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