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 static void step_forward(struct page_vma_mapped_walk *pvmw, unsigned long size)
 115 {
 116         pvmw->address = (pvmw->address + size) & ~(size - 1);
 117         if (!pvmw->address)
 118                 pvmw->address = ULONG_MAX;
 119 }
 120
 121 /**
 122  * page_vma_mapped_walk - check if @pvmw->page is mapped in @pvmw->vma at
 123  * @pvmw->address
 124  * @pvmw: pointer to struct page_vma_mapped_walk. page, vma, address and flags
 125  * must be set. pmd, pte and ptl must be NULL.
 126  *
 127  * Returns true if the page is mapped in the vma. @pvmw->pmd and @pvmw->pte point
 128  * to relevant page table entries. @pvmw->ptl is locked. @pvmw->address is
 129  * adjusted if needed (for PTE-mapped THPs).
 130  *
 131  * If @pvmw->pmd is set but @pvmw->pte is not, you have found PMD-mapped page
 132  * (usually THP). For PTE-mapped THP, you should run page_vma_mapped_walk() in
 133  * a loop to find all PTEs that map the THP.
 134  *
 135  * For HugeTLB pages, @pvmw->pte is set to the relevant page table entry
 136  * regardless of which page table level the page is mapped at. @pvmw->pmd is
 137  * NULL.
 138  *
 139  * Retruns false if there are no more page table entries for the page in
 140  * the vma. @pvmw->ptl is unlocked and @pvmw->pte is unmapped.
 141  *
 142  * If you need to stop the walk before page_vma_mapped_walk() returned false,
 143  * use page_vma_mapped_walk_done(). It will do the housekeeping.
 144  */
 145 bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
 146 {
 147         struct mm_struct *mm = pvmw->vma->vm_mm;
 148         struct page *page = pvmw->page;
 149         unsigned long end;
 150         pgd_t *pgd;
 151         p4d_t *p4d;
 152         pud_t *pud;
 153         pmd_t pmde;
 154
 155         /* The only possible pmd mapping has been handled on last iteration */
 156         if (pvmw->pmd && !pvmw->pte)
 157                 return not_found(pvmw);
 158
 159         if (unlikely(PageHuge(page))) {
 160                 /* The only possible mapping was handled on last iteration */
 161                 if (pvmw->pte)
 162                         return not_found(pvmw);
 163
 164                 /* when pud is not present, pte will be NULL */
 165                 pvmw->pte = huge_pte_offset(mm, pvmw->address,
 166                                             PAGE_SIZE << compound_order(page));
 167                 if (!pvmw->pte)
 168                         return false;
 169
 170                 pvmw->ptl = huge_pte_lockptr(page_hstate(page), mm, pvmw->pte);
 171                 spin_lock(pvmw->ptl);
 172                 if (!check_pte(pvmw))
 173                         return not_found(pvmw);
 174                 return true;
 175         }
 176
 177         /*
 178          * Seek to next pte only makes sense for THP.
 179          * But more important than that optimization, is to filter out
 180          * any PageKsm page: whose page->index misleads vma_address()
 181          * and vma_address_end() to disaster.
 182          */
 183         end = PageTransCompound(page) ?
 184                 vma_address_end(page, pvmw->vma) :
 185                 pvmw->address + PAGE_SIZE;
 186         if (pvmw->pte)
 187                 goto next_pte;
 188 restart:
 189         do {
 190                 pgd = pgd_offset(mm, pvmw->address);
 191                 if (!pgd_present(*pgd)) {
 192                         step_forward(pvmw, PGDIR_SIZE);
 193                         continue;
 194                 }
 195                 p4d = p4d_offset(pgd, pvmw->address);
 196                 if (!p4d_present(*p4d)) {
 197                         step_forward(pvmw, P4D_SIZE);
 198                         continue;
 199                 }
 200                 pud = pud_offset(p4d, pvmw->address);
 201                 if (!pud_present(*pud)) {
 202                         step_forward(pvmw, PUD_SIZE);
 203                         continue;
 204                 }
 205
 206                 pvmw->pmd = pmd_offset(pud, pvmw->address);
 207                 /*
 208                  * Make sure the pmd value isn't cached in a register by the
 209                  * compiler and used as a stale value after we've observed a
 210                  * subsequent update.
 211                  */
 212                 pmde = READ_ONCE(*pvmw->pmd);
 213
 214                 if (pmd_trans_huge(pmde) || is_pmd_migration_entry(pmde)) {
 215                         pvmw->ptl = pmd_lock(mm, pvmw->pmd);
 216                         pmde = *pvmw->pmd;
 217                         if (likely(pmd_trans_huge(pmde))) {
 218                                 if (pvmw->flags & PVMW_MIGRATION)
 219                                         return not_found(pvmw);
 220                                 if (pmd_page(pmde) != page)
 221                                         return not_found(pvmw);
 222                                 return true;
 223                         }
 224                         if (!pmd_present(pmde)) {
 225                                 swp_entry_t entry;
 226
 227                                 if (!thp_migration_supported() ||
 228                                     !(pvmw->flags & PVMW_MIGRATION))
 229                                         return not_found(pvmw);
 230                                 entry = pmd_to_swp_entry(pmde);
 231                                 if (!is_migration_entry(entry) ||
 232                                     migration_entry_to_page(entry) != page)
 233                                         return not_found(pvmw);
 234                                 return true;
 235                         }
 236                         /* THP pmd was split under us: handle on pte level */
 237                         spin_unlock(pvmw->ptl);
 238                         pvmw->ptl = NULL;
 239                 } else if (!pmd_present(pmde)) {
 240                         /*
 241                          * If PVMW_SYNC, take and drop THP pmd lock so that we
 242                          * cannot return prematurely, while zap_huge_pmd() has
 243                          * cleared *pmd but not decremented compound_mapcount().
 244                          */
 245                         if ((pvmw->flags & PVMW_SYNC) &&
 246                             PageTransCompound(page)) {
 247                                 spinlock_t *ptl = pmd_lock(mm, pvmw->pmd);
 248
 249                                 spin_unlock(ptl);
 250                         }
 251                         step_forward(pvmw, PMD_SIZE);
 252                         continue;
 253                 }
 254                 if (!map_pte(pvmw))
 255                         goto next_pte;
 256 this_pte:
 257                 if (check_pte(pvmw))
 258                         return true;
 259 next_pte:
 260                 do {
 261                         pvmw->address += PAGE_SIZE;
 262                         if (pvmw->address >= end)
 263                                 return not_found(pvmw);
 264                         /* Did we cross page table boundary? */
 265                         if ((pvmw->address & (PMD_SIZE - PAGE_SIZE)) == 0) {
 266                                 if (pvmw->ptl) {
 267                                         spin_unlock(pvmw->ptl);
 268                                         pvmw->ptl = NULL;
 269                                 }
 270                                 pte_unmap(pvmw->pte);
 271                                 pvmw->pte = NULL;
 272                                 goto restart;
 273                         }
 274                         pvmw->pte++;
 275                         if ((pvmw->flags & PVMW_SYNC) && !pvmw->ptl) {
 276                                 pvmw->ptl = pte_lockptr(mm, pvmw->pmd);
 277                                 spin_lock(pvmw->ptl);
 278                         }
 279                 } while (pte_none(*pvmw->pte));
 280
 281                 if (!pvmw->ptl) {
 282                         pvmw->ptl = pte_lockptr(mm, pvmw->pmd);
 283                         spin_lock(pvmw->ptl);
 284                 }
 285                 goto this_pte;
 286         } while (pvmw->address < end);
 287
 288         return false;
 289 }
 290
 291 /**
 292  * page_mapped_in_vma - check whether a page is really mapped in a VMA
 293  * @page: the page to test
 294  * @vma: the VMA to test
 295  *
 296  * Returns 1 if the page is mapped into the page tables of the VMA, 0
 297  * if the page is not mapped into the page tables of this VMA.  Only
 298  * valid for normal file or anonymous VMAs.
 299  */
 300 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
 301 {
 302         struct page_vma_mapped_walk pvmw = {
 303                 .page = page,
 304                 .vma = vma,
 305                 .flags = PVMW_SYNC,
 306         };
 307
 308         pvmw.address = vma_address(page, vma);
 309         if (pvmw.address == -EFAULT)
 310                 return 0;
 311         if (!page_vma_mapped_walk(&pvmw))
 312                 return 0;
 313         page_vma_mapped_walk_done(&pvmw);
 314         return 1;
 315 }