Linux Audio

Check our new training course

Loading...
Note: File does not exist in v6.13.7.
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Kernel-based Virtual Machine driver for Linux
   4 *
   5 * This module enables machines with Intel VT-x extensions to run virtual
   6 * machines without emulation or binary translation.
   7 *
   8 * MMU support
   9 *
  10 * Copyright (C) 2006 Qumranet, Inc.
  11 * Copyright 2010 Red Hat, Inc. and/or its affiliates.
  12 *
  13 * Authors:
  14 *   Yaniv Kamay  <yaniv@qumranet.com>
  15 *   Avi Kivity   <avi@qumranet.com>
  16 */
  17
  18#include "irq.h"
  19#include "mmu.h"
  20#include "x86.h"
  21#include "kvm_cache_regs.h"
  22#include "cpuid.h"
  23
  24#include <linux/kvm_host.h>
  25#include <linux/types.h>
  26#include <linux/string.h>
  27#include <linux/mm.h>
  28#include <linux/highmem.h>
  29#include <linux/moduleparam.h>
  30#include <linux/export.h>
  31#include <linux/swap.h>
  32#include <linux/hugetlb.h>
  33#include <linux/compiler.h>
  34#include <linux/srcu.h>
  35#include <linux/slab.h>
  36#include <linux/sched/signal.h>
  37#include <linux/uaccess.h>
  38#include <linux/hash.h>
  39#include <linux/kern_levels.h>
  40#include <linux/kthread.h>
  41
  42#include <asm/page.h>
  43#include <asm/pat.h>
  44#include <asm/cmpxchg.h>
  45#include <asm/e820/api.h>
  46#include <asm/io.h>
  47#include <asm/vmx.h>
  48#include <asm/kvm_page_track.h>
  49#include "trace.h"
  50
  51extern bool itlb_multihit_kvm_mitigation;
  52
  53static int __read_mostly nx_huge_pages = -1;
  54#ifdef CONFIG_PREEMPT_RT
  55/* Recovery can cause latency spikes, disable it for PREEMPT_RT.  */
  56static uint __read_mostly nx_huge_pages_recovery_ratio = 0;
  57#else
  58static uint __read_mostly nx_huge_pages_recovery_ratio = 60;
  59#endif
  60
  61static int set_nx_huge_pages(const char *val, const struct kernel_param *kp);
  62static int set_nx_huge_pages_recovery_ratio(const char *val, const struct kernel_param *kp);
  63
  64static struct kernel_param_ops nx_huge_pages_ops = {
  65	.set = set_nx_huge_pages,
  66	.get = param_get_bool,
  67};
  68
  69static struct kernel_param_ops nx_huge_pages_recovery_ratio_ops = {
  70	.set = set_nx_huge_pages_recovery_ratio,
  71	.get = param_get_uint,
  72};
  73
  74module_param_cb(nx_huge_pages, &nx_huge_pages_ops, &nx_huge_pages, 0644);
  75__MODULE_PARM_TYPE(nx_huge_pages, "bool");
  76module_param_cb(nx_huge_pages_recovery_ratio, &nx_huge_pages_recovery_ratio_ops,
  77		&nx_huge_pages_recovery_ratio, 0644);
  78__MODULE_PARM_TYPE(nx_huge_pages_recovery_ratio, "uint");
  79
  80/*
  81 * When setting this variable to true it enables Two-Dimensional-Paging
  82 * where the hardware walks 2 page tables:
  83 * 1. the guest-virtual to guest-physical
  84 * 2. while doing 1. it walks guest-physical to host-physical
  85 * If the hardware supports that we don't need to do shadow paging.
  86 */
  87bool tdp_enabled = false;
  88
  89enum {
  90	AUDIT_PRE_PAGE_FAULT,
  91	AUDIT_POST_PAGE_FAULT,
  92	AUDIT_PRE_PTE_WRITE,
  93	AUDIT_POST_PTE_WRITE,
  94	AUDIT_PRE_SYNC,
  95	AUDIT_POST_SYNC
  96};
  97
  98#undef MMU_DEBUG
  99
 100#ifdef MMU_DEBUG
 101static bool dbg = 0;
 102module_param(dbg, bool, 0644);
 103
 104#define pgprintk(x...) do { if (dbg) printk(x); } while (0)
 105#define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
 106#define MMU_WARN_ON(x) WARN_ON(x)
 107#else
 108#define pgprintk(x...) do { } while (0)
 109#define rmap_printk(x...) do { } while (0)
 110#define MMU_WARN_ON(x) do { } while (0)
 111#endif
 112
 113#define PTE_PREFETCH_NUM		8
 114
 115#define PT_FIRST_AVAIL_BITS_SHIFT 10
 116#define PT64_SECOND_AVAIL_BITS_SHIFT 54
 117
 118/*
 119 * The mask used to denote special SPTEs, which can be either MMIO SPTEs or
 120 * Access Tracking SPTEs.
 121 */
 122#define SPTE_SPECIAL_MASK (3ULL << 52)
 123#define SPTE_AD_ENABLED_MASK (0ULL << 52)
 124#define SPTE_AD_DISABLED_MASK (1ULL << 52)
 125#define SPTE_AD_WRPROT_ONLY_MASK (2ULL << 52)
 126#define SPTE_MMIO_MASK (3ULL << 52)
 127
 128#define PT64_LEVEL_BITS 9
 129
 130#define PT64_LEVEL_SHIFT(level) \
 131		(PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
 132
 133#define PT64_INDEX(address, level)\
 134	(((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 135
 136
 137#define PT32_LEVEL_BITS 10
 138
 139#define PT32_LEVEL_SHIFT(level) \
 140		(PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 141
 142#define PT32_LVL_OFFSET_MASK(level) \
 143	(PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 144						* PT32_LEVEL_BITS))) - 1))
 145
 146#define PT32_INDEX(address, level)\
 147	(((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 148
 149
 150#ifdef CONFIG_DYNAMIC_PHYSICAL_MASK
 151#define PT64_BASE_ADDR_MASK (physical_mask & ~(u64)(PAGE_SIZE-1))
 152#else
 153#define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 154#endif
 155#define PT64_LVL_ADDR_MASK(level) \
 156	(PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 157						* PT64_LEVEL_BITS))) - 1))
 158#define PT64_LVL_OFFSET_MASK(level) \
 159	(PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 160						* PT64_LEVEL_BITS))) - 1))
 161
 162#define PT32_BASE_ADDR_MASK PAGE_MASK
 163#define PT32_DIR_BASE_ADDR_MASK \
 164	(PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 165#define PT32_LVL_ADDR_MASK(level) \
 166	(PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 167					    * PT32_LEVEL_BITS))) - 1))
 168
 169#define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | shadow_user_mask \
 170			| shadow_x_mask | shadow_nx_mask | shadow_me_mask)
 171
 172#define ACC_EXEC_MASK    1
 173#define ACC_WRITE_MASK   PT_WRITABLE_MASK
 174#define ACC_USER_MASK    PT_USER_MASK
 175#define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 176
 177/* The mask for the R/X bits in EPT PTEs */
 178#define PT64_EPT_READABLE_MASK			0x1ull
 179#define PT64_EPT_EXECUTABLE_MASK		0x4ull
 180
 181#include <trace/events/kvm.h>
 182
 183#define SPTE_HOST_WRITEABLE	(1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 184#define SPTE_MMU_WRITEABLE	(1ULL << (PT_FIRST_AVAIL_BITS_SHIFT + 1))
 185
 186#define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 187
 188/* make pte_list_desc fit well in cache line */
 189#define PTE_LIST_EXT 3
 190
 191/*
 192 * Return values of handle_mmio_page_fault and mmu.page_fault:
 193 * RET_PF_RETRY: let CPU fault again on the address.
 194 * RET_PF_EMULATE: mmio page fault, emulate the instruction directly.
 195 *
 196 * For handle_mmio_page_fault only:
 197 * RET_PF_INVALID: the spte is invalid, let the real page fault path update it.
 198 */
 199enum {
 200	RET_PF_RETRY = 0,
 201	RET_PF_EMULATE = 1,
 202	RET_PF_INVALID = 2,
 203};
 204
 205struct pte_list_desc {
 206	u64 *sptes[PTE_LIST_EXT];
 207	struct pte_list_desc *more;
 208};
 209
 210struct kvm_shadow_walk_iterator {
 211	u64 addr;
 212	hpa_t shadow_addr;
 213	u64 *sptep;
 214	int level;
 215	unsigned index;
 216};
 217
 218static const union kvm_mmu_page_role mmu_base_role_mask = {
 219	.cr0_wp = 1,
 220	.gpte_is_8_bytes = 1,
 221	.nxe = 1,
 222	.smep_andnot_wp = 1,
 223	.smap_andnot_wp = 1,
 224	.smm = 1,
 225	.guest_mode = 1,
 226	.ad_disabled = 1,
 227};
 228
 229#define for_each_shadow_entry_using_root(_vcpu, _root, _addr, _walker)     \
 230	for (shadow_walk_init_using_root(&(_walker), (_vcpu),              \
 231					 (_root), (_addr));                \
 232	     shadow_walk_okay(&(_walker));			           \
 233	     shadow_walk_next(&(_walker)))
 234
 235#define for_each_shadow_entry(_vcpu, _addr, _walker)            \
 236	for (shadow_walk_init(&(_walker), _vcpu, _addr);	\
 237	     shadow_walk_okay(&(_walker));			\
 238	     shadow_walk_next(&(_walker)))
 239
 240#define for_each_shadow_entry_lockless(_vcpu, _addr, _walker, spte)	\
 241	for (shadow_walk_init(&(_walker), _vcpu, _addr);		\
 242	     shadow_walk_okay(&(_walker)) &&				\
 243		({ spte = mmu_spte_get_lockless(_walker.sptep); 1; });	\
 244	     __shadow_walk_next(&(_walker), spte))
 245
 246static struct kmem_cache *pte_list_desc_cache;
 247static struct kmem_cache *mmu_page_header_cache;
 248static struct percpu_counter kvm_total_used_mmu_pages;
 249
 250static u64 __read_mostly shadow_nx_mask;
 251static u64 __read_mostly shadow_x_mask;	/* mutual exclusive with nx_mask */
 252static u64 __read_mostly shadow_user_mask;
 253static u64 __read_mostly shadow_accessed_mask;
 254static u64 __read_mostly shadow_dirty_mask;
 255static u64 __read_mostly shadow_mmio_mask;
 256static u64 __read_mostly shadow_mmio_value;
 257static u64 __read_mostly shadow_mmio_access_mask;
 258static u64 __read_mostly shadow_present_mask;
 259static u64 __read_mostly shadow_me_mask;
 260
 261/*
 262 * SPTEs used by MMUs without A/D bits are marked with SPTE_AD_DISABLED_MASK;
 263 * shadow_acc_track_mask is the set of bits to be cleared in non-accessed
 264 * pages.
 265 */
 266static u64 __read_mostly shadow_acc_track_mask;
 267
 268/*
 269 * The mask/shift to use for saving the original R/X bits when marking the PTE
 270 * as not-present for access tracking purposes. We do not save the W bit as the
 271 * PTEs being access tracked also need to be dirty tracked, so the W bit will be
 272 * restored only when a write is attempted to the page.
 273 */
 274static const u64 shadow_acc_track_saved_bits_mask = PT64_EPT_READABLE_MASK |
 275						    PT64_EPT_EXECUTABLE_MASK;
 276static const u64 shadow_acc_track_saved_bits_shift = PT64_SECOND_AVAIL_BITS_SHIFT;
 277
 278/*
 279 * This mask must be set on all non-zero Non-Present or Reserved SPTEs in order
 280 * to guard against L1TF attacks.
 281 */
 282static u64 __read_mostly shadow_nonpresent_or_rsvd_mask;
 283
 284/*
 285 * The number of high-order 1 bits to use in the mask above.
 286 */
 287static const u64 shadow_nonpresent_or_rsvd_mask_len = 5;
 288
 289/*
 290 * In some cases, we need to preserve the GFN of a non-present or reserved
 291 * SPTE when we usurp the upper five bits of the physical address space to
 292 * defend against L1TF, e.g. for MMIO SPTEs.  To preserve the GFN, we'll
 293 * shift bits of the GFN that overlap with shadow_nonpresent_or_rsvd_mask
 294 * left into the reserved bits, i.e. the GFN in the SPTE will be split into
 295 * high and low parts.  This mask covers the lower bits of the GFN.
 296 */
 297static u64 __read_mostly shadow_nonpresent_or_rsvd_lower_gfn_mask;
 298
 299/*
 300 * The number of non-reserved physical address bits irrespective of features
 301 * that repurpose legal bits, e.g. MKTME.
 302 */
 303static u8 __read_mostly shadow_phys_bits;
 304
 305static void mmu_spte_set(u64 *sptep, u64 spte);
 306static bool is_executable_pte(u64 spte);
 307static union kvm_mmu_page_role
 308kvm_mmu_calc_root_page_role(struct kvm_vcpu *vcpu);
 309
 310#define CREATE_TRACE_POINTS
 311#include "mmutrace.h"
 312
 313
 314static inline bool kvm_available_flush_tlb_with_range(void)
 315{
 316	return kvm_x86_ops->tlb_remote_flush_with_range;
 317}
 318
 319static void kvm_flush_remote_tlbs_with_range(struct kvm *kvm,
 320		struct kvm_tlb_range *range)
 321{
 322	int ret = -ENOTSUPP;
 323
 324	if (range && kvm_x86_ops->tlb_remote_flush_with_range)
 325		ret = kvm_x86_ops->tlb_remote_flush_with_range(kvm, range);
 326
 327	if (ret)
 328		kvm_flush_remote_tlbs(kvm);
 329}
 330
 331static void kvm_flush_remote_tlbs_with_address(struct kvm *kvm,
 332		u64 start_gfn, u64 pages)
 333{
 334	struct kvm_tlb_range range;
 335
 336	range.start_gfn = start_gfn;
 337	range.pages = pages;
 338
 339	kvm_flush_remote_tlbs_with_range(kvm, &range);
 340}
 341
 342void kvm_mmu_set_mmio_spte_mask(u64 mmio_mask, u64 mmio_value, u64 access_mask)
 343{
 344	BUG_ON((u64)(unsigned)access_mask != access_mask);
 345	BUG_ON((mmio_mask & mmio_value) != mmio_value);
 346	shadow_mmio_value = mmio_value | SPTE_MMIO_MASK;
 347	shadow_mmio_mask = mmio_mask | SPTE_SPECIAL_MASK;
 348	shadow_mmio_access_mask = access_mask;
 349}
 350EXPORT_SYMBOL_GPL(kvm_mmu_set_mmio_spte_mask);
 351
 352static bool is_mmio_spte(u64 spte)
 353{
 354	return (spte & shadow_mmio_mask) == shadow_mmio_value;
 355}
 356
 357static inline bool sp_ad_disabled(struct kvm_mmu_page *sp)
 358{
 359	return sp->role.ad_disabled;
 360}
 361
 362static inline bool kvm_vcpu_ad_need_write_protect(struct kvm_vcpu *vcpu)
 363{
 364	/*
 365	 * When using the EPT page-modification log, the GPAs in the log
 366	 * would come from L2 rather than L1.  Therefore, we need to rely
 367	 * on write protection to record dirty pages.  This also bypasses
 368	 * PML, since writes now result in a vmexit.
 369	 */
 370	return vcpu->arch.mmu == &vcpu->arch.guest_mmu;
 371}
 372
 373static inline bool spte_ad_enabled(u64 spte)
 374{
 375	MMU_WARN_ON(is_mmio_spte(spte));
 376	return (spte & SPTE_SPECIAL_MASK) != SPTE_AD_DISABLED_MASK;
 377}
 378
 379static inline bool spte_ad_need_write_protect(u64 spte)
 380{
 381	MMU_WARN_ON(is_mmio_spte(spte));
 382	return (spte & SPTE_SPECIAL_MASK) != SPTE_AD_ENABLED_MASK;
 383}
 384
 385static bool is_nx_huge_page_enabled(void)
 386{
 387	return READ_ONCE(nx_huge_pages);
 388}
 389
 390static inline u64 spte_shadow_accessed_mask(u64 spte)
 391{
 392	MMU_WARN_ON(is_mmio_spte(spte));
 393	return spte_ad_enabled(spte) ? shadow_accessed_mask : 0;
 394}
 395
 396static inline u64 spte_shadow_dirty_mask(u64 spte)
 397{
 398	MMU_WARN_ON(is_mmio_spte(spte));
 399	return spte_ad_enabled(spte) ? shadow_dirty_mask : 0;
 400}
 401
 402static inline bool is_access_track_spte(u64 spte)
 403{
 404	return !spte_ad_enabled(spte) && (spte & shadow_acc_track_mask) == 0;
 405}
 406
 407/*
 408 * Due to limited space in PTEs, the MMIO generation is a 19 bit subset of
 409 * the memslots generation and is derived as follows:
 410 *
 411 * Bits 0-8 of the MMIO generation are propagated to spte bits 3-11
 412 * Bits 9-18 of the MMIO generation are propagated to spte bits 52-61
 413 *
 414 * The KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS flag is intentionally not included in
 415 * the MMIO generation number, as doing so would require stealing a bit from
 416 * the "real" generation number and thus effectively halve the maximum number
 417 * of MMIO generations that can be handled before encountering a wrap (which
 418 * requires a full MMU zap).  The flag is instead explicitly queried when
 419 * checking for MMIO spte cache hits.
 420 */
 421#define MMIO_SPTE_GEN_MASK		GENMASK_ULL(18, 0)
 422
 423#define MMIO_SPTE_GEN_LOW_START		3
 424#define MMIO_SPTE_GEN_LOW_END		11
 425#define MMIO_SPTE_GEN_LOW_MASK		GENMASK_ULL(MMIO_SPTE_GEN_LOW_END, \
 426						    MMIO_SPTE_GEN_LOW_START)
 427
 428#define MMIO_SPTE_GEN_HIGH_START	52
 429#define MMIO_SPTE_GEN_HIGH_END		61
 430#define MMIO_SPTE_GEN_HIGH_MASK		GENMASK_ULL(MMIO_SPTE_GEN_HIGH_END, \
 431						    MMIO_SPTE_GEN_HIGH_START)
 432static u64 generation_mmio_spte_mask(u64 gen)
 433{
 434	u64 mask;
 435
 436	WARN_ON(gen & ~MMIO_SPTE_GEN_MASK);
 437
 438	mask = (gen << MMIO_SPTE_GEN_LOW_START) & MMIO_SPTE_GEN_LOW_MASK;
 439	mask |= (gen << MMIO_SPTE_GEN_HIGH_START) & MMIO_SPTE_GEN_HIGH_MASK;
 440	return mask;
 441}
 442
 443static u64 get_mmio_spte_generation(u64 spte)
 444{
 445	u64 gen;
 446
 447	spte &= ~shadow_mmio_mask;
 448
 449	gen = (spte & MMIO_SPTE_GEN_LOW_MASK) >> MMIO_SPTE_GEN_LOW_START;
 450	gen |= (spte & MMIO_SPTE_GEN_HIGH_MASK) >> MMIO_SPTE_GEN_HIGH_START;
 451	return gen;
 452}
 453
 454static void mark_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, u64 gfn,
 455			   unsigned access)
 456{
 457	u64 gen = kvm_vcpu_memslots(vcpu)->generation & MMIO_SPTE_GEN_MASK;
 458	u64 mask = generation_mmio_spte_mask(gen);
 459	u64 gpa = gfn << PAGE_SHIFT;
 460
 461	access &= shadow_mmio_access_mask;
 462	mask |= shadow_mmio_value | access;
 463	mask |= gpa | shadow_nonpresent_or_rsvd_mask;
 464	mask |= (gpa & shadow_nonpresent_or_rsvd_mask)
 465		<< shadow_nonpresent_or_rsvd_mask_len;
 466
 467	trace_mark_mmio_spte(sptep, gfn, access, gen);
 468	mmu_spte_set(sptep, mask);
 469}
 470
 471static gfn_t get_mmio_spte_gfn(u64 spte)
 472{
 473	u64 gpa = spte & shadow_nonpresent_or_rsvd_lower_gfn_mask;
 474
 475	gpa |= (spte >> shadow_nonpresent_or_rsvd_mask_len)
 476	       & shadow_nonpresent_or_rsvd_mask;
 477
 478	return gpa >> PAGE_SHIFT;
 479}
 480
 481static unsigned get_mmio_spte_access(u64 spte)
 482{
 483	return spte & shadow_mmio_access_mask;
 484}
 485
 486static bool set_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, gfn_t gfn,
 487			  kvm_pfn_t pfn, unsigned access)
 488{
 489	if (unlikely(is_noslot_pfn(pfn))) {
 490		mark_mmio_spte(vcpu, sptep, gfn, access);
 491		return true;
 492	}
 493
 494	return false;
 495}
 496
 497static bool check_mmio_spte(struct kvm_vcpu *vcpu, u64 spte)
 498{
 499	u64 kvm_gen, spte_gen, gen;
 500
 501	gen = kvm_vcpu_memslots(vcpu)->generation;
 502	if (unlikely(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS))
 503		return false;
 504
 505	kvm_gen = gen & MMIO_SPTE_GEN_MASK;
 506	spte_gen = get_mmio_spte_generation(spte);
 507
 508	trace_check_mmio_spte(spte, kvm_gen, spte_gen);
 509	return likely(kvm_gen == spte_gen);
 510}
 511
 512/*
 513 * Sets the shadow PTE masks used by the MMU.
 514 *
 515 * Assumptions:
 516 *  - Setting either @accessed_mask or @dirty_mask requires setting both
 517 *  - At least one of @accessed_mask or @acc_track_mask must be set
 518 */
 519void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 520		u64 dirty_mask, u64 nx_mask, u64 x_mask, u64 p_mask,
 521		u64 acc_track_mask, u64 me_mask)
 522{
 523	BUG_ON(!dirty_mask != !accessed_mask);
 524	BUG_ON(!accessed_mask && !acc_track_mask);
 525	BUG_ON(acc_track_mask & SPTE_SPECIAL_MASK);
 526
 527	shadow_user_mask = user_mask;
 528	shadow_accessed_mask = accessed_mask;
 529	shadow_dirty_mask = dirty_mask;
 530	shadow_nx_mask = nx_mask;
 531	shadow_x_mask = x_mask;
 532	shadow_present_mask = p_mask;
 533	shadow_acc_track_mask = acc_track_mask;
 534	shadow_me_mask = me_mask;
 535}
 536EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 537
 538static u8 kvm_get_shadow_phys_bits(void)
 539{
 540	/*
 541	 * boot_cpu_data.x86_phys_bits is reduced when MKTME is detected
 542	 * in CPU detection code, but MKTME treats those reduced bits as
 543	 * 'keyID' thus they are not reserved bits. Therefore for MKTME
 544	 * we should still return physical address bits reported by CPUID.
 545	 */
 546	if (!boot_cpu_has(X86_FEATURE_TME) ||
 547	    WARN_ON_ONCE(boot_cpu_data.extended_cpuid_level < 0x80000008))
 548		return boot_cpu_data.x86_phys_bits;
 549
 550	return cpuid_eax(0x80000008) & 0xff;
 551}
 552
 553static void kvm_mmu_reset_all_pte_masks(void)
 554{
 555	u8 low_phys_bits;
 556
 557	shadow_user_mask = 0;
 558	shadow_accessed_mask = 0;
 559	shadow_dirty_mask = 0;
 560	shadow_nx_mask = 0;
 561	shadow_x_mask = 0;
 562	shadow_mmio_mask = 0;
 563	shadow_present_mask = 0;
 564	shadow_acc_track_mask = 0;
 565
 566	shadow_phys_bits = kvm_get_shadow_phys_bits();
 567
 568	/*
 569	 * If the CPU has 46 or less physical address bits, then set an
 570	 * appropriate mask to guard against L1TF attacks. Otherwise, it is
 571	 * assumed that the CPU is not vulnerable to L1TF.
 572	 *
 573	 * Some Intel CPUs address the L1 cache using more PA bits than are
 574	 * reported by CPUID. Use the PA width of the L1 cache when possible
 575	 * to achieve more effective mitigation, e.g. if system RAM overlaps
 576	 * the most significant bits of legal physical address space.
 577	 */
 578	shadow_nonpresent_or_rsvd_mask = 0;
 579	low_phys_bits = boot_cpu_data.x86_cache_bits;
 580	if (boot_cpu_data.x86_cache_bits <
 581	    52 - shadow_nonpresent_or_rsvd_mask_len) {
 582		shadow_nonpresent_or_rsvd_mask =
 583			rsvd_bits(boot_cpu_data.x86_cache_bits -
 584				  shadow_nonpresent_or_rsvd_mask_len,
 585				  boot_cpu_data.x86_cache_bits - 1);
 586		low_phys_bits -= shadow_nonpresent_or_rsvd_mask_len;
 587	} else
 588		WARN_ON_ONCE(boot_cpu_has_bug(X86_BUG_L1TF));
 589
 590	shadow_nonpresent_or_rsvd_lower_gfn_mask =
 591		GENMASK_ULL(low_phys_bits - 1, PAGE_SHIFT);
 592}
 593
 594static int is_cpuid_PSE36(void)
 595{
 596	return 1;
 597}
 598
 599static int is_nx(struct kvm_vcpu *vcpu)
 600{
 601	return vcpu->arch.efer & EFER_NX;
 602}
 603
 604static int is_shadow_present_pte(u64 pte)
 605{
 606	return (pte != 0) && !is_mmio_spte(pte);
 607}
 608
 609static int is_large_pte(u64 pte)
 610{
 611	return pte & PT_PAGE_SIZE_MASK;
 612}
 613
 614static int is_last_spte(u64 pte, int level)
 615{
 616	if (level == PT_PAGE_TABLE_LEVEL)
 617		return 1;
 618	if (is_large_pte(pte))
 619		return 1;
 620	return 0;
 621}
 622
 623static bool is_executable_pte(u64 spte)
 624{
 625	return (spte & (shadow_x_mask | shadow_nx_mask)) == shadow_x_mask;
 626}
 627
 628static kvm_pfn_t spte_to_pfn(u64 pte)
 629{
 630	return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 631}
 632
 633static gfn_t pse36_gfn_delta(u32 gpte)
 634{
 635	int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 636
 637	return (gpte & PT32_DIR_PSE36_MASK) << shift;
 638}
 639
 640#ifdef CONFIG_X86_64
 641static void __set_spte(u64 *sptep, u64 spte)
 642{
 643	WRITE_ONCE(*sptep, spte);
 644}
 645
 646static void __update_clear_spte_fast(u64 *sptep, u64 spte)
 647{
 648	WRITE_ONCE(*sptep, spte);
 649}
 650
 651static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
 652{
 653	return xchg(sptep, spte);
 654}
 655
 656static u64 __get_spte_lockless(u64 *sptep)
 657{
 658	return READ_ONCE(*sptep);
 659}
 660#else
 661union split_spte {
 662	struct {
 663		u32 spte_low;
 664		u32 spte_high;
 665	};
 666	u64 spte;
 667};
 668
 669static void count_spte_clear(u64 *sptep, u64 spte)
 670{
 671	struct kvm_mmu_page *sp =  page_header(__pa(sptep));
 672
 673	if (is_shadow_present_pte(spte))
 674		return;
 675
 676	/* Ensure the spte is completely set before we increase the count */
 677	smp_wmb();
 678	sp->clear_spte_count++;
 679}
 680
 681static void __set_spte(u64 *sptep, u64 spte)
 682{
 683	union split_spte *ssptep, sspte;
 684
 685	ssptep = (union split_spte *)sptep;
 686	sspte = (union split_spte)spte;
 687
 688	ssptep->spte_high = sspte.spte_high;
 689
 690	/*
 691	 * If we map the spte from nonpresent to present, We should store
 692	 * the high bits firstly, then set present bit, so cpu can not
 693	 * fetch this spte while we are setting the spte.
 694	 */
 695	smp_wmb();
 696
 697	WRITE_ONCE(ssptep->spte_low, sspte.spte_low);
 698}
 699
 700static void __update_clear_spte_fast(u64 *sptep, u64 spte)
 701{
 702	union split_spte *ssptep, sspte;
 703
 704	ssptep = (union split_spte *)sptep;
 705	sspte = (union split_spte)spte;
 706
 707	WRITE_ONCE(ssptep->spte_low, sspte.spte_low);
 708
 709	/*
 710	 * If we map the spte from present to nonpresent, we should clear
 711	 * present bit firstly to avoid vcpu fetch the old high bits.
 712	 */
 713	smp_wmb();
 714
 715	ssptep->spte_high = sspte.spte_high;
 716	count_spte_clear(sptep, spte);
 717}
 718
 719static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
 720{
 721	union split_spte *ssptep, sspte, orig;
 722
 723	ssptep = (union split_spte *)sptep;
 724	sspte = (union split_spte)spte;
 725
 726	/* xchg acts as a barrier before the setting of the high bits */
 727	orig.spte_low = xchg(&ssptep->spte_low, sspte.spte_low);
 728	orig.spte_high = ssptep->spte_high;
 729	ssptep->spte_high = sspte.spte_high;
 730	count_spte_clear(sptep, spte);
 731
 732	return orig.spte;
 733}
 734
 735/*
 736 * The idea using the light way get the spte on x86_32 guest is from
 737 * gup_get_pte (mm/gup.c).
 738 *
 739 * An spte tlb flush may be pending, because kvm_set_pte_rmapp
 740 * coalesces them and we are running out of the MMU lock.  Therefore
 741 * we need to protect against in-progress updates of the spte.
 742 *
 743 * Reading the spte while an update is in progress may get the old value
 744 * for the high part of the spte.  The race is fine for a present->non-present
 745 * change (because the high part of the spte is ignored for non-present spte),
 746 * but for a present->present change we must reread the spte.
 747 *
 748 * All such changes are done in two steps (present->non-present and
 749 * non-present->present), hence it is enough to count the number of
 750 * present->non-present updates: if it changed while reading the spte,
 751 * we might have hit the race.  This is done using clear_spte_count.
 752 */
 753static u64 __get_spte_lockless(u64 *sptep)
 754{
 755	struct kvm_mmu_page *sp =  page_header(__pa(sptep));
 756	union split_spte spte, *orig = (union split_spte *)sptep;
 757	int count;
 758
 759retry:
 760	count = sp->clear_spte_count;
 761	smp_rmb();
 762
 763	spte.spte_low = orig->spte_low;
 764	smp_rmb();
 765
 766	spte.spte_high = orig->spte_high;
 767	smp_rmb();
 768
 769	if (unlikely(spte.spte_low != orig->spte_low ||
 770	      count != sp->clear_spte_count))
 771		goto retry;
 772
 773	return spte.spte;
 774}
 775#endif
 776
 777static bool spte_can_locklessly_be_made_writable(u64 spte)
 778{
 779	return (spte & (SPTE_HOST_WRITEABLE | SPTE_MMU_WRITEABLE)) ==
 780		(SPTE_HOST_WRITEABLE | SPTE_MMU_WRITEABLE);
 781}
 782
 783static bool spte_has_volatile_bits(u64 spte)
 784{
 785	if (!is_shadow_present_pte(spte))
 786		return false;
 787
 788	/*
 789	 * Always atomically update spte if it can be updated
 790	 * out of mmu-lock, it can ensure dirty bit is not lost,
 791	 * also, it can help us to get a stable is_writable_pte()
 792	 * to ensure tlb flush is not missed.
 793	 */
 794	if (spte_can_locklessly_be_made_writable(spte) ||
 795	    is_access_track_spte(spte))
 796		return true;
 797
 798	if (spte_ad_enabled(spte)) {
 799		if ((spte & shadow_accessed_mask) == 0 ||
 800	    	    (is_writable_pte(spte) && (spte & shadow_dirty_mask) == 0))
 801			return true;
 802	}
 803
 804	return false;
 805}
 806
 807static bool is_accessed_spte(u64 spte)
 808{
 809	u64 accessed_mask = spte_shadow_accessed_mask(spte);
 810
 811	return accessed_mask ? spte & accessed_mask
 812			     : !is_access_track_spte(spte);
 813}
 814
 815static bool is_dirty_spte(u64 spte)
 816{
 817	u64 dirty_mask = spte_shadow_dirty_mask(spte);
 818
 819	return dirty_mask ? spte & dirty_mask : spte & PT_WRITABLE_MASK;
 820}
 821
 822/* Rules for using mmu_spte_set:
 823 * Set the sptep from nonpresent to present.
 824 * Note: the sptep being assigned *must* be either not present
 825 * or in a state where the hardware will not attempt to update
 826 * the spte.
 827 */
 828static void mmu_spte_set(u64 *sptep, u64 new_spte)
 829{
 830	WARN_ON(is_shadow_present_pte(*sptep));
 831	__set_spte(sptep, new_spte);
 832}
 833
 834/*
 835 * Update the SPTE (excluding the PFN), but do not track changes in its
 836 * accessed/dirty status.
 837 */
 838static u64 mmu_spte_update_no_track(u64 *sptep, u64 new_spte)
 839{
 840	u64 old_spte = *sptep;
 841
 842	WARN_ON(!is_shadow_present_pte(new_spte));
 843
 844	if (!is_shadow_present_pte(old_spte)) {
 845		mmu_spte_set(sptep, new_spte);
 846		return old_spte;
 847	}
 848
 849	if (!spte_has_volatile_bits(old_spte))
 850		__update_clear_spte_fast(sptep, new_spte);
 851	else
 852		old_spte = __update_clear_spte_slow(sptep, new_spte);
 853
 854	WARN_ON(spte_to_pfn(old_spte) != spte_to_pfn(new_spte));
 855
 856	return old_spte;
 857}
 858
 859/* Rules for using mmu_spte_update:
 860 * Update the state bits, it means the mapped pfn is not changed.
 861 *
 862 * Whenever we overwrite a writable spte with a read-only one we
 863 * should flush remote TLBs. Otherwise rmap_write_protect
 864 * will find a read-only spte, even though the writable spte
 865 * might be cached on a CPU's TLB, the return value indicates this
 866 * case.
 867 *
 868 * Returns true if the TLB needs to be flushed
 869 */
 870static bool mmu_spte_update(u64 *sptep, u64 new_spte)
 871{
 872	bool flush = false;
 873	u64 old_spte = mmu_spte_update_no_track(sptep, new_spte);
 874
 875	if (!is_shadow_present_pte(old_spte))
 876		return false;
 877
 878	/*
 879	 * For the spte updated out of mmu-lock is safe, since
 880	 * we always atomically update it, see the comments in
 881	 * spte_has_volatile_bits().
 882	 */
 883	if (spte_can_locklessly_be_made_writable(old_spte) &&
 884	      !is_writable_pte(new_spte))
 885		flush = true;
 886
 887	/*
 888	 * Flush TLB when accessed/dirty states are changed in the page tables,
 889	 * to guarantee consistency between TLB and page tables.
 890	 */
 891
 892	if (is_accessed_spte(old_spte) && !is_accessed_spte(new_spte)) {
 893		flush = true;
 894		kvm_set_pfn_accessed(spte_to_pfn(old_spte));
 895	}
 896
 897	if (is_dirty_spte(old_spte) && !is_dirty_spte(new_spte)) {
 898		flush = true;
 899		kvm_set_pfn_dirty(spte_to_pfn(old_spte));
 900	}
 901
 902	return flush;
 903}
 904
 905/*
 906 * Rules for using mmu_spte_clear_track_bits:
 907 * It sets the sptep from present to nonpresent, and track the
 908 * state bits, it is used to clear the last level sptep.
 909 * Returns non-zero if the PTE was previously valid.
 910 */
 911static int mmu_spte_clear_track_bits(u64 *sptep)
 912{
 913	kvm_pfn_t pfn;
 914	u64 old_spte = *sptep;
 915
 916	if (!spte_has_volatile_bits(old_spte))
 917		__update_clear_spte_fast(sptep, 0ull);
 918	else
 919		old_spte = __update_clear_spte_slow(sptep, 0ull);
 920
 921	if (!is_shadow_present_pte(old_spte))
 922		return 0;
 923
 924	pfn = spte_to_pfn(old_spte);
 925
 926	/*
 927	 * KVM does not hold the refcount of the page used by
 928	 * kvm mmu, before reclaiming the page, we should
 929	 * unmap it from mmu first.
 930	 */
 931	WARN_ON(!kvm_is_reserved_pfn(pfn) && !page_count(pfn_to_page(pfn)));
 932
 933	if (is_accessed_spte(old_spte))
 934		kvm_set_pfn_accessed(pfn);
 935
 936	if (is_dirty_spte(old_spte))
 937		kvm_set_pfn_dirty(pfn);
 938
 939	return 1;
 940}
 941
 942/*
 943 * Rules for using mmu_spte_clear_no_track:
 944 * Directly clear spte without caring the state bits of sptep,
 945 * it is used to set the upper level spte.
 946 */
 947static void mmu_spte_clear_no_track(u64 *sptep)
 948{
 949	__update_clear_spte_fast(sptep, 0ull);
 950}
 951
 952static u64 mmu_spte_get_lockless(u64 *sptep)
 953{
 954	return __get_spte_lockless(sptep);
 955}
 956
 957static u64 mark_spte_for_access_track(u64 spte)
 958{
 959	if (spte_ad_enabled(spte))
 960		return spte & ~shadow_accessed_mask;
 961
 962	if (is_access_track_spte(spte))
 963		return spte;
 964
 965	/*
 966	 * Making an Access Tracking PTE will result in removal of write access
 967	 * from the PTE. So, verify that we will be able to restore the write
 968	 * access in the fast page fault path later on.
 969	 */
 970	WARN_ONCE((spte & PT_WRITABLE_MASK) &&
 971		  !spte_can_locklessly_be_made_writable(spte),
 972		  "kvm: Writable SPTE is not locklessly dirty-trackable\n");
 973
 974	WARN_ONCE(spte & (shadow_acc_track_saved_bits_mask <<
 975			  shadow_acc_track_saved_bits_shift),
 976		  "kvm: Access Tracking saved bit locations are not zero\n");
 977
 978	spte |= (spte & shadow_acc_track_saved_bits_mask) <<
 979		shadow_acc_track_saved_bits_shift;
 980	spte &= ~shadow_acc_track_mask;
 981
 982	return spte;
 983}
 984
 985/* Restore an acc-track PTE back to a regular PTE */
 986static u64 restore_acc_track_spte(u64 spte)
 987{
 988	u64 new_spte = spte;
 989	u64 saved_bits = (spte >> shadow_acc_track_saved_bits_shift)
 990			 & shadow_acc_track_saved_bits_mask;
 991
 992	WARN_ON_ONCE(spte_ad_enabled(spte));
 993	WARN_ON_ONCE(!is_access_track_spte(spte));
 994
 995	new_spte &= ~shadow_acc_track_mask;
 996	new_spte &= ~(shadow_acc_track_saved_bits_mask <<
 997		      shadow_acc_track_saved_bits_shift);
 998	new_spte |= saved_bits;
 999
1000	return new_spte;
1001}
1002
1003/* Returns the Accessed status of the PTE and resets it at the same time. */
1004static bool mmu_spte_age(u64 *sptep)
1005{
1006	u64 spte = mmu_spte_get_lockless(sptep);
1007
1008	if (!is_accessed_spte(spte))
1009		return false;
1010
1011	if (spte_ad_enabled(spte)) {
1012		clear_bit((ffs(shadow_accessed_mask) - 1),
1013			  (unsigned long *)sptep);
1014	} else {
1015		/*
1016		 * Capture the dirty status of the page, so that it doesn't get
1017		 * lost when the SPTE is marked for access tracking.
1018		 */
1019		if (is_writable_pte(spte))
1020			kvm_set_pfn_dirty(spte_to_pfn(spte));
1021
1022		spte = mark_spte_for_access_track(spte);
1023		mmu_spte_update_no_track(sptep, spte);
1024	}
1025
1026	return true;
1027}
1028
1029static void walk_shadow_page_lockless_begin(struct kvm_vcpu *vcpu)
1030{
1031	/*
1032	 * Prevent page table teardown by making any free-er wait during
1033	 * kvm_flush_remote_tlbs() IPI to all active vcpus.
1034	 */
1035	local_irq_disable();
1036
1037	/*
1038	 * Make sure a following spte read is not reordered ahead of the write
1039	 * to vcpu->mode.
1040	 */
1041	smp_store_mb(vcpu->mode, READING_SHADOW_PAGE_TABLES);
1042}
1043
1044static void walk_shadow_page_lockless_end(struct kvm_vcpu *vcpu)
1045{
1046	/*
1047	 * Make sure the write to vcpu->mode is not reordered in front of
1048	 * reads to sptes.  If it does, kvm_mmu_commit_zap_page() can see us
1049	 * OUTSIDE_GUEST_MODE and proceed to free the shadow page table.
1050	 */
1051	smp_store_release(&vcpu->mode, OUTSIDE_GUEST_MODE);
1052	local_irq_enable();
1053}
1054
1055static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
1056				  struct kmem_cache *base_cache, int min)
1057{
1058	void *obj;
1059
1060	if (cache->nobjs >= min)
1061		return 0;
1062	while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
1063		obj = kmem_cache_zalloc(base_cache, GFP_KERNEL_ACCOUNT);
1064		if (!obj)
1065			return cache->nobjs >= min ? 0 : -ENOMEM;
1066		cache->objects[cache->nobjs++] = obj;
1067	}
1068	return 0;
1069}
1070
1071static int mmu_memory_cache_free_objects(struct kvm_mmu_memory_cache *cache)
1072{
1073	return cache->nobjs;
1074}
1075
1076static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc,
1077				  struct kmem_cache *cache)
1078{
1079	while (mc->nobjs)
1080		kmem_cache_free(cache, mc->objects[--mc->nobjs]);
1081}
1082
1083static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
1084				       int min)
1085{
1086	void *page;
1087
1088	if (cache->nobjs >= min)
1089		return 0;
1090	while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
1091		page = (void *)__get_free_page(GFP_KERNEL_ACCOUNT);
1092		if (!page)
1093			return cache->nobjs >= min ? 0 : -ENOMEM;
1094		cache->objects[cache->nobjs++] = page;
1095	}
1096	return 0;
1097}
1098
1099static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
1100{
1101	while (mc->nobjs)
1102		free_page((unsigned long)mc->objects[--mc->nobjs]);
1103}
1104
1105static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
1106{
1107	int r;
1108
1109	r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache,
1110				   pte_list_desc_cache, 8 + PTE_PREFETCH_NUM);
1111	if (r)
1112		goto out;
1113	r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
1114	if (r)
1115		goto out;
1116	r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
1117				   mmu_page_header_cache, 4);
1118out:
1119	return r;
1120}
1121
1122static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1123{
1124	mmu_free_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache,
1125				pte_list_desc_cache);
1126	mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
1127	mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache,
1128				mmu_page_header_cache);
1129}
1130
1131static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
1132{
1133	void *p;
1134
1135	BUG_ON(!mc->nobjs);
1136	p = mc->objects[--mc->nobjs];
1137	return p;
1138}
1139
1140static struct pte_list_desc *mmu_alloc_pte_list_desc(struct kvm_vcpu *vcpu)
1141{
1142	return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_list_desc_cache);
1143}
1144
1145static void mmu_free_pte_list_desc(struct pte_list_desc *pte_list_desc)
1146{
1147	kmem_cache_free(pte_list_desc_cache, pte_list_desc);
1148}
1149
1150static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index)
1151{
1152	if (!sp->role.direct)
1153		return sp->gfns[index];
1154
1155	return sp->gfn + (index << ((sp->role.level - 1) * PT64_LEVEL_BITS));
1156}
1157
1158static void kvm_mmu_page_set_gfn(struct kvm_mmu_page *sp, int index, gfn_t gfn)
1159{
1160	if (!sp->role.direct) {
1161		sp->gfns[index] = gfn;
1162		return;
1163	}
1164
1165	if (WARN_ON(gfn != kvm_mmu_page_get_gfn(sp, index)))
1166		pr_err_ratelimited("gfn mismatch under direct page %llx "
1167				   "(expected %llx, got %llx)\n",
1168				   sp->gfn,
1169				   kvm_mmu_page_get_gfn(sp, index), gfn);
1170}
1171
1172/*
1173 * Return the pointer to the large page information for a given gfn,
1174 * handling slots that are not large page aligned.
1175 */
1176static struct kvm_lpage_info *lpage_info_slot(gfn_t gfn,
1177					      struct kvm_memory_slot *slot,
1178					      int level)
1179{
1180	unsigned long idx;
1181
1182	idx = gfn_to_index(gfn, slot->base_gfn, level);
1183	return &slot->arch.lpage_info[level - 2][idx];
1184}
1185
1186static void update_gfn_disallow_lpage_count(struct kvm_memory_slot *slot,
1187					    gfn_t gfn, int count)
1188{
1189	struct kvm_lpage_info *linfo;
1190	int i;
1191
1192	for (i = PT_DIRECTORY_LEVEL; i <= PT_MAX_HUGEPAGE_LEVEL; ++i) {
1193		linfo = lpage_info_slot(gfn, slot, i);
1194		linfo->disallow_lpage += count;
1195		WARN_ON(linfo->disallow_lpage < 0);
1196	}
1197}
1198
1199void kvm_mmu_gfn_disallow_lpage(struct kvm_memory_slot *slot, gfn_t gfn)
1200{
1201	update_gfn_disallow_lpage_count(slot, gfn, 1);
1202}
1203
1204void kvm_mmu_gfn_allow_lpage(struct kvm_memory_slot *slot, gfn_t gfn)
1205{
1206	update_gfn_disallow_lpage_count(slot, gfn, -1);
1207}
1208
1209static void account_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp)
1210{
1211	struct kvm_memslots *slots;
1212	struct kvm_memory_slot *slot;
1213	gfn_t gfn;
1214
1215	kvm->arch.indirect_shadow_pages++;
1216	gfn = sp->gfn;
1217	slots = kvm_memslots_for_spte_role(kvm, sp->role);
1218	slot = __gfn_to_memslot(slots, gfn);
1219
1220	/* the non-leaf shadow pages are keeping readonly. */
1221	if (sp->role.level > PT_PAGE_TABLE_LEVEL)
1222		return kvm_slot_page_track_add_page(kvm, slot, gfn,
1223						    KVM_PAGE_TRACK_WRITE);
1224
1225	kvm_mmu_gfn_disallow_lpage(slot, gfn);
1226}
1227
1228static void account_huge_nx_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1229{
1230	if (sp->lpage_disallowed)
1231		return;
1232
1233	++kvm->stat.nx_lpage_splits;
1234	list_add_tail(&sp->lpage_disallowed_link,
1235		      &kvm->arch.lpage_disallowed_mmu_pages);
1236	sp->lpage_disallowed = true;
1237}
1238
1239static void unaccount_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp)
1240{
1241	struct kvm_memslots *slots;
1242	struct kvm_memory_slot *slot;
1243	gfn_t gfn;
1244
1245	kvm->arch.indirect_shadow_pages--;
1246	gfn = sp->gfn;
1247	slots = kvm_memslots_for_spte_role(kvm, sp->role);
1248	slot = __gfn_to_memslot(slots, gfn);
1249	if (sp->role.level > PT_PAGE_TABLE_LEVEL)
1250		return kvm_slot_page_track_remove_page(kvm, slot, gfn,
1251						       KVM_PAGE_TRACK_WRITE);
1252
1253	kvm_mmu_gfn_allow_lpage(slot, gfn);
1254}
1255
1256static void unaccount_huge_nx_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1257{
1258	--kvm->stat.nx_lpage_splits;
1259	sp->lpage_disallowed = false;
1260	list_del(&sp->lpage_disallowed_link);
1261}
1262
1263static bool __mmu_gfn_lpage_is_disallowed(gfn_t gfn, int level,
1264					  struct kvm_memory_slot *slot)
1265{
1266	struct kvm_lpage_info *linfo;
1267
1268	if (slot) {
1269		linfo = lpage_info_slot(gfn, slot, level);
1270		return !!linfo->disallow_lpage;
1271	}
1272
1273	return true;
1274}
1275
1276static bool mmu_gfn_lpage_is_disallowed(struct kvm_vcpu *vcpu, gfn_t gfn,
1277					int level)
1278{
1279	struct kvm_memory_slot *slot;
1280
1281	slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1282	return __mmu_gfn_lpage_is_disallowed(gfn, level, slot);
1283}
1284
1285static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
1286{
1287	unsigned long page_size;
1288	int i, ret = 0;
1289
1290	page_size = kvm_host_page_size(kvm, gfn);
1291
1292	for (i = PT_PAGE_TABLE_LEVEL; i <= PT_MAX_HUGEPAGE_LEVEL; ++i) {
1293		if (page_size >= KVM_HPAGE_SIZE(i))
1294			ret = i;
1295		else
1296			break;
1297	}
1298
1299	return ret;
1300}
1301
1302static inline bool memslot_valid_for_gpte(struct kvm_memory_slot *slot,
1303					  bool no_dirty_log)
1304{
1305	if (!slot || slot->flags & KVM_MEMSLOT_INVALID)
1306		return false;
1307	if (no_dirty_log && slot->dirty_bitmap)
1308		return false;
1309
1310	return true;
1311}
1312
1313static struct kvm_memory_slot *
1314gfn_to_memslot_dirty_bitmap(struct kvm_vcpu *vcpu, gfn_t gfn,
1315			    bool no_dirty_log)
1316{
1317	struct kvm_memory_slot *slot;
1318
1319	slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1320	if (!memslot_valid_for_gpte(slot, no_dirty_log))
1321		slot = NULL;
1322
1323	return slot;
1324}
1325
1326static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn,
1327			 bool *force_pt_level)
1328{
1329	int host_level, level, max_level;
1330	struct kvm_memory_slot *slot;
1331
1332	if (unlikely(*force_pt_level))
1333		return PT_PAGE_TABLE_LEVEL;
1334
1335	slot = kvm_vcpu_gfn_to_memslot(vcpu, large_gfn);
1336	*force_pt_level = !memslot_valid_for_gpte(slot, true);
1337	if (unlikely(*force_pt_level))
1338		return PT_PAGE_TABLE_LEVEL;
1339
1340	host_level = host_mapping_level(vcpu->kvm, large_gfn);
1341
1342	if (host_level == PT_PAGE_TABLE_LEVEL)
1343		return host_level;
1344
1345	max_level = min(kvm_x86_ops->get_lpage_level(), host_level);
1346
1347	for (level = PT_DIRECTORY_LEVEL; level <= max_level; ++level)
1348		if (__mmu_gfn_lpage_is_disallowed(large_gfn, level, slot))
1349			break;
1350
1351	return level - 1;
1352}
1353
1354/*
1355 * About rmap_head encoding:
1356 *
1357 * If the bit zero of rmap_head->val is clear, then it points to the only spte
1358 * in this rmap chain. Otherwise, (rmap_head->val & ~1) points to a struct
1359 * pte_list_desc containing more mappings.
1360 */
1361
1362/*
1363 * Returns the number of pointers in the rmap chain, not counting the new one.
1364 */
1365static int pte_list_add(struct kvm_vcpu *vcpu, u64 *spte,
1366			struct kvm_rmap_head *rmap_head)
1367{
1368	struct pte_list_desc *desc;
1369	int i, count = 0;
1370
1371	if (!rmap_head->val) {
1372		rmap_printk("pte_list_add: %p %llx 0->1\n", spte, *spte);
1373		rmap_head->val = (unsigned long)spte;
1374	} else if (!(rmap_head->val & 1)) {
1375		rmap_printk("pte_list_add: %p %llx 1->many\n", spte, *spte);
1376		desc = mmu_alloc_pte_list_desc(vcpu);
1377		desc->sptes[0] = (u64 *)rmap_head->val;
1378		desc->sptes[1] = spte;
1379		rmap_head->val = (unsigned long)desc | 1;
1380		++count;
1381	} else {
1382		rmap_printk("pte_list_add: %p %llx many->many\n", spte, *spte);
1383		desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
1384		while (desc->sptes[PTE_LIST_EXT-1] && desc->more) {
1385			desc = desc->more;
1386			count += PTE_LIST_EXT;
1387		}
1388		if (desc->sptes[PTE_LIST_EXT-1]) {
1389			desc->more = mmu_alloc_pte_list_desc(vcpu);
1390			desc = desc->more;
1391		}
1392		for (i = 0; desc->sptes[i]; ++i)
1393			++count;
1394		desc->sptes[i] = spte;
1395	}
1396	return count;
1397}
1398
1399static void
1400pte_list_desc_remove_entry(struct kvm_rmap_head *rmap_head,
1401			   struct pte_list_desc *desc, int i,
1402			   struct pte_list_desc *prev_desc)
1403{
1404	int j;
1405
1406	for (j = PTE_LIST_EXT - 1; !desc->sptes[j] && j > i; --j)
1407		;
1408	desc->sptes[i] = desc->sptes[j];
1409	desc->sptes[j] = NULL;
1410	if (j != 0)
1411		return;
1412	if (!prev_desc && !desc->more)
1413		rmap_head->val = (unsigned long)desc->sptes[0];
1414	else
1415		if (prev_desc)
1416			prev_desc->more = desc->more;
1417		else
1418			rmap_head->val = (unsigned long)desc->more | 1;
1419	mmu_free_pte_list_desc(desc);
1420}
1421
1422static void __pte_list_remove(u64 *spte, struct kvm_rmap_head *rmap_head)
1423{
1424	struct pte_list_desc *desc;
1425	struct pte_list_desc *prev_desc;
1426	int i;
1427
1428	if (!rmap_head->val) {
1429		pr_err("%s: %p 0->BUG\n", __func__, spte);
1430		BUG();
1431	} else if (!(rmap_head->val & 1)) {
1432		rmap_printk("%s:  %p 1->0\n", __func__, spte);
1433		if ((u64 *)rmap_head->val != spte) {
1434			pr_err("%s:  %p 1->BUG\n", __func__, spte);
1435			BUG();
1436		}
1437		rmap_head->val = 0;
1438	} else {
1439		rmap_printk("%s:  %p many->many\n", __func__, spte);
1440		desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
1441		prev_desc = NULL;
1442		while (desc) {
1443			for (i = 0; i < PTE_LIST_EXT && desc->sptes[i]; ++i) {
1444				if (desc->sptes[i] == spte) {
1445					pte_list_desc_remove_entry(rmap_head,
1446							desc, i, prev_desc);
1447					return;
1448				}
1449			}
1450			prev_desc = desc;
1451			desc = desc->more;
1452		}
1453		pr_err("%s: %p many->many\n", __func__, spte);
1454		BUG();
1455	}
1456}
1457
1458static void pte_list_remove(struct kvm_rmap_head *rmap_head, u64 *sptep)
1459{
1460	mmu_spte_clear_track_bits(sptep);
1461	__pte_list_remove(sptep, rmap_head);
1462}
1463
1464static struct kvm_rmap_head *__gfn_to_rmap(gfn_t gfn, int level,
1465					   struct kvm_memory_slot *slot)
1466{
1467	unsigned long idx;
1468
1469	idx = gfn_to_index(gfn, slot->base_gfn, level);
1470	return &slot->arch.rmap[level - PT_PAGE_TABLE_LEVEL][idx];
1471}
1472
1473static struct kvm_rmap_head *gfn_to_rmap(struct kvm *kvm, gfn_t gfn,
1474					 struct kvm_mmu_page *sp)
1475{
1476	struct kvm_memslots *slots;
1477	struct kvm_memory_slot *slot;
1478
1479	slots = kvm_memslots_for_spte_role(kvm, sp->role);
1480	slot = __gfn_to_memslot(slots, gfn);
1481	return __gfn_to_rmap(gfn, sp->role.level, slot);
1482}
1483
1484static bool rmap_can_add(struct kvm_vcpu *vcpu)
1485{
1486	struct kvm_mmu_memory_cache *cache;
1487
1488	cache = &vcpu->arch.mmu_pte_list_desc_cache;
1489	return mmu_memory_cache_free_objects(cache);
1490}
1491
1492static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
1493{
1494	struct kvm_mmu_page *sp;
1495	struct kvm_rmap_head *rmap_head;
1496
1497	sp = page_header(__pa(spte));
1498	kvm_mmu_page_set_gfn(sp, spte - sp->spt, gfn);
1499	rmap_head = gfn_to_rmap(vcpu->kvm, gfn, sp);
1500	return pte_list_add(vcpu, spte, rmap_head);
1501}
1502
1503static void rmap_remove(struct kvm *kvm, u64 *spte)
1504{
1505	struct kvm_mmu_page *sp;
1506	gfn_t gfn;
1507	struct kvm_rmap_head *rmap_head;
1508
1509	sp = page_header(__pa(spte));
1510	gfn = kvm_mmu_page_get_gfn(sp, spte - sp->spt);
1511	rmap_head = gfn_to_rmap(kvm, gfn, sp);
1512	__pte_list_remove(spte, rmap_head);
1513}
1514
1515/*
1516 * Used by the following functions to iterate through the sptes linked by a
1517 * rmap.  All fields are private and not assumed to be used outside.
1518 */
1519struct rmap_iterator {
1520	/* private fields */
1521	struct pte_list_desc *desc;	/* holds the sptep if not NULL */
1522	int pos;			/* index of the sptep */
1523};
1524
1525/*
1526 * Iteration must be started by this function.  This should also be used after
1527 * removing/dropping sptes from the rmap link because in such cases the
1528 * information in the itererator may not be valid.
1529 *
1530 * Returns sptep if found, NULL otherwise.
1531 */
1532static u64 *rmap_get_first(struct kvm_rmap_head *rmap_head,
1533			   struct rmap_iterator *iter)
1534{
1535	u64 *sptep;
1536
1537	if (!rmap_head->val)
1538		return NULL;
1539
1540	if (!(rmap_head->val & 1)) {
1541		iter->desc = NULL;
1542		sptep = (u64 *)rmap_head->val;
1543		goto out;
1544	}
1545
1546	iter->desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
1547	iter->pos = 0;
1548	sptep = iter->desc->sptes[iter->pos];
1549out:
1550	BUG_ON(!is_shadow_present_pte(*sptep));
1551	return sptep;
1552}
1553
1554/*
1555 * Must be used with a valid iterator: e.g. after rmap_get_first().
1556 *
1557 * Returns sptep if found, NULL otherwise.
1558 */
1559static u64 *rmap_get_next(struct rmap_iterator *iter)
1560{
1561	u64 *sptep;
1562
1563	if (iter->desc) {
1564		if (iter->pos < PTE_LIST_EXT - 1) {
1565			++iter->pos;
1566			sptep = iter->desc->sptes[iter->pos];
1567			if (sptep)
1568				goto out;
1569		}
1570
1571		iter->desc = iter->desc->more;
1572
1573		if (iter->desc) {
1574			iter->pos = 0;
1575			/* desc->sptes[0] cannot be NULL */
1576			sptep = iter->desc->sptes[iter->pos];
1577			goto out;
1578		}
1579	}
1580
1581	return NULL;
1582out:
1583	BUG_ON(!is_shadow_present_pte(*sptep));
1584	return sptep;
1585}
1586
1587#define for_each_rmap_spte(_rmap_head_, _iter_, _spte_)			\
1588	for (_spte_ = rmap_get_first(_rmap_head_, _iter_);		\
1589	     _spte_; _spte_ = rmap_get_next(_iter_))
1590
1591static void drop_spte(struct kvm *kvm, u64 *sptep)
1592{
1593	if (mmu_spte_clear_track_bits(sptep))
1594		rmap_remove(kvm, sptep);
1595}
1596
1597
1598static bool __drop_large_spte(struct kvm *kvm, u64 *sptep)
1599{
1600	if (is_large_pte(*sptep)) {
1601		WARN_ON(page_header(__pa(sptep))->role.level ==
1602			PT_PAGE_TABLE_LEVEL);
1603		drop_spte(kvm, sptep);
1604		--kvm->stat.lpages;
1605		return true;
1606	}
1607
1608	return false;
1609}
1610
1611static void drop_large_spte(struct kvm_vcpu *vcpu, u64 *sptep)
1612{
1613	if (__drop_large_spte(vcpu->kvm, sptep)) {
1614		struct kvm_mmu_page *sp = page_header(__pa(sptep));
1615
1616		kvm_flush_remote_tlbs_with_address(vcpu->kvm, sp->gfn,
1617			KVM_PAGES_PER_HPAGE(sp->role.level));
1618	}
1619}
1620
1621/*
1622 * Write-protect on the specified @sptep, @pt_protect indicates whether
1623 * spte write-protection is caused by protecting shadow page table.
1624 *
1625 * Note: write protection is difference between dirty logging and spte
1626 * protection:
1627 * - for dirty logging, the spte can be set to writable at anytime if
1628 *   its dirty bitmap is properly set.
1629 * - for spte protection, the spte can be writable only after unsync-ing
1630 *   shadow page.
1631 *
1632 * Return true if tlb need be flushed.
1633 */
1634static bool spte_write_protect(u64 *sptep, bool pt_protect)
1635{
1636	u64 spte = *sptep;
1637
1638	if (!is_writable_pte(spte) &&
1639	      !(pt_protect && spte_can_locklessly_be_made_writable(spte)))
1640		return false;
1641
1642	rmap_printk("rmap_write_protect: spte %p %llx\n", sptep, *sptep);
1643
1644	if (pt_protect)
1645		spte &= ~SPTE_MMU_WRITEABLE;
1646	spte = spte & ~PT_WRITABLE_MASK;
1647
1648	return mmu_spte_update(sptep, spte);
1649}
1650
1651static bool __rmap_write_protect(struct kvm *kvm,
1652				 struct kvm_rmap_head *rmap_head,
1653				 bool pt_protect)
1654{
1655	u64 *sptep;
1656	struct rmap_iterator iter;
1657	bool flush = false;
1658
1659	for_each_rmap_spte(rmap_head, &iter, sptep)
1660		flush |= spte_write_protect(sptep, pt_protect);
1661
1662	return flush;
1663}
1664
1665static bool spte_clear_dirty(u64 *sptep)
1666{
1667	u64 spte = *sptep;
1668
1669	rmap_printk("rmap_clear_dirty: spte %p %llx\n", sptep, *sptep);
1670
1671	MMU_WARN_ON(!spte_ad_enabled(spte));
1672	spte &= ~shadow_dirty_mask;
1673	return mmu_spte_update(sptep, spte);
1674}
1675
1676static bool spte_wrprot_for_clear_dirty(u64 *sptep)
1677{
1678	bool was_writable = test_and_clear_bit(PT_WRITABLE_SHIFT,
1679					       (unsigned long *)sptep);
1680	if (was_writable && !spte_ad_enabled(*sptep))
1681		kvm_set_pfn_dirty(spte_to_pfn(*sptep));
1682
1683	return was_writable;
1684}
1685
1686/*
1687 * Gets the GFN ready for another round of dirty logging by clearing the
1688 *	- D bit on ad-enabled SPTEs, and
1689 *	- W bit on ad-disabled SPTEs.
1690 * Returns true iff any D or W bits were cleared.
1691 */
1692static bool __rmap_clear_dirty(struct kvm *kvm, struct kvm_rmap_head *rmap_head)
1693{
1694	u64 *sptep;
1695	struct rmap_iterator iter;
1696	bool flush = false;
1697
1698	for_each_rmap_spte(rmap_head, &iter, sptep)
1699		if (spte_ad_need_write_protect(*sptep))
1700			flush |= spte_wrprot_for_clear_dirty(sptep);
1701		else
1702			flush |= spte_clear_dirty(sptep);
1703
1704	return flush;
1705}
1706
1707static bool spte_set_dirty(u64 *sptep)
1708{
1709	u64 spte = *sptep;
1710
1711	rmap_printk("rmap_set_dirty: spte %p %llx\n", sptep, *sptep);
1712
1713	/*
1714	 * Similar to the !kvm_x86_ops->slot_disable_log_dirty case,
1715	 * do not bother adding back write access to pages marked
1716	 * SPTE_AD_WRPROT_ONLY_MASK.
1717	 */
1718	spte |= shadow_dirty_mask;
1719
1720	return mmu_spte_update(sptep, spte);
1721}
1722
1723static bool __rmap_set_dirty(struct kvm *kvm, struct kvm_rmap_head *rmap_head)
1724{
1725	u64 *sptep;
1726	struct rmap_iterator iter;
1727	bool flush = false;
1728
1729	for_each_rmap_spte(rmap_head, &iter, sptep)
1730		if (spte_ad_enabled(*sptep))
1731			flush |= spte_set_dirty(sptep);
1732
1733	return flush;
1734}
1735
1736/**
1737 * kvm_mmu_write_protect_pt_masked - write protect selected PT level pages
1738 * @kvm: kvm instance
1739 * @slot: slot to protect
1740 * @gfn_offset: start of the BITS_PER_LONG pages we care about
1741 * @mask: indicates which pages we should protect
1742 *
1743 * Used when we do not need to care about huge page mappings: e.g. during dirty
1744 * logging we do not have any such mappings.
1745 */
1746static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1747				     struct kvm_memory_slot *slot,
1748				     gfn_t gfn_offset, unsigned long mask)
1749{
1750	struct kvm_rmap_head *rmap_head;
1751
1752	while (mask) {
1753		rmap_head = __gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask),
1754					  PT_PAGE_TABLE_LEVEL, slot);
1755		__rmap_write_protect(kvm, rmap_head, false);
1756
1757		/* clear the first set bit */
1758		mask &= mask - 1;
1759	}
1760}
1761
1762/**
1763 * kvm_mmu_clear_dirty_pt_masked - clear MMU D-bit for PT level pages, or write
1764 * protect the page if the D-bit isn't supported.
1765 * @kvm: kvm instance
1766 * @slot: slot to clear D-bit
1767 * @gfn_offset: start of the BITS_PER_LONG pages we care about
1768 * @mask: indicates which pages we should clear D-bit
1769 *
1770 * Used for PML to re-log the dirty GPAs after userspace querying dirty_bitmap.
1771 */
1772void kvm_mmu_clear_dirty_pt_masked(struct kvm *kvm,
1773				     struct kvm_memory_slot *slot,
1774				     gfn_t gfn_offset, unsigned long mask)
1775{
1776	struct kvm_rmap_head *rmap_head;
1777
1778	while (mask) {
1779		rmap_head = __gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask),
1780					  PT_PAGE_TABLE_LEVEL, slot);
1781		__rmap_clear_dirty(kvm, rmap_head);
1782
1783		/* clear the first set bit */
1784		mask &= mask - 1;
1785	}
1786}
1787EXPORT_SYMBOL_GPL(kvm_mmu_clear_dirty_pt_masked);
1788
1789/**
1790 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1791 * PT level pages.
1792 *
1793 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1794 * enable dirty logging for them.
1795 *
1796 * Used when we do not need to care about huge page mappings: e.g. during dirty
1797 * logging we do not have any such mappings.
1798 */
1799void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1800				struct kvm_memory_slot *slot,
1801				gfn_t gfn_offset, unsigned long mask)
1802{
1803	if (kvm_x86_ops->enable_log_dirty_pt_masked)
1804		kvm_x86_ops->enable_log_dirty_pt_masked(kvm, slot, gfn_offset,
1805				mask);
1806	else
1807		kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1808}
1809
1810/**
1811 * kvm_arch_write_log_dirty - emulate dirty page logging
1812 * @vcpu: Guest mode vcpu
1813 *
1814 * Emulate arch specific page modification logging for the
1815 * nested hypervisor
1816 */
1817int kvm_arch_write_log_dirty(struct kvm_vcpu *vcpu)
1818{
1819	if (kvm_x86_ops->write_log_dirty)
1820		return kvm_x86_ops->write_log_dirty(vcpu);
1821
1822	return 0;
1823}
1824
1825bool kvm_mmu_slot_gfn_write_protect(struct kvm *kvm,
1826				    struct kvm_memory_slot *slot, u64 gfn)
1827{
1828	struct kvm_rmap_head *rmap_head;
1829	int i;
1830	bool write_protected = false;
1831
1832	for (i = PT_PAGE_TABLE_LEVEL; i <= PT_MAX_HUGEPAGE_LEVEL; ++i) {
1833		rmap_head = __gfn_to_rmap(gfn, i, slot);
1834		write_protected |= __rmap_write_protect(kvm, rmap_head, true);
1835	}
1836
1837	return write_protected;
1838}
1839
1840static bool rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
1841{
1842	struct kvm_memory_slot *slot;
1843
1844	slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1845	return kvm_mmu_slot_gfn_write_protect(vcpu->kvm, slot, gfn);
1846}
1847
1848static bool kvm_zap_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head)
1849{
1850	u64 *sptep;
1851	struct rmap_iterator iter;
1852	bool flush = false;
1853
1854	while ((sptep = rmap_get_first(rmap_head, &iter))) {
1855		rmap_printk("%s: spte %p %llx.\n", __func__, sptep, *sptep);
1856
1857		pte_list_remove(rmap_head, sptep);
1858		flush = true;
1859	}
1860
1861	return flush;
1862}
1863
1864static int kvm_unmap_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1865			   struct kvm_memory_slot *slot, gfn_t gfn, int level,
1866			   unsigned long data)
1867{
1868	return kvm_zap_rmapp(kvm, rmap_head);
1869}
1870
1871static int kvm_set_pte_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1872			     struct kvm_memory_slot *slot, gfn_t gfn, int level,
1873			     unsigned long data)
1874{
1875	u64 *sptep;
1876	struct rmap_iterator iter;
1877	int need_flush = 0;
1878	u64 new_spte;
1879	pte_t *ptep = (pte_t *)data;
1880	kvm_pfn_t new_pfn;
1881
1882	WARN_ON(pte_huge(*ptep));
1883	new_pfn = pte_pfn(*ptep);
1884
1885restart:
1886	for_each_rmap_spte(rmap_head, &iter, sptep) {
1887		rmap_printk("kvm_set_pte_rmapp: spte %p %llx gfn %llx (%d)\n",
1888			    sptep, *sptep, gfn, level);
1889
1890		need_flush = 1;
1891
1892		if (pte_write(*ptep)) {
1893			pte_list_remove(rmap_head, sptep);
1894			goto restart;
1895		} else {
1896			new_spte = *sptep & ~PT64_BASE_ADDR_MASK;
1897			new_spte |= (u64)new_pfn << PAGE_SHIFT;
1898
1899			new_spte &= ~PT_WRITABLE_MASK;
1900			new_spte &= ~SPTE_HOST_WRITEABLE;
1901
1902			new_spte = mark_spte_for_access_track(new_spte);
1903
1904			mmu_spte_clear_track_bits(sptep);
1905			mmu_spte_set(sptep, new_spte);
1906		}
1907	}
1908
1909	if (need_flush && kvm_available_flush_tlb_with_range()) {
1910		kvm_flush_remote_tlbs_with_address(kvm, gfn, 1);
1911		return 0;
1912	}
1913
1914	return need_flush;
1915}
1916
1917struct slot_rmap_walk_iterator {
1918	/* input fields. */
1919	struct kvm_memory_slot *slot;
1920	gfn_t start_gfn;
1921	gfn_t end_gfn;
1922	int start_level;
1923	int end_level;
1924
1925	/* output fields. */
1926	gfn_t gfn;
1927	struct kvm_rmap_head *rmap;
1928	int level;
1929
1930	/* private field. */
1931	struct kvm_rmap_head *end_rmap;
1932};
1933
1934static void
1935rmap_walk_init_level(struct slot_rmap_walk_iterator *iterator, int level)
1936{
1937	iterator->level = level;
1938	iterator->gfn = iterator->start_gfn;
1939	iterator->rmap = __gfn_to_rmap(iterator->gfn, level, iterator->slot);
1940	iterator->end_rmap = __gfn_to_rmap(iterator->end_gfn, level,
1941					   iterator->slot);
1942}
1943
1944static void
1945slot_rmap_walk_init(struct slot_rmap_walk_iterator *iterator,
1946		    struct kvm_memory_slot *slot, int start_level,
1947		    int end_level, gfn_t start_gfn, gfn_t end_gfn)
1948{
1949	iterator->slot = slot;
1950	iterator->start_level = start_level;
1951	iterator->end_level = end_level;
1952	iterator->start_gfn = start_gfn;
1953	iterator->end_gfn = end_gfn;
1954
1955	rmap_walk_init_level(iterator, iterator->start_level);
1956}
1957
1958static bool slot_rmap_walk_okay(struct slot_rmap_walk_iterator *iterator)
1959{
1960	return !!iterator->rmap;
1961}
1962
1963static void slot_rmap_walk_next(struct slot_rmap_walk_iterator *iterator)
1964{
1965	if (++iterator->rmap <= iterator->end_rmap) {
1966		iterator->gfn += (1UL << KVM_HPAGE_GFN_SHIFT(iterator->level));
1967		return;
1968	}
1969
1970	if (++iterator->level > iterator->end_level) {
1971		iterator->rmap = NULL;
1972		return;
1973	}
1974
1975	rmap_walk_init_level(iterator, iterator->level);
1976}
1977
1978#define for_each_slot_rmap_range(_slot_, _start_level_, _end_level_,	\
1979	   _start_gfn, _end_gfn, _iter_)				\
1980	for (slot_rmap_walk_init(_iter_, _slot_, _start_level_,		\
1981				 _end_level_, _start_gfn, _end_gfn);	\
1982	     slot_rmap_walk_okay(_iter_);				\
1983	     slot_rmap_walk_next(_iter_))
1984
1985static int kvm_handle_hva_range(struct kvm *kvm,
1986				unsigned long start,
1987				unsigned long end,
1988				unsigned long data,
1989				int (*handler)(struct kvm *kvm,
1990					       struct kvm_rmap_head *rmap_head,
1991					       struct kvm_memory_slot *slot,
1992					       gfn_t gfn,
1993					       int level,
1994					       unsigned long data))
1995{
1996	struct kvm_memslots *slots;
1997	struct kvm_memory_slot *memslot;
1998	struct slot_rmap_walk_iterator iterator;
1999	int ret = 0;
2000	int i;
2001
2002	for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
2003		slots = __kvm_memslots(kvm, i);
2004		kvm_for_each_memslot(memslot, slots) {
2005			unsigned long hva_start, hva_end;
2006			gfn_t gfn_start, gfn_end;
2007
2008			hva_start = max(start, memslot->userspace_addr);
2009			hva_end = min(end, memslot->userspace_addr +
2010				      (memslot->npages << PAGE_SHIFT));
2011			if (hva_start >= hva_end)
2012				continue;
2013			/*
2014			 * {gfn(page) | page intersects with [hva_start, hva_end)} =
2015			 * {gfn_start, gfn_start+1, ..., gfn_end-1}.
2016			 */
2017			gfn_start = hva_to_gfn_memslot(hva_start, memslot);
2018			gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
2019
2020			for_each_slot_rmap_range(memslot, PT_PAGE_TABLE_LEVEL,
2021						 PT_MAX_HUGEPAGE_LEVEL,
2022						 gfn_start, gfn_end - 1,
2023						 &iterator)
2024				ret |= handler(kvm, iterator.rmap, memslot,
2025					       iterator.gfn, iterator.level, data);
2026		}
2027	}
2028
2029	return ret;
2030}
2031
2032static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
2033			  unsigned long data,
2034			  int (*handler)(struct kvm *kvm,
2035					 struct kvm_rmap_head *rmap_head,
2036					 struct kvm_memory_slot *slot,
2037					 gfn_t gfn, int level,
2038					 unsigned long data))
2039{
2040	return kvm_handle_hva_range(kvm, hva, hva + 1, data, handler);
2041}
2042
2043int kvm_unmap_hva_range(struct kvm *kvm, unsigned long start, unsigned long end)
2044{
2045	return kvm_handle_hva_range(kvm, start, end, 0, kvm_unmap_rmapp);
2046}
2047
2048int kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
2049{
2050	return kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
2051}
2052
2053static int kvm_age_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
2054			 struct kvm_memory_slot *slot, gfn_t gfn, int level,
2055			 unsigned long data)
2056{
2057	u64 *sptep;
2058	struct rmap_iterator uninitialized_var(iter);
2059	int young = 0;
2060
2061	for_each_rmap_spte(rmap_head, &iter, sptep)
2062		young |= mmu_spte_age(sptep);
2063
2064	trace_kvm_age_page(gfn, level, slot, young);
2065	return young;
2066}
2067
2068static int kvm_test_age_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
2069			      struct kvm_memory_slot *slot, gfn_t gfn,
2070			      int level, unsigned long data)
2071{
2072	u64 *sptep;
2073	struct rmap_iterator iter;
2074
2075	for_each_rmap_spte(rmap_head, &iter, sptep)
2076		if (is_accessed_spte(*sptep))
2077			return 1;
2078	return 0;
2079}
2080
2081#define RMAP_RECYCLE_THRESHOLD 1000
2082
2083static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
2084{
2085	struct kvm_rmap_head *rmap_head;
2086	struct kvm_mmu_page *sp;
2087
2088	sp = page_header(__pa(spte));
2089
2090	rmap_head = gfn_to_rmap(vcpu->kvm, gfn, sp);
2091
2092	kvm_unmap_rmapp(vcpu->kvm, rmap_head, NULL, gfn, sp->role.level, 0);
2093	kvm_flush_remote_tlbs_with_address(vcpu->kvm, sp->gfn,
2094			KVM_PAGES_PER_HPAGE(sp->role.level));
2095}
2096
2097int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
2098{
2099	return kvm_handle_hva_range(kvm, start, end, 0, kvm_age_rmapp);
2100}
2101
2102int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
2103{
2104	return kvm_handle_hva(kvm, hva, 0, kvm_test_age_rmapp);
2105}
2106
2107#ifdef MMU_DEBUG
2108static int is_empty_shadow_page(u64 *spt)
2109{
2110	u64 *pos;
2111	u64 *end;
2112
2113	for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
2114		if (is_shadow_present_pte(*pos)) {
2115			printk(KERN_ERR "%s: %p %llx\n", __func__,
2116			       pos, *pos);
2117			return 0;
2118		}
2119	return 1;
2120}
2121#endif
2122
2123/*
2124 * This value is the sum of all of the kvm instances's
2125 * kvm->arch.n_used_mmu_pages values.  We need a global,
2126 * aggregate version in order to make the slab shrinker
2127 * faster
2128 */
2129static inline void kvm_mod_used_mmu_pages(struct kvm *kvm, unsigned long nr)
2130{
2131	kvm->arch.n_used_mmu_pages += nr;
2132	percpu_counter_add(&kvm_total_used_mmu_pages, nr);
2133}
2134
2135static void kvm_mmu_free_page(struct kvm_mmu_page *sp)
2136{
2137	MMU_WARN_ON(!is_empty_shadow_page(sp->spt));
2138	hlist_del(&sp->hash_link);
2139	list_del(&sp->link);
2140	free_page((unsigned long)sp->spt);
2141	if (!sp->role.direct)
2142		free_page((unsigned long)sp->gfns);
2143	kmem_cache_free(mmu_page_header_cache, sp);
2144}
2145
2146static unsigned kvm_page_table_hashfn(gfn_t gfn)
2147{
2148	return hash_64(gfn, KVM_MMU_HASH_SHIFT);
2149}
2150
2151static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
2152				    struct kvm_mmu_page *sp, u64 *parent_pte)
2153{
2154	if (!parent_pte)
2155		return;
2156
2157	pte_list_add(vcpu, parent_pte, &sp->parent_ptes);
2158}
2159
2160static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
2161				       u64 *parent_pte)
2162{
2163	__pte_list_remove(parent_pte, &sp->parent_ptes);
2164}
2165
2166static void drop_parent_pte(struct kvm_mmu_page *sp,
2167			    u64 *parent_pte)
2168{
2169	mmu_page_remove_parent_pte(sp, parent_pte);
2170	mmu_spte_clear_no_track(parent_pte);
2171}
2172
2173static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu, int direct)
2174{
2175	struct kvm_mmu_page *sp;
2176
2177	sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache);
2178	sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache);
2179	if (!direct)
2180		sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache);
2181	set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
2182
2183	/*
2184	 * active_mmu_pages must be a FIFO list, as kvm_zap_obsolete_pages()
2185	 * depends on valid pages being added to the head of the list.  See
2186	 * comments in kvm_zap_obsolete_pages().
2187	 */
2188	sp->mmu_valid_gen = vcpu->kvm->arch.mmu_valid_gen;
2189	list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
2190	kvm_mod_used_mmu_pages(vcpu->kvm, +1);
2191	return sp;
2192}
2193
2194static void mark_unsync(u64 *spte);
2195static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
2196{
2197	u64 *sptep;
2198	struct rmap_iterator iter;
2199
2200	for_each_rmap_spte(&sp->parent_ptes, &iter, sptep) {
2201		mark_unsync(sptep);
2202	}
2203}
2204
2205static void mark_unsync(u64 *spte)
2206{
2207	struct kvm_mmu_page *sp;
2208	unsigned int index;
2209
2210	sp = page_header(__pa(spte));
2211	index = spte - sp->spt;
2212	if (__test_and_set_bit(index, sp->unsync_child_bitmap))
2213		return;
2214	if (sp->unsync_children++)
2215		return;
2216	kvm_mmu_mark_parents_unsync(sp);
2217}
2218
2219static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
2220			       struct kvm_mmu_page *sp)
2221{
2222	return 0;
2223}
2224
2225static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva, hpa_t root)
2226{
2227}
2228
2229static void nonpaging_update_pte(struct kvm_vcpu *vcpu,
2230				 struct kvm_mmu_page *sp, u64 *spte,
2231				 const void *pte)
2232{
2233	WARN_ON(1);
2234}
2235
2236#define KVM_PAGE_ARRAY_NR 16
2237
2238struct kvm_mmu_pages {
2239	struct mmu_page_and_offset {
2240		struct kvm_mmu_page *sp;
2241		unsigned int idx;
2242	} page[KVM_PAGE_ARRAY_NR];
2243	unsigned int nr;
2244};
2245
2246static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
2247			 int idx)
2248{
2249	int i;
2250
2251	if (sp->unsync)
2252		for (i=0; i < pvec->nr; i++)
2253			if (pvec->page[i].sp == sp)
2254				return 0;
2255
2256	pvec->page[pvec->nr].sp = sp;
2257	pvec->page[pvec->nr].idx = idx;
2258	pvec->nr++;
2259	return (pvec->nr == KVM_PAGE_ARRAY_NR);
2260}
2261
2262static inline void clear_unsync_child_bit(struct kvm_mmu_page *sp, int idx)
2263{
2264	--sp->unsync_children;
2265	WARN_ON((int)sp->unsync_children < 0);
2266	__clear_bit(idx, sp->unsync_child_bitmap);
2267}
2268
2269static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
2270			   struct kvm_mmu_pages *pvec)
2271{
2272	int i, ret, nr_unsync_leaf = 0;
2273
2274	for_each_set_bit(i, sp->unsync_child_bitmap, 512) {
2275		struct kvm_mmu_page *child;
2276		u64 ent = sp->spt[i];
2277
2278		if (!is_shadow_present_pte(ent) || is_large_pte(ent)) {
2279			clear_unsync_child_bit(sp, i);
2280			continue;
2281		}
2282
2283		child = page_header(ent & PT64_BASE_ADDR_MASK);
2284
2285		if (child->unsync_children) {
2286			if (mmu_pages_add(pvec, child, i))
2287				return -ENOSPC;
2288
2289			ret = __mmu_unsync_walk(child, pvec);
2290			if (!ret) {
2291				clear_unsync_child_bit(sp, i);
2292				continue;
2293			} else if (ret > 0) {
2294				nr_unsync_leaf += ret;
2295			} else
2296				return ret;
2297		} else if (child->unsync) {
2298			nr_unsync_leaf++;
2299			if (mmu_pages_add(pvec, child, i))
2300				return -ENOSPC;
2301		} else
2302			clear_unsync_child_bit(sp, i);
2303	}
2304
2305	return nr_unsync_leaf;
2306}
2307
2308#define INVALID_INDEX (-1)
2309
2310static int mmu_unsync_walk(struct kvm_mmu_page *sp,
2311			   struct kvm_mmu_pages *pvec)
2312{
2313	pvec->nr = 0;
2314	if (!sp->unsync_children)
2315		return 0;
2316
2317	mmu_pages_add(pvec, sp, INVALID_INDEX);
2318	return __mmu_unsync_walk(sp, pvec);
2319}
2320
2321static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
2322{
2323	WARN_ON(!sp->unsync);
2324	trace_kvm_mmu_sync_page(sp);
2325	sp->unsync = 0;
2326	--kvm->stat.mmu_unsync;
2327}
2328
2329static bool kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
2330				     struct list_head *invalid_list);
2331static void kvm_mmu_commit_zap_page(struct kvm *kvm,
2332				    struct list_head *invalid_list);
2333
2334
2335#define for_each_valid_sp(_kvm, _sp, _gfn)				\
2336	hlist_for_each_entry(_sp,					\
2337	  &(_kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(_gfn)], hash_link) \
2338		if (is_obsolete_sp((_kvm), (_sp))) {			\
2339		} else
2340
2341#define for_each_gfn_indirect_valid_sp(_kvm, _sp, _gfn)			\
2342	for_each_valid_sp(_kvm, _sp, _gfn)				\
2343		if ((_sp)->gfn != (_gfn) || (_sp)->role.direct) {} else
2344
2345static inline bool is_ept_sp(struct kvm_mmu_page *sp)
2346{
2347	return sp->role.cr0_wp && sp->role.smap_andnot_wp;
2348}
2349
2350/* @sp->gfn should be write-protected at the call site */
2351static bool __kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
2352			    struct list_head *invalid_list)
2353{
2354	if ((!is_ept_sp(sp) && sp->role.gpte_is_8_bytes != !!is_pae(vcpu)) ||
2355	    vcpu->arch.mmu->sync_page(vcpu, sp) == 0) {
2356		kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
2357		return false;
2358	}
2359
2360	return true;
2361}
2362
2363static bool kvm_mmu_remote_flush_or_zap(struct kvm *kvm,
2364					struct list_head *invalid_list,
2365					bool remote_flush)
2366{
2367	if (!remote_flush && list_empty(invalid_list))
2368		return false;
2369
2370	if (!list_empty(invalid_list))
2371		kvm_mmu_commit_zap_page(kvm, invalid_list);
2372	else
2373		kvm_flush_remote_tlbs(kvm);
2374	return true;
2375}
2376
2377static void kvm_mmu_flush_or_zap(struct kvm_vcpu *vcpu,
2378				 struct list_head *invalid_list,
2379				 bool remote_flush, bool local_flush)
2380{
2381	if (kvm_mmu_remote_flush_or_zap(vcpu->kvm, invalid_list, remote_flush))
2382		return;
2383
2384	if (local_flush)
2385		kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
2386}
2387
2388#ifdef CONFIG_KVM_MMU_AUDIT
2389#include "mmu_audit.c"
2390#else
2391static void kvm_mmu_audit(struct kvm_vcpu *vcpu, int point) { }
2392static void mmu_audit_disable(void) { }
2393#endif
2394
2395static bool is_obsolete_sp(struct kvm *kvm, struct kvm_mmu_page *sp)
2396{
2397	return sp->role.invalid ||
2398	       unlikely(sp->mmu_valid_gen != kvm->arch.mmu_valid_gen);
2399}
2400
2401static bool kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
2402			 struct list_head *invalid_list)
2403{
2404	kvm_unlink_unsync_page(vcpu->kvm, sp);
2405	return __kvm_sync_page(vcpu, sp, invalid_list);
2406}
2407
2408/* @gfn should be write-protected at the call site */
2409static bool kvm_sync_pages(struct kvm_vcpu *vcpu, gfn_t gfn,
2410			   struct list_head *invalid_list)
2411{
2412	struct kvm_mmu_page *s;
2413	bool ret = false;
2414
2415	for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn) {
2416		if (!s->unsync)
2417			continue;
2418
2419		WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
2420		ret |= kvm_sync_page(vcpu, s, invalid_list);
2421	}
2422
2423	return ret;
2424}
2425
2426struct mmu_page_path {
2427	struct kvm_mmu_page *parent[PT64_ROOT_MAX_LEVEL];
2428	unsigned int idx[PT64_ROOT_MAX_LEVEL];
2429};
2430
2431#define for_each_sp(pvec, sp, parents, i)			\
2432		for (i = mmu_pages_first(&pvec, &parents);	\
2433			i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});	\
2434			i = mmu_pages_next(&pvec, &parents, i))
2435
2436static int mmu_pages_next(struct kvm_mmu_pages *pvec,
2437			  struct mmu_page_path *parents,
2438			  int i)
2439{
2440	int n;
2441
2442	for (n = i+1; n < pvec->nr; n++) {
2443		struct kvm_mmu_page *sp = pvec->page[n].sp;
2444		unsigned idx = pvec->page[n].idx;
2445		int level = sp->role.level;
2446
2447		parents->idx[level-1] = idx;
2448		if (level == PT_PAGE_TABLE_LEVEL)
2449			break;
2450
2451		parents->parent[level-2] = sp;
2452	}
2453
2454	return n;
2455}
2456
2457static int mmu_pages_first(struct kvm_mmu_pages *pvec,
2458			   struct mmu_page_path *parents)
2459{
2460	struct kvm_mmu_page *sp;
2461	int level;
2462
2463	if (pvec->nr == 0)
2464		return 0;
2465
2466	WARN_ON(pvec->page[0].idx != INVALID_INDEX);
2467
2468	sp = pvec->page[0].sp;
2469	level = sp->role.level;
2470	WARN_ON(level == PT_PAGE_TABLE_LEVEL);
2471
2472	parents->parent[level-2] = sp;
2473
2474	/* Also set up a sentinel.  Further entries in pvec are all
2475	 * children of sp, so this element is never overwritten.
2476	 */
2477	parents->parent[level-1] = NULL;
2478	return mmu_pages_next(pvec, parents, 0);
2479}
2480
2481static void mmu_pages_clear_parents(struct mmu_page_path *parents)
2482{
2483	struct kvm_mmu_page *sp;
2484	unsigned int level = 0;
2485
2486	do {
2487		unsigned int idx = parents->idx[level];
2488		sp = parents->parent[level];
2489		if (!sp)
2490			return;
2491
2492		WARN_ON(idx == INVALID_INDEX);
2493		clear_unsync_child_bit(sp, idx);
2494		level++;
2495	} while (!sp->unsync_children);
2496}
2497
2498static void mmu_sync_children(struct kvm_vcpu *vcpu,
2499			      struct kvm_mmu_page *parent)
2500{
2501	int i;
2502	struct kvm_mmu_page *sp;
2503	struct mmu_page_path parents;
2504	struct kvm_mmu_pages pages;
2505	LIST_HEAD(invalid_list);
2506	bool flush = false;
2507
2508	while (mmu_unsync_walk(parent, &pages)) {
2509		bool protected = false;
2510
2511		for_each_sp(pages, sp, parents, i)
2512			protected |= rmap_write_protect(vcpu, sp->gfn);
2513
2514		if (protected) {
2515			kvm_flush_remote_tlbs(vcpu->kvm);
2516			flush = false;
2517		}
2518
2519		for_each_sp(pages, sp, parents, i) {
2520			flush |= kvm_sync_page(vcpu, sp, &invalid_list);
2521			mmu_pages_clear_parents(&parents);
2522		}
2523		if (need_resched() || spin_needbreak(&vcpu->kvm->mmu_lock)) {
2524			kvm_mmu_flush_or_zap(vcpu, &invalid_list, false, flush);
2525			cond_resched_lock(&vcpu->kvm->mmu_lock);
2526			flush = false;
2527		}
2528	}
2529
2530	kvm_mmu_flush_or_zap(vcpu, &invalid_list, false, flush);
2531}
2532
2533static void __clear_sp_write_flooding_count(struct kvm_mmu_page *sp)
2534{
2535	atomic_set(&sp->write_flooding_count,  0);
2536}
2537
2538static void clear_sp_write_flooding_count(u64 *spte)
2539{
2540	struct kvm_mmu_page *sp =  page_header(__pa(spte));
2541
2542	__clear_sp_write_flooding_count(sp);
2543}
2544
2545static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
2546					     gfn_t gfn,
2547					     gva_t gaddr,
2548					     unsigned level,
2549					     int direct,
2550					     unsigned access)
2551{
2552	union kvm_mmu_page_role role;
2553	unsigned quadrant;
2554	struct kvm_mmu_page *sp;
2555	bool need_sync = false;
2556	bool flush = false;
2557	int collisions = 0;
2558	LIST_HEAD(invalid_list);
2559
2560	role = vcpu->arch.mmu->mmu_role.base;
2561	role.level = level;
2562	role.direct = direct;
2563	if (role.direct)
2564		role.gpte_is_8_bytes = true;
2565	role.access = access;
2566	if (!vcpu->arch.mmu->direct_map
2567	    && vcpu->arch.mmu->root_level <= PT32_ROOT_LEVEL) {
2568		quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
2569		quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
2570		role.quadrant = quadrant;
2571	}
2572	for_each_valid_sp(vcpu->kvm, sp, gfn) {
2573		if (sp->gfn != gfn) {
2574			collisions++;
2575			continue;
2576		}
2577
2578		if (!need_sync && sp->unsync)
2579			need_sync = true;
2580
2581		if (sp->role.word != role.word)
2582			continue;
2583
2584		if (sp->unsync) {
2585			/* The page is good, but __kvm_sync_page might still end
2586			 * up zapping it.  If so, break in order to rebuild it.
2587			 */
2588			if (!__kvm_sync_page(vcpu, sp, &invalid_list))
2589				break;
2590
2591			WARN_ON(!list_empty(&invalid_list));
2592			kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
2593		}
2594
2595		if (sp->unsync_children)
2596			kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
2597
2598		__clear_sp_write_flooding_count(sp);
2599		trace_kvm_mmu_get_page(sp, false);
2600		goto out;
2601	}
2602
2603	++vcpu->kvm->stat.mmu_cache_miss;
2604
2605	sp = kvm_mmu_alloc_page(vcpu, direct);
2606
2607	sp->gfn = gfn;
2608	sp->role = role;
2609	hlist_add_head(&sp->hash_link,
2610		&vcpu->kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)]);
2611	if (!direct) {
2612		/*
2613		 * we should do write protection before syncing pages
2614		 * otherwise the content of the synced shadow page may
2615		 * be inconsistent with guest page table.
2616		 */
2617		account_shadowed(vcpu->kvm, sp);
2618		if (level == PT_PAGE_TABLE_LEVEL &&
2619		      rmap_write_protect(vcpu, gfn))
2620			kvm_flush_remote_tlbs_with_address(vcpu->kvm, gfn, 1);
2621
2622		if (level > PT_PAGE_TABLE_LEVEL && need_sync)
2623			flush |= kvm_sync_pages(vcpu, gfn, &invalid_list);
2624	}
2625	clear_page(sp->spt);
2626	trace_kvm_mmu_get_page(sp, true);
2627
2628	kvm_mmu_flush_or_zap(vcpu, &invalid_list, false, flush);
2629out:
2630	if (collisions > vcpu->kvm->stat.max_mmu_page_hash_collisions)
2631		vcpu->kvm->stat.max_mmu_page_hash_collisions = collisions;
2632	return sp;
2633}
2634
2635static void shadow_walk_init_using_root(struct kvm_shadow_walk_iterator *iterator,
2636					struct kvm_vcpu *vcpu, hpa_t root,
2637					u64 addr)
2638{
2639	iterator->addr = addr;
2640	iterator->shadow_addr = root;
2641	iterator->level = vcpu->arch.mmu->shadow_root_level;
2642
2643	if (iterator->level == PT64_ROOT_4LEVEL &&
2644	    vcpu->arch.mmu->root_level < PT64_ROOT_4LEVEL &&
2645	    !vcpu->arch.mmu->direct_map)
2646		--iterator->level;
2647
2648	if (iterator->level == PT32E_ROOT_LEVEL) {
2649		/*
2650		 * prev_root is currently only used for 64-bit hosts. So only
2651		 * the active root_hpa is valid here.
2652		 */
2653		BUG_ON(root != vcpu->arch.mmu->root_hpa);
2654
2655		iterator->shadow_addr
2656			= vcpu->arch.mmu->pae_root[(addr >> 30) & 3];
2657		iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
2658		--iterator->level;
2659		if (!iterator->shadow_addr)
2660			iterator->level = 0;
2661	}
2662}
2663
2664static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
2665			     struct kvm_vcpu *vcpu, u64 addr)
2666{
2667	shadow_walk_init_using_root(iterator, vcpu, vcpu->arch.mmu->root_hpa,
2668				    addr);
2669}
2670
2671static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
2672{
2673	if (iterator->level < PT_PAGE_TABLE_LEVEL)
2674		return false;
2675
2676	iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
2677	iterator->sptep	= ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
2678	return true;
2679}
2680
2681static void __shadow_walk_next(struct kvm_shadow_walk_iterator *iterator,
2682			       u64 spte)
2683{
2684	if (is_last_spte(spte, iterator->level)) {
2685		iterator->level = 0;
2686		return;
2687	}
2688
2689	iterator->shadow_addr = spte & PT64_BASE_ADDR_MASK;
2690	--iterator->level;
2691}
2692
2693static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
2694{
2695	__shadow_walk_next(iterator, *iterator->sptep);
2696}
2697
2698static void link_shadow_page(struct kvm_vcpu *vcpu, u64 *sptep,
2699			     struct kvm_mmu_page *sp)
2700{
2701	u64 spte;
2702
2703	BUILD_BUG_ON(VMX_EPT_WRITABLE_MASK != PT_WRITABLE_MASK);
2704
2705	spte = __pa(sp->spt) | shadow_present_mask | PT_WRITABLE_MASK |
2706	       shadow_user_mask | shadow_x_mask | shadow_me_mask;
2707
2708	if (sp_ad_disabled(sp))
2709		spte |= SPTE_AD_DISABLED_MASK;
2710	else
2711		spte |= shadow_accessed_mask;
2712
2713	mmu_spte_set(sptep, spte);
2714
2715	mmu_page_add_parent_pte(vcpu, sp, sptep);
2716
2717	if (sp->unsync_children || sp->unsync)
2718		mark_unsync(sptep);
2719}
2720
2721static void validate_direct_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2722				   unsigned direct_access)
2723{
2724	if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) {
2725		struct kvm_mmu_page *child;
2726
2727		/*
2728		 * For the direct sp, if the guest pte's dirty bit
2729		 * changed form clean to dirty, it will corrupt the
2730		 * sp's access: allow writable in the read-only sp,
2731		 * so we should update the spte at this point to get
2732		 * a new sp with the correct access.
2733		 */
2734		child = page_header(*sptep & PT64_BASE_ADDR_MASK);
2735		if (child->role.access == direct_access)
2736			return;
2737
2738		drop_parent_pte(child, sptep);
2739		kvm_flush_remote_tlbs_with_address(vcpu->kvm, child->gfn, 1);
2740	}
2741}
2742
2743static bool mmu_page_zap_pte(struct kvm *kvm, struct kvm_mmu_page *sp,
2744			     u64 *spte)
2745{
2746	u64 pte;
2747	struct kvm_mmu_page *child;
2748
2749	pte = *spte;
2750	if (is_shadow_present_pte(pte)) {
2751		if (is_last_spte(pte, sp->role.level)) {
2752			drop_spte(kvm, spte);
2753			if (is_large_pte(pte))
2754				--kvm->stat.lpages;
2755		} else {
2756			child = page_header(pte & PT64_BASE_ADDR_MASK);
2757			drop_parent_pte(child, spte);
2758		}
2759		return true;
2760	}
2761
2762	if (is_mmio_spte(pte))
2763		mmu_spte_clear_no_track(spte);
2764
2765	return false;
2766}
2767
2768static void kvm_mmu_page_unlink_children(struct kvm *kvm,
2769					 struct kvm_mmu_page *sp)
2770{
2771	unsigned i;
2772
2773	for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2774		mmu_page_zap_pte(kvm, sp, sp->spt + i);
2775}
2776
2777static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
2778{
2779	u64 *sptep;
2780	struct rmap_iterator iter;
2781
2782	while ((sptep = rmap_get_first(&sp->parent_ptes, &iter)))
2783		drop_parent_pte(sp, sptep);
2784}
2785
2786static int mmu_zap_unsync_children(struct kvm *kvm,
2787				   struct kvm_mmu_page *parent,
2788				   struct list_head *invalid_list)
2789{
2790	int i, zapped = 0;
2791	struct mmu_page_path parents;
2792	struct kvm_mmu_pages pages;
2793
2794	if (parent->role.level == PT_PAGE_TABLE_LEVEL)
2795		return 0;
2796
2797	while (mmu_unsync_walk(parent, &pages)) {
2798		struct kvm_mmu_page *sp;
2799
2800		for_each_sp(pages, sp, parents, i) {
2801			kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
2802			mmu_pages_clear_parents(&parents);
2803			zapped++;
2804		}
2805	}
2806
2807	return zapped;
2808}
2809
2810static bool __kvm_mmu_prepare_zap_page(struct kvm *kvm,
2811				       struct kvm_mmu_page *sp,
2812				       struct list_head *invalid_list,
2813				       int *nr_zapped)
2814{
2815	bool list_unstable;
2816
2817	trace_kvm_mmu_prepare_zap_page(sp);
2818	++kvm->stat.mmu_shadow_zapped;
2819	*nr_zapped = mmu_zap_unsync_children(kvm, sp, invalid_list);
2820	kvm_mmu_page_unlink_children(kvm, sp);
2821	kvm_mmu_unlink_parents(kvm, sp);
2822
2823	/* Zapping children means active_mmu_pages has become unstable. */
2824	list_unstable = *nr_zapped;
2825
2826	if (!sp->role.invalid && !sp->role.direct)
2827		unaccount_shadowed(kvm, sp);
2828
2829	if (sp->unsync)
2830		kvm_unlink_unsync_page(kvm, sp);
2831	if (!sp->root_count) {
2832		/* Count self */
2833		(*nr_zapped)++;
2834		list_move(&sp->link, invalid_list);
2835		kvm_mod_used_mmu_pages(kvm, -1);
2836	} else {
2837		list_move(&sp->link, &kvm->arch.active_mmu_pages);
2838
2839		/*
2840		 * Obsolete pages cannot be used on any vCPUs, see the comment
2841		 * in kvm_mmu_zap_all_fast().  Note, is_obsolete_sp() also
2842		 * treats invalid shadow pages as being obsolete.
2843		 */
2844		if (!is_obsolete_sp(kvm, sp))
2845			kvm_reload_remote_mmus(kvm);
2846	}
2847
2848	if (sp->lpage_disallowed)
2849		unaccount_huge_nx_page(kvm, sp);
2850
2851	sp->role.invalid = 1;
2852	return list_unstable;
2853}
2854
2855static bool kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
2856				     struct list_head *invalid_list)
2857{
2858	int nr_zapped;
2859
2860	__kvm_mmu_prepare_zap_page(kvm, sp, invalid_list, &nr_zapped);
2861	return nr_zapped;
2862}
2863
2864static void kvm_mmu_commit_zap_page(struct kvm *kvm,
2865				    struct list_head *invalid_list)
2866{
2867	struct kvm_mmu_page *sp, *nsp;
2868
2869	if (list_empty(invalid_list))
2870		return;
2871
2872	/*
2873	 * We need to make sure everyone sees our modifications to
2874	 * the page tables and see changes to vcpu->mode here. The barrier
2875	 * in the kvm_flush_remote_tlbs() achieves this. This pairs
2876	 * with vcpu_enter_guest and walk_shadow_page_lockless_begin/end.
2877	 *
2878	 * In addition, kvm_flush_remote_tlbs waits for all vcpus to exit
2879	 * guest mode and/or lockless shadow page table walks.
2880	 */
2881	kvm_flush_remote_tlbs(kvm);
2882
2883	list_for_each_entry_safe(sp, nsp, invalid_list, link) {
2884		WARN_ON(!sp->role.invalid || sp->root_count);
2885		kvm_mmu_free_page(sp);
2886	}
2887}
2888
2889static bool prepare_zap_oldest_mmu_page(struct kvm *kvm,
2890					struct list_head *invalid_list)
2891{
2892	struct kvm_mmu_page *sp;
2893
2894	if (list_empty(&kvm->arch.active_mmu_pages))
2895		return false;
2896
2897	sp = list_last_entry(&kvm->arch.active_mmu_pages,
2898			     struct kvm_mmu_page, link);
2899	return kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
2900}
2901
2902/*
2903 * Changing the number of mmu pages allocated to the vm
2904 * Note: if goal_nr_mmu_pages is too small, you will get dead lock
2905 */
2906void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned long goal_nr_mmu_pages)
2907{
2908	LIST_HEAD(invalid_list);
2909
2910	spin_lock(&kvm->mmu_lock);
2911
2912	if (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages) {
2913		/* Need to free some mmu pages to achieve the goal. */
2914		while (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages)
2915			if (!prepare_zap_oldest_mmu_page(kvm, &invalid_list))
2916				break;
2917
2918		kvm_mmu_commit_zap_page(kvm, &invalid_list);
2919		goal_nr_mmu_pages = kvm->arch.n_used_mmu_pages;
2920	}
2921
2922	kvm->arch.n_max_mmu_pages = goal_nr_mmu_pages;
2923
2924	spin_unlock(&kvm->mmu_lock);
2925}
2926
2927int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
2928{
2929	struct kvm_mmu_page *sp;
2930	LIST_HEAD(invalid_list);
2931	int r;
2932
2933	pgprintk("%s: looking for gfn %llx\n", __func__, gfn);
2934	r = 0;
2935	spin_lock(&kvm->mmu_lock);
2936	for_each_gfn_indirect_valid_sp(kvm, sp, gfn) {
2937		pgprintk("%s: gfn %llx role %x\n", __func__, gfn,
2938			 sp->role.word);
2939		r = 1;
2940		kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
2941	}
2942	kvm_mmu_commit_zap_page(kvm, &invalid_list);
2943	spin_unlock(&kvm->mmu_lock);
2944
2945	return r;
2946}
2947EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page);
2948
2949static void kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
2950{
2951	trace_kvm_mmu_unsync_page(sp);
2952	++vcpu->kvm->stat.mmu_unsync;
2953	sp->unsync = 1;
2954
2955	kvm_mmu_mark_parents_unsync(sp);
2956}
2957
2958static bool mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
2959				   bool can_unsync)
2960{
2961	struct kvm_mmu_page *sp;
2962
2963	if (kvm_page_track_is_active(vcpu, gfn, KVM_PAGE_TRACK_WRITE))
2964		return true;
2965
2966	for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn) {
2967		if (!can_unsync)
2968			return true;
2969
2970		if (sp->unsync)
2971			continue;
2972
2973		WARN_ON(sp->role.level != PT_PAGE_TABLE_LEVEL);
2974		kvm_unsync_page(vcpu, sp);
2975	}
2976
2977	/*
2978	 * We need to ensure that the marking of unsync pages is visible
2979	 * before the SPTE is updated to allow writes because
2980	 * kvm_mmu_sync_roots() checks the unsync flags without holding
2981	 * the MMU lock and so can race with this. If the SPTE was updated
2982	 * before the page had been marked as unsync-ed, something like the
2983	 * following could happen:
2984	 *
2985	 * CPU 1                    CPU 2
2986	 * ---------------------------------------------------------------------
2987	 * 1.2 Host updates SPTE
2988	 *     to be writable
2989	 *                      2.1 Guest writes a GPTE for GVA X.
2990	 *                          (GPTE being in the guest page table shadowed
2991	 *                           by the SP from CPU 1.)
2992	 *                          This reads SPTE during the page table walk.
2993	 *                          Since SPTE.W is read as 1, there is no
2994	 *                          fault.
2995	 *
2996	 *                      2.2 Guest issues TLB flush.
2997	 *                          That causes a VM Exit.
2998	 *
2999	 *                      2.3 kvm_mmu_sync_pages() reads sp->unsync.
3000	 *                          Since it is false, so it just returns.
3001	 *
3002	 *                      2.4 Guest accesses GVA X.
3003	 *                          Since the mapping in the SP was not updated,
3004	 *                          so the old mapping for GVA X incorrectly
3005	 *                          gets used.
3006	 * 1.1 Host marks SP
3007	 *     as unsync
3008	 *     (sp->unsync = true)
3009	 *
3010	 * The write barrier below ensures that 1.1 happens before 1.2 and thus
3011	 * the situation in 2.4 does not arise. The implicit barrier in 2.2
3012	 * pairs with this write barrier.
3013	 */
3014	smp_wmb();
3015
3016	return false;
3017}
3018
3019static bool kvm_is_mmio_pfn(kvm_pfn_t pfn)
3020{
3021	if (pfn_valid(pfn))
3022		return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) &&
3023			/*
3024			 * Some reserved pages, such as those from NVDIMM
3025			 * DAX devices, are not for MMIO, and can be mapped
3026			 * with cached memory type for better performance.
3027			 * However, the above check misconceives those pages
3028			 * as MMIO, and results in KVM mapping them with UC
3029			 * memory type, which would hurt the performance.
3030			 * Therefore, we check the host memory type in addition
3031			 * and only treat UC/UC-/WC pages as MMIO.
3032			 */
3033			(!pat_enabled() || pat_pfn_immune_to_uc_mtrr(pfn));
3034
3035	return !e820__mapped_raw_any(pfn_to_hpa(pfn),
3036				     pfn_to_hpa(pfn + 1) - 1,
3037				     E820_TYPE_RAM);
3038}
3039
3040/* Bits which may be returned by set_spte() */
3041#define SET_SPTE_WRITE_PROTECTED_PT	BIT(0)
3042#define SET_SPTE_NEED_REMOTE_TLB_FLUSH	BIT(1)
3043
3044static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
3045		    unsigned pte_access, int level,
3046		    gfn_t gfn, kvm_pfn_t pfn, bool speculative,
3047		    bool can_unsync, bool host_writable)
3048{
3049	u64 spte = 0;
3050	int ret = 0;
3051	struct kvm_mmu_page *sp;
3052
3053	if (set_mmio_spte(vcpu, sptep, gfn, pfn, pte_access))
3054		return 0;
3055
3056	sp = page_header(__pa(sptep));
3057	if (sp_ad_disabled(sp))
3058		spte |= SPTE_AD_DISABLED_MASK;
3059	else if (kvm_vcpu_ad_need_write_protect(vcpu))
3060		spte |= SPTE_AD_WRPROT_ONLY_MASK;
3061
3062	/*
3063	 * For the EPT case, shadow_present_mask is 0 if hardware
3064	 * supports exec-only page table entries.  In that case,
3065	 * ACC_USER_MASK and shadow_user_mask are used to represent
3066	 * read access.  See FNAME(gpte_access) in paging_tmpl.h.
3067	 */
3068	spte |= shadow_present_mask;
3069	if (!speculative)
3070		spte |= spte_shadow_accessed_mask(spte);
3071
3072	if (level > PT_PAGE_TABLE_LEVEL && (pte_access & ACC_EXEC_MASK) &&
3073	    is_nx_huge_page_enabled()) {
3074		pte_access &= ~ACC_EXEC_MASK;
3075	}
3076
3077	if (pte_access & ACC_EXEC_MASK)
3078		spte |= shadow_x_mask;
3079	else
3080		spte |= shadow_nx_mask;
3081
3082	if (pte_access & ACC_USER_MASK)
3083		spte |= shadow_user_mask;
3084
3085	if (level > PT_PAGE_TABLE_LEVEL)
3086		spte |= PT_PAGE_SIZE_MASK;
3087	if (tdp_enabled)
3088		spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
3089			kvm_is_mmio_pfn(pfn));
3090
3091	if (host_writable)
3092		spte |= SPTE_HOST_WRITEABLE;
3093	else
3094		pte_access &= ~ACC_WRITE_MASK;
3095
3096	if (!kvm_is_mmio_pfn(pfn))
3097		spte |= shadow_me_mask;
3098
3099	spte |= (u64)pfn << PAGE_SHIFT;
3100
3101	if (pte_access & ACC_WRITE_MASK) {
3102
3103		/*
3104		 * Other vcpu creates new sp in the window between
3105		 * mapping_level() and acquiring mmu-lock. We can
3106		 * allow guest to retry the access, the mapping can
3107		 * be fixed if guest refault.
3108		 */
3109		if (level > PT_PAGE_TABLE_LEVEL &&
3110		    mmu_gfn_lpage_is_disallowed(vcpu, gfn, level))
3111			goto done;
3112
3113		spte |= PT_WRITABLE_MASK | SPTE_MMU_WRITEABLE;
3114
3115		/*
3116		 * Optimization: for pte sync, if spte was writable the hash
3117		 * lookup is unnecessary (and expensive). Write protection
3118		 * is responsibility of mmu_get_page / kvm_sync_page.
3119		 * Same reasoning can be applied to dirty page accounting.
3120		 */
3121		if (!can_unsync && is_writable_pte(*sptep))
3122			goto set_pte;
3123
3124		if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
3125			pgprintk("%s: found shadow page for %llx, marking ro\n",
3126				 __func__, gfn);
3127			ret |= SET_SPTE_WRITE_PROTECTED_PT;
3128			pte_access &= ~ACC_WRITE_MASK;
3129			spte &= ~(PT_WRITABLE_MASK | SPTE_MMU_WRITEABLE);
3130		}
3131	}
3132
3133	if (pte_access & ACC_WRITE_MASK) {
3134		kvm_vcpu_mark_page_dirty(vcpu, gfn);
3135		spte |= spte_shadow_dirty_mask(spte);
3136	}
3137
3138	if (speculative)
3139		spte = mark_spte_for_access_track(spte);
3140
3141set_pte:
3142	if (mmu_spte_update(sptep, spte))
3143		ret |= SET_SPTE_NEED_REMOTE_TLB_FLUSH;
3144done:
3145	return ret;
3146}
3147
3148static int mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep, unsigned pte_access,
3149			int write_fault, int level, gfn_t gfn, kvm_pfn_t pfn,
3150		       	bool speculative, bool host_writable)
3151{
3152	int was_rmapped = 0;
3153	int rmap_count;
3154	int set_spte_ret;
3155	int ret = RET_PF_RETRY;
3156	bool flush = false;
3157
3158	pgprintk("%s: spte %llx write_fault %d gfn %llx\n", __func__,
3159		 *sptep, write_fault, gfn);
3160
3161	if (is_shadow_present_pte(*sptep)) {
3162		/*
3163		 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
3164		 * the parent of the now unreachable PTE.
3165		 */
3166		if (level > PT_PAGE_TABLE_LEVEL &&
3167		    !is_large_pte(*sptep)) {
3168			struct kvm_mmu_page *child;
3169			u64 pte = *sptep;
3170
3171			child = page_header(pte & PT64_BASE_ADDR_MASK);
3172			drop_parent_pte(child, sptep);
3173			flush = true;
3174		} else if (pfn != spte_to_pfn(*sptep)) {
3175			pgprintk("hfn old %llx new %llx\n",
3176				 spte_to_pfn(*sptep), pfn);
3177			drop_spte(vcpu->kvm, sptep);
3178			flush = true;
3179		} else
3180			was_rmapped = 1;
3181	}
3182
3183	set_spte_ret = set_spte(vcpu, sptep, pte_access, level, gfn, pfn,
3184				speculative, true, host_writable);
3185	if (set_spte_ret & SET_SPTE_WRITE_PROTECTED_PT) {
3186		if (write_fault)
3187			ret = RET_PF_EMULATE;
3188		kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
3189	}
3190
3191	if (set_spte_ret & SET_SPTE_NEED_REMOTE_TLB_FLUSH || flush)
3192		kvm_flush_remote_tlbs_with_address(vcpu->kvm, gfn,
3193				KVM_PAGES_PER_HPAGE(level));
3194
3195	if (unlikely(is_mmio_spte(*sptep)))
3196		ret = RET_PF_EMULATE;
3197
3198	pgprintk("%s: setting spte %llx\n", __func__, *sptep);
3199	trace_kvm_mmu_set_spte(level, gfn, sptep);
3200	if (!was_rmapped && is_large_pte(*sptep))
3201		++vcpu->kvm->stat.lpages;
3202
3203	if (is_shadow_present_pte(*sptep)) {
3204		if (!was_rmapped) {
3205			rmap_count = rmap_add(vcpu, sptep, gfn);
3206			if (rmap_count > RMAP_RECYCLE_THRESHOLD)
3207				rmap_recycle(vcpu, sptep, gfn);
3208		}
3209	}
3210
3211	return ret;
3212}
3213
3214static kvm_pfn_t pte_prefetch_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn,
3215				     bool no_dirty_log)
3216{
3217	struct kvm_memory_slot *slot;
3218
3219	slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, no_dirty_log);
3220	if (!slot)
3221		return KVM_PFN_ERR_FAULT;
3222
3223	return gfn_to_pfn_memslot_atomic(slot, gfn);
3224}
3225
3226static int direct_pte_prefetch_many(struct kvm_vcpu *vcpu,
3227				    struct kvm_mmu_page *sp,
3228				    u64 *start, u64 *end)
3229{
3230	struct page *pages[PTE_PREFETCH_NUM];
3231	struct kvm_memory_slot *slot;
3232	unsigned access = sp->role.access;
3233	int i, ret;
3234	gfn_t gfn;
3235
3236	gfn = kvm_mmu_page_get_gfn(sp, start - sp->spt);
3237	slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, access & ACC_WRITE_MASK);
3238	if (!slot)
3239		return -1;
3240
3241	ret = gfn_to_page_many_atomic(slot, gfn, pages, end - start);
3242	if (ret <= 0)
3243		return -1;
3244
3245	for (i = 0; i < ret; i++, gfn++, start++) {
3246		mmu_set_spte(vcpu, start, access, 0, sp->role.level, gfn,
3247			     page_to_pfn(pages[i]), true, true);
3248		put_page(pages[i]);
3249	}
3250
3251	return 0;
3252}
3253
3254static void __direct_pte_prefetch(struct kvm_vcpu *vcpu,
3255				  struct kvm_mmu_page *sp, u64 *sptep)
3256{
3257	u64 *spte, *start = NULL;
3258	int i;
3259
3260	WARN_ON(!sp->role.direct);
3261
3262	i = (sptep - sp->spt) & ~(PTE_PREFETCH_NUM - 1);
3263	spte = sp->spt + i;
3264
3265	for (i = 0; i < PTE_PREFETCH_NUM; i++, spte++) {
3266		if (is_shadow_present_pte(*spte) || spte == sptep) {
3267			if (!start)
3268				continue;
3269			if (direct_pte_prefetch_many(vcpu, sp, start, spte) < 0)
3270				break;
3271			start = NULL;
3272		} else if (!start)
3273			start = spte;
3274	}
3275}
3276
3277static void direct_pte_prefetch(struct kvm_vcpu *vcpu, u64 *sptep)
3278{
3279	struct kvm_mmu_page *sp;
3280
3281	sp = page_header(__pa(sptep));
3282
3283	/*
3284	 * Without accessed bits, there's no way to distinguish between
3285	 * actually accessed translations and prefetched, so disable pte
3286	 * prefetch if accessed bits aren't available.
3287	 */
3288	if (sp_ad_disabled(sp))
3289		return;
3290
3291	if (sp->role.level > PT_PAGE_TABLE_LEVEL)
3292		return;
3293
3294	__direct_pte_prefetch(vcpu, sp, sptep);
3295}
3296
3297static void disallowed_hugepage_adjust(struct kvm_shadow_walk_iterator it,
3298				       gfn_t gfn, kvm_pfn_t *pfnp, int *levelp)
3299{
3300	int level = *levelp;
3301	u64 spte = *it.sptep;
3302
3303	if (it.level == level && level > PT_PAGE_TABLE_LEVEL &&
3304	    is_nx_huge_page_enabled() &&
3305	    is_shadow_present_pte(spte) &&
3306	    !is_large_pte(spte)) {
3307		/*
3308		 * A small SPTE exists for this pfn, but FNAME(fetch)
3309		 * and __direct_map would like to create a large PTE
3310		 * instead: just force them to go down another level,
3311		 * patching back for them into pfn the next 9 bits of
3312		 * the address.
3313		 */
3314		u64 page_mask = KVM_PAGES_PER_HPAGE(level) - KVM_PAGES_PER_HPAGE(level - 1);
3315		*pfnp |= gfn & page_mask;
3316		(*levelp)--;
3317	}
3318}
3319
3320static int __direct_map(struct kvm_vcpu *vcpu, gpa_t gpa, int write,
3321			int map_writable, int level, kvm_pfn_t pfn,
3322			bool prefault, bool lpage_disallowed)
3323{
3324	struct kvm_shadow_walk_iterator it;
3325	struct kvm_mmu_page *sp;
3326	int ret;
3327	gfn_t gfn = gpa >> PAGE_SHIFT;
3328	gfn_t base_gfn = gfn;
3329
3330	if (!VALID_PAGE(vcpu->arch.mmu->root_hpa))
3331		return RET_PF_RETRY;
3332
3333	trace_kvm_mmu_spte_requested(gpa, level, pfn);
3334	for_each_shadow_entry(vcpu, gpa, it) {
3335		/*
3336		 * We cannot overwrite existing page tables with an NX
3337		 * large page, as the leaf could be executable.
3338		 */
3339		disallowed_hugepage_adjust(it, gfn, &pfn, &level);
3340
3341		base_gfn = gfn & ~(KVM_PAGES_PER_HPAGE(it.level) - 1);
3342		if (it.level == level)
3343			break;
3344
3345		drop_large_spte(vcpu, it.sptep);
3346		if (!is_shadow_present_pte(*it.sptep)) {
3347			sp = kvm_mmu_get_page(vcpu, base_gfn, it.addr,
3348					      it.level - 1, true, ACC_ALL);
3349
3350			link_shadow_page(vcpu, it.sptep, sp);
3351			if (lpage_disallowed)
3352				account_huge_nx_page(vcpu->kvm, sp);
3353		}
3354	}
3355
3356	ret = mmu_set_spte(vcpu, it.sptep, ACC_ALL,
3357			   write, level, base_gfn, pfn, prefault,
3358			   map_writable);
3359	direct_pte_prefetch(vcpu, it.sptep);
3360	++vcpu->stat.pf_fixed;
3361	return ret;
3362}
3363
3364static void kvm_send_hwpoison_signal(unsigned long address, struct task_struct *tsk)
3365{
3366	send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, PAGE_SHIFT, tsk);
3367}
3368
3369static int kvm_handle_bad_page(struct kvm_vcpu *vcpu, gfn_t gfn, kvm_pfn_t pfn)
3370{
3371	/*
3372	 * Do not cache the mmio info caused by writing the readonly gfn
3373	 * into the spte otherwise read access on readonly gfn also can
3374	 * caused mmio page fault and treat it as mmio access.
3375	 */
3376	if (pfn == KVM_PFN_ERR_RO_FAULT)
3377		return RET_PF_EMULATE;
3378
3379	if (pfn == KVM_PFN_ERR_HWPOISON) {
3380		kvm_send_hwpoison_signal(kvm_vcpu_gfn_to_hva(vcpu, gfn), current);
3381		return RET_PF_RETRY;
3382	}
3383
3384	return -EFAULT;
3385}
3386
3387static void transparent_hugepage_adjust(struct kvm_vcpu *vcpu,
3388					gfn_t gfn, kvm_pfn_t *pfnp,
3389					int *levelp)
3390{
3391	kvm_pfn_t pfn = *pfnp;
3392	int level = *levelp;
3393
3394	/*
3395	 * Check if it's a transparent hugepage. If this would be an
3396	 * hugetlbfs page, level wouldn't be set to
3397	 * PT_PAGE_TABLE_LEVEL and there would be no adjustment done
3398	 * here.
3399	 */
3400	if (!is_error_noslot_pfn(pfn) && !kvm_is_reserved_pfn(pfn) &&
3401	    !kvm_is_zone_device_pfn(pfn) && level == PT_PAGE_TABLE_LEVEL &&
3402	    PageTransCompoundMap(pfn_to_page(pfn)) &&
3403	    !mmu_gfn_lpage_is_disallowed(vcpu, gfn, PT_DIRECTORY_LEVEL)) {
3404		unsigned long mask;
3405		/*
3406		 * mmu_notifier_retry was successful and we hold the
3407		 * mmu_lock here, so the pmd can't become splitting
3408		 * from under us, and in turn
3409		 * __split_huge_page_refcount() can't run from under
3410		 * us and we can safely transfer the refcount from
3411		 * PG_tail to PG_head as we switch the pfn to tail to
3412		 * head.
3413		 */
3414		*levelp = level = PT_DIRECTORY_LEVEL;
3415		mask = KVM_PAGES_PER_HPAGE(level) - 1;
3416		VM_BUG_ON((gfn & mask) != (pfn & mask));
3417		if (pfn & mask) {
3418			kvm_release_pfn_clean(pfn);
3419			pfn &= ~mask;
3420			kvm_get_pfn(pfn);
3421			*pfnp = pfn;
3422		}
3423	}
3424}
3425
3426static bool handle_abnormal_pfn(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn,
3427				kvm_pfn_t pfn, unsigned access, int *ret_val)
3428{
3429	/* The pfn is invalid, report the error! */
3430	if (unlikely(is_error_pfn(pfn))) {
3431		*ret_val = kvm_handle_bad_page(vcpu, gfn, pfn);
3432		return true;
3433	}
3434
3435	if (unlikely(is_noslot_pfn(pfn)))
3436		vcpu_cache_mmio_info(vcpu, gva, gfn,
3437				     access & shadow_mmio_access_mask);
3438
3439	return false;
3440}
3441
3442static bool page_fault_can_be_fast(u32 error_code)
3443{
3444	/*
3445	 * Do not fix the mmio spte with invalid generation number which
3446	 * need to be updated by slow page fault path.
3447	 */
3448	if (unlikely(error_code & PFERR_RSVD_MASK))
3449		return false;
3450
3451	/* See if the page fault is due to an NX violation */
3452	if (unlikely(((error_code & (PFERR_FETCH_MASK | PFERR_PRESENT_MASK))
3453		      == (PFERR_FETCH_MASK | PFERR_PRESENT_MASK))))
3454		return false;
3455
3456	/*
3457	 * #PF can be fast if:
3458	 * 1. The shadow page table entry is not present, which could mean that
3459	 *    the fault is potentially caused by access tracking (if enabled).
3460	 * 2. The shadow page table entry is present and the fault
3461	 *    is caused by write-protect, that means we just need change the W
3462	 *    bit of the spte which can be done out of mmu-lock.
3463	 *
3464	 * However, if access tracking is disabled we know that a non-present
3465	 * page must be a genuine page fault where we have to create a new SPTE.
3466	 * So, if access tracking is disabled, we return true only for write
3467	 * accesses to a present page.
3468	 */
3469
3470	return shadow_acc_track_mask != 0 ||
3471	       ((error_code & (PFERR_WRITE_MASK | PFERR_PRESENT_MASK))
3472		== (PFERR_WRITE_MASK | PFERR_PRESENT_MASK));
3473}
3474
3475/*
3476 * Returns true if the SPTE was fixed successfully. Otherwise,
3477 * someone else modified the SPTE from its original value.
3478 */
3479static bool
3480fast_pf_fix_direct_spte(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
3481			u64 *sptep, u64 old_spte, u64 new_spte)
3482{
3483	gfn_t gfn;
3484
3485	WARN_ON(!sp->role.direct);
3486
3487	/*
3488	 * Theoretically we could also set dirty bit (and flush TLB) here in
3489	 * order to eliminate unnecessary PML logging. See comments in
3490	 * set_spte. But fast_page_fault is very unlikely to happen with PML
3491	 * enabled, so we do not do this. This might result in the same GPA
3492	 * to be logged in PML buffer again when the write really happens, and
3493	 * eventually to be called by mark_page_dirty twice. But it's also no
3494	 * harm. This also avoids the TLB flush needed after setting dirty bit
3495	 * so non-PML cases won't be impacted.
3496	 *
3497	 * Compare with set_spte where instead shadow_dirty_mask is set.
3498	 */
3499	if (cmpxchg64(sptep, old_spte, new_spte) != old_spte)
3500		return false;
3501
3502	if (is_writable_pte(new_spte) && !is_writable_pte(old_spte)) {
3503		/*
3504		 * The gfn of direct spte is stable since it is
3505		 * calculated by sp->gfn.
3506		 */
3507		gfn = kvm_mmu_page_get_gfn(sp, sptep - sp->spt);
3508		kvm_vcpu_mark_page_dirty(vcpu, gfn);
3509	}
3510
3511	return true;
3512}
3513
3514static bool is_access_allowed(u32 fault_err_code, u64 spte)
3515{
3516	if (fault_err_code & PFERR_FETCH_MASK)
3517		return is_executable_pte(spte);
3518
3519	if (fault_err_code & PFERR_WRITE_MASK)
3520		return is_writable_pte(spte);
3521
3522	/* Fault was on Read access */
3523	return spte & PT_PRESENT_MASK;
3524}
3525
3526/*
3527 * Return value:
3528 * - true: let the vcpu to access on the same address again.
3529 * - false: let the real page fault path to fix it.
3530 */
3531static bool fast_page_fault(struct kvm_vcpu *vcpu, gva_t gva, int level,
3532			    u32 error_code)
3533{
3534	struct kvm_shadow_walk_iterator iterator;
3535	struct kvm_mmu_page *sp;
3536	bool fault_handled = false;
3537	u64 spte = 0ull;
3538	uint retry_count = 0;
3539
3540	if (!VALID_PAGE(vcpu->arch.mmu->root_hpa))
3541		return false;
3542
3543	if (!page_fault_can_be_fast(error_code))
3544		return false;
3545
3546	walk_shadow_page_lockless_begin(vcpu);
3547
3548	do {
3549		u64 new_spte;
3550
3551		for_each_shadow_entry_lockless(vcpu, gva, iterator, spte)
3552			if (!is_shadow_present_pte(spte) ||
3553			    iterator.level < level)
3554				break;
3555
3556		sp = page_header(__pa(iterator.sptep));
3557		if (!is_last_spte(spte, sp->role.level))
3558			break;
3559
3560		/*
3561		 * Check whether the memory access that caused the fault would
3562		 * still cause it if it were to be performed right now. If not,
3563		 * then this is a spurious fault caused by TLB lazily flushed,
3564		 * or some other CPU has already fixed the PTE after the
3565		 * current CPU took the fault.
3566		 *
3567		 * Need not check the access of upper level table entries since
3568		 * they are always ACC_ALL.
3569		 */
3570		if (is_access_allowed(error_code, spte)) {
3571			fault_handled = true;
3572			break;
3573		}
3574
3575		new_spte = spte;
3576
3577		if (is_access_track_spte(spte))
3578			new_spte = restore_acc_track_spte(new_spte);
3579
3580		/*
3581		 * Currently, to simplify the code, write-protection can
3582		 * be removed in the fast path only if the SPTE was
3583		 * write-protected for dirty-logging or access tracking.
3584		 */
3585		if ((error_code & PFERR_WRITE_MASK) &&
3586		    spte_can_locklessly_be_made_writable(spte))
3587		{
3588			new_spte |= PT_WRITABLE_MASK;
3589
3590			/*
3591			 * Do not fix write-permission on the large spte.  Since
3592			 * we only dirty the first page into the dirty-bitmap in
3593			 * fast_pf_fix_direct_spte(), other pages are missed
3594			 * if its slot has dirty logging enabled.
3595			 *
3596			 * Instead, we let the slow page fault path create a
3597			 * normal spte to fix the access.
3598			 *
3599			 * See the comments in kvm_arch_commit_memory_region().
3600			 */
3601			if (sp->role.level > PT_PAGE_TABLE_LEVEL)
3602				break;
3603		}
3604
3605		/* Verify that the fault can be handled in the fast path */
3606		if (new_spte == spte ||
3607		    !is_access_allowed(error_code, new_spte))
3608			break;
3609
3610		/*
3611		 * Currently, fast page fault only works for direct mapping
3612		 * since the gfn is not stable for indirect shadow page. See
3613		 * Documentation/virt/kvm/locking.txt to get more detail.
3614		 */
3615		fault_handled = fast_pf_fix_direct_spte(vcpu, sp,
3616							iterator.sptep, spte,
3617							new_spte);
3618		if (fault_handled)
3619			break;
3620
3621		if (++retry_count > 4) {
3622			printk_once(KERN_WARNING
3623				"kvm: Fast #PF retrying more than 4 times.\n");
3624			break;
3625		}
3626
3627	} while (true);
3628
3629	trace_fast_page_fault(vcpu, gva, error_code, iterator.sptep,
3630			      spte, fault_handled);
3631	walk_shadow_page_lockless_end(vcpu);
3632
3633	return fault_handled;
3634}
3635
3636static bool try_async_pf(struct kvm_vcpu *vcpu, bool prefault, gfn_t gfn,
3637			 gva_t gva, kvm_pfn_t *pfn, bool write, bool *writable);
3638static int make_mmu_pages_available(struct kvm_vcpu *vcpu);
3639
3640static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, u32 error_code,
3641			 gfn_t gfn, bool prefault)
3642{
3643	int r;
3644	int level;
3645	bool force_pt_level;
3646	kvm_pfn_t pfn;
3647	unsigned long mmu_seq;
3648	bool map_writable, write = error_code & PFERR_WRITE_MASK;
3649	bool lpage_disallowed = (error_code & PFERR_FETCH_MASK) &&
3650				is_nx_huge_page_enabled();
3651
3652	force_pt_level = lpage_disallowed;
3653	level = mapping_level(vcpu, gfn, &force_pt_level);
3654	if (likely(!force_pt_level)) {
3655		/*
3656		 * This path builds a PAE pagetable - so we can map
3657		 * 2mb pages at maximum. Therefore check if the level
3658		 * is larger than that.
3659		 */
3660		if (level > PT_DIRECTORY_LEVEL)
3661			level = PT_DIRECTORY_LEVEL;
3662
3663		gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
3664	}
3665
3666	if (fast_page_fault(vcpu, v, level, error_code))
3667		return RET_PF_RETRY;
3668
3669	mmu_seq = vcpu->kvm->mmu_notifier_seq;
3670	smp_rmb();
3671
3672	if (try_async_pf(vcpu, prefault, gfn, v, &pfn, write, &map_writable))
3673		return RET_PF_RETRY;
3674
3675	if (handle_abnormal_pfn(vcpu, v, gfn, pfn, ACC_ALL, &r))
3676		return r;
3677
3678	r = RET_PF_RETRY;
3679	spin_lock(&vcpu->kvm->mmu_lock);
3680	if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
3681		goto out_unlock;
3682	if (make_mmu_pages_available(vcpu) < 0)
3683		goto out_unlock;
3684	if (likely(!force_pt_level))
3685		transparent_hugepage_adjust(vcpu, gfn, &pfn, &level);
3686	r = __direct_map(vcpu, v, write, map_writable, level, pfn,
3687			 prefault, false);
3688out_unlock:
3689	spin_unlock(&vcpu->kvm->mmu_lock);
3690	kvm_release_pfn_clean(pfn);
3691	return r;
3692}
3693
3694static void mmu_free_root_page(struct kvm *kvm, hpa_t *root_hpa,
3695			       struct list_head *invalid_list)
3696{
3697	struct kvm_mmu_page *sp;
3698
3699	if (!VALID_PAGE(*root_hpa))
3700		return;
3701
3702	sp = page_header(*root_hpa & PT64_BASE_ADDR_MASK);
3703	--sp->root_count;
3704	if (!sp->root_count && sp->role.invalid)
3705		kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
3706
3707	*root_hpa = INVALID_PAGE;
3708}
3709
3710/* roots_to_free must be some combination of the KVM_MMU_ROOT_* flags */
3711void kvm_mmu_free_roots(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
3712			ulong roots_to_free)
3713{
3714	int i;
3715	LIST_HEAD(invalid_list);
3716	bool free_active_root = roots_to_free & KVM_MMU_ROOT_CURRENT;
3717
3718	BUILD_BUG_ON(KVM_MMU_NUM_PREV_ROOTS >= BITS_PER_LONG);
3719
3720	/* Before acquiring the MMU lock, see if we need to do any real work. */
3721	if (!(free_active_root && VALID_PAGE(mmu->root_hpa))) {
3722		for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
3723			if ((roots_to_free & KVM_MMU_ROOT_PREVIOUS(i)) &&
3724			    VALID_PAGE(mmu->prev_roots[i].hpa))
3725				break;
3726
3727		if (i == KVM_MMU_NUM_PREV_ROOTS)
3728			return;
3729	}
3730
3731	spin_lock(&vcpu->kvm->mmu_lock);
3732
3733	for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
3734		if (roots_to_free & KVM_MMU_ROOT_PREVIOUS(i))
3735			mmu_free_root_page(vcpu->kvm, &mmu->prev_roots[i].hpa,
3736					   &invalid_list);
3737
3738	if (free_active_root) {
3739		if (mmu->shadow_root_level >= PT64_ROOT_4LEVEL &&
3740		    (mmu->root_level >= PT64_ROOT_4LEVEL || mmu->direct_map)) {
3741			mmu_free_root_page(vcpu->kvm, &mmu->root_hpa,
3742					   &invalid_list);
3743		} else {
3744			for (i = 0; i < 4; ++i)
3745				if (mmu->pae_root[i] != 0)
3746					mmu_free_root_page(vcpu->kvm,
3747							   &mmu->pae_root[i],
3748							   &invalid_list);
3749			mmu->root_hpa = INVALID_PAGE;
3750		}
3751		mmu->root_cr3 = 0;
3752	}
3753
3754	kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
3755	spin_unlock(&vcpu->kvm->mmu_lock);
3756}
3757EXPORT_SYMBOL_GPL(kvm_mmu_free_roots);
3758
3759static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
3760{
3761	int ret = 0;
3762
3763	if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
3764		kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
3765		ret = 1;
3766	}
3767
3768	return ret;
3769}
3770
3771static int mmu_alloc_direct_roots(struct kvm_vcpu *vcpu)
3772{
3773	struct kvm_mmu_page *sp;
3774	unsigned i;
3775
3776	if (vcpu->arch.mmu->shadow_root_level >= PT64_ROOT_4LEVEL) {
3777		spin_lock(&vcpu->kvm->mmu_lock);
3778		if(make_mmu_pages_available(vcpu) < 0) {
3779			spin_unlock(&vcpu->kvm->mmu_lock);
3780			return -ENOSPC;
3781		}
3782		sp = kvm_mmu_get_page(vcpu, 0, 0,
3783				vcpu->arch.mmu->shadow_root_level, 1, ACC_ALL);
3784		++sp->root_count;
3785		spin_unlock(&vcpu->kvm->mmu_lock);
3786		vcpu->arch.mmu->root_hpa = __pa(sp->spt);
3787	} else if (vcpu->arch.mmu->shadow_root_level == PT32E_ROOT_LEVEL) {
3788		for (i = 0; i < 4; ++i) {
3789			hpa_t root = vcpu->arch.mmu->pae_root[i];
3790
3791			MMU_WARN_ON(VALID_PAGE(root));
3792			spin_lock(&vcpu->kvm->mmu_lock);
3793			if (make_mmu_pages_available(vcpu) < 0) {
3794				spin_unlock(&vcpu->kvm->mmu_lock);
3795				return -ENOSPC;
3796			}
3797			sp = kvm_mmu_get_page(vcpu, i << (30 - PAGE_SHIFT),
3798					i << 30, PT32_ROOT_LEVEL, 1, ACC_ALL);
3799			root = __pa(sp->spt);
3800			++sp->root_count;
3801			spin_unlock(&vcpu->kvm->mmu_lock);
3802			vcpu->arch.mmu->pae_root[i] = root | PT_PRESENT_MASK;
3803		}
3804		vcpu->arch.mmu->root_hpa = __pa(vcpu->arch.mmu->pae_root);
3805	} else
3806		BUG();
3807	vcpu->arch.mmu->root_cr3 = vcpu->arch.mmu->get_cr3(vcpu);
3808
3809	return 0;
3810}
3811
3812static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu)
3813{
3814	struct kvm_mmu_page *sp;
3815	u64 pdptr, pm_mask;
3816	gfn_t root_gfn, root_cr3;
3817	int i;
3818
3819	root_cr3 = vcpu->arch.mmu->get_cr3(vcpu);
3820	root_gfn = root_cr3 >> PAGE_SHIFT;
3821
3822	if (mmu_check_root(vcpu, root_gfn))
3823		return 1;
3824
3825	/*
3826	 * Do we shadow a long mode page table? If so we need to
3827	 * write-protect the guests page table root.
3828	 */
3829	if (vcpu->arch.mmu->root_level >= PT64_ROOT_4LEVEL) {
3830		hpa_t root = vcpu->arch.mmu->root_hpa;
3831
3832		MMU_WARN_ON(VALID_PAGE(root));
3833
3834		spin_lock(&vcpu->kvm->mmu_lock);
3835		if (make_mmu_pages_available(vcpu) < 0) {
3836			spin_unlock(&vcpu->kvm->mmu_lock);
3837			return -ENOSPC;
3838		}
3839		sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
3840				vcpu->arch.mmu->shadow_root_level, 0, ACC_ALL);
3841		root = __pa(sp->spt);
3842		++sp->root_count;
3843		spin_unlock(&vcpu->kvm->mmu_lock);
3844		vcpu->arch.mmu->root_hpa = root;
3845		goto set_root_cr3;
3846	}
3847
3848	/*
3849	 * We shadow a 32 bit page table. This may be a legacy 2-level
3850	 * or a PAE 3-level page table. In either case we need to be aware that
3851	 * the shadow page table may be a PAE or a long mode page table.
3852	 */
3853	pm_mask = PT_PRESENT_MASK;
3854	if (vcpu->arch.mmu->shadow_root_level == PT64_ROOT_4LEVEL)
3855		pm_mask |= PT_ACCESSED_MASK | PT_WRITABLE_MASK | PT_USER_MASK;
3856
3857	for (i = 0; i < 4; ++i) {
3858		hpa_t root = vcpu->arch.mmu->pae_root[i];
3859
3860		MMU_WARN_ON(VALID_PAGE(root));
3861		if (vcpu->arch.mmu->root_level == PT32E_ROOT_LEVEL) {
3862			pdptr = vcpu->arch.mmu->get_pdptr(vcpu, i);
3863			if (!(pdptr & PT_PRESENT_MASK)) {
3864				vcpu->arch.mmu->pae_root[i] = 0;
3865				continue;
3866			}
3867			root_gfn = pdptr >> PAGE_SHIFT;
3868			if (mmu_check_root(vcpu, root_gfn))
3869				return 1;
3870		}
3871		spin_lock(&vcpu->kvm->mmu_lock);
3872		if (make_mmu_pages_available(vcpu) < 0) {
3873			spin_unlock(&vcpu->kvm->mmu_lock);
3874			return -ENOSPC;
3875		}
3876		sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30, PT32_ROOT_LEVEL,
3877				      0, ACC_ALL);
3878		root = __pa(sp->spt);
3879		++sp->root_count;
3880		spin_unlock(&vcpu->kvm->mmu_lock);
3881
3882		vcpu->arch.mmu->pae_root[i] = root | pm_mask;
3883	}
3884	vcpu->arch.mmu->root_hpa = __pa(vcpu->arch.mmu->pae_root);
3885
3886	/*
3887	 * If we shadow a 32 bit page table with a long mode page
3888	 * table we enter this path.
3889	 */
3890	if (vcpu->arch.mmu->shadow_root_level == PT64_ROOT_4LEVEL) {
3891		if (vcpu->arch.mmu->lm_root == NULL) {
3892			/*
3893			 * The additional page necessary for this is only
3894			 * allocated on demand.
3895			 */
3896
3897			u64 *lm_root;
3898
3899			lm_root = (void*)get_zeroed_page(GFP_KERNEL_ACCOUNT);
3900			if (lm_root == NULL)
3901				return 1;
3902
3903			lm_root[0] = __pa(vcpu->arch.mmu->pae_root) | pm_mask;
3904
3905			vcpu->arch.mmu->lm_root = lm_root;
3906		}
3907
3908		vcpu->arch.mmu->root_hpa = __pa(vcpu->arch.mmu->lm_root);
3909	}
3910
3911set_root_cr3:
3912	vcpu->arch.mmu->root_cr3 = root_cr3;
3913
3914	return 0;
3915}
3916
3917static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
3918{
3919	if (vcpu->arch.mmu->direct_map)
3920		return mmu_alloc_direct_roots(vcpu);
3921	else
3922		return mmu_alloc_shadow_roots(vcpu);
3923}
3924
3925void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
3926{
3927	int i;
3928	struct kvm_mmu_page *sp;
3929
3930	if (vcpu->arch.mmu->direct_map)
3931		return;
3932
3933	if (!VALID_PAGE(vcpu->arch.mmu->root_hpa))
3934		return;
3935
3936	vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY);
3937
3938	if (vcpu->arch.mmu->root_level >= PT64_ROOT_4LEVEL) {
3939		hpa_t root = vcpu->arch.mmu->root_hpa;
3940		sp = page_header(root);
3941
3942		/*
3943		 * Even if another CPU was marking the SP as unsync-ed
3944		 * simultaneously, any guest page table changes are not
3945		 * guaranteed to be visible anyway until this VCPU issues a TLB
3946		 * flush strictly after those changes are made. We only need to
3947		 * ensure that the other CPU sets these flags before any actual
3948		 * changes to the page tables are made. The comments in
3949		 * mmu_need_write_protect() describe what could go wrong if this
3950		 * requirement isn't satisfied.
3951		 */
3952		if (!smp_load_acquire(&sp->unsync) &&
3953		    !smp_load_acquire(&sp->unsync_children))
3954			return;
3955
3956		spin_lock(&vcpu->kvm->mmu_lock);
3957		kvm_mmu_audit(vcpu, AUDIT_PRE_SYNC);
3958
3959		mmu_sync_children(vcpu, sp);
3960
3961		kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
3962		spin_unlock(&vcpu->kvm->mmu_lock);
3963		return;
3964	}
3965
3966	spin_lock(&vcpu->kvm->mmu_lock);
3967	kvm_mmu_audit(vcpu, AUDIT_PRE_SYNC);
3968
3969	for (i = 0; i < 4; ++i) {
3970		hpa_t root = vcpu->arch.mmu->pae_root[i];
3971
3972		if (root && VALID_PAGE(root)) {
3973			root &= PT64_BASE_ADDR_MASK;
3974			sp = page_header(root);
3975			mmu_sync_children(vcpu, sp);
3976		}
3977	}
3978
3979	kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
3980	spin_unlock(&vcpu->kvm->mmu_lock);
3981}
3982EXPORT_SYMBOL_GPL(kvm_mmu_sync_roots);
3983
3984static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
3985				  u32 access, struct x86_exception *exception)
3986{
3987	if (exception)
3988		exception->error_code = 0;
3989	return vaddr;
3990}
3991
3992static gpa_t nonpaging_gva_to_gpa_nested(struct kvm_vcpu *vcpu, gva_t vaddr,
3993					 u32 access,
3994					 struct x86_exception *exception)
3995{
3996	if (exception)
3997		exception->error_code = 0;
3998	return vcpu->arch.nested_mmu.translate_gpa(vcpu, vaddr, access, exception);
3999}
4000
4001static bool
4002__is_rsvd_bits_set(struct rsvd_bits_validate *rsvd_check, u64 pte, int level)
4003{
4004	int bit7 = (pte >> 7) & 1, low6 = pte & 0x3f;
4005
4006	return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) |
4007		((rsvd_check->bad_mt_xwr & (1ull << low6)) != 0);
4008}
4009
4010static bool is_rsvd_bits_set(struct kvm_mmu *mmu, u64 gpte, int level)
4011{
4012	return __is_rsvd_bits_set(&mmu->guest_rsvd_check, gpte, level);
4013}
4014
4015static bool is_shadow_zero_bits_set(struct kvm_mmu *mmu, u64 spte, int level)
4016{
4017	return __is_rsvd_bits_set(&mmu->shadow_zero_check, spte, level);
4018}
4019
4020static bool mmio_info_in_cache(struct kvm_vcpu *vcpu, u64 addr, bool direct)
4021{
4022	/*
4023	 * A nested guest cannot use the MMIO cache if it is using nested
4024	 * page tables, because cr2 is a nGPA while the cache stores GPAs.
4025	 */
4026	if (mmu_is_nested(vcpu))
4027		return false;
4028
4029	if (direct)
4030		return vcpu_match_mmio_gpa(vcpu, addr);
4031
4032	return vcpu_match_mmio_gva(vcpu, addr);
4033}
4034
4035/* return true if reserved bit is detected on spte. */
4036static bool
4037walk_shadow_page_get_mmio_spte(struct kvm_vcpu *vcpu, u64 addr, u64 *sptep)
4038{
4039	struct kvm_shadow_walk_iterator iterator;
4040	u64 sptes[PT64_ROOT_MAX_LEVEL], spte = 0ull;
4041	int root, leaf;
4042	bool reserved = false;
4043
4044	if (!VALID_PAGE(vcpu->arch.mmu->root_hpa))
4045		goto exit;
4046
4047	walk_shadow_page_lockless_begin(vcpu);
4048
4049	for (shadow_walk_init(&iterator, vcpu, addr),
4050		 leaf = root = iterator.level;
4051	     shadow_walk_okay(&iterator);
4052	     __shadow_walk_next(&iterator, spte)) {
4053		spte = mmu_spte_get_lockless(iterator.sptep);
4054
4055		sptes[leaf - 1] = spte;
4056		leaf--;
4057
4058		if (!is_shadow_present_pte(spte))
4059			break;
4060
4061		reserved |= is_shadow_zero_bits_set(vcpu->arch.mmu, spte,
4062						    iterator.level);
4063	}
4064
4065	walk_shadow_page_lockless_end(vcpu);
4066
4067	if (reserved) {
4068		pr_err("%s: detect reserved bits on spte, addr 0x%llx, dump hierarchy:\n",
4069		       __func__, addr);
4070		while (root > leaf) {
4071			pr_err("------ spte 0x%llx level %d.\n",
4072			       sptes[root - 1], root);
4073			root--;
4074		}
4075	}
4076exit:
4077	*sptep = spte;
4078	return reserved;
4079}
4080
4081static int handle_mmio_page_fault(struct kvm_vcpu *vcpu, u64 addr, bool direct)
4082{
4083	u64 spte;
4084	bool reserved;
4085
4086	if (mmio_info_in_cache(vcpu, addr, direct))
4087		return RET_PF_EMULATE;
4088
4089	reserved = walk_shadow_page_get_mmio_spte(vcpu, addr, &spte);
4090	if (WARN_ON(reserved))
4091		return -EINVAL;
4092
4093	if (is_mmio_spte(spte)) {
4094		gfn_t gfn = get_mmio_spte_gfn(spte);
4095		unsigned access = get_mmio_spte_access(spte);
4096
4097		if (!check_mmio_spte(vcpu, spte))
4098			return RET_PF_INVALID;
4099
4100		if (direct)
4101			addr = 0;
4102
4103		trace_handle_mmio_page_fault(addr, gfn, access);
4104		vcpu_cache_mmio_info(vcpu, addr, gfn, access);
4105		return RET_PF_EMULATE;
4106	}
4107
4108	/*
4109	 * If the page table is zapped by other cpus, let CPU fault again on
4110	 * the address.
4111	 */
4112	return RET_PF_RETRY;
4113}
4114
4115static bool page_fault_handle_page_track(struct kvm_vcpu *vcpu,
4116					 u32 error_code, gfn_t gfn)
4117{
4118	if (unlikely(error_code & PFERR_RSVD_MASK))
4119		return false;
4120
4121	if (!(error_code & PFERR_PRESENT_MASK) ||
4122	      !(error_code & PFERR_WRITE_MASK))
4123		return false;
4124
4125	/*
4126	 * guest is writing the page which is write tracked which can
4127	 * not be fixed by page fault handler.
4128	 */
4129	if (kvm_page_track_is_active(vcpu, gfn, KVM_PAGE_TRACK_WRITE))
4130		return true;
4131
4132	return false;
4133}
4134
4135static void shadow_page_table_clear_flood(struct kvm_vcpu *vcpu, gva_t addr)
4136{
4137	struct kvm_shadow_walk_iterator iterator;
4138	u64 spte;
4139
4140	if (!VALID_PAGE(vcpu->arch.mmu->root_hpa))
4141		return;
4142
4143	walk_shadow_page_lockless_begin(vcpu);
4144	for_each_shadow_entry_lockless(vcpu, addr, iterator, spte) {
4145		clear_sp_write_flooding_count(iterator.sptep);
4146		if (!is_shadow_present_pte(spte))
4147			break;
4148	}
4149	walk_shadow_page_lockless_end(vcpu);
4150}
4151
4152static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
4153				u32 error_code, bool prefault)
4154{
4155	gfn_t gfn = gva >> PAGE_SHIFT;
4156	int r;
4157
4158	pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
4159
4160	if (page_fault_handle_page_track(vcpu, error_code, gfn))
4161		return RET_PF_EMULATE;
4162
4163	r = mmu_topup_memory_caches(vcpu);
4164	if (r)
4165		return r;
4166
4167	MMU_WARN_ON(!VALID_PAGE(vcpu->arch.mmu->root_hpa));
4168
4169
4170	return nonpaging_map(vcpu, gva & PAGE_MASK,
4171			     error_code, gfn, prefault);
4172}
4173
4174static int kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn)
4175{
4176	struct kvm_arch_async_pf arch;
4177
4178	arch.token = (vcpu->arch.apf.id++ << 12) | vcpu->vcpu_id;
4179	arch.gfn = gfn;
4180	arch.direct_map = vcpu->arch.mmu->direct_map;
4181	arch.cr3 = vcpu->arch.mmu->get_cr3(vcpu);
4182
4183	return kvm_setup_async_pf(vcpu, gva, kvm_vcpu_gfn_to_hva(vcpu, gfn), &arch);
4184}
4185
4186static bool try_async_pf(struct kvm_vcpu *vcpu, bool prefault, gfn_t gfn,
4187			 gva_t gva, kvm_pfn_t *pfn, bool write, bool *writable)
4188{
4189	struct kvm_memory_slot *slot;
4190	bool async;
4191
4192	/*
4193	 * Don't expose private memslots to L2.
4194	 */
4195	if (is_guest_mode(vcpu) && !kvm_is_visible_gfn(vcpu->kvm, gfn)) {
4196		*pfn = KVM_PFN_NOSLOT;
4197		return false;
4198	}
4199
4200	slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
4201	async = false;
4202	*pfn = __gfn_to_pfn_memslot(slot, gfn, false, &async, write, writable);
4203	if (!async)
4204		return false; /* *pfn has correct page already */
4205
4206	if (!prefault && kvm_can_do_async_pf(vcpu)) {
4207		trace_kvm_try_async_get_page(gva, gfn);
4208		if (kvm_find_async_pf_gfn(vcpu, gfn)) {
4209			trace_kvm_async_pf_doublefault(gva, gfn);
4210			kvm_make_request(KVM_REQ_APF_HALT, vcpu);
4211			return true;
4212		} else if (kvm_arch_setup_async_pf(vcpu, gva, gfn))
4213			return true;
4214	}
4215
4216	*pfn = __gfn_to_pfn_memslot(slot, gfn, false, NULL, write, writable);
4217	return false;
4218}
4219
4220int kvm_handle_page_fault(struct kvm_vcpu *vcpu, u64 error_code,
4221				u64 fault_address, char *insn, int insn_len)
4222{
4223	int r = 1;
4224
4225	vcpu->arch.l1tf_flush_l1d = true;
4226	switch (vcpu->arch.apf.host_apf_reason) {
4227	default:
4228		trace_kvm_page_fault(fault_address, error_code);
4229
4230		if (kvm_event_needs_reinjection(vcpu))
4231			kvm_mmu_unprotect_page_virt(vcpu, fault_address);
4232		r = kvm_mmu_page_fault(vcpu, fault_address, error_code, insn,
4233				insn_len);
4234		break;
4235	case KVM_PV_REASON_PAGE_NOT_PRESENT:
4236		vcpu->arch.apf.host_apf_reason = 0;
4237		local_irq_disable();
4238		kvm_async_pf_task_wait(fault_address, 0);
4239		local_irq_enable();
4240		break;
4241	case KVM_PV_REASON_PAGE_READY:
4242		vcpu->arch.apf.host_apf_reason = 0;
4243		local_irq_disable();
4244		kvm_async_pf_task_wake(fault_address);
4245		local_irq_enable();
4246		break;
4247	}
4248	return r;
4249}
4250EXPORT_SYMBOL_GPL(kvm_handle_page_fault);
4251
4252static bool
4253check_hugepage_cache_consistency(struct kvm_vcpu *vcpu, gfn_t gfn, int level)
4254{
4255	int page_num = KVM_PAGES_PER_HPAGE(level);
4256
4257	gfn &= ~(page_num - 1);
4258
4259	return kvm_mtrr_check_gfn_range_consistency(vcpu, gfn, page_num);
4260}
4261
4262static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa, u32 error_code,
4263			  bool prefault)
4264{
4265	kvm_pfn_t pfn;
4266	int r;
4267	int level;
4268	bool force_pt_level;
4269	gfn_t gfn = gpa >> PAGE_SHIFT;
4270	unsigned long mmu_seq;
4271	int write = error_code & PFERR_WRITE_MASK;
4272	bool map_writable;
4273	bool lpage_disallowed = (error_code & PFERR_FETCH_MASK) &&
4274				is_nx_huge_page_enabled();
4275
4276	MMU_WARN_ON(!VALID_PAGE(vcpu->arch.mmu->root_hpa));
4277
4278	if (page_fault_handle_page_track(vcpu, error_code, gfn))
4279		return RET_PF_EMULATE;
4280
4281	r = mmu_topup_memory_caches(vcpu);
4282	if (r)
4283		return r;
4284
4285	force_pt_level =
4286		lpage_disallowed ||
4287		!check_hugepage_cache_consistency(vcpu, gfn, PT_DIRECTORY_LEVEL);
4288	level = mapping_level(vcpu, gfn, &force_pt_level);
4289	if (likely(!force_pt_level)) {
4290		if (level > PT_DIRECTORY_LEVEL &&
4291		    !check_hugepage_cache_consistency(vcpu, gfn, level))
4292			level = PT_DIRECTORY_LEVEL;
4293		gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
4294	}
4295
4296	if (fast_page_fault(vcpu, gpa, level, error_code))
4297		return RET_PF_RETRY;
4298
4299	mmu_seq = vcpu->kvm->mmu_notifier_seq;
4300	smp_rmb();
4301
4302	if (try_async_pf(vcpu, prefault, gfn, gpa, &pfn, write, &map_writable))
4303		return RET_PF_RETRY;
4304
4305	if (handle_abnormal_pfn(vcpu, 0, gfn, pfn, ACC_ALL, &r))
4306		return r;
4307
4308	r = RET_PF_RETRY;
4309	spin_lock(&vcpu->kvm->mmu_lock);
4310	if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
4311		goto out_unlock;
4312	if (make_mmu_pages_available(vcpu) < 0)
4313		goto out_unlock;
4314	if (likely(!force_pt_level))
4315		transparent_hugepage_adjust(vcpu, gfn, &pfn, &level);
4316	r = __direct_map(vcpu, gpa, write, map_writable, level, pfn,
4317			 prefault, lpage_disallowed);
4318out_unlock:
4319	spin_unlock(&vcpu->kvm->mmu_lock);
4320	kvm_release_pfn_clean(pfn);
4321	return r;
4322}
4323
4324static void nonpaging_init_context(struct kvm_vcpu *vcpu,
4325				   struct kvm_mmu *context)
4326{
4327	context->page_fault = nonpaging_page_fault;
4328	context->gva_to_gpa = nonpaging_gva_to_gpa;
4329	context->sync_page = nonpaging_sync_page;
4330	context->invlpg = nonpaging_invlpg;
4331	context->update_pte = nonpaging_update_pte;
4332	context->root_level = 0;
4333	context->shadow_root_level = PT32E_ROOT_LEVEL;
4334	context->direct_map = true;
4335	context->nx = false;
4336}
4337
4338/*
4339 * Find out if a previously cached root matching the new CR3/role is available.
4340 * The current root is also inserted into the cache.
4341 * If a matching root was found, it is assigned to kvm_mmu->root_hpa and true is
4342 * returned.
4343 * Otherwise, the LRU root from the cache is assigned to kvm_mmu->root_hpa and
4344 * false is returned. This root should now be freed by the caller.
4345 */
4346static bool cached_root_available(struct kvm_vcpu *vcpu, gpa_t new_cr3,
4347				  union kvm_mmu_page_role new_role)
4348{
4349	uint i;
4350	struct kvm_mmu_root_info root;
4351	struct kvm_mmu *mmu = vcpu->arch.mmu;
4352
4353	root.cr3 = mmu->root_cr3;
4354	root.hpa = mmu->root_hpa;
4355
4356	for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
4357		swap(root, mmu->prev_roots[i]);
4358
4359		if (new_cr3 == root.cr3 && VALID_PAGE(root.hpa) &&
4360		    page_header(root.hpa) != NULL &&
4361		    new_role.word == page_header(root.hpa)->role.word)
4362			break;
4363	}
4364
4365	mmu->root_hpa = root.hpa;
4366	mmu->root_cr3 = root.cr3;
4367
4368	return i < KVM_MMU_NUM_PREV_ROOTS;
4369}
4370
4371static bool fast_cr3_switch(struct kvm_vcpu *vcpu, gpa_t new_cr3,
4372			    union kvm_mmu_page_role new_role,
4373			    bool skip_tlb_flush)
4374{
4375	struct kvm_mmu *mmu = vcpu->arch.mmu;
4376
4377	/*
4378	 * For now, limit the fast switch to 64-bit hosts+VMs in order to avoid
4379	 * having to deal with PDPTEs. We may add support for 32-bit hosts/VMs
4380	 * later if necessary.
4381	 */
4382	if (mmu->shadow_root_level >= PT64_ROOT_4LEVEL &&
4383	    mmu->root_level >= PT64_ROOT_4LEVEL) {
4384		if (mmu_check_root(vcpu, new_cr3 >> PAGE_SHIFT))
4385			return false;
4386
4387		if (cached_root_available(vcpu, new_cr3, new_role)) {
4388			/*
4389			 * It is possible that the cached previous root page is
4390			 * obsolete because of a change in the MMU generation
4391			 * number. However, changing the generation number is
4392			 * accompanied by KVM_REQ_MMU_RELOAD, which will free
4393			 * the root set here and allocate a new one.
4394			 */
4395			kvm_make_request(KVM_REQ_LOAD_CR3, vcpu);
4396			if (!skip_tlb_flush) {
4397				kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
4398				kvm_x86_ops->tlb_flush(vcpu, true);
4399			}
4400
4401			/*
4402			 * The last MMIO access's GVA and GPA are cached in the
4403			 * VCPU. When switching to a new CR3, that GVA->GPA
4404			 * mapping may no longer be valid. So clear any cached
4405			 * MMIO info even when we don't need to sync the shadow
4406			 * page tables.
4407			 */
4408			vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY);
4409
4410			__clear_sp_write_flooding_count(
4411				page_header(mmu->root_hpa));
4412
4413			return true;
4414		}
4415	}
4416
4417	return false;
4418}
4419
4420static void __kvm_mmu_new_cr3(struct kvm_vcpu *vcpu, gpa_t new_cr3,
4421			      union kvm_mmu_page_role new_role,
4422			      bool skip_tlb_flush)
4423{
4424	if (!fast_cr3_switch(vcpu, new_cr3, new_role, skip_tlb_flush))
4425		kvm_mmu_free_roots(vcpu, vcpu->arch.mmu,
4426				   KVM_MMU_ROOT_CURRENT);
4427}
4428
4429void kvm_mmu_new_cr3(struct kvm_vcpu *vcpu, gpa_t new_cr3, bool skip_tlb_flush)
4430{
4431	__kvm_mmu_new_cr3(vcpu, new_cr3, kvm_mmu_calc_root_page_role(vcpu),
4432			  skip_tlb_flush);
4433}
4434EXPORT_SYMBOL_GPL(kvm_mmu_new_cr3);
4435
4436static unsigned long get_cr3(struct kvm_vcpu *vcpu)
4437{
4438	return kvm_read_cr3(vcpu);
4439}
4440
4441static void inject_page_fault(struct kvm_vcpu *vcpu,
4442			      struct x86_exception *fault)
4443{
4444	vcpu->arch.mmu->inject_page_fault(vcpu, fault);
4445}
4446
4447static bool sync_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, gfn_t gfn,
4448			   unsigned access, int *nr_present)
4449{
4450	if (unlikely(is_mmio_spte(*sptep))) {
4451		if (gfn != get_mmio_spte_gfn(*sptep)) {
4452			mmu_spte_clear_no_track(sptep);
4453			return true;
4454		}
4455
4456		(*nr_present)++;
4457		mark_mmio_spte(vcpu, sptep, gfn, access);
4458		return true;
4459	}
4460
4461	return false;
4462}
4463
4464static inline bool is_last_gpte(struct kvm_mmu *mmu,
4465				unsigned level, unsigned gpte)
4466{
4467	/*
4468	 * The RHS has bit 7 set iff level < mmu->last_nonleaf_level.
4469	 * If it is clear, there are no large pages at this level, so clear
4470	 * PT_PAGE_SIZE_MASK in gpte if that is the case.
4471	 */
4472	gpte &= level - mmu->last_nonleaf_level;
4473
4474	/*
4475	 * PT_PAGE_TABLE_LEVEL always terminates.  The RHS has bit 7 set
4476	 * iff level <= PT_PAGE_TABLE_LEVEL, which for our purpose means
4477	 * level == PT_PAGE_TABLE_LEVEL; set PT_PAGE_SIZE_MASK in gpte then.
4478	 */
4479	gpte |= level - PT_PAGE_TABLE_LEVEL - 1;
4480
4481	return gpte & PT_PAGE_SIZE_MASK;
4482}
4483
4484#define PTTYPE_EPT 18 /* arbitrary */
4485#define PTTYPE PTTYPE_EPT
4486#include "paging_tmpl.h"
4487#undef PTTYPE
4488
4489#define PTTYPE 64
4490#include "paging_tmpl.h"
4491#undef PTTYPE
4492
4493#define PTTYPE 32
4494#include "paging_tmpl.h"
4495#undef PTTYPE
4496
4497static void
4498__reset_rsvds_bits_mask(struct kvm_vcpu *vcpu,
4499			struct rsvd_bits_validate *rsvd_check,
4500			int maxphyaddr, int level, bool nx, bool gbpages,
4501			bool pse, bool amd)
4502{
4503	u64 exb_bit_rsvd = 0;
4504	u64 gbpages_bit_rsvd = 0;
4505	u64 nonleaf_bit8_rsvd = 0;
4506
4507	rsvd_check->bad_mt_xwr = 0;
4508
4509	if (!nx)
4510		exb_bit_rsvd = rsvd_bits(63, 63);
4511	if (!gbpages)
4512		gbpages_bit_rsvd = rsvd_bits(7, 7);
4513
4514	/*
4515	 * Non-leaf PML4Es and PDPEs reserve bit 8 (which would be the G bit for
4516	 * leaf entries) on AMD CPUs only.
4517	 */
4518	if (amd)
4519		nonleaf_bit8_rsvd = rsvd_bits(8, 8);
4520
4521	switch (level) {
4522	case PT32_ROOT_LEVEL:
4523		/* no rsvd bits for 2 level 4K page table entries */
4524		rsvd_check->rsvd_bits_mask[0][1] = 0;
4525		rsvd_check->rsvd_bits_mask[0][0] = 0;
4526		rsvd_check->rsvd_bits_mask[1][0] =
4527			rsvd_check->rsvd_bits_mask[0][0];
4528
4529		if (!pse) {
4530			rsvd_check->rsvd_bits_mask[1][1] = 0;
4531			break;
4532		}
4533
4534		if (is_cpuid_PSE36())
4535			/* 36bits PSE 4MB page */
4536			rsvd_check->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
4537		else
4538			/* 32 bits PSE 4MB page */
4539			rsvd_check->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
4540		break;
4541	case PT32E_ROOT_LEVEL:
4542		rsvd_check->rsvd_bits_mask[0][2] =
4543			rsvd_bits(maxphyaddr, 63) |
4544			rsvd_bits(5, 8) | rsvd_bits(1, 2);	/* PDPTE */
4545		rsvd_check->rsvd_bits_mask[0][1] = exb_bit_rsvd |
4546			rsvd_bits(maxphyaddr, 62);	/* PDE */
4547		rsvd_check->rsvd_bits_mask[0][0] = exb_bit_rsvd |
4548			rsvd_bits(maxphyaddr, 62); 	/* PTE */
4549		rsvd_check->rsvd_bits_mask[1][1] = exb_bit_rsvd |
4550			rsvd_bits(maxphyaddr, 62) |
4551			rsvd_bits(13, 20);		/* large page */
4552		rsvd_check->rsvd_bits_mask[1][0] =
4553			rsvd_check->rsvd_bits_mask[0][0];
4554		break;
4555	case PT64_ROOT_5LEVEL:
4556		rsvd_check->rsvd_bits_mask[0][4] = exb_bit_rsvd |
4557			nonleaf_bit8_rsvd | rsvd_bits(7, 7) |
4558			rsvd_bits(maxphyaddr, 51);
4559		rsvd_check->rsvd_bits_mask[1][4] =
4560			rsvd_check->rsvd_bits_mask[0][4];
4561		/* fall through */
4562	case PT64_ROOT_4LEVEL:
4563		rsvd_check->rsvd_bits_mask[0][3] = exb_bit_rsvd |
4564			nonleaf_bit8_rsvd | rsvd_bits(7, 7) |
4565			rsvd_bits(maxphyaddr, 51);
4566		rsvd_check->rsvd_bits_mask[0][2] = exb_bit_rsvd |
4567			nonleaf_bit8_rsvd | gbpages_bit_rsvd |
4568			rsvd_bits(maxphyaddr, 51);
4569		rsvd_check->rsvd_bits_mask[0][1] = exb_bit_rsvd |
4570			rsvd_bits(maxphyaddr, 51);
4571		rsvd_check->rsvd_bits_mask[0][0] = exb_bit_rsvd |
4572			rsvd_bits(maxphyaddr, 51);
4573		rsvd_check->rsvd_bits_mask[1][3] =
4574			rsvd_check->rsvd_bits_mask[0][3];
4575		rsvd_check->rsvd_bits_mask[1][2] = exb_bit_rsvd |
4576			gbpages_bit_rsvd | rsvd_bits(maxphyaddr, 51) |
4577			rsvd_bits(13, 29);
4578		rsvd_check->rsvd_bits_mask[1][1] = exb_bit_rsvd |
4579			rsvd_bits(maxphyaddr, 51) |
4580			rsvd_bits(13, 20);		/* large page */
4581		rsvd_check->rsvd_bits_mask[1][0] =
4582			rsvd_check->rsvd_bits_mask[0][0];
4583		break;
4584	}
4585}
4586
4587static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu,
4588				  struct kvm_mmu *context)
4589{
4590	__reset_rsvds_bits_mask(vcpu, &context->guest_rsvd_check,
4591				cpuid_maxphyaddr(vcpu), context->root_level,
4592				context->nx,
4593				guest_cpuid_has(vcpu, X86_FEATURE_GBPAGES),
4594				is_pse(vcpu), guest_cpuid_is_amd(vcpu));
4595}
4596
4597static void
4598__reset_rsvds_bits_mask_ept(struct rsvd_bits_validate *rsvd_check,
4599			    int maxphyaddr, bool execonly)
4600{
4601	u64 bad_mt_xwr;
4602
4603	rsvd_check->rsvd_bits_mask[0][4] =
4604		rsvd_bits(maxphyaddr, 51) | rsvd_bits(3, 7);
4605	rsvd_check->rsvd_bits_mask[0][3] =
4606		rsvd_bits(maxphyaddr, 51) | rsvd_bits(3, 7);
4607	rsvd_check->rsvd_bits_mask[0][2] =
4608		rsvd_bits(maxphyaddr, 51) | rsvd_bits(3, 6);
4609	rsvd_check->rsvd_bits_mask[0][1] =
4610		rsvd_bits(maxphyaddr, 51) | rsvd_bits(3, 6);
4611	rsvd_check->rsvd_bits_mask[0][0] = rsvd_bits(maxphyaddr, 51);
4612
4613	/* large page */
4614	rsvd_check->rsvd_bits_mask[1][4] = rsvd_check->rsvd_bits_mask[0][4];
4615	rsvd_check->rsvd_bits_mask[1][3] = rsvd_check->rsvd_bits_mask[0][3];
4616	rsvd_check->rsvd_bits_mask[1][2] =
4617		rsvd_bits(maxphyaddr, 51) | rsvd_bits(12, 29);
4618	rsvd_check->rsvd_bits_mask[1][1] =
4619		rsvd_bits(maxphyaddr, 51) | rsvd_bits(12, 20);
4620	rsvd_check->rsvd_bits_mask[1][0] = rsvd_check->rsvd_bits_mask[0][0];
4621
4622	bad_mt_xwr = 0xFFull << (2 * 8);	/* bits 3..5 must not be 2 */
4623	bad_mt_xwr |= 0xFFull << (3 * 8);	/* bits 3..5 must not be 3 */
4624	bad_mt_xwr |= 0xFFull << (7 * 8);	/* bits 3..5 must not be 7 */
4625	bad_mt_xwr |= REPEAT_BYTE(1ull << 2);	/* bits 0..2 must not be 010 */
4626	bad_mt_xwr |= REPEAT_BYTE(1ull << 6);	/* bits 0..2 must not be 110 */
4627	if (!execonly) {
4628		/* bits 0..2 must not be 100 unless VMX capabilities allow it */
4629		bad_mt_xwr |= REPEAT_BYTE(1ull << 4);
4630	}
4631	rsvd_check->bad_mt_xwr = bad_mt_xwr;
4632}
4633
4634static void reset_rsvds_bits_mask_ept(struct kvm_vcpu *vcpu,
4635		struct kvm_mmu *context, bool execonly)
4636{
4637	__reset_rsvds_bits_mask_ept(&context->guest_rsvd_check,
4638				    cpuid_maxphyaddr(vcpu), execonly);
4639}
4640
4641/*
4642 * the page table on host is the shadow page table for the page
4643 * table in guest or amd nested guest, its mmu features completely
4644 * follow the features in guest.
4645 */
4646void
4647reset_shadow_zero_bits_mask(struct kvm_vcpu *vcpu, struct kvm_mmu *context)
4648{
4649	bool uses_nx = context->nx ||
4650		context->mmu_role.base.smep_andnot_wp;
4651	struct rsvd_bits_validate *shadow_zero_check;
4652	int i;
4653
4654	/*
4655	 * Passing "true" to the last argument is okay; it adds a check
4656	 * on bit 8 of the SPTEs which KVM doesn't use anyway.
4657	 */
4658	shadow_zero_check = &context->shadow_zero_check;
4659	__reset_rsvds_bits_mask(vcpu, shadow_zero_check,
4660				shadow_phys_bits,
4661				context->shadow_root_level, uses_nx,
4662				guest_cpuid_has(vcpu, X86_FEATURE_GBPAGES),
4663				is_pse(vcpu), true);
4664
4665	if (!shadow_me_mask)
4666		return;
4667
4668	for (i = context->shadow_root_level; --i >= 0;) {
4669		shadow_zero_check->rsvd_bits_mask[0][i] &= ~shadow_me_mask;
4670		shadow_zero_check->rsvd_bits_mask[1][i] &= ~shadow_me_mask;
4671	}
4672
4673}
4674EXPORT_SYMBOL_GPL(reset_shadow_zero_bits_mask);
4675
4676static inline bool boot_cpu_is_amd(void)
4677{
4678	WARN_ON_ONCE(!tdp_enabled);
4679	return shadow_x_mask == 0;
4680}
4681
4682/*
4683 * the direct page table on host, use as much mmu features as
4684 * possible, however, kvm currently does not do execution-protection.
4685 */
4686static void
4687reset_tdp_shadow_zero_bits_mask(struct kvm_vcpu *vcpu,
4688				struct kvm_mmu *context)
4689{
4690	struct rsvd_bits_validate *shadow_zero_check;
4691	int i;
4692
4693	shadow_zero_check = &context->shadow_zero_check;
4694
4695	if (boot_cpu_is_amd())
4696		__reset_rsvds_bits_mask(vcpu, shadow_zero_check,
4697					shadow_phys_bits,
4698					context->shadow_root_level, false,
4699					boot_cpu_has(X86_FEATURE_GBPAGES),
4700					true, true);
4701	else
4702		__reset_rsvds_bits_mask_ept(shadow_zero_check,
4703					    shadow_phys_bits,
4704					    false);
4705
4706	if (!shadow_me_mask)
4707		return;
4708
4709	for (i = context->shadow_root_level; --i >= 0;) {
4710		shadow_zero_check->rsvd_bits_mask[0][i] &= ~shadow_me_mask;
4711		shadow_zero_check->rsvd_bits_mask[1][i] &= ~shadow_me_mask;
4712	}
4713}
4714
4715/*
4716 * as the comments in reset_shadow_zero_bits_mask() except it
4717 * is the shadow page table for intel nested guest.
4718 */
4719static void
4720reset_ept_shadow_zero_bits_mask(struct kvm_vcpu *vcpu,
4721				struct kvm_mmu *context, bool execonly)
4722{
4723	__reset_rsvds_bits_mask_ept(&context->shadow_zero_check,
4724				    shadow_phys_bits, execonly);
4725}
4726
4727#define BYTE_MASK(access) \
4728	((1 & (access) ? 2 : 0) | \
4729	 (2 & (access) ? 4 : 0) | \
4730	 (3 & (access) ? 8 : 0) | \
4731	 (4 & (access) ? 16 : 0) | \
4732	 (5 & (access) ? 32 : 0) | \
4733	 (6 & (access) ? 64 : 0) | \
4734	 (7 & (access) ? 128 : 0))
4735
4736
4737static void update_permission_bitmask(struct kvm_vcpu *vcpu,
4738				      struct kvm_mmu *mmu, bool ept)
4739{
4740	unsigned byte;
4741
4742	const u8 x = BYTE_MASK(ACC_EXEC_MASK);
4743	const u8 w = BYTE_MASK(ACC_WRITE_MASK);
4744	const u8 u = BYTE_MASK(ACC_USER_MASK);
4745
4746	bool cr4_smep = kvm_read_cr4_bits(vcpu, X86_CR4_SMEP) != 0;
4747	bool cr4_smap = kvm_read_cr4_bits(vcpu, X86_CR4_SMAP) != 0;
4748	bool cr0_wp = is_write_protection(vcpu);
4749
4750	for (byte = 0; byte < ARRAY_SIZE(mmu->permissions); ++byte) {
4751		unsigned pfec = byte << 1;
4752
4753		/*
4754		 * Each "*f" variable has a 1 bit for each UWX value
4755		 * that causes a fault with the given PFEC.
4756		 */
4757
4758		/* Faults from writes to non-writable pages */
4759		u8 wf = (pfec & PFERR_WRITE_MASK) ? (u8)~w : 0;
4760		/* Faults from user mode accesses to supervisor pages */
4761		u8 uf = (pfec & PFERR_USER_MASK) ? (u8)~u : 0;
4762		/* Faults from fetches of non-executable pages*/
4763		u8 ff = (pfec & PFERR_FETCH_MASK) ? (u8)~x : 0;
4764		/* Faults from kernel mode fetches of user pages */
4765		u8 smepf = 0;
4766		/* Faults from kernel mode accesses of user pages */
4767		u8 smapf = 0;
4768
4769		if (!ept) {
4770			/* Faults from kernel mode accesses to user pages */
4771			u8 kf = (pfec & PFERR_USER_MASK) ? 0 : u;
4772
4773			/* Not really needed: !nx will cause pte.nx to fault */
4774			if (!mmu->nx)
4775				ff = 0;
4776
4777			/* Allow supervisor writes if !cr0.wp */
4778			if (!cr0_wp)
4779				wf = (pfec & PFERR_USER_MASK) ? wf : 0;
4780
4781			/* Disallow supervisor fetches of user code if cr4.smep */
4782			if (cr4_smep)
4783				smepf = (pfec & PFERR_FETCH_MASK) ? kf : 0;
4784
4785			/*
4786			 * SMAP:kernel-mode data accesses from user-mode
4787			 * mappings should fault. A fault is considered
4788			 * as a SMAP violation if all of the following
4789			 * conditions are true:
4790			 *   - X86_CR4_SMAP is set in CR4
4791			 *   - A user page is accessed
4792			 *   - The access is not a fetch
4793			 *   - Page fault in kernel mode
4794			 *   - if CPL = 3 or X86_EFLAGS_AC is clear
4795			 *
4796			 * Here, we cover the first three conditions.
4797			 * The fourth is computed dynamically in permission_fault();
4798			 * PFERR_RSVD_MASK bit will be set in PFEC if the access is
4799			 * *not* subject to SMAP restrictions.
4800			 */
4801			if (cr4_smap)
4802				smapf = (pfec & (PFERR_RSVD_MASK|PFERR_FETCH_MASK)) ? 0 : kf;
4803		}
4804
4805		mmu->permissions[byte] = ff | uf | wf | smepf | smapf;
4806	}
4807}
4808
4809/*
4810* PKU is an additional mechanism by which the paging controls access to
4811* user-mode addresses based on the value in the PKRU register.  Protection
4812* key violations are reported through a bit in the page fault error code.
4813* Unlike other bits of the error code, the PK bit is not known at the
4814* call site of e.g. gva_to_gpa; it must be computed directly in
4815* permission_fault based on two bits of PKRU, on some machine state (CR4,
4816* CR0, EFER, CPL), and on other bits of the error code and the page tables.
4817*
4818* In particular the following conditions come from the error code, the
4819* page tables and the machine state:
4820* - PK is always zero unless CR4.PKE=1 and EFER.LMA=1
4821* - PK is always zero if RSVD=1 (reserved bit set) or F=1 (instruction fetch)
4822* - PK is always zero if U=0 in the page tables
4823* - PKRU.WD is ignored if CR0.WP=0 and the access is a supervisor access.
4824*
4825* The PKRU bitmask caches the result of these four conditions.  The error
4826* code (minus the P bit) and the page table's U bit form an index into the
4827* PKRU bitmask.  Two bits of the PKRU bitmask are then extracted and ANDed
4828* with the two bits of the PKRU register corresponding to the protection key.
4829* For the first three conditions above the bits will be 00, thus masking
4830* away both AD and WD.  For all reads or if the last condition holds, WD
4831* only will be masked away.
4832*/
4833static void update_pkru_bitmask(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
4834				bool ept)
4835{
4836	unsigned bit;
4837	bool wp;
4838
4839	if (ept) {
4840		mmu->pkru_mask = 0;
4841		return;
4842	}
4843
4844	/* PKEY is enabled only if CR4.PKE and EFER.LMA are both set. */
4845	if (!kvm_read_cr4_bits(vcpu, X86_CR4_PKE) || !is_long_mode(vcpu)) {
4846		mmu->pkru_mask = 0;
4847		return;
4848	}
4849
4850	wp = is_write_protection(vcpu);
4851
4852	for (bit = 0; bit < ARRAY_SIZE(mmu->permissions); ++bit) {
4853		unsigned pfec, pkey_bits;
4854		bool check_pkey, check_write, ff, uf, wf, pte_user;
4855
4856		pfec = bit << 1;
4857		ff = pfec & PFERR_FETCH_MASK;
4858		uf = pfec & PFERR_USER_MASK;
4859		wf = pfec & PFERR_WRITE_MASK;
4860
4861		/* PFEC.RSVD is replaced by ACC_USER_MASK. */
4862		pte_user = pfec & PFERR_RSVD_MASK;
4863
4864		/*
4865		 * Only need to check the access which is not an
4866		 * instruction fetch and is to a user page.
4867		 */
4868		check_pkey = (!ff && pte_user);
4869		/*
4870		 * write access is controlled by PKRU if it is a
4871		 * user access or CR0.WP = 1.
4872		 */
4873		check_write = check_pkey && wf && (uf || wp);
4874
4875		/* PKRU.AD stops both read and write access. */
4876		pkey_bits = !!check_pkey;
4877		/* PKRU.WD stops write access. */
4878		pkey_bits |= (!!check_write) << 1;
4879
4880		mmu->pkru_mask |= (pkey_bits & 3) << pfec;
4881	}
4882}
4883
4884static void update_last_nonleaf_level(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu)
4885{
4886	unsigned root_level = mmu->root_level;
4887
4888	mmu->last_nonleaf_level = root_level;
4889	if (root_level == PT32_ROOT_LEVEL && is_pse(vcpu))
4890		mmu->last_nonleaf_level++;
4891}
4892
4893static void paging64_init_context_common(struct kvm_vcpu *vcpu,
4894					 struct kvm_mmu *context,
4895					 int level)
4896{
4897	context->nx = is_nx(vcpu);
4898	context->root_level = level;
4899
4900	reset_rsvds_bits_mask(vcpu, context);
4901	update_permission_bitmask(vcpu, context, false);
4902	update_pkru_bitmask(vcpu, context, false);
4903	update_last_nonleaf_level(vcpu, context);
4904
4905	MMU_WARN_ON(!is_pae(vcpu));
4906	context->page_fault = paging64_page_fault;
4907	context->gva_to_gpa = paging64_gva_to_gpa;
4908	context->sync_page = paging64_sync_page;
4909	context->invlpg = paging64_invlpg;
4910	context->update_pte = paging64_update_pte;
4911	context->shadow_root_level = level;
4912	context->direct_map = false;
4913}
4914
4915static void paging64_init_context(struct kvm_vcpu *vcpu,
4916				  struct kvm_mmu *context)
4917{
4918	int root_level = is_la57_mode(vcpu) ?
4919			 PT64_ROOT_5LEVEL : PT64_ROOT_4LEVEL;
4920
4921	paging64_init_context_common(vcpu, context, root_level);
4922}
4923
4924static void paging32_init_context(struct kvm_vcpu *vcpu,
4925				  struct kvm_mmu *context)
4926{
4927	context->nx = false;
4928	context->root_level = PT32_ROOT_LEVEL;
4929
4930	reset_rsvds_bits_mask(vcpu, context);
4931	update_permission_bitmask(vcpu, context, false);
4932	update_pkru_bitmask(vcpu, context, false);
4933	update_last_nonleaf_level(vcpu, context);
4934
4935	context->page_fault = paging32_page_fault;
4936	context->gva_to_gpa = paging32_gva_to_gpa;
4937	context->sync_page = paging32_sync_page;
4938	context->invlpg = paging32_invlpg;
4939	context->update_pte = paging32_update_pte;
4940	context->shadow_root_level = PT32E_ROOT_LEVEL;
4941	context->direct_map = false;
4942}
4943
4944static void paging32E_init_context(struct kvm_vcpu *vcpu,
4945				   struct kvm_mmu *context)
4946{
4947	paging64_init_context_common(vcpu, context, PT32E_ROOT_LEVEL);
4948}
4949
4950static union kvm_mmu_extended_role kvm_calc_mmu_role_ext(struct kvm_vcpu *vcpu)
4951{
4952	union kvm_mmu_extended_role ext = {0};
4953
4954	ext.cr0_pg = !!is_paging(vcpu);
4955	ext.cr4_pae = !!is_pae(vcpu);
4956	ext.cr4_smep = !!kvm_read_cr4_bits(vcpu, X86_CR4_SMEP);
4957	ext.cr4_smap = !!kvm_read_cr4_bits(vcpu, X86_CR4_SMAP);
4958	ext.cr4_pse = !!is_pse(vcpu);
4959	ext.cr4_pke = !!kvm_read_cr4_bits(vcpu, X86_CR4_PKE);
4960	ext.cr4_la57 = !!kvm_read_cr4_bits(vcpu, X86_CR4_LA57);
4961	ext.maxphyaddr = cpuid_maxphyaddr(vcpu);
4962
4963	ext.valid = 1;
4964
4965	return ext;
4966}
4967
4968static union kvm_mmu_role kvm_calc_mmu_role_common(struct kvm_vcpu *vcpu,
4969						   bool base_only)
4970{
4971	union kvm_mmu_role role = {0};
4972
4973	role.base.access = ACC_ALL;
4974	role.base.nxe = !!is_nx(vcpu);
4975	role.base.cr0_wp = is_write_protection(vcpu);
4976	role.base.smm = is_smm(vcpu);
4977	role.base.guest_mode = is_guest_mode(vcpu);
4978
4979	if (base_only)
4980		return role;
4981
4982	role.ext = kvm_calc_mmu_role_ext(vcpu);
4983
4984	return role;
4985}
4986
4987static union kvm_mmu_role
4988kvm_calc_tdp_mmu_root_page_role(struct kvm_vcpu *vcpu, bool base_only)
4989{
4990	union kvm_mmu_role role = kvm_calc_mmu_role_common(vcpu, base_only);
4991
4992	role.base.ad_disabled = (shadow_accessed_mask == 0);
4993	role.base.level = kvm_x86_ops->get_tdp_level(vcpu);
4994	role.base.direct = true;
4995	role.base.gpte_is_8_bytes = true;
4996
4997	return role;
4998}
4999
5000static void init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
5001{
5002	struct kvm_mmu *context = vcpu->arch.mmu;
5003	union kvm_mmu_role new_role =
5004		kvm_calc_tdp_mmu_root_page_role(vcpu, false);
5005
5006	new_role.base.word &= mmu_base_role_mask.word;
5007	if (new_role.as_u64 == context->mmu_role.as_u64)
5008		return;
5009
5010	context->mmu_role.as_u64 = new_role.as_u64;
5011	context->page_fault = tdp_page_fault;
5012	context->sync_page = nonpaging_sync_page;
5013	context->invlpg = nonpaging_invlpg;
5014	context->update_pte = nonpaging_update_pte;
5015	context->shadow_root_level = kvm_x86_ops->get_tdp_level(vcpu);
5016	context->direct_map = true;
5017	context->set_cr3 = kvm_x86_ops->set_tdp_cr3;
5018	context->get_cr3 = get_cr3;
5019	context->get_pdptr = kvm_pdptr_read;
5020	context->inject_page_fault = kvm_inject_page_fault;
5021
5022	if (!is_paging(vcpu)) {
5023		context->nx = false;
5024		context->gva_to_gpa = nonpaging_gva_to_gpa;
5025		context->root_level = 0;
5026	} else if (is_long_mode(vcpu)) {
5027		context->nx = is_nx(vcpu);
5028		context->root_level = is_la57_mode(vcpu) ?
5029				PT64_ROOT_5LEVEL : PT64_ROOT_4LEVEL;
5030		reset_rsvds_bits_mask(vcpu, context);
5031		context->gva_to_gpa = paging64_gva_to_gpa;
5032	} else if (is_pae(vcpu)) {
5033		context->nx = is_nx(vcpu);
5034		context->root_level = PT32E_ROOT_LEVEL;
5035		reset_rsvds_bits_mask(vcpu, context);
5036		context->gva_to_gpa = paging64_gva_to_gpa;
5037	} else {
5038		context->nx = false;
5039		context->root_level = PT32_ROOT_LEVEL;
5040		reset_rsvds_bits_mask(vcpu, context);
5041		context->gva_to_gpa = paging32_gva_to_gpa;
5042	}
5043
5044	update_permission_bitmask(vcpu, context, false);
5045	update_pkru_bitmask(vcpu, context, false);
5046	update_last_nonleaf_level(vcpu, context);
5047	reset_tdp_shadow_zero_bits_mask(vcpu, context);
5048}
5049
5050static union kvm_mmu_role
5051kvm_calc_shadow_mmu_root_page_role(struct kvm_vcpu *vcpu, bool base_only)
5052{
5053	union kvm_mmu_role role = kvm_calc_mmu_role_common(vcpu, base_only);
5054
5055	role.base.smep_andnot_wp = role.ext.cr4_smep &&
5056		!is_write_protection(vcpu);
5057	role.base.smap_andnot_wp = role.ext.cr4_smap &&
5058		!is_write_protection(vcpu);
5059	role.base.direct = !is_paging(vcpu);
5060	role.base.gpte_is_8_bytes = !!is_pae(vcpu);
5061
5062	if (!is_long_mode(vcpu))
5063		role.base.level = PT32E_ROOT_LEVEL;
5064	else if (is_la57_mode(vcpu))
5065		role.base.level = PT64_ROOT_5LEVEL;
5066	else
5067		role.base.level = PT64_ROOT_4LEVEL;
5068
5069	return role;
5070}
5071
5072void kvm_init_shadow_mmu(struct kvm_vcpu *vcpu)
5073{
5074	struct kvm_mmu *context = vcpu->arch.mmu;
5075	union kvm_mmu_role new_role =
5076		kvm_calc_shadow_mmu_root_page_role(vcpu, false);
5077
5078	new_role.base.word &= mmu_base_role_mask.word;
5079	if (new_role.as_u64 == context->mmu_role.as_u64)
5080		return;
5081
5082	if (!is_paging(vcpu))
5083		nonpaging_init_context(vcpu, context);
5084	else if (is_long_mode(vcpu))
5085		paging64_init_context(vcpu, context);
5086	else if (is_pae(vcpu))
5087		paging32E_init_context(vcpu, context);
5088	else
5089		paging32_init_context(vcpu, context);
5090
5091	context->mmu_role.as_u64 = new_role.as_u64;
5092	reset_shadow_zero_bits_mask(vcpu, context);
5093}
5094EXPORT_SYMBOL_GPL(kvm_init_shadow_mmu);
5095
5096static union kvm_mmu_role
5097kvm_calc_shadow_ept_root_page_role(struct kvm_vcpu *vcpu, bool accessed_dirty,
5098				   bool execonly)
5099{
5100	union kvm_mmu_role role = {0};
5101
5102	/* SMM flag is inherited from root_mmu */
5103	role.base.smm = vcpu->arch.root_mmu.mmu_role.base.smm;
5104
5105	role.base.level = PT64_ROOT_4LEVEL;
5106	role.base.gpte_is_8_bytes = true;
5107	role.base.direct = false;
5108	role.base.ad_disabled = !accessed_dirty;
5109	role.base.guest_mode = true;
5110	role.base.access = ACC_ALL;
5111
5112	/*
5113	 * WP=1 and NOT_WP=1 is an impossible combination, use WP and the
5114	 * SMAP variation to denote shadow EPT entries.
5115	 */
5116	role.base.cr0_wp = true;
5117	role.base.smap_andnot_wp = true;
5118
5119	role.ext = kvm_calc_mmu_role_ext(vcpu);
5120	role.ext.execonly = execonly;
5121
5122	return role;
5123}
5124
5125void kvm_init_shadow_ept_mmu(struct kvm_vcpu *vcpu, bool execonly,
5126			     bool accessed_dirty, gpa_t new_eptp)
5127{
5128	struct kvm_mmu *context = vcpu->arch.mmu;
5129	union kvm_mmu_role new_role =
5130		kvm_calc_shadow_ept_root_page_role(vcpu, accessed_dirty,
5131						   execonly);
5132
5133	__kvm_mmu_new_cr3(vcpu, new_eptp, new_role.base, false);
5134
5135	new_role.base.word &= mmu_base_role_mask.word;
5136	if (new_role.as_u64 == context->mmu_role.as_u64)
5137		return;
5138
5139	context->shadow_root_level = PT64_ROOT_4LEVEL;
5140
5141	context->nx = true;
5142	context->ept_ad = accessed_dirty;
5143	context->page_fault = ept_page_fault;
5144	context->gva_to_gpa = ept_gva_to_gpa;
5145	context->sync_page = ept_sync_page;
5146	context->invlpg = ept_invlpg;
5147	context->update_pte = ept_update_pte;
5148	context->root_level = PT64_ROOT_4LEVEL;
5149	context->direct_map = false;
5150	context->mmu_role.as_u64 = new_role.as_u64;
5151
5152	update_permission_bitmask(vcpu, context, true);
5153	update_pkru_bitmask(vcpu, context, true);
5154	update_last_nonleaf_level(vcpu, context);
5155	reset_rsvds_bits_mask_ept(vcpu, context, execonly);
5156	reset_ept_shadow_zero_bits_mask(vcpu, context, execonly);
5157}
5158EXPORT_SYMBOL_GPL(kvm_init_shadow_ept_mmu);
5159
5160static void init_kvm_softmmu(struct kvm_vcpu *vcpu)
5161{
5162	struct kvm_mmu *context = vcpu->arch.mmu;
5163
5164	kvm_init_shadow_mmu(vcpu);
5165	context->set_cr3           = kvm_x86_ops->set_cr3;
5166	context->get_cr3           = get_cr3;
5167	context->get_pdptr         = kvm_pdptr_read;
5168	context->inject_page_fault = kvm_inject_page_fault;
5169}
5170
5171static void init_kvm_nested_mmu(struct kvm_vcpu *vcpu)
5172{
5173	union kvm_mmu_role new_role = kvm_calc_mmu_role_common(vcpu, false);
5174	struct kvm_mmu *g_context = &vcpu->arch.nested_mmu;
5175
5176	new_role.base.word &= mmu_base_role_mask.word;
5177	if (new_role.as_u64 == g_context->mmu_role.as_u64)
5178		return;
5179
5180	g_context->mmu_role.as_u64 = new_role.as_u64;
5181	g_context->get_cr3           = get_cr3;
5182	g_context->get_pdptr         = kvm_pdptr_read;
5183	g_context->inject_page_fault = kvm_inject_page_fault;
5184
5185	/*
5186	 * Note that arch.mmu->gva_to_gpa translates l2_gpa to l1_gpa using
5187	 * L1's nested page tables (e.g. EPT12). The nested translation
5188	 * of l2_gva to l1_gpa is done by arch.nested_mmu.gva_to_gpa using
5189	 * L2's page tables as the first level of translation and L1's
5190	 * nested page tables as the second level of translation. Basically
5191	 * the gva_to_gpa functions between mmu and nested_mmu are swapped.
5192	 */
5193	if (!is_paging(vcpu)) {
5194		g_context->nx = false;
5195		g_context->root_level = 0;
5196		g_context->gva_to_gpa = nonpaging_gva_to_gpa_nested;
5197	} else if (is_long_mode(vcpu)) {
5198		g_context->nx = is_nx(vcpu);
5199		g_context->root_level = is_la57_mode(vcpu) ?
5200					PT64_ROOT_5LEVEL : PT64_ROOT_4LEVEL;
5201		reset_rsvds_bits_mask(vcpu, g_context);
5202		g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
5203	} else if (is_pae(vcpu)) {
5204		g_context->nx = is_nx(vcpu);
5205		g_context->root_level = PT32E_ROOT_LEVEL;
5206		reset_rsvds_bits_mask(vcpu, g_context);
5207		g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
5208	} else {
5209		g_context->nx = false;
5210		g_context->root_level = PT32_ROOT_LEVEL;
5211		reset_rsvds_bits_mask(vcpu, g_context);
5212		g_context->gva_to_gpa = paging32_gva_to_gpa_nested;
5213	}
5214
5215	update_permission_bitmask(vcpu, g_context, false);
5216	update_pkru_bitmask(vcpu, g_context, false);
5217	update_last_nonleaf_level(vcpu, g_context);
5218}
5219
5220void kvm_init_mmu(struct kvm_vcpu *vcpu, bool reset_roots)
5221{
5222	if (reset_roots) {
5223		uint i;
5224
5225		vcpu->arch.mmu->root_hpa = INVALID_PAGE;
5226
5227		for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
5228			vcpu->arch.mmu->prev_roots[i] = KVM_MMU_ROOT_INFO_INVALID;
5229	}
5230
5231	if (mmu_is_nested(vcpu))
5232		init_kvm_nested_mmu(vcpu);
5233	else if (tdp_enabled)
5234		init_kvm_tdp_mmu(vcpu);
5235	else
5236		init_kvm_softmmu(vcpu);
5237}
5238EXPORT_SYMBOL_GPL(kvm_init_mmu);
5239
5240static union kvm_mmu_page_role
5241kvm_mmu_calc_root_page_role(struct kvm_vcpu *vcpu)
5242{
5243	union kvm_mmu_role role;
5244
5245	if (tdp_enabled)
5246		role = kvm_calc_tdp_mmu_root_page_role(vcpu, true);
5247	else
5248		role = kvm_calc_shadow_mmu_root_page_role(vcpu, true);
5249
5250	return role.base;
5251}
5252
5253void kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
5254{
5255	kvm_mmu_unload(vcpu);
5256	kvm_init_mmu(vcpu, true);
5257}
5258EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
5259
5260int kvm_mmu_load(struct kvm_vcpu *vcpu)
5261{
5262	int r;
5263
5264	r = mmu_topup_memory_caches(vcpu);
5265	if (r)
5266		goto out;
5267	r = mmu_alloc_roots(vcpu);
5268	kvm_mmu_sync_roots(vcpu);
5269	if (r)
5270		goto out;
5271	kvm_mmu_load_cr3(vcpu);
5272	kvm_x86_ops->tlb_flush(vcpu, true);
5273out:
5274	return r;
5275}
5276EXPORT_SYMBOL_GPL(kvm_mmu_load);
5277
5278void kvm_mmu_unload(struct kvm_vcpu *vcpu)
5279{
5280	kvm_mmu_free_roots(vcpu, &vcpu->arch.root_mmu, KVM_MMU_ROOTS_ALL);
5281	WARN_ON(VALID_PAGE(vcpu->arch.root_mmu.root_hpa));
5282	kvm_mmu_free_roots(vcpu, &vcpu->arch.guest_mmu, KVM_MMU_ROOTS_ALL);
5283	WARN_ON(VALID_PAGE(vcpu->arch.guest_mmu.root_hpa));
5284}
5285EXPORT_SYMBOL_GPL(kvm_mmu_unload);
5286
5287static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
5288				  struct kvm_mmu_page *sp, u64 *spte,
5289				  const void *new)
5290{
5291	if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
5292		++vcpu->kvm->stat.mmu_pde_zapped;
5293		return;
5294        }
5295
5296	++vcpu->kvm->stat.mmu_pte_updated;
5297	vcpu->arch.mmu->update_pte(vcpu, sp, spte, new);
5298}
5299
5300static bool need_remote_flush(u64 old, u64 new)
5301{
5302	if (!is_shadow_present_pte(old))
5303		return false;
5304	if (!is_shadow_present_pte(new))
5305		return true;
5306	if ((old ^ new) & PT64_BASE_ADDR_MASK)
5307		return true;
5308	old ^= shadow_nx_mask;
5309	new ^= shadow_nx_mask;
5310	return (old & ~new & PT64_PERM_MASK) != 0;
5311}
5312
5313static u64 mmu_pte_write_fetch_gpte(struct kvm_vcpu *vcpu, gpa_t *gpa,
5314				    int *bytes)
5315{
5316	u64 gentry = 0;
5317	int r;
5318
5319	/*
5320	 * Assume that the pte write on a page table of the same type
5321	 * as the current vcpu paging mode since we update the sptes only
5322	 * when they have the same mode.
5323	 */
5324	if (is_pae(vcpu) && *bytes == 4) {
5325		/* Handle a 32-bit guest writing two halves of a 64-bit gpte */
5326		*gpa &= ~(gpa_t)7;
5327		*bytes = 8;
5328	}
5329
5330	if (*bytes == 4 || *bytes == 8) {
5331		r = kvm_vcpu_read_guest_atomic(vcpu, *gpa, &gentry, *bytes);
5332		if (r)
5333			gentry = 0;
5334	}
5335
5336	return gentry;
5337}
5338
5339/*
5340 * If we're seeing too many writes to a page, it may no longer be a page table,
5341 * or we may be forking, in which case it is better to unmap the page.
5342 */
5343static bool detect_write_flooding(struct kvm_mmu_page *sp)
5344{
5345	/*
5346	 * Skip write-flooding detected for the sp whose level is 1, because
5347	 * it can become unsync, then the guest page is not write-protected.
5348	 */
5349	if (sp->role.level == PT_PAGE_TABLE_LEVEL)
5350		return false;
5351
5352	atomic_inc(&sp->write_flooding_count);
5353	return atomic_read(&sp->write_flooding_count) >= 3;
5354}
5355
5356/*
5357 * Misaligned accesses are too much trouble to fix up; also, they usually
5358 * indicate a page is not used as a page table.
5359 */
5360static bool detect_write_misaligned(struct kvm_mmu_page *sp, gpa_t gpa,
5361				    int bytes)
5362{
5363	unsigned offset, pte_size, misaligned;
5364
5365	pgprintk("misaligned: gpa %llx bytes %d role %x\n",
5366		 gpa, bytes, sp->role.word);
5367
5368	offset = offset_in_page(gpa);
5369	pte_size = sp->role.gpte_is_8_bytes ? 8 : 4;
5370
5371	/*
5372	 * Sometimes, the OS only writes the last one bytes to update status
5373	 * bits, for example, in linux, andb instruction is used in clear_bit().
5374	 */
5375	if (!(offset & (pte_size - 1)) && bytes == 1)
5376		return false;
5377
5378	misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
5379	misaligned |= bytes < 4;
5380
5381	return misaligned;
5382}
5383
5384static u64 *get_written_sptes(struct kvm_mmu_page *sp, gpa_t gpa, int *nspte)
5385{
5386	unsigned page_offset, quadrant;
5387	u64 *spte;
5388	int level;
5389
5390	page_offset = offset_in_page(gpa);
5391	level = sp->role.level;
5392	*nspte = 1;
5393	if (!sp->role.gpte_is_8_bytes) {
5394		page_offset <<= 1;	/* 32->64 */
5395		/*
5396		 * A 32-bit pde maps 4MB while the shadow pdes map
5397		 * only 2MB.  So we need to double the offset again
5398		 * and zap two pdes instead of one.
5399		 */
5400		if (level == PT32_ROOT_LEVEL) {
5401			page_offset &= ~7; /* kill rounding error */
5402			page_offset <<= 1;
5403			*nspte = 2;
5404		}
5405		quadrant = page_offset >> PAGE_SHIFT;
5406		page_offset &= ~PAGE_MASK;
5407		if (quadrant != sp->role.quadrant)
5408			return NULL;
5409	}
5410
5411	spte = &sp->spt[page_offset / sizeof(*spte)];
5412	return spte;
5413}
5414
5415static void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
5416			      const u8 *new, int bytes,
5417			      struct kvm_page_track_notifier_node *node)
5418{
5419	gfn_t gfn = gpa >> PAGE_SHIFT;
5420	struct kvm_mmu_page *sp;
5421	LIST_HEAD(invalid_list);
5422	u64 entry, gentry, *spte;
5423	int npte;
5424	bool remote_flush, local_flush;
5425
5426	/*
5427	 * If we don't have indirect shadow pages, it means no page is
5428	 * write-protected, so we can exit simply.
5429	 */
5430	if (!READ_ONCE(vcpu->kvm->arch.indirect_shadow_pages))
5431		return;
5432
5433	remote_flush = local_flush = false;
5434
5435	pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
5436
5437	/*
5438	 * No need to care whether allocation memory is successful
5439	 * or not since pte prefetch is skiped if it does not have
5440	 * enough objects in the cache.
5441	 */
5442	mmu_topup_memory_caches(vcpu);
5443
5444	spin_lock(&vcpu->kvm->mmu_lock);
5445
5446	gentry = mmu_pte_write_fetch_gpte(vcpu, &gpa, &bytes);
5447
5448	++vcpu->kvm->stat.mmu_pte_write;
5449	kvm_mmu_audit(vcpu, AUDIT_PRE_PTE_WRITE);
5450
5451	for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn) {
5452		if (detect_write_misaligned(sp, gpa, bytes) ||
5453		      detect_write_flooding(sp)) {
5454			kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
5455			++vcpu->kvm->stat.mmu_flooded;
5456			continue;
5457		}
5458
5459		spte = get_written_sptes(sp, gpa, &npte);
5460		if (!spte)
5461			continue;
5462
5463		local_flush = true;
5464		while (npte--) {
5465			u32 base_role = vcpu->arch.mmu->mmu_role.base.word;
5466
5467			entry = *spte;
5468			mmu_page_zap_pte(vcpu->kvm, sp, spte);
5469			if (gentry &&
5470			      !((sp->role.word ^ base_role)
5471			      & mmu_base_role_mask.word) && rmap_can_add(vcpu))
5472				mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
5473			if (need_remote_flush(entry, *spte))
5474				remote_flush = true;
5475			++spte;
5476		}
5477	}
5478	kvm_mmu_flush_or_zap(vcpu, &invalid_list, remote_flush, local_flush);
5479	kvm_mmu_audit(vcpu, AUDIT_POST_PTE_WRITE);
5480	spin_unlock(&vcpu->kvm->mmu_lock);
5481}
5482
5483int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
5484{
5485	gpa_t gpa;
5486	int r;
5487
5488	if (vcpu->arch.mmu->direct_map)
5489		return 0;
5490
5491	gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
5492
5493	r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
5494
5495	return r;
5496}
5497EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
5498
5499static int make_mmu_pages_available(struct kvm_vcpu *vcpu)
5500{
5501	LIST_HEAD(invalid_list);
5502
5503	if (likely(kvm_mmu_available_pages(vcpu->kvm) >= KVM_MIN_FREE_MMU_PAGES))
5504		return 0;
5505
5506	while (kvm_mmu_available_pages(vcpu->kvm) < KVM_REFILL_PAGES) {
5507		if (!prepare_zap_oldest_mmu_page(vcpu->kvm, &invalid_list))
5508			break;
5509
5510		++vcpu->kvm->stat.mmu_recycled;
5511	}
5512	kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
5513
5514	if (!kvm_mmu_available_pages(vcpu->kvm))
5515		return -ENOSPC;
5516	return 0;
5517}
5518
5519int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u64 error_code,
5520		       void *insn, int insn_len)
5521{
5522	int r, emulation_type = 0;
5523	bool direct = vcpu->arch.mmu->direct_map;
5524
5525	/* With shadow page tables, fault_address contains a GVA or nGPA.  */
5526	if (vcpu->arch.mmu->direct_map) {
5527		vcpu->arch.gpa_available = true;
5528		vcpu->arch.gpa_val = cr2;
5529	}
5530
5531	r = RET_PF_INVALID;
5532	if (unlikely(error_code & PFERR_RSVD_MASK)) {
5533		r = handle_mmio_page_fault(vcpu, cr2, direct);
5534		if (r == RET_PF_EMULATE)
5535			goto emulate;
5536	}
5537
5538	if (r == RET_PF_INVALID) {
5539		r = vcpu->arch.mmu->page_fault(vcpu, cr2,
5540					       lower_32_bits(error_code),
5541					       false);
5542		WARN_ON(r == RET_PF_INVALID);
5543	}
5544
5545	if (r == RET_PF_RETRY)
5546		return 1;
5547	if (r < 0)
5548		return r;
5549
5550	/*
5551	 * Before emulating the instruction, check if the error code
5552	 * was due to a RO violation while translating the guest page.
5553	 * This can occur when using nested virtualization with nested
5554	 * paging in both guests. If true, we simply unprotect the page
5555	 * and resume the guest.
5556	 */
5557	if (vcpu->arch.mmu->direct_map &&
5558	    (error_code & PFERR_NESTED_GUEST_PAGE) == PFERR_NESTED_GUEST_PAGE) {
5559		kvm_mmu_unprotect_page(vcpu->kvm, gpa_to_gfn(cr2));
5560		return 1;
5561	}
5562
5563	/*
5564	 * vcpu->arch.mmu.page_fault returned RET_PF_EMULATE, but we can still
5565	 * optimistically try to just unprotect the page and let the processor
5566	 * re-execute the instruction that caused the page fault.  Do not allow
5567	 * retrying MMIO emulation, as it's not only pointless but could also
5568	 * cause us to enter an infinite loop because the processor will keep
5569	 * faulting on the non-existent MMIO address.  Retrying an instruction
5570	 * from a nested guest is also pointless and dangerous as we are only
5571	 * explicitly shadowing L1's page tables, i.e. unprotecting something
5572	 * for L1 isn't going to magically fix whatever issue cause L2 to fail.
5573	 */
5574	if (!mmio_info_in_cache(vcpu, cr2, direct) && !is_guest_mode(vcpu))
5575		emulation_type = EMULTYPE_ALLOW_RETRY;
5576emulate:
5577	/*
5578	 * On AMD platforms, under certain conditions insn_len may be zero on #NPF.
5579	 * This can happen if a guest gets a page-fault on data access but the HW
5580	 * table walker is not able to read the instruction page (e.g instruction
5581	 * page is not present in memory). In those cases we simply restart the
5582	 * guest, with the exception of AMD Erratum 1096 which is unrecoverable.
5583	 */
5584	if (unlikely(insn && !insn_len)) {
5585		if (!kvm_x86_ops->need_emulation_on_page_fault(vcpu))
5586			return 1;
5587	}
5588
5589	return x86_emulate_instruction(vcpu, cr2, emulation_type, insn,
5590				       insn_len);
5591}
5592EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
5593
5594void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
5595{
5596	struct kvm_mmu *mmu = vcpu->arch.mmu;
5597	int i;
5598
5599	/* INVLPG on a * non-canonical address is a NOP according to the SDM.  */
5600	if (is_noncanonical_address(gva, vcpu))
5601		return;
5602
5603	mmu->invlpg(vcpu, gva, mmu->root_hpa);
5604
5605	/*
5606	 * INVLPG is required to invalidate any global mappings for the VA,
5607	 * irrespective of PCID. Since it would take us roughly similar amount
5608	 * of work to determine whether any of the prev_root mappings of the VA
5609	 * is marked global, or to just sync it blindly, so we might as well
5610	 * just always sync it.
5611	 *
5612	 * Mappings not reachable via the current cr3 or the prev_roots will be
5613	 * synced when switching to that cr3, so nothing needs to be done here
5614	 * for them.
5615	 */
5616	for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
5617		if (VALID_PAGE(mmu->prev_roots[i].hpa))
5618			mmu->invlpg(vcpu, gva, mmu->prev_roots[i].hpa);
5619
5620	kvm_x86_ops->tlb_flush_gva(vcpu, gva);
5621	++vcpu->stat.invlpg;
5622}
5623EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
5624
5625void kvm_mmu_invpcid_gva(struct kvm_vcpu *vcpu, gva_t gva, unsigned long pcid)
5626{
5627	struct kvm_mmu *mmu = vcpu->arch.mmu;
5628	bool tlb_flush = false;
5629	uint i;
5630
5631	if (pcid == kvm_get_active_pcid(vcpu)) {
5632		mmu->invlpg(vcpu, gva, mmu->root_hpa);
5633		tlb_flush = true;
5634	}
5635
5636	for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
5637		if (VALID_PAGE(mmu->prev_roots[i].hpa) &&
5638		    pcid == kvm_get_pcid(vcpu, mmu->prev_roots[i].cr3)) {
5639			mmu->invlpg(vcpu, gva, mmu->prev_roots[i].hpa);
5640			tlb_flush = true;
5641		}
5642	}
5643
5644	if (tlb_flush)
5645		kvm_x86_ops->tlb_flush_gva(vcpu, gva);
5646
5647	++vcpu->stat.invlpg;
5648
5649	/*
5650	 * Mappings not reachable via the current cr3 or the prev_roots will be
5651	 * synced when switching to that cr3, so nothing needs to be done here
5652	 * for them.
5653	 */
5654}
5655EXPORT_SYMBOL_GPL(kvm_mmu_invpcid_gva);
5656
5657void kvm_enable_tdp(void)
5658{
5659	tdp_enabled = true;
5660}
5661EXPORT_SYMBOL_GPL(kvm_enable_tdp);
5662
5663void kvm_disable_tdp(void)
5664{
5665	tdp_enabled = false;
5666}
5667EXPORT_SYMBOL_GPL(kvm_disable_tdp);
5668
5669
5670/* The return value indicates if tlb flush on all vcpus is needed. */
5671typedef bool (*slot_level_handler) (struct kvm *kvm, struct kvm_rmap_head *rmap_head);
5672
5673/* The caller should hold mmu-lock before calling this function. */
5674static __always_inline bool
5675slot_handle_level_range(struct kvm *kvm, struct kvm_memory_slot *memslot,
5676			slot_level_handler fn, int start_level, int end_level,
5677			gfn_t start_gfn, gfn_t end_gfn, bool lock_flush_tlb)
5678{
5679	struct slot_rmap_walk_iterator iterator;
5680	bool flush = false;
5681
5682	for_each_slot_rmap_range(memslot, start_level, end_level, start_gfn,
5683			end_gfn, &iterator) {
5684		if (iterator.rmap)
5685			flush |= fn(kvm, iterator.rmap);
5686
5687		if (need_resched() || spin_needbreak(&kvm->mmu_lock)) {
5688			if (flush && lock_flush_tlb) {
5689				kvm_flush_remote_tlbs_with_address(kvm,
5690						start_gfn,
5691						iterator.gfn - start_gfn + 1);
5692				flush = false;
5693			}
5694			cond_resched_lock(&kvm->mmu_lock);
5695		}
5696	}
5697
5698	if (flush && lock_flush_tlb) {
5699		kvm_flush_remote_tlbs_with_address(kvm, start_gfn,
5700						   end_gfn - start_gfn + 1);
5701		flush = false;
5702	}
5703
5704	return flush;
5705}
5706
5707static __always_inline bool
5708slot_handle_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
5709		  slot_level_handler fn, int start_level, int end_level,
5710		  bool lock_flush_tlb)
5711{
5712	return slot_handle_level_range(kvm, memslot, fn, start_level,
5713			end_level, memslot->base_gfn,
5714			memslot->base_gfn + memslot->npages - 1,
5715			lock_flush_tlb);
5716}
5717
5718static __always_inline bool
5719slot_handle_all_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
5720		      slot_level_handler fn, bool lock_flush_tlb)
5721{
5722	return slot_handle_level(kvm, memslot, fn, PT_PAGE_TABLE_LEVEL,
5723				 PT_MAX_HUGEPAGE_LEVEL, lock_flush_tlb);
5724}
5725
5726static __always_inline bool
5727slot_handle_large_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
5728			slot_level_handler fn, bool lock_flush_tlb)
5729{
5730	return slot_handle_level(kvm, memslot, fn, PT_PAGE_TABLE_LEVEL + 1,
5731				 PT_MAX_HUGEPAGE_LEVEL, lock_flush_tlb);
5732}
5733
5734static __always_inline bool
5735slot_handle_leaf(struct kvm *kvm, struct kvm_memory_slot *memslot,
5736		 slot_level_handler fn, bool lock_flush_tlb)
5737{
5738	return slot_handle_level(kvm, memslot, fn, PT_PAGE_TABLE_LEVEL,
5739				 PT_PAGE_TABLE_LEVEL, lock_flush_tlb);
5740}
5741
5742static void free_mmu_pages(struct kvm_mmu *mmu)
5743{
5744	free_page((unsigned long)mmu->pae_root);
5745	free_page((unsigned long)mmu->lm_root);
5746}
5747
5748static int alloc_mmu_pages(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu)
5749{
5750	struct page *page;
5751	int i;
5752
5753	/*
5754	 * When using PAE paging, the four PDPTEs are treated as 'root' pages,
5755	 * while the PDP table is a per-vCPU construct that's allocated at MMU
5756	 * creation.  When emulating 32-bit mode, cr3 is only 32 bits even on
5757	 * x86_64.  Therefore we need to allocate the PDP table in the first
5758	 * 4GB of memory, which happens to fit the DMA32 zone.  Except for
5759	 * SVM's 32-bit NPT support, TDP paging doesn't use PAE paging and can
5760	 * skip allocating the PDP table.
5761	 */
5762	if (tdp_enabled && kvm_x86_ops->get_tdp_level(vcpu) > PT32E_ROOT_LEVEL)
5763		return 0;
5764
5765	page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_DMA32);
5766	if (!page)
5767		return -ENOMEM;
5768
5769	mmu->pae_root = page_address(page);
5770	for (i = 0; i < 4; ++i)
5771		mmu->pae_root[i] = INVALID_PAGE;
5772
5773	return 0;
5774}
5775
5776int kvm_mmu_create(struct kvm_vcpu *vcpu)
5777{
5778	uint i;
5779	int ret;
5780
5781	vcpu->arch.mmu = &vcpu->arch.root_mmu;
5782	vcpu->arch.walk_mmu = &vcpu->arch.root_mmu;
5783
5784	vcpu->arch.root_mmu.root_hpa = INVALID_PAGE;
5785	vcpu->arch.root_mmu.root_cr3 = 0;
5786	vcpu->arch.root_mmu.translate_gpa = translate_gpa;
5787	for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
5788		vcpu->arch.root_mmu.prev_roots[i] = KVM_MMU_ROOT_INFO_INVALID;
5789
5790	vcpu->arch.guest_mmu.root_hpa = INVALID_PAGE;
5791	vcpu->arch.guest_mmu.root_cr3 = 0;
5792	vcpu->arch.guest_mmu.translate_gpa = translate_gpa;
5793	for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
5794		vcpu->arch.guest_mmu.prev_roots[i] = KVM_MMU_ROOT_INFO_INVALID;
5795
5796	vcpu->arch.nested_mmu.translate_gpa = translate_nested_gpa;
5797
5798	ret = alloc_mmu_pages(vcpu, &vcpu->arch.guest_mmu);
5799	if (ret)
5800		return ret;
5801
5802	ret = alloc_mmu_pages(vcpu, &vcpu->arch.root_mmu);
5803	if (ret)
5804		goto fail_allocate_root;
5805
5806	return ret;
5807 fail_allocate_root:
5808	free_mmu_pages(&vcpu->arch.guest_mmu);
5809	return ret;
5810}
5811
5812#define BATCH_ZAP_PAGES	10
5813static void kvm_zap_obsolete_pages(struct kvm *kvm)
5814{
5815	struct kvm_mmu_page *sp, *node;
5816	int nr_zapped, batch = 0;
5817
5818restart:
5819	list_for_each_entry_safe_reverse(sp, node,
5820	      &kvm->arch.active_mmu_pages, link) {
5821		/*
5822		 * No obsolete valid page exists before a newly created page
5823		 * since active_mmu_pages is a FIFO list.
5824		 */
5825		if (!is_obsolete_sp(kvm, sp))
5826			break;
5827
5828		/*
5829		 * Skip invalid pages with a non-zero root count, zapping pages
5830		 * with a non-zero root count will never succeed, i.e. the page
5831		 * will get thrown back on active_mmu_pages and we'll get stuck
5832		 * in an infinite loop.
5833		 */
5834		if (sp->role.invalid && sp->root_count)
5835			continue;
5836
5837		/*
5838		 * No need to flush the TLB since we're only zapping shadow
5839		 * pages with an obsolete generation number and all vCPUS have
5840		 * loaded a new root, i.e. the shadow pages being zapped cannot
5841		 * be in active use by the guest.
5842		 */
5843		if (batch >= BATCH_ZAP_PAGES &&
5844		    cond_resched_lock(&kvm->mmu_lock)) {
5845			batch = 0;
5846			goto restart;
5847		}
5848
5849		if (__kvm_mmu_prepare_zap_page(kvm, sp,
5850				&kvm->arch.zapped_obsolete_pages, &nr_zapped)) {
5851			batch += nr_zapped;
5852			goto restart;
5853		}
5854	}
5855
5856	/*
5857	 * Trigger a remote TLB flush before freeing the page tables to ensure
5858	 * KVM is not in the middle of a lockless shadow page table walk, which
5859	 * may reference the pages.
5860	 */
5861	kvm_mmu_commit_zap_page(kvm, &kvm->arch.zapped_obsolete_pages);
5862}
5863
5864/*
5865 * Fast invalidate all shadow pages and use lock-break technique
5866 * to zap obsolete pages.
5867 *
5868 * It's required when memslot is being deleted or VM is being
5869 * destroyed, in these cases, we should ensure that KVM MMU does
5870 * not use any resource of the being-deleted slot or all slots
5871 * after calling the function.
5872 */
5873static void kvm_mmu_zap_all_fast(struct kvm *kvm)
5874{
5875	lockdep_assert_held(&kvm->slots_lock);
5876
5877	spin_lock(&kvm->mmu_lock);
5878	trace_kvm_mmu_zap_all_fast(kvm);
5879
5880	/*
5881	 * Toggle mmu_valid_gen between '0' and '1'.  Because slots_lock is
5882	 * held for the entire duration of zapping obsolete pages, it's
5883	 * impossible for there to be multiple invalid generations associated
5884	 * with *valid* shadow pages at any given time, i.e. there is exactly
5885	 * one valid generation and (at most) one invalid generation.
5886	 */
5887	kvm->arch.mmu_valid_gen = kvm->arch.mmu_valid_gen ? 0 : 1;
5888
5889	/*
5890	 * Notify all vcpus to reload its shadow page table and flush TLB.
5891	 * Then all vcpus will switch to new shadow page table with the new
5892	 * mmu_valid_gen.
5893	 *
5894	 * Note: we need to do this under the protection of mmu_lock,
5895	 * otherwise, vcpu would purge shadow page but miss tlb flush.
5896	 */
5897	kvm_reload_remote_mmus(kvm);
5898
5899	kvm_zap_obsolete_pages(kvm);
5900	spin_unlock(&kvm->mmu_lock);
5901}
5902
5903static bool kvm_has_zapped_obsolete_pages(struct kvm *kvm)
5904{
5905	return unlikely(!list_empty_careful(&kvm->arch.zapped_obsolete_pages));
5906}
5907
5908static void kvm_mmu_invalidate_zap_pages_in_memslot(struct kvm *kvm,
5909			struct kvm_memory_slot *slot,
5910			struct kvm_page_track_notifier_node *node)
5911{
5912	kvm_mmu_zap_all_fast(kvm);
5913}
5914
5915void kvm_mmu_init_vm(struct kvm *kvm)
5916{
5917	struct kvm_page_track_notifier_node *node = &kvm->arch.mmu_sp_tracker;
5918
5919	node->track_write = kvm_mmu_pte_write;
5920	node->track_flush_slot = kvm_mmu_invalidate_zap_pages_in_memslot;
5921	kvm_page_track_register_notifier(kvm, node);
5922}
5923
5924void kvm_mmu_uninit_vm(struct kvm *kvm)
5925{
5926	struct kvm_page_track_notifier_node *node = &kvm->arch.mmu_sp_tracker;
5927
5928	kvm_page_track_unregister_notifier(kvm, node);
5929}
5930
5931void kvm_zap_gfn_range(struct kvm *kvm, gfn_t gfn_start, gfn_t gfn_end)
5932{
5933	struct kvm_memslots *slots;
5934	struct kvm_memory_slot *memslot;
5935	int i;
5936
5937	spin_lock(&kvm->mmu_lock);
5938	for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
5939		slots = __kvm_memslots(kvm, i);
5940		kvm_for_each_memslot(memslot, slots) {
5941			gfn_t start, end;
5942
5943			start = max(gfn_start, memslot->base_gfn);
5944			end = min(gfn_end, memslot->base_gfn + memslot->npages);
5945			if (start >= end)
5946				continue;
5947
5948			slot_handle_level_range(kvm, memslot, kvm_zap_rmapp,
5949						PT_PAGE_TABLE_LEVEL, PT_MAX_HUGEPAGE_LEVEL,
5950						start, end - 1, true);
5951		}
5952	}
5953
5954	spin_unlock(&kvm->mmu_lock);
5955}
5956
5957static bool slot_rmap_write_protect(struct kvm *kvm,
5958				    struct kvm_rmap_head *rmap_head)
5959{
5960	return __rmap_write_protect(kvm, rmap_head, false);
5961}
5962
5963void kvm_mmu_slot_remove_write_access(struct kvm *kvm,
5964				      struct kvm_memory_slot *memslot)
5965{
5966	bool flush;
5967
5968	spin_lock(&kvm->mmu_lock);
5969	flush = slot_handle_all_level(kvm, memslot, slot_rmap_write_protect,
5970				      false);
5971	spin_unlock(&kvm->mmu_lock);
5972
5973	/*
5974	 * kvm_mmu_slot_remove_write_access() and kvm_vm_ioctl_get_dirty_log()
5975	 * which do tlb flush out of mmu-lock should be serialized by
5976	 * kvm->slots_lock otherwise tlb flush would be missed.
5977	 */
5978	lockdep_assert_held(&kvm->slots_lock);
5979
5980	/*
5981	 * We can flush all the TLBs out of the mmu lock without TLB
5982	 * corruption since we just change the spte from writable to
5983	 * readonly so that we only need to care the case of changing
5984	 * spte from present to present (changing the spte from present
5985	 * to nonpresent will flush all the TLBs immediately), in other
5986	 * words, the only case we care is mmu_spte_update() where we
5987	 * have checked SPTE_HOST_WRITEABLE | SPTE_MMU_WRITEABLE
5988	 * instead of PT_WRITABLE_MASK, that means it does not depend
5989	 * on PT_WRITABLE_MASK anymore.
5990	 */
5991	if (flush)
5992		kvm_flush_remote_tlbs_with_address(kvm, memslot->base_gfn,
5993			memslot->npages);
5994}
5995
5996static bool kvm_mmu_zap_collapsible_spte(struct kvm *kvm,
5997					 struct kvm_rmap_head *rmap_head)
5998{
5999	u64 *sptep;
6000	struct rmap_iterator iter;
6001	int need_tlb_flush = 0;
6002	kvm_pfn_t pfn;
6003	struct kvm_mmu_page *sp;
6004
6005restart:
6006	for_each_rmap_spte(rmap_head, &iter, sptep) {
6007		sp = page_header(__pa(sptep));
6008		pfn = spte_to_pfn(*sptep);
6009
6010		/*
6011		 * We cannot do huge page mapping for indirect shadow pages,
6012		 * which are found on the last rmap (level = 1) when not using
6013		 * tdp; such shadow pages are synced with the page table in
6014		 * the guest, and the guest page table is using 4K page size
6015		 * mapping if the indirect sp has level = 1.
6016		 */
6017		if (sp->role.direct && !kvm_is_reserved_pfn(pfn) &&
6018		    !kvm_is_zone_device_pfn(pfn) &&
6019		    PageTransCompoundMap(pfn_to_page(pfn))) {
6020			pte_list_remove(rmap_head, sptep);
6021
6022			if (kvm_available_flush_tlb_with_range())
6023				kvm_flush_remote_tlbs_with_address(kvm, sp->gfn,
6024					KVM_PAGES_PER_HPAGE(sp->role.level));
6025			else
6026				need_tlb_flush = 1;
6027
6028			goto restart;
6029		}
6030	}
6031
6032	return need_tlb_flush;
6033}
6034
6035void kvm_mmu_zap_collapsible_sptes(struct kvm *kvm,
6036				   const struct kvm_memory_slot *memslot)
6037{
6038	/* FIXME: const-ify all uses of struct kvm_memory_slot.  */
6039	spin_lock(&kvm->mmu_lock);
6040	slot_handle_leaf(kvm, (struct kvm_memory_slot *)memslot,
6041			 kvm_mmu_zap_collapsible_spte, true);
6042	spin_unlock(&kvm->mmu_lock);
6043}
6044
6045void kvm_mmu_slot_leaf_clear_dirty(struct kvm *kvm,
6046				   struct kvm_memory_slot *memslot)
6047{
6048	bool flush;
6049
6050	spin_lock(&kvm->mmu_lock);
6051	flush = slot_handle_leaf(kvm, memslot, __rmap_clear_dirty, false);
6052	spin_unlock(&kvm->mmu_lock);
6053
6054	lockdep_assert_held(&kvm->slots_lock);
6055
6056	/*
6057	 * It's also safe to flush TLBs out of mmu lock here as currently this
6058	 * function is only used for dirty logging, in which case flushing TLB
6059	 * out of mmu lock also guarantees no dirty pages will be lost in
6060	 * dirty_bitmap.
6061	 */
6062	if (flush)
6063		kvm_flush_remote_tlbs_with_address(kvm, memslot->base_gfn,
6064				memslot->npages);
6065}
6066EXPORT_SYMBOL_GPL(kvm_mmu_slot_leaf_clear_dirty);
6067
6068void kvm_mmu_slot_largepage_remove_write_access(struct kvm *kvm,
6069					struct kvm_memory_slot *memslot)
6070{
6071	bool flush;
6072
6073	spin_lock(&kvm->mmu_lock);
6074	flush = slot_handle_large_level(kvm, memslot, slot_rmap_write_protect,
6075					false);
6076	spin_unlock(&kvm->mmu_lock);
6077
6078	/* see kvm_mmu_slot_remove_write_access */
6079	lockdep_assert_held(&kvm->slots_lock);
6080
6081	if (flush)
6082		kvm_flush_remote_tlbs_with_address(kvm, memslot->base_gfn,
6083				memslot->npages);
6084}
6085EXPORT_SYMBOL_GPL(kvm_mmu_slot_largepage_remove_write_access);
6086
6087void kvm_mmu_slot_set_dirty(struct kvm *kvm,
6088			    struct kvm_memory_slot *memslot)
6089{
6090	bool flush;
6091
6092	spin_lock(&kvm->mmu_lock);
6093	flush = slot_handle_all_level(kvm, memslot, __rmap_set_dirty, false);
6094	spin_unlock(&kvm->mmu_lock);
6095
6096	lockdep_assert_held(&kvm->slots_lock);
6097
6098	/* see kvm_mmu_slot_leaf_clear_dirty */
6099	if (flush)
6100		kvm_flush_remote_tlbs_with_address(kvm, memslot->base_gfn,
6101				memslot->npages);
6102}
6103EXPORT_SYMBOL_GPL(kvm_mmu_slot_set_dirty);
6104
6105void kvm_mmu_zap_all(struct kvm *kvm)
6106{
6107	struct kvm_mmu_page *sp, *node;
6108	LIST_HEAD(invalid_list);
6109	int ign;
6110
6111	spin_lock(&kvm->mmu_lock);
6112restart:
6113	list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link) {
6114		if (sp->role.invalid && sp->root_count)
6115			continue;
6116		if (__kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list, &ign))
6117			goto restart;
6118		if (cond_resched_lock(&kvm->mmu_lock))
6119			goto restart;
6120	}
6121
6122	kvm_mmu_commit_zap_page(kvm, &invalid_list);
6123	spin_unlock(&kvm->mmu_lock);
6124}
6125
6126void kvm_mmu_invalidate_mmio_sptes(struct kvm *kvm, u64 gen)
6127{
6128	WARN_ON(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS);
6129
6130	gen &= MMIO_SPTE_GEN_MASK;
6131
6132	/*
6133	 * Generation numbers are incremented in multiples of the number of
6134	 * address spaces in order to provide unique generations across all
6135	 * address spaces.  Strip what is effectively the address space
6136	 * modifier prior to checking for a wrap of the MMIO generation so
6137	 * that a wrap in any address space is detected.
6138	 */
6139	gen &= ~((u64)KVM_ADDRESS_SPACE_NUM - 1);
6140
6141	/*
6142	 * The very rare case: if the MMIO generation number has wrapped,
6143	 * zap all shadow pages.
6144	 */
6145	if (unlikely(gen == 0)) {
6146		kvm_debug_ratelimited("kvm: zapping shadow pages for mmio generation wraparound\n");
6147		kvm_mmu_zap_all_fast(kvm);
6148	}
6149}
6150
6151static unsigned long
6152mmu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
6153{
6154	struct kvm *kvm;
6155	int nr_to_scan = sc->nr_to_scan;
6156	unsigned long freed = 0;
6157
6158	mutex_lock(&kvm_lock);
6159
6160	list_for_each_entry(kvm, &vm_list, vm_list) {
6161		int idx;
6162		LIST_HEAD(invalid_list);
6163
6164		/*
6165		 * Never scan more than sc->nr_to_scan VM instances.
6166		 * Will not hit this condition practically since we do not try
6167		 * to shrink more than one VM and it is very unlikely to see
6168		 * !n_used_mmu_pages so many times.
6169		 */
6170		if (!nr_to_scan--)
6171			break;
6172		/*
6173		 * n_used_mmu_pages is accessed without holding kvm->mmu_lock
6174		 * here. We may skip a VM instance errorneosly, but we do not
6175		 * want to shrink a VM that only started to populate its MMU
6176		 * anyway.
6177		 */
6178		if (!kvm->arch.n_used_mmu_pages &&
6179		    !kvm_has_zapped_obsolete_pages(kvm))
6180			continue;
6181
6182		idx = srcu_read_lock(&kvm->srcu);
6183		spin_lock(&kvm->mmu_lock);
6184
6185		if (kvm_has_zapped_obsolete_pages(kvm)) {
6186			kvm_mmu_commit_zap_page(kvm,
6187			      &kvm->arch.zapped_obsolete_pages);
6188			goto unlock;
6189		}
6190
6191		if (prepare_zap_oldest_mmu_page(kvm, &invalid_list))
6192			freed++;
6193		kvm_mmu_commit_zap_page(kvm, &invalid_list);
6194
6195unlock:
6196		spin_unlock(&kvm->mmu_lock);
6197		srcu_read_unlock(&kvm->srcu, idx);
6198
6199		/*
6200		 * unfair on small ones
6201		 * per-vm shrinkers cry out
6202		 * sadness comes quickly
6203		 */
6204		list_move_tail(&kvm->vm_list, &vm_list);
6205		break;
6206	}
6207
6208	mutex_unlock(&kvm_lock);
6209	return freed;
6210}
6211
6212static unsigned long
6213mmu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
6214{
6215	return percpu_counter_read_positive(&kvm_total_used_mmu_pages);
6216}
6217
6218static struct shrinker mmu_shrinker = {
6219	.count_objects = mmu_shrink_count,
6220	.scan_objects = mmu_shrink_scan,
6221	.seeks = DEFAULT_SEEKS * 10,
6222};
6223
6224static void mmu_destroy_caches(void)
6225{
6226	kmem_cache_destroy(pte_list_desc_cache);
6227	kmem_cache_destroy(mmu_page_header_cache);
6228}
6229
6230static void kvm_set_mmio_spte_mask(void)
6231{
6232	u64 mask;
6233
6234	/*
6235	 * Set the reserved bits and the present bit of an paging-structure
6236	 * entry to generate page fault with PFER.RSV = 1.
6237	 */
6238
6239	/*
6240	 * Mask the uppermost physical address bit, which would be reserved as
6241	 * long as the supported physical address width is less than 52.
6242	 */
6243	mask = 1ull << 51;
6244
6245	/* Set the present bit. */
6246	mask |= 1ull;
6247
6248	/*
6249	 * If reserved bit is not supported, clear the present bit to disable
6250	 * mmio page fault.
6251	 */
6252	if (IS_ENABLED(CONFIG_X86_64) && shadow_phys_bits == 52)
6253		mask &= ~1ull;
6254
6255	kvm_mmu_set_mmio_spte_mask(mask, mask, ACC_WRITE_MASK | ACC_USER_MASK);
6256}
6257
6258static bool get_nx_auto_mode(void)
6259{
6260	/* Return true when CPU has the bug, and mitigations are ON */
6261	return boot_cpu_has_bug(X86_BUG_ITLB_MULTIHIT) && !cpu_mitigations_off();
6262}
6263
6264static void __set_nx_huge_pages(bool val)
6265{
6266	nx_huge_pages = itlb_multihit_kvm_mitigation = val;
6267}
6268
6269static int set_nx_huge_pages(const char *val, const struct kernel_param *kp)
6270{
6271	bool old_val = nx_huge_pages;
6272	bool new_val;
6273
6274	/* In "auto" mode deploy workaround only if CPU has the bug. */
6275	if (sysfs_streq(val, "off"))
6276		new_val = 0;
6277	else if (sysfs_streq(val, "force"))
6278		new_val = 1;
6279	else if (sysfs_streq(val, "auto"))
6280		new_val = get_nx_auto_mode();
6281	else if (strtobool(val, &new_val) < 0)
6282		return -EINVAL;
6283
6284	__set_nx_huge_pages(new_val);
6285
6286	if (new_val != old_val) {
6287		struct kvm *kvm;
6288
6289		mutex_lock(&kvm_lock);
6290
6291		list_for_each_entry(kvm, &vm_list, vm_list) {
6292			mutex_lock(&kvm->slots_lock);
6293			kvm_mmu_zap_all_fast(kvm);
6294			mutex_unlock(&kvm->slots_lock);
6295
6296			wake_up_process(kvm->arch.nx_lpage_recovery_thread);
6297		}
6298		mutex_unlock(&kvm_lock);
6299	}
6300
6301	return 0;
6302}
6303
6304int kvm_mmu_module_init(void)
6305{
6306	int ret = -ENOMEM;
6307
6308	if (nx_huge_pages == -1)
6309		__set_nx_huge_pages(get_nx_auto_mode());
6310
6311	/*
6312	 * MMU roles use union aliasing which is, generally speaking, an
6313	 * undefined behavior. However, we supposedly know how compilers behave
6314	 * and the current status quo is unlikely to change. Guardians below are
6315	 * supposed to let us know if the assumption becomes false.
6316	 */
6317	BUILD_BUG_ON(sizeof(union kvm_mmu_page_role) != sizeof(u32));
6318	BUILD_BUG_ON(sizeof(union kvm_mmu_extended_role) != sizeof(u32));
6319	BUILD_BUG_ON(sizeof(union kvm_mmu_role) != sizeof(u64));
6320
6321	kvm_mmu_reset_all_pte_masks();
6322
6323	kvm_set_mmio_spte_mask();
6324
6325	pte_list_desc_cache = kmem_cache_create("pte_list_desc",
6326					    sizeof(struct pte_list_desc),
6327					    0, SLAB_ACCOUNT, NULL);
6328	if (!pte_list_desc_cache)
6329		goto out;
6330
6331	mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
6332						  sizeof(struct kvm_mmu_page),
6333						  0, SLAB_ACCOUNT, NULL);
6334	if (!mmu_page_header_cache)
6335		goto out;
6336
6337	if (percpu_counter_init(&kvm_total_used_mmu_pages, 0, GFP_KERNEL))
6338		goto out;
6339
6340	ret = register_shrinker(&mmu_shrinker);
6341	if (ret)
6342		goto out;
6343
6344	return 0;
6345
6346out:
6347	mmu_destroy_caches();
6348	return ret;
6349}
6350
6351/*
6352 * Calculate mmu pages needed for kvm.
6353 */
6354unsigned long kvm_mmu_calculate_default_mmu_pages(struct kvm *kvm)
6355{
6356	unsigned long nr_mmu_pages;
6357	unsigned long nr_pages = 0;
6358	struct kvm_memslots *slots;
6359	struct kvm_memory_slot *memslot;
6360	int i;
6361
6362	for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
6363		slots = __kvm_memslots(kvm, i);
6364
6365		kvm_for_each_memslot(memslot, slots)
6366			nr_pages += memslot->npages;
6367	}
6368
6369	nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
6370	nr_mmu_pages = max(nr_mmu_pages, KVM_MIN_ALLOC_MMU_PAGES);
6371
6372	return nr_mmu_pages;
6373}
6374
6375void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
6376{
6377	kvm_mmu_unload(vcpu);
6378	free_mmu_pages(&vcpu->arch.root_mmu);
6379	free_mmu_pages(&vcpu->arch.guest_mmu);
6380	mmu_free_memory_caches(vcpu);
6381}
6382
6383void kvm_mmu_module_exit(void)
6384{
6385	mmu_destroy_caches();
6386	percpu_counter_destroy(&kvm_total_used_mmu_pages);
6387	unregister_shrinker(&mmu_shrinker);
6388	mmu_audit_disable();
6389}
6390
6391static int set_nx_huge_pages_recovery_ratio(const char *val, const struct kernel_param *kp)
6392{
6393	unsigned int old_val;
6394	int err;
6395
6396	old_val = nx_huge_pages_recovery_ratio;
6397	err = param_set_uint(val, kp);
6398	if (err)
6399		return err;
6400
6401	if (READ_ONCE(nx_huge_pages) &&
6402	    !old_val && nx_huge_pages_recovery_ratio) {
6403		struct kvm *kvm;
6404
6405		mutex_lock(&kvm_lock);
6406
6407		list_for_each_entry(kvm, &vm_list, vm_list)
6408			wake_up_process(kvm->arch.nx_lpage_recovery_thread);
6409
6410		mutex_unlock(&kvm_lock);
6411	}
6412
6413	return err;
6414}
6415
6416static void kvm_recover_nx_lpages(struct kvm *kvm)
6417{
6418	int rcu_idx;
6419	struct kvm_mmu_page *sp;
6420	unsigned int ratio;
6421	LIST_HEAD(invalid_list);
6422	ulong to_zap;
6423
6424	rcu_idx = srcu_read_lock(&kvm->srcu);
6425	spin_lock(&kvm->mmu_lock);
6426
6427	ratio = READ_ONCE(nx_huge_pages_recovery_ratio);
6428	to_zap = ratio ? DIV_ROUND_UP(kvm->stat.nx_lpage_splits, ratio) : 0;
6429	while (to_zap && !list_empty(&kvm->arch.lpage_disallowed_mmu_pages)) {
6430		/*
6431		 * We use a separate list instead of just using active_mmu_pages
6432		 * because the number of lpage_disallowed pages is expected to
6433		 * be relatively small compared to the total.
6434		 */
6435		sp = list_first_entry(&kvm->arch.lpage_disallowed_mmu_pages,
6436				      struct kvm_mmu_page,
6437				      lpage_disallowed_link);
6438		WARN_ON_ONCE(!sp->lpage_disallowed);
6439		kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
6440		WARN_ON_ONCE(sp->lpage_disallowed);
6441
6442		if (!--to_zap || need_resched() || spin_needbreak(&kvm->mmu_lock)) {
6443			kvm_mmu_commit_zap_page(kvm, &invalid_list);
6444			if (to_zap)
6445				cond_resched_lock(&kvm->mmu_lock);
6446		}
6447	}
6448
6449	spin_unlock(&kvm->mmu_lock);
6450	srcu_read_unlock(&kvm->srcu, rcu_idx);
6451}
6452
6453static long get_nx_lpage_recovery_timeout(u64 start_time)
6454{
6455	return READ_ONCE(nx_huge_pages) && READ_ONCE(nx_huge_pages_recovery_ratio)
6456		? start_time + 60 * HZ - get_jiffies_64()
6457		: MAX_SCHEDULE_TIMEOUT;
6458}
6459
6460static int kvm_nx_lpage_recovery_worker(struct kvm *kvm, uintptr_t data)
6461{
6462	u64 start_time;
6463	long remaining_time;
6464
6465	while (true) {
6466		start_time = get_jiffies_64();
6467		remaining_time = get_nx_lpage_recovery_timeout(start_time);
6468
6469		set_current_state(TASK_INTERRUPTIBLE);
6470		while (!kthread_should_stop() && remaining_time > 0) {
6471			schedule_timeout(remaining_time);
6472			remaining_time = get_nx_lpage_recovery_timeout(start_time);
6473			set_current_state(TASK_INTERRUPTIBLE);
6474		}
6475
6476		set_current_state(TASK_RUNNING);
6477
6478		if (kthread_should_stop())
6479			return 0;
6480
6481		kvm_recover_nx_lpages(kvm);
6482	}
6483}
6484
6485int kvm_mmu_post_init_vm(struct kvm *kvm)
6486{
6487	int err;
6488
6489	err = kvm_vm_create_worker_thread(kvm, kvm_nx_lpage_recovery_worker, 0,
6490					  "kvm-nx-lpage-recovery",
6491					  &kvm->arch.nx_lpage_recovery_thread);
6492	if (!err)
6493		kthread_unpark(kvm->arch.nx_lpage_recovery_thread);
6494
6495	return err;
6496}
6497
6498void kvm_mmu_pre_destroy_vm(struct kvm *kvm)
6499{
6500	if (kvm->arch.nx_lpage_recovery_thread)
6501		kthread_stop(kvm->arch.nx_lpage_recovery_thread);
6502}