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