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 "sysemu/sysemu.h"
  20 #include "sysemu/kvm.h"
  21 #include "kvm_arm.h"
  22 #include "cpu.h"
  23 #include "internals.h"
  24 #include "hw/arm/arm.h"
  25 #include "exec/memattrs.h"
  26 #include "exec/address-spaces.h"
  27 #include "hw/boards.h"
  28 #include "qemu/log.h"
  29
  30 const KVMCapabilityInfo kvm_arch_required_capabilities[] = {
  31     KVM_CAP_LAST_INFO
  32 };
  33
  34 static bool cap_has_mp_state;
  35
  36 int kvm_arm_vcpu_init(CPUState *cs)
  37 {
  38     ARMCPU *cpu = ARM_CPU(cs);
  39     struct kvm_vcpu_init init;
  40
  41     init.target = cpu->kvm_target;
  42     memcpy(init.features, cpu->kvm_init_features, sizeof(init.features));
  43
  44     return kvm_vcpu_ioctl(cs, KVM_ARM_VCPU_INIT, &init);
  45 }
  46
  47 bool kvm_arm_create_scratch_host_vcpu(const uint32_t *cpus_to_try,
  48                                       int *fdarray,
  49                                       struct kvm_vcpu_init *init)
  50 {
  51     int ret, kvmfd = -1, vmfd = -1, cpufd = -1;
  52
  53     kvmfd = qemu_open("/dev/kvm", O_RDWR);
  54     if (kvmfd < 0) {
  55         goto err;
  56     }
  57     vmfd = ioctl(kvmfd, KVM_CREATE_VM, 0);
  58     if (vmfd < 0) {
  59         goto err;
  60     }
  61     cpufd = ioctl(vmfd, KVM_CREATE_VCPU, 0);
  62     if (cpufd < 0) {
  63         goto err;
  64     }
  65
  66     if (!init) {
  67         /* Caller doesn't want the VCPU to be initialized, so skip it */
  68         goto finish;
  69     }
  70
  71     ret = ioctl(vmfd, KVM_ARM_PREFERRED_TARGET, init);
  72     if (ret >= 0) {
  73         ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, init);
  74         if (ret < 0) {
  75             goto err;
  76         }
  77     } else if (cpus_to_try) {
  78         /* Old kernel which doesn't know about the
  79          * PREFERRED_TARGET ioctl: we know it will only support
  80          * creating one kind of guest CPU which is its preferred
  81          * CPU type.
  82          */
  83         while (*cpus_to_try != QEMU_KVM_ARM_TARGET_NONE) {
  84             init->target = *cpus_to_try++;
  85             memset(init->features, 0, sizeof(init->features));
  86             ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, init);
  87             if (ret >= 0) {
  88                 break;
  89             }
  90         }
  91         if (ret < 0) {
  92             goto err;
  93         }
  94     } else {
  95         /* Treat a NULL cpus_to_try argument the same as an empty
  96          * list, which means we will fail the call since this must
  97          * be an old kernel which doesn't support PREFERRED_TARGET.
  98          */
  99         goto err;
 100     }
 101
 102 finish:
 103     fdarray[0] = kvmfd;
 104     fdarray[1] = vmfd;
 105     fdarray[2] = cpufd;
 106
 107     return true;
 108
 109 err:
 110     if (cpufd >= 0) {
 111         close(cpufd);
 112     }
 113     if (vmfd >= 0) {
 114         close(vmfd);
 115     }
 116     if (kvmfd >= 0) {
 117         close(kvmfd);
 118     }
 119
 120     return false;
 121 }
 122
 123 void kvm_arm_destroy_scratch_host_vcpu(int *fdarray)
 124 {
 125     int i;
 126
 127     for (i = 2; i >= 0; i--) {
 128         close(fdarray[i]);
 129     }
 130 }
 131
 132 static void kvm_arm_host_cpu_class_init(ObjectClass *oc, void *data)
 133 {
 134     ARMHostCPUClass *ahcc = ARM_HOST_CPU_CLASS(oc);
 135
 136     /* All we really need to set up for the 'host' CPU
 137      * is the feature bits -- we rely on the fact that the
 138      * various ID register values in ARMCPU are only used for
 139      * TCG CPUs.
 140      */
 141     if (!kvm_arm_get_host_cpu_features(ahcc)) {
 142         fprintf(stderr, "Failed to retrieve host CPU features!\n");
 143         abort();
 144     }
 145 }
 146
 147 static void kvm_arm_host_cpu_initfn(Object *obj)
 148 {
 149     ARMHostCPUClass *ahcc = ARM_HOST_CPU_GET_CLASS(obj);
 150     ARMCPU *cpu = ARM_CPU(obj);
 151     CPUARMState *env = &cpu->env;
 152
 153     cpu->kvm_target = ahcc->target;
 154     cpu->dtb_compatible = ahcc->dtb_compatible;
 155     env->features = ahcc->features;
 156 }
 157
 158 static const TypeInfo host_arm_cpu_type_info = {
 159     .name = TYPE_ARM_HOST_CPU,
 160 #ifdef TARGET_AARCH64
 161     .parent = TYPE_AARCH64_CPU,
 162 #else
 163     .parent = TYPE_ARM_CPU,
 164 #endif
 165     .instance_init = kvm_arm_host_cpu_initfn,
 166     .class_init = kvm_arm_host_cpu_class_init,
 167     .class_size = sizeof(ARMHostCPUClass),
 168 };
 169
 170 int kvm_arch_init(MachineState *ms, KVMState *s)
 171 {
 172     /* For ARM interrupt delivery is always asynchronous,
 173      * whether we are using an in-kernel VGIC or not.
 174      */
 175     kvm_async_interrupts_allowed = true;
 176
 177     cap_has_mp_state = kvm_check_extension(s, KVM_CAP_MP_STATE);
 178
 179     type_register_static(&host_arm_cpu_type_info);
 180
 181     return 0;
 182 }
 183
 184 unsigned long kvm_arch_vcpu_id(CPUState *cpu)
 185 {
 186     return cpu->cpu_index;
 187 }
 188
 189 /* We track all the KVM devices which need their memory addresses
 190  * passing to the kernel in a list of these structures.
 191  * When board init is complete we run through the list and
 192  * tell the kernel the base addresses of the memory regions.
 193  * We use a MemoryListener to track mapping and unmapping of
 194  * the regions during board creation, so the board models don't
 195  * need to do anything special for the KVM case.
 196  */
 197 typedef struct KVMDevice {
 198     struct kvm_arm_device_addr kda;
 199     struct kvm_device_attr kdattr;
 200     MemoryRegion *mr;
 201     QSLIST_ENTRY(KVMDevice) entries;
 202     int dev_fd;
 203 } KVMDevice;
 204
 205 static QSLIST_HEAD(kvm_devices_head, KVMDevice) kvm_devices_head;
 206
 207 static void kvm_arm_devlistener_add(MemoryListener *listener,
 208                                     MemoryRegionSection *section)
 209 {
 210     KVMDevice *kd;
 211
 212     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
 213         if (section->mr == kd->mr) {
 214             kd->kda.addr = section->offset_within_address_space;
 215         }
 216     }
 217 }
 218
 219 static void kvm_arm_devlistener_del(MemoryListener *listener,
 220                                     MemoryRegionSection *section)
 221 {
 222     KVMDevice *kd;
 223
 224     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
 225         if (section->mr == kd->mr) {
 226             kd->kda.addr = -1;
 227         }
 228     }
 229 }
 230
 231 static MemoryListener devlistener = {
 232     .region_add = kvm_arm_devlistener_add,
 233     .region_del = kvm_arm_devlistener_del,
 234 };
 235
 236 static void kvm_arm_set_device_addr(KVMDevice *kd)
 237 {
 238     struct kvm_device_attr *attr = &kd->kdattr;
 239     int ret;
 240
 241     /* If the device control API is available and we have a device fd on the
 242      * KVMDevice struct, let's use the newer API
 243      */
 244     if (kd->dev_fd >= 0) {
 245         uint64_t addr = kd->kda.addr;
 246         attr->addr = (uintptr_t)&addr;
 247         ret = kvm_device_ioctl(kd->dev_fd, KVM_SET_DEVICE_ATTR, attr);
 248     } else {
 249         ret = kvm_vm_ioctl(kvm_state, KVM_ARM_SET_DEVICE_ADDR, &kd->kda);
 250     }
 251
 252     if (ret < 0) {
 253         fprintf(stderr, "Failed to set device address: %s\n",
 254                 strerror(-ret));
 255         abort();
 256     }
 257 }
 258
 259 static void kvm_arm_machine_init_done(Notifier *notifier, void *data)
 260 {
 261     KVMDevice *kd, *tkd;
 262
 263     memory_listener_unregister(&devlistener);
 264     QSLIST_FOREACH_SAFE(kd, &kvm_devices_head, entries, tkd) {
 265         if (kd->kda.addr != -1) {
 266             kvm_arm_set_device_addr(kd);
 267         }
 268         memory_region_unref(kd->mr);
 269         g_free(kd);
 270     }
 271 }
 272
 273 static Notifier notify = {
 274     .notify = kvm_arm_machine_init_done,
 275 };
 276
 277 void kvm_arm_register_device(MemoryRegion *mr, uint64_t devid, uint64_t group,
 278                              uint64_t attr, int dev_fd)
 279 {
 280     KVMDevice *kd;
 281
 282     if (!kvm_irqchip_in_kernel()) {
 283         return;
 284     }
 285
 286     if (QSLIST_EMPTY(&kvm_devices_head)) {
 287         memory_listener_register(&devlistener, &address_space_memory);
 288         qemu_add_machine_init_done_notifier(&notify);
 289     }
 290     kd = g_new0(KVMDevice, 1);
 291     kd->mr = mr;
 292     kd->kda.id = devid;
 293     kd->kda.addr = -1;
 294     kd->kdattr.flags = 0;
 295     kd->kdattr.group = group;
 296     kd->kdattr.attr = attr;
 297     kd->dev_fd = dev_fd;
 298     QSLIST_INSERT_HEAD(&kvm_devices_head, kd, entries);
 299     memory_region_ref(kd->mr);
 300 }
 301
 302 static int compare_u64(const void *a, const void *b)
 303 {
 304     if (*(uint64_t *)a > *(uint64_t *)b) {
 305         return 1;
 306     }
 307     if (*(uint64_t *)a < *(uint64_t *)b) {
 308         return -1;
 309     }
 310     return 0;
 311 }
 312
 313 /* Initialize the CPUState's cpreg list according to the kernel's
 314  * definition of what CPU registers it knows about (and throw away
 315  * the previous TCG-created cpreg list).
 316  */
 317 int kvm_arm_init_cpreg_list(ARMCPU *cpu)
 318 {
 319     struct kvm_reg_list rl;
 320     struct kvm_reg_list *rlp;
 321     int i, ret, arraylen;
 322     CPUState *cs = CPU(cpu);
 323
 324     rl.n = 0;
 325     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, &rl);
 326     if (ret != -E2BIG) {
 327         return ret;
 328     }
 329     rlp = g_malloc(sizeof(struct kvm_reg_list) + rl.n * sizeof(uint64_t));
 330     rlp->n = rl.n;
 331     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, rlp);
 332     if (ret) {
 333         goto out;
 334     }
 335     /* Sort the list we get back from the kernel, since cpreg_tuples
 336      * must be in strictly ascending order.
 337      */
 338     qsort(&rlp->reg, rlp->n, sizeof(rlp->reg[0]), compare_u64);
 339
 340     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 341         if (!kvm_arm_reg_syncs_via_cpreg_list(rlp->reg[i])) {
 342             continue;
 343         }
 344         switch (rlp->reg[i] & KVM_REG_SIZE_MASK) {
 345         case KVM_REG_SIZE_U32:
 346         case KVM_REG_SIZE_U64:
 347             break;
 348         default:
 349             fprintf(stderr, "Can't handle size of register in kernel list\n");
 350             ret = -EINVAL;
 351             goto out;
 352         }
 353
 354         arraylen++;
 355     }
 356
 357     cpu->cpreg_indexes = g_renew(uint64_t, cpu->cpreg_indexes, arraylen);
 358     cpu->cpreg_values = g_renew(uint64_t, cpu->cpreg_values, arraylen);
 359     cpu->cpreg_vmstate_indexes = g_renew(uint64_t, cpu->cpreg_vmstate_indexes,
 360                                          arraylen);
 361     cpu->cpreg_vmstate_values = g_renew(uint64_t, cpu->cpreg_vmstate_values,
 362                                         arraylen);
 363     cpu->cpreg_array_len = arraylen;
 364     cpu->cpreg_vmstate_array_len = arraylen;
 365
 366     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 367         uint64_t regidx = rlp->reg[i];
 368         if (!kvm_arm_reg_syncs_via_cpreg_list(regidx)) {
 369             continue;
 370         }
 371         cpu->cpreg_indexes[arraylen] = regidx;
 372         arraylen++;
 373     }
 374     assert(cpu->cpreg_array_len == arraylen);
 375
 376     if (!write_kvmstate_to_list(cpu)) {
 377         /* Shouldn't happen unless kernel is inconsistent about
 378          * what registers exist.
 379          */
 380         fprintf(stderr, "Initial read of kernel register state failed\n");
 381         ret = -EINVAL;
 382         goto out;
 383     }
 384
 385 out:
 386     g_free(rlp);
 387     return ret;
 388 }
 389
 390 bool write_kvmstate_to_list(ARMCPU *cpu)
 391 {
 392     CPUState *cs = CPU(cpu);
 393     int i;
 394     bool ok = true;
 395
 396     for (i = 0; i < cpu->cpreg_array_len; i++) {
 397         struct kvm_one_reg r;
 398         uint64_t regidx = cpu->cpreg_indexes[i];
 399         uint32_t v32;
 400         int ret;
 401
 402         r.id = regidx;
 403
 404         switch (regidx & KVM_REG_SIZE_MASK) {
 405         case KVM_REG_SIZE_U32:
 406             r.addr = (uintptr_t)&v32;
 407             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 408             if (!ret) {
 409                 cpu->cpreg_values[i] = v32;
 410             }
 411             break;
 412         case KVM_REG_SIZE_U64:
 413             r.addr = (uintptr_t)(cpu->cpreg_values + i);
 414             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 415             break;
 416         default:
 417             abort();
 418         }
 419         if (ret) {
 420             ok = false;
 421         }
 422     }
 423     return ok;
 424 }
 425
 426 bool write_list_to_kvmstate(ARMCPU *cpu, int level)
 427 {
 428     CPUState *cs = CPU(cpu);
 429     int i;
 430     bool ok = true;
 431
 432     for (i = 0; i < cpu->cpreg_array_len; i++) {
 433         struct kvm_one_reg r;
 434         uint64_t regidx = cpu->cpreg_indexes[i];
 435         uint32_t v32;
 436         int ret;
 437
 438         if (kvm_arm_cpreg_level(regidx) > level) {
 439             continue;
 440         }
 441
 442         r.id = regidx;
 443         switch (regidx & KVM_REG_SIZE_MASK) {
 444         case KVM_REG_SIZE_U32:
 445             v32 = cpu->cpreg_values[i];
 446             r.addr = (uintptr_t)&v32;
 447             break;
 448         case KVM_REG_SIZE_U64:
 449             r.addr = (uintptr_t)(cpu->cpreg_values + i);
 450             break;
 451         default:
 452             abort();
 453         }
 454         ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
 455         if (ret) {
 456             /* We might fail for "unknown register" and also for
 457              * "you tried to set a register which is constant with
 458              * a different value from what it actually contains".
 459              */
 460             ok = false;
 461         }
 462     }
 463     return ok;
 464 }
 465
 466 void kvm_arm_reset_vcpu(ARMCPU *cpu)
 467 {
 468     int ret;
 469
 470     /* Re-init VCPU so that all registers are set to
 471      * their respective reset values.
 472      */
 473     ret = kvm_arm_vcpu_init(CPU(cpu));
 474     if (ret < 0) {
 475         fprintf(stderr, "kvm_arm_vcpu_init failed: %s\n", strerror(-ret));
 476         abort();
 477     }
 478     if (!write_kvmstate_to_list(cpu)) {
 479         fprintf(stderr, "write_kvmstate_to_list failed\n");
 480         abort();
 481     }
 482 }
 483
 484 /*
 485  * Update KVM's MP_STATE based on what QEMU thinks it is
 486  */
 487 int kvm_arm_sync_mpstate_to_kvm(ARMCPU *cpu)
 488 {
 489     if (cap_has_mp_state) {
 490         struct kvm_mp_state mp_state = {
 491             .mp_state =
 492             cpu->powered_off ? KVM_MP_STATE_STOPPED : KVM_MP_STATE_RUNNABLE
 493         };
 494         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_SET_MP_STATE, &mp_state);
 495         if (ret) {
 496             fprintf(stderr, "%s: failed to set MP_STATE %d/%s\n",
 497                     __func__, ret, strerror(-ret));
 498             return -1;
 499         }
 500     }
 501
 502     return 0;
 503 }
 504
 505 /*
 506  * Sync the KVM MP_STATE into QEMU
 507  */
 508 int kvm_arm_sync_mpstate_to_qemu(ARMCPU *cpu)
 509 {
 510     if (cap_has_mp_state) {
 511         struct kvm_mp_state mp_state;
 512         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_GET_MP_STATE, &mp_state);
 513         if (ret) {
 514             fprintf(stderr, "%s: failed to get MP_STATE %d/%s\n",
 515                     __func__, ret, strerror(-ret));
 516             abort();
 517         }
 518         cpu->powered_off = (mp_state.mp_state == KVM_MP_STATE_STOPPED);
 519     }
 520
 521     return 0;
 522 }
 523
 524 void kvm_arch_pre_run(CPUState *cs, struct kvm_run *run)
 525 {
 526 }
 527
 528 MemTxAttrs kvm_arch_post_run(CPUState *cs, struct kvm_run *run)
 529 {
 530     return MEMTXATTRS_UNSPECIFIED;
 531 }
 532
 533
 534 int kvm_arch_handle_exit(CPUState *cs, struct kvm_run *run)
 535 {
 536     int ret = 0;
 537
 538     switch (run->exit_reason) {
 539     case KVM_EXIT_DEBUG:
 540         if (kvm_arm_handle_debug(cs, &run->debug.arch)) {
 541             ret = EXCP_DEBUG;
 542         } /* otherwise return to guest */
 543         break;
 544     default:
 545         qemu_log_mask(LOG_UNIMP, "%s: un-handled exit reason %d\n",
 546                       __func__, run->exit_reason);
 547         break;
 548     }
 549     return ret;
 550 }
 551
 552 bool kvm_arch_stop_on_emulation_error(CPUState *cs)
 553 {
 554     return true;
 555 }
 556
 557 int kvm_arch_process_async_events(CPUState *cs)
 558 {
 559     return 0;
 560 }
 561
 562 int kvm_arch_on_sigbus_vcpu(CPUState *cs, int code, void *addr)
 563 {
 564     return 1;
 565 }
 566
 567 int kvm_arch_on_sigbus(int code, void *addr)
 568 {
 569     return 1;
 570 }
 571
 572 /* The #ifdef protections are until 32bit headers are imported and can
 573  * be removed once both 32 and 64 bit reach feature parity.
 574  */
 575 void kvm_arch_update_guest_debug(CPUState *cs, struct kvm_guest_debug *dbg)
 576 {
 577 #ifdef KVM_GUESTDBG_USE_SW_BP
 578     if (kvm_sw_breakpoints_active(cs)) {
 579         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_SW_BP;
 580     }
 581 #endif
 582 #ifdef KVM_GUESTDBG_USE_HW
 583     if (kvm_arm_hw_debug_active(cs)) {
 584         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_HW;
 585         kvm_arm_copy_hw_debug_data(&dbg->arch);
 586     }
 587 #endif
 588 }
 589
 590 void kvm_arch_init_irq_routing(KVMState *s)
 591 {
 592 }
 593
 594 int kvm_arch_irqchip_create(MachineState *ms, KVMState *s)
 595 {
 596      if (machine_kernel_irqchip_split(ms)) {
 597          perror("-machine kernel_irqchip=split is not supported on ARM.");
 598          exit(1);
 599     }
 600
 601     /* If we can create the VGIC using the newer device control API, we
 602      * let the device do this when it initializes itself, otherwise we
 603      * fall back to the old API */
 604     return kvm_check_extension(s, KVM_CAP_DEVICE_CTRL);
 605 }
 606
 607 int kvm_arm_vgic_probe(void)
 608 {
 609     if (kvm_create_device(kvm_state,
 610                           KVM_DEV_TYPE_ARM_VGIC_V3, true) == 0) {
 611         return 3;
 612     } else if (kvm_create_device(kvm_state,
 613                                  KVM_DEV_TYPE_ARM_VGIC_V2, true) == 0) {
 614         return 2;
 615     } else {
 616         return 0;
 617     }
 618 }
 619
 620 int kvm_arch_fixup_msi_route(struct kvm_irq_routing_entry *route,
 621                              uint64_t address, uint32_t data, PCIDevice *dev)
 622 {
 623     return 0;
 624 }
 625
 626 int kvm_arch_add_msi_route_post(struct kvm_irq_routing_entry *route,
 627                                 int vector, PCIDevice *dev)
 628 {
 629     return 0;
 630 }
 631
 632 int kvm_arch_release_virq_post(int virq)
 633 {
 634     return 0;
 635 }
 636
 637 int kvm_arch_msi_data_to_gsi(uint32_t data)
 638 {
 639     return (data - 32) & 0xffff;
 640 }