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