mm/hmm.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Copyright 2013 Red Hat Inc.
   4  *
   5  * Authors: Jérôme Glisse <jglisse@redhat.com>
   6  */
   7 /*
   8  * Refer to include/linux/hmm.h for information about heterogeneous memory
   9  * management or HMM for short.
  10  */
  11 #include <linux/pagewalk.h>
  12 #include <linux/hmm.h>
  13 #include <linux/init.h>
  14 #include <linux/rmap.h>
  15 #include <linux/swap.h>
  16 #include <linux/slab.h>
  17 #include <linux/sched.h>
  18 #include <linux/mmzone.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/swapops.h>
  21 #include <linux/hugetlb.h>
  22 #include <linux/memremap.h>
  23 #include <linux/sched/mm.h>
  24 #include <linux/jump_label.h>
  25 #include <linux/dma-mapping.h>
  26 #include <linux/mmu_notifier.h>
  27 #include <linux/memory_hotplug.h>
  28
  29 struct hmm_vma_walk {
  30         struct hmm_range        *range;
  31         unsigned long           last;
  32 };
  33
  34 enum {
  35         HMM_NEED_FAULT = 1 << 0,
  36         HMM_NEED_WRITE_FAULT = 1 << 1,
  37         HMM_NEED_ALL_BITS = HMM_NEED_FAULT | HMM_NEED_WRITE_FAULT,
  38 };
  39
  40 static int hmm_pfns_fill(unsigned long addr, unsigned long end,
  41                          struct hmm_range *range, unsigned long cpu_flags)
  42 {
  43         unsigned long i = (addr - range->start) >> PAGE_SHIFT;
  44
  45         for (; addr < end; addr += PAGE_SIZE, i++)
  46                 range->hmm_pfns[i] = cpu_flags;
  47         return 0;
  48 }
  49
  50 /*
  51  * hmm_vma_fault() - fault in a range lacking valid pmd or pte(s)
  52  * @addr: range virtual start address (inclusive)
  53  * @end: range virtual end address (exclusive)
  54  * @required_fault: HMM_NEED_* flags
  55  * @walk: mm_walk structure
  56  * Return: -EBUSY after page fault, or page fault error
  57  *
  58  * This function will be called whenever pmd_none() or pte_none() returns true,
  59  * or whenever there is no page directory covering the virtual address range.
  60  */
  61 static int hmm_vma_fault(unsigned long addr, unsigned long end,
  62                          unsigned int required_fault, struct mm_walk *walk)
  63 {
  64         struct hmm_vma_walk *hmm_vma_walk = walk->private;
  65         struct vm_area_struct *vma = walk->vma;
  66         unsigned int fault_flags = FAULT_FLAG_REMOTE;
  67
  68         WARN_ON_ONCE(!required_fault);
  69         hmm_vma_walk->last = addr;
  70
  71         if (required_fault & HMM_NEED_WRITE_FAULT) {
  72                 if (!(vma->vm_flags & VM_WRITE))
  73                         return -EPERM;
  74                 fault_flags |= FAULT_FLAG_WRITE;
  75         }
  76
  77         for (; addr < end; addr += PAGE_SIZE)
  78                 if (handle_mm_fault(vma, addr, fault_flags) & VM_FAULT_ERROR)
  79                         return -EFAULT;
  80         return -EBUSY;
  81 }
  82
  83 static unsigned int hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
  84                                        unsigned long pfn_req_flags,
  85                                        unsigned long cpu_flags)
  86 {
  87         struct hmm_range *range = hmm_vma_walk->range;
  88
  89         /*
  90          * So we not only consider the individual per page request we also
  91          * consider the default flags requested for the range. The API can
  92          * be used 2 ways. The first one where the HMM user coalesces
  93          * multiple page faults into one request and sets flags per pfn for
  94          * those faults. The second one where the HMM user wants to pre-
  95          * fault a range with specific flags. For the latter one it is a
  96          * waste to have the user pre-fill the pfn arrays with a default
  97          * flags value.
  98          */
  99         pfn_req_flags &= range->pfn_flags_mask;
 100         pfn_req_flags |= range->default_flags;
 101
 102         /* We aren't ask to do anything ... */
 103         if (!(pfn_req_flags & HMM_PFN_REQ_FAULT))
 104                 return 0;
 105
 106         /* Need to write fault ? */
 107         if ((pfn_req_flags & HMM_PFN_REQ_WRITE) &&
 108             !(cpu_flags & HMM_PFN_WRITE))
 109                 return HMM_NEED_FAULT | HMM_NEED_WRITE_FAULT;
 110
 111         /* If CPU page table is not valid then we need to fault */
 112         if (!(cpu_flags & HMM_PFN_VALID))
 113                 return HMM_NEED_FAULT;
 114         return 0;
 115 }
 116
 117 static unsigned int
 118 hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 119                      const unsigned long hmm_pfns[], unsigned long npages,
 120                      unsigned long cpu_flags)
 121 {
 122         struct hmm_range *range = hmm_vma_walk->range;
 123         unsigned int required_fault = 0;
 124         unsigned long i;
 125
 126         /*
 127          * If the default flags do not request to fault pages, and the mask does
 128          * not allow for individual pages to be faulted, then
 129          * hmm_pte_need_fault() will always return 0.
 130          */
 131         if (!((range->default_flags | range->pfn_flags_mask) &
 132               HMM_PFN_REQ_FAULT))
 133                 return 0;
 134
 135         for (i = 0; i < npages; ++i) {
 136                 required_fault |= hmm_pte_need_fault(hmm_vma_walk, hmm_pfns[i],
 137                                                      cpu_flags);
 138                 if (required_fault == HMM_NEED_ALL_BITS)
 139                         return required_fault;
 140         }
 141         return required_fault;
 142 }
 143
 144 static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
 145                              __always_unused int depth, struct mm_walk *walk)
 146 {
 147         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 148         struct hmm_range *range = hmm_vma_walk->range;
 149         unsigned int required_fault;
 150         unsigned long i, npages;
 151         unsigned long *hmm_pfns;
 152
 153         i = (addr - range->start) >> PAGE_SHIFT;
 154         npages = (end - addr) >> PAGE_SHIFT;
 155         hmm_pfns = &range->hmm_pfns[i];
 156         required_fault =
 157                 hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0);
 158         if (!walk->vma) {
 159                 if (required_fault)
 160                         return -EFAULT;
 161                 return hmm_pfns_fill(addr, end, range, HMM_PFN_ERROR);
 162         }
 163         if (required_fault)
 164                 return hmm_vma_fault(addr, end, required_fault, walk);
 165         return hmm_pfns_fill(addr, end, range, 0);
 166 }
 167
 168 static inline unsigned long pmd_to_hmm_pfn_flags(struct hmm_range *range,
 169                                                  pmd_t pmd)
 170 {
 171         if (pmd_protnone(pmd))
 172                 return 0;
 173         return pmd_write(pmd) ? (HMM_PFN_VALID | HMM_PFN_WRITE) : HMM_PFN_VALID;
 174 }
 175
 176 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 177 static int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr,
 178                               unsigned long end, unsigned long hmm_pfns[],
 179                               pmd_t pmd)
 180 {
 181         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 182         struct hmm_range *range = hmm_vma_walk->range;
 183         unsigned long pfn, npages, i;
 184         unsigned int required_fault;
 185         unsigned long cpu_flags;
 186
 187         npages = (end - addr) >> PAGE_SHIFT;
 188         cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
 189         required_fault =
 190                 hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, cpu_flags);
 191         if (required_fault)
 192                 return hmm_vma_fault(addr, end, required_fault, walk);
 193
 194         pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
 195         for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++)
 196                 hmm_pfns[i] = pfn | cpu_flags;
 197         return 0;
 198 }
 199 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 200 /* stub to allow the code below to compile */
 201 int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr,
 202                 unsigned long end, unsigned long hmm_pfns[], pmd_t pmd);
 203 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 204
 205 static inline bool hmm_is_device_private_entry(struct hmm_range *range,
 206                 swp_entry_t entry)
 207 {
 208         return is_device_private_entry(entry) &&
 209                 device_private_entry_to_page(entry)->pgmap->owner ==
 210                 range->dev_private_owner;
 211 }
 212
 213 static inline unsigned long pte_to_hmm_pfn_flags(struct hmm_range *range,
 214                                                  pte_t pte)
 215 {
 216         if (pte_none(pte) || !pte_present(pte) || pte_protnone(pte))
 217                 return 0;
 218         return pte_write(pte) ? (HMM_PFN_VALID | HMM_PFN_WRITE) : HMM_PFN_VALID;
 219 }
 220
 221 static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
 222                               unsigned long end, pmd_t *pmdp, pte_t *ptep,
 223                               unsigned long *hmm_pfn)
 224 {
 225         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 226         struct hmm_range *range = hmm_vma_walk->range;
 227         unsigned int required_fault;
 228         unsigned long cpu_flags;
 229         pte_t pte = *ptep;
 230         uint64_t pfn_req_flags = *hmm_pfn;
 231
 232         if (pte_none(pte)) {
 233                 required_fault =
 234                         hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0);
 235                 if (required_fault)
 236                         goto fault;
 237                 *hmm_pfn = 0;
 238                 return 0;
 239         }
 240
 241         if (!pte_present(pte)) {
 242                 swp_entry_t entry = pte_to_swp_entry(pte);
 243
 244                 /*
 245                  * Never fault in device private pages pages, but just report
 246                  * the PFN even if not present.
 247                  */
 248                 if (hmm_is_device_private_entry(range, entry)) {
 249                         cpu_flags = HMM_PFN_VALID;
 250                         if (is_write_device_private_entry(entry))
 251                                 cpu_flags |= HMM_PFN_WRITE;
 252                         *hmm_pfn = device_private_entry_to_pfn(entry) |
 253                                         cpu_flags;
 254                         return 0;
 255                 }
 256
 257                 required_fault =
 258                         hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0);
 259                 if (!required_fault) {
 260                         *hmm_pfn = 0;
 261                         return 0;
 262                 }
 263
 264                 if (!non_swap_entry(entry))
 265                         goto fault;
 266
 267                 if (is_migration_entry(entry)) {
 268                         pte_unmap(ptep);
 269                         hmm_vma_walk->last = addr;
 270                         migration_entry_wait(walk->mm, pmdp, addr);
 271                         return -EBUSY;
 272                 }
 273
 274                 /* Report error for everything else */
 275                 pte_unmap(ptep);
 276                 return -EFAULT;
 277         }
 278
 279         cpu_flags = pte_to_hmm_pfn_flags(range, pte);
 280         required_fault =
 281                 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, cpu_flags);
 282         if (required_fault)
 283                 goto fault;
 284
 285         /*
 286          * Since each architecture defines a struct page for the zero page, just
 287          * fall through and treat it like a normal page.
 288          */
 289         if (pte_special(pte) && !is_zero_pfn(pte_pfn(pte))) {
 290                 if (hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0)) {
 291                         pte_unmap(ptep);
 292                         return -EFAULT;
 293                 }
 294                 *hmm_pfn = HMM_PFN_ERROR;
 295                 return 0;
 296         }
 297
 298         *hmm_pfn = pte_pfn(pte) | cpu_flags;
 299         return 0;
 300
 301 fault:
 302         pte_unmap(ptep);
 303         /* Fault any virtual address we were asked to fault */
 304         return hmm_vma_fault(addr, end, required_fault, walk);
 305 }
 306
 307 static int hmm_vma_walk_pmd(pmd_t *pmdp,
 308                             unsigned long start,
 309                             unsigned long end,
 310                             struct mm_walk *walk)
 311 {
 312         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 313         struct hmm_range *range = hmm_vma_walk->range;
 314         unsigned long *hmm_pfns =
 315                 &range->hmm_pfns[(start - range->start) >> PAGE_SHIFT];
 316         unsigned long npages = (end - start) >> PAGE_SHIFT;
 317         unsigned long addr = start;
 318         pte_t *ptep;
 319         pmd_t pmd;
 320
 321 again:
 322         pmd = READ_ONCE(*pmdp);
 323         if (pmd_none(pmd))
 324                 return hmm_vma_walk_hole(start, end, -1, walk);
 325
 326         if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
 327                 if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0)) {
 328                         hmm_vma_walk->last = addr;
 329                         pmd_migration_entry_wait(walk->mm, pmdp);
 330                         return -EBUSY;
 331                 }
 332                 return hmm_pfns_fill(start, end, range, 0);
 333         }
 334
 335         if (!pmd_present(pmd)) {
 336                 if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0))
 337                         return -EFAULT;
 338                 return hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
 339         }
 340
 341         if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
 342                 /*
 343                  * No need to take pmd_lock here, even if some other thread
 344                  * is splitting the huge pmd we will get that event through
 345                  * mmu_notifier callback.
 346                  *
 347                  * So just read pmd value and check again it's a transparent
 348                  * huge or device mapping one and compute corresponding pfn
 349                  * values.
 350                  */
 351                 pmd = pmd_read_atomic(pmdp);
 352                 barrier();
 353                 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
 354                         goto again;
 355
 356                 return hmm_vma_handle_pmd(walk, addr, end, hmm_pfns, pmd);
 357         }
 358
 359         /*
 360          * We have handled all the valid cases above ie either none, migration,
 361          * huge or transparent huge. At this point either it is a valid pmd
 362          * entry pointing to pte directory or it is a bad pmd that will not
 363          * recover.
 364          */
 365         if (pmd_bad(pmd)) {
 366                 if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0))
 367                         return -EFAULT;
 368                 return hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
 369         }
 370
 371         ptep = pte_offset_map(pmdp, addr);
 372         for (; addr < end; addr += PAGE_SIZE, ptep++, hmm_pfns++) {
 373                 int r;
 374
 375                 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, hmm_pfns);
 376                 if (r) {
 377                         /* hmm_vma_handle_pte() did pte_unmap() */
 378                         return r;
 379                 }
 380         }
 381         pte_unmap(ptep - 1);
 382         return 0;
 383 }
 384
 385 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && \
 386     defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
 387 static inline unsigned long pud_to_hmm_pfn_flags(struct hmm_range *range,
 388                                                  pud_t pud)
 389 {
 390         if (!pud_present(pud))
 391                 return 0;
 392         return pud_write(pud) ? (HMM_PFN_VALID | HMM_PFN_WRITE) : HMM_PFN_VALID;
 393 }
 394
 395 static int hmm_vma_walk_pud(pud_t *pudp, unsigned long start, unsigned long end,
 396                 struct mm_walk *walk)
 397 {
 398         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 399         struct hmm_range *range = hmm_vma_walk->range;
 400         unsigned long addr = start;
 401         pud_t pud;
 402         int ret = 0;
 403         spinlock_t *ptl = pud_trans_huge_lock(pudp, walk->vma);
 404
 405         if (!ptl)
 406                 return 0;
 407
 408         /* Normally we don't want to split the huge page */
 409         walk->action = ACTION_CONTINUE;
 410
 411         pud = READ_ONCE(*pudp);
 412         if (pud_none(pud)) {
 413                 spin_unlock(ptl);
 414                 return hmm_vma_walk_hole(start, end, -1, walk);
 415         }
 416
 417         if (pud_huge(pud) && pud_devmap(pud)) {
 418                 unsigned long i, npages, pfn;
 419                 unsigned int required_fault;
 420                 unsigned long *hmm_pfns;
 421                 unsigned long cpu_flags;
 422
 423                 if (!pud_present(pud)) {
 424                         spin_unlock(ptl);
 425                         return hmm_vma_walk_hole(start, end, -1, walk);
 426                 }
 427
 428                 i = (addr - range->start) >> PAGE_SHIFT;
 429                 npages = (end - addr) >> PAGE_SHIFT;
 430                 hmm_pfns = &range->hmm_pfns[i];
 431
 432                 cpu_flags = pud_to_hmm_pfn_flags(range, pud);
 433                 required_fault = hmm_range_need_fault(hmm_vma_walk, hmm_pfns,
 434                                                       npages, cpu_flags);
 435                 if (required_fault) {
 436                         spin_unlock(ptl);
 437                         return hmm_vma_fault(addr, end, required_fault, walk);
 438                 }
 439
 440                 pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 441                 for (i = 0; i < npages; ++i, ++pfn)
 442                         hmm_pfns[i] = pfn | cpu_flags;
 443                 goto out_unlock;
 444         }
 445
 446         /* Ask for the PUD to be split */
 447         walk->action = ACTION_SUBTREE;
 448
 449 out_unlock:
 450         spin_unlock(ptl);
 451         return ret;
 452 }
 453 #else
 454 #define hmm_vma_walk_pud        NULL
 455 #endif
 456
 457 #ifdef CONFIG_HUGETLB_PAGE
 458 static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask,
 459                                       unsigned long start, unsigned long end,
 460                                       struct mm_walk *walk)
 461 {
 462         unsigned long addr = start, i, pfn;
 463         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 464         struct hmm_range *range = hmm_vma_walk->range;
 465         struct vm_area_struct *vma = walk->vma;
 466         unsigned int required_fault;
 467         unsigned long pfn_req_flags;
 468         unsigned long cpu_flags;
 469         spinlock_t *ptl;
 470         pte_t entry;
 471
 472         ptl = huge_pte_lock(hstate_vma(vma), walk->mm, pte);
 473         entry = huge_ptep_get(pte);
 474
 475         i = (start - range->start) >> PAGE_SHIFT;
 476         pfn_req_flags = range->hmm_pfns[i];
 477         cpu_flags = pte_to_hmm_pfn_flags(range, entry);
 478         required_fault =
 479                 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, cpu_flags);
 480         if (required_fault) {
 481                 spin_unlock(ptl);
 482                 return hmm_vma_fault(addr, end, required_fault, walk);
 483         }
 484
 485         pfn = pte_pfn(entry) + ((start & ~hmask) >> PAGE_SHIFT);
 486         for (; addr < end; addr += PAGE_SIZE, i++, pfn++)
 487                 range->hmm_pfns[i] = pfn | cpu_flags;
 488
 489         spin_unlock(ptl);
 490         return 0;
 491 }
 492 #else
 493 #define hmm_vma_walk_hugetlb_entry NULL
 494 #endif /* CONFIG_HUGETLB_PAGE */
 495
 496 static int hmm_vma_walk_test(unsigned long start, unsigned long end,
 497                              struct mm_walk *walk)
 498 {
 499         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 500         struct hmm_range *range = hmm_vma_walk->range;
 501         struct vm_area_struct *vma = walk->vma;
 502
 503         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP | VM_MIXEDMAP)) &&
 504             vma->vm_flags & VM_READ)
 505                 return 0;
 506
 507         /*
 508          * vma ranges that don't have struct page backing them or map I/O
 509          * devices directly cannot be handled by hmm_range_fault().
 510          *
 511          * If the vma does not allow read access, then assume that it does not
 512          * allow write access either. HMM does not support architectures that
 513          * allow write without read.
 514          *
 515          * If a fault is requested for an unsupported range then it is a hard
 516          * failure.
 517          */
 518         if (hmm_range_need_fault(hmm_vma_walk,
 519                                  range->hmm_pfns +
 520                                          ((start - range->start) >> PAGE_SHIFT),
 521                                  (end - start) >> PAGE_SHIFT, 0))
 522                 return -EFAULT;
 523
 524         hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
 525
 526         /* Skip this vma and continue processing the next vma. */
 527         return 1;
 528 }
 529
 530 static const struct mm_walk_ops hmm_walk_ops = {
 531         .pud_entry      = hmm_vma_walk_pud,
 532         .pmd_entry      = hmm_vma_walk_pmd,
 533         .pte_hole       = hmm_vma_walk_hole,
 534         .hugetlb_entry  = hmm_vma_walk_hugetlb_entry,
 535         .test_walk      = hmm_vma_walk_test,
 536 };
 537
 538 /**
 539  * hmm_range_fault - try to fault some address in a virtual address range
 540  * @range:      argument structure
 541  *
 542  * Returns 0 on success or one of the following error codes:
 543  *
 544  * -EINVAL:     Invalid arguments or mm or virtual address is in an invalid vma
 545  *              (e.g., device file vma).
 546  * -ENOMEM:     Out of memory.
 547  * -EPERM:      Invalid permission (e.g., asking for write and range is read
 548  *              only).
 549  * -EBUSY:      The range has been invalidated and the caller needs to wait for
 550  *              the invalidation to finish.
 551  * -EFAULT:     A page was requested to be valid and could not be made valid
 552  *              ie it has no backing VMA or it is illegal to access
 553  *
 554  * This is similar to get_user_pages(), except that it can read the page tables
 555  * without mutating them (ie causing faults).
 556  */
 557 int hmm_range_fault(struct hmm_range *range)
 558 {
 559         struct hmm_vma_walk hmm_vma_walk = {
 560                 .range = range,
 561                 .last = range->start,
 562         };
 563         struct mm_struct *mm = range->notifier->mm;
 564         int ret;
 565
 566         mmap_assert_locked(mm);
 567
 568         do {
 569                 /* If range is no longer valid force retry. */
 570                 if (mmu_interval_check_retry(range->notifier,
 571                                              range->notifier_seq))
 572                         return -EBUSY;
 573                 ret = walk_page_range(mm, hmm_vma_walk.last, range->end,
 574                                       &hmm_walk_ops, &hmm_vma_walk);
 575                 /*
 576                  * When -EBUSY is returned the loop restarts with
 577                  * hmm_vma_walk.last set to an address that has not been stored
 578                  * in pfns. All entries < last in the pfn array are set to their
 579                  * output, and all >= are still at their input values.
 580                  */
 581         } while (ret == -EBUSY);
 582         return ret;
 583 }
 584 EXPORT_SYMBOL(hmm_range_fault);