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                             uint64_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 int hmm_pfns_bad(unsigned long addr,
 328                         unsigned long end,
 329                         struct mm_walk *walk)
 330 {
 331         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 332         struct hmm_range *range = hmm_vma_walk->range;
 333         uint64_t *pfns = range->pfns;
 334         unsigned long i;
 335
 336         i = (addr - range->start) >> PAGE_SHIFT;
 337         for (; addr < end; addr += PAGE_SIZE, i++)
 338                 pfns[i] = HMM_PFN_ERROR;
 339
 340         return 0;
 341 }
 342
 343 /*
 344  * hmm_vma_walk_hole() - handle a range lacking valid pmd or pte(s)
 345  * @start: range virtual start address (inclusive)
 346  * @end: range virtual end address (exclusive)
 347  * @walk: mm_walk structure
 348  * Returns: 0 on success, -EAGAIN after page fault, or page fault error
 349  *
 350  * This function will be called whenever pmd_none() or pte_none() returns true,
 351  * or whenever there is no page directory covering the virtual address range.
 352  */
 353 static int hmm_vma_walk_hole(unsigned long addr,
 354                              unsigned long end,
 355                              struct mm_walk *walk)
 356 {
 357         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 358         struct hmm_range *range = hmm_vma_walk->range;
 359         uint64_t *pfns = range->pfns;
 360         unsigned long i;
 361
 362         hmm_vma_walk->last = addr;
 363         i = (addr - range->start) >> PAGE_SHIFT;
 364         for (; addr < end; addr += PAGE_SIZE, i++) {
 365                 pfns[i] = 0;
 366                 if (hmm_vma_walk->fault) {
 367                         int ret;
 368
 369                         ret = hmm_vma_do_fault(walk, addr, &pfns[i]);
 370                         if (ret != -EAGAIN)
 371                                 return ret;
 372                 }
 373         }
 374
 375         return hmm_vma_walk->fault ? -EAGAIN : 0;
 376 }
 377
 378 static int hmm_vma_handle_pmd(struct mm_walk *walk,
 379                               unsigned long addr,
 380                               unsigned long end,
 381                               uint64_t *pfns,
 382                               pmd_t pmd)
 383 {
 384         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 385         unsigned long pfn, i;
 386         uint64_t flag = 0;
 387
 388         if (pmd_protnone(pmd))
 389                 return hmm_vma_walk_hole(addr, end, walk);
 390
 391         if ((hmm_vma_walk->fault & hmm_vma_walk->write) && !pmd_write(pmd))
 392                 return hmm_vma_walk_hole(addr, end, walk);
 393
 394         pfn = pmd_pfn(pmd) + pte_index(addr);
 395         flag |= pmd_write(pmd) ? HMM_PFN_WRITE : 0;
 396         for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++)
 397                 pfns[i] = hmm_pfn_from_pfn(pfn) | flag;
 398         hmm_vma_walk->last = end;
 399         return 0;
 400 }
 401
 402 static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
 403                               unsigned long end, pmd_t *pmdp, pte_t *ptep,
 404                               uint64_t *pfn)
 405 {
 406         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 407         struct vm_area_struct *vma = walk->vma;
 408         pte_t pte = *ptep;
 409
 410         *pfn = 0;
 411
 412         if (pte_none(pte)) {
 413                 *pfn = 0;
 414                 if (hmm_vma_walk->fault)
 415                         goto fault;
 416                 return 0;
 417         }
 418
 419         if (!pte_present(pte)) {
 420                 swp_entry_t entry = pte_to_swp_entry(pte);
 421
 422                 if (!non_swap_entry(entry)) {
 423                         if (hmm_vma_walk->fault)
 424                                 goto fault;
 425                         return 0;
 426                 }
 427
 428                 /*
 429                  * This is a special swap entry, ignore migration, use
 430                  * device and report anything else as error.
 431                  */
 432                 if (is_device_private_entry(entry)) {
 433                         *pfn = hmm_pfn_from_pfn(swp_offset(entry));
 434                         if (is_write_device_private_entry(entry)) {
 435                                 *pfn |= HMM_PFN_WRITE;
 436                         } else if ((hmm_vma_walk->fault & hmm_vma_walk->write))
 437                                 goto fault;
 438                         *pfn |= HMM_PFN_DEVICE_PRIVATE;
 439                         return 0;
 440                 }
 441
 442                 if (is_migration_entry(entry)) {
 443                         if (hmm_vma_walk->fault) {
 444                                 pte_unmap(ptep);
 445                                 hmm_vma_walk->last = addr;
 446                                 migration_entry_wait(vma->vm_mm,
 447                                                 pmdp, addr);
 448                                 return -EAGAIN;
 449                         }
 450                         return 0;
 451                 }
 452
 453                 /* Report error for everything else */
 454                 *pfn = HMM_PFN_ERROR;
 455                 return -EFAULT;
 456         }
 457
 458         if ((hmm_vma_walk->fault & hmm_vma_walk->write) && !pte_write(pte))
 459                 goto fault;
 460
 461         *pfn = hmm_pfn_from_pfn(pte_pfn(pte));
 462         *pfn |= pte_write(pte) ? HMM_PFN_WRITE : 0;
 463         return 0;
 464
 465 fault:
 466         pte_unmap(ptep);
 467         /* Fault any virtual address we were asked to fault */
 468         return hmm_vma_walk_hole(addr, end, walk);
 469 }
 470
 471 static int hmm_vma_walk_pmd(pmd_t *pmdp,
 472                             unsigned long start,
 473                             unsigned long end,
 474                             struct mm_walk *walk)
 475 {
 476         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 477         struct hmm_range *range = hmm_vma_walk->range;
 478         uint64_t *pfns = range->pfns;
 479         unsigned long addr = start, i;
 480         pte_t *ptep;
 481
 482         i = (addr - range->start) >> PAGE_SHIFT;
 483
 484 again:
 485         if (pmd_none(*pmdp))
 486                 return hmm_vma_walk_hole(start, end, walk);
 487
 488         if (pmd_huge(*pmdp) && (range->vma->vm_flags & VM_HUGETLB))
 489                 return hmm_pfns_bad(start, end, walk);
 490
 491         if (pmd_devmap(*pmdp) || pmd_trans_huge(*pmdp)) {
 492                 pmd_t pmd;
 493
 494                 /*
 495                  * No need to take pmd_lock here, even if some other threads
 496                  * is splitting the huge pmd we will get that event through
 497                  * mmu_notifier callback.
 498                  *
 499                  * So just read pmd value and check again its a transparent
 500                  * huge or device mapping one and compute corresponding pfn
 501                  * values.
 502                  */
 503                 pmd = pmd_read_atomic(pmdp);
 504                 barrier();
 505                 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
 506                         goto again;
 507
 508                 return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd);
 509         }
 510
 511         if (pmd_bad(*pmdp))
 512                 return hmm_pfns_bad(start, end, walk);
 513
 514         ptep = pte_offset_map(pmdp, addr);
 515         for (; addr < end; addr += PAGE_SIZE, ptep++, i++) {
 516                 int r;
 517
 518                 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]);
 519                 if (r) {
 520                         /* hmm_vma_handle_pte() did unmap pte directory */
 521                         hmm_vma_walk->last = addr;
 522                         return r;
 523                 }
 524         }
 525         pte_unmap(ptep - 1);
 526
 527         hmm_vma_walk->last = addr;
 528         return 0;
 529 }
 530
 531 static void hmm_pfns_clear(uint64_t *pfns,
 532                            unsigned long addr,
 533                            unsigned long end)
 534 {
 535         for (; addr < end; addr += PAGE_SIZE, pfns++)
 536                 *pfns = 0;
 537 }
 538
 539 static void hmm_pfns_special(struct hmm_range *range)
 540 {
 541         unsigned long addr = range->start, i = 0;
 542
 543         for (; addr < range->end; addr += PAGE_SIZE, i++)
 544                 range->pfns[i] = HMM_PFN_SPECIAL;
 545 }
 546
 547 /*
 548  * hmm_vma_get_pfns() - snapshot CPU page table for a range of virtual addresses
 549  * @range: range being snapshotted
 550  * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid
 551  *          vma permission, 0 success
 552  *
 553  * This snapshots the CPU page table for a range of virtual addresses. Snapshot
 554  * validity is tracked by range struct. See hmm_vma_range_done() for further
 555  * information.
 556  *
 557  * The range struct is initialized here. It tracks the CPU page table, but only
 558  * if the function returns success (0), in which case the caller must then call
 559  * hmm_vma_range_done() to stop CPU page table update tracking on this range.
 560  *
 561  * NOT CALLING hmm_vma_range_done() IF FUNCTION RETURNS 0 WILL LEAD TO SERIOUS
 562  * MEMORY CORRUPTION ! YOU HAVE BEEN WARNED !
 563  */
 564 int hmm_vma_get_pfns(struct hmm_range *range)
 565 {
 566         struct vm_area_struct *vma = range->vma;
 567         struct hmm_vma_walk hmm_vma_walk;
 568         struct mm_walk mm_walk;
 569         struct hmm *hmm;
 570
 571         /* Sanity check, this really should not happen ! */
 572         if (range->start < vma->vm_start || range->start >= vma->vm_end)
 573                 return -EINVAL;
 574         if (range->end < vma->vm_start || range->end > vma->vm_end)
 575                 return -EINVAL;
 576
 577         hmm = hmm_register(vma->vm_mm);
 578         if (!hmm)
 579                 return -ENOMEM;
 580         /* Caller must have registered a mirror, via hmm_mirror_register() ! */
 581         if (!hmm->mmu_notifier.ops)
 582                 return -EINVAL;
 583
 584         /* FIXME support hugetlb fs */
 585         if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) {
 586                 hmm_pfns_special(range);
 587                 return -EINVAL;
 588         }
 589
 590         if (!(vma->vm_flags & VM_READ)) {
 591                 /*
 592                  * If vma do not allow read access, then assume that it does
 593                  * not allow write access, either. Architecture that allow
 594                  * write without read access are not supported by HMM, because
 595                  * operations such has atomic access would not work.
 596                  */
 597                 hmm_pfns_clear(range->pfns, range->start, range->end);
 598                 return -EPERM;
 599         }
 600
 601         /* Initialize range to track CPU page table update */
 602         spin_lock(&hmm->lock);
 603         range->valid = true;
 604         list_add_rcu(&range->list, &hmm->ranges);
 605         spin_unlock(&hmm->lock);
 606
 607         hmm_vma_walk.fault = false;
 608         hmm_vma_walk.range = range;
 609         mm_walk.private = &hmm_vma_walk;
 610
 611         mm_walk.vma = vma;
 612         mm_walk.mm = vma->vm_mm;
 613         mm_walk.pte_entry = NULL;
 614         mm_walk.test_walk = NULL;
 615         mm_walk.hugetlb_entry = NULL;
 616         mm_walk.pmd_entry = hmm_vma_walk_pmd;
 617         mm_walk.pte_hole = hmm_vma_walk_hole;
 618
 619         walk_page_range(range->start, range->end, &mm_walk);
 620         return 0;
 621 }
 622 EXPORT_SYMBOL(hmm_vma_get_pfns);
 623
 624 /*
 625  * hmm_vma_range_done() - stop tracking change to CPU page table over a range
 626  * @range: range being tracked
 627  * Returns: false if range data has been invalidated, true otherwise
 628  *
 629  * Range struct is used to track updates to the CPU page table after a call to
 630  * either hmm_vma_get_pfns() or hmm_vma_fault(). Once the device driver is done
 631  * using the data,  or wants to lock updates to the data it got from those
 632  * functions, it must call the hmm_vma_range_done() function, which will then
 633  * stop tracking CPU page table updates.
 634  *
 635  * Note that device driver must still implement general CPU page table update
 636  * tracking either by using hmm_mirror (see hmm_mirror_register()) or by using
 637  * the mmu_notifier API directly.
 638  *
 639  * CPU page table update tracking done through hmm_range is only temporary and
 640  * to be used while trying to duplicate CPU page table contents for a range of
 641  * virtual addresses.
 642  *
 643  * There are two ways to use this :
 644  * again:
 645  *   hmm_vma_get_pfns(range); or hmm_vma_fault(...);
 646  *   trans = device_build_page_table_update_transaction(pfns);
 647  *   device_page_table_lock();
 648  *   if (!hmm_vma_range_done(range)) {
 649  *     device_page_table_unlock();
 650  *     goto again;
 651  *   }
 652  *   device_commit_transaction(trans);
 653  *   device_page_table_unlock();
 654  *
 655  * Or:
 656  *   hmm_vma_get_pfns(range); or hmm_vma_fault(...);
 657  *   device_page_table_lock();
 658  *   hmm_vma_range_done(range);
 659  *   device_update_page_table(range->pfns);
 660  *   device_page_table_unlock();
 661  */
 662 bool hmm_vma_range_done(struct hmm_range *range)
 663 {
 664         unsigned long npages = (range->end - range->start) >> PAGE_SHIFT;
 665         struct hmm *hmm;
 666
 667         if (range->end <= range->start) {
 668                 BUG();
 669                 return false;
 670         }
 671
 672         hmm = hmm_register(range->vma->vm_mm);
 673         if (!hmm) {
 674                 memset(range->pfns, 0, sizeof(*range->pfns) * npages);
 675                 return false;
 676         }
 677
 678         spin_lock(&hmm->lock);
 679         list_del_rcu(&range->list);
 680         spin_unlock(&hmm->lock);
 681
 682         return range->valid;
 683 }
 684 EXPORT_SYMBOL(hmm_vma_range_done);
 685
 686 /*
 687  * hmm_vma_fault() - try to fault some address in a virtual address range
 688  * @range: range being faulted
 689  * @write: is it a write fault
 690  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
 691  * Returns: 0 success, error otherwise (-EAGAIN means mmap_sem have been drop)
 692  *
 693  * This is similar to a regular CPU page fault except that it will not trigger
 694  * any memory migration if the memory being faulted is not accessible by CPUs.
 695  *
 696  * On error, for one virtual address in the range, the function will mark the
 697  * corresponding HMM pfn entry with an error flag.
 698  *
 699  * Expected use pattern:
 700  * retry:
 701  *   down_read(&mm->mmap_sem);
 702  *   // Find vma and address device wants to fault, initialize hmm_pfn_t
 703  *   // array accordingly
 704  *   ret = hmm_vma_fault(range, write, block);
 705  *   switch (ret) {
 706  *   case -EAGAIN:
 707  *     hmm_vma_range_done(range);
 708  *     // You might want to rate limit or yield to play nicely, you may
 709  *     // also commit any valid pfn in the array assuming that you are
 710  *     // getting true from hmm_vma_range_monitor_end()
 711  *     goto retry;
 712  *   case 0:
 713  *     break;
 714  *   case -ENOMEM:
 715  *   case -EINVAL:
 716  *   case -EPERM:
 717  *   default:
 718  *     // Handle error !
 719  *     up_read(&mm->mmap_sem)
 720  *     return;
 721  *   }
 722  *   // Take device driver lock that serialize device page table update
 723  *   driver_lock_device_page_table_update();
 724  *   hmm_vma_range_done(range);
 725  *   // Commit pfns we got from hmm_vma_fault()
 726  *   driver_unlock_device_page_table_update();
 727  *   up_read(&mm->mmap_sem)
 728  *
 729  * YOU MUST CALL hmm_vma_range_done() AFTER THIS FUNCTION RETURN SUCCESS (0)
 730  * BEFORE FREEING THE range struct OR YOU WILL HAVE SERIOUS MEMORY CORRUPTION !
 731  *
 732  * YOU HAVE BEEN WARNED !
 733  */
 734 int hmm_vma_fault(struct hmm_range *range, bool write, bool block)
 735 {
 736         struct vm_area_struct *vma = range->vma;
 737         unsigned long start = range->start;
 738         struct hmm_vma_walk hmm_vma_walk;
 739         struct mm_walk mm_walk;
 740         struct hmm *hmm;
 741         int ret;
 742
 743         /* Sanity check, this really should not happen ! */
 744         if (range->start < vma->vm_start || range->start >= vma->vm_end)
 745                 return -EINVAL;
 746         if (range->end < vma->vm_start || range->end > vma->vm_end)
 747                 return -EINVAL;
 748
 749         hmm = hmm_register(vma->vm_mm);
 750         if (!hmm) {
 751                 hmm_pfns_clear(range->pfns, range->start, range->end);
 752                 return -ENOMEM;
 753         }
 754         /* Caller must have registered a mirror using hmm_mirror_register() */
 755         if (!hmm->mmu_notifier.ops)
 756                 return -EINVAL;
 757
 758         /* FIXME support hugetlb fs */
 759         if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) {
 760                 hmm_pfns_special(range);
 761                 return -EINVAL;
 762         }
 763
 764         if (!(vma->vm_flags & VM_READ)) {
 765                 /*
 766                  * If vma do not allow read access, then assume that it does
 767                  * not allow write access, either. Architecture that allow
 768                  * write without read access are not supported by HMM, because
 769                  * operations such has atomic access would not work.
 770                  */
 771                 hmm_pfns_clear(range->pfns, range->start, range->end);
 772                 return -EPERM;
 773         }
 774
 775         /* Initialize range to track CPU page table update */
 776         spin_lock(&hmm->lock);
 777         range->valid = true;
 778         list_add_rcu(&range->list, &hmm->ranges);
 779         spin_unlock(&hmm->lock);
 780
 781         hmm_vma_walk.fault = true;
 782         hmm_vma_walk.write = write;
 783         hmm_vma_walk.block = block;
 784         hmm_vma_walk.range = range;
 785         mm_walk.private = &hmm_vma_walk;
 786         hmm_vma_walk.last = range->start;
 787
 788         mm_walk.vma = vma;
 789         mm_walk.mm = vma->vm_mm;
 790         mm_walk.pte_entry = NULL;
 791         mm_walk.test_walk = NULL;
 792         mm_walk.hugetlb_entry = NULL;
 793         mm_walk.pmd_entry = hmm_vma_walk_pmd;
 794         mm_walk.pte_hole = hmm_vma_walk_hole;
 795
 796         do {
 797                 ret = walk_page_range(start, range->end, &mm_walk);
 798                 start = hmm_vma_walk.last;
 799         } while (ret == -EAGAIN);
 800
 801         if (ret) {
 802                 unsigned long i;
 803
 804                 i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
 805                 hmm_pfns_clear(&range->pfns[i], hmm_vma_walk.last, range->end);
 806                 hmm_vma_range_done(range);
 807         }
 808         return ret;
 809 }
 810 EXPORT_SYMBOL(hmm_vma_fault);
 811 #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
 812
 813
 814 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) ||  IS_ENABLED(CONFIG_DEVICE_PUBLIC)
 815 struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
 816                                        unsigned long addr)
 817 {
 818         struct page *page;
 819
 820         page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
 821         if (!page)
 822                 return NULL;
 823         lock_page(page);
 824         return page;
 825 }
 826 EXPORT_SYMBOL(hmm_vma_alloc_locked_page);
 827
 828
 829 static void hmm_devmem_ref_release(struct percpu_ref *ref)
 830 {
 831         struct hmm_devmem *devmem;
 832
 833         devmem = container_of(ref, struct hmm_devmem, ref);
 834         complete(&devmem->completion);
 835 }
 836
 837 static void hmm_devmem_ref_exit(void *data)
 838 {
 839         struct percpu_ref *ref = data;
 840         struct hmm_devmem *devmem;
 841
 842         devmem = container_of(ref, struct hmm_devmem, ref);
 843         percpu_ref_exit(ref);
 844         devm_remove_action(devmem->device, &hmm_devmem_ref_exit, data);
 845 }
 846
 847 static void hmm_devmem_ref_kill(void *data)
 848 {
 849         struct percpu_ref *ref = data;
 850         struct hmm_devmem *devmem;
 851
 852         devmem = container_of(ref, struct hmm_devmem, ref);
 853         percpu_ref_kill(ref);
 854         wait_for_completion(&devmem->completion);
 855         devm_remove_action(devmem->device, &hmm_devmem_ref_kill, data);
 856 }
 857
 858 static int hmm_devmem_fault(struct vm_area_struct *vma,
 859                             unsigned long addr,
 860                             const struct page *page,
 861                             unsigned int flags,
 862                             pmd_t *pmdp)
 863 {
 864         struct hmm_devmem *devmem = page->pgmap->data;
 865
 866         return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp);
 867 }
 868
 869 static void hmm_devmem_free(struct page *page, void *data)
 870 {
 871         struct hmm_devmem *devmem = data;
 872
 873         devmem->ops->free(devmem, page);
 874 }
 875
 876 static DEFINE_MUTEX(hmm_devmem_lock);
 877 static RADIX_TREE(hmm_devmem_radix, GFP_KERNEL);
 878
 879 static void hmm_devmem_radix_release(struct resource *resource)
 880 {
 881         resource_size_t key, align_start, align_size;
 882
 883         align_start = resource->start & ~(PA_SECTION_SIZE - 1);
 884         align_size = ALIGN(resource_size(resource), PA_SECTION_SIZE);
 885
 886         mutex_lock(&hmm_devmem_lock);
 887         for (key = resource->start;
 888              key <= resource->end;
 889              key += PA_SECTION_SIZE)
 890                 radix_tree_delete(&hmm_devmem_radix, key >> PA_SECTION_SHIFT);
 891         mutex_unlock(&hmm_devmem_lock);
 892 }
 893
 894 static void hmm_devmem_release(struct device *dev, void *data)
 895 {
 896         struct hmm_devmem *devmem = data;
 897         struct resource *resource = devmem->resource;
 898         unsigned long start_pfn, npages;
 899         struct zone *zone;
 900         struct page *page;
 901
 902         if (percpu_ref_tryget_live(&devmem->ref)) {
 903                 dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
 904                 percpu_ref_put(&devmem->ref);
 905         }
 906
 907         /* pages are dead and unused, undo the arch mapping */
 908         start_pfn = (resource->start & ~(PA_SECTION_SIZE - 1)) >> PAGE_SHIFT;
 909         npages = ALIGN(resource_size(resource), PA_SECTION_SIZE) >> PAGE_SHIFT;
 910
 911         page = pfn_to_page(start_pfn);
 912         zone = page_zone(page);
 913
 914         mem_hotplug_begin();
 915         if (resource->desc == IORES_DESC_DEVICE_PRIVATE_MEMORY)
 916                 __remove_pages(zone, start_pfn, npages, NULL);
 917         else
 918                 arch_remove_memory(start_pfn << PAGE_SHIFT,
 919                                    npages << PAGE_SHIFT, NULL);
 920         mem_hotplug_done();
 921
 922         hmm_devmem_radix_release(resource);
 923 }
 924
 925 static struct hmm_devmem *hmm_devmem_find(resource_size_t phys)
 926 {
 927         WARN_ON_ONCE(!rcu_read_lock_held());
 928
 929         return radix_tree_lookup(&hmm_devmem_radix, phys >> PA_SECTION_SHIFT);
 930 }
 931
 932 static int hmm_devmem_pages_create(struct hmm_devmem *devmem)
 933 {
 934         resource_size_t key, align_start, align_size, align_end;
 935         struct device *device = devmem->device;
 936         int ret, nid, is_ram;
 937         unsigned long pfn;
 938
 939         align_start = devmem->resource->start & ~(PA_SECTION_SIZE - 1);
 940         align_size = ALIGN(devmem->resource->start +
 941                            resource_size(devmem->resource),
 942                            PA_SECTION_SIZE) - align_start;
 943
 944         is_ram = region_intersects(align_start, align_size,
 945                                    IORESOURCE_SYSTEM_RAM,
 946                                    IORES_DESC_NONE);
 947         if (is_ram == REGION_MIXED) {
 948                 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
 949                                 __func__, devmem->resource);
 950                 return -ENXIO;
 951         }
 952         if (is_ram == REGION_INTERSECTS)
 953                 return -ENXIO;
 954
 955         if (devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY)
 956                 devmem->pagemap.type = MEMORY_DEVICE_PUBLIC;
 957         else
 958                 devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
 959
 960         devmem->pagemap.res = *devmem->resource;
 961         devmem->pagemap.page_fault = hmm_devmem_fault;
 962         devmem->pagemap.page_free = hmm_devmem_free;
 963         devmem->pagemap.dev = devmem->device;
 964         devmem->pagemap.ref = &devmem->ref;
 965         devmem->pagemap.data = devmem;
 966
 967         mutex_lock(&hmm_devmem_lock);
 968         align_end = align_start + align_size - 1;
 969         for (key = align_start; key <= align_end; key += PA_SECTION_SIZE) {
 970                 struct hmm_devmem *dup;
 971
 972                 rcu_read_lock();
 973                 dup = hmm_devmem_find(key);
 974                 rcu_read_unlock();
 975                 if (dup) {
 976                         dev_err(device, "%s: collides with mapping for %s\n",
 977                                 __func__, dev_name(dup->device));
 978                         mutex_unlock(&hmm_devmem_lock);
 979                         ret = -EBUSY;
 980                         goto error;
 981                 }
 982                 ret = radix_tree_insert(&hmm_devmem_radix,
 983                                         key >> PA_SECTION_SHIFT,
 984                                         devmem);
 985                 if (ret) {
 986                         dev_err(device, "%s: failed: %d\n", __func__, ret);
 987                         mutex_unlock(&hmm_devmem_lock);
 988                         goto error_radix;
 989                 }
 990         }
 991         mutex_unlock(&hmm_devmem_lock);
 992
 993         nid = dev_to_node(device);
 994         if (nid < 0)
 995                 nid = numa_mem_id();
 996
 997         mem_hotplug_begin();
 998         /*
 999          * For device private memory we call add_pages() as we only need to
1000          * allocate and initialize struct page for the device memory. More-
1001          * over the device memory is un-accessible thus we do not want to
1002          * create a linear mapping for the memory like arch_add_memory()
1003          * would do.
1004          *
1005          * For device public memory, which is accesible by the CPU, we do
1006          * want the linear mapping and thus use arch_add_memory().
1007          */
1008         if (devmem->pagemap.type == MEMORY_DEVICE_PUBLIC)
1009                 ret = arch_add_memory(nid, align_start, align_size, NULL,
1010                                 false);
1011         else
1012                 ret = add_pages(nid, align_start >> PAGE_SHIFT,
1013                                 align_size >> PAGE_SHIFT, NULL, false);
1014         if (ret) {
1015                 mem_hotplug_done();
1016                 goto error_add_memory;
1017         }
1018         move_pfn_range_to_zone(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
1019                                 align_start >> PAGE_SHIFT,
1020                                 align_size >> PAGE_SHIFT, NULL);
1021         mem_hotplug_done();
1022
1023         for (pfn = devmem->pfn_first; pfn < devmem->pfn_last; pfn++) {
1024                 struct page *page = pfn_to_page(pfn);
1025
1026                 page->pgmap = &devmem->pagemap;
1027         }
1028         return 0;
1029
1030 error_add_memory:
1031         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
1032 error_radix:
1033         hmm_devmem_radix_release(devmem->resource);
1034 error:
1035         return ret;
1036 }
1037
1038 static int hmm_devmem_match(struct device *dev, void *data, void *match_data)
1039 {
1040         struct hmm_devmem *devmem = data;
1041
1042         return devmem->resource == match_data;
1043 }
1044
1045 static void hmm_devmem_pages_remove(struct hmm_devmem *devmem)
1046 {
1047         devres_release(devmem->device, &hmm_devmem_release,
1048                        &hmm_devmem_match, devmem->resource);
1049 }
1050
1051 /*
1052  * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory
1053  *
1054  * @ops: memory event device driver callback (see struct hmm_devmem_ops)
1055  * @device: device struct to bind the resource too
1056  * @size: size in bytes of the device memory to add
1057  * Returns: pointer to new hmm_devmem struct ERR_PTR otherwise
1058  *
1059  * This function first finds an empty range of physical address big enough to
1060  * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which
1061  * in turn allocates struct pages. It does not do anything beyond that; all
1062  * events affecting the memory will go through the various callbacks provided
1063  * by hmm_devmem_ops struct.
1064  *
1065  * Device driver should call this function during device initialization and
1066  * is then responsible of memory management. HMM only provides helpers.
1067  */
1068 struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
1069                                   struct device *device,
1070                                   unsigned long size)
1071 {
1072         struct hmm_devmem *devmem;
1073         resource_size_t addr;
1074         int ret;
1075
1076         static_branch_enable(&device_private_key);
1077
1078         devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem),
1079                                    GFP_KERNEL, dev_to_node(device));
1080         if (!devmem)
1081                 return ERR_PTR(-ENOMEM);
1082
1083         init_completion(&devmem->completion);
1084         devmem->pfn_first = -1UL;
1085         devmem->pfn_last = -1UL;
1086         devmem->resource = NULL;
1087         devmem->device = device;
1088         devmem->ops = ops;
1089
1090         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1091                               0, GFP_KERNEL);
1092         if (ret)
1093                 goto error_percpu_ref;
1094
1095         ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref);
1096         if (ret)
1097                 goto error_devm_add_action;
1098
1099         size = ALIGN(size, PA_SECTION_SIZE);
1100         addr = min((unsigned long)iomem_resource.end,
1101                    (1UL << MAX_PHYSMEM_BITS) - 1);
1102         addr = addr - size + 1UL;
1103
1104         /*
1105          * FIXME add a new helper to quickly walk resource tree and find free
1106          * range
1107          *
1108          * FIXME what about ioport_resource resource ?
1109          */
1110         for (; addr > size && addr >= iomem_resource.start; addr -= size) {
1111                 ret = region_intersects(addr, size, 0, IORES_DESC_NONE);
1112                 if (ret != REGION_DISJOINT)
1113                         continue;
1114
1115                 devmem->resource = devm_request_mem_region(device, addr, size,
1116                                                            dev_name(device));
1117                 if (!devmem->resource) {
1118                         ret = -ENOMEM;
1119                         goto error_no_resource;
1120                 }
1121                 break;
1122         }
1123         if (!devmem->resource) {
1124                 ret = -ERANGE;
1125                 goto error_no_resource;
1126         }
1127
1128         devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY;
1129         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1130         devmem->pfn_last = devmem->pfn_first +
1131                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1132
1133         ret = hmm_devmem_pages_create(devmem);
1134         if (ret)
1135                 goto error_pages;
1136
1137         devres_add(device, devmem);
1138
1139         ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref);
1140         if (ret) {
1141                 hmm_devmem_remove(devmem);
1142                 return ERR_PTR(ret);
1143         }
1144
1145         return devmem;
1146
1147 error_pages:
1148         devm_release_mem_region(device, devmem->resource->start,
1149                                 resource_size(devmem->resource));
1150 error_no_resource:
1151 error_devm_add_action:
1152         hmm_devmem_ref_kill(&devmem->ref);
1153         hmm_devmem_ref_exit(&devmem->ref);
1154 error_percpu_ref:
1155         devres_free(devmem);
1156         return ERR_PTR(ret);
1157 }
1158 EXPORT_SYMBOL(hmm_devmem_add);
1159
1160 struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops,
1161                                            struct device *device,
1162                                            struct resource *res)
1163 {
1164         struct hmm_devmem *devmem;
1165         int ret;
1166
1167         if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY)
1168                 return ERR_PTR(-EINVAL);
1169
1170         static_branch_enable(&device_private_key);
1171
1172         devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem),
1173                                    GFP_KERNEL, dev_to_node(device));
1174         if (!devmem)
1175                 return ERR_PTR(-ENOMEM);
1176
1177         init_completion(&devmem->completion);
1178         devmem->pfn_first = -1UL;
1179         devmem->pfn_last = -1UL;
1180         devmem->resource = res;
1181         devmem->device = device;
1182         devmem->ops = ops;
1183
1184         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1185                               0, GFP_KERNEL);
1186         if (ret)
1187                 goto error_percpu_ref;
1188
1189         ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref);
1190         if (ret)
1191                 goto error_devm_add_action;
1192
1193
1194         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1195         devmem->pfn_last = devmem->pfn_first +
1196                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1197
1198         ret = hmm_devmem_pages_create(devmem);
1199         if (ret)
1200                 goto error_devm_add_action;
1201
1202         devres_add(device, devmem);
1203
1204         ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref);
1205         if (ret) {
1206                 hmm_devmem_remove(devmem);
1207                 return ERR_PTR(ret);
1208         }
1209
1210         return devmem;
1211
1212 error_devm_add_action:
1213         hmm_devmem_ref_kill(&devmem->ref);
1214         hmm_devmem_ref_exit(&devmem->ref);
1215 error_percpu_ref:
1216         devres_free(devmem);
1217         return ERR_PTR(ret);
1218 }
1219 EXPORT_SYMBOL(hmm_devmem_add_resource);
1220
1221 /*
1222  * hmm_devmem_remove() - remove device memory (kill and free ZONE_DEVICE)
1223  *
1224  * @devmem: hmm_devmem struct use to track and manage the ZONE_DEVICE memory
1225  *
1226  * This will hot-unplug memory that was hotplugged by hmm_devmem_add on behalf
1227  * of the device driver. It will free struct page and remove the resource that
1228  * reserved the physical address range for this device memory.
1229  */
1230 void hmm_devmem_remove(struct hmm_devmem *devmem)
1231 {
1232         resource_size_t start, size;
1233         struct device *device;
1234         bool cdm = false;
1235
1236         if (!devmem)
1237                 return;
1238
1239         device = devmem->device;
1240         start = devmem->resource->start;
1241         size = resource_size(devmem->resource);
1242
1243         cdm = devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY;
1244         hmm_devmem_ref_kill(&devmem->ref);
1245         hmm_devmem_ref_exit(&devmem->ref);
1246         hmm_devmem_pages_remove(devmem);
1247
1248         if (!cdm)
1249                 devm_release_mem_region(device, start, size);
1250 }
1251 EXPORT_SYMBOL(hmm_devmem_remove);
1252
1253 /*
1254  * A device driver that wants to handle multiple devices memory through a
1255  * single fake device can use hmm_device to do so. This is purely a helper
1256  * and it is not needed to make use of any HMM functionality.
1257  */
1258 #define HMM_DEVICE_MAX 256
1259
1260 static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX);
1261 static DEFINE_SPINLOCK(hmm_device_lock);
1262 static struct class *hmm_device_class;
1263 static dev_t hmm_device_devt;
1264
1265 static void hmm_device_release(struct device *device)
1266 {
1267         struct hmm_device *hmm_device;
1268
1269         hmm_device = container_of(device, struct hmm_device, device);
1270         spin_lock(&hmm_device_lock);
1271         clear_bit(hmm_device->minor, hmm_device_mask);
1272         spin_unlock(&hmm_device_lock);
1273
1274         kfree(hmm_device);
1275 }
1276
1277 struct hmm_device *hmm_device_new(void *drvdata)
1278 {
1279         struct hmm_device *hmm_device;
1280
1281         hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL);
1282         if (!hmm_device)
1283                 return ERR_PTR(-ENOMEM);
1284
1285         spin_lock(&hmm_device_lock);
1286         hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX);
1287         if (hmm_device->minor >= HMM_DEVICE_MAX) {
1288                 spin_unlock(&hmm_device_lock);
1289                 kfree(hmm_device);
1290                 return ERR_PTR(-EBUSY);
1291         }
1292         set_bit(hmm_device->minor, hmm_device_mask);
1293         spin_unlock(&hmm_device_lock);
1294
1295         dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor);
1296         hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt),
1297                                         hmm_device->minor);
1298         hmm_device->device.release = hmm_device_release;
1299         dev_set_drvdata(&hmm_device->device, drvdata);
1300         hmm_device->device.class = hmm_device_class;
1301         device_initialize(&hmm_device->device);
1302
1303         return hmm_device;
1304 }
1305 EXPORT_SYMBOL(hmm_device_new);
1306
1307 void hmm_device_put(struct hmm_device *hmm_device)
1308 {
1309         put_device(&hmm_device->device);
1310 }
1311 EXPORT_SYMBOL(hmm_device_put);
1312
1313 static int __init hmm_init(void)
1314 {
1315         int ret;
1316
1317         ret = alloc_chrdev_region(&hmm_device_devt, 0,
1318                                   HMM_DEVICE_MAX,
1319                                   "hmm_device");
1320         if (ret)
1321                 return ret;
1322
1323         hmm_device_class = class_create(THIS_MODULE, "hmm_device");
1324         if (IS_ERR(hmm_device_class)) {
1325                 unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX);
1326                 return PTR_ERR(hmm_device_class);
1327         }
1328         return 0;
1329 }
1330
1331 device_initcall(hmm_init);
1332 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */