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