]> git.karo-electronics.de Git - karo-tx-linux.git/blob - arch/x86/mm/pageattr.c
Merge tag 'sunxi-fixes-for-4.12' of https://git.kernel.org/pub/scm/linux/kernel/git...
[karo-tx-linux.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/sched.h>
8 #include <linux/mm.h>
9 #include <linux/interrupt.h>
10 #include <linux/seq_file.h>
11 #include <linux/debugfs.h>
12 #include <linux/pfn.h>
13 #include <linux/percpu.h>
14 #include <linux/gfp.h>
15 #include <linux/pci.h>
16 #include <linux/vmalloc.h>
17
18 #include <asm/e820/api.h>
19 #include <asm/processor.h>
20 #include <asm/tlbflush.h>
21 #include <asm/sections.h>
22 #include <asm/setup.h>
23 #include <linux/uaccess.h>
24 #include <asm/pgalloc.h>
25 #include <asm/proto.h>
26 #include <asm/pat.h>
27 #include <asm/set_memory.h>
28
29 /*
30  * The current flushing context - we pass it instead of 5 arguments:
31  */
32 struct cpa_data {
33         unsigned long   *vaddr;
34         pgd_t           *pgd;
35         pgprot_t        mask_set;
36         pgprot_t        mask_clr;
37         unsigned long   numpages;
38         int             flags;
39         unsigned long   pfn;
40         unsigned        force_split : 1;
41         int             curpage;
42         struct page     **pages;
43 };
44
45 /*
46  * Serialize cpa() (for !DEBUG_PAGEALLOC which uses large identity mappings)
47  * using cpa_lock. So that we don't allow any other cpu, with stale large tlb
48  * entries change the page attribute in parallel to some other cpu
49  * splitting a large page entry along with changing the attribute.
50  */
51 static DEFINE_SPINLOCK(cpa_lock);
52
53 #define CPA_FLUSHTLB 1
54 #define CPA_ARRAY 2
55 #define CPA_PAGES_ARRAY 4
56
57 #ifdef CONFIG_PROC_FS
58 static unsigned long direct_pages_count[PG_LEVEL_NUM];
59
60 void update_page_count(int level, unsigned long pages)
61 {
62         /* Protect against CPA */
63         spin_lock(&pgd_lock);
64         direct_pages_count[level] += pages;
65         spin_unlock(&pgd_lock);
66 }
67
68 static void split_page_count(int level)
69 {
70         if (direct_pages_count[level] == 0)
71                 return;
72
73         direct_pages_count[level]--;
74         direct_pages_count[level - 1] += PTRS_PER_PTE;
75 }
76
77 void arch_report_meminfo(struct seq_file *m)
78 {
79         seq_printf(m, "DirectMap4k:    %8lu kB\n",
80                         direct_pages_count[PG_LEVEL_4K] << 2);
81 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
82         seq_printf(m, "DirectMap2M:    %8lu kB\n",
83                         direct_pages_count[PG_LEVEL_2M] << 11);
84 #else
85         seq_printf(m, "DirectMap4M:    %8lu kB\n",
86                         direct_pages_count[PG_LEVEL_2M] << 12);
87 #endif
88         if (direct_gbpages)
89                 seq_printf(m, "DirectMap1G:    %8lu kB\n",
90                         direct_pages_count[PG_LEVEL_1G] << 20);
91 }
92 #else
93 static inline void split_page_count(int level) { }
94 #endif
95
96 #ifdef CONFIG_X86_64
97
98 static inline unsigned long highmap_start_pfn(void)
99 {
100         return __pa_symbol(_text) >> PAGE_SHIFT;
101 }
102
103 static inline unsigned long highmap_end_pfn(void)
104 {
105         /* Do not reference physical address outside the kernel. */
106         return __pa_symbol(roundup(_brk_end, PMD_SIZE) - 1) >> PAGE_SHIFT;
107 }
108
109 #endif
110
111 static inline int
112 within(unsigned long addr, unsigned long start, unsigned long end)
113 {
114         return addr >= start && addr < end;
115 }
116
117 static inline int
118 within_inclusive(unsigned long addr, unsigned long start, unsigned long end)
119 {
120         return addr >= start && addr <= end;
121 }
122
123 /*
124  * Flushing functions
125  */
126
127 /**
128  * clflush_cache_range - flush a cache range with clflush
129  * @vaddr:      virtual start address
130  * @size:       number of bytes to flush
131  *
132  * clflushopt is an unordered instruction which needs fencing with mfence or
133  * sfence to avoid ordering issues.
134  */
135 void clflush_cache_range(void *vaddr, unsigned int size)
136 {
137         const unsigned long clflush_size = boot_cpu_data.x86_clflush_size;
138         void *p = (void *)((unsigned long)vaddr & ~(clflush_size - 1));
139         void *vend = vaddr + size;
140
141         if (p >= vend)
142                 return;
143
144         mb();
145
146         for (; p < vend; p += clflush_size)
147                 clflushopt(p);
148
149         mb();
150 }
151 EXPORT_SYMBOL_GPL(clflush_cache_range);
152
153 static void __cpa_flush_all(void *arg)
154 {
155         unsigned long cache = (unsigned long)arg;
156
157         /*
158          * Flush all to work around Errata in early athlons regarding
159          * large page flushing.
160          */
161         __flush_tlb_all();
162
163         if (cache && boot_cpu_data.x86 >= 4)
164                 wbinvd();
165 }
166
167 static void cpa_flush_all(unsigned long cache)
168 {
169         BUG_ON(irqs_disabled());
170
171         on_each_cpu(__cpa_flush_all, (void *) cache, 1);
172 }
173
174 static void __cpa_flush_range(void *arg)
175 {
176         /*
177          * We could optimize that further and do individual per page
178          * tlb invalidates for a low number of pages. Caveat: we must
179          * flush the high aliases on 64bit as well.
180          */
181         __flush_tlb_all();
182 }
183
184 static void cpa_flush_range(unsigned long start, int numpages, int cache)
185 {
186         unsigned int i, level;
187         unsigned long addr;
188
189         BUG_ON(irqs_disabled());
190         WARN_ON(PAGE_ALIGN(start) != start);
191
192         on_each_cpu(__cpa_flush_range, NULL, 1);
193
194         if (!cache)
195                 return;
196
197         /*
198          * We only need to flush on one CPU,
199          * clflush is a MESI-coherent instruction that
200          * will cause all other CPUs to flush the same
201          * cachelines:
202          */
203         for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
204                 pte_t *pte = lookup_address(addr, &level);
205
206                 /*
207                  * Only flush present addresses:
208                  */
209                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
210                         clflush_cache_range((void *) addr, PAGE_SIZE);
211         }
212 }
213
214 static void cpa_flush_array(unsigned long *start, int numpages, int cache,
215                             int in_flags, struct page **pages)
216 {
217         unsigned int i, level;
218 #ifdef CONFIG_PREEMPT
219         /*
220          * Avoid wbinvd() because it causes latencies on all CPUs,
221          * regardless of any CPU isolation that may be in effect.
222          *
223          * This should be extended for CAT enabled systems independent of
224          * PREEMPT because wbinvd() does not respect the CAT partitions and
225          * this is exposed to unpriviledged users through the graphics
226          * subsystem.
227          */
228         unsigned long do_wbinvd = 0;
229 #else
230         unsigned long do_wbinvd = cache && numpages >= 1024; /* 4M threshold */
231 #endif
232
233         BUG_ON(irqs_disabled());
234
235         on_each_cpu(__cpa_flush_all, (void *) do_wbinvd, 1);
236
237         if (!cache || do_wbinvd)
238                 return;
239
240         /*
241          * We only need to flush on one CPU,
242          * clflush is a MESI-coherent instruction that
243          * will cause all other CPUs to flush the same
244          * cachelines:
245          */
246         for (i = 0; i < numpages; i++) {
247                 unsigned long addr;
248                 pte_t *pte;
249
250                 if (in_flags & CPA_PAGES_ARRAY)
251                         addr = (unsigned long)page_address(pages[i]);
252                 else
253                         addr = start[i];
254
255                 pte = lookup_address(addr, &level);
256
257                 /*
258                  * Only flush present addresses:
259                  */
260                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
261                         clflush_cache_range((void *)addr, PAGE_SIZE);
262         }
263 }
264
265 /*
266  * Certain areas of memory on x86 require very specific protection flags,
267  * for example the BIOS area or kernel text. Callers don't always get this
268  * right (again, ioremap() on BIOS memory is not uncommon) so this function
269  * checks and fixes these known static required protection bits.
270  */
271 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
272                                    unsigned long pfn)
273 {
274         pgprot_t forbidden = __pgprot(0);
275
276         /*
277          * The BIOS area between 640k and 1Mb needs to be executable for
278          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
279          */
280 #ifdef CONFIG_PCI_BIOS
281         if (pcibios_enabled && within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
282                 pgprot_val(forbidden) |= _PAGE_NX;
283 #endif
284
285         /*
286          * The kernel text needs to be executable for obvious reasons
287          * Does not cover __inittext since that is gone later on. On
288          * 64bit we do not enforce !NX on the low mapping
289          */
290         if (within(address, (unsigned long)_text, (unsigned long)_etext))
291                 pgprot_val(forbidden) |= _PAGE_NX;
292
293         /*
294          * The .rodata section needs to be read-only. Using the pfn
295          * catches all aliases.
296          */
297         if (within(pfn, __pa_symbol(__start_rodata) >> PAGE_SHIFT,
298                    __pa_symbol(__end_rodata) >> PAGE_SHIFT))
299                 pgprot_val(forbidden) |= _PAGE_RW;
300
301 #if defined(CONFIG_X86_64)
302         /*
303          * Once the kernel maps the text as RO (kernel_set_to_readonly is set),
304          * kernel text mappings for the large page aligned text, rodata sections
305          * will be always read-only. For the kernel identity mappings covering
306          * the holes caused by this alignment can be anything that user asks.
307          *
308          * This will preserve the large page mappings for kernel text/data
309          * at no extra cost.
310          */
311         if (kernel_set_to_readonly &&
312             within(address, (unsigned long)_text,
313                    (unsigned long)__end_rodata_hpage_align)) {
314                 unsigned int level;
315
316                 /*
317                  * Don't enforce the !RW mapping for the kernel text mapping,
318                  * if the current mapping is already using small page mapping.
319                  * No need to work hard to preserve large page mappings in this
320                  * case.
321                  *
322                  * This also fixes the Linux Xen paravirt guest boot failure
323                  * (because of unexpected read-only mappings for kernel identity
324                  * mappings). In this paravirt guest case, the kernel text
325                  * mapping and the kernel identity mapping share the same
326                  * page-table pages. Thus we can't really use different
327                  * protections for the kernel text and identity mappings. Also,
328                  * these shared mappings are made of small page mappings.
329                  * Thus this don't enforce !RW mapping for small page kernel
330                  * text mapping logic will help Linux Xen parvirt guest boot
331                  * as well.
332                  */
333                 if (lookup_address(address, &level) && (level != PG_LEVEL_4K))
334                         pgprot_val(forbidden) |= _PAGE_RW;
335         }
336 #endif
337
338         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
339
340         return prot;
341 }
342
343 /*
344  * Lookup the page table entry for a virtual address in a specific pgd.
345  * Return a pointer to the entry and the level of the mapping.
346  */
347 pte_t *lookup_address_in_pgd(pgd_t *pgd, unsigned long address,
348                              unsigned int *level)
349 {
350         p4d_t *p4d;
351         pud_t *pud;
352         pmd_t *pmd;
353
354         *level = PG_LEVEL_NONE;
355
356         if (pgd_none(*pgd))
357                 return NULL;
358
359         p4d = p4d_offset(pgd, address);
360         if (p4d_none(*p4d))
361                 return NULL;
362
363         *level = PG_LEVEL_512G;
364         if (p4d_large(*p4d) || !p4d_present(*p4d))
365                 return (pte_t *)p4d;
366
367         pud = pud_offset(p4d, address);
368         if (pud_none(*pud))
369                 return NULL;
370
371         *level = PG_LEVEL_1G;
372         if (pud_large(*pud) || !pud_present(*pud))
373                 return (pte_t *)pud;
374
375         pmd = pmd_offset(pud, address);
376         if (pmd_none(*pmd))
377                 return NULL;
378
379         *level = PG_LEVEL_2M;
380         if (pmd_large(*pmd) || !pmd_present(*pmd))
381                 return (pte_t *)pmd;
382
383         *level = PG_LEVEL_4K;
384
385         return pte_offset_kernel(pmd, address);
386 }
387
388 /*
389  * Lookup the page table entry for a virtual address. Return a pointer
390  * to the entry and the level of the mapping.
391  *
392  * Note: We return pud and pmd either when the entry is marked large
393  * or when the present bit is not set. Otherwise we would return a
394  * pointer to a nonexisting mapping.
395  */
396 pte_t *lookup_address(unsigned long address, unsigned int *level)
397 {
398         return lookup_address_in_pgd(pgd_offset_k(address), address, level);
399 }
400 EXPORT_SYMBOL_GPL(lookup_address);
401
402 static pte_t *_lookup_address_cpa(struct cpa_data *cpa, unsigned long address,
403                                   unsigned int *level)
404 {
405         if (cpa->pgd)
406                 return lookup_address_in_pgd(cpa->pgd + pgd_index(address),
407                                                address, level);
408
409         return lookup_address(address, level);
410 }
411
412 /*
413  * Lookup the PMD entry for a virtual address. Return a pointer to the entry
414  * or NULL if not present.
415  */
416 pmd_t *lookup_pmd_address(unsigned long address)
417 {
418         pgd_t *pgd;
419         p4d_t *p4d;
420         pud_t *pud;
421
422         pgd = pgd_offset_k(address);
423         if (pgd_none(*pgd))
424                 return NULL;
425
426         p4d = p4d_offset(pgd, address);
427         if (p4d_none(*p4d) || p4d_large(*p4d) || !p4d_present(*p4d))
428                 return NULL;
429
430         pud = pud_offset(p4d, address);
431         if (pud_none(*pud) || pud_large(*pud) || !pud_present(*pud))
432                 return NULL;
433
434         return pmd_offset(pud, address);
435 }
436
437 /*
438  * This is necessary because __pa() does not work on some
439  * kinds of memory, like vmalloc() or the alloc_remap()
440  * areas on 32-bit NUMA systems.  The percpu areas can
441  * end up in this kind of memory, for instance.
442  *
443  * This could be optimized, but it is only intended to be
444  * used at inititalization time, and keeping it
445  * unoptimized should increase the testing coverage for
446  * the more obscure platforms.
447  */
448 phys_addr_t slow_virt_to_phys(void *__virt_addr)
449 {
450         unsigned long virt_addr = (unsigned long)__virt_addr;
451         phys_addr_t phys_addr;
452         unsigned long offset;
453         enum pg_level level;
454         pte_t *pte;
455
456         pte = lookup_address(virt_addr, &level);
457         BUG_ON(!pte);
458
459         /*
460          * pXX_pfn() returns unsigned long, which must be cast to phys_addr_t
461          * before being left-shifted PAGE_SHIFT bits -- this trick is to
462          * make 32-PAE kernel work correctly.
463          */
464         switch (level) {
465         case PG_LEVEL_1G:
466                 phys_addr = (phys_addr_t)pud_pfn(*(pud_t *)pte) << PAGE_SHIFT;
467                 offset = virt_addr & ~PUD_PAGE_MASK;
468                 break;
469         case PG_LEVEL_2M:
470                 phys_addr = (phys_addr_t)pmd_pfn(*(pmd_t *)pte) << PAGE_SHIFT;
471                 offset = virt_addr & ~PMD_PAGE_MASK;
472                 break;
473         default:
474                 phys_addr = (phys_addr_t)pte_pfn(*pte) << PAGE_SHIFT;
475                 offset = virt_addr & ~PAGE_MASK;
476         }
477
478         return (phys_addr_t)(phys_addr | offset);
479 }
480 EXPORT_SYMBOL_GPL(slow_virt_to_phys);
481
482 /*
483  * Set the new pmd in all the pgds we know about:
484  */
485 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
486 {
487         /* change init_mm */
488         set_pte_atomic(kpte, pte);
489 #ifdef CONFIG_X86_32
490         if (!SHARED_KERNEL_PMD) {
491                 struct page *page;
492
493                 list_for_each_entry(page, &pgd_list, lru) {
494                         pgd_t *pgd;
495                         p4d_t *p4d;
496                         pud_t *pud;
497                         pmd_t *pmd;
498
499                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
500                         p4d = p4d_offset(pgd, address);
501                         pud = pud_offset(p4d, address);
502                         pmd = pmd_offset(pud, address);
503                         set_pte_atomic((pte_t *)pmd, pte);
504                 }
505         }
506 #endif
507 }
508
509 static int
510 try_preserve_large_page(pte_t *kpte, unsigned long address,
511                         struct cpa_data *cpa)
512 {
513         unsigned long nextpage_addr, numpages, pmask, psize, addr, pfn, old_pfn;
514         pte_t new_pte, old_pte, *tmp;
515         pgprot_t old_prot, new_prot, req_prot;
516         int i, do_split = 1;
517         enum pg_level level;
518
519         if (cpa->force_split)
520                 return 1;
521
522         spin_lock(&pgd_lock);
523         /*
524          * Check for races, another CPU might have split this page
525          * up already:
526          */
527         tmp = _lookup_address_cpa(cpa, address, &level);
528         if (tmp != kpte)
529                 goto out_unlock;
530
531         switch (level) {
532         case PG_LEVEL_2M:
533                 old_prot = pmd_pgprot(*(pmd_t *)kpte);
534                 old_pfn = pmd_pfn(*(pmd_t *)kpte);
535                 break;
536         case PG_LEVEL_1G:
537                 old_prot = pud_pgprot(*(pud_t *)kpte);
538                 old_pfn = pud_pfn(*(pud_t *)kpte);
539                 break;
540         default:
541                 do_split = -EINVAL;
542                 goto out_unlock;
543         }
544
545         psize = page_level_size(level);
546         pmask = page_level_mask(level);
547
548         /*
549          * Calculate the number of pages, which fit into this large
550          * page starting at address:
551          */
552         nextpage_addr = (address + psize) & pmask;
553         numpages = (nextpage_addr - address) >> PAGE_SHIFT;
554         if (numpages < cpa->numpages)
555                 cpa->numpages = numpages;
556
557         /*
558          * We are safe now. Check whether the new pgprot is the same:
559          * Convert protection attributes to 4k-format, as cpa->mask* are set
560          * up accordingly.
561          */
562         old_pte = *kpte;
563         req_prot = pgprot_large_2_4k(old_prot);
564
565         pgprot_val(req_prot) &= ~pgprot_val(cpa->mask_clr);
566         pgprot_val(req_prot) |= pgprot_val(cpa->mask_set);
567
568         /*
569          * req_prot is in format of 4k pages. It must be converted to large
570          * page format: the caching mode includes the PAT bit located at
571          * different bit positions in the two formats.
572          */
573         req_prot = pgprot_4k_2_large(req_prot);
574
575         /*
576          * Set the PSE and GLOBAL flags only if the PRESENT flag is
577          * set otherwise pmd_present/pmd_huge will return true even on
578          * a non present pmd. The canon_pgprot will clear _PAGE_GLOBAL
579          * for the ancient hardware that doesn't support it.
580          */
581         if (pgprot_val(req_prot) & _PAGE_PRESENT)
582                 pgprot_val(req_prot) |= _PAGE_PSE | _PAGE_GLOBAL;
583         else
584                 pgprot_val(req_prot) &= ~(_PAGE_PSE | _PAGE_GLOBAL);
585
586         req_prot = canon_pgprot(req_prot);
587
588         /*
589          * old_pfn points to the large page base pfn. So we need
590          * to add the offset of the virtual address:
591          */
592         pfn = old_pfn + ((address & (psize - 1)) >> PAGE_SHIFT);
593         cpa->pfn = pfn;
594
595         new_prot = static_protections(req_prot, address, pfn);
596
597         /*
598          * We need to check the full range, whether
599          * static_protection() requires a different pgprot for one of
600          * the pages in the range we try to preserve:
601          */
602         addr = address & pmask;
603         pfn = old_pfn;
604         for (i = 0; i < (psize >> PAGE_SHIFT); i++, addr += PAGE_SIZE, pfn++) {
605                 pgprot_t chk_prot = static_protections(req_prot, addr, pfn);
606
607                 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
608                         goto out_unlock;
609         }
610
611         /*
612          * If there are no changes, return. maxpages has been updated
613          * above:
614          */
615         if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
616                 do_split = 0;
617                 goto out_unlock;
618         }
619
620         /*
621          * We need to change the attributes. Check, whether we can
622          * change the large page in one go. We request a split, when
623          * the address is not aligned and the number of pages is
624          * smaller than the number of pages in the large page. Note
625          * that we limited the number of possible pages already to
626          * the number of pages in the large page.
627          */
628         if (address == (address & pmask) && cpa->numpages == (psize >> PAGE_SHIFT)) {
629                 /*
630                  * The address is aligned and the number of pages
631                  * covers the full page.
632                  */
633                 new_pte = pfn_pte(old_pfn, new_prot);
634                 __set_pmd_pte(kpte, address, new_pte);
635                 cpa->flags |= CPA_FLUSHTLB;
636                 do_split = 0;
637         }
638
639 out_unlock:
640         spin_unlock(&pgd_lock);
641
642         return do_split;
643 }
644
645 static int
646 __split_large_page(struct cpa_data *cpa, pte_t *kpte, unsigned long address,
647                    struct page *base)
648 {
649         pte_t *pbase = (pte_t *)page_address(base);
650         unsigned long ref_pfn, pfn, pfninc = 1;
651         unsigned int i, level;
652         pte_t *tmp;
653         pgprot_t ref_prot;
654
655         spin_lock(&pgd_lock);
656         /*
657          * Check for races, another CPU might have split this page
658          * up for us already:
659          */
660         tmp = _lookup_address_cpa(cpa, address, &level);
661         if (tmp != kpte) {
662                 spin_unlock(&pgd_lock);
663                 return 1;
664         }
665
666         paravirt_alloc_pte(&init_mm, page_to_pfn(base));
667
668         switch (level) {
669         case PG_LEVEL_2M:
670                 ref_prot = pmd_pgprot(*(pmd_t *)kpte);
671                 /* clear PSE and promote PAT bit to correct position */
672                 ref_prot = pgprot_large_2_4k(ref_prot);
673                 ref_pfn = pmd_pfn(*(pmd_t *)kpte);
674                 break;
675
676         case PG_LEVEL_1G:
677                 ref_prot = pud_pgprot(*(pud_t *)kpte);
678                 ref_pfn = pud_pfn(*(pud_t *)kpte);
679                 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
680
681                 /*
682                  * Clear the PSE flags if the PRESENT flag is not set
683                  * otherwise pmd_present/pmd_huge will return true
684                  * even on a non present pmd.
685                  */
686                 if (!(pgprot_val(ref_prot) & _PAGE_PRESENT))
687                         pgprot_val(ref_prot) &= ~_PAGE_PSE;
688                 break;
689
690         default:
691                 spin_unlock(&pgd_lock);
692                 return 1;
693         }
694
695         /*
696          * Set the GLOBAL flags only if the PRESENT flag is set
697          * otherwise pmd/pte_present will return true even on a non
698          * present pmd/pte. The canon_pgprot will clear _PAGE_GLOBAL
699          * for the ancient hardware that doesn't support it.
700          */
701         if (pgprot_val(ref_prot) & _PAGE_PRESENT)
702                 pgprot_val(ref_prot) |= _PAGE_GLOBAL;
703         else
704                 pgprot_val(ref_prot) &= ~_PAGE_GLOBAL;
705
706         /*
707          * Get the target pfn from the original entry:
708          */
709         pfn = ref_pfn;
710         for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
711                 set_pte(&pbase[i], pfn_pte(pfn, canon_pgprot(ref_prot)));
712
713         if (virt_addr_valid(address)) {
714                 unsigned long pfn = PFN_DOWN(__pa(address));
715
716                 if (pfn_range_is_mapped(pfn, pfn + 1))
717                         split_page_count(level);
718         }
719
720         /*
721          * Install the new, split up pagetable.
722          *
723          * We use the standard kernel pagetable protections for the new
724          * pagetable protections, the actual ptes set above control the
725          * primary protection behavior:
726          */
727         __set_pmd_pte(kpte, address, mk_pte(base, __pgprot(_KERNPG_TABLE)));
728
729         /*
730          * Intel Atom errata AAH41 workaround.
731          *
732          * The real fix should be in hw or in a microcode update, but
733          * we also probabilistically try to reduce the window of having
734          * a large TLB mixed with 4K TLBs while instruction fetches are
735          * going on.
736          */
737         __flush_tlb_all();
738         spin_unlock(&pgd_lock);
739
740         return 0;
741 }
742
743 static int split_large_page(struct cpa_data *cpa, pte_t *kpte,
744                             unsigned long address)
745 {
746         struct page *base;
747
748         if (!debug_pagealloc_enabled())
749                 spin_unlock(&cpa_lock);
750         base = alloc_pages(GFP_KERNEL | __GFP_NOTRACK, 0);
751         if (!debug_pagealloc_enabled())
752                 spin_lock(&cpa_lock);
753         if (!base)
754                 return -ENOMEM;
755
756         if (__split_large_page(cpa, kpte, address, base))
757                 __free_page(base);
758
759         return 0;
760 }
761
762 static bool try_to_free_pte_page(pte_t *pte)
763 {
764         int i;
765
766         for (i = 0; i < PTRS_PER_PTE; i++)
767                 if (!pte_none(pte[i]))
768                         return false;
769
770         free_page((unsigned long)pte);
771         return true;
772 }
773
774 static bool try_to_free_pmd_page(pmd_t *pmd)
775 {
776         int i;
777
778         for (i = 0; i < PTRS_PER_PMD; i++)
779                 if (!pmd_none(pmd[i]))
780                         return false;
781
782         free_page((unsigned long)pmd);
783         return true;
784 }
785
786 static bool unmap_pte_range(pmd_t *pmd, unsigned long start, unsigned long end)
787 {
788         pte_t *pte = pte_offset_kernel(pmd, start);
789
790         while (start < end) {
791                 set_pte(pte, __pte(0));
792
793                 start += PAGE_SIZE;
794                 pte++;
795         }
796
797         if (try_to_free_pte_page((pte_t *)pmd_page_vaddr(*pmd))) {
798                 pmd_clear(pmd);
799                 return true;
800         }
801         return false;
802 }
803
804 static void __unmap_pmd_range(pud_t *pud, pmd_t *pmd,
805                               unsigned long start, unsigned long end)
806 {
807         if (unmap_pte_range(pmd, start, end))
808                 if (try_to_free_pmd_page((pmd_t *)pud_page_vaddr(*pud)))
809                         pud_clear(pud);
810 }
811
812 static void unmap_pmd_range(pud_t *pud, unsigned long start, unsigned long end)
813 {
814         pmd_t *pmd = pmd_offset(pud, start);
815
816         /*
817          * Not on a 2MB page boundary?
818          */
819         if (start & (PMD_SIZE - 1)) {
820                 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
821                 unsigned long pre_end = min_t(unsigned long, end, next_page);
822
823                 __unmap_pmd_range(pud, pmd, start, pre_end);
824
825                 start = pre_end;
826                 pmd++;
827         }
828
829         /*
830          * Try to unmap in 2M chunks.
831          */
832         while (end - start >= PMD_SIZE) {
833                 if (pmd_large(*pmd))
834                         pmd_clear(pmd);
835                 else
836                         __unmap_pmd_range(pud, pmd, start, start + PMD_SIZE);
837
838                 start += PMD_SIZE;
839                 pmd++;
840         }
841
842         /*
843          * 4K leftovers?
844          */
845         if (start < end)
846                 return __unmap_pmd_range(pud, pmd, start, end);
847
848         /*
849          * Try again to free the PMD page if haven't succeeded above.
850          */
851         if (!pud_none(*pud))
852                 if (try_to_free_pmd_page((pmd_t *)pud_page_vaddr(*pud)))
853                         pud_clear(pud);
854 }
855
856 static void unmap_pud_range(p4d_t *p4d, unsigned long start, unsigned long end)
857 {
858         pud_t *pud = pud_offset(p4d, start);
859
860         /*
861          * Not on a GB page boundary?
862          */
863         if (start & (PUD_SIZE - 1)) {
864                 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
865                 unsigned long pre_end   = min_t(unsigned long, end, next_page);
866
867                 unmap_pmd_range(pud, start, pre_end);
868
869                 start = pre_end;
870                 pud++;
871         }
872
873         /*
874          * Try to unmap in 1G chunks?
875          */
876         while (end - start >= PUD_SIZE) {
877
878                 if (pud_large(*pud))
879                         pud_clear(pud);
880                 else
881                         unmap_pmd_range(pud, start, start + PUD_SIZE);
882
883                 start += PUD_SIZE;
884                 pud++;
885         }
886
887         /*
888          * 2M leftovers?
889          */
890         if (start < end)
891                 unmap_pmd_range(pud, start, end);
892
893         /*
894          * No need to try to free the PUD page because we'll free it in
895          * populate_pgd's error path
896          */
897 }
898
899 static int alloc_pte_page(pmd_t *pmd)
900 {
901         pte_t *pte = (pte_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
902         if (!pte)
903                 return -1;
904
905         set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
906         return 0;
907 }
908
909 static int alloc_pmd_page(pud_t *pud)
910 {
911         pmd_t *pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
912         if (!pmd)
913                 return -1;
914
915         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
916         return 0;
917 }
918
919 static void populate_pte(struct cpa_data *cpa,
920                          unsigned long start, unsigned long end,
921                          unsigned num_pages, pmd_t *pmd, pgprot_t pgprot)
922 {
923         pte_t *pte;
924
925         pte = pte_offset_kernel(pmd, start);
926
927         /*
928          * Set the GLOBAL flags only if the PRESENT flag is
929          * set otherwise pte_present will return true even on
930          * a non present pte. The canon_pgprot will clear
931          * _PAGE_GLOBAL for the ancient hardware that doesn't
932          * support it.
933          */
934         if (pgprot_val(pgprot) & _PAGE_PRESENT)
935                 pgprot_val(pgprot) |= _PAGE_GLOBAL;
936         else
937                 pgprot_val(pgprot) &= ~_PAGE_GLOBAL;
938
939         pgprot = canon_pgprot(pgprot);
940
941         while (num_pages-- && start < end) {
942                 set_pte(pte, pfn_pte(cpa->pfn, pgprot));
943
944                 start    += PAGE_SIZE;
945                 cpa->pfn++;
946                 pte++;
947         }
948 }
949
950 static long populate_pmd(struct cpa_data *cpa,
951                          unsigned long start, unsigned long end,
952                          unsigned num_pages, pud_t *pud, pgprot_t pgprot)
953 {
954         long cur_pages = 0;
955         pmd_t *pmd;
956         pgprot_t pmd_pgprot;
957
958         /*
959          * Not on a 2M boundary?
960          */
961         if (start & (PMD_SIZE - 1)) {
962                 unsigned long pre_end = start + (num_pages << PAGE_SHIFT);
963                 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
964
965                 pre_end   = min_t(unsigned long, pre_end, next_page);
966                 cur_pages = (pre_end - start) >> PAGE_SHIFT;
967                 cur_pages = min_t(unsigned int, num_pages, cur_pages);
968
969                 /*
970                  * Need a PTE page?
971                  */
972                 pmd = pmd_offset(pud, start);
973                 if (pmd_none(*pmd))
974                         if (alloc_pte_page(pmd))
975                                 return -1;
976
977                 populate_pte(cpa, start, pre_end, cur_pages, pmd, pgprot);
978
979                 start = pre_end;
980         }
981
982         /*
983          * We mapped them all?
984          */
985         if (num_pages == cur_pages)
986                 return cur_pages;
987
988         pmd_pgprot = pgprot_4k_2_large(pgprot);
989
990         while (end - start >= PMD_SIZE) {
991
992                 /*
993                  * We cannot use a 1G page so allocate a PMD page if needed.
994                  */
995                 if (pud_none(*pud))
996                         if (alloc_pmd_page(pud))
997                                 return -1;
998
999                 pmd = pmd_offset(pud, start);
1000
1001                 set_pmd(pmd, __pmd(cpa->pfn << PAGE_SHIFT | _PAGE_PSE |
1002                                    massage_pgprot(pmd_pgprot)));
1003
1004                 start     += PMD_SIZE;
1005                 cpa->pfn  += PMD_SIZE >> PAGE_SHIFT;
1006                 cur_pages += PMD_SIZE >> PAGE_SHIFT;
1007         }
1008
1009         /*
1010          * Map trailing 4K pages.
1011          */
1012         if (start < end) {
1013                 pmd = pmd_offset(pud, start);
1014                 if (pmd_none(*pmd))
1015                         if (alloc_pte_page(pmd))
1016                                 return -1;
1017
1018                 populate_pte(cpa, start, end, num_pages - cur_pages,
1019                              pmd, pgprot);
1020         }
1021         return num_pages;
1022 }
1023
1024 static int populate_pud(struct cpa_data *cpa, unsigned long start, p4d_t *p4d,
1025                         pgprot_t pgprot)
1026 {
1027         pud_t *pud;
1028         unsigned long end;
1029         long cur_pages = 0;
1030         pgprot_t pud_pgprot;
1031
1032         end = start + (cpa->numpages << PAGE_SHIFT);
1033
1034         /*
1035          * Not on a Gb page boundary? => map everything up to it with
1036          * smaller pages.
1037          */
1038         if (start & (PUD_SIZE - 1)) {
1039                 unsigned long pre_end;
1040                 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
1041
1042                 pre_end   = min_t(unsigned long, end, next_page);
1043                 cur_pages = (pre_end - start) >> PAGE_SHIFT;
1044                 cur_pages = min_t(int, (int)cpa->numpages, cur_pages);
1045
1046                 pud = pud_offset(p4d, start);
1047
1048                 /*
1049                  * Need a PMD page?
1050                  */
1051                 if (pud_none(*pud))
1052                         if (alloc_pmd_page(pud))
1053                                 return -1;
1054
1055                 cur_pages = populate_pmd(cpa, start, pre_end, cur_pages,
1056                                          pud, pgprot);
1057                 if (cur_pages < 0)
1058                         return cur_pages;
1059
1060                 start = pre_end;
1061         }
1062
1063         /* We mapped them all? */
1064         if (cpa->numpages == cur_pages)
1065                 return cur_pages;
1066
1067         pud = pud_offset(p4d, start);
1068         pud_pgprot = pgprot_4k_2_large(pgprot);
1069
1070         /*
1071          * Map everything starting from the Gb boundary, possibly with 1G pages
1072          */
1073         while (boot_cpu_has(X86_FEATURE_GBPAGES) && end - start >= PUD_SIZE) {
1074                 set_pud(pud, __pud(cpa->pfn << PAGE_SHIFT | _PAGE_PSE |
1075                                    massage_pgprot(pud_pgprot)));
1076
1077                 start     += PUD_SIZE;
1078                 cpa->pfn  += PUD_SIZE >> PAGE_SHIFT;
1079                 cur_pages += PUD_SIZE >> PAGE_SHIFT;
1080                 pud++;
1081         }
1082
1083         /* Map trailing leftover */
1084         if (start < end) {
1085                 long tmp;
1086
1087                 pud = pud_offset(p4d, start);
1088                 if (pud_none(*pud))
1089                         if (alloc_pmd_page(pud))
1090                                 return -1;
1091
1092                 tmp = populate_pmd(cpa, start, end, cpa->numpages - cur_pages,
1093                                    pud, pgprot);
1094                 if (tmp < 0)
1095                         return cur_pages;
1096
1097                 cur_pages += tmp;
1098         }
1099         return cur_pages;
1100 }
1101
1102 /*
1103  * Restrictions for kernel page table do not necessarily apply when mapping in
1104  * an alternate PGD.
1105  */
1106 static int populate_pgd(struct cpa_data *cpa, unsigned long addr)
1107 {
1108         pgprot_t pgprot = __pgprot(_KERNPG_TABLE);
1109         pud_t *pud = NULL;      /* shut up gcc */
1110         p4d_t *p4d;
1111         pgd_t *pgd_entry;
1112         long ret;
1113
1114         pgd_entry = cpa->pgd + pgd_index(addr);
1115
1116         if (pgd_none(*pgd_entry)) {
1117                 p4d = (p4d_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
1118                 if (!p4d)
1119                         return -1;
1120
1121                 set_pgd(pgd_entry, __pgd(__pa(p4d) | _KERNPG_TABLE));
1122         }
1123
1124         /*
1125          * Allocate a PUD page and hand it down for mapping.
1126          */
1127         p4d = p4d_offset(pgd_entry, addr);
1128         if (p4d_none(*p4d)) {
1129                 pud = (pud_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
1130                 if (!pud)
1131                         return -1;
1132
1133                 set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE));
1134         }
1135
1136         pgprot_val(pgprot) &= ~pgprot_val(cpa->mask_clr);
1137         pgprot_val(pgprot) |=  pgprot_val(cpa->mask_set);
1138
1139         ret = populate_pud(cpa, addr, p4d, pgprot);
1140         if (ret < 0) {
1141                 /*
1142                  * Leave the PUD page in place in case some other CPU or thread
1143                  * already found it, but remove any useless entries we just
1144                  * added to it.
1145                  */
1146                 unmap_pud_range(p4d, addr,
1147                                 addr + (cpa->numpages << PAGE_SHIFT));
1148                 return ret;
1149         }
1150
1151         cpa->numpages = ret;
1152         return 0;
1153 }
1154
1155 static int __cpa_process_fault(struct cpa_data *cpa, unsigned long vaddr,
1156                                int primary)
1157 {
1158         if (cpa->pgd) {
1159                 /*
1160                  * Right now, we only execute this code path when mapping
1161                  * the EFI virtual memory map regions, no other users
1162                  * provide a ->pgd value. This may change in the future.
1163                  */
1164                 return populate_pgd(cpa, vaddr);
1165         }
1166
1167         /*
1168          * Ignore all non primary paths.
1169          */
1170         if (!primary) {
1171                 cpa->numpages = 1;
1172                 return 0;
1173         }
1174
1175         /*
1176          * Ignore the NULL PTE for kernel identity mapping, as it is expected
1177          * to have holes.
1178          * Also set numpages to '1' indicating that we processed cpa req for
1179          * one virtual address page and its pfn. TBD: numpages can be set based
1180          * on the initial value and the level returned by lookup_address().
1181          */
1182         if (within(vaddr, PAGE_OFFSET,
1183                    PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
1184                 cpa->numpages = 1;
1185                 cpa->pfn = __pa(vaddr) >> PAGE_SHIFT;
1186                 return 0;
1187         } else {
1188                 WARN(1, KERN_WARNING "CPA: called for zero pte. "
1189                         "vaddr = %lx cpa->vaddr = %lx\n", vaddr,
1190                         *cpa->vaddr);
1191
1192                 return -EFAULT;
1193         }
1194 }
1195
1196 static int __change_page_attr(struct cpa_data *cpa, int primary)
1197 {
1198         unsigned long address;
1199         int do_split, err;
1200         unsigned int level;
1201         pte_t *kpte, old_pte;
1202
1203         if (cpa->flags & CPA_PAGES_ARRAY) {
1204                 struct page *page = cpa->pages[cpa->curpage];
1205                 if (unlikely(PageHighMem(page)))
1206                         return 0;
1207                 address = (unsigned long)page_address(page);
1208         } else if (cpa->flags & CPA_ARRAY)
1209                 address = cpa->vaddr[cpa->curpage];
1210         else
1211                 address = *cpa->vaddr;
1212 repeat:
1213         kpte = _lookup_address_cpa(cpa, address, &level);
1214         if (!kpte)
1215                 return __cpa_process_fault(cpa, address, primary);
1216
1217         old_pte = *kpte;
1218         if (pte_none(old_pte))
1219                 return __cpa_process_fault(cpa, address, primary);
1220
1221         if (level == PG_LEVEL_4K) {
1222                 pte_t new_pte;
1223                 pgprot_t new_prot = pte_pgprot(old_pte);
1224                 unsigned long pfn = pte_pfn(old_pte);
1225
1226                 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
1227                 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
1228
1229                 new_prot = static_protections(new_prot, address, pfn);
1230
1231                 /*
1232                  * Set the GLOBAL flags only if the PRESENT flag is
1233                  * set otherwise pte_present will return true even on
1234                  * a non present pte. The canon_pgprot will clear
1235                  * _PAGE_GLOBAL for the ancient hardware that doesn't
1236                  * support it.
1237                  */
1238                 if (pgprot_val(new_prot) & _PAGE_PRESENT)
1239                         pgprot_val(new_prot) |= _PAGE_GLOBAL;
1240                 else
1241                         pgprot_val(new_prot) &= ~_PAGE_GLOBAL;
1242
1243                 /*
1244                  * We need to keep the pfn from the existing PTE,
1245                  * after all we're only going to change it's attributes
1246                  * not the memory it points to
1247                  */
1248                 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
1249                 cpa->pfn = pfn;
1250                 /*
1251                  * Do we really change anything ?
1252                  */
1253                 if (pte_val(old_pte) != pte_val(new_pte)) {
1254                         set_pte_atomic(kpte, new_pte);
1255                         cpa->flags |= CPA_FLUSHTLB;
1256                 }
1257                 cpa->numpages = 1;
1258                 return 0;
1259         }
1260
1261         /*
1262          * Check, whether we can keep the large page intact
1263          * and just change the pte:
1264          */
1265         do_split = try_preserve_large_page(kpte, address, cpa);
1266         /*
1267          * When the range fits into the existing large page,
1268          * return. cp->numpages and cpa->tlbflush have been updated in
1269          * try_large_page:
1270          */
1271         if (do_split <= 0)
1272                 return do_split;
1273
1274         /*
1275          * We have to split the large page:
1276          */
1277         err = split_large_page(cpa, kpte, address);
1278         if (!err) {
1279                 /*
1280                  * Do a global flush tlb after splitting the large page
1281                  * and before we do the actual change page attribute in the PTE.
1282                  *
1283                  * With out this, we violate the TLB application note, that says
1284                  * "The TLBs may contain both ordinary and large-page
1285                  *  translations for a 4-KByte range of linear addresses. This
1286                  *  may occur if software modifies the paging structures so that
1287                  *  the page size used for the address range changes. If the two
1288                  *  translations differ with respect to page frame or attributes
1289                  *  (e.g., permissions), processor behavior is undefined and may
1290                  *  be implementation-specific."
1291                  *
1292                  * We do this global tlb flush inside the cpa_lock, so that we
1293                  * don't allow any other cpu, with stale tlb entries change the
1294                  * page attribute in parallel, that also falls into the
1295                  * just split large page entry.
1296                  */
1297                 flush_tlb_all();
1298                 goto repeat;
1299         }
1300
1301         return err;
1302 }
1303
1304 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
1305
1306 static int cpa_process_alias(struct cpa_data *cpa)
1307 {
1308         struct cpa_data alias_cpa;
1309         unsigned long laddr = (unsigned long)__va(cpa->pfn << PAGE_SHIFT);
1310         unsigned long vaddr;
1311         int ret;
1312
1313         if (!pfn_range_is_mapped(cpa->pfn, cpa->pfn + 1))
1314                 return 0;
1315
1316         /*
1317          * No need to redo, when the primary call touched the direct
1318          * mapping already:
1319          */
1320         if (cpa->flags & CPA_PAGES_ARRAY) {
1321                 struct page *page = cpa->pages[cpa->curpage];
1322                 if (unlikely(PageHighMem(page)))
1323                         return 0;
1324                 vaddr = (unsigned long)page_address(page);
1325         } else if (cpa->flags & CPA_ARRAY)
1326                 vaddr = cpa->vaddr[cpa->curpage];
1327         else
1328                 vaddr = *cpa->vaddr;
1329
1330         if (!(within(vaddr, PAGE_OFFSET,
1331                     PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT)))) {
1332
1333                 alias_cpa = *cpa;
1334                 alias_cpa.vaddr = &laddr;
1335                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1336
1337                 ret = __change_page_attr_set_clr(&alias_cpa, 0);
1338                 if (ret)
1339                         return ret;
1340         }
1341
1342 #ifdef CONFIG_X86_64
1343         /*
1344          * If the primary call didn't touch the high mapping already
1345          * and the physical address is inside the kernel map, we need
1346          * to touch the high mapped kernel as well:
1347          */
1348         if (!within(vaddr, (unsigned long)_text, _brk_end) &&
1349             within_inclusive(cpa->pfn, highmap_start_pfn(),
1350                              highmap_end_pfn())) {
1351                 unsigned long temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) +
1352                                                __START_KERNEL_map - phys_base;
1353                 alias_cpa = *cpa;
1354                 alias_cpa.vaddr = &temp_cpa_vaddr;
1355                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1356
1357                 /*
1358                  * The high mapping range is imprecise, so ignore the
1359                  * return value.
1360                  */
1361                 __change_page_attr_set_clr(&alias_cpa, 0);
1362         }
1363 #endif
1364
1365         return 0;
1366 }
1367
1368 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
1369 {
1370         unsigned long numpages = cpa->numpages;
1371         int ret;
1372
1373         while (numpages) {
1374                 /*
1375                  * Store the remaining nr of pages for the large page
1376                  * preservation check.
1377                  */
1378                 cpa->numpages = numpages;
1379                 /* for array changes, we can't use large page */
1380                 if (cpa->flags & (CPA_ARRAY | CPA_PAGES_ARRAY))
1381                         cpa->numpages = 1;
1382
1383                 if (!debug_pagealloc_enabled())
1384                         spin_lock(&cpa_lock);
1385                 ret = __change_page_attr(cpa, checkalias);
1386                 if (!debug_pagealloc_enabled())
1387                         spin_unlock(&cpa_lock);
1388                 if (ret)
1389                         return ret;
1390
1391                 if (checkalias) {
1392                         ret = cpa_process_alias(cpa);
1393                         if (ret)
1394                                 return ret;
1395                 }
1396
1397                 /*
1398                  * Adjust the number of pages with the result of the
1399                  * CPA operation. Either a large page has been
1400                  * preserved or a single page update happened.
1401                  */
1402                 BUG_ON(cpa->numpages > numpages || !cpa->numpages);
1403                 numpages -= cpa->numpages;
1404                 if (cpa->flags & (CPA_PAGES_ARRAY | CPA_ARRAY))
1405                         cpa->curpage++;
1406                 else
1407                         *cpa->vaddr += cpa->numpages * PAGE_SIZE;
1408
1409         }
1410         return 0;
1411 }
1412
1413 static int change_page_attr_set_clr(unsigned long *addr, int numpages,
1414                                     pgprot_t mask_set, pgprot_t mask_clr,
1415                                     int force_split, int in_flag,
1416                                     struct page **pages)
1417 {
1418         struct cpa_data cpa;
1419         int ret, cache, checkalias;
1420         unsigned long baddr = 0;
1421
1422         memset(&cpa, 0, sizeof(cpa));
1423
1424         /*
1425          * Check, if we are requested to change a not supported
1426          * feature:
1427          */
1428         mask_set = canon_pgprot(mask_set);
1429         mask_clr = canon_pgprot(mask_clr);
1430         if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
1431                 return 0;
1432
1433         /* Ensure we are PAGE_SIZE aligned */
1434         if (in_flag & CPA_ARRAY) {
1435                 int i;
1436                 for (i = 0; i < numpages; i++) {
1437                         if (addr[i] & ~PAGE_MASK) {
1438                                 addr[i] &= PAGE_MASK;
1439                                 WARN_ON_ONCE(1);
1440                         }
1441                 }
1442         } else if (!(in_flag & CPA_PAGES_ARRAY)) {
1443                 /*
1444                  * in_flag of CPA_PAGES_ARRAY implies it is aligned.
1445                  * No need to cehck in that case
1446                  */
1447                 if (*addr & ~PAGE_MASK) {
1448                         *addr &= PAGE_MASK;
1449                         /*
1450                          * People should not be passing in unaligned addresses:
1451                          */
1452                         WARN_ON_ONCE(1);
1453                 }
1454                 /*
1455                  * Save address for cache flush. *addr is modified in the call
1456                  * to __change_page_attr_set_clr() below.
1457                  */
1458                 baddr = *addr;
1459         }
1460
1461         /* Must avoid aliasing mappings in the highmem code */
1462         kmap_flush_unused();
1463
1464         vm_unmap_aliases();
1465
1466         cpa.vaddr = addr;
1467         cpa.pages = pages;
1468         cpa.numpages = numpages;
1469         cpa.mask_set = mask_set;
1470         cpa.mask_clr = mask_clr;
1471         cpa.flags = 0;
1472         cpa.curpage = 0;
1473         cpa.force_split = force_split;
1474
1475         if (in_flag & (CPA_ARRAY | CPA_PAGES_ARRAY))
1476                 cpa.flags |= in_flag;
1477
1478         /* No alias checking for _NX bit modifications */
1479         checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
1480
1481         ret = __change_page_attr_set_clr(&cpa, checkalias);
1482
1483         /*
1484          * Check whether we really changed something:
1485          */
1486         if (!(cpa.flags & CPA_FLUSHTLB))
1487                 goto out;
1488
1489         /*
1490          * No need to flush, when we did not set any of the caching
1491          * attributes:
1492          */
1493         cache = !!pgprot2cachemode(mask_set);
1494
1495         /*
1496          * On success we use CLFLUSH, when the CPU supports it to
1497          * avoid the WBINVD. If the CPU does not support it and in the
1498          * error case we fall back to cpa_flush_all (which uses
1499          * WBINVD):
1500          */
1501         if (!ret && boot_cpu_has(X86_FEATURE_CLFLUSH)) {
1502                 if (cpa.flags & (CPA_PAGES_ARRAY | CPA_ARRAY)) {
1503                         cpa_flush_array(addr, numpages, cache,
1504                                         cpa.flags, pages);
1505                 } else
1506                         cpa_flush_range(baddr, numpages, cache);
1507         } else
1508                 cpa_flush_all(cache);
1509
1510 out:
1511         return ret;
1512 }
1513
1514 static inline int change_page_attr_set(unsigned long *addr, int numpages,
1515                                        pgprot_t mask, int array)
1516 {
1517         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
1518                 (array ? CPA_ARRAY : 0), NULL);
1519 }
1520
1521 static inline int change_page_attr_clear(unsigned long *addr, int numpages,
1522                                          pgprot_t mask, int array)
1523 {
1524         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
1525                 (array ? CPA_ARRAY : 0), NULL);
1526 }
1527
1528 static inline int cpa_set_pages_array(struct page **pages, int numpages,
1529                                        pgprot_t mask)
1530 {
1531         return change_page_attr_set_clr(NULL, numpages, mask, __pgprot(0), 0,
1532                 CPA_PAGES_ARRAY, pages);
1533 }
1534
1535 static inline int cpa_clear_pages_array(struct page **pages, int numpages,
1536                                          pgprot_t mask)
1537 {
1538         return change_page_attr_set_clr(NULL, numpages, __pgprot(0), mask, 0,
1539                 CPA_PAGES_ARRAY, pages);
1540 }
1541
1542 int _set_memory_uc(unsigned long addr, int numpages)
1543 {
1544         /*
1545          * for now UC MINUS. see comments in ioremap_nocache()
1546          * If you really need strong UC use ioremap_uc(), but note
1547          * that you cannot override IO areas with set_memory_*() as
1548          * these helpers cannot work with IO memory.
1549          */
1550         return change_page_attr_set(&addr, numpages,
1551                                     cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1552                                     0);
1553 }
1554
1555 int set_memory_uc(unsigned long addr, int numpages)
1556 {
1557         int ret;
1558
1559         /*
1560          * for now UC MINUS. see comments in ioremap_nocache()
1561          */
1562         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1563                               _PAGE_CACHE_MODE_UC_MINUS, NULL);
1564         if (ret)
1565                 goto out_err;
1566
1567         ret = _set_memory_uc(addr, numpages);
1568         if (ret)
1569                 goto out_free;
1570
1571         return 0;
1572
1573 out_free:
1574         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1575 out_err:
1576         return ret;
1577 }
1578 EXPORT_SYMBOL(set_memory_uc);
1579
1580 static int _set_memory_array(unsigned long *addr, int addrinarray,
1581                 enum page_cache_mode new_type)
1582 {
1583         enum page_cache_mode set_type;
1584         int i, j;
1585         int ret;
1586
1587         for (i = 0; i < addrinarray; i++) {
1588                 ret = reserve_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE,
1589                                         new_type, NULL);
1590                 if (ret)
1591                         goto out_free;
1592         }
1593
1594         /* If WC, set to UC- first and then WC */
1595         set_type = (new_type == _PAGE_CACHE_MODE_WC) ?
1596                                 _PAGE_CACHE_MODE_UC_MINUS : new_type;
1597
1598         ret = change_page_attr_set(addr, addrinarray,
1599                                    cachemode2pgprot(set_type), 1);
1600
1601         if (!ret && new_type == _PAGE_CACHE_MODE_WC)
1602                 ret = change_page_attr_set_clr(addr, addrinarray,
1603                                                cachemode2pgprot(
1604                                                 _PAGE_CACHE_MODE_WC),
1605                                                __pgprot(_PAGE_CACHE_MASK),
1606                                                0, CPA_ARRAY, NULL);
1607         if (ret)
1608                 goto out_free;
1609
1610         return 0;
1611
1612 out_free:
1613         for (j = 0; j < i; j++)
1614                 free_memtype(__pa(addr[j]), __pa(addr[j]) + PAGE_SIZE);
1615
1616         return ret;
1617 }
1618
1619 int set_memory_array_uc(unsigned long *addr, int addrinarray)
1620 {
1621         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_MODE_UC_MINUS);
1622 }
1623 EXPORT_SYMBOL(set_memory_array_uc);
1624
1625 int set_memory_array_wc(unsigned long *addr, int addrinarray)
1626 {
1627         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_MODE_WC);
1628 }
1629 EXPORT_SYMBOL(set_memory_array_wc);
1630
1631 int set_memory_array_wt(unsigned long *addr, int addrinarray)
1632 {
1633         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_MODE_WT);
1634 }
1635 EXPORT_SYMBOL_GPL(set_memory_array_wt);
1636
1637 int _set_memory_wc(unsigned long addr, int numpages)
1638 {
1639         int ret;
1640         unsigned long addr_copy = addr;
1641
1642         ret = change_page_attr_set(&addr, numpages,
1643                                    cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1644                                    0);
1645         if (!ret) {
1646                 ret = change_page_attr_set_clr(&addr_copy, numpages,
1647                                                cachemode2pgprot(
1648                                                 _PAGE_CACHE_MODE_WC),
1649                                                __pgprot(_PAGE_CACHE_MASK),
1650                                                0, 0, NULL);
1651         }
1652         return ret;
1653 }
1654
1655 int set_memory_wc(unsigned long addr, int numpages)
1656 {
1657         int ret;
1658
1659         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1660                 _PAGE_CACHE_MODE_WC, NULL);
1661         if (ret)
1662                 return ret;
1663
1664         ret = _set_memory_wc(addr, numpages);
1665         if (ret)
1666                 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1667
1668         return ret;
1669 }
1670 EXPORT_SYMBOL(set_memory_wc);
1671
1672 int _set_memory_wt(unsigned long addr, int numpages)
1673 {
1674         return change_page_attr_set(&addr, numpages,
1675                                     cachemode2pgprot(_PAGE_CACHE_MODE_WT), 0);
1676 }
1677
1678 int set_memory_wt(unsigned long addr, int numpages)
1679 {
1680         int ret;
1681
1682         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1683                               _PAGE_CACHE_MODE_WT, NULL);
1684         if (ret)
1685                 return ret;
1686
1687         ret = _set_memory_wt(addr, numpages);
1688         if (ret)
1689                 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1690
1691         return ret;
1692 }
1693 EXPORT_SYMBOL_GPL(set_memory_wt);
1694
1695 int _set_memory_wb(unsigned long addr, int numpages)
1696 {
1697         /* WB cache mode is hard wired to all cache attribute bits being 0 */
1698         return change_page_attr_clear(&addr, numpages,
1699                                       __pgprot(_PAGE_CACHE_MASK), 0);
1700 }
1701
1702 int set_memory_wb(unsigned long addr, int numpages)
1703 {
1704         int ret;
1705
1706         ret = _set_memory_wb(addr, numpages);
1707         if (ret)
1708                 return ret;
1709
1710         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1711         return 0;
1712 }
1713 EXPORT_SYMBOL(set_memory_wb);
1714
1715 int set_memory_array_wb(unsigned long *addr, int addrinarray)
1716 {
1717         int i;
1718         int ret;
1719
1720         /* WB cache mode is hard wired to all cache attribute bits being 0 */
1721         ret = change_page_attr_clear(addr, addrinarray,
1722                                       __pgprot(_PAGE_CACHE_MASK), 1);
1723         if (ret)
1724                 return ret;
1725
1726         for (i = 0; i < addrinarray; i++)
1727                 free_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE);
1728
1729         return 0;
1730 }
1731 EXPORT_SYMBOL(set_memory_array_wb);
1732
1733 int set_memory_x(unsigned long addr, int numpages)
1734 {
1735         if (!(__supported_pte_mask & _PAGE_NX))
1736                 return 0;
1737
1738         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1739 }
1740 EXPORT_SYMBOL(set_memory_x);
1741
1742 int set_memory_nx(unsigned long addr, int numpages)
1743 {
1744         if (!(__supported_pte_mask & _PAGE_NX))
1745                 return 0;
1746
1747         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1748 }
1749 EXPORT_SYMBOL(set_memory_nx);
1750
1751 int set_memory_ro(unsigned long addr, int numpages)
1752 {
1753         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1754 }
1755
1756 int set_memory_rw(unsigned long addr, int numpages)
1757 {
1758         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
1759 }
1760
1761 int set_memory_np(unsigned long addr, int numpages)
1762 {
1763         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
1764 }
1765
1766 int set_memory_4k(unsigned long addr, int numpages)
1767 {
1768         return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1769                                         __pgprot(0), 1, 0, NULL);
1770 }
1771
1772 int set_pages_uc(struct page *page, int numpages)
1773 {
1774         unsigned long addr = (unsigned long)page_address(page);
1775
1776         return set_memory_uc(addr, numpages);
1777 }
1778 EXPORT_SYMBOL(set_pages_uc);
1779
1780 static int _set_pages_array(struct page **pages, int addrinarray,
1781                 enum page_cache_mode new_type)
1782 {
1783         unsigned long start;
1784         unsigned long end;
1785         enum page_cache_mode set_type;
1786         int i;
1787         int free_idx;
1788         int ret;
1789
1790         for (i = 0; i < addrinarray; i++) {
1791                 if (PageHighMem(pages[i]))
1792                         continue;
1793                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1794                 end = start + PAGE_SIZE;
1795                 if (reserve_memtype(start, end, new_type, NULL))
1796                         goto err_out;
1797         }
1798
1799         /* If WC, set to UC- first and then WC */
1800         set_type = (new_type == _PAGE_CACHE_MODE_WC) ?
1801                                 _PAGE_CACHE_MODE_UC_MINUS : new_type;
1802
1803         ret = cpa_set_pages_array(pages, addrinarray,
1804                                   cachemode2pgprot(set_type));
1805         if (!ret && new_type == _PAGE_CACHE_MODE_WC)
1806                 ret = change_page_attr_set_clr(NULL, addrinarray,
1807                                                cachemode2pgprot(
1808                                                 _PAGE_CACHE_MODE_WC),
1809                                                __pgprot(_PAGE_CACHE_MASK),
1810                                                0, CPA_PAGES_ARRAY, pages);
1811         if (ret)
1812                 goto err_out;
1813         return 0; /* Success */
1814 err_out:
1815         free_idx = i;
1816         for (i = 0; i < free_idx; i++) {
1817                 if (PageHighMem(pages[i]))
1818                         continue;
1819                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1820                 end = start + PAGE_SIZE;
1821                 free_memtype(start, end);
1822         }
1823         return -EINVAL;
1824 }
1825
1826 int set_pages_array_uc(struct page **pages, int addrinarray)
1827 {
1828         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_MODE_UC_MINUS);
1829 }
1830 EXPORT_SYMBOL(set_pages_array_uc);
1831
1832 int set_pages_array_wc(struct page **pages, int addrinarray)
1833 {
1834         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_MODE_WC);
1835 }
1836 EXPORT_SYMBOL(set_pages_array_wc);
1837
1838 int set_pages_array_wt(struct page **pages, int addrinarray)
1839 {
1840         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_MODE_WT);
1841 }
1842 EXPORT_SYMBOL_GPL(set_pages_array_wt);
1843
1844 int set_pages_wb(struct page *page, int numpages)
1845 {
1846         unsigned long addr = (unsigned long)page_address(page);
1847
1848         return set_memory_wb(addr, numpages);
1849 }
1850 EXPORT_SYMBOL(set_pages_wb);
1851
1852 int set_pages_array_wb(struct page **pages, int addrinarray)
1853 {
1854         int retval;
1855         unsigned long start;
1856         unsigned long end;
1857         int i;
1858
1859         /* WB cache mode is hard wired to all cache attribute bits being 0 */
1860         retval = cpa_clear_pages_array(pages, addrinarray,
1861                         __pgprot(_PAGE_CACHE_MASK));
1862         if (retval)
1863                 return retval;
1864
1865         for (i = 0; i < addrinarray; i++) {
1866                 if (PageHighMem(pages[i]))
1867                         continue;
1868                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1869                 end = start + PAGE_SIZE;
1870                 free_memtype(start, end);
1871         }
1872
1873         return 0;
1874 }
1875 EXPORT_SYMBOL(set_pages_array_wb);
1876
1877 int set_pages_x(struct page *page, int numpages)
1878 {
1879         unsigned long addr = (unsigned long)page_address(page);
1880
1881         return set_memory_x(addr, numpages);
1882 }
1883 EXPORT_SYMBOL(set_pages_x);
1884
1885 int set_pages_nx(struct page *page, int numpages)
1886 {
1887         unsigned long addr = (unsigned long)page_address(page);
1888
1889         return set_memory_nx(addr, numpages);
1890 }
1891 EXPORT_SYMBOL(set_pages_nx);
1892
1893 int set_pages_ro(struct page *page, int numpages)
1894 {
1895         unsigned long addr = (unsigned long)page_address(page);
1896
1897         return set_memory_ro(addr, numpages);
1898 }
1899
1900 int set_pages_rw(struct page *page, int numpages)
1901 {
1902         unsigned long addr = (unsigned long)page_address(page);
1903
1904         return set_memory_rw(addr, numpages);
1905 }
1906
1907 #ifdef CONFIG_DEBUG_PAGEALLOC
1908
1909 static int __set_pages_p(struct page *page, int numpages)
1910 {
1911         unsigned long tempaddr = (unsigned long) page_address(page);
1912         struct cpa_data cpa = { .vaddr = &tempaddr,
1913                                 .pgd = NULL,
1914                                 .numpages = numpages,
1915                                 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1916                                 .mask_clr = __pgprot(0),
1917                                 .flags = 0};
1918
1919         /*
1920          * No alias checking needed for setting present flag. otherwise,
1921          * we may need to break large pages for 64-bit kernel text
1922          * mappings (this adds to complexity if we want to do this from
1923          * atomic context especially). Let's keep it simple!
1924          */
1925         return __change_page_attr_set_clr(&cpa, 0);
1926 }
1927
1928 static int __set_pages_np(struct page *page, int numpages)
1929 {
1930         unsigned long tempaddr = (unsigned long) page_address(page);
1931         struct cpa_data cpa = { .vaddr = &tempaddr,
1932                                 .pgd = NULL,
1933                                 .numpages = numpages,
1934                                 .mask_set = __pgprot(0),
1935                                 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1936                                 .flags = 0};
1937
1938         /*
1939          * No alias checking needed for setting not present flag. otherwise,
1940          * we may need to break large pages for 64-bit kernel text
1941          * mappings (this adds to complexity if we want to do this from
1942          * atomic context especially). Let's keep it simple!
1943          */
1944         return __change_page_attr_set_clr(&cpa, 0);
1945 }
1946
1947 void __kernel_map_pages(struct page *page, int numpages, int enable)
1948 {
1949         if (PageHighMem(page))
1950                 return;
1951         if (!enable) {
1952                 debug_check_no_locks_freed(page_address(page),
1953                                            numpages * PAGE_SIZE);
1954         }
1955
1956         /*
1957          * The return value is ignored as the calls cannot fail.
1958          * Large pages for identity mappings are not used at boot time
1959          * and hence no memory allocations during large page split.
1960          */
1961         if (enable)
1962                 __set_pages_p(page, numpages);
1963         else
1964                 __set_pages_np(page, numpages);
1965
1966         /*
1967          * We should perform an IPI and flush all tlbs,
1968          * but that can deadlock->flush only current cpu:
1969          */
1970         __flush_tlb_all();
1971
1972         arch_flush_lazy_mmu_mode();
1973 }
1974
1975 #ifdef CONFIG_HIBERNATION
1976
1977 bool kernel_page_present(struct page *page)
1978 {
1979         unsigned int level;
1980         pte_t *pte;
1981
1982         if (PageHighMem(page))
1983                 return false;
1984
1985         pte = lookup_address((unsigned long)page_address(page), &level);
1986         return (pte_val(*pte) & _PAGE_PRESENT);
1987 }
1988
1989 #endif /* CONFIG_HIBERNATION */
1990
1991 #endif /* CONFIG_DEBUG_PAGEALLOC */
1992
1993 int kernel_map_pages_in_pgd(pgd_t *pgd, u64 pfn, unsigned long address,
1994                             unsigned numpages, unsigned long page_flags)
1995 {
1996         int retval = -EINVAL;
1997
1998         struct cpa_data cpa = {
1999                 .vaddr = &address,
2000                 .pfn = pfn,
2001                 .pgd = pgd,
2002                 .numpages = numpages,
2003                 .mask_set = __pgprot(0),
2004                 .mask_clr = __pgprot(0),
2005                 .flags = 0,
2006         };
2007
2008         if (!(__supported_pte_mask & _PAGE_NX))
2009                 goto out;
2010
2011         if (!(page_flags & _PAGE_NX))
2012                 cpa.mask_clr = __pgprot(_PAGE_NX);
2013
2014         if (!(page_flags & _PAGE_RW))
2015                 cpa.mask_clr = __pgprot(_PAGE_RW);
2016
2017         cpa.mask_set = __pgprot(_PAGE_PRESENT | page_flags);
2018
2019         retval = __change_page_attr_set_clr(&cpa, 0);
2020         __flush_tlb_all();
2021
2022 out:
2023         return retval;
2024 }
2025
2026 /*
2027  * The testcases use internal knowledge of the implementation that shouldn't
2028  * be exposed to the rest of the kernel. Include these directly here.
2029  */
2030 #ifdef CONFIG_CPA_DEBUG
2031 #include "pageattr-test.c"
2032 #endif