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                                     !is_device_exclusive_entry(entry))
  46                                         return false;
  47                         } else if (!pte_present(*pvmw->pte))
  48                                 return false;
  49                 }
  50         }
  51         pvmw->ptl = pte_lockptr(pvmw->vma->vm_mm, pvmw->pmd);
  52         spin_lock(pvmw->ptl);
  53         return true;
  54 }
  55
  56 static inline bool pfn_is_match(struct page *page, unsigned long pfn)
  57 {
  58         unsigned long page_pfn = page_to_pfn(page);
  59
  60         /* normal page and hugetlbfs page */
  61         if (!PageTransCompound(page) || PageHuge(page))
  62                 return page_pfn == pfn;
  63
  64         /* THP can be referenced by any subpage */
  65         return pfn >= page_pfn && pfn - page_pfn < thp_nr_pages(page);
  66 }
  67
  68 /**
  69  * check_pte - check if @pvmw->page is mapped at the @pvmw->pte
  70  * @pvmw: page_vma_mapped_walk struct, includes a pair pte and page for checking
  71  *
  72  * page_vma_mapped_walk() found a place where @pvmw->page is *potentially*
  73  * mapped. check_pte() has to validate this.
  74  *
  75  * pvmw->pte may point to empty PTE, swap PTE or PTE pointing to
  76  * arbitrary page.
  77  *
  78  * If PVMW_MIGRATION flag is set, returns true if @pvmw->pte contains migration
  79  * entry that points to @pvmw->page or any subpage in case of THP.
  80  *
  81  * If PVMW_MIGRATION flag is not set, returns true if pvmw->pte points to
  82  * pvmw->page or any subpage in case of THP.
  83  *
  84  * Otherwise, return false.
  85  *
  86  */
  87 static bool check_pte(struct page_vma_mapped_walk *pvmw)
  88 {
  89         unsigned long pfn;
  90
  91         if (pvmw->flags & PVMW_MIGRATION) {
  92                 swp_entry_t entry;
  93                 if (!is_swap_pte(*pvmw->pte))
  94                         return false;
  95                 entry = pte_to_swp_entry(*pvmw->pte);
  96
  97                 if (!is_migration_entry(entry) &&
  98                     !is_device_exclusive_entry(entry))
  99                         return false;
 100
 101                 pfn = swp_offset(entry);
 102         } else if (is_swap_pte(*pvmw->pte)) {
 103                 swp_entry_t entry;
 104
 105                 /* Handle un-addressable ZONE_DEVICE memory */
 106                 entry = pte_to_swp_entry(*pvmw->pte);
 107                 if (!is_device_private_entry(entry) &&
 108                     !is_device_exclusive_entry(entry))
 109                         return false;
 110
 111                 pfn = swp_offset(entry);
 112         } else {
 113                 if (!pte_present(*pvmw->pte))
 114                         return false;
 115
 116                 pfn = pte_pfn(*pvmw->pte);
 117         }
 118
 119         return pfn_is_match(pvmw->page, pfn);
 120 }
 121
 122 static void step_forward(struct page_vma_mapped_walk *pvmw, unsigned long size)
 123 {
 124         pvmw->address = (pvmw->address + size) & ~(size - 1);
 125         if (!pvmw->address)
 126                 pvmw->address = ULONG_MAX;
 127 }
 128
 129 /**
 130  * page_vma_mapped_walk - check if @pvmw->page is mapped in @pvmw->vma at
 131  * @pvmw->address
 132  * @pvmw: pointer to struct page_vma_mapped_walk. page, vma, address and flags
 133  * must be set. pmd, pte and ptl must be NULL.
 134  *
 135  * Returns true if the page is mapped in the vma. @pvmw->pmd and @pvmw->pte point
 136  * to relevant page table entries. @pvmw->ptl is locked. @pvmw->address is
 137  * adjusted if needed (for PTE-mapped THPs).
 138  *
 139  * If @pvmw->pmd is set but @pvmw->pte is not, you have found PMD-mapped page
 140  * (usually THP). For PTE-mapped THP, you should run page_vma_mapped_walk() in
 141  * a loop to find all PTEs that map the THP.
 142  *
 143  * For HugeTLB pages, @pvmw->pte is set to the relevant page table entry
 144  * regardless of which page table level the page is mapped at. @pvmw->pmd is
 145  * NULL.
 146  *
 147  * Returns false if there are no more page table entries for the page in
 148  * the vma. @pvmw->ptl is unlocked and @pvmw->pte is unmapped.
 149  *
 150  * If you need to stop the walk before page_vma_mapped_walk() returned false,
 151  * use page_vma_mapped_walk_done(). It will do the housekeeping.
 152  */
 153 bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
 154 {
 155         struct mm_struct *mm = pvmw->vma->vm_mm;
 156         struct page *page = pvmw->page;
 157         unsigned long end;
 158         pgd_t *pgd;
 159         p4d_t *p4d;
 160         pud_t *pud;
 161         pmd_t pmde;
 162
 163         /* The only possible pmd mapping has been handled on last iteration */
 164         if (pvmw->pmd && !pvmw->pte)
 165                 return not_found(pvmw);
 166
 167         if (unlikely(PageHuge(page))) {
 168                 /* The only possible mapping was handled on last iteration */
 169                 if (pvmw->pte)
 170                         return not_found(pvmw);
 171
 172                 /* when pud is not present, pte will be NULL */
 173                 pvmw->pte = huge_pte_offset(mm, pvmw->address, page_size(page));
 174                 if (!pvmw->pte)
 175                         return false;
 176
 177                 pvmw->ptl = huge_pte_lockptr(page_hstate(page), mm, pvmw->pte);
 178                 spin_lock(pvmw->ptl);
 179                 if (!check_pte(pvmw))
 180                         return not_found(pvmw);
 181                 return true;
 182         }
 183
 184         /*
 185          * Seek to next pte only makes sense for THP.
 186          * But more important than that optimization, is to filter out
 187          * any PageKsm page: whose page->index misleads vma_address()
 188          * and vma_address_end() to disaster.
 189          */
 190         end = PageTransCompound(page) ?
 191                 vma_address_end(page, pvmw->vma) :
 192                 pvmw->address + PAGE_SIZE;
 193         if (pvmw->pte)
 194                 goto next_pte;
 195 restart:
 196         do {
 197                 pgd = pgd_offset(mm, pvmw->address);
 198                 if (!pgd_present(*pgd)) {
 199                         step_forward(pvmw, PGDIR_SIZE);
 200                         continue;
 201                 }
 202                 p4d = p4d_offset(pgd, pvmw->address);
 203                 if (!p4d_present(*p4d)) {
 204                         step_forward(pvmw, P4D_SIZE);
 205                         continue;
 206                 }
 207                 pud = pud_offset(p4d, pvmw->address);
 208                 if (!pud_present(*pud)) {
 209                         step_forward(pvmw, PUD_SIZE);
 210                         continue;
 211                 }
 212
 213                 pvmw->pmd = pmd_offset(pud, pvmw->address);
 214                 /*
 215                  * Make sure the pmd value isn't cached in a register by the
 216                  * compiler and used as a stale value after we've observed a
 217                  * subsequent update.
 218                  */
 219                 pmde = READ_ONCE(*pvmw->pmd);
 220
 221                 if (pmd_trans_huge(pmde) || is_pmd_migration_entry(pmde)) {
 222                         pvmw->ptl = pmd_lock(mm, pvmw->pmd);
 223                         pmde = *pvmw->pmd;
 224                         if (likely(pmd_trans_huge(pmde))) {
 225                                 if (pvmw->flags & PVMW_MIGRATION)
 226                                         return not_found(pvmw);
 227                                 if (pmd_page(pmde) != page)
 228                                         return not_found(pvmw);
 229                                 return true;
 230                         }
 231                         if (!pmd_present(pmde)) {
 232                                 swp_entry_t entry;
 233
 234                                 if (!thp_migration_supported() ||
 235                                     !(pvmw->flags & PVMW_MIGRATION))
 236                                         return not_found(pvmw);
 237                                 entry = pmd_to_swp_entry(pmde);
 238                                 if (!is_migration_entry(entry) ||
 239                                     pfn_swap_entry_to_page(entry) != page)
 240                                         return not_found(pvmw);
 241                                 return true;
 242                         }
 243                         /* THP pmd was split under us: handle on pte level */
 244                         spin_unlock(pvmw->ptl);
 245                         pvmw->ptl = NULL;
 246                 } else if (!pmd_present(pmde)) {
 247                         /*
 248                          * If PVMW_SYNC, take and drop THP pmd lock so that we
 249                          * cannot return prematurely, while zap_huge_pmd() has
 250                          * cleared *pmd but not decremented compound_mapcount().
 251                          */
 252                         if ((pvmw->flags & PVMW_SYNC) &&
 253                             PageTransCompound(page)) {
 254                                 spinlock_t *ptl = pmd_lock(mm, pvmw->pmd);
 255
 256                                 spin_unlock(ptl);
 257                         }
 258                         step_forward(pvmw, PMD_SIZE);
 259                         continue;
 260                 }
 261                 if (!map_pte(pvmw))
 262                         goto next_pte;
 263 this_pte:
 264                 if (check_pte(pvmw))
 265                         return true;
 266 next_pte:
 267                 do {
 268                         pvmw->address += PAGE_SIZE;
 269                         if (pvmw->address >= end)
 270                                 return not_found(pvmw);
 271                         /* Did we cross page table boundary? */
 272                         if ((pvmw->address & (PMD_SIZE - PAGE_SIZE)) == 0) {
 273                                 if (pvmw->ptl) {
 274                                         spin_unlock(pvmw->ptl);
 275                                         pvmw->ptl = NULL;
 276                                 }
 277                                 pte_unmap(pvmw->pte);
 278                                 pvmw->pte = NULL;
 279                                 goto restart;
 280                         }
 281                         pvmw->pte++;
 282                         if ((pvmw->flags & PVMW_SYNC) && !pvmw->ptl) {
 283                                 pvmw->ptl = pte_lockptr(mm, pvmw->pmd);
 284                                 spin_lock(pvmw->ptl);
 285                         }
 286                 } while (pte_none(*pvmw->pte));
 287
 288                 if (!pvmw->ptl) {
 289                         pvmw->ptl = pte_lockptr(mm, pvmw->pmd);
 290                         spin_lock(pvmw->ptl);
 291                 }
 292                 goto this_pte;
 293         } while (pvmw->address < end);
 294
 295         return false;
 296 }
 297
 298 /**
 299  * page_mapped_in_vma - check whether a page is really mapped in a VMA
 300  * @page: the page to test
 301  * @vma: the VMA to test
 302  *
 303  * Returns 1 if the page is mapped into the page tables of the VMA, 0
 304  * if the page is not mapped into the page tables of this VMA.  Only
 305  * valid for normal file or anonymous VMAs.
 306  */
 307 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
 308 {
 309         struct page_vma_mapped_walk pvmw = {
 310                 .page = page,
 311                 .vma = vma,
 312                 .flags = PVMW_SYNC,
 313         };
 314
 315         pvmw.address = vma_address(page, vma);
 316         if (pvmw.address == -EFAULT)
 317                 return 0;
 318         if (!page_vma_mapped_walk(&pvmw))
 319                 return 0;
 320         page_vma_mapped_walk_done(&pvmw);
 321         return 1;
 322 }