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 <linux/sched/mm.h>
  24 #include <linux/sched/signal.h>
  25
  26 #include <asm/iommu.h>
  27 #include <asm/tce.h>
  28 #include <asm/mmu_context.h>
  29
  30 #define DRIVER_VERSION  "0.1"
  31 #define DRIVER_AUTHOR   "aik@ozlabs.ru"
  32 #define DRIVER_DESC     "VFIO IOMMU SPAPR TCE"
  33
  34 static void tce_iommu_detach_group(void *iommu_data,
  35                 struct iommu_group *iommu_group);
  36
  37 static long try_increment_locked_vm(struct mm_struct *mm, long npages)
  38 {
  39         long ret = 0, locked, lock_limit;
  40
  41         if (WARN_ON_ONCE(!mm))
  42                 return -EPERM;
  43
  44         if (!npages)
  45                 return 0;
  46
  47         down_write(&mm->mmap_sem);
  48         locked = mm->locked_vm + npages;
  49         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
  50         if (locked > lock_limit && !capable(CAP_IPC_LOCK))
  51                 ret = -ENOMEM;
  52         else
  53                 mm->locked_vm += npages;
  54
  55         pr_debug("[%d] RLIMIT_MEMLOCK +%ld %ld/%ld%s\n", current->pid,
  56                         npages << PAGE_SHIFT,
  57                         mm->locked_vm << PAGE_SHIFT,
  58                         rlimit(RLIMIT_MEMLOCK),
  59                         ret ? " - exceeded" : "");
  60
  61         up_write(&mm->mmap_sem);
  62
  63         return ret;
  64 }
  65
  66 static void decrement_locked_vm(struct mm_struct *mm, long npages)
  67 {
  68         if (!mm || !npages)
  69                 return;
  70
  71         down_write(&mm->mmap_sem);
  72         if (WARN_ON_ONCE(npages > mm->locked_vm))
  73                 npages = mm->locked_vm;
  74         mm->locked_vm -= npages;
  75         pr_debug("[%d] RLIMIT_MEMLOCK -%ld %ld/%ld\n", current->pid,
  76                         npages << PAGE_SHIFT,
  77                         mm->locked_vm << PAGE_SHIFT,
  78                         rlimit(RLIMIT_MEMLOCK));
  79         up_write(&mm->mmap_sem);
  80 }
  81
  82 /*
  83  * VFIO IOMMU fd for SPAPR_TCE IOMMU implementation
  84  *
  85  * This code handles mapping and unmapping of user data buffers
  86  * into DMA'ble space using the IOMMU
  87  */
  88
  89 struct tce_iommu_group {
  90         struct list_head next;
  91         struct iommu_group *grp;
  92 };
  93
  94 /*
  95  * A container needs to remember which preregistered region  it has
  96  * referenced to do proper cleanup at the userspace process exit.
  97  */
  98 struct tce_iommu_prereg {
  99         struct list_head next;
 100         struct mm_iommu_table_group_mem_t *mem;
 101 };
 102
 103 /*
 104  * The container descriptor supports only a single group per container.
 105  * Required by the API as the container is not supplied with the IOMMU group
 106  * at the moment of initialization.
 107  */
 108 struct tce_container {
 109         struct mutex lock;
 110         bool enabled;
 111         bool v2;
 112         bool def_window_pending;
 113         unsigned long locked_pages;
 114         struct mm_struct *mm;
 115         struct iommu_table *tables[IOMMU_TABLE_GROUP_MAX_TABLES];
 116         struct list_head group_list;
 117         struct list_head prereg_list;
 118 };
 119
 120 static long tce_iommu_mm_set(struct tce_container *container)
 121 {
 122         if (container->mm) {
 123                 if (container->mm == current->mm)
 124                         return 0;
 125                 return -EPERM;
 126         }
 127         BUG_ON(!current->mm);
 128         container->mm = current->mm;
 129         atomic_inc(&container->mm->mm_count);
 130
 131         return 0;
 132 }
 133
 134 static long tce_iommu_prereg_free(struct tce_container *container,
 135                 struct tce_iommu_prereg *tcemem)
 136 {
 137         long ret;
 138
 139         ret = mm_iommu_put(container->mm, tcemem->mem);
 140         if (ret)
 141                 return ret;
 142
 143         list_del(&tcemem->next);
 144         kfree(tcemem);
 145
 146         return 0;
 147 }
 148
 149 static long tce_iommu_unregister_pages(struct tce_container *container,
 150                 __u64 vaddr, __u64 size)
 151 {
 152         struct mm_iommu_table_group_mem_t *mem;
 153         struct tce_iommu_prereg *tcemem;
 154         bool found = false;
 155
 156         if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK))
 157                 return -EINVAL;
 158
 159         mem = mm_iommu_find(container->mm, vaddr, size >> PAGE_SHIFT);
 160         if (!mem)
 161                 return -ENOENT;
 162
 163         list_for_each_entry(tcemem, &container->prereg_list, next) {
 164                 if (tcemem->mem == mem) {
 165                         found = true;
 166                         break;
 167                 }
 168         }
 169
 170         if (!found)
 171                 return -ENOENT;
 172
 173         return tce_iommu_prereg_free(container, tcemem);
 174 }
 175
 176 static long tce_iommu_register_pages(struct tce_container *container,
 177                 __u64 vaddr, __u64 size)
 178 {
 179         long ret = 0;
 180         struct mm_iommu_table_group_mem_t *mem = NULL;
 181         struct tce_iommu_prereg *tcemem;
 182         unsigned long entries = size >> PAGE_SHIFT;
 183
 184         if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK) ||
 185                         ((vaddr + size) < vaddr))
 186                 return -EINVAL;
 187
 188         mem = mm_iommu_find(container->mm, vaddr, entries);
 189         if (mem) {
 190                 list_for_each_entry(tcemem, &container->prereg_list, next) {
 191                         if (tcemem->mem == mem)
 192                                 return -EBUSY;
 193                 }
 194         }
 195
 196         ret = mm_iommu_get(container->mm, vaddr, entries, &mem);
 197         if (ret)
 198                 return ret;
 199
 200         tcemem = kzalloc(sizeof(*tcemem), GFP_KERNEL);
 201         if (!tcemem) {
 202                 mm_iommu_put(container->mm, mem);
 203                 return -ENOMEM;
 204         }
 205
 206         tcemem->mem = mem;
 207         list_add(&tcemem->next, &container->prereg_list);
 208
 209         container->enabled = true;
 210
 211         return 0;
 212 }
 213
 214 static bool tce_page_is_contained(struct page *page, unsigned page_shift)
 215 {
 216         /*
 217          * Check that the TCE table granularity is not bigger than the size of
 218          * a page we just found. Otherwise the hardware can get access to
 219          * a bigger memory chunk that it should.
 220          */
 221         return (PAGE_SHIFT + compound_order(compound_head(page))) >= page_shift;
 222 }
 223
 224 static inline bool tce_groups_attached(struct tce_container *container)
 225 {
 226         return !list_empty(&container->group_list);
 227 }
 228
 229 static long tce_iommu_find_table(struct tce_container *container,
 230                 phys_addr_t ioba, struct iommu_table **ptbl)
 231 {
 232         long i;
 233
 234         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
 235                 struct iommu_table *tbl = container->tables[i];
 236
 237                 if (tbl) {
 238                         unsigned long entry = ioba >> tbl->it_page_shift;
 239                         unsigned long start = tbl->it_offset;
 240                         unsigned long end = start + tbl->it_size;
 241
 242                         if ((start <= entry) && (entry < end)) {
 243                                 *ptbl = tbl;
 244                                 return i;
 245                         }
 246                 }
 247         }
 248
 249         return -1;
 250 }
 251
 252 static int tce_iommu_find_free_table(struct tce_container *container)
 253 {
 254         int i;
 255
 256         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
 257                 if (!container->tables[i])
 258                         return i;
 259         }
 260
 261         return -ENOSPC;
 262 }
 263
 264 static int tce_iommu_enable(struct tce_container *container)
 265 {
 266         int ret = 0;
 267         unsigned long locked;
 268         struct iommu_table_group *table_group;
 269         struct tce_iommu_group *tcegrp;
 270
 271         if (container->enabled)
 272                 return -EBUSY;
 273
 274         /*
 275          * When userspace pages are mapped into the IOMMU, they are effectively
 276          * locked memory, so, theoretically, we need to update the accounting
 277          * of locked pages on each map and unmap.  For powerpc, the map unmap
 278          * paths can be very hot, though, and the accounting would kill
 279          * performance, especially since it would be difficult to impossible
 280          * to handle the accounting in real mode only.
 281          *
 282          * To address that, rather than precisely accounting every page, we
 283          * instead account for a worst case on locked memory when the iommu is
 284          * enabled and disabled.  The worst case upper bound on locked memory
 285          * is the size of the whole iommu window, which is usually relatively
 286          * small (compared to total memory sizes) on POWER hardware.
 287          *
 288          * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits,
 289          * that would effectively kill the guest at random points, much better
 290          * enforcing the limit based on the max that the guest can map.
 291          *
 292          * Unfortunately at the moment it counts whole tables, no matter how
 293          * much memory the guest has. I.e. for 4GB guest and 4 IOMMU groups
 294          * each with 2GB DMA window, 8GB will be counted here. The reason for
 295          * this is that we cannot tell here the amount of RAM used by the guest
 296          * as this information is only available from KVM and VFIO is
 297          * KVM agnostic.
 298          *
 299          * So we do not allow enabling a container without a group attached
 300          * as there is no way to know how much we should increment
 301          * the locked_vm counter.
 302          */
 303         if (!tce_groups_attached(container))
 304                 return -ENODEV;
 305
 306         tcegrp = list_first_entry(&container->group_list,
 307                         struct tce_iommu_group, next);
 308         table_group = iommu_group_get_iommudata(tcegrp->grp);
 309         if (!table_group)
 310                 return -ENODEV;
 311
 312         if (!table_group->tce32_size)
 313                 return -EPERM;
 314
 315         ret = tce_iommu_mm_set(container);
 316         if (ret)
 317                 return ret;
 318
 319         locked = table_group->tce32_size >> PAGE_SHIFT;
 320         ret = try_increment_locked_vm(container->mm, locked);
 321         if (ret)
 322                 return ret;
 323
 324         container->locked_pages = locked;
 325
 326         container->enabled = true;
 327
 328         return ret;
 329 }
 330
 331 static void tce_iommu_disable(struct tce_container *container)
 332 {
 333         if (!container->enabled)
 334                 return;
 335
 336         container->enabled = false;
 337
 338         BUG_ON(!container->mm);
 339         decrement_locked_vm(container->mm, container->locked_pages);
 340 }
 341
 342 static void *tce_iommu_open(unsigned long arg)
 343 {
 344         struct tce_container *container;
 345
 346         if ((arg != VFIO_SPAPR_TCE_IOMMU) && (arg != VFIO_SPAPR_TCE_v2_IOMMU)) {
 347                 pr_err("tce_vfio: Wrong IOMMU type\n");
 348                 return ERR_PTR(-EINVAL);
 349         }
 350
 351         container = kzalloc(sizeof(*container), GFP_KERNEL);
 352         if (!container)
 353                 return ERR_PTR(-ENOMEM);
 354
 355         mutex_init(&container->lock);
 356         INIT_LIST_HEAD_RCU(&container->group_list);
 357         INIT_LIST_HEAD_RCU(&container->prereg_list);
 358
 359         container->v2 = arg == VFIO_SPAPR_TCE_v2_IOMMU;
 360
 361         return container;
 362 }
 363
 364 static int tce_iommu_clear(struct tce_container *container,
 365                 struct iommu_table *tbl,
 366                 unsigned long entry, unsigned long pages);
 367 static void tce_iommu_free_table(struct tce_container *container,
 368                 struct iommu_table *tbl);
 369
 370 static void tce_iommu_release(void *iommu_data)
 371 {
 372         struct tce_container *container = iommu_data;
 373         struct tce_iommu_group *tcegrp;
 374         long i;
 375
 376         while (tce_groups_attached(container)) {
 377                 tcegrp = list_first_entry(&container->group_list,
 378                                 struct tce_iommu_group, next);
 379                 tce_iommu_detach_group(iommu_data, tcegrp->grp);
 380         }
 381
 382         /*
 383          * If VFIO created a table, it was not disposed
 384          * by tce_iommu_detach_group() so do it now.
 385          */
 386         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
 387                 struct iommu_table *tbl = container->tables[i];
 388
 389                 if (!tbl)
 390                         continue;
 391
 392                 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
 393                 tce_iommu_free_table(container, tbl);
 394         }
 395
 396         while (!list_empty(&container->prereg_list)) {
 397                 struct tce_iommu_prereg *tcemem;
 398
 399                 tcemem = list_first_entry(&container->prereg_list,
 400                                 struct tce_iommu_prereg, next);
 401                 WARN_ON_ONCE(tce_iommu_prereg_free(container, tcemem));
 402         }
 403
 404         tce_iommu_disable(container);
 405         if (container->mm)
 406                 mmdrop(container->mm);
 407         mutex_destroy(&container->lock);
 408
 409         kfree(container);
 410 }
 411
 412 static void tce_iommu_unuse_page(struct tce_container *container,
 413                 unsigned long hpa)
 414 {
 415         struct page *page;
 416
 417         page = pfn_to_page(hpa >> PAGE_SHIFT);
 418         put_page(page);
 419 }
 420
 421 static int tce_iommu_prereg_ua_to_hpa(struct tce_container *container,
 422                 unsigned long tce, unsigned long shift,
 423                 unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)
 424 {
 425         long ret = 0;
 426         struct mm_iommu_table_group_mem_t *mem;
 427
 428         mem = mm_iommu_lookup(container->mm, tce, 1ULL << shift);
 429         if (!mem)
 430                 return -EINVAL;
 431
 432         ret = mm_iommu_ua_to_hpa(mem, tce, shift, phpa);
 433         if (ret)
 434                 return -EINVAL;
 435
 436         *pmem = mem;
 437
 438         return 0;
 439 }
 440
 441 static void tce_iommu_unuse_page_v2(struct tce_container *container,
 442                 struct iommu_table *tbl, unsigned long entry)
 443 {
 444         struct mm_iommu_table_group_mem_t *mem = NULL;
 445         int ret;
 446         unsigned long hpa = 0;
 447         __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry);
 448
 449         if (!pua)
 450                 return;
 451
 452         ret = tce_iommu_prereg_ua_to_hpa(container, be64_to_cpu(*pua),
 453                         tbl->it_page_shift, &hpa, &mem);
 454         if (ret)
 455                 pr_debug("%s: tce %llx at #%lx was not cached, ret=%d\n",
 456                                 __func__, be64_to_cpu(*pua), entry, ret);
 457         if (mem)
 458                 mm_iommu_mapped_dec(mem);
 459
 460         *pua = cpu_to_be64(0);
 461 }
 462
 463 static int tce_iommu_clear(struct tce_container *container,
 464                 struct iommu_table *tbl,
 465                 unsigned long entry, unsigned long pages)
 466 {
 467         unsigned long oldhpa;
 468         long ret;
 469         enum dma_data_direction direction;
 470
 471         for ( ; pages; --pages, ++entry) {
 472                 cond_resched();
 473
 474                 direction = DMA_NONE;
 475                 oldhpa = 0;
 476                 ret = iommu_tce_xchg(tbl, entry, &oldhpa, &direction);
 477                 if (ret)
 478                         continue;
 479
 480                 if (direction == DMA_NONE)
 481                         continue;
 482
 483                 if (container->v2) {
 484                         tce_iommu_unuse_page_v2(container, tbl, entry);
 485                         continue;
 486                 }
 487
 488                 tce_iommu_unuse_page(container, oldhpa);
 489         }
 490
 491         return 0;
 492 }
 493
 494 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
 495 {
 496         struct page *page = NULL;
 497         enum dma_data_direction direction = iommu_tce_direction(tce);
 498
 499         if (get_user_pages_fast(tce & PAGE_MASK, 1,
 500                         direction != DMA_TO_DEVICE, &page) != 1)
 501                 return -EFAULT;
 502
 503         *hpa = __pa((unsigned long) page_address(page));
 504
 505         return 0;
 506 }
 507
 508 static long tce_iommu_build(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                 unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
 520
 521                 ret = tce_iommu_use_page(tce, &hpa);
 522                 if (ret)
 523                         break;
 524
 525                 page = pfn_to_page(hpa >> PAGE_SHIFT);
 526                 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
 527                         ret = -EPERM;
 528                         break;
 529                 }
 530
 531                 hpa |= offset;
 532                 dirtmp = direction;
 533                 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
 534                 if (ret) {
 535                         tce_iommu_unuse_page(container, hpa);
 536                         pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
 537                                         __func__, entry << tbl->it_page_shift,
 538                                         tce, ret);
 539                         break;
 540                 }
 541
 542                 if (dirtmp != DMA_NONE)
 543                         tce_iommu_unuse_page(container, hpa);
 544
 545                 tce += IOMMU_PAGE_SIZE(tbl);
 546         }
 547
 548         if (ret)
 549                 tce_iommu_clear(container, tbl, entry, i);
 550
 551         return ret;
 552 }
 553
 554 static long tce_iommu_build_v2(struct tce_container *container,
 555                 struct iommu_table *tbl,
 556                 unsigned long entry, unsigned long tce, unsigned long pages,
 557                 enum dma_data_direction direction)
 558 {
 559         long i, ret = 0;
 560         struct page *page;
 561         unsigned long hpa;
 562         enum dma_data_direction dirtmp;
 563
 564         for (i = 0; i < pages; ++i) {
 565                 struct mm_iommu_table_group_mem_t *mem = NULL;
 566                 __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry + i);
 567
 568                 ret = tce_iommu_prereg_ua_to_hpa(container,
 569                                 tce, tbl->it_page_shift, &hpa, &mem);
 570                 if (ret)
 571                         break;
 572
 573                 page = pfn_to_page(hpa >> PAGE_SHIFT);
 574                 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
 575                         ret = -EPERM;
 576                         break;
 577                 }
 578
 579                 /* Preserve offset within IOMMU page */
 580                 hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
 581                 dirtmp = direction;
 582
 583                 /* The registered region is being unregistered */
 584                 if (mm_iommu_mapped_inc(mem))
 585                         break;
 586
 587                 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
 588                 if (ret) {
 589                         /* dirtmp cannot be DMA_NONE here */
 590                         tce_iommu_unuse_page_v2(container, tbl, entry + i);
 591                         pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
 592                                         __func__, entry << tbl->it_page_shift,
 593                                         tce, ret);
 594                         break;
 595                 }
 596
 597                 if (dirtmp != DMA_NONE)
 598                         tce_iommu_unuse_page_v2(container, tbl, entry + i);
 599
 600                 *pua = cpu_to_be64(tce);
 601
 602                 tce += IOMMU_PAGE_SIZE(tbl);
 603         }
 604
 605         if (ret)
 606                 tce_iommu_clear(container, tbl, entry, i);
 607
 608         return ret;
 609 }
 610
 611 static long tce_iommu_create_table(struct tce_container *container,
 612                         struct iommu_table_group *table_group,
 613                         int num,
 614                         __u32 page_shift,
 615                         __u64 window_size,
 616                         __u32 levels,
 617                         struct iommu_table **ptbl)
 618 {
 619         long ret, table_size;
 620
 621         table_size = table_group->ops->get_table_size(page_shift, window_size,
 622                         levels);
 623         if (!table_size)
 624                 return -EINVAL;
 625
 626         ret = try_increment_locked_vm(container->mm, table_size >> PAGE_SHIFT);
 627         if (ret)
 628                 return ret;
 629
 630         ret = table_group->ops->create_table(table_group, num,
 631                         page_shift, window_size, levels, ptbl);
 632
 633         WARN_ON(!ret && !(*ptbl)->it_ops->free);
 634         WARN_ON(!ret && ((*ptbl)->it_allocated_size > table_size));
 635
 636         return ret;
 637 }
 638
 639 static void tce_iommu_free_table(struct tce_container *container,
 640                 struct iommu_table *tbl)
 641 {
 642         unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
 643
 644         iommu_tce_table_put(tbl);
 645         decrement_locked_vm(container->mm, pages);
 646 }
 647
 648 static long tce_iommu_create_window(struct tce_container *container,
 649                 __u32 page_shift, __u64 window_size, __u32 levels,
 650                 __u64 *start_addr)
 651 {
 652         struct tce_iommu_group *tcegrp;
 653         struct iommu_table_group *table_group;
 654         struct iommu_table *tbl = NULL;
 655         long ret, num;
 656
 657         num = tce_iommu_find_free_table(container);
 658         if (num < 0)
 659                 return num;
 660
 661         /* Get the first group for ops::create_table */
 662         tcegrp = list_first_entry(&container->group_list,
 663                         struct tce_iommu_group, next);
 664         table_group = iommu_group_get_iommudata(tcegrp->grp);
 665         if (!table_group)
 666                 return -EFAULT;
 667
 668         if (!(table_group->pgsizes & (1ULL << page_shift)))
 669                 return -EINVAL;
 670
 671         if (!table_group->ops->set_window || !table_group->ops->unset_window ||
 672                         !table_group->ops->get_table_size ||
 673                         !table_group->ops->create_table)
 674                 return -EPERM;
 675
 676         /* Create TCE table */
 677         ret = tce_iommu_create_table(container, table_group, num,
 678                         page_shift, window_size, levels, &tbl);
 679         if (ret)
 680                 return ret;
 681
 682         BUG_ON(!tbl->it_ops->free);
 683
 684         /*
 685          * Program the table to every group.
 686          * Groups have been tested for compatibility at the attach time.
 687          */
 688         list_for_each_entry(tcegrp, &container->group_list, next) {
 689                 table_group = iommu_group_get_iommudata(tcegrp->grp);
 690
 691                 ret = table_group->ops->set_window(table_group, num, tbl);
 692                 if (ret)
 693                         goto unset_exit;
 694         }
 695
 696         container->tables[num] = tbl;
 697
 698         /* Return start address assigned by platform in create_table() */
 699         *start_addr = tbl->it_offset << tbl->it_page_shift;
 700
 701         return 0;
 702
 703 unset_exit:
 704         list_for_each_entry(tcegrp, &container->group_list, next) {
 705                 table_group = iommu_group_get_iommudata(tcegrp->grp);
 706                 table_group->ops->unset_window(table_group, num);
 707         }
 708         tce_iommu_free_table(container, tbl);
 709
 710         return ret;
 711 }
 712
 713 static long tce_iommu_remove_window(struct tce_container *container,
 714                 __u64 start_addr)
 715 {
 716         struct iommu_table_group *table_group = NULL;
 717         struct iommu_table *tbl;
 718         struct tce_iommu_group *tcegrp;
 719         int num;
 720
 721         num = tce_iommu_find_table(container, start_addr, &tbl);
 722         if (num < 0)
 723                 return -EINVAL;
 724
 725         BUG_ON(!tbl->it_size);
 726
 727         /* Detach groups from IOMMUs */
 728         list_for_each_entry(tcegrp, &container->group_list, next) {
 729                 table_group = iommu_group_get_iommudata(tcegrp->grp);
 730
 731                 /*
 732                  * SPAPR TCE IOMMU exposes the default DMA window to
 733                  * the guest via dma32_window_start/size of
 734                  * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
 735                  * the userspace to remove this window, some do not so
 736                  * here we check for the platform capability.
 737                  */
 738                 if (!table_group->ops || !table_group->ops->unset_window)
 739                         return -EPERM;
 740
 741                 table_group->ops->unset_window(table_group, num);
 742         }
 743
 744         /* Free table */
 745         tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
 746         tce_iommu_free_table(container, tbl);
 747         container->tables[num] = NULL;
 748
 749         return 0;
 750 }
 751
 752 static long tce_iommu_create_default_window(struct tce_container *container)
 753 {
 754         long ret;
 755         __u64 start_addr = 0;
 756         struct tce_iommu_group *tcegrp;
 757         struct iommu_table_group *table_group;
 758
 759         if (!container->def_window_pending)
 760                 return 0;
 761
 762         if (!tce_groups_attached(container))
 763                 return -ENODEV;
 764
 765         tcegrp = list_first_entry(&container->group_list,
 766                         struct tce_iommu_group, next);
 767         table_group = iommu_group_get_iommudata(tcegrp->grp);
 768         if (!table_group)
 769                 return -ENODEV;
 770
 771         ret = tce_iommu_create_window(container, IOMMU_PAGE_SHIFT_4K,
 772                         table_group->tce32_size, 1, &start_addr);
 773         WARN_ON_ONCE(!ret && start_addr);
 774
 775         if (!ret)
 776                 container->def_window_pending = false;
 777
 778         return ret;
 779 }
 780
 781 static long tce_iommu_ioctl(void *iommu_data,
 782                                  unsigned int cmd, unsigned long arg)
 783 {
 784         struct tce_container *container = iommu_data;
 785         unsigned long minsz, ddwsz;
 786         long ret;
 787
 788         switch (cmd) {
 789         case VFIO_CHECK_EXTENSION:
 790                 switch (arg) {
 791                 case VFIO_SPAPR_TCE_IOMMU:
 792                 case VFIO_SPAPR_TCE_v2_IOMMU:
 793                         ret = 1;
 794                         break;
 795                 default:
 796                         ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
 797                         break;
 798                 }
 799
 800                 return (ret < 0) ? 0 : ret;
 801         }
 802
 803         /*
 804          * Sanity check to prevent one userspace from manipulating
 805          * another userspace mm.
 806          */
 807         BUG_ON(!container);
 808         if (container->mm && container->mm != current->mm)
 809                 return -EPERM;
 810
 811         switch (cmd) {
 812         case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
 813                 struct vfio_iommu_spapr_tce_info info;
 814                 struct tce_iommu_group *tcegrp;
 815                 struct iommu_table_group *table_group;
 816
 817                 if (!tce_groups_attached(container))
 818                         return -ENXIO;
 819
 820                 tcegrp = list_first_entry(&container->group_list,
 821                                 struct tce_iommu_group, next);
 822                 table_group = iommu_group_get_iommudata(tcegrp->grp);
 823
 824                 if (!table_group)
 825                         return -ENXIO;
 826
 827                 minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
 828                                 dma32_window_size);
 829
 830                 if (copy_from_user(&info, (void __user *)arg, minsz))
 831                         return -EFAULT;
 832
 833                 if (info.argsz < minsz)
 834                         return -EINVAL;
 835
 836                 info.dma32_window_start = table_group->tce32_start;
 837                 info.dma32_window_size = table_group->tce32_size;
 838                 info.flags = 0;
 839                 memset(&info.ddw, 0, sizeof(info.ddw));
 840
 841                 if (table_group->max_dynamic_windows_supported &&
 842                                 container->v2) {
 843                         info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
 844                         info.ddw.pgsizes = table_group->pgsizes;
 845                         info.ddw.max_dynamic_windows_supported =
 846                                 table_group->max_dynamic_windows_supported;
 847                         info.ddw.levels = table_group->max_levels;
 848                 }
 849
 850                 ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
 851
 852                 if (info.argsz >= ddwsz)
 853                         minsz = ddwsz;
 854
 855                 if (copy_to_user((void __user *)arg, &info, minsz))
 856                         return -EFAULT;
 857
 858                 return 0;
 859         }
 860         case VFIO_IOMMU_MAP_DMA: {
 861                 struct vfio_iommu_type1_dma_map param;
 862                 struct iommu_table *tbl = NULL;
 863                 long num;
 864                 enum dma_data_direction direction;
 865
 866                 if (!container->enabled)
 867                         return -EPERM;
 868
 869                 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
 870
 871                 if (copy_from_user(&param, (void __user *)arg, minsz))
 872                         return -EFAULT;
 873
 874                 if (param.argsz < minsz)
 875                         return -EINVAL;
 876
 877                 if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
 878                                 VFIO_DMA_MAP_FLAG_WRITE))
 879                         return -EINVAL;
 880
 881                 ret = tce_iommu_create_default_window(container);
 882                 if (ret)
 883                         return ret;
 884
 885                 num = tce_iommu_find_table(container, param.iova, &tbl);
 886                 if (num < 0)
 887                         return -ENXIO;
 888
 889                 if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
 890                                 (param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
 891                         return -EINVAL;
 892
 893                 /* iova is checked by the IOMMU API */
 894                 if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
 895                         if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
 896                                 direction = DMA_BIDIRECTIONAL;
 897                         else
 898                                 direction = DMA_TO_DEVICE;
 899                 } else {
 900                         if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
 901                                 direction = DMA_FROM_DEVICE;
 902                         else
 903                                 return -EINVAL;
 904                 }
 905
 906                 ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
 907                 if (ret)
 908                         return ret;
 909
 910                 if (container->v2)
 911                         ret = tce_iommu_build_v2(container, tbl,
 912                                         param.iova >> tbl->it_page_shift,
 913                                         param.vaddr,
 914                                         param.size >> tbl->it_page_shift,
 915                                         direction);
 916                 else
 917                         ret = tce_iommu_build(container, tbl,
 918                                         param.iova >> tbl->it_page_shift,
 919                                         param.vaddr,
 920                                         param.size >> tbl->it_page_shift,
 921                                         direction);
 922
 923                 iommu_flush_tce(tbl);
 924
 925                 return ret;
 926         }
 927         case VFIO_IOMMU_UNMAP_DMA: {
 928                 struct vfio_iommu_type1_dma_unmap param;
 929                 struct iommu_table *tbl = NULL;
 930                 long num;
 931
 932                 if (!container->enabled)
 933                         return -EPERM;
 934
 935                 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
 936                                 size);
 937
 938                 if (copy_from_user(&param, (void __user *)arg, minsz))
 939                         return -EFAULT;
 940
 941                 if (param.argsz < minsz)
 942                         return -EINVAL;
 943
 944                 /* No flag is supported now */
 945                 if (param.flags)
 946                         return -EINVAL;
 947
 948                 ret = tce_iommu_create_default_window(container);
 949                 if (ret)
 950                         return ret;
 951
 952                 num = tce_iommu_find_table(container, param.iova, &tbl);
 953                 if (num < 0)
 954                         return -ENXIO;
 955
 956                 if (param.size & ~IOMMU_PAGE_MASK(tbl))
 957                         return -EINVAL;
 958
 959                 ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
 960                                 param.size >> tbl->it_page_shift);
 961                 if (ret)
 962                         return ret;
 963
 964                 ret = tce_iommu_clear(container, tbl,
 965                                 param.iova >> tbl->it_page_shift,
 966                                 param.size >> tbl->it_page_shift);
 967                 iommu_flush_tce(tbl);
 968
 969                 return ret;
 970         }
 971         case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
 972                 struct vfio_iommu_spapr_register_memory param;
 973
 974                 if (!container->v2)
 975                         break;
 976
 977                 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
 978                                 size);
 979
 980                 ret = tce_iommu_mm_set(container);
 981                 if (ret)
 982                         return ret;
 983
 984                 if (copy_from_user(&param, (void __user *)arg, minsz))
 985                         return -EFAULT;
 986
 987                 if (param.argsz < minsz)
 988                         return -EINVAL;
 989
 990                 /* No flag is supported now */
 991                 if (param.flags)
 992                         return -EINVAL;
 993
 994                 mutex_lock(&container->lock);
 995                 ret = tce_iommu_register_pages(container, param.vaddr,
 996                                 param.size);
 997                 mutex_unlock(&container->lock);
 998
 999                 return ret;
1000         }
1001         case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
1002                 struct vfio_iommu_spapr_register_memory param;
1003
1004                 if (!container->v2)
1005                         break;
1006
1007                 if (!container->mm)
1008                         return -EPERM;
1009
1010                 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1011                                 size);
1012
1013                 if (copy_from_user(&param, (void __user *)arg, minsz))
1014                         return -EFAULT;
1015
1016                 if (param.argsz < minsz)
1017                         return -EINVAL;
1018
1019                 /* No flag is supported now */
1020                 if (param.flags)
1021                         return -EINVAL;
1022
1023                 mutex_lock(&container->lock);
1024                 ret = tce_iommu_unregister_pages(container, param.vaddr,
1025                                 param.size);
1026                 mutex_unlock(&container->lock);
1027
1028                 return ret;
1029         }
1030         case VFIO_IOMMU_ENABLE:
1031                 if (container->v2)
1032                         break;
1033
1034                 mutex_lock(&container->lock);
1035                 ret = tce_iommu_enable(container);
1036                 mutex_unlock(&container->lock);
1037                 return ret;
1038
1039
1040         case VFIO_IOMMU_DISABLE:
1041                 if (container->v2)
1042                         break;
1043
1044                 mutex_lock(&container->lock);
1045                 tce_iommu_disable(container);
1046                 mutex_unlock(&container->lock);
1047                 return 0;
1048
1049         case VFIO_EEH_PE_OP: {
1050                 struct tce_iommu_group *tcegrp;
1051
1052                 ret = 0;
1053                 list_for_each_entry(tcegrp, &container->group_list, next) {
1054                         ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
1055                                         cmd, arg);
1056                         if (ret)
1057                                 return ret;
1058                 }
1059                 return ret;
1060         }
1061
1062         case VFIO_IOMMU_SPAPR_TCE_CREATE: {
1063                 struct vfio_iommu_spapr_tce_create create;
1064
1065                 if (!container->v2)
1066                         break;
1067
1068                 ret = tce_iommu_mm_set(container);
1069                 if (ret)
1070                         return ret;
1071
1072                 if (!tce_groups_attached(container))
1073                         return -ENXIO;
1074
1075                 minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
1076                                 start_addr);
1077
1078                 if (copy_from_user(&create, (void __user *)arg, minsz))
1079                         return -EFAULT;
1080
1081                 if (create.argsz < minsz)
1082                         return -EINVAL;
1083
1084                 if (create.flags)
1085                         return -EINVAL;
1086
1087                 mutex_lock(&container->lock);
1088
1089                 ret = tce_iommu_create_default_window(container);
1090                 if (!ret)
1091                         ret = tce_iommu_create_window(container,
1092                                         create.page_shift,
1093                                         create.window_size, create.levels,
1094                                         &create.start_addr);
1095
1096                 mutex_unlock(&container->lock);
1097
1098                 if (!ret && copy_to_user((void __user *)arg, &create, minsz))
1099                         ret = -EFAULT;
1100
1101                 return ret;
1102         }
1103         case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1104                 struct vfio_iommu_spapr_tce_remove remove;
1105
1106                 if (!container->v2)
1107                         break;
1108
1109                 ret = tce_iommu_mm_set(container);
1110                 if (ret)
1111                         return ret;
1112
1113                 if (!tce_groups_attached(container))
1114                         return -ENXIO;
1115
1116                 minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1117                                 start_addr);
1118
1119                 if (copy_from_user(&remove, (void __user *)arg, minsz))
1120                         return -EFAULT;
1121
1122                 if (remove.argsz < minsz)
1123                         return -EINVAL;
1124
1125                 if (remove.flags)
1126                         return -EINVAL;
1127
1128                 if (container->def_window_pending && !remove.start_addr) {
1129                         container->def_window_pending = false;
1130                         return 0;
1131                 }
1132
1133                 mutex_lock(&container->lock);
1134
1135                 ret = tce_iommu_remove_window(container, remove.start_addr);
1136
1137                 mutex_unlock(&container->lock);
1138
1139                 return ret;
1140         }
1141         }
1142
1143         return -ENOTTY;
1144 }
1145
1146 static void tce_iommu_release_ownership(struct tce_container *container,
1147                 struct iommu_table_group *table_group)
1148 {
1149         int i;
1150
1151         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1152                 struct iommu_table *tbl = container->tables[i];
1153
1154                 if (!tbl)
1155                         continue;
1156
1157                 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1158                 if (tbl->it_map)
1159                         iommu_release_ownership(tbl);
1160
1161                 container->tables[i] = NULL;
1162         }
1163 }
1164
1165 static int tce_iommu_take_ownership(struct tce_container *container,
1166                 struct iommu_table_group *table_group)
1167 {
1168         int i, j, rc = 0;
1169
1170         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1171                 struct iommu_table *tbl = table_group->tables[i];
1172
1173                 if (!tbl || !tbl->it_map)
1174                         continue;
1175
1176                 rc = iommu_take_ownership(tbl);
1177                 if (rc) {
1178                         for (j = 0; j < i; ++j)
1179                                 iommu_release_ownership(
1180                                                 table_group->tables[j]);
1181
1182                         return rc;
1183                 }
1184         }
1185
1186         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1187                 container->tables[i] = table_group->tables[i];
1188
1189         return 0;
1190 }
1191
1192 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1193                 struct iommu_table_group *table_group)
1194 {
1195         long i;
1196
1197         if (!table_group->ops->unset_window) {
1198                 WARN_ON_ONCE(1);
1199                 return;
1200         }
1201
1202         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1203                 table_group->ops->unset_window(table_group, i);
1204
1205         table_group->ops->release_ownership(table_group);
1206 }
1207
1208 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1209                 struct iommu_table_group *table_group)
1210 {
1211         long i, ret = 0;
1212
1213         if (!table_group->ops->create_table || !table_group->ops->set_window ||
1214                         !table_group->ops->release_ownership) {
1215                 WARN_ON_ONCE(1);
1216                 return -EFAULT;
1217         }
1218
1219         table_group->ops->take_ownership(table_group);
1220
1221         /* Set all windows to the new group */
1222         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1223                 struct iommu_table *tbl = container->tables[i];
1224
1225                 if (!tbl)
1226                         continue;
1227
1228                 ret = table_group->ops->set_window(table_group, i, tbl);
1229                 if (ret)
1230                         goto release_exit;
1231         }
1232
1233         return 0;
1234
1235 release_exit:
1236         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1237                 table_group->ops->unset_window(table_group, i);
1238
1239         table_group->ops->release_ownership(table_group);
1240
1241         return ret;
1242 }
1243
1244 static int tce_iommu_attach_group(void *iommu_data,
1245                 struct iommu_group *iommu_group)
1246 {
1247         int ret;
1248         struct tce_container *container = iommu_data;
1249         struct iommu_table_group *table_group;
1250         struct tce_iommu_group *tcegrp = NULL;
1251
1252         mutex_lock(&container->lock);
1253
1254         /* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1255                         iommu_group_id(iommu_group), iommu_group); */
1256         table_group = iommu_group_get_iommudata(iommu_group);
1257         if (!table_group) {
1258                 ret = -ENODEV;
1259                 goto unlock_exit;
1260         }
1261
1262         if (tce_groups_attached(container) && (!table_group->ops ||
1263                         !table_group->ops->take_ownership ||
1264                         !table_group->ops->release_ownership)) {
1265                 ret = -EBUSY;
1266                 goto unlock_exit;
1267         }
1268
1269         /* Check if new group has the same iommu_ops (i.e. compatible) */
1270         list_for_each_entry(tcegrp, &container->group_list, next) {
1271                 struct iommu_table_group *table_group_tmp;
1272
1273                 if (tcegrp->grp == iommu_group) {
1274                         pr_warn("tce_vfio: Group %d is already attached\n",
1275                                         iommu_group_id(iommu_group));
1276                         ret = -EBUSY;
1277                         goto unlock_exit;
1278                 }
1279                 table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1280                 if (table_group_tmp->ops->create_table !=
1281                                 table_group->ops->create_table) {
1282                         pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1283                                         iommu_group_id(iommu_group),
1284                                         iommu_group_id(tcegrp->grp));
1285                         ret = -EPERM;
1286                         goto unlock_exit;
1287                 }
1288         }
1289
1290         tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1291         if (!tcegrp) {
1292                 ret = -ENOMEM;
1293                 goto unlock_exit;
1294         }
1295
1296         if (!table_group->ops || !table_group->ops->take_ownership ||
1297                         !table_group->ops->release_ownership) {
1298                 if (container->v2) {
1299                         ret = -EPERM;
1300                         goto unlock_exit;
1301                 }
1302                 ret = tce_iommu_take_ownership(container, table_group);
1303         } else {
1304                 if (!container->v2) {
1305                         ret = -EPERM;
1306                         goto unlock_exit;
1307                 }
1308                 ret = tce_iommu_take_ownership_ddw(container, table_group);
1309                 if (!tce_groups_attached(container) && !container->tables[0])
1310                         container->def_window_pending = true;
1311         }
1312
1313         if (!ret) {
1314                 tcegrp->grp = iommu_group;
1315                 list_add(&tcegrp->next, &container->group_list);
1316         }
1317
1318 unlock_exit:
1319         if (ret && tcegrp)
1320                 kfree(tcegrp);
1321
1322         mutex_unlock(&container->lock);
1323
1324         return ret;
1325 }
1326
1327 static void tce_iommu_detach_group(void *iommu_data,
1328                 struct iommu_group *iommu_group)
1329 {
1330         struct tce_container *container = iommu_data;
1331         struct iommu_table_group *table_group;
1332         bool found = false;
1333         struct tce_iommu_group *tcegrp;
1334
1335         mutex_lock(&container->lock);
1336
1337         list_for_each_entry(tcegrp, &container->group_list, next) {
1338                 if (tcegrp->grp == iommu_group) {
1339                         found = true;
1340                         break;
1341                 }
1342         }
1343
1344         if (!found) {
1345                 pr_warn("tce_vfio: detaching unattached group #%u\n",
1346                                 iommu_group_id(iommu_group));
1347                 goto unlock_exit;
1348         }
1349
1350         list_del(&tcegrp->next);
1351         kfree(tcegrp);
1352
1353         table_group = iommu_group_get_iommudata(iommu_group);
1354         BUG_ON(!table_group);
1355
1356         if (!table_group->ops || !table_group->ops->release_ownership)
1357                 tce_iommu_release_ownership(container, table_group);
1358         else
1359                 tce_iommu_release_ownership_ddw(container, table_group);
1360
1361 unlock_exit:
1362         mutex_unlock(&container->lock);
1363 }
1364
1365 const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1366         .name           = "iommu-vfio-powerpc",
1367         .owner          = THIS_MODULE,
1368         .open           = tce_iommu_open,
1369         .release        = tce_iommu_release,
1370         .ioctl          = tce_iommu_ioctl,
1371         .attach_group   = tce_iommu_attach_group,
1372         .detach_group   = tce_iommu_detach_group,
1373 };
1374
1375 static int __init tce_iommu_init(void)
1376 {
1377         return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1378 }
1379
1380 static void __exit tce_iommu_cleanup(void)
1381 {
1382         vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1383 }
1384
1385 module_init(tce_iommu_init);
1386 module_exit(tce_iommu_cleanup);
1387
1388 MODULE_VERSION(DRIVER_VERSION);
1389 MODULE_LICENSE("GPL v2");
1390 MODULE_AUTHOR(DRIVER_AUTHOR);
1391 MODULE_DESCRIPTION(DRIVER_DESC);
1392