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/pgtable.h>
7#include <asm/tlb.h>
8#include <asm/fixmap.h>
9#include <asm/mtrr.h>
10
11#define PGALLOC_GFP (GFP_KERNEL_ACCOUNT | __GFP_ZERO)
12
13#ifdef CONFIG_HIGHPTE
14#define PGALLOC_USER_GFP __GFP_HIGHMEM
15#else
16#define PGALLOC_USER_GFP 0
17#endif
18
19gfp_t __userpte_alloc_gfp = PGALLOC_GFP | PGALLOC_USER_GFP;
20
21pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
22{
23 return (pte_t *)__get_free_page(PGALLOC_GFP & ~__GFP_ACCOUNT);
24}
25
26pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
27{
28 struct page *pte;
29
30 pte = alloc_pages(__userpte_alloc_gfp, 0);
31 if (!pte)
32 return NULL;
33 if (!pgtable_page_ctor(pte)) {
34 __free_page(pte);
35 return NULL;
36 }
37 return pte;
38}
39
40static int __init setup_userpte(char *arg)
41{
42 if (!arg)
43 return -EINVAL;
44
45 /*
46 * "userpte=nohigh" disables allocation of user pagetables in
47 * high memory.
48 */
49 if (strcmp(arg, "nohigh") == 0)
50 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
51 else
52 return -EINVAL;
53 return 0;
54}
55early_param("userpte", setup_userpte);
56
57void ___pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
58{
59 pgtable_page_dtor(pte);
60 paravirt_release_pte(page_to_pfn(pte));
61 tlb_remove_table(tlb, pte);
62}
63
64#if CONFIG_PGTABLE_LEVELS > 2
65void ___pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
66{
67 struct page *page = virt_to_page(pmd);
68 paravirt_release_pmd(__pa(pmd) >> PAGE_SHIFT);
69 /*
70 * NOTE! For PAE, any changes to the top page-directory-pointer-table
71 * entries need a full cr3 reload to flush.
72 */
73#ifdef CONFIG_X86_PAE
74 tlb->need_flush_all = 1;
75#endif
76 pgtable_pmd_page_dtor(page);
77 tlb_remove_table(tlb, page);
78}
79
80#if CONFIG_PGTABLE_LEVELS > 3
81void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
82{
83 paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
84 tlb_remove_table(tlb, virt_to_page(pud));
85}
86
87#if CONFIG_PGTABLE_LEVELS > 4
88void ___p4d_free_tlb(struct mmu_gather *tlb, p4d_t *p4d)
89{
90 paravirt_release_p4d(__pa(p4d) >> PAGE_SHIFT);
91 tlb_remove_table(tlb, virt_to_page(p4d));
92}
93#endif /* CONFIG_PGTABLE_LEVELS > 4 */
94#endif /* CONFIG_PGTABLE_LEVELS > 3 */
95#endif /* CONFIG_PGTABLE_LEVELS > 2 */
96
97static inline void pgd_list_add(pgd_t *pgd)
98{
99 struct page *page = virt_to_page(pgd);
100
101 list_add(&page->lru, &pgd_list);
102}
103
104static inline void pgd_list_del(pgd_t *pgd)
105{
106 struct page *page = virt_to_page(pgd);
107
108 list_del(&page->lru);
109}
110
111#define UNSHARED_PTRS_PER_PGD \
112 (SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
113
114
115static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
116{
117 BUILD_BUG_ON(sizeof(virt_to_page(pgd)->index) < sizeof(mm));
118 virt_to_page(pgd)->index = (pgoff_t)mm;
119}
120
121struct mm_struct *pgd_page_get_mm(struct page *page)
122{
123 return (struct mm_struct *)page->index;
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
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
202#endif /* CONFIG_X86_PAE */
203
204static void free_pmds(struct mm_struct *mm, pmd_t *pmds[])
205{
206 int i;
207
208 for(i = 0; i < PREALLOCATED_PMDS; i++)
209 if (pmds[i]) {
210 pgtable_pmd_page_dtor(virt_to_page(pmds[i]));
211 free_page((unsigned long)pmds[i]);
212 mm_dec_nr_pmds(mm);
213 }
214}
215
216static int preallocate_pmds(struct mm_struct *mm, pmd_t *pmds[])
217{
218 int i;
219 bool failed = false;
220 gfp_t gfp = PGALLOC_GFP;
221
222 if (mm == &init_mm)
223 gfp &= ~__GFP_ACCOUNT;
224
225 for(i = 0; i < PREALLOCATED_PMDS; i++) {
226 pmd_t *pmd = (pmd_t *)__get_free_page(gfp);
227 if (!pmd)
228 failed = true;
229 if (pmd && !pgtable_pmd_page_ctor(virt_to_page(pmd))) {
230 free_page((unsigned long)pmd);
231 pmd = NULL;
232 failed = true;
233 }
234 if (pmd)
235 mm_inc_nr_pmds(mm);
236 pmds[i] = pmd;
237 }
238
239 if (failed) {
240 free_pmds(mm, pmds);
241 return -ENOMEM;
242 }
243
244 return 0;
245}
246
247/*
248 * Mop up any pmd pages which may still be attached to the pgd.
249 * Normally they will be freed by munmap/exit_mmap, but any pmd we
250 * preallocate which never got a corresponding vma will need to be
251 * freed manually.
252 */
253static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
254{
255 int i;
256
257 for(i = 0; i < PREALLOCATED_PMDS; i++) {
258 pgd_t pgd = pgdp[i];
259
260 if (pgd_val(pgd) != 0) {
261 pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
262
263 pgdp[i] = native_make_pgd(0);
264
265 paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
266 pmd_free(mm, pmd);
267 mm_dec_nr_pmds(mm);
268 }
269 }
270}
271
272static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
273{
274 p4d_t *p4d;
275 pud_t *pud;
276 int i;
277
278 if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
279 return;
280
281 p4d = p4d_offset(pgd, 0);
282 pud = pud_offset(p4d, 0);
283
284 for (i = 0; i < PREALLOCATED_PMDS; i++, pud++) {
285 pmd_t *pmd = pmds[i];
286
287 if (i >= KERNEL_PGD_BOUNDARY)
288 memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
289 sizeof(pmd_t) * PTRS_PER_PMD);
290
291 pud_populate(mm, pud, pmd);
292 }
293}
294
295/*
296 * Xen paravirt assumes pgd table should be in one page. 64 bit kernel also
297 * assumes that pgd should be in one page.
298 *
299 * But kernel with PAE paging that is not running as a Xen domain
300 * only needs to allocate 32 bytes for pgd instead of one page.
301 */
302#ifdef CONFIG_X86_PAE
303
304#include <linux/slab.h>
305
306#define PGD_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
307#define PGD_ALIGN 32
308
309static struct kmem_cache *pgd_cache;
310
311static int __init pgd_cache_init(void)
312{
313 /*
314 * When PAE kernel is running as a Xen domain, it does not use
315 * shared kernel pmd. And this requires a whole page for pgd.
316 */
317 if (!SHARED_KERNEL_PMD)
318 return 0;
319
320 /*
321 * when PAE kernel is not running as a Xen domain, it uses
322 * shared kernel pmd. Shared kernel pmd does not require a whole
323 * page for pgd. We are able to just allocate a 32-byte for pgd.
324 * During boot time, we create a 32-byte slab for pgd table allocation.
325 */
326 pgd_cache = kmem_cache_create("pgd_cache", PGD_SIZE, PGD_ALIGN,
327 SLAB_PANIC, NULL);
328 if (!pgd_cache)
329 return -ENOMEM;
330
331 return 0;
332}
333core_initcall(pgd_cache_init);
334
335static inline pgd_t *_pgd_alloc(void)
336{
337 /*
338 * If no SHARED_KERNEL_PMD, PAE kernel is running as a Xen domain.
339 * We allocate one page for pgd.
340 */
341 if (!SHARED_KERNEL_PMD)
342 return (pgd_t *)__get_free_page(PGALLOC_GFP);
343
344 /*
345 * Now PAE kernel is not running as a Xen domain. We can allocate
346 * a 32-byte slab for pgd to save memory space.
347 */
348 return kmem_cache_alloc(pgd_cache, PGALLOC_GFP);
349}
350
351static inline void _pgd_free(pgd_t *pgd)
352{
353 if (!SHARED_KERNEL_PMD)
354 free_page((unsigned long)pgd);
355 else
356 kmem_cache_free(pgd_cache, pgd);
357}
358#else
359
360static inline pgd_t *_pgd_alloc(void)
361{
362 return (pgd_t *)__get_free_pages(PGALLOC_GFP, PGD_ALLOCATION_ORDER);
363}
364
365static inline void _pgd_free(pgd_t *pgd)
366{
367 free_pages((unsigned long)pgd, PGD_ALLOCATION_ORDER);
368}
369#endif /* CONFIG_X86_PAE */
370
371pgd_t *pgd_alloc(struct mm_struct *mm)
372{
373 pgd_t *pgd;
374 pmd_t *pmds[PREALLOCATED_PMDS];
375
376 pgd = _pgd_alloc();
377
378 if (pgd == NULL)
379 goto out;
380
381 mm->pgd = pgd;
382
383 if (preallocate_pmds(mm, pmds) != 0)
384 goto out_free_pgd;
385
386 if (paravirt_pgd_alloc(mm) != 0)
387 goto out_free_pmds;
388
389 /*
390 * Make sure that pre-populating the pmds is atomic with
391 * respect to anything walking the pgd_list, so that they
392 * never see a partially populated pgd.
393 */
394 spin_lock(&pgd_lock);
395
396 pgd_ctor(mm, pgd);
397 pgd_prepopulate_pmd(mm, pgd, pmds);
398
399 spin_unlock(&pgd_lock);
400
401 return pgd;
402
403out_free_pmds:
404 free_pmds(mm, pmds);
405out_free_pgd:
406 _pgd_free(pgd);
407out:
408 return NULL;
409}
410
411void pgd_free(struct mm_struct *mm, pgd_t *pgd)
412{
413 pgd_mop_up_pmds(mm, pgd);
414 pgd_dtor(pgd);
415 paravirt_pgd_free(mm, pgd);
416 _pgd_free(pgd);
417}
418
419/*
420 * Used to set accessed or dirty bits in the page table entries
421 * on other architectures. On x86, the accessed and dirty bits
422 * are tracked by hardware. However, do_wp_page calls this function
423 * to also make the pte writeable at the same time the dirty bit is
424 * set. In that case we do actually need to write the PTE.
425 */
426int ptep_set_access_flags(struct vm_area_struct *vma,
427 unsigned long address, pte_t *ptep,
428 pte_t entry, int dirty)
429{
430 int changed = !pte_same(*ptep, entry);
431
432 if (changed && dirty)
433 *ptep = entry;
434
435 return changed;
436}
437
438#ifdef CONFIG_TRANSPARENT_HUGEPAGE
439int pmdp_set_access_flags(struct vm_area_struct *vma,
440 unsigned long address, pmd_t *pmdp,
441 pmd_t entry, int dirty)
442{
443 int changed = !pmd_same(*pmdp, entry);
444
445 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
446
447 if (changed && dirty) {
448 *pmdp = entry;
449 /*
450 * We had a write-protection fault here and changed the pmd
451 * to to more permissive. No need to flush the TLB for that,
452 * #PF is architecturally guaranteed to do that and in the
453 * worst-case we'll generate a spurious fault.
454 */
455 }
456
457 return changed;
458}
459
460int pudp_set_access_flags(struct vm_area_struct *vma, unsigned long address,
461 pud_t *pudp, pud_t entry, int dirty)
462{
463 int changed = !pud_same(*pudp, entry);
464
465 VM_BUG_ON(address & ~HPAGE_PUD_MASK);
466
467 if (changed && dirty) {
468 *pudp = entry;
469 /*
470 * We had a write-protection fault here and changed the pud
471 * to to more permissive. No need to flush the TLB for that,
472 * #PF is architecturally guaranteed to do that and in the
473 * worst-case we'll generate a spurious fault.
474 */
475 }
476
477 return changed;
478}
479#endif
480
481int ptep_test_and_clear_young(struct vm_area_struct *vma,
482 unsigned long addr, pte_t *ptep)
483{
484 int ret = 0;
485
486 if (pte_young(*ptep))
487 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
488 (unsigned long *) &ptep->pte);
489
490 return ret;
491}
492
493#ifdef CONFIG_TRANSPARENT_HUGEPAGE
494int pmdp_test_and_clear_young(struct vm_area_struct *vma,
495 unsigned long addr, pmd_t *pmdp)
496{
497 int ret = 0;
498
499 if (pmd_young(*pmdp))
500 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
501 (unsigned long *)pmdp);
502
503 return ret;
504}
505int pudp_test_and_clear_young(struct vm_area_struct *vma,
506 unsigned long addr, pud_t *pudp)
507{
508 int ret = 0;
509
510 if (pud_young(*pudp))
511 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
512 (unsigned long *)pudp);
513
514 return ret;
515}
516#endif
517
518int ptep_clear_flush_young(struct vm_area_struct *vma,
519 unsigned long address, pte_t *ptep)
520{
521 /*
522 * On x86 CPUs, clearing the accessed bit without a TLB flush
523 * doesn't cause data corruption. [ It could cause incorrect
524 * page aging and the (mistaken) reclaim of hot pages, but the
525 * chance of that should be relatively low. ]
526 *
527 * So as a performance optimization don't flush the TLB when
528 * clearing the accessed bit, it will eventually be flushed by
529 * a context switch or a VM operation anyway. [ In the rare
530 * event of it not getting flushed for a long time the delay
531 * shouldn't really matter because there's no real memory
532 * pressure for swapout to react to. ]
533 */
534 return ptep_test_and_clear_young(vma, address, ptep);
535}
536
537#ifdef CONFIG_TRANSPARENT_HUGEPAGE
538int pmdp_clear_flush_young(struct vm_area_struct *vma,
539 unsigned long address, pmd_t *pmdp)
540{
541 int young;
542
543 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
544
545 young = pmdp_test_and_clear_young(vma, address, pmdp);
546 if (young)
547 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
548
549 return young;
550}
551#endif
552
553/**
554 * reserve_top_address - reserves a hole in the top of kernel address space
555 * @reserve - size of hole to reserve
556 *
557 * Can be used to relocate the fixmap area and poke a hole in the top
558 * of kernel address space to make room for a hypervisor.
559 */
560void __init reserve_top_address(unsigned long reserve)
561{
562#ifdef CONFIG_X86_32
563 BUG_ON(fixmaps_set > 0);
564 __FIXADDR_TOP = round_down(-reserve, 1 << PMD_SHIFT) - PAGE_SIZE;
565 printk(KERN_INFO "Reserving virtual address space above 0x%08lx (rounded to 0x%08lx)\n",
566 -reserve, __FIXADDR_TOP + PAGE_SIZE);
567#endif
568}
569
570int fixmaps_set;
571
572void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
573{
574 unsigned long address = __fix_to_virt(idx);
575
576 if (idx >= __end_of_fixed_addresses) {
577 BUG();
578 return;
579 }
580 set_pte_vaddr(address, pte);
581 fixmaps_set++;
582}
583
584void native_set_fixmap(enum fixed_addresses idx, phys_addr_t phys,
585 pgprot_t flags)
586{
587 /* Sanitize 'prot' against any unsupported bits: */
588 pgprot_val(flags) &= __default_kernel_pte_mask;
589
590 __native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
591}
592
593#ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
594#ifdef CONFIG_X86_5LEVEL
595/**
596 * p4d_set_huge - setup kernel P4D mapping
597 *
598 * No 512GB pages yet -- always return 0
599 */
600int p4d_set_huge(p4d_t *p4d, phys_addr_t addr, pgprot_t prot)
601{
602 return 0;
603}
604
605/**
606 * p4d_clear_huge - clear kernel P4D mapping when it is set
607 *
608 * No 512GB pages yet -- always return 0
609 */
610int p4d_clear_huge(p4d_t *p4d)
611{
612 return 0;
613}
614#endif
615
616/**
617 * pud_set_huge - setup kernel PUD mapping
618 *
619 * MTRRs can override PAT memory types with 4KiB granularity. Therefore, this
620 * function sets up a huge page only if any of the following conditions are met:
621 *
622 * - MTRRs are disabled, or
623 *
624 * - MTRRs are enabled and the range is completely covered by a single MTRR, or
625 *
626 * - MTRRs are enabled and the corresponding MTRR memory type is WB, which
627 * has no effect on the requested PAT memory type.
628 *
629 * Callers should try to decrease page size (1GB -> 2MB -> 4K) if the bigger
630 * page mapping attempt fails.
631 *
632 * Returns 1 on success and 0 on failure.
633 */
634int pud_set_huge(pud_t *pud, phys_addr_t addr, pgprot_t prot)
635{
636 u8 mtrr, uniform;
637
638 mtrr = mtrr_type_lookup(addr, addr + PUD_SIZE, &uniform);
639 if ((mtrr != MTRR_TYPE_INVALID) && (!uniform) &&
640 (mtrr != MTRR_TYPE_WRBACK))
641 return 0;
642
643 /* Bail out if we are we on a populated non-leaf entry: */
644 if (pud_present(*pud) && !pud_huge(*pud))
645 return 0;
646
647 prot = pgprot_4k_2_large(prot);
648
649 set_pte((pte_t *)pud, pfn_pte(
650 (u64)addr >> PAGE_SHIFT,
651 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
652
653 return 1;
654}
655
656/**
657 * pmd_set_huge - setup kernel PMD mapping
658 *
659 * See text over pud_set_huge() above.
660 *
661 * Returns 1 on success and 0 on failure.
662 */
663int pmd_set_huge(pmd_t *pmd, phys_addr_t addr, pgprot_t prot)
664{
665 u8 mtrr, uniform;
666
667 mtrr = mtrr_type_lookup(addr, addr + PMD_SIZE, &uniform);
668 if ((mtrr != MTRR_TYPE_INVALID) && (!uniform) &&
669 (mtrr != MTRR_TYPE_WRBACK)) {
670 pr_warn_once("%s: Cannot satisfy [mem %#010llx-%#010llx] with a huge-page mapping due to MTRR override.\n",
671 __func__, addr, addr + PMD_SIZE);
672 return 0;
673 }
674
675 /* Bail out if we are we on a populated non-leaf entry: */
676 if (pmd_present(*pmd) && !pmd_huge(*pmd))
677 return 0;
678
679 prot = pgprot_4k_2_large(prot);
680
681 set_pte((pte_t *)pmd, pfn_pte(
682 (u64)addr >> PAGE_SHIFT,
683 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
684
685 return 1;
686}
687
688/**
689 * pud_clear_huge - clear kernel PUD mapping when it is set
690 *
691 * Returns 1 on success and 0 on failure (no PUD map is found).
692 */
693int pud_clear_huge(pud_t *pud)
694{
695 if (pud_large(*pud)) {
696 pud_clear(pud);
697 return 1;
698 }
699
700 return 0;
701}
702
703/**
704 * pmd_clear_huge - clear kernel PMD mapping when it is set
705 *
706 * Returns 1 on success and 0 on failure (no PMD map is found).
707 */
708int pmd_clear_huge(pmd_t *pmd)
709{
710 if (pmd_large(*pmd)) {
711 pmd_clear(pmd);
712 return 1;
713 }
714
715 return 0;
716}
717
718/**
719 * pud_free_pmd_page - Clear pud entry and free pmd page.
720 * @pud: Pointer to a PUD.
721 *
722 * Context: The pud range has been unmaped and TLB purged.
723 * Return: 1 if clearing the entry succeeded. 0 otherwise.
724 */
725int pud_free_pmd_page(pud_t *pud)
726{
727 pmd_t *pmd;
728 int i;
729
730 if (pud_none(*pud))
731 return 1;
732
733 pmd = (pmd_t *)pud_page_vaddr(*pud);
734
735 for (i = 0; i < PTRS_PER_PMD; i++)
736 if (!pmd_free_pte_page(&pmd[i]))
737 return 0;
738
739 pud_clear(pud);
740 free_page((unsigned long)pmd);
741
742 return 1;
743}
744
745/**
746 * pmd_free_pte_page - Clear pmd entry and free pte page.
747 * @pmd: Pointer to a PMD.
748 *
749 * Context: The pmd range has been unmaped and TLB purged.
750 * Return: 1 if clearing the entry succeeded. 0 otherwise.
751 */
752int pmd_free_pte_page(pmd_t *pmd)
753{
754 pte_t *pte;
755
756 if (pmd_none(*pmd))
757 return 1;
758
759 pte = (pte_t *)pmd_page_vaddr(*pmd);
760 pmd_clear(pmd);
761 free_page((unsigned long)pte);
762
763 return 1;
764}
765#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#include <asm/mtrr.h>
8
9#define PGALLOC_GFP (GFP_KERNEL_ACCOUNT | __GFP_NOTRACK | __GFP_ZERO)
10
11#ifdef CONFIG_HIGHPTE
12#define PGALLOC_USER_GFP __GFP_HIGHMEM
13#else
14#define PGALLOC_USER_GFP 0
15#endif
16
17gfp_t __userpte_alloc_gfp = PGALLOC_GFP | PGALLOC_USER_GFP;
18
19pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
20{
21 return (pte_t *)__get_free_page(PGALLOC_GFP & ~__GFP_ACCOUNT);
22}
23
24pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
25{
26 struct page *pte;
27
28 pte = alloc_pages(__userpte_alloc_gfp, 0);
29 if (!pte)
30 return NULL;
31 if (!pgtable_page_ctor(pte)) {
32 __free_page(pte);
33 return NULL;
34 }
35 return pte;
36}
37
38static int __init setup_userpte(char *arg)
39{
40 if (!arg)
41 return -EINVAL;
42
43 /*
44 * "userpte=nohigh" disables allocation of user pagetables in
45 * high memory.
46 */
47 if (strcmp(arg, "nohigh") == 0)
48 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
49 else
50 return -EINVAL;
51 return 0;
52}
53early_param("userpte", setup_userpte);
54
55void ___pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
56{
57 pgtable_page_dtor(pte);
58 paravirt_release_pte(page_to_pfn(pte));
59 tlb_remove_page(tlb, pte);
60}
61
62#if CONFIG_PGTABLE_LEVELS > 2
63void ___pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
64{
65 struct page *page = virt_to_page(pmd);
66 paravirt_release_pmd(__pa(pmd) >> PAGE_SHIFT);
67 /*
68 * NOTE! For PAE, any changes to the top page-directory-pointer-table
69 * entries need a full cr3 reload to flush.
70 */
71#ifdef CONFIG_X86_PAE
72 tlb->need_flush_all = 1;
73#endif
74 pgtable_pmd_page_dtor(page);
75 tlb_remove_page(tlb, page);
76}
77
78#if CONFIG_PGTABLE_LEVELS > 3
79void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
80{
81 paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
82 tlb_remove_page(tlb, virt_to_page(pud));
83}
84#endif /* CONFIG_PGTABLE_LEVELS > 3 */
85#endif /* CONFIG_PGTABLE_LEVELS > 2 */
86
87static inline void pgd_list_add(pgd_t *pgd)
88{
89 struct page *page = virt_to_page(pgd);
90
91 list_add(&page->lru, &pgd_list);
92}
93
94static inline void pgd_list_del(pgd_t *pgd)
95{
96 struct page *page = virt_to_page(pgd);
97
98 list_del(&page->lru);
99}
100
101#define UNSHARED_PTRS_PER_PGD \
102 (SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
103
104
105static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
106{
107 BUILD_BUG_ON(sizeof(virt_to_page(pgd)->index) < sizeof(mm));
108 virt_to_page(pgd)->index = (pgoff_t)mm;
109}
110
111struct mm_struct *pgd_page_get_mm(struct page *page)
112{
113 return (struct mm_struct *)page->index;
114}
115
116static void pgd_ctor(struct mm_struct *mm, pgd_t *pgd)
117{
118 /* If the pgd points to a shared pagetable level (either the
119 ptes in non-PAE, or shared PMD in PAE), then just copy the
120 references from swapper_pg_dir. */
121 if (CONFIG_PGTABLE_LEVELS == 2 ||
122 (CONFIG_PGTABLE_LEVELS == 3 && SHARED_KERNEL_PMD) ||
123 CONFIG_PGTABLE_LEVELS == 4) {
124 clone_pgd_range(pgd + KERNEL_PGD_BOUNDARY,
125 swapper_pg_dir + KERNEL_PGD_BOUNDARY,
126 KERNEL_PGD_PTRS);
127 }
128
129 /* list required to sync kernel mapping updates */
130 if (!SHARED_KERNEL_PMD) {
131 pgd_set_mm(pgd, mm);
132 pgd_list_add(pgd);
133 }
134}
135
136static void pgd_dtor(pgd_t *pgd)
137{
138 if (SHARED_KERNEL_PMD)
139 return;
140
141 spin_lock(&pgd_lock);
142 pgd_list_del(pgd);
143 spin_unlock(&pgd_lock);
144}
145
146/*
147 * List of all pgd's needed for non-PAE so it can invalidate entries
148 * in both cached and uncached pgd's; not needed for PAE since the
149 * kernel pmd is shared. If PAE were not to share the pmd a similar
150 * tactic would be needed. This is essentially codepath-based locking
151 * against pageattr.c; it is the unique case in which a valid change
152 * of kernel pagetables can't be lazily synchronized by vmalloc faults.
153 * vmalloc faults work because attached pagetables are never freed.
154 * -- nyc
155 */
156
157#ifdef CONFIG_X86_PAE
158/*
159 * In PAE mode, we need to do a cr3 reload (=tlb flush) when
160 * updating the top-level pagetable entries to guarantee the
161 * processor notices the update. Since this is expensive, and
162 * all 4 top-level entries are used almost immediately in a
163 * new process's life, we just pre-populate them here.
164 *
165 * Also, if we're in a paravirt environment where the kernel pmd is
166 * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
167 * and initialize the kernel pmds here.
168 */
169#define PREALLOCATED_PMDS UNSHARED_PTRS_PER_PGD
170
171void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
172{
173 paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
174
175 /* Note: almost everything apart from _PAGE_PRESENT is
176 reserved at the pmd (PDPT) level. */
177 set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
178
179 /*
180 * According to Intel App note "TLBs, Paging-Structure Caches,
181 * and Their Invalidation", April 2007, document 317080-001,
182 * section 8.1: in PAE mode we explicitly have to flush the
183 * TLB via cr3 if the top-level pgd is changed...
184 */
185 flush_tlb_mm(mm);
186}
187#else /* !CONFIG_X86_PAE */
188
189/* No need to prepopulate any pagetable entries in non-PAE modes. */
190#define PREALLOCATED_PMDS 0
191
192#endif /* CONFIG_X86_PAE */
193
194static void free_pmds(struct mm_struct *mm, pmd_t *pmds[])
195{
196 int i;
197
198 for(i = 0; i < PREALLOCATED_PMDS; i++)
199 if (pmds[i]) {
200 pgtable_pmd_page_dtor(virt_to_page(pmds[i]));
201 free_page((unsigned long)pmds[i]);
202 mm_dec_nr_pmds(mm);
203 }
204}
205
206static int preallocate_pmds(struct mm_struct *mm, pmd_t *pmds[])
207{
208 int i;
209 bool failed = false;
210 gfp_t gfp = PGALLOC_GFP;
211
212 if (mm == &init_mm)
213 gfp &= ~__GFP_ACCOUNT;
214
215 for(i = 0; i < PREALLOCATED_PMDS; i++) {
216 pmd_t *pmd = (pmd_t *)__get_free_page(gfp);
217 if (!pmd)
218 failed = true;
219 if (pmd && !pgtable_pmd_page_ctor(virt_to_page(pmd))) {
220 free_page((unsigned long)pmd);
221 pmd = NULL;
222 failed = true;
223 }
224 if (pmd)
225 mm_inc_nr_pmds(mm);
226 pmds[i] = pmd;
227 }
228
229 if (failed) {
230 free_pmds(mm, pmds);
231 return -ENOMEM;
232 }
233
234 return 0;
235}
236
237/*
238 * Mop up any pmd pages which may still be attached to the pgd.
239 * Normally they will be freed by munmap/exit_mmap, but any pmd we
240 * preallocate which never got a corresponding vma will need to be
241 * freed manually.
242 */
243static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
244{
245 int i;
246
247 for(i = 0; i < PREALLOCATED_PMDS; i++) {
248 pgd_t pgd = pgdp[i];
249
250 if (pgd_val(pgd) != 0) {
251 pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
252
253 pgdp[i] = native_make_pgd(0);
254
255 paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
256 pmd_free(mm, pmd);
257 mm_dec_nr_pmds(mm);
258 }
259 }
260}
261
262static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
263{
264 pud_t *pud;
265 int i;
266
267 if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
268 return;
269
270 pud = pud_offset(pgd, 0);
271
272 for (i = 0; i < PREALLOCATED_PMDS; i++, pud++) {
273 pmd_t *pmd = pmds[i];
274
275 if (i >= KERNEL_PGD_BOUNDARY)
276 memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
277 sizeof(pmd_t) * PTRS_PER_PMD);
278
279 pud_populate(mm, pud, pmd);
280 }
281}
282
283/*
284 * Xen paravirt assumes pgd table should be in one page. 64 bit kernel also
285 * assumes that pgd should be in one page.
286 *
287 * But kernel with PAE paging that is not running as a Xen domain
288 * only needs to allocate 32 bytes for pgd instead of one page.
289 */
290#ifdef CONFIG_X86_PAE
291
292#include <linux/slab.h>
293
294#define PGD_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
295#define PGD_ALIGN 32
296
297static struct kmem_cache *pgd_cache;
298
299static int __init pgd_cache_init(void)
300{
301 /*
302 * When PAE kernel is running as a Xen domain, it does not use
303 * shared kernel pmd. And this requires a whole page for pgd.
304 */
305 if (!SHARED_KERNEL_PMD)
306 return 0;
307
308 /*
309 * when PAE kernel is not running as a Xen domain, it uses
310 * shared kernel pmd. Shared kernel pmd does not require a whole
311 * page for pgd. We are able to just allocate a 32-byte for pgd.
312 * During boot time, we create a 32-byte slab for pgd table allocation.
313 */
314 pgd_cache = kmem_cache_create("pgd_cache", PGD_SIZE, PGD_ALIGN,
315 SLAB_PANIC, NULL);
316 if (!pgd_cache)
317 return -ENOMEM;
318
319 return 0;
320}
321core_initcall(pgd_cache_init);
322
323static inline pgd_t *_pgd_alloc(void)
324{
325 /*
326 * If no SHARED_KERNEL_PMD, PAE kernel is running as a Xen domain.
327 * We allocate one page for pgd.
328 */
329 if (!SHARED_KERNEL_PMD)
330 return (pgd_t *)__get_free_page(PGALLOC_GFP);
331
332 /*
333 * Now PAE kernel is not running as a Xen domain. We can allocate
334 * a 32-byte slab for pgd to save memory space.
335 */
336 return kmem_cache_alloc(pgd_cache, PGALLOC_GFP);
337}
338
339static inline void _pgd_free(pgd_t *pgd)
340{
341 if (!SHARED_KERNEL_PMD)
342 free_page((unsigned long)pgd);
343 else
344 kmem_cache_free(pgd_cache, pgd);
345}
346#else
347static inline pgd_t *_pgd_alloc(void)
348{
349 return (pgd_t *)__get_free_page(PGALLOC_GFP);
350}
351
352static inline void _pgd_free(pgd_t *pgd)
353{
354 free_page((unsigned long)pgd);
355}
356#endif /* CONFIG_X86_PAE */
357
358pgd_t *pgd_alloc(struct mm_struct *mm)
359{
360 pgd_t *pgd;
361 pmd_t *pmds[PREALLOCATED_PMDS];
362
363 pgd = _pgd_alloc();
364
365 if (pgd == NULL)
366 goto out;
367
368 mm->pgd = pgd;
369
370 if (preallocate_pmds(mm, pmds) != 0)
371 goto out_free_pgd;
372
373 if (paravirt_pgd_alloc(mm) != 0)
374 goto out_free_pmds;
375
376 /*
377 * Make sure that pre-populating the pmds is atomic with
378 * respect to anything walking the pgd_list, so that they
379 * never see a partially populated pgd.
380 */
381 spin_lock(&pgd_lock);
382
383 pgd_ctor(mm, pgd);
384 pgd_prepopulate_pmd(mm, pgd, pmds);
385
386 spin_unlock(&pgd_lock);
387
388 return pgd;
389
390out_free_pmds:
391 free_pmds(mm, pmds);
392out_free_pgd:
393 _pgd_free(pgd);
394out:
395 return NULL;
396}
397
398void pgd_free(struct mm_struct *mm, pgd_t *pgd)
399{
400 pgd_mop_up_pmds(mm, pgd);
401 pgd_dtor(pgd);
402 paravirt_pgd_free(mm, pgd);
403 _pgd_free(pgd);
404}
405
406/*
407 * Used to set accessed or dirty bits in the page table entries
408 * on other architectures. On x86, the accessed and dirty bits
409 * are tracked by hardware. However, do_wp_page calls this function
410 * to also make the pte writeable at the same time the dirty bit is
411 * set. In that case we do actually need to write the PTE.
412 */
413int ptep_set_access_flags(struct vm_area_struct *vma,
414 unsigned long address, pte_t *ptep,
415 pte_t entry, int dirty)
416{
417 int changed = !pte_same(*ptep, entry);
418
419 if (changed && dirty) {
420 *ptep = entry;
421 pte_update(vma->vm_mm, address, ptep);
422 }
423
424 return changed;
425}
426
427#ifdef CONFIG_TRANSPARENT_HUGEPAGE
428int pmdp_set_access_flags(struct vm_area_struct *vma,
429 unsigned long address, pmd_t *pmdp,
430 pmd_t entry, int dirty)
431{
432 int changed = !pmd_same(*pmdp, entry);
433
434 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
435
436 if (changed && dirty) {
437 *pmdp = entry;
438 /*
439 * We had a write-protection fault here and changed the pmd
440 * to to more permissive. No need to flush the TLB for that,
441 * #PF is architecturally guaranteed to do that and in the
442 * worst-case we'll generate a spurious fault.
443 */
444 }
445
446 return changed;
447}
448#endif
449
450int ptep_test_and_clear_young(struct vm_area_struct *vma,
451 unsigned long addr, pte_t *ptep)
452{
453 int ret = 0;
454
455 if (pte_young(*ptep))
456 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
457 (unsigned long *) &ptep->pte);
458
459 if (ret)
460 pte_update(vma->vm_mm, addr, ptep);
461
462 return ret;
463}
464
465#ifdef CONFIG_TRANSPARENT_HUGEPAGE
466int pmdp_test_and_clear_young(struct vm_area_struct *vma,
467 unsigned long addr, pmd_t *pmdp)
468{
469 int ret = 0;
470
471 if (pmd_young(*pmdp))
472 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
473 (unsigned long *)pmdp);
474
475 return ret;
476}
477#endif
478
479int ptep_clear_flush_young(struct vm_area_struct *vma,
480 unsigned long address, pte_t *ptep)
481{
482 /*
483 * On x86 CPUs, clearing the accessed bit without a TLB flush
484 * doesn't cause data corruption. [ It could cause incorrect
485 * page aging and the (mistaken) reclaim of hot pages, but the
486 * chance of that should be relatively low. ]
487 *
488 * So as a performance optimization don't flush the TLB when
489 * clearing the accessed bit, it will eventually be flushed by
490 * a context switch or a VM operation anyway. [ In the rare
491 * event of it not getting flushed for a long time the delay
492 * shouldn't really matter because there's no real memory
493 * pressure for swapout to react to. ]
494 */
495 return ptep_test_and_clear_young(vma, address, ptep);
496}
497
498#ifdef CONFIG_TRANSPARENT_HUGEPAGE
499int pmdp_clear_flush_young(struct vm_area_struct *vma,
500 unsigned long address, pmd_t *pmdp)
501{
502 int young;
503
504 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
505
506 young = pmdp_test_and_clear_young(vma, address, pmdp);
507 if (young)
508 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
509
510 return young;
511}
512#endif
513
514/**
515 * reserve_top_address - reserves a hole in the top of kernel address space
516 * @reserve - size of hole to reserve
517 *
518 * Can be used to relocate the fixmap area and poke a hole in the top
519 * of kernel address space to make room for a hypervisor.
520 */
521void __init reserve_top_address(unsigned long reserve)
522{
523#ifdef CONFIG_X86_32
524 BUG_ON(fixmaps_set > 0);
525 __FIXADDR_TOP = round_down(-reserve, 1 << PMD_SHIFT) - PAGE_SIZE;
526 printk(KERN_INFO "Reserving virtual address space above 0x%08lx (rounded to 0x%08lx)\n",
527 -reserve, __FIXADDR_TOP + PAGE_SIZE);
528#endif
529}
530
531int fixmaps_set;
532
533void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
534{
535 unsigned long address = __fix_to_virt(idx);
536
537 if (idx >= __end_of_fixed_addresses) {
538 BUG();
539 return;
540 }
541 set_pte_vaddr(address, pte);
542 fixmaps_set++;
543}
544
545void native_set_fixmap(enum fixed_addresses idx, phys_addr_t phys,
546 pgprot_t flags)
547{
548 __native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
549}
550
551#ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
552/**
553 * pud_set_huge - setup kernel PUD mapping
554 *
555 * MTRRs can override PAT memory types with 4KiB granularity. Therefore, this
556 * function sets up a huge page only if any of the following conditions are met:
557 *
558 * - MTRRs are disabled, or
559 *
560 * - MTRRs are enabled and the range is completely covered by a single MTRR, or
561 *
562 * - MTRRs are enabled and the corresponding MTRR memory type is WB, which
563 * has no effect on the requested PAT memory type.
564 *
565 * Callers should try to decrease page size (1GB -> 2MB -> 4K) if the bigger
566 * page mapping attempt fails.
567 *
568 * Returns 1 on success and 0 on failure.
569 */
570int pud_set_huge(pud_t *pud, phys_addr_t addr, pgprot_t prot)
571{
572 u8 mtrr, uniform;
573
574 mtrr = mtrr_type_lookup(addr, addr + PUD_SIZE, &uniform);
575 if ((mtrr != MTRR_TYPE_INVALID) && (!uniform) &&
576 (mtrr != MTRR_TYPE_WRBACK))
577 return 0;
578
579 prot = pgprot_4k_2_large(prot);
580
581 set_pte((pte_t *)pud, pfn_pte(
582 (u64)addr >> PAGE_SHIFT,
583 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
584
585 return 1;
586}
587
588/**
589 * pmd_set_huge - setup kernel PMD mapping
590 *
591 * See text over pud_set_huge() above.
592 *
593 * Returns 1 on success and 0 on failure.
594 */
595int pmd_set_huge(pmd_t *pmd, phys_addr_t addr, pgprot_t prot)
596{
597 u8 mtrr, uniform;
598
599 mtrr = mtrr_type_lookup(addr, addr + PMD_SIZE, &uniform);
600 if ((mtrr != MTRR_TYPE_INVALID) && (!uniform) &&
601 (mtrr != MTRR_TYPE_WRBACK)) {
602 pr_warn_once("%s: Cannot satisfy [mem %#010llx-%#010llx] with a huge-page mapping due to MTRR override.\n",
603 __func__, addr, addr + PMD_SIZE);
604 return 0;
605 }
606
607 prot = pgprot_4k_2_large(prot);
608
609 set_pte((pte_t *)pmd, pfn_pte(
610 (u64)addr >> PAGE_SHIFT,
611 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
612
613 return 1;
614}
615
616/**
617 * pud_clear_huge - clear kernel PUD mapping when it is set
618 *
619 * Returns 1 on success and 0 on failure (no PUD map is found).
620 */
621int pud_clear_huge(pud_t *pud)
622{
623 if (pud_large(*pud)) {
624 pud_clear(pud);
625 return 1;
626 }
627
628 return 0;
629}
630
631/**
632 * pmd_clear_huge - clear kernel PMD mapping when it is set
633 *
634 * Returns 1 on success and 0 on failure (no PMD map is found).
635 */
636int pmd_clear_huge(pmd_t *pmd)
637{
638 if (pmd_large(*pmd)) {
639 pmd_clear(pmd);
640 return 1;
641 }
642
643 return 0;
644}
645#endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */