Merge branch 'fix/asoc' of git://git.kernel.org/pub/scm/linux/kernel/git/tiwai/sound-2.6
[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/mm.h>
10 #include <linux/interrupt.h>
11 #include <linux/seq_file.h>
12 #include <linux/debugfs.h>
13 #include <linux/pfn.h>
14 #include <linux/percpu.h>
15 #include <linux/gfp.h>
16 #include <linux/pci.h>
17
18 #include <asm/e820.h>
19 #include <asm/processor.h>
20 #include <asm/tlbflush.h>
21 #include <asm/sections.h>
22 #include <asm/setup.h>
23 #include <asm/uaccess.h>
24 #include <asm/pgalloc.h>
25 #include <asm/proto.h>
26 #include <asm/pat.h>
27
28 /*
29  * The current flushing context - we pass it instead of 5 arguments:
30  */
31 struct cpa_data {
32         unsigned long   *vaddr;
33         pgprot_t        mask_set;
34         pgprot_t        mask_clr;
35         int             numpages;
36         int             flags;
37         unsigned long   pfn;
38         unsigned        force_split : 1;
39         int             curpage;
40         struct page     **pages;
41 };
42
43 /*
44  * Serialize cpa() (for !DEBUG_PAGEALLOC which uses large identity mappings)
45  * using cpa_lock. So that we don't allow any other cpu, with stale large tlb
46  * entries change the page attribute in parallel to some other cpu
47  * splitting a large page entry along with changing the attribute.
48  */
49 static DEFINE_SPINLOCK(cpa_lock);
50
51 #define CPA_FLUSHTLB 1
52 #define CPA_ARRAY 2
53 #define CPA_PAGES_ARRAY 4
54
55 #ifdef CONFIG_PROC_FS
56 static unsigned long direct_pages_count[PG_LEVEL_NUM];
57
58 void update_page_count(int level, unsigned long pages)
59 {
60         unsigned long flags;
61
62         /* Protect against CPA */
63         spin_lock_irqsave(&pgd_lock, flags);
64         direct_pages_count[level] += pages;
65         spin_unlock_irqrestore(&pgd_lock, flags);
66 }
67
68 static void split_page_count(int level)
69 {
70         direct_pages_count[level]--;
71         direct_pages_count[level - 1] += PTRS_PER_PTE;
72 }
73
74 void arch_report_meminfo(struct seq_file *m)
75 {
76         seq_printf(m, "DirectMap4k:    %8lu kB\n",
77                         direct_pages_count[PG_LEVEL_4K] << 2);
78 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
79         seq_printf(m, "DirectMap2M:    %8lu kB\n",
80                         direct_pages_count[PG_LEVEL_2M] << 11);
81 #else
82         seq_printf(m, "DirectMap4M:    %8lu kB\n",
83                         direct_pages_count[PG_LEVEL_2M] << 12);
84 #endif
85 #ifdef CONFIG_X86_64
86         if (direct_gbpages)
87                 seq_printf(m, "DirectMap1G:    %8lu kB\n",
88                         direct_pages_count[PG_LEVEL_1G] << 20);
89 #endif
90 }
91 #else
92 static inline void split_page_count(int level) { }
93 #endif
94
95 #ifdef CONFIG_X86_64
96
97 static inline unsigned long highmap_start_pfn(void)
98 {
99         return __pa(_text) >> PAGE_SHIFT;
100 }
101
102 static inline unsigned long highmap_end_pfn(void)
103 {
104         return __pa(roundup(_brk_end, PMD_SIZE)) >> PAGE_SHIFT;
105 }
106
107 #endif
108
109 #ifdef CONFIG_DEBUG_PAGEALLOC
110 # define debug_pagealloc 1
111 #else
112 # define debug_pagealloc 0
113 #endif
114
115 static inline int
116 within(unsigned long addr, unsigned long start, unsigned long end)
117 {
118         return addr >= start && addr < end;
119 }
120
121 /*
122  * Flushing functions
123  */
124
125 /**
126  * clflush_cache_range - flush a cache range with clflush
127  * @addr:       virtual start address
128  * @size:       number of bytes to flush
129  *
130  * clflush is an unordered instruction which needs fencing with mfence
131  * to avoid ordering issues.
132  */
133 void clflush_cache_range(void *vaddr, unsigned int size)
134 {
135         void *vend = vaddr + size - 1;
136
137         mb();
138
139         for (; vaddr < vend; vaddr += boot_cpu_data.x86_clflush_size)
140                 clflush(vaddr);
141         /*
142          * Flush any possible final partial cacheline:
143          */
144         clflush(vend);
145
146         mb();
147 }
148 EXPORT_SYMBOL_GPL(clflush_cache_range);
149
150 static void __cpa_flush_all(void *arg)
151 {
152         unsigned long cache = (unsigned long)arg;
153
154         /*
155          * Flush all to work around Errata in early athlons regarding
156          * large page flushing.
157          */
158         __flush_tlb_all();
159
160         if (cache && boot_cpu_data.x86 >= 4)
161                 wbinvd();
162 }
163
164 static void cpa_flush_all(unsigned long cache)
165 {
166         BUG_ON(irqs_disabled());
167
168         on_each_cpu(__cpa_flush_all, (void *) cache, 1);
169 }
170
171 static void __cpa_flush_range(void *arg)
172 {
173         /*
174          * We could optimize that further and do individual per page
175          * tlb invalidates for a low number of pages. Caveat: we must
176          * flush the high aliases on 64bit as well.
177          */
178         __flush_tlb_all();
179 }
180
181 static void cpa_flush_range(unsigned long start, int numpages, int cache)
182 {
183         unsigned int i, level;
184         unsigned long addr;
185
186         BUG_ON(irqs_disabled());
187         WARN_ON(PAGE_ALIGN(start) != start);
188
189         on_each_cpu(__cpa_flush_range, NULL, 1);
190
191         if (!cache)
192                 return;
193
194         /*
195          * We only need to flush on one CPU,
196          * clflush is a MESI-coherent instruction that
197          * will cause all other CPUs to flush the same
198          * cachelines:
199          */
200         for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
201                 pte_t *pte = lookup_address(addr, &level);
202
203                 /*
204                  * Only flush present addresses:
205                  */
206                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
207                         clflush_cache_range((void *) addr, PAGE_SIZE);
208         }
209 }
210
211 static void cpa_flush_array(unsigned long *start, int numpages, int cache,
212                             int in_flags, struct page **pages)
213 {
214         unsigned int i, level;
215         unsigned long do_wbinvd = cache && numpages >= 1024; /* 4M threshold */
216
217         BUG_ON(irqs_disabled());
218
219         on_each_cpu(__cpa_flush_all, (void *) do_wbinvd, 1);
220
221         if (!cache || do_wbinvd)
222                 return;
223
224         /*
225          * We only need to flush on one CPU,
226          * clflush is a MESI-coherent instruction that
227          * will cause all other CPUs to flush the same
228          * cachelines:
229          */
230         for (i = 0; i < numpages; i++) {
231                 unsigned long addr;
232                 pte_t *pte;
233
234                 if (in_flags & CPA_PAGES_ARRAY)
235                         addr = (unsigned long)page_address(pages[i]);
236                 else
237                         addr = start[i];
238
239                 pte = lookup_address(addr, &level);
240
241                 /*
242                  * Only flush present addresses:
243                  */
244                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
245                         clflush_cache_range((void *)addr, PAGE_SIZE);
246         }
247 }
248
249 /*
250  * Certain areas of memory on x86 require very specific protection flags,
251  * for example the BIOS area or kernel text. Callers don't always get this
252  * right (again, ioremap() on BIOS memory is not uncommon) so this function
253  * checks and fixes these known static required protection bits.
254  */
255 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
256                                    unsigned long pfn)
257 {
258         pgprot_t forbidden = __pgprot(0);
259
260         /*
261          * The BIOS area between 640k and 1Mb needs to be executable for
262          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
263          */
264 #ifdef CONFIG_PCI_BIOS
265         if (pcibios_enabled && within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
266                 pgprot_val(forbidden) |= _PAGE_NX;
267 #endif
268
269         /*
270          * The kernel text needs to be executable for obvious reasons
271          * Does not cover __inittext since that is gone later on. On
272          * 64bit we do not enforce !NX on the low mapping
273          */
274         if (within(address, (unsigned long)_text, (unsigned long)_etext))
275                 pgprot_val(forbidden) |= _PAGE_NX;
276
277         /*
278          * The .rodata section needs to be read-only. Using the pfn
279          * catches all aliases.
280          */
281         if (within(pfn, __pa((unsigned long)__start_rodata) >> PAGE_SHIFT,
282                    __pa((unsigned long)__end_rodata) >> PAGE_SHIFT))
283                 pgprot_val(forbidden) |= _PAGE_RW;
284
285 #if defined(CONFIG_X86_64) && defined(CONFIG_DEBUG_RODATA)
286         /*
287          * Once the kernel maps the text as RO (kernel_set_to_readonly is set),
288          * kernel text mappings for the large page aligned text, rodata sections
289          * will be always read-only. For the kernel identity mappings covering
290          * the holes caused by this alignment can be anything that user asks.
291          *
292          * This will preserve the large page mappings for kernel text/data
293          * at no extra cost.
294          */
295         if (kernel_set_to_readonly &&
296             within(address, (unsigned long)_text,
297                    (unsigned long)__end_rodata_hpage_align)) {
298                 unsigned int level;
299
300                 /*
301                  * Don't enforce the !RW mapping for the kernel text mapping,
302                  * if the current mapping is already using small page mapping.
303                  * No need to work hard to preserve large page mappings in this
304                  * case.
305                  *
306                  * This also fixes the Linux Xen paravirt guest boot failure
307                  * (because of unexpected read-only mappings for kernel identity
308                  * mappings). In this paravirt guest case, the kernel text
309                  * mapping and the kernel identity mapping share the same
310                  * page-table pages. Thus we can't really use different
311                  * protections for the kernel text and identity mappings. Also,
312                  * these shared mappings are made of small page mappings.
313                  * Thus this don't enforce !RW mapping for small page kernel
314                  * text mapping logic will help Linux Xen parvirt guest boot
315                  * aswell.
316                  */
317                 if (lookup_address(address, &level) && (level != PG_LEVEL_4K))
318                         pgprot_val(forbidden) |= _PAGE_RW;
319         }
320 #endif
321
322         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
323
324         return prot;
325 }
326
327 /*
328  * Lookup the page table entry for a virtual address. Return a pointer
329  * to the entry and the level of the mapping.
330  *
331  * Note: We return pud and pmd either when the entry is marked large
332  * or when the present bit is not set. Otherwise we would return a
333  * pointer to a nonexisting mapping.
334  */
335 pte_t *lookup_address(unsigned long address, unsigned int *level)
336 {
337         pgd_t *pgd = pgd_offset_k(address);
338         pud_t *pud;
339         pmd_t *pmd;
340
341         *level = PG_LEVEL_NONE;
342
343         if (pgd_none(*pgd))
344                 return NULL;
345
346         pud = pud_offset(pgd, address);
347         if (pud_none(*pud))
348                 return NULL;
349
350         *level = PG_LEVEL_1G;
351         if (pud_large(*pud) || !pud_present(*pud))
352                 return (pte_t *)pud;
353
354         pmd = pmd_offset(pud, address);
355         if (pmd_none(*pmd))
356                 return NULL;
357
358         *level = PG_LEVEL_2M;
359         if (pmd_large(*pmd) || !pmd_present(*pmd))
360                 return (pte_t *)pmd;
361
362         *level = PG_LEVEL_4K;
363
364         return pte_offset_kernel(pmd, address);
365 }
366 EXPORT_SYMBOL_GPL(lookup_address);
367
368 /*
369  * Set the new pmd in all the pgds we know about:
370  */
371 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
372 {
373         /* change init_mm */
374         set_pte_atomic(kpte, pte);
375 #ifdef CONFIG_X86_32
376         if (!SHARED_KERNEL_PMD) {
377                 struct page *page;
378
379                 list_for_each_entry(page, &pgd_list, lru) {
380                         pgd_t *pgd;
381                         pud_t *pud;
382                         pmd_t *pmd;
383
384                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
385                         pud = pud_offset(pgd, address);
386                         pmd = pmd_offset(pud, address);
387                         set_pte_atomic((pte_t *)pmd, pte);
388                 }
389         }
390 #endif
391 }
392
393 static int
394 try_preserve_large_page(pte_t *kpte, unsigned long address,
395                         struct cpa_data *cpa)
396 {
397         unsigned long nextpage_addr, numpages, pmask, psize, flags, addr, pfn;
398         pte_t new_pte, old_pte, *tmp;
399         pgprot_t old_prot, new_prot, req_prot;
400         int i, do_split = 1;
401         unsigned int level;
402
403         if (cpa->force_split)
404                 return 1;
405
406         spin_lock_irqsave(&pgd_lock, flags);
407         /*
408          * Check for races, another CPU might have split this page
409          * up already:
410          */
411         tmp = lookup_address(address, &level);
412         if (tmp != kpte)
413                 goto out_unlock;
414
415         switch (level) {
416         case PG_LEVEL_2M:
417                 psize = PMD_PAGE_SIZE;
418                 pmask = PMD_PAGE_MASK;
419                 break;
420 #ifdef CONFIG_X86_64
421         case PG_LEVEL_1G:
422                 psize = PUD_PAGE_SIZE;
423                 pmask = PUD_PAGE_MASK;
424                 break;
425 #endif
426         default:
427                 do_split = -EINVAL;
428                 goto out_unlock;
429         }
430
431         /*
432          * Calculate the number of pages, which fit into this large
433          * page starting at address:
434          */
435         nextpage_addr = (address + psize) & pmask;
436         numpages = (nextpage_addr - address) >> PAGE_SHIFT;
437         if (numpages < cpa->numpages)
438                 cpa->numpages = numpages;
439
440         /*
441          * We are safe now. Check whether the new pgprot is the same:
442          */
443         old_pte = *kpte;
444         old_prot = new_prot = req_prot = pte_pgprot(old_pte);
445
446         pgprot_val(req_prot) &= ~pgprot_val(cpa->mask_clr);
447         pgprot_val(req_prot) |= pgprot_val(cpa->mask_set);
448
449         /*
450          * old_pte points to the large page base address. So we need
451          * to add the offset of the virtual address:
452          */
453         pfn = pte_pfn(old_pte) + ((address & (psize - 1)) >> PAGE_SHIFT);
454         cpa->pfn = pfn;
455
456         new_prot = static_protections(req_prot, address, pfn);
457
458         /*
459          * We need to check the full range, whether
460          * static_protection() requires a different pgprot for one of
461          * the pages in the range we try to preserve:
462          */
463         addr = address & pmask;
464         pfn = pte_pfn(old_pte);
465         for (i = 0; i < (psize >> PAGE_SHIFT); i++, addr += PAGE_SIZE, pfn++) {
466                 pgprot_t chk_prot = static_protections(req_prot, addr, pfn);
467
468                 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
469                         goto out_unlock;
470         }
471
472         /*
473          * If there are no changes, return. maxpages has been updated
474          * above:
475          */
476         if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
477                 do_split = 0;
478                 goto out_unlock;
479         }
480
481         /*
482          * We need to change the attributes. Check, whether we can
483          * change the large page in one go. We request a split, when
484          * the address is not aligned and the number of pages is
485          * smaller than the number of pages in the large page. Note
486          * that we limited the number of possible pages already to
487          * the number of pages in the large page.
488          */
489         if (address == (address & pmask) && cpa->numpages == (psize >> PAGE_SHIFT)) {
490                 /*
491                  * The address is aligned and the number of pages
492                  * covers the full page.
493                  */
494                 new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
495                 __set_pmd_pte(kpte, address, new_pte);
496                 cpa->flags |= CPA_FLUSHTLB;
497                 do_split = 0;
498         }
499
500 out_unlock:
501         spin_unlock_irqrestore(&pgd_lock, flags);
502
503         return do_split;
504 }
505
506 static int split_large_page(pte_t *kpte, unsigned long address)
507 {
508         unsigned long flags, pfn, pfninc = 1;
509         unsigned int i, level;
510         pte_t *pbase, *tmp;
511         pgprot_t ref_prot;
512         struct page *base;
513
514         if (!debug_pagealloc)
515                 spin_unlock(&cpa_lock);
516         base = alloc_pages(GFP_KERNEL | __GFP_NOTRACK, 0);
517         if (!debug_pagealloc)
518                 spin_lock(&cpa_lock);
519         if (!base)
520                 return -ENOMEM;
521
522         spin_lock_irqsave(&pgd_lock, flags);
523         /*
524          * Check for races, another CPU might have split this page
525          * up for us already:
526          */
527         tmp = lookup_address(address, &level);
528         if (tmp != kpte)
529                 goto out_unlock;
530
531         pbase = (pte_t *)page_address(base);
532         paravirt_alloc_pte(&init_mm, page_to_pfn(base));
533         ref_prot = pte_pgprot(pte_clrhuge(*kpte));
534         /*
535          * If we ever want to utilize the PAT bit, we need to
536          * update this function to make sure it's converted from
537          * bit 12 to bit 7 when we cross from the 2MB level to
538          * the 4K level:
539          */
540         WARN_ON_ONCE(pgprot_val(ref_prot) & _PAGE_PAT_LARGE);
541
542 #ifdef CONFIG_X86_64
543         if (level == PG_LEVEL_1G) {
544                 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
545                 pgprot_val(ref_prot) |= _PAGE_PSE;
546         }
547 #endif
548
549         /*
550          * Get the target pfn from the original entry:
551          */
552         pfn = pte_pfn(*kpte);
553         for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
554                 set_pte(&pbase[i], pfn_pte(pfn, ref_prot));
555
556         if (address >= (unsigned long)__va(0) &&
557                 address < (unsigned long)__va(max_low_pfn_mapped << PAGE_SHIFT))
558                 split_page_count(level);
559
560 #ifdef CONFIG_X86_64
561         if (address >= (unsigned long)__va(1UL<<32) &&
562                 address < (unsigned long)__va(max_pfn_mapped << PAGE_SHIFT))
563                 split_page_count(level);
564 #endif
565
566         /*
567          * Install the new, split up pagetable.
568          *
569          * We use the standard kernel pagetable protections for the new
570          * pagetable protections, the actual ptes set above control the
571          * primary protection behavior:
572          */
573         __set_pmd_pte(kpte, address, mk_pte(base, __pgprot(_KERNPG_TABLE)));
574
575         /*
576          * Intel Atom errata AAH41 workaround.
577          *
578          * The real fix should be in hw or in a microcode update, but
579          * we also probabilistically try to reduce the window of having
580          * a large TLB mixed with 4K TLBs while instruction fetches are
581          * going on.
582          */
583         __flush_tlb_all();
584
585         base = NULL;
586
587 out_unlock:
588         /*
589          * If we dropped out via the lookup_address check under
590          * pgd_lock then stick the page back into the pool:
591          */
592         if (base)
593                 __free_page(base);
594         spin_unlock_irqrestore(&pgd_lock, flags);
595
596         return 0;
597 }
598
599 static int __cpa_process_fault(struct cpa_data *cpa, unsigned long vaddr,
600                                int primary)
601 {
602         /*
603          * Ignore all non primary paths.
604          */
605         if (!primary)
606                 return 0;
607
608         /*
609          * Ignore the NULL PTE for kernel identity mapping, as it is expected
610          * to have holes.
611          * Also set numpages to '1' indicating that we processed cpa req for
612          * one virtual address page and its pfn. TBD: numpages can be set based
613          * on the initial value and the level returned by lookup_address().
614          */
615         if (within(vaddr, PAGE_OFFSET,
616                    PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
617                 cpa->numpages = 1;
618                 cpa->pfn = __pa(vaddr) >> PAGE_SHIFT;
619                 return 0;
620         } else {
621                 WARN(1, KERN_WARNING "CPA: called for zero pte. "
622                         "vaddr = %lx cpa->vaddr = %lx\n", vaddr,
623                         *cpa->vaddr);
624
625                 return -EFAULT;
626         }
627 }
628
629 static int __change_page_attr(struct cpa_data *cpa, int primary)
630 {
631         unsigned long address;
632         int do_split, err;
633         unsigned int level;
634         pte_t *kpte, old_pte;
635
636         if (cpa->flags & CPA_PAGES_ARRAY) {
637                 struct page *page = cpa->pages[cpa->curpage];
638                 if (unlikely(PageHighMem(page)))
639                         return 0;
640                 address = (unsigned long)page_address(page);
641         } else if (cpa->flags & CPA_ARRAY)
642                 address = cpa->vaddr[cpa->curpage];
643         else
644                 address = *cpa->vaddr;
645 repeat:
646         kpte = lookup_address(address, &level);
647         if (!kpte)
648                 return __cpa_process_fault(cpa, address, primary);
649
650         old_pte = *kpte;
651         if (!pte_val(old_pte))
652                 return __cpa_process_fault(cpa, address, primary);
653
654         if (level == PG_LEVEL_4K) {
655                 pte_t new_pte;
656                 pgprot_t new_prot = pte_pgprot(old_pte);
657                 unsigned long pfn = pte_pfn(old_pte);
658
659                 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
660                 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
661
662                 new_prot = static_protections(new_prot, address, pfn);
663
664                 /*
665                  * We need to keep the pfn from the existing PTE,
666                  * after all we're only going to change it's attributes
667                  * not the memory it points to
668                  */
669                 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
670                 cpa->pfn = pfn;
671                 /*
672                  * Do we really change anything ?
673                  */
674                 if (pte_val(old_pte) != pte_val(new_pte)) {
675                         set_pte_atomic(kpte, new_pte);
676                         cpa->flags |= CPA_FLUSHTLB;
677                 }
678                 cpa->numpages = 1;
679                 return 0;
680         }
681
682         /*
683          * Check, whether we can keep the large page intact
684          * and just change the pte:
685          */
686         do_split = try_preserve_large_page(kpte, address, cpa);
687         /*
688          * When the range fits into the existing large page,
689          * return. cp->numpages and cpa->tlbflush have been updated in
690          * try_large_page:
691          */
692         if (do_split <= 0)
693                 return do_split;
694
695         /*
696          * We have to split the large page:
697          */
698         err = split_large_page(kpte, address);
699         if (!err) {
700                 /*
701                  * Do a global flush tlb after splitting the large page
702                  * and before we do the actual change page attribute in the PTE.
703                  *
704                  * With out this, we violate the TLB application note, that says
705                  * "The TLBs may contain both ordinary and large-page
706                  *  translations for a 4-KByte range of linear addresses. This
707                  *  may occur if software modifies the paging structures so that
708                  *  the page size used for the address range changes. If the two
709                  *  translations differ with respect to page frame or attributes
710                  *  (e.g., permissions), processor behavior is undefined and may
711                  *  be implementation-specific."
712                  *
713                  * We do this global tlb flush inside the cpa_lock, so that we
714                  * don't allow any other cpu, with stale tlb entries change the
715                  * page attribute in parallel, that also falls into the
716                  * just split large page entry.
717                  */
718                 flush_tlb_all();
719                 goto repeat;
720         }
721
722         return err;
723 }
724
725 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
726
727 static int cpa_process_alias(struct cpa_data *cpa)
728 {
729         struct cpa_data alias_cpa;
730         unsigned long laddr = (unsigned long)__va(cpa->pfn << PAGE_SHIFT);
731         unsigned long vaddr;
732         int ret;
733
734         if (cpa->pfn >= max_pfn_mapped)
735                 return 0;
736
737 #ifdef CONFIG_X86_64
738         if (cpa->pfn >= max_low_pfn_mapped && cpa->pfn < (1UL<<(32-PAGE_SHIFT)))
739                 return 0;
740 #endif
741         /*
742          * No need to redo, when the primary call touched the direct
743          * mapping already:
744          */
745         if (cpa->flags & CPA_PAGES_ARRAY) {
746                 struct page *page = cpa->pages[cpa->curpage];
747                 if (unlikely(PageHighMem(page)))
748                         return 0;
749                 vaddr = (unsigned long)page_address(page);
750         } else if (cpa->flags & CPA_ARRAY)
751                 vaddr = cpa->vaddr[cpa->curpage];
752         else
753                 vaddr = *cpa->vaddr;
754
755         if (!(within(vaddr, PAGE_OFFSET,
756                     PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT)))) {
757
758                 alias_cpa = *cpa;
759                 alias_cpa.vaddr = &laddr;
760                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
761
762                 ret = __change_page_attr_set_clr(&alias_cpa, 0);
763                 if (ret)
764                         return ret;
765         }
766
767 #ifdef CONFIG_X86_64
768         /*
769          * If the primary call didn't touch the high mapping already
770          * and the physical address is inside the kernel map, we need
771          * to touch the high mapped kernel as well:
772          */
773         if (!within(vaddr, (unsigned long)_text, _brk_end) &&
774             within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn())) {
775                 unsigned long temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) +
776                                                __START_KERNEL_map - phys_base;
777                 alias_cpa = *cpa;
778                 alias_cpa.vaddr = &temp_cpa_vaddr;
779                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
780
781                 /*
782                  * The high mapping range is imprecise, so ignore the
783                  * return value.
784                  */
785                 __change_page_attr_set_clr(&alias_cpa, 0);
786         }
787 #endif
788
789         return 0;
790 }
791
792 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
793 {
794         int ret, numpages = cpa->numpages;
795
796         while (numpages) {
797                 /*
798                  * Store the remaining nr of pages for the large page
799                  * preservation check.
800                  */
801                 cpa->numpages = numpages;
802                 /* for array changes, we can't use large page */
803                 if (cpa->flags & (CPA_ARRAY | CPA_PAGES_ARRAY))
804                         cpa->numpages = 1;
805
806                 if (!debug_pagealloc)
807                         spin_lock(&cpa_lock);
808                 ret = __change_page_attr(cpa, checkalias);
809                 if (!debug_pagealloc)
810                         spin_unlock(&cpa_lock);
811                 if (ret)
812                         return ret;
813
814                 if (checkalias) {
815                         ret = cpa_process_alias(cpa);
816                         if (ret)
817                                 return ret;
818                 }
819
820                 /*
821                  * Adjust the number of pages with the result of the
822                  * CPA operation. Either a large page has been
823                  * preserved or a single page update happened.
824                  */
825                 BUG_ON(cpa->numpages > numpages);
826                 numpages -= cpa->numpages;
827                 if (cpa->flags & (CPA_PAGES_ARRAY | CPA_ARRAY))
828                         cpa->curpage++;
829                 else
830                         *cpa->vaddr += cpa->numpages * PAGE_SIZE;
831
832         }
833         return 0;
834 }
835
836 static inline int cache_attr(pgprot_t attr)
837 {
838         return pgprot_val(attr) &
839                 (_PAGE_PAT | _PAGE_PAT_LARGE | _PAGE_PWT | _PAGE_PCD);
840 }
841
842 static int change_page_attr_set_clr(unsigned long *addr, int numpages,
843                                     pgprot_t mask_set, pgprot_t mask_clr,
844                                     int force_split, int in_flag,
845                                     struct page **pages)
846 {
847         struct cpa_data cpa;
848         int ret, cache, checkalias;
849         unsigned long baddr = 0;
850
851         /*
852          * Check, if we are requested to change a not supported
853          * feature:
854          */
855         mask_set = canon_pgprot(mask_set);
856         mask_clr = canon_pgprot(mask_clr);
857         if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
858                 return 0;
859
860         /* Ensure we are PAGE_SIZE aligned */
861         if (in_flag & CPA_ARRAY) {
862                 int i;
863                 for (i = 0; i < numpages; i++) {
864                         if (addr[i] & ~PAGE_MASK) {
865                                 addr[i] &= PAGE_MASK;
866                                 WARN_ON_ONCE(1);
867                         }
868                 }
869         } else if (!(in_flag & CPA_PAGES_ARRAY)) {
870                 /*
871                  * in_flag of CPA_PAGES_ARRAY implies it is aligned.
872                  * No need to cehck in that case
873                  */
874                 if (*addr & ~PAGE_MASK) {
875                         *addr &= PAGE_MASK;
876                         /*
877                          * People should not be passing in unaligned addresses:
878                          */
879                         WARN_ON_ONCE(1);
880                 }
881                 /*
882                  * Save address for cache flush. *addr is modified in the call
883                  * to __change_page_attr_set_clr() below.
884                  */
885                 baddr = *addr;
886         }
887
888         /* Must avoid aliasing mappings in the highmem code */
889         kmap_flush_unused();
890
891         vm_unmap_aliases();
892
893         cpa.vaddr = addr;
894         cpa.pages = pages;
895         cpa.numpages = numpages;
896         cpa.mask_set = mask_set;
897         cpa.mask_clr = mask_clr;
898         cpa.flags = 0;
899         cpa.curpage = 0;
900         cpa.force_split = force_split;
901
902         if (in_flag & (CPA_ARRAY | CPA_PAGES_ARRAY))
903                 cpa.flags |= in_flag;
904
905         /* No alias checking for _NX bit modifications */
906         checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
907
908         ret = __change_page_attr_set_clr(&cpa, checkalias);
909
910         /*
911          * Check whether we really changed something:
912          */
913         if (!(cpa.flags & CPA_FLUSHTLB))
914                 goto out;
915
916         /*
917          * No need to flush, when we did not set any of the caching
918          * attributes:
919          */
920         cache = cache_attr(mask_set);
921
922         /*
923          * On success we use clflush, when the CPU supports it to
924          * avoid the wbindv. If the CPU does not support it and in the
925          * error case we fall back to cpa_flush_all (which uses
926          * wbindv):
927          */
928         if (!ret && cpu_has_clflush) {
929                 if (cpa.flags & (CPA_PAGES_ARRAY | CPA_ARRAY)) {
930                         cpa_flush_array(addr, numpages, cache,
931                                         cpa.flags, pages);
932                 } else
933                         cpa_flush_range(baddr, numpages, cache);
934         } else
935                 cpa_flush_all(cache);
936
937 out:
938         return ret;
939 }
940
941 static inline int change_page_attr_set(unsigned long *addr, int numpages,
942                                        pgprot_t mask, int array)
943 {
944         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
945                 (array ? CPA_ARRAY : 0), NULL);
946 }
947
948 static inline int change_page_attr_clear(unsigned long *addr, int numpages,
949                                          pgprot_t mask, int array)
950 {
951         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
952                 (array ? CPA_ARRAY : 0), NULL);
953 }
954
955 static inline int cpa_set_pages_array(struct page **pages, int numpages,
956                                        pgprot_t mask)
957 {
958         return change_page_attr_set_clr(NULL, numpages, mask, __pgprot(0), 0,
959                 CPA_PAGES_ARRAY, pages);
960 }
961
962 static inline int cpa_clear_pages_array(struct page **pages, int numpages,
963                                          pgprot_t mask)
964 {
965         return change_page_attr_set_clr(NULL, numpages, __pgprot(0), mask, 0,
966                 CPA_PAGES_ARRAY, pages);
967 }
968
969 int _set_memory_uc(unsigned long addr, int numpages)
970 {
971         /*
972          * for now UC MINUS. see comments in ioremap_nocache()
973          */
974         return change_page_attr_set(&addr, numpages,
975                                     __pgprot(_PAGE_CACHE_UC_MINUS), 0);
976 }
977
978 int set_memory_uc(unsigned long addr, int numpages)
979 {
980         int ret;
981
982         /*
983          * for now UC MINUS. see comments in ioremap_nocache()
984          */
985         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
986                             _PAGE_CACHE_UC_MINUS, NULL);
987         if (ret)
988                 goto out_err;
989
990         ret = _set_memory_uc(addr, numpages);
991         if (ret)
992                 goto out_free;
993
994         return 0;
995
996 out_free:
997         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
998 out_err:
999         return ret;
1000 }
1001 EXPORT_SYMBOL(set_memory_uc);
1002
1003 int _set_memory_array(unsigned long *addr, int addrinarray,
1004                 unsigned long new_type)
1005 {
1006         int i, j;
1007         int ret;
1008
1009         /*
1010          * for now UC MINUS. see comments in ioremap_nocache()
1011          */
1012         for (i = 0; i < addrinarray; i++) {
1013                 ret = reserve_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE,
1014                                         new_type, NULL);
1015                 if (ret)
1016                         goto out_free;
1017         }
1018
1019         ret = change_page_attr_set(addr, addrinarray,
1020                                     __pgprot(_PAGE_CACHE_UC_MINUS), 1);
1021
1022         if (!ret && new_type == _PAGE_CACHE_WC)
1023                 ret = change_page_attr_set_clr(addr, addrinarray,
1024                                                __pgprot(_PAGE_CACHE_WC),
1025                                                __pgprot(_PAGE_CACHE_MASK),
1026                                                0, CPA_ARRAY, NULL);
1027         if (ret)
1028                 goto out_free;
1029
1030         return 0;
1031
1032 out_free:
1033         for (j = 0; j < i; j++)
1034                 free_memtype(__pa(addr[j]), __pa(addr[j]) + PAGE_SIZE);
1035
1036         return ret;
1037 }
1038
1039 int set_memory_array_uc(unsigned long *addr, int addrinarray)
1040 {
1041         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_UC_MINUS);
1042 }
1043 EXPORT_SYMBOL(set_memory_array_uc);
1044
1045 int set_memory_array_wc(unsigned long *addr, int addrinarray)
1046 {
1047         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_WC);
1048 }
1049 EXPORT_SYMBOL(set_memory_array_wc);
1050
1051 int _set_memory_wc(unsigned long addr, int numpages)
1052 {
1053         int ret;
1054         unsigned long addr_copy = addr;
1055
1056         ret = change_page_attr_set(&addr, numpages,
1057                                     __pgprot(_PAGE_CACHE_UC_MINUS), 0);
1058         if (!ret) {
1059                 ret = change_page_attr_set_clr(&addr_copy, numpages,
1060                                                __pgprot(_PAGE_CACHE_WC),
1061                                                __pgprot(_PAGE_CACHE_MASK),
1062                                                0, 0, NULL);
1063         }
1064         return ret;
1065 }
1066
1067 int set_memory_wc(unsigned long addr, int numpages)
1068 {
1069         int ret;
1070
1071         if (!pat_enabled)
1072                 return set_memory_uc(addr, numpages);
1073
1074         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1075                 _PAGE_CACHE_WC, NULL);
1076         if (ret)
1077                 goto out_err;
1078
1079         ret = _set_memory_wc(addr, numpages);
1080         if (ret)
1081                 goto out_free;
1082
1083         return 0;
1084
1085 out_free:
1086         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1087 out_err:
1088         return ret;
1089 }
1090 EXPORT_SYMBOL(set_memory_wc);
1091
1092 int _set_memory_wb(unsigned long addr, int numpages)
1093 {
1094         return change_page_attr_clear(&addr, numpages,
1095                                       __pgprot(_PAGE_CACHE_MASK), 0);
1096 }
1097
1098 int set_memory_wb(unsigned long addr, int numpages)
1099 {
1100         int ret;
1101
1102         ret = _set_memory_wb(addr, numpages);
1103         if (ret)
1104                 return ret;
1105
1106         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1107         return 0;
1108 }
1109 EXPORT_SYMBOL(set_memory_wb);
1110
1111 int set_memory_array_wb(unsigned long *addr, int addrinarray)
1112 {
1113         int i;
1114         int ret;
1115
1116         ret = change_page_attr_clear(addr, addrinarray,
1117                                       __pgprot(_PAGE_CACHE_MASK), 1);
1118         if (ret)
1119                 return ret;
1120
1121         for (i = 0; i < addrinarray; i++)
1122                 free_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE);
1123
1124         return 0;
1125 }
1126 EXPORT_SYMBOL(set_memory_array_wb);
1127
1128 int set_memory_x(unsigned long addr, int numpages)
1129 {
1130         if (!(__supported_pte_mask & _PAGE_NX))
1131                 return 0;
1132
1133         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1134 }
1135 EXPORT_SYMBOL(set_memory_x);
1136
1137 int set_memory_nx(unsigned long addr, int numpages)
1138 {
1139         if (!(__supported_pte_mask & _PAGE_NX))
1140                 return 0;
1141
1142         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1143 }
1144 EXPORT_SYMBOL(set_memory_nx);
1145
1146 int set_memory_ro(unsigned long addr, int numpages)
1147 {
1148         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1149 }
1150 EXPORT_SYMBOL_GPL(set_memory_ro);
1151
1152 int set_memory_rw(unsigned long addr, int numpages)
1153 {
1154         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
1155 }
1156 EXPORT_SYMBOL_GPL(set_memory_rw);
1157
1158 int set_memory_np(unsigned long addr, int numpages)
1159 {
1160         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
1161 }
1162
1163 int set_memory_4k(unsigned long addr, int numpages)
1164 {
1165         return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1166                                         __pgprot(0), 1, 0, NULL);
1167 }
1168
1169 int set_pages_uc(struct page *page, int numpages)
1170 {
1171         unsigned long addr = (unsigned long)page_address(page);
1172
1173         return set_memory_uc(addr, numpages);
1174 }
1175 EXPORT_SYMBOL(set_pages_uc);
1176
1177 static int _set_pages_array(struct page **pages, int addrinarray,
1178                 unsigned long new_type)
1179 {
1180         unsigned long start;
1181         unsigned long end;
1182         int i;
1183         int free_idx;
1184         int ret;
1185
1186         for (i = 0; i < addrinarray; i++) {
1187                 if (PageHighMem(pages[i]))
1188                         continue;
1189                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1190                 end = start + PAGE_SIZE;
1191                 if (reserve_memtype(start, end, new_type, NULL))
1192                         goto err_out;
1193         }
1194
1195         ret = cpa_set_pages_array(pages, addrinarray,
1196                         __pgprot(_PAGE_CACHE_UC_MINUS));
1197         if (!ret && new_type == _PAGE_CACHE_WC)
1198                 ret = change_page_attr_set_clr(NULL, addrinarray,
1199                                                __pgprot(_PAGE_CACHE_WC),
1200                                                __pgprot(_PAGE_CACHE_MASK),
1201                                                0, CPA_PAGES_ARRAY, pages);
1202         if (ret)
1203                 goto err_out;
1204         return 0; /* Success */
1205 err_out:
1206         free_idx = i;
1207         for (i = 0; i < free_idx; i++) {
1208                 if (PageHighMem(pages[i]))
1209                         continue;
1210                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1211                 end = start + PAGE_SIZE;
1212                 free_memtype(start, end);
1213         }
1214         return -EINVAL;
1215 }
1216
1217 int set_pages_array_uc(struct page **pages, int addrinarray)
1218 {
1219         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_UC_MINUS);
1220 }
1221 EXPORT_SYMBOL(set_pages_array_uc);
1222
1223 int set_pages_array_wc(struct page **pages, int addrinarray)
1224 {
1225         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_WC);
1226 }
1227 EXPORT_SYMBOL(set_pages_array_wc);
1228
1229 int set_pages_wb(struct page *page, int numpages)
1230 {
1231         unsigned long addr = (unsigned long)page_address(page);
1232
1233         return set_memory_wb(addr, numpages);
1234 }
1235 EXPORT_SYMBOL(set_pages_wb);
1236
1237 int set_pages_array_wb(struct page **pages, int addrinarray)
1238 {
1239         int retval;
1240         unsigned long start;
1241         unsigned long end;
1242         int i;
1243
1244         retval = cpa_clear_pages_array(pages, addrinarray,
1245                         __pgprot(_PAGE_CACHE_MASK));
1246         if (retval)
1247                 return retval;
1248
1249         for (i = 0; i < addrinarray; i++) {
1250                 if (PageHighMem(pages[i]))
1251                         continue;
1252                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1253                 end = start + PAGE_SIZE;
1254                 free_memtype(start, end);
1255         }
1256
1257         return 0;
1258 }
1259 EXPORT_SYMBOL(set_pages_array_wb);
1260
1261 int set_pages_x(struct page *page, int numpages)
1262 {
1263         unsigned long addr = (unsigned long)page_address(page);
1264
1265         return set_memory_x(addr, numpages);
1266 }
1267 EXPORT_SYMBOL(set_pages_x);
1268
1269 int set_pages_nx(struct page *page, int numpages)
1270 {
1271         unsigned long addr = (unsigned long)page_address(page);
1272
1273         return set_memory_nx(addr, numpages);
1274 }
1275 EXPORT_SYMBOL(set_pages_nx);
1276
1277 int set_pages_ro(struct page *page, int numpages)
1278 {
1279         unsigned long addr = (unsigned long)page_address(page);
1280
1281         return set_memory_ro(addr, numpages);
1282 }
1283
1284 int set_pages_rw(struct page *page, int numpages)
1285 {
1286         unsigned long addr = (unsigned long)page_address(page);
1287
1288         return set_memory_rw(addr, numpages);
1289 }
1290
1291 #ifdef CONFIG_DEBUG_PAGEALLOC
1292
1293 static int __set_pages_p(struct page *page, int numpages)
1294 {
1295         unsigned long tempaddr = (unsigned long) page_address(page);
1296         struct cpa_data cpa = { .vaddr = &tempaddr,
1297                                 .numpages = numpages,
1298                                 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1299                                 .mask_clr = __pgprot(0),
1300                                 .flags = 0};
1301
1302         /*
1303          * No alias checking needed for setting present flag. otherwise,
1304          * we may need to break large pages for 64-bit kernel text
1305          * mappings (this adds to complexity if we want to do this from
1306          * atomic context especially). Let's keep it simple!
1307          */
1308         return __change_page_attr_set_clr(&cpa, 0);
1309 }
1310
1311 static int __set_pages_np(struct page *page, int numpages)
1312 {
1313         unsigned long tempaddr = (unsigned long) page_address(page);
1314         struct cpa_data cpa = { .vaddr = &tempaddr,
1315                                 .numpages = numpages,
1316                                 .mask_set = __pgprot(0),
1317                                 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1318                                 .flags = 0};
1319
1320         /*
1321          * No alias checking needed for setting not present flag. otherwise,
1322          * we may need to break large pages for 64-bit kernel text
1323          * mappings (this adds to complexity if we want to do this from
1324          * atomic context especially). Let's keep it simple!
1325          */
1326         return __change_page_attr_set_clr(&cpa, 0);
1327 }
1328
1329 void kernel_map_pages(struct page *page, int numpages, int enable)
1330 {
1331         if (PageHighMem(page))
1332                 return;
1333         if (!enable) {
1334                 debug_check_no_locks_freed(page_address(page),
1335                                            numpages * PAGE_SIZE);
1336         }
1337
1338         /*
1339          * If page allocator is not up yet then do not call c_p_a():
1340          */
1341         if (!debug_pagealloc_enabled)
1342                 return;
1343
1344         /*
1345          * The return value is ignored as the calls cannot fail.
1346          * Large pages for identity mappings are not used at boot time
1347          * and hence no memory allocations during large page split.
1348          */
1349         if (enable)
1350                 __set_pages_p(page, numpages);
1351         else
1352                 __set_pages_np(page, numpages);
1353
1354         /*
1355          * We should perform an IPI and flush all tlbs,
1356          * but that can deadlock->flush only current cpu:
1357          */
1358         __flush_tlb_all();
1359 }
1360
1361 #ifdef CONFIG_HIBERNATION
1362
1363 bool kernel_page_present(struct page *page)
1364 {
1365         unsigned int level;
1366         pte_t *pte;
1367
1368         if (PageHighMem(page))
1369                 return false;
1370
1371         pte = lookup_address((unsigned long)page_address(page), &level);
1372         return (pte_val(*pte) & _PAGE_PRESENT);
1373 }
1374
1375 #endif /* CONFIG_HIBERNATION */
1376
1377 #endif /* CONFIG_DEBUG_PAGEALLOC */
1378
1379 /*
1380  * The testcases use internal knowledge of the implementation that shouldn't
1381  * be exposed to the rest of the kernel. Include these directly here.
1382  */
1383 #ifdef CONFIG_CPA_DEBUG
1384 #include "pageattr-test.c"
1385 #endif