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 "qemu/fifo8.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     Fifo8 incoming_fifo;
  84     MemoryRegion ivshmem_mmio;
  85
  86     /* We might need to register the BAR before we actually have the memory.
  87      * So prepare a container MemoryRegion for the BAR immediately and
  88      * add a subregion when we have the memory.
  89      */
  90     MemoryRegion bar;
  91     MemoryRegion ivshmem;
  92     uint64_t ivshmem_size; /* size of shared memory region */
  93     uint32_t ivshmem_64bit;
  94
  95     Peer *peers;
  96     int nb_peers;               /* space in @peers[] */
  97
  98     int vm_id;
  99     uint32_t vectors;
 100     uint32_t features;
 101     MSIVector *msi_vectors;
 102
 103     Error *migration_blocker;
 104
 105     char * shmobj;
 106     char * sizearg;
 107     char * role;
 108     int role_val;   /* scalar to avoid multiple string comparisons */
 109 } IVShmemState;
 110
 111 /* registers for the Inter-VM shared memory device */
 112 enum ivshmem_registers {
 113     INTRMASK = 0,
 114     INTRSTATUS = 4,
 115     IVPOSITION = 8,
 116     DOORBELL = 12,
 117 };
 118
 119 static inline uint32_t ivshmem_has_feature(IVShmemState *ivs,
 120                                                     unsigned int feature) {
 121     return (ivs->features & (1 << feature));
 122 }
 123
 124 static void ivshmem_update_irq(IVShmemState *s)
 125 {
 126     PCIDevice *d = PCI_DEVICE(s);
 127     uint32_t isr = s->intrstatus & s->intrmask;
 128
 129     /* No INTx with msi=on, whether the guest enabled MSI-X or not */
 130     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 131         return;
 132     }
 133
 134     /* don't print ISR resets */
 135     if (isr) {
 136         IVSHMEM_DPRINTF("Set IRQ to %d (%04x %04x)\n",
 137                         isr ? 1 : 0, s->intrstatus, s->intrmask);
 138     }
 139
 140     pci_set_irq(d, isr != 0);
 141 }
 142
 143 static void ivshmem_IntrMask_write(IVShmemState *s, uint32_t val)
 144 {
 145     IVSHMEM_DPRINTF("IntrMask write(w) val = 0x%04x\n", val);
 146
 147     s->intrmask = val;
 148     ivshmem_update_irq(s);
 149 }
 150
 151 static uint32_t ivshmem_IntrMask_read(IVShmemState *s)
 152 {
 153     uint32_t ret = s->intrmask;
 154
 155     IVSHMEM_DPRINTF("intrmask read(w) val = 0x%04x\n", ret);
 156     return ret;
 157 }
 158
 159 static void ivshmem_IntrStatus_write(IVShmemState *s, uint32_t val)
 160 {
 161     IVSHMEM_DPRINTF("IntrStatus write(w) val = 0x%04x\n", val);
 162
 163     s->intrstatus = val;
 164     ivshmem_update_irq(s);
 165 }
 166
 167 static uint32_t ivshmem_IntrStatus_read(IVShmemState *s)
 168 {
 169     uint32_t ret = s->intrstatus;
 170
 171     /* reading ISR clears all interrupts */
 172     s->intrstatus = 0;
 173     ivshmem_update_irq(s);
 174     return ret;
 175 }
 176
 177 static void ivshmem_io_write(void *opaque, hwaddr addr,
 178                              uint64_t val, unsigned size)
 179 {
 180     IVShmemState *s = opaque;
 181
 182     uint16_t dest = val >> 16;
 183     uint16_t vector = val & 0xff;
 184
 185     addr &= 0xfc;
 186
 187     IVSHMEM_DPRINTF("writing to addr " TARGET_FMT_plx "\n", addr);
 188     switch (addr)
 189     {
 190         case INTRMASK:
 191             ivshmem_IntrMask_write(s, val);
 192             break;
 193
 194         case INTRSTATUS:
 195             ivshmem_IntrStatus_write(s, val);
 196             break;
 197
 198         case DOORBELL:
 199             /* check that dest VM ID is reasonable */
 200             if (dest >= s->nb_peers) {
 201                 IVSHMEM_DPRINTF("Invalid destination VM ID (%d)\n", dest);
 202                 break;
 203             }
 204
 205             /* check doorbell range */
 206             if (vector < s->peers[dest].nb_eventfds) {
 207                 IVSHMEM_DPRINTF("Notifying VM %d on vector %d\n", dest, vector);
 208                 event_notifier_set(&s->peers[dest].eventfds[vector]);
 209             } else {
 210                 IVSHMEM_DPRINTF("Invalid destination vector %d on VM %d\n",
 211                                 vector, dest);
 212             }
 213             break;
 214         default:
 215             IVSHMEM_DPRINTF("Unhandled write " TARGET_FMT_plx "\n", addr);
 216     }
 217 }
 218
 219 static uint64_t ivshmem_io_read(void *opaque, hwaddr addr,
 220                                 unsigned size)
 221 {
 222
 223     IVShmemState *s = opaque;
 224     uint32_t ret;
 225
 226     switch (addr)
 227     {
 228         case INTRMASK:
 229             ret = ivshmem_IntrMask_read(s);
 230             break;
 231
 232         case INTRSTATUS:
 233             ret = ivshmem_IntrStatus_read(s);
 234             break;
 235
 236         case IVPOSITION:
 237             ret = s->vm_id;
 238             break;
 239
 240         default:
 241             IVSHMEM_DPRINTF("why are we reading " TARGET_FMT_plx "\n", addr);
 242             ret = 0;
 243     }
 244
 245     return ret;
 246 }
 247
 248 static const MemoryRegionOps ivshmem_mmio_ops = {
 249     .read = ivshmem_io_read,
 250     .write = ivshmem_io_write,
 251     .endianness = DEVICE_NATIVE_ENDIAN,
 252     .impl = {
 253         .min_access_size = 4,
 254         .max_access_size = 4,
 255     },
 256 };
 257
 258 static int ivshmem_can_receive(void * opaque)
 259 {
 260     return sizeof(int64_t);
 261 }
 262
 263 static void ivshmem_vector_notify(void *opaque)
 264 {
 265     MSIVector *entry = opaque;
 266     PCIDevice *pdev = entry->pdev;
 267     IVShmemState *s = IVSHMEM(pdev);
 268     int vector = entry - s->msi_vectors;
 269     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 270
 271     if (!event_notifier_test_and_clear(n)) {
 272         return;
 273     }
 274
 275     IVSHMEM_DPRINTF("interrupt on vector %p %d\n", pdev, vector);
 276     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 277         if (msix_enabled(pdev)) {
 278             msix_notify(pdev, vector);
 279         }
 280     } else {
 281         ivshmem_IntrStatus_write(s, 1);
 282     }
 283 }
 284
 285 static int ivshmem_vector_unmask(PCIDevice *dev, unsigned vector,
 286                                  MSIMessage msg)
 287 {
 288     IVShmemState *s = IVSHMEM(dev);
 289     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 290     MSIVector *v = &s->msi_vectors[vector];
 291     int ret;
 292
 293     IVSHMEM_DPRINTF("vector unmask %p %d\n", dev, vector);
 294
 295     ret = kvm_irqchip_update_msi_route(kvm_state, v->virq, msg, dev);
 296     if (ret < 0) {
 297         return ret;
 298     }
 299
 300     return kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, n, NULL, v->virq);
 301 }
 302
 303 static void ivshmem_vector_mask(PCIDevice *dev, unsigned vector)
 304 {
 305     IVShmemState *s = IVSHMEM(dev);
 306     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 307     int ret;
 308
 309     IVSHMEM_DPRINTF("vector mask %p %d\n", dev, vector);
 310
 311     ret = kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state, n,
 312                                                 s->msi_vectors[vector].virq);
 313     if (ret != 0) {
 314         error_report("remove_irqfd_notifier_gsi failed");
 315     }
 316 }
 317
 318 static void ivshmem_vector_poll(PCIDevice *dev,
 319                                 unsigned int vector_start,
 320                                 unsigned int vector_end)
 321 {
 322     IVShmemState *s = IVSHMEM(dev);
 323     unsigned int vector;
 324
 325     IVSHMEM_DPRINTF("vector poll %p %d-%d\n", dev, vector_start, vector_end);
 326
 327     vector_end = MIN(vector_end, s->vectors);
 328
 329     for (vector = vector_start; vector < vector_end; vector++) {
 330         EventNotifier *notifier = &s->peers[s->vm_id].eventfds[vector];
 331
 332         if (!msix_is_masked(dev, vector)) {
 333             continue;
 334         }
 335
 336         if (event_notifier_test_and_clear(notifier)) {
 337             msix_set_pending(dev, vector);
 338         }
 339     }
 340 }
 341
 342 static void watch_vector_notifier(IVShmemState *s, EventNotifier *n,
 343                                  int vector)
 344 {
 345     int eventfd = event_notifier_get_fd(n);
 346
 347     assert(!s->msi_vectors[vector].pdev);
 348     s->msi_vectors[vector].pdev = PCI_DEVICE(s);
 349
 350     qemu_set_fd_handler(eventfd, ivshmem_vector_notify,
 351                         NULL, &s->msi_vectors[vector]);
 352 }
 353
 354 static int check_shm_size(IVShmemState *s, int fd, Error **errp)
 355 {
 356     /* check that the guest isn't going to try and map more memory than the
 357      * the object has allocated return -1 to indicate error */
 358
 359     struct stat buf;
 360
 361     if (fstat(fd, &buf) < 0) {
 362         error_setg(errp, "exiting: fstat on fd %d failed: %s",
 363                    fd, strerror(errno));
 364         return -1;
 365     }
 366
 367     if (s->ivshmem_size > buf.st_size) {
 368         error_setg(errp, "Requested memory size greater"
 369                    " than shared object size (%" PRIu64 " > %" PRIu64")",
 370                    s->ivshmem_size, (uint64_t)buf.st_size);
 371         return -1;
 372     } else {
 373         return 0;
 374     }
 375 }
 376
 377 /* create the shared memory BAR when we are not using the server, so we can
 378  * create the BAR and map the memory immediately */
 379 static int create_shared_memory_BAR(IVShmemState *s, int fd, uint8_t attr,
 380                                     Error **errp)
 381 {
 382     void * ptr;
 383
 384     ptr = mmap(0, s->ivshmem_size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
 385     if (ptr == MAP_FAILED) {
 386         error_setg_errno(errp, errno, "Failed to mmap shared memory");
 387         return -1;
 388     }
 389
 390     memory_region_init_ram_ptr(&s->ivshmem, OBJECT(s), "ivshmem.bar2",
 391                                s->ivshmem_size, ptr);
 392     qemu_set_ram_fd(memory_region_get_ram_addr(&s->ivshmem), fd);
 393     vmstate_register_ram(&s->ivshmem, DEVICE(s));
 394     memory_region_add_subregion(&s->bar, 0, &s->ivshmem);
 395
 396     /* region for shared memory */
 397     pci_register_bar(PCI_DEVICE(s), 2, attr, &s->bar);
 398
 399     return 0;
 400 }
 401
 402 static void ivshmem_add_eventfd(IVShmemState *s, int posn, int i)
 403 {
 404     memory_region_add_eventfd(&s->ivshmem_mmio,
 405                               DOORBELL,
 406                               4,
 407                               true,
 408                               (posn << 16) | i,
 409                               &s->peers[posn].eventfds[i]);
 410 }
 411
 412 static void ivshmem_del_eventfd(IVShmemState *s, int posn, int i)
 413 {
 414     memory_region_del_eventfd(&s->ivshmem_mmio,
 415                               DOORBELL,
 416                               4,
 417                               true,
 418                               (posn << 16) | i,
 419                               &s->peers[posn].eventfds[i]);
 420 }
 421
 422 static void close_peer_eventfds(IVShmemState *s, int posn)
 423 {
 424     int i, n;
 425
 426     assert(posn >= 0 && posn < s->nb_peers);
 427     n = s->peers[posn].nb_eventfds;
 428
 429     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
 430         memory_region_transaction_begin();
 431         for (i = 0; i < n; i++) {
 432             ivshmem_del_eventfd(s, posn, i);
 433         }
 434         memory_region_transaction_commit();
 435     }
 436
 437     for (i = 0; i < n; i++) {
 438         event_notifier_cleanup(&s->peers[posn].eventfds[i]);
 439     }
 440
 441     g_free(s->peers[posn].eventfds);
 442     s->peers[posn].nb_eventfds = 0;
 443 }
 444
 445 static void resize_peers(IVShmemState *s, int nb_peers)
 446 {
 447     int old_nb_peers = s->nb_peers;
 448     int i;
 449
 450     assert(nb_peers > old_nb_peers);
 451     IVSHMEM_DPRINTF("bumping storage to %d peers\n", nb_peers);
 452
 453     s->peers = g_realloc(s->peers, nb_peers * sizeof(Peer));
 454     s->nb_peers = nb_peers;
 455
 456     for (i = old_nb_peers; i < nb_peers; i++) {
 457         s->peers[i].eventfds = g_new0(EventNotifier, s->vectors);
 458         s->peers[i].nb_eventfds = 0;
 459     }
 460 }
 461
 462 static bool fifo_update_and_get(IVShmemState *s, const uint8_t *buf, int size,
 463                                 void *data, size_t len)
 464 {
 465     const uint8_t *p;
 466     uint32_t num;
 467
 468     assert(len <= sizeof(int64_t)); /* limitation of the fifo */
 469     if (fifo8_is_empty(&s->incoming_fifo) && size == len) {
 470         memcpy(data, buf, size);
 471         return true;
 472     }
 473
 474     IVSHMEM_DPRINTF("short read of %d bytes\n", size);
 475
 476     num = MIN(size, sizeof(int64_t) - fifo8_num_used(&s->incoming_fifo));
 477     fifo8_push_all(&s->incoming_fifo, buf, num);
 478
 479     if (fifo8_num_used(&s->incoming_fifo) < len) {
 480         assert(num == 0);
 481         return false;
 482     }
 483
 484     size -= num;
 485     buf += num;
 486     p = fifo8_pop_buf(&s->incoming_fifo, len, &num);
 487     assert(num == len);
 488
 489     memcpy(data, p, len);
 490
 491     if (size > 0) {
 492         fifo8_push_all(&s->incoming_fifo, buf, size);
 493     }
 494
 495     return true;
 496 }
 497
 498 static bool fifo_update_and_get_i64(IVShmemState *s,
 499                                     const uint8_t *buf, int size, int64_t *i64)
 500 {
 501     if (fifo_update_and_get(s, buf, size, i64, sizeof(*i64))) {
 502         *i64 = GINT64_FROM_LE(*i64);
 503         return true;
 504     }
 505
 506     return false;
 507 }
 508
 509 static void ivshmem_add_kvm_msi_virq(IVShmemState *s, int vector,
 510                                      Error **errp)
 511 {
 512     PCIDevice *pdev = PCI_DEVICE(s);
 513     MSIMessage msg = msix_get_message(pdev, vector);
 514     int ret;
 515
 516     IVSHMEM_DPRINTF("ivshmem_add_kvm_msi_virq vector:%d\n", vector);
 517     assert(!s->msi_vectors[vector].pdev);
 518
 519     ret = kvm_irqchip_add_msi_route(kvm_state, msg, pdev);
 520     if (ret < 0) {
 521         error_setg(errp, "kvm_irqchip_add_msi_route failed");
 522         return;
 523     }
 524
 525     s->msi_vectors[vector].virq = ret;
 526     s->msi_vectors[vector].pdev = pdev;
 527 }
 528
 529 static void setup_interrupt(IVShmemState *s, int vector, Error **errp)
 530 {
 531     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 532     bool with_irqfd = kvm_msi_via_irqfd_enabled() &&
 533         ivshmem_has_feature(s, IVSHMEM_MSI);
 534     PCIDevice *pdev = PCI_DEVICE(s);
 535     Error *err = NULL;
 536
 537     IVSHMEM_DPRINTF("setting up interrupt for vector: %d\n", vector);
 538
 539     if (!with_irqfd) {
 540         IVSHMEM_DPRINTF("with eventfd\n");
 541         watch_vector_notifier(s, n, vector);
 542     } else if (msix_enabled(pdev)) {
 543         IVSHMEM_DPRINTF("with irqfd\n");
 544         ivshmem_add_kvm_msi_virq(s, vector, &err);
 545         if (err) {
 546             error_propagate(errp, err);
 547             return;
 548         }
 549
 550         if (!msix_is_masked(pdev, vector)) {
 551             kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, n, NULL,
 552                                                s->msi_vectors[vector].virq);
 553             /* TODO handle error */
 554         }
 555     } else {
 556         /* it will be delayed until msix is enabled, in write_config */
 557         IVSHMEM_DPRINTF("with irqfd, delayed until msix enabled\n");
 558     }
 559 }
 560
 561 static void process_msg_shmem(IVShmemState *s, int fd, Error **errp)
 562 {
 563     Error *err = NULL;
 564     void *ptr;
 565
 566     if (memory_region_is_mapped(&s->ivshmem)) {
 567         error_setg(errp, "server sent unexpected shared memory message");
 568         close(fd);
 569         return;
 570     }
 571
 572     if (check_shm_size(s, fd, &err) == -1) {
 573         error_propagate(errp, err);
 574         close(fd);
 575         return;
 576     }
 577
 578     /* mmap the region and map into the BAR2 */
 579     ptr = mmap(0, s->ivshmem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 580     if (ptr == MAP_FAILED) {
 581         error_setg_errno(errp, errno, "Failed to mmap shared memory");
 582         close(fd);
 583         return;
 584     }
 585     memory_region_init_ram_ptr(&s->ivshmem, OBJECT(s),
 586                                "ivshmem.bar2", s->ivshmem_size, ptr);
 587     qemu_set_ram_fd(memory_region_get_ram_addr(&s->ivshmem), fd);
 588     vmstate_register_ram(&s->ivshmem, DEVICE(s));
 589     memory_region_add_subregion(&s->bar, 0, &s->ivshmem);
 590 }
 591
 592 static void process_msg_disconnect(IVShmemState *s, uint16_t posn,
 593                                    Error **errp)
 594 {
 595     IVSHMEM_DPRINTF("posn %d has gone away\n", posn);
 596     if (posn >= s->nb_peers || posn == s->vm_id) {
 597         error_setg(errp, "invalid peer %d", posn);
 598         return;
 599     }
 600     close_peer_eventfds(s, posn);
 601 }
 602
 603 static void process_msg_connect(IVShmemState *s, uint16_t posn, int fd,
 604                                 Error **errp)
 605 {
 606     Peer *peer = &s->peers[posn];
 607     int vector;
 608
 609     /*
 610      * The N-th connect message for this peer comes with the file
 611      * descriptor for vector N-1.  Count messages to find the vector.
 612      */
 613     if (peer->nb_eventfds >= s->vectors) {
 614         error_setg(errp, "Too many eventfd received, device has %d vectors",
 615                    s->vectors);
 616         close(fd);
 617         return;
 618     }
 619     vector = peer->nb_eventfds++;
 620
 621     IVSHMEM_DPRINTF("eventfds[%d][%d] = %d\n", posn, vector, fd);
 622     event_notifier_init_fd(&peer->eventfds[vector], fd);
 623     fcntl_setfl(fd, O_NONBLOCK); /* msix/irqfd poll non block */
 624
 625     if (posn == s->vm_id) {
 626         setup_interrupt(s, vector, errp);
 627         /* TODO do we need to handle the error? */
 628     }
 629
 630     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
 631         ivshmem_add_eventfd(s, posn, vector);
 632     }
 633 }
 634
 635 static void process_msg(IVShmemState *s, int64_t msg, int fd, Error **errp)
 636 {
 637     IVSHMEM_DPRINTF("posn is %" PRId64 ", fd is %d\n", msg, fd);
 638
 639     if (msg < -1 || msg > IVSHMEM_MAX_PEERS) {
 640         error_setg(errp, "server sent invalid message %" PRId64, msg);
 641         close(fd);
 642         return;
 643     }
 644
 645     if (msg == -1) {
 646         process_msg_shmem(s, fd, errp);
 647         return;
 648     }
 649
 650     if (msg >= s->nb_peers) {
 651         resize_peers(s, msg + 1);
 652     }
 653
 654     if (fd >= 0) {
 655         process_msg_connect(s, msg, fd, errp);
 656     } else {
 657         process_msg_disconnect(s, msg, errp);
 658     }
 659 }
 660
 661 static void ivshmem_read(void *opaque, const uint8_t *buf, int size)
 662 {
 663     IVShmemState *s = opaque;
 664     Error *err = NULL;
 665     int fd;
 666     int64_t msg;
 667
 668     if (!fifo_update_and_get_i64(s, buf, size, &msg)) {
 669         return;
 670     }
 671
 672     fd = qemu_chr_fe_get_msgfd(s->server_chr);
 673     IVSHMEM_DPRINTF("posn is %" PRId64 ", fd is %d\n", msg, fd);
 674
 675     process_msg(s, msg, fd, &err);
 676     if (err) {
 677         error_report_err(err);
 678     }
 679 }
 680
 681 static int64_t ivshmem_recv_msg(IVShmemState *s, int *pfd, Error **errp)
 682 {
 683     int64_t msg;
 684     int n, ret;
 685
 686     n = 0;
 687     do {
 688         ret = qemu_chr_fe_read_all(s->server_chr, (uint8_t *)&msg + n,
 689                                  sizeof(msg) - n);
 690         if (ret < 0 && ret != -EINTR) {
 691             error_setg_errno(errp, -ret, "read from server failed");
 692             return INT64_MIN;
 693         }
 694         n += ret;
 695     } while (n < sizeof(msg));
 696
 697     *pfd = qemu_chr_fe_get_msgfd(s->server_chr);
 698     return msg;
 699 }
 700
 701 static void ivshmem_recv_setup(IVShmemState *s, Error **errp)
 702 {
 703     Error *err = NULL;
 704     int64_t msg;
 705     int fd;
 706
 707     msg = ivshmem_recv_msg(s, &fd, &err);
 708     if (err) {
 709         error_propagate(errp, err);
 710         return;
 711     }
 712     if (msg != IVSHMEM_PROTOCOL_VERSION) {
 713         error_setg(errp, "server sent version %" PRId64 ", expecting %d",
 714                    msg, IVSHMEM_PROTOCOL_VERSION);
 715         return;
 716     }
 717     if (fd != -1) {
 718         error_setg(errp, "server sent invalid version message");
 719         return;
 720     }
 721
 722     /*
 723      * ivshmem-server sends the remaining initial messages in a fixed
 724      * order, but the device has always accepted them in any order.
 725      * Stay as compatible as practical, just in case people use
 726      * servers that behave differently.
 727      */
 728
 729     /*
 730      * ivshmem_device_spec.txt has always required the ID message
 731      * right here, and ivshmem-server has always complied.  However,
 732      * older versions of the device accepted it out of order, but
 733      * broke when an interrupt setup message arrived before it.
 734      */
 735     msg = ivshmem_recv_msg(s, &fd, &err);
 736     if (err) {
 737         error_propagate(errp, err);
 738         return;
 739     }
 740     if (fd != -1 || msg < 0 || msg > IVSHMEM_MAX_PEERS) {
 741         error_setg(errp, "server sent invalid ID message");
 742         return;
 743     }
 744     s->vm_id = msg;
 745
 746     /*
 747      * Receive more messages until we got shared memory.
 748      */
 749     do {
 750         msg = ivshmem_recv_msg(s, &fd, &err);
 751         if (err) {
 752             error_propagate(errp, err);
 753             return;
 754         }
 755         process_msg(s, msg, fd, &err);
 756         if (err) {
 757             error_propagate(errp, err);
 758             return;
 759         }
 760     } while (msg != -1);
 761
 762     /*
 763      * This function must either map the shared memory or fail.  The
 764      * loop above ensures that: it terminates normally only after it
 765      * successfully processed the server's shared memory message.
 766      * Assert that actually mapped the shared memory:
 767      */
 768     assert(memory_region_is_mapped(&s->ivshmem));
 769 }
 770
 771 /* Select the MSI-X vectors used by device.
 772  * ivshmem maps events to vectors statically, so
 773  * we just enable all vectors on init and after reset. */
 774 static void ivshmem_msix_vector_use(IVShmemState *s)
 775 {
 776     PCIDevice *d = PCI_DEVICE(s);
 777     int i;
 778
 779     for (i = 0; i < s->vectors; i++) {
 780         msix_vector_use(d, i);
 781     }
 782 }
 783
 784 static void ivshmem_reset(DeviceState *d)
 785 {
 786     IVShmemState *s = IVSHMEM(d);
 787
 788     s->intrstatus = 0;
 789     s->intrmask = 0;
 790     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 791         ivshmem_msix_vector_use(s);
 792     }
 793 }
 794
 795 static int ivshmem_setup_interrupts(IVShmemState *s)
 796 {
 797     /* allocate QEMU callback data for receiving interrupts */
 798     s->msi_vectors = g_malloc0(s->vectors * sizeof(MSIVector));
 799
 800     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 801         if (msix_init_exclusive_bar(PCI_DEVICE(s), s->vectors, 1)) {
 802             return -1;
 803         }
 804
 805         IVSHMEM_DPRINTF("msix initialized (%d vectors)\n", s->vectors);
 806         ivshmem_msix_vector_use(s);
 807     }
 808
 809     return 0;
 810 }
 811
 812 static void ivshmem_enable_irqfd(IVShmemState *s)
 813 {
 814     PCIDevice *pdev = PCI_DEVICE(s);
 815     int i;
 816
 817     for (i = 0; i < s->peers[s->vm_id].nb_eventfds; i++) {
 818         Error *err = NULL;
 819
 820         ivshmem_add_kvm_msi_virq(s, i, &err);
 821         if (err) {
 822             error_report_err(err);
 823             /* TODO do we need to handle the error? */
 824         }
 825     }
 826
 827     if (msix_set_vector_notifiers(pdev,
 828                                   ivshmem_vector_unmask,
 829                                   ivshmem_vector_mask,
 830                                   ivshmem_vector_poll)) {
 831         error_report("ivshmem: msix_set_vector_notifiers failed");
 832     }
 833 }
 834
 835 static void ivshmem_remove_kvm_msi_virq(IVShmemState *s, int vector)
 836 {
 837     IVSHMEM_DPRINTF("ivshmem_remove_kvm_msi_virq vector:%d\n", vector);
 838
 839     if (s->msi_vectors[vector].pdev == NULL) {
 840         return;
 841     }
 842
 843     /* it was cleaned when masked in the frontend. */
 844     kvm_irqchip_release_virq(kvm_state, s->msi_vectors[vector].virq);
 845
 846     s->msi_vectors[vector].pdev = NULL;
 847 }
 848
 849 static void ivshmem_disable_irqfd(IVShmemState *s)
 850 {
 851     PCIDevice *pdev = PCI_DEVICE(s);
 852     int i;
 853
 854     for (i = 0; i < s->peers[s->vm_id].nb_eventfds; i++) {
 855         ivshmem_remove_kvm_msi_virq(s, i);
 856     }
 857
 858     msix_unset_vector_notifiers(pdev);
 859 }
 860
 861 static void ivshmem_write_config(PCIDevice *pdev, uint32_t address,
 862                                  uint32_t val, int len)
 863 {
 864     IVShmemState *s = IVSHMEM(pdev);
 865     int is_enabled, was_enabled = msix_enabled(pdev);
 866
 867     pci_default_write_config(pdev, address, val, len);
 868     is_enabled = msix_enabled(pdev);
 869
 870     if (kvm_msi_via_irqfd_enabled()) {
 871         if (!was_enabled && is_enabled) {
 872             ivshmem_enable_irqfd(s);
 873         } else if (was_enabled && !is_enabled) {
 874             ivshmem_disable_irqfd(s);
 875         }
 876     }
 877 }
 878
 879 static void pci_ivshmem_realize(PCIDevice *dev, Error **errp)
 880 {
 881     IVShmemState *s = IVSHMEM(dev);
 882     Error *err = NULL;
 883     uint8_t *pci_conf;
 884     uint8_t attr = PCI_BASE_ADDRESS_SPACE_MEMORY |
 885         PCI_BASE_ADDRESS_MEM_PREFETCH;
 886
 887     if (!!s->server_chr + !!s->shmobj + !!s->hostmem != 1) {
 888         error_setg(errp,
 889                    "You must specify either 'shm', 'chardev' or 'x-memdev'");
 890         return;
 891     }
 892
 893     if (s->hostmem) {
 894         MemoryRegion *mr;
 895
 896         if (s->sizearg) {
 897             g_warning("size argument ignored with hostmem");
 898         }
 899
 900         mr = host_memory_backend_get_memory(s->hostmem, &error_abort);
 901         s->ivshmem_size = memory_region_size(mr);
 902     } else if (s->sizearg == NULL) {
 903         s->ivshmem_size = 4 << 20; /* 4 MB default */
 904     } else {
 905         char *end;
 906         int64_t size = qemu_strtosz(s->sizearg, &end);
 907         if (size < 0 || *end != '\0' || !is_power_of_2(size)) {
 908             error_setg(errp, "Invalid size %s", s->sizearg);
 909             return;
 910         }
 911         s->ivshmem_size = size;
 912     }
 913
 914     /* IRQFD requires MSI */
 915     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD) &&
 916         !ivshmem_has_feature(s, IVSHMEM_MSI)) {
 917         error_setg(errp, "ioeventfd/irqfd requires MSI");
 918         return;
 919     }
 920
 921     /* check that role is reasonable */
 922     if (s->role) {
 923         if (strncmp(s->role, "peer", 5) == 0) {
 924             s->role_val = IVSHMEM_PEER;
 925         } else if (strncmp(s->role, "master", 7) == 0) {
 926             s->role_val = IVSHMEM_MASTER;
 927         } else {
 928             error_setg(errp, "'role' must be 'peer' or 'master'");
 929             return;
 930         }
 931     } else {
 932         s->role_val = IVSHMEM_MASTER; /* default */
 933     }
 934
 935     pci_conf = dev->config;
 936     pci_conf[PCI_COMMAND] = PCI_COMMAND_IO | PCI_COMMAND_MEMORY;
 937
 938     /*
 939      * Note: we don't use INTx with IVSHMEM_MSI at all, so this is a
 940      * bald-faced lie then.  But it's a backwards compatible lie.
 941      */
 942     pci_config_set_interrupt_pin(pci_conf, 1);
 943
 944     memory_region_init_io(&s->ivshmem_mmio, OBJECT(s), &ivshmem_mmio_ops, s,
 945                           "ivshmem-mmio", IVSHMEM_REG_BAR_SIZE);
 946
 947     /* region for registers*/
 948     pci_register_bar(dev, 0, PCI_BASE_ADDRESS_SPACE_MEMORY,
 949                      &s->ivshmem_mmio);
 950
 951     memory_region_init(&s->bar, OBJECT(s), "ivshmem-bar2-container", s->ivshmem_size);
 952     if (s->ivshmem_64bit) {
 953         attr |= PCI_BASE_ADDRESS_MEM_TYPE_64;
 954     }
 955
 956     if (s->hostmem != NULL) {
 957         MemoryRegion *mr;
 958
 959         IVSHMEM_DPRINTF("using hostmem\n");
 960
 961         mr = host_memory_backend_get_memory(MEMORY_BACKEND(s->hostmem),
 962                                             &error_abort);
 963         vmstate_register_ram(mr, DEVICE(s));
 964         memory_region_add_subregion(&s->bar, 0, mr);
 965         pci_register_bar(PCI_DEVICE(s), 2, attr, &s->bar);
 966     } else if (s->server_chr != NULL) {
 967         IVSHMEM_DPRINTF("using shared memory server (socket = %s)\n",
 968                         s->server_chr->filename);
 969
 970         /* we allocate enough space for 16 peers and grow as needed */
 971         resize_peers(s, 16);
 972
 973         pci_register_bar(dev, 2, attr, &s->bar);
 974
 975         /*
 976          * Receive setup messages from server synchronously.
 977          * Older versions did it asynchronously, but that creates a
 978          * number of entertaining race conditions.
 979          */
 980         ivshmem_recv_setup(s, &err);
 981         if (err) {
 982             error_propagate(errp, err);
 983             return;
 984         }
 985
 986         qemu_chr_add_handlers(s->server_chr, ivshmem_can_receive,
 987                               ivshmem_read, NULL, s);
 988
 989         if (ivshmem_setup_interrupts(s) < 0) {
 990             error_setg(errp, "failed to initialize interrupts");
 991             return;
 992         }
 993     } else {
 994         /* just map the file immediately, we're not using a server */
 995         int fd;
 996
 997         IVSHMEM_DPRINTF("using shm_open (shm object = %s)\n", s->shmobj);
 998
 999         /* try opening with O_EXCL and if it succeeds zero the memory
1000          * by truncating to 0 */
1001         if ((fd = shm_open(s->shmobj, O_CREAT|O_RDWR|O_EXCL,
1002                         S_IRWXU|S_IRWXG|S_IRWXO)) > 0) {
1003            /* truncate file to length PCI device's memory */
1004             if (ftruncate(fd, s->ivshmem_size) != 0) {
1005                 error_report("could not truncate shared file");
1006             }
1007
1008         } else if ((fd = shm_open(s->shmobj, O_CREAT|O_RDWR,
1009                         S_IRWXU|S_IRWXG|S_IRWXO)) < 0) {
1010             error_setg(errp, "could not open shared file");
1011             return;
1012         }
1013
1014         if (check_shm_size(s, fd, errp) == -1) {
1015             return;
1016         }
1017
1018         create_shared_memory_BAR(s, fd, attr, &err);
1019         if (err) {
1020             error_propagate(errp, err);
1021             return;
1022         }
1023     }
1024
1025     fifo8_create(&s->incoming_fifo, sizeof(int64_t));
1026
1027     if (s->role_val == IVSHMEM_PEER) {
1028         error_setg(&s->migration_blocker,
1029                    "Migration is disabled when using feature 'peer mode' in device 'ivshmem'");
1030         migrate_add_blocker(s->migration_blocker);
1031     }
1032 }
1033
1034 static void pci_ivshmem_exit(PCIDevice *dev)
1035 {
1036     IVShmemState *s = IVSHMEM(dev);
1037     int i;
1038
1039     fifo8_destroy(&s->incoming_fifo);
1040
1041     if (s->migration_blocker) {
1042         migrate_del_blocker(s->migration_blocker);
1043         error_free(s->migration_blocker);
1044     }
1045
1046     if (memory_region_is_mapped(&s->ivshmem)) {
1047         if (!s->hostmem) {
1048             void *addr = memory_region_get_ram_ptr(&s->ivshmem);
1049             int fd;
1050
1051             if (munmap(addr, s->ivshmem_size) == -1) {
1052                 error_report("Failed to munmap shared memory %s",
1053                              strerror(errno));
1054             }
1055
1056             fd = qemu_get_ram_fd(memory_region_get_ram_addr(&s->ivshmem));
1057             if (fd != -1) {
1058                 close(fd);
1059             }
1060         }
1061
1062         vmstate_unregister_ram(&s->ivshmem, DEVICE(dev));
1063         memory_region_del_subregion(&s->bar, &s->ivshmem);
1064     }
1065
1066     if (s->peers) {
1067         for (i = 0; i < s->nb_peers; i++) {
1068             close_peer_eventfds(s, i);
1069         }
1070         g_free(s->peers);
1071     }
1072
1073     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
1074         msix_uninit_exclusive_bar(dev);
1075     }
1076
1077     g_free(s->msi_vectors);
1078 }
1079
1080 static bool test_msix(void *opaque, int version_id)
1081 {
1082     IVShmemState *s = opaque;
1083
1084     return ivshmem_has_feature(s, IVSHMEM_MSI);
1085 }
1086
1087 static bool test_no_msix(void *opaque, int version_id)
1088 {
1089     return !test_msix(opaque, version_id);
1090 }
1091
1092 static int ivshmem_pre_load(void *opaque)
1093 {
1094     IVShmemState *s = opaque;
1095
1096     if (s->role_val == IVSHMEM_PEER) {
1097         error_report("'peer' devices are not migratable");
1098         return -EINVAL;
1099     }
1100
1101     return 0;
1102 }
1103
1104 static int ivshmem_post_load(void *opaque, int version_id)
1105 {
1106     IVShmemState *s = opaque;
1107
1108     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
1109         ivshmem_msix_vector_use(s);
1110     }
1111     return 0;
1112 }
1113
1114 static int ivshmem_load_old(QEMUFile *f, void *opaque, int version_id)
1115 {
1116     IVShmemState *s = opaque;
1117     PCIDevice *pdev = PCI_DEVICE(s);
1118     int ret;
1119
1120     IVSHMEM_DPRINTF("ivshmem_load_old\n");
1121
1122     if (version_id != 0) {
1123         return -EINVAL;
1124     }
1125
1126     if (s->role_val == IVSHMEM_PEER) {
1127         error_report("'peer' devices are not migratable");
1128         return -EINVAL;
1129     }
1130
1131     ret = pci_device_load(pdev, f);
1132     if (ret) {
1133         return ret;
1134     }
1135
1136     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
1137         msix_load(pdev, f);
1138         ivshmem_msix_vector_use(s);
1139     } else {
1140         s->intrstatus = qemu_get_be32(f);
1141         s->intrmask = qemu_get_be32(f);
1142     }
1143
1144     return 0;
1145 }
1146
1147 static const VMStateDescription ivshmem_vmsd = {
1148     .name = "ivshmem",
1149     .version_id = 1,
1150     .minimum_version_id = 1,
1151     .pre_load = ivshmem_pre_load,
1152     .post_load = ivshmem_post_load,
1153     .fields = (VMStateField[]) {
1154         VMSTATE_PCI_DEVICE(parent_obj, IVShmemState),
1155
1156         VMSTATE_MSIX_TEST(parent_obj, IVShmemState, test_msix),
1157         VMSTATE_UINT32_TEST(intrstatus, IVShmemState, test_no_msix),
1158         VMSTATE_UINT32_TEST(intrmask, IVShmemState, test_no_msix),
1159
1160         VMSTATE_END_OF_LIST()
1161     },
1162     .load_state_old = ivshmem_load_old,
1163     .minimum_version_id_old = 0
1164 };
1165
1166 static Property ivshmem_properties[] = {
1167     DEFINE_PROP_CHR("chardev", IVShmemState, server_chr),
1168     DEFINE_PROP_STRING("size", IVShmemState, sizearg),
1169     DEFINE_PROP_UINT32("vectors", IVShmemState, vectors, 1),
1170     DEFINE_PROP_BIT("ioeventfd", IVShmemState, features, IVSHMEM_IOEVENTFD, false),
1171     DEFINE_PROP_BIT("msi", IVShmemState, features, IVSHMEM_MSI, true),
1172     DEFINE_PROP_STRING("shm", IVShmemState, shmobj),
1173     DEFINE_PROP_STRING("role", IVShmemState, role),
1174     DEFINE_PROP_UINT32("use64", IVShmemState, ivshmem_64bit, 1),
1175     DEFINE_PROP_END_OF_LIST(),
1176 };
1177
1178 static void ivshmem_class_init(ObjectClass *klass, void *data)
1179 {
1180     DeviceClass *dc = DEVICE_CLASS(klass);
1181     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
1182
1183     k->realize = pci_ivshmem_realize;
1184     k->exit = pci_ivshmem_exit;
1185     k->config_write = ivshmem_write_config;
1186     k->vendor_id = PCI_VENDOR_ID_IVSHMEM;
1187     k->device_id = PCI_DEVICE_ID_IVSHMEM;
1188     k->class_id = PCI_CLASS_MEMORY_RAM;
1189     dc->reset = ivshmem_reset;
1190     dc->props = ivshmem_properties;
1191     dc->vmsd = &ivshmem_vmsd;
1192     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
1193     dc->desc = "Inter-VM shared memory";
1194 }
1195
1196 static void ivshmem_check_memdev_is_busy(Object *obj, const char *name,
1197                                          Object *val, Error **errp)
1198 {
1199     MemoryRegion *mr;
1200
1201     mr = host_memory_backend_get_memory(MEMORY_BACKEND(val), &error_abort);
1202     if (memory_region_is_mapped(mr)) {
1203         char *path = object_get_canonical_path_component(val);
1204         error_setg(errp, "can't use already busy memdev: %s", path);
1205         g_free(path);
1206     } else {
1207         qdev_prop_allow_set_link_before_realize(obj, name, val, errp);
1208     }
1209 }
1210
1211 static void ivshmem_init(Object *obj)
1212 {
1213     IVShmemState *s = IVSHMEM(obj);
1214
1215     object_property_add_link(obj, "x-memdev", TYPE_MEMORY_BACKEND,
1216                              (Object **)&s->hostmem,
1217                              ivshmem_check_memdev_is_busy,
1218                              OBJ_PROP_LINK_UNREF_ON_RELEASE,
1219                              &error_abort);
1220 }
1221
1222 static const TypeInfo ivshmem_info = {
1223     .name          = TYPE_IVSHMEM,
1224     .parent        = TYPE_PCI_DEVICE,
1225     .instance_size = sizeof(IVShmemState),
1226     .instance_init = ivshmem_init,
1227     .class_init    = ivshmem_class_init,
1228 };
1229
1230 static void ivshmem_register_types(void)
1231 {
1232     type_register_static(&ivshmem_info);
1233 }
1234
1235 type_init(ivshmem_register_types)