x86, mm: Use new pagetable helpers in try_preserve_large_page()
[firefly-linux-kernel-4.4.55.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         /* Protect against CPA */
61         spin_lock(&pgd_lock);
62         direct_pages_count[level] += pages;
63         spin_unlock(&pgd_lock);
64 }
65
66 static void split_page_count(int level)
67 {
68         direct_pages_count[level]--;
69         direct_pages_count[level - 1] += PTRS_PER_PTE;
70 }
71
72 void arch_report_meminfo(struct seq_file *m)
73 {
74         seq_printf(m, "DirectMap4k:    %8lu kB\n",
75                         direct_pages_count[PG_LEVEL_4K] << 2);
76 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
77         seq_printf(m, "DirectMap2M:    %8lu kB\n",
78                         direct_pages_count[PG_LEVEL_2M] << 11);
79 #else
80         seq_printf(m, "DirectMap4M:    %8lu kB\n",
81                         direct_pages_count[PG_LEVEL_2M] << 12);
82 #endif
83 #ifdef CONFIG_X86_64
84         if (direct_gbpages)
85                 seq_printf(m, "DirectMap1G:    %8lu kB\n",
86                         direct_pages_count[PG_LEVEL_1G] << 20);
87 #endif
88 }
89 #else
90 static inline void split_page_count(int level) { }
91 #endif
92
93 #ifdef CONFIG_X86_64
94
95 static inline unsigned long highmap_start_pfn(void)
96 {
97         return __pa_symbol(_text) >> PAGE_SHIFT;
98 }
99
100 static inline unsigned long highmap_end_pfn(void)
101 {
102         return __pa_symbol(roundup(_brk_end, PMD_SIZE)) >> PAGE_SHIFT;
103 }
104
105 #endif
106
107 #ifdef CONFIG_DEBUG_PAGEALLOC
108 # define debug_pagealloc 1
109 #else
110 # define debug_pagealloc 0
111 #endif
112
113 static inline int
114 within(unsigned long addr, unsigned long start, unsigned long end)
115 {
116         return addr >= start && addr < end;
117 }
118
119 /*
120  * Flushing functions
121  */
122
123 /**
124  * clflush_cache_range - flush a cache range with clflush
125  * @vaddr:      virtual start address
126  * @size:       number of bytes to flush
127  *
128  * clflush is an unordered instruction which needs fencing with mfence
129  * to avoid ordering issues.
130  */
131 void clflush_cache_range(void *vaddr, unsigned int size)
132 {
133         void *vend = vaddr + size - 1;
134
135         mb();
136
137         for (; vaddr < vend; vaddr += boot_cpu_data.x86_clflush_size)
138                 clflush(vaddr);
139         /*
140          * Flush any possible final partial cacheline:
141          */
142         clflush(vend);
143
144         mb();
145 }
146 EXPORT_SYMBOL_GPL(clflush_cache_range);
147
148 static void __cpa_flush_all(void *arg)
149 {
150         unsigned long cache = (unsigned long)arg;
151
152         /*
153          * Flush all to work around Errata in early athlons regarding
154          * large page flushing.
155          */
156         __flush_tlb_all();
157
158         if (cache && boot_cpu_data.x86 >= 4)
159                 wbinvd();
160 }
161
162 static void cpa_flush_all(unsigned long cache)
163 {
164         BUG_ON(irqs_disabled());
165
166         on_each_cpu(__cpa_flush_all, (void *) cache, 1);
167 }
168
169 static void __cpa_flush_range(void *arg)
170 {
171         /*
172          * We could optimize that further and do individual per page
173          * tlb invalidates for a low number of pages. Caveat: we must
174          * flush the high aliases on 64bit as well.
175          */
176         __flush_tlb_all();
177 }
178
179 static void cpa_flush_range(unsigned long start, int numpages, int cache)
180 {
181         unsigned int i, level;
182         unsigned long addr;
183
184         BUG_ON(irqs_disabled());
185         WARN_ON(PAGE_ALIGN(start) != start);
186
187         on_each_cpu(__cpa_flush_range, NULL, 1);
188
189         if (!cache)
190                 return;
191
192         /*
193          * We only need to flush on one CPU,
194          * clflush is a MESI-coherent instruction that
195          * will cause all other CPUs to flush the same
196          * cachelines:
197          */
198         for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
199                 pte_t *pte = lookup_address(addr, &level);
200
201                 /*
202                  * Only flush present addresses:
203                  */
204                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
205                         clflush_cache_range((void *) addr, PAGE_SIZE);
206         }
207 }
208
209 static void cpa_flush_array(unsigned long *start, int numpages, int cache,
210                             int in_flags, struct page **pages)
211 {
212         unsigned int i, level;
213         unsigned long do_wbinvd = cache && numpages >= 1024; /* 4M threshold */
214
215         BUG_ON(irqs_disabled());
216
217         on_each_cpu(__cpa_flush_all, (void *) do_wbinvd, 1);
218
219         if (!cache || do_wbinvd)
220                 return;
221
222         /*
223          * We only need to flush on one CPU,
224          * clflush is a MESI-coherent instruction that
225          * will cause all other CPUs to flush the same
226          * cachelines:
227          */
228         for (i = 0; i < numpages; i++) {
229                 unsigned long addr;
230                 pte_t *pte;
231
232                 if (in_flags & CPA_PAGES_ARRAY)
233                         addr = (unsigned long)page_address(pages[i]);
234                 else
235                         addr = start[i];
236
237                 pte = lookup_address(addr, &level);
238
239                 /*
240                  * Only flush present addresses:
241                  */
242                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
243                         clflush_cache_range((void *)addr, PAGE_SIZE);
244         }
245 }
246
247 /*
248  * Certain areas of memory on x86 require very specific protection flags,
249  * for example the BIOS area or kernel text. Callers don't always get this
250  * right (again, ioremap() on BIOS memory is not uncommon) so this function
251  * checks and fixes these known static required protection bits.
252  */
253 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
254                                    unsigned long pfn)
255 {
256         pgprot_t forbidden = __pgprot(0);
257
258         /*
259          * The BIOS area between 640k and 1Mb needs to be executable for
260          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
261          */
262 #ifdef CONFIG_PCI_BIOS
263         if (pcibios_enabled && within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
264                 pgprot_val(forbidden) |= _PAGE_NX;
265 #endif
266
267         /*
268          * The kernel text needs to be executable for obvious reasons
269          * Does not cover __inittext since that is gone later on. On
270          * 64bit we do not enforce !NX on the low mapping
271          */
272         if (within(address, (unsigned long)_text, (unsigned long)_etext))
273                 pgprot_val(forbidden) |= _PAGE_NX;
274
275         /*
276          * The .rodata section needs to be read-only. Using the pfn
277          * catches all aliases.
278          */
279         if (within(pfn, __pa_symbol(__start_rodata) >> PAGE_SHIFT,
280                    __pa_symbol(__end_rodata) >> PAGE_SHIFT))
281                 pgprot_val(forbidden) |= _PAGE_RW;
282
283 #if defined(CONFIG_X86_64) && defined(CONFIG_DEBUG_RODATA)
284         /*
285          * Once the kernel maps the text as RO (kernel_set_to_readonly is set),
286          * kernel text mappings for the large page aligned text, rodata sections
287          * will be always read-only. For the kernel identity mappings covering
288          * the holes caused by this alignment can be anything that user asks.
289          *
290          * This will preserve the large page mappings for kernel text/data
291          * at no extra cost.
292          */
293         if (kernel_set_to_readonly &&
294             within(address, (unsigned long)_text,
295                    (unsigned long)__end_rodata_hpage_align)) {
296                 unsigned int level;
297
298                 /*
299                  * Don't enforce the !RW mapping for the kernel text mapping,
300                  * if the current mapping is already using small page mapping.
301                  * No need to work hard to preserve large page mappings in this
302                  * case.
303                  *
304                  * This also fixes the Linux Xen paravirt guest boot failure
305                  * (because of unexpected read-only mappings for kernel identity
306                  * mappings). In this paravirt guest case, the kernel text
307                  * mapping and the kernel identity mapping share the same
308                  * page-table pages. Thus we can't really use different
309                  * protections for the kernel text and identity mappings. Also,
310                  * these shared mappings are made of small page mappings.
311                  * Thus this don't enforce !RW mapping for small page kernel
312                  * text mapping logic will help Linux Xen parvirt guest boot
313                  * as well.
314                  */
315                 if (lookup_address(address, &level) && (level != PG_LEVEL_4K))
316                         pgprot_val(forbidden) |= _PAGE_RW;
317         }
318 #endif
319
320         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
321
322         return prot;
323 }
324
325 /*
326  * Lookup the page table entry for a virtual address. Return a pointer
327  * to the entry and the level of the mapping.
328  *
329  * Note: We return pud and pmd either when the entry is marked large
330  * or when the present bit is not set. Otherwise we would return a
331  * pointer to a nonexisting mapping.
332  */
333 pte_t *lookup_address(unsigned long address, unsigned int *level)
334 {
335         pgd_t *pgd = pgd_offset_k(address);
336         pud_t *pud;
337         pmd_t *pmd;
338
339         *level = PG_LEVEL_NONE;
340
341         if (pgd_none(*pgd))
342                 return NULL;
343
344         pud = pud_offset(pgd, address);
345         if (pud_none(*pud))
346                 return NULL;
347
348         *level = PG_LEVEL_1G;
349         if (pud_large(*pud) || !pud_present(*pud))
350                 return (pte_t *)pud;
351
352         pmd = pmd_offset(pud, address);
353         if (pmd_none(*pmd))
354                 return NULL;
355
356         *level = PG_LEVEL_2M;
357         if (pmd_large(*pmd) || !pmd_present(*pmd))
358                 return (pte_t *)pmd;
359
360         *level = PG_LEVEL_4K;
361
362         return pte_offset_kernel(pmd, address);
363 }
364 EXPORT_SYMBOL_GPL(lookup_address);
365
366 /*
367  * Set the new pmd in all the pgds we know about:
368  */
369 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
370 {
371         /* change init_mm */
372         set_pte_atomic(kpte, pte);
373 #ifdef CONFIG_X86_32
374         if (!SHARED_KERNEL_PMD) {
375                 struct page *page;
376
377                 list_for_each_entry(page, &pgd_list, lru) {
378                         pgd_t *pgd;
379                         pud_t *pud;
380                         pmd_t *pmd;
381
382                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
383                         pud = pud_offset(pgd, address);
384                         pmd = pmd_offset(pud, address);
385                         set_pte_atomic((pte_t *)pmd, pte);
386                 }
387         }
388 #endif
389 }
390
391 static int
392 try_preserve_large_page(pte_t *kpte, unsigned long address,
393                         struct cpa_data *cpa)
394 {
395         unsigned long nextpage_addr, numpages, pmask, psize, addr, pfn;
396         pte_t new_pte, old_pte, *tmp;
397         pgprot_t old_prot, new_prot, req_prot;
398         int i, do_split = 1;
399         enum pg_level level;
400
401         if (cpa->force_split)
402                 return 1;
