]> www.pilppa.org Git - linux-2.6-omap-h63xx.git/blob - arch/x86/mm/pageattr.c
x86, generic: CPA add statistics about state of direct mapping v4
[linux-2.6-omap-h63xx.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 static unsigned long direct_pages_count[PG_LEVEL_NUM];
38
39 void __meminit update_page_count(int level, unsigned long pages)
40 {
41 #ifdef CONFIG_PROC_FS
42         unsigned long flags;
43         /* Protect against CPA */
44         spin_lock_irqsave(&pgd_lock, flags);
45         direct_pages_count[level] += pages;
46         spin_unlock_irqrestore(&pgd_lock, flags);
47 #endif
48 }
49
50 #ifdef CONFIG_X86_64
51
52 static inline unsigned long highmap_start_pfn(void)
53 {
54         return __pa(_text) >> PAGE_SHIFT;
55 }
56
57 static inline unsigned long highmap_end_pfn(void)
58 {
59         return __pa(round_up((unsigned long)_end, PMD_SIZE)) >> PAGE_SHIFT;
60 }
61
62 #endif
63
64 #ifdef CONFIG_DEBUG_PAGEALLOC
65 # define debug_pagealloc 1
66 #else
67 # define debug_pagealloc 0
68 #endif
69
70 static inline int
71 within(unsigned long addr, unsigned long start, unsigned long end)
72 {
73         return addr >= start && addr < end;
74 }
75
76 /*
77  * Flushing functions
78  */
79
80 /**
81  * clflush_cache_range - flush a cache range with clflush
82  * @addr:       virtual start address
83  * @size:       number of bytes to flush
84  *
85  * clflush is an unordered instruction which needs fencing with mfence
86  * to avoid ordering issues.
87  */
88 void clflush_cache_range(void *vaddr, unsigned int size)
89 {
90         void *vend = vaddr + size - 1;
91
92         mb();
93
94         for (; vaddr < vend; vaddr += boot_cpu_data.x86_clflush_size)
95                 clflush(vaddr);
96         /*
97          * Flush any possible final partial cacheline:
98          */
99         clflush(vend);
100
101         mb();
102 }
103
104 static void __cpa_flush_all(void *arg)
105 {
106         unsigned long cache = (unsigned long)arg;
107
108         /*
109          * Flush all to work around Errata in early athlons regarding
110          * large page flushing.
111          */
112         __flush_tlb_all();
113
114         if (cache && boot_cpu_data.x86_model >= 4)
115                 wbinvd();
116 }
117
118 static void cpa_flush_all(unsigned long cache)
119 {
120         BUG_ON(irqs_disabled());
121
122         on_each_cpu(__cpa_flush_all, (void *) cache, 1, 1);
123 }
124
125 static void __cpa_flush_range(void *arg)
126 {
127         /*
128          * We could optimize that further and do individual per page
129          * tlb invalidates for a low number of pages. Caveat: we must
130          * flush the high aliases on 64bit as well.
131          */
132         __flush_tlb_all();
133 }
134
135 static void cpa_flush_range(unsigned long start, int numpages, int cache)
136 {
137         unsigned int i, level;
138         unsigned long addr;
139
140         BUG_ON(irqs_disabled());
141         WARN_ON(PAGE_ALIGN(start) != start);
142
143         on_each_cpu(__cpa_flush_range, NULL, 1, 1);
144
145         if (!cache)
146                 return;
147
148         /*
149          * We only need to flush on one CPU,
150          * clflush is a MESI-coherent instruction that
151          * will cause all other CPUs to flush the same
152          * cachelines:
153          */
154         for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
155                 pte_t *pte = lookup_address(addr, &level);
156
157                 /*
158                  * Only flush present addresses:
159                  */
160                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
161                         clflush_cache_range((void *) addr, PAGE_SIZE);
162         }
163 }
164
165 /*
166  * Certain areas of memory on x86 require very specific protection flags,
167  * for example the BIOS area or kernel text. Callers don't always get this
168  * right (again, ioremap() on BIOS memory is not uncommon) so this function
169  * checks and fixes these known static required protection bits.
170  */
171 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
172                                    unsigned long pfn)
173 {
174         pgprot_t forbidden = __pgprot(0);
175
176         /*
177          * The BIOS area between 640k and 1Mb needs to be executable for
178          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
179          */
180         if (within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
181                 pgprot_val(forbidden) |= _PAGE_NX;
182
183         /*
184          * The kernel text needs to be executable for obvious reasons
185          * Does not cover __inittext since that is gone later on. On
186          * 64bit we do not enforce !NX on the low mapping
187          */
188         if (within(address, (unsigned long)_text, (unsigned long)_etext))
189                 pgprot_val(forbidden) |= _PAGE_NX;
190
191         /*
192          * The .rodata section needs to be read-only. Using the pfn
193          * catches all aliases.
194          */
195         if (within(pfn, __pa((unsigned long)__start_rodata) >> PAGE_SHIFT,
196                    __pa((unsigned long)__end_rodata) >> PAGE_SHIFT))
197                 pgprot_val(forbidden) |= _PAGE_RW;
198
199         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
200
201         return prot;
202 }
203
204 /*
205  * Lookup the page table entry for a virtual address. Return a pointer
206  * to the entry and the level of the mapping.
207  *
208  * Note: We return pud and pmd either when the entry is marked large
209  * or when the present bit is not set. Otherwise we would return a
210  * pointer to a nonexisting mapping.
211  */
212 pte_t *lookup_address(unsigned long address, unsigned int *level)
213 {
214         pgd_t *pgd = pgd_offset_k(address);
215         pud_t *pud;
216         pmd_t *pmd;
217
218         *level = PG_LEVEL_NONE;
219
220         if (pgd_none(*pgd))
221                 return NULL;
222
223         pud = pud_offset(pgd, address);
224         if (pud_none(*pud))
225                 return NULL;
226
227         *level = PG_LEVEL_1G;
228         if (pud_large(*pud) || !pud_present(*pud))
229                 return (pte_t *)pud;
230
231         pmd = pmd_offset(pud, address);
232         if (pmd_none(*pmd))
233                 return NULL;
234
235         *level = PG_LEVEL_2M;
236         if (pmd_large(*pmd) || !pmd_present(*pmd))
237                 return (pte_t *)pmd;
238
239         *level = PG_LEVEL_4K;
240
241         return pte_offset_kernel(pmd, address);
242 }
243
244 /*
245  * Set the new pmd in all the pgds we know about:
246  */
247 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
248 {
249         /* change init_mm */
250         set_pte_atomic(kpte, pte);
251 #ifdef CONFIG_X86_32
252         if (!SHARED_KERNEL_PMD) {
253                 struct page *page;
254
255                 list_for_each_entry(page, &pgd_list, lru) {
256                         pgd_t *pgd;
257                         pud_t *pud;
258                         pmd_t *pmd;
259
260                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
261                         pud = pud_offset(pgd, address);
262                         pmd = pmd_offset(pud, address);
263                         set_pte_atomic((pte_t *)pmd, pte);
264                 }
265         }
266 #endif
267 }
268
269 static int
270 try_preserve_large_page(pte_t *kpte, unsigned long address,
271                         struct cpa_data *cpa)
272 {
273         unsigned long nextpage_addr, numpages, pmask, psize, flags, addr, pfn;
274         pte_t new_pte, old_pte, *tmp;
275         pgprot_t old_prot, new_prot;
276         int i, do_split = 1;
277         unsigned int level;
278
279         if (cpa->force_split)
280                 return 1;
281
282         spin_lock_irqsave(&pgd_lock, flags);
283         /*
284          * Check for races, another CPU might have split this page
285          * up already:
286          */
287         tmp = lookup_address(address, &level);
288         if (tmp != kpte)
289                 goto out_unlock;
290
291         switch (level) {
292         case PG_LEVEL_2M:
293                 psize = PMD_PAGE_SIZE;
294                 pmask = PMD_PAGE_MASK;
295                 break;
296 #ifdef CONFIG_X86_64
297         case PG_LEVEL_1G:
298                 psize = PUD_PAGE_SIZE;
299                 pmask = PUD_PAGE_MASK;
300                 break;
301 #endif
302         default:
303                 do_split = -EINVAL;
304                 goto out_unlock;
305         }
306
307         /*
308          * Calculate the number of pages, which fit into this large
309          * page starting at address:
310          */
311         nextpage_addr = (address + psize) & pmask;
312         numpages = (nextpage_addr - address) >> PAGE_SHIFT;
313         if (numpages < cpa->numpages)
314                 cpa->numpages = numpages;
315
316         /*
317          * We are safe now. Check whether the new pgprot is the same:
318          */
319         old_pte = *kpte;
320         old_prot = new_prot = pte_pgprot(old_pte);
321
322         pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
323         pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
324
325         /*
326          * old_pte points to the large page base address. So we need
327          * to add the offset of the virtual address:
328          */
329         pfn = pte_pfn(old_pte) + ((address & (psize - 1)) >> PAGE_SHIFT);
330         cpa->pfn = pfn;
331
332         new_prot = static_protections(new_prot, address, pfn);
333
334         /*
335          * We need to check the full range, whether
336          * static_protection() requires a different pgprot for one of
337          * the pages in the range we try to preserve:
338          */
339         addr = address + PAGE_SIZE;
340         pfn++;
341         for (i = 1; i < cpa->numpages; i++, addr += PAGE_SIZE, pfn++) {
342                 pgprot_t chk_prot = static_protections(new_prot, addr, pfn);
343
344                 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
345                         goto out_unlock;
346         }
347
348         /*
349          * If there are no changes, return. maxpages has been updated
350          * above:
351          */
352         if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
353                 do_split = 0;
354                 goto out_unlock;
355         }
356
357         /*
358          * We need to change the attributes. Check, whether we can
359          * change the large page in one go. We request a split, when
360          * the address is not aligned and the number of pages is
361          * smaller than the number of pages in the large page. Note
362          * that we limited the number of possible pages already to
363          * the number of pages in the large page.
364          */
365         if (address == (nextpage_addr - psize) && cpa->numpages == numpages) {
366                 /*
367                  * The address is aligned and the number of pages
368                  * covers the full page.
369                  */
370                 new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
371                 __set_pmd_pte(kpte, address, new_pte);
372                 cpa->flushtlb = 1;
373                 do_split = 0;
374         }
375
376 out_unlock:
377         spin_unlock_irqrestore(&pgd_lock, flags);
378
379         return do_split;
380 }
381
382 static LIST_HEAD(page_pool);
383 static unsigned long pool_size, pool_pages, pool_low;
384 static unsigned long pool_used, pool_failed;
385
386 static void cpa_fill_pool(struct page **ret)
387 {
388         gfp_t gfp = GFP_KERNEL;
389         unsigned long flags;
390         struct page *p;
391
392         /*
393          * Avoid recursion (on debug-pagealloc) and also signal
394          * our priority to get to these pagetables:
395          */
396         if (current->flags & PF_MEMALLOC)
397                 return;
398         current->flags |= PF_MEMALLOC;
399
400         /*
401          * Allocate atomically from atomic contexts:
402          */
403         if (in_atomic() || irqs_disabled() || debug_pagealloc)
404                 gfp =  GFP_ATOMIC | __GFP_NORETRY | __GFP_NOWARN;
405
406         while (pool_pages < pool_size || (ret && !*ret)) {
407                 p = alloc_pages(gfp, 0);
408                 if (!p) {
409                         pool_failed++;
410                         break;
411                 }
412                 /*
413                  * If the call site needs a page right now, provide it:
414                  */
415                 if (ret && !*ret) {
416                         *ret = p;
417                         continue;
418                 }
419                 spin_lock_irqsave(&pgd_lock, flags);
420                 list_add(&p->lru, &page_pool);
421                 pool_pages++;
422                 spin_unlock_irqrestore(&pgd_lock, flags);
423         }
424
425         current->flags &= ~PF_MEMALLOC;
426 }
427
428 #define SHIFT_MB                (20 - PAGE_SHIFT)
429 #define ROUND_MB_GB             ((1 << 10) - 1)
430 #define SHIFT_MB_GB             10
431 #define POOL_PAGES_PER_GB       16
432
433 void __init cpa_init(void)
434 {
435         struct sysinfo si;
436         unsigned long gb;
437
438         si_meminfo(&si);
439         /*
440          * Calculate the number of pool pages:
441          *
442          * Convert totalram (nr of pages) to MiB and round to the next
443          * GiB. Shift MiB to Gib and multiply the result by
444          * POOL_PAGES_PER_GB:
445          */
446         if (debug_pagealloc) {
447                 gb = ((si.totalram >> SHIFT_MB) + ROUND_MB_GB) >> SHIFT_MB_GB;
448                 pool_size = POOL_PAGES_PER_GB * gb;
449         } else {
450                 pool_size = 1;
451         }
452         pool_low = pool_size;
453
454         cpa_fill_pool(NULL);
455         printk(KERN_DEBUG
456                "CPA: page pool initialized %lu of %lu pages preallocated\n",
457                pool_pages, pool_size);
458 }
459
460 static int split_large_page(pte_t *kpte, unsigned long address)
461 {
462         unsigned long flags, pfn, pfninc = 1;
463         unsigned int i, level;
464         pte_t *pbase, *tmp;
465         pgprot_t ref_prot;
466         struct page *base;
467
468         /*
469          * Get a page from the pool. The pool list is protected by the
470          * pgd_lock, which we have to take anyway for the split
471          * operation:
472          */
473         spin_lock_irqsave(&pgd_lock, flags);
474         if (list_empty(&page_pool)) {
475                 spin_unlock_irqrestore(&pgd_lock, flags);
476                 base = NULL;
477                 cpa_fill_pool(&base);
478                 if (!base)
479                         return -ENOMEM;
480                 spin_lock_irqsave(&pgd_lock, flags);
481         } else {
482                 base = list_first_entry(&page_pool, struct page, lru);
483                 list_del(&base->lru);
484                 pool_pages--;
485
486                 if (pool_pages < pool_low)
487                         pool_low = pool_pages;
488         }
489
490         /*
491          * Check for races, another CPU might have split this page
492          * up for us already:
493          */
494         tmp = lookup_address(address, &level);
495         if (tmp != kpte)
496                 goto out_unlock;
497
498         pbase = (pte_t *)page_address(base);
499         paravirt_alloc_pte(&init_mm, page_to_pfn(base));
500         ref_prot = pte_pgprot(pte_clrhuge(*kpte));
501
502 #ifdef CONFIG_X86_64
503         if (level == PG_LEVEL_1G) {
504                 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
505                 pgprot_val(ref_prot) |= _PAGE_PSE;
506         }
507 #endif
508
509         /*
510          * Get the target pfn from the original entry:
511          */
512         pfn = pte_pfn(*kpte);
513         for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
514                 set_pte(&pbase[i], pfn_pte(pfn, ref_prot));
515
516         if (address >= (unsigned long)__va(0) &&
517                 address < (unsigned long)__va(max_pfn_mapped << PAGE_SHIFT)) {
518                 direct_pages_count[level]--;
519                 direct_pages_count[level - 1] += PTRS_PER_PTE;
520         }
521
522         /*
523          * Install the new, split up pagetable. Important details here:
524          *
525          * On Intel the NX bit of all levels must be cleared to make a
526          * page executable. See section 4.13.2 of Intel 64 and IA-32
527          * Architectures Software Developer's Manual).
528          *
529          * Mark the entry present. The current mapping might be
530          * set to not present, which we preserved above.
531          */
532         ref_prot = pte_pgprot(pte_mkexec(pte_clrhuge(*kpte)));
533         pgprot_val(ref_prot) |= _PAGE_PRESENT;
534         __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
535         base = NULL;
536
537 out_unlock:
538         /*
539          * If we dropped out via the lookup_address check under
540          * pgd_lock then stick the page back into the pool:
541          */
542         if (base) {
543                 list_add(&base->lru, &page_pool);
544                 pool_pages++;
545         } else
546                 pool_used++;
547         spin_unlock_irqrestore(&pgd_lock, flags);
548
549         return 0;
550 }
551
552 static int __change_page_attr(struct cpa_data *cpa, int primary)
553 {
554         unsigned long address = cpa->vaddr;
555         int do_split, err;
556         unsigned int level;
557         pte_t *kpte, old_pte;
558
559 repeat:
560         kpte = lookup_address(address, &level);
561         if (!kpte)
562                 return 0;
563
564         old_pte = *kpte;
565         if (!pte_val(old_pte)) {
566                 if (!primary)
567                         return 0;
568                 printk(KERN_WARNING "CPA: called for zero pte. "
569                        "vaddr = %lx cpa->vaddr = %lx\n", address,
570                        cpa->vaddr);
571                 WARN_ON(1);
572                 return -EINVAL;
573         }
574
575         if (level == PG_LEVEL_4K) {
576                 pte_t new_pte;
577                 pgprot_t new_prot = pte_pgprot(old_pte);
578                 unsigned long pfn = pte_pfn(old_pte);
579
580                 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
581                 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
582
583                 new_prot = static_protections(new_prot, address, pfn);
584
585                 /*
586                  * We need to keep the pfn from the existing PTE,
587                  * after all we're only going to change it's attributes
588                  * not the memory it points to
589                  */
590                 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
591                 cpa->pfn = pfn;
592                 /*
593                  * Do we really change anything ?
594                  */
595                 if (pte_val(old_pte) != pte_val(new_pte)) {
596                         set_pte_atomic(kpte, new_pte);
597                         cpa->flushtlb = 1;
598                 }
599                 cpa->numpages = 1;
600                 return 0;
601         }
602
603         /*
604          * Check, whether we can keep the large page intact
605          * and just change the pte:
606          */
607         do_split = try_preserve_large_page(kpte, address, cpa);
608         /*
609          * When the range fits into the existing large page,
610          * return. cp->numpages and cpa->tlbflush have been updated in
611          * try_large_page:
612          */
613         if (do_split <= 0)
614                 return do_split;
615
616         /*
617          * We have to split the large page:
618          */
619         err = split_large_page(kpte, address);
620         if (!err) {
621                 cpa->flushtlb = 1;
622                 goto repeat;
623         }
624
625         return err;
626 }
627
628 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
629
630 static int cpa_process_alias(struct cpa_data *cpa)
631 {
632         struct cpa_data alias_cpa;
633         int ret = 0;
634
635         if (cpa->pfn > max_pfn_mapped)
636                 return 0;
637
638         /*
639          * No need to redo, when the primary call touched the direct
640          * mapping already:
641          */
642         if (!within(cpa->vaddr, PAGE_OFFSET,
643                     PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
644
645                 alias_cpa = *cpa;
646                 alias_cpa.vaddr = (unsigned long) __va(cpa->pfn << PAGE_SHIFT);
647
648                 ret = __change_page_attr_set_clr(&alias_cpa, 0);
649         }
650
651 #ifdef CONFIG_X86_64
652         if (ret)
653                 return ret;
654         /*
655          * No need to redo, when the primary call touched the high
656          * mapping already:
657          */
658         if (within(cpa->vaddr, (unsigned long) _text, (unsigned long) _end))
659                 return 0;
660
661         /*
662          * If the physical address is inside the kernel map, we need
663          * to touch the high mapped kernel as well:
664          */
665         if (!within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn()))
666                 return 0;
667
668         alias_cpa = *cpa;
669         alias_cpa.vaddr =
670                 (cpa->pfn << PAGE_SHIFT) + __START_KERNEL_map - phys_base;
671
672         /*
673          * The high mapping range is imprecise, so ignore the return value.
674          */
675         __change_page_attr_set_clr(&alias_cpa, 0);
676 #endif
677         return ret;
678 }
679
680 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
681 {
682         int ret, numpages = cpa->numpages;
683
684         while (numpages) {
685                 /*
686                  * Store the remaining nr of pages for the large page
687                  * preservation check.
688                  */
689                 cpa->numpages = numpages;
690
691                 ret = __change_page_attr(cpa, checkalias);
692                 if (ret)
693                         return ret;
694
695                 if (checkalias) {
696                         ret = cpa_process_alias(cpa);
697                         if (ret)
698                                 return ret;
699                 }
700
701                 /*
702                  * Adjust the number of pages with the result of the
703                  * CPA operation. Either a large page has been
704                  * preserved or a single page update happened.
705                  */
706                 BUG_ON(cpa->numpages > numpages);
707                 numpages -= cpa->numpages;
708                 cpa->vaddr += cpa->numpages * PAGE_SIZE;
709         }
710         return 0;
711 }
712
713 static inline int cache_attr(pgprot_t attr)
714 {
715         return pgprot_val(attr) &
716                 (_PAGE_PAT | _PAGE_PAT_LARGE | _PAGE_PWT | _PAGE_PCD);
717 }
718
719 static int change_page_attr_set_clr(unsigned long addr, int numpages,
720                                     pgprot_t mask_set, pgprot_t mask_clr,
721                                     int force_split)
722 {
723         struct cpa_data cpa;
724         int ret, cache, checkalias;
725
726         /*
727          * Check, if we are requested to change a not supported
728          * feature:
729          */
730         mask_set = canon_pgprot(mask_set);
731         mask_clr = canon_pgprot(mask_clr);
732         if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
733                 return 0;
734
735         /* Ensure we are PAGE_SIZE aligned */
736         if (addr & ~PAGE_MASK) {
737                 addr &= PAGE_MASK;
738                 /*
739                  * People should not be passing in unaligned addresses:
740                  */
741                 WARN_ON_ONCE(1);
742         }
743
744         cpa.vaddr = addr;
745         cpa.numpages = numpages;
746         cpa.mask_set = mask_set;
747         cpa.mask_clr = mask_clr;
748         cpa.flushtlb = 0;
749         cpa.force_split = force_split;
750
751         /* No alias checking for _NX bit modifications */
752         checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
753
754         ret = __change_page_attr_set_clr(&cpa, checkalias);
755
756         /*
757          * Check whether we really changed something:
758          */
759         if (!cpa.flushtlb)
760                 goto out;
761
762         /*
763          * No need to flush, when we did not set any of the caching
764          * attributes:
765          */
766         cache = cache_attr(mask_set);
767
768         /*
769          * On success we use clflush, when the CPU supports it to
770          * avoid the wbindv. If the CPU does not support it and in the
771          * error case we fall back to cpa_flush_all (which uses
772          * wbindv):
773          */
774         if (!ret && cpu_has_clflush)
775                 cpa_flush_range(addr, numpages, cache);
776         else
777                 cpa_flush_all(cache);
778
779 out:
780         cpa_fill_pool(NULL);
781
782         return ret;
783 }
784
785 static inline int change_page_attr_set(unsigned long addr, int numpages,
786                                        pgprot_t mask)
787 {
788         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0);
789 }
790
791 static inline int change_page_attr_clear(unsigned long addr, int numpages,
792                                          pgprot_t mask)
793 {
794         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0);
795 }
796
797 int _set_memory_uc(unsigned long addr, int numpages)
798 {
799         /*
800          * for now UC MINUS. see comments in ioremap_nocache()
801          */
802         return change_page_attr_set(addr, numpages,
803                                     __pgprot(_PAGE_CACHE_UC_MINUS));
804 }
805
806 int set_memory_uc(unsigned long addr, int numpages)
807 {
808         /*
809          * for now UC MINUS. see comments in ioremap_nocache()
810          */
811         if (reserve_memtype(addr, addr + numpages * PAGE_SIZE,
812                             _PAGE_CACHE_UC_MINUS, NULL))
813                 return -EINVAL;
814
815         return _set_memory_uc(addr, numpages);
816 }
817 EXPORT_SYMBOL(set_memory_uc);
818
819 int _set_memory_wc(unsigned long addr, int numpages)
820 {
821         return change_page_attr_set(addr, numpages,
822                                     __pgprot(_PAGE_CACHE_WC));
823 }
824
825 int set_memory_wc(unsigned long addr, int numpages)
826 {
827         if (!pat_wc_enabled)
828                 return set_memory_uc(addr, numpages);
829
830         if (reserve_memtype(addr, addr + numpages * PAGE_SIZE,
831                 _PAGE_CACHE_WC, NULL))
832                 return -EINVAL;
833
834         return _set_memory_wc(addr, numpages);
835 }
836 EXPORT_SYMBOL(set_memory_wc);
837
838 int _set_memory_wb(unsigned long addr, int numpages)
839 {
840         return change_page_attr_clear(addr, numpages,
841                                       __pgprot(_PAGE_CACHE_MASK));
842 }
843
844 int set_memory_wb(unsigned long addr, int numpages)
845 {
846         free_memtype(addr, addr + numpages * PAGE_SIZE);
847
848         return _set_memory_wb(addr, numpages);
849 }
850 EXPORT_SYMBOL(set_memory_wb);
851
852 int set_memory_x(unsigned long addr, int numpages)
853 {
854         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_NX));
855 }
856 EXPORT_SYMBOL(set_memory_x);
857
858 int set_memory_nx(unsigned long addr, int numpages)
859 {
860         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_NX));
861 }
862 EXPORT_SYMBOL(set_memory_nx);
863
864 int set_memory_ro(unsigned long addr, int numpages)
865 {
866         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_RW));
867 }
868
869 int set_memory_rw(unsigned long addr, int numpages)
870 {
871         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_RW));
872 }
873
874 int set_memory_np(unsigned long addr, int numpages)
875 {
876         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_PRESENT));
877 }
878
879 int set_memory_4k(unsigned long addr, int numpages)
880 {
881         return change_page_attr_set_clr(addr, numpages, __pgprot(0),
882                                         __pgprot(0), 1);
883 }
884
885 int set_pages_uc(struct page *page, int numpages)
886 {
887         unsigned long addr = (unsigned long)page_address(page);
888
889         return set_memory_uc(addr, numpages);
890 }
891 EXPORT_SYMBOL(set_pages_uc);
892
893 int set_pages_wb(struct page *page, int numpages)
894 {
895         unsigned long addr = (unsigned long)page_address(page);
896
897         return set_memory_wb(addr, numpages);
898 }
899 EXPORT_SYMBOL(set_pages_wb);
900
901 int set_pages_x(struct page *page, int numpages)
902 {
903         unsigned long addr = (unsigned long)page_address(page);
904
905         return set_memory_x(addr, numpages);
906 }
907 EXPORT_SYMBOL(set_pages_x);
908
909 int set_pages_nx(struct page *page, int numpages)
910 {
911         unsigned long addr = (unsigned long)page_address(page);
912
913         return set_memory_nx(addr, numpages);
914 }
915 EXPORT_SYMBOL(set_pages_nx);
916
917 int set_pages_ro(struct page *page, int numpages)
918 {
919         unsigned long addr = (unsigned long)page_address(page);
920
921         return set_memory_ro(addr, numpages);
922 }
923
924 int set_pages_rw(struct page *page, int numpages)
925 {
926         unsigned long addr = (unsigned long)page_address(page);
927
928         return set_memory_rw(addr, numpages);
929 }
930
931 #ifdef CONFIG_DEBUG_PAGEALLOC
932
933 static int __set_pages_p(struct page *page, int numpages)
934 {
935         struct cpa_data cpa = { .vaddr = (unsigned long) page_address(page),
936                                 .numpages = numpages,
937                                 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
938                                 .mask_clr = __pgprot(0)};
939
940         return __change_page_attr_set_clr(&cpa, 1);
941 }
942
943 static int __set_pages_np(struct page *page, int numpages)
944 {
945         struct cpa_data cpa = { .vaddr = (unsigned long) page_address(page),
946                                 .numpages = numpages,
947                                 .mask_set = __pgprot(0),
948                                 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW)};
949
950         return __change_page_attr_set_clr(&cpa, 1);
951 }
952
953 void kernel_map_pages(struct page *page, int numpages, int enable)
954 {
955         if (PageHighMem(page))
956                 return;
957         if (!enable) {
958                 debug_check_no_locks_freed(page_address(page),
959                                            numpages * PAGE_SIZE);
960         }
961
962         /*
963          * If page allocator is not up yet then do not call c_p_a():
964          */
965         if (!debug_pagealloc_enabled)
966                 return;
967
968         /*
969          * The return value is ignored as the calls cannot fail.
970          * Large pages are kept enabled at boot time, and are
971          * split up quickly with DEBUG_PAGEALLOC. If a splitup
972          * fails here (due to temporary memory shortage) no damage
973          * is done because we just keep the largepage intact up
974          * to the next attempt when it will likely be split up:
975          */
976         if (enable)
977                 __set_pages_p(page, numpages);
978         else
979                 __set_pages_np(page, numpages);
980
981         /*
982          * We should perform an IPI and flush all tlbs,
983          * but that can deadlock->flush only current cpu:
984          */
985         __flush_tlb_all();
986
987         /*
988          * Try to refill the page pool here. We can do this only after
989          * the tlb flush.
990          */
991         cpa_fill_pool(NULL);
992 }
993
994 #ifdef CONFIG_DEBUG_FS
995 static int dpa_show(struct seq_file *m, void *v)
996 {
997         seq_puts(m, "DEBUG_PAGEALLOC\n");
998         seq_printf(m, "pool_size     : %lu\n", pool_size);
999         seq_printf(m, "pool_pages    : %lu\n", pool_pages);
1000         seq_printf(m, "pool_low      : %lu\n", pool_low);
1001         seq_printf(m, "pool_used     : %lu\n", pool_used);
1002         seq_printf(m, "pool_failed   : %lu\n", pool_failed);
1003
1004         return 0;
1005 }
1006
1007 static int dpa_open(struct inode *inode, struct file *filp)
1008 {
1009         return single_open(filp, dpa_show, NULL);
1010 }
1011
1012 static const struct file_operations dpa_fops = {
1013         .open           = dpa_open,
1014         .read           = seq_read,
1015         .llseek         = seq_lseek,
1016         .release        = single_release,
1017 };
1018
1019 static int __init debug_pagealloc_proc_init(void)
1020 {
1021         struct dentry *de;
1022
1023         de = debugfs_create_file("debug_pagealloc", 0600, NULL, NULL,
1024                                  &dpa_fops);
1025         if (!de)
1026                 return -ENOMEM;
1027
1028         return 0;
1029 }
1030 __initcall(debug_pagealloc_proc_init);
1031 #endif
1032
1033 #ifdef CONFIG_HIBERNATION
1034
1035 bool kernel_page_present(struct page *page)
1036 {
1037         unsigned int level;
1038         pte_t *pte;
1039
1040         if (PageHighMem(page))
1041                 return false;
1042
1043         pte = lookup_address((unsigned long)page_address(page), &level);
1044         return (pte_val(*pte) & _PAGE_PRESENT);
1045 }
1046
1047 #endif /* CONFIG_HIBERNATION */
1048
1049 #endif /* CONFIG_DEBUG_PAGEALLOC */
1050
1051 #ifdef CONFIG_PROC_FS
1052 int arch_report_meminfo(char *page)
1053 {
1054         int n;
1055         n = sprintf(page, "DirectMap4k:  %8lu\n"
1056                           "DirectMap2M:  %8lu\n",
1057                         direct_pages_count[PG_LEVEL_4K],
1058                         direct_pages_count[PG_LEVEL_2M]);
1059 #ifdef CONFIG_X86_64
1060         n += sprintf(page + n, "DirectMap1G:  %8lu\n",
1061                         direct_pages_count[PG_LEVEL_1G]);
1062 #endif
1063         return n;
1064 }
1065 #endif
1066
1067 /*
1068  * The testcases use internal knowledge of the implementation that shouldn't
1069  * be exposed to the rest of the kernel. Include these directly here.
1070  */
1071 #ifdef CONFIG_CPA_DEBUG
1072 #include "pageattr-test.c"
1073 #endif