Merge branch 'topic/pcm-estrpipe-in-pm' into for-linus
[pandora-kernel.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 #include <linux/pfn.h>
15
16 #include <asm/e820.h>
17 #include <asm/processor.h>
18 #include <asm/tlbflush.h>
19 #include <asm/sections.h>
20 #include <asm/setup.h>
21 #include <asm/uaccess.h>
22 #include <asm/pgalloc.h>
23 #include <asm/proto.h>
24 #include <asm/pat.h>
25
26 /*
27  * The current flushing context - we pass it instead of 5 arguments:
28  */
29 struct cpa_data {
30         unsigned long   *vaddr;
31         pgprot_t        mask_set;
32         pgprot_t        mask_clr;
33         int             numpages;
34         int             flags;
35         unsigned long   pfn;
36         unsigned        force_split : 1;
37         int             curpage;
38         struct page     **pages;
39 };
40
41 /*
42  * Serialize cpa() (for !DEBUG_PAGEALLOC which uses large identity mappings)
43  * using cpa_lock. So that we don't allow any other cpu, with stale large tlb
44  * entries change the page attribute in parallel to some other cpu
45  * splitting a large page entry along with changing the attribute.
46  */
47 static DEFINE_SPINLOCK(cpa_lock);
48
49 #define CPA_FLUSHTLB 1
50 #define CPA_ARRAY 2
51 #define CPA_PAGES_ARRAY 4
52
53 #ifdef CONFIG_PROC_FS
54 static unsigned long direct_pages_count[PG_LEVEL_NUM];
55
56 void update_page_count(int level, unsigned long pages)
57 {
58         unsigned long flags;
59
60         /* Protect against CPA */
61         spin_lock_irqsave(&pgd_lock, flags);
62         direct_pages_count[level] += pages;
63         spin_unlock_irqrestore(&pgd_lock, flags);
64 }
65
66 static void split_page_count(int level)
67 {
68         direct_pages_count[level]--;
69         direct_pages_count[level - 1] += PTRS_PER_PTE;
70 }
71
72 void arch_report_meminfo(struct seq_file *m)
73 {
74         seq_printf(m, "DirectMap4k:    %8lu kB\n",
75                         direct_pages_count[PG_LEVEL_4K] << 2);
76 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
77         seq_printf(m, "DirectMap2M:    %8lu kB\n",
78                         direct_pages_count[PG_LEVEL_2M] << 11);
79 #else
80         seq_printf(m, "DirectMap4M:    %8lu kB\n",
81                         direct_pages_count[PG_LEVEL_2M] << 12);
82 #endif
83 #ifdef CONFIG_X86_64
84         if (direct_gbpages)
85                 seq_printf(m, "DirectMap1G:    %8lu kB\n",
86                         direct_pages_count[PG_LEVEL_1G] << 20);
87 #endif
88 }
89 #else
90 static inline void split_page_count(int level) { }
91 #endif
92
93 #ifdef CONFIG_X86_64
94
95 static inline unsigned long highmap_start_pfn(void)
96 {
97         return __pa(_text) >> PAGE_SHIFT;
98 }
99
100 static inline unsigned long highmap_end_pfn(void)
101 {
102         return __pa(roundup(_brk_end, PMD_SIZE)) >> PAGE_SHIFT;
103 }
104
105 #endif
106
107 #ifdef CONFIG_DEBUG_PAGEALLOC
108 # define debug_pagealloc 1
109 #else
110 # define debug_pagealloc 0
111 #endif
112
113 static inline int
114 within(unsigned long addr, unsigned long start, unsigned long end)
115 {
116         return addr >= start && addr < end;
117 }
118
119 /*
120  * Flushing functions
121  */
122
123 /**
124  * clflush_cache_range - flush a cache range with clflush
125  * @addr:       virtual start address
126  * @size:       number of bytes to flush
127  *
128  * clflush is an unordered instruction which needs fencing with mfence
129  * to avoid ordering issues.
130  */
131 void clflush_cache_range(void *vaddr, unsigned int size)
132 {
133         void *vend = vaddr + size - 1;
134
135         mb();
136
137         for (; vaddr < vend; vaddr += boot_cpu_data.x86_clflush_size)
138                 clflush(vaddr);
139         /*
140          * Flush any possible final partial cacheline:
141          */
142         clflush(vend);
143
144         mb();
145 }
146
147 static void __cpa_flush_all(void *arg)
148 {
149         unsigned long cache = (unsigned long)arg;
150
151         /*
152          * Flush all to work around Errata in early athlons regarding
153          * large page flushing.
154          */
155         __flush_tlb_all();
156
157         if (cache && boot_cpu_data.x86 >= 4)
158                 wbinvd();
159 }
160
161 static void cpa_flush_all(unsigned long cache)
162 {
163         BUG_ON(irqs_disabled());
164
165         on_each_cpu(__cpa_flush_all, (void *) cache, 1);
166 }
167
168 static void __cpa_flush_range(void *arg)
169 {
170         /*
171          * We could optimize that further and do individual per page
172          * tlb invalidates for a low number of pages. Caveat: we must
173          * flush the high aliases on 64bit as well.
174          */
175         __flush_tlb_all();
176 }
177
178 static void cpa_flush_range(unsigned long start, int numpages, int cache)
179 {
180         unsigned int i, level;
181         unsigned long addr;
182
183         BUG_ON(irqs_disabled());
184         WARN_ON(PAGE_ALIGN(start) != start);
185
186         on_each_cpu(__cpa_flush_range, NULL, 1);
187
188         if (!cache)
189                 return;
190
191         /*
192          * We only need to flush on one CPU,
193          * clflush is a MESI-coherent instruction that
194          * will cause all other CPUs to flush the same
195          * cachelines:
196          */
197         for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
198                 pte_t *pte = lookup_address(addr, &level);
199
200                 /*
201                  * Only flush present addresses:
202                  */
203                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
204                         clflush_cache_range((void *) addr, PAGE_SIZE);
205         }
206 }
207
208 static void cpa_flush_array(unsigned long *start, int numpages, int cache,
209                             int in_flags, struct page **pages)
210 {
211         unsigned int i, level;
212         unsigned long do_wbinvd = cache && numpages >= 1024; /* 4M threshold */
213
214         BUG_ON(irqs_disabled());
215
216         on_each_cpu(__cpa_flush_all, (void *) do_wbinvd, 1);
217
218         if (!cache || do_wbinvd)
219                 return;
220
221         /*
222          * We only need to flush on one CPU,
223          * clflush is a MESI-coherent instruction that
224          * will cause all other CPUs to flush the same
225          * cachelines:
226          */
227         for (i = 0; i < numpages; i++) {
228                 unsigned long addr;
229                 pte_t *pte;
230
231                 if (in_flags & CPA_PAGES_ARRAY)
232                         addr = (unsigned long)page_address(pages[i]);
233                 else
234                         addr = start[i];
235
236                 pte = lookup_address(addr, &level);
237
238                 /*
239                  * Only flush present addresses:
240                  */
241                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
242                         clflush_cache_range((void *)addr, PAGE_SIZE);
243         }
244 }
245
246 /*
247  * Certain areas of memory on x86 require very specific protection flags,
248  * for example the BIOS area or kernel text. Callers don't always get this
249  * right (again, ioremap() on BIOS memory is not uncommon) so this function
250  * checks and fixes these known static required protection bits.
251  */
252 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
253                                    unsigned long pfn)
254 {
255         pgprot_t forbidden = __pgprot(0);
256
257         /*
258          * The BIOS area between 640k and 1Mb needs to be executable for
259          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
260          */
261         if (within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
262                 pgprot_val(forbidden) |= _PAGE_NX;
263
264         /*
265          * The kernel text needs to be executable for obvious reasons
266          * Does not cover __inittext since that is gone later on. On
267          * 64bit we do not enforce !NX on the low mapping
268          */
269         if (within(address, (unsigned long)_text, (unsigned long)_etext))
270                 pgprot_val(forbidden) |= _PAGE_NX;
271
272         /*
273          * The .rodata section needs to be read-only. Using the pfn
274          * catches all aliases.
275          */
276         if (within(pfn, __pa((unsigned long)__start_rodata) >> PAGE_SHIFT,
277                    __pa((unsigned long)__end_rodata) >> PAGE_SHIFT))
278                 pgprot_val(forbidden) |= _PAGE_RW;
279
280         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
281
282         return prot;
283 }
284
285 /*
286  * Lookup the page table entry for a virtual address. Return a pointer
287  * to the entry and the level of the mapping.
288  *
289  * Note: We return pud and pmd either when the entry is marked large
290  * or when the present bit is not set. Otherwise we would return a
291  * pointer to a nonexisting mapping.
292  */
293 pte_t *lookup_address(unsigned long address, unsigned int *level)
294 {
295         pgd_t *pgd = pgd_offset_k(address);
296         pud_t *pud;
297         pmd_t *pmd;
298
299         *level = PG_LEVEL_NONE;
300
301         if (pgd_none(*pgd))
302                 return NULL;
303
304         pud = pud_offset(pgd, address);
305         if (pud_none(*pud))
306                 return NULL;
307
308         *level = PG_LEVEL_1G;
309         if (pud_large(*pud) || !pud_present(*pud))
310                 return (pte_t *)pud;
311
312         pmd = pmd_offset(pud, address);
313         if (pmd_none(*pmd))
314                 return NULL;
315
316         *level = PG_LEVEL_2M;
317         if (pmd_large(*pmd) || !pmd_present(*pmd))
318                 return (pte_t *)pmd;
319
320         *level = PG_LEVEL_4K;
321
322         return pte_offset_kernel(pmd, address);
323 }
324 EXPORT_SYMBOL_GPL(lookup_address);
325
326 /*
327  * Set the new pmd in all the pgds we know about:
328  */
329 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
330 {
331         /* change init_mm */
332         set_pte_atomic(kpte, pte);
333 #ifdef CONFIG_X86_32
334         if (!SHARED_KERNEL_PMD) {
335                 struct page *page;
336
337                 list_for_each_entry(page, &pgd_list, lru) {
338                         pgd_t *pgd;
339                         pud_t *pud;
340                         pmd_t *pmd;
341
342                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
343                         pud = pud_offset(pgd, address);
344                         pmd = pmd_offset(pud, address);
345                         set_pte_atomic((pte_t *)pmd, pte);
346                 }
347         }
348 #endif
349 }
350
351 static int
352 try_preserve_large_page(pte_t *kpte, unsigned long address,
353                         struct cpa_data *cpa)
354 {
355         unsigned long nextpage_addr, numpages, pmask, psize, flags, addr, pfn;
356         pte_t new_pte, old_pte, *tmp;
357         pgprot_t old_prot, new_prot;
358         int i, do_split = 1;
359         unsigned int level;
360
361         if (cpa->force_split)
362                 return 1;
363
364         spin_lock_irqsave(&pgd_lock, flags);
365         /*
366          * Check for races, another CPU might have split this page
367          * up already:
368          */
369         tmp = lookup_address(address, &level);
370         if (tmp != kpte)
371                 goto out_unlock;
372
373         switch (level) {
374         case PG_LEVEL_2M:
375                 psize = PMD_PAGE_SIZE;
376                 pmask = PMD_PAGE_MASK;
377                 break;
378 #ifdef CONFIG_X86_64
379         case PG_LEVEL_1G:
380                 psize = PUD_PAGE_SIZE;
381                 pmask = PUD_PAGE_MASK;
382                 break;
383 #endif
384         default:
385                 do_split = -EINVAL;
386                 goto out_unlock;
387         }
388
389         /*
390          * Calculate the number of pages, which fit into this large
391          * page starting at address:
392          */
393         nextpage_addr = (address + psize) & pmask;
394         numpages = (nextpage_addr - address) >> PAGE_SHIFT;
395         if (numpages < cpa->numpages)
396                 cpa->numpages = numpages;
397
398         /*
399          * We are safe now. Check whether the new pgprot is the same:
400          */
401         old_pte = *kpte;
402         old_prot = new_prot = pte_pgprot(old_pte);
403
404         pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
405         pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
406
407         /*
408          * old_pte points to the large page base address. So we need
409          * to add the offset of the virtual address:
410          */
411         pfn = pte_pfn(old_pte) + ((address & (psize - 1)) >> PAGE_SHIFT);
412         cpa->pfn = pfn;
413
414         new_prot = static_protections(new_prot, address, pfn);
415
416         /*
417          * We need to check the full range, whether
418          * static_protection() requires a different pgprot for one of
419          * the pages in the range we try to preserve:
420          */
421         addr = address + PAGE_SIZE;
422         pfn++;
423         for (i = 1; i < cpa->numpages; i++, addr += PAGE_SIZE, pfn++) {
424                 pgprot_t chk_prot = static_protections(new_prot, addr, pfn);
425
426                 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
427                         goto out_unlock;
428         }
429
430         /*
431          * If there are no changes, return. maxpages has been updated
432          * above:
433          */
434         if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
435                 do_split = 0;
436                 goto out_unlock;
437         }
438
439         /*
440          * We need to change the attributes. Check, whether we can
441          * change the large page in one go. We request a split, when
442          * the address is not aligned and the number of pages is
443          * smaller than the number of pages in the large page. Note
444          * that we limited the number of possible pages already to
445          * the number of pages in the large page.
446          */
447         if (address == (nextpage_addr - psize) && cpa->numpages == numpages) {
448                 /*
449                  * The address is aligned and the number of pages
450                  * covers the full page.
451                  */
452                 new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
453                 __set_pmd_pte(kpte, address, new_pte);
454                 cpa->flags |= CPA_FLUSHTLB;
455                 do_split = 0;
456         }
457
458 out_unlock:
459         spin_unlock_irqrestore(&pgd_lock, flags);
460
461         return do_split;
462 }
463
464 static int split_large_page(pte_t *kpte, unsigned long address)
465 {
466         unsigned long flags, pfn, pfninc = 1;
467         unsigned int i, level;
468         pte_t *pbase, *tmp;
469         pgprot_t ref_prot;
470         struct page *base;
471
472         if (!debug_pagealloc)
473                 spin_unlock(&cpa_lock);
474         base = alloc_pages(GFP_KERNEL | __GFP_NOTRACK, 0);
475         if (!debug_pagealloc)
476                 spin_lock(&cpa_lock);
477         if (!base)
478                 return -ENOMEM;
479
480         spin_lock_irqsave(&pgd_lock, flags);
481         /*
482          * Check for races, another CPU might have split this page
483          * up for us already:
484          */
485         tmp = lookup_address(address, &level);
486         if (tmp != kpte)
487                 goto out_unlock;
488
489         pbase = (pte_t *)page_address(base);
490         paravirt_alloc_pte(&init_mm, page_to_pfn(base));
491         ref_prot = pte_pgprot(pte_clrhuge(*kpte));
492         /*
493          * If we ever want to utilize the PAT bit, we need to
494          * update this function to make sure it's converted from
495          * bit 12 to bit 7 when we cross from the 2MB level to
496          * the 4K level:
497          */
498         WARN_ON_ONCE(pgprot_val(ref_prot) & _PAGE_PAT_LARGE);
499
500 #ifdef CONFIG_X86_64
501         if (level == PG_LEVEL_1G) {
502                 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
503                 pgprot_val(ref_prot) |= _PAGE_PSE;
504         }
505 #endif
506
507         /*
508          * Get the target pfn from the original entry:
509          */
510         pfn = pte_pfn(*kpte);
511         for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
512                 set_pte(&pbase[i], pfn_pte(pfn, ref_prot));
513
514         if (address >= (unsigned long)__va(0) &&
515                 address < (unsigned long)__va(max_low_pfn_mapped << PAGE_SHIFT))
516                 split_page_count(level);
517
518 #ifdef CONFIG_X86_64
519         if (address >= (unsigned long)__va(1UL<<32) &&
520                 address < (unsigned long)__va(max_pfn_mapped << PAGE_SHIFT))
521                 split_page_count(level);
522 #endif
523
524         /*
525          * Install the new, split up pagetable.
526          *
527          * We use the standard kernel pagetable protections for the new
528          * pagetable protections, the actual ptes set above control the
529          * primary protection behavior:
530          */
531         __set_pmd_pte(kpte, address, mk_pte(base, __pgprot(_KERNPG_TABLE)));
532
533         /*
534          * Intel Atom errata AAH41 workaround.
535          *
536          * The real fix should be in hw or in a microcode update, but
537          * we also probabilistically try to reduce the window of having
538          * a large TLB mixed with 4K TLBs while instruction fetches are
539          * going on.
540          */
541         __flush_tlb_all();
542
543         base = NULL;
544
545 out_unlock:
546         /*
547          * If we dropped out via the lookup_address check under
548          * pgd_lock then stick the page back into the pool:
549          */
550         if (base)
551                 __free_page(base);
552         spin_unlock_irqrestore(&pgd_lock, flags);
553
554         return 0;
555 }
556
557 static int __cpa_process_fault(struct cpa_data *cpa, unsigned long vaddr,
558                                int primary)
559 {
560         /*
561          * Ignore all non primary paths.
562          */
563         if (!primary)
564                 return 0;
565
566         /*
567          * Ignore the NULL PTE for kernel identity mapping, as it is expected
568          * to have holes.
569          * Also set numpages to '1' indicating that we processed cpa req for
570          * one virtual address page and its pfn. TBD: numpages can be set based
571          * on the initial value and the level returned by lookup_address().
572          */
573         if (within(vaddr, PAGE_OFFSET,
574                    PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
575                 cpa->numpages = 1;
576                 cpa->pfn = __pa(vaddr) >> PAGE_SHIFT;
577                 return 0;
578         } else {
579                 WARN(1, KERN_WARNING "CPA: called for zero pte. "
580                         "vaddr = %lx cpa->vaddr = %lx\n", vaddr,
581                         *cpa->vaddr);
582
583                 return -EFAULT;
584         }
585 }
586
587 static int __change_page_attr(struct cpa_data *cpa, int primary)
588 {
589         unsigned long address;
590         int do_split, err;
591         unsigned int level;
592         pte_t *kpte, old_pte;
593
594         if (cpa->flags & CPA_PAGES_ARRAY) {
595                 struct page *page = cpa->pages[cpa->curpage];
596                 if (unlikely(PageHighMem(page)))
597                         return 0;
598                 address = (unsigned long)page_address(page);
599         } else if (cpa->flags & CPA_ARRAY)
600                 address = cpa->vaddr[cpa->curpage];
601         else
602                 address = *cpa->vaddr;
603 repeat:
604         kpte = lookup_address(address, &level);
605         if (!kpte)
606                 return __cpa_process_fault(cpa, address, primary);
607
608         old_pte = *kpte;
609         if (!pte_val(old_pte))
610                 return __cpa_process_fault(cpa, address, primary);
611
612         if (level == PG_LEVEL_4K) {
613                 pte_t new_pte;
614                 pgprot_t new_prot = pte_pgprot(old_pte);
615                 unsigned long pfn = pte_pfn(old_pte);
616
617                 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
618                 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
619
620                 new_prot = static_protections(new_prot, address, pfn);
621
622                 /*
623                  * We need to keep the pfn from the existing PTE,
624                  * after all we're only going to change it's attributes
625                  * not the memory it points to
626                  */
627                 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
628                 cpa->pfn = pfn;
629                 /*
630                  * Do we really change anything ?
631                  */
632                 if (pte_val(old_pte) != pte_val(new_pte)) {
633                         set_pte_atomic(kpte, new_pte);
634                         cpa->flags |= CPA_FLUSHTLB;
635                 }
636                 cpa->numpages = 1;
637                 return 0;
638         }
639
640         /*
641          * Check, whether we can keep the large page intact
642          * and just change the pte:
643          */
644         do_split = try_preserve_large_page(kpte, address, cpa);
645         /*
646          * When the range fits into the existing large page,
647          * return. cp->numpages and cpa->tlbflush have been updated in
648          * try_large_page:
649          */
650         if (do_split <= 0)
651                 return do_split;
652
653         /*
654          * We have to split the large page:
655          */
656         err = split_large_page(kpte, address);
657         if (!err) {
658                 /*
659                  * Do a global flush tlb after splitting the large page
660                  * and before we do the actual change page attribute in the PTE.
661                  *
662                  * With out this, we violate the TLB application note, that says
663                  * "The TLBs may contain both ordinary and large-page
664                  *  translations for a 4-KByte range of linear addresses. This
665                  *  may occur if software modifies the paging structures so that
666                  *  the page size used for the address range changes. If the two
667                  *  translations differ with respect to page frame or attributes
668                  *  (e.g., permissions), processor behavior is undefined and may
669                  *  be implementation-specific."
670                  *
671                  * We do this global tlb flush inside the cpa_lock, so that we
672                  * don't allow any other cpu, with stale tlb entries change the
673                  * page attribute in parallel, that also falls into the
674                  * just split large page entry.
675                  */
676                 flush_tlb_all();
677                 goto repeat;
678         }
679
680         return err;
681 }
682
683 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
684
685 static int cpa_process_alias(struct cpa_data *cpa)
686 {
687         struct cpa_data alias_cpa;
688         unsigned long laddr = (unsigned long)__va(cpa->pfn << PAGE_SHIFT);
689         unsigned long vaddr, remapped;
690         int ret;
691
692         if (cpa->pfn >= max_pfn_mapped)
693                 return 0;
694
695 #ifdef CONFIG_X86_64
696         if (cpa->pfn >= max_low_pfn_mapped && cpa->pfn < (1UL<<(32-PAGE_SHIFT)))
697                 return 0;
698 #endif
699         /*
700          * No need to redo, when the primary call touched the direct
701          * mapping already:
702          */
703         if (cpa->flags & CPA_PAGES_ARRAY) {
704                 struct page *page = cpa->pages[cpa->curpage];
705                 if (unlikely(PageHighMem(page)))
706                         return 0;
707                 vaddr = (unsigned long)page_address(page);
708         } else if (cpa->flags & CPA_ARRAY)
709                 vaddr = cpa->vaddr[cpa->curpage];
710         else
711                 vaddr = *cpa->vaddr;
712
713         if (!(within(vaddr, PAGE_OFFSET,
714                     PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT)))) {
715
716                 alias_cpa = *cpa;
717                 alias_cpa.vaddr = &laddr;
718                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
719
720                 ret = __change_page_attr_set_clr(&alias_cpa, 0);
721                 if (ret)
722                         return ret;
723         }
724
725 #ifdef CONFIG_X86_64
726         /*
727          * If the primary call didn't touch the high mapping already
728          * and the physical address is inside the kernel map, we need
729          * to touch the high mapped kernel as well:
730          */
731         if (!within(vaddr, (unsigned long)_text, _brk_end) &&
732             within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn())) {
733                 unsigned long temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) +
734                                                __START_KERNEL_map - phys_base;
735                 alias_cpa = *cpa;
736                 alias_cpa.vaddr = &temp_cpa_vaddr;
737                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
738
739                 /*
740                  * The high mapping range is imprecise, so ignore the
741                  * return value.
742                  */
743                 __change_page_attr_set_clr(&alias_cpa, 0);
744         }
745 #endif
746
747         /*
748          * If the PMD page was partially used for per-cpu remapping,
749          * the recycled area needs to be split and modified.  Because
750          * the area is always proper subset of a PMD page
751          * cpa->numpages is guaranteed to be 1 for these areas, so
752          * there's no need to loop over and check for further remaps.
753          */
754         remapped = (unsigned long)pcpu_lpage_remapped((void *)laddr);
755         if (remapped) {
756                 WARN_ON(cpa->numpages > 1);
757                 alias_cpa = *cpa;
758                 alias_cpa.vaddr = &remapped;
759                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
760                 ret = __change_page_attr_set_clr(&alias_cpa, 0);
761                 if (ret)
762                         return ret;
763         }
764
765         return 0;
766 }
767
768 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
769 {
770         int ret, numpages = cpa->numpages;
771
772         while (numpages) {
773                 /*
774                  * Store the remaining nr of pages for the large page
775                  * preservation check.
776                  */
777                 cpa->numpages = numpages;
778                 /* for array changes, we can't use large page */
779                 if (cpa->flags & (CPA_ARRAY | CPA_PAGES_ARRAY))
780                         cpa->numpages = 1;
781
782                 if (!debug_pagealloc)
783                         spin_lock(&cpa_lock);
784                 ret = __change_page_attr(cpa, checkalias);
785                 if (!debug_pagealloc)
786                         spin_unlock(&cpa_lock);
787                 if (ret)
788                         return ret;
789
790                 if (checkalias) {
791                         ret = cpa_process_alias(cpa);
792                         if (ret)
793                                 return ret;
794                 }
795
796                 /*
797                  * Adjust the number of pages with the result of the
798                  * CPA operation. Either a large page has been
799                  * preserved or a single page update happened.
800                  */
801                 BUG_ON(cpa->numpages > numpages);
802                 numpages -= cpa->numpages;
803                 if (cpa->flags & (CPA_PAGES_ARRAY | CPA_ARRAY))
804                         cpa->curpage++;
805                 else
806                         *cpa->vaddr += cpa->numpages * PAGE_SIZE;
807
808         }
809         return 0;
810 }
811
812 static inline int cache_attr(pgprot_t attr)
813 {
814         return pgprot_val(attr) &
815                 (_PAGE_PAT | _PAGE_PAT_LARGE | _PAGE_PWT | _PAGE_PCD);
816 }
817
818 static int change_page_attr_set_clr(unsigned long *addr, int numpages,
819                                     pgprot_t mask_set, pgprot_t mask_clr,
820                                     int force_split, int in_flag,
821                                     struct page **pages)
822 {
823         struct cpa_data cpa;
824         int ret, cache, checkalias;
825
826         /*
827          * Check, if we are requested to change a not supported
828          * feature:
829          */
830         mask_set = canon_pgprot(mask_set);
831         mask_clr = canon_pgprot(mask_clr);
832         if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
833                 return 0;
834
835         /* Ensure we are PAGE_SIZE aligned */
836         if (in_flag & CPA_ARRAY) {
837                 int i;
838                 for (i = 0; i < numpages; i++) {
839                         if (addr[i] & ~PAGE_MASK) {
840                                 addr[i] &= PAGE_MASK;
841                                 WARN_ON_ONCE(1);
842                         }
843                 }
844         } else if (!(in_flag & CPA_PAGES_ARRAY)) {
845                 /*
846                  * in_flag of CPA_PAGES_ARRAY implies it is aligned.
847                  * No need to cehck in that case
848                  */
849                 if (*addr & ~PAGE_MASK) {
850                         *addr &= PAGE_MASK;
851                         /*
852                          * People should not be passing in unaligned addresses:
853                          */
854                         WARN_ON_ONCE(1);
855                 }
856         }
857
858         /* Must avoid aliasing mappings in the highmem code */
859         kmap_flush_unused();
860
861         vm_unmap_aliases();
862
863         cpa.vaddr = addr;
864         cpa.pages = pages;
865         cpa.numpages = numpages;
866         cpa.mask_set = mask_set;
867         cpa.mask_clr = mask_clr;
868         cpa.flags = 0;
869         cpa.curpage = 0;
870         cpa.force_split = force_split;
871
872         if (in_flag & (CPA_ARRAY | CPA_PAGES_ARRAY))
873                 cpa.flags |= in_flag;
874
875         /* No alias checking for _NX bit modifications */
876         checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
877
878         ret = __change_page_attr_set_clr(&cpa, checkalias);
879
880         /*
881          * Check whether we really changed something:
882          */
883         if (!(cpa.flags & CPA_FLUSHTLB))
884                 goto out;
885
886         /*
887          * No need to flush, when we did not set any of the caching
888          * attributes:
889          */
890         cache = cache_attr(mask_set);
891
892         /*
893          * On success we use clflush, when the CPU supports it to
894          * avoid the wbindv. If the CPU does not support it and in the
895          * error case we fall back to cpa_flush_all (which uses
896          * wbindv):
897          */
898         if (!ret && cpu_has_clflush) {
899                 if (cpa.flags & (CPA_PAGES_ARRAY | CPA_ARRAY)) {
900                         cpa_flush_array(addr, numpages, cache,
901                                         cpa.flags, pages);
902                 } else
903                         cpa_flush_range(*addr, numpages, cache);
904         } else
905                 cpa_flush_all(cache);
906
907 out:
908         return ret;
909 }
910
911 static inline int change_page_attr_set(unsigned long *addr, int numpages,
912                                        pgprot_t mask, int array)
913 {
914         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
915                 (array ? CPA_ARRAY : 0), NULL);
916 }
917
918 static inline int change_page_attr_clear(unsigned long *addr, int numpages,
919                                          pgprot_t mask, int array)
920 {
921         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
922                 (array ? CPA_ARRAY : 0), NULL);
923 }
924
925 static inline int cpa_set_pages_array(struct page **pages, int numpages,
926                                        pgprot_t mask)
927 {
928         return change_page_attr_set_clr(NULL, numpages, mask, __pgprot(0), 0,
929                 CPA_PAGES_ARRAY, pages);
930 }
931
932 static inline int cpa_clear_pages_array(struct page **pages, int numpages,
933                                          pgprot_t mask)
934 {
935         return change_page_attr_set_clr(NULL, numpages, __pgprot(0), mask, 0,
936                 CPA_PAGES_ARRAY, pages);
937 }
938
939 int _set_memory_uc(unsigned long addr, int numpages)
940 {
941         /*
942          * for now UC MINUS. see comments in ioremap_nocache()
943          */
944         return change_page_attr_set(&addr, numpages,
945                                     __pgprot(_PAGE_CACHE_UC_MINUS), 0);
946 }
947
948 int set_memory_uc(unsigned long addr, int numpages)
949 {
950         int ret;
951
952         /*
953          * for now UC MINUS. see comments in ioremap_nocache()
954          */
955         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
956                             _PAGE_CACHE_UC_MINUS, NULL);
957         if (ret)
958                 goto out_err;
959
960         ret = _set_memory_uc(addr, numpages);
