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  * @sequence: we track updates to the CPU page table with a sequence number
  47  * @ranges: list of range being snapshotted
  48  * @mirrors: list of mirrors for this mm
  49  * @mmu_notifier: mmu notifier to track updates to CPU page table
  50  * @mirrors_sem: read/write semaphore protecting the mirrors list
  51  */
  52 struct hmm {
  53         struct mm_struct        *mm;
  54         spinlock_t              lock;
  55         atomic_t                sequence;
  56         struct list_head        ranges;
  57         struct list_head        mirrors;
  58         struct mmu_notifier     mmu_notifier;
  59         struct rw_semaphore     mirrors_sem;
  60 };
  61
  62 /*
  63  * hmm_register - register HMM against an mm (HMM internal)
  64  *
  65  * @mm: mm struct to attach to
  66  *
  67  * This is not intended to be used directly by device drivers. It allocates an
  68  * HMM struct if mm does not have one, and initializes it.
  69  */
  70 static struct hmm *hmm_register(struct mm_struct *mm)
  71 {
  72         struct hmm *hmm = READ_ONCE(mm->hmm);
  73         bool cleanup = false;
  74
  75         /*
  76          * The hmm struct can only be freed once the mm_struct goes away,
  77          * hence we should always have pre-allocated an new hmm struct
  78          * above.
  79          */
  80         if (hmm)
  81                 return hmm;
  82
  83         hmm = kmalloc(sizeof(*hmm), GFP_KERNEL);
  84         if (!hmm)
  85                 return NULL;
  86         INIT_LIST_HEAD(&hmm->mirrors);
  87         init_rwsem(&hmm->mirrors_sem);
  88         atomic_set(&hmm->sequence, 0);
  89         hmm->mmu_notifier.ops = NULL;
  90         INIT_LIST_HEAD(&hmm->ranges);
  91         spin_lock_init(&hmm->lock);
  92         hmm->mm = mm;
  93
  94         spin_lock(&mm->page_table_lock);
  95         if (!mm->hmm)
  96                 mm->hmm = hmm;
  97         else
  98                 cleanup = true;
  99         spin_unlock(&mm->page_table_lock);
 100
 101         if (cleanup)
 102                 goto error;
 103
 104         /*
 105          * We should only get here if hold the mmap_sem in write mode ie on
 106          * registration of first mirror through hmm_mirror_register()
 107          */
 108         hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops;
 109         if (__mmu_notifier_register(&hmm->mmu_notifier, mm))
 110                 goto error_mm;
 111
 112         return mm->hmm;
 113
 114 error_mm:
 115         spin_lock(&mm->page_table_lock);
 116         if (mm->hmm == hmm)
 117                 mm->hmm = NULL;
 118         spin_unlock(&mm->page_table_lock);
 119 error:
 120         kfree(hmm);
 121         return NULL;
 122 }
 123
 124 void hmm_mm_destroy(struct mm_struct *mm)
 125 {
 126         kfree(mm->hmm);
 127 }
 128
 129 static int hmm_invalidate_range(struct hmm *hmm,
 130                                 const struct hmm_update *update)
 131 {
 132         struct hmm_mirror *mirror;
 133         struct hmm_range *range;
 134
 135         spin_lock(&hmm->lock);
 136         list_for_each_entry(range, &hmm->ranges, list) {
 137                 unsigned long addr, idx, npages;
 138
 139                 if (update->end < range->start || update->start >= range->end)
 140                         continue;
 141
 142                 range->valid = false;
 143                 addr = max(update->start, range->start);
 144                 idx = (addr - range->start) >> PAGE_SHIFT;
 145                 npages = (min(range->end, update->end) - addr) >> PAGE_SHIFT;
 146                 memset(&range->pfns[idx], 0, sizeof(*range->pfns) * npages);
 147         }
 148         spin_unlock(&hmm->lock);
 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 *hmm = mm->hmm;
 198
 199         VM_BUG_ON(!hmm);
 200
 201         atomic_inc(&hmm->sequence);
 202
 203         return 0;
 204 }
 205
 206 static void hmm_invalidate_range_end(struct mmu_notifier *mn,
 207                                      struct mm_struct *mm,
 208                                      unsigned long start,
 209                                      unsigned long end)
 210 {
 211         struct hmm_update update;
 212         struct hmm *hmm = mm->hmm;
 213
 214         VM_BUG_ON(!hmm);
 215
 216         update.start = start;
 217         update.end = end;
 218         update.event = HMM_UPDATE_INVALIDATE;
 219         update.blockable = true;
 220         hmm_invalidate_range(hmm, &update);
 221 }
 222
 223 static const struct mmu_notifier_ops hmm_mmu_notifier_ops = {
 224         .release                = hmm_release,
 225         .invalidate_range_start = hmm_invalidate_range_start,
 226         .invalidate_range_end   = hmm_invalidate_range_end,
 227 };
 228
 229 /*
 230  * hmm_mirror_register() - register a mirror against an mm
 231  *
 232  * @mirror: new mirror struct to register
 233  * @mm: mm to register against
 234  *
 235  * To start mirroring a process address space, the device driver must register
 236  * an HMM mirror struct.
 237  *
 238  * THE mm->mmap_sem MUST BE HELD IN WRITE MODE !
 239  */
 240 int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm)
 241 {
 242         /* Sanity check */
 243         if (!mm || !mirror || !mirror->ops)
 244                 return -EINVAL;
 245
 246 again:
 247         mirror->hmm = hmm_register(mm);
 248         if (!mirror->hmm)
 249                 return -ENOMEM;
 250
 251         down_write(&mirror->hmm->mirrors_sem);
 252         if (mirror->hmm->mm == NULL) {
 253                 /*
 254                  * A racing hmm_mirror_unregister() is about to destroy the hmm
 255                  * struct. Try again to allocate a new one.
 256                  */
 257                 up_write(&mirror->hmm->mirrors_sem);
 258                 mirror->hmm = NULL;
 259                 goto again;
 260         } else {
 261                 list_add(&mirror->list, &mirror->hmm->mirrors);
 262                 up_write(&mirror->hmm->mirrors_sem);
 263         }
 264
 265         return 0;
 266 }
 267 EXPORT_SYMBOL(hmm_mirror_register);
 268
 269 /*
 270  * hmm_mirror_unregister() - unregister a mirror
 271  *
 272  * @mirror: new mirror struct to register
 273  *
 274  * Stop mirroring a process address space, and cleanup.
 275  */
 276 void hmm_mirror_unregister(struct hmm_mirror *mirror)
 277 {
 278         bool should_unregister = false;
 279         struct mm_struct *mm;
 280         struct hmm *hmm;
 281
 282         if (mirror->hmm == NULL)
 283                 return;
 284
 285         hmm = mirror->hmm;
 286         down_write(&hmm->mirrors_sem);
 287         list_del_init(&mirror->list);
 288         should_unregister = list_empty(&hmm->mirrors);
 289         mirror->hmm = NULL;
 290         mm = hmm->mm;
 291         hmm->mm = NULL;
 292         up_write(&hmm->mirrors_sem);
 293
 294         if (!should_unregister || mm == NULL)
 295                 return;
 296
 297         mmu_notifier_unregister_no_release(&hmm->mmu_notifier, mm);
 298
 299         spin_lock(&mm->page_table_lock);
 300         if (mm->hmm == hmm)
 301                 mm->hmm = NULL;
 302         spin_unlock(&mm->page_table_lock);
 303
 304         kfree(hmm);
 305 }
 306 EXPORT_SYMBOL(hmm_mirror_unregister);
 307
 308 struct hmm_vma_walk {
 309         struct hmm_range        *range;
 310         unsigned long           last;
 311         bool                    fault;
 312         bool                    block;
 313 };
 314
 315 static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr,
 316                             bool write_fault, uint64_t *pfn)
 317 {
 318         unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE;
 319         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 320         struct hmm_range *range = hmm_vma_walk->range;
 321         struct vm_area_struct *vma = walk->vma;
 322         vm_fault_t ret;
 323
 324         flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY;
 325         flags |= write_fault ? FAULT_FLAG_WRITE : 0;
 326         ret = handle_mm_fault(vma, addr, flags);
 327         if (ret & VM_FAULT_RETRY)
 328                 return -EBUSY;
 329         if (ret & VM_FAULT_ERROR) {
 330                 *pfn = range->values[HMM_PFN_ERROR];
 331                 return -EFAULT;
 332         }
 333
 334         return -EAGAIN;
 335 }
 336
 337 static int hmm_pfns_bad(unsigned long addr,
 338                         unsigned long end,
 339                         struct mm_walk *walk)
 340 {
 341         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 342         struct hmm_range *range = hmm_vma_walk->range;
 343         uint64_t *pfns = range->pfns;
 344         unsigned long i;
 345
 346         i = (addr - range->start) >> PAGE_SHIFT;
 347         for (; addr < end; addr += PAGE_SIZE, i++)
 348                 pfns[i] = range->values[HMM_PFN_ERROR];
 349
 350         return 0;
 351 }
 352
 353 /*
 354  * hmm_vma_walk_hole() - handle a range lacking valid pmd or pte(s)
 355  * @start: range virtual start address (inclusive)
 356  * @end: range virtual end address (exclusive)
 357  * @fault: should we fault or not ?
 358  * @write_fault: write fault ?
 359  * @walk: mm_walk structure
 360  * Returns: 0 on success, -EAGAIN after page fault, or page fault error
 361  *
 362  * This function will be called whenever pmd_none() or pte_none() returns true,
 363  * or whenever there is no page directory covering the virtual address range.
 364  */
 365 static int hmm_vma_walk_hole_(unsigned long addr, unsigned long end,
 366                               bool fault, bool write_fault,
 367                               struct mm_walk *walk)
 368 {
 369         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 370         struct hmm_range *range = hmm_vma_walk->range;
 371         uint64_t *pfns = range->pfns;
 372         unsigned long i;
 373
 374         hmm_vma_walk->last = addr;
 375         i = (addr - range->start) >> PAGE_SHIFT;
 376         for (; addr < end; addr += PAGE_SIZE, i++) {
 377                 pfns[i] = range->values[HMM_PFN_NONE];
 378                 if (fault || write_fault) {
 379                         int ret;
 380
 381                         ret = hmm_vma_do_fault(walk, addr, write_fault,
 382                                                &pfns[i]);
 383                         if (ret != -EAGAIN)
 384                                 return ret;
 385                 }
 386         }
 387
 388         return (fault || write_fault) ? -EAGAIN : 0;
 389 }
 390
 391 static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 392                                       uint64_t pfns, uint64_t cpu_flags,
 393                                       bool *fault, bool *write_fault)
 394 {
 395         struct hmm_range *range = hmm_vma_walk->range;
 396
 397         *fault = *write_fault = false;
 398         if (!hmm_vma_walk->fault)
 399                 return;
 400
 401         /* We aren't ask to do anything ... */
 402         if (!(pfns & range->flags[HMM_PFN_VALID]))
 403                 return;
 404         /* If this is device memory than only fault if explicitly requested */
 405         if ((cpu_flags & range->flags[HMM_PFN_DEVICE_PRIVATE])) {
 406                 /* Do we fault on device memory ? */
 407                 if (pfns & range->flags[HMM_PFN_DEVICE_PRIVATE]) {
 408                         *write_fault = pfns & range->flags[HMM_PFN_WRITE];
 409                         *fault = true;
 410                 }
 411                 return;
 412         }
 413
 414         /* If CPU page table is not valid then we need to fault */
 415         *fault = !(cpu_flags & range->flags[HMM_PFN_VALID]);
 416         /* Need to write fault ? */
 417         if ((pfns & range->flags[HMM_PFN_WRITE]) &&
 418             !(cpu_flags & range->flags[HMM_PFN_WRITE])) {
 419                 *write_fault = true;
 420                 *fault = true;
 421         }
 422 }
 423
 424 static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 425                                  const uint64_t *pfns, unsigned long npages,
 426                                  uint64_t cpu_flags, bool *fault,
 427                                  bool *write_fault)
 428 {
 429         unsigned long i;
 430
 431         if (!hmm_vma_walk->fault) {
 432                 *fault = *write_fault = false;
 433                 return;
 434         }
 435
 436         for (i = 0; i < npages; ++i) {
 437                 hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags,
 438                                    fault, write_fault);
 439                 if ((*fault) || (*write_fault))
 440                         return;
 441         }
 442 }
 443
 444 static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
 445                              struct mm_walk *walk)
 446 {
 447         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 448         struct hmm_range *range = hmm_vma_walk->range;
 449         bool fault, write_fault;
 450         unsigned long i, npages;
 451         uint64_t *pfns;
 452
 453         i = (addr - range->start) >> PAGE_SHIFT;
 454         npages = (end - addr) >> PAGE_SHIFT;
 455         pfns = &range->pfns[i];
 456         hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 457                              0, &fault, &write_fault);
 458         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 459 }
 460
 461 static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd)
 462 {
 463         if (pmd_protnone(pmd))
 464                 return 0;
 465         return pmd_write(pmd) ? range->flags[HMM_PFN_VALID] |
 466                                 range->flags[HMM_PFN_WRITE] :
 467                                 range->flags[HMM_PFN_VALID];
 468 }
 469
 470 static int hmm_vma_handle_pmd(struct mm_walk *walk,
 471                               unsigned long addr,
 472                               unsigned long end,
 473                               uint64_t *pfns,
 474                               pmd_t pmd)
 475 {
 476         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 477         struct hmm_range *range = hmm_vma_walk->range;
 478         unsigned long pfn, npages, i;
 479         bool fault, write_fault;
 480         uint64_t cpu_flags;
 481
 482         npages = (end - addr) >> PAGE_SHIFT;
 483         cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
 484         hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags,
 485                              &fault, &write_fault);
 486
 487         if (pmd_protnone(pmd) || fault || write_fault)
 488                 return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 489
 490         pfn = pmd_pfn(pmd) + pte_index(addr);
 491         for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++)
 492                 pfns[i] = hmm_pfn_from_pfn(range, pfn) | cpu_flags;
 493         hmm_vma_walk->last = end;
 494         return 0;
 495 }
 496
 497 static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte)
 498 {
 499         if (pte_none(pte) || !pte_present(pte))
 500                 return 0;
 501         return pte_write(pte) ? range->flags[HMM_PFN_VALID] |
 502                                 range->flags[HMM_PFN_WRITE] :
 503                                 range->flags[HMM_PFN_VALID];
 504 }
 505
 506 static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
 507                               unsigned long end, pmd_t *pmdp, pte_t *ptep,
 508                               uint64_t *pfn)
 509 {
 510         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 511         struct hmm_range *range = hmm_vma_walk->range;
 512         struct vm_area_struct *vma = walk->vma;
 513         bool fault, write_fault;
 514         uint64_t cpu_flags;
 515         pte_t pte = *ptep;
 516         uint64_t orig_pfn = *pfn;
 517
 518         *pfn = range->values[HMM_PFN_NONE];
 519         cpu_flags = pte_to_hmm_pfn_flags(range, pte);
 520         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 521                            &fault, &write_fault);
 522
 523         if (pte_none(pte)) {
 524                 if (fault || write_fault)
 525                         goto fault;
 526                 return 0;
 527         }
 528
 529         if (!pte_present(pte)) {
 530                 swp_entry_t entry = pte_to_swp_entry(pte);
 531
 532                 if (!non_swap_entry(entry)) {
 533                         if (fault || write_fault)
 534                                 goto fault;
 535                         return 0;
 536                 }
 537
 538                 /*
 539                  * This is a special swap entry, ignore migration, use
 540                  * device and report anything else as error.
 541                  */
 542                 if (is_device_private_entry(entry)) {
 543                         cpu_flags = range->flags[HMM_PFN_VALID] |
 544                                 range->flags[HMM_PFN_DEVICE_PRIVATE];
 545                         cpu_flags |= is_write_device_private_entry(entry) ?
 546                                 range->flags[HMM_PFN_WRITE] : 0;
 547                         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 548                                            &fault, &write_fault);
 549                         if (fault || write_fault)
 550                                 goto fault;
 551                         *pfn = hmm_pfn_from_pfn(range, swp_offset(entry));
 552                         *pfn |= cpu_flags;
 553                         return 0;
 554                 }
 555
 556                 if (is_migration_entry(entry)) {
 557                         if (fault || write_fault) {
 558                                 pte_unmap(ptep);
 559                                 hmm_vma_walk->last = addr;
 560                                 migration_entry_wait(vma->vm_mm,
 561                                                      pmdp, addr);
 562                                 return -EAGAIN;
 563                         }
 564                         return 0;
 565                 }
 566
 567                 /* Report error for everything else */
 568                 *pfn = range->values[HMM_PFN_ERROR];
 569                 return -EFAULT;
 570         }
 571
 572         if (fault || write_fault)
 573                 goto fault;
 574
 575         *pfn = hmm_pfn_from_pfn(range, pte_pfn(pte)) | cpu_flags;
 576         return 0;
 577
 578 fault:
 579         pte_unmap(ptep);
 580         /* Fault any virtual address we were asked to fault */
 581         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 582 }
 583
 584 static int hmm_vma_walk_pmd(pmd_t *pmdp,
 585                             unsigned long start,
 586                             unsigned long end,
 587                             struct mm_walk *walk)
 588 {
 589         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 590         struct hmm_range *range = hmm_vma_walk->range;
 591         struct vm_area_struct *vma = walk->vma;
 592         uint64_t *pfns = range->pfns;
 593         unsigned long addr = start, i;
 594         pte_t *ptep;
 595         pmd_t pmd;
 596
 597
 598 again:
 599         pmd = READ_ONCE(*pmdp);
 600         if (pmd_none(pmd))
 601                 return hmm_vma_walk_hole(start, end, walk);
 602
 603         if (pmd_huge(pmd) && (range->vma->vm_flags & VM_HUGETLB))
 604                 return hmm_pfns_bad(start, end, walk);
 605
 606         if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
 607                 bool fault, write_fault;
 608                 unsigned long npages;
 609                 uint64_t *pfns;
 610
 611                 i = (addr - range->start) >> PAGE_SHIFT;
 612                 npages = (end - addr) >> PAGE_SHIFT;
 613                 pfns = &range->pfns[i];
 614
 615                 hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 616                                      0, &fault, &write_fault);
 617                 if (fault || write_fault) {
 618                         hmm_vma_walk->last = addr;
 619                         pmd_migration_entry_wait(vma->vm_mm, pmdp);
 620                         return -EAGAIN;
 621                 }
 622                 return 0;
 623         } else if (!pmd_present(pmd))
 624                 return hmm_pfns_bad(start, end, walk);
 625
 626         if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
 627                 /*
 628                  * No need to take pmd_lock here, even if some other threads
 629                  * is splitting the huge pmd we will get that event through
 630                  * mmu_notifier callback.
 631                  *
 632                  * So just read pmd value and check again its a transparent
 633                  * huge or device mapping one and compute corresponding pfn
 634                  * values.
 635                  */
 636                 pmd = pmd_read_atomic(pmdp);
 637                 barrier();
 638                 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
 639                         goto again;
 640
 641                 i = (addr - range->start) >> PAGE_SHIFT;
 642                 return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd);
 643         }
 644
 645         /*
 646          * We have handled all the valid case above ie either none, migration,
 647          * huge or transparent huge. At this point either it is a valid pmd
 648          * entry pointing to pte directory or it is a bad pmd that will not
 649          * recover.
 650          */
 651         if (pmd_bad(pmd))
 652                 return hmm_pfns_bad(start, end, walk);
 653
 654         ptep = pte_offset_map(pmdp, addr);
 655         i = (addr - range->start) >> PAGE_SHIFT;
 656         for (; addr < end; addr += PAGE_SIZE, ptep++, i++) {
 657                 int r;
 658
 659                 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]);
 660                 if (r) {
 661                         /* hmm_vma_handle_pte() did unmap pte directory */
 662                         hmm_vma_walk->last = addr;
 663                         return r;
 664                 }
 665         }
 666         pte_unmap(ptep - 1);
 667
 668         hmm_vma_walk->last = addr;
 669         return 0;
 670 }
 671
 672 static void hmm_pfns_clear(struct hmm_range *range,
 673                            uint64_t *pfns,
 674                            unsigned long addr,
 675                            unsigned long end)
 676 {
 677         for (; addr < end; addr += PAGE_SIZE, pfns++)
 678                 *pfns = range->values[HMM_PFN_NONE];
 679 }
 680
 681 static void hmm_pfns_special(struct hmm_range *range)
 682 {
 683         unsigned long addr = range->start, i = 0;
 684
 685         for (; addr < range->end; addr += PAGE_SIZE, i++)
 686                 range->pfns[i] = range->values[HMM_PFN_SPECIAL];
 687 }
 688
 689 /*
 690  * hmm_vma_get_pfns() - snapshot CPU page table for a range of virtual addresses
 691  * @range: range being snapshotted
 692  * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid
 693  *          vma permission, 0 success
 694  *
 695  * This snapshots the CPU page table for a range of virtual addresses. Snapshot
 696  * validity is tracked by range struct. See hmm_vma_range_done() for further
 697  * information.
 698  *
 699  * The range struct is initialized here. It tracks the CPU page table, but only
 700  * if the function returns success (0), in which case the caller must then call
 701  * hmm_vma_range_done() to stop CPU page table update tracking on this range.
 702  *
 703  * NOT CALLING hmm_vma_range_done() IF FUNCTION RETURNS 0 WILL LEAD TO SERIOUS
 704  * MEMORY CORRUPTION ! YOU HAVE BEEN WARNED !
 705  */
 706 int hmm_vma_get_pfns(struct hmm_range *range)
 707 {
 708         struct vm_area_struct *vma = range->vma;
 709         struct hmm_vma_walk hmm_vma_walk;
 710         struct mm_walk mm_walk;
 711         struct hmm *hmm;
 712
 713         /* Sanity check, this really should not happen ! */
 714         if (range->start < vma->vm_start || range->start >= vma->vm_end)
 715                 return -EINVAL;
 716         if (range->end < vma->vm_start || range->end > vma->vm_end)
 717                 return -EINVAL;
 718
 719         hmm = hmm_register(vma->vm_mm);
 720         if (!hmm)
 721                 return -ENOMEM;
 722         /* Caller must have registered a mirror, via hmm_mirror_register() ! */
 723         if (!hmm->mmu_notifier.ops)
 724                 return -EINVAL;
 725
 726         /* FIXME support hugetlb fs */
 727         if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) ||
 728                         vma_is_dax(vma)) {
 729                 hmm_pfns_special(range);
 730                 return -EINVAL;
 731         }
 732
 733         if (!(vma->vm_flags & VM_READ)) {
 734                 /*
 735                  * If vma do not allow read access, then assume that it does
 736                  * not allow write access, either. Architecture that allow
 737                  * write without read access are not supported by HMM, because
 738                  * operations such has atomic access would not work.
 739                  */
 740                 hmm_pfns_clear(range, range->pfns, range->start, range->end);
 741                 return -EPERM;
 742         }
 743
 744         /* Initialize range to track CPU page table update */
 745         spin_lock(&hmm->lock);
 746         range->valid = true;
 747         list_add_rcu(&range->list, &hmm->ranges);
 748         spin_unlock(&hmm->lock);
 749
 750         hmm_vma_walk.fault = false;
 751         hmm_vma_walk.range = range;
 752         mm_walk.private = &hmm_vma_walk;
 753
 754         mm_walk.vma = vma;
 755         mm_walk.mm = vma->vm_mm;
 756         mm_walk.pte_entry = NULL;
 757         mm_walk.test_walk = NULL;
 758         mm_walk.hugetlb_entry = NULL;
 759         mm_walk.pmd_entry = hmm_vma_walk_pmd;
 760         mm_walk.pte_hole = hmm_vma_walk_hole;
 761
 762         walk_page_range(range->start, range->end, &mm_walk);
 763         return 0;
 764 }
 765 EXPORT_SYMBOL(hmm_vma_get_pfns);
 766
 767 /*
 768  * hmm_vma_range_done() - stop tracking change to CPU page table over a range
 769  * @range: range being tracked
 770  * Returns: false if range data has been invalidated, true otherwise
 771  *
 772  * Range struct is used to track updates to the CPU page table after a call to
 773  * either hmm_vma_get_pfns() or hmm_vma_fault(). Once the device driver is done
 774  * using the data,  or wants to lock updates to the data it got from those
 775  * functions, it must call the hmm_vma_range_done() function, which will then
 776  * stop tracking CPU page table updates.
 777  *
 778  * Note that device driver must still implement general CPU page table update
 779  * tracking either by using hmm_mirror (see hmm_mirror_register()) or by using
 780  * the mmu_notifier API directly.
 781  *
 782  * CPU page table update tracking done through hmm_range is only temporary and
 783  * to be used while trying to duplicate CPU page table contents for a range of
 784  * virtual addresses.
 785  *
 786  * There are two ways to use this :
 787  * again:
 788  *   hmm_vma_get_pfns(range); or hmm_vma_fault(...);
 789  *   trans = device_build_page_table_update_transaction(pfns);
 790  *   device_page_table_lock();
 791  *   if (!hmm_vma_range_done(range)) {
 792  *     device_page_table_unlock();
 793  *     goto again;
 794  *   }
 795  *   device_commit_transaction(trans);
 796  *   device_page_table_unlock();
 797  *
 798  * Or:
 799  *   hmm_vma_get_pfns(range); or hmm_vma_fault(...);
 800  *   device_page_table_lock();
 801  *   hmm_vma_range_done(range);
 802  *   device_update_page_table(range->pfns);
 803  *   device_page_table_unlock();
 804  */
 805 bool hmm_vma_range_done(struct hmm_range *range)
 806 {
 807         unsigned long npages = (range->end - range->start) >> PAGE_SHIFT;
 808         struct hmm *hmm;
 809
 810         if (range->end <= range->start) {
 811                 BUG();
 812                 return false;
 813         }
 814
 815         hmm = hmm_register(range->vma->vm_mm);
 816         if (!hmm) {
 817                 memset(range->pfns, 0, sizeof(*range->pfns) * npages);
 818                 return false;
 819         }
 820
 821         spin_lock(&hmm->lock);
 822         list_del_rcu(&range->list);
 823         spin_unlock(&hmm->lock);
 824
 825         return range->valid;
 826 }
 827 EXPORT_SYMBOL(hmm_vma_range_done);
 828
 829 /*
 830  * hmm_vma_fault() - try to fault some address in a virtual address range
 831  * @range: range being faulted
 832  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
 833  * Returns: 0 success, error otherwise (-EAGAIN means mmap_sem have been drop)
 834  *
 835  * This is similar to a regular CPU page fault except that it will not trigger
 836  * any memory migration if the memory being faulted is not accessible by CPUs.
 837  *
 838  * On error, for one virtual address in the range, the function will mark the
 839  * corresponding HMM pfn entry with an error flag.
 840  *
 841  * Expected use pattern:
 842  * retry:
 843  *   down_read(&mm->mmap_sem);
 844  *   // Find vma and address device wants to fault, initialize hmm_pfn_t
 845  *   // array accordingly
 846  *   ret = hmm_vma_fault(range, write, block);
 847  *   switch (ret) {
 848  *   case -EAGAIN:
 849  *     hmm_vma_range_done(range);
 850  *     // You might want to rate limit or yield to play nicely, you may
 851  *     // also commit any valid pfn in the array assuming that you are
 852  *     // getting true from hmm_vma_range_monitor_end()
 853  *     goto retry;
 854  *   case 0:
 855  *     break;
 856  *   case -ENOMEM:
 857  *   case -EINVAL:
 858  *   case -EPERM:
 859  *   default:
 860  *     // Handle error !
 861  *     up_read(&mm->mmap_sem)
 862  *     return;
 863  *   }
 864  *   // Take device driver lock that serialize device page table update
 865  *   driver_lock_device_page_table_update();
 866  *   hmm_vma_range_done(range);
 867  *   // Commit pfns we got from hmm_vma_fault()
 868  *   driver_unlock_device_page_table_update();
 869  *   up_read(&mm->mmap_sem)
 870  *
 871  * YOU MUST CALL hmm_vma_range_done() AFTER THIS FUNCTION RETURN SUCCESS (0)
 872  * BEFORE FREEING THE range struct OR YOU WILL HAVE SERIOUS MEMORY CORRUPTION !
 873  *
 874  * YOU HAVE BEEN WARNED !
 875  */
 876 int hmm_vma_fault(struct hmm_range *range, bool block)
 877 {
 878         struct vm_area_struct *vma = range->vma;
 879         unsigned long start = range->start;
 880         struct hmm_vma_walk hmm_vma_walk;
 881         struct mm_walk mm_walk;
 882         struct hmm *hmm;
 883         int ret;
 884
 885         /* Sanity check, this really should not happen ! */
 886         if (range->start < vma->vm_start || range->start >= vma->vm_end)
 887                 return -EINVAL;
 888         if (range->end < vma->vm_start || range->end > vma->vm_end)
 889                 return -EINVAL;
 890
 891         hmm = hmm_register(vma->vm_mm);
 892         if (!hmm) {
 893                 hmm_pfns_clear(range, range->pfns, range->start, range->end);
 894                 return -ENOMEM;
 895         }
 896         /* Caller must have registered a mirror using hmm_mirror_register() */
 897         if (!hmm->mmu_notifier.ops)
 898                 return -EINVAL;
 899
 900         /* FIXME support hugetlb fs */
 901         if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) ||
 902                         vma_is_dax(vma)) {
 903                 hmm_pfns_special(range);
 904                 return -EINVAL;
 905         }
 906
 907         if (!(vma->vm_flags & VM_READ)) {
 908                 /*
 909                  * If vma do not allow read access, then assume that it does
 910                  * not allow write access, either. Architecture that allow
 911                  * write without read access are not supported by HMM, because
 912                  * operations such has atomic access would not work.
 913                  */
 914                 hmm_pfns_clear(range, range->pfns, range->start, range->end);
 915                 return -EPERM;
 916         }
 917
 918         /* Initialize range to track CPU page table update */
 919         spin_lock(&hmm->lock);
 920         range->valid = true;
 921         list_add_rcu(&range->list, &hmm->ranges);
 922         spin_unlock(&hmm->lock);
 923
 924         hmm_vma_walk.fault = true;
 925         hmm_vma_walk.block = block;
 926         hmm_vma_walk.range = range;
 927         mm_walk.private = &hmm_vma_walk;
 928         hmm_vma_walk.last = range->start;
 929
 930         mm_walk.vma = vma;
 931         mm_walk.mm = vma->vm_mm;
 932         mm_walk.pte_entry = NULL;
 933         mm_walk.test_walk = NULL;
 934         mm_walk.hugetlb_entry = NULL;
 935         mm_walk.pmd_entry = hmm_vma_walk_pmd;
 936         mm_walk.pte_hole = hmm_vma_walk_hole;
 937
 938         do {
 939                 ret = walk_page_range(start, range->end, &mm_walk);
 940                 start = hmm_vma_walk.last;
 941         } while (ret == -EAGAIN);
 942
 943         if (ret) {
 944                 unsigned long i;
 945
 946                 i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
 947                 hmm_pfns_clear(range, &range->pfns[i], hmm_vma_walk.last,
 948                                range->end);
 949                 hmm_vma_range_done(range);
 950         }
 951         return ret;
 952 }
 953 EXPORT_SYMBOL(hmm_vma_fault);
 954 #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
 955
 956
 957 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) ||  IS_ENABLED(CONFIG_DEVICE_PUBLIC)
 958 struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
 959                                        unsigned long addr)
 960 {
 961         struct page *page;
 962
 963         page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
 964         if (!page)
 965                 return NULL;
 966         lock_page(page);
 967         return page;
 968 }
 969 EXPORT_SYMBOL(hmm_vma_alloc_locked_page);
 970
 971
 972 static void hmm_devmem_ref_release(struct percpu_ref *ref)
 973 {
 974         struct hmm_devmem *devmem;
 975
 976         devmem = container_of(ref, struct hmm_devmem, ref);
 977         complete(&devmem->completion);
 978 }
 979
 980 static void hmm_devmem_ref_exit(void *data)
 981 {
 982         struct percpu_ref *ref = data;
 983         struct hmm_devmem *devmem;
 984
 985         devmem = container_of(ref, struct hmm_devmem, ref);
 986         percpu_ref_exit(ref);
 987         devm_remove_action(devmem->device, &hmm_devmem_ref_exit, data);
 988 }
 989
 990 static void hmm_devmem_ref_kill(void *data)
 991 {
 992         struct percpu_ref *ref = data;
 993         struct hmm_devmem *devmem;
 994
 995         devmem = container_of(ref, struct hmm_devmem, ref);
 996         percpu_ref_kill(ref);
 997         wait_for_completion(&devmem->completion);
 998         devm_remove_action(devmem->device, &hmm_devmem_ref_kill, data);
 999 }
1000
1001 static int hmm_devmem_fault(struct vm_area_struct *vma,
1002                             unsigned long addr,
1003                             const struct page *page,
1004                             unsigned int flags,
1005                             pmd_t *pmdp)
1006 {
1007         struct hmm_devmem *devmem = page->pgmap->data;
1008
1009         return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp);
1010 }
1011
1012 static void hmm_devmem_free(struct page *page, void *data)
1013 {
1014         struct hmm_devmem *devmem = data;
1015
1016         page->mapping = NULL;
1017
1018         devmem->ops->free(devmem, page);
1019 }
1020
1021 static DEFINE_MUTEX(hmm_devmem_lock);
1022 static RADIX_TREE(hmm_devmem_radix, GFP_KERNEL);
1023
1024 static void hmm_devmem_radix_release(struct resource *resource)
1025 {
1026         resource_size_t key;
1027
1028         mutex_lock(&hmm_devmem_lock);
1029         for (key = resource->start;
1030              key <= resource->end;
1031              key += PA_SECTION_SIZE)
1032                 radix_tree_delete(&hmm_devmem_radix, key >> PA_SECTION_SHIFT);
1033         mutex_unlock(&hmm_devmem_lock);
1034 }
1035
1036 static void hmm_devmem_release(struct device *dev, void *data)
1037 {
1038         struct hmm_devmem *devmem = data;
1039         struct resource *resource = devmem->resource;
1040         unsigned long start_pfn, npages;
1041         struct zone *zone;
1042         struct page *page;
1043
1044         if (percpu_ref_tryget_live(&devmem->ref)) {
1045                 dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
1046                 percpu_ref_put(&devmem->ref);
1047         }
1048
1049         /* pages are dead and unused, undo the arch mapping */
1050         start_pfn = (resource->start & ~(PA_SECTION_SIZE - 1)) >> PAGE_SHIFT;
1051         npages = ALIGN(resource_size(resource), PA_SECTION_SIZE) >> PAGE_SHIFT;
1052
1053         page = pfn_to_page(start_pfn);
1054         zone = page_zone(page);
1055
1056         mem_hotplug_begin();
1057         if (resource->desc == IORES_DESC_DEVICE_PRIVATE_MEMORY)
1058                 __remove_pages(zone, start_pfn, npages, NULL);
1059         else
1060                 arch_remove_memory(start_pfn << PAGE_SHIFT,
1061                                    npages << PAGE_SHIFT, NULL);
1062         mem_hotplug_done();
1063
1064         hmm_devmem_radix_release(resource);
1065 }
1066
1067 static int hmm_devmem_pages_create(struct hmm_devmem *devmem)
1068 {
1069         resource_size_t key, align_start, align_size, align_end;
1070         struct device *device = devmem->device;
1071         int ret, nid, is_ram;
1072
1073         align_start = devmem->resource->start & ~(PA_SECTION_SIZE - 1);
1074         align_size = ALIGN(devmem->resource->start +
1075                            resource_size(devmem->resource),
1076                            PA_SECTION_SIZE) - align_start;
1077
1078         is_ram = region_intersects(align_start, align_size,
1079                                    IORESOURCE_SYSTEM_RAM,
1080                                    IORES_DESC_NONE);
1081         if (is_ram == REGION_MIXED) {
1082                 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
1083                                 __func__, devmem->resource);
1084                 return -ENXIO;
1085         }
1086         if (is_ram == REGION_INTERSECTS)
1087                 return -ENXIO;
1088
1089         if (devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY)
1090                 devmem->pagemap.type = MEMORY_DEVICE_PUBLIC;
1091         else
1092                 devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
1093
1094         devmem->pagemap.res = *devmem->resource;
1095         devmem->pagemap.page_fault = hmm_devmem_fault;
1096         devmem->pagemap.page_free = hmm_devmem_free;
1097         devmem->pagemap.dev = devmem->device;
1098         devmem->pagemap.ref = &devmem->ref;
1099         devmem->pagemap.data = devmem;
1100
1101         mutex_lock(&hmm_devmem_lock);
1102         align_end = align_start + align_size - 1;
1103         for (key = align_start; key <= align_end; key += PA_SECTION_SIZE) {
1104                 struct hmm_devmem *dup;
1105
1106                 dup = radix_tree_lookup(&hmm_devmem_radix,
1107                                         key >> PA_SECTION_SHIFT);
1108                 if (dup) {
1109                         dev_err(device, "%s: collides with mapping for %s\n",
1110                                 __func__, dev_name(dup->device));
1111                         mutex_unlock(&hmm_devmem_lock);
1112                         ret = -EBUSY;
1113                         goto error;
1114                 }
1115                 ret = radix_tree_insert(&hmm_devmem_radix,
1116                                         key >> PA_SECTION_SHIFT,
1117                                         devmem);
1118                 if (ret) {
1119                         dev_err(device, "%s: failed: %d\n", __func__, ret);
1120                         mutex_unlock(&hmm_devmem_lock);
1121                         goto error_radix;
1122                 }
1123         }
1124         mutex_unlock(&hmm_devmem_lock);
1125
1126         nid = dev_to_node(device);
1127         if (nid < 0)
1128                 nid = numa_mem_id();
1129
1130         mem_hotplug_begin();
1131         /*
1132          * For device private memory we call add_pages() as we only need to
1133          * allocate and initialize struct page for the device memory. More-
1134          * over the device memory is un-accessible thus we do not want to
1135          * create a linear mapping for the memory like arch_add_memory()
1136          * would do.
1137          *
1138          * For device public memory, which is accesible by the CPU, we do
1139          * want the linear mapping and thus use arch_add_memory().
1140          */
1141         if (devmem->pagemap.type == MEMORY_DEVICE_PUBLIC)
1142                 ret = arch_add_memory(nid, align_start, align_size, NULL,
1143                                 false);
1144         else
1145                 ret = add_pages(nid, align_start >> PAGE_SHIFT,
1146                                 align_size >> PAGE_SHIFT, NULL, false);
1147         if (ret) {
1148                 mem_hotplug_done();
1149                 goto error_add_memory;
1150         }
1151         move_pfn_range_to_zone(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
1152                                 align_start >> PAGE_SHIFT,
1153                                 align_size >> PAGE_SHIFT, NULL);
1154         mem_hotplug_done();
1155
1156         /*
1157          * Initialization of the pages has been deferred until now in order
1158          * to allow us to do the work while not holding the hotplug lock.
1159          */
1160         memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
1161                                 align_start >> PAGE_SHIFT,
1162                                 align_size >> PAGE_SHIFT, &devmem->pagemap);
1163
1164         return 0;
1165
1166 error_add_memory:
1167         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
1168 error_radix:
1169         hmm_devmem_radix_release(devmem->resource);
1170 error:
1171         return ret;
1172 }
1173
1174 static int hmm_devmem_match(struct device *dev, void *data, void *match_data)
1175 {
1176         struct hmm_devmem *devmem = data;
1177
1178         return devmem->resource == match_data;
1179 }
1180
1181 static void hmm_devmem_pages_remove(struct hmm_devmem *devmem)
1182 {
1183         devres_release(devmem->device, &hmm_devmem_release,
1184                        &hmm_devmem_match, devmem->resource);
1185 }
1186
1187 /*
1188  * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory
1189  *
1190  * @ops: memory event device driver callback (see struct hmm_devmem_ops)
1191  * @device: device struct to bind the resource too
1192  * @size: size in bytes of the device memory to add
1193  * Returns: pointer to new hmm_devmem struct ERR_PTR otherwise
1194  *
1195  * This function first finds an empty range of physical address big enough to
1196  * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which
1197  * in turn allocates struct pages. It does not do anything beyond that; all
1198  * events affecting the memory will go through the various callbacks provided
1199  * by hmm_devmem_ops struct.
1200  *
1201  * Device driver should call this function during device initialization and
1202  * is then responsible of memory management. HMM only provides helpers.
1203  */
1204 struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
1205                                   struct device *device,
1206                                   unsigned long size)
1207 {
1208         struct hmm_devmem *devmem;
1209         resource_size_t addr;
1210         int ret;
1211
1212         dev_pagemap_get_ops();
1213
1214         devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem),
1215                                    GFP_KERNEL, dev_to_node(device));
1216         if (!devmem)
1217                 return ERR_PTR(-ENOMEM);
1218
1219         init_completion(&devmem->completion);
1220         devmem->pfn_first = -1UL;
1221         devmem->pfn_last = -1UL;
1222         devmem->resource = NULL;
1223         devmem->device = device;
1224         devmem->ops = ops;
1225
1226         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1227                               0, GFP_KERNEL);
1228         if (ret)
1229                 goto error_percpu_ref;
1230
1231         ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref);
1232         if (ret)
1233                 goto error_devm_add_action;
1234
1235         size = ALIGN(size, PA_SECTION_SIZE);
1236         addr = min((unsigned long)iomem_resource.end,
1237                    (1UL << MAX_PHYSMEM_BITS) - 1);
1238         addr = addr - size + 1UL;
1239
1240         /*
1241          * FIXME add a new helper to quickly walk resource tree and find free
1242          * range
1243          *
1244          * FIXME what about ioport_resource resource ?
1245          */
1246         for (; addr > size && addr >= iomem_resource.start; addr -= size) {
1247                 ret = region_intersects(addr, size, 0, IORES_DESC_NONE);
1248                 if (ret != REGION_DISJOINT)
1249                         continue;
1250
1251                 devmem->resource = devm_request_mem_region(device, addr, size,
1252                                                            dev_name(device));
1253                 if (!devmem->resource) {
1254                         ret = -ENOMEM;
1255                         goto error_no_resource;
1256                 }
1257                 break;
1258         }
1259         if (!devmem->resource) {
1260                 ret = -ERANGE;
1261                 goto error_no_resource;
1262         }
1263
1264         devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY;
1265         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1266         devmem->pfn_last = devmem->pfn_first +
1267                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1268
1269         ret = hmm_devmem_pages_create(devmem);
1270         if (ret)
1271                 goto error_pages;
1272
1273         devres_add(device, devmem);
1274
1275         ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref);
1276         if (ret) {
1277                 hmm_devmem_remove(devmem);
1278                 return ERR_PTR(ret);
1279         }
1280
1281         return devmem;
1282
1283 error_pages:
1284         devm_release_mem_region(device, devmem->resource->start,
1285                                 resource_size(devmem->resource));
1286 error_no_resource:
1287 error_devm_add_action:
1288         hmm_devmem_ref_kill(&devmem->ref);
1289         hmm_devmem_ref_exit(&devmem->ref);
1290 error_percpu_ref:
1291         devres_free(devmem);
1292         return ERR_PTR(ret);
1293 }
1294 EXPORT_SYMBOL(hmm_devmem_add);
1295
1296 struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops,
1297                                            struct device *device,
1298                                            struct resource *res)
1299 {
1300         struct hmm_devmem *devmem;
1301         int ret;
1302
1303         if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY)
1304                 return ERR_PTR(-EINVAL);
1305
1306         dev_pagemap_get_ops();
1307
1308         devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem),
1309                                    GFP_KERNEL, dev_to_node(device));
1310         if (!devmem)
1311                 return ERR_PTR(-ENOMEM);
1312
1313         init_completion(&devmem->completion);
1314         devmem->pfn_first = -1UL;
1315         devmem->pfn_last = -1UL;
1316         devmem->resource = res;
1317         devmem->device = device;
1318         devmem->ops = ops;
1319
1320         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1321                               0, GFP_KERNEL);
1322         if (ret)
1323                 goto error_percpu_ref;
1324
1325         ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref);
1326         if (ret)
1327                 goto error_devm_add_action;
1328
1329
1330         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1331         devmem->pfn_last = devmem->pfn_first +
1332                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1333
1334         ret = hmm_devmem_pages_create(devmem);
1335         if (ret)
1336                 goto error_devm_add_action;
1337
1338         devres_add(device, devmem);
1339
1340         ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref);
1341         if (ret) {
1342                 hmm_devmem_remove(devmem);
1343                 return ERR_PTR(ret);
1344         }
1345
1346         return devmem;
1347
1348 error_devm_add_action:
1349         hmm_devmem_ref_kill(&devmem->ref);
1350         hmm_devmem_ref_exit(&devmem->ref);
1351 error_percpu_ref:
1352         devres_free(devmem);
1353         return ERR_PTR(ret);
1354 }
1355 EXPORT_SYMBOL(hmm_devmem_add_resource);
1356
1357 /*
1358  * hmm_devmem_remove() - remove device memory (kill and free ZONE_DEVICE)
1359  *
1360  * @devmem: hmm_devmem struct use to track and manage the ZONE_DEVICE memory
1361  *
1362  * This will hot-unplug memory that was hotplugged by hmm_devmem_add on behalf
1363  * of the device driver. It will free struct page and remove the resource that
1364  * reserved the physical address range for this device memory.
1365  */
1366 void hmm_devmem_remove(struct hmm_devmem *devmem)
1367 {
1368         resource_size_t start, size;
1369         struct device *device;
1370         bool cdm = false;
1371
1372         if (!devmem)
1373                 return;
1374
1375         device = devmem->device;
1376         start = devmem->resource->start;
1377         size = resource_size(devmem->resource);
1378
1379         cdm = devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY;
1380         hmm_devmem_ref_kill(&devmem->ref);
1381         hmm_devmem_ref_exit(&devmem->ref);
1382         hmm_devmem_pages_remove(devmem);
1383
1384         if (!cdm)
1385                 devm_release_mem_region(device, start, size);
1386 }
1387 EXPORT_SYMBOL(hmm_devmem_remove);
1388
1389 /*
1390  * A device driver that wants to handle multiple devices memory through a
1391  * single fake device can use hmm_device to do so. This is purely a helper
1392  * and it is not needed to make use of any HMM functionality.
1393  */
1394 #define HMM_DEVICE_MAX 256
1395
1396 static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX);
1397 static DEFINE_SPINLOCK(hmm_device_lock);
1398 static struct class *hmm_device_class;
1399 static dev_t hmm_device_devt;
1400
1401 static void hmm_device_release(struct device *device)
1402 {
1403         struct hmm_device *hmm_device;
1404
1405         hmm_device = container_of(device, struct hmm_device, device);
1406         spin_lock(&hmm_device_lock);
1407         clear_bit(hmm_device->minor, hmm_device_mask);
1408         spin_unlock(&hmm_device_lock);
1409
1410         kfree(hmm_device);
1411 }
1412
1413 struct hmm_device *hmm_device_new(void *drvdata)
1414 {
1415         struct hmm_device *hmm_device;
1416
1417         hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL);
1418         if (!hmm_device)
1419                 return ERR_PTR(-ENOMEM);
1420
1421         spin_lock(&hmm_device_lock);
1422         hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX);
1423         if (hmm_device->minor >= HMM_DEVICE_MAX) {
1424                 spin_unlock(&hmm_device_lock);
1425                 kfree(hmm_device);
1426                 return ERR_PTR(-EBUSY);
1427         }
1428         set_bit(hmm_device->minor, hmm_device_mask);
1429         spin_unlock(&hmm_device_lock);
1430
1431         dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor);
1432         hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt),
1433                                         hmm_device->minor);
1434         hmm_device->device.release = hmm_device_release;
1435         dev_set_drvdata(&hmm_device->device, drvdata);
1436         hmm_device->device.class = hmm_device_class;
1437         device_initialize(&hmm_device->device);
1438
1439         return hmm_device;
1440 }
1441 EXPORT_SYMBOL(hmm_device_new);
1442
1443 void hmm_device_put(struct hmm_device *hmm_device)
1444 {
1445         put_device(&hmm_device->device);
1446 }
1447 EXPORT_SYMBOL(hmm_device_put);
1448
1449 static int __init hmm_init(void)
1450 {
1451         int ret;
1452
1453         ret = alloc_chrdev_region(&hmm_device_devt, 0,
1454                                   HMM_DEVICE_MAX,
1455                                   "hmm_device");
1456         if (ret)
1457                 return ret;
1458
1459         hmm_device_class = class_create(THIS_MODULE, "hmm_device");
1460         if (IS_ERR(hmm_device_class)) {
1461                 unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX);
1462                 return PTR_ERR(hmm_device_class);
1463         }
1464         return 0;
1465 }
1466
1467 device_initcall(hmm_init);
1468 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */