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
   9 #include "x86.h"
  10 #include "xen.h"
  11 #include "lapic.h"
  12 #include "hyperv.h"
  13
  14 #include <linux/eventfd.h>
  15 #include <linux/kvm_host.h>
  16 #include <linux/sched/stat.h>
  17
  18 #include <trace/events/kvm.h>
  19 #include <xen/interface/xen.h>
  20 #include <xen/interface/vcpu.h>
  21 #include <xen/interface/event_channel.h>
  22 #include <xen/interface/sched.h>
  23
  24 #include "trace.h"
  25
  26 static int kvm_xen_set_evtchn(struct kvm_xen_evtchn *xe, struct kvm *kvm);
  27 static int kvm_xen_setattr_evtchn(struct kvm *kvm, struct kvm_xen_hvm_attr *data);
  28 static bool kvm_xen_hcall_evtchn_send(struct kvm_vcpu *vcpu, u64 param, u64 *r);
  29
  30 DEFINE_STATIC_KEY_DEFERRED_FALSE(kvm_xen_enabled, HZ);
  31
  32 static int kvm_xen_shared_info_init(struct kvm *kvm, gfn_t gfn)
  33 {
  34         struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
  35         struct pvclock_wall_clock *wc;
  36         gpa_t gpa = gfn_to_gpa(gfn);
  37         u32 *wc_sec_hi;
  38         u32 wc_version;
  39         u64 wall_nsec;
  40         int ret = 0;
  41         int idx = srcu_read_lock(&kvm->srcu);
  42
  43         if (gfn == GPA_INVALID) {
  44                 kvm_gfn_to_pfn_cache_destroy(kvm, gpc);
  45                 goto out;
  46         }
  47
  48         do {
  49                 ret = kvm_gfn_to_pfn_cache_init(kvm, gpc, NULL, KVM_HOST_USES_PFN,
  50                                                 gpa, PAGE_SIZE);
  51                 if (ret)
  52                         goto out;
  53
  54                 /*
  55                  * This code mirrors kvm_write_wall_clock() except that it writes
  56                  * directly through the pfn cache and doesn't mark the page dirty.
  57                  */
  58                 wall_nsec = ktime_get_real_ns() - get_kvmclock_ns(kvm);
  59
  60                 /* It could be invalid again already, so we need to check */
  61                 read_lock_irq(&gpc->lock);
  62
  63                 if (gpc->valid)
  64                         break;
  65
  66                 read_unlock_irq(&gpc->lock);
  67         } while (1);
  68
  69         /* Paranoia checks on the 32-bit struct layout */
  70         BUILD_BUG_ON(offsetof(struct compat_shared_info, wc) != 0x900);
  71         BUILD_BUG_ON(offsetof(struct compat_shared_info, arch.wc_sec_hi) != 0x924);
  72         BUILD_BUG_ON(offsetof(struct pvclock_vcpu_time_info, version) != 0);
  73
  74 #ifdef CONFIG_X86_64
  75         /* Paranoia checks on the 64-bit struct layout */
  76         BUILD_BUG_ON(offsetof(struct shared_info, wc) != 0xc00);
  77         BUILD_BUG_ON(offsetof(struct shared_info, wc_sec_hi) != 0xc0c);
  78
  79         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
  80                 struct shared_info *shinfo = gpc->khva;
  81
  82                 wc_sec_hi = &shinfo->wc_sec_hi;
  83                 wc = &shinfo->wc;
  84         } else
  85 #endif
  86         {
  87                 struct compat_shared_info *shinfo = gpc->khva;
  88
  89                 wc_sec_hi = &shinfo->arch.wc_sec_hi;
  90                 wc = &shinfo->wc;
  91         }
  92
  93         /* Increment and ensure an odd value */
  94         wc_version = wc->version = (wc->version + 1) | 1;
  95         smp_wmb();
  96
  97         wc->nsec = do_div(wall_nsec,  1000000000);
  98         wc->sec = (u32)wall_nsec;
  99         *wc_sec_hi = wall_nsec >> 32;
 100         smp_wmb();
 101
 102         wc->version = wc_version + 1;
 103         read_unlock_irq(&gpc->lock);
 104
 105         kvm_make_all_cpus_request(kvm, KVM_REQ_MASTERCLOCK_UPDATE);
 106
 107 out:
 108         srcu_read_unlock(&kvm->srcu, idx);
 109         return ret;
 110 }
 111
 112 void kvm_xen_inject_timer_irqs(struct kvm_vcpu *vcpu)
 113 {
 114         if (atomic_read(&vcpu->arch.xen.timer_pending) > 0) {
 115                 struct kvm_xen_evtchn e;
 116
 117                 e.vcpu_id = vcpu->vcpu_id;
 118                 e.vcpu_idx = vcpu->vcpu_idx;
 119                 e.port = vcpu->arch.xen.timer_virq;
 120                 e.priority = KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL;
 121
 122                 kvm_xen_set_evtchn(&e, vcpu->kvm);
 123
 124                 vcpu->arch.xen.timer_expires = 0;
 125                 atomic_set(&vcpu->arch.xen.timer_pending, 0);
 126         }
 127 }
 128
 129 static enum hrtimer_restart xen_timer_callback(struct hrtimer *timer)
 130 {
 131         struct kvm_vcpu *vcpu = container_of(timer, struct kvm_vcpu,
 132                                              arch.xen.timer);
 133         if (atomic_read(&vcpu->arch.xen.timer_pending))
 134                 return HRTIMER_NORESTART;
 135
 136         atomic_inc(&vcpu->arch.xen.timer_pending);
 137         kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
 138         kvm_vcpu_kick(vcpu);
 139
 140         return HRTIMER_NORESTART;
 141 }
 142
 143 static void kvm_xen_start_timer(struct kvm_vcpu *vcpu, u64 guest_abs, s64 delta_ns)
 144 {
 145         atomic_set(&vcpu->arch.xen.timer_pending, 0);
 146         vcpu->arch.xen.timer_expires = guest_abs;
 147
 148         if (delta_ns <= 0) {
 149                 xen_timer_callback(&vcpu->arch.xen.timer);
 150         } else {
 151                 ktime_t ktime_now = ktime_get();
 152                 hrtimer_start(&vcpu->arch.xen.timer,
 153                               ktime_add_ns(ktime_now, delta_ns),
 154                               HRTIMER_MODE_ABS_HARD);
 155         }
 156 }
 157
 158 static void kvm_xen_stop_timer(struct kvm_vcpu *vcpu)
 159 {
 160         hrtimer_cancel(&vcpu->arch.xen.timer);
 161         vcpu->arch.xen.timer_expires = 0;
 162         atomic_set(&vcpu->arch.xen.timer_pending, 0);
 163 }
 164
 165 static void kvm_xen_init_timer(struct kvm_vcpu *vcpu)
 166 {
 167         hrtimer_init(&vcpu->arch.xen.timer, CLOCK_MONOTONIC,
 168                      HRTIMER_MODE_ABS_HARD);
 169         vcpu->arch.xen.timer.function = xen_timer_callback;
 170 }
 171
 172 static void kvm_xen_update_runstate(struct kvm_vcpu *v, int state)
 173 {
 174         struct kvm_vcpu_xen *vx = &v->arch.xen;
 175         u64 now = get_kvmclock_ns(v->kvm);
 176         u64 delta_ns = now - vx->runstate_entry_time;
 177         u64 run_delay = current->sched_info.run_delay;
 178
 179         if (unlikely(!vx->runstate_entry_time))
 180                 vx->current_runstate = RUNSTATE_offline;
 181
 182         /*
 183          * Time waiting for the scheduler isn't "stolen" if the
 184          * vCPU wasn't running anyway.
 185          */
 186         if (vx->current_runstate == RUNSTATE_running) {
 187                 u64 steal_ns = run_delay - vx->last_steal;
 188
 189                 delta_ns -= steal_ns;
 190
 191                 vx->runstate_times[RUNSTATE_runnable] += steal_ns;
 192         }
 193         vx->last_steal = run_delay;
 194
 195         vx->runstate_times[vx->current_runstate] += delta_ns;
 196         vx->current_runstate = state;
 197         vx->runstate_entry_time = now;
 198 }
 199
 200 void kvm_xen_update_runstate_guest(struct kvm_vcpu *v, int state)
 201 {
 202         struct kvm_vcpu_xen *vx = &v->arch.xen;
 203         struct gfn_to_pfn_cache *gpc = &vx->runstate_cache;
 204         uint64_t *user_times;
 205         unsigned long flags;
 206         size_t user_len;
 207         int *user_state;
 208
 209         kvm_xen_update_runstate(v, state);
 210
 211         if (!vx->runstate_cache.active)
 212                 return;
 213
 214         if (IS_ENABLED(CONFIG_64BIT) && v->kvm->arch.xen.long_mode)
 215                 user_len = sizeof(struct vcpu_runstate_info);
 216         else
 217                 user_len = sizeof(struct compat_vcpu_runstate_info);
 218
 219         read_lock_irqsave(&gpc->lock, flags);
 220         while (!kvm_gfn_to_pfn_cache_check(v->kvm, gpc, gpc->gpa,
 221                                            user_len)) {
 222                 read_unlock_irqrestore(&gpc->lock, flags);
 223
 224                 /* When invoked from kvm_sched_out() we cannot sleep */
 225                 if (state == RUNSTATE_runnable)
 226                         return;
 227
 228                 if (kvm_gfn_to_pfn_cache_refresh(v->kvm, gpc, gpc->gpa, user_len))
 229                         return;
 230
 231                 read_lock_irqsave(&gpc->lock, flags);
 232         }
 233
 234         /*
 235          * The only difference between 32-bit and 64-bit versions of the
 236          * runstate struct us the alignment of uint64_t in 32-bit, which
 237          * means that the 64-bit version has an additional 4 bytes of
 238          * padding after the first field 'state'.
 239          *
 240          * So we use 'int __user *user_state' to point to the state field,
 241          * and 'uint64_t __user *user_times' for runstate_entry_time. So
 242          * the actual array of time[] in each state starts at user_times[1].
 243          */
 244         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state) != 0);
 245         BUILD_BUG_ON(offsetof(struct compat_vcpu_runstate_info, state) != 0);
 246         BUILD_BUG_ON(sizeof(struct compat_vcpu_runstate_info) != 0x2c);
 247 #ifdef CONFIG_X86_64
 248         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state_entry_time) !=
 249                      offsetof(struct compat_vcpu_runstate_info, state_entry_time) + 4);
 250         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, time) !=
 251                      offsetof(struct compat_vcpu_runstate_info, time) + 4);
 252 #endif
 253
 254         user_state = gpc->khva;
 255
 256         if (IS_ENABLED(CONFIG_64BIT) && v->kvm->arch.xen.long_mode)
 257                 user_times = gpc->khva + offsetof(struct vcpu_runstate_info,
 258                                                   state_entry_time);
 259         else
 260                 user_times = gpc->khva + offsetof(struct compat_vcpu_runstate_info,
 261                                                   state_entry_time);
 262
 263         /*
 264          * First write the updated state_entry_time at the appropriate
 265          * location determined by 'offset'.
 266          */
 267         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, state_entry_time) !=
 268                      sizeof(user_times[0]));
 269         BUILD_BUG_ON(sizeof_field(struct compat_vcpu_runstate_info, state_entry_time) !=
 270                      sizeof(user_times[0]));
 271
 272         user_times[0] = vx->runstate_entry_time | XEN_RUNSTATE_UPDATE;
 273         smp_wmb();
 274
 275         /*
 276          * Next, write the new runstate. This is in the *same* place
 277          * for 32-bit and 64-bit guests, asserted here for paranoia.
 278          */
 279         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state) !=
 280                      offsetof(struct compat_vcpu_runstate_info, state));
 281         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, state) !=
 282                      sizeof(vx->current_runstate));
 283         BUILD_BUG_ON(sizeof_field(struct compat_vcpu_runstate_info, state) !=
 284                      sizeof(vx->current_runstate));
 285
 286         *user_state = vx->current_runstate;
 287
 288         /*
 289          * Write the actual runstate times immediately after the
 290          * runstate_entry_time.
 291          */
 292         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state_entry_time) !=
 293                      offsetof(struct vcpu_runstate_info, time) - sizeof(u64));
 294         BUILD_BUG_ON(offsetof(struct compat_vcpu_runstate_info, state_entry_time) !=
 295                      offsetof(struct compat_vcpu_runstate_info, time) - sizeof(u64));
 296         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, time) !=
 297                      sizeof_field(struct compat_vcpu_runstate_info, time));
 298         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, time) !=
 299                      sizeof(vx->runstate_times));
 300
 301         memcpy(user_times + 1, vx->runstate_times, sizeof(vx->runstate_times));
 302         smp_wmb();
 303
 304         /*
 305          * Finally, clear the XEN_RUNSTATE_UPDATE bit in the guest's
 306          * runstate_entry_time field.
 307          */
 308         user_times[0] &= ~XEN_RUNSTATE_UPDATE;
 309         smp_wmb();
 310
 311         read_unlock_irqrestore(&gpc->lock, flags);
 312
 313         mark_page_dirty_in_slot(v->kvm, gpc->memslot, gpc->gpa >> PAGE_SHIFT);
 314 }
 315
 316 /*
 317  * On event channel delivery, the vcpu_info may not have been accessible.
 318  * In that case, there are bits in vcpu->arch.xen.evtchn_pending_sel which
 319  * need to be marked into the vcpu_info (and evtchn_upcall_pending set).
 320  * Do so now that we can sleep in the context of the vCPU to bring the
 321  * page in, and refresh the pfn cache for it.
 322  */
 323 void kvm_xen_inject_pending_events(struct kvm_vcpu *v)
 324 {
 325         unsigned long evtchn_pending_sel = READ_ONCE(v->arch.xen.evtchn_pending_sel);
 326         struct gfn_to_pfn_cache *gpc = &v->arch.xen.vcpu_info_cache;
 327         unsigned long flags;
 328
 329         if (!evtchn_pending_sel)
 330                 return;
 331
 332         /*
 333          * Yes, this is an open-coded loop. But that's just what put_user()
 334          * does anyway. Page it in and retry the instruction. We're just a
 335          * little more honest about it.
 336          */
 337         read_lock_irqsave(&gpc->lock, flags);
 338         while (!kvm_gfn_to_pfn_cache_check(v->kvm, gpc, gpc->gpa,
 339                                            sizeof(struct vcpu_info))) {
 340                 read_unlock_irqrestore(&gpc->lock, flags);
 341
 342                 if (kvm_gfn_to_pfn_cache_refresh(v->kvm, gpc, gpc->gpa,
 343                                                  sizeof(struct vcpu_info)))
 344                         return;
 345
 346                 read_lock_irqsave(&gpc->lock, flags);
 347         }
 348
 349         /* Now gpc->khva is a valid kernel address for the vcpu_info */
 350         if (IS_ENABLED(CONFIG_64BIT) && v->kvm->arch.xen.long_mode) {
 351                 struct vcpu_info *vi = gpc->khva;
 352
 353                 asm volatile(LOCK_PREFIX "orq %0, %1\n"
 354                              "notq %0\n"
 355                              LOCK_PREFIX "andq %0, %2\n"
 356                              : "=r" (evtchn_pending_sel),
 357                                "+m" (vi->evtchn_pending_sel),
 358                                "+m" (v->arch.xen.evtchn_pending_sel)
 359                              : "0" (evtchn_pending_sel));
 360                 WRITE_ONCE(vi->evtchn_upcall_pending, 1);
 361         } else {
 362                 u32 evtchn_pending_sel32 = evtchn_pending_sel;
 363                 struct compat_vcpu_info *vi = gpc->khva;
 364
 365                 asm volatile(LOCK_PREFIX "orl %0, %1\n"
 366                              "notl %0\n"
 367                              LOCK_PREFIX "andl %0, %2\n"
 368                              : "=r" (evtchn_pending_sel32),
 369                                "+m" (vi->evtchn_pending_sel),
 370                                "+m" (v->arch.xen.evtchn_pending_sel)
 371                              : "0" (evtchn_pending_sel32));
 372                 WRITE_ONCE(vi->evtchn_upcall_pending, 1);
 373         }
 374         read_unlock_irqrestore(&gpc->lock, flags);
 375
 376         mark_page_dirty_in_slot(v->kvm, gpc->memslot, gpc->gpa >> PAGE_SHIFT);
 377 }
 378
 379 int __kvm_xen_has_interrupt(struct kvm_vcpu *v)
 380 {
 381         struct gfn_to_pfn_cache *gpc = &v->arch.xen.vcpu_info_cache;
 382         unsigned long flags;
 383         u8 rc = 0;
 384
 385         /*
 386          * If the global upcall vector (HVMIRQ_callback_vector) is set and
 387          * the vCPU's evtchn_upcall_pending flag is set, the IRQ is pending.
 388          */
 389
 390         /* No need for compat handling here */
 391         BUILD_BUG_ON(offsetof(struct vcpu_info, evtchn_upcall_pending) !=
 392                      offsetof(struct compat_vcpu_info, evtchn_upcall_pending));
 393         BUILD_BUG_ON(sizeof(rc) !=
 394                      sizeof_field(struct vcpu_info, evtchn_upcall_pending));
 395         BUILD_BUG_ON(sizeof(rc) !=
 396                      sizeof_field(struct compat_vcpu_info, evtchn_upcall_pending));
 397
 398         read_lock_irqsave(&gpc->lock, flags);
 399         while (!kvm_gfn_to_pfn_cache_check(v->kvm, gpc, gpc->gpa,
 400                                            sizeof(struct vcpu_info))) {
 401                 read_unlock_irqrestore(&gpc->lock, flags);
 402
 403                 /*
 404                  * This function gets called from kvm_vcpu_block() after setting the
 405                  * task to TASK_INTERRUPTIBLE, to see if it needs to wake immediately
 406                  * from a HLT. So we really mustn't sleep. If the page ended up absent
 407                  * at that point, just return 1 in order to trigger an immediate wake,
 408                  * and we'll end up getting called again from a context where we *can*
 409                  * fault in the page and wait for it.
 410                  */
 411                 if (in_atomic() || !task_is_running(current))
 412                         return 1;
 413
 414                 if (kvm_gfn_to_pfn_cache_refresh(v->kvm, gpc, gpc->gpa,
 415                                                  sizeof(struct vcpu_info))) {
 416                         /*
 417                          * If this failed, userspace has screwed up the
 418                          * vcpu_info mapping. No interrupts for you.
 419                          */
 420                         return 0;
 421                 }
 422                 read_lock_irqsave(&gpc->lock, flags);
 423         }
 424
 425         rc = ((struct vcpu_info *)gpc->khva)->evtchn_upcall_pending;
 426         read_unlock_irqrestore(&gpc->lock, flags);
 427         return rc;
 428 }
 429
 430 int kvm_xen_hvm_set_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 431 {
 432         int r = -ENOENT;
 433
 434
 435         switch (data->type) {
 436         case KVM_XEN_ATTR_TYPE_LONG_MODE:
 437                 if (!IS_ENABLED(CONFIG_64BIT) && data->u.long_mode) {
 438                         r = -EINVAL;
 439                 } else {
 440                         mutex_lock(&kvm->lock);
 441                         kvm->arch.xen.long_mode = !!data->u.long_mode;
 442                         mutex_unlock(&kvm->lock);
 443                         r = 0;
 444                 }
 445                 break;
 446
 447         case KVM_XEN_ATTR_TYPE_SHARED_INFO:
 448                 mutex_lock(&kvm->lock);
 449                 r = kvm_xen_shared_info_init(kvm, data->u.shared_info.gfn);
 450                 mutex_unlock(&kvm->lock);
 451                 break;
 452
 453         case KVM_XEN_ATTR_TYPE_UPCALL_VECTOR:
 454                 if (data->u.vector && data->u.vector < 0x10)
 455                         r = -EINVAL;
 456                 else {
 457                         mutex_lock(&kvm->lock);
 458                         kvm->arch.xen.upcall_vector = data->u.vector;
 459                         mutex_unlock(&kvm->lock);
 460                         r = 0;
 461                 }
 462                 break;
 463
 464         case KVM_XEN_ATTR_TYPE_EVTCHN:
 465                 r = kvm_xen_setattr_evtchn(kvm, data);
 466                 break;
 467
 468         default:
 469                 break;
 470         }
 471
 472         return r;
 473 }
 474
 475 int kvm_xen_hvm_get_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 476 {
 477         int r = -ENOENT;
 478
 479         mutex_lock(&kvm->lock);
 480
 481         switch (data->type) {
 482         case KVM_XEN_ATTR_TYPE_LONG_MODE:
 483                 data->u.long_mode = kvm->arch.xen.long_mode;
 484                 r = 0;
 485                 break;
 486
 487         case KVM_XEN_ATTR_TYPE_SHARED_INFO:
 488                 if (kvm->arch.xen.shinfo_cache.active)
 489                         data->u.shared_info.gfn = gpa_to_gfn(kvm->arch.xen.shinfo_cache.gpa);
 490                 else
 491                         data->u.shared_info.gfn = GPA_INVALID;
 492                 r = 0;
 493                 break;
 494
 495         case KVM_XEN_ATTR_TYPE_UPCALL_VECTOR:
 496                 data->u.vector = kvm->arch.xen.upcall_vector;
 497                 r = 0;
 498                 break;
 499
 500         default:
 501                 break;
 502         }
 503
 504         mutex_unlock(&kvm->lock);
 505         return r;
 506 }
 507
 508 int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 509 {
 510         int idx, r = -ENOENT;
 511
 512         mutex_lock(&vcpu->kvm->lock);
 513         idx = srcu_read_lock(&vcpu->kvm->srcu);
 514
 515         switch (data->type) {
 516         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
 517                 /* No compat necessary here. */
 518                 BUILD_BUG_ON(sizeof(struct vcpu_info) !=
 519                              sizeof(struct compat_vcpu_info));
 520                 BUILD_BUG_ON(offsetof(struct vcpu_info, time) !=
 521                              offsetof(struct compat_vcpu_info, time));
 522
 523                 if (data->u.gpa == GPA_INVALID) {
 524                         kvm_gfn_to_pfn_cache_destroy(vcpu->kvm, &vcpu->arch.xen.vcpu_info_cache);
 525                         r = 0;
 526                         break;
 527                 }
 528
 529                 r = kvm_gfn_to_pfn_cache_init(vcpu->kvm,
 530                                               &vcpu->arch.xen.vcpu_info_cache,
 531                                               NULL, KVM_HOST_USES_PFN, data->u.gpa,
 532                                               sizeof(struct vcpu_info));
 533                 if (!r)
 534                         kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
 535
 536                 break;
 537
 538         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_TIME_INFO:
 539                 if (data->u.gpa == GPA_INVALID) {
 540                         kvm_gfn_to_pfn_cache_destroy(vcpu->kvm,
 541                                                      &vcpu->arch.xen.vcpu_time_info_cache);
 542                         r = 0;
 543                         break;
 544                 }
 545
 546                 r = kvm_gfn_to_pfn_cache_init(vcpu->kvm,
 547                                               &vcpu->arch.xen.vcpu_time_info_cache,
 548                                               NULL, KVM_HOST_USES_PFN, data->u.gpa,
 549                                               sizeof(struct pvclock_vcpu_time_info));
 550                 if (!r)
 551                         kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
 552                 break;
 553
 554         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADDR:
 555                 if (!sched_info_on()) {
 556                         r = -EOPNOTSUPP;
 557                         break;
 558                 }
 559                 if (data->u.gpa == GPA_INVALID) {
 560                         kvm_gfn_to_pfn_cache_destroy(vcpu->kvm,
 561                                                      &vcpu->arch.xen.runstate_cache);
 562                         r = 0;
 563                         break;
 564                 }
 565
 566                 r = kvm_gfn_to_pfn_cache_init(vcpu->kvm,
 567                                               &vcpu->arch.xen.runstate_cache,
 568                                               NULL, KVM_HOST_USES_PFN, data->u.gpa,
 569                                               sizeof(struct vcpu_runstate_info));
 570                 break;
 571
 572         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_CURRENT:
 573                 if (!sched_info_on()) {
 574                         r = -EOPNOTSUPP;
 575                         break;
 576                 }
 577                 if (data->u.runstate.state > RUNSTATE_offline) {
 578                         r = -EINVAL;
 579                         break;
 580                 }
 581
 582                 kvm_xen_update_runstate(vcpu, data->u.runstate.state);
 583                 r = 0;
 584                 break;
 585
 586         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_DATA:
 587                 if (!sched_info_on()) {
 588                         r = -EOPNOTSUPP;
 589                         break;
 590                 }
 591                 if (data->u.runstate.state > RUNSTATE_offline) {
 592                         r = -EINVAL;
 593                         break;
 594                 }
 595                 if (data->u.runstate.state_entry_time !=
 596                     (data->u.runstate.time_running +
 597                      data->u.runstate.time_runnable +
 598                      data->u.runstate.time_blocked +
 599                      data->u.runstate.time_offline)) {
 600                         r = -EINVAL;
 601                         break;
 602                 }
 603                 if (get_kvmclock_ns(vcpu->kvm) <
 604                     data->u.runstate.state_entry_time) {
 605                         r = -EINVAL;
 606                         break;
 607                 }
 608
 609                 vcpu->arch.xen.current_runstate = data->u.runstate.state;
 610                 vcpu->arch.xen.runstate_entry_time =
 611                         data->u.runstate.state_entry_time;
 612                 vcpu->arch.xen.runstate_times[RUNSTATE_running] =
 613                         data->u.runstate.time_running;
 614                 vcpu->arch.xen.runstate_times[RUNSTATE_runnable] =
 615                         data->u.runstate.time_runnable;
 616                 vcpu->arch.xen.runstate_times[RUNSTATE_blocked] =
 617                         data->u.runstate.time_blocked;
 618                 vcpu->arch.xen.runstate_times[RUNSTATE_offline] =
 619                         data->u.runstate.time_offline;
 620                 vcpu->arch.xen.last_steal = current->sched_info.run_delay;
 621                 r = 0;
 622                 break;
 623
 624         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADJUST:
 625                 if (!sched_info_on()) {
 626                         r = -EOPNOTSUPP;
 627                         break;
 628                 }
 629                 if (data->u.runstate.state > RUNSTATE_offline &&
 630                     data->u.runstate.state != (u64)-1) {
 631                         r = -EINVAL;
 632                         break;
 633                 }
 634                 /* The adjustment must add up */
 635                 if (data->u.runstate.state_entry_time !=
 636                     (data->u.runstate.time_running +
 637                      data->u.runstate.time_runnable +
 638                      data->u.runstate.time_blocked +
 639                      data->u.runstate.time_offline)) {
 640                         r = -EINVAL;
 641                         break;
 642                 }
 643
 644                 if (get_kvmclock_ns(vcpu->kvm) <
 645                     (vcpu->arch.xen.runstate_entry_time +
 646                      data->u.runstate.state_entry_time)) {
 647                         r = -EINVAL;
 648                         break;
 649                 }
 650
 651                 vcpu->arch.xen.runstate_entry_time +=
 652                         data->u.runstate.state_entry_time;
 653                 vcpu->arch.xen.runstate_times[RUNSTATE_running] +=
 654                         data->u.runstate.time_running;
 655                 vcpu->arch.xen.runstate_times[RUNSTATE_runnable] +=
 656                         data->u.runstate.time_runnable;
 657                 vcpu->arch.xen.runstate_times[RUNSTATE_blocked] +=
 658                         data->u.runstate.time_blocked;
 659                 vcpu->arch.xen.runstate_times[RUNSTATE_offline] +=
 660                         data->u.runstate.time_offline;
 661
 662                 if (data->u.runstate.state <= RUNSTATE_offline)
 663                         kvm_xen_update_runstate(vcpu, data->u.runstate.state);
 664                 r = 0;
 665                 break;
 666
 667         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_ID:
 668                 if (data->u.vcpu_id >= KVM_MAX_VCPUS)
 669                         r = -EINVAL;
 670                 else {
 671                         vcpu->arch.xen.vcpu_id = data->u.vcpu_id;
 672                         r = 0;
 673                 }
 674                 break;
 675
 676         case KVM_XEN_VCPU_ATTR_TYPE_TIMER:
 677                 if (data->u.timer.port) {
 678                         if (data->u.timer.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL) {
 679                                 r = -EINVAL;
 680                                 break;
 681                         }
 682                         vcpu->arch.xen.timer_virq = data->u.timer.port;
 683                         kvm_xen_init_timer(vcpu);
 684
 685                         /* Restart the timer if it's set */
 686                         if (data->u.timer.expires_ns)
 687                                 kvm_xen_start_timer(vcpu, data->u.timer.expires_ns,
 688                                                     data->u.timer.expires_ns -
 689                                                     get_kvmclock_ns(vcpu->kvm));
 690                 } else if (kvm_xen_timer_enabled(vcpu)) {
 691                         kvm_xen_stop_timer(vcpu);
 692                         vcpu->arch.xen.timer_virq = 0;
 693                 }
 694
 695                 r = 0;
 696                 break;
 697
 698         default:
 699                 break;
 700         }
 701
 702         srcu_read_unlock(&vcpu->kvm->srcu, idx);
 703         mutex_unlock(&vcpu->kvm->lock);
 704         return r;
 705 }
 706
 707 int kvm_xen_vcpu_get_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 708 {
 709         int r = -ENOENT;
 710
 711         mutex_lock(&vcpu->kvm->lock);
 712
 713         switch (data->type) {
 714         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
 715                 if (vcpu->arch.xen.vcpu_info_cache.active)
 716                         data->u.gpa = vcpu->arch.xen.vcpu_info_cache.gpa;
 717                 else
 718                         data->u.gpa = GPA_INVALID;
 719                 r = 0;
 720                 break;
 721
 722         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_TIME_INFO:
 723                 if (vcpu->arch.xen.vcpu_time_info_cache.active)
 724                         data->u.gpa = vcpu->arch.xen.vcpu_time_info_cache.gpa;
 725                 else
 726                         data->u.gpa = GPA_INVALID;
 727                 r = 0;
 728                 break;
 729
 730         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADDR:
 731                 if (!sched_info_on()) {
 732                         r = -EOPNOTSUPP;
 733                         break;
 734                 }
 735                 if (vcpu->arch.xen.runstate_cache.active) {
 736                         data->u.gpa = vcpu->arch.xen.runstate_cache.gpa;
 737                         r = 0;
 738                 }
 739                 break;
 740
 741         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_CURRENT:
 742                 if (!sched_info_on()) {
 743                         r = -EOPNOTSUPP;
 744                         break;
 745                 }
 746                 data->u.runstate.state = vcpu->arch.xen.current_runstate;
 747                 r = 0;
 748                 break;
 749
 750         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_DATA:
 751                 if (!sched_info_on()) {
 752                         r = -EOPNOTSUPP;
 753                         break;
 754                 }
 755                 data->u.runstate.state = vcpu->arch.xen.current_runstate;
 756                 data->u.runstate.state_entry_time =
 757                         vcpu->arch.xen.runstate_entry_time;
 758                 data->u.runstate.time_running =
 759                         vcpu->arch.xen.runstate_times[RUNSTATE_running];
 760                 data->u.runstate.time_runnable =
 761                         vcpu->arch.xen.runstate_times[RUNSTATE_runnable];
 762                 data->u.runstate.time_blocked =
 763                         vcpu->arch.xen.runstate_times[RUNSTATE_blocked];
 764                 data->u.runstate.time_offline =
 765                         vcpu->arch.xen.runstate_times[RUNSTATE_offline];
 766                 r = 0;
 767                 break;
 768
 769         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADJUST:
 770                 r = -EINVAL;
 771                 break;
 772
 773         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_ID:
 774                 data->u.vcpu_id = vcpu->arch.xen.vcpu_id;
 775                 r = 0;
 776                 break;
 777
 778         case KVM_XEN_VCPU_ATTR_TYPE_TIMER:
 779                 data->u.timer.port = vcpu->arch.xen.timer_virq;
 780                 data->u.timer.priority = KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL;
 781                 data->u.timer.expires_ns = vcpu->arch.xen.timer_expires;
 782                 r = 0;
 783                 break;
 784
 785         default:
 786                 break;
 787         }
 788
 789         mutex_unlock(&vcpu->kvm->lock);
 790         return r;
 791 }
 792
 793 int kvm_xen_write_hypercall_page(struct kvm_vcpu *vcpu, u64 data)
 794 {
 795         struct kvm *kvm = vcpu->kvm;
 796         u32 page_num = data & ~PAGE_MASK;
 797         u64 page_addr = data & PAGE_MASK;
 798         bool lm = is_long_mode(vcpu);
 799
 800         /* Latch long_mode for shared_info pages etc. */
 801         vcpu->kvm->arch.xen.long_mode = lm;
 802
 803         /*
 804          * If Xen hypercall intercept is enabled, fill the hypercall
 805          * page with VMCALL/VMMCALL instructions since that's what
 806          * we catch. Else the VMM has provided the hypercall pages
 807          * with instructions of its own choosing, so use those.
 808          */
 809         if (kvm_xen_hypercall_enabled(kvm)) {
 810                 u8 instructions[32];
 811                 int i;
 812
 813                 if (page_num)
 814                         return 1;
 815
 816                 /* mov imm32, %eax */
 817                 instructions[0] = 0xb8;
 818
 819                 /* vmcall / vmmcall */
 820                 static_call(kvm_x86_patch_hypercall)(vcpu, instructions + 5);
 821
 822                 /* ret */
 823                 instructions[8] = 0xc3;
 824
 825                 /* int3 to pad */
 826                 memset(instructions + 9, 0xcc, sizeof(instructions) - 9);
 827
 828                 for (i = 0; i < PAGE_SIZE / sizeof(instructions); i++) {
 829                         *(u32 *)&instructions[1] = i;
 830                         if (kvm_vcpu_write_guest(vcpu,
 831                                                  page_addr + (i * sizeof(instructions)),
 832                                                  instructions, sizeof(instructions)))
 833                                 return 1;
 834                 }
 835         } else {
 836                 /*
 837                  * Note, truncation is a non-issue as 'lm' is guaranteed to be
 838                  * false for a 32-bit kernel, i.e. when hva_t is only 4 bytes.
 839                  */
 840                 hva_t blob_addr = lm ? kvm->arch.xen_hvm_config.blob_addr_64
 841                                      : kvm->arch.xen_hvm_config.blob_addr_32;
 842                 u8 blob_size = lm ? kvm->arch.xen_hvm_config.blob_size_64
 843                                   : kvm->arch.xen_hvm_config.blob_size_32;
 844                 u8 *page;
 845
 846                 if (page_num >= blob_size)
 847                         return 1;
 848
 849                 blob_addr += page_num * PAGE_SIZE;
 850
 851                 page = memdup_user((u8 __user *)blob_addr, PAGE_SIZE);
 852                 if (IS_ERR(page))
 853                         return PTR_ERR(page);
 854
 855                 if (kvm_vcpu_write_guest(vcpu, page_addr, page, PAGE_SIZE)) {
 856                         kfree(page);
 857                         return 1;
 858                 }
 859         }
 860         return 0;
 861 }
 862
 863 int kvm_xen_hvm_config(struct kvm *kvm, struct kvm_xen_hvm_config *xhc)
 864 {
 865         if (xhc->flags & ~KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL)
 866                 return -EINVAL;
 867
 868         /*
 869          * With hypercall interception the kernel generates its own
 870          * hypercall page so it must not be provided.
 871          */
 872         if ((xhc->flags & KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL) &&
 873             (xhc->blob_addr_32 || xhc->blob_addr_64 ||
 874              xhc->blob_size_32 || xhc->blob_size_64))
 875                 return -EINVAL;
 876
 877         mutex_lock(&kvm->lock);
 878
 879         if (xhc->msr && !kvm->arch.xen_hvm_config.msr)
 880                 static_branch_inc(&kvm_xen_enabled.key);
 881         else if (!xhc->msr && kvm->arch.xen_hvm_config.msr)
 882                 static_branch_slow_dec_deferred(&kvm_xen_enabled);
 883
 884         memcpy(&kvm->arch.xen_hvm_config, xhc, sizeof(*xhc));
 885
 886         mutex_unlock(&kvm->lock);
 887         return 0;
 888 }
 889
 890 static int kvm_xen_hypercall_set_result(struct kvm_vcpu *vcpu, u64 result)
 891 {
 892         kvm_rax_write(vcpu, result);
 893         return kvm_skip_emulated_instruction(vcpu);
 894 }
 895
 896 static int kvm_xen_hypercall_complete_userspace(struct kvm_vcpu *vcpu)
 897 {
 898         struct kvm_run *run = vcpu->run;
 899
 900         if (unlikely(!kvm_is_linear_rip(vcpu, vcpu->arch.xen.hypercall_rip)))
 901                 return 1;
 902
 903         return kvm_xen_hypercall_set_result(vcpu, run->xen.u.hcall.result);
 904 }
 905
 906 static bool kvm_xen_hcall_sched_op(struct kvm_vcpu *vcpu, int cmd, u64 param, u64 *r)
 907 {
 908         switch (cmd) {
 909         case SCHEDOP_yield:
 910                 kvm_vcpu_on_spin(vcpu, true);
 911                 *r = 0;
 912                 return true;
 913         default:
 914                 break;
 915         }
 916
 917         return false;
 918 }
 919
 920 struct compat_vcpu_set_singleshot_timer {
 921     uint64_t timeout_abs_ns;
 922     uint32_t flags;
 923 } __attribute__((packed));
 924
 925 static bool kvm_xen_hcall_vcpu_op(struct kvm_vcpu *vcpu, bool longmode, int cmd,
 926                                   int vcpu_id, u64 param, u64 *r)
 927 {
 928         struct vcpu_set_singleshot_timer oneshot;
 929         s64 delta;
 930         gpa_t gpa;
 931         int idx;
 932
 933         if (!kvm_xen_timer_enabled(vcpu))
 934                 return false;
 935
 936         switch (cmd) {
 937         case VCPUOP_set_singleshot_timer:
 938                 if (vcpu->arch.xen.vcpu_id != vcpu_id) {
 939                         *r = -EINVAL;
 940                         return true;
 941                 }
 942                 idx = srcu_read_lock(&vcpu->kvm->srcu);
 943                 gpa = kvm_mmu_gva_to_gpa_system(vcpu, param, NULL);
 944                 srcu_read_unlock(&vcpu->kvm->srcu, idx);
 945
 946                 /*
 947                  * The only difference for 32-bit compat is the 4 bytes of
 948                  * padding after the interesting part of the structure. So
 949                  * for a faithful emulation of Xen we have to *try* to copy
 950                  * the padding and return -EFAULT if we can't. Otherwise we
 951                  * might as well just have copied the 12-byte 32-bit struct.
 952                  */
 953                 BUILD_BUG_ON(offsetof(struct compat_vcpu_set_singleshot_timer, timeout_abs_ns) !=
 954                              offsetof(struct vcpu_set_singleshot_timer, timeout_abs_ns));
 955                 BUILD_BUG_ON(sizeof_field(struct compat_vcpu_set_singleshot_timer, timeout_abs_ns) !=
 956                              sizeof_field(struct vcpu_set_singleshot_timer, timeout_abs_ns));
 957                 BUILD_BUG_ON(offsetof(struct compat_vcpu_set_singleshot_timer, flags) !=
 958                              offsetof(struct vcpu_set_singleshot_timer, flags));
 959                 BUILD_BUG_ON(sizeof_field(struct compat_vcpu_set_singleshot_timer, flags) !=
 960                              sizeof_field(struct vcpu_set_singleshot_timer, flags));
 961
 962                 if (!gpa ||
 963                     kvm_vcpu_read_guest(vcpu, gpa, &oneshot, longmode ? sizeof(oneshot) :
 964                                         sizeof(struct compat_vcpu_set_singleshot_timer))) {
 965                         *r = -EFAULT;
 966                         return true;
 967                 }
 968
 969                 delta = oneshot.timeout_abs_ns - get_kvmclock_ns(vcpu->kvm);
 970                 if ((oneshot.flags & VCPU_SSHOTTMR_future) && delta < 0) {
 971                         *r = -ETIME;
 972                         return true;
 973                 }
 974
 975                 kvm_xen_start_timer(vcpu, oneshot.timeout_abs_ns, delta);
 976                 *r = 0;
 977                 return true;
 978
 979         case VCPUOP_stop_singleshot_timer:
 980                 if (vcpu->arch.xen.vcpu_id != vcpu_id) {
 981                         *r = -EINVAL;
 982                         return true;
 983                 }
 984                 kvm_xen_stop_timer(vcpu);
 985                 *r = 0;
 986                 return true;
 987         }
 988
 989         return false;
 990 }
 991
 992 static bool kvm_xen_hcall_set_timer_op(struct kvm_vcpu *vcpu, uint64_t timeout,
 993                                        u64 *r)
 994 {
 995         if (!kvm_xen_timer_enabled(vcpu))
 996                 return false;
 997
 998         if (timeout) {
 999                 uint64_t guest_now = get_kvmclock_ns(vcpu->kvm);
1000                 int64_t delta = timeout - guest_now;
1001
1002                 /* Xen has a 'Linux workaround' in do_set_timer_op() which
1003                  * checks for negative absolute timeout values (caused by
1004                  * integer overflow), and for values about 13 days in the
1005                  * future (2^50ns) which would be caused by jiffies
1006                  * overflow. For those cases, it sets the timeout 100ms in
1007                  * the future (not *too* soon, since if a guest really did
1008                  * set a long timeout on purpose we don't want to keep
1009                  * churning CPU time by waking it up).
1010                  */
1011                 if (unlikely((int64_t)timeout < 0 ||
1012                              (delta > 0 && (uint32_t) (delta >> 50) != 0))) {
1013                         delta = 100 * NSEC_PER_MSEC;
1014                         timeout = guest_now + delta;
1015                 }
1016
1017                 kvm_xen_start_timer(vcpu, timeout, delta);
1018         } else {
1019                 kvm_xen_stop_timer(vcpu);
1020         }
1021
1022         *r = 0;
1023         return true;
1024 }
1025
1026 int kvm_xen_hypercall(struct kvm_vcpu *vcpu)
1027 {
1028         bool longmode;
1029         u64 input, params[6], r = -ENOSYS;
1030         bool handled = false;
1031
1032         input = (u64)kvm_register_read(vcpu, VCPU_REGS_RAX);
1033
1034         /* Hyper-V hypercalls get bit 31 set in EAX */
1035         if ((input & 0x80000000) &&
1036             kvm_hv_hypercall_enabled(vcpu))
1037                 return kvm_hv_hypercall(vcpu);
1038
1039         longmode = is_64_bit_hypercall(vcpu);
1040         if (!longmode) {
1041                 params[0] = (u32)kvm_rbx_read(vcpu);
1042                 params[1] = (u32)kvm_rcx_read(vcpu);
1043                 params[2] = (u32)kvm_rdx_read(vcpu);
1044                 params[3] = (u32)kvm_rsi_read(vcpu);
1045                 params[4] = (u32)kvm_rdi_read(vcpu);
1046                 params[5] = (u32)kvm_rbp_read(vcpu);
1047         }
1048 #ifdef CONFIG_X86_64
1049         else {
1050                 params[0] = (u64)kvm_rdi_read(vcpu);
1051                 params[1] = (u64)kvm_rsi_read(vcpu);
1052                 params[2] = (u64)kvm_rdx_read(vcpu);
1053                 params[3] = (u64)kvm_r10_read(vcpu);
1054                 params[4] = (u64)kvm_r8_read(vcpu);
1055                 params[5] = (u64)kvm_r9_read(vcpu);
1056         }
1057 #endif
1058         trace_kvm_xen_hypercall(input, params[0], params[1], params[2],
1059                                 params[3], params[4], params[5]);
1060
1061         switch (input) {
1062         case __HYPERVISOR_event_channel_op:
1063                 if (params[0] == EVTCHNOP_send)
1064                         handled = kvm_xen_hcall_evtchn_send(vcpu, params[1], &r);
1065                 break;
1066         case __HYPERVISOR_sched_op:
1067                 handled = kvm_xen_hcall_sched_op(vcpu, params[0], params[1], &r);
1068                 break;
1069         case __HYPERVISOR_vcpu_op:
1070                 handled = kvm_xen_hcall_vcpu_op(vcpu, longmode, params[0], params[1],
1071                                                 params[2], &r);
1072                 break;
1073         case __HYPERVISOR_set_timer_op: {
1074                 u64 timeout = params[0];
1075                 /* In 32-bit mode, the 64-bit timeout is in two 32-bit params. */
1076                 if (!longmode)
1077                         timeout |= params[1] << 32;
1078                 handled = kvm_xen_hcall_set_timer_op(vcpu, timeout, &r);
1079                 break;
1080         }
1081         default:
1082                 break;
1083         }
1084
1085         if (handled)
1086                 return kvm_xen_hypercall_set_result(vcpu, r);
1087
1088         vcpu->run->exit_reason = KVM_EXIT_XEN;
1089         vcpu->run->xen.type = KVM_EXIT_XEN_HCALL;
1090         vcpu->run->xen.u.hcall.longmode = longmode;
1091         vcpu->run->xen.u.hcall.cpl = static_call(kvm_x86_get_cpl)(vcpu);
1092         vcpu->run->xen.u.hcall.input = input;
1093         vcpu->run->xen.u.hcall.params[0] = params[0];
1094         vcpu->run->xen.u.hcall.params[1] = params[1];
1095         vcpu->run->xen.u.hcall.params[2] = params[2];
1096         vcpu->run->xen.u.hcall.params[3] = params[3];
1097         vcpu->run->xen.u.hcall.params[4] = params[4];
1098         vcpu->run->xen.u.hcall.params[5] = params[5];
1099         vcpu->arch.xen.hypercall_rip = kvm_get_linear_rip(vcpu);
1100         vcpu->arch.complete_userspace_io =
1101                 kvm_xen_hypercall_complete_userspace;
1102
1103         return 0;
1104 }
1105
1106 static inline int max_evtchn_port(struct kvm *kvm)
1107 {
1108         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode)
1109                 return EVTCHN_2L_NR_CHANNELS;
1110         else
1111                 return COMPAT_EVTCHN_2L_NR_CHANNELS;
1112 }
1113
1114 /*
1115  * The return value from this function is propagated to kvm_set_irq() API,
1116  * so it returns:
1117  *  < 0   Interrupt was ignored (masked or not delivered for other reasons)
1118  *  = 0   Interrupt was coalesced (previous irq is still pending)
1119  *  > 0   Number of CPUs interrupt was delivered to
1120  *
1121  * It is also called directly from kvm_arch_set_irq_inatomic(), where the
1122  * only check on its return value is a comparison with -EWOULDBLOCK'.
1123  */
1124 int kvm_xen_set_evtchn_fast(struct kvm_xen_evtchn *xe, struct kvm *kvm)
1125 {
1126         struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
1127         struct kvm_vcpu *vcpu;
1128         unsigned long *pending_bits, *mask_bits;
1129         unsigned long flags;
1130         int port_word_bit;
1131         bool kick_vcpu = false;
1132         int vcpu_idx, idx, rc;
1133
1134         vcpu_idx = READ_ONCE(xe->vcpu_idx);
1135         if (vcpu_idx >= 0)
1136                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
1137         else {
1138                 vcpu = kvm_get_vcpu_by_id(kvm, xe->vcpu_id);
1139                 if (!vcpu)
1140                         return -EINVAL;
1141                 WRITE_ONCE(xe->vcpu_idx, kvm_vcpu_get_idx(vcpu));
1142         }
1143
1144         if (!vcpu->arch.xen.vcpu_info_cache.active)
1145                 return -EINVAL;
1146
1147         if (xe->port >= max_evtchn_port(kvm))
1148                 return -EINVAL;
1149
1150         rc = -EWOULDBLOCK;
1151
1152         idx = srcu_read_lock(&kvm->srcu);
1153
1154         read_lock_irqsave(&gpc->lock, flags);
1155         if (!kvm_gfn_to_pfn_cache_check(kvm, gpc, gpc->gpa, PAGE_SIZE))
1156                 goto out_rcu;
1157
1158         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
1159                 struct shared_info *shinfo = gpc->khva;
1160                 pending_bits = (unsigned long *)&shinfo->evtchn_pending;
1161                 mask_bits = (unsigned long *)&shinfo->evtchn_mask;
1162                 port_word_bit = xe->port / 64;
1163         } else {
1164                 struct compat_shared_info *shinfo = gpc->khva;
1165                 pending_bits = (unsigned long *)&shinfo->evtchn_pending;
1166                 mask_bits = (unsigned long *)&shinfo->evtchn_mask;
1167                 port_word_bit = xe->port / 32;
1168         }
1169
1170         /*
1171          * If this port wasn't already set, and if it isn't masked, then
1172          * we try to set the corresponding bit in the in-kernel shadow of
1173          * evtchn_pending_sel for the target vCPU. And if *that* wasn't
1174          * already set, then we kick the vCPU in question to write to the
1175          * *real* evtchn_pending_sel in its own guest vcpu_info struct.
1176          */
1177         if (test_and_set_bit(xe->port, pending_bits)) {
1178                 rc = 0; /* It was already raised */
1179         } else if (test_bit(xe->port, mask_bits)) {
1180                 rc = -ENOTCONN; /* Masked */
1181         } else {
1182                 rc = 1; /* Delivered to the bitmap in shared_info. */
1183                 /* Now switch to the vCPU's vcpu_info to set the index and pending_sel */
1184                 read_unlock_irqrestore(&gpc->lock, flags);
1185                 gpc = &vcpu->arch.xen.vcpu_info_cache;
1186
1187                 read_lock_irqsave(&gpc->lock, flags);
1188                 if (!kvm_gfn_to_pfn_cache_check(kvm, gpc, gpc->gpa, sizeof(struct vcpu_info))) {
1189                         /*
1190                          * Could not access the vcpu_info. Set the bit in-kernel
1191                          * and prod the vCPU to deliver it for itself.
1192                          */
1193                         if (!test_and_set_bit(port_word_bit, &vcpu->arch.xen.evtchn_pending_sel))
1194                                 kick_vcpu = true;
1195                         goto out_rcu;
1196                 }
1197
1198                 if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
1199                         struct vcpu_info *vcpu_info = gpc->khva;
1200                         if (!test_and_set_bit(port_word_bit, &vcpu_info->evtchn_pending_sel)) {
1201                                 WRITE_ONCE(vcpu_info->evtchn_upcall_pending, 1);
1202                                 kick_vcpu = true;
1203                         }
1204                 } else {
1205                         struct compat_vcpu_info *vcpu_info = gpc->khva;
1206                         if (!test_and_set_bit(port_word_bit,
1207                                               (unsigned long *)&vcpu_info->evtchn_pending_sel)) {
1208                                 WRITE_ONCE(vcpu_info->evtchn_upcall_pending, 1);
1209                                 kick_vcpu = true;
1210                         }
1211                 }
1212         }
1213
1214  out_rcu:
1215         read_unlock_irqrestore(&gpc->lock, flags);
1216         srcu_read_unlock(&kvm->srcu, idx);
1217
1218         if (kick_vcpu) {
1219                 kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
1220                 kvm_vcpu_kick(vcpu);
1221         }
1222
1223         return rc;
1224 }
1225
1226 static int kvm_xen_set_evtchn(struct kvm_xen_evtchn *xe, struct kvm *kvm)
1227 {
1228         bool mm_borrowed = false;
1229         int rc;
1230
1231         rc = kvm_xen_set_evtchn_fast(xe, kvm);
1232         if (rc != -EWOULDBLOCK)
1233                 return rc;
1234
1235         if (current->mm != kvm->mm) {
1236                 /*
1237                  * If not on a thread which already belongs to this KVM,
1238                  * we'd better be in the irqfd workqueue.
1239                  */
1240                 if (WARN_ON_ONCE(current->mm))
1241                         return -EINVAL;
1242
1243                 kthread_use_mm(kvm->mm);
1244                 mm_borrowed = true;
1245         }
1246
1247         /*
1248          * For the irqfd workqueue, using the main kvm->lock mutex is
1249          * fine since this function is invoked from kvm_set_irq() with
1250          * no other lock held, no srcu. In future if it will be called
1251          * directly from a vCPU thread (e.g. on hypercall for an IPI)
1252          * then it may need to switch to using a leaf-node mutex for
1253          * serializing the shared_info mapping.
1254          */
1255         mutex_lock(&kvm->lock);
1256
1257         /*
1258          * It is theoretically possible for the page to be unmapped
1259          * and the MMU notifier to invalidate the shared_info before
1260          * we even get to use it. In that case, this looks like an
1261          * infinite loop. It was tempting to do it via the userspace
1262          * HVA instead... but that just *hides* the fact that it's
1263          * an infinite loop, because if a fault occurs and it waits
1264          * for the page to come back, it can *still* immediately
1265          * fault and have to wait again, repeatedly.
1266          *
1267          * Conversely, the page could also have been reinstated by
1268          * another thread before we even obtain the mutex above, so
1269          * check again *first* before remapping it.
1270          */
1271         do {
1272                 struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
1273                 int idx;
1274
1275                 rc = kvm_xen_set_evtchn_fast(xe, kvm);
1276                 if (rc != -EWOULDBLOCK)
1277                         break;
1278
1279                 idx = srcu_read_lock(&kvm->srcu);
1280                 rc = kvm_gfn_to_pfn_cache_refresh(kvm, gpc, gpc->gpa, PAGE_SIZE);
1281                 srcu_read_unlock(&kvm->srcu, idx);
1282         } while(!rc);
1283
1284         mutex_unlock(&kvm->lock);
1285
1286         if (mm_borrowed)
1287                 kthread_unuse_mm(kvm->mm);
1288
1289         return rc;
1290 }
1291
1292 /* This is the version called from kvm_set_irq() as the .set function */
1293 static int evtchn_set_fn(struct kvm_kernel_irq_routing_entry *e, struct kvm *kvm,
1294                          int irq_source_id, int level, bool line_status)
1295 {
1296         if (!level)
1297                 return -EINVAL;
1298
1299         return kvm_xen_set_evtchn(&e->xen_evtchn, kvm);
1300 }
1301
1302 /*
1303  * Set up an event channel interrupt from the KVM IRQ routing table.
1304  * Used for e.g. PIRQ from passed through physical devices.
1305  */
1306 int kvm_xen_setup_evtchn(struct kvm *kvm,
1307                          struct kvm_kernel_irq_routing_entry *e,
1308                          const struct kvm_irq_routing_entry *ue)
1309
1310 {
1311         struct kvm_vcpu *vcpu;
1312
1313         if (ue->u.xen_evtchn.port >= max_evtchn_port(kvm))
1314                 return -EINVAL;
1315
1316         /* We only support 2 level event channels for now */
1317         if (ue->u.xen_evtchn.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1318                 return -EINVAL;
1319
1320         /*
1321          * Xen gives us interesting mappings from vCPU index to APIC ID,
1322          * which means kvm_get_vcpu_by_id() has to iterate over all vCPUs
1323          * to find it. Do that once at setup time, instead of every time.
1324          * But beware that on live update / live migration, the routing
1325          * table might be reinstated before the vCPU threads have finished
1326          * recreating their vCPUs.
1327          */
1328         vcpu = kvm_get_vcpu_by_id(kvm, ue->u.xen_evtchn.vcpu);
1329         if (vcpu)
1330                 e->xen_evtchn.vcpu_idx = kvm_vcpu_get_idx(vcpu);
1331         else
1332                 e->xen_evtchn.vcpu_idx = -1;
1333
1334         e->xen_evtchn.port = ue->u.xen_evtchn.port;
1335         e->xen_evtchn.vcpu_id = ue->u.xen_evtchn.vcpu;
1336         e->xen_evtchn.priority = ue->u.xen_evtchn.priority;
1337         e->set = evtchn_set_fn;
1338
1339         return 0;
1340 }
1341
1342 /*
1343  * Explicit event sending from userspace with KVM_XEN_HVM_EVTCHN_SEND ioctl.
1344  */
1345 int kvm_xen_hvm_evtchn_send(struct kvm *kvm, struct kvm_irq_routing_xen_evtchn *uxe)
1346 {
1347         struct kvm_xen_evtchn e;
1348         int ret;
1349
1350         if (!uxe->port || uxe->port >= max_evtchn_port(kvm))
1351                 return -EINVAL;
1352
1353         /* We only support 2 level event channels for now */
1354         if (uxe->priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1355                 return -EINVAL;
1356
1357         e.port = uxe->port;
1358         e.vcpu_id = uxe->vcpu;
1359         e.vcpu_idx = -1;
1360         e.priority = uxe->priority;
1361
1362         ret = kvm_xen_set_evtchn(&e, kvm);
1363
1364         /*
1365          * None of that 'return 1 if it actually got delivered' nonsense.
1366          * We don't care if it was masked (-ENOTCONN) either.
1367          */
1368         if (ret > 0 || ret == -ENOTCONN)
1369                 ret = 0;
1370
1371         return ret;
1372 }
1373
1374 /*
1375  * Support for *outbound* event channel events via the EVTCHNOP_send hypercall.
1376  */
1377 struct evtchnfd {
1378         u32 send_port;
1379         u32 type;
1380         union {
1381                 struct kvm_xen_evtchn port;
1382                 struct {
1383                         u32 port; /* zero */
1384                         struct eventfd_ctx *ctx;
1385                 } eventfd;
1386         } deliver;
1387 };
1388
1389 /*
1390  * Update target vCPU or priority for a registered sending channel.
1391  */
1392 static int kvm_xen_eventfd_update(struct kvm *kvm,
1393                                   struct kvm_xen_hvm_attr *data)
1394 {
1395         u32 port = data->u.evtchn.send_port;
1396         struct evtchnfd *evtchnfd;
1397
1398         if (!port || port >= max_evtchn_port(kvm))
1399                 return -EINVAL;
1400
1401         mutex_lock(&kvm->lock);
1402         evtchnfd = idr_find(&kvm->arch.xen.evtchn_ports, port);
1403         mutex_unlock(&kvm->lock);
1404
1405         if (!evtchnfd)
1406                 return -ENOENT;
1407
1408         /* For an UPDATE, nothing may change except the priority/vcpu */
1409         if (evtchnfd->type != data->u.evtchn.type)
1410                 return -EINVAL;
1411
1412         /*
1413          * Port cannot change, and if it's zero that was an eventfd
1414          * which can't be changed either.
1415          */
1416         if (!evtchnfd->deliver.port.port ||
1417             evtchnfd->deliver.port.port != data->u.evtchn.deliver.port.port)
1418                 return -EINVAL;
1419
1420         /* We only support 2 level event channels for now */
1421         if (data->u.evtchn.deliver.port.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1422                 return -EINVAL;
1423
1424         mutex_lock(&kvm->lock);
1425         evtchnfd->deliver.port.priority = data->u.evtchn.deliver.port.priority;
1426         if (evtchnfd->deliver.port.vcpu_id != data->u.evtchn.deliver.port.vcpu) {
1427                 evtchnfd->deliver.port.vcpu_id = data->u.evtchn.deliver.port.vcpu;
1428                 evtchnfd->deliver.port.vcpu_idx = -1;
1429         }
1430         mutex_unlock(&kvm->lock);
1431         return 0;
1432 }
1433
1434 /*
1435  * Configure the target (eventfd or local port delivery) for sending on
1436  * a given event channel.
1437  */
1438 static int kvm_xen_eventfd_assign(struct kvm *kvm,
1439                                   struct kvm_xen_hvm_attr *data)
1440 {
1441         u32 port = data->u.evtchn.send_port;
1442         struct eventfd_ctx *eventfd = NULL;
1443         struct evtchnfd *evtchnfd = NULL;
1444         int ret = -EINVAL;
1445
1446         if (!port || port >= max_evtchn_port(kvm))
1447                 return -EINVAL;
1448
1449         evtchnfd = kzalloc(sizeof(struct evtchnfd), GFP_KERNEL);
1450         if (!evtchnfd)
1451                 return -ENOMEM;
1452
1453         switch(data->u.evtchn.type) {
1454         case EVTCHNSTAT_ipi:
1455                 /* IPI  must map back to the same port# */
1456                 if (data->u.evtchn.deliver.port.port != data->u.evtchn.send_port)
1457                         goto out; /* -EINVAL */
1458                 break;
1459
1460         case EVTCHNSTAT_interdomain:
1461                 if (data->u.evtchn.deliver.port.port) {
1462                         if (data->u.evtchn.deliver.port.port >= max_evtchn_port(kvm))
1463                                 goto out; /* -EINVAL */
1464                 } else {
1465                         eventfd = eventfd_ctx_fdget(data->u.evtchn.deliver.eventfd.fd);
1466                         if (IS_ERR(eventfd)) {
1467                                 ret = PTR_ERR(eventfd);
1468                                 goto out;
1469                         }
1470                 }
1471                 break;
1472
1473         case EVTCHNSTAT_virq:
1474         case EVTCHNSTAT_closed:
1475         case EVTCHNSTAT_unbound:
1476         case EVTCHNSTAT_pirq:
1477         default: /* Unknown event channel type */
1478                 goto out; /* -EINVAL */
1479         }
1480
1481         evtchnfd->send_port = data->u.evtchn.send_port;
1482         evtchnfd->type = data->u.evtchn.type;
1483         if (eventfd) {
1484                 evtchnfd->deliver.eventfd.ctx = eventfd;
1485         } else {
1486                 /* We only support 2 level event channels for now */
1487                 if (data->u.evtchn.deliver.port.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1488                         goto out; /* -EINVAL; */
1489
1490                 evtchnfd->deliver.port.port = data->u.evtchn.deliver.port.port;
1491                 evtchnfd->deliver.port.vcpu_id = data->u.evtchn.deliver.port.vcpu;
1492                 evtchnfd->deliver.port.vcpu_idx = -1;
1493                 evtchnfd->deliver.port.priority = data->u.evtchn.deliver.port.priority;
1494         }
1495
1496         mutex_lock(&kvm->lock);
1497         ret = idr_alloc(&kvm->arch.xen.evtchn_ports, evtchnfd, port, port + 1,
1498                         GFP_KERNEL);
1499         mutex_unlock(&kvm->lock);
1500         if (ret >= 0)
1501                 return 0;
1502
1503         if (ret == -ENOSPC)
1504                 ret = -EEXIST;
1505 out:
1506         if (eventfd)
1507                 eventfd_ctx_put(eventfd);
1508         kfree(evtchnfd);
1509         return ret;
1510 }
1511
1512 static int kvm_xen_eventfd_deassign(struct kvm *kvm, u32 port)
1513 {
1514         struct evtchnfd *evtchnfd;
1515
1516         mutex_lock(&kvm->lock);
1517         evtchnfd = idr_remove(&kvm->arch.xen.evtchn_ports, port);
1518         mutex_unlock(&kvm->lock);
1519
1520         if (!evtchnfd)
1521                 return -ENOENT;
1522
1523         if (kvm)
1524                 synchronize_srcu(&kvm->srcu);
1525         if (!evtchnfd->deliver.port.port)
1526                 eventfd_ctx_put(evtchnfd->deliver.eventfd.ctx);
1527         kfree(evtchnfd);
1528         return 0;
1529 }
1530
1531 static int kvm_xen_eventfd_reset(struct kvm *kvm)
1532 {
1533         struct evtchnfd *evtchnfd;
1534         int i;
1535
1536         mutex_lock(&kvm->lock);
1537         idr_for_each_entry(&kvm->arch.xen.evtchn_ports, evtchnfd, i) {
1538                 idr_remove(&kvm->arch.xen.evtchn_ports, evtchnfd->send_port);
1539                 synchronize_srcu(&kvm->srcu);
1540                 if (!evtchnfd->deliver.port.port)
1541                         eventfd_ctx_put(evtchnfd->deliver.eventfd.ctx);
1542                 kfree(evtchnfd);
1543         }
1544         mutex_unlock(&kvm->lock);
1545
1546         return 0;
1547 }
1548
1549 static int kvm_xen_setattr_evtchn(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
1550 {
1551         u32 port = data->u.evtchn.send_port;
1552
1553         if (data->u.evtchn.flags == KVM_XEN_EVTCHN_RESET)
1554                 return kvm_xen_eventfd_reset(kvm);
1555
1556         if (!port || port >= max_evtchn_port(kvm))
1557                 return -EINVAL;
1558
1559         if (data->u.evtchn.flags == KVM_XEN_EVTCHN_DEASSIGN)
1560                 return kvm_xen_eventfd_deassign(kvm, port);
1561         if (data->u.evtchn.flags == KVM_XEN_EVTCHN_UPDATE)
1562                 return kvm_xen_eventfd_update(kvm, data);
1563         if (data->u.evtchn.flags)
1564                 return -EINVAL;
1565
1566         return kvm_xen_eventfd_assign(kvm, data);
1567 }
1568
1569 static bool kvm_xen_hcall_evtchn_send(struct kvm_vcpu *vcpu, u64 param, u64 *r)
1570 {
1571         struct evtchnfd *evtchnfd;
1572         struct evtchn_send send;
1573         gpa_t gpa;
1574         int idx;
1575
1576         idx = srcu_read_lock(&vcpu->kvm->srcu);
1577         gpa = kvm_mmu_gva_to_gpa_system(vcpu, param, NULL);
1578         srcu_read_unlock(&vcpu->kvm->srcu, idx);
1579
1580         if (!gpa || kvm_vcpu_read_guest(vcpu, gpa, &send, sizeof(send))) {
1581                 *r = -EFAULT;
1582                 return true;
1583         }
1584
1585         /* The evtchn_ports idr is protected by vcpu->kvm->srcu */
1586         evtchnfd = idr_find(&vcpu->kvm->arch.xen.evtchn_ports, send.port);
1587         if (!evtchnfd)
1588                 return false;
1589
1590         if (evtchnfd->deliver.port.port) {
1591                 int ret = kvm_xen_set_evtchn(&evtchnfd->deliver.port, vcpu->kvm);
1592                 if (ret < 0 && ret != -ENOTCONN)
1593                         return false;
1594         } else {
1595                 eventfd_signal(evtchnfd->deliver.eventfd.ctx, 1);
1596         }
1597
1598         *r = 0;
1599         return true;
1600 }
1601
1602 void kvm_xen_init_vcpu(struct kvm_vcpu *vcpu)
1603 {
1604         vcpu->arch.xen.vcpu_id = vcpu->vcpu_idx;
1605 }
1606
1607 void kvm_xen_destroy_vcpu(struct kvm_vcpu *vcpu)
1608 {
1609         if (kvm_xen_timer_enabled(vcpu))
1610                 kvm_xen_stop_timer(vcpu);
1611
1612         kvm_gfn_to_pfn_cache_destroy(vcpu->kvm,
1613                                      &vcpu->arch.xen.runstate_cache);
1614         kvm_gfn_to_pfn_cache_destroy(vcpu->kvm,
1615                                      &vcpu->arch.xen.vcpu_info_cache);
1616         kvm_gfn_to_pfn_cache_destroy(vcpu->kvm,
1617                                      &vcpu->arch.xen.vcpu_time_info_cache);
1618 }
1619
1620 void kvm_xen_init_vm(struct kvm *kvm)
1621 {
1622         idr_init(&kvm->arch.xen.evtchn_ports);
1623 }
1624
1625 void kvm_xen_destroy_vm(struct kvm *kvm)
1626 {
1627         struct evtchnfd *evtchnfd;
1628         int i;
1629
1630         kvm_gfn_to_pfn_cache_destroy(kvm, &kvm->arch.xen.shinfo_cache);
1631
1632         idr_for_each_entry(&kvm->arch.xen.evtchn_ports, evtchnfd, i) {
1633                 if (!evtchnfd->deliver.port.port)
1634                         eventfd_ctx_put(evtchnfd->deliver.eventfd.ctx);
1635                 kfree(evtchnfd);
1636         }
1637         idr_destroy(&kvm->arch.xen.evtchn_ports);
1638
1639         if (kvm->arch.xen_hvm_config.msr)
1640                 static_branch_slow_dec_deferred(&kvm_xen_enabled);
1641 }