403
404         spin_lock(&pgd_lock);
405         /*
406          * Check for races, another CPU might have split this page
407          * up already:
408          */
409         tmp = lookup_address(address, &level);
410         if (tmp != kpte)
411                 goto out_unlock;
412
413         switch (level) {
414         case PG_LEVEL_2M:
415 #ifdef CONFIG_X86_64
416         case PG_LEVEL_1G:
417 #endif
418                 psize = page_level_size(level);
419                 pmask = page_level_mask(level);
420                 break;
421         default:
422                 do_split = -EINVAL;
423                 goto out_unlock;
424         }
425
426         /*
427          * Calculate the number of pages, which fit into this large
428          * page starting at address:
429          */
430         nextpage_addr = (address + psize) & pmask;
431         numpages = (nextpage_addr - address) >> PAGE_SHIFT;
432         if (numpages < cpa->numpages)
433                 cpa->numpages = numpages;
434
435         /*
436          * We are safe now. Check whether the new pgprot is the same:
437          */
438         old_pte = *kpte;
439         old_prot = new_prot = req_prot = pte_pgprot(old_pte);
440
441         pgprot_val(req_prot) &= ~pgprot_val(cpa->mask_clr);
442         pgprot_val(req_prot) |= pgprot_val(cpa->mask_set);
443
444         /*
445          * old_pte points to the large page base address. So we need
446          * to add the offset of the virtual address:
447          */
448         pfn = pte_pfn(old_pte) + ((address & (psize - 1)) >> PAGE_SHIFT);
449         cpa->pfn = pfn;
450
451         new_prot = static_protections(req_prot, address, pfn);
452
453         /*
454          * We need to check the full range, whether
455          * static_protection() requires a different pgprot for one of
456          * the pages in the range we try to preserve:
457          */
458         addr = address & pmask;
459         pfn = pte_pfn(old_pte);
460         for (i = 0; i < (psize >> PAGE_SHIFT); i++, addr += PAGE_SIZE, pfn++) {
461                 pgprot_t chk_prot = static_protections(req_prot, addr, pfn);
462
463                 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
464                         goto out_unlock;
465         }
466
467         /*
468          * If there are no changes, return. maxpages has been updated
469          * above:
470          */
471         if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
472                 do_split = 0;
473                 goto out_unlock;
474         }
475
476         /*
477          * We need to change the attributes. Check, whether we can
478          * change the large page in one go. We request a split, when
479          * the address is not aligned and the number of pages is
480          * smaller than the number of pages in the large page. Note
481          * that we limited the number of possible pages already to
482          * the number of pages in the large page.
483          */
484         if (address == (address & pmask) && cpa->numpages == (psize >> PAGE_SHIFT)) {
485                 /*
486                  * The address is aligned and the number of pages
487                  * covers the full page.
488                  */
489                 new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
490                 __set_pmd_pte(kpte, address, new_pte);
491                 cpa->flags |= CPA_FLUSHTLB;
492                 do_split = 0;
493         }
494
495 out_unlock:
496         spin_unlock(&pgd_lock);
497
498         return do_split;
499 }
500
501 static int split_large_page(pte_t *kpte, unsigned long address)
502 {
503         unsigned long pfn, pfninc = 1;
504         unsigned int i, level;
505         pte_t *pbase, *tmp;
506         pgprot_t ref_prot;
507         struct page *base;
508
509         if (!debug_pagealloc)
510                 spin_unlock(&cpa_lock);
511         base = alloc_pages(GFP_KERNEL | __GFP_NOTRACK, 0);
512         if (!debug_pagealloc)
513                 spin_lock(&cpa_lock);
514         if (!base)
515                 return -ENOMEM;
516
517         spin_lock(&pgd_lock);
518         /*
519          * Check for races, another CPU might have split this page
520          * up for us already:
521          */
522         tmp = lookup_address(address, &level);
523         if (tmp != kpte)
524                 goto out_unlock;
525
526         pbase = (pte_t *)page_address(base);
527         paravirt_alloc_pte(&init_mm, page_to_pfn(base));
528         ref_prot = pte_pgprot(pte_clrhuge(*kpte));
529         /*
530          * If we ever want to utilize the PAT bit, we need to
531          * update this function to make sure it's converted from
532          * bit 12 to bit 7 when we cross from the 2MB level to
533          * the 4K level:
534          */
535         WARN_ON_ONCE(pgprot_val(ref_prot) & _PAGE_PAT_LARGE);
536
537 #ifdef CONFIG_X86_64
538         if (level == PG_LEVEL_1G) {
539                 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
540                 pgprot_val(ref_prot) |= _PAGE_PSE;
541         }
542 #endif
543
544         /*
545          * Get the target pfn from the original entry:
546          */
547         pfn = pte_pfn(*kpte);
548         for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
549                 set_pte(&pbase[i], pfn_pte(pfn, ref_prot));
550
551         if (address >= (unsigned long)__va(0) &&
552                 address < (unsigned long)__va(max_low_pfn_mapped << PAGE_SHIFT))
553                 split_page_count(level);
554
555 #ifdef CONFIG_X86_64
556         if (address >= (unsigned long)__va(1UL<<32) &&
557                 address < (unsigned long)__va(max_pfn_mapped << PAGE_SHIFT))
558                 split_page_count(level);
559 #endif
560
561         /*
562          * Install the new, split up pagetable.
563          *
564          * We use the standard kernel pagetable protections for the new
565          * pagetable protections, the actual ptes set above control the
566          * primary protection behavior:
567          */
568         __set_pmd_pte(kpte, address, mk_pte(base, __pgprot(_KERNPG_TABLE)));
569
570         /*
571          * Intel Atom errata AAH41 workaround.
572          *
573          * The real fix should be in hw or in a microcode update, but
574          * we also probabilistically try to reduce the window of having
575          * a large TLB mixed with 4K TLBs while instruction fetches are
576          * going on.
577          */
578         __flush_tlb_all();
579
580         base = NULL;
581
582 out_unlock:
583         /*
584          * If we dropped out via the lookup_address check under
585          * pgd_lock then stick the page back into the pool:
586          */
587         if (base)
588                 __free_page(base);
589         spin_unlock(&pgd_lock);
590
591         return 0;
592 }
593
594 static int __cpa_process_fault(struct cpa_data *cpa, unsigned long vaddr,
595                                int primary)
596 {
597         /*
598          * Ignore all non primary paths.
599          */
600         if (!primary)
601                 return 0;
602
603         /*
604          * Ignore the NULL PTE for kernel identity mapping, as it is expected
605          * to have holes.
606          * Also set numpages to '1' indicating that we processed cpa req for
607          * one virtual address page and its pfn. TBD: numpages can be set based
608          * on the initial value and the level returned by lookup_address().
609          */
610         if (within(vaddr, PAGE_OFFSET,
611                    PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
612                 cpa->numpages = 1;
613                 cpa->pfn = __pa(vaddr) >> PAGE_SHIFT;
614                 return 0;
615         } else {
616                 WARN(1, KERN_WARNING "CPA: called for zero pte. "
617                         "vaddr = %lx cpa->vaddr = %lx\n", vaddr,
618                         *cpa->vaddr);
619
620                 return -EFAULT;
621         }
622 }
623
624 static int __change_page_attr(struct cpa_data *cpa, int primary)
625 {
626         unsigned long address;
627         int do_split, err;
628         unsigned int level;
629         pte_t *kpte, old_pte;
630
631         if (cpa->flags & CPA_PAGES_ARRAY) {
632                 struct page *page = cpa->pages[cpa->curpage];
633                 if (unlikely(PageHighMem(page)))
634                         return 0;
635                 address = (unsigned long)page_address(page);
636         } else if (cpa->flags & CPA_ARRAY)
637                 address = cpa->vaddr[cpa->curpage];
638         else
639                 address = *cpa->vaddr;
640 repeat:
641         kpte = lookup_address(address, &level);
642         if (!kpte)
643                 return __cpa_process_fault(cpa, address, primary);
644
645         old_pte = *kpte;
646         if (!pte_val(old_pte))
647                 return __cpa_process_fault(cpa, address, primary);
648
649         if (level == PG_LEVEL_4K) {
650                 pte_t new_pte;
651                 pgprot_t new_prot = pte_pgprot(old_pte);
652                 unsigned long pfn = pte_pfn(old_pte);
653
654                 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
655                 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
656
657                 new_prot = static_protections(new_prot, address, pfn);
658
659                 /*
660                  * We need to keep the pfn from the existing PTE,
661                  * after all we're only going to change it's attributes
662                  * not the memory it points to
663                  */
664                 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
665                 cpa->pfn = pfn;
666                 /*
667                  * Do we really change anything ?
668                  */
669                 if (pte_val(old_pte) != pte_val(new_pte)) {
670                         set_pte_atomic(kpte, new_pte);
671                         cpa->flags |= CPA_FLUSHTLB;
672                 }
673                 cpa->numpages = 1;
674                 return 0;
675         }
676
677         /*
678          * Check, whether we can keep the large page intact
679          * and just change the pte:
680          */
681         do_split = try_preserve_large_page(kpte, address, cpa);
682         /*
683          * When the range fits into the existing large page,
684          * return. cp->numpages and cpa->tlbflush have been updated in
685          * try_large_page:
686          */
687         if (do_split <= 0)
688                 return do_split;
689
690         /*
691          * We have to split the large page:
692          */
693         err = split_large_page(kpte, address);
694         if (!err) {
695                 /*
696                  * Do a global flush tlb after splitting the large page
697                  * and before we do the actual change page attribute in the PTE.
698                  *
699                  * With out this, we violate the TLB application note, that says
700                  * "The TLBs may contain both ordinary and large-page
701                  *  translations for a 4-KByte range of linear addresses. This
702                  *  may occur if software modifies the paging structures so that
703                  *  the page size used for the address range changes. If the two
704                  *  translations differ with respect to page frame or attributes
705                  *  (e.g., permissions), processor behavior is undefined and may
706                  *  be implementation-specific."
707                  *
708                  * We do this global tlb flush inside the cpa_lock, so that we
709                  * don't allow any other cpu, with stale tlb entries change the
710                  * page attribute in parallel, that also falls into the
711                  * just split large page entry.
712                  */
713                 flush_tlb_all();
714                 goto repeat;
715         }
716
717         return err;
718 }
719
720 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
721
722 static int cpa_process_alias(struct cpa_data *cpa)
723 {
724         struct cpa_data alias_cpa;
725         unsigned long laddr = (unsigned long)__va(cpa->pfn << PAGE_SHIFT);
726         unsigned long vaddr;
727         int ret;
728
729         if (cpa->pfn >= max_pfn_mapped)
730                 return 0;
731
732 #ifdef CONFIG_X86_64
733         if (cpa->pfn >= max_low_pfn_mapped && cpa->pfn < (1UL<<(32-PAGE_SHIFT)))
734                 return 0;
735 #endif
736         /*
737          * No need to redo, when the primary call touched the direct
738          * mapping already:
739          */
740         if (cpa->flags & CPA_PAGES_ARRAY) {
741                 struct page *page = cpa->pages[cpa->curpage];
742                 if (unlikely(PageHighMem(page)))
743                         return 0;
744                 vaddr = (unsigned long)page_address(page);
745         } else if (cpa->flags & CPA_ARRAY)
746                 vaddr = cpa->vaddr[cpa->curpage];
747         else
748                 vaddr = *cpa->vaddr;
749
750         if (!(within(vaddr, PAGE_OFFSET,
751                     PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT)))) {
752
753                 alias_cpa = *cpa;
754                 alias_cpa.vaddr = &laddr;
755                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
756
757                 ret = __change_page_attr_set_clr(&alias_cpa, 0);
758                 if (ret)
759                         return ret;
760         }
761
762 #ifdef CONFIG_X86_64
763         /*
764          * If the primary call didn't touch the high mapping already
765          * and the physical address is inside the kernel map, we need
766          * to touch the high mapped kernel as well:
767          */
768         if (!within(vaddr, (unsigned long)_text, _brk_end) &&
769             within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn())) {
770                 unsigned long temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) +
771                                                __START_KERNEL_map - phys_base;
772                 alias_cpa = *cpa;
773                 alias_cpa.vaddr = &temp_cpa_vaddr;
774                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
775
776                 /*
777                  * The high mapping range is imprecise, so ignore the
778                  * return value.
779                  */
780                 __change_page_attr_set_clr(&alias_cpa, 0);
781         }
782 #endif
783
784         return 0;
785 }
786
787 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
788 {
789         int ret, numpages = cpa->numpages;
790
791         while (numpages) {
792                 /*
793                  * Store the remaining nr of pages for the large page
794                  * preservation check.
795                  */
796                 cpa->numpages = numpages;
797                 /* for array changes, we can't use large page */
798                 if (cpa->flags & (CPA_ARRAY | CPA_PAGES_ARRAY))
799                         cpa->numpages = 1;
800
801                 if (!debug_pagealloc)
802                         spin_lock(&cpa_lock);
803                 ret = __change_page_attr(cpa, checkalias);
804                 if (!debug_pagealloc)
805                         spin_unlock(&cpa_lock);
806                 if (ret)
807                         return ret;
808
809                 if (checkalias) {
810                         ret = cpa_process_alias(cpa);
811                         if (ret)
812                                 return ret;
813                 }
814
815                 /*
816                  * Adjust the number of pages with the result of the
817                  * CPA operation. Either a large page has been
818                  * preserved or a single page update happened.
819                  */
820                 BUG_ON(cpa->numpages > numpages);
821                 numpages -= cpa->numpages;
822                 if (cpa->flags & (CPA_PAGES_ARRAY | CPA_ARRAY))
823                         cpa->curpage++;
824                 else
825                         *cpa->vaddr += cpa->numpages * PAGE_SIZE;
826
827         }
828         return 0;
829 }
830
831 static inline int cache_attr(pgprot_t attr)
832 {
833         return pgprot_val(attr) &
834                 (_PAGE_PAT | _PAGE_PAT_LARGE | _PAGE_PWT | _PAGE_PCD);
835 }
836
837 static int change_page_attr_set_clr(unsigned long *addr, int numpages,
838                                     pgprot_t mask_set, pgprot_t mask_clr,
839                                     int force_split, int in_flag,
840                                     struct page **pages)
841 {
842         struct cpa_data cpa;
843         int ret, cache, checkalias;
844         unsigned long baddr = 0;
845
846         /*
847          * Check, if we are requested to change a not supported
848          * feature:
849          */
850         mask_set = canon_pgprot(mask_set);
851         mask_clr = canon_pgprot(mask_clr);
852         if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
853                 return 0;
854
855         /* Ensure we are PAGE_SIZE aligned */
856         if (in_flag & CPA_ARRAY) {
857                 int i;
858                 for (i = 0; i < numpages; i++) {
859                         if (addr[i] & ~PAGE_MASK) {
860                                 addr[i] &= PAGE_MASK;
861                                 WARN_ON_ONCE(1);
862                         }
863                 }
864         } else if (!(in_flag & CPA_PAGES_ARRAY)) {
865                 /*
866                  * in_flag of CPA_PAGES_ARRAY implies it is aligned.
867                  * No need to cehck in that case
868                  */
869                 if (*addr & ~PAGE_MASK) {
870                         *addr &= PAGE_MASK;
871                         /*
872                          * People should not be passing in unaligned addresses:
873                          */
874                         WARN_ON_ONCE(1);
875                 }
876                 /*
877                  * Save address for cache flush. *addr is modified in the call
878                  * to __change_page_attr_set_clr() below.
879                  */
880                 baddr = *addr;
881         }
882
883         /* Must avoid aliasing mappings in the highmem code */
884         kmap_flush_unused();
885
886         vm_unmap_aliases();
887
888         cpa.vaddr = addr;
889         cpa.pages = pages;
890         cpa.numpages = numpages;
891         cpa.mask_set = mask_set;
892         cpa.mask_clr = mask_clr;
893         cpa.flags = 0;
894         cpa.curpage = 0;
895         cpa.force_split = force_split;
896
897         if (in_flag & (CPA_ARRAY | CPA_PAGES_ARRAY))
898                 cpa.flags |= in_flag;
899
900         /* No alias checking for _NX bit modifications */
901         checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
902
903         ret = __change_page_attr_set_clr(&cpa, checkalias);
904
905         /*
906          * Check whether we really changed something:
907          */
908         if (!(cpa.flags & CPA_FLUSHTLB))
909                 goto out;
910
911         /*
912          * No need to flush, when we did not set any of the caching
913          * attributes:
914          */
915         cache = cache_attr(mask_set);
916
917         /*
918          * On success we use clflush, when the CPU supports it to
919          * avoid the wbindv. If the CPU does not support it and in the
920          * error case we fall back to cpa_flush_all (which uses
921          * wbindv):
922          */
923         if (!ret && cpu_has_clflush) {
924                 if (cpa.flags & (CPA_PAGES_ARRAY | CPA_ARRAY)) {
925                         cpa_flush_array(addr, numpages, cache,
926                                         cpa.flags, pages);
927                 } else
928                         cpa_flush_range(baddr, numpages, cache);
929         } else
930                 cpa_flush_all(cache);
931
932 out:
933         return ret;
934 }
935
936 static inline int change_page_attr_set(unsigned long *addr, int numpages,
937                                        pgprot_t mask, int array)
938 {
939         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
940                 (array ? CPA_ARRAY : 0), NULL);
941 }
942
943 static inline int change_page_attr_clear(unsigned long *addr, int numpages,
944                                          pgprot_t mask, int array)
945 {
946         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
947                 (array ? CPA_ARRAY : 0), NULL);
948 }
949
950 static inline int cpa_set_pages_array(struct page **pages, int numpages,
951                                        pgprot_t mask)
952 {
953         return change_page_attr_set_clr(NULL, numpages, mask, __pgprot(0), 0,
954                 CPA_PAGES_ARRAY, pages);
955 }
956
957 static inline int cpa_clear_pages_array(struct page **pages, int numpages,
958                                          pgprot_t mask)
959 {
960         return change_page_attr_set_clr(NULL, numpages, __pgprot(0), mask, 0,
961                 CPA_PAGES_ARRAY, pages);
962 }
963
964 int _set_memory_uc(unsigned long addr, int numpages)
965 {
966         /*
967          * for now UC MINUS. see comments in ioremap_nocache()
968          */
969         return change_page_attr_set(&addr, numpages,
970                                     __pgprot(_PAGE_CACHE_UC_MINUS), 0);
971 }
972
973 int set_memory_uc(unsigned long addr, int numpages)
974 {
975         int ret;
976
977         /*
978          * for now UC MINUS. see comments in ioremap_nocache()
979          */
980         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
981                             _PAGE_CACHE_UC_MINUS, NULL);
982         if (ret)
983                 goto out_err;
984
985         ret = _set_memory_uc(addr, numpages);
986         if (ret)
987                 goto out_free;
988
989         return 0;
990
991 out_free:
992         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
993 out_err:
994         return ret;
995 }
996 EXPORT_SYMBOL(set_memory_uc);
997
998 static int _set_memory_array(unsigned long *addr, int addrinarray,
999                 unsigned long new_type)
1000 {
1001         int i, j;
1002         int ret;
1003
1004         /*
1005          * for now UC MINUS. see comments in ioremap_nocache()
1006          */
1007         for (i = 0; i < addrinarray; i++) {
1008                 ret = reserve_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE,
1009                                         new_type, NULL);
1010                 if (ret)
1011                         goto out_free;
1012         }
1013
1014         ret = change_page_attr_set(addr, addrinarray,
1015                                     __pgprot(_PAGE_CACHE_UC_MINUS), 1);
1016
1017         if (!ret && new_type == _PAGE_CACHE_WC)
1018                 ret = change_page_attr_set_clr(addr, addrinarray,
1019                                                __pgprot(_PAGE_CACHE_WC),
1020                                                __pgprot(_PAGE_CACHE_MASK),
1021                                                0, CPA_ARRAY, NULL);
1022         if (ret)
1023                 goto out_free;
1024
1025         return 0;
1026
1027 out_free:
1028         for (j = 0; j < i; j++)
1029                 free_memtype(__pa(addr[j]), __pa(addr[j]) + PAGE_SIZE);
1030
1031         return ret;
1032 }
1033
1034 int set_memory_array_uc(unsigned long *addr, int addrinarray)
1035 {
1036         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_UC_MINUS);
1037 }
1038 EXPORT_SYMBOL(set_memory_array_uc);
1039
1040 int set_memory_array_wc(unsigned long *addr, int addrinarray)
1041 {
1042         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_WC);
1043 }
1044 EXPORT_SYMBOL(set_memory_array_wc);
1045
1046 int _set_memory_wc(unsigned long addr, int numpages)
1047 {
1048         int ret;
1049         unsigned long addr_copy = addr;
1050
1051         ret = change_page_attr_set(&addr, numpages,
1052                                     __pgprot(_PAGE_CACHE_UC_MINUS), 0);
1053         if (!ret) {
1054                 ret = change_page_attr_set_clr(&addr_copy, numpages,
1055                                                __pgprot(_PAGE_CACHE_WC),
1056                                                __pgprot(_PAGE_CACHE_MASK),
1057                                                0, 0, NULL);
1058         }
1059         return ret;
1060 }
1061
1062 int set_memory_wc(unsigned long addr, int numpages)
1063 {
1064         int ret;
1065
1066         if (!pat_enabled)
1067                 return set_memory_uc(addr, numpages);
1068
1069         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1070                 _PAGE_CACHE_WC, NULL);
1071         if (ret)
1072                 goto out_err;
1073
1074         ret = _set_memory_wc(addr, numpages);
1075         if (ret)
1076                 goto out_free;
1077
1078         return 0;
1079
1080 out_free:
1081         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1082 out_err:
1083         return ret;
1084 }
1085 EXPORT_SYMBOL(set_memory_wc);
1086
1087 int _set_memory_wb(unsigned long addr, int numpages)
1088 {
1089         return change_page_attr_clear(&addr, numpages,
1090                                       __pgprot(_PAGE_CACHE_MASK), 0);
1091 }
1092
1093 int set_memory_wb(unsigned long addr, int numpages)
1094 {
1095         int ret;
1096
1097         ret = _set_memory_wb(addr, numpages);
1098         if (ret)
1099                 return ret;
1100
1101         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1102         return 0;
1103 }
1104 EXPORT_SYMBOL(set_memory_wb);
1105
1106 int set_memory_array_wb(unsigned long *addr, int addrinarray)
1107 {
1108         int i;
1109         int ret;
1110
1111         ret = change_page_attr_clear(addr, addrinarray,
1112                                       __pgprot(_PAGE_CACHE_MASK), 1);
1113         if (ret)
1114                 return ret;
1115
1116         for (i = 0; i < addrinarray; i++)
1117                 free_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE);
1118
1119         return 0;
1120 }
1121 EXPORT_SYMBOL(set_memory_array_wb);
1122
1123 int set_memory_x(unsigned long addr, int numpages)
1124 {
1125         if (!(__supported_pte_mask & _PAGE_NX))
1126                 return 0;
1127
1128         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1129 }
1130 EXPORT_SYMBOL(set_memory_x);
1131
1132 int set_memory_nx(unsigned long addr, int numpages)
1133 {
1134         if (!(__supported_pte_mask & _PAGE_NX))
1135                 return 0;
1136
1137         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1138 }
1139 EXPORT_SYMBOL(set_memory_nx);
1140
1141 int set_memory_ro(unsigned long addr, int numpages)
1142 {
1143         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1144 }
1145 EXPORT_SYMBOL_GPL(set_memory_ro);
1146
1147 int set_memory_rw(unsigned long addr, int numpages)
1148 {
1149         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
1150 }
1151 EXPORT_SYMBOL_GPL(set_memory_rw);
1152
1153 int set_memory_np(unsigned long addr, int numpages)
1154 {
1155         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
1156 }
1157
1158 int set_memory_4k(unsigned long addr, int numpages)
1159 {
1160         return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1161                                         __pgprot(0), 1, 0, NULL);
1162 }
1163
1164 int set_pages_uc(struct page *page, int numpages)
1165 {
1166         unsigned long addr = (unsigned long)page_address(page);
1167
1168         return set_memory_uc(addr, numpages);
1169 }
1170 EXPORT_SYMBOL(set_pages_uc);
1171
1172 static int _set_pages_array(struct page **pages, int addrinarray,
1173                 unsigned long new_type)
1174 {
1175         unsigned long start;
1176         unsigned long end;
1177         int i;
1178         int free_idx;
1179         int ret;
1180
1181         for (i = 0; i < addrinarray; i++) {
1182                 if (PageHighMem(pages[i]))
1183                         continue;
1184                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1185                 end = start + PAGE_SIZE;
1186                 if (reserve_memtype(start, end, new_type, NULL))
1187                         goto err_out;
1188         }
1189
1190         ret = cpa_set_pages_array(pages, addrinarray,
1191                         __pgprot(_PAGE_CACHE_UC_MINUS));
1192         if (!ret && new_type == _PAGE_CACHE_WC)
1193                 ret = change_page_attr_set_clr(NULL, addrinarray,
1194                                                __pgprot(_PAGE_CACHE_WC),
1195                                                __pgprot(_PAGE_CACHE_MASK),
1196                                                0, CPA_PAGES_ARRAY, pages);
1197         if (ret)
1198                 goto err_out;
1199         return 0; /* Success */
1200 err_out:
1201         free_idx = i;
1202         for (i = 0; i < free_idx; i++) {
1203                 if (PageHighMem(pages[i]))
1204                         continue;
1205                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1206                 end = start + PAGE_SIZE;
1207                 free_memtype(start, end);
1208         }
1209         return -EINVAL;
1210 }
1211
1212 int set_pages_array_uc(struct page **pages, int addrinarray)
1213 {
1214         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_UC_MINUS);
1215 }
1216 EXPORT_SYMBOL(set_pages_array_uc);
1217
1218 int set_pages_array_wc(struct page **pages, int addrinarray)
1219 {
1220         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_WC);
1221 }
1222 EXPORT_SYMBOL(set_pages_array_wc);
1223
1224 int set_pages_wb(struct page *page, int numpages)
1225 {
1226         unsigned long addr = (unsigned long)page_address(page);
1227
1228         return set_memory_wb(addr, numpages);
1229 }
1230 EXPORT_SYMBOL(set_pages_wb);
1231
1232 int set_pages_array_wb(struct page **pages, int addrinarray)
1233 {
1234         int retval;
1235         unsigned long start;
1236         unsigned long end;
1237         int i;
1238
1239         retval = cpa_clear_pages_array(pages, addrinarray,
1240                         __pgprot(_PAGE_CACHE_MASK));
1241         if (retval)
1242                 return retval;
1243
1244         for (i = 0; i < addrinarray; i++) {
1245                 if (PageHighMem(pages[i]))
1246                         continue;
1247                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1248                 end = start + PAGE_SIZE;
1249                 free_memtype(start, end);
1250         }
1251
1252         return 0;
1253 }
1254 EXPORT_SYMBOL(set_pages_array_wb);
1255
1256 int set_pages_x(struct page *page, int numpages)
1257 {
1258         unsigned long addr = (unsigned long)page_address(page);
1259
1260         return set_memory_x(addr, numpages);
1261 }
1262 EXPORT_SYMBOL(set_pages_x);
1263
1264 int set_pages_nx(struct page *page, int numpages)
1265 {
1266         unsigned long addr = (unsigned long)page_address(page);
1267
1268         return set_memory_nx(addr, numpages);
1269 }
1270 EXPORT_SYMBOL(set_pages_nx);
1271
1272 int set_pages_ro(struct page *page, int numpages)
1273 {
1274         unsigned long addr = (unsigned long)page_address(page);
1275
1276         return set_memory_ro(addr, numpages);
1277 }
1278
1279 int set_pages_rw(struct page *page, int numpages)
1280 {
1281         unsigned long addr = (unsigned long)page_address(page);
1282
1283         return set_memory_rw(addr, numpages);
1284 }
1285
1286 #ifdef CONFIG_DEBUG_PAGEALLOC
1287
1288 static int __set_pages_p(struct page *page, int numpages)
1289 {
1290         unsigned long tempaddr = (unsigned long) page_address(page);
1291         struct cpa_data cpa = { .vaddr = &tempaddr,
1292                                 .numpages = numpages,
1293                                 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1294                                 .mask_clr = __pgprot(0),
1295                                 .flags = 0};
1296
1297         /*
1298          * No alias checking needed for setting present flag. otherwise,
1299          * we may need to break large pages for 64-bit kernel text
1300          * mappings (this adds to complexity if we want to do this from
1301          * atomic context especially). Let's keep it simple!
1302          */
1303         return __change_page_attr_set_clr(&cpa, 0);
1304 }
1305
1306 static int __set_pages_np(struct page *page, int numpages)
1307 {
1308         unsigned long tempaddr = (unsigned long) page_address(page);
1309         struct cpa_data cpa = { .vaddr = &tempaddr,
1310                                 .numpages = numpages,
1311                                 .mask_set = __pgprot(0),
1312                                 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1313                                 .flags = 0};
1314
1315         /*
1316          * No alias checking needed for setting not present flag. otherwise,
1317          * we may need to break large pages for 64-bit kernel text
1318          * mappings (this adds to complexity if we want to do this from
1319          * atomic context especially). Let's keep it simple!
1320          */
1321         return __change_page_attr_set_clr(&cpa, 0);
1322 }
1323
1324 void kernel_map_pages(struct page *page, int numpages, int enable)
1325 {
1326         if (PageHighMem(page))
1327                 return;
1328         if (!enable) {
1329                 debug_check_no_locks_freed(page_address(page),
1330                                            numpages * PAGE_SIZE);
1331         }
1332
1333         /*
1334          * The return value is ignored as the calls cannot fail.
1335          * Large pages for identity mappings are not used at boot time
1336          * and hence no memory allocations during large page split.
1337          */
1338         if (enable)
1339                 __set_pages_p(page, numpages);
1340         else
1341                 __set_pages_np(page, numpages);
1342
1343         /*
1344          * We should perform an IPI and flush all tlbs,
1345          * but that can deadlock->flush only current cpu:
1346          */
1347         __flush_tlb_all();
1348 }
1349
1350 #ifdef CONFIG_HIBERNATION
1351
1352 bool kernel_page_present(struct page *page)
1353 {
1354         unsigned int level;
1355         pte_t *pte;
1356
1357         if (PageHighMem(page))
1358                 return false;
1359
1360         pte = lookup_address((unsigned long)page_address(page), &level);
1361         return (pte_val(*pte) & _PAGE_PRESENT);
1362 }
1363
1364 #endif /* CONFIG_HIBERNATION */
1365
1366 #endif /* CONFIG_DEBUG_PAGEALLOC */
1367
1368 /*
1369  * The testcases use internal knowledge of the implementation that shouldn't
1370  * be exposed to the rest of the kernel. Include these directly here.
1371  */
1372 #ifdef CONFIG_CPA_DEBUG
1373 #include "pageattr-test.c"
1374 #endif