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