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 defined(CONFIG_DEVICE_PRIVATE) || defined(CONFIG_DEVICE_PUBLIC)
  39 /*
  40  * Device private memory see HMM (Documentation/vm/hmm.txt) or hmm.h
  41  */
  42 DEFINE_STATIC_KEY_FALSE(device_private_key);
  43 EXPORT_SYMBOL(device_private_key);
  44 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */
  45
  46
  47 #if IS_ENABLED(CONFIG_HMM_MIRROR)
  48 static const struct mmu_notifier_ops hmm_mmu_notifier_ops;
  49
  50 /*
  51  * struct hmm - HMM per mm struct
  52  *
  53  * @mm: mm struct this HMM struct is bound to
  54  * @lock: lock protecting ranges list
  55  * @sequence: we track updates to the CPU page table with a sequence number
  56  * @ranges: list of range being snapshotted
  57  * @mirrors: list of mirrors for this mm
  58  * @mmu_notifier: mmu notifier to track updates to CPU page table
  59  * @mirrors_sem: read/write semaphore protecting the mirrors list
  60  */
  61 struct hmm {
  62         struct mm_struct        *mm;
  63         spinlock_t              lock;
  64         atomic_t                sequence;
  65         struct list_head        ranges;
  66         struct list_head        mirrors;
  67         struct mmu_notifier     mmu_notifier;
  68         struct rw_semaphore     mirrors_sem;
  69 };
  70
  71 /*
  72  * hmm_register - register HMM against an mm (HMM internal)
  73  *
  74  * @mm: mm struct to attach to
  75  *
  76  * This is not intended to be used directly by device drivers. It allocates an
  77  * HMM struct if mm does not have one, and initializes it.
  78  */
  79 static struct hmm *hmm_register(struct mm_struct *mm)
  80 {
  81         struct hmm *hmm = READ_ONCE(mm->hmm);
  82         bool cleanup = false;
  83
  84         /*
  85          * The hmm struct can only be freed once the mm_struct goes away,
  86          * hence we should always have pre-allocated an new hmm struct
  87          * above.
  88          */
  89         if (hmm)
  90                 return hmm;
  91
  92         hmm = kmalloc(sizeof(*hmm), GFP_KERNEL);
  93         if (!hmm)
  94                 return NULL;
  95         INIT_LIST_HEAD(&hmm->mirrors);
  96         init_rwsem(&hmm->mirrors_sem);
  97         atomic_set(&hmm->sequence, 0);
  98         hmm->mmu_notifier.ops = NULL;
  99         INIT_LIST_HEAD(&hmm->ranges);
 100         spin_lock_init(&hmm->lock);
 101         hmm->mm = mm;
 102
 103         /*
 104          * We should only get here if hold the mmap_sem in write mode ie on
 105          * registration of first mirror through hmm_mirror_register()
 106          */
 107         hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops;
 108         if (__mmu_notifier_register(&hmm->mmu_notifier, mm)) {
 109                 kfree(hmm);
 110                 return NULL;
 111         }
 112
 113         spin_lock(&mm->page_table_lock);
 114         if (!mm->hmm)
 115                 mm->hmm = hmm;
 116         else
 117                 cleanup = true;
 118         spin_unlock(&mm->page_table_lock);
 119
 120         if (cleanup) {
 121                 mmu_notifier_unregister(&hmm->mmu_notifier, mm);
 122                 kfree(hmm);
 123         }
 124
 125         return mm->hmm;
 126 }
 127
 128 void hmm_mm_destroy(struct mm_struct *mm)
 129 {
 130         kfree(mm->hmm);
 131 }
 132
 133 static void hmm_invalidate_range(struct hmm *hmm,
 134                                  enum hmm_update_type action,
 135                                  unsigned long start,
 136                                  unsigned long end)
 137 {
 138         struct hmm_mirror *mirror;
 139         struct hmm_range *range;
 140
 141         spin_lock(&hmm->lock);
 142         list_for_each_entry(range, &hmm->ranges, list) {
 143                 unsigned long addr, idx, npages;
 144
 145                 if (end < range->start || start >= range->end)
 146                         continue;
 147
 148                 range->valid = false;
 149                 addr = max(start, range->start);
 150                 idx = (addr - range->start) >> PAGE_SHIFT;
 151                 npages = (min(range->end, end) - addr) >> PAGE_SHIFT;
 152                 memset(&range->pfns[idx], 0, sizeof(*range->pfns) * npages);
 153         }
 154         spin_unlock(&hmm->lock);
 155
 156         down_read(&hmm->mirrors_sem);
 157         list_for_each_entry(mirror, &hmm->mirrors, list)
 158                 mirror->ops->sync_cpu_device_pagetables(mirror, action,
 159                                                         start, end);
 160         up_read(&hmm->mirrors_sem);
 161 }
 162
 163 static void hmm_release(struct mmu_notifier *mn, struct mm_struct *mm)
 164 {
 165         struct hmm_mirror *mirror;
 166         struct hmm *hmm = mm->hmm;
 167
 168         down_write(&hmm->mirrors_sem);
 169         mirror = list_first_entry_or_null(&hmm->mirrors, struct hmm_mirror,
 170                                           list);
 171         while (mirror) {
 172                 list_del_init(&mirror->list);
 173                 if (mirror->ops->release) {
 174                         /*
 175                          * Drop mirrors_sem so callback can wait on any pending
 176                          * work that might itself trigger mmu_notifier callback
 177                          * and thus would deadlock with us.
 178                          */
 179                         up_write(&hmm->mirrors_sem);
 180                         mirror->ops->release(mirror);
 181                         down_write(&hmm->mirrors_sem);
 182                 }
 183                 mirror = list_first_entry_or_null(&hmm->mirrors,
 184                                                   struct hmm_mirror, list);
 185         }
 186         up_write(&hmm->mirrors_sem);
 187 }
 188
 189 static void hmm_invalidate_range_start(struct mmu_notifier *mn,
 190                                        struct mm_struct *mm,
 191                                        unsigned long start,
 192                                        unsigned long end)
 193 {
 194         struct hmm *hmm = mm->hmm;
 195
 196         VM_BUG_ON(!hmm);
 197
 198         atomic_inc(&hmm->sequence);
 199 }
 200
 201 static void hmm_invalidate_range_end(struct mmu_notifier *mn,
 202                                      struct mm_struct *mm,
 203                                      unsigned long start,
 204                                      unsigned long end)
 205 {
 206         struct hmm *hmm = mm->hmm;
 207
 208         VM_BUG_ON(!hmm);
 209
 210         hmm_invalidate_range(mm->hmm, HMM_UPDATE_INVALIDATE, start, end);
 211 }
 212
 213 static const struct mmu_notifier_ops hmm_mmu_notifier_ops = {
 214         .release                = hmm_release,
 215         .invalidate_range_start = hmm_invalidate_range_start,
 216         .invalidate_range_end   = hmm_invalidate_range_end,
 217 };
 218
 219 /*
 220  * hmm_mirror_register() - register a mirror against an mm
 221  *
 222  * @mirror: new mirror struct to register
 223  * @mm: mm to register against
 224  *
 225  * To start mirroring a process address space, the device driver must register
 226  * an HMM mirror struct.
