mm/hmm.c

   1 /*
   2  * Copyright 2013 Red Hat Inc.
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * Authors: JÃ©rÃ´me Glisse <jglisse@redhat.com>
  15  */
  16 /*
  17  * Refer to include/linux/hmm.h for information about heterogeneous memory
  18  * management or HMM for short.
  19  */
  20 #include <linux/mm.h>
  21 #include <linux/hmm.h>
  22 #include <linux/init.h>
  23 #include <linux/rmap.h>
  24 #include <linux/swap.h>
  25 #include <linux/slab.h>
  26 #include <linux/sched.h>
  27 #include <linux/mmzone.h>
  28 #include <linux/pagemap.h>
  29 #include <linux/swapops.h>
  30 #include <linux/hugetlb.h>
  31 #include <linux/memremap.h>
  32 #include <linux/jump_label.h>
  33 #include <linux/mmu_notifier.h>
  34 #include <linux/memory_hotplug.h>
  35
  36 #define PA_SECTION_SIZE (1UL << PA_SECTION_SHIFT)
  37
  38 #if defined(CONFIG_DEVICE_PRIVATE) || defined(CONFIG_DEVICE_PUBLIC)
  39 /*
  40  * Device private memory see HMM (Documentation/vm/hmm.txt) or hmm.h
  41  */
  42 DEFINE_STATIC_KEY_FALSE(device_private_key);
  43 EXPORT_SYMBOL(device_private_key);
  44 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */
  45
  46
  47 #if IS_ENABLED(CONFIG_HMM_MIRROR)
  48 static const struct mmu_notifier_ops hmm_mmu_notifier_ops;
  49
  50 /*
  51  * struct hmm - HMM per mm struct
  52  *
  53  * @mm: mm struct this HMM struct is bound to
  54  * @lock: lock protecting ranges list
  55  * @sequence: we track updates to the CPU page table with a sequence number
  56  * @ranges: list of range being snapshotted
  57  * @mirrors: list of mirrors for this mm
  58  * @mmu_notifier: mmu notifier to track updates to CPU page table
  59  * @mirrors_sem: read/write semaphore protecting the mirrors list
  60  */
  61 struct hmm {
  62         struct mm_struct        *mm;
  63         spinlock_t              lock;
  64         atomic_t                sequence;
  65         struct list_head        ranges;
  66         struct list_head        mirrors;
  67         struct mmu_notifier     mmu_notifier;
  68         struct rw_semaphore     mirrors_sem;
  69 };
  70
  71 /*
  72  * hmm_register - register HMM against an mm (HMM internal)
  73  *
  74  * @mm: mm struct to attach to
  75  *
  76  * This is not intended to be used directly by device drivers. It allocates an
  77  * HMM struct if mm does not have one, and initializes it.
  78  */
  79 static struct hmm *hmm_register(struct mm_struct *mm)
  80 {
  81         struct hmm *hmm = READ_ONCE(mm->hmm);
  82         bool cleanup = false;
  83
  84         /*
  85          * The hmm struct can only be freed once the mm_struct goes away,
  86          * hence we should always have pre-allocated an new hmm struct
  87          * above.
  88          */
  89         if (hmm)
  90                 return hmm;
  91
  92         hmm = kmalloc(sizeof(*hmm), GFP_KERNEL);
  93         if (!hmm)
  94                 return NULL;
  95         INIT_LIST_HEAD(&hmm->mirrors);
  96         init_rwsem(&hmm->mirrors_sem);
  97         atomic_set(&hmm->sequence, 0);
  98         hmm->mmu_notifier.ops = NULL;
  99         INIT_LIST_HEAD(&hmm->ranges);
 100         spin_lock_init(&hmm->lock);
 101         hmm->mm = mm;
 102
 103         /*
 104          * We should only get here if hold the mmap_sem in write mode ie on
 105          * registration of first mirror through hmm_mirror_register()
 106          */
 107         hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops;
 108         if (__mmu_notifier_register(&hmm->mmu_notifier, mm)) {
 109                 kfree(hmm);
 110                 return NULL;
 111         }
 112
 113         spin_lock(&mm->page_table_lock);
 114         if (!mm->hmm)
 115                 mm->hmm = hmm;
 116         else
 117                 cleanup = true;
 118         spin_unlock(&mm->page_table_lock);
 119
 120         if (cleanup) {
 121                 mmu_notifier_unregister(&hmm->mmu_notifier, mm);
 122                 kfree(hmm);
 123         }
 124
 125         return mm->hmm;
 126 }
 127
 128 void hmm_mm_destroy(struct mm_struct *mm)
 129 {
 130         kfree(mm->hmm);
 131 }
 132
 133 static void hmm_invalidate_range(struct hmm *hmm,
 134                                  enum hmm_update_type action,
 135                                  unsigned long start,
 136                                  unsigned long end)
 137 {
 138         struct hmm_mirror *mirror;
 139         struct hmm_range *range;
 140
 141         spin_lock(&hmm->lock);
 142         list_for_each_entry(range, &hmm->ranges, list) {
 143                 unsigned long addr, idx, npages;
 144
 145                 if (end < range->start || start >= range->end)
 146                         continue;
 147
 148                 range->valid = false;
 149                 addr = max(start, range->start);
 150                 idx = (addr - range->start) >> PAGE_SHIFT;
 151                 npages = (min(range->end, end) - addr) >> PAGE_SHIFT;
 152                 memset(&range->pfns[idx], 0, sizeof(*range->pfns) * npages);
 153         }
 154         spin_unlock(&hmm->lock);
 155
 156         down_read(&hmm->mirrors_sem);
 157         list_for_each_entry(mirror, &hmm->mirrors, list)
 158                 mirror->ops->sync_cpu_device_pagetables(mirror, action,
 159                                                         start, end);
 160         up_read(&hmm->mirrors_sem);
 161 }
 162
 163 static void hmm_invalidate_range_start(struct mmu_notifier *mn,
 164                                        struct mm_struct *mm,
 165                                        unsigned long start,
 166                                        unsigned long end)
 167 {
 168         struct hmm *hmm = mm->hmm;
 169
 170         VM_BUG_ON(!hmm);
 171
 172         atomic_inc(&hmm->sequence);
 173 }
 174
 175 static void hmm_invalidate_range_end(struct mmu_notifier *mn,
 176                                      struct mm_struct *mm,
 177                                      unsigned long start,
 178                                      unsigned long end)
 179 {
 180         struct hmm *hmm = mm->hmm;
 181
 182         VM_BUG_ON(!hmm);
 183
 184         hmm_invalidate_range(mm->hmm, HMM_UPDATE_INVALIDATE, start, end);
 185 }
 186
 187 static const struct mmu_notifier_ops hmm_mmu_notifier_ops = {
 188         .invalidate_range_start = hmm_invalidate_range_start,
 189         .invalidate_range_end   = hmm_invalidate_range_end,
 190 };
 191
 192 /*
 193  * hmm_mirror_register() - register a mirror against an mm
 194  *
 195  * @mirror: new mirror struct to register
 196  * @mm: mm to register against
 197  *
 198  * To start mirroring a process address space, the device driver must register
 199  * an HMM mirror struct.
 200  *
 201  * THE mm->mmap_sem MUST BE HELD IN WRITE MODE !
 202  */
 203 int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm)
 204 {
 205         /* Sanity check */
 206         if (!mm || !mirror || !mirror->ops)
 207                 return -EINVAL;
 208
 209         mirror->hmm = hmm_register(mm);
 210         if (!mirror->hmm)
 211                 return -ENOMEM;
 212
 213         down_write(&mirror->hmm->mirrors_sem);
 214         list_add(&mirror->list, &mirror->hmm->mirrors);
 215         up_write(&mirror->hmm->mirrors_sem);
 216
 217         return 0;
 218 }
 219 EXPORT_SYMBOL(hmm_mirror_register);
 220
 221 /*
 222  * hmm_mirror_unregister() - unregister a mirror
 223  *
 224  * @mirror: new mirror struct to register
 225  *
 226  * Stop mirroring a process address space, and cleanup.
 227  */
 228 void hmm_mirror_unregister(struct hmm_mirror *mirror)
 229 {
 230         struct hmm *hmm = mirror->hmm;
 231
 232         down_write(&hmm->mirrors_sem);
 233         list_del(&mirror->list);
 234         up_write(&hmm->mirrors_sem);
 235 }
 236 EXPORT_SYMBOL(hmm_mirror_unregister);
 237
 238 struct hmm_vma_walk {
 239         struct hmm_range        *range;
 240         unsigned long           last;
 241         bool                    fault;
 242         bool                    block;
 243         bool                    write;
 244 };
 245
 246 static int hmm_vma_do_fault(struct mm_walk *walk,
 247                             unsigned long addr,
 248                             hmm_pfn_t *pfn)
 249 {
 250         unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE;
 251         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 252         struct vm_area_struct *vma = walk->vma;
 253         int r;
 254
 255         flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY;
 256         flags |= hmm_vma_walk->write ? FAULT_FLAG_WRITE : 0;
 257         r = handle_mm_fault(vma, addr, flags);
 258         if (r & VM_FAULT_RETRY)
 259                 return -EBUSY;
 260         if (r & VM_FAULT_ERROR) {
 261                 *pfn = HMM_PFN_ERROR;
 262                 return -EFAULT;
 263         }
 264
 265         return -EAGAIN;
 266 }
 267
 268 static void hmm_pfns_special(hmm_pfn_t *pfns,
 269                              unsigned long addr,
 270                              unsigned long end)
 271 {
 272         for (; addr < end; addr += PAGE_SIZE, pfns++)
 273                 *pfns = HMM_PFN_SPECIAL;
 274 }
 275
 276 static int hmm_pfns_bad(unsigned long addr,
 277                         unsigned long end,
 278                         struct mm_walk *walk)
 279 {
 280         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 281         struct hmm_range *range = hmm_vma_walk->range;
 282         hmm_pfn_t *pfns = range->pfns;
 283         unsigned long i;
 284
 285         i = (addr - range->start) >> PAGE_SHIFT;
 286         for (; addr < end; addr += PAGE_SIZE, i++)
 287                 pfns[i] = HMM_PFN_ERROR;
 288
 289         return 0;
 290 }
 291
 292 static void hmm_pfns_clear(hmm_pfn_t *pfns,
 293                            unsigned long addr,
 294                            unsigned long end)
 295 {
 296         for (; addr < end; addr += PAGE_SIZE, pfns++)
 297                 *pfns = 0;
 298 }
 299
 300 static int hmm_vma_walk_hole(unsigned long addr,
 301                              unsigned long end,
 302                              struct mm_walk *walk)
 303 {
 304         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 305         struct hmm_range *range = hmm_vma_walk->range;
 306         hmm_pfn_t *pfns = range->pfns;
 307         unsigned long i;
 308
 309         hmm_vma_walk->last = addr;
 310         i = (addr - range->start) >> PAGE_SHIFT;
 311         for (; addr < end; addr += PAGE_SIZE, i++) {
 312                 pfns[i] = HMM_PFN_EMPTY;
 313                 if (hmm_vma_walk->fault) {
 314                         int ret;
 315
 316                         ret = hmm_vma_do_fault(walk, addr, &pfns[i]);
 317                         if (ret != -EAGAIN)
 318                                 return ret;
 319                 }
 320         }
 321
 322         return hmm_vma_walk->fault ? -EAGAIN : 0;
 323 }
 324
 325 static int hmm_vma_walk_clear(unsigned long addr,
 326                               unsigned long end,
 327                               struct mm_walk *walk)
 328 {
 329         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 330         struct hmm_range *range = hmm_vma_walk->range;
 331         hmm_pfn_t *pfns = range->pfns;
 332         unsigned long i;
 333
 334         hmm_vma_walk->last = addr;
 335         i = (addr - range->start) >> PAGE_SHIFT;
 336         for (; addr < end; addr += PAGE_SIZE, i++) {
 337                 pfns[i] = 0;
 338                 if (hmm_vma_walk->fault) {
 339                         int ret;
 340
 341                         ret = hmm_vma_do_fault(walk, addr, &pfns[i]);
 342                         if (ret != -EAGAIN)
 343                                 return ret;
 344                 }
 345         }
 346
 347         return hmm_vma_walk->fault ? -EAGAIN : 0;
 348 }
 349
 350 static int hmm_vma_walk_pmd(pmd_t *pmdp,
 351                             unsigned long start,
 352                             unsigned long end,
 353                             struct mm_walk *walk)
 354 {
 355         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 356         struct hmm_range *range = hmm_vma_walk->range;
 357         struct vm_area_struct *vma = walk->vma;
 358         hmm_pfn_t *pfns = range->pfns;
 359         unsigned long addr = start, i;
 360         bool write_fault;
 361         hmm_pfn_t flag;
 362         pte_t *ptep;
 363
 364         i = (addr - range->start) >> PAGE_SHIFT;
 365         flag = vma->vm_flags & VM_READ ? HMM_PFN_READ : 0;
 366         write_fault = hmm_vma_walk->fault & hmm_vma_walk->write;
 367
 368 again:
 369         if (pmd_none(*pmdp))
 370                 return hmm_vma_walk_hole(start, end, walk);
 371
 372         if (pmd_huge(*pmdp) && vma->vm_flags & VM_HUGETLB)
 373                 return hmm_pfns_bad(start, end, walk);
 374
 375         if (pmd_devmap(*pmdp) || pmd_trans_huge(*pmdp)) {
 376                 unsigned long pfn;
 377                 pmd_t pmd;
 378
 379                 /*
 380                  * No need to take pmd_lock here, even if some other threads
 381                  * is splitting the huge pmd we will get that event through
 382                  * mmu_notifier callback.
 383                  *
 384                  * So just read pmd value and check again its a transparent
 385                  * huge or device mapping one and compute corresponding pfn
 386                  * values.
 387                  */
 388                 pmd = pmd_read_atomic(pmdp);
 389                 barrier();
 390                 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
 391                         goto again;
 392                 if (pmd_protnone(pmd))
 393                         return hmm_vma_walk_clear(start, end, walk);
 394
 395                 if (write_fault && !pmd_write(pmd))
 396                         return hmm_vma_walk_clear(start, end, walk);
 397
 398                 pfn = pmd_pfn(pmd) + pte_index(addr);
 399                 flag |= pmd_write(pmd) ? HMM_PFN_WRITE : 0;
 400                 for (; addr < end; addr += PAGE_SIZE, i++, pfn++)
 401                         pfns[i] = hmm_pfn_t_from_pfn(pfn) | flag;
 402                 return 0;
 403         }
 404
 405         if (pmd_bad(*pmdp))
 406                 return hmm_pfns_bad(start, end, walk);
 407
 408         ptep = pte_offset_map(pmdp, addr);
 409         for (; addr < end; addr += PAGE_SIZE, ptep++, i++) {
 410                 pte_t pte = *ptep;
 411
 412                 pfns[i] = 0;
 413
 414                 if (pte_none(pte)) {
 415                         pfns[i] = HMM_PFN_EMPTY;
 416                         if (hmm_vma_walk->fault)
 417                                 goto fault;
 418                         continue;
 419                 }
 420
 421                 if (!pte_present(pte)) {
 422                         swp_entry_t entry;
 423
 424                         if (!non_swap_entry(entry)) {
 425                                 if (hmm_vma_walk->fault)
 426                                         goto fault;
 427                                 continue;
 428                         }
 429
 430                         entry = pte_to_swp_entry(pte);
 431
 432                         /*
 433                          * This is a special swap entry, ignore migration, use
 434                          * device and report anything else as error.
 435                          */
 436                         if (is_device_private_entry(entry)) {
 437                                 pfns[i] = hmm_pfn_t_from_pfn(swp_offset(entry));
 438                                 if (is_write_device_private_entry(entry)) {
 439                                         pfns[i] |= HMM_PFN_WRITE;
 440                                 } else if (write_fault)
 441                                         goto fault;
 442                                 pfns[i] |= HMM_PFN_DEVICE_UNADDRESSABLE;
 443                                 pfns[i] |= flag;
 444                         } else if (is_migration_entry(entry)) {
 445                                 if (hmm_vma_walk->fault) {
 446                                         pte_unmap(ptep);
 447                                         hmm_vma_walk->last = addr;
 448                                         migration_entry_wait(vma->vm_mm,
 449                                                              pmdp, addr);
 450                                         return -EAGAIN;
 451                                 }
 452                                 continue;
 453                         } else {
 454                                 /* Report error for everything else */
 455                                 pfns[i] = HMM_PFN_ERROR;
 456                         }
 457                         continue;
 458                 }
 459
 460                 if (write_fault && !pte_write(pte))
 461                         goto fault;
 462
 463                 pfns[i] = hmm_pfn_t_from_pfn(pte_pfn(pte)) | flag;
 464                 pfns[i] |= pte_write(pte) ? HMM_PFN_WRITE : 0;
 465                 continue;
 466
 467 fault:
 468                 pte_unmap(ptep);
 469                 /* Fault all pages in range */
 470                 return hmm_vma_walk_clear(start, end, walk);
 471         }
 472         pte_unmap(ptep - 1);
 473
 474         return 0;
 475 }
 476
 477 /*
 478  * hmm_vma_get_pfns() - snapshot CPU page table for a range of virtual addresses
 479  * @vma: virtual memory area containing the virtual address range
 480  * @range: used to track snapshot validity
 481  * @start: range virtual start address (inclusive)
 482  * @end: range virtual end address (exclusive)
 483  * @entries: array of hmm_pfn_t: provided by the caller, filled in by function
 484  * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, 0 success
 485  *
 486  * This snapshots the CPU page table for a range of virtual addresses. Snapshot
 487  * validity is tracked by range struct. See hmm_vma_range_done() for further
 488  * information.
 489  *
 490  * The range struct is initialized here. It tracks the CPU page table, but only
 491  * if the function returns success (0), in which case the caller must then call
 492  * hmm_vma_range_done() to stop CPU page table update tracking on this range.
 493  *
 494  * NOT CALLING hmm_vma_range_done() IF FUNCTION RETURNS 0 WILL LEAD TO SERIOUS
 495  * MEMORY CORRUPTION ! YOU HAVE BEEN WARNED !
 496  */
 497 int hmm_vma_get_pfns(struct vm_area_struct *vma,
 498                      struct hmm_range *range,
 499                      unsigned long start,
 500                      unsigned long end,
 501                      hmm_pfn_t *pfns)
 502 {
 503         struct hmm_vma_walk hmm_vma_walk;
 504         struct mm_walk mm_walk;
 505         struct hmm *hmm;
 506
 507         /* FIXME support hugetlb fs */
 508         if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) {
 509                 hmm_pfns_special(pfns, start, end);
 510                 return -EINVAL;
 511         }
 512
 513         /* Sanity check, this really should not happen ! */
 514         if (start < vma->vm_start || start >= vma->vm_end)
 515                 return -EINVAL;
 516         if (end < vma->vm_start || end > vma->vm_end)
 517                 return -EINVAL;
 518
 519         hmm = hmm_register(vma->vm_mm);
 520         if (!hmm)
 521                 return -ENOMEM;
 522         /* Caller must have registered a mirror, via hmm_mirror_register() ! */
 523         if (!hmm->mmu_notifier.ops)
 524                 return -EINVAL;
 525
 526         /* Initialize range to track CPU page table update */
 527         range->start = start;
 528         range->pfns = pfns;
 529         range->end = end;
 530         spin_lock(&hmm->lock);
 531         range->valid = true;
 532         list_add_rcu(&range->list, &hmm->ranges);
 533         spin_unlock(&hmm->lock);
 534
 535         hmm_vma_walk.fault = false;
 536         hmm_vma_walk.range = range;
 537         mm_walk.private = &hmm_vma_walk;
 538
 539         mm_walk.vma = vma;
 540         mm_walk.mm = vma->vm_mm;
 541         mm_walk.pte_entry = NULL;
 542         mm_walk.test_walk = NULL;
 543         mm_walk.hugetlb_entry = NULL;
 544         mm_walk.pmd_entry = hmm_vma_walk_pmd;
 545         mm_walk.pte_hole = hmm_vma_walk_hole;
 546
 547         walk_page_range(start, end, &mm_walk);
 548         return 0;
 549 }
 550 EXPORT_SYMBOL(hmm_vma_get_pfns);
 551
 552 /*
 553  * hmm_vma_range_done() - stop tracking change to CPU page table over a range
 554  * @vma: virtual memory area containing the virtual address range
 555  * @range: range being tracked
 556  * Returns: false if range data has been invalidated, true otherwise
 557  *
 558  * Range struct is used to track updates to the CPU page table after a call to
 559  * either hmm_vma_get_pfns() or hmm_vma_fault(). Once the device driver is done
 560  * using the data,  or wants to lock updates to the data it got from those
 561  * functions, it must call the hmm_vma_range_done() function, which will then
 562  * stop tracking CPU page table updates.
 563  *
 564  * Note that device driver must still implement general CPU page table update
 565  * tracking either by using hmm_mirror (see hmm_mirror_register()) or by using
 566  * the mmu_notifier API directly.
 567  *
 568  * CPU page table update tracking done through hmm_range is only temporary and
 569  * to be used while trying to duplicate CPU page table contents for a range of
 570  * virtual addresses.
 571  *
 572  * There are two ways to use this :
 573  * again:
 574  *   hmm_vma_get_pfns(vma, range, start, end, pfns); or hmm_vma_fault(...);
 575  *   trans = device_build_page_table_update_transaction(pfns);
 576  *   device_page_table_lock();
 577  *   if (!hmm_vma_range_done(vma, range)) {
 578  *     device_page_table_unlock();
 579  *     goto again;
 580  *   }
 581  *   device_commit_transaction(trans);
 582  *   device_page_table_unlock();
 583  *
 584  * Or:
 585  *   hmm_vma_get_pfns(vma, range, start, end, pfns); or hmm_vma_fault(...);
 586  *   device_page_table_lock();
 587  *   hmm_vma_range_done(vma, range);
 588  *   device_update_page_table(pfns);
 589  *   device_page_table_unlock();
 590  */
 591 bool hmm_vma_range_done(struct vm_area_struct *vma, struct hmm_range *range)
 592 {
 593         unsigned long npages = (range->end - range->start) >> PAGE_SHIFT;
 594         struct hmm *hmm;
 595
 596         if (range->end <= range->start) {
 597                 BUG();
 598                 return false;
 599         }
 600
 601         hmm = hmm_register(vma->vm_mm);
 602         if (!hmm) {
 603                 memset(range->pfns, 0, sizeof(*range->pfns) * npages);
 604                 return false;
 605         }
 606
 607         spin_lock(&hmm->lock);
 608         list_del_rcu(&range->list);
 609         spin_unlock(&hmm->lock);
 610
 611         return range->valid;
 612 }
 613 EXPORT_SYMBOL(hmm_vma_range_done);
 614
 615 /*
 616  * hmm_vma_fault() - try to fault some address in a virtual address range
 617  * @vma: virtual memory area containing the virtual address range
 618  * @range: use to track pfns array content validity
 619  * @start: fault range virtual start address (inclusive)
 620  * @end: fault range virtual end address (exclusive)
 621  * @pfns: array of hmm_pfn_t, only entry with fault flag set will be faulted
 622  * @write: is it a write fault
 623  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
 624  * Returns: 0 success, error otherwise (-EAGAIN means mmap_sem have been drop)
 625  *
 626  * This is similar to a regular CPU page fault except that it will not trigger
 627  * any memory migration if the memory being faulted is not accessible by CPUs.
 628  *
 629  * On error, for one virtual address in the range, the function will set the
 630  * hmm_pfn_t error flag for the corresponding pfn entry.
 631  *
 632  * Expected use pattern:
 633  * retry:
 634  *   down_read(&mm->mmap_sem);
 635  *   // Find vma and address device wants to fault, initialize hmm_pfn_t
 636  *   // array accordingly
 637  *   ret = hmm_vma_fault(vma, start, end, pfns, allow_retry);
 638  *   switch (ret) {
 639  *   case -EAGAIN:
 640  *     hmm_vma_range_done(vma, range);
 641  *     // You might want to rate limit or yield to play nicely, you may
 642  *     // also commit any valid pfn in the array assuming that you are
 643  *     // getting true from hmm_vma_range_monitor_end()
 644  *     goto retry;
 645  *   case 0:
 646  *     break;
 647  *   default:
 648  *     // Handle error !
 649  *     up_read(&mm->mmap_sem)
 650  *     return;
 651  *   }
 652  *   // Take device driver lock that serialize device page table update
 653  *   driver_lock_device_page_table_update();
 654  *   hmm_vma_range_done(vma, range);
 655  *   // Commit pfns we got from hmm_vma_fault()
 656  *   driver_unlock_device_page_table_update();
 657  *   up_read(&mm->mmap_sem)
 658  *
 659  * YOU MUST CALL hmm_vma_range_done() AFTER THIS FUNCTION RETURN SUCCESS (0)
 660  * BEFORE FREEING THE range struct OR YOU WILL HAVE SERIOUS MEMORY CORRUPTION !
 661  *
 662  * YOU HAVE BEEN WARNED !
 663  */
 664 int hmm_vma_fault(struct vm_area_struct *vma,
 665                   struct hmm_range *range,
 666                   unsigned long start,
 667                   unsigned long end,
 668                   hmm_pfn_t *pfns,
 669                   bool write,
 670                   bool block)
 671 {
 672         struct hmm_vma_walk hmm_vma_walk;
 673         struct mm_walk mm_walk;
 674         struct hmm *hmm;
 675         int ret;
 676
 677         /* Sanity check, this really should not happen ! */
 678         if (start < vma->vm_start || start >= vma->vm_end)
 679                 return -EINVAL;
 680         if (end < vma->vm_start || end > vma->vm_end)
 681                 return -EINVAL;
 682
 683         hmm = hmm_register(vma->vm_mm);
 684         if (!hmm) {
 685                 hmm_pfns_clear(pfns, start, end);
 686                 return -ENOMEM;
 687         }
 688         /* Caller must have registered a mirror using hmm_mirror_register() */
 689         if (!hmm->mmu_notifier.ops)
 690                 return -EINVAL;
 691
 692         /* Initialize range to track CPU page table update */
 693         range->start = start;
 694         range->pfns = pfns;
 695         range->end = end;
 696         spin_lock(&hmm->lock);
 697         range->valid = true;
 698         list_add_rcu(&range->list, &hmm->ranges);
 699         spin_unlock(&hmm->lock);
 700
 701         /* FIXME support hugetlb fs */
 702         if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) {
 703                 hmm_pfns_special(pfns, start, end);
 704                 return 0;
 705         }
 706
 707         hmm_vma_walk.fault = true;
 708         hmm_vma_walk.write = write;
 709         hmm_vma_walk.block = block;
 710         hmm_vma_walk.range = range;
 711         mm_walk.private = &hmm_vma_walk;
 712         hmm_vma_walk.last = range->start;
 713
 714         mm_walk.vma = vma;
 715         mm_walk.mm = vma->vm_mm;
 716         mm_walk.pte_entry = NULL;
 717         mm_walk.test_walk = NULL;
 718         mm_walk.hugetlb_entry = NULL;
 719         mm_walk.pmd_entry = hmm_vma_walk_pmd;
 720         mm_walk.pte_hole = hmm_vma_walk_hole;
 721
 722         do {
 723                 ret = walk_page_range(start, end, &mm_walk);
 724                 start = hmm_vma_walk.last;
 725         } while (ret == -EAGAIN);
 726
 727         if (ret) {
 728                 unsigned long i;
 729
 730                 i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
 731                 hmm_pfns_clear(&pfns[i], hmm_vma_walk.last, end);
 732                 hmm_vma_range_done(vma, range);
 733         }
 734         return ret;
 735 }
 736 EXPORT_SYMBOL(hmm_vma_fault);
 737 #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
 738
 739
 740 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) ||  IS_ENABLED(CONFIG_DEVICE_PUBLIC)
 741 struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
 742                                        unsigned long addr)
 743 {
 744         struct page *page;
 745
 746         page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
 747         if (!page)
 748                 return NULL;
 749         lock_page(page);
 750         return page;
 751 }
 752 EXPORT_SYMBOL(hmm_vma_alloc_locked_page);
 753
 754
 755 static void hmm_devmem_ref_release(struct percpu_ref *ref)
 756 {
 757         struct hmm_devmem *devmem;
 758
 759         devmem = container_of(ref, struct hmm_devmem, ref);
 760         complete(&devmem->completion);
 761 }
 762
 763 static void hmm_devmem_ref_exit(void *data)
 764 {
 765         struct percpu_ref *ref = data;
 766         struct hmm_devmem *devmem;
 767
 768         devmem = container_of(ref, struct hmm_devmem, ref);
 769         percpu_ref_exit(ref);
 770 }
 771
 772 static void hmm_devmem_ref_kill(void *data)
 773 {
 774         struct percpu_ref *ref = data;
 775         struct hmm_devmem *devmem;
 776
 777         devmem = container_of(ref, struct hmm_devmem, ref);
 778         percpu_ref_kill(ref);
 779         wait_for_completion(&devmem->completion);
 780 }
 781
 782 static int hmm_devmem_fault(struct vm_area_struct *vma,
 783                             unsigned long addr,
 784                             const struct page *page,
 785                             unsigned int flags,
 786                             pmd_t *pmdp)
 787 {
 788         struct hmm_devmem *devmem = page->pgmap->data;
 789
 790         return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp);
 791 }
 792
 793 static void hmm_devmem_free(struct page *page, void *data)
 794 {
 795         struct hmm_devmem *devmem = data;
 796
 797         devmem->ops->free(devmem, page);
 798 }
 799
 800 static DEFINE_MUTEX(hmm_devmem_lock);
 801 static RADIX_TREE(hmm_devmem_radix, GFP_KERNEL);
 802
 803 static void hmm_devmem_radix_release(struct resource *resource)
 804 {
 805         resource_size_t key, align_start, align_size, align_end;
 806
 807         align_start = resource->start & ~(PA_SECTION_SIZE - 1);
 808         align_size = ALIGN(resource_size(resource), PA_SECTION_SIZE);
 809         align_end = align_start + align_size - 1;
 810
 811         mutex_lock(&hmm_devmem_lock);
 812         for (key = resource->start;
 813              key <= resource->end;
 814              key += PA_SECTION_SIZE)
 815                 radix_tree_delete(&hmm_devmem_radix, key >> PA_SECTION_SHIFT);
 816         mutex_unlock(&hmm_devmem_lock);
 817 }
 818
 819 static void hmm_devmem_release(void *data)
 820 {
 821         struct hmm_devmem *devmem = data;
 822         struct resource *resource = devmem->resource;
 823         unsigned long start_pfn, npages;
 824         struct zone *zone;
 825         struct page *page;
 826
 827         /* pages are dead and unused, undo the arch mapping */
 828         start_pfn = (resource->start & ~(PA_SECTION_SIZE - 1)) >> PAGE_SHIFT;
 829         npages = ALIGN(resource_size(resource), PA_SECTION_SIZE) >> PAGE_SHIFT;
 830
 831         page = pfn_to_page(start_pfn);
 832         zone = page_zone(page);
 833
 834         mem_hotplug_begin();
 835         if (resource->desc == IORES_DESC_DEVICE_PRIVATE_MEMORY)
 836                 __remove_pages(zone, start_pfn, npages);
 837         else
 838                 arch_remove_memory(start_pfn << PAGE_SHIFT,
 839                                    npages << PAGE_SHIFT);
 840         mem_hotplug_done();
 841
 842         hmm_devmem_radix_release(resource);
 843 }
 844
 845 static struct hmm_devmem *hmm_devmem_find(resource_size_t phys)
 846 {
 847         WARN_ON_ONCE(!rcu_read_lock_held());
 848
 849         return radix_tree_lookup(&hmm_devmem_radix, phys >> PA_SECTION_SHIFT);
 850 }
 851
 852 static int hmm_devmem_pages_create(struct hmm_devmem *devmem)
 853 {
 854         resource_size_t key, align_start, align_size, align_end;
 855         struct device *device = devmem->device;
 856         int ret, nid, is_ram;
 857         unsigned long pfn;
 858
 859         align_start = devmem->resource->start & ~(PA_SECTION_SIZE - 1);
 860         align_size = ALIGN(devmem->resource->start +
 861                            resource_size(devmem->resource),
 862                            PA_SECTION_SIZE) - align_start;
 863
 864         is_ram = region_intersects(align_start, align_size,
 865                                    IORESOURCE_SYSTEM_RAM,
 866                                    IORES_DESC_NONE);
 867         if (is_ram == REGION_MIXED) {
 868                 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
 869                                 __func__, devmem->resource);
 870                 return -ENXIO;
 871         }
 872         if (is_ram == REGION_INTERSECTS)
 873                 return -ENXIO;
 874
 875         if (devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY)
 876                 devmem->pagemap.type = MEMORY_DEVICE_PUBLIC;
 877         else
 878                 devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
 879
 880         devmem->pagemap.res = devmem->resource;
 881         devmem->pagemap.page_fault = hmm_devmem_fault;
 882         devmem->pagemap.page_free = hmm_devmem_free;
 883         devmem->pagemap.dev = devmem->device;
 884         devmem->pagemap.ref = &devmem->ref;
 885         devmem->pagemap.data = devmem;
 886
 887         mutex_lock(&hmm_devmem_lock);
 888         align_end = align_start + align_size - 1;
 889         for (key = align_start; key <= align_end; key += PA_SECTION_SIZE) {
 890                 struct hmm_devmem *dup;
 891
 892                 rcu_read_lock();
 893                 dup = hmm_devmem_find(key);
 894                 rcu_read_unlock();
 895                 if (dup) {
 896                         dev_err(device, "%s: collides with mapping for %s\n",
 897                                 __func__, dev_name(dup->device));
 898                         mutex_unlock(&hmm_devmem_lock);
 899                         ret = -EBUSY;
 900                         goto error;
 901                 }
 902                 ret = radix_tree_insert(&hmm_devmem_radix,
 903                                         key >> PA_SECTION_SHIFT,
 904                                         devmem);
 905                 if (ret) {
 906                         dev_err(device, "%s: failed: %d\n", __func__, ret);
 907                         mutex_unlock(&hmm_devmem_lock);
 908                         goto error_radix;
 909                 }
 910         }
 911         mutex_unlock(&hmm_devmem_lock);
 912
 913         nid = dev_to_node(device);
 914         if (nid < 0)
 915                 nid = numa_mem_id();
 916
 917         mem_hotplug_begin();
 918         /*
 919          * For device private memory we call add_pages() as we only need to
 920          * allocate and initialize struct page for the device memory. More-
 921          * over the device memory is un-accessible thus we do not want to
 922          * create a linear mapping for the memory like arch_add_memory()
 923          * would do.
 924          *
 925          * For device public memory, which is accesible by the CPU, we do
 926          * want the linear mapping and thus use arch_add_memory().
 927          */
 928         if (devmem->pagemap.type == MEMORY_DEVICE_PUBLIC)
 929                 ret = arch_add_memory(nid, align_start, align_size, false);
 930         else
 931                 ret = add_pages(nid, align_start >> PAGE_SHIFT,
 932                                 align_size >> PAGE_SHIFT, false);
 933         if (ret) {
 934                 mem_hotplug_done();
 935                 goto error_add_memory;
 936         }
 937         move_pfn_range_to_zone(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
 938                                 align_start >> PAGE_SHIFT,
 939                                 align_size >> PAGE_SHIFT);
 940         mem_hotplug_done();
 941
 942         for (pfn = devmem->pfn_first; pfn < devmem->pfn_last; pfn++) {
 943                 struct page *page = pfn_to_page(pfn);
 944
 945                 page->pgmap = &devmem->pagemap;
 946         }
 947         return 0;
 948
 949 error_add_memory:
 950         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
 951 error_radix:
 952         hmm_devmem_radix_release(devmem->resource);
 953 error:
 954         return ret;
 955 }
 956
 957 /*
 958  * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory
 959  *
 960  * @ops: memory event device driver callback (see struct hmm_devmem_ops)
 961  * @device: device struct to bind the resource too
 962  * @size: size in bytes of the device memory to add
 963  * Returns: pointer to new hmm_devmem struct ERR_PTR otherwise
 964  *
 965  * This function first finds an empty range of physical address big enough to
 966  * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which
 967  * in turn allocates struct pages. It does not do anything beyond that; all
 968  * events affecting the memory will go through the various callbacks provided
 969  * by hmm_devmem_ops struct.
 970  *
 971  * Device driver should call this function during device initialization and
 972  * is then responsible of memory management. HMM only provides helpers.
 973  */
 974 struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
 975                                   struct device *device,
 976                                   unsigned long size)
 977 {
 978         struct hmm_devmem *devmem;
 979         resource_size_t addr;
 980         int ret;
 981
 982         static_branch_enable(&device_private_key);
 983
 984         devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL);
 985         if (!devmem)
 986                 return ERR_PTR(-ENOMEM);
 987
 988         init_completion(&devmem->completion);
 989         devmem->pfn_first = -1UL;
 990         devmem->pfn_last = -1UL;
 991         devmem->resource = NULL;
 992         devmem->device = device;
 993         devmem->ops = ops;
 994
 995         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
 996                               0, GFP_KERNEL);
 997         if (ret)
 998                 return ERR_PTR(ret);
 999
1000         ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit, &devmem->ref);
1001         if (ret)
1002                 return ERR_PTR(ret);
1003
1004         size = ALIGN(size, PA_SECTION_SIZE);
1005         addr = min((unsigned long)iomem_resource.end,
1006                    (1UL << MAX_PHYSMEM_BITS) - 1);
1007         addr = addr - size + 1UL;
1008
1009         /*
1010          * FIXME add a new helper to quickly walk resource tree and find free
1011          * range
1012          *
1013          * FIXME what about ioport_resource resource ?
1014          */
1015         for (; addr > size && addr >= iomem_resource.start; addr -= size) {
1016                 ret = region_intersects(addr, size, 0, IORES_DESC_NONE);
1017                 if (ret != REGION_DISJOINT)
1018                         continue;
1019
1020                 devmem->resource = devm_request_mem_region(device, addr, size,
1021                                                            dev_name(device));
1022                 if (!devmem->resource)
1023                         return ERR_PTR(-ENOMEM);
1024                 break;
1025         }
1026         if (!devmem->resource)
1027                 return ERR_PTR(-ERANGE);
1028
1029         devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY;
1030         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1031         devmem->pfn_last = devmem->pfn_first +
1032                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1033
1034         ret = hmm_devmem_pages_create(devmem);
1035         if (ret)
1036                 return ERR_PTR(ret);
1037
1038         ret = devm_add_action_or_reset(device, hmm_devmem_release, devmem);
1039         if (ret)
1040                 return ERR_PTR(ret);
1041
1042         return devmem;
1043 }
1044 EXPORT_SYMBOL_GPL(hmm_devmem_add);
1045
1046 struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops,
1047                                            struct device *device,
1048                                            struct resource *res)
1049 {
1050         struct hmm_devmem *devmem;
1051         int ret;
1052
1053         if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY)
1054                 return ERR_PTR(-EINVAL);
1055
1056         static_branch_enable(&device_private_key);
1057
1058         devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL);
1059         if (!devmem)
1060                 return ERR_PTR(-ENOMEM);
1061
1062         init_completion(&devmem->completion);
1063         devmem->pfn_first = -1UL;
1064         devmem->pfn_last = -1UL;
1065         devmem->resource = res;
1066         devmem->device = device;
1067         devmem->ops = ops;
1068
1069         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1070                               0, GFP_KERNEL);
1071         if (ret)
1072                 return ERR_PTR(ret);
1073
1074         ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit,
1075                         &devmem->ref);
1076         if (ret)
1077                 return ERR_PTR(ret);
1078
1079         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1080         devmem->pfn_last = devmem->pfn_first +
1081                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1082
1083         ret = hmm_devmem_pages_create(devmem);
1084         if (ret)
1085                 return ERR_PTR(ret);
1086
1087         ret = devm_add_action_or_reset(device, hmm_devmem_release, devmem);
1088         if (ret)
1089                 return ERR_PTR(ret);
1090
1091         ret = devm_add_action_or_reset(device, hmm_devmem_ref_kill,
1092                         &devmem->ref);
1093         if (ret)
1094                 return ERR_PTR(ret);
1095
1096         return devmem;
1097 }
1098 EXPORT_SYMBOL_GPL(hmm_devmem_add_resource);
1099
1100 /*
1101  * A device driver that wants to handle multiple devices memory through a
1102  * single fake device can use hmm_device to do so. This is purely a helper
1103  * and it is not needed to make use of any HMM functionality.
1104  */
1105 #define HMM_DEVICE_MAX 256
1106
1107 static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX);
1108 static DEFINE_SPINLOCK(hmm_device_lock);
1109 static struct class *hmm_device_class;
1110 static dev_t hmm_device_devt;
1111
1112 static void hmm_device_release(struct device *device)
1113 {
1114         struct hmm_device *hmm_device;
1115
1116         hmm_device = container_of(device, struct hmm_device, device);
1117         spin_lock(&hmm_device_lock);
1118         clear_bit(hmm_device->minor, hmm_device_mask);
1119         spin_unlock(&hmm_device_lock);
1120
1121         kfree(hmm_device);
1122 }
1123
1124 struct hmm_device *hmm_device_new(void *drvdata)
1125 {
1126         struct hmm_device *hmm_device;
1127
1128         hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL);
1129         if (!hmm_device)
1130                 return ERR_PTR(-ENOMEM);
1131
1132         spin_lock(&hmm_device_lock);
1133         hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX);
1134         if (hmm_device->minor >= HMM_DEVICE_MAX) {
1135                 spin_unlock(&hmm_device_lock);
1136                 kfree(hmm_device);
1137                 return ERR_PTR(-EBUSY);
1138         }
1139         set_bit(hmm_device->minor, hmm_device_mask);
1140         spin_unlock(&hmm_device_lock);
1141
1142         dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor);
1143         hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt),
1144                                         hmm_device->minor);
1145         hmm_device->device.release = hmm_device_release;
1146         dev_set_drvdata(&hmm_device->device, drvdata);
1147         hmm_device->device.class = hmm_device_class;
1148         device_initialize(&hmm_device->device);
1149
1150         return hmm_device;
1151 }
1152 EXPORT_SYMBOL(hmm_device_new);
1153
1154 void hmm_device_put(struct hmm_device *hmm_device)
1155 {
1156         put_device(&hmm_device->device);
1157 }
1158 EXPORT_SYMBOL(hmm_device_put);
1159
1160 static int __init hmm_init(void)
1161 {
1162         int ret;
1163
1164         ret = alloc_chrdev_region(&hmm_device_devt, 0,
1165                                   HMM_DEVICE_MAX,
1166                                   "hmm_device");
1167         if (ret)
1168                 return ret;
1169
1170         hmm_device_class = class_create(THIS_MODULE, "hmm_device");
1171         if (IS_ERR(hmm_device_class)) {
1172                 unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX);
1173                 return PTR_ERR(hmm_device_class);
1174         }
1175         return 0;
1176 }
1177
1178 device_initcall(hmm_init);
1179 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */