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, spinlock_t **ptlp)
  17 {
  18         pte_t ptent;
  19
  20         if (pvmw->flags & PVMW_SYNC) {
  21                 /* Use the stricter lookup */
  22                 pvmw->pte = pte_offset_map_lock(pvmw->vma->vm_mm, pvmw->pmd,
  23                                                 pvmw->address, &pvmw->ptl);
  24                 *ptlp = pvmw->ptl;
  25                 return !!pvmw->pte;
  26         }
  27
  28         /*
  29          * It is important to return the ptl corresponding to pte,
  30          * in case *pvmw->pmd changes underneath us; so we need to
  31          * return it even when choosing not to lock, in case caller
  32          * proceeds to loop over next ptes, and finds a match later.
  33          * Though, in most cases, page lock already protects this.
  34          */
  35         pvmw->pte = pte_offset_map_nolock(pvmw->vma->vm_mm, pvmw->pmd,
  36                                           pvmw->address, ptlp);
  37         if (!pvmw->pte)
  38                 return false;
  39
  40         ptent = ptep_get(pvmw->pte);
  41
  42         if (pvmw->flags & PVMW_MIGRATION) {
  43                 if (!is_swap_pte(ptent))
  44                         return false;
  45         } else if (is_swap_pte(ptent)) {
  46                 swp_entry_t entry;
  47                 /*
  48                  * Handle un-addressable ZONE_DEVICE memory.
  49                  *
  50                  * We get here when we are trying to unmap a private
  51                  * device page from the process address space. Such
  52                  * page is not CPU accessible and thus is mapped as
  53                  * a special swap entry, nonetheless it still does
  54                  * count as a valid regular mapping for the page
  55                  * (and is accounted as such in page maps count).
  56                  *
  57                  * So handle this special case as if it was a normal
  58                  * page mapping ie lock CPU page table and return true.
  59                  *
  60                  * For more details on device private memory see HMM
  61                  * (include/linux/hmm.h or mm/hmm.c).
  62                  */
  63                 entry = pte_to_swp_entry(ptent);
  64                 if (!is_device_private_entry(entry) &&
  65                     !is_device_exclusive_entry(entry))
  66                         return false;
  67         } else if (!pte_present(ptent)) {
  68                 return false;
  69         }
  70         pvmw->ptl = *ptlp;
  71         spin_lock(pvmw->ptl);
  72         return true;
  73 }
  74
  75 /**
  76  * check_pte - check if @pvmw->page is mapped at the @pvmw->pte
  77  * @pvmw: page_vma_mapped_walk struct, includes a pair pte and page for checking
  78  *
  79  * page_vma_mapped_walk() found a place where @pvmw->page is *potentially*
  80  * mapped. check_pte() has to validate this.
  81  *
  82  * pvmw->pte may point to empty PTE, swap PTE or PTE pointing to
  83  * arbitrary page.
  84  *
  85  * If PVMW_MIGRATION flag is set, returns true if @pvmw->pte contains migration
  86  * entry that points to @pvmw->page or any subpage in case of THP.
  87  *
  88  * If PVMW_MIGRATION flag is not set, returns true if pvmw->pte points to
  89  * pvmw->page or any subpage in case of THP.
  90  *
  91  * Otherwise, return false.
  92  *
  93  */
  94 static bool check_pte(struct page_vma_mapped_walk *pvmw)
  95 {
  96         unsigned long pfn;
  97         pte_t ptent = ptep_get(pvmw->pte);
  98
  99         if (pvmw->flags & PVMW_MIGRATION) {
 100                 swp_entry_t entry;
 101                 if (!is_swap_pte(ptent))
 102                         return false;
 103                 entry = pte_to_swp_entry(ptent);
 104
 105                 if (!is_migration_entry(entry) &&
 106                     !is_device_exclusive_entry(entry))
 107                         return false;
 108
 109                 pfn = swp_offset_pfn(entry);
 110         } else if (is_swap_pte(ptent)) {
 111                 swp_entry_t entry;
 112
 113                 /* Handle un-addressable ZONE_DEVICE memory */
 114                 entry = pte_to_swp_entry(ptent);
 115                 if (!is_device_private_entry(entry) &&
 116                     !is_device_exclusive_entry(entry))
 117                         return false;
 118
 119                 pfn = swp_offset_pfn(entry);
 120         } else {
 121                 if (!pte_present(ptent))
 122                         return false;
 123
 124                 pfn = pte_pfn(ptent);
 125         }
 126
 127         return (pfn - pvmw->pfn) < pvmw->nr_pages;
 128 }
 129
 130 /* Returns true if the two ranges overlap.  Careful to not overflow. */
 131 static bool check_pmd(unsigned long pfn, struct page_vma_mapped_walk *pvmw)
 132 {
 133         if ((pfn + HPAGE_PMD_NR - 1) < pvmw->pfn)
 134                 return false;
 135         if (pfn > pvmw->pfn + pvmw->nr_pages - 1)
 136                 return false;
 137         return true;
 138 }
 139
 140 static void step_forward(struct page_vma_mapped_walk *pvmw, unsigned long size)
 141 {
 142         pvmw->address = (pvmw->address + size) & ~(size - 1);
 143         if (!pvmw->address)
 144                 pvmw->address = ULONG_MAX;
 145 }
 146
 147 /**
 148  * page_vma_mapped_walk - check if @pvmw->pfn is mapped in @pvmw->vma at
 149  * @pvmw->address
 150  * @pvmw: pointer to struct page_vma_mapped_walk. page, vma, address and flags
 151  * must be set. pmd, pte and ptl must be NULL.
 152  *
 153  * Returns true if the page is mapped in the vma. @pvmw->pmd and @pvmw->pte point
 154  * to relevant page table entries. @pvmw->ptl is locked. @pvmw->address is
 155  * adjusted if needed (for PTE-mapped THPs).
 156  *
 157  * If @pvmw->pmd is set but @pvmw->pte is not, you have found PMD-mapped page
 158  * (usually THP). For PTE-mapped THP, you should run page_vma_mapped_walk() in
 159  * a loop to find all PTEs that map the THP.
 160  *
 161  * For HugeTLB pages, @pvmw->pte is set to the relevant page table entry
 162  * regardless of which page table level the page is mapped at. @pvmw->pmd is
 163  * NULL.
 164  *
 165  * Returns false if there are no more page table entries for the page in
 166  * the vma. @pvmw->ptl is unlocked and @pvmw->pte is unmapped.
 167  *
 168  * If you need to stop the walk before page_vma_mapped_walk() returned false,
 169  * use page_vma_mapped_walk_done(). It will do the housekeeping.
 170  */
 171 bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
 172 {
 173         struct vm_area_struct *vma = pvmw->vma;
 174         struct mm_struct *mm = vma->vm_mm;
 175         unsigned long end;
 176         spinlock_t *ptl;
 177         pgd_t *pgd;
 178         p4d_t *p4d;
 179         pud_t *pud;
 180         pmd_t pmde;
 181
 182         /* The only possible pmd mapping has been handled on last iteration */
 183         if (pvmw->pmd && !pvmw->pte)
 184                 return not_found(pvmw);
 185
 186         if (unlikely(is_vm_hugetlb_page(vma))) {
 187                 struct hstate *hstate = hstate_vma(vma);
 188                 unsigned long size = huge_page_size(hstate);
 189                 /* The only possible mapping was handled on last iteration */
 190                 if (pvmw->pte)
 191                         return not_found(pvmw);
 192                 /*
 193                  * All callers that get here will already hold the
 194                  * i_mmap_rwsem.  Therefore, no additional locks need to be
 195                  * taken before calling hugetlb_walk().
 196                  */
 197                 pvmw->pte = hugetlb_walk(vma, pvmw->address, size);
 198                 if (!pvmw->pte)
 199                         return false;
 200
 201                 pvmw->ptl = huge_pte_lock(hstate, mm, pvmw->pte);
 202                 if (!check_pte(pvmw))
 203                         return not_found(pvmw);
 204                 return true;
 205         }
 206
 207         end = vma_address_end(pvmw);
 208         if (pvmw->pte)
 209                 goto next_pte;
 210 restart:
 211         do {
 212                 pgd = pgd_offset(mm, pvmw->address);
 213                 if (!pgd_present(*pgd)) {
 214                         step_forward(pvmw, PGDIR_SIZE);
 215                         continue;
 216                 }
 217                 p4d = p4d_offset(pgd, pvmw->address);
 218                 if (!p4d_present(*p4d)) {
 219                         step_forward(pvmw, P4D_SIZE);
 220                         continue;
 221                 }
 222                 pud = pud_offset(p4d, pvmw->address);
 223                 if (!pud_present(*pud)) {
 224                         step_forward(pvmw, PUD_SIZE);
 225                         continue;
 226                 }
 227
 228                 pvmw->pmd = pmd_offset(pud, pvmw->address);
 229                 /*
 230                  * Make sure the pmd value isn't cached in a register by the
 231                  * compiler and used as a stale value after we've observed a
 232                  * subsequent update.
 233                  */
 234                 pmde = pmdp_get_lockless(pvmw->pmd);
 235
 236                 if (pmd_trans_huge(pmde) || is_pmd_migration_entry(pmde) ||
 237                     (pmd_present(pmde) && pmd_devmap(pmde))) {
 238                         pvmw->ptl = pmd_lock(mm, pvmw->pmd);
 239                         pmde = *pvmw->pmd;
 240                         if (!pmd_present(pmde)) {
 241                                 swp_entry_t entry;
 242
 243                                 if (!thp_migration_supported() ||
 244                                     !(pvmw->flags & PVMW_MIGRATION))
 245                                         return not_found(pvmw);
 246                                 entry = pmd_to_swp_entry(pmde);
 247                                 if (!is_migration_entry(entry) ||
 248                                     !check_pmd(swp_offset_pfn(entry), pvmw))
 249                                         return not_found(pvmw);
 250                                 return true;
 251                         }
 252                         if (likely(pmd_trans_huge(pmde) || pmd_devmap(pmde))) {
 253                                 if (pvmw->flags & PVMW_MIGRATION)
 254                                         return not_found(pvmw);
 255                                 if (!check_pmd(pmd_pfn(pmde), pvmw))
 256                                         return not_found(pvmw);
 257                                 return true;
 258                         }
 259                         /* THP pmd was split under us: handle on pte level */
 260                         spin_unlock(pvmw->ptl);
 261                         pvmw->ptl = NULL;
 262                 } else if (!pmd_present(pmde)) {
 263                         /*
 264                          * If PVMW_SYNC, take and drop THP pmd lock so that we
 265                          * cannot return prematurely, while zap_huge_pmd() has
 266                          * cleared *pmd but not decremented compound_mapcount().
 267                          */
 268                         if ((pvmw->flags & PVMW_SYNC) &&
 269                             transhuge_vma_suitable(vma, pvmw->address) &&
 270                             (pvmw->nr_pages >= HPAGE_PMD_NR)) {
 271                                 spinlock_t *ptl = pmd_lock(mm, pvmw->pmd);
 272
 273                                 spin_unlock(ptl);
 274                         }
 275                         step_forward(pvmw, PMD_SIZE);
 276                         continue;
 277                 }
 278                 if (!map_pte(pvmw, &ptl)) {
 279                         if (!pvmw->pte)
 280                                 goto restart;
 281                         goto next_pte;
 282                 }
 283 this_pte:
 284                 if (check_pte(pvmw))
 285                         return true;
 286 next_pte:
 287                 do {
 288                         pvmw->address += PAGE_SIZE;
 289                         if (pvmw->address >= end)
 290                                 return not_found(pvmw);
 291                         /* Did we cross page table boundary? */
 292                         if ((pvmw->address & (PMD_SIZE - PAGE_SIZE)) == 0) {
 293                                 if (pvmw->ptl) {
 294                                         spin_unlock(pvmw->ptl);
 295                                         pvmw->ptl = NULL;
 296                                 }
 297                                 pte_unmap(pvmw->pte);
 298                                 pvmw->pte = NULL;
 299                                 goto restart;
 300                         }
 301                         pvmw->pte++;
 302                 } while (pte_none(ptep_get(pvmw->pte)));
 303
 304                 if (!pvmw->ptl) {
 305                         pvmw->ptl = ptl;
 306                         spin_lock(pvmw->ptl);
 307                 }
 308                 goto this_pte;
 309         } while (pvmw->address < end);
 310
 311         return false;
 312 }
 313
 314 /**
 315  * page_mapped_in_vma - check whether a page is really mapped in a VMA
 316  * @page: the page to test
 317  * @vma: the VMA to test
 318  *
 319  * Returns 1 if the page is mapped into the page tables of the VMA, 0
 320  * if the page is not mapped into the page tables of this VMA.  Only
 321  * valid for normal file or anonymous VMAs.
 322  */
 323 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
 324 {
 325         struct page_vma_mapped_walk pvmw = {
 326                 .pfn = page_to_pfn(page),
 327                 .nr_pages = 1,
 328                 .vma = vma,
 329                 .flags = PVMW_SYNC,
 330         };
 331
 332         pvmw.address = vma_address(page, vma);
 333         if (pvmw.address == -EFAULT)
 334                 return 0;
 335         if (!page_vma_mapped_walk(&pvmw))
 336                 return 0;
 337         page_vma_mapped_walk_done(&pvmw);
 338         return 1;
 339 }