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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include "x86.h"
  11 #include "xen.h"
  12 #include "hyperv.h"
  13 #include "lapic.h"
  14
  15 #include <linux/eventfd.h>
  16 #include <linux/kvm_host.h>
  17 #include <linux/sched/stat.h>
  18
  19 #include <trace/events/kvm.h>
  20 #include <xen/interface/xen.h>
  21 #include <xen/interface/vcpu.h>
  22 #include <xen/interface/version.h>
  23 #include <xen/interface/event_channel.h>
  24 #include <xen/interface/sched.h>
  25
  26 #include <asm/xen/cpuid.h>
  27
  28 #include "cpuid.h"
  29 #include "trace.h"
  30
  31 static int kvm_xen_set_evtchn(struct kvm_xen_evtchn *xe, struct kvm *kvm);
  32 static int kvm_xen_setattr_evtchn(struct kvm *kvm, struct kvm_xen_hvm_attr *data);
  33 static bool kvm_xen_hcall_evtchn_send(struct kvm_vcpu *vcpu, u64 param, u64 *r);
  34
  35 DEFINE_STATIC_KEY_DEFERRED_FALSE(kvm_xen_enabled, HZ);
  36
  37 static int kvm_xen_shared_info_init(struct kvm *kvm, gfn_t gfn)
  38 {
  39         struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
  40         struct pvclock_wall_clock *wc;
  41         gpa_t gpa = gfn_to_gpa(gfn);
  42         u32 *wc_sec_hi;
  43         u32 wc_version;
  44         u64 wall_nsec;
  45         int ret = 0;
  46         int idx = srcu_read_lock(&kvm->srcu);
  47
  48         if (gfn == KVM_XEN_INVALID_GFN) {
  49                 kvm_gpc_deactivate(gpc);
  50                 goto out;
  51         }
  52
  53         do {
  54                 ret = kvm_gpc_activate(gpc, gpa, PAGE_SIZE);
  55                 if (ret)
  56                         goto out;
  57
  58                 /*
  59                  * This code mirrors kvm_write_wall_clock() except that it writes
  60                  * directly through the pfn cache and doesn't mark the page dirty.
  61                  */
  62                 wall_nsec = kvm_get_wall_clock_epoch(kvm);
  63
  64                 /* It could be invalid again already, so we need to check */
  65                 read_lock_irq(&gpc->lock);
  66
  67                 if (gpc->valid)
  68                         break;
  69
  70                 read_unlock_irq(&gpc->lock);
  71         } while (1);
  72
  73         /* Paranoia checks on the 32-bit struct layout */
  74         BUILD_BUG_ON(offsetof(struct compat_shared_info, wc) != 0x900);
  75         BUILD_BUG_ON(offsetof(struct compat_shared_info, arch.wc_sec_hi) != 0x924);
  76         BUILD_BUG_ON(offsetof(struct pvclock_vcpu_time_info, version) != 0);
  77
  78 #ifdef CONFIG_X86_64
  79         /* Paranoia checks on the 64-bit struct layout */
  80         BUILD_BUG_ON(offsetof(struct shared_info, wc) != 0xc00);
  81         BUILD_BUG_ON(offsetof(struct shared_info, wc_sec_hi) != 0xc0c);
  82
  83         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
  84                 struct shared_info *shinfo = gpc->khva;
  85
  86                 wc_sec_hi = &shinfo->wc_sec_hi;
  87                 wc = &shinfo->wc;
  88         } else
  89 #endif
  90         {
  91                 struct compat_shared_info *shinfo = gpc->khva;
  92
  93                 wc_sec_hi = &shinfo->arch.wc_sec_hi;
  94                 wc = &shinfo->wc;
  95         }
  96
  97         /* Increment and ensure an odd value */
  98         wc_version = wc->version = (wc->version + 1) | 1;
  99         smp_wmb();
 100
 101         wc->nsec = do_div(wall_nsec, NSEC_PER_SEC);
 102         wc->sec = (u32)wall_nsec;
 103         *wc_sec_hi = wall_nsec >> 32;
 104         smp_wmb();
 105
 106         wc->version = wc_version + 1;
 107         read_unlock_irq(&gpc->lock);
 108
 109         kvm_make_all_cpus_request(kvm, KVM_REQ_MASTERCLOCK_UPDATE);
 110
 111 out:
 112         srcu_read_unlock(&kvm->srcu, idx);
 113         return ret;
 114 }
 115
 116 void kvm_xen_inject_timer_irqs(struct kvm_vcpu *vcpu)
 117 {
 118         if (atomic_read(&vcpu->arch.xen.timer_pending) > 0) {
 119                 struct kvm_xen_evtchn e;
 120
 121                 e.vcpu_id = vcpu->vcpu_id;
 122                 e.vcpu_idx = vcpu->vcpu_idx;
 123                 e.port = vcpu->arch.xen.timer_virq;
 124                 e.priority = KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL;
 125
 126                 kvm_xen_set_evtchn(&e, vcpu->kvm);
 127
 128                 vcpu->arch.xen.timer_expires = 0;
 129                 atomic_set(&vcpu->arch.xen.timer_pending, 0);
 130         }
 131 }
 132
 133 static enum hrtimer_restart xen_timer_callback(struct hrtimer *timer)
 134 {
 135         struct kvm_vcpu *vcpu = container_of(timer, struct kvm_vcpu,
 136                                              arch.xen.timer);
 137         struct kvm_xen_evtchn e;
 138         int rc;
 139
 140         if (atomic_read(&vcpu->arch.xen.timer_pending))
 141                 return HRTIMER_NORESTART;
 142
 143         e.vcpu_id = vcpu->vcpu_id;
 144         e.vcpu_idx = vcpu->vcpu_idx;
 145         e.port = vcpu->arch.xen.timer_virq;
 146         e.priority = KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL;
 147
 148         rc = kvm_xen_set_evtchn_fast(&e, vcpu->kvm);
 149         if (rc != -EWOULDBLOCK) {
 150                 vcpu->arch.xen.timer_expires = 0;
 151                 return HRTIMER_NORESTART;
 152         }
 153
 154         atomic_inc(&vcpu->arch.xen.timer_pending);
 155         kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
 156         kvm_vcpu_kick(vcpu);
 157
 158         return HRTIMER_NORESTART;
 159 }
 160
 161 static void kvm_xen_start_timer(struct kvm_vcpu *vcpu, u64 guest_abs, s64 delta_ns)
 162 {
 163         /*
 164          * Avoid races with the old timer firing. Checking timer_expires
 165          * to avoid calling hrtimer_cancel() will only have false positives
 166          * so is fine.
 167          */
 168         if (vcpu->arch.xen.timer_expires)
 169                 hrtimer_cancel(&vcpu->arch.xen.timer);
 170
 171         atomic_set(&vcpu->arch.xen.timer_pending, 0);
 172         vcpu->arch.xen.timer_expires = guest_abs;
 173
 174         if (delta_ns <= 0) {
 175                 xen_timer_callback(&vcpu->arch.xen.timer);
 176         } else {
 177                 ktime_t ktime_now = ktime_get();
 178                 hrtimer_start(&vcpu->arch.xen.timer,
 179                               ktime_add_ns(ktime_now, delta_ns),
 180                               HRTIMER_MODE_ABS_HARD);
 181         }
 182 }
 183
 184 static void kvm_xen_stop_timer(struct kvm_vcpu *vcpu)
 185 {
 186         hrtimer_cancel(&vcpu->arch.xen.timer);
 187         vcpu->arch.xen.timer_expires = 0;
 188         atomic_set(&vcpu->arch.xen.timer_pending, 0);
 189 }
 190
 191 static void kvm_xen_init_timer(struct kvm_vcpu *vcpu)
 192 {
 193         hrtimer_init(&vcpu->arch.xen.timer, CLOCK_MONOTONIC,
 194                      HRTIMER_MODE_ABS_HARD);
 195         vcpu->arch.xen.timer.function = xen_timer_callback;
 196 }
 197
 198 static void kvm_xen_update_runstate_guest(struct kvm_vcpu *v, bool atomic)
 199 {
 200         struct kvm_vcpu_xen *vx = &v->arch.xen;
 201         struct gfn_to_pfn_cache *gpc1 = &vx->runstate_cache;
 202         struct gfn_to_pfn_cache *gpc2 = &vx->runstate2_cache;
 203         size_t user_len, user_len1, user_len2;
 204         struct vcpu_runstate_info rs;
 205         unsigned long flags;
 206         size_t times_ofs;
 207         uint8_t *update_bit = NULL;
 208         uint64_t entry_time;
 209         uint64_t *rs_times;
 210         int *rs_state;
 211
 212         /*
 213          * The only difference between 32-bit and 64-bit versions of the
 214          * runstate struct is the alignment of uint64_t in 32-bit, which
 215          * means that the 64-bit version has an additional 4 bytes of
 216          * padding after the first field 'state'. Let's be really really
 217          * paranoid about that, and matching it with our internal data
 218          * structures that we memcpy into it...
 219          */
 220         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state) != 0);
 221         BUILD_BUG_ON(offsetof(struct compat_vcpu_runstate_info, state) != 0);
 222         BUILD_BUG_ON(sizeof(struct compat_vcpu_runstate_info) != 0x2c);
 223 #ifdef CONFIG_X86_64
 224         /*
 225          * The 64-bit structure has 4 bytes of padding before 'state_entry_time'
 226          * so each subsequent field is shifted by 4, and it's 4 bytes longer.
 227          */
 228         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state_entry_time) !=
 229                      offsetof(struct compat_vcpu_runstate_info, state_entry_time) + 4);
 230         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, time) !=
 231                      offsetof(struct compat_vcpu_runstate_info, time) + 4);
 232         BUILD_BUG_ON(sizeof(struct vcpu_runstate_info) != 0x2c + 4);
 233 #endif
 234         /*
 235          * The state field is in the same place at the start of both structs,
 236          * and is the same size (int) as vx->current_runstate.
 237          */
 238         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state) !=
 239                      offsetof(struct compat_vcpu_runstate_info, state));
 240         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, state) !=
 241                      sizeof(vx->current_runstate));
 242         BUILD_BUG_ON(sizeof_field(struct compat_vcpu_runstate_info, state) !=
 243                      sizeof(vx->current_runstate));
 244
 245         /*
 246          * The state_entry_time field is 64 bits in both versions, and the
 247          * XEN_RUNSTATE_UPDATE flag is in the top bit, which given that x86
 248          * is little-endian means that it's in the last *byte* of the word.
 249          * That detail is important later.
 250          */
 251         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, state_entry_time) !=
 252                      sizeof(uint64_t));
 253         BUILD_BUG_ON(sizeof_field(struct compat_vcpu_runstate_info, state_entry_time) !=
 254                      sizeof(uint64_t));
 255         BUILD_BUG_ON((XEN_RUNSTATE_UPDATE >> 56) != 0x80);
 256
 257         /*
 258          * The time array is four 64-bit quantities in both versions, matching
 259          * the vx->runstate_times and immediately following state_entry_time.
 260          */
 261         BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state_entry_time) !=
 262                      offsetof(struct vcpu_runstate_info, time) - sizeof(uint64_t));
 263         BUILD_BUG_ON(offsetof(struct compat_vcpu_runstate_info, state_entry_time) !=
 264                      offsetof(struct compat_vcpu_runstate_info, time) - sizeof(uint64_t));
 265         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, time) !=
 266                      sizeof_field(struct compat_vcpu_runstate_info, time));
 267         BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, time) !=
 268                      sizeof(vx->runstate_times));
 269
 270         if (IS_ENABLED(CONFIG_64BIT) && v->kvm->arch.xen.long_mode) {
 271                 user_len = sizeof(struct vcpu_runstate_info);
 272                 times_ofs = offsetof(struct vcpu_runstate_info,
 273                                      state_entry_time);
 274         } else {
 275                 user_len = sizeof(struct compat_vcpu_runstate_info);
 276                 times_ofs = offsetof(struct compat_vcpu_runstate_info,
 277                                      state_entry_time);
 278         }
 279
 280         /*
 281          * There are basically no alignment constraints. The guest can set it
 282          * up so it crosses from one page to the next, and at arbitrary byte
 283          * alignment (and the 32-bit ABI doesn't align the 64-bit integers
 284          * anyway, even if the overall struct had been 64-bit aligned).
 285          */
 286         if ((gpc1->gpa & ~PAGE_MASK) + user_len >= PAGE_SIZE) {
 287                 user_len1 = PAGE_SIZE - (gpc1->gpa & ~PAGE_MASK);
 288                 user_len2 = user_len - user_len1;
 289         } else {
 290                 user_len1 = user_len;
 291                 user_len2 = 0;
 292         }
 293         BUG_ON(user_len1 + user_len2 != user_len);
 294
 295  retry:
 296         /*
 297          * Attempt to obtain the GPC lock on *both* (if there are two)
 298          * gfn_to_pfn caches that cover the region.
 299          */
 300         if (atomic) {
 301                 local_irq_save(flags);
 302                 if (!read_trylock(&gpc1->lock)) {
 303                         local_irq_restore(flags);
 304                         return;
 305                 }
 306         } else {
 307                 read_lock_irqsave(&gpc1->lock, flags);
 308         }
 309         while (!kvm_gpc_check(gpc1, user_len1)) {
 310                 read_unlock_irqrestore(&gpc1->lock, flags);
 311
 312                 /* When invoked from kvm_sched_out() we cannot sleep */
 313                 if (atomic)
 314                         return;
 315
 316                 if (kvm_gpc_refresh(gpc1, user_len1))
 317                         return;
 318
 319                 read_lock_irqsave(&gpc1->lock, flags);
 320         }
 321
 322         if (likely(!user_len2)) {
 323                 /*
 324                  * Set up three pointers directly to the runstate_info
 325                  * struct in the guest (via the GPC).
 326                  *
 327                  *  • @rs_state   → state field
 328                  *  • @rs_times   → state_entry_time field.
 329                  *  • @update_bit → last byte of state_entry_time, which
 330                  *                  contains the XEN_RUNSTATE_UPDATE bit.
 331                  */
 332                 rs_state = gpc1->khva;
 333                 rs_times = gpc1->khva + times_ofs;
 334                 if (v->kvm->arch.xen.runstate_update_flag)
 335                         update_bit = ((void *)(&rs_times[1])) - 1;
 336         } else {
 337                 /*
 338                  * The guest's runstate_info is split across two pages and we
 339                  * need to hold and validate both GPCs simultaneously. We can
 340                  * declare a lock ordering GPC1 > GPC2 because nothing else
 341                  * takes them more than one at a time. Set a subclass on the
 342                  * gpc1 lock to make lockdep shut up about it.
 343                  */
 344                 lock_set_subclass(&gpc1->lock.dep_map, 1, _THIS_IP_);
 345                 if (atomic) {
 346                         if (!read_trylock(&gpc2->lock)) {
 347                                 read_unlock_irqrestore(&gpc1->lock, flags);
 348                                 return;
 349                         }
 350                 } else {
 351                         read_lock(&gpc2->lock);
 352                 }
 353
 354                 if (!kvm_gpc_check(gpc2, user_len2)) {
 355                         read_unlock(&gpc2->lock);
 356                         read_unlock_irqrestore(&gpc1->lock, flags);
 357
 358                         /* When invoked from kvm_sched_out() we cannot sleep */
 359                         if (atomic)
 360                                 return;
 361
 362                         /*
 363                          * Use kvm_gpc_activate() here because if the runstate
 364                          * area was configured in 32-bit mode and only extends
 365                          * to the second page now because the guest changed to
 366                          * 64-bit mode, the second GPC won't have been set up.
 367                          */
 368                         if (kvm_gpc_activate(gpc2, gpc1->gpa + user_len1,
 369                                              user_len2))
 370                                 return;
 371
 372                         /*
 373                          * We dropped the lock on GPC1 so we have to go all the
 374                          * way back and revalidate that too.
 375                          */
 376                         goto retry;
 377                 }
 378
 379                 /*
 380                  * In this case, the runstate_info struct will be assembled on
 381                  * the kernel stack (compat or not as appropriate) and will
 382                  * be copied to GPC1/GPC2 with a dual memcpy. Set up the three
 383                  * rs pointers accordingly.
 384                  */
 385                 rs_times = &rs.state_entry_time;
 386
 387                 /*
 388                  * The rs_state pointer points to the start of what we'll
 389                  * copy to the guest, which in the case of a compat guest
 390                  * is the 32-bit field that the compiler thinks is padding.
 391                  */
 392                 rs_state = ((void *)rs_times) - times_ofs;
 393
 394                 /*
 395                  * The update_bit is still directly in the guest memory,
 396                  * via one GPC or the other.
 397                  */
 398                 if (v->kvm->arch.xen.runstate_update_flag) {
 399                         if (user_len1 >= times_ofs + sizeof(uint64_t))
 400                                 update_bit = gpc1->khva + times_ofs +
 401                                         sizeof(uint64_t) - 1;
 402                         else
 403                                 update_bit = gpc2->khva + times_ofs +
 404                                         sizeof(uint64_t) - 1 - user_len1;
 405                 }
 406
 407 #ifdef CONFIG_X86_64
 408                 /*
 409                  * Don't leak kernel memory through the padding in the 64-bit
 410                  * version of the struct.
 411                  */
 412                 memset(&rs, 0, offsetof(struct vcpu_runstate_info, state_entry_time));
 413 #endif
 414         }
 415
 416         /*
 417          * First, set the XEN_RUNSTATE_UPDATE bit in the top bit of the
 418          * state_entry_time field, directly in the guest. We need to set
 419          * that (and write-barrier) before writing to the rest of the
 420          * structure, and clear it last. Just as Xen does, we address the
 421          * single *byte* in which it resides because it might be in a
 422          * different cache line to the rest of the 64-bit word, due to
 423          * the (lack of) alignment constraints.
 424          */
 425         entry_time = vx->runstate_entry_time;
 426         if (update_bit) {
 427                 entry_time |= XEN_RUNSTATE_UPDATE;
 428                 *update_bit = (vx->runstate_entry_time | XEN_RUNSTATE_UPDATE) >> 56;
 429                 smp_wmb();
 430         }
 431
 432         /*
 433          * Now assemble the actual structure, either on our kernel stack
 434          * or directly in the guest according to how the rs_state and
 435          * rs_times pointers were set up above.
 436          */
 437         *rs_state = vx->current_runstate;
 438         rs_times[0] = entry_time;
 439         memcpy(rs_times + 1, vx->runstate_times, sizeof(vx->runstate_times));
 440
 441         /* For the split case, we have to then copy it to the guest. */
 442         if (user_len2) {
 443                 memcpy(gpc1->khva, rs_state, user_len1);
 444                 memcpy(gpc2->khva, ((void *)rs_state) + user_len1, user_len2);
 445         }
 446         smp_wmb();
 447
 448         /* Finally, clear the XEN_RUNSTATE_UPDATE bit. */
 449         if (update_bit) {
 450                 entry_time &= ~XEN_RUNSTATE_UPDATE;
 451                 *update_bit = entry_time >> 56;
 452                 smp_wmb();
 453         }
 454
 455         if (user_len2)
 456                 read_unlock(&gpc2->lock);
 457
 458         read_unlock_irqrestore(&gpc1->lock, flags);
 459
 460         mark_page_dirty_in_slot(v->kvm, gpc1->memslot, gpc1->gpa >> PAGE_SHIFT);
 461         if (user_len2)
 462                 mark_page_dirty_in_slot(v->kvm, gpc2->memslot, gpc2->gpa >> PAGE_SHIFT);
 463 }
 464
 465 void kvm_xen_update_runstate(struct kvm_vcpu *v, int state)
 466 {
 467         struct kvm_vcpu_xen *vx = &v->arch.xen;
 468         u64 now = get_kvmclock_ns(v->kvm);
 469         u64 delta_ns = now - vx->runstate_entry_time;
 470         u64 run_delay = current->sched_info.run_delay;
 471
 472         if (unlikely(!vx->runstate_entry_time))
 473                 vx->current_runstate = RUNSTATE_offline;
 474
 475         /*
 476          * Time waiting for the scheduler isn't "stolen" if the
 477          * vCPU wasn't running anyway.
 478          */
 479         if (vx->current_runstate == RUNSTATE_running) {
 480                 u64 steal_ns = run_delay - vx->last_steal;
 481
 482                 delta_ns -= steal_ns;
 483
 484                 vx->runstate_times[RUNSTATE_runnable] += steal_ns;
 485         }
 486         vx->last_steal = run_delay;
 487
 488         vx->runstate_times[vx->current_runstate] += delta_ns;
 489         vx->current_runstate = state;
 490         vx->runstate_entry_time = now;
 491
 492         if (vx->runstate_cache.active)
 493                 kvm_xen_update_runstate_guest(v, state == RUNSTATE_runnable);
 494 }
 495
 496 void kvm_xen_inject_vcpu_vector(struct kvm_vcpu *v)
 497 {
 498         struct kvm_lapic_irq irq = { };
 499         int r;
 500
 501         irq.dest_id = v->vcpu_id;
 502         irq.vector = v->arch.xen.upcall_vector;
 503         irq.dest_mode = APIC_DEST_PHYSICAL;
 504         irq.shorthand = APIC_DEST_NOSHORT;
 505         irq.delivery_mode = APIC_DM_FIXED;
 506         irq.level = 1;
 507
 508         /* The fast version will always work for physical unicast */
 509         WARN_ON_ONCE(!kvm_irq_delivery_to_apic_fast(v->kvm, NULL, &irq, &r, NULL));
 510 }
 511
 512 /*
 513  * On event channel delivery, the vcpu_info may not have been accessible.
 514  * In that case, there are bits in vcpu->arch.xen.evtchn_pending_sel which
 515  * need to be marked into the vcpu_info (and evtchn_upcall_pending set).
 516  * Do so now that we can sleep in the context of the vCPU to bring the
 517  * page in, and refresh the pfn cache for it.
 518  */
 519 void kvm_xen_inject_pending_events(struct kvm_vcpu *v)
 520 {
 521         unsigned long evtchn_pending_sel = READ_ONCE(v->arch.xen.evtchn_pending_sel);
 522         struct gfn_to_pfn_cache *gpc = &v->arch.xen.vcpu_info_cache;
 523         unsigned long flags;
 524
 525         if (!evtchn_pending_sel)
 526                 return;
 527
 528         /*
 529          * Yes, this is an open-coded loop. But that's just what put_user()
 530          * does anyway. Page it in and retry the instruction. We're just a
 531          * little more honest about it.
 532          */
 533         read_lock_irqsave(&gpc->lock, flags);
 534         while (!kvm_gpc_check(gpc, sizeof(struct vcpu_info))) {
 535                 read_unlock_irqrestore(&gpc->lock, flags);
 536
 537                 if (kvm_gpc_refresh(gpc, sizeof(struct vcpu_info)))
 538                         return;
 539
 540                 read_lock_irqsave(&gpc->lock, flags);
 541         }
 542
 543         /* Now gpc->khva is a valid kernel address for the vcpu_info */
 544         if (IS_ENABLED(CONFIG_64BIT) && v->kvm->arch.xen.long_mode) {
 545                 struct vcpu_info *vi = gpc->khva;
 546
 547                 asm volatile(LOCK_PREFIX "orq %0, %1\n"
 548                              "notq %0\n"
 549                              LOCK_PREFIX "andq %0, %2\n"
 550                              : "=r" (evtchn_pending_sel),
 551                                "+m" (vi->evtchn_pending_sel),
 552                                "+m" (v->arch.xen.evtchn_pending_sel)
 553                              : "0" (evtchn_pending_sel));
 554                 WRITE_ONCE(vi->evtchn_upcall_pending, 1);
 555         } else {
 556                 u32 evtchn_pending_sel32 = evtchn_pending_sel;
 557                 struct compat_vcpu_info *vi = gpc->khva;
 558
 559                 asm volatile(LOCK_PREFIX "orl %0, %1\n"
 560                              "notl %0\n"
 561                              LOCK_PREFIX "andl %0, %2\n"
 562                              : "=r" (evtchn_pending_sel32),
 563                                "+m" (vi->evtchn_pending_sel),
 564                                "+m" (v->arch.xen.evtchn_pending_sel)
 565                              : "0" (evtchn_pending_sel32));
 566                 WRITE_ONCE(vi->evtchn_upcall_pending, 1);
 567         }
 568         read_unlock_irqrestore(&gpc->lock, flags);
 569
 570         /* For the per-vCPU lapic vector, deliver it as MSI. */
 571         if (v->arch.xen.upcall_vector)
 572                 kvm_xen_inject_vcpu_vector(v);
 573
 574         mark_page_dirty_in_slot(v->kvm, gpc->memslot, gpc->gpa >> PAGE_SHIFT);
 575 }
 576
 577 int __kvm_xen_has_interrupt(struct kvm_vcpu *v)
 578 {
 579         struct gfn_to_pfn_cache *gpc = &v->arch.xen.vcpu_info_cache;
 580         unsigned long flags;
 581         u8 rc = 0;
 582
 583         /*
 584          * If the global upcall vector (HVMIRQ_callback_vector) is set and
 585          * the vCPU's evtchn_upcall_pending flag is set, the IRQ is pending.
 586          */
 587
 588         /* No need for compat handling here */
 589         BUILD_BUG_ON(offsetof(struct vcpu_info, evtchn_upcall_pending) !=
 590                      offsetof(struct compat_vcpu_info, evtchn_upcall_pending));
 591         BUILD_BUG_ON(sizeof(rc) !=
 592                      sizeof_field(struct vcpu_info, evtchn_upcall_pending));
 593         BUILD_BUG_ON(sizeof(rc) !=
 594                      sizeof_field(struct compat_vcpu_info, evtchn_upcall_pending));
 595
 596         read_lock_irqsave(&gpc->lock, flags);
 597         while (!kvm_gpc_check(gpc, sizeof(struct vcpu_info))) {
 598                 read_unlock_irqrestore(&gpc->lock, flags);
 599
 600                 /*
 601                  * This function gets called from kvm_vcpu_block() after setting the
 602                  * task to TASK_INTERRUPTIBLE, to see if it needs to wake immediately
 603                  * from a HLT. So we really mustn't sleep. If the page ended up absent
 604                  * at that point, just return 1 in order to trigger an immediate wake,
 605                  * and we'll end up getting called again from a context where we *can*
 606                  * fault in the page and wait for it.
 607                  */
 608                 if (in_atomic() || !task_is_running(current))
 609                         return 1;
 610
 611                 if (kvm_gpc_refresh(gpc, sizeof(struct vcpu_info))) {
 612                         /*
 613                          * If this failed, userspace has screwed up the
 614                          * vcpu_info mapping. No interrupts for you.
 615                          */
 616                         return 0;
 617                 }
 618                 read_lock_irqsave(&gpc->lock, flags);
 619         }
 620
 621         rc = ((struct vcpu_info *)gpc->khva)->evtchn_upcall_pending;
 622         read_unlock_irqrestore(&gpc->lock, flags);
 623         return rc;
 624 }
 625
 626 int kvm_xen_hvm_set_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 627 {
 628         int r = -ENOENT;
 629
 630
 631         switch (data->type) {
 632         case KVM_XEN_ATTR_TYPE_LONG_MODE:
 633                 if (!IS_ENABLED(CONFIG_64BIT) && data->u.long_mode) {
 634                         r = -EINVAL;
 635                 } else {
 636                         mutex_lock(&kvm->arch.xen.xen_lock);
 637                         kvm->arch.xen.long_mode = !!data->u.long_mode;
 638                         mutex_unlock(&kvm->arch.xen.xen_lock);
 639                         r = 0;
 640                 }
 641                 break;
 642
 643         case KVM_XEN_ATTR_TYPE_SHARED_INFO:
 644                 mutex_lock(&kvm->arch.xen.xen_lock);
 645                 r = kvm_xen_shared_info_init(kvm, data->u.shared_info.gfn);
 646                 mutex_unlock(&kvm->arch.xen.xen_lock);
 647                 break;
 648
 649         case KVM_XEN_ATTR_TYPE_UPCALL_VECTOR:
 650                 if (data->u.vector && data->u.vector < 0x10)
 651                         r = -EINVAL;
 652                 else {
 653                         mutex_lock(&kvm->arch.xen.xen_lock);
 654                         kvm->arch.xen.upcall_vector = data->u.vector;
 655                         mutex_unlock(&kvm->arch.xen.xen_lock);
 656                         r = 0;
 657                 }
 658                 break;
 659
 660         case KVM_XEN_ATTR_TYPE_EVTCHN:
 661                 r = kvm_xen_setattr_evtchn(kvm, data);
 662                 break;
 663
 664         case KVM_XEN_ATTR_TYPE_XEN_VERSION:
 665                 mutex_lock(&kvm->arch.xen.xen_lock);
 666                 kvm->arch.xen.xen_version = data->u.xen_version;
 667                 mutex_unlock(&kvm->arch.xen.xen_lock);
 668                 r = 0;
 669                 break;
 670
 671         case KVM_XEN_ATTR_TYPE_RUNSTATE_UPDATE_FLAG:
 672                 if (!sched_info_on()) {
 673                         r = -EOPNOTSUPP;
 674                         break;
 675                 }
 676                 mutex_lock(&kvm->arch.xen.xen_lock);
 677                 kvm->arch.xen.runstate_update_flag = !!data->u.runstate_update_flag;
 678                 mutex_unlock(&kvm->arch.xen.xen_lock);
 679                 r = 0;
 680                 break;
 681
 682         default:
 683                 break;
 684         }
 685
 686         return r;
 687 }
 688
 689 int kvm_xen_hvm_get_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 690 {
 691         int r = -ENOENT;
 692
 693         mutex_lock(&kvm->arch.xen.xen_lock);
 694
 695         switch (data->type) {
 696         case KVM_XEN_ATTR_TYPE_LONG_MODE:
 697                 data->u.long_mode = kvm->arch.xen.long_mode;
 698                 r = 0;
 699                 break;
 700
 701         case KVM_XEN_ATTR_TYPE_SHARED_INFO:
 702                 if (kvm->arch.xen.shinfo_cache.active)
 703                         data->u.shared_info.gfn = gpa_to_gfn(kvm->arch.xen.shinfo_cache.gpa);
 704                 else
 705                         data->u.shared_info.gfn = KVM_XEN_INVALID_GFN;
 706                 r = 0;
 707                 break;
 708
 709         case KVM_XEN_ATTR_TYPE_UPCALL_VECTOR:
 710                 data->u.vector = kvm->arch.xen.upcall_vector;
 711                 r = 0;
 712                 break;
 713
 714         case KVM_XEN_ATTR_TYPE_XEN_VERSION:
 715                 data->u.xen_version = kvm->arch.xen.xen_version;
 716                 r = 0;
 717                 break;
 718
 719         case KVM_XEN_ATTR_TYPE_RUNSTATE_UPDATE_FLAG:
 720                 if (!sched_info_on()) {
 721                         r = -EOPNOTSUPP;
 722                         break;
 723                 }
 724                 data->u.runstate_update_flag = kvm->arch.xen.runstate_update_flag;
 725                 r = 0;
 726                 break;
 727
 728         default:
 729                 break;
 730         }
 731
 732         mutex_unlock(&kvm->arch.xen.xen_lock);
 733         return r;
 734 }
 735
 736 int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 737 {
 738         int idx, r = -ENOENT;
 739
 740         mutex_lock(&vcpu->kvm->arch.xen.xen_lock);
 741         idx = srcu_read_lock(&vcpu->kvm->srcu);
 742
 743         switch (data->type) {
 744         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
 745                 /* No compat necessary here. */
 746                 BUILD_BUG_ON(sizeof(struct vcpu_info) !=
 747                              sizeof(struct compat_vcpu_info));
 748                 BUILD_BUG_ON(offsetof(struct vcpu_info, time) !=
 749                              offsetof(struct compat_vcpu_info, time));
 750
 751                 if (data->u.gpa == KVM_XEN_INVALID_GPA) {
 752                         kvm_gpc_deactivate(&vcpu->arch.xen.vcpu_info_cache);
 753                         r = 0;
 754                         break;
 755                 }
 756
 757                 r = kvm_gpc_activate(&vcpu->arch.xen.vcpu_info_cache,
 758                                      data->u.gpa, sizeof(struct vcpu_info));
 759                 if (!r)
 760                         kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
 761
 762                 break;
 763
 764         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_TIME_INFO:
 765                 if (data->u.gpa == KVM_XEN_INVALID_GPA) {
 766                         kvm_gpc_deactivate(&vcpu->arch.xen.vcpu_time_info_cache);
 767                         r = 0;
 768                         break;
 769                 }
 770
 771                 r = kvm_gpc_activate(&vcpu->arch.xen.vcpu_time_info_cache,
 772                                      data->u.gpa,
 773                                      sizeof(struct pvclock_vcpu_time_info));
 774                 if (!r)
 775                         kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
 776                 break;
 777
 778         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADDR: {
 779                 size_t sz, sz1, sz2;
 780
 781                 if (!sched_info_on()) {
 782                         r = -EOPNOTSUPP;
 783                         break;
 784                 }
 785                 if (data->u.gpa == KVM_XEN_INVALID_GPA) {
 786                         r = 0;
 787                 deactivate_out:
 788                         kvm_gpc_deactivate(&vcpu->arch.xen.runstate_cache);
 789                         kvm_gpc_deactivate(&vcpu->arch.xen.runstate2_cache);
 790                         break;
 791                 }
 792
 793                 /*
 794                  * If the guest switches to 64-bit mode after setting the runstate
 795                  * address, that's actually OK. kvm_xen_update_runstate_guest()
 796                  * will cope.
 797                  */
 798                 if (IS_ENABLED(CONFIG_64BIT) && vcpu->kvm->arch.xen.long_mode)
 799                         sz = sizeof(struct vcpu_runstate_info);
 800                 else
 801                         sz = sizeof(struct compat_vcpu_runstate_info);
 802
 803                 /* How much fits in the (first) page? */
 804                 sz1 = PAGE_SIZE - (data->u.gpa & ~PAGE_MASK);
 805                 r = kvm_gpc_activate(&vcpu->arch.xen.runstate_cache,
 806                                      data->u.gpa, sz1);
 807                 if (r)
 808                         goto deactivate_out;
 809
 810                 /* Either map the second page, or deactivate the second GPC */
 811                 if (sz1 >= sz) {
 812                         kvm_gpc_deactivate(&vcpu->arch.xen.runstate2_cache);
 813                 } else {
 814                         sz2 = sz - sz1;
 815                         BUG_ON((data->u.gpa + sz1) & ~PAGE_MASK);
 816                         r = kvm_gpc_activate(&vcpu->arch.xen.runstate2_cache,
 817                                              data->u.gpa + sz1, sz2);
 818                         if (r)
 819                                 goto deactivate_out;
 820                 }
 821
 822                 kvm_xen_update_runstate_guest(vcpu, false);
 823                 break;
 824         }
 825         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_CURRENT:
 826                 if (!sched_info_on()) {
 827                         r = -EOPNOTSUPP;
 828                         break;
 829                 }
 830                 if (data->u.runstate.state > RUNSTATE_offline) {
 831                         r = -EINVAL;
 832                         break;
 833                 }
 834
 835                 kvm_xen_update_runstate(vcpu, data->u.runstate.state);
 836                 r = 0;
 837                 break;
 838
 839         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_DATA:
 840                 if (!sched_info_on()) {
 841                         r = -EOPNOTSUPP;
 842                         break;
 843                 }
 844                 if (data->u.runstate.state > RUNSTATE_offline) {
 845                         r = -EINVAL;
 846                         break;
 847                 }
 848                 if (data->u.runstate.state_entry_time !=
 849                     (data->u.runstate.time_running +
 850                      data->u.runstate.time_runnable +
 851                      data->u.runstate.time_blocked +
 852                      data->u.runstate.time_offline)) {
 853                         r = -EINVAL;
 854                         break;
 855                 }
 856                 if (get_kvmclock_ns(vcpu->kvm) <
 857                     data->u.runstate.state_entry_time) {
 858                         r = -EINVAL;
 859                         break;
 860                 }
 861
 862                 vcpu->arch.xen.current_runstate = data->u.runstate.state;
 863                 vcpu->arch.xen.runstate_entry_time =
 864                         data->u.runstate.state_entry_time;
 865                 vcpu->arch.xen.runstate_times[RUNSTATE_running] =
 866                         data->u.runstate.time_running;
 867                 vcpu->arch.xen.runstate_times[RUNSTATE_runnable] =
 868                         data->u.runstate.time_runnable;
 869                 vcpu->arch.xen.runstate_times[RUNSTATE_blocked] =
 870                         data->u.runstate.time_blocked;
 871                 vcpu->arch.xen.runstate_times[RUNSTATE_offline] =
 872                         data->u.runstate.time_offline;
 873                 vcpu->arch.xen.last_steal = current->sched_info.run_delay;
 874                 r = 0;
 875                 break;
 876
 877         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADJUST:
 878                 if (!sched_info_on()) {
 879                         r = -EOPNOTSUPP;
 880                         break;
 881                 }
 882                 if (data->u.runstate.state > RUNSTATE_offline &&
 883                     data->u.runstate.state != (u64)-1) {
 884                         r = -EINVAL;
 885                         break;
 886                 }
 887                 /* The adjustment must add up */
 888                 if (data->u.runstate.state_entry_time !=
 889                     (data->u.runstate.time_running +
 890                      data->u.runstate.time_runnable +
 891                      data->u.runstate.time_blocked +
 892                      data->u.runstate.time_offline)) {
 893                         r = -EINVAL;
 894                         break;
 895                 }
 896
 897                 if (get_kvmclock_ns(vcpu->kvm) <
 898                     (vcpu->arch.xen.runstate_entry_time +
 899                      data->u.runstate.state_entry_time)) {
 900                         r = -EINVAL;
 901                         break;
 902                 }
 903
 904                 vcpu->arch.xen.runstate_entry_time +=
 905                         data->u.runstate.state_entry_time;
 906                 vcpu->arch.xen.runstate_times[RUNSTATE_running] +=
 907                         data->u.runstate.time_running;
 908                 vcpu->arch.xen.runstate_times[RUNSTATE_runnable] +=
 909                         data->u.runstate.time_runnable;
 910                 vcpu->arch.xen.runstate_times[RUNSTATE_blocked] +=
 911                         data->u.runstate.time_blocked;
 912                 vcpu->arch.xen.runstate_times[RUNSTATE_offline] +=
 913                         data->u.runstate.time_offline;
 914
 915                 if (data->u.runstate.state <= RUNSTATE_offline)
 916                         kvm_xen_update_runstate(vcpu, data->u.runstate.state);
 917                 else if (vcpu->arch.xen.runstate_cache.active)
 918                         kvm_xen_update_runstate_guest(vcpu, false);
 919                 r = 0;
 920                 break;
 921
 922         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_ID:
 923                 if (data->u.vcpu_id >= KVM_MAX_VCPUS)
 924                         r = -EINVAL;
 925                 else {
 926                         vcpu->arch.xen.vcpu_id = data->u.vcpu_id;
 927                         r = 0;
 928                 }
 929                 break;
 930
 931         case KVM_XEN_VCPU_ATTR_TYPE_TIMER:
 932                 if (data->u.timer.port &&
 933                     data->u.timer.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL) {
 934                         r = -EINVAL;
 935                         break;
 936                 }
 937
 938                 if (!vcpu->arch.xen.timer.function)
 939                         kvm_xen_init_timer(vcpu);
 940
 941                 /* Stop the timer (if it's running) before changing the vector */
 942                 kvm_xen_stop_timer(vcpu);
 943                 vcpu->arch.xen.timer_virq = data->u.timer.port;
 944
 945                 /* Start the timer if the new value has a valid vector+expiry. */
 946                 if (data->u.timer.port && data->u.timer.expires_ns)
 947                         kvm_xen_start_timer(vcpu, data->u.timer.expires_ns,
 948                                             data->u.timer.expires_ns -
 949                                             get_kvmclock_ns(vcpu->kvm));
 950
 951                 r = 0;
 952                 break;
 953
 954         case KVM_XEN_VCPU_ATTR_TYPE_UPCALL_VECTOR:
 955                 if (data->u.vector && data->u.vector < 0x10)
 956                         r = -EINVAL;
 957                 else {
 958                         vcpu->arch.xen.upcall_vector = data->u.vector;
 959                         r = 0;
 960                 }
 961                 break;
 962
 963         default:
 964                 break;
 965         }
 966
 967         srcu_read_unlock(&vcpu->kvm->srcu, idx);
 968         mutex_unlock(&vcpu->kvm->arch.xen.xen_lock);
 969         return r;
 970 }
 971
 972 int kvm_xen_vcpu_get_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 973 {
 974         int r = -ENOENT;
 975
 976         mutex_lock(&vcpu->kvm->arch.xen.xen_lock);
 977
 978         switch (data->type) {
 979         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
 980                 if (vcpu->arch.xen.vcpu_info_cache.active)
 981                         data->u.gpa = vcpu->arch.xen.vcpu_info_cache.gpa;
 982                 else
 983                         data->u.gpa = KVM_XEN_INVALID_GPA;
 984                 r = 0;
 985                 break;
 986
 987         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_TIME_INFO:
 988                 if (vcpu->arch.xen.vcpu_time_info_cache.active)
 989                         data->u.gpa = vcpu->arch.xen.vcpu_time_info_cache.gpa;
 990                 else
 991                         data->u.gpa = KVM_XEN_INVALID_GPA;
 992                 r = 0;
 993                 break;
 994
 995         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADDR:
 996                 if (!sched_info_on()) {
 997                         r = -EOPNOTSUPP;
 998                         break;
 999                 }
1000                 if (vcpu->arch.xen.runstate_cache.active) {
1001                         data->u.gpa = vcpu->arch.xen.runstate_cache.gpa;
1002                         r = 0;
1003                 }
1004                 break;
1005
1006         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_CURRENT:
1007                 if (!sched_info_on()) {
1008                         r = -EOPNOTSUPP;
1009                         break;
1010                 }
1011                 data->u.runstate.state = vcpu->arch.xen.current_runstate;
1012                 r = 0;
1013                 break;
1014
1015         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_DATA:
1016                 if (!sched_info_on()) {
1017                         r = -EOPNOTSUPP;
1018                         break;
1019                 }
1020                 data->u.runstate.state = vcpu->arch.xen.current_runstate;
1021                 data->u.runstate.state_entry_time =
1022                         vcpu->arch.xen.runstate_entry_time;
1023                 data->u.runstate.time_running =
1024                         vcpu->arch.xen.runstate_times[RUNSTATE_running];
1025                 data->u.runstate.time_runnable =
1026                         vcpu->arch.xen.runstate_times[RUNSTATE_runnable];
1027                 data->u.runstate.time_blocked =
1028                         vcpu->arch.xen.runstate_times[RUNSTATE_blocked];
1029                 data->u.runstate.time_offline =
1030                         vcpu->arch.xen.runstate_times[RUNSTATE_offline];
1031                 r = 0;
1032                 break;
1033
1034         case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADJUST:
1035                 r = -EINVAL;
1036                 break;
1037
1038         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_ID:
1039                 data->u.vcpu_id = vcpu->arch.xen.vcpu_id;
1040                 r = 0;
1041                 break;
1042
1043         case KVM_XEN_VCPU_ATTR_TYPE_TIMER:
1044                 /*
1045                  * Ensure a consistent snapshot of state is captured, with a
1046                  * timer either being pending, or the event channel delivered
1047                  * to the corresponding bit in the shared_info. Not still
1048                  * lurking in the timer_pending flag for deferred delivery.
1049                  * Purely as an optimisation, if the timer_expires field is
1050                  * zero, that means the timer isn't active (or even in the
1051                  * timer_pending flag) and there is no need to cancel it.
1052                  */
1053                 if (vcpu->arch.xen.timer_expires) {
1054                         hrtimer_cancel(&vcpu->arch.xen.timer);
1055                         kvm_xen_inject_timer_irqs(vcpu);
1056                 }
1057
1058                 data->u.timer.port = vcpu->arch.xen.timer_virq;
1059                 data->u.timer.priority = KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL;
1060                 data->u.timer.expires_ns = vcpu->arch.xen.timer_expires;
1061
1062                 /*
1063                  * The hrtimer may trigger and raise the IRQ immediately,
1064                  * while the returned state causes it to be set up and
1065                  * raised again on the destination system after migration.
1066                  * That's fine, as the guest won't even have had a chance
1067                  * to run and handle the interrupt. Asserting an already
1068                  * pending event channel is idempotent.
1069                  */
1070                 if (vcpu->arch.xen.timer_expires)
1071                         hrtimer_start_expires(&vcpu->arch.xen.timer,
1072                                               HRTIMER_MODE_ABS_HARD);
1073
1074                 r = 0;
1075                 break;
1076
1077         case KVM_XEN_VCPU_ATTR_TYPE_UPCALL_VECTOR:
1078                 data->u.vector = vcpu->arch.xen.upcall_vector;
1079                 r = 0;
1080                 break;
1081
1082         default:
1083                 break;
1084         }
1085
1086         mutex_unlock(&vcpu->kvm->arch.xen.xen_lock);
1087         return r;
1088 }
1089
1090 int kvm_xen_write_hypercall_page(struct kvm_vcpu *vcpu, u64 data)
1091 {
1092         struct kvm *kvm = vcpu->kvm;
1093         u32 page_num = data & ~PAGE_MASK;
1094         u64 page_addr = data & PAGE_MASK;
1095         bool lm = is_long_mode(vcpu);
1096
1097         /* Latch long_mode for shared_info pages etc. */
1098         vcpu->kvm->arch.xen.long_mode = lm;
1099
1100         /*
1101          * If Xen hypercall intercept is enabled, fill the hypercall
1102          * page with VMCALL/VMMCALL instructions since that's what
1103          * we catch. Else the VMM has provided the hypercall pages
1104          * with instructions of its own choosing, so use those.
1105          */
1106         if (kvm_xen_hypercall_enabled(kvm)) {
1107                 u8 instructions[32];
1108                 int i;
1109
1110                 if (page_num)
1111                         return 1;
1112
1113                 /* mov imm32, %eax */
1114                 instructions[0] = 0xb8;
1115
1116                 /* vmcall / vmmcall */
1117                 static_call(kvm_x86_patch_hypercall)(vcpu, instructions + 5);
1118
1119                 /* ret */
1120                 instructions[8] = 0xc3;
1121
1122                 /* int3 to pad */
1123                 memset(instructions + 9, 0xcc, sizeof(instructions) - 9);
1124
1125                 for (i = 0; i < PAGE_SIZE / sizeof(instructions); i++) {
1126                         *(u32 *)&instructions[1] = i;
1127                         if (kvm_vcpu_write_guest(vcpu,
1128                                                  page_addr + (i * sizeof(instructions)),
1129                                                  instructions, sizeof(instructions)))
1130                                 return 1;
1131                 }
1132         } else {
1133                 /*
1134                  * Note, truncation is a non-issue as 'lm' is guaranteed to be
1135                  * false for a 32-bit kernel, i.e. when hva_t is only 4 bytes.
1136                  */
1137                 hva_t blob_addr = lm ? kvm->arch.xen_hvm_config.blob_addr_64
1138                                      : kvm->arch.xen_hvm_config.blob_addr_32;
1139                 u8 blob_size = lm ? kvm->arch.xen_hvm_config.blob_size_64
1140                                   : kvm->arch.xen_hvm_config.blob_size_32;
1141                 u8 *page;
1142                 int ret;
1143
1144                 if (page_num >= blob_size)
1145                         return 1;
1146
1147                 blob_addr += page_num * PAGE_SIZE;
1148
1149                 page = memdup_user((u8 __user *)blob_addr, PAGE_SIZE);
1150                 if (IS_ERR(page))
1151                         return PTR_ERR(page);
1152
1153                 ret = kvm_vcpu_write_guest(vcpu, page_addr, page, PAGE_SIZE);
1154                 kfree(page);
1155                 if (ret)
1156                         return 1;
1157         }
1158         return 0;
1159 }
1160
1161 int kvm_xen_hvm_config(struct kvm *kvm, struct kvm_xen_hvm_config *xhc)
1162 {
1163         /* Only some feature flags need to be *enabled* by userspace */
1164         u32 permitted_flags = KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL |
1165                 KVM_XEN_HVM_CONFIG_EVTCHN_SEND |
1166                 KVM_XEN_HVM_CONFIG_PVCLOCK_TSC_UNSTABLE;
1167         u32 old_flags;
1168
1169         if (xhc->flags & ~permitted_flags)
1170                 return -EINVAL;
1171
1172         /*
1173          * With hypercall interception the kernel generates its own
1174          * hypercall page so it must not be provided.
1175          */
1176         if ((xhc->flags & KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL) &&
1177             (xhc->blob_addr_32 || xhc->blob_addr_64 ||
1178              xhc->blob_size_32 || xhc->blob_size_64))
1179                 return -EINVAL;
1180
1181         mutex_lock(&kvm->arch.xen.xen_lock);
1182
1183         if (xhc->msr && !kvm->arch.xen_hvm_config.msr)
1184                 static_branch_inc(&kvm_xen_enabled.key);
1185         else if (!xhc->msr && kvm->arch.xen_hvm_config.msr)
1186                 static_branch_slow_dec_deferred(&kvm_xen_enabled);
1187
1188         old_flags = kvm->arch.xen_hvm_config.flags;
1189         memcpy(&kvm->arch.xen_hvm_config, xhc, sizeof(*xhc));
1190
1191         mutex_unlock(&kvm->arch.xen.xen_lock);
1192
1193         if ((old_flags ^ xhc->flags) & KVM_XEN_HVM_CONFIG_PVCLOCK_TSC_UNSTABLE)
1194                 kvm_make_all_cpus_request(kvm, KVM_REQ_CLOCK_UPDATE);
1195
1196         return 0;
1197 }
1198
1199 static int kvm_xen_hypercall_set_result(struct kvm_vcpu *vcpu, u64 result)
1200 {
1201         kvm_rax_write(vcpu, result);
1202         return kvm_skip_emulated_instruction(vcpu);
1203 }
1204
1205 static int kvm_xen_hypercall_complete_userspace(struct kvm_vcpu *vcpu)
1206 {
1207         struct kvm_run *run = vcpu->run;
1208
1209         if (unlikely(!kvm_is_linear_rip(vcpu, vcpu->arch.xen.hypercall_rip)))
1210                 return 1;
1211
1212         return kvm_xen_hypercall_set_result(vcpu, run->xen.u.hcall.result);
1213 }
1214
1215 static inline int max_evtchn_port(struct kvm *kvm)
1216 {
1217         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode)
1218                 return EVTCHN_2L_NR_CHANNELS;
1219         else
1220                 return COMPAT_EVTCHN_2L_NR_CHANNELS;
1221 }
1222
1223 static bool wait_pending_event(struct kvm_vcpu *vcpu, int nr_ports,
1224                                evtchn_port_t *ports)
1225 {
1226         struct kvm *kvm = vcpu->kvm;
1227         struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
1228         unsigned long *pending_bits;
1229         unsigned long flags;
1230         bool ret = true;
1231         int idx, i;
1232
1233         idx = srcu_read_lock(&kvm->srcu);
1234         read_lock_irqsave(&gpc->lock, flags);
1235         if (!kvm_gpc_check(gpc, PAGE_SIZE))
1236                 goto out_rcu;
1237
1238         ret = false;
1239         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
1240                 struct shared_info *shinfo = gpc->khva;
1241                 pending_bits = (unsigned long *)&shinfo->evtchn_pending;
1242         } else {
1243                 struct compat_shared_info *shinfo = gpc->khva;
1244                 pending_bits = (unsigned long *)&shinfo->evtchn_pending;
1245         }
1246
1247         for (i = 0; i < nr_ports; i++) {
1248                 if (test_bit(ports[i], pending_bits)) {
1249                         ret = true;
1250                         break;
1251                 }
1252         }
1253
1254  out_rcu:
1255         read_unlock_irqrestore(&gpc->lock, flags);
1256         srcu_read_unlock(&kvm->srcu, idx);
1257
1258         return ret;
1259 }
1260
1261 static bool kvm_xen_schedop_poll(struct kvm_vcpu *vcpu, bool longmode,
1262                                  u64 param, u64 *r)
1263 {
1264         struct sched_poll sched_poll;
1265         evtchn_port_t port, *ports;
1266         struct x86_exception e;
1267         int i;
1268
1269         if (!lapic_in_kernel(vcpu) ||
1270             !(vcpu->kvm->arch.xen_hvm_config.flags & KVM_XEN_HVM_CONFIG_EVTCHN_SEND))
1271                 return false;
1272
1273         if (IS_ENABLED(CONFIG_64BIT) && !longmode) {
1274                 struct compat_sched_poll sp32;
1275
1276                 /* Sanity check that the compat struct definition is correct */
1277                 BUILD_BUG_ON(sizeof(sp32) != 16);
1278
1279                 if (kvm_read_guest_virt(vcpu, param, &sp32, sizeof(sp32), &e)) {
1280                         *r = -EFAULT;
1281                         return true;
1282                 }
1283
1284                 /*
1285                  * This is a 32-bit pointer to an array of evtchn_port_t which
1286                  * are uint32_t, so once it's converted no further compat
1287                  * handling is needed.
1288                  */
1289                 sched_poll.ports = (void *)(unsigned long)(sp32.ports);
1290                 sched_poll.nr_ports = sp32.nr_ports;
1291                 sched_poll.timeout = sp32.timeout;
1292         } else {
1293                 if (kvm_read_guest_virt(vcpu, param, &sched_poll,
1294                                         sizeof(sched_poll), &e)) {
1295                         *r = -EFAULT;
1296                         return true;
1297                 }
1298         }
1299
1300         if (unlikely(sched_poll.nr_ports > 1)) {
1301                 /* Xen (unofficially) limits number of pollers to 128 */
1302                 if (sched_poll.nr_ports > 128) {
1303                         *r = -EINVAL;
1304                         return true;
1305                 }
1306
1307                 ports = kmalloc_array(sched_poll.nr_ports,
1308                                       sizeof(*ports), GFP_KERNEL);
1309                 if (!ports) {
1310                         *r = -ENOMEM;
1311                         return true;
1312                 }
1313         } else
1314                 ports = &port;
1315
1316         if (kvm_read_guest_virt(vcpu, (gva_t)sched_poll.ports, ports,
1317                                 sched_poll.nr_ports * sizeof(*ports), &e)) {
1318                 *r = -EFAULT;
1319                 return true;
1320         }
1321
1322         for (i = 0; i < sched_poll.nr_ports; i++) {
1323                 if (ports[i] >= max_evtchn_port(vcpu->kvm)) {
1324                         *r = -EINVAL;
1325                         goto out;
1326                 }
1327         }
1328
1329         if (sched_poll.nr_ports == 1)
1330                 vcpu->arch.xen.poll_evtchn = port;
1331         else
1332                 vcpu->arch.xen.poll_evtchn = -1;
1333
1334         set_bit(vcpu->vcpu_idx, vcpu->kvm->arch.xen.poll_mask);
1335
1336         if (!wait_pending_event(vcpu, sched_poll.nr_ports, ports)) {
1337                 vcpu->arch.mp_state = KVM_MP_STATE_HALTED;
1338
1339                 if (sched_poll.timeout)
1340                         mod_timer(&vcpu->arch.xen.poll_timer,
1341                                   jiffies + nsecs_to_jiffies(sched_poll.timeout));
1342
1343                 kvm_vcpu_halt(vcpu);
1344
1345                 if (sched_poll.timeout)
1346                         del_timer(&vcpu->arch.xen.poll_timer);
1347
1348                 vcpu->arch.mp_state = KVM_MP_STATE_RUNNABLE;
1349         }
1350
1351         vcpu->arch.xen.poll_evtchn = 0;
1352         *r = 0;
1353 out:
1354         /* Really, this is only needed in case of timeout */
1355         clear_bit(vcpu->vcpu_idx, vcpu->kvm->arch.xen.poll_mask);
1356
1357         if (unlikely(sched_poll.nr_ports > 1))
1358                 kfree(ports);
1359         return true;
1360 }
1361
1362 static void cancel_evtchn_poll(struct timer_list *t)
1363 {
1364         struct kvm_vcpu *vcpu = from_timer(vcpu, t, arch.xen.poll_timer);
1365
1366         kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
1367         kvm_vcpu_kick(vcpu);
1368 }
1369
1370 static bool kvm_xen_hcall_sched_op(struct kvm_vcpu *vcpu, bool longmode,
1371                                    int cmd, u64 param, u64 *r)
1372 {
1373         switch (cmd) {
1374         case SCHEDOP_poll:
1375                 if (kvm_xen_schedop_poll(vcpu, longmode, param, r))
1376                         return true;
1377                 fallthrough;
1378         case SCHEDOP_yield:
1379                 kvm_vcpu_on_spin(vcpu, true);
1380                 *r = 0;
1381                 return true;
1382         default:
1383                 break;
1384         }
1385
1386         return false;
1387 }
1388
1389 struct compat_vcpu_set_singleshot_timer {
1390     uint64_t timeout_abs_ns;
1391     uint32_t flags;
1392 } __attribute__((packed));
1393
1394 static bool kvm_xen_hcall_vcpu_op(struct kvm_vcpu *vcpu, bool longmode, int cmd,
1395                                   int vcpu_id, u64 param, u64 *r)
1396 {
1397         struct vcpu_set_singleshot_timer oneshot;
1398         struct x86_exception e;
1399         s64 delta;
1400
1401         if (!kvm_xen_timer_enabled(vcpu))
1402                 return false;
1403
1404         switch (cmd) {
1405         case VCPUOP_set_singleshot_timer:
1406                 if (vcpu->arch.xen.vcpu_id != vcpu_id) {
1407                         *r = -EINVAL;
1408                         return true;
1409                 }
1410
1411                 /*
1412                  * The only difference for 32-bit compat is the 4 bytes of
1413                  * padding after the interesting part of the structure. So
1414                  * for a faithful emulation of Xen we have to *try* to copy
1415                  * the padding and return -EFAULT if we can't. Otherwise we
1416                  * might as well just have copied the 12-byte 32-bit struct.
1417                  */
1418                 BUILD_BUG_ON(offsetof(struct compat_vcpu_set_singleshot_timer, timeout_abs_ns) !=
1419                              offsetof(struct vcpu_set_singleshot_timer, timeout_abs_ns));
1420                 BUILD_BUG_ON(sizeof_field(struct compat_vcpu_set_singleshot_timer, timeout_abs_ns) !=
1421                              sizeof_field(struct vcpu_set_singleshot_timer, timeout_abs_ns));
1422                 BUILD_BUG_ON(offsetof(struct compat_vcpu_set_singleshot_timer, flags) !=
1423                              offsetof(struct vcpu_set_singleshot_timer, flags));
1424                 BUILD_BUG_ON(sizeof_field(struct compat_vcpu_set_singleshot_timer, flags) !=
1425                              sizeof_field(struct vcpu_set_singleshot_timer, flags));
1426
1427                 if (kvm_read_guest_virt(vcpu, param, &oneshot, longmode ? sizeof(oneshot) :
1428                                         sizeof(struct compat_vcpu_set_singleshot_timer), &e)) {
1429                         *r = -EFAULT;
1430                         return true;
1431                 }
1432
1433                 /* A delta <= 0 results in an immediate callback, which is what we want */
1434                 delta = oneshot.timeout_abs_ns - get_kvmclock_ns(vcpu->kvm);
1435                 kvm_xen_start_timer(vcpu, oneshot.timeout_abs_ns, delta);
1436                 *r = 0;
1437                 return true;
1438
1439         case VCPUOP_stop_singleshot_timer:
1440                 if (vcpu->arch.xen.vcpu_id != vcpu_id) {
1441                         *r = -EINVAL;
1442                         return true;
1443                 }
1444                 kvm_xen_stop_timer(vcpu);
1445                 *r = 0;
1446                 return true;
1447         }
1448
1449         return false;
1450 }
1451
1452 static bool kvm_xen_hcall_set_timer_op(struct kvm_vcpu *vcpu, uint64_t timeout,
1453                                        u64 *r)
1454 {
1455         if (!kvm_xen_timer_enabled(vcpu))
1456                 return false;
1457
1458         if (timeout) {
1459                 uint64_t guest_now = get_kvmclock_ns(vcpu->kvm);
1460                 int64_t delta = timeout - guest_now;
1461
1462                 /* Xen has a 'Linux workaround' in do_set_timer_op() which
1463                  * checks for negative absolute timeout values (caused by
1464                  * integer overflow), and for values about 13 days in the
1465                  * future (2^50ns) which would be caused by jiffies
1466                  * overflow. For those cases, it sets the timeout 100ms in
1467                  * the future (not *too* soon, since if a guest really did
1468                  * set a long timeout on purpose we don't want to keep
1469                  * churning CPU time by waking it up).
1470                  */
1471                 if (unlikely((int64_t)timeout < 0 ||
1472                              (delta > 0 && (uint32_t) (delta >> 50) != 0))) {
1473                         delta = 100 * NSEC_PER_MSEC;
1474                         timeout = guest_now + delta;
1475                 }
1476
1477                 kvm_xen_start_timer(vcpu, timeout, delta);
1478         } else {
1479                 kvm_xen_stop_timer(vcpu);
1480         }
1481
1482         *r = 0;
1483         return true;
1484 }
1485
1486 int kvm_xen_hypercall(struct kvm_vcpu *vcpu)
1487 {
1488         bool longmode;
1489         u64 input, params[6], r = -ENOSYS;
1490         bool handled = false;
1491         u8 cpl;
1492
1493         input = (u64)kvm_register_read(vcpu, VCPU_REGS_RAX);
1494
1495         /* Hyper-V hypercalls get bit 31 set in EAX */
1496         if ((input & 0x80000000) &&
1497             kvm_hv_hypercall_enabled(vcpu))
1498                 return kvm_hv_hypercall(vcpu);
1499
1500         longmode = is_64_bit_hypercall(vcpu);
1501         if (!longmode) {
1502                 params[0] = (u32)kvm_rbx_read(vcpu);
1503                 params[1] = (u32)kvm_rcx_read(vcpu);
1504                 params[2] = (u32)kvm_rdx_read(vcpu);
1505                 params[3] = (u32)kvm_rsi_read(vcpu);
1506                 params[4] = (u32)kvm_rdi_read(vcpu);
1507                 params[5] = (u32)kvm_rbp_read(vcpu);
1508         }
1509 #ifdef CONFIG_X86_64
1510         else {
1511                 params[0] = (u64)kvm_rdi_read(vcpu);
1512                 params[1] = (u64)kvm_rsi_read(vcpu);
1513                 params[2] = (u64)kvm_rdx_read(vcpu);
1514                 params[3] = (u64)kvm_r10_read(vcpu);
1515                 params[4] = (u64)kvm_r8_read(vcpu);
1516                 params[5] = (u64)kvm_r9_read(vcpu);
1517         }
1518 #endif
1519         cpl = static_call(kvm_x86_get_cpl)(vcpu);
1520         trace_kvm_xen_hypercall(cpl, input, params[0], params[1], params[2],
1521                                 params[3], params[4], params[5]);
1522
1523         /*
1524          * Only allow hypercall acceleration for CPL0. The rare hypercalls that
1525          * are permitted in guest userspace can be handled by the VMM.
1526          */
1527         if (unlikely(cpl > 0))
1528                 goto handle_in_userspace;
1529
1530         switch (input) {
1531         case __HYPERVISOR_xen_version:
1532                 if (params[0] == XENVER_version && vcpu->kvm->arch.xen.xen_version) {
1533                         r = vcpu->kvm->arch.xen.xen_version;
1534                         handled = true;
1535                 }
1536                 break;
1537         case __HYPERVISOR_event_channel_op:
1538                 if (params[0] == EVTCHNOP_send)
1539                         handled = kvm_xen_hcall_evtchn_send(vcpu, params[1], &r);
1540                 break;
1541         case __HYPERVISOR_sched_op:
1542                 handled = kvm_xen_hcall_sched_op(vcpu, longmode, params[0],
1543                                                  params[1], &r);
1544                 break;
1545         case __HYPERVISOR_vcpu_op:
1546                 handled = kvm_xen_hcall_vcpu_op(vcpu, longmode, params[0], params[1],
1547                                                 params[2], &r);
1548                 break;
1549         case __HYPERVISOR_set_timer_op: {
1550                 u64 timeout = params[0];
1551                 /* In 32-bit mode, the 64-bit timeout is in two 32-bit params. */
1552                 if (!longmode)
1553                         timeout |= params[1] << 32;
1554                 handled = kvm_xen_hcall_set_timer_op(vcpu, timeout, &r);
1555                 break;
1556         }
1557         default:
1558                 break;
1559         }
1560
1561         if (handled)
1562                 return kvm_xen_hypercall_set_result(vcpu, r);
1563
1564 handle_in_userspace:
1565         vcpu->run->exit_reason = KVM_EXIT_XEN;
1566         vcpu->run->xen.type = KVM_EXIT_XEN_HCALL;
1567         vcpu->run->xen.u.hcall.longmode = longmode;
1568         vcpu->run->xen.u.hcall.cpl = cpl;
1569         vcpu->run->xen.u.hcall.input = input;
1570         vcpu->run->xen.u.hcall.params[0] = params[0];
1571         vcpu->run->xen.u.hcall.params[1] = params[1];
1572         vcpu->run->xen.u.hcall.params[2] = params[2];
1573         vcpu->run->xen.u.hcall.params[3] = params[3];
1574         vcpu->run->xen.u.hcall.params[4] = params[4];
1575         vcpu->run->xen.u.hcall.params[5] = params[5];
1576         vcpu->arch.xen.hypercall_rip = kvm_get_linear_rip(vcpu);
1577         vcpu->arch.complete_userspace_io =
1578                 kvm_xen_hypercall_complete_userspace;
1579
1580         return 0;
1581 }
1582
1583 static void kvm_xen_check_poller(struct kvm_vcpu *vcpu, int port)
1584 {
1585         int poll_evtchn = vcpu->arch.xen.poll_evtchn;
1586
1587         if ((poll_evtchn == port || poll_evtchn == -1) &&
1588             test_and_clear_bit(vcpu->vcpu_idx, vcpu->kvm->arch.xen.poll_mask)) {
1589                 kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
1590                 kvm_vcpu_kick(vcpu);
1591         }
1592 }
1593
1594 /*
1595  * The return value from this function is propagated to kvm_set_irq() API,
1596  * so it returns:
1597  *  < 0   Interrupt was ignored (masked or not delivered for other reasons)
1598  *  = 0   Interrupt was coalesced (previous irq is still pending)
1599  *  > 0   Number of CPUs interrupt was delivered to
1600  *
1601  * It is also called directly from kvm_arch_set_irq_inatomic(), where the
1602  * only check on its return value is a comparison with -EWOULDBLOCK'.
1603  */
1604 int kvm_xen_set_evtchn_fast(struct kvm_xen_evtchn *xe, struct kvm *kvm)
1605 {
1606         struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
1607         struct kvm_vcpu *vcpu;
1608         unsigned long *pending_bits, *mask_bits;
1609         unsigned long flags;
1610         int port_word_bit;
1611         bool kick_vcpu = false;
1612         int vcpu_idx, idx, rc;
1613
1614         vcpu_idx = READ_ONCE(xe->vcpu_idx);
1615         if (vcpu_idx >= 0)
1616                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
1617         else {
1618                 vcpu = kvm_get_vcpu_by_id(kvm, xe->vcpu_id);
1619                 if (!vcpu)
1620                         return -EINVAL;
1621                 WRITE_ONCE(xe->vcpu_idx, vcpu->vcpu_idx);
1622         }
1623
1624         if (!vcpu->arch.xen.vcpu_info_cache.active)
1625                 return -EINVAL;
1626
1627         if (xe->port >= max_evtchn_port(kvm))
1628                 return -EINVAL;
1629
1630         rc = -EWOULDBLOCK;
1631
1632         idx = srcu_read_lock(&kvm->srcu);
1633
1634         read_lock_irqsave(&gpc->lock, flags);
1635         if (!kvm_gpc_check(gpc, PAGE_SIZE))
1636                 goto out_rcu;
1637
1638         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
1639                 struct shared_info *shinfo = gpc->khva;
1640                 pending_bits = (unsigned long *)&shinfo->evtchn_pending;
1641                 mask_bits = (unsigned long *)&shinfo->evtchn_mask;
1642                 port_word_bit = xe->port / 64;
1643         } else {
1644                 struct compat_shared_info *shinfo = gpc->khva;
1645                 pending_bits = (unsigned long *)&shinfo->evtchn_pending;
1646                 mask_bits = (unsigned long *)&shinfo->evtchn_mask;
1647                 port_word_bit = xe->port / 32;
1648         }
1649
1650         /*
1651          * If this port wasn't already set, and if it isn't masked, then
1652          * we try to set the corresponding bit in the in-kernel shadow of
1653          * evtchn_pending_sel for the target vCPU. And if *that* wasn't
1654          * already set, then we kick the vCPU in question to write to the
1655          * *real* evtchn_pending_sel in its own guest vcpu_info struct.
1656          */
1657         if (test_and_set_bit(xe->port, pending_bits)) {
1658                 rc = 0; /* It was already raised */
1659         } else if (test_bit(xe->port, mask_bits)) {
1660                 rc = -ENOTCONN; /* Masked */
1661                 kvm_xen_check_poller(vcpu, xe->port);
1662         } else {
1663                 rc = 1; /* Delivered to the bitmap in shared_info. */
1664                 /* Now switch to the vCPU's vcpu_info to set the index and pending_sel */
1665                 read_unlock_irqrestore(&gpc->lock, flags);
1666                 gpc = &vcpu->arch.xen.vcpu_info_cache;
1667
1668                 read_lock_irqsave(&gpc->lock, flags);
1669                 if (!kvm_gpc_check(gpc, sizeof(struct vcpu_info))) {
1670                         /*
1671                          * Could not access the vcpu_info. Set the bit in-kernel
1672                          * and prod the vCPU to deliver it for itself.
1673                          */
1674                         if (!test_and_set_bit(port_word_bit, &vcpu->arch.xen.evtchn_pending_sel))
1675                                 kick_vcpu = true;
1676                         goto out_rcu;
1677                 }
1678
1679                 if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
1680                         struct vcpu_info *vcpu_info = gpc->khva;
1681                         if (!test_and_set_bit(port_word_bit, &vcpu_info->evtchn_pending_sel)) {
1682                                 WRITE_ONCE(vcpu_info->evtchn_upcall_pending, 1);
1683                                 kick_vcpu = true;
1684                         }
1685                 } else {
1686                         struct compat_vcpu_info *vcpu_info = gpc->khva;
1687                         if (!test_and_set_bit(port_word_bit,
1688                                               (unsigned long *)&vcpu_info->evtchn_pending_sel)) {
1689                                 WRITE_ONCE(vcpu_info->evtchn_upcall_pending, 1);
1690                                 kick_vcpu = true;
1691                         }
1692                 }
1693
1694                 /* For the per-vCPU lapic vector, deliver it as MSI. */
1695                 if (kick_vcpu && vcpu->arch.xen.upcall_vector) {
1696                         kvm_xen_inject_vcpu_vector(vcpu);
1697                         kick_vcpu = false;
1698                 }
1699         }
1700
1701  out_rcu:
1702         read_unlock_irqrestore(&gpc->lock, flags);
1703         srcu_read_unlock(&kvm->srcu, idx);
1704
1705         if (kick_vcpu) {
1706                 kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
1707                 kvm_vcpu_kick(vcpu);
1708         }
1709
1710         return rc;
1711 }
1712
1713 static int kvm_xen_set_evtchn(struct kvm_xen_evtchn *xe, struct kvm *kvm)
1714 {
1715         bool mm_borrowed = false;
1716         int rc;
1717
1718         rc = kvm_xen_set_evtchn_fast(xe, kvm);
1719         if (rc != -EWOULDBLOCK)
1720                 return rc;
1721
1722         if (current->mm != kvm->mm) {
1723                 /*
1724                  * If not on a thread which already belongs to this KVM,
1725                  * we'd better be in the irqfd workqueue.
1726                  */
1727                 if (WARN_ON_ONCE(current->mm))
1728                         return -EINVAL;
1729
1730                 kthread_use_mm(kvm->mm);
1731                 mm_borrowed = true;
1732         }
1733
1734         mutex_lock(&kvm->arch.xen.xen_lock);
1735
1736         /*
1737          * It is theoretically possible for the page to be unmapped
1738          * and the MMU notifier to invalidate the shared_info before
1739          * we even get to use it. In that case, this looks like an
1740          * infinite loop. It was tempting to do it via the userspace
1741          * HVA instead... but that just *hides* the fact that it's
1742          * an infinite loop, because if a fault occurs and it waits
1743          * for the page to come back, it can *still* immediately
1744          * fault and have to wait again, repeatedly.
1745          *
1746          * Conversely, the page could also have been reinstated by
1747          * another thread before we even obtain the mutex above, so
1748          * check again *first* before remapping it.
1749          */
1750         do {
1751                 struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
1752                 int idx;
1753
1754                 rc = kvm_xen_set_evtchn_fast(xe, kvm);
1755                 if (rc != -EWOULDBLOCK)
1756                         break;
1757
1758                 idx = srcu_read_lock(&kvm->srcu);
1759                 rc = kvm_gpc_refresh(gpc, PAGE_SIZE);
1760                 srcu_read_unlock(&kvm->srcu, idx);
1761         } while(!rc);
1762
1763         mutex_unlock(&kvm->arch.xen.xen_lock);
1764
1765         if (mm_borrowed)
1766                 kthread_unuse_mm(kvm->mm);
1767
1768         return rc;
1769 }
1770
1771 /* This is the version called from kvm_set_irq() as the .set function */
1772 static int evtchn_set_fn(struct kvm_kernel_irq_routing_entry *e, struct kvm *kvm,
1773                          int irq_source_id, int level, bool line_status)
1774 {
1775         if (!level)
1776                 return -EINVAL;
1777
1778         return kvm_xen_set_evtchn(&e->xen_evtchn, kvm);
1779 }
1780
1781 /*
1782  * Set up an event channel interrupt from the KVM IRQ routing table.
1783  * Used for e.g. PIRQ from passed through physical devices.
1784  */
1785 int kvm_xen_setup_evtchn(struct kvm *kvm,
1786                          struct kvm_kernel_irq_routing_entry *e,
1787                          const struct kvm_irq_routing_entry *ue)
1788
1789 {
1790         struct kvm_vcpu *vcpu;
1791
1792         if (ue->u.xen_evtchn.port >= max_evtchn_port(kvm))
1793                 return -EINVAL;
1794
1795         /* We only support 2 level event channels for now */
1796         if (ue->u.xen_evtchn.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1797                 return -EINVAL;
1798
1799         /*
1800          * Xen gives us interesting mappings from vCPU index to APIC ID,
1801          * which means kvm_get_vcpu_by_id() has to iterate over all vCPUs
1802          * to find it. Do that once at setup time, instead of every time.
1803          * But beware that on live update / live migration, the routing
1804          * table might be reinstated before the vCPU threads have finished
1805          * recreating their vCPUs.
1806          */
1807         vcpu = kvm_get_vcpu_by_id(kvm, ue->u.xen_evtchn.vcpu);
1808         if (vcpu)
1809                 e->xen_evtchn.vcpu_idx = vcpu->vcpu_idx;
1810         else
1811                 e->xen_evtchn.vcpu_idx = -1;
1812
1813         e->xen_evtchn.port = ue->u.xen_evtchn.port;
1814         e->xen_evtchn.vcpu_id = ue->u.xen_evtchn.vcpu;
1815         e->xen_evtchn.priority = ue->u.xen_evtchn.priority;
1816         e->set = evtchn_set_fn;
1817
1818         return 0;
1819 }
1820
1821 /*
1822  * Explicit event sending from userspace with KVM_XEN_HVM_EVTCHN_SEND ioctl.
1823  */
1824 int kvm_xen_hvm_evtchn_send(struct kvm *kvm, struct kvm_irq_routing_xen_evtchn *uxe)
1825 {
1826         struct kvm_xen_evtchn e;
1827         int ret;
1828
1829         if (!uxe->port || uxe->port >= max_evtchn_port(kvm))
1830                 return -EINVAL;
1831
1832         /* We only support 2 level event channels for now */
1833         if (uxe->priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1834                 return -EINVAL;
1835
1836         e.port = uxe->port;
1837         e.vcpu_id = uxe->vcpu;
1838         e.vcpu_idx = -1;
1839         e.priority = uxe->priority;
1840
1841         ret = kvm_xen_set_evtchn(&e, kvm);
1842
1843         /*
1844          * None of that 'return 1 if it actually got delivered' nonsense.
1845          * We don't care if it was masked (-ENOTCONN) either.
1846          */
1847         if (ret > 0 || ret == -ENOTCONN)
1848                 ret = 0;
1849
1850         return ret;
1851 }
1852
1853 /*
1854  * Support for *outbound* event channel events via the EVTCHNOP_send hypercall.
1855  */
1856 struct evtchnfd {
1857         u32 send_port;
1858         u32 type;
1859         union {
1860                 struct kvm_xen_evtchn port;
1861                 struct {
1862                         u32 port; /* zero */
1863                         struct eventfd_ctx *ctx;
1864                 } eventfd;
1865         } deliver;
1866 };
1867
1868 /*
1869  * Update target vCPU or priority for a registered sending channel.
1870  */
1871 static int kvm_xen_eventfd_update(struct kvm *kvm,
1872                                   struct kvm_xen_hvm_attr *data)
1873 {
1874         u32 port = data->u.evtchn.send_port;
1875         struct evtchnfd *evtchnfd;
1876         int ret;
1877
1878         /* Protect writes to evtchnfd as well as the idr lookup.  */
1879         mutex_lock(&kvm->arch.xen.xen_lock);
1880         evtchnfd = idr_find(&kvm->arch.xen.evtchn_ports, port);
1881
1882         ret = -ENOENT;
1883         if (!evtchnfd)
1884                 goto out_unlock;
1885
1886         /* For an UPDATE, nothing may change except the priority/vcpu */
1887         ret = -EINVAL;
1888         if (evtchnfd->type != data->u.evtchn.type)
1889                 goto out_unlock;
1890
1891         /*
1892          * Port cannot change, and if it's zero that was an eventfd
1893          * which can't be changed either.
1894          */
1895         if (!evtchnfd->deliver.port.port ||
1896             evtchnfd->deliver.port.port != data->u.evtchn.deliver.port.port)
1897                 goto out_unlock;
1898
1899         /* We only support 2 level event channels for now */
1900         if (data->u.evtchn.deliver.port.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1901                 goto out_unlock;
1902
1903         evtchnfd->deliver.port.priority = data->u.evtchn.deliver.port.priority;
1904         if (evtchnfd->deliver.port.vcpu_id != data->u.evtchn.deliver.port.vcpu) {
1905                 evtchnfd->deliver.port.vcpu_id = data->u.evtchn.deliver.port.vcpu;
1906                 evtchnfd->deliver.port.vcpu_idx = -1;
1907         }
1908         ret = 0;
1909 out_unlock:
1910         mutex_unlock(&kvm->arch.xen.xen_lock);
1911         return ret;
1912 }
1913
1914 /*
1915  * Configure the target (eventfd or local port delivery) for sending on
1916  * a given event channel.
1917  */
1918 static int kvm_xen_eventfd_assign(struct kvm *kvm,
1919                                   struct kvm_xen_hvm_attr *data)
1920 {
1921         u32 port = data->u.evtchn.send_port;
1922         struct eventfd_ctx *eventfd = NULL;
1923         struct evtchnfd *evtchnfd;
1924         int ret = -EINVAL;
1925
1926         evtchnfd = kzalloc(sizeof(struct evtchnfd), GFP_KERNEL);
1927         if (!evtchnfd)
1928                 return -ENOMEM;
1929
1930         switch(data->u.evtchn.type) {
1931         case EVTCHNSTAT_ipi:
1932                 /* IPI  must map back to the same port# */
1933                 if (data->u.evtchn.deliver.port.port != data->u.evtchn.send_port)
1934                         goto out_noeventfd; /* -EINVAL */
1935                 break;
1936
1937         case EVTCHNSTAT_interdomain:
1938                 if (data->u.evtchn.deliver.port.port) {
1939                         if (data->u.evtchn.deliver.port.port >= max_evtchn_port(kvm))
1940                                 goto out_noeventfd; /* -EINVAL */
1941                 } else {
1942                         eventfd = eventfd_ctx_fdget(data->u.evtchn.deliver.eventfd.fd);
1943                         if (IS_ERR(eventfd)) {
1944                                 ret = PTR_ERR(eventfd);
1945                                 goto out_noeventfd;
1946                         }
1947                 }
1948                 break;
1949
1950         case EVTCHNSTAT_virq:
1951         case EVTCHNSTAT_closed:
1952         case EVTCHNSTAT_unbound:
1953         case EVTCHNSTAT_pirq:
1954         default: /* Unknown event channel type */
1955                 goto out; /* -EINVAL */
1956         }
1957
1958         evtchnfd->send_port = data->u.evtchn.send_port;
1959         evtchnfd->type = data->u.evtchn.type;
1960         if (eventfd) {
1961                 evtchnfd->deliver.eventfd.ctx = eventfd;
1962         } else {
1963                 /* We only support 2 level event channels for now */
1964                 if (data->u.evtchn.deliver.port.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1965                         goto out; /* -EINVAL; */
1966
1967                 evtchnfd->deliver.port.port = data->u.evtchn.deliver.port.port;
1968                 evtchnfd->deliver.port.vcpu_id = data->u.evtchn.deliver.port.vcpu;
1969                 evtchnfd->deliver.port.vcpu_idx = -1;
1970                 evtchnfd->deliver.port.priority = data->u.evtchn.deliver.port.priority;
1971         }
1972
1973         mutex_lock(&kvm->arch.xen.xen_lock);
1974         ret = idr_alloc(&kvm->arch.xen.evtchn_ports, evtchnfd, port, port + 1,
1975                         GFP_KERNEL);
1976         mutex_unlock(&kvm->arch.xen.xen_lock);
1977         if (ret >= 0)
1978                 return 0;
1979
1980         if (ret == -ENOSPC)
1981                 ret = -EEXIST;
1982 out:
1983         if (eventfd)
1984                 eventfd_ctx_put(eventfd);
1985 out_noeventfd:
1986         kfree(evtchnfd);
1987         return ret;
1988 }
1989
1990 static int kvm_xen_eventfd_deassign(struct kvm *kvm, u32 port)
1991 {
1992         struct evtchnfd *evtchnfd;
1993
1994         mutex_lock(&kvm->arch.xen.xen_lock);
1995         evtchnfd = idr_remove(&kvm->arch.xen.evtchn_ports, port);
1996         mutex_unlock(&kvm->arch.xen.xen_lock);
1997
1998         if (!evtchnfd)
1999                 return -ENOENT;
2000
2001         synchronize_srcu(&kvm->srcu);
2002         if (!evtchnfd->deliver.port.port)
2003                 eventfd_ctx_put(evtchnfd->deliver.eventfd.ctx);
2004         kfree(evtchnfd);
2005         return 0;
2006 }
2007
2008 static int kvm_xen_eventfd_reset(struct kvm *kvm)
2009 {
2010         struct evtchnfd *evtchnfd, **all_evtchnfds;
2011         int i;
2012         int n = 0;
2013
2014         mutex_lock(&kvm->arch.xen.xen_lock);
2015
2016         /*
2017          * Because synchronize_srcu() cannot be called inside the
2018          * critical section, first collect all the evtchnfd objects
2019          * in an array as they are removed from evtchn_ports.
2020          */
2021         idr_for_each_entry(&kvm->arch.xen.evtchn_ports, evtchnfd, i)
2022                 n++;
2023
2024         all_evtchnfds = kmalloc_array(n, sizeof(struct evtchnfd *), GFP_KERNEL);
2025         if (!all_evtchnfds) {
2026                 mutex_unlock(&kvm->arch.xen.xen_lock);
2027                 return -ENOMEM;
2028         }
2029
2030         n = 0;
2031         idr_for_each_entry(&kvm->arch.xen.evtchn_ports, evtchnfd, i) {
2032                 all_evtchnfds[n++] = evtchnfd;
2033                 idr_remove(&kvm->arch.xen.evtchn_ports, evtchnfd->send_port);
2034         }
2035         mutex_unlock(&kvm->arch.xen.xen_lock);
2036
2037         synchronize_srcu(&kvm->srcu);
2038
2039         while (n--) {
2040                 evtchnfd = all_evtchnfds[n];
2041                 if (!evtchnfd->deliver.port.port)
2042                         eventfd_ctx_put(evtchnfd->deliver.eventfd.ctx);
2043                 kfree(evtchnfd);
2044         }
2045         kfree(all_evtchnfds);
2046
2047         return 0;
2048 }
2049
2050 static int kvm_xen_setattr_evtchn(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
2051 {
2052         u32 port = data->u.evtchn.send_port;
2053
2054         if (data->u.evtchn.flags == KVM_XEN_EVTCHN_RESET)
2055                 return kvm_xen_eventfd_reset(kvm);
2056
2057         if (!port || port >= max_evtchn_port(kvm))
2058                 return -EINVAL;
2059
2060         if (data->u.evtchn.flags == KVM_XEN_EVTCHN_DEASSIGN)
2061                 return kvm_xen_eventfd_deassign(kvm, port);
2062         if (data->u.evtchn.flags == KVM_XEN_EVTCHN_UPDATE)
2063                 return kvm_xen_eventfd_update(kvm, data);
2064         if (data->u.evtchn.flags)
2065                 return -EINVAL;
2066
2067         return kvm_xen_eventfd_assign(kvm, data);
2068 }
2069
2070 static bool kvm_xen_hcall_evtchn_send(struct kvm_vcpu *vcpu, u64 param, u64 *r)
2071 {
2072         struct evtchnfd *evtchnfd;
2073         struct evtchn_send send;
2074         struct x86_exception e;
2075
2076         /* Sanity check: this structure is the same for 32-bit and 64-bit */
2077         BUILD_BUG_ON(sizeof(send) != 4);
2078         if (kvm_read_guest_virt(vcpu, param, &send, sizeof(send), &e)) {
2079                 *r = -EFAULT;
2080                 return true;
2081         }
2082
2083         /*
2084          * evtchnfd is protected by kvm->srcu; the idr lookup instead
2085          * is protected by RCU.
2086          */
2087         rcu_read_lock();
2088         evtchnfd = idr_find(&vcpu->kvm->arch.xen.evtchn_ports, send.port);
2089         rcu_read_unlock();
2090         if (!evtchnfd)
2091                 return false;
2092
2093         if (evtchnfd->deliver.port.port) {
2094                 int ret = kvm_xen_set_evtchn(&evtchnfd->deliver.port, vcpu->kvm);
2095                 if (ret < 0 && ret != -ENOTCONN)
2096                         return false;
2097         } else {
2098                 eventfd_signal(evtchnfd->deliver.eventfd.ctx);
2099         }
2100
2101         *r = 0;
2102         return true;
2103 }
2104
2105 void kvm_xen_init_vcpu(struct kvm_vcpu *vcpu)
2106 {
2107         vcpu->arch.xen.vcpu_id = vcpu->vcpu_idx;
2108         vcpu->arch.xen.poll_evtchn = 0;
2109
2110         timer_setup(&vcpu->arch.xen.poll_timer, cancel_evtchn_poll, 0);
2111
2112         kvm_gpc_init(&vcpu->arch.xen.runstate_cache, vcpu->kvm, NULL,
2113                      KVM_HOST_USES_PFN);
2114         kvm_gpc_init(&vcpu->arch.xen.runstate2_cache, vcpu->kvm, NULL,
2115                      KVM_HOST_USES_PFN);
2116         kvm_gpc_init(&vcpu->arch.xen.vcpu_info_cache, vcpu->kvm, NULL,
2117                      KVM_HOST_USES_PFN);
2118         kvm_gpc_init(&vcpu->arch.xen.vcpu_time_info_cache, vcpu->kvm, NULL,
2119                      KVM_HOST_USES_PFN);
2120 }
2121
2122 void kvm_xen_destroy_vcpu(struct kvm_vcpu *vcpu)
2123 {
2124         if (kvm_xen_timer_enabled(vcpu))
2125                 kvm_xen_stop_timer(vcpu);
2126
2127         kvm_gpc_deactivate(&vcpu->arch.xen.runstate_cache);
2128         kvm_gpc_deactivate(&vcpu->arch.xen.runstate2_cache);
2129         kvm_gpc_deactivate(&vcpu->arch.xen.vcpu_info_cache);
2130         kvm_gpc_deactivate(&vcpu->arch.xen.vcpu_time_info_cache);
2131
2132         del_timer_sync(&vcpu->arch.xen.poll_timer);
2133 }
2134
2135 void kvm_xen_update_tsc_info(struct kvm_vcpu *vcpu)
2136 {
2137         struct kvm_cpuid_entry2 *entry;
2138         u32 function;
2139
2140         if (!vcpu->arch.xen.cpuid.base)
2141                 return;
2142
2143         function = vcpu->arch.xen.cpuid.base | XEN_CPUID_LEAF(3);
2144         if (function > vcpu->arch.xen.cpuid.limit)
2145                 return;
2146
2147         entry = kvm_find_cpuid_entry_index(vcpu, function, 1);
2148         if (entry) {
2149                 entry->ecx = vcpu->arch.hv_clock.tsc_to_system_mul;
2150                 entry->edx = vcpu->arch.hv_clock.tsc_shift;
2151         }
2152
2153         entry = kvm_find_cpuid_entry_index(vcpu, function, 2);
2154         if (entry)
2155                 entry->eax = vcpu->arch.hw_tsc_khz;
2156 }
2157
2158 void kvm_xen_init_vm(struct kvm *kvm)
2159 {
2160         mutex_init(&kvm->arch.xen.xen_lock);
2161         idr_init(&kvm->arch.xen.evtchn_ports);
2162         kvm_gpc_init(&kvm->arch.xen.shinfo_cache, kvm, NULL, KVM_HOST_USES_PFN);
2163 }
2164
2165 void kvm_xen_destroy_vm(struct kvm *kvm)
2166 {
2167         struct evtchnfd *evtchnfd;
2168         int i;
2169
2170         kvm_gpc_deactivate(&kvm->arch.xen.shinfo_cache);
2171
2172         idr_for_each_entry(&kvm->arch.xen.evtchn_ports, evtchnfd, i) {
2173                 if (!evtchnfd->deliver.port.port)
2174                         eventfd_ctx_put(evtchnfd->deliver.eventfd.ctx);
2175                 kfree(evtchnfd);
2176         }
2177         idr_destroy(&kvm->arch.xen.evtchn_ports);
2178
2179         if (kvm->arch.xen_hvm_config.msr)
2180                 static_branch_slow_dec_deferred(&kvm_xen_enabled);
2181 }