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 "hyperv.h"
  12
  13 #include <linux/kvm_host.h>
  14 #include <linux/sched/stat.h>
  15
  16 #include <trace/events/kvm.h>
  17 #include <xen/interface/xen.h>
  18 #include <xen/interface/vcpu.h>
  19 #include <xen/interface/event_channel.h>
  20
  21 #include "trace.h"
  22
  23 DEFINE_STATIC_KEY_DEFERRED_FALSE(kvm_xen_enabled, HZ);
  24
  25 static int kvm_xen_shared_info_init(struct kvm *kvm, gfn_t gfn)
  26 {
  27         struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
  28         struct pvclock_wall_clock *wc;
  29         gpa_t gpa = gfn_to_gpa(gfn);
  30         u32 *wc_sec_hi;
  31         u32 wc_version;
  32         u64 wall_nsec;
  33         int ret = 0;
  34         int idx = srcu_read_lock(&kvm->srcu);
  35
  36         if (gfn == GPA_INVALID) {
  37                 kvm_gfn_to_pfn_cache_destroy(kvm, gpc);
  38                 goto out;
  39         }
  40
  41         do {
  42                 ret = kvm_gfn_to_pfn_cache_init(kvm, gpc, NULL, false, true,
  43                                                 gpa, PAGE_SIZE, false);
  44                 if (ret)
  45                         goto out;
  46
  47                 /*
  48                  * This code mirrors kvm_write_wall_clock() except that it writes
  49                  * directly through the pfn cache and doesn't mark the page dirty.
  50                  */
  51                 wall_nsec = ktime_get_real_ns() - get_kvmclock_ns(kvm);
  52
  53                 /* It could be invalid again already, so we need to check */
  54                 read_lock_irq(&gpc->lock);
  55
  56                 if (gpc->valid)
  57                         break;
  58
  59                 read_unlock_irq(&gpc->lock);
  60         } while (1);
  61
  62         /* Paranoia checks on the 32-bit struct layout */
  63         BUILD_BUG_ON(offsetof(struct compat_shared_info, wc) != 0x900);
  64         BUILD_BUG_ON(offsetof(struct compat_shared_info, arch.wc_sec_hi) != 0x924);
  65         BUILD_BUG_ON(offsetof(struct pvclock_vcpu_time_info, version) != 0);
  66
  67 #ifdef CONFIG_X86_64
  68         /* Paranoia checks on the 64-bit struct layout */
  69         BUILD_BUG_ON(offsetof(struct shared_info, wc) != 0xc00);
  70         BUILD_BUG_ON(offsetof(struct shared_info, wc_sec_hi) != 0xc0c);
  71
  72         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
  73                 struct shared_info *shinfo = gpc->khva;
  74
  75                 wc_sec_hi = &shinfo->wc_sec_hi;
  76                 wc = &shinfo->wc;
  77         } else
  78 #endif
  79         {
  80                 struct compat_shared_info *shinfo = gpc->khva;
  81
  82                 wc_sec_hi = &shinfo->arch.wc_sec_hi;
  83                 wc = &shinfo->wc;
  84         }
  85
  86         /* Increment and ensure an odd value */
  87         wc_version = wc->version = (wc->version + 1) | 1;
  88         smp_wmb();
  89
  90         wc->nsec = do_div(wall_nsec,  1000000000);
  91         wc->sec = (u32)wall_nsec;
  92         *wc_sec_hi = wall_nsec >> 32;
  93         smp_wmb();
  94
  95         wc->version = wc_version + 1;
  96         read_unlock_irq(&gpc->lock);
  97
  98         kvm_make_all_cpus_request(kvm, KVM_REQ_MASTERCLOCK_UPDATE);
  99
 100 out:
 101         srcu_read_unlock(&kvm->srcu, idx);
 102         return ret;
 103 }
 104
 105 static void kvm_xen_update_runstate(struct kvm_vcpu *v, int state)
 106 {
 107         struct kvm_vcpu_xen *vx = &v->arch.xen;
 108         u64 now = get_kvmclock_ns(v->kvm);
 109         u64 delta_ns = now - vx->runstate_entry_time;
 110         u64 run_delay = current->sched_info.run_delay;
 111
 112         if (unlikely(!vx->runstate_entry_time))
 113                 vx->current_runstate = RUNSTATE_offline;
 114
 115         /*
 116          * Time waiting for the scheduler isn't "stolen" if the
 117          * vCPU wasn't running anyway.
 118          */
 119         if (vx->current_runstate == RUNSTATE_running) {
 120                 u64 steal_ns = run_delay - vx->last_steal;
 121
 122                 delta_ns -= steal_ns;
 123
 124                 vx->runstate_times[RUNSTATE_runnable] += steal_ns;
 125         }
 126         vx->last_steal = run_delay;
 127
 128         vx->runstate_times[vx->current_runstate] += delta_ns;
 129         vx->current_runstate = state;
 130         vx->runstate_entry_time = now;
 131 }
 132
 133 void kvm_xen_update_runstate_guest(struct kvm_vcpu *v, int state)
 134 {
 135         struct kvm_vcpu_xen *vx = &v->arch.xen;
 136         uint64_t state_entry_time;
 137         unsigned int offset;
 138
 139         kvm_xen_update_runstate(v, state);
 140
 141         if (!vx->runstate_set)
 142                 return;
 143
 144         BUILD_BUG_ON(sizeof(struct compat_vcpu_runstate_info) != 0x2c);
 145
 146         offset = offsetof(struct compat_vcpu_runstate_info, state_entry_time);
 147 #ifdef CONFIG_X86_64
 148         /*
 149          * The only difference is alignment of uint64_t in 32-bit.
 150          * So the first field 'state' is accessed directly using
 151          * offsetof() (where its offset happens to be zero), while the
 152          * remaining fields which are all uint64_t, start at 'offset'
 153          * which we tweak here by adding 4.
 154          */
 155         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state_entry_time) !=
 156                      offsetof(struct compat_vcpu_runstate_info, state_entry_time) + 4);
 157         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, time) !=
 158                      offsetof(struct compat_vcpu_runstate_info, time) + 4);
 159
 160         if (v->kvm->arch.xen.long_mode)
 161                 offset = offsetof(struct vcpu_runstate_info, state_entry_time);
 162 #endif
 163         /*
 164          * First write the updated state_entry_time at the appropriate
 165          * location determined by 'offset'.
 166          */
 167         state_entry_time = vx->runstate_entry_time;
 168         state_entry_time |= XEN_RUNSTATE_UPDATE;
 169
 170         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, state_entry_time) !=
 171                      sizeof(state_entry_time));
 172         BUILD_BUG_ON(sizeof_field(struct compat_vcpu_runstate_info, state_entry_time) !=
 173                      sizeof(state_entry_time));
 174
 175         if (kvm_write_guest_offset_cached(v->kvm, &v->arch.xen.runstate_cache,
 176                                           &state_entry_time, offset,
 177                                           sizeof(state_entry_time)))
 178                 return;
 179         smp_wmb();
 180
 181         /*
 182          * Next, write the new runstate. This is in the *same* place
 183          * for 32-bit and 64-bit guests, asserted here for paranoia.
 184          */
 185         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state) !=
 186                      offsetof(struct compat_vcpu_runstate_info, state));
 187         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, state) !=
 188                      sizeof(vx->current_runstate));
 189         BUILD_BUG_ON(sizeof_field(struct compat_vcpu_runstate_info, state) !=
 190                      sizeof(vx->current_runstate));
 191
 192         if (kvm_write_guest_offset_cached(v->kvm, &v->arch.xen.runstate_cache,
 193                                           &vx->current_runstate,
 194                                           offsetof(struct vcpu_runstate_info, state),
 195                                           sizeof(vx->current_runstate)))
 196                 return;
 197
 198         /*
 199          * Write the actual runstate times immediately after the
 200          * runstate_entry_time.
 201          */
 202         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state_entry_time) !=
 203                      offsetof(struct vcpu_runstate_info, time) - sizeof(u64));
 204         BUILD_BUG_ON(offsetof(struct compat_vcpu_runstate_info, state_entry_time) !=
 205                      offsetof(struct compat_vcpu_runstate_info, time) - sizeof(u64));
 206         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, time) !=
 207                      sizeof_field(struct compat_vcpu_runstate_info, time));
 208         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, time) !=
 209                      sizeof(vx->runstate_times));
 210
 211         if (kvm_write_guest_offset_cached(v->kvm, &v->arch.xen.runstate_cache,
 212                                           &vx->runstate_times[0],
 213                                           offset + sizeof(u64),
 214                                           sizeof(vx->runstate_times)))
 215                 return;
 216
 217         smp_wmb();
 218
 219         /*
 220          * Finally, clear the XEN_RUNSTATE_UPDATE bit in the guest's
 221          * runstate_entry_time field.
 222          */
 223
 224         state_entry_time &= ~XEN_RUNSTATE_UPDATE;
 225         if (kvm_write_guest_offset_cached(v->kvm, &v->arch.xen.runstate_cache,
 226                                           &state_entry_time, offset,
 227                                           sizeof(state_entry_time)))
 228                 return;
 229 }
 230
 231 int __kvm_xen_has_interrupt(struct kvm_vcpu *v)
 232 {
 233         unsigned long evtchn_pending_sel = READ_ONCE(v->arch.xen.evtchn_pending_sel);
 234         bool atomic = in_atomic() || !task_is_running(current);
 235         int err;
 236         u8 rc = 0;
 237
 238         /*
 239          * If the global upcall vector (HVMIRQ_callback_vector) is set and
 240          * the vCPU's evtchn_upcall_pending flag is set, the IRQ is pending.
 241          */
 242         struct gfn_to_hva_cache *ghc = &v->arch.xen.vcpu_info_cache;
 243         struct kvm_memslots *slots = kvm_memslots(v->kvm);
 244         bool ghc_valid = slots->generation == ghc->generation &&
 245                 !kvm_is_error_hva(ghc->hva) && ghc->memslot;
 246
 247         unsigned int offset = offsetof(struct vcpu_info, evtchn_upcall_pending);
 248
 249         /* No need for compat handling here */
 250         BUILD_BUG_ON(offsetof(struct vcpu_info, evtchn_upcall_pending) !=
 251                      offsetof(struct compat_vcpu_info, evtchn_upcall_pending));
 252         BUILD_BUG_ON(sizeof(rc) !=
 253                      sizeof_field(struct vcpu_info, evtchn_upcall_pending));
 254         BUILD_BUG_ON(sizeof(rc) !=
 255                      sizeof_field(struct compat_vcpu_info, evtchn_upcall_pending));
 256
 257         /*
 258          * For efficiency, this mirrors the checks for using the valid
 259          * cache in kvm_read_guest_offset_cached(), but just uses
 260          * __get_user() instead. And falls back to the slow path.
 261          */
 262         if (!evtchn_pending_sel && ghc_valid) {
 263                 /* Fast path */
 264                 pagefault_disable();
 265                 err = __get_user(rc, (u8 __user *)ghc->hva + offset);
 266                 pagefault_enable();
 267                 if (!err)
 268                         return rc;
 269         }
 270
 271         /* Slow path */
 272
 273         /*
 274          * This function gets called from kvm_vcpu_block() after setting the
 275          * task to TASK_INTERRUPTIBLE, to see if it needs to wake immediately
 276          * from a HLT. So we really mustn't sleep. If the page ended up absent
 277          * at that point, just return 1 in order to trigger an immediate wake,
 278          * and we'll end up getting called again from a context where we *can*
 279          * fault in the page and wait for it.
 280          */
 281         if (atomic)
 282                 return 1;
 283
 284         if (!ghc_valid) {
 285                 err = kvm_gfn_to_hva_cache_init(v->kvm, ghc, ghc->gpa, ghc->len);
 286                 if (err || !ghc->memslot) {
 287                         /*
 288                          * If this failed, userspace has screwed up the
 289                          * vcpu_info mapping. No interrupts for you.
 290                          */
 291                         return 0;
 292                 }
 293         }
 294
 295         /*
 296          * Now we have a valid (protected by srcu) userspace HVA in
 297          * ghc->hva which points to the struct vcpu_info. If there
 298          * are any bits in the in-kernel evtchn_pending_sel then
 299          * we need to write those to the guest vcpu_info and set
 300          * its evtchn_upcall_pending flag. If there aren't any bits
 301          * to add, we only want to *check* evtchn_upcall_pending.
 302          */
 303         if (evtchn_pending_sel) {
 304                 bool long_mode = v->kvm->arch.xen.long_mode;
 305
 306                 if (!user_access_begin((void __user *)ghc->hva, sizeof(struct vcpu_info)))
 307                         return 0;
 308
 309                 if (IS_ENABLED(CONFIG_64BIT) && long_mode) {
 310                         struct vcpu_info __user *vi = (void __user *)ghc->hva;
 311
 312                         /* Attempt to set the evtchn_pending_sel bits in the
 313                          * guest, and if that succeeds then clear the same
 314                          * bits in the in-kernel version. */
 315                         asm volatile("1:\t" LOCK_PREFIX "orq %0, %1\n"
 316                                      "\tnotq %0\n"
 317                                      "\t" LOCK_PREFIX "andq %0, %2\n"
 318                                      "2:\n"
 319                                      _ASM_EXTABLE_UA(1b, 2b)
 320                                      : "=r" (evtchn_pending_sel),
 321                                        "+m" (vi->evtchn_pending_sel),
 322                                        "+m" (v->arch.xen.evtchn_pending_sel)
 323                                      : "0" (evtchn_pending_sel));
 324                 } else {
 325                         struct compat_vcpu_info __user *vi = (void __user *)ghc->hva;
 326                         u32 evtchn_pending_sel32 = evtchn_pending_sel;
 327
 328                         /* Attempt to set the evtchn_pending_sel bits in the
 329                          * guest, and if that succeeds then clear the same
 330                          * bits in the in-kernel version. */
 331                         asm volatile("1:\t" LOCK_PREFIX "orl %0, %1\n"
 332                                      "\tnotl %0\n"
 333                                      "\t" LOCK_PREFIX "andl %0, %2\n"
 334                                      "2:\n"
 335                                      _ASM_EXTABLE_UA(1b, 2b)
 336                                      : "=r" (evtchn_pending_sel32),
 337                                        "+m" (vi->evtchn_pending_sel),
 338                                        "+m" (v->arch.xen.evtchn_pending_sel)
 339                                      : "0" (evtchn_pending_sel32));
 340                 }
 341                 rc = 1;
 342                 unsafe_put_user(rc, (u8 __user *)ghc->hva + offset, err);
 343
 344         err:
 345                 user_access_end();
 346
 347                 mark_page_dirty_in_slot(v->kvm, ghc->memslot, ghc->gpa >> PAGE_SHIFT);
 348         } else {
 349                 __get_user(rc, (u8 __user *)ghc->hva + offset);
 350         }
 351
 352         return rc;
 353 }
 354
 355 int kvm_xen_hvm_set_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 356 {
 357         int r = -ENOENT;
 358
 359         mutex_lock(&kvm->lock);
 360
 361         switch (data->type) {
 362         case KVM_XEN_ATTR_TYPE_LONG_MODE:
 363                 if (!IS_ENABLED(CONFIG_64BIT) && data->u.long_mode) {
 364                         r = -EINVAL;
 365                 } else {
 366                         kvm->arch.xen.long_mode = !!data->u.long_mode;
 367                         r = 0;
 368                 }
 369                 break;
 370
 371         case KVM_XEN_ATTR_TYPE_SHARED_INFO:
 372                 r = kvm_xen_shared_info_init(kvm, data->u.shared_info.gfn);
 373                 break;
 374
 375         case KVM_XEN_ATTR_TYPE_UPCALL_VECTOR:
 376                 if (data->u.vector && data->u.vector < 0x10)
 377                         r = -EINVAL;
 378                 else {
 379                         kvm->arch.xen.upcall_vector = data->u.vector;
 380                         r = 0;
 381                 }
 382                 break;
 383
 384         default:
 385                 break;
 386         }
 387
 388         mutex_unlock(&kvm->lock);
 389         return r;
 390 }
 391
 392 int kvm_xen_hvm_get_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 393 {
 394         int r = -ENOENT;
 395
 396         mutex_lock(&kvm->lock);
 397
 398         switch (data->type) {
 399         case KVM_XEN_ATTR_TYPE_LONG_MODE:
 400                 data->u.long_mode = kvm->arch.xen.long_mode;
 401                 r = 0;
 402                 break;
 403
 404         case KVM_XEN_ATTR_TYPE_SHARED_INFO:
 405                 if (kvm->arch.xen.shinfo_cache.active)
 406                         data->u.shared_info.gfn = gpa_to_gfn(kvm->arch.xen.shinfo_cache.gpa);
 407                 else
 408                         data->u.shared_info.gfn = GPA_INVALID;
 409                 r = 0;
 410                 break;
 411
 412         case KVM_XEN_ATTR_TYPE_UPCALL_VECTOR:
 413                 data->u.vector = kvm->arch.xen.upcall_vector;
 414                 r = 0;
 415                 break;
 416
 417         default:
 418                 break;
 419         }
 420
 421         mutex_unlock(&kvm->lock);
 422         return r;
 423 }
 424
 425 int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 426 {
 427         int idx, r = -ENOENT;
 428
 429         mutex_lock(&vcpu->kvm->lock);
 430         idx = srcu_read_lock(&vcpu->kvm->srcu);
 431
 432         switch (data->type) {
 433         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
 434                 /* No compat necessary here. */
 435                 BUILD_BUG_ON(sizeof(struct vcpu_info) !=
 436                              sizeof(struct compat_vcpu_info));
 437                 BUILD_BUG_ON(offsetof(struct vcpu_info, time) !=
 438                              offsetof(struct compat_vcpu_info, time));
 439
 440                 if (data->u.gpa == GPA_INVALID) {
 441                         vcpu->arch.xen.vcpu_info_set = false;
 442                         r = 0;
 443                         break;
 444                 }
 445
 446                 r = kvm_gfn_to_hva_cache_init(vcpu->kvm,
 447                                               &vcpu->arch.xen.vcpu_info_cache,
 448                                               data->u.gpa,
 449                                               sizeof(struct vcpu_info));
 450                 if (!r) {
 451                         vcpu->arch.xen.vcpu_info_set = true;
 452                         kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
 453                 }
 454                 break;
 455
 456         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_TIME_INFO:
 457                 if (data->u.gpa == GPA_INVALID) {
 458                         vcpu->arch.xen.vcpu_time_info_set = false;
 459                         r = 0;
 460                         break;
 461                 }
 462
 463                 r = kvm_gfn_to_hva_cache_init(vcpu->kvm,
 464                                               &vcpu->arch.xen.vcpu_time_info_cache,
 465                                               data->u.gpa,
 466                                               sizeof(struct pvclock_vcpu_time_info));
 467                 if (!r) {
 468                         vcpu->arch.xen.vcpu_time_info_set = true;
 469                         kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
 470                 }
 471                 break;
 472
 473         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADDR:
 474                 if (!sched_info_on()) {
 475                         r = -EOPNOTSUPP;
 476                         break;
 477                 }
 478                 if (data->u.gpa == GPA_INVALID) {
 479                         vcpu->arch.xen.runstate_set = false;
 480                         r = 0;
 481                         break;
 482                 }
 483
 484                 r = kvm_gfn_to_hva_cache_init(vcpu->kvm,
 485                                               &vcpu->arch.xen.runstate_cache,
 486                                               data->u.gpa,
 487                                               sizeof(struct vcpu_runstate_info));
 488                 if (!r) {
 489                         vcpu->arch.xen.runstate_set = true;
 490                 }
 491                 break;
 492
 493         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_CURRENT:
 494                 if (!sched_info_on()) {
 495                         r = -EOPNOTSUPP;
 496                         break;
 497                 }
 498                 if (data->u.runstate.state > RUNSTATE_offline) {
 499                         r = -EINVAL;
 500                         break;
 501                 }
 502
 503                 kvm_xen_update_runstate(vcpu, data->u.runstate.state);
 504                 r = 0;
 505                 break;
 506
 507         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_DATA:
 508                 if (!sched_info_on()) {
 509                         r = -EOPNOTSUPP;
 510                         break;
 511                 }
 512                 if (data->u.runstate.state > RUNSTATE_offline) {
 513                         r = -EINVAL;
 514                         break;
 515                 }
 516                 if (data->u.runstate.state_entry_time !=
 517                     (data->u.runstate.time_running +
 518                      data->u.runstate.time_runnable +
 519                      data->u.runstate.time_blocked +
 520                      data->u.runstate.time_offline)) {
 521                         r = -EINVAL;
 522                         break;
 523                 }
 524                 if (get_kvmclock_ns(vcpu->kvm) <
 525                     data->u.runstate.state_entry_time) {
 526                         r = -EINVAL;
 527                         break;
 528                 }
 529
 530                 vcpu->arch.xen.current_runstate = data->u.runstate.state;
 531                 vcpu->arch.xen.runstate_entry_time =
 532                         data->u.runstate.state_entry_time;
 533                 vcpu->arch.xen.runstate_times[RUNSTATE_running] =
 534                         data->u.runstate.time_running;
 535                 vcpu->arch.xen.runstate_times[RUNSTATE_runnable] =
 536                         data->u.runstate.time_runnable;
 537                 vcpu->arch.xen.runstate_times[RUNSTATE_blocked] =
 538                         data->u.runstate.time_blocked;
 539                 vcpu->arch.xen.runstate_times[RUNSTATE_offline] =
 540                         data->u.runstate.time_offline;
 541                 vcpu->arch.xen.last_steal = current->sched_info.run_delay;
 542                 r = 0;
 543                 break;
 544
 545         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADJUST:
 546                 if (!sched_info_on()) {
 547                         r = -EOPNOTSUPP;
 548                         break;
 549                 }
 550                 if (data->u.runstate.state > RUNSTATE_offline &&
 551                     data->u.runstate.state != (u64)-1) {
 552                         r = -EINVAL;
 553                         break;
 554                 }
 555                 /* The adjustment must add up */
 556                 if (data->u.runstate.state_entry_time !=
 557                     (data->u.runstate.time_running +
 558                      data->u.runstate.time_runnable +
 559                      data->u.runstate.time_blocked +
 560                      data->u.runstate.time_offline)) {
 561                         r = -EINVAL;
 562                         break;
 563                 }
 564
 565                 if (get_kvmclock_ns(vcpu->kvm) <
 566                     (vcpu->arch.xen.runstate_entry_time +
 567                      data->u.runstate.state_entry_time)) {
 568                         r = -EINVAL;
 569                         break;
 570                 }
 571
 572                 vcpu->arch.xen.runstate_entry_time +=
 573                         data->u.runstate.state_entry_time;
 574                 vcpu->arch.xen.runstate_times[RUNSTATE_running] +=
 575                         data->u.runstate.time_running;
 576                 vcpu->arch.xen.runstate_times[RUNSTATE_runnable] +=
 577                         data->u.runstate.time_runnable;
 578                 vcpu->arch.xen.runstate_times[RUNSTATE_blocked] +=
 579                         data->u.runstate.time_blocked;
 580                 vcpu->arch.xen.runstate_times[RUNSTATE_offline] +=
 581                         data->u.runstate.time_offline;
 582
 583                 if (data->u.runstate.state <= RUNSTATE_offline)
 584                         kvm_xen_update_runstate(vcpu, data->u.runstate.state);
 585                 r = 0;
 586                 break;
 587
 588         default:
 589                 break;
 590         }
 591
 592         srcu_read_unlock(&vcpu->kvm->srcu, idx);
 593         mutex_unlock(&vcpu->kvm->lock);
 594         return r;
 595 }
 596
 597 int kvm_xen_vcpu_get_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 598 {
 599         int r = -ENOENT;
 600
 601         mutex_lock(&vcpu->kvm->lock);
 602
 603         switch (data->type) {
 604         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
 605                 if (vcpu->arch.xen.vcpu_info_set)
 606                         data->u.gpa = vcpu->arch.xen.vcpu_info_cache.gpa;
 607                 else
 608                         data->u.gpa = GPA_INVALID;
 609                 r = 0;
 610                 break;
 611
 612         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_TIME_INFO:
 613                 if (vcpu->arch.xen.vcpu_time_info_set)
 614                         data->u.gpa = vcpu->arch.xen.vcpu_time_info_cache.gpa;
 615                 else
 616                         data->u.gpa = GPA_INVALID;
 617                 r = 0;
 618                 break;
 619
 620         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADDR:
 621                 if (!sched_info_on()) {
 622                         r = -EOPNOTSUPP;
 623                         break;
 624                 }
 625                 if (vcpu->arch.xen.runstate_set) {
 626                         data->u.gpa = vcpu->arch.xen.runstate_cache.gpa;
 627                         r = 0;
 628                 }
 629                 break;
 630
 631         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_CURRENT:
 632                 if (!sched_info_on()) {
 633                         r = -EOPNOTSUPP;
 634                         break;
 635                 }
 636                 data->u.runstate.state = vcpu->arch.xen.current_runstate;
 637                 r = 0;
 638                 break;
 639
 640         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_DATA:
 641                 if (!sched_info_on()) {
 642                         r = -EOPNOTSUPP;
 643                         break;
 644                 }
 645                 data->u.runstate.state = vcpu->arch.xen.current_runstate;
 646                 data->u.runstate.state_entry_time =
 647                         vcpu->arch.xen.runstate_entry_time;
 648                 data->u.runstate.time_running =
 649                         vcpu->arch.xen.runstate_times[RUNSTATE_running];
 650                 data->u.runstate.time_runnable =
 651                         vcpu->arch.xen.runstate_times[RUNSTATE_runnable];
 652                 data->u.runstate.time_blocked =
 653                         vcpu->arch.xen.runstate_times[RUNSTATE_blocked];
 654                 data->u.runstate.time_offline =
 655                         vcpu->arch.xen.runstate_times[RUNSTATE_offline];
 656                 r = 0;
 657                 break;
 658
 659         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADJUST:
 660                 r = -EINVAL;
 661                 break;
 662
 663         default:
 664                 break;
 665         }
 666
 667         mutex_unlock(&vcpu->kvm->lock);
 668         return r;
 669 }
 670
 671 int kvm_xen_write_hypercall_page(struct kvm_vcpu *vcpu, u64 data)
 672 {
 673         struct kvm *kvm = vcpu->kvm;
 674         u32 page_num = data & ~PAGE_MASK;
 675         u64 page_addr = data & PAGE_MASK;
 676         bool lm = is_long_mode(vcpu);
 677
 678         /* Latch long_mode for shared_info pages etc. */
 679         vcpu->kvm->arch.xen.long_mode = lm;
 680
 681         /*
 682          * If Xen hypercall intercept is enabled, fill the hypercall
 683          * page with VMCALL/VMMCALL instructions since that's what
 684          * we catch. Else the VMM has provided the hypercall pages
 685          * with instructions of its own choosing, so use those.
 686          */
 687         if (kvm_xen_hypercall_enabled(kvm)) {
 688                 u8 instructions[32];
 689                 int i;
 690
 691                 if (page_num)
 692                         return 1;
 693
 694                 /* mov imm32, %eax */
 695                 instructions[0] = 0xb8;
 696
 697                 /* vmcall / vmmcall */
 698                 kvm_x86_ops.patch_hypercall(vcpu, instructions + 5);
 699
 700                 /* ret */
 701                 instructions[8] = 0xc3;
 702
 703                 /* int3 to pad */
 704                 memset(instructions + 9, 0xcc, sizeof(instructions) - 9);
 705
 706                 for (i = 0; i < PAGE_SIZE / sizeof(instructions); i++) {
 707                         *(u32 *)&instructions[1] = i;
 708                         if (kvm_vcpu_write_guest(vcpu,
 709                                                  page_addr + (i * sizeof(instructions)),
 710                                                  instructions, sizeof(instructions)))
 711                                 return 1;
 712                 }
 713         } else {
 714                 /*
 715                  * Note, truncation is a non-issue as 'lm' is guaranteed to be
 716                  * false for a 32-bit kernel, i.e. when hva_t is only 4 bytes.
 717                  */
 718                 hva_t blob_addr = lm ? kvm->arch.xen_hvm_config.blob_addr_64
 719                                      : kvm->arch.xen_hvm_config.blob_addr_32;
 720                 u8 blob_size = lm ? kvm->arch.xen_hvm_config.blob_size_64
 721                                   : kvm->arch.xen_hvm_config.blob_size_32;
 722                 u8 *page;
 723
 724                 if (page_num >= blob_size)
 725                         return 1;
 726
 727                 blob_addr += page_num * PAGE_SIZE;
 728
 729                 page = memdup_user((u8 __user *)blob_addr, PAGE_SIZE);
 730                 if (IS_ERR(page))
 731                         return PTR_ERR(page);
 732
 733                 if (kvm_vcpu_write_guest(vcpu, page_addr, page, PAGE_SIZE)) {
 734                         kfree(page);
 735                         return 1;
 736                 }
 737         }
 738         return 0;
 739 }
 740
 741 int kvm_xen_hvm_config(struct kvm *kvm, struct kvm_xen_hvm_config *xhc)
 742 {
 743         if (xhc->flags & ~KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL)
 744                 return -EINVAL;
 745
 746         /*
 747          * With hypercall interception the kernel generates its own
 748          * hypercall page so it must not be provided.
 749          */
 750         if ((xhc->flags & KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL) &&
 751             (xhc->blob_addr_32 || xhc->blob_addr_64 ||
 752              xhc->blob_size_32 || xhc->blob_size_64))
 753                 return -EINVAL;
 754
 755         mutex_lock(&kvm->lock);
 756
 757         if (xhc->msr && !kvm->arch.xen_hvm_config.msr)
 758                 static_branch_inc(&kvm_xen_enabled.key);
 759         else if (!xhc->msr && kvm->arch.xen_hvm_config.msr)
 760                 static_branch_slow_dec_deferred(&kvm_xen_enabled);
 761
 762         memcpy(&kvm->arch.xen_hvm_config, xhc, sizeof(*xhc));
 763
 764         mutex_unlock(&kvm->lock);
 765         return 0;
 766 }
 767
 768 void kvm_xen_init_vm(struct kvm *kvm)
 769 {
 770 }
 771
 772 void kvm_xen_destroy_vm(struct kvm *kvm)
 773 {
 774         kvm_gfn_to_pfn_cache_destroy(kvm, &kvm->arch.xen.shinfo_cache);
 775
 776         if (kvm->arch.xen_hvm_config.msr)
 777                 static_branch_slow_dec_deferred(&kvm_xen_enabled);
 778 }
 779
 780 static int kvm_xen_hypercall_set_result(struct kvm_vcpu *vcpu, u64 result)
 781 {
 782         kvm_rax_write(vcpu, result);
 783         return kvm_skip_emulated_instruction(vcpu);
 784 }
 785
 786 static int kvm_xen_hypercall_complete_userspace(struct kvm_vcpu *vcpu)
 787 {
 788         struct kvm_run *run = vcpu->run;
 789
 790         if (unlikely(!kvm_is_linear_rip(vcpu, vcpu->arch.xen.hypercall_rip)))
 791                 return 1;
 792
 793         return kvm_xen_hypercall_set_result(vcpu, run->xen.u.hcall.result);
 794 }
 795
 796 int kvm_xen_hypercall(struct kvm_vcpu *vcpu)
 797 {
 798         bool longmode;
 799         u64 input, params[6];
 800
 801         input = (u64)kvm_register_read(vcpu, VCPU_REGS_RAX);
 802
 803         /* Hyper-V hypercalls get bit 31 set in EAX */
 804         if ((input & 0x80000000) &&
 805             kvm_hv_hypercall_enabled(vcpu))
 806                 return kvm_hv_hypercall(vcpu);
 807
 808         longmode = is_64_bit_hypercall(vcpu);
 809         if (!longmode) {
 810                 params[0] = (u32)kvm_rbx_read(vcpu);
 811                 params[1] = (u32)kvm_rcx_read(vcpu);
 812                 params[2] = (u32)kvm_rdx_read(vcpu);
 813                 params[3] = (u32)kvm_rsi_read(vcpu);
 814                 params[4] = (u32)kvm_rdi_read(vcpu);
 815                 params[5] = (u32)kvm_rbp_read(vcpu);
 816         }
 817 #ifdef CONFIG_X86_64
 818         else {
 819                 params[0] = (u64)kvm_rdi_read(vcpu);
 820                 params[1] = (u64)kvm_rsi_read(vcpu);
 821                 params[2] = (u64)kvm_rdx_read(vcpu);
 822                 params[3] = (u64)kvm_r10_read(vcpu);
 823                 params[4] = (u64)kvm_r8_read(vcpu);
 824                 params[5] = (u64)kvm_r9_read(vcpu);
 825         }
 826 #endif
 827         trace_kvm_xen_hypercall(input, params[0], params[1], params[2],
 828                                 params[3], params[4], params[5]);
 829
 830         vcpu->run->exit_reason = KVM_EXIT_XEN;
 831         vcpu->run->xen.type = KVM_EXIT_XEN_HCALL;
 832         vcpu->run->xen.u.hcall.longmode = longmode;
 833         vcpu->run->xen.u.hcall.cpl = kvm_x86_ops.get_cpl(vcpu);
 834         vcpu->run->xen.u.hcall.input = input;
 835         vcpu->run->xen.u.hcall.params[0] = params[0];
 836         vcpu->run->xen.u.hcall.params[1] = params[1];
 837         vcpu->run->xen.u.hcall.params[2] = params[2];
 838         vcpu->run->xen.u.hcall.params[3] = params[3];
 839         vcpu->run->xen.u.hcall.params[4] = params[4];
 840         vcpu->run->xen.u.hcall.params[5] = params[5];
 841         vcpu->arch.xen.hypercall_rip = kvm_get_linear_rip(vcpu);
 842         vcpu->arch.complete_userspace_io =
 843                 kvm_xen_hypercall_complete_userspace;
 844
 845         return 0;
 846 }
 847
 848 static inline int max_evtchn_port(struct kvm *kvm)
 849 {
 850         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode)
 851                 return EVTCHN_2L_NR_CHANNELS;
 852         else
 853                 return COMPAT_EVTCHN_2L_NR_CHANNELS;
 854 }
 855
 856 /*
 857  * This follows the kvm_set_irq() API, so it returns:
 858  *  < 0   Interrupt was ignored (masked or not delivered for other reasons)
 859  *  = 0   Interrupt was coalesced (previous irq is still pending)
 860  *  > 0   Number of CPUs interrupt was delivered to
 861  */
 862 int kvm_xen_set_evtchn_fast(struct kvm_kernel_irq_routing_entry *e,
 863                             struct kvm *kvm)
 864 {
 865         struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
 866         struct kvm_vcpu *vcpu;
 867         unsigned long *pending_bits, *mask_bits;
 868         unsigned long flags;
 869         int port_word_bit;
 870         bool kick_vcpu = false;
 871         int idx;
 872         int rc;
 873
 874         vcpu = kvm_get_vcpu_by_id(kvm, e->xen_evtchn.vcpu);
 875         if (!vcpu)
 876                 return -1;
 877
 878         if (!vcpu->arch.xen.vcpu_info_set)
 879                 return -1;
 880
 881         if (e->xen_evtchn.port >= max_evtchn_port(kvm))
 882                 return -1;
 883
 884         rc = -EWOULDBLOCK;
 885         read_lock_irqsave(&gpc->lock, flags);
 886
 887         idx = srcu_read_lock(&kvm->srcu);
 888         if (!kvm_gfn_to_pfn_cache_check(kvm, gpc, gpc->gpa, PAGE_SIZE))
 889                 goto out_rcu;
 890
 891         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
 892                 struct shared_info *shinfo = gpc->khva;
 893                 pending_bits = (unsigned long *)&shinfo->evtchn_pending;
 894                 mask_bits = (unsigned long *)&shinfo->evtchn_mask;
 895                 port_word_bit = e->xen_evtchn.port / 64;
 896         } else {
 897                 struct compat_shared_info *shinfo = gpc->khva;
 898                 pending_bits = (unsigned long *)&shinfo->evtchn_pending;
 899                 mask_bits = (unsigned long *)&shinfo->evtchn_mask;
 900                 port_word_bit = e->xen_evtchn.port / 32;
 901         }
 902
 903         /*
 904          * If this port wasn't already set, and if it isn't masked, then
 905          * we try to set the corresponding bit in the in-kernel shadow of
 906          * evtchn_pending_sel for the target vCPU. And if *that* wasn't
 907          * already set, then we kick the vCPU in question to write to the
 908          * *real* evtchn_pending_sel in its own guest vcpu_info struct.
 909          */
 910         if (test_and_set_bit(e->xen_evtchn.port, pending_bits)) {
 911                 rc = 0; /* It was already raised */
 912         } else if (test_bit(e->xen_evtchn.port, mask_bits)) {
 913                 rc = -1; /* Masked */
 914         } else {
 915                 rc = 1; /* Delivered. But was the vCPU waking already? */
 916                 if (!test_and_set_bit(port_word_bit, &vcpu->arch.xen.evtchn_pending_sel))
 917                         kick_vcpu = true;
 918         }
 919
 920  out_rcu:
 921         srcu_read_unlock(&kvm->srcu, idx);
 922         read_unlock_irqrestore(&gpc->lock, flags);
 923
 924         if (kick_vcpu) {
 925                 kvm_make_request(KVM_REQ_EVENT, vcpu);
 926                 kvm_vcpu_kick(vcpu);
 927         }
 928
 929         return rc;
 930 }
 931
 932 /* This is the version called from kvm_set_irq() as the .set function */
 933 static int evtchn_set_fn(struct kvm_kernel_irq_routing_entry *e, struct kvm *kvm,
 934                          int irq_source_id, int level, bool line_status)
 935 {
 936         bool mm_borrowed = false;
 937         int rc;
 938
 939         if (!level)
 940                 return -1;
 941
 942         rc = kvm_xen_set_evtchn_fast(e, kvm);
 943         if (rc != -EWOULDBLOCK)
 944                 return rc;
 945
 946         if (current->mm != kvm->mm) {
 947                 /*
 948                  * If not on a thread which already belongs to this KVM,
 949                  * we'd better be in the irqfd workqueue.
 950                  */
 951                 if (WARN_ON_ONCE(current->mm))
 952                         return -EINVAL;
 953
 954                 kthread_use_mm(kvm->mm);
 955                 mm_borrowed = true;
 956         }
 957
 958         /*
 959          * For the irqfd workqueue, using the main kvm->lock mutex is
 960          * fine since this function is invoked from kvm_set_irq() with
 961          * no other lock held, no srcu. In future if it will be called
 962          * directly from a vCPU thread (e.g. on hypercall for an IPI)
 963          * then it may need to switch to using a leaf-node mutex for
 964          * serializing the shared_info mapping.
 965          */
 966         mutex_lock(&kvm->lock);
 967
 968         /*
 969          * It is theoretically possible for the page to be unmapped
 970          * and the MMU notifier to invalidate the shared_info before
 971          * we even get to use it. In that case, this looks like an
 972          * infinite loop. It was tempting to do it via the userspace
 973          * HVA instead... but that just *hides* the fact that it's
 974          * an infinite loop, because if a fault occurs and it waits
 975          * for the page to come back, it can *still* immediately
 976          * fault and have to wait again, repeatedly.
 977          *
 978          * Conversely, the page could also have been reinstated by
 979          * another thread before we even obtain the mutex above, so
 980          * check again *first* before remapping it.
 981          */
 982         do {
 983                 struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
 984                 int idx;
 985
 986                 rc = kvm_xen_set_evtchn_fast(e, kvm);
 987                 if (rc != -EWOULDBLOCK)
 988                         break;
 989
 990                 idx = srcu_read_lock(&kvm->srcu);
 991                 rc = kvm_gfn_to_pfn_cache_refresh(kvm, gpc, gpc->gpa,
 992                                                   PAGE_SIZE, false);
 993                 srcu_read_unlock(&kvm->srcu, idx);
 994         } while(!rc);
 995
 996         mutex_unlock(&kvm->lock);
 997
 998         if (mm_borrowed)
 999                 kthread_unuse_mm(kvm->mm);
1000
1001         return rc;
1002 }
1003
1004 int kvm_xen_setup_evtchn(struct kvm *kvm,
1005                          struct kvm_kernel_irq_routing_entry *e,
1006                          const struct kvm_irq_routing_entry *ue)
1007
1008 {
1009         if (ue->u.xen_evtchn.port >= max_evtchn_port(kvm))
1010                 return -EINVAL;
1011
1012         /* We only support 2 level event channels for now */
1013         if (ue->u.xen_evtchn.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1014                 return -EINVAL;
1015
1016         e->xen_evtchn.port = ue->u.xen_evtchn.port;
1017         e->xen_evtchn.vcpu = ue->u.xen_evtchn.vcpu;
1018         e->xen_evtchn.priority = ue->u.xen_evtchn.priority;
1019         e->set = evtchn_set_fn;
1020
1021         return 0;
1022 }