961         if (ret)
962                 goto out_free;
963
964         return 0;
965
966 out_free:
967         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
968 out_err:
969         return ret;
970 }
971 EXPORT_SYMBOL(set_memory_uc);
972
973 int set_memory_array_uc(unsigned long *addr, int addrinarray)
974 {
975         int i, j;
976         int ret;
977
978         /*
979          * for now UC MINUS. see comments in ioremap_nocache()
980          */
981         for (i = 0; i < addrinarray; i++) {
982                 ret = reserve_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE,
983                                         _PAGE_CACHE_UC_MINUS, NULL);
984                 if (ret)
985                         goto out_free;
986         }
987
988         ret = change_page_attr_set(addr, addrinarray,
989                                     __pgprot(_PAGE_CACHE_UC_MINUS), 1);
990         if (ret)
991                 goto out_free;
992
993         return 0;
994
995 out_free:
996         for (j = 0; j < i; j++)
997                 free_memtype(__pa(addr[j]), __pa(addr[j]) + PAGE_SIZE);
998
999         return ret;
1000 }
1001 EXPORT_SYMBOL(set_memory_array_uc);
1002
1003 int _set_memory_wc(unsigned long addr, int numpages)
1004 {
1005         int ret;
1006         unsigned long addr_copy = addr;
1007
1008         ret = change_page_attr_set(&addr, numpages,
1009                                     __pgprot(_PAGE_CACHE_UC_MINUS), 0);
1010         if (!ret) {
1011                 ret = change_page_attr_set_clr(&addr_copy, numpages,
1012                                                __pgprot(_PAGE_CACHE_WC),
1013                                                __pgprot(_PAGE_CACHE_MASK),
1014                                                0, 0, NULL);
1015         }
1016         return ret;
1017 }
1018
1019 int set_memory_wc(unsigned long addr, int numpages)
1020 {
1021         int ret;
1022
1023         if (!pat_enabled)
1024                 return set_memory_uc(addr, numpages);
1025
1026         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1027                 _PAGE_CACHE_WC, NULL);
1028         if (ret)
1029                 goto out_err;
1030
1031         ret = _set_memory_wc(addr, numpages);
1032         if (ret)
1033                 goto out_free;
1034
1035         return 0;
1036
1037 out_free:
1038         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1039 out_err:
1040         return ret;
1041 }
1042 EXPORT_SYMBOL(set_memory_wc);
1043
1044 int _set_memory_wb(unsigned long addr, int numpages)
1045 {
1046         return change_page_attr_clear(&addr, numpages,
1047                                       __pgprot(_PAGE_CACHE_MASK), 0);
1048 }
1049
1050 int set_memory_wb(unsigned long addr, int numpages)
1051 {
1052         int ret;
1053
1054         ret = _set_memory_wb(addr, numpages);
1055         if (ret)
1056                 return ret;
1057
1058         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1059         return 0;
1060 }
1061 EXPORT_SYMBOL(set_memory_wb);
1062
1063 int set_memory_array_wb(unsigned long *addr, int addrinarray)
1064 {
1065         int i;
1066         int ret;
1067
1068         ret = change_page_attr_clear(addr, addrinarray,
1069                                       __pgprot(_PAGE_CACHE_MASK), 1);
1070         if (ret)
1071                 return ret;
1072
1073         for (i = 0; i < addrinarray; i++)
1074                 free_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE);
1075
1076         return 0;
1077 }
1078 EXPORT_SYMBOL(set_memory_array_wb);
1079
1080 int set_memory_x(unsigned long addr, int numpages)
1081 {
1082         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1083 }
1084 EXPORT_SYMBOL(set_memory_x);
1085
1086 int set_memory_nx(unsigned long addr, int numpages)
1087 {
1088         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1089 }
1090 EXPORT_SYMBOL(set_memory_nx);
1091
1092 int set_memory_ro(unsigned long addr, int numpages)
1093 {
1094         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1095 }
1096 EXPORT_SYMBOL_GPL(set_memory_ro);
1097
1098 int set_memory_rw(unsigned long addr, int numpages)
1099 {
1100         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
1101 }
1102 EXPORT_SYMBOL_GPL(set_memory_rw);
1103
1104 int set_memory_np(unsigned long addr, int numpages)
1105 {
1106         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
1107 }
1108
1109 int set_memory_4k(unsigned long addr, int numpages)
1110 {
1111         return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1112                                         __pgprot(0), 1, 0, NULL);
1113 }
1114
1115 int set_pages_uc(struct page *page, int numpages)
1116 {
1117         unsigned long addr = (unsigned long)page_address(page);
1118
1119         return set_memory_uc(addr, numpages);
1120 }
1121 EXPORT_SYMBOL(set_pages_uc);
1122
1123 int set_pages_array_uc(struct page **pages, int addrinarray)
1124 {
1125         unsigned long start;
1126         unsigned long end;
1127         int i;
1128         int free_idx;
1129
1130         for (i = 0; i < addrinarray; i++) {
1131                 if (PageHighMem(pages[i]))
1132                         continue;
1133                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1134                 end = start + PAGE_SIZE;
1135                 if (reserve_memtype(start, end, _PAGE_CACHE_UC_MINUS, NULL))
1136                         goto err_out;
1137         }
1138
1139         if (cpa_set_pages_array(pages, addrinarray,
1140                         __pgprot(_PAGE_CACHE_UC_MINUS)) == 0) {
1141                 return 0; /* Success */
1142         }
1143 err_out:
1144         free_idx = i;
1145         for (i = 0; i < free_idx; i++) {
1146                 if (PageHighMem(pages[i]))
1147                         continue;
1148                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1149                 end = start + PAGE_SIZE;
1150                 free_memtype(start, end);
1151         }
1152         return -EINVAL;
1153 }
1154 EXPORT_SYMBOL(set_pages_array_uc);
1155
1156 int set_pages_wb(struct page *page, int numpages)
1157 {
1158         unsigned long addr = (unsigned long)page_address(page);
1159
1160         return set_memory_wb(addr, numpages);
1161 }
1162 EXPORT_SYMBOL(set_pages_wb);
1163
1164 int set_pages_array_wb(struct page **pages, int addrinarray)
1165 {
1166         int retval;
1167         unsigned long start;
1168         unsigned long end;
1169         int i;
1170
1171         retval = cpa_clear_pages_array(pages, addrinarray,
1172                         __pgprot(_PAGE_CACHE_MASK));
1173         if (retval)
1174                 return retval;
1175
1176         for (i = 0; i < addrinarray; i++) {
1177                 if (PageHighMem(pages[i]))
1178                         continue;
1179                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1180                 end = start + PAGE_SIZE;
1181                 free_memtype(start, end);
1182         }
1183
1184         return 0;
1185 }
1186 EXPORT_SYMBOL(set_pages_array_wb);
1187
1188 int set_pages_x(struct page *page, int numpages)
1189 {
1190         unsigned long addr = (unsigned long)page_address(page);
1191
1192         return set_memory_x(addr, numpages);
1193 }
1194 EXPORT_SYMBOL(set_pages_x);
1195
1196 int set_pages_nx(struct page *page, int numpages)
1197 {
1198         unsigned long addr = (unsigned long)page_address(page);
1199
1200         return set_memory_nx(addr, numpages);
1201 }
1202 EXPORT_SYMBOL(set_pages_nx);
1203
1204 int set_pages_ro(struct page *page, int numpages)
1205 {
1206         unsigned long addr = (unsigned long)page_address(page);
1207
1208         return set_memory_ro(addr, numpages);
1209 }
1210
1211 int set_pages_rw(struct page *page, int numpages)
1212 {
1213         unsigned long addr = (unsigned long)page_address(page);
1214
1215         return set_memory_rw(addr, numpages);
1216 }
1217
1218 #ifdef CONFIG_DEBUG_PAGEALLOC
1219
1220 static int __set_pages_p(struct page *page, int numpages)
1221 {
1222         unsigned long tempaddr = (unsigned long) page_address(page);
1223         struct cpa_data cpa = { .vaddr = &tempaddr,
1224                                 .numpages = numpages,
1225                                 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1226                                 .mask_clr = __pgprot(0),
1227                                 .flags = 0};
1228
1229         /*
1230          * No alias checking needed for setting present flag. otherwise,
1231          * we may need to break large pages for 64-bit kernel text
1232          * mappings (this adds to complexity if we want to do this from
1233          * atomic context especially). Let's keep it simple!
1234          */
1235         return __change_page_attr_set_clr(&cpa, 0);
1236 }
1237
1238 static int __set_pages_np(struct page *page, int numpages)
1239 {
1240         unsigned long tempaddr = (unsigned long) page_address(page);
1241         struct cpa_data cpa = { .vaddr = &tempaddr,
1242                                 .numpages = numpages,
1243                                 .mask_set = __pgprot(0),
1244                                 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1245                                 .flags = 0};
1246
1247         /*
1248          * No alias checking needed for setting not present flag. otherwise,
1249          * we may need to break large pages for 64-bit kernel text
1250          * mappings (this adds to complexity if we want to do this from
1251          * atomic context especially). Let's keep it simple!
1252          */
1253         return __change_page_attr_set_clr(&cpa, 0);
1254 }
1255
1256 void kernel_map_pages(struct page *page, int numpages, int enable)
1257 {
1258         if (PageHighMem(page))
1259                 return;
1260         if (!enable) {
1261                 debug_check_no_locks_freed(page_address(page),
1262                                            numpages * PAGE_SIZE);
1263         }
1264
1265         /*
1266          * If page allocator is not up yet then do not call c_p_a():
1267          */
1268         if (!debug_pagealloc_enabled)
1269                 return;
1270
1271         /*
1272          * The return value is ignored as the calls cannot fail.
1273          * Large pages for identity mappings are not used at boot time
1274          * and hence no memory allocations during large page split.
1275          */
1276         if (enable)
1277                 __set_pages_p(page, numpages);
1278         else
1279                 __set_pages_np(page, numpages);
1280
1281         /*
1282          * We should perform an IPI and flush all tlbs,
1283          * but that can deadlock->flush only current cpu:
1284          */
1285         __flush_tlb_all();
1286 }
1287
1288 #ifdef CONFIG_HIBERNATION
1289
1290 bool kernel_page_present(struct page *page)
1291 {
1292         unsigned int level;
1293         pte_t *pte;
1294
1295         if (PageHighMem(page))
1296                 return false;
1297
1298         pte = lookup_address((unsigned long)page_address(page), &level);
1299         return (pte_val(*pte) & _PAGE_PRESENT);
1300 }
1301
1302 #endif /* CONFIG_HIBERNATION */
1303
1304 #endif /* CONFIG_DEBUG_PAGEALLOC */
1305
1306 /*
1307  * The testcases use internal knowledge of the implementation that shouldn't
1308  * be exposed to the rest of the kernel. Include these directly here.
1309  */
1310 #ifdef CONFIG_CPA_DEBUG
1311 #include "pageattr-test.c"
1312 #endif