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