1 // SPDX-License-Identifier: GPL-2.0
4 #include <linux/hugetlb.h>
5 #include <asm/pgalloc.h>
6 #include <asm/pgtable.h>
8 #include <asm/fixmap.h>
11 #ifdef CONFIG_DYNAMIC_PHYSICAL_MASK
12 phys_addr_t physical_mask __ro_after_init = (1ULL << __PHYSICAL_MASK_SHIFT) - 1;
13 EXPORT_SYMBOL(physical_mask);
16 #define PGALLOC_GFP (GFP_KERNEL_ACCOUNT | __GFP_ZERO)
19 #define PGALLOC_USER_GFP __GFP_HIGHMEM
21 #define PGALLOC_USER_GFP 0
24 gfp_t __userpte_alloc_gfp = PGALLOC_GFP | PGALLOC_USER_GFP;
26 pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
28 return (pte_t *)__get_free_page(PGALLOC_GFP & ~__GFP_ACCOUNT);
31 pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
35 pte = alloc_pages(__userpte_alloc_gfp, 0);
38 if (!pgtable_page_ctor(pte)) {
45 static int __init setup_userpte(char *arg)
51 * "userpte=nohigh" disables allocation of user pagetables in
54 if (strcmp(arg, "nohigh") == 0)
55 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
60 early_param("userpte", setup_userpte);
62 void ___pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
64 pgtable_page_dtor(pte);
65 paravirt_release_pte(page_to_pfn(pte));
66 tlb_remove_table(tlb, pte);
69 #if CONFIG_PGTABLE_LEVELS > 2
70 void ___pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
72 struct page *page = virt_to_page(pmd);
73 paravirt_release_pmd(__pa(pmd) >> PAGE_SHIFT);
75 * NOTE! For PAE, any changes to the top page-directory-pointer-table
76 * entries need a full cr3 reload to flush.
79 tlb->need_flush_all = 1;
81 pgtable_pmd_page_dtor(page);
82 tlb_remove_table(tlb, page);
85 #if CONFIG_PGTABLE_LEVELS > 3
86 void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
88 paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
89 tlb_remove_table(tlb, virt_to_page(pud));
92 #if CONFIG_PGTABLE_LEVELS > 4
93 void ___p4d_free_tlb(struct mmu_gather *tlb, p4d_t *p4d)
95 paravirt_release_p4d(__pa(p4d) >> PAGE_SHIFT);
96 tlb_remove_table(tlb, virt_to_page(p4d));
98 #endif /* CONFIG_PGTABLE_LEVELS > 4 */
99 #endif /* CONFIG_PGTABLE_LEVELS > 3 */
100 #endif /* CONFIG_PGTABLE_LEVELS > 2 */
102 static inline void pgd_list_add(pgd_t *pgd)
104 struct page *page = virt_to_page(pgd);
106 list_add(&page->lru, &pgd_list);
109 static inline void pgd_list_del(pgd_t *pgd)
111 struct page *page = virt_to_page(pgd);
113 list_del(&page->lru);
116 #define UNSHARED_PTRS_PER_PGD \
117 (SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
120 static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
122 virt_to_page(pgd)->pt_mm = mm;
125 struct mm_struct *pgd_page_get_mm(struct page *page)
130 static void pgd_ctor(struct mm_struct *mm, pgd_t *pgd)
132 /* If the pgd points to a shared pagetable level (either the
133 ptes in non-PAE, or shared PMD in PAE), then just copy the
134 references from swapper_pg_dir. */
135 if (CONFIG_PGTABLE_LEVELS == 2 ||
136 (CONFIG_PGTABLE_LEVELS == 3 && SHARED_KERNEL_PMD) ||
137 CONFIG_PGTABLE_LEVELS >= 4) {
138 clone_pgd_range(pgd + KERNEL_PGD_BOUNDARY,
139 swapper_pg_dir + KERNEL_PGD_BOUNDARY,
143 /* list required to sync kernel mapping updates */
144 if (!SHARED_KERNEL_PMD) {
150 static void pgd_dtor(pgd_t *pgd)
152 if (SHARED_KERNEL_PMD)
155 spin_lock(&pgd_lock);
157 spin_unlock(&pgd_lock);
161 * List of all pgd's needed for non-PAE so it can invalidate entries
162 * in both cached and uncached pgd's; not needed for PAE since the
163 * kernel pmd is shared. If PAE were not to share the pmd a similar
164 * tactic would be needed. This is essentially codepath-based locking
165 * against pageattr.c; it is the unique case in which a valid change
166 * of kernel pagetables can't be lazily synchronized by vmalloc faults.
167 * vmalloc faults work because attached pagetables are never freed.
171 #ifdef CONFIG_X86_PAE
173 * In PAE mode, we need to do a cr3 reload (=tlb flush) when
174 * updating the top-level pagetable entries to guarantee the
175 * processor notices the update. Since this is expensive, and
176 * all 4 top-level entries are used almost immediately in a
177 * new process's life, we just pre-populate them here.
179 * Also, if we're in a paravirt environment where the kernel pmd is
180 * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
181 * and initialize the kernel pmds here.
183 #define PREALLOCATED_PMDS UNSHARED_PTRS_PER_PGD
185 void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
187 paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
189 /* Note: almost everything apart from _PAGE_PRESENT is
190 reserved at the pmd (PDPT) level. */
191 set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
194 * According to Intel App note "TLBs, Paging-Structure Caches,
195 * and Their Invalidation", April 2007, document 317080-001,
196 * section 8.1: in PAE mode we explicitly have to flush the
197 * TLB via cr3 if the top-level pgd is changed...
201 #else /* !CONFIG_X86_PAE */
203 /* No need to prepopulate any pagetable entries in non-PAE modes. */
204 #define PREALLOCATED_PMDS 0
206 #endif /* CONFIG_X86_PAE */
208 static void free_pmds(struct mm_struct *mm, pmd_t *pmds[])
212 for(i = 0; i < PREALLOCATED_PMDS; i++)
214 pgtable_pmd_page_dtor(virt_to_page(pmds[i]));
215 free_page((unsigned long)pmds[i]);
220 static int preallocate_pmds(struct mm_struct *mm, pmd_t *pmds[])
224 gfp_t gfp = PGALLOC_GFP;
227 gfp &= ~__GFP_ACCOUNT;
229 for(i = 0; i < PREALLOCATED_PMDS; i++) {
230 pmd_t *pmd = (pmd_t *)__get_free_page(gfp);
233 if (pmd && !pgtable_pmd_page_ctor(virt_to_page(pmd))) {
234 free_page((unsigned long)pmd);
252 * Mop up any pmd pages which may still be attached to the pgd.
253 * Normally they will be freed by munmap/exit_mmap, but any pmd we
254 * preallocate which never got a corresponding vma will need to be
257 static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
261 for(i = 0; i < PREALLOCATED_PMDS; i++) {
264 if (pgd_val(pgd) != 0) {
265 pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
267 pgdp[i] = native_make_pgd(0);
269 paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
276 static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
282 if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
285 p4d = p4d_offset(pgd, 0);
286 pud = pud_offset(p4d, 0);
288 for (i = 0; i < PREALLOCATED_PMDS; i++, pud++) {
289 pmd_t *pmd = pmds[i];
291 if (i >= KERNEL_PGD_BOUNDARY)
292 memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
293 sizeof(pmd_t) * PTRS_PER_PMD);
295 pud_populate(mm, pud, pmd);
300 * Xen paravirt assumes pgd table should be in one page. 64 bit kernel also
301 * assumes that pgd should be in one page.
303 * But kernel with PAE paging that is not running as a Xen domain
304 * only needs to allocate 32 bytes for pgd instead of one page.
306 #ifdef CONFIG_X86_PAE
308 #include <linux/slab.h>
310 #define PGD_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
313 static struct kmem_cache *pgd_cache;
315 static int __init pgd_cache_init(void)
318 * When PAE kernel is running as a Xen domain, it does not use
319 * shared kernel pmd. And this requires a whole page for pgd.
321 if (!SHARED_KERNEL_PMD)
325 * when PAE kernel is not running as a Xen domain, it uses
326 * shared kernel pmd. Shared kernel pmd does not require a whole
327 * page for pgd. We are able to just allocate a 32-byte for pgd.
328 * During boot time, we create a 32-byte slab for pgd table allocation.
330 pgd_cache = kmem_cache_create("pgd_cache", PGD_SIZE, PGD_ALIGN,
337 core_initcall(pgd_cache_init);
339 static inline pgd_t *_pgd_alloc(void)
342 * If no SHARED_KERNEL_PMD, PAE kernel is running as a Xen domain.
343 * We allocate one page for pgd.
345 if (!SHARED_KERNEL_PMD)
346 return (pgd_t *)__get_free_pages(PGALLOC_GFP,
347 PGD_ALLOCATION_ORDER);
350 * Now PAE kernel is not running as a Xen domain. We can allocate
351 * a 32-byte slab for pgd to save memory space.
353 return kmem_cache_alloc(pgd_cache, PGALLOC_GFP);
356 static inline void _pgd_free(pgd_t *pgd)
358 if (!SHARED_KERNEL_PMD)
359 free_pages((unsigned long)pgd, PGD_ALLOCATION_ORDER);
361 kmem_cache_free(pgd_cache, pgd);
365 static inline pgd_t *_pgd_alloc(void)
367 return (pgd_t *)__get_free_pages(PGALLOC_GFP, PGD_ALLOCATION_ORDER);
370 static inline void _pgd_free(pgd_t *pgd)
372 free_pages((unsigned long)pgd, PGD_ALLOCATION_ORDER);
374 #endif /* CONFIG_X86_PAE */
376 pgd_t *pgd_alloc(struct mm_struct *mm)
379 pmd_t *pmds[PREALLOCATED_PMDS];
388 if (preallocate_pmds(mm, pmds) != 0)
391 if (paravirt_pgd_alloc(mm) != 0)
395 * Make sure that pre-populating the pmds is atomic with
396 * respect to anything walking the pgd_list, so that they
397 * never see a partially populated pgd.
399 spin_lock(&pgd_lock);
402 pgd_prepopulate_pmd(mm, pgd, pmds);
404 spin_unlock(&pgd_lock);
416 void pgd_free(struct mm_struct *mm, pgd_t *pgd)
418 pgd_mop_up_pmds(mm, pgd);
420 paravirt_pgd_free(mm, pgd);
425 * Used to set accessed or dirty bits in the page table entries
426 * on other architectures. On x86, the accessed and dirty bits
427 * are tracked by hardware. However, do_wp_page calls this function
428 * to also make the pte writeable at the same time the dirty bit is
429 * set. In that case we do actually need to write the PTE.
431 int ptep_set_access_flags(struct vm_area_struct *vma,
432 unsigned long address, pte_t *ptep,
433 pte_t entry, int dirty)
435 int changed = !pte_same(*ptep, entry);
437 if (changed && dirty)
443 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
444 int pmdp_set_access_flags(struct vm_area_struct *vma,
445 unsigned long address, pmd_t *pmdp,
446 pmd_t entry, int dirty)
448 int changed = !pmd_same(*pmdp, entry);
450 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
452 if (changed && dirty) {
455 * We had a write-protection fault here and changed the pmd
456 * to to more permissive. No need to flush the TLB for that,
457 * #PF is architecturally guaranteed to do that and in the
458 * worst-case we'll generate a spurious fault.
465 int pudp_set_access_flags(struct vm_area_struct *vma, unsigned long address,
466 pud_t *pudp, pud_t entry, int dirty)
468 int changed = !pud_same(*pudp, entry);
470 VM_BUG_ON(address & ~HPAGE_PUD_MASK);
472 if (changed && dirty) {
475 * We had a write-protection fault here and changed the pud
476 * to to more permissive. No need to flush the TLB for that,
477 * #PF is architecturally guaranteed to do that and in the
478 * worst-case we'll generate a spurious fault.
486 int ptep_test_and_clear_young(struct vm_area_struct *vma,
487 unsigned long addr, pte_t *ptep)
491 if (pte_young(*ptep))
492 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
493 (unsigned long *) &ptep->pte);
498 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
499 int pmdp_test_and_clear_young(struct vm_area_struct *vma,
500 unsigned long addr, pmd_t *pmdp)
504 if (pmd_young(*pmdp))
505 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
506 (unsigned long *)pmdp);
510 int pudp_test_and_clear_young(struct vm_area_struct *vma,
511 unsigned long addr, pud_t *pudp)
515 if (pud_young(*pudp))
516 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
517 (unsigned long *)pudp);
523 int ptep_clear_flush_young(struct vm_area_struct *vma,
524 unsigned long address, pte_t *ptep)
527 * On x86 CPUs, clearing the accessed bit without a TLB flush
528 * doesn't cause data corruption. [ It could cause incorrect
529 * page aging and the (mistaken) reclaim of hot pages, but the
530 * chance of that should be relatively low. ]
532 * So as a performance optimization don't flush the TLB when
533 * clearing the accessed bit, it will eventually be flushed by
534 * a context switch or a VM operation anyway. [ In the rare
535 * event of it not getting flushed for a long time the delay
536 * shouldn't really matter because there's no real memory
537 * pressure for swapout to react to. ]
539 return ptep_test_and_clear_young(vma, address, ptep);
542 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
543 int pmdp_clear_flush_young(struct vm_area_struct *vma,
544 unsigned long address, pmd_t *pmdp)
548 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
550 young = pmdp_test_and_clear_young(vma, address, pmdp);
552 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
559 * reserve_top_address - reserves a hole in the top of kernel address space
560 * @reserve - size of hole to reserve
562 * Can be used to relocate the fixmap area and poke a hole in the top
563 * of kernel address space to make room for a hypervisor.
565 void __init reserve_top_address(unsigned long reserve)
568 BUG_ON(fixmaps_set > 0);
569 __FIXADDR_TOP = round_down(-reserve, 1 << PMD_SHIFT) - PAGE_SIZE;
570 printk(KERN_INFO "Reserving virtual address space above 0x%08lx (rounded to 0x%08lx)\n",
571 -reserve, __FIXADDR_TOP + PAGE_SIZE);
577 void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
579 unsigned long address = __fix_to_virt(idx);
581 if (idx >= __end_of_fixed_addresses) {
585 set_pte_vaddr(address, pte);
589 void native_set_fixmap(enum fixed_addresses idx, phys_addr_t phys,
592 /* Sanitize 'prot' against any unsupported bits: */
593 pgprot_val(flags) &= __default_kernel_pte_mask;
595 __native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
598 #ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
599 #ifdef CONFIG_X86_5LEVEL
601 * p4d_set_huge - setup kernel P4D mapping
603 * No 512GB pages yet -- always return 0
605 int p4d_set_huge(p4d_t *p4d, phys_addr_t addr, pgprot_t prot)
611 * p4d_clear_huge - clear kernel P4D mapping when it is set
613 * No 512GB pages yet -- always return 0
615 int p4d_clear_huge(p4d_t *p4d)
622 * pud_set_huge - setup kernel PUD mapping
624 * MTRRs can override PAT memory types with 4KiB granularity. Therefore, this
625 * function sets up a huge page only if any of the following conditions are met:
627 * - MTRRs are disabled, or
629 * - MTRRs are enabled and the range is completely covered by a single MTRR, or
631 * - MTRRs are enabled and the corresponding MTRR memory type is WB, which
632 * has no effect on the requested PAT memory type.
634 * Callers should try to decrease page size (1GB -> 2MB -> 4K) if the bigger
635 * page mapping attempt fails.
637 * Returns 1 on success and 0 on failure.
639 int pud_set_huge(pud_t *pud, phys_addr_t addr, pgprot_t prot)
643 mtrr = mtrr_type_lookup(addr, addr + PUD_SIZE, &uniform);
644 if ((mtrr != MTRR_TYPE_INVALID) && (!uniform) &&
645 (mtrr != MTRR_TYPE_WRBACK))
648 /* Bail out if we are we on a populated non-leaf entry: */
649 if (pud_present(*pud) && !pud_huge(*pud))
652 prot = pgprot_4k_2_large(prot);
654 set_pte((pte_t *)pud, pfn_pte(
655 (u64)addr >> PAGE_SHIFT,
656 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
662 * pmd_set_huge - setup kernel PMD mapping
664 * See text over pud_set_huge() above.
666 * Returns 1 on success and 0 on failure.
668 int pmd_set_huge(pmd_t *pmd, phys_addr_t addr, pgprot_t prot)
672 mtrr = mtrr_type_lookup(addr, addr + PMD_SIZE, &uniform);
673 if ((mtrr != MTRR_TYPE_INVALID) && (!uniform) &&
674 (mtrr != MTRR_TYPE_WRBACK)) {
675 pr_warn_once("%s: Cannot satisfy [mem %#010llx-%#010llx] with a huge-page mapping due to MTRR override.\n",
676 __func__, addr, addr + PMD_SIZE);
680 /* Bail out if we are we on a populated non-leaf entry: */
681 if (pmd_present(*pmd) && !pmd_huge(*pmd))
684 prot = pgprot_4k_2_large(prot);
686 set_pte((pte_t *)pmd, pfn_pte(
687 (u64)addr >> PAGE_SHIFT,
688 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
694 * pud_clear_huge - clear kernel PUD mapping when it is set
696 * Returns 1 on success and 0 on failure (no PUD map is found).
698 int pud_clear_huge(pud_t *pud)
700 if (pud_large(*pud)) {
709 * pmd_clear_huge - clear kernel PMD mapping when it is set
711 * Returns 1 on success and 0 on failure (no PMD map is found).
713 int pmd_clear_huge(pmd_t *pmd)
715 if (pmd_large(*pmd)) {
724 * pud_free_pmd_page - Clear pud entry and free pmd page.
725 * @pud: Pointer to a PUD.
727 * Context: The pud range has been unmaped and TLB purged.
728 * Return: 1 if clearing the entry succeeded. 0 otherwise.
730 int pud_free_pmd_page(pud_t *pud)
738 pmd = (pmd_t *)pud_page_vaddr(*pud);
740 for (i = 0; i < PTRS_PER_PMD; i++)
741 if (!pmd_free_pte_page(&pmd[i]))
745 free_page((unsigned long)pmd);
751 * pmd_free_pte_page - Clear pmd entry and free pte page.
752 * @pmd: Pointer to a PMD.
754 * Context: The pmd range has been unmaped and TLB purged.
755 * Return: 1 if clearing the entry succeeded. 0 otherwise.
757 int pmd_free_pte_page(pmd_t *pmd)
764 pte = (pte_t *)pmd_page_vaddr(*pmd);
766 free_page((unsigned long)pte);
770 #endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */