drivers/vfio/vfio_iommu_spapr_tce.c

   1 /*
   2  * VFIO: IOMMU DMA mapping support for TCE on POWER
   3  *
   4  * Copyright (C) 2013 IBM Corp.  All rights reserved.
   5  *     Author: Alexey Kardashevskiy <aik@ozlabs.ru>
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License version 2 as
   9  * published by the Free Software Foundation.
  10  *
  11  * Derived from original vfio_iommu_type1.c:
  12  * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
  13  *     Author: Alex Williamson <alex.williamson@redhat.com>
  14  */
  15
  16 #include <linux/module.h>
  17 #include <linux/pci.h>
  18 #include <linux/slab.h>
  19 #include <linux/uaccess.h>
  20 #include <linux/err.h>
  21 #include <linux/vfio.h>
  22 #include <linux/vmalloc.h>
  23 #include <asm/iommu.h>
  24 #include <asm/tce.h>
  25 #include <asm/mmu_context.h>
  26
  27 #define DRIVER_VERSION  "0.1"
  28 #define DRIVER_AUTHOR   "aik@ozlabs.ru"
  29 #define DRIVER_DESC     "VFIO IOMMU SPAPR TCE"
  30
  31 static void tce_iommu_detach_group(void *iommu_data,
  32                 struct iommu_group *iommu_group);
  33
  34 static long try_increment_locked_vm(struct mm_struct *mm, long npages)
  35 {
  36         long ret = 0, locked, lock_limit;
  37
  38         if (WARN_ON_ONCE(!mm))
  39                 return -EPERM;
  40
  41         if (!npages)
  42                 return 0;
  43
  44         down_write(&mm->mmap_sem);
  45         locked = mm->locked_vm + npages;
  46         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
  47         if (locked > lock_limit && !capable(CAP_IPC_LOCK))
  48                 ret = -ENOMEM;
  49         else
  50                 mm->locked_vm += npages;
  51
  52         pr_debug("[%d] RLIMIT_MEMLOCK +%ld %ld/%ld%s\n", current->pid,
  53                         npages << PAGE_SHIFT,
  54                         mm->locked_vm << PAGE_SHIFT,
  55                         rlimit(RLIMIT_MEMLOCK),
  56                         ret ? " - exceeded" : "");
  57
  58         up_write(&mm->mmap_sem);
  59
  60         return ret;
  61 }
  62
  63 static void decrement_locked_vm(struct mm_struct *mm, long npages)
  64 {
  65         if (!mm || !npages)
  66                 return;
  67
  68         down_write(&mm->mmap_sem);
  69         if (WARN_ON_ONCE(npages > mm->locked_vm))
  70                 npages = mm->locked_vm;
  71         mm->locked_vm -= npages;
  72         pr_debug("[%d] RLIMIT_MEMLOCK -%ld %ld/%ld\n", current->pid,
  73                         npages << PAGE_SHIFT,
  74                         mm->locked_vm << PAGE_SHIFT,
  75                         rlimit(RLIMIT_MEMLOCK));
  76         up_write(&mm->mmap_sem);
  77 }
  78
  79 /*
  80  * VFIO IOMMU fd for SPAPR_TCE IOMMU implementation
  81  *
  82  * This code handles mapping and unmapping of user data buffers
  83  * into DMA'ble space using the IOMMU
  84  */
  85
  86 struct tce_iommu_group {
  87         struct list_head next;
  88         struct iommu_group *grp;
  89 };
  90
  91 /*
  92  * A container needs to remember which preregistered region  it has
  93  * referenced to do proper cleanup at the userspace process exit.
  94  */
  95 struct tce_iommu_prereg {
  96         struct list_head next;
  97         struct mm_iommu_table_group_mem_t *mem;
  98 };
  99
 100 /*
 101  * The container descriptor supports only a single group per container.
 102  * Required by the API as the container is not supplied with the IOMMU group
 103  * at the moment of initialization.
 104  */
 105 struct tce_container {
 106         struct mutex lock;
 107         bool enabled;
 108         bool v2;
 109         bool def_window_pending;
 110         unsigned long locked_pages;
 111         struct mm_struct *mm;
 112         struct iommu_table *tables[IOMMU_TABLE_GROUP_MAX_TABLES];
 113         struct list_head group_list;
 114         struct list_head prereg_list;
 115 };
 116
 117 static long tce_iommu_mm_set(struct tce_container *container)
 118 {
 119         if (container->mm) {
 120                 if (container->mm == current->mm)
 121                         return 0;
 122                 return -EPERM;
 123         }
 124         BUG_ON(!current->mm);
 125         container->mm = current->mm;
 126         atomic_inc(&container->mm->mm_count);
 127
 128         return 0;
 129 }
 130
 131 static long tce_iommu_prereg_free(struct tce_container *container,
 132                 struct tce_iommu_prereg *tcemem)
 133 {
 134         long ret;
 135
 136         ret = mm_iommu_put(container->mm, tcemem->mem);
 137         if (ret)
 138                 return ret;
 139
 140         list_del(&tcemem->next);
 141         kfree(tcemem);
 142
 143         return 0;
 144 }
 145
 146 static long tce_iommu_unregister_pages(struct tce_container *container,
 147                 __u64 vaddr, __u64 size)
 148 {
 149         struct mm_iommu_table_group_mem_t *mem;
 150         struct tce_iommu_prereg *tcemem;
 151         bool found = false;
 152
 153         if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK))
 154                 return -EINVAL;
 155
 156         mem = mm_iommu_find(container->mm, vaddr, size >> PAGE_SHIFT);
 157         if (!mem)
 158                 return -ENOENT;
 159
 160         list_for_each_entry(tcemem, &container->prereg_list, next) {
 161                 if (tcemem->mem == mem) {
 162                         found = true;
 163                         break;
 164                 }
 165         }
 166
 167         if (!found)
 168                 return -ENOENT;
 169
 170         return tce_iommu_prereg_free(container, tcemem);
 171 }
 172
 173 static long tce_iommu_register_pages(struct tce_container *container,
 174                 __u64 vaddr, __u64 size)
 175 {
 176         long ret = 0;
 177         struct mm_iommu_table_group_mem_t *mem = NULL;
 178         struct tce_iommu_prereg *tcemem;
 179         unsigned long entries = size >> PAGE_SHIFT;
 180
 181         if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK) ||
 182                         ((vaddr + size) < vaddr))
 183                 return -EINVAL;
 184
 185         mem = mm_iommu_find(container->mm, vaddr, entries);
 186         if (mem) {
 187                 list_for_each_entry(tcemem, &container->prereg_list, next) {
 188                         if (tcemem->mem == mem)
 189                                 return -EBUSY;
 190                 }
 191         }
 192
 193         ret = mm_iommu_get(container->mm, vaddr, entries, &mem);
 194         if (ret)
 195                 return ret;
 196
 197         tcemem = kzalloc(sizeof(*tcemem), GFP_KERNEL);
 198         if (!tcemem) {
 199                 mm_iommu_put(container->mm, mem);
 200                 return -ENOMEM;
 201         }
 202
 203         tcemem->mem = mem;
 204         list_add(&tcemem->next, &container->prereg_list);
 205
 206         container->enabled = true;
 207
 208         return 0;
 209 }
 210
 211 static long tce_iommu_userspace_view_alloc(struct iommu_table *tbl,
 212                 struct mm_struct *mm)
 213 {
 214         unsigned long cb = _ALIGN_UP(sizeof(tbl->it_userspace[0]) *
 215                         tbl->it_size, PAGE_SIZE);
 216         unsigned long *uas;
 217         long ret;
 218
 219         BUG_ON(tbl->it_userspace);
 220
 221         ret = try_increment_locked_vm(mm, cb >> PAGE_SHIFT);
 222         if (ret)
 223                 return ret;
 224
 225         uas = vzalloc(cb);
 226         if (!uas) {
 227                 decrement_locked_vm(mm, cb >> PAGE_SHIFT);
 228                 return -ENOMEM;
 229         }
 230         tbl->it_userspace = uas;
 231
 232         return 0;
 233 }
 234
 235 static void tce_iommu_userspace_view_free(struct iommu_table *tbl,
 236                 struct mm_struct *mm)
 237 {
 238         unsigned long cb = _ALIGN_UP(sizeof(tbl->it_userspace[0]) *
 239                         tbl->it_size, PAGE_SIZE);
 240
 241         if (!tbl->it_userspace)
 242                 return;
 243
 244         vfree(tbl->it_userspace);
 245         tbl->it_userspace = NULL;
 246         decrement_locked_vm(mm, cb >> PAGE_SHIFT);
 247 }
 248
 249 static bool tce_page_is_contained(struct page *page, unsigned page_shift)
 250 {
 251         /*
 252          * Check that the TCE table granularity is not bigger than the size of
 253          * a page we just found. Otherwise the hardware can get access to
 254          * a bigger memory chunk that it should.
 255          */
 256         return (PAGE_SHIFT + compound_order(compound_head(page))) >= page_shift;
 257 }
 258
 259 static inline bool tce_groups_attached(struct tce_container *container)
 260 {
 261         return !list_empty(&container->group_list);
 262 }
 263
 264 static long tce_iommu_find_table(struct tce_container *container,
 265                 phys_addr_t ioba, struct iommu_table **ptbl)
 266 {
 267         long i;
 268
 269         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
 270                 struct iommu_table *tbl = container->tables[i];
 271
 272                 if (tbl) {
 273                         unsigned long entry = ioba >> tbl->it_page_shift;
 274                         unsigned long start = tbl->it_offset;
 275                         unsigned long end = start + tbl->it_size;
 276
 277                         if ((start <= entry) && (entry < end)) {
 278                                 *ptbl = tbl;
 279                                 return i;
 280                         }
 281                 }
 282         }
 283
 284         return -1;
 285 }
 286
 287 static int tce_iommu_find_free_table(struct tce_container *container)
 288 {
 289         int i;
 290
 291         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
 292                 if (!container->tables[i])
 293                         return i;
 294         }
 295
 296         return -ENOSPC;
 297 }
 298
 299 static int tce_iommu_enable(struct tce_container *container)
 300 {
 301         int ret = 0;
 302         unsigned long locked;
 303         struct iommu_table_group *table_group;
 304         struct tce_iommu_group *tcegrp;
 305
 306         if (container->enabled)
 307                 return -EBUSY;
 308
 309         /*
 310          * When userspace pages are mapped into the IOMMU, they are effectively
 311          * locked memory, so, theoretically, we need to update the accounting
 312          * of locked pages on each map and unmap.  For powerpc, the map unmap
 313          * paths can be very hot, though, and the accounting would kill
 314          * performance, especially since it would be difficult to impossible
 315          * to handle the accounting in real mode only.
 316          *
 317          * To address that, rather than precisely accounting every page, we
 318          * instead account for a worst case on locked memory when the iommu is
 319          * enabled and disabled.  The worst case upper bound on locked memory
 320          * is the size of the whole iommu window, which is usually relatively
 321          * small (compared to total memory sizes) on POWER hardware.
 322          *
 323          * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits,
 324          * that would effectively kill the guest at random points, much better
 325          * enforcing the limit based on the max that the guest can map.
 326          *
 327          * Unfortunately at the moment it counts whole tables, no matter how
 328          * much memory the guest has. I.e. for 4GB guest and 4 IOMMU groups
 329          * each with 2GB DMA window, 8GB will be counted here. The reason for
 330          * this is that we cannot tell here the amount of RAM used by the guest
 331          * as this information is only available from KVM and VFIO is
 332          * KVM agnostic.
 333          *
 334          * So we do not allow enabling a container without a group attached
 335          * as there is no way to know how much we should increment
 336          * the locked_vm counter.
 337          */
 338         if (!tce_groups_attached(container))
 339                 return -ENODEV;
 340
 341         tcegrp = list_first_entry(&container->group_list,
 342                         struct tce_iommu_group, next);
 343         table_group = iommu_group_get_iommudata(tcegrp->grp);
 344         if (!table_group)
 345                 return -ENODEV;
 346
 347         if (!table_group->tce32_size)
 348                 return -EPERM;
 349
 350         ret = tce_iommu_mm_set(container);
 351         if (ret)
 352                 return ret;
 353
 354         locked = table_group->tce32_size >> PAGE_SHIFT;
 355         ret = try_increment_locked_vm(container->mm, locked);
 356         if (ret)
 357                 return ret;
 358
 359         container->locked_pages = locked;
 360
 361         container->enabled = true;
 362
 363         return ret;
 364 }
 365
 366 static void tce_iommu_disable(struct tce_container *container)
 367 {
 368         if (!container->enabled)
 369                 return;
 370
 371         container->enabled = false;
 372
 373         BUG_ON(!container->mm);
 374         decrement_locked_vm(container->mm, container->locked_pages);
 375 }
 376
 377 static void *tce_iommu_open(unsigned long arg)
 378 {
 379         struct tce_container *container;
 380
 381         if ((arg != VFIO_SPAPR_TCE_IOMMU) && (arg != VFIO_SPAPR_TCE_v2_IOMMU)) {
 382                 pr_err("tce_vfio: Wrong IOMMU type\n");
 383                 return ERR_PTR(-EINVAL);
 384         }
 385
 386         container = kzalloc(sizeof(*container), GFP_KERNEL);
 387         if (!container)
 388                 return ERR_PTR(-ENOMEM);
 389
 390         mutex_init(&container->lock);
 391         INIT_LIST_HEAD_RCU(&container->group_list);
 392         INIT_LIST_HEAD_RCU(&container->prereg_list);
 393
 394         container->v2 = arg == VFIO_SPAPR_TCE_v2_IOMMU;
 395
 396         return container;
 397 }
 398
 399 static int tce_iommu_clear(struct tce_container *container,
 400                 struct iommu_table *tbl,
 401                 unsigned long entry, unsigned long pages);
 402 static void tce_iommu_free_table(struct tce_container *container,
 403                 struct iommu_table *tbl);
 404
 405 static void tce_iommu_release(void *iommu_data)
 406 {
 407         struct tce_container *container = iommu_data;
 408         struct tce_iommu_group *tcegrp;
 409         struct tce_iommu_prereg *tcemem, *tmtmp;
 410         long i;
 411
 412         while (tce_groups_attached(container)) {
 413                 tcegrp = list_first_entry(&container->group_list,
 414                                 struct tce_iommu_group, next);
 415                 tce_iommu_detach_group(iommu_data, tcegrp->grp);
 416         }
 417
 418         /*
 419          * If VFIO created a table, it was not disposed
 420          * by tce_iommu_detach_group() so do it now.
 421          */
 422         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
 423                 struct iommu_table *tbl = container->tables[i];
 424
 425                 if (!tbl)
 426                         continue;
 427
 428                 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
 429                 tce_iommu_free_table(container, tbl);
 430         }
 431
 432         list_for_each_entry_safe(tcemem, tmtmp, &container->prereg_list, next)
 433                 WARN_ON(tce_iommu_prereg_free(container, tcemem));
 434
 435         tce_iommu_disable(container);
 436         if (container->mm)
 437                 mmdrop(container->mm);
 438         mutex_destroy(&container->lock);
 439
 440         kfree(container);
 441 }
 442
 443 static void tce_iommu_unuse_page(struct tce_container *container,
 444                 unsigned long hpa)
 445 {
 446         struct page *page;
 447
 448         page = pfn_to_page(hpa >> PAGE_SHIFT);
 449         put_page(page);
 450 }
 451
 452 static int tce_iommu_prereg_ua_to_hpa(struct tce_container *container,
 453                 unsigned long tce, unsigned long size,
 454                 unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)
 455 {
 456         long ret = 0;
 457         struct mm_iommu_table_group_mem_t *mem;
 458
 459         mem = mm_iommu_lookup(container->mm, tce, size);
 460         if (!mem)
 461                 return -EINVAL;
 462
 463         ret = mm_iommu_ua_to_hpa(mem, tce, phpa);
 464         if (ret)
 465                 return -EINVAL;
 466
 467         *pmem = mem;
 468
 469         return 0;
 470 }
 471
 472 static void tce_iommu_unuse_page_v2(struct tce_container *container,
 473                 struct iommu_table *tbl, unsigned long entry)
 474 {
 475         struct mm_iommu_table_group_mem_t *mem = NULL;
 476         int ret;
 477         unsigned long hpa = 0;
 478         unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry);
 479
 480         if (!pua)
 481                 return;
 482
 483         ret = tce_iommu_prereg_ua_to_hpa(container, *pua, IOMMU_PAGE_SIZE(tbl),
 484                         &hpa, &mem);
 485         if (ret)
 486                 pr_debug("%s: tce %lx at #%lx was not cached, ret=%d\n",
 487                                 __func__, *pua, entry, ret);
 488         if (mem)
 489                 mm_iommu_mapped_dec(mem);
 490
 491         *pua = 0;
 492 }
 493
 494 static int tce_iommu_clear(struct tce_container *container,
 495                 struct iommu_table *tbl,
 496                 unsigned long entry, unsigned long pages)
 497 {
 498         unsigned long oldhpa;
 499         long ret;
 500         enum dma_data_direction direction;
 501
 502         for ( ; pages; --pages, ++entry) {
 503                 direction = DMA_NONE;
 504                 oldhpa = 0;
 505                 ret = iommu_tce_xchg(tbl, entry, &oldhpa, &direction);
 506                 if (ret)
 507                         continue;
 508
 509                 if (direction == DMA_NONE)
 510                         continue;
 511
 512                 if (container->v2) {
 513                         tce_iommu_unuse_page_v2(container, tbl, entry);
 514                         continue;
 515                 }
 516
 517                 tce_iommu_unuse_page(container, oldhpa);
 518         }
 519
 520         return 0;
 521 }
 522
 523 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
 524 {
 525         struct page *page = NULL;
 526         enum dma_data_direction direction = iommu_tce_direction(tce);
 527
 528         if (get_user_pages_fast(tce & PAGE_MASK, 1,
 529                         direction != DMA_TO_DEVICE, &page) != 1)
 530                 return -EFAULT;
 531
 532         *hpa = __pa((unsigned long) page_address(page));
 533
 534         return 0;
 535 }
 536
 537 static long tce_iommu_build(struct tce_container *container,
 538                 struct iommu_table *tbl,
 539                 unsigned long entry, unsigned long tce, unsigned long pages,
 540                 enum dma_data_direction direction)
 541 {
 542         long i, ret = 0;
 543         struct page *page;
 544         unsigned long hpa;
 545         enum dma_data_direction dirtmp;
 546
 547         for (i = 0; i < pages; ++i) {
 548                 unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
 549
 550                 ret = tce_iommu_use_page(tce, &hpa);
 551                 if (ret)
 552                         break;
 553
 554                 page = pfn_to_page(hpa >> PAGE_SHIFT);
 555                 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
 556                         ret = -EPERM;
 557                         break;
 558                 }
 559
 560                 hpa |= offset;
 561                 dirtmp = direction;
 562                 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
 563                 if (ret) {
 564                         tce_iommu_unuse_page(container, hpa);
 565                         pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
 566                                         __func__, entry << tbl->it_page_shift,
 567                                         tce, ret);
 568                         break;
 569                 }
 570
 571                 if (dirtmp != DMA_NONE)
 572                         tce_iommu_unuse_page(container, hpa);
 573
 574                 tce += IOMMU_PAGE_SIZE(tbl);
 575         }
 576
 577         if (ret)
 578                 tce_iommu_clear(container, tbl, entry, i);
 579
 580         return ret;
 581 }
 582
 583 static long tce_iommu_build_v2(struct tce_container *container,
 584                 struct iommu_table *tbl,
 585                 unsigned long entry, unsigned long tce, unsigned long pages,
 586                 enum dma_data_direction direction)
 587 {
 588         long i, ret = 0;
 589         struct page *page;
 590         unsigned long hpa;
 591         enum dma_data_direction dirtmp;
 592
 593         if (!tbl->it_userspace) {
 594                 ret = tce_iommu_userspace_view_alloc(tbl, container->mm);
 595                 if (ret)
 596                         return ret;
 597         }
 598
 599         for (i = 0; i < pages; ++i) {
 600                 struct mm_iommu_table_group_mem_t *mem = NULL;
 601                 unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl,
 602                                 entry + i);
 603
 604                 ret = tce_iommu_prereg_ua_to_hpa(container,
 605                                 tce, IOMMU_PAGE_SIZE(tbl), &hpa, &mem);
 606                 if (ret)
 607                         break;
 608
 609                 page = pfn_to_page(hpa >> PAGE_SHIFT);
 610                 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
 611                         ret = -EPERM;
 612                         break;
 613                 }
 614
 615                 /* Preserve offset within IOMMU page */
 616                 hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
 617                 dirtmp = direction;
 618
 619                 /* The registered region is being unregistered */
 620                 if (mm_iommu_mapped_inc(mem))
 621                         break;
 622
 623                 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
 624                 if (ret) {
 625                         /* dirtmp cannot be DMA_NONE here */
 626                         tce_iommu_unuse_page_v2(container, tbl, entry + i);
 627                         pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
 628                                         __func__, entry << tbl->it_page_shift,
 629                                         tce, ret);
 630                         break;
 631                 }
 632
 633                 if (dirtmp != DMA_NONE)
 634                         tce_iommu_unuse_page_v2(container, tbl, entry + i);
 635
 636                 *pua = tce;
 637
 638                 tce += IOMMU_PAGE_SIZE(tbl);
 639         }
 640
 641         if (ret)
 642                 tce_iommu_clear(container, tbl, entry, i);
 643
 644         return ret;
 645 }
 646
 647 static long tce_iommu_create_table(struct tce_container *container,
 648                         struct iommu_table_group *table_group,
 649                         int num,
 650                         __u32 page_shift,
 651                         __u64 window_size,
 652                         __u32 levels,
 653                         struct iommu_table **ptbl)
 654 {
 655         long ret, table_size;
 656
 657         table_size = table_group->ops->get_table_size(page_shift, window_size,
 658                         levels);
 659         if (!table_size)
 660                 return -EINVAL;
 661
 662         ret = try_increment_locked_vm(container->mm, table_size >> PAGE_SHIFT);
 663         if (ret)
 664                 return ret;
 665
 666         ret = table_group->ops->create_table(table_group, num,
 667                         page_shift, window_size, levels, ptbl);
 668
 669         WARN_ON(!ret && !(*ptbl)->it_ops->free);
 670         WARN_ON(!ret && ((*ptbl)->it_allocated_size != table_size));
 671
 672         return ret;
 673 }
 674
 675 static void tce_iommu_free_table(struct tce_container *container,
 676                 struct iommu_table *tbl)
 677 {
 678         unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
 679
 680         tce_iommu_userspace_view_free(tbl, container->mm);
 681         tbl->it_ops->free(tbl);
 682         decrement_locked_vm(container->mm, pages);
 683 }
 684
 685 static long tce_iommu_create_window(struct tce_container *container,
 686                 __u32 page_shift, __u64 window_size, __u32 levels,
 687                 __u64 *start_addr)
 688 {
 689         struct tce_iommu_group *tcegrp;
 690         struct iommu_table_group *table_group;
 691         struct iommu_table *tbl = NULL;
 692         long ret, num;
 693
 694         num = tce_iommu_find_free_table(container);
 695         if (num < 0)
 696                 return num;
 697
 698         /* Get the first group for ops::create_table */
 699         tcegrp = list_first_entry(&container->group_list,
 700                         struct tce_iommu_group, next);
 701         table_group = iommu_group_get_iommudata(tcegrp->grp);
 702         if (!table_group)
 703                 return -EFAULT;
 704
 705         if (!(table_group->pgsizes & (1ULL << page_shift)))
 706                 return -EINVAL;
 707
 708         if (!table_group->ops->set_window || !table_group->ops->unset_window ||
 709                         !table_group->ops->get_table_size ||
 710                         !table_group->ops->create_table)
 711                 return -EPERM;
 712
 713         /* Create TCE table */
 714         ret = tce_iommu_create_table(container, table_group, num,
 715                         page_shift, window_size, levels, &tbl);
 716         if (ret)
 717                 return ret;
 718
 719         BUG_ON(!tbl->it_ops->free);
 720
 721         /*
 722          * Program the table to every group.
 723          * Groups have been tested for compatibility at the attach time.
 724          */
 725         list_for_each_entry(tcegrp, &container->group_list, next) {
 726                 table_group = iommu_group_get_iommudata(tcegrp->grp);
 727
 728                 ret = table_group->ops->set_window(table_group, num, tbl);
 729                 if (ret)
 730                         goto unset_exit;
 731         }
 732
 733         container->tables[num] = tbl;
 734
 735         /* Return start address assigned by platform in create_table() */
 736         *start_addr = tbl->it_offset << tbl->it_page_shift;
 737
 738         return 0;
 739
 740 unset_exit:
 741         list_for_each_entry(tcegrp, &container->group_list, next) {
 742                 table_group = iommu_group_get_iommudata(tcegrp->grp);
 743                 table_group->ops->unset_window(table_group, num);
 744         }
 745         tce_iommu_free_table(container, tbl);
 746
 747         return ret;
 748 }
 749
 750 static long tce_iommu_remove_window(struct tce_container *container,
 751                 __u64 start_addr)
 752 {
 753         struct iommu_table_group *table_group = NULL;
 754         struct iommu_table *tbl;
 755         struct tce_iommu_group *tcegrp;
 756         int num;
 757
 758         num = tce_iommu_find_table(container, start_addr, &tbl);
 759         if (num < 0)
 760                 return -EINVAL;
 761
 762         BUG_ON(!tbl->it_size);
 763
 764         /* Detach groups from IOMMUs */
 765         list_for_each_entry(tcegrp, &container->group_list, next) {
 766                 table_group = iommu_group_get_iommudata(tcegrp->grp);
 767
 768                 /*
 769                  * SPAPR TCE IOMMU exposes the default DMA window to
 770                  * the guest via dma32_window_start/size of
 771                  * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
 772                  * the userspace to remove this window, some do not so
 773                  * here we check for the platform capability.
 774                  */
 775                 if (!table_group->ops || !table_group->ops->unset_window)
 776                         return -EPERM;
 777
 778                 table_group->ops->unset_window(table_group, num);
 779         }
 780
 781         /* Free table */
 782         tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
 783         tce_iommu_free_table(container, tbl);
 784         container->tables[num] = NULL;
 785
 786         return 0;
 787 }
 788
 789 static long tce_iommu_create_default_window(struct tce_container *container)
 790 {
 791         long ret;
 792         __u64 start_addr = 0;
 793         struct tce_iommu_group *tcegrp;
 794         struct iommu_table_group *table_group;
 795
 796         if (!container->def_window_pending)
 797                 return 0;
 798
 799         if (!tce_groups_attached(container))
 800                 return -ENODEV;
 801
 802         tcegrp = list_first_entry(&container->group_list,
 803                         struct tce_iommu_group, next);
 804         table_group = iommu_group_get_iommudata(tcegrp->grp);
 805         if (!table_group)
 806                 return -ENODEV;
 807
 808         ret = tce_iommu_create_window(container, IOMMU_PAGE_SHIFT_4K,
 809                         table_group->tce32_size, 1, &start_addr);
 810         WARN_ON_ONCE(!ret && start_addr);
 811
 812         if (!ret)
 813                 container->def_window_pending = false;
 814
 815         return ret;
 816 }
 817
 818 static long tce_iommu_ioctl(void *iommu_data,
 819                                  unsigned int cmd, unsigned long arg)
 820 {
 821         struct tce_container *container = iommu_data;
 822         unsigned long minsz, ddwsz;
 823         long ret;
 824
 825         switch (cmd) {
 826         case VFIO_CHECK_EXTENSION:
 827                 switch (arg) {
 828                 case VFIO_SPAPR_TCE_IOMMU:
 829                 case VFIO_SPAPR_TCE_v2_IOMMU:
 830                         ret = 1;
 831                         break;
 832                 default:
 833                         ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
 834                         break;
 835                 }
 836
 837                 return (ret < 0) ? 0 : ret;
 838         }
 839
 840         /*
 841          * Sanity check to prevent one userspace from manipulating
 842          * another userspace mm.
 843          */
 844         BUG_ON(!container);
 845         if (container->mm && container->mm != current->mm)
 846                 return -EPERM;
 847
 848         switch (cmd) {
 849         case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
 850                 struct vfio_iommu_spapr_tce_info info;
 851                 struct tce_iommu_group *tcegrp;
 852                 struct iommu_table_group *table_group;
 853
 854                 if (!tce_groups_attached(container))
 855                         return -ENXIO;
 856
 857                 tcegrp = list_first_entry(&container->group_list,
 858                                 struct tce_iommu_group, next);
 859                 table_group = iommu_group_get_iommudata(tcegrp->grp);
 860
 861                 if (!table_group)
 862                         return -ENXIO;
 863
 864                 minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
 865                                 dma32_window_size);
 866
 867                 if (copy_from_user(&info, (void __user *)arg, minsz))
 868                         return -EFAULT;
 869
 870                 if (info.argsz < minsz)
 871                         return -EINVAL;
 872
 873                 info.dma32_window_start = table_group->tce32_start;
 874                 info.dma32_window_size = table_group->tce32_size;
 875                 info.flags = 0;
 876                 memset(&info.ddw, 0, sizeof(info.ddw));
 877
 878                 if (table_group->max_dynamic_windows_supported &&
 879                                 container->v2) {
 880                         info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
 881                         info.ddw.pgsizes = table_group->pgsizes;
 882                         info.ddw.max_dynamic_windows_supported =
 883                                 table_group->max_dynamic_windows_supported;
 884                         info.ddw.levels = table_group->max_levels;
 885                 }
 886
 887                 ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
 888
 889                 if (info.argsz >= ddwsz)
 890                         minsz = ddwsz;
 891
 892                 if (copy_to_user((void __user *)arg, &info, minsz))
 893                         return -EFAULT;
 894
 895                 return 0;
 896         }
 897         case VFIO_IOMMU_MAP_DMA: {
 898                 struct vfio_iommu_type1_dma_map param;
 899                 struct iommu_table *tbl = NULL;
 900                 long num;
 901                 enum dma_data_direction direction;
 902
 903                 if (!container->enabled)
 904                         return -EPERM;
 905
 906                 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
 907
 908                 if (copy_from_user(&param, (void __user *)arg, minsz))
 909                         return -EFAULT;
 910
 911                 if (param.argsz < minsz)
 912                         return -EINVAL;
 913
 914                 if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
 915                                 VFIO_DMA_MAP_FLAG_WRITE))
 916                         return -EINVAL;
 917
 918                 ret = tce_iommu_create_default_window(container);
 919                 if (ret)
 920                         return ret;
 921
 922                 num = tce_iommu_find_table(container, param.iova, &tbl);
 923                 if (num < 0)
 924                         return -ENXIO;
 925
 926                 if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
 927                                 (param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
 928                         return -EINVAL;
 929
 930                 /* iova is checked by the IOMMU API */
 931                 if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
 932                         if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
 933                                 direction = DMA_BIDIRECTIONAL;
 934                         else
 935                                 direction = DMA_TO_DEVICE;
 936                 } else {
 937                         if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
 938                                 direction = DMA_FROM_DEVICE;
 939                         else
 940                                 return -EINVAL;
 941                 }
 942
 943                 ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
 944                 if (ret)
 945                         return ret;
 946
 947                 if (container->v2)
 948                         ret = tce_iommu_build_v2(container, tbl,
 949                                         param.iova >> tbl->it_page_shift,
 950                                         param.vaddr,
 951                                         param.size >> tbl->it_page_shift,
 952                                         direction);
 953                 else
 954                         ret = tce_iommu_build(container, tbl,
 955                                         param.iova >> tbl->it_page_shift,
 956                                         param.vaddr,
 957                                         param.size >> tbl->it_page_shift,
 958                                         direction);
 959
 960                 iommu_flush_tce(tbl);
 961
 962                 return ret;
 963         }
 964         case VFIO_IOMMU_UNMAP_DMA: {
 965                 struct vfio_iommu_type1_dma_unmap param;
 966                 struct iommu_table *tbl = NULL;
 967                 long num;
 968
 969                 if (!container->enabled)
 970                         return -EPERM;
 971
 972                 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
 973                                 size);
 974
 975                 if (copy_from_user(&param, (void __user *)arg, minsz))
 976                         return -EFAULT;
 977
 978                 if (param.argsz < minsz)
 979                         return -EINVAL;
 980
 981                 /* No flag is supported now */
 982                 if (param.flags)
 983                         return -EINVAL;
 984
 985                 ret = tce_iommu_create_default_window(container);
 986                 if (ret)
 987                         return ret;
 988
 989                 num = tce_iommu_find_table(container, param.iova, &tbl);
 990                 if (num < 0)
 991                         return -ENXIO;
 992
 993                 if (param.size & ~IOMMU_PAGE_MASK(tbl))
 994                         return -EINVAL;
 995
 996                 ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
 997                                 param.size >> tbl->it_page_shift);
 998                 if (ret)
 999                         return ret;
1000
1001                 ret = tce_iommu_clear(container, tbl,
1002                                 param.iova >> tbl->it_page_shift,
1003                                 param.size >> tbl->it_page_shift);
1004                 iommu_flush_tce(tbl);
1005
1006                 return ret;
1007         }
1008         case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
1009                 struct vfio_iommu_spapr_register_memory param;
1010
1011                 if (!container->v2)
1012                         break;
1013
1014                 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1015                                 size);
1016
1017                 ret = tce_iommu_mm_set(container);
1018                 if (ret)
1019                         return ret;
1020
1021                 if (copy_from_user(&param, (void __user *)arg, minsz))
1022                         return -EFAULT;
1023
1024                 if (param.argsz < minsz)
1025                         return -EINVAL;
1026
1027                 /* No flag is supported now */
1028                 if (param.flags)
1029                         return -EINVAL;
1030
1031                 mutex_lock(&container->lock);
1032                 ret = tce_iommu_register_pages(container, param.vaddr,
1033                                 param.size);
1034                 mutex_unlock(&container->lock);
1035
1036                 return ret;
1037         }
1038         case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
1039                 struct vfio_iommu_spapr_register_memory param;
1040
1041                 if (!container->v2)
1042                         break;
1043
1044                 if (!container->mm)
1045                         return -EPERM;
1046
1047                 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1048                                 size);
1049
1050                 if (copy_from_user(&param, (void __user *)arg, minsz))
1051                         return -EFAULT;
1052
1053                 if (param.argsz < minsz)
1054                         return -EINVAL;
1055
1056                 /* No flag is supported now */
1057                 if (param.flags)
1058                         return -EINVAL;
1059
1060                 mutex_lock(&container->lock);
1061                 ret = tce_iommu_unregister_pages(container, param.vaddr,
1062                                 param.size);
1063                 mutex_unlock(&container->lock);
1064
1065                 return ret;
1066         }
1067         case VFIO_IOMMU_ENABLE:
1068                 if (container->v2)
1069                         break;
1070
1071                 mutex_lock(&container->lock);
1072                 ret = tce_iommu_enable(container);
1073                 mutex_unlock(&container->lock);
1074                 return ret;
1075
1076
1077         case VFIO_IOMMU_DISABLE:
1078                 if (container->v2)
1079                         break;
1080
1081                 mutex_lock(&container->lock);
1082                 tce_iommu_disable(container);
1083                 mutex_unlock(&container->lock);
1084                 return 0;
1085
1086         case VFIO_EEH_PE_OP: {
1087                 struct tce_iommu_group *tcegrp;
1088
1089                 ret = 0;
1090                 list_for_each_entry(tcegrp, &container->group_list, next) {
1091                         ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
1092                                         cmd, arg);
1093                         if (ret)
1094                                 return ret;
1095                 }
1096                 return ret;
1097         }
1098
1099         case VFIO_IOMMU_SPAPR_TCE_CREATE: {
1100                 struct vfio_iommu_spapr_tce_create create;
1101
1102                 if (!container->v2)
1103                         break;
1104
1105                 ret = tce_iommu_mm_set(container);
1106                 if (ret)
1107                         return ret;
1108
1109                 if (!tce_groups_attached(container))
1110                         return -ENXIO;
1111
1112                 minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
1113                                 start_addr);
1114
1115                 if (copy_from_user(&create, (void __user *)arg, minsz))
1116                         return -EFAULT;
1117
1118                 if (create.argsz < minsz)
1119                         return -EINVAL;
1120
1121                 if (create.flags)
1122                         return -EINVAL;
1123
1124                 mutex_lock(&container->lock);
1125
1126                 ret = tce_iommu_create_default_window(container);
1127                 if (!ret)
1128                         ret = tce_iommu_create_window(container,
1129                                         create.page_shift,
1130                                         create.window_size, create.levels,
1131                                         &create.start_addr);
1132
1133                 mutex_unlock(&container->lock);
1134
1135                 if (!ret && copy_to_user((void __user *)arg, &create, minsz))
1136                         ret = -EFAULT;
1137
1138                 return ret;
1139         }
1140         case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1141                 struct vfio_iommu_spapr_tce_remove remove;
1142
1143                 if (!container->v2)
1144                         break;
1145
1146                 ret = tce_iommu_mm_set(container);
1147                 if (ret)
1148                         return ret;
1149
1150                 if (!tce_groups_attached(container))
1151                         return -ENXIO;
1152
1153                 minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1154                                 start_addr);
1155
1156                 if (copy_from_user(&remove, (void __user *)arg, minsz))
1157                         return -EFAULT;
1158
1159                 if (remove.argsz < minsz)
1160                         return -EINVAL;
1161
1162                 if (remove.flags)
1163                         return -EINVAL;
1164
1165                 if (container->def_window_pending && !remove.start_addr) {
1166                         container->def_window_pending = false;
1167                         return 0;
1168                 }
1169
1170                 mutex_lock(&container->lock);
1171
1172                 ret = tce_iommu_remove_window(container, remove.start_addr);
1173
1174                 mutex_unlock(&container->lock);
1175
1176                 return ret;
1177         }
1178         }
1179
1180         return -ENOTTY;
1181 }
1182
1183 static void tce_iommu_release_ownership(struct tce_container *container,
1184                 struct iommu_table_group *table_group)
1185 {
1186         int i;
1187
1188         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1189                 struct iommu_table *tbl = container->tables[i];
1190
1191                 if (!tbl)
1192                         continue;
1193
1194                 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1195                 tce_iommu_userspace_view_free(tbl, container->mm);
1196                 if (tbl->it_map)
1197                         iommu_release_ownership(tbl);
1198
1199                 container->tables[i] = NULL;
1200         }
1201 }
1202
1203 static int tce_iommu_take_ownership(struct tce_container *container,
1204                 struct iommu_table_group *table_group)
1205 {
1206         int i, j, rc = 0;
1207
1208         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1209                 struct iommu_table *tbl = table_group->tables[i];
1210
1211                 if (!tbl || !tbl->it_map)
1212                         continue;
1213
1214                 rc = iommu_take_ownership(tbl);
1215                 if (rc) {
1216                         for (j = 0; j < i; ++j)
1217                                 iommu_release_ownership(
1218                                                 table_group->tables[j]);
1219
1220                         return rc;
1221                 }
1222         }
1223
1224         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1225                 container->tables[i] = table_group->tables[i];
1226
1227         return 0;
1228 }
1229
1230 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1231                 struct iommu_table_group *table_group)
1232 {
1233         long i;
1234
1235         if (!table_group->ops->unset_window) {
1236                 WARN_ON_ONCE(1);
1237                 return;
1238         }
1239
1240         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1241                 table_group->ops->unset_window(table_group, i);
1242
1243         table_group->ops->release_ownership(table_group);
1244 }
1245
1246 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1247                 struct iommu_table_group *table_group)
1248 {
1249         long i, ret = 0;
1250
1251         if (!table_group->ops->create_table || !table_group->ops->set_window ||
1252                         !table_group->ops->release_ownership) {
1253                 WARN_ON_ONCE(1);
1254                 return -EFAULT;
1255         }
1256
1257         table_group->ops->take_ownership(table_group);
1258
1259         /* Set all windows to the new group */
1260         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1261                 struct iommu_table *tbl = container->tables[i];
1262
1263                 if (!tbl)
1264                         continue;
1265
1266                 ret = table_group->ops->set_window(table_group, i, tbl);
1267                 if (ret)
1268                         goto release_exit;
1269         }
1270
1271         return 0;
1272
1273 release_exit:
1274         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1275                 table_group->ops->unset_window(table_group, i);
1276
1277         table_group->ops->release_ownership(table_group);
1278
1279         return ret;
1280 }
1281
1282 static int tce_iommu_attach_group(void *iommu_data,
1283                 struct iommu_group *iommu_group)
1284 {
1285         int ret;
1286         struct tce_container *container = iommu_data;
1287         struct iommu_table_group *table_group;
1288         struct tce_iommu_group *tcegrp = NULL;
1289
1290         mutex_lock(&container->lock);
1291
1292         /* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1293                         iommu_group_id(iommu_group), iommu_group); */
1294         table_group = iommu_group_get_iommudata(iommu_group);
1295         if (!table_group) {
1296                 ret = -ENODEV;
1297                 goto unlock_exit;
1298         }
1299
1300         if (tce_groups_attached(container) && (!table_group->ops ||
1301                         !table_group->ops->take_ownership ||
1302                         !table_group->ops->release_ownership)) {
1303                 ret = -EBUSY;
1304                 goto unlock_exit;
1305         }
1306
1307         /* Check if new group has the same iommu_ops (i.e. compatible) */
1308         list_for_each_entry(tcegrp, &container->group_list, next) {
1309                 struct iommu_table_group *table_group_tmp;
1310
1311                 if (tcegrp->grp == iommu_group) {
1312                         pr_warn("tce_vfio: Group %d is already attached\n",
1313                                         iommu_group_id(iommu_group));
1314                         ret = -EBUSY;
1315                         goto unlock_exit;
1316                 }
1317                 table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1318                 if (table_group_tmp->ops->create_table !=
1319                                 table_group->ops->create_table) {
1320                         pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1321                                         iommu_group_id(iommu_group),
1322                                         iommu_group_id(tcegrp->grp));
1323                         ret = -EPERM;
1324                         goto unlock_exit;
1325                 }
1326         }
1327
1328         tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1329         if (!tcegrp) {
1330                 ret = -ENOMEM;
1331                 goto unlock_exit;
1332         }
1333
1334         if (!table_group->ops || !table_group->ops->take_ownership ||
1335                         !table_group->ops->release_ownership) {
1336                 if (container->v2) {
1337                         ret = -EPERM;
1338                         goto unlock_exit;
1339                 }
1340                 ret = tce_iommu_take_ownership(container, table_group);
1341         } else {
1342                 if (!container->v2) {
1343                         ret = -EPERM;
1344                         goto unlock_exit;
1345                 }
1346                 ret = tce_iommu_take_ownership_ddw(container, table_group);
1347                 if (!tce_groups_attached(container) && !container->tables[0])
1348                         container->def_window_pending = true;
1349         }
1350
1351         if (!ret) {
1352                 tcegrp->grp = iommu_group;
1353                 list_add(&tcegrp->next, &container->group_list);
1354         }
1355
1356 unlock_exit:
1357         if (ret && tcegrp)
1358                 kfree(tcegrp);
1359
1360         mutex_unlock(&container->lock);
1361
1362         return ret;
1363 }
1364
1365 static void tce_iommu_detach_group(void *iommu_data,
1366                 struct iommu_group *iommu_group)
1367 {
1368         struct tce_container *container = iommu_data;
1369         struct iommu_table_group *table_group;
1370         bool found = false;
1371         struct tce_iommu_group *tcegrp;
1372
1373         mutex_lock(&container->lock);
1374
1375         list_for_each_entry(tcegrp, &container->group_list, next) {
1376                 if (tcegrp->grp == iommu_group) {
1377                         found = true;
1378                         break;
1379                 }
1380         }
1381
1382         if (!found) {
1383                 pr_warn("tce_vfio: detaching unattached group #%u\n",
1384                                 iommu_group_id(iommu_group));
1385                 goto unlock_exit;
1386         }
1387
1388         list_del(&tcegrp->next);
1389         kfree(tcegrp);
1390
1391         table_group = iommu_group_get_iommudata(iommu_group);
1392         BUG_ON(!table_group);
1393
1394         if (!table_group->ops || !table_group->ops->release_ownership)
1395                 tce_iommu_release_ownership(container, table_group);
1396         else
1397                 tce_iommu_release_ownership_ddw(container, table_group);
1398
1399 unlock_exit:
1400         mutex_unlock(&container->lock);
1401 }
1402
1403 const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1404         .name           = "iommu-vfio-powerpc",
1405         .owner          = THIS_MODULE,
1406         .open           = tce_iommu_open,
1407         .release        = tce_iommu_release,
1408         .ioctl          = tce_iommu_ioctl,
1409         .attach_group   = tce_iommu_attach_group,
1410         .detach_group   = tce_iommu_detach_group,
1411 };
1412
1413 static int __init tce_iommu_init(void)
1414 {
1415         return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1416 }
1417
1418 static void __exit tce_iommu_cleanup(void)
1419 {
1420         vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1421 }
1422
1423 module_init(tce_iommu_init);
1424 module_exit(tce_iommu_cleanup);
1425
1426 MODULE_VERSION(DRIVER_VERSION);
1427 MODULE_LICENSE("GPL v2");
1428 MODULE_AUTHOR(DRIVER_AUTHOR);
1429 MODULE_DESCRIPTION(DRIVER_DESC);
1430