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 "sysemu/qtest.h"
  39 #include "qapi/visitor.h"
  40
  41 #include "hw/misc/ivshmem.h"
  42 #include "qom/object.h"
  43
  44 #define PCI_VENDOR_ID_IVSHMEM   PCI_VENDOR_ID_REDHAT_QUMRANET
  45 #define PCI_DEVICE_ID_IVSHMEM   0x1110
  46
  47 #define IVSHMEM_MAX_PEERS UINT16_MAX
  48 #define IVSHMEM_IOEVENTFD   0
  49 #define IVSHMEM_MSI     1
  50
  51 #define IVSHMEM_REG_BAR_SIZE 0x100
  52
  53 #define IVSHMEM_DEBUG 0
  54 #define IVSHMEM_DPRINTF(fmt, ...)                       \
  55     do {                                                \
  56         if (IVSHMEM_DEBUG) {                            \
  57             printf("IVSHMEM: " fmt, ## __VA_ARGS__);    \
  58         }                                               \
  59     } while (0)
  60
  61 #define TYPE_IVSHMEM_COMMON "ivshmem-common"
  62 typedef struct IVShmemState IVShmemState;
  63 DECLARE_INSTANCE_CHECKER(IVShmemState, IVSHMEM_COMMON,
  64                          TYPE_IVSHMEM_COMMON)
  65
  66 #define TYPE_IVSHMEM_PLAIN "ivshmem-plain"
  67 DECLARE_INSTANCE_CHECKER(IVShmemState, IVSHMEM_PLAIN,
  68                          TYPE_IVSHMEM_PLAIN)
  69
  70 #define TYPE_IVSHMEM_DOORBELL "ivshmem-doorbell"
  71 DECLARE_INSTANCE_CHECKER(IVShmemState, IVSHMEM_DOORBELL,
  72                          TYPE_IVSHMEM_DOORBELL)
  73
  74 #define TYPE_IVSHMEM "ivshmem"
  75 DECLARE_INSTANCE_CHECKER(IVShmemState, IVSHMEM,
  76                          TYPE_IVSHMEM)
  77
  78 typedef struct Peer {
  79     int nb_eventfds;
  80     EventNotifier *eventfds;
  81 } Peer;
  82
  83 typedef struct MSIVector {
  84     PCIDevice *pdev;
  85     int virq;
  86     bool unmasked;
  87 } MSIVector;
  88
  89 struct IVShmemState {
  90     /*< private >*/
  91     PCIDevice parent_obj;
  92     /*< public >*/
  93
  94     uint32_t features;
  95
  96     /* exactly one of these two may be set */
  97     HostMemoryBackend *hostmem; /* with interrupts */
  98     CharBackend server_chr; /* without interrupts */
  99
 100     /* registers */
 101     uint32_t intrmask;
 102     uint32_t intrstatus;
 103     int vm_id;
 104
 105     /* BARs */
 106     MemoryRegion ivshmem_mmio;  /* BAR 0 (registers) */
 107     MemoryRegion *ivshmem_bar2; /* BAR 2 (shared memory) */
 108     MemoryRegion server_bar2;   /* used with server_chr */
 109
 110     /* interrupt support */
 111     Peer *peers;
 112     int nb_peers;               /* space in @peers[] */
 113     uint32_t vectors;
 114     MSIVector *msi_vectors;
 115     uint64_t msg_buf;           /* buffer for receiving server messages */
 116     int msg_buffered_bytes;     /* #bytes in @msg_buf */
 117
 118     /* migration stuff */
 119     OnOffAuto master;
 120     Error *migration_blocker;
 121 };
 122
 123 /* registers for the Inter-VM shared memory device */
 124 enum ivshmem_registers {
 125     INTRMASK = 0,
 126     INTRSTATUS = 4,
 127     IVPOSITION = 8,
 128     DOORBELL = 12,
 129 };
 130
 131 static inline uint32_t ivshmem_has_feature(IVShmemState *ivs,
 132                                                     unsigned int feature) {
 133     return (ivs->features & (1 << feature));
 134 }
 135
 136 static inline bool ivshmem_is_master(IVShmemState *s)
 137 {
 138     assert(s->master != ON_OFF_AUTO_AUTO);
 139     return s->master == ON_OFF_AUTO_ON;
 140 }
 141
 142 static void ivshmem_IntrMask_write(IVShmemState *s, uint32_t val)
 143 {
 144     IVSHMEM_DPRINTF("IntrMask write(w) val = 0x%04x\n", val);
 145
 146     s->intrmask = val;
 147 }
 148
 149 static uint32_t ivshmem_IntrMask_read(IVShmemState *s)
 150 {
 151     uint32_t ret = s->intrmask;
 152
 153     IVSHMEM_DPRINTF("intrmask read(w) val = 0x%04x\n", ret);
 154     return ret;
 155 }
 156
 157 static void ivshmem_IntrStatus_write(IVShmemState *s, uint32_t val)
 158 {
 159     IVSHMEM_DPRINTF("IntrStatus write(w) val = 0x%04x\n", val);
 160
 161     s->intrstatus = val;
 162 }
 163
 164 static uint32_t ivshmem_IntrStatus_read(IVShmemState *s)
 165 {
 166     uint32_t ret = s->intrstatus;
 167
 168     /* reading ISR clears all interrupts */
 169     s->intrstatus = 0;
 170     return ret;
 171 }
 172
 173 static void ivshmem_io_write(void *opaque, hwaddr addr,
 174                              uint64_t val, unsigned size)
 175 {
 176     IVShmemState *s = opaque;
 177
 178     uint16_t dest = val >> 16;
 179     uint16_t vector = val & 0xff;
 180
 181     addr &= 0xfc;
 182
 183     IVSHMEM_DPRINTF("writing to addr " TARGET_FMT_plx "\n", addr);
 184     switch (addr)
 185     {
 186         case INTRMASK:
 187             ivshmem_IntrMask_write(s, val);
 188             break;
 189
 190         case INTRSTATUS:
 191             ivshmem_IntrStatus_write(s, val);
 192             break;
 193
 194         case DOORBELL:
 195             /* check that dest VM ID is reasonable */
 196             if (dest >= s->nb_peers) {
 197                 IVSHMEM_DPRINTF("Invalid destination VM ID (%d)\n", dest);
 198                 break;
 199             }
 200
 201             /* check doorbell range */
 202             if (vector < s->peers[dest].nb_eventfds) {
 203                 IVSHMEM_DPRINTF("Notifying VM %d on vector %d\n", dest, vector);
 204                 event_notifier_set(&s->peers[dest].eventfds[vector]);
 205             } else {
 206                 IVSHMEM_DPRINTF("Invalid destination vector %d on VM %d\n",
 207                                 vector, dest);
 208             }
 209             break;
 210         default:
 211             IVSHMEM_DPRINTF("Unhandled write " TARGET_FMT_plx "\n", addr);
 212     }
 213 }
 214
 215 static uint64_t ivshmem_io_read(void *opaque, hwaddr addr,
 216                                 unsigned size)
 217 {
 218
 219     IVShmemState *s = opaque;
 220     uint32_t ret;
 221
 222     switch (addr)
 223     {
 224         case INTRMASK:
 225             ret = ivshmem_IntrMask_read(s);
 226             break;
 227
 228         case INTRSTATUS:
 229             ret = ivshmem_IntrStatus_read(s);
 230             break;
 231
 232         case IVPOSITION:
 233             ret = s->vm_id;
 234             break;
 235
 236         default:
 237             IVSHMEM_DPRINTF("why are we reading " TARGET_FMT_plx "\n", addr);
 238             ret = 0;
 239     }
 240
 241     return ret;
 242 }
 243
 244 static const MemoryRegionOps ivshmem_mmio_ops = {
 245     .read = ivshmem_io_read,
 246     .write = ivshmem_io_write,
 247     .endianness = DEVICE_NATIVE_ENDIAN,
 248     .impl = {
 249         .min_access_size = 4,
 250         .max_access_size = 4,
 251     },
 252 };
 253
 254 static void ivshmem_vector_notify(void *opaque)
 255 {
 256     MSIVector *entry = opaque;
 257     PCIDevice *pdev = entry->pdev;
 258     IVShmemState *s = IVSHMEM_COMMON(pdev);
 259     int vector = entry - s->msi_vectors;
 260     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 261
 262     if (!event_notifier_test_and_clear(n)) {
 263         return;
 264     }
 265
 266     IVSHMEM_DPRINTF("interrupt on vector %p %d\n", pdev, vector);
 267     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 268         if (msix_enabled(pdev)) {
 269             msix_notify(pdev, vector);
 270         }
 271     } else {
 272         ivshmem_IntrStatus_write(s, 1);
 273     }
 274 }
 275
 276 static int ivshmem_vector_unmask(PCIDevice *dev, unsigned vector,
 277                                  MSIMessage msg)
 278 {
 279     IVShmemState *s = IVSHMEM_COMMON(dev);
 280     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 281     MSIVector *v = &s->msi_vectors[vector];
 282     int ret;
 283
 284     IVSHMEM_DPRINTF("vector unmask %p %d\n", dev, vector);
 285     if (!v->pdev) {
 286         error_report("ivshmem: vector %d route does not exist", vector);
 287         return -EINVAL;
 288     }
 289     assert(!v->unmasked);
 290
 291     ret = kvm_irqchip_update_msi_route(kvm_state, v->virq, msg, dev);
 292     if (ret < 0) {
 293         return ret;
 294     }
 295     kvm_irqchip_commit_routes(kvm_state);
 296
 297     ret = kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, n, NULL, v->virq);
 298     if (ret < 0) {
 299         return ret;
 300     }
 301     v->unmasked = true;
 302
 303     return 0;
 304 }
 305
 306 static void ivshmem_vector_mask(PCIDevice *dev, unsigned vector)
 307 {
 308     IVShmemState *s = IVSHMEM_COMMON(dev);
 309     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 310     MSIVector *v = &s->msi_vectors[vector];
 311     int ret;
 312
 313     IVSHMEM_DPRINTF("vector mask %p %d\n", dev, vector);
 314     if (!v->pdev) {
 315         error_report("ivshmem: vector %d route does not exist", vector);
 316         return;
 317     }
 318     assert(v->unmasked);
 319
 320     ret = kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state, n, v->virq);
 321     if (ret < 0) {
 322         error_report("remove_irqfd_notifier_gsi failed");
 323         return;
 324     }
 325     v->unmasked = false;
 326 }
 327
 328 static void ivshmem_vector_poll(PCIDevice *dev,
 329                                 unsigned int vector_start,
 330                                 unsigned int vector_end)
 331 {
 332     IVShmemState *s = IVSHMEM_COMMON(dev);
 333     unsigned int vector;
 334
 335     IVSHMEM_DPRINTF("vector poll %p %d-%d\n", dev, vector_start, vector_end);
 336
 337     vector_end = MIN(vector_end, s->vectors);
 338
 339     for (vector = vector_start; vector < vector_end; vector++) {
 340         EventNotifier *notifier = &s->peers[s->vm_id].eventfds[vector];
 341
 342         if (!msix_is_masked(dev, vector)) {
 343             continue;
 344         }
 345
 346         if (event_notifier_test_and_clear(notifier)) {
 347             msix_set_pending(dev, vector);
 348         }
 349     }
 350 }
 351
 352 static void watch_vector_notifier(IVShmemState *s, EventNotifier *n,
 353                                  int vector)
 354 {
 355     int eventfd = event_notifier_get_fd(n);
 356
 357     assert(!s->msi_vectors[vector].pdev);
 358     s->msi_vectors[vector].pdev = PCI_DEVICE(s);
 359
 360     qemu_set_fd_handler(eventfd, ivshmem_vector_notify,
 361                         NULL, &s->msi_vectors[vector]);
 362 }
 363
 364 static void ivshmem_add_eventfd(IVShmemState *s, int posn, int i)
 365 {
 366     memory_region_add_eventfd(&s->ivshmem_mmio,
 367                               DOORBELL,
 368                               4,
 369                               true,
 370                               (posn << 16) | i,
 371                               &s->peers[posn].eventfds[i]);
 372 }
 373
 374 static void ivshmem_del_eventfd(IVShmemState *s, int posn, int i)
 375 {
 376     memory_region_del_eventfd(&s->ivshmem_mmio,
 377                               DOORBELL,
 378                               4,
 379                               true,
 380                               (posn << 16) | i,
 381                               &s->peers[posn].eventfds[i]);
 382 }
 383
 384 static void close_peer_eventfds(IVShmemState *s, int posn)
 385 {
 386     int i, n;
 387
 388     assert(posn >= 0 && posn < s->nb_peers);
 389     n = s->peers[posn].nb_eventfds;
 390
 391     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
 392         memory_region_transaction_begin();
 393         for (i = 0; i < n; i++) {
 394             ivshmem_del_eventfd(s, posn, i);
 395         }
 396         memory_region_transaction_commit();
 397     }
 398
 399     for (i = 0; i < n; i++) {
 400         event_notifier_cleanup(&s->peers[posn].eventfds[i]);
 401     }
 402
 403     g_free(s->peers[posn].eventfds);
 404     s->peers[posn].nb_eventfds = 0;
 405 }
 406
 407 static void resize_peers(IVShmemState *s, int nb_peers)
 408 {
 409     int old_nb_peers = s->nb_peers;
 410     int i;
 411
 412     assert(nb_peers > old_nb_peers);
 413     IVSHMEM_DPRINTF("bumping storage to %d peers\n", nb_peers);
 414
 415     s->peers = g_realloc(s->peers, nb_peers * sizeof(Peer));
 416     s->nb_peers = nb_peers;
 417
 418     for (i = old_nb_peers; i < nb_peers; i++) {
 419         s->peers[i].eventfds = g_new0(EventNotifier, s->vectors);
 420         s->peers[i].nb_eventfds = 0;
 421     }
 422 }
 423
 424 static void ivshmem_add_kvm_msi_virq(IVShmemState *s, int vector,
 425                                      Error **errp)
 426 {
 427     PCIDevice *pdev = PCI_DEVICE(s);
 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     ret = kvm_irqchip_add_msi_route(kvm_state, vector, pdev);
 434     if (ret < 0) {
 435         error_setg(errp, "kvm_irqchip_add_msi_route failed");
 436         return;
 437     }
 438
 439     s->msi_vectors[vector].virq = ret;
 440     s->msi_vectors[vector].pdev = pdev;
 441 }
 442
 443 static void setup_interrupt(IVShmemState *s, int vector, Error **errp)
 444 {
 445     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 446     bool with_irqfd = kvm_msi_via_irqfd_enabled() &&
 447         ivshmem_has_feature(s, IVSHMEM_MSI);
 448     PCIDevice *pdev = PCI_DEVICE(s);
 449     Error *err = NULL;
 450
 451     IVSHMEM_DPRINTF("setting up interrupt for vector: %d\n", vector);
 452
 453     if (!with_irqfd) {
 454         IVSHMEM_DPRINTF("with eventfd\n");
 455         watch_vector_notifier(s, n, vector);
 456     } else if (msix_enabled(pdev)) {
 457         IVSHMEM_DPRINTF("with irqfd\n");
 458         ivshmem_add_kvm_msi_virq(s, vector, &err);
 459         if (err) {
 460             error_propagate(errp, err);
 461             return;
 462         }
 463
 464         if (!msix_is_masked(pdev, vector)) {
 465             kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, n, NULL,
 466                                                s->msi_vectors[vector].virq);
 467             /* TODO handle error */
 468         }
 469     } else {
 470         /* it will be delayed until msix is enabled, in write_config */
 471         IVSHMEM_DPRINTF("with irqfd, delayed until msix enabled\n");
 472     }
 473 }
 474
 475 static void process_msg_shmem(IVShmemState *s, int fd, Error **errp)
 476 {
 477     Error *local_err = NULL;
 478     struct stat buf;
 479     size_t size;
 480
 481     if (s->ivshmem_bar2) {
 482         error_setg(errp, "server sent unexpected shared memory message");
 483         close(fd);
 484         return;
 485     }
 486
 487     if (fstat(fd, &buf) < 0) {
 488         error_setg_errno(errp, errno,
 489             "can't determine size of shared memory sent by server");
 490         close(fd);
 491         return;
 492     }
 493
 494     size = buf.st_size;
 495
 496     /* mmap the region and map into the BAR2 */
 497     memory_region_init_ram_from_fd(&s->server_bar2, OBJECT(s),
 498                                    "ivshmem.bar2", size, true, fd, 0,
 499                                    &local_err);
 500     if (local_err) {
 501         error_propagate(errp, local_err);
 502         return;
 503     }
 504
 505     s->ivshmem_bar2 = &s->server_bar2;
 506 }
 507
 508 static void process_msg_disconnect(IVShmemState *s, uint16_t posn,
 509                                    Error **errp)
 510 {
 511     IVSHMEM_DPRINTF("posn %d has gone away\n", posn);
 512     if (posn >= s->nb_peers || posn == s->vm_id) {
 513         error_setg(errp, "invalid peer %d", posn);
 514         return;
 515     }
 516     close_peer_eventfds(s, posn);
 517 }
 518
 519 static void process_msg_connect(IVShmemState *s, uint16_t posn, int fd,
 520                                 Error **errp)
 521 {
 522     Peer *peer = &s->peers[posn];
 523     int vector;
 524
 525     /*
 526      * The N-th connect message for this peer comes with the file
 527      * descriptor for vector N-1.  Count messages to find the vector.
 528      */
 529     if (peer->nb_eventfds >= s->vectors) {
 530         error_setg(errp, "Too many eventfd received, device has %d vectors",
 531                    s->vectors);
 532         close(fd);
 533         return;
 534     }
 535     vector = peer->nb_eventfds++;
 536
 537     IVSHMEM_DPRINTF("eventfds[%d][%d] = %d\n", posn, vector, fd);
 538     event_notifier_init_fd(&peer->eventfds[vector], fd);
 539     fcntl_setfl(fd, O_NONBLOCK); /* msix/irqfd poll non block */
 540
 541     if (posn == s->vm_id) {
 542         setup_interrupt(s, vector, errp);
 543         /* TODO do we need to handle the error? */
 544     }
 545
 546     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
 547         ivshmem_add_eventfd(s, posn, vector);
 548     }
 549 }
 550
 551 static void process_msg(IVShmemState *s, int64_t msg, int fd, Error **errp)
 552 {
 553     IVSHMEM_DPRINTF("posn is %" PRId64 ", fd is %d\n", msg, fd);
 554
 555     if (msg < -1 || msg > IVSHMEM_MAX_PEERS) {
 556         error_setg(errp, "server sent invalid message %" PRId64, msg);
 557         close(fd);
 558         return;
 559     }
 560
 561     if (msg == -1) {
 562         process_msg_shmem(s, fd, errp);
 563         return;
 564     }
 565
 566     if (msg >= s->nb_peers) {
 567         resize_peers(s, msg + 1);
 568     }
 569
 570     if (fd >= 0) {
 571         process_msg_connect(s, msg, fd, errp);
 572     } else {
 573         process_msg_disconnect(s, msg, errp);
 574     }
 575 }
 576
 577 static int ivshmem_can_receive(void *opaque)
 578 {
 579     IVShmemState *s = opaque;
 580
 581     assert(s->msg_buffered_bytes < sizeof(s->msg_buf));
 582     return sizeof(s->msg_buf) - s->msg_buffered_bytes;
 583 }
 584
 585 static void ivshmem_read(void *opaque, const uint8_t *buf, int size)
 586 {
 587     IVShmemState *s = opaque;
 588     Error *err = NULL;
 589     int fd;
 590     int64_t msg;
 591
 592     assert(size >= 0 && s->msg_buffered_bytes + size <= sizeof(s->msg_buf));
 593     memcpy((unsigned char *)&s->msg_buf + s->msg_buffered_bytes, buf, size);
 594     s->msg_buffered_bytes += size;
 595     if (s->msg_buffered_bytes < sizeof(s->msg_buf)) {
 596         return;
 597     }
 598     msg = le64_to_cpu(s->msg_buf);
 599     s->msg_buffered_bytes = 0;
 600
 601     fd = qemu_chr_fe_get_msgfd(&s->server_chr);
 602
 603     process_msg(s, msg, fd, &err);
 604     if (err) {
 605         error_report_err(err);
 606     }
 607 }
 608
 609 static int64_t ivshmem_recv_msg(IVShmemState *s, int *pfd, Error **errp)
 610 {
 611     int64_t msg;
 612     int n, ret;
 613
 614     n = 0;
 615     do {
 616         ret = qemu_chr_fe_read_all(&s->server_chr, (uint8_t *)&msg + n,
 617                                    sizeof(msg) - n);
 618         if (ret < 0) {
 619             if (ret == -EINTR) {
 620                 continue;
 621             }
 622             error_setg_errno(errp, -ret, "read from server failed");
 623             return INT64_MIN;
 624         }
 625         n += ret;
 626     } while (n < sizeof(msg));
 627
 628     *pfd = qemu_chr_fe_get_msgfd(&s->server_chr);
 629     return le64_to_cpu(msg);
 630 }
 631
 632 static void ivshmem_recv_setup(IVShmemState *s, Error **errp)
 633 {
 634     Error *err = NULL;
 635     int64_t msg;
 636     int fd;
 637
 638     msg = ivshmem_recv_msg(s, &fd, &err);
 639     if (err) {
 640         error_propagate(errp, err);
 641         return;
 642     }
 643     if (msg != IVSHMEM_PROTOCOL_VERSION) {
 644         error_setg(errp, "server sent version %" PRId64 ", expecting %d",
 645                    msg, IVSHMEM_PROTOCOL_VERSION);
 646         return;
 647     }
 648     if (fd != -1) {
 649         error_setg(errp, "server sent invalid version message");
 650         return;
 651     }
 652
 653     /*
 654      * ivshmem-server sends the remaining initial messages in a fixed
 655      * order, but the device has always accepted them in any order.
 656      * Stay as compatible as practical, just in case people use
 657      * servers that behave differently.
 658      */
 659
 660     /*
 661      * ivshmem_device_spec.txt has always required the ID message
 662      * right here, and ivshmem-server has always complied.  However,
 663      * older versions of the device accepted it out of order, but
 664      * broke when an interrupt setup message arrived before it.
 665      */
 666     msg = ivshmem_recv_msg(s, &fd, &err);
 667     if (err) {
 668         error_propagate(errp, err);
 669         return;
 670     }
 671     if (fd != -1 || msg < 0 || msg > IVSHMEM_MAX_PEERS) {
 672         error_setg(errp, "server sent invalid ID message");
 673         return;
 674     }
 675     s->vm_id = msg;
 676
 677     /*
 678      * Receive more messages until we got shared memory.
 679      */
 680     do {
 681         msg = ivshmem_recv_msg(s, &fd, &err);
 682         if (err) {
 683             error_propagate(errp, err);
 684             return;
 685         }
 686         process_msg(s, msg, fd, &err);
 687         if (err) {
 688             error_propagate(errp, err);
 689             return;
 690         }
 691     } while (msg != -1);
 692
 693     /*
 694      * This function must either map the shared memory or fail.  The
 695      * loop above ensures that: it terminates normally only after it
 696      * successfully processed the server's shared memory message.
 697      * Assert that actually mapped the shared memory:
 698      */
 699     assert(s->ivshmem_bar2);
 700 }
 701
 702 /* Select the MSI-X vectors used by device.
 703  * ivshmem maps events to vectors statically, so
 704  * we just enable all vectors on init and after reset. */
 705 static void ivshmem_msix_vector_use(IVShmemState *s)
 706 {
 707     PCIDevice *d = PCI_DEVICE(s);
 708     int i;
 709
 710     for (i = 0; i < s->vectors; i++) {
 711         msix_vector_use(d, i);
 712     }
 713 }
 714
 715 static void ivshmem_disable_irqfd(IVShmemState *s);
 716
 717 static void ivshmem_reset(DeviceState *d)
 718 {
 719     IVShmemState *s = IVSHMEM_COMMON(d);
 720
 721     ivshmem_disable_irqfd(s);
 722
 723     s->intrstatus = 0;
 724     s->intrmask = 0;
 725     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 726         ivshmem_msix_vector_use(s);
 727     }
 728 }
 729
 730 static int ivshmem_setup_interrupts(IVShmemState *s, Error **errp)
 731 {
 732     /* allocate QEMU callback data for receiving interrupts */
 733     s->msi_vectors = g_malloc0(s->vectors * sizeof(MSIVector));
 734
 735     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 736         if (msix_init_exclusive_bar(PCI_DEVICE(s), s->vectors, 1, errp)) {
 737             return -1;
 738         }
 739
 740         IVSHMEM_DPRINTF("msix initialized (%d vectors)\n", s->vectors);
 741         ivshmem_msix_vector_use(s);
 742     }
 743
 744     return 0;
 745 }
 746
 747 static void ivshmem_remove_kvm_msi_virq(IVShmemState *s, int vector)
 748 {
 749     IVSHMEM_DPRINTF("ivshmem_remove_kvm_msi_virq vector:%d\n", vector);
 750
 751     if (s->msi_vectors[vector].pdev == NULL) {
 752         return;
 753     }
 754
 755     /* it was cleaned when masked in the frontend. */
 756     kvm_irqchip_release_virq(kvm_state, s->msi_vectors[vector].virq);
 757
 758     s->msi_vectors[vector].pdev = NULL;
 759 }
 760
 761 static void ivshmem_enable_irqfd(IVShmemState *s)
 762 {
 763     PCIDevice *pdev = PCI_DEVICE(s);
 764     int i;
 765
 766     for (i = 0; i < s->peers[s->vm_id].nb_eventfds; i++) {
 767         Error *err = NULL;
 768
 769         ivshmem_add_kvm_msi_virq(s, i, &err);
 770         if (err) {
 771             error_report_err(err);
 772             goto undo;
 773         }
 774     }
 775
 776     if (msix_set_vector_notifiers(pdev,
 777                                   ivshmem_vector_unmask,
 778                                   ivshmem_vector_mask,
 779                                   ivshmem_vector_poll)) {
 780         error_report("ivshmem: msix_set_vector_notifiers failed");
 781         goto undo;
 782     }
 783     return;
 784
 785 undo:
 786     while (--i >= 0) {
 787         ivshmem_remove_kvm_msi_virq(s, i);
 788     }
 789 }
 790
 791 static void ivshmem_disable_irqfd(IVShmemState *s)
 792 {
 793     PCIDevice *pdev = PCI_DEVICE(s);
 794     int i;
 795
 796     if (!pdev->msix_vector_use_notifier) {
 797         return;
 798     }
 799
 800     msix_unset_vector_notifiers(pdev);
 801
 802     for (i = 0; i < s->peers[s->vm_id].nb_eventfds; i++) {
 803         /*
 804          * MSI-X is already disabled here so msix_unset_vector_notifiers()
 805          * didn't call our release notifier.  Do it now to keep our masks and
 806          * unmasks balanced.
 807          */
 808         if (s->msi_vectors[i].unmasked) {
 809             ivshmem_vector_mask(pdev, i);
 810         }
 811         ivshmem_remove_kvm_msi_virq(s, i);
 812     }
 813
 814 }
 815
 816 static void ivshmem_write_config(PCIDevice *pdev, uint32_t address,
 817                                  uint32_t val, int len)
 818 {
 819     IVShmemState *s = IVSHMEM_COMMON(pdev);
 820     int is_enabled, was_enabled = msix_enabled(pdev);
 821
 822     pci_default_write_config(pdev, address, val, len);
 823     is_enabled = msix_enabled(pdev);
 824
 825     if (kvm_msi_via_irqfd_enabled()) {
 826         if (!was_enabled && is_enabled) {
 827             ivshmem_enable_irqfd(s);
 828         } else if (was_enabled && !is_enabled) {
 829             ivshmem_disable_irqfd(s);
 830         }
 831     }
 832 }
 833
 834 static void ivshmem_common_realize(PCIDevice *dev, Error **errp)
 835 {
 836     IVShmemState *s = IVSHMEM_COMMON(dev);
 837     Error *err = NULL;
 838     uint8_t *pci_conf;
 839
 840     /* IRQFD requires MSI */
 841     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD) &&
 842         !ivshmem_has_feature(s, IVSHMEM_MSI)) {
 843         error_setg(errp, "ioeventfd/irqfd requires MSI");
 844         return;
 845     }
 846
 847     pci_conf = dev->config;
 848     pci_conf[PCI_COMMAND] = PCI_COMMAND_IO | PCI_COMMAND_MEMORY;
 849
 850     memory_region_init_io(&s->ivshmem_mmio, OBJECT(s), &ivshmem_mmio_ops, s,
 851                           "ivshmem-mmio", IVSHMEM_REG_BAR_SIZE);
 852
 853     /* region for registers*/
 854     pci_register_bar(dev, 0, PCI_BASE_ADDRESS_SPACE_MEMORY,
 855                      &s->ivshmem_mmio);
 856
 857     if (s->hostmem != NULL) {
 858         IVSHMEM_DPRINTF("using hostmem\n");
 859
 860         s->ivshmem_bar2 = host_memory_backend_get_memory(s->hostmem);
 861         host_memory_backend_set_mapped(s->hostmem, true);
 862     } else {
 863         Chardev *chr = qemu_chr_fe_get_driver(&s->server_chr);
 864         assert(chr);
 865
 866         IVSHMEM_DPRINTF("using shared memory server (socket = %s)\n",
 867                         chr->filename);
 868
 869         /* we allocate enough space for 16 peers and grow as needed */
 870         resize_peers(s, 16);
 871
 872         /*
 873          * Receive setup messages from server synchronously.
 874          * Older versions did it asynchronously, but that creates a
 875          * number of entertaining race conditions.
 876          */
 877         ivshmem_recv_setup(s, &err);
 878         if (err) {
 879             error_propagate(errp, err);
 880             return;
 881         }
 882
 883         if (s->master == ON_OFF_AUTO_ON && s->vm_id != 0) {
 884             error_setg(errp,
 885                        "master must connect to the server before any peers");
 886             return;
 887         }
 888
 889         qemu_chr_fe_set_handlers(&s->server_chr, ivshmem_can_receive,
 890                                  ivshmem_read, NULL, NULL, s, NULL, true);
 891
 892         if (ivshmem_setup_interrupts(s, errp) < 0) {
 893             error_prepend(errp, "Failed to initialize interrupts: ");
 894             return;
 895         }
 896     }
 897
 898     if (s->master == ON_OFF_AUTO_AUTO) {
 899         s->master = s->vm_id == 0 ? ON_OFF_AUTO_ON : ON_OFF_AUTO_OFF;
 900     }
 901
 902     if (!ivshmem_is_master(s)) {
 903         error_setg(&s->migration_blocker,
 904                    "Migration is disabled when using feature 'peer mode' in device 'ivshmem'");
 905         if (migrate_add_blocker(s->migration_blocker, errp) < 0) {
 906             error_free(s->migration_blocker);
 907             return;
 908         }
 909     }
 910
 911     vmstate_register_ram(s->ivshmem_bar2, DEVICE(s));
 912     pci_register_bar(PCI_DEVICE(s), 2,
 913                      PCI_BASE_ADDRESS_SPACE_MEMORY |
 914                      PCI_BASE_ADDRESS_MEM_PREFETCH |
 915                      PCI_BASE_ADDRESS_MEM_TYPE_64,
 916                      s->ivshmem_bar2);
 917 }
 918
 919 static void ivshmem_exit(PCIDevice *dev)
 920 {
 921     IVShmemState *s = IVSHMEM_COMMON(dev);
 922     int i;
 923
 924     if (s->migration_blocker) {
 925         migrate_del_blocker(s->migration_blocker);
 926         error_free(s->migration_blocker);
 927     }
 928
 929     if (memory_region_is_mapped(s->ivshmem_bar2)) {
 930         if (!s->hostmem) {
 931             void *addr = memory_region_get_ram_ptr(s->ivshmem_bar2);
 932             int fd;
 933
 934             if (munmap(addr, memory_region_size(s->ivshmem_bar2) == -1)) {
 935                 error_report("Failed to munmap shared memory %s",
 936                              strerror(errno));
 937             }
 938
 939             fd = memory_region_get_fd(s->ivshmem_bar2);
 940             close(fd);
 941         }
 942
 943         vmstate_unregister_ram(s->ivshmem_bar2, DEVICE(dev));
 944     }
 945
 946     if (s->hostmem) {
 947         host_memory_backend_set_mapped(s->hostmem, false);
 948     }
 949
 950     if (s->peers) {
 951         for (i = 0; i < s->nb_peers; i++) {
 952             close_peer_eventfds(s, i);
 953         }
 954         g_free(s->peers);
 955     }
 956
 957     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 958         msix_uninit_exclusive_bar(dev);
 959     }
 960
 961     g_free(s->msi_vectors);
 962 }
 963
 964 static int ivshmem_pre_load(void *opaque)
 965 {
 966     IVShmemState *s = opaque;
 967
 968     if (!ivshmem_is_master(s)) {
 969         error_report("'peer' devices are not migratable");
 970         return -EINVAL;
 971     }
 972
 973     return 0;
 974 }
 975
 976 static int ivshmem_post_load(void *opaque, int version_id)
 977 {
 978     IVShmemState *s = opaque;
 979
 980     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 981         ivshmem_msix_vector_use(s);
 982     }
 983     return 0;
 984 }
 985
 986 static void ivshmem_common_class_init(ObjectClass *klass, void *data)
 987 {
 988     DeviceClass *dc = DEVICE_CLASS(klass);
 989     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
 990
 991     k->realize = ivshmem_common_realize;
 992     k->exit = ivshmem_exit;
 993     k->config_write = ivshmem_write_config;
 994     k->vendor_id = PCI_VENDOR_ID_IVSHMEM;
 995     k->device_id = PCI_DEVICE_ID_IVSHMEM;
 996     k->class_id = PCI_CLASS_MEMORY_RAM;
 997     k->revision = 1;
 998     dc->reset = ivshmem_reset;
 999     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
1000     dc->desc = "Inter-VM shared memory";
1001 }
1002
1003 static const TypeInfo ivshmem_common_info = {
1004     .name          = TYPE_IVSHMEM_COMMON,
1005     .parent        = TYPE_PCI_DEVICE,
1006     .instance_size = sizeof(IVShmemState),
1007     .abstract      = true,
1008     .class_init    = ivshmem_common_class_init,
1009     .interfaces = (InterfaceInfo[]) {
1010         { INTERFACE_CONVENTIONAL_PCI_DEVICE },
1011         { },
1012     },
1013 };
1014
1015 static const VMStateDescription ivshmem_plain_vmsd = {
1016     .name = TYPE_IVSHMEM_PLAIN,
1017     .version_id = 0,
1018     .minimum_version_id = 0,
1019     .pre_load = ivshmem_pre_load,
1020     .post_load = ivshmem_post_load,
1021     .fields = (VMStateField[]) {
1022         VMSTATE_PCI_DEVICE(parent_obj, IVShmemState),
1023         VMSTATE_UINT32(intrstatus, IVShmemState),
1024         VMSTATE_UINT32(intrmask, IVShmemState),
1025         VMSTATE_END_OF_LIST()
1026     },
1027 };
1028
1029 static Property ivshmem_plain_properties[] = {
1030     DEFINE_PROP_ON_OFF_AUTO("master", IVShmemState, master, ON_OFF_AUTO_OFF),
1031     DEFINE_PROP_LINK("memdev", IVShmemState, hostmem, TYPE_MEMORY_BACKEND,
1032                      HostMemoryBackend *),
1033     DEFINE_PROP_END_OF_LIST(),
1034 };
1035
1036 static void ivshmem_plain_realize(PCIDevice *dev, Error **errp)
1037 {
1038     IVShmemState *s = IVSHMEM_COMMON(dev);
1039
1040     if (!s->hostmem) {
1041         error_setg(errp, "You must specify a 'memdev'");
1042         return;
1043     } else if (host_memory_backend_is_mapped(s->hostmem)) {
1044         error_setg(errp, "can't use already busy memdev: %s",
1045                    object_get_canonical_path_component(OBJECT(s->hostmem)));
1046         return;
1047     }
1048
1049     ivshmem_common_realize(dev, errp);
1050 }
1051
1052 static void ivshmem_plain_class_init(ObjectClass *klass, void *data)
1053 {
1054     DeviceClass *dc = DEVICE_CLASS(klass);
1055     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
1056
1057     k->realize = ivshmem_plain_realize;
1058     device_class_set_props(dc, ivshmem_plain_properties);
1059     dc->vmsd = &ivshmem_plain_vmsd;
1060 }
1061
1062 static const TypeInfo ivshmem_plain_info = {
1063     .name          = TYPE_IVSHMEM_PLAIN,
1064     .parent        = TYPE_IVSHMEM_COMMON,
1065     .instance_size = sizeof(IVShmemState),
1066     .class_init    = ivshmem_plain_class_init,
1067 };
1068
1069 static const VMStateDescription ivshmem_doorbell_vmsd = {
1070     .name = TYPE_IVSHMEM_DOORBELL,
1071     .version_id = 0,
1072     .minimum_version_id = 0,
1073     .pre_load = ivshmem_pre_load,
1074     .post_load = ivshmem_post_load,
1075     .fields = (VMStateField[]) {
1076         VMSTATE_PCI_DEVICE(parent_obj, IVShmemState),
1077         VMSTATE_MSIX(parent_obj, IVShmemState),
1078         VMSTATE_UINT32(intrstatus, IVShmemState),
1079         VMSTATE_UINT32(intrmask, IVShmemState),
1080         VMSTATE_END_OF_LIST()
1081     },
1082 };
1083
1084 static Property ivshmem_doorbell_properties[] = {
1085     DEFINE_PROP_CHR("chardev", IVShmemState, server_chr),
1086     DEFINE_PROP_UINT32("vectors", IVShmemState, vectors, 1),
1087     DEFINE_PROP_BIT("ioeventfd", IVShmemState, features, IVSHMEM_IOEVENTFD,
1088                     true),
1089     DEFINE_PROP_ON_OFF_AUTO("master", IVShmemState, master, ON_OFF_AUTO_OFF),
1090     DEFINE_PROP_END_OF_LIST(),
1091 };
1092
1093 static void ivshmem_doorbell_init(Object *obj)
1094 {
1095     IVShmemState *s = IVSHMEM_DOORBELL(obj);
1096
1097     s->features |= (1 << IVSHMEM_MSI);
1098 }
1099
1100 static void ivshmem_doorbell_realize(PCIDevice *dev, Error **errp)
1101 {
1102     IVShmemState *s = IVSHMEM_COMMON(dev);
1103
1104     if (!qemu_chr_fe_backend_connected(&s->server_chr)) {
1105         error_setg(errp, "You must specify a 'chardev'");
1106         return;
1107     }
1108
1109     ivshmem_common_realize(dev, errp);
1110 }
1111
1112 static void ivshmem_doorbell_class_init(ObjectClass *klass, void *data)
1113 {
1114     DeviceClass *dc = DEVICE_CLASS(klass);
1115     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
1116
1117     k->realize = ivshmem_doorbell_realize;
1118     device_class_set_props(dc, ivshmem_doorbell_properties);
1119     dc->vmsd = &ivshmem_doorbell_vmsd;
1120 }
1121
1122 static const TypeInfo ivshmem_doorbell_info = {
1123     .name          = TYPE_IVSHMEM_DOORBELL,
1124     .parent        = TYPE_IVSHMEM_COMMON,
1125     .instance_size = sizeof(IVShmemState),
1126     .instance_init = ivshmem_doorbell_init,
1127     .class_init    = ivshmem_doorbell_class_init,
1128 };
1129
1130 static void ivshmem_register_types(void)
1131 {
1132     type_register_static(&ivshmem_common_info);
1133     type_register_static(&ivshmem_plain_info);
1134     type_register_static(&ivshmem_doorbell_info);
1135 }
1136
1137 type_init(ivshmem_register_types)