = (pud_t *)pgd_page_vaddr(*pgd);
928 		remove_pud_table(pud, addr, next, direct);
929 	}
930 
931 	flush_tlb_all();
932 }
933 
934 void __ref vmemmap_free(unsigned long start, unsigned long end)
935 {
936 	remove_pagetable(start, end, false);
937 }
938 
939 #ifdef CONFIG_MEMORY_HOTREMOVE
940 static void __meminit
941 kernel_physical_mapping_remove(unsigned long start, unsigned long end)
942 {
943 	start = (unsigned long)__va(start);
944 	end = (unsigned long)__va(end);
945 
946 	remove_pagetable(start, end, true);
947 }
948 
949 int __ref arch_remove_memory(u64 start, u64 size)
950 {
951 	unsigned long start_pfn = start >> PAGE_SHIFT;
952 	unsigned long nr_pages = size >> PAGE_SHIFT;
953 	struct page *page = pfn_to_page(start_pfn);
954 	struct vmem_altmap *altmap;
955 	struct zone *zone;
956 	int ret;
957 
958 	/* With altmap the first mapped page is offset from @start */
959 	altmap = to_vmem_altmap((unsigned long) page);
960 	if (altmap)
961 		page += vmem_altmap_offset(altmap);
962 	zone = page_zone(page);
963 	ret = __remove_pages(zone, start_pfn, nr_pages);
964 	WARN_ON_ONCE(ret);
965 	kernel_physical_mapping_remove(start, start + size);
966 
967 	return ret;
968 }
969 #endif
970 #endif /* CONFIG_MEMORY_HOTPLUG */
971 
972 static struct kcore_list kcore_vsyscall;
973 
974 static void __init register_page_bootmem_info(void)
975 {
976 #ifdef CONFIG_NUMA
977 	int i;
978 
979 	for_each_online_node(i)
980 		register_page_bootmem_info_node(NODE_DATA(i));
981 #endif
982 }
983 
984 void __init mem_init(void)
985 {
986 	pci_iommu_alloc();
987 
988 	/* clear_bss() already clear the empty_zero_page */
989 
990 	register_page_bootmem_info();
991 
992 	/* this will put all memory onto the freelists */
993 	free_all_bootmem();
994 	after_bootmem = 1;
995 
996 	/* Register memory areas for /proc/kcore */
997 	kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR,
998 			 PAGE_SIZE, KCORE_OTHER);
999 
1000 	mem_init_print_info(NULL);
1001 }
1002 
1003 int kernel_set_to_readonly;
1004 
1005 void set_kernel_text_rw(void)
1006 {
1007 	unsigned long start = PFN_ALIGN(_text);
1008 	unsigned long end = PFN_ALIGN(__stop___ex_table);
1009 
1010 	if (!kernel_set_to_readonly)
1011 		return;
1012 
1013 	pr_debug("Set kernel text: %lx - %lx for read write\n",
1014 		 start, end);
1015 
1016 	/*
1017 	 * Make the kernel identity mapping for text RW. Kernel text
1018 	 * mapping will always be RO. Refer to the comment in
1019 	 * static_protections() in pageattr.c
1020 	 */
1021 	set_memory_rw(start, (end - start) >> PAGE_SHIFT);
1022 }
1023 
1024 void set_kernel_text_ro(void)
1025 {
1026 	unsigned long start = PFN_ALIGN(_text);
1027 	unsigned long end = PFN_ALIGN(__stop___ex_table);
1028 
1029 	if (!kernel_set_to_readonly)
1030 		return;
1031 
1032 	pr_debug("Set kernel text: %lx - %lx for read only\n",
1033 		 start, end);
1034 
1035 	/*
1036 	 * Set the kernel identity mapping for text RO.
1037 	 */
1038 	set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1039 }
1040 
1041 void mark_rodata_ro(void)
1042 {
1043 	unsigned long start = PFN_ALIGN(_text);
1044 	unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1045 	unsigned long end = (unsigned long) &__end_rodata_hpage_align;
1046 	unsigned long text_end = PFN_ALIGN(&__stop___ex_table);
1047 	unsigned long rodata_end = PFN_ALIGN(&__end_rodata);
1048 	unsigned long all_end;
1049 
1050 	printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1051 	       (end - start) >> 10);
1052 	set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1053 
1054 	kernel_set_to_readonly = 1;
1055 
1056 	/*
1057 	 * The rodata/data/bss/brk section (but not the kernel text!)
1058 	 * should also be not-executable.
1059 	 *
1060 	 * We align all_end to PMD_SIZE because the existing mapping
1061 	 * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1062 	 * split the PMD and the reminder between _brk_end and the end
1063 	 * of the PMD will remain mapped executable.
1064 	 *
1065 	 * Any PMD which was setup after the one which covers _brk_end
1066 	 * has been zapped already via cleanup_highmem().
1067 	 */
1068 	all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1069 	set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
1070 
1071 #ifdef CONFIG_CPA_DEBUG
1072 	printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1073 	set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1074 
1075 	printk(KERN_INFO "Testing CPA: again\n");
1076 	set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1077 #endif
1078 
1079 	free_init_pages("unused kernel",
1080 			(unsigned long) __va(__pa_symbol(text_end)),
1081 			(unsigned long) __va(__pa_symbol(rodata_start)));
1082 	free_init_pages("unused kernel",
1083 			(unsigned long) __va(__pa_symbol(rodata_end)),
1084 			(unsigned long) __va(__pa_symbol(_sdata)));
1085 
1086 	debug_checkwx();
1087 }
1088 
1089 int kern_addr_valid(unsigned long addr)
1090 {
1091 	unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1092 	pgd_t *pgd;
1093 	pud_t *pud;
1094 	pmd_t *pmd;
1095 	pte_t *pte;
1096 
1097 	if (above != 0 && above != -1UL)
1098 		return 0;
1099 
1100 	pgd = pgd_offset_k(addr);
1101 	if (pgd_none(*pgd))
1102 		return 0;
1103 
1104 	pud = pud_offset(pgd, addr);
1105 	if (pud_none(*pud))
1106 		return 0;
1107 
1108 	if (pud_large(*pud))
1109 		return pfn_valid(pud_pfn(*pud));
1110 
1111 	pmd = pmd_offset(pud, addr);
1112 	if (pmd_none(*pmd))
1113 		return 0;
1114 
1115 	if (pmd_large(*pmd))
1116 		return pfn_valid(pmd_pfn(*pmd));
1117 
1118 	pte = pte_offset_kernel(pmd, addr);
1119 	if (pte_none(*pte))
1120 		return 0;
1121 
1122 	return pfn_valid(pte_pfn(*pte));
1123 }
1124 
1125 static unsigned long probe_memory_block_size(void)
1126 {
1127 	unsigned long bz = MIN_MEMORY_BLOCK_SIZE;
1128 
1129 	/* if system is UV or has 64GB of RAM or more, use large blocks */
1130 	if (is_uv_system() || ((max_pfn << PAGE_SHIFT) >= (64UL << 30)))
1131 		bz = 2UL << 30; /* 2GB */
1132 
1133 	pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20);
1134 
1135 	return bz;
1136 }
1137 
1138 static unsigned long memory_block_size_probed;
1139 unsigned long memory_block_size_bytes(void)
1140 {
1141 	if (!memory_block_size_probed)
1142 		memory_block_size_probed = probe_memory_block_size();
1143 
1144 	return memory_block_size_probed;
1145 }
1146 
1147 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1148 /*
1149  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1150  */
1151 static long __meminitdata addr_start, addr_end;
1152 static void __meminitdata *p_start, *p_end;
1153 static int __meminitdata node_start;
1154 
1155 static int __meminit vmemmap_populate_hugepages(unsigned long start,
1156 		unsigned long end, int node, struct vmem_altmap *altmap)
1157 {
1158 	unsigned long addr;
1159 	unsigned long next;
1160 	pgd_t *pgd;
1161 	pud_t *pud;
1162 	pmd_t *pmd;
1163 
1164 	for (addr = start; addr < end; addr = next) {
1165 		next = pmd_addr_end(addr, end);
1166 
1167 		pgd = vmemmap_pgd_populate(addr, node);
1168 		if (!pgd)
1169 			return -ENOMEM;
1170 
1171 		pud = vmemmap_pud_populate(pgd, addr, node);
1172 		if (!pud)
1173 			return -ENOMEM;
1174 
1175 		pmd = pmd_offset(pud, addr);
1176 		if (pmd_none(*pmd)) {
1177 			void *p;
1178 
1179 			p = __vmemmap_alloc_block_buf(PMD_SIZE, node, altmap);
1180 			if (p) {
1181 				pte_t entry;
1182 
1183 				entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1184 						PAGE_KERNEL_LARGE);
1185 				set_pmd(pmd, __pmd(pte_val(entry)));
1186 
1187 				/* check to see if we have contiguous blocks */
1188 				if (p_end != p || node_start != node) {
1189 					if (p_start)
1190 						pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1191 						       addr_start, addr_end-1, p_start, p_end-1, node_start);
1192 					addr_start = addr;
1193 					node_start = node;
1194 					p_start = p;
1195 				}
1196 
1197 				addr_end = addr + PMD_SIZE;
1198 				p_end = p + PMD_SIZE;
1199 				continue;
1200 			} else if (altmap)
1201 				return -ENOMEM; /* no fallback */
1202 		} else if (pmd_large(*pmd)) {
1203 			vmemmap_verify((pte_t *)pmd, node, addr, next);
1204 			continue;
1205 		}
1206 		pr_warn_once("vmemmap: falling back to regular page backing\n");
1207 		if (vmemmap_populate_basepages(addr, next, node))
1208 			return -ENOMEM;
1209 	}
1210 	return 0;
1211 }
1212 
1213 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
1214 {
1215 	struct vmem_altmap *altmap = to_vmem_altmap(start);
1216 	int err;
1217 
1218 	if (boot_cpu_has(X86_FEATURE_PSE))
1219 		err = vmemmap_populate_hugepages(start, end, node, altmap);
1220 	else if (altmap) {
1221 		pr_err_once("%s: no cpu support for altmap allocations\n",
1222 				__func__);
1223 		err = -ENOMEM;
1224 	} else
1225 		err = vmemmap_populate_basepages(start, end, node);
1226 	if (!err)
1227 		sync_global_pgds(start, end - 1);
1228 	return err;
1229 }
1230 
1231 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1232 void register_page_bootmem_memmap(unsigned long section_nr,
1233 				  struct page *start_page, unsigned long size)
1234 {
1235 	unsigned long addr = (unsigned long)start_page;
1236 	unsigned long end = (unsigned long)(start_page + size);
1237 	unsigned long next;
1238 	pgd_t *pgd;
1239 	pud_t *pud;
1240 	pmd_t *pmd;
1241 	unsigned int nr_pages;
1242 	struct page *page;
1243 
1244 	for (; addr < end; addr = next) {
1245 		pte_t *pte = NULL;
1246 
1247 		pgd = pgd_offset_k(addr);
1248 		if (pgd_none(*pgd)) {
1249 			next = (addr + PAGE_SIZE) & PAGE_MASK;
1250 			continue;
1251 		}
1252 		get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1253 
1254 		pud = pud_offset(pgd, addr);
1255 		if (pud_none(*pud)) {
1256 			next = (addr + PAGE_SIZE) & PAGE_MASK;
1257 			continue;
1258 		}
1259 		get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1260 
1261 		if (!boot_cpu_has(X86_FEATURE_PSE)) {
1262 			next = (addr + PAGE_SIZE) & PAGE_MASK;
1263 			pmd = pmd_offset(pud, addr);
1264 			if (pmd_none(*pmd))
1265 				continue;
1266 			get_page_bootmem(section_nr, pmd_page(*pmd),
1267 					 MIX_SECTION_INFO);
1268 
1269 			pte = pte_offset_kernel(pmd, addr);
1270 			if (pte_none(*pte))
1271 				continue;
1272 			get_page_bootmem(section_nr, pte_page(*pte),
1273 					 SECTION_INFO);
1274 		} else {
1275 			next = pmd_addr_end(addr, end);
1276 
1277 			pmd = pmd_offset(pud, addr);
1278 			if (pmd_none(*pmd))
1279 				continue;
1280 
1281 			nr_pages = 1 << (get_order(PMD_SIZE));
1282 			page = pmd_page(*pmd);
1283 			while (nr_pages--)
1284 				get_page_bootmem(section_nr, page++,
1285 						 SECTION_INFO);
1286 		}
1287 	}
1288 }
1289 #endif
1290 
1291 void __meminit vmemmap_populate_print_last(void)
1292 {
1293 	if (p_start) {
1294 		pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1295 			addr_start, addr_end-1, p_start, p_end-1, node_start);
1296 		p_start = NULL;
1297 		p_end = NULL;
1298 		node_start = 0;
1299 	}
1300 }
1301 #endif
1302