mm/hmm.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Copyright 2013 Red Hat Inc.
   4  *
   5  * Authors: Jérôme Glisse <jglisse@redhat.com>
   6  */
   7 /*
   8  * Refer to include/linux/hmm.h for information about heterogeneous memory
   9  * management or HMM for short.
  10  */
  11 #include <linux/mm.h>
  12 #include <linux/hmm.h>
  13 #include <linux/init.h>
  14 #include <linux/rmap.h>
  15 #include <linux/swap.h>
  16 #include <linux/slab.h>
  17 #include <linux/sched.h>
  18 #include <linux/mmzone.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/swapops.h>
  21 #include <linux/hugetlb.h>
  22 #include <linux/memremap.h>
  23 #include <linux/sched/mm.h>
  24 #include <linux/jump_label.h>
  25 #include <linux/dma-mapping.h>
  26 #include <linux/mmu_notifier.h>
  27 #include <linux/memory_hotplug.h>
  28
  29 #define PA_SECTION_SIZE (1UL << PA_SECTION_SHIFT)
  30
  31 #if IS_ENABLED(CONFIG_HMM_MIRROR)
  32 static const struct mmu_notifier_ops hmm_mmu_notifier_ops;
  33
  34 /**
  35  * hmm_get_or_create - register HMM against an mm (HMM internal)
  36  *
  37  * @mm: mm struct to attach to
  38  * Returns: returns an HMM object, either by referencing the existing
  39  *          (per-process) object, or by creating a new one.
  40  *
  41  * This is not intended to be used directly by device drivers. If mm already
  42  * has an HMM struct then it get a reference on it and returns it. Otherwise
  43  * it allocates an HMM struct, initializes it, associate it with the mm and
  44  * returns it.
  45  */
  46 static struct hmm *hmm_get_or_create(struct mm_struct *mm)
  47 {
  48         struct hmm *hmm;
  49
  50         lockdep_assert_held_exclusive(&mm->mmap_sem);
  51
  52         /* Abuse the page_table_lock to also protect mm->hmm. */
  53         spin_lock(&mm->page_table_lock);
  54         hmm = mm->hmm;
  55         if (mm->hmm && kref_get_unless_zero(&mm->hmm->kref))
  56                 goto out_unlock;
  57         spin_unlock(&mm->page_table_lock);
  58
  59         hmm = kmalloc(sizeof(*hmm), GFP_KERNEL);
  60         if (!hmm)
  61                 return NULL;
  62         init_waitqueue_head(&hmm->wq);
  63         INIT_LIST_HEAD(&hmm->mirrors);
  64         init_rwsem(&hmm->mirrors_sem);
  65         hmm->mmu_notifier.ops = NULL;
  66         INIT_LIST_HEAD(&hmm->ranges);
  67         mutex_init(&hmm->lock);
  68         kref_init(&hmm->kref);
  69         hmm->notifiers = 0;
  70         hmm->dead = false;
  71         hmm->mm = mm;
  72
  73         hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops;
  74         if (__mmu_notifier_register(&hmm->mmu_notifier, mm)) {
  75                 kfree(hmm);
  76                 return NULL;
  77         }
  78
  79         mmgrab(hmm->mm);
  80
  81         /*
  82          * We hold the exclusive mmap_sem here so we know that mm->hmm is
  83          * still NULL or 0 kref, and is safe to update.
  84          */
  85         spin_lock(&mm->page_table_lock);
  86         mm->hmm = hmm;
  87
  88 out_unlock:
  89         spin_unlock(&mm->page_table_lock);
  90         return hmm;
  91 }
  92
  93 static void hmm_free_rcu(struct rcu_head *rcu)
  94 {
  95         struct hmm *hmm = container_of(rcu, struct hmm, rcu);
  96
  97         mmdrop(hmm->mm);
  98         kfree(hmm);
  99 }
 100
 101 static void hmm_free(struct kref *kref)
 102 {
 103         struct hmm *hmm = container_of(kref, struct hmm, kref);
 104
 105         spin_lock(&hmm->mm->page_table_lock);
 106         if (hmm->mm->hmm == hmm)
 107                 hmm->mm->hmm = NULL;
 108         spin_unlock(&hmm->mm->page_table_lock);
 109
 110         mmu_notifier_unregister_no_release(&hmm->mmu_notifier, hmm->mm);
 111         mmu_notifier_call_srcu(&hmm->rcu, hmm_free_rcu);
 112 }
 113
 114 static inline void hmm_put(struct hmm *hmm)
 115 {
 116         kref_put(&hmm->kref, hmm_free);
 117 }
 118
 119 static void hmm_release(struct mmu_notifier *mn, struct mm_struct *mm)
 120 {
 121         struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier);
 122         struct hmm_mirror *mirror;
 123         struct hmm_range *range;
 124
 125         /* Bail out if hmm is in the process of being freed */
 126         if (!kref_get_unless_zero(&hmm->kref))
 127                 return;
 128
 129         /* Report this HMM as dying. */
 130         hmm->dead = true;
 131
 132         /* Wake-up everyone waiting on any range. */
 133         mutex_lock(&hmm->lock);
 134         list_for_each_entry(range, &hmm->ranges, list)
 135                 range->valid = false;
 136         wake_up_all(&hmm->wq);
 137         mutex_unlock(&hmm->lock);
 138
 139         down_write(&hmm->mirrors_sem);
 140         mirror = list_first_entry_or_null(&hmm->mirrors, struct hmm_mirror,
 141                                           list);
 142         while (mirror) {
 143                 list_del_init(&mirror->list);
 144                 if (mirror->ops->release) {
 145                         /*
 146                          * Drop mirrors_sem so the release callback can wait
 147                          * on any pending work that might itself trigger a
 148                          * mmu_notifier callback and thus would deadlock with
 149                          * us.
 150                          */
 151                         up_write(&hmm->mirrors_sem);
 152                         mirror->ops->release(mirror);
 153                         down_write(&hmm->mirrors_sem);
 154                 }
 155                 mirror = list_first_entry_or_null(&hmm->mirrors,
 156                                                   struct hmm_mirror, list);
 157         }
 158         up_write(&hmm->mirrors_sem);
 159
 160         hmm_put(hmm);
 161 }
 162
 163 static int hmm_invalidate_range_start(struct mmu_notifier *mn,
 164                         const struct mmu_notifier_range *nrange)
 165 {
 166         struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier);
 167         struct hmm_mirror *mirror;
 168         struct hmm_update update;
 169         struct hmm_range *range;
 170         int ret = 0;
 171
 172         if (!kref_get_unless_zero(&hmm->kref))
 173                 return 0;
 174
 175         update.start = nrange->start;
 176         update.end = nrange->end;
 177         update.event = HMM_UPDATE_INVALIDATE;
 178         update.blockable = mmu_notifier_range_blockable(nrange);
 179
 180         if (mmu_notifier_range_blockable(nrange))
 181                 mutex_lock(&hmm->lock);
 182         else if (!mutex_trylock(&hmm->lock)) {
 183                 ret = -EAGAIN;
 184                 goto out;
 185         }
 186         hmm->notifiers++;
 187         list_for_each_entry(range, &hmm->ranges, list) {
 188                 if (update.end < range->start || update.start >= range->end)
 189                         continue;
 190
 191                 range->valid = false;
 192         }
 193         mutex_unlock(&hmm->lock);
 194
 195         if (mmu_notifier_range_blockable(nrange))
 196                 down_read(&hmm->mirrors_sem);
 197         else if (!down_read_trylock(&hmm->mirrors_sem)) {
 198                 ret = -EAGAIN;
 199                 goto out;
 200         }
 201         list_for_each_entry(mirror, &hmm->mirrors, list) {
 202                 int ret;
 203
 204                 ret = mirror->ops->sync_cpu_device_pagetables(mirror, &update);
 205                 if (!update.blockable && ret == -EAGAIN)
 206                         break;
 207         }
 208         up_read(&hmm->mirrors_sem);
 209
 210 out:
 211         hmm_put(hmm);
 212         return ret;
 213 }
 214
 215 static void hmm_invalidate_range_end(struct mmu_notifier *mn,
 216                         const struct mmu_notifier_range *nrange)
 217 {
 218         struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier);
 219
 220         if (!kref_get_unless_zero(&hmm->kref))
 221                 return;
 222
 223         mutex_lock(&hmm->lock);
 224         hmm->notifiers--;
 225         if (!hmm->notifiers) {
 226                 struct hmm_range *range;
 227
 228                 list_for_each_entry(range, &hmm->ranges, list) {
 229                         if (range->valid)
 230                                 continue;
 231                         range->valid = true;
 232                 }
 233                 wake_up_all(&hmm->wq);
 234         }
 235         mutex_unlock(&hmm->lock);
 236
 237         hmm_put(hmm);
 238 }
 239
 240 static const struct mmu_notifier_ops hmm_mmu_notifier_ops = {
 241         .release                = hmm_release,
 242         .invalidate_range_start = hmm_invalidate_range_start,
 243         .invalidate_range_end   = hmm_invalidate_range_end,
 244 };
 245
 246 /*
 247  * hmm_mirror_register() - register a mirror against an mm
 248  *
 249  * @mirror: new mirror struct to register
 250  * @mm: mm to register against
 251  * Return: 0 on success, -ENOMEM if no memory, -EINVAL if invalid arguments
 252  *
 253  * To start mirroring a process address space, the device driver must register
 254  * an HMM mirror struct.
 255  *
 256  * THE mm->mmap_sem MUST BE HELD IN WRITE MODE !
 257  */
 258 int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm)
 259 {
 260         /* Sanity check */
 261         if (!mm || !mirror || !mirror->ops)
 262                 return -EINVAL;
 263
 264         mirror->hmm = hmm_get_or_create(mm);
 265         if (!mirror->hmm)
 266                 return -ENOMEM;
 267
 268         down_write(&mirror->hmm->mirrors_sem);
 269         list_add(&mirror->list, &mirror->hmm->mirrors);
 270         up_write(&mirror->hmm->mirrors_sem);
 271
 272         return 0;
 273 }
 274 EXPORT_SYMBOL(hmm_mirror_register);
 275
 276 /*
 277  * hmm_mirror_unregister() - unregister a mirror
 278  *
 279  * @mirror: mirror struct to unregister
 280  *
 281  * Stop mirroring a process address space, and cleanup.
 282  */
 283 void hmm_mirror_unregister(struct hmm_mirror *mirror)
 284 {
 285         struct hmm *hmm = READ_ONCE(mirror->hmm);
 286
 287         if (hmm == NULL)
 288                 return;
 289
 290         down_write(&hmm->mirrors_sem);
 291         list_del_init(&mirror->list);
 292         /* To protect us against double unregister ... */
 293         mirror->hmm = NULL;
 294         up_write(&hmm->mirrors_sem);
 295
 296         hmm_put(hmm);
 297 }
 298 EXPORT_SYMBOL(hmm_mirror_unregister);
 299
 300 struct hmm_vma_walk {
 301         struct hmm_range        *range;
 302         struct dev_pagemap      *pgmap;
 303         unsigned long           last;
 304         bool                    fault;
 305         bool                    block;
 306 };
 307
 308 static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr,
 309                             bool write_fault, uint64_t *pfn)
 310 {
 311         unsigned int flags = FAULT_FLAG_REMOTE;
 312         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 313         struct hmm_range *range = hmm_vma_walk->range;
 314         struct vm_area_struct *vma = walk->vma;
 315         vm_fault_t ret;
 316
 317         flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY;
 318         flags |= write_fault ? FAULT_FLAG_WRITE : 0;
 319         ret = handle_mm_fault(vma, addr, flags);
 320         if (ret & VM_FAULT_RETRY)
 321                 return -EAGAIN;
 322         if (ret & VM_FAULT_ERROR) {
 323                 *pfn = range->values[HMM_PFN_ERROR];
 324                 return -EFAULT;
 325         }
 326
 327         return -EBUSY;
 328 }
 329
 330 static int hmm_pfns_bad(unsigned long addr,
 331                         unsigned long end,
 332                         struct mm_walk *walk)
 333 {
 334         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 335         struct hmm_range *range = hmm_vma_walk->range;
 336         uint64_t *pfns = range->pfns;
 337         unsigned long i;
 338
 339         i = (addr - range->start) >> PAGE_SHIFT;
 340         for (; addr < end; addr += PAGE_SIZE, i++)
 341                 pfns[i] = range->values[HMM_PFN_ERROR];
 342
 343         return 0;
 344 }
 345
 346 /*
 347  * hmm_vma_walk_hole() - handle a range lacking valid pmd or pte(s)
 348  * @start: range virtual start address (inclusive)
 349  * @end: range virtual end address (exclusive)
 350  * @fault: should we fault or not ?
 351  * @write_fault: write fault ?
 352  * @walk: mm_walk structure
 353  * Return: 0 on success, -EBUSY after page fault, or page fault error
 354  *
 355  * This function will be called whenever pmd_none() or pte_none() returns true,
 356  * or whenever there is no page directory covering the virtual address range.
 357  */
 358 static int hmm_vma_walk_hole_(unsigned long addr, unsigned long end,
 359                               bool fault, bool write_fault,
 360                               struct mm_walk *walk)
 361 {
 362         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 363         struct hmm_range *range = hmm_vma_walk->range;
 364         uint64_t *pfns = range->pfns;
 365         unsigned long i, page_size;
 366
 367         hmm_vma_walk->last = addr;
 368         page_size = hmm_range_page_size(range);
 369         i = (addr - range->start) >> range->page_shift;
 370
 371         for (; addr < end; addr += page_size, i++) {
 372                 pfns[i] = range->values[HMM_PFN_NONE];
 373                 if (fault || write_fault) {
 374                         int ret;
 375
 376                         ret = hmm_vma_do_fault(walk, addr, write_fault,
 377                                                &pfns[i]);
 378                         if (ret != -EBUSY)
 379                                 return ret;
 380                 }
 381         }
 382
 383         return (fault || write_fault) ? -EBUSY : 0;
 384 }
 385
 386 static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 387                                       uint64_t pfns, uint64_t cpu_flags,
 388                                       bool *fault, bool *write_fault)
 389 {
 390         struct hmm_range *range = hmm_vma_walk->range;
 391
 392         if (!hmm_vma_walk->fault)
 393                 return;
 394
 395         /*
 396          * So we not only consider the individual per page request we also
 397          * consider the default flags requested for the range. The API can
 398          * be use in 2 fashions. The first one where the HMM user coalesce
 399          * multiple page fault into one request and set flags per pfns for
 400          * of those faults. The second one where the HMM user want to pre-
 401          * fault a range with specific flags. For the latter one it is a
 402          * waste to have the user pre-fill the pfn arrays with a default
 403          * flags value.
 404          */
 405         pfns = (pfns & range->pfn_flags_mask) | range->default_flags;
 406
 407         /* We aren't ask to do anything ... */
 408         if (!(pfns & range->flags[HMM_PFN_VALID]))
 409                 return;
 410         /* If this is device memory than only fault if explicitly requested */
 411         if ((cpu_flags & range->flags[HMM_PFN_DEVICE_PRIVATE])) {
 412                 /* Do we fault on device memory ? */
 413                 if (pfns & range->flags[HMM_PFN_DEVICE_PRIVATE]) {
 414                         *write_fault = pfns & range->flags[HMM_PFN_WRITE];
 415                         *fault = true;
 416                 }
 417                 return;
 418         }
 419
 420         /* If CPU page table is not valid then we need to fault */
 421         *fault = !(cpu_flags & range->flags[HMM_PFN_VALID]);
 422         /* Need to write fault ? */
 423         if ((pfns & range->flags[HMM_PFN_WRITE]) &&
 424             !(cpu_flags & range->flags[HMM_PFN_WRITE])) {
 425                 *write_fault = true;
 426                 *fault = true;
 427         }
 428 }
 429
 430 static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 431                                  const uint64_t *pfns, unsigned long npages,
 432                                  uint64_t cpu_flags, bool *fault,
 433                                  bool *write_fault)
 434 {
 435         unsigned long i;
 436
 437         if (!hmm_vma_walk->fault) {
 438                 *fault = *write_fault = false;
 439                 return;
 440         }
 441
 442         *fault = *write_fault = false;
 443         for (i = 0; i < npages; ++i) {
 444                 hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags,
 445                                    fault, write_fault);
 446                 if ((*write_fault))
 447                         return;
 448         }
 449 }
 450
 451 static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
 452                              struct mm_walk *walk)
 453 {
 454         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 455         struct hmm_range *range = hmm_vma_walk->range;
 456         bool fault, write_fault;
 457         unsigned long i, npages;
 458         uint64_t *pfns;
 459
 460         i = (addr - range->start) >> PAGE_SHIFT;
 461         npages = (end - addr) >> PAGE_SHIFT;
 462         pfns = &range->pfns[i];
 463         hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 464                              0, &fault, &write_fault);
 465         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 466 }
 467
 468 static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd)
 469 {
 470         if (pmd_protnone(pmd))
 471                 return 0;
 472         return pmd_write(pmd) ? range->flags[HMM_PFN_VALID] |
 473                                 range->flags[HMM_PFN_WRITE] :
 474                                 range->flags[HMM_PFN_VALID];
 475 }
 476
 477 static inline uint64_t pud_to_hmm_pfn_flags(struct hmm_range *range, pud_t pud)
 478 {
 479         if (!pud_present(pud))
 480                 return 0;
 481         return pud_write(pud) ? range->flags[HMM_PFN_VALID] |
 482                                 range->flags[HMM_PFN_WRITE] :
 483                                 range->flags[HMM_PFN_VALID];
 484 }
 485
 486 static int hmm_vma_handle_pmd(struct mm_walk *walk,
 487                               unsigned long addr,
 488                               unsigned long end,
 489                               uint64_t *pfns,
 490                               pmd_t pmd)
 491 {
 492 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 493         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 494         struct hmm_range *range = hmm_vma_walk->range;
 495         unsigned long pfn, npages, i;
 496         bool fault, write_fault;
 497         uint64_t cpu_flags;
 498
 499         npages = (end - addr) >> PAGE_SHIFT;
 500         cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
 501         hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags,
 502                              &fault, &write_fault);
 503
 504         if (pmd_protnone(pmd) || fault || write_fault)
 505                 return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 506
 507         pfn = pmd_pfn(pmd) + pte_index(addr);
 508         for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++) {
 509                 if (pmd_devmap(pmd)) {
 510                         hmm_vma_walk->pgmap = get_dev_pagemap(pfn,
 511                                               hmm_vma_walk->pgmap);
 512                         if (unlikely(!hmm_vma_walk->pgmap))
 513                                 return -EBUSY;
 514                 }
 515                 pfns[i] = hmm_device_entry_from_pfn(range, pfn) | cpu_flags;
 516         }
 517         if (hmm_vma_walk->pgmap) {
 518                 put_dev_pagemap(hmm_vma_walk->pgmap);
 519                 hmm_vma_walk->pgmap = NULL;
 520         }
 521         hmm_vma_walk->last = end;
 522         return 0;
 523 #else
 524         /* If THP is not enabled then we should never reach that code ! */
 525         return -EINVAL;
 526 #endif
 527 }
 528
 529 static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte)
 530 {
 531         if (pte_none(pte) || !pte_present(pte) || pte_protnone(pte))
 532                 return 0;
 533         return pte_write(pte) ? range->flags[HMM_PFN_VALID] |
 534                                 range->flags[HMM_PFN_WRITE] :
 535                                 range->flags[HMM_PFN_VALID];
 536 }
 537
 538 static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
 539                               unsigned long end, pmd_t *pmdp, pte_t *ptep,
 540                               uint64_t *pfn)
 541 {
 542         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 543         struct hmm_range *range = hmm_vma_walk->range;
 544         struct vm_area_struct *vma = walk->vma;
 545         bool fault, write_fault;
 546         uint64_t cpu_flags;
 547         pte_t pte = *ptep;
 548         uint64_t orig_pfn = *pfn;
 549
 550         *pfn = range->values[HMM_PFN_NONE];
 551         fault = write_fault = false;
 552
 553         if (pte_none(pte)) {
 554                 hmm_pte_need_fault(hmm_vma_walk, orig_pfn, 0,
 555                                    &fault, &write_fault);
 556                 if (fault || write_fault)
 557                         goto fault;
 558                 return 0;
 559         }
 560
 561         if (!pte_present(pte)) {
 562                 swp_entry_t entry = pte_to_swp_entry(pte);
 563
 564                 if (!non_swap_entry(entry)) {
 565                         if (fault || write_fault)
 566                                 goto fault;
 567                         return 0;
 568                 }
 569
 570                 /*
 571                  * This is a special swap entry, ignore migration, use
 572                  * device and report anything else as error.
 573                  */
 574                 if (is_device_private_entry(entry)) {
 575                         cpu_flags = range->flags[HMM_PFN_VALID] |
 576                                 range->flags[HMM_PFN_DEVICE_PRIVATE];
 577                         cpu_flags |= is_write_device_private_entry(entry) ?
 578                                 range->flags[HMM_PFN_WRITE] : 0;
 579                         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 580                                            &fault, &write_fault);
 581                         if (fault || write_fault)
 582                                 goto fault;
 583                         *pfn = hmm_device_entry_from_pfn(range,
 584                                             swp_offset(entry));
 585                         *pfn |= cpu_flags;
 586                         return 0;
 587                 }
 588
 589                 if (is_migration_entry(entry)) {
 590                         if (fault || write_fault) {
 591                                 pte_unmap(ptep);
 592                                 hmm_vma_walk->last = addr;
 593                                 migration_entry_wait(vma->vm_mm,
 594                                                      pmdp, addr);
 595                                 return -EBUSY;
 596                         }
 597                         return 0;
 598                 }
 599
 600                 /* Report error for everything else */
 601                 *pfn = range->values[HMM_PFN_ERROR];
 602                 return -EFAULT;
 603         } else {
 604                 cpu_flags = pte_to_hmm_pfn_flags(range, pte);
 605                 hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 606                                    &fault, &write_fault);
 607         }
 608
 609         if (fault || write_fault)
 610                 goto fault;
 611
 612         if (pte_devmap(pte)) {
 613                 hmm_vma_walk->pgmap = get_dev_pagemap(pte_pfn(pte),
 614                                               hmm_vma_walk->pgmap);
 615                 if (unlikely(!hmm_vma_walk->pgmap))
 616                         return -EBUSY;
 617         } else if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && pte_special(pte)) {
 618                 *pfn = range->values[HMM_PFN_SPECIAL];
 619                 return -EFAULT;
 620         }
 621
 622         *pfn = hmm_device_entry_from_pfn(range, pte_pfn(pte)) | cpu_flags;
 623         return 0;
 624
 625 fault:
 626         if (hmm_vma_walk->pgmap) {
 627                 put_dev_pagemap(hmm_vma_walk->pgmap);
 628                 hmm_vma_walk->pgmap = NULL;
 629         }
 630         pte_unmap(ptep);
 631         /* Fault any virtual address we were asked to fault */
 632         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 633 }
 634
 635 static int hmm_vma_walk_pmd(pmd_t *pmdp,
 636                             unsigned long start,
 637                             unsigned long end,
 638                             struct mm_walk *walk)
 639 {
 640         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 641         struct hmm_range *range = hmm_vma_walk->range;
 642         struct vm_area_struct *vma = walk->vma;
 643         uint64_t *pfns = range->pfns;
 644         unsigned long addr = start, i;
 645         pte_t *ptep;
 646         pmd_t pmd;
 647
 648
 649 again:
 650         pmd = READ_ONCE(*pmdp);
 651         if (pmd_none(pmd))
 652                 return hmm_vma_walk_hole(start, end, walk);
 653
 654         if (pmd_huge(pmd) && (range->vma->vm_flags & VM_HUGETLB))
 655                 return hmm_pfns_bad(start, end, walk);
 656
 657         if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
 658                 bool fault, write_fault;
 659                 unsigned long npages;
 660                 uint64_t *pfns;
 661
 662                 i = (addr - range->start) >> PAGE_SHIFT;
 663                 npages = (end - addr) >> PAGE_SHIFT;
 664                 pfns = &range->pfns[i];
 665
 666                 hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 667                                      0, &fault, &write_fault);
 668                 if (fault || write_fault) {
 669                         hmm_vma_walk->last = addr;
 670                         pmd_migration_entry_wait(vma->vm_mm, pmdp);
 671                         return -EBUSY;
 672                 }
 673                 return 0;
 674         } else if (!pmd_present(pmd))
 675                 return hmm_pfns_bad(start, end, walk);
 676
 677         if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
 678                 /*
 679                  * No need to take pmd_lock here, even if some other threads
 680                  * is splitting the huge pmd we will get that event through
 681                  * mmu_notifier callback.
 682                  *
 683                  * So just read pmd value and check again its a transparent
 684                  * huge or device mapping one and compute corresponding pfn
 685                  * values.
 686                  */
 687                 pmd = pmd_read_atomic(pmdp);
 688                 barrier();
 689                 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
 690                         goto again;
 691
 692                 i = (addr - range->start) >> PAGE_SHIFT;
 693                 return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd);
 694         }
 695
 696         /*
 697          * We have handled all the valid case above ie either none, migration,
 698          * huge or transparent huge. At this point either it is a valid pmd
 699          * entry pointing to pte directory or it is a bad pmd that will not
 700          * recover.
 701          */
 702         if (pmd_bad(pmd))
 703                 return hmm_pfns_bad(start, end, walk);
 704
 705         ptep = pte_offset_map(pmdp, addr);
 706         i = (addr - range->start) >> PAGE_SHIFT;
 707         for (; addr < end; addr += PAGE_SIZE, ptep++, i++) {
 708                 int r;
 709
 710                 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]);
 711                 if (r) {
 712                         /* hmm_vma_handle_pte() did unmap pte directory */
 713                         hmm_vma_walk->last = addr;
 714                         return r;
 715                 }
 716         }
 717         if (hmm_vma_walk->pgmap) {
 718                 /*
 719                  * We do put_dev_pagemap() here and not in hmm_vma_handle_pte()
 720                  * so that we can leverage get_dev_pagemap() optimization which
 721                  * will not re-take a reference on a pgmap if we already have
 722                  * one.
 723                  */
 724                 put_dev_pagemap(hmm_vma_walk->pgmap);
 725                 hmm_vma_walk->pgmap = NULL;
 726         }
 727         pte_unmap(ptep - 1);
 728
 729         hmm_vma_walk->last = addr;
 730         return 0;
 731 }
 732
 733 static int hmm_vma_walk_pud(pud_t *pudp,
 734                             unsigned long start,
 735                             unsigned long end,
 736                             struct mm_walk *walk)
 737 {
 738         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 739         struct hmm_range *range = hmm_vma_walk->range;
 740         unsigned long addr = start, next;
 741         pmd_t *pmdp;
 742         pud_t pud;
 743         int ret;
 744
 745 again:
 746         pud = READ_ONCE(*pudp);
 747         if (pud_none(pud))
 748                 return hmm_vma_walk_hole(start, end, walk);
 749
 750         if (pud_huge(pud) && pud_devmap(pud)) {
 751                 unsigned long i, npages, pfn;
 752                 uint64_t *pfns, cpu_flags;
 753                 bool fault, write_fault;
 754
 755                 if (!pud_present(pud))
 756                         return hmm_vma_walk_hole(start, end, walk);
 757
 758                 i = (addr - range->start) >> PAGE_SHIFT;
 759                 npages = (end - addr) >> PAGE_SHIFT;
 760                 pfns = &range->pfns[i];
 761
 762                 cpu_flags = pud_to_hmm_pfn_flags(range, pud);
 763                 hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 764                                      cpu_flags, &fault, &write_fault);
 765                 if (fault || write_fault)
 766                         return hmm_vma_walk_hole_(addr, end, fault,
 767                                                 write_fault, walk);
 768
 769                 pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 770                 for (i = 0; i < npages; ++i, ++pfn) {
 771                         hmm_vma_walk->pgmap = get_dev_pagemap(pfn,
 772                                               hmm_vma_walk->pgmap);
 773                         if (unlikely(!hmm_vma_walk->pgmap))
 774                                 return -EBUSY;
 775                         pfns[i] = hmm_device_entry_from_pfn(range, pfn) |
 776                                   cpu_flags;
 777                 }
 778                 if (hmm_vma_walk->pgmap) {
 779                         put_dev_pagemap(hmm_vma_walk->pgmap);
 780                         hmm_vma_walk->pgmap = NULL;
 781                 }
 782                 hmm_vma_walk->last = end;
 783                 return 0;
 784         }
 785
 786         split_huge_pud(walk->vma, pudp, addr);
 787         if (pud_none(*pudp))
 788                 goto again;
 789
 790         pmdp = pmd_offset(pudp, addr);
 791         do {
 792                 next = pmd_addr_end(addr, end);
 793                 ret = hmm_vma_walk_pmd(pmdp, addr, next, walk);
 794                 if (ret)
 795                         return ret;
 796         } while (pmdp++, addr = next, addr != end);
 797
 798         return 0;
 799 }
 800
 801 static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask,
 802                                       unsigned long start, unsigned long end,
 803                                       struct mm_walk *walk)
 804 {
 805 #ifdef CONFIG_HUGETLB_PAGE
 806         unsigned long addr = start, i, pfn, mask, size, pfn_inc;
 807         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 808         struct hmm_range *range = hmm_vma_walk->range;
 809         struct vm_area_struct *vma = walk->vma;
 810         struct hstate *h = hstate_vma(vma);
 811         uint64_t orig_pfn, cpu_flags;
 812         bool fault, write_fault;
 813         spinlock_t *ptl;
 814         pte_t entry;
 815         int ret = 0;
 816
 817         size = 1UL << huge_page_shift(h);
 818         mask = size - 1;
 819         if (range->page_shift != PAGE_SHIFT) {
 820                 /* Make sure we are looking at full page. */
 821                 if (start & mask)
 822                         return -EINVAL;
 823                 if (end < (start + size))
 824                         return -EINVAL;
 825                 pfn_inc = size >> PAGE_SHIFT;
 826         } else {
 827                 pfn_inc = 1;
 828                 size = PAGE_SIZE;
 829         }
 830
 831
 832         ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
 833         entry = huge_ptep_get(pte);
 834
 835         i = (start - range->start) >> range->page_shift;
 836         orig_pfn = range->pfns[i];
 837         range->pfns[i] = range->values[HMM_PFN_NONE];
 838         cpu_flags = pte_to_hmm_pfn_flags(range, entry);
 839         fault = write_fault = false;
 840         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 841                            &fault, &write_fault);
 842         if (fault || write_fault) {
 843                 ret = -ENOENT;
 844                 goto unlock;
 845         }
 846
 847         pfn = pte_pfn(entry) + ((start & mask) >> range->page_shift);
 848         for (; addr < end; addr += size, i++, pfn += pfn_inc)
 849                 range->pfns[i] = hmm_device_entry_from_pfn(range, pfn) |
 850                                  cpu_flags;
 851         hmm_vma_walk->last = end;
 852
 853 unlock:
 854         spin_unlock(ptl);
 855
 856         if (ret == -ENOENT)
 857                 return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 858
 859         return ret;
 860 #else /* CONFIG_HUGETLB_PAGE */
 861         return -EINVAL;
 862 #endif
 863 }
 864
 865 static void hmm_pfns_clear(struct hmm_range *range,
 866                            uint64_t *pfns,
 867                            unsigned long addr,
 868                            unsigned long end)
 869 {
 870         for (; addr < end; addr += PAGE_SIZE, pfns++)
 871                 *pfns = range->values[HMM_PFN_NONE];
 872 }
 873
 874 /*
 875  * hmm_range_register() - start tracking change to CPU page table over a range
 876  * @range: range
 877  * @mm: the mm struct for the range of virtual address
 878  * @start: start virtual address (inclusive)
 879  * @end: end virtual address (exclusive)
 880  * @page_shift: expect page shift for the range
 881  * Returns 0 on success, -EFAULT if the address space is no longer valid
 882  *
 883  * Track updates to the CPU page table see include/linux/hmm.h
 884  */
 885 int hmm_range_register(struct hmm_range *range,
 886                        struct hmm_mirror *mirror,
 887                        unsigned long start,
 888                        unsigned long end,
 889                        unsigned page_shift)
 890 {
 891         unsigned long mask = ((1UL << page_shift) - 1UL);
 892         struct hmm *hmm = mirror->hmm;
 893
 894         range->valid = false;
 895         range->hmm = NULL;
 896
 897         if ((start & mask) || (end & mask))
 898                 return -EINVAL;
 899         if (start >= end)
 900                 return -EINVAL;
 901
 902         range->page_shift = page_shift;
 903         range->start = start;
 904         range->end = end;
 905
 906         /* Check if hmm_mm_destroy() was call. */
 907         if (hmm->mm == NULL || hmm->dead)
 908                 return -EFAULT;
 909
 910         /* Initialize range to track CPU page table updates. */
 911         mutex_lock(&hmm->lock);
 912
 913         range->hmm = hmm;
 914         kref_get(&hmm->kref);
 915         list_add(&range->list, &hmm->ranges);
 916
 917         /*
 918          * If there are any concurrent notifiers we have to wait for them for
 919          * the range to be valid (see hmm_range_wait_until_valid()).
 920          */
 921         if (!hmm->notifiers)
 922                 range->valid = true;
 923         mutex_unlock(&hmm->lock);
 924
 925         return 0;
 926 }
 927 EXPORT_SYMBOL(hmm_range_register);
 928
 929 /*
 930  * hmm_range_unregister() - stop tracking change to CPU page table over a range
 931  * @range: range
 932  *
 933  * Range struct is used to track updates to the CPU page table after a call to
 934  * hmm_range_register(). See include/linux/hmm.h for how to use it.
 935  */
 936 void hmm_range_unregister(struct hmm_range *range)
 937 {
 938         struct hmm *hmm = range->hmm;
 939
 940         /* Sanity check this really should not happen. */
 941         if (hmm == NULL || range->end <= range->start)
 942                 return;
 943
 944         mutex_lock(&hmm->lock);
 945         list_del(&range->list);
 946         mutex_unlock(&hmm->lock);
 947
 948         /* Drop reference taken by hmm_range_register() */
 949         range->valid = false;
 950         hmm_put(hmm);
 951         range->hmm = NULL;
 952 }
 953 EXPORT_SYMBOL(hmm_range_unregister);
 954
 955 /*
 956  * hmm_range_snapshot() - snapshot CPU page table for a range
 957  * @range: range
 958  * Return: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid
 959  *          permission (for instance asking for write and range is read only),
 960  *          -EAGAIN if you need to retry, -EFAULT invalid (ie either no valid
 961  *          vma or it is illegal to access that range), number of valid pages
 962  *          in range->pfns[] (from range start address).
 963  *
 964  * This snapshots the CPU page table for a range of virtual addresses. Snapshot
 965  * validity is tracked by range struct. See in include/linux/hmm.h for example
 966  * on how to use.
 967  */
 968 long hmm_range_snapshot(struct hmm_range *range)
 969 {
 970         const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
 971         unsigned long start = range->start, end;
 972         struct hmm_vma_walk hmm_vma_walk;
 973         struct hmm *hmm = range->hmm;
 974         struct vm_area_struct *vma;
 975         struct mm_walk mm_walk;
 976
 977         /* Check if hmm_mm_destroy() was call. */
 978         if (hmm->mm == NULL || hmm->dead)
 979                 return -EFAULT;
 980
 981         do {
 982                 /* If range is no longer valid force retry. */
 983                 if (!range->valid)
 984                         return -EAGAIN;
 985
 986                 vma = find_vma(hmm->mm, start);
 987                 if (vma == NULL || (vma->vm_flags & device_vma))
 988                         return -EFAULT;
 989
 990                 if (is_vm_hugetlb_page(vma)) {
 991                         if (huge_page_shift(hstate_vma(vma)) !=
 992                                     range->page_shift &&
 993                             range->page_shift != PAGE_SHIFT)
 994                                 return -EINVAL;
 995                 } else {
 996                         if (range->page_shift != PAGE_SHIFT)
 997                                 return -EINVAL;
 998                 }
 999
1000                 if (!(vma->vm_flags & VM_READ)) {
1001                         /*
1002                          * If vma do not allow read access, then assume that it
1003                          * does not allow write access, either. HMM does not
1004                          * support architecture that allow write without read.
1005                          */
1006                         hmm_pfns_clear(range, range->pfns,
1007                                 range->start, range->end);
1008                         return -EPERM;
1009                 }
1010
1011                 range->vma = vma;
1012                 hmm_vma_walk.pgmap = NULL;
1013                 hmm_vma_walk.last = start;
1014                 hmm_vma_walk.fault = false;
1015                 hmm_vma_walk.range = range;
1016                 mm_walk.private = &hmm_vma_walk;
1017                 end = min(range->end, vma->vm_end);
1018
1019                 mm_walk.vma = vma;
1020                 mm_walk.mm = vma->vm_mm;
1021                 mm_walk.pte_entry = NULL;
1022                 mm_walk.test_walk = NULL;
1023                 mm_walk.hugetlb_entry = NULL;
1024                 mm_walk.pud_entry = hmm_vma_walk_pud;
1025                 mm_walk.pmd_entry = hmm_vma_walk_pmd;
1026                 mm_walk.pte_hole = hmm_vma_walk_hole;
1027                 mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry;
1028
1029                 walk_page_range(start, end, &mm_walk);
1030                 start = end;
1031         } while (start < range->end);
1032
1033         return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
1034 }
1035 EXPORT_SYMBOL(hmm_range_snapshot);
1036
1037 /*
1038  * hmm_range_fault() - try to fault some address in a virtual address range
1039  * @range: range being faulted
1040  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
1041  * Return: number of valid pages in range->pfns[] (from range start
1042  *          address). This may be zero. If the return value is negative,
1043  *          then one of the following values may be returned:
1044  *
1045  *           -EINVAL  invalid arguments or mm or virtual address are in an
1046  *                    invalid vma (for instance device file vma).
1047  *           -ENOMEM: Out of memory.
1048  *           -EPERM:  Invalid permission (for instance asking for write and
1049  *                    range is read only).
1050  *           -EAGAIN: If you need to retry and mmap_sem was drop. This can only
1051  *                    happens if block argument is false.
1052  *           -EBUSY:  If the the range is being invalidated and you should wait
1053  *                    for invalidation to finish.
1054  *           -EFAULT: Invalid (ie either no valid vma or it is illegal to access
1055  *                    that range), number of valid pages in range->pfns[] (from
1056  *                    range start address).
1057  *
1058  * This is similar to a regular CPU page fault except that it will not trigger
1059  * any memory migration if the memory being faulted is not accessible by CPUs
1060  * and caller does not ask for migration.
1061  *
1062  * On error, for one virtual address in the range, the function will mark the
1063  * corresponding HMM pfn entry with an error flag.
1064  */
1065 long hmm_range_fault(struct hmm_range *range, bool block)
1066 {
1067         const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
1068         unsigned long start = range->start, end;
1069         struct hmm_vma_walk hmm_vma_walk;
1070         struct hmm *hmm = range->hmm;
1071         struct vm_area_struct *vma;
1072         struct mm_walk mm_walk;
1073         int ret;
1074
1075         /* Check if hmm_mm_destroy() was call. */
1076         if (hmm->mm == NULL || hmm->dead)
1077                 return -EFAULT;
1078
1079         do {
1080                 /* If range is no longer valid force retry. */
1081                 if (!range->valid) {
1082                         up_read(&hmm->mm->mmap_sem);
1083                         return -EAGAIN;
1084                 }
1085
1086                 vma = find_vma(hmm->mm, start);
1087                 if (vma == NULL || (vma->vm_flags & device_vma))
1088                         return -EFAULT;
1089
1090                 if (is_vm_hugetlb_page(vma)) {
1091                         if (huge_page_shift(hstate_vma(vma)) !=
1092                             range->page_shift &&
1093                             range->page_shift != PAGE_SHIFT)
1094                                 return -EINVAL;
1095                 } else {
1096                         if (range->page_shift != PAGE_SHIFT)
1097                                 return -EINVAL;
1098                 }
1099
1100                 if (!(vma->vm_flags & VM_READ)) {
1101                         /*
1102                          * If vma do not allow read access, then assume that it
1103                          * does not allow write access, either. HMM does not
1104                          * support architecture that allow write without read.
1105                          */
1106                         hmm_pfns_clear(range, range->pfns,
1107                                 range->start, range->end);
1108                         return -EPERM;
1109                 }
1110
1111                 range->vma = vma;
1112                 hmm_vma_walk.pgmap = NULL;
1113                 hmm_vma_walk.last = start;
1114                 hmm_vma_walk.fault = true;
1115                 hmm_vma_walk.block = block;
1116                 hmm_vma_walk.range = range;
1117                 mm_walk.private = &hmm_vma_walk;
1118                 end = min(range->end, vma->vm_end);
1119
1120                 mm_walk.vma = vma;
1121                 mm_walk.mm = vma->vm_mm;
1122                 mm_walk.pte_entry = NULL;
1123                 mm_walk.test_walk = NULL;
1124                 mm_walk.hugetlb_entry = NULL;
1125                 mm_walk.pud_entry = hmm_vma_walk_pud;
1126                 mm_walk.pmd_entry = hmm_vma_walk_pmd;
1127                 mm_walk.pte_hole = hmm_vma_walk_hole;
1128                 mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry;
1129
1130                 do {
1131                         ret = walk_page_range(start, end, &mm_walk);
1132                         start = hmm_vma_walk.last;
1133
1134                         /* Keep trying while the range is valid. */
1135                 } while (ret == -EBUSY && range->valid);
1136
1137                 if (ret) {
1138                         unsigned long i;
1139
1140                         i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
1141                         hmm_pfns_clear(range, &range->pfns[i],
1142                                 hmm_vma_walk.last, range->end);
1143                         return ret;
1144                 }
1145                 start = end;
1146
1147         } while (start < range->end);
1148
1149         return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
1150 }
1151 EXPORT_SYMBOL(hmm_range_fault);
1152
1153 /**
1154  * hmm_range_dma_map() - hmm_range_fault() and dma map page all in one.
1155  * @range: range being faulted
1156  * @device: device against to dma map page to
1157  * @daddrs: dma address of mapped pages
1158  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
1159  * Return: number of pages mapped on success, -EAGAIN if mmap_sem have been
1160  *          drop and you need to try again, some other error value otherwise
1161  *
1162  * Note same usage pattern as hmm_range_fault().
1163  */
1164 long hmm_range_dma_map(struct hmm_range *range,
1165                        struct device *device,
1166                        dma_addr_t *daddrs,
1167                        bool block)
1168 {
1169         unsigned long i, npages, mapped;
1170         long ret;
1171
1172         ret = hmm_range_fault(range, block);
1173         if (ret <= 0)
1174                 return ret ? ret : -EBUSY;
1175
1176         npages = (range->end - range->start) >> PAGE_SHIFT;
1177         for (i = 0, mapped = 0; i < npages; ++i) {
1178                 enum dma_data_direction dir = DMA_TO_DEVICE;
1179                 struct page *page;
1180
1181                 /*
1182                  * FIXME need to update DMA API to provide invalid DMA address
1183                  * value instead of a function to test dma address value. This
1184                  * would remove lot of dumb code duplicated accross many arch.
1185                  *
1186                  * For now setting it to 0 here is good enough as the pfns[]
1187                  * value is what is use to check what is valid and what isn't.
1188                  */
1189                 daddrs[i] = 0;
1190
1191                 page = hmm_device_entry_to_page(range, range->pfns[i]);
1192                 if (page == NULL)
1193                         continue;
1194
1195                 /* Check if range is being invalidated */
1196                 if (!range->valid) {
1197                         ret = -EBUSY;
1198                         goto unmap;
1199                 }
1200
1201                 /* If it is read and write than map bi-directional. */
1202                 if (range->pfns[i] & range->flags[HMM_PFN_WRITE])
1203                         dir = DMA_BIDIRECTIONAL;
1204
1205                 daddrs[i] = dma_map_page(device, page, 0, PAGE_SIZE, dir);
1206                 if (dma_mapping_error(device, daddrs[i])) {
1207                         ret = -EFAULT;
1208                         goto unmap;
1209                 }
1210
1211                 mapped++;
1212         }
1213
1214         return mapped;
1215
1216 unmap:
1217         for (npages = i, i = 0; (i < npages) && mapped; ++i) {
1218                 enum dma_data_direction dir = DMA_TO_DEVICE;
1219                 struct page *page;
1220
1221                 page = hmm_device_entry_to_page(range, range->pfns[i]);
1222                 if (page == NULL)
1223                         continue;
1224
1225                 if (dma_mapping_error(device, daddrs[i]))
1226                         continue;
1227
1228                 /* If it is read and write than map bi-directional. */
1229                 if (range->pfns[i] & range->flags[HMM_PFN_WRITE])
1230                         dir = DMA_BIDIRECTIONAL;
1231
1232                 dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir);
1233                 mapped--;
1234         }
1235
1236         return ret;
1237 }
1238 EXPORT_SYMBOL(hmm_range_dma_map);
1239
1240 /**
1241  * hmm_range_dma_unmap() - unmap range of that was map with hmm_range_dma_map()
1242  * @range: range being unmapped
1243  * @vma: the vma against which the range (optional)
1244  * @device: device against which dma map was done
1245  * @daddrs: dma address of mapped pages
1246  * @dirty: dirty page if it had the write flag set
1247  * Return: number of page unmapped on success, -EINVAL otherwise
1248  *
1249  * Note that caller MUST abide by mmu notifier or use HMM mirror and abide
1250  * to the sync_cpu_device_pagetables() callback so that it is safe here to
1251  * call set_page_dirty(). Caller must also take appropriate locks to avoid
1252  * concurrent mmu notifier or sync_cpu_device_pagetables() to make progress.
1253  */
1254 long hmm_range_dma_unmap(struct hmm_range *range,
1255                          struct vm_area_struct *vma,
1256                          struct device *device,
1257                          dma_addr_t *daddrs,
1258                          bool dirty)
1259 {
1260         unsigned long i, npages;
1261         long cpages = 0;
1262
1263         /* Sanity check. */
1264         if (range->end <= range->start)
1265                 return -EINVAL;
1266         if (!daddrs)
1267                 return -EINVAL;
1268         if (!range->pfns)
1269                 return -EINVAL;
1270
1271         npages = (range->end - range->start) >> PAGE_SHIFT;
1272         for (i = 0; i < npages; ++i) {
1273                 enum dma_data_direction dir = DMA_TO_DEVICE;
1274                 struct page *page;
1275
1276                 page = hmm_device_entry_to_page(range, range->pfns[i]);
1277                 if (page == NULL)
1278                         continue;
1279
1280                 /* If it is read and write than map bi-directional. */
1281                 if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) {
1282                         dir = DMA_BIDIRECTIONAL;
1283
1284                         /*
1285                          * See comments in function description on why it is
1286                          * safe here to call set_page_dirty()
1287                          */
1288                         if (dirty)
1289                                 set_page_dirty(page);
1290                 }
1291
1292                 /* Unmap and clear pfns/dma address */
1293                 dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir);
1294                 range->pfns[i] = range->values[HMM_PFN_NONE];
1295                 /* FIXME see comments in hmm_vma_dma_map() */
1296                 daddrs[i] = 0;
1297                 cpages++;
1298         }
1299
1300         return cpages;
1301 }
1302 EXPORT_SYMBOL(hmm_range_dma_unmap);
1303 #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
1304
1305
1306 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) ||  IS_ENABLED(CONFIG_DEVICE_PUBLIC)
1307 struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
1308                                        unsigned long addr)
1309 {
1310         struct page *page;
1311
1312         page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
1313         if (!page)
1314                 return NULL;
1315         lock_page(page);
1316         return page;
1317 }
1318 EXPORT_SYMBOL(hmm_vma_alloc_locked_page);
1319
1320
1321 static void hmm_devmem_ref_release(struct percpu_ref *ref)
1322 {
1323         struct hmm_devmem *devmem;
1324
1325         devmem = container_of(ref, struct hmm_devmem, ref);
1326         complete(&devmem->completion);
1327 }
1328
1329 static void hmm_devmem_ref_exit(void *data)
1330 {
1331         struct percpu_ref *ref = data;
1332         struct hmm_devmem *devmem;
1333
1334         devmem = container_of(ref, struct hmm_devmem, ref);
1335         wait_for_completion(&devmem->completion);
1336         percpu_ref_exit(ref);
1337 }
1338
1339 static void hmm_devmem_ref_kill(struct percpu_ref *ref)
1340 {
1341         percpu_ref_kill(ref);
1342 }
1343
1344 static vm_fault_t hmm_devmem_fault(struct vm_area_struct *vma,
1345                             unsigned long addr,
1346                             const struct page *page,
1347                             unsigned int flags,
1348                             pmd_t *pmdp)
1349 {
1350         struct hmm_devmem *devmem = page->pgmap->data;
1351
1352         return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp);
1353 }
1354
1355 static void hmm_devmem_free(struct page *page, void *data)
1356 {
1357         struct hmm_devmem *devmem = data;
1358
1359         page->mapping = NULL;
1360
1361         devmem->ops->free(devmem, page);
1362 }
1363
1364 /*
1365  * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory
1366  *
1367  * @ops: memory event device driver callback (see struct hmm_devmem_ops)
1368  * @device: device struct to bind the resource too
1369  * @size: size in bytes of the device memory to add
1370  * Return: pointer to new hmm_devmem struct ERR_PTR otherwise
1371  *
1372  * This function first finds an empty range of physical address big enough to
1373  * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which
1374  * in turn allocates struct pages. It does not do anything beyond that; all
1375  * events affecting the memory will go through the various callbacks provided
1376  * by hmm_devmem_ops struct.
1377  *
1378  * Device driver should call this function during device initialization and
1379  * is then responsible of memory management. HMM only provides helpers.
1380  */
1381 struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
1382                                   struct device *device,
1383                                   unsigned long size)
1384 {
1385         struct hmm_devmem *devmem;
1386         resource_size_t addr;
1387         void *result;
1388         int ret;
1389
1390         dev_pagemap_get_ops();
1391
1392         devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL);
1393         if (!devmem)
1394                 return ERR_PTR(-ENOMEM);
1395
1396         init_completion(&devmem->completion);
1397         devmem->pfn_first = -1UL;
1398         devmem->pfn_last = -1UL;
1399         devmem->resource = NULL;
1400         devmem->device = device;
1401         devmem->ops = ops;
1402
1403         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1404                               0, GFP_KERNEL);
1405         if (ret)
1406                 return ERR_PTR(ret);
1407
1408         ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit, &devmem->ref);
1409         if (ret)
1410                 return ERR_PTR(ret);
1411
1412         size = ALIGN(size, PA_SECTION_SIZE);
1413         addr = min((unsigned long)iomem_resource.end,
1414                    (1UL << MAX_PHYSMEM_BITS) - 1);
1415         addr = addr - size + 1UL;
1416
1417         /*
1418          * FIXME add a new helper to quickly walk resource tree and find free
1419          * range
1420          *
1421          * FIXME what about ioport_resource resource ?
1422          */
1423         for (; addr > size && addr >= iomem_resource.start; addr -= size) {
1424                 ret = region_intersects(addr, size, 0, IORES_DESC_NONE);
1425                 if (ret != REGION_DISJOINT)
1426                         continue;
1427
1428                 devmem->resource = devm_request_mem_region(device, addr, size,
1429                                                            dev_name(device));
1430                 if (!devmem->resource)
1431                         return ERR_PTR(-ENOMEM);
1432                 break;
1433         }
1434         if (!devmem->resource)
1435                 return ERR_PTR(-ERANGE);
1436
1437         devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY;
1438         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1439         devmem->pfn_last = devmem->pfn_first +
1440                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1441         devmem->page_fault = hmm_devmem_fault;
1442
1443         devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
1444         devmem->pagemap.res = *devmem->resource;
1445         devmem->pagemap.page_free = hmm_devmem_free;
1446         devmem->pagemap.altmap_valid = false;
1447         devmem->pagemap.ref = &devmem->ref;
1448         devmem->pagemap.data = devmem;
1449         devmem->pagemap.kill = hmm_devmem_ref_kill;
1450
1451         result = devm_memremap_pages(devmem->device, &devmem->pagemap);
1452         if (IS_ERR(result))
1453                 return result;
1454         return devmem;
1455 }
1456 EXPORT_SYMBOL_GPL(hmm_devmem_add);
1457
1458 struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops,
1459                                            struct device *device,
1460                                            struct resource *res)
1461 {
1462         struct hmm_devmem *devmem;
1463         void *result;
1464         int ret;
1465
1466         if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY)
1467                 return ERR_PTR(-EINVAL);
1468
1469         dev_pagemap_get_ops();
1470
1471         devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL);
1472         if (!devmem)
1473                 return ERR_PTR(-ENOMEM);
1474
1475         init_completion(&devmem->completion);
1476         devmem->pfn_first = -1UL;
1477         devmem->pfn_last = -1UL;
1478         devmem->resource = res;
1479         devmem->device = device;
1480         devmem->ops = ops;
1481
1482         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1483                               0, GFP_KERNEL);
1484         if (ret)
1485                 return ERR_PTR(ret);
1486
1487         ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit,
1488                         &devmem->ref);
1489         if (ret)
1490                 return ERR_PTR(ret);
1491
1492         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1493         devmem->pfn_last = devmem->pfn_first +
1494                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1495         devmem->page_fault = hmm_devmem_fault;
1496
1497         devmem->pagemap.type = MEMORY_DEVICE_PUBLIC;
1498         devmem->pagemap.res = *devmem->resource;
1499         devmem->pagemap.page_free = hmm_devmem_free;
1500         devmem->pagemap.altmap_valid = false;
1501         devmem->pagemap.ref = &devmem->ref;
1502         devmem->pagemap.data = devmem;
1503         devmem->pagemap.kill = hmm_devmem_ref_kill;
1504
1505         result = devm_memremap_pages(devmem->device, &devmem->pagemap);
1506         if (IS_ERR(result))
1507                 return result;
1508         return devmem;
1509 }
1510 EXPORT_SYMBOL_GPL(hmm_devmem_add_resource);
1511
1512 /*
1513  * A device driver that wants to handle multiple devices memory through a
1514  * single fake device can use hmm_device to do so. This is purely a helper
1515  * and it is not needed to make use of any HMM functionality.
1516  */
1517 #define HMM_DEVICE_MAX 256
1518
1519 static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX);
1520 static DEFINE_SPINLOCK(hmm_device_lock);
1521 static struct class *hmm_device_class;
1522 static dev_t hmm_device_devt;
1523
1524 static void hmm_device_release(struct device *device)
1525 {
1526         struct hmm_device *hmm_device;
1527
1528         hmm_device = container_of(device, struct hmm_device, device);
1529         spin_lock(&hmm_device_lock);
1530         clear_bit(hmm_device->minor, hmm_device_mask);
1531         spin_unlock(&hmm_device_lock);
1532
1533         kfree(hmm_device);
1534 }
1535
1536 struct hmm_device *hmm_device_new(void *drvdata)
1537 {
1538         struct hmm_device *hmm_device;
1539
1540         hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL);
1541         if (!hmm_device)
1542                 return ERR_PTR(-ENOMEM);
1543
1544         spin_lock(&hmm_device_lock);
1545         hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX);
1546         if (hmm_device->minor >= HMM_DEVICE_MAX) {
1547                 spin_unlock(&hmm_device_lock);
1548                 kfree(hmm_device);
1549                 return ERR_PTR(-EBUSY);
1550         }
1551         set_bit(hmm_device->minor, hmm_device_mask);
1552         spin_unlock(&hmm_device_lock);
1553
1554         dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor);
1555         hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt),
1556                                         hmm_device->minor);
1557         hmm_device->device.release = hmm_device_release;
1558         dev_set_drvdata(&hmm_device->device, drvdata);
1559         hmm_device->device.class = hmm_device_class;
1560         device_initialize(&hmm_device->device);
1561
1562         return hmm_device;
1563 }
1564 EXPORT_SYMBOL(hmm_device_new);
1565
1566 void hmm_device_put(struct hmm_device *hmm_device)
1567 {
1568         put_device(&hmm_device->device);
1569 }
1570 EXPORT_SYMBOL(hmm_device_put);
1571
1572 static int __init hmm_init(void)
1573 {
1574         int ret;
1575
1576         ret = alloc_chrdev_region(&hmm_device_devt, 0,
1577                                   HMM_DEVICE_MAX,
1578                                   "hmm_device");
1579         if (ret)
1580                 return ret;
1581
1582         hmm_device_class = class_create(THIS_MODULE, "hmm_device");
1583         if (IS_ERR(hmm_device_class)) {
1584                 unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX);
1585                 return PTR_ERR(hmm_device_class);
1586         }
1587         return 0;
1588 }
1589
1590 device_initcall(hmm_init);
1591 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */