mm/page_vma_mapped.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/mm.h>
   3 #include <linux/rmap.h>
   4 #include <linux/hugetlb.h>
   5 #include <linux/swap.h>
   6 #include <linux/swapops.h>
   7
   8 #include "internal.h"
   9
  10 static inline bool not_found(struct page_vma_mapped_walk *pvmw)
  11 {
  12         page_vma_mapped_walk_done(pvmw);
  13         return false;
  14 }
  15
  16 static bool map_pte(struct page_vma_mapped_walk *pvmw)
  17 {
  18         pvmw->pte = pte_offset_map(pvmw->pmd, pvmw->address);
  19         if (!(pvmw->flags & PVMW_SYNC)) {
  20                 if (pvmw->flags & PVMW_MIGRATION) {
  21                         if (!is_swap_pte(*pvmw->pte))
  22                                 return false;
  23                 } else {
  24                         /*
  25                          * We get here when we are trying to unmap a private
  26                          * device page from the process address space. Such
  27                          * page is not CPU accessible and thus is mapped as
  28                          * a special swap entry, nonetheless it still does
  29                          * count as a valid regular mapping for the page (and
  30                          * is accounted as such in page maps count).
  31                          *
  32                          * So handle this special case as if it was a normal
  33                          * page mapping ie lock CPU page table and returns
  34                          * true.
  35                          *
  36                          * For more details on device private memory see HMM
  37                          * (include/linux/hmm.h or mm/hmm.c).
  38                          */
  39                         if (is_swap_pte(*pvmw->pte)) {
  40                                 swp_entry_t entry;
  41
  42                                 /* Handle un-addressable ZONE_DEVICE memory */
  43                                 entry = pte_to_swp_entry(*pvmw->pte);
  44                                 if (!is_device_private_entry(entry))
  45                                         return false;
  46                         } else if (!pte_present(*pvmw->pte))
  47                                 return false;
  48                 }
  49         }
  50         pvmw->ptl = pte_lockptr(pvmw->vma->vm_mm, pvmw->pmd);
  51         spin_lock(pvmw->ptl);
  52         return true;
  53 }
  54
  55 static inline bool pfn_is_match(struct page *page, unsigned long pfn)
  56 {
  57         unsigned long page_pfn = page_to_pfn(page);
  58
  59         /* normal page and hugetlbfs page */
  60         if (!PageTransCompound(page) || PageHuge(page))
  61                 return page_pfn == pfn;
  62
  63         /* THP can be referenced by any subpage */
  64         return pfn >= page_pfn && pfn - page_pfn < hpage_nr_pages(page);
  65 }
  66
  67 /**
  68  * check_pte - check if @pvmw->page is mapped at the @pvmw->pte
  69  *
  70  * page_vma_mapped_walk() found a place where @pvmw->page is *potentially*
  71  * mapped. check_pte() has to validate this.
  72  *
  73  * @pvmw->pte may point to empty PTE, swap PTE or PTE pointing to arbitrary
  74  * page.
  75  *
  76  * If PVMW_MIGRATION flag is set, returns true if @pvmw->pte contains migration
  77  * entry that points to @pvmw->page or any subpage in case of THP.
  78  *
  79  * If PVMW_MIGRATION flag is not set, returns true if @pvmw->pte points to
  80  * @pvmw->page or any subpage in case of THP.
  81  *
  82  * Otherwise, return false.
  83  *
  84  */
  85 static bool check_pte(struct page_vma_mapped_walk *pvmw)
  86 {
  87         unsigned long pfn;
  88
  89         if (pvmw->flags & PVMW_MIGRATION) {
  90                 swp_entry_t entry;
  91                 if (!is_swap_pte(*pvmw->pte))
  92                         return false;
  93                 entry = pte_to_swp_entry(*pvmw->pte);
  94
  95                 if (!is_migration_entry(entry))
  96                         return false;
  97
  98                 pfn = migration_entry_to_pfn(entry);
  99         } else if (is_swap_pte(*pvmw->pte)) {
 100                 swp_entry_t entry;
 101
 102                 /* Handle un-addressable ZONE_DEVICE memory */
 103                 entry = pte_to_swp_entry(*pvmw->pte);
 104                 if (!is_device_private_entry(entry))
 105                         return false;
 106
 107                 pfn = device_private_entry_to_pfn(entry);
 108         } else {
 109                 if (!pte_present(*pvmw->pte))
 110                         return false;
 111
 112                 pfn = pte_pfn(*pvmw->pte);
 113         }
 114
 115         return pfn_is_match(pvmw->page, pfn);
 116 }
 117
 118 /**
 119  * page_vma_mapped_walk - check if @pvmw->page is mapped in @pvmw->vma at
 120  * @pvmw->address
 121  * @pvmw: pointer to struct page_vma_mapped_walk. page, vma, address and flags
 122  * must be set. pmd, pte and ptl must be NULL.
 123  *
 124  * Returns true if the page is mapped in the vma. @pvmw->pmd and @pvmw->pte point
 125  * to relevant page table entries. @pvmw->ptl is locked. @pvmw->address is
 126  * adjusted if needed (for PTE-mapped THPs).
 127  *
 128  * If @pvmw->pmd is set but @pvmw->pte is not, you have found PMD-mapped page
 129  * (usually THP). For PTE-mapped THP, you should run page_vma_mapped_walk() in
 130  * a loop to find all PTEs that map the THP.
 131  *
 132  * For HugeTLB pages, @pvmw->pte is set to the relevant page table entry
 133  * regardless of which page table level the page is mapped at. @pvmw->pmd is
 134  * NULL.
 135  *
 136  * Retruns false if there are no more page table entries for the page in
 137  * the vma. @pvmw->ptl is unlocked and @pvmw->pte is unmapped.
 138  *
 139  * If you need to stop the walk before page_vma_mapped_walk() returned false,
 140  * use page_vma_mapped_walk_done(). It will do the housekeeping.
 141  */
 142 bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
 143 {
 144         struct mm_struct *mm = pvmw->vma->vm_mm;
 145         struct page *page = pvmw->page;
 146         pgd_t *pgd;
 147         p4d_t *p4d;
 148         pud_t *pud;
 149         pmd_t pmde;
 150
 151         /* The only possible pmd mapping has been handled on last iteration */
 152         if (pvmw->pmd && !pvmw->pte)
 153                 return not_found(pvmw);
 154
 155         if (pvmw->pte)
 156                 goto next_pte;
 157
 158         if (unlikely(PageHuge(pvmw->page))) {
 159                 /* when pud is not present, pte will be NULL */
 160                 pvmw->pte = huge_pte_offset(mm, pvmw->address, page_size(page));
 161                 if (!pvmw->pte)
 162                         return false;
 163
 164                 pvmw->ptl = huge_pte_lockptr(page_hstate(page), mm, pvmw->pte);
 165                 spin_lock(pvmw->ptl);
 166                 if (!check_pte(pvmw))
 167                         return not_found(pvmw);
 168                 return true;
 169         }
 170 restart:
 171         pgd = pgd_offset(mm, pvmw->address);
 172         if (!pgd_present(*pgd))
 173                 return false;
 174         p4d = p4d_offset(pgd, pvmw->address);
 175         if (!p4d_present(*p4d))
 176                 return false;
 177         pud = pud_offset(p4d, pvmw->address);
 178         if (!pud_present(*pud))
 179                 return false;
 180         pvmw->pmd = pmd_offset(pud, pvmw->address);
 181         /*
 182          * Make sure the pmd value isn't cached in a register by the
 183          * compiler and used as a stale value after we've observed a
 184          * subsequent update.
 185          */
 186         pmde = READ_ONCE(*pvmw->pmd);
 187         if (pmd_trans_huge(pmde) || is_pmd_migration_entry(pmde)) {
 188                 pvmw->ptl = pmd_lock(mm, pvmw->pmd);
 189                 if (likely(pmd_trans_huge(*pvmw->pmd))) {
 190                         if (pvmw->flags & PVMW_MIGRATION)
 191                                 return not_found(pvmw);
 192                         if (pmd_page(*pvmw->pmd) != page)
 193                                 return not_found(pvmw);
 194                         return true;
 195                 } else if (!pmd_present(*pvmw->pmd)) {
 196                         if (thp_migration_supported()) {
 197                                 if (!(pvmw->flags & PVMW_MIGRATION))
 198                                         return not_found(pvmw);
 199                                 if (is_migration_entry(pmd_to_swp_entry(*pvmw->pmd))) {
 200                                         swp_entry_t entry = pmd_to_swp_entry(*pvmw->pmd);
 201
 202                                         if (migration_entry_to_page(entry) != page)
 203                                                 return not_found(pvmw);
 204                                         return true;
 205                                 }
 206                         }
 207                         return not_found(pvmw);
 208                 } else {
 209                         /* THP pmd was split under us: handle on pte level */
 210                         spin_unlock(pvmw->ptl);
 211                         pvmw->ptl = NULL;
 212                 }
 213         } else if (!pmd_present(pmde)) {
 214                 return false;
 215         }
 216         if (!map_pte(pvmw))
 217                 goto next_pte;
 218         while (1) {
 219                 if (check_pte(pvmw))
 220                         return true;
 221 next_pte:
 222                 /* Seek to next pte only makes sense for THP */
 223                 if (!PageTransHuge(pvmw->page) || PageHuge(pvmw->page))
 224                         return not_found(pvmw);
 225                 do {
 226                         pvmw->address += PAGE_SIZE;
 227                         if (pvmw->address >= pvmw->vma->vm_end ||
 228                             pvmw->address >=
 229                                         __vma_address(pvmw->page, pvmw->vma) +
 230                                         hpage_nr_pages(pvmw->page) * PAGE_SIZE)
 231                                 return not_found(pvmw);
 232                         /* Did we cross page table boundary? */
 233                         if (pvmw->address % PMD_SIZE == 0) {
 234                                 pte_unmap(pvmw->pte);
 235                                 if (pvmw->ptl) {
 236                                         spin_unlock(pvmw->ptl);
 237                                         pvmw->ptl = NULL;
 238                                 }
 239                                 goto restart;
 240                         } else {
 241                                 pvmw->pte++;
 242                         }
 243                 } while (pte_none(*pvmw->pte));
 244
 245                 if (!pvmw->ptl) {
 246                         pvmw->ptl = pte_lockptr(mm, pvmw->pmd);
 247                         spin_lock(pvmw->ptl);
 248                 }
 249         }
 250 }
 251
 252 /**
 253  * page_mapped_in_vma - check whether a page is really mapped in a VMA
 254  * @page: the page to test
 255  * @vma: the VMA to test
 256  *
 257  * Returns 1 if the page is mapped into the page tables of the VMA, 0
 258  * if the page is not mapped into the page tables of this VMA.  Only
 259  * valid for normal file or anonymous VMAs.
 260  */
 261 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
 262 {
 263         struct page_vma_mapped_walk pvmw = {
 264                 .page = page,
 265                 .vma = vma,
 266                 .flags = PVMW_SYNC,
 267         };
 268         unsigned long start, end;
 269
 270         start = __vma_address(page, vma);
 271         end = start + PAGE_SIZE * (hpage_nr_pages(page) - 1);
 272
 273         if (unlikely(end < vma->vm_start || start >= vma->vm_end))
 274                 return 0;
 275         pvmw.address = max(start, vma->vm_start);
 276         if (!page_vma_mapped_walk(&pvmw))
 277                 return 0;
 278         page_vma_mapped_walk_done(&pvmw);
 279         return 1;
 280 }