Loading...
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 pgtable_pte_page_dtor(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 page *page = virt_to_page(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 pgtable_pmd_page_dtor(page);
73 paravirt_tlb_remove_table(tlb, page);
74}
75
76#if CONFIG_PGTABLE_LEVELS > 3
77void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
78{
79 paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
80 paravirt_tlb_remove_table(tlb, virt_to_page(pud));
81}
82
83#if CONFIG_PGTABLE_LEVELS > 4
84void ___p4d_free_tlb(struct mmu_gather *tlb, p4d_t *p4d)
85{
86 paravirt_release_p4d(__pa(p4d) >> PAGE_SHIFT);
87 paravirt_tlb_remove_table(tlb, virt_to_page(p4d));
88}
89#endif /* CONFIG_PGTABLE_LEVELS > 4 */
90#endif /* CONFIG_PGTABLE_LEVELS > 3 */
91#endif /* CONFIG_PGTABLE_LEVELS > 2 */
92
93static inline void pgd_list_add(pgd_t *pgd)
94{
95 struct page *page = virt_to_page(pgd);
96
97 list_add(&page->lru, &pgd_list);
98}
99
100static inline void pgd_list_del(pgd_t *pgd)
101{
102 struct page *page = virt_to_page(pgd);
103
104 list_del(&page->lru);
105}
106
107#define UNSHARED_PTRS_PER_PGD \
108 (SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
109#define MAX_UNSHARED_PTRS_PER_PGD \
110 max_t(size_t, KERNEL_PGD_BOUNDARY, PTRS_PER_PGD)
111
112
113static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
114{
115 virt_to_page(pgd)->pt_mm = mm;
116}
117
118struct mm_struct *pgd_page_get_mm(struct page *page)
119{
120 return page->pt_mm;
121}
122
123static void pgd_ctor(struct mm_struct *mm, pgd_t *pgd)
124{
125 /* If the pgd points to a shared pagetable level (either the
126 ptes in non-PAE, or shared PMD in PAE), then just copy the
127 references from swapper_pg_dir. */
128 if (CONFIG_PGTABLE_LEVELS == 2 ||
129 (CONFIG_PGTABLE_LEVELS == 3 && SHARED_KERNEL_PMD) ||
130 CONFIG_PGTABLE_LEVELS >= 4) {
131 clone_pgd_range(pgd + KERNEL_PGD_BOUNDARY,
132 swapper_pg_dir + KERNEL_PGD_BOUNDARY,
133 KERNEL_PGD_PTRS);
134 }
135
136 /* list required to sync kernel mapping updates */
137 if (!SHARED_KERNEL_PMD) {
138 pgd_set_mm(pgd, mm);
139 pgd_list_add(pgd);
140 }
141}
142
143static void pgd_dtor(pgd_t *pgd)
144{
145 if (SHARED_KERNEL_PMD)
146 return;
147
148 spin_lock(&pgd_lock);
149 pgd_list_del(pgd);
150 spin_unlock(&pgd_lock);
151}
152
153/*
154 * List of all pgd's needed for non-PAE so it can invalidate entries
155 * in both cached and uncached pgd's; not needed for PAE since the
156 * kernel pmd is shared. If PAE were not to share the pmd a similar
157 * tactic would be needed. This is essentially codepath-based locking
158 * against pageattr.c; it is the unique case in which a valid change
159 * of kernel pagetables can't be lazily synchronized by vmalloc faults.
160 * vmalloc faults work because attached pagetables are never freed.
161 * -- nyc
162 */
163
164#ifdef CONFIG_X86_PAE
165/*
166 * In PAE mode, we need to do a cr3 reload (=tlb flush) when
167 * updating the top-level pagetable entries to guarantee the
168 * processor notices the update. Since this is expensive, and
169 * all 4 top-level entries are used almost immediately in a
170 * new process's life, we just pre-populate them here.
171 *
172 * Also, if we're in a paravirt environment where the kernel pmd is
173 * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
174 * and initialize the kernel pmds here.
175 */
176#define PREALLOCATED_PMDS UNSHARED_PTRS_PER_PGD
177#define MAX_PREALLOCATED_PMDS MAX_UNSHARED_PTRS_PER_PGD
178
179/*
180 * We allocate separate PMDs for the kernel part of the user page-table
181 * when PTI is enabled. We need them to map the per-process LDT into the
182 * user-space page-table.
183 */
184#define PREALLOCATED_USER_PMDS (boot_cpu_has(X86_FEATURE_PTI) ? \
185 KERNEL_PGD_PTRS : 0)
186#define MAX_PREALLOCATED_USER_PMDS KERNEL_PGD_PTRS
187
188void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
189{
190 paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
191
192 /* Note: almost everything apart from _PAGE_PRESENT is
193 reserved at the pmd (PDPT) level. */
194 set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
195
196 /*
197 * According to Intel App note "TLBs, Paging-Structure Caches,
198 * and Their Invalidation", April 2007, document 317080-001,
199 * section 8.1: in PAE mode we explicitly have to flush the
200 * TLB via cr3 if the top-level pgd is changed...
201 */
202 flush_tlb_mm(mm);
203}
204#else /* !CONFIG_X86_PAE */
205
206/* No need to prepopulate any pagetable entries in non-PAE modes. */
207#define PREALLOCATED_PMDS 0
208#define MAX_PREALLOCATED_PMDS 0
209#define PREALLOCATED_USER_PMDS 0
210#define MAX_PREALLOCATED_USER_PMDS 0
211#endif /* CONFIG_X86_PAE */
212
213static void free_pmds(struct mm_struct *mm, pmd_t *pmds[], int count)
214{
215 int i;
216
217 for (i = 0; i < count; i++)
218 if (pmds[i]) {
219 pgtable_pmd_page_dtor(virt_to_page(pmds[i]));
220 free_page((unsigned long)pmds[i]);
221 mm_dec_nr_pmds(mm);
222 }
223}
224
225static int preallocate_pmds(struct mm_struct *mm, pmd_t *pmds[], int count)
226{
227 int i;
228 bool failed = false;
229 gfp_t gfp = GFP_PGTABLE_USER;
230
231 if (mm == &init_mm)
232 gfp &= ~__GFP_ACCOUNT;
233
234 for (i = 0; i < count; i++) {
235 pmd_t *pmd = (pmd_t *)__get_free_page(gfp);
236 if (!pmd)
237 failed = true;
238 if (pmd && !pgtable_pmd_page_ctor(virt_to_page(pmd))) {
239 free_page((unsigned long)pmd);
240 pmd = NULL;
241 failed = true;
242 }
243 if (pmd)
244 mm_inc_nr_pmds(mm);
245 pmds[i] = pmd;
246 }
247
248 if (failed) {
249 free_pmds(mm, pmds, count);
250 return -ENOMEM;
251 }
252
253 return 0;
254}
255
256/*
257 * Mop up any pmd pages which may still be attached to the pgd.
258 * Normally they will be freed by munmap/exit_mmap, but any pmd we
259 * preallocate which never got a corresponding vma will need to be
260 * freed manually.
261 */
262static void mop_up_one_pmd(struct mm_struct *mm, pgd_t *pgdp)
263{
264 pgd_t pgd = *pgdp;
265
266 if (pgd_val(pgd) != 0) {
267 pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
268
269 pgd_clear(pgdp);
270
271 paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
272 pmd_free(mm, pmd);
273 mm_dec_nr_pmds(mm);
274 }
275}
276
277static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
278{
279 int i;
280
281 for (i = 0; i < PREALLOCATED_PMDS; i++)
282 mop_up_one_pmd(mm, &pgdp[i]);
283
284#ifdef CONFIG_PAGE_TABLE_ISOLATION
285
286 if (!boot_cpu_has(X86_FEATURE_PTI))
287 return;
288
289 pgdp = kernel_to_user_pgdp(pgdp);
290
291 for (i = 0; i < PREALLOCATED_USER_PMDS; i++)
292 mop_up_one_pmd(mm, &pgdp[i + KERNEL_PGD_BOUNDARY]);
293#endif
294}
295
296static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
297{
298 p4d_t *p4d;
299 pud_t *pud;
300 int i;
301
302 if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
303 return;
304
305 p4d = p4d_offset(pgd, 0);
306 pud = pud_offset(p4d, 0);
307
308 for (i = 0; i < PREALLOCATED_PMDS; i++, pud++) {
309 pmd_t *pmd = pmds[i];
310
311 if (i >= KERNEL_PGD_BOUNDARY)
312 memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
313 sizeof(pmd_t) * PTRS_PER_PMD);
314
315 pud_populate(mm, pud, pmd);
316 }
317}
318
319#ifdef CONFIG_PAGE_TABLE_ISOLATION
320static void pgd_prepopulate_user_pmd(struct mm_struct *mm,
321 pgd_t *k_pgd, pmd_t *pmds[])
322{
323 pgd_t *s_pgd = kernel_to_user_pgdp(swapper_pg_dir);
324 pgd_t *u_pgd = kernel_to_user_pgdp(k_pgd);
325 p4d_t *u_p4d;
326 pud_t *u_pud;
327 int i;
328
329 u_p4d = p4d_offset(u_pgd, 0);
330 u_pud = pud_offset(u_p4d, 0);
331
332 s_pgd += KERNEL_PGD_BOUNDARY;
333 u_pud += KERNEL_PGD_BOUNDARY;
334
335 for (i = 0; i < PREALLOCATED_USER_PMDS; i++, u_pud++, s_pgd++) {
336 pmd_t *pmd = pmds[i];
337
338 memcpy(pmd, (pmd_t *)pgd_page_vaddr(*s_pgd),
339 sizeof(pmd_t) * PTRS_PER_PMD);
340
341 pud_populate(mm, u_pud, pmd);
342 }
343
344}
345#else
346static void pgd_prepopulate_user_pmd(struct mm_struct *mm,
347 pgd_t *k_pgd, pmd_t *pmds[])
348{
349}
350#endif
351/*
352 * Xen paravirt assumes pgd table should be in one page. 64 bit kernel also
353 * assumes that pgd should be in one page.
354 *
355 * But kernel with PAE paging that is not running as a Xen domain
356 * only needs to allocate 32 bytes for pgd instead of one page.
357 */
358#ifdef CONFIG_X86_PAE
359
360#include <linux/slab.h>
361
362#define PGD_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
363#define PGD_ALIGN 32
364
365static struct kmem_cache *pgd_cache;
366
367void __init pgtable_cache_init(void)
368{
369 /*
370 * When PAE kernel is running as a Xen domain, it does not use
371 * shared kernel pmd. And this requires a whole page for pgd.
372 */
373 if (!SHARED_KERNEL_PMD)
374 return;
375
376 /*
377 * when PAE kernel is not running as a Xen domain, it uses
378 * shared kernel pmd. Shared kernel pmd does not require a whole
379 * page for pgd. We are able to just allocate a 32-byte for pgd.
380 * During boot time, we create a 32-byte slab for pgd table allocation.
381 */
382 pgd_cache = kmem_cache_create("pgd_cache", PGD_SIZE, PGD_ALIGN,
383 SLAB_PANIC, NULL);
384}
385
386static inline pgd_t *_pgd_alloc(void)
387{
388 /*
389 * If no SHARED_KERNEL_PMD, PAE kernel is running as a Xen domain.
390 * We allocate one page for pgd.
391 */
392 if (!SHARED_KERNEL_PMD)
393 return (pgd_t *)__get_free_pages(GFP_PGTABLE_USER,
394 PGD_ALLOCATION_ORDER);
395
396 /*
397 * Now PAE kernel is not running as a Xen domain. We can allocate
398 * a 32-byte slab for pgd to save memory space.
399 */
400 return kmem_cache_alloc(pgd_cache, GFP_PGTABLE_USER);
401}
402
403static inline void _pgd_free(pgd_t *pgd)
404{
405 if (!SHARED_KERNEL_PMD)
406 free_pages((unsigned long)pgd, PGD_ALLOCATION_ORDER);
407 else
408 kmem_cache_free(pgd_cache, pgd);
409}
410#else
411
412static inline pgd_t *_pgd_alloc(void)
413{
414 return (pgd_t *)__get_free_pages(GFP_PGTABLE_USER,
415 PGD_ALLOCATION_ORDER);
416}
417
418static inline void _pgd_free(pgd_t *pgd)
419{
420 free_pages((unsigned long)pgd, PGD_ALLOCATION_ORDER);
421}
422#endif /* CONFIG_X86_PAE */
423
424pgd_t *pgd_alloc(struct mm_struct *mm)
425{
426 pgd_t *pgd;
427 pmd_t *u_pmds[MAX_PREALLOCATED_USER_PMDS];
428 pmd_t *pmds[MAX_PREALLOCATED_PMDS];
429
430 pgd = _pgd_alloc();
431
432 if (pgd == NULL)
433 goto out;
434
435 mm->pgd = pgd;
436
437 if (preallocate_pmds(mm, pmds, PREALLOCATED_PMDS) != 0)
438 goto out_free_pgd;
439
440 if (preallocate_pmds(mm, u_pmds, PREALLOCATED_USER_PMDS) != 0)
441 goto out_free_pmds;
442
443 if (paravirt_pgd_alloc(mm) != 0)
444 goto out_free_user_pmds;
445
446 /*
447 * Make sure that pre-populating the pmds is atomic with
448 * respect to anything walking the pgd_list, so that they
449 * never see a partially populated pgd.
450 */
451 spin_lock(&pgd_lock);
452
453 pgd_ctor(mm, pgd);
454 pgd_prepopulate_pmd(mm, pgd, pmds);
455 pgd_prepopulate_user_pmd(mm, pgd, u_pmds);
456
457 spin_unlock(&pgd_lock);
458
459 return pgd;
460
461out_free_user_pmds:
462 free_pmds(mm, u_pmds, PREALLOCATED_USER_PMDS);
463out_free_pmds:
464 free_pmds(mm, pmds, PREALLOCATED_PMDS);
465out_free_pgd:
466 _pgd_free(pgd);
467out:
468 return NULL;
469}
470
471void pgd_free(struct mm_struct *mm, pgd_t *pgd)
472{
473 pgd_mop_up_pmds(mm, pgd);
474 pgd_dtor(pgd);
475 paravirt_pgd_free(mm, pgd);
476 _pgd_free(pgd);
477}
478
479/*
480 * Used to set accessed or dirty bits in the page table entries
481 * on other architectures. On x86, the accessed and dirty bits
482 * are tracked by hardware. However, do_wp_page calls this function
483 * to also make the pte writeable at the same time the dirty bit is
484 * set. In that case we do actually need to write the PTE.
485 */
486int ptep_set_access_flags(struct vm_area_struct *vma,
487 unsigned long address, pte_t *ptep,
488 pte_t entry, int dirty)
489{
490 int changed = !pte_same(*ptep, entry);
491
492 if (changed && dirty)
493 set_pte(ptep, entry);
494
495 return changed;
496}
497
498#ifdef CONFIG_TRANSPARENT_HUGEPAGE
499int pmdp_set_access_flags(struct vm_area_struct *vma,
500 unsigned long address, pmd_t *pmdp,
501 pmd_t entry, int dirty)
502{
503 int changed = !pmd_same(*pmdp, entry);
504
505 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
506
507 if (changed && dirty) {
508 set_pmd(pmdp, entry);
509 /*
510 * We had a write-protection fault here and changed the pmd
511 * to to more permissive. No need to flush the TLB for that,
512 * #PF is architecturally guaranteed to do that and in the
513 * worst-case we'll generate a spurious fault.
514 */
515 }
516
517 return changed;
518}
519
520int pudp_set_access_flags(struct vm_area_struct *vma, unsigned long address,
521 pud_t *pudp, pud_t entry, int dirty)
522{
523 int changed = !pud_same(*pudp, entry);
524
525 VM_BUG_ON(address & ~HPAGE_PUD_MASK);
526
527 if (changed && dirty) {
528 set_pud(pudp, entry);
529 /*
530 * We had a write-protection fault here and changed the pud
531 * to to more permissive. No need to flush the TLB for that,
532 * #PF is architecturally guaranteed to do that and in the
533 * worst-case we'll generate a spurious fault.
534 */
535 }
536
537 return changed;
538}
539#endif
540
541int ptep_test_and_clear_young(struct vm_area_struct *vma,
542 unsigned long addr, pte_t *ptep)
543{
544 int ret = 0;
545
546 if (pte_young(*ptep))
547 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
548 (unsigned long *) &ptep->pte);
549
550 return ret;
551}
552
553#ifdef CONFIG_TRANSPARENT_HUGEPAGE
554int pmdp_test_and_clear_young(struct vm_area_struct *vma,
555 unsigned long addr, pmd_t *pmdp)
556{
557 int ret = 0;
558
559 if (pmd_young(*pmdp))
560 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
561 (unsigned long *)pmdp);
562
563 return ret;
564}
565int pudp_test_and_clear_young(struct vm_area_struct *vma,
566 unsigned long addr, pud_t *pudp)
567{
568 int ret = 0;
569
570 if (pud_young(*pudp))
571 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
572 (unsigned long *)pudp);
573
574 return ret;
575}
576#endif
577
578int ptep_clear_flush_young(struct vm_area_struct *vma,
579 unsigned long address, pte_t *ptep)
580{
581 /*
582 * On x86 CPUs, clearing the accessed bit without a TLB flush
583 * doesn't cause data corruption. [ It could cause incorrect
584 * page aging and the (mistaken) reclaim of hot pages, but the
585 * chance of that should be relatively low. ]
586 *
587 * So as a performance optimization don't flush the TLB when
588 * clearing the accessed bit, it will eventually be flushed by
589 * a context switch or a VM operation anyway. [ In the rare
590 * event of it not getting flushed for a long time the delay
591 * shouldn't really matter because there's no real memory
592 * pressure for swapout to react to. ]
593 */
594 return ptep_test_and_clear_young(vma, address, ptep);
595}
596
597#ifdef CONFIG_TRANSPARENT_HUGEPAGE
598int pmdp_clear_flush_young(struct vm_area_struct *vma,
599 unsigned long address, pmd_t *pmdp)
600{
601 int young;
602
603 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
604
605 young = pmdp_test_and_clear_young(vma, address, pmdp);
606 if (young)
607 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
608
609 return young;
610}
611#endif
612
613/**
614 * reserve_top_address - reserves a hole in the top of kernel address space
615 * @reserve - size of hole to reserve
616 *
617 * Can be used to relocate the fixmap area and poke a hole in the top
618 * of kernel address space to make room for a hypervisor.
619 */
620void __init reserve_top_address(unsigned long reserve)
621{
622#ifdef CONFIG_X86_32
623 BUG_ON(fixmaps_set > 0);
624 __FIXADDR_TOP = round_down(-reserve, 1 << PMD_SHIFT) - PAGE_SIZE;
625 printk(KERN_INFO "Reserving virtual address space above 0x%08lx (rounded to 0x%08lx)\n",
626 -reserve, __FIXADDR_TOP + PAGE_SIZE);
627#endif
628}
629
630int fixmaps_set;
631
632void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
633{
634 unsigned long address = __fix_to_virt(idx);
635
636#ifdef CONFIG_X86_64
637 /*
638 * Ensure that the static initial page tables are covering the
639 * fixmap completely.
640 */
641 BUILD_BUG_ON(__end_of_permanent_fixed_addresses >
642 (FIXMAP_PMD_NUM * PTRS_PER_PTE));
643#endif
644
645 if (idx >= __end_of_fixed_addresses) {
646 BUG();
647 return;
648 }
649 set_pte_vaddr(address, pte);
650 fixmaps_set++;
651}
652
653void native_set_fixmap(unsigned /* enum fixed_addresses */ idx,
654 phys_addr_t phys, pgprot_t flags)
655{
656 /* Sanitize 'prot' against any unsupported bits: */
657 pgprot_val(flags) &= __default_kernel_pte_mask;
658
659 __native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
660}
661
662#ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
663#ifdef CONFIG_X86_5LEVEL
664/**
665 * p4d_set_huge - setup kernel P4D mapping
666 *
667 * No 512GB pages yet -- always return 0
668 */
669int p4d_set_huge(p4d_t *p4d, phys_addr_t addr, pgprot_t prot)
670{
671 return 0;
672}
673
674/**
675 * p4d_clear_huge - clear kernel P4D mapping when it is set
676 *
677 * No 512GB pages yet -- always return 0
678 */
679int p4d_clear_huge(p4d_t *p4d)
680{
681 return 0;
682}
683#endif
684
685/**
686 * pud_set_huge - setup kernel PUD mapping
687 *
688 * MTRRs can override PAT memory types with 4KiB granularity. Therefore, this
689 * function sets up a huge page only if any of the following conditions are met:
690 *
691 * - MTRRs are disabled, or
692 *
693 * - MTRRs are enabled and the range is completely covered by a single MTRR, or
694 *
695 * - MTRRs are enabled and the corresponding MTRR memory type is WB, which
696 * has no effect on the requested PAT memory type.
697 *
698 * Callers should try to decrease page size (1GB -> 2MB -> 4K) if the bigger
699 * page mapping attempt fails.
700 *
701 * Returns 1 on success and 0 on failure.
702 */
703int pud_set_huge(pud_t *pud, phys_addr_t addr, pgprot_t prot)
704{
705 u8 mtrr, uniform;
706
707 mtrr = mtrr_type_lookup(addr, addr + PUD_SIZE, &uniform);
708 if ((mtrr != MTRR_TYPE_INVALID) && (!uniform) &&
709 (mtrr != MTRR_TYPE_WRBACK))
710 return 0;
711
712 /* Bail out if we are we on a populated non-leaf entry: */
713 if (pud_present(*pud) && !pud_huge(*pud))
714 return 0;
715
716 set_pte((pte_t *)pud, pfn_pte(
717 (u64)addr >> PAGE_SHIFT,
718 __pgprot(protval_4k_2_large(pgprot_val(prot)) | _PAGE_PSE)));
719
720 return 1;
721}
722
723/**
724 * pmd_set_huge - setup kernel PMD mapping
725 *
726 * See text over pud_set_huge() above.
727 *
728 * Returns 1 on success and 0 on failure.
729 */
730int pmd_set_huge(pmd_t *pmd, phys_addr_t addr, pgprot_t prot)
731{
732 u8 mtrr, uniform;
733
734 mtrr = mtrr_type_lookup(addr, addr + PMD_SIZE, &uniform);
735 if ((mtrr != MTRR_TYPE_INVALID) && (!uniform) &&
736 (mtrr != MTRR_TYPE_WRBACK)) {
737 pr_warn_once("%s: Cannot satisfy [mem %#010llx-%#010llx] with a huge-page mapping due to MTRR override.\n",
738 __func__, addr, addr + PMD_SIZE);
739 return 0;
740 }
741
742 /* Bail out if we are we on a populated non-leaf entry: */
743 if (pmd_present(*pmd) && !pmd_huge(*pmd))
744 return 0;
745
746 set_pte((pte_t *)pmd, pfn_pte(
747 (u64)addr >> PAGE_SHIFT,
748 __pgprot(protval_4k_2_large(pgprot_val(prot)) | _PAGE_PSE)));
749
750 return 1;
751}
752
753/**
754 * pud_clear_huge - clear kernel PUD mapping when it is set
755 *
756 * Returns 1 on success and 0 on failure (no PUD map is found).
757 */
758int pud_clear_huge(pud_t *pud)
759{
760 if (pud_large(*pud)) {
761 pud_clear(pud);
762 return 1;
763 }
764
765 return 0;
766}
767
768/**
769 * pmd_clear_huge - clear kernel PMD mapping when it is set
770 *
771 * Returns 1 on success and 0 on failure (no PMD map is found).
772 */
773int pmd_clear_huge(pmd_t *pmd)
774{
775 if (pmd_large(*pmd)) {
776 pmd_clear(pmd);
777 return 1;
778 }
779
780 return 0;
781}
782
783/*
784 * Until we support 512GB pages, skip them in the vmap area.
785 */
786int p4d_free_pud_page(p4d_t *p4d, unsigned long addr)
787{
788 return 0;
789}
790
791#ifdef CONFIG_X86_64
792/**
793 * pud_free_pmd_page - Clear pud entry and free pmd page.
794 * @pud: Pointer to a PUD.
795 * @addr: Virtual address associated with pud.
796 *
797 * Context: The pud range has been unmapped and TLB purged.
798 * Return: 1 if clearing the entry succeeded. 0 otherwise.
799 *
800 * NOTE: Callers must allow a single page allocation.
801 */
802int pud_free_pmd_page(pud_t *pud, unsigned long addr)
803{
804 pmd_t *pmd, *pmd_sv;
805 pte_t *pte;
806 int i;
807
808 pmd = (pmd_t *)pud_page_vaddr(*pud);
809 pmd_sv = (pmd_t *)__get_free_page(GFP_KERNEL);
810 if (!pmd_sv)
811 return 0;
812
813 for (i = 0; i < PTRS_PER_PMD; i++) {
814 pmd_sv[i] = pmd[i];
815 if (!pmd_none(pmd[i]))
816 pmd_clear(&pmd[i]);
817 }
818
819 pud_clear(pud);
820
821 /* INVLPG to clear all paging-structure caches */
822 flush_tlb_kernel_range(addr, addr + PAGE_SIZE-1);
823
824 for (i = 0; i < PTRS_PER_PMD; i++) {
825 if (!pmd_none(pmd_sv[i])) {
826 pte = (pte_t *)pmd_page_vaddr(pmd_sv[i]);
827 free_page((unsigned long)pte);
828 }
829 }
830
831 free_page((unsigned long)pmd_sv);
832 free_page((unsigned long)pmd);
833
834 return 1;
835}
836
837/**
838 * pmd_free_pte_page - Clear pmd entry and free pte page.
839 * @pmd: Pointer to a PMD.
840 * @addr: Virtual address associated with pmd.
841 *
842 * Context: The pmd range has been unmapped and TLB purged.
843 * Return: 1 if clearing the entry succeeded. 0 otherwise.
844 */
845int pmd_free_pte_page(pmd_t *pmd, unsigned long addr)
846{
847 pte_t *pte;
848
849 pte = (pte_t *)pmd_page_vaddr(*pmd);
850 pmd_clear(pmd);
851
852 /* INVLPG to clear all paging-structure caches */
853 flush_tlb_kernel_range(addr, addr + PAGE_SIZE-1);
854
855 free_page((unsigned long)pte);
856
857 return 1;
858}
859
860#else /* !CONFIG_X86_64 */
861
862int pud_free_pmd_page(pud_t *pud, unsigned long addr)
863{
864 return pud_none(*pud);
865}
866
867/*
868 * Disable free page handling on x86-PAE. This assures that ioremap()
869 * does not update sync'd pmd entries. See vmalloc_sync_one().
870 */
871int pmd_free_pte_page(pmd_t *pmd, unsigned long addr)
872{
873 return pmd_none(*pmd);
874}
875
876#endif /* CONFIG_X86_64 */
877#endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */
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}