1 // SPDX-License-Identifier: GPL-2.0
4 #include <asm/pgalloc.h>
5 #include <asm/pgtable.h>
7 #include <asm/fixmap.h>
10 #define PGALLOC_GFP (GFP_KERNEL_ACCOUNT | __GFP_NOTRACK | __GFP_ZERO)
13 #define PGALLOC_USER_GFP __GFP_HIGHMEM
15 #define PGALLOC_USER_GFP 0
18 gfp_t __userpte_alloc_gfp = PGALLOC_GFP | PGALLOC_USER_GFP;
20 pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
22 return (pte_t *)__get_free_page(PGALLOC_GFP & ~__GFP_ACCOUNT);
25 pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
29 pte = alloc_pages(__userpte_alloc_gfp, 0);
32 if (!pgtable_page_ctor(pte)) {
39 static int __init setup_userpte(char *arg)
45 * "userpte=nohigh" disables allocation of user pagetables in
48 if (strcmp(arg, "nohigh") == 0)
49 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
54 early_param("userpte", setup_userpte);
56 void ___pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
58 pgtable_page_dtor(pte);
59 paravirt_release_pte(page_to_pfn(pte));
60 tlb_remove_table(tlb, pte);
63 #if CONFIG_PGTABLE_LEVELS > 2
64 void ___pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
66 struct page *page = virt_to_page(pmd);
67 paravirt_release_pmd(__pa(pmd) >> PAGE_SHIFT);
69 * NOTE! For PAE, any changes to the top page-directory-pointer-table
70 * entries need a full cr3 reload to flush.
73 tlb->need_flush_all = 1;
75 pgtable_pmd_page_dtor(page);
76 tlb_remove_table(tlb, page);
79 #if CONFIG_PGTABLE_LEVELS > 3
80 void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
82 paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
83 tlb_remove_table(tlb, virt_to_page(pud));
86 #if CONFIG_PGTABLE_LEVELS > 4
87 void ___p4d_free_tlb(struct mmu_gather *tlb, p4d_t *p4d)
89 paravirt_release_p4d(__pa(p4d) >> PAGE_SHIFT);
90 tlb_remove_table(tlb, virt_to_page(p4d));
92 #endif /* CONFIG_PGTABLE_LEVELS > 4 */
93 #endif /* CONFIG_PGTABLE_LEVELS > 3 */
94 #endif /* CONFIG_PGTABLE_LEVELS > 2 */
96 static inline void pgd_list_add(pgd_t *pgd)
98 struct page *page = virt_to_page(pgd);
100 list_add(&page->lru, &pgd_list);
103 static inline void pgd_list_del(pgd_t *pgd)
105 struct page *page = virt_to_page(pgd);
107 list_del(&page->lru);
110 #define UNSHARED_PTRS_PER_PGD \
111 (SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
114 static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
116 BUILD_BUG_ON(sizeof(virt_to_page(pgd)->index) < sizeof(mm));
117 virt_to_page(pgd)->index = (pgoff_t)mm;
120 struct mm_struct *pgd_page_get_mm(struct page *page)
122 return (struct mm_struct *)page->index;
125 static void pgd_ctor(struct mm_struct *mm, pgd_t *pgd)
127 /* If the pgd points to a shared pagetable level (either the
128 ptes in non-PAE, or shared PMD in PAE), then just copy the
129 references from swapper_pg_dir. */
130 if (CONFIG_PGTABLE_LEVELS == 2 ||
131 (CONFIG_PGTABLE_LEVELS == 3 && SHARED_KERNEL_PMD) ||
132 CONFIG_PGTABLE_LEVELS >= 4) {
133 clone_pgd_range(pgd + KERNEL_PGD_BOUNDARY,
134 swapper_pg_dir + KERNEL_PGD_BOUNDARY,
138 /* list required to sync kernel mapping updates */
139 if (!SHARED_KERNEL_PMD) {
145 static void pgd_dtor(pgd_t *pgd)
147 if (SHARED_KERNEL_PMD)
150 spin_lock(&pgd_lock);
152 spin_unlock(&pgd_lock);
156 * List of all pgd's needed for non-PAE so it can invalidate entries
157 * in both cached and uncached pgd's; not needed for PAE since the
158 * kernel pmd is shared. If PAE were not to share the pmd a similar
159 * tactic would be needed. This is essentially codepath-based locking
160 * against pageattr.c; it is the unique case in which a valid change
161 * of kernel pagetables can't be lazily synchronized by vmalloc faults.
162 * vmalloc faults work because attached pagetables are never freed.
166 #ifdef CONFIG_X86_PAE
168 * In PAE mode, we need to do a cr3 reload (=tlb flush) when
169 * updating the top-level pagetable entries to guarantee the
170 * processor notices the update. Since this is expensive, and
171 * all 4 top-level entries are used almost immediately in a
172 * new process's life, we just pre-populate them here.
174 * Also, if we're in a paravirt environment where the kernel pmd is
175 * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
176 * and initialize the kernel pmds here.
178 #define PREALLOCATED_PMDS UNSHARED_PTRS_PER_PGD
180 void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
182 paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
184 /* Note: almost everything apart from _PAGE_PRESENT is
185 reserved at the pmd (PDPT) level. */
186 set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
189 * According to Intel App note "TLBs, Paging-Structure Caches,
190 * and Their Invalidation", April 2007, document 317080-001,
191 * section 8.1: in PAE mode we explicitly have to flush the
192 * TLB via cr3 if the top-level pgd is changed...
196 #else /* !CONFIG_X86_PAE */
198 /* No need to prepopulate any pagetable entries in non-PAE modes. */
199 #define PREALLOCATED_PMDS 0
201 #endif /* CONFIG_X86_PAE */
203 static void free_pmds(struct mm_struct *mm, pmd_t *pmds[])
207 for(i = 0; i < PREALLOCATED_PMDS; i++)
209 pgtable_pmd_page_dtor(virt_to_page(pmds[i]));
210 free_page((unsigned long)pmds[i]);
215 static int preallocate_pmds(struct mm_struct *mm, pmd_t *pmds[])
219 gfp_t gfp = PGALLOC_GFP;
222 gfp &= ~__GFP_ACCOUNT;
224 for(i = 0; i < PREALLOCATED_PMDS; i++) {
225 pmd_t *pmd = (pmd_t *)__get_free_page(gfp);
228 if (pmd && !pgtable_pmd_page_ctor(virt_to_page(pmd))) {
229 free_page((unsigned long)pmd);
247 * Mop up any pmd pages which may still be attached to the pgd.
248 * Normally they will be freed by munmap/exit_mmap, but any pmd we
249 * preallocate which never got a corresponding vma will need to be
252 static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
256 for(i = 0; i < PREALLOCATED_PMDS; i++) {
259 if (pgd_val(pgd) != 0) {
260 pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
262 pgdp[i] = native_make_pgd(0);
264 paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
271 static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
277 if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
280 p4d = p4d_offset(pgd, 0);
281 pud = pud_offset(p4d, 0);
283 for (i = 0; i < PREALLOCATED_PMDS; i++, pud++) {
284 pmd_t *pmd = pmds[i];
286 if (i >= KERNEL_PGD_BOUNDARY)
287 memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
288 sizeof(pmd_t) * PTRS_PER_PMD);
290 pud_populate(mm, pud, pmd);
295 * Xen paravirt assumes pgd table should be in one page. 64 bit kernel also
296 * assumes that pgd should be in one page.
298 * But kernel with PAE paging that is not running as a Xen domain
299 * only needs to allocate 32 bytes for pgd instead of one page.
301 #ifdef CONFIG_X86_PAE
303 #include <linux/slab.h>
305 #define PGD_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
308 static struct kmem_cache *pgd_cache;
310 static int __init pgd_cache_init(void)
313 * When PAE kernel is running as a Xen domain, it does not use
314 * shared kernel pmd. And this requires a whole page for pgd.
316 if (!SHARED_KERNEL_PMD)
320 * when PAE kernel is not running as a Xen domain, it uses
321 * shared kernel pmd. Shared kernel pmd does not require a whole
322 * page for pgd. We are able to just allocate a 32-byte for pgd.
323 * During boot time, we create a 32-byte slab for pgd table allocation.
325 pgd_cache = kmem_cache_create("pgd_cache", PGD_SIZE, PGD_ALIGN,
332 core_initcall(pgd_cache_init);
334 static inline pgd_t *_pgd_alloc(void)
337 * If no SHARED_KERNEL_PMD, PAE kernel is running as a Xen domain.
338 * We allocate one page for pgd.
340 if (!SHARED_KERNEL_PMD)
341 return (pgd_t *)__get_free_page(PGALLOC_GFP);
344 * Now PAE kernel is not running as a Xen domain. We can allocate
345 * a 32-byte slab for pgd to save memory space.
347 return kmem_cache_alloc(pgd_cache, PGALLOC_GFP);
350 static inline void _pgd_free(pgd_t *pgd)
352 if (!SHARED_KERNEL_PMD)
353 free_page((unsigned long)pgd);
355 kmem_cache_free(pgd_cache, pgd);
358 static inline pgd_t *_pgd_alloc(void)
360 return (pgd_t *)__get_free_page(PGALLOC_GFP);
363 static inline void _pgd_free(pgd_t *pgd)
365 free_page((unsigned long)pgd);
367 #endif /* CONFIG_X86_PAE */
369 pgd_t *pgd_alloc(struct mm_struct *mm)
372 pmd_t *pmds[PREALLOCATED_PMDS];
381 if (preallocate_pmds(mm, pmds) != 0)
384 if (paravirt_pgd_alloc(mm) != 0)
388 * Make sure that pre-populating the pmds is atomic with
389 * respect to anything walking the pgd_list, so that they
390 * never see a partially populated pgd.
392 spin_lock(&pgd_lock);
395 pgd_prepopulate_pmd(mm, pgd, pmds);
397 spin_unlock(&pgd_lock);
409 void pgd_free(struct mm_struct *mm, pgd_t *pgd)
411 pgd_mop_up_pmds(mm, pgd);
413 paravirt_pgd_free(mm, pgd);
418 * Used to set accessed or dirty bits in the page table entries
419 * on other architectures. On x86, the accessed and dirty bits
420 * are tracked by hardware. However, do_wp_page calls this function
421 * to also make the pte writeable at the same time the dirty bit is
422 * set. In that case we do actually need to write the PTE.
424 int ptep_set_access_flags(struct vm_area_struct *vma,
425 unsigned long address, pte_t *ptep,
426 pte_t entry, int dirty)
428 int changed = !pte_same(*ptep, entry);
430 if (changed && dirty)
436 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
437 int pmdp_set_access_flags(struct vm_area_struct *vma,
438 unsigned long address, pmd_t *pmdp,
439 pmd_t entry, int dirty)
441 int changed = !pmd_same(*pmdp, entry);
443 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
445 if (changed && dirty) {
448 * We had a write-protection fault here and changed the pmd
449 * to to more permissive. No need to flush the TLB for that,
450 * #PF is architecturally guaranteed to do that and in the
451 * worst-case we'll generate a spurious fault.
458 int pudp_set_access_flags(struct vm_area_struct *vma, unsigned long address,
459 pud_t *pudp, pud_t entry, int dirty)
461 int changed = !pud_same(*pudp, entry);
463 VM_BUG_ON(address & ~HPAGE_PUD_MASK);
465 if (changed && dirty) {
468 * We had a write-protection fault here and changed the pud
469 * to to more permissive. No need to flush the TLB for that,
470 * #PF is architecturally guaranteed to do that and in the
471 * worst-case we'll generate a spurious fault.
479 int ptep_test_and_clear_young(struct vm_area_struct *vma,
480 unsigned long addr, pte_t *ptep)
484 if (pte_young(*ptep))
485 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
486 (unsigned long *) &ptep->pte);
491 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
492 int pmdp_test_and_clear_young(struct vm_area_struct *vma,
493 unsigned long addr, pmd_t *pmdp)
497 if (pmd_young(*pmdp))
498 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
499 (unsigned long *)pmdp);
503 int pudp_test_and_clear_young(struct vm_area_struct *vma,
504 unsigned long addr, pud_t *pudp)
508 if (pud_young(*pudp))
509 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
510 (unsigned long *)pudp);
516 int ptep_clear_flush_young(struct vm_area_struct *vma,
517 unsigned long address, pte_t *ptep)
520 * On x86 CPUs, clearing the accessed bit without a TLB flush
521 * doesn't cause data corruption. [ It could cause incorrect
522 * page aging and the (mistaken) reclaim of hot pages, but the
523 * chance of that should be relatively low. ]
525 * So as a performance optimization don't flush the TLB when
526 * clearing the accessed bit, it will eventually be flushed by
527 * a context switch or a VM operation anyway. [ In the rare
528 * event of it not getting flushed for a long time the delay
529 * shouldn't really matter because there's no real memory
530 * pressure for swapout to react to. ]
532 return ptep_test_and_clear_young(vma, address, ptep);
535 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
536 int pmdp_clear_flush_young(struct vm_area_struct *vma,
537 unsigned long address, pmd_t *pmdp)
541 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
543 young = pmdp_test_and_clear_young(vma, address, pmdp);
545 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
552 * reserve_top_address - reserves a hole in the top of kernel address space
553 * @reserve - size of hole to reserve
555 * Can be used to relocate the fixmap area and poke a hole in the top
556 * of kernel address space to make room for a hypervisor.
558 void __init reserve_top_address(unsigned long reserve)
561 BUG_ON(fixmaps_set > 0);
562 __FIXADDR_TOP = round_down(-reserve, 1 << PMD_SHIFT) - PAGE_SIZE;
563 printk(KERN_INFO "Reserving virtual address space above 0x%08lx (rounded to 0x%08lx)\n",
564 -reserve, __FIXADDR_TOP + PAGE_SIZE);
570 void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
572 unsigned long address = __fix_to_virt(idx);
574 if (idx >= __end_of_fixed_addresses) {
578 set_pte_vaddr(address, pte);
582 void native_set_fixmap(enum fixed_addresses idx, phys_addr_t phys,
585 __native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
588 #ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
589 #ifdef CONFIG_X86_5LEVEL
591 * p4d_set_huge - setup kernel P4D mapping
593 * No 512GB pages yet -- always return 0
595 int p4d_set_huge(p4d_t *p4d, phys_addr_t addr, pgprot_t prot)
601 * p4d_clear_huge - clear kernel P4D mapping when it is set
603 * No 512GB pages yet -- always return 0
605 int p4d_clear_huge(p4d_t *p4d)
612 * pud_set_huge - setup kernel PUD mapping
614 * MTRRs can override PAT memory types with 4KiB granularity. Therefore, this
615 * function sets up a huge page only if any of the following conditions are met:
617 * - MTRRs are disabled, or
619 * - MTRRs are enabled and the range is completely covered by a single MTRR, or
621 * - MTRRs are enabled and the corresponding MTRR memory type is WB, which
622 * has no effect on the requested PAT memory type.
624 * Callers should try to decrease page size (1GB -> 2MB -> 4K) if the bigger
625 * page mapping attempt fails.
627 * Returns 1 on success and 0 on failure.
629 int pud_set_huge(pud_t *pud, phys_addr_t addr, pgprot_t prot)
633 mtrr = mtrr_type_lookup(addr, addr + PUD_SIZE, &uniform);
634 if ((mtrr != MTRR_TYPE_INVALID) && (!uniform) &&
635 (mtrr != MTRR_TYPE_WRBACK))
638 prot = pgprot_4k_2_large(prot);
640 set_pte((pte_t *)pud, pfn_pte(
641 (u64)addr >> PAGE_SHIFT,
642 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
648 * pmd_set_huge - setup kernel PMD mapping
650 * See text over pud_set_huge() above.
652 * Returns 1 on success and 0 on failure.
654 int pmd_set_huge(pmd_t *pmd, phys_addr_t addr, pgprot_t prot)
658 mtrr = mtrr_type_lookup(addr, addr + PMD_SIZE, &uniform);
659 if ((mtrr != MTRR_TYPE_INVALID) && (!uniform) &&
660 (mtrr != MTRR_TYPE_WRBACK)) {
661 pr_warn_once("%s: Cannot satisfy [mem %#010llx-%#010llx] with a huge-page mapping due to MTRR override.\n",
662 __func__, addr, addr + PMD_SIZE);
666 prot = pgprot_4k_2_large(prot);
668 set_pte((pte_t *)pmd, pfn_pte(
669 (u64)addr >> PAGE_SHIFT,
670 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
676 * pud_clear_huge - clear kernel PUD mapping when it is set
678 * Returns 1 on success and 0 on failure (no PUD map is found).
680 int pud_clear_huge(pud_t *pud)
682 if (pud_large(*pud)) {
691 * pmd_clear_huge - clear kernel PMD mapping when it is set
693 * Returns 1 on success and 0 on failure (no PMD map is found).
695 int pmd_clear_huge(pmd_t *pmd)
697 if (pmd_large(*pmd)) {
704 #endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */