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// SPDX-License-Identifier: GPL-2.0
2#include <linux/mm.h>
3#include <linux/gfp.h>
4#include <linux/hugetlb.h>
5#include <asm/pgalloc.h>
6#include <asm/pgtable.h>
7#include <asm/tlb.h>
8#include <asm/fixmap.h>
9#include <asm/mtrr.h>
10
11#ifdef CONFIG_DYNAMIC_PHYSICAL_MASK
12phys_addr_t physical_mask __ro_after_init = (1ULL << __PHYSICAL_MASK_SHIFT) - 1;
13EXPORT_SYMBOL(physical_mask);
14#endif
15
16#ifdef CONFIG_HIGHPTE
17#define PGTABLE_HIGHMEM __GFP_HIGHMEM
18#else
19#define PGTABLE_HIGHMEM 0
20#endif
21
22gfp_t __userpte_alloc_gfp = GFP_PGTABLE_USER | PGTABLE_HIGHMEM;
23
24pgtable_t pte_alloc_one(struct mm_struct *mm)
25{
26 return __pte_alloc_one(mm, __userpte_alloc_gfp);
27}
28
29static int __init setup_userpte(char *arg)
30{
31 if (!arg)
32 return -EINVAL;
33
34 /*
35 * "userpte=nohigh" disables allocation of user pagetables in
36 * high memory.
37 */
38 if (strcmp(arg, "nohigh") == 0)
39 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
40 else
41 return -EINVAL;
42 return 0;
43}
44early_param("userpte", setup_userpte);
45
46void ___pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
47{
48 pgtable_pte_page_dtor(pte);
49 paravirt_release_pte(page_to_pfn(pte));
50 paravirt_tlb_remove_table(tlb, pte);
51}
52
53#if CONFIG_PGTABLE_LEVELS > 2
54void ___pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
55{
56 struct page *page = virt_to_page(pmd);
57 paravirt_release_pmd(__pa(pmd) >> PAGE_SHIFT);
58 /*
59 * NOTE! For PAE, any changes to the top page-directory-pointer-table
60 * entries need a full cr3 reload to flush.
61 */
62#ifdef CONFIG_X86_PAE
63 tlb->need_flush_all = 1;
64#endif
65 pgtable_pmd_page_dtor(page);
66 paravirt_tlb_remove_table(tlb, page);
67}
68
69#if CONFIG_PGTABLE_LEVELS > 3
70void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
71{
72 paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
73 paravirt_tlb_remove_table(tlb, virt_to_page(pud));
74}
75
76#if CONFIG_PGTABLE_LEVELS > 4
77void ___p4d_free_tlb(struct mmu_gather *tlb, p4d_t *p4d)
78{
79 paravirt_release_p4d(__pa(p4d) >> PAGE_SHIFT);
80 paravirt_tlb_remove_table(tlb, virt_to_page(p4d));
81}
82#endif /* CONFIG_PGTABLE_LEVELS > 4 */
83#endif /* CONFIG_PGTABLE_LEVELS > 3 */
84#endif /* CONFIG_PGTABLE_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#define MAX_UNSHARED_PTRS_PER_PGD \
103 max_t(size_t, KERNEL_PGD_BOUNDARY, PTRS_PER_PGD)
104
105
106static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
107{
108 virt_to_page(pgd)->pt_mm = mm;
109}
110
111struct mm_struct *pgd_page_get_mm(struct page *page)
112{
113 return page->pt_mm;
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#define MAX_PREALLOCATED_PMDS MAX_UNSHARED_PTRS_PER_PGD
171
172/*
173 * We allocate separate PMDs for the kernel part of the user page-table
174 * when PTI is enabled. We need them to map the per-process LDT into the
175 * user-space page-table.
176 */
177#define PREALLOCATED_USER_PMDS (boot_cpu_has(X86_FEATURE_PTI) ? \
178 KERNEL_PGD_PTRS : 0)
179#define MAX_PREALLOCATED_USER_PMDS KERNEL_PGD_PTRS
180
181void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
182{
183 paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
184
185 /* Note: almost everything apart from _PAGE_PRESENT is
186 reserved at the pmd (PDPT) level. */
187 set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
188
189 /*
190 * According to Intel App note "TLBs, Paging-Structure Caches,
191 * and Their Invalidation", April 2007, document 317080-001,
192 * section 8.1: in PAE mode we explicitly have to flush the
193 * TLB via cr3 if the top-level pgd is changed...
194 */
195 flush_tlb_mm(mm);
196}
197#else /* !CONFIG_X86_PAE */
198
199/* No need to prepopulate any pagetable entries in non-PAE modes. */
200#define PREALLOCATED_PMDS 0
201#define MAX_PREALLOCATED_PMDS 0
202#define PREALLOCATED_USER_PMDS 0
203#define MAX_PREALLOCATED_USER_PMDS 0
204#endif /* CONFIG_X86_PAE */
205
206static void free_pmds(struct mm_struct *mm, pmd_t *pmds[], int count)
207{
208 int i;
209
210 for (i = 0; i < count; i++)
211 if (pmds[i]) {
212 pgtable_pmd_page_dtor(virt_to_page(pmds[i]));
213 free_page((unsigned long)pmds[i]);
214 mm_dec_nr_pmds(mm);
215 }
216}
217
218static int preallocate_pmds(struct mm_struct *mm, pmd_t *pmds[], int count)
219{
220 int i;
221 bool failed = false;
222 gfp_t gfp = GFP_PGTABLE_USER;
223
224 if (mm == &init_mm)
225 gfp &= ~__GFP_ACCOUNT;
226
227 for (i = 0; i < count; i++) {
228 pmd_t *pmd = (pmd_t *)__get_free_page(gfp);
229 if (!pmd)
230 failed = true;
231 if (pmd && !pgtable_pmd_page_ctor(virt_to_page(pmd))) {
232 free_page((unsigned long)pmd);
233 pmd = NULL;
234 failed = true;
235 }
236 if (pmd)
237 mm_inc_nr_pmds(mm);
238 pmds[i] = pmd;
239 }
240
241 if (failed) {
242 free_pmds(mm, pmds, count);
243 return -ENOMEM;
244 }
245
246 return 0;
247}
248
249/*
250 * Mop up any pmd pages which may still be attached to the pgd.
251 * Normally they will be freed by munmap/exit_mmap, but any pmd we
252 * preallocate which never got a corresponding vma will need to be
253 * freed manually.
254 */
255static void mop_up_one_pmd(struct mm_struct *mm, pgd_t *pgdp)
256{
257 pgd_t pgd = *pgdp;
258
259 if (pgd_val(pgd) != 0) {
260 pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
261
262 pgd_clear(pgdp);
263
264 paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
265 pmd_free(mm, pmd);
266 mm_dec_nr_pmds(mm);
267 }
268}
269
270static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
271{
272 int i;
273
274 for (i = 0; i < PREALLOCATED_PMDS; i++)
275 mop_up_one_pmd(mm, &pgdp[i]);
276
277#ifdef CONFIG_PAGE_TABLE_ISOLATION
278
279 if (!boot_cpu_has(X86_FEATURE_PTI))
280 return;
281
282 pgdp = kernel_to_user_pgdp(pgdp);
283
284 for (i = 0; i < PREALLOCATED_USER_PMDS; i++)
285 mop_up_one_pmd(mm, &pgdp[i + KERNEL_PGD_BOUNDARY]);
286#endif
287}
288
289static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
290{
291 p4d_t *p4d;
292 pud_t *pud;
293 int i;
294
295 if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
296 return;
297
298 p4d = p4d_offset(pgd, 0);
299 pud = pud_offset(p4d, 0);
300
301 for (i = 0; i < PREALLOCATED_PMDS; i++, pud++) {
302 pmd_t *pmd = pmds[i];
303
304 if (i >= KERNEL_PGD_BOUNDARY)
305 memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
306 sizeof(pmd_t) * PTRS_PER_PMD);
307
308 pud_populate(mm, pud, pmd);
309 }
310}
311
312#ifdef CONFIG_PAGE_TABLE_ISOLATION
313static void pgd_prepopulate_user_pmd(struct mm_struct *mm,
314 pgd_t *k_pgd, pmd_t *pmds[])
315{
316 pgd_t *s_pgd = kernel_to_user_pgdp(swapper_pg_dir);
317 pgd_t *u_pgd = kernel_to_user_pgdp(k_pgd);
318 p4d_t *u_p4d;
319 pud_t *u_pud;
320 int i;
321
322 u_p4d = p4d_offset(u_pgd, 0);
323 u_pud = pud_offset(u_p4d, 0);
324
325 s_pgd += KERNEL_PGD_BOUNDARY;
326 u_pud += KERNEL_PGD_BOUNDARY;
327
328 for (i = 0; i < PREALLOCATED_USER_PMDS; i++, u_pud++, s_pgd++) {
329 pmd_t *pmd = pmds[i];
330
331 memcpy(pmd, (pmd_t *)pgd_page_vaddr(*s_pgd),
332 sizeof(pmd_t) * PTRS_PER_PMD);
333
334 pud_populate(mm, u_pud, pmd);
335 }
336
337}
338#else
339static void pgd_prepopulate_user_pmd(struct mm_struct *mm,
340 pgd_t *k_pgd, pmd_t *pmds[])
341{
342}
343#endif
344/*
345 * Xen paravirt assumes pgd table should be in one page. 64 bit kernel also
346 * assumes that pgd should be in one page.
347 *
348 * But kernel with PAE paging that is not running as a Xen domain
349 * only needs to allocate 32 bytes for pgd instead of one page.
350 */
351#ifdef CONFIG_X86_PAE
352
353#include <linux/slab.h>
354
355#define PGD_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
356#define PGD_ALIGN 32
357
358static struct kmem_cache *pgd_cache;
359
360void __init pgtable_cache_init(void)
361{
362 /*
363 * When PAE kernel is running as a Xen domain, it does not use
364 * shared kernel pmd. And this requires a whole page for pgd.
365 */
366 if (!SHARED_KERNEL_PMD)
367 return;
368
369 /*
370 * when PAE kernel is not running as a Xen domain, it uses
371 * shared kernel pmd. Shared kernel pmd does not require a whole
372 * page for pgd. We are able to just allocate a 32-byte for pgd.
373 * During boot time, we create a 32-byte slab for pgd table allocation.
374 */
375 pgd_cache = kmem_cache_create("pgd_cache", PGD_SIZE, PGD_ALIGN,
376 SLAB_PANIC, NULL);
377}
378
379static inline pgd_t *_pgd_alloc(void)
380{
381 /*
382 * If no SHARED_KERNEL_PMD, PAE kernel is running as a Xen domain.
383 * We allocate one page for pgd.
384 */
385 if (!SHARED_KERNEL_PMD)
386 return (pgd_t *)__get_free_pages(GFP_PGTABLE_USER,
387 PGD_ALLOCATION_ORDER);
388
389 /*
390 * Now PAE kernel is not running as a Xen domain. We can allocate
391 * a 32-byte slab for pgd to save memory space.
392 */
393 return kmem_cache_alloc(pgd_cache, GFP_PGTABLE_USER);
394}
395
396static inline void _pgd_free(pgd_t *pgd)
397{
398 if (!SHARED_KERNEL_PMD)
399 free_pages((unsigned long)pgd, PGD_ALLOCATION_ORDER);
400 else
401 kmem_cache_free(pgd_cache, pgd);
402}
403#else
404
405static inline pgd_t *_pgd_alloc(void)
406{
407 return (pgd_t *)__get_free_pages(GFP_PGTABLE_USER,
408 PGD_ALLOCATION_ORDER);
409}
410
411static inline void _pgd_free(pgd_t *pgd)
412{
413 free_pages((unsigned long)pgd, PGD_ALLOCATION_ORDER);
414}
415#endif /* CONFIG_X86_PAE */
416
417pgd_t *pgd_alloc(struct mm_struct *mm)
418{
419 pgd_t *pgd;
420 pmd_t *u_pmds[MAX_PREALLOCATED_USER_PMDS];
421 pmd_t *pmds[MAX_PREALLOCATED_PMDS];
422
423 pgd = _pgd_alloc();
424
425 if (pgd == NULL)
426 goto out;
427
428 mm->pgd = pgd;
429
430 if (preallocate_pmds(mm, pmds, PREALLOCATED_PMDS) != 0)
431 goto out_free_pgd;
432
433 if (preallocate_pmds(mm, u_pmds, PREALLOCATED_USER_PMDS) != 0)
434 goto out_free_pmds;
435
436 if (paravirt_pgd_alloc(mm) != 0)
437 goto out_free_user_pmds;
438
439 /*
440 * Make sure that pre-populating the pmds is atomic with
441 * respect to anything walking the pgd_list, so that they
442 * never see a partially populated pgd.
443 */
444 spin_lock(&pgd_lock);
445
446 pgd_ctor(mm, pgd);
447 pgd_prepopulate_pmd(mm, pgd, pmds);
448 pgd_prepopulate_user_pmd(mm, pgd, u_pmds);
449
450 spin_unlock(&pgd_lock);
451
452 return pgd;
453
454out_free_user_pmds:
455 free_pmds(mm, u_pmds, PREALLOCATED_USER_PMDS);
456out_free_pmds:
457 free_pmds(mm, pmds, PREALLOCATED_PMDS);
458out_free_pgd:
459 _pgd_free(pgd);
460out:
461 return NULL;
462}
463
464void pgd_free(struct mm_struct *mm, pgd_t *pgd)
465{
466 pgd_mop_up_pmds(mm, pgd);
467 pgd_dtor(pgd);
468 paravirt_pgd_free(mm, pgd);
469 _pgd_free(pgd);
470}
471
472/*
473 * Used to set accessed or dirty bits in the page table entries
474 * on other architectures. On x86, the accessed and dirty bits
475 * are tracked by hardware. However, do_wp_page calls this function
476 * to also make the pte writeable at the same time the dirty bit is
477 * set. In that case we do actually need to write the PTE.
478 */
479int ptep_set_access_flags(struct vm_area_struct *vma,
480 unsigned long address, pte_t *ptep,
481 pte_t entry, int dirty)
482{
483 int changed = !pte_same(*ptep, entry);
484
485 if (changed && dirty)
486 set_pte(ptep, entry);
487
488 return changed;
489}
490
491#ifdef CONFIG_TRANSPARENT_HUGEPAGE
492int pmdp_set_access_flags(struct vm_area_struct *vma,
493 unsigned long address, pmd_t *pmdp,
494 pmd_t entry, int dirty)
495{
496 int changed = !pmd_same(*pmdp, entry);
497
498 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
499
500 if (changed && dirty) {
501 set_pmd(pmdp, entry);
502 /*
503 * We had a write-protection fault here and changed the pmd
504 * to to more permissive. No need to flush the TLB for that,
505 * #PF is architecturally guaranteed to do that and in the
506 * worst-case we'll generate a spurious fault.
507 */
508 }
509
510 return changed;
511}
512
513int pudp_set_access_flags(struct vm_area_struct *vma, unsigned long address,
514 pud_t *pudp, pud_t entry, int dirty)
515{
516 int changed = !pud_same(*pudp, entry);
517
518 VM_BUG_ON(address & ~HPAGE_PUD_MASK);
519
520 if (changed && dirty) {
521 set_pud(pudp, entry);
522 /*
523 * We had a write-protection fault here and changed the pud
524 * to to more permissive. No need to flush the TLB for that,
525 * #PF is architecturally guaranteed to do that and in the
526 * worst-case we'll generate a spurious fault.
527 */
528 }
529
530 return changed;
531}
532#endif
533
534int ptep_test_and_clear_young(struct vm_area_struct *vma,
535 unsigned long addr, pte_t *ptep)
536{
537 int ret = 0;
538
539 if (pte_young(*ptep))
540 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
541 (unsigned long *) &ptep->pte);
542
543 return ret;
544}
545
546#ifdef CONFIG_TRANSPARENT_HUGEPAGE
547int pmdp_test_and_clear_young(struct vm_area_struct *vma,
548 unsigned long addr, pmd_t *pmdp)
549{
550 int ret = 0;
551
552 if (pmd_young(*pmdp))
553 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
554 (unsigned long *)pmdp);
555
556 return ret;
557}
558int pudp_test_and_clear_young(struct vm_area_struct *vma,
559 unsigned long addr, pud_t *pudp)
560{
561 int ret = 0;
562
563 if (pud_young(*pudp))
564 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
565 (unsigned long *)pudp);
566
567 return ret;
568}
569#endif
570
571int ptep_clear_flush_young(struct vm_area_struct *vma,
572 unsigned long address, pte_t *ptep)
573{
574 /*
575 * On x86 CPUs, clearing the accessed bit without a TLB flush
576 * doesn't cause data corruption. [ It could cause incorrect
577 * page aging and the (mistaken) reclaim of hot pages, but the
578 * chance of that should be relatively low. ]
579 *
580 * So as a performance optimization don't flush the TLB when
581 * clearing the accessed bit, it will eventually be flushed by
582 * a context switch or a VM operation anyway. [ In the rare
583 * event of it not getting flushed for a long time the delay
584 * shouldn't really matter because there's no real memory
585 * pressure for swapout to react to. ]
586 */
587 return ptep_test_and_clear_young(vma, address, ptep);
588}
589
590#ifdef CONFIG_TRANSPARENT_HUGEPAGE
591int pmdp_clear_flush_young(struct vm_area_struct *vma,
592 unsigned long address, pmd_t *pmdp)
593{
594 int young;
595
596 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
597
598 young = pmdp_test_and_clear_young(vma, address, pmdp);
599 if (young)
600 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
601
602 return young;
603}
604#endif
605
606/**
607 * reserve_top_address - reserves a hole in the top of kernel address space
608 * @reserve - size of hole to reserve
609 *
610 * Can be used to relocate the fixmap area and poke a hole in the top
611 * of kernel address space to make room for a hypervisor.
612 */
613void __init reserve_top_address(unsigned long reserve)
614{
615#ifdef CONFIG_X86_32
616 BUG_ON(fixmaps_set > 0);
617 __FIXADDR_TOP = round_down(-reserve, 1 << PMD_SHIFT) - PAGE_SIZE;
618 printk(KERN_INFO "Reserving virtual address space above 0x%08lx (rounded to 0x%08lx)\n",
619 -reserve, __FIXADDR_TOP + PAGE_SIZE);
620#endif
621}
622
623int fixmaps_set;
624
625void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
626{
627 unsigned long address = __fix_to_virt(idx);
628
629#ifdef CONFIG_X86_64
630 /*
631 * Ensure that the static initial page tables are covering the
632 * fixmap completely.
633 */
634 BUILD_BUG_ON(__end_of_permanent_fixed_addresses >
635 (FIXMAP_PMD_NUM * PTRS_PER_PTE));
636#endif
637
638 if (idx >= __end_of_fixed_addresses) {
639 BUG();
640 return;
641 }
642 set_pte_vaddr(address, pte);
643 fixmaps_set++;
644}
645
646void native_set_fixmap(enum fixed_addresses idx, phys_addr_t phys,
647 pgprot_t flags)
648{
649 /* Sanitize 'prot' against any unsupported bits: */
650 pgprot_val(flags) &= __default_kernel_pte_mask;
651
652 __native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
653}
654
655#ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
656#ifdef CONFIG_X86_5LEVEL
657/**
658 * p4d_set_huge - setup kernel P4D mapping
659 *
660 * No 512GB pages yet -- always return 0
661 */
662int p4d_set_huge(p4d_t *p4d, phys_addr_t addr, pgprot_t prot)
663{
664 return 0;
665}
666
667/**
668 * p4d_clear_huge - clear kernel P4D mapping when it is set
669 *
670 * No 512GB pages yet -- always return 0
671 */
672int p4d_clear_huge(p4d_t *p4d)
673{
674 return 0;
675}
676#endif
677
678/**
679 * pud_set_huge - setup kernel PUD mapping
680 *
681 * MTRRs can override PAT memory types with 4KiB granularity. Therefore, this
682 * function sets up a huge page only if any of the following conditions are met:
683 *
684 * - MTRRs are disabled, or
685 *
686 * - MTRRs are enabled and the range is completely covered by a single MTRR, or
687 *
688 * - MTRRs are enabled and the corresponding MTRR memory type is WB, which
689 * has no effect on the requested PAT memory type.
690 *
691 * Callers should try to decrease page size (1GB -> 2MB -> 4K) if the bigger
692 * page mapping attempt fails.
693 *
694 * Returns 1 on success and 0 on failure.
695 */
696int pud_set_huge(pud_t *pud, phys_addr_t addr, pgprot_t prot)
697{
698 u8 mtrr, uniform;
699
700 mtrr = mtrr_type_lookup(addr, addr + PUD_SIZE, &uniform);
701 if ((mtrr != MTRR_TYPE_INVALID) && (!uniform) &&
702 (mtrr != MTRR_TYPE_WRBACK))
703 return 0;
704
705 /* Bail out if we are we on a populated non-leaf entry: */
706 if (pud_present(*pud) && !pud_huge(*pud))
707 return 0;
708
709 prot = pgprot_4k_2_large(prot);
710
711 set_pte((pte_t *)pud, pfn_pte(
712 (u64)addr >> PAGE_SHIFT,
713 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
714
715 return 1;
716}
717
718/**
719 * pmd_set_huge - setup kernel PMD mapping
720 *
721 * See text over pud_set_huge() above.
722 *
723 * Returns 1 on success and 0 on failure.
724 */
725int pmd_set_huge(pmd_t *pmd, phys_addr_t addr, pgprot_t prot)
726{
727 u8 mtrr, uniform;
728
729 mtrr = mtrr_type_lookup(addr, addr + PMD_SIZE, &uniform);
730 if ((mtrr != MTRR_TYPE_INVALID) && (!uniform) &&
731 (mtrr != MTRR_TYPE_WRBACK)) {
732 pr_warn_once("%s: Cannot satisfy [mem %#010llx-%#010llx] with a huge-page mapping due to MTRR override.\n",
733 __func__, addr, addr + PMD_SIZE);
734 return 0;
735 }
736
737 /* Bail out if we are we on a populated non-leaf entry: */
738 if (pmd_present(*pmd) && !pmd_huge(*pmd))
739 return 0;
740
741 prot = pgprot_4k_2_large(prot);
742
743 set_pte((pte_t *)pmd, pfn_pte(
744 (u64)addr >> PAGE_SHIFT,
745 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
746
747 return 1;
748}
749
750/**
751 * pud_clear_huge - clear kernel PUD mapping when it is set
752 *
753 * Returns 1 on success and 0 on failure (no PUD map is found).
754 */
755int pud_clear_huge(pud_t *pud)
756{
757 if (pud_large(*pud)) {
758 pud_clear(pud);
759 return 1;
760 }
761
762 return 0;
763}
764
765/**
766 * pmd_clear_huge - clear kernel PMD mapping when it is set
767 *
768 * Returns 1 on success and 0 on failure (no PMD map is found).
769 */
770int pmd_clear_huge(pmd_t *pmd)
771{
772 if (pmd_large(*pmd)) {
773 pmd_clear(pmd);
774 return 1;
775 }
776
777 return 0;
778}
779
780/*
781 * Until we support 512GB pages, skip them in the vmap area.
782 */
783int p4d_free_pud_page(p4d_t *p4d, unsigned long addr)
784{
785 return 0;
786}
787
788#ifdef CONFIG_X86_64
789/**
790 * pud_free_pmd_page - Clear pud entry and free pmd page.
791 * @pud: Pointer to a PUD.
792 * @addr: Virtual address associated with pud.
793 *
794 * Context: The pud range has been unmapped and TLB purged.
795 * Return: 1 if clearing the entry succeeded. 0 otherwise.
796 *
797 * NOTE: Callers must allow a single page allocation.
798 */
799int pud_free_pmd_page(pud_t *pud, unsigned long addr)
800{
801 pmd_t *pmd, *pmd_sv;
802 pte_t *pte;
803 int i;
804
805 pmd = (pmd_t *)pud_page_vaddr(*pud);
806 pmd_sv = (pmd_t *)__get_free_page(GFP_KERNEL);
807 if (!pmd_sv)
808 return 0;
809
810 for (i = 0; i < PTRS_PER_PMD; i++) {
811 pmd_sv[i] = pmd[i];
812 if (!pmd_none(pmd[i]))
813 pmd_clear(&pmd[i]);
814 }
815
816 pud_clear(pud);
817
818 /* INVLPG to clear all paging-structure caches */
819 flush_tlb_kernel_range(addr, addr + PAGE_SIZE-1);
820
821 for (i = 0; i < PTRS_PER_PMD; i++) {
822 if (!pmd_none(pmd_sv[i])) {
823 pte = (pte_t *)pmd_page_vaddr(pmd_sv[i]);
824 free_page((unsigned long)pte);
825 }
826 }
827
828 free_page((unsigned long)pmd_sv);
829 free_page((unsigned long)pmd);
830
831 return 1;
832}
833
834/**
835 * pmd_free_pte_page - Clear pmd entry and free pte page.
836 * @pmd: Pointer to a PMD.
837 * @addr: Virtual address associated with pmd.
838 *
839 * Context: The pmd range has been unmapped and TLB purged.
840 * Return: 1 if clearing the entry succeeded. 0 otherwise.
841 */
842int pmd_free_pte_page(pmd_t *pmd, unsigned long addr)
843{
844 pte_t *pte;
845
846 pte = (pte_t *)pmd_page_vaddr(*pmd);
847 pmd_clear(pmd);
848
849 /* INVLPG to clear all paging-structure caches */
850 flush_tlb_kernel_range(addr, addr + PAGE_SIZE-1);
851
852 free_page((unsigned long)pte);
853
854 return 1;
855}
856
857#else /* !CONFIG_X86_64 */
858
859int pud_free_pmd_page(pud_t *pud, unsigned long addr)
860{
861 return pud_none(*pud);
862}
863
864/*
865 * Disable free page handling on x86-PAE. This assures that ioremap()
866 * does not update sync'd pmd entries. See vmalloc_sync_one().
867 */
868int pmd_free_pte_page(pmd_t *pmd, unsigned long addr)
869{
870 return pmd_none(*pmd);
871}
872
873#endif /* CONFIG_X86_64 */
874#endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */