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