Loading...
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
16gfp_t __userpte_alloc_gfp = PGALLOC_GFP | PGALLOC_USER_GFP;
17
18pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
19{
20 return (pte_t *)__get_free_page(PGALLOC_GFP);
21}
22
23pgtable_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
37static 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}
52early_param("userpte", setup_userpte);
53
54void ___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
62void ___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
78void ___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
86static 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
93static 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
104static 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
110struct mm_struct *pgd_page_get_mm(struct page *page)
111{
112 return (struct mm_struct *)page->index;
113}
114
115static 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
135static 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
170void 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
193static 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
204static 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 */
235static 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
253static 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
274pgd_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
306out_free_pmds:
307 free_pmds(pmds);
308out_free_pgd:
309 free_page((unsigned long)pgd);
310out:
311 return NULL;
312}
313
314void 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 */
329int 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
344int 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
367int 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
383int 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
399int 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
412int 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
426void 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 */
448void __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
458int fixmaps_set;
459
460void __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
472void 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}
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#include <asm/mtrr.h>
8
9#define PGALLOC_GFP (GFP_KERNEL_ACCOUNT | __GFP_NOTRACK | __GFP_ZERO)
10
11#ifdef CONFIG_HIGHPTE
12#define PGALLOC_USER_GFP __GFP_HIGHMEM
13#else
14#define PGALLOC_USER_GFP 0
15#endif
16
17gfp_t __userpte_alloc_gfp = PGALLOC_GFP | PGALLOC_USER_GFP;
18
19pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
20{
21 return (pte_t *)__get_free_page(PGALLOC_GFP & ~__GFP_ACCOUNT);
22}
23
24pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
25{
26 struct page *pte;
27
28 pte = alloc_pages(__userpte_alloc_gfp, 0);
29 if (!pte)
30 return NULL;
31 if (!pgtable_page_ctor(pte)) {
32 __free_page(pte);
33 return NULL;
34 }
35 return pte;
36}
37
38static int __init setup_userpte(char *arg)
39{
40 if (!arg)
41 return -EINVAL;
42
43 /*
44 * "userpte=nohigh" disables allocation of user pagetables in
45 * high memory.
46 */
47 if (strcmp(arg, "nohigh") == 0)
48 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
49 else
50 return -EINVAL;
51 return 0;
52}
53early_param("userpte", setup_userpte);
54
55void ___pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
56{
57 pgtable_page_dtor(pte);
58 paravirt_release_pte(page_to_pfn(pte));
59 tlb_remove_page(tlb, pte);
60}
61
62#if CONFIG_PGTABLE_LEVELS > 2
63void ___pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
64{
65 struct page *page = virt_to_page(pmd);
66 paravirt_release_pmd(__pa(pmd) >> PAGE_SHIFT);
67 /*
68 * NOTE! For PAE, any changes to the top page-directory-pointer-table
69 * entries need a full cr3 reload to flush.
70 */
71#ifdef CONFIG_X86_PAE
72 tlb->need_flush_all = 1;
73#endif
74 pgtable_pmd_page_dtor(page);
75 tlb_remove_page(tlb, page);
76}
77
78#if CONFIG_PGTABLE_LEVELS > 3
79void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
80{
81 paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
82 tlb_remove_page(tlb, virt_to_page(pud));
83}
84#endif /* CONFIG_PGTABLE_LEVELS > 3 */
85#endif /* CONFIG_PGTABLE_LEVELS > 2 */
86
87static inline void pgd_list_add(pgd_t *pgd)
88{
89 struct page *page = virt_to_page(pgd);
90
91 list_add(&page->lru, &pgd_list);
92}
93
94static inline void pgd_list_del(pgd_t *pgd)
95{
96 struct page *page = virt_to_page(pgd);
97
98 list_del(&page->lru);
99}
100
101#define UNSHARED_PTRS_PER_PGD \
102 (SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
103
104
105static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
106{
107 BUILD_BUG_ON(sizeof(virt_to_page(pgd)->index) < sizeof(mm));
108 virt_to_page(pgd)->index = (pgoff_t)mm;
109}
110
111struct mm_struct *pgd_page_get_mm(struct page *page)
112{
113 return (struct mm_struct *)page->index;
114}
115
116static void pgd_ctor(struct mm_struct *mm, pgd_t *pgd)
117{
118 /* If the pgd points to a shared pagetable level (either the
119 ptes in non-PAE, or shared PMD in PAE), then just copy the
120 references from swapper_pg_dir. */
121 if (CONFIG_PGTABLE_LEVELS == 2 ||
122 (CONFIG_PGTABLE_LEVELS == 3 && SHARED_KERNEL_PMD) ||
123 CONFIG_PGTABLE_LEVELS == 4) {
124 clone_pgd_range(pgd + KERNEL_PGD_BOUNDARY,
125 swapper_pg_dir + KERNEL_PGD_BOUNDARY,
126 KERNEL_PGD_PTRS);
127 }
128
129 /* list required to sync kernel mapping updates */
130 if (!SHARED_KERNEL_PMD) {
131 pgd_set_mm(pgd, mm);
132 pgd_list_add(pgd);
133 }
134}
135
136static void pgd_dtor(pgd_t *pgd)
137{
138 if (SHARED_KERNEL_PMD)
139 return;
140
141 spin_lock(&pgd_lock);
142 pgd_list_del(pgd);
143 spin_unlock(&pgd_lock);
144}
145
146/*
147 * List of all pgd's needed for non-PAE so it can invalidate entries
148 * in both cached and uncached pgd's; not needed for PAE since the
149 * kernel pmd is shared. If PAE were not to share the pmd a similar
150 * tactic would be needed. This is essentially codepath-based locking
151 * against pageattr.c; it is the unique case in which a valid change
152 * of kernel pagetables can't be lazily synchronized by vmalloc faults.
153 * vmalloc faults work because attached pagetables are never freed.
154 * -- nyc
155 */
156
157#ifdef CONFIG_X86_PAE
158/*
159 * In PAE mode, we need to do a cr3 reload (=tlb flush) when
160 * updating the top-level pagetable entries to guarantee the
161 * processor notices the update. Since this is expensive, and
162 * all 4 top-level entries are used almost immediately in a
163 * new process's life, we just pre-populate them here.
164 *
165 * Also, if we're in a paravirt environment where the kernel pmd is
166 * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
167 * and initialize the kernel pmds here.
168 */
169#define PREALLOCATED_PMDS UNSHARED_PTRS_PER_PGD
170
171void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
172{
173 paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
174
175 /* Note: almost everything apart from _PAGE_PRESENT is
176 reserved at the pmd (PDPT) level. */
177 set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
178
179 /*
180 * According to Intel App note "TLBs, Paging-Structure Caches,
181 * and Their Invalidation", April 2007, document 317080-001,
182 * section 8.1: in PAE mode we explicitly have to flush the
183 * TLB via cr3 if the top-level pgd is changed...
184 */
185 flush_tlb_mm(mm);
186}
187#else /* !CONFIG_X86_PAE */
188
189/* No need to prepopulate any pagetable entries in non-PAE modes. */
190#define PREALLOCATED_PMDS 0
191
192#endif /* CONFIG_X86_PAE */
193
194static void free_pmds(struct mm_struct *mm, pmd_t *pmds[])
195{
196 int i;
197
198 for(i = 0; i < PREALLOCATED_PMDS; i++)
199 if (pmds[i]) {
200 pgtable_pmd_page_dtor(virt_to_page(pmds[i]));
201 free_page((unsigned long)pmds[i]);
202 mm_dec_nr_pmds(mm);
203 }
204}
205
206static int preallocate_pmds(struct mm_struct *mm, pmd_t *pmds[])
207{
208 int i;
209 bool failed = false;
210 gfp_t gfp = PGALLOC_GFP;
211
212 if (mm == &init_mm)
213 gfp &= ~__GFP_ACCOUNT;
214
215 for(i = 0; i < PREALLOCATED_PMDS; i++) {
216 pmd_t *pmd = (pmd_t *)__get_free_page(gfp);
217 if (!pmd)
218 failed = true;
219 if (pmd && !pgtable_pmd_page_ctor(virt_to_page(pmd))) {
220 free_page((unsigned long)pmd);
221 pmd = NULL;
222 failed = true;
223 }
224 if (pmd)
225 mm_inc_nr_pmds(mm);
226 pmds[i] = pmd;
227 }
228
229 if (failed) {
230 free_pmds(mm, pmds);
231 return -ENOMEM;
232 }
233
234 return 0;
235}
236
237/*
238 * Mop up any pmd pages which may still be attached to the pgd.
239 * Normally they will be freed by munmap/exit_mmap, but any pmd we
240 * preallocate which never got a corresponding vma will need to be
241 * freed manually.
242 */
243static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
244{
245 int i;
246
247 for(i = 0; i < PREALLOCATED_PMDS; i++) {
248 pgd_t pgd = pgdp[i];
249
250 if (pgd_val(pgd) != 0) {
251 pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
252
253 pgdp[i] = native_make_pgd(0);
254
255 paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
256 pmd_free(mm, pmd);
257 mm_dec_nr_pmds(mm);
258 }
259 }
260}
261
262static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
263{
264 pud_t *pud;
265 int i;
266
267 if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
268 return;
269
270 pud = pud_offset(pgd, 0);
271
272 for (i = 0; i < PREALLOCATED_PMDS; i++, pud++) {
273 pmd_t *pmd = pmds[i];
274
275 if (i >= KERNEL_PGD_BOUNDARY)
276 memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
277 sizeof(pmd_t) * PTRS_PER_PMD);
278
279 pud_populate(mm, pud, pmd);
280 }
281}
282
283/*
284 * Xen paravirt assumes pgd table should be in one page. 64 bit kernel also
285 * assumes that pgd should be in one page.
286 *
287 * But kernel with PAE paging that is not running as a Xen domain
288 * only needs to allocate 32 bytes for pgd instead of one page.
289 */
290#ifdef CONFIG_X86_PAE
291
292#include <linux/slab.h>
293
294#define PGD_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
295#define PGD_ALIGN 32
296
297static struct kmem_cache *pgd_cache;
298
299static int __init pgd_cache_init(void)
300{
301 /*
302 * When PAE kernel is running as a Xen domain, it does not use
303 * shared kernel pmd. And this requires a whole page for pgd.
304 */
305 if (!SHARED_KERNEL_PMD)
306 return 0;
307
308 /*
309 * when PAE kernel is not running as a Xen domain, it uses
310 * shared kernel pmd. Shared kernel pmd does not require a whole
311 * page for pgd. We are able to just allocate a 32-byte for pgd.
312 * During boot time, we create a 32-byte slab for pgd table allocation.
313 */
314 pgd_cache = kmem_cache_create("pgd_cache", PGD_SIZE, PGD_ALIGN,
315 SLAB_PANIC, NULL);
316 if (!pgd_cache)
317 return -ENOMEM;
318
319 return 0;
320}
321core_initcall(pgd_cache_init);
322
323static inline pgd_t *_pgd_alloc(void)
324{
325 /*
326 * If no SHARED_KERNEL_PMD, PAE kernel is running as a Xen domain.
327 * We allocate one page for pgd.
328 */
329 if (!SHARED_KERNEL_PMD)
330 return (pgd_t *)__get_free_page(PGALLOC_GFP);
331
332 /*
333 * Now PAE kernel is not running as a Xen domain. We can allocate
334 * a 32-byte slab for pgd to save memory space.
335 */
336 return kmem_cache_alloc(pgd_cache, PGALLOC_GFP);
337}
338
339static inline void _pgd_free(pgd_t *pgd)
340{
341 if (!SHARED_KERNEL_PMD)
342 free_page((unsigned long)pgd);
343 else
344 kmem_cache_free(pgd_cache, pgd);
345}
346#else
347static inline pgd_t *_pgd_alloc(void)
348{
349 return (pgd_t *)__get_free_page(PGALLOC_GFP);
350}
351
352static inline void _pgd_free(pgd_t *pgd)
353{
354 free_page((unsigned long)pgd);
355}
356#endif /* CONFIG_X86_PAE */
357
358pgd_t *pgd_alloc(struct mm_struct *mm)
359{
360 pgd_t *pgd;
361 pmd_t *pmds[PREALLOCATED_PMDS];
362
363 pgd = _pgd_alloc();
364
365 if (pgd == NULL)
366 goto out;
367
368 mm->pgd = pgd;
369
370 if (preallocate_pmds(mm, pmds) != 0)
371 goto out_free_pgd;
372
373 if (paravirt_pgd_alloc(mm) != 0)
374 goto out_free_pmds;
375
376 /*
377 * Make sure that pre-populating the pmds is atomic with
378 * respect to anything walking the pgd_list, so that they
379 * never see a partially populated pgd.
380 */
381 spin_lock(&pgd_lock);
382
383 pgd_ctor(mm, pgd);
384 pgd_prepopulate_pmd(mm, pgd, pmds);
385
386 spin_unlock(&pgd_lock);
387
388 return pgd;
389
390out_free_pmds:
391 free_pmds(mm, pmds);
392out_free_pgd:
393 _pgd_free(pgd);
394out:
395 return NULL;
396}
397
398void pgd_free(struct mm_struct *mm, pgd_t *pgd)
399{
400 pgd_mop_up_pmds(mm, pgd);
401 pgd_dtor(pgd);
402 paravirt_pgd_free(mm, pgd);
403 _pgd_free(pgd);
404}
405
406/*
407 * Used to set accessed or dirty bits in the page table entries
408 * on other architectures. On x86, the accessed and dirty bits
409 * are tracked by hardware. However, do_wp_page calls this function
410 * to also make the pte writeable at the same time the dirty bit is
411 * set. In that case we do actually need to write the PTE.
412 */
413int ptep_set_access_flags(struct vm_area_struct *vma,
414 unsigned long address, pte_t *ptep,
415 pte_t entry, int dirty)
416{
417 int changed = !pte_same(*ptep, entry);
418
419 if (changed && dirty) {
420 *ptep = entry;
421 pte_update(vma->vm_mm, address, ptep);
422 }
423
424 return changed;
425}
426
427#ifdef CONFIG_TRANSPARENT_HUGEPAGE
428int pmdp_set_access_flags(struct vm_area_struct *vma,
429 unsigned long address, pmd_t *pmdp,
430 pmd_t entry, int dirty)
431{
432 int changed = !pmd_same(*pmdp, entry);
433
434 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
435
436 if (changed && dirty) {
437 *pmdp = entry;
438 /*
439 * We had a write-protection fault here and changed the pmd
440 * to to more permissive. No need to flush the TLB for that,
441 * #PF is architecturally guaranteed to do that and in the
442 * worst-case we'll generate a spurious fault.
443 */
444 }
445
446 return changed;
447}
448#endif
449
450int ptep_test_and_clear_young(struct vm_area_struct *vma,
451 unsigned long addr, pte_t *ptep)
452{
453 int ret = 0;
454
455 if (pte_young(*ptep))
456 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
457 (unsigned long *) &ptep->pte);
458
459 if (ret)
460 pte_update(vma->vm_mm, addr, ptep);
461
462 return ret;
463}
464
465#ifdef CONFIG_TRANSPARENT_HUGEPAGE
466int pmdp_test_and_clear_young(struct vm_area_struct *vma,
467 unsigned long addr, pmd_t *pmdp)
468{
469 int ret = 0;
470
471 if (pmd_young(*pmdp))
472 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
473 (unsigned long *)pmdp);
474
475 return ret;
476}
477#endif
478
479int ptep_clear_flush_young(struct vm_area_struct *vma,
480 unsigned long address, pte_t *ptep)
481{
482 /*
483 * On x86 CPUs, clearing the accessed bit without a TLB flush
484 * doesn't cause data corruption. [ It could cause incorrect
485 * page aging and the (mistaken) reclaim of hot pages, but the
486 * chance of that should be relatively low. ]
487 *
488 * So as a performance optimization don't flush the TLB when
489 * clearing the accessed bit, it will eventually be flushed by
490 * a context switch or a VM operation anyway. [ In the rare
491 * event of it not getting flushed for a long time the delay
492 * shouldn't really matter because there's no real memory
493 * pressure for swapout to react to. ]
494 */
495 return ptep_test_and_clear_young(vma, address, ptep);
496}
497
498#ifdef CONFIG_TRANSPARENT_HUGEPAGE
499int pmdp_clear_flush_young(struct vm_area_struct *vma,
500 unsigned long address, pmd_t *pmdp)
501{
502 int young;
503
504 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
505
506 young = pmdp_test_and_clear_young(vma, address, pmdp);
507 if (young)
508 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
509
510 return young;
511}
512#endif
513
514/**
515 * reserve_top_address - reserves a hole in the top of kernel address space
516 * @reserve - size of hole to reserve
517 *
518 * Can be used to relocate the fixmap area and poke a hole in the top
519 * of kernel address space to make room for a hypervisor.
520 */
521void __init reserve_top_address(unsigned long reserve)
522{
523#ifdef CONFIG_X86_32
524 BUG_ON(fixmaps_set > 0);
525 __FIXADDR_TOP = round_down(-reserve, 1 << PMD_SHIFT) - PAGE_SIZE;
526 printk(KERN_INFO "Reserving virtual address space above 0x%08lx (rounded to 0x%08lx)\n",
527 -reserve, __FIXADDR_TOP + PAGE_SIZE);
528#endif
529}
530
531int fixmaps_set;
532
533void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
534{
535 unsigned long address = __fix_to_virt(idx);
536
537 if (idx >= __end_of_fixed_addresses) {
538 BUG();
539 return;
540 }
541 set_pte_vaddr(address, pte);
542 fixmaps_set++;
543}
544
545void native_set_fixmap(enum fixed_addresses idx, phys_addr_t phys,
546 pgprot_t flags)
547{
548 __native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
549}
550
551#ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
552/**
553 * pud_set_huge - setup kernel PUD mapping
554 *
555 * MTRRs can override PAT memory types with 4KiB granularity. Therefore, this
556 * function sets up a huge page only if any of the following conditions are met:
557 *
558 * - MTRRs are disabled, or
559 *
560 * - MTRRs are enabled and the range is completely covered by a single MTRR, or
561 *
562 * - MTRRs are enabled and the corresponding MTRR memory type is WB, which
563 * has no effect on the requested PAT memory type.
564 *
565 * Callers should try to decrease page size (1GB -> 2MB -> 4K) if the bigger
566 * page mapping attempt fails.
567 *
568 * Returns 1 on success and 0 on failure.
569 */
570int pud_set_huge(pud_t *pud, phys_addr_t addr, pgprot_t prot)
571{
572 u8 mtrr, uniform;
573
574 mtrr = mtrr_type_lookup(addr, addr + PUD_SIZE, &uniform);
575 if ((mtrr != MTRR_TYPE_INVALID) && (!uniform) &&
576 (mtrr != MTRR_TYPE_WRBACK))
577 return 0;
578
579 prot = pgprot_4k_2_large(prot);
580
581 set_pte((pte_t *)pud, pfn_pte(
582 (u64)addr >> PAGE_SHIFT,
583 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
584
585 return 1;
586}
587
588/**
589 * pmd_set_huge - setup kernel PMD mapping
590 *
591 * See text over pud_set_huge() above.
592 *
593 * Returns 1 on success and 0 on failure.
594 */
595int pmd_set_huge(pmd_t *pmd, phys_addr_t addr, pgprot_t prot)
596{
597 u8 mtrr, uniform;
598
599 mtrr = mtrr_type_lookup(addr, addr + PMD_SIZE, &uniform);
600 if ((mtrr != MTRR_TYPE_INVALID) && (!uniform) &&
601 (mtrr != MTRR_TYPE_WRBACK)) {
602 pr_warn_once("%s: Cannot satisfy [mem %#010llx-%#010llx] with a huge-page mapping due to MTRR override.\n",
603 __func__, addr, addr + PMD_SIZE);
604 return 0;
605 }
606
607 prot = pgprot_4k_2_large(prot);
608
609 set_pte((pte_t *)pmd, pfn_pte(
610 (u64)addr >> PAGE_SHIFT,
611 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
612
613 return 1;
614}
615
616/**
617 * pud_clear_huge - clear kernel PUD mapping when it is set
618 *
619 * Returns 1 on success and 0 on failure (no PUD map is found).
620 */
621int pud_clear_huge(pud_t *pud)
622{
623 if (pud_large(*pud)) {
624 pud_clear(pud);
625 return 1;
626 }
627
628 return 0;
629}
630
631/**
632 * pmd_clear_huge - clear kernel PMD mapping when it is set
633 *
634 * Returns 1 on success and 0 on failure (no PMD map is found).
635 */
636int pmd_clear_huge(pmd_t *pmd)
637{
638 if (pmd_large(*pmd)) {
639 pmd_clear(pmd);
640 return 1;
641 }
642
643 return 0;
644}
645#endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */