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