mm, memory_hotplug: add nid parameter to arch_remove_memory
[sfrench/cifs-2.6.git] / arch / x86 / mm / init_64.c
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/memblock.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/memory.h>
30 #include <linux/memory_hotplug.h>
31 #include <linux/memremap.h>
32 #include <linux/nmi.h>
33 #include <linux/gfp.h>
34 #include <linux/kcore.h>
35
36 #include <asm/processor.h>
37 #include <asm/bios_ebda.h>
38 #include <linux/uaccess.h>
39 #include <asm/pgtable.h>
40 #include <asm/pgalloc.h>
41 #include <asm/dma.h>
42 #include <asm/fixmap.h>
43 #include <asm/e820/api.h>
44 #include <asm/apic.h>
45 #include <asm/tlb.h>
46 #include <asm/mmu_context.h>
47 #include <asm/proto.h>
48 #include <asm/smp.h>
49 #include <asm/sections.h>
50 #include <asm/kdebug.h>
51 #include <asm/numa.h>
52 #include <asm/set_memory.h>
53 #include <asm/init.h>
54 #include <asm/uv/uv.h>
55 #include <asm/setup.h>
56
57 #include "mm_internal.h"
58
59 #include "ident_map.c"
60
61 /*
62  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
63  * physical space so we can cache the place of the first one and move
64  * around without checking the pgd every time.
65  */
66
67 /* Bits supported by the hardware: */
68 pteval_t __supported_pte_mask __read_mostly = ~0;
69 /* Bits allowed in normal kernel mappings: */
70 pteval_t __default_kernel_pte_mask __read_mostly = ~0;
71 EXPORT_SYMBOL_GPL(__supported_pte_mask);
72 /* Used in PAGE_KERNEL_* macros which are reasonably used out-of-tree: */
73 EXPORT_SYMBOL(__default_kernel_pte_mask);
74
75 int force_personality32;
76
77 /*
78  * noexec32=on|off
79  * Control non executable heap for 32bit processes.
80  * To control the stack too use noexec=off
81  *
82  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
83  * off  PROT_READ implies PROT_EXEC
84  */
85 static int __init nonx32_setup(char *str)
86 {
87         if (!strcmp(str, "on"))
88                 force_personality32 &= ~READ_IMPLIES_EXEC;
89         else if (!strcmp(str, "off"))
90                 force_personality32 |= READ_IMPLIES_EXEC;
91         return 1;
92 }
93 __setup("noexec32=", nonx32_setup);
94
95 static void sync_global_pgds_l5(unsigned long start, unsigned long end)
96 {
97         unsigned long addr;
98
99         for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
100                 const pgd_t *pgd_ref = pgd_offset_k(addr);
101                 struct page *page;
102
103                 /* Check for overflow */
104                 if (addr < start)
105                         break;
106
107                 if (pgd_none(*pgd_ref))
108                         continue;
109
110                 spin_lock(&pgd_lock);
111                 list_for_each_entry(page, &pgd_list, lru) {
112                         pgd_t *pgd;
113                         spinlock_t *pgt_lock;
114
115                         pgd = (pgd_t *)page_address(page) + pgd_index(addr);
116                         /* the pgt_lock only for Xen */
117                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
118                         spin_lock(pgt_lock);
119
120                         if (!pgd_none(*pgd_ref) && !pgd_none(*pgd))
121                                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
122
123                         if (pgd_none(*pgd))
124                                 set_pgd(pgd, *pgd_ref);
125
126                         spin_unlock(pgt_lock);
127                 }
128                 spin_unlock(&pgd_lock);
129         }
130 }
131
132 static void sync_global_pgds_l4(unsigned long start, unsigned long end)
133 {
134         unsigned long addr;
135
136         for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
137                 pgd_t *pgd_ref = pgd_offset_k(addr);
138                 const p4d_t *p4d_ref;
139                 struct page *page;
140
141                 /*
142                  * With folded p4d, pgd_none() is always false, we need to
143                  * handle synchonization on p4d level.
144                  */
145                 MAYBE_BUILD_BUG_ON(pgd_none(*pgd_ref));
146                 p4d_ref = p4d_offset(pgd_ref, addr);
147
148                 if (p4d_none(*p4d_ref))
149                         continue;
150
151                 spin_lock(&pgd_lock);
152                 list_for_each_entry(page, &pgd_list, lru) {
153                         pgd_t *pgd;
154                         p4d_t *p4d;
155                         spinlock_t *pgt_lock;
156
157                         pgd = (pgd_t *)page_address(page) + pgd_index(addr);
158                         p4d = p4d_offset(pgd, addr);
159                         /* the pgt_lock only for Xen */
160                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
161                         spin_lock(pgt_lock);
162
163                         if (!p4d_none(*p4d_ref) && !p4d_none(*p4d))
164                                 BUG_ON(p4d_page_vaddr(*p4d)
165                                        != p4d_page_vaddr(*p4d_ref));
166
167                         if (p4d_none(*p4d))
168                                 set_p4d(p4d, *p4d_ref);
169
170                         spin_unlock(pgt_lock);
171                 }
172                 spin_unlock(&pgd_lock);
173         }
174 }
175
176 /*
177  * When memory was added make sure all the processes MM have
178  * suitable PGD entries in the local PGD level page.
179  */
180 void sync_global_pgds(unsigned long start, unsigned long end)
181 {
182         if (pgtable_l5_enabled())
183                 sync_global_pgds_l5(start, end);
184         else
185                 sync_global_pgds_l4(start, end);
186 }
187
188 /*
189  * NOTE: This function is marked __ref because it calls __init function
190  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
191  */
192 static __ref void *spp_getpage(void)
193 {
194         void *ptr;
195
196         if (after_bootmem)
197                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
198         else
199                 ptr = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
200
201         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
202                 panic("set_pte_phys: cannot allocate page data %s\n",
203                         after_bootmem ? "after bootmem" : "");
204         }
205
206         pr_debug("spp_getpage %p\n", ptr);
207
208         return ptr;
209 }
210
211 static p4d_t *fill_p4d(pgd_t *pgd, unsigned long vaddr)
212 {
213         if (pgd_none(*pgd)) {
214                 p4d_t *p4d = (p4d_t *)spp_getpage();
215                 pgd_populate(&init_mm, pgd, p4d);
216                 if (p4d != p4d_offset(pgd, 0))
217                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
218                                p4d, p4d_offset(pgd, 0));
219         }
220         return p4d_offset(pgd, vaddr);
221 }
222
223 static pud_t *fill_pud(p4d_t *p4d, unsigned long vaddr)
224 {
225         if (p4d_none(*p4d)) {
226                 pud_t *pud = (pud_t *)spp_getpage();
227                 p4d_populate(&init_mm, p4d, pud);
228                 if (pud != pud_offset(p4d, 0))
229                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
230                                pud, pud_offset(p4d, 0));
231         }
232         return pud_offset(p4d, vaddr);
233 }
234
235 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
236 {
237         if (pud_none(*pud)) {
238                 pmd_t *pmd = (pmd_t *) spp_getpage();
239                 pud_populate(&init_mm, pud, pmd);
240                 if (pmd != pmd_offset(pud, 0))
241                         printk(KERN_ERR "PAGETABLE BUG #02! %p <-> %p\n",
242                                pmd, pmd_offset(pud, 0));
243         }
244         return pmd_offset(pud, vaddr);
245 }
246
247 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
248 {
249         if (pmd_none(*pmd)) {
250                 pte_t *pte = (pte_t *) spp_getpage();
251                 pmd_populate_kernel(&init_mm, pmd, pte);
252                 if (pte != pte_offset_kernel(pmd, 0))
253                         printk(KERN_ERR "PAGETABLE BUG #03!\n");
254         }
255         return pte_offset_kernel(pmd, vaddr);
256 }
257
258 static void __set_pte_vaddr(pud_t *pud, unsigned long vaddr, pte_t new_pte)
259 {
260         pmd_t *pmd = fill_pmd(pud, vaddr);
261         pte_t *pte = fill_pte(pmd, vaddr);
262
263         set_pte(pte, new_pte);
264
265         /*
266          * It's enough to flush this one mapping.
267          * (PGE mappings get flushed as well)
268          */
269         __flush_tlb_one_kernel(vaddr);
270 }
271
272 void set_pte_vaddr_p4d(p4d_t *p4d_page, unsigned long vaddr, pte_t new_pte)
273 {
274         p4d_t *p4d = p4d_page + p4d_index(vaddr);
275         pud_t *pud = fill_pud(p4d, vaddr);
276
277         __set_pte_vaddr(pud, vaddr, new_pte);
278 }
279
280 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
281 {
282         pud_t *pud = pud_page + pud_index(vaddr);
283
284         __set_pte_vaddr(pud, vaddr, new_pte);
285 }
286
287 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
288 {
289         pgd_t *pgd;
290         p4d_t *p4d_page;
291
292         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
293
294         pgd = pgd_offset_k(vaddr);
295         if (pgd_none(*pgd)) {
296                 printk(KERN_ERR
297                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
298                 return;
299         }
300
301         p4d_page = p4d_offset(pgd, 0);
302         set_pte_vaddr_p4d(p4d_page, vaddr, pteval);
303 }
304
305 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
306 {
307         pgd_t *pgd;
308         p4d_t *p4d;
309         pud_t *pud;
310
311         pgd = pgd_offset_k(vaddr);
312         p4d = fill_p4d(pgd, vaddr);
313         pud = fill_pud(p4d, vaddr);
314         return fill_pmd(pud, vaddr);
315 }
316
317 pte_t * __init populate_extra_pte(unsigned long vaddr)
318 {
319         pmd_t *pmd;
320
321         pmd = populate_extra_pmd(vaddr);
322         return fill_pte(pmd, vaddr);
323 }
324
325 /*
326  * Create large page table mappings for a range of physical addresses.
327  */
328 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
329                                         enum page_cache_mode cache)
330 {
331         pgd_t *pgd;
332         p4d_t *p4d;
333         pud_t *pud;
334         pmd_t *pmd;
335         pgprot_t prot;
336
337         pgprot_val(prot) = pgprot_val(PAGE_KERNEL_LARGE) |
338                 pgprot_val(pgprot_4k_2_large(cachemode2pgprot(cache)));
339         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
340         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
341                 pgd = pgd_offset_k((unsigned long)__va(phys));
342                 if (pgd_none(*pgd)) {
343                         p4d = (p4d_t *) spp_getpage();
344                         set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE |
345                                                 _PAGE_USER));
346                 }
347                 p4d = p4d_offset(pgd, (unsigned long)__va(phys));
348                 if (p4d_none(*p4d)) {
349                         pud = (pud_t *) spp_getpage();
350                         set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE |
351                                                 _PAGE_USER));
352                 }
353                 pud = pud_offset(p4d, (unsigned long)__va(phys));
354                 if (pud_none(*pud)) {
355                         pmd = (pmd_t *) spp_getpage();
356                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
357                                                 _PAGE_USER));
358                 }
359                 pmd = pmd_offset(pud, phys);
360                 BUG_ON(!pmd_none(*pmd));
361                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
362         }
363 }
364
365 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
366 {
367         __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_WB);
368 }
369
370 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
371 {
372         __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_UC);
373 }
374
375 /*
376  * The head.S code sets up the kernel high mapping:
377  *
378  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
379  *
380  * phys_base holds the negative offset to the kernel, which is added
381  * to the compile time generated pmds. This results in invalid pmds up
382  * to the point where we hit the physaddr 0 mapping.
383  *
384  * We limit the mappings to the region from _text to _brk_end.  _brk_end
385  * is rounded up to the 2MB boundary. This catches the invalid pmds as
386  * well, as they are located before _text:
387  */
388 void __init cleanup_highmap(void)
389 {
390         unsigned long vaddr = __START_KERNEL_map;
391         unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
392         unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
393         pmd_t *pmd = level2_kernel_pgt;
394
395         /*
396          * Native path, max_pfn_mapped is not set yet.
397          * Xen has valid max_pfn_mapped set in
398          *      arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
399          */
400         if (max_pfn_mapped)
401                 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
402
403         for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
404                 if (pmd_none(*pmd))
405                         continue;
406                 if (vaddr < (unsigned long) _text || vaddr > end)
407                         set_pmd(pmd, __pmd(0));
408         }
409 }
410
411 /*
412  * Create PTE level page table mapping for physical addresses.
413  * It returns the last physical address mapped.
414  */
415 static unsigned long __meminit
416 phys_pte_init(pte_t *pte_page, unsigned long paddr, unsigned long paddr_end,
417               pgprot_t prot)
418 {
419         unsigned long pages = 0, paddr_next;
420         unsigned long paddr_last = paddr_end;
421         pte_t *pte;
422         int i;
423
424         pte = pte_page + pte_index(paddr);
425         i = pte_index(paddr);
426
427         for (; i < PTRS_PER_PTE; i++, paddr = paddr_next, pte++) {
428                 paddr_next = (paddr & PAGE_MASK) + PAGE_SIZE;
429                 if (paddr >= paddr_end) {
430                         if (!after_bootmem &&
431                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
432                                              E820_TYPE_RAM) &&
433                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
434                                              E820_TYPE_RESERVED_KERN))
435                                 set_pte_safe(pte, __pte(0));
436                         continue;
437                 }
438
439                 /*
440                  * We will re-use the existing mapping.
441                  * Xen for example has some special requirements, like mapping
442                  * pagetable pages as RO. So assume someone who pre-setup
443                  * these mappings are more intelligent.
444                  */
445                 if (!pte_none(*pte)) {
446                         if (!after_bootmem)
447                                 pages++;
448                         continue;
449                 }
450
451                 if (0)
452                         pr_info("   pte=%p addr=%lx pte=%016lx\n", pte, paddr,
453                                 pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL).pte);
454                 pages++;
455                 set_pte_safe(pte, pfn_pte(paddr >> PAGE_SHIFT, prot));
456                 paddr_last = (paddr & PAGE_MASK) + PAGE_SIZE;
457         }
458
459         update_page_count(PG_LEVEL_4K, pages);
460
461         return paddr_last;
462 }
463
464 /*
465  * Create PMD level page table mapping for physical addresses. The virtual
466  * and physical address have to be aligned at this level.
467  * It returns the last physical address mapped.
468  */
469 static unsigned long __meminit
470 phys_pmd_init(pmd_t *pmd_page, unsigned long paddr, unsigned long paddr_end,
471               unsigned long page_size_mask, pgprot_t prot)
472 {
473         unsigned long pages = 0, paddr_next;
474         unsigned long paddr_last = paddr_end;
475
476         int i = pmd_index(paddr);
477
478         for (; i < PTRS_PER_PMD; i++, paddr = paddr_next) {
479                 pmd_t *pmd = pmd_page + pmd_index(paddr);
480                 pte_t *pte;
481                 pgprot_t new_prot = prot;
482
483                 paddr_next = (paddr & PMD_MASK) + PMD_SIZE;
484                 if (paddr >= paddr_end) {
485                         if (!after_bootmem &&
486                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
487                                              E820_TYPE_RAM) &&
488                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
489                                              E820_TYPE_RESERVED_KERN))
490                                 set_pmd_safe(pmd, __pmd(0));
491                         continue;
492                 }
493
494                 if (!pmd_none(*pmd)) {
495                         if (!pmd_large(*pmd)) {
496                                 spin_lock(&init_mm.page_table_lock);
497                                 pte = (pte_t *)pmd_page_vaddr(*pmd);
498                                 paddr_last = phys_pte_init(pte, paddr,
499                                                            paddr_end, prot);
500                                 spin_unlock(&init_mm.page_table_lock);
501                                 continue;
502                         }
503                         /*
504                          * If we are ok with PG_LEVEL_2M mapping, then we will
505                          * use the existing mapping,
506                          *
507                          * Otherwise, we will split the large page mapping but
508                          * use the same existing protection bits except for
509                          * large page, so that we don't violate Intel's TLB
510                          * Application note (317080) which says, while changing
511                          * the page sizes, new and old translations should
512                          * not differ with respect to page frame and
513                          * attributes.
514                          */
515                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
516                                 if (!after_bootmem)
517                                         pages++;
518                                 paddr_last = paddr_next;
519                                 continue;
520                         }
521                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
522                 }
523
524                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
525                         pages++;
526                         spin_lock(&init_mm.page_table_lock);
527                         set_pte_safe((pte_t *)pmd,
528                                 pfn_pte((paddr & PMD_MASK) >> PAGE_SHIFT,
529                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
530                         spin_unlock(&init_mm.page_table_lock);
531                         paddr_last = paddr_next;
532                         continue;
533                 }
534
535                 pte = alloc_low_page();
536                 paddr_last = phys_pte_init(pte, paddr, paddr_end, new_prot);
537
538                 spin_lock(&init_mm.page_table_lock);
539                 pmd_populate_kernel_safe(&init_mm, pmd, pte);
540                 spin_unlock(&init_mm.page_table_lock);
541         }
542         update_page_count(PG_LEVEL_2M, pages);
543         return paddr_last;
544 }
545
546 /*
547  * Create PUD level page table mapping for physical addresses. The virtual
548  * and physical address do not have to be aligned at this level. KASLR can
549  * randomize virtual addresses up to this level.
550  * It returns the last physical address mapped.
551  */
552 static unsigned long __meminit
553 phys_pud_init(pud_t *pud_page, unsigned long paddr, unsigned long paddr_end,
554               unsigned long page_size_mask)
555 {
556         unsigned long pages = 0, paddr_next;
557         unsigned long paddr_last = paddr_end;
558         unsigned long vaddr = (unsigned long)__va(paddr);
559         int i = pud_index(vaddr);
560
561         for (; i < PTRS_PER_PUD; i++, paddr = paddr_next) {
562                 pud_t *pud;
563                 pmd_t *pmd;
564                 pgprot_t prot = PAGE_KERNEL;
565
566                 vaddr = (unsigned long)__va(paddr);
567                 pud = pud_page + pud_index(vaddr);
568                 paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
569
570                 if (paddr >= paddr_end) {
571                         if (!after_bootmem &&
572                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
573                                              E820_TYPE_RAM) &&
574                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
575                                              E820_TYPE_RESERVED_KERN))
576                                 set_pud_safe(pud, __pud(0));
577                         continue;
578                 }
579
580                 if (!pud_none(*pud)) {
581                         if (!pud_large(*pud)) {
582                                 pmd = pmd_offset(pud, 0);
583                                 paddr_last = phys_pmd_init(pmd, paddr,
584                                                            paddr_end,
585                                                            page_size_mask,
586                                                            prot);
587                                 continue;
588                         }
589                         /*
590                          * If we are ok with PG_LEVEL_1G mapping, then we will
591                          * use the existing mapping.
592                          *
593                          * Otherwise, we will split the gbpage mapping but use
594                          * the same existing protection  bits except for large
595                          * page, so that we don't violate Intel's TLB
596                          * Application note (317080) which says, while changing
597                          * the page sizes, new and old translations should
598                          * not differ with respect to page frame and
599                          * attributes.
600                          */
601                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
602                                 if (!after_bootmem)
603                                         pages++;
604                                 paddr_last = paddr_next;
605                                 continue;
606                         }
607                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
608                 }
609
610                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
611                         pages++;
612                         spin_lock(&init_mm.page_table_lock);
613                         set_pte_safe((pte_t *)pud,
614                                 pfn_pte((paddr & PUD_MASK) >> PAGE_SHIFT,
615                                         PAGE_KERNEL_LARGE));
616                         spin_unlock(&init_mm.page_table_lock);
617                         paddr_last = paddr_next;
618                         continue;
619                 }
620
621                 pmd = alloc_low_page();
622                 paddr_last = phys_pmd_init(pmd, paddr, paddr_end,
623                                            page_size_mask, prot);
624
625                 spin_lock(&init_mm.page_table_lock);
626                 pud_populate_safe(&init_mm, pud, pmd);
627                 spin_unlock(&init_mm.page_table_lock);
628         }
629
630         update_page_count(PG_LEVEL_1G, pages);
631
632         return paddr_last;
633 }
634
635 static unsigned long __meminit
636 phys_p4d_init(p4d_t *p4d_page, unsigned long paddr, unsigned long paddr_end,
637               unsigned long page_size_mask)
638 {
639         unsigned long paddr_next, paddr_last = paddr_end;
640         unsigned long vaddr = (unsigned long)__va(paddr);
641         int i = p4d_index(vaddr);
642
643         if (!pgtable_l5_enabled())
644                 return phys_pud_init((pud_t *) p4d_page, paddr, paddr_end, page_size_mask);
645
646         for (; i < PTRS_PER_P4D; i++, paddr = paddr_next) {
647                 p4d_t *p4d;
648                 pud_t *pud;
649
650                 vaddr = (unsigned long)__va(paddr);
651                 p4d = p4d_page + p4d_index(vaddr);
652                 paddr_next = (paddr & P4D_MASK) + P4D_SIZE;
653
654                 if (paddr >= paddr_end) {
655                         if (!after_bootmem &&
656                             !e820__mapped_any(paddr & P4D_MASK, paddr_next,
657                                              E820_TYPE_RAM) &&
658                             !e820__mapped_any(paddr & P4D_MASK, paddr_next,
659                                              E820_TYPE_RESERVED_KERN))
660                                 set_p4d_safe(p4d, __p4d(0));
661                         continue;
662                 }
663
664                 if (!p4d_none(*p4d)) {
665                         pud = pud_offset(p4d, 0);
666                         paddr_last = phys_pud_init(pud, paddr,
667                                         paddr_end,
668                                         page_size_mask);
669                         continue;
670                 }
671
672                 pud = alloc_low_page();
673                 paddr_last = phys_pud_init(pud, paddr, paddr_end,
674                                            page_size_mask);
675
676                 spin_lock(&init_mm.page_table_lock);
677                 p4d_populate_safe(&init_mm, p4d, pud);
678                 spin_unlock(&init_mm.page_table_lock);
679         }
680
681         return paddr_last;
682 }
683
684 /*
685  * Create page table mapping for the physical memory for specific physical
686  * addresses. The virtual and physical addresses have to be aligned on PMD level
687  * down. It returns the last physical address mapped.
688  */
689 unsigned long __meminit
690 kernel_physical_mapping_init(unsigned long paddr_start,
691                              unsigned long paddr_end,
692                              unsigned long page_size_mask)
693 {
694         bool pgd_changed = false;
695         unsigned long vaddr, vaddr_start, vaddr_end, vaddr_next, paddr_last;
696
697         paddr_last = paddr_end;
698         vaddr = (unsigned long)__va(paddr_start);
699         vaddr_end = (unsigned long)__va(paddr_end);
700         vaddr_start = vaddr;
701
702         for (; vaddr < vaddr_end; vaddr = vaddr_next) {
703                 pgd_t *pgd = pgd_offset_k(vaddr);
704                 p4d_t *p4d;
705
706                 vaddr_next = (vaddr & PGDIR_MASK) + PGDIR_SIZE;
707
708                 if (pgd_val(*pgd)) {
709                         p4d = (p4d_t *)pgd_page_vaddr(*pgd);
710                         paddr_last = phys_p4d_init(p4d, __pa(vaddr),
711                                                    __pa(vaddr_end),
712                                                    page_size_mask);
713                         continue;
714                 }
715
716                 p4d = alloc_low_page();
717                 paddr_last = phys_p4d_init(p4d, __pa(vaddr), __pa(vaddr_end),
718                                            page_size_mask);
719
720                 spin_lock(&init_mm.page_table_lock);
721                 if (pgtable_l5_enabled())
722                         pgd_populate_safe(&init_mm, pgd, p4d);
723                 else
724                         p4d_populate_safe(&init_mm, p4d_offset(pgd, vaddr), (pud_t *) p4d);
725                 spin_unlock(&init_mm.page_table_lock);
726                 pgd_changed = true;
727         }
728
729         if (pgd_changed)
730                 sync_global_pgds(vaddr_start, vaddr_end - 1);
731
732         return paddr_last;
733 }
734
735 #ifndef CONFIG_NUMA
736 void __init initmem_init(void)
737 {
738         memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
739 }
740 #endif
741
742 void __init paging_init(void)
743 {
744         sparse_memory_present_with_active_regions(MAX_NUMNODES);
745         sparse_init();
746
747         /*
748          * clear the default setting with node 0
749          * note: don't use nodes_clear here, that is really clearing when
750          *       numa support is not compiled in, and later node_set_state
751          *       will not set it back.
752          */
753         node_clear_state(0, N_MEMORY);
754         if (N_MEMORY != N_NORMAL_MEMORY)
755                 node_clear_state(0, N_NORMAL_MEMORY);
756
757         zone_sizes_init();
758 }
759
760 /*
761  * Memory hotplug specific functions
762  */
763 #ifdef CONFIG_MEMORY_HOTPLUG
764 /*
765  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
766  * updating.
767  */
768 static void update_end_of_memory_vars(u64 start, u64 size)
769 {
770         unsigned long end_pfn = PFN_UP(start + size);
771
772         if (end_pfn > max_pfn) {
773                 max_pfn = end_pfn;
774                 max_low_pfn = end_pfn;
775                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
776         }
777 }
778
779 int add_pages(int nid, unsigned long start_pfn, unsigned long nr_pages,
780                 struct vmem_altmap *altmap, bool want_memblock)
781 {
782         int ret;
783
784         ret = __add_pages(nid, start_pfn, nr_pages, altmap, want_memblock);
785         WARN_ON_ONCE(ret);
786
787         /* update max_pfn, max_low_pfn and high_memory */
788         update_end_of_memory_vars(start_pfn << PAGE_SHIFT,
789                                   nr_pages << PAGE_SHIFT);
790
791         return ret;
792 }
793
794 int arch_add_memory(int nid, u64 start, u64 size, struct vmem_altmap *altmap,
795                 bool want_memblock)
796 {
797         unsigned long start_pfn = start >> PAGE_SHIFT;
798         unsigned long nr_pages = size >> PAGE_SHIFT;
799
800         init_memory_mapping(start, start + size);
801
802         return add_pages(nid, start_pfn, nr_pages, altmap, want_memblock);
803 }
804
805 #define PAGE_INUSE 0xFD
806
807 static void __meminit free_pagetable(struct page *page, int order)
808 {
809         unsigned long magic;
810         unsigned int nr_pages = 1 << order;
811
812         /* bootmem page has reserved flag */
813         if (PageReserved(page)) {
814                 __ClearPageReserved(page);
815
816                 magic = (unsigned long)page->freelist;
817                 if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
818                         while (nr_pages--)
819                                 put_page_bootmem(page++);
820                 } else
821                         while (nr_pages--)
822                                 free_reserved_page(page++);
823         } else
824                 free_pages((unsigned long)page_address(page), order);
825 }
826
827 static void __meminit free_hugepage_table(struct page *page,
828                 struct vmem_altmap *altmap)
829 {
830         if (altmap)
831                 vmem_altmap_free(altmap, PMD_SIZE / PAGE_SIZE);
832         else
833                 free_pagetable(page, get_order(PMD_SIZE));
834 }
835
836 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
837 {
838         pte_t *pte;
839         int i;
840
841         for (i = 0; i < PTRS_PER_PTE; i++) {
842                 pte = pte_start + i;
843                 if (!pte_none(*pte))
844                         return;
845         }
846
847         /* free a pte talbe */
848         free_pagetable(pmd_page(*pmd), 0);
849         spin_lock(&init_mm.page_table_lock);
850         pmd_clear(pmd);
851         spin_unlock(&init_mm.page_table_lock);
852 }
853
854 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
855 {
856         pmd_t *pmd;
857         int i;
858
859         for (i = 0; i < PTRS_PER_PMD; i++) {
860                 pmd = pmd_start + i;
861                 if (!pmd_none(*pmd))
862                         return;
863         }
864
865         /* free a pmd talbe */
866         free_pagetable(pud_page(*pud), 0);
867         spin_lock(&init_mm.page_table_lock);
868         pud_clear(pud);
869         spin_unlock(&init_mm.page_table_lock);
870 }
871
872 static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
873 {
874         pud_t *pud;
875         int i;
876
877         for (i = 0; i < PTRS_PER_PUD; i++) {
878                 pud = pud_start + i;
879                 if (!pud_none(*pud))
880                         return;
881         }
882
883         /* free a pud talbe */
884         free_pagetable(p4d_page(*p4d), 0);
885         spin_lock(&init_mm.page_table_lock);
886         p4d_clear(p4d);
887         spin_unlock(&init_mm.page_table_lock);
888 }
889
890 static void __meminit
891 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
892                  bool direct)
893 {
894         unsigned long next, pages = 0;
895         pte_t *pte;
896         void *page_addr;
897         phys_addr_t phys_addr;
898
899         pte = pte_start + pte_index(addr);
900         for (; addr < end; addr = next, pte++) {
901                 next = (addr + PAGE_SIZE) & PAGE_MASK;
902                 if (next > end)
903                         next = end;
904
905                 if (!pte_present(*pte))
906                         continue;
907
908                 /*
909                  * We mapped [0,1G) memory as identity mapping when
910                  * initializing, in arch/x86/kernel/head_64.S. These
911                  * pagetables cannot be removed.
912                  */
913                 phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
914                 if (phys_addr < (phys_addr_t)0x40000000)
915                         return;
916
917                 if (PAGE_ALIGNED(addr) && PAGE_ALIGNED(next)) {
918                         /*
919                          * Do not free direct mapping pages since they were
920                          * freed when offlining, or simplely not in use.
921                          */
922                         if (!direct)
923                                 free_pagetable(pte_page(*pte), 0);
924
925                         spin_lock(&init_mm.page_table_lock);
926                         pte_clear(&init_mm, addr, pte);
927                         spin_unlock(&init_mm.page_table_lock);
928
929                         /* For non-direct mapping, pages means nothing. */
930                         pages++;
931                 } else {
932                         /*
933                          * If we are here, we are freeing vmemmap pages since
934                          * direct mapped memory ranges to be freed are aligned.
935                          *
936                          * If we are not removing the whole page, it means
937                          * other page structs in this page are being used and
938                          * we canot remove them. So fill the unused page_structs
939                          * with 0xFD, and remove the page when it is wholly
940                          * filled with 0xFD.
941                          */
942                         memset((void *)addr, PAGE_INUSE, next - addr);
943
944                         page_addr = page_address(pte_page(*pte));
945                         if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) {
946                                 free_pagetable(pte_page(*pte), 0);
947
948                                 spin_lock(&init_mm.page_table_lock);
949                                 pte_clear(&init_mm, addr, pte);
950                                 spin_unlock(&init_mm.page_table_lock);
951                         }
952                 }
953         }
954
955         /* Call free_pte_table() in remove_pmd_table(). */
956         flush_tlb_all();
957         if (direct)
958                 update_page_count(PG_LEVEL_4K, -pages);
959 }
960
961 static void __meminit
962 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
963                  bool direct, struct vmem_altmap *altmap)
964 {
965         unsigned long next, pages = 0;
966         pte_t *pte_base;
967         pmd_t *pmd;
968         void *page_addr;
969
970         pmd = pmd_start + pmd_index(addr);
971         for (; addr < end; addr = next, pmd++) {
972                 next = pmd_addr_end(addr, end);
973
974                 if (!pmd_present(*pmd))
975                         continue;
976
977                 if (pmd_large(*pmd)) {
978                         if (IS_ALIGNED(addr, PMD_SIZE) &&
979                             IS_ALIGNED(next, PMD_SIZE)) {
980                                 if (!direct)
981                                         free_hugepage_table(pmd_page(*pmd),
982                                                             altmap);
983
984                                 spin_lock(&init_mm.page_table_lock);
985                                 pmd_clear(pmd);
986                                 spin_unlock(&init_mm.page_table_lock);
987                                 pages++;
988                         } else {
989                                 /* If here, we are freeing vmemmap pages. */
990                                 memset((void *)addr, PAGE_INUSE, next - addr);
991
992                                 page_addr = page_address(pmd_page(*pmd));
993                                 if (!memchr_inv(page_addr, PAGE_INUSE,
994                                                 PMD_SIZE)) {
995                                         free_hugepage_table(pmd_page(*pmd),
996                                                             altmap);
997
998                                         spin_lock(&init_mm.page_table_lock);
999                                         pmd_clear(pmd);
1000                                         spin_unlock(&init_mm.page_table_lock);
1001                                 }
1002                         }
1003
1004                         continue;
1005                 }
1006
1007                 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
1008                 remove_pte_table(pte_base, addr, next, direct);
1009                 free_pte_table(pte_base, pmd);
1010         }
1011
1012         /* Call free_pmd_table() in remove_pud_table(). */
1013         if (direct)
1014                 update_page_count(PG_LEVEL_2M, -pages);
1015 }
1016
1017 static void __meminit
1018 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
1019                  struct vmem_altmap *altmap, bool direct)
1020 {
1021         unsigned long next, pages = 0;
1022         pmd_t *pmd_base;
1023         pud_t *pud;
1024         void *page_addr;
1025
1026         pud = pud_start + pud_index(addr);
1027         for (; addr < end; addr = next, pud++) {
1028                 next = pud_addr_end(addr, end);
1029
1030                 if (!pud_present(*pud))
1031                         continue;
1032
1033                 if (pud_large(*pud)) {
1034                         if (IS_ALIGNED(addr, PUD_SIZE) &&
1035                             IS_ALIGNED(next, PUD_SIZE)) {
1036                                 if (!direct)
1037                                         free_pagetable(pud_page(*pud),
1038                                                        get_order(PUD_SIZE));
1039
1040                                 spin_lock(&init_mm.page_table_lock);
1041                                 pud_clear(pud);
1042                                 spin_unlock(&init_mm.page_table_lock);
1043                                 pages++;
1044                         } else {
1045                                 /* If here, we are freeing vmemmap pages. */
1046                                 memset((void *)addr, PAGE_INUSE, next - addr);
1047
1048                                 page_addr = page_address(pud_page(*pud));
1049                                 if (!memchr_inv(page_addr, PAGE_INUSE,
1050                                                 PUD_SIZE)) {
1051                                         free_pagetable(pud_page(*pud),
1052                                                        get_order(PUD_SIZE));
1053
1054                                         spin_lock(&init_mm.page_table_lock);
1055                                         pud_clear(pud);
1056                                         spin_unlock(&init_mm.page_table_lock);
1057                                 }
1058                         }
1059
1060                         continue;
1061                 }
1062
1063                 pmd_base = pmd_offset(pud, 0);
1064                 remove_pmd_table(pmd_base, addr, next, direct, altmap);
1065                 free_pmd_table(pmd_base, pud);
1066         }
1067
1068         if (direct)
1069                 update_page_count(PG_LEVEL_1G, -pages);
1070 }
1071
1072 static void __meminit
1073 remove_p4d_table(p4d_t *p4d_start, unsigned long addr, unsigned long end,
1074                  struct vmem_altmap *altmap, bool direct)
1075 {
1076         unsigned long next, pages = 0;
1077         pud_t *pud_base;
1078         p4d_t *p4d;
1079
1080         p4d = p4d_start + p4d_index(addr);
1081         for (; addr < end; addr = next, p4d++) {
1082                 next = p4d_addr_end(addr, end);
1083
1084                 if (!p4d_present(*p4d))
1085                         continue;
1086
1087                 BUILD_BUG_ON(p4d_large(*p4d));
1088
1089                 pud_base = pud_offset(p4d, 0);
1090                 remove_pud_table(pud_base, addr, next, altmap, direct);
1091                 /*
1092                  * For 4-level page tables we do not want to free PUDs, but in the
1093                  * 5-level case we should free them. This code will have to change
1094                  * to adapt for boot-time switching between 4 and 5 level page tables.
1095                  */
1096                 if (pgtable_l5_enabled())
1097                         free_pud_table(pud_base, p4d);
1098         }
1099
1100         if (direct)
1101                 update_page_count(PG_LEVEL_512G, -pages);
1102 }
1103
1104 /* start and end are both virtual address. */
1105 static void __meminit
1106 remove_pagetable(unsigned long start, unsigned long end, bool direct,
1107                 struct vmem_altmap *altmap)
1108 {
1109         unsigned long next;
1110         unsigned long addr;
1111         pgd_t *pgd;
1112         p4d_t *p4d;
1113
1114         for (addr = start; addr < end; addr = next) {
1115                 next = pgd_addr_end(addr, end);
1116
1117                 pgd = pgd_offset_k(addr);
1118                 if (!pgd_present(*pgd))
1119                         continue;
1120
1121                 p4d = p4d_offset(pgd, 0);
1122                 remove_p4d_table(p4d, addr, next, altmap, direct);
1123         }
1124
1125         flush_tlb_all();
1126 }
1127
1128 void __ref vmemmap_free(unsigned long start, unsigned long end,
1129                 struct vmem_altmap *altmap)
1130 {
1131         remove_pagetable(start, end, false, altmap);
1132 }
1133
1134 #ifdef CONFIG_MEMORY_HOTREMOVE
1135 static void __meminit
1136 kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1137 {
1138         start = (unsigned long)__va(start);
1139         end = (unsigned long)__va(end);
1140
1141         remove_pagetable(start, end, true, NULL);
1142 }
1143
1144 int __ref arch_remove_memory(int nid, u64 start, u64 size,
1145                                 struct vmem_altmap *altmap)
1146 {
1147         unsigned long start_pfn = start >> PAGE_SHIFT;
1148         unsigned long nr_pages = size >> PAGE_SHIFT;
1149         struct page *page = pfn_to_page(start_pfn);
1150         struct zone *zone;
1151         int ret;
1152
1153         /* With altmap the first mapped page is offset from @start */
1154         if (altmap)
1155                 page += vmem_altmap_offset(altmap);
1156         zone = page_zone(page);
1157         ret = __remove_pages(zone, start_pfn, nr_pages, altmap);
1158         WARN_ON_ONCE(ret);
1159         kernel_physical_mapping_remove(start, start + size);
1160
1161         return ret;
1162 }
1163 #endif
1164 #endif /* CONFIG_MEMORY_HOTPLUG */
1165
1166 static struct kcore_list kcore_vsyscall;
1167
1168 static void __init register_page_bootmem_info(void)
1169 {
1170 #ifdef CONFIG_NUMA
1171         int i;
1172
1173         for_each_online_node(i)
1174                 register_page_bootmem_info_node(NODE_DATA(i));
1175 #endif
1176 }
1177
1178 void __init mem_init(void)
1179 {
1180         pci_iommu_alloc();
1181
1182         /* clear_bss() already clear the empty_zero_page */
1183
1184         /* this will put all memory onto the freelists */
1185         memblock_free_all();
1186         after_bootmem = 1;
1187         x86_init.hyper.init_after_bootmem();
1188
1189         /*
1190          * Must be done after boot memory is put on freelist, because here we
1191          * might set fields in deferred struct pages that have not yet been
1192          * initialized, and memblock_free_all() initializes all the reserved
1193          * deferred pages for us.
1194          */
1195         register_page_bootmem_info();
1196
1197         /* Register memory areas for /proc/kcore */
1198         if (get_gate_vma(&init_mm))
1199                 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR, PAGE_SIZE, KCORE_USER);
1200
1201         mem_init_print_info(NULL);
1202 }
1203
1204 int kernel_set_to_readonly;
1205
1206 void set_kernel_text_rw(void)
1207 {
1208         unsigned long start = PFN_ALIGN(_text);
1209         unsigned long end = PFN_ALIGN(__stop___ex_table);
1210
1211         if (!kernel_set_to_readonly)
1212                 return;
1213
1214         pr_debug("Set kernel text: %lx - %lx for read write\n",
1215                  start, end);
1216
1217         /*
1218          * Make the kernel identity mapping for text RW. Kernel text
1219          * mapping will always be RO. Refer to the comment in
1220          * static_protections() in pageattr.c
1221          */
1222         set_memory_rw(start, (end - start) >> PAGE_SHIFT);
1223 }
1224
1225 void set_kernel_text_ro(void)
1226 {
1227         unsigned long start = PFN_ALIGN(_text);
1228         unsigned long end = PFN_ALIGN(__stop___ex_table);
1229
1230         if (!kernel_set_to_readonly)
1231                 return;
1232
1233         pr_debug("Set kernel text: %lx - %lx for read only\n",
1234                  start, end);
1235
1236         /*
1237          * Set the kernel identity mapping for text RO.
1238          */
1239         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1240 }
1241
1242 void mark_rodata_ro(void)
1243 {
1244         unsigned long start = PFN_ALIGN(_text);
1245         unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1246         unsigned long end = (unsigned long) &__end_rodata_hpage_align;
1247         unsigned long text_end = PFN_ALIGN(&__stop___ex_table);
1248         unsigned long rodata_end = PFN_ALIGN(&__end_rodata);
1249         unsigned long all_end;
1250
1251         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1252                (end - start) >> 10);
1253         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1254
1255         kernel_set_to_readonly = 1;
1256
1257         /*
1258          * The rodata/data/bss/brk section (but not the kernel text!)
1259          * should also be not-executable.
1260          *
1261          * We align all_end to PMD_SIZE because the existing mapping
1262          * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1263          * split the PMD and the reminder between _brk_end and the end
1264          * of the PMD will remain mapped executable.
1265          *
1266          * Any PMD which was setup after the one which covers _brk_end
1267          * has been zapped already via cleanup_highmem().
1268          */
1269         all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1270         set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
1271
1272 #ifdef CONFIG_CPA_DEBUG
1273         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1274         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1275
1276         printk(KERN_INFO "Testing CPA: again\n");
1277         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1278 #endif
1279
1280         free_kernel_image_pages((void *)text_end, (void *)rodata_start);
1281         free_kernel_image_pages((void *)rodata_end, (void *)_sdata);
1282
1283         debug_checkwx();
1284 }
1285
1286 int kern_addr_valid(unsigned long addr)
1287 {
1288         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1289         pgd_t *pgd;
1290         p4d_t *p4d;
1291         pud_t *pud;
1292         pmd_t *pmd;
1293         pte_t *pte;
1294
1295         if (above != 0 && above != -1UL)
1296                 return 0;
1297
1298         pgd = pgd_offset_k(addr);
1299         if (pgd_none(*pgd))
1300                 return 0;
1301
1302         p4d = p4d_offset(pgd, addr);
1303         if (p4d_none(*p4d))
1304                 return 0;
1305
1306         pud = pud_offset(p4d, addr);
1307         if (pud_none(*pud))
1308                 return 0;
1309
1310         if (pud_large(*pud))
1311                 return pfn_valid(pud_pfn(*pud));
1312
1313         pmd = pmd_offset(pud, addr);
1314         if (pmd_none(*pmd))
1315                 return 0;
1316
1317         if (pmd_large(*pmd))
1318                 return pfn_valid(pmd_pfn(*pmd));
1319
1320         pte = pte_offset_kernel(pmd, addr);
1321         if (pte_none(*pte))
1322                 return 0;
1323
1324         return pfn_valid(pte_pfn(*pte));
1325 }
1326
1327 /*
1328  * Block size is the minimum amount of memory which can be hotplugged or
1329  * hotremoved. It must be power of two and must be equal or larger than
1330  * MIN_MEMORY_BLOCK_SIZE.
1331  */
1332 #define MAX_BLOCK_SIZE (2UL << 30)
1333
1334 /* Amount of ram needed to start using large blocks */
1335 #define MEM_SIZE_FOR_LARGE_BLOCK (64UL << 30)
1336
1337 /* Adjustable memory block size */
1338 static unsigned long set_memory_block_size;
1339 int __init set_memory_block_size_order(unsigned int order)
1340 {
1341         unsigned long size = 1UL << order;
1342
1343         if (size > MEM_SIZE_FOR_LARGE_BLOCK || size < MIN_MEMORY_BLOCK_SIZE)
1344                 return -EINVAL;
1345
1346         set_memory_block_size = size;
1347         return 0;
1348 }
1349
1350 static unsigned long probe_memory_block_size(void)
1351 {
1352         unsigned long boot_mem_end = max_pfn << PAGE_SHIFT;
1353         unsigned long bz;
1354
1355         /* If memory block size has been set, then use it */
1356         bz = set_memory_block_size;
1357         if (bz)
1358                 goto done;
1359
1360         /* Use regular block if RAM is smaller than MEM_SIZE_FOR_LARGE_BLOCK */
1361         if (boot_mem_end < MEM_SIZE_FOR_LARGE_BLOCK) {
1362                 bz = MIN_MEMORY_BLOCK_SIZE;
1363                 goto done;
1364         }
1365
1366         /* Find the largest allowed block size that aligns to memory end */
1367         for (bz = MAX_BLOCK_SIZE; bz > MIN_MEMORY_BLOCK_SIZE; bz >>= 1) {
1368                 if (IS_ALIGNED(boot_mem_end, bz))
1369                         break;
1370         }
1371 done:
1372         pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20);
1373
1374         return bz;
1375 }
1376
1377 static unsigned long memory_block_size_probed;
1378 unsigned long memory_block_size_bytes(void)
1379 {
1380         if (!memory_block_size_probed)
1381                 memory_block_size_probed = probe_memory_block_size();
1382
1383         return memory_block_size_probed;
1384 }
1385
1386 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1387 /*
1388  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1389  */
1390 static long __meminitdata addr_start, addr_end;
1391 static void __meminitdata *p_start, *p_end;
1392 static int __meminitdata node_start;
1393
1394 static int __meminit vmemmap_populate_hugepages(unsigned long start,
1395                 unsigned long end, int node, struct vmem_altmap *altmap)
1396 {
1397         unsigned long addr;
1398         unsigned long next;
1399         pgd_t *pgd;
1400         p4d_t *p4d;
1401         pud_t *pud;
1402         pmd_t *pmd;
1403
1404         for (addr = start; addr < end; addr = next) {
1405                 next = pmd_addr_end(addr, end);
1406
1407                 pgd = vmemmap_pgd_populate(addr, node);
1408                 if (!pgd)
1409                         return -ENOMEM;
1410
1411                 p4d = vmemmap_p4d_populate(pgd, addr, node);
1412                 if (!p4d)
1413                         return -ENOMEM;
1414
1415                 pud = vmemmap_pud_populate(p4d, addr, node);
1416                 if (!pud)
1417                         return -ENOMEM;
1418
1419                 pmd = pmd_offset(pud, addr);
1420                 if (pmd_none(*pmd)) {
1421                         void *p;
1422
1423                         if (altmap)
1424                                 p = altmap_alloc_block_buf(PMD_SIZE, altmap);
1425                         else
1426                                 p = vmemmap_alloc_block_buf(PMD_SIZE, node);
1427                         if (p) {
1428                                 pte_t entry;
1429
1430                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1431                                                 PAGE_KERNEL_LARGE);
1432                                 set_pmd(pmd, __pmd(pte_val(entry)));
1433
1434                                 /* check to see if we have contiguous blocks */
1435                                 if (p_end != p || node_start != node) {
1436                                         if (p_start)
1437                                                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1438                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1439                                         addr_start = addr;
1440                                         node_start = node;
1441                                         p_start = p;
1442                                 }
1443
1444                                 addr_end = addr + PMD_SIZE;
1445                                 p_end = p + PMD_SIZE;
1446                                 continue;
1447                         } else if (altmap)
1448                                 return -ENOMEM; /* no fallback */
1449                 } else if (pmd_large(*pmd)) {
1450                         vmemmap_verify((pte_t *)pmd, node, addr, next);
1451                         continue;
1452                 }
1453                 if (vmemmap_populate_basepages(addr, next, node))
1454                         return -ENOMEM;
1455         }
1456         return 0;
1457 }
1458
1459 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1460                 struct vmem_altmap *altmap)
1461 {
1462         int err;
1463
1464         if (boot_cpu_has(X86_FEATURE_PSE))
1465                 err = vmemmap_populate_hugepages(start, end, node, altmap);
1466         else if (altmap) {
1467                 pr_err_once("%s: no cpu support for altmap allocations\n",
1468                                 __func__);
1469                 err = -ENOMEM;
1470         } else
1471                 err = vmemmap_populate_basepages(start, end, node);
1472         if (!err)
1473                 sync_global_pgds(start, end - 1);
1474         return err;
1475 }
1476
1477 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1478 void register_page_bootmem_memmap(unsigned long section_nr,
1479                                   struct page *start_page, unsigned long nr_pages)
1480 {
1481         unsigned long addr = (unsigned long)start_page;
1482         unsigned long end = (unsigned long)(start_page + nr_pages);
1483         unsigned long next;
1484         pgd_t *pgd;
1485         p4d_t *p4d;
1486         pud_t *pud;
1487         pmd_t *pmd;
1488         unsigned int nr_pmd_pages;
1489         struct page *page;
1490
1491         for (; addr < end; addr = next) {
1492                 pte_t *pte = NULL;
1493
1494                 pgd = pgd_offset_k(addr);
1495                 if (pgd_none(*pgd)) {
1496                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1497                         continue;
1498                 }
1499                 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1500
1501                 p4d = p4d_offset(pgd, addr);
1502                 if (p4d_none(*p4d)) {
1503                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1504                         continue;
1505                 }
1506                 get_page_bootmem(section_nr, p4d_page(*p4d), MIX_SECTION_INFO);
1507
1508                 pud = pud_offset(p4d, addr);
1509                 if (pud_none(*pud)) {
1510                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1511                         continue;
1512                 }
1513                 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1514
1515                 if (!boot_cpu_has(X86_FEATURE_PSE)) {
1516                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1517                         pmd = pmd_offset(pud, addr);
1518                         if (pmd_none(*pmd))
1519                                 continue;
1520                         get_page_bootmem(section_nr, pmd_page(*pmd),
1521                                          MIX_SECTION_INFO);
1522
1523                         pte = pte_offset_kernel(pmd, addr);
1524                         if (pte_none(*pte))
1525                                 continue;
1526                         get_page_bootmem(section_nr, pte_page(*pte),
1527                                          SECTION_INFO);
1528                 } else {
1529                         next = pmd_addr_end(addr, end);
1530
1531                         pmd = pmd_offset(pud, addr);
1532                         if (pmd_none(*pmd))
1533                                 continue;
1534
1535                         nr_pmd_pages = 1 << get_order(PMD_SIZE);
1536                         page = pmd_page(*pmd);
1537                         while (nr_pmd_pages--)
1538                                 get_page_bootmem(section_nr, page++,
1539                                                  SECTION_INFO);
1540                 }
1541         }
1542 }
1543 #endif
1544
1545 void __meminit vmemmap_populate_print_last(void)
1546 {
1547         if (p_start) {
1548                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1549                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1550                 p_start = NULL;
1551                 p_end = NULL;
1552                 node_start = 0;
1553         }
1554 }
1555 #endif