target-arm/kvm32.c

   1 /*
   2  * ARM implementation of KVM hooks, 32 bit specific code.
   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 static inline void set_feature(uint64_t *features, int feature)
  28 {
  29     *features |= 1ULL << feature;
  30 }
  31
  32 bool kvm_arm_get_host_cpu_features(ARMHostCPUClass *ahcc)
  33 {
  34     /* Identify the feature bits corresponding to the host CPU, and
  35      * fill out the ARMHostCPUClass fields accordingly. To do this
  36      * we have to create a scratch VM, create a single CPU inside it,
  37      * and then query that CPU for the relevant ID registers.
  38      */
  39     int i, ret, fdarray[3];
  40     uint32_t midr, id_pfr0, id_isar0, mvfr1;
  41     uint64_t features = 0;
  42     /* Old kernels may not know about the PREFERRED_TARGET ioctl: however
  43      * we know these will only support creating one kind of guest CPU,
  44      * which is its preferred CPU type.
  45      */
  46     static const uint32_t cpus_to_try[] = {
  47         QEMU_KVM_ARM_TARGET_CORTEX_A15,
  48         QEMU_KVM_ARM_TARGET_NONE
  49     };
  50     struct kvm_vcpu_init init;
  51     struct kvm_one_reg idregs[] = {
  52         {
  53             .id = KVM_REG_ARM | KVM_REG_SIZE_U32
  54             | ENCODE_CP_REG(15, 0, 0, 0, 0, 0),
  55             .addr = (uintptr_t)&midr,
  56         },
  57         {
  58             .id = KVM_REG_ARM | KVM_REG_SIZE_U32
  59             | ENCODE_CP_REG(15, 0, 0, 1, 0, 0),
  60             .addr = (uintptr_t)&id_pfr0,
  61         },
  62         {
  63             .id = KVM_REG_ARM | KVM_REG_SIZE_U32
  64             | ENCODE_CP_REG(15, 0, 0, 2, 0, 0),
  65             .addr = (uintptr_t)&id_isar0,
  66         },
  67         {
  68             .id = KVM_REG_ARM | KVM_REG_SIZE_U32
  69             | KVM_REG_ARM_VFP | KVM_REG_ARM_VFP_MVFR1,
  70             .addr = (uintptr_t)&mvfr1,
  71         },
  72     };
  73
  74     if (!kvm_arm_create_scratch_host_vcpu(cpus_to_try, fdarray, &init)) {
  75         return false;
  76     }
  77
  78     ahcc->target = init.target;
  79
  80     /* This is not strictly blessed by the device tree binding docs yet,
  81      * but in practice the kernel does not care about this string so
  82      * there is no point maintaining an KVM_ARM_TARGET_* -> string table.
  83      */
  84     ahcc->dtb_compatible = "arm,arm-v7";
  85
  86     for (i = 0; i < ARRAY_SIZE(idregs); i++) {
  87         ret = ioctl(fdarray[2], KVM_GET_ONE_REG, &idregs[i]);
  88         if (ret) {
  89             break;
  90         }
  91     }
  92
  93     kvm_arm_destroy_scratch_host_vcpu(fdarray);
  94
  95     if (ret) {
  96         return false;
  97     }
  98
  99     /* Now we've retrieved all the register information we can
 100      * set the feature bits based on the ID register fields.
 101      * We can assume any KVM supporting CPU is at least a v7
 102      * with VFPv3, LPAE and the generic timers; this in turn implies
 103      * most of the other feature bits, but a few must be tested.
 104      */
 105     set_feature(&features, ARM_FEATURE_V7);
 106     set_feature(&features, ARM_FEATURE_VFP3);
 107     set_feature(&features, ARM_FEATURE_LPAE);
 108     set_feature(&features, ARM_FEATURE_GENERIC_TIMER);
 109
 110     switch (extract32(id_isar0, 24, 4)) {
 111     case 1:
 112         set_feature(&features, ARM_FEATURE_THUMB_DIV);
 113         break;
 114     case 2:
 115         set_feature(&features, ARM_FEATURE_ARM_DIV);
 116         set_feature(&features, ARM_FEATURE_THUMB_DIV);
 117         break;
 118     default:
 119         break;
 120     }
 121
 122     if (extract32(id_pfr0, 12, 4) == 1) {
 123         set_feature(&features, ARM_FEATURE_THUMB2EE);
 124     }
 125     if (extract32(mvfr1, 20, 4) == 1) {
 126         set_feature(&features, ARM_FEATURE_VFP_FP16);
 127     }
 128     if (extract32(mvfr1, 12, 4) == 1) {
 129         set_feature(&features, ARM_FEATURE_NEON);
 130     }
 131     if (extract32(mvfr1, 28, 4) == 1) {
 132         /* FMAC support implies VFPv4 */
 133         set_feature(&features, ARM_FEATURE_VFP4);
 134     }
 135
 136     ahcc->features = features;
 137
 138     return true;
 139 }
 140
 141 static bool reg_syncs_via_tuple_list(uint64_t regidx)
 142 {
 143     /* Return true if the regidx is a register we should synchronize
 144      * via the cpreg_tuples array (ie is not a core reg we sync by
 145      * hand in kvm_arch_get/put_registers())
 146      */
 147     switch (regidx & KVM_REG_ARM_COPROC_MASK) {
 148     case KVM_REG_ARM_CORE:
 149     case KVM_REG_ARM_VFP:
 150         return false;
 151     default:
 152         return true;
 153     }
 154 }
 155
 156 static int compare_u64(const void *a, const void *b)
 157 {
 158     if (*(uint64_t *)a > *(uint64_t *)b) {
 159         return 1;
 160     }
 161     if (*(uint64_t *)a < *(uint64_t *)b) {
 162         return -1;
 163     }
 164     return 0;
 165 }
 166
 167 int kvm_arch_init_vcpu(CPUState *cs)
 168 {
 169     int i, ret, arraylen;
 170     uint64_t v;
 171     struct kvm_one_reg r;
 172     struct kvm_reg_list rl;
 173     struct kvm_reg_list *rlp;
 174     ARMCPU *cpu = ARM_CPU(cs);
 175
 176     if (cpu->kvm_target == QEMU_KVM_ARM_TARGET_NONE) {
 177         fprintf(stderr, "KVM is not supported for this guest CPU type\n");
 178         return -EINVAL;
 179     }
 180
 181     /* Determine init features for this CPU */
 182     memset(cpu->kvm_init_features, 0, sizeof(cpu->kvm_init_features));
 183     if (cpu->start_powered_off) {
 184         cpu->kvm_init_features[0] |= 1 << KVM_ARM_VCPU_POWER_OFF;
 185     }
 186     if (kvm_check_extension(cs->kvm_state, KVM_CAP_ARM_PSCI_0_2)) {
 187         cpu->psci_version = 2;
 188         cpu->kvm_init_features[0] |= 1 << KVM_ARM_VCPU_PSCI_0_2;
 189     }
 190
 191     /* Do KVM_ARM_VCPU_INIT ioctl */
 192     ret = kvm_arm_vcpu_init(cs);
 193     if (ret) {
 194         return ret;
 195     }
 196
 197     /* Query the kernel to make sure it supports 32 VFP
 198      * registers: QEMU's "cortex-a15" CPU is always a
 199      * VFP-D32 core. The simplest way to do this is just
 200      * to attempt to read register d31.
 201      */
 202     r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP | 31;
 203     r.addr = (uintptr_t)(&v);
 204     ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 205     if (ret == -ENOENT) {
 206         return -EINVAL;
 207     }
 208
 209     /* Populate the cpreg list based on the kernel's idea
 210      * of what registers exist (and throw away the TCG-created list).
 211      */
 212     rl.n = 0;
 213     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, &rl);
 214     if (ret != -E2BIG) {
 215         return ret;
 216     }
 217     rlp = g_malloc(sizeof(struct kvm_reg_list) + rl.n * sizeof(uint64_t));
 218     rlp->n = rl.n;
 219     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, rlp);
 220     if (ret) {
 221         goto out;
 222     }
 223     /* Sort the list we get back from the kernel, since cpreg_tuples
 224      * must be in strictly ascending order.
 225      */
 226     qsort(&rlp->reg, rlp->n, sizeof(rlp->reg[0]), compare_u64);
 227
 228     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 229         if (!reg_syncs_via_tuple_list(rlp->reg[i])) {
 230             continue;
 231         }
 232         switch (rlp->reg[i] & KVM_REG_SIZE_MASK) {
 233         case KVM_REG_SIZE_U32:
 234         case KVM_REG_SIZE_U64:
 235             break;
 236         default:
 237             fprintf(stderr, "Can't handle size of register in kernel list\n");
 238             ret = -EINVAL;
 239             goto out;
 240         }
 241
 242         arraylen++;
 243     }
 244
 245     cpu->cpreg_indexes = g_renew(uint64_t, cpu->cpreg_indexes, arraylen);
 246     cpu->cpreg_values = g_renew(uint64_t, cpu->cpreg_values, arraylen);
 247     cpu->cpreg_vmstate_indexes = g_renew(uint64_t, cpu->cpreg_vmstate_indexes,
 248                                          arraylen);
 249     cpu->cpreg_vmstate_values = g_renew(uint64_t, cpu->cpreg_vmstate_values,
 250                                         arraylen);
 251     cpu->cpreg_array_len = arraylen;
 252     cpu->cpreg_vmstate_array_len = arraylen;
 253
 254     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 255         uint64_t regidx = rlp->reg[i];
 256         if (!reg_syncs_via_tuple_list(regidx)) {
 257             continue;
 258         }
 259         cpu->cpreg_indexes[arraylen] = regidx;
 260         arraylen++;
 261     }
 262     assert(cpu->cpreg_array_len == arraylen);
 263
 264     if (!write_kvmstate_to_list(cpu)) {
 265         /* Shouldn't happen unless kernel is inconsistent about
 266          * what registers exist.
 267          */
 268         fprintf(stderr, "Initial read of kernel register state failed\n");
 269         ret = -EINVAL;
 270         goto out;
 271     }
 272
 273 out:
 274     g_free(rlp);
 275     return ret;
 276 }
 277
 278 typedef struct Reg {
 279     uint64_t id;
 280     int offset;
 281 } Reg;
 282
 283 #define COREREG(KERNELNAME, QEMUFIELD)                       \
 284     {                                                        \
 285         KVM_REG_ARM | KVM_REG_SIZE_U32 |                     \
 286         KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(KERNELNAME), \
 287         offsetof(CPUARMState, QEMUFIELD)                     \
 288     }
 289
 290 #define VFPSYSREG(R)                                       \
 291     {                                                      \
 292         KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP | \
 293         KVM_REG_ARM_VFP_##R,                               \
 294         offsetof(CPUARMState, vfp.xregs[ARM_VFP_##R])      \
 295     }
 296
 297 /* Like COREREG, but handle fields which are in a uint64_t in CPUARMState. */
 298 #define COREREG64(KERNELNAME, QEMUFIELD)                     \
 299     {                                                        \
 300         KVM_REG_ARM | KVM_REG_SIZE_U32 |                     \
 301         KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(KERNELNAME), \
 302         offsetoflow32(CPUARMState, QEMUFIELD)                \
 303     }
 304
 305 static const Reg regs[] = {
 306     /* R0_usr .. R14_usr */
 307     COREREG(usr_regs.uregs[0], regs[0]),
 308     COREREG(usr_regs.uregs[1], regs[1]),
 309     COREREG(usr_regs.uregs[2], regs[2]),
 310     COREREG(usr_regs.uregs[3], regs[3]),
 311     COREREG(usr_regs.uregs[4], regs[4]),
 312     COREREG(usr_regs.uregs[5], regs[5]),
 313     COREREG(usr_regs.uregs[6], regs[6]),
 314     COREREG(usr_regs.uregs[7], regs[7]),
 315     COREREG(usr_regs.uregs[8], usr_regs[0]),
 316     COREREG(usr_regs.uregs[9], usr_regs[1]),
 317     COREREG(usr_regs.uregs[10], usr_regs[2]),
 318     COREREG(usr_regs.uregs[11], usr_regs[3]),
 319     COREREG(usr_regs.uregs[12], usr_regs[4]),
 320     COREREG(usr_regs.uregs[13], banked_r13[0]),
 321     COREREG(usr_regs.uregs[14], banked_r14[0]),
 322     /* R13, R14, SPSR for SVC, ABT, UND, IRQ banks */
 323     COREREG(svc_regs[0], banked_r13[1]),
 324     COREREG(svc_regs[1], banked_r14[1]),
 325     COREREG64(svc_regs[2], banked_spsr[1]),
 326     COREREG(abt_regs[0], banked_r13[2]),
 327     COREREG(abt_regs[1], banked_r14[2]),
 328     COREREG64(abt_regs[2], banked_spsr[2]),
 329     COREREG(und_regs[0], banked_r13[3]),
 330     COREREG(und_regs[1], banked_r14[3]),
 331     COREREG64(und_regs[2], banked_spsr[3]),
 332     COREREG(irq_regs[0], banked_r13[4]),
 333     COREREG(irq_regs[1], banked_r14[4]),
 334     COREREG64(irq_regs[2], banked_spsr[4]),
 335     /* R8_fiq .. R14_fiq and SPSR_fiq */
 336     COREREG(fiq_regs[0], fiq_regs[0]),
 337     COREREG(fiq_regs[1], fiq_regs[1]),
 338     COREREG(fiq_regs[2], fiq_regs[2]),
 339     COREREG(fiq_regs[3], fiq_regs[3]),
 340     COREREG(fiq_regs[4], fiq_regs[4]),
 341     COREREG(fiq_regs[5], banked_r13[5]),
 342     COREREG(fiq_regs[6], banked_r14[5]),
 343     COREREG64(fiq_regs[7], banked_spsr[5]),
 344     /* R15 */
 345     COREREG(usr_regs.uregs[15], regs[15]),
 346     /* VFP system registers */
 347     VFPSYSREG(FPSID),
 348     VFPSYSREG(MVFR1),
 349     VFPSYSREG(MVFR0),
 350     VFPSYSREG(FPEXC),
 351     VFPSYSREG(FPINST),
 352     VFPSYSREG(FPINST2),
 353 };
 354
 355 int kvm_arch_put_registers(CPUState *cs, int level)
 356 {
 357     ARMCPU *cpu = ARM_CPU(cs);
 358     CPUARMState *env = &cpu->env;
 359     struct kvm_one_reg r;
 360     int mode, bn;
 361     int ret, i;
 362     uint32_t cpsr, fpscr;
 363
 364     /* Make sure the banked regs are properly set */
 365     mode = env->uncached_cpsr & CPSR_M;
 366     bn = bank_number(mode);
 367     if (mode == ARM_CPU_MODE_FIQ) {
 368         memcpy(env->fiq_regs, env->regs + 8, 5 * sizeof(uint32_t));
 369     } else {
 370         memcpy(env->usr_regs, env->regs + 8, 5 * sizeof(uint32_t));
 371     }
 372     env->banked_r13[bn] = env->regs[13];
 373     env->banked_r14[bn] = env->regs[14];
 374     env->banked_spsr[bn] = env->spsr;
 375
 376     /* Now we can safely copy stuff down to the kernel */
 377     for (i = 0; i < ARRAY_SIZE(regs); i++) {
 378         r.id = regs[i].id;
 379         r.addr = (uintptr_t)(env) + regs[i].offset;
 380         ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
 381         if (ret) {
 382             return ret;
 383         }
 384     }
 385
 386     /* Special cases which aren't a single CPUARMState field */
 387     cpsr = cpsr_read(env);
 388     r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 |
 389         KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(usr_regs.ARM_cpsr);
 390     r.addr = (uintptr_t)(&cpsr);
 391     ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
 392     if (ret) {
 393         return ret;
 394     }
 395
 396     /* VFP registers */
 397     r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP;
 398     for (i = 0; i < 32; i++) {
 399         r.addr = (uintptr_t)(&env->vfp.regs[i]);
 400         ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
 401         if (ret) {
 402             return ret;
 403         }
 404         r.id++;
 405     }
 406
 407     r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP |
 408         KVM_REG_ARM_VFP_FPSCR;
 409     fpscr = vfp_get_fpscr(env);
 410     r.addr = (uintptr_t)&fpscr;
 411     ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
 412     if (ret) {
 413         return ret;
 414     }
 415
 416     /* Note that we do not call write_cpustate_to_list()
 417      * here, so we are only writing the tuple list back to
 418      * KVM. This is safe because nothing can change the
 419      * CPUARMState cp15 fields (in particular gdb accesses cannot)
 420      * and so there are no changes to sync. In fact syncing would
 421      * be wrong at this point: for a constant register where TCG and
 422      * KVM disagree about its value, the preceding write_list_to_cpustate()
 423      * would not have had any effect on the CPUARMState value (since the
 424      * register is read-only), and a write_cpustate_to_list() here would
 425      * then try to write the TCG value back into KVM -- this would either
 426      * fail or incorrectly change the value the guest sees.
 427      *
 428      * If we ever want to allow the user to modify cp15 registers via
 429      * the gdb stub, we would need to be more clever here (for instance
 430      * tracking the set of registers kvm_arch_get_registers() successfully
 431      * managed to update the CPUARMState with, and only allowing those
 432      * to be written back up into the kernel).
 433      */
 434     if (!write_list_to_kvmstate(cpu)) {
 435         return EINVAL;
 436     }
 437
 438     return ret;
 439 }
 440
 441 int kvm_arch_get_registers(CPUState *cs)
 442 {
 443     ARMCPU *cpu = ARM_CPU(cs);
 444     CPUARMState *env = &cpu->env;
 445     struct kvm_one_reg r;
 446     int mode, bn;
 447     int ret, i;
 448     uint32_t cpsr, fpscr;
 449
 450     for (i = 0; i < ARRAY_SIZE(regs); i++) {
 451         r.id = regs[i].id;
 452         r.addr = (uintptr_t)(env) + regs[i].offset;
 453         ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 454         if (ret) {
 455             return ret;
 456         }
 457     }
 458
 459     /* Special cases which aren't a single CPUARMState field */
 460     r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 |
 461         KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(usr_regs.ARM_cpsr);
 462     r.addr = (uintptr_t)(&cpsr);
 463     ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 464     if (ret) {
 465         return ret;
 466     }
 467     cpsr_write(env, cpsr, 0xffffffff);
 468
 469     /* Make sure the current mode regs are properly set */
 470     mode = env->uncached_cpsr & CPSR_M;
 471     bn = bank_number(mode);
 472     if (mode == ARM_CPU_MODE_FIQ) {
 473         memcpy(env->regs + 8, env->fiq_regs, 5 * sizeof(uint32_t));
 474     } else {
 475         memcpy(env->regs + 8, env->usr_regs, 5 * sizeof(uint32_t));
 476     }
 477     env->regs[13] = env->banked_r13[bn];
 478     env->regs[14] = env->banked_r14[bn];
 479     env->spsr = env->banked_spsr[bn];
 480
 481     /* VFP registers */
 482     r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP;
 483     for (i = 0; i < 32; i++) {
 484         r.addr = (uintptr_t)(&env->vfp.regs[i]);
 485         ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 486         if (ret) {
 487             return ret;
 488         }
 489         r.id++;
 490     }
 491
 492     r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP |
 493         KVM_REG_ARM_VFP_FPSCR;
 494     r.addr = (uintptr_t)&fpscr;
 495     ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 496     if (ret) {
 497         return ret;
 498     }
 499     vfp_set_fpscr(env, fpscr);
 500
 501     if (!write_kvmstate_to_list(cpu)) {
 502         return EINVAL;
 503     }
 504     /* Note that it's OK to have registers which aren't in CPUState,
 505      * so we can ignore a failure return here.
 506      */
 507     write_list_to_cpustate(cpu);
 508
 509     return 0;
 510 }
 511
 512 void kvm_arm_reset_vcpu(ARMCPU *cpu)
 513 {
 514     /* Re-init VCPU so that all registers are set to
 515      * their respective reset values.
 516      */
 517     kvm_arm_vcpu_init(CPU(cpu));
 518     write_kvmstate_to_list(cpu);
 519 }