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 /**
  57  * check_pte - check if @pvmw->page is mapped at the @pvmw->pte
  58  * @pvmw: page_vma_mapped_walk struct, includes a pair pte and page for checking
  59  *
  60  * page_vma_mapped_walk() found a place where @pvmw->page is *potentially*
  61  * mapped. check_pte() has to validate this.
  62  *
  63  * pvmw->pte may point to empty PTE, swap PTE or PTE pointing to
  64  * arbitrary page.
  65  *
  66  * If PVMW_MIGRATION flag is set, returns true if @pvmw->pte contains migration
  67  * entry that points to @pvmw->page or any subpage in case of THP.
  68  *
  69  * If PVMW_MIGRATION flag is not set, returns true if pvmw->pte points to
  70  * pvmw->page or any subpage in case of THP.
  71  *
  72  * Otherwise, return false.
  73  *
  74  */
  75 static bool check_pte(struct page_vma_mapped_walk *pvmw)
  76 {
  77         unsigned long pfn;
  78
  79         if (pvmw->flags & PVMW_MIGRATION) {
  80                 swp_entry_t entry;
  81                 if (!is_swap_pte(*pvmw->pte))
  82                         return false;
  83                 entry = pte_to_swp_entry(*pvmw->pte);
  84
  85                 if (!is_migration_entry(entry) &&
  86                     !is_device_exclusive_entry(entry))
  87                         return false;
  88
  89                 pfn = swp_offset_pfn(entry);
  90         } else if (is_swap_pte(*pvmw->pte)) {
  91                 swp_entry_t entry;
  92
  93                 /* Handle un-addressable ZONE_DEVICE memory */
  94                 entry = pte_to_swp_entry(*pvmw->pte);
  95                 if (!is_device_private_entry(entry) &&
  96                     !is_device_exclusive_entry(entry))
  97                         return false;
  98
  99                 pfn = swp_offset_pfn(entry);
 100         } else {
 101                 if (!pte_present(*pvmw->pte))
 102                         return false;
 103
 104                 pfn = pte_pfn(*pvmw->pte);
 105         }
 106
 107         return (pfn - pvmw->pfn) < pvmw->nr_pages;
 108 }
 109
 110 /* Returns true if the two ranges overlap.  Careful to not overflow. */
 111 static bool check_pmd(unsigned long pfn, struct page_vma_mapped_walk *pvmw)
 112 {
 113         if ((pfn + HPAGE_PMD_NR - 1) < pvmw->pfn)
 114                 return false;
 115         if (pfn > pvmw->pfn + pvmw->nr_pages - 1)
 116                 return false;
 117         return true;
 118 }
 119
 120 static void step_forward(struct page_vma_mapped_walk *pvmw, unsigned long size)
 121 {
 122         pvmw->address = (pvmw->address + size) & ~(size - 1);
 123         if (!pvmw->address)
 124                 pvmw->address = ULONG_MAX;
 125 }
 126
 127 /**
 128  * page_vma_mapped_walk - check if @pvmw->pfn is mapped in @pvmw->vma at
 129  * @pvmw->address
 130  * @pvmw: pointer to struct page_vma_mapped_walk. page, vma, address and flags
 131  * must be set. pmd, pte and ptl must be NULL.
 132  *
 133  * Returns true if the page is mapped in the vma. @pvmw->pmd and @pvmw->pte point
 134  * to relevant page table entries. @pvmw->ptl is locked. @pvmw->address is
 135  * adjusted if needed (for PTE-mapped THPs).
 136  *
 137  * If @pvmw->pmd is set but @pvmw->pte is not, you have found PMD-mapped page
 138  * (usually THP). For PTE-mapped THP, you should run page_vma_mapped_walk() in
 139  * a loop to find all PTEs that map the THP.
 140  *
 141  * For HugeTLB pages, @pvmw->pte is set to the relevant page table entry
 142  * regardless of which page table level the page is mapped at. @pvmw->pmd is
 143  * NULL.
 144  *
 145  * Returns false if there are no more page table entries for the page in
 146  * the vma. @pvmw->ptl is unlocked and @pvmw->pte is unmapped.
 147  *
 148  * If you need to stop the walk before page_vma_mapped_walk() returned false,
 149  * use page_vma_mapped_walk_done(). It will do the housekeeping.
 150  */
 151 bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
 152 {
 153         struct vm_area_struct *vma = pvmw->vma;
 154         struct mm_struct *mm = vma->vm_mm;
 155         unsigned long end;
 156         pgd_t *pgd;
 157         p4d_t *p4d;
 158         pud_t *pud;
 159         pmd_t pmde;
 160
 161         /* The only possible pmd mapping has been handled on last iteration */
 162         if (pvmw->pmd && !pvmw->pte)
 163                 return not_found(pvmw);
 164
 165         if (unlikely(is_vm_hugetlb_page(vma))) {
 166                 struct hstate *hstate = hstate_vma(vma);
 167                 unsigned long size = huge_page_size(hstate);
 168                 /* The only possible mapping was handled on last iteration */
 169                 if (pvmw->pte)
 170                         return not_found(pvmw);
 171                 /*
 172                  * All callers that get here will already hold the
 173                  * i_mmap_rwsem.  Therefore, no additional locks need to be
 174                  * taken before calling hugetlb_walk().
 175                  */
 176                 pvmw->pte = hugetlb_walk(vma, pvmw->address, size);
 177                 if (!pvmw->pte)
 178                         return false;
 179
 180                 pvmw->ptl = huge_pte_lock(hstate, mm, pvmw->pte);
 181                 if (!check_pte(pvmw))
 182                         return not_found(pvmw);
 183                 return true;
 184         }
 185
 186         end = vma_address_end(pvmw);
 187         if (pvmw->pte)
 188                 goto next_pte;
 189 restart:
 190         do {
 191                 pgd = pgd_offset(mm, pvmw->address);
 192                 if (!pgd_present(*pgd)) {
 193                         step_forward(pvmw, PGDIR_SIZE);
 194                         continue;
 195                 }
 196                 p4d = p4d_offset(pgd, pvmw->address);
 197                 if (!p4d_present(*p4d)) {
 198                         step_forward(pvmw, P4D_SIZE);
 199                         continue;
 200                 }
 201                 pud = pud_offset(p4d, pvmw->address);
 202                 if (!pud_present(*pud)) {
 203                         step_forward(pvmw, PUD_SIZE);
 204                         continue;
 205                 }
 206
 207                 pvmw->pmd = pmd_offset(pud, pvmw->address);
 208                 /*
 209                  * Make sure the pmd value isn't cached in a register by the
 210                  * compiler and used as a stale value after we've observed a
 211                  * subsequent update.
 212                  */
 213                 pmde = READ_ONCE(*pvmw->pmd);
 214
 215                 if (pmd_trans_huge(pmde) || is_pmd_migration_entry(pmde) ||
 216                     (pmd_present(pmde) && pmd_devmap(pmde))) {
 217                         pvmw->ptl = pmd_lock(mm, pvmw->pmd);
 218                         pmde = *pvmw->pmd;
 219                         if (!pmd_present(pmde)) {
 220                                 swp_entry_t entry;
 221
 222                                 if (!thp_migration_supported() ||
 223                                     !(pvmw->flags & PVMW_MIGRATION))
 224                                         return not_found(pvmw);
 225                                 entry = pmd_to_swp_entry(pmde);
 226                                 if (!is_migration_entry(entry) ||
 227                                     !check_pmd(swp_offset_pfn(entry), pvmw))
 228                                         return not_found(pvmw);
 229                                 return true;
 230                         }
 231                         if (likely(pmd_trans_huge(pmde) || pmd_devmap(pmde))) {
 232                                 if (pvmw->flags & PVMW_MIGRATION)
 233                                         return not_found(pvmw);
 234                                 if (!check_pmd(pmd_pfn(pmde), pvmw))
 235                                         return not_found(pvmw);
 236                                 return true;
 237                         }
 238                         /* THP pmd was split under us: handle on pte level */
 239                         spin_unlock(pvmw->ptl);
 240                         pvmw->ptl = NULL;
 241                 } else if (!pmd_present(pmde)) {
 242                         /*
 243                          * If PVMW_SYNC, take and drop THP pmd lock so that we
 244                          * cannot return prematurely, while zap_huge_pmd() has
 245                          * cleared *pmd but not decremented compound_mapcount().
 246                          */
 247                         if ((pvmw->flags & PVMW_SYNC) &&
 248                             transhuge_vma_suitable(vma, pvmw->address) &&
 249                             (pvmw->nr_pages >= HPAGE_PMD_NR)) {
 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                 .pfn = page_to_pfn(page),
 307                 .nr_pages = 1,
 308                 .vma = vma,
 309                 .flags = PVMW_SYNC,
 310         };
 311
 312         pvmw.address = vma_address(page, vma);
 313         if (pvmw.address == -EFAULT)
 314                 return 0;
 315         if (!page_vma_mapped_walk(&pvmw))
 316                 return 0;
 317         page_vma_mapped_walk_done(&pvmw);
 318         return 1;
 319 }