drivers/vfio/vfio_iommu_spapr_tce.c

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