arch/x86/kvm/xen.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright © 2019 Oracle and/or its affiliates. All rights reserved.
   4  * Copyright © 2020 Amazon.com, Inc. or its affiliates. All Rights Reserved.
   5  *
   6  * KVM Xen emulation
   7  */
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include "x86.h"
  11 #include "xen.h"
  12 #include "hyperv.h"
  13 #include "lapic.h"
  14
  15 #include <linux/eventfd.h>
  16 #include <linux/kvm_host.h>
  17 #include <linux/sched/stat.h>
  18
  19 #include <trace/events/kvm.h>
  20 #include <xen/interface/xen.h>
  21 #include <xen/interface/vcpu.h>
  22 #include <xen/interface/version.h>
  23 #include <xen/interface/event_channel.h>
  24 #include <xen/interface/sched.h>
  25
  26 #include <asm/xen/cpuid.h>
  27
  28 #include "cpuid.h"
  29 #include "trace.h"
  30
  31 static int kvm_xen_set_evtchn(struct kvm_xen_evtchn *xe, struct kvm *kvm);
  32 static int kvm_xen_setattr_evtchn(struct kvm *kvm, struct kvm_xen_hvm_attr *data);
  33 static bool kvm_xen_hcall_evtchn_send(struct kvm_vcpu *vcpu, u64 param, u64 *r);
  34
  35 DEFINE_STATIC_KEY_DEFERRED_FALSE(kvm_xen_enabled, HZ);
  36
  37 static int kvm_xen_shared_info_init(struct kvm *kvm, gfn_t gfn)
  38 {
  39         struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
  40         struct pvclock_wall_clock *wc;
  41         gpa_t gpa = gfn_to_gpa(gfn);
  42         u32 *wc_sec_hi;
  43         u32 wc_version;
  44         u64 wall_nsec;
  45         int ret = 0;
  46         int idx = srcu_read_lock(&kvm->srcu);
  47
  48         if (gfn == KVM_XEN_INVALID_GFN) {
  49                 kvm_gpc_deactivate(gpc);
  50                 goto out;
  51         }
  52
  53         do {
  54                 ret = kvm_gpc_activate(gpc, gpa, PAGE_SIZE);
  55                 if (ret)
  56                         goto out;
  57
  58                 /*
  59                  * This code mirrors kvm_write_wall_clock() except that it writes
  60                  * directly through the pfn cache and doesn't mark the page dirty.
  61                  */
  62                 wall_nsec = ktime_get_real_ns() - get_kvmclock_ns(kvm);
  63
  64                 /* It could be invalid again already, so we need to check */
  65                 read_lock_irq(&gpc->lock);
  66
  67                 if (gpc->valid)
  68                         break;
  69
  70                 read_unlock_irq(&gpc->lock);
  71         } while (1);
  72
  73         /* Paranoia checks on the 32-bit struct layout */
  74         BUILD_BUG_ON(offsetof(struct compat_shared_info, wc) != 0x900);
  75         BUILD_BUG_ON(offsetof(struct compat_shared_info, arch.wc_sec_hi) != 0x924);
  76         BUILD_BUG_ON(offsetof(struct pvclock_vcpu_time_info, version) != 0);
  77
  78 #ifdef CONFIG_X86_64
  79         /* Paranoia checks on the 64-bit struct layout */
  80         BUILD_BUG_ON(offsetof(struct shared_info, wc) != 0xc00);
  81         BUILD_BUG_ON(offsetof(struct shared_info, wc_sec_hi) != 0xc0c);
  82
  83         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
  84                 struct shared_info *shinfo = gpc->khva;
  85
  86                 wc_sec_hi = &shinfo->wc_sec_hi;
  87                 wc = &shinfo->wc;
  88         } else
  89 #endif
  90         {
  91                 struct compat_shared_info *shinfo = gpc->khva;
  92
  93                 wc_sec_hi = &shinfo->arch.wc_sec_hi;
  94                 wc = &shinfo->wc;
  95         }
  96
  97         /* Increment and ensure an odd value */
  98         wc_version = wc->version = (wc->version + 1) | 1;
  99         smp_wmb();
 100
 101         wc->nsec = do_div(wall_nsec,  1000000000);
 102         wc->sec = (u32)wall_nsec;
 103         *wc_sec_hi = wall_nsec >> 32;
 104         smp_wmb();
 105
 106         wc->version = wc_version + 1;
 107         read_unlock_irq(&gpc->lock);
 108
 109         kvm_make_all_cpus_request(kvm, KVM_REQ_MASTERCLOCK_UPDATE);
 110
 111 out:
 112         srcu_read_unlock(&kvm->srcu, idx);
 113         return ret;
 114 }
 115
 116 void kvm_xen_inject_timer_irqs(struct kvm_vcpu *vcpu)
 117 {
 118         if (atomic_read(&vcpu->arch.xen.timer_pending) > 0) {
 119                 struct kvm_xen_evtchn e;
 120
 121                 e.vcpu_id = vcpu->vcpu_id;
 122                 e.vcpu_idx = vcpu->vcpu_idx;
 123                 e.port = vcpu->arch.xen.timer_virq;
 124                 e.priority = KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL;
 125
 126                 kvm_xen_set_evtchn(&e, vcpu->kvm);
 127
 128                 vcpu->arch.xen.timer_expires = 0;
 129                 atomic_set(&vcpu->arch.xen.timer_pending, 0);
 130         }
 131 }
 132
 133 static enum hrtimer_restart xen_timer_callback(struct hrtimer *timer)
 134 {
 135         struct kvm_vcpu *vcpu = container_of(timer, struct kvm_vcpu,
 136                                              arch.xen.timer);
 137         if (atomic_read(&vcpu->arch.xen.timer_pending))
 138                 return HRTIMER_NORESTART;
 139
 140         atomic_inc(&vcpu->arch.xen.timer_pending);
 141         kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
 142         kvm_vcpu_kick(vcpu);
 143
 144         return HRTIMER_NORESTART;
 145 }
 146
 147 static void kvm_xen_start_timer(struct kvm_vcpu *vcpu, u64 guest_abs, s64 delta_ns)
 148 {
 149         atomic_set(&vcpu->arch.xen.timer_pending, 0);
 150         vcpu->arch.xen.timer_expires = guest_abs;
 151
 152         if (delta_ns <= 0) {
 153                 xen_timer_callback(&vcpu->arch.xen.timer);
 154         } else {
 155                 ktime_t ktime_now = ktime_get();
 156                 hrtimer_start(&vcpu->arch.xen.timer,
 157                               ktime_add_ns(ktime_now, delta_ns),
 158                               HRTIMER_MODE_ABS_HARD);
 159         }
 160 }
 161
 162 static void kvm_xen_stop_timer(struct kvm_vcpu *vcpu)
 163 {
 164         hrtimer_cancel(&vcpu->arch.xen.timer);
 165         vcpu->arch.xen.timer_expires = 0;
 166         atomic_set(&vcpu->arch.xen.timer_pending, 0);
 167 }
 168
 169 static void kvm_xen_init_timer(struct kvm_vcpu *vcpu)
 170 {
 171         hrtimer_init(&vcpu->arch.xen.timer, CLOCK_MONOTONIC,
 172                      HRTIMER_MODE_ABS_HARD);
 173         vcpu->arch.xen.timer.function = xen_timer_callback;
 174 }
 175
 176 static void kvm_xen_update_runstate_guest(struct kvm_vcpu *v, bool atomic)
 177 {
 178         struct kvm_vcpu_xen *vx = &v->arch.xen;
 179         struct gfn_to_pfn_cache *gpc1 = &vx->runstate_cache;
 180         struct gfn_to_pfn_cache *gpc2 = &vx->runstate2_cache;
 181         size_t user_len, user_len1, user_len2;
 182         struct vcpu_runstate_info rs;
 183         unsigned long flags;
 184         size_t times_ofs;
 185         uint8_t *update_bit = NULL;
 186         uint64_t entry_time;
 187         uint64_t *rs_times;
 188         int *rs_state;
 189
 190         /*
 191          * The only difference between 32-bit and 64-bit versions of the
 192          * runstate struct is the alignment of uint64_t in 32-bit, which
 193          * means that the 64-bit version has an additional 4 bytes of
 194          * padding after the first field 'state'. Let's be really really
 195          * paranoid about that, and matching it with our internal data
 196          * structures that we memcpy into it...
 197          */
 198         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state) != 0);
 199         BUILD_BUG_ON(offsetof(struct compat_vcpu_runstate_info, state) != 0);
 200         BUILD_BUG_ON(sizeof(struct compat_vcpu_runstate_info) != 0x2c);
 201 #ifdef CONFIG_X86_64
 202         /*
 203          * The 64-bit structure has 4 bytes of padding before 'state_entry_time'
 204          * so each subsequent field is shifted by 4, and it's 4 bytes longer.
 205          */
 206         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state_entry_time) !=
 207                      offsetof(struct compat_vcpu_runstate_info, state_entry_time) + 4);
 208         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, time) !=
 209                      offsetof(struct compat_vcpu_runstate_info, time) + 4);
 210         BUILD_BUG_ON(sizeof(struct vcpu_runstate_info) != 0x2c + 4);
 211 #endif
 212         /*
 213          * The state field is in the same place at the start of both structs,
 214          * and is the same size (int) as vx->current_runstate.
 215          */
 216         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state) !=
 217                      offsetof(struct compat_vcpu_runstate_info, state));
 218         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, state) !=
 219                      sizeof(vx->current_runstate));
 220         BUILD_BUG_ON(sizeof_field(struct compat_vcpu_runstate_info, state) !=
 221                      sizeof(vx->current_runstate));
 222
 223         /*
 224          * The state_entry_time field is 64 bits in both versions, and the
 225          * XEN_RUNSTATE_UPDATE flag is in the top bit, which given that x86
 226          * is little-endian means that it's in the last *byte* of the word.
 227          * That detail is important later.
 228          */
 229         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, state_entry_time) !=
 230                      sizeof(uint64_t));
 231         BUILD_BUG_ON(sizeof_field(struct compat_vcpu_runstate_info, state_entry_time) !=
 232                      sizeof(uint64_t));
 233         BUILD_BUG_ON((XEN_RUNSTATE_UPDATE >> 56) != 0x80);
 234
 235         /*
 236          * The time array is four 64-bit quantities in both versions, matching
 237          * the vx->runstate_times and immediately following state_entry_time.
 238          */
 239         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state_entry_time) !=
 240                      offsetof(struct vcpu_runstate_info, time) - sizeof(uint64_t));
 241         BUILD_BUG_ON(offsetof(struct compat_vcpu_runstate_info, state_entry_time) !=
 242                      offsetof(struct compat_vcpu_runstate_info, time) - sizeof(uint64_t));
 243         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, time) !=
 244                      sizeof_field(struct compat_vcpu_runstate_info, time));
 245         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, time) !=
 246                      sizeof(vx->runstate_times));
 247
 248         if (IS_ENABLED(CONFIG_64BIT) && v->kvm->arch.xen.long_mode) {
 249                 user_len = sizeof(struct vcpu_runstate_info);
 250                 times_ofs = offsetof(struct vcpu_runstate_info,
 251                                      state_entry_time);
 252         } else {
 253                 user_len = sizeof(struct compat_vcpu_runstate_info);
 254                 times_ofs = offsetof(struct compat_vcpu_runstate_info,
 255                                      state_entry_time);
 256         }
 257
 258         /*
 259          * There are basically no alignment constraints. The guest can set it
 260          * up so it crosses from one page to the next, and at arbitrary byte
 261          * alignment (and the 32-bit ABI doesn't align the 64-bit integers
 262          * anyway, even if the overall struct had been 64-bit aligned).
 263          */
 264         if ((gpc1->gpa & ~PAGE_MASK) + user_len >= PAGE_SIZE) {
 265                 user_len1 = PAGE_SIZE - (gpc1->gpa & ~PAGE_MASK);
 266                 user_len2 = user_len - user_len1;
 267         } else {
 268                 user_len1 = user_len;
 269                 user_len2 = 0;
 270         }
 271         BUG_ON(user_len1 + user_len2 != user_len);
 272
 273  retry:
 274         /*
 275          * Attempt to obtain the GPC lock on *both* (if there are two)
 276          * gfn_to_pfn caches that cover the region.
 277          */
 278         if (atomic) {
 279                 local_irq_save(flags);
 280                 if (!read_trylock(&gpc1->lock)) {
 281                         local_irq_restore(flags);
 282                         return;
 283                 }
 284         } else {
 285                 read_lock_irqsave(&gpc1->lock, flags);
 286         }
 287         while (!kvm_gpc_check(gpc1, user_len1)) {
 288                 read_unlock_irqrestore(&gpc1->lock, flags);
 289
 290                 /* When invoked from kvm_sched_out() we cannot sleep */
 291                 if (atomic)
 292                         return;
 293
 294                 if (kvm_gpc_refresh(gpc1, user_len1))
 295                         return;
 296
 297                 read_lock_irqsave(&gpc1->lock, flags);
 298         }
 299
 300         if (likely(!user_len2)) {
 301                 /*
 302                  * Set up three pointers directly to the runstate_info
 303                  * struct in the guest (via the GPC).
 304                  *
 305                  *  • @rs_state   → state field
 306                  *  • @rs_times   → state_entry_time field.
 307                  *  • @update_bit → last byte of state_entry_time, which
 308                  *                  contains the XEN_RUNSTATE_UPDATE bit.
 309                  */
 310                 rs_state = gpc1->khva;
 311                 rs_times = gpc1->khva + times_ofs;
 312                 if (v->kvm->arch.xen.runstate_update_flag)
 313                         update_bit = ((void *)(&rs_times[1])) - 1;
 314         } else {
 315                 /*
 316                  * The guest's runstate_info is split across two pages and we
 317                  * need to hold and validate both GPCs simultaneously. We can
 318                  * declare a lock ordering GPC1 > GPC2 because nothing else
 319                  * takes them more than one at a time. Set a subclass on the
 320                  * gpc1 lock to make lockdep shut up about it.
 321                  */
 322                 lock_set_subclass(&gpc1->lock.dep_map, 1, _THIS_IP_);
 323                 if (atomic) {
 324                         if (!read_trylock(&gpc2->lock)) {
 325                                 read_unlock_irqrestore(&gpc1->lock, flags);
 326                                 return;
 327                         }
 328                 } else {
 329                         read_lock(&gpc2->lock);
 330                 }
 331
 332                 if (!kvm_gpc_check(gpc2, user_len2)) {
 333                         read_unlock(&gpc2->lock);
 334                         read_unlock_irqrestore(&gpc1->lock, flags);
 335
 336                         /* When invoked from kvm_sched_out() we cannot sleep */
 337                         if (atomic)
 338                                 return;
 339
 340                         /*
 341                          * Use kvm_gpc_activate() here because if the runstate
 342                          * area was configured in 32-bit mode and only extends
 343                          * to the second page now because the guest changed to
 344                          * 64-bit mode, the second GPC won't have been set up.
 345                          */
 346                         if (kvm_gpc_activate(gpc2, gpc1->gpa + user_len1,
 347                                              user_len2))
 348                                 return;
 349
 350                         /*
 351                          * We dropped the lock on GPC1 so we have to go all the
 352                          * way back and revalidate that too.
 353                          */
 354                         goto retry;
 355                 }
 356
 357                 /*
 358                  * In this case, the runstate_info struct will be assembled on
 359                  * the kernel stack (compat or not as appropriate) and will
 360                  * be copied to GPC1/GPC2 with a dual memcpy. Set up the three
 361                  * rs pointers accordingly.
 362                  */
 363                 rs_times = &rs.state_entry_time;
 364
 365                 /*
 366                  * The rs_state pointer points to the start of what we'll
 367                  * copy to the guest, which in the case of a compat guest
 368                  * is the 32-bit field that the compiler thinks is padding.
 369                  */
 370                 rs_state = ((void *)rs_times) - times_ofs;
 371
 372                 /*
 373                  * The update_bit is still directly in the guest memory,
 374                  * via one GPC or the other.
 375                  */
 376                 if (v->kvm->arch.xen.runstate_update_flag) {
 377                         if (user_len1 >= times_ofs + sizeof(uint64_t))
 378                                 update_bit = gpc1->khva + times_ofs +
 379                                         sizeof(uint64_t) - 1;
 380                         else
 381                                 update_bit = gpc2->khva + times_ofs +
 382                                         sizeof(uint64_t) - 1 - user_len1;
 383                 }
 384
 385 #ifdef CONFIG_X86_64
 386                 /*
 387                  * Don't leak kernel memory through the padding in the 64-bit
 388                  * version of the struct.
 389                  */
 390                 memset(&rs, 0, offsetof(struct vcpu_runstate_info, state_entry_time));
 391 #endif
 392         }
 393
 394         /*
 395          * First, set the XEN_RUNSTATE_UPDATE bit in the top bit of the
 396          * state_entry_time field, directly in the guest. We need to set
 397          * that (and write-barrier) before writing to the rest of the
 398          * structure, and clear it last. Just as Xen does, we address the
 399          * single *byte* in which it resides because it might be in a
 400          * different cache line to the rest of the 64-bit word, due to
 401          * the (lack of) alignment constraints.
 402          */
 403         entry_time = vx->runstate_entry_time;
 404         if (update_bit) {
 405                 entry_time |= XEN_RUNSTATE_UPDATE;
 406                 *update_bit = (vx->runstate_entry_time | XEN_RUNSTATE_UPDATE) >> 56;
 407                 smp_wmb();
 408         }
 409
 410         /*
 411          * Now assemble the actual structure, either on our kernel stack
 412          * or directly in the guest according to how the rs_state and
 413          * rs_times pointers were set up above.
 414          */
 415         *rs_state = vx->current_runstate;
 416         rs_times[0] = entry_time;
 417         memcpy(rs_times + 1, vx->runstate_times, sizeof(vx->runstate_times));
 418
 419         /* For the split case, we have to then copy it to the guest. */
 420         if (user_len2) {
 421                 memcpy(gpc1->khva, rs_state, user_len1);
 422                 memcpy(gpc2->khva, ((void *)rs_state) + user_len1, user_len2);
 423         }
 424         smp_wmb();
 425
 426         /* Finally, clear the XEN_RUNSTATE_UPDATE bit. */
 427         if (update_bit) {
 428                 entry_time &= ~XEN_RUNSTATE_UPDATE;
 429                 *update_bit = entry_time >> 56;
 430                 smp_wmb();
 431         }
 432
 433         if (user_len2)
 434                 read_unlock(&gpc2->lock);
 435
 436         read_unlock_irqrestore(&gpc1->lock, flags);
 437
 438         mark_page_dirty_in_slot(v->kvm, gpc1->memslot, gpc1->gpa >> PAGE_SHIFT);
 439         if (user_len2)
 440                 mark_page_dirty_in_slot(v->kvm, gpc2->memslot, gpc2->gpa >> PAGE_SHIFT);
 441 }
 442
 443 void kvm_xen_update_runstate(struct kvm_vcpu *v, int state)
 444 {
 445         struct kvm_vcpu_xen *vx = &v->arch.xen;
 446         u64 now = get_kvmclock_ns(v->kvm);
 447         u64 delta_ns = now - vx->runstate_entry_time;
 448         u64 run_delay = current->sched_info.run_delay;
 449
 450         if (unlikely(!vx->runstate_entry_time))
 451                 vx->current_runstate = RUNSTATE_offline;
 452
 453         /*
 454          * Time waiting for the scheduler isn't "stolen" if the
 455          * vCPU wasn't running anyway.
 456          */
 457         if (vx->current_runstate == RUNSTATE_running) {
 458                 u64 steal_ns = run_delay - vx->last_steal;
 459
 460                 delta_ns -= steal_ns;
 461
 462                 vx->runstate_times[RUNSTATE_runnable] += steal_ns;
 463         }
 464         vx->last_steal = run_delay;
 465
 466         vx->runstate_times[vx->current_runstate] += delta_ns;
 467         vx->current_runstate = state;
 468         vx->runstate_entry_time = now;
 469
 470         if (vx->runstate_cache.active)
 471                 kvm_xen_update_runstate_guest(v, state == RUNSTATE_runnable);
 472 }
 473
 474 static void kvm_xen_inject_vcpu_vector(struct kvm_vcpu *v)
 475 {
 476         struct kvm_lapic_irq irq = { };
 477         int r;
 478
 479         irq.dest_id = v->vcpu_id;
 480         irq.vector = v->arch.xen.upcall_vector;
 481         irq.dest_mode = APIC_DEST_PHYSICAL;
 482         irq.shorthand = APIC_DEST_NOSHORT;
 483         irq.delivery_mode = APIC_DM_FIXED;
 484         irq.level = 1;
 485
 486         /* The fast version will always work for physical unicast */
 487         WARN_ON_ONCE(!kvm_irq_delivery_to_apic_fast(v->kvm, NULL, &irq, &r, NULL));
 488 }
 489
 490 /*
 491  * On event channel delivery, the vcpu_info may not have been accessible.
 492  * In that case, there are bits in vcpu->arch.xen.evtchn_pending_sel which
 493  * need to be marked into the vcpu_info (and evtchn_upcall_pending set).
 494  * Do so now that we can sleep in the context of the vCPU to bring the
 495  * page in, and refresh the pfn cache for it.
 496  */
 497 void kvm_xen_inject_pending_events(struct kvm_vcpu *v)
 498 {
 499         unsigned long evtchn_pending_sel = READ_ONCE(v->arch.xen.evtchn_pending_sel);
 500         struct gfn_to_pfn_cache *gpc = &v->arch.xen.vcpu_info_cache;
 501         unsigned long flags;
 502
 503         if (!evtchn_pending_sel)
 504                 return;
 505
 506         /*
 507          * Yes, this is an open-coded loop. But that's just what put_user()
 508          * does anyway. Page it in and retry the instruction. We're just a
 509          * little more honest about it.
