arch/powerpc/kvm/book3s_64_vio_hv.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/hugetlb.h>
  17 #include <linux/list.h>
  18 #include <linux/stringify.h>
  19
  20 #include <asm/kvm_ppc.h>
  21 #include <asm/kvm_book3s.h>
  22 #include <asm/book3s/64/mmu-hash.h>
  23 #include <asm/mmu_context.h>
  24 #include <asm/hvcall.h>
  25 #include <asm/synch.h>
  26 #include <asm/ppc-opcode.h>
  27 #include <asm/udbg.h>
  28 #include <asm/iommu.h>
  29 #include <asm/tce.h>
  30 #include <asm/pte-walk.h>
  31
  32 #ifdef CONFIG_BUG
  33
  34 #define WARN_ON_ONCE_RM(condition)      ({                      \
  35         static bool __section(".data.unlikely") __warned;       \
  36         int __ret_warn_once = !!(condition);                    \
  37                                                                 \
  38         if (unlikely(__ret_warn_once && !__warned)) {           \
  39                 __warned = true;                                \
  40                 pr_err("WARN_ON_ONCE_RM: (%s) at %s:%u\n",      \
  41                                 __stringify(condition),         \
  42                                 __func__, __LINE__);            \
  43                 dump_stack();                                   \
  44         }                                                       \
  45         unlikely(__ret_warn_once);                              \
  46 })
  47
  48 #else
  49
  50 #define WARN_ON_ONCE_RM(condition) ({                           \
  51         int __ret_warn_on = !!(condition);                      \
  52         unlikely(__ret_warn_on);                                \
  53 })
  54
  55 #endif
  56
  57 /*
  58  * Finds a TCE table descriptor by LIOBN.
  59  *
  60  * WARNING: This will be called in real or virtual mode on HV KVM and virtual
  61  *          mode on PR KVM
  62  */
  63 struct kvmppc_spapr_tce_table *kvmppc_find_table(struct kvm *kvm,
  64                 unsigned long liobn)
  65 {
  66         struct kvmppc_spapr_tce_table *stt;
  67
  68         list_for_each_entry_lockless(stt, &kvm->arch.spapr_tce_tables, list)
  69                 if (stt->liobn == liobn)
  70                         return stt;
  71
  72         return NULL;
  73 }
  74 EXPORT_SYMBOL_GPL(kvmppc_find_table);
  75
  76 #ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE
  77 static long kvmppc_rm_tce_to_ua(struct kvm *kvm,
  78                                 unsigned long tce, unsigned long *ua)
  79 {
  80         unsigned long gfn = tce >> PAGE_SHIFT;
  81         struct kvm_memory_slot *memslot;
  82
  83         memslot = search_memslots(kvm_memslots_raw(kvm), gfn);
  84         if (!memslot)
  85                 return -EINVAL;
  86
  87         *ua = __gfn_to_hva_memslot(memslot, gfn) |
  88                 (tce & ~(PAGE_MASK | TCE_PCI_READ | TCE_PCI_WRITE));
  89
  90         return 0;
  91 }
  92
  93 /*
  94  * Validates TCE address.
  95  * At the moment flags and page mask are validated.
  96  * As the host kernel does not access those addresses (just puts them
  97  * to the table and user space is supposed to process them), we can skip
  98  * checking other things (such as TCE is a guest RAM address or the page
  99  * was actually allocated).
 100  */
 101 static long kvmppc_rm_tce_validate(struct kvmppc_spapr_tce_table *stt,
 102                 unsigned long tce)
 103 {
 104         unsigned long gpa = tce & ~(TCE_PCI_READ | TCE_PCI_WRITE);
 105         enum dma_data_direction dir = iommu_tce_direction(tce);
 106         struct kvmppc_spapr_tce_iommu_table *stit;
 107         unsigned long ua = 0;
 108
 109         /* Allow userspace to poison TCE table */
 110         if (dir == DMA_NONE)
 111                 return H_SUCCESS;
 112
 113         if (iommu_tce_check_gpa(stt->page_shift, gpa))
 114                 return H_PARAMETER;
 115
 116         if (kvmppc_rm_tce_to_ua(stt->kvm, tce, &ua))
 117                 return H_TOO_HARD;
 118
 119         list_for_each_entry_lockless(stit, &stt->iommu_tables, next) {
 120                 unsigned long hpa = 0;
 121                 struct mm_iommu_table_group_mem_t *mem;
 122                 long shift = stit->tbl->it_page_shift;
 123
 124                 mem = mm_iommu_lookup_rm(stt->kvm->mm, ua, 1ULL << shift);
 125                 if (!mem)
 126                         return H_TOO_HARD;
 127
 128                 if (mm_iommu_ua_to_hpa_rm(mem, ua, shift, &hpa))
 129                         return H_TOO_HARD;
 130         }
 131
 132         return H_SUCCESS;
 133 }
 134
 135 /* Note on the use of page_address() in real mode,
 136  *
 137  * It is safe to use page_address() in real mode on ppc64 because
 138  * page_address() is always defined as lowmem_page_address()
 139  * which returns __va(PFN_PHYS(page_to_pfn(page))) which is arithmetic
 140  * operation and does not access page struct.
 141  *
 142  * Theoretically page_address() could be defined different
 143  * but either WANT_PAGE_VIRTUAL or HASHED_PAGE_VIRTUAL
 144  * would have to be enabled.
 145  * WANT_PAGE_VIRTUAL is never enabled on ppc32/ppc64,
 146  * HASHED_PAGE_VIRTUAL could be enabled for ppc32 only and only
 147  * if CONFIG_HIGHMEM is defined. As CONFIG_SPARSEMEM_VMEMMAP
 148  * is not expected to be enabled on ppc32, page_address()
 149  * is safe for ppc32 as well.
 150  *
 151  * WARNING: This will be called in real-mode on HV KVM and virtual
 152  *          mode on PR KVM
 153  */
 154 static u64 *kvmppc_page_address(struct page *page)
 155 {
 156 #if defined(HASHED_PAGE_VIRTUAL) || defined(WANT_PAGE_VIRTUAL)
 157 #error TODO: fix to avoid page_address() here
 158 #endif
 159         return (u64 *) page_address(page);
 160 }
 161
 162 /*
 163  * Handles TCE requests for emulated devices.
 164  * Puts guest TCE values to the table and expects user space to convert them.
 165  * Cannot fail so kvmppc_rm_tce_validate must be called before it.
 166  */
 167 static void kvmppc_rm_tce_put(struct kvmppc_spapr_tce_table *stt,
 168                 unsigned long idx, unsigned long tce)
 169 {
 170         struct page *page;
 171         u64 *tbl;
 172
 173         idx -= stt->offset;
 174         page = stt->pages[idx / TCES_PER_PAGE];
 175         /*
 176          * page must not be NULL in real mode,
 177          * kvmppc_rm_ioba_validate() must have taken care of this.
 178          */
 179         WARN_ON_ONCE_RM(!page);
 180         tbl = kvmppc_page_address(page);
 181
 182         tbl[idx % TCES_PER_PAGE] = tce;
 183 }
 184
 185 /*
 186  * TCEs pages are allocated in kvmppc_rm_tce_put() which won't be able to do so
 187  * in real mode.
 188  * Check if kvmppc_rm_tce_put() can succeed in real mode, i.e. a TCEs page is
 189  * allocated or not required (when clearing a tce entry).
 190  */
 191 static long kvmppc_rm_ioba_validate(struct kvmppc_spapr_tce_table *stt,
 192                 unsigned long ioba, unsigned long npages, bool clearing)
 193 {
 194         unsigned long i, idx, sttpage, sttpages;
 195         unsigned long ret = kvmppc_ioba_validate(stt, ioba, npages);
 196
 197         if (ret)
 198                 return ret;
 199         /*
 200          * clearing==true says kvmppc_rm_tce_put won't be allocating pages
 201          * for empty tces.
 202          */
 203         if (clearing)
 204                 return H_SUCCESS;
 205
 206         idx = (ioba >> stt->page_shift) - stt->offset;
 207         sttpage = idx / TCES_PER_PAGE;
 208         sttpages = ALIGN(idx % TCES_PER_PAGE + npages, TCES_PER_PAGE) /
 209                         TCES_PER_PAGE;
 210         for (i = sttpage; i < sttpage + sttpages; ++i)
 211                 if (!stt->pages[i])
 212                         return H_TOO_HARD;
 213
 214         return H_SUCCESS;
 215 }
 216
 217 static long iommu_tce_xchg_no_kill_rm(struct mm_struct *mm,
 218                 struct iommu_table *tbl,
 219                 unsigned long entry, unsigned long *hpa,
 220                 enum dma_data_direction *direction)
 221 {
 222         long ret;
 223
 224         ret = tbl->it_ops->xchg_no_kill(tbl, entry, hpa, direction, true);
 225
 226         if (!ret && ((*direction == DMA_FROM_DEVICE) ||
 227                                 (*direction == DMA_BIDIRECTIONAL))) {
 228                 __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl, entry);
 229                 /*
 230                  * kvmppc_rm_tce_iommu_do_map() updates the UA cache after
 231                  * calling this so we still get here a valid UA.
 232                  */
 233                 if (pua && *pua)
 234                         mm_iommu_ua_mark_dirty_rm(mm, be64_to_cpu(*pua));
 235         }
 236
 237         return ret;
 238 }
 239
 240 static void iommu_tce_kill_rm(struct iommu_table *tbl,
 241                 unsigned long entry, unsigned long pages)
 242 {
 243         if (tbl->it_ops->tce_kill)
 244                 tbl->it_ops->tce_kill(tbl, entry, pages, true);
 245 }
 246
 247 static void kvmppc_rm_clear_tce(struct kvm *kvm, struct iommu_table *tbl,
 248                 unsigned long entry)
 249 {
 250         unsigned long hpa = 0;
 251         enum dma_data_direction dir = DMA_NONE;
 252
 253         iommu_tce_xchg_no_kill_rm(kvm->mm, tbl, entry, &hpa, &dir);
 254 }
 255
 256 static long kvmppc_rm_tce_iommu_mapped_dec(struct kvm *kvm,
 257                 struct iommu_table *tbl, unsigned long entry)
 258 {
 259         struct mm_iommu_table_group_mem_t *mem = NULL;
 260         const unsigned long pgsize = 1ULL << tbl->it_page_shift;
 261         __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl, entry);
 262
 263         if (!pua)
 264                 /* it_userspace allocation might be delayed */
 265                 return H_TOO_HARD;
 266
 267         mem = mm_iommu_lookup_rm(kvm->mm, be64_to_cpu(*pua), pgsize);
 268         if (!mem)
 269                 return H_TOO_HARD;
 270
 271         mm_iommu_mapped_dec(mem);
 272
 273         *pua = cpu_to_be64(0);
 274
 275         return H_SUCCESS;
 276 }
 277
 278 static long kvmppc_rm_tce_iommu_do_unmap(struct kvm *kvm,
 279                 struct iommu_table *tbl, unsigned long entry)
 280 {
 281         enum dma_data_direction dir = DMA_NONE;
 282         unsigned long hpa = 0;
 283         long ret;
 284
 285         if (iommu_tce_xchg_no_kill_rm(kvm->mm, tbl, entry, &hpa, &dir))
 286                 /*
 287                  * real mode xchg can fail if struct page crosses
 288                  * a page boundary
 289                  */
 290                 return H_TOO_HARD;
 291
 292         if (dir == DMA_NONE)
 293                 return H_SUCCESS;
 294
 295         ret = kvmppc_rm_tce_iommu_mapped_dec(kvm, tbl, entry);
 296         if (ret)
 297                 iommu_tce_xchg_no_kill_rm(kvm->mm, tbl, entry, &hpa, &dir);
 298
 299         return ret;
 300 }
 301
 302 static long kvmppc_rm_tce_iommu_unmap(struct kvm *kvm,
 303                 struct kvmppc_spapr_tce_table *stt, struct iommu_table *tbl,
 304                 unsigned long entry)
 305 {
 306         unsigned long i, ret = H_SUCCESS;
 307         unsigned long subpages = 1ULL << (stt->page_shift - tbl->it_page_shift);
 308         unsigned long io_entry = entry * subpages;
 309
 310         for (i = 0; i < subpages; ++i) {
 311                 ret = kvmppc_rm_tce_iommu_do_unmap(kvm, tbl, io_entry + i);
 312                 if (ret != H_SUCCESS)
 313                         break;
 314         }
 315
 316         return ret;
 317 }
 318
 319 static long kvmppc_rm_tce_iommu_do_map(struct kvm *kvm, struct iommu_table *tbl,
 320                 unsigned long entry, unsigned long ua,
 321                 enum dma_data_direction dir)
 322 {
 323         long ret;
 324         unsigned long hpa = 0;
 325         __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl, entry);
 326         struct mm_iommu_table_group_mem_t *mem;
 327
 328         if (!pua)
 329                 /* it_userspace allocation might be delayed */
 330                 return H_TOO_HARD;
 331
 332         mem = mm_iommu_lookup_rm(kvm->mm, ua, 1ULL << tbl->it_page_shift);
 333         if (!mem)
 334                 return H_TOO_HARD;
 335
 336         if (WARN_ON_ONCE_RM(mm_iommu_ua_to_hpa_rm(mem, ua, tbl->it_page_shift,
 337                         &hpa)))
 338                 return H_TOO_HARD;
 339
 340         if (WARN_ON_ONCE_RM(mm_iommu_mapped_inc(mem)))
 341                 return H_TOO_HARD;
 342
 343         ret = iommu_tce_xchg_no_kill_rm(kvm->mm, tbl, entry, &hpa, &dir);
 344         if (ret) {
 345                 mm_iommu_mapped_dec(mem);
 346                 /*
 347                  * real mode xchg can fail if struct page crosses
 348                  * a page boundary
 349                  */
 350                 return H_TOO_HARD;
 351         }
 352
 353         if (dir != DMA_NONE)
 354                 kvmppc_rm_tce_iommu_mapped_dec(kvm, tbl, entry);
 355
 356         *pua = cpu_to_be64(ua);
 357
 358         return 0;
 359 }
 360
 361 static long kvmppc_rm_tce_iommu_map(struct kvm *kvm,
 362                 struct kvmppc_spapr_tce_table *stt, struct iommu_table *tbl,
 363                 unsigned long entry, unsigned long ua,
 364                 enum dma_data_direction dir)
 365 {
 366         unsigned long i, pgoff, ret = H_SUCCESS;
 367         unsigned long subpages = 1ULL << (stt->page_shift - tbl->it_page_shift);
 368         unsigned long io_entry = entry * subpages;
 369
 370         for (i = 0, pgoff = 0; i < subpages;
 371                         ++i, pgoff += IOMMU_PAGE_SIZE(tbl)) {
 372
 373                 ret = kvmppc_rm_tce_iommu_do_map(kvm, tbl,
 374                                 io_entry + i, ua + pgoff, dir);
 375                 if (ret != H_SUCCESS)
 376                         break;
 377         }
 378
 379         return ret;
 380 }
 381
 382 long kvmppc_rm_h_put_tce(struct kvm_vcpu *vcpu, unsigned long liobn,
 383                 unsigned long ioba, unsigned long tce)
 384 {
 385         struct kvmppc_spapr_tce_table *stt;
 386         long ret;
 387         struct kvmppc_spapr_tce_iommu_table *stit;
 388         unsigned long entry, ua = 0;
 389         enum dma_data_direction dir;
 390
 391         /* udbg_printf("H_PUT_TCE(): liobn=0x%lx ioba=0x%lx, tce=0x%lx\n", */
 392         /*          liobn, ioba, tce); */
 393
 394         /* For radix, we might be in virtual mode, so punt */
 395         if (kvm_is_radix(vcpu->kvm))
 396                 return H_TOO_HARD;
 397
 398         stt = kvmppc_find_table(vcpu->kvm, liobn);
 399         if (!stt)
 400                 return H_TOO_HARD;
 401
 402         ret = kvmppc_rm_ioba_validate(stt, ioba, 1, tce == 0);
 403         if (ret != H_SUCCESS)
 404                 return ret;
 405
 406         ret = kvmppc_rm_tce_validate(stt, tce);
 407         if (ret != H_SUCCESS)
 408                 return ret;
 409
 410         dir = iommu_tce_direction(tce);
 411         if ((dir != DMA_NONE) && kvmppc_rm_tce_to_ua(vcpu->kvm, tce, &ua))
 412                 return H_PARAMETER;
 413
 414         entry = ioba >> stt->page_shift;
 415
 416         list_for_each_entry_lockless(stit, &stt->iommu_tables, next) {
 417                 if (dir == DMA_NONE)
 418                         ret = kvmppc_rm_tce_iommu_unmap(vcpu->kvm, stt,
 419                                         stit->tbl, entry);
 420                 else
 421                         ret = kvmppc_rm_tce_iommu_map(vcpu->kvm, stt,
 422                                         stit->tbl, entry, ua, dir);
 423
 424                 iommu_tce_kill_rm(stit->tbl, entry, 1);
 425
 426                 if (ret != H_SUCCESS) {
 427                         kvmppc_rm_clear_tce(vcpu->kvm, stit->tbl, entry);
 428                         return ret;
 429                 }
 430         }
 431
 432         kvmppc_rm_tce_put(stt, entry, tce);
 433
 434         return H_SUCCESS;
 435 }
 436
 437 static long kvmppc_rm_ua_to_hpa(struct kvm_vcpu *vcpu, unsigned long mmu_seq,
 438                                 unsigned long ua, unsigned long *phpa)
 439 {
 440         pte_t *ptep, pte;
 441         unsigned shift = 0;
 442
 443         /*
 444          * Called in real mode with MSR_EE = 0. We are safe here.
 445          * It is ok to do the lookup with arch.pgdir here, because
 446          * we are doing this on secondary cpus and current task there
 447          * is not the hypervisor. Also this is safe against THP in the
 448          * host, because an IPI to primary thread will wait for the secondary
 449          * to exit which will agains result in the below page table walk
 450          * to finish.
 451          */
 452         /* an rmap lock won't make it safe. because that just ensure hash
 453          * page table entries are removed with rmap lock held. After that
 454          * mmu notifier returns and we go ahead and removing ptes from Qemu page table.
 455          */
 456         ptep = find_kvm_host_pte(vcpu->kvm, mmu_seq, ua, &shift);
 457         if (!ptep)
 458                 return -ENXIO;
 459
 460         pte = READ_ONCE(*ptep);
 461         if (!pte_present(pte))
 462                 return -ENXIO;
 463
 464         if (!shift)
 465                 shift = PAGE_SHIFT;
 466
 467         /* Avoid handling anything potentially complicated in realmode */
 468         if (shift > PAGE_SHIFT)
 469                 return -EAGAIN;
 470
 471         if (!pte_young(pte))
 472                 return -EAGAIN;
 473
 474         *phpa = (pte_pfn(pte) << PAGE_SHIFT) | (ua & ((1ULL << shift) - 1)) |
 475                         (ua & ~PAGE_MASK);
 476
 477         return 0;
 478 }
 479
 480 long kvmppc_rm_h_put_tce_indirect(struct kvm_vcpu *vcpu,
 481                 unsigned long liobn, unsigned long ioba,
 482                 unsigned long tce_list, unsigned long npages)
 483 {
 484         struct kvm *kvm = vcpu->kvm;
 485         struct kvmppc_spapr_tce_table *stt;
 486         long i, ret = H_SUCCESS;
 487         unsigned long tces, entry, ua = 0;
 488         unsigned long mmu_seq;
 489         bool prereg = false;
 490         struct kvmppc_spapr_tce_iommu_table *stit;
 491
 492         /* For radix, we might be in virtual mode, so punt */
 493         if (kvm_is_radix(vcpu->kvm))
 494                 return H_TOO_HARD;
 495
 496         /*
 497          * used to check for invalidations in progress
 498          */
 499         mmu_seq = kvm->mmu_notifier_seq;
 500         smp_rmb();
 501
 502         stt = kvmppc_find_table(vcpu->kvm, liobn);
 503         if (!stt)
 504                 return H_TOO_HARD;
 505
 506         entry = ioba >> stt->page_shift;
 507         /*
 508          * The spec says that the maximum size of the list is 512 TCEs
 509          * so the whole table addressed resides in 4K page
 510          */
 511         if (npages > 512)
 512                 return H_PARAMETER;
 513
 514         if (tce_list & (SZ_4K - 1))
 515                 return H_PARAMETER;
 516
 517         ret = kvmppc_rm_ioba_validate(stt, ioba, npages, false);
 518         if (ret != H_SUCCESS)
 519                 return ret;
 520
 521         if (mm_iommu_preregistered(vcpu->kvm->mm)) {
 522                 /*
 523                  * We get here if guest memory was pre-registered which
 524                  * is normally VFIO case and gpa->hpa translation does not
 525                  * depend on hpt.
 526                  */
 527                 struct mm_iommu_table_group_mem_t *mem;
 528
 529                 if (kvmppc_rm_tce_to_ua(vcpu->kvm, tce_list, &ua))
 530                         return H_TOO_HARD;
 531
 532                 mem = mm_iommu_lookup_rm(vcpu->kvm->mm, ua, IOMMU_PAGE_SIZE_4K);
 533                 if (mem)
 534                         prereg = mm_iommu_ua_to_hpa_rm(mem, ua,
 535                                         IOMMU_PAGE_SHIFT_4K, &tces) == 0;
 536         }
 537
 538         if (!prereg) {
 539                 /*
 540                  * This is usually a case of a guest with emulated devices only
 541                  * when TCE list is not in preregistered memory.
 542                  * We do not require memory to be preregistered in this case
 543                  * so lock rmap and do __find_linux_pte_or_hugepte().
 544                  */
 545                 if (kvmppc_rm_tce_to_ua(vcpu->kvm, tce_list, &ua))
 546                         return H_TOO_HARD;
 547
 548                 arch_spin_lock(&kvm->mmu_lock.rlock.raw_lock);
 549                 if (kvmppc_rm_ua_to_hpa(vcpu, mmu_seq, ua, &tces)) {
 550                         ret = H_TOO_HARD;
 551                         goto unlock_exit;
 552                 }
 553         }
 554
 555         for (i = 0; i < npages; ++i) {
 556                 unsigned long tce = be64_to_cpu(((u64 *)tces)[i]);
 557
 558                 ret = kvmppc_rm_tce_validate(stt, tce);
 559                 if (ret != H_SUCCESS)
 560                         goto unlock_exit;
 561         }
 562
 563         for (i = 0; i < npages; ++i) {
 564                 unsigned long tce = be64_to_cpu(((u64 *)tces)[i]);
 565
 566                 ua = 0;
 567                 if (kvmppc_rm_tce_to_ua(vcpu->kvm, tce, &ua)) {
 568                         ret = H_PARAMETER;
 569                         goto invalidate_exit;
 570                 }
 571
 572                 list_for_each_entry_lockless(stit, &stt->iommu_tables, next) {
 573                         ret = kvmppc_rm_tce_iommu_map(vcpu->kvm, stt,
 574                                         stit->tbl, entry + i, ua,
 575                                         iommu_tce_direction(tce));
 576
 577                         if (ret != H_SUCCESS) {
 578                                 kvmppc_rm_clear_tce(vcpu->kvm, stit->tbl,
 579                                                 entry);
 580                                 goto invalidate_exit;
 581                         }
 582                 }
 583
 584                 kvmppc_rm_tce_put(stt, entry + i, tce);
 585         }
 586
 587 invalidate_exit:
 588         list_for_each_entry_lockless(stit, &stt->iommu_tables, next)
 589                 iommu_tce_kill_rm(stit->tbl, entry, npages);
 590
 591 unlock_exit:
 592         if (!prereg)
 593                 arch_spin_unlock(&kvm->mmu_lock.rlock.raw_lock);
 594         return ret;
 595 }
 596
 597 long kvmppc_rm_h_stuff_tce(struct kvm_vcpu *vcpu,
 598                 unsigned long liobn, unsigned long ioba,
 599                 unsigned long tce_value, unsigned long npages)
 600 {
 601         struct kvmppc_spapr_tce_table *stt;
 602         long i, ret;
 603         struct kvmppc_spapr_tce_iommu_table *stit;
 604
 605         /* For radix, we might be in virtual mode, so punt */
 606         if (kvm_is_radix(vcpu->kvm))
 607                 return H_TOO_HARD;
 608
 609         stt = kvmppc_find_table(vcpu->kvm, liobn);
 610         if (!stt)
 611                 return H_TOO_HARD;
 612
 613         ret = kvmppc_rm_ioba_validate(stt, ioba, npages, tce_value == 0);
 614         if (ret != H_SUCCESS)
 615                 return ret;
 616
 617         /* Check permission bits only to allow userspace poison TCE for debug */
 618         if (tce_value & (TCE_PCI_WRITE | TCE_PCI_READ))
 619                 return H_PARAMETER;
 620
 621         list_for_each_entry_lockless(stit, &stt->iommu_tables, next) {
 622                 unsigned long entry = ioba >> stt->page_shift;
 623
 624                 for (i = 0; i < npages; ++i) {
 625                         ret = kvmppc_rm_tce_iommu_unmap(vcpu->kvm, stt,
 626                                         stit->tbl, entry + i);
 627
 628                         if (ret == H_SUCCESS)
 629                                 continue;
 630
 631                         if (ret == H_TOO_HARD)
 632                                 goto invalidate_exit;
 633
 634                         WARN_ON_ONCE_RM(1);
 635                         kvmppc_rm_clear_tce(vcpu->kvm, stit->tbl, entry);
 636                 }
 637         }
 638
 639         for (i = 0; i < npages; ++i, ioba += (1ULL << stt->page_shift))
 640                 kvmppc_rm_tce_put(stt, ioba >> stt->page_shift, tce_value);
 641
 642 invalidate_exit:
 643         list_for_each_entry_lockless(stit, &stt->iommu_tables, next)
 644                 iommu_tce_kill_rm(stit->tbl, ioba >> stt->page_shift, npages);
 645
 646         return ret;
 647 }
 648
 649 /* This can be called in either virtual mode or real mode */
 650 long kvmppc_h_get_tce(struct kvm_vcpu *vcpu, unsigned long liobn,
 651                       unsigned long ioba)
 652 {
 653         struct kvmppc_spapr_tce_table *stt;
 654         long ret;
 655         unsigned long idx;
 656         struct page *page;
 657         u64 *tbl;
 658
 659         stt = kvmppc_find_table(vcpu->kvm, liobn);
 660         if (!stt)
 661                 return H_TOO_HARD;
 662
 663         ret = kvmppc_ioba_validate(stt, ioba, 1);
 664         if (ret != H_SUCCESS)
 665                 return ret;
 666
 667         idx = (ioba >> stt->page_shift) - stt->offset;
 668         page = stt->pages[idx / TCES_PER_PAGE];
 669         if (!page) {
 670                 vcpu->arch.regs.gpr[4] = 0;
 671                 return H_SUCCESS;
 672         }
 673         tbl = (u64 *)page_address(page);
 674
 675         vcpu->arch.regs.gpr[4] = tbl[idx % TCES_PER_PAGE];
 676
 677         return H_SUCCESS;
 678 }
 679 EXPORT_SYMBOL_GPL(kvmppc_h_get_tce);
 680
 681 #endif /* KVM_BOOK3S_HV_POSSIBLE */