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(long npages)
  35 {
  36         long ret = 0, locked, lock_limit;
  37
  38         if (!current || !current->mm)
  39                 return -ESRCH; /* process exited */
  40
  41         if (!npages)
  42                 return 0;
  43
  44         down_write(&current->mm->mmap_sem);
  45         locked = current->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                 current->mm->locked_vm += npages;
  51
  52         pr_debug("[%d] RLIMIT_MEMLOCK +%ld %ld/%ld%s\n", current->pid,
  53                         npages << PAGE_SHIFT,
  54                         current->mm->locked_vm << PAGE_SHIFT,
  55                         rlimit(RLIMIT_MEMLOCK),
  56                         ret ? " - exceeded" : "");
  57
  58         up_write(&current->mm->mmap_sem);
  59
  60         return ret;
  61 }
  62
  63 static void decrement_locked_vm(long npages)
  64 {
  65         if (!current || !current->mm || !npages)
  66                 return; /* process exited */
  67
  68         down_write(&current->mm->mmap_sem);
  69         if (WARN_ON_ONCE(npages > current->mm->locked_vm))
  70                 npages = current->mm->locked_vm;
  71         current->mm->locked_vm -= npages;
  72         pr_debug("[%d] RLIMIT_MEMLOCK -%ld %ld/%ld\n", current->pid,
  73                         npages << PAGE_SHIFT,
  74                         current->mm->locked_vm << PAGE_SHIFT,
  75                         rlimit(RLIMIT_MEMLOCK));
  76         up_write(&current->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  * The container descriptor supports only a single group per container.
  93  * Required by the API as the container is not supplied with the IOMMU group
  94  * at the moment of initialization.
  95  */
  96 struct tce_container {
  97         struct mutex lock;
  98         bool enabled;
  99         bool v2;
 100         unsigned long locked_pages;
 101         struct iommu_table *tables[IOMMU_TABLE_GROUP_MAX_TABLES];
 102         struct list_head group_list;
 103 };
 104
 105 static long tce_iommu_unregister_pages(struct tce_container *container,
 106                 __u64 vaddr, __u64 size)
 107 {
 108         struct mm_iommu_table_group_mem_t *mem;
 109
 110         if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK))
 111                 return -EINVAL;
 112
 113         mem = mm_iommu_find(vaddr, size >> PAGE_SHIFT);
 114         if (!mem)
 115                 return -ENOENT;
 116
 117         return mm_iommu_put(mem);
 118 }
 119
 120 static long tce_iommu_register_pages(struct tce_container *container,
 121                 __u64 vaddr, __u64 size)
 122 {
 123         long ret = 0;
 124         struct mm_iommu_table_group_mem_t *mem = NULL;
 125         unsigned long entries = size >> PAGE_SHIFT;
 126
 127         if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK) ||
 128                         ((vaddr + size) < vaddr))
 129                 return -EINVAL;
 130
 131         ret = mm_iommu_get(vaddr, entries, &mem);
 132         if (ret)
 133                 return ret;
 134
 135         container->enabled = true;
 136
 137         return 0;
 138 }
 139
 140 static long tce_iommu_userspace_view_alloc(struct iommu_table *tbl)
 141 {
 142         unsigned long cb = _ALIGN_UP(sizeof(tbl->it_userspace[0]) *
 143                         tbl->it_size, PAGE_SIZE);
 144         unsigned long *uas;
 145         long ret;
 146
 147         BUG_ON(tbl->it_userspace);
 148
 149         ret = try_increment_locked_vm(cb >> PAGE_SHIFT);
 150         if (ret)
 151                 return ret;
 152
 153         uas = vzalloc(cb);
 154         if (!uas) {
 155                 decrement_locked_vm(cb >> PAGE_SHIFT);
 156                 return -ENOMEM;
 157         }
 158         tbl->it_userspace = uas;
 159
 160         return 0;
 161 }
 162
 163 static void tce_iommu_userspace_view_free(struct iommu_table *tbl)
 164 {
 165         unsigned long cb = _ALIGN_UP(sizeof(tbl->it_userspace[0]) *
 166                         tbl->it_size, PAGE_SIZE);
 167
 168         if (!tbl->it_userspace)
 169                 return;
 170
 171         vfree(tbl->it_userspace);
 172         tbl->it_userspace = NULL;
 173         decrement_locked_vm(cb >> PAGE_SHIFT);
 174 }
 175
 176 static bool tce_page_is_contained(struct page *page, unsigned page_shift)
 177 {
 178         /*
 179          * Check that the TCE table granularity is not bigger than the size of
 180          * a page we just found. Otherwise the hardware can get access to
 181          * a bigger memory chunk that it should.
 182          */
 183         return (PAGE_SHIFT + compound_order(compound_head(page))) >= page_shift;
 184 }
 185
 186 static inline bool tce_groups_attached(struct tce_container *container)
 187 {
 188         return !list_empty(&container->group_list);
 189 }
 190
 191 static long tce_iommu_find_table(struct tce_container *container,
 192                 phys_addr_t ioba, struct iommu_table **ptbl)
 193 {
 194         long i;
 195
 196         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
 197                 struct iommu_table *tbl = container->tables[i];
 198
 199                 if (tbl) {
 200                         unsigned long entry = ioba >> tbl->it_page_shift;
 201                         unsigned long start = tbl->it_offset;
 202                         unsigned long end = start + tbl->it_size;
 203
 204                         if ((start <= entry) && (entry < end)) {
 205                                 *ptbl = tbl;
 206                                 return i;
 207                         }
 208                 }
 209         }
 210
 211         return -1;
 212 }
 213
 214 static int tce_iommu_enable(struct tce_container *container)
 215 {
 216         int ret = 0;
 217         unsigned long locked;
 218         struct iommu_table_group *table_group;
 219         struct tce_iommu_group *tcegrp;
 220
 221         if (!current->mm)
 222                 return -ESRCH; /* process exited */
 223
 224         if (container->enabled)
 225                 return -EBUSY;
 226
 227         /*
 228          * When userspace pages are mapped into the IOMMU, they are effectively
 229          * locked memory, so, theoretically, we need to update the accounting
 230          * of locked pages on each map and unmap.  For powerpc, the map unmap
 231          * paths can be very hot, though, and the accounting would kill
 232          * performance, especially since it would be difficult to impossible
 233          * to handle the accounting in real mode only.
 234          *
 235          * To address that, rather than precisely accounting every page, we
 236          * instead account for a worst case on locked memory when the iommu is
 237          * enabled and disabled.  The worst case upper bound on locked memory
 238          * is the size of the whole iommu window, which is usually relatively
 239          * small (compared to total memory sizes) on POWER hardware.
 240          *
 241          * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits,
 242          * that would effectively kill the guest at random points, much better
 243          * enforcing the limit based on the max that the guest can map.
 244          *
 245          * Unfortunately at the moment it counts whole tables, no matter how
 246          * much memory the guest has. I.e. for 4GB guest and 4 IOMMU groups
 247          * each with 2GB DMA window, 8GB will be counted here. The reason for
 248          * this is that we cannot tell here the amount of RAM used by the guest
 249          * as this information is only available from KVM and VFIO is
 250          * KVM agnostic.
 251          *
 252          * So we do not allow enabling a container without a group attached
 253          * as there is no way to know how much we should increment
 254          * the locked_vm counter.
 255          */
 256         if (!tce_groups_attached(container))
 257                 return -ENODEV;
 258
 259         tcegrp = list_first_entry(&container->group_list,
 260                         struct tce_iommu_group, next);
 261         table_group = iommu_group_get_iommudata(tcegrp->grp);
 262         if (!table_group)
 263                 return -ENODEV;
 264
 265         if (!table_group->tce32_size)
 266                 return -EPERM;
 267
 268         locked = table_group->tce32_size >> PAGE_SHIFT;
 269         ret = try_increment_locked_vm(locked);
 270         if (ret)
 271                 return ret;
 272
 273         container->locked_pages = locked;
 274
 275         container->enabled = true;
 276
 277         return ret;
 278 }
 279
 280 static void tce_iommu_disable(struct tce_container *container)
 281 {
 282         if (!container->enabled)
 283                 return;
 284
 285         container->enabled = false;
 286
 287         if (!current->mm)
 288                 return;
 289
 290         decrement_locked_vm(container->locked_pages);
 291 }
 292
 293 static void *tce_iommu_open(unsigned long arg)
 294 {
 295         struct tce_container *container;
 296
 297         if ((arg != VFIO_SPAPR_TCE_IOMMU) && (arg != VFIO_SPAPR_TCE_v2_IOMMU)) {
 298                 pr_err("tce_vfio: Wrong IOMMU type\n");
 299                 return ERR_PTR(-EINVAL);
 300         }
 301
 302         container = kzalloc(sizeof(*container), GFP_KERNEL);
 303         if (!container)
 304                 return ERR_PTR(-ENOMEM);
 305
 306         mutex_init(&container->lock);
 307         INIT_LIST_HEAD_RCU(&container->group_list);
 308
 309         container->v2 = arg == VFIO_SPAPR_TCE_v2_IOMMU;
 310
 311         return container;
 312 }
 313
 314 static int tce_iommu_clear(struct tce_container *container,
 315                 struct iommu_table *tbl,
 316                 unsigned long entry, unsigned long pages);
 317 static void tce_iommu_free_table(struct iommu_table *tbl);
 318
 319 static void tce_iommu_release(void *iommu_data)
 320 {
 321         struct tce_container *container = iommu_data;
 322         struct iommu_table_group *table_group;
 323         struct tce_iommu_group *tcegrp;
 324         long i;
 325
 326         while (tce_groups_attached(container)) {
 327                 tcegrp = list_first_entry(&container->group_list,
 328                                 struct tce_iommu_group, next);
 329                 table_group = iommu_group_get_iommudata(tcegrp->grp);
 330                 tce_iommu_detach_group(iommu_data, tcegrp->grp);
 331         }
 332
 333         /*
 334          * If VFIO created a table, it was not disposed
 335          * by tce_iommu_detach_group() so do it now.
 336          */
 337         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
 338                 struct iommu_table *tbl = container->tables[i];
 339
 340                 if (!tbl)
 341                         continue;
 342
 343                 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
 344                 tce_iommu_free_table(tbl);
 345         }
 346
 347         tce_iommu_disable(container);
 348         mutex_destroy(&container->lock);
 349
 350         kfree(container);
 351 }
 352
 353 static void tce_iommu_unuse_page(struct tce_container *container,
 354                 unsigned long hpa)
 355 {
 356         struct page *page;
 357
 358         page = pfn_to_page(hpa >> PAGE_SHIFT);
 359         put_page(page);
 360 }
 361
 362 static int tce_iommu_prereg_ua_to_hpa(unsigned long tce, unsigned long size,
 363                 unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)
 364 {
 365         long ret = 0;
 366         struct mm_iommu_table_group_mem_t *mem;
 367
 368         mem = mm_iommu_lookup(tce, size);
 369         if (!mem)
 370                 return -EINVAL;
 371
 372         ret = mm_iommu_ua_to_hpa(mem, tce, phpa);
 373         if (ret)
 374                 return -EINVAL;
 375
 376         *pmem = mem;
 377
 378         return 0;
 379 }
 380
 381 static void tce_iommu_unuse_page_v2(struct iommu_table *tbl,
 382                 unsigned long entry)
 383 {
 384         struct mm_iommu_table_group_mem_t *mem = NULL;
 385         int ret;
 386         unsigned long hpa = 0;
 387         unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry);
 388
 389         if (!pua || !current || !current->mm)
 390                 return;
 391
 392         ret = tce_iommu_prereg_ua_to_hpa(*pua, IOMMU_PAGE_SIZE(tbl),
 393                         &hpa, &mem);
 394         if (ret)
 395                 pr_debug("%s: tce %lx at #%lx was not cached, ret=%d\n",
 396                                 __func__, *pua, entry, ret);
 397         if (mem)
 398                 mm_iommu_mapped_dec(mem);
 399
 400         *pua = 0;
 401 }
 402
 403 static int tce_iommu_clear(struct tce_container *container,
 404                 struct iommu_table *tbl,
 405                 unsigned long entry, unsigned long pages)
 406 {
 407         unsigned long oldhpa;
 408         long ret;
 409         enum dma_data_direction direction;
 410
 411         for ( ; pages; --pages, ++entry) {
 412                 direction = DMA_NONE;
 413                 oldhpa = 0;
 414                 ret = iommu_tce_xchg(tbl, entry, &oldhpa, &direction);
 415                 if (ret)
 416                         continue;
 417
 418                 if (direction == DMA_NONE)
 419                         continue;
 420
 421                 if (container->v2) {
 422                         tce_iommu_unuse_page_v2(tbl, entry);
 423                         continue;
 424                 }
 425
 426                 tce_iommu_unuse_page(container, oldhpa);
 427         }
 428
 429         return 0;
 430 }
 431
 432 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
 433 {
 434         struct page *page = NULL;
 435         enum dma_data_direction direction = iommu_tce_direction(tce);
 436
 437         if (get_user_pages_fast(tce & PAGE_MASK, 1,
 438                         direction != DMA_TO_DEVICE, &page) != 1)
 439                 return -EFAULT;
 440
 441         *hpa = __pa((unsigned long) page_address(page));
 442
 443         return 0;
 444 }
 445
 446 static long tce_iommu_build(struct tce_container *container,
 447                 struct iommu_table *tbl,
 448                 unsigned long entry, unsigned long tce, unsigned long pages,
 449                 enum dma_data_direction direction)
 450 {
 451         long i, ret = 0;
 452         struct page *page;
 453         unsigned long hpa;
 454         enum dma_data_direction dirtmp;
 455
 456         for (i = 0; i < pages; ++i) {
 457                 unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
 458
 459                 ret = tce_iommu_use_page(tce, &hpa);
 460                 if (ret)
 461                         break;
 462
 463                 page = pfn_to_page(hpa >> PAGE_SHIFT);
 464                 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
 465                         ret = -EPERM;
 466                         break;
 467                 }
 468
 469                 hpa |= offset;
 470                 dirtmp = direction;
 471                 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
 472                 if (ret) {
 473                         tce_iommu_unuse_page(container, hpa);
 474                         pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
 475                                         __func__, entry << tbl->it_page_shift,
 476                                         tce, ret);
 477                         break;
 478                 }
 479
 480                 if (dirtmp != DMA_NONE)
 481                         tce_iommu_unuse_page(container, hpa);
 482
 483                 tce += IOMMU_PAGE_SIZE(tbl);
 484         }
 485
 486         if (ret)
 487                 tce_iommu_clear(container, tbl, entry, i);
 488
 489         return ret;
 490 }
 491
 492 static long tce_iommu_build_v2(struct tce_container *container,
 493                 struct iommu_table *tbl,
 494                 unsigned long entry, unsigned long tce, unsigned long pages,
 495                 enum dma_data_direction direction)
 496 {
 497         long i, ret = 0;
 498         struct page *page;
 499         unsigned long hpa;
 500         enum dma_data_direction dirtmp;
 501
 502         for (i = 0; i < pages; ++i) {
 503                 struct mm_iommu_table_group_mem_t *mem = NULL;
 504                 unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl,
 505                                 entry + i);
 506
 507                 ret = tce_iommu_prereg_ua_to_hpa(tce, IOMMU_PAGE_SIZE(tbl),
 508                                 &hpa, &mem);
 509                 if (ret)
 510                         break;
 511
 512                 page = pfn_to_page(hpa >> PAGE_SHIFT);
 513                 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
 514                         ret = -EPERM;
 515                         break;
 516                 }
 517
 518                 /* Preserve offset within IOMMU page */
 519                 hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
 520                 dirtmp = direction;
 521
 522                 /* The registered region is being unregistered */
 523                 if (mm_iommu_mapped_inc(mem))
 524                         break;
 525
 526                 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
 527                 if (ret) {
 528                         /* dirtmp cannot be DMA_NONE here */
 529                         tce_iommu_unuse_page_v2(tbl, entry + i);
 530                         pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
 531                                         __func__, entry << tbl->it_page_shift,
 532                                         tce, ret);
 533                         break;
 534                 }
 535
 536                 if (dirtmp != DMA_NONE)
 537                         tce_iommu_unuse_page_v2(tbl, entry + i);
 538
 539                 *pua = tce;
 540
 541                 tce += IOMMU_PAGE_SIZE(tbl);
 542         }
 543
 544         if (ret)
 545                 tce_iommu_clear(container, tbl, entry, i);
 546
 547         return ret;
 548 }
 549
 550 static long tce_iommu_create_table(struct tce_container *container,
 551                         struct iommu_table_group *table_group,
 552                         int num,
 553                         __u32 page_shift,
 554                         __u64 window_size,
 555                         __u32 levels,
 556                         struct iommu_table **ptbl)
 557 {
 558         long ret, table_size;
 559
 560         table_size = table_group->ops->get_table_size(page_shift, window_size,
 561                         levels);
 562         if (!table_size)
 563                 return -EINVAL;
 564
 565         ret = try_increment_locked_vm(table_size >> PAGE_SHIFT);
 566         if (ret)
 567                 return ret;
 568
 569         ret = table_group->ops->create_table(table_group, num,
 570                         page_shift, window_size, levels, ptbl);
 571
 572         WARN_ON(!ret && !(*ptbl)->it_ops->free);
 573         WARN_ON(!ret && ((*ptbl)->it_allocated_size != table_size));
 574
 575         if (!ret && container->v2) {
 576                 ret = tce_iommu_userspace_view_alloc(*ptbl);
 577                 if (ret)
 578                         (*ptbl)->it_ops->free(*ptbl);
 579         }
 580
 581         if (ret)
 582                 decrement_locked_vm(table_size >> PAGE_SHIFT);
 583
 584         return ret;
 585 }
 586
 587 static void tce_iommu_free_table(struct iommu_table *tbl)
 588 {
 589         unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
 590
 591         tce_iommu_userspace_view_free(tbl);
 592         tbl->it_ops->free(tbl);
 593         decrement_locked_vm(pages);
 594 }
 595
 596 static long tce_iommu_ioctl(void *iommu_data,
 597                                  unsigned int cmd, unsigned long arg)
 598 {
 599         struct tce_container *container = iommu_data;
 600         unsigned long minsz;
 601         long ret;
 602
 603         switch (cmd) {
 604         case VFIO_CHECK_EXTENSION:
 605                 switch (arg) {
 606                 case VFIO_SPAPR_TCE_IOMMU:
 607                 case VFIO_SPAPR_TCE_v2_IOMMU:
 608                         ret = 1;
 609                         break;
 610                 default:
 611                         ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
 612                         break;
 613                 }
 614
 615                 return (ret < 0) ? 0 : ret;
 616
 617         case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
 618                 struct vfio_iommu_spapr_tce_info info;
 619                 struct tce_iommu_group *tcegrp;
 620                 struct iommu_table_group *table_group;
 621
 622                 if (!tce_groups_attached(container))
 623                         return -ENXIO;
 624
 625                 tcegrp = list_first_entry(&container->group_list,
 626                                 struct tce_iommu_group, next);
 627                 table_group = iommu_group_get_iommudata(tcegrp->grp);
 628
 629                 if (!table_group)
 630                         return -ENXIO;
 631
 632                 minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
 633                                 dma32_window_size);
 634
 635                 if (copy_from_user(&info, (void __user *)arg, minsz))
 636                         return -EFAULT;
 637
 638                 if (info.argsz < minsz)
 639                         return -EINVAL;
 640
 641                 info.dma32_window_start = table_group->tce32_start;
 642                 info.dma32_window_size = table_group->tce32_size;
 643                 info.flags = 0;
 644
 645                 if (copy_to_user((void __user *)arg, &info, minsz))
 646                         return -EFAULT;
 647
 648                 return 0;
 649         }
 650         case VFIO_IOMMU_MAP_DMA: {
 651                 struct vfio_iommu_type1_dma_map param;
 652                 struct iommu_table *tbl = NULL;
 653                 long num;
 654                 enum dma_data_direction direction;
 655
 656                 if (!container->enabled)
 657                         return -EPERM;
 658
 659                 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
 660
 661                 if (copy_from_user(&param, (void __user *)arg, minsz))
 662                         return -EFAULT;
 663
 664                 if (param.argsz < minsz)
 665                         return -EINVAL;
 666
 667                 if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
 668                                 VFIO_DMA_MAP_FLAG_WRITE))
 669                         return -EINVAL;
 670
 671                 num = tce_iommu_find_table(container, param.iova, &tbl);
 672                 if (num < 0)
 673                         return -ENXIO;
 674
 675                 if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
 676                                 (param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
 677                         return -EINVAL;
 678
 679                 /* iova is checked by the IOMMU API */
 680                 if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
 681                         if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
 682                                 direction = DMA_BIDIRECTIONAL;
 683                         else
 684                                 direction = DMA_TO_DEVICE;
 685                 } else {
 686                         if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
 687                                 direction = DMA_FROM_DEVICE;
 688                         else
 689                                 return -EINVAL;
 690                 }
 691
 692                 ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
 693                 if (ret)
 694                         return ret;
 695
 696                 if (container->v2)
 697                         ret = tce_iommu_build_v2(container, tbl,
 698                                         param.iova >> tbl->it_page_shift,
 699                                         param.vaddr,
 700                                         param.size >> tbl->it_page_shift,
 701                                         direction);
 702                 else
 703                         ret = tce_iommu_build(container, tbl,
 704                                         param.iova >> tbl->it_page_shift,
 705                                         param.vaddr,
 706                                         param.size >> tbl->it_page_shift,
 707                                         direction);
 708
 709                 iommu_flush_tce(tbl);
 710
 711                 return ret;
 712         }
 713         case VFIO_IOMMU_UNMAP_DMA: {
 714                 struct vfio_iommu_type1_dma_unmap param;
 715                 struct iommu_table *tbl = NULL;
 716                 long num;
 717
 718                 if (!container->enabled)
 719                         return -EPERM;
 720
 721                 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
 722                                 size);
 723
 724                 if (copy_from_user(&param, (void __user *)arg, minsz))
 725                         return -EFAULT;
 726
 727                 if (param.argsz < minsz)
 728                         return -EINVAL;
 729
 730                 /* No flag is supported now */
 731                 if (param.flags)
 732                         return -EINVAL;
 733
 734                 num = tce_iommu_find_table(container, param.iova, &tbl);
 735                 if (num < 0)
 736                         return -ENXIO;
 737
 738                 if (param.size & ~IOMMU_PAGE_MASK(tbl))
 739                         return -EINVAL;
 740
 741                 ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
 742                                 param.size >> tbl->it_page_shift);
 743                 if (ret)
 744                         return ret;
 745
 746                 ret = tce_iommu_clear(container, tbl,
 747                                 param.iova >> tbl->it_page_shift,
 748                                 param.size >> tbl->it_page_shift);
 749                 iommu_flush_tce(tbl);
 750
 751                 return ret;
 752         }
 753         case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
 754                 struct vfio_iommu_spapr_register_memory param;
 755
 756                 if (!container->v2)
 757                         break;
 758
 759                 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
 760                                 size);
 761
 762                 if (copy_from_user(&param, (void __user *)arg, minsz))
 763                         return -EFAULT;
 764
 765                 if (param.argsz < minsz)
 766                         return -EINVAL;
 767
 768                 /* No flag is supported now */
 769                 if (param.flags)
 770                         return -EINVAL;
 771
 772                 mutex_lock(&container->lock);
 773                 ret = tce_iommu_register_pages(container, param.vaddr,
 774                                 param.size);
 775                 mutex_unlock(&container->lock);
 776
 777                 return ret;
 778         }
 779         case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
 780                 struct vfio_iommu_spapr_register_memory param;
 781
 782                 if (!container->v2)
 783                         break;
 784
 785                 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
 786                                 size);
 787
 788                 if (copy_from_user(&param, (void __user *)arg, minsz))
 789                         return -EFAULT;
 790
 791                 if (param.argsz < minsz)
 792                         return -EINVAL;
 793
 794                 /* No flag is supported now */
 795                 if (param.flags)
 796                         return -EINVAL;
 797
 798                 mutex_lock(&container->lock);
 799                 ret = tce_iommu_unregister_pages(container, param.vaddr,
 800                                 param.size);
 801                 mutex_unlock(&container->lock);
 802
 803                 return ret;
 804         }
 805         case VFIO_IOMMU_ENABLE:
 806                 if (container->v2)
 807                         break;
 808
 809                 mutex_lock(&container->lock);
 810                 ret = tce_iommu_enable(container);
 811                 mutex_unlock(&container->lock);
 812                 return ret;
 813
 814
 815         case VFIO_IOMMU_DISABLE:
 816                 if (container->v2)
 817                         break;
 818
 819                 mutex_lock(&container->lock);
 820                 tce_iommu_disable(container);
 821                 mutex_unlock(&container->lock);
 822                 return 0;
 823
 824         case VFIO_EEH_PE_OP: {
 825                 struct tce_iommu_group *tcegrp;
 826
 827                 ret = 0;
 828                 list_for_each_entry(tcegrp, &container->group_list, next) {
 829                         ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
 830                                         cmd, arg);
 831                         if (ret)
 832                                 return ret;
 833                 }
 834                 return ret;
 835         }
 836
 837         }
 838
 839         return -ENOTTY;
 840 }
 841
 842 static void tce_iommu_release_ownership(struct tce_container *container,
 843                 struct iommu_table_group *table_group)
 844 {
 845         int i;
 846
 847         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
 848                 struct iommu_table *tbl = container->tables[i];
 849
 850                 if (!tbl)
 851                         continue;
 852
 853                 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
 854                 tce_iommu_userspace_view_free(tbl);
 855                 if (tbl->it_map)
 856                         iommu_release_ownership(tbl);
 857
 858                 container->tables[i] = NULL;
 859         }
 860 }
 861
 862 static int tce_iommu_take_ownership(struct tce_container *container,
 863                 struct iommu_table_group *table_group)
 864 {
 865         int i, j, rc = 0;
 866
 867         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
 868                 struct iommu_table *tbl = table_group->tables[i];
 869
 870                 if (!tbl || !tbl->it_map)
 871                         continue;
 872
 873                 rc = tce_iommu_userspace_view_alloc(tbl);
 874                 if (!rc)
 875                         rc = iommu_take_ownership(tbl);
 876
 877                 if (rc) {
 878                         for (j = 0; j < i; ++j)
 879                                 iommu_release_ownership(
 880                                                 table_group->tables[j]);
 881
 882                         return rc;
 883                 }
 884         }
 885
 886         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
 887                 container->tables[i] = table_group->tables[i];
 888
 889         return 0;
 890 }
 891
 892 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
 893                 struct iommu_table_group *table_group)
 894 {
 895         long i;
 896
 897         if (!table_group->ops->unset_window) {
 898                 WARN_ON_ONCE(1);
 899                 return;
 900         }
 901
 902         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
 903                 table_group->ops->unset_window(table_group, i);
 904
 905         table_group->ops->release_ownership(table_group);
 906 }
 907
 908 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
 909                 struct iommu_table_group *table_group)
 910 {
 911         long i, ret = 0;
 912         struct iommu_table *tbl = NULL;
 913
 914         if (!table_group->ops->create_table || !table_group->ops->set_window ||
 915                         !table_group->ops->release_ownership) {
 916                 WARN_ON_ONCE(1);
 917                 return -EFAULT;
 918         }
 919
 920         table_group->ops->take_ownership(table_group);
 921
 922         /*
 923          * If it the first group attached, check if there is
 924          * a default DMA window and create one if none as
 925          * the userspace expects it to exist.
 926          */
 927         if (!tce_groups_attached(container) && !container->tables[0]) {
 928                 ret = tce_iommu_create_table(container,
 929                                 table_group,
 930                                 0, /* window number */
 931                                 IOMMU_PAGE_SHIFT_4K,
 932                                 table_group->tce32_size,
 933                                 1, /* default levels */
 934                                 &tbl);
 935                 if (ret)
 936                         goto release_exit;
 937                 else
 938                         container->tables[0] = tbl;
 939         }
 940
 941         /* Set all windows to the new group */
 942         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
 943                 tbl = container->tables[i];
 944
 945                 if (!tbl)
 946                         continue;
 947
 948                 /* Set the default window to a new group */
 949                 ret = table_group->ops->set_window(table_group, i, tbl);
 950                 if (ret)
 951                         goto release_exit;
 952         }
 953
 954         return 0;
 955
 956 release_exit:
 957         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
 958                 table_group->ops->unset_window(table_group, i);
 959
 960         table_group->ops->release_ownership(table_group);
 961
 962         return ret;
 963 }
 964
 965 static int tce_iommu_attach_group(void *iommu_data,
 966                 struct iommu_group *iommu_group)
 967 {
 968         int ret;
 969         struct tce_container *container = iommu_data;
 970         struct iommu_table_group *table_group;
 971         struct tce_iommu_group *tcegrp = NULL;
 972
 973         mutex_lock(&container->lock);
 974
 975         /* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
 976                         iommu_group_id(iommu_group), iommu_group); */
 977         table_group = iommu_group_get_iommudata(iommu_group);
 978
 979         if (tce_groups_attached(container) && (!table_group->ops ||
 980                         !table_group->ops->take_ownership ||
 981                         !table_group->ops->release_ownership)) {
 982                 ret = -EBUSY;
 983                 goto unlock_exit;
 984         }
 985
 986         /* Check if new group has the same iommu_ops (i.e. compatible) */
 987         list_for_each_entry(tcegrp, &container->group_list, next) {
 988                 struct iommu_table_group *table_group_tmp;
 989
 990                 if (tcegrp->grp == iommu_group) {
 991                         pr_warn("tce_vfio: Group %d is already attached\n",
 992                                         iommu_group_id(iommu_group));
 993                         ret = -EBUSY;
 994                         goto unlock_exit;
 995                 }
 996                 table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
 997                 if (table_group_tmp->ops != table_group->ops) {
 998                         pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
 999                                         iommu_group_id(iommu_group),
1000                                         iommu_group_id(tcegrp->grp));
1001                         ret = -EPERM;
1002                         goto unlock_exit;
1003                 }
1004         }
1005
1006         tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1007         if (!tcegrp) {
1008                 ret = -ENOMEM;
1009                 goto unlock_exit;
1010         }
1011
1012         if (!table_group->ops || !table_group->ops->take_ownership ||
1013                         !table_group->ops->release_ownership)
1014                 ret = tce_iommu_take_ownership(container, table_group);
1015         else
1016                 ret = tce_iommu_take_ownership_ddw(container, table_group);
1017
1018         if (!ret) {
1019                 tcegrp->grp = iommu_group;
1020                 list_add(&tcegrp->next, &container->group_list);
1021         }
1022
1023 unlock_exit:
1024         if (ret && tcegrp)
1025                 kfree(tcegrp);
1026
1027         mutex_unlock(&container->lock);
1028
1029         return ret;
1030 }
1031
1032 static void tce_iommu_detach_group(void *iommu_data,
1033                 struct iommu_group *iommu_group)
1034 {
1035         struct tce_container *container = iommu_data;
1036         struct iommu_table_group *table_group;
1037         bool found = false;
1038         struct tce_iommu_group *tcegrp;
1039
1040         mutex_lock(&container->lock);
1041
1042         list_for_each_entry(tcegrp, &container->group_list, next) {
1043                 if (tcegrp->grp == iommu_group) {
1044                         found = true;
1045                         break;
1046                 }
1047         }
1048
1049         if (!found) {
1050                 pr_warn("tce_vfio: detaching unattached group #%u\n",
1051                                 iommu_group_id(iommu_group));
1052                 goto unlock_exit;
1053         }
1054
1055         list_del(&tcegrp->next);
1056         kfree(tcegrp);
1057
1058         table_group = iommu_group_get_iommudata(iommu_group);
1059         BUG_ON(!table_group);
1060
1061         if (!table_group->ops || !table_group->ops->release_ownership)
1062                 tce_iommu_release_ownership(container, table_group);
1063         else
1064                 tce_iommu_release_ownership_ddw(container, table_group);
1065
1066 unlock_exit:
1067         mutex_unlock(&container->lock);
1068 }
1069
1070 const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1071         .name           = "iommu-vfio-powerpc",
1072         .owner          = THIS_MODULE,
1073         .open           = tce_iommu_open,
1074         .release        = tce_iommu_release,
1075         .ioctl          = tce_iommu_ioctl,
1076         .attach_group   = tce_iommu_attach_group,
1077         .detach_group   = tce_iommu_detach_group,
1078 };
1079
1080 static int __init tce_iommu_init(void)
1081 {
1082         return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1083 }
1084
1085 static void __exit tce_iommu_cleanup(void)
1086 {
1087         vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1088 }
1089
1090 module_init(tce_iommu_init);
1091 module_exit(tce_iommu_cleanup);
1092
1093 MODULE_VERSION(DRIVER_VERSION);
1094 MODULE_LICENSE("GPL v2");
1095 MODULE_AUTHOR(DRIVER_AUTHOR);
1096 MODULE_DESCRIPTION(DRIVER_DESC);
1097