mm/hmm.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Copyright 2013 Red Hat Inc.
   4  *
   5  * Authors: Jérôme Glisse <[email protected]>
   6  */
   7 /*
   8  * Refer to include/linux/hmm.h for information about heterogeneous memory
   9  * management or HMM for short.
  10  */
  11 #include <linux/pagewalk.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 static struct mmu_notifier *hmm_alloc_notifier(struct mm_struct *mm)
  30 {
  31         struct hmm *hmm;
  32
  33         hmm = kzalloc(sizeof(*hmm), GFP_KERNEL);
  34         if (!hmm)
  35                 return ERR_PTR(-ENOMEM);
  36
  37         init_waitqueue_head(&hmm->wq);
  38         INIT_LIST_HEAD(&hmm->mirrors);
  39         init_rwsem(&hmm->mirrors_sem);
  40         INIT_LIST_HEAD(&hmm->ranges);
  41         spin_lock_init(&hmm->ranges_lock);
  42         hmm->notifiers = 0;
  43         return &hmm->mmu_notifier;
  44 }
  45
  46 static void hmm_free_notifier(struct mmu_notifier *mn)
  47 {
  48         struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier);
  49
  50         WARN_ON(!list_empty(&hmm->ranges));
  51         WARN_ON(!list_empty(&hmm->mirrors));
  52         kfree(hmm);
  53 }
  54
  55 static void hmm_release(struct mmu_notifier *mn, struct mm_struct *mm)
  56 {
  57         struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier);
  58         struct hmm_mirror *mirror;
  59
  60         /*
  61          * Since hmm_range_register() holds the mmget() lock hmm_release() is
  62          * prevented as long as a range exists.
  63          */
  64         WARN_ON(!list_empty_careful(&hmm->ranges));
  65
  66         down_read(&hmm->mirrors_sem);
  67         list_for_each_entry(mirror, &hmm->mirrors, list) {
  68                 /*
  69                  * Note: The driver is not allowed to trigger
  70                  * hmm_mirror_unregister() from this thread.
  71                  */
  72                 if (mirror->ops->release)
  73                         mirror->ops->release(mirror);
  74         }
  75         up_read(&hmm->mirrors_sem);
  76 }
  77
  78 static void notifiers_decrement(struct hmm *hmm)
  79 {
  80         unsigned long flags;
  81
  82         spin_lock_irqsave(&hmm->ranges_lock, flags);
  83         hmm->notifiers--;
  84         if (!hmm->notifiers) {
  85                 struct hmm_range *range;
  86
  87                 list_for_each_entry(range, &hmm->ranges, list) {
  88                         if (range->valid)
  89                                 continue;
  90                         range->valid = true;
  91                 }
  92                 wake_up_all(&hmm->wq);
  93         }
  94         spin_unlock_irqrestore(&hmm->ranges_lock, flags);
  95 }
  96
  97 static int hmm_invalidate_range_start(struct mmu_notifier *mn,
  98                         const struct mmu_notifier_range *nrange)
  99 {
 100         struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier);
 101         struct hmm_mirror *mirror;
 102         struct hmm_range *range;
 103         unsigned long flags;
 104         int ret = 0;
 105
 106         spin_lock_irqsave(&hmm->ranges_lock, flags);
 107         hmm->notifiers++;
 108         list_for_each_entry(range, &hmm->ranges, list) {
 109                 if (nrange->end < range->start || nrange->start >= range->end)
 110                         continue;
 111
 112                 range->valid = false;
 113         }
 114         spin_unlock_irqrestore(&hmm->ranges_lock, flags);
 115
 116         if (mmu_notifier_range_blockable(nrange))
 117                 down_read(&hmm->mirrors_sem);
 118         else if (!down_read_trylock(&hmm->mirrors_sem)) {
 119                 ret = -EAGAIN;
 120                 goto out;
 121         }
 122
 123         list_for_each_entry(mirror, &hmm->mirrors, list) {
 124                 int rc;
 125
 126                 rc = mirror->ops->sync_cpu_device_pagetables(mirror, nrange);
 127                 if (rc) {
 128                         if (WARN_ON(mmu_notifier_range_blockable(nrange) ||
 129                             rc != -EAGAIN))
 130                                 continue;
 131                         ret = -EAGAIN;
 132                         break;
 133                 }
 134         }
 135         up_read(&hmm->mirrors_sem);
 136
 137 out:
 138         if (ret)
 139                 notifiers_decrement(hmm);
 140         return ret;
 141 }
 142
 143 static void hmm_invalidate_range_end(struct mmu_notifier *mn,
 144                         const struct mmu_notifier_range *nrange)
 145 {
 146         struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier);
 147
 148         notifiers_decrement(hmm);
 149 }
 150
 151 static const struct mmu_notifier_ops hmm_mmu_notifier_ops = {
 152         .release                = hmm_release,
 153         .invalidate_range_start = hmm_invalidate_range_start,
 154         .invalidate_range_end   = hmm_invalidate_range_end,
 155         .alloc_notifier         = hmm_alloc_notifier,
 156         .free_notifier          = hmm_free_notifier,
 157 };
 158
 159 /*
 160  * hmm_mirror_register() - register a mirror against an mm
 161  *
 162  * @mirror: new mirror struct to register
 163  * @mm: mm to register against
 164  * Return: 0 on success, -ENOMEM if no memory, -EINVAL if invalid arguments
 165  *
 166  * To start mirroring a process address space, the device driver must register
 167  * an HMM mirror struct.
 168  *
 169  * The caller cannot unregister the hmm_mirror while any ranges are
 170  * registered.
 171  *
 172  * Callers using this function must put a call to mmu_notifier_synchronize()
 173  * in their module exit functions.
 174  */
 175 int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm)
 176 {
 177         struct mmu_notifier *mn;
 178
 179         lockdep_assert_held_write(&mm->mmap_sem);
 180
 181         /* Sanity check */
 182         if (!mm || !mirror || !mirror->ops)
 183                 return -EINVAL;
 184
 185         mn = mmu_notifier_get_locked(&hmm_mmu_notifier_ops, mm);
 186         if (IS_ERR(mn))
 187                 return PTR_ERR(mn);
 188         mirror->hmm = container_of(mn, struct hmm, mmu_notifier);
 189
 190         down_write(&mirror->hmm->mirrors_sem);
 191         list_add(&mirror->list, &mirror->hmm->mirrors);
 192         up_write(&mirror->hmm->mirrors_sem);
 193
 194         return 0;
 195 }
 196 EXPORT_SYMBOL(hmm_mirror_register);
 197
 198 /*
 199  * hmm_mirror_unregister() - unregister a mirror
 200  *
 201  * @mirror: mirror struct to unregister
 202  *
 203  * Stop mirroring a process address space, and cleanup.
 204  */
 205 void hmm_mirror_unregister(struct hmm_mirror *mirror)
 206 {
 207         struct hmm *hmm = mirror->hmm;
 208
 209         down_write(&hmm->mirrors_sem);
 210         list_del(&mirror->list);
 211         up_write(&hmm->mirrors_sem);
 212         mmu_notifier_put(&hmm->mmu_notifier);
 213 }
 214 EXPORT_SYMBOL(hmm_mirror_unregister);
 215
 216 struct hmm_vma_walk {
 217         struct hmm_range        *range;
 218         struct dev_pagemap      *pgmap;
 219         unsigned long           last;
 220         unsigned int            flags;
 221 };
 222
 223 static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr,
 224                             bool write_fault, uint64_t *pfn)
 225 {
 226         unsigned int flags = FAULT_FLAG_REMOTE;
 227         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 228         struct hmm_range *range = hmm_vma_walk->range;
 229         struct vm_area_struct *vma = walk->vma;
 230         vm_fault_t ret;
 231
 232         if (!vma)
 233                 goto err;
 234
 235         if (hmm_vma_walk->flags & HMM_FAULT_ALLOW_RETRY)
 236                 flags |= FAULT_FLAG_ALLOW_RETRY;
 237         if (write_fault)
 238                 flags |= FAULT_FLAG_WRITE;
 239
 240         ret = handle_mm_fault(vma, addr, flags);
 241         if (ret & VM_FAULT_RETRY) {
 242                 /* Note, handle_mm_fault did up_read(&mm->mmap_sem)) */
 243                 return -EAGAIN;
 244         }
 245         if (ret & VM_FAULT_ERROR)
 246                 goto err;
 247
 248         return -EBUSY;
 249
 250 err:
 251         *pfn = range->values[HMM_PFN_ERROR];
 252         return -EFAULT;
 253 }
 254
 255 static int hmm_pfns_bad(unsigned long addr,
 256                         unsigned long end,
 257                         struct mm_walk *walk)
 258 {
 259         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 260         struct hmm_range *range = hmm_vma_walk->range;
 261         uint64_t *pfns = range->pfns;
 262         unsigned long i;
 263
 264         i = (addr - range->start) >> PAGE_SHIFT;
 265         for (; addr < end; addr += PAGE_SIZE, i++)
 266                 pfns[i] = range->values[HMM_PFN_ERROR];
 267
 268         return 0;
 269 }
 270
 271 /*
 272  * hmm_vma_walk_hole_() - handle a range lacking valid pmd or pte(s)
 273  * @addr: range virtual start address (inclusive)
 274  * @end: range virtual end address (exclusive)
 275  * @fault: should we fault or not ?
 276  * @write_fault: write fault ?
 277  * @walk: mm_walk structure
 278  * Return: 0 on success, -EBUSY after page fault, or page fault error
 279  *
 280  * This function will be called whenever pmd_none() or pte_none() returns true,
 281  * or whenever there is no page directory covering the virtual address range.
 282  */
 283 static int hmm_vma_walk_hole_(unsigned long addr, unsigned long end,
 284                               bool fault, bool write_fault,
 285                               struct mm_walk *walk)
 286 {
 287         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 288         struct hmm_range *range = hmm_vma_walk->range;
 289         uint64_t *pfns = range->pfns;
 290         unsigned long i;
 291
 292         hmm_vma_walk->last = addr;
 293         i = (addr - range->start) >> PAGE_SHIFT;
 294
 295         if (write_fault && walk->vma && !(walk->vma->vm_flags & VM_WRITE))
 296                 return -EPERM;
 297
 298         for (; addr < end; addr += PAGE_SIZE, i++) {
 299                 pfns[i] = range->values[HMM_PFN_NONE];
 300                 if (fault || write_fault) {
 301                         int ret;
 302
 303                         ret = hmm_vma_do_fault(walk, addr, write_fault,
 304                                                &pfns[i]);
 305                         if (ret != -EBUSY)
 306                                 return ret;
 307                 }
 308         }
 309
 310         return (fault || write_fault) ? -EBUSY : 0;
 311 }
 312
 313 static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 314                                       uint64_t pfns, uint64_t cpu_flags,
 315                                       bool *fault, bool *write_fault)
 316 {
 317         struct hmm_range *range = hmm_vma_walk->range;
 318
 319         if (hmm_vma_walk->flags & HMM_FAULT_SNAPSHOT)
 320                 return;
 321
 322         /*
 323          * So we not only consider the individual per page request we also
 324          * consider the default flags requested for the range. The API can
 325          * be used 2 ways. The first one where the HMM user coalesces
 326          * multiple page faults into one request and sets flags per pfn for
 327          * those faults. The second one where the HMM user wants to pre-
 328          * fault a range with specific flags. For the latter one it is a
 329          * waste to have the user pre-fill the pfn arrays with a default
 330          * flags value.
 331          */
 332         pfns = (pfns & range->pfn_flags_mask) | range->default_flags;
 333
 334         /* We aren't ask to do anything ... */
 335         if (!(pfns & range->flags[HMM_PFN_VALID]))
 336                 return;
 337         /* If this is device memory then only fault if explicitly requested */
 338         if ((cpu_flags & range->flags[HMM_PFN_DEVICE_PRIVATE])) {
 339                 /* Do we fault on device memory ? */
 340                 if (pfns & range->flags[HMM_PFN_DEVICE_PRIVATE]) {
 341                         *write_fault = pfns & range->flags[HMM_PFN_WRITE];
 342                         *fault = true;
 343                 }
 344                 return;
 345         }
 346
 347         /* If CPU page table is not valid then we need to fault */
 348         *fault = !(cpu_flags & range->flags[HMM_PFN_VALID]);
 349         /* Need to write fault ? */
 350         if ((pfns & range->flags[HMM_PFN_WRITE]) &&
 351             !(cpu_flags & range->flags[HMM_PFN_WRITE])) {
 352                 *write_fault = true;
 353                 *fault = true;
 354         }
 355 }
 356
 357 static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 358                                  const uint64_t *pfns, unsigned long npages,
 359                                  uint64_t cpu_flags, bool *fault,
 360                                  bool *write_fault)
 361 {
 362         unsigned long i;
 363
 364         if (hmm_vma_walk->flags & HMM_FAULT_SNAPSHOT) {
 365                 *fault = *write_fault = false;
 366                 return;
 367         }
 368
 369         *fault = *write_fault = false;
 370         for (i = 0; i < npages; ++i) {
 371                 hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags,
 372                                    fault, write_fault);
 373                 if ((*write_fault))
 374                         return;
 375         }
 376 }
 377
 378 static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
 379                              struct mm_walk *walk)
 380 {
 381         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 382         struct hmm_range *range = hmm_vma_walk->range;
 383         bool fault, write_fault;
 384         unsigned long i, npages;
 385         uint64_t *pfns;
 386
 387         i = (addr - range->start) >> PAGE_SHIFT;
 388         npages = (end - addr) >> PAGE_SHIFT;
 389         pfns = &range->pfns[i];
 390         hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 391                              0, &fault, &write_fault);
 392         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 393 }
 394
 395 static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd)
 396 {
 397         if (pmd_protnone(pmd))
 398                 return 0;
 399         return pmd_write(pmd) ? range->flags[HMM_PFN_VALID] |
 400                                 range->flags[HMM_PFN_WRITE] :
 401                                 range->flags[HMM_PFN_VALID];
 402 }
 403
 404 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 405 static int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr,
 406                 unsigned long end, uint64_t *pfns, pmd_t pmd)
 407 {
 408         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 409         struct hmm_range *range = hmm_vma_walk->range;
 410         unsigned long pfn, npages, i;
 411         bool fault, write_fault;
 412         uint64_t cpu_flags;
 413
 414         npages = (end - addr) >> PAGE_SHIFT;
 415         cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
 416         hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags,
 417                              &fault, &write_fault);
 418
 419         if (pmd_protnone(pmd) || fault || write_fault)
 420                 return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 421
 422         pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
 423         for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++) {
 424                 if (pmd_devmap(pmd)) {
 425                         hmm_vma_walk->pgmap = get_dev_pagemap(pfn,
 426                                               hmm_vma_walk->pgmap);
 427                         if (unlikely(!hmm_vma_walk->pgmap))
 428                                 return -EBUSY;
 429                 }
 430                 pfns[i] = hmm_device_entry_from_pfn(range, pfn) | cpu_flags;
 431         }
 432         if (hmm_vma_walk->pgmap) {
 433                 put_dev_pagemap(hmm_vma_walk->pgmap);
 434                 hmm_vma_walk->pgmap = NULL;
 435         }
 436         hmm_vma_walk->last = end;
 437         return 0;
 438 }
 439 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 440 /* stub to allow the code below to compile */
 441 int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr,
 442                 unsigned long end, uint64_t *pfns, pmd_t pmd);
 443 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 444
 445 static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte)
 446 {
 447         if (pte_none(pte) || !pte_present(pte) || pte_protnone(pte))
 448                 return 0;
 449         return pte_write(pte) ? range->flags[HMM_PFN_VALID] |
 450                                 range->flags[HMM_PFN_WRITE] :
 451                                 range->flags[HMM_PFN_VALID];
 452 }
 453
 454 static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
 455                               unsigned long end, pmd_t *pmdp, pte_t *ptep,
 456                               uint64_t *pfn)
 457 {
 458         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 459         struct hmm_range *range = hmm_vma_walk->range;
 460         bool fault, write_fault;
 461         uint64_t cpu_flags;
 462         pte_t pte = *ptep;
 463         uint64_t orig_pfn = *pfn;
 464
 465         *pfn = range->values[HMM_PFN_NONE];
 466         fault = write_fault = false;
 467
 468         if (pte_none(pte)) {
 469                 hmm_pte_need_fault(hmm_vma_walk, orig_pfn, 0,
 470                                    &fault, &write_fault);
 471                 if (fault || write_fault)
 472                         goto fault;
 473                 return 0;
 474         }
 475
 476         if (!pte_present(pte)) {
 477                 swp_entry_t entry = pte_to_swp_entry(pte);
 478
 479                 if (!non_swap_entry(entry)) {
 480                         cpu_flags = pte_to_hmm_pfn_flags(range, pte);
 481                         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 482                                            &fault, &write_fault);
 483                         if (fault || write_fault)
 484                                 goto fault;
 485                         return 0;
 486                 }
 487
 488                 /*
 489                  * This is a special swap entry, ignore migration, use
 490                  * device and report anything else as error.
 491                  */
 492                 if (is_device_private_entry(entry)) {
 493                         cpu_flags = range->flags[HMM_PFN_VALID] |
 494                                 range->flags[HMM_PFN_DEVICE_PRIVATE];
 495                         cpu_flags |= is_write_device_private_entry(entry) ?
 496                                 range->flags[HMM_PFN_WRITE] : 0;
 497                         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 498                                            &fault, &write_fault);
 499                         if (fault || write_fault)
 500                                 goto fault;
 501                         *pfn = hmm_device_entry_from_pfn(range,
 502                                             swp_offset(entry));
 503                         *pfn |= cpu_flags;
 504                         return 0;
 505                 }
 506
 507                 if (is_migration_entry(entry)) {
 508                         if (fault || write_fault) {
 509                                 pte_unmap(ptep);
 510                                 hmm_vma_walk->last = addr;
 511                                 migration_entry_wait(walk->mm, pmdp, addr);
 512                                 return -EBUSY;
 513                         }
 514                         return 0;
 515                 }
 516
 517                 /* Report error for everything else */
 518                 *pfn = range->values[HMM_PFN_ERROR];
 519                 return -EFAULT;
 520         } else {
 521                 cpu_flags = pte_to_hmm_pfn_flags(range, pte);
 522                 hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 523                                    &fault, &write_fault);
 524         }
 525
 526         if (fault || write_fault)
 527                 goto fault;
 528
 529         if (pte_devmap(pte)) {
 530                 hmm_vma_walk->pgmap = get_dev_pagemap(pte_pfn(pte),
 531                                               hmm_vma_walk->pgmap);
 532                 if (unlikely(!hmm_vma_walk->pgmap))
 533                         return -EBUSY;
 534         } else if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && pte_special(pte)) {
 535                 *pfn = range->values[HMM_PFN_SPECIAL];
 536                 return -EFAULT;
 537         }
 538
 539         *pfn = hmm_device_entry_from_pfn(range, pte_pfn(pte)) | cpu_flags;
 540         return 0;
 541
 542 fault:
 543         if (hmm_vma_walk->pgmap) {
 544                 put_dev_pagemap(hmm_vma_walk->pgmap);
 545                 hmm_vma_walk->pgmap = NULL;
 546         }
 547         pte_unmap(ptep);
 548         /* Fault any virtual address we were asked to fault */
 549         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 550 }
 551
 552 static int hmm_vma_walk_pmd(pmd_t *pmdp,
 553                             unsigned long start,
 554                             unsigned long end,
 555                             struct mm_walk *walk)
 556 {
 557         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 558         struct hmm_range *range = hmm_vma_walk->range;
 559         uint64_t *pfns = range->pfns;
 560         unsigned long addr = start, i;
 561         pte_t *ptep;
 562         pmd_t pmd;
 563
 564 again:
 565         pmd = READ_ONCE(*pmdp);
 566         if (pmd_none(pmd))
 567                 return hmm_vma_walk_hole(start, end, walk);
 568
 569         if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
 570                 bool fault, write_fault;
 571                 unsigned long npages;
 572                 uint64_t *pfns;
 573
 574                 i = (addr - range->start) >> PAGE_SHIFT;
 575                 npages = (end - addr) >> PAGE_SHIFT;
 576                 pfns = &range->pfns[i];
 577
 578                 hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 579                                      0, &fault, &write_fault);
 580                 if (fault || write_fault) {
 581                         hmm_vma_walk->last = addr;
 582                         pmd_migration_entry_wait(walk->mm, pmdp);
 583                         return -EBUSY;
 584                 }
 585                 return 0;
 586         } else if (!pmd_present(pmd))
 587                 return hmm_pfns_bad(start, end, walk);
 588
 589         if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
 590                 /*
 591                  * No need to take pmd_lock here, even if some other thread
 592                  * is splitting the huge pmd we will get that event through
 593                  * mmu_notifier callback.
 594                  *
 595                  * So just read pmd value and check again it's a transparent
 596                  * huge or device mapping one and compute corresponding pfn
 597                  * values.
 598                  */
 599                 pmd = pmd_read_atomic(pmdp);
 600                 barrier();
 601                 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
 602                         goto again;
 603
 604                 i = (addr - range->start) >> PAGE_SHIFT;
 605                 return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd);
 606         }
 607
 608         /*
 609          * We have handled all the valid cases above ie either none, migration,
 610          * huge or transparent huge. At this point either it is a valid pmd
 611          * entry pointing to pte directory or it is a bad pmd that will not
 612          * recover.
 613          */
 614         if (pmd_bad(pmd))
 615                 return hmm_pfns_bad(start, end, walk);
 616
 617         ptep = pte_offset_map(pmdp, addr);
 618         i = (addr - range->start) >> PAGE_SHIFT;
 619         for (; addr < end; addr += PAGE_SIZE, ptep++, i++) {
 620                 int r;
 621
 622                 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]);
 623                 if (r) {
 624                         /* hmm_vma_handle_pte() did unmap pte directory */
 625                         hmm_vma_walk->last = addr;
 626                         return r;
 627                 }
 628         }
 629         if (hmm_vma_walk->pgmap) {
 630                 /*
 631                  * We do put_dev_pagemap() here and not in hmm_vma_handle_pte()
 632                  * so that we can leverage get_dev_pagemap() optimization which
 633                  * will not re-take a reference on a pgmap if we already have
 634                  * one.
 635                  */
 636                 put_dev_pagemap(hmm_vma_walk->pgmap);
 637                 hmm_vma_walk->pgmap = NULL;
 638         }
 639         pte_unmap(ptep - 1);
 640
 641         hmm_vma_walk->last = addr;
 642         return 0;
 643 }
 644
 645 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && \
 646     defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
 647 static inline uint64_t pud_to_hmm_pfn_flags(struct hmm_range *range, pud_t pud)
 648 {
 649         if (!pud_present(pud))
 650                 return 0;
 651         return pud_write(pud) ? range->flags[HMM_PFN_VALID] |
 652                                 range->flags[HMM_PFN_WRITE] :
 653                                 range->flags[HMM_PFN_VALID];
 654 }
 655
 656 static int hmm_vma_walk_pud(pud_t *pudp, unsigned long start, unsigned long end,
 657                 struct mm_walk *walk)
 658 {
 659         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 660         struct hmm_range *range = hmm_vma_walk->range;
 661         unsigned long addr = start, next;
 662         pmd_t *pmdp;
 663         pud_t pud;
 664         int ret;
 665
 666 again:
 667         pud = READ_ONCE(*pudp);
 668         if (pud_none(pud))
 669                 return hmm_vma_walk_hole(start, end, walk);
 670
 671         if (pud_huge(pud) && pud_devmap(pud)) {
 672                 unsigned long i, npages, pfn;
 673                 uint64_t *pfns, cpu_flags;
 674                 bool fault, write_fault;
 675
 676                 if (!pud_present(pud))
 677                         return hmm_vma_walk_hole(start, end, walk);
 678
 679                 i = (addr - range->start) >> PAGE_SHIFT;
 680                 npages = (end - addr) >> PAGE_SHIFT;
 681                 pfns = &range->pfns[i];
 682
 683                 cpu_flags = pud_to_hmm_pfn_flags(range, pud);
 684                 hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 685                                      cpu_flags, &fault, &write_fault);
 686                 if (fault || write_fault)
 687                         return hmm_vma_walk_hole_(addr, end, fault,
 688                                                 write_fault, walk);
 689
 690                 pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 691                 for (i = 0; i < npages; ++i, ++pfn) {
 692                         hmm_vma_walk->pgmap = get_dev_pagemap(pfn,
 693                                               hmm_vma_walk->pgmap);
 694                         if (unlikely(!hmm_vma_walk->pgmap))
 695                                 return -EBUSY;
 696                         pfns[i] = hmm_device_entry_from_pfn(range, pfn) |
 697                                   cpu_flags;
 698                 }
 699                 if (hmm_vma_walk->pgmap) {
 700                         put_dev_pagemap(hmm_vma_walk->pgmap);
 701                         hmm_vma_walk->pgmap = NULL;
 702                 }
 703                 hmm_vma_walk->last = end;
 704                 return 0;
 705         }
 706
 707         split_huge_pud(walk->vma, pudp, addr);
 708         if (pud_none(*pudp))
 709                 goto again;
 710
 711         pmdp = pmd_offset(pudp, addr);
 712         do {
 713                 next = pmd_addr_end(addr, end);
 714                 ret = hmm_vma_walk_pmd(pmdp, addr, next, walk);
 715                 if (ret)
 716                         return ret;
 717         } while (pmdp++, addr = next, addr != end);
 718
 719         return 0;
 720 }
 721 #else
 722 #define hmm_vma_walk_pud        NULL
 723 #endif
 724
 725 #ifdef CONFIG_HUGETLB_PAGE
 726 static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask,
 727                                       unsigned long start, unsigned long end,
 728                                       struct mm_walk *walk)
 729 {
 730         unsigned long addr = start, i, pfn;
 731         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 732         struct hmm_range *range = hmm_vma_walk->range;
 733         struct vm_area_struct *vma = walk->vma;
 734         uint64_t orig_pfn, cpu_flags;
 735         bool fault, write_fault;
 736         spinlock_t *ptl;
 737         pte_t entry;
 738         int ret = 0;
 739
 740         ptl = huge_pte_lock(hstate_vma(vma), walk->mm, pte);
 741         entry = huge_ptep_get(pte);
 742
 743         i = (start - range->start) >> PAGE_SHIFT;
 744         orig_pfn = range->pfns[i];
 745         range->pfns[i] = range->values[HMM_PFN_NONE];
 746         cpu_flags = pte_to_hmm_pfn_flags(range, entry);
 747         fault = write_fault = false;
 748         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 749                            &fault, &write_fault);
 750         if (fault || write_fault) {
 751                 ret = -ENOENT;
 752                 goto unlock;
 753         }
 754
 755         pfn = pte_pfn(entry) + ((start & ~hmask) >> PAGE_SHIFT);
 756         for (; addr < end; addr += PAGE_SIZE, i++, pfn++)
 757                 range->pfns[i] = hmm_device_entry_from_pfn(range, pfn) |
 758                                  cpu_flags;
 759         hmm_vma_walk->last = end;
 760
 761 unlock:
 762         spin_unlock(ptl);
 763
 764         if (ret == -ENOENT)
 765                 return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 766
 767         return ret;
 768 }
 769 #else
 770 #define hmm_vma_walk_hugetlb_entry NULL
 771 #endif /* CONFIG_HUGETLB_PAGE */
 772
 773 static void hmm_pfns_clear(struct hmm_range *range,
 774                            uint64_t *pfns,
 775                            unsigned long addr,
 776                            unsigned long end)
 777 {
 778         for (; addr < end; addr += PAGE_SIZE, pfns++)
 779                 *pfns = range->values[HMM_PFN_NONE];
 780 }
 781
 782 /*
 783  * hmm_range_register() - start tracking change to CPU page table over a range
 784  * @range: range
 785  * @mm: the mm struct for the range of virtual address
 786  *
 787  * Return: 0 on success, -EFAULT if the address space is no longer valid
 788  *
 789  * Track updates to the CPU page table see include/linux/hmm.h
 790  */
 791 int hmm_range_register(struct hmm_range *range, struct hmm_mirror *mirror)
 792 {
 793         struct hmm *hmm = mirror->hmm;
 794         unsigned long flags;
 795
 796         range->valid = false;
 797         range->hmm = NULL;
 798
 799         if ((range->start & (PAGE_SIZE - 1)) || (range->end & (PAGE_SIZE - 1)))
 800                 return -EINVAL;
 801         if (range->start >= range->end)
 802                 return -EINVAL;
 803
 804         /* Prevent hmm_release() from running while the range is valid */
 805         if (!mmget_not_zero(hmm->mmu_notifier.mm))
 806                 return -EFAULT;
 807
 808         /* Initialize range to track CPU page table updates. */
 809         spin_lock_irqsave(&hmm->ranges_lock, flags);
 810
 811         range->hmm = hmm;
 812         list_add(&range->list, &hmm->ranges);
 813
 814         /*
 815          * If there are any concurrent notifiers we have to wait for them for
 816          * the range to be valid (see hmm_range_wait_until_valid()).
 817          */
 818         if (!hmm->notifiers)
 819                 range->valid = true;
 820         spin_unlock_irqrestore(&hmm->ranges_lock, flags);
 821
 822         return 0;
 823 }
 824 EXPORT_SYMBOL(hmm_range_register);
 825
 826 /*
 827  * hmm_range_unregister() - stop tracking change to CPU page table over a range
 828  * @range: range
 829  *
 830  * Range struct is used to track updates to the CPU page table after a call to
 831  * hmm_range_register(). See include/linux/hmm.h for how to use it.
 832  */
 833 void hmm_range_unregister(struct hmm_range *range)
 834 {
 835         struct hmm *hmm = range->hmm;
 836         unsigned long flags;
 837
 838         spin_lock_irqsave(&hmm->ranges_lock, flags);
 839         list_del_init(&range->list);
 840         spin_unlock_irqrestore(&hmm->ranges_lock, flags);
 841
 842         /* Drop reference taken by hmm_range_register() */
 843         mmput(hmm->mmu_notifier.mm);
 844
 845         /*
 846          * The range is now invalid and the ref on the hmm is dropped, so
 847          * poison the pointer.  Leave other fields in place, for the caller's
 848          * use.
 849          */
 850         range->valid = false;
 851         memset(&range->hmm, POISON_INUSE, sizeof(range->hmm));
 852 }
 853 EXPORT_SYMBOL(hmm_range_unregister);
 854
 855 static const struct mm_walk_ops hmm_walk_ops = {
 856         .pud_entry      = hmm_vma_walk_pud,
 857         .pmd_entry      = hmm_vma_walk_pmd,
 858         .pte_hole       = hmm_vma_walk_hole,
 859         .hugetlb_entry  = hmm_vma_walk_hugetlb_entry,
 860 };
 861
 862 /**
 863  * hmm_range_fault - try to fault some address in a virtual address range
 864  * @range:      range being faulted
 865  * @flags:      HMM_FAULT_* flags
 866  *
 867  * Return: the number of valid pages in range->pfns[] (from range start
 868  * address), which may be zero.  On error one of the following status codes
 869  * can be returned:
 870  *
 871  * -EINVAL:     Invalid arguments or mm or virtual address is in an invalid vma
 872  *              (e.g., device file vma).
 873  * -ENOMEM:     Out of memory.
 874  * -EPERM:      Invalid permission (e.g., asking for write and range is read
 875  *              only).
 876  * -EAGAIN:     A page fault needs to be retried and mmap_sem was dropped.
 877  * -EBUSY:      The range has been invalidated and the caller needs to wait for
 878  *              the invalidation to finish.
 879  * -EFAULT:     Invalid (i.e., either no valid vma or it is illegal to access
 880  *              that range) number of valid pages in range->pfns[] (from
 881  *              range start address).
 882  *
 883  * This is similar to a regular CPU page fault except that it will not trigger
 884  * any memory migration if the memory being faulted is not accessible by CPUs
 885  * and caller does not ask for migration.
 886  *
 887  * On error, for one virtual address in the range, the function will mark the
 888  * corresponding HMM pfn entry with an error flag.
 889  */
 890 long hmm_range_fault(struct hmm_range *range, unsigned int flags)
 891 {
 892         const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
 893         unsigned long start = range->start, end;
 894         struct hmm_vma_walk hmm_vma_walk;
 895         struct hmm *hmm = range->hmm;
 896         struct vm_area_struct *vma;
 897         int ret;
 898
 899         lockdep_assert_held(&hmm->mmu_notifier.mm->mmap_sem);
 900
 901         do {
 902                 /* If range is no longer valid force retry. */
 903                 if (!range->valid)
 904                         return -EBUSY;
 905
 906                 vma = find_vma(hmm->mmu_notifier.mm, start);
 907                 if (vma == NULL || (vma->vm_flags & device_vma))
 908                         return -EFAULT;
 909
 910                 if (!(vma->vm_flags & VM_READ)) {
 911                         /*
 912                          * If vma do not allow read access, then assume that it
 913                          * does not allow write access, either. HMM does not
 914                          * support architecture that allow write without read.
 915                          */
 916                         hmm_pfns_clear(range, range->pfns,
 917                                 range->start, range->end);
 918                         return -EPERM;
 919                 }
 920
 921                 hmm_vma_walk.pgmap = NULL;
 922                 hmm_vma_walk.last = start;
 923                 hmm_vma_walk.flags = flags;
 924                 hmm_vma_walk.range = range;
 925                 end = min(range->end, vma->vm_end);
 926
 927                 walk_page_range(vma->vm_mm, start, end, &hmm_walk_ops,
 928                                 &hmm_vma_walk);
 929
 930                 do {
 931                         ret = walk_page_range(vma->vm_mm, start, end,
 932                                         &hmm_walk_ops, &hmm_vma_walk);
 933                         start = hmm_vma_walk.last;
 934
 935                         /* Keep trying while the range is valid. */
 936                 } while (ret == -EBUSY && range->valid);
 937
 938                 if (ret) {
 939                         unsigned long i;
 940
 941                         i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
 942                         hmm_pfns_clear(range, &range->pfns[i],
 943                                 hmm_vma_walk.last, range->end);
 944                         return ret;
 945                 }
 946                 start = end;
 947
 948         } while (start < range->end);
 949
 950         return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
 951 }
 952 EXPORT_SYMBOL(hmm_range_fault);
 953
 954 /**
 955  * hmm_range_dma_map - hmm_range_fault() and dma map page all in one.
 956  * @range:      range being faulted
 957  * @device:     device to map page to
 958  * @daddrs:     array of dma addresses for the mapped pages
 959  * @flags:      HMM_FAULT_*
 960  *
 961  * Return: the number of pages mapped on success (including zero), or any
 962  * status return from hmm_range_fault() otherwise.
 963  */
 964 long hmm_range_dma_map(struct hmm_range *range, struct device *device,
 965                 dma_addr_t *daddrs, unsigned int flags)
 966 {
 967         unsigned long i, npages, mapped;
 968         long ret;
 969
 970         ret = hmm_range_fault(range, flags);
 971         if (ret <= 0)
 972                 return ret ? ret : -EBUSY;
 973
 974         npages = (range->end - range->start) >> PAGE_SHIFT;
 975         for (i = 0, mapped = 0; i < npages; ++i) {
 976                 enum dma_data_direction dir = DMA_TO_DEVICE;
 977                 struct page *page;
 978
 979                 /*
 980                  * FIXME need to update DMA API to provide invalid DMA address
 981                  * value instead of a function to test dma address value. This
 982                  * would remove lot of dumb code duplicated accross many arch.
 983                  *
 984                  * For now setting it to 0 here is good enough as the pfns[]
 985                  * value is what is use to check what is valid and what isn't.
 986                  */
 987                 daddrs[i] = 0;
 988
 989                 page = hmm_device_entry_to_page(range, range->pfns[i]);
 990                 if (page == NULL)
 991                         continue;
 992
 993                 /* Check if range is being invalidated */
 994                 if (!range->valid) {
 995                         ret = -EBUSY;
 996                         goto unmap;
 997                 }
 998
 999                 /* If it is read and write than map bi-directional. */
1000                 if (range->pfns[i] & range->flags[HMM_PFN_WRITE])
1001                         dir = DMA_BIDIRECTIONAL;
1002
1003                 daddrs[i] = dma_map_page(device, page, 0, PAGE_SIZE, dir);
1004                 if (dma_mapping_error(device, daddrs[i])) {
1005                         ret = -EFAULT;
1006                         goto unmap;
1007                 }
1008
1009                 mapped++;
1010         }
1011
1012         return mapped;
1013
1014 unmap:
1015         for (npages = i, i = 0; (i < npages) && mapped; ++i) {
1016                 enum dma_data_direction dir = DMA_TO_DEVICE;
1017                 struct page *page;
1018
1019                 page = hmm_device_entry_to_page(range, range->pfns[i]);
1020                 if (page == NULL)
1021                         continue;
1022
1023                 if (dma_mapping_error(device, daddrs[i]))
1024                         continue;
1025
1026                 /* If it is read and write than map bi-directional. */
1027                 if (range->pfns[i] & range->flags[HMM_PFN_WRITE])
1028                         dir = DMA_BIDIRECTIONAL;
1029
1030                 dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir);
1031                 mapped--;
1032         }
1033
1034         return ret;
1035 }
1036 EXPORT_SYMBOL(hmm_range_dma_map);
1037
1038 /**
1039  * hmm_range_dma_unmap() - unmap range of that was map with hmm_range_dma_map()
1040  * @range: range being unmapped
1041  * @device: device against which dma map was done
1042  * @daddrs: dma address of mapped pages
1043  * @dirty: dirty page if it had the write flag set
1044  * Return: number of page unmapped on success, -EINVAL otherwise
1045  *
1046  * Note that caller MUST abide by mmu notifier or use HMM mirror and abide
1047  * to the sync_cpu_device_pagetables() callback so that it is safe here to
1048  * call set_page_dirty(). Caller must also take appropriate locks to avoid
1049  * concurrent mmu notifier or sync_cpu_device_pagetables() to make progress.
1050  */
1051 long hmm_range_dma_unmap(struct hmm_range *range,
1052                          struct device *device,
1053                          dma_addr_t *daddrs,
1054                          bool dirty)
1055 {
1056         unsigned long i, npages;
1057         long cpages = 0;
1058
1059         /* Sanity check. */
1060         if (range->end <= range->start)
1061                 return -EINVAL;
1062         if (!daddrs)
1063                 return -EINVAL;
1064         if (!range->pfns)
1065                 return -EINVAL;
1066
1067         npages = (range->end - range->start) >> PAGE_SHIFT;
1068         for (i = 0; i < npages; ++i) {
1069                 enum dma_data_direction dir = DMA_TO_DEVICE;
1070                 struct page *page;
1071
1072                 page = hmm_device_entry_to_page(range, range->pfns[i]);
1073                 if (page == NULL)
1074                         continue;
1075
1076                 /* If it is read and write than map bi-directional. */
1077                 if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) {
1078                         dir = DMA_BIDIRECTIONAL;
1079
1080                         /*
1081                          * See comments in function description on why it is
1082                          * safe here to call set_page_dirty()
1083                          */
1084                         if (dirty)
1085                                 set_page_dirty(page);
1086                 }
1087
1088                 /* Unmap and clear pfns/dma address */
1089                 dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir);
1090                 range->pfns[i] = range->values[HMM_PFN_NONE];
1091                 /* FIXME see comments in hmm_vma_dma_map() */
1092                 daddrs[i] = 0;
1093                 cpages++;
1094         }
1095
1096         return cpages;
1097 }
1098 EXPORT_SYMBOL(hmm_range_dma_unmap);