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, KVM_HOST_USES_PFN,
  43                                                 gpa, PAGE_SIZE);
  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         struct gfn_to_hva_cache *ghc = &vx->runstate_cache;
 137         struct kvm_memslots *slots = kvm_memslots(v->kvm);
 138         bool atomic = (state == RUNSTATE_runnable);
 139         uint64_t state_entry_time;
 140         int __user *user_state;
 141         uint64_t __user *user_times;
 142
 143         kvm_xen_update_runstate(v, state);
 144
 145         if (!vx->runstate_set)
 146                 return;
 147
 148         if (unlikely(slots->generation != ghc->generation || kvm_is_error_hva(ghc->hva)) &&
 149             kvm_gfn_to_hva_cache_init(v->kvm, ghc, ghc->gpa, ghc->len))
 150                 return;
 151
 152         /* We made sure it fits in a single page */
 153         BUG_ON(!ghc->memslot);
 154
 155         if (atomic)
 156                 pagefault_disable();
 157
 158         /*
 159          * The only difference between 32-bit and 64-bit versions of the
 160          * runstate struct us the alignment of uint64_t in 32-bit, which
 161          * means that the 64-bit version has an additional 4 bytes of
 162          * padding after the first field 'state'.
 163          *
 164          * So we use 'int __user *user_state' to point to the state field,
 165          * and 'uint64_t __user *user_times' for runstate_entry_time. So
 166          * the actual array of time[] in each state starts at user_times[1].
 167          */
 168         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state) != 0);
 169         BUILD_BUG_ON(offsetof(struct compat_vcpu_runstate_info, state) != 0);
 170         user_state = (int __user *)ghc->hva;
 171
 172         BUILD_BUG_ON(sizeof(struct compat_vcpu_runstate_info) != 0x2c);
 173
 174         user_times = (uint64_t __user *)(ghc->hva +
 175                                          offsetof(struct compat_vcpu_runstate_info,
 176                                                   state_entry_time));
 177 #ifdef CONFIG_X86_64
 178         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state_entry_time) !=
 179                      offsetof(struct compat_vcpu_runstate_info, state_entry_time) + 4);
 180         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, time) !=
 181                      offsetof(struct compat_vcpu_runstate_info, time) + 4);
 182
 183         if (v->kvm->arch.xen.long_mode)
 184                 user_times = (uint64_t __user *)(ghc->hva +
 185                                                  offsetof(struct vcpu_runstate_info,
 186                                                           state_entry_time));
 187 #endif
 188         /*
 189          * First write the updated state_entry_time at the appropriate
 190          * location determined by 'offset'.
 191          */
 192         state_entry_time = vx->runstate_entry_time;
 193         state_entry_time |= XEN_RUNSTATE_UPDATE;
 194
 195         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, state_entry_time) !=
 196                      sizeof(state_entry_time));
 197         BUILD_BUG_ON(sizeof_field(struct compat_vcpu_runstate_info, state_entry_time) !=
 198                      sizeof(state_entry_time));
 199
 200         if (__put_user(state_entry_time, user_times))
 201                 goto out;
 202         smp_wmb();
 203
 204         /*
 205          * Next, write the new runstate. This is in the *same* place
 206          * for 32-bit and 64-bit guests, asserted here for paranoia.
 207          */
 208         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state) !=
 209                      offsetof(struct compat_vcpu_runstate_info, state));
 210         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, state) !=
 211                      sizeof(vx->current_runstate));
 212         BUILD_BUG_ON(sizeof_field(struct compat_vcpu_runstate_info, state) !=
 213                      sizeof(vx->current_runstate));
 214
 215         if (__put_user(vx->current_runstate, user_state))
 216                 goto out;
 217
 218         /*
 219          * Write the actual runstate times immediately after the
 220          * runstate_entry_time.
 221          */
 222         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state_entry_time) !=
 223                      offsetof(struct vcpu_runstate_info, time) - sizeof(u64));
 224         BUILD_BUG_ON(offsetof(struct compat_vcpu_runstate_info, state_entry_time) !=
 225                      offsetof(struct compat_vcpu_runstate_info, time) - sizeof(u64));
 226         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, time) !=
 227                      sizeof_field(struct compat_vcpu_runstate_info, time));
 228         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, time) !=
 229                      sizeof(vx->runstate_times));
 230
 231         if (__copy_to_user(user_times + 1, vx->runstate_times, sizeof(vx->runstate_times)))
 232                 goto out;
 233         smp_wmb();
 234
 235         /*
 236          * Finally, clear the XEN_RUNSTATE_UPDATE bit in the guest's
 237          * runstate_entry_time field.
 238          */
 239         state_entry_time &= ~XEN_RUNSTATE_UPDATE;
 240         __put_user(state_entry_time, user_times);
 241         smp_wmb();
 242
 243  out:
 244         mark_page_dirty_in_slot(v->kvm, ghc->memslot, ghc->gpa >> PAGE_SHIFT);
 245
 246         if (atomic)
 247                 pagefault_enable();
 248 }
 249
 250 int __kvm_xen_has_interrupt(struct kvm_vcpu *v)
 251 {
 252         unsigned long evtchn_pending_sel = READ_ONCE(v->arch.xen.evtchn_pending_sel);
 253         bool atomic = in_atomic() || !task_is_running(current);
 254         int err;
 255         u8 rc = 0;
 256
 257         /*
 258          * If the global upcall vector (HVMIRQ_callback_vector) is set and
 259          * the vCPU's evtchn_upcall_pending flag is set, the IRQ is pending.
 260          */
 261         struct gfn_to_hva_cache *ghc = &v->arch.xen.vcpu_info_cache;
 262         struct kvm_memslots *slots = kvm_memslots(v->kvm);
 263         bool ghc_valid = slots->generation == ghc->generation &&
 264                 !kvm_is_error_hva(ghc->hva) && ghc->memslot;
 265
 266         unsigned int offset = offsetof(struct vcpu_info, evtchn_upcall_pending);
 267
 268         /* No need for compat handling here */
 269         BUILD_BUG_ON(offsetof(struct vcpu_info, evtchn_upcall_pending) !=
 270                      offsetof(struct compat_vcpu_info, evtchn_upcall_pending));
 271         BUILD_BUG_ON(sizeof(rc) !=
 272                      sizeof_field(struct vcpu_info, evtchn_upcall_pending));
 273         BUILD_BUG_ON(sizeof(rc) !=
 274                      sizeof_field(struct compat_vcpu_info, evtchn_upcall_pending));
 275
 276         /*
 277          * For efficiency, this mirrors the checks for using the valid
 278          * cache in kvm_read_guest_offset_cached(), but just uses
 279          * __get_user() instead. And falls back to the slow path.
 280          */
 281         if (!evtchn_pending_sel && ghc_valid) {
 282                 /* Fast path */
 283                 pagefault_disable();
 284                 err = __get_user(rc, (u8 __user *)ghc->hva + offset);
 285                 pagefault_enable();
 286                 if (!err)
 287                         return rc;
 288         }
 289
 290         /* Slow path */
 291
 292         /*
 293          * This function gets called from kvm_vcpu_block() after setting the
 294          * task to TASK_INTERRUPTIBLE, to see if it needs to wake immediately
 295          * from a HLT. So we really mustn't sleep. If the page ended up absent
 296          * at that point, just return 1 in order to trigger an immediate wake,
 297          * and we'll end up getting called again from a context where we *can*
 298          * fault in the page and wait for it.
 299          */
 300         if (atomic)
 301                 return 1;
 302
 303         if (!ghc_valid) {
 304                 err = kvm_gfn_to_hva_cache_init(v->kvm, ghc, ghc->gpa, ghc->len);
 305                 if (err || !ghc->memslot) {
 306                         /*
 307                          * If this failed, userspace has screwed up the
 308                          * vcpu_info mapping. No interrupts for you.
 309                          */
 310                         return 0;
 311                 }
 312         }
 313
 314         /*
 315          * Now we have a valid (protected by srcu) userspace HVA in
 316          * ghc->hva which points to the struct vcpu_info. If there
 317          * are any bits in the in-kernel evtchn_pending_sel then
 318          * we need to write those to the guest vcpu_info and set
 319          * its evtchn_upcall_pending flag. If there aren't any bits
 320          * to add, we only want to *check* evtchn_upcall_pending.
 321          */
 322         if (evtchn_pending_sel) {
 323                 bool long_mode = v->kvm->arch.xen.long_mode;
 324
 325                 if (!user_access_begin((void __user *)ghc->hva, sizeof(struct vcpu_info)))
 326                         return 0;
 327
 328                 if (IS_ENABLED(CONFIG_64BIT) && long_mode) {
 329                         struct vcpu_info __user *vi = (void __user *)ghc->hva;
 330
 331                         /* Attempt to set the evtchn_pending_sel bits in the
 332                          * guest, and if that succeeds then clear the same
 333                          * bits in the in-kernel version. */
 334                         asm volatile("1:\t" LOCK_PREFIX "orq %0, %1\n"
 335                                      "\tnotq %0\n"
 336                                      "\t" LOCK_PREFIX "andq %0, %2\n"
 337                                      "2:\n"
 338                                      _ASM_EXTABLE_UA(1b, 2b)
 339                                      : "=r" (evtchn_pending_sel),
 340                                        "+m" (vi->evtchn_pending_sel),
 341                                        "+m" (v->arch.xen.evtchn_pending_sel)
 342                                      : "0" (evtchn_pending_sel));
 343                 } else {
 344                         struct compat_vcpu_info __user *vi = (void __user *)ghc->hva;
 345                         u32 evtchn_pending_sel32 = evtchn_pending_sel;
 346
 347                         /* Attempt to set the evtchn_pending_sel bits in the
 348                          * guest, and if that succeeds then clear the same
 349                          * bits in the in-kernel version. */
 350                         asm volatile("1:\t" LOCK_PREFIX "orl %0, %1\n"
 351                                      "\tnotl %0\n"
 352                                      "\t" LOCK_PREFIX "andl %0, %2\n"
 353                                      "2:\n"
 354                                      _ASM_EXTABLE_UA(1b, 2b)
 355                                      : "=r" (evtchn_pending_sel32),
 356                                        "+m" (vi->evtchn_pending_sel),
 357                                        "+m" (v->arch.xen.evtchn_pending_sel)
 358                                      : "0" (evtchn_pending_sel32));
 359                 }
 360                 rc = 1;
 361                 unsafe_put_user(rc, (u8 __user *)ghc->hva + offset, err);
 362
 363         err:
 364                 user_access_end();
 365
 366                 mark_page_dirty_in_slot(v->kvm, ghc->memslot, ghc->gpa >> PAGE_SHIFT);
 367         } else {
 368                 __get_user(rc, (u8 __user *)ghc->hva + offset);
 369         }
 370
 371         return rc;
 372 }
 373
 374 int kvm_xen_hvm_set_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 375 {
 376         int r = -ENOENT;
 377
 378         mutex_lock(&kvm->lock);
 379
 380         switch (data->type) {
 381         case KVM_XEN_ATTR_TYPE_LONG_MODE:
 382                 if (!IS_ENABLED(CONFIG_64BIT) && data->u.long_mode) {
 383                         r = -EINVAL;
 384                 } else {
 385                         kvm->arch.xen.long_mode = !!data->u.long_mode;
 386                         r = 0;
 387                 }
 388                 break;
 389
 390         case KVM_XEN_ATTR_TYPE_SHARED_INFO:
 391                 r = kvm_xen_shared_info_init(kvm, data->u.shared_info.gfn);
 392                 break;
 393
 394         case KVM_XEN_ATTR_TYPE_UPCALL_VECTOR:
 395                 if (data->u.vector && data->u.vector < 0x10)
 396                         r = -EINVAL;
 397                 else {
 398                         kvm->arch.xen.upcall_vector = data->u.vector;
 399                         r = 0;
 400                 }
 401                 break;
 402
 403         default:
 404                 break;
 405         }
 406
 407         mutex_unlock(&kvm->lock);
 408         return r;
 409 }
 410
 411 int kvm_xen_hvm_get_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 412 {
 413         int r = -ENOENT;
 414
 415         mutex_lock(&kvm->lock);
 416
 417         switch (data->type) {
 418         case KVM_XEN_ATTR_TYPE_LONG_MODE:
 419                 data->u.long_mode = kvm->arch.xen.long_mode;
 420                 r = 0;
 421                 break;
 422
 423         case KVM_XEN_ATTR_TYPE_SHARED_INFO:
 424                 if (kvm->arch.xen.shinfo_cache.active)
 425                         data->u.shared_info.gfn = gpa_to_gfn(kvm->arch.xen.shinfo_cache.gpa);
 426                 else
 427                         data->u.shared_info.gfn = GPA_INVALID;
 428                 r = 0;
 429                 break;
 430
 431         case KVM_XEN_ATTR_TYPE_UPCALL_VECTOR:
 432                 data->u.vector = kvm->arch.xen.upcall_vector;
 433                 r = 0;
 434                 break;
 435
 436         default:
 437                 break;
 438         }
 439
 440         mutex_unlock(&kvm->lock);
 441         return r;
 442 }
 443
 444 int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 445 {
 446         int idx, r = -ENOENT;
 447
 448         mutex_lock(&vcpu->kvm->lock);
 449         idx = srcu_read_lock(&vcpu->kvm->srcu);
 450
 451         switch (data->type) {
 452         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
 453                 /* No compat necessary here. */
 454                 BUILD_BUG_ON(sizeof(struct vcpu_info) !=
 455                              sizeof(struct compat_vcpu_info));
 456                 BUILD_BUG_ON(offsetof(struct vcpu_info, time) !=
 457                              offsetof(struct compat_vcpu_info, time));
 458
 459                 if (data->u.gpa == GPA_INVALID) {
 460                         vcpu->arch.xen.vcpu_info_set = false;
 461                         r = 0;
 462                         break;
 463                 }
 464
 465                 /* It must fit within a single page */
 466                 if ((data->u.gpa & ~PAGE_MASK) + sizeof(struct vcpu_info) > PAGE_SIZE) {
 467                         r = -EINVAL;
 468                         break;
 469                 }
 470
 471                 r = kvm_gfn_to_hva_cache_init(vcpu->kvm,
 472                                               &vcpu->arch.xen.vcpu_info_cache,
 473                                               data->u.gpa,
 474                                               sizeof(struct vcpu_info));
 475                 if (!r) {
 476                         vcpu->arch.xen.vcpu_info_set = true;
 477                         kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
 478                 }
 479                 break;
 480
 481         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_TIME_INFO:
 482                 if (data->u.gpa == GPA_INVALID) {
 483                         vcpu->arch.xen.vcpu_time_info_set = false;
 484                         r = 0;
 485                         break;
 486                 }
 487
 488                 /* It must fit within a single page */
 489                 if ((data->u.gpa & ~PAGE_MASK) + sizeof(struct pvclock_vcpu_time_info) > PAGE_SIZE) {
 490                         r = -EINVAL;
 491                         break;
 492                 }
 493
 494                 r = kvm_gfn_to_hva_cache_init(vcpu->kvm,
 495                                               &vcpu->arch.xen.vcpu_time_info_cache,
 496                                               data->u.gpa,
 497                                               sizeof(struct pvclock_vcpu_time_info));
 498                 if (!r) {
 499                         vcpu->arch.xen.vcpu_time_info_set = true;
 500                         kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
 501                 }
 502                 break;
 503
 504         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADDR:
 505                 if (!sched_info_on()) {
 506                         r = -EOPNOTSUPP;
 507                         break;
 508                 }
 509                 if (data->u.gpa == GPA_INVALID) {
 510                         vcpu->arch.xen.runstate_set = false;
 511                         r = 0;
 512                         break;
 513                 }
 514
 515                 /* It must fit within a single page */
 516                 if ((data->u.gpa & ~PAGE_MASK) + sizeof(struct vcpu_runstate_info) > PAGE_SIZE) {
 517                         r = -EINVAL;
 518                         break;
 519                 }
 520
 521                 r = kvm_gfn_to_hva_cache_init(vcpu->kvm,
 522                                               &vcpu->arch.xen.runstate_cache,
 523                                               data->u.gpa,
 524                                               sizeof(struct vcpu_runstate_info));
 525                 if (!r) {
 526                         vcpu->arch.xen.runstate_set = true;
 527                 }
 528                 break;
 529
 530         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_CURRENT:
 531                 if (!sched_info_on()) {
 532                         r = -EOPNOTSUPP;
 533                         break;
 534                 }
 535                 if (data->u.runstate.state > RUNSTATE_offline) {
 536                         r = -EINVAL;
 537                         break;
 538                 }
 539
 540                 kvm_xen_update_runstate(vcpu, data->u.runstate.state);
 541                 r = 0;
 542                 break;
 543
 544         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_DATA:
 545                 if (!sched_info_on()) {
 546                         r = -EOPNOTSUPP;
 547                         break;
 548                 }
 549                 if (data->u.runstate.state > RUNSTATE_offline) {
 550                         r = -EINVAL;
 551                         break;
 552                 }
 553                 if (data->u.runstate.state_entry_time !=
 554                     (data->u.runstate.time_running +
 555                      data->u.runstate.time_runnable +
 556                      data->u.runstate.time_blocked +
 557                      data->u.runstate.time_offline)) {
 558                         r = -EINVAL;
 559                         break;
 560                 }
 561                 if (get_kvmclock_ns(vcpu->kvm) <
 562                     data->u.runstate.state_entry_time) {
 563                         r = -EINVAL;
 564                         break;
 565                 }
 566
 567                 vcpu->arch.xen.current_runstate = data->u.runstate.state;
 568                 vcpu->arch.xen.runstate_entry_time =
 569                         data->u.runstate.state_entry_time;
 570                 vcpu->arch.xen.runstate_times[RUNSTATE_running] =
 571                         data->u.runstate.time_running;
 572                 vcpu->arch.xen.runstate_times[RUNSTATE_runnable] =
 573                         data->u.runstate.time_runnable;
 574                 vcpu->arch.xen.runstate_times[RUNSTATE_blocked] =
 575                         data->u.runstate.time_blocked;
 576                 vcpu->arch.xen.runstate_times[RUNSTATE_offline] =
 577                         data->u.runstate.time_offline;
 578                 vcpu->arch.xen.last_steal = current->sched_info.run_delay;
 579                 r = 0;
 580                 break;
 581
 582         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADJUST:
 583                 if (!sched_info_on()) {
 584                         r = -EOPNOTSUPP;
 585                         break;
 586                 }
 587                 if (data->u.runstate.state > RUNSTATE_offline &&
 588                     data->u.runstate.state != (u64)-1) {
 589                         r = -EINVAL;
 590                         break;
 591                 }
 592                 /* The adjustment must add up */
 593                 if (data->u.runstate.state_entry_time !=
 594                     (data->u.runstate.time_running +
 595                      data->u.runstate.time_runnable +
 596                      data->u.runstate.time_blocked +
 597                      data->u.runstate.time_offline)) {
 598                         r = -EINVAL;
 599                         break;
 600                 }
 601
 602                 if (get_kvmclock_ns(vcpu->kvm) <
 603                     (vcpu->arch.xen.runstate_entry_time +
 604                      data->u.runstate.state_entry_time)) {
 605                         r = -EINVAL;
 606                         break;
 607                 }
 608
 609                 vcpu->arch.xen.runstate_entry_time +=
 610                         data->u.runstate.state_entry_time;
 611                 vcpu->arch.xen.runstate_times[RUNSTATE_running] +=
 612                         data->u.runstate.time_running;
 613                 vcpu->arch.xen.runstate_times[RUNSTATE_runnable] +=
 614                         data->u.runstate.time_runnable;
 615                 vcpu->arch.xen.runstate_times[RUNSTATE_blocked] +=
 616                         data->u.runstate.time_blocked;
 617                 vcpu->arch.xen.runstate_times[RUNSTATE_offline] +=
 618                         data->u.runstate.time_offline;
 619
 620                 if (data->u.runstate.state <= RUNSTATE_offline)
 621                         kvm_xen_update_runstate(vcpu, data->u.runstate.state);
 622                 r = 0;
 623                 break;
 624
 625         default:
 626                 break;
 627         }
 628
 629         srcu_read_unlock(&vcpu->kvm->srcu, idx);
 630         mutex_unlock(&vcpu->kvm->lock);
 631         return r;
 632 }
 633
 634 int kvm_xen_vcpu_get_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 635 {
 636         int r = -ENOENT;
 637
 638         mutex_lock(&vcpu->kvm->lock);
 639
 640         switch (data->type) {
 641         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
 642                 if (vcpu->arch.xen.vcpu_info_set)
 643                         data->u.gpa = vcpu->arch.xen.vcpu_info_cache.gpa;
 644                 else
 645                         data->u.gpa = GPA_INVALID;
 646                 r = 0;
 647                 break;
 648
 649         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_TIME_INFO:
 650                 if (vcpu->arch.xen.vcpu_time_info_set)
 651                         data->u.gpa = vcpu->arch.xen.vcpu_time_info_cache.gpa;
 652                 else
 653                         data->u.gpa = GPA_INVALID;
 654                 r = 0;
 655                 break;
 656
 657         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADDR:
 658                 if (!sched_info_on()) {
 659                         r = -EOPNOTSUPP;
 660                         break;
 661                 }
 662                 if (vcpu->arch.xen.runstate_set) {
 663                         data->u.gpa = vcpu->arch.xen.runstate_cache.gpa;
 664                         r = 0;
 665                 }
 666                 break;
 667
 668         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_CURRENT:
 669                 if (!sched_info_on()) {
 670                         r = -EOPNOTSUPP;
 671                         break;
 672                 }
 673                 data->u.runstate.state = vcpu->arch.xen.current_runstate;
 674                 r = 0;
 675                 break;
 676
 677         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_DATA:
 678                 if (!sched_info_on()) {
 679                         r = -EOPNOTSUPP;
 680                         break;
 681                 }
 682                 data->u.runstate.state = vcpu->arch.xen.current_runstate;
 683                 data->u.runstate.state_entry_time =
 684                         vcpu->arch.xen.runstate_entry_time;
 685                 data->u.runstate.time_running =
 686                         vcpu->arch.xen.runstate_times[RUNSTATE_running];
 687                 data->u.runstate.time_runnable =
 688                         vcpu->arch.xen.runstate_times[RUNSTATE_runnable];
 689                 data->u.runstate.time_blocked =
 690                         vcpu->arch.xen.runstate_times[RUNSTATE_blocked];
 691                 data->u.runstate.time_offline =
 692                         vcpu->arch.xen.runstate_times[RUNSTATE_offline];
 693                 r = 0;
 694                 break;
 695
 696         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADJUST:
 697                 r = -EINVAL;
 698                 break;
 699
 700         default:
 701                 break;
 702         }
 703
 704         mutex_unlock(&vcpu->kvm->lock);
 705         return r;
 706 }
 707
 708 int kvm_xen_write_hypercall_page(struct kvm_vcpu *vcpu, u64 data)
 709 {
 710         struct kvm *kvm = vcpu->kvm;
 711         u32 page_num = data & ~PAGE_MASK;
 712         u64 page_addr = data & PAGE_MASK;
 713         bool lm = is_long_mode(vcpu);
 714
 715         /* Latch long_mode for shared_info pages etc. */
 716         vcpu->kvm->arch.xen.long_mode = lm;
 717
 718         /*
 719          * If Xen hypercall intercept is enabled, fill the hypercall
 720          * page with VMCALL/VMMCALL instructions since that's what
 721          * we catch. Else the VMM has provided the hypercall pages
 722          * with instructions of its own choosing, so use those.
 723          */
 724         if (kvm_xen_hypercall_enabled(kvm)) {
 725                 u8 instructions[32];
 726                 int i;
 727
 728                 if (page_num)
 729                         return 1;
 730
 731                 /* mov imm32, %eax */
 732                 instructions[0] = 0xb8;
 733
 734                 /* vmcall / vmmcall */
 735                 static_call(kvm_x86_patch_hypercall)(vcpu, instructions + 5);
 736
 737                 /* ret */
 738                 instructions[8] = 0xc3;
 739
 740                 /* int3 to pad */
 741                 memset(instructions + 9, 0xcc, sizeof(instructions) - 9);
 742
 743                 for (i = 0; i < PAGE_SIZE / sizeof(instructions); i++) {
 744                         *(u32 *)&instructions[1] = i;
 745                         if (kvm_vcpu_write_guest(vcpu,
 746                                                  page_addr + (i * sizeof(instructions)),
 747                                                  instructions, sizeof(instructions)))
 748                                 return 1;
 749                 }
 750         } else {
 751                 /*
 752                  * Note, truncation is a non-issue as 'lm' is guaranteed to be
 753                  * false for a 32-bit kernel, i.e. when hva_t is only 4 bytes.
 754                  */
 755                 hva_t blob_addr = lm ? kvm->arch.xen_hvm_config.blob_addr_64
 756                                      : kvm->arch.xen_hvm_config.blob_addr_32;
 757                 u8 blob_size = lm ? kvm->arch.xen_hvm_config.blob_size_64
 758                                   : kvm->arch.xen_hvm_config.blob_size_32;
 759                 u8 *page;
 760
 761                 if (page_num >= blob_size)
 762                         return 1;
 763
 764                 blob_addr += page_num * PAGE_SIZE;
 765
 766                 page = memdup_user((u8 __user *)blob_addr, PAGE_SIZE);
 767                 if (IS_ERR(page))
 768                         return PTR_ERR(page);
 769
 770                 if (kvm_vcpu_write_guest(vcpu, page_addr, page, PAGE_SIZE)) {
 771                         kfree(page);
 772                         return 1;
 773                 }
 774         }
 775         return 0;
 776 }
 777
 778 int kvm_xen_hvm_config(struct kvm *kvm, struct kvm_xen_hvm_config *xhc)
 779 {
 780         if (xhc->flags & ~KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL)
 781                 return -EINVAL;
 782
 783         /*
 784          * With hypercall interception the kernel generates its own
 785          * hypercall page so it must not be provided.
 786          */
 787         if ((xhc->flags & KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL) &&
 788             (xhc->blob_addr_32 || xhc->blob_addr_64 ||
 789              xhc->blob_size_32 || xhc->blob_size_64))
 790                 return -EINVAL;
 791
 792         mutex_lock(&kvm->lock);
 793
 794         if (xhc->msr && !kvm->arch.xen_hvm_config.msr)
 795                 static_branch_inc(&kvm_xen_enabled.key);
 796         else if (!xhc->msr && kvm->arch.xen_hvm_config.msr)
 797                 static_branch_slow_dec_deferred(&kvm_xen_enabled);
 798
 799         memcpy(&kvm->arch.xen_hvm_config, xhc, sizeof(*xhc));
 800
 801         mutex_unlock(&kvm->lock);
 802         return 0;
 803 }
 804
 805 void kvm_xen_init_vm(struct kvm *kvm)
 806 {
 807 }
 808
 809 void kvm_xen_destroy_vm(struct kvm *kvm)
 810 {
 811         kvm_gfn_to_pfn_cache_destroy(kvm, &kvm->arch.xen.shinfo_cache);
 812
 813         if (kvm->arch.xen_hvm_config.msr)
 814                 static_branch_slow_dec_deferred(&kvm_xen_enabled);
 815 }
 816
 817 static int kvm_xen_hypercall_set_result(struct kvm_vcpu *vcpu, u64 result)
 818 {
 819         kvm_rax_write(vcpu, result);
 820         return kvm_skip_emulated_instruction(vcpu);
 821 }
 822
 823 static int kvm_xen_hypercall_complete_userspace(struct kvm_vcpu *vcpu)
 824 {
 825         struct kvm_run *run = vcpu->run;
 826
 827         if (unlikely(!kvm_is_linear_rip(vcpu, vcpu->arch.xen.hypercall_rip)))
 828                 return 1;
 829
 830         return kvm_xen_hypercall_set_result(vcpu, run->xen.u.hcall.result);
 831 }
 832
 833 int kvm_xen_hypercall(struct kvm_vcpu *vcpu)
 834 {
 835         bool longmode;
 836         u64 input, params[6];
 837
 838         input = (u64)kvm_register_read(vcpu, VCPU_REGS_RAX);
 839
 840         /* Hyper-V hypercalls get bit 31 set in EAX */
 841         if ((input & 0x80000000) &&
 842             kvm_hv_hypercall_enabled(vcpu))
 843                 return kvm_hv_hypercall(vcpu);
 844
 845         longmode = is_64_bit_hypercall(vcpu);
 846         if (!longmode) {
 847                 params[0] = (u32)kvm_rbx_read(vcpu);
 848                 params[1] = (u32)kvm_rcx_read(vcpu);
 849                 params[2] = (u32)kvm_rdx_read(vcpu);
 850                 params[3] = (u32)kvm_rsi_read(vcpu);
 851                 params[4] = (u32)kvm_rdi_read(vcpu);
 852                 params[5] = (u32)kvm_rbp_read(vcpu);
 853         }
 854 #ifdef CONFIG_X86_64
 855         else {
 856                 params[0] = (u64)kvm_rdi_read(vcpu);
 857                 params[1] = (u64)kvm_rsi_read(vcpu);
 858                 params[2] = (u64)kvm_rdx_read(vcpu);
 859                 params[3] = (u64)kvm_r10_read(vcpu);
 860                 params[4] = (u64)kvm_r8_read(vcpu);
 861                 params[5] = (u64)kvm_r9_read(vcpu);
 862         }
 863 #endif
 864         trace_kvm_xen_hypercall(input, params[0], params[1], params[2],
 865                                 params[3], params[4], params[5]);
 866
 867         vcpu->run->exit_reason = KVM_EXIT_XEN;
 868         vcpu->run->xen.type = KVM_EXIT_XEN_HCALL;
 869         vcpu->run->xen.u.hcall.longmode = longmode;
 870         vcpu->run->xen.u.hcall.cpl = static_call(kvm_x86_get_cpl)(vcpu);
 871         vcpu->run->xen.u.hcall.input = input;
 872         vcpu->run->xen.u.hcall.params[0] = params[0];
 873         vcpu->run->xen.u.hcall.params[1] = params[1];
 874         vcpu->run->xen.u.hcall.params[2] = params[2];
 875         vcpu->run->xen.u.hcall.params[3] = params[3];
 876         vcpu->run->xen.u.hcall.params[4] = params[4];
 877         vcpu->run->xen.u.hcall.params[5] = params[5];
 878         vcpu->arch.xen.hypercall_rip = kvm_get_linear_rip(vcpu);
 879         vcpu->arch.complete_userspace_io =
 880                 kvm_xen_hypercall_complete_userspace;
 881
 882         return 0;
 883 }
 884
 885 static inline int max_evtchn_port(struct kvm *kvm)
 886 {
 887         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode)
 888                 return EVTCHN_2L_NR_CHANNELS;
 889         else
 890                 return COMPAT_EVTCHN_2L_NR_CHANNELS;
 891 }
 892
 893 /*
 894  * This follows the kvm_set_irq() API, so it returns:
 895  *  < 0   Interrupt was ignored (masked or not delivered for other reasons)
 896  *  = 0   Interrupt was coalesced (previous irq is still pending)
 897  *  > 0   Number of CPUs interrupt was delivered to
 898  */
 899 int kvm_xen_set_evtchn_fast(struct kvm_kernel_irq_routing_entry *e,
 900                             struct kvm *kvm)
 901 {
 902         struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
 903         struct kvm_vcpu *vcpu;
 904         unsigned long *pending_bits, *mask_bits;
 905         unsigned long flags;
 906         int port_word_bit;
 907         bool kick_vcpu = false;
 908         int idx;
 909         int rc;
 910
 911         vcpu = kvm_get_vcpu_by_id(kvm, e->xen_evtchn.vcpu);
 912         if (!vcpu)
 913                 return -1;
 914
 915         if (!vcpu->arch.xen.vcpu_info_set)
 916                 return -1;
 917
 918         if (e->xen_evtchn.port >= max_evtchn_port(kvm))
 919                 return -1;
 920
 921         rc = -EWOULDBLOCK;
 922         read_lock_irqsave(&gpc->lock, flags);
 923
 924         idx = srcu_read_lock(&kvm->srcu);
 925         if (!kvm_gfn_to_pfn_cache_check(kvm, gpc, gpc->gpa, PAGE_SIZE))
 926                 goto out_rcu;
 927
 928         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
 929                 struct shared_info *shinfo = gpc->khva;
 930                 pending_bits = (unsigned long *)&shinfo->evtchn_pending;
 931                 mask_bits = (unsigned long *)&shinfo->evtchn_mask;
 932                 port_word_bit = e->xen_evtchn.port / 64;
 933         } else {
 934                 struct compat_shared_info *shinfo = gpc->khva;
 935                 pending_bits = (unsigned long *)&shinfo->evtchn_pending;
 936                 mask_bits = (unsigned long *)&shinfo->evtchn_mask;
 937                 port_word_bit = e->xen_evtchn.port / 32;
 938         }
 939
 940         /*
 941          * If this port wasn't already set, and if it isn't masked, then
 942          * we try to set the corresponding bit in the in-kernel shadow of
 943          * evtchn_pending_sel for the target vCPU. And if *that* wasn't
 944          * already set, then we kick the vCPU in question to write to the
 945          * *real* evtchn_pending_sel in its own guest vcpu_info struct.
 946          */
 947         if (test_and_set_bit(e->xen_evtchn.port, pending_bits)) {
 948                 rc = 0; /* It was already raised */
 949         } else if (test_bit(e->xen_evtchn.port, mask_bits)) {
 950                 rc = -1; /* Masked */
 951         } else {
 952                 rc = 1; /* Delivered. But was the vCPU waking already? */
 953                 if (!test_and_set_bit(port_word_bit, &vcpu->arch.xen.evtchn_pending_sel))
 954                         kick_vcpu = true;
 955         }
 956
 957  out_rcu:
 958         srcu_read_unlock(&kvm->srcu, idx);
 959         read_unlock_irqrestore(&gpc->lock, flags);
 960
 961         if (kick_vcpu) {
 962                 kvm_make_request(KVM_REQ_EVENT, vcpu);
 963                 kvm_vcpu_kick(vcpu);
 964         }
 965
 966         return rc;
 967 }
 968
 969 /* This is the version called from kvm_set_irq() as the .set function */
 970 static int evtchn_set_fn(struct kvm_kernel_irq_routing_entry *e, struct kvm *kvm,
 971                          int irq_source_id, int level, bool line_status)
 972 {
 973         bool mm_borrowed = false;
 974         int rc;
 975
 976         if (!level)
 977                 return -1;
 978
 979         rc = kvm_xen_set_evtchn_fast(e, kvm);
 980         if (rc != -EWOULDBLOCK)
 981                 return rc;
 982
 983         if (current->mm != kvm->mm) {
 984                 /*
 985                  * If not on a thread which already belongs to this KVM,
 986                  * we'd better be in the irqfd workqueue.
 987                  */
 988                 if (WARN_ON_ONCE(current->mm))
 989                         return -EINVAL;
 990
 991                 kthread_use_mm(kvm->mm);
 992                 mm_borrowed = true;
 993         }
 994
 995         /*
 996          * For the irqfd workqueue, using the main kvm->lock mutex is
 997          * fine since this function is invoked from kvm_set_irq() with
 998          * no other lock held, no srcu. In future if it will be called
 999          * directly from a vCPU thread (e.g. on hypercall for an IPI)
1000          * then it may need to switch to using a leaf-node mutex for
1001          * serializing the shared_info mapping.
1002          */
1003         mutex_lock(&kvm->lock);
1004
1005         /*
1006          * It is theoretically possible for the page to be unmapped
1007          * and the MMU notifier to invalidate the shared_info before
1008          * we even get to use it. In that case, this looks like an
1009          * infinite loop. It was tempting to do it via the userspace
1010          * HVA instead... but that just *hides* the fact that it's
1011          * an infinite loop, because if a fault occurs and it waits
1012          * for the page to come back, it can *still* immediately
1013          * fault and have to wait again, repeatedly.
1014          *
1015          * Conversely, the page could also have been reinstated by
1016          * another thread before we even obtain the mutex above, so
1017          * check again *first* before remapping it.
1018          */
1019         do {
1020                 struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
1021                 int idx;
1022
1023                 rc = kvm_xen_set_evtchn_fast(e, kvm);
1024                 if (rc != -EWOULDBLOCK)
1025                         break;
1026
1027                 idx = srcu_read_lock(&kvm->srcu);
1028                 rc = kvm_gfn_to_pfn_cache_refresh(kvm, gpc, gpc->gpa, PAGE_SIZE);
1029                 srcu_read_unlock(&kvm->srcu, idx);
1030         } while(!rc);
1031
1032         mutex_unlock(&kvm->lock);
1033
1034         if (mm_borrowed)
1035                 kthread_unuse_mm(kvm->mm);
1036
1037         return rc;
1038 }
1039
1040 int kvm_xen_setup_evtchn(struct kvm *kvm,
1041                          struct kvm_kernel_irq_routing_entry *e,
1042                          const struct kvm_irq_routing_entry *ue)
1043
1044 {
1045         if (ue->u.xen_evtchn.port >= max_evtchn_port(kvm))
1046                 return -EINVAL;
1047
1048         /* We only support 2 level event channels for now */
1049         if (ue->u.xen_evtchn.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1050                 return -EINVAL;
1051
1052         e->xen_evtchn.port = ue->u.xen_evtchn.port;
1053         e->xen_evtchn.vcpu = ue->u.xen_evtchn.vcpu;
1054         e->xen_evtchn.priority = ue->u.xen_evtchn.priority;
1055         e->set = evtchn_set_fn;
1056
1057         return 0;
1058 }