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 #include "qemu/osdep.h"
  20 #include "hw/hw.h"
  21 #include "hw/i386/pc.h"
  22 #include "hw/pci/pci.h"
  23 #include "hw/pci/msi.h"
  24 #include "hw/pci/msix.h"
  25 #include "sysemu/kvm.h"
  26 #include "migration/migration.h"
  27 #include "qemu/error-report.h"
  28 #include "qemu/event_notifier.h"
  29 #include "qom/object_interfaces.h"
  30 #include "sysemu/char.h"
  31 #include "sysemu/hostmem.h"
  32 #include "qapi/visitor.h"
  33 #include "exec/ram_addr.h"
  34
  35 #include "hw/misc/ivshmem.h"
  36
  37 #include <sys/mman.h>
  38
  39 #define PCI_VENDOR_ID_IVSHMEM   PCI_VENDOR_ID_REDHAT_QUMRANET
  40 #define PCI_DEVICE_ID_IVSHMEM   0x1110
  41
  42 #define IVSHMEM_MAX_PEERS UINT16_MAX
  43 #define IVSHMEM_IOEVENTFD   0
  44 #define IVSHMEM_MSI     1
  45
  46 #define IVSHMEM_PEER    0
  47 #define IVSHMEM_MASTER  1
  48
  49 #define IVSHMEM_REG_BAR_SIZE 0x100
  50
  51 #define IVSHMEM_DEBUG 0
  52 #define IVSHMEM_DPRINTF(fmt, ...)                       \
  53     do {                                                \
  54         if (IVSHMEM_DEBUG) {                            \
  55             printf("IVSHMEM: " fmt, ## __VA_ARGS__);    \
  56         }                                               \
  57     } while (0)
  58
  59 #define TYPE_IVSHMEM "ivshmem"
  60 #define IVSHMEM(obj) \
  61     OBJECT_CHECK(IVShmemState, (obj), TYPE_IVSHMEM)
  62
  63 typedef struct Peer {
  64     int nb_eventfds;
  65     EventNotifier *eventfds;
  66 } Peer;
  67
  68 typedef struct MSIVector {
  69     PCIDevice *pdev;
  70     int virq;
  71 } MSIVector;
  72
  73 typedef struct IVShmemState {
  74     /*< private >*/
  75     PCIDevice parent_obj;
  76     /*< public >*/
  77
  78     HostMemoryBackend *hostmem;
  79     uint32_t intrmask;
  80     uint32_t intrstatus;
  81
  82     CharDriverState *server_chr;
  83     MemoryRegion ivshmem_mmio;
  84
  85     MemoryRegion *ivshmem_bar2; /* BAR 2 (shared memory) */
  86     MemoryRegion server_bar2;   /* used with server_chr */
  87     size_t ivshmem_size; /* size of shared memory region */
  88     uint32_t ivshmem_64bit;
  89
  90     Peer *peers;
  91     int nb_peers;               /* space in @peers[] */
  92
  93     int vm_id;
  94     uint32_t vectors;
  95     uint32_t features;
  96     MSIVector *msi_vectors;
  97     uint64_t msg_buf;           /* buffer for receiving server messages */
  98     int msg_buffered_bytes;     /* #bytes in @msg_buf */
  99
 100     Error *migration_blocker;
 101
 102     char * shmobj;
 103     char * sizearg;
 104     char * role;
 105     int role_val;   /* scalar to avoid multiple string comparisons */
 106 } IVShmemState;
 107
 108 /* registers for the Inter-VM shared memory device */
 109 enum ivshmem_registers {
 110     INTRMASK = 0,
 111     INTRSTATUS = 4,
 112     IVPOSITION = 8,
 113     DOORBELL = 12,
 114 };
 115
 116 static inline uint32_t ivshmem_has_feature(IVShmemState *ivs,
 117                                                     unsigned int feature) {
 118     return (ivs->features & (1 << feature));
 119 }
 120
 121 static void ivshmem_update_irq(IVShmemState *s)
 122 {
 123     PCIDevice *d = PCI_DEVICE(s);
 124     uint32_t isr = s->intrstatus & s->intrmask;
 125
 126     /* No INTx with msi=on, whether the guest enabled MSI-X or not */
 127     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 128         return;
 129     }
 130
 131     /* don't print ISR resets */
 132     if (isr) {
 133         IVSHMEM_DPRINTF("Set IRQ to %d (%04x %04x)\n",
 134                         isr ? 1 : 0, s->intrstatus, s->intrmask);
 135     }
 136
 137     pci_set_irq(d, isr != 0);
 138 }
 139
 140 static void ivshmem_IntrMask_write(IVShmemState *s, uint32_t val)
 141 {
 142     IVSHMEM_DPRINTF("IntrMask write(w) val = 0x%04x\n", val);
 143
 144     s->intrmask = val;
 145     ivshmem_update_irq(s);
 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     ivshmem_update_irq(s);
 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     ivshmem_update_irq(s);
 171     return ret;
 172 }
 173
 174 static void ivshmem_io_write(void *opaque, hwaddr addr,
 175                              uint64_t val, unsigned size)
 176 {
 177     IVShmemState *s = opaque;
 178
 179     uint16_t dest = val >> 16;
 180     uint16_t vector = val & 0xff;
 181
 182     addr &= 0xfc;
 183
 184     IVSHMEM_DPRINTF("writing to addr " TARGET_FMT_plx "\n", addr);
 185     switch (addr)
 186     {
 187         case INTRMASK:
 188             ivshmem_IntrMask_write(s, val);
 189             break;
 190
 191         case INTRSTATUS:
 192             ivshmem_IntrStatus_write(s, val);
 193             break;
 194
 195         case DOORBELL:
 196             /* check that dest VM ID is reasonable */
 197             if (dest >= s->nb_peers) {
 198                 IVSHMEM_DPRINTF("Invalid destination VM ID (%d)\n", dest);
 199                 break;
 200             }
 201
 202             /* check doorbell range */
 203             if (vector < s->peers[dest].nb_eventfds) {
 204                 IVSHMEM_DPRINTF("Notifying VM %d on vector %d\n", dest, vector);
 205                 event_notifier_set(&s->peers[dest].eventfds[vector]);
 206             } else {
 207                 IVSHMEM_DPRINTF("Invalid destination vector %d on VM %d\n",
 208                                 vector, dest);
 209             }
 210             break;
 211         default:
 212             IVSHMEM_DPRINTF("Unhandled write " TARGET_FMT_plx "\n", addr);
 213     }
 214 }
 215
 216 static uint64_t ivshmem_io_read(void *opaque, hwaddr addr,
 217                                 unsigned size)
 218 {
 219
 220     IVShmemState *s = opaque;
 221     uint32_t ret;
 222
 223     switch (addr)
 224     {
 225         case INTRMASK:
 226             ret = ivshmem_IntrMask_read(s);
 227             break;
 228
 229         case INTRSTATUS:
 230             ret = ivshmem_IntrStatus_read(s);
 231             break;
 232
 233         case IVPOSITION:
 234             ret = s->vm_id;
 235             break;
 236
 237         default:
 238             IVSHMEM_DPRINTF("why are we reading " TARGET_FMT_plx "\n", addr);
 239             ret = 0;
 240     }
 241
 242     return ret;
 243 }
 244
 245 static const MemoryRegionOps ivshmem_mmio_ops = {
 246     .read = ivshmem_io_read,
 247     .write = ivshmem_io_write,
 248     .endianness = DEVICE_NATIVE_ENDIAN,
 249     .impl = {
 250         .min_access_size = 4,
 251         .max_access_size = 4,
 252     },
 253 };
 254
 255 static void ivshmem_vector_notify(void *opaque)
 256 {
 257     MSIVector *entry = opaque;
 258     PCIDevice *pdev = entry->pdev;
 259     IVShmemState *s = IVSHMEM(pdev);
 260     int vector = entry - s->msi_vectors;
 261     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 262
 263     if (!event_notifier_test_and_clear(n)) {
 264         return;
 265     }
 266
 267     IVSHMEM_DPRINTF("interrupt on vector %p %d\n", pdev, vector);
 268     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 269         if (msix_enabled(pdev)) {
 270             msix_notify(pdev, vector);
 271         }
 272     } else {
 273         ivshmem_IntrStatus_write(s, 1);
 274     }
 275 }
 276
 277 static int ivshmem_vector_unmask(PCIDevice *dev, unsigned vector,
 278                                  MSIMessage msg)
 279 {
 280     IVShmemState *s = IVSHMEM(dev);
 281     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 282     MSIVector *v = &s->msi_vectors[vector];
 283     int ret;
 284
 285     IVSHMEM_DPRINTF("vector unmask %p %d\n", dev, vector);
 286
 287     ret = kvm_irqchip_update_msi_route(kvm_state, v->virq, msg, dev);
 288     if (ret < 0) {
 289         return ret;
 290     }
 291
 292     return kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, n, NULL, v->virq);
 293 }
 294
 295 static void ivshmem_vector_mask(PCIDevice *dev, unsigned vector)
 296 {
 297     IVShmemState *s = IVSHMEM(dev);
 298     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 299     int ret;
 300
 301     IVSHMEM_DPRINTF("vector mask %p %d\n", dev, vector);
 302
 303     ret = kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state, n,
 304                                                 s->msi_vectors[vector].virq);
 305     if (ret != 0) {
 306         error_report("remove_irqfd_notifier_gsi failed");
 307     }
 308 }
 309
 310 static void ivshmem_vector_poll(PCIDevice *dev,
 311                                 unsigned int vector_start,
 312                                 unsigned int vector_end)
 313 {
 314     IVShmemState *s = IVSHMEM(dev);
 315     unsigned int vector;
 316
 317     IVSHMEM_DPRINTF("vector poll %p %d-%d\n", dev, vector_start, vector_end);
 318
 319     vector_end = MIN(vector_end, s->vectors);
 320
 321     for (vector = vector_start; vector < vector_end; vector++) {
 322         EventNotifier *notifier = &s->peers[s->vm_id].eventfds[vector];
 323
 324         if (!msix_is_masked(dev, vector)) {
 325             continue;
 326         }
 327
 328         if (event_notifier_test_and_clear(notifier)) {
 329             msix_set_pending(dev, vector);
 330         }
 331     }
 332 }
 333
 334 static void watch_vector_notifier(IVShmemState *s, EventNotifier *n,
 335                                  int vector)
 336 {
 337     int eventfd = event_notifier_get_fd(n);
 338
 339     assert(!s->msi_vectors[vector].pdev);
 340     s->msi_vectors[vector].pdev = PCI_DEVICE(s);
 341
 342     qemu_set_fd_handler(eventfd, ivshmem_vector_notify,
 343                         NULL, &s->msi_vectors[vector]);
 344 }
 345
 346 static void ivshmem_add_eventfd(IVShmemState *s, int posn, int i)
 347 {
 348     memory_region_add_eventfd(&s->ivshmem_mmio,
 349                               DOORBELL,
 350                               4,
 351                               true,
 352                               (posn << 16) | i,
 353                               &s->peers[posn].eventfds[i]);
 354 }
 355
 356 static void ivshmem_del_eventfd(IVShmemState *s, int posn, int i)
 357 {
 358     memory_region_del_eventfd(&s->ivshmem_mmio,
 359                               DOORBELL,
 360                               4,
 361                               true,
 362                               (posn << 16) | i,
 363                               &s->peers[posn].eventfds[i]);
 364 }
 365
 366 static void close_peer_eventfds(IVShmemState *s, int posn)
 367 {
 368     int i, n;
 369
 370     assert(posn >= 0 && posn < s->nb_peers);
 371     n = s->peers[posn].nb_eventfds;
 372
 373     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
 374         memory_region_transaction_begin();
 375         for (i = 0; i < n; i++) {
 376             ivshmem_del_eventfd(s, posn, i);
 377         }
 378         memory_region_transaction_commit();
 379     }
 380
 381     for (i = 0; i < n; i++) {
 382         event_notifier_cleanup(&s->peers[posn].eventfds[i]);
 383     }
 384
 385     g_free(s->peers[posn].eventfds);
 386     s->peers[posn].nb_eventfds = 0;
 387 }
 388
 389 static void resize_peers(IVShmemState *s, int nb_peers)
 390 {
 391     int old_nb_peers = s->nb_peers;
 392     int i;
 393
 394     assert(nb_peers > old_nb_peers);
 395     IVSHMEM_DPRINTF("bumping storage to %d peers\n", nb_peers);
 396
 397     s->peers = g_realloc(s->peers, nb_peers * sizeof(Peer));
 398     s->nb_peers = nb_peers;
 399
 400     for (i = old_nb_peers; i < nb_peers; i++) {
 401         s->peers[i].eventfds = g_new0(EventNotifier, s->vectors);
 402         s->peers[i].nb_eventfds = 0;
 403     }
 404 }
 405
 406 static void ivshmem_add_kvm_msi_virq(IVShmemState *s, int vector,
 407                                      Error **errp)
 408 {
 409     PCIDevice *pdev = PCI_DEVICE(s);
 410     MSIMessage msg = msix_get_message(pdev, vector);
 411     int ret;
 412
 413     IVSHMEM_DPRINTF("ivshmem_add_kvm_msi_virq vector:%d\n", vector);
 414     assert(!s->msi_vectors[vector].pdev);
 415
 416     ret = kvm_irqchip_add_msi_route(kvm_state, msg, pdev);
 417     if (ret < 0) {
 418         error_setg(errp, "kvm_irqchip_add_msi_route failed");
 419         return;
 420     }
 421
 422     s->msi_vectors[vector].virq = ret;
 423     s->msi_vectors[vector].pdev = pdev;
 424 }
 425
 426 static void setup_interrupt(IVShmemState *s, int vector, Error **errp)
 427 {
 428     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 429     bool with_irqfd = kvm_msi_via_irqfd_enabled() &&
 430         ivshmem_has_feature(s, IVSHMEM_MSI);
 431     PCIDevice *pdev = PCI_DEVICE(s);
 432     Error *err = NULL;
 433
 434     IVSHMEM_DPRINTF("setting up interrupt for vector: %d\n", vector);
 435
 436     if (!with_irqfd) {
 437         IVSHMEM_DPRINTF("with eventfd\n");
 438         watch_vector_notifier(s, n, vector);
 439     } else if (msix_enabled(pdev)) {
 440         IVSHMEM_DPRINTF("with irqfd\n");
 441         ivshmem_add_kvm_msi_virq(s, vector, &err);
 442         if (err) {
 443             error_propagate(errp, err);
 444             return;
 445         }
 446
 447         if (!msix_is_masked(pdev, vector)) {
 448             kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, n, NULL,
 449                                                s->msi_vectors[vector].virq);
 450             /* TODO handle error */
 451         }
 452     } else {
 453         /* it will be delayed until msix is enabled, in write_config */
 454         IVSHMEM_DPRINTF("with irqfd, delayed until msix enabled\n");
 455     }
 456 }
 457
 458 static void process_msg_shmem(IVShmemState *s, int fd, Error **errp)
 459 {
 460     struct stat buf;
 461     void *ptr;
 462
 463     if (s->ivshmem_bar2) {
 464         error_setg(errp, "server sent unexpected shared memory message");
 465         close(fd);
 466         return;
 467     }
 468
 469     if (fstat(fd, &buf) < 0) {
 470         error_setg_errno(errp, errno,
 471             "can't determine size of shared memory sent by server");
 472         close(fd);
 473         return;
 474     }
 475
 476     if (s->ivshmem_size > buf.st_size) {
 477         error_setg(errp, "server sent only %zd bytes of shared memory",
 478                    (size_t)buf.st_size);
 479         close(fd);
 480         return;
 481     }
 482
 483     /* mmap the region and map into the BAR2 */
 484     ptr = mmap(0, s->ivshmem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 485     if (ptr == MAP_FAILED) {
 486         error_setg_errno(errp, errno, "Failed to mmap shared memory");
 487         close(fd);
 488         return;
 489     }
 490     memory_region_init_ram_ptr(&s->server_bar2, OBJECT(s),
 491                                "ivshmem.bar2", s->ivshmem_size, ptr);
 492     qemu_set_ram_fd(memory_region_get_ram_addr(&s->server_bar2), fd);
 493     s->ivshmem_bar2 = &s->server_bar2;
 494 }
 495
 496 static void process_msg_disconnect(IVShmemState *s, uint16_t posn,
 497                                    Error **errp)
 498 {
 499     IVSHMEM_DPRINTF("posn %d has gone away\n", posn);
 500     if (posn >= s->nb_peers || posn == s->vm_id) {
 501         error_setg(errp, "invalid peer %d", posn);
 502         return;
 503     }
 504     close_peer_eventfds(s, posn);
 505 }
 506
 507 static void process_msg_connect(IVShmemState *s, uint16_t posn, int fd,
 508                                 Error **errp)
 509 {
 510     Peer *peer = &s->peers[posn];
 511     int vector;
 512
 513     /*
 514      * The N-th connect message for this peer comes with the file
 515      * descriptor for vector N-1.  Count messages to find the vector.
 516      */
 517     if (peer->nb_eventfds >= s->vectors) {
 518         error_setg(errp, "Too many eventfd received, device has %d vectors",
 519                    s->vectors);
 520         close(fd);
 521         return;
 522     }
 523     vector = peer->nb_eventfds++;
 524
 525     IVSHMEM_DPRINTF("eventfds[%d][%d] = %d\n", posn, vector, fd);
 526     event_notifier_init_fd(&peer->eventfds[vector], fd);
 527     fcntl_setfl(fd, O_NONBLOCK); /* msix/irqfd poll non block */
 528
 529     if (posn == s->vm_id) {
 530         setup_interrupt(s, vector, errp);
 531         /* TODO do we need to handle the error? */
 532     }
 533
 534     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
 535         ivshmem_add_eventfd(s, posn, vector);
 536     }
 537 }
 538
 539 static void process_msg(IVShmemState *s, int64_t msg, int fd, Error **errp)
 540 {
 541     IVSHMEM_DPRINTF("posn is %" PRId64 ", fd is %d\n", msg, fd);
 542
 543     if (msg < -1 || msg > IVSHMEM_MAX_PEERS) {
 544         error_setg(errp, "server sent invalid message %" PRId64, msg);
 545         close(fd);
 546         return;
 547     }
 548
 549     if (msg == -1) {
 550         process_msg_shmem(s, fd, errp);
 551         return;
 552     }
 553
 554     if (msg >= s->nb_peers) {
 555         resize_peers(s, msg + 1);
 556     }
 557
 558     if (fd >= 0) {
 559         process_msg_connect(s, msg, fd, errp);
 560     } else {
 561         process_msg_disconnect(s, msg, errp);
 562     }
 563 }
 564
 565 static int ivshmem_can_receive(void *opaque)
 566 {
 567     IVShmemState *s = opaque;
 568
 569     assert(s->msg_buffered_bytes < sizeof(s->msg_buf));
 570     return sizeof(s->msg_buf) - s->msg_buffered_bytes;
 571 }
 572
 573 static void ivshmem_read(void *opaque, const uint8_t *buf, int size)
 574 {
 575     IVShmemState *s = opaque;
 576     Error *err = NULL;
 577     int fd;
 578     int64_t msg;
 579
 580     assert(size >= 0 && s->msg_buffered_bytes + size <= sizeof(s->msg_buf));
 581     memcpy((unsigned char *)&s->msg_buf + s->msg_buffered_bytes, buf, size);
 582     s->msg_buffered_bytes += size;
 583     if (s->msg_buffered_bytes < sizeof(s->msg_buf)) {
 584         return;
 585     }
 586     msg = le64_to_cpu(s->msg_buf);
 587     s->msg_buffered_bytes = 0;
 588
 589     fd = qemu_chr_fe_get_msgfd(s->server_chr);
 590     IVSHMEM_DPRINTF("posn is %" PRId64 ", fd is %d\n", msg, fd);
 591
 592     process_msg(s, msg, fd, &err);
 593     if (err) {
 594         error_report_err(err);
 595     }
 596 }
 597
 598 static int64_t ivshmem_recv_msg(IVShmemState *s, int *pfd, Error **errp)
 599 {
 600     int64_t msg;
 601     int n, ret;
 602
 603     n = 0;
 604     do {
 605         ret = qemu_chr_fe_read_all(s->server_chr, (uint8_t *)&msg + n,
 606                                  sizeof(msg) - n);
 607         if (ret < 0 && ret != -EINTR) {
 608             error_setg_errno(errp, -ret, "read from server failed");
 609             return INT64_MIN;
 610         }
 611         n += ret;
 612     } while (n < sizeof(msg));
 613
 614     *pfd = qemu_chr_fe_get_msgfd(s->server_chr);
 615     return msg;
 616 }
 617
 618 static void ivshmem_recv_setup(IVShmemState *s, Error **errp)
 619 {
 620     Error *err = NULL;
 621     int64_t msg;
 622     int fd;
 623
 624     msg = ivshmem_recv_msg(s, &fd, &err);
 625     if (err) {
 626         error_propagate(errp, err);
 627         return;
 628     }
 629     if (msg != IVSHMEM_PROTOCOL_VERSION) {
 630         error_setg(errp, "server sent version %" PRId64 ", expecting %d",
 631                    msg, IVSHMEM_PROTOCOL_VERSION);
 632         return;
 633     }
 634     if (fd != -1) {
 635         error_setg(errp, "server sent invalid version message");
 636         return;
 637     }
 638
 639     /*
 640      * ivshmem-server sends the remaining initial messages in a fixed
 641      * order, but the device has always accepted them in any order.
 642      * Stay as compatible as practical, just in case people use
 643      * servers that behave differently.
 644      */
 645
 646     /*
 647      * ivshmem_device_spec.txt has always required the ID message
 648      * right here, and ivshmem-server has always complied.  However,
 649      * older versions of the device accepted it out of order, but
 650      * broke when an interrupt setup message arrived before it.
 651      */
 652     msg = ivshmem_recv_msg(s, &fd, &err);
 653     if (err) {
 654         error_propagate(errp, err);
 655         return;
 656     }
 657     if (fd != -1 || msg < 0 || msg > IVSHMEM_MAX_PEERS) {
 658         error_setg(errp, "server sent invalid ID message");
 659         return;
 660     }
 661     s->vm_id = msg;
 662
 663     /*
 664      * Receive more messages until we got shared memory.
 665      */
 666     do {
 667         msg = ivshmem_recv_msg(s, &fd, &err);
 668         if (err) {
 669             error_propagate(errp, err);
 670             return;
 671         }
 672         process_msg(s, msg, fd, &err);
 673         if (err) {
 674             error_propagate(errp, err);
 675             return;
 676         }
 677     } while (msg != -1);
 678
 679     /*
 680      * This function must either map the shared memory or fail.  The
 681      * loop above ensures that: it terminates normally only after it
 682      * successfully processed the server's shared memory message.
 683      * Assert that actually mapped the shared memory:
 684      */
 685     assert(s->ivshmem_bar2);
 686 }
 687
 688 /* Select the MSI-X vectors used by device.
 689  * ivshmem maps events to vectors statically, so
 690  * we just enable all vectors on init and after reset. */
 691 static void ivshmem_msix_vector_use(IVShmemState *s)
 692 {
 693     PCIDevice *d = PCI_DEVICE(s);
 694     int i;
 695
 696     for (i = 0; i < s->vectors; i++) {
 697         msix_vector_use(d, i);
 698     }
 699 }
 700
 701 static void ivshmem_reset(DeviceState *d)
 702 {
 703     IVShmemState *s = IVSHMEM(d);
 704
 705     s->intrstatus = 0;
 706     s->intrmask = 0;
 707     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 708         ivshmem_msix_vector_use(s);
 709     }
 710 }
 711
 712 static int ivshmem_setup_interrupts(IVShmemState *s)
 713 {
 714     /* allocate QEMU callback data for receiving interrupts */
 715     s->msi_vectors = g_malloc0(s->vectors * sizeof(MSIVector));
 716
 717     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 718         if (msix_init_exclusive_bar(PCI_DEVICE(s), s->vectors, 1)) {
 719             return -1;
 720         }
 721
 722         IVSHMEM_DPRINTF("msix initialized (%d vectors)\n", s->vectors);
 723         ivshmem_msix_vector_use(s);
 724     }
 725
 726     return 0;
 727 }
 728
 729 static void ivshmem_enable_irqfd(IVShmemState *s)
 730 {
 731     PCIDevice *pdev = PCI_DEVICE(s);
 732     int i;
 733
 734     for (i = 0; i < s->peers[s->vm_id].nb_eventfds; i++) {
 735         Error *err = NULL;
 736
 737         ivshmem_add_kvm_msi_virq(s, i, &err);
 738         if (err) {
 739             error_report_err(err);
 740             /* TODO do we need to handle the error? */
 741         }
 742     }
 743
 744     if (msix_set_vector_notifiers(pdev,
 745                                   ivshmem_vector_unmask,
 746                                   ivshmem_vector_mask,
 747                                   ivshmem_vector_poll)) {
 748         error_report("ivshmem: msix_set_vector_notifiers failed");
 749     }
 750 }
 751
 752 static void ivshmem_remove_kvm_msi_virq(IVShmemState *s, int vector)
 753 {
 754     IVSHMEM_DPRINTF("ivshmem_remove_kvm_msi_virq vector:%d\n", vector);
 755
 756     if (s->msi_vectors[vector].pdev == NULL) {
 757         return;
 758     }
 759
 760     /* it was cleaned when masked in the frontend. */
 761     kvm_irqchip_release_virq(kvm_state, s->msi_vectors[vector].virq);
 762
 763     s->msi_vectors[vector].pdev = NULL;
 764 }
 765
 766 static void ivshmem_disable_irqfd(IVShmemState *s)
 767 {
 768     PCIDevice *pdev = PCI_DEVICE(s);
 769     int i;
 770
 771     for (i = 0; i < s->peers[s->vm_id].nb_eventfds; i++) {
 772         ivshmem_remove_kvm_msi_virq(s, i);
 773     }
 774
 775     msix_unset_vector_notifiers(pdev);
 776 }
 777
 778 static void ivshmem_write_config(PCIDevice *pdev, uint32_t address,
 779                                  uint32_t val, int len)
 780 {
 781     IVShmemState *s = IVSHMEM(pdev);
 782     int is_enabled, was_enabled = msix_enabled(pdev);
 783
 784     pci_default_write_config(pdev, address, val, len);
 785     is_enabled = msix_enabled(pdev);
 786
 787     if (kvm_msi_via_irqfd_enabled()) {
 788         if (!was_enabled && is_enabled) {
 789             ivshmem_enable_irqfd(s);
 790         } else if (was_enabled && !is_enabled) {
 791             ivshmem_disable_irqfd(s);
 792         }
 793     }
 794 }
 795
 796 static void desugar_shm(IVShmemState *s)
 797 {
 798     Object *obj;
 799     char *path;
 800
 801     obj = object_new("memory-backend-file");
 802     path = g_strdup_printf("/dev/shm/%s", s->shmobj);
 803     object_property_set_str(obj, path, "mem-path", &error_abort);
 804     g_free(path);
 805     object_property_set_int(obj, s->ivshmem_size, "size", &error_abort);
 806     object_property_set_bool(obj, true, "share", &error_abort);
 807     object_property_add_child(OBJECT(s), "internal-shm-backend", obj,
 808                               &error_abort);
 809     user_creatable_complete(obj, &error_abort);
 810     s->hostmem = MEMORY_BACKEND(obj);
 811 }
 812
 813 static void pci_ivshmem_realize(PCIDevice *dev, Error **errp)
 814 {
 815     IVShmemState *s = IVSHMEM(dev);
 816     Error *err = NULL;
 817     uint8_t *pci_conf;
 818     uint8_t attr = PCI_BASE_ADDRESS_SPACE_MEMORY |
 819         PCI_BASE_ADDRESS_MEM_PREFETCH;
 820
 821     if (!!s->server_chr + !!s->shmobj + !!s->hostmem != 1) {
 822         error_setg(errp,
 823                    "You must specify either 'shm', 'chardev' or 'x-memdev'");
 824         return;
 825     }
 826
 827     if (s->hostmem) {
 828         MemoryRegion *mr;
 829
 830         if (s->sizearg) {
 831             g_warning("size argument ignored with hostmem");
 832         }
 833
 834         mr = host_memory_backend_get_memory(s->hostmem, &error_abort);
 835         s->ivshmem_size = memory_region_size(mr);
 836     } else if (s->sizearg == NULL) {
 837         s->ivshmem_size = 4 << 20; /* 4 MB default */
 838     } else {
 839         char *end;
 840         int64_t size = qemu_strtosz(s->sizearg, &end);
 841         if (size < 0 || (size_t)size != size || *end != '\0'
 842             || !is_power_of_2(size)) {
 843             error_setg(errp, "Invalid size %s", s->sizearg);
 844             return;
 845         }
 846         s->ivshmem_size = size;
 847     }
 848
 849     /* IRQFD requires MSI */
 850     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD) &&
 851         !ivshmem_has_feature(s, IVSHMEM_MSI)) {
 852         error_setg(errp, "ioeventfd/irqfd requires MSI");
 853         return;
 854     }
 855
 856     /* check that role is reasonable */
 857     if (s->role) {
 858         if (strncmp(s->role, "peer", 5) == 0) {
 859             s->role_val = IVSHMEM_PEER;
 860         } else if (strncmp(s->role, "master", 7) == 0) {
 861             s->role_val = IVSHMEM_MASTER;
 862         } else {
 863             error_setg(errp, "'role' must be 'peer' or 'master'");
 864             return;
 865         }
 866     } else {
 867         s->role_val = IVSHMEM_MASTER; /* default */
 868     }
 869
 870     pci_conf = dev->config;
 871     pci_conf[PCI_COMMAND] = PCI_COMMAND_IO | PCI_COMMAND_MEMORY;
 872
 873     /*
 874      * Note: we don't use INTx with IVSHMEM_MSI at all, so this is a
 875      * bald-faced lie then.  But it's a backwards compatible lie.
 876      */
 877     pci_config_set_interrupt_pin(pci_conf, 1);
 878
 879     memory_region_init_io(&s->ivshmem_mmio, OBJECT(s), &ivshmem_mmio_ops, s,
 880                           "ivshmem-mmio", IVSHMEM_REG_BAR_SIZE);
 881
 882     /* region for registers*/
 883     pci_register_bar(dev, 0, PCI_BASE_ADDRESS_SPACE_MEMORY,
 884                      &s->ivshmem_mmio);
 885
 886     if (s->ivshmem_64bit) {
 887         attr |= PCI_BASE_ADDRESS_MEM_TYPE_64;
 888     }
 889
 890     if (s->shmobj) {
 891         desugar_shm(s);
 892     }
 893
 894     if (s->hostmem != NULL) {
 895         IVSHMEM_DPRINTF("using hostmem\n");
 896
 897         s->ivshmem_bar2 = host_memory_backend_get_memory(s->hostmem,
 898                                                          &error_abort);
 899     } else {
 900         IVSHMEM_DPRINTF("using shared memory server (socket = %s)\n",
 901                         s->server_chr->filename);
 902
 903         /* we allocate enough space for 16 peers and grow as needed */
 904         resize_peers(s, 16);
 905
 906         /*
 907          * Receive setup messages from server synchronously.
 908          * Older versions did it asynchronously, but that creates a
 909          * number of entertaining race conditions.
 910          */
 911         ivshmem_recv_setup(s, &err);
 912         if (err) {
 913             error_propagate(errp, err);
 914             return;
 915         }
 916
 917         qemu_chr_add_handlers(s->server_chr, ivshmem_can_receive,
 918                               ivshmem_read, NULL, s);
 919
 920         if (ivshmem_setup_interrupts(s) < 0) {
 921             error_setg(errp, "failed to initialize interrupts");
 922             return;
 923         }
 924     }
 925
 926     vmstate_register_ram(s->ivshmem_bar2, DEVICE(s));
 927     pci_register_bar(PCI_DEVICE(s), 2, attr, s->ivshmem_bar2);
 928
 929     if (s->role_val == IVSHMEM_PEER) {
 930         error_setg(&s->migration_blocker,
 931                    "Migration is disabled when using feature 'peer mode' in device 'ivshmem'");
 932         migrate_add_blocker(s->migration_blocker);
 933     }
 934 }
 935
 936 static void pci_ivshmem_exit(PCIDevice *dev)
 937 {
 938     IVShmemState *s = IVSHMEM(dev);
 939     int i;
 940
 941     if (s->migration_blocker) {
 942         migrate_del_blocker(s->migration_blocker);
 943         error_free(s->migration_blocker);
 944     }
 945
 946     if (memory_region_is_mapped(s->ivshmem_bar2)) {
 947         if (!s->hostmem) {
 948             void *addr = memory_region_get_ram_ptr(s->ivshmem_bar2);
 949             int fd;
 950
 951             if (munmap(addr, s->ivshmem_size) == -1) {
 952                 error_report("Failed to munmap shared memory %s",
 953                              strerror(errno));
 954             }
 955
 956             fd = qemu_get_ram_fd(memory_region_get_ram_addr(s->ivshmem_bar2));
 957             close(fd);
 958         }
 959
 960         vmstate_unregister_ram(s->ivshmem_bar2, DEVICE(dev));
 961     }
 962
 963     if (s->peers) {
 964         for (i = 0; i < s->nb_peers; i++) {
 965             close_peer_eventfds(s, i);
 966         }
 967         g_free(s->peers);
 968     }
 969
 970     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 971         msix_uninit_exclusive_bar(dev);
 972     }
 973
 974     g_free(s->msi_vectors);
 975 }
 976
 977 static bool test_msix(void *opaque, int version_id)
 978 {
 979     IVShmemState *s = opaque;
 980
 981     return ivshmem_has_feature(s, IVSHMEM_MSI);
 982 }
 983
 984 static bool test_no_msix(void *opaque, int version_id)
 985 {
 986     return !test_msix(opaque, version_id);
 987 }
 988
 989 static int ivshmem_pre_load(void *opaque)
 990 {
 991     IVShmemState *s = opaque;
 992
 993     if (s->role_val == IVSHMEM_PEER) {
 994         error_report("'peer' devices are not migratable");
 995         return -EINVAL;
 996     }
 997
 998     return 0;
 999 }
1000
1001 static int ivshmem_post_load(void *opaque, int version_id)
1002 {
1003     IVShmemState *s = opaque;
1004
1005     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
1006         ivshmem_msix_vector_use(s);
1007     }
1008     return 0;
1009 }
1010
1011 static int ivshmem_load_old(QEMUFile *f, void *opaque, int version_id)
1012 {
1013     IVShmemState *s = opaque;
1014     PCIDevice *pdev = PCI_DEVICE(s);
1015     int ret;
1016
1017     IVSHMEM_DPRINTF("ivshmem_load_old\n");
1018
1019     if (version_id != 0) {
1020         return -EINVAL;
1021     }
1022
1023     if (s->role_val == IVSHMEM_PEER) {
1024         error_report("'peer' devices are not migratable");
1025         return -EINVAL;
1026     }
1027
1028     ret = pci_device_load(pdev, f);
1029     if (ret) {
1030         return ret;
1031     }
1032
1033     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
1034         msix_load(pdev, f);
1035         ivshmem_msix_vector_use(s);
1036     } else {
1037         s->intrstatus = qemu_get_be32(f);
1038         s->intrmask = qemu_get_be32(f);
1039     }
1040
1041     return 0;
1042 }
1043
1044 static const VMStateDescription ivshmem_vmsd = {
1045     .name = "ivshmem",
1046     .version_id = 1,
1047     .minimum_version_id = 1,
1048     .pre_load = ivshmem_pre_load,
1049     .post_load = ivshmem_post_load,
1050     .fields = (VMStateField[]) {
1051         VMSTATE_PCI_DEVICE(parent_obj, IVShmemState),
1052
1053         VMSTATE_MSIX_TEST(parent_obj, IVShmemState, test_msix),
1054         VMSTATE_UINT32_TEST(intrstatus, IVShmemState, test_no_msix),
1055         VMSTATE_UINT32_TEST(intrmask, IVShmemState, test_no_msix),
1056
1057         VMSTATE_END_OF_LIST()
1058     },
1059     .load_state_old = ivshmem_load_old,
1060     .minimum_version_id_old = 0
1061 };
1062
1063 static Property ivshmem_properties[] = {
1064     DEFINE_PROP_CHR("chardev", IVShmemState, server_chr),
1065     DEFINE_PROP_STRING("size", IVShmemState, sizearg),
1066     DEFINE_PROP_UINT32("vectors", IVShmemState, vectors, 1),
1067     DEFINE_PROP_BIT("ioeventfd", IVShmemState, features, IVSHMEM_IOEVENTFD, false),
1068     DEFINE_PROP_BIT("msi", IVShmemState, features, IVSHMEM_MSI, true),
1069     DEFINE_PROP_STRING("shm", IVShmemState, shmobj),
1070     DEFINE_PROP_STRING("role", IVShmemState, role),
1071     DEFINE_PROP_UINT32("use64", IVShmemState, ivshmem_64bit, 1),
1072     DEFINE_PROP_END_OF_LIST(),
1073 };
1074
1075 static void ivshmem_class_init(ObjectClass *klass, void *data)
1076 {
1077     DeviceClass *dc = DEVICE_CLASS(klass);
1078     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
1079
1080     k->realize = pci_ivshmem_realize;
1081     k->exit = pci_ivshmem_exit;
1082     k->config_write = ivshmem_write_config;
1083     k->vendor_id = PCI_VENDOR_ID_IVSHMEM;
1084     k->device_id = PCI_DEVICE_ID_IVSHMEM;
1085     k->class_id = PCI_CLASS_MEMORY_RAM;
1086     dc->reset = ivshmem_reset;
1087     dc->props = ivshmem_properties;
1088     dc->vmsd = &ivshmem_vmsd;
1089     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
1090     dc->desc = "Inter-VM shared memory";
1091 }
1092
1093 static void ivshmem_check_memdev_is_busy(Object *obj, const char *name,
1094                                          Object *val, Error **errp)
1095 {
1096     MemoryRegion *mr;
1097
1098     mr = host_memory_backend_get_memory(MEMORY_BACKEND(val), &error_abort);
1099     if (memory_region_is_mapped(mr)) {
1100         char *path = object_get_canonical_path_component(val);
1101         error_setg(errp, "can't use already busy memdev: %s", path);
1102         g_free(path);
1103     } else {
1104         qdev_prop_allow_set_link_before_realize(obj, name, val, errp);
1105     }
1106 }
1107
1108 static void ivshmem_init(Object *obj)
1109 {
1110     IVShmemState *s = IVSHMEM(obj);
1111
1112     object_property_add_link(obj, "x-memdev", TYPE_MEMORY_BACKEND,
1113                              (Object **)&s->hostmem,
1114                              ivshmem_check_memdev_is_busy,
1115                              OBJ_PROP_LINK_UNREF_ON_RELEASE,
1116                              &error_abort);
1117 }
1118
1119 static const TypeInfo ivshmem_info = {
1120     .name          = TYPE_IVSHMEM,
1121     .parent        = TYPE_PCI_DEVICE,
1122     .instance_size = sizeof(IVShmemState),
1123     .instance_init = ivshmem_init,
1124     .class_init    = ivshmem_class_init,
1125 };
1126
1127 static void ivshmem_register_types(void)
1128 {
1129     type_register_static(&ivshmem_info);
1130 }
1131
1132 type_init(ivshmem_register_types)