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 static void stage2_unmap_memslot(struct kvm *kvm,
 615                                  struct kvm_memory_slot *memslot)
 616 {
 617         hva_t hva = memslot->userspace_addr;
 618         phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
 619         phys_addr_t size = PAGE_SIZE * memslot->npages;
 620         hva_t reg_end = hva + size;
 621
 622         /*
 623          * A memory region could potentially cover multiple VMAs, and any holes
 624          * between them, so iterate over all of them to find out if we should
 625          * unmap any of them.
 626          *
 627          *     +--------------------------------------------+
 628          * +---------------+----------------+   +----------------+
 629          * |   : VMA 1     |      VMA 2     |   |    VMA 3  :    |
 630          * +---------------+----------------+   +----------------+
 631          *     |               memory region                |
 632          *     +--------------------------------------------+
 633          */
 634         do {
 635                 struct vm_area_struct *vma = find_vma(current->mm, hva);
 636                 hva_t vm_start, vm_end;
 637
 638                 if (!vma || vma->vm_start >= reg_end)
 639                         break;
 640
 641                 /*
 642                  * Take the intersection of this VMA with the memory region
 643                  */
 644                 vm_start = max(hva, vma->vm_start);
 645                 vm_end = min(reg_end, vma->vm_end);
 646
 647                 if (!(vma->vm_flags & VM_PFNMAP)) {
 648                         gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
 649                         unmap_stage2_range(kvm, gpa, vm_end - vm_start);
 650                 }
 651                 hva = vm_end;
 652         } while (hva < reg_end);
 653 }
 654
 655 /**
 656  * stage2_unmap_vm - Unmap Stage-2 RAM mappings
 657  * @kvm: The struct kvm pointer
 658  *
 659  * Go through the memregions and unmap any reguler RAM
 660  * backing memory already mapped to the VM.
 661  */
 662 void stage2_unmap_vm(struct kvm *kvm)
 663 {
 664         struct kvm_memslots *slots;
 665         struct kvm_memory_slot *memslot;
 666         int idx;
 667
 668         idx = srcu_read_lock(&kvm->srcu);
 669         spin_lock(&kvm->mmu_lock);
 670
 671         slots = kvm_memslots(kvm);
 672         kvm_for_each_memslot(memslot, slots)
 673                 stage2_unmap_memslot(kvm, memslot);
 674
 675         spin_unlock(&kvm->mmu_lock);
 676         srcu_read_unlock(&kvm->srcu, idx);
 677 }
 678
 679 /**
 680  * kvm_free_stage2_pgd - free all stage-2 tables
 681  * @kvm:        The KVM struct pointer for the VM.
 682  *
 683  * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
 684  * underlying level-2 and level-3 tables before freeing the actual level-1 table
 685  * and setting the struct pointer to NULL.
 686  *
 687  * Note we don't need locking here as this is only called when the VM is
 688  * destroyed, which can only be done once.
 689  */
 690 void kvm_free_stage2_pgd(struct kvm *kvm)
 691 {
 692         if (kvm->arch.pgd == NULL)
 693                 return;
 694
 695         unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
 696         kvm_free_hwpgd(kvm);
 697         if (KVM_PREALLOC_LEVEL > 0)
 698                 kfree(kvm->arch.pgd);
 699         else
 700                 free_pages((unsigned long)kvm->arch.pgd, S2_PGD_ORDER);
 701         kvm->arch.pgd = NULL;
 702 }
 703
 704 static pud_t *stage2_get_pud(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
 705                              phys_addr_t addr)
 706 {
 707         pgd_t *pgd;
 708         pud_t *pud;
 709
 710         pgd = kvm->arch.pgd + pgd_index(addr);
 711         if (WARN_ON(pgd_none(*pgd))) {
 712                 if (!cache)
 713                         return NULL;
 714                 pud = mmu_memory_cache_alloc(cache);
 715                 pgd_populate(NULL, pgd, pud);
 716                 get_page(virt_to_page(pgd));
 717         }
 718
 719         return pud_offset(pgd, addr);
 720 }
 721
 722 static pmd_t *stage2_get_pmd(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
 723                              phys_addr_t addr)
 724 {
 725         pud_t *pud;
 726         pmd_t *pmd;
 727
 728         pud = stage2_get_pud(kvm, cache, addr);
 729         if (pud_none(*pud)) {
 730                 if (!cache)
 731                         return NULL;
 732                 pmd = mmu_memory_cache_alloc(cache);
 733                 pud_populate(NULL, pud, pmd);
 734                 get_page(virt_to_page(pud));
 735         }
 736
 737         return pmd_offset(pud, addr);
 738 }
 739
 740 static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
 741                                *cache, phys_addr_t addr, const pmd_t *new_pmd)
 742 {
 743         pmd_t *pmd, old_pmd;
 744
 745         pmd = stage2_get_pmd(kvm, cache, addr);
 746         VM_BUG_ON(!pmd);
 747
 748         /*
 749          * Mapping in huge pages should only happen through a fault.  If a
 750          * page is merged into a transparent huge page, the individual
 751          * subpages of that huge page should be unmapped through MMU
 752          * notifiers before we get here.
 753          *
 754          * Merging of CompoundPages is not supported; they should become
 755          * splitting first, unmapped, merged, and mapped back in on-demand.
 756          */
 757         VM_BUG_ON(pmd_present(*pmd) && pmd_pfn(*pmd) != pmd_pfn(*new_pmd));
 758
 759         old_pmd = *pmd;
 760         kvm_set_pmd(pmd, *new_pmd);
 761         if (pmd_present(old_pmd))
 762                 kvm_tlb_flush_vmid_ipa(kvm, addr);
 763         else
 764                 get_page(virt_to_page(pmd));
 765         return 0;
 766 }
 767
 768 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
 769                           phys_addr_t addr, const pte_t *new_pte, bool iomap)
 770 {
 771         pmd_t *pmd;
 772         pte_t *pte, old_pte;
 773
 774         /* Create stage-2 page table mapping - Levels 0 and 1 */
 775         pmd = stage2_get_pmd(kvm, cache, addr);
 776         if (!pmd) {
 777                 /*
 778                  * Ignore calls from kvm_set_spte_hva for unallocated
 779                  * address ranges.
 780                  */
 781                 return 0;
 782         }
 783
 784         /* Create stage-2 page mappings - Level 2 */
 785         if (pmd_none(*pmd)) {
 786                 if (!cache)
 787                         return 0; /* ignore calls from kvm_set_spte_hva */
 788                 pte = mmu_memory_cache_alloc(cache);
 789                 kvm_clean_pte(pte);
 790                 pmd_populate_kernel(NULL, pmd, pte);
 791                 get_page(virt_to_page(pmd));
 792         }
 793
 794         pte = pte_offset_kernel(pmd, addr);
 795
 796         if (iomap && pte_present(*pte))
 797                 return -EFAULT;
 798
 799         /* Create 2nd stage page table mapping - Level 3 */
 800         old_pte = *pte;
 801         kvm_set_pte(pte, *new_pte);
 802         if (pte_present(old_pte))
 803                 kvm_tlb_flush_vmid_ipa(kvm, addr);
 804         else
 805                 get_page(virt_to_page(pte));
 806
 807         return 0;
 808 }
 809
 810 /**
 811  * kvm_phys_addr_ioremap - map a device range to guest IPA
 812  *
 813  * @kvm:        The KVM pointer
 814  * @guest_ipa:  The IPA at which to insert the mapping
 815  * @pa:         The physical address of the device
 816  * @size:       The size of the mapping
 817  */
 818 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
 819                           phys_addr_t pa, unsigned long size, bool writable)
 820 {
 821         phys_addr_t addr, end;
 822         int ret = 0;
 823         unsigned long pfn;
 824         struct kvm_mmu_memory_cache cache = { 0, };
 825
 826         end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
 827         pfn = __phys_to_pfn(pa);
 828
 829         for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
 830                 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
 831
 832                 if (writable)
 833                         kvm_set_s2pte_writable(&pte);
 834
 835                 ret = mmu_topup_memory_cache(&cache, KVM_MMU_CACHE_MIN_PAGES,
 836                                                 KVM_NR_MEM_OBJS);
 837                 if (ret)
 838                         goto out;
 839                 spin_lock(&kvm->mmu_lock);
 840                 ret = stage2_set_pte(kvm, &cache, addr, &pte, true);
 841                 spin_unlock(&kvm->mmu_lock);
 842                 if (ret)
 843                         goto out;
 844
 845                 pfn++;
 846         }
 847
 848 out:
 849         mmu_free_memory_cache(&cache);
 850         return ret;
 851 }
 852
 853 static bool transparent_hugepage_adjust(pfn_t *pfnp, phys_addr_t *ipap)
 854 {
 855         pfn_t pfn = *pfnp;
 856         gfn_t gfn = *ipap >> PAGE_SHIFT;
 857
 858         if (PageTransCompound(pfn_to_page(pfn))) {
 859                 unsigned long mask;
 860                 /*
 861                  * The address we faulted on is backed by a transparent huge
 862                  * page.  However, because we map the compound huge page and
 863                  * not the individual tail page, we need to transfer the
 864                  * refcount to the head page.  We have to be careful that the
 865                  * THP doesn't start to split while we are adjusting the
 866                  * refcounts.
 867                  *
 868                  * We are sure this doesn't happen, because mmu_notifier_retry
 869                  * was successful and we are holding the mmu_lock, so if this
 870                  * THP is trying to split, it will be blocked in the mmu
 871                  * notifier before touching any of the pages, specifically
 872                  * before being able to call __split_huge_page_refcount().
 873                  *
 874                  * We can therefore safely transfer the refcount from PG_tail
 875                  * to PG_head and switch the pfn from a tail page to the head
 876                  * page accordingly.
 877                  */
 878                 mask = PTRS_PER_PMD - 1;
 879                 VM_BUG_ON((gfn & mask) != (pfn & mask));
 880                 if (pfn & mask) {
 881                         *ipap &= PMD_MASK;
 882                         kvm_release_pfn_clean(pfn);
 883                         pfn &= ~mask;
 884                         kvm_get_pfn(pfn);
 885                         *pfnp = pfn;
 886                 }
 887
 888                 return true;
 889         }
 890
 891         return false;
 892 }
 893
 894 static bool kvm_is_write_fault(struct kvm_vcpu *vcpu)
 895 {
 896         if (kvm_vcpu_trap_is_iabt(vcpu))
 897                 return false;
 898
 899         return kvm_vcpu_dabt_iswrite(vcpu);
 900 }
 901
 902 static bool kvm_is_device_pfn(unsigned long pfn)
 903 {
 904         return !pfn_valid(pfn);
 905 }
 906
 907 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
 908                           struct kvm_memory_slot *memslot, unsigned long hva,
 909                           unsigned long fault_status)
 910 {
 911         int ret;
 912         bool write_fault, writable, hugetlb = false, force_pte = false;
 913         unsigned long mmu_seq;
 914         gfn_t gfn = fault_ipa >> PAGE_SHIFT;
 915         struct kvm *kvm = vcpu->kvm;
 916         struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
 917         struct vm_area_struct *vma;
 918         pfn_t pfn;
 919         pgprot_t mem_type = PAGE_S2;
 920         bool fault_ipa_uncached;
 921
 922         write_fault = kvm_is_write_fault(vcpu);
 923         if (fault_status == FSC_PERM && !write_fault) {
 924                 kvm_err("Unexpected L2 read permission error\n");
 925                 return -EFAULT;
 926         }
 927
 928         /* Let's check if we will get back a huge page backed by hugetlbfs */
 929         down_read(&current->mm->mmap_sem);
 930         vma = find_vma_intersection(current->mm, hva, hva + 1);
 931         if (unlikely(!vma)) {
 932                 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
 933                 up_read(&current->mm->mmap_sem);
 934                 return -EFAULT;
 935         }
 936
 937         if (is_vm_hugetlb_page(vma)) {
 938                 hugetlb = true;
 939                 gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
 940         } else {
 941                 /*
 942                  * Pages belonging to memslots that don't have the same
 943                  * alignment for userspace and IPA cannot be mapped using
 944                  * block descriptors even if the pages belong to a THP for
 945                  * the process, because the stage-2 block descriptor will
 946                  * cover more than a single THP and we loose atomicity for
 947                  * unmapping, updates, and splits of the THP or other pages
 948                  * in the stage-2 block range.
 949                  */
 950                 if ((memslot->userspace_addr & ~PMD_MASK) !=
 951                     ((memslot->base_gfn << PAGE_SHIFT) & ~PMD_MASK))
 952                         force_pte = true;
 953         }
 954         up_read(&current->mm->mmap_sem);
 955
 956         /* We need minimum second+third level pages */
 957         ret = mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES,
 958                                      KVM_NR_MEM_OBJS);
 959         if (ret)
 960                 return ret;
 961
 962         mmu_seq = vcpu->kvm->mmu_notifier_seq;
 963         /*
 964          * Ensure the read of mmu_notifier_seq happens before we call
 965          * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
 966          * the page we just got a reference to gets unmapped before we have a
 967          * chance to grab the mmu_lock, which ensure that if the page gets
 968          * unmapped afterwards, the call to kvm_unmap_hva will take it away
 969          * from us again properly. This smp_rmb() interacts with the smp_wmb()
 970          * in kvm_mmu_notifier_invalidate_<page|range_end>.
 971          */
 972         smp_rmb();
 973
 974         pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
 975         if (is_error_pfn(pfn))
 976                 return -EFAULT;
 977
 978         if (kvm_is_device_pfn(pfn))
 979                 mem_type = PAGE_S2_DEVICE;
 980
 981         spin_lock(&kvm->mmu_lock);
 982         if (mmu_notifier_retry(kvm, mmu_seq))
 983                 goto out_unlock;
 984         if (!hugetlb && !force_pte)
 985                 hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
 986
 987         fault_ipa_uncached = memslot->flags & KVM_MEMSLOT_INCOHERENT;
 988
 989         if (hugetlb) {
 990                 pmd_t new_pmd = pfn_pmd(pfn, mem_type);
 991                 new_pmd = pmd_mkhuge(new_pmd);
 992                 if (writable) {
 993                         kvm_set_s2pmd_writable(&new_pmd);
 994                         kvm_set_pfn_dirty(pfn);
 995                 }
 996                 coherent_cache_guest_page(vcpu, hva & PMD_MASK, PMD_SIZE,
 997                                           fault_ipa_uncached);
 998                 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
 999         } else {
1000                 pte_t new_pte = pfn_pte(pfn, mem_type);
1001                 if (writable) {
1002                         kvm_set_s2pte_writable(&new_pte);
1003                         kvm_set_pfn_dirty(pfn);
1004                 }
1005                 coherent_cache_guest_page(vcpu, hva, PAGE_SIZE,
1006                                           fault_ipa_uncached);
1007                 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte,
1008                         pgprot_val(mem_type) == pgprot_val(PAGE_S2_DEVICE));
1009         }
1010
1011
1012 out_unlock:
1013         spin_unlock(&kvm->mmu_lock);
1014         kvm_release_pfn_clean(pfn);
1015         return ret;
1016 }
1017
1018 /**
1019  * kvm_handle_guest_abort - handles all 2nd stage aborts
1020  * @vcpu:       the VCPU pointer
1021  * @run:        the kvm_run structure
1022  *
1023  * Any abort that gets to the host is almost guaranteed to be caused by a
1024  * missing second stage translation table entry, which can mean that either the
1025  * guest simply needs more memory and we must allocate an appropriate page or it
1026  * can mean that the guest tried to access I/O memory, which is emulated by user
1027  * space. The distinction is based on the IPA causing the fault and whether this
1028  * memory region has been registered as standard RAM by user space.
1029  */
1030 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
1031 {
1032         unsigned long fault_status;
1033         phys_addr_t fault_ipa;
1034         struct kvm_memory_slot *memslot;
1035         unsigned long hva;
1036         bool is_iabt, write_fault, writable;
1037         gfn_t gfn;
1038         int ret, idx;
1039
1040         is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1041         fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1042
1043         trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
1044                               kvm_vcpu_get_hfar(vcpu), fault_ipa);
1045
1046         /* Check the stage-2 fault is trans. fault or write fault */
1047         fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1048         if (fault_status != FSC_FAULT && fault_status != FSC_PERM) {
1049                 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1050                         kvm_vcpu_trap_get_class(vcpu),
1051                         (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1052                         (unsigned long)kvm_vcpu_get_hsr(vcpu));
1053                 return -EFAULT;
1054         }
1055
1056         idx = srcu_read_lock(&vcpu->kvm->srcu);
1057
1058         gfn = fault_ipa >> PAGE_SHIFT;
1059         memslot = gfn_to_memslot(vcpu->kvm, gfn);
1060         hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1061         write_fault = kvm_is_write_fault(vcpu);
1062         if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1063                 if (is_iabt) {
1064                         /* Prefetch Abort on I/O address */
1065                         kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1066                         ret = 1;
1067                         goto out_unlock;
1068                 }
1069
1070                 /*
1071                  * The IPA is reported as [MAX:12], so we need to
1072                  * complement it with the bottom 12 bits from the
1073                  * faulting VA. This is always 12 bits, irrespective
1074                  * of the page size.
1075                  */
1076                 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1077                 ret = io_mem_abort(vcpu, run, fault_ipa);
1078                 goto out_unlock;
1079         }
1080
1081         /* Userspace should not be able to register out-of-bounds IPAs */
1082         VM_BUG_ON(fault_ipa >= KVM_PHYS_SIZE);
1083
1084         ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1085         if (ret == 0)
1086                 ret = 1;
1087 out_unlock:
1088         srcu_read_unlock(&vcpu->kvm->srcu, idx);
1089         return ret;
1090 }
1091
1092 static void handle_hva_to_gpa(struct kvm *kvm,
1093                               unsigned long start,
1094                               unsigned long end,
1095                               void (*handler)(struct kvm *kvm,
1096                                               gpa_t gpa, void *data),
1097                               void *data)
1098 {
1099         struct kvm_memslots *slots;
1100         struct kvm_memory_slot *memslot;
1101
1102         slots = kvm_memslots(kvm);
1103
1104         /* we only care about the pages that the guest sees */
1105         kvm_for_each_memslot(memslot, slots) {
1106                 unsigned long hva_start, hva_end;
1107                 gfn_t gfn, gfn_end;
1108
1109                 hva_start = max(start, memslot->userspace_addr);
1110                 hva_end = min(end, memslot->userspace_addr +
1111                                         (memslot->npages << PAGE_SHIFT));
1112                 if (hva_start >= hva_end)
1113                         continue;
1114
1115                 /*
1116                  * {gfn(page) | page intersects with [hva_start, hva_end)} =
1117                  * {gfn_start, gfn_start+1, ..., gfn_end-1}.
1118                  */
1119                 gfn = hva_to_gfn_memslot(hva_start, memslot);
1120                 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
1121
1122                 for (; gfn < gfn_end; ++gfn) {
1123                         gpa_t gpa = gfn << PAGE_SHIFT;
1124                         handler(kvm, gpa, data);
1125                 }
1126         }
1127 }
1128
1129 static void kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
1130 {
1131         unmap_stage2_range(kvm, gpa, PAGE_SIZE);
1132 }
1133
1134 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1135 {
1136         unsigned long end = hva + PAGE_SIZE;
1137
1138         if (!kvm->arch.pgd)
1139                 return 0;
1140
1141         trace_kvm_unmap_hva(hva);
1142         handle_hva_to_gpa(kvm, hva, end, &kvm_unmap_hva_handler, NULL);
1143         return 0;
1144 }
1145
1146 int kvm_unmap_hva_range(struct kvm *kvm,
1147                         unsigned long start, unsigned long end)
1148 {
1149         if (!kvm->arch.pgd)
1150                 return 0;
1151
1152         trace_kvm_unmap_hva_range(start, end);
1153         handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
1154         return 0;
1155 }
1156
1157 static void kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, void *data)
1158 {
1159         pte_t *pte = (pte_t *)data;
1160
1161         stage2_set_pte(kvm, NULL, gpa, pte, false);
1162 }
1163
1164
1165 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1166 {
1167         unsigned long end = hva + PAGE_SIZE;
1168         pte_t stage2_pte;
1169
1170         if (!kvm->arch.pgd)
1171                 return;
1172
1173         trace_kvm_set_spte_hva(hva);
1174         stage2_pte = pfn_pte(pte_pfn(pte), PAGE_S2);
1175         handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
1176 }
1177
1178 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1179 {
1180         mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
1181 }
1182
1183 phys_addr_t kvm_mmu_get_httbr(void)
1184 {
1185         return virt_to_phys(hyp_pgd);
1186 }
1187
1188 phys_addr_t kvm_mmu_get_boot_httbr(void)
1189 {
1190         return virt_to_phys(boot_hyp_pgd);
1191 }
1192
1193 phys_addr_t kvm_get_idmap_vector(void)
1194 {
1195         return hyp_idmap_vector;
1196 }
1197
1198 int kvm_mmu_init(void)
1199 {
1200         int err;
1201
1202         hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
1203         hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
1204         hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
1205
1206         if ((hyp_idmap_start ^ hyp_idmap_end) & PAGE_MASK) {
1207                 /*
1208                  * Our init code is crossing a page boundary. Allocate
1209                  * a bounce page, copy the code over and use that.
1210                  */
1211                 size_t len = __hyp_idmap_text_end - __hyp_idmap_text_start;
1212                 phys_addr_t phys_base;
1213
1214                 init_bounce_page = (void *)__get_free_page(GFP_KERNEL);
1215                 if (!init_bounce_page) {
1216                         kvm_err("Couldn't allocate HYP init bounce page\n");
1217                         err = -ENOMEM;
1218                         goto out;
1219                 }
1220
1221                 memcpy(init_bounce_page, __hyp_idmap_text_start, len);
1222                 /*
1223                  * Warning: the code we just copied to the bounce page
1224                  * must be flushed to the point of coherency.
1225                  * Otherwise, the data may be sitting in L2, and HYP
1226                  * mode won't be able to observe it as it runs with
1227                  * caches off at that point.
1228                  */
1229                 kvm_flush_dcache_to_poc(init_bounce_page, len);
1230
1231                 phys_base = kvm_virt_to_phys(init_bounce_page);
1232                 hyp_idmap_vector += phys_base - hyp_idmap_start;
1233                 hyp_idmap_start = phys_base;
1234                 hyp_idmap_end = phys_base + len;
1235
1236                 kvm_info("Using HYP init bounce page @%lx\n",
1237                          (unsigned long)phys_base);
1238         }
1239
1240         hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1241         boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1242
1243         if (!hyp_pgd || !boot_hyp_pgd) {
1244                 kvm_err("Hyp mode PGD not allocated\n");
1245                 err = -ENOMEM;
1246                 goto out;
1247         }
1248
1249         /* Create the idmap in the boot page tables */
1250         err =   __create_hyp_mappings(boot_hyp_pgd,
1251                                       hyp_idmap_start, hyp_idmap_end,
1252                                       __phys_to_pfn(hyp_idmap_start),
1253                                       PAGE_HYP);
1254
1255         if (err) {
1256                 kvm_err("Failed to idmap %lx-%lx\n",
1257                         hyp_idmap_start, hyp_idmap_end);
1258                 goto out;
1259         }
1260
1261         /* Map the very same page at the trampoline VA */
1262         err =   __create_hyp_mappings(boot_hyp_pgd,
1263                                       TRAMPOLINE_VA, TRAMPOLINE_VA + PAGE_SIZE,
1264                                       __phys_to_pfn(hyp_idmap_start),
1265                                       PAGE_HYP);
1266         if (err) {
1267                 kvm_err("Failed to map trampoline @%lx into boot HYP pgd\n",
1268                         TRAMPOLINE_VA);
1269                 goto out;
1270         }
1271
1272         /* Map the same page again into the runtime page tables */
1273         err =   __create_hyp_mappings(hyp_pgd,
1274                                       TRAMPOLINE_VA, TRAMPOLINE_VA + PAGE_SIZE,
1275                                       __phys_to_pfn(hyp_idmap_start),
1276                                       PAGE_HYP);
1277         if (err) {
1278                 kvm_err("Failed to map trampoline @%lx into runtime HYP pgd\n",
1279                         TRAMPOLINE_VA);
1280                 goto out;
1281         }
1282
1283         return 0;
1284 out:
1285         free_hyp_pgds();
1286         return err;
1287 }
1288
1289 void kvm_arch_commit_memory_region(struct kvm *kvm,
1290                                    struct kvm_userspace_memory_region *mem,
1291                                    const struct kvm_memory_slot *old,
1292                                    enum kvm_mr_change change)
1293 {
1294 }
1295
1296 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1297                                    struct kvm_memory_slot *memslot,
1298                                    struct kvm_userspace_memory_region *mem,
1299                                    enum kvm_mr_change change)
1300 {
1301         hva_t hva = mem->userspace_addr;
1302         hva_t reg_end = hva + mem->memory_size;
1303         bool writable = !(mem->flags & KVM_MEM_READONLY);
1304         int ret = 0;
1305
1306         if (change != KVM_MR_CREATE && change != KVM_MR_MOVE)
1307                 return 0;
1308
1309         /*
1310          * Prevent userspace from creating a memory region outside of the IPA
1311          * space addressable by the KVM guest IPA space.
1312          */
1313         if (memslot->base_gfn + memslot->npages >=
1314             (KVM_PHYS_SIZE >> PAGE_SHIFT))
1315                 return -EFAULT;
1316
1317         /*
1318          * A memory region could potentially cover multiple VMAs, and any holes
1319          * between them, so iterate over all of them to find out if we can map
1320          * any of them right now.
1321          *
1322          *     +--------------------------------------------+
1323          * +---------------+----------------+   +----------------+
1324          * |   : VMA 1     |      VMA 2     |   |    VMA 3  :    |
1325          * +---------------+----------------+   +----------------+
1326          *     |               memory region                |
1327          *     +--------------------------------------------+
1328          */
1329         do {
1330                 struct vm_area_struct *vma = find_vma(current->mm, hva);
1331                 hva_t vm_start, vm_end;
1332
1333                 if (!vma || vma->vm_start >= reg_end)
1334                         break;
1335
1336                 /*
1337                  * Mapping a read-only VMA is only allowed if the
1338                  * memory region is configured as read-only.
1339                  */
1340                 if (writable && !(vma->vm_flags & VM_WRITE)) {
1341                         ret = -EPERM;
1342                         break;
1343                 }
1344
1345                 /*
1346                  * Take the intersection of this VMA with the memory region
1347                  */
1348                 vm_start = max(hva, vma->vm_start);
1349                 vm_end = min(reg_end, vma->vm_end);
1350
1351                 if (vma->vm_flags & VM_PFNMAP) {
1352                         gpa_t gpa = mem->guest_phys_addr +
1353                                     (vm_start - mem->userspace_addr);
1354                         phys_addr_t pa = (vma->vm_pgoff << PAGE_SHIFT) +
1355                                          vm_start - vma->vm_start;
1356
1357                         ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
1358                                                     vm_end - vm_start,
1359                                                     writable);
1360                         if (ret)
1361                                 break;
1362                 }
1363                 hva = vm_end;
1364         } while (hva < reg_end);
1365
1366         spin_lock(&kvm->mmu_lock);
1367         if (ret)
1368                 unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
1369         else
1370                 stage2_flush_memslot(kvm, memslot);
1371         spin_unlock(&kvm->mmu_lock);
1372         return ret;
1373 }
1374
1375 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
1376                            struct kvm_memory_slot *dont)
1377 {
1378 }
1379
1380 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
1381                             unsigned long npages)
1382 {
1383         /*
1384          * Readonly memslots are not incoherent with the caches by definition,
1385          * but in practice, they are used mostly to emulate ROMs or NOR flashes
1386          * that the guest may consider devices and hence map as uncached.
1387          * To prevent incoherency issues in these cases, tag all readonly
1388          * regions as incoherent.
1389          */
1390         if (slot->flags & KVM_MEM_READONLY)
1391                 slot->flags |= KVM_MEMSLOT_INCOHERENT;
1392         return 0;
1393 }
1394
1395 void kvm_arch_memslots_updated(struct kvm *kvm)
1396 {
1397 }
1398
1399 void kvm_arch_flush_shadow_all(struct kvm *kvm)
1400 {
1401 }
1402
1403 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
1404                                    struct kvm_memory_slot *slot)
1405 {
1406         gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
1407         phys_addr_t size = slot->npages << PAGE_SHIFT;
1408
1409         spin_lock(&kvm->mmu_lock);
1410         unmap_stage2_range(kvm, gpa, size);
1411         spin_unlock(&kvm->mmu_lock);
1412 }