arch/powerpc/kvm/book3s_64_vio.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *
   4  * Copyright 2010 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
   5  * Copyright 2011 David Gibson, IBM Corporation <dwg@au1.ibm.com>
   6  * Copyright 2016 Alexey Kardashevskiy, IBM Corporation <aik@au1.ibm.com>
   7  */
   8
   9 #include <linux/types.h>
  10 #include <linux/string.h>
  11 #include <linux/kvm.h>
  12 #include <linux/kvm_host.h>
  13 #include <linux/highmem.h>
  14 #include <linux/gfp.h>
  15 #include <linux/slab.h>
  16 #include <linux/sched/signal.h>
  17 #include <linux/hugetlb.h>
  18 #include <linux/list.h>
  19 #include <linux/anon_inodes.h>
  20 #include <linux/iommu.h>
  21 #include <linux/file.h>
  22 #include <linux/mm.h>
  23
  24 #include <asm/kvm_ppc.h>
  25 #include <asm/kvm_book3s.h>
  26 #include <asm/book3s/64/mmu-hash.h>
  27 #include <asm/hvcall.h>
  28 #include <asm/synch.h>
  29 #include <asm/ppc-opcode.h>
  30 #include <asm/udbg.h>
  31 #include <asm/iommu.h>
  32 #include <asm/tce.h>
  33 #include <asm/mmu_context.h>
  34
  35 static unsigned long kvmppc_tce_pages(unsigned long iommu_pages)
  36 {
  37         return ALIGN(iommu_pages * sizeof(u64), PAGE_SIZE) / PAGE_SIZE;
  38 }
  39
  40 static unsigned long kvmppc_stt_pages(unsigned long tce_pages)
  41 {
  42         unsigned long stt_bytes = sizeof(struct kvmppc_spapr_tce_table) +
  43                         (tce_pages * sizeof(struct page *));
  44
  45         return tce_pages + ALIGN(stt_bytes, PAGE_SIZE) / PAGE_SIZE;
  46 }
  47
  48 static void kvm_spapr_tce_iommu_table_free(struct rcu_head *head)
  49 {
  50         struct kvmppc_spapr_tce_iommu_table *stit = container_of(head,
  51                         struct kvmppc_spapr_tce_iommu_table, rcu);
  52
  53         iommu_tce_table_put(stit->tbl);
  54
  55         kfree(stit);
  56 }
  57
  58 static void kvm_spapr_tce_liobn_put(struct kref *kref)
  59 {
  60         struct kvmppc_spapr_tce_iommu_table *stit = container_of(kref,
  61                         struct kvmppc_spapr_tce_iommu_table, kref);
  62
  63         list_del_rcu(&stit->next);
  64
  65         call_rcu(&stit->rcu, kvm_spapr_tce_iommu_table_free);
  66 }
  67
  68 extern void kvm_spapr_tce_release_iommu_group(struct kvm *kvm,
  69                 struct iommu_group *grp)
  70 {
  71         int i;
  72         struct kvmppc_spapr_tce_table *stt;
  73         struct kvmppc_spapr_tce_iommu_table *stit, *tmp;
  74         struct iommu_table_group *table_group = NULL;
  75
  76         rcu_read_lock();
  77         list_for_each_entry_rcu(stt, &kvm->arch.spapr_tce_tables, list) {
  78
  79                 table_group = iommu_group_get_iommudata(grp);
  80                 if (WARN_ON(!table_group))
  81                         continue;
  82
  83                 list_for_each_entry_safe(stit, tmp, &stt->iommu_tables, next) {
  84                         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
  85                                 if (table_group->tables[i] != stit->tbl)
  86                                         continue;
  87
  88                                 kref_put(&stit->kref, kvm_spapr_tce_liobn_put);
  89                         }
  90                 }
  91                 cond_resched_rcu();
  92         }
  93         rcu_read_unlock();
  94 }
  95
  96 extern long kvm_spapr_tce_attach_iommu_group(struct kvm *kvm, int tablefd,
  97                 struct iommu_group *grp)
  98 {
  99         struct kvmppc_spapr_tce_table *stt = NULL;
 100         bool found = false;
 101         struct iommu_table *tbl = NULL;
 102         struct iommu_table_group *table_group;
 103         long i;
 104         struct kvmppc_spapr_tce_iommu_table *stit;
 105         struct fd f;
 106
 107         f = fdget(tablefd);
 108         if (!f.file)
 109                 return -EBADF;
 110
 111         rcu_read_lock();
 112         list_for_each_entry_rcu(stt, &kvm->arch.spapr_tce_tables, list) {
 113                 if (stt == f.file->private_data) {
 114                         found = true;
 115                         break;
 116                 }
 117         }
 118         rcu_read_unlock();
 119
 120         fdput(f);
 121
 122         if (!found)
 123                 return -EINVAL;
 124
 125         table_group = iommu_group_get_iommudata(grp);
 126         if (WARN_ON(!table_group))
 127                 return -EFAULT;
 128
 129         for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
 130                 struct iommu_table *tbltmp = table_group->tables[i];
 131
 132                 if (!tbltmp)
 133                         continue;
 134                 /* Make sure hardware table parameters are compatible */
 135                 if ((tbltmp->it_page_shift <= stt->page_shift) &&
 136                                 (tbltmp->it_offset << tbltmp->it_page_shift ==
 137                                  stt->offset << stt->page_shift) &&
 138                                 (tbltmp->it_size << tbltmp->it_page_shift >=
 139                                  stt->size << stt->page_shift)) {
 140                         /*
 141                          * Reference the table to avoid races with
 142                          * add/remove DMA windows.
 143                          */
 144                         tbl = iommu_tce_table_get(tbltmp);
 145                         break;
 146                 }
 147         }
 148         if (!tbl)
 149                 return -EINVAL;
 150
 151         rcu_read_lock();
 152         list_for_each_entry_rcu(stit, &stt->iommu_tables, next) {
 153                 if (tbl != stit->tbl)
 154                         continue;
 155
 156                 if (!kref_get_unless_zero(&stit->kref)) {
 157                         /* stit is being destroyed */
 158                         iommu_tce_table_put(tbl);
 159                         rcu_read_unlock();
 160                         return -ENOTTY;
 161                 }
 162                 /*
 163                  * The table is already known to this KVM, we just increased
 164                  * its KVM reference counter and can return.
 165                  */
 166                 rcu_read_unlock();
 167                 return 0;
 168         }
 169         rcu_read_unlock();
 170
 171         stit = kzalloc(sizeof(*stit), GFP_KERNEL);
 172         if (!stit) {
 173                 iommu_tce_table_put(tbl);
 174                 return -ENOMEM;
 175         }
 176
 177         stit->tbl = tbl;
 178         kref_init(&stit->kref);
 179
 180         list_add_rcu(&stit->next, &stt->iommu_tables);
 181
 182         return 0;
 183 }
 184
 185 static void release_spapr_tce_table(struct rcu_head *head)
 186 {
 187         struct kvmppc_spapr_tce_table *stt = container_of(head,
 188                         struct kvmppc_spapr_tce_table, rcu);
 189         unsigned long i, npages = kvmppc_tce_pages(stt->size);
 190
 191         for (i = 0; i < npages; i++)
 192                 if (stt->pages[i])
 193                         __free_page(stt->pages[i]);
 194
 195         kfree(stt);
 196 }
 197
 198 static struct page *kvm_spapr_get_tce_page(struct kvmppc_spapr_tce_table *stt,
 199                 unsigned long sttpage)
 200 {
 201         struct page *page = stt->pages[sttpage];
 202
 203         if (page)
 204                 return page;
 205
 206         mutex_lock(&stt->alloc_lock);
 207         page = stt->pages[sttpage];
 208         if (!page) {
 209                 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 210                 WARN_ON_ONCE(!page);
 211                 if (page)
 212                         stt->pages[sttpage] = page;
 213         }
 214         mutex_unlock(&stt->alloc_lock);
 215
 216         return page;
 217 }
 218
 219 static vm_fault_t kvm_spapr_tce_fault(struct vm_fault *vmf)
 220 {
 221         struct kvmppc_spapr_tce_table *stt = vmf->vma->vm_file->private_data;
 222         struct page *page;
 223
 224         if (vmf->pgoff >= kvmppc_tce_pages(stt->size))
 225                 return VM_FAULT_SIGBUS;
 226
 227         page = kvm_spapr_get_tce_page(stt, vmf->pgoff);
 228         if (!page)
 229                 return VM_FAULT_OOM;
 230
 231         get_page(page);
 232         vmf->page = page;
 233         return 0;
 234 }
 235
 236 static const struct vm_operations_struct kvm_spapr_tce_vm_ops = {
 237         .fault = kvm_spapr_tce_fault,
 238 };
 239
 240 static int kvm_spapr_tce_mmap(struct file *file, struct vm_area_struct *vma)
 241 {
 242         vma->vm_ops = &kvm_spapr_tce_vm_ops;
 243         return 0;
 244 }
 245
 246 static int kvm_spapr_tce_release(struct inode *inode, struct file *filp)
 247 {
 248         struct kvmppc_spapr_tce_table *stt = filp->private_data;
 249         struct kvmppc_spapr_tce_iommu_table *stit, *tmp;
 250         struct kvm *kvm = stt->kvm;
 251
 252         mutex_lock(&kvm->lock);
 253         list_del_rcu(&stt->list);
 254         mutex_unlock(&kvm->lock);
 255
 256         list_for_each_entry_safe(stit, tmp, &stt->iommu_tables, next) {
 257                 WARN_ON(!kref_read(&stit->kref));
 258                 while (1) {
 259                         if (kref_put(&stit->kref, kvm_spapr_tce_liobn_put))
 260                                 break;
 261                 }
 262         }
 263
 264         account_locked_vm(kvm->mm,
 265                 kvmppc_stt_pages(kvmppc_tce_pages(stt->size)), false);
 266
 267         kvm_put_kvm(stt->kvm);
 268
 269         call_rcu(&stt->rcu, release_spapr_tce_table);
 270
 271         return 0;
 272 }
 273
 274 static const struct file_operations kvm_spapr_tce_fops = {
 275         .mmap           = kvm_spapr_tce_mmap,
 276         .release        = kvm_spapr_tce_release,
 277 };
 278
 279 long kvm_vm_ioctl_create_spapr_tce(struct kvm *kvm,
 280                                    struct kvm_create_spapr_tce_64 *args)
 281 {
 282         struct kvmppc_spapr_tce_table *stt = NULL;
 283         struct kvmppc_spapr_tce_table *siter;
 284         struct mm_struct *mm = kvm->mm;
 285         unsigned long npages, size = args->size;
 286         int ret;
 287
 288         if (!args->size || args->page_shift < 12 || args->page_shift > 34 ||
 289                 (args->offset + args->size > (ULLONG_MAX >> args->page_shift)))
 290                 return -EINVAL;
 291
 292         npages = kvmppc_tce_pages(size);
 293         ret = account_locked_vm(mm, kvmppc_stt_pages(npages), true);
 294         if (ret)
 295                 return ret;
 296
 297         ret = -ENOMEM;
 298         stt = kzalloc(struct_size(stt, pages, npages), GFP_KERNEL);
 299         if (!stt)
 300                 goto fail_acct;
 301
 302         stt->liobn = args->liobn;
 303         stt->page_shift = args->page_shift;
 304         stt->offset = args->offset;
 305         stt->size = size;
 306         stt->kvm = kvm;
 307         mutex_init(&stt->alloc_lock);
 308         INIT_LIST_HEAD_RCU(&stt->iommu_tables);
 309
 310         mutex_lock(&kvm->lock);
 311
 312         /* Check this LIOBN hasn't been previously allocated */
 313         ret = 0;
 314         list_for_each_entry(siter, &kvm->arch.spapr_tce_tables, list) {
 315                 if (siter->liobn == args->liobn) {
 316                         ret = -EBUSY;
 317                         break;
 318                 }
 319         }
 320
 321         kvm_get_kvm(kvm);
 322         if (!ret)
 323                 ret = anon_inode_getfd("kvm-spapr-tce", &kvm_spapr_tce_fops,
 324                                        stt, O_RDWR | O_CLOEXEC);
 325
 326         if (ret >= 0)
 327                 list_add_rcu(&stt->list, &kvm->arch.spapr_tce_tables);
 328         else
 329                 kvm_put_kvm_no_destroy(kvm);
 330
 331         mutex_unlock(&kvm->lock);
 332
 333         if (ret >= 0)
 334                 return ret;
 335
 336         kfree(stt);
 337  fail_acct:
 338         account_locked_vm(mm, kvmppc_stt_pages(npages), false);
 339         return ret;
 340 }
 341
 342 static long kvmppc_tce_to_ua(struct kvm *kvm, unsigned long tce,
 343                 unsigned long *ua)
 344 {
 345         unsigned long gfn = tce >> PAGE_SHIFT;
 346         struct kvm_memory_slot *memslot;
 347
 348         memslot = __gfn_to_memslot(kvm_memslots(kvm), gfn);
 349         if (!memslot)
 350                 return -EINVAL;
 351
 352         *ua = __gfn_to_hva_memslot(memslot, gfn) |
 353                 (tce & ~(PAGE_MASK | TCE_PCI_READ | TCE_PCI_WRITE));
 354
 355         return 0;
 356 }
 357
 358 static long kvmppc_tce_validate(struct kvmppc_spapr_tce_table *stt,
 359                 unsigned long tce)
 360 {
 361         unsigned long gpa = tce & ~(TCE_PCI_READ | TCE_PCI_WRITE);
 362         enum dma_data_direction dir = iommu_tce_direction(tce);
 363         struct kvmppc_spapr_tce_iommu_table *stit;
 364         unsigned long ua = 0;
 365
 366         /* Allow userspace to poison TCE table */
 367         if (dir == DMA_NONE)
 368                 return H_SUCCESS;
 369
 370         if (iommu_tce_check_gpa(stt->page_shift, gpa))
 371                 return H_TOO_HARD;
 372
 373         if (kvmppc_tce_to_ua(stt->kvm, tce, &ua))
 374                 return H_TOO_HARD;
 375
 376         rcu_read_lock();
 377         list_for_each_entry_rcu(stit, &stt->iommu_tables, next) {
 378                 unsigned long hpa = 0;
 379                 struct mm_iommu_table_group_mem_t *mem;
 380                 long shift = stit->tbl->it_page_shift;
 381
 382                 mem = mm_iommu_lookup(stt->kvm->mm, ua, 1ULL << shift);
 383                 if (!mem || mm_iommu_ua_to_hpa(mem, ua, shift, &hpa)) {
 384                         rcu_read_unlock();
 385                         return H_TOO_HARD;
 386                 }
 387         }
 388         rcu_read_unlock();
 389
 390         return H_SUCCESS;
 391 }
 392
 393 /*
 394  * Handles TCE requests for emulated devices.
 395  * Puts guest TCE values to the table and expects user space to convert them.
 396  * Cannot fail so kvmppc_tce_validate must be called before it.
 397  */
 398 static void kvmppc_tce_put(struct kvmppc_spapr_tce_table *stt,
 399                 unsigned long idx, unsigned long tce)
 400 {
 401         struct page *page;
 402         u64 *tbl;
 403         unsigned long sttpage;
 404
 405         idx -= stt->offset;
 406         sttpage = idx / TCES_PER_PAGE;
 407         page = stt->pages[sttpage];
 408
 409         if (!page) {
 410                 /* We allow any TCE, not just with read|write permissions */
 411                 if (!tce)
 412                         return;
 413
 414                 page = kvm_spapr_get_tce_page(stt, sttpage);
 415                 if (!page)
 416                         return;
 417         }
 418         tbl = page_to_virt(page);
 419
 420         tbl[idx % TCES_PER_PAGE] = tce;
 421 }
 422
 423 static void kvmppc_clear_tce(struct mm_struct *mm, struct iommu_table *tbl,
 424                 unsigned long entry)
 425 {
 426         unsigned long hpa = 0;
 427         enum dma_data_direction dir = DMA_NONE;
 428
 429         iommu_tce_xchg_no_kill(mm, tbl, entry, &hpa, &dir);
 430 }
 431
 432 static long kvmppc_tce_iommu_mapped_dec(struct kvm *kvm,
 433                 struct iommu_table *tbl, unsigned long entry)
 434 {
 435         struct mm_iommu_table_group_mem_t *mem = NULL;
 436         const unsigned long pgsize = 1ULL << tbl->it_page_shift;
 437         __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl, entry);
 438
 439         if (!pua)
 440                 return H_SUCCESS;
 441
 442         mem = mm_iommu_lookup(kvm->mm, be64_to_cpu(*pua), pgsize);
 443         if (!mem)
 444                 return H_TOO_HARD;
 445
 446         mm_iommu_mapped_dec(mem);
 447
 448         *pua = cpu_to_be64(0);
 449
 450         return H_SUCCESS;
 451 }
 452
 453 static long kvmppc_tce_iommu_do_unmap(struct kvm *kvm,
 454                 struct iommu_table *tbl, unsigned long entry)
 455 {
 456         enum dma_data_direction dir = DMA_NONE;
 457         unsigned long hpa = 0;
 458         long ret;
 459
 460         if (WARN_ON_ONCE(iommu_tce_xchg_no_kill(kvm->mm, tbl, entry, &hpa,
 461                                         &dir)))
 462                 return H_TOO_HARD;
 463
 464         if (dir == DMA_NONE)
 465                 return H_SUCCESS;
 466
 467         ret = kvmppc_tce_iommu_mapped_dec(kvm, tbl, entry);
 468         if (ret != H_SUCCESS)
 469                 iommu_tce_xchg_no_kill(kvm->mm, tbl, entry, &hpa, &dir);
 470
 471         return ret;
 472 }
 473
 474 static long kvmppc_tce_iommu_unmap(struct kvm *kvm,
 475                 struct kvmppc_spapr_tce_table *stt, struct iommu_table *tbl,
 476                 unsigned long entry)
 477 {
 478         unsigned long i, ret = H_SUCCESS;
 479         unsigned long subpages = 1ULL << (stt->page_shift - tbl->it_page_shift);
 480         unsigned long io_entry = entry * subpages;
 481
 482         for (i = 0; i < subpages; ++i) {
 483                 ret = kvmppc_tce_iommu_do_unmap(kvm, tbl, io_entry + i);
 484                 if (ret != H_SUCCESS)
 485                         break;
 486         }
 487
 488         return ret;
 489 }
 490
 491 static long kvmppc_tce_iommu_do_map(struct kvm *kvm, struct iommu_table *tbl,
 492                 unsigned long entry, unsigned long ua,
 493                 enum dma_data_direction dir)
 494 {
 495         long ret;
 496         unsigned long hpa;
 497         __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry);
 498         struct mm_iommu_table_group_mem_t *mem;
 499
 500         if (!pua)
 501                 /* it_userspace allocation might be delayed */
 502                 return H_TOO_HARD;
 503
 504         mem = mm_iommu_lookup(kvm->mm, ua, 1ULL << tbl->it_page_shift);
 505         if (!mem)
 506                 /* This only handles v2 IOMMU type, v1 is handled via ioctl() */
 507                 return H_TOO_HARD;
 508
 509         if (WARN_ON_ONCE(mm_iommu_ua_to_hpa(mem, ua, tbl->it_page_shift, &hpa)))
 510                 return H_TOO_HARD;
 511
 512         if (mm_iommu_mapped_inc(mem))
 513                 return H_TOO_HARD;
 514
 515         ret = iommu_tce_xchg_no_kill(kvm->mm, tbl, entry, &hpa, &dir);
 516         if (WARN_ON_ONCE(ret)) {
 517                 mm_iommu_mapped_dec(mem);
 518                 return H_TOO_HARD;
 519         }
 520
 521         if (dir != DMA_NONE)
 522                 kvmppc_tce_iommu_mapped_dec(kvm, tbl, entry);
 523
 524         *pua = cpu_to_be64(ua);
 525
 526         return 0;
 527 }
 528
 529 static long kvmppc_tce_iommu_map(struct kvm *kvm,
 530                 struct kvmppc_spapr_tce_table *stt, struct iommu_table *tbl,
 531                 unsigned long entry, unsigned long ua,
 532                 enum dma_data_direction dir)
 533 {
 534         unsigned long i, pgoff, ret = H_SUCCESS;
 535         unsigned long subpages = 1ULL << (stt->page_shift - tbl->it_page_shift);
 536         unsigned long io_entry = entry * subpages;
 537
 538         for (i = 0, pgoff = 0; i < subpages;
 539                         ++i, pgoff += IOMMU_PAGE_SIZE(tbl)) {
 540
 541                 ret = kvmppc_tce_iommu_do_map(kvm, tbl,
 542                                 io_entry + i, ua + pgoff, dir);
 543                 if (ret != H_SUCCESS)
 544                         break;
 545         }
 546
 547         return ret;
 548 }
 549
 550 long kvmppc_h_put_tce(struct kvm_vcpu *vcpu, unsigned long liobn,
 551                       unsigned long ioba, unsigned long tce)
 552 {
 553         struct kvmppc_spapr_tce_table *stt;
 554         long ret, idx;
 555         struct kvmppc_spapr_tce_iommu_table *stit;
 556         unsigned long entry, ua = 0;
 557         enum dma_data_direction dir;
 558
 559         /* udbg_printf("H_PUT_TCE(): liobn=0x%lx ioba=0x%lx, tce=0x%lx\n", */
 560         /*          liobn, ioba, tce); */
 561
 562         stt = kvmppc_find_table(vcpu->kvm, liobn);
 563         if (!stt)
 564                 return H_TOO_HARD;
 565
 566         ret = kvmppc_ioba_validate(stt, ioba, 1);
 567         if (ret != H_SUCCESS)
 568                 return ret;
 569
 570         idx = srcu_read_lock(&vcpu->kvm->srcu);
 571
 572         ret = kvmppc_tce_validate(stt, tce);
 573         if (ret != H_SUCCESS)
 574                 goto unlock_exit;
 575
 576         dir = iommu_tce_direction(tce);
 577
 578         if ((dir != DMA_NONE) && kvmppc_tce_to_ua(vcpu->kvm, tce, &ua)) {
 579                 ret = H_PARAMETER;
 580                 goto unlock_exit;
 581         }
 582
 583         entry = ioba >> stt->page_shift;
 584
 585         list_for_each_entry_lockless(stit, &stt->iommu_tables, next) {
 586                 if (dir == DMA_NONE)
 587                         ret = kvmppc_tce_iommu_unmap(vcpu->kvm, stt,
 588                                         stit->tbl, entry);
 589                 else
 590                         ret = kvmppc_tce_iommu_map(vcpu->kvm, stt, stit->tbl,
 591                                         entry, ua, dir);
 592
 593                 iommu_tce_kill(stit->tbl, entry, 1);
 594
 595                 if (ret != H_SUCCESS) {
 596                         kvmppc_clear_tce(vcpu->kvm->mm, stit->tbl, entry);
 597                         goto unlock_exit;
 598                 }
 599         }
 600
 601         kvmppc_tce_put(stt, entry, tce);
 602
 603 unlock_exit:
 604         srcu_read_unlock(&vcpu->kvm->srcu, idx);
 605
 606         return ret;
 607 }
 608 EXPORT_SYMBOL_GPL(kvmppc_h_put_tce);
 609
 610 long kvmppc_h_put_tce_indirect(struct kvm_vcpu *vcpu,
 611                 unsigned long liobn, unsigned long ioba,
 612                 unsigned long tce_list, unsigned long npages)
 613 {
 614         struct kvmppc_spapr_tce_table *stt;
 615         long i, ret = H_SUCCESS, idx;
 616         unsigned long entry, ua = 0;
 617         u64 __user *tces;
 618         u64 tce;
 619         struct kvmppc_spapr_tce_iommu_table *stit;
 620
 621         stt = kvmppc_find_table(vcpu->kvm, liobn);
 622         if (!stt)
 623                 return H_TOO_HARD;
 624
 625         entry = ioba >> stt->page_shift;
 626         /*
 627          * SPAPR spec says that the maximum size of the list is 512 TCEs
 628          * so the whole table fits in 4K page
 629          */
 630         if (npages > 512)
 631                 return H_PARAMETER;
 632
 633         if (tce_list & (SZ_4K - 1))
 634                 return H_PARAMETER;
 635
 636         ret = kvmppc_ioba_validate(stt, ioba, npages);
 637         if (ret != H_SUCCESS)
 638                 return ret;
 639
 640         idx = srcu_read_lock(&vcpu->kvm->srcu);
 641         if (kvmppc_tce_to_ua(vcpu->kvm, tce_list, &ua)) {
 642                 ret = H_TOO_HARD;
 643                 goto unlock_exit;
 644         }
 645         tces = (u64 __user *) ua;
 646
 647         for (i = 0; i < npages; ++i) {
 648                 if (get_user(tce, tces + i)) {
 649                         ret = H_TOO_HARD;
 650                         goto unlock_exit;
 651                 }
 652                 tce = be64_to_cpu(tce);
 653
 654                 ret = kvmppc_tce_validate(stt, tce);
 655                 if (ret != H_SUCCESS)
 656                         goto unlock_exit;
 657         }
 658
 659         for (i = 0; i < npages; ++i) {
 660                 /*
 661                  * This looks unsafe, because we validate, then regrab
 662                  * the TCE from userspace which could have been changed by
 663                  * another thread.
 664                  *
 665                  * But it actually is safe, because the relevant checks will be
 666                  * re-executed in the following code.  If userspace tries to
 667                  * change this dodgily it will result in a messier failure mode
 668                  * but won't threaten the host.
 669                  */
 670                 if (get_user(tce, tces + i)) {
 671                         ret = H_TOO_HARD;
 672                         goto invalidate_exit;
 673                 }
 674                 tce = be64_to_cpu(tce);
 675
 676                 if (kvmppc_tce_to_ua(vcpu->kvm, tce, &ua)) {
 677                         ret = H_PARAMETER;
 678                         goto invalidate_exit;
 679                 }
 680
 681                 list_for_each_entry_lockless(stit, &stt->iommu_tables, next) {
 682                         ret = kvmppc_tce_iommu_map(vcpu->kvm, stt,
 683                                         stit->tbl, entry + i, ua,
 684                                         iommu_tce_direction(tce));
 685
 686                         if (ret != H_SUCCESS) {
 687                                 kvmppc_clear_tce(vcpu->kvm->mm, stit->tbl,
 688                                                 entry);
 689                                 goto invalidate_exit;
 690                         }
 691                 }
 692
 693                 kvmppc_tce_put(stt, entry + i, tce);
 694         }
 695
 696 invalidate_exit:
 697         list_for_each_entry_lockless(stit, &stt->iommu_tables, next)
 698                 iommu_tce_kill(stit->tbl, entry, npages);
 699
 700 unlock_exit:
 701         srcu_read_unlock(&vcpu->kvm->srcu, idx);
 702
 703         return ret;
 704 }
 705 EXPORT_SYMBOL_GPL(kvmppc_h_put_tce_indirect);
 706
 707 long kvmppc_h_stuff_tce(struct kvm_vcpu *vcpu,
 708                 unsigned long liobn, unsigned long ioba,
 709                 unsigned long tce_value, unsigned long npages)
 710 {
 711         struct kvmppc_spapr_tce_table *stt;
 712         long i, ret;
 713         struct kvmppc_spapr_tce_iommu_table *stit;
 714
 715         stt = kvmppc_find_table(vcpu->kvm, liobn);
 716         if (!stt)
 717                 return H_TOO_HARD;
 718
 719         ret = kvmppc_ioba_validate(stt, ioba, npages);
 720         if (ret != H_SUCCESS)
 721                 return ret;
 722
 723         /* Check permission bits only to allow userspace poison TCE for debug */
 724         if (tce_value & (TCE_PCI_WRITE | TCE_PCI_READ))
 725                 return H_PARAMETER;
 726
 727         list_for_each_entry_lockless(stit, &stt->iommu_tables, next) {
 728                 unsigned long entry = ioba >> stt->page_shift;
 729
 730                 for (i = 0; i < npages; ++i) {
 731                         ret = kvmppc_tce_iommu_unmap(vcpu->kvm, stt,
 732                                         stit->tbl, entry + i);
 733
 734                         if (ret == H_SUCCESS)
 735                                 continue;
 736
 737                         if (ret == H_TOO_HARD)
 738                                 goto invalidate_exit;
 739
 740                         WARN_ON_ONCE(1);
 741                         kvmppc_clear_tce(vcpu->kvm->mm, stit->tbl, entry);
 742                 }
 743         }
 744
 745         for (i = 0; i < npages; ++i, ioba += (1ULL << stt->page_shift))
 746                 kvmppc_tce_put(stt, ioba >> stt->page_shift, tce_value);
 747
 748 invalidate_exit:
 749         list_for_each_entry_lockless(stit, &stt->iommu_tables, next)
 750                 iommu_tce_kill(stit->tbl, ioba >> stt->page_shift, npages);
 751
 752         return ret;
 753 }
 754 EXPORT_SYMBOL_GPL(kvmppc_h_stuff_tce);