mm/hmm.c

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