mm/hmm.c

   1 /*
   2  * Copyright 2013 Red Hat Inc.
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * Authors: Jérôme Glisse <jglisse@redhat.com>
  15  */
  16 /*
  17  * Refer to include/linux/hmm.h for information about heterogeneous memory
  18  * management or HMM for short.
  19  */
  20 #include <linux/mm.h>
  21 #include <linux/hmm.h>
  22 #include <linux/init.h>
  23 #include <linux/rmap.h>
  24 #include <linux/swap.h>
  25 #include <linux/slab.h>
  26 #include <linux/sched.h>
  27 #include <linux/mmzone.h>
  28 #include <linux/pagemap.h>
  29 #include <linux/swapops.h>
  30 #include <linux/hugetlb.h>
  31 #include <linux/memremap.h>
  32 #include <linux/jump_label.h>
  33 #include <linux/mmu_notifier.h>
  34 #include <linux/memory_hotplug.h>
  35
  36 #define PA_SECTION_SIZE (1UL << PA_SECTION_SHIFT)
  37
  38 #if IS_ENABLED(CONFIG_HMM_MIRROR)
  39 static const struct mmu_notifier_ops hmm_mmu_notifier_ops;
  40
  41 static inline struct hmm *mm_get_hmm(struct mm_struct *mm)
  42 {
  43         struct hmm *hmm = READ_ONCE(mm->hmm);
  44
  45         if (hmm && kref_get_unless_zero(&hmm->kref))
  46                 return hmm;
  47
  48         return NULL;
  49 }
  50
  51 /**
  52  * hmm_get_or_create - register HMM against an mm (HMM internal)
  53  *
  54  * @mm: mm struct to attach to
  55  * Returns: returns an HMM object, either by referencing the existing
  56  *          (per-process) object, or by creating a new one.
  57  *
  58  * This is not intended to be used directly by device drivers. If mm already
  59  * has an HMM struct then it get a reference on it and returns it. Otherwise
  60  * it allocates an HMM struct, initializes it, associate it with the mm and
  61  * returns it.
  62  */
  63 static struct hmm *hmm_get_or_create(struct mm_struct *mm)
  64 {
  65         struct hmm *hmm = mm_get_hmm(mm);
  66         bool cleanup = false;
  67
  68         if (hmm)
  69                 return hmm;
  70
  71         hmm = kmalloc(sizeof(*hmm), GFP_KERNEL);
  72         if (!hmm)
  73                 return NULL;
  74         init_waitqueue_head(&hmm->wq);
  75         INIT_LIST_HEAD(&hmm->mirrors);
  76         init_rwsem(&hmm->mirrors_sem);
  77         hmm->mmu_notifier.ops = NULL;
  78         INIT_LIST_HEAD(&hmm->ranges);
  79         mutex_init(&hmm->lock);
  80         kref_init(&hmm->kref);
  81         hmm->notifiers = 0;
  82         hmm->dead = false;
  83         hmm->mm = mm;
  84
  85         spin_lock(&mm->page_table_lock);
  86         if (!mm->hmm)
  87                 mm->hmm = hmm;
  88         else
  89                 cleanup = true;
  90         spin_unlock(&mm->page_table_lock);
  91
  92         if (cleanup)
  93                 goto error;
  94
  95         /*
  96          * We should only get here if hold the mmap_sem in write mode ie on
  97          * registration of first mirror through hmm_mirror_register()
  98          */
  99         hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops;
 100         if (__mmu_notifier_register(&hmm->mmu_notifier, mm))
 101                 goto error_mm;
 102
 103         return hmm;
 104
 105 error_mm:
 106         spin_lock(&mm->page_table_lock);
 107         if (mm->hmm == hmm)
 108                 mm->hmm = NULL;
 109         spin_unlock(&mm->page_table_lock);
 110 error:
 111         kfree(hmm);
 112         return NULL;
 113 }
 114
 115 static void hmm_free(struct kref *kref)
 116 {
 117         struct hmm *hmm = container_of(kref, struct hmm, kref);
 118         struct mm_struct *mm = hmm->mm;
 119
 120         mmu_notifier_unregister_no_release(&hmm->mmu_notifier, mm);
 121
 122         spin_lock(&mm->page_table_lock);
 123         if (mm->hmm == hmm)
 124                 mm->hmm = NULL;
 125         spin_unlock(&mm->page_table_lock);
 126
 127         kfree(hmm);
 128 }
 129
 130 static inline void hmm_put(struct hmm *hmm)
 131 {
 132         kref_put(&hmm->kref, hmm_free);
 133 }
 134
 135 void hmm_mm_destroy(struct mm_struct *mm)
 136 {
 137         struct hmm *hmm;
 138
 139         spin_lock(&mm->page_table_lock);
 140         hmm = mm_get_hmm(mm);
 141         mm->hmm = NULL;
 142         if (hmm) {
 143                 hmm->mm = NULL;
 144                 hmm->dead = true;
 145                 spin_unlock(&mm->page_table_lock);
 146                 hmm_put(hmm);
 147                 return;
 148         }
 149
 150         spin_unlock(&mm->page_table_lock);
 151 }
 152
 153 static void hmm_release(struct mmu_notifier *mn, struct mm_struct *mm)
 154 {
 155         struct hmm *hmm = mm_get_hmm(mm);
 156         struct hmm_mirror *mirror;
 157         struct hmm_range *range;
 158
 159         /* Report this HMM as dying. */
 160         hmm->dead = true;
 161
 162         /* Wake-up everyone waiting on any range. */
 163         mutex_lock(&hmm->lock);
 164         list_for_each_entry(range, &hmm->ranges, list) {
 165                 range->valid = false;
 166         }
 167         wake_up_all(&hmm->wq);
 168         mutex_unlock(&hmm->lock);
 169
 170         down_write(&hmm->mirrors_sem);
 171         mirror = list_first_entry_or_null(&hmm->mirrors, struct hmm_mirror,
 172                                           list);
 173         while (mirror) {
 174                 list_del_init(&mirror->list);
 175                 if (mirror->ops->release) {
 176                         /*
 177                          * Drop mirrors_sem so callback can wait on any pending
 178                          * work that might itself trigger mmu_notifier callback
 179                          * and thus would deadlock with us.
 180                          */
 181                         up_write(&hmm->mirrors_sem);
 182                         mirror->ops->release(mirror);
 183                         down_write(&hmm->mirrors_sem);
 184                 }
 185                 mirror = list_first_entry_or_null(&hmm->mirrors,
 186                                                   struct hmm_mirror, list);
 187         }
 188         up_write(&hmm->mirrors_sem);
 189
 190         hmm_put(hmm);
 191 }
 192
 193 static int hmm_invalidate_range_start(struct mmu_notifier *mn,
 194                         const struct mmu_notifier_range *nrange)
 195 {
 196         struct hmm *hmm = mm_get_hmm(nrange->mm);
 197         struct hmm_mirror *mirror;
 198         struct hmm_update update;
 199         struct hmm_range *range;
 200         int ret = 0;
 201
 202         VM_BUG_ON(!hmm);
 203
 204         update.start = nrange->start;
 205         update.end = nrange->end;
 206         update.event = HMM_UPDATE_INVALIDATE;
 207         update.blockable = nrange->blockable;
 208
 209         if (nrange->blockable)
 210                 mutex_lock(&hmm->lock);
 211         else if (!mutex_trylock(&hmm->lock)) {
 212                 ret = -EAGAIN;
 213                 goto out;
 214         }
 215         hmm->notifiers++;
 216         list_for_each_entry(range, &hmm->ranges, list) {
 217                 if (update.end < range->start || update.start >= range->end)
 218                         continue;
 219
 220                 range->valid = false;
 221         }
 222         mutex_unlock(&hmm->lock);
 223
 224         if (nrange->blockable)
 225                 down_read(&hmm->mirrors_sem);
 226         else if (!down_read_trylock(&hmm->mirrors_sem)) {
 227                 ret = -EAGAIN;
 228                 goto out;
 229         }
 230         list_for_each_entry(mirror, &hmm->mirrors, list) {
 231                 int ret;
 232
 233                 ret = mirror->ops->sync_cpu_device_pagetables(mirror, &update);
 234                 if (!update.blockable && ret == -EAGAIN) {
 235                         up_read(&hmm->mirrors_sem);
 236                         ret = -EAGAIN;
 237                         goto out;
 238                 }
 239         }
 240         up_read(&hmm->mirrors_sem);
 241
 242 out:
 243         hmm_put(hmm);
 244         return ret;
 245 }
 246
 247 static void hmm_invalidate_range_end(struct mmu_notifier *mn,
 248                         const struct mmu_notifier_range *nrange)
 249 {
 250         struct hmm *hmm = mm_get_hmm(nrange->mm);
 251
 252         VM_BUG_ON(!hmm);
 253
 254         mutex_lock(&hmm->lock);
 255         hmm->notifiers--;
 256         if (!hmm->notifiers) {
 257                 struct hmm_range *range;
 258
 259                 list_for_each_entry(range, &hmm->ranges, list) {
 260                         if (range->valid)
 261                                 continue;
 262                         range->valid = true;
 263                 }
 264                 wake_up_all(&hmm->wq);
 265         }
 266         mutex_unlock(&hmm->lock);
 267
 268         hmm_put(hmm);
 269 }
 270
 271 static const struct mmu_notifier_ops hmm_mmu_notifier_ops = {
 272         .release                = hmm_release,
 273         .invalidate_range_start = hmm_invalidate_range_start,
 274         .invalidate_range_end   = hmm_invalidate_range_end,
 275 };
 276
 277 /*
 278  * hmm_mirror_register() - register a mirror against an mm
 279  *
 280  * @mirror: new mirror struct to register
 281  * @mm: mm to register against
 282  *
 283  * To start mirroring a process address space, the device driver must register
 284  * an HMM mirror struct.
 285  *
 286  * THE mm->mmap_sem MUST BE HELD IN WRITE MODE !
 287  */
 288 int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm)
 289 {
 290         /* Sanity check */
 291         if (!mm || !mirror || !mirror->ops)
 292                 return -EINVAL;
 293
 294         mirror->hmm = hmm_get_or_create(mm);
 295         if (!mirror->hmm)
 296                 return -ENOMEM;
 297
 298         down_write(&mirror->hmm->mirrors_sem);
 299         list_add(&mirror->list, &mirror->hmm->mirrors);
 300         up_write(&mirror->hmm->mirrors_sem);
 301
 302         return 0;
 303 }
 304 EXPORT_SYMBOL(hmm_mirror_register);
 305
 306 /*
 307  * hmm_mirror_unregister() - unregister a mirror
 308  *
 309  * @mirror: new mirror struct to register
 310  *
 311  * Stop mirroring a process address space, and cleanup.
 312  */
 313 void hmm_mirror_unregister(struct hmm_mirror *mirror)
 314 {
 315         struct hmm *hmm = READ_ONCE(mirror->hmm);
 316
 317         if (hmm == NULL)
 318                 return;
 319
 320         down_write(&hmm->mirrors_sem);
 321         list_del_init(&mirror->list);
 322         /* To protect us against double unregister ... */
 323         mirror->hmm = NULL;
 324         up_write(&hmm->mirrors_sem);
 325
 326         hmm_put(hmm);
 327 }
 328 EXPORT_SYMBOL(hmm_mirror_unregister);
 329
 330 struct hmm_vma_walk {
 331         struct hmm_range        *range;
 332         unsigned long           last;
 333         bool                    fault;
 334         bool                    block;
 335 };
 336
 337 static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr,
 338                             bool write_fault, uint64_t *pfn)
 339 {
 340         unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE;
 341         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 342         struct hmm_range *range = hmm_vma_walk->range;
 343         struct vm_area_struct *vma = walk->vma;
 344         vm_fault_t ret;
 345
 346         flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY;
 347         flags |= write_fault ? FAULT_FLAG_WRITE : 0;
 348         ret = handle_mm_fault(vma, addr, flags);
 349         if (ret & VM_FAULT_RETRY)
 350                 return -EAGAIN;
 351         if (ret & VM_FAULT_ERROR) {
 352                 *pfn = range->values[HMM_PFN_ERROR];
 353                 return -EFAULT;
 354         }
 355
 356         return -EBUSY;
 357 }
 358
 359 static int hmm_pfns_bad(unsigned long addr,
 360                         unsigned long end,
 361                         struct mm_walk *walk)
 362 {
 363         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 364         struct hmm_range *range = hmm_vma_walk->range;
 365         uint64_t *pfns = range->pfns;
 366         unsigned long i;
 367
 368         i = (addr - range->start) >> PAGE_SHIFT;
 369         for (; addr < end; addr += PAGE_SIZE, i++)
 370                 pfns[i] = range->values[HMM_PFN_ERROR];
 371
 372         return 0;
 373 }
 374
 375 /*
 376  * hmm_vma_walk_hole() - handle a range lacking valid pmd or pte(s)
 377  * @start: range virtual start address (inclusive)
 378  * @end: range virtual end address (exclusive)
 379  * @fault: should we fault or not ?
 380  * @write_fault: write fault ?
 381  * @walk: mm_walk structure
 382  * Returns: 0 on success, -EBUSY after page fault, or page fault error
 383  *
 384  * This function will be called whenever pmd_none() or pte_none() returns true,
 385  * or whenever there is no page directory covering the virtual address range.
 386  */
 387 static int hmm_vma_walk_hole_(unsigned long addr, unsigned long end,
 388                               bool fault, bool write_fault,
 389                               struct mm_walk *walk)
 390 {
 391         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 392         struct hmm_range *range = hmm_vma_walk->range;
 393         uint64_t *pfns = range->pfns;
 394         unsigned long i, page_size;
 395
 396         hmm_vma_walk->last = addr;
 397         page_size = hmm_range_page_size(range);
 398         i = (addr - range->start) >> range->page_shift;
 399
 400         for (; addr < end; addr += page_size, i++) {
 401                 pfns[i] = range->values[HMM_PFN_NONE];
 402                 if (fault || write_fault) {
 403                         int ret;
 404
 405                         ret = hmm_vma_do_fault(walk, addr, write_fault,
 406                                                &pfns[i]);
 407                         if (ret != -EBUSY)
 408                                 return ret;
 409                 }
 410         }
 411
 412         return (fault || write_fault) ? -EBUSY : 0;
 413 }
 414
 415 static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 416                                       uint64_t pfns, uint64_t cpu_flags,
 417                                       bool *fault, bool *write_fault)
 418 {
 419         struct hmm_range *range = hmm_vma_walk->range;
 420
 421         if (!hmm_vma_walk->fault)
 422                 return;
 423
 424         /*
 425          * So we not only consider the individual per page request we also
 426          * consider the default flags requested for the range. The API can
 427          * be use in 2 fashions. The first one where the HMM user coalesce
 428          * multiple page fault into one request and set flags per pfns for
 429          * of those faults. The second one where the HMM user want to pre-
 430          * fault a range with specific flags. For the latter one it is a
 431          * waste to have the user pre-fill the pfn arrays with a default
 432          * flags value.
 433          */
 434         pfns = (pfns & range->pfn_flags_mask) | range->default_flags;
 435
 436         /* We aren't ask to do anything ... */
 437         if (!(pfns & range->flags[HMM_PFN_VALID]))
 438                 return;
 439         /* If this is device memory than only fault if explicitly requested */
 440         if ((cpu_flags & range->flags[HMM_PFN_DEVICE_PRIVATE])) {
 441                 /* Do we fault on device memory ? */
 442                 if (pfns & range->flags[HMM_PFN_DEVICE_PRIVATE]) {
 443                         *write_fault = pfns & range->flags[HMM_PFN_WRITE];
 444                         *fault = true;
 445                 }
 446                 return;
 447         }
 448
 449         /* If CPU page table is not valid then we need to fault */
 450         *fault = !(cpu_flags & range->flags[HMM_PFN_VALID]);
 451         /* Need to write fault ? */
 452         if ((pfns & range->flags[HMM_PFN_WRITE]) &&
 453             !(cpu_flags & range->flags[HMM_PFN_WRITE])) {
 454                 *write_fault = true;
 455                 *fault = true;
 456         }
 457 }
 458
 459 static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 460                                  const uint64_t *pfns, unsigned long npages,
 461                                  uint64_t cpu_flags, bool *fault,
 462                                  bool *write_fault)
 463 {
 464         unsigned long i;
 465
 466         if (!hmm_vma_walk->fault) {
 467                 *fault = *write_fault = false;
 468                 return;
 469         }
 470
 471         *fault = *write_fault = false;
 472         for (i = 0; i < npages; ++i) {
 473                 hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags,
 474                                    fault, write_fault);
 475                 if ((*write_fault))
 476                         return;
 477         }
 478 }
 479
 480 static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
 481                              struct mm_walk *walk)
 482 {
 483         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 484         struct hmm_range *range = hmm_vma_walk->range;
 485         bool fault, write_fault;
 486         unsigned long i, npages;
 487         uint64_t *pfns;
 488
 489         i = (addr - range->start) >> PAGE_SHIFT;
 490         npages = (end - addr) >> PAGE_SHIFT;
 491         pfns = &range->pfns[i];
 492         hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 493                              0, &fault, &write_fault);
 494         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 495 }
 496
 497 static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd)
 498 {
 499         if (pmd_protnone(pmd))
 500                 return 0;
 501         return pmd_write(pmd) ? 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         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 513         struct hmm_range *range = hmm_vma_walk->range;
 514         unsigned long pfn, npages, i;
 515         bool fault, write_fault;
 516         uint64_t cpu_flags;
 517
 518         npages = (end - addr) >> PAGE_SHIFT;
 519         cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
 520         hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags,
 521                              &fault, &write_fault);
 522
 523         if (pmd_protnone(pmd) || fault || write_fault)
 524                 return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 525
 526         pfn = pmd_pfn(pmd) + pte_index(addr);
 527         for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++)
 528                 pfns[i] = hmm_pfn_from_pfn(range, pfn) | cpu_flags;
 529         hmm_vma_walk->last = end;
 530         return 0;
 531 }
 532
 533 static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte)
 534 {
 535         if (pte_none(pte) || !pte_present(pte))
 536                 return 0;
 537         return pte_write(pte) ? range->flags[HMM_PFN_VALID] |
 538                                 range->flags[HMM_PFN_WRITE] :
 539                                 range->flags[HMM_PFN_VALID];
 540 }
 541
 542 static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
 543                               unsigned long end, pmd_t *pmdp, pte_t *ptep,
 544                               uint64_t *pfn)
 545 {
 546         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 547         struct hmm_range *range = hmm_vma_walk->range;
 548         struct vm_area_struct *vma = walk->vma;
 549         bool fault, write_fault;
 550         uint64_t cpu_flags;
 551         pte_t pte = *ptep;
 552         uint64_t orig_pfn = *pfn;
 553
 554         *pfn = range->values[HMM_PFN_NONE];
 555         fault = write_fault = false;
 556
 557         if (pte_none(pte)) {
 558                 hmm_pte_need_fault(hmm_vma_walk, orig_pfn, 0,
 559                                    &fault, &write_fault);
 560                 if (fault || write_fault)
 561                         goto fault;
 562                 return 0;
 563         }
 564
 565         if (!pte_present(pte)) {
 566                 swp_entry_t entry = pte_to_swp_entry(pte);
 567
 568                 if (!non_swap_entry(entry)) {
 569                         if (fault || write_fault)
 570                                 goto fault;
 571                         return 0;
 572                 }
 573
 574                 /*
 575                  * This is a special swap entry, ignore migration, use
 576                  * device and report anything else as error.
 577                  */
 578                 if (is_device_private_entry(entry)) {
 579                         cpu_flags = range->flags[HMM_PFN_VALID] |
 580                                 range->flags[HMM_PFN_DEVICE_PRIVATE];
 581                         cpu_flags |= is_write_device_private_entry(entry) ?
 582                                 range->flags[HMM_PFN_WRITE] : 0;
 583                         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 584                                            &fault, &write_fault);
 585                         if (fault || write_fault)
 586                                 goto fault;
 587                         *pfn = hmm_pfn_from_pfn(range, swp_offset(entry));
 588                         *pfn |= cpu_flags;
 589                         return 0;
 590                 }
 591
 592                 if (is_migration_entry(entry)) {
 593                         if (fault || write_fault) {
 594                                 pte_unmap(ptep);
 595                                 hmm_vma_walk->last = addr;
 596                                 migration_entry_wait(vma->vm_mm,
 597                                                      pmdp, addr);
 598                                 return -EBUSY;
 599                         }
 600                         return 0;
 601                 }
 602
 603                 /* Report error for everything else */
 604                 *pfn = range->values[HMM_PFN_ERROR];
 605                 return -EFAULT;
 606         } else {
 607                 cpu_flags = pte_to_hmm_pfn_flags(range, pte);
 608                 hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 609                                    &fault, &write_fault);
 610         }
 611
 612         if (fault || write_fault)
 613                 goto fault;
 614
 615         *pfn = hmm_pfn_from_pfn(range, pte_pfn(pte)) | cpu_flags;
 616         return 0;
 617
 618 fault:
 619         pte_unmap(ptep);
 620         /* Fault any virtual address we were asked to fault */
 621         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 622 }
 623
 624 static int hmm_vma_walk_pmd(pmd_t *pmdp,
 625                             unsigned long start,
 626                             unsigned long end,
 627                             struct mm_walk *walk)
 628 {
 629         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 630         struct hmm_range *range = hmm_vma_walk->range;
 631         struct vm_area_struct *vma = walk->vma;
 632         uint64_t *pfns = range->pfns;
 633         unsigned long addr = start, i;
 634         pte_t *ptep;
 635         pmd_t pmd;
 636
 637
 638 again:
 639         pmd = READ_ONCE(*pmdp);
 640         if (pmd_none(pmd))
 641                 return hmm_vma_walk_hole(start, end, walk);
 642
 643         if (pmd_huge(pmd) && (range->vma->vm_flags & VM_HUGETLB))
 644                 return hmm_pfns_bad(start, end, walk);
 645
 646         if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
 647                 bool fault, write_fault;
 648                 unsigned long npages;
 649                 uint64_t *pfns;
 650
 651                 i = (addr - range->start) >> PAGE_SHIFT;
 652                 npages = (end - addr) >> PAGE_SHIFT;
 653                 pfns = &range->pfns[i];
 654
 655                 hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 656                                      0, &fault, &write_fault);
 657                 if (fault || write_fault) {
 658                         hmm_vma_walk->last = addr;
 659                         pmd_migration_entry_wait(vma->vm_mm, pmdp);
 660                         return -EBUSY;
 661                 }
 662                 return 0;
 663         } else if (!pmd_present(pmd))
 664                 return hmm_pfns_bad(start, end, walk);
 665
 666         if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
 667                 /*
 668                  * No need to take pmd_lock here, even if some other threads
 669                  * is splitting the huge pmd we will get that event through
 670                  * mmu_notifier callback.
 671                  *
 672                  * So just read pmd value and check again its a transparent
 673                  * huge or device mapping one and compute corresponding pfn
 674                  * values.
 675                  */
 676                 pmd = pmd_read_atomic(pmdp);
 677                 barrier();
 678                 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
 679                         goto again;
 680
 681                 i = (addr - range->start) >> PAGE_SHIFT;
 682                 return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd);
 683         }
 684
 685         /*
 686          * We have handled all the valid case above ie either none, migration,
 687          * huge or transparent huge. At this point either it is a valid pmd
 688          * entry pointing to pte directory or it is a bad pmd that will not
 689          * recover.
 690          */
 691         if (pmd_bad(pmd))
 692                 return hmm_pfns_bad(start, end, walk);
 693
 694         ptep = pte_offset_map(pmdp, addr);
 695         i = (addr - range->start) >> PAGE_SHIFT;
 696         for (; addr < end; addr += PAGE_SIZE, ptep++, i++) {
 697                 int r;
 698
 699                 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]);
 700                 if (r) {
 701                         /* hmm_vma_handle_pte() did unmap pte directory */
 702                         hmm_vma_walk->last = addr;
 703                         return r;
 704                 }
 705         }
 706         pte_unmap(ptep - 1);
 707
 708         hmm_vma_walk->last = addr;
 709         return 0;
 710 }
 711
 712 static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask,
 713                                       unsigned long start, unsigned long end,
 714                                       struct mm_walk *walk)
 715 {
 716 #ifdef CONFIG_HUGETLB_PAGE
 717         unsigned long addr = start, i, pfn, mask, size, pfn_inc;
 718         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 719         struct hmm_range *range = hmm_vma_walk->range;
 720         struct vm_area_struct *vma = walk->vma;
 721         struct hstate *h = hstate_vma(vma);
 722         uint64_t orig_pfn, cpu_flags;
 723         bool fault, write_fault;
 724         spinlock_t *ptl;
 725         pte_t entry;
 726         int ret = 0;
 727
 728         size = 1UL << huge_page_shift(h);
 729         mask = size - 1;
 730         if (range->page_shift != PAGE_SHIFT) {
 731                 /* Make sure we are looking at full page. */
 732                 if (start & mask)
 733                         return -EINVAL;
 734                 if (end < (start + size))
 735                         return -EINVAL;
 736                 pfn_inc = size >> PAGE_SHIFT;
 737         } else {
 738                 pfn_inc = 1;
 739                 size = PAGE_SIZE;
 740         }
 741
 742
 743         ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
 744         entry = huge_ptep_get(pte);
 745
 746         i = (start - range->start) >> range->page_shift;
 747         orig_pfn = range->pfns[i];
 748         range->pfns[i] = range->values[HMM_PFN_NONE];
 749         cpu_flags = pte_to_hmm_pfn_flags(range, entry);
 750         fault = write_fault = false;
 751         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 752                            &fault, &write_fault);
 753         if (fault || write_fault) {
 754                 ret = -ENOENT;
 755                 goto unlock;
 756         }
 757
 758         pfn = pte_pfn(entry) + ((start & mask) >> range->page_shift);
 759         for (; addr < end; addr += size, i++, pfn += pfn_inc)
 760                 range->pfns[i] = hmm_pfn_from_pfn(range, pfn) | cpu_flags;
 761         hmm_vma_walk->last = end;
 762
 763 unlock:
 764         spin_unlock(ptl);
 765
 766         if (ret == -ENOENT)
 767                 return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 768
 769         return ret;
 770 #else /* CONFIG_HUGETLB_PAGE */
 771         return -EINVAL;
 772 #endif
 773 }
 774
 775 static void hmm_pfns_clear(struct hmm_range *range,
 776                            uint64_t *pfns,
 777                            unsigned long addr,
 778                            unsigned long end)
 779 {
 780         for (; addr < end; addr += PAGE_SIZE, pfns++)
 781                 *pfns = range->values[HMM_PFN_NONE];
 782 }
 783
 784 static void hmm_pfns_special(struct hmm_range *range)
 785 {
 786         unsigned long addr = range->start, i = 0;
 787
 788         for (; addr < range->end; addr += PAGE_SIZE, i++)
 789                 range->pfns[i] = range->values[HMM_PFN_SPECIAL];
 790 }
 791
 792 /*
 793  * hmm_range_register() - start tracking change to CPU page table over a range
 794  * @range: range
 795  * @mm: the mm struct for the range of virtual address
 796  * @start: start virtual address (inclusive)
 797  * @end: end virtual address (exclusive)
 798  * @page_shift: expect page shift for the range
 799  * Returns 0 on success, -EFAULT if the address space is no longer valid
 800  *
 801  * Track updates to the CPU page table see include/linux/hmm.h
 802  */
 803 int hmm_range_register(struct hmm_range *range,
 804                        struct mm_struct *mm,
 805                        unsigned long start,
 806                        unsigned long end,
 807                        unsigned page_shift)
 808 {
 809         unsigned long mask = ((1UL << page_shift) - 1UL);
 810
 811         range->valid = false;
 812         range->hmm = NULL;
 813
 814         if ((start & mask) || (end & mask))
 815                 return -EINVAL;
 816         if (start >= end)
 817                 return -EINVAL;
 818
 819         range->page_shift = page_shift;
 820         range->start = start;
 821         range->end = end;
 822
 823         range->hmm = hmm_get_or_create(mm);
 824         if (!range->hmm)
 825                 return -EFAULT;
 826
 827         /* Check if hmm_mm_destroy() was call. */
 828         if (range->hmm->mm == NULL || range->hmm->dead) {
 829                 hmm_put(range->hmm);
 830                 return -EFAULT;
 831         }
 832
 833         /* Initialize range to track CPU page table update */
 834         mutex_lock(&range->hmm->lock);
 835
 836         list_add_rcu(&range->list, &range->hmm->ranges);
 837
 838         /*
 839          * If there are any concurrent notifiers we have to wait for them for
 840          * the range to be valid (see hmm_range_wait_until_valid()).
 841          */
 842         if (!range->hmm->notifiers)
 843                 range->valid = true;
 844         mutex_unlock(&range->hmm->lock);
 845
 846         return 0;
 847 }
 848 EXPORT_SYMBOL(hmm_range_register);
 849
 850 /*
 851  * hmm_range_unregister() - stop tracking change to CPU page table over a range
 852  * @range: range
 853  *
 854  * Range struct is used to track updates to the CPU page table after a call to
 855  * hmm_range_register(). See include/linux/hmm.h for how to use it.
 856  */
 857 void hmm_range_unregister(struct hmm_range *range)
 858 {
 859         /* Sanity check this really should not happen. */
 860         if (range->hmm == NULL || range->end <= range->start)
 861                 return;
 862
 863         mutex_lock(&range->hmm->lock);
 864         list_del_rcu(&range->list);
 865         mutex_unlock(&range->hmm->lock);
 866
 867         /* Drop reference taken by hmm_range_register() */
 868         range->valid = false;
 869         hmm_put(range->hmm);
 870         range->hmm = NULL;
 871 }
 872 EXPORT_SYMBOL(hmm_range_unregister);
 873
 874 /*
 875  * hmm_range_snapshot() - snapshot CPU page table for a range
 876  * @range: range
 877  * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid
 878  *          permission (for instance asking for write and range is read only),
 879  *          -EAGAIN if you need to retry, -EFAULT invalid (ie either no valid
 880  *          vma or it is illegal to access that range), number of valid pages
 881  *          in range->pfns[] (from range start address).
 882  *
 883  * This snapshots the CPU page table for a range of virtual addresses. Snapshot
 884  * validity is tracked by range struct. See in include/linux/hmm.h for example
 885  * on how to use.
 886  */
 887 long hmm_range_snapshot(struct hmm_range *range)
 888 {
 889         const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
 890         unsigned long start = range->start, end;
 891         struct hmm_vma_walk hmm_vma_walk;
 892         struct hmm *hmm = range->hmm;
 893         struct vm_area_struct *vma;
 894         struct mm_walk mm_walk;
 895
 896         /* Check if hmm_mm_destroy() was call. */
 897         if (hmm->mm == NULL || hmm->dead)
 898                 return -EFAULT;
 899
 900         do {
 901                 /* If range is no longer valid force retry. */
 902                 if (!range->valid)
 903                         return -EAGAIN;
 904
 905                 vma = find_vma(hmm->mm, start);
 906                 if (vma == NULL || (vma->vm_flags & device_vma))
 907                         return -EFAULT;
 908
 909                 /* FIXME support dax */
 910                 if (vma_is_dax(vma)) {
 911                         hmm_pfns_special(range);
 912                         return -EINVAL;
 913                 }
 914
 915                 if (is_vm_hugetlb_page(vma)) {
 916                         struct hstate *h = hstate_vma(vma);
 917
 918                         if (huge_page_shift(h) != range->page_shift &&
 919                             range->page_shift != PAGE_SHIFT)
 920                                 return -EINVAL;
 921                 } else {
 922                         if (range->page_shift != PAGE_SHIFT)
 923                                 return -EINVAL;
 924                 }
 925
 926                 if (!(vma->vm_flags & VM_READ)) {
 927                         /*
 928                          * If vma do not allow read access, then assume that it
 929                          * does not allow write access, either. HMM does not
 930                          * support architecture that allow write without read.
 931                          */
 932                         hmm_pfns_clear(range, range->pfns,
 933                                 range->start, range->end);
 934                         return -EPERM;
 935                 }
 936
 937                 range->vma = vma;
 938                 hmm_vma_walk.last = start;
 939                 hmm_vma_walk.fault = false;
 940                 hmm_vma_walk.range = range;
 941                 mm_walk.private = &hmm_vma_walk;
 942                 end = min(range->end, vma->vm_end);
 943
 944                 mm_walk.vma = vma;
 945                 mm_walk.mm = vma->vm_mm;
 946                 mm_walk.pte_entry = NULL;
 947                 mm_walk.test_walk = NULL;
 948                 mm_walk.hugetlb_entry = NULL;
 949                 mm_walk.pmd_entry = hmm_vma_walk_pmd;
 950                 mm_walk.pte_hole = hmm_vma_walk_hole;
 951                 mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry;
 952
 953                 walk_page_range(start, end, &mm_walk);
 954                 start = end;
 955         } while (start < range->end);
 956
 957         return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
 958 }
 959 EXPORT_SYMBOL(hmm_range_snapshot);
 960
 961 /*
 962  * hmm_range_fault() - try to fault some address in a virtual address range
 963  * @range: range being faulted
 964  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
 965  * Returns: number of valid pages in range->pfns[] (from range start
 966  *          address). This may be zero. If the return value is negative,
 967  *          then one of the following values may be returned:
 968  *
 969  *           -EINVAL  invalid arguments or mm or virtual address are in an
 970  *                    invalid vma (for instance device file vma).
 971  *           -ENOMEM: Out of memory.
 972  *           -EPERM:  Invalid permission (for instance asking for write and
 973  *                    range is read only).
 974  *           -EAGAIN: If you need to retry and mmap_sem was drop. This can only
 975  *                    happens if block argument is false.
 976  *           -EBUSY:  If the the range is being invalidated and you should wait
 977  *                    for invalidation to finish.
 978  *           -EFAULT: Invalid (ie either no valid vma or it is illegal to access
 979  *                    that range), number of valid pages in range->pfns[] (from
 980  *                    range start address).
 981  *
 982  * This is similar to a regular CPU page fault except that it will not trigger
 983  * any memory migration if the memory being faulted is not accessible by CPUs
 984  * and caller does not ask for migration.
 985  *
 986  * On error, for one virtual address in the range, the function will mark the
 987  * corresponding HMM pfn entry with an error flag.
 988  */
 989 long hmm_range_fault(struct hmm_range *range, bool block)
 990 {
 991         const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
 992         unsigned long start = range->start, end;
 993         struct hmm_vma_walk hmm_vma_walk;
 994         struct hmm *hmm = range->hmm;
 995         struct vm_area_struct *vma;
 996         struct mm_walk mm_walk;
 997         int ret;
 998
 999         /* Check if hmm_mm_destroy() was call. */
1000         if (hmm->mm == NULL || hmm->dead)
1001                 return -EFAULT;
1002
1003         do {
1004                 /* If range is no longer valid force retry. */
1005                 if (!range->valid) {
1006                         up_read(&hmm->mm->mmap_sem);
1007                         return -EAGAIN;
1008                 }
1009
1010                 vma = find_vma(hmm->mm, start);
1011                 if (vma == NULL || (vma->vm_flags & device_vma))
1012                         return -EFAULT;
1013
1014                 /* FIXME support dax */
1015                 if (vma_is_dax(vma)) {
1016                         hmm_pfns_special(range);
1017                         return -EINVAL;
1018                 }
1019
1020                 if (is_vm_hugetlb_page(vma)) {
1021                         if (huge_page_shift(hstate_vma(vma)) !=
1022                             range->page_shift &&
1023                             range->page_shift != PAGE_SHIFT)
1024                                 return -EINVAL;
1025                 } else {
1026                         if (range->page_shift != PAGE_SHIFT)
1027                                 return -EINVAL;
1028                 }
1029
1030                 if (!(vma->vm_flags & VM_READ)) {
1031                         /*
1032                          * If vma do not allow read access, then assume that it
1033                          * does not allow write access, either. HMM does not
1034                          * support architecture that allow write without read.
1035                          */
1036                         hmm_pfns_clear(range, range->pfns,
1037                                 range->start, range->end);
1038                         return -EPERM;
1039                 }
1040
1041                 range->vma = vma;
1042                 hmm_vma_walk.last = start;
1043                 hmm_vma_walk.fault = true;
1044                 hmm_vma_walk.block = block;
1045                 hmm_vma_walk.range = range;
1046                 mm_walk.private = &hmm_vma_walk;
1047                 end = min(range->end, vma->vm_end);
1048
1049                 mm_walk.vma = vma;
1050                 mm_walk.mm = vma->vm_mm;
1051                 mm_walk.pte_entry = NULL;
1052                 mm_walk.test_walk = NULL;
1053                 mm_walk.hugetlb_entry = NULL;
1054                 mm_walk.pmd_entry = hmm_vma_walk_pmd;
1055                 mm_walk.pte_hole = hmm_vma_walk_hole;
1056                 mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry;
1057
1058                 do {
1059                         ret = walk_page_range(start, end, &mm_walk);
1060                         start = hmm_vma_walk.last;
1061
1062                         /* Keep trying while the range is valid. */
1063                 } while (ret == -EBUSY && range->valid);
1064
1065                 if (ret) {
1066                         unsigned long i;
1067
1068                         i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
1069                         hmm_pfns_clear(range, &range->pfns[i],
1070                                 hmm_vma_walk.last, range->end);
1071                         return ret;
1072                 }
1073                 start = end;
1074
1075         } while (start < range->end);
1076
1077         return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
1078 }
1079 EXPORT_SYMBOL(hmm_range_fault);
1080 #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
1081
1082
1083 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) ||  IS_ENABLED(CONFIG_DEVICE_PUBLIC)
1084 struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
1085                                        unsigned long addr)
1086 {
1087         struct page *page;
1088
1089         page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
1090         if (!page)
1091                 return NULL;
1092         lock_page(page);
1093         return page;
1094 }
1095 EXPORT_SYMBOL(hmm_vma_alloc_locked_page);
1096
1097
1098 static void hmm_devmem_ref_release(struct percpu_ref *ref)
1099 {
1100         struct hmm_devmem *devmem;
1101
1102         devmem = container_of(ref, struct hmm_devmem, ref);
1103         complete(&devmem->completion);
1104 }
1105
1106 static void hmm_devmem_ref_exit(void *data)
1107 {
1108         struct percpu_ref *ref = data;
1109         struct hmm_devmem *devmem;
1110
1111         devmem = container_of(ref, struct hmm_devmem, ref);
1112         wait_for_completion(&devmem->completion);
1113         percpu_ref_exit(ref);
1114 }
1115
1116 static void hmm_devmem_ref_kill(struct percpu_ref *ref)
1117 {
1118         percpu_ref_kill(ref);
1119 }
1120
1121 static vm_fault_t hmm_devmem_fault(struct vm_area_struct *vma,
1122                             unsigned long addr,
1123                             const struct page *page,
1124                             unsigned int flags,
1125                             pmd_t *pmdp)
1126 {
1127         struct hmm_devmem *devmem = page->pgmap->data;
1128
1129         return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp);
1130 }
1131
1132 static void hmm_devmem_free(struct page *page, void *data)
1133 {
1134         struct hmm_devmem *devmem = data;
1135
1136         page->mapping = NULL;
1137
1138         devmem->ops->free(devmem, page);
1139 }
1140
1141 /*
1142  * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory
1143  *
1144  * @ops: memory event device driver callback (see struct hmm_devmem_ops)
1145  * @device: device struct to bind the resource too
1146  * @size: size in bytes of the device memory to add
1147  * Returns: pointer to new hmm_devmem struct ERR_PTR otherwise
1148  *
1149  * This function first finds an empty range of physical address big enough to
1150  * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which
1151  * in turn allocates struct pages. It does not do anything beyond that; all
1152  * events affecting the memory will go through the various callbacks provided
1153  * by hmm_devmem_ops struct.
1154  *
1155  * Device driver should call this function during device initialization and
1156  * is then responsible of memory management. HMM only provides helpers.
1157  */
1158 struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
1159                                   struct device *device,
1160                                   unsigned long size)
1161 {
1162         struct hmm_devmem *devmem;
1163         resource_size_t addr;
1164         void *result;
1165         int ret;
1166
1167         dev_pagemap_get_ops();
1168
1169         devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL);
1170         if (!devmem)
1171                 return ERR_PTR(-ENOMEM);
1172
1173         init_completion(&devmem->completion);
1174         devmem->pfn_first = -1UL;
1175         devmem->pfn_last = -1UL;
1176         devmem->resource = NULL;
1177         devmem->device = device;
1178         devmem->ops = ops;
1179
1180         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1181                               0, GFP_KERNEL);
1182         if (ret)
1183                 return ERR_PTR(ret);
1184
1185         ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit, &devmem->ref);
1186         if (ret)
1187                 return ERR_PTR(ret);
1188
1189         size = ALIGN(size, PA_SECTION_SIZE);
1190         addr = min((unsigned long)iomem_resource.end,
1191                    (1UL << MAX_PHYSMEM_BITS) - 1);
1192         addr = addr - size + 1UL;
1193
1194         /*
1195          * FIXME add a new helper to quickly walk resource tree and find free
1196          * range
1197          *
1198          * FIXME what about ioport_resource resource ?
1199          */
1200         for (; addr > size && addr >= iomem_resource.start; addr -= size) {
1201                 ret = region_intersects(addr, size, 0, IORES_DESC_NONE);
1202                 if (ret != REGION_DISJOINT)
1203                         continue;
1204
1205                 devmem->resource = devm_request_mem_region(device, addr, size,
1206                                                            dev_name(device));
1207                 if (!devmem->resource)
1208                         return ERR_PTR(-ENOMEM);
1209                 break;
1210         }
1211         if (!devmem->resource)
1212                 return ERR_PTR(-ERANGE);
1213
1214         devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY;
1215         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1216         devmem->pfn_last = devmem->pfn_first +
1217                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1218         devmem->page_fault = hmm_devmem_fault;
1219
1220         devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
1221         devmem->pagemap.res = *devmem->resource;
1222         devmem->pagemap.page_free = hmm_devmem_free;
1223         devmem->pagemap.altmap_valid = false;
1224         devmem->pagemap.ref = &devmem->ref;
1225         devmem->pagemap.data = devmem;
1226         devmem->pagemap.kill = hmm_devmem_ref_kill;
1227
1228         result = devm_memremap_pages(devmem->device, &devmem->pagemap);
1229         if (IS_ERR(result))
1230                 return result;
1231         return devmem;
1232 }
1233 EXPORT_SYMBOL_GPL(hmm_devmem_add);
1234
1235 struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops,
1236                                            struct device *device,
1237                                            struct resource *res)
1238 {
1239         struct hmm_devmem *devmem;
1240         void *result;
1241         int ret;
1242
1243         if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY)
1244                 return ERR_PTR(-EINVAL);
1245
1246         dev_pagemap_get_ops();
1247
1248         devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL);
1249         if (!devmem)
1250                 return ERR_PTR(-ENOMEM);
1251
1252         init_completion(&devmem->completion);
1253         devmem->pfn_first = -1UL;
1254         devmem->pfn_last = -1UL;
1255         devmem->resource = res;
1256         devmem->device = device;
1257         devmem->ops = ops;
1258
1259         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1260                               0, GFP_KERNEL);
1261         if (ret)
1262                 return ERR_PTR(ret);
1263
1264         ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit,
1265                         &devmem->ref);
1266         if (ret)
1267                 return ERR_PTR(ret);
1268
1269         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1270         devmem->pfn_last = devmem->pfn_first +
1271                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1272         devmem->page_fault = hmm_devmem_fault;
1273
1274         devmem->pagemap.type = MEMORY_DEVICE_PUBLIC;
1275         devmem->pagemap.res = *devmem->resource;
1276         devmem->pagemap.page_free = hmm_devmem_free;
1277         devmem->pagemap.altmap_valid = false;
1278         devmem->pagemap.ref = &devmem->ref;
1279         devmem->pagemap.data = devmem;
1280         devmem->pagemap.kill = hmm_devmem_ref_kill;
1281
1282         result = devm_memremap_pages(devmem->device, &devmem->pagemap);
1283         if (IS_ERR(result))
1284                 return result;
1285         return devmem;
1286 }
1287 EXPORT_SYMBOL_GPL(hmm_devmem_add_resource);
1288
1289 /*
1290  * A device driver that wants to handle multiple devices memory through a
1291  * single fake device can use hmm_device to do so. This is purely a helper
1292  * and it is not needed to make use of any HMM functionality.
1293  */
1294 #define HMM_DEVICE_MAX 256
1295
1296 static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX);
1297 static DEFINE_SPINLOCK(hmm_device_lock);
1298 static struct class *hmm_device_class;
1299 static dev_t hmm_device_devt;
1300
1301 static void hmm_device_release(struct device *device)
1302 {
1303         struct hmm_device *hmm_device;
1304
1305         hmm_device = container_of(device, struct hmm_device, device);
1306         spin_lock(&hmm_device_lock);
1307         clear_bit(hmm_device->minor, hmm_device_mask);
1308         spin_unlock(&hmm_device_lock);
1309
1310         kfree(hmm_device);
1311 }
1312
1313 struct hmm_device *hmm_device_new(void *drvdata)
1314 {
1315         struct hmm_device *hmm_device;
1316
1317         hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL);
1318         if (!hmm_device)
1319                 return ERR_PTR(-ENOMEM);
1320
1321         spin_lock(&hmm_device_lock);
1322         hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX);
1323         if (hmm_device->minor >= HMM_DEVICE_MAX) {
1324                 spin_unlock(&hmm_device_lock);
1325                 kfree(hmm_device);
1326                 return ERR_PTR(-EBUSY);
1327         }
1328         set_bit(hmm_device->minor, hmm_device_mask);
1329         spin_unlock(&hmm_device_lock);
1330
1331         dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor);
1332         hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt),
1333                                         hmm_device->minor);
1334         hmm_device->device.release = hmm_device_release;
1335         dev_set_drvdata(&hmm_device->device, drvdata);
1336         hmm_device->device.class = hmm_device_class;
1337         device_initialize(&hmm_device->device);
1338
1339         return hmm_device;
1340 }
1341 EXPORT_SYMBOL(hmm_device_new);
1342
1343 void hmm_device_put(struct hmm_device *hmm_device)
1344 {
1345         put_device(&hmm_device->device);
1346 }
1347 EXPORT_SYMBOL(hmm_device_put);
1348
1349 static int __init hmm_init(void)
1350 {
1351         int ret;
1352
1353         ret = alloc_chrdev_region(&hmm_device_devt, 0,
1354                                   HMM_DEVICE_MAX,
1355                                   "hmm_device");
1356         if (ret)
1357                 return ret;
1358
1359         hmm_device_class = class_create(THIS_MODULE, "hmm_device");
1360         if (IS_ERR(hmm_device_class)) {
1361                 unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX);
1362                 return PTR_ERR(hmm_device_class);
1363         }
1364         return 0;
1365 }
1366
1367 device_initcall(hmm_init);
1368 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */