mm/hmm.c

   1 /*
   2  * Copyright 2013 Red Hat Inc.
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * Authors: Jérôme Glisse <jglisse@redhat.com>
  15  */
  16 /*
  17  * Refer to include/linux/hmm.h for information about heterogeneous memory
  18  * management or HMM for short.
  19  */
  20 #include <linux/mm.h>
  21 #include <linux/hmm.h>
  22 #include <linux/init.h>
  23 #include <linux/rmap.h>
  24 #include <linux/swap.h>
  25 #include <linux/slab.h>
  26 #include <linux/sched.h>
  27 #include <linux/mmzone.h>
  28 #include <linux/pagemap.h>
  29 #include <linux/swapops.h>
  30 #include <linux/hugetlb.h>
  31 #include <linux/memremap.h>
  32 #include <linux/jump_label.h>
  33 #include <linux/mmu_notifier.h>
  34 #include <linux/memory_hotplug.h>
  35
  36 #define PA_SECTION_SIZE (1UL << PA_SECTION_SHIFT)
  37
  38 #if IS_ENABLED(CONFIG_HMM_MIRROR)
  39 static const struct mmu_notifier_ops hmm_mmu_notifier_ops;
  40
  41 /*
  42  * struct hmm - HMM per mm struct
  43  *
  44  * @mm: mm struct this HMM struct is bound to
  45  * @lock: lock protecting ranges list
  46  * @sequence: we track updates to the CPU page table with a sequence number
  47  * @ranges: list of range being snapshotted
  48  * @mirrors: list of mirrors for this mm
  49  * @mmu_notifier: mmu notifier to track updates to CPU page table
  50  * @mirrors_sem: read/write semaphore protecting the mirrors list
  51  */
  52 struct hmm {
  53         struct mm_struct        *mm;
  54         spinlock_t              lock;
  55         atomic_t                sequence;
  56         struct list_head        ranges;
  57         struct list_head        mirrors;
  58         struct mmu_notifier     mmu_notifier;
  59         struct rw_semaphore     mirrors_sem;
  60 };
  61
  62 /*
  63  * hmm_register - register HMM against an mm (HMM internal)
  64  *
  65  * @mm: mm struct to attach to
  66  *
  67  * This is not intended to be used directly by device drivers. It allocates an
  68  * HMM struct if mm does not have one, and initializes it.
  69  */
  70 static struct hmm *hmm_register(struct mm_struct *mm)
  71 {
  72         struct hmm *hmm = READ_ONCE(mm->hmm);
  73         bool cleanup = false;
  74
  75         /*
  76          * The hmm struct can only be freed once the mm_struct goes away,
  77          * hence we should always have pre-allocated an new hmm struct
  78          * above.
  79          */
  80         if (hmm)
  81                 return hmm;
  82
  83         hmm = kmalloc(sizeof(*hmm), GFP_KERNEL);
  84         if (!hmm)
  85                 return NULL;
  86         INIT_LIST_HEAD(&hmm->mirrors);
  87         init_rwsem(&hmm->mirrors_sem);
  88         atomic_set(&hmm->sequence, 0);
  89         hmm->mmu_notifier.ops = NULL;
  90         INIT_LIST_HEAD(&hmm->ranges);
  91         spin_lock_init(&hmm->lock);
  92         hmm->mm = mm;
  93
  94         spin_lock(&mm->page_table_lock);
  95         if (!mm->hmm)
  96                 mm->hmm = hmm;
  97         else
  98                 cleanup = true;
  99         spin_unlock(&mm->page_table_lock);
 100
 101         if (cleanup)
 102                 goto error;
 103
 104         /*
 105          * We should only get here if hold the mmap_sem in write mode ie on
 106          * registration of first mirror through hmm_mirror_register()
 107          */
 108         hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops;
 109         if (__mmu_notifier_register(&hmm->mmu_notifier, mm))
 110                 goto error_mm;
 111
 112         return mm->hmm;
 113
 114 error_mm:
 115         spin_lock(&mm->page_table_lock);
 116         if (mm->hmm == hmm)
 117                 mm->hmm = NULL;
 118         spin_unlock(&mm->page_table_lock);
 119 error:
 120         kfree(hmm);
 121         return NULL;
 122 }
 123
 124 void hmm_mm_destroy(struct mm_struct *mm)
 125 {
 126         kfree(mm->hmm);
 127 }
 128
 129 static void hmm_invalidate_range(struct hmm *hmm,
 130                                  enum hmm_update_type action,
 131                                  unsigned long start,
 132                                  unsigned long end)
 133 {
 134         struct hmm_mirror *mirror;
 135         struct hmm_range *range;
 136
 137         spin_lock(&hmm->lock);
 138         list_for_each_entry(range, &hmm->ranges, list) {
 139                 unsigned long addr, idx, npages;
 140
 141                 if (end < range->start || start >= range->end)
 142                         continue;
 143
 144                 range->valid = false;
 145                 addr = max(start, range->start);
 146                 idx = (addr - range->start) >> PAGE_SHIFT;
 147                 npages = (min(range->end, end) - addr) >> PAGE_SHIFT;
 148                 memset(&range->pfns[idx], 0, sizeof(*range->pfns) * npages);
 149         }
 150         spin_unlock(&hmm->lock);
 151
 152         down_read(&hmm->mirrors_sem);
 153         list_for_each_entry(mirror, &hmm->mirrors, list)
 154                 mirror->ops->sync_cpu_device_pagetables(mirror, action,
 155                                                         start, end);
 156         up_read(&hmm->mirrors_sem);
 157 }
 158
 159 static void hmm_release(struct mmu_notifier *mn, struct mm_struct *mm)
 160 {
 161         struct hmm_mirror *mirror;
 162         struct hmm *hmm = mm->hmm;
 163
 164         down_write(&hmm->mirrors_sem);
 165         mirror = list_first_entry_or_null(&hmm->mirrors, struct hmm_mirror,
 166                                           list);
 167         while (mirror) {
 168                 list_del_init(&mirror->list);
 169                 if (mirror->ops->release) {
 170                         /*
 171                          * Drop mirrors_sem so callback can wait on any pending
 172                          * work that might itself trigger mmu_notifier callback
 173                          * and thus would deadlock with us.
 174                          */
 175                         up_write(&hmm->mirrors_sem);
 176                         mirror->ops->release(mirror);
 177                         down_write(&hmm->mirrors_sem);
 178                 }
 179                 mirror = list_first_entry_or_null(&hmm->mirrors,
 180                                                   struct hmm_mirror, list);
 181         }
 182         up_write(&hmm->mirrors_sem);
 183 }
 184
 185 static int hmm_invalidate_range_start(struct mmu_notifier *mn,
 186                                        struct mm_struct *mm,
 187                                        unsigned long start,
 188                                        unsigned long end,
 189                                        bool blockable)
 190 {
 191         struct hmm *hmm = mm->hmm;
 192
 193         VM_BUG_ON(!hmm);
 194
 195         atomic_inc(&hmm->sequence);
 196
 197         return 0;
 198 }
 199
 200 static void hmm_invalidate_range_end(struct mmu_notifier *mn,
 201                                      struct mm_struct *mm,
 202                                      unsigned long start,
 203                                      unsigned long end)
 204 {
 205         struct hmm *hmm = mm->hmm;
 206
 207         VM_BUG_ON(!hmm);
 208
 209         hmm_invalidate_range(mm->hmm, HMM_UPDATE_INVALIDATE, start, end);
 210 }
 211
 212 static const struct mmu_notifier_ops hmm_mmu_notifier_ops = {
 213         .release                = hmm_release,
 214         .invalidate_range_start = hmm_invalidate_range_start,
 215         .invalidate_range_end   = hmm_invalidate_range_end,
 216 };
 217
 218 /*
 219  * hmm_mirror_register() - register a mirror against an mm
 220  *
 221  * @mirror: new mirror struct to register
 222  * @mm: mm to register against
 223  *
 224  * To start mirroring a process address space, the device driver must register
 225  * an HMM mirror struct.
 226  *
 227  * THE mm->mmap_sem MUST BE HELD IN WRITE MODE !
 228  */
 229 int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm)
 230 {
 231         /* Sanity check */
 232         if (!mm || !mirror || !mirror->ops)
 233                 return -EINVAL;
 234
 235 again:
 236         mirror->hmm = hmm_register(mm);
 237         if (!mirror->hmm)
 238                 return -ENOMEM;
 239
 240         down_write(&mirror->hmm->mirrors_sem);
 241         if (mirror->hmm->mm == NULL) {
 242                 /*
 243                  * A racing hmm_mirror_unregister() is about to destroy the hmm
 244                  * struct. Try again to allocate a new one.
 245                  */
 246                 up_write(&mirror->hmm->mirrors_sem);
 247                 mirror->hmm = NULL;
 248                 goto again;
 249         } else {
 250                 list_add(&mirror->list, &mirror->hmm->mirrors);
 251                 up_write(&mirror->hmm->mirrors_sem);
 252         }
 253
 254         return 0;
 255 }
 256 EXPORT_SYMBOL(hmm_mirror_register);
 257
 258 /*
 259  * hmm_mirror_unregister() - unregister a mirror
 260  *
 261  * @mirror: new mirror struct to register
 262  *
 263  * Stop mirroring a process address space, and cleanup.
 264  */
 265 void hmm_mirror_unregister(struct hmm_mirror *mirror)
 266 {
 267         bool should_unregister = false;
 268         struct mm_struct *mm;
 269         struct hmm *hmm;
 270
 271         if (mirror->hmm == NULL)
 272                 return;
 273
 274         hmm = mirror->hmm;
 275         down_write(&hmm->mirrors_sem);
 276         list_del_init(&mirror->list);
 277         should_unregister = list_empty(&hmm->mirrors);
 278         mirror->hmm = NULL;
 279         mm = hmm->mm;
 280         hmm->mm = NULL;
 281         up_write(&hmm->mirrors_sem);
 282
 283         if (!should_unregister || mm == NULL)
 284                 return;
 285
 286         mmu_notifier_unregister_no_release(&hmm->mmu_notifier, mm);
 287
 288         spin_lock(&mm->page_table_lock);
 289         if (mm->hmm == hmm)
 290                 mm->hmm = NULL;
 291         spin_unlock(&mm->page_table_lock);
 292
 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 };
 303
 304 static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr,
 305                             bool write_fault, uint64_t *pfn)
 306 {
 307         unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE;
 308         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 309         struct hmm_range *range = hmm_vma_walk->range;
 310         struct vm_area_struct *vma = walk->vma;
 311         vm_fault_t ret;
 312
 313         flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY;
 314         flags |= write_fault ? FAULT_FLAG_WRITE : 0;
 315         ret = handle_mm_fault(vma, addr, flags);
 316         if (ret & VM_FAULT_RETRY)
 317                 return -EBUSY;
 318         if (ret & VM_FAULT_ERROR) {
 319                 *pfn = range->values[HMM_PFN_ERROR];
 320                 return -EFAULT;
 321         }
 322
 323         return -EAGAIN;
 324 }
 325
 326 static int hmm_pfns_bad(unsigned long addr,
 327                         unsigned long end,
 328                         struct mm_walk *walk)
 329 {
 330         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 331         struct hmm_range *range = hmm_vma_walk->range;
 332         uint64_t *pfns = range->pfns;
 333         unsigned long i;
 334
 335         i = (addr - range->start) >> PAGE_SHIFT;
 336         for (; addr < end; addr += PAGE_SIZE, i++)
 337                 pfns[i] = range->values[HMM_PFN_ERROR];
 338
 339         return 0;
 340 }
 341
 342 /*
 343  * hmm_vma_walk_hole() - handle a range lacking valid pmd or pte(s)
 344  * @start: range virtual start address (inclusive)
 345  * @end: range virtual end address (exclusive)
 346  * @fault: should we fault or not ?
 347  * @write_fault: write fault ?
 348  * @walk: mm_walk structure
 349  * Returns: 0 on success, -EAGAIN after page fault, or page fault error
 350  *
 351  * This function will be called whenever pmd_none() or pte_none() returns true,
 352  * or whenever there is no page directory covering the virtual address range.
 353  */
 354 static int hmm_vma_walk_hole_(unsigned long addr, unsigned long end,
 355                               bool fault, bool write_fault,
 356                               struct mm_walk *walk)
 357 {
 358         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 359         struct hmm_range *range = hmm_vma_walk->range;
 360         uint64_t *pfns = range->pfns;
 361         unsigned long i;
 362
 363         hmm_vma_walk->last = addr;
 364         i = (addr - range->start) >> PAGE_SHIFT;
 365         for (; addr < end; addr += PAGE_SIZE, i++) {
 366                 pfns[i] = range->values[HMM_PFN_NONE];
 367                 if (fault || write_fault) {
 368                         int ret;
 369
 370                         ret = hmm_vma_do_fault(walk, addr, write_fault,
 371                                                &pfns[i]);
 372                         if (ret != -EAGAIN)
 373                                 return ret;
 374                 }
 375         }
 376
 377         return (fault || write_fault) ? -EAGAIN : 0;
 378 }
 379
 380 static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 381                                       uint64_t pfns, uint64_t cpu_flags,
 382                                       bool *fault, bool *write_fault)
 383 {
 384         struct hmm_range *range = hmm_vma_walk->range;
 385
 386         *fault = *write_fault = false;
 387         if (!hmm_vma_walk->fault)
 388                 return;
 389
 390         /* We aren't ask to do anything ... */
 391         if (!(pfns & range->flags[HMM_PFN_VALID]))
 392                 return;
 393         /* If this is device memory than only fault if explicitly requested */
 394         if ((cpu_flags & range->flags[HMM_PFN_DEVICE_PRIVATE])) {
 395                 /* Do we fault on device memory ? */
 396                 if (pfns & range->flags[HMM_PFN_DEVICE_PRIVATE]) {
 397                         *write_fault = pfns & range->flags[HMM_PFN_WRITE];
 398                         *fault = true;
 399                 }
 400                 return;
 401         }
 402
 403         /* If CPU page table is not valid then we need to fault */
 404         *fault = !(cpu_flags & range->flags[HMM_PFN_VALID]);
 405         /* Need to write fault ? */
 406         if ((pfns & range->flags[HMM_PFN_WRITE]) &&
 407             !(cpu_flags & range->flags[HMM_PFN_WRITE])) {
 408                 *write_fault = true;
 409                 *fault = true;
 410         }
 411 }
 412
 413 static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 414                                  const uint64_t *pfns, unsigned long npages,
 415                                  uint64_t cpu_flags, bool *fault,
 416                                  bool *write_fault)
 417 {
 418         unsigned long i;
 419
 420         if (!hmm_vma_walk->fault) {
 421                 *fault = *write_fault = false;
 422                 return;
 423         }
 424
 425         for (i = 0; i < npages; ++i) {
 426                 hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags,
 427                                    fault, write_fault);
 428                 if ((*fault) || (*write_fault))
 429                         return;
 430         }
 431 }
 432
 433 static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
 434                              struct mm_walk *walk)
 435 {
 436         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 437         struct hmm_range *range = hmm_vma_walk->range;
 438         bool fault, write_fault;
 439         unsigned long i, npages;
 440         uint64_t *pfns;
 441
 442         i = (addr - range->start) >> PAGE_SHIFT;
 443         npages = (end - addr) >> PAGE_SHIFT;
 444         pfns = &range->pfns[i];
 445         hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 446                              0, &fault, &write_fault);
 447         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 448 }
 449
 450 static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd)
 451 {
 452         if (pmd_protnone(pmd))
 453                 return 0;
 454         return pmd_write(pmd) ? range->flags[HMM_PFN_VALID] |
 455                                 range->flags[HMM_PFN_WRITE] :
 456                                 range->flags[HMM_PFN_VALID];
 457 }
 458
 459 static int hmm_vma_handle_pmd(struct mm_walk *walk,
 460                               unsigned long addr,
 461                               unsigned long end,
 462                               uint64_t *pfns,
 463                               pmd_t pmd)
 464 {
 465         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 466         struct hmm_range *range = hmm_vma_walk->range;
 467         unsigned long pfn, npages, i;
 468         bool fault, write_fault;
 469         uint64_t cpu_flags;
 470
 471         npages = (end - addr) >> PAGE_SHIFT;
 472         cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
 473         hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags,
 474                              &fault, &write_fault);
 475
 476         if (pmd_protnone(pmd) || fault || write_fault)
 477                 return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 478
 479         pfn = pmd_pfn(pmd) + pte_index(addr);
 480         for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++)
 481                 pfns[i] = hmm_pfn_from_pfn(range, pfn) | cpu_flags;
 482         hmm_vma_walk->last = end;
 483         return 0;
 484 }
 485
 486 static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte)
 487 {
 488         if (pte_none(pte) || !pte_present(pte))
 489                 return 0;
 490         return pte_write(pte) ? range->flags[HMM_PFN_VALID] |
 491                                 range->flags[HMM_PFN_WRITE] :
 492                                 range->flags[HMM_PFN_VALID];
 493 }
 494
 495 static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
 496                               unsigned long end, pmd_t *pmdp, pte_t *ptep,
 497                               uint64_t *pfn)
 498 {
 499         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 500         struct hmm_range *range = hmm_vma_walk->range;
 501         struct vm_area_struct *vma = walk->vma;
 502         bool fault, write_fault;
 503         uint64_t cpu_flags;
 504         pte_t pte = *ptep;
 505         uint64_t orig_pfn = *pfn;
 506
 507         *pfn = range->values[HMM_PFN_NONE];
 508         cpu_flags = pte_to_hmm_pfn_flags(range, pte);
 509         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 510                            &fault, &write_fault);
 511
 512         if (pte_none(pte)) {
 513                 if (fault || write_fault)
 514                         goto fault;
 515                 return 0;
 516         }
 517
 518         if (!pte_present(pte)) {
 519                 swp_entry_t entry = pte_to_swp_entry(pte);
 520
 521                 if (!non_swap_entry(entry)) {
 522                         if (fault || write_fault)
 523                                 goto fault;
 524                         return 0;
 525                 }
 526
 527                 /*
 528                  * This is a special swap entry, ignore migration, use
 529                  * device and report anything else as error.
 530                  */
 531                 if (is_device_private_entry(entry)) {
 532                         cpu_flags = range->flags[HMM_PFN_VALID] |
 533                                 range->flags[HMM_PFN_DEVICE_PRIVATE];
 534                         cpu_flags |= is_write_device_private_entry(entry) ?
 535                                 range->flags[HMM_PFN_WRITE] : 0;
 536                         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 537                                            &fault, &write_fault);
 538                         if (fault || write_fault)
 539                                 goto fault;
 540                         *pfn = hmm_pfn_from_pfn(range, swp_offset(entry));
 541                         *pfn |= cpu_flags;
 542                         return 0;
 543                 }
 544
 545                 if (is_migration_entry(entry)) {
 546                         if (fault || write_fault) {
 547                                 pte_unmap(ptep);
 548                                 hmm_vma_walk->last = addr;
 549                                 migration_entry_wait(vma->vm_mm,
 550                                                      pmdp, addr);
 551                                 return -EAGAIN;
 552                         }
 553                         return 0;
 554                 }
 555
 556                 /* Report error for everything else */
 557                 *pfn = range->values[HMM_PFN_ERROR];
 558                 return -EFAULT;
 559         }
 560
 561         if (fault || write_fault)
 562                 goto fault;
 563
 564         *pfn = hmm_pfn_from_pfn(range, pte_pfn(pte)) | cpu_flags;
 565         return 0;
 566
 567 fault:
 568         pte_unmap(ptep);
 569         /* Fault any virtual address we were asked to fault */
 570         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 571 }
 572
 573 static int hmm_vma_walk_pmd(pmd_t *pmdp,
 574                             unsigned long start,
 575                             unsigned long end,
 576                             struct mm_walk *walk)
 577 {
 578         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 579         struct hmm_range *range = hmm_vma_walk->range;
 580         struct vm_area_struct *vma = walk->vma;
 581         uint64_t *pfns = range->pfns;
 582         unsigned long addr = start, i;
 583         pte_t *ptep;
 584         pmd_t pmd;
 585
 586
 587 again:
 588         pmd = READ_ONCE(*pmdp);
 589         if (pmd_none(pmd))
 590                 return hmm_vma_walk_hole(start, end, walk);
 591
 592         if (pmd_huge(pmd) && (range->vma->vm_flags & VM_HUGETLB))
 593                 return hmm_pfns_bad(start, end, walk);
 594
 595         if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
 596                 bool fault, write_fault;
 597                 unsigned long npages;
 598                 uint64_t *pfns;
 599
 600                 i = (addr - range->start) >> PAGE_SHIFT;
 601                 npages = (end - addr) >> PAGE_SHIFT;
 602                 pfns = &range->pfns[i];
 603
 604                 hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 605                                      0, &fault, &write_fault);
 606                 if (fault || write_fault) {
 607                         hmm_vma_walk->last = addr;
 608                         pmd_migration_entry_wait(vma->vm_mm, pmdp);
 609                         return -EAGAIN;
 610                 }
 611                 return 0;
 612         } else if (!pmd_present(pmd))
 613                 return hmm_pfns_bad(start, end, walk);
 614
 615         if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
 616                 /*
 617                  * No need to take pmd_lock here, even if some other threads
 618                  * is splitting the huge pmd we will get that event through
 619                  * mmu_notifier callback.
 620                  *
 621                  * So just read pmd value and check again its a transparent
 622                  * huge or device mapping one and compute corresponding pfn
 623                  * values.
 624                  */
 625                 pmd = pmd_read_atomic(pmdp);
 626                 barrier();
 627                 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
 628                         goto again;
 629
 630                 i = (addr - range->start) >> PAGE_SHIFT;
 631                 return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd);
 632         }
 633
 634         /*
 635          * We have handled all the valid case above ie either none, migration,
 636          * huge or transparent huge. At this point either it is a valid pmd
 637          * entry pointing to pte directory or it is a bad pmd that will not
 638          * recover.
 639          */
 640         if (pmd_bad(pmd))
 641                 return hmm_pfns_bad(start, end, walk);
 642
 643         ptep = pte_offset_map(pmdp, addr);
 644         i = (addr - range->start) >> PAGE_SHIFT;
 645         for (; addr < end; addr += PAGE_SIZE, ptep++, i++) {
 646                 int r;
 647
 648                 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]);
 649                 if (r) {
 650                         /* hmm_vma_handle_pte() did unmap pte directory */
 651                         hmm_vma_walk->last = addr;
 652                         return r;
 653                 }
 654         }
 655         pte_unmap(ptep - 1);
 656
 657         hmm_vma_walk->last = addr;
 658         return 0;
 659 }
 660
 661 static void hmm_pfns_clear(struct hmm_range *range,
 662                            uint64_t *pfns,
 663                            unsigned long addr,
 664                            unsigned long end)
 665 {
 666         for (; addr < end; addr += PAGE_SIZE, pfns++)
 667                 *pfns = range->values[HMM_PFN_NONE];
 668 }
 669
 670 static void hmm_pfns_special(struct hmm_range *range)
 671 {
 672         unsigned long addr = range->start, i = 0;
 673
 674         for (; addr < range->end; addr += PAGE_SIZE, i++)
 675                 range->pfns[i] = range->values[HMM_PFN_SPECIAL];
 676 }
 677
 678 /*
 679  * hmm_vma_get_pfns() - snapshot CPU page table for a range of virtual addresses
 680  * @range: range being snapshotted
 681  * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid
 682  *          vma permission, 0 success
 683  *
 684  * This snapshots the CPU page table for a range of virtual addresses. Snapshot
 685  * validity is tracked by range struct. See hmm_vma_range_done() for further
 686  * information.
 687  *
 688  * The range struct is initialized here. It tracks the CPU page table, but only
 689  * if the function returns success (0), in which case the caller must then call
 690  * hmm_vma_range_done() to stop CPU page table update tracking on this range.
 691  *
 692  * NOT CALLING hmm_vma_range_done() IF FUNCTION RETURNS 0 WILL LEAD TO SERIOUS
 693  * MEMORY CORRUPTION ! YOU HAVE BEEN WARNED !
 694  */
 695 int hmm_vma_get_pfns(struct hmm_range *range)
 696 {
 697         struct vm_area_struct *vma = range->vma;
 698         struct hmm_vma_walk hmm_vma_walk;
 699         struct mm_walk mm_walk;
 700         struct hmm *hmm;
 701
 702         /* Sanity check, this really should not happen ! */
 703         if (range->start < vma->vm_start || range->start >= vma->vm_end)
 704                 return -EINVAL;
 705         if (range->end < vma->vm_start || range->end > vma->vm_end)
 706                 return -EINVAL;
 707
 708         hmm = hmm_register(vma->vm_mm);
 709         if (!hmm)
 710                 return -ENOMEM;
 711         /* Caller must have registered a mirror, via hmm_mirror_register() ! */
 712         if (!hmm->mmu_notifier.ops)
 713                 return -EINVAL;
 714
 715         /* FIXME support hugetlb fs */
 716         if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) ||
 717                         vma_is_dax(vma)) {
 718                 hmm_pfns_special(range);
 719                 return -EINVAL;
 720         }
 721
 722         if (!(vma->vm_flags & VM_READ)) {
 723                 /*
 724                  * If vma do not allow read access, then assume that it does
 725                  * not allow write access, either. Architecture that allow
 726                  * write without read access are not supported by HMM, because
 727                  * operations such has atomic access would not work.
 728                  */
 729                 hmm_pfns_clear(range, range->pfns, range->start, range->end);
 730                 return -EPERM;
 731         }
 732
 733         /* Initialize range to track CPU page table update */
 734         spin_lock(&hmm->lock);
 735         range->valid = true;
 736         list_add_rcu(&range->list, &hmm->ranges);
 737         spin_unlock(&hmm->lock);
 738
 739         hmm_vma_walk.fault = false;
 740         hmm_vma_walk.range = range;
 741         mm_walk.private = &hmm_vma_walk;
 742
 743         mm_walk.vma = vma;
 744         mm_walk.mm = vma->vm_mm;
 745         mm_walk.pte_entry = NULL;
 746         mm_walk.test_walk = NULL;
 747         mm_walk.hugetlb_entry = NULL;
 748         mm_walk.pmd_entry = hmm_vma_walk_pmd;
 749         mm_walk.pte_hole = hmm_vma_walk_hole;
 750
 751         walk_page_range(range->start, range->end, &mm_walk);
 752         return 0;
 753 }
 754 EXPORT_SYMBOL(hmm_vma_get_pfns);
 755
 756 /*
 757  * hmm_vma_range_done() - stop tracking change to CPU page table over a range
 758  * @range: range being tracked
 759  * Returns: false if range data has been invalidated, true otherwise
 760  *
 761  * Range struct is used to track updates to the CPU page table after a call to
 762  * either hmm_vma_get_pfns() or hmm_vma_fault(). Once the device driver is done
 763  * using the data,  or wants to lock updates to the data it got from those
 764  * functions, it must call the hmm_vma_range_done() function, which will then
 765  * stop tracking CPU page table updates.
 766  *
 767  * Note that device driver must still implement general CPU page table update
 768  * tracking either by using hmm_mirror (see hmm_mirror_register()) or by using
 769  * the mmu_notifier API directly.
 770  *
 771  * CPU page table update tracking done through hmm_range is only temporary and
 772  * to be used while trying to duplicate CPU page table contents for a range of
 773  * virtual addresses.
 774  *
 775  * There are two ways to use this :
 776  * again:
 777  *   hmm_vma_get_pfns(range); or hmm_vma_fault(...);
 778  *   trans = device_build_page_table_update_transaction(pfns);
 779  *   device_page_table_lock();
 780  *   if (!hmm_vma_range_done(range)) {
 781  *     device_page_table_unlock();
 782  *     goto again;
 783  *   }
 784  *   device_commit_transaction(trans);
 785  *   device_page_table_unlock();
 786  *
 787  * Or:
 788  *   hmm_vma_get_pfns(range); or hmm_vma_fault(...);
 789  *   device_page_table_lock();
 790  *   hmm_vma_range_done(range);
 791  *   device_update_page_table(range->pfns);
 792  *   device_page_table_unlock();
 793  */
 794 bool hmm_vma_range_done(struct hmm_range *range)
 795 {
 796         unsigned long npages = (range->end - range->start) >> PAGE_SHIFT;
 797         struct hmm *hmm;
 798
 799         if (range->end <= range->start) {
 800                 BUG();
 801                 return false;
 802         }
 803
 804         hmm = hmm_register(range->vma->vm_mm);
 805         if (!hmm) {
 806                 memset(range->pfns, 0, sizeof(*range->pfns) * npages);
 807                 return false;
 808         }
 809
 810         spin_lock(&hmm->lock);
 811         list_del_rcu(&range->list);
 812         spin_unlock(&hmm->lock);
 813
 814         return range->valid;
 815 }
 816 EXPORT_SYMBOL(hmm_vma_range_done);
 817
 818 /*
 819  * hmm_vma_fault() - try to fault some address in a virtual address range
 820  * @range: range being faulted
 821  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
 822  * Returns: 0 success, error otherwise (-EAGAIN means mmap_sem have been drop)
 823  *
 824  * This is similar to a regular CPU page fault except that it will not trigger
 825  * any memory migration if the memory being faulted is not accessible by CPUs.
 826  *
 827  * On error, for one virtual address in the range, the function will mark the
 828  * corresponding HMM pfn entry with an error flag.
 829  *
 830  * Expected use pattern:
 831  * retry:
 832  *   down_read(&mm->mmap_sem);
 833  *   // Find vma and address device wants to fault, initialize hmm_pfn_t
 834  *   // array accordingly
 835  *   ret = hmm_vma_fault(range, write, block);
 836  *   switch (ret) {
 837  *   case -EAGAIN:
 838  *     hmm_vma_range_done(range);
 839  *     // You might want to rate limit or yield to play nicely, you may
 840  *     // also commit any valid pfn in the array assuming that you are
 841  *     // getting true from hmm_vma_range_monitor_end()
 842  *     goto retry;
 843  *   case 0:
 844  *     break;
 845  *   case -ENOMEM:
 846  *   case -EINVAL:
 847  *   case -EPERM:
 848  *   default:
 849  *     // Handle error !
 850  *     up_read(&mm->mmap_sem)
 851  *     return;
 852  *   }
 853  *   // Take device driver lock that serialize device page table update
 854  *   driver_lock_device_page_table_update();
 855  *   hmm_vma_range_done(range);
 856  *   // Commit pfns we got from hmm_vma_fault()
 857  *   driver_unlock_device_page_table_update();
 858  *   up_read(&mm->mmap_sem)
 859  *
 860  * YOU MUST CALL hmm_vma_range_done() AFTER THIS FUNCTION RETURN SUCCESS (0)
 861  * BEFORE FREEING THE range struct OR YOU WILL HAVE SERIOUS MEMORY CORRUPTION !
 862  *
 863  * YOU HAVE BEEN WARNED !
 864  */
 865 int hmm_vma_fault(struct hmm_range *range, bool block)
 866 {
 867         struct vm_area_struct *vma = range->vma;
 868         unsigned long start = range->start;
 869         struct hmm_vma_walk hmm_vma_walk;
 870         struct mm_walk mm_walk;
 871         struct hmm *hmm;
 872         int ret;
 873
 874         /* Sanity check, this really should not happen ! */
 875         if (range->start < vma->vm_start || range->start >= vma->vm_end)
 876                 return -EINVAL;
 877         if (range->end < vma->vm_start || range->end > vma->vm_end)
 878                 return -EINVAL;
 879
 880         hmm = hmm_register(vma->vm_mm);
 881         if (!hmm) {
 882                 hmm_pfns_clear(range, range->pfns, range->start, range->end);
 883                 return -ENOMEM;
 884         }
 885         /* Caller must have registered a mirror using hmm_mirror_register() */
 886         if (!hmm->mmu_notifier.ops)
 887                 return -EINVAL;
 888
 889         /* FIXME support hugetlb fs */
 890         if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) ||
 891                         vma_is_dax(vma)) {
 892                 hmm_pfns_special(range);
 893                 return -EINVAL;
 894         }
 895
 896         if (!(vma->vm_flags & VM_READ)) {
 897                 /*
 898                  * If vma do not allow read access, then assume that it does
 899                  * not allow write access, either. Architecture that allow
 900                  * write without read access are not supported by HMM, because
 901                  * operations such has atomic access would not work.
 902                  */
 903                 hmm_pfns_clear(range, range->pfns, range->start, range->end);
 904                 return -EPERM;
 905         }
 906
 907         /* Initialize range to track CPU page table update */
 908         spin_lock(&hmm->lock);
 909         range->valid = true;
 910         list_add_rcu(&range->list, &hmm->ranges);
 911         spin_unlock(&hmm->lock);
 912
 913         hmm_vma_walk.fault = true;
 914         hmm_vma_walk.block = block;
 915         hmm_vma_walk.range = range;
 916         mm_walk.private = &hmm_vma_walk;
 917         hmm_vma_walk.last = range->start;
 918
 919         mm_walk.vma = vma;
 920         mm_walk.mm = vma->vm_mm;
 921         mm_walk.pte_entry = NULL;
 922         mm_walk.test_walk = NULL;
 923         mm_walk.hugetlb_entry = NULL;
 924         mm_walk.pmd_entry = hmm_vma_walk_pmd;
 925         mm_walk.pte_hole = hmm_vma_walk_hole;
 926
 927         do {
 928                 ret = walk_page_range(start, range->end, &mm_walk);
 929                 start = hmm_vma_walk.last;
 930         } while (ret == -EAGAIN);
 931
 932         if (ret) {
 933                 unsigned long i;
 934
 935                 i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
 936                 hmm_pfns_clear(range, &range->pfns[i], hmm_vma_walk.last,
 937                                range->end);
 938                 hmm_vma_range_done(range);
 939         }
 940         return ret;
 941 }
 942 EXPORT_SYMBOL(hmm_vma_fault);
 943 #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
 944
 945
 946 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) ||  IS_ENABLED(CONFIG_DEVICE_PUBLIC)
 947 struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
 948                                        unsigned long addr)
 949 {
 950         struct page *page;
 951
 952         page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
 953         if (!page)
 954                 return NULL;
 955         lock_page(page);
 956         return page;
 957 }
 958 EXPORT_SYMBOL(hmm_vma_alloc_locked_page);
 959
 960
 961 static void hmm_devmem_ref_release(struct percpu_ref *ref)
 962 {
 963         struct hmm_devmem *devmem;
 964
 965         devmem = container_of(ref, struct hmm_devmem, ref);
 966         complete(&devmem->completion);
 967 }
 968
 969 static void hmm_devmem_ref_exit(void *data)
 970 {
 971         struct percpu_ref *ref = data;
 972         struct hmm_devmem *devmem;
 973
 974         devmem = container_of(ref, struct hmm_devmem, ref);
 975         percpu_ref_exit(ref);
 976         devm_remove_action(devmem->device, &hmm_devmem_ref_exit, data);
 977 }
 978
 979 static void hmm_devmem_ref_kill(void *data)
 980 {
 981         struct percpu_ref *ref = data;
 982         struct hmm_devmem *devmem;
 983
 984         devmem = container_of(ref, struct hmm_devmem, ref);
 985         percpu_ref_kill(ref);
 986         wait_for_completion(&devmem->completion);
 987         devm_remove_action(devmem->device, &hmm_devmem_ref_kill, data);
 988 }
 989
 990 static int hmm_devmem_fault(struct vm_area_struct *vma,
 991                             unsigned long addr,
 992                             const struct page *page,
 993                             unsigned int flags,
 994                             pmd_t *pmdp)
 995 {
 996         struct hmm_devmem *devmem = page->pgmap->data;
 997
 998         return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp);
 999 }
1000
1001 static void hmm_devmem_free(struct page *page, void *data)
1002 {
1003         struct hmm_devmem *devmem = data;
1004
1005         page->mapping = NULL;
1006
1007         devmem->ops->free(devmem, page);
1008 }
1009
1010 static DEFINE_MUTEX(hmm_devmem_lock);
1011 static RADIX_TREE(hmm_devmem_radix, GFP_KERNEL);
1012
1013 static void hmm_devmem_radix_release(struct resource *resource)
1014 {
1015         resource_size_t key;
1016
1017         mutex_lock(&hmm_devmem_lock);
1018         for (key = resource->start;
1019              key <= resource->end;
1020              key += PA_SECTION_SIZE)
1021                 radix_tree_delete(&hmm_devmem_radix, key >> PA_SECTION_SHIFT);
1022         mutex_unlock(&hmm_devmem_lock);
1023 }
1024
1025 static void hmm_devmem_release(struct device *dev, void *data)
1026 {
1027         struct hmm_devmem *devmem = data;
1028         struct resource *resource = devmem->resource;
1029         unsigned long start_pfn, npages;
1030         struct zone *zone;
1031         struct page *page;
1032
1033         if (percpu_ref_tryget_live(&devmem->ref)) {
1034                 dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
1035                 percpu_ref_put(&devmem->ref);
1036         }
1037
1038         /* pages are dead and unused, undo the arch mapping */
1039         start_pfn = (resource->start & ~(PA_SECTION_SIZE - 1)) >> PAGE_SHIFT;
1040         npages = ALIGN(resource_size(resource), PA_SECTION_SIZE) >> PAGE_SHIFT;
1041
1042         page = pfn_to_page(start_pfn);
1043         zone = page_zone(page);
1044
1045         mem_hotplug_begin();
1046         if (resource->desc == IORES_DESC_DEVICE_PRIVATE_MEMORY)
1047                 __remove_pages(zone, start_pfn, npages, NULL);
1048         else
1049                 arch_remove_memory(start_pfn << PAGE_SHIFT,
1050                                    npages << PAGE_SHIFT, NULL);
1051         mem_hotplug_done();
1052
1053         hmm_devmem_radix_release(resource);
1054 }
1055
1056 static int hmm_devmem_pages_create(struct hmm_devmem *devmem)
1057 {
1058         resource_size_t key, align_start, align_size, align_end;
1059         struct device *device = devmem->device;
1060         int ret, nid, is_ram;
1061
1062         align_start = devmem->resource->start & ~(PA_SECTION_SIZE - 1);
1063         align_size = ALIGN(devmem->resource->start +
1064                            resource_size(devmem->resource),
1065                            PA_SECTION_SIZE) - align_start;
1066
1067         is_ram = region_intersects(align_start, align_size,
1068                                    IORESOURCE_SYSTEM_RAM,
1069                                    IORES_DESC_NONE);
1070         if (is_ram == REGION_MIXED) {
1071                 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
1072                                 __func__, devmem->resource);
1073                 return -ENXIO;
1074         }
1075         if (is_ram == REGION_INTERSECTS)
1076                 return -ENXIO;
1077
1078         if (devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY)
1079                 devmem->pagemap.type = MEMORY_DEVICE_PUBLIC;
1080         else
1081                 devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
1082
1083         devmem->pagemap.res = *devmem->resource;
1084         devmem->pagemap.page_fault = hmm_devmem_fault;
1085         devmem->pagemap.page_free = hmm_devmem_free;
1086         devmem->pagemap.dev = devmem->device;
1087         devmem->pagemap.ref = &devmem->ref;
1088         devmem->pagemap.data = devmem;
1089
1090         mutex_lock(&hmm_devmem_lock);
1091         align_end = align_start + align_size - 1;
1092         for (key = align_start; key <= align_end; key += PA_SECTION_SIZE) {
1093                 struct hmm_devmem *dup;
1094
1095                 dup = radix_tree_lookup(&hmm_devmem_radix,
1096                                         key >> PA_SECTION_SHIFT);
1097                 if (dup) {
1098                         dev_err(device, "%s: collides with mapping for %s\n",
1099                                 __func__, dev_name(dup->device));
1100                         mutex_unlock(&hmm_devmem_lock);
1101                         ret = -EBUSY;
1102                         goto error;
1103                 }
1104                 ret = radix_tree_insert(&hmm_devmem_radix,
1105                                         key >> PA_SECTION_SHIFT,
1106                                         devmem);
1107                 if (ret) {
1108                         dev_err(device, "%s: failed: %d\n", __func__, ret);
1109                         mutex_unlock(&hmm_devmem_lock);
1110                         goto error_radix;
1111                 }
1112         }
1113         mutex_unlock(&hmm_devmem_lock);
1114
1115         nid = dev_to_node(device);
1116         if (nid < 0)
1117                 nid = numa_mem_id();
1118
1119         mem_hotplug_begin();
1120         /*
1121          * For device private memory we call add_pages() as we only need to
1122          * allocate and initialize struct page for the device memory. More-
1123          * over the device memory is un-accessible thus we do not want to
1124          * create a linear mapping for the memory like arch_add_memory()
1125          * would do.
1126          *
1127          * For device public memory, which is accesible by the CPU, we do
1128          * want the linear mapping and thus use arch_add_memory().
1129          */
1130         if (devmem->pagemap.type == MEMORY_DEVICE_PUBLIC)
1131                 ret = arch_add_memory(nid, align_start, align_size, NULL,
1132                                 false);
1133         else
1134                 ret = add_pages(nid, align_start >> PAGE_SHIFT,
1135                                 align_size >> PAGE_SHIFT, NULL, false);
1136         if (ret) {
1137                 mem_hotplug_done();
1138                 goto error_add_memory;
1139         }
1140         move_pfn_range_to_zone(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
1141                                 align_start >> PAGE_SHIFT,
1142                                 align_size >> PAGE_SHIFT, NULL);
1143         mem_hotplug_done();
1144
1145         /*
1146          * Initialization of the pages has been deferred until now in order
1147          * to allow us to do the work while not holding the hotplug lock.
1148          */
1149         memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
1150                                 align_start >> PAGE_SHIFT,
1151                                 align_size >> PAGE_SHIFT, &devmem->pagemap);
1152
1153         return 0;
1154
1155 error_add_memory:
1156         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
1157 error_radix:
1158         hmm_devmem_radix_release(devmem->resource);
1159 error:
1160         return ret;
1161 }
1162
1163 static int hmm_devmem_match(struct device *dev, void *data, void *match_data)
1164 {
1165         struct hmm_devmem *devmem = data;
1166
1167         return devmem->resource == match_data;
1168 }
1169
1170 static void hmm_devmem_pages_remove(struct hmm_devmem *devmem)
1171 {
1172         devres_release(devmem->device, &hmm_devmem_release,
1173                        &hmm_devmem_match, devmem->resource);
1174 }
1175
1176 /*
1177  * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory
1178  *
1179  * @ops: memory event device driver callback (see struct hmm_devmem_ops)
1180  * @device: device struct to bind the resource too
1181  * @size: size in bytes of the device memory to add
1182  * Returns: pointer to new hmm_devmem struct ERR_PTR otherwise
1183  *
1184  * This function first finds an empty range of physical address big enough to
1185  * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which
1186  * in turn allocates struct pages. It does not do anything beyond that; all
1187  * events affecting the memory will go through the various callbacks provided
1188  * by hmm_devmem_ops struct.
1189  *
1190  * Device driver should call this function during device initialization and
1191  * is then responsible of memory management. HMM only provides helpers.
1192  */
1193 struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
1194                                   struct device *device,
1195                                   unsigned long size)
1196 {
1197         struct hmm_devmem *devmem;
1198         resource_size_t addr;
1199         int ret;
1200
1201         dev_pagemap_get_ops();
1202
1203         devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem),
1204                                    GFP_KERNEL, dev_to_node(device));
1205         if (!devmem)
1206                 return ERR_PTR(-ENOMEM);
1207
1208         init_completion(&devmem->completion);
1209         devmem->pfn_first = -1UL;
1210         devmem->pfn_last = -1UL;
1211         devmem->resource = NULL;
1212         devmem->device = device;
1213         devmem->ops = ops;
1214
1215         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1216                               0, GFP_KERNEL);
1217         if (ret)
1218                 goto error_percpu_ref;
1219
1220         ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref);
1221         if (ret)
1222                 goto error_devm_add_action;
1223
1224         size = ALIGN(size, PA_SECTION_SIZE);
1225         addr = min((unsigned long)iomem_resource.end,
1226                    (1UL << MAX_PHYSMEM_BITS) - 1);
1227         addr = addr - size + 1UL;
1228
1229         /*
1230          * FIXME add a new helper to quickly walk resource tree and find free
1231          * range
1232          *
1233          * FIXME what about ioport_resource resource ?
1234          */
1235         for (; addr > size && addr >= iomem_resource.start; addr -= size) {
1236                 ret = region_intersects(addr, size, 0, IORES_DESC_NONE);
1237                 if (ret != REGION_DISJOINT)
1238                         continue;
1239
1240                 devmem->resource = devm_request_mem_region(device, addr, size,
1241                                                            dev_name(device));
1242                 if (!devmem->resource) {
1243                         ret = -ENOMEM;
1244                         goto error_no_resource;
1245                 }
1246                 break;
1247         }
1248         if (!devmem->resource) {
1249                 ret = -ERANGE;
1250                 goto error_no_resource;
1251         }
1252
1253         devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY;
1254         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1255         devmem->pfn_last = devmem->pfn_first +
1256                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1257
1258         ret = hmm_devmem_pages_create(devmem);
1259         if (ret)
1260                 goto error_pages;
1261
1262         devres_add(device, devmem);
1263
1264         ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref);
1265         if (ret) {
1266                 hmm_devmem_remove(devmem);
1267                 return ERR_PTR(ret);
1268         }
1269
1270         return devmem;
1271
1272 error_pages:
1273         devm_release_mem_region(device, devmem->resource->start,
1274                                 resource_size(devmem->resource));
1275 error_no_resource:
1276 error_devm_add_action:
1277         hmm_devmem_ref_kill(&devmem->ref);
1278         hmm_devmem_ref_exit(&devmem->ref);
1279 error_percpu_ref:
1280         devres_free(devmem);
1281         return ERR_PTR(ret);
1282 }
1283 EXPORT_SYMBOL(hmm_devmem_add);
1284
1285 struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops,
1286                                            struct device *device,
1287                                            struct resource *res)
1288 {
1289         struct hmm_devmem *devmem;
1290         int ret;
1291
1292         if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY)
1293                 return ERR_PTR(-EINVAL);
1294
1295         dev_pagemap_get_ops();
1296
1297         devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem),
1298                                    GFP_KERNEL, dev_to_node(device));
1299         if (!devmem)
1300                 return ERR_PTR(-ENOMEM);
1301
1302         init_completion(&devmem->completion);
1303         devmem->pfn_first = -1UL;
1304         devmem->pfn_last = -1UL;
1305         devmem->resource = res;
1306         devmem->device = device;
1307         devmem->ops = ops;
1308
1309         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1310                               0, GFP_KERNEL);
1311         if (ret)
1312                 goto error_percpu_ref;
1313
1314         ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref);
1315         if (ret)
1316                 goto error_devm_add_action;
1317
1318
1319         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1320         devmem->pfn_last = devmem->pfn_first +
1321                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1322
1323         ret = hmm_devmem_pages_create(devmem);
1324         if (ret)
1325                 goto error_devm_add_action;
1326
1327         devres_add(device, devmem);
1328
1329         ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref);
1330         if (ret) {
1331                 hmm_devmem_remove(devmem);
1332                 return ERR_PTR(ret);
1333         }
1334
1335         return devmem;
1336
1337 error_devm_add_action:
1338         hmm_devmem_ref_kill(&devmem->ref);
1339         hmm_devmem_ref_exit(&devmem->ref);
1340 error_percpu_ref:
1341         devres_free(devmem);
1342         return ERR_PTR(ret);
1343 }
1344 EXPORT_SYMBOL(hmm_devmem_add_resource);
1345
1346 /*
1347  * hmm_devmem_remove() - remove device memory (kill and free ZONE_DEVICE)
1348  *
1349  * @devmem: hmm_devmem struct use to track and manage the ZONE_DEVICE memory
1350  *
1351  * This will hot-unplug memory that was hotplugged by hmm_devmem_add on behalf
1352  * of the device driver. It will free struct page and remove the resource that
1353  * reserved the physical address range for this device memory.
1354  */
1355 void hmm_devmem_remove(struct hmm_devmem *devmem)
1356 {
1357         resource_size_t start, size;
1358         struct device *device;
1359         bool cdm = false;
1360
1361         if (!devmem)
1362                 return;
1363
1364         device = devmem->device;
1365         start = devmem->resource->start;
1366         size = resource_size(devmem->resource);
1367
1368         cdm = devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY;
1369         hmm_devmem_ref_kill(&devmem->ref);
1370         hmm_devmem_ref_exit(&devmem->ref);
1371         hmm_devmem_pages_remove(devmem);
1372
1373         if (!cdm)
1374                 devm_release_mem_region(device, start, size);
1375 }
1376 EXPORT_SYMBOL(hmm_devmem_remove);
1377
1378 /*
1379  * A device driver that wants to handle multiple devices memory through a
1380  * single fake device can use hmm_device to do so. This is purely a helper
1381  * and it is not needed to make use of any HMM functionality.
1382  */
1383 #define HMM_DEVICE_MAX 256
1384
1385 static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX);
1386 static DEFINE_SPINLOCK(hmm_device_lock);
1387 static struct class *hmm_device_class;
1388 static dev_t hmm_device_devt;
1389
1390 static void hmm_device_release(struct device *device)
1391 {
1392         struct hmm_device *hmm_device;
1393
1394         hmm_device = container_of(device, struct hmm_device, device);
1395         spin_lock(&hmm_device_lock);
1396         clear_bit(hmm_device->minor, hmm_device_mask);
1397         spin_unlock(&hmm_device_lock);
1398
1399         kfree(hmm_device);
1400 }
1401
1402 struct hmm_device *hmm_device_new(void *drvdata)
1403 {
1404         struct hmm_device *hmm_device;
1405
1406         hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL);
1407         if (!hmm_device)
1408                 return ERR_PTR(-ENOMEM);
1409
1410         spin_lock(&hmm_device_lock);
1411         hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX);
1412         if (hmm_device->minor >= HMM_DEVICE_MAX) {
1413                 spin_unlock(&hmm_device_lock);
1414                 kfree(hmm_device);
1415                 return ERR_PTR(-EBUSY);
1416         }
1417         set_bit(hmm_device->minor, hmm_device_mask);
1418         spin_unlock(&hmm_device_lock);
1419
1420         dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor);
1421         hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt),
1422                                         hmm_device->minor);
1423         hmm_device->device.release = hmm_device_release;
1424         dev_set_drvdata(&hmm_device->device, drvdata);
1425         hmm_device->device.class = hmm_device_class;
1426         device_initialize(&hmm_device->device);
1427
1428         return hmm_device;
1429 }
1430 EXPORT_SYMBOL(hmm_device_new);
1431
1432 void hmm_device_put(struct hmm_device *hmm_device)
1433 {
1434         put_device(&hmm_device->device);
1435 }
1436 EXPORT_SYMBOL(hmm_device_put);
1437
1438 static int __init hmm_init(void)
1439 {
1440         int ret;
1441
1442         ret = alloc_chrdev_region(&hmm_device_devt, 0,
1443                                   HMM_DEVICE_MAX,
1444                                   "hmm_device");
1445         if (ret)
1446                 return ret;
1447
1448         hmm_device_class = class_create(THIS_MODULE, "hmm_device");
1449         if (IS_ERR(hmm_device_class)) {
1450                 unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX);
1451                 return PTR_ERR(hmm_device_class);
1452         }
1453         return 0;
1454 }
1455
1456 device_initcall(hmm_init);
1457 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */