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_release(struct mmu_notifier *mn, struct mm_struct *mm)
 164 {
 165         struct hmm_mirror *mirror;
 166         struct hmm *hmm = mm->hmm;
 167
 168         down_write(&hmm->mirrors_sem);
 169         mirror = list_first_entry_or_null(&hmm->mirrors, struct hmm_mirror,
 170                                           list);
 171         while (mirror) {
 172                 list_del_init(&mirror->list);
 173                 if (mirror->ops->release) {
 174                         /*
 175                          * Drop mirrors_sem so callback can wait on any pending
 176                          * work that might itself trigger mmu_notifier callback
 177                          * and thus would deadlock with us.
 178                          */
 179                         up_write(&hmm->mirrors_sem);
 180                         mirror->ops->release(mirror);
 181                         down_write(&hmm->mirrors_sem);
 182                 }
 183                 mirror = list_first_entry_or_null(&hmm->mirrors,
 184                                                   struct hmm_mirror, list);
 185         }
 186         up_write(&hmm->mirrors_sem);
 187 }
 188
 189 static void hmm_invalidate_range_start(struct mmu_notifier *mn,
 190                                        struct mm_struct *mm,
 191                                        unsigned long start,
 192                                        unsigned long end)
 193 {
 194         struct hmm *hmm = mm->hmm;
 195
 196         VM_BUG_ON(!hmm);
 197
 198         atomic_inc(&hmm->sequence);
 199 }
 200
 201 static void hmm_invalidate_range_end(struct mmu_notifier *mn,
 202                                      struct mm_struct *mm,
 203                                      unsigned long start,
 204                                      unsigned long end)
 205 {
 206         struct hmm *hmm = mm->hmm;
 207
 208         VM_BUG_ON(!hmm);
 209
 210         hmm_invalidate_range(mm->hmm, HMM_UPDATE_INVALIDATE, start, end);
 211 }
 212
 213 static const struct mmu_notifier_ops hmm_mmu_notifier_ops = {
 214         .release                = hmm_release,
 215         .invalidate_range_start = hmm_invalidate_range_start,
 216         .invalidate_range_end   = hmm_invalidate_range_end,
 217 };
 218
 219 /*
 220  * hmm_mirror_register() - register a mirror against an mm
 221  *
 222  * @mirror: new mirror struct to register
 223  * @mm: mm to register against
 224  *
 225  * To start mirroring a process address space, the device driver must register
 226  * an HMM mirror struct.
 227  *
 228  * THE mm->mmap_sem MUST BE HELD IN WRITE MODE !
 229  */
 230 int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm)
 231 {
 232         /* Sanity check */
 233         if (!mm || !mirror || !mirror->ops)
 234                 return -EINVAL;
 235
 236 again:
 237         mirror->hmm = hmm_register(mm);
 238         if (!mirror->hmm)
 239                 return -ENOMEM;
 240
 241         down_write(&mirror->hmm->mirrors_sem);
 242         if (mirror->hmm->mm == NULL) {
 243                 /*
 244                  * A racing hmm_mirror_unregister() is about to destroy the hmm
 245                  * struct. Try again to allocate a new one.
 246                  */
 247                 up_write(&mirror->hmm->mirrors_sem);
 248                 mirror->hmm = NULL;
 249                 goto again;
 250         } else {
 251                 list_add(&mirror->list, &mirror->hmm->mirrors);
 252                 up_write(&mirror->hmm->mirrors_sem);
 253         }
 254
 255         return 0;
 256 }
 257 EXPORT_SYMBOL(hmm_mirror_register);
 258
 259 /*
 260  * hmm_mirror_unregister() - unregister a mirror
 261  *
 262  * @mirror: new mirror struct to register
 263  *
 264  * Stop mirroring a process address space, and cleanup.
 265  */
 266 void hmm_mirror_unregister(struct hmm_mirror *mirror)
 267 {
 268         bool should_unregister = false;
 269         struct mm_struct *mm;
 270         struct hmm *hmm;
 271
 272         if (mirror->hmm == NULL)
 273                 return;
 274
 275         hmm = mirror->hmm;
 276         down_write(&hmm->mirrors_sem);
 277         list_del_init(&mirror->list);
 278         should_unregister = list_empty(&hmm->mirrors);
 279         mirror->hmm = NULL;
 280         mm = hmm->mm;
 281         hmm->mm = NULL;
 282         up_write(&hmm->mirrors_sem);
 283
 284         if (!should_unregister || mm == NULL)
 285                 return;
 286
 287         spin_lock(&mm->page_table_lock);
 288         if (mm->hmm == hmm)
 289                 mm->hmm = NULL;
 290         spin_unlock(&mm->page_table_lock);
 291
 292         mmu_notifier_unregister_no_release(&hmm->mmu_notifier, mm);
 293         kfree(hmm);
 294 }
 295 EXPORT_SYMBOL(hmm_mirror_unregister);
 296
 297 struct hmm_vma_walk {
 298         struct hmm_range        *range;
 299         unsigned long           last;
 300         bool                    fault;
 301         bool                    block;
 302         bool                    write;
 303 };
 304
 305 static int hmm_vma_do_fault(struct mm_walk *walk,
 306                             unsigned long addr,
 307                             hmm_pfn_t *pfn)
 308 {
 309         unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE;
 310         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 311         struct vm_area_struct *vma = walk->vma;
 312         int r;
 313
 314         flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY;
 315         flags |= hmm_vma_walk->write ? FAULT_FLAG_WRITE : 0;
 316         r = handle_mm_fault(vma, addr, flags);
 317         if (r & VM_FAULT_RETRY)
 318                 return -EBUSY;
 319         if (r & VM_FAULT_ERROR) {
 320                 *pfn = HMM_PFN_ERROR;
 321                 return -EFAULT;
 322         }
 323
 324         return -EAGAIN;
 325 }
 326
 327 static void hmm_pfns_special(hmm_pfn_t *pfns,
 328                              unsigned long addr,
 329                              unsigned long end)
 330 {
 331         for (; addr < end; addr += PAGE_SIZE, pfns++)
 332                 *pfns = HMM_PFN_SPECIAL;
 333 }
 334
 335 static int hmm_pfns_bad(unsigned long addr,
 336                         unsigned long end,
 337                         struct mm_walk *walk)
 338 {
 339         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 340         struct hmm_range *range = hmm_vma_walk->range;
 341         hmm_pfn_t *pfns = range->pfns;
 342         unsigned long i;
 343
 344         i = (addr - range->start) >> PAGE_SHIFT;
 345         for (; addr < end; addr += PAGE_SIZE, i++)
 346                 pfns[i] = HMM_PFN_ERROR;
 347
 348         return 0;
 349 }
 350
 351 static void hmm_pfns_clear(hmm_pfn_t *pfns,
 352                            unsigned long addr,
 353                            unsigned long end)
 354 {
 355         for (; addr < end; addr += PAGE_SIZE, pfns++)
 356                 *pfns = 0;
 357 }
 358
 359 static int hmm_vma_walk_hole(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         hmm_pfn_t *pfns = range->pfns;
 366         unsigned long i;
 367
 368         hmm_vma_walk->last = addr;
 369         i = (addr - range->start) >> PAGE_SHIFT;
 370         for (; addr < end; addr += PAGE_SIZE, i++) {
 371                 pfns[i] = HMM_PFN_EMPTY;
 372                 if (hmm_vma_walk->fault) {
 373                         int ret;
 374
 375                         ret = hmm_vma_do_fault(walk, addr, &pfns[i]);
 376                         if (ret != -EAGAIN)
 377                                 return ret;
 378                 }
 379         }
 380
 381         return hmm_vma_walk->fault ? -EAGAIN : 0;
 382 }
 383
 384 static int hmm_vma_walk_clear(unsigned long addr,
 385                               unsigned long end,
 386                               struct mm_walk *walk)
 387 {
 388         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 389         struct hmm_range *range = hmm_vma_walk->range;
 390         hmm_pfn_t *pfns = range->pfns;
 391         unsigned long i;
 392
 393         hmm_vma_walk->last = addr;
 394         i = (addr - range->start) >> PAGE_SHIFT;
 395         for (; addr < end; addr += PAGE_SIZE, i++) {
 396                 pfns[i] = 0;
 397                 if (hmm_vma_walk->fault) {
 398                         int ret;
 399
 400                         ret = hmm_vma_do_fault(walk, addr, &pfns[i]);
 401                         if (ret != -EAGAIN)
 402                                 return ret;
 403                 }
 404         }
 405
 406         return hmm_vma_walk->fault ? -EAGAIN : 0;
 407 }
 408
 409 static int hmm_vma_walk_pmd(pmd_t *pmdp,
 410                             unsigned long start,
 411                             unsigned long end,
 412                             struct mm_walk *walk)
 413 {
 414         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 415         struct hmm_range *range = hmm_vma_walk->range;
 416         struct vm_area_struct *vma = walk->vma;
 417         hmm_pfn_t *pfns = range->pfns;
 418         unsigned long addr = start, i;
 419         bool write_fault;
 420         hmm_pfn_t flag;
 421         pte_t *ptep;
 422
 423         i = (addr - range->start) >> PAGE_SHIFT;
 424         flag = vma->vm_flags & VM_READ ? HMM_PFN_READ : 0;
 425         write_fault = hmm_vma_walk->fault & hmm_vma_walk->write;
 426
 427 again:
 428         if (pmd_none(*pmdp))
 429                 return hmm_vma_walk_hole(start, end, walk);
 430
 431         if (pmd_huge(*pmdp) && vma->vm_flags & VM_HUGETLB)
 432                 return hmm_pfns_bad(start, end, walk);
 433
 434         if (pmd_devmap(*pmdp) || pmd_trans_huge(*pmdp)) {
 435                 unsigned long pfn;
 436                 pmd_t pmd;
 437
 438                 /*
 439                  * No need to take pmd_lock here, even if some other threads
 440                  * is splitting the huge pmd we will get that event through
 441                  * mmu_notifier callback.
 442                  *
 443                  * So just read pmd value and check again its a transparent
 444                  * huge or device mapping one and compute corresponding pfn
 445                  * values.
 446                  */
 447                 pmd = pmd_read_atomic(pmdp);
 448                 barrier();
 449                 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
 450                         goto again;
 451                 if (pmd_protnone(pmd))
 452                         return hmm_vma_walk_clear(start, end, walk);
 453
 454                 if (write_fault && !pmd_write(pmd))
 455                         return hmm_vma_walk_clear(start, end, walk);
 456
 457                 pfn = pmd_pfn(pmd) + pte_index(addr);
 458                 flag |= pmd_write(pmd) ? HMM_PFN_WRITE : 0;
 459                 for (; addr < end; addr += PAGE_SIZE, i++, pfn++)
 460                         pfns[i] = hmm_pfn_t_from_pfn(pfn) | flag;
 461                 return 0;
 462         }
 463
 464         if (pmd_bad(*pmdp))
 465                 return hmm_pfns_bad(start, end, walk);
 466
 467         ptep = pte_offset_map(pmdp, addr);
 468         for (; addr < end; addr += PAGE_SIZE, ptep++, i++) {
 469                 pte_t pte = *ptep;
 470
 471                 pfns[i] = 0;
 472
 473                 if (pte_none(pte)) {
 474                         pfns[i] = HMM_PFN_EMPTY;
 475                         if (hmm_vma_walk->fault)
 476                                 goto fault;
 477                         continue;
 478                 }
 479
 480                 if (!pte_present(pte)) {
 481                         swp_entry_t entry = pte_to_swp_entry(pte);
 482
 483                         if (!non_swap_entry(entry)) {
 484                                 if (hmm_vma_walk->fault)
 485                                         goto fault;
 486                                 continue;
 487                         }
 488
 489                         /*
 490                          * This is a special swap entry, ignore migration, use
 491                          * device and report anything else as error.
 492                          */
 493                         if (is_device_private_entry(entry)) {
 494                                 pfns[i] = hmm_pfn_t_from_pfn(swp_offset(entry));
 495                                 if (is_write_device_private_entry(entry)) {
 496                                         pfns[i] |= HMM_PFN_WRITE;
 497                                 } else if (write_fault)
 498                                         goto fault;
 499                                 pfns[i] |= HMM_PFN_DEVICE_UNADDRESSABLE;
 500                                 pfns[i] |= flag;
 501                         } else if (is_migration_entry(entry)) {
 502                                 if (hmm_vma_walk->fault) {
 503                                         pte_unmap(ptep);
 504                                         hmm_vma_walk->last = addr;
 505                                         migration_entry_wait(vma->vm_mm,
 506                                                              pmdp, addr);
 507                                         return -EAGAIN;
 508                                 }
 509                                 continue;
 510                         } else {
 511                                 /* Report error for everything else */
 512                                 pfns[i] = HMM_PFN_ERROR;
 513                         }
 514                         continue;
 515                 }
 516
 517                 if (write_fault && !pte_write(pte))
 518                         goto fault;
 519
 520                 pfns[i] = hmm_pfn_t_from_pfn(pte_pfn(pte)) | flag;
 521                 pfns[i] |= pte_write(pte) ? HMM_PFN_WRITE : 0;
 522                 continue;
 523
 524 fault:
 525                 pte_unmap(ptep);
 526                 /* Fault all pages in range */
 527                 return hmm_vma_walk_clear(start, end, walk);
 528         }
 529         pte_unmap(ptep - 1);
 530
 531         return 0;
 532 }
 533
 534 /*
 535  * hmm_vma_get_pfns() - snapshot CPU page table for a range of virtual addresses
 536  * @vma: virtual memory area containing the virtual address range
 537  * @range: used to track snapshot validity
 538  * @start: range virtual start address (inclusive)
 539  * @end: range virtual end address (exclusive)
 540  * @entries: array of hmm_pfn_t: provided by the caller, filled in by function
 541  * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, 0 success
 542  *
 543  * This snapshots the CPU page table for a range of virtual addresses. Snapshot
 544  * validity is tracked by range struct. See hmm_vma_range_done() for further
 545  * information.
 546  *
 547  * The range struct is initialized here. It tracks the CPU page table, but only
 548  * if the function returns success (0), in which case the caller must then call
 549  * hmm_vma_range_done() to stop CPU page table update tracking on this range.
 550  *
 551  * NOT CALLING hmm_vma_range_done() IF FUNCTION RETURNS 0 WILL LEAD TO SERIOUS
 552  * MEMORY CORRUPTION ! YOU HAVE BEEN WARNED !
 553  */
 554 int hmm_vma_get_pfns(struct vm_area_struct *vma,
 555                      struct hmm_range *range,
 556                      unsigned long start,
 557                      unsigned long end,
 558                      hmm_pfn_t *pfns)
 559 {
 560         struct hmm_vma_walk hmm_vma_walk;
 561         struct mm_walk mm_walk;
 562         struct hmm *hmm;
 563
 564         /* FIXME support hugetlb fs */
 565         if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) {
 566                 hmm_pfns_special(pfns, start, end);
 567                 return -EINVAL;
 568         }
 569
 570         /* Sanity check, this really should not happen ! */
 571         if (start < vma->vm_start || start >= vma->vm_end)
 572                 return -EINVAL;
 573         if (end < vma->vm_start || end > vma->vm_end)
 574                 return -EINVAL;
 575
 576         hmm = hmm_register(vma->vm_mm);
 577         if (!hmm)
 578                 return -ENOMEM;
 579         /* Caller must have registered a mirror, via hmm_mirror_register() ! */
 580         if (!hmm->mmu_notifier.ops)
 581                 return -EINVAL;
 582
 583         /* Initialize range to track CPU page table update */
 584         range->start = start;
 585         range->pfns = pfns;
 586         range->end = end;
 587         spin_lock(&hmm->lock);
 588         range->valid = true;
 589         list_add_rcu(&range->list, &hmm->ranges);
 590         spin_unlock(&hmm->lock);
 591
 592         hmm_vma_walk.fault = false;
 593         hmm_vma_walk.range = range;
 594         mm_walk.private = &hmm_vma_walk;
 595
 596         mm_walk.vma = vma;
 597         mm_walk.mm = vma->vm_mm;
 598         mm_walk.pte_entry = NULL;
 599         mm_walk.test_walk = NULL;
 600         mm_walk.hugetlb_entry = NULL;
 601         mm_walk.pmd_entry = hmm_vma_walk_pmd;
 602         mm_walk.pte_hole = hmm_vma_walk_hole;
 603
 604         walk_page_range(start, end, &mm_walk);
 605         return 0;
 606 }
 607 EXPORT_SYMBOL(hmm_vma_get_pfns);
 608
 609 /*
 610  * hmm_vma_range_done() - stop tracking change to CPU page table over a range
 611  * @vma: virtual memory area containing the virtual address range
 612  * @range: range being tracked
 613  * Returns: false if range data has been invalidated, true otherwise
 614  *
 615  * Range struct is used to track updates to the CPU page table after a call to
 616  * either hmm_vma_get_pfns() or hmm_vma_fault(). Once the device driver is done
 617  * using the data,  or wants to lock updates to the data it got from those
 618  * functions, it must call the hmm_vma_range_done() function, which will then
 619  * stop tracking CPU page table updates.
 620  *
 621  * Note that device driver must still implement general CPU page table update
 622  * tracking either by using hmm_mirror (see hmm_mirror_register()) or by using
 623  * the mmu_notifier API directly.
 624  *
 625  * CPU page table update tracking done through hmm_range is only temporary and
 626  * to be used while trying to duplicate CPU page table contents for a range of
 627  * virtual addresses.
 628  *
 629  * There are two ways to use this :
 630  * again:
 631  *   hmm_vma_get_pfns(vma, range, start, end, pfns); or hmm_vma_fault(...);
 632  *   trans = device_build_page_table_update_transaction(pfns);
 633  *   device_page_table_lock();
 634  *   if (!hmm_vma_range_done(vma, range)) {
 635  *     device_page_table_unlock();
 636  *     goto again;
 637  *   }
 638  *   device_commit_transaction(trans);
 639  *   device_page_table_unlock();
 640  *
 641  * Or:
 642  *   hmm_vma_get_pfns(vma, range, start, end, pfns); or hmm_vma_fault(...);
 643  *   device_page_table_lock();
 644  *   hmm_vma_range_done(vma, range);
 645  *   device_update_page_table(pfns);
 646  *   device_page_table_unlock();
 647  */
 648 bool hmm_vma_range_done(struct vm_area_struct *vma, struct hmm_range *range)
 649 {
 650         unsigned long npages = (range->end - range->start) >> PAGE_SHIFT;
 651         struct hmm *hmm;
 652
 653         if (range->end <= range->start) {
 654                 BUG();
 655                 return false;
 656         }
 657
 658         hmm = hmm_register(vma->vm_mm);
 659         if (!hmm) {
 660                 memset(range->pfns, 0, sizeof(*range->pfns) * npages);
 661                 return false;
 662         }
 663
 664         spin_lock(&hmm->lock);
 665         list_del_rcu(&range->list);
 666         spin_unlock(&hmm->lock);
 667
 668         return range->valid;
 669 }
 670 EXPORT_SYMBOL(hmm_vma_range_done);
 671
 672 /*
 673  * hmm_vma_fault() - try to fault some address in a virtual address range
 674  * @vma: virtual memory area containing the virtual address range
 675  * @range: use to track pfns array content validity
 676  * @start: fault range virtual start address (inclusive)
 677  * @end: fault range virtual end address (exclusive)
 678  * @pfns: array of hmm_pfn_t, only entry with fault flag set will be faulted
 679  * @write: is it a write fault
 680  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
 681  * Returns: 0 success, error otherwise (-EAGAIN means mmap_sem have been drop)
 682  *
 683  * This is similar to a regular CPU page fault except that it will not trigger
 684  * any memory migration if the memory being faulted is not accessible by CPUs.
 685  *
 686  * On error, for one virtual address in the range, the function will set the
 687  * hmm_pfn_t error flag for the corresponding pfn entry.
 688  *
 689  * Expected use pattern:
 690  * retry:
 691  *   down_read(&mm->mmap_sem);
 692  *   // Find vma and address device wants to fault, initialize hmm_pfn_t
 693  *   // array accordingly
 694  *   ret = hmm_vma_fault(vma, start, end, pfns, allow_retry);
 695  *   switch (ret) {
 696  *   case -EAGAIN:
 697  *     hmm_vma_range_done(vma, range);
 698  *     // You might want to rate limit or yield to play nicely, you may
 699  *     // also commit any valid pfn in the array assuming that you are
 700  *     // getting true from hmm_vma_range_monitor_end()
 701  *     goto retry;
 702  *   case 0:
 703  *     break;
 704  *   default:
 705  *     // Handle error !
 706  *     up_read(&mm->mmap_sem)
 707  *     return;
 708  *   }
 709  *   // Take device driver lock that serialize device page table update
 710  *   driver_lock_device_page_table_update();
 711  *   hmm_vma_range_done(vma, range);
 712  *   // Commit pfns we got from hmm_vma_fault()
 713  *   driver_unlock_device_page_table_update();
 714  *   up_read(&mm->mmap_sem)
 715  *
 716  * YOU MUST CALL hmm_vma_range_done() AFTER THIS FUNCTION RETURN SUCCESS (0)
 717  * BEFORE FREEING THE range struct OR YOU WILL HAVE SERIOUS MEMORY CORRUPTION !
 718  *
 719  * YOU HAVE BEEN WARNED !
 720  */
 721 int hmm_vma_fault(struct vm_area_struct *vma,
 722                   struct hmm_range *range,
 723                   unsigned long start,
 724                   unsigned long end,
 725                   hmm_pfn_t *pfns,
 726                   bool write,
 727                   bool block)
 728 {
 729         struct hmm_vma_walk hmm_vma_walk;
 730         struct mm_walk mm_walk;
 731         struct hmm *hmm;
 732         int ret;
 733
 734         /* Sanity check, this really should not happen ! */
 735         if (start < vma->vm_start || start >= vma->vm_end)
 736                 return -EINVAL;
 737         if (end < vma->vm_start || end > vma->vm_end)
 738                 return -EINVAL;
 739
 740         hmm = hmm_register(vma->vm_mm);
 741         if (!hmm) {
 742                 hmm_pfns_clear(pfns, start, end);
 743                 return -ENOMEM;
 744         }
 745         /* Caller must have registered a mirror using hmm_mirror_register() */
 746         if (!hmm->mmu_notifier.ops)
 747                 return -EINVAL;
 748
 749         /* Initialize range to track CPU page table update */
 750         range->start = start;
 751         range->pfns = pfns;
 752         range->end = end;
 753         spin_lock(&hmm->lock);
 754         range->valid = true;
 755         list_add_rcu(&range->list, &hmm->ranges);
 756         spin_unlock(&hmm->lock);
 757
 758         /* FIXME support hugetlb fs */
 759         if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) {
 760                 hmm_pfns_special(pfns, start, end);
 761                 return 0;
 762         }
 763
 764         hmm_vma_walk.fault = true;
 765         hmm_vma_walk.write = write;
 766         hmm_vma_walk.block = block;
 767         hmm_vma_walk.range = range;
 768         mm_walk.private = &hmm_vma_walk;
 769         hmm_vma_walk.last = range->start;
 770
 771         mm_walk.vma = vma;
 772         mm_walk.mm = vma->vm_mm;
 773         mm_walk.pte_entry = NULL;
 774         mm_walk.test_walk = NULL;
 775         mm_walk.hugetlb_entry = NULL;
 776         mm_walk.pmd_entry = hmm_vma_walk_pmd;
 777         mm_walk.pte_hole = hmm_vma_walk_hole;
 778
 779         do {
 780                 ret = walk_page_range(start, end, &mm_walk);
 781                 start = hmm_vma_walk.last;
 782         } while (ret == -EAGAIN);
 783
 784         if (ret) {
 785                 unsigned long i;
 786
 787                 i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
 788                 hmm_pfns_clear(&pfns[i], hmm_vma_walk.last, end);
 789                 hmm_vma_range_done(vma, range);
 790         }
 791         return ret;
 792 }
 793 EXPORT_SYMBOL(hmm_vma_fault);
 794 #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
 795
 796
 797 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) ||  IS_ENABLED(CONFIG_DEVICE_PUBLIC)
 798 struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
 799                                        unsigned long addr)
 800 {
 801         struct page *page;
 802
 803         page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
 804         if (!page)
 805                 return NULL;
 806         lock_page(page);
 807         return page;
 808 }
 809 EXPORT_SYMBOL(hmm_vma_alloc_locked_page);
 810
 811
 812 static void hmm_devmem_ref_release(struct percpu_ref *ref)
 813 {
 814         struct hmm_devmem *devmem;
 815
 816         devmem = container_of(ref, struct hmm_devmem, ref);
 817         complete(&devmem->completion);
 818 }
 819
 820 static void hmm_devmem_ref_exit(void *data)
 821 {
 822         struct percpu_ref *ref = data;
 823         struct hmm_devmem *devmem;
 824
 825         devmem = container_of(ref, struct hmm_devmem, ref);
 826         percpu_ref_exit(ref);
 827         devm_remove_action(devmem->device, &hmm_devmem_ref_exit, data);
 828 }
 829
 830 static void hmm_devmem_ref_kill(void *data)
 831 {
 832         struct percpu_ref *ref = data;
 833         struct hmm_devmem *devmem;
 834
 835         devmem = container_of(ref, struct hmm_devmem, ref);
 836         percpu_ref_kill(ref);
 837         wait_for_completion(&devmem->completion);
 838         devm_remove_action(devmem->device, &hmm_devmem_ref_kill, data);
 839 }
 840
 841 static int hmm_devmem_fault(struct vm_area_struct *vma,
 842                             unsigned long addr,
 843                             const struct page *page,
 844                             unsigned int flags,
 845                             pmd_t *pmdp)
 846 {
 847         struct hmm_devmem *devmem = page->pgmap->data;
 848
 849         return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp);
 850 }
 851
 852 static void hmm_devmem_free(struct page *page, void *data)
 853 {
 854         struct hmm_devmem *devmem = data;
 855
 856         devmem->ops->free(devmem, page);
 857 }
 858
 859 static DEFINE_MUTEX(hmm_devmem_lock);
 860 static RADIX_TREE(hmm_devmem_radix, GFP_KERNEL);
 861
 862 static void hmm_devmem_radix_release(struct resource *resource)
 863 {
 864         resource_size_t key, align_start, align_size;
 865
 866         align_start = resource->start & ~(PA_SECTION_SIZE - 1);
 867         align_size = ALIGN(resource_size(resource), PA_SECTION_SIZE);
 868
 869         mutex_lock(&hmm_devmem_lock);
 870         for (key = resource->start;
 871              key <= resource->end;
 872              key += PA_SECTION_SIZE)
 873                 radix_tree_delete(&hmm_devmem_radix, key >> PA_SECTION_SHIFT);
 874         mutex_unlock(&hmm_devmem_lock);
 875 }
 876
 877 static void hmm_devmem_release(struct device *dev, void *data)
 878 {
 879         struct hmm_devmem *devmem = data;
 880         struct resource *resource = devmem->resource;
 881         unsigned long start_pfn, npages;
 882         struct zone *zone;
 883         struct page *page;
 884
 885         if (percpu_ref_tryget_live(&devmem->ref)) {
 886                 dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
 887                 percpu_ref_put(&devmem->ref);
 888         }
 889
 890         /* pages are dead and unused, undo the arch mapping */
 891         start_pfn = (resource->start & ~(PA_SECTION_SIZE - 1)) >> PAGE_SHIFT;
 892         npages = ALIGN(resource_size(resource), PA_SECTION_SIZE) >> PAGE_SHIFT;
 893
 894         page = pfn_to_page(start_pfn);
 895         zone = page_zone(page);
 896
 897         mem_hotplug_begin();
 898         if (resource->desc == IORES_DESC_DEVICE_PRIVATE_MEMORY)
 899                 __remove_pages(zone, start_pfn, npages, NULL);
 900         else
 901                 arch_remove_memory(start_pfn << PAGE_SHIFT,
 902                                    npages << PAGE_SHIFT, NULL);
 903         mem_hotplug_done();
 904
 905         hmm_devmem_radix_release(resource);
 906 }
 907
 908 static struct hmm_devmem *hmm_devmem_find(resource_size_t phys)
 909 {
 910         WARN_ON_ONCE(!rcu_read_lock_held());
 911
 912         return radix_tree_lookup(&hmm_devmem_radix, phys >> PA_SECTION_SHIFT);
 913 }
 914
 915 static int hmm_devmem_pages_create(struct hmm_devmem *devmem)
 916 {
 917         resource_size_t key, align_start, align_size, align_end;
 918         struct device *device = devmem->device;
 919         int ret, nid, is_ram;
 920         unsigned long pfn;
 921
 922         align_start = devmem->resource->start & ~(PA_SECTION_SIZE - 1);
 923         align_size = ALIGN(devmem->resource->start +
 924                            resource_size(devmem->resource),
 925                            PA_SECTION_SIZE) - align_start;
 926
 927         is_ram = region_intersects(align_start, align_size,
 928                                    IORESOURCE_SYSTEM_RAM,
 929                                    IORES_DESC_NONE);
 930         if (is_ram == REGION_MIXED) {
 931                 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
 932                                 __func__, devmem->resource);
 933                 return -ENXIO;
 934         }
 935         if (is_ram == REGION_INTERSECTS)
 936                 return -ENXIO;
 937
 938         if (devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY)
 939                 devmem->pagemap.type = MEMORY_DEVICE_PUBLIC;
 940         else
 941                 devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
 942
 943         devmem->pagemap.res = *devmem->resource;
 944         devmem->pagemap.page_fault = hmm_devmem_fault;
 945         devmem->pagemap.page_free = hmm_devmem_free;
 946         devmem->pagemap.dev = devmem->device;
 947         devmem->pagemap.ref = &devmem->ref;
 948         devmem->pagemap.data = devmem;
 949
 950         mutex_lock(&hmm_devmem_lock);
 951         align_end = align_start + align_size - 1;
 952         for (key = align_start; key <= align_end; key += PA_SECTION_SIZE) {
 953                 struct hmm_devmem *dup;
 954
 955                 rcu_read_lock();
 956                 dup = hmm_devmem_find(key);
 957                 rcu_read_unlock();
 958                 if (dup) {
 959                         dev_err(device, "%s: collides with mapping for %s\n",
 960                                 __func__, dev_name(dup->device));
 961                         mutex_unlock(&hmm_devmem_lock);
 962                         ret = -EBUSY;
 963                         goto error;
 964                 }
 965                 ret = radix_tree_insert(&hmm_devmem_radix,
 966                                         key >> PA_SECTION_SHIFT,
 967                                         devmem);
 968                 if (ret) {
 969                         dev_err(device, "%s: failed: %d\n", __func__, ret);
 970                         mutex_unlock(&hmm_devmem_lock);
 971                         goto error_radix;
 972                 }
 973         }
 974         mutex_unlock(&hmm_devmem_lock);
 975
 976         nid = dev_to_node(device);
 977         if (nid < 0)
 978                 nid = numa_mem_id();
 979
 980         mem_hotplug_begin();
 981         /*
 982          * For device private memory we call add_pages() as we only need to
 983          * allocate and initialize struct page for the device memory. More-
 984          * over the device memory is un-accessible thus we do not want to
 985          * create a linear mapping for the memory like arch_add_memory()
 986          * would do.
 987          *
 988          * For device public memory, which is accesible by the CPU, we do
 989          * want the linear mapping and thus use arch_add_memory().
 990          */
 991         if (devmem->pagemap.type == MEMORY_DEVICE_PUBLIC)
 992                 ret = arch_add_memory(nid, align_start, align_size, NULL,
 993                                 false);
 994         else
 995                 ret = add_pages(nid, align_start >> PAGE_SHIFT,
 996                                 align_size >> PAGE_SHIFT, NULL, false);
 997         if (ret) {
 998                 mem_hotplug_done();
 999                 goto error_add_memory;
1000         }
1001         move_pfn_range_to_zone(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
1002                                 align_start >> PAGE_SHIFT,
1003                                 align_size >> PAGE_SHIFT, NULL);
1004         mem_hotplug_done();
1005
1006         for (pfn = devmem->pfn_first; pfn < devmem->pfn_last; pfn++) {
1007                 struct page *page = pfn_to_page(pfn);
1008
1009                 page->pgmap = &devmem->pagemap;
1010         }
1011         return 0;
1012
1013 error_add_memory:
1014         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
1015 error_radix:
1016         hmm_devmem_radix_release(devmem->resource);
1017 error:
1018         return ret;
1019 }
1020
1021 static int hmm_devmem_match(struct device *dev, void *data, void *match_data)
1022 {
1023         struct hmm_devmem *devmem = data;
1024
1025         return devmem->resource == match_data;
1026 }
1027
1028 static void hmm_devmem_pages_remove(struct hmm_devmem *devmem)
1029 {
1030         devres_release(devmem->device, &hmm_devmem_release,
1031                        &hmm_devmem_match, devmem->resource);
1032 }
1033
1034 /*
1035  * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory
1036  *
1037  * @ops: memory event device driver callback (see struct hmm_devmem_ops)
1038  * @device: device struct to bind the resource too
1039  * @size: size in bytes of the device memory to add
1040  * Returns: pointer to new hmm_devmem struct ERR_PTR otherwise
1041  *
1042  * This function first finds an empty range of physical address big enough to
1043  * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which
1044  * in turn allocates struct pages. It does not do anything beyond that; all
1045  * events affecting the memory will go through the various callbacks provided
1046  * by hmm_devmem_ops struct.
1047  *
1048  * Device driver should call this function during device initialization and
1049  * is then responsible of memory management. HMM only provides helpers.
1050  */
1051 struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
1052                                   struct device *device,
1053                                   unsigned long size)
1054 {
1055         struct hmm_devmem *devmem;
1056         resource_size_t addr;
1057         int ret;
1058
1059         static_branch_enable(&device_private_key);
1060
1061         devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem),
1062                                    GFP_KERNEL, dev_to_node(device));
1063         if (!devmem)
1064                 return ERR_PTR(-ENOMEM);
1065
1066         init_completion(&devmem->completion);
1067         devmem->pfn_first = -1UL;
1068         devmem->pfn_last = -1UL;
1069         devmem->resource = NULL;
1070         devmem->device = device;
1071         devmem->ops = ops;
1072
1073         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1074                               0, GFP_KERNEL);
1075         if (ret)
1076                 goto error_percpu_ref;
1077
1078         ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref);
1079         if (ret)
1080                 goto error_devm_add_action;
1081
1082         size = ALIGN(size, PA_SECTION_SIZE);
1083         addr = min((unsigned long)iomem_resource.end,
1084                    (1UL << MAX_PHYSMEM_BITS) - 1);
1085         addr = addr - size + 1UL;
1086
1087         /*
1088          * FIXME add a new helper to quickly walk resource tree and find free
1089          * range
1090          *
1091          * FIXME what about ioport_resource resource ?
1092          */
1093         for (; addr > size && addr >= iomem_resource.start; addr -= size) {
1094                 ret = region_intersects(addr, size, 0, IORES_DESC_NONE);
1095                 if (ret != REGION_DISJOINT)
1096                         continue;
1097
1098                 devmem->resource = devm_request_mem_region(device, addr, size,
1099                                                            dev_name(device));
1100                 if (!devmem->resource) {
1101                         ret = -ENOMEM;
1102                         goto error_no_resource;
1103                 }
1104                 break;
1105         }
1106         if (!devmem->resource) {
1107                 ret = -ERANGE;
1108                 goto error_no_resource;
1109         }
1110
1111         devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY;
1112         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1113         devmem->pfn_last = devmem->pfn_first +
1114                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1115
1116         ret = hmm_devmem_pages_create(devmem);
1117         if (ret)
1118                 goto error_pages;
1119
1120         devres_add(device, devmem);
1121
1122         ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref);
1123         if (ret) {
1124                 hmm_devmem_remove(devmem);
1125                 return ERR_PTR(ret);
1126         }
1127
1128         return devmem;
1129
1130 error_pages:
1131         devm_release_mem_region(device, devmem->resource->start,
1132                                 resource_size(devmem->resource));
1133 error_no_resource:
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);
1142
1143 struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops,
1144                                            struct device *device,
1145                                            struct resource *res)
1146 {
1147         struct hmm_devmem *devmem;
1148         int ret;
1149
1150         if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY)
1151                 return ERR_PTR(-EINVAL);
1152
1153         static_branch_enable(&device_private_key);
1154
1155         devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem),
1156                                    GFP_KERNEL, dev_to_node(device));
1157         if (!devmem)
1158                 return ERR_PTR(-ENOMEM);
1159
1160         init_completion(&devmem->completion);
1161         devmem->pfn_first = -1UL;
1162         devmem->pfn_last = -1UL;
1163         devmem->resource = res;
1164         devmem->device = device;
1165         devmem->ops = ops;
1166
1167         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1168                               0, GFP_KERNEL);
1169         if (ret)
1170                 goto error_percpu_ref;
1171
1172         ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref);
1173         if (ret)
1174                 goto error_devm_add_action;
1175
1176
1177         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1178         devmem->pfn_last = devmem->pfn_first +
1179                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1180
1181         ret = hmm_devmem_pages_create(devmem);
1182         if (ret)
1183                 goto error_devm_add_action;
1184
1185         devres_add(device, devmem);
1186
1187         ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref);
1188         if (ret) {
1189                 hmm_devmem_remove(devmem);
1190                 return ERR_PTR(ret);
1191         }
1192
1193         return devmem;
1194
1195 error_devm_add_action:
1196         hmm_devmem_ref_kill(&devmem->ref);
1197         hmm_devmem_ref_exit(&devmem->ref);
1198 error_percpu_ref:
1199         devres_free(devmem);
1200         return ERR_PTR(ret);
1201 }
1202 EXPORT_SYMBOL(hmm_devmem_add_resource);
1203
1204 /*
1205  * hmm_devmem_remove() - remove device memory (kill and free ZONE_DEVICE)
1206  *
1207  * @devmem: hmm_devmem struct use to track and manage the ZONE_DEVICE memory
1208  *
1209  * This will hot-unplug memory that was hotplugged by hmm_devmem_add on behalf
1210  * of the device driver. It will free struct page and remove the resource that
1211  * reserved the physical address range for this device memory.
1212  */
1213 void hmm_devmem_remove(struct hmm_devmem *devmem)
1214 {
1215         resource_size_t start, size;
1216         struct device *device;
1217         bool cdm = false;
1218
1219         if (!devmem)
1220                 return;
1221
1222         device = devmem->device;
1223         start = devmem->resource->start;
1224         size = resource_size(devmem->resource);
1225
1226         cdm = devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY;
1227         hmm_devmem_ref_kill(&devmem->ref);
1228         hmm_devmem_ref_exit(&devmem->ref);
1229         hmm_devmem_pages_remove(devmem);
1230
1231         if (!cdm)
1232                 devm_release_mem_region(device, start, size);
1233 }
1234 EXPORT_SYMBOL(hmm_devmem_remove);
1235
1236 /*
1237  * A device driver that wants to handle multiple devices memory through a
1238  * single fake device can use hmm_device to do so. This is purely a helper
1239  * and it is not needed to make use of any HMM functionality.
1240  */
1241 #define HMM_DEVICE_MAX 256
1242
1243 static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX);
1244 static DEFINE_SPINLOCK(hmm_device_lock);
1245 static struct class *hmm_device_class;
1246 static dev_t hmm_device_devt;
1247
1248 static void hmm_device_release(struct device *device)
1249 {
1250         struct hmm_device *hmm_device;
1251
1252         hmm_device = container_of(device, struct hmm_device, device);
1253         spin_lock(&hmm_device_lock);
1254         clear_bit(hmm_device->minor, hmm_device_mask);
1255         spin_unlock(&hmm_device_lock);
1256
1257         kfree(hmm_device);
1258 }
1259
1260 struct hmm_device *hmm_device_new(void *drvdata)
1261 {
1262         struct hmm_device *hmm_device;
1263
1264         hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL);
1265         if (!hmm_device)
1266                 return ERR_PTR(-ENOMEM);
1267
1268         spin_lock(&hmm_device_lock);
1269         hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX);
1270         if (hmm_device->minor >= HMM_DEVICE_MAX) {
1271                 spin_unlock(&hmm_device_lock);
1272                 kfree(hmm_device);
1273                 return ERR_PTR(-EBUSY);
1274         }
1275         set_bit(hmm_device->minor, hmm_device_mask);
1276         spin_unlock(&hmm_device_lock);
1277
1278         dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor);
1279         hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt),
1280                                         hmm_device->minor);
1281         hmm_device->device.release = hmm_device_release;
1282         dev_set_drvdata(&hmm_device->device, drvdata);
1283         hmm_device->device.class = hmm_device_class;
1284         device_initialize(&hmm_device->device);
1285
1286         return hmm_device;
1287 }
1288 EXPORT_SYMBOL(hmm_device_new);
1289
1290 void hmm_device_put(struct hmm_device *hmm_device)
1291 {
1292         put_device(&hmm_device->device);
1293 }
1294 EXPORT_SYMBOL(hmm_device_put);
1295
1296 static int __init hmm_init(void)
1297 {
1298         int ret;
1299
1300         ret = alloc_chrdev_region(&hmm_device_devt, 0,
1301                                   HMM_DEVICE_MAX,
1302                                   "hmm_device");
1303         if (ret)
1304                 return ret;
1305
1306         hmm_device_class = class_create(THIS_MODULE, "hmm_device");
1307         if (IS_ERR(hmm_device_class)) {
1308                 unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX);
1309                 return PTR_ERR(hmm_device_class);
1310         }
1311         return 0;
1312 }
1313
1314 device_initcall(hmm_init);
1315 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */