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