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/timer.h"
  17 #include "qemu/error-report.h"
  18 #include "qemu/main-loop.h"
  19 #include "qom/object.h"
  20 #include "qapi/error.h"
  21 #include "sysemu/sysemu.h"
  22 #include "sysemu/kvm.h"
  23 #include "sysemu/kvm_int.h"
  24 #include "kvm_arm.h"
  25 #include "cpu.h"
  26 #include "trace.h"
  27 #include "internals.h"
  28 #include "hw/pci/pci.h"
  29 #include "exec/memattrs.h"
  30 #include "exec/address-spaces.h"
  31 #include "hw/boards.h"
  32 #include "hw/irq.h"
  33 #include "qemu/log.h"
  34
  35 const KVMCapabilityInfo kvm_arch_required_capabilities[] = {
  36     KVM_CAP_LAST_INFO
  37 };
  38
  39 static bool cap_has_mp_state;
  40 static bool cap_has_inject_serror_esr;
  41 static bool cap_has_inject_ext_dabt;
  42
  43 static ARMHostCPUFeatures arm_host_cpu_features;
  44
  45 int kvm_arm_vcpu_init(CPUState *cs)
  46 {
  47     ARMCPU *cpu = ARM_CPU(cs);
  48     struct kvm_vcpu_init init;
  49
  50     init.target = cpu->kvm_target;
  51     memcpy(init.features, cpu->kvm_init_features, sizeof(init.features));
  52
  53     return kvm_vcpu_ioctl(cs, KVM_ARM_VCPU_INIT, &init);
  54 }
  55
  56 int kvm_arm_vcpu_finalize(CPUState *cs, int feature)
  57 {
  58     return kvm_vcpu_ioctl(cs, KVM_ARM_VCPU_FINALIZE, &feature);
  59 }
  60
  61 void kvm_arm_init_serror_injection(CPUState *cs)
  62 {
  63     cap_has_inject_serror_esr = kvm_check_extension(cs->kvm_state,
  64                                     KVM_CAP_ARM_INJECT_SERROR_ESR);
  65 }
  66
  67 bool kvm_arm_create_scratch_host_vcpu(const uint32_t *cpus_to_try,
  68                                       int *fdarray,
  69                                       struct kvm_vcpu_init *init)
  70 {
  71     int ret = 0, kvmfd = -1, vmfd = -1, cpufd = -1;
  72     int max_vm_pa_size;
  73
  74     kvmfd = qemu_open_old("/dev/kvm", O_RDWR);
  75     if (kvmfd < 0) {
  76         goto err;
  77     }
  78     max_vm_pa_size = ioctl(kvmfd, KVM_CHECK_EXTENSION, KVM_CAP_ARM_VM_IPA_SIZE);
  79     if (max_vm_pa_size < 0) {
  80         max_vm_pa_size = 0;
  81     }
  82     vmfd = ioctl(kvmfd, KVM_CREATE_VM, max_vm_pa_size);
  83     if (vmfd < 0) {
  84         goto err;
  85     }
  86     cpufd = ioctl(vmfd, KVM_CREATE_VCPU, 0);
  87     if (cpufd < 0) {
  88         goto err;
  89     }
  90
  91     if (!init) {
  92         /* Caller doesn't want the VCPU to be initialized, so skip it */
  93         goto finish;
  94     }
  95
  96     if (init->target == -1) {
  97         struct kvm_vcpu_init preferred;
  98
  99         ret = ioctl(vmfd, KVM_ARM_PREFERRED_TARGET, &preferred);
 100         if (!ret) {
 101             init->target = preferred.target;
 102         }
 103     }
 104     if (ret >= 0) {
 105         ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, init);
 106         if (ret < 0) {
 107             goto err;
 108         }
 109     } else if (cpus_to_try) {
 110         /* Old kernel which doesn't know about the
 111          * PREFERRED_TARGET ioctl: we know it will only support
 112          * creating one kind of guest CPU which is its preferred
 113          * CPU type.
 114          */
 115         struct kvm_vcpu_init try;
 116
 117         while (*cpus_to_try != QEMU_KVM_ARM_TARGET_NONE) {
 118             try.target = *cpus_to_try++;
 119             memcpy(try.features, init->features, sizeof(init->features));
 120             ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, &try);
 121             if (ret >= 0) {
 122                 break;
 123             }
 124         }
 125         if (ret < 0) {
 126             goto err;
 127         }
 128         init->target = try.target;
 129     } else {
 130         /* Treat a NULL cpus_to_try argument the same as an empty
 131          * list, which means we will fail the call since this must
 132          * be an old kernel which doesn't support PREFERRED_TARGET.
 133          */
 134         goto err;
 135     }
 136
 137 finish:
 138     fdarray[0] = kvmfd;
 139     fdarray[1] = vmfd;
 140     fdarray[2] = cpufd;
 141
 142     return true;
 143
 144 err:
 145     if (cpufd >= 0) {
 146         close(cpufd);
 147     }
 148     if (vmfd >= 0) {
 149         close(vmfd);
 150     }
 151     if (kvmfd >= 0) {
 152         close(kvmfd);
 153     }
 154
 155     return false;
 156 }
 157
 158 void kvm_arm_destroy_scratch_host_vcpu(int *fdarray)
 159 {
 160     int i;
 161
 162     for (i = 2; i >= 0; i--) {
 163         close(fdarray[i]);
 164     }
 165 }
 166
 167 void kvm_arm_set_cpu_features_from_host(ARMCPU *cpu)
 168 {
 169     CPUARMState *env = &cpu->env;
 170
 171     if (!arm_host_cpu_features.dtb_compatible) {
 172         if (!kvm_enabled() ||
 173             !kvm_arm_get_host_cpu_features(&arm_host_cpu_features)) {
 174             /* We can't report this error yet, so flag that we need to
 175              * in arm_cpu_realizefn().
 176              */
 177             cpu->kvm_target = QEMU_KVM_ARM_TARGET_NONE;
 178             cpu->host_cpu_probe_failed = true;
 179             return;
 180         }
 181     }
 182
 183     cpu->kvm_target = arm_host_cpu_features.target;
 184     cpu->dtb_compatible = arm_host_cpu_features.dtb_compatible;
 185     cpu->isar = arm_host_cpu_features.isar;
 186     env->features = arm_host_cpu_features.features;
 187 }
 188
 189 static bool kvm_no_adjvtime_get(Object *obj, Error **errp)
 190 {
 191     return !ARM_CPU(obj)->kvm_adjvtime;
 192 }
 193
 194 static void kvm_no_adjvtime_set(Object *obj, bool value, Error **errp)
 195 {
 196     ARM_CPU(obj)->kvm_adjvtime = !value;
 197 }
 198
 199 static bool kvm_steal_time_get(Object *obj, Error **errp)
 200 {
 201     return ARM_CPU(obj)->kvm_steal_time != ON_OFF_AUTO_OFF;
 202 }
 203
 204 static void kvm_steal_time_set(Object *obj, bool value, Error **errp)
 205 {
 206     ARM_CPU(obj)->kvm_steal_time = value ? ON_OFF_AUTO_ON : ON_OFF_AUTO_OFF;
 207 }
 208
 209 /* KVM VCPU properties should be prefixed with "kvm-". */
 210 void kvm_arm_add_vcpu_properties(Object *obj)
 211 {
 212     ARMCPU *cpu = ARM_CPU(obj);
 213     CPUARMState *env = &cpu->env;
 214
 215     if (arm_feature(env, ARM_FEATURE_GENERIC_TIMER)) {
 216         cpu->kvm_adjvtime = true;
 217         object_property_add_bool(obj, "kvm-no-adjvtime", kvm_no_adjvtime_get,
 218                                  kvm_no_adjvtime_set);
 219         object_property_set_description(obj, "kvm-no-adjvtime",
 220                                         "Set on to disable the adjustment of "
 221                                         "the virtual counter. VM stopped time "
 222                                         "will be counted.");
 223     }
 224
 225     cpu->kvm_steal_time = ON_OFF_AUTO_AUTO;
 226     object_property_add_bool(obj, "kvm-steal-time", kvm_steal_time_get,
 227                              kvm_steal_time_set);
 228     object_property_set_description(obj, "kvm-steal-time",
 229                                     "Set off to disable KVM steal time.");
 230 }
 231
 232 bool kvm_arm_pmu_supported(void)
 233 {
 234     return kvm_check_extension(kvm_state, KVM_CAP_ARM_PMU_V3);
 235 }
 236
 237 int kvm_arm_get_max_vm_ipa_size(MachineState *ms, bool *fixed_ipa)
 238 {
 239     KVMState *s = KVM_STATE(ms->accelerator);
 240     int ret;
 241
 242     ret = kvm_check_extension(s, KVM_CAP_ARM_VM_IPA_SIZE);
 243     *fixed_ipa = ret <= 0;
 244
 245     return ret > 0 ? ret : 40;
 246 }
 247
 248 int kvm_arch_init(MachineState *ms, KVMState *s)
 249 {
 250     int ret = 0;
 251     /* For ARM interrupt delivery is always asynchronous,
 252      * whether we are using an in-kernel VGIC or not.
 253      */
 254     kvm_async_interrupts_allowed = true;
 255
 256     /*
 257      * PSCI wakes up secondary cores, so we always need to
 258      * have vCPUs waiting in kernel space
 259      */
 260     kvm_halt_in_kernel_allowed = true;
 261
 262     cap_has_mp_state = kvm_check_extension(s, KVM_CAP_MP_STATE);
 263
 264     if (ms->smp.cpus > 256 &&
 265         !kvm_check_extension(s, KVM_CAP_ARM_IRQ_LINE_LAYOUT_2)) {
 266         error_report("Using more than 256 vcpus requires a host kernel "
 267                      "with KVM_CAP_ARM_IRQ_LINE_LAYOUT_2");
 268         ret = -EINVAL;
 269     }
 270
 271     if (kvm_check_extension(s, KVM_CAP_ARM_NISV_TO_USER)) {
 272         if (kvm_vm_enable_cap(s, KVM_CAP_ARM_NISV_TO_USER, 0)) {
 273             error_report("Failed to enable KVM_CAP_ARM_NISV_TO_USER cap");
 274         } else {
 275             /* Set status for supporting the external dabt injection */
 276             cap_has_inject_ext_dabt = kvm_check_extension(s,
 277                                     KVM_CAP_ARM_INJECT_EXT_DABT);
 278         }
 279     }
 280
 281     return ret;
 282 }
 283
 284 unsigned long kvm_arch_vcpu_id(CPUState *cpu)
 285 {
 286     return cpu->cpu_index;
 287 }
 288
 289 /* We track all the KVM devices which need their memory addresses
 290  * passing to the kernel in a list of these structures.
 291  * When board init is complete we run through the list and
 292  * tell the kernel the base addresses of the memory regions.
 293  * We use a MemoryListener to track mapping and unmapping of
 294  * the regions during board creation, so the board models don't
 295  * need to do anything special for the KVM case.
 296  *
 297  * Sometimes the address must be OR'ed with some other fields
 298  * (for example for KVM_VGIC_V3_ADDR_TYPE_REDIST_REGION).
 299  * @kda_addr_ormask aims at storing the value of those fields.
 300  */
 301 typedef struct KVMDevice {
 302     struct kvm_arm_device_addr kda;
 303     struct kvm_device_attr kdattr;
 304     uint64_t kda_addr_ormask;
 305     MemoryRegion *mr;
 306     QSLIST_ENTRY(KVMDevice) entries;
 307     int dev_fd;
 308 } KVMDevice;
 309
 310 static QSLIST_HEAD(, KVMDevice) kvm_devices_head;
 311
 312 static void kvm_arm_devlistener_add(MemoryListener *listener,
 313                                     MemoryRegionSection *section)
 314 {
 315     KVMDevice *kd;
 316
 317     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
 318         if (section->mr == kd->mr) {
 319             kd->kda.addr = section->offset_within_address_space;
 320         }
 321     }
 322 }
 323
 324 static void kvm_arm_devlistener_del(MemoryListener *listener,
 325                                     MemoryRegionSection *section)
 326 {
 327     KVMDevice *kd;
 328
 329     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
 330         if (section->mr == kd->mr) {
 331             kd->kda.addr = -1;
 332         }
 333     }
 334 }
 335
 336 static MemoryListener devlistener = {
 337     .name = "kvm-arm",
 338     .region_add = kvm_arm_devlistener_add,
 339     .region_del = kvm_arm_devlistener_del,
 340 };
 341
 342 static void kvm_arm_set_device_addr(KVMDevice *kd)
 343 {
 344     struct kvm_device_attr *attr = &kd->kdattr;
 345     int ret;
 346
 347     /* If the device control API is available and we have a device fd on the
 348      * KVMDevice struct, let's use the newer API
 349      */
 350     if (kd->dev_fd >= 0) {
 351         uint64_t addr = kd->kda.addr;
 352
 353         addr |= kd->kda_addr_ormask;
 354         attr->addr = (uintptr_t)&addr;
 355         ret = kvm_device_ioctl(kd->dev_fd, KVM_SET_DEVICE_ATTR, attr);
 356     } else {
 357         ret = kvm_vm_ioctl(kvm_state, KVM_ARM_SET_DEVICE_ADDR, &kd->kda);
 358     }
 359
 360     if (ret < 0) {
 361         fprintf(stderr, "Failed to set device address: %s\n",
 362                 strerror(-ret));
 363         abort();
 364     }
 365 }
 366
 367 static void kvm_arm_machine_init_done(Notifier *notifier, void *data)
 368 {
 369     KVMDevice *kd, *tkd;
 370
 371     QSLIST_FOREACH_SAFE(kd, &kvm_devices_head, entries, tkd) {
 372         if (kd->kda.addr != -1) {
 373             kvm_arm_set_device_addr(kd);
 374         }
 375         memory_region_unref(kd->mr);
 376         QSLIST_REMOVE_HEAD(&kvm_devices_head, entries);
 377         g_free(kd);
 378     }
 379     memory_listener_unregister(&devlistener);
 380 }
 381
 382 static Notifier notify = {
 383     .notify = kvm_arm_machine_init_done,
 384 };
 385
 386 void kvm_arm_register_device(MemoryRegion *mr, uint64_t devid, uint64_t group,
 387                              uint64_t attr, int dev_fd, uint64_t addr_ormask)
 388 {
 389     KVMDevice *kd;
 390
 391     if (!kvm_irqchip_in_kernel()) {
 392         return;
 393     }
 394
 395     if (QSLIST_EMPTY(&kvm_devices_head)) {
 396         memory_listener_register(&devlistener, &address_space_memory);
 397         qemu_add_machine_init_done_notifier(&notify);
 398     }
 399     kd = g_new0(KVMDevice, 1);
 400     kd->mr = mr;
 401     kd->kda.id = devid;
 402     kd->kda.addr = -1;
 403     kd->kdattr.flags = 0;
 404     kd->kdattr.group = group;
 405     kd->kdattr.attr = attr;
 406     kd->dev_fd = dev_fd;
 407     kd->kda_addr_ormask = addr_ormask;
 408     QSLIST_INSERT_HEAD(&kvm_devices_head, kd, entries);
 409     memory_region_ref(kd->mr);
 410 }
 411
 412 static int compare_u64(const void *a, const void *b)
 413 {
 414     if (*(uint64_t *)a > *(uint64_t *)b) {
 415         return 1;
 416     }
 417     if (*(uint64_t *)a < *(uint64_t *)b) {
 418         return -1;
 419     }
 420     return 0;
 421 }
 422
 423 /*
 424  * cpreg_values are sorted in ascending order by KVM register ID
 425  * (see kvm_arm_init_cpreg_list). This allows us to cheaply find
 426  * the storage for a KVM register by ID with a binary search.
 427  */
 428 static uint64_t *kvm_arm_get_cpreg_ptr(ARMCPU *cpu, uint64_t regidx)
 429 {
 430     uint64_t *res;
 431
 432     res = bsearch(&regidx, cpu->cpreg_indexes, cpu->cpreg_array_len,
 433                   sizeof(uint64_t), compare_u64);
 434     assert(res);
 435
 436     return &cpu->cpreg_values[res - cpu->cpreg_indexes];
 437 }
 438
 439 /* Initialize the ARMCPU cpreg list according to the kernel's
 440  * definition of what CPU registers it knows about (and throw away
 441  * the previous TCG-created cpreg list).
 442  */
 443 int kvm_arm_init_cpreg_list(ARMCPU *cpu)
 444 {
 445     struct kvm_reg_list rl;
 446     struct kvm_reg_list *rlp;
 447     int i, ret, arraylen;
 448     CPUState *cs = CPU(cpu);
 449
 450     rl.n = 0;
 451     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, &rl);
 452     if (ret != -E2BIG) {
 453         return ret;
 454     }
 455     rlp = g_malloc(sizeof(struct kvm_reg_list) + rl.n * sizeof(uint64_t));
 456     rlp->n = rl.n;
 457     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, rlp);
 458     if (ret) {
 459         goto out;
 460     }
 461     /* Sort the list we get back from the kernel, since cpreg_tuples
 462      * must be in strictly ascending order.
 463      */
 464     qsort(&rlp->reg, rlp->n, sizeof(rlp->reg[0]), compare_u64);
 465
 466     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 467         if (!kvm_arm_reg_syncs_via_cpreg_list(rlp->reg[i])) {
 468             continue;
 469         }
 470         switch (rlp->reg[i] & KVM_REG_SIZE_MASK) {
 471         case KVM_REG_SIZE_U32:
 472         case KVM_REG_SIZE_U64:
 473             break;
 474         default:
 475             fprintf(stderr, "Can't handle size of register in kernel list\n");
 476             ret = -EINVAL;
 477             goto out;
 478         }
 479
 480         arraylen++;
 481     }
 482
 483     cpu->cpreg_indexes = g_renew(uint64_t, cpu->cpreg_indexes, arraylen);
 484     cpu->cpreg_values = g_renew(uint64_t, cpu->cpreg_values, arraylen);
 485     cpu->cpreg_vmstate_indexes = g_renew(uint64_t, cpu->cpreg_vmstate_indexes,
 486                                          arraylen);
 487     cpu->cpreg_vmstate_values = g_renew(uint64_t, cpu->cpreg_vmstate_values,
 488                                         arraylen);
 489     cpu->cpreg_array_len = arraylen;
 490     cpu->cpreg_vmstate_array_len = arraylen;
 491
 492     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 493         uint64_t regidx = rlp->reg[i];
 494         if (!kvm_arm_reg_syncs_via_cpreg_list(regidx)) {
 495             continue;
 496         }
 497         cpu->cpreg_indexes[arraylen] = regidx;
 498         arraylen++;
 499     }
 500     assert(cpu->cpreg_array_len == arraylen);
 501
 502     if (!write_kvmstate_to_list(cpu)) {
 503         /* Shouldn't happen unless kernel is inconsistent about
 504          * what registers exist.
 505          */
 506         fprintf(stderr, "Initial read of kernel register state failed\n");
 507         ret = -EINVAL;
 508         goto out;
 509     }
 510
 511 out:
 512     g_free(rlp);
 513     return ret;
 514 }
 515
 516 bool write_kvmstate_to_list(ARMCPU *cpu)
 517 {
 518     CPUState *cs = CPU(cpu);
 519     int i;
 520     bool ok = true;
 521
 522     for (i = 0; i < cpu->cpreg_array_len; i++) {
 523         struct kvm_one_reg r;
 524         uint64_t regidx = cpu->cpreg_indexes[i];
 525         uint32_t v32;
 526         int ret;
 527
 528         r.id = regidx;
 529
 530         switch (regidx & KVM_REG_SIZE_MASK) {
 531         case KVM_REG_SIZE_U32:
 532             r.addr = (uintptr_t)&v32;
 533             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 534             if (!ret) {
 535                 cpu->cpreg_values[i] = v32;
 536             }
 537             break;
 538         case KVM_REG_SIZE_U64:
 539             r.addr = (uintptr_t)(cpu->cpreg_values + i);
 540             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 541             break;
 542         default:
 543             g_assert_not_reached();
 544         }
 545         if (ret) {
 546             ok = false;
 547         }
 548     }
 549     return ok;
 550 }
 551
 552 bool write_list_to_kvmstate(ARMCPU *cpu, int level)
 553 {
 554     CPUState *cs = CPU(cpu);
 555     int i;
 556     bool ok = true;
 557
 558     for (i = 0; i < cpu->cpreg_array_len; i++) {
 559         struct kvm_one_reg r;
 560         uint64_t regidx = cpu->cpreg_indexes[i];
 561         uint32_t v32;
 562         int ret;
 563
 564         if (kvm_arm_cpreg_level(regidx) > level) {
 565             continue;
 566         }
 567
 568         r.id = regidx;
 569         switch (regidx & KVM_REG_SIZE_MASK) {
 570         case KVM_REG_SIZE_U32:
 571             v32 = cpu->cpreg_values[i];
 572             r.addr = (uintptr_t)&v32;
 573             break;
 574         case KVM_REG_SIZE_U64:
 575             r.addr = (uintptr_t)(cpu->cpreg_values + i);
 576             break;
 577         default:
 578             g_assert_not_reached();
 579         }
 580         ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
 581         if (ret) {
 582             /* We might fail for "unknown register" and also for
 583              * "you tried to set a register which is constant with
 584              * a different value from what it actually contains".
 585              */
 586             ok = false;
 587         }
 588     }
 589     return ok;
 590 }
 591
 592 void kvm_arm_cpu_pre_save(ARMCPU *cpu)
 593 {
 594     /* KVM virtual time adjustment */
 595     if (cpu->kvm_vtime_dirty) {
 596         *kvm_arm_get_cpreg_ptr(cpu, KVM_REG_ARM_TIMER_CNT) = cpu->kvm_vtime;
 597     }
 598 }
 599
 600 void kvm_arm_cpu_post_load(ARMCPU *cpu)
 601 {
 602     /* KVM virtual time adjustment */
 603     if (cpu->kvm_adjvtime) {
 604         cpu->kvm_vtime = *kvm_arm_get_cpreg_ptr(cpu, KVM_REG_ARM_TIMER_CNT);
 605         cpu->kvm_vtime_dirty = true;
 606     }
 607 }
 608
 609 void kvm_arm_reset_vcpu(ARMCPU *cpu)
 610 {
 611     int ret;
 612
 613     /* Re-init VCPU so that all registers are set to
 614      * their respective reset values.
 615      */
 616     ret = kvm_arm_vcpu_init(CPU(cpu));
 617     if (ret < 0) {
 618         fprintf(stderr, "kvm_arm_vcpu_init failed: %s\n", strerror(-ret));
 619         abort();
 620     }
 621     if (!write_kvmstate_to_list(cpu)) {
 622         fprintf(stderr, "write_kvmstate_to_list failed\n");
 623         abort();
 624     }
 625     /*
 626      * Sync the reset values also into the CPUState. This is necessary
 627      * because the next thing we do will be a kvm_arch_put_registers()
 628      * which will update the list values from the CPUState before copying
 629      * the list values back to KVM. It's OK to ignore failure returns here
 630      * for the same reason we do so in kvm_arch_get_registers().
 631      */
 632     write_list_to_cpustate(cpu);
 633 }
 634
 635 /*
 636  * Update KVM's MP_STATE based on what QEMU thinks it is
 637  */
 638 int kvm_arm_sync_mpstate_to_kvm(ARMCPU *cpu)
 639 {
 640     if (cap_has_mp_state) {
 641         struct kvm_mp_state mp_state = {
 642             .mp_state = (cpu->power_state == PSCI_OFF) ?
 643             KVM_MP_STATE_STOPPED : KVM_MP_STATE_RUNNABLE
 644         };
 645         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_SET_MP_STATE, &mp_state);
 646         if (ret) {
 647             fprintf(stderr, "%s: failed to set MP_STATE %d/%s\n",
 648                     __func__, ret, strerror(-ret));
 649             return -1;
 650         }
 651     }
 652
 653     return 0;
 654 }
 655
 656 /*
 657  * Sync the KVM MP_STATE into QEMU
 658  */
 659 int kvm_arm_sync_mpstate_to_qemu(ARMCPU *cpu)
 660 {
 661     if (cap_has_mp_state) {
 662         struct kvm_mp_state mp_state;
 663         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_GET_MP_STATE, &mp_state);
 664         if (ret) {
 665             fprintf(stderr, "%s: failed to get MP_STATE %d/%s\n",
 666                     __func__, ret, strerror(-ret));
 667             abort();
 668         }
 669         cpu->power_state = (mp_state.mp_state == KVM_MP_STATE_STOPPED) ?
 670             PSCI_OFF : PSCI_ON;
 671     }
 672
 673     return 0;
 674 }
 675
 676 void kvm_arm_get_virtual_time(CPUState *cs)
 677 {
 678     ARMCPU *cpu = ARM_CPU(cs);
 679     struct kvm_one_reg reg = {
 680         .id = KVM_REG_ARM_TIMER_CNT,
 681         .addr = (uintptr_t)&cpu->kvm_vtime,
 682     };
 683     int ret;
 684
 685     if (cpu->kvm_vtime_dirty) {
 686         return;
 687     }
 688
 689     ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
 690     if (ret) {
 691         error_report("Failed to get KVM_REG_ARM_TIMER_CNT");
 692         abort();
 693     }
 694
 695     cpu->kvm_vtime_dirty = true;
 696 }
 697
 698 void kvm_arm_put_virtual_time(CPUState *cs)
 699 {
 700     ARMCPU *cpu = ARM_CPU(cs);
 701     struct kvm_one_reg reg = {
 702         .id = KVM_REG_ARM_TIMER_CNT,
 703         .addr = (uintptr_t)&cpu->kvm_vtime,
 704     };
 705     int ret;
 706
 707     if (!cpu->kvm_vtime_dirty) {
 708         return;
 709     }
 710
 711     ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
 712     if (ret) {
 713         error_report("Failed to set KVM_REG_ARM_TIMER_CNT");
 714         abort();
 715     }
 716
 717     cpu->kvm_vtime_dirty = false;
 718 }
 719
 720 int kvm_put_vcpu_events(ARMCPU *cpu)
 721 {
 722     CPUARMState *env = &cpu->env;
 723     struct kvm_vcpu_events events;
 724     int ret;
 725
 726     if (!kvm_has_vcpu_events()) {
 727         return 0;
 728     }
 729
 730     memset(&events, 0, sizeof(events));
 731     events.exception.serror_pending = env->serror.pending;
 732
 733     /* Inject SError to guest with specified syndrome if host kernel
 734      * supports it, otherwise inject SError without syndrome.
 735      */
 736     if (cap_has_inject_serror_esr) {
 737         events.exception.serror_has_esr = env->serror.has_esr;
 738         events.exception.serror_esr = env->serror.esr;
 739     }
 740
 741     ret = kvm_vcpu_ioctl(CPU(cpu), KVM_SET_VCPU_EVENTS, &events);
 742     if (ret) {
 743         error_report("failed to put vcpu events");
 744     }
 745
 746     return ret;
 747 }
 748
 749 int kvm_get_vcpu_events(ARMCPU *cpu)
 750 {
 751     CPUARMState *env = &cpu->env;
 752     struct kvm_vcpu_events events;
 753     int ret;
 754
 755     if (!kvm_has_vcpu_events()) {
 756         return 0;
 757     }
 758
 759     memset(&events, 0, sizeof(events));
 760     ret = kvm_vcpu_ioctl(CPU(cpu), KVM_GET_VCPU_EVENTS, &events);
 761     if (ret) {
 762         error_report("failed to get vcpu events");
 763         return ret;
 764     }
 765
 766     env->serror.pending = events.exception.serror_pending;
 767     env->serror.has_esr = events.exception.serror_has_esr;
 768     env->serror.esr = events.exception.serror_esr;
 769
 770     return 0;
 771 }
 772
 773 void kvm_arch_pre_run(CPUState *cs, struct kvm_run *run)
 774 {
 775     ARMCPU *cpu = ARM_CPU(cs);
 776     CPUARMState *env = &cpu->env;
 777
 778     if (unlikely(env->ext_dabt_raised)) {
 779         /*
 780          * Verifying that the ext DABT has been properly injected,
 781          * otherwise risking indefinitely re-running the faulting instruction
 782          * Covering a very narrow case for kernels 5.5..5.5.4
 783          * when injected abort was misconfigured to be
 784          * an IMPLEMENTATION DEFINED exception (for 32-bit EL1)
 785          */
 786         if (!arm_feature(env, ARM_FEATURE_AARCH64) &&
 787             unlikely(!kvm_arm_verify_ext_dabt_pending(cs))) {
 788
 789             error_report("Data abort exception with no valid ISS generated by "
 790                    "guest memory access. KVM unable to emulate faulting "
 791                    "instruction. Failed to inject an external data abort "
 792                    "into the guest.");
 793             abort();
 794        }
 795        /* Clear the status */
 796        env->ext_dabt_raised = 0;
 797     }
 798 }
 799
 800 MemTxAttrs kvm_arch_post_run(CPUState *cs, struct kvm_run *run)
 801 {
 802     ARMCPU *cpu;
 803     uint32_t switched_level;
 804
 805     if (kvm_irqchip_in_kernel()) {
 806         /*
 807          * We only need to sync timer states with user-space interrupt
 808          * controllers, so return early and save cycles if we don't.
 809          */
 810         return MEMTXATTRS_UNSPECIFIED;
 811     }
 812
 813     cpu = ARM_CPU(cs);
 814
 815     /* Synchronize our shadowed in-kernel device irq lines with the kvm ones */
 816     if (run->s.regs.device_irq_level != cpu->device_irq_level) {
 817         switched_level = cpu->device_irq_level ^ run->s.regs.device_irq_level;
 818
 819         qemu_mutex_lock_iothread();
 820
 821         if (switched_level & KVM_ARM_DEV_EL1_VTIMER) {
 822             qemu_set_irq(cpu->gt_timer_outputs[GTIMER_VIRT],
 823                          !!(run->s.regs.device_irq_level &
 824                             KVM_ARM_DEV_EL1_VTIMER));
 825             switched_level &= ~KVM_ARM_DEV_EL1_VTIMER;
 826         }
 827
 828         if (switched_level & KVM_ARM_DEV_EL1_PTIMER) {
 829             qemu_set_irq(cpu->gt_timer_outputs[GTIMER_PHYS],
 830                          !!(run->s.regs.device_irq_level &
 831                             KVM_ARM_DEV_EL1_PTIMER));
 832             switched_level &= ~KVM_ARM_DEV_EL1_PTIMER;
 833         }
 834
 835         if (switched_level & KVM_ARM_DEV_PMU) {
 836             qemu_set_irq(cpu->pmu_interrupt,
 837                          !!(run->s.regs.device_irq_level & KVM_ARM_DEV_PMU));
 838             switched_level &= ~KVM_ARM_DEV_PMU;
 839         }
 840
 841         if (switched_level) {
 842             qemu_log_mask(LOG_UNIMP, "%s: unhandled in-kernel device IRQ %x\n",
 843                           __func__, switched_level);
 844         }
 845
 846         /* We also mark unknown levels as processed to not waste cycles */
 847         cpu->device_irq_level = run->s.regs.device_irq_level;
 848         qemu_mutex_unlock_iothread();
 849     }
 850
 851     return MEMTXATTRS_UNSPECIFIED;
 852 }
 853
 854 void kvm_arm_vm_state_change(void *opaque, bool running, RunState state)
 855 {
 856     CPUState *cs = opaque;
 857     ARMCPU *cpu = ARM_CPU(cs);
 858
 859     if (running) {
 860         if (cpu->kvm_adjvtime) {
 861             kvm_arm_put_virtual_time(cs);
 862         }
 863     } else {
 864         if (cpu->kvm_adjvtime) {
 865             kvm_arm_get_virtual_time(cs);
 866         }
 867     }
 868 }
 869
 870 /**
 871  * kvm_arm_handle_dabt_nisv:
 872  * @cs: CPUState
 873  * @esr_iss: ISS encoding (limited) for the exception from Data Abort
 874  *           ISV bit set to '0b0' -> no valid instruction syndrome
 875  * @fault_ipa: faulting address for the synchronous data abort
 876  *
 877  * Returns: 0 if the exception has been handled, < 0 otherwise
 878  */
 879 static int kvm_arm_handle_dabt_nisv(CPUState *cs, uint64_t esr_iss,
 880                                     uint64_t fault_ipa)
 881 {
 882     ARMCPU *cpu = ARM_CPU(cs);
 883     CPUARMState *env = &cpu->env;
 884     /*
 885      * Request KVM to inject the external data abort into the guest
 886      */
 887     if (cap_has_inject_ext_dabt) {
 888         struct kvm_vcpu_events events = { };
 889         /*
 890          * The external data abort event will be handled immediately by KVM
 891          * using the address fault that triggered the exit on given VCPU.
 892          * Requesting injection of the external data abort does not rely
 893          * on any other VCPU state. Therefore, in this particular case, the VCPU
 894          * synchronization can be exceptionally skipped.
 895          */
 896         events.exception.ext_dabt_pending = 1;
 897         /* KVM_CAP_ARM_INJECT_EXT_DABT implies KVM_CAP_VCPU_EVENTS */
 898         if (!kvm_vcpu_ioctl(cs, KVM_SET_VCPU_EVENTS, &events)) {
 899             env->ext_dabt_raised = 1;
 900             return 0;
 901         }
 902     } else {
 903         error_report("Data abort exception triggered by guest memory access "
 904                      "at physical address: 0x"  TARGET_FMT_lx,
 905                      (target_ulong)fault_ipa);
 906         error_printf("KVM unable to emulate faulting instruction.\n");
 907     }
 908     return -1;
 909 }
 910
 911 int kvm_arch_handle_exit(CPUState *cs, struct kvm_run *run)
 912 {
 913     int ret = 0;
 914
 915     switch (run->exit_reason) {
 916     case KVM_EXIT_DEBUG:
 917         if (kvm_arm_handle_debug(cs, &run->debug.arch)) {
 918             ret = EXCP_DEBUG;
 919         } /* otherwise return to guest */
 920         break;
 921     case KVM_EXIT_ARM_NISV:
 922         /* External DABT with no valid iss to decode */
 923         ret = kvm_arm_handle_dabt_nisv(cs, run->arm_nisv.esr_iss,
 924                                        run->arm_nisv.fault_ipa);
 925         break;
 926     default:
 927         qemu_log_mask(LOG_UNIMP, "%s: un-handled exit reason %d\n",
 928                       __func__, run->exit_reason);
 929         break;
 930     }
 931     return ret;
 932 }
 933
 934 bool kvm_arch_stop_on_emulation_error(CPUState *cs)
 935 {
 936     return true;
 937 }
 938
 939 int kvm_arch_process_async_events(CPUState *cs)
 940 {
 941     return 0;
 942 }
 943
 944 void kvm_arch_update_guest_debug(CPUState *cs, struct kvm_guest_debug *dbg)
 945 {
 946     if (kvm_sw_breakpoints_active(cs)) {
 947         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_SW_BP;
 948     }
 949     if (kvm_arm_hw_debug_active(cs)) {
 950         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_HW;
 951         kvm_arm_copy_hw_debug_data(&dbg->arch);
 952     }
 953 }
 954
 955 void kvm_arch_init_irq_routing(KVMState *s)
 956 {
 957 }
 958
 959 int kvm_arch_irqchip_create(KVMState *s)
 960 {
 961     if (kvm_kernel_irqchip_split()) {
 962         error_report("-machine kernel_irqchip=split is not supported on ARM.");
 963         exit(1);
 964     }
 965
 966     /* If we can create the VGIC using the newer device control API, we
 967      * let the device do this when it initializes itself, otherwise we
 968      * fall back to the old API */
 969     return kvm_check_extension(s, KVM_CAP_DEVICE_CTRL);
 970 }
 971
 972 int kvm_arm_vgic_probe(void)
 973 {
 974     int val = 0;
 975
 976     if (kvm_create_device(kvm_state,
 977                           KVM_DEV_TYPE_ARM_VGIC_V3, true) == 0) {
 978         val |= KVM_ARM_VGIC_V3;
 979     }
 980     if (kvm_create_device(kvm_state,
 981                           KVM_DEV_TYPE_ARM_VGIC_V2, true) == 0) {
 982         val |= KVM_ARM_VGIC_V2;
 983     }
 984     return val;
 985 }
 986
 987 int kvm_arm_set_irq(int cpu, int irqtype, int irq, int level)
 988 {
 989     int kvm_irq = (irqtype << KVM_ARM_IRQ_TYPE_SHIFT) | irq;
 990     int cpu_idx1 = cpu % 256;
 991     int cpu_idx2 = cpu / 256;
 992
 993     kvm_irq |= (cpu_idx1 << KVM_ARM_IRQ_VCPU_SHIFT) |
 994                (cpu_idx2 << KVM_ARM_IRQ_VCPU2_SHIFT);
 995
 996     return kvm_set_irq(kvm_state, kvm_irq, !!level);
 997 }
 998
 999 int kvm_arch_fixup_msi_route(struct kvm_irq_routing_entry *route,
1000                              uint64_t address, uint32_t data, PCIDevice *dev)
1001 {
1002     AddressSpace *as = pci_device_iommu_address_space(dev);
1003     hwaddr xlat, len, doorbell_gpa;
1004     MemoryRegionSection mrs;
1005     MemoryRegion *mr;
1006
1007     if (as == &address_space_memory) {
1008         return 0;
1009     }
1010
1011     /* MSI doorbell address is translated by an IOMMU */
1012
1013     RCU_READ_LOCK_GUARD();
1014
1015     mr = address_space_translate(as, address, &xlat, &len, true,
1016                                  MEMTXATTRS_UNSPECIFIED);
1017
1018     if (!mr) {
1019         return 1;
1020     }
1021
1022     mrs = memory_region_find(mr, xlat, 1);
1023
1024     if (!mrs.mr) {
1025         return 1;
1026     }
1027
1028     doorbell_gpa = mrs.offset_within_address_space;
1029     memory_region_unref(mrs.mr);
1030
1031     route->u.msi.address_lo = doorbell_gpa;
1032     route->u.msi.address_hi = doorbell_gpa >> 32;
1033
1034     trace_kvm_arm_fixup_msi_route(address, doorbell_gpa);
1035
1036     return 0;
1037 }
1038
1039 int kvm_arch_add_msi_route_post(struct kvm_irq_routing_entry *route,
1040                                 int vector, PCIDevice *dev)
1041 {
1042     return 0;
1043 }
1044
1045 int kvm_arch_release_virq_post(int virq)
1046 {
1047     return 0;
1048 }
1049
1050 int kvm_arch_msi_data_to_gsi(uint32_t data)
1051 {
1052     return (data - 32) & 0xffff;
1053 }
1054
1055 bool kvm_arch_cpu_check_are_resettable(void)
1056 {
1057     return true;
1058 }