Merge branch 'clockevents/fixes' of git://git.linaro.org/people/daniel.lezcano/linux...
[linux-drm-fsl-dcu.git] / arch / x86 / mm / pgtable.c
1 #include <linux/mm.h>
2 #include <linux/gfp.h>
3 #include <asm/pgalloc.h>
4 #include <asm/pgtable.h>
5 #include <asm/tlb.h>
6 #include <asm/fixmap.h>
7
8 #define PGALLOC_GFP GFP_KERNEL | __GFP_NOTRACK | __GFP_REPEAT | __GFP_ZERO
9
10 #ifdef CONFIG_HIGHPTE
11 #define PGALLOC_USER_GFP __GFP_HIGHMEM
12 #else
13 #define PGALLOC_USER_GFP 0
14 #endif
15
16 gfp_t __userpte_alloc_gfp = PGALLOC_GFP | PGALLOC_USER_GFP;
17
18 pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
19 {
20         return (pte_t *)__get_free_page(PGALLOC_GFP);
21 }
22
23 pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
24 {
25         struct page *pte;
26
27         pte = alloc_pages(__userpte_alloc_gfp, 0);
28         if (!pte)
29                 return NULL;
30         if (!pgtable_page_ctor(pte)) {
31                 __free_page(pte);
32                 return NULL;
33         }
34         return pte;
35 }
36
37 static int __init setup_userpte(char *arg)
38 {
39         if (!arg)
40                 return -EINVAL;
41
42         /*
43          * "userpte=nohigh" disables allocation of user pagetables in
44          * high memory.
45          */
46         if (strcmp(arg, "nohigh") == 0)
47                 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
48         else
49                 return -EINVAL;
50         return 0;
51 }
52 early_param("userpte", setup_userpte);
53
54 void ___pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
55 {
56         pgtable_page_dtor(pte);
57         paravirt_release_pte(page_to_pfn(pte));
58         tlb_remove_page(tlb, pte);
59 }
60
61 #if PAGETABLE_LEVELS > 2
62 void ___pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
63 {
64         struct page *page = virt_to_page(pmd);
65         paravirt_release_pmd(__pa(pmd) >> PAGE_SHIFT);
66         /*
67          * NOTE! For PAE, any changes to the top page-directory-pointer-table
68          * entries need a full cr3 reload to flush.
69          */
70 #ifdef CONFIG_X86_PAE
71         tlb->need_flush_all = 1;
72 #endif
73         pgtable_pmd_page_dtor(page);
74         tlb_remove_page(tlb, page);
75 }
76
77 #if PAGETABLE_LEVELS > 3
78 void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
79 {
80         paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
81         tlb_remove_page(tlb, virt_to_page(pud));
82 }
83 #endif  /* PAGETABLE_LEVELS > 3 */
84 #endif  /* PAGETABLE_LEVELS > 2 */
85
86 static inline void pgd_list_add(pgd_t *pgd)
87 {
88         struct page *page = virt_to_page(pgd);
89
90         list_add(&page->lru, &pgd_list);
91 }
92
93 static inline void pgd_list_del(pgd_t *pgd)
94 {
95         struct page *page = virt_to_page(pgd);
96
97         list_del(&page->lru);
98 }
99
100 #define UNSHARED_PTRS_PER_PGD                           \
101         (SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
102
103
104 static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
105 {
106         BUILD_BUG_ON(sizeof(virt_to_page(pgd)->index) < sizeof(mm));
107         virt_to_page(pgd)->index = (pgoff_t)mm;
108 }
109
110 struct mm_struct *pgd_page_get_mm(struct page *page)
111 {
112         return (struct mm_struct *)page->index;
113 }
114
115 static void pgd_ctor(struct mm_struct *mm, pgd_t *pgd)
116 {
117         /* If the pgd points to a shared pagetable level (either the
118            ptes in non-PAE, or shared PMD in PAE), then just copy the
119            references from swapper_pg_dir. */
120         if (PAGETABLE_LEVELS == 2 ||
121             (PAGETABLE_LEVELS == 3 && SHARED_KERNEL_PMD) ||
122             PAGETABLE_LEVELS == 4) {
123                 clone_pgd_range(pgd + KERNEL_PGD_BOUNDARY,
124                                 swapper_pg_dir + KERNEL_PGD_BOUNDARY,
125                                 KERNEL_PGD_PTRS);
126         }
127
128         /* list required to sync kernel mapping updates */
129         if (!SHARED_KERNEL_PMD) {
130                 pgd_set_mm(pgd, mm);
131                 pgd_list_add(pgd);
132         }
133 }
134
135 static void pgd_dtor(pgd_t *pgd)
136 {
137         if (SHARED_KERNEL_PMD)
138                 return;
139
140         spin_lock(&pgd_lock);
141         pgd_list_del(pgd);
142         spin_unlock(&pgd_lock);
143 }
144
145 /*
146  * List of all pgd's needed for non-PAE so it can invalidate entries
147  * in both cached and uncached pgd's; not needed for PAE since the
148  * kernel pmd is shared. If PAE were not to share the pmd a similar
149  * tactic would be needed. This is essentially codepath-based locking
150  * against pageattr.c; it is the unique case in which a valid change
151  * of kernel pagetables can't be lazily synchronized by vmalloc faults.
152  * vmalloc faults work because attached pagetables are never freed.
153  * -- nyc
154  */
155
156 #ifdef CONFIG_X86_PAE
157 /*
158  * In PAE mode, we need to do a cr3 reload (=tlb flush) when
159  * updating the top-level pagetable entries to guarantee the
160  * processor notices the update.  Since this is expensive, and
161  * all 4 top-level entries are used almost immediately in a
162  * new process's life, we just pre-populate them here.
163  *
164  * Also, if we're in a paravirt environment where the kernel pmd is
165  * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
166  * and initialize the kernel pmds here.
167  */
168 #define PREALLOCATED_PMDS       UNSHARED_PTRS_PER_PGD
169
170 void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
171 {
172         paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
173
174         /* Note: almost everything apart from _PAGE_PRESENT is
175            reserved at the pmd (PDPT) level. */
176         set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
177
178         /*
179          * According to Intel App note "TLBs, Paging-Structure Caches,
180          * and Their Invalidation", April 2007, document 317080-001,
181          * section 8.1: in PAE mode we explicitly have to flush the
182          * TLB via cr3 if the top-level pgd is changed...
183          */
184         flush_tlb_mm(mm);
185 }
186 #else  /* !CONFIG_X86_PAE */
187
188 /* No need to prepopulate any pagetable entries in non-PAE modes. */
189 #define PREALLOCATED_PMDS       0
190
191 #endif  /* CONFIG_X86_PAE */
192
193 static void free_pmds(pmd_t *pmds[])
194 {
195         int i;
196
197         for(i = 0; i < PREALLOCATED_PMDS; i++)
198                 if (pmds[i]) {
199                         pgtable_pmd_page_dtor(virt_to_page(pmds[i]));
200                         free_page((unsigned long)pmds[i]);
201                 }
202 }
203
204 static int preallocate_pmds(pmd_t *pmds[])
205 {
206         int i;
207         bool failed = false;
208
209         for(i = 0; i < PREALLOCATED_PMDS; i++) {
210                 pmd_t *pmd = (pmd_t *)__get_free_page(PGALLOC_GFP);
211                 if (!pmd)
212                         failed = true;
213                 if (pmd && !pgtable_pmd_page_ctor(virt_to_page(pmd))) {
214                         free_page((unsigned long)pmd);
215                         pmd = NULL;
216                         failed = true;
217                 }
218                 pmds[i] = pmd;
219         }
220
221         if (failed) {
222                 free_pmds(pmds);
223                 return -ENOMEM;
224         }
225
226         return 0;
227 }
228
229 /*
230  * Mop up any pmd pages which may still be attached to the pgd.
231  * Normally they will be freed by munmap/exit_mmap, but any pmd we
232  * preallocate which never got a corresponding vma will need to be
233  * freed manually.
234  */
235 static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
236 {
237         int i;
238
239         for(i = 0; i < PREALLOCATED_PMDS; i++) {
240                 pgd_t pgd = pgdp[i];
241
242                 if (pgd_val(pgd) != 0) {
243                         pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
244
245                         pgdp[i] = native_make_pgd(0);
246
247                         paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
248                         pmd_free(mm, pmd);
249                 }
250         }
251 }
252
253 static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
254 {
255         pud_t *pud;
256         int i;
257
258         if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
259                 return;
260
261         pud = pud_offset(pgd, 0);
262
263         for (i = 0; i < PREALLOCATED_PMDS; i++, pud++) {
264                 pmd_t *pmd = pmds[i];
265
266                 if (i >= KERNEL_PGD_BOUNDARY)
267                         memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
268                                sizeof(pmd_t) * PTRS_PER_PMD);
269
270                 pud_populate(mm, pud, pmd);
271         }
272 }
273
274 pgd_t *pgd_alloc(struct mm_struct *mm)
275 {
276         pgd_t *pgd;
277         pmd_t *pmds[PREALLOCATED_PMDS];
278
279         pgd = (pgd_t *)__get_free_page(PGALLOC_GFP);
280
281         if (pgd == NULL)
282                 goto out;
283
284         mm->pgd = pgd;
285
286         if (preallocate_pmds(pmds) != 0)
287                 goto out_free_pgd;
288
289         if (paravirt_pgd_alloc(mm) != 0)
290                 goto out_free_pmds;
291
292         /*
293          * Make sure that pre-populating the pmds is atomic with
294          * respect to anything walking the pgd_list, so that they
295          * never see a partially populated pgd.
296          */
297         spin_lock(&pgd_lock);
298
299         pgd_ctor(mm, pgd);
300         pgd_prepopulate_pmd(mm, pgd, pmds);
301
302         spin_unlock(&pgd_lock);
303
304         return pgd;
305
306 out_free_pmds:
307         free_pmds(pmds);
308 out_free_pgd:
309         free_page((unsigned long)pgd);
310 out:
311         return NULL;
312 }
313
314 void pgd_free(struct mm_struct *mm, pgd_t *pgd)
315 {
316         pgd_mop_up_pmds(mm, pgd);
317         pgd_dtor(pgd);
318         paravirt_pgd_free(mm, pgd);
319         free_page((unsigned long)pgd);
320 }
321
322 /*
323  * Used to set accessed or dirty bits in the page table entries
324  * on other architectures. On x86, the accessed and dirty bits
325  * are tracked by hardware. However, do_wp_page calls this function
326  * to also make the pte writeable at the same time the dirty bit is
327  * set. In that case we do actually need to write the PTE.
328  */
329 int ptep_set_access_flags(struct vm_area_struct *vma,
330                           unsigned long address, pte_t *ptep,
331                           pte_t entry, int dirty)
332 {
333         int changed = !pte_same(*ptep, entry);
334
335         if (changed && dirty) {
336                 *ptep = entry;
337                 pte_update_defer(vma->vm_mm, address, ptep);
338         }
339
340         return changed;
341 }
342
343 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
344 int pmdp_set_access_flags(struct vm_area_struct *vma,
345                           unsigned long address, pmd_t *pmdp,
346                           pmd_t entry, int dirty)
347 {
348         int changed = !pmd_same(*pmdp, entry);
349
350         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
351
352         if (changed && dirty) {
353                 *pmdp = entry;
354                 pmd_update_defer(vma->vm_mm, address, pmdp);
355                 /*
356                  * We had a write-protection fault here and changed the pmd
357                  * to to more permissive. No need to flush the TLB for that,
358                  * #PF is architecturally guaranteed to do that and in the
359                  * worst-case we'll generate a spurious fault.
360                  */
361         }
362
363         return changed;
364 }
365 #endif
366
367 int ptep_test_and_clear_young(struct vm_area_struct *vma,
368                               unsigned long addr, pte_t *ptep)
369 {
370         int ret = 0;
371
372         if (pte_young(*ptep))
373                 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
374                                          (unsigned long *) &ptep->pte);
375
376         if (ret)
377                 pte_update(vma->vm_mm, addr, ptep);
378
379         return ret;
380 }
381
382 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
383 int pmdp_test_and_clear_young(struct vm_area_struct *vma,
384                               unsigned long addr, pmd_t *pmdp)
385 {
386         int ret = 0;
387
388         if (pmd_young(*pmdp))
389                 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
390                                          (unsigned long *)pmdp);
391
392         if (ret)
393                 pmd_update(vma->vm_mm, addr, pmdp);
394
395         return ret;
396 }
397 #endif
398
399 int ptep_clear_flush_young(struct vm_area_struct *vma,
400                            unsigned long address, pte_t *ptep)
401 {
402         int young;
403
404         young = ptep_test_and_clear_young(vma, address, ptep);
405         if (young)
406                 flush_tlb_page(vma, address);
407
408         return young;
409 }
410
411 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
412 int pmdp_clear_flush_young(struct vm_area_struct *vma,
413                            unsigned long address, pmd_t *pmdp)
414 {
415         int young;
416
417         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
418
419         young = pmdp_test_and_clear_young(vma, address, pmdp);
420         if (young)
421                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
422
423         return young;
424 }
425
426 void pmdp_splitting_flush(struct vm_area_struct *vma,
427                           unsigned long address, pmd_t *pmdp)
428 {
429         int set;
430         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
431         set = !test_and_set_bit(_PAGE_BIT_SPLITTING,
432                                 (unsigned long *)pmdp);
433         if (set) {
434                 pmd_update(vma->vm_mm, address, pmdp);
435                 /* need tlb flush only to serialize against gup-fast */
436                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
437         }
438 }
439 #endif
440
441 /**
442  * reserve_top_address - reserves a hole in the top of kernel address space
443  * @reserve - size of hole to reserve
444  *
445  * Can be used to relocate the fixmap area and poke a hole in the top
446  * of kernel address space to make room for a hypervisor.
447  */
448 void __init reserve_top_address(unsigned long reserve)
449 {
450 #ifdef CONFIG_X86_32
451         BUG_ON(fixmaps_set > 0);
452         printk(KERN_INFO "Reserving virtual address space above 0x%08x\n",
453                (int)-reserve);
454         __FIXADDR_TOP = -reserve - PAGE_SIZE;
455 #endif
456 }
457
458 int fixmaps_set;
459
460 void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
461 {
462         unsigned long address = __fix_to_virt(idx);
463
464         if (idx >= __end_of_fixed_addresses) {
465                 BUG();
466                 return;
467         }
468         set_pte_vaddr(address, pte);
469         fixmaps_set++;
470 }
471
472 void native_set_fixmap(enum fixed_addresses idx, phys_addr_t phys,
473                        pgprot_t flags)
474 {
475         __native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
476 }