target/arm/kvm.c

   1 /*
   2  * ARM implementation of KVM hooks
   3  *
   4  * Copyright Christoffer Dall 2009-2010
   5  *
   6  * This work is licensed under the terms of the GNU GPL, version 2 or later.
   7  * See the COPYING file in the top-level directory.
   8  *
   9  */
  10
  11 #include "qemu/osdep.h"
  12 #include <sys/ioctl.h>
  13
  14 #include <linux/kvm.h>
  15
  16 #include "qemu-common.h"
  17 #include "qemu/timer.h"
  18 #include "qemu/error-report.h"
  19 #include "qemu/main-loop.h"
  20 #include "sysemu/sysemu.h"
  21 #include "sysemu/kvm.h"
  22 #include "sysemu/kvm_int.h"
  23 #include "kvm_arm.h"
  24 #include "cpu.h"
  25 #include "trace.h"
  26 #include "internals.h"
  27 #include "hw/pci/pci.h"
  28 #include "exec/memattrs.h"
  29 #include "exec/address-spaces.h"
  30 #include "hw/boards.h"
  31 #include "hw/irq.h"
  32 #include "qemu/log.h"
  33
  34 const KVMCapabilityInfo kvm_arch_required_capabilities[] = {
  35     KVM_CAP_LAST_INFO
  36 };
  37
  38 static bool cap_has_mp_state;
  39 static bool cap_has_inject_serror_esr;
  40
  41 static ARMHostCPUFeatures arm_host_cpu_features;
  42
  43 int kvm_arm_vcpu_init(CPUState *cs)
  44 {
  45     ARMCPU *cpu = ARM_CPU(cs);
  46     struct kvm_vcpu_init init;
  47
  48     init.target = cpu->kvm_target;
  49     memcpy(init.features, cpu->kvm_init_features, sizeof(init.features));
  50
  51     return kvm_vcpu_ioctl(cs, KVM_ARM_VCPU_INIT, &init);
  52 }
  53
  54 int kvm_arm_vcpu_finalize(CPUState *cs, int feature)
  55 {
  56     return kvm_vcpu_ioctl(cs, KVM_ARM_VCPU_FINALIZE, &feature);
  57 }
  58
  59 void kvm_arm_init_serror_injection(CPUState *cs)
  60 {
  61     cap_has_inject_serror_esr = kvm_check_extension(cs->kvm_state,
  62                                     KVM_CAP_ARM_INJECT_SERROR_ESR);
  63 }
  64
  65 bool kvm_arm_create_scratch_host_vcpu(const uint32_t *cpus_to_try,
  66                                       int *fdarray,
  67                                       struct kvm_vcpu_init *init)
  68 {
  69     int ret = 0, kvmfd = -1, vmfd = -1, cpufd = -1;
  70
  71     kvmfd = qemu_open("/dev/kvm", O_RDWR);
  72     if (kvmfd < 0) {
  73         goto err;
  74     }
  75     vmfd = ioctl(kvmfd, KVM_CREATE_VM, 0);
  76     if (vmfd < 0) {
  77         goto err;
  78     }
  79     cpufd = ioctl(vmfd, KVM_CREATE_VCPU, 0);
  80     if (cpufd < 0) {
  81         goto err;
  82     }
  83
  84     if (!init) {
  85         /* Caller doesn't want the VCPU to be initialized, so skip it */
  86         goto finish;
  87     }
  88
  89     if (init->target == -1) {
  90         struct kvm_vcpu_init preferred;
  91
  92         ret = ioctl(vmfd, KVM_ARM_PREFERRED_TARGET, &preferred);
  93         if (!ret) {
  94             init->target = preferred.target;
  95         }
  96     }
  97     if (ret >= 0) {
  98         ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, init);
  99         if (ret < 0) {
 100             goto err;
 101         }
 102     } else if (cpus_to_try) {
 103         /* Old kernel which doesn't know about the
 104          * PREFERRED_TARGET ioctl: we know it will only support
 105          * creating one kind of guest CPU which is its preferred
 106          * CPU type.
 107          */
 108         struct kvm_vcpu_init try;
 109
 110         while (*cpus_to_try != QEMU_KVM_ARM_TARGET_NONE) {
 111             try.target = *cpus_to_try++;
 112             memcpy(try.features, init->features, sizeof(init->features));
 113             ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, &try);
 114             if (ret >= 0) {
 115                 break;
 116             }
 117         }
 118         if (ret < 0) {
 119             goto err;
 120         }
 121         init->target = try.target;
 122     } else {
 123         /* Treat a NULL cpus_to_try argument the same as an empty
 124          * list, which means we will fail the call since this must
 125          * be an old kernel which doesn't support PREFERRED_TARGET.
 126          */
 127         goto err;
 128     }
 129
 130 finish:
 131     fdarray[0] = kvmfd;
 132     fdarray[1] = vmfd;
 133     fdarray[2] = cpufd;
 134
 135     return true;
 136
 137 err:
 138     if (cpufd >= 0) {
 139         close(cpufd);
 140     }
 141     if (vmfd >= 0) {
 142         close(vmfd);
 143     }
 144     if (kvmfd >= 0) {
 145         close(kvmfd);
 146     }
 147
 148     return false;
 149 }
 150
 151 void kvm_arm_destroy_scratch_host_vcpu(int *fdarray)
 152 {
 153     int i;
 154
 155     for (i = 2; i >= 0; i--) {
 156         close(fdarray[i]);
 157     }
 158 }
 159
 160 void kvm_arm_set_cpu_features_from_host(ARMCPU *cpu)
 161 {
 162     CPUARMState *env = &cpu->env;
 163
 164     if (!arm_host_cpu_features.dtb_compatible) {
 165         if (!kvm_enabled() ||
 166             !kvm_arm_get_host_cpu_features(&arm_host_cpu_features)) {
 167             /* We can't report this error yet, so flag that we need to
 168              * in arm_cpu_realizefn().
 169              */
 170             cpu->kvm_target = QEMU_KVM_ARM_TARGET_NONE;
 171             cpu->host_cpu_probe_failed = true;
 172             return;
 173         }
 174     }
 175
 176     cpu->kvm_target = arm_host_cpu_features.target;
 177     cpu->dtb_compatible = arm_host_cpu_features.dtb_compatible;
 178     cpu->isar = arm_host_cpu_features.isar;
 179     env->features = arm_host_cpu_features.features;
 180 }
 181
 182 bool kvm_arm_pmu_supported(CPUState *cpu)
 183 {
 184     KVMState *s = KVM_STATE(current_machine->accelerator);
 185
 186     return kvm_check_extension(s, KVM_CAP_ARM_PMU_V3);
 187 }
 188
 189 int kvm_arm_get_max_vm_ipa_size(MachineState *ms)
 190 {
 191     KVMState *s = KVM_STATE(ms->accelerator);
 192     int ret;
 193
 194     ret = kvm_check_extension(s, KVM_CAP_ARM_VM_IPA_SIZE);
 195     return ret > 0 ? ret : 40;
 196 }
 197
 198 int kvm_arch_init(MachineState *ms, KVMState *s)
 199 {
 200     int ret = 0;
 201     /* For ARM interrupt delivery is always asynchronous,
 202      * whether we are using an in-kernel VGIC or not.
 203      */
 204     kvm_async_interrupts_allowed = true;
 205
 206     /*
 207      * PSCI wakes up secondary cores, so we always need to
 208      * have vCPUs waiting in kernel space
 209      */
 210     kvm_halt_in_kernel_allowed = true;
 211
 212     cap_has_mp_state = kvm_check_extension(s, KVM_CAP_MP_STATE);
 213
 214     if (ms->smp.cpus > 256 &&
 215         !kvm_check_extension(s, KVM_CAP_ARM_IRQ_LINE_LAYOUT_2)) {
 216         error_report("Using more than 256 vcpus requires a host kernel "
 217                      "with KVM_CAP_ARM_IRQ_LINE_LAYOUT_2");
 218         ret = -EINVAL;
 219     }
 220
 221     return ret;
 222 }
 223
 224 unsigned long kvm_arch_vcpu_id(CPUState *cpu)
 225 {
 226     return cpu->cpu_index;
 227 }
 228
 229 /* We track all the KVM devices which need their memory addresses
 230  * passing to the kernel in a list of these structures.
 231  * When board init is complete we run through the list and
 232  * tell the kernel the base addresses of the memory regions.
 233  * We use a MemoryListener to track mapping and unmapping of
 234  * the regions during board creation, so the board models don't
 235  * need to do anything special for the KVM case.
 236  *
 237  * Sometimes the address must be OR'ed with some other fields
 238  * (for example for KVM_VGIC_V3_ADDR_TYPE_REDIST_REGION).
 239  * @kda_addr_ormask aims at storing the value of those fields.
 240  */
 241 typedef struct KVMDevice {
 242     struct kvm_arm_device_addr kda;
 243     struct kvm_device_attr kdattr;
 244     uint64_t kda_addr_ormask;
 245     MemoryRegion *mr;
 246     QSLIST_ENTRY(KVMDevice) entries;
 247     int dev_fd;
 248 } KVMDevice;
 249
 250 static QSLIST_HEAD(, KVMDevice) kvm_devices_head;
 251
 252 static void kvm_arm_devlistener_add(MemoryListener *listener,
 253                                     MemoryRegionSection *section)
 254 {
 255     KVMDevice *kd;
 256
 257     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
 258         if (section->mr == kd->mr) {
 259             kd->kda.addr = section->offset_within_address_space;
 260         }
 261     }
 262 }
 263
 264 static void kvm_arm_devlistener_del(MemoryListener *listener,
 265                                     MemoryRegionSection *section)
 266 {
 267     KVMDevice *kd;
 268
 269     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
 270         if (section->mr == kd->mr) {
 271             kd->kda.addr = -1;
 272         }
 273     }
 274 }
 275
 276 static MemoryListener devlistener = {
 277     .region_add = kvm_arm_devlistener_add,
 278     .region_del = kvm_arm_devlistener_del,
 279 };
 280
 281 static void kvm_arm_set_device_addr(KVMDevice *kd)
 282 {
 283     struct kvm_device_attr *attr = &kd->kdattr;
 284     int ret;
 285
 286     /* If the device control API is available and we have a device fd on the
 287      * KVMDevice struct, let's use the newer API
 288      */
 289     if (kd->dev_fd >= 0) {
 290         uint64_t addr = kd->kda.addr;
 291
 292         addr |= kd->kda_addr_ormask;
 293         attr->addr = (uintptr_t)&addr;
 294         ret = kvm_device_ioctl(kd->dev_fd, KVM_SET_DEVICE_ATTR, attr);
 295     } else {
 296         ret = kvm_vm_ioctl(kvm_state, KVM_ARM_SET_DEVICE_ADDR, &kd->kda);
 297     }
 298
 299     if (ret < 0) {
 300         fprintf(stderr, "Failed to set device address: %s\n",
 301                 strerror(-ret));
 302         abort();
 303     }
 304 }
 305
 306 static void kvm_arm_machine_init_done(Notifier *notifier, void *data)
 307 {
 308     KVMDevice *kd, *tkd;
 309
 310     QSLIST_FOREACH_SAFE(kd, &kvm_devices_head, entries, tkd) {
 311         if (kd->kda.addr != -1) {
 312             kvm_arm_set_device_addr(kd);
 313         }
 314         memory_region_unref(kd->mr);
 315         QSLIST_REMOVE_HEAD(&kvm_devices_head, entries);
 316         g_free(kd);
 317     }
 318     memory_listener_unregister(&devlistener);
 319 }
 320
 321 static Notifier notify = {
 322     .notify = kvm_arm_machine_init_done,
 323 };
 324
 325 void kvm_arm_register_device(MemoryRegion *mr, uint64_t devid, uint64_t group,
 326                              uint64_t attr, int dev_fd, uint64_t addr_ormask)
 327 {
 328     KVMDevice *kd;
 329
 330     if (!kvm_irqchip_in_kernel()) {
 331         return;
 332     }
 333
 334     if (QSLIST_EMPTY(&kvm_devices_head)) {
 335         memory_listener_register(&devlistener, &address_space_memory);
 336         qemu_add_machine_init_done_notifier(&notify);
 337     }
 338     kd = g_new0(KVMDevice, 1);
 339     kd->mr = mr;
 340     kd->kda.id = devid;
 341     kd->kda.addr = -1;
 342     kd->kdattr.flags = 0;
 343     kd->kdattr.group = group;
 344     kd->kdattr.attr = attr;
 345     kd->dev_fd = dev_fd;
 346     kd->kda_addr_ormask = addr_ormask;
 347     QSLIST_INSERT_HEAD(&kvm_devices_head, kd, entries);
 348     memory_region_ref(kd->mr);
 349 }
 350
 351 static int compare_u64(const void *a, const void *b)
 352 {
 353     if (*(uint64_t *)a > *(uint64_t *)b) {
 354         return 1;
 355     }
 356     if (*(uint64_t *)a < *(uint64_t *)b) {
 357         return -1;
 358     }
 359     return 0;
 360 }
 361
 362 /* Initialize the ARMCPU cpreg list according to the kernel's
 363  * definition of what CPU registers it knows about (and throw away
 364  * the previous TCG-created cpreg list).
 365  */
 366 int kvm_arm_init_cpreg_list(ARMCPU *cpu)
 367 {
 368     struct kvm_reg_list rl;
 369     struct kvm_reg_list *rlp;
 370     int i, ret, arraylen;
 371     CPUState *cs = CPU(cpu);
 372
 373     rl.n = 0;
 374     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, &rl);
 375     if (ret != -E2BIG) {
 376         return ret;
 377     }
 378     rlp = g_malloc(sizeof(struct kvm_reg_list) + rl.n * sizeof(uint64_t));
 379     rlp->n = rl.n;
 380     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, rlp);
 381     if (ret) {
 382         goto out;
 383     }
 384     /* Sort the list we get back from the kernel, since cpreg_tuples
 385      * must be in strictly ascending order.
 386      */
 387     qsort(&rlp->reg, rlp->n, sizeof(rlp->reg[0]), compare_u64);
 388
 389     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 390         if (!kvm_arm_reg_syncs_via_cpreg_list(rlp->reg[i])) {
 391             continue;
 392         }
 393         switch (rlp->reg[i] & KVM_REG_SIZE_MASK) {
 394         case KVM_REG_SIZE_U32:
 395         case KVM_REG_SIZE_U64:
 396             break;
 397         default:
 398             fprintf(stderr, "Can't handle size of register in kernel list\n");
 399             ret = -EINVAL;
 400             goto out;
 401         }
 402
 403         arraylen++;
 404     }
 405
 406     cpu->cpreg_indexes = g_renew(uint64_t, cpu->cpreg_indexes, arraylen);
 407     cpu->cpreg_values = g_renew(uint64_t, cpu->cpreg_values, arraylen);
 408     cpu->cpreg_vmstate_indexes = g_renew(uint64_t, cpu->cpreg_vmstate_indexes,
 409                                          arraylen);
 410     cpu->cpreg_vmstate_values = g_renew(uint64_t, cpu->cpreg_vmstate_values,
 411                                         arraylen);
 412     cpu->cpreg_array_len = arraylen;
 413     cpu->cpreg_vmstate_array_len = arraylen;
 414
 415     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 416         uint64_t regidx = rlp->reg[i];
 417         if (!kvm_arm_reg_syncs_via_cpreg_list(regidx)) {
 418             continue;
 419         }
 420         cpu->cpreg_indexes[arraylen] = regidx;
 421         arraylen++;
 422     }
 423     assert(cpu->cpreg_array_len == arraylen);
 424
 425     if (!write_kvmstate_to_list(cpu)) {
 426         /* Shouldn't happen unless kernel is inconsistent about
 427          * what registers exist.
 428          */
 429         fprintf(stderr, "Initial read of kernel register state failed\n");
 430         ret = -EINVAL;
 431         goto out;
 432     }
 433
 434 out:
 435     g_free(rlp);
 436     return ret;
 437 }
 438
 439 bool write_kvmstate_to_list(ARMCPU *cpu)
 440 {
 441     CPUState *cs = CPU(cpu);
 442     int i;
 443     bool ok = true;
 444
 445     for (i = 0; i < cpu->cpreg_array_len; i++) {
 446         struct kvm_one_reg r;
 447         uint64_t regidx = cpu->cpreg_indexes[i];
 448         uint32_t v32;
 449         int ret;
 450
 451         r.id = regidx;
 452
 453         switch (regidx & KVM_REG_SIZE_MASK) {
 454         case KVM_REG_SIZE_U32:
 455             r.addr = (uintptr_t)&v32;
 456             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 457             if (!ret) {
 458                 cpu->cpreg_values[i] = v32;
 459             }
 460             break;
 461         case KVM_REG_SIZE_U64:
 462             r.addr = (uintptr_t)(cpu->cpreg_values + i);
 463             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 464             break;
 465         default:
 466             abort();
 467         }
 468         if (ret) {
 469             ok = false;
 470         }
 471     }
 472     return ok;
 473 }
 474
 475 bool write_list_to_kvmstate(ARMCPU *cpu, int level)
 476 {
 477     CPUState *cs = CPU(cpu);
 478     int i;
 479     bool ok = true;
 480
 481     for (i = 0; i < cpu->cpreg_array_len; i++) {
 482         struct kvm_one_reg r;
 483         uint64_t regidx = cpu->cpreg_indexes[i];
 484         uint32_t v32;
 485         int ret;
 486
 487         if (kvm_arm_cpreg_level(regidx) > level) {
 488             continue;
 489         }
 490
 491         r.id = regidx;
 492         switch (regidx & KVM_REG_SIZE_MASK) {
 493         case KVM_REG_SIZE_U32:
 494             v32 = cpu->cpreg_values[i];
 495             r.addr = (uintptr_t)&v32;
 496             break;
 497         case KVM_REG_SIZE_U64:
 498             r.addr = (uintptr_t)(cpu->cpreg_values + i);
 499             break;
 500         default:
 501             abort();
 502         }
 503         ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
 504         if (ret) {
 505             /* We might fail for "unknown register" and also for
 506              * "you tried to set a register which is constant with
 507              * a different value from what it actually contains".
 508              */
 509             ok = false;
 510         }
 511     }
 512     return ok;
 513 }
 514
 515 void kvm_arm_reset_vcpu(ARMCPU *cpu)
 516 {
 517     int ret;
 518
 519     /* Re-init VCPU so that all registers are set to
 520      * their respective reset values.
 521      */
 522     ret = kvm_arm_vcpu_init(CPU(cpu));
 523     if (ret < 0) {
 524         fprintf(stderr, "kvm_arm_vcpu_init failed: %s\n", strerror(-ret));
 525         abort();
 526     }
 527     if (!write_kvmstate_to_list(cpu)) {
 528         fprintf(stderr, "write_kvmstate_to_list failed\n");
 529         abort();
 530     }
 531     /*
 532      * Sync the reset values also into the CPUState. This is necessary
 533      * because the next thing we do will be a kvm_arch_put_registers()
 534      * which will update the list values from the CPUState before copying
 535      * the list values back to KVM. It's OK to ignore failure returns here
 536      * for the same reason we do so in kvm_arch_get_registers().
 537      */
 538     write_list_to_cpustate(cpu);
 539 }
 540
 541 /*
 542  * Update KVM's MP_STATE based on what QEMU thinks it is
 543  */
 544 int kvm_arm_sync_mpstate_to_kvm(ARMCPU *cpu)
 545 {
 546     if (cap_has_mp_state) {
 547         struct kvm_mp_state mp_state = {
 548             .mp_state = (cpu->power_state == PSCI_OFF) ?
 549             KVM_MP_STATE_STOPPED : KVM_MP_STATE_RUNNABLE
 550         };
 551         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_SET_MP_STATE, &mp_state);
 552         if (ret) {
 553             fprintf(stderr, "%s: failed to set MP_STATE %d/%s\n",
 554                     __func__, ret, strerror(-ret));
 555             return -1;
 556         }
 557     }
 558
 559     return 0;
 560 }
 561
 562 /*
 563  * Sync the KVM MP_STATE into QEMU
 564  */
 565 int kvm_arm_sync_mpstate_to_qemu(ARMCPU *cpu)
 566 {
 567     if (cap_has_mp_state) {
 568         struct kvm_mp_state mp_state;
 569         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_GET_MP_STATE, &mp_state);
 570         if (ret) {
 571             fprintf(stderr, "%s: failed to get MP_STATE %d/%s\n",
 572                     __func__, ret, strerror(-ret));
 573             abort();
 574         }
 575         cpu->power_state = (mp_state.mp_state == KVM_MP_STATE_STOPPED) ?
 576             PSCI_OFF : PSCI_ON;
 577     }
 578
 579     return 0;
 580 }
 581
 582 int kvm_put_vcpu_events(ARMCPU *cpu)
 583 {
 584     CPUARMState *env = &cpu->env;
 585     struct kvm_vcpu_events events;
 586     int ret;
 587
 588     if (!kvm_has_vcpu_events()) {
 589         return 0;
 590     }
 591
 592     memset(&events, 0, sizeof(events));
 593     events.exception.serror_pending = env->serror.pending;
 594
 595     /* Inject SError to guest with specified syndrome if host kernel
 596      * supports it, otherwise inject SError without syndrome.
 597      */
 598     if (cap_has_inject_serror_esr) {
 599         events.exception.serror_has_esr = env->serror.has_esr;
 600         events.exception.serror_esr = env->serror.esr;
 601     }
 602
 603     ret = kvm_vcpu_ioctl(CPU(cpu), KVM_SET_VCPU_EVENTS, &events);
 604     if (ret) {
 605         error_report("failed to put vcpu events");
 606     }
 607
 608     return ret;
 609 }
 610
 611 int kvm_get_vcpu_events(ARMCPU *cpu)
 612 {
 613     CPUARMState *env = &cpu->env;
 614     struct kvm_vcpu_events events;
 615     int ret;
 616
 617     if (!kvm_has_vcpu_events()) {
 618         return 0;
 619     }
 620
 621     memset(&events, 0, sizeof(events));
 622     ret = kvm_vcpu_ioctl(CPU(cpu), KVM_GET_VCPU_EVENTS, &events);
 623     if (ret) {
 624         error_report("failed to get vcpu events");
 625         return ret;
 626     }
 627
 628     env->serror.pending = events.exception.serror_pending;
 629     env->serror.has_esr = events.exception.serror_has_esr;
 630     env->serror.esr = events.exception.serror_esr;
 631
 632     return 0;
 633 }
 634
 635 void kvm_arch_pre_run(CPUState *cs, struct kvm_run *run)
 636 {
 637 }
 638
 639 MemTxAttrs kvm_arch_post_run(CPUState *cs, struct kvm_run *run)
 640 {
 641     ARMCPU *cpu;
 642     uint32_t switched_level;
 643
 644     if (kvm_irqchip_in_kernel()) {
 645         /*
 646          * We only need to sync timer states with user-space interrupt
 647          * controllers, so return early and save cycles if we don't.
 648          */
 649         return MEMTXATTRS_UNSPECIFIED;
 650     }
 651
 652     cpu = ARM_CPU(cs);
 653
 654     /* Synchronize our shadowed in-kernel device irq lines with the kvm ones */
 655     if (run->s.regs.device_irq_level != cpu->device_irq_level) {
 656         switched_level = cpu->device_irq_level ^ run->s.regs.device_irq_level;
 657
 658         qemu_mutex_lock_iothread();
 659
 660         if (switched_level & KVM_ARM_DEV_EL1_VTIMER) {
 661             qemu_set_irq(cpu->gt_timer_outputs[GTIMER_VIRT],
 662                          !!(run->s.regs.device_irq_level &
 663                             KVM_ARM_DEV_EL1_VTIMER));
 664             switched_level &= ~KVM_ARM_DEV_EL1_VTIMER;
 665         }
 666
 667         if (switched_level & KVM_ARM_DEV_EL1_PTIMER) {
 668             qemu_set_irq(cpu->gt_timer_outputs[GTIMER_PHYS],
 669                          !!(run->s.regs.device_irq_level &
 670                             KVM_ARM_DEV_EL1_PTIMER));
 671             switched_level &= ~KVM_ARM_DEV_EL1_PTIMER;
 672         }
 673
 674         if (switched_level & KVM_ARM_DEV_PMU) {
 675             qemu_set_irq(cpu->pmu_interrupt,
 676                          !!(run->s.regs.device_irq_level & KVM_ARM_DEV_PMU));
 677             switched_level &= ~KVM_ARM_DEV_PMU;
 678         }
 679
 680         if (switched_level) {
 681             qemu_log_mask(LOG_UNIMP, "%s: unhandled in-kernel device IRQ %x\n",
 682                           __func__, switched_level);
 683         }
 684
 685         /* We also mark unknown levels as processed to not waste cycles */
 686         cpu->device_irq_level = run->s.regs.device_irq_level;
 687         qemu_mutex_unlock_iothread();
 688     }
 689
 690     return MEMTXATTRS_UNSPECIFIED;
 691 }
 692
 693
 694 int kvm_arch_handle_exit(CPUState *cs, struct kvm_run *run)
 695 {
 696     int ret = 0;
 697
 698     switch (run->exit_reason) {
 699     case KVM_EXIT_DEBUG:
 700         if (kvm_arm_handle_debug(cs, &run->debug.arch)) {
 701             ret = EXCP_DEBUG;
 702         } /* otherwise return to guest */
 703         break;
 704     default:
 705         qemu_log_mask(LOG_UNIMP, "%s: un-handled exit reason %d\n",
 706                       __func__, run->exit_reason);
 707         break;
 708     }
 709     return ret;
 710 }
 711
 712 bool kvm_arch_stop_on_emulation_error(CPUState *cs)
 713 {
 714     return true;
 715 }
 716
 717 int kvm_arch_process_async_events(CPUState *cs)
 718 {
 719     return 0;
 720 }
 721
 722 /* The #ifdef protections are until 32bit headers are imported and can
 723  * be removed once both 32 and 64 bit reach feature parity.
 724  */
 725 void kvm_arch_update_guest_debug(CPUState *cs, struct kvm_guest_debug *dbg)
 726 {
 727 #ifdef KVM_GUESTDBG_USE_SW_BP
 728     if (kvm_sw_breakpoints_active(cs)) {
 729         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_SW_BP;
 730     }
 731 #endif
 732 #ifdef KVM_GUESTDBG_USE_HW
 733     if (kvm_arm_hw_debug_active(cs)) {
 734         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_HW;
 735         kvm_arm_copy_hw_debug_data(&dbg->arch);
 736     }
 737 #endif
 738 }
 739
 740 void kvm_arch_init_irq_routing(KVMState *s)
 741 {
 742 }
 743
 744 int kvm_arch_irqchip_create(MachineState *ms, KVMState *s)
 745 {
 746      if (machine_kernel_irqchip_split(ms)) {
 747          perror("-machine kernel_irqchip=split is not supported on ARM.");
 748          exit(1);
 749     }
 750
 751     /* If we can create the VGIC using the newer device control API, we
 752      * let the device do this when it initializes itself, otherwise we
 753      * fall back to the old API */
 754     return kvm_check_extension(s, KVM_CAP_DEVICE_CTRL);
 755 }
 756
 757 int kvm_arm_vgic_probe(void)
 758 {
 759     if (kvm_create_device(kvm_state,
 760                           KVM_DEV_TYPE_ARM_VGIC_V3, true) == 0) {
 761         return 3;
 762     } else if (kvm_create_device(kvm_state,
 763                                  KVM_DEV_TYPE_ARM_VGIC_V2, true) == 0) {
 764         return 2;
 765     } else {
 766         return 0;
 767     }
 768 }
 769
 770 int kvm_arm_set_irq(int cpu, int irqtype, int irq, int level)
 771 {
 772     int kvm_irq = (irqtype << KVM_ARM_IRQ_TYPE_SHIFT) | irq;
 773     int cpu_idx1 = cpu % 256;
 774     int cpu_idx2 = cpu / 256;
 775
 776     kvm_irq |= (cpu_idx1 << KVM_ARM_IRQ_VCPU_SHIFT) |
 777                (cpu_idx2 << KVM_ARM_IRQ_VCPU2_SHIFT);
 778
 779     return kvm_set_irq(kvm_state, kvm_irq, !!level);
 780 }
 781
 782 int kvm_arch_fixup_msi_route(struct kvm_irq_routing_entry *route,
 783                              uint64_t address, uint32_t data, PCIDevice *dev)
 784 {
 785     AddressSpace *as = pci_device_iommu_address_space(dev);
 786     hwaddr xlat, len, doorbell_gpa;
 787     MemoryRegionSection mrs;
 788     MemoryRegion *mr;
 789     int ret = 1;
 790
 791     if (as == &address_space_memory) {
 792         return 0;
 793     }
 794
 795     /* MSI doorbell address is translated by an IOMMU */
 796
 797     rcu_read_lock();
 798     mr = address_space_translate(as, address, &xlat, &len, true,
 799                                  MEMTXATTRS_UNSPECIFIED);
 800     if (!mr) {
 801         goto unlock;
 802     }
 803     mrs = memory_region_find(mr, xlat, 1);
 804     if (!mrs.mr) {
 805         goto unlock;
 806     }
 807
 808     doorbell_gpa = mrs.offset_within_address_space;
 809     memory_region_unref(mrs.mr);
 810
 811     route->u.msi.address_lo = doorbell_gpa;
 812     route->u.msi.address_hi = doorbell_gpa >> 32;
 813
 814     trace_kvm_arm_fixup_msi_route(address, doorbell_gpa);
 815
 816     ret = 0;
 817
 818 unlock:
 819     rcu_read_unlock();
 820     return ret;
 821 }
 822
 823 int kvm_arch_add_msi_route_post(struct kvm_irq_routing_entry *route,
 824                                 int vector, PCIDevice *dev)
 825 {
 826     return 0;
 827 }
 828
 829 int kvm_arch_release_virq_post(int virq)
 830 {
 831     return 0;
 832 }
 833
 834 int kvm_arch_msi_data_to_gsi(uint32_t data)
 835 {
 836     return (data - 32) & 0xffff;
 837 }