lib/test_hmm.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * This is a module to test the HMM (Heterogeneous Memory Management)
   4  * mirror and zone device private memory migration APIs of the kernel.
   5  * Userspace programs can register with the driver to mirror their own address
   6  * space and can use the device to read/write any valid virtual address.
   7  */
   8 #include <linux/init.h>
   9 #include <linux/fs.h>
  10 #include <linux/mm.h>
  11 #include <linux/module.h>
  12 #include <linux/kernel.h>
  13 #include <linux/cdev.h>
  14 #include <linux/device.h>
  15 #include <linux/mutex.h>
  16 #include <linux/rwsem.h>
  17 #include <linux/sched.h>
  18 #include <linux/slab.h>
  19 #include <linux/highmem.h>
  20 #include <linux/delay.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/hmm.h>
  23 #include <linux/vmalloc.h>
  24 #include <linux/swap.h>
  25 #include <linux/swapops.h>
  26 #include <linux/sched/mm.h>
  27 #include <linux/platform_device.h>
  28
  29 #include "test_hmm_uapi.h"
  30
  31 #define DMIRROR_NDEVICES                2
  32 #define DMIRROR_RANGE_FAULT_TIMEOUT     1000
  33 #define DEVMEM_CHUNK_SIZE               (256 * 1024 * 1024U)
  34 #define DEVMEM_CHUNKS_RESERVE           16
  35
  36 static const struct dev_pagemap_ops dmirror_devmem_ops;
  37 static const struct mmu_interval_notifier_ops dmirror_min_ops;
  38 static dev_t dmirror_dev;
  39 static struct page *dmirror_zero_page;
  40
  41 struct dmirror_device;
  42
  43 struct dmirror_bounce {
  44         void                    *ptr;
  45         unsigned long           size;
  46         unsigned long           addr;
  47         unsigned long           cpages;
  48 };
  49
  50 #define DPT_XA_TAG_WRITE 3UL
  51
  52 /*
  53  * Data structure to track address ranges and register for mmu interval
  54  * notifier updates.
  55  */
  56 struct dmirror_interval {
  57         struct mmu_interval_notifier    notifier;
  58         struct dmirror                  *dmirror;
  59 };
  60
  61 /*
  62  * Data attached to the open device file.
  63  * Note that it might be shared after a fork().
  64  */
  65 struct dmirror {
  66         struct dmirror_device           *mdevice;
  67         struct xarray                   pt;
  68         struct mmu_interval_notifier    notifier;
  69         struct mutex                    mutex;
  70 };
  71
  72 /*
  73  * ZONE_DEVICE pages for migration and simulating device memory.
  74  */
  75 struct dmirror_chunk {
  76         struct dev_pagemap      pagemap;
  77         struct dmirror_device   *mdevice;
  78 };
  79
  80 /*
  81  * Per device data.
  82  */
  83 struct dmirror_device {
  84         struct cdev             cdevice;
  85         struct hmm_devmem       *devmem;
  86
  87         unsigned int            devmem_capacity;
  88         unsigned int            devmem_count;
  89         struct dmirror_chunk    **devmem_chunks;
  90         struct mutex            devmem_lock;    /* protects the above */
  91
  92         unsigned long           calloc;
  93         unsigned long           cfree;
  94         struct page             *free_pages;
  95         spinlock_t              lock;           /* protects the above */
  96 };
  97
  98 static struct dmirror_device dmirror_devices[DMIRROR_NDEVICES];
  99
 100 static int dmirror_bounce_init(struct dmirror_bounce *bounce,
 101                                unsigned long addr,
 102                                unsigned long size)
 103 {
 104         bounce->addr = addr;
 105         bounce->size = size;
 106         bounce->cpages = 0;
 107         bounce->ptr = vmalloc(size);
 108         if (!bounce->ptr)
 109                 return -ENOMEM;
 110         return 0;
 111 }
 112
 113 static void dmirror_bounce_fini(struct dmirror_bounce *bounce)
 114 {
 115         vfree(bounce->ptr);
 116 }
 117
 118 static int dmirror_fops_open(struct inode *inode, struct file *filp)
 119 {
 120         struct cdev *cdev = inode->i_cdev;
 121         struct dmirror *dmirror;
 122         int ret;
 123
 124         /* Mirror this process address space */
 125         dmirror = kzalloc(sizeof(*dmirror), GFP_KERNEL);
 126         if (dmirror == NULL)
 127                 return -ENOMEM;
 128
 129         dmirror->mdevice = container_of(cdev, struct dmirror_device, cdevice);
 130         mutex_init(&dmirror->mutex);
 131         xa_init(&dmirror->pt);
 132
 133         ret = mmu_interval_notifier_insert(&dmirror->notifier, current->mm,
 134                                 0, ULONG_MAX & PAGE_MASK, &dmirror_min_ops);
 135         if (ret) {
 136                 kfree(dmirror);
 137                 return ret;
 138         }
 139
 140         filp->private_data = dmirror;
 141         return 0;
 142 }
 143
 144 static int dmirror_fops_release(struct inode *inode, struct file *filp)
 145 {
 146         struct dmirror *dmirror = filp->private_data;
 147
 148         mmu_interval_notifier_remove(&dmirror->notifier);
 149         xa_destroy(&dmirror->pt);
 150         kfree(dmirror);
 151         return 0;
 152 }
 153
 154 static struct dmirror_device *dmirror_page_to_device(struct page *page)
 155
 156 {
 157         return container_of(page->pgmap, struct dmirror_chunk,
 158                             pagemap)->mdevice;
 159 }
 160
 161 static int dmirror_do_fault(struct dmirror *dmirror, struct hmm_range *range)
 162 {
 163         unsigned long *pfns = range->hmm_pfns;
 164         unsigned long pfn;
 165
 166         for (pfn = (range->start >> PAGE_SHIFT);
 167              pfn < (range->end >> PAGE_SHIFT);
 168              pfn++, pfns++) {
 169                 struct page *page;
 170                 void *entry;
 171
 172                 /*
 173                  * Since we asked for hmm_range_fault() to populate pages,
 174                  * it shouldn't return an error entry on success.
 175                  */
 176                 WARN_ON(*pfns & HMM_PFN_ERROR);
 177                 WARN_ON(!(*pfns & HMM_PFN_VALID));
 178
 179                 page = hmm_pfn_to_page(*pfns);
 180                 WARN_ON(!page);
 181
 182                 entry = page;
 183                 if (*pfns & HMM_PFN_WRITE)
 184                         entry = xa_tag_pointer(entry, DPT_XA_TAG_WRITE);
 185                 else if (WARN_ON(range->default_flags & HMM_PFN_WRITE))
 186                         return -EFAULT;
 187                 entry = xa_store(&dmirror->pt, pfn, entry, GFP_ATOMIC);
 188                 if (xa_is_err(entry))
 189                         return xa_err(entry);
 190         }
 191
 192         return 0;
 193 }
 194
 195 static void dmirror_do_update(struct dmirror *dmirror, unsigned long start,
 196                               unsigned long end)
 197 {
 198         unsigned long pfn;
 199         void *entry;
 200
 201         /*
 202          * The XArray doesn't hold references to pages since it relies on
 203          * the mmu notifier to clear page pointers when they become stale.
 204          * Therefore, it is OK to just clear the entry.
 205          */
 206         xa_for_each_range(&dmirror->pt, pfn, entry, start >> PAGE_SHIFT,
 207                           end >> PAGE_SHIFT)
 208                 xa_erase(&dmirror->pt, pfn);
 209 }
 210
 211 static bool dmirror_interval_invalidate(struct mmu_interval_notifier *mni,
 212                                 const struct mmu_notifier_range *range,
 213                                 unsigned long cur_seq)
 214 {
 215         struct dmirror *dmirror = container_of(mni, struct dmirror, notifier);
 216
 217         if (mmu_notifier_range_blockable(range))
 218                 mutex_lock(&dmirror->mutex);
 219         else if (!mutex_trylock(&dmirror->mutex))
 220                 return false;
 221
 222         mmu_interval_set_seq(mni, cur_seq);
 223         dmirror_do_update(dmirror, range->start, range->end);
 224
 225         mutex_unlock(&dmirror->mutex);
 226         return true;
 227 }
 228
 229 static const struct mmu_interval_notifier_ops dmirror_min_ops = {
 230         .invalidate = dmirror_interval_invalidate,
 231 };
 232
 233 static int dmirror_range_fault(struct dmirror *dmirror,
 234                                 struct hmm_range *range)
 235 {
 236         struct mm_struct *mm = dmirror->notifier.mm;
 237         unsigned long timeout =
 238                 jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
 239         int ret;
 240
 241         while (true) {
 242                 if (time_after(jiffies, timeout)) {
 243                         ret = -EBUSY;
 244                         goto out;
 245                 }
 246
 247                 range->notifier_seq = mmu_interval_read_begin(range->notifier);
 248                 mmap_read_lock(mm);
 249                 ret = hmm_range_fault(range);
 250                 mmap_read_unlock(mm);
 251                 if (ret) {
 252                         if (ret == -EBUSY)
 253                                 continue;
 254                         goto out;
 255                 }
 256
 257                 mutex_lock(&dmirror->mutex);
 258                 if (mmu_interval_read_retry(range->notifier,
 259                                             range->notifier_seq)) {
 260                         mutex_unlock(&dmirror->mutex);
 261                         continue;
 262                 }
 263                 break;
 264         }
 265
 266         ret = dmirror_do_fault(dmirror, range);
 267
 268         mutex_unlock(&dmirror->mutex);
 269 out:
 270         return ret;
 271 }
 272
 273 static int dmirror_fault(struct dmirror *dmirror, unsigned long start,
 274                          unsigned long end, bool write)
 275 {
 276         struct mm_struct *mm = dmirror->notifier.mm;
 277         unsigned long addr;
 278         unsigned long pfns[64];
 279         struct hmm_range range = {
 280                 .notifier = &dmirror->notifier,
 281                 .hmm_pfns = pfns,
 282                 .pfn_flags_mask = 0,
 283                 .default_flags =
 284                         HMM_PFN_REQ_FAULT | (write ? HMM_PFN_REQ_WRITE : 0),
 285                 .dev_private_owner = dmirror->mdevice,
 286         };
 287         int ret = 0;
 288
 289         /* Since the mm is for the mirrored process, get a reference first. */
 290         if (!mmget_not_zero(mm))
 291                 return 0;
 292
 293         for (addr = start; addr < end; addr = range.end) {
 294                 range.start = addr;
 295                 range.end = min(addr + (ARRAY_SIZE(pfns) << PAGE_SHIFT), end);
 296
 297                 ret = dmirror_range_fault(dmirror, &range);
 298                 if (ret)
 299                         break;
 300         }
 301
 302         mmput(mm);
 303         return ret;
 304 }
 305
 306 static int dmirror_do_read(struct dmirror *dmirror, unsigned long start,
 307                            unsigned long end, struct dmirror_bounce *bounce)
 308 {
 309         unsigned long pfn;
 310         void *ptr;
 311
 312         ptr = bounce->ptr + ((start - bounce->addr) & PAGE_MASK);
 313
 314         for (pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++) {
 315                 void *entry;
 316                 struct page *page;
 317                 void *tmp;
 318
 319                 entry = xa_load(&dmirror->pt, pfn);
 320                 page = xa_untag_pointer(entry);
 321                 if (!page)
 322                         return -ENOENT;
 323
 324                 tmp = kmap(page);
 325                 memcpy(ptr, tmp, PAGE_SIZE);
 326                 kunmap(page);
 327
 328                 ptr += PAGE_SIZE;
 329                 bounce->cpages++;
 330         }
 331
 332         return 0;
 333 }
 334
 335 static int dmirror_read(struct dmirror *dmirror, struct hmm_dmirror_cmd *cmd)
 336 {
 337         struct dmirror_bounce bounce;
 338         unsigned long start, end;
 339         unsigned long size = cmd->npages << PAGE_SHIFT;
 340         int ret;
 341
 342         start = cmd->addr;
 343         end = start + size;
 344         if (end < start)
 345                 return -EINVAL;
 346
 347         ret = dmirror_bounce_init(&bounce, start, size);
 348         if (ret)
 349                 return ret;
 350
 351         while (1) {
 352                 mutex_lock(&dmirror->mutex);
 353                 ret = dmirror_do_read(dmirror, start, end, &bounce);
 354                 mutex_unlock(&dmirror->mutex);
 355                 if (ret != -ENOENT)
 356                         break;
 357
 358                 start = cmd->addr + (bounce.cpages << PAGE_SHIFT);
 359                 ret = dmirror_fault(dmirror, start, end, false);
 360                 if (ret)
 361                         break;
 362                 cmd->faults++;
 363         }
 364
 365         if (ret == 0) {
 366                 if (copy_to_user(u64_to_user_ptr(cmd->ptr), bounce.ptr,
 367                                  bounce.size))
 368                         ret = -EFAULT;
 369         }
 370         cmd->cpages = bounce.cpages;
 371         dmirror_bounce_fini(&bounce);
 372         return ret;
 373 }
 374
 375 static int dmirror_do_write(struct dmirror *dmirror, unsigned long start,
 376                             unsigned long end, struct dmirror_bounce *bounce)
 377 {
 378         unsigned long pfn;
 379         void *ptr;
 380
 381         ptr = bounce->ptr + ((start - bounce->addr) & PAGE_MASK);
 382
 383         for (pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++) {
 384                 void *entry;
 385                 struct page *page;
 386                 void *tmp;
 387
 388                 entry = xa_load(&dmirror->pt, pfn);
 389                 page = xa_untag_pointer(entry);
 390                 if (!page || xa_pointer_tag(entry) != DPT_XA_TAG_WRITE)
 391                         return -ENOENT;
 392
 393                 tmp = kmap(page);
 394                 memcpy(tmp, ptr, PAGE_SIZE);
 395                 kunmap(page);
 396
 397                 ptr += PAGE_SIZE;
 398                 bounce->cpages++;
 399         }
 400
 401         return 0;
 402 }
 403
 404 static int dmirror_write(struct dmirror *dmirror, struct hmm_dmirror_cmd *cmd)
 405 {
 406         struct dmirror_bounce bounce;
 407         unsigned long start, end;
 408         unsigned long size = cmd->npages << PAGE_SHIFT;
 409         int ret;
 410
 411         start = cmd->addr;
 412         end = start + size;
 413         if (end < start)
 414                 return -EINVAL;
 415
 416         ret = dmirror_bounce_init(&bounce, start, size);
 417         if (ret)
 418                 return ret;
 419         if (copy_from_user(bounce.ptr, u64_to_user_ptr(cmd->ptr),
 420                            bounce.size)) {
 421                 ret = -EFAULT;
 422                 goto fini;
 423         }
 424
 425         while (1) {
 426                 mutex_lock(&dmirror->mutex);
 427                 ret = dmirror_do_write(dmirror, start, end, &bounce);
 428                 mutex_unlock(&dmirror->mutex);
 429                 if (ret != -ENOENT)
 430                         break;
 431
 432                 start = cmd->addr + (bounce.cpages << PAGE_SHIFT);
 433                 ret = dmirror_fault(dmirror, start, end, true);
 434                 if (ret)
 435                         break;
 436                 cmd->faults++;
 437         }
 438
 439 fini:
 440         cmd->cpages = bounce.cpages;
 441         dmirror_bounce_fini(&bounce);
 442         return ret;
 443 }
 444
 445 static bool dmirror_allocate_chunk(struct dmirror_device *mdevice,
 446                                    struct page **ppage)
 447 {
 448         struct dmirror_chunk *devmem;
 449         struct resource *res;
 450         unsigned long pfn;
 451         unsigned long pfn_first;
 452         unsigned long pfn_last;
 453         void *ptr;
 454
 455         mutex_lock(&mdevice->devmem_lock);
 456
 457         if (mdevice->devmem_count == mdevice->devmem_capacity) {
 458                 struct dmirror_chunk **new_chunks;
 459                 unsigned int new_capacity;
 460
 461                 new_capacity = mdevice->devmem_capacity +
 462                                 DEVMEM_CHUNKS_RESERVE;
 463                 new_chunks = krealloc(mdevice->devmem_chunks,
 464                                 sizeof(new_chunks[0]) * new_capacity,
 465                                 GFP_KERNEL);
 466                 if (!new_chunks)
 467                         goto err;
 468                 mdevice->devmem_capacity = new_capacity;
 469                 mdevice->devmem_chunks = new_chunks;
 470         }
 471
 472         res = request_free_mem_region(&iomem_resource, DEVMEM_CHUNK_SIZE,
 473                                         "hmm_dmirror");
 474         if (IS_ERR(res))
 475                 goto err;
 476
 477         devmem = kzalloc(sizeof(*devmem), GFP_KERNEL);
 478         if (!devmem)
 479                 goto err_release;
 480
 481         devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
 482         devmem->pagemap.res = *res;
 483         devmem->pagemap.ops = &dmirror_devmem_ops;
 484         devmem->pagemap.owner = mdevice;
 485
 486         ptr = memremap_pages(&devmem->pagemap, numa_node_id());
 487         if (IS_ERR(ptr))
 488                 goto err_free;
 489
 490         devmem->mdevice = mdevice;
 491         pfn_first = devmem->pagemap.res.start >> PAGE_SHIFT;
 492         pfn_last = pfn_first +
 493                 (resource_size(&devmem->pagemap.res) >> PAGE_SHIFT);
 494         mdevice->devmem_chunks[mdevice->devmem_count++] = devmem;
 495
 496         mutex_unlock(&mdevice->devmem_lock);
 497
 498         pr_info("added new %u MB chunk (total %u chunks, %u MB) PFNs [0x%lx 0x%lx)\n",
 499                 DEVMEM_CHUNK_SIZE / (1024 * 1024),
 500                 mdevice->devmem_count,
 501                 mdevice->devmem_count * (DEVMEM_CHUNK_SIZE / (1024 * 1024)),
 502                 pfn_first, pfn_last);
 503
 504         spin_lock(&mdevice->lock);
 505         for (pfn = pfn_first; pfn < pfn_last; pfn++) {
 506                 struct page *page = pfn_to_page(pfn);
 507
 508                 page->zone_device_data = mdevice->free_pages;
 509                 mdevice->free_pages = page;
 510         }
 511         if (ppage) {
 512                 *ppage = mdevice->free_pages;
 513                 mdevice->free_pages = (*ppage)->zone_device_data;
 514                 mdevice->calloc++;
 515         }
 516         spin_unlock(&mdevice->lock);
 517
 518         return true;
 519
 520 err_free:
 521         kfree(devmem);
 522 err_release:
 523         release_mem_region(devmem->pagemap.res.start,
 524                            resource_size(&devmem->pagemap.res));
 525 err:
 526         mutex_unlock(&mdevice->devmem_lock);
 527         return false;
 528 }
 529
 530 static struct page *dmirror_devmem_alloc_page(struct dmirror_device *mdevice)
 531 {
 532         struct page *dpage = NULL;
 533         struct page *rpage;
 534
 535         /*
 536          * This is a fake device so we alloc real system memory to store
 537          * our device memory.
 538          */
 539         rpage = alloc_page(GFP_HIGHUSER);
 540         if (!rpage)
 541                 return NULL;
 542
 543         spin_lock(&mdevice->lock);
 544
 545         if (mdevice->free_pages) {
 546                 dpage = mdevice->free_pages;
 547                 mdevice->free_pages = dpage->zone_device_data;
 548                 mdevice->calloc++;
 549                 spin_unlock(&mdevice->lock);
 550         } else {
 551                 spin_unlock(&mdevice->lock);
 552                 if (!dmirror_allocate_chunk(mdevice, &dpage))
 553                         goto error;
 554         }
 555
 556         dpage->zone_device_data = rpage;
 557         get_page(dpage);
 558         lock_page(dpage);
 559         return dpage;
 560
 561 error:
 562         __free_page(rpage);
 563         return NULL;
 564 }
 565
 566 static void dmirror_migrate_alloc_and_copy(struct migrate_vma *args,
 567                                            struct dmirror *dmirror)
 568 {
 569         struct dmirror_device *mdevice = dmirror->mdevice;
 570         const unsigned long *src = args->src;
 571         unsigned long *dst = args->dst;
 572         unsigned long addr;
 573
 574         for (addr = args->start; addr < args->end; addr += PAGE_SIZE,
 575                                                    src++, dst++) {
 576                 struct page *spage;
 577                 struct page *dpage;
 578                 struct page *rpage;
 579
 580                 if (!(*src & MIGRATE_PFN_MIGRATE))
 581                         continue;
 582
 583                 /*
 584                  * Note that spage might be NULL which is OK since it is an
 585                  * unallocated pte_none() or read-only zero page.
 586                  */
 587                 spage = migrate_pfn_to_page(*src);
 588
 589                 /*
 590                  * Don't migrate device private pages from our own driver or
 591                  * others. For our own we would do a device private memory copy
 592                  * not a migration and for others, we would need to fault the
 593                  * other device's page into system memory first.
 594                  */
 595                 if (spage && is_zone_device_page(spage))
 596                         continue;
 597
 598                 dpage = dmirror_devmem_alloc_page(mdevice);
 599                 if (!dpage)
 600                         continue;
 601
 602                 rpage = dpage->zone_device_data;
 603                 if (spage)
 604                         copy_highpage(rpage, spage);
 605                 else
 606                         clear_highpage(rpage);
 607
 608                 /*
 609                  * Normally, a device would use the page->zone_device_data to
 610                  * point to the mirror but here we use it to hold the page for
 611                  * the simulated device memory and that page holds the pointer
 612                  * to the mirror.
 613                  */
 614                 rpage->zone_device_data = dmirror;
 615
 616                 *dst = migrate_pfn(page_to_pfn(dpage)) |
 617                             MIGRATE_PFN_LOCKED;
 618                 if ((*src & MIGRATE_PFN_WRITE) ||
 619                     (!spage && args->vma->vm_flags & VM_WRITE))
 620                         *dst |= MIGRATE_PFN_WRITE;
 621         }
 622 }
 623
 624 static int dmirror_migrate_finalize_and_map(struct migrate_vma *args,
 625                                             struct dmirror *dmirror)
 626 {
 627         unsigned long start = args->start;
 628         unsigned long end = args->end;
 629         const unsigned long *src = args->src;
 630         const unsigned long *dst = args->dst;
 631         unsigned long pfn;
 632
 633         /* Map the migrated pages into the device's page tables. */
 634         mutex_lock(&dmirror->mutex);
 635
 636         for (pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++,
 637                                                                 src++, dst++) {
 638                 struct page *dpage;
 639                 void *entry;
 640
 641                 if (!(*src & MIGRATE_PFN_MIGRATE))
 642                         continue;
 643
 644                 dpage = migrate_pfn_to_page(*dst);
 645                 if (!dpage)
 646                         continue;
 647
 648                 /*
 649                  * Store the page that holds the data so the page table
 650                  * doesn't have to deal with ZONE_DEVICE private pages.
 651                  */
 652                 entry = dpage->zone_device_data;
 653                 if (*dst & MIGRATE_PFN_WRITE)
 654                         entry = xa_tag_pointer(entry, DPT_XA_TAG_WRITE);
 655                 entry = xa_store(&dmirror->pt, pfn, entry, GFP_ATOMIC);
 656                 if (xa_is_err(entry)) {
 657                         mutex_unlock(&dmirror->mutex);
 658                         return xa_err(entry);
 659                 }
 660         }
 661
 662         mutex_unlock(&dmirror->mutex);
 663         return 0;
 664 }
 665
 666 static int dmirror_migrate(struct dmirror *dmirror,
 667                            struct hmm_dmirror_cmd *cmd)
 668 {
 669         unsigned long start, end, addr;
 670         unsigned long size = cmd->npages << PAGE_SHIFT;
 671         struct mm_struct *mm = dmirror->notifier.mm;
 672         struct vm_area_struct *vma;
 673         unsigned long src_pfns[64];
 674         unsigned long dst_pfns[64];
 675         struct dmirror_bounce bounce;
 676         struct migrate_vma args;
 677         unsigned long next;
 678         int ret;
 679
 680         start = cmd->addr;
 681         end = start + size;
 682         if (end < start)
 683                 return -EINVAL;
 684
 685         /* Since the mm is for the mirrored process, get a reference first. */
 686         if (!mmget_not_zero(mm))
 687                 return -EINVAL;
 688
 689         mmap_read_lock(mm);
 690         for (addr = start; addr < end; addr = next) {
 691                 vma = find_vma(mm, addr);
 692                 if (!vma || addr < vma->vm_start ||
 693                     !(vma->vm_flags & VM_READ)) {
 694                         ret = -EINVAL;
 695                         goto out;
 696                 }
 697                 next = min(end, addr + (ARRAY_SIZE(src_pfns) << PAGE_SHIFT));
 698                 if (next > vma->vm_end)
 699                         next = vma->vm_end;
 700
 701                 args.vma = vma;
 702                 args.src = src_pfns;
 703                 args.dst = dst_pfns;
 704                 args.start = addr;
 705                 args.end = next;
 706                 args.src_owner = NULL;
 707                 ret = migrate_vma_setup(&args);
 708                 if (ret)
 709                         goto out;
 710
 711                 dmirror_migrate_alloc_and_copy(&args, dmirror);
 712                 migrate_vma_pages(&args);
 713                 dmirror_migrate_finalize_and_map(&args, dmirror);
 714                 migrate_vma_finalize(&args);
 715         }
 716         mmap_read_unlock(mm);
 717         mmput(mm);
 718
 719         /* Return the migrated data for verification. */
 720         ret = dmirror_bounce_init(&bounce, start, size);
 721         if (ret)
 722                 return ret;
 723         mutex_lock(&dmirror->mutex);
 724         ret = dmirror_do_read(dmirror, start, end, &bounce);
 725         mutex_unlock(&dmirror->mutex);
 726         if (ret == 0) {
 727                 if (copy_to_user(u64_to_user_ptr(cmd->ptr), bounce.ptr,
 728                                  bounce.size))
 729                         ret = -EFAULT;
 730         }
 731         cmd->cpages = bounce.cpages;
 732         dmirror_bounce_fini(&bounce);
 733         return ret;
 734
 735 out:
 736         mmap_read_unlock(mm);
 737         mmput(mm);
 738         return ret;
 739 }
 740
 741 static void dmirror_mkentry(struct dmirror *dmirror, struct hmm_range *range,
 742                             unsigned char *perm, unsigned long entry)
 743 {
 744         struct page *page;
 745
 746         if (entry & HMM_PFN_ERROR) {
 747                 *perm = HMM_DMIRROR_PROT_ERROR;
 748                 return;
 749         }
 750         if (!(entry & HMM_PFN_VALID)) {
 751                 *perm = HMM_DMIRROR_PROT_NONE;
 752                 return;
 753         }
 754
 755         page = hmm_pfn_to_page(entry);
 756         if (is_device_private_page(page)) {
 757                 /* Is the page migrated to this device or some other? */
 758                 if (dmirror->mdevice == dmirror_page_to_device(page))
 759                         *perm = HMM_DMIRROR_PROT_DEV_PRIVATE_LOCAL;
 760                 else
 761                         *perm = HMM_DMIRROR_PROT_DEV_PRIVATE_REMOTE;
 762         } else if (is_zero_pfn(page_to_pfn(page)))
 763                 *perm = HMM_DMIRROR_PROT_ZERO;
 764         else
 765                 *perm = HMM_DMIRROR_PROT_NONE;
 766         if (entry & HMM_PFN_WRITE)
 767                 *perm |= HMM_DMIRROR_PROT_WRITE;
 768         else
 769                 *perm |= HMM_DMIRROR_PROT_READ;
 770 }
 771
 772 static bool dmirror_snapshot_invalidate(struct mmu_interval_notifier *mni,
 773                                 const struct mmu_notifier_range *range,
 774                                 unsigned long cur_seq)
 775 {
 776         struct dmirror_interval *dmi =
 777                 container_of(mni, struct dmirror_interval, notifier);
 778         struct dmirror *dmirror = dmi->dmirror;
 779
 780         if (mmu_notifier_range_blockable(range))
 781                 mutex_lock(&dmirror->mutex);
 782         else if (!mutex_trylock(&dmirror->mutex))
 783                 return false;
 784
 785         /*
 786          * Snapshots only need to set the sequence number since any
 787          * invalidation in the interval invalidates the whole snapshot.
 788          */
 789         mmu_interval_set_seq(mni, cur_seq);
 790
 791         mutex_unlock(&dmirror->mutex);
 792         return true;
 793 }
 794
 795 static const struct mmu_interval_notifier_ops dmirror_mrn_ops = {
 796         .invalidate = dmirror_snapshot_invalidate,
 797 };
 798
 799 static int dmirror_range_snapshot(struct dmirror *dmirror,
 800                                   struct hmm_range *range,
 801                                   unsigned char *perm)
 802 {
 803         struct mm_struct *mm = dmirror->notifier.mm;
 804         struct dmirror_interval notifier;
 805         unsigned long timeout =
 806                 jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
 807         unsigned long i;
 808         unsigned long n;
 809         int ret = 0;
 810
 811         notifier.dmirror = dmirror;
 812         range->notifier = &notifier.notifier;
 813
 814         ret = mmu_interval_notifier_insert(range->notifier, mm,
 815                         range->start, range->end - range->start,
 816                         &dmirror_mrn_ops);
 817         if (ret)
 818                 return ret;
 819
 820         while (true) {
 821                 if (time_after(jiffies, timeout)) {
 822                         ret = -EBUSY;
 823                         goto out;
 824                 }
 825
 826                 range->notifier_seq = mmu_interval_read_begin(range->notifier);
 827
 828                 mmap_read_lock(mm);
 829                 ret = hmm_range_fault(range);
 830                 mmap_read_unlock(mm);
 831                 if (ret) {
 832                         if (ret == -EBUSY)
 833                                 continue;
 834                         goto out;
 835                 }
 836
 837                 mutex_lock(&dmirror->mutex);
 838                 if (mmu_interval_read_retry(range->notifier,
 839                                             range->notifier_seq)) {
 840                         mutex_unlock(&dmirror->mutex);
 841                         continue;
 842                 }
 843                 break;
 844         }
 845
 846         n = (range->end - range->start) >> PAGE_SHIFT;
 847         for (i = 0; i < n; i++)
 848                 dmirror_mkentry(dmirror, range, perm + i, range->hmm_pfns[i]);
 849
 850         mutex_unlock(&dmirror->mutex);
 851 out:
 852         mmu_interval_notifier_remove(range->notifier);
 853         return ret;
 854 }
 855
 856 static int dmirror_snapshot(struct dmirror *dmirror,
 857                             struct hmm_dmirror_cmd *cmd)
 858 {
 859         struct mm_struct *mm = dmirror->notifier.mm;
 860         unsigned long start, end;
 861         unsigned long size = cmd->npages << PAGE_SHIFT;
 862         unsigned long addr;
 863         unsigned long next;
 864         unsigned long pfns[64];
 865         unsigned char perm[64];
 866         char __user *uptr;
 867         struct hmm_range range = {
 868                 .hmm_pfns = pfns,
 869                 .dev_private_owner = dmirror->mdevice,
 870         };
 871         int ret = 0;
 872
 873         start = cmd->addr;
 874         end = start + size;
 875         if (end < start)
 876                 return -EINVAL;
 877
 878         /* Since the mm is for the mirrored process, get a reference first. */
 879         if (!mmget_not_zero(mm))
 880                 return -EINVAL;
 881
 882         /*
 883          * Register a temporary notifier to detect invalidations even if it
 884          * overlaps with other mmu_interval_notifiers.
 885          */
 886         uptr = u64_to_user_ptr(cmd->ptr);
 887         for (addr = start; addr < end; addr = next) {
 888                 unsigned long n;
 889
 890                 next = min(addr + (ARRAY_SIZE(pfns) << PAGE_SHIFT), end);
 891                 range.start = addr;
 892                 range.end = next;
 893
 894                 ret = dmirror_range_snapshot(dmirror, &range, perm);
 895                 if (ret)
 896                         break;
 897
 898                 n = (range.end - range.start) >> PAGE_SHIFT;
 899                 if (copy_to_user(uptr, perm, n)) {
 900                         ret = -EFAULT;
 901                         break;
 902                 }
 903
 904                 cmd->cpages += n;
 905                 uptr += n;
 906         }
 907         mmput(mm);
 908
 909         return ret;
 910 }
 911
 912 static long dmirror_fops_unlocked_ioctl(struct file *filp,
 913                                         unsigned int command,
 914                                         unsigned long arg)
 915 {
 916         void __user *uarg = (void __user *)arg;
 917         struct hmm_dmirror_cmd cmd;
 918         struct dmirror *dmirror;
 919         int ret;
 920
 921         dmirror = filp->private_data;
 922         if (!dmirror)
 923                 return -EINVAL;
 924
 925         if (copy_from_user(&cmd, uarg, sizeof(cmd)))
 926                 return -EFAULT;
 927
 928         if (cmd.addr & ~PAGE_MASK)
 929                 return -EINVAL;
 930         if (cmd.addr >= (cmd.addr + (cmd.npages << PAGE_SHIFT)))
 931                 return -EINVAL;
 932
 933         cmd.cpages = 0;
 934         cmd.faults = 0;
 935
 936         switch (command) {
 937         case HMM_DMIRROR_READ:
 938                 ret = dmirror_read(dmirror, &cmd);
 939                 break;
 940
 941         case HMM_DMIRROR_WRITE:
 942                 ret = dmirror_write(dmirror, &cmd);
 943                 break;
 944
 945         case HMM_DMIRROR_MIGRATE:
 946                 ret = dmirror_migrate(dmirror, &cmd);
 947                 break;
 948
 949         case HMM_DMIRROR_SNAPSHOT:
 950                 ret = dmirror_snapshot(dmirror, &cmd);
 951                 break;
 952
 953         default:
 954                 return -EINVAL;
 955         }
 956         if (ret)
 957                 return ret;
 958
 959         if (copy_to_user(uarg, &cmd, sizeof(cmd)))
 960                 return -EFAULT;
 961
 962         return 0;
 963 }
 964
 965 static const struct file_operations dmirror_fops = {
 966         .open           = dmirror_fops_open,
 967         .release        = dmirror_fops_release,
 968         .unlocked_ioctl = dmirror_fops_unlocked_ioctl,
 969         .llseek         = default_llseek,
 970         .owner          = THIS_MODULE,
 971 };
 972
 973 static void dmirror_devmem_free(struct page *page)
 974 {
 975         struct page *rpage = page->zone_device_data;
 976         struct dmirror_device *mdevice;
 977
 978         if (rpage)
 979                 __free_page(rpage);
 980
 981         mdevice = dmirror_page_to_device(page);
 982
 983         spin_lock(&mdevice->lock);
 984         mdevice->cfree++;
 985         page->zone_device_data = mdevice->free_pages;
 986         mdevice->free_pages = page;
 987         spin_unlock(&mdevice->lock);
 988 }
 989
 990 static vm_fault_t dmirror_devmem_fault_alloc_and_copy(struct migrate_vma *args,
 991                                                 struct dmirror_device *mdevice)
 992 {
 993         const unsigned long *src = args->src;
 994         unsigned long *dst = args->dst;
 995         unsigned long start = args->start;
 996         unsigned long end = args->end;
 997         unsigned long addr;
 998
 999         for (addr = start; addr < end; addr += PAGE_SIZE,
1000                                        src++, dst++) {
1001                 struct page *dpage, *spage;
1002
1003                 spage = migrate_pfn_to_page(*src);
1004                 if (!spage || !(*src & MIGRATE_PFN_MIGRATE))
1005                         continue;
1006                 spage = spage->zone_device_data;
1007
1008                 dpage = alloc_page_vma(GFP_HIGHUSER_MOVABLE, args->vma, addr);
1009                 if (!dpage)
1010                         continue;
1011
1012                 lock_page(dpage);
1013                 copy_highpage(dpage, spage);
1014                 *dst = migrate_pfn(page_to_pfn(dpage)) | MIGRATE_PFN_LOCKED;
1015                 if (*src & MIGRATE_PFN_WRITE)
1016                         *dst |= MIGRATE_PFN_WRITE;
1017         }
1018         return 0;
1019 }
1020
1021 static void dmirror_devmem_fault_finalize_and_map(struct migrate_vma *args,
1022                                                   struct dmirror *dmirror)
1023 {
1024         /* Invalidate the device's page table mapping. */
1025         mutex_lock(&dmirror->mutex);
1026         dmirror_do_update(dmirror, args->start, args->end);
1027         mutex_unlock(&dmirror->mutex);
1028 }
1029
1030 static vm_fault_t dmirror_devmem_fault(struct vm_fault *vmf)
1031 {
1032         struct migrate_vma args;
1033         unsigned long src_pfns;
1034         unsigned long dst_pfns;
1035         struct page *rpage;
1036         struct dmirror *dmirror;
1037         vm_fault_t ret;
1038
1039         /*
1040          * Normally, a device would use the page->zone_device_data to point to
1041          * the mirror but here we use it to hold the page for the simulated
1042          * device memory and that page holds the pointer to the mirror.
1043          */
1044         rpage = vmf->page->zone_device_data;
1045         dmirror = rpage->zone_device_data;
1046
1047         /* FIXME demonstrate how we can adjust migrate range */
1048         args.vma = vmf->vma;
1049         args.start = vmf->address;
1050         args.end = args.start + PAGE_SIZE;
1051         args.src = &src_pfns;
1052         args.dst = &dst_pfns;
1053         args.src_owner = dmirror->mdevice;
1054
1055         if (migrate_vma_setup(&args))
1056                 return VM_FAULT_SIGBUS;
1057
1058         ret = dmirror_devmem_fault_alloc_and_copy(&args, dmirror->mdevice);
1059         if (ret)
1060                 return ret;
1061         migrate_vma_pages(&args);
1062         dmirror_devmem_fault_finalize_and_map(&args, dmirror);
1063         migrate_vma_finalize(&args);
1064         return 0;
1065 }
1066
1067 static const struct dev_pagemap_ops dmirror_devmem_ops = {
1068         .page_free      = dmirror_devmem_free,
1069         .migrate_to_ram = dmirror_devmem_fault,
1070 };
1071
1072 static int dmirror_device_init(struct dmirror_device *mdevice, int id)
1073 {
1074         dev_t dev;
1075         int ret;
1076
1077         dev = MKDEV(MAJOR(dmirror_dev), id);
1078         mutex_init(&mdevice->devmem_lock);
1079         spin_lock_init(&mdevice->lock);
1080
1081         cdev_init(&mdevice->cdevice, &dmirror_fops);
1082         mdevice->cdevice.owner = THIS_MODULE;
1083         ret = cdev_add(&mdevice->cdevice, dev, 1);
1084         if (ret)
1085                 return ret;
1086
1087         /* Build a list of free ZONE_DEVICE private struct pages */
1088         dmirror_allocate_chunk(mdevice, NULL);
1089
1090         return 0;
1091 }
1092
1093 static void dmirror_device_remove(struct dmirror_device *mdevice)
1094 {
1095         unsigned int i;
1096
1097         if (mdevice->devmem_chunks) {
1098                 for (i = 0; i < mdevice->devmem_count; i++) {
1099                         struct dmirror_chunk *devmem =
1100                                 mdevice->devmem_chunks[i];
1101
1102                         memunmap_pages(&devmem->pagemap);
1103                         release_mem_region(devmem->pagemap.res.start,
1104                                            resource_size(&devmem->pagemap.res));
1105                         kfree(devmem);
1106                 }
1107                 kfree(mdevice->devmem_chunks);
1108         }
1109
1110         cdev_del(&mdevice->cdevice);
1111 }
1112
1113 static int __init hmm_dmirror_init(void)
1114 {
1115         int ret;
1116         int id;
1117
1118         ret = alloc_chrdev_region(&dmirror_dev, 0, DMIRROR_NDEVICES,
1119                                   "HMM_DMIRROR");
1120         if (ret)
1121                 goto err_unreg;
1122
1123         for (id = 0; id < DMIRROR_NDEVICES; id++) {
1124                 ret = dmirror_device_init(dmirror_devices + id, id);
1125                 if (ret)
1126                         goto err_chrdev;
1127         }
1128
1129         /*
1130          * Allocate a zero page to simulate a reserved page of device private
1131          * memory which is always zero. The zero_pfn page isn't used just to
1132          * make the code here simpler (i.e., we need a struct page for it).
1133          */
1134         dmirror_zero_page = alloc_page(GFP_HIGHUSER | __GFP_ZERO);
1135         if (!dmirror_zero_page) {
1136                 ret = -ENOMEM;
1137                 goto err_chrdev;
1138         }
1139
1140         pr_info("HMM test module loaded. This is only for testing HMM.\n");
1141         return 0;
1142
1143 err_chrdev:
1144         while (--id >= 0)
1145                 dmirror_device_remove(dmirror_devices + id);
1146         unregister_chrdev_region(dmirror_dev, DMIRROR_NDEVICES);
1147 err_unreg:
1148         return ret;
1149 }
1150
1151 static void __exit hmm_dmirror_exit(void)
1152 {
1153         int id;
1154
1155         if (dmirror_zero_page)
1156                 __free_page(dmirror_zero_page);
1157         for (id = 0; id < DMIRROR_NDEVICES; id++)
1158                 dmirror_device_remove(dmirror_devices + id);
1159         unregister_chrdev_region(dmirror_dev, DMIRROR_NDEVICES);
1160 }
1161
1162 module_init(hmm_dmirror_init);
1163 module_exit(hmm_dmirror_exit);
1164 MODULE_LICENSE("GPL");