arch/mips/kvm/mmu.c

   1 /*
   2  * This file is subject to the terms and conditions of the GNU General Public
   3  * License.  See the file "COPYING" in the main directory of this archive
   4  * for more details.
   5  *
   6  * KVM/MIPS MMU handling in the KVM module.
   7  *
   8  * Copyright (C) 2012  MIPS Technologies, Inc.  All rights reserved.
   9  * Authors: Sanjay Lal <sanjayl@kymasys.com>
  10  */
  11
  12 #include <linux/highmem.h>
  13 #include <linux/kvm_host.h>
  14 #include <linux/uaccess.h>
  15 #include <asm/mmu_context.h>
  16 #include <asm/pgalloc.h>
  17
  18 /*
  19  * KVM_MMU_CACHE_MIN_PAGES is the number of GPA page table translation levels
  20  * for which pages need to be cached.
  21  */
  22 #if defined(__PAGETABLE_PMD_FOLDED)
  23 #define KVM_MMU_CACHE_MIN_PAGES 1
  24 #else
  25 #define KVM_MMU_CACHE_MIN_PAGES 2
  26 #endif
  27
  28 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
  29 {
  30         kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
  31 }
  32
  33 /**
  34  * kvm_pgd_init() - Initialise KVM GPA page directory.
  35  * @page:       Pointer to page directory (PGD) for KVM GPA.
  36  *
  37  * Initialise a KVM GPA page directory with pointers to the invalid table, i.e.
  38  * representing no mappings. This is similar to pgd_init(), however it
  39  * initialises all the page directory pointers, not just the ones corresponding
  40  * to the userland address space (since it is for the guest physical address
  41  * space rather than a virtual address space).
  42  */
  43 static void kvm_pgd_init(void *page)
  44 {
  45         unsigned long *p, *end;
  46         unsigned long entry;
  47
  48 #ifdef __PAGETABLE_PMD_FOLDED
  49         entry = (unsigned long)invalid_pte_table;
  50 #else
  51         entry = (unsigned long)invalid_pmd_table;
  52 #endif
  53
  54         p = (unsigned long *)page;
  55         end = p + PTRS_PER_PGD;
  56
  57         do {
  58                 p[0] = entry;
  59                 p[1] = entry;
  60                 p[2] = entry;
  61                 p[3] = entry;
  62                 p[4] = entry;
  63                 p += 8;
  64                 p[-3] = entry;
  65                 p[-2] = entry;
  66                 p[-1] = entry;
  67         } while (p != end);
  68 }
  69
  70 /**
  71  * kvm_pgd_alloc() - Allocate and initialise a KVM GPA page directory.
  72  *
  73  * Allocate a blank KVM GPA page directory (PGD) for representing guest physical
  74  * to host physical page mappings.
  75  *
  76  * Returns:     Pointer to new KVM GPA page directory.
  77  *              NULL on allocation failure.
  78  */
  79 pgd_t *kvm_pgd_alloc(void)
  80 {
  81         pgd_t *ret;
  82
  83         ret = (pgd_t *)__get_free_pages(GFP_KERNEL, PGD_ORDER);
  84         if (ret)
  85                 kvm_pgd_init(ret);
  86
  87         return ret;
  88 }
  89
  90 /**
  91  * kvm_mips_walk_pgd() - Walk page table with optional allocation.
  92  * @pgd:        Page directory pointer.
  93  * @addr:       Address to index page table using.
  94  * @cache:      MMU page cache to allocate new page tables from, or NULL.
  95  *
  96  * Walk the page tables pointed to by @pgd to find the PTE corresponding to the
  97  * address @addr. If page tables don't exist for @addr, they will be created
  98  * from the MMU cache if @cache is not NULL.
  99  *
 100  * Returns:     Pointer to pte_t corresponding to @addr.
 101  *              NULL if a page table doesn't exist for @addr and !@cache.
 102  *              NULL if a page table allocation failed.
 103  */
 104 static pte_t *kvm_mips_walk_pgd(pgd_t *pgd, struct kvm_mmu_memory_cache *cache,
 105                                 unsigned long addr)
 106 {
 107         p4d_t *p4d;
 108         pud_t *pud;
 109         pmd_t *pmd;
 110
 111         pgd += pgd_index(addr);
 112         if (pgd_none(*pgd)) {
 113                 /* Not used on MIPS yet */
 114                 BUG();
 115                 return NULL;
 116         }
 117         p4d = p4d_offset(pgd, addr);
 118         pud = pud_offset(p4d, addr);
 119         if (pud_none(*pud)) {
 120                 pmd_t *new_pmd;
 121
 122                 if (!cache)
 123                         return NULL;
 124                 new_pmd = kvm_mmu_memory_cache_alloc(cache);
 125                 pmd_init((unsigned long)new_pmd,
 126                          (unsigned long)invalid_pte_table);
 127                 pud_populate(NULL, pud, new_pmd);
 128         }
 129         pmd = pmd_offset(pud, addr);
 130         if (pmd_none(*pmd)) {
 131                 pte_t *new_pte;
 132
 133                 if (!cache)
 134                         return NULL;
 135                 new_pte = kvm_mmu_memory_cache_alloc(cache);
 136                 clear_page(new_pte);
 137                 pmd_populate_kernel(NULL, pmd, new_pte);
 138         }
 139         return pte_offset_kernel(pmd, addr);
 140 }
 141
 142 /* Caller must hold kvm->mm_lock */
 143 static pte_t *kvm_mips_pte_for_gpa(struct kvm *kvm,
 144                                    struct kvm_mmu_memory_cache *cache,
 145                                    unsigned long addr)
 146 {
 147         return kvm_mips_walk_pgd(kvm->arch.gpa_mm.pgd, cache, addr);
 148 }
 149
 150 /*
 151  * kvm_mips_flush_gpa_{pte,pmd,pud,pgd,pt}.
 152  * Flush a range of guest physical address space from the VM's GPA page tables.
 153  */
 154
 155 static bool kvm_mips_flush_gpa_pte(pte_t *pte, unsigned long start_gpa,
 156                                    unsigned long end_gpa)
 157 {
 158         int i_min = pte_index(start_gpa);
 159         int i_max = pte_index(end_gpa);
 160         bool safe_to_remove = (i_min == 0 && i_max == PTRS_PER_PTE - 1);
 161         int i;
 162
 163         for (i = i_min; i <= i_max; ++i) {
 164                 if (!pte_present(pte[i]))
 165                         continue;
 166
 167                 set_pte(pte + i, __pte(0));
 168         }
 169         return safe_to_remove;
 170 }
 171
 172 static bool kvm_mips_flush_gpa_pmd(pmd_t *pmd, unsigned long start_gpa,
 173                                    unsigned long end_gpa)
 174 {
 175         pte_t *pte;
 176         unsigned long end = ~0ul;
 177         int i_min = pmd_index(start_gpa);
 178         int i_max = pmd_index(end_gpa);
 179         bool safe_to_remove = (i_min == 0 && i_max == PTRS_PER_PMD - 1);
 180         int i;
 181
 182         for (i = i_min; i <= i_max; ++i, start_gpa = 0) {
 183                 if (!pmd_present(pmd[i]))
 184                         continue;
 185
 186                 pte = pte_offset_kernel(pmd + i, 0);
 187                 if (i == i_max)
 188                         end = end_gpa;
 189
 190                 if (kvm_mips_flush_gpa_pte(pte, start_gpa, end)) {
 191                         pmd_clear(pmd + i);
 192                         pte_free_kernel(NULL, pte);
 193                 } else {
 194                         safe_to_remove = false;
 195                 }
 196         }
 197         return safe_to_remove;
 198 }
 199
 200 static bool kvm_mips_flush_gpa_pud(pud_t *pud, unsigned long start_gpa,
 201                                    unsigned long end_gpa)
 202 {
 203         pmd_t *pmd;
 204         unsigned long end = ~0ul;
 205         int i_min = pud_index(start_gpa);
 206         int i_max = pud_index(end_gpa);
 207         bool safe_to_remove = (i_min == 0 && i_max == PTRS_PER_PUD - 1);
 208         int i;
 209
 210         for (i = i_min; i <= i_max; ++i, start_gpa = 0) {
 211                 if (!pud_present(pud[i]))
 212                         continue;
 213
 214                 pmd = pmd_offset(pud + i, 0);
 215                 if (i == i_max)
 216                         end = end_gpa;
 217
 218                 if (kvm_mips_flush_gpa_pmd(pmd, start_gpa, end)) {
 219                         pud_clear(pud + i);
 220                         pmd_free(NULL, pmd);
 221                 } else {
 222                         safe_to_remove = false;
 223                 }
 224         }
 225         return safe_to_remove;
 226 }
 227
 228 static bool kvm_mips_flush_gpa_pgd(pgd_t *pgd, unsigned long start_gpa,
 229                                    unsigned long end_gpa)
 230 {
 231         p4d_t *p4d;
 232         pud_t *pud;
 233         unsigned long end = ~0ul;
 234         int i_min = pgd_index(start_gpa);
 235         int i_max = pgd_index(end_gpa);
 236         bool safe_to_remove = (i_min == 0 && i_max == PTRS_PER_PGD - 1);
 237         int i;
 238
 239         for (i = i_min; i <= i_max; ++i, start_gpa = 0) {
 240                 if (!pgd_present(pgd[i]))
 241                         continue;
 242
 243                 p4d = p4d_offset(pgd, 0);
 244                 pud = pud_offset(p4d + i, 0);
 245                 if (i == i_max)
 246                         end = end_gpa;
 247
 248                 if (kvm_mips_flush_gpa_pud(pud, start_gpa, end)) {
 249                         pgd_clear(pgd + i);
 250                         pud_free(NULL, pud);
 251                 } else {
 252                         safe_to_remove = false;
 253                 }
 254         }
 255         return safe_to_remove;
 256 }
 257
 258 /**
 259  * kvm_mips_flush_gpa_pt() - Flush a range of guest physical addresses.
 260  * @kvm:        KVM pointer.
 261  * @start_gfn:  Guest frame number of first page in GPA range to flush.
 262  * @end_gfn:    Guest frame number of last page in GPA range to flush.
 263  *
 264  * Flushes a range of GPA mappings from the GPA page tables.
 265  *
 266  * The caller must hold the @kvm->mmu_lock spinlock.
 267  *
 268  * Returns:     Whether its safe to remove the top level page directory because
 269  *              all lower levels have been removed.
 270  */
 271 bool kvm_mips_flush_gpa_pt(struct kvm *kvm, gfn_t start_gfn, gfn_t end_gfn)
 272 {
 273         return kvm_mips_flush_gpa_pgd(kvm->arch.gpa_mm.pgd,
 274                                       start_gfn << PAGE_SHIFT,
 275                                       end_gfn << PAGE_SHIFT);
 276 }
 277
 278 #define BUILD_PTE_RANGE_OP(name, op)                                    \
 279 static int kvm_mips_##name##_pte(pte_t *pte, unsigned long start,       \
 280                                  unsigned long end)                     \
 281 {                                                                       \
 282         int ret = 0;                                                    \
 283         int i_min = pte_index(start);                           \
 284         int i_max = pte_index(end);                                     \
 285         int i;                                                          \
 286         pte_t old, new;                                                 \
 287                                                                         \
 288         for (i = i_min; i <= i_max; ++i) {                              \
 289                 if (!pte_present(pte[i]))                               \
 290                         continue;                                       \
 291                                                                         \
 292                 old = pte[i];                                           \
 293                 new = op(old);                                          \
 294                 if (pte_val(new) == pte_val(old))                       \
 295                         continue;                                       \
 296                 set_pte(pte + i, new);                                  \
 297                 ret = 1;                                                \
 298         }                                                               \
 299         return ret;                                                     \
 300 }                                                                       \
 301                                                                         \
 302 /* returns true if anything was done */                                 \
 303 static int kvm_mips_##name##_pmd(pmd_t *pmd, unsigned long start,       \
 304                                  unsigned long end)                     \
 305 {                                                                       \
 306         int ret = 0;                                                    \
 307         pte_t *pte;                                                     \
 308         unsigned long cur_end = ~0ul;                                   \
 309         int i_min = pmd_index(start);                           \
 310         int i_max = pmd_index(end);                                     \
 311         int i;                                                          \
 312                                                                         \
 313         for (i = i_min; i <= i_max; ++i, start = 0) {                   \
 314                 if (!pmd_present(pmd[i]))                               \
 315                         continue;                                       \
 316                                                                         \
 317                 pte = pte_offset_kernel(pmd + i, 0);                            \
 318                 if (i == i_max)                                         \
 319                         cur_end = end;                                  \
 320                                                                         \
 321                 ret |= kvm_mips_##name##_pte(pte, start, cur_end);      \
 322         }                                                               \
 323         return ret;                                                     \
 324 }                                                                       \
 325                                                                         \
 326 static int kvm_mips_##name##_pud(pud_t *pud, unsigned long start,       \
 327                                  unsigned long end)                     \
 328 {                                                                       \
 329         int ret = 0;                                                    \
 330         pmd_t *pmd;                                                     \
 331         unsigned long cur_end = ~0ul;                                   \
 332         int i_min = pud_index(start);                           \
 333         int i_max = pud_index(end);                                     \
 334         int i;                                                          \
 335                                                                         \
 336         for (i = i_min; i <= i_max; ++i, start = 0) {                   \
 337                 if (!pud_present(pud[i]))                               \
 338                         continue;                                       \
 339                                                                         \
 340                 pmd = pmd_offset(pud + i, 0);                           \
 341                 if (i == i_max)                                         \
 342                         cur_end = end;                                  \
 343                                                                         \
 344                 ret |= kvm_mips_##name##_pmd(pmd, start, cur_end);      \
 345         }                                                               \
 346         return ret;                                                     \
 347 }                                                                       \
 348                                                                         \
 349 static int kvm_mips_##name##_pgd(pgd_t *pgd, unsigned long start,       \
 350                                  unsigned long end)                     \
 351 {                                                                       \
 352         int ret = 0;                                                    \
 353         p4d_t *p4d;                                                     \
 354         pud_t *pud;                                                     \
 355         unsigned long cur_end = ~0ul;                                   \
 356         int i_min = pgd_index(start);                                   \
 357         int i_max = pgd_index(end);                                     \
 358         int i;                                                          \
 359                                                                         \
 360         for (i = i_min; i <= i_max; ++i, start = 0) {                   \
 361                 if (!pgd_present(pgd[i]))                               \
 362                         continue;                                       \
 363                                                                         \
 364                 p4d = p4d_offset(pgd, 0);                               \
 365                 pud = pud_offset(p4d + i, 0);                           \
 366                 if (i == i_max)                                         \
 367                         cur_end = end;                                  \
 368                                                                         \
 369                 ret |= kvm_mips_##name##_pud(pud, start, cur_end);      \
 370         }                                                               \
 371         return ret;                                                     \
 372 }
 373
 374 /*
 375  * kvm_mips_mkclean_gpa_pt.
 376  * Mark a range of guest physical address space clean (writes fault) in the VM's
 377  * GPA page table to allow dirty page tracking.
 378  */
 379
 380 BUILD_PTE_RANGE_OP(mkclean, pte_mkclean)
 381
 382 /**
 383  * kvm_mips_mkclean_gpa_pt() - Make a range of guest physical addresses clean.
 384  * @kvm:        KVM pointer.
 385  * @start_gfn:  Guest frame number of first page in GPA range to flush.
 386  * @end_gfn:    Guest frame number of last page in GPA range to flush.
 387  *
 388  * Make a range of GPA mappings clean so that guest writes will fault and
 389  * trigger dirty page logging.
 390  *
 391  * The caller must hold the @kvm->mmu_lock spinlock.
 392  *
 393  * Returns:     Whether any GPA mappings were modified, which would require
 394  *              derived mappings (GVA page tables & TLB enties) to be
 395  *              invalidated.
 396  */
 397 int kvm_mips_mkclean_gpa_pt(struct kvm *kvm, gfn_t start_gfn, gfn_t end_gfn)
 398 {
 399         return kvm_mips_mkclean_pgd(kvm->arch.gpa_mm.pgd,
 400                                     start_gfn << PAGE_SHIFT,
 401                                     end_gfn << PAGE_SHIFT);
 402 }
 403
 404 /**
 405  * kvm_arch_mmu_enable_log_dirty_pt_masked() - write protect dirty pages
 406  * @kvm:        The KVM pointer
 407  * @slot:       The memory slot associated with mask
 408  * @gfn_offset: The gfn offset in memory slot
 409  * @mask:       The mask of dirty pages at offset 'gfn_offset' in this memory
 410  *              slot to be write protected
 411  *
 412  * Walks bits set in mask write protects the associated pte's. Caller must
 413  * acquire @kvm->mmu_lock.
 414  */
 415 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
 416                 struct kvm_memory_slot *slot,
 417                 gfn_t gfn_offset, unsigned long mask)
 418 {
 419         gfn_t base_gfn = slot->base_gfn + gfn_offset;
 420         gfn_t start = base_gfn +  __ffs(mask);
 421         gfn_t end = base_gfn + __fls(mask);
 422
 423         kvm_mips_mkclean_gpa_pt(kvm, start, end);
 424 }
 425
 426 /*
 427  * kvm_mips_mkold_gpa_pt.
 428  * Mark a range of guest physical address space old (all accesses fault) in the
 429  * VM's GPA page table to allow detection of commonly used pages.
 430  */
 431
 432 BUILD_PTE_RANGE_OP(mkold, pte_mkold)
 433
 434 static int kvm_mips_mkold_gpa_pt(struct kvm *kvm, gfn_t start_gfn,
 435                                  gfn_t end_gfn)
 436 {
 437         return kvm_mips_mkold_pgd(kvm->arch.gpa_mm.pgd,
 438                                   start_gfn << PAGE_SHIFT,
 439                                   end_gfn << PAGE_SHIFT);
 440 }
 441
 442 static int handle_hva_to_gpa(struct kvm *kvm,
 443                              unsigned long start,
 444                              unsigned long end,
 445                              int (*handler)(struct kvm *kvm, gfn_t gfn,
 446                                             gpa_t gfn_end,
 447                                             struct kvm_memory_slot *memslot,
 448                                             void *data),
 449                              void *data)
 450 {
 451         struct kvm_memslots *slots;
 452         struct kvm_memory_slot *memslot;
 453         int ret = 0;
 454
 455         slots = kvm_memslots(kvm);
 456
 457         /* we only care about the pages that the guest sees */
 458         kvm_for_each_memslot(memslot, slots) {
 459                 unsigned long hva_start, hva_end;
 460                 gfn_t gfn, gfn_end;
 461
 462                 hva_start = max(start, memslot->userspace_addr);
 463                 hva_end = min(end, memslot->userspace_addr +
 464                                         (memslot->npages << PAGE_SHIFT));
 465                 if (hva_start >= hva_end)
 466                         continue;
 467
 468                 /*
 469                  * {gfn(page) | page intersects with [hva_start, hva_end)} =
 470                  * {gfn_start, gfn_start+1, ..., gfn_end-1}.
 471                  */
 472                 gfn = hva_to_gfn_memslot(hva_start, memslot);
 473                 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
 474
 475                 ret |= handler(kvm, gfn, gfn_end, memslot, data);
 476         }
 477
 478         return ret;
 479 }
 480
 481
 482 static int kvm_unmap_hva_handler(struct kvm *kvm, gfn_t gfn, gfn_t gfn_end,
 483                                  struct kvm_memory_slot *memslot, void *data)
 484 {
 485         kvm_mips_flush_gpa_pt(kvm, gfn, gfn_end);
 486         return 1;
 487 }
 488
 489 int kvm_unmap_hva_range(struct kvm *kvm, unsigned long start, unsigned long end,
 490                         unsigned flags)
 491 {
 492         handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
 493
 494         kvm_mips_callbacks->flush_shadow_all(kvm);
 495         return 0;
 496 }
 497
 498 static int kvm_set_spte_handler(struct kvm *kvm, gfn_t gfn, gfn_t gfn_end,
 499                                 struct kvm_memory_slot *memslot, void *data)
 500 {
 501         gpa_t gpa = gfn << PAGE_SHIFT;
 502         pte_t hva_pte = *(pte_t *)data;
 503         pte_t *gpa_pte = kvm_mips_pte_for_gpa(kvm, NULL, gpa);
 504         pte_t old_pte;
 505
 506         if (!gpa_pte)
 507                 return 0;
 508
 509         /* Mapping may need adjusting depending on memslot flags */
 510         old_pte = *gpa_pte;
 511         if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES && !pte_dirty(old_pte))
 512                 hva_pte = pte_mkclean(hva_pte);
 513         else if (memslot->flags & KVM_MEM_READONLY)
 514                 hva_pte = pte_wrprotect(hva_pte);
 515
 516         set_pte(gpa_pte, hva_pte);
 517
 518         /* Replacing an absent or old page doesn't need flushes */
 519         if (!pte_present(old_pte) || !pte_young(old_pte))
 520                 return 0;
 521
 522         /* Pages swapped, aged, moved, or cleaned require flushes */
 523         return !pte_present(hva_pte) ||
 524                !pte_young(hva_pte) ||
 525                pte_pfn(old_pte) != pte_pfn(hva_pte) ||
 526                (pte_dirty(old_pte) && !pte_dirty(hva_pte));
 527 }
 528
 529 int kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
 530 {
 531         unsigned long end = hva + PAGE_SIZE;
 532         int ret;
 533
 534         ret = handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &pte);
 535         if (ret)
 536                 kvm_mips_callbacks->flush_shadow_all(kvm);
 537         return 0;
 538 }
 539
 540 static int kvm_age_hva_handler(struct kvm *kvm, gfn_t gfn, gfn_t gfn_end,
 541                                struct kvm_memory_slot *memslot, void *data)
 542 {
 543         return kvm_mips_mkold_gpa_pt(kvm, gfn, gfn_end);
 544 }
 545
 546 static int kvm_test_age_hva_handler(struct kvm *kvm, gfn_t gfn, gfn_t gfn_end,
 547                                     struct kvm_memory_slot *memslot, void *data)
 548 {
 549         gpa_t gpa = gfn << PAGE_SHIFT;
 550         pte_t *gpa_pte = kvm_mips_pte_for_gpa(kvm, NULL, gpa);
 551
 552         if (!gpa_pte)
 553                 return 0;
 554         return pte_young(*gpa_pte);
 555 }
 556
 557 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
 558 {
 559         return handle_hva_to_gpa(kvm, start, end, kvm_age_hva_handler, NULL);
 560 }
 561
 562 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
 563 {
 564         return handle_hva_to_gpa(kvm, hva, hva, kvm_test_age_hva_handler, NULL);
 565 }
 566
 567 /**
 568  * _kvm_mips_map_page_fast() - Fast path GPA fault handler.
 569  * @vcpu:               VCPU pointer.
 570  * @gpa:                Guest physical address of fault.
 571  * @write_fault:        Whether the fault was due to a write.
 572  * @out_entry:          New PTE for @gpa (written on success unless NULL).
 573  * @out_buddy:          New PTE for @gpa's buddy (written on success unless
 574  *                      NULL).
 575  *
 576  * Perform fast path GPA fault handling, doing all that can be done without
 577  * calling into KVM. This handles marking old pages young (for idle page
 578  * tracking), and dirtying of clean pages (for dirty page logging).
 579  *
 580  * Returns:     0 on success, in which case we can update derived mappings and
 581  *              resume guest execution.
 582  *              -EFAULT on failure due to absent GPA mapping or write to
 583  *              read-only page, in which case KVM must be consulted.
 584  */
 585 static int _kvm_mips_map_page_fast(struct kvm_vcpu *vcpu, unsigned long gpa,
 586                                    bool write_fault,
 587                                    pte_t *out_entry, pte_t *out_buddy)
 588 {
 589         struct kvm *kvm = vcpu->kvm;
 590         gfn_t gfn = gpa >> PAGE_SHIFT;
 591         pte_t *ptep;
 592         kvm_pfn_t pfn = 0;      /* silence bogus GCC warning */
 593         bool pfn_valid = false;
 594         int ret = 0;
 595
 596         spin_lock(&kvm->mmu_lock);
 597
 598         /* Fast path - just check GPA page table for an existing entry */
 599         ptep = kvm_mips_pte_for_gpa(kvm, NULL, gpa);
 600         if (!ptep || !pte_present(*ptep)) {
 601                 ret = -EFAULT;
 602                 goto out;
 603         }
 604
 605         /* Track access to pages marked old */
 606         if (!pte_young(*ptep)) {
 607                 set_pte(ptep, pte_mkyoung(*ptep));
 608                 pfn = pte_pfn(*ptep);
 609                 pfn_valid = true;
 610                 /* call kvm_set_pfn_accessed() after unlock */
 611         }
 612         if (write_fault && !pte_dirty(*ptep)) {
 613                 if (!pte_write(*ptep)) {
 614                         ret = -EFAULT;
 615                         goto out;
 616                 }
 617
 618                 /* Track dirtying of writeable pages */
 619                 set_pte(ptep, pte_mkdirty(*ptep));
 620                 pfn = pte_pfn(*ptep);
 621                 mark_page_dirty(kvm, gfn);
 622                 kvm_set_pfn_dirty(pfn);
 623         }
 624
 625         if (out_entry)
 626                 *out_entry = *ptep;
 627         if (out_buddy)
 628                 *out_buddy = *ptep_buddy(ptep);
 629
 630 out:
 631         spin_unlock(&kvm->mmu_lock);
 632         if (pfn_valid)
 633                 kvm_set_pfn_accessed(pfn);
 634         return ret;
 635 }
 636
 637 /**
 638  * kvm_mips_map_page() - Map a guest physical page.
 639  * @vcpu:               VCPU pointer.
 640  * @gpa:                Guest physical address of fault.
 641  * @write_fault:        Whether the fault was due to a write.
 642  * @out_entry:          New PTE for @gpa (written on success unless NULL).
 643  * @out_buddy:          New PTE for @gpa's buddy (written on success unless
 644  *                      NULL).
 645  *
 646  * Handle GPA faults by creating a new GPA mapping (or updating an existing
 647  * one).
 648  *
 649  * This takes care of marking pages young or dirty (idle/dirty page tracking),
 650  * asking KVM for the corresponding PFN, and creating a mapping in the GPA page
 651  * tables. Derived mappings (GVA page tables and TLBs) must be handled by the
 652  * caller.
 653  *
 654  * Returns:     0 on success, in which case the caller may use the @out_entry
 655  *              and @out_buddy PTEs to update derived mappings and resume guest
 656  *              execution.
 657  *              -EFAULT if there is no memory region at @gpa or a write was
 658  *              attempted to a read-only memory region. This is usually handled
 659  *              as an MMIO access.
 660  */
 661 static int kvm_mips_map_page(struct kvm_vcpu *vcpu, unsigned long gpa,
 662                              bool write_fault,
 663                              pte_t *out_entry, pte_t *out_buddy)
 664 {
 665         struct kvm *kvm = vcpu->kvm;
 666         struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
 667         gfn_t gfn = gpa >> PAGE_SHIFT;
 668         int srcu_idx, err;
 669         kvm_pfn_t pfn;
 670         pte_t *ptep, entry;
 671         bool writeable;
 672         unsigned long prot_bits;
 673         unsigned long mmu_seq;
 674
 675         /* Try the fast path to handle old / clean pages */
 676         srcu_idx = srcu_read_lock(&kvm->srcu);
 677         err = _kvm_mips_map_page_fast(vcpu, gpa, write_fault, out_entry,
 678                                       out_buddy);
 679         if (!err)
 680                 goto out;
 681
 682         /* We need a minimum of cached pages ready for page table creation */
 683         err = kvm_mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES);
 684         if (err)
 685                 goto out;
 686
 687 retry:
 688         /*
 689          * Used to check for invalidations in progress, of the pfn that is
 690          * returned by pfn_to_pfn_prot below.
 691          */
 692         mmu_seq = kvm->mmu_notifier_seq;
 693         /*
 694          * Ensure the read of mmu_notifier_seq isn't reordered with PTE reads in
 695          * gfn_to_pfn_prot() (which calls get_user_pages()), so that we don't
 696          * risk the page we get a reference to getting unmapped before we have a
 697          * chance to grab the mmu_lock without mmu_notifier_retry() noticing.
 698          *
 699          * This smp_rmb() pairs with the effective smp_wmb() of the combination
 700          * of the pte_unmap_unlock() after the PTE is zapped, and the
 701          * spin_lock() in kvm_mmu_notifier_invalidate_<page|range_end>() before
 702          * mmu_notifier_seq is incremented.
 703          */
 704         smp_rmb();
 705
 706         /* Slow path - ask KVM core whether we can access this GPA */
 707         pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writeable);
 708         if (is_error_noslot_pfn(pfn)) {
 709                 err = -EFAULT;
 710                 goto out;
 711         }
 712
 713         spin_lock(&kvm->mmu_lock);
 714         /* Check if an invalidation has taken place since we got pfn */
 715         if (mmu_notifier_retry(kvm, mmu_seq)) {
 716                 /*
 717                  * This can happen when mappings are changed asynchronously, but
 718                  * also synchronously if a COW is triggered by
 719                  * gfn_to_pfn_prot().
 720                  */
 721                 spin_unlock(&kvm->mmu_lock);
 722                 kvm_release_pfn_clean(pfn);
 723                 goto retry;
 724         }
 725
 726         /* Ensure page tables are allocated */
 727         ptep = kvm_mips_pte_for_gpa(kvm, memcache, gpa);
 728
 729         /* Set up the PTE */
 730         prot_bits = _PAGE_PRESENT | __READABLE | _page_cachable_default;
 731         if (writeable) {
 732                 prot_bits |= _PAGE_WRITE;
 733                 if (write_fault) {
 734                         prot_bits |= __WRITEABLE;
 735                         mark_page_dirty(kvm, gfn);
 736                         kvm_set_pfn_dirty(pfn);
 737                 }
 738         }
 739         entry = pfn_pte(pfn, __pgprot(prot_bits));
 740
 741         /* Write the PTE */
 742         set_pte(ptep, entry);
 743
 744         err = 0;
 745         if (out_entry)
 746                 *out_entry = *ptep;
 747         if (out_buddy)
 748                 *out_buddy = *ptep_buddy(ptep);
 749
 750         spin_unlock(&kvm->mmu_lock);
 751         kvm_release_pfn_clean(pfn);
 752         kvm_set_pfn_accessed(pfn);
 753 out:
 754         srcu_read_unlock(&kvm->srcu, srcu_idx);
 755         return err;
 756 }
 757
 758 static pte_t *kvm_trap_emul_pte_for_gva(struct kvm_vcpu *vcpu,
 759                                         unsigned long addr)
 760 {
 761         struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
 762         pgd_t *pgdp;
 763         int ret;
 764
 765         /* We need a minimum of cached pages ready for page table creation */
 766         ret = kvm_mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES);
 767         if (ret)
 768                 return NULL;
 769
 770         if (KVM_GUEST_KERNEL_MODE(vcpu))
 771                 pgdp = vcpu->arch.guest_kernel_mm.pgd;
 772         else
 773                 pgdp = vcpu->arch.guest_user_mm.pgd;
 774
 775         return kvm_mips_walk_pgd(pgdp, memcache, addr);
 776 }
 777
 778 void kvm_trap_emul_invalidate_gva(struct kvm_vcpu *vcpu, unsigned long addr,
 779                                   bool user)
 780 {
 781         pgd_t *pgdp;
 782         pte_t *ptep;
 783
 784         addr &= PAGE_MASK << 1;
 785
 786         pgdp = vcpu->arch.guest_kernel_mm.pgd;
 787         ptep = kvm_mips_walk_pgd(pgdp, NULL, addr);
 788         if (ptep) {
 789                 ptep[0] = pfn_pte(0, __pgprot(0));
 790                 ptep[1] = pfn_pte(0, __pgprot(0));
 791         }
 792
 793         if (user) {
 794                 pgdp = vcpu->arch.guest_user_mm.pgd;
 795                 ptep = kvm_mips_walk_pgd(pgdp, NULL, addr);
 796                 if (ptep) {
 797                         ptep[0] = pfn_pte(0, __pgprot(0));
 798                         ptep[1] = pfn_pte(0, __pgprot(0));
 799                 }
 800         }
 801 }
 802
 803 /*
 804  * kvm_mips_flush_gva_{pte,pmd,pud,pgd,pt}.
 805  * Flush a range of guest physical address space from the VM's GPA page tables.
 806  */
 807
 808 static bool kvm_mips_flush_gva_pte(pte_t *pte, unsigned long start_gva,
 809                                    unsigned long end_gva)
 810 {
 811         int i_min = pte_index(start_gva);
 812         int i_max = pte_index(end_gva);
 813         bool safe_to_remove = (i_min == 0 && i_max == PTRS_PER_PTE - 1);
 814         int i;
 815
 816         /*
 817          * There's no freeing to do, so there's no point clearing individual
 818          * entries unless only part of the last level page table needs flushing.
 819          */
 820         if (safe_to_remove)
 821                 return true;
 822
 823         for (i = i_min; i <= i_max; ++i) {
 824                 if (!pte_present(pte[i]))
 825                         continue;
 826
 827                 set_pte(pte + i, __pte(0));
 828         }
 829         return false;
 830 }
 831
 832 static bool kvm_mips_flush_gva_pmd(pmd_t *pmd, unsigned long start_gva,
 833                                    unsigned long end_gva)
 834 {
 835         pte_t *pte;
 836         unsigned long end = ~0ul;
 837         int i_min = pmd_index(start_gva);
 838         int i_max = pmd_index(end_gva);
 839         bool safe_to_remove = (i_min == 0 && i_max == PTRS_PER_PMD - 1);
 840         int i;
 841
 842         for (i = i_min; i <= i_max; ++i, start_gva = 0) {
 843                 if (!pmd_present(pmd[i]))
 844                         continue;
 845
 846                 pte = pte_offset_kernel(pmd + i, 0);
 847                 if (i == i_max)
 848                         end = end_gva;
 849
 850                 if (kvm_mips_flush_gva_pte(pte, start_gva, end)) {
 851                         pmd_clear(pmd + i);
 852                         pte_free_kernel(NULL, pte);
 853                 } else {
 854                         safe_to_remove = false;
 855                 }
 856         }
 857         return safe_to_remove;
 858 }
 859
 860 static bool kvm_mips_flush_gva_pud(pud_t *pud, unsigned long start_gva,
 861                                    unsigned long end_gva)
 862 {
 863         pmd_t *pmd;
 864         unsigned long end = ~0ul;
 865         int i_min = pud_index(start_gva);
 866         int i_max = pud_index(end_gva);
 867         bool safe_to_remove = (i_min == 0 && i_max == PTRS_PER_PUD - 1);
 868         int i;
 869
 870         for (i = i_min; i <= i_max; ++i, start_gva = 0) {
 871                 if (!pud_present(pud[i]))
 872                         continue;
 873
 874                 pmd = pmd_offset(pud + i, 0);
 875                 if (i == i_max)
 876                         end = end_gva;
 877
 878                 if (kvm_mips_flush_gva_pmd(pmd, start_gva, end)) {
 879                         pud_clear(pud + i);
 880                         pmd_free(NULL, pmd);
 881                 } else {
 882                         safe_to_remove = false;
 883                 }
 884         }
 885         return safe_to_remove;
 886 }
 887
 888 static bool kvm_mips_flush_gva_pgd(pgd_t *pgd, unsigned long start_gva,
 889                                    unsigned long end_gva)
 890 {
 891         p4d_t *p4d;
 892         pud_t *pud;
 893         unsigned long end = ~0ul;
 894         int i_min = pgd_index(start_gva);
 895         int i_max = pgd_index(end_gva);
 896         bool safe_to_remove = (i_min == 0 && i_max == PTRS_PER_PGD - 1);
 897         int i;
 898
 899         for (i = i_min; i <= i_max; ++i, start_gva = 0) {
 900                 if (!pgd_present(pgd[i]))
 901                         continue;
 902
 903                 p4d = p4d_offset(pgd, 0);
 904                 pud = pud_offset(p4d + i, 0);
 905                 if (i == i_max)
 906                         end = end_gva;
 907
 908                 if (kvm_mips_flush_gva_pud(pud, start_gva, end)) {
 909                         pgd_clear(pgd + i);
 910                         pud_free(NULL, pud);
 911                 } else {
 912                         safe_to_remove = false;
 913                 }
 914         }
 915         return safe_to_remove;
 916 }
 917
 918 void kvm_mips_flush_gva_pt(pgd_t *pgd, enum kvm_mips_flush flags)
 919 {
 920         if (flags & KMF_GPA) {
 921                 /* all of guest virtual address space could be affected */
 922                 if (flags & KMF_KERN)
 923                         /* useg, kseg0, seg2/3 */
 924                         kvm_mips_flush_gva_pgd(pgd, 0, 0x7fffffff);
 925                 else
 926                         /* useg */
 927                         kvm_mips_flush_gva_pgd(pgd, 0, 0x3fffffff);
 928         } else {
 929                 /* useg */
 930                 kvm_mips_flush_gva_pgd(pgd, 0, 0x3fffffff);
 931
 932                 /* kseg2/3 */
 933                 if (flags & KMF_KERN)
 934                         kvm_mips_flush_gva_pgd(pgd, 0x60000000, 0x7fffffff);
 935         }
 936 }
 937
 938 static pte_t kvm_mips_gpa_pte_to_gva_unmapped(pte_t pte)
 939 {
 940         /*
 941          * Don't leak writeable but clean entries from GPA page tables. We don't
 942          * want the normal Linux tlbmod handler to handle dirtying when KVM
 943          * accesses guest memory.
 944          */
 945         if (!pte_dirty(pte))
 946                 pte = pte_wrprotect(pte);
 947
 948         return pte;
 949 }
 950
 951 static pte_t kvm_mips_gpa_pte_to_gva_mapped(pte_t pte, long entrylo)
 952 {
 953         /* Guest EntryLo overrides host EntryLo */
 954         if (!(entrylo & ENTRYLO_D))
 955                 pte = pte_mkclean(pte);
 956
 957         return kvm_mips_gpa_pte_to_gva_unmapped(pte);
 958 }
 959
 960 #ifdef CONFIG_KVM_MIPS_VZ
 961 int kvm_mips_handle_vz_root_tlb_fault(unsigned long badvaddr,
 962                                       struct kvm_vcpu *vcpu,
 963                                       bool write_fault)
 964 {
 965         int ret;
 966
 967         ret = kvm_mips_map_page(vcpu, badvaddr, write_fault, NULL, NULL);
 968         if (ret)
 969                 return ret;
 970
 971         /* Invalidate this entry in the TLB */
 972         return kvm_vz_host_tlb_inv(vcpu, badvaddr);
 973 }
 974 #endif
 975
 976 /* XXXKYMA: Must be called with interrupts disabled */
 977 int kvm_mips_handle_kseg0_tlb_fault(unsigned long badvaddr,
 978                                     struct kvm_vcpu *vcpu,
 979                                     bool write_fault)
 980 {
 981         unsigned long gpa;
 982         pte_t pte_gpa[2], *ptep_gva;
 983         int idx;
 984
 985         if (KVM_GUEST_KSEGX(badvaddr) != KVM_GUEST_KSEG0) {
 986                 kvm_err("%s: Invalid BadVaddr: %#lx\n", __func__, badvaddr);
 987                 kvm_mips_dump_host_tlbs();
 988                 return -1;
 989         }
 990
 991         /* Get the GPA page table entry */
 992         gpa = KVM_GUEST_CPHYSADDR(badvaddr);
 993         idx = (badvaddr >> PAGE_SHIFT) & 1;
 994         if (kvm_mips_map_page(vcpu, gpa, write_fault, &pte_gpa[idx],
 995                               &pte_gpa[!idx]) < 0)
 996                 return -1;
 997
 998         /* Get the GVA page table entry */
 999         ptep_gva = kvm_trap_emul_pte_for_gva(vcpu, badvaddr & ~PAGE_SIZE);
1000         if (!ptep_gva) {
1001                 kvm_err("No ptep for gva %lx\n", badvaddr);
1002                 return -1;
1003         }
1004
1005         /* Copy a pair of entries from GPA page table to GVA page table */
1006         ptep_gva[0] = kvm_mips_gpa_pte_to_gva_unmapped(pte_gpa[0]);
1007         ptep_gva[1] = kvm_mips_gpa_pte_to_gva_unmapped(pte_gpa[1]);
1008
1009         /* Invalidate this entry in the TLB, guest kernel ASID only */
1010         kvm_mips_host_tlb_inv(vcpu, badvaddr, false, true);
1011         return 0;
1012 }
1013
1014 int kvm_mips_handle_mapped_seg_tlb_fault(struct kvm_vcpu *vcpu,
1015                                          struct kvm_mips_tlb *tlb,
1016                                          unsigned long gva,
1017                                          bool write_fault)
1018 {
1019         struct kvm *kvm = vcpu->kvm;
1020         long tlb_lo[2];
1021         pte_t pte_gpa[2], *ptep_buddy, *ptep_gva;
1022         unsigned int idx = TLB_LO_IDX(*tlb, gva);
1023         bool kernel = KVM_GUEST_KERNEL_MODE(vcpu);
1024
1025         tlb_lo[0] = tlb->tlb_lo[0];
1026         tlb_lo[1] = tlb->tlb_lo[1];
1027
1028         /*
1029          * The commpage address must not be mapped to anything else if the guest
1030          * TLB contains entries nearby, or commpage accesses will break.
1031          */
1032         if (!((gva ^ KVM_GUEST_COMMPAGE_ADDR) & VPN2_MASK & (PAGE_MASK << 1)))
1033                 tlb_lo[TLB_LO_IDX(*tlb, KVM_GUEST_COMMPAGE_ADDR)] = 0;
1034
1035         /* Get the GPA page table entry */
1036         if (kvm_mips_map_page(vcpu, mips3_tlbpfn_to_paddr(tlb_lo[idx]),
1037                               write_fault, &pte_gpa[idx], NULL) < 0)
1038                 return -1;
1039
1040         /* And its GVA buddy's GPA page table entry if it also exists */
1041         pte_gpa[!idx] = pfn_pte(0, __pgprot(0));
1042         if (tlb_lo[!idx] & ENTRYLO_V) {
1043                 spin_lock(&kvm->mmu_lock);
1044                 ptep_buddy = kvm_mips_pte_for_gpa(kvm, NULL,
1045                                         mips3_tlbpfn_to_paddr(tlb_lo[!idx]));
1046                 if (ptep_buddy)
1047                         pte_gpa[!idx] = *ptep_buddy;
1048                 spin_unlock(&kvm->mmu_lock);
1049         }
1050
1051         /* Get the GVA page table entry pair */
1052         ptep_gva = kvm_trap_emul_pte_for_gva(vcpu, gva & ~PAGE_SIZE);
1053         if (!ptep_gva) {
1054                 kvm_err("No ptep for gva %lx\n", gva);
1055                 return -1;
1056         }
1057
1058         /* Copy a pair of entries from GPA page table to GVA page table */
1059         ptep_gva[0] = kvm_mips_gpa_pte_to_gva_mapped(pte_gpa[0], tlb_lo[0]);
1060         ptep_gva[1] = kvm_mips_gpa_pte_to_gva_mapped(pte_gpa[1], tlb_lo[1]);
1061
1062         /* Invalidate this entry in the TLB, current guest mode ASID only */
1063         kvm_mips_host_tlb_inv(vcpu, gva, !kernel, kernel);
1064
1065         kvm_debug("@ %#lx tlb_lo0: 0x%08lx tlb_lo1: 0x%08lx\n", vcpu->arch.pc,
1066                   tlb->tlb_lo[0], tlb->tlb_lo[1]);
1067
1068         return 0;
1069 }
1070
1071 int kvm_mips_handle_commpage_tlb_fault(unsigned long badvaddr,
1072                                        struct kvm_vcpu *vcpu)
1073 {
1074         kvm_pfn_t pfn;
1075         pte_t *ptep;
1076
1077         ptep = kvm_trap_emul_pte_for_gva(vcpu, badvaddr);
1078         if (!ptep) {
1079                 kvm_err("No ptep for commpage %lx\n", badvaddr);
1080                 return -1;
1081         }
1082
1083         pfn = PFN_DOWN(virt_to_phys(vcpu->arch.kseg0_commpage));
1084         /* Also set valid and dirty, so refill handler doesn't have to */
1085         *ptep = pte_mkyoung(pte_mkdirty(pfn_pte(pfn, PAGE_SHARED)));
1086
1087         /* Invalidate this entry in the TLB, guest kernel ASID only */
1088         kvm_mips_host_tlb_inv(vcpu, badvaddr, false, true);
1089         return 0;
1090 }
1091
1092 /**
1093  * kvm_mips_migrate_count() - Migrate timer.
1094  * @vcpu:       Virtual CPU.
1095  *
1096  * Migrate CP0_Count hrtimer to the current CPU by cancelling and restarting it
1097  * if it was running prior to being cancelled.
1098  *
1099  * Must be called when the VCPU is migrated to a different CPU to ensure that
1100  * timer expiry during guest execution interrupts the guest and causes the
1101  * interrupt to be delivered in a timely manner.
1102  */
1103 static void kvm_mips_migrate_count(struct kvm_vcpu *vcpu)
1104 {
1105         if (hrtimer_cancel(&vcpu->arch.comparecount_timer))
1106                 hrtimer_restart(&vcpu->arch.comparecount_timer);
1107 }
1108
1109 /* Restore ASID once we are scheduled back after preemption */
1110 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
1111 {
1112         unsigned long flags;
1113
1114         kvm_debug("%s: vcpu %p, cpu: %d\n", __func__, vcpu, cpu);
1115
1116         local_irq_save(flags);
1117
1118         vcpu->cpu = cpu;
1119         if (vcpu->arch.last_sched_cpu != cpu) {
1120                 kvm_debug("[%d->%d]KVM VCPU[%d] switch\n",
1121                           vcpu->arch.last_sched_cpu, cpu, vcpu->vcpu_id);
1122                 /*
1123                  * Migrate the timer interrupt to the current CPU so that it
1124                  * always interrupts the guest and synchronously triggers a
1125                  * guest timer interrupt.
1126                  */
1127                 kvm_mips_migrate_count(vcpu);
1128         }
1129
1130         /* restore guest state to registers */
1131         kvm_mips_callbacks->vcpu_load(vcpu, cpu);
1132
1133         local_irq_restore(flags);
1134 }
1135
1136 /* ASID can change if another task is scheduled during preemption */
1137 void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
1138 {
1139         unsigned long flags;
1140         int cpu;
1141
1142         local_irq_save(flags);
1143
1144         cpu = smp_processor_id();
1145         vcpu->arch.last_sched_cpu = cpu;
1146         vcpu->cpu = -1;
1147
1148         /* save guest state in registers */
1149         kvm_mips_callbacks->vcpu_put(vcpu, cpu);
1150
1151         local_irq_restore(flags);
1152 }
1153
1154 /**
1155  * kvm_trap_emul_gva_fault() - Safely attempt to handle a GVA access fault.
1156  * @vcpu:       Virtual CPU.
1157  * @gva:        Guest virtual address to be accessed.
1158  * @write:      True if write attempted (must be dirtied and made writable).
1159  *
1160  * Safely attempt to handle a GVA fault, mapping GVA pages if necessary, and
1161  * dirtying the page if @write so that guest instructions can be modified.
1162  *
1163  * Returns:     KVM_MIPS_MAPPED on success.
1164  *              KVM_MIPS_GVA if bad guest virtual address.
1165  *              KVM_MIPS_GPA if bad guest physical address.
1166  *              KVM_MIPS_TLB if guest TLB not present.
1167  *              KVM_MIPS_TLBINV if guest TLB present but not valid.
1168  *              KVM_MIPS_TLBMOD if guest TLB read only.
1169  */
1170 enum kvm_mips_fault_result kvm_trap_emul_gva_fault(struct kvm_vcpu *vcpu,
1171                                                    unsigned long gva,
1172                                                    bool write)
1173 {
1174         struct mips_coproc *cop0 = vcpu->arch.cop0;
1175         struct kvm_mips_tlb *tlb;
1176         int index;
1177
1178         if (KVM_GUEST_KSEGX(gva) == KVM_GUEST_KSEG0) {
1179                 if (kvm_mips_handle_kseg0_tlb_fault(gva, vcpu, write) < 0)
1180                         return KVM_MIPS_GPA;
1181         } else if ((KVM_GUEST_KSEGX(gva) < KVM_GUEST_KSEG0) ||
1182                    KVM_GUEST_KSEGX(gva) == KVM_GUEST_KSEG23) {
1183                 /* Address should be in the guest TLB */
1184                 index = kvm_mips_guest_tlb_lookup(vcpu, (gva & VPN2_MASK) |
1185                           (kvm_read_c0_guest_entryhi(cop0) & KVM_ENTRYHI_ASID));
1186                 if (index < 0)
1187                         return KVM_MIPS_TLB;
1188                 tlb = &vcpu->arch.guest_tlb[index];
1189
1190                 /* Entry should be valid, and dirty for writes */
1191                 if (!TLB_IS_VALID(*tlb, gva))
1192                         return KVM_MIPS_TLBINV;
1193                 if (write && !TLB_IS_DIRTY(*tlb, gva))
1194                         return KVM_MIPS_TLBMOD;
1195
1196                 if (kvm_mips_handle_mapped_seg_tlb_fault(vcpu, tlb, gva, write))
1197                         return KVM_MIPS_GPA;
1198         } else {
1199                 return KVM_MIPS_GVA;
1200         }
1201
1202         return KVM_MIPS_MAPPED;
1203 }
1204
1205 int kvm_get_inst(u32 *opc, struct kvm_vcpu *vcpu, u32 *out)
1206 {
1207         int err;
1208
1209         if (WARN(IS_ENABLED(CONFIG_KVM_MIPS_VZ),
1210                  "Expect BadInstr/BadInstrP registers to be used with VZ\n"))
1211                 return -EINVAL;
1212
1213 retry:
1214         kvm_trap_emul_gva_lockless_begin(vcpu);
1215         err = get_user(*out, opc);
1216         kvm_trap_emul_gva_lockless_end(vcpu);
1217
1218         if (unlikely(err)) {
1219                 /*
1220                  * Try to handle the fault, maybe we just raced with a GVA
1221                  * invalidation.
1222                  */
1223                 err = kvm_trap_emul_gva_fault(vcpu, (unsigned long)opc,
1224                                               false);
1225                 if (unlikely(err)) {
1226                         kvm_err("%s: illegal address: %p\n",
1227                                 __func__, opc);
1228                         return -EFAULT;
1229                 }
1230
1231                 /* Hopefully it'll work now */
1232                 goto retry;
1233         }
1234         return 0;
1235 }