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 #include <sys/mman.h>
  14
  15 #include <linux/kvm.h>
  16
  17 #include "qemu-common.h"
  18 #include "qemu/timer.h"
  19 #include "qemu/error-report.h"
  20 #include "sysemu/sysemu.h"
  21 #include "sysemu/kvm.h"
  22 #include "kvm_arm.h"
  23 #include "cpu.h"
  24 #include "internals.h"
  25 #include "hw/arm/arm.h"
  26 #include "exec/memattrs.h"
  27 #include "hw/boards.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     ret = ioctl(vmfd, KVM_ARM_PREFERRED_TARGET, init);
  66     if (ret >= 0) {
  67         ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, init);
  68         if (ret < 0) {
  69             goto err;
  70         }
  71     } else {
  72         /* Old kernel which doesn't know about the
  73          * PREFERRED_TARGET ioctl: we know it will only support
  74          * creating one kind of guest CPU which is its preferred
  75          * CPU type.
  76          */
  77         while (*cpus_to_try != QEMU_KVM_ARM_TARGET_NONE) {
  78             init->target = *cpus_to_try++;
  79             memset(init->features, 0, sizeof(init->features));
  80             ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, init);
  81             if (ret >= 0) {
  82                 break;
  83             }
  84         }
  85         if (ret < 0) {
  86             goto err;
  87         }
  88     }
  89
  90     fdarray[0] = kvmfd;
  91     fdarray[1] = vmfd;
  92     fdarray[2] = cpufd;
  93
  94     return true;
  95
  96 err:
  97     if (cpufd >= 0) {
  98         close(cpufd);
  99     }
 100     if (vmfd >= 0) {
 101         close(vmfd);
 102     }
 103     if (kvmfd >= 0) {
 104         close(kvmfd);
 105     }
 106
 107     return false;
 108 }
 109
 110 void kvm_arm_destroy_scratch_host_vcpu(int *fdarray)
 111 {
 112     int i;
 113
 114     for (i = 2; i >= 0; i--) {
 115         close(fdarray[i]);
 116     }
 117 }
 118
 119 static void kvm_arm_host_cpu_class_init(ObjectClass *oc, void *data)
 120 {
 121     ARMHostCPUClass *ahcc = ARM_HOST_CPU_CLASS(oc);
 122
 123     /* All we really need to set up for the 'host' CPU
 124      * is the feature bits -- we rely on the fact that the
 125      * various ID register values in ARMCPU are only used for
 126      * TCG CPUs.
 127      */
 128     if (!kvm_arm_get_host_cpu_features(ahcc)) {
 129         fprintf(stderr, "Failed to retrieve host CPU features!\n");
 130         abort();
 131     }
 132 }
 133
 134 static void kvm_arm_host_cpu_initfn(Object *obj)
 135 {
 136     ARMHostCPUClass *ahcc = ARM_HOST_CPU_GET_CLASS(obj);
 137     ARMCPU *cpu = ARM_CPU(obj);
 138     CPUARMState *env = &cpu->env;
 139
 140     cpu->kvm_target = ahcc->target;
 141     cpu->dtb_compatible = ahcc->dtb_compatible;
 142     env->features = ahcc->features;
 143 }
 144
 145 static const TypeInfo host_arm_cpu_type_info = {
 146     .name = TYPE_ARM_HOST_CPU,
 147 #ifdef TARGET_AARCH64
 148     .parent = TYPE_AARCH64_CPU,
 149 #else
 150     .parent = TYPE_ARM_CPU,
 151 #endif
 152     .instance_init = kvm_arm_host_cpu_initfn,
 153     .class_init = kvm_arm_host_cpu_class_init,
 154     .class_size = sizeof(ARMHostCPUClass),
 155 };
 156
 157 int kvm_arch_init(MachineState *ms, KVMState *s)
 158 {
 159     /* For ARM interrupt delivery is always asynchronous,
 160      * whether we are using an in-kernel VGIC or not.
 161      */
 162     kvm_async_interrupts_allowed = true;
 163
 164     cap_has_mp_state = kvm_check_extension(s, KVM_CAP_MP_STATE);
 165
 166     type_register_static(&host_arm_cpu_type_info);
 167
 168     return 0;
 169 }
 170
 171 unsigned long kvm_arch_vcpu_id(CPUState *cpu)
 172 {
 173     return cpu->cpu_index;
 174 }
 175
 176 /* We track all the KVM devices which need their memory addresses
 177  * passing to the kernel in a list of these structures.
 178  * When board init is complete we run through the list and
 179  * tell the kernel the base addresses of the memory regions.
 180  * We use a MemoryListener to track mapping and unmapping of
 181  * the regions during board creation, so the board models don't
 182  * need to do anything special for the KVM case.
 183  */
 184 typedef struct KVMDevice {
 185     struct kvm_arm_device_addr kda;
 186     struct kvm_device_attr kdattr;
 187     MemoryRegion *mr;
 188     QSLIST_ENTRY(KVMDevice) entries;
 189     int dev_fd;
 190 } KVMDevice;
 191
 192 static QSLIST_HEAD(kvm_devices_head, KVMDevice) kvm_devices_head;
 193
 194 static void kvm_arm_devlistener_add(MemoryListener *listener,
 195                                     MemoryRegionSection *section)
 196 {
 197     KVMDevice *kd;
 198
 199     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
 200         if (section->mr == kd->mr) {
 201             kd->kda.addr = section->offset_within_address_space;
 202         }
 203     }
 204 }
 205
 206 static void kvm_arm_devlistener_del(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 = -1;
 214         }
 215     }
 216 }
 217
 218 static MemoryListener devlistener = {
 219     .region_add = kvm_arm_devlistener_add,
 220     .region_del = kvm_arm_devlistener_del,
 221 };
 222
 223 static void kvm_arm_set_device_addr(KVMDevice *kd)
 224 {
 225     struct kvm_device_attr *attr = &kd->kdattr;
 226     int ret;
 227
 228     /* If the device control API is available and we have a device fd on the
 229      * KVMDevice struct, let's use the newer API
 230      */
 231     if (kd->dev_fd >= 0) {
 232         uint64_t addr = kd->kda.addr;
 233         attr->addr = (uintptr_t)&addr;
 234         ret = kvm_device_ioctl(kd->dev_fd, KVM_SET_DEVICE_ATTR, attr);
 235     } else {
 236         ret = kvm_vm_ioctl(kvm_state, KVM_ARM_SET_DEVICE_ADDR, &kd->kda);
 237     }
 238
 239     if (ret < 0) {
 240         fprintf(stderr, "Failed to set device address: %s\n",
 241                 strerror(-ret));
 242         abort();
 243     }
 244 }
 245
 246 static void kvm_arm_machine_init_done(Notifier *notifier, void *data)
 247 {
 248     KVMDevice *kd, *tkd;
 249
 250     memory_listener_unregister(&devlistener);
 251     QSLIST_FOREACH_SAFE(kd, &kvm_devices_head, entries, tkd) {
 252         if (kd->kda.addr != -1) {
 253             kvm_arm_set_device_addr(kd);
 254         }
 255         memory_region_unref(kd->mr);
 256         g_free(kd);
 257     }
 258 }
 259
 260 static Notifier notify = {
 261     .notify = kvm_arm_machine_init_done,
 262 };
 263
 264 void kvm_arm_register_device(MemoryRegion *mr, uint64_t devid, uint64_t group,
 265                              uint64_t attr, int dev_fd)
 266 {
 267     KVMDevice *kd;
 268
 269     if (!kvm_irqchip_in_kernel()) {
 270         return;
 271     }
 272
 273     if (QSLIST_EMPTY(&kvm_devices_head)) {
 274         memory_listener_register(&devlistener, NULL);
 275         qemu_add_machine_init_done_notifier(&notify);
 276     }
 277     kd = g_new0(KVMDevice, 1);
 278     kd->mr = mr;
 279     kd->kda.id = devid;
 280     kd->kda.addr = -1;
 281     kd->kdattr.flags = 0;
 282     kd->kdattr.group = group;
 283     kd->kdattr.attr = attr;
 284     kd->dev_fd = dev_fd;
 285     QSLIST_INSERT_HEAD(&kvm_devices_head, kd, entries);
 286     memory_region_ref(kd->mr);
 287 }
 288
 289 static int compare_u64(const void *a, const void *b)
 290 {
 291     if (*(uint64_t *)a > *(uint64_t *)b) {
 292         return 1;
 293     }
 294     if (*(uint64_t *)a < *(uint64_t *)b) {
 295         return -1;
 296     }
 297     return 0;
 298 }
 299
 300 /* Initialize the CPUState's cpreg list according to the kernel's
 301  * definition of what CPU registers it knows about (and throw away
 302  * the previous TCG-created cpreg list).
 303  */
 304 int kvm_arm_init_cpreg_list(ARMCPU *cpu)
 305 {
 306     struct kvm_reg_list rl;
 307     struct kvm_reg_list *rlp;
 308     int i, ret, arraylen;
 309     CPUState *cs = CPU(cpu);
 310
 311     rl.n = 0;
 312     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, &rl);
 313     if (ret != -E2BIG) {
 314         return ret;
 315     }
 316     rlp = g_malloc(sizeof(struct kvm_reg_list) + rl.n * sizeof(uint64_t));
 317     rlp->n = rl.n;
 318     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, rlp);
 319     if (ret) {
 320         goto out;
 321     }
 322     /* Sort the list we get back from the kernel, since cpreg_tuples
 323      * must be in strictly ascending order.
 324      */
 325     qsort(&rlp->reg, rlp->n, sizeof(rlp->reg[0]), compare_u64);
 326
 327     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 328         if (!kvm_arm_reg_syncs_via_cpreg_list(rlp->reg[i])) {
 329             continue;
 330         }
 331         switch (rlp->reg[i] & KVM_REG_SIZE_MASK) {
 332         case KVM_REG_SIZE_U32:
 333         case KVM_REG_SIZE_U64:
 334             break;
 335         default:
 336             fprintf(stderr, "Can't handle size of register in kernel list\n");
 337             ret = -EINVAL;
 338             goto out;
 339         }
 340
 341         arraylen++;
 342     }
 343
 344     cpu->cpreg_indexes = g_renew(uint64_t, cpu->cpreg_indexes, arraylen);
 345     cpu->cpreg_values = g_renew(uint64_t, cpu->cpreg_values, arraylen);
 346     cpu->cpreg_vmstate_indexes = g_renew(uint64_t, cpu->cpreg_vmstate_indexes,
 347                                          arraylen);
 348     cpu->cpreg_vmstate_values = g_renew(uint64_t, cpu->cpreg_vmstate_values,
 349                                         arraylen);
 350     cpu->cpreg_array_len = arraylen;
 351     cpu->cpreg_vmstate_array_len = arraylen;
 352
 353     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 354         uint64_t regidx = rlp->reg[i];
 355         if (!kvm_arm_reg_syncs_via_cpreg_list(regidx)) {
 356             continue;
 357         }
 358         cpu->cpreg_indexes[arraylen] = regidx;
 359         arraylen++;
 360     }
 361     assert(cpu->cpreg_array_len == arraylen);
 362
 363     if (!write_kvmstate_to_list(cpu)) {
 364         /* Shouldn't happen unless kernel is inconsistent about
 365          * what registers exist.
 366          */
 367         fprintf(stderr, "Initial read of kernel register state failed\n");
 368         ret = -EINVAL;
 369         goto out;
 370     }
 371
 372 out:
 373     g_free(rlp);
 374     return ret;
 375 }
 376
 377 bool write_kvmstate_to_list(ARMCPU *cpu)
 378 {
 379     CPUState *cs = CPU(cpu);
 380     int i;
 381     bool ok = true;
 382
 383     for (i = 0; i < cpu->cpreg_array_len; i++) {
 384         struct kvm_one_reg r;
 385         uint64_t regidx = cpu->cpreg_indexes[i];
 386         uint32_t v32;
 387         int ret;
 388
 389         r.id = regidx;
 390
 391         switch (regidx & KVM_REG_SIZE_MASK) {
 392         case KVM_REG_SIZE_U32:
 393             r.addr = (uintptr_t)&v32;
 394             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 395             if (!ret) {
 396                 cpu->cpreg_values[i] = v32;
 397             }
 398             break;
 399         case KVM_REG_SIZE_U64:
 400             r.addr = (uintptr_t)(cpu->cpreg_values + i);
 401             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 402             break;
 403         default:
 404             abort();
 405         }
 406         if (ret) {
 407             ok = false;
 408         }
 409     }
 410     return ok;
 411 }
 412
 413 bool write_list_to_kvmstate(ARMCPU *cpu, int level)
 414 {
 415     CPUState *cs = CPU(cpu);
 416     int i;
 417     bool ok = true;
 418
 419     for (i = 0; i < cpu->cpreg_array_len; i++) {
 420         struct kvm_one_reg r;
 421         uint64_t regidx = cpu->cpreg_indexes[i];
 422         uint32_t v32;
 423         int ret;
 424
 425         if (kvm_arm_cpreg_level(regidx) > level) {
 426             continue;
 427         }
 428
 429         r.id = regidx;
 430         switch (regidx & KVM_REG_SIZE_MASK) {
 431         case KVM_REG_SIZE_U32:
 432             v32 = cpu->cpreg_values[i];
 433             r.addr = (uintptr_t)&v32;
 434             break;
 435         case KVM_REG_SIZE_U64:
 436             r.addr = (uintptr_t)(cpu->cpreg_values + i);
 437             break;
 438         default:
 439             abort();
 440         }
 441         ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
 442         if (ret) {
 443             /* We might fail for "unknown register" and also for
 444              * "you tried to set a register which is constant with
 445              * a different value from what it actually contains".
 446              */
 447             ok = false;
 448         }
 449     }
 450     return ok;
 451 }
 452
 453 void kvm_arm_reset_vcpu(ARMCPU *cpu)
 454 {
 455     int ret;
 456
 457     /* Re-init VCPU so that all registers are set to
 458      * their respective reset values.
 459      */
 460     ret = kvm_arm_vcpu_init(CPU(cpu));
 461     if (ret < 0) {
 462         fprintf(stderr, "kvm_arm_vcpu_init failed: %s\n", strerror(-ret));
 463         abort();
 464     }
 465     if (!write_kvmstate_to_list(cpu)) {
 466         fprintf(stderr, "write_kvmstate_to_list failed\n");
 467         abort();
 468     }
 469 }
 470
 471 /*
 472  * Update KVM's MP_STATE based on what QEMU thinks it is
 473  */
 474 int kvm_arm_sync_mpstate_to_kvm(ARMCPU *cpu)
 475 {
 476     if (cap_has_mp_state) {
 477         struct kvm_mp_state mp_state = {
 478             .mp_state =
 479             cpu->powered_off ? KVM_MP_STATE_STOPPED : KVM_MP_STATE_RUNNABLE
 480         };
 481         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_SET_MP_STATE, &mp_state);
 482         if (ret) {
 483             fprintf(stderr, "%s: failed to set MP_STATE %d/%s\n",
 484                     __func__, ret, strerror(-ret));
 485             return -1;
 486         }
 487     }
 488
 489     return 0;
 490 }
 491
 492 /*
 493  * Sync the KVM MP_STATE into QEMU
 494  */
 495 int kvm_arm_sync_mpstate_to_qemu(ARMCPU *cpu)
 496 {
 497     if (cap_has_mp_state) {
 498         struct kvm_mp_state mp_state;
 499         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_GET_MP_STATE, &mp_state);
 500         if (ret) {
 501             fprintf(stderr, "%s: failed to get MP_STATE %d/%s\n",
 502                     __func__, ret, strerror(-ret));
 503             abort();
 504         }
 505         cpu->powered_off = (mp_state.mp_state == KVM_MP_STATE_STOPPED);
 506     }
 507
 508     return 0;
 509 }
 510
 511 void kvm_arch_pre_run(CPUState *cs, struct kvm_run *run)
 512 {
 513 }
 514
 515 MemTxAttrs kvm_arch_post_run(CPUState *cs, struct kvm_run *run)
 516 {
 517     return MEMTXATTRS_UNSPECIFIED;
 518 }
 519
 520
 521 int kvm_arch_handle_exit(CPUState *cs, struct kvm_run *run)
 522 {
 523     int ret = 0;
 524
 525     switch (run->exit_reason) {
 526     case KVM_EXIT_DEBUG:
 527         if (kvm_arm_handle_debug(cs, &run->debug.arch)) {
 528             ret = EXCP_DEBUG;
 529         } /* otherwise return to guest */
 530         break;
 531     default:
 532         qemu_log_mask(LOG_UNIMP, "%s: un-handled exit reason %d\n",
 533                       __func__, run->exit_reason);
 534         break;
 535     }
 536     return ret;
 537 }
 538
 539 bool kvm_arch_stop_on_emulation_error(CPUState *cs)
 540 {
 541     return true;
 542 }
 543
 544 int kvm_arch_process_async_events(CPUState *cs)
 545 {
 546     return 0;
 547 }
 548
 549 int kvm_arch_on_sigbus_vcpu(CPUState *cs, int code, void *addr)
 550 {
 551     return 1;
 552 }
 553
 554 int kvm_arch_on_sigbus(int code, void *addr)
 555 {
 556     return 1;
 557 }
 558
 559 /* The #ifdef protections are until 32bit headers are imported and can
 560  * be removed once both 32 and 64 bit reach feature parity.
 561  */
 562 void kvm_arch_update_guest_debug(CPUState *cs, struct kvm_guest_debug *dbg)
 563 {
 564 #ifdef KVM_GUESTDBG_USE_SW_BP
 565     if (kvm_sw_breakpoints_active(cs)) {
 566         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_SW_BP;
 567     }
 568 #endif
 569 #ifdef KVM_GUESTDBG_USE_HW
 570     if (kvm_arm_hw_debug_active(cs)) {
 571         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_HW;
 572         kvm_arm_copy_hw_debug_data(&dbg->arch);
 573     }
 574 #endif
 575 }
 576
 577 void kvm_arch_init_irq_routing(KVMState *s)
 578 {
 579 }
 580
 581 int kvm_arch_irqchip_create(MachineState *ms, KVMState *s)
 582 {
 583      if (machine_kernel_irqchip_split(ms)) {
 584          perror("-machine kernel_irqchip=split is not supported on ARM.");
 585          exit(1);
 586     }
 587
 588     /* If we can create the VGIC using the newer device control API, we
 589      * let the device do this when it initializes itself, otherwise we
 590      * fall back to the old API */
 591     return kvm_check_extension(s, KVM_CAP_DEVICE_CTRL);
 592 }
 593
 594 int kvm_arm_vgic_probe(void)
 595 {
 596     if (kvm_create_device(kvm_state,
 597                           KVM_DEV_TYPE_ARM_VGIC_V3, true) == 0) {
 598         return 3;
 599     } else if (kvm_create_device(kvm_state,
 600                                  KVM_DEV_TYPE_ARM_VGIC_V2, true) == 0) {
 601         return 2;
 602     } else {
 603         return 0;
 604     }
 605 }
 606
 607 int kvm_arch_fixup_msi_route(struct kvm_irq_routing_entry *route,
 608                              uint64_t address, uint32_t data, PCIDevice *dev)
 609 {
 610     return 0;
 611 }
 612
 613 int kvm_arch_msi_data_to_gsi(uint32_t data)
 614 {
 615     return (data - 32) & 0xffff;
 616 }