 227  *
 228  * THE mm->mmap_sem MUST BE HELD IN WRITE MODE !
 229  */
 230 int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm)
 231 {
 232         /* Sanity check */
 233         if (!mm || !mirror || !mirror->ops)
 234                 return -EINVAL;
 235
 236 again:
 237         mirror->hmm = hmm_register(mm);
 238         if (!mirror->hmm)
 239                 return -ENOMEM;
 240
 241         down_write(&mirror->hmm->mirrors_sem);
 242         if (mirror->hmm->mm == NULL) {
 243                 /*
 244                  * A racing hmm_mirror_unregister() is about to destroy the hmm
 245                  * struct. Try again to allocate a new one.
 246                  */
 247                 up_write(&mirror->hmm->mirrors_sem);
 248                 mirror->hmm = NULL;
 249                 goto again;
 250         } else {
 251                 list_add(&mirror->list, &mirror->hmm->mirrors);
 252                 up_write(&mirror->hmm->mirrors_sem);
 253         }
 254
 255         return 0;
 256 }
 257 EXPORT_SYMBOL(hmm_mirror_register);
 258
 259 /*
 260  * hmm_mirror_unregister() - unregister a mirror
 261  *
 262  * @mirror: new mirror struct to register
 263  *
 264  * Stop mirroring a process address space, and cleanup.
 265  */
 266 void hmm_mirror_unregister(struct hmm_mirror *mirror)
 267 {
 268         bool should_unregister = false;
 269         struct mm_struct *mm;
 270         struct hmm *hmm;
 271
 272         if (mirror->hmm == NULL)
 273                 return;
 274
 275         hmm = mirror->hmm;
 276         down_write(&hmm->mirrors_sem);
 277         list_del_init(&mirror->list);
 278         should_unregister = list_empty(&hmm->mirrors);
 279         mirror->hmm = NULL;
 280         mm = hmm->mm;
 281         hmm->mm = NULL;
 282         up_write(&hmm->mirrors_sem);
 283
 284         if (!should_unregister || mm == NULL)
 285                 return;
 286
 287         spin_lock(&mm->page_table_lock);
 288         if (mm->hmm == hmm)
 289                 mm->hmm = NULL;
 290         spin_unlock(&mm->page_table_lock);
 291
 292         mmu_notifier_unregister_no_release(&hmm->mmu_notifier, mm);
 293         kfree(hmm);
 294 }
 295 EXPORT_SYMBOL(hmm_mirror_unregister);
 296
 297 struct hmm_vma_walk {
 298         struct hmm_range        *range;
 299         unsigned long           last;
 300         bool                    fault;
 301         bool                    block;
 302 };
 303
 304 static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr,
 305                             bool write_fault, uint64_t *pfn)
 306 {
 307         unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE;
 308         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 309         struct hmm_range *range = hmm_vma_walk->range;
 310         struct vm_area_struct *vma = walk->vma;
 311         int r;
 312
 313         flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY;
 314         flags |= write_fault ? FAULT_FLAG_WRITE : 0;
 315         r = handle_mm_fault(vma, addr, flags);
 316         if (r & VM_FAULT_RETRY)
 317                 return -EBUSY;
 318         if (r & VM_FAULT_ERROR) {
 319                 *pfn = range->values[HMM_PFN_ERROR];
 320                 return -EFAULT;
 321         }
 322
 323         return -EAGAIN;
 324 }
 325
 326 static int hmm_pfns_bad(unsigned long addr,
 327                         unsigned long end,
 328                         struct mm_walk *walk)
 329 {
 330         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 331         struct hmm_range *range = hmm_vma_walk->range;
 332         uint64_t *pfns = range->pfns;
 333         unsigned long i;
 334
 335         i = (addr - range->start) >> PAGE_SHIFT;
 336         for (; addr < end; addr += PAGE_SIZE, i++)
 337                 pfns[i] = range->values[HMM_PFN_ERROR];
 338
 339         return 0;
 340 }
 341
 342 /*
 343  * hmm_vma_walk_hole() - handle a range lacking valid pmd or pte(s)
 344  * @start: range virtual start address (inclusive)
 345  * @end: range virtual end address (exclusive)
 346  * @fault: should we fault or not ?
 347  * @write_fault: write fault ?
 348  * @walk: mm_walk structure
 349  * Returns: 0 on success, -EAGAIN after page fault, or page fault error
 350  *
 351  * This function will be called whenever pmd_none() or pte_none() returns true,
 352  * or whenever there is no page directory covering the virtual address range.
 353  */
 354 static int hmm_vma_walk_hole_(unsigned long addr, unsigned long end,
 355                               bool fault, bool write_fault,
 356                               struct mm_walk *walk)
 357 {
 358         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 359         struct hmm_range *range = hmm_vma_walk->range;
 360         uint64_t *pfns = range->pfns;
 361         unsigned long i;
 362
 363         hmm_vma_walk->last = addr;
 364         i = (addr - range->start) >> PAGE_SHIFT;
 365         for (; addr < end; addr += PAGE_SIZE, i++) {
 366                 pfns[i] = range->values[HMM_PFN_NONE];
 367                 if (fault || write_fault) {
 368                         int ret;
 369
 370                         ret = hmm_vma_do_fault(walk, addr, write_fault,
 371                                                &pfns[i]);
 372                         if (ret != -EAGAIN)
 373                                 return ret;
 374                 }
 375         }
 376
 377         return (fault || write_fault) ? -EAGAIN : 0;
 378 }
 379
 380 static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 381                                       uint64_t pfns, uint64_t cpu_flags,
 382                                       bool *fault, bool *write_fault)
 383 {
 384         struct hmm_range *range = hmm_vma_walk->range;
 385
 386         *fault = *write_fault = false;
 387         if (!hmm_vma_walk->fault)
 388                 return;
 389
 390         /* We aren't ask to do anything ... */
 391         if (!(pfns & range->flags[HMM_PFN_VALID]))
 392                 return;
 393         /* If this is device memory than only fault if explicitly requested */
 394         if ((cpu_flags & range->flags[HMM_PFN_DEVICE_PRIVATE])) {
 395                 /* Do we fault on device memory ? */
 396                 if (pfns & range->flags[HMM_PFN_DEVICE_PRIVATE]) {
 397                         *write_fault = pfns & range->flags[HMM_PFN_WRITE];
 398                         *fault = true;
 399                 }
 400                 return;
 401         }
 402
 403         /* If CPU page table is not valid then we need to fault */
 404         *fault = !(cpu_flags & range->flags[HMM_PFN_VALID]);
 405         /* Need to write fault ? */
 406         if ((pfns & range->flags[HMM_PFN_WRITE]) &&
 407             !(cpu_flags & range->flags[HMM_PFN_WRITE])) {
 408                 *write_fault = true;
 409                 *fault = true;
 410         }
 411 }
 412
 413 static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 414                                  const uint64_t *pfns, unsigned long npages,
 415                                  uint64_t cpu_flags, bool *fault,
 416                                  bool *write_fault)
 417 {
 418         unsigned long i;
 419
 420         if (!hmm_vma_walk->fault) {
 421                 *fault = *write_fault = false;
 422                 return;
 423         }
 424
 425         for (i = 0; i < npages; ++i) {
 426                 hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags,
 427                                    fault, write_fault);
 428                 if ((*fault) || (*write_fault))
 429                         return;
 430         }
 431 }
 432
 433 static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
 434                              struct mm_walk *walk)
 435 {
 436         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 437         struct hmm_range *range = hmm_vma_walk->range;
 438         bool fault, write_fault;
 439         unsigned long i, npages;
 440         uint64_t *pfns;
 441
 442         i = (addr - range->start) >> PAGE_SHIFT;
 443         npages = (end - addr) >> PAGE_SHIFT;
 444         pfns = &range->pfns[i];
 445         hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 446                              0, &fault, &write_fault);
 447         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 448 }
 449
 450 static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd)
 451 {
 452         if (pmd_protnone(pmd))
 453                 return 0;
 454         return pmd_write(pmd) ? range->flags[HMM_PFN_VALID] |
 455                                 range->flags[HMM_PFN_WRITE] :
 456                                 range->flags[HMM_PFN_VALID];
 457 }
 458
 459 static int hmm_vma_handle_pmd(struct mm_walk *walk,
 460                               unsigned long addr,
 461                               unsigned long end,
 462                               uint64_t *pfns,
 463                               pmd_t pmd)
 464 {
 465         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 466         struct hmm_range *range = hmm_vma_walk->range;
 467         unsigned long pfn, npages, i;
 468         bool fault, write_fault;
 469         uint64_t cpu_flags;
 470
 471         npages = (end - addr) >> PAGE_SHIFT;
 472         cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
 473         hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags,
 474                              &fault, &write_fault);
 475
 476         if (pmd_protnone(pmd) || fault || write_fault)
 477                 return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 478
 479         pfn = pmd_pfn(pmd) + pte_index(addr);
 480         for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++)
 481                 pfns[i] = hmm_pfn_from_pfn(range, pfn) | cpu_flags;
 482         hmm_vma_walk->last = end;
 483         return 0;
 484 }
 485
 486 static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte)
 487 {
 488         if (pte_none(pte) || !pte_present(pte))
 489                 return 0;
 490         return pte_write(pte) ? range->flags[HMM_PFN_VALID] |
 491                                 range->flags[HMM_PFN_WRITE] :
 492                                 range->flags[HMM_PFN_VALID];
 493 }
 494
 495 static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
 496                               unsigned long end, pmd_t *pmdp, pte_t *ptep,
 497                               uint64_t *pfn)
 498 {
 499         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 500         struct hmm_range *range = hmm_vma_walk->range;
 501         struct vm_area_struct *vma = walk->vma;
 502         bool fault, write_fault;
 503         uint64_t cpu_flags;
 504         pte_t pte = *ptep;
 505         uint64_t orig_pfn = *pfn;
 506
 507         *pfn = range->values[HMM_PFN_NONE];
 508         cpu_flags = pte_to_hmm_pfn_flags(range, pte);
 509         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 510                            &fault, &write_fault);
 511
 512         if (pte_none(pte)) {
 513                 if (fault || write_fault)
 514                         goto fault;
 515                 return 0;
 516         }
 517
 518         if (!pte_present(pte)) {
 519                 swp_entry_t entry = pte_to_swp_entry(pte);
 520
 521                 if (!non_swap_entry(entry)) {
 522                         if (fault || write_fault)
 523                                 goto fault;
 524                         return 0;
 525                 }
 526
 527                 /*
 528                  * This is a special swap entry, ignore migration, use
 529                  * device and report anything else as error.
 530                  */
 531                 if (is_device_private_entry(entry)) {
 532                         cpu_flags = range->flags[HMM_PFN_VALID] |
 533                                 range->flags[HMM_PFN_DEVICE_PRIVATE];
 534                         cpu_flags |= is_write_device_private_entry(entry) ?
 535                                 range->flags[HMM_PFN_WRITE] : 0;
 536                         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 537                                            &fault, &write_fault);
 538                         if (fault || write_fault)
 539                                 goto fault;
 540                         *pfn = hmm_pfn_from_pfn(range, swp_offset(entry));
 541                         *pfn |= cpu_flags;
 542                         return 0;
 543                 }
 544
 545                 if (is_migration_entry(entry)) {
 546                         if (fault || write_fault) {
 547                                 pte_unmap(ptep);
 548                                 hmm_vma_walk->last = addr;
 549                                 migration_entry_wait(vma->vm_mm,
 550                                                      pmdp, addr);
 551                                 return -EAGAIN;
 552                         }
 553                         return 0;
 554                 }
 555
 556                 /* Report error for everything else */
 557                 *pfn = range->values[HMM_PFN_ERROR];
 558                 return -EFAULT;
 559         }
 560
 561         if (fault || write_fault)
 562                 goto fault;
 563
 564         *pfn = hmm_pfn_from_pfn(range, pte_pfn(pte)) | cpu_flags;
 565         return 0;
 566
 567 fault:
 568         pte_unmap(ptep);
 569         /* Fault any virtual address we were asked to fault */
 570         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 571 }
 572
 573 static int hmm_vma_walk_pmd(pmd_t *pmdp,
 574                             unsigned long start,
 575                             unsigned long end,
 576                             struct mm_walk *walk)
 577 {
 578         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 579         struct hmm_range *range = hmm_vma_walk->range;
 580         uint64_t *pfns = range->pfns;
 581         unsigned long addr = start, i;
 582         pte_t *ptep;
 583
 584         i = (addr - range->start) >> PAGE_SHIFT;
 585
 586 again:
 587         if (pmd_none(*pmdp))
 588                 return hmm_vma_walk_hole(start, end, walk);
 589
 590         if (pmd_huge(*pmdp) && (range->vma->vm_flags & VM_HUGETLB))
 591                 return hmm_pfns_bad(start, end, walk);
 592
 593         if (pmd_devmap(*pmdp) || pmd_trans_huge(*pmdp)) {
 594                 pmd_t pmd;
 595
 596                 /*
 597                  * No need to take pmd_lock here, even if some other threads
 598                  * is splitting the huge pmd we will get that event through
 599                  * mmu_notifier callback.
 600                  *
 601                  * So just read pmd value and check again its a transparent
 602                  * huge or device mapping one and compute corresponding pfn
 603                  * values.
 604                  */
 605                 pmd = pmd_read_atomic(pmdp);
 606                 barrier();
 607                 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
 608                         goto again;
 609
 610                 return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd);
 611         }
 612
 613         if (pmd_bad(*pmdp))
 614                 return hmm_pfns_bad(start, end, walk);
 615
 616         ptep = pte_offset_map(pmdp, addr);
 617         for (; addr < end; addr += PAGE_SIZE, ptep++, i++) {
 618                 int r;
 619
 620                 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]);
 621                 if (r) {
 622                         /* hmm_vma_handle_pte() did unmap pte directory */
 623                         hmm_vma_walk->last = addr;
 624                         return r;
 625                 }
 626         }
 627         pte_unmap(ptep - 1);
 628
 629         hmm_vma_walk->last = addr;
 630         return 0;
 631 }
 632
 633 static void hmm_pfns_clear(struct hmm_range *range,
 634                            uint64_t *pfns,
 635                            unsigned long addr,
 636                            unsigned long end)
 637 {
 638         for (; addr < end; addr += PAGE_SIZE, pfns++)
 639                 *pfns = range->values[HMM_PFN_NONE];
 640 }
 641
 642 static void hmm_pfns_special(struct hmm_range *range)
 643 {
 644         unsigned long addr = range->start, i = 0;
 645
 646         for (; addr < range->end; addr += PAGE_SIZE, i++)
 647                 range->pfns[i] = range->values[HMM_PFN_SPECIAL];
 648 }
 649
 650 /*
 651  * hmm_vma_get_pfns() - snapshot CPU page table for a range of virtual addresses
 652  * @range: range being snapshotted
 653  * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid
 654  *          vma permission, 0 success
 655  *
 656  * This snapshots the CPU page table for a range of virtual addresses. Snapshot
 657  * validity is tracked by range struct. See hmm_vma_range_done() for further
 658  * information.
 659  *
 660  * The range struct is initialized here. It tracks the CPU page table, but only
 661  * if the function returns success (0), in which case the caller must then call
 662  * hmm_vma_range_done() to stop CPU page table update tracking on this range.
 663  *
 664  * NOT CALLING hmm_vma_range_done() IF FUNCTION RETURNS 0 WILL LEAD TO SERIOUS
 665  * MEMORY CORRUPTION ! YOU HAVE BEEN WARNED !
 666  */
 667 int hmm_vma_get_pfns(struct hmm_range *range)
 668 {
 669         struct vm_area_struct *vma = range->vma;
 670         struct hmm_vma_walk hmm_vma_walk;
 671         struct mm_walk mm_walk;
 672         struct hmm *hmm;
 673
 674         /* Sanity check, this really should not happen ! */
 675         if (range->start < vma->vm_start || range->start >= vma->vm_end)
 676                 return -EINVAL;
 677         if (range->end < vma->vm_start || range->end > vma->vm_end)
 678                 return -EINVAL;
 679
 680         hmm = hmm_register(vma->vm_mm);
 681         if (!hmm)
 682                 return -ENOMEM;
 683         /* Caller must have registered a mirror, via hmm_mirror_register() ! */
 684         if (!hmm->mmu_notifier.ops)
 685                 return -EINVAL;
 686
 687         /* FIXME support hugetlb fs */
 688         if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) {
 689                 hmm_pfns_special(range);
 690                 return -EINVAL;
 691         }
 692
 693         if (!(vma->vm_flags & VM_READ)) {
 694                 /*
 695                  * If vma do not allow read access, then assume that it does
 696                  * not allow write access, either. Architecture that allow
 697                  * write without read access are not supported by HMM, because
 698                  * operations such has atomic access would not work.
 699                  */
 700                 hmm_pfns_clear(range, range->pfns, range->start, range->end);
 701                 return -EPERM;
 702         }
 703
 704         /* Initialize range to track CPU page table update */
 705         spin_lock(&hmm->lock);
 706         range->valid = true;
 707         list_add_rcu(&range->list, &hmm->ranges);
 708         spin_unlock(&hmm->lock);
 709
 710         hmm_vma_walk.fault = false;
 711         hmm_vma_walk.range = range;
 712         mm_walk.private = &hmm_vma_walk;
 713
 714         mm_walk.vma = vma;
 715         mm_walk.mm = vma->vm_mm;
 716         mm_walk.pte_entry = NULL;
 717         mm_walk.test_walk = NULL;
 718         mm_walk.hugetlb_entry = NULL;
 719         mm_walk.pmd_entry = hmm_vma_walk_pmd;
 720         mm_walk.pte_hole = hmm_vma_walk_hole;
 721
 722         walk_page_range(range->start, range->end, &mm_walk);
 723         return 0;
 724 }
 725 EXPORT_SYMBOL(hmm_vma_get_pfns);
 726
 727 /*
 728  * hmm_vma_range_done() - stop tracking change to CPU page table over a range
 729  * @range: range being tracked
 730  * Returns: false if range data has been invalidated, true otherwise
 731  *
 732  * Range struct is used to track updates to the CPU page table after a call to
 733  * either hmm_vma_get_pfns() or hmm_vma_fault(). Once the device driver is done
 734  * using the data,  or wants to lock updates to the data it got from those
 735  * functions, it must call the hmm_vma_range_done() function, which will then
 736  * stop tracking CPU page table updates.
 737  *
 738  * Note that device driver must still implement general CPU page table update
 739  * tracking either by using hmm_mirror (see hmm_mirror_register()) or by using
 740  * the mmu_notifier API directly.
 741  *
 742  * CPU page table update tracking done through hmm_range is only temporary and
 743  * to be used while trying to duplicate CPU page table contents for a range of
 744  * virtual addresses.
 745  *
 746  * There are two ways to use this :
 747  * again:
 748  *   hmm_vma_get_pfns(range); or hmm_vma_fault(...);
 749  *   trans = device_build_page_table_update_transaction(pfns);
 750  *   device_page_table_lock();
 751  *   if (!hmm_vma_range_done(range)) {
 752  *     device_page_table_unlock();
 753  *     goto again;
 754  *   }
 755  *   device_commit_transaction(trans);
 756  *   device_page_table_unlock();
 757  *
 758  * Or:
 759  *   hmm_vma_get_pfns(range); or hmm_vma_fault(...);
 760  *   device_page_table_lock();
 761  *   hmm_vma_range_done(range);
 762  *   device_update_page_table(range->pfns);
 763  *   device_page_table_unlock();
 764  */
 765 bool hmm_vma_range_done(struct hmm_range *range)
 766 {
 767         unsigned long npages = (range->end - range->start) >> PAGE_SHIFT;
 768         struct hmm *hmm;
 769
 770         if (range->end <= range->start) {
 771                 BUG();
 772                 return false;
 773         }
 774
 775         hmm = hmm_register(range->vma->vm_mm);
 776         if (!hmm) {
 777                 memset(range->pfns, 0, sizeof(*range->pfns) * npages);
 778                 return false;
 779         }
 780
 781         spin_lock(&hmm->lock);
 782         list_del_rcu(&range->list);
 783         spin_unlock(&hmm->lock);
 784
 785         return range->valid;
 786 }
 787 EXPORT_SYMBOL(hmm_vma_range_done);
 788
 789 /*
 790  * hmm_vma_fault() - try to fault some address in a virtual address range
 791  * @range: range being faulted
 792  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
 793  * Returns: 0 success, error otherwise (-EAGAIN means mmap_sem have been drop)
 794  *
 795  * This is similar to a regular CPU page fault except that it will not trigger
 796  * any memory migration if the memory being faulted is not accessible by CPUs.
 797  *
 798  * On error, for one virtual address in the range, the function will mark the
 799  * corresponding HMM pfn entry with an error flag.
 800  *
 801  * Expected use pattern:
 802  * retry:
 803  *   down_read(&mm->mmap_sem);
 804  *   // Find vma and address device wants to fault, initialize hmm_pfn_t
 805  *   // array accordingly
 806  *   ret = hmm_vma_fault(range, write, block);
 807  *   switch (ret) {
 808  *   case -EAGAIN:
 809  *     hmm_vma_range_done(range);
 810  *     // You might want to rate limit or yield to play nicely, you may
 811  *     // also commit any valid pfn in the array assuming that you are
 812  *     // getting true from hmm_vma_range_monitor_end()
 813  *     goto retry;
 814  *   case 0:
 815  *     break;
 816  *   case -ENOMEM:
 817  *   case -EINVAL:
 818  *   case -EPERM:
 819  *   default:
 820  *     // Handle error !
 821  *     up_read(&mm->mmap_sem)
 822  *     return;
 823  *   }
 824  *   // Take device driver lock that serialize device page table update
 825  *   driver_lock_device_page_table_update();
 826  *   hmm_vma_range_done(range);
 827  *   // Commit pfns we got from hmm_vma_fault()
 828  *   driver_unlock_device_page_table_update();
 829  *   up_read(&mm->mmap_sem)
 830  *
 831  * YOU MUST CALL hmm_vma_range_done() AFTER THIS FUNCTION RETURN SUCCESS (0)
 832  * BEFORE FREEING THE range struct OR YOU WILL HAVE SERIOUS MEMORY CORRUPTION !
 833  *
 834  * YOU HAVE BEEN WARNED !
 835  */
 836 int hmm_vma_fault(struct hmm_range *range, bool block)
 837 {
 838         struct vm_area_struct *vma = range->vma;
 839         unsigned long start = range->start;
 840         struct hmm_vma_walk hmm_vma_walk;
 841         struct mm_walk mm_walk;
 842         struct hmm *hmm;
 843         int ret;
 844
 845         /* Sanity check, this really should not happen ! */
 846         if (range->start < vma->vm_start || range->start >= vma->vm_end)
 847                 return -EINVAL;
 848         if (range->end < vma->vm_start || range->end > vma->vm_end)
 849                 return -EINVAL;
 850
 851         hmm = hmm_register(vma->vm_mm);
 852         if (!hmm) {
 853                 hmm_pfns_clear(range, range->pfns, range->start, range->end);
 854                 return -ENOMEM;
 855         }
 856         /* Caller must have registered a mirror using hmm_mirror_register() */
 857         if (!hmm->mmu_notifier.ops)
 858                 return -EINVAL;
 859
 860         /* FIXME support hugetlb fs */
 861         if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) {
 862                 hmm_pfns_special(range);
 863                 return -EINVAL;
 864         }
 865
 866         if (!(vma->vm_flags & VM_READ)) {
 867                 /*
 868                  * If vma do not allow read access, then assume that it does
 869                  * not allow write access, either. Architecture that allow
 870                  * write without read access are not supported by HMM, because
 871                  * operations such has atomic access would not work.
 872                  */
 873                 hmm_pfns_clear(range, range->pfns, range->start, range->end);
 874                 return -EPERM;
 875         }
 876
 877         /* Initialize range to track CPU page table update */
 878         spin_lock(&hmm->lock);
 879         range->valid = true;
 880         list_add_rcu(&range->list, &hmm->ranges);
 881         spin_unlock(&hmm->lock);
 882
 883         hmm_vma_walk.fault = true;
 884         hmm_vma_walk.block = block;
 885         hmm_vma_walk.range = range;
 886         mm_walk.private = &hmm_vma_walk;
 887         hmm_vma_walk.last = range->start;
 888
 889         mm_walk.vma = vma;
 890         mm_walk.mm = vma->vm_mm;
 891         mm_walk.pte_entry = NULL;
 892         mm_walk.test_walk = NULL;
 893         mm_walk.hugetlb_entry = NULL;
 894         mm_walk.pmd_entry = hmm_vma_walk_pmd;
 895         mm_walk.pte_hole = hmm_vma_walk_hole;
 896
 897         do {
 898                 ret = walk_page_range(start, range->end, &mm_walk);
 899                 start = hmm_vma_walk.last;
 900         } while (ret == -EAGAIN);
 901
 902         if (ret) {
 903                 unsigned long i;
 904
 905                 i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
 906                 hmm_pfns_clear(range, &range->pfns[i], hmm_vma_walk.last,
 907                                range->end);
 908                 hmm_vma_range_done(range);
 909         }
 910         return ret;
 911 }
 912 EXPORT_SYMBOL(hmm_vma_fault);
 913 #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
 914
 915
 916 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) ||  IS_ENABLED(CONFIG_DEVICE_PUBLIC)
 917 struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
 918                                        unsigned long addr)
 919 {
 920         struct page *page;
 921
 922         page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
 923         if (!page)
 924                 return NULL;
 925         lock_page(page);
 926         return page;
 927 }
 928 EXPORT_SYMBOL(hmm_vma_alloc_locked_page);
 929
 930
 931 static void hmm_devmem_ref_release(struct percpu_ref *ref)
 932 {
 933         struct hmm_devmem *devmem;
 934
 935         devmem = container_of(ref, struct hmm_devmem, ref);
 936         complete(&devmem->completion);
 937 }
 938
 939 static void hmm_devmem_ref_exit(void *data)
 940 {
 941         struct percpu_ref *ref = data;
 942         struct hmm_devmem *devmem;
 943
 944         devmem = container_of(ref, struct hmm_devmem, ref);
 945         percpu_ref_exit(ref);
 946         devm_remove_action(devmem->device, &hmm_devmem_ref_exit, data);
 947 }
 948
 949 static void hmm_devmem_ref_kill(void *data)
 950 {
 951         struct percpu_ref *ref = data;
 952         struct hmm_devmem *devmem;
 953
 954         devmem = container_of(ref, struct hmm_devmem, ref);
 955         percpu_ref_kill(ref);
 956         wait_for_completion(&devmem->completion);
 957         devm_remove_action(devmem->device, &hmm_devmem_ref_kill, data);
 958 }
 959
 960 static int hmm_devmem_fault(struct vm_area_struct *vma,
 961                             unsigned long addr,
 962                             const struct page *page,
 963                             unsigned int flags,
 964                             pmd_t *pmdp)
 965 {
 966         struct hmm_devmem *devmem = page->pgmap->data;
 967
 968         return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp);
 969 }
 970
 971 static void hmm_devmem_free(struct page *page, void *data)
 972 {
 973         struct hmm_devmem *devmem = data;
 974
 975         devmem->ops->free(devmem, page);
 976 }
 977
 978 static DEFINE_MUTEX(hmm_devmem_lock);
 979 static RADIX_TREE(hmm_devmem_radix, GFP_KERNEL);
 980
 981 static void hmm_devmem_radix_release(struct resource *resource)
 982 {
 983         resource_size_t key, align_start, align_size;
 984
 985         align_start = resource->start & ~(PA_SECTION_SIZE - 1);
 986         align_size = ALIGN(resource_size(resource), PA_SECTION_SIZE);
 987
 988         mutex_lock(&hmm_devmem_lock);
 989         for (key = resource->start;
 990              key <= resource->end;
 991              key += PA_SECTION_SIZE)
 992                 radix_tree_delete(&hmm_devmem_radix, key >> PA_SECTION_SHIFT);
 993         mutex_unlock(&hmm_devmem_lock);
 994 }
 995
 996 static void hmm_devmem_release(struct device *dev, void *data)
 997 {
 998         struct hmm_devmem *devmem = data;
 999         struct resource *resource = devmem->resource;
1000         unsigned long start_pfn, npages;
1001         struct zone *zone;
1002         struct page *page;
1003
1004         if (percpu_ref_tryget_live(&devmem->ref)) {
1005                 dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
1006                 percpu_ref_put(&devmem->ref);
1007         }
1008
1009         /* pages are dead and unused, undo the arch mapping */
1010         start_pfn = (resource->start & ~(PA_SECTION_SIZE - 1)) >> PAGE_SHIFT;
1011         npages = ALIGN(resource_size(resource), PA_SECTION_SIZE) >> PAGE_SHIFT;
1012
1013         page = pfn_to_page(start_pfn);
1014         zone = page_zone(page);
1015
1016         mem_hotplug_begin();
1017         if (resource->desc == IORES_DESC_DEVICE_PRIVATE_MEMORY)
1018                 __remove_pages(zone, start_pfn, npages, NULL);
1019         else
1020                 arch_remove_memory(start_pfn << PAGE_SHIFT,
1021                                    npages << PAGE_SHIFT, NULL);
1022         mem_hotplug_done();
1023
1024         hmm_devmem_radix_release(resource);
1025 }
1026
1027 static int hmm_devmem_pages_create(struct hmm_devmem *devmem)
1028 {
1029         resource_size_t key, align_start, align_size, align_end;
1030         struct device *device = devmem->device;
1031         int ret, nid, is_ram;
1032         unsigned long pfn;
1033
1034         align_start = devmem->resource->start & ~(PA_SECTION_SIZE - 1);
1035         align_size = ALIGN(devmem->resource->start +
1036                            resource_size(devmem->resource),
1037                            PA_SECTION_SIZE) - align_start;
1038
1039         is_ram = region_intersects(align_start, align_size,
1040                                    IORESOURCE_SYSTEM_RAM,
1041                                    IORES_DESC_NONE);
1042         if (is_ram == REGION_MIXED) {
1043                 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
1044                                 __func__, devmem->resource);
1045                 return -ENXIO;
1046         }
1047         if (is_ram == REGION_INTERSECTS)
1048                 return -ENXIO;
1049
1050         if (devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY)
1051                 devmem->pagemap.type = MEMORY_DEVICE_PUBLIC;
1052         else
1053                 devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
1054
1055         devmem->pagemap.res = *devmem->resource;
1056         devmem->pagemap.page_fault = hmm_devmem_fault;
1057         devmem->pagemap.page_free = hmm_devmem_free;
1058         devmem->pagemap.dev = devmem->device;
1059         devmem->pagemap.ref = &devmem->ref;
1060         devmem->pagemap.data = devmem;
1061
1062         mutex_lock(&hmm_devmem_lock);
1063         align_end = align_start + align_size - 1;
1064         for (key = align_start; key <= align_end; key += PA_SECTION_SIZE) {
1065                 struct hmm_devmem *dup;
1066
1067                 dup = radix_tree_lookup(&hmm_devmem_radix,
1068                                         key >> PA_SECTION_SHIFT);
1069                 if (dup) {
1070                         dev_err(device, "%s: collides with mapping for %s\n",
1071                                 __func__, dev_name(dup->device));
1072                         mutex_unlock(&hmm_devmem_lock);
1073                         ret = -EBUSY;
1074                         goto error;
1075                 }
1076                 ret = radix_tree_insert(&hmm_devmem_radix,
1077                                         key >> PA_SECTION_SHIFT,
1078                                         devmem);
1079                 if (ret) {
1080                         dev_err(device, "%s: failed: %d\n", __func__, ret);
1081                         mutex_unlock(&hmm_devmem_lock);
1082                         goto error_radix;
1083                 }
1084         }
1085         mutex_unlock(&hmm_devmem_lock);
1086
1087         nid = dev_to_node(device);
1088         if (nid < 0)
1089                 nid = numa_mem_id();
1090
1091         mem_hotplug_begin();
1092         /*
1093          * For device private memory we call add_pages() as we only need to
1094          * allocate and initialize struct page for the device memory. More-
1095          * over the device memory is un-accessible thus we do not want to
1096          * create a linear mapping for the memory like arch_add_memory()
1097          * would do.
1098          *
1099          * For device public memory, which is accesible by the CPU, we do
1100          * want the linear mapping and thus use arch_add_memory().
1101          */
1102         if (devmem->pagemap.type == MEMORY_DEVICE_PUBLIC)
1103                 ret = arch_add_memory(nid, align_start, align_size, NULL,
1104                                 false);
1105         else
1106                 ret = add_pages(nid, align_start >> PAGE_SHIFT,
1107                                 align_size >> PAGE_SHIFT, NULL, false);
1108         if (ret) {
1109                 mem_hotplug_done();
1110                 goto error_add_memory;
1111         }
1112         move_pfn_range_to_zone(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
1113                                 align_start >> PAGE_SHIFT,
1114                                 align_size >> PAGE_SHIFT, NULL);
1115         mem_hotplug_done();
1116
1117         for (pfn = devmem->pfn_first; pfn < devmem->pfn_last; pfn++) {
1118                 struct page *page = pfn_to_page(pfn);
1119
1120                 page->pgmap = &devmem->pagemap;
1121         }
1122         return 0;
1123
1124 error_add_memory:
1125         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
1126 error_radix:
1127         hmm_devmem_radix_release(devmem->resource);
1128 error:
1129         return ret;
1130 }
1131
1132 static int hmm_devmem_match(struct device *dev, void *data, void *match_data)
1133 {
1134         struct hmm_devmem *devmem = data;
1135
1136         return devmem->resource == match_data;
1137 }
1138
1139 static void hmm_devmem_pages_remove(struct hmm_devmem *devmem)
1140 {
1141         devres_release(devmem->device, &hmm_devmem_release,
1142                        &hmm_devmem_match, devmem->resource);
1143 }
1144
1145 /*
1146  * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory
1147  *
1148  * @ops: memory event device driver callback (see struct hmm_devmem_ops)
1149  * @device: device struct to bind the resource too
1150  * @size: size in bytes of the device memory to add
1151  * Returns: pointer to new hmm_devmem struct ERR_PTR otherwise
1152  *
1153  * This function first finds an empty range of physical address big enough to
1154  * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which
1155  * in turn allocates struct pages. It does not do anything beyond that; all
1156  * events affecting the memory will go through the various callbacks provided
1157  * by hmm_devmem_ops struct.
1158  *
1159  * Device driver should call this function during device initialization and
1160  * is then responsible of memory management. HMM only provides helpers.
1161  */
1162 struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
1163                                   struct device *device,
1164                                   unsigned long size)
1165 {
1166         struct hmm_devmem *devmem;
1167         resource_size_t addr;
1168         int ret;
1169
1170         static_branch_enable(&device_private_key);
1171
1172         devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem),
1173                                    GFP_KERNEL, dev_to_node(device));
1174         if (!devmem)
1175                 return ERR_PTR(-ENOMEM);
1176
1177         init_completion(&devmem->completion);
1178         devmem->pfn_first = -1UL;
1179         devmem->pfn_last = -1UL;
1180         devmem->resource = NULL;
1181         devmem->device = device;
1182         devmem->ops = ops;
1183
1184         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1185                               0, GFP_KERNEL);
1186         if (ret)
1187                 goto error_percpu_ref;
1188
1189         ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref);
1190         if (ret)
1191                 goto error_devm_add_action;
1192
1193         size = ALIGN(size, PA_SECTION_SIZE);
1194         addr = min((unsigned long)iomem_resource.end,
1195                    (1UL << MAX_PHYSMEM_BITS) - 1);
1196         addr = addr - size + 1UL;
1197
1198         /*
1199          * FIXME add a new helper to quickly walk resource tree and find free
1200          * range
1201          *
1202          * FIXME what about ioport_resource resource ?
1203          */
1204         for (; addr > size && addr >= iomem_resource.start; addr -= size) {
1205                 ret = region_intersects(addr, size, 0, IORES_DESC_NONE);
1206                 if (ret != REGION_DISJOINT)
1207                         continue;
1208
1209                 devmem->resource = devm_request_mem_region(device, addr, size,
1210                                                            dev_name(device));
1211                 if (!devmem->resource) {
1212                         ret = -ENOMEM;
1213                         goto error_no_resource;
1214                 }
1215                 break;
1216         }
1217         if (!devmem->resource) {
1218                 ret = -ERANGE;
1219                 goto error_no_resource;
1220         }
1221
1222         devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY;
1223         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1224         devmem->pfn_last = devmem->pfn_first +
1225                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1226
1227         ret = hmm_devmem_pages_create(devmem);
1228         if (ret)
1229                 goto error_pages;
1230
1231         devres_add(device, devmem);
1232
1233         ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref);
1234         if (ret) {
1235                 hmm_devmem_remove(devmem);
1236                 return ERR_PTR(ret);
1237         }
1238
1239         return devmem;
1240
1241 error_pages:
1242         devm_release_mem_region(device, devmem->resource->start,
1243                                 resource_size(devmem->resource));
1244 error_no_resource:
1245 error_devm_add_action:
1246         hmm_devmem_ref_kill(&devmem->ref);
1247         hmm_devmem_ref_exit(&devmem->ref);
1248 error_percpu_ref:
1249         devres_free(devmem);
1250         return ERR_PTR(ret);
1251 }
1252 EXPORT_SYMBOL(hmm_devmem_add);
1253
1254 struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops,
1255                                            struct device *device,
1256                                            struct resource *res)
1257 {
1258         struct hmm_devmem *devmem;
1259         int ret;
1260
1261         if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY)
1262                 return ERR_PTR(-EINVAL);
1263
1264         static_branch_enable(&device_private_key);
1265
1266         devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem),
1267                                    GFP_KERNEL, dev_to_node(device));
1268         if (!devmem)
1269                 return ERR_PTR(-ENOMEM);
1270
1271         init_completion(&devmem->completion);
1272         devmem->pfn_first = -1UL;
1273         devmem->pfn_last = -1UL;
1274         devmem->resource = res;
1275         devmem->device = device;
1276         devmem->ops = ops;
1277
1278         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1279                               0, GFP_KERNEL);
1280         if (ret)
1281                 goto error_percpu_ref;
1282
1283         ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref);
1284         if (ret)
1285                 goto error_devm_add_action;
1286
1287
1288         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1289         devmem->pfn_last = devmem->pfn_first +
1290                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1291
1292         ret = hmm_devmem_pages_create(devmem);
1293         if (ret)
1294                 goto error_devm_add_action;
1295
1296         devres_add(device, devmem);
1297
1298         ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref);
1299         if (ret) {
1300                 hmm_devmem_remove(devmem);
1301                 return ERR_PTR(ret);
1302         }
1303
1304         return devmem;
1305
1306 error_devm_add_action:
1307         hmm_devmem_ref_kill(&devmem->ref);
1308         hmm_devmem_ref_exit(&devmem->ref);
1309 error_percpu_ref:
1310         devres_free(devmem);
1311         return ERR_PTR(ret);
1312 }
1313 EXPORT_SYMBOL(hmm_devmem_add_resource);
1314
1315 /*
1316  * hmm_devmem_remove() - remove device memory (kill and free ZONE_DEVICE)
1317  *
1318  * @devmem: hmm_devmem struct use to track and manage the ZONE_DEVICE memory
1319  *
1320  * This will hot-unplug memory that was hotplugged by hmm_devmem_add on behalf
1321  * of the device driver. It will free struct page and remove the resource that
1322  * reserved the physical address range for this device memory.
1323  */
1324 void hmm_devmem_remove(struct hmm_devmem *devmem)
1325 {
1326         resource_size_t start, size;
1327         struct device *device;
1328         bool cdm = false;
1329
1330         if (!devmem)
1331                 return;
1332
1333         device = devmem->device;
1334         start = devmem->resource->start;
1335         size = resource_size(devmem->resource);
1336
1337         cdm = devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY;
1338         hmm_devmem_ref_kill(&devmem->ref);
1339         hmm_devmem_ref_exit(&devmem->ref);
1340         hmm_devmem_pages_remove(devmem);
1341
1342         if (!cdm)
1343                 devm_release_mem_region(device, start, size);
1344 }
1345 EXPORT_SYMBOL(hmm_devmem_remove);
1346
1347 /*
1348  * A device driver that wants to handle multiple devices memory through a
1349  * single fake device can use hmm_device to do so. This is purely a helper
1350  * and it is not needed to make use of any HMM functionality.
1351  */
1352 #define HMM_DEVICE_MAX 256
1353
1354 static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX);
1355 static DEFINE_SPINLOCK(hmm_device_lock);
1356 static struct class *hmm_device_class;
1357 static dev_t hmm_device_devt;
1358
1359 static void hmm_device_release(struct device *device)
1360 {
1361         struct hmm_device *hmm_device;
1362
1363         hmm_device = container_of(device, struct hmm_device, device);
1364         spin_lock(&hmm_device_lock);
1365         clear_bit(hmm_device->minor, hmm_device_mask);
1366         spin_unlock(&hmm_device_lock);
1367
1368         kfree(hmm_device);
1369 }
1370
1371 struct hmm_device *hmm_device_new(void *drvdata)
1372 {
1373         struct hmm_device *hmm_device;
1374
1375         hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL);
1376         if (!hmm_device)
1377                 return ERR_PTR(-ENOMEM);
1378
1379         spin_lock(&hmm_device_lock);
1380         hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX);
1381         if (hmm_device->minor >= HMM_DEVICE_MAX) {
1382                 spin_unlock(&hmm_device_lock);
1383                 kfree(hmm_device);
1384                 return ERR_PTR(-EBUSY);
1385         }
1386         set_bit(hmm_device->minor, hmm_device_mask);
1387         spin_unlock(&hmm_device_lock);
1388
1389         dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor);
1390         hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt),
1391                                         hmm_device->minor);
1392         hmm_device->device.release = hmm_device_release;
1393         dev_set_drvdata(&hmm_device->device, drvdata);
1394         hmm_device->device.class = hmm_device_class;
1395         device_initialize(&hmm_device->device);
1396
1397         return hmm_device;
1398 }
1399 EXPORT_SYMBOL(hmm_device_new);
1400
1401 void hmm_device_put(struct hmm_device *hmm_device)
1402 {
1403         put_device(&hmm_device->device);
1404 }
1405 EXPORT_SYMBOL(hmm_device_put);
1406
1407 static int __init hmm_init(void)
1408 {
1409         int ret;
1410
1411         ret = alloc_chrdev_region(&hmm_device_devt, 0,
1412                                   HMM_DEVICE_MAX,
1413                                   "hmm_device");
1414         if (ret)
1415                 return ret;
1416
1417         hmm_device_class = class_create(THIS_MODULE, "hmm_device");
1418         if (IS_ERR(hmm_device_class)) {
1419                 unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX);
1420                 return PTR_ERR(hmm_device_class);
1421         }
1422         return 0;
1423 }
1424
1425 device_initcall(hmm_init);
1426 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */