hw/hyperv/hyperv.c

   1 /*
   2  * Hyper-V guest/hypervisor interaction
   3  *
   4  * Copyright (c) 2015-2018 Virtuozzo International GmbH.
   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 #include "qemu/osdep.h"
  11 #include "qemu/main-loop.h"
  12 #include "qemu/module.h"
  13 #include "qapi/error.h"
  14 #include "exec/address-spaces.h"
  15 #include "sysemu/kvm.h"
  16 #include "qemu/bitops.h"
  17 #include "qemu/error-report.h"
  18 #include "qemu/lockable.h"
  19 #include "qemu/queue.h"
  20 #include "qemu/rcu.h"
  21 #include "qemu/rcu_queue.h"
  22 #include "hw/hyperv/hyperv.h"
  23
  24 typedef struct SynICState {
  25     DeviceState parent_obj;
  26
  27     CPUState *cs;
  28
  29     bool enabled;
  30     hwaddr msg_page_addr;
  31     hwaddr event_page_addr;
  32     MemoryRegion msg_page_mr;
  33     MemoryRegion event_page_mr;
  34     struct hyperv_message_page *msg_page;
  35     struct hyperv_event_flags_page *event_page;
  36 } SynICState;
  37
  38 #define TYPE_SYNIC "hyperv-synic"
  39 #define SYNIC(obj) OBJECT_CHECK(SynICState, (obj), TYPE_SYNIC)
  40
  41 static SynICState *get_synic(CPUState *cs)
  42 {
  43     return SYNIC(object_resolve_path_component(OBJECT(cs), "synic"));
  44 }
  45
  46 static void synic_update(SynICState *synic, bool enable,
  47                          hwaddr msg_page_addr, hwaddr event_page_addr)
  48 {
  49
  50     synic->enabled = enable;
  51     if (synic->msg_page_addr != msg_page_addr) {
  52         if (synic->msg_page_addr) {
  53             memory_region_del_subregion(get_system_memory(),
  54                                         &synic->msg_page_mr);
  55         }
  56         if (msg_page_addr) {
  57             memory_region_add_subregion(get_system_memory(), msg_page_addr,
  58                                         &synic->msg_page_mr);
  59         }
  60         synic->msg_page_addr = msg_page_addr;
  61     }
  62     if (synic->event_page_addr != event_page_addr) {
  63         if (synic->event_page_addr) {
  64             memory_region_del_subregion(get_system_memory(),
  65                                         &synic->event_page_mr);
  66         }
  67         if (event_page_addr) {
  68             memory_region_add_subregion(get_system_memory(), event_page_addr,
  69                                         &synic->event_page_mr);
  70         }
  71         synic->event_page_addr = event_page_addr;
  72     }
  73 }
  74
  75 void hyperv_synic_update(CPUState *cs, bool enable,
  76                          hwaddr msg_page_addr, hwaddr event_page_addr)
  77 {
  78     SynICState *synic = get_synic(cs);
  79
  80     if (!synic) {
  81         return;
  82     }
  83
  84     synic_update(synic, enable, msg_page_addr, event_page_addr);
  85 }
  86
  87 static void synic_realize(DeviceState *dev, Error **errp)
  88 {
  89     Object *obj = OBJECT(dev);
  90     SynICState *synic = SYNIC(dev);
  91     char *msgp_name, *eventp_name;
  92     uint32_t vp_index;
  93
  94     /* memory region names have to be globally unique */
  95     vp_index = hyperv_vp_index(synic->cs);
  96     msgp_name = g_strdup_printf("synic-%u-msg-page", vp_index);
  97     eventp_name = g_strdup_printf("synic-%u-event-page", vp_index);
  98
  99     memory_region_init_ram(&synic->msg_page_mr, obj, msgp_name,
 100                            sizeof(*synic->msg_page), &error_abort);
 101     memory_region_init_ram(&synic->event_page_mr, obj, eventp_name,
 102                            sizeof(*synic->event_page), &error_abort);
 103     synic->msg_page = memory_region_get_ram_ptr(&synic->msg_page_mr);
 104     synic->event_page = memory_region_get_ram_ptr(&synic->event_page_mr);
 105
 106     g_free(msgp_name);
 107     g_free(eventp_name);
 108 }
 109 static void synic_reset(DeviceState *dev)
 110 {
 111     SynICState *synic = SYNIC(dev);
 112     memset(synic->msg_page, 0, sizeof(*synic->msg_page));
 113     memset(synic->event_page, 0, sizeof(*synic->event_page));
 114     synic_update(synic, false, 0, 0);
 115 }
 116
 117 static void synic_class_init(ObjectClass *klass, void *data)
 118 {
 119     DeviceClass *dc = DEVICE_CLASS(klass);
 120
 121     dc->realize = synic_realize;
 122     dc->reset = synic_reset;
 123     dc->user_creatable = false;
 124 }
 125
 126 void hyperv_synic_add(CPUState *cs)
 127 {
 128     Object *obj;
 129     SynICState *synic;
 130
 131     obj = object_new(TYPE_SYNIC);
 132     synic = SYNIC(obj);
 133     synic->cs = cs;
 134     object_property_add_child(OBJECT(cs), "synic", obj, &error_abort);
 135     object_unref(obj);
 136     object_property_set_bool(obj, true, "realized", &error_abort);
 137 }
 138
 139 void hyperv_synic_reset(CPUState *cs)
 140 {
 141     SynICState *synic = get_synic(cs);
 142
 143     if (synic) {
 144         device_legacy_reset(DEVICE(synic));
 145     }
 146 }
 147
 148 static const TypeInfo synic_type_info = {
 149     .name = TYPE_SYNIC,
 150     .parent = TYPE_DEVICE,
 151     .instance_size = sizeof(SynICState),
 152     .class_init = synic_class_init,
 153 };
 154
 155 static void synic_register_types(void)
 156 {
 157     type_register_static(&synic_type_info);
 158 }
 159
 160 type_init(synic_register_types)
 161
 162 /*
 163  * KVM has its own message producers (SynIC timers).  To guarantee
 164  * serialization with both KVM vcpu and the guest cpu, the messages are first
 165  * staged in an intermediate area and then posted to the SynIC message page in
 166  * the vcpu thread.
 167  */
 168 typedef struct HvSintStagedMessage {
 169     /* message content staged by hyperv_post_msg */
 170     struct hyperv_message msg;
 171     /* callback + data (r/o) to complete the processing in a BH */
 172     HvSintMsgCb cb;
 173     void *cb_data;
 174     /* message posting status filled by cpu_post_msg */
 175     int status;
 176     /* passing the buck: */
 177     enum {
 178         /* initial state */
 179         HV_STAGED_MSG_FREE,
 180         /*
 181          * hyperv_post_msg (e.g. in main loop) grabs the staged area (FREE ->
 182          * BUSY), copies msg, and schedules cpu_post_msg on the assigned cpu
 183          */
 184         HV_STAGED_MSG_BUSY,
 185         /*
 186          * cpu_post_msg (vcpu thread) tries to copy staged msg to msg slot,
 187          * notify the guest, records the status, marks the posting done (BUSY
 188          * -> POSTED), and schedules sint_msg_bh BH
 189          */
 190         HV_STAGED_MSG_POSTED,
 191         /*
 192          * sint_msg_bh (BH) verifies that the posting is done, runs the
 193          * callback, and starts over (POSTED -> FREE)
 194          */
 195     } state;
 196 } HvSintStagedMessage;
 197
 198 struct HvSintRoute {
 199     uint32_t sint;
 200     SynICState *synic;
 201     int gsi;
 202     EventNotifier sint_set_notifier;
 203     EventNotifier sint_ack_notifier;
 204
 205     HvSintStagedMessage *staged_msg;
 206
 207     unsigned refcount;
 208 };
 209
 210 static CPUState *hyperv_find_vcpu(uint32_t vp_index)
 211 {
 212     CPUState *cs = qemu_get_cpu(vp_index);
 213     assert(hyperv_vp_index(cs) == vp_index);
 214     return cs;
 215 }
 216
 217 /*
 218  * BH to complete the processing of a staged message.
 219  */
 220 static void sint_msg_bh(void *opaque)
 221 {
 222     HvSintRoute *sint_route = opaque;
 223     HvSintStagedMessage *staged_msg = sint_route->staged_msg;
 224
 225     if (atomic_read(&staged_msg->state) != HV_STAGED_MSG_POSTED) {
 226         /* status nor ready yet (spurious ack from guest?), ignore */
 227         return;
 228     }
 229
 230     staged_msg->cb(staged_msg->cb_data, staged_msg->status);
 231     staged_msg->status = 0;
 232
 233     /* staged message processing finished, ready to start over */
 234     atomic_set(&staged_msg->state, HV_STAGED_MSG_FREE);
 235     /* drop the reference taken in hyperv_post_msg */
 236     hyperv_sint_route_unref(sint_route);
 237 }
 238
 239 /*
 240  * Worker to transfer the message from the staging area into the SynIC message
 241  * page in vcpu context.
 242  */
 243 static void cpu_post_msg(CPUState *cs, run_on_cpu_data data)
 244 {
 245     HvSintRoute *sint_route = data.host_ptr;
 246     HvSintStagedMessage *staged_msg = sint_route->staged_msg;
 247     SynICState *synic = sint_route->synic;
 248     struct hyperv_message *dst_msg;
 249     bool wait_for_sint_ack = false;
 250
 251     assert(staged_msg->state == HV_STAGED_MSG_BUSY);
 252
 253     if (!synic->enabled || !synic->msg_page_addr) {
 254         staged_msg->status = -ENXIO;
 255         goto posted;
 256     }
 257
 258     dst_msg = &synic->msg_page->slot[sint_route->sint];
 259
 260     if (dst_msg->header.message_type != HV_MESSAGE_NONE) {
 261         dst_msg->header.message_flags |= HV_MESSAGE_FLAG_PENDING;
 262         staged_msg->status = -EAGAIN;
 263         wait_for_sint_ack = true;
 264     } else {
 265         memcpy(dst_msg, &staged_msg->msg, sizeof(*dst_msg));
 266         staged_msg->status = hyperv_sint_route_set_sint(sint_route);
 267     }
 268
 269     memory_region_set_dirty(&synic->msg_page_mr, 0, sizeof(*synic->msg_page));
 270
 271 posted:
 272     atomic_set(&staged_msg->state, HV_STAGED_MSG_POSTED);
 273     /*
 274      * Notify the msg originator of the progress made; if the slot was busy we
 275      * set msg_pending flag in it so it will be the guest who will do EOM and
 276      * trigger the notification from KVM via sint_ack_notifier
 277      */
 278     if (!wait_for_sint_ack) {
 279         aio_bh_schedule_oneshot(qemu_get_aio_context(), sint_msg_bh,
 280                                 sint_route);
 281     }
 282 }
 283
 284 /*
 285  * Post a Hyper-V message to the staging area, for delivery to guest in the
 286  * vcpu thread.
 287  */
 288 int hyperv_post_msg(HvSintRoute *sint_route, struct hyperv_message *src_msg)
 289 {
 290     HvSintStagedMessage *staged_msg = sint_route->staged_msg;
 291
 292     assert(staged_msg);
 293
 294     /* grab the staging area */
 295     if (atomic_cmpxchg(&staged_msg->state, HV_STAGED_MSG_FREE,
 296                        HV_STAGED_MSG_BUSY) != HV_STAGED_MSG_FREE) {
 297         return -EAGAIN;
 298     }
 299
 300     memcpy(&staged_msg->msg, src_msg, sizeof(*src_msg));
 301
 302     /* hold a reference on sint_route until the callback is finished */
 303     hyperv_sint_route_ref(sint_route);
 304
 305     /* schedule message posting attempt in vcpu thread */
 306     async_run_on_cpu(sint_route->synic->cs, cpu_post_msg,
 307                      RUN_ON_CPU_HOST_PTR(sint_route));
 308     return 0;
 309 }
 310
 311 static void sint_ack_handler(EventNotifier *notifier)
 312 {
 313     HvSintRoute *sint_route = container_of(notifier, HvSintRoute,
 314                                            sint_ack_notifier);
 315     event_notifier_test_and_clear(notifier);
 316
 317     /*
 318      * the guest consumed the previous message so complete the current one with
 319      * -EAGAIN and let the msg originator retry
 320      */
 321     aio_bh_schedule_oneshot(qemu_get_aio_context(), sint_msg_bh, sint_route);
 322 }
 323
 324 /*
 325  * Set given event flag for a given sint on a given vcpu, and signal the sint.
 326  */
 327 int hyperv_set_event_flag(HvSintRoute *sint_route, unsigned eventno)
 328 {
 329     int ret;
 330     SynICState *synic = sint_route->synic;
 331     unsigned long *flags, set_mask;
 332     unsigned set_idx;
 333
 334     if (eventno > HV_EVENT_FLAGS_COUNT) {
 335         return -EINVAL;
 336     }
 337     if (!synic->enabled || !synic->event_page_addr) {
 338         return -ENXIO;
 339     }
 340
 341     set_idx = BIT_WORD(eventno);
 342     set_mask = BIT_MASK(eventno);
 343     flags = synic->event_page->slot[sint_route->sint].flags;
 344
 345     if ((atomic_fetch_or(&flags[set_idx], set_mask) & set_mask) != set_mask) {
 346         memory_region_set_dirty(&synic->event_page_mr, 0,
 347                                 sizeof(*synic->event_page));
 348         ret = hyperv_sint_route_set_sint(sint_route);
 349     } else {
 350         ret = 0;
 351     }
 352     return ret;
 353 }
 354
 355 HvSintRoute *hyperv_sint_route_new(uint32_t vp_index, uint32_t sint,
 356                                    HvSintMsgCb cb, void *cb_data)
 357 {
 358     HvSintRoute *sint_route;
 359     EventNotifier *ack_notifier;
 360     int r, gsi;
 361     CPUState *cs;
 362     SynICState *synic;
 363
 364     cs = hyperv_find_vcpu(vp_index);
 365     if (!cs) {
 366         return NULL;
 367     }
 368
 369     synic = get_synic(cs);
 370     if (!synic) {
 371         return NULL;
 372     }
 373
 374     sint_route = g_new0(HvSintRoute, 1);
 375     r = event_notifier_init(&sint_route->sint_set_notifier, false);
 376     if (r) {
 377         goto err;
 378     }
 379
 380
 381     ack_notifier = cb ? &sint_route->sint_ack_notifier : NULL;
 382     if (ack_notifier) {
 383         sint_route->staged_msg = g_new0(HvSintStagedMessage, 1);
 384         sint_route->staged_msg->cb = cb;
 385         sint_route->staged_msg->cb_data = cb_data;
 386
 387         r = event_notifier_init(ack_notifier, false);
 388         if (r) {
 389             goto err_sint_set_notifier;
 390         }
 391
 392         event_notifier_set_handler(ack_notifier, sint_ack_handler);
 393     }
 394
 395     gsi = kvm_irqchip_add_hv_sint_route(kvm_state, vp_index, sint);
 396     if (gsi < 0) {
 397         goto err_gsi;
 398     }
 399
 400     r = kvm_irqchip_add_irqfd_notifier_gsi(kvm_state,
 401                                            &sint_route->sint_set_notifier,
 402                                            ack_notifier, gsi);
 403     if (r) {
 404         goto err_irqfd;
 405     }
 406     sint_route->gsi = gsi;
 407     sint_route->synic = synic;
 408     sint_route->sint = sint;
 409     sint_route->refcount = 1;
 410
 411     return sint_route;
 412
 413 err_irqfd:
 414     kvm_irqchip_release_virq(kvm_state, gsi);
 415 err_gsi:
 416     if (ack_notifier) {
 417         event_notifier_set_handler(ack_notifier, NULL);
 418         event_notifier_cleanup(ack_notifier);
 419         g_free(sint_route->staged_msg);
 420     }
 421 err_sint_set_notifier:
 422     event_notifier_cleanup(&sint_route->sint_set_notifier);
 423 err:
 424     g_free(sint_route);
 425
 426     return NULL;
 427 }
 428
 429 void hyperv_sint_route_ref(HvSintRoute *sint_route)
 430 {
 431     sint_route->refcount++;
 432 }
 433
 434 void hyperv_sint_route_unref(HvSintRoute *sint_route)
 435 {
 436     if (!sint_route) {
 437         return;
 438     }
 439
 440     assert(sint_route->refcount > 0);
 441
 442     if (--sint_route->refcount) {
 443         return;
 444     }
 445
 446     kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state,
 447                                           &sint_route->sint_set_notifier,
 448                                           sint_route->gsi);
 449     kvm_irqchip_release_virq(kvm_state, sint_route->gsi);
 450     if (sint_route->staged_msg) {
 451         event_notifier_set_handler(&sint_route->sint_ack_notifier, NULL);
 452         event_notifier_cleanup(&sint_route->sint_ack_notifier);
 453         g_free(sint_route->staged_msg);
 454     }
 455     event_notifier_cleanup(&sint_route->sint_set_notifier);
 456     g_free(sint_route);
 457 }
 458
 459 int hyperv_sint_route_set_sint(HvSintRoute *sint_route)
 460 {
 461     return event_notifier_set(&sint_route->sint_set_notifier);
 462 }
 463
 464 typedef struct MsgHandler {
 465     struct rcu_head rcu;
 466     QLIST_ENTRY(MsgHandler) link;
 467     uint32_t conn_id;
 468     HvMsgHandler handler;
 469     void *data;
 470 } MsgHandler;
 471
 472 typedef struct EventFlagHandler {
 473     struct rcu_head rcu;
 474     QLIST_ENTRY(EventFlagHandler) link;
 475     uint32_t conn_id;
 476     EventNotifier *notifier;
 477 } EventFlagHandler;
 478
 479 static QLIST_HEAD(, MsgHandler) msg_handlers;
 480 static QLIST_HEAD(, EventFlagHandler) event_flag_handlers;
 481 static QemuMutex handlers_mutex;
 482
 483 static void __attribute__((constructor)) hv_init(void)
 484 {
 485     QLIST_INIT(&msg_handlers);
 486     QLIST_INIT(&event_flag_handlers);
 487     qemu_mutex_init(&handlers_mutex);
 488 }
 489
 490 int hyperv_set_msg_handler(uint32_t conn_id, HvMsgHandler handler, void *data)
 491 {
 492     int ret;
 493     MsgHandler *mh;
 494
 495     QEMU_LOCK_GUARD(&handlers_mutex);
 496     QLIST_FOREACH(mh, &msg_handlers, link) {
 497         if (mh->conn_id == conn_id) {
 498             if (handler) {
 499                 ret = -EEXIST;
 500             } else {
 501                 QLIST_REMOVE_RCU(mh, link);
 502                 g_free_rcu(mh, rcu);
 503                 ret = 0;
 504             }
 505             return ret;
 506         }
 507     }
 508
 509     if (handler) {
 510         mh = g_new(MsgHandler, 1);
 511         mh->conn_id = conn_id;
 512         mh->handler = handler;
 513         mh->data = data;
 514         QLIST_INSERT_HEAD_RCU(&msg_handlers, mh, link);
 515         ret = 0;
 516     } else {
 517         ret = -ENOENT;
 518     }
 519
 520     return ret;
 521 }
 522
 523 uint16_t hyperv_hcall_post_message(uint64_t param, bool fast)
 524 {
 525     uint16_t ret;
 526     hwaddr len;
 527     struct hyperv_post_message_input *msg;
 528     MsgHandler *mh;
 529
 530     if (fast) {
 531         return HV_STATUS_INVALID_HYPERCALL_CODE;
 532     }
 533     if (param & (__alignof__(*msg) - 1)) {
 534         return HV_STATUS_INVALID_ALIGNMENT;
 535     }
 536
 537     len = sizeof(*msg);
 538     msg = cpu_physical_memory_map(param, &len, 0);
 539     if (len < sizeof(*msg)) {
 540         ret = HV_STATUS_INSUFFICIENT_MEMORY;
 541         goto unmap;
 542     }
 543     if (msg->payload_size > sizeof(msg->payload)) {
 544         ret = HV_STATUS_INVALID_HYPERCALL_INPUT;
 545         goto unmap;
 546     }
 547
 548     ret = HV_STATUS_INVALID_CONNECTION_ID;
 549     WITH_RCU_READ_LOCK_GUARD() {
 550         QLIST_FOREACH_RCU(mh, &msg_handlers, link) {
 551             if (mh->conn_id == (msg->connection_id & HV_CONNECTION_ID_MASK)) {
 552                 ret = mh->handler(msg, mh->data);
 553                 break;
 554             }
 555         }
 556     }
 557
 558 unmap:
 559     cpu_physical_memory_unmap(msg, len, 0, 0);
 560     return ret;
 561 }
 562
 563 static int set_event_flag_handler(uint32_t conn_id, EventNotifier *notifier)
 564 {
 565     int ret;
 566     EventFlagHandler *handler;
 567
 568     QEMU_LOCK_GUARD(&handlers_mutex);
 569     QLIST_FOREACH(handler, &event_flag_handlers, link) {
 570         if (handler->conn_id == conn_id) {
 571             if (notifier) {
 572                 ret = -EEXIST;
 573             } else {
 574                 QLIST_REMOVE_RCU(handler, link);
 575                 g_free_rcu(handler, rcu);
 576                 ret = 0;
 577             }
 578             return ret;
 579         }
 580     }
 581
 582     if (notifier) {
 583         handler = g_new(EventFlagHandler, 1);
 584         handler->conn_id = conn_id;
 585         handler->notifier = notifier;
 586         QLIST_INSERT_HEAD_RCU(&event_flag_handlers, handler, link);
 587         ret = 0;
 588     } else {
 589         ret = -ENOENT;
 590     }
 591
 592     return ret;
 593 }
 594
 595 static bool process_event_flags_userspace;
 596
 597 int hyperv_set_event_flag_handler(uint32_t conn_id, EventNotifier *notifier)
 598 {
 599     if (!process_event_flags_userspace &&
 600         !kvm_check_extension(kvm_state, KVM_CAP_HYPERV_EVENTFD)) {
 601         process_event_flags_userspace = true;
 602
 603         warn_report("Hyper-V event signaling is not supported by this kernel; "
 604                     "using slower userspace hypercall processing");
 605     }
 606
 607     if (!process_event_flags_userspace) {
 608         struct kvm_hyperv_eventfd hvevfd = {
 609             .conn_id = conn_id,
 610             .fd = notifier ? event_notifier_get_fd(notifier) : -1,
 611             .flags = notifier ? 0 : KVM_HYPERV_EVENTFD_DEASSIGN,
 612         };
 613
 614         return kvm_vm_ioctl(kvm_state, KVM_HYPERV_EVENTFD, &hvevfd);
 615     }
 616     return set_event_flag_handler(conn_id, notifier);
 617 }
 618
 619 uint16_t hyperv_hcall_signal_event(uint64_t param, bool fast)
 620 {
 621     EventFlagHandler *handler;
 622
 623     if (unlikely(!fast)) {
 624         hwaddr addr = param;
 625
 626         if (addr & (__alignof__(addr) - 1)) {
 627             return HV_STATUS_INVALID_ALIGNMENT;
 628         }
 629
 630         param = ldq_phys(&address_space_memory, addr);
 631     }
 632
 633     /*
 634      * Per spec, bits 32-47 contain the extra "flag number".  However, we
 635      * have no use for it, and in all known usecases it is zero, so just
 636      * report lookup failure if it isn't.
 637      */
 638     if (param & 0xffff00000000ULL) {
 639         return HV_STATUS_INVALID_PORT_ID;
 640     }
 641     /* remaining bits are reserved-zero */
 642     if (param & ~HV_CONNECTION_ID_MASK) {
 643         return HV_STATUS_INVALID_HYPERCALL_INPUT;
 644     }
 645
 646     RCU_READ_LOCK_GUARD();
 647     QLIST_FOREACH_RCU(handler, &event_flag_handlers, link) {
 648         if (handler->conn_id == param) {
 649             event_notifier_set(handler->notifier);
 650             return 0;
 651         }
 652     }
 653     return HV_STATUS_INVALID_CONNECTION_ID;
 654 }