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