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/mm.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/jump_label.h>
  24 #include <linux/dma-mapping.h>
  25 #include <linux/mmu_notifier.h>
  26 #include <linux/memory_hotplug.h>
  27
  28 #define PA_SECTION_SIZE (1UL << PA_SECTION_SHIFT)
  29
  30 #if IS_ENABLED(CONFIG_HMM_MIRROR)
  31 static const struct mmu_notifier_ops hmm_mmu_notifier_ops;
  32
  33 static inline struct hmm *mm_get_hmm(struct mm_struct *mm)
  34 {
  35         struct hmm *hmm = READ_ONCE(mm->hmm);
  36
  37         if (hmm && kref_get_unless_zero(&hmm->kref))
  38                 return hmm;
  39
  40         return NULL;
  41 }
  42
  43 /**
  44  * hmm_get_or_create - register HMM against an mm (HMM internal)
  45  *
  46  * @mm: mm struct to attach to
  47  * Returns: returns an HMM object, either by referencing the existing
  48  *          (per-process) object, or by creating a new one.
  49  *
  50  * This is not intended to be used directly by device drivers. If mm already
  51  * has an HMM struct then it get a reference on it and returns it. Otherwise
  52  * it allocates an HMM struct, initializes it, associate it with the mm and
  53  * returns it.
  54  */
  55 static struct hmm *hmm_get_or_create(struct mm_struct *mm)
  56 {
  57         struct hmm *hmm = mm_get_hmm(mm);
  58         bool cleanup = false;
  59
  60         if (hmm)
  61                 return hmm;
  62
  63         hmm = kmalloc(sizeof(*hmm), GFP_KERNEL);
  64         if (!hmm)
  65                 return NULL;
  66         init_waitqueue_head(&hmm->wq);
  67         INIT_LIST_HEAD(&hmm->mirrors);
  68         init_rwsem(&hmm->mirrors_sem);
  69         hmm->mmu_notifier.ops = NULL;
  70         INIT_LIST_HEAD(&hmm->ranges);
  71         mutex_init(&hmm->lock);
  72         kref_init(&hmm->kref);
  73         hmm->notifiers = 0;
  74         hmm->dead = false;
  75         hmm->mm = mm;
  76
  77         spin_lock(&mm->page_table_lock);
  78         if (!mm->hmm)
  79                 mm->hmm = hmm;
  80         else
  81                 cleanup = true;
  82         spin_unlock(&mm->page_table_lock);
  83
  84         if (cleanup)
  85                 goto error;
  86
  87         /*
  88          * We should only get here if hold the mmap_sem in write mode ie on
  89          * registration of first mirror through hmm_mirror_register()
  90          */
  91         hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops;
  92         if (__mmu_notifier_register(&hmm->mmu_notifier, mm))
  93                 goto error_mm;
  94
  95         return hmm;
  96
  97 error_mm:
  98         spin_lock(&mm->page_table_lock);
  99         if (mm->hmm == hmm)
 100                 mm->hmm = NULL;
 101         spin_unlock(&mm->page_table_lock);
 102 error:
 103         kfree(hmm);
 104         return NULL;
 105 }
 106
 107 static void hmm_free(struct kref *kref)
 108 {
 109         struct hmm *hmm = container_of(kref, struct hmm, kref);
 110         struct mm_struct *mm = hmm->mm;
 111
 112         mmu_notifier_unregister_no_release(&hmm->mmu_notifier, mm);
 113
 114         spin_lock(&mm->page_table_lock);
 115         if (mm->hmm == hmm)
 116                 mm->hmm = NULL;
 117         spin_unlock(&mm->page_table_lock);
 118
 119         kfree(hmm);
 120 }
 121
 122 static inline void hmm_put(struct hmm *hmm)
 123 {
 124         kref_put(&hmm->kref, hmm_free);
 125 }
 126
 127 void hmm_mm_destroy(struct mm_struct *mm)
 128 {
 129         struct hmm *hmm;
 130
 131         spin_lock(&mm->page_table_lock);
 132         hmm = mm_get_hmm(mm);
 133         mm->hmm = NULL;
 134         if (hmm) {
 135                 hmm->mm = NULL;
 136                 hmm->dead = true;
 137                 spin_unlock(&mm->page_table_lock);
 138                 hmm_put(hmm);
 139                 return;
 140         }
 141
 142         spin_unlock(&mm->page_table_lock);
 143 }
 144
 145 static void hmm_release(struct mmu_notifier *mn, struct mm_struct *mm)
 146 {
 147         struct hmm *hmm = mm_get_hmm(mm);
 148         struct hmm_mirror *mirror;
 149         struct hmm_range *range;
 150
 151         /* Report this HMM as dying. */
 152         hmm->dead = true;
 153
 154         /* Wake-up everyone waiting on any range. */
 155         mutex_lock(&hmm->lock);
 156         list_for_each_entry(range, &hmm->ranges, list)
 157                 range->valid = false;
 158         wake_up_all(&hmm->wq);
 159         mutex_unlock(&hmm->lock);
 160
 161         down_write(&hmm->mirrors_sem);
 162         mirror = list_first_entry_or_null(&hmm->mirrors, struct hmm_mirror,
 163                                           list);
 164         while (mirror) {
 165                 list_del_init(&mirror->list);
 166                 if (mirror->ops->release) {
 167                         /*
 168                          * Drop mirrors_sem so the release callback can wait
 169                          * on any pending work that might itself trigger a
 170                          * mmu_notifier callback and thus would deadlock with
 171                          * us.
 172                          */
 173                         up_write(&hmm->mirrors_sem);
 174                         mirror->ops->release(mirror);
 175                         down_write(&hmm->mirrors_sem);
 176                 }
 177                 mirror = list_first_entry_or_null(&hmm->mirrors,
 178                                                   struct hmm_mirror, list);
 179         }
 180         up_write(&hmm->mirrors_sem);
 181
 182         hmm_put(hmm);
 183 }
 184
 185 static int hmm_invalidate_range_start(struct mmu_notifier *mn,
 186                         const struct mmu_notifier_range *nrange)
 187 {
 188         struct hmm *hmm = mm_get_hmm(nrange->mm);
 189         struct hmm_mirror *mirror;
 190         struct hmm_update update;
 191         struct hmm_range *range;
 192         int ret = 0;
 193
 194         VM_BUG_ON(!hmm);
 195
 196         update.start = nrange->start;
 197         update.end = nrange->end;
 198         update.event = HMM_UPDATE_INVALIDATE;
 199         update.blockable = mmu_notifier_range_blockable(nrange);
 200
 201         if (mmu_notifier_range_blockable(nrange))
 202                 mutex_lock(&hmm->lock);
 203         else if (!mutex_trylock(&hmm->lock)) {
 204                 ret = -EAGAIN;
 205                 goto out;
 206         }
 207         hmm->notifiers++;
 208         list_for_each_entry(range, &hmm->ranges, list) {
 209                 if (update.end < range->start || update.start >= range->end)
 210                         continue;
 211
 212                 range->valid = false;
 213         }
 214         mutex_unlock(&hmm->lock);
 215
 216         if (mmu_notifier_range_blockable(nrange))
 217                 down_read(&hmm->mirrors_sem);
 218         else if (!down_read_trylock(&hmm->mirrors_sem)) {
 219                 ret = -EAGAIN;
 220                 goto out;
 221         }
 222         list_for_each_entry(mirror, &hmm->mirrors, list) {
 223                 int ret;
 224
 225                 ret = mirror->ops->sync_cpu_device_pagetables(mirror, &update);
 226                 if (!update.blockable && ret == -EAGAIN)
 227                         break;
 228         }
 229         up_read(&hmm->mirrors_sem);
 230
 231 out:
 232         hmm_put(hmm);
 233         return ret;
 234 }
 235
 236 static void hmm_invalidate_range_end(struct mmu_notifier *mn,
 237                         const struct mmu_notifier_range *nrange)
 238 {
 239         struct hmm *hmm = mm_get_hmm(nrange->mm);
 240
 241         VM_BUG_ON(!hmm);
 242
 243         mutex_lock(&hmm->lock);
 244         hmm->notifiers--;
 245         if (!hmm->notifiers) {
 246                 struct hmm_range *range;
 247
 248                 list_for_each_entry(range, &hmm->ranges, list) {
 249                         if (range->valid)
 250                                 continue;
 251                         range->valid = true;
 252                 }
 253                 wake_up_all(&hmm->wq);
 254         }
 255         mutex_unlock(&hmm->lock);
 256
 257         hmm_put(hmm);
 258 }
 259
 260 static const struct mmu_notifier_ops hmm_mmu_notifier_ops = {
 261         .release                = hmm_release,
 262         .invalidate_range_start = hmm_invalidate_range_start,
 263         .invalidate_range_end   = hmm_invalidate_range_end,
 264 };
 265
 266 /*
 267  * hmm_mirror_register() - register a mirror against an mm
 268  *
 269  * @mirror: new mirror struct to register
 270  * @mm: mm to register against
 271  * Return: 0 on success, -ENOMEM if no memory, -EINVAL if invalid arguments
 272  *
 273  * To start mirroring a process address space, the device driver must register
 274  * an HMM mirror struct.
 275  *
 276  * THE mm->mmap_sem MUST BE HELD IN WRITE MODE !
 277  */
 278 int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm)
 279 {
 280         /* Sanity check */
 281         if (!mm || !mirror || !mirror->ops)
 282                 return -EINVAL;
 283
 284         mirror->hmm = hmm_get_or_create(mm);
 285         if (!mirror->hmm)
 286                 return -ENOMEM;
 287
 288         down_write(&mirror->hmm->mirrors_sem);
 289         list_add(&mirror->list, &mirror->hmm->mirrors);
 290         up_write(&mirror->hmm->mirrors_sem);
 291
 292         return 0;
 293 }
 294 EXPORT_SYMBOL(hmm_mirror_register);
 295
 296 /*
 297  * hmm_mirror_unregister() - unregister a mirror
 298  *
 299  * @mirror: mirror struct to unregister
 300  *
 301  * Stop mirroring a process address space, and cleanup.
 302  */
 303 void hmm_mirror_unregister(struct hmm_mirror *mirror)
 304 {
 305         struct hmm *hmm = READ_ONCE(mirror->hmm);
 306
 307         if (hmm == NULL)
 308                 return;
 309
 310         down_write(&hmm->mirrors_sem);
 311         list_del_init(&mirror->list);
 312         /* To protect us against double unregister ... */
 313         mirror->hmm = NULL;
 314         up_write(&hmm->mirrors_sem);
 315
 316         hmm_put(hmm);
 317 }
 318 EXPORT_SYMBOL(hmm_mirror_unregister);
 319
 320 struct hmm_vma_walk {
 321         struct hmm_range        *range;
 322         struct dev_pagemap      *pgmap;
 323         unsigned long           last;
 324         bool                    fault;
 325         bool                    block;
 326 };
 327
 328 static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr,
 329                             bool write_fault, uint64_t *pfn)
 330 {
 331         unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE;
 332         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 333         struct hmm_range *range = hmm_vma_walk->range;
 334         struct vm_area_struct *vma = walk->vma;
 335         vm_fault_t ret;
 336
 337         flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY;
 338         flags |= write_fault ? FAULT_FLAG_WRITE : 0;
 339         ret = handle_mm_fault(vma, addr, flags);
 340         if (ret & VM_FAULT_RETRY)
 341                 return -EAGAIN;
 342         if (ret & VM_FAULT_ERROR) {
 343                 *pfn = range->values[HMM_PFN_ERROR];
 344                 return -EFAULT;
 345         }
 346
 347         return -EBUSY;
 348 }
 349
 350 static int hmm_pfns_bad(unsigned long addr,
 351                         unsigned long end,
 352                         struct mm_walk *walk)
 353 {
 354         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 355         struct hmm_range *range = hmm_vma_walk->range;
 356         uint64_t *pfns = range->pfns;
 357         unsigned long i;
 358
 359         i = (addr - range->start) >> PAGE_SHIFT;
 360         for (; addr < end; addr += PAGE_SIZE, i++)
 361                 pfns[i] = range->values[HMM_PFN_ERROR];
 362
 363         return 0;
 364 }
 365
 366 /*
 367  * hmm_vma_walk_hole() - handle a range lacking valid pmd or pte(s)
 368  * @start: range virtual start address (inclusive)
 369  * @end: range virtual end address (exclusive)
 370  * @fault: should we fault or not ?
 371  * @write_fault: write fault ?
 372  * @walk: mm_walk structure
 373  * Return: 0 on success, -EBUSY after page fault, or page fault error
 374  *
 375  * This function will be called whenever pmd_none() or pte_none() returns true,
 376  * or whenever there is no page directory covering the virtual address range.
 377  */
 378 static int hmm_vma_walk_hole_(unsigned long addr, unsigned long end,
 379                               bool fault, bool write_fault,
 380                               struct mm_walk *walk)
 381 {
 382         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 383         struct hmm_range *range = hmm_vma_walk->range;
 384         uint64_t *pfns = range->pfns;
 385         unsigned long i, page_size;
 386
 387         hmm_vma_walk->last = addr;
 388         page_size = hmm_range_page_size(range);
 389         i = (addr - range->start) >> range->page_shift;
 390
 391         for (; addr < end; addr += page_size, i++) {
 392                 pfns[i] = range->values[HMM_PFN_NONE];
 393                 if (fault || write_fault) {
 394                         int ret;
 395
 396                         ret = hmm_vma_do_fault(walk, addr, write_fault,
 397                                                &pfns[i]);
 398                         if (ret != -EBUSY)
 399                                 return ret;
 400                 }
 401         }
 402
 403         return (fault || write_fault) ? -EBUSY : 0;
 404 }
 405
 406 static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 407                                       uint64_t pfns, uint64_t cpu_flags,
 408                                       bool *fault, bool *write_fault)
 409 {
 410         struct hmm_range *range = hmm_vma_walk->range;
 411
 412         if (!hmm_vma_walk->fault)
 413                 return;
 414
 415         /*
 416          * So we not only consider the individual per page request we also
 417          * consider the default flags requested for the range. The API can
 418          * be use in 2 fashions. The first one where the HMM user coalesce
 419          * multiple page fault into one request and set flags per pfns for
 420          * of those faults. The second one where the HMM user want to pre-
 421          * fault a range with specific flags. For the latter one it is a
 422          * waste to have the user pre-fill the pfn arrays with a default
 423          * flags value.
 424          */
 425         pfns = (pfns & range->pfn_flags_mask) | range->default_flags;
 426
 427         /* We aren't ask to do anything ... */
 428         if (!(pfns & range->flags[HMM_PFN_VALID]))
 429                 return;
 430         /* If this is device memory than only fault if explicitly requested */
 431         if ((cpu_flags & range->flags[HMM_PFN_DEVICE_PRIVATE])) {
 432                 /* Do we fault on device memory ? */
 433                 if (pfns & range->flags[HMM_PFN_DEVICE_PRIVATE]) {
 434                         *write_fault = pfns & range->flags[HMM_PFN_WRITE];
 435                         *fault = true;
 436                 }
 437                 return;
 438         }
 439
 440         /* If CPU page table is not valid then we need to fault */
 441         *fault = !(cpu_flags & range->flags[HMM_PFN_VALID]);
 442         /* Need to write fault ? */
 443         if ((pfns & range->flags[HMM_PFN_WRITE]) &&
 444             !(cpu_flags & range->flags[HMM_PFN_WRITE])) {
 445                 *write_fault = true;
 446                 *fault = true;
 447         }
 448 }
 449
 450 static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 451                                  const uint64_t *pfns, unsigned long npages,
 452                                  uint64_t cpu_flags, bool *fault,
 453                                  bool *write_fault)
 454 {
 455         unsigned long i;
 456
 457         if (!hmm_vma_walk->fault) {
 458                 *fault = *write_fault = false;
 459                 return;
 460         }
 461
 462         *fault = *write_fault = false;
 463         for (i = 0; i < npages; ++i) {
 464                 hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags,
 465                                    fault, write_fault);
 466                 if ((*write_fault))
 467                         return;
 468         }
 469 }
 470
 471 static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
 472                              struct mm_walk *walk)
 473 {
 474         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 475         struct hmm_range *range = hmm_vma_walk->range;
 476         bool fault, write_fault;
 477         unsigned long i, npages;
 478         uint64_t *pfns;
 479
 480         i = (addr - range->start) >> PAGE_SHIFT;
 481         npages = (end - addr) >> PAGE_SHIFT;
 482         pfns = &range->pfns[i];
 483         hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 484                              0, &fault, &write_fault);
 485         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 486 }
 487
 488 static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd)
 489 {
 490         if (pmd_protnone(pmd))
 491                 return 0;
 492         return pmd_write(pmd) ? range->flags[HMM_PFN_VALID] |
 493                                 range->flags[HMM_PFN_WRITE] :
 494                                 range->flags[HMM_PFN_VALID];
 495 }
 496
 497 static inline uint64_t pud_to_hmm_pfn_flags(struct hmm_range *range, pud_t pud)
 498 {
 499         if (!pud_present(pud))
 500                 return 0;
 501         return pud_write(pud) ? range->flags[HMM_PFN_VALID] |
 502                                 range->flags[HMM_PFN_WRITE] :
 503                                 range->flags[HMM_PFN_VALID];
 504 }
 505
 506 static int hmm_vma_handle_pmd(struct mm_walk *walk,
 507                               unsigned long addr,
 508                               unsigned long end,
 509                               uint64_t *pfns,
 510                               pmd_t pmd)
 511 {
 512 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 513         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 514         struct hmm_range *range = hmm_vma_walk->range;
 515         unsigned long pfn, npages, i;
 516         bool fault, write_fault;
 517         uint64_t cpu_flags;
 518
 519         npages = (end - addr) >> PAGE_SHIFT;
 520         cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
 521         hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags,
 522                              &fault, &write_fault);
 523
 524         if (pmd_protnone(pmd) || fault || write_fault)
 525                 return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 526
 527         pfn = pmd_pfn(pmd) + pte_index(addr);
 528         for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++) {
 529                 if (pmd_devmap(pmd)) {
 530                         hmm_vma_walk->pgmap = get_dev_pagemap(pfn,
 531                                               hmm_vma_walk->pgmap);
 532                         if (unlikely(!hmm_vma_walk->pgmap))
 533                                 return -EBUSY;
 534                 }
 535                 pfns[i] = hmm_device_entry_from_pfn(range, pfn) | cpu_flags;
 536         }
 537         if (hmm_vma_walk->pgmap) {
 538                 put_dev_pagemap(hmm_vma_walk->pgmap);
 539                 hmm_vma_walk->pgmap = NULL;
 540         }
 541         hmm_vma_walk->last = end;
 542         return 0;
 543 #else
 544         /* If THP is not enabled then we should never reach that code ! */
 545         return -EINVAL;
 546 #endif
 547 }
 548
 549 static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte)
 550 {
 551         if (pte_none(pte) || !pte_present(pte))
 552                 return 0;
 553         return pte_write(pte) ? range->flags[HMM_PFN_VALID] |
 554                                 range->flags[HMM_PFN_WRITE] :
 555                                 range->flags[HMM_PFN_VALID];
 556 }
 557
 558 static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
 559                               unsigned long end, pmd_t *pmdp, pte_t *ptep,
 560                               uint64_t *pfn)
 561 {
 562         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 563         struct hmm_range *range = hmm_vma_walk->range;
 564         struct vm_area_struct *vma = walk->vma;
 565         bool fault, write_fault;
 566         uint64_t cpu_flags;
 567         pte_t pte = *ptep;
 568         uint64_t orig_pfn = *pfn;
 569
 570         *pfn = range->values[HMM_PFN_NONE];
 571         fault = write_fault = false;
 572
 573         if (pte_none(pte)) {
 574                 hmm_pte_need_fault(hmm_vma_walk, orig_pfn, 0,
 575                                    &fault, &write_fault);
 576                 if (fault || write_fault)
 577                         goto fault;
 578                 return 0;
 579         }
 580
 581         if (!pte_present(pte)) {
 582                 swp_entry_t entry = pte_to_swp_entry(pte);
 583
 584                 if (!non_swap_entry(entry)) {
 585                         if (fault || write_fault)
 586                                 goto fault;
 587                         return 0;
 588                 }
 589
 590                 /*
 591                  * This is a special swap entry, ignore migration, use
 592                  * device and report anything else as error.
 593                  */
 594                 if (is_device_private_entry(entry)) {
 595                         cpu_flags = range->flags[HMM_PFN_VALID] |
 596                                 range->flags[HMM_PFN_DEVICE_PRIVATE];
 597                         cpu_flags |= is_write_device_private_entry(entry) ?
 598                                 range->flags[HMM_PFN_WRITE] : 0;
 599                         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 600                                            &fault, &write_fault);
 601                         if (fault || write_fault)
 602                                 goto fault;
 603                         *pfn = hmm_device_entry_from_pfn(range,
 604                                             swp_offset(entry));
 605                         *pfn |= cpu_flags;
 606                         return 0;
 607                 }
 608
 609                 if (is_migration_entry(entry)) {
 610                         if (fault || write_fault) {
 611                                 pte_unmap(ptep);
 612                                 hmm_vma_walk->last = addr;
 613                                 migration_entry_wait(vma->vm_mm,
 614                                                      pmdp, addr);
 615                                 return -EBUSY;
 616                         }
 617                         return 0;
 618                 }
 619
 620                 /* Report error for everything else */
 621                 *pfn = range->values[HMM_PFN_ERROR];
 622                 return -EFAULT;
 623         } else {
 624                 cpu_flags = pte_to_hmm_pfn_flags(range, pte);
 625                 hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 626                                    &fault, &write_fault);
 627         }
 628
 629         if (fault || write_fault)
 630                 goto fault;
 631
 632         if (pte_devmap(pte)) {
 633                 hmm_vma_walk->pgmap = get_dev_pagemap(pte_pfn(pte),
 634                                               hmm_vma_walk->pgmap);
 635                 if (unlikely(!hmm_vma_walk->pgmap))
 636                         return -EBUSY;
 637         } else if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && pte_special(pte)) {
 638                 *pfn = range->values[HMM_PFN_SPECIAL];
 639                 return -EFAULT;
 640         }
 641
 642         *pfn = hmm_device_entry_from_pfn(range, pte_pfn(pte)) | cpu_flags;
 643         return 0;
 644
 645 fault:
 646         if (hmm_vma_walk->pgmap) {
 647                 put_dev_pagemap(hmm_vma_walk->pgmap);
 648                 hmm_vma_walk->pgmap = NULL;
 649         }
 650         pte_unmap(ptep);
 651         /* Fault any virtual address we were asked to fault */
 652         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 653 }
 654
 655 static int hmm_vma_walk_pmd(pmd_t *pmdp,
 656                             unsigned long start,
 657                             unsigned long end,
 658                             struct mm_walk *walk)
 659 {
 660         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 661         struct hmm_range *range = hmm_vma_walk->range;
 662         struct vm_area_struct *vma = walk->vma;
 663         uint64_t *pfns = range->pfns;
 664         unsigned long addr = start, i;
 665         pte_t *ptep;
 666         pmd_t pmd;
 667
 668
 669 again:
 670         pmd = READ_ONCE(*pmdp);
 671         if (pmd_none(pmd))
 672                 return hmm_vma_walk_hole(start, end, walk);
 673
 674         if (pmd_huge(pmd) && (range->vma->vm_flags & VM_HUGETLB))
 675                 return hmm_pfns_bad(start, end, walk);
 676
 677         if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
 678                 bool fault, write_fault;
 679                 unsigned long npages;
 680                 uint64_t *pfns;
 681
 682                 i = (addr - range->start) >> PAGE_SHIFT;
 683                 npages = (end - addr) >> PAGE_SHIFT;
 684                 pfns = &range->pfns[i];
 685
 686                 hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 687                                      0, &fault, &write_fault);
 688                 if (fault || write_fault) {
 689                         hmm_vma_walk->last = addr;
 690                         pmd_migration_entry_wait(vma->vm_mm, pmdp);
 691                         return -EBUSY;
 692                 }
 693                 return 0;
 694         } else if (!pmd_present(pmd))
 695                 return hmm_pfns_bad(start, end, walk);
 696
 697         if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
 698                 /*
 699                  * No need to take pmd_lock here, even if some other threads
 700                  * is splitting the huge pmd we will get that event through
 701                  * mmu_notifier callback.
 702                  *
 703                  * So just read pmd value and check again its a transparent
 704                  * huge or device mapping one and compute corresponding pfn
 705                  * values.
 706                  */
 707                 pmd = pmd_read_atomic(pmdp);
 708                 barrier();
 709                 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
 710                         goto again;
 711
 712                 i = (addr - range->start) >> PAGE_SHIFT;
 713                 return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd);
 714         }
 715
 716         /*
 717          * We have handled all the valid case above ie either none, migration,
 718          * huge or transparent huge. At this point either it is a valid pmd
 719          * entry pointing to pte directory or it is a bad pmd that will not
 720          * recover.
 721          */
 722         if (pmd_bad(pmd))
 723                 return hmm_pfns_bad(start, end, walk);
 724
 725         ptep = pte_offset_map(pmdp, addr);
 726         i = (addr - range->start) >> PAGE_SHIFT;
 727         for (; addr < end; addr += PAGE_SIZE, ptep++, i++) {
 728                 int r;
 729
 730                 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]);
 731                 if (r) {
 732                         /* hmm_vma_handle_pte() did unmap pte directory */
 733                         hmm_vma_walk->last = addr;
 734                         return r;
 735                 }
 736         }
 737         if (hmm_vma_walk->pgmap) {
 738                 /*
 739                  * We do put_dev_pagemap() here and not in hmm_vma_handle_pte()
 740                  * so that we can leverage get_dev_pagemap() optimization which
 741                  * will not re-take a reference on a pgmap if we already have
 742                  * one.
 743                  */
 744                 put_dev_pagemap(hmm_vma_walk->pgmap);
 745                 hmm_vma_walk->pgmap = NULL;
 746         }
 747         pte_unmap(ptep - 1);
 748
 749         hmm_vma_walk->last = addr;
 750         return 0;
 751 }
 752
 753 static int hmm_vma_walk_pud(pud_t *pudp,
 754                             unsigned long start,
 755                             unsigned long end,
 756                             struct mm_walk *walk)
 757 {
 758         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 759         struct hmm_range *range = hmm_vma_walk->range;
 760         unsigned long addr = start, next;
 761         pmd_t *pmdp;
 762         pud_t pud;
 763         int ret;
 764
 765 again:
 766         pud = READ_ONCE(*pudp);
 767         if (pud_none(pud))
 768                 return hmm_vma_walk_hole(start, end, walk);
 769
 770         if (pud_huge(pud) && pud_devmap(pud)) {
 771                 unsigned long i, npages, pfn;
 772                 uint64_t *pfns, cpu_flags;
 773                 bool fault, write_fault;
 774
 775                 if (!pud_present(pud))
 776                         return hmm_vma_walk_hole(start, end, walk);
 777
 778                 i = (addr - range->start) >> PAGE_SHIFT;
 779                 npages = (end - addr) >> PAGE_SHIFT;
 780                 pfns = &range->pfns[i];
 781
 782                 cpu_flags = pud_to_hmm_pfn_flags(range, pud);
 783                 hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 784                                      cpu_flags, &fault, &write_fault);
 785                 if (fault || write_fault)
 786                         return hmm_vma_walk_hole_(addr, end, fault,
 787                                                 write_fault, walk);
 788
 789                 pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 790                 for (i = 0; i < npages; ++i, ++pfn) {
 791                         hmm_vma_walk->pgmap = get_dev_pagemap(pfn,
 792                                               hmm_vma_walk->pgmap);
 793                         if (unlikely(!hmm_vma_walk->pgmap))
 794                                 return -EBUSY;
 795                         pfns[i] = hmm_device_entry_from_pfn(range, pfn) |
 796                                   cpu_flags;
 797                 }
 798                 if (hmm_vma_walk->pgmap) {
 799                         put_dev_pagemap(hmm_vma_walk->pgmap);
 800                         hmm_vma_walk->pgmap = NULL;
 801                 }
 802                 hmm_vma_walk->last = end;
 803                 return 0;
 804         }
 805
 806         split_huge_pud(walk->vma, pudp, addr);
 807         if (pud_none(*pudp))
 808                 goto again;
 809
 810         pmdp = pmd_offset(pudp, addr);
 811         do {
 812                 next = pmd_addr_end(addr, end);
 813                 ret = hmm_vma_walk_pmd(pmdp, addr, next, walk);
 814                 if (ret)
 815                         return ret;
 816         } while (pmdp++, addr = next, addr != end);
 817
 818         return 0;
 819 }
 820
 821 static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask,
 822                                       unsigned long start, unsigned long end,
 823                                       struct mm_walk *walk)
 824 {
 825 #ifdef CONFIG_HUGETLB_PAGE
 826         unsigned long addr = start, i, pfn, mask, size, pfn_inc;
 827         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 828         struct hmm_range *range = hmm_vma_walk->range;
 829         struct vm_area_struct *vma = walk->vma;
 830         struct hstate *h = hstate_vma(vma);
 831         uint64_t orig_pfn, cpu_flags;
 832         bool fault, write_fault;
 833         spinlock_t *ptl;
 834         pte_t entry;
 835         int ret = 0;
 836
 837         size = 1UL << huge_page_shift(h);
 838         mask = size - 1;
 839         if (range->page_shift != PAGE_SHIFT) {
 840                 /* Make sure we are looking at full page. */
 841                 if (start & mask)
 842                         return -EINVAL;
 843                 if (end < (start + size))
 844                         return -EINVAL;
 845                 pfn_inc = size >> PAGE_SHIFT;
 846         } else {
 847                 pfn_inc = 1;
 848                 size = PAGE_SIZE;
 849         }
 850
 851
 852         ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
 853         entry = huge_ptep_get(pte);
 854
 855         i = (start - range->start) >> range->page_shift;
 856         orig_pfn = range->pfns[i];
 857         range->pfns[i] = range->values[HMM_PFN_NONE];
 858         cpu_flags = pte_to_hmm_pfn_flags(range, entry);
 859         fault = write_fault = false;
 860         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 861                            &fault, &write_fault);
 862         if (fault || write_fault) {
 863                 ret = -ENOENT;
 864                 goto unlock;
 865         }
 866
 867         pfn = pte_pfn(entry) + ((start & mask) >> range->page_shift);
 868         for (; addr < end; addr += size, i++, pfn += pfn_inc)
 869                 range->pfns[i] = hmm_device_entry_from_pfn(range, pfn) |
 870                                  cpu_flags;
 871         hmm_vma_walk->last = end;
 872
 873 unlock:
 874         spin_unlock(ptl);
 875
 876         if (ret == -ENOENT)
 877                 return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 878
 879         return ret;
 880 #else /* CONFIG_HUGETLB_PAGE */
 881         return -EINVAL;
 882 #endif
 883 }
 884
 885 static void hmm_pfns_clear(struct hmm_range *range,
 886                            uint64_t *pfns,
 887                            unsigned long addr,
 888                            unsigned long end)
 889 {
 890         for (; addr < end; addr += PAGE_SIZE, pfns++)
 891                 *pfns = range->values[HMM_PFN_NONE];
 892 }
 893
 894 /*
 895  * hmm_range_register() - start tracking change to CPU page table over a range
 896  * @range: range
 897  * @mm: the mm struct for the range of virtual address
 898  * @start: start virtual address (inclusive)
 899  * @end: end virtual address (exclusive)
 900  * @page_shift: expect page shift for the range
 901  * Returns 0 on success, -EFAULT if the address space is no longer valid
 902  *
 903  * Track updates to the CPU page table see include/linux/hmm.h
 904  */
 905 int hmm_range_register(struct hmm_range *range,
 906                        struct mm_struct *mm,
 907                        unsigned long start,
 908                        unsigned long end,
 909                        unsigned page_shift)
 910 {
 911         unsigned long mask = ((1UL << page_shift) - 1UL);
 912         struct hmm *hmm;
 913
 914         range->valid = false;
 915         range->hmm = NULL;
 916
 917         if ((start & mask) || (end & mask))
 918                 return -EINVAL;
 919         if (start >= end)
 920                 return -EINVAL;
 921
 922         range->page_shift = page_shift;
 923         range->start = start;
 924         range->end = end;
 925
 926         hmm = hmm_get_or_create(mm);
 927         if (!hmm)
 928                 return -EFAULT;
 929
 930         /* Check if hmm_mm_destroy() was call. */
 931         if (hmm->mm == NULL || hmm->dead) {
 932                 hmm_put(hmm);
 933                 return -EFAULT;
 934         }
 935
 936         /* Initialize range to track CPU page table updates. */
 937         mutex_lock(&hmm->lock);
 938
 939         range->hmm = hmm;
 940         list_add_rcu(&range->list, &hmm->ranges);
 941
 942         /*
 943          * If there are any concurrent notifiers we have to wait for them for
 944          * the range to be valid (see hmm_range_wait_until_valid()).
 945          */
 946         if (!hmm->notifiers)
 947                 range->valid = true;
 948         mutex_unlock(&hmm->lock);
 949
 950         return 0;
 951 }
 952 EXPORT_SYMBOL(hmm_range_register);
 953
 954 /*
 955  * hmm_range_unregister() - stop tracking change to CPU page table over a range
 956  * @range: range
 957  *
 958  * Range struct is used to track updates to the CPU page table after a call to
 959  * hmm_range_register(). See include/linux/hmm.h for how to use it.
 960  */
 961 void hmm_range_unregister(struct hmm_range *range)
 962 {
 963         struct hmm *hmm = range->hmm;
 964
 965         /* Sanity check this really should not happen. */
 966         if (hmm == NULL || range->end <= range->start)
 967                 return;
 968
 969         mutex_lock(&hmm->lock);
 970         list_del_rcu(&range->list);
 971         mutex_unlock(&hmm->lock);
 972
 973         /* Drop reference taken by hmm_range_register() */
 974         range->valid = false;
 975         hmm_put(hmm);
 976         range->hmm = NULL;
 977 }
 978 EXPORT_SYMBOL(hmm_range_unregister);
 979
 980 /*
 981  * hmm_range_snapshot() - snapshot CPU page table for a range
 982  * @range: range
 983  * Return: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid
 984  *          permission (for instance asking for write and range is read only),
 985  *          -EAGAIN if you need to retry, -EFAULT invalid (ie either no valid
 986  *          vma or it is illegal to access that range), number of valid pages
 987  *          in range->pfns[] (from range start address).
 988  *
 989  * This snapshots the CPU page table for a range of virtual addresses. Snapshot
 990  * validity is tracked by range struct. See in include/linux/hmm.h for example
 991  * on how to use.
 992  */
 993 long hmm_range_snapshot(struct hmm_range *range)
 994 {
 995         const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
 996         unsigned long start = range->start, end;
 997         struct hmm_vma_walk hmm_vma_walk;
 998         struct hmm *hmm = range->hmm;
 999         struct vm_area_struct *vma;
1000         struct mm_walk mm_walk;
1001
1002         /* Check if hmm_mm_destroy() was call. */
1003         if (hmm->mm == NULL || hmm->dead)
1004                 return -EFAULT;
1005
1006         do {
1007                 /* If range is no longer valid force retry. */
1008                 if (!range->valid)
1009                         return -EAGAIN;
1010
1011                 vma = find_vma(hmm->mm, start);
1012                 if (vma == NULL || (vma->vm_flags & device_vma))
1013                         return -EFAULT;
1014
1015                 if (is_vm_hugetlb_page(vma)) {
1016                         if (huge_page_shift(hstate_vma(vma)) !=
1017                                     range->page_shift &&
1018                             range->page_shift != PAGE_SHIFT)
1019                                 return -EINVAL;
1020                 } else {
1021                         if (range->page_shift != PAGE_SHIFT)
1022                                 return -EINVAL;
1023                 }
1024
1025                 if (!(vma->vm_flags & VM_READ)) {
1026                         /*
1027                          * If vma do not allow read access, then assume that it
1028                          * does not allow write access, either. HMM does not
1029                          * support architecture that allow write without read.
1030                          */
1031                         hmm_pfns_clear(range, range->pfns,
1032                                 range->start, range->end);
1033                         return -EPERM;
1034                 }
1035
1036                 range->vma = vma;
1037                 hmm_vma_walk.pgmap = NULL;
1038                 hmm_vma_walk.last = start;
1039                 hmm_vma_walk.fault = false;
1040                 hmm_vma_walk.range = range;
1041                 mm_walk.private = &hmm_vma_walk;
1042                 end = min(range->end, vma->vm_end);
1043
1044                 mm_walk.vma = vma;
1045                 mm_walk.mm = vma->vm_mm;
1046                 mm_walk.pte_entry = NULL;
1047                 mm_walk.test_walk = NULL;
1048                 mm_walk.hugetlb_entry = NULL;
1049                 mm_walk.pud_entry = hmm_vma_walk_pud;
1050                 mm_walk.pmd_entry = hmm_vma_walk_pmd;
1051                 mm_walk.pte_hole = hmm_vma_walk_hole;
1052                 mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry;
1053
1054                 walk_page_range(start, end, &mm_walk);
1055                 start = end;
1056         } while (start < range->end);
1057
1058         return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
1059 }
1060 EXPORT_SYMBOL(hmm_range_snapshot);
1061
1062 /*
1063  * hmm_range_fault() - try to fault some address in a virtual address range
1064  * @range: range being faulted
1065  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
1066  * Return: number of valid pages in range->pfns[] (from range start
1067  *          address). This may be zero. If the return value is negative,
1068  *          then one of the following values may be returned:
1069  *
1070  *           -EINVAL  invalid arguments or mm or virtual address are in an
1071  *                    invalid vma (for instance device file vma).
1072  *           -ENOMEM: Out of memory.
1073  *           -EPERM:  Invalid permission (for instance asking for write and
1074  *                    range is read only).
1075  *           -EAGAIN: If you need to retry and mmap_sem was drop. This can only
1076  *                    happens if block argument is false.
1077  *           -EBUSY:  If the the range is being invalidated and you should wait
1078  *                    for invalidation to finish.
1079  *           -EFAULT: Invalid (ie either no valid vma or it is illegal to access
1080  *                    that range), number of valid pages in range->pfns[] (from
1081  *                    range start address).
1082  *
1083  * This is similar to a regular CPU page fault except that it will not trigger
1084  * any memory migration if the memory being faulted is not accessible by CPUs
1085  * and caller does not ask for migration.
1086  *
1087  * On error, for one virtual address in the range, the function will mark the
1088  * corresponding HMM pfn entry with an error flag.
1089  */
1090 long hmm_range_fault(struct hmm_range *range, bool block)
1091 {
1092         const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
1093         unsigned long start = range->start, end;
1094         struct hmm_vma_walk hmm_vma_walk;
1095         struct hmm *hmm = range->hmm;
1096         struct vm_area_struct *vma;
1097         struct mm_walk mm_walk;
1098         int ret;
1099
1100         /* Check if hmm_mm_destroy() was call. */
1101         if (hmm->mm == NULL || hmm->dead)
1102                 return -EFAULT;
1103
1104         do {
1105                 /* If range is no longer valid force retry. */
1106                 if (!range->valid) {
1107                         up_read(&hmm->mm->mmap_sem);
1108                         return -EAGAIN;
1109                 }
1110
1111                 vma = find_vma(hmm->mm, start);
1112                 if (vma == NULL || (vma->vm_flags & device_vma))
1113                         return -EFAULT;
1114
1115                 if (is_vm_hugetlb_page(vma)) {
1116                         if (huge_page_shift(hstate_vma(vma)) !=
1117                             range->page_shift &&
1118                             range->page_shift != PAGE_SHIFT)
1119                                 return -EINVAL;
1120                 } else {
1121                         if (range->page_shift != PAGE_SHIFT)
1122                                 return -EINVAL;
1123                 }
1124
1125                 if (!(vma->vm_flags & VM_READ)) {
1126                         /*
1127                          * If vma do not allow read access, then assume that it
1128                          * does not allow write access, either. HMM does not
1129                          * support architecture that allow write without read.
1130                          */
1131                         hmm_pfns_clear(range, range->pfns,
1132                                 range->start, range->end);
1133                         return -EPERM;
1134                 }
1135
1136                 range->vma = vma;
1137                 hmm_vma_walk.pgmap = NULL;
1138                 hmm_vma_walk.last = start;
1139                 hmm_vma_walk.fault = true;
1140                 hmm_vma_walk.block = block;
1141                 hmm_vma_walk.range = range;
1142                 mm_walk.private = &hmm_vma_walk;
1143                 end = min(range->end, vma->vm_end);
1144
1145                 mm_walk.vma = vma;
1146                 mm_walk.mm = vma->vm_mm;
1147                 mm_walk.pte_entry = NULL;
1148                 mm_walk.test_walk = NULL;
1149                 mm_walk.hugetlb_entry = NULL;
1150                 mm_walk.pud_entry = hmm_vma_walk_pud;
1151                 mm_walk.pmd_entry = hmm_vma_walk_pmd;
1152                 mm_walk.pte_hole = hmm_vma_walk_hole;
1153                 mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry;
1154
1155                 do {
1156                         ret = walk_page_range(start, end, &mm_walk);
1157                         start = hmm_vma_walk.last;
1158
1159                         /* Keep trying while the range is valid. */
1160                 } while (ret == -EBUSY && range->valid);
1161
1162                 if (ret) {
1163                         unsigned long i;
1164
1165                         i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
1166                         hmm_pfns_clear(range, &range->pfns[i],
1167                                 hmm_vma_walk.last, range->end);
1168                         return ret;
1169                 }
1170                 start = end;
1171
1172         } while (start < range->end);
1173
1174         return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
1175 }
1176 EXPORT_SYMBOL(hmm_range_fault);
1177
1178 /**
1179  * hmm_range_dma_map() - hmm_range_fault() and dma map page all in one.
1180  * @range: range being faulted
1181  * @device: device against to dma map page to
1182  * @daddrs: dma address of mapped pages
1183  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
1184  * Return: number of pages mapped on success, -EAGAIN if mmap_sem have been
1185  *          drop and you need to try again, some other error value otherwise
1186  *
1187  * Note same usage pattern as hmm_range_fault().
1188  */
1189 long hmm_range_dma_map(struct hmm_range *range,
1190                        struct device *device,
1191                        dma_addr_t *daddrs,
1192                        bool block)
1193 {
1194         unsigned long i, npages, mapped;
1195         long ret;
1196
1197         ret = hmm_range_fault(range, block);
1198         if (ret <= 0)
1199                 return ret ? ret : -EBUSY;
1200
1201         npages = (range->end - range->start) >> PAGE_SHIFT;
1202         for (i = 0, mapped = 0; i < npages; ++i) {
1203                 enum dma_data_direction dir = DMA_TO_DEVICE;
1204                 struct page *page;
1205
1206                 /*
1207                  * FIXME need to update DMA API to provide invalid DMA address
1208                  * value instead of a function to test dma address value. This
1209                  * would remove lot of dumb code duplicated accross many arch.
1210                  *
1211                  * For now setting it to 0 here is good enough as the pfns[]
1212                  * value is what is use to check what is valid and what isn't.
1213                  */
1214                 daddrs[i] = 0;
1215
1216                 page = hmm_device_entry_to_page(range, range->pfns[i]);
1217                 if (page == NULL)
1218                         continue;
1219
1220                 /* Check if range is being invalidated */
1221                 if (!range->valid) {
1222                         ret = -EBUSY;
1223                         goto unmap;
1224                 }
1225
1226                 /* If it is read and write than map bi-directional. */
1227                 if (range->pfns[i] & range->flags[HMM_PFN_WRITE])
1228                         dir = DMA_BIDIRECTIONAL;
1229
1230                 daddrs[i] = dma_map_page(device, page, 0, PAGE_SIZE, dir);
1231                 if (dma_mapping_error(device, daddrs[i])) {
1232                         ret = -EFAULT;
1233                         goto unmap;
1234                 }
1235
1236                 mapped++;
1237         }
1238
1239         return mapped;
1240
1241 unmap:
1242         for (npages = i, i = 0; (i < npages) && mapped; ++i) {
1243                 enum dma_data_direction dir = DMA_TO_DEVICE;
1244                 struct page *page;
1245
1246                 page = hmm_device_entry_to_page(range, range->pfns[i]);
1247                 if (page == NULL)
1248                         continue;
1249
1250                 if (dma_mapping_error(device, daddrs[i]))
1251                         continue;
1252
1253                 /* If it is read and write than map bi-directional. */
1254                 if (range->pfns[i] & range->flags[HMM_PFN_WRITE])
1255                         dir = DMA_BIDIRECTIONAL;
1256
1257                 dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir);
1258                 mapped--;
1259         }
1260
1261         return ret;
1262 }
1263 EXPORT_SYMBOL(hmm_range_dma_map);
1264
1265 /**
1266  * hmm_range_dma_unmap() - unmap range of that was map with hmm_range_dma_map()
1267  * @range: range being unmapped
1268  * @vma: the vma against which the range (optional)
1269  * @device: device against which dma map was done
1270  * @daddrs: dma address of mapped pages
1271  * @dirty: dirty page if it had the write flag set
1272  * Return: number of page unmapped on success, -EINVAL otherwise
1273  *
1274  * Note that caller MUST abide by mmu notifier or use HMM mirror and abide
1275  * to the sync_cpu_device_pagetables() callback so that it is safe here to
1276  * call set_page_dirty(). Caller must also take appropriate locks to avoid
1277  * concurrent mmu notifier or sync_cpu_device_pagetables() to make progress.
1278  */
1279 long hmm_range_dma_unmap(struct hmm_range *range,
1280                          struct vm_area_struct *vma,
1281                          struct device *device,
1282                          dma_addr_t *daddrs,
1283                          bool dirty)
1284 {
1285         unsigned long i, npages;
1286         long cpages = 0;
1287
1288         /* Sanity check. */
1289         if (range->end <= range->start)
1290                 return -EINVAL;
1291         if (!daddrs)
1292                 return -EINVAL;
1293         if (!range->pfns)
1294                 return -EINVAL;
1295
1296         npages = (range->end - range->start) >> PAGE_SHIFT;
1297         for (i = 0; i < npages; ++i) {
1298                 enum dma_data_direction dir = DMA_TO_DEVICE;
1299                 struct page *page;
1300
1301                 page = hmm_device_entry_to_page(range, range->pfns[i]);
1302                 if (page == NULL)
1303                         continue;
1304
1305                 /* If it is read and write than map bi-directional. */
1306                 if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) {
1307                         dir = DMA_BIDIRECTIONAL;
1308
1309                         /*
1310                          * See comments in function description on why it is
1311                          * safe here to call set_page_dirty()
1312                          */
1313                         if (dirty)
1314                                 set_page_dirty(page);
1315                 }
1316
1317                 /* Unmap and clear pfns/dma address */
1318                 dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir);
1319                 range->pfns[i] = range->values[HMM_PFN_NONE];
1320                 /* FIXME see comments in hmm_vma_dma_map() */
1321                 daddrs[i] = 0;
1322                 cpages++;
1323         }
1324
1325         return cpages;
1326 }
1327 EXPORT_SYMBOL(hmm_range_dma_unmap);
1328 #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
1329
1330
1331 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) ||  IS_ENABLED(CONFIG_DEVICE_PUBLIC)
1332 struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
1333                                        unsigned long addr)
1334 {
1335         struct page *page;
1336
1337         page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
1338         if (!page)
1339                 return NULL;
1340         lock_page(page);
1341         return page;
1342 }
1343 EXPORT_SYMBOL(hmm_vma_alloc_locked_page);
1344
1345
1346 static void hmm_devmem_ref_release(struct percpu_ref *ref)
1347 {
1348         struct hmm_devmem *devmem;
1349
1350         devmem = container_of(ref, struct hmm_devmem, ref);
1351         complete(&devmem->completion);
1352 }
1353
1354 static void hmm_devmem_ref_exit(void *data)
1355 {
1356         struct percpu_ref *ref = data;
1357         struct hmm_devmem *devmem;
1358
1359         devmem = container_of(ref, struct hmm_devmem, ref);
1360         wait_for_completion(&devmem->completion);
1361         percpu_ref_exit(ref);
1362 }
1363
1364 static void hmm_devmem_ref_kill(struct percpu_ref *ref)
1365 {
1366         percpu_ref_kill(ref);
1367 }
1368
1369 static vm_fault_t hmm_devmem_fault(struct vm_area_struct *vma,
1370                             unsigned long addr,
1371                             const struct page *page,
1372                             unsigned int flags,
1373                             pmd_t *pmdp)
1374 {
1375         struct hmm_devmem *devmem = page->pgmap->data;
1376
1377         return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp);
1378 }
1379
1380 static void hmm_devmem_free(struct page *page, void *data)
1381 {
1382         struct hmm_devmem *devmem = data;
1383
1384         page->mapping = NULL;
1385
1386         devmem->ops->free(devmem, page);
1387 }
1388
1389 /*
1390  * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory
1391  *
1392  * @ops: memory event device driver callback (see struct hmm_devmem_ops)
1393  * @device: device struct to bind the resource too
1394  * @size: size in bytes of the device memory to add
1395  * Return: pointer to new hmm_devmem struct ERR_PTR otherwise
1396  *
1397  * This function first finds an empty range of physical address big enough to
1398  * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which
1399  * in turn allocates struct pages. It does not do anything beyond that; all
1400  * events affecting the memory will go through the various callbacks provided
1401  * by hmm_devmem_ops struct.
1402  *
1403  * Device driver should call this function during device initialization and
1404  * is then responsible of memory management. HMM only provides helpers.
1405  */
1406 struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
1407                                   struct device *device,
1408                                   unsigned long size)
1409 {
1410         struct hmm_devmem *devmem;
1411         resource_size_t addr;
1412         void *result;
1413         int ret;
1414
1415         dev_pagemap_get_ops();
1416
1417         devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL);
1418         if (!devmem)
1419                 return ERR_PTR(-ENOMEM);
1420
1421         init_completion(&devmem->completion);
1422         devmem->pfn_first = -1UL;
1423         devmem->pfn_last = -1UL;
1424         devmem->resource = NULL;
1425         devmem->device = device;
1426         devmem->ops = ops;
1427
1428         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1429                               0, GFP_KERNEL);
1430         if (ret)
1431                 return ERR_PTR(ret);
1432
1433         ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit, &devmem->ref);
1434         if (ret)
1435                 return ERR_PTR(ret);
1436
1437         size = ALIGN(size, PA_SECTION_SIZE);
1438         addr = min((unsigned long)iomem_resource.end,
1439                    (1UL << MAX_PHYSMEM_BITS) - 1);
1440         addr = addr - size + 1UL;
1441
1442         /*
1443          * FIXME add a new helper to quickly walk resource tree and find free
1444          * range
1445          *
1446          * FIXME what about ioport_resource resource ?
1447          */
1448         for (; addr > size && addr >= iomem_resource.start; addr -= size) {
1449                 ret = region_intersects(addr, size, 0, IORES_DESC_NONE);
1450                 if (ret != REGION_DISJOINT)
1451                         continue;
1452
1453                 devmem->resource = devm_request_mem_region(device, addr, size,
1454                                                            dev_name(device));
1455                 if (!devmem->resource)
1456                         return ERR_PTR(-ENOMEM);
1457                 break;
1458         }
1459         if (!devmem->resource)
1460                 return ERR_PTR(-ERANGE);
1461
1462         devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY;
1463         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1464         devmem->pfn_last = devmem->pfn_first +
1465                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1466         devmem->page_fault = hmm_devmem_fault;
1467
1468         devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
1469         devmem->pagemap.res = *devmem->resource;
1470         devmem->pagemap.page_free = hmm_devmem_free;
1471         devmem->pagemap.altmap_valid = false;
1472         devmem->pagemap.ref = &devmem->ref;
1473         devmem->pagemap.data = devmem;
1474         devmem->pagemap.kill = hmm_devmem_ref_kill;
1475
1476         result = devm_memremap_pages(devmem->device, &devmem->pagemap);
1477         if (IS_ERR(result))
1478                 return result;
1479         return devmem;
1480 }
1481 EXPORT_SYMBOL_GPL(hmm_devmem_add);
1482
1483 struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops,
1484                                            struct device *device,
1485                                            struct resource *res)
1486 {
1487         struct hmm_devmem *devmem;
1488         void *result;
1489         int ret;
1490
1491         if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY)
1492                 return ERR_PTR(-EINVAL);
1493
1494         dev_pagemap_get_ops();
1495
1496         devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL);
1497         if (!devmem)
1498                 return ERR_PTR(-ENOMEM);
1499
1500         init_completion(&devmem->completion);
1501         devmem->pfn_first = -1UL;
1502         devmem->pfn_last = -1UL;
1503         devmem->resource = res;
1504         devmem->device = device;
1505         devmem->ops = ops;
1506
1507         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1508                               0, GFP_KERNEL);
1509         if (ret)
1510                 return ERR_PTR(ret);
1511
1512         ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit,
1513                         &devmem->ref);
1514         if (ret)
1515                 return ERR_PTR(ret);
1516
1517         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1518         devmem->pfn_last = devmem->pfn_first +
1519                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1520         devmem->page_fault = hmm_devmem_fault;
1521
1522         devmem->pagemap.type = MEMORY_DEVICE_PUBLIC;
1523         devmem->pagemap.res = *devmem->resource;
1524         devmem->pagemap.page_free = hmm_devmem_free;
1525         devmem->pagemap.altmap_valid = false;
1526         devmem->pagemap.ref = &devmem->ref;
1527         devmem->pagemap.data = devmem;
1528         devmem->pagemap.kill = hmm_devmem_ref_kill;
1529
1530         result = devm_memremap_pages(devmem->device, &devmem->pagemap);
1531         if (IS_ERR(result))
1532                 return result;
1533         return devmem;
1534 }
1535 EXPORT_SYMBOL_GPL(hmm_devmem_add_resource);
1536
1537 /*
1538  * A device driver that wants to handle multiple devices memory through a
1539  * single fake device can use hmm_device to do so. This is purely a helper
1540  * and it is not needed to make use of any HMM functionality.
1541  */
1542 #define HMM_DEVICE_MAX 256
1543
1544 static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX);
1545 static DEFINE_SPINLOCK(hmm_device_lock);
1546 static struct class *hmm_device_class;
1547 static dev_t hmm_device_devt;
1548
1549 static void hmm_device_release(struct device *device)
1550 {
1551         struct hmm_device *hmm_device;
1552
1553         hmm_device = container_of(device, struct hmm_device, device);
1554         spin_lock(&hmm_device_lock);
1555         clear_bit(hmm_device->minor, hmm_device_mask);
1556         spin_unlock(&hmm_device_lock);
1557
1558         kfree(hmm_device);
1559 }
1560
1561 struct hmm_device *hmm_device_new(void *drvdata)
1562 {
1563         struct hmm_device *hmm_device;
1564
1565         hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL);
1566         if (!hmm_device)
1567                 return ERR_PTR(-ENOMEM);
1568
1569         spin_lock(&hmm_device_lock);
1570         hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX);
1571         if (hmm_device->minor >= HMM_DEVICE_MAX) {
1572                 spin_unlock(&hmm_device_lock);
1573                 kfree(hmm_device);
1574                 return ERR_PTR(-EBUSY);
1575         }
1576         set_bit(hmm_device->minor, hmm_device_mask);
1577         spin_unlock(&hmm_device_lock);
1578
1579         dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor);
1580         hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt),
1581                                         hmm_device->minor);
1582         hmm_device->device.release = hmm_device_release;
1583         dev_set_drvdata(&hmm_device->device, drvdata);
1584         hmm_device->device.class = hmm_device_class;
1585         device_initialize(&hmm_device->device);
1586
1587         return hmm_device;
1588 }
1589 EXPORT_SYMBOL(hmm_device_new);
1590
1591 void hmm_device_put(struct hmm_device *hmm_device)
1592 {
1593         put_device(&hmm_device->device);
1594 }
1595 EXPORT_SYMBOL(hmm_device_put);
1596
1597 static int __init hmm_init(void)
1598 {
1599         int ret;
1600
1601         ret = alloc_chrdev_region(&hmm_device_devt, 0,
1602                                   HMM_DEVICE_MAX,
1603                                   "hmm_device");
1604         if (ret)
1605                 return ret;
1606
1607         hmm_device_class = class_create(THIS_MODULE, "hmm_device");
1608         if (IS_ERR(hmm_device_class)) {
1609                 unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX);
1610                 return PTR_ERR(hmm_device_class);
1611         }
1612         return 0;
1613 }
1614
1615 device_initcall(hmm_init);
1616 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */