linux/virt/kvm/arm/mmu.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 2012 - Virtual Open Systems and Columbia University
   4 * Author: Christoffer Dall <c.dall@virtualopensystems.com>
   5 */
   6
   7#include <linux/mman.h>
   8#include <linux/kvm_host.h>
   9#include <linux/io.h>
  10#include <linux/hugetlb.h>
  11#include <linux/sched/signal.h>
  12#include <trace/events/kvm.h>
  13#include <asm/pgalloc.h>
  14#include <asm/cacheflush.h>
  15#include <asm/kvm_arm.h>
  16#include <asm/kvm_mmu.h>
  17#include <asm/kvm_mmio.h>
  18#include <asm/kvm_ras.h>
  19#include <asm/kvm_asm.h>
  20#include <asm/kvm_emulate.h>
  21#include <asm/virt.h>
  22
  23#include "trace.h"
  24
  25static pgd_t *boot_hyp_pgd;
  26static pgd_t *hyp_pgd;
  27static pgd_t *merged_hyp_pgd;
  28static DEFINE_MUTEX(kvm_hyp_pgd_mutex);
  29
  30static unsigned long hyp_idmap_start;
  31static unsigned long hyp_idmap_end;
  32static phys_addr_t hyp_idmap_vector;
  33
  34static unsigned long io_map_base;
  35
  36#define hyp_pgd_order get_order(PTRS_PER_PGD * sizeof(pgd_t))
  37
  38#define KVM_S2PTE_FLAG_IS_IOMAP         (1UL << 0)
  39#define KVM_S2_FLAG_LOGGING_ACTIVE      (1UL << 1)
  40
  41static bool memslot_is_logging(struct kvm_memory_slot *memslot)
  42{
  43        return memslot->dirty_bitmap && !(memslot->flags & KVM_MEM_READONLY);
  44}
  45
  46/**
  47 * kvm_flush_remote_tlbs() - flush all VM TLB entries for v7/8
  48 * @kvm:        pointer to kvm structure.
  49 *
  50 * Interface to HYP function to flush all VM TLB entries
  51 */
  52void kvm_flush_remote_tlbs(struct kvm *kvm)
  53{
  54        kvm_call_hyp(__kvm_tlb_flush_vmid, kvm);
  55}
  56
  57static void kvm_tlb_flush_vmid_ipa(struct kvm *kvm, phys_addr_t ipa)
  58{
  59        kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, kvm, ipa);
  60}
  61
  62/*
  63 * D-Cache management functions. They take the page table entries by
  64 * value, as they are flushing the cache using the kernel mapping (or
  65 * kmap on 32bit).
  66 */
  67static void kvm_flush_dcache_pte(pte_t pte)
  68{
  69        __kvm_flush_dcache_pte(pte);
  70}
  71
  72static void kvm_flush_dcache_pmd(pmd_t pmd)
  73{
  74        __kvm_flush_dcache_pmd(pmd);
  75}
  76
  77static void kvm_flush_dcache_pud(pud_t pud)
  78{
  79        __kvm_flush_dcache_pud(pud);
  80}
  81
  82static bool kvm_is_device_pfn(unsigned long pfn)
  83{
  84        return !pfn_valid(pfn);
  85}
  86
  87/**
  88 * stage2_dissolve_pmd() - clear and flush huge PMD entry
  89 * @kvm:        pointer to kvm structure.
  90 * @addr:       IPA
  91 * @pmd:        pmd pointer for IPA
  92 *
  93 * Function clears a PMD entry, flushes addr 1st and 2nd stage TLBs.
  94 */
  95static void stage2_dissolve_pmd(struct kvm *kvm, phys_addr_t addr, pmd_t *pmd)
  96{
  97        if (!pmd_thp_or_huge(*pmd))
  98                return;
  99
 100        pmd_clear(pmd);
 101        kvm_tlb_flush_vmid_ipa(kvm, addr);
 102        put_page(virt_to_page(pmd));
 103}
 104
 105/**
 106 * stage2_dissolve_pud() - clear and flush huge PUD entry
 107 * @kvm:        pointer to kvm structure.
 108 * @addr:       IPA
 109 * @pud:        pud pointer for IPA
 110 *
 111 * Function clears a PUD entry, flushes addr 1st and 2nd stage TLBs.
 112 */
 113static void stage2_dissolve_pud(struct kvm *kvm, phys_addr_t addr, pud_t *pudp)
 114{
 115        if (!stage2_pud_huge(kvm, *pudp))
 116                return;
 117
 118        stage2_pud_clear(kvm, pudp);
 119        kvm_tlb_flush_vmid_ipa(kvm, addr);
 120        put_page(virt_to_page(pudp));
 121}
 122
 123static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 124                                  int min, int max)
 125{
 126        void *page;
 127
 128        BUG_ON(max > KVM_NR_MEM_OBJS);
 129        if (cache->nobjs >= min)
 130                return 0;
 131        while (cache->nobjs < max) {
 132                page = (void *)__get_free_page(GFP_PGTABLE_USER);
 133                if (!page)
 134                        return -ENOMEM;
 135                cache->objects[cache->nobjs++] = page;
 136        }
 137        return 0;
 138}
 139
 140static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 141{
 142        while (mc->nobjs)
 143                free_page((unsigned long)mc->objects[--mc->nobjs]);
 144}
 145
 146static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
 147{
 148        void *p;
 149
 150        BUG_ON(!mc || !mc->nobjs);
 151        p = mc->objects[--mc->nobjs];
 152        return p;
 153}
 154
 155static void clear_stage2_pgd_entry(struct kvm *kvm, pgd_t *pgd, phys_addr_t addr)
 156{
 157        pud_t *pud_table __maybe_unused = stage2_pud_offset(kvm, pgd, 0UL);
 158        stage2_pgd_clear(kvm, pgd);
 159        kvm_tlb_flush_vmid_ipa(kvm, addr);
 160        stage2_pud_free(kvm, pud_table);
 161        put_page(virt_to_page(pgd));
 162}
 163
 164static void clear_stage2_pud_entry(struct kvm *kvm, pud_t *pud, phys_addr_t addr)
 165{
 166        pmd_t *pmd_table __maybe_unused = stage2_pmd_offset(kvm, pud, 0);
 167        VM_BUG_ON(stage2_pud_huge(kvm, *pud));
 168        stage2_pud_clear(kvm, pud);
 169        kvm_tlb_flush_vmid_ipa(kvm, addr);
 170        stage2_pmd_free(kvm, pmd_table);
 171        put_page(virt_to_page(pud));
 172}
 173
 174static void clear_stage2_pmd_entry(struct kvm *kvm, pmd_t *pmd, phys_addr_t addr)
 175{
 176        pte_t *pte_table = pte_offset_kernel(pmd, 0);
 177        VM_BUG_ON(pmd_thp_or_huge(*pmd));
 178        pmd_clear(pmd);
 179        kvm_tlb_flush_vmid_ipa(kvm, addr);
 180        free_page((unsigned long)pte_table);
 181        put_page(virt_to_page(pmd));
 182}
 183
 184static inline void kvm_set_pte(pte_t *ptep, pte_t new_pte)
 185{
 186        WRITE_ONCE(*ptep, new_pte);
 187        dsb(ishst);
 188}
 189
 190static inline void kvm_set_pmd(pmd_t *pmdp, pmd_t new_pmd)
 191{
 192        WRITE_ONCE(*pmdp, new_pmd);
 193        dsb(ishst);
 194}
 195
 196static inline void kvm_pmd_populate(pmd_t *pmdp, pte_t *ptep)
 197{
 198        kvm_set_pmd(pmdp, kvm_mk_pmd(ptep));
 199}
 200
 201static inline void kvm_pud_populate(pud_t *pudp, pmd_t *pmdp)
 202{
 203        WRITE_ONCE(*pudp, kvm_mk_pud(pmdp));
 204        dsb(ishst);
 205}
 206
 207static inline void kvm_pgd_populate(pgd_t *pgdp, pud_t *pudp)
 208{
 209        WRITE_ONCE(*pgdp, kvm_mk_pgd(pudp));
 210        dsb(ishst);
 211}
 212
 213/*
 214 * Unmapping vs dcache management:
 215 *
 216 * If a guest maps certain memory pages as uncached, all writes will
 217 * bypass the data cache and go directly to RAM.  However, the CPUs
 218 * can still speculate reads (not writes) and fill cache lines with
 219 * data.
 220 *
 221 * Those cache lines will be *clean* cache lines though, so a
 222 * clean+invalidate operation is equivalent to an invalidate
 223 * operation, because no cache lines are marked dirty.
 224 *
 225 * Those clean cache lines could be filled prior to an uncached write
 226 * by the guest, and the cache coherent IO subsystem would therefore
 227 * end up writing old data to disk.
 228 *
 229 * This is why right after unmapping a page/section and invalidating
 230 * the corresponding TLBs, we call kvm_flush_dcache_p*() to make sure
 231 * the IO subsystem will never hit in the cache.
 232 *
 233 * This is all avoided on systems that have ARM64_HAS_STAGE2_FWB, as
 234 * we then fully enforce cacheability of RAM, no matter what the guest
 235 * does.
 236 */
 237static void unmap_stage2_ptes(struct kvm *kvm, pmd_t *pmd,
 238                       phys_addr_t addr, phys_addr_t end)
 239{
 240        phys_addr_t start_addr = addr;
 241        pte_t *pte, *start_pte;
 242
 243        start_pte = pte = pte_offset_kernel(pmd, addr);
 244        do {
 245                if (!pte_none(*pte)) {
 246                        pte_t old_pte = *pte;
 247
 248                        kvm_set_pte(pte, __pte(0));
 249                        kvm_tlb_flush_vmid_ipa(kvm, addr);
 250
 251                        /* No need to invalidate the cache for device mappings */
 252                        if (!kvm_is_device_pfn(pte_pfn(old_pte)))
 253                                kvm_flush_dcache_pte(old_pte);
 254
 255                        put_page(virt_to_page(pte));
 256                }
 257        } while (pte++, addr += PAGE_SIZE, addr != end);
 258
 259        if (stage2_pte_table_empty(kvm, start_pte))
 260                clear_stage2_pmd_entry(kvm, pmd, start_addr);
 261}
 262
 263static void unmap_stage2_pmds(struct kvm *kvm, pud_t *pud,
 264                       phys_addr_t addr, phys_addr_t end)
 265{
 266        phys_addr_t next, start_addr = addr;
 267        pmd_t *pmd, *start_pmd;
 268
 269        start_pmd = pmd = stage2_pmd_offset(kvm, pud, addr);
 270        do {
 271                next = stage2_pmd_addr_end(kvm, addr, end);
 272                if (!pmd_none(*pmd)) {
 273                        if (pmd_thp_or_huge(*pmd)) {
 274                                pmd_t old_pmd = *pmd;
 275
 276                                pmd_clear(pmd);
 277                                kvm_tlb_flush_vmid_ipa(kvm, addr);
 278
 279                                kvm_flush_dcache_pmd(old_pmd);
 280
 281                                put_page(virt_to_page(pmd));
 282                        } else {
 283                                unmap_stage2_ptes(kvm, pmd, addr, next);
 284                        }
 285                }
 286        } while (pmd++, addr = next, addr != end);
 287
 288        if (stage2_pmd_table_empty(kvm, start_pmd))
 289                clear_stage2_pud_entry(kvm, pud, start_addr);
 290}
 291
 292static void unmap_stage2_puds(struct kvm *kvm, pgd_t *pgd,
 293                       phys_addr_t addr, phys_addr_t end)
 294{
 295        phys_addr_t next, start_addr = addr;
 296        pud_t *pud, *start_pud;
 297
 298        start_pud = pud = stage2_pud_offset(kvm, pgd, addr);
 299        do {
 300                next = stage2_pud_addr_end(kvm, addr, end);
 301                if (!stage2_pud_none(kvm, *pud)) {
 302                        if (stage2_pud_huge(kvm, *pud)) {
 303                                pud_t old_pud = *pud;
 304
 305                                stage2_pud_clear(kvm, pud);
 306                                kvm_tlb_flush_vmid_ipa(kvm, addr);
 307                                kvm_flush_dcache_pud(old_pud);
 308                                put_page(virt_to_page(pud));
 309                        } else {
 310                                unmap_stage2_pmds(kvm, pud, addr, next);
 311                        }
 312                }
 313        } while (pud++, addr = next, addr != end);
 314
 315        if (stage2_pud_table_empty(kvm, start_pud))
 316                clear_stage2_pgd_entry(kvm, pgd, start_addr);
 317}
 318
 319/**
 320 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
 321 * @kvm:   The VM pointer
 322 * @start: The intermediate physical base address of the range to unmap
 323 * @size:  The size of the area to unmap
 324 *
 325 * Clear a range of stage-2 mappings, lowering the various ref-counts.  Must
 326 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
 327 * destroying the VM), otherwise another faulting VCPU may come in and mess
 328 * with things behind our backs.
 329 */
 330static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
 331{
 332        pgd_t *pgd;
 333        phys_addr_t addr = start, end = start + size;
 334        phys_addr_t next;
 335
 336        assert_spin_locked(&kvm->mmu_lock);
 337        WARN_ON(size & ~PAGE_MASK);
 338
 339        pgd = kvm->arch.pgd + stage2_pgd_index(kvm, addr);
 340        do {
 341                /*
 342                 * Make sure the page table is still active, as another thread
 343                 * could have possibly freed the page table, while we released
 344                 * the lock.
 345                 */
 346                if (!READ_ONCE(kvm->arch.pgd))
 347                        break;
 348                next = stage2_pgd_addr_end(kvm, addr, end);
 349                if (!stage2_pgd_none(kvm, *pgd))
 350                        unmap_stage2_puds(kvm, pgd, addr, next);
 351                /*
 352                 * If the range is too large, release the kvm->mmu_lock
 353                 * to prevent starvation and lockup detector warnings.
 354                 */
 355                if (next != end)
 356                        cond_resched_lock(&kvm->mmu_lock);
 357        } while (pgd++, addr = next, addr != end);
 358}
 359
 360static void stage2_flush_ptes(struct kvm *kvm, pmd_t *pmd,
 361                              phys_addr_t addr, phys_addr_t end)
 362{
 363        pte_t *pte;
 364
 365        pte = pte_offset_kernel(pmd, addr);
 366        do {
 367                if (!pte_none(*pte) && !kvm_is_device_pfn(pte_pfn(*pte)))
 368                        kvm_flush_dcache_pte(*pte);
 369        } while (pte++, addr += PAGE_SIZE, addr != end);
 370}
 371
 372static void stage2_flush_pmds(struct kvm *kvm, pud_t *pud,
 373                              phys_addr_t addr, phys_addr_t end)
 374{
 375        pmd_t *pmd;
 376        phys_addr_t next;
 377
 378        pmd = stage2_pmd_offset(kvm, pud, addr);
 379        do {
 380                next = stage2_pmd_addr_end(kvm, addr, end);
 381                if (!pmd_none(*pmd)) {
 382                        if (pmd_thp_or_huge(*pmd))
 383                                kvm_flush_dcache_pmd(*pmd);
 384                        else
 385                                stage2_flush_ptes(kvm, pmd, addr, next);
 386                }
 387        } while (pmd++, addr = next, addr != end);
 388}
 389
 390static void stage2_flush_puds(struct kvm *kvm, pgd_t *pgd,
 391                              phys_addr_t addr, phys_addr_t end)
 392{
 393        pud_t *pud;
 394        phys_addr_t next;
 395
 396        pud = stage2_pud_offset(kvm, pgd, addr);
 397        do {
 398                next = stage2_pud_addr_end(kvm, addr, end);
 399                if (!stage2_pud_none(kvm, *pud)) {
 400                        if (stage2_pud_huge(kvm, *pud))
 401                                kvm_flush_dcache_pud(*pud);
 402                        else
 403                                stage2_flush_pmds(kvm, pud, addr, next);
 404                }
 405        } while (pud++, addr = next, addr != end);
 406}
 407
 408static void stage2_flush_memslot(struct kvm *kvm,
 409                                 struct kvm_memory_slot *memslot)
 410{
 411        phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
 412        phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
 413        phys_addr_t next;
 414        pgd_t *pgd;
 415
 416        pgd = kvm->arch.pgd + stage2_pgd_index(kvm, addr);
 417        do {
 418                next = stage2_pgd_addr_end(kvm, addr, end);
 419                if (!stage2_pgd_none(kvm, *pgd))
 420                        stage2_flush_puds(kvm, pgd, addr, next);
 421        } while (pgd++, addr = next, addr != end);
 422}
 423
 424/**
 425 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
 426 * @kvm: The struct kvm pointer
 427 *
 428 * Go through the stage 2 page tables and invalidate any cache lines
 429 * backing memory already mapped to the VM.
 430 */
 431static void stage2_flush_vm(struct kvm *kvm)
 432{
 433        struct kvm_memslots *slots;
 434        struct kvm_memory_slot *memslot;
 435        int idx;
 436
 437        idx = srcu_read_lock(&kvm->srcu);
 438        spin_lock(&kvm->mmu_lock);
 439
 440        slots = kvm_memslots(kvm);
 441        kvm_for_each_memslot(memslot, slots)
 442                stage2_flush_memslot(kvm, memslot);
 443
 444        spin_unlock(&kvm->mmu_lock);
 445        srcu_read_unlock(&kvm->srcu, idx);
 446}
 447
 448static void clear_hyp_pgd_entry(pgd_t *pgd)
 449{
 450        pud_t *pud_table __maybe_unused = pud_offset(pgd, 0UL);
 451        pgd_clear(pgd);
 452        pud_free(NULL, pud_table);
 453        put_page(virt_to_page(pgd));
 454}
 455
 456static void clear_hyp_pud_entry(pud_t *pud)
 457{
 458        pmd_t *pmd_table __maybe_unused = pmd_offset(pud, 0);
 459        VM_BUG_ON(pud_huge(*pud));
 460        pud_clear(pud);
 461        pmd_free(NULL, pmd_table);
 462        put_page(virt_to_page(pud));
 463}
 464
 465static void clear_hyp_pmd_entry(pmd_t *pmd)
 466{
 467        pte_t *pte_table = pte_offset_kernel(pmd, 0);
 468        VM_BUG_ON(pmd_thp_or_huge(*pmd));
 469        pmd_clear(pmd);
 470        pte_free_kernel(NULL, pte_table);
 471        put_page(virt_to_page(pmd));
 472}
 473
 474static void unmap_hyp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
 475{
 476        pte_t *pte, *start_pte;
 477
 478        start_pte = pte = pte_offset_kernel(pmd, addr);
 479        do {
 480                if (!pte_none(*pte)) {
 481                        kvm_set_pte(pte, __pte(0));
 482                        put_page(virt_to_page(pte));
 483                }
 484        } while (pte++, addr += PAGE_SIZE, addr != end);
 485
 486        if (hyp_pte_table_empty(start_pte))
 487                clear_hyp_pmd_entry(pmd);
 488}
 489
 490static void unmap_hyp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
 491{
 492        phys_addr_t next;
 493        pmd_t *pmd, *start_pmd;
 494
 495        start_pmd = pmd = pmd_offset(pud, addr);
 496        do {
 497                next = pmd_addr_end(addr, end);
 498                /* Hyp doesn't use huge pmds */
 499                if (!pmd_none(*pmd))
 500                        unmap_hyp_ptes(pmd, addr, next);
 501        } while (pmd++, addr = next, addr != end);
 502
 503        if (hyp_pmd_table_empty(start_pmd))
 504                clear_hyp_pud_entry(pud);
 505}
 506
 507static void unmap_hyp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
 508{
 509        phys_addr_t next;
 510        pud_t *pud, *start_pud;
 511
 512        start_pud = pud = pud_offset(pgd, addr);
 513        do {
 514                next = pud_addr_end(addr, end);
 515                /* Hyp doesn't use huge puds */
 516                if (!pud_none(*pud))
 517                        unmap_hyp_pmds(pud, addr, next);
 518        } while (pud++, addr = next, addr != end);
 519
 520        if (hyp_pud_table_empty(start_pud))
 521                clear_hyp_pgd_entry(pgd);
 522}
 523
 524static unsigned int kvm_pgd_index(unsigned long addr, unsigned int ptrs_per_pgd)
 525{
 526        return (addr >> PGDIR_SHIFT) & (ptrs_per_pgd - 1);
 527}
 528
 529static void __unmap_hyp_range(pgd_t *pgdp, unsigned long ptrs_per_pgd,
 530                              phys_addr_t start, u64 size)
 531{
 532        pgd_t *pgd;
 533        phys_addr_t addr = start, end = start + size;
 534        phys_addr_t next;
 535
 536        /*
 537         * We don't unmap anything from HYP, except at the hyp tear down.
 538         * Hence, we don't have to invalidate the TLBs here.
 539         */
 540        pgd = pgdp + kvm_pgd_index(addr, ptrs_per_pgd);
 541        do {
 542                next = pgd_addr_end(addr, end);
 543                if (!pgd_none(*pgd))
 544                        unmap_hyp_puds(pgd, addr, next);
 545        } while (pgd++, addr = next, addr != end);
 546}
 547
 548static void unmap_hyp_range(pgd_t *pgdp, phys_addr_t start, u64 size)
 549{
 550        __unmap_hyp_range(pgdp, PTRS_PER_PGD, start, size);
 551}
 552
 553static void unmap_hyp_idmap_range(pgd_t *pgdp, phys_addr_t start, u64 size)
 554{
 555        __unmap_hyp_range(pgdp, __kvm_idmap_ptrs_per_pgd(), start, size);
 556}
 557
 558/**
 559 * free_hyp_pgds - free Hyp-mode page tables
 560 *
 561 * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
 562 * therefore contains either mappings in the kernel memory area (above
 563 * PAGE_OFFSET), or device mappings in the idmap range.
 564 *
 565 * boot_hyp_pgd should only map the idmap range, and is only used in
 566 * the extended idmap case.
 567 */
 568void free_hyp_pgds(void)
 569{
 570        pgd_t *id_pgd;
 571
 572        mutex_lock(&kvm_hyp_pgd_mutex);
 573
 574        id_pgd = boot_hyp_pgd ? boot_hyp_pgd : hyp_pgd;
 575
 576        if (id_pgd) {
 577                /* In case we never called hyp_mmu_init() */
 578                if (!io_map_base)
 579                        io_map_base = hyp_idmap_start;
 580                unmap_hyp_idmap_range(id_pgd, io_map_base,
 581                                      hyp_idmap_start + PAGE_SIZE - io_map_base);
 582        }
 583
 584        if (boot_hyp_pgd) {
 585                free_pages((unsigned long)boot_hyp_pgd, hyp_pgd_order);
 586                boot_hyp_pgd = NULL;
 587        }
 588
 589        if (hyp_pgd) {
 590                unmap_hyp_range(hyp_pgd, kern_hyp_va(PAGE_OFFSET),
 591                                (uintptr_t)high_memory - PAGE_OFFSET);
 592
 593                free_pages((unsigned long)hyp_pgd, hyp_pgd_order);
 594                hyp_pgd = NULL;
 595        }
 596        if (merged_hyp_pgd) {
 597                clear_page(merged_hyp_pgd);
 598                free_page((unsigned long)merged_hyp_pgd);
 599                merged_hyp_pgd = NULL;
 600        }
 601
 602        mutex_unlock(&kvm_hyp_pgd_mutex);
 603}
 604
 605static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
 606                                    unsigned long end, unsigned long pfn,
 607                                    pgprot_t prot)
 608{
 609        pte_t *pte;
 610        unsigned long addr;
 611
 612        addr = start;
 613        do {
 614                pte = pte_offset_kernel(pmd, addr);
 615                kvm_set_pte(pte, kvm_pfn_pte(pfn, prot));
 616                get_page(virt_to_page(pte));
 617                pfn++;
 618        } while (addr += PAGE_SIZE, addr != end);
 619}
 620
 621static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
 622                                   unsigned long end, unsigned long pfn,
 623                                   pgprot_t prot)
 624{
 625        pmd_t *pmd;
 626        pte_t *pte;
 627        unsigned long addr, next;
 628
 629        addr = start;
 630        do {
 631                pmd = pmd_offset(pud, addr);
 632
 633                BUG_ON(pmd_sect(*pmd));
 634
 635                if (pmd_none(*pmd)) {
 636                        pte = pte_alloc_one_kernel(NULL);
 637                        if (!pte) {
 638                                kvm_err("Cannot allocate Hyp pte\n");
 639                                return -ENOMEM;
 640                        }
 641                        kvm_pmd_populate(pmd, pte);
 642                        get_page(virt_to_page(pmd));
 643                }
 644
 645                next = pmd_addr_end(addr, end);
 646
 647                create_hyp_pte_mappings(pmd, addr, next, pfn, prot);
 648                pfn += (next - addr) >> PAGE_SHIFT;
 649        } while (addr = next, addr != end);
 650
 651        return 0;
 652}
 653
 654static int create_hyp_pud_mappings(pgd_t *pgd, unsigned long start,
 655                                   unsigned long end, unsigned long pfn,
 656                                   pgprot_t prot)
 657{
 658        pud_t *pud;
 659        pmd_t *pmd;
 660        unsigned long addr, next;
 661        int ret;
 662
 663        addr = start;
 664        do {
 665                pud = pud_offset(pgd, addr);
 666
 667                if (pud_none_or_clear_bad(pud)) {
 668                        pmd = pmd_alloc_one(NULL, addr);
 669                        if (!pmd) {
 670                                kvm_err("Cannot allocate Hyp pmd\n");
 671                                return -ENOMEM;
 672                        }
 673                        kvm_pud_populate(pud, pmd);
 674                        get_page(virt_to_page(pud));
 675                }
 676
 677                next = pud_addr_end(addr, end);
 678                ret = create_hyp_pmd_mappings(pud, addr, next, pfn, prot);
 679                if (ret)
 680                        return ret;
 681                pfn += (next - addr) >> PAGE_SHIFT;
 682        } while (addr = next, addr != end);
 683
 684        return 0;
 685}
 686
 687static int __create_hyp_mappings(pgd_t *pgdp, unsigned long ptrs_per_pgd,
 688                                 unsigned long start, unsigned long end,
 689                                 unsigned long pfn, pgprot_t prot)
 690{
 691        pgd_t *pgd;
 692        pud_t *pud;
 693        unsigned long addr, next;
 694        int err = 0;
 695
 696        mutex_lock(&kvm_hyp_pgd_mutex);
 697        addr = start & PAGE_MASK;
 698        end = PAGE_ALIGN(end);
 699        do {
 700                pgd = pgdp + kvm_pgd_index(addr, ptrs_per_pgd);
 701
 702                if (pgd_none(*pgd)) {
 703                        pud = pud_alloc_one(NULL, addr);
 704                        if (!pud) {
 705                                kvm_err("Cannot allocate Hyp pud\n");
 706                                err = -ENOMEM;
 707                                goto out;
 708                        }
 709                        kvm_pgd_populate(pgd, pud);
 710                        get_page(virt_to_page(pgd));
 711                }
 712
 713                next = pgd_addr_end(addr, end);
 714                err = create_hyp_pud_mappings(pgd, addr, next, pfn, prot);
 715                if (err)
 716                        goto out;
 717                pfn += (next - addr) >> PAGE_SHIFT;
 718        } while (addr = next, addr != end);
 719out:
 720        mutex_unlock(&kvm_hyp_pgd_mutex);
 721        return err;
 722}
 723
 724static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
 725{
 726        if (!is_vmalloc_addr(kaddr)) {
 727                BUG_ON(!virt_addr_valid(kaddr));
 728                return __pa(kaddr);
 729        } else {
 730                return page_to_phys(vmalloc_to_page(kaddr)) +
 731                       offset_in_page(kaddr);
 732        }
 733}
 734
 735/**
 736 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
 737 * @from:       The virtual kernel start address of the range
 738 * @to:         The virtual kernel end address of the range (exclusive)
 739 * @prot:       The protection to be applied to this range
 740 *
 741 * The same virtual address as the kernel virtual address is also used
 742 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
 743 * physical pages.
 744 */
 745int create_hyp_mappings(void *from, void *to, pgprot_t prot)
 746{
 747        phys_addr_t phys_addr;
 748        unsigned long virt_addr;
 749        unsigned long start = kern_hyp_va((unsigned long)from);
 750        unsigned long end = kern_hyp_va((unsigned long)to);
 751
 752        if (is_kernel_in_hyp_mode())
 753                return 0;
 754
 755        start = start & PAGE_MASK;
 756        end = PAGE_ALIGN(end);
 757
 758        for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
 759                int err;
 760
 761                phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
 762                err = __create_hyp_mappings(hyp_pgd, PTRS_PER_PGD,
 763                                            virt_addr, virt_addr + PAGE_SIZE,
 764                                            __phys_to_pfn(phys_addr),
 765                                            prot);
 766                if (err)
 767                        return err;
 768        }
 769
 770        return 0;
 771}
 772
 773static int __create_hyp_private_mapping(phys_addr_t phys_addr, size_t size,
 774                                        unsigned long *haddr, pgprot_t prot)
 775{
 776        pgd_t *pgd = hyp_pgd;
 777        unsigned long base;
 778        int ret = 0;
 779
 780        mutex_lock(&kvm_hyp_pgd_mutex);
 781
 782        /*
 783         * This assumes that we we have enough space below the idmap
 784         * page to allocate our VAs. If not, the check below will
 785         * kick. A potential alternative would be to detect that
 786         * overflow and switch to an allocation above the idmap.
 787         *
 788         * The allocated size is always a multiple of PAGE_SIZE.
 789         */
 790        size = PAGE_ALIGN(size + offset_in_page(phys_addr));
 791        base = io_map_base - size;
 792
 793        /*
 794         * Verify that BIT(VA_BITS - 1) hasn't been flipped by
 795         * allocating the new area, as it would indicate we've
 796         * overflowed the idmap/IO address range.
 797         */
 798        if ((base ^ io_map_base) & BIT(VA_BITS - 1))
 799                ret = -ENOMEM;
 800        else
 801                io_map_base = base;
 802
 803        mutex_unlock(&kvm_hyp_pgd_mutex);
 804
 805        if (ret)
 806                goto out;
 807
 808        if (__kvm_cpu_uses_extended_idmap())
 809                pgd = boot_hyp_pgd;
 810
 811        ret = __create_hyp_mappings(pgd, __kvm_idmap_ptrs_per_pgd(),
 812                                    base, base + size,
 813                                    __phys_to_pfn(phys_addr), prot);
 814        if (ret)
 815                goto out;
 816
 817        *haddr = base + offset_in_page(phys_addr);
 818
 819out:
 820        return ret;
 821}
 822
 823/**
 824 * create_hyp_io_mappings - Map IO into both kernel and HYP
 825 * @phys_addr:  The physical start address which gets mapped
 826 * @size:       Size of the region being mapped
 827 * @kaddr:      Kernel VA for this mapping
 828 * @haddr:      HYP VA for this mapping
 829 */
 830int create_hyp_io_mappings(phys_addr_t phys_addr, size_t size,
 831                           void __iomem **kaddr,
 832                           void __iomem **haddr)
 833{
 834        unsigned long addr;
 835        int ret;
 836
 837        *kaddr = ioremap(phys_addr, size);
 838        if (!*kaddr)
 839                return -ENOMEM;
 840
 841        if (is_kernel_in_hyp_mode()) {
 842                *haddr = *kaddr;
 843                return 0;
 844        }
 845
 846        ret = __create_hyp_private_mapping(phys_addr, size,
 847                                           &addr, PAGE_HYP_DEVICE);
 848        if (ret) {
 849                iounmap(*kaddr);
 850                *kaddr = NULL;
 851                *haddr = NULL;
 852                return ret;
 853        }
 854
 855        *haddr = (void __iomem *)addr;
 856        return 0;
 857}
 858
 859/**
 860 * create_hyp_exec_mappings - Map an executable range into HYP
 861 * @phys_addr:  The physical start address which gets mapped
 862 * @size:       Size of the region being mapped
 863 * @haddr:      HYP VA for this mapping
 864 */
 865int create_hyp_exec_mappings(phys_addr_t phys_addr, size_t size,
 866                             void **haddr)
 867{
 868        unsigned long addr;
 869        int ret;
 870
 871        BUG_ON(is_kernel_in_hyp_mode());
 872
 873        ret = __create_hyp_private_mapping(phys_addr, size,
 874                                           &addr, PAGE_HYP_EXEC);
 875        if (ret) {
 876                *haddr = NULL;
 877                return ret;
 878        }
 879
 880        *haddr = (void *)addr;
 881        return 0;
 882}
 883
 884/**
 885 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
 886 * @kvm:        The KVM struct pointer for the VM.
 887 *
 888 * Allocates only the stage-2 HW PGD level table(s) of size defined by
 889 * stage2_pgd_size(kvm).
 890 *
 891 * Note we don't need locking here as this is only called when the VM is
 892 * created, which can only be done once.
 893 */
 894int kvm_alloc_stage2_pgd(struct kvm *kvm)
 895{
 896        phys_addr_t pgd_phys;
 897        pgd_t *pgd;
 898
 899        if (kvm->arch.pgd != NULL) {
 900                kvm_err("kvm_arch already initialized?\n");
 901                return -EINVAL;
 902        }
 903
 904        /* Allocate the HW PGD, making sure that each page gets its own refcount */
 905        pgd = alloc_pages_exact(stage2_pgd_size(kvm), GFP_KERNEL | __GFP_ZERO);
 906        if (!pgd)
 907                return -ENOMEM;
 908
 909        pgd_phys = virt_to_phys(pgd);
 910        if (WARN_ON(pgd_phys & ~kvm_vttbr_baddr_mask(kvm)))
 911                return -EINVAL;
 912
 913        kvm->arch.pgd = pgd;
 914        kvm->arch.pgd_phys = pgd_phys;
 915        return 0;
 916}
 917
 918static void stage2_unmap_memslot(struct kvm *kvm,
 919                                 struct kvm_memory_slot *memslot)
 920{
 921        hva_t hva = memslot->userspace_addr;
 922        phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
 923        phys_addr_t size = PAGE_SIZE * memslot->npages;
 924        hva_t reg_end = hva + size;
 925
 926        /*
 927         * A memory region could potentially cover multiple VMAs, and any holes
 928         * between them, so iterate over all of them to find out if we should
 929         * unmap any of them.
 930         *
 931         *     +--------------------------------------------+
 932         * +---------------+----------------+   +----------------+
 933         * |   : VMA 1     |      VMA 2     |   |    VMA 3  :    |
 934         * +---------------+----------------+   +----------------+
 935         *     |               memory region                |
 936         *     +--------------------------------------------+
 937         */
 938        do {
 939                struct vm_area_struct *vma = find_vma(current->mm, hva);
 940                hva_t vm_start, vm_end;
 941
 942                if (!vma || vma->vm_start >= reg_end)
 943                        break;
 944
 945                /*
 946                 * Take the intersection of this VMA with the memory region
 947                 */
 948                vm_start = max(hva, vma->vm_start);
 949                vm_end = min(reg_end, vma->vm_end);
 950
 951                if (!(vma->vm_flags & VM_PFNMAP)) {
 952                        gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
 953                        unmap_stage2_range(kvm, gpa, vm_end - vm_start);
 954                }
 955                hva = vm_end;
 956        } while (hva < reg_end);
 957}
 958
 959/**
 960 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
 961 * @kvm: The struct kvm pointer
 962 *
 963 * Go through the memregions and unmap any reguler RAM
 964 * backing memory already mapped to the VM.
 965 */
 966void stage2_unmap_vm(struct kvm *kvm)
 967{
 968        struct kvm_memslots *slots;
 969        struct kvm_memory_slot *memslot;
 970        int idx;
 971
 972        idx = srcu_read_lock(&kvm->srcu);
 973        down_read(&current->mm->mmap_sem);
 974        spin_lock(&kvm->mmu_lock);
 975
 976        slots = kvm_memslots(kvm);
 977        kvm_for_each_memslot(memslot, slots)
 978                stage2_unmap_memslot(kvm, memslot);
 979
 980        spin_unlock(&kvm->mmu_lock);
 981        up_read(&current->mm->mmap_sem);
 982        srcu_read_unlock(&kvm->srcu, idx);
 983}
 984
 985/**
 986 * kvm_free_stage2_pgd - free all stage-2 tables
 987 * @kvm:        The KVM struct pointer for the VM.
 988 *
 989 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
 990 * underlying level-2 and level-3 tables before freeing the actual level-1 table
 991 * and setting the struct pointer to NULL.
 992 */
 993void kvm_free_stage2_pgd(struct kvm *kvm)
 994{
 995        void *pgd = NULL;
 996
 997        spin_lock(&kvm->mmu_lock);
 998        if (kvm->arch.pgd) {
 999                unmap_stage2_range(kvm, 0, kvm_phys_size(kvm));
1000                pgd = READ_ONCE(kvm->arch.pgd);
1001                kvm->arch.pgd = NULL;
1002                kvm->arch.pgd_phys = 0;
1003        }
1004        spin_unlock(&kvm->mmu_lock);
1005
1006        /* Free the HW pgd, one page at a time */
1007        if (pgd)
1008                free_pages_exact(pgd, stage2_pgd_size(kvm));
1009}
1010
1011static pud_t *stage2_get_pud(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
1012                             phys_addr_t addr)
1013{
1014        pgd_t *pgd;
1015        pud_t *pud;
1016
1017        pgd = kvm->arch.pgd + stage2_pgd_index(kvm, addr);
1018        if (stage2_pgd_none(kvm, *pgd)) {
1019                if (!cache)
1020                        return NULL;
1021                pud = mmu_memory_cache_alloc(cache);
1022                stage2_pgd_populate(kvm, pgd, pud);
1023                get_page(virt_to_page(pgd));
1024        }
1025
1026        return stage2_pud_offset(kvm, pgd, addr);
1027}
1028
1029static pmd_t *stage2_get_pmd(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
1030                             phys_addr_t addr)
1031{
1032        pud_t *pud;
1033        pmd_t *pmd;
1034
1035        pud = stage2_get_pud(kvm, cache, addr);
1036        if (!pud || stage2_pud_huge(kvm, *pud))
1037                return NULL;
1038
1039        if (stage2_pud_none(kvm, *pud)) {
1040                if (!cache)
1041                        return NULL;
1042                pmd = mmu_memory_cache_alloc(cache);
1043                stage2_pud_populate(kvm, pud, pmd);
1044                get_page(virt_to_page(pud));
1045        }
1046
1047        return stage2_pmd_offset(kvm, pud, addr);
1048}
1049
1050static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
1051                               *cache, phys_addr_t addr, const pmd_t *new_pmd)
1052{
1053        pmd_t *pmd, old_pmd;
1054
1055retry:
1056        pmd = stage2_get_pmd(kvm, cache, addr);
1057        VM_BUG_ON(!pmd);
1058
1059        old_pmd = *pmd;
1060        /*
1061         * Multiple vcpus faulting on the same PMD entry, can
1062         * lead to them sequentially updating the PMD with the
1063         * same value. Following the break-before-make
1064         * (pmd_clear() followed by tlb_flush()) process can
1065         * hinder forward progress due to refaults generated
1066         * on missing translations.
1067         *
1068         * Skip updating the page table if the entry is
1069         * unchanged.
1070         */
1071        if (pmd_val(old_pmd) == pmd_val(*new_pmd))
1072                return 0;
1073
1074        if (pmd_present(old_pmd)) {
1075                /*
1076                 * If we already have PTE level mapping for this block,
1077                 * we must unmap it to avoid inconsistent TLB state and
1078                 * leaking the table page. We could end up in this situation
1079                 * if the memory slot was marked for dirty logging and was
1080                 * reverted, leaving PTE level mappings for the pages accessed
1081                 * during the period. So, unmap the PTE level mapping for this
1082                 * block and retry, as we could have released the upper level
1083                 * table in the process.
1084                 *
1085                 * Normal THP split/merge follows mmu_notifier callbacks and do
1086                 * get handled accordingly.
1087                 */
1088                if (!pmd_thp_or_huge(old_pmd)) {
1089                        unmap_stage2_range(kvm, addr & S2_PMD_MASK, S2_PMD_SIZE);
1090                        goto retry;
1091                }
1092                /*
1093                 * Mapping in huge pages should only happen through a
1094                 * fault.  If a page is merged into a transparent huge
1095                 * page, the individual subpages of that huge page
1096                 * should be unmapped through MMU notifiers before we
1097                 * get here.
1098                 *
1099                 * Merging of CompoundPages is not supported; they
1100                 * should become splitting first, unmapped, merged,
1101                 * and mapped back in on-demand.
1102                 */
1103                WARN_ON_ONCE(pmd_pfn(old_pmd) != pmd_pfn(*new_pmd));
1104                pmd_clear(pmd);
1105                kvm_tlb_flush_vmid_ipa(kvm, addr);
1106        } else {
1107                get_page(virt_to_page(pmd));
1108        }
1109
1110        kvm_set_pmd(pmd, *new_pmd);
1111        return 0;
1112}
1113
1114static int stage2_set_pud_huge(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
1115                               phys_addr_t addr, const pud_t *new_pudp)
1116{
1117        pud_t *pudp, old_pud;
1118
1119retry:
1120        pudp = stage2_get_pud(kvm, cache, addr);
1121        VM_BUG_ON(!pudp);
1122
1123        old_pud = *pudp;
1124
1125        /*
1126         * A large number of vcpus faulting on the same stage 2 entry,
1127         * can lead to a refault due to the stage2_pud_clear()/tlb_flush().
1128         * Skip updating the page tables if there is no change.
1129         */
1130        if (pud_val(old_pud) == pud_val(*new_pudp))
1131                return 0;
1132
1133        if (stage2_pud_present(kvm, old_pud)) {
1134                /*
1135                 * If we already have table level mapping for this block, unmap
1136                 * the range for this block and retry.
1137                 */
1138                if (!stage2_pud_huge(kvm, old_pud)) {
1139                        unmap_stage2_range(kvm, addr & S2_PUD_MASK, S2_PUD_SIZE);
1140                        goto retry;
1141                }
1142
1143                WARN_ON_ONCE(kvm_pud_pfn(old_pud) != kvm_pud_pfn(*new_pudp));
1144                stage2_pud_clear(kvm, pudp);
1145                kvm_tlb_flush_vmid_ipa(kvm, addr);
1146        } else {
1147                get_page(virt_to_page(pudp));
1148        }
1149
1150        kvm_set_pud(pudp, *new_pudp);
1151        return 0;
1152}
1153
1154/*
1155 * stage2_get_leaf_entry - walk the stage2 VM page tables and return
1156 * true if a valid and present leaf-entry is found. A pointer to the
1157 * leaf-entry is returned in the appropriate level variable - pudpp,
1158 * pmdpp, ptepp.
1159 */
1160static bool stage2_get_leaf_entry(struct kvm *kvm, phys_addr_t addr,
1161                                  pud_t **pudpp, pmd_t **pmdpp, pte_t **ptepp)
1162{
1163        pud_t *pudp;
1164        pmd_t *pmdp;
1165        pte_t *ptep;
1166
1167        *pudpp = NULL;
1168        *pmdpp = NULL;
1169        *ptepp = NULL;
1170
1171        pudp = stage2_get_pud(kvm, NULL, addr);
1172        if (!pudp || stage2_pud_none(kvm, *pudp) || !stage2_pud_present(kvm, *pudp))
1173                return false;
1174
1175        if (stage2_pud_huge(kvm, *pudp)) {
1176                *pudpp = pudp;
1177                return true;
1178        }
1179
1180        pmdp = stage2_pmd_offset(kvm, pudp, addr);
1181        if (!pmdp || pmd_none(*pmdp) || !pmd_present(*pmdp))
1182                return false;
1183
1184        if (pmd_thp_or_huge(*pmdp)) {
1185                *pmdpp = pmdp;
1186                return true;
1187        }
1188
1189        ptep = pte_offset_kernel(pmdp, addr);
1190        if (!ptep || pte_none(*ptep) || !pte_present(*ptep))
1191                return false;
1192
1193        *ptepp = ptep;
1194        return true;
1195}
1196
1197static bool stage2_is_exec(struct kvm *kvm, phys_addr_t addr)
1198{
1199        pud_t *pudp;
1200        pmd_t *pmdp;
1201        pte_t *ptep;
1202        bool found;
1203
1204        found = stage2_get_leaf_entry(kvm, addr, &pudp, &pmdp, &ptep);
1205        if (!found)
1206                return false;
1207
1208        if (pudp)
1209                return kvm_s2pud_exec(pudp);
1210        else if (pmdp)
1211                return kvm_s2pmd_exec(pmdp);
1212        else
1213                return kvm_s2pte_exec(ptep);
1214}
1215
1216static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
1217                          phys_addr_t addr, const pte_t *new_pte,
1218                          unsigned long flags)
1219{
1220        pud_t *pud;
1221        pmd_t *pmd;
1222        pte_t *pte, old_pte;
1223        bool iomap = flags & KVM_S2PTE_FLAG_IS_IOMAP;
1224        bool logging_active = flags & KVM_S2_FLAG_LOGGING_ACTIVE;
1225
1226        VM_BUG_ON(logging_active && !cache);
1227
1228        /* Create stage-2 page table mapping - Levels 0 and 1 */
1229        pud = stage2_get_pud(kvm, cache, addr);
1230        if (!pud) {
1231                /*
1232                 * Ignore calls from kvm_set_spte_hva for unallocated
1233                 * address ranges.
1234                 */
1235                return 0;
1236        }
1237
1238        /*
1239         * While dirty page logging - dissolve huge PUD, then continue
1240         * on to allocate page.
1241         */
1242        if (logging_active)
1243                stage2_dissolve_pud(kvm, addr, pud);
1244
1245        if (stage2_pud_none(kvm, *pud)) {
1246                if (!cache)
1247                        return 0; /* ignore calls from kvm_set_spte_hva */
1248                pmd = mmu_memory_cache_alloc(cache);
1249                stage2_pud_populate(kvm, pud, pmd);
1250                get_page(virt_to_page(pud));
1251        }
1252
1253        pmd = stage2_pmd_offset(kvm, pud, addr);
1254        if (!pmd) {
1255                /*
1256                 * Ignore calls from kvm_set_spte_hva for unallocated
1257                 * address ranges.
1258                 */
1259                return 0;
1260        }
1261
1262        /*
1263         * While dirty page logging - dissolve huge PMD, then continue on to
1264         * allocate page.
1265         */
1266        if (logging_active)
1267                stage2_dissolve_pmd(kvm, addr, pmd);
1268
1269        /* Create stage-2 page mappings - Level 2 */
1270        if (pmd_none(*pmd)) {
1271                if (!cache)
1272                        return 0; /* ignore calls from kvm_set_spte_hva */
1273                pte = mmu_memory_cache_alloc(cache);
1274                kvm_pmd_populate(pmd, pte);
1275                get_page(virt_to_page(pmd));
1276        }
1277
1278        pte = pte_offset_kernel(pmd, addr);
1279
1280        if (iomap && pte_present(*pte))
1281                return -EFAULT;
1282
1283        /* Create 2nd stage page table mapping - Level 3 */
1284        old_pte = *pte;
1285        if (pte_present(old_pte)) {
1286                /* Skip page table update if there is no change */
1287                if (pte_val(old_pte) == pte_val(*new_pte))
1288                        return 0;
1289
1290                kvm_set_pte(pte, __pte(0));
1291                kvm_tlb_flush_vmid_ipa(kvm, addr);
1292        } else {
1293                get_page(virt_to_page(pte));
1294        }
1295
1296        kvm_set_pte(pte, *new_pte);
1297        return 0;
1298}
1299
1300#ifndef __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
1301static int stage2_ptep_test_and_clear_young(pte_t *pte)
1302{
1303        if (pte_young(*pte)) {
1304                *pte = pte_mkold(*pte);
1305                return 1;
1306        }
1307        return 0;
1308}
1309#else
1310static int stage2_ptep_test_and_clear_young(pte_t *pte)
1311{
1312        return __ptep_test_and_clear_young(pte);
1313}
1314#endif
1315
1316static int stage2_pmdp_test_and_clear_young(pmd_t *pmd)
1317{
1318        return stage2_ptep_test_and_clear_young((pte_t *)pmd);
1319}
1320
1321static int stage2_pudp_test_and_clear_young(pud_t *pud)
1322{
1323        return stage2_ptep_test_and_clear_young((pte_t *)pud);
1324}
1325
1326/**
1327 * kvm_phys_addr_ioremap - map a device range to guest IPA
1328 *
1329 * @kvm:        The KVM pointer
1330 * @guest_ipa:  The IPA at which to insert the mapping
1331 * @pa:         The physical address of the device
1332 * @size:       The size of the mapping
1333 */
1334int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
1335                          phys_addr_t pa, unsigned long size, bool writable)
1336{
1337        phys_addr_t addr, end;
1338        int ret = 0;
1339        unsigned long pfn;
1340        struct kvm_mmu_memory_cache cache = { 0, };
1341
1342        end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
1343        pfn = __phys_to_pfn(pa);
1344
1345        for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
1346                pte_t pte = kvm_pfn_pte(pfn, PAGE_S2_DEVICE);
1347
1348                if (writable)
1349                        pte = kvm_s2pte_mkwrite(pte);
1350
1351                ret = mmu_topup_memory_cache(&cache,
1352                                             kvm_mmu_cache_min_pages(kvm),
1353                                             KVM_NR_MEM_OBJS);
1354                if (ret)
1355                        goto out;
1356                spin_lock(&kvm->mmu_lock);
1357                ret = stage2_set_pte(kvm, &cache, addr, &pte,
1358                                                KVM_S2PTE_FLAG_IS_IOMAP);
1359                spin_unlock(&kvm->mmu_lock);
1360                if (ret)
1361                        goto out;
1362
1363                pfn++;
1364        }
1365
1366out:
1367        mmu_free_memory_cache(&cache);
1368        return ret;
1369}
1370
1371static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
1372{
1373        kvm_pfn_t pfn = *pfnp;
1374        gfn_t gfn = *ipap >> PAGE_SHIFT;
1375        struct page *page = pfn_to_page(pfn);
1376
1377        /*
1378         * PageTransCompoundMap() returns true for THP and
1379         * hugetlbfs. Make sure the adjustment is done only for THP
1380         * pages.
1381         */
1382        if (!PageHuge(page) && PageTransCompoundMap(page)) {
1383                unsigned long mask;
1384                /*
1385                 * The address we faulted on is backed by a transparent huge
1386                 * page.  However, because we map the compound huge page and
1387                 * not the individual tail page, we need to transfer the
1388                 * refcount to the head page.  We have to be careful that the
1389                 * THP doesn't start to split while we are adjusting the
1390                 * refcounts.
1391                 *
1392                 * We are sure this doesn't happen, because mmu_notifier_retry
1393                 * was successful and we are holding the mmu_lock, so if this
1394                 * THP is trying to split, it will be blocked in the mmu
1395                 * notifier before touching any of the pages, specifically
1396                 * before being able to call __split_huge_page_refcount().
1397                 *
1398                 * We can therefore safely transfer the refcount from PG_tail
1399                 * to PG_head and switch the pfn from a tail page to the head
1400                 * page accordingly.
1401                 */
1402                mask = PTRS_PER_PMD - 1;
1403                VM_BUG_ON((gfn & mask) != (pfn & mask));
1404                if (pfn & mask) {
1405                        *ipap &= PMD_MASK;
1406                        kvm_release_pfn_clean(pfn);
1407                        pfn &= ~mask;
1408                        kvm_get_pfn(pfn);
1409                        *pfnp = pfn;
1410                }
1411
1412                return true;
1413        }
1414
1415        return false;
1416}
1417
1418/**
1419 * stage2_wp_ptes - write protect PMD range
1420 * @pmd:        pointer to pmd entry
1421 * @addr:       range start address
1422 * @end:        range end address
1423 */
1424static void stage2_wp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
1425{
1426        pte_t *pte;
1427
1428        pte = pte_offset_kernel(pmd, addr);
1429        do {
1430                if (!pte_none(*pte)) {
1431                        if (!kvm_s2pte_readonly(pte))
1432                                kvm_set_s2pte_readonly(pte);
1433                }
1434        } while (pte++, addr += PAGE_SIZE, addr != end);
1435}
1436
1437/**
1438 * stage2_wp_pmds - write protect PUD range
1439 * kvm:         kvm instance for the VM
1440 * @pud:        pointer to pud entry
1441 * @addr:       range start address
1442 * @end:        range end address
1443 */
1444static void stage2_wp_pmds(struct kvm *kvm, pud_t *pud,
1445                           phys_addr_t addr, phys_addr_t end)
1446{
1447        pmd_t *pmd;
1448        phys_addr_t next;
1449
1450        pmd = stage2_pmd_offset(kvm, pud, addr);
1451
1452        do {
1453                next = stage2_pmd_addr_end(kvm, addr, end);
1454                if (!pmd_none(*pmd)) {
1455                        if (pmd_thp_or_huge(*pmd)) {
1456                                if (!kvm_s2pmd_readonly(pmd))
1457                                        kvm_set_s2pmd_readonly(pmd);
1458                        } else {
1459                                stage2_wp_ptes(pmd, addr, next);
1460                        }
1461                }
1462        } while (pmd++, addr = next, addr != end);
1463}
1464
1465/**
1466 * stage2_wp_puds - write protect PGD range
1467 * @pgd:        pointer to pgd entry
1468 * @addr:       range start address
1469 * @end:        range end address
1470 */
1471static void  stage2_wp_puds(struct kvm *kvm, pgd_t *pgd,
1472                            phys_addr_t addr, phys_addr_t end)
1473{
1474        pud_t *pud;
1475        phys_addr_t next;
1476
1477        pud = stage2_pud_offset(kvm, pgd, addr);
1478        do {
1479                next = stage2_pud_addr_end(kvm, addr, end);
1480                if (!stage2_pud_none(kvm, *pud)) {
1481                        if (stage2_pud_huge(kvm, *pud)) {
1482                                if (!kvm_s2pud_readonly(pud))
1483                                        kvm_set_s2pud_readonly(pud);
1484                        } else {
1485                                stage2_wp_pmds(kvm, pud, addr, next);
1486                        }
1487                }
1488        } while (pud++, addr = next, addr != end);
1489}
1490
1491/**
1492 * stage2_wp_range() - write protect stage2 memory region range
1493 * @kvm:        The KVM pointer
1494 * @addr:       Start address of range
1495 * @end:        End address of range
1496 */
1497static void stage2_wp_range(struct kvm *kvm, phys_addr_t addr, phys_addr_t end)
1498{
1499        pgd_t *pgd;
1500        phys_addr_t next;
1501
1502        pgd = kvm->arch.pgd + stage2_pgd_index(kvm, addr);
1503        do {
1504                /*
1505                 * Release kvm_mmu_lock periodically if the memory region is
1506                 * large. Otherwise, we may see kernel panics with
1507                 * CONFIG_DETECT_HUNG_TASK, CONFIG_LOCKUP_DETECTOR,
1508                 * CONFIG_LOCKDEP. Additionally, holding the lock too long
1509                 * will also starve other vCPUs. We have to also make sure
1510                 * that the page tables are not freed while we released
1511                 * the lock.
1512                 */
1513                cond_resched_lock(&kvm->mmu_lock);
1514                if (!READ_ONCE(kvm->arch.pgd))
1515                        break;
1516                next = stage2_pgd_addr_end(kvm, addr, end);
1517                if (stage2_pgd_present(kvm, *pgd))
1518                        stage2_wp_puds(kvm, pgd, addr, next);
1519        } while (pgd++, addr = next, addr != end);
1520}
1521
1522/**
1523 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1524 * @kvm:        The KVM pointer
1525 * @slot:       The memory slot to write protect
1526 *
1527 * Called to start logging dirty pages after memory region
1528 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1529 * all present PUD, PMD and PTEs are write protected in the memory region.
1530 * Afterwards read of dirty page log can be called.
1531 *
1532 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1533 * serializing operations for VM memory regions.
1534 */
1535void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1536{
1537        struct kvm_memslots *slots = kvm_memslots(kvm);
1538        struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1539        phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
1540        phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1541
1542        spin_lock(&kvm->mmu_lock);
1543        stage2_wp_range(kvm, start, end);
1544        spin_unlock(&kvm->mmu_lock);
1545        kvm_flush_remote_tlbs(kvm);
1546}
1547
1548/**
1549 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
1550 * @kvm:        The KVM pointer
1551 * @slot:       The memory slot associated with mask
1552 * @gfn_offset: The gfn offset in memory slot
1553 * @mask:       The mask of dirty pages at offset 'gfn_offset' in this memory
1554 *              slot to be write protected
1555 *
1556 * Walks bits set in mask write protects the associated pte's. Caller must
1557 * acquire kvm_mmu_lock.
1558 */
1559static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1560                struct kvm_memory_slot *slot,
1561                gfn_t gfn_offset, unsigned long mask)
1562{
1563        phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1564        phys_addr_t start = (base_gfn +  __ffs(mask)) << PAGE_SHIFT;
1565        phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1566
1567        stage2_wp_range(kvm, start, end);
1568}
1569
1570/*
1571 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1572 * dirty pages.
1573 *
1574 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1575 * enable dirty logging for them.
1576 */
1577void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1578                struct kvm_memory_slot *slot,
1579                gfn_t gfn_offset, unsigned long mask)
1580{
1581        kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1582}
1583
1584static void clean_dcache_guest_page(kvm_pfn_t pfn, unsigned long size)
1585{
1586        __clean_dcache_guest_page(pfn, size);
1587}
1588
1589static void invalidate_icache_guest_page(kvm_pfn_t pfn, unsigned long size)
1590{
1591        __invalidate_icache_guest_page(pfn, size);
1592}
1593
1594static void kvm_send_hwpoison_signal(unsigned long address,
1595                                     struct vm_area_struct *vma)
1596{
1597        short lsb;
1598
1599        if (is_vm_hugetlb_page(vma))
1600                lsb = huge_page_shift(hstate_vma(vma));
1601        else
1602                lsb = PAGE_SHIFT;
1603
1604        send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb, current);
1605}
1606
1607static bool fault_supports_stage2_huge_mapping(struct kvm_memory_slot *memslot,
1608                                               unsigned long hva,
1609                                               unsigned long map_size)
1610{
1611        gpa_t gpa_start;
1612        hva_t uaddr_start, uaddr_end;
1613        size_t size;
1614
1615        size = memslot->npages * PAGE_SIZE;
1616
1617        gpa_start = memslot->base_gfn << PAGE_SHIFT;
1618
1619        uaddr_start = memslot->userspace_addr;
1620        uaddr_end = uaddr_start + size;
1621
1622        /*
1623         * Pages belonging to memslots that don't have the same alignment
1624         * within a PMD/PUD for userspace and IPA cannot be mapped with stage-2
1625         * PMD/PUD entries, because we'll end up mapping the wrong pages.
1626         *
1627         * Consider a layout like the following:
1628         *
1629         *    memslot->userspace_addr:
1630         *    +-----+--------------------+--------------------+---+
1631         *    |abcde|fgh  Stage-1 block  |    Stage-1 block tv|xyz|
1632         *    +-----+--------------------+--------------------+---+
1633         *
1634         *    memslot->base_gfn << PAGE_SIZE:
1635         *      +---+--------------------+--------------------+-----+
1636         *      |abc|def  Stage-2 block  |    Stage-2 block   |tvxyz|
1637         *      +---+--------------------+--------------------+-----+
1638         *
1639         * If we create those stage-2 blocks, we'll end up with this incorrect
1640         * mapping:
1641         *   d -> f
1642         *   e -> g
1643         *   f -> h
1644         */
1645        if ((gpa_start & (map_size - 1)) != (uaddr_start & (map_size - 1)))
1646                return false;
1647
1648        /*
1649         * Next, let's make sure we're not trying to map anything not covered
1650         * by the memslot. This means we have to prohibit block size mappings
1651         * for the beginning and end of a non-block aligned and non-block sized
1652         * memory slot (illustrated by the head and tail parts of the
1653         * userspace view above containing pages 'abcde' and 'xyz',
1654         * respectively).
1655         *
1656         * Note that it doesn't matter if we do the check using the
1657         * userspace_addr or the base_gfn, as both are equally aligned (per
1658         * the check above) and equally sized.
1659         */
1660        return (hva & ~(map_size - 1)) >= uaddr_start &&
1661               (hva & ~(map_size - 1)) + map_size <= uaddr_end;
1662}
1663
1664static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1665                          struct kvm_memory_slot *memslot, unsigned long hva,
1666                          unsigned long fault_status)
1667{
1668        int ret;
1669        bool write_fault, writable, force_pte = false;
1670        bool exec_fault, needs_exec;
1671        unsigned long mmu_seq;
1672        gfn_t gfn = fault_ipa >> PAGE_SHIFT;
1673        struct kvm *kvm = vcpu->kvm;
1674        struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1675        struct vm_area_struct *vma;
1676        kvm_pfn_t pfn;
1677        pgprot_t mem_type = PAGE_S2;
1678        bool logging_active = memslot_is_logging(memslot);
1679        unsigned long vma_pagesize, flags = 0;
1680
1681        write_fault = kvm_is_write_fault(vcpu);
1682        exec_fault = kvm_vcpu_trap_is_iabt(vcpu);
1683        VM_BUG_ON(write_fault && exec_fault);
1684
1685        if (fault_status == FSC_PERM && !write_fault && !exec_fault) {
1686                kvm_err("Unexpected L2 read permission error\n");
1687                return -EFAULT;
1688        }
1689
1690        /* Let's check if we will get back a huge page backed by hugetlbfs */
1691        down_read(&current->mm->mmap_sem);
1692        vma = find_vma_intersection(current->mm, hva, hva + 1);
1693        if (unlikely(!vma)) {
1694                kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1695                up_read(&current->mm->mmap_sem);
1696                return -EFAULT;
1697        }
1698
1699        vma_pagesize = vma_kernel_pagesize(vma);
1700        if (logging_active ||
1701            !fault_supports_stage2_huge_mapping(memslot, hva, vma_pagesize)) {
1702                force_pte = true;
1703                vma_pagesize = PAGE_SIZE;
1704        }
1705
1706        /*
1707         * The stage2 has a minimum of 2 level table (For arm64 see
1708         * kvm_arm_setup_stage2()). Hence, we are guaranteed that we can
1709         * use PMD_SIZE huge mappings (even when the PMD is folded into PGD).
1710         * As for PUD huge maps, we must make sure that we have at least
1711         * 3 levels, i.e, PMD is not folded.
1712         */
1713        if (vma_pagesize == PMD_SIZE ||
1714            (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm)))
1715                gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT;
1716        up_read(&current->mm->mmap_sem);
1717
1718        /* We need minimum second+third level pages */
1719        ret = mmu_topup_memory_cache(memcache, kvm_mmu_cache_min_pages(kvm),
1720                                     KVM_NR_MEM_OBJS);
1721        if (ret)
1722                return ret;
1723
1724        mmu_seq = vcpu->kvm->mmu_notifier_seq;
1725        /*
1726         * Ensure the read of mmu_notifier_seq happens before we call
1727         * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1728         * the page we just got a reference to gets unmapped before we have a
1729         * chance to grab the mmu_lock, which ensure that if the page gets
1730         * unmapped afterwards, the call to kvm_unmap_hva will take it away
1731         * from us again properly. This smp_rmb() interacts with the smp_wmb()
1732         * in kvm_mmu_notifier_invalidate_<page|range_end>.
1733         */
1734        smp_rmb();
1735
1736        pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
1737        if (pfn == KVM_PFN_ERR_HWPOISON) {
1738                kvm_send_hwpoison_signal(hva, vma);
1739                return 0;
1740        }
1741        if (is_error_noslot_pfn(pfn))
1742                return -EFAULT;
1743
1744        if (kvm_is_device_pfn(pfn)) {
1745                mem_type = PAGE_S2_DEVICE;
1746                flags |= KVM_S2PTE_FLAG_IS_IOMAP;
1747        } else if (logging_active) {
1748                /*
1749                 * Faults on pages in a memslot with logging enabled
1750                 * should not be mapped with huge pages (it introduces churn
1751                 * and performance degradation), so force a pte mapping.
1752                 */
1753                flags |= KVM_S2_FLAG_LOGGING_ACTIVE;
1754
1755                /*
1756                 * Only actually map the page as writable if this was a write
1757                 * fault.
1758                 */
1759                if (!write_fault)
1760                        writable = false;
1761        }
1762
1763        spin_lock(&kvm->mmu_lock);
1764        if (mmu_notifier_retry(kvm, mmu_seq))
1765                goto out_unlock;
1766
1767        if (vma_pagesize == PAGE_SIZE && !force_pte) {
1768                /*
1769                 * Only PMD_SIZE transparent hugepages(THP) are
1770                 * currently supported. This code will need to be
1771                 * updated to support other THP sizes.
1772                 *
1773                 * Make sure the host VA and the guest IPA are sufficiently
1774                 * aligned and that the block is contained within the memslot.
1775                 */
1776                if (fault_supports_stage2_huge_mapping(memslot, hva, PMD_SIZE) &&
1777                    transparent_hugepage_adjust(&pfn, &fault_ipa))
1778                        vma_pagesize = PMD_SIZE;
1779        }
1780
1781        if (writable)
1782                kvm_set_pfn_dirty(pfn);
1783
1784        if (fault_status != FSC_PERM)
1785                clean_dcache_guest_page(pfn, vma_pagesize);
1786
1787        if (exec_fault)
1788                invalidate_icache_guest_page(pfn, vma_pagesize);
1789
1790        /*
1791         * If we took an execution fault we have made the
1792         * icache/dcache coherent above and should now let the s2
1793         * mapping be executable.
1794         *
1795         * Write faults (!exec_fault && FSC_PERM) are orthogonal to
1796         * execute permissions, and we preserve whatever we have.
1797         */
1798        needs_exec = exec_fault ||
1799                (fault_status == FSC_PERM && stage2_is_exec(kvm, fault_ipa));
1800
1801        if (vma_pagesize == PUD_SIZE) {
1802                pud_t new_pud = kvm_pfn_pud(pfn, mem_type);
1803
1804                new_pud = kvm_pud_mkhuge(new_pud);
1805                if (writable)
1806                        new_pud = kvm_s2pud_mkwrite(new_pud);
1807
1808                if (needs_exec)
1809                        new_pud = kvm_s2pud_mkexec(new_pud);
1810
1811                ret = stage2_set_pud_huge(kvm, memcache, fault_ipa, &new_pud);
1812        } else if (vma_pagesize == PMD_SIZE) {
1813                pmd_t new_pmd = kvm_pfn_pmd(pfn, mem_type);
1814
1815                new_pmd = kvm_pmd_mkhuge(new_pmd);
1816
1817                if (writable)
1818                        new_pmd = kvm_s2pmd_mkwrite(new_pmd);
1819
1820                if (needs_exec)
1821                        new_pmd = kvm_s2pmd_mkexec(new_pmd);
1822
1823                ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
1824        } else {
1825                pte_t new_pte = kvm_pfn_pte(pfn, mem_type);
1826
1827                if (writable) {
1828                        new_pte = kvm_s2pte_mkwrite(new_pte);
1829                        mark_page_dirty(kvm, gfn);
1830                }
1831
1832                if (needs_exec)
1833                        new_pte = kvm_s2pte_mkexec(new_pte);
1834
1835                ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte, flags);
1836        }
1837
1838out_unlock:
1839        spin_unlock(&kvm->mmu_lock);
1840        kvm_set_pfn_accessed(pfn);
1841        kvm_release_pfn_clean(pfn);
1842        return ret;
1843}
1844
1845/*
1846 * Resolve the access fault by making the page young again.
1847 * Note that because the faulting entry is guaranteed not to be
1848 * cached in the TLB, we don't need to invalidate anything.
1849 * Only the HW Access Flag updates are supported for Stage 2 (no DBM),
1850 * so there is no need for atomic (pte|pmd)_mkyoung operations.
1851 */
1852static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1853{
1854        pud_t *pud;
1855        pmd_t *pmd;
1856        pte_t *pte;
1857        kvm_pfn_t pfn;
1858        bool pfn_valid = false;
1859
1860        trace_kvm_access_fault(fault_ipa);
1861
1862        spin_lock(&vcpu->kvm->mmu_lock);
1863
1864        if (!stage2_get_leaf_entry(vcpu->kvm, fault_ipa, &pud, &pmd, &pte))
1865                goto out;
1866
1867        if (pud) {              /* HugeTLB */
1868                *pud = kvm_s2pud_mkyoung(*pud);
1869                pfn = kvm_pud_pfn(*pud);
1870                pfn_valid = true;
1871        } else  if (pmd) {      /* THP, HugeTLB */
1872                *pmd = pmd_mkyoung(*pmd);
1873                pfn = pmd_pfn(*pmd);
1874                pfn_valid = true;
1875        } else {
1876                *pte = pte_mkyoung(*pte);       /* Just a page... */
1877                pfn = pte_pfn(*pte);
1878                pfn_valid = true;
1879        }
1880
1881out:
1882        spin_unlock(&vcpu->kvm->mmu_lock);
1883        if (pfn_valid)
1884                kvm_set_pfn_accessed(pfn);
1885}
1886
1887/**
1888 * kvm_handle_guest_abort - handles all 2nd stage aborts
1889 * @vcpu:       the VCPU pointer
1890 * @run:        the kvm_run structure
1891 *
1892 * Any abort that gets to the host is almost guaranteed to be caused by a
1893 * missing second stage translation table entry, which can mean that either the
1894 * guest simply needs more memory and we must allocate an appropriate page or it
1895 * can mean that the guest tried to access I/O memory, which is emulated by user
1896 * space. The distinction is based on the IPA causing the fault and whether this
1897 * memory region has been registered as standard RAM by user space.
1898 */
1899int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
1900{
1901        unsigned long fault_status;
1902        phys_addr_t fault_ipa;
1903        struct kvm_memory_slot *memslot;
1904        unsigned long hva;
1905        bool is_iabt, write_fault, writable;
1906        gfn_t gfn;
1907        int ret, idx;
1908
1909        fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1910
1911        fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1912        is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1913
1914        /* Synchronous External Abort? */
1915        if (kvm_vcpu_dabt_isextabt(vcpu)) {
1916                /*
1917                 * For RAS the host kernel may handle this abort.
1918                 * There is no need to pass the error into the guest.
1919                 */
1920                if (!kvm_handle_guest_sea(fault_ipa, kvm_vcpu_get_hsr(vcpu)))
1921                        return 1;
1922
1923                if (unlikely(!is_iabt)) {
1924                        kvm_inject_vabt(vcpu);
1925                        return 1;
1926                }
1927        }
1928
1929        trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
1930                              kvm_vcpu_get_hfar(vcpu), fault_ipa);
1931
1932        /* Check the stage-2 fault is trans. fault or write fault */
1933        if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1934            fault_status != FSC_ACCESS) {
1935                kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1936                        kvm_vcpu_trap_get_class(vcpu),
1937                        (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1938                        (unsigned long)kvm_vcpu_get_hsr(vcpu));
1939                return -EFAULT;
1940        }
1941
1942        idx = srcu_read_lock(&vcpu->kvm->srcu);
1943
1944        gfn = fault_ipa >> PAGE_SHIFT;
1945        memslot = gfn_to_memslot(vcpu->kvm, gfn);
1946        hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1947        write_fault = kvm_is_write_fault(vcpu);
1948        if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1949                if (is_iabt) {
1950                        /* Prefetch Abort on I/O address */
1951                        kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1952                        ret = 1;
1953                        goto out_unlock;
1954                }
1955
1956                /*
1957                 * Check for a cache maintenance operation. Since we
1958                 * ended-up here, we know it is outside of any memory
1959                 * slot. But we can't find out if that is for a device,
1960                 * or if the guest is just being stupid. The only thing
1961                 * we know for sure is that this range cannot be cached.
1962                 *
1963                 * So let's assume that the guest is just being
1964                 * cautious, and skip the instruction.
1965                 */
1966                if (kvm_vcpu_dabt_is_cm(vcpu)) {
1967                        kvm_skip_instr(vcpu, kvm_vcpu_trap_il_is32bit(vcpu));
1968                        ret = 1;
1969                        goto out_unlock;
1970                }
1971
1972                /*
1973                 * The IPA is reported as [MAX:12], so we need to
1974                 * complement it with the bottom 12 bits from the
1975                 * faulting VA. This is always 12 bits, irrespective
1976                 * of the page size.
1977                 */
1978                fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1979                ret = io_mem_abort(vcpu, run, fault_ipa);
1980                goto out_unlock;
1981        }
1982
1983        /* Userspace should not be able to register out-of-bounds IPAs */
1984        VM_BUG_ON(fault_ipa >= kvm_phys_size(vcpu->kvm));
1985
1986        if (fault_status == FSC_ACCESS) {
1987                handle_access_fault(vcpu, fault_ipa);
1988                ret = 1;
1989                goto out_unlock;
1990        }
1991
1992        ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1993        if (ret == 0)
1994                ret = 1;
1995out_unlock:
1996        srcu_read_unlock(&vcpu->kvm->srcu, idx);
1997        return ret;
1998}
1999
2000static int handle_hva_to_gpa(struct kvm *kvm,
2001                             unsigned long start,
2002                             unsigned long end,
2003                             int (*handler)(struct kvm *kvm,
2004                                            gpa_t gpa, u64 size,
2005                                            void *data),
2006                             void *data)
2007{
2008        struct kvm_memslots *slots;
2009        struct kvm_memory_slot *memslot;
2010        int ret = 0;
2011
2012        slots = kvm_memslots(kvm);
2013
2014        /* we only care about the pages that the guest sees */
2015        kvm_for_each_memslot(memslot, slots) {
2016                unsigned long hva_start, hva_end;
2017                gfn_t gpa;
2018
2019                hva_start = max(start, memslot->userspace_addr);
2020                hva_end = min(end, memslot->userspace_addr +
2021                                        (memslot->npages << PAGE_SHIFT));
2022                if (hva_start >= hva_end)
2023                        continue;
2024
2025                gpa = hva_to_gfn_memslot(hva_start, memslot) << PAGE_SHIFT;
2026                ret |= handler(kvm, gpa, (u64)(hva_end - hva_start), data);
2027        }
2028
2029        return ret;
2030}
2031
2032static int kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
2033{
2034        unmap_stage2_range(kvm, gpa, size);
2035        return 0;
2036}
2037
2038int kvm_unmap_hva_range(struct kvm *kvm,
2039                        unsigned long start, unsigned long end)
2040{
2041        if (!kvm->arch.pgd)
2042                return 0;
2043
2044        trace_kvm_unmap_hva_range(start, end);
2045        handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
2046        return 0;
2047}
2048
2049static int kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
2050{
2051        pte_t *pte = (pte_t *)data;
2052
2053        WARN_ON(size != PAGE_SIZE);
2054        /*
2055         * We can always call stage2_set_pte with KVM_S2PTE_FLAG_LOGGING_ACTIVE
2056         * flag clear because MMU notifiers will have unmapped a huge PMD before
2057         * calling ->change_pte() (which in turn calls kvm_set_spte_hva()) and
2058         * therefore stage2_set_pte() never needs to clear out a huge PMD
2059         * through this calling path.
2060         */
2061        stage2_set_pte(kvm, NULL, gpa, pte, 0);
2062        return 0;
2063}
2064
2065
2066int kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
2067{
2068        unsigned long end = hva + PAGE_SIZE;
2069        kvm_pfn_t pfn = pte_pfn(pte);
2070        pte_t stage2_pte;
2071
2072        if (!kvm->arch.pgd)
2073                return 0;
2074
2075        trace_kvm_set_spte_hva(hva);
2076
2077        /*
2078         * We've moved a page around, probably through CoW, so let's treat it
2079         * just like a translation fault and clean the cache to the PoC.
2080         */
2081        clean_dcache_guest_page(pfn, PAGE_SIZE);
2082        stage2_pte = kvm_pfn_pte(pfn, PAGE_S2);
2083        handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
2084
2085        return 0;
2086}
2087
2088static int kvm_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
2089{
2090        pud_t *pud;
2091        pmd_t *pmd;
2092        pte_t *pte;
2093
2094        WARN_ON(size != PAGE_SIZE && size != PMD_SIZE && size != PUD_SIZE);
2095        if (!stage2_get_leaf_entry(kvm, gpa, &pud, &pmd, &pte))
2096                return 0;
2097
2098        if (pud)
2099                return stage2_pudp_test_and_clear_young(pud);
2100        else if (pmd)
2101                return stage2_pmdp_test_and_clear_young(pmd);
2102        else
2103                return stage2_ptep_test_and_clear_young(pte);
2104}
2105
2106static int kvm_test_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
2107{
2108        pud_t *pud;
2109        pmd_t *pmd;
2110        pte_t *pte;
2111
2112        WARN_ON(size != PAGE_SIZE && size != PMD_SIZE && size != PUD_SIZE);
2113        if (!stage2_get_leaf_entry(kvm, gpa, &pud, &pmd, &pte))
2114                return 0;
2115
2116        if (pud)
2117                return kvm_s2pud_young(*pud);
2118        else if (pmd)
2119                return pmd_young(*pmd);
2120        else
2121                return pte_young(*pte);
2122}
2123
2124int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
2125{
2126        if (!kvm->arch.pgd)
2127                return 0;
2128        trace_kvm_age_hva(start, end);
2129        return handle_hva_to_gpa(kvm, start, end, kvm_age_hva_handler, NULL);
2130}
2131
2132int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
2133{
2134        if (!kvm->arch.pgd)
2135                return 0;
2136        trace_kvm_test_age_hva(hva);
2137        return handle_hva_to_gpa(kvm, hva, hva, kvm_test_age_hva_handler, NULL);
2138}
2139
2140void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
2141{
2142        mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
2143}
2144
2145phys_addr_t kvm_mmu_get_httbr(void)
2146{
2147        if (__kvm_cpu_uses_extended_idmap())
2148                return virt_to_phys(merged_hyp_pgd);
2149        else
2150                return virt_to_phys(hyp_pgd);
2151}
2152
2153phys_addr_t kvm_get_idmap_vector(void)
2154{
2155        return hyp_idmap_vector;
2156}
2157
2158static int kvm_map_idmap_text(pgd_t *pgd)
2159{
2160        int err;
2161
2162        /* Create the idmap in the boot page tables */
2163        err =   __create_hyp_mappings(pgd, __kvm_idmap_ptrs_per_pgd(),
2164                                      hyp_idmap_start, hyp_idmap_end,
2165                                      __phys_to_pfn(hyp_idmap_start),
2166                                      PAGE_HYP_EXEC);
2167        if (err)
2168                kvm_err("Failed to idmap %lx-%lx\n",
2169                        hyp_idmap_start, hyp_idmap_end);
2170
2171        return err;
2172}
2173
2174int kvm_mmu_init(void)
2175{
2176        int err;
2177
2178        hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
2179        hyp_idmap_start = ALIGN_DOWN(hyp_idmap_start, PAGE_SIZE);
2180        hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
2181        hyp_idmap_end = ALIGN(hyp_idmap_end, PAGE_SIZE);
2182        hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
2183
2184        /*
2185         * We rely on the linker script to ensure at build time that the HYP
2186         * init code does not cross a page boundary.
2187         */
2188        BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
2189
2190        kvm_debug("IDMAP page: %lx\n", hyp_idmap_start);
2191        kvm_debug("HYP VA range: %lx:%lx\n",
2192                  kern_hyp_va(PAGE_OFFSET),
2193                  kern_hyp_va((unsigned long)high_memory - 1));
2194
2195        if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
2196            hyp_idmap_start <  kern_hyp_va((unsigned long)high_memory - 1) &&
2197            hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
2198                /*
2199                 * The idmap page is intersecting with the VA space,
2200                 * it is not safe to continue further.
2201                 */
2202                kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
2203                err = -EINVAL;
2204                goto out;
2205        }
2206
2207        hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
2208        if (!hyp_pgd) {
2209                kvm_err("Hyp mode PGD not allocated\n");
2210                err = -ENOMEM;
2211                goto out;
2212        }
2213
2214        if (__kvm_cpu_uses_extended_idmap()) {
2215                boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
2216                                                         hyp_pgd_order);
2217                if (!boot_hyp_pgd) {
2218                        kvm_err("Hyp boot PGD not allocated\n");
2219                        err = -ENOMEM;
2220                        goto out;
2221                }
2222
2223                err = kvm_map_idmap_text(boot_hyp_pgd);
2224                if (err)
2225                        goto out;
2226
2227                merged_hyp_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
2228                if (!merged_hyp_pgd) {
2229                        kvm_err("Failed to allocate extra HYP pgd\n");
2230                        goto out;
2231                }
2232                __kvm_extend_hypmap(boot_hyp_pgd, hyp_pgd, merged_hyp_pgd,
2233                                    hyp_idmap_start);
2234        } else {
2235                err = kvm_map_idmap_text(hyp_pgd);
2236                if (err)
2237                        goto out;
2238        }
2239
2240        io_map_base = hyp_idmap_start;
2241        return 0;
2242out:
2243        free_hyp_pgds();
2244        return err;
2245}
2246
2247void kvm_arch_commit_memory_region(struct kvm *kvm,
2248                                   const struct kvm_userspace_memory_region *mem,
2249                                   const struct kvm_memory_slot *old,
2250                                   const struct kvm_memory_slot *new,
2251                                   enum kvm_mr_change change)
2252{
2253        /*
2254         * At this point memslot has been committed and there is an
2255         * allocated dirty_bitmap[], dirty pages will be be tracked while the
2256         * memory slot is write protected.
2257         */
2258        if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
2259                kvm_mmu_wp_memory_region(kvm, mem->slot);
2260}
2261
2262int kvm_arch_prepare_memory_region(struct kvm *kvm,
2263                                   struct kvm_memory_slot *memslot,
2264                                   const struct kvm_userspace_memory_region *mem,
2265                                   enum kvm_mr_change change)
2266{
2267        hva_t hva = mem->userspace_addr;
2268        hva_t reg_end = hva + mem->memory_size;
2269        bool writable = !(mem->flags & KVM_MEM_READONLY);
2270        int ret = 0;
2271
2272        if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
2273                        change != KVM_MR_FLAGS_ONLY)
2274                return 0;
2275
2276        /*
2277         * Prevent userspace from creating a memory region outside of the IPA
2278         * space addressable by the KVM guest IPA space.
2279         */
2280        if (memslot->base_gfn + memslot->npages >=
2281            (kvm_phys_size(kvm) >> PAGE_SHIFT))
2282                return -EFAULT;
2283
2284        down_read(&current->mm->mmap_sem);
2285        /*
2286         * A memory region could potentially cover multiple VMAs, and any holes
2287         * between them, so iterate over all of them to find out if we can map
2288         * any of them right now.
2289         *
2290         *     +--------------------------------------------+
2291         * +---------------+----------------+   +----------------+
2292         * |   : VMA 1     |      VMA 2     |   |    VMA 3  :    |
2293         * +---------------+----------------+   +----------------+
2294         *     |               memory region                |
2295         *     +--------------------------------------------+
2296         */
2297        do {
2298                struct vm_area_struct *vma = find_vma(current->mm, hva);
2299                hva_t vm_start, vm_end;
2300
2301                if (!vma || vma->vm_start >= reg_end)
2302                        break;
2303
2304                /*
2305                 * Mapping a read-only VMA is only allowed if the
2306                 * memory region is configured as read-only.
2307                 */
2308                if (writable && !(vma->vm_flags & VM_WRITE)) {
2309                        ret = -EPERM;
2310                        break;
2311                }
2312
2313                /*
2314                 * Take the intersection of this VMA with the memory region
2315                 */
2316                vm_start = max(hva, vma->vm_start);
2317                vm_end = min(reg_end, vma->vm_end);
2318
2319                if (vma->vm_flags & VM_PFNMAP) {
2320                        gpa_t gpa = mem->guest_phys_addr +
2321                                    (vm_start - mem->userspace_addr);
2322                        phys_addr_t pa;
2323
2324                        pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
2325                        pa += vm_start - vma->vm_start;
2326
2327                        /* IO region dirty page logging not allowed */
2328                        if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES) {
2329                                ret = -EINVAL;
2330                                goto out;
2331                        }
2332
2333                        ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
2334                                                    vm_end - vm_start,
2335                                                    writable);
2336                        if (ret)
2337                                break;
2338                }
2339                hva = vm_end;
2340        } while (hva < reg_end);
2341
2342        if (change == KVM_MR_FLAGS_ONLY)
2343                goto out;
2344
2345        spin_lock(&kvm->mmu_lock);
2346        if (ret)
2347                unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
2348        else
2349                stage2_flush_memslot(kvm, memslot);
2350        spin_unlock(&kvm->mmu_lock);
2351out:
2352        up_read(&current->mm->mmap_sem);
2353        return ret;
2354}
2355
2356void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
2357                           struct kvm_memory_slot *dont)
2358{
2359}
2360
2361int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
2362                            unsigned long npages)
2363{
2364        return 0;
2365}
2366
2367void kvm_arch_memslots_updated(struct kvm *kvm, u64 gen)
2368{
2369}
2370
2371void kvm_arch_flush_shadow_all(struct kvm *kvm)
2372{
2373        kvm_free_stage2_pgd(kvm);
2374}
2375
2376void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
2377                                   struct kvm_memory_slot *slot)
2378{
2379        gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
2380        phys_addr_t size = slot->npages << PAGE_SHIFT;
2381
2382        spin_lock(&kvm->mmu_lock);
2383        unmap_stage2_range(kvm, gpa, size);
2384        spin_unlock(&kvm->mmu_lock);
2385}
2386
2387/*
2388 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
2389 *
2390 * Main problems:
2391 * - S/W ops are local to a CPU (not broadcast)
2392 * - We have line migration behind our back (speculation)
2393 * - System caches don't support S/W at all (damn!)
2394 *
2395 * In the face of the above, the best we can do is to try and convert
2396 * S/W ops to VA ops. Because the guest is not allowed to infer the
2397 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
2398 * which is a rather good thing for us.
2399 *
2400 * Also, it is only used when turning caches on/off ("The expected
2401 * usage of the cache maintenance instructions that operate by set/way
2402 * is associated with the cache maintenance instructions associated
2403 * with the powerdown and powerup of caches, if this is required by
2404 * the implementation.").
2405 *
2406 * We use the following policy:
2407 *
2408 * - If we trap a S/W operation, we enable VM trapping to detect
2409 *   caches being turned on/off, and do a full clean.
2410 *
2411 * - We flush the caches on both caches being turned on and off.
2412 *
2413 * - Once the caches are enabled, we stop trapping VM ops.
2414 */
2415void kvm_set_way_flush(struct kvm_vcpu *vcpu)
2416{
2417        unsigned long hcr = *vcpu_hcr(vcpu);
2418
2419        /*
2420         * If this is the first time we do a S/W operation
2421         * (i.e. HCR_TVM not set) flush the whole memory, and set the
2422         * VM trapping.
2423         *
2424         * Otherwise, rely on the VM trapping to wait for the MMU +
2425         * Caches to be turned off. At that point, we'll be able to
2426         * clean the caches again.
2427         */
2428        if (!(hcr & HCR_TVM)) {
2429                trace_kvm_set_way_flush(*vcpu_pc(vcpu),
2430                                        vcpu_has_cache_enabled(vcpu));
2431                stage2_flush_vm(vcpu->kvm);
2432                *vcpu_hcr(vcpu) = hcr | HCR_TVM;
2433        }
2434}
2435
2436void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
2437{
2438        bool now_enabled = vcpu_has_cache_enabled(vcpu);
2439
2440        /*
2441         * If switching the MMU+caches on, need to invalidate the caches.
2442         * If switching it off, need to clean the caches.
2443         * Clean + invalidate does the trick always.
2444         */
2445        if (now_enabled != was_enabled)
2446                stage2_flush_vm(vcpu->kvm);
2447
2448        /* Caches are now on, stop trapping VM ops (until a S/W op) */
2449        if (now_enabled)
2450                *vcpu_hcr(vcpu) &= ~HCR_TVM;
2451
2452        trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);
2453}
2454