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