virt/kvm/arm/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 <linux/sched/signal.h>
  24 #include <trace/events/kvm.h>
  25 #include <asm/pgalloc.h>
  26 #include <asm/cacheflush.h>
  27 #include <asm/kvm_arm.h>
  28 #include <asm/kvm_mmu.h>
  29 #include <asm/kvm_mmio.h>
  30 #include <asm/kvm_asm.h>
  31 #include <asm/kvm_emulate.h>
  32 #include <asm/virt.h>
  33 #include <asm/system_misc.h>
  34
  35 #include "trace.h"
  36
  37 static pgd_t *boot_hyp_pgd;
  38 static pgd_t *hyp_pgd;
  39 static pgd_t *merged_hyp_pgd;
  40 static DEFINE_MUTEX(kvm_hyp_pgd_mutex);
  41
  42 static unsigned long hyp_idmap_start;
  43 static unsigned long hyp_idmap_end;
  44 static phys_addr_t hyp_idmap_vector;
  45
  46 #define S2_PGD_SIZE     (PTRS_PER_S2_PGD * sizeof(pgd_t))
  47 #define hyp_pgd_order get_order(PTRS_PER_PGD * sizeof(pgd_t))
  48
  49 #define KVM_S2PTE_FLAG_IS_IOMAP         (1UL << 0)
  50 #define KVM_S2_FLAG_LOGGING_ACTIVE      (1UL << 1)
  51
  52 static bool memslot_is_logging(struct kvm_memory_slot *memslot)
  53 {
  54         return memslot->dirty_bitmap && !(memslot->flags & KVM_MEM_READONLY);
  55 }
  56
  57 /**
  58  * kvm_flush_remote_tlbs() - flush all VM TLB entries for v7/8
  59  * @kvm:        pointer to kvm structure.
  60  *
  61  * Interface to HYP function to flush all VM TLB entries
  62  */
  63 void kvm_flush_remote_tlbs(struct kvm *kvm)
  64 {
  65         kvm_call_hyp(__kvm_tlb_flush_vmid, kvm);
  66 }
  67
  68 static void kvm_tlb_flush_vmid_ipa(struct kvm *kvm, phys_addr_t ipa)
  69 {
  70         kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, kvm, ipa);
  71 }
  72
  73 /*
  74  * D-Cache management functions. They take the page table entries by
  75  * value, as they are flushing the cache using the kernel mapping (or
  76  * kmap on 32bit).
  77  */
  78 static void kvm_flush_dcache_pte(pte_t pte)
  79 {
  80         __kvm_flush_dcache_pte(pte);
  81 }
  82
  83 static void kvm_flush_dcache_pmd(pmd_t pmd)
  84 {
  85         __kvm_flush_dcache_pmd(pmd);
  86 }
  87
  88 static void kvm_flush_dcache_pud(pud_t pud)
  89 {
  90         __kvm_flush_dcache_pud(pud);
  91 }
  92
  93 static bool kvm_is_device_pfn(unsigned long pfn)
  94 {
  95         return !pfn_valid(pfn);
  96 }
  97
  98 /**
  99  * stage2_dissolve_pmd() - clear and flush huge PMD entry
 100  * @kvm:        pointer to kvm structure.
 101  * @addr:       IPA
 102  * @pmd:        pmd pointer for IPA
 103  *
 104  * Function clears a PMD entry, flushes addr 1st and 2nd stage TLBs. Marks all
 105  * pages in the range dirty.
 106  */
 107 static void stage2_dissolve_pmd(struct kvm *kvm, phys_addr_t addr, pmd_t *pmd)
 108 {
 109         if (!pmd_thp_or_huge(*pmd))
 110                 return;
 111
 112         pmd_clear(pmd);
 113         kvm_tlb_flush_vmid_ipa(kvm, addr);
 114         put_page(virt_to_page(pmd));
 115 }
 116
 117 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 118                                   int min, int max)
 119 {
 120         void *page;
 121
 122         BUG_ON(max > KVM_NR_MEM_OBJS);
 123         if (cache->nobjs >= min)
 124                 return 0;
 125         while (cache->nobjs < max) {
 126                 page = (void *)__get_free_page(PGALLOC_GFP);
 127                 if (!page)
 128                         return -ENOMEM;
 129                 cache->objects[cache->nobjs++] = page;
 130         }
 131         return 0;
 132 }
 133
 134 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 135 {
 136         while (mc->nobjs)
 137                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 138 }
 139
 140 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
 141 {
 142         void *p;
 143
 144         BUG_ON(!mc || !mc->nobjs);
 145         p = mc->objects[--mc->nobjs];
 146         return p;
 147 }
 148
 149 static void clear_stage2_pgd_entry(struct kvm *kvm, pgd_t *pgd, phys_addr_t addr)
 150 {
 151         pud_t *pud_table __maybe_unused = stage2_pud_offset(pgd, 0UL);
 152         stage2_pgd_clear(pgd);
 153         kvm_tlb_flush_vmid_ipa(kvm, addr);
 154         stage2_pud_free(pud_table);
 155         put_page(virt_to_page(pgd));
 156 }
 157
 158 static void clear_stage2_pud_entry(struct kvm *kvm, pud_t *pud, phys_addr_t addr)
 159 {
 160         pmd_t *pmd_table __maybe_unused = stage2_pmd_offset(pud, 0);
 161         VM_BUG_ON(stage2_pud_huge(*pud));
 162         stage2_pud_clear(pud);
 163         kvm_tlb_flush_vmid_ipa(kvm, addr);
 164         stage2_pmd_free(pmd_table);
 165         put_page(virt_to_page(pud));
 166 }
 167
 168 static void clear_stage2_pmd_entry(struct kvm *kvm, pmd_t *pmd, phys_addr_t addr)
 169 {
 170         pte_t *pte_table = pte_offset_kernel(pmd, 0);
 171         VM_BUG_ON(pmd_thp_or_huge(*pmd));
 172         pmd_clear(pmd);
 173         kvm_tlb_flush_vmid_ipa(kvm, addr);
 174         pte_free_kernel(NULL, pte_table);
 175         put_page(virt_to_page(pmd));
 176 }
 177
 178 /*
 179  * Unmapping vs dcache management:
 180  *
 181  * If a guest maps certain memory pages as uncached, all writes will
 182  * bypass the data cache and go directly to RAM.  However, the CPUs
 183  * can still speculate reads (not writes) and fill cache lines with
 184  * data.
 185  *
 186  * Those cache lines will be *clean* cache lines though, so a
 187  * clean+invalidate operation is equivalent to an invalidate
 188  * operation, because no cache lines are marked dirty.
 189  *
 190  * Those clean cache lines could be filled prior to an uncached write
 191  * by the guest, and the cache coherent IO subsystem would therefore
 192  * end up writing old data to disk.
 193  *
 194  * This is why right after unmapping a page/section and invalidating
 195  * the corresponding TLBs, we call kvm_flush_dcache_p*() to make sure
 196  * the IO subsystem will never hit in the cache.
 197  */
 198 static void unmap_stage2_ptes(struct kvm *kvm, pmd_t *pmd,
 199                        phys_addr_t addr, phys_addr_t end)
 200 {
 201         phys_addr_t start_addr = addr;
 202         pte_t *pte, *start_pte;
 203
 204         start_pte = pte = pte_offset_kernel(pmd, addr);
 205         do {
 206                 if (!pte_none(*pte)) {
 207                         pte_t old_pte = *pte;
 208
 209                         kvm_set_pte(pte, __pte(0));
 210                         kvm_tlb_flush_vmid_ipa(kvm, addr);
 211
 212                         /* No need to invalidate the cache for device mappings */
 213                         if (!kvm_is_device_pfn(pte_pfn(old_pte)))
 214                                 kvm_flush_dcache_pte(old_pte);
 215
 216                         put_page(virt_to_page(pte));
 217                 }
 218         } while (pte++, addr += PAGE_SIZE, addr != end);
 219
 220         if (stage2_pte_table_empty(start_pte))
 221                 clear_stage2_pmd_entry(kvm, pmd, start_addr);
 222 }
 223
 224 static void unmap_stage2_pmds(struct kvm *kvm, pud_t *pud,
 225                        phys_addr_t addr, phys_addr_t end)
 226 {
 227         phys_addr_t next, start_addr = addr;
 228         pmd_t *pmd, *start_pmd;
 229
 230         start_pmd = pmd = stage2_pmd_offset(pud, addr);
 231         do {
 232                 next = stage2_pmd_addr_end(addr, end);
 233                 if (!pmd_none(*pmd)) {
 234                         if (pmd_thp_or_huge(*pmd)) {
 235                                 pmd_t old_pmd = *pmd;
 236
 237                                 pmd_clear(pmd);
 238                                 kvm_tlb_flush_vmid_ipa(kvm, addr);
 239
 240                                 kvm_flush_dcache_pmd(old_pmd);
 241
 242                                 put_page(virt_to_page(pmd));
 243                         } else {
 244                                 unmap_stage2_ptes(kvm, pmd, addr, next);
 245                         }
 246                 }
 247         } while (pmd++, addr = next, addr != end);
 248
 249         if (stage2_pmd_table_empty(start_pmd))
 250                 clear_stage2_pud_entry(kvm, pud, start_addr);
 251 }
 252
 253 static void unmap_stage2_puds(struct kvm *kvm, pgd_t *pgd,
 254                        phys_addr_t addr, phys_addr_t end)
 255 {
 256         phys_addr_t next, start_addr = addr;
 257         pud_t *pud, *start_pud;
 258
 259         start_pud = pud = stage2_pud_offset(pgd, addr);
 260         do {
 261                 next = stage2_pud_addr_end(addr, end);
 262                 if (!stage2_pud_none(*pud)) {
 263                         if (stage2_pud_huge(*pud)) {
 264                                 pud_t old_pud = *pud;
 265
 266                                 stage2_pud_clear(pud);
 267                                 kvm_tlb_flush_vmid_ipa(kvm, addr);
 268                                 kvm_flush_dcache_pud(old_pud);
 269                                 put_page(virt_to_page(pud));
 270                         } else {
 271                                 unmap_stage2_pmds(kvm, pud, addr, next);
 272                         }
 273                 }
 274         } while (pud++, addr = next, addr != end);
 275
 276         if (stage2_pud_table_empty(start_pud))
 277                 clear_stage2_pgd_entry(kvm, pgd, start_addr);
 278 }
 279
 280 /**
 281  * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
 282  * @kvm:   The VM pointer
 283  * @start: The intermediate physical base address of the range to unmap
 284  * @size:  The size of the area to unmap
 285  *
 286  * Clear a range of stage-2 mappings, lowering the various ref-counts.  Must
 287  * be called while holding mmu_lock (unless for freeing the stage2 pgd before
 288  * destroying the VM), otherwise another faulting VCPU may come in and mess
 289  * with things behind our backs.
 290  */
 291 static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
 292 {
 293         pgd_t *pgd;
 294         phys_addr_t addr = start, end = start + size;
 295         phys_addr_t next;
 296
 297         assert_spin_locked(&kvm->mmu_lock);
 298         pgd = kvm->arch.pgd + stage2_pgd_index(addr);
 299         do {
 300                 /*
 301                  * Make sure the page table is still active, as another thread
 302                  * could have possibly freed the page table, while we released
 303                  * the lock.
 304                  */
 305                 if (!READ_ONCE(kvm->arch.pgd))
 306                         break;
 307                 next = stage2_pgd_addr_end(addr, end);
 308                 if (!stage2_pgd_none(*pgd))
 309                         unmap_stage2_puds(kvm, pgd, addr, next);
 310                 /*
 311                  * If the range is too large, release the kvm->mmu_lock
 312                  * to prevent starvation and lockup detector warnings.
 313                  */
 314                 if (next != end)
 315                         cond_resched_lock(&kvm->mmu_lock);
 316         } while (pgd++, addr = next, addr != end);
 317 }
 318
 319 static void stage2_flush_ptes(struct kvm *kvm, pmd_t *pmd,
 320                               phys_addr_t addr, phys_addr_t end)
 321 {
 322         pte_t *pte;
 323
 324         pte = pte_offset_kernel(pmd, addr);
 325         do {
 326                 if (!pte_none(*pte) && !kvm_is_device_pfn(pte_pfn(*pte)))
 327                         kvm_flush_dcache_pte(*pte);
 328         } while (pte++, addr += PAGE_SIZE, addr != end);
 329 }
 330
 331 static void stage2_flush_pmds(struct kvm *kvm, pud_t *pud,
 332                               phys_addr_t addr, phys_addr_t end)
 333 {
 334         pmd_t *pmd;
 335         phys_addr_t next;
 336
 337         pmd = stage2_pmd_offset(pud, addr);
 338         do {
 339                 next = stage2_pmd_addr_end(addr, end);
 340                 if (!pmd_none(*pmd)) {
 341                         if (pmd_thp_or_huge(*pmd))
 342                                 kvm_flush_dcache_pmd(*pmd);
 343                         else
 344                                 stage2_flush_ptes(kvm, pmd, addr, next);
 345                 }
 346         } while (pmd++, addr = next, addr != end);
 347 }
 348
 349 static void stage2_flush_puds(struct kvm *kvm, pgd_t *pgd,
 350                               phys_addr_t addr, phys_addr_t end)
 351 {
 352         pud_t *pud;
 353         phys_addr_t next;
 354
 355         pud = stage2_pud_offset(pgd, addr);
 356         do {
 357                 next = stage2_pud_addr_end(addr, end);
 358                 if (!stage2_pud_none(*pud)) {
 359                         if (stage2_pud_huge(*pud))
 360                                 kvm_flush_dcache_pud(*pud);
 361                         else
 362                                 stage2_flush_pmds(kvm, pud, addr, next);
 363                 }
 364         } while (pud++, addr = next, addr != end);
 365 }
 366
 367 static void stage2_flush_memslot(struct kvm *kvm,
 368                                  struct kvm_memory_slot *memslot)
 369 {
 370         phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
 371         phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
 372         phys_addr_t next;
 373         pgd_t *pgd;
 374
 375         pgd = kvm->arch.pgd + stage2_pgd_index(addr);
 376         do {
 377                 next = stage2_pgd_addr_end(addr, end);
 378                 stage2_flush_puds(kvm, pgd, addr, next);
 379         } while (pgd++, addr = next, addr != end);
 380 }
 381
 382 /**
 383  * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
 384  * @kvm: The struct kvm pointer
 385  *
 386  * Go through the stage 2 page tables and invalidate any cache lines
 387  * backing memory already mapped to the VM.
 388  */
 389 static void stage2_flush_vm(struct kvm *kvm)
 390 {
 391         struct kvm_memslots *slots;
 392         struct kvm_memory_slot *memslot;
 393         int idx;
 394
 395         idx = srcu_read_lock(&kvm->srcu);
 396         spin_lock(&kvm->mmu_lock);
 397
 398         slots = kvm_memslots(kvm);
 399         kvm_for_each_memslot(memslot, slots)
 400                 stage2_flush_memslot(kvm, memslot);
 401
 402         spin_unlock(&kvm->mmu_lock);
 403         srcu_read_unlock(&kvm->srcu, idx);
 404 }
 405
 406 static void clear_hyp_pgd_entry(pgd_t *pgd)
 407 {
 408         pud_t *pud_table __maybe_unused = pud_offset(pgd, 0UL);
 409         pgd_clear(pgd);
 410         pud_free(NULL, pud_table);
 411         put_page(virt_to_page(pgd));
 412 }
 413
 414 static void clear_hyp_pud_entry(pud_t *pud)
 415 {
 416         pmd_t *pmd_table __maybe_unused = pmd_offset(pud, 0);
 417         VM_BUG_ON(pud_huge(*pud));
 418         pud_clear(pud);
 419         pmd_free(NULL, pmd_table);
 420         put_page(virt_to_page(pud));
 421 }
 422
 423 static void clear_hyp_pmd_entry(pmd_t *pmd)
 424 {
 425         pte_t *pte_table = pte_offset_kernel(pmd, 0);
 426         VM_BUG_ON(pmd_thp_or_huge(*pmd));
 427         pmd_clear(pmd);
 428         pte_free_kernel(NULL, pte_table);
 429         put_page(virt_to_page(pmd));
 430 }
 431
 432 static void unmap_hyp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
 433 {
 434         pte_t *pte, *start_pte;
 435
 436         start_pte = pte = pte_offset_kernel(pmd, addr);
 437         do {
 438                 if (!pte_none(*pte)) {
 439                         kvm_set_pte(pte, __pte(0));
 440                         put_page(virt_to_page(pte));
 441                 }
 442         } while (pte++, addr += PAGE_SIZE, addr != end);
 443
 444         if (hyp_pte_table_empty(start_pte))
 445                 clear_hyp_pmd_entry(pmd);
 446 }
 447
 448 static void unmap_hyp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
 449 {
 450         phys_addr_t next;
 451         pmd_t *pmd, *start_pmd;
 452
 453         start_pmd = pmd = pmd_offset(pud, addr);
 454         do {
 455                 next = pmd_addr_end(addr, end);
 456                 /* Hyp doesn't use huge pmds */
 457                 if (!pmd_none(*pmd))
 458                         unmap_hyp_ptes(pmd, addr, next);
 459         } while (pmd++, addr = next, addr != end);
 460
 461         if (hyp_pmd_table_empty(start_pmd))
 462                 clear_hyp_pud_entry(pud);
 463 }
 464
 465 static void unmap_hyp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
 466 {
 467         phys_addr_t next;
 468         pud_t *pud, *start_pud;
 469
 470         start_pud = pud = pud_offset(pgd, addr);
 471         do {
 472                 next = pud_addr_end(addr, end);
 473                 /* Hyp doesn't use huge puds */
 474                 if (!pud_none(*pud))
 475                         unmap_hyp_pmds(pud, addr, next);
 476         } while (pud++, addr = next, addr != end);
 477
 478         if (hyp_pud_table_empty(start_pud))
 479                 clear_hyp_pgd_entry(pgd);
 480 }
 481
 482 static void unmap_hyp_range(pgd_t *pgdp, phys_addr_t start, u64 size)
 483 {
 484         pgd_t *pgd;
 485         phys_addr_t addr = start, end = start + size;
 486         phys_addr_t next;
 487
 488         /*
 489          * We don't unmap anything from HYP, except at the hyp tear down.
 490          * Hence, we don't have to invalidate the TLBs here.
 491          */
 492         pgd = pgdp + pgd_index(addr);
 493         do {
 494                 next = pgd_addr_end(addr, end);
 495                 if (!pgd_none(*pgd))
 496                         unmap_hyp_puds(pgd, addr, next);
 497         } while (pgd++, addr = next, addr != end);
 498 }
 499
 500 /**
 501  * free_hyp_pgds - free Hyp-mode page tables
 502  *
 503  * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
 504  * therefore contains either mappings in the kernel memory area (above
 505  * PAGE_OFFSET), or device mappings in the vmalloc range (from
 506  * VMALLOC_START to VMALLOC_END).
 507  *
 508  * boot_hyp_pgd should only map two pages for the init code.
 509  */
 510 void free_hyp_pgds(void)
 511 {
 512         unsigned long addr;
 513
 514         mutex_lock(&kvm_hyp_pgd_mutex);
 515
 516         if (boot_hyp_pgd) {
 517                 unmap_hyp_range(boot_hyp_pgd, hyp_idmap_start, PAGE_SIZE);
 518                 free_pages((unsigned long)boot_hyp_pgd, hyp_pgd_order);
 519                 boot_hyp_pgd = NULL;
 520         }
 521
 522         if (hyp_pgd) {
 523                 unmap_hyp_range(hyp_pgd, hyp_idmap_start, PAGE_SIZE);
 524                 for (addr = PAGE_OFFSET; virt_addr_valid(addr); addr += PGDIR_SIZE)
 525                         unmap_hyp_range(hyp_pgd, kern_hyp_va(addr), PGDIR_SIZE);
 526                 for (addr = VMALLOC_START; is_vmalloc_addr((void*)addr); addr += PGDIR_SIZE)
 527                         unmap_hyp_range(hyp_pgd, kern_hyp_va(addr), PGDIR_SIZE);
 528
 529                 free_pages((unsigned long)hyp_pgd, hyp_pgd_order);
 530                 hyp_pgd = NULL;
 531         }
 532         if (merged_hyp_pgd) {
 533                 clear_page(merged_hyp_pgd);
 534                 free_page((unsigned long)merged_hyp_pgd);
 535                 merged_hyp_pgd = NULL;
 536         }
 537
 538         mutex_unlock(&kvm_hyp_pgd_mutex);
 539 }
 540
 541 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
 542                                     unsigned long end, unsigned long pfn,
 543                                     pgprot_t prot)
 544 {
 545         pte_t *pte;
 546         unsigned long addr;
 547
 548         addr = start;
 549         do {
 550                 pte = pte_offset_kernel(pmd, addr);
 551                 kvm_set_pte(pte, pfn_pte(pfn, prot));
 552                 get_page(virt_to_page(pte));
 553                 kvm_flush_dcache_to_poc(pte, sizeof(*pte));
 554                 pfn++;
 555         } while (addr += PAGE_SIZE, addr != end);
 556 }
 557
 558 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
 559                                    unsigned long end, unsigned long pfn,
 560                                    pgprot_t prot)
 561 {
 562         pmd_t *pmd;
 563         pte_t *pte;
 564         unsigned long addr, next;
 565
 566         addr = start;
 567         do {
 568                 pmd = pmd_offset(pud, addr);
 569
 570                 BUG_ON(pmd_sect(*pmd));
 571
 572                 if (pmd_none(*pmd)) {
 573                         pte = pte_alloc_one_kernel(NULL, addr);
 574                         if (!pte) {
 575                                 kvm_err("Cannot allocate Hyp pte\n");
 576                                 return -ENOMEM;
 577                         }
 578                         pmd_populate_kernel(NULL, pmd, pte);
 579                         get_page(virt_to_page(pmd));
 580                         kvm_flush_dcache_to_poc(pmd, sizeof(*pmd));
 581                 }
 582
 583                 next = pmd_addr_end(addr, end);
 584
 585                 create_hyp_pte_mappings(pmd, addr, next, pfn, prot);
 586                 pfn += (next - addr) >> PAGE_SHIFT;
 587         } while (addr = next, addr != end);
 588
 589         return 0;
 590 }
 591
 592 static int create_hyp_pud_mappings(pgd_t *pgd, unsigned long start,
 593                                    unsigned long end, unsigned long pfn,
 594                                    pgprot_t prot)
 595 {
 596         pud_t *pud;
 597         pmd_t *pmd;
 598         unsigned long addr, next;
 599         int ret;
 600
 601         addr = start;
 602         do {
 603                 pud = pud_offset(pgd, addr);
 604
 605                 if (pud_none_or_clear_bad(pud)) {
 606                         pmd = pmd_alloc_one(NULL, addr);
 607                         if (!pmd) {
 608                                 kvm_err("Cannot allocate Hyp pmd\n");
 609                                 return -ENOMEM;
 610                         }
 611                         pud_populate(NULL, pud, pmd);
 612                         get_page(virt_to_page(pud));
 613                         kvm_flush_dcache_to_poc(pud, sizeof(*pud));
 614                 }
 615
 616                 next = pud_addr_end(addr, end);
 617                 ret = create_hyp_pmd_mappings(pud, addr, next, pfn, prot);
 618                 if (ret)
 619                         return ret;
 620                 pfn += (next - addr) >> PAGE_SHIFT;
 621         } while (addr = next, addr != end);
 622
 623         return 0;
 624 }
 625
 626 static int __create_hyp_mappings(pgd_t *pgdp,
 627                                  unsigned long start, unsigned long end,
 628                                  unsigned long pfn, pgprot_t prot)
 629 {
 630         pgd_t *pgd;
 631         pud_t *pud;
 632         unsigned long addr, next;
 633         int err = 0;
 634
 635         mutex_lock(&kvm_hyp_pgd_mutex);
 636         addr = start & PAGE_MASK;
 637         end = PAGE_ALIGN(end);
 638         do {
 639                 pgd = pgdp + pgd_index(addr);
 640
 641                 if (pgd_none(*pgd)) {
 642                         pud = pud_alloc_one(NULL, addr);
 643                         if (!pud) {
 644                                 kvm_err("Cannot allocate Hyp pud\n");
 645                                 err = -ENOMEM;
 646                                 goto out;
 647                         }
 648                         pgd_populate(NULL, pgd, pud);
 649                         get_page(virt_to_page(pgd));
 650                         kvm_flush_dcache_to_poc(pgd, sizeof(*pgd));
 651                 }
 652
 653                 next = pgd_addr_end(addr, end);
 654                 err = create_hyp_pud_mappings(pgd, addr, next, pfn, prot);
 655                 if (err)
 656                         goto out;
 657                 pfn += (next - addr) >> PAGE_SHIFT;
 658         } while (addr = next, addr != end);
 659 out:
 660         mutex_unlock(&kvm_hyp_pgd_mutex);
 661         return err;
 662 }
 663
 664 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
 665 {
 666         if (!is_vmalloc_addr(kaddr)) {
 667                 BUG_ON(!virt_addr_valid(kaddr));
 668                 return __pa(kaddr);
 669         } else {
 670                 return page_to_phys(vmalloc_to_page(kaddr)) +
 671                        offset_in_page(kaddr);
 672         }
 673 }
 674
 675 /**
 676  * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
 677  * @from:       The virtual kernel start address of the range
 678  * @to:         The virtual kernel end address of the range (exclusive)
 679  * @prot:       The protection to be applied to this range
 680  *
 681  * The same virtual address as the kernel virtual address is also used
 682  * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
 683  * physical pages.
 684  */
 685 int create_hyp_mappings(void *from, void *to, pgprot_t prot)
 686 {
 687         phys_addr_t phys_addr;
 688         unsigned long virt_addr;
 689         unsigned long start = kern_hyp_va((unsigned long)from);
 690         unsigned long end = kern_hyp_va((unsigned long)to);
 691
 692         if (is_kernel_in_hyp_mode())
 693                 return 0;
 694
 695         start = start & PAGE_MASK;
 696         end = PAGE_ALIGN(end);
 697
 698         for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
 699                 int err;
 700
 701                 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
 702                 err = __create_hyp_mappings(hyp_pgd, virt_addr,
 703                                             virt_addr + PAGE_SIZE,
 704                                             __phys_to_pfn(phys_addr),
 705                                             prot);
 706                 if (err)
 707                         return err;
 708         }
 709
 710         return 0;
 711 }
 712
 713 /**
 714  * create_hyp_io_mappings - duplicate a kernel IO mapping into Hyp mode
 715  * @from:       The kernel start VA of the range
 716  * @to:         The kernel end VA of the range (exclusive)
 717  * @phys_addr:  The physical start address which gets mapped
 718  *
 719  * The resulting HYP VA is the same as the kernel VA, modulo
 720  * HYP_PAGE_OFFSET.
 721  */
 722 int create_hyp_io_mappings(void *from, void *to, phys_addr_t phys_addr)
 723 {
 724         unsigned long start = kern_hyp_va((unsigned long)from);
 725         unsigned long end = kern_hyp_va((unsigned long)to);
 726
 727         if (is_kernel_in_hyp_mode())
 728                 return 0;
 729
 730         /* Check for a valid kernel IO mapping */
 731         if (!is_vmalloc_addr(from) || !is_vmalloc_addr(to - 1))
 732                 return -EINVAL;
 733
 734         return __create_hyp_mappings(hyp_pgd, start, end,
 735                                      __phys_to_pfn(phys_addr), PAGE_HYP_DEVICE);
 736 }
 737
 738 /**
 739  * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
 740  * @kvm:        The KVM struct pointer for the VM.
 741  *
 742  * Allocates only the stage-2 HW PGD level table(s) (can support either full
 743  * 40-bit input addresses or limited to 32-bit input addresses). Clears the
 744  * allocated pages.
 745  *
 746  * Note we don't need locking here as this is only called when the VM is
 747  * created, which can only be done once.
 748  */
 749 int kvm_alloc_stage2_pgd(struct kvm *kvm)
 750 {
 751         pgd_t *pgd;
 752
 753         if (kvm->arch.pgd != NULL) {
 754                 kvm_err("kvm_arch already initialized?\n");
 755                 return -EINVAL;
 756         }
 757
 758         /* Allocate the HW PGD, making sure that each page gets its own refcount */
 759         pgd = alloc_pages_exact(S2_PGD_SIZE, GFP_KERNEL | __GFP_ZERO);
 760         if (!pgd)
 761                 return -ENOMEM;
 762
 763         kvm->arch.pgd = pgd;
 764         return 0;
 765 }
 766
 767 static void stage2_unmap_memslot(struct kvm *kvm,
 768                                  struct kvm_memory_slot *memslot)
 769 {
 770         hva_t hva = memslot->userspace_addr;
 771         phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
 772         phys_addr_t size = PAGE_SIZE * memslot->npages;
 773         hva_t reg_end = hva + size;
 774
 775         /*
 776          * A memory region could potentially cover multiple VMAs, and any holes
 777          * between them, so iterate over all of them to find out if we should
 778          * unmap any of them.
 779          *
 780          *     +--------------------------------------------+
 781          * +---------------+----------------+   +----------------+
 782          * |   : VMA 1     |      VMA 2     |   |    VMA 3  :    |
 783          * +---------------+----------------+   +----------------+
 784          *     |               memory region                |
 785          *     +--------------------------------------------+
 786          */
 787         do {
 788                 struct vm_area_struct *vma = find_vma(current->mm, hva);
 789                 hva_t vm_start, vm_end;
 790
 791                 if (!vma || vma->vm_start >= reg_end)
 792                         break;
 793
 794                 /*
 795                  * Take the intersection of this VMA with the memory region
 796                  */
 797                 vm_start = max(hva, vma->vm_start);
 798                 vm_end = min(reg_end, vma->vm_end);
 799
 800                 if (!(vma->vm_flags & VM_PFNMAP)) {
 801                         gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
 802                         unmap_stage2_range(kvm, gpa, vm_end - vm_start);
 803                 }
 804                 hva = vm_end;
 805         } while (hva < reg_end);
 806 }
 807
 808 /**
 809  * stage2_unmap_vm - Unmap Stage-2 RAM mappings
 810  * @kvm: The struct kvm pointer
 811  *
 812  * Go through the memregions and unmap any reguler RAM
 813  * backing memory already mapped to the VM.
 814  */
 815 void stage2_unmap_vm(struct kvm *kvm)
 816 {
 817         struct kvm_memslots *slots;
 818         struct kvm_memory_slot *memslot;
 819         int idx;
 820
 821         idx = srcu_read_lock(&kvm->srcu);
 822         down_read(&current->mm->mmap_sem);
 823         spin_lock(&kvm->mmu_lock);
 824
 825         slots = kvm_memslots(kvm);
 826         kvm_for_each_memslot(memslot, slots)
 827                 stage2_unmap_memslot(kvm, memslot);
 828
 829         spin_unlock(&kvm->mmu_lock);
 830         up_read(&current->mm->mmap_sem);
 831         srcu_read_unlock(&kvm->srcu, idx);
 832 }
 833
 834 /**
 835  * kvm_free_stage2_pgd - free all stage-2 tables
 836  * @kvm:        The KVM struct pointer for the VM.
 837  *
 838  * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
 839  * underlying level-2 and level-3 tables before freeing the actual level-1 table
 840  * and setting the struct pointer to NULL.
 841  */
 842 void kvm_free_stage2_pgd(struct kvm *kvm)
 843 {
 844         void *pgd = NULL;
 845
 846         spin_lock(&kvm->mmu_lock);
 847         if (kvm->arch.pgd) {
 848                 unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
 849                 pgd = READ_ONCE(kvm->arch.pgd);
 850                 kvm->arch.pgd = NULL;
 851         }
 852         spin_unlock(&kvm->mmu_lock);
 853
 854         /* Free the HW pgd, one page at a time */
 855         if (pgd)
 856                 free_pages_exact(pgd, S2_PGD_SIZE);
 857 }
 858
 859 static pud_t *stage2_get_pud(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
 860                              phys_addr_t addr)
 861 {
 862         pgd_t *pgd;
 863         pud_t *pud;
 864
 865         pgd = kvm->arch.pgd + stage2_pgd_index(addr);
 866         if (WARN_ON(stage2_pgd_none(*pgd))) {
 867                 if (!cache)
 868                         return NULL;
 869                 pud = mmu_memory_cache_alloc(cache);
 870                 stage2_pgd_populate(pgd, pud);
 871                 get_page(virt_to_page(pgd));
 872         }
 873
 874         return stage2_pud_offset(pgd, addr);
 875 }
 876
 877 static pmd_t *stage2_get_pmd(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
 878                              phys_addr_t addr)
 879 {
 880         pud_t *pud;
 881         pmd_t *pmd;
 882
 883         pud = stage2_get_pud(kvm, cache, addr);
 884         if (!pud)
 885                 return NULL;
 886
 887         if (stage2_pud_none(*pud)) {
 888                 if (!cache)
 889                         return NULL;
 890                 pmd = mmu_memory_cache_alloc(cache);
 891                 stage2_pud_populate(pud, pmd);
 892                 get_page(virt_to_page(pud));
 893         }
 894
 895         return stage2_pmd_offset(pud, addr);
 896 }
 897
 898 static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
 899                                *cache, phys_addr_t addr, const pmd_t *new_pmd)
 900 {
 901         pmd_t *pmd, old_pmd;
 902
 903         pmd = stage2_get_pmd(kvm, cache, addr);
 904         VM_BUG_ON(!pmd);
 905
 906         /*
 907          * Mapping in huge pages should only happen through a fault.  If a
 908          * page is merged into a transparent huge page, the individual
 909          * subpages of that huge page should be unmapped through MMU
 910          * notifiers before we get here.
 911          *
 912          * Merging of CompoundPages is not supported; they should become
 913          * splitting first, unmapped, merged, and mapped back in on-demand.
 914          */
 915         VM_BUG_ON(pmd_present(*pmd) && pmd_pfn(*pmd) != pmd_pfn(*new_pmd));
 916
 917         old_pmd = *pmd;
 918         if (pmd_present(old_pmd)) {
 919                 pmd_clear(pmd);
 920                 kvm_tlb_flush_vmid_ipa(kvm, addr);
 921         } else {
 922                 get_page(virt_to_page(pmd));
 923         }
 924
 925         kvm_set_pmd(pmd, *new_pmd);
 926         return 0;
 927 }
 928
 929 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
 930                           phys_addr_t addr, const pte_t *new_pte,
 931                           unsigned long flags)
 932 {
 933         pmd_t *pmd;
 934         pte_t *pte, old_pte;
 935         bool iomap = flags & KVM_S2PTE_FLAG_IS_IOMAP;
 936         bool logging_active = flags & KVM_S2_FLAG_LOGGING_ACTIVE;
 937
 938         VM_BUG_ON(logging_active && !cache);
 939
 940         /* Create stage-2 page table mapping - Levels 0 and 1 */
 941         pmd = stage2_get_pmd(kvm, cache, addr);
 942         if (!pmd) {
 943                 /*
 944                  * Ignore calls from kvm_set_spte_hva for unallocated
 945                  * address ranges.
 946                  */
 947                 return 0;
 948         }
 949
 950         /*
 951          * While dirty page logging - dissolve huge PMD, then continue on to
 952          * allocate page.
 953          */
 954         if (logging_active)
 955                 stage2_dissolve_pmd(kvm, addr, pmd);
 956
 957         /* Create stage-2 page mappings - Level 2 */
 958         if (pmd_none(*pmd)) {
 959                 if (!cache)
 960                         return 0; /* ignore calls from kvm_set_spte_hva */
 961                 pte = mmu_memory_cache_alloc(cache);
 962                 pmd_populate_kernel(NULL, pmd, pte);
 963                 get_page(virt_to_page(pmd));
 964         }
 965
 966         pte = pte_offset_kernel(pmd, addr);
 967
 968         if (iomap && pte_present(*pte))
 969                 return -EFAULT;
 970
 971         /* Create 2nd stage page table mapping - Level 3 */
 972         old_pte = *pte;
 973         if (pte_present(old_pte)) {
 974                 kvm_set_pte(pte, __pte(0));
 975                 kvm_tlb_flush_vmid_ipa(kvm, addr);
 976         } else {
 977                 get_page(virt_to_page(pte));
 978         }
 979
 980         kvm_set_pte(pte, *new_pte);
 981         return 0;
 982 }
 983
 984 #ifndef __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
 985 static int stage2_ptep_test_and_clear_young(pte_t *pte)
 986 {
 987         if (pte_young(*pte)) {
 988                 *pte = pte_mkold(*pte);
 989                 return 1;
 990         }
 991         return 0;
 992 }
 993 #else
 994 static int stage2_ptep_test_and_clear_young(pte_t *pte)
 995 {
 996         return __ptep_test_and_clear_young(pte);
 997 }
 998 #endif
 999
1000 static int stage2_pmdp_test_and_clear_young(pmd_t *pmd)
1001 {
1002         return stage2_ptep_test_and_clear_young((pte_t *)pmd);
1003 }
1004
1005 /**
1006  * kvm_phys_addr_ioremap - map a device range to guest IPA
1007  *
1008  * @kvm:        The KVM pointer
1009  * @guest_ipa:  The IPA at which to insert the mapping
1010  * @pa:         The physical address of the device
1011  * @size:       The size of the mapping
1012  */
1013 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
1014                           phys_addr_t pa, unsigned long size, bool writable)
1015 {
1016         phys_addr_t addr, end;
1017         int ret = 0;
1018         unsigned long pfn;
1019         struct kvm_mmu_memory_cache cache = { 0, };
1020
1021         end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
1022         pfn = __phys_to_pfn(pa);
1023
1024         for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
1025                 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
1026
1027                 if (writable)
1028                         pte = kvm_s2pte_mkwrite(pte);
1029
1030                 ret = mmu_topup_memory_cache(&cache, KVM_MMU_CACHE_MIN_PAGES,
1031                                                 KVM_NR_MEM_OBJS);
1032                 if (ret)
1033                         goto out;
1034                 spin_lock(&kvm->mmu_lock);
1035                 ret = stage2_set_pte(kvm, &cache, addr, &pte,
1036                                                 KVM_S2PTE_FLAG_IS_IOMAP);
1037                 spin_unlock(&kvm->mmu_lock);
1038                 if (ret)
1039                         goto out;
1040
1041                 pfn++;
1042         }
1043
1044 out:
1045         mmu_free_memory_cache(&cache);
1046         return ret;
1047 }
1048
1049 static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
1050 {
1051         kvm_pfn_t pfn = *pfnp;
1052         gfn_t gfn = *ipap >> PAGE_SHIFT;
1053
1054         if (PageTransCompoundMap(pfn_to_page(pfn))) {
1055                 unsigned long mask;
1056                 /*
1057                  * The address we faulted on is backed by a transparent huge
1058                  * page.  However, because we map the compound huge page and
1059                  * not the individual tail page, we need to transfer the
1060                  * refcount to the head page.  We have to be careful that the
1061                  * THP doesn't start to split while we are adjusting the
1062                  * refcounts.
1063                  *
1064                  * We are sure this doesn't happen, because mmu_notifier_retry
1065                  * was successful and we are holding the mmu_lock, so if this
1066                  * THP is trying to split, it will be blocked in the mmu
1067                  * notifier before touching any of the pages, specifically
1068                  * before being able to call __split_huge_page_refcount().
1069                  *
1070                  * We can therefore safely transfer the refcount from PG_tail
1071                  * to PG_head and switch the pfn from a tail page to the head
1072                  * page accordingly.
1073                  */
1074                 mask = PTRS_PER_PMD - 1;
1075                 VM_BUG_ON((gfn & mask) != (pfn & mask));
1076                 if (pfn & mask) {
1077                         *ipap &= PMD_MASK;
1078                         kvm_release_pfn_clean(pfn);
1079                         pfn &= ~mask;
1080                         kvm_get_pfn(pfn);
1081                         *pfnp = pfn;
1082                 }
1083
1084                 return true;
1085         }
1086
1087         return false;
1088 }
1089
1090 static bool kvm_is_write_fault(struct kvm_vcpu *vcpu)
1091 {
1092         if (kvm_vcpu_trap_is_iabt(vcpu))
1093                 return false;
1094
1095         return kvm_vcpu_dabt_iswrite(vcpu);
1096 }
1097
1098 /**
1099  * stage2_wp_ptes - write protect PMD range
1100  * @pmd:        pointer to pmd entry
1101  * @addr:       range start address
1102  * @end:        range end address
1103  */
1104 static void stage2_wp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
1105 {
1106         pte_t *pte;
1107
1108         pte = pte_offset_kernel(pmd, addr);
1109         do {
1110                 if (!pte_none(*pte)) {
1111                         if (!kvm_s2pte_readonly(pte))
1112                                 kvm_set_s2pte_readonly(pte);
1113                 }
1114         } while (pte++, addr += PAGE_SIZE, addr != end);
1115 }
1116
1117 /**
1118  * stage2_wp_pmds - write protect PUD range
1119  * @pud:        pointer to pud entry
1120  * @addr:       range start address
1121  * @end:        range end address
1122  */
1123 static void stage2_wp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
1124 {
1125         pmd_t *pmd;
1126         phys_addr_t next;
1127
1128         pmd = stage2_pmd_offset(pud, addr);
1129
1130         do {
1131                 next = stage2_pmd_addr_end(addr, end);
1132                 if (!pmd_none(*pmd)) {
1133                         if (pmd_thp_or_huge(*pmd)) {
1134                                 if (!kvm_s2pmd_readonly(pmd))
1135                                         kvm_set_s2pmd_readonly(pmd);
1136                         } else {
1137                                 stage2_wp_ptes(pmd, addr, next);
1138                         }
1139                 }
1140         } while (pmd++, addr = next, addr != end);
1141 }
1142
1143 /**
1144   * stage2_wp_puds - write protect PGD range
1145   * @pgd:       pointer to pgd entry
1146   * @addr:      range start address
1147   * @end:       range end address
1148   *
1149   * Process PUD entries, for a huge PUD we cause a panic.
1150   */
1151 static void  stage2_wp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
1152 {
1153         pud_t *pud;
1154         phys_addr_t next;
1155
1156         pud = stage2_pud_offset(pgd, addr);
1157         do {
1158                 next = stage2_pud_addr_end(addr, end);
1159                 if (!stage2_pud_none(*pud)) {
1160                         /* TODO:PUD not supported, revisit later if supported */
1161                         BUG_ON(stage2_pud_huge(*pud));
1162                         stage2_wp_pmds(pud, addr, next);
1163                 }
1164         } while (pud++, addr = next, addr != end);
1165 }
1166
1167 /**
1168  * stage2_wp_range() - write protect stage2 memory region range
1169  * @kvm:        The KVM pointer
1170  * @addr:       Start address of range
1171  * @end:        End address of range
1172  */
1173 static void stage2_wp_range(struct kvm *kvm, phys_addr_t addr, phys_addr_t end)
1174 {
1175         pgd_t *pgd;
1176         phys_addr_t next;
1177
1178         pgd = kvm->arch.pgd + stage2_pgd_index(addr);
1179         do {
1180                 /*
1181                  * Release kvm_mmu_lock periodically if the memory region is
1182                  * large. Otherwise, we may see kernel panics with
1183                  * CONFIG_DETECT_HUNG_TASK, CONFIG_LOCKUP_DETECTOR,
1184                  * CONFIG_LOCKDEP. Additionally, holding the lock too long
1185                  * will also starve other vCPUs. We have to also make sure
1186                  * that the page tables are not freed while we released
1187                  * the lock.
1188                  */
1189                 cond_resched_lock(&kvm->mmu_lock);
1190                 if (!READ_ONCE(kvm->arch.pgd))
1191                         break;
1192                 next = stage2_pgd_addr_end(addr, end);
1193                 if (stage2_pgd_present(*pgd))
1194                         stage2_wp_puds(pgd, addr, next);
1195         } while (pgd++, addr = next, addr != end);
1196 }
1197
1198 /**
1199  * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1200  * @kvm:        The KVM pointer
1201  * @slot:       The memory slot to write protect
1202  *
1203  * Called to start logging dirty pages after memory region
1204  * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1205  * all present PMD and PTEs are write protected in the memory region.
1206  * Afterwards read of dirty page log can be called.
1207  *
1208  * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1209  * serializing operations for VM memory regions.
1210  */
1211 void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1212 {
1213         struct kvm_memslots *slots = kvm_memslots(kvm);
1214         struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1215         phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
1216         phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1217
1218         spin_lock(&kvm->mmu_lock);
1219         stage2_wp_range(kvm, start, end);
1220         spin_unlock(&kvm->mmu_lock);
1221         kvm_flush_remote_tlbs(kvm);
1222 }
1223
1224 /**
1225  * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
1226  * @kvm:        The KVM pointer
1227  * @slot:       The memory slot associated with mask
1228  * @gfn_offset: The gfn offset in memory slot
1229  * @mask:       The mask of dirty pages at offset 'gfn_offset' in this memory
1230  *              slot to be write protected
1231  *
1232  * Walks bits set in mask write protects the associated pte's. Caller must
1233  * acquire kvm_mmu_lock.
1234  */
1235 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1236                 struct kvm_memory_slot *slot,
1237                 gfn_t gfn_offset, unsigned long mask)
1238 {
1239         phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1240         phys_addr_t start = (base_gfn +  __ffs(mask)) << PAGE_SHIFT;
1241         phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1242
1243         stage2_wp_range(kvm, start, end);
1244 }
1245
1246 /*
1247  * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1248  * dirty pages.
1249  *
1250  * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1251  * enable dirty logging for them.
1252  */
1253 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1254                 struct kvm_memory_slot *slot,
1255                 gfn_t gfn_offset, unsigned long mask)
1256 {
1257         kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1258 }
1259
1260 static void coherent_cache_guest_page(struct kvm_vcpu *vcpu, kvm_pfn_t pfn,
1261                                       unsigned long size)
1262 {
1263         __coherent_cache_guest_page(vcpu, pfn, size);
1264 }
1265
1266 static void kvm_send_hwpoison_signal(unsigned long address,
1267                                      struct vm_area_struct *vma)
1268 {
1269         siginfo_t info;
1270
1271         info.si_signo   = SIGBUS;
1272         info.si_errno   = 0;
1273         info.si_code    = BUS_MCEERR_AR;
1274         info.si_addr    = (void __user *)address;
1275
1276         if (is_vm_hugetlb_page(vma))
1277                 info.si_addr_lsb = huge_page_shift(hstate_vma(vma));
1278         else
1279                 info.si_addr_lsb = PAGE_SHIFT;
1280
1281         send_sig_info(SIGBUS, &info, current);
1282 }
1283
1284 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1285                           struct kvm_memory_slot *memslot, unsigned long hva,
1286                           unsigned long fault_status)
1287 {
1288         int ret;
1289         bool write_fault, writable, hugetlb = false, force_pte = false;
1290         unsigned long mmu_seq;
1291         gfn_t gfn = fault_ipa >> PAGE_SHIFT;
1292         struct kvm *kvm = vcpu->kvm;
1293         struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1294         struct vm_area_struct *vma;
1295         kvm_pfn_t pfn;
1296         pgprot_t mem_type = PAGE_S2;
1297         bool logging_active = memslot_is_logging(memslot);
1298         unsigned long flags = 0;
1299
1300         write_fault = kvm_is_write_fault(vcpu);
1301         if (fault_status == FSC_PERM && !write_fault) {
1302                 kvm_err("Unexpected L2 read permission error\n");
1303                 return -EFAULT;
1304         }
1305
1306         /* Let's check if we will get back a huge page backed by hugetlbfs */
1307         down_read(&current->mm->mmap_sem);
1308         vma = find_vma_intersection(current->mm, hva, hva + 1);
1309         if (unlikely(!vma)) {
1310                 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1311                 up_read(&current->mm->mmap_sem);
1312                 return -EFAULT;
1313         }
1314
1315         if (is_vm_hugetlb_page(vma) && !logging_active) {
1316                 hugetlb = true;
1317                 gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
1318         } else {
1319                 /*
1320                  * Pages belonging to memslots that don't have the same
1321                  * alignment for userspace and IPA cannot be mapped using
1322                  * block descriptors even if the pages belong to a THP for
1323                  * the process, because the stage-2 block descriptor will
1324                  * cover more than a single THP and we loose atomicity for
1325                  * unmapping, updates, and splits of the THP or other pages
1326                  * in the stage-2 block range.
1327                  */
1328                 if ((memslot->userspace_addr & ~PMD_MASK) !=
1329                     ((memslot->base_gfn << PAGE_SHIFT) & ~PMD_MASK))
1330                         force_pte = true;
1331         }
1332         up_read(&current->mm->mmap_sem);
1333
1334         /* We need minimum second+third level pages */
1335         ret = mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES,
1336                                      KVM_NR_MEM_OBJS);
1337         if (ret)
1338                 return ret;
1339
1340         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1341         /*
1342          * Ensure the read of mmu_notifier_seq happens before we call
1343          * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1344          * the page we just got a reference to gets unmapped before we have a
1345          * chance to grab the mmu_lock, which ensure that if the page gets
1346          * unmapped afterwards, the call to kvm_unmap_hva will take it away
1347          * from us again properly. This smp_rmb() interacts with the smp_wmb()
1348          * in kvm_mmu_notifier_invalidate_<page|range_end>.
1349          */
1350         smp_rmb();
1351
1352         pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
1353         if (pfn == KVM_PFN_ERR_HWPOISON) {
1354                 kvm_send_hwpoison_signal(hva, vma);
1355                 return 0;
1356         }
1357         if (is_error_noslot_pfn(pfn))
1358                 return -EFAULT;
1359
1360         if (kvm_is_device_pfn(pfn)) {
1361                 mem_type = PAGE_S2_DEVICE;
1362                 flags |= KVM_S2PTE_FLAG_IS_IOMAP;
1363         } else if (logging_active) {
1364                 /*
1365                  * Faults on pages in a memslot with logging enabled
1366                  * should not be mapped with huge pages (it introduces churn
1367                  * and performance degradation), so force a pte mapping.
1368                  */
1369                 force_pte = true;
1370                 flags |= KVM_S2_FLAG_LOGGING_ACTIVE;
1371
1372                 /*
1373                  * Only actually map the page as writable if this was a write
1374                  * fault.
1375                  */
1376                 if (!write_fault)
1377                         writable = false;
1378         }
1379
1380         spin_lock(&kvm->mmu_lock);
1381         if (mmu_notifier_retry(kvm, mmu_seq))
1382                 goto out_unlock;
1383
1384         if (!hugetlb && !force_pte)
1385                 hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
1386
1387         if (hugetlb) {
1388                 pmd_t new_pmd = pfn_pmd(pfn, mem_type);
1389                 new_pmd = pmd_mkhuge(new_pmd);
1390                 if (writable) {
1391                         new_pmd = kvm_s2pmd_mkwrite(new_pmd);
1392                         kvm_set_pfn_dirty(pfn);
1393                 }
1394                 coherent_cache_guest_page(vcpu, pfn, PMD_SIZE);
1395                 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
1396         } else {
1397                 pte_t new_pte = pfn_pte(pfn, mem_type);
1398
1399                 if (writable) {
1400                         new_pte = kvm_s2pte_mkwrite(new_pte);
1401                         kvm_set_pfn_dirty(pfn);
1402                         mark_page_dirty(kvm, gfn);
1403                 }
1404                 coherent_cache_guest_page(vcpu, pfn, PAGE_SIZE);
1405                 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte, flags);
1406         }
1407
1408 out_unlock:
1409         spin_unlock(&kvm->mmu_lock);
1410         kvm_set_pfn_accessed(pfn);
1411         kvm_release_pfn_clean(pfn);
1412         return ret;
1413 }
1414
1415 /*
1416  * Resolve the access fault by making the page young again.
1417  * Note that because the faulting entry is guaranteed not to be
1418  * cached in the TLB, we don't need to invalidate anything.
1419  * Only the HW Access Flag updates are supported for Stage 2 (no DBM),
1420  * so there is no need for atomic (pte|pmd)_mkyoung operations.
1421  */
1422 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1423 {
1424         pmd_t *pmd;
1425         pte_t *pte;
1426         kvm_pfn_t pfn;
1427         bool pfn_valid = false;
1428
1429         trace_kvm_access_fault(fault_ipa);
1430
1431         spin_lock(&vcpu->kvm->mmu_lock);
1432
1433         pmd = stage2_get_pmd(vcpu->kvm, NULL, fault_ipa);
1434         if (!pmd || pmd_none(*pmd))     /* Nothing there */
1435                 goto out;
1436
1437         if (pmd_thp_or_huge(*pmd)) {    /* THP, HugeTLB */
1438                 *pmd = pmd_mkyoung(*pmd);
1439                 pfn = pmd_pfn(*pmd);
1440                 pfn_valid = true;
1441                 goto out;
1442         }
1443
1444         pte = pte_offset_kernel(pmd, fault_ipa);
1445         if (pte_none(*pte))             /* Nothing there either */
1446                 goto out;
1447
1448         *pte = pte_mkyoung(*pte);       /* Just a page... */
1449         pfn = pte_pfn(*pte);
1450         pfn_valid = true;
1451 out:
1452         spin_unlock(&vcpu->kvm->mmu_lock);
1453         if (pfn_valid)
1454                 kvm_set_pfn_accessed(pfn);
1455 }
1456
1457 static bool is_abort_sea(unsigned long fault_status)
1458 {
1459         switch (fault_status) {
1460         case FSC_SEA:
1461         case FSC_SEA_TTW0:
1462         case FSC_SEA_TTW1:
1463         case FSC_SEA_TTW2:
1464         case FSC_SEA_TTW3:
1465         case FSC_SECC:
1466         case FSC_SECC_TTW0:
1467         case FSC_SECC_TTW1:
1468         case FSC_SECC_TTW2:
1469         case FSC_SECC_TTW3:
1470                 return true;
1471         default:
1472                 return false;
1473         }
1474 }
1475
1476 /**
1477  * kvm_handle_guest_abort - handles all 2nd stage aborts
1478  * @vcpu:       the VCPU pointer
1479  * @run:        the kvm_run structure
1480  *
1481  * Any abort that gets to the host is almost guaranteed to be caused by a
1482  * missing second stage translation table entry, which can mean that either the
1483  * guest simply needs more memory and we must allocate an appropriate page or it
1484  * can mean that the guest tried to access I/O memory, which is emulated by user
1485  * space. The distinction is based on the IPA causing the fault and whether this
1486  * memory region has been registered as standard RAM by user space.
1487  */
1488 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
1489 {
1490         unsigned long fault_status;
1491         phys_addr_t fault_ipa;
1492         struct kvm_memory_slot *memslot;
1493         unsigned long hva;
1494         bool is_iabt, write_fault, writable;
1495         gfn_t gfn;
1496         int ret, idx;
1497
1498         fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1499
1500         fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1501
1502         /*
1503          * The host kernel will handle the synchronous external abort. There
1504          * is no need to pass the error into the guest.
1505          */
1506         if (is_abort_sea(fault_status)) {
1507                 if (!handle_guest_sea(fault_ipa, kvm_vcpu_get_hsr(vcpu)))
1508                         return 1;
1509         }
1510
1511         is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1512         if (unlikely(!is_iabt && kvm_vcpu_dabt_isextabt(vcpu))) {
1513                 kvm_inject_vabt(vcpu);
1514                 return 1;
1515         }
1516
1517         trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
1518                               kvm_vcpu_get_hfar(vcpu), fault_ipa);
1519
1520         /* Check the stage-2 fault is trans. fault or write fault */
1521         if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1522             fault_status != FSC_ACCESS) {
1523                 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1524                         kvm_vcpu_trap_get_class(vcpu),
1525                         (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1526                         (unsigned long)kvm_vcpu_get_hsr(vcpu));
1527                 return -EFAULT;
1528         }
1529
1530         idx = srcu_read_lock(&vcpu->kvm->srcu);
1531
1532         gfn = fault_ipa >> PAGE_SHIFT;
1533         memslot = gfn_to_memslot(vcpu->kvm, gfn);
1534         hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1535         write_fault = kvm_is_write_fault(vcpu);
1536         if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1537                 if (is_iabt) {
1538                         /* Prefetch Abort on I/O address */
1539                         kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1540                         ret = 1;
1541                         goto out_unlock;
1542                 }
1543
1544                 /*
1545                  * Check for a cache maintenance operation. Since we
1546                  * ended-up here, we know it is outside of any memory
1547                  * slot. But we can't find out if that is for a device,
1548                  * or if the guest is just being stupid. The only thing
1549                  * we know for sure is that this range cannot be cached.
1550                  *
1551                  * So let's assume that the guest is just being
1552                  * cautious, and skip the instruction.
1553                  */
1554                 if (kvm_vcpu_dabt_is_cm(vcpu)) {
1555                         kvm_skip_instr(vcpu, kvm_vcpu_trap_il_is32bit(vcpu));
1556                         ret = 1;
1557                         goto out_unlock;
1558                 }
1559
1560                 /*
1561                  * The IPA is reported as [MAX:12], so we need to
1562                  * complement it with the bottom 12 bits from the
1563                  * faulting VA. This is always 12 bits, irrespective
1564                  * of the page size.
1565                  */
1566                 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1567                 ret = io_mem_abort(vcpu, run, fault_ipa);
1568                 goto out_unlock;
1569         }
1570
1571         /* Userspace should not be able to register out-of-bounds IPAs */
1572         VM_BUG_ON(fault_ipa >= KVM_PHYS_SIZE);
1573
1574         if (fault_status == FSC_ACCESS) {
1575                 handle_access_fault(vcpu, fault_ipa);
1576                 ret = 1;
1577                 goto out_unlock;
1578         }
1579
1580         ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1581         if (ret == 0)
1582                 ret = 1;
1583 out_unlock:
1584         srcu_read_unlock(&vcpu->kvm->srcu, idx);
1585         return ret;
1586 }
1587
1588 static int handle_hva_to_gpa(struct kvm *kvm,
1589                              unsigned long start,
1590                              unsigned long end,
1591                              int (*handler)(struct kvm *kvm,
1592                                             gpa_t gpa, u64 size,
1593                                             void *data),
1594                              void *data)
1595 {
1596         struct kvm_memslots *slots;
1597         struct kvm_memory_slot *memslot;
1598         int ret = 0;
1599
1600         slots = kvm_memslots(kvm);
1601
1602         /* we only care about the pages that the guest sees */
1603         kvm_for_each_memslot(memslot, slots) {
1604                 unsigned long hva_start, hva_end;
1605                 gfn_t gpa;
1606
1607                 hva_start = max(start, memslot->userspace_addr);
1608                 hva_end = min(end, memslot->userspace_addr +
1609                                         (memslot->npages << PAGE_SHIFT));
1610                 if (hva_start >= hva_end)
1611                         continue;
1612
1613                 gpa = hva_to_gfn_memslot(hva_start, memslot) << PAGE_SHIFT;
1614                 ret |= handler(kvm, gpa, (u64)(hva_end - hva_start), data);
1615         }
1616
1617         return ret;
1618 }
1619
1620 static int kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1621 {
1622         unmap_stage2_range(kvm, gpa, size);
1623         return 0;
1624 }
1625
1626 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1627 {
1628         unsigned long end = hva + PAGE_SIZE;
1629
1630         if (!kvm->arch.pgd)
1631                 return 0;
1632
1633         trace_kvm_unmap_hva(hva);
1634         handle_hva_to_gpa(kvm, hva, end, &kvm_unmap_hva_handler, NULL);
1635         return 0;
1636 }
1637
1638 int kvm_unmap_hva_range(struct kvm *kvm,
1639                         unsigned long start, unsigned long end)
1640 {
1641         if (!kvm->arch.pgd)
1642                 return 0;
1643
1644         trace_kvm_unmap_hva_range(start, end);
1645         handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
1646         return 0;
1647 }
1648
1649 static int kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1650 {
1651         pte_t *pte = (pte_t *)data;
1652
1653         WARN_ON(size != PAGE_SIZE);
1654         /*
1655          * We can always call stage2_set_pte with KVM_S2PTE_FLAG_LOGGING_ACTIVE
1656          * flag clear because MMU notifiers will have unmapped a huge PMD before
1657          * calling ->change_pte() (which in turn calls kvm_set_spte_hva()) and
1658          * therefore stage2_set_pte() never needs to clear out a huge PMD
1659          * through this calling path.
1660          */
1661         stage2_set_pte(kvm, NULL, gpa, pte, 0);
1662         return 0;
1663 }
1664
1665
1666 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1667 {
1668         unsigned long end = hva + PAGE_SIZE;
1669         pte_t stage2_pte;
1670
1671         if (!kvm->arch.pgd)
1672                 return;
1673
1674         trace_kvm_set_spte_hva(hva);
1675         stage2_pte = pfn_pte(pte_pfn(pte), PAGE_S2);
1676         handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
1677 }
1678
1679 static int kvm_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1680 {
1681         pmd_t *pmd;
1682         pte_t *pte;
1683
1684         WARN_ON(size != PAGE_SIZE && size != PMD_SIZE);
1685         pmd = stage2_get_pmd(kvm, NULL, gpa);
1686         if (!pmd || pmd_none(*pmd))     /* Nothing there */
1687                 return 0;
1688
1689         if (pmd_thp_or_huge(*pmd))      /* THP, HugeTLB */
1690                 return stage2_pmdp_test_and_clear_young(pmd);
1691
1692         pte = pte_offset_kernel(pmd, gpa);
1693         if (pte_none(*pte))
1694                 return 0;
1695
1696         return stage2_ptep_test_and_clear_young(pte);
1697 }
1698
1699 static int kvm_test_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1700 {
1701         pmd_t *pmd;
1702         pte_t *pte;
1703
1704         WARN_ON(size != PAGE_SIZE && size != PMD_SIZE);
1705         pmd = stage2_get_pmd(kvm, NULL, gpa);
1706         if (!pmd || pmd_none(*pmd))     /* Nothing there */
1707                 return 0;
1708
1709         if (pmd_thp_or_huge(*pmd))              /* THP, HugeTLB */
1710                 return pmd_young(*pmd);
1711
1712         pte = pte_offset_kernel(pmd, gpa);
1713         if (!pte_none(*pte))            /* Just a page... */
1714                 return pte_young(*pte);
1715
1716         return 0;
1717 }
1718
1719 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
1720 {
1721         if (!kvm->arch.pgd)
1722                 return 0;
1723         trace_kvm_age_hva(start, end);
1724         return handle_hva_to_gpa(kvm, start, end, kvm_age_hva_handler, NULL);
1725 }
1726
1727 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1728 {
1729         if (!kvm->arch.pgd)
1730                 return 0;
1731         trace_kvm_test_age_hva(hva);
1732         return handle_hva_to_gpa(kvm, hva, hva, kvm_test_age_hva_handler, NULL);
1733 }
1734
1735 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1736 {
1737         mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
1738 }
1739
1740 phys_addr_t kvm_mmu_get_httbr(void)
1741 {
1742         if (__kvm_cpu_uses_extended_idmap())
1743                 return virt_to_phys(merged_hyp_pgd);
1744         else
1745                 return virt_to_phys(hyp_pgd);
1746 }
1747
1748 phys_addr_t kvm_get_idmap_vector(void)
1749 {
1750         return hyp_idmap_vector;
1751 }
1752
1753 static int kvm_map_idmap_text(pgd_t *pgd)
1754 {
1755         int err;
1756
1757         /* Create the idmap in the boot page tables */
1758         err =   __create_hyp_mappings(pgd,
1759                                       hyp_idmap_start, hyp_idmap_end,
1760                                       __phys_to_pfn(hyp_idmap_start),
1761                                       PAGE_HYP_EXEC);
1762         if (err)
1763                 kvm_err("Failed to idmap %lx-%lx\n",
1764                         hyp_idmap_start, hyp_idmap_end);
1765
1766         return err;
1767 }
1768
1769 int kvm_mmu_init(void)
1770 {
1771         int err;
1772
1773         hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
1774         hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
1775         hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
1776
1777         /*
1778          * We rely on the linker script to ensure at build time that the HYP
1779          * init code does not cross a page boundary.
1780          */
1781         BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
1782
1783         kvm_info("IDMAP page: %lx\n", hyp_idmap_start);
1784         kvm_info("HYP VA range: %lx:%lx\n",
1785                  kern_hyp_va(PAGE_OFFSET), kern_hyp_va(~0UL));
1786
1787         if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
1788             hyp_idmap_start <  kern_hyp_va(~0UL) &&
1789             hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
1790                 /*
1791                  * The idmap page is intersecting with the VA space,
1792                  * it is not safe to continue further.
1793                  */
1794                 kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
1795                 err = -EINVAL;
1796                 goto out;
1797         }
1798
1799         hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1800         if (!hyp_pgd) {
1801                 kvm_err("Hyp mode PGD not allocated\n");
1802                 err = -ENOMEM;
1803                 goto out;
1804         }
1805
1806         if (__kvm_cpu_uses_extended_idmap()) {
1807                 boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
1808                                                          hyp_pgd_order);
1809                 if (!boot_hyp_pgd) {
1810                         kvm_err("Hyp boot PGD not allocated\n");
1811                         err = -ENOMEM;
1812                         goto out;
1813                 }
1814
1815                 err = kvm_map_idmap_text(boot_hyp_pgd);
1816                 if (err)
1817                         goto out;
1818
1819                 merged_hyp_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
1820                 if (!merged_hyp_pgd) {
1821                         kvm_err("Failed to allocate extra HYP pgd\n");
1822                         goto out;
1823                 }
1824                 __kvm_extend_hypmap(boot_hyp_pgd, hyp_pgd, merged_hyp_pgd,
1825                                     hyp_idmap_start);
1826         } else {
1827                 err = kvm_map_idmap_text(hyp_pgd);
1828                 if (err)
1829                         goto out;
1830         }
1831
1832         return 0;
1833 out:
1834         free_hyp_pgds();
1835         return err;
1836 }
1837
1838 void kvm_arch_commit_memory_region(struct kvm *kvm,
1839                                    const struct kvm_userspace_memory_region *mem,
1840                                    const struct kvm_memory_slot *old,
1841                                    const struct kvm_memory_slot *new,
1842                                    enum kvm_mr_change change)
1843 {
1844         /*
1845          * At this point memslot has been committed and there is an
1846          * allocated dirty_bitmap[], dirty pages will be be tracked while the
1847          * memory slot is write protected.
1848          */
1849         if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
1850                 kvm_mmu_wp_memory_region(kvm, mem->slot);
1851 }
1852
1853 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1854                                    struct kvm_memory_slot *memslot,
1855                                    const struct kvm_userspace_memory_region *mem,
1856                                    enum kvm_mr_change change)
1857 {
1858         hva_t hva = mem->userspace_addr;
1859         hva_t reg_end = hva + mem->memory_size;
1860         bool writable = !(mem->flags & KVM_MEM_READONLY);
1861         int ret = 0;
1862
1863         if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
1864                         change != KVM_MR_FLAGS_ONLY)
1865                 return 0;
1866
1867         /*
1868          * Prevent userspace from creating a memory region outside of the IPA
1869          * space addressable by the KVM guest IPA space.
1870          */
1871         if (memslot->base_gfn + memslot->npages >=
1872             (KVM_PHYS_SIZE >> PAGE_SHIFT))
1873                 return -EFAULT;
1874
1875         down_read(&current->mm->mmap_sem);
1876         /*
1877          * A memory region could potentially cover multiple VMAs, and any holes
1878          * between them, so iterate over all of them to find out if we can map
1879          * any of them right now.
1880          *
1881          *     +--------------------------------------------+
1882          * +---------------+----------------+   +----------------+
1883          * |   : VMA 1     |      VMA 2     |   |    VMA 3  :    |
1884          * +---------------+----------------+   +----------------+
1885          *     |               memory region                |
1886          *     +--------------------------------------------+
1887          */
1888         do {
1889                 struct vm_area_struct *vma = find_vma(current->mm, hva);
1890                 hva_t vm_start, vm_end;
1891
1892                 if (!vma || vma->vm_start >= reg_end)
1893                         break;
1894
1895                 /*
1896                  * Mapping a read-only VMA is only allowed if the
1897                  * memory region is configured as read-only.
1898                  */
1899                 if (writable && !(vma->vm_flags & VM_WRITE)) {
1900                         ret = -EPERM;
1901                         break;
1902                 }
1903
1904                 /*
1905                  * Take the intersection of this VMA with the memory region
1906                  */
1907                 vm_start = max(hva, vma->vm_start);
1908                 vm_end = min(reg_end, vma->vm_end);
1909
1910                 if (vma->vm_flags & VM_PFNMAP) {
1911                         gpa_t gpa = mem->guest_phys_addr +
1912                                     (vm_start - mem->userspace_addr);
1913                         phys_addr_t pa;
1914
1915                         pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
1916                         pa += vm_start - vma->vm_start;
1917
1918                         /* IO region dirty page logging not allowed */
1919                         if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1920                                 ret = -EINVAL;
1921                                 goto out;
1922                         }
1923
1924                         ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
1925                                                     vm_end - vm_start,
1926                                                     writable);
1927                         if (ret)
1928                                 break;
1929                 }
1930                 hva = vm_end;
1931         } while (hva < reg_end);
1932
1933         if (change == KVM_MR_FLAGS_ONLY)
1934                 goto out;
1935
1936         spin_lock(&kvm->mmu_lock);
1937         if (ret)
1938                 unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
1939         else
1940                 stage2_flush_memslot(kvm, memslot);
1941         spin_unlock(&kvm->mmu_lock);
1942 out:
1943         up_read(&current->mm->mmap_sem);
1944         return ret;
1945 }
1946
1947 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
1948                            struct kvm_memory_slot *dont)
1949 {
1950 }
1951
1952 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
1953                             unsigned long npages)
1954 {
1955         return 0;
1956 }
1957
1958 void kvm_arch_memslots_updated(struct kvm *kvm, struct kvm_memslots *slots)
1959 {
1960 }
1961
1962 void kvm_arch_flush_shadow_all(struct kvm *kvm)
1963 {
1964         kvm_free_stage2_pgd(kvm);
1965 }
1966
1967 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
1968                                    struct kvm_memory_slot *slot)
1969 {
1970         gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
1971         phys_addr_t size = slot->npages << PAGE_SHIFT;
1972
1973         spin_lock(&kvm->mmu_lock);
1974         unmap_stage2_range(kvm, gpa, size);
1975         spin_unlock(&kvm->mmu_lock);
1976 }
1977
1978 /*
1979  * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
1980  *
1981  * Main problems:
1982  * - S/W ops are local to a CPU (not broadcast)
1983  * - We have line migration behind our back (speculation)
1984  * - System caches don't support S/W at all (damn!)
1985  *
1986  * In the face of the above, the best we can do is to try and convert
1987  * S/W ops to VA ops. Because the guest is not allowed to infer the
1988  * S/W to PA mapping, it can only use S/W to nuke the whole cache,
1989  * which is a rather good thing for us.
1990  *
1991  * Also, it is only used when turning caches on/off ("The expected
1992  * usage of the cache maintenance instructions that operate by set/way
1993  * is associated with the cache maintenance instructions associated
1994  * with the powerdown and powerup of caches, if this is required by
1995  * the implementation.").
1996  *
1997  * We use the following policy:
1998  *
1999  * - If we trap a S/W operation, we enable VM trapping to detect
2000  *   caches being turned on/off, and do a full clean.
2001  *
2002  * - We flush the caches on both caches being turned on and off.
2003  *
2004  * - Once the caches are enabled, we stop trapping VM ops.
2005  */
2006 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
2007 {
2008         unsigned long hcr = vcpu_get_hcr(vcpu);
2009
2010         /*
2011          * If this is the first time we do a S/W operation
2012          * (i.e. HCR_TVM not set) flush the whole memory, and set the
2013          * VM trapping.
2014          *
2015          * Otherwise, rely on the VM trapping to wait for the MMU +
2016          * Caches to be turned off. At that point, we'll be able to
2017          * clean the caches again.
2018          */
2019         if (!(hcr & HCR_TVM)) {
2020                 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
2021                                         vcpu_has_cache_enabled(vcpu));
2022                 stage2_flush_vm(vcpu->kvm);
2023                 vcpu_set_hcr(vcpu, hcr | HCR_TVM);
2024         }
2025 }
2026
2027 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
2028 {
2029         bool now_enabled = vcpu_has_cache_enabled(vcpu);
2030
2031         /*
2032          * If switching the MMU+caches on, need to invalidate the caches.
2033          * If switching it off, need to clean the caches.
2034          * Clean + invalidate does the trick always.
2035          */
2036         if (now_enabled != was_enabled)
2037                 stage2_flush_vm(vcpu->kvm);
2038
2039         /* Caches are now on, stop trapping VM ops (until a S/W op) */
2040         if (now_enabled)
2041                 vcpu_set_hcr(vcpu, vcpu_get_hcr(vcpu) & ~HCR_TVM);
2042
2043         trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);
2044 }