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 "qom/object.h"
  21 #include "qapi/error.h"
  22 #include "sysemu/sysemu.h"
  23 #include "sysemu/kvm.h"
  24 #include "sysemu/kvm_int.h"
  25 #include "kvm_arm.h"
  26 #include "cpu.h"
  27 #include "trace.h"
  28 #include "internals.h"
  29 #include "hw/pci/pci.h"
  30 #include "exec/memattrs.h"
  31 #include "exec/address-spaces.h"
  32 #include "hw/boards.h"
  33 #include "hw/irq.h"
  34 #include "qemu/log.h"
  35
  36 const KVMCapabilityInfo kvm_arch_required_capabilities[] = {
  37     KVM_CAP_LAST_INFO
  38 };
  39
  40 static bool cap_has_mp_state;
  41 static bool cap_has_inject_serror_esr;
  42 static bool cap_has_inject_ext_dabt;
  43
  44 static ARMHostCPUFeatures arm_host_cpu_features;
  45
  46 int kvm_arm_vcpu_init(CPUState *cs)
  47 {
  48     ARMCPU *cpu = ARM_CPU(cs);
  49     struct kvm_vcpu_init init;
  50
  51     init.target = cpu->kvm_target;
  52     memcpy(init.features, cpu->kvm_init_features, sizeof(init.features));
  53
  54     return kvm_vcpu_ioctl(cs, KVM_ARM_VCPU_INIT, &init);
  55 }
  56
  57 int kvm_arm_vcpu_finalize(CPUState *cs, int feature)
  58 {
  59     return kvm_vcpu_ioctl(cs, KVM_ARM_VCPU_FINALIZE, &feature);
  60 }
  61
  62 void kvm_arm_init_serror_injection(CPUState *cs)
  63 {
  64     cap_has_inject_serror_esr = kvm_check_extension(cs->kvm_state,
  65                                     KVM_CAP_ARM_INJECT_SERROR_ESR);
  66 }
  67
  68 bool kvm_arm_create_scratch_host_vcpu(const uint32_t *cpus_to_try,
  69                                       int *fdarray,
  70                                       struct kvm_vcpu_init *init)
  71 {
  72     int ret = 0, kvmfd = -1, vmfd = -1, cpufd = -1;
  73     int max_vm_pa_size;
  74
  75     kvmfd = qemu_open_old("/dev/kvm", O_RDWR);
  76     if (kvmfd < 0) {
  77         goto err;
  78     }
  79     max_vm_pa_size = ioctl(kvmfd, KVM_CHECK_EXTENSION, KVM_CAP_ARM_VM_IPA_SIZE);
  80     if (max_vm_pa_size < 0) {
  81         max_vm_pa_size = 0;
  82     }
  83     vmfd = ioctl(kvmfd, KVM_CREATE_VM, max_vm_pa_size);
  84     if (vmfd < 0) {
  85         goto err;
  86     }
  87     cpufd = ioctl(vmfd, KVM_CREATE_VCPU, 0);
  88     if (cpufd < 0) {
  89         goto err;
  90     }
  91
  92     if (!init) {
  93         /* Caller doesn't want the VCPU to be initialized, so skip it */
  94         goto finish;
  95     }
  96
  97     if (init->target == -1) {
  98         struct kvm_vcpu_init preferred;
  99
 100         ret = ioctl(vmfd, KVM_ARM_PREFERRED_TARGET, &preferred);
 101         if (!ret) {
 102             init->target = preferred.target;
 103         }
 104     }
 105     if (ret >= 0) {
 106         ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, init);
 107         if (ret < 0) {
 108             goto err;
 109         }
 110     } else if (cpus_to_try) {
 111         /* Old kernel which doesn't know about the
 112          * PREFERRED_TARGET ioctl: we know it will only support
 113          * creating one kind of guest CPU which is its preferred
 114          * CPU type.
 115          */
 116         struct kvm_vcpu_init try;
 117
 118         while (*cpus_to_try != QEMU_KVM_ARM_TARGET_NONE) {
 119             try.target = *cpus_to_try++;
 120             memcpy(try.features, init->features, sizeof(init->features));
 121             ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, &try);
 122             if (ret >= 0) {
 123                 break;
 124             }
 125         }
 126         if (ret < 0) {
 127             goto err;
 128         }
 129         init->target = try.target;
 130     } else {
 131         /* Treat a NULL cpus_to_try argument the same as an empty
 132          * list, which means we will fail the call since this must
 133          * be an old kernel which doesn't support PREFERRED_TARGET.
 134          */
 135         goto err;
 136     }
 137
 138 finish:
 139     fdarray[0] = kvmfd;
 140     fdarray[1] = vmfd;
 141     fdarray[2] = cpufd;
 142
 143     return true;
 144
 145 err:
 146     if (cpufd >= 0) {
 147         close(cpufd);
 148     }
 149     if (vmfd >= 0) {
 150         close(vmfd);
 151     }
 152     if (kvmfd >= 0) {
 153         close(kvmfd);
 154     }
 155
 156     return false;
 157 }
 158
 159 void kvm_arm_destroy_scratch_host_vcpu(int *fdarray)
 160 {
 161     int i;
 162
 163     for (i = 2; i >= 0; i--) {
 164         close(fdarray[i]);
 165     }
 166 }
 167
 168 void kvm_arm_set_cpu_features_from_host(ARMCPU *cpu)
 169 {
 170     CPUARMState *env = &cpu->env;
 171
 172     if (!arm_host_cpu_features.dtb_compatible) {
 173         if (!kvm_enabled() ||
 174             !kvm_arm_get_host_cpu_features(&arm_host_cpu_features)) {
 175             /* We can't report this error yet, so flag that we need to
 176              * in arm_cpu_realizefn().
 177              */
 178             cpu->kvm_target = QEMU_KVM_ARM_TARGET_NONE;
 179             cpu->host_cpu_probe_failed = true;
 180             return;
 181         }
 182     }
 183
 184     cpu->kvm_target = arm_host_cpu_features.target;
 185     cpu->dtb_compatible = arm_host_cpu_features.dtb_compatible;
 186     cpu->isar = arm_host_cpu_features.isar;
 187     env->features = arm_host_cpu_features.features;
 188 }
 189
 190 static bool kvm_no_adjvtime_get(Object *obj, Error **errp)
 191 {
 192     return !ARM_CPU(obj)->kvm_adjvtime;
 193 }
 194
 195 static void kvm_no_adjvtime_set(Object *obj, bool value, Error **errp)
 196 {
 197     ARM_CPU(obj)->kvm_adjvtime = !value;
 198 }
 199
 200 static bool kvm_steal_time_get(Object *obj, Error **errp)
 201 {
 202     return ARM_CPU(obj)->kvm_steal_time != ON_OFF_AUTO_OFF;
 203 }
 204
 205 static void kvm_steal_time_set(Object *obj, bool value, Error **errp)
 206 {
 207     ARM_CPU(obj)->kvm_steal_time = value ? ON_OFF_AUTO_ON : ON_OFF_AUTO_OFF;
 208 }
 209
 210 /* KVM VCPU properties should be prefixed with "kvm-". */
 211 void kvm_arm_add_vcpu_properties(Object *obj)
 212 {
 213     ARMCPU *cpu = ARM_CPU(obj);
 214     CPUARMState *env = &cpu->env;
 215
 216     if (arm_feature(env, ARM_FEATURE_GENERIC_TIMER)) {
 217         cpu->kvm_adjvtime = true;
 218         object_property_add_bool(obj, "kvm-no-adjvtime", kvm_no_adjvtime_get,
 219                                  kvm_no_adjvtime_set);
 220         object_property_set_description(obj, "kvm-no-adjvtime",
 221                                         "Set on to disable the adjustment of "
 222                                         "the virtual counter. VM stopped time "
 223                                         "will be counted.");
 224     }
 225
 226     cpu->kvm_steal_time = ON_OFF_AUTO_AUTO;
 227     object_property_add_bool(obj, "kvm-steal-time", kvm_steal_time_get,
 228                              kvm_steal_time_set);
 229     object_property_set_description(obj, "kvm-steal-time",
 230                                     "Set off to disable KVM steal time.");
 231 }
 232
 233 bool kvm_arm_pmu_supported(void)
 234 {
 235     return kvm_check_extension(kvm_state, KVM_CAP_ARM_PMU_V3);
 236 }
 237
 238 int kvm_arm_get_max_vm_ipa_size(MachineState *ms, bool *fixed_ipa)
 239 {
 240     KVMState *s = KVM_STATE(ms->accelerator);
 241     int ret;
 242
 243     ret = kvm_check_extension(s, KVM_CAP_ARM_VM_IPA_SIZE);
 244     *fixed_ipa = ret <= 0;
 245
 246     return ret > 0 ? ret : 40;
 247 }
 248
 249 int kvm_arch_init(MachineState *ms, KVMState *s)
 250 {
 251     int ret = 0;
 252     /* For ARM interrupt delivery is always asynchronous,
 253      * whether we are using an in-kernel VGIC or not.
 254      */
 255     kvm_async_interrupts_allowed = true;
 256
 257     /*
 258      * PSCI wakes up secondary cores, so we always need to
 259      * have vCPUs waiting in kernel space
 260      */
 261     kvm_halt_in_kernel_allowed = true;
 262
 263     cap_has_mp_state = kvm_check_extension(s, KVM_CAP_MP_STATE);
 264
 265     if (ms->smp.cpus > 256 &&
 266         !kvm_check_extension(s, KVM_CAP_ARM_IRQ_LINE_LAYOUT_2)) {
 267         error_report("Using more than 256 vcpus requires a host kernel "
 268                      "with KVM_CAP_ARM_IRQ_LINE_LAYOUT_2");
 269         ret = -EINVAL;
 270     }
 271
 272     if (kvm_check_extension(s, KVM_CAP_ARM_NISV_TO_USER)) {
 273         if (kvm_vm_enable_cap(s, KVM_CAP_ARM_NISV_TO_USER, 0)) {
 274             error_report("Failed to enable KVM_CAP_ARM_NISV_TO_USER cap");
 275         } else {
 276             /* Set status for supporting the external dabt injection */
 277             cap_has_inject_ext_dabt = kvm_check_extension(s,
 278                                     KVM_CAP_ARM_INJECT_EXT_DABT);
 279         }
 280     }
 281
 282     return ret;
 283 }
 284
 285 unsigned long kvm_arch_vcpu_id(CPUState *cpu)
 286 {
 287     return cpu->cpu_index;
 288 }
 289
 290 /* We track all the KVM devices which need their memory addresses
 291  * passing to the kernel in a list of these structures.
 292  * When board init is complete we run through the list and
 293  * tell the kernel the base addresses of the memory regions.
 294  * We use a MemoryListener to track mapping and unmapping of
 295  * the regions during board creation, so the board models don't
 296  * need to do anything special for the KVM case.
 297  *
 298  * Sometimes the address must be OR'ed with some other fields
 299  * (for example for KVM_VGIC_V3_ADDR_TYPE_REDIST_REGION).
 300  * @kda_addr_ormask aims at storing the value of those fields.
 301  */
 302 typedef struct KVMDevice {
 303     struct kvm_arm_device_addr kda;
 304     struct kvm_device_attr kdattr;
 305     uint64_t kda_addr_ormask;
 306     MemoryRegion *mr;
 307     QSLIST_ENTRY(KVMDevice) entries;
 308     int dev_fd;
 309 } KVMDevice;
 310
 311 static QSLIST_HEAD(, KVMDevice) kvm_devices_head;
 312
 313 static void kvm_arm_devlistener_add(MemoryListener *listener,
 314                                     MemoryRegionSection *section)
 315 {
 316     KVMDevice *kd;
 317
 318     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
 319         if (section->mr == kd->mr) {
 320             kd->kda.addr = section->offset_within_address_space;
 321         }
 322     }
 323 }
 324
 325 static void kvm_arm_devlistener_del(MemoryListener *listener,
 326                                     MemoryRegionSection *section)
 327 {
 328     KVMDevice *kd;
 329
 330     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
 331         if (section->mr == kd->mr) {
 332             kd->kda.addr = -1;
 333         }
 334     }
 335 }
 336
 337 static MemoryListener devlistener = {
 338     .name = "kvm-arm",
 339     .region_add = kvm_arm_devlistener_add,
 340     .region_del = kvm_arm_devlistener_del,
 341 };
 342
 343 static void kvm_arm_set_device_addr(KVMDevice *kd)
 344 {
 345     struct kvm_device_attr *attr = &kd->kdattr;
 346     int ret;
 347
 348     /* If the device control API is available and we have a device fd on the
 349      * KVMDevice struct, let's use the newer API
 350      */
 351     if (kd->dev_fd >= 0) {
 352         uint64_t addr = kd->kda.addr;
 353
 354         addr |= kd->kda_addr_ormask;
 355         attr->addr = (uintptr_t)&addr;
 356         ret = kvm_device_ioctl(kd->dev_fd, KVM_SET_DEVICE_ATTR, attr);
 357     } else {
 358         ret = kvm_vm_ioctl(kvm_state, KVM_ARM_SET_DEVICE_ADDR, &kd->kda);
 359     }
 360
 361     if (ret < 0) {
 362         fprintf(stderr, "Failed to set device address: %s\n",
 363                 strerror(-ret));
 364         abort();
 365     }
 366 }
 367
 368 static void kvm_arm_machine_init_done(Notifier *notifier, void *data)
 369 {
 370     KVMDevice *kd, *tkd;
 371
 372     QSLIST_FOREACH_SAFE(kd, &kvm_devices_head, entries, tkd) {
 373         if (kd->kda.addr != -1) {
 374             kvm_arm_set_device_addr(kd);
 375         }
 376         memory_region_unref(kd->mr);
 377         QSLIST_REMOVE_HEAD(&kvm_devices_head, entries);
 378         g_free(kd);
 379     }
 380     memory_listener_unregister(&devlistener);
 381 }
 382
 383 static Notifier notify = {
 384     .notify = kvm_arm_machine_init_done,
 385 };
 386
 387 void kvm_arm_register_device(MemoryRegion *mr, uint64_t devid, uint64_t group,
 388                              uint64_t attr, int dev_fd, uint64_t addr_ormask)
 389 {
 390     KVMDevice *kd;
 391
 392     if (!kvm_irqchip_in_kernel()) {
 393         return;
 394     }
 395
 396     if (QSLIST_EMPTY(&kvm_devices_head)) {
 397         memory_listener_register(&devlistener, &address_space_memory);
 398         qemu_add_machine_init_done_notifier(&notify);
 399     }
 400     kd = g_new0(KVMDevice, 1);
 401     kd->mr = mr;
 402     kd->kda.id = devid;
 403     kd->kda.addr = -1;
 404     kd->kdattr.flags = 0;
 405     kd->kdattr.group = group;
 406     kd->kdattr.attr = attr;
 407     kd->dev_fd = dev_fd;
 408     kd->kda_addr_ormask = addr_ormask;
 409     QSLIST_INSERT_HEAD(&kvm_devices_head, kd, entries);
 410     memory_region_ref(kd->mr);
 411 }
 412
 413 static int compare_u64(const void *a, const void *b)
 414 {
 415     if (*(uint64_t *)a > *(uint64_t *)b) {
 416         return 1;
 417     }
 418     if (*(uint64_t *)a < *(uint64_t *)b) {
 419         return -1;
 420     }
 421     return 0;
 422 }
 423
 424 /*
 425  * cpreg_values are sorted in ascending order by KVM register ID
 426  * (see kvm_arm_init_cpreg_list). This allows us to cheaply find
 427  * the storage for a KVM register by ID with a binary search.
 428  */
 429 static uint64_t *kvm_arm_get_cpreg_ptr(ARMCPU *cpu, uint64_t regidx)
 430 {
 431     uint64_t *res;
 432
 433     res = bsearch(&regidx, cpu->cpreg_indexes, cpu->cpreg_array_len,
 434                   sizeof(uint64_t), compare_u64);
 435     assert(res);
 436
 437     return &cpu->cpreg_values[res - cpu->cpreg_indexes];
 438 }
 439
 440 /* Initialize the ARMCPU cpreg list according to the kernel's
 441  * definition of what CPU registers it knows about (and throw away
 442  * the previous TCG-created cpreg list).
 443  */
 444 int kvm_arm_init_cpreg_list(ARMCPU *cpu)
 445 {
 446     struct kvm_reg_list rl;
 447     struct kvm_reg_list *rlp;
 448     int i, ret, arraylen;
 449     CPUState *cs = CPU(cpu);
 450
 451     rl.n = 0;
 452     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, &rl);
 453     if (ret != -E2BIG) {
 454         return ret;
 455     }
 456     rlp = g_malloc(sizeof(struct kvm_reg_list) + rl.n * sizeof(uint64_t));
 457     rlp->n = rl.n;
 458     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, rlp);
 459     if (ret) {
 460         goto out;
 461     }
 462     /* Sort the list we get back from the kernel, since cpreg_tuples
 463      * must be in strictly ascending order.
 464      */
 465     qsort(&rlp->reg, rlp->n, sizeof(rlp->reg[0]), compare_u64);
 466
 467     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 468         if (!kvm_arm_reg_syncs_via_cpreg_list(rlp->reg[i])) {
 469             continue;
 470         }
 471         switch (rlp->reg[i] & KVM_REG_SIZE_MASK) {
 472         case KVM_REG_SIZE_U32:
 473         case KVM_REG_SIZE_U64:
 474             break;
 475         default:
 476             fprintf(stderr, "Can't handle size of register in kernel list\n");
 477             ret = -EINVAL;
 478             goto out;
 479         }
 480
 481         arraylen++;
 482     }
 483
 484     cpu->cpreg_indexes = g_renew(uint64_t, cpu->cpreg_indexes, arraylen);
 485     cpu->cpreg_values = g_renew(uint64_t, cpu->cpreg_values, arraylen);
 486     cpu->cpreg_vmstate_indexes = g_renew(uint64_t, cpu->cpreg_vmstate_indexes,
 487                                          arraylen);
 488     cpu->cpreg_vmstate_values = g_renew(uint64_t, cpu->cpreg_vmstate_values,
 489                                         arraylen);
 490     cpu->cpreg_array_len = arraylen;
 491     cpu->cpreg_vmstate_array_len = arraylen;
 492
 493     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 494         uint64_t regidx = rlp->reg[i];
 495         if (!kvm_arm_reg_syncs_via_cpreg_list(regidx)) {
 496             continue;
 497         }
 498         cpu->cpreg_indexes[arraylen] = regidx;
 499         arraylen++;
 500     }
 501     assert(cpu->cpreg_array_len == arraylen);
 502
 503     if (!write_kvmstate_to_list(cpu)) {
 504         /* Shouldn't happen unless kernel is inconsistent about
 505          * what registers exist.
 506          */
 507         fprintf(stderr, "Initial read of kernel register state failed\n");
 508         ret = -EINVAL;
 509         goto out;
 510     }
 511
 512 out:
 513     g_free(rlp);
 514     return ret;
 515 }
 516
 517 bool write_kvmstate_to_list(ARMCPU *cpu)
 518 {
 519     CPUState *cs = CPU(cpu);
 520     int i;
 521     bool ok = true;
 522
 523     for (i = 0; i < cpu->cpreg_array_len; i++) {
 524         struct kvm_one_reg r;
 525         uint64_t regidx = cpu->cpreg_indexes[i];
 526         uint32_t v32;
 527         int ret;
 528
 529         r.id = regidx;
 530
 531         switch (regidx & KVM_REG_SIZE_MASK) {
 532         case KVM_REG_SIZE_U32:
 533             r.addr = (uintptr_t)&v32;
 534             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 535             if (!ret) {
 536                 cpu->cpreg_values[i] = v32;
 537             }
 538             break;
 539         case KVM_REG_SIZE_U64:
 540             r.addr = (uintptr_t)(cpu->cpreg_values + i);
 541             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 542             break;
 543         default:
 544             abort();
 545         }
 546         if (ret) {
 547             ok = false;
 548         }
 549     }
 550     return ok;
 551 }
 552
 553 bool write_list_to_kvmstate(ARMCPU *cpu, int level)
 554 {
 555     CPUState *cs = CPU(cpu);
 556     int i;
 557     bool ok = true;
 558
 559     for (i = 0; i < cpu->cpreg_array_len; i++) {
 560         struct kvm_one_reg r;
 561         uint64_t regidx = cpu->cpreg_indexes[i];
 562         uint32_t v32;
 563         int ret;
 564
 565         if (kvm_arm_cpreg_level(regidx) > level) {
 566             continue;
 567         }
 568
 569         r.id = regidx;
 570         switch (regidx & KVM_REG_SIZE_MASK) {
 571         case KVM_REG_SIZE_U32:
 572             v32 = cpu->cpreg_values[i];
 573             r.addr = (uintptr_t)&v32;
 574             break;
 575         case KVM_REG_SIZE_U64:
 576             r.addr = (uintptr_t)(cpu->cpreg_values + i);
 577             break;
 578         default:
 579             abort();
 580         }
 581         ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
 582         if (ret) {
 583             /* We might fail for "unknown register" and also for
 584              * "you tried to set a register which is constant with
 585              * a different value from what it actually contains".
 586              */
 587             ok = false;
 588         }
 589     }
 590     return ok;
 591 }
 592
 593 void kvm_arm_cpu_pre_save(ARMCPU *cpu)
 594 {
 595     /* KVM virtual time adjustment */
 596     if (cpu->kvm_vtime_dirty) {
 597         *kvm_arm_get_cpreg_ptr(cpu, KVM_REG_ARM_TIMER_CNT) = cpu->kvm_vtime;
 598     }
 599 }
 600
 601 void kvm_arm_cpu_post_load(ARMCPU *cpu)
 602 {
 603     /* KVM virtual time adjustment */
 604     if (cpu->kvm_adjvtime) {
 605         cpu->kvm_vtime = *kvm_arm_get_cpreg_ptr(cpu, KVM_REG_ARM_TIMER_CNT);
 606         cpu->kvm_vtime_dirty = true;
 607     }
 608 }
 609
 610 void kvm_arm_reset_vcpu(ARMCPU *cpu)
 611 {
 612     int ret;
 613
 614     /* Re-init VCPU so that all registers are set to
 615      * their respective reset values.
 616      */
 617     ret = kvm_arm_vcpu_init(CPU(cpu));
 618     if (ret < 0) {
 619         fprintf(stderr, "kvm_arm_vcpu_init failed: %s\n", strerror(-ret));
 620         abort();
 621     }
 622     if (!write_kvmstate_to_list(cpu)) {
 623         fprintf(stderr, "write_kvmstate_to_list failed\n");
 624         abort();
 625     }
 626     /*
 627      * Sync the reset values also into the CPUState. This is necessary
 628      * because the next thing we do will be a kvm_arch_put_registers()
 629      * which will update the list values from the CPUState before copying
 630      * the list values back to KVM. It's OK to ignore failure returns here
 631      * for the same reason we do so in kvm_arch_get_registers().
 632      */
 633     write_list_to_cpustate(cpu);
 634 }
 635
 636 /*
 637  * Update KVM's MP_STATE based on what QEMU thinks it is
 638  */
 639 int kvm_arm_sync_mpstate_to_kvm(ARMCPU *cpu)
 640 {
 641     if (cap_has_mp_state) {
 642         struct kvm_mp_state mp_state = {
 643             .mp_state = (cpu->power_state == PSCI_OFF) ?
 644             KVM_MP_STATE_STOPPED : KVM_MP_STATE_RUNNABLE
 645         };
 646         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_SET_MP_STATE, &mp_state);
 647         if (ret) {
 648             fprintf(stderr, "%s: failed to set MP_STATE %d/%s\n",
 649                     __func__, ret, strerror(-ret));
 650             return -1;
 651         }
 652     }
 653
 654     return 0;
 655 }
 656
 657 /*
 658  * Sync the KVM MP_STATE into QEMU
 659  */
 660 int kvm_arm_sync_mpstate_to_qemu(ARMCPU *cpu)
 661 {
 662     if (cap_has_mp_state) {
 663         struct kvm_mp_state mp_state;
 664         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_GET_MP_STATE, &mp_state);
 665         if (ret) {
 666             fprintf(stderr, "%s: failed to get MP_STATE %d/%s\n",
 667                     __func__, ret, strerror(-ret));
 668             abort();
 669         }
 670         cpu->power_state = (mp_state.mp_state == KVM_MP_STATE_STOPPED) ?
 671             PSCI_OFF : PSCI_ON;
 672     }
 673
 674     return 0;
 675 }
 676
 677 void kvm_arm_get_virtual_time(CPUState *cs)
 678 {
 679     ARMCPU *cpu = ARM_CPU(cs);
 680     struct kvm_one_reg reg = {
 681         .id = KVM_REG_ARM_TIMER_CNT,
 682         .addr = (uintptr_t)&cpu->kvm_vtime,
 683     };
 684     int ret;
 685
 686     if (cpu->kvm_vtime_dirty) {
 687         return;
 688     }
 689
 690     ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
 691     if (ret) {
 692         error_report("Failed to get KVM_REG_ARM_TIMER_CNT");
 693         abort();
 694     }
 695
 696     cpu->kvm_vtime_dirty = true;
 697 }
 698
 699 void kvm_arm_put_virtual_time(CPUState *cs)
 700 {
 701     ARMCPU *cpu = ARM_CPU(cs);
 702     struct kvm_one_reg reg = {
 703         .id = KVM_REG_ARM_TIMER_CNT,
 704         .addr = (uintptr_t)&cpu->kvm_vtime,
 705     };
 706     int ret;
 707
 708     if (!cpu->kvm_vtime_dirty) {
 709         return;
 710     }
 711
 712     ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
 713     if (ret) {
 714         error_report("Failed to set KVM_REG_ARM_TIMER_CNT");
 715         abort();
 716     }
 717
 718     cpu->kvm_vtime_dirty = false;
 719 }
 720
 721 int kvm_put_vcpu_events(ARMCPU *cpu)
 722 {
 723     CPUARMState *env = &cpu->env;
 724     struct kvm_vcpu_events events;
 725     int ret;
 726
 727     if (!kvm_has_vcpu_events()) {
 728         return 0;
 729     }
 730
 731     memset(&events, 0, sizeof(events));
 732     events.exception.serror_pending = env->serror.pending;
 733
 734     /* Inject SError to guest with specified syndrome if host kernel
 735      * supports it, otherwise inject SError without syndrome.
 736      */
 737     if (cap_has_inject_serror_esr) {
 738         events.exception.serror_has_esr = env->serror.has_esr;
 739         events.exception.serror_esr = env->serror.esr;
 740     }
 741
 742     ret = kvm_vcpu_ioctl(CPU(cpu), KVM_SET_VCPU_EVENTS, &events);
 743     if (ret) {
 744         error_report("failed to put vcpu events");
 745     }
 746
 747     return ret;
 748 }
 749
 750 int kvm_get_vcpu_events(ARMCPU *cpu)
 751 {
 752     CPUARMState *env = &cpu->env;
 753     struct kvm_vcpu_events events;
 754     int ret;
 755
 756     if (!kvm_has_vcpu_events()) {
 757         return 0;
 758     }
 759
 760     memset(&events, 0, sizeof(events));
 761     ret = kvm_vcpu_ioctl(CPU(cpu), KVM_GET_VCPU_EVENTS, &events);
 762     if (ret) {
 763         error_report("failed to get vcpu events");
 764         return ret;
 765     }
 766
 767     env->serror.pending = events.exception.serror_pending;
 768     env->serror.has_esr = events.exception.serror_has_esr;
 769     env->serror.esr = events.exception.serror_esr;
 770
 771     return 0;
 772 }
 773
 774 void kvm_arch_pre_run(CPUState *cs, struct kvm_run *run)
 775 {
 776     ARMCPU *cpu = ARM_CPU(cs);
 777     CPUARMState *env = &cpu->env;
 778
 779     if (unlikely(env->ext_dabt_raised)) {
 780         /*
 781          * Verifying that the ext DABT has been properly injected,
 782          * otherwise risking indefinitely re-running the faulting instruction
 783          * Covering a very narrow case for kernels 5.5..5.5.4
 784          * when injected abort was misconfigured to be
 785          * an IMPLEMENTATION DEFINED exception (for 32-bit EL1)
 786          */
 787         if (!arm_feature(env, ARM_FEATURE_AARCH64) &&
 788             unlikely(!kvm_arm_verify_ext_dabt_pending(cs))) {
 789
 790             error_report("Data abort exception with no valid ISS generated by "
 791                    "guest memory access. KVM unable to emulate faulting "
 792                    "instruction. Failed to inject an external data abort "
 793                    "into the guest.");
 794             abort();
 795        }
 796        /* Clear the status */
 797        env->ext_dabt_raised = 0;
 798     }
 799 }
 800
 801 MemTxAttrs kvm_arch_post_run(CPUState *cs, struct kvm_run *run)
 802 {
 803     ARMCPU *cpu;
 804     uint32_t switched_level;
 805
 806     if (kvm_irqchip_in_kernel()) {
 807         /*
 808          * We only need to sync timer states with user-space interrupt
 809          * controllers, so return early and save cycles if we don't.
 810          */
 811         return MEMTXATTRS_UNSPECIFIED;
 812     }
 813
 814     cpu = ARM_CPU(cs);
 815
 816     /* Synchronize our shadowed in-kernel device irq lines with the kvm ones */
 817     if (run->s.regs.device_irq_level != cpu->device_irq_level) {
 818         switched_level = cpu->device_irq_level ^ run->s.regs.device_irq_level;
 819
 820         qemu_mutex_lock_iothread();
 821
 822         if (switched_level & KVM_ARM_DEV_EL1_VTIMER) {
 823             qemu_set_irq(cpu->gt_timer_outputs[GTIMER_VIRT],
 824                          !!(run->s.regs.device_irq_level &
 825                             KVM_ARM_DEV_EL1_VTIMER));
 826             switched_level &= ~KVM_ARM_DEV_EL1_VTIMER;
 827         }
 828
 829         if (switched_level & KVM_ARM_DEV_EL1_PTIMER) {
 830             qemu_set_irq(cpu->gt_timer_outputs[GTIMER_PHYS],
 831                          !!(run->s.regs.device_irq_level &
 832                             KVM_ARM_DEV_EL1_PTIMER));
 833             switched_level &= ~KVM_ARM_DEV_EL1_PTIMER;
 834         }
 835
 836         if (switched_level & KVM_ARM_DEV_PMU) {
 837             qemu_set_irq(cpu->pmu_interrupt,
 838                          !!(run->s.regs.device_irq_level & KVM_ARM_DEV_PMU));
 839             switched_level &= ~KVM_ARM_DEV_PMU;
 840         }
 841
 842         if (switched_level) {
 843             qemu_log_mask(LOG_UNIMP, "%s: unhandled in-kernel device IRQ %x\n",
 844                           __func__, switched_level);
 845         }
 846
 847         /* We also mark unknown levels as processed to not waste cycles */
 848         cpu->device_irq_level = run->s.regs.device_irq_level;
 849         qemu_mutex_unlock_iothread();
 850     }
 851
 852     return MEMTXATTRS_UNSPECIFIED;
 853 }
 854
 855 void kvm_arm_vm_state_change(void *opaque, bool running, RunState state)
 856 {
 857     CPUState *cs = opaque;
 858     ARMCPU *cpu = ARM_CPU(cs);
 859
 860     if (running) {
 861         if (cpu->kvm_adjvtime) {
 862             kvm_arm_put_virtual_time(cs);
 863         }
 864     } else {
 865         if (cpu->kvm_adjvtime) {
 866             kvm_arm_get_virtual_time(cs);
 867         }
 868     }
 869 }
 870
 871 /**
 872  * kvm_arm_handle_dabt_nisv:
 873  * @cs: CPUState
 874  * @esr_iss: ISS encoding (limited) for the exception from Data Abort
 875  *           ISV bit set to '0b0' -> no valid instruction syndrome
 876  * @fault_ipa: faulting address for the synchronous data abort
 877  *
 878  * Returns: 0 if the exception has been handled, < 0 otherwise
 879  */
 880 static int kvm_arm_handle_dabt_nisv(CPUState *cs, uint64_t esr_iss,
 881                                     uint64_t fault_ipa)
 882 {
 883     ARMCPU *cpu = ARM_CPU(cs);
 884     CPUARMState *env = &cpu->env;
 885     /*
 886      * Request KVM to inject the external data abort into the guest
 887      */
 888     if (cap_has_inject_ext_dabt) {
 889         struct kvm_vcpu_events events = { };
 890         /*
 891          * The external data abort event will be handled immediately by KVM
 892          * using the address fault that triggered the exit on given VCPU.
 893          * Requesting injection of the external data abort does not rely
 894          * on any other VCPU state. Therefore, in this particular case, the VCPU
 895          * synchronization can be exceptionally skipped.
 896          */
 897         events.exception.ext_dabt_pending = 1;
 898         /* KVM_CAP_ARM_INJECT_EXT_DABT implies KVM_CAP_VCPU_EVENTS */
 899         if (!kvm_vcpu_ioctl(cs, KVM_SET_VCPU_EVENTS, &events)) {
 900             env->ext_dabt_raised = 1;
 901             return 0;
 902         }
 903     } else {
 904         error_report("Data abort exception triggered by guest memory access "
 905                      "at physical address: 0x"  TARGET_FMT_lx,
 906                      (target_ulong)fault_ipa);
 907         error_printf("KVM unable to emulate faulting instruction.\n");
 908     }
 909     return -1;
 910 }
 911
 912 int kvm_arch_handle_exit(CPUState *cs, struct kvm_run *run)
 913 {
 914     int ret = 0;
 915
 916     switch (run->exit_reason) {
 917     case KVM_EXIT_DEBUG:
 918         if (kvm_arm_handle_debug(cs, &run->debug.arch)) {
 919             ret = EXCP_DEBUG;
 920         } /* otherwise return to guest */
 921         break;
 922     case KVM_EXIT_ARM_NISV:
 923         /* External DABT with no valid iss to decode */
 924         ret = kvm_arm_handle_dabt_nisv(cs, run->arm_nisv.esr_iss,
 925                                        run->arm_nisv.fault_ipa);
 926         break;
 927     default:
 928         qemu_log_mask(LOG_UNIMP, "%s: un-handled exit reason %d\n",
 929                       __func__, run->exit_reason);
 930         break;
 931     }
 932     return ret;
 933 }
 934
 935 bool kvm_arch_stop_on_emulation_error(CPUState *cs)
 936 {
 937     return true;
 938 }
 939
 940 int kvm_arch_process_async_events(CPUState *cs)
 941 {
 942     return 0;
 943 }
 944
 945 void kvm_arch_update_guest_debug(CPUState *cs, struct kvm_guest_debug *dbg)
 946 {
 947     if (kvm_sw_breakpoints_active(cs)) {
 948         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_SW_BP;
 949     }
 950     if (kvm_arm_hw_debug_active(cs)) {
 951         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_HW;
 952         kvm_arm_copy_hw_debug_data(&dbg->arch);
 953     }
 954 }
 955
 956 void kvm_arch_init_irq_routing(KVMState *s)
 957 {
 958 }
 959
 960 int kvm_arch_irqchip_create(KVMState *s)
 961 {
 962     if (kvm_kernel_irqchip_split()) {
 963         perror("-machine kernel_irqchip=split is not supported on ARM.");
 964         exit(1);
 965     }
 966
 967     /* If we can create the VGIC using the newer device control API, we
 968      * let the device do this when it initializes itself, otherwise we
 969      * fall back to the old API */
 970     return kvm_check_extension(s, KVM_CAP_DEVICE_CTRL);
 971 }
 972
 973 int kvm_arm_vgic_probe(void)
 974 {
 975     int val = 0;
 976
 977     if (kvm_create_device(kvm_state,
 978                           KVM_DEV_TYPE_ARM_VGIC_V3, true) == 0) {
 979         val |= KVM_ARM_VGIC_V3;
 980     }
 981     if (kvm_create_device(kvm_state,
 982                           KVM_DEV_TYPE_ARM_VGIC_V2, true) == 0) {
 983         val |= KVM_ARM_VGIC_V2;
 984     }
 985     return val;
 986 }
 987
 988 int kvm_arm_set_irq(int cpu, int irqtype, int irq, int level)
 989 {
 990     int kvm_irq = (irqtype << KVM_ARM_IRQ_TYPE_SHIFT) | irq;
 991     int cpu_idx1 = cpu % 256;
 992     int cpu_idx2 = cpu / 256;
 993
 994     kvm_irq |= (cpu_idx1 << KVM_ARM_IRQ_VCPU_SHIFT) |
 995                (cpu_idx2 << KVM_ARM_IRQ_VCPU2_SHIFT);
 996
 997     return kvm_set_irq(kvm_state, kvm_irq, !!level);
 998 }
 999
1000 int kvm_arch_fixup_msi_route(struct kvm_irq_routing_entry *route,
1001                              uint64_t address, uint32_t data, PCIDevice *dev)
1002 {
1003     AddressSpace *as = pci_device_iommu_address_space(dev);
1004     hwaddr xlat, len, doorbell_gpa;
1005     MemoryRegionSection mrs;
1006     MemoryRegion *mr;
1007
1008     if (as == &address_space_memory) {
1009         return 0;
1010     }
1011
1012     /* MSI doorbell address is translated by an IOMMU */
1013
1014     RCU_READ_LOCK_GUARD();
1015
1016     mr = address_space_translate(as, address, &xlat, &len, true,
1017                                  MEMTXATTRS_UNSPECIFIED);
1018
1019     if (!mr) {
1020         return 1;
1021     }
1022
1023     mrs = memory_region_find(mr, xlat, 1);
1024
1025     if (!mrs.mr) {
1026         return 1;
1027     }
1028
1029     doorbell_gpa = mrs.offset_within_address_space;
1030     memory_region_unref(mrs.mr);
1031
1032     route->u.msi.address_lo = doorbell_gpa;
1033     route->u.msi.address_hi = doorbell_gpa >> 32;
1034
1035     trace_kvm_arm_fixup_msi_route(address, doorbell_gpa);
1036
1037     return 0;
1038 }
1039
1040 int kvm_arch_add_msi_route_post(struct kvm_irq_routing_entry *route,
1041                                 int vector, PCIDevice *dev)
1042 {
1043     return 0;
1044 }
1045
1046 int kvm_arch_release_virq_post(int virq)
1047 {
1048     return 0;
1049 }
1050
1051 int kvm_arch_msi_data_to_gsi(uint32_t data)
1052 {
1053     return (data - 32) & 0xffff;
1054 }
1055
1056 bool kvm_arch_cpu_check_are_resettable(void)
1057 {
1058     return true;
1059 }