mm/hmm.c

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