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  *
  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 static void step_forward(struct page_vma_mapped_walk *pvmw, unsigned long size)
 119 {
 120         pvmw->address = (pvmw->address + size) & ~(size - 1);
 121         if (!pvmw->address)
 122                 pvmw->address = ULONG_MAX;
 123 }
 124
 125 /**
 126  * page_vma_mapped_walk - check if @pvmw->page is mapped in @pvmw->vma at
 127  * @pvmw->address
 128  * @pvmw: pointer to struct page_vma_mapped_walk. page, vma, address and flags
 129  * must be set. pmd, pte and ptl must be NULL.
 130  *
 131  * Returns true if the page is mapped in the vma. @pvmw->pmd and @pvmw->pte point
 132  * to relevant page table entries. @pvmw->ptl is locked. @pvmw->address is
 133  * adjusted if needed (for PTE-mapped THPs).
 134  *
 135  * If @pvmw->pmd is set but @pvmw->pte is not, you have found PMD-mapped page
 136  * (usually THP). For PTE-mapped THP, you should run page_vma_mapped_walk() in
 137  * a loop to find all PTEs that map the THP.
 138  *
 139  * For HugeTLB pages, @pvmw->pte is set to the relevant page table entry
 140  * regardless of which page table level the page is mapped at. @pvmw->pmd is
 141  * NULL.
 142  *
 143  * Retruns false if there are no more page table entries for the page in
 144  * the vma. @pvmw->ptl is unlocked and @pvmw->pte is unmapped.
 145  *
 146  * If you need to stop the walk before page_vma_mapped_walk() returned false,
 147  * use page_vma_mapped_walk_done(). It will do the housekeeping.
 148  */
 149 bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
 150 {
 151         struct mm_struct *mm = pvmw->vma->vm_mm;
 152         struct page *page = pvmw->page;
 153         unsigned long end;
 154         pgd_t *pgd;
 155         p4d_t *p4d;
 156         pud_t *pud;
 157         pmd_t pmde;
 158
 159         /* The only possible pmd mapping has been handled on last iteration */
 160         if (pvmw->pmd && !pvmw->pte)
 161                 return not_found(pvmw);
 162
 163         if (unlikely(PageHuge(page))) {
 164                 /* The only possible mapping was handled on last iteration */
 165                 if (pvmw->pte)
 166                         return not_found(pvmw);
 167
 168                 /* when pud is not present, pte will be NULL */
 169                 pvmw->pte = huge_pte_offset(mm, pvmw->address, page_size(page));
 170                 if (!pvmw->pte)
 171                         return false;
 172
 173                 pvmw->ptl = huge_pte_lockptr(page_hstate(page), mm, pvmw->pte);
 174                 spin_lock(pvmw->ptl);
 175                 if (!check_pte(pvmw))
 176                         return not_found(pvmw);
 177                 return true;
 178         }
 179
 180         /*
 181          * Seek to next pte only makes sense for THP.
 182          * But more important than that optimization, is to filter out
 183          * any PageKsm page: whose page->index misleads vma_address()
 184          * and vma_address_end() to disaster.
 185          */
 186         end = PageTransCompound(page) ?
 187                 vma_address_end(page, pvmw->vma) :
 188                 pvmw->address + PAGE_SIZE;
 189         if (pvmw->pte)
 190                 goto next_pte;
 191 restart:
 192         do {
 193                 pgd = pgd_offset(mm, pvmw->address);
 194                 if (!pgd_present(*pgd)) {
 195                         step_forward(pvmw, PGDIR_SIZE);
 196                         continue;
 197                 }
 198                 p4d = p4d_offset(pgd, pvmw->address);
 199                 if (!p4d_present(*p4d)) {
 200                         step_forward(pvmw, P4D_SIZE);
 201                         continue;
 202                 }
 203                 pud = pud_offset(p4d, pvmw->address);
 204                 if (!pud_present(*pud)) {
 205                         step_forward(pvmw, PUD_SIZE);
 206                         continue;
 207                 }
 208
 209                 pvmw->pmd = pmd_offset(pud, pvmw->address);
 210                 /*
 211                  * Make sure the pmd value isn't cached in a register by the
 212                  * compiler and used as a stale value after we've observed a
 213                  * subsequent update.
 214                  */
 215                 pmde = READ_ONCE(*pvmw->pmd);
 216
 217                 if (pmd_trans_huge(pmde) || is_pmd_migration_entry(pmde)) {
 218                         pvmw->ptl = pmd_lock(mm, pvmw->pmd);
 219                         pmde = *pvmw->pmd;
 220                         if (likely(pmd_trans_huge(pmde))) {
 221                                 if (pvmw->flags & PVMW_MIGRATION)
 222                                         return not_found(pvmw);
 223                                 if (pmd_page(pmde) != page)
 224                                         return not_found(pvmw);
 225                                 return true;
 226                         }
 227                         if (!pmd_present(pmde)) {
 228                                 swp_entry_t entry;
 229
 230                                 if (!thp_migration_supported() ||
 231                                     !(pvmw->flags & PVMW_MIGRATION))
 232                                         return not_found(pvmw);
 233                                 entry = pmd_to_swp_entry(pmde);
 234                                 if (!is_migration_entry(entry) ||
 235                                     migration_entry_to_page(entry) != page)
 236                                         return not_found(pvmw);
 237                                 return true;
 238                         }
 239                         /* THP pmd was split under us: handle on pte level */
 240                         spin_unlock(pvmw->ptl);
 241                         pvmw->ptl = NULL;
 242                 } else if (!pmd_present(pmde)) {
 243                         /*
 244                          * If PVMW_SYNC, take and drop THP pmd lock so that we
 245                          * cannot return prematurely, while zap_huge_pmd() has
 246                          * cleared *pmd but not decremented compound_mapcount().
 247                          */
 248                         if ((pvmw->flags & PVMW_SYNC) &&
 249                             PageTransCompound(page)) {
 250                                 spinlock_t *ptl = pmd_lock(mm, pvmw->pmd);
 251
 252                                 spin_unlock(ptl);
 253                         }
 254                         step_forward(pvmw, PMD_SIZE);
 255                         continue;
 256                 }
 257                 if (!map_pte(pvmw))
 258                         goto next_pte;
 259 this_pte:
 260                 if (check_pte(pvmw))
 261                         return true;
 262 next_pte:
 263                 do {
 264                         pvmw->address += PAGE_SIZE;
 265                         if (pvmw->address >= end)
 266                                 return not_found(pvmw);
 267                         /* Did we cross page table boundary? */
 268                         if ((pvmw->address & (PMD_SIZE - PAGE_SIZE)) == 0) {
 269                                 if (pvmw->ptl) {
 270                                         spin_unlock(pvmw->ptl);
 271                                         pvmw->ptl = NULL;
 272                                 }
 273                                 pte_unmap(pvmw->pte);
 274                                 pvmw->pte = NULL;
 275                                 goto restart;
 276                         }
 277                         pvmw->pte++;
 278                         if ((pvmw->flags & PVMW_SYNC) && !pvmw->ptl) {
 279                                 pvmw->ptl = pte_lockptr(mm, pvmw->pmd);
 280                                 spin_lock(pvmw->ptl);
 281                         }
 282                 } while (pte_none(*pvmw->pte));
 283
 284                 if (!pvmw->ptl) {
 285                         pvmw->ptl = pte_lockptr(mm, pvmw->pmd);
 286                         spin_lock(pvmw->ptl);
 287                 }
 288                 goto this_pte;
 289         } while (pvmw->address < end);
 290
 291         return false;
 292 }
 293
 294 /**
 295  * page_mapped_in_vma - check whether a page is really mapped in a VMA
 296  * @page: the page to test
 297  * @vma: the VMA to test
 298  *
 299  * Returns 1 if the page is mapped into the page tables of the VMA, 0
 300  * if the page is not mapped into the page tables of this VMA.  Only
 301  * valid for normal file or anonymous VMAs.
 302  */
 303 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
 304 {
 305         struct page_vma_mapped_walk pvmw = {
 306                 .page = page,
 307                 .vma = vma,
 308                 .flags = PVMW_SYNC,
 309         };
 310
 311         pvmw.address = vma_address(page, vma);
 312         if (pvmw.address == -EFAULT)
 313                 return 0;
 314         if (!page_vma_mapped_walk(&pvmw))
 315                 return 0;
 316         page_vma_mapped_walk_done(&pvmw);
 317         return 1;
 318 }