 510          */
 511         read_lock_irqsave(&gpc->lock, flags);
 512         while (!kvm_gpc_check(gpc, sizeof(struct vcpu_info))) {
 513                 read_unlock_irqrestore(&gpc->lock, flags);
 514
 515                 if (kvm_gpc_refresh(gpc, sizeof(struct vcpu_info)))
 516                         return;
 517
 518                 read_lock_irqsave(&gpc->lock, flags);
 519         }
 520
 521         /* Now gpc->khva is a valid kernel address for the vcpu_info */
 522         if (IS_ENABLED(CONFIG_64BIT) && v->kvm->arch.xen.long_mode) {
 523                 struct vcpu_info *vi = gpc->khva;
 524
 525                 asm volatile(LOCK_PREFIX "orq %0, %1\n"
 526                              "notq %0\n"
 527                              LOCK_PREFIX "andq %0, %2\n"
 528                              : "=r" (evtchn_pending_sel),
 529                                "+m" (vi->evtchn_pending_sel),
 530                                "+m" (v->arch.xen.evtchn_pending_sel)
 531                              : "0" (evtchn_pending_sel));
 532                 WRITE_ONCE(vi->evtchn_upcall_pending, 1);
 533         } else {
 534                 u32 evtchn_pending_sel32 = evtchn_pending_sel;
 535                 struct compat_vcpu_info *vi = gpc->khva;
 536
 537                 asm volatile(LOCK_PREFIX "orl %0, %1\n"
 538                              "notl %0\n"
 539                              LOCK_PREFIX "andl %0, %2\n"
 540                              : "=r" (evtchn_pending_sel32),
 541                                "+m" (vi->evtchn_pending_sel),
 542                                "+m" (v->arch.xen.evtchn_pending_sel)
 543                              : "0" (evtchn_pending_sel32));
 544                 WRITE_ONCE(vi->evtchn_upcall_pending, 1);
 545         }
 546         read_unlock_irqrestore(&gpc->lock, flags);
 547
 548         /* For the per-vCPU lapic vector, deliver it as MSI. */
 549         if (v->arch.xen.upcall_vector)
 550                 kvm_xen_inject_vcpu_vector(v);
 551
 552         mark_page_dirty_in_slot(v->kvm, gpc->memslot, gpc->gpa >> PAGE_SHIFT);
 553 }
 554
 555 int __kvm_xen_has_interrupt(struct kvm_vcpu *v)
 556 {
 557         struct gfn_to_pfn_cache *gpc = &v->arch.xen.vcpu_info_cache;
 558         unsigned long flags;
 559         u8 rc = 0;
 560
 561         /*
 562          * If the global upcall vector (HVMIRQ_callback_vector) is set and
 563          * the vCPU's evtchn_upcall_pending flag is set, the IRQ is pending.
 564          */
 565
 566         /* No need for compat handling here */
 567         BUILD_BUG_ON(offsetof(struct vcpu_info, evtchn_upcall_pending) !=
 568                      offsetof(struct compat_vcpu_info, evtchn_upcall_pending));
 569         BUILD_BUG_ON(sizeof(rc) !=
 570                      sizeof_field(struct vcpu_info, evtchn_upcall_pending));
 571         BUILD_BUG_ON(sizeof(rc) !=
 572                      sizeof_field(struct compat_vcpu_info, evtchn_upcall_pending));
 573
 574         read_lock_irqsave(&gpc->lock, flags);
 575         while (!kvm_gpc_check(gpc, sizeof(struct vcpu_info))) {
 576                 read_unlock_irqrestore(&gpc->lock, flags);
 577
 578                 /*
 579                  * This function gets called from kvm_vcpu_block() after setting the
 580                  * task to TASK_INTERRUPTIBLE, to see if it needs to wake immediately
 581                  * from a HLT. So we really mustn't sleep. If the page ended up absent
 582                  * at that point, just return 1 in order to trigger an immediate wake,
 583                  * and we'll end up getting called again from a context where we *can*
 584                  * fault in the page and wait for it.
 585                  */
 586                 if (in_atomic() || !task_is_running(current))
 587                         return 1;
 588
 589                 if (kvm_gpc_refresh(gpc, sizeof(struct vcpu_info))) {
 590                         /*
 591                          * If this failed, userspace has screwed up the
 592                          * vcpu_info mapping. No interrupts for you.
 593                          */
 594                         return 0;
 595                 }
 596                 read_lock_irqsave(&gpc->lock, flags);
 597         }
 598
 599         rc = ((struct vcpu_info *)gpc->khva)->evtchn_upcall_pending;
 600         read_unlock_irqrestore(&gpc->lock, flags);
 601         return rc;
 602 }
 603
 604 int kvm_xen_hvm_set_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 605 {
 606         int r = -ENOENT;
 607
 608
 609         switch (data->type) {
 610         case KVM_XEN_ATTR_TYPE_LONG_MODE:
 611                 if (!IS_ENABLED(CONFIG_64BIT) && data->u.long_mode) {
 612                         r = -EINVAL;
 613                 } else {
 614                         mutex_lock(&kvm->arch.xen.xen_lock);
 615                         kvm->arch.xen.long_mode = !!data->u.long_mode;
 616                         mutex_unlock(&kvm->arch.xen.xen_lock);
 617                         r = 0;
 618                 }
 619                 break;
 620
 621         case KVM_XEN_ATTR_TYPE_SHARED_INFO:
 622                 mutex_lock(&kvm->arch.xen.xen_lock);
 623                 r = kvm_xen_shared_info_init(kvm, data->u.shared_info.gfn);
 624                 mutex_unlock(&kvm->arch.xen.xen_lock);
 625                 break;
 626
 627         case KVM_XEN_ATTR_TYPE_UPCALL_VECTOR:
 628                 if (data->u.vector && data->u.vector < 0x10)
 629                         r = -EINVAL;
 630                 else {
 631                         mutex_lock(&kvm->arch.xen.xen_lock);
 632                         kvm->arch.xen.upcall_vector = data->u.vector;
 633                         mutex_unlock(&kvm->arch.xen.xen_lock);
 634                         r = 0;
 635                 }
 636                 break;
 637
 638         case KVM_XEN_ATTR_TYPE_EVTCHN:
 639                 r = kvm_xen_setattr_evtchn(kvm, data);
 640                 break;
 641
 642         case KVM_XEN_ATTR_TYPE_XEN_VERSION:
 643                 mutex_lock(&kvm->arch.xen.xen_lock);
 644                 kvm->arch.xen.xen_version = data->u.xen_version;
 645                 mutex_unlock(&kvm->arch.xen.xen_lock);
 646                 r = 0;
 647                 break;
 648
 649         case KVM_XEN_ATTR_TYPE_RUNSTATE_UPDATE_FLAG:
 650                 if (!sched_info_on()) {
 651                         r = -EOPNOTSUPP;
 652                         break;
 653                 }
 654                 mutex_lock(&kvm->arch.xen.xen_lock);
 655                 kvm->arch.xen.runstate_update_flag = !!data->u.runstate_update_flag;
 656                 mutex_unlock(&kvm->arch.xen.xen_lock);
 657                 r = 0;
 658                 break;
 659
 660         default:
 661                 break;
 662         }
 663
 664         return r;
 665 }
 666
 667 int kvm_xen_hvm_get_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 668 {
 669         int r = -ENOENT;
 670
 671         mutex_lock(&kvm->arch.xen.xen_lock);
 672
 673         switch (data->type) {
 674         case KVM_XEN_ATTR_TYPE_LONG_MODE:
 675                 data->u.long_mode = kvm->arch.xen.long_mode;
 676                 r = 0;
 677                 break;
 678
 679         case KVM_XEN_ATTR_TYPE_SHARED_INFO:
 680                 if (kvm->arch.xen.shinfo_cache.active)
 681                         data->u.shared_info.gfn = gpa_to_gfn(kvm->arch.xen.shinfo_cache.gpa);
 682                 else
 683                         data->u.shared_info.gfn = KVM_XEN_INVALID_GFN;
 684                 r = 0;
 685                 break;
 686
 687         case KVM_XEN_ATTR_TYPE_UPCALL_VECTOR:
 688                 data->u.vector = kvm->arch.xen.upcall_vector;
 689                 r = 0;
 690                 break;
 691
 692         case KVM_XEN_ATTR_TYPE_XEN_VERSION:
 693                 data->u.xen_version = kvm->arch.xen.xen_version;
 694                 r = 0;
 695                 break;
 696
 697         case KVM_XEN_ATTR_TYPE_RUNSTATE_UPDATE_FLAG:
 698                 if (!sched_info_on()) {
 699                         r = -EOPNOTSUPP;
 700                         break;
 701                 }
 702                 data->u.runstate_update_flag = kvm->arch.xen.runstate_update_flag;
 703                 r = 0;
 704                 break;
 705
 706         default:
 707                 break;
 708         }
 709
 710         mutex_unlock(&kvm->arch.xen.xen_lock);
 711         return r;
 712 }
 713
 714 int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 715 {
 716         int idx, r = -ENOENT;
 717
 718         mutex_lock(&vcpu->kvm->arch.xen.xen_lock);
 719         idx = srcu_read_lock(&vcpu->kvm->srcu);
 720
 721         switch (data->type) {
 722         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
 723                 /* No compat necessary here. */
 724                 BUILD_BUG_ON(sizeof(struct vcpu_info) !=
 725                              sizeof(struct compat_vcpu_info));
 726                 BUILD_BUG_ON(offsetof(struct vcpu_info, time) !=
 727                              offsetof(struct compat_vcpu_info, time));
 728
 729                 if (data->u.gpa == KVM_XEN_INVALID_GPA) {
 730                         kvm_gpc_deactivate(&vcpu->arch.xen.vcpu_info_cache);
 731                         r = 0;
 732                         break;
 733                 }
 734
 735                 r = kvm_gpc_activate(&vcpu->arch.xen.vcpu_info_cache,
 736                                      data->u.gpa, sizeof(struct vcpu_info));
 737                 if (!r)
 738                         kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
 739
 740                 break;
 741
 742         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_TIME_INFO:
 743                 if (data->u.gpa == KVM_XEN_INVALID_GPA) {
 744                         kvm_gpc_deactivate(&vcpu->arch.xen.vcpu_time_info_cache);
 745                         r = 0;
 746                         break;
 747                 }
 748
 749                 r = kvm_gpc_activate(&vcpu->arch.xen.vcpu_time_info_cache,
 750                                      data->u.gpa,
 751                                      sizeof(struct pvclock_vcpu_time_info));
 752                 if (!r)
 753                         kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
 754                 break;
 755
 756         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADDR: {
 757                 size_t sz, sz1, sz2;
 758
 759                 if (!sched_info_on()) {
 760                         r = -EOPNOTSUPP;
 761                         break;
 762                 }
 763                 if (data->u.gpa == KVM_XEN_INVALID_GPA) {
 764                         r = 0;
 765                 deactivate_out:
 766                         kvm_gpc_deactivate(&vcpu->arch.xen.runstate_cache);
 767                         kvm_gpc_deactivate(&vcpu->arch.xen.runstate2_cache);
 768                         break;
 769                 }
 770
 771                 /*
 772                  * If the guest switches to 64-bit mode after setting the runstate
 773                  * address, that's actually OK. kvm_xen_update_runstate_guest()
 774                  * will cope.
 775                  */
 776                 if (IS_ENABLED(CONFIG_64BIT) && vcpu->kvm->arch.xen.long_mode)
 777                         sz = sizeof(struct vcpu_runstate_info);
 778                 else
 779                         sz = sizeof(struct compat_vcpu_runstate_info);
 780
 781                 /* How much fits in the (first) page? */
 782                 sz1 = PAGE_SIZE - (data->u.gpa & ~PAGE_MASK);
 783                 r = kvm_gpc_activate(&vcpu->arch.xen.runstate_cache,
 784                                      data->u.gpa, sz1);
 785                 if (r)
 786                         goto deactivate_out;
 787
 788                 /* Either map the second page, or deactivate the second GPC */
 789                 if (sz1 >= sz) {
 790                         kvm_gpc_deactivate(&vcpu->arch.xen.runstate2_cache);
 791                 } else {
 792                         sz2 = sz - sz1;
 793                         BUG_ON((data->u.gpa + sz1) & ~PAGE_MASK);
 794                         r = kvm_gpc_activate(&vcpu->arch.xen.runstate2_cache,
 795                                              data->u.gpa + sz1, sz2);
 796                         if (r)
 797                                 goto deactivate_out;
 798                 }
 799
 800                 kvm_xen_update_runstate_guest(vcpu, false);
 801                 break;
 802         }
 803         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_CURRENT:
 804                 if (!sched_info_on()) {
 805                         r = -EOPNOTSUPP;
 806                         break;
 807                 }
 808                 if (data->u.runstate.state > RUNSTATE_offline) {
 809                         r = -EINVAL;
 810                         break;
 811                 }
 812
 813                 kvm_xen_update_runstate(vcpu, data->u.runstate.state);
 814                 r = 0;
 815                 break;
 816
 817         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_DATA:
 818                 if (!sched_info_on()) {
 819                         r = -EOPNOTSUPP;
 820                         break;
 821                 }
 822                 if (data->u.runstate.state > RUNSTATE_offline) {
 823                         r = -EINVAL;
 824                         break;
 825                 }
 826                 if (data->u.runstate.state_entry_time !=
 827                     (data->u.runstate.time_running +
 828                      data->u.runstate.time_runnable +
 829                      data->u.runstate.time_blocked +
 830                      data->u.runstate.time_offline)) {
 831                         r = -EINVAL;
 832                         break;
 833                 }
 834                 if (get_kvmclock_ns(vcpu->kvm) <
 835                     data->u.runstate.state_entry_time) {
 836                         r = -EINVAL;
 837                         break;
 838                 }
 839
 840                 vcpu->arch.xen.current_runstate = data->u.runstate.state;
 841                 vcpu->arch.xen.runstate_entry_time =
 842                         data->u.runstate.state_entry_time;
 843                 vcpu->arch.xen.runstate_times[RUNSTATE_running] =
 844                         data->u.runstate.time_running;
 845                 vcpu->arch.xen.runstate_times[RUNSTATE_runnable] =
 846                         data->u.runstate.time_runnable;
 847                 vcpu->arch.xen.runstate_times[RUNSTATE_blocked] =
 848                         data->u.runstate.time_blocked;
 849                 vcpu->arch.xen.runstate_times[RUNSTATE_offline] =
 850                         data->u.runstate.time_offline;
 851                 vcpu->arch.xen.last_steal = current->sched_info.run_delay;
 852                 r = 0;
 853                 break;
 854
 855         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADJUST:
 856                 if (!sched_info_on()) {
 857                         r = -EOPNOTSUPP;
 858                         break;
 859                 }
 860                 if (data->u.runstate.state > RUNSTATE_offline &&
 861                     data->u.runstate.state != (u64)-1) {
 862                         r = -EINVAL;
 863                         break;
 864                 }
 865                 /* The adjustment must add up */
 866                 if (data->u.runstate.state_entry_time !=
 867                     (data->u.runstate.time_running +
 868                      data->u.runstate.time_runnable +
 869                      data->u.runstate.time_blocked +
 870                      data->u.runstate.time_offline)) {
 871                         r = -EINVAL;
 872                         break;
 873                 }
 874
 875                 if (get_kvmclock_ns(vcpu->kvm) <
 876                     (vcpu->arch.xen.runstate_entry_time +
 877                      data->u.runstate.state_entry_time)) {
 878                         r = -EINVAL;
 879                         break;
 880                 }
 881
 882                 vcpu->arch.xen.runstate_entry_time +=
 883                         data->u.runstate.state_entry_time;
 884                 vcpu->arch.xen.runstate_times[RUNSTATE_running] +=
 885                         data->u.runstate.time_running;
 886                 vcpu->arch.xen.runstate_times[RUNSTATE_runnable] +=
 887                         data->u.runstate.time_runnable;
 888                 vcpu->arch.xen.runstate_times[RUNSTATE_blocked] +=
 889                         data->u.runstate.time_blocked;
 890                 vcpu->arch.xen.runstate_times[RUNSTATE_offline] +=
 891                         data->u.runstate.time_offline;
 892
 893                 if (data->u.runstate.state <= RUNSTATE_offline)
 894                         kvm_xen_update_runstate(vcpu, data->u.runstate.state);
 895                 else if (vcpu->arch.xen.runstate_cache.active)
 896                         kvm_xen_update_runstate_guest(vcpu, false);
 897                 r = 0;
 898                 break;
 899
 900         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_ID:
 901                 if (data->u.vcpu_id >= KVM_MAX_VCPUS)
 902                         r = -EINVAL;
 903                 else {
 904                         vcpu->arch.xen.vcpu_id = data->u.vcpu_id;
 905                         r = 0;
 906                 }
 907                 break;
 908
 909         case KVM_XEN_VCPU_ATTR_TYPE_TIMER:
 910                 if (data->u.timer.port &&
 911                     data->u.timer.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL) {
 912                         r = -EINVAL;
 913                         break;
 914                 }
 915
 916                 if (!vcpu->arch.xen.timer.function)
 917                         kvm_xen_init_timer(vcpu);
 918
 919                 /* Stop the timer (if it's running) before changing the vector */
 920                 kvm_xen_stop_timer(vcpu);
 921                 vcpu->arch.xen.timer_virq = data->u.timer.port;
 922
 923                 /* Start the timer if the new value has a valid vector+expiry. */
 924                 if (data->u.timer.port && data->u.timer.expires_ns)
 925                         kvm_xen_start_timer(vcpu, data->u.timer.expires_ns,
 926                                             data->u.timer.expires_ns -
 927                                             get_kvmclock_ns(vcpu->kvm));
 928
 929                 r = 0;
 930                 break;
 931
 932         case KVM_XEN_VCPU_ATTR_TYPE_UPCALL_VECTOR:
 933                 if (data->u.vector && data->u.vector < 0x10)
 934                         r = -EINVAL;
 935                 else {
 936                         vcpu->arch.xen.upcall_vector = data->u.vector;
 937                         r = 0;
 938                 }
 939                 break;
 940
 941         default:
 942                 break;
 943         }
 944
 945         srcu_read_unlock(&vcpu->kvm->srcu, idx);
 946         mutex_unlock(&vcpu->kvm->arch.xen.xen_lock);
 947         return r;
 948 }
 949
 950 int kvm_xen_vcpu_get_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 951 {
 952         int r = -ENOENT;
 953
 954         mutex_lock(&vcpu->kvm->arch.xen.xen_lock);
 955
 956         switch (data->type) {
 957         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
 958                 if (vcpu->arch.xen.vcpu_info_cache.active)
 959                         data->u.gpa = vcpu->arch.xen.vcpu_info_cache.gpa;
 960                 else
 961                         data->u.gpa = KVM_XEN_INVALID_GPA;
 962                 r = 0;
 963                 break;
 964
 965         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_TIME_INFO:
 966                 if (vcpu->arch.xen.vcpu_time_info_cache.active)
 967                         data->u.gpa = vcpu->arch.xen.vcpu_time_info_cache.gpa;
 968                 else
 969                         data->u.gpa = KVM_XEN_INVALID_GPA;
 970                 r = 0;
 971                 break;
 972
 973         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADDR:
 974                 if (!sched_info_on()) {
 975                         r = -EOPNOTSUPP;
 976                         break;
 977                 }
 978                 if (vcpu->arch.xen.runstate_cache.active) {
 979                         data->u.gpa = vcpu->arch.xen.runstate_cache.gpa;
 980                         r = 0;
 981                 }
 982                 break;
 983
 984         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_CURRENT:
 985                 if (!sched_info_on()) {
 986                         r = -EOPNOTSUPP;
 987                         break;
 988                 }
 989                 data->u.runstate.state = vcpu->arch.xen.current_runstate;
 990                 r = 0;
 991                 break;
 992
 993         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_DATA:
 994                 if (!sched_info_on()) {
 995                         r = -EOPNOTSUPP;
 996                         break;
 997                 }
 998                 data->u.runstate.state = vcpu->arch.xen.current_runstate;
 999                 data->u.runstate.state_entry_time =
1000                         vcpu->arch.xen.runstate_entry_time;
1001                 data->u.runstate.time_running =
1002                         vcpu->arch.xen.runstate_times[RUNSTATE_running];
1003                 data->u.runstate.time_runnable =
1004                         vcpu->arch.xen.runstate_times[RUNSTATE_runnable];
1005                 data->u.runstate.time_blocked =
1006                         vcpu->arch.xen.runstate_times[RUNSTATE_blocked];
1007                 data->u.runstate.time_offline =
1008                         vcpu->arch.xen.runstate_times[RUNSTATE_offline];
1009                 r = 0;
1010                 break;
1011
1012         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADJUST:
1013                 r = -EINVAL;
1014                 break;
1015
1016         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_ID:
1017                 data->u.vcpu_id = vcpu->arch.xen.vcpu_id;
1018                 r = 0;
1019                 break;
1020
1021         case KVM_XEN_VCPU_ATTR_TYPE_TIMER:
1022                 data->u.timer.port = vcpu->arch.xen.timer_virq;
1023                 data->u.timer.priority = KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL;
1024                 data->u.timer.expires_ns = vcpu->arch.xen.timer_expires;
1025                 r = 0;
1026                 break;
1027
1028         case KVM_XEN_VCPU_ATTR_TYPE_UPCALL_VECTOR:
1029                 data->u.vector = vcpu->arch.xen.upcall_vector;
1030                 r = 0;
1031                 break;
1032
1033         default:
1034                 break;
1035         }
1036
1037         mutex_unlock(&vcpu->kvm->arch.xen.xen_lock);
1038         return r;
1039 }
1040
1041 int kvm_xen_write_hypercall_page(struct kvm_vcpu *vcpu, u64 data)
1042 {
1043         struct kvm *kvm = vcpu->kvm;
1044         u32 page_num = data & ~PAGE_MASK;
1045         u64 page_addr = data & PAGE_MASK;
1046         bool lm = is_long_mode(vcpu);
1047
1048         /* Latch long_mode for shared_info pages etc. */
1049         vcpu->kvm->arch.xen.long_mode = lm;
1050
1051         /*
1052          * If Xen hypercall intercept is enabled, fill the hypercall
1053          * page with VMCALL/VMMCALL instructions since that's what
1054          * we catch. Else the VMM has provided the hypercall pages
1055          * with instructions of its own choosing, so use those.
1056          */
1057         if (kvm_xen_hypercall_enabled(kvm)) {
1058                 u8 instructions[32];
1059                 int i;
1060
1061                 if (page_num)
1062                         return 1;
1063
1064                 /* mov imm32, %eax */
1065                 instructions[0] = 0xb8;
1066
1067                 /* vmcall / vmmcall */
1068                 static_call(kvm_x86_patch_hypercall)(vcpu, instructions + 5);
1069
1070                 /* ret */
1071                 instructions[8] = 0xc3;
1072
1073                 /* int3 to pad */
1074                 memset(instructions + 9, 0xcc, sizeof(instructions) - 9);
1075
1076                 for (i = 0; i < PAGE_SIZE / sizeof(instructions); i++) {
1077                         *(u32 *)&instructions[1] = i;
1078                         if (kvm_vcpu_write_guest(vcpu,
1079                                                  page_addr + (i * sizeof(instructions)),
1080                                                  instructions, sizeof(instructions)))
1081                                 return 1;
1082                 }
1083         } else {
1084                 /*
1085                  * Note, truncation is a non-issue as 'lm' is guaranteed to be
1086                  * false for a 32-bit kernel, i.e. when hva_t is only 4 bytes.
1087                  */
1088                 hva_t blob_addr = lm ? kvm->arch.xen_hvm_config.blob_addr_64
1089                                      : kvm->arch.xen_hvm_config.blob_addr_32;
1090                 u8 blob_size = lm ? kvm->arch.xen_hvm_config.blob_size_64
1091                                   : kvm->arch.xen_hvm_config.blob_size_32;
1092                 u8 *page;
1093                 int ret;
1094
1095                 if (page_num >= blob_size)
1096                         return 1;
1097
1098                 blob_addr += page_num * PAGE_SIZE;
1099
1100                 page = memdup_user((u8 __user *)blob_addr, PAGE_SIZE);
1101                 if (IS_ERR(page))
1102                         return PTR_ERR(page);
1103
1104                 ret = kvm_vcpu_write_guest(vcpu, page_addr, page, PAGE_SIZE);
1105                 kfree(page);
1106                 if (ret)
1107                         return 1;
1108         }
1109         return 0;
1110 }
1111
1112 int kvm_xen_hvm_config(struct kvm *kvm, struct kvm_xen_hvm_config *xhc)
1113 {
1114         /* Only some feature flags need to be *enabled* by userspace */
1115         u32 permitted_flags = KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL |
1116                 KVM_XEN_HVM_CONFIG_EVTCHN_SEND;
1117
1118         if (xhc->flags & ~permitted_flags)
1119                 return -EINVAL;
1120
1121         /*
1122          * With hypercall interception the kernel generates its own
1123          * hypercall page so it must not be provided.
1124          */
1125         if ((xhc->flags & KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL) &&
1126             (xhc->blob_addr_32 || xhc->blob_addr_64 ||
1127              xhc->blob_size_32 || xhc->blob_size_64))
1128                 return -EINVAL;
1129
1130         mutex_lock(&kvm->arch.xen.xen_lock);
1131
1132         if (xhc->msr && !kvm->arch.xen_hvm_config.msr)
1133                 static_branch_inc(&kvm_xen_enabled.key);
1134         else if (!xhc->msr && kvm->arch.xen_hvm_config.msr)
1135                 static_branch_slow_dec_deferred(&kvm_xen_enabled);
1136
1137         memcpy(&kvm->arch.xen_hvm_config, xhc, sizeof(*xhc));
1138
1139         mutex_unlock(&kvm->arch.xen.xen_lock);
1140         return 0;
1141 }
1142
1143 static int kvm_xen_hypercall_set_result(struct kvm_vcpu *vcpu, u64 result)
1144 {
1145         kvm_rax_write(vcpu, result);
1146         return kvm_skip_emulated_instruction(vcpu);
1147 }
1148
1149 static int kvm_xen_hypercall_complete_userspace(struct kvm_vcpu *vcpu)
1150 {
1151         struct kvm_run *run = vcpu->run;
1152
1153         if (unlikely(!kvm_is_linear_rip(vcpu, vcpu->arch.xen.hypercall_rip)))
1154                 return 1;
1155
1156         return kvm_xen_hypercall_set_result(vcpu, run->xen.u.hcall.result);
1157 }
1158
1159 static inline int max_evtchn_port(struct kvm *kvm)
1160 {
1161         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode)
1162                 return EVTCHN_2L_NR_CHANNELS;
1163         else
1164                 return COMPAT_EVTCHN_2L_NR_CHANNELS;
1165 }
1166
1167 static bool wait_pending_event(struct kvm_vcpu *vcpu, int nr_ports,
1168                                evtchn_port_t *ports)
1169 {
1170         struct kvm *kvm = vcpu->kvm;
1171         struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
1172         unsigned long *pending_bits;
1173         unsigned long flags;
1174         bool ret = true;
1175         int idx, i;
1176
1177         idx = srcu_read_lock(&kvm->srcu);
1178         read_lock_irqsave(&gpc->lock, flags);
1179         if (!kvm_gpc_check(gpc, PAGE_SIZE))
1180                 goto out_rcu;
1181
1182         ret = false;
1183         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
1184                 struct shared_info *shinfo = gpc->khva;
1185                 pending_bits = (unsigned long *)&shinfo->evtchn_pending;
1186         } else {
1187                 struct compat_shared_info *shinfo = gpc->khva;
1188                 pending_bits = (unsigned long *)&shinfo->evtchn_pending;
1189         }
1190
1191         for (i = 0; i < nr_ports; i++) {
1192                 if (test_bit(ports[i], pending_bits)) {
1193                         ret = true;
1194                         break;
1195                 }
1196         }
1197
1198  out_rcu:
1199         read_unlock_irqrestore(&gpc->lock, flags);
1200         srcu_read_unlock(&kvm->srcu, idx);
1201
1202         return ret;
1203 }
1204
1205 static bool kvm_xen_schedop_poll(struct kvm_vcpu *vcpu, bool longmode,
1206                                  u64 param, u64 *r)
1207 {
1208         struct sched_poll sched_poll;
1209         evtchn_port_t port, *ports;
1210         struct x86_exception e;
1211         int i;
1212
1213         if (!lapic_in_kernel(vcpu) ||
1214             !(vcpu->kvm->arch.xen_hvm_config.flags & KVM_XEN_HVM_CONFIG_EVTCHN_SEND))
1215                 return false;
1216
1217         if (IS_ENABLED(CONFIG_64BIT) && !longmode) {
1218                 struct compat_sched_poll sp32;
1219
1220                 /* Sanity check that the compat struct definition is correct */
1221                 BUILD_BUG_ON(sizeof(sp32) != 16);
1222
1223                 if (kvm_read_guest_virt(vcpu, param, &sp32, sizeof(sp32), &e)) {
1224                         *r = -EFAULT;
1225                         return true;
1226                 }
1227
1228                 /*
1229                  * This is a 32-bit pointer to an array of evtchn_port_t which
1230                  * are uint32_t, so once it's converted no further compat
1231                  * handling is needed.
1232                  */
1233                 sched_poll.ports = (void *)(unsigned long)(sp32.ports);
1234                 sched_poll.nr_ports = sp32.nr_ports;
1235                 sched_poll.timeout = sp32.timeout;
1236         } else {
1237                 if (kvm_read_guest_virt(vcpu, param, &sched_poll,
1238                                         sizeof(sched_poll), &e)) {
1239                         *r = -EFAULT;
1240                         return true;
1241                 }
1242         }
1243
1244         if (unlikely(sched_poll.nr_ports > 1)) {
1245                 /* Xen (unofficially) limits number of pollers to 128 */
1246                 if (sched_poll.nr_ports > 128) {
1247                         *r = -EINVAL;
1248                         return true;
1249                 }
1250
1251                 ports = kmalloc_array(sched_poll.nr_ports,
1252                                       sizeof(*ports), GFP_KERNEL);
1253                 if (!ports) {
1254                         *r = -ENOMEM;
1255                         return true;
1256                 }
1257         } else
1258                 ports = &port;
1259
1260         if (kvm_read_guest_virt(vcpu, (gva_t)sched_poll.ports, ports,
1261                                 sched_poll.nr_ports * sizeof(*ports), &e)) {
1262                 *r = -EFAULT;
1263                 return true;
1264         }
1265
1266         for (i = 0; i < sched_poll.nr_ports; i++) {
1267                 if (ports[i] >= max_evtchn_port(vcpu->kvm)) {
1268                         *r = -EINVAL;
1269                         goto out;
1270                 }
1271         }
1272
1273         if (sched_poll.nr_ports == 1)
1274                 vcpu->arch.xen.poll_evtchn = port;
1275         else
1276                 vcpu->arch.xen.poll_evtchn = -1;
1277
1278         set_bit(vcpu->vcpu_idx, vcpu->kvm->arch.xen.poll_mask);
1279
1280         if (!wait_pending_event(vcpu, sched_poll.nr_ports, ports)) {
1281                 vcpu->arch.mp_state = KVM_MP_STATE_HALTED;
1282
1283                 if (sched_poll.timeout)
1284                         mod_timer(&vcpu->arch.xen.poll_timer,
1285                                   jiffies + nsecs_to_jiffies(sched_poll.timeout));
1286
1287                 kvm_vcpu_halt(vcpu);
1288
1289                 if (sched_poll.timeout)
1290                         del_timer(&vcpu->arch.xen.poll_timer);
1291
1292                 vcpu->arch.mp_state = KVM_MP_STATE_RUNNABLE;
1293         }
1294
1295         vcpu->arch.xen.poll_evtchn = 0;
1296         *r = 0;
1297 out:
1298         /* Really, this is only needed in case of timeout */
1299         clear_bit(vcpu->vcpu_idx, vcpu->kvm->arch.xen.poll_mask);
1300
1301         if (unlikely(sched_poll.nr_ports > 1))
1302                 kfree(ports);
1303         return true;
1304 }
1305
1306 static void cancel_evtchn_poll(struct timer_list *t)
1307 {
1308         struct kvm_vcpu *vcpu = from_timer(vcpu, t, arch.xen.poll_timer);
1309
1310         kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
1311         kvm_vcpu_kick(vcpu);
1312 }
1313
1314 static bool kvm_xen_hcall_sched_op(struct kvm_vcpu *vcpu, bool longmode,
1315                                    int cmd, u64 param, u64 *r)
1316 {
1317         switch (cmd) {
1318         case SCHEDOP_poll:
1319                 if (kvm_xen_schedop_poll(vcpu, longmode, param, r))
1320                         return true;
1321                 fallthrough;
1322         case SCHEDOP_yield:
1323                 kvm_vcpu_on_spin(vcpu, true);
1324                 *r = 0;
1325                 return true;
1326         default:
1327                 break;
1328         }
1329
1330         return false;
1331 }
1332
1333 struct compat_vcpu_set_singleshot_timer {
1334     uint64_t timeout_abs_ns;
1335     uint32_t flags;
1336 } __attribute__((packed));
1337
1338 static bool kvm_xen_hcall_vcpu_op(struct kvm_vcpu *vcpu, bool longmode, int cmd,
1339                                   int vcpu_id, u64 param, u64 *r)
1340 {
1341         struct vcpu_set_singleshot_timer oneshot;
1342         struct x86_exception e;
1343         s64 delta;
1344
1345         if (!kvm_xen_timer_enabled(vcpu))
1346                 return false;
1347
1348         switch (cmd) {
1349         case VCPUOP_set_singleshot_timer:
1350                 if (vcpu->arch.xen.vcpu_id != vcpu_id) {
1351                         *r = -EINVAL;
1352                         return true;
1353                 }
1354
1355                 /*
1356                  * The only difference for 32-bit compat is the 4 bytes of
1357                  * padding after the interesting part of the structure. So
1358                  * for a faithful emulation of Xen we have to *try* to copy
1359                  * the padding and return -EFAULT if we can't. Otherwise we
1360                  * might as well just have copied the 12-byte 32-bit struct.
1361                  */
1362                 BUILD_BUG_ON(offsetof(struct compat_vcpu_set_singleshot_timer, timeout_abs_ns) !=
1363                              offsetof(struct vcpu_set_singleshot_timer, timeout_abs_ns));
1364                 BUILD_BUG_ON(sizeof_field(struct compat_vcpu_set_singleshot_timer, timeout_abs_ns) !=
1365                              sizeof_field(struct vcpu_set_singleshot_timer, timeout_abs_ns));
1366                 BUILD_BUG_ON(offsetof(struct compat_vcpu_set_singleshot_timer, flags) !=
1367                              offsetof(struct vcpu_set_singleshot_timer, flags));
1368                 BUILD_BUG_ON(sizeof_field(struct compat_vcpu_set_singleshot_timer, flags) !=
1369                              sizeof_field(struct vcpu_set_singleshot_timer, flags));
1370
1371                 if (kvm_read_guest_virt(vcpu, param, &oneshot, longmode ? sizeof(oneshot) :
1372                                         sizeof(struct compat_vcpu_set_singleshot_timer), &e)) {
1373                         *r = -EFAULT;
1374                         return true;
1375                 }
1376
1377                 delta = oneshot.timeout_abs_ns - get_kvmclock_ns(vcpu->kvm);
1378                 if ((oneshot.flags & VCPU_SSHOTTMR_future) && delta < 0) {
1379                         *r = -ETIME;
1380                         return true;
1381                 }
1382
1383                 kvm_xen_start_timer(vcpu, oneshot.timeout_abs_ns, delta);
1384                 *r = 0;
1385                 return true;
1386
1387         case VCPUOP_stop_singleshot_timer:
1388                 if (vcpu->arch.xen.vcpu_id != vcpu_id) {
1389                         *r = -EINVAL;
1390                         return true;
1391                 }
1392                 kvm_xen_stop_timer(vcpu);
1393                 *r = 0;
1394                 return true;
1395         }
1396
1397         return false;
1398 }
1399
1400 static bool kvm_xen_hcall_set_timer_op(struct kvm_vcpu *vcpu, uint64_t timeout,
1401                                        u64 *r)
1402 {
1403         if (!kvm_xen_timer_enabled(vcpu))
1404                 return false;
1405
1406         if (timeout) {
1407                 uint64_t guest_now = get_kvmclock_ns(vcpu->kvm);
1408                 int64_t delta = timeout - guest_now;
1409
1410                 /* Xen has a 'Linux workaround' in do_set_timer_op() which
1411                  * checks for negative absolute timeout values (caused by
1412                  * integer overflow), and for values about 13 days in the
1413                  * future (2^50ns) which would be caused by jiffies
1414                  * overflow. For those cases, it sets the timeout 100ms in
1415                  * the future (not *too* soon, since if a guest really did
1416                  * set a long timeout on purpose we don't want to keep
1417                  * churning CPU time by waking it up).
1418                  */
1419                 if (unlikely((int64_t)timeout < 0 ||
1420                              (delta > 0 && (uint32_t) (delta >> 50) != 0))) {
1421                         delta = 100 * NSEC_PER_MSEC;
1422                         timeout = guest_now + delta;
1423                 }
1424
1425                 kvm_xen_start_timer(vcpu, timeout, delta);
1426         } else {
1427                 kvm_xen_stop_timer(vcpu);
1428         }
1429
1430         *r = 0;
1431         return true;
1432 }
1433
1434 int kvm_xen_hypercall(struct kvm_vcpu *vcpu)
1435 {
1436         bool longmode;
1437         u64 input, params[6], r = -ENOSYS;
1438         bool handled = false;
1439         u8 cpl;
1440
1441         input = (u64)kvm_register_read(vcpu, VCPU_REGS_RAX);
1442
1443         /* Hyper-V hypercalls get bit 31 set in EAX */
1444         if ((input & 0x80000000) &&
1445             kvm_hv_hypercall_enabled(vcpu))
1446                 return kvm_hv_hypercall(vcpu);
1447
1448         longmode = is_64_bit_hypercall(vcpu);
1449         if (!longmode) {
1450                 params[0] = (u32)kvm_rbx_read(vcpu);
1451                 params[1] = (u32)kvm_rcx_read(vcpu);
1452                 params[2] = (u32)kvm_rdx_read(vcpu);
1453                 params[3] = (u32)kvm_rsi_read(vcpu);
1454                 params[4] = (u32)kvm_rdi_read(vcpu);
1455                 params[5] = (u32)kvm_rbp_read(vcpu);
1456         }
1457 #ifdef CONFIG_X86_64
1458         else {
1459                 params[0] = (u64)kvm_rdi_read(vcpu);
1460                 params[1] = (u64)kvm_rsi_read(vcpu);
1461                 params[2] = (u64)kvm_rdx_read(vcpu);
1462                 params[3] = (u64)kvm_r10_read(vcpu);
1463                 params[4] = (u64)kvm_r8_read(vcpu);
1464                 params[5] = (u64)kvm_r9_read(vcpu);
1465         }
1466 #endif
1467         cpl = static_call(kvm_x86_get_cpl)(vcpu);
1468         trace_kvm_xen_hypercall(cpl, input, params[0], params[1], params[2],
1469                                 params[3], params[4], params[5]);
1470
1471         /*
1472          * Only allow hypercall acceleration for CPL0. The rare hypercalls that
1473          * are permitted in guest userspace can be handled by the VMM.
1474          */
1475         if (unlikely(cpl > 0))
1476                 goto handle_in_userspace;
1477
1478         switch (input) {
1479         case __HYPERVISOR_xen_version:
1480                 if (params[0] == XENVER_version && vcpu->kvm->arch.xen.xen_version) {
1481                         r = vcpu->kvm->arch.xen.xen_version;
1482                         handled = true;
1483                 }
1484                 break;
1485         case __HYPERVISOR_event_channel_op:
1486                 if (params[0] == EVTCHNOP_send)
1487                         handled = kvm_xen_hcall_evtchn_send(vcpu, params[1], &r);
1488                 break;
1489         case __HYPERVISOR_sched_op:
1490                 handled = kvm_xen_hcall_sched_op(vcpu, longmode, params[0],
1491                                                  params[1], &r);
1492                 break;
1493         case __HYPERVISOR_vcpu_op:
1494                 handled = kvm_xen_hcall_vcpu_op(vcpu, longmode, params[0], params[1],
1495                                                 params[2], &r);
1496                 break;
1497         case __HYPERVISOR_set_timer_op: {
1498                 u64 timeout = params[0];
1499                 /* In 32-bit mode, the 64-bit timeout is in two 32-bit params. */
1500                 if (!longmode)
1501                         timeout |= params[1] << 32;
1502                 handled = kvm_xen_hcall_set_timer_op(vcpu, timeout, &r);
1503                 break;
1504         }
1505         default:
1506                 break;
1507         }
1508
1509         if (handled)
1510                 return kvm_xen_hypercall_set_result(vcpu, r);
1511
1512 handle_in_userspace:
1513         vcpu->run->exit_reason = KVM_EXIT_XEN;
1514         vcpu->run->xen.type = KVM_EXIT_XEN_HCALL;
1515         vcpu->run->xen.u.hcall.longmode = longmode;
1516         vcpu->run->xen.u.hcall.cpl = cpl;
1517         vcpu->run->xen.u.hcall.input = input;
1518         vcpu->run->xen.u.hcall.params[0] = params[0];
1519         vcpu->run->xen.u.hcall.params[1] = params[1];
1520         vcpu->run->xen.u.hcall.params[2] = params[2];
1521         vcpu->run->xen.u.hcall.params[3] = params[3];
1522         vcpu->run->xen.u.hcall.params[4] = params[4];
1523         vcpu->run->xen.u.hcall.params[5] = params[5];
1524         vcpu->arch.xen.hypercall_rip = kvm_get_linear_rip(vcpu);
1525         vcpu->arch.complete_userspace_io =
1526                 kvm_xen_hypercall_complete_userspace;
1527
1528         return 0;
1529 }
1530
1531 static void kvm_xen_check_poller(struct kvm_vcpu *vcpu, int port)
1532 {
1533         int poll_evtchn = vcpu->arch.xen.poll_evtchn;
1534
1535         if ((poll_evtchn == port || poll_evtchn == -1) &&
1536             test_and_clear_bit(vcpu->vcpu_idx, vcpu->kvm->arch.xen.poll_mask)) {
1537                 kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
1538                 kvm_vcpu_kick(vcpu);
1539         }
1540 }
1541
1542 /*
1543  * The return value from this function is propagated to kvm_set_irq() API,
1544  * so it returns:
1545  *  < 0   Interrupt was ignored (masked or not delivered for other reasons)
1546  *  = 0   Interrupt was coalesced (previous irq is still pending)
1547  *  > 0   Number of CPUs interrupt was delivered to
1548  *
1549  * It is also called directly from kvm_arch_set_irq_inatomic(), where the
1550  * only check on its return value is a comparison with -EWOULDBLOCK'.
1551  */
1552 int kvm_xen_set_evtchn_fast(struct kvm_xen_evtchn *xe, struct kvm *kvm)
1553 {
1554         struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
1555         struct kvm_vcpu *vcpu;
1556         unsigned long *pending_bits, *mask_bits;
1557         unsigned long flags;
1558         int port_word_bit;
1559         bool kick_vcpu = false;
1560         int vcpu_idx, idx, rc;
1561
1562         vcpu_idx = READ_ONCE(xe->vcpu_idx);
1563         if (vcpu_idx >= 0)
1564                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
1565         else {
1566                 vcpu = kvm_get_vcpu_by_id(kvm, xe->vcpu_id);
1567                 if (!vcpu)
1568                         return -EINVAL;
1569                 WRITE_ONCE(xe->vcpu_idx, vcpu->vcpu_idx);
1570         }
1571
1572         if (!vcpu->arch.xen.vcpu_info_cache.active)
1573                 return -EINVAL;
1574
1575         if (xe->port >= max_evtchn_port(kvm))
1576                 return -EINVAL;
1577
1578         rc = -EWOULDBLOCK;
1579
1580         idx = srcu_read_lock(&kvm->srcu);
1581
1582         read_lock_irqsave(&gpc->lock, flags);
1583         if (!kvm_gpc_check(gpc, PAGE_SIZE))
1584                 goto out_rcu;
1585
1586         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
1587                 struct shared_info *shinfo = gpc->khva;
1588                 pending_bits = (unsigned long *)&shinfo->evtchn_pending;
1589                 mask_bits = (unsigned long *)&shinfo->evtchn_mask;
1590                 port_word_bit = xe->port / 64;
1591         } else {
1592                 struct compat_shared_info *shinfo = gpc->khva;
1593                 pending_bits = (unsigned long *)&shinfo->evtchn_pending;
1594                 mask_bits = (unsigned long *)&shinfo->evtchn_mask;
1595                 port_word_bit = xe->port / 32;
1596         }
1597
1598         /*
1599          * If this port wasn't already set, and if it isn't masked, then
1600          * we try to set the corresponding bit in the in-kernel shadow of
1601          * evtchn_pending_sel for the target vCPU. And if *that* wasn't
1602          * already set, then we kick the vCPU in question to write to the
1603          * *real* evtchn_pending_sel in its own guest vcpu_info struct.
1604          */
1605         if (test_and_set_bit(xe->port, pending_bits)) {
1606                 rc = 0; /* It was already raised */
1607         } else if (test_bit(xe->port, mask_bits)) {
1608                 rc = -ENOTCONN; /* Masked */
1609                 kvm_xen_check_poller(vcpu, xe->port);
1610         } else {
1611                 rc = 1; /* Delivered to the bitmap in shared_info. */
1612                 /* Now switch to the vCPU's vcpu_info to set the index and pending_sel */
1613                 read_unlock_irqrestore(&gpc->lock, flags);
1614                 gpc = &vcpu->arch.xen.vcpu_info_cache;
1615
1616                 read_lock_irqsave(&gpc->lock, flags);
1617                 if (!kvm_gpc_check(gpc, sizeof(struct vcpu_info))) {
1618                         /*
1619                          * Could not access the vcpu_info. Set the bit in-kernel
1620                          * and prod the vCPU to deliver it for itself.
1621                          */
1622                         if (!test_and_set_bit(port_word_bit, &vcpu->arch.xen.evtchn_pending_sel))
1623                                 kick_vcpu = true;
1624                         goto out_rcu;
1625                 }
1626
1627                 if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
1628                         struct vcpu_info *vcpu_info = gpc->khva;
1629                         if (!test_and_set_bit(port_word_bit, &vcpu_info->evtchn_pending_sel)) {
1630                                 WRITE_ONCE(vcpu_info->evtchn_upcall_pending, 1);
1631                                 kick_vcpu = true;
1632                         }
1633                 } else {
1634                         struct compat_vcpu_info *vcpu_info = gpc->khva;
1635                         if (!test_and_set_bit(port_word_bit,
1636                                               (unsigned long *)&vcpu_info->evtchn_pending_sel)) {
1637                                 WRITE_ONCE(vcpu_info->evtchn_upcall_pending, 1);
1638                                 kick_vcpu = true;
1639                         }
1640                 }
1641
1642                 /* For the per-vCPU lapic vector, deliver it as MSI. */
1643                 if (kick_vcpu && vcpu->arch.xen.upcall_vector) {
1644                         kvm_xen_inject_vcpu_vector(vcpu);
1645                         kick_vcpu = false;
1646                 }
1647         }
1648
1649  out_rcu:
1650         read_unlock_irqrestore(&gpc->lock, flags);
1651         srcu_read_unlock(&kvm->srcu, idx);
1652
1653         if (kick_vcpu) {
1654                 kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
1655                 kvm_vcpu_kick(vcpu);
1656         }
1657
1658         return rc;
1659 }
1660
1661 static int kvm_xen_set_evtchn(struct kvm_xen_evtchn *xe, struct kvm *kvm)
1662 {
1663         bool mm_borrowed = false;
1664         int rc;
1665
1666         rc = kvm_xen_set_evtchn_fast(xe, kvm);
1667         if (rc != -EWOULDBLOCK)
1668                 return rc;
1669
1670         if (current->mm != kvm->mm) {
1671                 /*
1672                  * If not on a thread which already belongs to this KVM,
1673                  * we'd better be in the irqfd workqueue.
1674                  */
1675                 if (WARN_ON_ONCE(current->mm))
1676                         return -EINVAL;
1677
1678                 kthread_use_mm(kvm->mm);
1679                 mm_borrowed = true;
1680         }
1681
1682         mutex_lock(&kvm->arch.xen.xen_lock);
1683
1684         /*
1685          * It is theoretically possible for the page to be unmapped
1686          * and the MMU notifier to invalidate the shared_info before
1687          * we even get to use it. In that case, this looks like an
1688          * infinite loop. It was tempting to do it via the userspace
1689          * HVA instead... but that just *hides* the fact that it's
1690          * an infinite loop, because if a fault occurs and it waits
1691          * for the page to come back, it can *still* immediately
1692          * fault and have to wait again, repeatedly.
1693          *
1694          * Conversely, the page could also have been reinstated by
1695          * another thread before we even obtain the mutex above, so
1696          * check again *first* before remapping it.
1697          */
1698         do {
1699                 struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
1700                 int idx;
1701
1702                 rc = kvm_xen_set_evtchn_fast(xe, kvm);
1703                 if (rc != -EWOULDBLOCK)
1704                         break;
1705
1706                 idx = srcu_read_lock(&kvm->srcu);
1707                 rc = kvm_gpc_refresh(gpc, PAGE_SIZE);
1708                 srcu_read_unlock(&kvm->srcu, idx);
1709         } while(!rc);
1710
1711         mutex_unlock(&kvm->arch.xen.xen_lock);
1712
1713         if (mm_borrowed)
1714                 kthread_unuse_mm(kvm->mm);
1715
1716         return rc;
1717 }
1718
1719 /* This is the version called from kvm_set_irq() as the .set function */
1720 static int evtchn_set_fn(struct kvm_kernel_irq_routing_entry *e, struct kvm *kvm,
1721                          int irq_source_id, int level, bool line_status)
1722 {
1723         if (!level)
1724                 return -EINVAL;
1725
1726         return kvm_xen_set_evtchn(&e->xen_evtchn, kvm);
1727 }
1728
1729 /*
1730  * Set up an event channel interrupt from the KVM IRQ routing table.
1731  * Used for e.g. PIRQ from passed through physical devices.
1732  */
1733 int kvm_xen_setup_evtchn(struct kvm *kvm,
1734                          struct kvm_kernel_irq_routing_entry *e,
1735                          const struct kvm_irq_routing_entry *ue)
1736
1737 {
1738         struct kvm_vcpu *vcpu;
1739
1740         if (ue->u.xen_evtchn.port >= max_evtchn_port(kvm))
1741                 return -EINVAL;
1742
1743         /* We only support 2 level event channels for now */
1744         if (ue->u.xen_evtchn.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1745                 return -EINVAL;
1746
1747         /*
1748          * Xen gives us interesting mappings from vCPU index to APIC ID,
1749          * which means kvm_get_vcpu_by_id() has to iterate over all vCPUs
1750          * to find it. Do that once at setup time, instead of every time.
1751          * But beware that on live update / live migration, the routing
1752          * table might be reinstated before the vCPU threads have finished
1753          * recreating their vCPUs.
1754          */
1755         vcpu = kvm_get_vcpu_by_id(kvm, ue->u.xen_evtchn.vcpu);
1756         if (vcpu)
1757                 e->xen_evtchn.vcpu_idx = vcpu->vcpu_idx;
1758         else
1759                 e->xen_evtchn.vcpu_idx = -1;
1760
1761         e->xen_evtchn.port = ue->u.xen_evtchn.port;
1762         e->xen_evtchn.vcpu_id = ue->u.xen_evtchn.vcpu;
1763         e->xen_evtchn.priority = ue->u.xen_evtchn.priority;
1764         e->set = evtchn_set_fn;
1765
1766         return 0;
1767 }
1768
1769 /*
1770  * Explicit event sending from userspace with KVM_XEN_HVM_EVTCHN_SEND ioctl.
1771  */
1772 int kvm_xen_hvm_evtchn_send(struct kvm *kvm, struct kvm_irq_routing_xen_evtchn *uxe)
1773 {
1774         struct kvm_xen_evtchn e;
1775         int ret;
1776
1777         if (!uxe->port || uxe->port >= max_evtchn_port(kvm))
1778                 return -EINVAL;
1779
1780         /* We only support 2 level event channels for now */
1781         if (uxe->priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1782                 return -EINVAL;
1783
1784         e.port = uxe->port;
1785         e.vcpu_id = uxe->vcpu;
1786         e.vcpu_idx = -1;
1787         e.priority = uxe->priority;
1788
1789         ret = kvm_xen_set_evtchn(&e, kvm);
1790
1791         /*
1792          * None of that 'return 1 if it actually got delivered' nonsense.
1793          * We don't care if it was masked (-ENOTCONN) either.
1794          */
1795         if (ret > 0 || ret == -ENOTCONN)
1796                 ret = 0;
1797
1798         return ret;
1799 }
1800
1801 /*
1802  * Support for *outbound* event channel events via the EVTCHNOP_send hypercall.
1803  */
1804 struct evtchnfd {
1805         u32 send_port;
1806         u32 type;
1807         union {
1808                 struct kvm_xen_evtchn port;
1809                 struct {
1810                         u32 port; /* zero */
1811                         struct eventfd_ctx *ctx;
1812                 } eventfd;
1813         } deliver;
1814 };
1815
1816 /*
1817  * Update target vCPU or priority for a registered sending channel.
1818  */
1819 static int kvm_xen_eventfd_update(struct kvm *kvm,
1820                                   struct kvm_xen_hvm_attr *data)
1821 {
1822         u32 port = data->u.evtchn.send_port;
1823         struct evtchnfd *evtchnfd;
1824         int ret;
1825
1826         /* Protect writes to evtchnfd as well as the idr lookup.  */
1827         mutex_lock(&kvm->arch.xen.xen_lock);
1828         evtchnfd = idr_find(&kvm->arch.xen.evtchn_ports, port);
1829
1830         ret = -ENOENT;
1831         if (!evtchnfd)
1832                 goto out_unlock;
1833
1834         /* For an UPDATE, nothing may change except the priority/vcpu */
1835         ret = -EINVAL;
1836         if (evtchnfd->type != data->u.evtchn.type)
1837                 goto out_unlock;
1838
1839         /*
1840          * Port cannot change, and if it's zero that was an eventfd
1841          * which can't be changed either.
1842          */
1843         if (!evtchnfd->deliver.port.port ||
1844             evtchnfd->deliver.port.port != data->u.evtchn.deliver.port.port)
1845                 goto out_unlock;
1846
1847         /* We only support 2 level event channels for now */
1848         if (data->u.evtchn.deliver.port.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1849                 goto out_unlock;
1850
1851         evtchnfd->deliver.port.priority = data->u.evtchn.deliver.port.priority;
1852         if (evtchnfd->deliver.port.vcpu_id != data->u.evtchn.deliver.port.vcpu) {
1853                 evtchnfd->deliver.port.vcpu_id = data->u.evtchn.deliver.port.vcpu;
1854                 evtchnfd->deliver.port.vcpu_idx = -1;
1855         }
1856         ret = 0;
1857 out_unlock:
1858         mutex_unlock(&kvm->arch.xen.xen_lock);
1859         return ret;
1860 }
1861
1862 /*
1863  * Configure the target (eventfd or local port delivery) for sending on
1864  * a given event channel.
1865  */
1866 static int kvm_xen_eventfd_assign(struct kvm *kvm,
1867                                   struct kvm_xen_hvm_attr *data)
1868 {
1869         u32 port = data->u.evtchn.send_port;
1870         struct eventfd_ctx *eventfd = NULL;
1871         struct evtchnfd *evtchnfd;
1872         int ret = -EINVAL;
1873
1874         evtchnfd = kzalloc(sizeof(struct evtchnfd), GFP_KERNEL);
1875         if (!evtchnfd)
1876                 return -ENOMEM;
1877
1878         switch(data->u.evtchn.type) {
1879         case EVTCHNSTAT_ipi:
1880                 /* IPI  must map back to the same port# */
1881                 if (data->u.evtchn.deliver.port.port != data->u.evtchn.send_port)
1882                         goto out_noeventfd; /* -EINVAL */
1883                 break;
1884
1885         case EVTCHNSTAT_interdomain:
1886                 if (data->u.evtchn.deliver.port.port) {
1887                         if (data->u.evtchn.deliver.port.port >= max_evtchn_port(kvm))
1888                                 goto out_noeventfd; /* -EINVAL */
1889                 } else {
1890                         eventfd = eventfd_ctx_fdget(data->u.evtchn.deliver.eventfd.fd);
1891                         if (IS_ERR(eventfd)) {
1892                                 ret = PTR_ERR(eventfd);
1893                                 goto out_noeventfd;
1894                         }
1895                 }
1896                 break;
1897
1898         case EVTCHNSTAT_virq:
1899         case EVTCHNSTAT_closed:
1900         case EVTCHNSTAT_unbound:
1901         case EVTCHNSTAT_pirq:
1902         default: /* Unknown event channel type */
1903                 goto out; /* -EINVAL */
1904         }
1905
1906         evtchnfd->send_port = data->u.evtchn.send_port;
1907         evtchnfd->type = data->u.evtchn.type;
1908         if (eventfd) {
1909                 evtchnfd->deliver.eventfd.ctx = eventfd;
1910         } else {
1911                 /* We only support 2 level event channels for now */
1912                 if (data->u.evtchn.deliver.port.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1913                         goto out; /* -EINVAL; */
1914
1915                 evtchnfd->deliver.port.port = data->u.evtchn.deliver.port.port;
1916                 evtchnfd->deliver.port.vcpu_id = data->u.evtchn.deliver.port.vcpu;
1917                 evtchnfd->deliver.port.vcpu_idx = -1;
1918                 evtchnfd->deliver.port.priority = data->u.evtchn.deliver.port.priority;
1919         }
1920
1921         mutex_lock(&kvm->arch.xen.xen_lock);
1922         ret = idr_alloc(&kvm->arch.xen.evtchn_ports, evtchnfd, port, port + 1,
1923                         GFP_KERNEL);
1924         mutex_unlock(&kvm->arch.xen.xen_lock);
1925         if (ret >= 0)
1926                 return 0;
1927
1928         if (ret == -ENOSPC)
1929                 ret = -EEXIST;
1930 out:
1931         if (eventfd)
1932                 eventfd_ctx_put(eventfd);
1933 out_noeventfd:
1934         kfree(evtchnfd);
1935         return ret;
1936 }
1937
1938 static int kvm_xen_eventfd_deassign(struct kvm *kvm, u32 port)
1939 {
1940         struct evtchnfd *evtchnfd;
1941
1942         mutex_lock(&kvm->arch.xen.xen_lock);
1943         evtchnfd = idr_remove(&kvm->arch.xen.evtchn_ports, port);
1944         mutex_unlock(&kvm->arch.xen.xen_lock);
1945
1946         if (!evtchnfd)
1947                 return -ENOENT;
1948
1949         synchronize_srcu(&kvm->srcu);
1950         if (!evtchnfd->deliver.port.port)
1951                 eventfd_ctx_put(evtchnfd->deliver.eventfd.ctx);
1952         kfree(evtchnfd);
1953         return 0;
1954 }
1955
1956 static int kvm_xen_eventfd_reset(struct kvm *kvm)
1957 {
1958         struct evtchnfd *evtchnfd, **all_evtchnfds;
1959         int i;
1960         int n = 0;
1961
1962         mutex_lock(&kvm->arch.xen.xen_lock);
1963
1964         /*
1965          * Because synchronize_srcu() cannot be called inside the
1966          * critical section, first collect all the evtchnfd objects
1967          * in an array as they are removed from evtchn_ports.
1968          */
1969         idr_for_each_entry(&kvm->arch.xen.evtchn_ports, evtchnfd, i)
1970                 n++;
1971
1972         all_evtchnfds = kmalloc_array(n, sizeof(struct evtchnfd *), GFP_KERNEL);
1973         if (!all_evtchnfds) {
1974                 mutex_unlock(&kvm->arch.xen.xen_lock);
1975                 return -ENOMEM;
1976         }
1977
1978         n = 0;
1979         idr_for_each_entry(&kvm->arch.xen.evtchn_ports, evtchnfd, i) {
1980                 all_evtchnfds[n++] = evtchnfd;
1981                 idr_remove(&kvm->arch.xen.evtchn_ports, evtchnfd->send_port);
1982         }
1983         mutex_unlock(&kvm->arch.xen.xen_lock);
1984
1985         synchronize_srcu(&kvm->srcu);
1986
1987         while (n--) {
1988                 evtchnfd = all_evtchnfds[n];
1989                 if (!evtchnfd->deliver.port.port)
1990                         eventfd_ctx_put(evtchnfd->deliver.eventfd.ctx);
1991                 kfree(evtchnfd);
1992         }
1993         kfree(all_evtchnfds);
1994
1995         return 0;
1996 }
1997
1998 static int kvm_xen_setattr_evtchn(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
1999 {
2000         u32 port = data->u.evtchn.send_port;
2001
2002         if (data->u.evtchn.flags == KVM_XEN_EVTCHN_RESET)
2003                 return kvm_xen_eventfd_reset(kvm);
2004
2005         if (!port || port >= max_evtchn_port(kvm))
2006                 return -EINVAL;
2007
2008         if (data->u.evtchn.flags == KVM_XEN_EVTCHN_DEASSIGN)
2009                 return kvm_xen_eventfd_deassign(kvm, port);
2010         if (data->u.evtchn.flags == KVM_XEN_EVTCHN_UPDATE)
2011                 return kvm_xen_eventfd_update(kvm, data);
2012         if (data->u.evtchn.flags)
2013                 return -EINVAL;
2014
2015         return kvm_xen_eventfd_assign(kvm, data);
2016 }
2017
2018 static bool kvm_xen_hcall_evtchn_send(struct kvm_vcpu *vcpu, u64 param, u64 *r)
2019 {
2020         struct evtchnfd *evtchnfd;
2021         struct evtchn_send send;
2022         struct x86_exception e;
2023
2024         /* Sanity check: this structure is the same for 32-bit and 64-bit */
2025         BUILD_BUG_ON(sizeof(send) != 4);
2026         if (kvm_read_guest_virt(vcpu, param, &send, sizeof(send), &e)) {
2027                 *r = -EFAULT;
2028                 return true;
2029         }
2030
2031         /*
2032          * evtchnfd is protected by kvm->srcu; the idr lookup instead
2033          * is protected by RCU.
2034          */
2035         rcu_read_lock();
2036         evtchnfd = idr_find(&vcpu->kvm->arch.xen.evtchn_ports, send.port);
2037         rcu_read_unlock();
2038         if (!evtchnfd)
2039                 return false;
2040
2041         if (evtchnfd->deliver.port.port) {
2042                 int ret = kvm_xen_set_evtchn(&evtchnfd->deliver.port, vcpu->kvm);
2043                 if (ret < 0 && ret != -ENOTCONN)
2044                         return false;
2045         } else {
2046                 eventfd_signal(evtchnfd->deliver.eventfd.ctx, 1);
2047         }
2048
2049         *r = 0;
2050         return true;
2051 }
2052
2053 void kvm_xen_init_vcpu(struct kvm_vcpu *vcpu)
2054 {
2055         vcpu->arch.xen.vcpu_id = vcpu->vcpu_idx;
2056         vcpu->arch.xen.poll_evtchn = 0;
2057
2058         timer_setup(&vcpu->arch.xen.poll_timer, cancel_evtchn_poll, 0);
2059
2060         kvm_gpc_init(&vcpu->arch.xen.runstate_cache, vcpu->kvm, NULL,
2061                      KVM_HOST_USES_PFN);
2062         kvm_gpc_init(&vcpu->arch.xen.runstate2_cache, vcpu->kvm, NULL,
2063                      KVM_HOST_USES_PFN);
2064         kvm_gpc_init(&vcpu->arch.xen.vcpu_info_cache, vcpu->kvm, NULL,
2065                      KVM_HOST_USES_PFN);
2066         kvm_gpc_init(&vcpu->arch.xen.vcpu_time_info_cache, vcpu->kvm, NULL,
2067                      KVM_HOST_USES_PFN);
2068 }
2069
2070 void kvm_xen_destroy_vcpu(struct kvm_vcpu *vcpu)
2071 {
2072         if (kvm_xen_timer_enabled(vcpu))
2073                 kvm_xen_stop_timer(vcpu);
2074
2075         kvm_gpc_deactivate(&vcpu->arch.xen.runstate_cache);
2076         kvm_gpc_deactivate(&vcpu->arch.xen.runstate2_cache);
2077         kvm_gpc_deactivate(&vcpu->arch.xen.vcpu_info_cache);
2078         kvm_gpc_deactivate(&vcpu->arch.xen.vcpu_time_info_cache);
2079
2080         del_timer_sync(&vcpu->arch.xen.poll_timer);
2081 }
2082
2083 void kvm_xen_update_tsc_info(struct kvm_vcpu *vcpu)
2084 {
2085         struct kvm_cpuid_entry2 *entry;
2086         u32 function;
2087
2088         if (!vcpu->arch.xen.cpuid.base)
2089                 return;
2090
2091         function = vcpu->arch.xen.cpuid.base | XEN_CPUID_LEAF(3);
2092         if (function > vcpu->arch.xen.cpuid.limit)
2093                 return;
2094
2095         entry = kvm_find_cpuid_entry_index(vcpu, function, 1);
2096         if (entry) {
2097                 entry->ecx = vcpu->arch.hv_clock.tsc_to_system_mul;
2098                 entry->edx = vcpu->arch.hv_clock.tsc_shift;
2099         }
2100
2101         entry = kvm_find_cpuid_entry_index(vcpu, function, 2);
2102         if (entry)
2103                 entry->eax = vcpu->arch.hw_tsc_khz;
2104 }
2105
2106 void kvm_xen_init_vm(struct kvm *kvm)
2107 {
2108         mutex_init(&kvm->arch.xen.xen_lock);
2109         idr_init(&kvm->arch.xen.evtchn_ports);
2110         kvm_gpc_init(&kvm->arch.xen.shinfo_cache, kvm, NULL, KVM_HOST_USES_PFN);
2111 }
2112
2113 void kvm_xen_destroy_vm(struct kvm *kvm)
2114 {
2115         struct evtchnfd *evtchnfd;
2116         int i;
2117
2118         kvm_gpc_deactivate(&kvm->arch.xen.shinfo_cache);
2119
2120         idr_for_each_entry(&kvm->arch.xen.evtchn_ports, evtchnfd, i) {
2121                 if (!evtchnfd->deliver.port.port)
2122                         eventfd_ctx_put(evtchnfd->deliver.eventfd.ctx);
2123                 kfree(evtchnfd);
2124         }
2125         idr_destroy(&kvm->arch.xen.evtchn_ports);
2126
2127         if (kvm->arch.xen_hvm_config.msr)
2128                 static_branch_slow_dec_deferred(&kvm_xen_enabled);
2129 }