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 < thp_nr_pages(page);
  65 }
  66
  67 /**
  68  * check_pte - check if @pvmw->page is mapped at the @pvmw->pte
  69  * @pvmw: page_vma_mapped_walk struct, includes a pair pte and page for checking
  70  *
  71  * page_vma_mapped_walk() found a place where @pvmw->page is *potentially*
  72  * mapped. check_pte() has to validate this.
  73  *
  74  * pvmw->pte may point to empty PTE, swap PTE or PTE pointing to
  75  * arbitrary page.
  76  *
  77  * If PVMW_MIGRATION flag is set, returns true if @pvmw->pte contains migration
  78  * entry that points to @pvmw->page or any subpage in case of THP.
  79  *
  80  * If PVMW_MIGRATION flag is not set, returns true if pvmw->pte points to
  81  * pvmw->page or any subpage in case of THP.
  82  *
  83  * Otherwise, return false.
  84  *
  85  */
  86 static bool check_pte(struct page_vma_mapped_walk *pvmw)
  87 {
  88         unsigned long pfn;
  89
  90         if (pvmw->flags & PVMW_MIGRATION) {
  91                 swp_entry_t entry;
  92                 if (!is_swap_pte(*pvmw->pte))
  93                         return false;
  94                 entry = pte_to_swp_entry(*pvmw->pte);
  95
  96                 if (!is_migration_entry(entry))
  97                         return false;
  98
  99                 pfn = migration_entry_to_pfn(entry);
 100         } else if (is_swap_pte(*pvmw->pte)) {
 101                 swp_entry_t entry;
 102
 103                 /* Handle un-addressable ZONE_DEVICE memory */
 104                 entry = pte_to_swp_entry(*pvmw->pte);
 105                 if (!is_device_private_entry(entry))
 106                         return false;
 107
 108                 pfn = device_private_entry_to_pfn(entry);
 109         } else {
 110                 if (!pte_present(*pvmw->pte))
 111                         return false;
 112
 113                 pfn = pte_pfn(*pvmw->pte);
 114         }
 115
 116         return pfn_is_match(pvmw->page, pfn);
 117 }
 118
 119 /**
 120  * page_vma_mapped_walk - check if @pvmw->page is mapped in @pvmw->vma at
 121  * @pvmw->address
 122  * @pvmw: pointer to struct page_vma_mapped_walk. page, vma, address and flags
 123  * must be set. pmd, pte and ptl must be NULL.
 124  *
 125  * Returns true if the page is mapped in the vma. @pvmw->pmd and @pvmw->pte point
 126  * to relevant page table entries. @pvmw->ptl is locked. @pvmw->address is
 127  * adjusted if needed (for PTE-mapped THPs).
 128  *
 129  * If @pvmw->pmd is set but @pvmw->pte is not, you have found PMD-mapped page
 130  * (usually THP). For PTE-mapped THP, you should run page_vma_mapped_walk() in
 131  * a loop to find all PTEs that map the THP.
 132  *
 133  * For HugeTLB pages, @pvmw->pte is set to the relevant page table entry
 134  * regardless of which page table level the page is mapped at. @pvmw->pmd is
 135  * NULL.
 136  *
 137  * Returns false if there are no more page table entries for the page in
 138  * the vma. @pvmw->ptl is unlocked and @pvmw->pte is unmapped.
 139  *
 140  * If you need to stop the walk before page_vma_mapped_walk() returned false,
 141  * use page_vma_mapped_walk_done(). It will do the housekeeping.
 142  */
 143 bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
 144 {
 145         struct mm_struct *mm = pvmw->vma->vm_mm;
 146         struct page *page = pvmw->page;
 147         unsigned long end;
 148         pgd_t *pgd;
 149         p4d_t *p4d;
 150         pud_t *pud;
 151         pmd_t pmde;
 152
 153         /* The only possible pmd mapping has been handled on last iteration */
 154         if (pvmw->pmd && !pvmw->pte)
 155                 return not_found(pvmw);
 156
 157         if (unlikely(PageHuge(page))) {
 158                 /* The only possible mapping was handled on last iteration */
 159                 if (pvmw->pte)
 160                         return not_found(pvmw);
 161
 162                 /* when pud is not present, pte will be NULL */
 163                 pvmw->pte = huge_pte_offset(mm, pvmw->address, page_size(page));
 164                 if (!pvmw->pte)
 165                         return false;
 166
 167                 pvmw->ptl = huge_pte_lockptr(page_hstate(page), mm, pvmw->pte);
 168                 spin_lock(pvmw->ptl);
 169                 if (!check_pte(pvmw))
 170                         return not_found(pvmw);
 171                 return true;
 172         }
 173
 174         if (pvmw->pte)
 175                 goto next_pte;
 176 restart:
 177         {
 178                 pgd = pgd_offset(mm, pvmw->address);
 179                 if (!pgd_present(*pgd))
 180                         return false;
 181                 p4d = p4d_offset(pgd, pvmw->address);
 182                 if (!p4d_present(*p4d))
 183                         return false;
 184                 pud = pud_offset(p4d, pvmw->address);
 185                 if (!pud_present(*pud))
 186                         return false;
 187
 188                 pvmw->pmd = pmd_offset(pud, pvmw->address);
 189                 /*
 190                  * Make sure the pmd value isn't cached in a register by the
 191                  * compiler and used as a stale value after we've observed a
 192                  * subsequent update.
 193                  */
 194                 pmde = READ_ONCE(*pvmw->pmd);
 195
 196                 if (pmd_trans_huge(pmde) || is_pmd_migration_entry(pmde)) {
 197                         pvmw->ptl = pmd_lock(mm, pvmw->pmd);
 198                         pmde = *pvmw->pmd;
 199                         if (likely(pmd_trans_huge(pmde))) {
 200                                 if (pvmw->flags & PVMW_MIGRATION)
 201                                         return not_found(pvmw);
 202                                 if (pmd_page(pmde) != page)
 203                                         return not_found(pvmw);
 204                                 return true;
 205                         }
 206                         if (!pmd_present(pmde)) {
 207                                 swp_entry_t entry;
 208
 209                                 if (!thp_migration_supported() ||
 210                                     !(pvmw->flags & PVMW_MIGRATION))
 211                                         return not_found(pvmw);
 212                                 entry = pmd_to_swp_entry(pmde);
 213                                 if (!is_migration_entry(entry) ||
 214                                     migration_entry_to_page(entry) != page)
 215                                         return not_found(pvmw);
 216                                 return true;
 217                         }
 218                         /* THP pmd was split under us: handle on pte level */
 219                         spin_unlock(pvmw->ptl);
 220                         pvmw->ptl = NULL;
 221                 } else if (!pmd_present(pmde)) {
 222                         /*
 223                          * If PVMW_SYNC, take and drop THP pmd lock so that we
 224                          * cannot return prematurely, while zap_huge_pmd() has
 225                          * cleared *pmd but not decremented compound_mapcount().
 226                          */
 227                         if ((pvmw->flags & PVMW_SYNC) &&
 228                             PageTransCompound(page)) {
 229                                 spinlock_t *ptl = pmd_lock(mm, pvmw->pmd);
 230
 231                                 spin_unlock(ptl);
 232                         }
 233                         return false;
 234                 }
 235                 if (!map_pte(pvmw))
 236                         goto next_pte;
 237 this_pte:
 238                 if (check_pte(pvmw))
 239                         return true;
 240 next_pte:
 241                 /* Seek to next pte only makes sense for THP */
 242                 if (!PageTransHuge(page))
 243                         return not_found(pvmw);
 244                 end = vma_address_end(page, pvmw->vma);
 245                 do {
 246                         pvmw->address += PAGE_SIZE;
 247                         if (pvmw->address >= end)
 248                                 return not_found(pvmw);
 249                         /* Did we cross page table boundary? */
 250                         if ((pvmw->address & (PMD_SIZE - PAGE_SIZE)) == 0) {
 251                                 if (pvmw->ptl) {
 252                                         spin_unlock(pvmw->ptl);
 253                                         pvmw->ptl = NULL;
 254                                 }
 255                                 pte_unmap(pvmw->pte);
 256                                 pvmw->pte = NULL;
 257                                 goto restart;
 258                         }
 259                         pvmw->pte++;
 260                 } while (pte_none(*pvmw->pte));
 261
 262                 if (!pvmw->ptl) {
 263                         pvmw->ptl = pte_lockptr(mm, pvmw->pmd);
 264                         spin_lock(pvmw->ptl);
 265                 }
 266                 goto this_pte;
 267         }
 268 }
 269
 270 /**
 271  * page_mapped_in_vma - check whether a page is really mapped in a VMA
 272  * @page: the page to test
 273  * @vma: the VMA to test
 274  *
 275  * Returns 1 if the page is mapped into the page tables of the VMA, 0
 276  * if the page is not mapped into the page tables of this VMA.  Only
 277  * valid for normal file or anonymous VMAs.
 278  */
 279 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
 280 {
 281         struct page_vma_mapped_walk pvmw = {
 282                 .page = page,
 283                 .vma = vma,
 284                 .flags = PVMW_SYNC,
 285         };
 286
 287         pvmw.address = vma_address(page, vma);
 288         if (pvmw.address == -EFAULT)
 289                 return 0;
 290         if (!page_vma_mapped_walk(&pvmw))
 291                 return 0;
 292         page_vma_mapped_walk_done(&pvmw);
 293         return 1;
 294 }