arch/arm/kvm/mmu.c

   1 /*
   2  * Copyright (C) 2012 - Virtual Open Systems and Columbia University
   3  * Author: Christoffer Dall <c.dall@virtualopensystems.com>
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License, version 2, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write to the Free Software
  16  * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
  17  */
  18
  19 #include <linux/mman.h>
  20 #include <linux/kvm_host.h>
  21 #include <linux/io.h>
  22 #include <linux/hugetlb.h>
  23 #include <trace/events/kvm.h>
  24 #include <asm/pgalloc.h>
  25 #include <asm/cacheflush.h>
  26 #include <asm/kvm_arm.h>
  27 #include <asm/kvm_mmu.h>
  28 #include <asm/kvm_mmio.h>
  29 #include <asm/kvm_asm.h>
  30 #include <asm/kvm_emulate.h>
  31
  32 #include "trace.h"
  33
  34 extern char  __hyp_idmap_text_start[], __hyp_idmap_text_end[];
  35
  36 static pgd_t *boot_hyp_pgd;
  37 static pgd_t *hyp_pgd;
  38 static DEFINE_MUTEX(kvm_hyp_pgd_mutex);
  39
  40 static void *init_bounce_page;
  41 static unsigned long hyp_idmap_start;
  42 static unsigned long hyp_idmap_end;
  43 static phys_addr_t hyp_idmap_vector;
  44
  45 #define hyp_pgd_order get_order(PTRS_PER_PGD * sizeof(pgd_t))
  46
  47 #define kvm_pmd_huge(_x)        (pmd_huge(_x) || pmd_trans_huge(_x))
  48
  49 static void kvm_tlb_flush_vmid_ipa(struct kvm *kvm, phys_addr_t ipa)
  50 {
  51         /*
  52          * This function also gets called when dealing with HYP page
  53          * tables. As HYP doesn't have an associated struct kvm (and
  54          * the HYP page tables are fairly static), we don't do
  55          * anything there.
  56          */
  57         if (kvm)
  58                 kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, kvm, ipa);
  59 }
  60
  61 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
  62                                   int min, int max)
  63 {
  64         void *page;
  65
  66         BUG_ON(max > KVM_NR_MEM_OBJS);
  67         if (cache->nobjs >= min)
  68                 return 0;
  69         while (cache->nobjs < max) {
  70                 page = (void *)__get_free_page(PGALLOC_GFP);
  71                 if (!page)
  72                         return -ENOMEM;
  73                 cache->objects[cache->nobjs++] = page;
  74         }
  75         return 0;
  76 }
  77
  78 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
  79 {
  80         while (mc->nobjs)
  81                 free_page((unsigned long)mc->objects[--mc->nobjs]);
  82 }
  83
  84 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
  85 {
  86         void *p;
  87
  88         BUG_ON(!mc || !mc->nobjs);
  89         p = mc->objects[--mc->nobjs];
  90         return p;
  91 }
  92
  93 static void clear_pgd_entry(struct kvm *kvm, pgd_t *pgd, phys_addr_t addr)
  94 {
  95         pud_t *pud_table __maybe_unused = pud_offset(pgd, 0);
  96         pgd_clear(pgd);
  97         kvm_tlb_flush_vmid_ipa(kvm, addr);
  98         pud_free(NULL, pud_table);
  99         put_page(virt_to_page(pgd));
 100 }
 101
 102 static void clear_pud_entry(struct kvm *kvm, pud_t *pud, phys_addr_t addr)
 103 {
 104         pmd_t *pmd_table = pmd_offset(pud, 0);
 105         VM_BUG_ON(pud_huge(*pud));
 106         pud_clear(pud);
 107         kvm_tlb_flush_vmid_ipa(kvm, addr);
 108         pmd_free(NULL, pmd_table);
 109         put_page(virt_to_page(pud));
 110 }
 111
 112 static void clear_pmd_entry(struct kvm *kvm, pmd_t *pmd, phys_addr_t addr)
 113 {
 114         pte_t *pte_table = pte_offset_kernel(pmd, 0);
 115         VM_BUG_ON(kvm_pmd_huge(*pmd));
 116         pmd_clear(pmd);
 117         kvm_tlb_flush_vmid_ipa(kvm, addr);
 118         pte_free_kernel(NULL, pte_table);
 119         put_page(virt_to_page(pmd));
 120 }
 121
 122 static void unmap_ptes(struct kvm *kvm, pmd_t *pmd,
 123                        phys_addr_t addr, phys_addr_t end)
 124 {
 125         phys_addr_t start_addr = addr;
 126         pte_t *pte, *start_pte;
 127
 128         start_pte = pte = pte_offset_kernel(pmd, addr);
 129         do {
 130                 if (!pte_none(*pte)) {
 131                         kvm_set_pte(pte, __pte(0));
 132                         put_page(virt_to_page(pte));
 133                         kvm_tlb_flush_vmid_ipa(kvm, addr);
 134                 }
 135         } while (pte++, addr += PAGE_SIZE, addr != end);
 136
 137         if (kvm_pte_table_empty(kvm, start_pte))
 138                 clear_pmd_entry(kvm, pmd, start_addr);
 139 }
 140
 141 static void unmap_pmds(struct kvm *kvm, pud_t *pud,
 142                        phys_addr_t addr, phys_addr_t end)
 143 {
 144         phys_addr_t next, start_addr = addr;
 145         pmd_t *pmd, *start_pmd;
 146
 147         start_pmd = pmd = pmd_offset(pud, addr);
 148         do {
 149                 next = kvm_pmd_addr_end(addr, end);
 150                 if (!pmd_none(*pmd)) {
 151                         if (kvm_pmd_huge(*pmd)) {
 152                                 pmd_clear(pmd);
 153                                 kvm_tlb_flush_vmid_ipa(kvm, addr);
 154                                 put_page(virt_to_page(pmd));
 155                         } else {
 156                                 unmap_ptes(kvm, pmd, addr, next);
 157                         }
 158                 }
 159         } while (pmd++, addr = next, addr != end);
 160
 161         if (kvm_pmd_table_empty(kvm, start_pmd))
 162                 clear_pud_entry(kvm, pud, start_addr);
 163 }
 164
 165 static void unmap_puds(struct kvm *kvm, pgd_t *pgd,
 166                        phys_addr_t addr, phys_addr_t end)
 167 {
 168         phys_addr_t next, start_addr = addr;
 169         pud_t *pud, *start_pud;
 170
 171         start_pud = pud = pud_offset(pgd, addr);
 172         do {
 173                 next = kvm_pud_addr_end(addr, end);
 174                 if (!pud_none(*pud)) {
 175                         if (pud_huge(*pud)) {
 176                                 pud_clear(pud);
 177                                 kvm_tlb_flush_vmid_ipa(kvm, addr);
 178                                 put_page(virt_to_page(pud));
 179                         } else {
 180                                 unmap_pmds(kvm, pud, addr, next);
 181                         }
 182                 }
 183         } while (pud++, addr = next, addr != end);
 184
 185         if (kvm_pud_table_empty(kvm, start_pud))
 186                 clear_pgd_entry(kvm, pgd, start_addr);
 187 }
 188
 189
 190 static void unmap_range(struct kvm *kvm, pgd_t *pgdp,
 191                         phys_addr_t start, u64 size)
 192 {
 193         pgd_t *pgd;
 194         phys_addr_t addr = start, end = start + size;
 195         phys_addr_t next;
 196
 197         pgd = pgdp + pgd_index(addr);
 198         do {
 199                 next = kvm_pgd_addr_end(addr, end);
 200                 unmap_puds(kvm, pgd, addr, next);
 201         } while (pgd++, addr = next, addr != end);
 202 }
 203
 204 static void stage2_flush_ptes(struct kvm *kvm, pmd_t *pmd,
 205                               phys_addr_t addr, phys_addr_t end)
 206 {
 207         pte_t *pte;
 208
 209         pte = pte_offset_kernel(pmd, addr);
 210         do {
 211                 if (!pte_none(*pte)) {
 212                         hva_t hva = gfn_to_hva(kvm, addr >> PAGE_SHIFT);
 213                         kvm_flush_dcache_to_poc((void*)hva, PAGE_SIZE);
 214                 }
 215         } while (pte++, addr += PAGE_SIZE, addr != end);
 216 }
 217
 218 static void stage2_flush_pmds(struct kvm *kvm, pud_t *pud,
 219                               phys_addr_t addr, phys_addr_t end)
 220 {
 221         pmd_t *pmd;
 222         phys_addr_t next;
 223
 224         pmd = pmd_offset(pud, addr);
 225         do {
 226                 next = kvm_pmd_addr_end(addr, end);
 227                 if (!pmd_none(*pmd)) {
 228                         if (kvm_pmd_huge(*pmd)) {
 229                                 hva_t hva = gfn_to_hva(kvm, addr >> PAGE_SHIFT);
 230                                 kvm_flush_dcache_to_poc((void*)hva, PMD_SIZE);
 231                         } else {
 232                                 stage2_flush_ptes(kvm, pmd, addr, next);
 233                         }
 234                 }
 235         } while (pmd++, addr = next, addr != end);
 236 }
 237
 238 static void stage2_flush_puds(struct kvm *kvm, pgd_t *pgd,
 239                               phys_addr_t addr, phys_addr_t end)
 240 {
 241         pud_t *pud;
 242         phys_addr_t next;
 243
 244         pud = pud_offset(pgd, addr);
 245         do {
 246                 next = kvm_pud_addr_end(addr, end);
 247                 if (!pud_none(*pud)) {
 248                         if (pud_huge(*pud)) {
 249                                 hva_t hva = gfn_to_hva(kvm, addr >> PAGE_SHIFT);
 250                                 kvm_flush_dcache_to_poc((void*)hva, PUD_SIZE);
 251                         } else {
 252                                 stage2_flush_pmds(kvm, pud, addr, next);
 253                         }
 254                 }
 255         } while (pud++, addr = next, addr != end);
 256 }
 257
 258 static void stage2_flush_memslot(struct kvm *kvm,
 259                                  struct kvm_memory_slot *memslot)
 260 {
 261         phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
 262         phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
 263         phys_addr_t next;
 264         pgd_t *pgd;
 265
 266         pgd = kvm->arch.pgd + pgd_index(addr);
 267         do {
 268                 next = kvm_pgd_addr_end(addr, end);
 269                 stage2_flush_puds(kvm, pgd, addr, next);
 270         } while (pgd++, addr = next, addr != end);
 271 }
 272
 273 /**
 274  * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
 275  * @kvm: The struct kvm pointer
 276  *
 277  * Go through the stage 2 page tables and invalidate any cache lines
 278  * backing memory already mapped to the VM.
 279  */
 280 void stage2_flush_vm(struct kvm *kvm)
 281 {
 282         struct kvm_memslots *slots;
 283         struct kvm_memory_slot *memslot;
 284         int idx;
 285
 286         idx = srcu_read_lock(&kvm->srcu);
 287         spin_lock(&kvm->mmu_lock);
 288
 289         slots = kvm_memslots(kvm);
 290         kvm_for_each_memslot(memslot, slots)
 291                 stage2_flush_memslot(kvm, memslot);
 292
 293         spin_unlock(&kvm->mmu_lock);
 294         srcu_read_unlock(&kvm->srcu, idx);
 295 }
 296
 297 /**
 298  * free_boot_hyp_pgd - free HYP boot page tables
 299  *
 300  * Free the HYP boot page tables. The bounce page is also freed.
 301  */
 302 void free_boot_hyp_pgd(void)
 303 {
 304         mutex_lock(&kvm_hyp_pgd_mutex);
 305
 306         if (boot_hyp_pgd) {
 307                 unmap_range(NULL, boot_hyp_pgd, hyp_idmap_start, PAGE_SIZE);
 308                 unmap_range(NULL, boot_hyp_pgd, TRAMPOLINE_VA, PAGE_SIZE);
 309                 free_pages((unsigned long)boot_hyp_pgd, hyp_pgd_order);
 310                 boot_hyp_pgd = NULL;
 311         }
 312
 313         if (hyp_pgd)
 314                 unmap_range(NULL, hyp_pgd, TRAMPOLINE_VA, PAGE_SIZE);
 315
 316         free_page((unsigned long)init_bounce_page);
 317         init_bounce_page = NULL;
 318
 319         mutex_unlock(&kvm_hyp_pgd_mutex);
 320 }
 321
 322 /**
 323  * free_hyp_pgds - free Hyp-mode page tables
 324  *
 325  * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
 326  * therefore contains either mappings in the kernel memory area (above
 327  * PAGE_OFFSET), or device mappings in the vmalloc range (from
 328  * VMALLOC_START to VMALLOC_END).
 329  *
 330  * boot_hyp_pgd should only map two pages for the init code.
 331  */
 332 void free_hyp_pgds(void)
 333 {
 334         unsigned long addr;
 335
 336         free_boot_hyp_pgd();
 337
 338         mutex_lock(&kvm_hyp_pgd_mutex);
 339
 340         if (hyp_pgd) {
 341                 for (addr = PAGE_OFFSET; virt_addr_valid(addr); addr += PGDIR_SIZE)
 342                         unmap_range(NULL, hyp_pgd, KERN_TO_HYP(addr), PGDIR_SIZE);
 343                 for (addr = VMALLOC_START; is_vmalloc_addr((void*)addr); addr += PGDIR_SIZE)
 344                         unmap_range(NULL, hyp_pgd, KERN_TO_HYP(addr), PGDIR_SIZE);
 345
 346                 free_pages((unsigned long)hyp_pgd, hyp_pgd_order);
 347                 hyp_pgd = NULL;
 348         }
 349
 350         mutex_unlock(&kvm_hyp_pgd_mutex);
 351 }
 352
 353 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
 354                                     unsigned long end, unsigned long pfn,
 355                                     pgprot_t prot)
 356 {
 357         pte_t *pte;
 358         unsigned long addr;
 359
 360         addr = start;
 361         do {
 362                 pte = pte_offset_kernel(pmd, addr);
 363                 kvm_set_pte(pte, pfn_pte(pfn, prot));
 364                 get_page(virt_to_page(pte));
 365                 kvm_flush_dcache_to_poc(pte, sizeof(*pte));
 366                 pfn++;
 367         } while (addr += PAGE_SIZE, addr != end);
 368 }
 369
 370 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
 371                                    unsigned long end, unsigned long pfn,
 372                                    pgprot_t prot)
 373 {
 374         pmd_t *pmd;
 375         pte_t *pte;
 376         unsigned long addr, next;
 377
 378         addr = start;
 379         do {
 380                 pmd = pmd_offset(pud, addr);
 381
 382                 BUG_ON(pmd_sect(*pmd));
 383
 384                 if (pmd_none(*pmd)) {
 385                         pte = pte_alloc_one_kernel(NULL, addr);
 386                         if (!pte) {
 387                                 kvm_err("Cannot allocate Hyp pte\n");
 388                                 return -ENOMEM;
 389                         }
 390                         pmd_populate_kernel(NULL, pmd, pte);
 391                         get_page(virt_to_page(pmd));
 392                         kvm_flush_dcache_to_poc(pmd, sizeof(*pmd));
 393                 }
 394
 395                 next = pmd_addr_end(addr, end);
 396
 397                 create_hyp_pte_mappings(pmd, addr, next, pfn, prot);
 398                 pfn += (next - addr) >> PAGE_SHIFT;
 399         } while (addr = next, addr != end);
 400
 401         return 0;
 402 }
 403
 404 static int create_hyp_pud_mappings(pgd_t *pgd, unsigned long start,
 405                                    unsigned long end, unsigned long pfn,
 406                                    pgprot_t prot)
 407 {
 408         pud_t *pud;
 409         pmd_t *pmd;
 410         unsigned long addr, next;
 411         int ret;
 412
 413         addr = start;
 414         do {
 415                 pud = pud_offset(pgd, addr);
 416
 417                 if (pud_none_or_clear_bad(pud)) {
 418                         pmd = pmd_alloc_one(NULL, addr);
 419                         if (!pmd) {
 420                                 kvm_err("Cannot allocate Hyp pmd\n");
 421                                 return -ENOMEM;
 422                         }
 423                         pud_populate(NULL, pud, pmd);
 424                         get_page(virt_to_page(pud));
 425                         kvm_flush_dcache_to_poc(pud, sizeof(*pud));
 426                 }
 427
 428                 next = pud_addr_end(addr, end);
 429                 ret = create_hyp_pmd_mappings(pud, addr, next, pfn, prot);
 430                 if (ret)
 431                         return ret;
 432                 pfn += (next - addr) >> PAGE_SHIFT;
 433         } while (addr = next, addr != end);
 434
 435         return 0;
 436 }
 437
 438 static int __create_hyp_mappings(pgd_t *pgdp,
 439                                  unsigned long start, unsigned long end,
 440                                  unsigned long pfn, pgprot_t prot)
 441 {
 442         pgd_t *pgd;
 443         pud_t *pud;
 444         unsigned long addr, next;
 445         int err = 0;
 446
 447         mutex_lock(&kvm_hyp_pgd_mutex);
 448         addr = start & PAGE_MASK;
 449         end = PAGE_ALIGN(end);
 450         do {
 451                 pgd = pgdp + pgd_index(addr);
 452
 453                 if (pgd_none(*pgd)) {
 454                         pud = pud_alloc_one(NULL, addr);
 455                         if (!pud) {
 456                                 kvm_err("Cannot allocate Hyp pud\n");
 457                                 err = -ENOMEM;
 458                                 goto out;
 459                         }
 460                         pgd_populate(NULL, pgd, pud);
 461                         get_page(virt_to_page(pgd));
 462                         kvm_flush_dcache_to_poc(pgd, sizeof(*pgd));
 463                 }
 464
 465                 next = pgd_addr_end(addr, end);
 466                 err = create_hyp_pud_mappings(pgd, addr, next, pfn, prot);
 467                 if (err)
 468                         goto out;
 469                 pfn += (next - addr) >> PAGE_SHIFT;
 470         } while (addr = next, addr != end);
 471 out:
 472         mutex_unlock(&kvm_hyp_pgd_mutex);
 473         return err;
 474 }
 475
 476 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
 477 {
 478         if (!is_vmalloc_addr(kaddr)) {
 479                 BUG_ON(!virt_addr_valid(kaddr));
 480                 return __pa(kaddr);
 481         } else {
 482                 return page_to_phys(vmalloc_to_page(kaddr)) +
 483                        offset_in_page(kaddr);
 484         }
 485 }
 486
 487 /**
 488  * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
 489  * @from:       The virtual kernel start address of the range
 490  * @to:         The virtual kernel end address of the range (exclusive)
 491  *
 492  * The same virtual address as the kernel virtual address is also used
 493  * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
 494  * physical pages.
 495  */
 496 int create_hyp_mappings(void *from, void *to)
 497 {
 498         phys_addr_t phys_addr;
 499         unsigned long virt_addr;
 500         unsigned long start = KERN_TO_HYP((unsigned long)from);
 501         unsigned long end = KERN_TO_HYP((unsigned long)to);
 502
 503         start = start & PAGE_MASK;
 504         end = PAGE_ALIGN(end);
 505
 506         for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
 507                 int err;
 508
 509                 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
 510                 err = __create_hyp_mappings(hyp_pgd, virt_addr,
 511                                             virt_addr + PAGE_SIZE,
 512                                             __phys_to_pfn(phys_addr),
 513                                             PAGE_HYP);
 514                 if (err)
 515                         return err;
 516         }
 517
 518         return 0;
 519 }
 520
 521 /**
 522  * create_hyp_io_mappings - duplicate a kernel IO mapping into Hyp mode
 523  * @from:       The kernel start VA of the range
 524  * @to:         The kernel end VA of the range (exclusive)
 525  * @phys_addr:  The physical start address which gets mapped
 526  *
 527  * The resulting HYP VA is the same as the kernel VA, modulo
 528  * HYP_PAGE_OFFSET.
 529  */
 530 int create_hyp_io_mappings(void *from, void *to, phys_addr_t phys_addr)
 531 {
 532         unsigned long start = KERN_TO_HYP((unsigned long)from);
 533         unsigned long end = KERN_TO_HYP((unsigned long)to);
 534
 535         /* Check for a valid kernel IO mapping */
 536         if (!is_vmalloc_addr(from) || !is_vmalloc_addr(to - 1))
 537                 return -EINVAL;
 538
 539         return __create_hyp_mappings(hyp_pgd, start, end,
 540                                      __phys_to_pfn(phys_addr), PAGE_HYP_DEVICE);
 541 }
 542
 543 /**
 544  * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
 545  * @kvm:        The KVM struct pointer for the VM.
 546  *
 547  * Allocates the 1st level table only of size defined by S2_PGD_ORDER (can
 548  * support either full 40-bit input addresses or limited to 32-bit input
 549  * addresses). Clears the allocated pages.
 550  *
 551  * Note we don't need locking here as this is only called when the VM is
 552  * created, which can only be done once.
 553  */
 554 int kvm_alloc_stage2_pgd(struct kvm *kvm)
 555 {
 556         int ret;
 557         pgd_t *pgd;
 558
 559         if (kvm->arch.pgd != NULL) {
 560                 kvm_err("kvm_arch already initialized?\n");
 561                 return -EINVAL;
 562         }
 563
 564         if (KVM_PREALLOC_LEVEL > 0) {
 565                 /*
 566                  * Allocate fake pgd for the page table manipulation macros to
 567                  * work.  This is not used by the hardware and we have no
 568                  * alignment requirement for this allocation.
 569                  */
 570                 pgd = (pgd_t *)kmalloc(PTRS_PER_S2_PGD * sizeof(pgd_t),
 571                                        GFP_KERNEL | __GFP_ZERO);
 572         } else {
 573                 /*
 574                  * Allocate actual first-level Stage-2 page table used by the
 575                  * hardware for Stage-2 page table walks.
 576                  */
 577                 pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, S2_PGD_ORDER);
 578         }
 579
 580         if (!pgd)
 581                 return -ENOMEM;
 582
 583         ret = kvm_prealloc_hwpgd(kvm, pgd);
 584         if (ret)
 585                 goto out_err;
 586
 587         kvm_clean_pgd(pgd);
 588         kvm->arch.pgd = pgd;
 589         return 0;
 590 out_err:
 591         if (KVM_PREALLOC_LEVEL > 0)
 592                 kfree(pgd);
 593         else
 594                 free_pages((unsigned long)pgd, S2_PGD_ORDER);
 595         return ret;
 596 }
 597
 598 /**
 599  * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
 600  * @kvm:   The VM pointer
 601  * @start: The intermediate physical base address of the range to unmap
 602  * @size:  The size of the area to unmap
 603  *
 604  * Clear a range of stage-2 mappings, lowering the various ref-counts.  Must
 605  * be called while holding mmu_lock (unless for freeing the stage2 pgd before
 606  * destroying the VM), otherwise another faulting VCPU may come in and mess
 607  * with things behind our backs.
 608  */
 609 static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
 610 {
 611         unmap_range(kvm, kvm->arch.pgd, start, size);
 612 }
 613
 614 /**
 615  * kvm_free_stage2_pgd - free all stage-2 tables
 616  * @kvm:        The KVM struct pointer for the VM.
 617  *
 618  * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
 619  * underlying level-2 and level-3 tables before freeing the actual level-1 table
 620  * and setting the struct pointer to NULL.
 621  *
 622  * Note we don't need locking here as this is only called when the VM is
 623  * destroyed, which can only be done once.
 624  */
 625 void kvm_free_stage2_pgd(struct kvm *kvm)
 626 {
 627         if (kvm->arch.pgd == NULL)
 628                 return;
 629
 630         unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
 631         kvm_free_hwpgd(kvm);
 632         if (KVM_PREALLOC_LEVEL > 0)
 633                 kfree(kvm->arch.pgd);
 634         else
 635                 free_pages((unsigned long)kvm->arch.pgd, S2_PGD_ORDER);
 636         kvm->arch.pgd = NULL;
 637 }
 638
 639 static pud_t *stage2_get_pud(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
 640                              phys_addr_t addr)
 641 {
 642         pgd_t *pgd;
 643         pud_t *pud;
 644
 645         pgd = kvm->arch.pgd + pgd_index(addr);
 646         if (WARN_ON(pgd_none(*pgd))) {
 647                 if (!cache)
 648                         return NULL;
 649                 pud = mmu_memory_cache_alloc(cache);
 650                 pgd_populate(NULL, pgd, pud);
 651                 get_page(virt_to_page(pgd));
 652         }
 653
 654         return pud_offset(pgd, addr);
 655 }
 656
 657 static pmd_t *stage2_get_pmd(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
 658                              phys_addr_t addr)
 659 {
 660         pud_t *pud;
 661         pmd_t *pmd;
 662
 663         pud = stage2_get_pud(kvm, cache, addr);
 664         if (pud_none(*pud)) {
 665                 if (!cache)
 666                         return NULL;
 667                 pmd = mmu_memory_cache_alloc(cache);
 668                 pud_populate(NULL, pud, pmd);
 669                 get_page(virt_to_page(pud));
 670         }
 671
 672         return pmd_offset(pud, addr);
 673 }
 674
 675 static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
 676                                *cache, phys_addr_t addr, const pmd_t *new_pmd)
 677 {
 678         pmd_t *pmd, old_pmd;
 679
 680         pmd = stage2_get_pmd(kvm, cache, addr);
 681         VM_BUG_ON(!pmd);
 682
 683         /*
 684          * Mapping in huge pages should only happen through a fault.  If a
 685          * page is merged into a transparent huge page, the individual
 686          * subpages of that huge page should be unmapped through MMU
 687          * notifiers before we get here.
 688          *
 689          * Merging of CompoundPages is not supported; they should become
 690          * splitting first, unmapped, merged, and mapped back in on-demand.
 691          */
 692         VM_BUG_ON(pmd_present(*pmd) && pmd_pfn(*pmd) != pmd_pfn(*new_pmd));
 693
 694         old_pmd = *pmd;
 695         kvm_set_pmd(pmd, *new_pmd);
 696         if (pmd_present(old_pmd))
 697                 kvm_tlb_flush_vmid_ipa(kvm, addr);
 698         else
 699                 get_page(virt_to_page(pmd));
 700         return 0;
 701 }
 702
 703 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
 704                           phys_addr_t addr, const pte_t *new_pte, bool iomap)
 705 {
 706         pmd_t *pmd;
 707         pte_t *pte, old_pte;
 708
 709         /* Create stage-2 page table mapping - Levels 0 and 1 */
 710         pmd = stage2_get_pmd(kvm, cache, addr);
 711         if (!pmd) {
 712                 /*
 713                  * Ignore calls from kvm_set_spte_hva for unallocated
 714                  * address ranges.
 715                  */
 716                 return 0;
 717         }
 718
 719         /* Create stage-2 page mappings - Level 2 */
 720         if (pmd_none(*pmd)) {
 721                 if (!cache)
 722                         return 0; /* ignore calls from kvm_set_spte_hva */
 723                 pte = mmu_memory_cache_alloc(cache);
 724                 kvm_clean_pte(pte);
 725                 pmd_populate_kernel(NULL, pmd, pte);
 726                 get_page(virt_to_page(pmd));
 727         }
 728
 729         pte = pte_offset_kernel(pmd, addr);
 730
 731         if (iomap && pte_present(*pte))
 732                 return -EFAULT;
 733
 734         /* Create 2nd stage page table mapping - Level 3 */
 735         old_pte = *pte;
 736         kvm_set_pte(pte, *new_pte);
 737         if (pte_present(old_pte))
 738                 kvm_tlb_flush_vmid_ipa(kvm, addr);
 739         else
 740                 get_page(virt_to_page(pte));
 741
 742         return 0;
 743 }
 744
 745 /**
 746  * kvm_phys_addr_ioremap - map a device range to guest IPA
 747  *
 748  * @kvm:        The KVM pointer
 749  * @guest_ipa:  The IPA at which to insert the mapping
 750  * @pa:         The physical address of the device
 751  * @size:       The size of the mapping
 752  */
 753 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
 754                           phys_addr_t pa, unsigned long size, bool writable)
 755 {
 756         phys_addr_t addr, end;
 757         int ret = 0;
 758         unsigned long pfn;
 759         struct kvm_mmu_memory_cache cache = { 0, };
 760
 761         end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
 762         pfn = __phys_to_pfn(pa);
 763
 764         for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
 765                 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
 766
 767                 if (writable)
 768                         kvm_set_s2pte_writable(&pte);
 769
 770                 ret = mmu_topup_memory_cache(&cache, KVM_MMU_CACHE_MIN_PAGES,
 771                                                 KVM_NR_MEM_OBJS);
 772                 if (ret)
 773                         goto out;
 774                 spin_lock(&kvm->mmu_lock);
 775                 ret = stage2_set_pte(kvm, &cache, addr, &pte, true);
 776                 spin_unlock(&kvm->mmu_lock);
 777                 if (ret)
 778                         goto out;
 779
 780                 pfn++;
 781         }
 782
 783 out:
 784         mmu_free_memory_cache(&cache);
 785         return ret;
 786 }
 787
 788 static bool transparent_hugepage_adjust(pfn_t *pfnp, phys_addr_t *ipap)
 789 {
 790         pfn_t pfn = *pfnp;
 791         gfn_t gfn = *ipap >> PAGE_SHIFT;
 792
 793         if (PageTransCompound(pfn_to_page(pfn))) {
 794                 unsigned long mask;
 795                 /*
 796                  * The address we faulted on is backed by a transparent huge
 797                  * page.  However, because we map the compound huge page and
 798                  * not the individual tail page, we need to transfer the
 799                  * refcount to the head page.  We have to be careful that the
 800                  * THP doesn't start to split while we are adjusting the
 801                  * refcounts.
 802                  *
 803                  * We are sure this doesn't happen, because mmu_notifier_retry
 804                  * was successful and we are holding the mmu_lock, so if this
 805                  * THP is trying to split, it will be blocked in the mmu
 806                  * notifier before touching any of the pages, specifically
 807                  * before being able to call __split_huge_page_refcount().
 808                  *
 809                  * We can therefore safely transfer the refcount from PG_tail
 810                  * to PG_head and switch the pfn from a tail page to the head
 811                  * page accordingly.
 812                  */
 813                 mask = PTRS_PER_PMD - 1;
 814                 VM_BUG_ON((gfn & mask) != (pfn & mask));
 815                 if (pfn & mask) {
 816                         *ipap &= PMD_MASK;
 817                         kvm_release_pfn_clean(pfn);
 818                         pfn &= ~mask;
 819                         kvm_get_pfn(pfn);
 820                         *pfnp = pfn;
 821                 }
 822
 823                 return true;
 824         }
 825
 826         return false;
 827 }
 828
 829 static bool kvm_is_write_fault(struct kvm_vcpu *vcpu)
 830 {
 831         if (kvm_vcpu_trap_is_iabt(vcpu))
 832                 return false;
 833
 834         return kvm_vcpu_dabt_iswrite(vcpu);
 835 }
 836
 837 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
 838                           struct kvm_memory_slot *memslot, unsigned long hva,
 839                           unsigned long fault_status)
 840 {
 841         int ret;
 842         bool write_fault, writable, hugetlb = false, force_pte = false;
 843         unsigned long mmu_seq;
 844         gfn_t gfn = fault_ipa >> PAGE_SHIFT;
 845         struct kvm *kvm = vcpu->kvm;
 846         struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
 847         struct vm_area_struct *vma;
 848         pfn_t pfn;
 849         pgprot_t mem_type = PAGE_S2;
 850
 851         write_fault = kvm_is_write_fault(vcpu);
 852         if (fault_status == FSC_PERM && !write_fault) {
 853                 kvm_err("Unexpected L2 read permission error\n");
 854                 return -EFAULT;
 855         }
 856
 857         /* Let's check if we will get back a huge page backed by hugetlbfs */
 858         down_read(&current->mm->mmap_sem);
 859         vma = find_vma_intersection(current->mm, hva, hva + 1);
 860         if (unlikely(!vma)) {
 861                 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
 862                 up_read(&current->mm->mmap_sem);
 863                 return -EFAULT;
 864         }
 865
 866         if (is_vm_hugetlb_page(vma)) {
 867                 hugetlb = true;
 868                 gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
 869         } else {
 870                 /*
 871                  * Pages belonging to memslots that don't have the same
 872                  * alignment for userspace and IPA cannot be mapped using
 873                  * block descriptors even if the pages belong to a THP for
 874                  * the process, because the stage-2 block descriptor will
 875                  * cover more than a single THP and we loose atomicity for
 876                  * unmapping, updates, and splits of the THP or other pages
 877                  * in the stage-2 block range.
 878                  */
 879                 if ((memslot->userspace_addr & ~PMD_MASK) !=
 880                     ((memslot->base_gfn << PAGE_SHIFT) & ~PMD_MASK))
 881                         force_pte = true;
 882         }
 883         up_read(&current->mm->mmap_sem);
 884
 885         /* We need minimum second+third level pages */
 886         ret = mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES,
 887                                      KVM_NR_MEM_OBJS);
 888         if (ret)
 889                 return ret;
 890
 891         mmu_seq = vcpu->kvm->mmu_notifier_seq;
 892         /*
 893          * Ensure the read of mmu_notifier_seq happens before we call
 894          * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
 895          * the page we just got a reference to gets unmapped before we have a
 896          * chance to grab the mmu_lock, which ensure that if the page gets
 897          * unmapped afterwards, the call to kvm_unmap_hva will take it away
 898          * from us again properly. This smp_rmb() interacts with the smp_wmb()
 899          * in kvm_mmu_notifier_invalidate_<page|range_end>.
 900          */
 901         smp_rmb();
 902
 903         pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
 904         if (is_error_pfn(pfn))
 905                 return -EFAULT;
 906
 907         if (kvm_is_mmio_pfn(pfn))
 908                 mem_type = PAGE_S2_DEVICE;
 909
 910         spin_lock(&kvm->mmu_lock);
 911         if (mmu_notifier_retry(kvm, mmu_seq))
 912                 goto out_unlock;
 913         if (!hugetlb && !force_pte)
 914                 hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
 915
 916         if (hugetlb) {
 917                 pmd_t new_pmd = pfn_pmd(pfn, mem_type);
 918                 new_pmd = pmd_mkhuge(new_pmd);
 919                 if (writable) {
 920                         kvm_set_s2pmd_writable(&new_pmd);
 921                         kvm_set_pfn_dirty(pfn);
 922                 }
 923                 coherent_cache_guest_page(vcpu, hva & PMD_MASK, PMD_SIZE);
 924                 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
 925         } else {
 926                 pte_t new_pte = pfn_pte(pfn, mem_type);
 927                 if (writable) {
 928                         kvm_set_s2pte_writable(&new_pte);
 929                         kvm_set_pfn_dirty(pfn);
 930                 }
 931                 coherent_cache_guest_page(vcpu, hva, PAGE_SIZE);
 932                 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte,
 933                         pgprot_val(mem_type) == pgprot_val(PAGE_S2_DEVICE));
 934         }
 935
 936
 937 out_unlock:
 938         spin_unlock(&kvm->mmu_lock);
 939         kvm_release_pfn_clean(pfn);
 940         return ret;
 941 }
 942
 943 /**
 944  * kvm_handle_guest_abort - handles all 2nd stage aborts
 945  * @vcpu:       the VCPU pointer
 946  * @run:        the kvm_run structure
 947  *
 948  * Any abort that gets to the host is almost guaranteed to be caused by a
 949  * missing second stage translation table entry, which can mean that either the
 950  * guest simply needs more memory and we must allocate an appropriate page or it
 951  * can mean that the guest tried to access I/O memory, which is emulated by user
 952  * space. The distinction is based on the IPA causing the fault and whether this
 953  * memory region has been registered as standard RAM by user space.
 954  */
 955 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
 956 {
 957         unsigned long fault_status;
 958         phys_addr_t fault_ipa;
 959         struct kvm_memory_slot *memslot;
 960         unsigned long hva;
 961         bool is_iabt, write_fault, writable;
 962         gfn_t gfn;
 963         int ret, idx;
 964
 965         is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
 966         fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
 967
 968         trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
 969                               kvm_vcpu_get_hfar(vcpu), fault_ipa);
 970
 971         /* Check the stage-2 fault is trans. fault or write fault */
 972         fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
 973         if (fault_status != FSC_FAULT && fault_status != FSC_PERM) {
 974                 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
 975                         kvm_vcpu_trap_get_class(vcpu),
 976                         (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
 977                         (unsigned long)kvm_vcpu_get_hsr(vcpu));
 978                 return -EFAULT;
 979         }
 980
 981         idx = srcu_read_lock(&vcpu->kvm->srcu);
 982
 983         gfn = fault_ipa >> PAGE_SHIFT;
 984         memslot = gfn_to_memslot(vcpu->kvm, gfn);
 985         hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
 986         write_fault = kvm_is_write_fault(vcpu);
 987         if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
 988                 if (is_iabt) {
 989                         /* Prefetch Abort on I/O address */
 990                         kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
 991                         ret = 1;
 992                         goto out_unlock;
 993                 }
 994
 995                 /*
 996                  * The IPA is reported as [MAX:12], so we need to
 997                  * complement it with the bottom 12 bits from the
 998                  * faulting VA. This is always 12 bits, irrespective
 999                  * of the page size.
1000                  */
1001                 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1002                 ret = io_mem_abort(vcpu, run, fault_ipa);
1003                 goto out_unlock;
1004         }
1005
1006         /* Userspace should not be able to register out-of-bounds IPAs */
1007         VM_BUG_ON(fault_ipa >= KVM_PHYS_SIZE);
1008
1009         ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1010         if (ret == 0)
1011                 ret = 1;
1012 out_unlock:
1013         srcu_read_unlock(&vcpu->kvm->srcu, idx);
1014         return ret;
1015 }
1016
1017 static void handle_hva_to_gpa(struct kvm *kvm,
1018                               unsigned long start,
1019                               unsigned long end,
1020                               void (*handler)(struct kvm *kvm,
1021                                               gpa_t gpa, void *data),
1022                               void *data)
1023 {
1024         struct kvm_memslots *slots;
1025         struct kvm_memory_slot *memslot;
1026
1027         slots = kvm_memslots(kvm);
1028
1029         /* we only care about the pages that the guest sees */
1030         kvm_for_each_memslot(memslot, slots) {
1031                 unsigned long hva_start, hva_end;
1032                 gfn_t gfn, gfn_end;
1033
1034                 hva_start = max(start, memslot->userspace_addr);
1035                 hva_end = min(end, memslot->userspace_addr +
1036                                         (memslot->npages << PAGE_SHIFT));
1037                 if (hva_start >= hva_end)
1038                         continue;
1039
1040                 /*
1041                  * {gfn(page) | page intersects with [hva_start, hva_end)} =
1042                  * {gfn_start, gfn_start+1, ..., gfn_end-1}.
1043                  */
1044                 gfn = hva_to_gfn_memslot(hva_start, memslot);
1045                 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
1046
1047                 for (; gfn < gfn_end; ++gfn) {
1048                         gpa_t gpa = gfn << PAGE_SHIFT;
1049                         handler(kvm, gpa, data);
1050                 }
1051         }
1052 }
1053
1054 static void kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
1055 {
1056         unmap_stage2_range(kvm, gpa, PAGE_SIZE);
1057 }
1058
1059 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1060 {
1061         unsigned long end = hva + PAGE_SIZE;
1062
1063         if (!kvm->arch.pgd)
1064                 return 0;
1065
1066         trace_kvm_unmap_hva(hva);
1067         handle_hva_to_gpa(kvm, hva, end, &kvm_unmap_hva_handler, NULL);
1068         return 0;
1069 }
1070
1071 int kvm_unmap_hva_range(struct kvm *kvm,
1072                         unsigned long start, unsigned long end)
1073 {
1074         if (!kvm->arch.pgd)
1075                 return 0;
1076
1077         trace_kvm_unmap_hva_range(start, end);
1078         handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
1079         return 0;
1080 }
1081
1082 static void kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, void *data)
1083 {
1084         pte_t *pte = (pte_t *)data;
1085
1086         stage2_set_pte(kvm, NULL, gpa, pte, false);
1087 }
1088
1089
1090 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1091 {
1092         unsigned long end = hva + PAGE_SIZE;
1093         pte_t stage2_pte;
1094
1095         if (!kvm->arch.pgd)
1096                 return;
1097
1098         trace_kvm_set_spte_hva(hva);
1099         stage2_pte = pfn_pte(pte_pfn(pte), PAGE_S2);
1100         handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
1101 }
1102
1103 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1104 {
1105         mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
1106 }
1107
1108 phys_addr_t kvm_mmu_get_httbr(void)
1109 {
1110         return virt_to_phys(hyp_pgd);
1111 }
1112
1113 phys_addr_t kvm_mmu_get_boot_httbr(void)
1114 {
1115         return virt_to_phys(boot_hyp_pgd);
1116 }
1117
1118 phys_addr_t kvm_get_idmap_vector(void)
1119 {
1120         return hyp_idmap_vector;
1121 }
1122
1123 int kvm_mmu_init(void)
1124 {
1125         int err;
1126
1127         hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
1128         hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
1129         hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
1130
1131         if ((hyp_idmap_start ^ hyp_idmap_end) & PAGE_MASK) {
1132                 /*
1133                  * Our init code is crossing a page boundary. Allocate
1134                  * a bounce page, copy the code over and use that.
1135                  */
1136                 size_t len = __hyp_idmap_text_end - __hyp_idmap_text_start;
1137                 phys_addr_t phys_base;
1138
1139                 init_bounce_page = (void *)__get_free_page(GFP_KERNEL);
1140                 if (!init_bounce_page) {
1141                         kvm_err("Couldn't allocate HYP init bounce page\n");
1142                         err = -ENOMEM;
1143                         goto out;
1144                 }
1145
1146                 memcpy(init_bounce_page, __hyp_idmap_text_start, len);
1147                 /*
1148                  * Warning: the code we just copied to the bounce page
1149                  * must be flushed to the point of coherency.
1150                  * Otherwise, the data may be sitting in L2, and HYP
1151                  * mode won't be able to observe it as it runs with
1152                  * caches off at that point.
1153                  */
1154                 kvm_flush_dcache_to_poc(init_bounce_page, len);
1155
1156                 phys_base = kvm_virt_to_phys(init_bounce_page);
1157                 hyp_idmap_vector += phys_base - hyp_idmap_start;
1158                 hyp_idmap_start = phys_base;
1159                 hyp_idmap_end = phys_base + len;
1160
1161                 kvm_info("Using HYP init bounce page @%lx\n",
1162                          (unsigned long)phys_base);
1163         }
1164
1165         hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1166         boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1167
1168         if (!hyp_pgd || !boot_hyp_pgd) {
1169                 kvm_err("Hyp mode PGD not allocated\n");
1170                 err = -ENOMEM;
1171                 goto out;
1172         }
1173
1174         /* Create the idmap in the boot page tables */
1175         err =   __create_hyp_mappings(boot_hyp_pgd,
1176                                       hyp_idmap_start, hyp_idmap_end,
1177                                       __phys_to_pfn(hyp_idmap_start),
1178                                       PAGE_HYP);
1179
1180         if (err) {
1181                 kvm_err("Failed to idmap %lx-%lx\n",
1182                         hyp_idmap_start, hyp_idmap_end);
1183                 goto out;
1184         }
1185
1186         /* Map the very same page at the trampoline VA */
1187         err =   __create_hyp_mappings(boot_hyp_pgd,
1188                                       TRAMPOLINE_VA, TRAMPOLINE_VA + PAGE_SIZE,
1189                                       __phys_to_pfn(hyp_idmap_start),
1190                                       PAGE_HYP);
1191         if (err) {
1192                 kvm_err("Failed to map trampoline @%lx into boot HYP pgd\n",
1193                         TRAMPOLINE_VA);
1194                 goto out;
1195         }
1196
1197         /* Map the same page again into the runtime page tables */
1198         err =   __create_hyp_mappings(hyp_pgd,
1199                                       TRAMPOLINE_VA, TRAMPOLINE_VA + PAGE_SIZE,
1200                                       __phys_to_pfn(hyp_idmap_start),
1201                                       PAGE_HYP);
1202         if (err) {
1203                 kvm_err("Failed to map trampoline @%lx into runtime HYP pgd\n",
1204                         TRAMPOLINE_VA);
1205                 goto out;
1206         }
1207
1208         return 0;
1209 out:
1210         free_hyp_pgds();
1211         return err;
1212 }
1213
1214 void kvm_arch_commit_memory_region(struct kvm *kvm,
1215                                    struct kvm_userspace_memory_region *mem,
1216                                    const struct kvm_memory_slot *old,
1217                                    enum kvm_mr_change change)
1218 {
1219 }
1220
1221 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1222                                    struct kvm_memory_slot *memslot,
1223                                    struct kvm_userspace_memory_region *mem,
1224                                    enum kvm_mr_change change)
1225 {
1226         hva_t hva = mem->userspace_addr;
1227         hva_t reg_end = hva + mem->memory_size;
1228         bool writable = !(mem->flags & KVM_MEM_READONLY);
1229         int ret = 0;
1230
1231         if (change != KVM_MR_CREATE && change != KVM_MR_MOVE)
1232                 return 0;
1233
1234         /*
1235          * Prevent userspace from creating a memory region outside of the IPA
1236          * space addressable by the KVM guest IPA space.
1237          */
1238         if (memslot->base_gfn + memslot->npages >=
1239             (KVM_PHYS_SIZE >> PAGE_SHIFT))
1240                 return -EFAULT;
1241
1242         /*
1243          * A memory region could potentially cover multiple VMAs, and any holes
1244          * between them, so iterate over all of them to find out if we can map
1245          * any of them right now.
1246          *
1247          *     +--------------------------------------------+
1248          * +---------------+----------------+   +----------------+
1249          * |   : VMA 1     |      VMA 2     |   |    VMA 3  :    |
1250          * +---------------+----------------+   +----------------+
1251          *     |               memory region                |
1252          *     +--------------------------------------------+
1253          */
1254         do {
1255                 struct vm_area_struct *vma = find_vma(current->mm, hva);
1256                 hva_t vm_start, vm_end;
1257
1258                 if (!vma || vma->vm_start >= reg_end)
1259                         break;
1260
1261                 /*
1262                  * Mapping a read-only VMA is only allowed if the
1263                  * memory region is configured as read-only.
1264                  */
1265                 if (writable && !(vma->vm_flags & VM_WRITE)) {
1266                         ret = -EPERM;
1267                         break;
1268                 }
1269
1270                 /*
1271                  * Take the intersection of this VMA with the memory region
1272                  */
1273                 vm_start = max(hva, vma->vm_start);
1274                 vm_end = min(reg_end, vma->vm_end);
1275
1276                 if (vma->vm_flags & VM_PFNMAP) {
1277                         gpa_t gpa = mem->guest_phys_addr +
1278                                     (vm_start - mem->userspace_addr);
1279                         phys_addr_t pa = (vma->vm_pgoff << PAGE_SHIFT) +
1280                                          vm_start - vma->vm_start;
1281
1282                         ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
1283                                                     vm_end - vm_start,
1284                                                     writable);
1285                         if (ret)
1286                                 break;
1287                 }
1288                 hva = vm_end;
1289         } while (hva < reg_end);
1290
1291         if (ret) {
1292                 spin_lock(&kvm->mmu_lock);
1293                 unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
1294                 spin_unlock(&kvm->mmu_lock);
1295         }
1296         return ret;
1297 }
1298
1299 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
1300                            struct kvm_memory_slot *dont)
1301 {
1302 }
1303
1304 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
1305                             unsigned long npages)
1306 {
1307         return 0;
1308 }
1309
1310 void kvm_arch_memslots_updated(struct kvm *kvm)
1311 {
1312 }
1313
1314 void kvm_arch_flush_shadow_all(struct kvm *kvm)
1315 {
1316 }
1317
1318 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
1319                                    struct kvm_memory_slot *slot)
1320 {
1321         gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
1322         phys_addr_t size = slot->npages << PAGE_SHIFT;
1323
1324         spin_lock(&kvm->mmu_lock);
1325         unmap_stage2_range(kvm, gpa, size);
1326         spin_unlock(&kvm->mmu_lock);
1327 }