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         struct tce_iommu_prereg *tcemem, *tmtmp;
 375         long i;
 376
 377         while (tce_groups_attached(container)) {
 378                 tcegrp = list_first_entry(&container->group_list,
 379                                 struct tce_iommu_group, next);
 380                 tce_iommu_detach_group(iommu_data, tcegrp->grp);
 381         }
 382
 383         /*
 384          * If VFIO created a table, it was not disposed
 385          * by tce_iommu_detach_group() so do it now.
 386          */
 387         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
 388                 struct iommu_table *tbl = container->tables[i];
 389
 390                 if (!tbl)
 391                         continue;
 392
 393                 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
 394                 tce_iommu_free_table(container, tbl);
 395         }
 396
 397         list_for_each_entry_safe(tcemem, tmtmp, &container->prereg_list, next)
 398                 WARN_ON(tce_iommu_prereg_free(container, tcemem));
 399
 400         tce_iommu_disable(container);
 401         if (container->mm)
 402                 mmdrop(container->mm);
 403         mutex_destroy(&container->lock);
 404
 405         kfree(container);
 406 }
 407
 408 static void tce_iommu_unuse_page(struct tce_container *container,
 409                 unsigned long hpa)
 410 {
 411         struct page *page;
 412
 413         page = pfn_to_page(hpa >> PAGE_SHIFT);
 414         put_page(page);
 415 }
 416
 417 static int tce_iommu_prereg_ua_to_hpa(struct tce_container *container,
 418                 unsigned long tce, unsigned long shift,
 419                 unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)
 420 {
 421         long ret = 0;
 422         struct mm_iommu_table_group_mem_t *mem;
 423
 424         mem = mm_iommu_lookup(container->mm, tce, 1ULL << shift);
 425         if (!mem)
 426                 return -EINVAL;
 427
 428         ret = mm_iommu_ua_to_hpa(mem, tce, shift, phpa);
 429         if (ret)
 430                 return -EINVAL;
 431
 432         *pmem = mem;
 433
 434         return 0;
 435 }
 436
 437 static void tce_iommu_unuse_page_v2(struct tce_container *container,
 438                 struct iommu_table *tbl, unsigned long entry)
 439 {
 440         struct mm_iommu_table_group_mem_t *mem = NULL;
 441         int ret;
 442         unsigned long hpa = 0;
 443         __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry);
 444
 445         if (!pua)
 446                 return;
 447
 448         ret = tce_iommu_prereg_ua_to_hpa(container, be64_to_cpu(*pua),
 449                         tbl->it_page_shift, &hpa, &mem);
 450         if (ret)
 451                 pr_debug("%s: tce %llx at #%lx was not cached, ret=%d\n",
 452                                 __func__, be64_to_cpu(*pua), entry, ret);
 453         if (mem)
 454                 mm_iommu_mapped_dec(mem);
 455
 456         *pua = cpu_to_be64(0);
 457 }
 458
 459 static int tce_iommu_clear(struct tce_container *container,
 460                 struct iommu_table *tbl,
 461                 unsigned long entry, unsigned long pages)
 462 {
 463         unsigned long oldhpa;
 464         long ret;
 465         enum dma_data_direction direction;
 466
 467         for ( ; pages; --pages, ++entry) {
 468                 cond_resched();
 469
 470                 direction = DMA_NONE;
 471                 oldhpa = 0;
 472                 ret = iommu_tce_xchg(tbl, entry, &oldhpa, &direction);
 473                 if (ret)
 474                         continue;
 475
 476                 if (direction == DMA_NONE)
 477                         continue;
 478
 479                 if (container->v2) {
 480                         tce_iommu_unuse_page_v2(container, tbl, entry);
 481                         continue;
 482                 }
 483
 484                 tce_iommu_unuse_page(container, oldhpa);
 485         }
 486
 487         return 0;
 488 }
 489
 490 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
 491 {
 492         struct page *page = NULL;
 493         enum dma_data_direction direction = iommu_tce_direction(tce);
 494
 495         if (get_user_pages_fast(tce & PAGE_MASK, 1,
 496                         direction != DMA_TO_DEVICE, &page) != 1)
 497                 return -EFAULT;
 498
 499         *hpa = __pa((unsigned long) page_address(page));
 500
 501         return 0;
 502 }
 503
 504 static long tce_iommu_build(struct tce_container *container,
 505                 struct iommu_table *tbl,
 506                 unsigned long entry, unsigned long tce, unsigned long pages,
 507                 enum dma_data_direction direction)
 508 {
 509         long i, ret = 0;
 510         struct page *page;
 511         unsigned long hpa;
 512         enum dma_data_direction dirtmp;
 513
 514         for (i = 0; i < pages; ++i) {
 515                 unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
 516
 517                 ret = tce_iommu_use_page(tce, &hpa);
 518                 if (ret)
 519                         break;
 520
 521                 page = pfn_to_page(hpa >> PAGE_SHIFT);
 522                 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
 523                         ret = -EPERM;
 524                         break;
 525                 }
 526
 527                 hpa |= offset;
 528                 dirtmp = direction;
 529                 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
 530                 if (ret) {
 531                         tce_iommu_unuse_page(container, hpa);
 532                         pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
 533                                         __func__, entry << tbl->it_page_shift,
 534                                         tce, ret);
 535                         break;
 536                 }
 537
 538                 if (dirtmp != DMA_NONE)
 539                         tce_iommu_unuse_page(container, hpa);
 540
 541                 tce += IOMMU_PAGE_SIZE(tbl);
 542         }
 543
 544         if (ret)
 545                 tce_iommu_clear(container, tbl, entry, i);
 546
 547         return ret;
 548 }
 549
 550 static long tce_iommu_build_v2(struct tce_container *container,
 551                 struct iommu_table *tbl,
 552                 unsigned long entry, unsigned long tce, unsigned long pages,
 553                 enum dma_data_direction direction)
 554 {
 555         long i, ret = 0;
 556         struct page *page;
 557         unsigned long hpa;
 558         enum dma_data_direction dirtmp;
 559
 560         for (i = 0; i < pages; ++i) {
 561                 struct mm_iommu_table_group_mem_t *mem = NULL;
 562                 __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry + i);
 563
 564                 ret = tce_iommu_prereg_ua_to_hpa(container,
 565                                 tce, tbl->it_page_shift, &hpa, &mem);
 566                 if (ret)
 567                         break;
 568
 569                 page = pfn_to_page(hpa >> PAGE_SHIFT);
 570                 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
 571                         ret = -EPERM;
 572                         break;
 573                 }
 574
 575                 /* Preserve offset within IOMMU page */
 576                 hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
 577                 dirtmp = direction;
 578
 579                 /* The registered region is being unregistered */
 580                 if (mm_iommu_mapped_inc(mem))
 581                         break;
 582
 583                 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
 584                 if (ret) {
 585                         /* dirtmp cannot be DMA_NONE here */
 586                         tce_iommu_unuse_page_v2(container, tbl, entry + i);
 587                         pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
 588                                         __func__, entry << tbl->it_page_shift,
 589                                         tce, ret);
 590                         break;
 591                 }
 592
 593                 if (dirtmp != DMA_NONE)
 594                         tce_iommu_unuse_page_v2(container, tbl, entry + i);
 595
 596                 *pua = cpu_to_be64(tce);
 597
 598                 tce += IOMMU_PAGE_SIZE(tbl);
 599         }
 600
 601         if (ret)
 602                 tce_iommu_clear(container, tbl, entry, i);
 603
 604         return ret;
 605 }
 606
 607 static long tce_iommu_create_table(struct tce_container *container,
 608                         struct iommu_table_group *table_group,
 609                         int num,
 610                         __u32 page_shift,
 611                         __u64 window_size,
 612                         __u32 levels,
 613                         struct iommu_table **ptbl)
 614 {
 615         long ret, table_size;
 616
 617         table_size = table_group->ops->get_table_size(page_shift, window_size,
 618                         levels);
 619         if (!table_size)
 620                 return -EINVAL;
 621
 622         ret = try_increment_locked_vm(container->mm, table_size >> PAGE_SHIFT);
 623         if (ret)
 624                 return ret;
 625
 626         ret = table_group->ops->create_table(table_group, num,
 627                         page_shift, window_size, levels, ptbl);
 628
 629         WARN_ON(!ret && !(*ptbl)->it_ops->free);
 630         WARN_ON(!ret && ((*ptbl)->it_allocated_size > table_size));
 631
 632         return ret;
 633 }
 634
 635 static void tce_iommu_free_table(struct tce_container *container,
 636                 struct iommu_table *tbl)
 637 {
 638         unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
 639
 640         iommu_tce_table_put(tbl);
 641         decrement_locked_vm(container->mm, pages);
 642 }
 643
 644 static long tce_iommu_create_window(struct tce_container *container,
 645                 __u32 page_shift, __u64 window_size, __u32 levels,
 646                 __u64 *start_addr)
 647 {
 648         struct tce_iommu_group *tcegrp;
 649         struct iommu_table_group *table_group;
 650         struct iommu_table *tbl = NULL;
 651         long ret, num;
 652
 653         num = tce_iommu_find_free_table(container);
 654         if (num < 0)
 655                 return num;
 656
 657         /* Get the first group for ops::create_table */
 658         tcegrp = list_first_entry(&container->group_list,
 659                         struct tce_iommu_group, next);
 660         table_group = iommu_group_get_iommudata(tcegrp->grp);
 661         if (!table_group)
 662                 return -EFAULT;
 663
 664         if (!(table_group->pgsizes & (1ULL << page_shift)))
 665                 return -EINVAL;
 666
 667         if (!table_group->ops->set_window || !table_group->ops->unset_window ||
 668                         !table_group->ops->get_table_size ||
 669                         !table_group->ops->create_table)
 670                 return -EPERM;
 671
 672         /* Create TCE table */
 673         ret = tce_iommu_create_table(container, table_group, num,
 674                         page_shift, window_size, levels, &tbl);
 675         if (ret)
 676                 return ret;
 677
 678         BUG_ON(!tbl->it_ops->free);
 679
 680         /*
 681          * Program the table to every group.
 682          * Groups have been tested for compatibility at the attach time.
 683          */
 684         list_for_each_entry(tcegrp, &container->group_list, next) {
 685                 table_group = iommu_group_get_iommudata(tcegrp->grp);
 686
 687                 ret = table_group->ops->set_window(table_group, num, tbl);
 688                 if (ret)
 689                         goto unset_exit;
 690         }
 691
 692         container->tables[num] = tbl;
 693
 694         /* Return start address assigned by platform in create_table() */
 695         *start_addr = tbl->it_offset << tbl->it_page_shift;
 696
 697         return 0;
 698
 699 unset_exit:
 700         list_for_each_entry(tcegrp, &container->group_list, next) {
 701                 table_group = iommu_group_get_iommudata(tcegrp->grp);
 702                 table_group->ops->unset_window(table_group, num);
 703         }
 704         tce_iommu_free_table(container, tbl);
 705
 706         return ret;
 707 }
 708
 709 static long tce_iommu_remove_window(struct tce_container *container,
 710                 __u64 start_addr)
 711 {
 712         struct iommu_table_group *table_group = NULL;
 713         struct iommu_table *tbl;
 714         struct tce_iommu_group *tcegrp;
 715         int num;
 716
 717         num = tce_iommu_find_table(container, start_addr, &tbl);
 718         if (num < 0)
 719                 return -EINVAL;
 720
 721         BUG_ON(!tbl->it_size);
 722
 723         /* Detach groups from IOMMUs */
 724         list_for_each_entry(tcegrp, &container->group_list, next) {
 725                 table_group = iommu_group_get_iommudata(tcegrp->grp);
 726
 727                 /*
 728                  * SPAPR TCE IOMMU exposes the default DMA window to
 729                  * the guest via dma32_window_start/size of
 730                  * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
 731                  * the userspace to remove this window, some do not so
 732                  * here we check for the platform capability.
 733                  */
 734                 if (!table_group->ops || !table_group->ops->unset_window)
 735                         return -EPERM;
 736
 737                 table_group->ops->unset_window(table_group, num);
 738         }
 739
 740         /* Free table */
 741         tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
 742         tce_iommu_free_table(container, tbl);
 743         container->tables[num] = NULL;
 744
 745         return 0;
 746 }
 747
 748 static long tce_iommu_create_default_window(struct tce_container *container)
 749 {
 750         long ret;
 751         __u64 start_addr = 0;
 752         struct tce_iommu_group *tcegrp;
 753         struct iommu_table_group *table_group;
 754
 755         if (!container->def_window_pending)
 756                 return 0;
 757
 758         if (!tce_groups_attached(container))
 759                 return -ENODEV;
 760
 761         tcegrp = list_first_entry(&container->group_list,
 762                         struct tce_iommu_group, next);
 763         table_group = iommu_group_get_iommudata(tcegrp->grp);
 764         if (!table_group)
 765                 return -ENODEV;
 766
 767         ret = tce_iommu_create_window(container, IOMMU_PAGE_SHIFT_4K,
 768                         table_group->tce32_size, 1, &start_addr);
 769         WARN_ON_ONCE(!ret && start_addr);
 770
 771         if (!ret)
 772                 container->def_window_pending = false;
 773
 774         return ret;
 775 }
 776
 777 static long tce_iommu_ioctl(void *iommu_data,
 778                                  unsigned int cmd, unsigned long arg)
 779 {
 780         struct tce_container *container = iommu_data;
 781         unsigned long minsz, ddwsz;
 782         long ret;
 783
 784         switch (cmd) {
 785         case VFIO_CHECK_EXTENSION:
 786                 switch (arg) {
 787                 case VFIO_SPAPR_TCE_IOMMU:
 788                 case VFIO_SPAPR_TCE_v2_IOMMU:
 789                         ret = 1;
 790                         break;
 791                 default:
 792                         ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
 793                         break;
 794                 }
 795
 796                 return (ret < 0) ? 0 : ret;
 797         }
 798
 799         /*
 800          * Sanity check to prevent one userspace from manipulating
 801          * another userspace mm.
 802          */
 803         BUG_ON(!container);
 804         if (container->mm && container->mm != current->mm)
 805                 return -EPERM;
 806
 807         switch (cmd) {
 808         case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
 809                 struct vfio_iommu_spapr_tce_info info;
 810                 struct tce_iommu_group *tcegrp;
 811                 struct iommu_table_group *table_group;
 812
 813                 if (!tce_groups_attached(container))
 814                         return -ENXIO;
 815
 816                 tcegrp = list_first_entry(&container->group_list,
 817                                 struct tce_iommu_group, next);
 818                 table_group = iommu_group_get_iommudata(tcegrp->grp);
 819
 820                 if (!table_group)
 821                         return -ENXIO;
 822
 823                 minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
 824                                 dma32_window_size);
 825
 826                 if (copy_from_user(&info, (void __user *)arg, minsz))
 827                         return -EFAULT;
 828
 829                 if (info.argsz < minsz)
 830                         return -EINVAL;
 831
 832                 info.dma32_window_start = table_group->tce32_start;
 833                 info.dma32_window_size = table_group->tce32_size;
 834                 info.flags = 0;
 835                 memset(&info.ddw, 0, sizeof(info.ddw));
 836
 837                 if (table_group->max_dynamic_windows_supported &&
 838                                 container->v2) {
 839                         info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
 840                         info.ddw.pgsizes = table_group->pgsizes;
 841                         info.ddw.max_dynamic_windows_supported =
 842                                 table_group->max_dynamic_windows_supported;
 843                         info.ddw.levels = table_group->max_levels;
 844                 }
 845
 846                 ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
 847
 848                 if (info.argsz >= ddwsz)
 849                         minsz = ddwsz;
 850
 851                 if (copy_to_user((void __user *)arg, &info, minsz))
 852                         return -EFAULT;
 853
 854                 return 0;
 855         }
 856         case VFIO_IOMMU_MAP_DMA: {
 857                 struct vfio_iommu_type1_dma_map param;
 858                 struct iommu_table *tbl = NULL;
 859                 long num;
 860                 enum dma_data_direction direction;
 861
 862                 if (!container->enabled)
 863                         return -EPERM;
 864
 865                 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
 866
 867                 if (copy_from_user(&param, (void __user *)arg, minsz))
 868                         return -EFAULT;
 869
 870                 if (param.argsz < minsz)
 871                         return -EINVAL;
 872
 873                 if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
 874                                 VFIO_DMA_MAP_FLAG_WRITE))
 875                         return -EINVAL;
 876
 877                 ret = tce_iommu_create_default_window(container);
 878                 if (ret)
 879                         return ret;
 880
 881                 num = tce_iommu_find_table(container, param.iova, &tbl);
 882                 if (num < 0)
 883                         return -ENXIO;
 884
 885                 if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
 886                                 (param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
 887                         return -EINVAL;
 888
 889                 /* iova is checked by the IOMMU API */
 890                 if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
 891                         if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
 892                                 direction = DMA_BIDIRECTIONAL;
 893                         else
 894                                 direction = DMA_TO_DEVICE;
 895                 } else {
 896                         if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
 897                                 direction = DMA_FROM_DEVICE;
 898                         else
 899                                 return -EINVAL;
 900                 }
 901
 902                 ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
 903                 if (ret)
 904                         return ret;
 905
 906                 if (container->v2)
 907                         ret = tce_iommu_build_v2(container, tbl,
 908                                         param.iova >> tbl->it_page_shift,
 909                                         param.vaddr,
 910                                         param.size >> tbl->it_page_shift,
 911                                         direction);
 912                 else
 913                         ret = tce_iommu_build(container, tbl,
 914                                         param.iova >> tbl->it_page_shift,
 915                                         param.vaddr,
 916                                         param.size >> tbl->it_page_shift,
 917                                         direction);
 918
 919                 iommu_flush_tce(tbl);
 920
 921                 return ret;
 922         }
 923         case VFIO_IOMMU_UNMAP_DMA: {
 924                 struct vfio_iommu_type1_dma_unmap param;
 925                 struct iommu_table *tbl = NULL;
 926                 long num;
 927
 928                 if (!container->enabled)
 929                         return -EPERM;
 930
 931                 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
 932                                 size);
 933
 934                 if (copy_from_user(&param, (void __user *)arg, minsz))
 935                         return -EFAULT;
 936
 937                 if (param.argsz < minsz)
 938                         return -EINVAL;
 939
 940                 /* No flag is supported now */
 941                 if (param.flags)
 942                         return -EINVAL;
 943
 944                 ret = tce_iommu_create_default_window(container);
 945                 if (ret)
 946                         return ret;
 947
 948                 num = tce_iommu_find_table(container, param.iova, &tbl);
 949                 if (num < 0)
 950                         return -ENXIO;
 951
 952                 if (param.size & ~IOMMU_PAGE_MASK(tbl))
 953                         return -EINVAL;
 954
 955                 ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
 956                                 param.size >> tbl->it_page_shift);
 957                 if (ret)
 958                         return ret;
 959
 960                 ret = tce_iommu_clear(container, tbl,
 961                                 param.iova >> tbl->it_page_shift,
 962                                 param.size >> tbl->it_page_shift);
 963                 iommu_flush_tce(tbl);
 964
 965                 return ret;
 966         }
 967         case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
 968                 struct vfio_iommu_spapr_register_memory param;
 969
 970                 if (!container->v2)
 971                         break;
 972
 973                 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
 974                                 size);
 975
 976                 ret = tce_iommu_mm_set(container);
 977                 if (ret)
 978                         return ret;
 979
 980                 if (copy_from_user(&param, (void __user *)arg, minsz))
 981                         return -EFAULT;
 982
 983                 if (param.argsz < minsz)
 984                         return -EINVAL;
 985
 986                 /* No flag is supported now */
 987                 if (param.flags)
 988                         return -EINVAL;
 989
 990                 mutex_lock(&container->lock);
 991                 ret = tce_iommu_register_pages(container, param.vaddr,
 992                                 param.size);
 993                 mutex_unlock(&container->lock);
 994
 995                 return ret;
 996         }
 997         case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
 998                 struct vfio_iommu_spapr_register_memory param;
 999
1000                 if (!container->v2)
1001                         break;
1002
1003                 if (!container->mm)
1004                         return -EPERM;
1005
1006                 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1007                                 size);
1008
1009                 if (copy_from_user(&param, (void __user *)arg, minsz))
1010                         return -EFAULT;
1011
1012                 if (param.argsz < minsz)
1013                         return -EINVAL;
1014
1015                 /* No flag is supported now */
1016                 if (param.flags)
1017                         return -EINVAL;
1018
1019                 mutex_lock(&container->lock);
1020                 ret = tce_iommu_unregister_pages(container, param.vaddr,
1021                                 param.size);
1022                 mutex_unlock(&container->lock);
1023
1024                 return ret;
1025         }
1026         case VFIO_IOMMU_ENABLE:
1027                 if (container->v2)
1028                         break;
1029
1030                 mutex_lock(&container->lock);
1031                 ret = tce_iommu_enable(container);
1032                 mutex_unlock(&container->lock);
1033                 return ret;
1034
1035
1036         case VFIO_IOMMU_DISABLE:
1037                 if (container->v2)
1038                         break;
1039
1040                 mutex_lock(&container->lock);
1041                 tce_iommu_disable(container);
1042                 mutex_unlock(&container->lock);
1043                 return 0;
1044
1045         case VFIO_EEH_PE_OP: {
1046                 struct tce_iommu_group *tcegrp;
1047
1048                 ret = 0;
1049                 list_for_each_entry(tcegrp, &container->group_list, next) {
1050                         ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
1051                                         cmd, arg);
1052                         if (ret)
1053                                 return ret;
1054                 }
1055                 return ret;
1056         }
1057
1058         case VFIO_IOMMU_SPAPR_TCE_CREATE: {
1059                 struct vfio_iommu_spapr_tce_create create;
1060
1061                 if (!container->v2)
1062                         break;
1063
1064                 ret = tce_iommu_mm_set(container);
1065                 if (ret)
1066                         return ret;
1067
1068                 if (!tce_groups_attached(container))
1069                         return -ENXIO;
1070
1071                 minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
1072                                 start_addr);
1073
1074                 if (copy_from_user(&create, (void __user *)arg, minsz))
1075                         return -EFAULT;
1076
1077                 if (create.argsz < minsz)
1078                         return -EINVAL;
1079
1080                 if (create.flags)
1081                         return -EINVAL;
1082
1083                 mutex_lock(&container->lock);
1084
1085                 ret = tce_iommu_create_default_window(container);
1086                 if (!ret)
1087                         ret = tce_iommu_create_window(container,
1088                                         create.page_shift,
1089                                         create.window_size, create.levels,
1090                                         &create.start_addr);
1091
1092                 mutex_unlock(&container->lock);
1093
1094                 if (!ret && copy_to_user((void __user *)arg, &create, minsz))
1095                         ret = -EFAULT;
1096
1097                 return ret;
1098         }
1099         case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1100                 struct vfio_iommu_spapr_tce_remove remove;
1101
1102                 if (!container->v2)
1103                         break;
1104
1105                 ret = tce_iommu_mm_set(container);
1106                 if (ret)
1107                         return ret;
1108
1109                 if (!tce_groups_attached(container))
1110                         return -ENXIO;
1111
1112                 minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1113                                 start_addr);
1114
1115                 if (copy_from_user(&remove, (void __user *)arg, minsz))
1116                         return -EFAULT;
1117
1118                 if (remove.argsz < minsz)
1119                         return -EINVAL;
1120
1121                 if (remove.flags)
1122                         return -EINVAL;
1123
1124                 if (container->def_window_pending && !remove.start_addr) {
1125                         container->def_window_pending = false;
1126                         return 0;
1127                 }
1128
1129                 mutex_lock(&container->lock);
1130
1131                 ret = tce_iommu_remove_window(container, remove.start_addr);
1132
1133                 mutex_unlock(&container->lock);
1134
1135                 return ret;
1136         }
1137         }
1138
1139         return -ENOTTY;
1140 }
1141
1142 static void tce_iommu_release_ownership(struct tce_container *container,
1143                 struct iommu_table_group *table_group)
1144 {
1145         int i;
1146
1147         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1148                 struct iommu_table *tbl = container->tables[i];
1149
1150                 if (!tbl)
1151                         continue;
1152
1153                 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1154                 if (tbl->it_map)
1155                         iommu_release_ownership(tbl);
1156
1157                 container->tables[i] = NULL;
1158         }
1159 }
1160
1161 static int tce_iommu_take_ownership(struct tce_container *container,
1162                 struct iommu_table_group *table_group)
1163 {
1164         int i, j, rc = 0;
1165
1166         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1167                 struct iommu_table *tbl = table_group->tables[i];
1168
1169                 if (!tbl || !tbl->it_map)
1170                         continue;
1171
1172                 rc = iommu_take_ownership(tbl);
1173                 if (rc) {
1174                         for (j = 0; j < i; ++j)
1175                                 iommu_release_ownership(
1176                                                 table_group->tables[j]);
1177
1178                         return rc;
1179                 }
1180         }
1181
1182         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1183                 container->tables[i] = table_group->tables[i];
1184
1185         return 0;
1186 }
1187
1188 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1189                 struct iommu_table_group *table_group)
1190 {
1191         long i;
1192
1193         if (!table_group->ops->unset_window) {
1194                 WARN_ON_ONCE(1);
1195                 return;
1196         }
1197
1198         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1199                 table_group->ops->unset_window(table_group, i);
1200
1201         table_group->ops->release_ownership(table_group);
1202 }
1203
1204 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1205                 struct iommu_table_group *table_group)
1206 {
1207         long i, ret = 0;
1208
1209         if (!table_group->ops->create_table || !table_group->ops->set_window ||
1210                         !table_group->ops->release_ownership) {
1211                 WARN_ON_ONCE(1);
1212                 return -EFAULT;
1213         }
1214
1215         table_group->ops->take_ownership(table_group);
1216
1217         /* Set all windows to the new group */
1218         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1219                 struct iommu_table *tbl = container->tables[i];
1220
1221                 if (!tbl)
1222                         continue;
1223
1224                 ret = table_group->ops->set_window(table_group, i, tbl);
1225                 if (ret)
1226                         goto release_exit;
1227         }
1228
1229         return 0;
1230
1231 release_exit:
1232         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1233                 table_group->ops->unset_window(table_group, i);
1234
1235         table_group->ops->release_ownership(table_group);
1236
1237         return ret;
1238 }
1239
1240 static int tce_iommu_attach_group(void *iommu_data,
1241                 struct iommu_group *iommu_group)
1242 {
1243         int ret;
1244         struct tce_container *container = iommu_data;
1245         struct iommu_table_group *table_group;
1246         struct tce_iommu_group *tcegrp = NULL;
1247
1248         mutex_lock(&container->lock);
1249
1250         /* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1251                         iommu_group_id(iommu_group), iommu_group); */
1252         table_group = iommu_group_get_iommudata(iommu_group);
1253         if (!table_group) {
1254                 ret = -ENODEV;
1255                 goto unlock_exit;
1256         }
1257
1258         if (tce_groups_attached(container) && (!table_group->ops ||
1259                         !table_group->ops->take_ownership ||
1260                         !table_group->ops->release_ownership)) {
1261                 ret = -EBUSY;
1262                 goto unlock_exit;
1263         }
1264
1265         /* Check if new group has the same iommu_ops (i.e. compatible) */
1266         list_for_each_entry(tcegrp, &container->group_list, next) {
1267                 struct iommu_table_group *table_group_tmp;
1268
1269                 if (tcegrp->grp == iommu_group) {
1270                         pr_warn("tce_vfio: Group %d is already attached\n",
1271                                         iommu_group_id(iommu_group));
1272                         ret = -EBUSY;
1273                         goto unlock_exit;
1274                 }
1275                 table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1276                 if (table_group_tmp->ops->create_table !=
1277                                 table_group->ops->create_table) {
1278                         pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1279                                         iommu_group_id(iommu_group),
1280                                         iommu_group_id(tcegrp->grp));
1281                         ret = -EPERM;
1282                         goto unlock_exit;
1283                 }
1284         }
1285
1286         tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1287         if (!tcegrp) {
1288                 ret = -ENOMEM;
1289                 goto unlock_exit;
1290         }
1291
1292         if (!table_group->ops || !table_group->ops->take_ownership ||
1293                         !table_group->ops->release_ownership) {
1294                 if (container->v2) {
1295                         ret = -EPERM;
1296                         goto unlock_exit;
1297                 }
1298                 ret = tce_iommu_take_ownership(container, table_group);
1299         } else {
1300                 if (!container->v2) {
1301                         ret = -EPERM;
1302                         goto unlock_exit;
1303                 }
1304                 ret = tce_iommu_take_ownership_ddw(container, table_group);
1305                 if (!tce_groups_attached(container) && !container->tables[0])
1306                         container->def_window_pending = true;
1307         }
1308
1309         if (!ret) {
1310                 tcegrp->grp = iommu_group;
1311                 list_add(&tcegrp->next, &container->group_list);
1312         }
1313
1314 unlock_exit:
1315         if (ret && tcegrp)
1316                 kfree(tcegrp);
1317
1318         mutex_unlock(&container->lock);
1319
1320         return ret;
1321 }
1322
1323 static void tce_iommu_detach_group(void *iommu_data,
1324                 struct iommu_group *iommu_group)
1325 {
1326         struct tce_container *container = iommu_data;
1327         struct iommu_table_group *table_group;
1328         bool found = false;
1329         struct tce_iommu_group *tcegrp;
1330
1331         mutex_lock(&container->lock);
1332
1333         list_for_each_entry(tcegrp, &container->group_list, next) {
1334                 if (tcegrp->grp == iommu_group) {
1335                         found = true;
1336                         break;
1337                 }
1338         }
1339
1340         if (!found) {
1341                 pr_warn("tce_vfio: detaching unattached group #%u\n",
1342                                 iommu_group_id(iommu_group));
1343                 goto unlock_exit;
1344         }
1345
1346         list_del(&tcegrp->next);
1347         kfree(tcegrp);
1348
1349         table_group = iommu_group_get_iommudata(iommu_group);
1350         BUG_ON(!table_group);
1351
1352         if (!table_group->ops || !table_group->ops->release_ownership)
1353                 tce_iommu_release_ownership(container, table_group);
1354         else
1355                 tce_iommu_release_ownership_ddw(container, table_group);
1356
1357 unlock_exit:
1358         mutex_unlock(&container->lock);
1359 }
1360
1361 const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1362         .name           = "iommu-vfio-powerpc",
1363         .owner          = THIS_MODULE,
1364         .open           = tce_iommu_open,
1365         .release        = tce_iommu_release,
1366         .ioctl          = tce_iommu_ioctl,
1367         .attach_group   = tce_iommu_attach_group,
1368         .detach_group   = tce_iommu_detach_group,
1369 };
1370
1371 static int __init tce_iommu_init(void)
1372 {
1373         return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1374 }
1375
1376 static void __exit tce_iommu_cleanup(void)
1377 {
1378         vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1379 }
1380
1381 module_init(tce_iommu_init);
1382 module_exit(tce_iommu_cleanup);
1383
1384 MODULE_VERSION(DRIVER_VERSION);
1385 MODULE_LICENSE("GPL v2");
1386 MODULE_AUTHOR(DRIVER_AUTHOR);
1387 MODULE_DESCRIPTION(DRIVER_DESC);
1388