arch/riscv/kvm/mmu.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2019 Western Digital Corporation or its affiliates.
   4  *
   5  * Authors:
   6  *     Anup Patel <anup.patel@wdc.com>
   7  */
   8
   9 #include <linux/bitops.h>
  10 #include <linux/errno.h>
  11 #include <linux/err.h>
  12 #include <linux/hugetlb.h>
  13 #include <linux/module.h>
  14 #include <linux/uaccess.h>
  15 #include <linux/vmalloc.h>
  16 #include <linux/kvm_host.h>
  17 #include <linux/sched/signal.h>
  18 #include <asm/csr.h>
  19 #include <asm/page.h>
  20 #include <asm/pgtable.h>
  21
  22 #ifdef CONFIG_64BIT
  23 static unsigned long gstage_mode = (HGATP_MODE_SV39X4 << HGATP_MODE_SHIFT);
  24 static unsigned long gstage_pgd_levels = 3;
  25 #define gstage_index_bits       9
  26 #else
  27 static unsigned long gstage_mode = (HGATP_MODE_SV32X4 << HGATP_MODE_SHIFT);
  28 static unsigned long gstage_pgd_levels = 2;
  29 #define gstage_index_bits       10
  30 #endif
  31
  32 #define gstage_pgd_xbits        2
  33 #define gstage_pgd_size (1UL << (HGATP_PAGE_SHIFT + gstage_pgd_xbits))
  34 #define gstage_gpa_bits (HGATP_PAGE_SHIFT + \
  35                          (gstage_pgd_levels * gstage_index_bits) + \
  36                          gstage_pgd_xbits)
  37 #define gstage_gpa_size ((gpa_t)(1ULL << gstage_gpa_bits))
  38
  39 #define gstage_pte_leaf(__ptep) \
  40         (pte_val(*(__ptep)) & (_PAGE_READ | _PAGE_WRITE | _PAGE_EXEC))
  41
  42 static inline unsigned long gstage_pte_index(gpa_t addr, u32 level)
  43 {
  44         unsigned long mask;
  45         unsigned long shift = HGATP_PAGE_SHIFT + (gstage_index_bits * level);
  46
  47         if (level == (gstage_pgd_levels - 1))
  48                 mask = (PTRS_PER_PTE * (1UL << gstage_pgd_xbits)) - 1;
  49         else
  50                 mask = PTRS_PER_PTE - 1;
  51
  52         return (addr >> shift) & mask;
  53 }
  54
  55 static inline unsigned long gstage_pte_page_vaddr(pte_t pte)
  56 {
  57         return (unsigned long)pfn_to_virt(pte_val(pte) >> _PAGE_PFN_SHIFT);
  58 }
  59
  60 static int gstage_page_size_to_level(unsigned long page_size, u32 *out_level)
  61 {
  62         u32 i;
  63         unsigned long psz = 1UL << 12;
  64
  65         for (i = 0; i < gstage_pgd_levels; i++) {
  66                 if (page_size == (psz << (i * gstage_index_bits))) {
  67                         *out_level = i;
  68                         return 0;
  69                 }
  70         }
  71
  72         return -EINVAL;
  73 }
  74
  75 static int gstage_level_to_page_order(u32 level, unsigned long *out_pgorder)
  76 {
  77         if (gstage_pgd_levels < level)
  78                 return -EINVAL;
  79
  80         *out_pgorder = 12 + (level * gstage_index_bits);
  81         return 0;
  82 }
  83
  84 static int gstage_level_to_page_size(u32 level, unsigned long *out_pgsize)
  85 {
  86         int rc;
  87         unsigned long page_order = PAGE_SHIFT;
  88
  89         rc = gstage_level_to_page_order(level, &page_order);
  90         if (rc)
  91                 return rc;
  92
  93         *out_pgsize = BIT(page_order);
  94         return 0;
  95 }
  96
  97 static bool gstage_get_leaf_entry(struct kvm *kvm, gpa_t addr,
  98                                   pte_t **ptepp, u32 *ptep_level)
  99 {
 100         pte_t *ptep;
 101         u32 current_level = gstage_pgd_levels - 1;
 102
 103         *ptep_level = current_level;
 104         ptep = (pte_t *)kvm->arch.pgd;
 105         ptep = &ptep[gstage_pte_index(addr, current_level)];
 106         while (ptep && pte_val(*ptep)) {
 107                 if (gstage_pte_leaf(ptep)) {
 108                         *ptep_level = current_level;
 109                         *ptepp = ptep;
 110                         return true;
 111                 }
 112
 113                 if (current_level) {
 114                         current_level--;
 115                         *ptep_level = current_level;
 116                         ptep = (pte_t *)gstage_pte_page_vaddr(*ptep);
 117                         ptep = &ptep[gstage_pte_index(addr, current_level)];
 118                 } else {
 119                         ptep = NULL;
 120                 }
 121         }
 122
 123         return false;
 124 }
 125
 126 static void gstage_remote_tlb_flush(struct kvm *kvm, u32 level, gpa_t addr)
 127 {
 128         unsigned long order = PAGE_SHIFT;
 129
 130         if (gstage_level_to_page_order(level, &order))
 131                 return;
 132         addr &= ~(BIT(order) - 1);
 133
 134         kvm_riscv_hfence_gvma_vmid_gpa(kvm, -1UL, 0, addr, BIT(order), order);
 135 }
 136
 137 static int gstage_set_pte(struct kvm *kvm, u32 level,
 138                            struct kvm_mmu_memory_cache *pcache,
 139                            gpa_t addr, const pte_t *new_pte)
 140 {
 141         u32 current_level = gstage_pgd_levels - 1;
 142         pte_t *next_ptep = (pte_t *)kvm->arch.pgd;
 143         pte_t *ptep = &next_ptep[gstage_pte_index(addr, current_level)];
 144
 145         if (current_level < level)
 146                 return -EINVAL;
 147
 148         while (current_level != level) {
 149                 if (gstage_pte_leaf(ptep))
 150                         return -EEXIST;
 151
 152                 if (!pte_val(*ptep)) {
 153                         if (!pcache)
 154                                 return -ENOMEM;
 155                         next_ptep = kvm_mmu_memory_cache_alloc(pcache);
 156                         if (!next_ptep)
 157                                 return -ENOMEM;
 158                         *ptep = pfn_pte(PFN_DOWN(__pa(next_ptep)),
 159                                         __pgprot(_PAGE_TABLE));
 160                 } else {
 161                         if (gstage_pte_leaf(ptep))
 162                                 return -EEXIST;
 163                         next_ptep = (pte_t *)gstage_pte_page_vaddr(*ptep);
 164                 }
 165
 166                 current_level--;
 167                 ptep = &next_ptep[gstage_pte_index(addr, current_level)];
 168         }
 169
 170         *ptep = *new_pte;
 171         if (gstage_pte_leaf(ptep))
 172                 gstage_remote_tlb_flush(kvm, current_level, addr);
 173
 174         return 0;
 175 }
 176
 177 static int gstage_map_page(struct kvm *kvm,
 178                            struct kvm_mmu_memory_cache *pcache,
 179                            gpa_t gpa, phys_addr_t hpa,
 180                            unsigned long page_size,
 181                            bool page_rdonly, bool page_exec)
 182 {
 183         int ret;
 184         u32 level = 0;
 185         pte_t new_pte;
 186         pgprot_t prot;
 187
 188         ret = gstage_page_size_to_level(page_size, &level);
 189         if (ret)
 190                 return ret;
 191
 192         /*
 193          * A RISC-V implementation can choose to either:
 194          * 1) Update 'A' and 'D' PTE bits in hardware
 195          * 2) Generate page fault when 'A' and/or 'D' bits are not set
 196          *    PTE so that software can update these bits.
 197          *
 198          * We support both options mentioned above. To achieve this, we
 199          * always set 'A' and 'D' PTE bits at time of creating G-stage
 200          * mapping. To support KVM dirty page logging with both options
 201          * mentioned above, we will write-protect G-stage PTEs to track
 202          * dirty pages.
 203          */
 204
 205         if (page_exec) {
 206                 if (page_rdonly)
 207                         prot = PAGE_READ_EXEC;
 208                 else
 209                         prot = PAGE_WRITE_EXEC;
 210         } else {
 211                 if (page_rdonly)
 212                         prot = PAGE_READ;
 213                 else
 214                         prot = PAGE_WRITE;
 215         }
 216         new_pte = pfn_pte(PFN_DOWN(hpa), prot);
 217         new_pte = pte_mkdirty(new_pte);
 218
 219         return gstage_set_pte(kvm, level, pcache, gpa, &new_pte);
 220 }
 221
 222 enum gstage_op {
 223         GSTAGE_OP_NOP = 0,      /* Nothing */
 224         GSTAGE_OP_CLEAR,        /* Clear/Unmap */
 225         GSTAGE_OP_WP,           /* Write-protect */
 226 };
 227
 228 static void gstage_op_pte(struct kvm *kvm, gpa_t addr,
 229                           pte_t *ptep, u32 ptep_level, enum gstage_op op)
 230 {
 231         int i, ret;
 232         pte_t *next_ptep;
 233         u32 next_ptep_level;
 234         unsigned long next_page_size, page_size;
 235
 236         ret = gstage_level_to_page_size(ptep_level, &page_size);
 237         if (ret)
 238                 return;
 239
 240         BUG_ON(addr & (page_size - 1));
 241
 242         if (!pte_val(*ptep))
 243                 return;
 244
 245         if (ptep_level && !gstage_pte_leaf(ptep)) {
 246                 next_ptep = (pte_t *)gstage_pte_page_vaddr(*ptep);
 247                 next_ptep_level = ptep_level - 1;
 248                 ret = gstage_level_to_page_size(next_ptep_level,
 249                                                 &next_page_size);
 250                 if (ret)
 251                         return;
 252
 253                 if (op == GSTAGE_OP_CLEAR)
 254                         set_pte(ptep, __pte(0));
 255                 for (i = 0; i < PTRS_PER_PTE; i++)
 256                         gstage_op_pte(kvm, addr + i * next_page_size,
 257                                         &next_ptep[i], next_ptep_level, op);
 258                 if (op == GSTAGE_OP_CLEAR)
 259                         put_page(virt_to_page(next_ptep));
 260         } else {
 261                 if (op == GSTAGE_OP_CLEAR)
 262                         set_pte(ptep, __pte(0));
 263                 else if (op == GSTAGE_OP_WP)
 264                         set_pte(ptep, __pte(pte_val(*ptep) & ~_PAGE_WRITE));
 265                 gstage_remote_tlb_flush(kvm, ptep_level, addr);
 266         }
 267 }
 268
 269 static void gstage_unmap_range(struct kvm *kvm, gpa_t start,
 270                                gpa_t size, bool may_block)
 271 {
 272         int ret;
 273         pte_t *ptep;
 274         u32 ptep_level;
 275         bool found_leaf;
 276         unsigned long page_size;
 277         gpa_t addr = start, end = start + size;
 278
 279         while (addr < end) {
 280                 found_leaf = gstage_get_leaf_entry(kvm, addr,
 281                                                    &ptep, &ptep_level);
 282                 ret = gstage_level_to_page_size(ptep_level, &page_size);
 283                 if (ret)
 284                         break;
 285
 286                 if (!found_leaf)
 287                         goto next;
 288
 289                 if (!(addr & (page_size - 1)) && ((end - addr) >= page_size))
 290                         gstage_op_pte(kvm, addr, ptep,
 291                                       ptep_level, GSTAGE_OP_CLEAR);
 292
 293 next:
 294                 addr += page_size;
 295
 296                 /*
 297                  * If the range is too large, release the kvm->mmu_lock
 298                  * to prevent starvation and lockup detector warnings.
 299                  */
 300                 if (may_block && addr < end)
 301                         cond_resched_lock(&kvm->mmu_lock);
 302         }
 303 }
 304
 305 static void gstage_wp_range(struct kvm *kvm, gpa_t start, gpa_t end)
 306 {
 307         int ret;
 308         pte_t *ptep;
 309         u32 ptep_level;
 310         bool found_leaf;
 311         gpa_t addr = start;
 312         unsigned long page_size;
 313
 314         while (addr < end) {
 315                 found_leaf = gstage_get_leaf_entry(kvm, addr,
 316                                                    &ptep, &ptep_level);
 317                 ret = gstage_level_to_page_size(ptep_level, &page_size);
 318                 if (ret)
 319                         break;
 320
 321                 if (!found_leaf)
 322                         goto next;
 323
 324                 if (!(addr & (page_size - 1)) && ((end - addr) >= page_size))
 325                         gstage_op_pte(kvm, addr, ptep,
 326                                       ptep_level, GSTAGE_OP_WP);
 327
 328 next:
 329                 addr += page_size;
 330         }
 331 }
 332
 333 static void gstage_wp_memory_region(struct kvm *kvm, int slot)
 334 {
 335         struct kvm_memslots *slots = kvm_memslots(kvm);
 336         struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
 337         phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
 338         phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
 339
 340         spin_lock(&kvm->mmu_lock);
 341         gstage_wp_range(kvm, start, end);
 342         spin_unlock(&kvm->mmu_lock);
 343         kvm_flush_remote_tlbs(kvm);
 344 }
 345
 346 static int gstage_ioremap(struct kvm *kvm, gpa_t gpa, phys_addr_t hpa,
 347                           unsigned long size, bool writable)
 348 {
 349         pte_t pte;
 350         int ret = 0;
 351         unsigned long pfn;
 352         phys_addr_t addr, end;
 353         struct kvm_mmu_memory_cache pcache;
 354
 355         memset(&pcache, 0, sizeof(pcache));
 356         pcache.gfp_zero = __GFP_ZERO;
 357
 358         end = (gpa + size + PAGE_SIZE - 1) & PAGE_MASK;
 359         pfn = __phys_to_pfn(hpa);
 360
 361         for (addr = gpa; addr < end; addr += PAGE_SIZE) {
 362                 pte = pfn_pte(pfn, PAGE_KERNEL);
 363
 364                 if (!writable)
 365                         pte = pte_wrprotect(pte);
 366
 367                 ret = kvm_mmu_topup_memory_cache(&pcache, gstage_pgd_levels);
 368                 if (ret)
 369                         goto out;
 370
 371                 spin_lock(&kvm->mmu_lock);
 372                 ret = gstage_set_pte(kvm, 0, &pcache, addr, &pte);
 373                 spin_unlock(&kvm->mmu_lock);
 374                 if (ret)
 375                         goto out;
 376
 377                 pfn++;
 378         }
 379
 380 out:
 381         kvm_mmu_free_memory_cache(&pcache);
 382         return ret;
 383 }
 384
 385 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
 386                                              struct kvm_memory_slot *slot,
 387                                              gfn_t gfn_offset,
 388                                              unsigned long mask)
 389 {
 390         phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
 391         phys_addr_t start = (base_gfn +  __ffs(mask)) << PAGE_SHIFT;
 392         phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
 393
 394         gstage_wp_range(kvm, start, end);
 395 }
 396
 397 void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
 398 {
 399 }
 400
 401 void kvm_arch_flush_remote_tlbs_memslot(struct kvm *kvm,
 402                                         const struct kvm_memory_slot *memslot)
 403 {
 404         kvm_flush_remote_tlbs(kvm);
 405 }
 406
 407 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free)
 408 {
 409 }
 410
 411 void kvm_arch_memslots_updated(struct kvm *kvm, u64 gen)
 412 {
 413 }
 414
 415 void kvm_arch_flush_shadow_all(struct kvm *kvm)
 416 {
 417         kvm_riscv_gstage_free_pgd(kvm);
 418 }
 419
 420 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
 421                                    struct kvm_memory_slot *slot)
 422 {
 423         gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
 424         phys_addr_t size = slot->npages << PAGE_SHIFT;
 425
 426         spin_lock(&kvm->mmu_lock);
 427         gstage_unmap_range(kvm, gpa, size, false);
 428         spin_unlock(&kvm->mmu_lock);
 429 }
 430
 431 void kvm_arch_commit_memory_region(struct kvm *kvm,
 432                                 struct kvm_memory_slot *old,
 433                                 const struct kvm_memory_slot *new,
 434                                 enum kvm_mr_change change)
 435 {
 436         /*
 437          * At this point memslot has been committed and there is an
 438          * allocated dirty_bitmap[], dirty pages will be tracked while
 439          * the memory slot is write protected.
 440          */
 441         if (change != KVM_MR_DELETE && new->flags & KVM_MEM_LOG_DIRTY_PAGES)
 442                 gstage_wp_memory_region(kvm, new->id);
 443 }
 444
 445 int kvm_arch_prepare_memory_region(struct kvm *kvm,
 446                                 const struct kvm_memory_slot *old,
 447                                 struct kvm_memory_slot *new,
 448                                 enum kvm_mr_change change)
 449 {
 450         hva_t hva, reg_end, size;
 451         gpa_t base_gpa;
 452         bool writable;
 453         int ret = 0;
 454
 455         if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
 456                         change != KVM_MR_FLAGS_ONLY)
 457                 return 0;
 458
 459         /*
 460          * Prevent userspace from creating a memory region outside of the GPA
 461          * space addressable by the KVM guest GPA space.
 462          */
 463         if ((new->base_gfn + new->npages) >=
 464             (gstage_gpa_size >> PAGE_SHIFT))
 465                 return -EFAULT;
 466
 467         hva = new->userspace_addr;
 468         size = new->npages << PAGE_SHIFT;
 469         reg_end = hva + size;
 470         base_gpa = new->base_gfn << PAGE_SHIFT;
 471         writable = !(new->flags & KVM_MEM_READONLY);
 472
 473         mmap_read_lock(current->mm);
 474
 475         /*
 476          * A memory region could potentially cover multiple VMAs, and
 477          * any holes between them, so iterate over all of them to find
 478          * out if we can map any of them right now.
 479          *
 480          *     +--------------------------------------------+
 481          * +---------------+----------------+   +----------------+
 482          * |   : VMA 1     |      VMA 2     |   |    VMA 3  :    |
 483          * +---------------+----------------+   +----------------+
 484          *     |               memory region                |
 485          *     +--------------------------------------------+
 486          */
 487         do {
 488                 struct vm_area_struct *vma = find_vma(current->mm, hva);
 489                 hva_t vm_start, vm_end;
 490
 491                 if (!vma || vma->vm_start >= reg_end)
 492                         break;
 493
 494                 /*
 495                  * Mapping a read-only VMA is only allowed if the
 496                  * memory region is configured as read-only.
 497                  */
 498                 if (writable && !(vma->vm_flags & VM_WRITE)) {
 499                         ret = -EPERM;
 500                         break;
 501                 }
 502
 503                 /* Take the intersection of this VMA with the memory region */
 504                 vm_start = max(hva, vma->vm_start);
 505                 vm_end = min(reg_end, vma->vm_end);
 506
 507                 if (vma->vm_flags & VM_PFNMAP) {
 508                         gpa_t gpa = base_gpa + (vm_start - hva);
 509                         phys_addr_t pa;
 510
 511                         pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
 512                         pa += vm_start - vma->vm_start;
 513
 514                         /* IO region dirty page logging not allowed */
 515                         if (new->flags & KVM_MEM_LOG_DIRTY_PAGES) {
 516                                 ret = -EINVAL;
 517                                 goto out;
 518                         }
 519
 520                         ret = gstage_ioremap(kvm, gpa, pa,
 521                                              vm_end - vm_start, writable);
 522                         if (ret)
 523                                 break;
 524                 }
 525                 hva = vm_end;
 526         } while (hva < reg_end);
 527
 528         if (change == KVM_MR_FLAGS_ONLY)
 529                 goto out;
 530
 531         spin_lock(&kvm->mmu_lock);
 532         if (ret)
 533                 gstage_unmap_range(kvm, base_gpa, size, false);
 534         spin_unlock(&kvm->mmu_lock);
 535
 536 out:
 537         mmap_read_unlock(current->mm);
 538         return ret;
 539 }
 540
 541 bool kvm_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range)
 542 {
 543         if (!kvm->arch.pgd)
 544                 return false;
 545
 546         gstage_unmap_range(kvm, range->start << PAGE_SHIFT,
 547                            (range->end - range->start) << PAGE_SHIFT,
 548                            range->may_block);
 549         return false;
 550 }
 551
 552 bool kvm_set_spte_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
 553 {
 554         int ret;
 555         kvm_pfn_t pfn = pte_pfn(range->pte);
 556
 557         if (!kvm->arch.pgd)
 558                 return false;
 559
 560         WARN_ON(range->end - range->start != 1);
 561
 562         ret = gstage_map_page(kvm, NULL, range->start << PAGE_SHIFT,
 563                               __pfn_to_phys(pfn), PAGE_SIZE, true, true);
 564         if (ret) {
 565                 kvm_debug("Failed to map G-stage page (error %d)\n", ret);
 566                 return true;
 567         }
 568
 569         return false;
 570 }
 571
 572 bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
 573 {
 574         pte_t *ptep;
 575         u32 ptep_level = 0;
 576         u64 size = (range->end - range->start) << PAGE_SHIFT;
 577
 578         if (!kvm->arch.pgd)
 579                 return false;
 580
 581         WARN_ON(size != PAGE_SIZE && size != PMD_SIZE && size != PGDIR_SIZE);
 582
 583         if (!gstage_get_leaf_entry(kvm, range->start << PAGE_SHIFT,
 584                                    &ptep, &ptep_level))
 585                 return false;
 586
 587         return ptep_test_and_clear_young(NULL, 0, ptep);
 588 }
 589
 590 bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
 591 {
 592         pte_t *ptep;
 593         u32 ptep_level = 0;
 594         u64 size = (range->end - range->start) << PAGE_SHIFT;
 595
 596         if (!kvm->arch.pgd)
 597                 return false;
 598
 599         WARN_ON(size != PAGE_SIZE && size != PMD_SIZE && size != PGDIR_SIZE);
 600
 601         if (!gstage_get_leaf_entry(kvm, range->start << PAGE_SHIFT,
 602                                    &ptep, &ptep_level))
 603                 return false;
 604
 605         return pte_young(*ptep);
 606 }
 607
 608 int kvm_riscv_gstage_map(struct kvm_vcpu *vcpu,
 609                          struct kvm_memory_slot *memslot,
 610                          gpa_t gpa, unsigned long hva, bool is_write)
 611 {
 612         int ret;
 613         kvm_pfn_t hfn;
 614         bool writeable;
 615         short vma_pageshift;
 616         gfn_t gfn = gpa >> PAGE_SHIFT;
 617         struct vm_area_struct *vma;
 618         struct kvm *kvm = vcpu->kvm;
 619         struct kvm_mmu_memory_cache *pcache = &vcpu->arch.mmu_page_cache;
 620         bool logging = (memslot->dirty_bitmap &&
 621                         !(memslot->flags & KVM_MEM_READONLY)) ? true : false;
 622         unsigned long vma_pagesize, mmu_seq;
 623
 624         mmap_read_lock(current->mm);
 625
 626         vma = find_vma_intersection(current->mm, hva, hva + 1);
 627         if (unlikely(!vma)) {
 628                 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
 629                 mmap_read_unlock(current->mm);
 630                 return -EFAULT;
 631         }
 632
 633         if (is_vm_hugetlb_page(vma))
 634                 vma_pageshift = huge_page_shift(hstate_vma(vma));
 635         else
 636                 vma_pageshift = PAGE_SHIFT;
 637         vma_pagesize = 1ULL << vma_pageshift;
 638         if (logging || (vma->vm_flags & VM_PFNMAP))
 639                 vma_pagesize = PAGE_SIZE;
 640
 641         if (vma_pagesize == PMD_SIZE || vma_pagesize == PGDIR_SIZE)
 642                 gfn = (gpa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT;
 643
 644         mmap_read_unlock(current->mm);
 645
 646         if (vma_pagesize != PGDIR_SIZE &&
 647             vma_pagesize != PMD_SIZE &&
 648             vma_pagesize != PAGE_SIZE) {
 649                 kvm_err("Invalid VMA page size 0x%lx\n", vma_pagesize);
 650                 return -EFAULT;
 651         }
 652
 653         /* We need minimum second+third level pages */
 654         ret = kvm_mmu_topup_memory_cache(pcache, gstage_pgd_levels);
 655         if (ret) {
 656                 kvm_err("Failed to topup G-stage cache\n");
 657                 return ret;
 658         }
 659
 660         mmu_seq = kvm->mmu_notifier_seq;
 661
 662         hfn = gfn_to_pfn_prot(kvm, gfn, is_write, &writeable);
 663         if (hfn == KVM_PFN_ERR_HWPOISON) {
 664                 send_sig_mceerr(BUS_MCEERR_AR, (void __user *)hva,
 665                                 vma_pageshift, current);
 666                 return 0;
 667         }
 668         if (is_error_noslot_pfn(hfn))
 669                 return -EFAULT;
 670
 671         /*
 672          * If logging is active then we allow writable pages only
 673          * for write faults.
 674          */
 675         if (logging && !is_write)
 676                 writeable = false;
 677
 678         spin_lock(&kvm->mmu_lock);
 679
 680         if (mmu_notifier_retry(kvm, mmu_seq))
 681                 goto out_unlock;
 682
 683         if (writeable) {
 684                 kvm_set_pfn_dirty(hfn);
 685                 mark_page_dirty(kvm, gfn);
 686                 ret = gstage_map_page(kvm, pcache, gpa, hfn << PAGE_SHIFT,
 687                                       vma_pagesize, false, true);
 688         } else {
 689                 ret = gstage_map_page(kvm, pcache, gpa, hfn << PAGE_SHIFT,
 690                                       vma_pagesize, true, true);
 691         }
 692
 693         if (ret)
 694                 kvm_err("Failed to map in G-stage\n");
 695
 696 out_unlock:
 697         spin_unlock(&kvm->mmu_lock);
 698         kvm_set_pfn_accessed(hfn);
 699         kvm_release_pfn_clean(hfn);
 700         return ret;
 701 }
 702
 703 int kvm_riscv_gstage_alloc_pgd(struct kvm *kvm)
 704 {
 705         struct page *pgd_page;
 706
 707         if (kvm->arch.pgd != NULL) {
 708                 kvm_err("kvm_arch already initialized?\n");
 709                 return -EINVAL;
 710         }
 711
 712         pgd_page = alloc_pages(GFP_KERNEL | __GFP_ZERO,
 713                                 get_order(gstage_pgd_size));
 714         if (!pgd_page)
 715                 return -ENOMEM;
 716         kvm->arch.pgd = page_to_virt(pgd_page);
 717         kvm->arch.pgd_phys = page_to_phys(pgd_page);
 718
 719         return 0;
 720 }
 721
 722 void kvm_riscv_gstage_free_pgd(struct kvm *kvm)
 723 {
 724         void *pgd = NULL;
 725
 726         spin_lock(&kvm->mmu_lock);
 727         if (kvm->arch.pgd) {
 728                 gstage_unmap_range(kvm, 0UL, gstage_gpa_size, false);
 729                 pgd = READ_ONCE(kvm->arch.pgd);
 730                 kvm->arch.pgd = NULL;
 731                 kvm->arch.pgd_phys = 0;
 732         }
 733         spin_unlock(&kvm->mmu_lock);
 734
 735         if (pgd)
 736                 free_pages((unsigned long)pgd, get_order(gstage_pgd_size));
 737 }
 738
 739 void kvm_riscv_gstage_update_hgatp(struct kvm_vcpu *vcpu)
 740 {
 741         unsigned long hgatp = gstage_mode;
 742         struct kvm_arch *k = &vcpu->kvm->arch;
 743
 744         hgatp |= (READ_ONCE(k->vmid.vmid) << HGATP_VMID_SHIFT) &
 745                  HGATP_VMID_MASK;
 746         hgatp |= (k->pgd_phys >> PAGE_SHIFT) & HGATP_PPN;
 747
 748         csr_write(CSR_HGATP, hgatp);
 749
 750         if (!kvm_riscv_gstage_vmid_bits())
 751                 kvm_riscv_local_hfence_gvma_all();
 752 }
 753
 754 void kvm_riscv_gstage_mode_detect(void)
 755 {
 756 #ifdef CONFIG_64BIT
 757         /* Try Sv57x4 G-stage mode */
 758         csr_write(CSR_HGATP, HGATP_MODE_SV57X4 << HGATP_MODE_SHIFT);
 759         if ((csr_read(CSR_HGATP) >> HGATP_MODE_SHIFT) == HGATP_MODE_SV57X4) {
 760                 gstage_mode = (HGATP_MODE_SV57X4 << HGATP_MODE_SHIFT);
 761                 gstage_pgd_levels = 5;
 762                 goto skip_sv48x4_test;
 763         }
 764
 765         /* Try Sv48x4 G-stage mode */
 766         csr_write(CSR_HGATP, HGATP_MODE_SV48X4 << HGATP_MODE_SHIFT);
 767         if ((csr_read(CSR_HGATP) >> HGATP_MODE_SHIFT) == HGATP_MODE_SV48X4) {
 768                 gstage_mode = (HGATP_MODE_SV48X4 << HGATP_MODE_SHIFT);
 769                 gstage_pgd_levels = 4;
 770         }
 771 skip_sv48x4_test:
 772
 773         csr_write(CSR_HGATP, 0);
 774         kvm_riscv_local_hfence_gvma_all();
 775 #endif
 776 }
 777
 778 unsigned long kvm_riscv_gstage_mode(void)
 779 {
 780         return gstage_mode >> HGATP_MODE_SHIFT;
 781 }
 782
 783 int kvm_riscv_gstage_gpa_bits(void)
 784 {
 785         return gstage_gpa_bits;
 786 }