mm/hmm.c

   1 /*
   2  * Copyright 2013 Red Hat Inc.
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * Authors: Jérôme Glisse <jglisse@redhat.com>
  15  */
  16 /*
  17  * Refer to include/linux/hmm.h for information about heterogeneous memory
  18  * management or HMM for short.
  19  */
  20 #include <linux/mm.h>
  21 #include <linux/hmm.h>
  22 #include <linux/init.h>
  23 #include <linux/rmap.h>
  24 #include <linux/swap.h>
  25 #include <linux/slab.h>
  26 #include <linux/sched.h>
  27 #include <linux/mmzone.h>
  28 #include <linux/pagemap.h>
  29 #include <linux/swapops.h>
  30 #include <linux/hugetlb.h>
  31 #include <linux/memremap.h>
  32 #include <linux/jump_label.h>
  33 #include <linux/mmu_notifier.h>
  34 #include <linux/memory_hotplug.h>
  35
  36 #define PA_SECTION_SIZE (1UL << PA_SECTION_SHIFT)
  37
  38 #if IS_ENABLED(CONFIG_HMM_MIRROR)
  39 static const struct mmu_notifier_ops hmm_mmu_notifier_ops;
  40
  41 static inline struct hmm *mm_get_hmm(struct mm_struct *mm)
  42 {
  43         struct hmm *hmm = READ_ONCE(mm->hmm);
  44
  45         if (hmm && kref_get_unless_zero(&hmm->kref))
  46                 return hmm;
  47
  48         return NULL;
  49 }
  50
  51 /**
  52  * hmm_get_or_create - register HMM against an mm (HMM internal)
  53  *
  54  * @mm: mm struct to attach to
  55  * Returns: returns an HMM object, either by referencing the existing
  56  *          (per-process) object, or by creating a new one.
  57  *
  58  * This is not intended to be used directly by device drivers. If mm already
  59  * has an HMM struct then it get a reference on it and returns it. Otherwise
  60  * it allocates an HMM struct, initializes it, associate it with the mm and
  61  * returns it.
  62  */
  63 static struct hmm *hmm_get_or_create(struct mm_struct *mm)
  64 {
  65         struct hmm *hmm = mm_get_hmm(mm);
  66         bool cleanup = false;
  67
  68         if (hmm)
  69                 return hmm;
  70
  71         hmm = kmalloc(sizeof(*hmm), GFP_KERNEL);
  72         if (!hmm)
  73                 return NULL;
  74         init_waitqueue_head(&hmm->wq);
  75         INIT_LIST_HEAD(&hmm->mirrors);
  76         init_rwsem(&hmm->mirrors_sem);
  77         hmm->mmu_notifier.ops = NULL;
  78         INIT_LIST_HEAD(&hmm->ranges);
  79         mutex_init(&hmm->lock);
  80         kref_init(&hmm->kref);
  81         hmm->notifiers = 0;
  82         hmm->dead = false;
  83         hmm->mm = mm;
  84
  85         spin_lock(&mm->page_table_lock);
  86         if (!mm->hmm)
  87                 mm->hmm = hmm;
  88         else
  89                 cleanup = true;
  90         spin_unlock(&mm->page_table_lock);
  91
  92         if (cleanup)
  93                 goto error;
  94
  95         /*
  96          * We should only get here if hold the mmap_sem in write mode ie on
  97          * registration of first mirror through hmm_mirror_register()
  98          */
  99         hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops;
 100         if (__mmu_notifier_register(&hmm->mmu_notifier, mm))
 101                 goto error_mm;
 102
 103         return hmm;
 104
 105 error_mm:
 106         spin_lock(&mm->page_table_lock);
 107         if (mm->hmm == hmm)
 108                 mm->hmm = NULL;
 109         spin_unlock(&mm->page_table_lock);
 110 error:
 111         kfree(hmm);
 112         return NULL;
 113 }
 114
 115 static void hmm_free(struct kref *kref)
 116 {
 117         struct hmm *hmm = container_of(kref, struct hmm, kref);
 118         struct mm_struct *mm = hmm->mm;
 119
 120         mmu_notifier_unregister_no_release(&hmm->mmu_notifier, mm);
 121
 122         spin_lock(&mm->page_table_lock);
 123         if (mm->hmm == hmm)
 124                 mm->hmm = NULL;
 125         spin_unlock(&mm->page_table_lock);
 126
 127         kfree(hmm);
 128 }
 129
 130 static inline void hmm_put(struct hmm *hmm)
 131 {
 132         kref_put(&hmm->kref, hmm_free);
 133 }
 134
 135 void hmm_mm_destroy(struct mm_struct *mm)
 136 {
 137         struct hmm *hmm;
 138
 139         spin_lock(&mm->page_table_lock);
 140         hmm = mm_get_hmm(mm);
 141         mm->hmm = NULL;
 142         if (hmm) {
 143                 hmm->mm = NULL;
 144                 hmm->dead = true;
 145                 spin_unlock(&mm->page_table_lock);
 146                 hmm_put(hmm);
 147                 return;
 148         }
 149
 150         spin_unlock(&mm->page_table_lock);
 151 }
 152
 153 static void hmm_release(struct mmu_notifier *mn, struct mm_struct *mm)
 154 {
 155         struct hmm *hmm = mm_get_hmm(mm);
 156         struct hmm_mirror *mirror;
 157         struct hmm_range *range;
 158
 159         /* Report this HMM as dying. */
 160         hmm->dead = true;
 161
 162         /* Wake-up everyone waiting on any range. */
 163         mutex_lock(&hmm->lock);
 164         list_for_each_entry(range, &hmm->ranges, list) {
 165                 range->valid = false;
 166         }
 167         wake_up_all(&hmm->wq);
 168         mutex_unlock(&hmm->lock);
 169
 170         down_write(&hmm->mirrors_sem);
 171         mirror = list_first_entry_or_null(&hmm->mirrors, struct hmm_mirror,
 172                                           list);
 173         while (mirror) {
 174                 list_del_init(&mirror->list);
 175                 if (mirror->ops->release) {
 176                         /*
 177                          * Drop mirrors_sem so callback can wait on any pending
 178                          * work that might itself trigger mmu_notifier callback
 179                          * and thus would deadlock with us.
 180                          */
 181                         up_write(&hmm->mirrors_sem);
 182                         mirror->ops->release(mirror);
 183                         down_write(&hmm->mirrors_sem);
 184                 }
 185                 mirror = list_first_entry_or_null(&hmm->mirrors,
 186                                                   struct hmm_mirror, list);
 187         }
 188         up_write(&hmm->mirrors_sem);
 189
 190         hmm_put(hmm);
 191 }
 192
 193 static int hmm_invalidate_range_start(struct mmu_notifier *mn,
 194                         const struct mmu_notifier_range *nrange)
 195 {
 196         struct hmm *hmm = mm_get_hmm(nrange->mm);
 197         struct hmm_mirror *mirror;
 198         struct hmm_update update;
 199         struct hmm_range *range;
 200         int ret = 0;
 201
 202         VM_BUG_ON(!hmm);
 203
 204         update.start = nrange->start;
 205         update.end = nrange->end;
 206         update.event = HMM_UPDATE_INVALIDATE;
 207         update.blockable = nrange->blockable;
 208
 209         if (nrange->blockable)
 210                 mutex_lock(&hmm->lock);
 211         else if (!mutex_trylock(&hmm->lock)) {
 212                 ret = -EAGAIN;
 213                 goto out;
 214         }
 215         hmm->notifiers++;
 216         list_for_each_entry(range, &hmm->ranges, list) {
 217                 if (update.end < range->start || update.start >= range->end)
 218                         continue;
 219
 220                 range->valid = false;
 221         }
 222         mutex_unlock(&hmm->lock);
 223
 224         if (nrange->blockable)
 225                 down_read(&hmm->mirrors_sem);
 226         else if (!down_read_trylock(&hmm->mirrors_sem)) {
 227                 ret = -EAGAIN;
 228                 goto out;
 229         }
 230         list_for_each_entry(mirror, &hmm->mirrors, list) {
 231                 int ret;
 232
 233                 ret = mirror->ops->sync_cpu_device_pagetables(mirror, &update);
 234                 if (!update.blockable && ret == -EAGAIN) {
 235                         up_read(&hmm->mirrors_sem);
 236                         ret = -EAGAIN;
 237                         goto out;
 238                 }
 239         }
 240         up_read(&hmm->mirrors_sem);
 241
 242 out:
 243         hmm_put(hmm);
 244         return ret;
 245 }
 246
 247 static void hmm_invalidate_range_end(struct mmu_notifier *mn,
 248                         const struct mmu_notifier_range *nrange)
 249 {
 250         struct hmm *hmm = mm_get_hmm(nrange->mm);
 251
 252         VM_BUG_ON(!hmm);
 253
 254         mutex_lock(&hmm->lock);
 255         hmm->notifiers--;
 256         if (!hmm->notifiers) {
 257                 struct hmm_range *range;
 258
 259                 list_for_each_entry(range, &hmm->ranges, list) {
 260                         if (range->valid)
 261                                 continue;
 262                         range->valid = true;
 263                 }
 264                 wake_up_all(&hmm->wq);
 265         }
 266         mutex_unlock(&hmm->lock);
 267
 268         hmm_put(hmm);
 269 }
 270
 271 static const struct mmu_notifier_ops hmm_mmu_notifier_ops = {
 272         .release                = hmm_release,
 273         .invalidate_range_start = hmm_invalidate_range_start,
 274         .invalidate_range_end   = hmm_invalidate_range_end,
 275 };
 276
 277 /*
 278  * hmm_mirror_register() - register a mirror against an mm
 279  *
 280  * @mirror: new mirror struct to register
 281  * @mm: mm to register against
 282  *
 283  * To start mirroring a process address space, the device driver must register
 284  * an HMM mirror struct.
 285  *
 286  * THE mm->mmap_sem MUST BE HELD IN WRITE MODE !
 287  */
 288 int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm)
 289 {
 290         /* Sanity check */
 291         if (!mm || !mirror || !mirror->ops)
 292                 return -EINVAL;
 293
 294         mirror->hmm = hmm_get_or_create(mm);
 295         if (!mirror->hmm)
 296                 return -ENOMEM;
 297
 298         down_write(&mirror->hmm->mirrors_sem);
 299         list_add(&mirror->list, &mirror->hmm->mirrors);
 300         up_write(&mirror->hmm->mirrors_sem);
 301
 302         return 0;
 303 }
 304 EXPORT_SYMBOL(hmm_mirror_register);
 305
 306 /*
 307  * hmm_mirror_unregister() - unregister a mirror
 308  *
 309  * @mirror: new mirror struct to register
 310  *
 311  * Stop mirroring a process address space, and cleanup.
 312  */
 313 void hmm_mirror_unregister(struct hmm_mirror *mirror)
 314 {
 315         struct hmm *hmm = READ_ONCE(mirror->hmm);
 316
 317         if (hmm == NULL)
 318                 return;
 319
 320         down_write(&hmm->mirrors_sem);
 321         list_del_init(&mirror->list);
 322         /* To protect us against double unregister ... */
 323         mirror->hmm = NULL;
 324         up_write(&hmm->mirrors_sem);
 325
 326         hmm_put(hmm);
 327 }
 328 EXPORT_SYMBOL(hmm_mirror_unregister);
 329
 330 struct hmm_vma_walk {
 331         struct hmm_range        *range;
 332         struct dev_pagemap      *pgmap;
 333         unsigned long           last;
 334         bool                    fault;
 335         bool                    block;
 336 };
 337
 338 static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr,
 339                             bool write_fault, uint64_t *pfn)
 340 {
 341         unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE;
 342         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 343         struct hmm_range *range = hmm_vma_walk->range;
 344         struct vm_area_struct *vma = walk->vma;
 345         vm_fault_t ret;
 346
 347         flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY;
 348         flags |= write_fault ? FAULT_FLAG_WRITE : 0;
 349         ret = handle_mm_fault(vma, addr, flags);
 350         if (ret & VM_FAULT_RETRY)
 351                 return -EAGAIN;
 352         if (ret & VM_FAULT_ERROR) {
 353                 *pfn = range->values[HMM_PFN_ERROR];
 354                 return -EFAULT;
 355         }
 356
 357         return -EBUSY;
 358 }
 359
 360 static int hmm_pfns_bad(unsigned long addr,
 361                         unsigned long end,
 362                         struct mm_walk *walk)
 363 {
 364         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 365         struct hmm_range *range = hmm_vma_walk->range;
 366         uint64_t *pfns = range->pfns;
 367         unsigned long i;
 368
 369         i = (addr - range->start) >> PAGE_SHIFT;
 370         for (; addr < end; addr += PAGE_SIZE, i++)
 371                 pfns[i] = range->values[HMM_PFN_ERROR];
 372
 373         return 0;
 374 }
 375
 376 /*
 377  * hmm_vma_walk_hole() - handle a range lacking valid pmd or pte(s)
 378  * @start: range virtual start address (inclusive)
 379  * @end: range virtual end address (exclusive)
 380  * @fault: should we fault or not ?
 381  * @write_fault: write fault ?
 382  * @walk: mm_walk structure
 383  * Returns: 0 on success, -EBUSY after page fault, or page fault error
 384  *
 385  * This function will be called whenever pmd_none() or pte_none() returns true,
 386  * or whenever there is no page directory covering the virtual address range.
 387  */
 388 static int hmm_vma_walk_hole_(unsigned long addr, unsigned long end,
 389                               bool fault, bool write_fault,
 390                               struct mm_walk *walk)
 391 {
 392         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 393         struct hmm_range *range = hmm_vma_walk->range;
 394         uint64_t *pfns = range->pfns;
 395         unsigned long i, page_size;
 396
 397         hmm_vma_walk->last = addr;
 398         page_size = hmm_range_page_size(range);
 399         i = (addr - range->start) >> range->page_shift;
 400
 401         for (; addr < end; addr += page_size, i++) {
 402                 pfns[i] = range->values[HMM_PFN_NONE];
 403                 if (fault || write_fault) {
 404                         int ret;
 405
 406                         ret = hmm_vma_do_fault(walk, addr, write_fault,
 407                                                &pfns[i]);
 408                         if (ret != -EBUSY)
 409                                 return ret;
 410                 }
 411         }
 412
 413         return (fault || write_fault) ? -EBUSY : 0;
 414 }
 415
 416 static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 417                                       uint64_t pfns, uint64_t cpu_flags,
 418                                       bool *fault, bool *write_fault)
 419 {
 420         struct hmm_range *range = hmm_vma_walk->range;
 421
 422         if (!hmm_vma_walk->fault)
 423                 return;
 424
 425         /*
 426          * So we not only consider the individual per page request we also
 427          * consider the default flags requested for the range. The API can
 428          * be use in 2 fashions. The first one where the HMM user coalesce
 429          * multiple page fault into one request and set flags per pfns for
 430          * of those faults. The second one where the HMM user want to pre-
 431          * fault a range with specific flags. For the latter one it is a
 432          * waste to have the user pre-fill the pfn arrays with a default
 433          * flags value.
 434          */
 435         pfns = (pfns & range->pfn_flags_mask) | range->default_flags;
 436
 437         /* We aren't ask to do anything ... */
 438         if (!(pfns & range->flags[HMM_PFN_VALID]))
 439                 return;
 440         /* If this is device memory than only fault if explicitly requested */
 441         if ((cpu_flags & range->flags[HMM_PFN_DEVICE_PRIVATE])) {
 442                 /* Do we fault on device memory ? */
 443                 if (pfns & range->flags[HMM_PFN_DEVICE_PRIVATE]) {
 444                         *write_fault = pfns & range->flags[HMM_PFN_WRITE];
 445                         *fault = true;
 446                 }
 447                 return;
 448         }
 449
 450         /* If CPU page table is not valid then we need to fault */
 451         *fault = !(cpu_flags & range->flags[HMM_PFN_VALID]);
 452         /* Need to write fault ? */
 453         if ((pfns & range->flags[HMM_PFN_WRITE]) &&
 454             !(cpu_flags & range->flags[HMM_PFN_WRITE])) {
 455                 *write_fault = true;
 456                 *fault = true;
 457         }
 458 }
 459
 460 static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 461                                  const uint64_t *pfns, unsigned long npages,
 462                                  uint64_t cpu_flags, bool *fault,
 463                                  bool *write_fault)
 464 {
 465         unsigned long i;
 466
 467         if (!hmm_vma_walk->fault) {
 468                 *fault = *write_fault = false;
 469                 return;
 470         }
 471
 472         *fault = *write_fault = false;
 473         for (i = 0; i < npages; ++i) {
 474                 hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags,
 475                                    fault, write_fault);
 476                 if ((*write_fault))
 477                         return;
 478         }
 479 }
 480
 481 static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
 482                              struct mm_walk *walk)
 483 {
 484         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 485         struct hmm_range *range = hmm_vma_walk->range;
 486         bool fault, write_fault;
 487         unsigned long i, npages;
 488         uint64_t *pfns;
 489
 490         i = (addr - range->start) >> PAGE_SHIFT;
 491         npages = (end - addr) >> PAGE_SHIFT;
 492         pfns = &range->pfns[i];
 493         hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 494                              0, &fault, &write_fault);
 495         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 496 }
 497
 498 static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd)
 499 {
 500         if (pmd_protnone(pmd))
 501                 return 0;
 502         return pmd_write(pmd) ? range->flags[HMM_PFN_VALID] |
 503                                 range->flags[HMM_PFN_WRITE] :
 504                                 range->flags[HMM_PFN_VALID];
 505 }
 506
 507 static inline uint64_t pud_to_hmm_pfn_flags(struct hmm_range *range, pud_t pud)
 508 {
 509         if (!pud_present(pud))
 510                 return 0;
 511         return pud_write(pud) ? range->flags[HMM_PFN_VALID] |
 512                                 range->flags[HMM_PFN_WRITE] :
 513                                 range->flags[HMM_PFN_VALID];
 514 }
 515
 516 static int hmm_vma_handle_pmd(struct mm_walk *walk,
 517                               unsigned long addr,
 518                               unsigned long end,
 519                               uint64_t *pfns,
 520                               pmd_t pmd)
 521 {
 522 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 523         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 524         struct hmm_range *range = hmm_vma_walk->range;
 525         unsigned long pfn, npages, i;
 526         bool fault, write_fault;
 527         uint64_t cpu_flags;
 528
 529         npages = (end - addr) >> PAGE_SHIFT;
 530         cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
 531         hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags,
 532                              &fault, &write_fault);
 533
 534         if (pmd_protnone(pmd) || fault || write_fault)
 535                 return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 536
 537         pfn = pmd_pfn(pmd) + pte_index(addr);
 538         for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++) {
 539                 if (pmd_devmap(pmd)) {
 540                         hmm_vma_walk->pgmap = get_dev_pagemap(pfn,
 541                                               hmm_vma_walk->pgmap);
 542                         if (unlikely(!hmm_vma_walk->pgmap))
 543                                 return -EBUSY;
 544                 }
 545                 pfns[i] = hmm_pfn_from_pfn(range, pfn) | cpu_flags;
 546         }
 547         if (hmm_vma_walk->pgmap) {
 548                 put_dev_pagemap(hmm_vma_walk->pgmap);
 549                 hmm_vma_walk->pgmap = NULL;
 550         }
 551         hmm_vma_walk->last = end;
 552         return 0;
 553 #else
 554         /* If THP is not enabled then we should never reach that code ! */
 555         return -EINVAL;
 556 #endif
 557 }
 558
 559 static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte)
 560 {
 561         if (pte_none(pte) || !pte_present(pte))
 562                 return 0;
 563         return pte_write(pte) ? range->flags[HMM_PFN_VALID] |
 564                                 range->flags[HMM_PFN_WRITE] :
 565                                 range->flags[HMM_PFN_VALID];
 566 }
 567
 568 static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
 569                               unsigned long end, pmd_t *pmdp, pte_t *ptep,
 570                               uint64_t *pfn)
 571 {
 572         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 573         struct hmm_range *range = hmm_vma_walk->range;
 574         struct vm_area_struct *vma = walk->vma;
 575         bool fault, write_fault;
 576         uint64_t cpu_flags;
 577         pte_t pte = *ptep;
 578         uint64_t orig_pfn = *pfn;
 579
 580         *pfn = range->values[HMM_PFN_NONE];
 581         fault = write_fault = false;
 582
 583         if (pte_none(pte)) {
 584                 hmm_pte_need_fault(hmm_vma_walk, orig_pfn, 0,
 585                                    &fault, &write_fault);
 586                 if (fault || write_fault)
 587                         goto fault;
 588                 return 0;
 589         }
 590
 591         if (!pte_present(pte)) {
 592                 swp_entry_t entry = pte_to_swp_entry(pte);
 593
 594                 if (!non_swap_entry(entry)) {
 595                         if (fault || write_fault)
 596                                 goto fault;
 597                         return 0;
 598                 }
 599
 600                 /*
 601                  * This is a special swap entry, ignore migration, use
 602                  * device and report anything else as error.
 603                  */
 604                 if (is_device_private_entry(entry)) {
 605                         cpu_flags = range->flags[HMM_PFN_VALID] |
 606                                 range->flags[HMM_PFN_DEVICE_PRIVATE];
 607                         cpu_flags |= is_write_device_private_entry(entry) ?
 608                                 range->flags[HMM_PFN_WRITE] : 0;
 609                         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 610                                            &fault, &write_fault);
 611                         if (fault || write_fault)
 612                                 goto fault;
 613                         *pfn = hmm_pfn_from_pfn(range, swp_offset(entry));
 614                         *pfn |= cpu_flags;
 615                         return 0;
 616                 }
 617
 618                 if (is_migration_entry(entry)) {
 619                         if (fault || write_fault) {
 620                                 pte_unmap(ptep);
 621                                 hmm_vma_walk->last = addr;
 622                                 migration_entry_wait(vma->vm_mm,
 623                                                      pmdp, addr);
 624                                 return -EBUSY;
 625                         }
 626                         return 0;
 627                 }
 628
 629                 /* Report error for everything else */
 630                 *pfn = range->values[HMM_PFN_ERROR];
 631                 return -EFAULT;
 632         } else {
 633                 cpu_flags = pte_to_hmm_pfn_flags(range, pte);
 634                 hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 635                                    &fault, &write_fault);
 636         }
 637
 638         if (fault || write_fault)
 639                 goto fault;
 640
 641         if (pte_devmap(pte)) {
 642                 hmm_vma_walk->pgmap = get_dev_pagemap(pte_pfn(pte),
 643                                               hmm_vma_walk->pgmap);
 644                 if (unlikely(!hmm_vma_walk->pgmap))
 645                         return -EBUSY;
 646         } else if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && pte_special(pte)) {
 647                 *pfn = range->values[HMM_PFN_SPECIAL];
 648                 return -EFAULT;
 649         }
 650
 651         *pfn = hmm_pfn_from_pfn(range, pte_pfn(pte)) | cpu_flags;
 652         return 0;
 653
 654 fault:
 655         if (hmm_vma_walk->pgmap) {
 656                 put_dev_pagemap(hmm_vma_walk->pgmap);
 657                 hmm_vma_walk->pgmap = NULL;
 658         }
 659         pte_unmap(ptep);
 660         /* Fault any virtual address we were asked to fault */
 661         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 662 }
 663
 664 static int hmm_vma_walk_pmd(pmd_t *pmdp,
 665                             unsigned long start,
 666                             unsigned long end,
 667                             struct mm_walk *walk)
 668 {
 669         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 670         struct hmm_range *range = hmm_vma_walk->range;
 671         struct vm_area_struct *vma = walk->vma;
 672         uint64_t *pfns = range->pfns;
 673         unsigned long addr = start, i;
 674         pte_t *ptep;
 675         pmd_t pmd;
 676
 677
 678 again:
 679         pmd = READ_ONCE(*pmdp);
 680         if (pmd_none(pmd))
 681                 return hmm_vma_walk_hole(start, end, walk);
 682
 683         if (pmd_huge(pmd) && (range->vma->vm_flags & VM_HUGETLB))
 684                 return hmm_pfns_bad(start, end, walk);
 685
 686         if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
 687                 bool fault, write_fault;
 688                 unsigned long npages;
 689                 uint64_t *pfns;
 690
 691                 i = (addr - range->start) >> PAGE_SHIFT;
 692                 npages = (end - addr) >> PAGE_SHIFT;
 693                 pfns = &range->pfns[i];
 694
 695                 hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 696                                      0, &fault, &write_fault);
 697                 if (fault || write_fault) {
 698                         hmm_vma_walk->last = addr;
 699                         pmd_migration_entry_wait(vma->vm_mm, pmdp);
 700                         return -EBUSY;
 701                 }
 702                 return 0;
 703         } else if (!pmd_present(pmd))
 704                 return hmm_pfns_bad(start, end, walk);
 705
 706         if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
 707                 /*
 708                  * No need to take pmd_lock here, even if some other threads
 709                  * is splitting the huge pmd we will get that event through
 710                  * mmu_notifier callback.
 711                  *
 712                  * So just read pmd value and check again its a transparent
 713                  * huge or device mapping one and compute corresponding pfn
 714                  * values.
 715                  */
 716                 pmd = pmd_read_atomic(pmdp);
 717                 barrier();
 718                 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
 719                         goto again;
 720
 721                 i = (addr - range->start) >> PAGE_SHIFT;
 722                 return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd);
 723         }
 724
 725         /*
 726          * We have handled all the valid case above ie either none, migration,
 727          * huge or transparent huge. At this point either it is a valid pmd
 728          * entry pointing to pte directory or it is a bad pmd that will not
 729          * recover.
 730          */
 731         if (pmd_bad(pmd))
 732                 return hmm_pfns_bad(start, end, walk);
 733
 734         ptep = pte_offset_map(pmdp, addr);
 735         i = (addr - range->start) >> PAGE_SHIFT;
 736         for (; addr < end; addr += PAGE_SIZE, ptep++, i++) {
 737                 int r;
 738
 739                 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]);
 740                 if (r) {
 741                         /* hmm_vma_handle_pte() did unmap pte directory */
 742                         hmm_vma_walk->last = addr;
 743                         return r;
 744                 }
 745         }
 746         if (hmm_vma_walk->pgmap) {
 747                 /*
 748                  * We do put_dev_pagemap() here and not in hmm_vma_handle_pte()
 749                  * so that we can leverage get_dev_pagemap() optimization which
 750                  * will not re-take a reference on a pgmap if we already have
 751                  * one.
 752                  */
 753                 put_dev_pagemap(hmm_vma_walk->pgmap);
 754                 hmm_vma_walk->pgmap = NULL;
 755         }
 756         pte_unmap(ptep - 1);
 757
 758         hmm_vma_walk->last = addr;
 759         return 0;
 760 }
 761
 762 static int hmm_vma_walk_pud(pud_t *pudp,
 763                             unsigned long start,
 764                             unsigned long end,
 765                             struct mm_walk *walk)
 766 {
 767         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 768         struct hmm_range *range = hmm_vma_walk->range;
 769         unsigned long addr = start, next;
 770         pmd_t *pmdp;
 771         pud_t pud;
 772         int ret;
 773
 774 again:
 775         pud = READ_ONCE(*pudp);
 776         if (pud_none(pud))
 777                 return hmm_vma_walk_hole(start, end, walk);
 778
 779         if (pud_huge(pud) && pud_devmap(pud)) {
 780                 unsigned long i, npages, pfn;
 781                 uint64_t *pfns, cpu_flags;
 782                 bool fault, write_fault;
 783
 784                 if (!pud_present(pud))
 785                         return hmm_vma_walk_hole(start, end, walk);
 786
 787                 i = (addr - range->start) >> PAGE_SHIFT;
 788                 npages = (end - addr) >> PAGE_SHIFT;
 789                 pfns = &range->pfns[i];
 790
 791                 cpu_flags = pud_to_hmm_pfn_flags(range, pud);
 792                 hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 793                                      cpu_flags, &fault, &write_fault);
 794                 if (fault || write_fault)
 795                         return hmm_vma_walk_hole_(addr, end, fault,
 796                                                 write_fault, walk);
 797
 798 #ifdef CONFIG_HUGETLB_PAGE
 799                 pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 800                 for (i = 0; i < npages; ++i, ++pfn) {
 801                         hmm_vma_walk->pgmap = get_dev_pagemap(pfn,
 802                                               hmm_vma_walk->pgmap);
 803                         if (unlikely(!hmm_vma_walk->pgmap))
 804                                 return -EBUSY;
 805                         pfns[i] = hmm_pfn_from_pfn(range, pfn) | cpu_flags;
 806                 }
 807                 if (hmm_vma_walk->pgmap) {
 808                         put_dev_pagemap(hmm_vma_walk->pgmap);
 809                         hmm_vma_walk->pgmap = NULL;
 810                 }
 811                 hmm_vma_walk->last = end;
 812                 return 0;
 813 #else
 814                 return -EINVAL;
 815 #endif
 816         }
 817
 818         split_huge_pud(walk->vma, pudp, addr);
 819         if (pud_none(*pudp))
 820                 goto again;
 821
 822         pmdp = pmd_offset(pudp, addr);
 823         do {
 824                 next = pmd_addr_end(addr, end);
 825                 ret = hmm_vma_walk_pmd(pmdp, addr, next, walk);
 826                 if (ret)
 827                         return ret;
 828         } while (pmdp++, addr = next, addr != end);
 829
 830         return 0;
 831 }
 832
 833 static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask,
 834                                       unsigned long start, unsigned long end,
 835                                       struct mm_walk *walk)
 836 {
 837 #ifdef CONFIG_HUGETLB_PAGE
 838         unsigned long addr = start, i, pfn, mask, size, pfn_inc;
 839         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 840         struct hmm_range *range = hmm_vma_walk->range;
 841         struct vm_area_struct *vma = walk->vma;
 842         struct hstate *h = hstate_vma(vma);
 843         uint64_t orig_pfn, cpu_flags;
 844         bool fault, write_fault;
 845         spinlock_t *ptl;
 846         pte_t entry;
 847         int ret = 0;
 848
 849         size = 1UL << huge_page_shift(h);
 850         mask = size - 1;
 851         if (range->page_shift != PAGE_SHIFT) {
 852                 /* Make sure we are looking at full page. */
 853                 if (start & mask)
 854                         return -EINVAL;
 855                 if (end < (start + size))
 856                         return -EINVAL;
 857                 pfn_inc = size >> PAGE_SHIFT;
 858         } else {
 859                 pfn_inc = 1;
 860                 size = PAGE_SIZE;
 861         }
 862
 863
 864         ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
 865         entry = huge_ptep_get(pte);
 866
 867         i = (start - range->start) >> range->page_shift;
 868         orig_pfn = range->pfns[i];
 869         range->pfns[i] = range->values[HMM_PFN_NONE];
 870         cpu_flags = pte_to_hmm_pfn_flags(range, entry);
 871         fault = write_fault = false;
 872         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 873                            &fault, &write_fault);
 874         if (fault || write_fault) {
 875                 ret = -ENOENT;
 876                 goto unlock;
 877         }
 878
 879         pfn = pte_pfn(entry) + ((start & mask) >> range->page_shift);
 880         for (; addr < end; addr += size, i++, pfn += pfn_inc)
 881                 range->pfns[i] = hmm_pfn_from_pfn(range, pfn) | cpu_flags;
 882         hmm_vma_walk->last = end;
 883
 884 unlock:
 885         spin_unlock(ptl);
 886
 887         if (ret == -ENOENT)
 888                 return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 889
 890         return ret;
 891 #else /* CONFIG_HUGETLB_PAGE */
 892         return -EINVAL;
 893 #endif
 894 }
 895
 896 static void hmm_pfns_clear(struct hmm_range *range,
 897                            uint64_t *pfns,
 898                            unsigned long addr,
 899                            unsigned long end)
 900 {
 901         for (; addr < end; addr += PAGE_SIZE, pfns++)
 902                 *pfns = range->values[HMM_PFN_NONE];
 903 }
 904
 905 /*
 906  * hmm_range_register() - start tracking change to CPU page table over a range
 907  * @range: range
 908  * @mm: the mm struct for the range of virtual address
 909  * @start: start virtual address (inclusive)
 910  * @end: end virtual address (exclusive)
 911  * @page_shift: expect page shift for the range
 912  * Returns 0 on success, -EFAULT if the address space is no longer valid
 913  *
 914  * Track updates to the CPU page table see include/linux/hmm.h
 915  */
 916 int hmm_range_register(struct hmm_range *range,
 917                        struct mm_struct *mm,
 918                        unsigned long start,
 919                        unsigned long end,
 920                        unsigned page_shift)
 921 {
 922         unsigned long mask = ((1UL << page_shift) - 1UL);
 923
 924         range->valid = false;
 925         range->hmm = NULL;
 926
 927         if ((start & mask) || (end & mask))
 928                 return -EINVAL;
 929         if (start >= end)
 930                 return -EINVAL;
 931
 932         range->page_shift = page_shift;
 933         range->start = start;
 934         range->end = end;
 935
 936         range->hmm = hmm_get_or_create(mm);
 937         if (!range->hmm)
 938                 return -EFAULT;
 939
 940         /* Check if hmm_mm_destroy() was call. */
 941         if (range->hmm->mm == NULL || range->hmm->dead) {
 942                 hmm_put(range->hmm);
 943                 return -EFAULT;
 944         }
 945
 946         /* Initialize range to track CPU page table update */
 947         mutex_lock(&range->hmm->lock);
 948
 949         list_add_rcu(&range->list, &range->hmm->ranges);
 950
 951         /*
 952          * If there are any concurrent notifiers we have to wait for them for
 953          * the range to be valid (see hmm_range_wait_until_valid()).
 954          */
 955         if (!range->hmm->notifiers)
 956                 range->valid = true;
 957         mutex_unlock(&range->hmm->lock);
 958
 959         return 0;
 960 }
 961 EXPORT_SYMBOL(hmm_range_register);
 962
 963 /*
 964  * hmm_range_unregister() - stop tracking change to CPU page table over a range
 965  * @range: range
 966  *
 967  * Range struct is used to track updates to the CPU page table after a call to
 968  * hmm_range_register(). See include/linux/hmm.h for how to use it.
 969  */
 970 void hmm_range_unregister(struct hmm_range *range)
 971 {
 972         /* Sanity check this really should not happen. */
 973         if (range->hmm == NULL || range->end <= range->start)
 974                 return;
 975
 976         mutex_lock(&range->hmm->lock);
 977         list_del_rcu(&range->list);
 978         mutex_unlock(&range->hmm->lock);
 979
 980         /* Drop reference taken by hmm_range_register() */
 981         range->valid = false;
 982         hmm_put(range->hmm);
 983         range->hmm = NULL;
 984 }
 985 EXPORT_SYMBOL(hmm_range_unregister);
 986
 987 /*
 988  * hmm_range_snapshot() - snapshot CPU page table for a range
 989  * @range: range
 990  * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid
 991  *          permission (for instance asking for write and range is read only),
 992  *          -EAGAIN if you need to retry, -EFAULT invalid (ie either no valid
 993  *          vma or it is illegal to access that range), number of valid pages
 994  *          in range->pfns[] (from range start address).
 995  *
 996  * This snapshots the CPU page table for a range of virtual addresses. Snapshot
 997  * validity is tracked by range struct. See in include/linux/hmm.h for example
 998  * on how to use.
 999  */
1000 long hmm_range_snapshot(struct hmm_range *range)
1001 {
1002         const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
1003         unsigned long start = range->start, end;
1004         struct hmm_vma_walk hmm_vma_walk;
1005         struct hmm *hmm = range->hmm;
1006         struct vm_area_struct *vma;
1007         struct mm_walk mm_walk;
1008
1009         /* Check if hmm_mm_destroy() was call. */
1010         if (hmm->mm == NULL || hmm->dead)
1011                 return -EFAULT;
1012
1013         do {
1014                 /* If range is no longer valid force retry. */
1015                 if (!range->valid)
1016                         return -EAGAIN;
1017
1018                 vma = find_vma(hmm->mm, start);
1019                 if (vma == NULL || (vma->vm_flags & device_vma))
1020                         return -EFAULT;
1021
1022                 if (is_vm_hugetlb_page(vma)) {
1023                         struct hstate *h = hstate_vma(vma);
1024
1025                         if (huge_page_shift(h) != range->page_shift &&
1026                             range->page_shift != PAGE_SHIFT)
1027                                 return -EINVAL;
1028                 } else {
1029                         if (range->page_shift != PAGE_SHIFT)
1030                                 return -EINVAL;
1031                 }
1032
1033                 if (!(vma->vm_flags & VM_READ)) {
1034                         /*
1035                          * If vma do not allow read access, then assume that it
1036                          * does not allow write access, either. HMM does not
1037                          * support architecture that allow write without read.
1038                          */
1039                         hmm_pfns_clear(range, range->pfns,
1040                                 range->start, range->end);
1041                         return -EPERM;
1042                 }
1043
1044                 range->vma = vma;
1045                 hmm_vma_walk.pgmap = NULL;
1046                 hmm_vma_walk.last = start;
1047                 hmm_vma_walk.fault = false;
1048                 hmm_vma_walk.range = range;
1049                 mm_walk.private = &hmm_vma_walk;
1050                 end = min(range->end, vma->vm_end);
1051
1052                 mm_walk.vma = vma;
1053                 mm_walk.mm = vma->vm_mm;
1054                 mm_walk.pte_entry = NULL;
1055                 mm_walk.test_walk = NULL;
1056                 mm_walk.hugetlb_entry = NULL;
1057                 mm_walk.pud_entry = hmm_vma_walk_pud;
1058                 mm_walk.pmd_entry = hmm_vma_walk_pmd;
1059                 mm_walk.pte_hole = hmm_vma_walk_hole;
1060                 mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry;
1061
1062                 walk_page_range(start, end, &mm_walk);
1063                 start = end;
1064         } while (start < range->end);
1065
1066         return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
1067 }
1068 EXPORT_SYMBOL(hmm_range_snapshot);
1069
1070 /*
1071  * hmm_range_fault() - try to fault some address in a virtual address range
1072  * @range: range being faulted
1073  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
1074  * Returns: number of valid pages in range->pfns[] (from range start
1075  *          address). This may be zero. If the return value is negative,
1076  *          then one of the following values may be returned:
1077  *
1078  *           -EINVAL  invalid arguments or mm or virtual address are in an
1079  *                    invalid vma (for instance device file vma).
1080  *           -ENOMEM: Out of memory.
1081  *           -EPERM:  Invalid permission (for instance asking for write and
1082  *                    range is read only).
1083  *           -EAGAIN: If you need to retry and mmap_sem was drop. This can only
1084  *                    happens if block argument is false.
1085  *           -EBUSY:  If the the range is being invalidated and you should wait
1086  *                    for invalidation to finish.
1087  *           -EFAULT: Invalid (ie either no valid vma or it is illegal to access
1088  *                    that range), number of valid pages in range->pfns[] (from
1089  *                    range start address).
1090  *
1091  * This is similar to a regular CPU page fault except that it will not trigger
1092  * any memory migration if the memory being faulted is not accessible by CPUs
1093  * and caller does not ask for migration.
1094  *
1095  * On error, for one virtual address in the range, the function will mark the
1096  * corresponding HMM pfn entry with an error flag.
1097  */
1098 long hmm_range_fault(struct hmm_range *range, bool block)
1099 {
1100         const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
1101         unsigned long start = range->start, end;
1102         struct hmm_vma_walk hmm_vma_walk;
1103         struct hmm *hmm = range->hmm;
1104         struct vm_area_struct *vma;
1105         struct mm_walk mm_walk;
1106         int ret;
1107
1108         /* Check if hmm_mm_destroy() was call. */
1109         if (hmm->mm == NULL || hmm->dead)
1110                 return -EFAULT;
1111
1112         do {
1113                 /* If range is no longer valid force retry. */
1114                 if (!range->valid) {
1115                         up_read(&hmm->mm->mmap_sem);
1116                         return -EAGAIN;
1117                 }
1118
1119                 vma = find_vma(hmm->mm, start);
1120                 if (vma == NULL || (vma->vm_flags & device_vma))
1121                         return -EFAULT;
1122
1123                 if (is_vm_hugetlb_page(vma)) {
1124                         if (huge_page_shift(hstate_vma(vma)) !=
1125                             range->page_shift &&
1126                             range->page_shift != PAGE_SHIFT)
1127                                 return -EINVAL;
1128                 } else {
1129                         if (range->page_shift != PAGE_SHIFT)
1130                                 return -EINVAL;
1131                 }
1132
1133                 if (!(vma->vm_flags & VM_READ)) {
1134                         /*
1135                          * If vma do not allow read access, then assume that it
1136                          * does not allow write access, either. HMM does not
1137                          * support architecture that allow write without read.
1138                          */
1139                         hmm_pfns_clear(range, range->pfns,
1140                                 range->start, range->end);
1141                         return -EPERM;
1142                 }
1143
1144                 range->vma = vma;
1145                 hmm_vma_walk.pgmap = NULL;
1146                 hmm_vma_walk.last = start;
1147                 hmm_vma_walk.fault = true;
1148                 hmm_vma_walk.block = block;
1149                 hmm_vma_walk.range = range;
1150                 mm_walk.private = &hmm_vma_walk;
1151                 end = min(range->end, vma->vm_end);
1152
1153                 mm_walk.vma = vma;
1154                 mm_walk.mm = vma->vm_mm;
1155                 mm_walk.pte_entry = NULL;
1156                 mm_walk.test_walk = NULL;
1157                 mm_walk.hugetlb_entry = NULL;
1158                 mm_walk.pud_entry = hmm_vma_walk_pud;
1159                 mm_walk.pmd_entry = hmm_vma_walk_pmd;
1160                 mm_walk.pte_hole = hmm_vma_walk_hole;
1161                 mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry;
1162
1163                 do {
1164                         ret = walk_page_range(start, end, &mm_walk);
1165                         start = hmm_vma_walk.last;
1166
1167                         /* Keep trying while the range is valid. */
1168                 } while (ret == -EBUSY && range->valid);
1169
1170                 if (ret) {
1171                         unsigned long i;
1172
1173                         i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
1174                         hmm_pfns_clear(range, &range->pfns[i],
1175                                 hmm_vma_walk.last, range->end);
1176                         return ret;
1177                 }
1178                 start = end;
1179
1180         } while (start < range->end);
1181
1182         return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
1183 }
1184 EXPORT_SYMBOL(hmm_range_fault);
1185 #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
1186
1187
1188 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) ||  IS_ENABLED(CONFIG_DEVICE_PUBLIC)
1189 struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
1190                                        unsigned long addr)
1191 {
1192         struct page *page;
1193
1194         page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
1195         if (!page)
1196                 return NULL;
1197         lock_page(page);
1198         return page;
1199 }
1200 EXPORT_SYMBOL(hmm_vma_alloc_locked_page);
1201
1202
1203 static void hmm_devmem_ref_release(struct percpu_ref *ref)
1204 {
1205         struct hmm_devmem *devmem;
1206
1207         devmem = container_of(ref, struct hmm_devmem, ref);
1208         complete(&devmem->completion);
1209 }
1210
1211 static void hmm_devmem_ref_exit(void *data)
1212 {
1213         struct percpu_ref *ref = data;
1214         struct hmm_devmem *devmem;
1215
1216         devmem = container_of(ref, struct hmm_devmem, ref);
1217         wait_for_completion(&devmem->completion);
1218         percpu_ref_exit(ref);
1219 }
1220
1221 static void hmm_devmem_ref_kill(struct percpu_ref *ref)
1222 {
1223         percpu_ref_kill(ref);
1224 }
1225
1226 static vm_fault_t hmm_devmem_fault(struct vm_area_struct *vma,
1227                             unsigned long addr,
1228                             const struct page *page,
1229                             unsigned int flags,
1230                             pmd_t *pmdp)
1231 {
1232         struct hmm_devmem *devmem = page->pgmap->data;
1233
1234         return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp);
1235 }
1236
1237 static void hmm_devmem_free(struct page *page, void *data)
1238 {
1239         struct hmm_devmem *devmem = data;
1240
1241         page->mapping = NULL;
1242
1243         devmem->ops->free(devmem, page);
1244 }
1245
1246 /*
1247  * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory
1248  *
1249  * @ops: memory event device driver callback (see struct hmm_devmem_ops)
1250  * @device: device struct to bind the resource too
1251  * @size: size in bytes of the device memory to add
1252  * Returns: pointer to new hmm_devmem struct ERR_PTR otherwise
1253  *
1254  * This function first finds an empty range of physical address big enough to
1255  * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which
1256  * in turn allocates struct pages. It does not do anything beyond that; all
1257  * events affecting the memory will go through the various callbacks provided
1258  * by hmm_devmem_ops struct.
1259  *
1260  * Device driver should call this function during device initialization and
1261  * is then responsible of memory management. HMM only provides helpers.
1262  */
1263 struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
1264                                   struct device *device,
1265                                   unsigned long size)
1266 {
1267         struct hmm_devmem *devmem;
1268         resource_size_t addr;
1269         void *result;
1270         int ret;
1271
1272         dev_pagemap_get_ops();
1273
1274         devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL);
1275         if (!devmem)
1276                 return ERR_PTR(-ENOMEM);
1277
1278         init_completion(&devmem->completion);
1279         devmem->pfn_first = -1UL;
1280         devmem->pfn_last = -1UL;
1281         devmem->resource = NULL;
1282         devmem->device = device;
1283         devmem->ops = ops;
1284
1285         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1286                               0, GFP_KERNEL);
1287         if (ret)
1288                 return ERR_PTR(ret);
1289
1290         ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit, &devmem->ref);
1291         if (ret)
1292                 return ERR_PTR(ret);
1293
1294         size = ALIGN(size, PA_SECTION_SIZE);
1295         addr = min((unsigned long)iomem_resource.end,
1296                    (1UL << MAX_PHYSMEM_BITS) - 1);
1297         addr = addr - size + 1UL;
1298
1299         /*
1300          * FIXME add a new helper to quickly walk resource tree and find free
1301          * range
1302          *
1303          * FIXME what about ioport_resource resource ?
1304          */
1305         for (; addr > size && addr >= iomem_resource.start; addr -= size) {
1306                 ret = region_intersects(addr, size, 0, IORES_DESC_NONE);
1307                 if (ret != REGION_DISJOINT)
1308                         continue;
1309
1310                 devmem->resource = devm_request_mem_region(device, addr, size,
1311                                                            dev_name(device));
1312                 if (!devmem->resource)
1313                         return ERR_PTR(-ENOMEM);
1314                 break;
1315         }
1316         if (!devmem->resource)
1317                 return ERR_PTR(-ERANGE);
1318
1319         devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY;
1320         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1321         devmem->pfn_last = devmem->pfn_first +
1322                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1323         devmem->page_fault = hmm_devmem_fault;
1324
1325         devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
1326         devmem->pagemap.res = *devmem->resource;
1327         devmem->pagemap.page_free = hmm_devmem_free;
1328         devmem->pagemap.altmap_valid = false;
1329         devmem->pagemap.ref = &devmem->ref;
1330         devmem->pagemap.data = devmem;
1331         devmem->pagemap.kill = hmm_devmem_ref_kill;
1332
1333         result = devm_memremap_pages(devmem->device, &devmem->pagemap);
1334         if (IS_ERR(result))
1335                 return result;
1336         return devmem;
1337 }
1338 EXPORT_SYMBOL_GPL(hmm_devmem_add);
1339
1340 struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops,
1341                                            struct device *device,
1342                                            struct resource *res)
1343 {
1344         struct hmm_devmem *devmem;
1345         void *result;
1346         int ret;
1347
1348         if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY)
1349                 return ERR_PTR(-EINVAL);
1350
1351         dev_pagemap_get_ops();
1352
1353         devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL);
1354         if (!devmem)
1355                 return ERR_PTR(-ENOMEM);
1356
1357         init_completion(&devmem->completion);
1358         devmem->pfn_first = -1UL;
1359         devmem->pfn_last = -1UL;
1360         devmem->resource = res;
1361         devmem->device = device;
1362         devmem->ops = ops;
1363
1364         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1365                               0, GFP_KERNEL);
1366         if (ret)
1367                 return ERR_PTR(ret);
1368
1369         ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit,
1370                         &devmem->ref);
1371         if (ret)
1372                 return ERR_PTR(ret);
1373
1374         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1375         devmem->pfn_last = devmem->pfn_first +
1376                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1377         devmem->page_fault = hmm_devmem_fault;
1378
1379         devmem->pagemap.type = MEMORY_DEVICE_PUBLIC;
1380         devmem->pagemap.res = *devmem->resource;
1381         devmem->pagemap.page_free = hmm_devmem_free;
1382         devmem->pagemap.altmap_valid = false;
1383         devmem->pagemap.ref = &devmem->ref;
1384         devmem->pagemap.data = devmem;
1385         devmem->pagemap.kill = hmm_devmem_ref_kill;
1386
1387         result = devm_memremap_pages(devmem->device, &devmem->pagemap);
1388         if (IS_ERR(result))
1389                 return result;
1390         return devmem;
1391 }
1392 EXPORT_SYMBOL_GPL(hmm_devmem_add_resource);
1393
1394 /*
1395  * A device driver that wants to handle multiple devices memory through a
1396  * single fake device can use hmm_device to do so. This is purely a helper
1397  * and it is not needed to make use of any HMM functionality.
1398  */
1399 #define HMM_DEVICE_MAX 256
1400
1401 static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX);
1402 static DEFINE_SPINLOCK(hmm_device_lock);
1403 static struct class *hmm_device_class;
1404 static dev_t hmm_device_devt;
1405
1406 static void hmm_device_release(struct device *device)
1407 {
1408         struct hmm_device *hmm_device;
1409
1410         hmm_device = container_of(device, struct hmm_device, device);
1411         spin_lock(&hmm_device_lock);
1412         clear_bit(hmm_device->minor, hmm_device_mask);
1413         spin_unlock(&hmm_device_lock);
1414
1415         kfree(hmm_device);
1416 }
1417
1418 struct hmm_device *hmm_device_new(void *drvdata)
1419 {
1420         struct hmm_device *hmm_device;
1421
1422         hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL);
1423         if (!hmm_device)
1424                 return ERR_PTR(-ENOMEM);
1425
1426         spin_lock(&hmm_device_lock);
1427         hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX);
1428         if (hmm_device->minor >= HMM_DEVICE_MAX) {
1429                 spin_unlock(&hmm_device_lock);
1430                 kfree(hmm_device);
1431                 return ERR_PTR(-EBUSY);
1432         }
1433         set_bit(hmm_device->minor, hmm_device_mask);
1434         spin_unlock(&hmm_device_lock);
1435
1436         dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor);
1437         hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt),
1438                                         hmm_device->minor);
1439         hmm_device->device.release = hmm_device_release;
1440         dev_set_drvdata(&hmm_device->device, drvdata);
1441         hmm_device->device.class = hmm_device_class;
1442         device_initialize(&hmm_device->device);
1443
1444         return hmm_device;
1445 }
1446 EXPORT_SYMBOL(hmm_device_new);
1447
1448 void hmm_device_put(struct hmm_device *hmm_device)
1449 {
1450         put_device(&hmm_device->device);
1451 }
1452 EXPORT_SYMBOL(hmm_device_put);
1453
1454 static int __init hmm_init(void)
1455 {
1456         int ret;
1457
1458         ret = alloc_chrdev_region(&hmm_device_devt, 0,
1459                                   HMM_DEVICE_MAX,
1460                                   "hmm_device");
1461         if (ret)
1462                 return ret;
1463
1464         hmm_device_class = class_create(THIS_MODULE, "hmm_device");
1465         if (IS_ERR(hmm_device_class)) {
1466                 unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX);
1467                 return PTR_ERR(hmm_device_class);
1468         }
1469         return 0;
1470 }
1471
1472 device_initcall(hmm_init);
1473 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */