x86, pat: avoid highmem cache attribute aliasing
[linux-2.6.git] / arch / x86 / mm / pageattr.c
1 /*
2  * Copyright 2002 Andi Kleen, SuSE Labs.
3  * Thanks to Ben LaHaise for precious feedback.
4  */
5 #include <linux/highmem.h>
6 #include <linux/bootmem.h>
7 #include <linux/module.h>
8 #include <linux/sched.h>
9 #include <linux/slab.h>
10 #include <linux/mm.h>
11 #include <linux/interrupt.h>
12 #include <linux/seq_file.h>
13 #include <linux/debugfs.h>
14
15 #include <asm/e820.h>
16 #include <asm/processor.h>
17 #include <asm/tlbflush.h>
18 #include <asm/sections.h>
19 #include <asm/uaccess.h>
20 #include <asm/pgalloc.h>
21 #include <asm/proto.h>
22 #include <asm/pat.h>
23
24 /*
25  * The current flushing context - we pass it instead of 5 arguments:
26  */
27 struct cpa_data {
28         unsigned long   vaddr;
29         pgprot_t        mask_set;
30         pgprot_t        mask_clr;
31         int             numpages;
32         int             flushtlb;
33         unsigned long   pfn;
34         unsigned        force_split : 1;
35 };
36
37 #ifdef CONFIG_PROC_FS
38 static unsigned long direct_pages_count[PG_LEVEL_NUM];
39
40 void update_page_count(int level, unsigned long pages)
41 {
42         unsigned long flags;
43
44         /* Protect against CPA */
45         spin_lock_irqsave(&pgd_lock, flags);
46         direct_pages_count[level] += pages;
47         spin_unlock_irqrestore(&pgd_lock, flags);
48 }
49
50 static void split_page_count(int level)
51 {
52         direct_pages_count[level]--;
53         direct_pages_count[level - 1] += PTRS_PER_PTE;
54 }
55
56 int arch_report_meminfo(char *page)
57 {
58         int n = sprintf(page, "DirectMap4k:  %8lu\n"
59                         "DirectMap2M:  %8lu\n",
60                         direct_pages_count[PG_LEVEL_4K],
61                         direct_pages_count[PG_LEVEL_2M]);
62 #ifdef CONFIG_X86_64
63         n += sprintf(page + n, "DirectMap1G:  %8lu\n",
64                      direct_pages_count[PG_LEVEL_1G]);
65 #endif
66         return n;
67 }
68 #else
69 static inline void split_page_count(int level) { }
70 #endif
71
72 #ifdef CONFIG_X86_64
73
74 static inline unsigned long highmap_start_pfn(void)
75 {
76         return __pa(_text) >> PAGE_SHIFT;
77 }
78
79 static inline unsigned long highmap_end_pfn(void)
80 {
81         return __pa(round_up((unsigned long)_end, PMD_SIZE)) >> PAGE_SHIFT;
82 }
83
84 #endif
85
86 #ifdef CONFIG_DEBUG_PAGEALLOC
87 # define debug_pagealloc 1
88 #else
89 # define debug_pagealloc 0
90 #endif
91
92 static inline int
93 within(unsigned long addr, unsigned long start, unsigned long end)
94 {
95         return addr >= start && addr < end;
96 }
97
98 /*
99  * Flushing functions
100  */
101
102 /**
103  * clflush_cache_range - flush a cache range with clflush
104  * @addr:       virtual start address
105  * @size:       number of bytes to flush
106  *
107  * clflush is an unordered instruction which needs fencing with mfence
108  * to avoid ordering issues.
109  */
110 void clflush_cache_range(void *vaddr, unsigned int size)
111 {
112         void *vend = vaddr + size - 1;
113
114         mb();
115
116         for (; vaddr < vend; vaddr += boot_cpu_data.x86_clflush_size)
117                 clflush(vaddr);
118         /*
119          * Flush any possible final partial cacheline:
120          */
121         clflush(vend);
122
123         mb();
124 }
125
126 static void __cpa_flush_all(void *arg)
127 {
128         unsigned long cache = (unsigned long)arg;
129
130         /*
131          * Flush all to work around Errata in early athlons regarding
132          * large page flushing.
133          */
134         __flush_tlb_all();
135
136         if (cache && boot_cpu_data.x86_model >= 4)
137                 wbinvd();
138 }
139
140 static void cpa_flush_all(unsigned long cache)
141 {
142         BUG_ON(irqs_disabled());
143
144         on_each_cpu(__cpa_flush_all, (void *) cache, 1);
145 }
146
147 static void __cpa_flush_range(void *arg)
148 {
149         /*
150          * We could optimize that further and do individual per page
151          * tlb invalidates for a low number of pages. Caveat: we must
152          * flush the high aliases on 64bit as well.
153          */
154         __flush_tlb_all();
155 }
156
157 static void cpa_flush_range(unsigned long start, int numpages, int cache)
158 {
159         unsigned int i, level;
160         unsigned long addr;
161
162         BUG_ON(irqs_disabled());
163         WARN_ON(PAGE_ALIGN(start) != start);
164
165         on_each_cpu(__cpa_flush_range, NULL, 1);
166
167         if (!cache)
168                 return;
169
170         /*
171          * We only need to flush on one CPU,
172          * clflush is a MESI-coherent instruction that
173          * will cause all other CPUs to flush the same
174          * cachelines:
175          */
176         for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
177                 pte_t *pte = lookup_address(addr, &level);
178
179                 /*
180                  * Only flush present addresses:
181                  */
182                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
183                         clflush_cache_range((void *) addr, PAGE_SIZE);
184         }
185 }
186
187 /*
188  * Certain areas of memory on x86 require very specific protection flags,
189  * for example the BIOS area or kernel text. Callers don't always get this
190  * right (again, ioremap() on BIOS memory is not uncommon) so this function
191  * checks and fixes these known static required protection bits.
192  */
193 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
194                                    unsigned long pfn)
195 {
196         pgprot_t forbidden = __pgprot(0);
197
198         /*
199          * The BIOS area between 640k and 1Mb needs to be executable for
200          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
201          */
202         if (within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
203                 pgprot_val(forbidden) |= _PAGE_NX;
204
205         /*
206          * The kernel text needs to be executable for obvious reasons
207          * Does not cover __inittext since that is gone later on. On
208          * 64bit we do not enforce !NX on the low mapping
209          */
210         if (within(address, (unsigned long)_text, (unsigned long)_etext))
211                 pgprot_val(forbidden) |= _PAGE_NX;
212
213         /*
214          * The .rodata section needs to be read-only. Using the pfn
215          * catches all aliases.
216          */
217         if (within(pfn, __pa((unsigned long)__start_rodata) >> PAGE_SHIFT,
218                    __pa((unsigned long)__end_rodata) >> PAGE_SHIFT))
219                 pgprot_val(forbidden) |= _PAGE_RW;
220
221         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
222
223         return prot;
224 }
225
226 /*
227  * Lookup the page table entry for a virtual address. Return a pointer
228  * to the entry and the level of the mapping.
229  *
230  * Note: We return pud and pmd either when the entry is marked large
231  * or when the present bit is not set. Otherwise we would return a
232  * pointer to a nonexisting mapping.
233  */
234 pte_t *lookup_address(unsigned long address, unsigned int *level)
235 {
236         pgd_t *pgd = pgd_offset_k(address);
237         pud_t *pud;
238         pmd_t *pmd;
239
240         *level = PG_LEVEL_NONE;
241
242         if (pgd_none(*pgd))
243                 return NULL;
244
245         pud = pud_offset(pgd, address);
246         if (pud_none(*pud))
247                 return NULL;
248
249         *level = PG_LEVEL_1G;
250         if (pud_large(*pud) || !pud_present(*pud))
251                 return (pte_t *)pud;
252
253         pmd = pmd_offset(pud, address);
254         if (pmd_none(*pmd))
255                 return NULL;
256
257         *level = PG_LEVEL_2M;
258         if (pmd_large(*pmd) || !pmd_present(*pmd))
259                 return (pte_t *)pmd;
260
261         *level = PG_LEVEL_4K;
262
263         return pte_offset_kernel(pmd, address);
264 }
265 EXPORT_SYMBOL_GPL(lookup_address);
266
267 /*
268  * Set the new pmd in all the pgds we know about:
269  */
270 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
271 {
272         /* change init_mm */
273         set_pte_atomic(kpte, pte);
274 #ifdef CONFIG_X86_32
275         if (!SHARED_KERNEL_PMD) {
276                 struct page *page;
277
278                 list_for_each_entry(page, &pgd_list, lru) {
279                         pgd_t *pgd;
280                         pud_t *pud;
281                         pmd_t *pmd;
282
283                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
284                         pud = pud_offset(pgd, address);
285                         pmd = pmd_offset(pud, address);
286                         set_pte_atomic((pte_t *)pmd, pte);
287                 }
288         }
289 #endif
290 }
291
292 static int
293 try_preserve_large_page(pte_t *kpte, unsigned long address,
294                         struct cpa_data *cpa)
295 {
296         unsigned long nextpage_addr, numpages, pmask, psize, flags, addr, pfn;
297         pte_t new_pte, old_pte, *tmp;
298         pgprot_t old_prot, new_prot;
299         int i, do_split = 1;
300         unsigned int level;
301
302         if (cpa->force_split)
303                 return 1;
304
305         spin_lock_irqsave(&pgd_lock, flags);
306         /*
307          * Check for races, another CPU might have split this page
308          * up already:
309          */
310         tmp = lookup_address(address, &level);
311         if (tmp != kpte)
312                 goto out_unlock;
313
314         switch (level) {
315         case PG_LEVEL_2M:
316                 psize = PMD_PAGE_SIZE;
317                 pmask = PMD_PAGE_MASK;
318                 break;
319 #ifdef CONFIG_X86_64
320         case PG_LEVEL_1G:
321                 psize = PUD_PAGE_SIZE;
322                 pmask = PUD_PAGE_MASK;
323                 break;
324 #endif
325         default:
326                 do_split = -EINVAL;
327                 goto out_unlock;
328         }
329
330         /*
331          * Calculate the number of pages, which fit into this large
332          * page starting at address:
333          */
334         nextpage_addr = (address + psize) & pmask;
335         numpages = (nextpage_addr - address) >> PAGE_SHIFT;
336         if (numpages < cpa->numpages)
337                 cpa->numpages = numpages;
338
339         /*
340          * We are safe now. Check whether the new pgprot is the same:
341          */
342         old_pte = *kpte;
343         old_prot = new_prot = pte_pgprot(old_pte);
344
345         pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
346         pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
347
348         /*
349          * old_pte points to the large page base address. So we need
350          * to add the offset of the virtual address:
351          */
352         pfn = pte_pfn(old_pte) + ((address & (psize - 1)) >> PAGE_SHIFT);
353         cpa->pfn = pfn;
354
355         new_prot = static_protections(new_prot, address, pfn);
356
357         /*
358          * We need to check the full range, whether
359          * static_protection() requires a different pgprot for one of
360          * the pages in the range we try to preserve:
361          */
362         addr = address + PAGE_SIZE;
363         pfn++;
364         for (i = 1; i < cpa->numpages; i++, addr += PAGE_SIZE, pfn++) {
365                 pgprot_t chk_prot = static_protections(new_prot, addr, pfn);
366
367                 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
368                         goto out_unlock;
369         }
370
371         /*
372          * If there are no changes, return. maxpages has been updated
373          * above:
374          */
375         if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
376                 do_split = 0;
377                 goto out_unlock;
378         }
379
380         /*
381          * We need to change the attributes. Check, whether we can
382          * change the large page in one go. We request a split, when
383          * the address is not aligned and the number of pages is
384          * smaller than the number of pages in the large page. Note
385          * that we limited the number of possible pages already to
386          * the number of pages in the large page.
387          */
388         if (address == (nextpage_addr - psize) && cpa->numpages == numpages) {
389                 /*
390                  * The address is aligned and the number of pages
391                  * covers the full page.
392                  */
393                 new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
394                 __set_pmd_pte(kpte, address, new_pte);
395                 cpa->flushtlb = 1;
396                 do_split = 0;
397         }
398
399 out_unlock:
400         spin_unlock_irqrestore(&pgd_lock, flags);
401
402         return do_split;
403 }
404
405 static LIST_HEAD(page_pool);
406 static unsigned long pool_size, pool_pages, pool_low;
407 static unsigned long pool_used, pool_failed;
408
409 static void cpa_fill_pool(struct page **ret)
410 {
411         gfp_t gfp = GFP_KERNEL;
412         unsigned long flags;
413         struct page *p;
414
415         /*
416          * Avoid recursion (on debug-pagealloc) and also signal
417          * our priority to get to these pagetables:
418          */
419         if (current->flags & PF_MEMALLOC)
420                 return;
421         current->flags |= PF_MEMALLOC;
422
423         /*
424          * Allocate atomically from atomic contexts:
425          */
426         if (in_atomic() || irqs_disabled() || debug_pagealloc)
427                 gfp =  GFP_ATOMIC | __GFP_NORETRY | __GFP_NOWARN;
428
429         while (pool_pages < pool_size || (ret && !*ret)) {
430                 p = alloc_pages(gfp, 0);
431                 if (!p) {
432                         pool_failed++;
433                         break;
434                 }
435                 /*
436                  * If the call site needs a page right now, provide it:
437                  */
438                 if (ret && !*ret) {
439                         *ret = p;
440                         continue;
441                 }
442                 spin_lock_irqsave(&pgd_lock, flags);
443                 list_add(&p->lru, &page_pool);
444                 pool_pages++;
445                 spin_unlock_irqrestore(&pgd_lock, flags);
446         }
447
448         current->flags &= ~PF_MEMALLOC;
449 }
450
451 #define SHIFT_MB                (20 - PAGE_SHIFT)
452 #define ROUND_MB_GB             ((1 << 10) - 1)
453 #define SHIFT_MB_GB             10
454 #define POOL_PAGES_PER_GB       16
455
456 void __init cpa_init(void)
457 {
458         struct sysinfo si;
459         unsigned long gb;
460
461         si_meminfo(&si);
462         /*
463          * Calculate the number of pool pages:
464          *
465          * Convert totalram (nr of pages) to MiB and round to the next
466          * GiB. Shift MiB to Gib and multiply the result by
467          * POOL_PAGES_PER_GB:
468          */
469         if (debug_pagealloc) {
470                 gb = ((si.totalram >> SHIFT_MB) + ROUND_MB_GB) >> SHIFT_MB_GB;
471                 pool_size = POOL_PAGES_PER_GB * gb;
472         } else {
473                 pool_size = 1;
474         }
475         pool_low = pool_size;
476
477         cpa_fill_pool(NULL);
478         printk(KERN_DEBUG
479                "CPA: page pool initialized %lu of %lu pages preallocated\n",
480                pool_pages, pool_size);
481 }
482
483 static int split_large_page(pte_t *kpte, unsigned long address)
484 {
485         unsigned long flags, pfn, pfninc = 1;
486         unsigned int i, level;
487         pte_t *pbase, *tmp;
488         pgprot_t ref_prot;
489         struct page *base;
490
491         /*
492          * Get a page from the pool. The pool list is protected by the
493          * pgd_lock, which we have to take anyway for the split
494          * operation:
495          */
496         spin_lock_irqsave(&pgd_lock, flags);
497         if (list_empty(&page_pool)) {
498                 spin_unlock_irqrestore(&pgd_lock, flags);
499                 base = NULL;
500                 cpa_fill_pool(&base);
501                 if (!base)
502                         return -ENOMEM;
503                 spin_lock_irqsave(&pgd_lock, flags);
504         } else {
505                 base = list_first_entry(&page_pool, struct page, lru);
506                 list_del(&base->lru);
507                 pool_pages--;
508
509                 if (pool_pages < pool_low)
510                         pool_low = pool_pages;
511         }
512
513         /*
514          * Check for races, another CPU might have split this page
515          * up for us already:
516          */
517         tmp = lookup_address(address, &level);
518         if (tmp != kpte)
519                 goto out_unlock;
520
521         pbase = (pte_t *)page_address(base);
522         paravirt_alloc_pte(&init_mm, page_to_pfn(base));
523         ref_prot = pte_pgprot(pte_clrhuge(*kpte));
524
525 #ifdef CONFIG_X86_64
526         if (level == PG_LEVEL_1G) {
527                 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
528                 pgprot_val(ref_prot) |= _PAGE_PSE;
529         }
530 #endif
531
532         /*
533          * Get the target pfn from the original entry:
534          */
535         pfn = pte_pfn(*kpte);
536         for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
537                 set_pte(&pbase[i], pfn_pte(pfn, ref_prot));
538
539         if (address >= (unsigned long)__va(0) &&
540                 address < (unsigned long)__va(max_low_pfn_mapped << PAGE_SHIFT))
541                 split_page_count(level);
542
543 #ifdef CONFIG_X86_64
544         if (address >= (unsigned long)__va(1UL<<32) &&
545                 address < (unsigned long)__va(max_pfn_mapped << PAGE_SHIFT))
546                 split_page_count(level);
547 #endif
548
549         /*
550          * Install the new, split up pagetable. Important details here:
551          *
552          * On Intel the NX bit of all levels must be cleared to make a
553          * page executable. See section 4.13.2 of Intel 64 and IA-32
554          * Architectures Software Developer's Manual).
555          *
556          * Mark the entry present. The current mapping might be
557          * set to not present, which we preserved above.
558          */
559         ref_prot = pte_pgprot(pte_mkexec(pte_clrhuge(*kpte)));
560         pgprot_val(ref_prot) |= _PAGE_PRESENT;
561         __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
562         base = NULL;
563
564 out_unlock:
565         /*
566          * If we dropped out via the lookup_address check under
567          * pgd_lock then stick the page back into the pool:
568          */
569         if (base) {
570                 list_add(&base->lru, &page_pool);
571                 pool_pages++;
572         } else
573                 pool_used++;
574         spin_unlock_irqrestore(&pgd_lock, flags);
575
576         return 0;
577 }
578
579 static int __change_page_attr(struct cpa_data *cpa, int primary)
580 {
581         unsigned long address = cpa->vaddr;
582         int do_split, err;
583         unsigned int level;
584         pte_t *kpte, old_pte;
585
586 repeat:
587         kpte = lookup_address(address, &level);
588         if (!kpte)
589                 return 0;
590
591         old_pte = *kpte;
592         if (!pte_val(old_pte)) {
593                 if (!primary)
594                         return 0;
595                 printk(KERN_WARNING "CPA: called for zero pte. "
596                        "vaddr = %lx cpa->vaddr = %lx\n", address,
597                        cpa->vaddr);
598                 WARN_ON(1);
599                 return -EINVAL;
600         }
601
602         if (level == PG_LEVEL_4K) {
603                 pte_t new_pte;
604                 pgprot_t new_prot = pte_pgprot(old_pte);
605                 unsigned long pfn = pte_pfn(old_pte);
606
607                 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
608                 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
609
610                 new_prot = static_protections(new_prot, address, pfn);
611
612                 /*
613                  * We need to keep the pfn from the existing PTE,
614                  * after all we're only going to change it's attributes
615                  * not the memory it points to
616                  */
617                 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
618                 cpa->pfn = pfn;
619                 /*
620                  * Do we really change anything ?
621                  */
622                 if (pte_val(old_pte) != pte_val(new_pte)) {
623                         set_pte_atomic(kpte, new_pte);
624                         cpa->flushtlb = 1;
625                 }
626                 cpa->numpages = 1;
627                 return 0;
628         }
629
630         /*
631          * Check, whether we can keep the large page intact
632          * and just change the pte:
633          */
634         do_split = try_preserve_large_page(kpte, address, cpa);
635         /*
636          * When the range fits into the existing large page,
637          * return. cp->numpages and cpa->tlbflush have been updated in
638          * try_large_page:
639          */
640         if (do_split <= 0)
641                 return do_split;
642
643         /*
644          * We have to split the large page:
645          */
646         err = split_large_page(kpte, address);
647         if (!err) {
648                 cpa->flushtlb = 1;
649                 goto repeat;
650         }
651
652         return err;
653 }
654
655 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
656
657 static int cpa_process_alias(struct cpa_data *cpa)
658 {
659         struct cpa_data alias_cpa;
660         int ret = 0;
661
662         if (cpa->pfn >= max_pfn_mapped)
663                 return 0;
664
665 #ifdef CONFIG_X86_64
666         if (cpa->pfn >= max_low_pfn_mapped && cpa->pfn < (1UL<<(32-PAGE_SHIFT)))
667                 return 0;
668 #endif
669         /*
670          * No need to redo, when the primary call touched the direct
671          * mapping already:
672          */
673         if (!(within(cpa->vaddr, PAGE_OFFSET,
674                     PAGE_OFFSET + (max_low_pfn_mapped << PAGE_SHIFT))
675 #ifdef CONFIG_X86_64
676                 || within(cpa->vaddr, PAGE_OFFSET + (1UL<<32),
677                     PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))
678 #endif
679         )) {
680
681                 alias_cpa = *cpa;
682                 alias_cpa.vaddr = (unsigned long) __va(cpa->pfn << PAGE_SHIFT);
683
684                 ret = __change_page_attr_set_clr(&alias_cpa, 0);
685         }
686
687 #ifdef CONFIG_X86_64
688         if (ret)
689                 return ret;
690         /*
691          * No need to redo, when the primary call touched the high
692          * mapping already:
693          */
694         if (within(cpa->vaddr, (unsigned long) _text, (unsigned long) _end))
695                 return 0;
696
697         /*
698          * If the physical address is inside the kernel map, we need
699          * to touch the high mapped kernel as well:
700          */
701         if (!within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn()))
702                 return 0;
703
704         alias_cpa = *cpa;
705         alias_cpa.vaddr =
706                 (cpa->pfn << PAGE_SHIFT) + __START_KERNEL_map - phys_base;
707
708         /*
709          * The high mapping range is imprecise, so ignore the return value.
710          */
711         __change_page_attr_set_clr(&alias_cpa, 0);
712 #endif
713         return ret;
714 }
715
716 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
717 {
718         int ret, numpages = cpa->numpages;
719
720         while (numpages) {
721                 /*
722                  * Store the remaining nr of pages for the large page
723                  * preservation check.
724                  */
725                 cpa->numpages = numpages;
726
727                 ret = __change_page_attr(cpa, checkalias);
728                 if (ret)
729                         return ret;
730
731                 if (checkalias) {
732                         ret = cpa_process_alias(cpa);
733                         if (ret)
734                                 return ret;
735                 }
736
737                 /*
738                  * Adjust the number of pages with the result of the
739                  * CPA operation. Either a large page has been
740                  * preserved or a single page update happened.
741                  */
742                 BUG_ON(cpa->numpages > numpages);
743                 numpages -= cpa->numpages;
744                 cpa->vaddr += cpa->numpages * PAGE_SIZE;
745         }
746         return 0;
747 }
748
749 static inline int cache_attr(pgprot_t attr)
750 {
751         return pgprot_val(attr) &
752                 (_PAGE_PAT | _PAGE_PAT_LARGE | _PAGE_PWT | _PAGE_PCD);
753 }
754
755 static int do_change_page_attr_set_clr(unsigned long addr, int numpages,
756                                     pgprot_t mask_set, pgprot_t mask_clr,
757                                     int force_split, int *tlb_flush)
758 {
759         struct cpa_data cpa;
760         int ret, checkalias;
761
762         /*
763          * Check, if we are requested to change a not supported
764          * feature:
765          */
766         mask_set = canon_pgprot(mask_set);
767         mask_clr = canon_pgprot(mask_clr);
768         if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
769                 return 0;
770
771         /* Ensure we are PAGE_SIZE aligned */
772         if (addr & ~PAGE_MASK) {
773                 addr &= PAGE_MASK;
774                 /*
775                  * People should not be passing in unaligned addresses:
776                  */
777                 WARN_ON_ONCE(1);
778         }
779
780         /* Must avoid aliasing mappings in the highmem code */
781         kmap_flush_unused();
782
783         cpa.vaddr = addr;
784         cpa.numpages = numpages;
785         cpa.mask_set = mask_set;
786         cpa.mask_clr = mask_clr;
787         cpa.flushtlb = 0;
788         cpa.force_split = force_split;
789
790         /* No alias checking for _NX bit modifications */
791         checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
792
793         ret = __change_page_attr_set_clr(&cpa, checkalias);
794
795         /*
796          * Check whether we really changed something:
797          */
798         *tlb_flush = cpa.flushtlb;
799         cpa_fill_pool(NULL);
800
801         return ret;
802 }
803
804 static int change_page_attr_set_clr(unsigned long addr, int numpages,
805                                     pgprot_t mask_set, pgprot_t mask_clr,
806                                     int force_split)
807 {
808         int cache, flush_cache = 0, ret;
809
810         ret = do_change_page_attr_set_clr(addr, numpages, mask_set, mask_clr,
811                 force_split, &flush_cache);
812         if (!flush_cache)
813                 goto out;
814         /*
815          * No need to flush, when we did not set any of the caching
816          * attributes:
817          */
818         cache = cache_attr(mask_set);
819
820         /*
821          * On success we use clflush, when the CPU supports it to
822          * avoid the wbindv. If the CPU does not support it and in the
823          * error case we fall back to cpa_flush_all (which uses
824          * wbindv):
825          */
826         if (!ret && cpu_has_clflush)
827                 cpa_flush_range(addr, numpages, cache);
828         else
829                 cpa_flush_all(cache);
830 out:
831         return ret;
832 }
833
834 static inline int change_page_attr_set(unsigned long addr, int numpages,
835                                        pgprot_t mask)
836 {
837         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0);
838 }
839
840 static inline int change_page_attr_clear(unsigned long addr, int numpages,
841                                          pgprot_t mask)
842 {
843         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0);
844 }
845
846 int _set_memory_uc(unsigned long addr, int numpages)
847 {
848         /*
849          * for now UC MINUS. see comments in ioremap_nocache()
850          */
851         return change_page_attr_set(addr, numpages,
852                                     __pgprot(_PAGE_CACHE_UC_MINUS));
853 }
854
855 int set_memory_uc(unsigned long addr, int numpages)
856 {
857         /*
858          * for now UC MINUS. see comments in ioremap_nocache()
859          */
860         if (reserve_memtype(addr, addr + numpages * PAGE_SIZE,
861                             _PAGE_CACHE_UC_MINUS, NULL))
862                 return -EINVAL;
863
864         return _set_memory_uc(addr, numpages);
865 }
866 EXPORT_SYMBOL(set_memory_uc);
867
868 int set_memory_uc_noflush(unsigned long addr, int numpages)
869 {
870         int flush;
871         /*
872          * for now UC MINUS. see comments in ioremap_nocache()
873          */
874         if (reserve_memtype(addr, addr + numpages * PAGE_SIZE,
875                             _PAGE_CACHE_UC_MINUS, NULL))
876                 return -EINVAL;
877         /*
878          * for now UC MINUS. see comments in ioremap_nocache()
879          */
880         return do_change_page_attr_set_clr(addr, numpages,
881                                     __pgprot(_PAGE_CACHE_UC_MINUS),
882                                         __pgprot(0), 0, &flush);
883 }
884 EXPORT_SYMBOL(set_memory_uc_noflush);
885
886 void set_memory_flush_all(void)
887 {
888         cpa_flush_all(1);
889 }
890 EXPORT_SYMBOL(set_memory_flush_all);
891
892 int _set_memory_wc(unsigned long addr, int numpages)
893 {
894         return change_page_attr_set(addr, numpages,
895                                     __pgprot(_PAGE_CACHE_WC));
896 }
897
898 int set_memory_wc(unsigned long addr, int numpages)
899 {
900         if (!pat_enabled)
901                 return set_memory_uc(addr, numpages);
902
903         if (reserve_memtype(addr, addr + numpages * PAGE_SIZE,
904                 _PAGE_CACHE_WC, NULL))
905                 return -EINVAL;
906
907         return _set_memory_wc(addr, numpages);
908 }
909 EXPORT_SYMBOL(set_memory_wc);
910
911 int _set_memory_wb(unsigned long addr, int numpages)
912 {
913         return change_page_attr_clear(addr, numpages,
914                                       __pgprot(_PAGE_CACHE_MASK));
915 }
916
917 int set_memory_wb(unsigned long addr, int numpages)
918 {
919         free_memtype(addr, addr + numpages * PAGE_SIZE);
920
921         return _set_memory_wb(addr, numpages);
922 }
923 EXPORT_SYMBOL(set_memory_wb);
924
925 int set_memory_x(unsigned long addr, int numpages)
926 {
927         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_NX));
928 }
929 EXPORT_SYMBOL(set_memory_x);
930
931 int set_memory_nx(unsigned long addr, int numpages)
932 {
933         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_NX));
934 }
935 EXPORT_SYMBOL(set_memory_nx);
936
937 int set_memory_ro(unsigned long addr, int numpages)
938 {
939         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_RW));
940 }
941
942 int set_memory_rw(unsigned long addr, int numpages)
943 {
944         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_RW));
945 }
946
947 int set_memory_np(unsigned long addr, int numpages)
948 {
949         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_PRESENT));
950 }
951
952 int set_memory_4k(unsigned long addr, int numpages)
953 {
954         return change_page_attr_set_clr(addr, numpages, __pgprot(0),
955                                         __pgprot(0), 1);
956 }
957
958 int set_pages_uc(struct page *page, int numpages)
959 {
960         unsigned long addr = (unsigned long)page_address(page);
961
962         return set_memory_uc(addr, numpages);
963 }
964 EXPORT_SYMBOL(set_pages_uc);
965
966 int set_pages_uc_noflush(struct page *page, int numpages)
967 {
968         unsigned long addr = (unsigned long)page_address(page);
969
970         return set_memory_uc_noflush(addr, numpages);
971 }
972 EXPORT_SYMBOL(set_pages_uc_noflush);
973
974 int set_pages_wb(struct page *page, int numpages)
975 {
976         unsigned long addr = (unsigned long)page_address(page);
977
978         return set_memory_wb(addr, numpages);
979 }
980 EXPORT_SYMBOL(set_pages_wb);
981
982 int set_pages_x(struct page *page, int numpages)
983 {
984         unsigned long addr = (unsigned long)page_address(page);
985
986         return set_memory_x(addr, numpages);
987 }
988 EXPORT_SYMBOL(set_pages_x);
989
990 int set_pages_nx(struct page *page, int numpages)
991 {
992         unsigned long addr = (unsigned long)page_address(page);
993
994         return set_memory_nx(addr, numpages);
995 }
996 EXPORT_SYMBOL(set_pages_nx);
997
998 int set_pages_ro(struct page *page, int numpages)
999 {
1000         unsigned long addr = (unsigned long)page_address(page);
1001
1002         return set_memory_ro(addr, numpages);
1003 }
1004
1005 int set_pages_rw(struct page *page, int numpages)
1006 {
1007         unsigned long addr = (unsigned long)page_address(page);
1008
1009         return set_memory_rw(addr, numpages);
1010 }
1011
1012 #ifdef CONFIG_DEBUG_PAGEALLOC
1013
1014 static int __set_pages_p(struct page *page, int numpages)
1015 {
1016         struct cpa_data cpa = { .vaddr = (unsigned long) page_address(page),
1017                                 .numpages = numpages,
1018                                 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1019                                 .mask_clr = __pgprot(0)};
1020
1021         return __change_page_attr_set_clr(&cpa, 1);
1022 }
1023
1024 static int __set_pages_np(struct page *page, int numpages)
1025 {
1026         struct cpa_data cpa = { .vaddr = (unsigned long) page_address(page),
1027                                 .numpages = numpages,
1028                                 .mask_set = __pgprot(0),
1029                                 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW)};
1030
1031         return __change_page_attr_set_clr(&cpa, 1);
1032 }
1033
1034 void kernel_map_pages(struct page *page, int numpages, int enable)
1035 {
1036         if (PageHighMem(page))
1037                 return;
1038         if (!enable) {
1039                 debug_check_no_locks_freed(page_address(page),
1040                                            numpages * PAGE_SIZE);
1041         }
1042
1043         /*
1044          * If page allocator is not up yet then do not call c_p_a():
1045          */
1046         if (!debug_pagealloc_enabled)
1047                 return;
1048
1049         /*
1050          * The return value is ignored as the calls cannot fail.
1051          * Large pages are kept enabled at boot time, and are
1052          * split up quickly with DEBUG_PAGEALLOC. If a splitup
1053          * fails here (due to temporary memory shortage) no damage
1054          * is done because we just keep the largepage intact up
1055          * to the next attempt when it will likely be split up:
1056          */
1057         if (enable)
1058                 __set_pages_p(page, numpages);
1059         else
1060                 __set_pages_np(page, numpages);
1061
1062         /*
1063          * We should perform an IPI and flush all tlbs,
1064          * but that can deadlock->flush only current cpu:
1065          */
1066         __flush_tlb_all();
1067
1068         /*
1069          * Try to refill the page pool here. We can do this only after
1070          * the tlb flush.
1071          */
1072         cpa_fill_pool(NULL);
1073 }
1074
1075 #ifdef CONFIG_DEBUG_FS
1076 static int dpa_show(struct seq_file *m, void *v)
1077 {
1078         seq_puts(m, "DEBUG_PAGEALLOC\n");
1079         seq_printf(m, "pool_size     : %lu\n", pool_size);
1080         seq_printf(m, "pool_pages    : %lu\n", pool_pages);
1081         seq_printf(m, "pool_low      : %lu\n", pool_low);
1082         seq_printf(m, "pool_used     : %lu\n", pool_used);
1083         seq_printf(m, "pool_failed   : %lu\n", pool_failed);
1084
1085         return 0;
1086 }
1087
1088 static int dpa_open(struct inode *inode, struct file *filp)
1089 {
1090         return single_open(filp, dpa_show, NULL);
1091 }
1092
1093 static const struct file_operations dpa_fops = {
1094         .open           = dpa_open,
1095         .read           = seq_read,
1096         .llseek         = seq_lseek,
1097         .release        = single_release,
1098 };
1099
1100 static int __init debug_pagealloc_proc_init(void)
1101 {
1102         struct dentry *de;
1103
1104         de = debugfs_create_file("debug_pagealloc", 0600, NULL, NULL,
1105                                  &dpa_fops);
1106         if (!de)
1107                 return -ENOMEM;
1108
1109         return 0;
1110 }
1111 __initcall(debug_pagealloc_proc_init);
1112 #endif
1113
1114 #ifdef CONFIG_HIBERNATION
1115
1116 bool kernel_page_present(struct page *page)
1117 {
1118         unsigned int level;
1119         pte_t *pte;
1120
1121         if (PageHighMem(page))
1122                 return false;
1123
1124         pte = lookup_address((unsigned long)page_address(page), &level);
1125         return (pte_val(*pte) & _PAGE_PRESENT);
1126 }
1127
1128 #endif /* CONFIG_HIBERNATION */
1129
1130 #endif /* CONFIG_DEBUG_PAGEALLOC */
1131
1132 /*
1133  * The testcases use internal knowledge of the implementation that shouldn't
1134  * be exposed to the rest of the kernel. Include these directly here.
1135  */
1136 #ifdef CONFIG_CPA_DEBUG
1137 #include "pageattr-test.c"
1138 #endif