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 <stdio.h>
  12 #include <sys/types.h>
  13 #include <sys/ioctl.h>
  14 #include <sys/mman.h>
  15
  16 #include <linux/kvm.h>
  17
  18 #include "qemu-common.h"
  19 #include "qemu/timer.h"
  20 #include "qemu/error-report.h"
  21 #include "sysemu/sysemu.h"
  22 #include "sysemu/kvm.h"
  23 #include "kvm_arm.h"
  24 #include "cpu.h"
  25 #include "internals.h"
  26 #include "hw/arm/arm.h"
  27 #include "exec/memattrs.h"
  28 #include "hw/boards.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     ret = ioctl(vmfd, KVM_ARM_PREFERRED_TARGET, init);
  67     if (ret >= 0) {
  68         ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, init);
  69         if (ret < 0) {
  70             goto err;
  71         }
  72     } else {
  73         /* Old kernel which doesn't know about the
  74          * PREFERRED_TARGET ioctl: we know it will only support
  75          * creating one kind of guest CPU which is its preferred
  76          * CPU type.
  77          */
  78         while (*cpus_to_try != QEMU_KVM_ARM_TARGET_NONE) {
  79             init->target = *cpus_to_try++;
  80             memset(init->features, 0, sizeof(init->features));
  81             ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, init);
  82             if (ret >= 0) {
  83                 break;
  84             }
  85         }
  86         if (ret < 0) {
  87             goto err;
  88         }
  89     }
  90
  91     fdarray[0] = kvmfd;
  92     fdarray[1] = vmfd;
  93     fdarray[2] = cpufd;
  94
  95     return true;
  96
  97 err:
  98     if (cpufd >= 0) {
  99         close(cpufd);
 100     }
 101     if (vmfd >= 0) {
 102         close(vmfd);
 103     }
 104     if (kvmfd >= 0) {
 105         close(kvmfd);
 106     }
 107
 108     return false;
 109 }
 110
 111 void kvm_arm_destroy_scratch_host_vcpu(int *fdarray)
 112 {
 113     int i;
 114
 115     for (i = 2; i >= 0; i--) {
 116         close(fdarray[i]);
 117     }
 118 }
 119
 120 static void kvm_arm_host_cpu_class_init(ObjectClass *oc, void *data)
 121 {
 122     ARMHostCPUClass *ahcc = ARM_HOST_CPU_CLASS(oc);
 123
 124     /* All we really need to set up for the 'host' CPU
 125      * is the feature bits -- we rely on the fact that the
 126      * various ID register values in ARMCPU are only used for
 127      * TCG CPUs.
 128      */
 129     if (!kvm_arm_get_host_cpu_features(ahcc)) {
 130         fprintf(stderr, "Failed to retrieve host CPU features!\n");
 131         abort();
 132     }
 133 }
 134
 135 static void kvm_arm_host_cpu_initfn(Object *obj)
 136 {
 137     ARMHostCPUClass *ahcc = ARM_HOST_CPU_GET_CLASS(obj);
 138     ARMCPU *cpu = ARM_CPU(obj);
 139     CPUARMState *env = &cpu->env;
 140
 141     cpu->kvm_target = ahcc->target;
 142     cpu->dtb_compatible = ahcc->dtb_compatible;
 143     env->features = ahcc->features;
 144 }
 145
 146 static const TypeInfo host_arm_cpu_type_info = {
 147     .name = TYPE_ARM_HOST_CPU,
 148 #ifdef TARGET_AARCH64
 149     .parent = TYPE_AARCH64_CPU,
 150 #else
 151     .parent = TYPE_ARM_CPU,
 152 #endif
 153     .instance_init = kvm_arm_host_cpu_initfn,
 154     .class_init = kvm_arm_host_cpu_class_init,
 155     .class_size = sizeof(ARMHostCPUClass),
 156 };
 157
 158 int kvm_arch_init(MachineState *ms, KVMState *s)
 159 {
 160     /* For ARM interrupt delivery is always asynchronous,
 161      * whether we are using an in-kernel VGIC or not.
 162      */
 163     kvm_async_interrupts_allowed = true;
 164
 165     cap_has_mp_state = kvm_check_extension(s, KVM_CAP_MP_STATE);
 166
 167     type_register_static(&host_arm_cpu_type_info);
 168
 169     return 0;
 170 }
 171
 172 unsigned long kvm_arch_vcpu_id(CPUState *cpu)
 173 {
 174     return cpu->cpu_index;
 175 }
 176
 177 /* We track all the KVM devices which need their memory addresses
 178  * passing to the kernel in a list of these structures.
 179  * When board init is complete we run through the list and
 180  * tell the kernel the base addresses of the memory regions.
 181  * We use a MemoryListener to track mapping and unmapping of
 182  * the regions during board creation, so the board models don't
 183  * need to do anything special for the KVM case.
 184  */
 185 typedef struct KVMDevice {
 186     struct kvm_arm_device_addr kda;
 187     struct kvm_device_attr kdattr;
 188     MemoryRegion *mr;
 189     QSLIST_ENTRY(KVMDevice) entries;
 190     int dev_fd;
 191 } KVMDevice;
 192
 193 static QSLIST_HEAD(kvm_devices_head, KVMDevice) kvm_devices_head;
 194
 195 static void kvm_arm_devlistener_add(MemoryListener *listener,
 196                                     MemoryRegionSection *section)
 197 {
 198     KVMDevice *kd;
 199
 200     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
 201         if (section->mr == kd->mr) {
 202             kd->kda.addr = section->offset_within_address_space;
 203         }
 204     }
 205 }
 206
 207 static void kvm_arm_devlistener_del(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 = -1;
 215         }
 216     }
 217 }
 218
 219 static MemoryListener devlistener = {
 220     .region_add = kvm_arm_devlistener_add,
 221     .region_del = kvm_arm_devlistener_del,
 222 };
 223
 224 static void kvm_arm_set_device_addr(KVMDevice *kd)
 225 {
 226     struct kvm_device_attr *attr = &kd->kdattr;
 227     int ret;
 228
 229     /* If the device control API is available and we have a device fd on the
 230      * KVMDevice struct, let's use the newer API
 231      */
 232     if (kd->dev_fd >= 0) {
 233         uint64_t addr = kd->kda.addr;
 234         attr->addr = (uintptr_t)&addr;
 235         ret = kvm_device_ioctl(kd->dev_fd, KVM_SET_DEVICE_ATTR, attr);
 236     } else {
 237         ret = kvm_vm_ioctl(kvm_state, KVM_ARM_SET_DEVICE_ADDR, &kd->kda);
 238     }
 239
 240     if (ret < 0) {
 241         fprintf(stderr, "Failed to set device address: %s\n",
 242                 strerror(-ret));
 243         abort();
 244     }
 245 }
 246
 247 static void kvm_arm_machine_init_done(Notifier *notifier, void *data)
 248 {
 249     KVMDevice *kd, *tkd;
 250
 251     memory_listener_unregister(&devlistener);
 252     QSLIST_FOREACH_SAFE(kd, &kvm_devices_head, entries, tkd) {
 253         if (kd->kda.addr != -1) {
 254             kvm_arm_set_device_addr(kd);
 255         }
 256         memory_region_unref(kd->mr);
 257         g_free(kd);
 258     }
 259 }
 260
 261 static Notifier notify = {
 262     .notify = kvm_arm_machine_init_done,
 263 };
 264
 265 void kvm_arm_register_device(MemoryRegion *mr, uint64_t devid, uint64_t group,
 266                              uint64_t attr, int dev_fd)
 267 {
 268     KVMDevice *kd;
 269
 270     if (!kvm_irqchip_in_kernel()) {
 271         return;
 272     }
 273
 274     if (QSLIST_EMPTY(&kvm_devices_head)) {
 275         memory_listener_register(&devlistener, NULL);
 276         qemu_add_machine_init_done_notifier(&notify);
 277     }
 278     kd = g_new0(KVMDevice, 1);
 279     kd->mr = mr;
 280     kd->kda.id = devid;
 281     kd->kda.addr = -1;
 282     kd->kdattr.flags = 0;
 283     kd->kdattr.group = group;
 284     kd->kdattr.attr = attr;
 285     kd->dev_fd = dev_fd;
 286     QSLIST_INSERT_HEAD(&kvm_devices_head, kd, entries);
 287     memory_region_ref(kd->mr);
 288 }
 289
 290 static int compare_u64(const void *a, const void *b)
 291 {
 292     if (*(uint64_t *)a > *(uint64_t *)b) {
 293         return 1;
 294     }
 295     if (*(uint64_t *)a < *(uint64_t *)b) {
 296         return -1;
 297     }
 298     return 0;
 299 }
 300
 301 /* Initialize the CPUState's cpreg list according to the kernel's
 302  * definition of what CPU registers it knows about (and throw away
 303  * the previous TCG-created cpreg list).
