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