hw/misc/ivshmem.c

   1 /*
   2  * Inter-VM Shared Memory PCI device.
   3  *
   4  * Author:
   5  *      Cam Macdonell <cam@cs.ualberta.ca>
   6  *
   7  * Based On: cirrus_vga.c
   8  *          Copyright (c) 2004 Fabrice Bellard
   9  *          Copyright (c) 2004 Makoto Suzuki (suzu)
  10  *
  11  *      and rtl8139.c
  12  *          Copyright (c) 2006 Igor Kovalenko
  13  *
  14  * This code is licensed under the GNU GPL v2.
  15  *
  16  * Contributions after 2012-01-13 are licensed under the terms of the
  17  * GNU GPL, version 2 or (at your option) any later version.
  18  */
  19
  20 #include "qemu/osdep.h"
  21 #include "qemu/units.h"
  22 #include "qapi/error.h"
  23 #include "qemu/cutils.h"
  24 #include "hw/pci/pci.h"
  25 #include "hw/qdev-properties.h"
  26 #include "hw/qdev-properties-system.h"
  27 #include "hw/pci/msi.h"
  28 #include "hw/pci/msix.h"
  29 #include "sysemu/kvm.h"
  30 #include "migration/blocker.h"
  31 #include "migration/vmstate.h"
  32 #include "qemu/error-report.h"
  33 #include "qemu/event_notifier.h"
  34 #include "qemu/module.h"
  35 #include "qom/object_interfaces.h"
  36 #include "chardev/char-fe.h"
  37 #include "sysemu/hostmem.h"
  38 #include "qapi/visitor.h"
  39
  40 #include "hw/misc/ivshmem.h"
  41 #include "qom/object.h"
  42
  43 #define PCI_VENDOR_ID_IVSHMEM   PCI_VENDOR_ID_REDHAT_QUMRANET
  44 #define PCI_DEVICE_ID_IVSHMEM   0x1110
  45
  46 #define IVSHMEM_MAX_PEERS UINT16_MAX
  47 #define IVSHMEM_IOEVENTFD   0
  48 #define IVSHMEM_MSI     1
  49
  50 #define IVSHMEM_REG_BAR_SIZE 0x100
  51
  52 #define IVSHMEM_DEBUG 0
  53 #define IVSHMEM_DPRINTF(fmt, ...)                       \
  54     do {                                                \
  55         if (IVSHMEM_DEBUG) {                            \
  56             printf("IVSHMEM: " fmt, ## __VA_ARGS__);    \
  57         }                                               \
  58     } while (0)
  59
  60 #define TYPE_IVSHMEM_COMMON "ivshmem-common"
  61 typedef struct IVShmemState IVShmemState;
  62 DECLARE_INSTANCE_CHECKER(IVShmemState, IVSHMEM_COMMON,
  63                          TYPE_IVSHMEM_COMMON)
  64
  65 #define TYPE_IVSHMEM_PLAIN "ivshmem-plain"
  66 DECLARE_INSTANCE_CHECKER(IVShmemState, IVSHMEM_PLAIN,
  67                          TYPE_IVSHMEM_PLAIN)
  68
  69 #define TYPE_IVSHMEM_DOORBELL "ivshmem-doorbell"
  70 DECLARE_INSTANCE_CHECKER(IVShmemState, IVSHMEM_DOORBELL,
  71                          TYPE_IVSHMEM_DOORBELL)
  72
  73 #define TYPE_IVSHMEM "ivshmem"
  74 DECLARE_INSTANCE_CHECKER(IVShmemState, IVSHMEM,
  75                          TYPE_IVSHMEM)
  76
  77 typedef struct Peer {
  78     int nb_eventfds;
  79     EventNotifier *eventfds;
  80 } Peer;
  81
  82 typedef struct MSIVector {
  83     PCIDevice *pdev;
  84     int virq;
  85     bool unmasked;
  86 } MSIVector;
  87
  88 struct IVShmemState {
  89     /*< private >*/
  90     PCIDevice parent_obj;
  91     /*< public >*/
  92
  93     uint32_t features;
  94
  95     /* exactly one of these two may be set */
  96     HostMemoryBackend *hostmem; /* with interrupts */
  97     CharBackend server_chr; /* without interrupts */
  98
  99     /* registers */
 100     uint32_t intrmask;
 101     uint32_t intrstatus;
 102     int vm_id;
 103
 104     /* BARs */
 105     MemoryRegion ivshmem_mmio;  /* BAR 0 (registers) */
 106     MemoryRegion *ivshmem_bar2; /* BAR 2 (shared memory) */
 107     MemoryRegion server_bar2;   /* used with server_chr */
 108
 109     /* interrupt support */
 110     Peer *peers;
 111     int nb_peers;               /* space in @peers[] */
 112     uint32_t vectors;
 113     MSIVector *msi_vectors;
 114     uint64_t msg_buf;           /* buffer for receiving server messages */
 115     int msg_buffered_bytes;     /* #bytes in @msg_buf */
 116
 117     /* migration stuff */
 118     OnOffAuto master;
 119     Error *migration_blocker;
 120 };
 121
 122 /* registers for the Inter-VM shared memory device */
 123 enum ivshmem_registers {
 124     INTRMASK = 0,
 125     INTRSTATUS = 4,
 126     IVPOSITION = 8,
 127     DOORBELL = 12,
 128 };
 129
 130 static inline uint32_t ivshmem_has_feature(IVShmemState *ivs,
 131                                                     unsigned int feature) {
 132     return (ivs->features & (1 << feature));
 133 }
 134
 135 static inline bool ivshmem_is_master(IVShmemState *s)
 136 {
 137     assert(s->master != ON_OFF_AUTO_AUTO);
 138     return s->master == ON_OFF_AUTO_ON;
 139 }
 140
 141 static void ivshmem_IntrMask_write(IVShmemState *s, uint32_t val)
 142 {
 143     IVSHMEM_DPRINTF("IntrMask write(w) val = 0x%04x\n", val);
 144
 145     s->intrmask = val;
 146 }
 147
 148 static uint32_t ivshmem_IntrMask_read(IVShmemState *s)
 149 {
 150     uint32_t ret = s->intrmask;
 151
 152     IVSHMEM_DPRINTF("intrmask read(w) val = 0x%04x\n", ret);
 153     return ret;
 154 }
 155
 156 static void ivshmem_IntrStatus_write(IVShmemState *s, uint32_t val)
 157 {
 158     IVSHMEM_DPRINTF("IntrStatus write(w) val = 0x%04x\n", val);
 159
 160     s->intrstatus = val;
 161 }
 162
 163 static uint32_t ivshmem_IntrStatus_read(IVShmemState *s)
 164 {
 165     uint32_t ret = s->intrstatus;
 166
 167     /* reading ISR clears all interrupts */
 168     s->intrstatus = 0;
 169     return ret;
 170 }
 171
 172 static void ivshmem_io_write(void *opaque, hwaddr addr,
 173                              uint64_t val, unsigned size)
 174 {
 175     IVShmemState *s = opaque;
 176
 177     uint16_t dest = val >> 16;
 178     uint16_t vector = val & 0xff;
 179
 180     addr &= 0xfc;
 181
 182     IVSHMEM_DPRINTF("writing to addr " TARGET_FMT_plx "\n", addr);
 183     switch (addr)
 184     {
 185         case INTRMASK:
 186             ivshmem_IntrMask_write(s, val);
 187             break;
 188
 189         case INTRSTATUS:
 190             ivshmem_IntrStatus_write(s, val);
 191             break;
 192
 193         case DOORBELL:
 194             /* check that dest VM ID is reasonable */
 195             if (dest >= s->nb_peers) {
 196                 IVSHMEM_DPRINTF("Invalid destination VM ID (%d)\n", dest);
 197                 break;
 198             }
 199
 200             /* check doorbell range */
 201             if (vector < s->peers[dest].nb_eventfds) {
 202                 IVSHMEM_DPRINTF("Notifying VM %d on vector %d\n", dest, vector);
 203                 event_notifier_set(&s->peers[dest].eventfds[vector]);
 204             } else {
 205                 IVSHMEM_DPRINTF("Invalid destination vector %d on VM %d\n",
 206                                 vector, dest);
 207             }
 208             break;
 209         default:
 210             IVSHMEM_DPRINTF("Unhandled write " TARGET_FMT_plx "\n", addr);
 211     }
 212 }
 213
 214 static uint64_t ivshmem_io_read(void *opaque, hwaddr addr,
 215                                 unsigned size)
 216 {
 217
 218     IVShmemState *s = opaque;
 219     uint32_t ret;
 220
 221     switch (addr)
 222     {
 223         case INTRMASK:
 224             ret = ivshmem_IntrMask_read(s);
 225             break;
 226
 227         case INTRSTATUS:
 228             ret = ivshmem_IntrStatus_read(s);
 229             break;
 230
 231         case IVPOSITION:
 232             ret = s->vm_id;
 233             break;
 234
 235         default:
 236             IVSHMEM_DPRINTF("why are we reading " TARGET_FMT_plx "\n", addr);
 237             ret = 0;
 238     }
 239
 240     return ret;
 241 }
 242
 243 static const MemoryRegionOps ivshmem_mmio_ops = {
 244     .read = ivshmem_io_read,
 245     .write = ivshmem_io_write,
 246     .endianness = DEVICE_LITTLE_ENDIAN,
 247     .impl = {
 248         .min_access_size = 4,
 249         .max_access_size = 4,
 250     },
 251 };
 252
 253 static void ivshmem_vector_notify(void *opaque)
 254 {
 255     MSIVector *entry = opaque;
 256     PCIDevice *pdev = entry->pdev;
 257     IVShmemState *s = IVSHMEM_COMMON(pdev);
 258     int vector = entry - s->msi_vectors;
 259     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 260
 261     if (!event_notifier_test_and_clear(n)) {
 262         return;
 263     }
 264
 265     IVSHMEM_DPRINTF("interrupt on vector %p %d\n", pdev, vector);
 266     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 267         if (msix_enabled(pdev)) {
 268             msix_notify(pdev, vector);
 269         }
 270     } else {
 271         ivshmem_IntrStatus_write(s, 1);
 272     }
 273 }
 274
 275 static int ivshmem_vector_unmask(PCIDevice *dev, unsigned vector,
 276                                  MSIMessage msg)
 277 {
 278     IVShmemState *s = IVSHMEM_COMMON(dev);
 279     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 280     MSIVector *v = &s->msi_vectors[vector];
 281     int ret;
 282
 283     IVSHMEM_DPRINTF("vector unmask %p %d\n", dev, vector);
 284     if (!v->pdev) {
 285         error_report("ivshmem: vector %d route does not exist", vector);
 286         return -EINVAL;
 287     }
 288     assert(!v->unmasked);
 289
 290     ret = kvm_irqchip_update_msi_route(kvm_state, v->virq, msg, dev);
 291     if (ret < 0) {
 292         return ret;
 293     }
 294     kvm_irqchip_commit_routes(kvm_state);
 295
 296     ret = kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, n, NULL, v->virq);
 297     if (ret < 0) {
 298         return ret;
 299     }
 300     v->unmasked = true;
 301
 302     return 0;
 303 }
 304
 305 static void ivshmem_vector_mask(PCIDevice *dev, unsigned vector)
 306 {
 307     IVShmemState *s = IVSHMEM_COMMON(dev);
 308     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 309     MSIVector *v = &s->msi_vectors[vector];
 310     int ret;
 311
 312     IVSHMEM_DPRINTF("vector mask %p %d\n", dev, vector);
 313     if (!v->pdev) {
 314         error_report("ivshmem: vector %d route does not exist", vector);
 315         return;
 316     }
 317     assert(v->unmasked);
 318
 319     ret = kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state, n, v->virq);
 320     if (ret < 0) {
 321         error_report("remove_irqfd_notifier_gsi failed");
 322         return;
 323     }
 324     v->unmasked = false;
 325 }
 326
 327 static void ivshmem_vector_poll(PCIDevice *dev,
 328                                 unsigned int vector_start,
 329                                 unsigned int vector_end)
 330 {
 331     IVShmemState *s = IVSHMEM_COMMON(dev);
 332     unsigned int vector;
 333
 334     IVSHMEM_DPRINTF("vector poll %p %d-%d\n", dev, vector_start, vector_end);
 335
 336     vector_end = MIN(vector_end, s->vectors);
 337
 338     for (vector = vector_start; vector < vector_end; vector++) {
 339         EventNotifier *notifier = &s->peers[s->vm_id].eventfds[vector];
 340
 341         if (!msix_is_masked(dev, vector)) {
 342             continue;
 343         }
 344
 345         if (event_notifier_test_and_clear(notifier)) {
 346             msix_set_pending(dev, vector);
 347         }
 348     }
 349 }
 350
 351 static void watch_vector_notifier(IVShmemState *s, EventNotifier *n,
 352                                  int vector)
 353 {
 354     int eventfd = event_notifier_get_fd(n);
 355
 356     assert(!s->msi_vectors[vector].pdev);
 357     s->msi_vectors[vector].pdev = PCI_DEVICE(s);
 358
 359     qemu_set_fd_handler(eventfd, ivshmem_vector_notify,
 360                         NULL, &s->msi_vectors[vector]);
 361 }
 362
 363 static void ivshmem_add_eventfd(IVShmemState *s, int posn, int i)
 364 {
 365     memory_region_add_eventfd(&s->ivshmem_mmio,
 366                               DOORBELL,
 367                               4,
 368                               true,
 369                               (posn << 16) | i,
 370                               &s->peers[posn].eventfds[i]);
 371 }
 372
 373 static void ivshmem_del_eventfd(IVShmemState *s, int posn, int i)
 374 {
 375     memory_region_del_eventfd(&s->ivshmem_mmio,
 376                               DOORBELL,
 377                               4,
 378                               true,
 379                               (posn << 16) | i,
 380                               &s->peers[posn].eventfds[i]);
 381 }
 382
 383 static void close_peer_eventfds(IVShmemState *s, int posn)
 384 {
 385     int i, n;
 386
 387     assert(posn >= 0 && posn < s->nb_peers);
 388     n = s->peers[posn].nb_eventfds;
 389
 390     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
 391         memory_region_transaction_begin();
 392         for (i = 0; i < n; i++) {
 393             ivshmem_del_eventfd(s, posn, i);
 394         }
 395         memory_region_transaction_commit();
 396     }
 397
 398     for (i = 0; i < n; i++) {
 399         event_notifier_cleanup(&s->peers[posn].eventfds[i]);
 400     }
 401
 402     g_free(s->peers[posn].eventfds);
 403     s->peers[posn].nb_eventfds = 0;
 404 }
 405
 406 static void resize_peers(IVShmemState *s, int nb_peers)
 407 {
 408     int old_nb_peers = s->nb_peers;
 409     int i;
 410
 411     assert(nb_peers > old_nb_peers);
 412     IVSHMEM_DPRINTF("bumping storage to %d peers\n", nb_peers);
 413
 414     s->peers = g_renew(Peer, s->peers, nb_peers);
 415     s->nb_peers = nb_peers;
 416
 417     for (i = old_nb_peers; i < nb_peers; i++) {
 418         s->peers[i].eventfds = g_new0(EventNotifier, s->vectors);
 419         s->peers[i].nb_eventfds = 0;
 420     }
 421 }
 422
 423 static void ivshmem_add_kvm_msi_virq(IVShmemState *s, int vector,
 424                                      Error **errp)
 425 {
 426     PCIDevice *pdev = PCI_DEVICE(s);
 427     KVMRouteChange c;
 428     int ret;
 429
 430     IVSHMEM_DPRINTF("ivshmem_add_kvm_msi_virq vector:%d\n", vector);
 431     assert(!s->msi_vectors[vector].pdev);
 432
 433     c = kvm_irqchip_begin_route_changes(kvm_state);
 434     ret = kvm_irqchip_add_msi_route(&c, vector, pdev);
 435     if (ret < 0) {
 436         error_setg(errp, "kvm_irqchip_add_msi_route failed");
 437         return;
 438     }
 439     kvm_irqchip_commit_route_changes(&c);
 440
 441     s->msi_vectors[vector].virq = ret;
 442     s->msi_vectors[vector].pdev = pdev;
 443 }
 444
 445 static void setup_interrupt(IVShmemState *s, int vector, Error **errp)
 446 {
 447     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 448     bool with_irqfd = kvm_msi_via_irqfd_enabled() &&
 449         ivshmem_has_feature(s, IVSHMEM_MSI);
 450     PCIDevice *pdev = PCI_DEVICE(s);
 451     Error *err = NULL;
 452
 453     IVSHMEM_DPRINTF("setting up interrupt for vector: %d\n", vector);
 454
 455     if (!with_irqfd) {
 456         IVSHMEM_DPRINTF("with eventfd\n");
 457         watch_vector_notifier(s, n, vector);
 458     } else if (msix_enabled(pdev)) {
 459         IVSHMEM_DPRINTF("with irqfd\n");
 460         ivshmem_add_kvm_msi_virq(s, vector, &err);
 461         if (err) {
 462             error_propagate(errp, err);
 463             return;
 464         }
 465
 466         if (!msix_is_masked(pdev, vector)) {
 467             kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, n, NULL,
 468                                                s->msi_vectors[vector].virq);
 469             /* TODO handle error */
 470         }
 471     } else {
 472         /* it will be delayed until msix is enabled, in write_config */
 473         IVSHMEM_DPRINTF("with irqfd, delayed until msix enabled\n");
 474     }
 475 }
 476
 477 static void process_msg_shmem(IVShmemState *s, int fd, Error **errp)
 478 {
 479     Error *local_err = NULL;
 480     struct stat buf;
 481     size_t size;
 482
 483     if (s->ivshmem_bar2) {
 484         error_setg(errp, "server sent unexpected shared memory message");
 485         close(fd);
 486         return;
 487     }
 488
 489     if (fstat(fd, &buf) < 0) {
 490         error_setg_errno(errp, errno,
 491             "can't determine size of shared memory sent by server");
 492         close(fd);
 493         return;
 494     }
 495
 496     size = buf.st_size;
 497
 498     /* mmap the region and map into the BAR2 */
 499     memory_region_init_ram_from_fd(&s->server_bar2, OBJECT(s), "ivshmem.bar2",
 500                                    size, RAM_SHARED, fd, 0, &local_err);
 501     if (local_err) {
 502         error_propagate(errp, local_err);
 503         return;
 504     }
 505
 506     s->ivshmem_bar2 = &s->server_bar2;
 507 }
 508
 509 static void process_msg_disconnect(IVShmemState *s, uint16_t posn,
 510                                    Error **errp)
 511 {
 512     IVSHMEM_DPRINTF("posn %d has gone away\n", posn);
 513     if (posn >= s->nb_peers || posn == s->vm_id) {
 514         error_setg(errp, "invalid peer %d", posn);
 515         return;
 516     }
 517     close_peer_eventfds(s, posn);
 518 }
 519
 520 static void process_msg_connect(IVShmemState *s, uint16_t posn, int fd,
 521                                 Error **errp)
 522 {
 523     Peer *peer = &s->peers[posn];
 524     int vector;
 525
 526     /*
 527      * The N-th connect message for this peer comes with the file
 528      * descriptor for vector N-1.  Count messages to find the vector.
 529      */
 530     if (peer->nb_eventfds >= s->vectors) {
 531         error_setg(errp, "Too many eventfd received, device has %d vectors",
 532                    s->vectors);
 533         close(fd);
 534         return;
 535     }
 536     vector = peer->nb_eventfds++;
 537
 538     IVSHMEM_DPRINTF("eventfds[%d][%d] = %d\n", posn, vector, fd);
 539     event_notifier_init_fd(&peer->eventfds[vector], fd);
 540     g_unix_set_fd_nonblocking(fd, true, NULL); /* msix/irqfd poll non block */
 541
 542     if (posn == s->vm_id) {
 543         setup_interrupt(s, vector, errp);
 544         /* TODO do we need to handle the error? */
 545     }
 546
 547     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
 548         ivshmem_add_eventfd(s, posn, vector);
 549     }
 550 }
 551
 552 static void process_msg(IVShmemState *s, int64_t msg, int fd, Error **errp)
 553 {
 554     IVSHMEM_DPRINTF("posn is %" PRId64 ", fd is %d\n", msg, fd);
 555
 556     if (msg < -1 || msg > IVSHMEM_MAX_PEERS) {
 557         error_setg(errp, "server sent invalid message %" PRId64, msg);
 558         close(fd);
 559         return;
 560     }
 561
 562     if (msg == -1) {
 563         process_msg_shmem(s, fd, errp);
 564         return;
 565     }
 566
 567     if (msg >= s->nb_peers) {
 568         resize_peers(s, msg + 1);
 569     }
 570
 571     if (fd >= 0) {
 572         process_msg_connect(s, msg, fd, errp);
 573     } else {
 574         process_msg_disconnect(s, msg, errp);
 575     }
 576 }
 577
 578 static int ivshmem_can_receive(void *opaque)
 579 {
 580     IVShmemState *s = opaque;
 581
 582     assert(s->msg_buffered_bytes < sizeof(s->msg_buf));
 583     return sizeof(s->msg_buf) - s->msg_buffered_bytes;
 584 }
 585
 586 static void ivshmem_read(void *opaque, const uint8_t *buf, int size)
 587 {
 588     IVShmemState *s = opaque;
 589     Error *err = NULL;
 590     int fd;
 591     int64_t msg;
 592
 593     assert(size >= 0 && s->msg_buffered_bytes + size <= sizeof(s->msg_buf));
 594     memcpy((unsigned char *)&s->msg_buf + s->msg_buffered_bytes, buf, size);
 595     s->msg_buffered_bytes += size;
 596     if (s->msg_buffered_bytes < sizeof(s->msg_buf)) {
 597         return;
 598     }
 599     msg = le64_to_cpu(s->msg_buf);
 600     s->msg_buffered_bytes = 0;
 601
 602     fd = qemu_chr_fe_get_msgfd(&s->server_chr);
 603
 604     process_msg(s, msg, fd, &err);
 605     if (err) {
 606         error_report_err(err);
 607     }
 608 }
 609
 610 static int64_t ivshmem_recv_msg(IVShmemState *s, int *pfd, Error **errp)
 611 {
 612     int64_t msg;
 613     int n, ret;
 614
 615     n = 0;
 616     do {
 617         ret = qemu_chr_fe_read_all(&s->server_chr, (uint8_t *)&msg + n,
 618                                    sizeof(msg) - n);
 619         if (ret < 0) {
 620             if (ret == -EINTR) {
 621                 continue;
 622             }
 623             error_setg_errno(errp, -ret, "read from server failed");
 624             return INT64_MIN;
 625         }
 626         n += ret;
 627     } while (n < sizeof(msg));
 628
 629     *pfd = qemu_chr_fe_get_msgfd(&s->server_chr);
 630     return le64_to_cpu(msg);
 631 }
 632
 633 static void ivshmem_recv_setup(IVShmemState *s, Error **errp)
 634 {
 635     Error *err = NULL;
 636     int64_t msg;
 637     int fd;
 638
 639     msg = ivshmem_recv_msg(s, &fd, &err);
 640     if (err) {
 641         error_propagate(errp, err);
 642         return;
 643     }
 644     if (msg != IVSHMEM_PROTOCOL_VERSION) {
 645         error_setg(errp, "server sent version %" PRId64 ", expecting %d",
 646                    msg, IVSHMEM_PROTOCOL_VERSION);
 647         return;
 648     }
 649     if (fd != -1) {
 650         error_setg(errp, "server sent invalid version message");
 651         return;
 652     }
 653
 654     /*
 655      * ivshmem-server sends the remaining initial messages in a fixed
 656      * order, but the device has always accepted them in any order.
 657      * Stay as compatible as practical, just in case people use
 658      * servers that behave differently.
 659      */
 660
 661     /*
 662      * ivshmem_device_spec.txt has always required the ID message
 663      * right here, and ivshmem-server has always complied.  However,
 664      * older versions of the device accepted it out of order, but
 665      * broke when an interrupt setup message arrived before it.
 666      */
 667     msg = ivshmem_recv_msg(s, &fd, &err);
 668     if (err) {
 669         error_propagate(errp, err);
 670         return;
 671     }
 672     if (fd != -1 || msg < 0 || msg > IVSHMEM_MAX_PEERS) {
 673         error_setg(errp, "server sent invalid ID message");
 674         return;
 675     }
 676     s->vm_id = msg;
 677
 678     /*
 679      * Receive more messages until we got shared memory.
 680      */
 681     do {
 682         msg = ivshmem_recv_msg(s, &fd, &err);
 683         if (err) {
 684             error_propagate(errp, err);
 685             return;
 686         }
 687         process_msg(s, msg, fd, &err);
 688         if (err) {
 689             error_propagate(errp, err);
 690             return;
 691         }
 692     } while (msg != -1);
 693
 694     /*
 695      * This function must either map the shared memory or fail.  The
 696      * loop above ensures that: it terminates normally only after it
 697      * successfully processed the server's shared memory message.
 698      * Assert that actually mapped the shared memory:
 699      */
 700     assert(s->ivshmem_bar2);
 701 }
 702
 703 /* Select the MSI-X vectors used by device.
 704  * ivshmem maps events to vectors statically, so
 705  * we just enable all vectors on init and after reset. */
 706 static void ivshmem_msix_vector_use(IVShmemState *s)
 707 {
 708     PCIDevice *d = PCI_DEVICE(s);
 709     int i;
 710
 711     for (i = 0; i < s->vectors; i++) {
 712         msix_vector_use(d, i);
 713     }
 714 }
 715
 716 static void ivshmem_disable_irqfd(IVShmemState *s);
 717
 718 static void ivshmem_reset(DeviceState *d)
 719 {
 720     IVShmemState *s = IVSHMEM_COMMON(d);
 721
 722     ivshmem_disable_irqfd(s);
 723
 724     s->intrstatus = 0;
 725     s->intrmask = 0;
 726     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 727         ivshmem_msix_vector_use(s);
 728     }
 729 }
 730
 731 static int ivshmem_setup_interrupts(IVShmemState *s, Error **errp)
 732 {
 733     /* allocate QEMU callback data for receiving interrupts */
 734     s->msi_vectors = g_new0(MSIVector, s->vectors);
 735
 736     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 737         if (msix_init_exclusive_bar(PCI_DEVICE(s), s->vectors, 1, errp)) {
 738             return -1;
 739         }
 740
 741         IVSHMEM_DPRINTF("msix initialized (%d vectors)\n", s->vectors);
 742         ivshmem_msix_vector_use(s);
 743     }
 744
 745     return 0;
 746 }
 747
 748 static void ivshmem_remove_kvm_msi_virq(IVShmemState *s, int vector)
 749 {
 750     IVSHMEM_DPRINTF("ivshmem_remove_kvm_msi_virq vector:%d\n", vector);
 751
 752     if (s->msi_vectors[vector].pdev == NULL) {
 753         return;
 754     }
 755
 756     /* it was cleaned when masked in the frontend. */
 757     kvm_irqchip_release_virq(kvm_state, s->msi_vectors[vector].virq);
 758
 759     s->msi_vectors[vector].pdev = NULL;
 760 }
 761
 762 static void ivshmem_enable_irqfd(IVShmemState *s)
 763 {
 764     PCIDevice *pdev = PCI_DEVICE(s);
 765     int i;
 766
 767     for (i = 0; i < s->peers[s->vm_id].nb_eventfds; i++) {
 768         Error *err = NULL;
 769
 770         ivshmem_add_kvm_msi_virq(s, i, &err);
 771         if (err) {
 772             error_report_err(err);
 773             goto undo;
 774         }
 775     }
 776
 777     if (msix_set_vector_notifiers(pdev,
 778                                   ivshmem_vector_unmask,
 779                                   ivshmem_vector_mask,
 780                                   ivshmem_vector_poll)) {
 781         error_report("ivshmem: msix_set_vector_notifiers failed");
 782         goto undo;
 783     }
 784     return;
 785
 786 undo:
 787     while (--i >= 0) {
 788         ivshmem_remove_kvm_msi_virq(s, i);
 789     }
 790 }
 791
 792 static void ivshmem_disable_irqfd(IVShmemState *s)
 793 {
 794     PCIDevice *pdev = PCI_DEVICE(s);
 795     int i;
 796
 797     if (!pdev->msix_vector_use_notifier) {
 798         return;
 799     }
 800
 801     msix_unset_vector_notifiers(pdev);
 802
 803     for (i = 0; i < s->peers[s->vm_id].nb_eventfds; i++) {
 804         /*
 805          * MSI-X is already disabled here so msix_unset_vector_notifiers()
 806          * didn't call our release notifier.  Do it now to keep our masks and
 807          * unmasks balanced.
 808          */
 809         if (s->msi_vectors[i].unmasked) {
 810             ivshmem_vector_mask(pdev, i);
 811         }
 812         ivshmem_remove_kvm_msi_virq(s, i);
 813     }
 814
 815 }
 816
 817 static void ivshmem_write_config(PCIDevice *pdev, uint32_t address,
 818                                  uint32_t val, int len)
 819 {
 820     IVShmemState *s = IVSHMEM_COMMON(pdev);
 821     int is_enabled, was_enabled = msix_enabled(pdev);
 822
 823     pci_default_write_config(pdev, address, val, len);
 824     is_enabled = msix_enabled(pdev);
 825
 826     if (kvm_msi_via_irqfd_enabled()) {
 827         if (!was_enabled && is_enabled) {
 828             ivshmem_enable_irqfd(s);
 829         } else if (was_enabled && !is_enabled) {
 830             ivshmem_disable_irqfd(s);
 831         }
 832     }
 833 }
 834
 835 static void ivshmem_common_realize(PCIDevice *dev, Error **errp)
 836 {
 837     IVShmemState *s = IVSHMEM_COMMON(dev);
 838     Error *err = NULL;
 839     uint8_t *pci_conf;
 840
 841     /* IRQFD requires MSI */
 842     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD) &&
 843         !ivshmem_has_feature(s, IVSHMEM_MSI)) {
 844         error_setg(errp, "ioeventfd/irqfd requires MSI");
 845         return;
 846     }
 847
 848     pci_conf = dev->config;
 849     pci_conf[PCI_COMMAND] = PCI_COMMAND_IO | PCI_COMMAND_MEMORY;
 850
 851     memory_region_init_io(&s->ivshmem_mmio, OBJECT(s), &ivshmem_mmio_ops, s,
 852                           "ivshmem-mmio", IVSHMEM_REG_BAR_SIZE);
 853
 854     /* region for registers*/
 855     pci_register_bar(dev, 0, PCI_BASE_ADDRESS_SPACE_MEMORY,
 856                      &s->ivshmem_mmio);
 857
 858     if (s->hostmem != NULL) {
 859         IVSHMEM_DPRINTF("using hostmem\n");
 860
 861         s->ivshmem_bar2 = host_memory_backend_get_memory(s->hostmem);
 862         host_memory_backend_set_mapped(s->hostmem, true);
 863     } else {
 864         Chardev *chr = qemu_chr_fe_get_driver(&s->server_chr);
 865         assert(chr);
 866
 867         IVSHMEM_DPRINTF("using shared memory server (socket = %s)\n",
 868                         chr->filename);
 869
 870         /* we allocate enough space for 16 peers and grow as needed */
 871         resize_peers(s, 16);
 872
 873         /*
 874          * Receive setup messages from server synchronously.
 875          * Older versions did it asynchronously, but that creates a
 876          * number of entertaining race conditions.
 877          */
 878         ivshmem_recv_setup(s, &err);
 879         if (err) {
 880             error_propagate(errp, err);
 881             return;
 882         }
 883
 884         if (s->master == ON_OFF_AUTO_ON && s->vm_id != 0) {
 885             error_setg(errp,
 886                        "master must connect to the server before any peers");
 887             return;
 888         }
 889
 890         qemu_chr_fe_set_handlers(&s->server_chr, ivshmem_can_receive,
 891                                  ivshmem_read, NULL, NULL, s, NULL, true);
 892
 893         if (ivshmem_setup_interrupts(s, errp) < 0) {
 894             error_prepend(errp, "Failed to initialize interrupts: ");
 895             return;
 896         }
 897     }
 898
 899     if (s->master == ON_OFF_AUTO_AUTO) {
 900         s->master = s->vm_id == 0 ? ON_OFF_AUTO_ON : ON_OFF_AUTO_OFF;
 901     }
 902
 903     if (!ivshmem_is_master(s)) {
 904         error_setg(&s->migration_blocker,
 905                    "Migration is disabled when using feature 'peer mode' in device 'ivshmem'");
 906         if (migrate_add_blocker(s->migration_blocker, errp) < 0) {
 907             error_free(s->migration_blocker);
 908             return;
 909         }
 910     }
 911
 912     vmstate_register_ram(s->ivshmem_bar2, DEVICE(s));
 913     pci_register_bar(PCI_DEVICE(s), 2,
 914                      PCI_BASE_ADDRESS_SPACE_MEMORY |
 915                      PCI_BASE_ADDRESS_MEM_PREFETCH |
 916                      PCI_BASE_ADDRESS_MEM_TYPE_64,
 917                      s->ivshmem_bar2);
 918 }
 919
 920 static void ivshmem_exit(PCIDevice *dev)
 921 {
 922     IVShmemState *s = IVSHMEM_COMMON(dev);
 923     int i;
 924
 925     if (s->migration_blocker) {
 926         migrate_del_blocker(s->migration_blocker);
 927         error_free(s->migration_blocker);
 928     }
 929
 930     if (memory_region_is_mapped(s->ivshmem_bar2)) {
 931         if (!s->hostmem) {
 932             void *addr = memory_region_get_ram_ptr(s->ivshmem_bar2);
 933             int fd;
 934
 935             if (munmap(addr, memory_region_size(s->ivshmem_bar2) == -1)) {
 936                 error_report("Failed to munmap shared memory %s",
 937                              strerror(errno));
 938             }
 939
 940             fd = memory_region_get_fd(s->ivshmem_bar2);
 941             close(fd);
 942         }
 943
 944         vmstate_unregister_ram(s->ivshmem_bar2, DEVICE(dev));
 945     }
 946
 947     if (s->hostmem) {
 948         host_memory_backend_set_mapped(s->hostmem, false);
 949     }
 950
 951     if (s->peers) {
 952         for (i = 0; i < s->nb_peers; i++) {
 953             close_peer_eventfds(s, i);
 954         }
 955         g_free(s->peers);
 956     }
 957
 958     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 959         msix_uninit_exclusive_bar(dev);
 960     }
 961
 962     g_free(s->msi_vectors);
 963 }
 964
 965 static int ivshmem_pre_load(void *opaque)
 966 {
 967     IVShmemState *s = opaque;
 968
 969     if (!ivshmem_is_master(s)) {
 970         error_report("'peer' devices are not migratable");
 971         return -EINVAL;
 972     }
 973
 974     return 0;
 975 }
 976
 977 static int ivshmem_post_load(void *opaque, int version_id)
 978 {
 979     IVShmemState *s = opaque;
 980
 981     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 982         ivshmem_msix_vector_use(s);
 983     }
 984     return 0;
 985 }
 986
 987 static void ivshmem_common_class_init(ObjectClass *klass, void *data)
 988 {
 989     DeviceClass *dc = DEVICE_CLASS(klass);
 990     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
 991
 992     k->realize = ivshmem_common_realize;
 993     k->exit = ivshmem_exit;
 994     k->config_write = ivshmem_write_config;
 995     k->vendor_id = PCI_VENDOR_ID_IVSHMEM;
 996     k->device_id = PCI_DEVICE_ID_IVSHMEM;
 997     k->class_id = PCI_CLASS_MEMORY_RAM;
 998     k->revision = 1;
 999     dc->reset = ivshmem_reset;
1000     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
1001     dc->desc = "Inter-VM shared memory";
1002 }
1003
1004 static const TypeInfo ivshmem_common_info = {
1005     .name          = TYPE_IVSHMEM_COMMON,
1006     .parent        = TYPE_PCI_DEVICE,
1007     .instance_size = sizeof(IVShmemState),
1008     .abstract      = true,
1009     .class_init    = ivshmem_common_class_init,
1010     .interfaces = (InterfaceInfo[]) {
1011         { INTERFACE_CONVENTIONAL_PCI_DEVICE },
1012         { },
1013     },
1014 };
1015
1016 static const VMStateDescription ivshmem_plain_vmsd = {
1017     .name = TYPE_IVSHMEM_PLAIN,
1018     .version_id = 0,
1019     .minimum_version_id = 0,
1020     .pre_load = ivshmem_pre_load,
1021     .post_load = ivshmem_post_load,
1022     .fields = (VMStateField[]) {
1023         VMSTATE_PCI_DEVICE(parent_obj, IVShmemState),
1024         VMSTATE_UINT32(intrstatus, IVShmemState),
1025         VMSTATE_UINT32(intrmask, IVShmemState),
1026         VMSTATE_END_OF_LIST()
1027     },
1028 };
1029
1030 static Property ivshmem_plain_properties[] = {
1031     DEFINE_PROP_ON_OFF_AUTO("master", IVShmemState, master, ON_OFF_AUTO_OFF),
1032     DEFINE_PROP_LINK("memdev", IVShmemState, hostmem, TYPE_MEMORY_BACKEND,
1033                      HostMemoryBackend *),
1034     DEFINE_PROP_END_OF_LIST(),
1035 };
1036
1037 static void ivshmem_plain_realize(PCIDevice *dev, Error **errp)
1038 {
1039     IVShmemState *s = IVSHMEM_COMMON(dev);
1040
1041     if (!s->hostmem) {
1042         error_setg(errp, "You must specify a 'memdev'");
1043         return;
1044     } else if (host_memory_backend_is_mapped(s->hostmem)) {
1045         error_setg(errp, "can't use already busy memdev: %s",
1046                    object_get_canonical_path_component(OBJECT(s->hostmem)));
1047         return;
1048     }
1049
1050     ivshmem_common_realize(dev, errp);
1051 }
1052
1053 static void ivshmem_plain_class_init(ObjectClass *klass, void *data)
1054 {
1055     DeviceClass *dc = DEVICE_CLASS(klass);
1056     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
1057
1058     k->realize = ivshmem_plain_realize;
1059     device_class_set_props(dc, ivshmem_plain_properties);
1060     dc->vmsd = &ivshmem_plain_vmsd;
1061 }
1062
1063 static const TypeInfo ivshmem_plain_info = {
1064     .name          = TYPE_IVSHMEM_PLAIN,
1065     .parent        = TYPE_IVSHMEM_COMMON,
1066     .instance_size = sizeof(IVShmemState),
1067     .class_init    = ivshmem_plain_class_init,
1068 };
1069
1070 static const VMStateDescription ivshmem_doorbell_vmsd = {
1071     .name = TYPE_IVSHMEM_DOORBELL,
1072     .version_id = 0,
1073     .minimum_version_id = 0,
1074     .pre_load = ivshmem_pre_load,
1075     .post_load = ivshmem_post_load,
1076     .fields = (VMStateField[]) {
1077         VMSTATE_PCI_DEVICE(parent_obj, IVShmemState),
1078         VMSTATE_MSIX(parent_obj, IVShmemState),
1079         VMSTATE_UINT32(intrstatus, IVShmemState),
1080         VMSTATE_UINT32(intrmask, IVShmemState),
1081         VMSTATE_END_OF_LIST()
1082     },
1083 };
1084
1085 static Property ivshmem_doorbell_properties[] = {
1086     DEFINE_PROP_CHR("chardev", IVShmemState, server_chr),
1087     DEFINE_PROP_UINT32("vectors", IVShmemState, vectors, 1),
1088     DEFINE_PROP_BIT("ioeventfd", IVShmemState, features, IVSHMEM_IOEVENTFD,
1089                     true),
1090     DEFINE_PROP_ON_OFF_AUTO("master", IVShmemState, master, ON_OFF_AUTO_OFF),
1091     DEFINE_PROP_END_OF_LIST(),
1092 };
1093
1094 static void ivshmem_doorbell_init(Object *obj)
1095 {
1096     IVShmemState *s = IVSHMEM_DOORBELL(obj);
1097
1098     s->features |= (1 << IVSHMEM_MSI);
1099 }
1100
1101 static void ivshmem_doorbell_realize(PCIDevice *dev, Error **errp)
1102 {
1103     IVShmemState *s = IVSHMEM_COMMON(dev);
1104
1105     if (!qemu_chr_fe_backend_connected(&s->server_chr)) {
1106         error_setg(errp, "You must specify a 'chardev'");
1107         return;
1108     }
1109
1110     ivshmem_common_realize(dev, errp);
1111 }
1112
1113 static void ivshmem_doorbell_class_init(ObjectClass *klass, void *data)
1114 {
1115     DeviceClass *dc = DEVICE_CLASS(klass);
1116     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
1117
1118     k->realize = ivshmem_doorbell_realize;
1119     device_class_set_props(dc, ivshmem_doorbell_properties);
1120     dc->vmsd = &ivshmem_doorbell_vmsd;
1121 }
1122
1123 static const TypeInfo ivshmem_doorbell_info = {
1124     .name          = TYPE_IVSHMEM_DOORBELL,
1125     .parent        = TYPE_IVSHMEM_COMMON,
1126     .instance_size = sizeof(IVShmemState),
1127     .instance_init = ivshmem_doorbell_init,
1128     .class_init    = ivshmem_doorbell_class_init,
1129 };
1130
1131 static void ivshmem_register_types(void)
1132 {
1133     type_register_static(&ivshmem_common_info);
1134     type_register_static(&ivshmem_plain_info);
1135     type_register_static(&ivshmem_doorbell_info);
1136 }
1137
1138 type_init(ivshmem_register_types)