 304  */
 305 int kvm_arm_init_cpreg_list(ARMCPU *cpu)
 306 {
 307     struct kvm_reg_list rl;
 308     struct kvm_reg_list *rlp;
 309     int i, ret, arraylen;
 310     CPUState *cs = CPU(cpu);
 311
 312     rl.n = 0;
 313     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, &rl);
 314     if (ret != -E2BIG) {
 315         return ret;
 316     }
 317     rlp = g_malloc(sizeof(struct kvm_reg_list) + rl.n * sizeof(uint64_t));
 318     rlp->n = rl.n;
 319     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, rlp);
 320     if (ret) {
 321         goto out;
 322     }
 323     /* Sort the list we get back from the kernel, since cpreg_tuples
 324      * must be in strictly ascending order.
 325      */
 326     qsort(&rlp->reg, rlp->n, sizeof(rlp->reg[0]), compare_u64);
 327
 328     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 329         if (!kvm_arm_reg_syncs_via_cpreg_list(rlp->reg[i])) {
 330             continue;
 331         }
 332         switch (rlp->reg[i] & KVM_REG_SIZE_MASK) {
 333         case KVM_REG_SIZE_U32:
 334         case KVM_REG_SIZE_U64:
 335             break;
 336         default:
 337             fprintf(stderr, "Can't handle size of register in kernel list\n");
 338             ret = -EINVAL;
 339             goto out;
 340         }
 341
 342         arraylen++;
 343     }
 344
 345     cpu->cpreg_indexes = g_renew(uint64_t, cpu->cpreg_indexes, arraylen);
 346     cpu->cpreg_values = g_renew(uint64_t, cpu->cpreg_values, arraylen);
 347     cpu->cpreg_vmstate_indexes = g_renew(uint64_t, cpu->cpreg_vmstate_indexes,
 348                                          arraylen);
 349     cpu->cpreg_vmstate_values = g_renew(uint64_t, cpu->cpreg_vmstate_values,
 350                                         arraylen);
 351     cpu->cpreg_array_len = arraylen;
 352     cpu->cpreg_vmstate_array_len = arraylen;
 353
 354     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 355         uint64_t regidx = rlp->reg[i];
 356         if (!kvm_arm_reg_syncs_via_cpreg_list(regidx)) {
 357             continue;
 358         }
 359         cpu->cpreg_indexes[arraylen] = regidx;
 360         arraylen++;
 361     }
 362     assert(cpu->cpreg_array_len == arraylen);
 363
 364     if (!write_kvmstate_to_list(cpu)) {
 365         /* Shouldn't happen unless kernel is inconsistent about
 366          * what registers exist.
 367          */
 368         fprintf(stderr, "Initial read of kernel register state failed\n");
 369         ret = -EINVAL;
 370         goto out;
 371     }
 372
 373 out:
 374     g_free(rlp);
 375     return ret;
 376 }
 377
 378 bool write_kvmstate_to_list(ARMCPU *cpu)
 379 {
 380     CPUState *cs = CPU(cpu);
 381     int i;
 382     bool ok = true;
 383
 384     for (i = 0; i < cpu->cpreg_array_len; i++) {
 385         struct kvm_one_reg r;
 386         uint64_t regidx = cpu->cpreg_indexes[i];
 387         uint32_t v32;
 388         int ret;
 389
 390         r.id = regidx;
 391
 392         switch (regidx & KVM_REG_SIZE_MASK) {
 393         case KVM_REG_SIZE_U32:
 394             r.addr = (uintptr_t)&v32;
 395             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 396             if (!ret) {
 397                 cpu->cpreg_values[i] = v32;
 398             }
 399             break;
 400         case KVM_REG_SIZE_U64:
 401             r.addr = (uintptr_t)(cpu->cpreg_values + i);
 402             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 403             break;
 404         default:
 405             abort();
 406         }
 407         if (ret) {
 408             ok = false;
 409         }
 410     }
 411     return ok;
 412 }
 413
 414 bool write_list_to_kvmstate(ARMCPU *cpu, int level)
 415 {
 416     CPUState *cs = CPU(cpu);
 417     int i;
 418     bool ok = true;
 419
 420     for (i = 0; i < cpu->cpreg_array_len; i++) {
 421         struct kvm_one_reg r;
 422         uint64_t regidx = cpu->cpreg_indexes[i];
 423         uint32_t v32;
 424         int ret;
 425
 426         if (kvm_arm_cpreg_level(regidx) > level) {
 427             continue;
 428         }
 429
 430         r.id = regidx;
 431         switch (regidx & KVM_REG_SIZE_MASK) {
 432         case KVM_REG_SIZE_U32:
 433             v32 = cpu->cpreg_values[i];
 434             r.addr = (uintptr_t)&v32;
 435             break;
 436         case KVM_REG_SIZE_U64:
 437             r.addr = (uintptr_t)(cpu->cpreg_values + i);
 438             break;
 439         default:
 440             abort();
 441         }
 442         ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
 443         if (ret) {
 444             /* We might fail for "unknown register" and also for
 445              * "you tried to set a register which is constant with
 446              * a different value from what it actually contains".
 447              */
 448             ok = false;
 449         }
 450     }
 451     return ok;
 452 }
 453
 454 void kvm_arm_reset_vcpu(ARMCPU *cpu)
 455 {
 456     int ret;
 457
 458     /* Re-init VCPU so that all registers are set to
 459      * their respective reset values.
 460      */
 461     ret = kvm_arm_vcpu_init(CPU(cpu));
 462     if (ret < 0) {
 463         fprintf(stderr, "kvm_arm_vcpu_init failed: %s\n", strerror(-ret));
 464         abort();
 465     }
 466     if (!write_kvmstate_to_list(cpu)) {
 467         fprintf(stderr, "write_kvmstate_to_list failed\n");
 468         abort();
 469     }
 470 }
 471
 472 /*
 473  * Update KVM's MP_STATE based on what QEMU thinks it is
 474  */
 475 int kvm_arm_sync_mpstate_to_kvm(ARMCPU *cpu)
 476 {
 477     if (cap_has_mp_state) {
 478         struct kvm_mp_state mp_state = {
 479             .mp_state =
 480             cpu->powered_off ? KVM_MP_STATE_STOPPED : KVM_MP_STATE_RUNNABLE
 481         };
 482         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_SET_MP_STATE, &mp_state);
 483         if (ret) {
 484             fprintf(stderr, "%s: failed to set MP_STATE %d/%s\n",
 485                     __func__, ret, strerror(-ret));
 486             return -1;
 487         }
 488     }
 489
 490     return 0;
 491 }
 492
 493 /*
 494  * Sync the KVM MP_STATE into QEMU
 495  */
 496 int kvm_arm_sync_mpstate_to_qemu(ARMCPU *cpu)
 497 {
 498     if (cap_has_mp_state) {
 499         struct kvm_mp_state mp_state;
 500         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_GET_MP_STATE, &mp_state);
 501         if (ret) {
 502             fprintf(stderr, "%s: failed to get MP_STATE %d/%s\n",
 503                     __func__, ret, strerror(-ret));
 504             abort();
 505         }
 506         cpu->powered_off = (mp_state.mp_state == KVM_MP_STATE_STOPPED);
 507     }
 508
 509     return 0;
 510 }
 511
 512 void kvm_arch_pre_run(CPUState *cs, struct kvm_run *run)
 513 {
 514 }
 515
 516 MemTxAttrs kvm_arch_post_run(CPUState *cs, struct kvm_run *run)
 517 {
 518     return MEMTXATTRS_UNSPECIFIED;
 519 }
 520
 521
 522 int kvm_arch_handle_exit(CPUState *cs, struct kvm_run *run)
 523 {
 524     int ret = 0;
 525
 526     switch (run->exit_reason) {
 527     case KVM_EXIT_DEBUG:
 528         if (kvm_arm_handle_debug(cs, &run->debug.arch)) {
 529             ret = EXCP_DEBUG;
 530         } /* otherwise return to guest */
 531         break;
 532     default:
 533         qemu_log_mask(LOG_UNIMP, "%s: un-handled exit reason %d\n",
 534                       __func__, run->exit_reason);
 535         break;
 536     }
 537     return ret;
 538 }
 539
 540 bool kvm_arch_stop_on_emulation_error(CPUState *cs)
 541 {
 542     return true;
 543 }
 544
 545 int kvm_arch_process_async_events(CPUState *cs)
 546 {
 547     return 0;
 548 }
 549
 550 int kvm_arch_on_sigbus_vcpu(CPUState *cs, int code, void *addr)
 551 {
 552     return 1;
 553 }
 554
 555 int kvm_arch_on_sigbus(int code, void *addr)
 556 {
 557     return 1;
 558 }
 559
 560 /* The #ifdef protections are until 32bit headers are imported and can
 561  * be removed once both 32 and 64 bit reach feature parity.
 562  */
 563 void kvm_arch_update_guest_debug(CPUState *cs, struct kvm_guest_debug *dbg)
 564 {
 565 #ifdef KVM_GUESTDBG_USE_SW_BP
 566     if (kvm_sw_breakpoints_active(cs)) {
 567         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_SW_BP;
 568     }
 569 #endif
 570 #ifdef KVM_GUESTDBG_USE_HW
 571     if (kvm_arm_hw_debug_active(cs)) {
 572         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_HW;
 573         kvm_arm_copy_hw_debug_data(&dbg->arch);
 574     }
 575 #endif
 576 }
 577
 578 void kvm_arch_init_irq_routing(KVMState *s)
 579 {
 580 }
 581
 582 int kvm_arch_irqchip_create(MachineState *ms, KVMState *s)
 583 {
 584      if (machine_kernel_irqchip_split(ms)) {
 585          perror("-machine kernel_irqchip=split is not supported on ARM.");
 586          exit(1);
 587     }
 588
 589     /* If we can create the VGIC using the newer device control API, we
 590      * let the device do this when it initializes itself, otherwise we
 591      * fall back to the old API */
 592     return kvm_check_extension(s, KVM_CAP_DEVICE_CTRL);
 593 }
 594
 595 int kvm_arm_vgic_probe(void)
 596 {
 597     if (kvm_create_device(kvm_state,
 598                           KVM_DEV_TYPE_ARM_VGIC_V3, true) == 0) {
 599         return 3;
 600     } else if (kvm_create_device(kvm_state,
 601                                  KVM_DEV_TYPE_ARM_VGIC_V2, true) == 0) {
 602         return 2;
 603     } else {
 604         return 0;
 605     }
 606 }
 607
 608 int kvm_arch_fixup_msi_route(struct kvm_irq_routing_entry *route,
 609                              uint64_t address, uint32_t data, PCIDevice *dev)
 610 {
 611     return 0;
 612 }
 613
 614 int kvm_arch_msi_data_to_gsi(uint32_t data)
 615 {
 616     return (data - 32) & 0xffff;
 617 }