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: JÃ©rÃ´me 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 defined(CONFIG_DEVICE_PRIVATE) || defined(CONFIG_DEVICE_PUBLIC)
  39 /*
  40  * Device private memory see HMM (Documentation/vm/hmm.txt) or hmm.h
  41  */
  42 DEFINE_STATIC_KEY_FALSE(device_private_key);
  43 EXPORT_SYMBOL(device_private_key);
  44 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */
  45
  46
  47 #if IS_ENABLED(CONFIG_HMM_MIRROR)
  48 static const struct mmu_notifier_ops hmm_mmu_notifier_ops;
  49
  50 /*
  51  * struct hmm - HMM per mm struct
  52  *
  53  * @mm: mm struct this HMM struct is bound to
  54  * @lock: lock protecting ranges list
  55  * @sequence: we track updates to the CPU page table with a sequence number
  56  * @ranges: list of range being snapshotted
  57  * @mirrors: list of mirrors for this mm
  58  * @mmu_notifier: mmu notifier to track updates to CPU page table
  59  * @mirrors_sem: read/write semaphore protecting the mirrors list
  60  */
  61 struct hmm {
  62         struct mm_struct        *mm;
  63         spinlock_t              lock;
  64         atomic_t                sequence;
  65         struct list_head        ranges;
  66         struct list_head        mirrors;
  67         struct mmu_notifier     mmu_notifier;
  68         struct rw_semaphore     mirrors_sem;
  69 };
  70
  71 /*
  72  * hmm_register - register HMM against an mm (HMM internal)
  73  *
  74  * @mm: mm struct to attach to
  75  *
  76  * This is not intended to be used directly by device drivers. It allocates an
  77  * HMM struct if mm does not have one, and initializes it.
  78  */
  79 static struct hmm *hmm_register(struct mm_struct *mm)
  80 {
  81         struct hmm *hmm = READ_ONCE(mm->hmm);
  82         bool cleanup = false;
  83
  84         /*
  85          * The hmm struct can only be freed once the mm_struct goes away,
  86          * hence we should always have pre-allocated an new hmm struct
  87          * above.
  88          */
  89         if (hmm)
  90                 return hmm;
  91
  92         hmm = kmalloc(sizeof(*hmm), GFP_KERNEL);
  93         if (!hmm)
  94                 return NULL;
  95         INIT_LIST_HEAD(&hmm->mirrors);
  96         init_rwsem(&hmm->mirrors_sem);
  97         atomic_set(&hmm->sequence, 0);
  98         hmm->mmu_notifier.ops = NULL;
  99         INIT_LIST_HEAD(&hmm->ranges);
 100         spin_lock_init(&hmm->lock);
 101         hmm->mm = mm;
 102
 103         /*
 104          * We should only get here if hold the mmap_sem in write mode ie on
 105          * registration of first mirror through hmm_mirror_register()
 106          */
 107         hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops;
 108         if (__mmu_notifier_register(&hmm->mmu_notifier, mm)) {
 109                 kfree(hmm);
 110                 return NULL;
 111         }
 112
 113         spin_lock(&mm->page_table_lock);
 114         if (!mm->hmm)
 115                 mm->hmm = hmm;
 116         else
 117                 cleanup = true;
 118         spin_unlock(&mm->page_table_lock);
 119
 120         if (cleanup) {
 121                 mmu_notifier_unregister(&hmm->mmu_notifier, mm);
 122                 kfree(hmm);
 123         }
 124
 125         return mm->hmm;
 126 }
 127
 128 void hmm_mm_destroy(struct mm_struct *mm)
 129 {
 130         kfree(mm->hmm);
 131 }
 132
 133 static void hmm_invalidate_range(struct hmm *hmm,
 134                                  enum hmm_update_type action,
 135                                  unsigned long start,
 136                                  unsigned long end)
 137 {
 138         struct hmm_mirror *mirror;
 139         struct hmm_range *range;
 140
 141         spin_lock(&hmm->lock);
 142         list_for_each_entry(range, &hmm->ranges, list) {
 143                 unsigned long addr, idx, npages;
 144
 145                 if (end < range->start || start >= range->end)
 146                         continue;
 147
 148                 range->valid = false;
 149                 addr = max(start, range->start);
 150                 idx = (addr - range->start) >> PAGE_SHIFT;
 151                 npages = (min(range->end, end) - addr) >> PAGE_SHIFT;
 152                 memset(&range->pfns[idx], 0, sizeof(*range->pfns) * npages);
 153         }
 154         spin_unlock(&hmm->lock);
 155
 156         down_read(&hmm->mirrors_sem);
 157         list_for_each_entry(mirror, &hmm->mirrors, list)
 158                 mirror->ops->sync_cpu_device_pagetables(mirror, action,
 159                                                         start, end);
 160         up_read(&hmm->mirrors_sem);
 161 }
 162
 163 static void hmm_invalidate_range_start(struct mmu_notifier *mn,
 164                                        struct mm_struct *mm,
 165                                        unsigned long start,
 166                                        unsigned long end)
 167 {
 168         struct hmm *hmm = mm->hmm;
 169
 170         VM_BUG_ON(!hmm);
 171
 172         atomic_inc(&hmm->sequence);
 173 }
 174
 175 static void hmm_invalidate_range_end(struct mmu_notifier *mn,
 176                                      struct mm_struct *mm,
 177                                      unsigned long start,
 178                                      unsigned long end)
 179 {
 180         struct hmm *hmm = mm->hmm;
 181
 182         VM_BUG_ON(!hmm);
 183
 184         hmm_invalidate_range(mm->hmm, HMM_UPDATE_INVALIDATE, start, end);
 185 }
 186
 187 static const struct mmu_notifier_ops hmm_mmu_notifier_ops = {
 188         .invalidate_range_start = hmm_invalidate_range_start,
 189         .invalidate_range_end   = hmm_invalidate_range_end,
 190 };
 191
 192 /*
 193  * hmm_mirror_register() - register a mirror against an mm
 194  *
 195  * @mirror: new mirror struct to register
 196  * @mm: mm to register against
 197  *
 198  * To start mirroring a process address space, the device driver must register
 199  * an HMM mirror struct.
 200  *
 201  * THE mm->mmap_sem MUST BE HELD IN WRITE MODE !
 202  */
 203 int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm)
 204 {
 205         /* Sanity check */
 206         if (!mm || !mirror || !mirror->ops)
 207                 return -EINVAL;
 208
 209         mirror->hmm = hmm_register(mm);
 210         if (!mirror->hmm)
 211                 return -ENOMEM;
 212
 213         down_write(&mirror->hmm->mirrors_sem);
 214         list_add(&mirror->list, &mirror->hmm->mirrors);
 215         up_write(&mirror->hmm->mirrors_sem);
 216
 217         return 0;
 218 }
 219 EXPORT_SYMBOL(hmm_mirror_register);
 220
 221 /*
 222  * hmm_mirror_unregister() - unregister a mirror
 223  *
 224  * @mirror: new mirror struct to register
 225  *
 226  * Stop mirroring a process address space, and cleanup.
 227  */
 228 void hmm_mirror_unregister(struct hmm_mirror *mirror)
 229 {
 230         struct hmm *hmm = mirror->hmm;
 231
 232         down_write(&hmm->mirrors_sem);
 233         list_del(&mirror->list);
 234         up_write(&hmm->mirrors_sem);
 235 }
 236 EXPORT_SYMBOL(hmm_mirror_unregister);
 237
 238 struct hmm_vma_walk {
 239         struct hmm_range        *range;
 240         unsigned long           last;
 241         bool                    fault;
 242         bool                    block;
 243         bool                    write;
 244 };
 245
 246 static int hmm_vma_do_fault(struct mm_walk *walk,
 247                             unsigned long addr,
 248                             hmm_pfn_t *pfn)
 249 {
 250         unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE;
 251         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 252         struct vm_area_struct *vma = walk->vma;
 253         int r;
 254
 255         flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY;
 256         flags |= hmm_vma_walk->write ? FAULT_FLAG_WRITE : 0;
 257         r = handle_mm_fault(vma, addr, flags);
 258         if (r & VM_FAULT_RETRY)
 259                 return -EBUSY;
 260         if (r & VM_FAULT_ERROR) {
 261                 *pfn = HMM_PFN_ERROR;
 262                 return -EFAULT;
 263         }
 264
 265         return -EAGAIN;
 266 }
 267
 268 static void hmm_pfns_special(hmm_pfn_t *pfns,
 269                              unsigned long addr,
 270                              unsigned long end)
 271 {
 272         for (; addr < end; addr += PAGE_SIZE, pfns++)
 273                 *pfns = HMM_PFN_SPECIAL;
 274 }
 275
 276 static int hmm_pfns_bad(unsigned long addr,
 277                         unsigned long end,
 278                         struct mm_walk *walk)
 279 {
 280         struct hmm_range *range = walk->private;
 281         hmm_pfn_t *pfns = range->pfns;
 282         unsigned long i;
 283
 284         i = (addr - range->start) >> PAGE_SHIFT;
 285         for (; addr < end; addr += PAGE_SIZE, i++)
 286                 pfns[i] = HMM_PFN_ERROR;
 287
 288         return 0;
 289 }
 290
 291 static void hmm_pfns_clear(hmm_pfn_t *pfns,
 292                            unsigned long addr,
 293                            unsigned long end)
 294 {
 295         for (; addr < end; addr += PAGE_SIZE, pfns++)
 296                 *pfns = 0;
 297 }
 298
 299 static int hmm_vma_walk_hole(unsigned long addr,
 300                              unsigned long end,
 301                              struct mm_walk *walk)
 302 {
 303         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 304         struct hmm_range *range = hmm_vma_walk->range;
 305         hmm_pfn_t *pfns = range->pfns;
 306         unsigned long i;
 307
 308         hmm_vma_walk->last = addr;
 309         i = (addr - range->start) >> PAGE_SHIFT;
 310         for (; addr < end; addr += PAGE_SIZE, i++) {
 311                 pfns[i] = HMM_PFN_EMPTY;
 312                 if (hmm_vma_walk->fault) {
 313                         int ret;
 314
 315                         ret = hmm_vma_do_fault(walk, addr, &pfns[i]);
 316                         if (ret != -EAGAIN)
 317                                 return ret;
 318                 }
 319         }
 320
 321         return hmm_vma_walk->fault ? -EAGAIN : 0;
 322 }
 323
 324 static int hmm_vma_walk_clear(unsigned long addr,
 325                               unsigned long end,
 326                               struct mm_walk *walk)
 327 {
 328         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 329         struct hmm_range *range = hmm_vma_walk->range;
 330         hmm_pfn_t *pfns = range->pfns;
 331         unsigned long i;
 332
 333         hmm_vma_walk->last = addr;
 334         i = (addr - range->start) >> PAGE_SHIFT;
 335         for (; addr < end; addr += PAGE_SIZE, i++) {
 336                 pfns[i] = 0;
 337                 if (hmm_vma_walk->fault) {
 338                         int ret;
 339
 340                         ret = hmm_vma_do_fault(walk, addr, &pfns[i]);
 341                         if (ret != -EAGAIN)
 342                                 return ret;
 343                 }
 344         }
 345
 346         return hmm_vma_walk->fault ? -EAGAIN : 0;
 347 }
 348
 349 static int hmm_vma_walk_pmd(pmd_t *pmdp,
 350                             unsigned long start,
 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         struct vm_area_struct *vma = walk->vma;
 357         hmm_pfn_t *pfns = range->pfns;
 358         unsigned long addr = start, i;
 359         bool write_fault;
 360         hmm_pfn_t flag;
 361         pte_t *ptep;
 362
 363         i = (addr - range->start) >> PAGE_SHIFT;
 364         flag = vma->vm_flags & VM_READ ? HMM_PFN_READ : 0;
 365         write_fault = hmm_vma_walk->fault & hmm_vma_walk->write;
 366
 367 again:
 368         if (pmd_none(*pmdp))
 369                 return hmm_vma_walk_hole(start, end, walk);
 370
 371         if (pmd_huge(*pmdp) && vma->vm_flags & VM_HUGETLB)
 372                 return hmm_pfns_bad(start, end, walk);
 373
 374         if (pmd_devmap(*pmdp) || pmd_trans_huge(*pmdp)) {
 375                 unsigned long pfn;
 376                 pmd_t pmd;
 377
 378                 /*
 379                  * No need to take pmd_lock here, even if some other threads
 380                  * is splitting the huge pmd we will get that event through
 381                  * mmu_notifier callback.
 382                  *
 383                  * So just read pmd value and check again its a transparent
 384                  * huge or device mapping one and compute corresponding pfn
 385                  * values.
 386                  */
 387                 pmd = pmd_read_atomic(pmdp);
 388                 barrier();
 389                 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
 390                         goto again;
 391                 if (pmd_protnone(pmd))
 392                         return hmm_vma_walk_clear(start, end, walk);
 393
 394                 if (write_fault && !pmd_write(pmd))
 395                         return hmm_vma_walk_clear(start, end, walk);
 396
 397                 pfn = pmd_pfn(pmd) + pte_index(addr);
 398                 flag |= pmd_write(pmd) ? HMM_PFN_WRITE : 0;
 399                 for (; addr < end; addr += PAGE_SIZE, i++, pfn++)
 400                         pfns[i] = hmm_pfn_t_from_pfn(pfn) | flag;
 401                 return 0;
 402         }
 403
 404         if (pmd_bad(*pmdp))
 405                 return hmm_pfns_bad(start, end, walk);
 406
 407         ptep = pte_offset_map(pmdp, addr);
 408         for (; addr < end; addr += PAGE_SIZE, ptep++, i++) {
 409                 pte_t pte = *ptep;
 410
 411                 pfns[i] = 0;
 412
 413                 if (pte_none(pte)) {
 414                         pfns[i] = HMM_PFN_EMPTY;
 415                         if (hmm_vma_walk->fault)
 416                                 goto fault;
 417                         continue;
 418                 }
 419
 420                 if (!pte_present(pte)) {
 421                         swp_entry_t entry = pte_to_swp_entry(pte);
 422
 423                         if (!non_swap_entry(entry)) {
 424                                 if (hmm_vma_walk->fault)
 425                                         goto fault;
 426                                 continue;
 427                         }
 428
 429                         /*
 430                          * This is a special swap entry, ignore migration, use
 431                          * device and report anything else as error.
 432                          */
 433                         if (is_device_private_entry(entry)) {
 434                                 pfns[i] = hmm_pfn_t_from_pfn(swp_offset(entry));
 435                                 if (is_write_device_private_entry(entry)) {
 436                                         pfns[i] |= HMM_PFN_WRITE;
 437                                 } else if (write_fault)
 438                                         goto fault;
 439                                 pfns[i] |= HMM_PFN_DEVICE_UNADDRESSABLE;
 440                                 pfns[i] |= flag;
 441                         } else if (is_migration_entry(entry)) {
 442                                 if (hmm_vma_walk->fault) {
 443                                         pte_unmap(ptep);
 444                                         hmm_vma_walk->last = addr;
 445                                         migration_entry_wait(vma->vm_mm,
 446                                                              pmdp, addr);
 447                                         return -EAGAIN;
 448                                 }
 449                                 continue;
 450                         } else {
 451                                 /* Report error for everything else */
 452                                 pfns[i] = HMM_PFN_ERROR;
 453                         }
 454                         continue;
 455                 }
 456
 457                 if (write_fault && !pte_write(pte))
 458                         goto fault;
 459
 460                 pfns[i] = hmm_pfn_t_from_pfn(pte_pfn(pte)) | flag;
 461                 pfns[i] |= pte_write(pte) ? HMM_PFN_WRITE : 0;
 462                 continue;
 463
 464 fault:
 465                 pte_unmap(ptep);
 466                 /* Fault all pages in range */
 467                 return hmm_vma_walk_clear(start, end, walk);
 468         }
 469         pte_unmap(ptep - 1);
 470
 471         return 0;
 472 }
 473
 474 /*
 475  * hmm_vma_get_pfns() - snapshot CPU page table for a range of virtual addresses
 476  * @vma: virtual memory area containing the virtual address range
 477  * @range: used to track snapshot validity
 478  * @start: range virtual start address (inclusive)
 479  * @end: range virtual end address (exclusive)
 480  * @entries: array of hmm_pfn_t: provided by the caller, filled in by function
 481  * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, 0 success
 482  *
 483  * This snapshots the CPU page table for a range of virtual addresses. Snapshot
 484  * validity is tracked by range struct. See hmm_vma_range_done() for further
 485  * information.
 486  *
 487  * The range struct is initialized here. It tracks the CPU page table, but only
 488  * if the function returns success (0), in which case the caller must then call
 489  * hmm_vma_range_done() to stop CPU page table update tracking on this range.
 490  *
 491  * NOT CALLING hmm_vma_range_done() IF FUNCTION RETURNS 0 WILL LEAD TO SERIOUS
 492  * MEMORY CORRUPTION ! YOU HAVE BEEN WARNED !
 493  */
 494 int hmm_vma_get_pfns(struct vm_area_struct *vma,
 495                      struct hmm_range *range,
 496                      unsigned long start,
 497                      unsigned long end,
 498                      hmm_pfn_t *pfns)
 499 {
 500         struct hmm_vma_walk hmm_vma_walk;
 501         struct mm_walk mm_walk;
 502         struct hmm *hmm;
 503
 504         /* FIXME support hugetlb fs */
 505         if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) {
 506                 hmm_pfns_special(pfns, start, end);
 507                 return -EINVAL;
 508         }
 509
 510         /* Sanity check, this really should not happen ! */
 511         if (start < vma->vm_start || start >= vma->vm_end)
 512                 return -EINVAL;
 513         if (end < vma->vm_start || end > vma->vm_end)
 514                 return -EINVAL;
 515
 516         hmm = hmm_register(vma->vm_mm);
 517         if (!hmm)
 518                 return -ENOMEM;
 519         /* Caller must have registered a mirror, via hmm_mirror_register() ! */
 520         if (!hmm->mmu_notifier.ops)
 521                 return -EINVAL;
 522
 523         /* Initialize range to track CPU page table update */
 524         range->start = start;
 525         range->pfns = pfns;
 526         range->end = end;
 527         spin_lock(&hmm->lock);
 528         range->valid = true;
 529         list_add_rcu(&range->list, &hmm->ranges);
 530         spin_unlock(&hmm->lock);
 531
 532         hmm_vma_walk.fault = false;
 533         hmm_vma_walk.range = range;
 534         mm_walk.private = &hmm_vma_walk;
 535
 536         mm_walk.vma = vma;
 537         mm_walk.mm = vma->vm_mm;
 538         mm_walk.pte_entry = NULL;
 539         mm_walk.test_walk = NULL;
 540         mm_walk.hugetlb_entry = NULL;
 541         mm_walk.pmd_entry = hmm_vma_walk_pmd;
 542         mm_walk.pte_hole = hmm_vma_walk_hole;
 543
 544         walk_page_range(start, end, &mm_walk);
 545         return 0;
 546 }
 547 EXPORT_SYMBOL(hmm_vma_get_pfns);
 548
 549 /*
 550  * hmm_vma_range_done() - stop tracking change to CPU page table over a range
 551  * @vma: virtual memory area containing the virtual address range
 552  * @range: range being tracked
 553  * Returns: false if range data has been invalidated, true otherwise
 554  *
 555  * Range struct is used to track updates to the CPU page table after a call to
 556  * either hmm_vma_get_pfns() or hmm_vma_fault(). Once the device driver is done
 557  * using the data,  or wants to lock updates to the data it got from those
 558  * functions, it must call the hmm_vma_range_done() function, which will then
 559  * stop tracking CPU page table updates.
 560  *
 561  * Note that device driver must still implement general CPU page table update
 562  * tracking either by using hmm_mirror (see hmm_mirror_register()) or by using
 563  * the mmu_notifier API directly.
 564  *
 565  * CPU page table update tracking done through hmm_range is only temporary and
 566  * to be used while trying to duplicate CPU page table contents for a range of
 567  * virtual addresses.
 568  *
 569  * There are two ways to use this :
 570  * again:
 571  *   hmm_vma_get_pfns(vma, range, start, end, pfns); or hmm_vma_fault(...);
 572  *   trans = device_build_page_table_update_transaction(pfns);
 573  *   device_page_table_lock();
 574  *   if (!hmm_vma_range_done(vma, range)) {
 575  *     device_page_table_unlock();
 576  *     goto again;
 577  *   }
 578  *   device_commit_transaction(trans);
 579  *   device_page_table_unlock();
 580  *
 581  * Or:
 582  *   hmm_vma_get_pfns(vma, range, start, end, pfns); or hmm_vma_fault(...);
 583  *   device_page_table_lock();
 584  *   hmm_vma_range_done(vma, range);
 585  *   device_update_page_table(pfns);
 586  *   device_page_table_unlock();
 587  */
 588 bool hmm_vma_range_done(struct vm_area_struct *vma, struct hmm_range *range)
 589 {
 590         unsigned long npages = (range->end - range->start) >> PAGE_SHIFT;
 591         struct hmm *hmm;
 592
 593         if (range->end <= range->start) {
 594                 BUG();
 595                 return false;
 596         }
 597
 598         hmm = hmm_register(vma->vm_mm);
 599         if (!hmm) {
 600                 memset(range->pfns, 0, sizeof(*range->pfns) * npages);
 601                 return false;
 602         }
 603
 604         spin_lock(&hmm->lock);
 605         list_del_rcu(&range->list);
 606         spin_unlock(&hmm->lock);
 607
 608         return range->valid;
 609 }
 610 EXPORT_SYMBOL(hmm_vma_range_done);
 611
 612 /*
 613  * hmm_vma_fault() - try to fault some address in a virtual address range
 614  * @vma: virtual memory area containing the virtual address range
 615  * @range: use to track pfns array content validity
 616  * @start: fault range virtual start address (inclusive)
 617  * @end: fault range virtual end address (exclusive)
 618  * @pfns: array of hmm_pfn_t, only entry with fault flag set will be faulted
 619  * @write: is it a write fault
 620  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
 621  * Returns: 0 success, error otherwise (-EAGAIN means mmap_sem have been drop)
 622  *
 623  * This is similar to a regular CPU page fault except that it will not trigger
 624  * any memory migration if the memory being faulted is not accessible by CPUs.
 625  *
 626  * On error, for one virtual address in the range, the function will set the
 627  * hmm_pfn_t error flag for the corresponding pfn entry.
 628  *
 629  * Expected use pattern:
 630  * retry:
 631  *   down_read(&mm->mmap_sem);
 632  *   // Find vma and address device wants to fault, initialize hmm_pfn_t
 633  *   // array accordingly
 634  *   ret = hmm_vma_fault(vma, start, end, pfns, allow_retry);
 635  *   switch (ret) {
 636  *   case -EAGAIN:
 637  *     hmm_vma_range_done(vma, range);
 638  *     // You might want to rate limit or yield to play nicely, you may
 639  *     // also commit any valid pfn in the array assuming that you are
 640  *     // getting true from hmm_vma_range_monitor_end()
 641  *     goto retry;
 642  *   case 0:
 643  *     break;
 644  *   default:
 645  *     // Handle error !
 646  *     up_read(&mm->mmap_sem)
 647  *     return;
 648  *   }
 649  *   // Take device driver lock that serialize device page table update
 650  *   driver_lock_device_page_table_update();
 651  *   hmm_vma_range_done(vma, range);
 652  *   // Commit pfns we got from hmm_vma_fault()
 653  *   driver_unlock_device_page_table_update();
 654  *   up_read(&mm->mmap_sem)
 655  *
 656  * YOU MUST CALL hmm_vma_range_done() AFTER THIS FUNCTION RETURN SUCCESS (0)
 657  * BEFORE FREEING THE range struct OR YOU WILL HAVE SERIOUS MEMORY CORRUPTION !
 658  *
 659  * YOU HAVE BEEN WARNED !
 660  */
 661 int hmm_vma_fault(struct vm_area_struct *vma,
 662                   struct hmm_range *range,
 663                   unsigned long start,
 664                   unsigned long end,
 665                   hmm_pfn_t *pfns,
 666                   bool write,
 667                   bool block)
 668 {
 669         struct hmm_vma_walk hmm_vma_walk;
 670         struct mm_walk mm_walk;
 671         struct hmm *hmm;
 672         int ret;
 673
 674         /* Sanity check, this really should not happen ! */
 675         if (start < vma->vm_start || start >= vma->vm_end)
 676                 return -EINVAL;
 677         if (end < vma->vm_start || end > vma->vm_end)
 678                 return -EINVAL;
 679
 680         hmm = hmm_register(vma->vm_mm);
 681         if (!hmm) {
 682                 hmm_pfns_clear(pfns, start, end);
 683                 return -ENOMEM;
 684         }
 685         /* Caller must have registered a mirror using hmm_mirror_register() */
 686         if (!hmm->mmu_notifier.ops)
 687                 return -EINVAL;
 688
 689         /* Initialize range to track CPU page table update */
 690         range->start = start;
 691         range->pfns = pfns;
 692         range->end = end;
 693         spin_lock(&hmm->lock);
 694         range->valid = true;
 695         list_add_rcu(&range->list, &hmm->ranges);
 696         spin_unlock(&hmm->lock);
 697
 698         /* FIXME support hugetlb fs */
 699         if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) {
 700                 hmm_pfns_special(pfns, start, end);
 701                 return 0;
 702         }
 703
 704         hmm_vma_walk.fault = true;
 705         hmm_vma_walk.write = write;
 706         hmm_vma_walk.block = block;
 707         hmm_vma_walk.range = range;
 708         mm_walk.private = &hmm_vma_walk;
 709         hmm_vma_walk.last = range->start;
 710
 711         mm_walk.vma = vma;
 712         mm_walk.mm = vma->vm_mm;
 713         mm_walk.pte_entry = NULL;
 714         mm_walk.test_walk = NULL;
 715         mm_walk.hugetlb_entry = NULL;
 716         mm_walk.pmd_entry = hmm_vma_walk_pmd;
 717         mm_walk.pte_hole = hmm_vma_walk_hole;
 718
 719         do {
 720                 ret = walk_page_range(start, end, &mm_walk);
 721                 start = hmm_vma_walk.last;
 722         } while (ret == -EAGAIN);
 723
 724         if (ret) {
 725                 unsigned long i;
 726
 727                 i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
 728                 hmm_pfns_clear(&pfns[i], hmm_vma_walk.last, end);
 729                 hmm_vma_range_done(vma, range);
 730         }
 731         return ret;
 732 }
 733 EXPORT_SYMBOL(hmm_vma_fault);
 734 #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
 735
 736
 737 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) ||  IS_ENABLED(CONFIG_DEVICE_PUBLIC)
 738 struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
 739                                        unsigned long addr)
 740 {
 741         struct page *page;
 742
 743         page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
 744         if (!page)
 745                 return NULL;
 746         lock_page(page);
 747         return page;
 748 }
 749 EXPORT_SYMBOL(hmm_vma_alloc_locked_page);
 750
 751
 752 static void hmm_devmem_ref_release(struct percpu_ref *ref)
 753 {
 754         struct hmm_devmem *devmem;
 755
 756         devmem = container_of(ref, struct hmm_devmem, ref);
 757         complete(&devmem->completion);
 758 }
 759
 760 static void hmm_devmem_ref_exit(void *data)
 761 {
 762         struct percpu_ref *ref = data;
 763         struct hmm_devmem *devmem;
 764
 765         devmem = container_of(ref, struct hmm_devmem, ref);
 766         percpu_ref_exit(ref);
 767         devm_remove_action(devmem->device, &hmm_devmem_ref_exit, data);
 768 }
 769
 770 static void hmm_devmem_ref_kill(void *data)
 771 {
 772         struct percpu_ref *ref = data;
 773         struct hmm_devmem *devmem;
 774
 775         devmem = container_of(ref, struct hmm_devmem, ref);
 776         percpu_ref_kill(ref);
 777         wait_for_completion(&devmem->completion);
 778         devm_remove_action(devmem->device, &hmm_devmem_ref_kill, data);
 779 }
 780
 781 static int hmm_devmem_fault(struct vm_area_struct *vma,
 782                             unsigned long addr,
 783                             const struct page *page,
 784                             unsigned int flags,
 785                             pmd_t *pmdp)
 786 {
 787         struct hmm_devmem *devmem = page->pgmap->data;
 788
 789         return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp);
 790 }
 791
 792 static void hmm_devmem_free(struct page *page, void *data)
 793 {
 794         struct hmm_devmem *devmem = data;
 795
 796         devmem->ops->free(devmem, page);
 797 }
 798
 799 static DEFINE_MUTEX(hmm_devmem_lock);
 800 static RADIX_TREE(hmm_devmem_radix, GFP_KERNEL);
 801
 802 static void hmm_devmem_radix_release(struct resource *resource)
 803 {
 804         resource_size_t key, align_start, align_size;
 805
 806         align_start = resource->start & ~(PA_SECTION_SIZE - 1);
 807         align_size = ALIGN(resource_size(resource), PA_SECTION_SIZE);
 808
 809         mutex_lock(&hmm_devmem_lock);
 810         for (key = resource->start;
 811              key <= resource->end;
 812              key += PA_SECTION_SIZE)
 813                 radix_tree_delete(&hmm_devmem_radix, key >> PA_SECTION_SHIFT);
 814         mutex_unlock(&hmm_devmem_lock);
 815 }
 816
 817 static void hmm_devmem_release(struct device *dev, void *data)
 818 {
 819         struct hmm_devmem *devmem = data;
 820         struct resource *resource = devmem->resource;
 821         unsigned long start_pfn, npages;
 822         struct zone *zone;
 823         struct page *page;
 824
 825         if (percpu_ref_tryget_live(&devmem->ref)) {
 826                 dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
 827                 percpu_ref_put(&devmem->ref);
 828         }
 829
 830         /* pages are dead and unused, undo the arch mapping */
 831         start_pfn = (resource->start & ~(PA_SECTION_SIZE - 1)) >> PAGE_SHIFT;
 832         npages = ALIGN(resource_size(resource), PA_SECTION_SIZE) >> PAGE_SHIFT;
 833
 834         page = pfn_to_page(start_pfn);
 835         zone = page_zone(page);
 836
 837         mem_hotplug_begin();
 838         if (resource->desc == IORES_DESC_DEVICE_PRIVATE_MEMORY)
 839                 __remove_pages(zone, start_pfn, npages);
 840         else
 841                 arch_remove_memory(start_pfn << PAGE_SHIFT,
 842                                    npages << PAGE_SHIFT);
 843         mem_hotplug_done();
 844
 845         hmm_devmem_radix_release(resource);
 846 }
 847
 848 static struct hmm_devmem *hmm_devmem_find(resource_size_t phys)
 849 {
 850         WARN_ON_ONCE(!rcu_read_lock_held());
 851
 852         return radix_tree_lookup(&hmm_devmem_radix, phys >> PA_SECTION_SHIFT);
 853 }
 854
 855 static int hmm_devmem_pages_create(struct hmm_devmem *devmem)
 856 {
 857         resource_size_t key, align_start, align_size, align_end;
 858         struct device *device = devmem->device;
 859         int ret, nid, is_ram;
 860         unsigned long pfn;
 861
 862         align_start = devmem->resource->start & ~(PA_SECTION_SIZE - 1);
 863         align_size = ALIGN(devmem->resource->start +
 864                            resource_size(devmem->resource),
 865                            PA_SECTION_SIZE) - align_start;
 866
 867         is_ram = region_intersects(align_start, align_size,
 868                                    IORESOURCE_SYSTEM_RAM,
 869                                    IORES_DESC_NONE);
 870         if (is_ram == REGION_MIXED) {
 871                 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
 872                                 __func__, devmem->resource);
 873                 return -ENXIO;
 874         }
 875         if (is_ram == REGION_INTERSECTS)
 876                 return -ENXIO;
 877
 878         if (devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY)
 879                 devmem->pagemap.type = MEMORY_DEVICE_PUBLIC;
 880         else
 881                 devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
 882
 883         devmem->pagemap.res = devmem->resource;
 884         devmem->pagemap.page_fault = hmm_devmem_fault;
 885         devmem->pagemap.page_free = hmm_devmem_free;
 886         devmem->pagemap.dev = devmem->device;
 887         devmem->pagemap.ref = &devmem->ref;
 888         devmem->pagemap.data = devmem;
 889
 890         mutex_lock(&hmm_devmem_lock);
 891         align_end = align_start + align_size - 1;
 892         for (key = align_start; key <= align_end; key += PA_SECTION_SIZE) {
 893                 struct hmm_devmem *dup;
 894
 895                 rcu_read_lock();
 896                 dup = hmm_devmem_find(key);
 897                 rcu_read_unlock();
 898                 if (dup) {
 899                         dev_err(device, "%s: collides with mapping for %s\n",
 900                                 __func__, dev_name(dup->device));
 901                         mutex_unlock(&hmm_devmem_lock);
 902                         ret = -EBUSY;
 903                         goto error;
 904                 }
 905                 ret = radix_tree_insert(&hmm_devmem_radix,
 906                                         key >> PA_SECTION_SHIFT,
 907                                         devmem);
 908                 if (ret) {
 909                         dev_err(device, "%s: failed: %d\n", __func__, ret);
 910                         mutex_unlock(&hmm_devmem_lock);
 911                         goto error_radix;
 912                 }
 913         }
 914         mutex_unlock(&hmm_devmem_lock);
 915
 916         nid = dev_to_node(device);
 917         if (nid < 0)
 918                 nid = numa_mem_id();
 919
 920         mem_hotplug_begin();
 921         /*
 922          * For device private memory we call add_pages() as we only need to
 923          * allocate and initialize struct page for the device memory. More-
 924          * over the device memory is un-accessible thus we do not want to
 925          * create a linear mapping for the memory like arch_add_memory()
 926          * would do.
 927          *
 928          * For device public memory, which is accesible by the CPU, we do
 929          * want the linear mapping and thus use arch_add_memory().
 930          */
 931         if (devmem->pagemap.type == MEMORY_DEVICE_PUBLIC)
 932                 ret = arch_add_memory(nid, align_start, align_size, false);
 933         else
 934                 ret = add_pages(nid, align_start >> PAGE_SHIFT,
 935                                 align_size >> PAGE_SHIFT, false);
 936         if (ret) {
 937                 mem_hotplug_done();
 938                 goto error_add_memory;
 939         }
 940         move_pfn_range_to_zone(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
 941                                 align_start >> PAGE_SHIFT,
 942                                 align_size >> PAGE_SHIFT);
 943         mem_hotplug_done();
 944
 945         for (pfn = devmem->pfn_first; pfn < devmem->pfn_last; pfn++) {
 946                 struct page *page = pfn_to_page(pfn);
 947
 948                 page->pgmap = &devmem->pagemap;
 949         }
 950         return 0;
 951
 952 error_add_memory:
 953         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
 954 error_radix:
 955         hmm_devmem_radix_release(devmem->resource);
 956 error:
 957         return ret;
 958 }
 959
 960 static int hmm_devmem_match(struct device *dev, void *data, void *match_data)
 961 {
 962         struct hmm_devmem *devmem = data;
 963
 964         return devmem->resource == match_data;
 965 }
 966
 967 static void hmm_devmem_pages_remove(struct hmm_devmem *devmem)
 968 {
 969         devres_release(devmem->device, &hmm_devmem_release,
 970                        &hmm_devmem_match, devmem->resource);
 971 }
 972
 973 /*
 974  * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory
 975  *
 976  * @ops: memory event device driver callback (see struct hmm_devmem_ops)
 977  * @device: device struct to bind the resource too
 978  * @size: size in bytes of the device memory to add
 979  * Returns: pointer to new hmm_devmem struct ERR_PTR otherwise
 980  *
 981  * This function first finds an empty range of physical address big enough to
 982  * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which
 983  * in turn allocates struct pages. It does not do anything beyond that; all
 984  * events affecting the memory will go through the various callbacks provided
 985  * by hmm_devmem_ops struct.
 986  *
 987  * Device driver should call this function during device initialization and
 988  * is then responsible of memory management. HMM only provides helpers.
 989  */
 990 struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
 991                                   struct device *device,
 992                                   unsigned long size)
 993 {
 994         struct hmm_devmem *devmem;
 995         resource_size_t addr;
 996         int ret;
 997
 998         static_branch_enable(&device_private_key);
 999
1000         devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem),
1001                                    GFP_KERNEL, dev_to_node(device));
1002         if (!devmem)
1003                 return ERR_PTR(-ENOMEM);
1004
1005         init_completion(&devmem->completion);
1006         devmem->pfn_first = -1UL;
1007         devmem->pfn_last = -1UL;
1008         devmem->resource = NULL;
1009         devmem->device = device;
1010         devmem->ops = ops;
1011
1012         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1013                               0, GFP_KERNEL);
1014         if (ret)
1015                 goto error_percpu_ref;
1016
1017         ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref);
1018         if (ret)
1019                 goto error_devm_add_action;
1020
1021         size = ALIGN(size, PA_SECTION_SIZE);
1022         addr = min((unsigned long)iomem_resource.end,
1023                    (1UL << MAX_PHYSMEM_BITS) - 1);
1024         addr = addr - size + 1UL;
1025
1026         /*
1027          * FIXME add a new helper to quickly walk resource tree and find free
1028          * range
1029          *
1030          * FIXME what about ioport_resource resource ?
1031          */
1032         for (; addr > size && addr >= iomem_resource.start; addr -= size) {
1033                 ret = region_intersects(addr, size, 0, IORES_DESC_NONE);
1034                 if (ret != REGION_DISJOINT)
1035                         continue;
1036
1037                 devmem->resource = devm_request_mem_region(device, addr, size,
1038                                                            dev_name(device));
1039                 if (!devmem->resource) {
1040                         ret = -ENOMEM;
1041                         goto error_no_resource;
1042                 }
1043                 break;
1044         }
1045         if (!devmem->resource) {
1046                 ret = -ERANGE;
1047                 goto error_no_resource;
1048         }
1049
1050         devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY;
1051         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1052         devmem->pfn_last = devmem->pfn_first +
1053                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1054
1055         ret = hmm_devmem_pages_create(devmem);
1056         if (ret)
1057                 goto error_pages;
1058
1059         devres_add(device, devmem);
1060
1061         ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref);
1062         if (ret) {
1063                 hmm_devmem_remove(devmem);
1064                 return ERR_PTR(ret);
1065         }
1066
1067         return devmem;
1068
1069 error_pages:
1070         devm_release_mem_region(device, devmem->resource->start,
1071                                 resource_size(devmem->resource));
1072 error_no_resource:
1073 error_devm_add_action:
1074         hmm_devmem_ref_kill(&devmem->ref);
1075         hmm_devmem_ref_exit(&devmem->ref);
1076 error_percpu_ref:
1077         devres_free(devmem);
1078         return ERR_PTR(ret);
1079 }
1080 EXPORT_SYMBOL(hmm_devmem_add);
1081
1082 struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops,
1083                                            struct device *device,
1084                                            struct resource *res)
1085 {
1086         struct hmm_devmem *devmem;
1087         int ret;
1088
1089         if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY)
1090                 return ERR_PTR(-EINVAL);
1091
1092         static_branch_enable(&device_private_key);
1093
1094         devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem),
1095                                    GFP_KERNEL, dev_to_node(device));
1096         if (!devmem)
1097                 return ERR_PTR(-ENOMEM);
1098
1099         init_completion(&devmem->completion);
1100         devmem->pfn_first = -1UL;
1101         devmem->pfn_last = -1UL;
1102         devmem->resource = res;
1103         devmem->device = device;
1104         devmem->ops = ops;
1105
1106         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1107                               0, GFP_KERNEL);
1108         if (ret)
1109                 goto error_percpu_ref;
1110
1111         ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref);
1112         if (ret)
1113                 goto error_devm_add_action;
1114
1115
1116         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1117         devmem->pfn_last = devmem->pfn_first +
1118                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1119
1120         ret = hmm_devmem_pages_create(devmem);
1121         if (ret)
1122                 goto error_devm_add_action;
1123
1124         devres_add(device, devmem);
1125
1126         ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref);
1127         if (ret) {
1128                 hmm_devmem_remove(devmem);
1129                 return ERR_PTR(ret);
1130         }
1131
1132         return devmem;
1133
1134 error_devm_add_action:
1135         hmm_devmem_ref_kill(&devmem->ref);
1136         hmm_devmem_ref_exit(&devmem->ref);
1137 error_percpu_ref:
1138         devres_free(devmem);
1139         return ERR_PTR(ret);
1140 }
1141 EXPORT_SYMBOL(hmm_devmem_add_resource);
1142
1143 /*
1144  * hmm_devmem_remove() - remove device memory (kill and free ZONE_DEVICE)
1145  *
1146  * @devmem: hmm_devmem struct use to track and manage the ZONE_DEVICE memory
1147  *
1148  * This will hot-unplug memory that was hotplugged by hmm_devmem_add on behalf
1149  * of the device driver. It will free struct page and remove the resource that
1150  * reserved the physical address range for this device memory.
1151  */
1152 void hmm_devmem_remove(struct hmm_devmem *devmem)
1153 {
1154         resource_size_t start, size;
1155         struct device *device;
1156         bool cdm = false;
1157
1158         if (!devmem)
1159                 return;
1160
1161         device = devmem->device;
1162         start = devmem->resource->start;
1163         size = resource_size(devmem->resource);
1164
1165         cdm = devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY;
1166         hmm_devmem_ref_kill(&devmem->ref);
1167         hmm_devmem_ref_exit(&devmem->ref);
1168         hmm_devmem_pages_remove(devmem);
1169
1170         if (!cdm)
1171                 devm_release_mem_region(device, start, size);
1172 }
1173 EXPORT_SYMBOL(hmm_devmem_remove);
1174
1175 /*
1176  * A device driver that wants to handle multiple devices memory through a
1177  * single fake device can use hmm_device to do so. This is purely a helper
1178  * and it is not needed to make use of any HMM functionality.
1179  */
1180 #define HMM_DEVICE_MAX 256
1181
1182 static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX);
1183 static DEFINE_SPINLOCK(hmm_device_lock);
1184 static struct class *hmm_device_class;
1185 static dev_t hmm_device_devt;
1186
1187 static void hmm_device_release(struct device *device)
1188 {
1189         struct hmm_device *hmm_device;
1190
1191         hmm_device = container_of(device, struct hmm_device, device);
1192         spin_lock(&hmm_device_lock);
1193         clear_bit(hmm_device->minor, hmm_device_mask);
1194         spin_unlock(&hmm_device_lock);
1195
1196         kfree(hmm_device);
1197 }
1198
1199 struct hmm_device *hmm_device_new(void *drvdata)
1200 {
1201         struct hmm_device *hmm_device;
1202
1203         hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL);
1204         if (!hmm_device)
1205                 return ERR_PTR(-ENOMEM);
1206
1207         spin_lock(&hmm_device_lock);
1208         hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX);
1209         if (hmm_device->minor >= HMM_DEVICE_MAX) {
1210                 spin_unlock(&hmm_device_lock);
1211                 kfree(hmm_device);
1212                 return ERR_PTR(-EBUSY);
1213         }
1214         set_bit(hmm_device->minor, hmm_device_mask);
1215         spin_unlock(&hmm_device_lock);
1216
1217         dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor);
1218         hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt),
1219                                         hmm_device->minor);
1220         hmm_device->device.release = hmm_device_release;
1221         dev_set_drvdata(&hmm_device->device, drvdata);
1222         hmm_device->device.class = hmm_device_class;
1223         device_initialize(&hmm_device->device);
1224
1225         return hmm_device;
1226 }
1227 EXPORT_SYMBOL(hmm_device_new);
1228
1229 void hmm_device_put(struct hmm_device *hmm_device)
1230 {
1231         put_device(&hmm_device->device);
1232 }
1233 EXPORT_SYMBOL(hmm_device_put);
1234
1235 static int __init hmm_init(void)
1236 {
1237         int ret;
1238
1239         ret = alloc_chrdev_region(&hmm_device_devt, 0,
1240                                   HMM_DEVICE_MAX,
1241                                   "hmm_device");
1242         if (ret)
1243                 return ret;
1244
1245         hmm_device_class = class_create(THIS_MODULE, "hmm_device");
1246         if (IS_ERR(hmm_device_class)) {
1247                 unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX);
1248                 return PTR_ERR(hmm_device_class);
1249         }
1250         return 0;
1251 }
1252
1253 device_initcall(hmm_init);
1254 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */