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 "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 "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
  26 static inline void set_feature(uint64_t *features, int feature)
  27 {
  28     *features |= 1ULL << feature;
  29 }
  30
  31 bool kvm_arm_get_host_cpu_features(ARMHostCPUClass *ahcc)
  32 {
  33     /* Identify the feature bits corresponding to the host CPU, and
  34      * fill out the ARMHostCPUClass fields accordingly. To do this
  35      * we have to create a scratch VM, create a single CPU inside it,
  36      * and then query that CPU for the relevant ID registers.
  37      */
  38     int i, ret, fdarray[3];
  39     uint32_t midr, id_pfr0, id_isar0, mvfr1;
  40     uint64_t features = 0;
  41     /* Old kernels may not know about the PREFERRED_TARGET ioctl: however
  42      * we know these will only support creating one kind of guest CPU,
  43      * which is its preferred CPU type.
  44      */
  45     static const uint32_t cpus_to_try[] = {
  46         QEMU_KVM_ARM_TARGET_CORTEX_A15,
  47         QEMU_KVM_ARM_TARGET_NONE
  48     };
  49     struct kvm_vcpu_init init;
  50     struct kvm_one_reg idregs[] = {
  51         {
  52             .id = KVM_REG_ARM | KVM_REG_SIZE_U32
  53             | ENCODE_CP_REG(15, 0, 0, 0, 0, 0, 0),
  54             .addr = (uintptr_t)&midr,
  55         },
  56         {
  57             .id = KVM_REG_ARM | KVM_REG_SIZE_U32
  58             | ENCODE_CP_REG(15, 0, 0, 0, 1, 0, 0),
  59             .addr = (uintptr_t)&id_pfr0,
  60         },
  61         {
  62             .id = KVM_REG_ARM | KVM_REG_SIZE_U32
  63             | ENCODE_CP_REG(15, 0, 0, 0, 2, 0, 0),
  64             .addr = (uintptr_t)&id_isar0,
  65         },
  66         {
  67             .id = KVM_REG_ARM | KVM_REG_SIZE_U32
  68             | KVM_REG_ARM_VFP | KVM_REG_ARM_VFP_MVFR1,
  69             .addr = (uintptr_t)&mvfr1,
  70         },
  71     };
  72
  73     if (!kvm_arm_create_scratch_host_vcpu(cpus_to_try, fdarray, &init)) {
  74         return false;
  75     }
  76
  77     ahcc->target = init.target;
  78
  79     /* This is not strictly blessed by the device tree binding docs yet,
  80      * but in practice the kernel does not care about this string so
  81      * there is no point maintaining an KVM_ARM_TARGET_* -> string table.
  82      */
  83     ahcc->dtb_compatible = "arm,arm-v7";
  84
  85     for (i = 0; i < ARRAY_SIZE(idregs); i++) {
  86         ret = ioctl(fdarray[2], KVM_GET_ONE_REG, &idregs[i]);
  87         if (ret) {
  88             break;
  89         }
  90     }
  91
  92     kvm_arm_destroy_scratch_host_vcpu(fdarray);
  93
  94     if (ret) {
  95         return false;
  96     }
  97
  98     /* Now we've retrieved all the register information we can
  99      * set the feature bits based on the ID register fields.
 100      * We can assume any KVM supporting CPU is at least a v7
 101      * with VFPv3, LPAE and the generic timers; this in turn implies
 102      * most of the other feature bits, but a few must be tested.
 103      */
 104     set_feature(&features, ARM_FEATURE_V7);
 105     set_feature(&features, ARM_FEATURE_VFP3);
 106     set_feature(&features, ARM_FEATURE_LPAE);
 107     set_feature(&features, ARM_FEATURE_GENERIC_TIMER);
 108
 109     switch (extract32(id_isar0, 24, 4)) {
 110     case 1:
 111         set_feature(&features, ARM_FEATURE_THUMB_DIV);
 112         break;
 113     case 2:
 114         set_feature(&features, ARM_FEATURE_ARM_DIV);
 115         set_feature(&features, ARM_FEATURE_THUMB_DIV);
 116         break;
 117     default:
 118         break;
 119     }
 120
 121     if (extract32(id_pfr0, 12, 4) == 1) {
 122         set_feature(&features, ARM_FEATURE_THUMB2EE);
 123     }
 124     if (extract32(mvfr1, 20, 4) == 1) {
 125         set_feature(&features, ARM_FEATURE_VFP_FP16);
 126     }
 127     if (extract32(mvfr1, 12, 4) == 1) {
 128         set_feature(&features, ARM_FEATURE_NEON);
 129     }
 130     if (extract32(mvfr1, 28, 4) == 1) {
 131         /* FMAC support implies VFPv4 */
 132         set_feature(&features, ARM_FEATURE_VFP4);
 133     }
 134
 135     ahcc->features = features;
 136
 137     return true;
 138 }
 139
 140 bool kvm_arm_reg_syncs_via_cpreg_list(uint64_t regidx)
 141 {
 142     /* Return true if the regidx is a register we should synchronize
 143      * via the cpreg_tuples array (ie is not a core reg we sync by
 144      * hand in kvm_arch_get/put_registers())
 145      */
 146     switch (regidx & KVM_REG_ARM_COPROC_MASK) {
 147     case KVM_REG_ARM_CORE:
 148     case KVM_REG_ARM_VFP:
 149         return false;
 150     default:
 151         return true;
 152     }
 153 }
 154
 155 typedef struct CPRegStateLevel {
 156     uint64_t regidx;
 157     int level;
 158 } CPRegStateLevel;
 159
 160 /* All coprocessor registers not listed in the following table are assumed to
 161  * be of the level KVM_PUT_RUNTIME_STATE. If a register should be written less
 162  * often, you must add it to this table with a state of either
 163  * KVM_PUT_RESET_STATE or KVM_PUT_FULL_STATE.
 164  */
 165 static const CPRegStateLevel non_runtime_cpregs[] = {
 166     { KVM_REG_ARM_TIMER_CNT, KVM_PUT_FULL_STATE },
 167 };
 168
 169 int kvm_arm_cpreg_level(uint64_t regidx)
 170 {
 171     int i;
 172
 173     for (i = 0; i < ARRAY_SIZE(non_runtime_cpregs); i++) {
 174         const CPRegStateLevel *l = &non_runtime_cpregs[i];
 175         if (l->regidx == regidx) {
 176             return l->level;
 177         }
 178     }
 179
 180     return KVM_PUT_RUNTIME_STATE;
 181 }
 182
 183 #define ARM_CPU_ID_MPIDR       0, 0, 0, 5
 184
 185 int kvm_arch_init_vcpu(CPUState *cs)
 186 {
 187     int ret;
 188     uint64_t v;
 189     uint32_t mpidr;
 190     struct kvm_one_reg r;
 191     ARMCPU *cpu = ARM_CPU(cs);
 192
 193     if (cpu->kvm_target == QEMU_KVM_ARM_TARGET_NONE) {
 194         fprintf(stderr, "KVM is not supported for this guest CPU type\n");
 195         return -EINVAL;
 196     }
 197
 198     /* Determine init features for this CPU */
 199     memset(cpu->kvm_init_features, 0, sizeof(cpu->kvm_init_features));
 200     if (cpu->start_powered_off) {
 201         cpu->kvm_init_features[0] |= 1 << KVM_ARM_VCPU_POWER_OFF;
 202     }
 203     if (kvm_check_extension(cs->kvm_state, KVM_CAP_ARM_PSCI_0_2)) {
 204         cpu->psci_version = 2;
 205         cpu->kvm_init_features[0] |= 1 << KVM_ARM_VCPU_PSCI_0_2;
 206     }
 207
 208     /* Do KVM_ARM_VCPU_INIT ioctl */
 209     ret = kvm_arm_vcpu_init(cs);
 210     if (ret) {
 211         return ret;
 212     }
 213
 214     /* Query the kernel to make sure it supports 32 VFP
 215      * registers: QEMU's "cortex-a15" CPU is always a
 216      * VFP-D32 core. The simplest way to do this is just
 217      * to attempt to read register d31.
 218      */
 219     r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP | 31;
 220     r.addr = (uintptr_t)(&v);
 221     ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 222     if (ret == -ENOENT) {
 223         return -EINVAL;
 224     }
 225
 226     /*
 227      * When KVM is in use, PSCI is emulated in-kernel and not by qemu.
 228      * Currently KVM has its own idea about MPIDR assignment, so we
 229      * override our defaults with what we get from KVM.
 230      */
 231     ret = kvm_get_one_reg(cs, ARM_CP15_REG32(ARM_CPU_ID_MPIDR), &mpidr);
 232     if (ret) {
 233         return ret;
 234     }
 235     cpu->mp_affinity = mpidr & ARM32_AFFINITY_MASK;
 236
 237     return kvm_arm_init_cpreg_list(cpu);
 238 }
 239
 240 typedef struct Reg {
 241     uint64_t id;
 242     int offset;
 243 } Reg;
 244
 245 #define COREREG(KERNELNAME, QEMUFIELD)                       \
 246     {                                                        \
 247         KVM_REG_ARM | KVM_REG_SIZE_U32 |                     \
 248         KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(KERNELNAME), \
 249         offsetof(CPUARMState, QEMUFIELD)                     \
 250     }
 251
 252 #define VFPSYSREG(R)                                       \
 253     {                                                      \
 254         KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP | \
 255         KVM_REG_ARM_VFP_##R,                               \
 256         offsetof(CPUARMState, vfp.xregs[ARM_VFP_##R])      \
 257     }
 258
 259 /* Like COREREG, but handle fields which are in a uint64_t in CPUARMState. */
 260 #define COREREG64(KERNELNAME, QEMUFIELD)                     \
 261     {                                                        \
 262         KVM_REG_ARM | KVM_REG_SIZE_U32 |                     \
 263         KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(KERNELNAME), \
 264         offsetoflow32(CPUARMState, QEMUFIELD)                \
 265     }
 266
 267 static const Reg regs[] = {
 268     /* R0_usr .. R14_usr */
 269     COREREG(usr_regs.uregs[0], regs[0]),
 270     COREREG(usr_regs.uregs[1], regs[1]),
 271     COREREG(usr_regs.uregs[2], regs[2]),
 272     COREREG(usr_regs.uregs[3], regs[3]),
 273     COREREG(usr_regs.uregs[4], regs[4]),
 274     COREREG(usr_regs.uregs[5], regs[5]),
 275     COREREG(usr_regs.uregs[6], regs[6]),
 276     COREREG(usr_regs.uregs[7], regs[7]),
 277     COREREG(usr_regs.uregs[8], usr_regs[0]),
 278     COREREG(usr_regs.uregs[9], usr_regs[1]),
 279     COREREG(usr_regs.uregs[10], usr_regs[2]),
 280     COREREG(usr_regs.uregs[11], usr_regs[3]),
 281     COREREG(usr_regs.uregs[12], usr_regs[4]),
 282     COREREG(usr_regs.uregs[13], banked_r13[BANK_USRSYS]),
 283     COREREG(usr_regs.uregs[14], banked_r14[BANK_USRSYS]),
 284     /* R13, R14, SPSR for SVC, ABT, UND, IRQ banks */
 285     COREREG(svc_regs[0], banked_r13[BANK_SVC]),
 286     COREREG(svc_regs[1], banked_r14[BANK_SVC]),
 287     COREREG64(svc_regs[2], banked_spsr[BANK_SVC]),
 288     COREREG(abt_regs[0], banked_r13[BANK_ABT]),
 289     COREREG(abt_regs[1], banked_r14[BANK_ABT]),
 290     COREREG64(abt_regs[2], banked_spsr[BANK_ABT]),
 291     COREREG(und_regs[0], banked_r13[BANK_UND]),
 292     COREREG(und_regs[1], banked_r14[BANK_UND]),
 293     COREREG64(und_regs[2], banked_spsr[BANK_UND]),
 294     COREREG(irq_regs[0], banked_r13[BANK_IRQ]),
 295     COREREG(irq_regs[1], banked_r14[BANK_IRQ]),
 296     COREREG64(irq_regs[2], banked_spsr[BANK_IRQ]),
 297     /* R8_fiq .. R14_fiq and SPSR_fiq */
 298     COREREG(fiq_regs[0], fiq_regs[0]),
 299     COREREG(fiq_regs[1], fiq_regs[1]),
 300     COREREG(fiq_regs[2], fiq_regs[2]),
 301     COREREG(fiq_regs[3], fiq_regs[3]),
 302     COREREG(fiq_regs[4], fiq_regs[4]),
 303     COREREG(fiq_regs[5], banked_r13[BANK_FIQ]),
 304     COREREG(fiq_regs[6], banked_r14[BANK_FIQ]),
 305     COREREG64(fiq_regs[7], banked_spsr[BANK_FIQ]),
 306     /* R15 */
 307     COREREG(usr_regs.uregs[15], regs[15]),
 308     /* VFP system registers */
 309     VFPSYSREG(FPSID),
 310     VFPSYSREG(MVFR1),
 311     VFPSYSREG(MVFR0),
 312     VFPSYSREG(FPEXC),
 313     VFPSYSREG(FPINST),
 314     VFPSYSREG(FPINST2),
 315 };
 316
 317 int kvm_arch_put_registers(CPUState *cs, int level)
 318 {
 319     ARMCPU *cpu = ARM_CPU(cs);
 320     CPUARMState *env = &cpu->env;
 321     struct kvm_one_reg r;
 322     int mode, bn;
 323     int ret, i;
 324     uint32_t cpsr, fpscr;
 325
 326     /* Make sure the banked regs are properly set */
 327     mode = env->uncached_cpsr & CPSR_M;
 328     bn = bank_number(mode);
 329     if (mode == ARM_CPU_MODE_FIQ) {
 330         memcpy(env->fiq_regs, env->regs + 8, 5 * sizeof(uint32_t));
 331     } else {
 332         memcpy(env->usr_regs, env->regs + 8, 5 * sizeof(uint32_t));
 333     }
 334     env->banked_r13[bn] = env->regs[13];
 335     env->banked_r14[bn] = env->regs[14];
 336     env->banked_spsr[bn] = env->spsr;
 337
 338     /* Now we can safely copy stuff down to the kernel */
 339     for (i = 0; i < ARRAY_SIZE(regs); i++) {
 340         r.id = regs[i].id;
 341         r.addr = (uintptr_t)(env) + regs[i].offset;
 342         ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
 343         if (ret) {
 344             return ret;
 345         }
 346     }
 347
 348     /* Special cases which aren't a single CPUARMState field */
 349     cpsr = cpsr_read(env);
 350     r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 |
 351         KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(usr_regs.ARM_cpsr);
 352     r.addr = (uintptr_t)(&cpsr);
 353     ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
 354     if (ret) {
 355         return ret;
 356     }
 357
 358     /* VFP registers */
 359     r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP;
 360     for (i = 0; i < 32; i++) {
 361         r.addr = (uintptr_t)(&env->vfp.regs[i]);
 362         ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
 363         if (ret) {
 364             return ret;
 365         }
 366         r.id++;
 367     }
 368
 369     r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP |
 370         KVM_REG_ARM_VFP_FPSCR;
 371     fpscr = vfp_get_fpscr(env);
 372     r.addr = (uintptr_t)&fpscr;
 373     ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
 374     if (ret) {
 375         return ret;
 376     }
 377
 378     /* Note that we do not call write_cpustate_to_list()
 379      * here, so we are only writing the tuple list back to
 380      * KVM. This is safe because nothing can change the
 381      * CPUARMState cp15 fields (in particular gdb accesses cannot)
 382      * and so there are no changes to sync. In fact syncing would
 383      * be wrong at this point: for a constant register where TCG and
 384      * KVM disagree about its value, the preceding write_list_to_cpustate()
 385      * would not have had any effect on the CPUARMState value (since the
 386      * register is read-only), and a write_cpustate_to_list() here would
 387      * then try to write the TCG value back into KVM -- this would either
 388      * fail or incorrectly change the value the guest sees.
 389      *
 390      * If we ever want to allow the user to modify cp15 registers via
 391      * the gdb stub, we would need to be more clever here (for instance
 392      * tracking the set of registers kvm_arch_get_registers() successfully
 393      * managed to update the CPUARMState with, and only allowing those
 394      * to be written back up into the kernel).
 395      */
 396     if (!write_list_to_kvmstate(cpu, level)) {
 397         return EINVAL;
 398     }
 399
 400     kvm_arm_sync_mpstate_to_kvm(cpu);
 401
 402     return ret;
 403 }
 404
 405 int kvm_arch_get_registers(CPUState *cs)
 406 {
 407     ARMCPU *cpu = ARM_CPU(cs);
 408     CPUARMState *env = &cpu->env;
 409     struct kvm_one_reg r;
 410     int mode, bn;
 411     int ret, i;
 412     uint32_t cpsr, fpscr;
 413
 414     for (i = 0; i < ARRAY_SIZE(regs); i++) {
 415         r.id = regs[i].id;
 416         r.addr = (uintptr_t)(env) + regs[i].offset;
 417         ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 418         if (ret) {
 419             return ret;
 420         }
 421     }
 422
 423     /* Special cases which aren't a single CPUARMState field */
 424     r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 |
 425         KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(usr_regs.ARM_cpsr);
 426     r.addr = (uintptr_t)(&cpsr);
 427     ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 428     if (ret) {
 429         return ret;
 430     }
 431     cpsr_write(env, cpsr, 0xffffffff, CPSRWriteRaw);
 432
 433     /* Make sure the current mode regs are properly set */
 434     mode = env->uncached_cpsr & CPSR_M;
 435     bn = bank_number(mode);
 436     if (mode == ARM_CPU_MODE_FIQ) {
 437         memcpy(env->regs + 8, env->fiq_regs, 5 * sizeof(uint32_t));
 438     } else {
 439         memcpy(env->regs + 8, env->usr_regs, 5 * sizeof(uint32_t));
 440     }
 441     env->regs[13] = env->banked_r13[bn];
 442     env->regs[14] = env->banked_r14[bn];
 443     env->spsr = env->banked_spsr[bn];
 444
 445     /* VFP registers */
 446     r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP;
 447     for (i = 0; i < 32; i++) {
 448         r.addr = (uintptr_t)(&env->vfp.regs[i]);
 449         ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 450         if (ret) {
 451             return ret;
 452         }
 453         r.id++;
 454     }
 455
 456     r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP |
 457         KVM_REG_ARM_VFP_FPSCR;
 458     r.addr = (uintptr_t)&fpscr;
 459     ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 460     if (ret) {
 461         return ret;
 462     }
 463     vfp_set_fpscr(env, fpscr);
 464
 465     if (!write_kvmstate_to_list(cpu)) {
 466         return EINVAL;
 467     }
 468     /* Note that it's OK to have registers which aren't in CPUState,
 469      * so we can ignore a failure return here.
 470      */
 471     write_list_to_cpustate(cpu);
 472
 473     kvm_arm_sync_mpstate_to_qemu(cpu);
 474
 475     return 0;
 476 }
 477
 478 int kvm_arch_insert_sw_breakpoint(CPUState *cs, struct kvm_sw_breakpoint *bp)
 479 {
 480     qemu_log_mask(LOG_UNIMP, "%s: guest debug not yet implemented\n", __func__);
 481     return -EINVAL;
 482 }
 483
 484 int kvm_arch_remove_sw_breakpoint(CPUState *cs, struct kvm_sw_breakpoint *bp)
 485 {
 486     qemu_log_mask(LOG_UNIMP, "%s: guest debug not yet implemented\n", __func__);
 487     return -EINVAL;
 488 }
 489
 490 bool kvm_arm_handle_debug(CPUState *cs, struct kvm_debug_exit_arch *debug_exit)
 491 {
 492     qemu_log_mask(LOG_UNIMP, "%s: guest debug not yet implemented\n", __func__);
 493     return false;
 494 }
 495
 496 int kvm_arch_insert_hw_breakpoint(target_ulong addr,
 497                                   target_ulong len, int type)
 498 {
 499     qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
 500     return -EINVAL;
 501 }
 502
 503 int kvm_arch_remove_hw_breakpoint(target_ulong addr,
 504                                   target_ulong len, int type)
 505 {
 506     qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
 507     return -EINVAL;
 508 }
 509
 510 void kvm_arch_remove_all_hw_breakpoints(void)
 511 {
 512     qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
 513 }
 514
 515 void kvm_arm_copy_hw_debug_data(struct kvm_guest_debug_arch *ptr)
 516 {
 517     qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
 518 }
 519
 520 bool kvm_arm_hw_debug_active(CPUState *cs)
 521 {
 522     return false;
 523 }