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 pgtable_page_ctor(pte);
30 return pte;
31}
32
33static int __init setup_userpte(char *arg)
34{
35 if (!arg)
36 return -EINVAL;
37
38 /*
39 * "userpte=nohigh" disables allocation of user pagetables in
40 * high memory.
41 */
42 if (strcmp(arg, "nohigh") == 0)
43 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
44 else
45 return -EINVAL;
46 return 0;
47}
48early_param("userpte", setup_userpte);
49
50void ___pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
51{
52 pgtable_page_dtor(pte);
53 paravirt_release_pte(page_to_pfn(pte));
54 tlb_remove_page(tlb, pte);
55}
56
57#if PAGETABLE_LEVELS > 2
58void ___pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
59{
60 paravirt_release_pmd(__pa(pmd) >> PAGE_SHIFT);
61 tlb_remove_page(tlb, virt_to_page(pmd));
62}
63
64#if PAGETABLE_LEVELS > 3
65void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
66{
67 paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
68 tlb_remove_page(tlb, virt_to_page(pud));
69}
70#endif /* PAGETABLE_LEVELS > 3 */
71#endif /* PAGETABLE_LEVELS > 2 */
72
73static inline void pgd_list_add(pgd_t *pgd)
74{
75 struct page *page = virt_to_page(pgd);
76
77 list_add(&page->lru, &pgd_list);
78}
79
80static inline void pgd_list_del(pgd_t *pgd)
81{
82 struct page *page = virt_to_page(pgd);
83
84 list_del(&page->lru);
85}
86
87#define UNSHARED_PTRS_PER_PGD \
88 (SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
89
90
91static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
92{
93 BUILD_BUG_ON(sizeof(virt_to_page(pgd)->index) < sizeof(mm));
94 virt_to_page(pgd)->index = (pgoff_t)mm;
95}
96
97struct mm_struct *pgd_page_get_mm(struct page *page)
98{
99 return (struct mm_struct *)page->index;
100}
101
102static void pgd_ctor(struct mm_struct *mm, pgd_t *pgd)
103{
104 /* If the pgd points to a shared pagetable level (either the
105 ptes in non-PAE, or shared PMD in PAE), then just copy the
106 references from swapper_pg_dir. */
107 if (PAGETABLE_LEVELS == 2 ||
108 (PAGETABLE_LEVELS == 3 && SHARED_KERNEL_PMD) ||
109 PAGETABLE_LEVELS == 4) {
110 clone_pgd_range(pgd + KERNEL_PGD_BOUNDARY,
111 swapper_pg_dir + KERNEL_PGD_BOUNDARY,
112 KERNEL_PGD_PTRS);
113 }
114
115 /* list required to sync kernel mapping updates */
116 if (!SHARED_KERNEL_PMD) {
117 pgd_set_mm(pgd, mm);
118 pgd_list_add(pgd);
119 }
120}
121
122static void pgd_dtor(pgd_t *pgd)
123{
124 if (SHARED_KERNEL_PMD)
125 return;
126
127 spin_lock(&pgd_lock);
128 pgd_list_del(pgd);
129 spin_unlock(&pgd_lock);
130}
131
132/*
133 * List of all pgd's needed for non-PAE so it can invalidate entries
134 * in both cached and uncached pgd's; not needed for PAE since the
135 * kernel pmd is shared. If PAE were not to share the pmd a similar
136 * tactic would be needed. This is essentially codepath-based locking
137 * against pageattr.c; it is the unique case in which a valid change
138 * of kernel pagetables can't be lazily synchronized by vmalloc faults.
139 * vmalloc faults work because attached pagetables are never freed.
140 * -- wli
141 */
142
143#ifdef CONFIG_X86_PAE
144/*
145 * In PAE mode, we need to do a cr3 reload (=tlb flush) when
146 * updating the top-level pagetable entries to guarantee the
147 * processor notices the update. Since this is expensive, and
148 * all 4 top-level entries are used almost immediately in a
149 * new process's life, we just pre-populate them here.
150 *
151 * Also, if we're in a paravirt environment where the kernel pmd is
152 * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
153 * and initialize the kernel pmds here.
154 */
155#define PREALLOCATED_PMDS UNSHARED_PTRS_PER_PGD
156
157void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
158{
159 paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
160
161 /* Note: almost everything apart from _PAGE_PRESENT is
162 reserved at the pmd (PDPT) level. */
163 set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
164
165 /*
166 * According to Intel App note "TLBs, Paging-Structure Caches,
167 * and Their Invalidation", April 2007, document 317080-001,
168 * section 8.1: in PAE mode we explicitly have to flush the
169 * TLB via cr3 if the top-level pgd is changed...
170 */
171 flush_tlb_mm(mm);
172}
173#else /* !CONFIG_X86_PAE */
174
175/* No need to prepopulate any pagetable entries in non-PAE modes. */
176#define PREALLOCATED_PMDS 0
177
178#endif /* CONFIG_X86_PAE */
179
180static void free_pmds(pmd_t *pmds[])
181{
182 int i;
183
184 for(i = 0; i < PREALLOCATED_PMDS; i++)
185 if (pmds[i])
186 free_page((unsigned long)pmds[i]);
187}
188
189static int preallocate_pmds(pmd_t *pmds[])
190{
191 int i;
192 bool failed = false;
193
194 for(i = 0; i < PREALLOCATED_PMDS; i++) {
195 pmd_t *pmd = (pmd_t *)__get_free_page(PGALLOC_GFP);
196 if (pmd == NULL)
197 failed = true;
198 pmds[i] = pmd;
199 }
200
201 if (failed) {
202 free_pmds(pmds);
203 return -ENOMEM;
204 }
205
206 return 0;
207}
208
209/*
210 * Mop up any pmd pages which may still be attached to the pgd.
211 * Normally they will be freed by munmap/exit_mmap, but any pmd we
212 * preallocate which never got a corresponding vma will need to be
213 * freed manually.
214 */
215static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
216{
217 int i;
218
219 for(i = 0; i < PREALLOCATED_PMDS; i++) {
220 pgd_t pgd = pgdp[i];
221
222 if (pgd_val(pgd) != 0) {
223 pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
224
225 pgdp[i] = native_make_pgd(0);
226
227 paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
228 pmd_free(mm, pmd);
229 }
230 }
231}
232
233static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
234{
235 pud_t *pud;
236 unsigned long addr;
237 int i;
238
239 if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
240 return;
241
242 pud = pud_offset(pgd, 0);
243
244 for (addr = i = 0; i < PREALLOCATED_PMDS;
245 i++, pud++, addr += PUD_SIZE) {
246 pmd_t *pmd = pmds[i];
247
248 if (i >= KERNEL_PGD_BOUNDARY)
249 memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
250 sizeof(pmd_t) * PTRS_PER_PMD);
251
252 pud_populate(mm, pud, pmd);
253 }
254}
255
256pgd_t *pgd_alloc(struct mm_struct *mm)
257{
258 pgd_t *pgd;
259 pmd_t *pmds[PREALLOCATED_PMDS];
260
261 pgd = (pgd_t *)__get_free_page(PGALLOC_GFP);
262
263 if (pgd == NULL)
264 goto out;
265
266 mm->pgd = pgd;
267
268 if (preallocate_pmds(pmds) != 0)
269 goto out_free_pgd;
270
271 if (paravirt_pgd_alloc(mm) != 0)
272 goto out_free_pmds;
273
274 /*
275 * Make sure that pre-populating the pmds is atomic with
276 * respect to anything walking the pgd_list, so that they
277 * never see a partially populated pgd.
278 */
279 spin_lock(&pgd_lock);
280
281 pgd_ctor(mm, pgd);
282 pgd_prepopulate_pmd(mm, pgd, pmds);
283
284 spin_unlock(&pgd_lock);
285
286 return pgd;
287
288out_free_pmds:
289 free_pmds(pmds);
290out_free_pgd:
291 free_page((unsigned long)pgd);
292out:
293 return NULL;
294}
295
296void pgd_free(struct mm_struct *mm, pgd_t *pgd)
297{
298 pgd_mop_up_pmds(mm, pgd);
299 pgd_dtor(pgd);
300 paravirt_pgd_free(mm, pgd);
301 free_page((unsigned long)pgd);
302}
303
304int ptep_set_access_flags(struct vm_area_struct *vma,
305 unsigned long address, pte_t *ptep,
306 pte_t entry, int dirty)
307{
308 int changed = !pte_same(*ptep, entry);
309
310 if (changed && dirty) {
311 *ptep = entry;
312 pte_update_defer(vma->vm_mm, address, ptep);
313 flush_tlb_page(vma, address);
314 }
315
316 return changed;
317}
318
319#ifdef CONFIG_TRANSPARENT_HUGEPAGE
320int pmdp_set_access_flags(struct vm_area_struct *vma,
321 unsigned long address, pmd_t *pmdp,
322 pmd_t entry, int dirty)
323{
324 int changed = !pmd_same(*pmdp, entry);
325
326 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
327
328 if (changed && dirty) {
329 *pmdp = entry;
330 pmd_update_defer(vma->vm_mm, address, pmdp);
331 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
332 }
333
334 return changed;
335}
336#endif
337
338int ptep_test_and_clear_young(struct vm_area_struct *vma,
339 unsigned long addr, pte_t *ptep)
340{
341 int ret = 0;
342
343 if (pte_young(*ptep))
344 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
345 (unsigned long *) &ptep->pte);
346
347 if (ret)
348 pte_update(vma->vm_mm, addr, ptep);
349
350 return ret;
351}
352
353#ifdef CONFIG_TRANSPARENT_HUGEPAGE
354int pmdp_test_and_clear_young(struct vm_area_struct *vma,
355 unsigned long addr, pmd_t *pmdp)
356{
357 int ret = 0;
358
359 if (pmd_young(*pmdp))
360 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
361 (unsigned long *)pmdp);
362
363 if (ret)
364 pmd_update(vma->vm_mm, addr, pmdp);
365
366 return ret;
367}
368#endif
369
370int ptep_clear_flush_young(struct vm_area_struct *vma,
371 unsigned long address, pte_t *ptep)
372{
373 int young;
374
375 young = ptep_test_and_clear_young(vma, address, ptep);
376 if (young)
377 flush_tlb_page(vma, address);
378
379 return young;
380}
381
382#ifdef CONFIG_TRANSPARENT_HUGEPAGE
383int pmdp_clear_flush_young(struct vm_area_struct *vma,
384 unsigned long address, pmd_t *pmdp)
385{
386 int young;
387
388 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
389
390 young = pmdp_test_and_clear_young(vma, address, pmdp);
391 if (young)
392 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
393
394 return young;
395}
396
397void pmdp_splitting_flush(struct vm_area_struct *vma,
398 unsigned long address, pmd_t *pmdp)
399{
400 int set;
401 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
402 set = !test_and_set_bit(_PAGE_BIT_SPLITTING,
403 (unsigned long *)pmdp);
404 if (set) {
405 pmd_update(vma->vm_mm, address, pmdp);
406 /* need tlb flush only to serialize against gup-fast */
407 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
408 }
409}
410#endif
411
412/**
413 * reserve_top_address - reserves a hole in the top of kernel address space
414 * @reserve - size of hole to reserve
415 *
416 * Can be used to relocate the fixmap area and poke a hole in the top
417 * of kernel address space to make room for a hypervisor.
418 */
419void __init reserve_top_address(unsigned long reserve)
420{
421#ifdef CONFIG_X86_32
422 BUG_ON(fixmaps_set > 0);
423 printk(KERN_INFO "Reserving virtual address space above 0x%08x\n",
424 (int)-reserve);
425 __FIXADDR_TOP = -reserve - PAGE_SIZE;
426#endif
427}
428
429int fixmaps_set;
430
431void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
432{
433 unsigned long address = __fix_to_virt(idx);
434
435 if (idx >= __end_of_fixed_addresses) {
436 BUG();
437 return;
438 }
439 set_pte_vaddr(address, pte);
440 fixmaps_set++;
441}
442
443void native_set_fixmap(enum fixed_addresses idx, phys_addr_t phys,
444 pgprot_t flags)
445{
446 __native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
447}
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/tlb.h>
7#include <asm/fixmap.h>
8#include <asm/mtrr.h>
9
10#ifdef CONFIG_DYNAMIC_PHYSICAL_MASK
11phys_addr_t physical_mask __ro_after_init = (1ULL << __PHYSICAL_MASK_SHIFT) - 1;
12EXPORT_SYMBOL(physical_mask);
13#endif
14
15#ifdef CONFIG_HIGHPTE
16#define PGTABLE_HIGHMEM __GFP_HIGHMEM
17#else
18#define PGTABLE_HIGHMEM 0
19#endif
20
21#ifndef CONFIG_PARAVIRT
22static inline
23void paravirt_tlb_remove_table(struct mmu_gather *tlb, void *table)
24{
25 tlb_remove_page(tlb, table);
26}
27#endif
28
29gfp_t __userpte_alloc_gfp = GFP_PGTABLE_USER | PGTABLE_HIGHMEM;
30
31pgtable_t pte_alloc_one(struct mm_struct *mm)
32{
33 return __pte_alloc_one(mm, __userpte_alloc_gfp);
34}
35
36static int __init setup_userpte(char *arg)
37{
38 if (!arg)
39 return -EINVAL;
40
41 /*
42 * "userpte=nohigh" disables allocation of user pagetables in
43 * high memory.
44 */
45 if (strcmp(arg, "nohigh") == 0)
46 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
47 else
48 return -EINVAL;
49 return 0;
50}
51early_param("userpte", setup_userpte);
52
53void ___pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
54{
55 pagetable_pte_dtor(page_ptdesc(pte));
56 paravirt_release_pte(page_to_pfn(pte));
57 paravirt_tlb_remove_table(tlb, pte);
58}
59
60#if CONFIG_PGTABLE_LEVELS > 2
61void ___pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
62{
63 struct ptdesc *ptdesc = virt_to_ptdesc(pmd);
64 paravirt_release_pmd(__pa(pmd) >> PAGE_SHIFT);
65 /*
66 * NOTE! For PAE, any changes to the top page-directory-pointer-table
67 * entries need a full cr3 reload to flush.
68 */
69#ifdef CONFIG_X86_PAE
70 tlb->need_flush_all = 1;
71#endif
72 pagetable_pmd_dtor(ptdesc);
73 paravirt_tlb_remove_table(tlb, ptdesc_page(ptdesc));
74}
75
76#if CONFIG_PGTABLE_LEVELS > 3
77void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
78{
79 struct ptdesc *ptdesc = virt_to_ptdesc(pud);
80
81 pagetable_pud_dtor(ptdesc);
82 paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
83 paravirt_tlb_remove_table(tlb, virt_to_page(pud));
84}
85
86#if CONFIG_PGTABLE_LEVELS > 4
87void ___p4d_free_tlb(struct mmu_gather *tlb, p4d_t *p4d)
88{
89 paravirt_release_p4d(__pa(p4d) >> PAGE_SHIFT);
90 paravirt_tlb_remove_table(tlb, virt_to_page(p4d));
91}
92#endif /* CONFIG_PGTABLE_LEVELS > 4 */
93#endif /* CONFIG_PGTABLE_LEVELS > 3 */
94#endif /* CONFIG_PGTABLE_LEVELS > 2 */
95
96static inline void pgd_list_add(pgd_t *pgd)
97{
98 struct ptdesc *ptdesc = virt_to_ptdesc(pgd);
99
100 list_add(&ptdesc->pt_list, &pgd_list);
101}
102
103static inline void pgd_list_del(pgd_t *pgd)
104{
105 struct ptdesc *ptdesc = virt_to_ptdesc(pgd);
106
107 list_del(&ptdesc->pt_list);
108}
109
110#define UNSHARED_PTRS_PER_PGD \
111 (SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
112#define MAX_UNSHARED_PTRS_PER_PGD \
113 MAX_T(size_t, KERNEL_PGD_BOUNDARY, PTRS_PER_PGD)
114
115
116static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
117{
118 virt_to_ptdesc(pgd)->pt_mm = mm;
119}
120
121struct mm_struct *pgd_page_get_mm(struct page *page)
122{
123 return page_ptdesc(page)->pt_mm;
124}
125
126static void pgd_ctor(struct mm_struct *mm, pgd_t *pgd)
127{
128 /* If the pgd points to a shared pagetable level (either the
129 ptes in non-PAE, or shared PMD in PAE), then just copy the
130 references from swapper_pg_dir. */
131 if (CONFIG_PGTABLE_LEVELS == 2 ||
132 (CONFIG_PGTABLE_LEVELS == 3 && SHARED_KERNEL_PMD) ||
133 CONFIG_PGTABLE_LEVELS >= 4) {
134 clone_pgd_range(pgd + KERNEL_PGD_BOUNDARY,
135 swapper_pg_dir + KERNEL_PGD_BOUNDARY,
136 KERNEL_PGD_PTRS);
137 }
138
139 /* list required to sync kernel mapping updates */
140 if (!SHARED_KERNEL_PMD) {
141 pgd_set_mm(pgd, mm);
142 pgd_list_add(pgd);
143 }
144}
145
146static void pgd_dtor(pgd_t *pgd)
147{
148 if (SHARED_KERNEL_PMD)
149 return;
150
151 spin_lock(&pgd_lock);
152 pgd_list_del(pgd);
153 spin_unlock(&pgd_lock);
154}
155
156/*
157 * List of all pgd's needed for non-PAE so it can invalidate entries
158 * in both cached and uncached pgd's; not needed for PAE since the
159 * kernel pmd is shared. If PAE were not to share the pmd a similar
160 * tactic would be needed. This is essentially codepath-based locking
161 * against pageattr.c; it is the unique case in which a valid change
162 * of kernel pagetables can't be lazily synchronized by vmalloc faults.
163 * vmalloc faults work because attached pagetables are never freed.
164 * -- nyc
165 */
166
167#ifdef CONFIG_X86_PAE
168/*
169 * In PAE mode, we need to do a cr3 reload (=tlb flush) when
170 * updating the top-level pagetable entries to guarantee the
171 * processor notices the update. Since this is expensive, and
172 * all 4 top-level entries are used almost immediately in a
173 * new process's life, we just pre-populate them here.
174 *
175 * Also, if we're in a paravirt environment where the kernel pmd is
176 * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
177 * and initialize the kernel pmds here.
178 */
179#define PREALLOCATED_PMDS UNSHARED_PTRS_PER_PGD
180#define MAX_PREALLOCATED_PMDS MAX_UNSHARED_PTRS_PER_PGD
181
182/*
183 * We allocate separate PMDs for the kernel part of the user page-table
184 * when PTI is enabled. We need them to map the per-process LDT into the
185 * user-space page-table.
186 */
187#define PREALLOCATED_USER_PMDS (boot_cpu_has(X86_FEATURE_PTI) ? \
188 KERNEL_PGD_PTRS : 0)
189#define MAX_PREALLOCATED_USER_PMDS KERNEL_PGD_PTRS
190
191void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
192{
193 paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
194
195 /* Note: almost everything apart from _PAGE_PRESENT is
196 reserved at the pmd (PDPT) level. */
197 set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
198
199 /*
200 * According to Intel App note "TLBs, Paging-Structure Caches,
201 * and Their Invalidation", April 2007, document 317080-001,
202 * section 8.1: in PAE mode we explicitly have to flush the
203 * TLB via cr3 if the top-level pgd is changed...
204 */
205 flush_tlb_mm(mm);
206}
207#else /* !CONFIG_X86_PAE */
208
209/* No need to prepopulate any pagetable entries in non-PAE modes. */
210#define PREALLOCATED_PMDS 0
211#define MAX_PREALLOCATED_PMDS 0
212#define PREALLOCATED_USER_PMDS 0
213#define MAX_PREALLOCATED_USER_PMDS 0
214#endif /* CONFIG_X86_PAE */
215
216static void free_pmds(struct mm_struct *mm, pmd_t *pmds[], int count)
217{
218 int i;
219 struct ptdesc *ptdesc;
220
221 for (i = 0; i < count; i++)
222 if (pmds[i]) {
223 ptdesc = virt_to_ptdesc(pmds[i]);
224
225 pagetable_pmd_dtor(ptdesc);
226 pagetable_free(ptdesc);
227 mm_dec_nr_pmds(mm);
228 }
229}
230
231static int preallocate_pmds(struct mm_struct *mm, pmd_t *pmds[], int count)
232{
233 int i;
234 bool failed = false;
235 gfp_t gfp = GFP_PGTABLE_USER;
236
237 if (mm == &init_mm)
238 gfp &= ~__GFP_ACCOUNT;
239 gfp &= ~__GFP_HIGHMEM;
240
241 for (i = 0; i < count; i++) {
242 pmd_t *pmd = NULL;
243 struct ptdesc *ptdesc = pagetable_alloc(gfp, 0);
244
245 if (!ptdesc)
246 failed = true;
247 if (ptdesc && !pagetable_pmd_ctor(ptdesc)) {
248 pagetable_free(ptdesc);
249 ptdesc = NULL;
250 failed = true;
251 }
252 if (ptdesc) {
253 mm_inc_nr_pmds(mm);
254 pmd = ptdesc_address(ptdesc);
255 }
256
257 pmds[i] = pmd;
258 }
259
260 if (failed) {
261 free_pmds(mm, pmds, count);
262 return -ENOMEM;
263 }
264
265 return 0;
266}
267
268/*
269 * Mop up any pmd pages which may still be attached to the pgd.
270 * Normally they will be freed by munmap/exit_mmap, but any pmd we
271 * preallocate which never got a corresponding vma will need to be
272 * freed manually.
273 */
274static void mop_up_one_pmd(struct mm_struct *mm, pgd_t *pgdp)
275{
276 pgd_t pgd = *pgdp;
277
278 if (pgd_val(pgd) != 0) {
279 pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
280
281 pgd_clear(pgdp);
282
283 paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
284 pmd_free(mm, pmd);
285 mm_dec_nr_pmds(mm);
286 }
287}
288
289static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
290{
291 int i;
292
293 for (i = 0; i < PREALLOCATED_PMDS; i++)
294 mop_up_one_pmd(mm, &pgdp[i]);
295
296#ifdef CONFIG_MITIGATION_PAGE_TABLE_ISOLATION
297
298 if (!boot_cpu_has(X86_FEATURE_PTI))
299 return;
300
301 pgdp = kernel_to_user_pgdp(pgdp);
302
303 for (i = 0; i < PREALLOCATED_USER_PMDS; i++)
304 mop_up_one_pmd(mm, &pgdp[i + KERNEL_PGD_BOUNDARY]);
305#endif
306}
307
308static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
309{
310 p4d_t *p4d;
311 pud_t *pud;
312 int i;
313
314 p4d = p4d_offset(pgd, 0);
315 pud = pud_offset(p4d, 0);
316
317 for (i = 0; i < PREALLOCATED_PMDS; i++, pud++) {
318 pmd_t *pmd = pmds[i];
319
320 if (i >= KERNEL_PGD_BOUNDARY)
321 memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
322 sizeof(pmd_t) * PTRS_PER_PMD);
323
324 pud_populate(mm, pud, pmd);
325 }
326}
327
328#ifdef CONFIG_MITIGATION_PAGE_TABLE_ISOLATION
329static void pgd_prepopulate_user_pmd(struct mm_struct *mm,
330 pgd_t *k_pgd, pmd_t *pmds[])
331{
332 pgd_t *s_pgd = kernel_to_user_pgdp(swapper_pg_dir);
333 pgd_t *u_pgd = kernel_to_user_pgdp(k_pgd);
334 p4d_t *u_p4d;
335 pud_t *u_pud;
336 int i;
337
338 u_p4d = p4d_offset(u_pgd, 0);
339 u_pud = pud_offset(u_p4d, 0);
340
341 s_pgd += KERNEL_PGD_BOUNDARY;
342 u_pud += KERNEL_PGD_BOUNDARY;
343
344 for (i = 0; i < PREALLOCATED_USER_PMDS; i++, u_pud++, s_pgd++) {
345 pmd_t *pmd = pmds[i];
346
347 memcpy(pmd, (pmd_t *)pgd_page_vaddr(*s_pgd),
348 sizeof(pmd_t) * PTRS_PER_PMD);
349
350 pud_populate(mm, u_pud, pmd);
351 }
352
353}
354#else
355static void pgd_prepopulate_user_pmd(struct mm_struct *mm,
356 pgd_t *k_pgd, pmd_t *pmds[])
357{
358}
359#endif
360/*
361 * Xen paravirt assumes pgd table should be in one page. 64 bit kernel also
362 * assumes that pgd should be in one page.
363 *
364 * But kernel with PAE paging that is not running as a Xen domain
365 * only needs to allocate 32 bytes for pgd instead of one page.
366 */
367#ifdef CONFIG_X86_PAE
368
369#include <linux/slab.h>
370
371#define PGD_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
372#define PGD_ALIGN 32
373
374static struct kmem_cache *pgd_cache;
375
376void __init pgtable_cache_init(void)
377{
378 /*
379 * When PAE kernel is running as a Xen domain, it does not use
380 * shared kernel pmd. And this requires a whole page for pgd.
381 */
382 if (!SHARED_KERNEL_PMD)
383 return;
384
385 /*
386 * when PAE kernel is not running as a Xen domain, it uses
387 * shared kernel pmd. Shared kernel pmd does not require a whole
388 * page for pgd. We are able to just allocate a 32-byte for pgd.
389 * During boot time, we create a 32-byte slab for pgd table allocation.
390 */
391 pgd_cache = kmem_cache_create("pgd_cache", PGD_SIZE, PGD_ALIGN,
392 SLAB_PANIC, NULL);
393}
394
395static inline pgd_t *_pgd_alloc(void)
396{
397 /*
398 * If no SHARED_KERNEL_PMD, PAE kernel is running as a Xen domain.
399 * We allocate one page for pgd.
400 */
401 if (!SHARED_KERNEL_PMD)
402 return (pgd_t *)__get_free_pages(GFP_PGTABLE_USER,
403 PGD_ALLOCATION_ORDER);
404
405 /*
406 * Now PAE kernel is not running as a Xen domain. We can allocate
407 * a 32-byte slab for pgd to save memory space.
408 */
409 return kmem_cache_alloc(pgd_cache, GFP_PGTABLE_USER);
410}
411
412static inline void _pgd_free(pgd_t *pgd)
413{
414 if (!SHARED_KERNEL_PMD)
415 free_pages((unsigned long)pgd, PGD_ALLOCATION_ORDER);
416 else
417 kmem_cache_free(pgd_cache, pgd);
418}
419#else
420
421static inline pgd_t *_pgd_alloc(void)
422{
423 return (pgd_t *)__get_free_pages(GFP_PGTABLE_USER,
424 PGD_ALLOCATION_ORDER);
425}
426
427static inline void _pgd_free(pgd_t *pgd)
428{
429 free_pages((unsigned long)pgd, PGD_ALLOCATION_ORDER);
430}
431#endif /* CONFIG_X86_PAE */
432
433pgd_t *pgd_alloc(struct mm_struct *mm)
434{
435 pgd_t *pgd;
436 pmd_t *u_pmds[MAX_PREALLOCATED_USER_PMDS];
437 pmd_t *pmds[MAX_PREALLOCATED_PMDS];
438
439 pgd = _pgd_alloc();
440
441 if (pgd == NULL)
442 goto out;
443
444 mm->pgd = pgd;
445
446 if (sizeof(pmds) != 0 &&
447 preallocate_pmds(mm, pmds, PREALLOCATED_PMDS) != 0)
448 goto out_free_pgd;
449
450 if (sizeof(u_pmds) != 0 &&
451 preallocate_pmds(mm, u_pmds, PREALLOCATED_USER_PMDS) != 0)
452 goto out_free_pmds;
453
454 if (paravirt_pgd_alloc(mm) != 0)
455 goto out_free_user_pmds;
456
457 /*
458 * Make sure that pre-populating the pmds is atomic with
459 * respect to anything walking the pgd_list, so that they
460 * never see a partially populated pgd.
461 */
462 spin_lock(&pgd_lock);
463
464 pgd_ctor(mm, pgd);
465 if (sizeof(pmds) != 0)
466 pgd_prepopulate_pmd(mm, pgd, pmds);
467
468 if (sizeof(u_pmds) != 0)
469 pgd_prepopulate_user_pmd(mm, pgd, u_pmds);
470
471 spin_unlock(&pgd_lock);
472
473 return pgd;
474
475out_free_user_pmds:
476 if (sizeof(u_pmds) != 0)
477 free_pmds(mm, u_pmds, PREALLOCATED_USER_PMDS);
478out_free_pmds:
479 if (sizeof(pmds) != 0)
480 free_pmds(mm, pmds, PREALLOCATED_PMDS);
481out_free_pgd:
482 _pgd_free(pgd);
483out:
484 return NULL;
485}
486
487void pgd_free(struct mm_struct *mm, pgd_t *pgd)
488{
489 pgd_mop_up_pmds(mm, pgd);
490 pgd_dtor(pgd);
491 paravirt_pgd_free(mm, pgd);
492 _pgd_free(pgd);
493}
494
495/*
496 * Used to set accessed or dirty bits in the page table entries
497 * on other architectures. On x86, the accessed and dirty bits
498 * are tracked by hardware. However, do_wp_page calls this function
499 * to also make the pte writeable at the same time the dirty bit is
500 * set. In that case we do actually need to write the PTE.
501 */
502int ptep_set_access_flags(struct vm_area_struct *vma,
503 unsigned long address, pte_t *ptep,
504 pte_t entry, int dirty)
505{
506 int changed = !pte_same(*ptep, entry);
507
508 if (changed && dirty)
509 set_pte(ptep, entry);
510
511 return changed;
512}
513
514#ifdef CONFIG_TRANSPARENT_HUGEPAGE
515int pmdp_set_access_flags(struct vm_area_struct *vma,
516 unsigned long address, pmd_t *pmdp,
517 pmd_t entry, int dirty)
518{
519 int changed = !pmd_same(*pmdp, entry);
520
521 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
522
523 if (changed && dirty) {
524 set_pmd(pmdp, entry);
525 /*
526 * We had a write-protection fault here and changed the pmd
527 * to to more permissive. No need to flush the TLB for that,
528 * #PF is architecturally guaranteed to do that and in the
529 * worst-case we'll generate a spurious fault.
530 */
531 }
532
533 return changed;
534}
535
536int pudp_set_access_flags(struct vm_area_struct *vma, unsigned long address,
537 pud_t *pudp, pud_t entry, int dirty)
538{
539 int changed = !pud_same(*pudp, entry);
540
541 VM_BUG_ON(address & ~HPAGE_PUD_MASK);
542
543 if (changed && dirty) {
544 set_pud(pudp, entry);
545 /*
546 * We had a write-protection fault here and changed the pud
547 * to to more permissive. No need to flush the TLB for that,
548 * #PF is architecturally guaranteed to do that and in the
549 * worst-case we'll generate a spurious fault.
550 */
551 }
552
553 return changed;
554}
555#endif
556
557int ptep_test_and_clear_young(struct vm_area_struct *vma,
558 unsigned long addr, pte_t *ptep)
559{
560 int ret = 0;
561
562 if (pte_young(*ptep))
563 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
564 (unsigned long *) &ptep->pte);
565
566 return ret;
567}
568
569#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG)
570int pmdp_test_and_clear_young(struct vm_area_struct *vma,
571 unsigned long addr, pmd_t *pmdp)
572{
573 int ret = 0;
574
575 if (pmd_young(*pmdp))
576 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
577 (unsigned long *)pmdp);
578
579 return ret;
580}
581#endif
582
583#ifdef CONFIG_TRANSPARENT_HUGEPAGE
584int pudp_test_and_clear_young(struct vm_area_struct *vma,
585 unsigned long addr, pud_t *pudp)
586{
587 int ret = 0;
588
589 if (pud_young(*pudp))
590 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
591 (unsigned long *)pudp);
592
593 return ret;
594}
595#endif
596
597int ptep_clear_flush_young(struct vm_area_struct *vma,
598 unsigned long address, pte_t *ptep)
599{
600 /*
601 * On x86 CPUs, clearing the accessed bit without a TLB flush
602 * doesn't cause data corruption. [ It could cause incorrect
603 * page aging and the (mistaken) reclaim of hot pages, but the
604 * chance of that should be relatively low. ]
605 *
606 * So as a performance optimization don't flush the TLB when
607 * clearing the accessed bit, it will eventually be flushed by
608 * a context switch or a VM operation anyway. [ In the rare
609 * event of it not getting flushed for a long time the delay
610 * shouldn't really matter because there's no real memory
611 * pressure for swapout to react to. ]
612 */
613 return ptep_test_and_clear_young(vma, address, ptep);
614}
615
616#ifdef CONFIG_TRANSPARENT_HUGEPAGE
617int pmdp_clear_flush_young(struct vm_area_struct *vma,
618 unsigned long address, pmd_t *pmdp)
619{
620 int young;
621
622 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
623
624 young = pmdp_test_and_clear_young(vma, address, pmdp);
625 if (young)
626 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
627
628 return young;
629}
630
631pmd_t pmdp_invalidate_ad(struct vm_area_struct *vma, unsigned long address,
632 pmd_t *pmdp)
633{
634 VM_WARN_ON_ONCE(!pmd_present(*pmdp));
635
636 /*
637 * No flush is necessary. Once an invalid PTE is established, the PTE's
638 * access and dirty bits cannot be updated.
639 */
640 return pmdp_establish(vma, address, pmdp, pmd_mkinvalid(*pmdp));
641}
642#endif
643
644#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
645 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
646pud_t pudp_invalidate(struct vm_area_struct *vma, unsigned long address,
647 pud_t *pudp)
648{
649 VM_WARN_ON_ONCE(!pud_present(*pudp));
650 pud_t old = pudp_establish(vma, address, pudp, pud_mkinvalid(*pudp));
651 flush_pud_tlb_range(vma, address, address + HPAGE_PUD_SIZE);
652 return old;
653}
654#endif
655
656/**
657 * reserve_top_address - reserves a hole in the top of kernel address space
658 * @reserve - size of hole to reserve
659 *
660 * Can be used to relocate the fixmap area and poke a hole in the top
661 * of kernel address space to make room for a hypervisor.
662 */
663void __init reserve_top_address(unsigned long reserve)
664{
665#ifdef CONFIG_X86_32
666 BUG_ON(fixmaps_set > 0);
667 __FIXADDR_TOP = round_down(-reserve, 1 << PMD_SHIFT) - PAGE_SIZE;
668 printk(KERN_INFO "Reserving virtual address space above 0x%08lx (rounded to 0x%08lx)\n",
669 -reserve, __FIXADDR_TOP + PAGE_SIZE);
670#endif
671}
672
673int fixmaps_set;
674
675void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
676{
677 unsigned long address = __fix_to_virt(idx);
678
679#ifdef CONFIG_X86_64
680 /*
681 * Ensure that the static initial page tables are covering the
682 * fixmap completely.
683 */
684 BUILD_BUG_ON(__end_of_permanent_fixed_addresses >
685 (FIXMAP_PMD_NUM * PTRS_PER_PTE));
686#endif
687
688 if (idx >= __end_of_fixed_addresses) {
689 BUG();
690 return;
691 }
692 set_pte_vaddr(address, pte);
693 fixmaps_set++;
694}
695
696void native_set_fixmap(unsigned /* enum fixed_addresses */ idx,
697 phys_addr_t phys, pgprot_t flags)
698{
699 /* Sanitize 'prot' against any unsupported bits: */
700 pgprot_val(flags) &= __default_kernel_pte_mask;
701
702 __native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
703}
704
705#ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
706#ifdef CONFIG_X86_5LEVEL
707/**
708 * p4d_set_huge - setup kernel P4D mapping
709 *
710 * No 512GB pages yet -- always return 0
711 */
712int p4d_set_huge(p4d_t *p4d, phys_addr_t addr, pgprot_t prot)
713{
714 return 0;
715}
716
717/**
718 * p4d_clear_huge - clear kernel P4D mapping when it is set
719 *
720 * No 512GB pages yet -- always return 0
721 */
722void p4d_clear_huge(p4d_t *p4d)
723{
724}
725#endif
726
727/**
728 * pud_set_huge - setup kernel PUD mapping
729 *
730 * MTRRs can override PAT memory types with 4KiB granularity. Therefore, this
731 * function sets up a huge page only if the complete range has the same MTRR
732 * caching mode.
733 *
734 * Callers should try to decrease page size (1GB -> 2MB -> 4K) if the bigger
735 * page mapping attempt fails.
736 *
737 * Returns 1 on success and 0 on failure.
738 */
739int pud_set_huge(pud_t *pud, phys_addr_t addr, pgprot_t prot)
740{
741 u8 uniform;
742
743 mtrr_type_lookup(addr, addr + PUD_SIZE, &uniform);
744 if (!uniform)
745 return 0;
746
747 /* Bail out if we are we on a populated non-leaf entry: */
748 if (pud_present(*pud) && !pud_leaf(*pud))
749 return 0;
750
751 set_pte((pte_t *)pud, pfn_pte(
752 (u64)addr >> PAGE_SHIFT,
753 __pgprot(protval_4k_2_large(pgprot_val(prot)) | _PAGE_PSE)));
754
755 return 1;
756}
757
758/**
759 * pmd_set_huge - setup kernel PMD mapping
760 *
761 * See text over pud_set_huge() above.
762 *
763 * Returns 1 on success and 0 on failure.
764 */
765int pmd_set_huge(pmd_t *pmd, phys_addr_t addr, pgprot_t prot)
766{
767 u8 uniform;
768
769 mtrr_type_lookup(addr, addr + PMD_SIZE, &uniform);
770 if (!uniform) {
771 pr_warn_once("%s: Cannot satisfy [mem %#010llx-%#010llx] with a huge-page mapping due to MTRR override.\n",
772 __func__, addr, addr + PMD_SIZE);
773 return 0;
774 }
775
776 /* Bail out if we are we on a populated non-leaf entry: */
777 if (pmd_present(*pmd) && !pmd_leaf(*pmd))
778 return 0;
779
780 set_pte((pte_t *)pmd, pfn_pte(
781 (u64)addr >> PAGE_SHIFT,
782 __pgprot(protval_4k_2_large(pgprot_val(prot)) | _PAGE_PSE)));
783
784 return 1;
785}
786
787/**
788 * pud_clear_huge - clear kernel PUD mapping when it is set
789 *
790 * Returns 1 on success and 0 on failure (no PUD map is found).
791 */
792int pud_clear_huge(pud_t *pud)
793{
794 if (pud_leaf(*pud)) {
795 pud_clear(pud);
796 return 1;
797 }
798
799 return 0;
800}
801
802/**
803 * pmd_clear_huge - clear kernel PMD mapping when it is set
804 *
805 * Returns 1 on success and 0 on failure (no PMD map is found).
806 */
807int pmd_clear_huge(pmd_t *pmd)
808{
809 if (pmd_leaf(*pmd)) {
810 pmd_clear(pmd);
811 return 1;
812 }
813
814 return 0;
815}
816
817#ifdef CONFIG_X86_64
818/**
819 * pud_free_pmd_page - Clear pud entry and free pmd page.
820 * @pud: Pointer to a PUD.
821 * @addr: Virtual address associated with pud.
822 *
823 * Context: The pud range has been unmapped and TLB purged.
824 * Return: 1 if clearing the entry succeeded. 0 otherwise.
825 *
826 * NOTE: Callers must allow a single page allocation.
827 */
828int pud_free_pmd_page(pud_t *pud, unsigned long addr)
829{
830 pmd_t *pmd, *pmd_sv;
831 pte_t *pte;
832 int i;
833
834 pmd = pud_pgtable(*pud);
835 pmd_sv = (pmd_t *)__get_free_page(GFP_KERNEL);
836 if (!pmd_sv)
837 return 0;
838
839 for (i = 0; i < PTRS_PER_PMD; i++) {
840 pmd_sv[i] = pmd[i];
841 if (!pmd_none(pmd[i]))
842 pmd_clear(&pmd[i]);
843 }
844
845 pud_clear(pud);
846
847 /* INVLPG to clear all paging-structure caches */
848 flush_tlb_kernel_range(addr, addr + PAGE_SIZE-1);
849
850 for (i = 0; i < PTRS_PER_PMD; i++) {
851 if (!pmd_none(pmd_sv[i])) {
852 pte = (pte_t *)pmd_page_vaddr(pmd_sv[i]);
853 free_page((unsigned long)pte);
854 }
855 }
856
857 free_page((unsigned long)pmd_sv);
858
859 pagetable_pmd_dtor(virt_to_ptdesc(pmd));
860 free_page((unsigned long)pmd);
861
862 return 1;
863}
864
865/**
866 * pmd_free_pte_page - Clear pmd entry and free pte page.
867 * @pmd: Pointer to a PMD.
868 * @addr: Virtual address associated with pmd.
869 *
870 * Context: The pmd range has been unmapped and TLB purged.
871 * Return: 1 if clearing the entry succeeded. 0 otherwise.
872 */
873int pmd_free_pte_page(pmd_t *pmd, unsigned long addr)
874{
875 pte_t *pte;
876
877 pte = (pte_t *)pmd_page_vaddr(*pmd);
878 pmd_clear(pmd);
879
880 /* INVLPG to clear all paging-structure caches */
881 flush_tlb_kernel_range(addr, addr + PAGE_SIZE-1);
882
883 free_page((unsigned long)pte);
884
885 return 1;
886}
887
888#else /* !CONFIG_X86_64 */
889
890/*
891 * Disable free page handling on x86-PAE. This assures that ioremap()
892 * does not update sync'd pmd entries. See vmalloc_sync_one().
893 */
894int pmd_free_pte_page(pmd_t *pmd, unsigned long addr)
895{
896 return pmd_none(*pmd);
897}
898
899#endif /* CONFIG_X86_64 */
900#endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */
901
902pte_t pte_mkwrite(pte_t pte, struct vm_area_struct *vma)
903{
904 if (vma->vm_flags & VM_SHADOW_STACK)
905 return pte_mkwrite_shstk(pte);
906
907 pte = pte_mkwrite_novma(pte);
908
909 return pte_clear_saveddirty(pte);
910}
911
912pmd_t pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
913{
914 if (vma->vm_flags & VM_SHADOW_STACK)
915 return pmd_mkwrite_shstk(pmd);
916
917 pmd = pmd_mkwrite_novma(pmd);
918
919 return pmd_clear_saveddirty(pmd);
920}
921
922void arch_check_zapped_pte(struct vm_area_struct *vma, pte_t pte)
923{
924 /*
925 * Hardware before shadow stack can (rarely) set Dirty=1
926 * on a Write=0 PTE. So the below condition
927 * only indicates a software bug when shadow stack is
928 * supported by the HW. This checking is covered in
929 * pte_shstk().
930 */
931 VM_WARN_ON_ONCE(!(vma->vm_flags & VM_SHADOW_STACK) &&
932 pte_shstk(pte));
933}
934
935void arch_check_zapped_pmd(struct vm_area_struct *vma, pmd_t pmd)
936{
937 /* See note in arch_check_zapped_pte() */
938 VM_WARN_ON_ONCE(!(vma->vm_flags & VM_SHADOW_STACK) &&
939 pmd_shstk(pmd));
940}
941
942void arch_check_zapped_pud(struct vm_area_struct *vma, pud_t pud)
943{
944 /* See note in arch_check_zapped_pte() */
945 VM_WARN_ON_ONCE(!(vma->vm_flags & VM_SHADOW_STACK) && pud_shstk(pud));
946}