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 G_MAXUINT16
  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; /* how many peers we have space for */
  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 /* accessing registers - based on rtl8139 */
 125 static void ivshmem_update_irq(IVShmemState *s)
 126 {
 127     PCIDevice *d = PCI_DEVICE(s);
 128     int isr;
 129     isr = (s->intrstatus & s->intrmask) & 0xffffffff;
 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
 146     ivshmem_update_irq(s);
 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
 155     return ret;
 156 }
 157
 158 static void ivshmem_IntrStatus_write(IVShmemState *s, uint32_t val)
 159 {
 160     IVSHMEM_DPRINTF("IntrStatus write(w) val = 0x%04x\n", val);
 161
 162     s->intrstatus = val;
 163
 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
 174     ivshmem_update_irq(s);
 175
 176     return ret;
 177 }
 178
 179 static void ivshmem_io_write(void *opaque, hwaddr addr,
 180                              uint64_t val, unsigned size)
 181 {
 182     IVShmemState *s = opaque;
 183
 184     uint16_t dest = val >> 16;
 185     uint16_t vector = val & 0xff;
 186
 187     addr &= 0xfc;
 188
 189     IVSHMEM_DPRINTF("writing to addr " TARGET_FMT_plx "\n", addr);
 190     switch (addr)
 191     {
 192         case INTRMASK:
 193             ivshmem_IntrMask_write(s, val);
 194             break;
 195
 196         case INTRSTATUS:
 197             ivshmem_IntrStatus_write(s, val);
 198             break;
 199
 200         case DOORBELL:
 201             /* check that dest VM ID is reasonable */
 202             if (dest >= s->nb_peers) {
 203                 IVSHMEM_DPRINTF("Invalid destination VM ID (%d)\n", dest);
 204                 break;
 205             }
 206
 207             /* check doorbell range */
 208             if (vector < s->peers[dest].nb_eventfds) {
 209                 IVSHMEM_DPRINTF("Notifying VM %d on vector %d\n", dest, vector);
 210                 event_notifier_set(&s->peers[dest].eventfds[vector]);
 211             } else {
 212                 IVSHMEM_DPRINTF("Invalid destination vector %d on VM %d\n",
 213                                 vector, dest);
 214             }
 215             break;
 216         default:
 217             IVSHMEM_DPRINTF("Unhandled write " TARGET_FMT_plx "\n", addr);
 218     }
 219 }
 220
 221 static uint64_t ivshmem_io_read(void *opaque, hwaddr addr,
 222                                 unsigned size)
 223 {
 224
 225     IVShmemState *s = opaque;
 226     uint32_t ret;
 227
 228     switch (addr)
 229     {
 230         case INTRMASK:
 231             ret = ivshmem_IntrMask_read(s);
 232             break;
 233
 234         case INTRSTATUS:
 235             ret = ivshmem_IntrStatus_read(s);
 236             break;
 237
 238         case IVPOSITION:
 239             /* return my VM ID if the memory is mapped */
 240             if (memory_region_is_mapped(&s->ivshmem)) {
 241                 ret = s->vm_id;
 242             } else {
 243                 ret = -1;
 244             }
 245             break;
 246
 247         default:
 248             IVSHMEM_DPRINTF("why are we reading " TARGET_FMT_plx "\n", addr);
 249             ret = 0;
 250     }
 251
 252     return ret;
 253 }
 254
 255 static const MemoryRegionOps ivshmem_mmio_ops = {
 256     .read = ivshmem_io_read,
 257     .write = ivshmem_io_write,
 258     .endianness = DEVICE_NATIVE_ENDIAN,
 259     .impl = {
 260         .min_access_size = 4,
 261         .max_access_size = 4,
 262     },
 263 };
 264
 265 static int ivshmem_can_receive(void * opaque)
 266 {
 267     return sizeof(int64_t);
 268 }
 269
 270 static void ivshmem_vector_notify(void *opaque)
 271 {
 272     MSIVector *entry = opaque;
 273     PCIDevice *pdev = entry->pdev;
 274     IVShmemState *s = IVSHMEM(pdev);
 275     int vector = entry - s->msi_vectors;
 276     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 277
 278     if (!event_notifier_test_and_clear(n)) {
 279         return;
 280     }
 281
 282     IVSHMEM_DPRINTF("interrupt on vector %p %d\n", pdev, vector);
 283     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 284         msix_notify(pdev, vector);
 285     } else {
 286         ivshmem_IntrStatus_write(s, 1);
 287     }
 288 }
 289
 290 static int ivshmem_vector_unmask(PCIDevice *dev, unsigned vector,
 291                                  MSIMessage msg)
 292 {
 293     IVShmemState *s = IVSHMEM(dev);
 294     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 295     MSIVector *v = &s->msi_vectors[vector];
 296     int ret;
 297
 298     IVSHMEM_DPRINTF("vector unmask %p %d\n", dev, vector);
 299
 300     ret = kvm_irqchip_update_msi_route(kvm_state, v->virq, msg, dev);
 301     if (ret < 0) {
 302         return ret;
 303     }
 304
 305     return kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, n, NULL, v->virq);
 306 }
 307
 308 static void ivshmem_vector_mask(PCIDevice *dev, unsigned vector)
 309 {
 310     IVShmemState *s = IVSHMEM(dev);
 311     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 312     int ret;
 313
 314     IVSHMEM_DPRINTF("vector mask %p %d\n", dev, vector);
 315
 316     ret = kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state, n,
 317                                                 s->msi_vectors[vector].virq);
 318     if (ret != 0) {
 319         error_report("remove_irqfd_notifier_gsi failed");
 320     }
 321 }
 322
 323 static void ivshmem_vector_poll(PCIDevice *dev,
 324                                 unsigned int vector_start,
 325                                 unsigned int vector_end)
 326 {
 327     IVShmemState *s = IVSHMEM(dev);
 328     unsigned int vector;
 329
 330     IVSHMEM_DPRINTF("vector poll %p %d-%d\n", dev, vector_start, vector_end);
 331
 332     vector_end = MIN(vector_end, s->vectors);
 333
 334     for (vector = vector_start; vector < vector_end; vector++) {
 335         EventNotifier *notifier = &s->peers[s->vm_id].eventfds[vector];
 336
 337         if (!msix_is_masked(dev, vector)) {
 338             continue;
 339         }
 340
 341         if (event_notifier_test_and_clear(notifier)) {
 342             msix_set_pending(dev, vector);
 343         }
 344     }
 345 }
 346
 347 static void watch_vector_notifier(IVShmemState *s, EventNotifier *n,
 348                                  int vector)
 349 {
 350     int eventfd = event_notifier_get_fd(n);
 351
 352     /* if MSI is supported we need multiple interrupts */
 353     s->msi_vectors[vector].pdev = PCI_DEVICE(s);
 354
 355     qemu_set_fd_handler(eventfd, ivshmem_vector_notify,
 356                         NULL, &s->msi_vectors[vector]);
 357 }
 358
 359 static int check_shm_size(IVShmemState *s, int fd, Error **errp)
 360 {
 361     /* check that the guest isn't going to try and map more memory than the
 362      * the object has allocated return -1 to indicate error */
 363
 364     struct stat buf;
 365
 366     if (fstat(fd, &buf) < 0) {
 367         error_setg(errp, "exiting: fstat on fd %d failed: %s",
 368                    fd, strerror(errno));
 369         return -1;
 370     }
 371
 372     if (s->ivshmem_size > buf.st_size) {
 373         error_setg(errp, "Requested memory size greater"
 374                    " than shared object size (%" PRIu64 " > %" PRIu64")",
 375                    s->ivshmem_size, (uint64_t)buf.st_size);
 376         return -1;
 377     } else {
 378         return 0;
 379     }
 380 }
 381
 382 /* create the shared memory BAR when we are not using the server, so we can
 383  * create the BAR and map the memory immediately */
 384 static int create_shared_memory_BAR(IVShmemState *s, int fd, uint8_t attr,
 385                                     Error **errp)
 386 {
 387     void * ptr;
 388
 389     ptr = mmap(0, s->ivshmem_size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
 390     if (ptr == MAP_FAILED) {
 391         error_setg_errno(errp, errno, "Failed to mmap shared memory");
 392         return -1;
 393     }
 394
 395     memory_region_init_ram_ptr(&s->ivshmem, OBJECT(s), "ivshmem.bar2",
 396                                s->ivshmem_size, ptr);
 397     qemu_set_ram_fd(memory_region_get_ram_addr(&s->ivshmem), fd);
 398     vmstate_register_ram(&s->ivshmem, DEVICE(s));
 399     memory_region_add_subregion(&s->bar, 0, &s->ivshmem);
 400
 401     /* region for shared memory */
 402     pci_register_bar(PCI_DEVICE(s), 2, attr, &s->bar);
 403
 404     return 0;
 405 }
 406
 407 static void ivshmem_add_eventfd(IVShmemState *s, int posn, int i)
 408 {
 409     memory_region_add_eventfd(&s->ivshmem_mmio,
 410                               DOORBELL,
 411                               4,
 412                               true,
 413                               (posn << 16) | i,
 414                               &s->peers[posn].eventfds[i]);
 415 }
 416
 417 static void ivshmem_del_eventfd(IVShmemState *s, int posn, int i)
 418 {
 419     memory_region_del_eventfd(&s->ivshmem_mmio,
 420                               DOORBELL,
 421                               4,
 422                               true,
 423                               (posn << 16) | i,
 424                               &s->peers[posn].eventfds[i]);
 425 }
 426
 427 static void close_peer_eventfds(IVShmemState *s, int posn)
 428 {
 429     int i, n;
 430
 431     if (!ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
 432         return;
 433     }
 434     if (posn < 0 || posn >= s->nb_peers) {
 435         error_report("invalid peer %d", posn);
 436         return;
 437     }
 438
 439     n = s->peers[posn].nb_eventfds;
 440
 441     memory_region_transaction_begin();
 442     for (i = 0; i < n; i++) {
 443         ivshmem_del_eventfd(s, posn, i);
 444     }
 445     memory_region_transaction_commit();
 446     for (i = 0; i < n; i++) {
 447         event_notifier_cleanup(&s->peers[posn].eventfds[i]);
 448     }
 449
 450     g_free(s->peers[posn].eventfds);
 451     s->peers[posn].nb_eventfds = 0;
 452 }
 453
 454 /* this function increase the dynamic storage need to store data about other
 455  * peers */
 456 static int resize_peers(IVShmemState *s, int new_min_size)
 457 {
 458
 459     int j, old_size;
 460
 461     /* limit number of max peers */
 462     if (new_min_size <= 0 || new_min_size > IVSHMEM_MAX_PEERS) {
 463         return -1;
 464     }
 465     if (new_min_size <= s->nb_peers) {
 466         return 0;
 467     }
 468
 469     old_size = s->nb_peers;
 470     s->nb_peers = new_min_size;
 471
 472     IVSHMEM_DPRINTF("bumping storage to %d peers\n", s->nb_peers);
 473
 474     s->peers = g_realloc(s->peers, s->nb_peers * sizeof(Peer));
 475
 476     for (j = old_size; j < s->nb_peers; j++) {
 477         s->peers[j].eventfds = g_new0(EventNotifier, s->vectors);
 478         s->peers[j].nb_eventfds = 0;
 479     }
 480
 481     return 0;
 482 }
 483
 484 static bool fifo_update_and_get(IVShmemState *s, const uint8_t *buf, int size,
 485                                 void *data, size_t len)
 486 {
 487     const uint8_t *p;
 488     uint32_t num;
 489
 490     assert(len <= sizeof(int64_t)); /* limitation of the fifo */
 491     if (fifo8_is_empty(&s->incoming_fifo) && size == len) {
 492         memcpy(data, buf, size);
 493         return true;
 494     }
 495
 496     IVSHMEM_DPRINTF("short read of %d bytes\n", size);
 497
 498     num = MIN(size, sizeof(int64_t) - fifo8_num_used(&s->incoming_fifo));
 499     fifo8_push_all(&s->incoming_fifo, buf, num);
 500
 501     if (fifo8_num_used(&s->incoming_fifo) < len) {
 502         assert(num == 0);
 503         return false;
 504     }
 505
 506     size -= num;
 507     buf += num;
 508     p = fifo8_pop_buf(&s->incoming_fifo, len, &num);
 509     assert(num == len);
 510
 511     memcpy(data, p, len);
 512
 513     if (size > 0) {
 514         fifo8_push_all(&s->incoming_fifo, buf, size);
 515     }
 516
 517     return true;
 518 }
 519
 520 static bool fifo_update_and_get_i64(IVShmemState *s,
 521                                     const uint8_t *buf, int size, int64_t *i64)
 522 {
 523     if (fifo_update_and_get(s, buf, size, i64, sizeof(*i64))) {
 524         *i64 = GINT64_FROM_LE(*i64);
 525         return true;
 526     }
 527
 528     return false;
 529 }
 530
 531 static int ivshmem_add_kvm_msi_virq(IVShmemState *s, int vector)
 532 {
 533     PCIDevice *pdev = PCI_DEVICE(s);
 534     MSIMessage msg = msix_get_message(pdev, vector);
 535     int ret;
 536
 537     IVSHMEM_DPRINTF("ivshmem_add_kvm_msi_virq vector:%d\n", vector);
 538
 539     if (s->msi_vectors[vector].pdev != NULL) {
 540         return 0;
 541     }
 542
 543     ret = kvm_irqchip_add_msi_route(kvm_state, msg, pdev);
 544     if (ret < 0) {
 545         error_report("ivshmem: kvm_irqchip_add_msi_route failed");
 546         return -1;
 547     }
 548
 549     s->msi_vectors[vector].virq = ret;
 550     s->msi_vectors[vector].pdev = pdev;
 551
 552     return 0;
 553 }
 554
 555 static void setup_interrupt(IVShmemState *s, int vector)
 556 {
 557     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 558     bool with_irqfd = kvm_msi_via_irqfd_enabled() &&
 559         ivshmem_has_feature(s, IVSHMEM_MSI);
 560     PCIDevice *pdev = PCI_DEVICE(s);
 561
 562     IVSHMEM_DPRINTF("setting up interrupt for vector: %d\n", vector);
 563
 564     if (!with_irqfd) {
 565         IVSHMEM_DPRINTF("with eventfd\n");
 566         watch_vector_notifier(s, n, vector);
 567     } else if (msix_enabled(pdev)) {
 568         IVSHMEM_DPRINTF("with irqfd\n");
 569         if (ivshmem_add_kvm_msi_virq(s, vector) < 0) {
 570             return;
 571         }
 572
 573         if (!msix_is_masked(pdev, vector)) {
 574             kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, n, NULL,
 575                                                s->msi_vectors[vector].virq);
 576         }
 577     } else {
 578         /* it will be delayed until msix is enabled, in write_config */
 579         IVSHMEM_DPRINTF("with irqfd, delayed until msix enabled\n");
 580     }
 581 }
 582
 583 static void ivshmem_read(void *opaque, const uint8_t *buf, int size)
 584 {
 585     IVShmemState *s = opaque;
 586     int incoming_fd;
 587     int new_eventfd;
 588     int64_t incoming_posn;
 589     Error *err = NULL;
 590     Peer *peer;
 591
 592     if (!fifo_update_and_get_i64(s, buf, size, &incoming_posn)) {
 593         return;
 594     }
 595
 596     if (incoming_posn < -1) {
 597         IVSHMEM_DPRINTF("invalid incoming_posn %" PRId64 "\n", incoming_posn);
 598         return;
 599     }
 600
 601     /* pick off s->server_chr->msgfd and store it, posn should accompany msg */
 602     incoming_fd = qemu_chr_fe_get_msgfd(s->server_chr);
 603     IVSHMEM_DPRINTF("posn is %" PRId64 ", fd is %d\n",
 604                     incoming_posn, incoming_fd);
 605
 606     /* make sure we have enough space for this peer */
 607     if (incoming_posn >= s->nb_peers) {
 608         if (resize_peers(s, incoming_posn + 1) < 0) {
 609             error_report("failed to resize peers array");
 610             if (incoming_fd != -1) {
 611                 close(incoming_fd);
 612             }
 613             return;
 614         }
 615     }
 616
 617     peer = &s->peers[incoming_posn];
 618
 619     if (incoming_fd == -1) {
 620         /* if posn is positive and unseen before then this is our posn*/
 621         if (incoming_posn >= 0 && s->vm_id == -1) {
 622             /* receive our posn */
 623             s->vm_id = incoming_posn;
 624         } else {
 625             /* otherwise an fd == -1 means an existing peer has gone away */
 626             IVSHMEM_DPRINTF("posn %" PRId64 " has gone away\n", incoming_posn);
 627             close_peer_eventfds(s, incoming_posn);
 628         }
 629         return;
 630     }
 631
 632     /* if the position is -1, then it's shared memory region fd */
 633     if (incoming_posn == -1) {
 634         void * map_ptr;
 635
 636         if (memory_region_is_mapped(&s->ivshmem)) {
 637             error_report("shm already initialized");
 638             close(incoming_fd);
 639             return;
 640         }
 641
 642         if (check_shm_size(s, incoming_fd, &err) == -1) {
 643             error_report_err(err);
 644             close(incoming_fd);
 645             return;
 646         }
 647
 648         /* mmap the region and map into the BAR2 */
 649         map_ptr = mmap(0, s->ivshmem_size, PROT_READ|PROT_WRITE, MAP_SHARED,
 650                                                             incoming_fd, 0);
 651         if (map_ptr == MAP_FAILED) {
 652             error_report("Failed to mmap shared memory %s", strerror(errno));
 653             close(incoming_fd);
 654             return;
 655         }
 656         memory_region_init_ram_ptr(&s->ivshmem, OBJECT(s),
 657                                    "ivshmem.bar2", s->ivshmem_size, map_ptr);
 658         qemu_set_ram_fd(memory_region_get_ram_addr(&s->ivshmem),
 659                         incoming_fd);
 660         vmstate_register_ram(&s->ivshmem, DEVICE(s));
 661
 662         IVSHMEM_DPRINTF("guest h/w addr = %p, size = %" PRIu64 "\n",
 663                         map_ptr, s->ivshmem_size);
 664
 665         memory_region_add_subregion(&s->bar, 0, &s->ivshmem);
 666
 667         return;
 668     }
 669
 670     /* each peer has an associated array of eventfds, and we keep
 671      * track of how many eventfds received so far */
 672     /* get a new eventfd: */
 673     if (peer->nb_eventfds >= s->vectors) {
 674         error_report("Too many eventfd received, device has %d vectors",
 675                      s->vectors);
 676         close(incoming_fd);
 677         return;
 678     }
 679
 680     new_eventfd = peer->nb_eventfds++;
 681
 682     /* this is an eventfd for a particular peer VM */
 683     IVSHMEM_DPRINTF("eventfds[%" PRId64 "][%d] = %d\n", incoming_posn,
 684                     new_eventfd, incoming_fd);
 685     event_notifier_init_fd(&peer->eventfds[new_eventfd], incoming_fd);
 686     fcntl_setfl(incoming_fd, O_NONBLOCK); /* msix/irqfd poll non block */
 687
 688     if (incoming_posn == s->vm_id) {
 689         setup_interrupt(s, new_eventfd);
 690     }
 691
 692     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
 693         ivshmem_add_eventfd(s, incoming_posn, new_eventfd);
 694     }
 695 }
 696
 697 static void ivshmem_check_version(void *opaque, const uint8_t * buf, int size)
 698 {
 699     IVShmemState *s = opaque;
 700     int tmp;
 701     int64_t version;
 702
 703     if (!fifo_update_and_get_i64(s, buf, size, &version)) {
 704         return;
 705     }
 706
 707     tmp = qemu_chr_fe_get_msgfd(s->server_chr);
 708     if (tmp != -1 || version != IVSHMEM_PROTOCOL_VERSION) {
 709         fprintf(stderr, "incompatible version, you are connecting to a ivshmem-"
 710                 "server using a different protocol please check your setup\n");
 711         qemu_chr_add_handlers(s->server_chr, NULL, NULL, NULL, s);
 712         return;
 713     }
 714
 715     IVSHMEM_DPRINTF("version check ok, switch to real chardev handler\n");
 716     qemu_chr_add_handlers(s->server_chr, ivshmem_can_receive, ivshmem_read,
 717                           NULL, s);
 718 }
 719
 720 /* Select the MSI-X vectors used by device.
 721  * ivshmem maps events to vectors statically, so
 722  * we just enable all vectors on init and after reset. */
 723 static void ivshmem_use_msix(IVShmemState * s)
 724 {
 725     PCIDevice *d = PCI_DEVICE(s);
 726     int i;
 727
 728     IVSHMEM_DPRINTF("%s, msix present: %d\n", __func__, msix_present(d));
 729     if (!msix_present(d)) {
 730         return;
 731     }
 732
 733     for (i = 0; i < s->vectors; i++) {
 734         msix_vector_use(d, i);
 735     }
 736 }
 737
 738 static void ivshmem_reset(DeviceState *d)
 739 {
 740     IVShmemState *s = IVSHMEM(d);
 741
 742     s->intrstatus = 0;
 743     s->intrmask = 0;
 744     ivshmem_use_msix(s);
 745 }
 746
 747 static int ivshmem_setup_interrupts(IVShmemState *s)
 748 {
 749     /* allocate QEMU callback data for receiving interrupts */
 750     s->msi_vectors = g_malloc0(s->vectors * sizeof(MSIVector));
 751
 752     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 753         if (msix_init_exclusive_bar(PCI_DEVICE(s), s->vectors, 1)) {
 754             return -1;
 755         }
 756
 757         IVSHMEM_DPRINTF("msix initialized (%d vectors)\n", s->vectors);
 758         ivshmem_use_msix(s);
 759     }
 760
 761     return 0;
 762 }
 763
 764 static void ivshmem_enable_irqfd(IVShmemState *s)
 765 {
 766     PCIDevice *pdev = PCI_DEVICE(s);
 767     int i;
 768
 769     for (i = 0; i < s->peers[s->vm_id].nb_eventfds; i++) {
 770         ivshmem_add_kvm_msi_virq(s, i);
 771     }
 772
 773     if (msix_set_vector_notifiers(pdev,
 774                                   ivshmem_vector_unmask,
 775                                   ivshmem_vector_mask,
 776                                   ivshmem_vector_poll)) {
 777         error_report("ivshmem: msix_set_vector_notifiers failed");
 778     }
 779 }
 780
 781 static void ivshmem_remove_kvm_msi_virq(IVShmemState *s, int vector)
 782 {
 783     IVSHMEM_DPRINTF("ivshmem_remove_kvm_msi_virq vector:%d\n", vector);
 784
 785     if (s->msi_vectors[vector].pdev == NULL) {
 786         return;
 787     }
 788
 789     /* it was cleaned when masked in the frontend. */
 790     kvm_irqchip_release_virq(kvm_state, s->msi_vectors[vector].virq);
 791
 792     s->msi_vectors[vector].pdev = NULL;
 793 }
 794
 795 static void ivshmem_disable_irqfd(IVShmemState *s)
 796 {
 797     PCIDevice *pdev = PCI_DEVICE(s);
 798     int i;
 799
 800     for (i = 0; i < s->peers[s->vm_id].nb_eventfds; i++) {
 801         ivshmem_remove_kvm_msi_virq(s, i);
 802     }
 803
 804     msix_unset_vector_notifiers(pdev);
 805 }
 806
 807 static void ivshmem_write_config(PCIDevice *pdev, uint32_t address,
 808                                  uint32_t val, int len)
 809 {
 810     IVShmemState *s = IVSHMEM(pdev);
 811     int is_enabled, was_enabled = msix_enabled(pdev);
 812
 813     pci_default_write_config(pdev, address, val, len);
 814     is_enabled = msix_enabled(pdev);
 815
 816     if (kvm_msi_via_irqfd_enabled() && s->vm_id != -1) {
 817         if (!was_enabled && is_enabled) {
 818             ivshmem_enable_irqfd(s);
 819         } else if (was_enabled && !is_enabled) {
 820             ivshmem_disable_irqfd(s);
 821         }
 822     }
 823 }
 824
 825 static void pci_ivshmem_realize(PCIDevice *dev, Error **errp)
 826 {
 827     IVShmemState *s = IVSHMEM(dev);
 828     Error *err = NULL;
 829     uint8_t *pci_conf;
 830     uint8_t attr = PCI_BASE_ADDRESS_SPACE_MEMORY |
 831         PCI_BASE_ADDRESS_MEM_PREFETCH;
 832
 833     if (!!s->server_chr + !!s->shmobj + !!s->hostmem != 1) {
 834         error_setg(errp,
 835                    "You must specify either 'shm', 'chardev' or 'x-memdev'");
 836         return;
 837     }
 838
 839     if (s->hostmem) {
 840         MemoryRegion *mr;
 841
 842         if (s->sizearg) {
 843             g_warning("size argument ignored with hostmem");
 844         }
 845
 846         mr = host_memory_backend_get_memory(s->hostmem, &error_abort);
 847         s->ivshmem_size = memory_region_size(mr);
 848     } else if (s->sizearg == NULL) {
 849         s->ivshmem_size = 4 << 20; /* 4 MB default */
 850     } else {
 851         char *end;
 852         int64_t size = qemu_strtosz(s->sizearg, &end);
 853         if (size < 0 || *end != '\0' || !is_power_of_2(size)) {
 854             error_setg(errp, "Invalid size %s", s->sizearg);
 855             return;
 856         }
 857         s->ivshmem_size = size;
 858     }
 859
 860     /* IRQFD requires MSI */
 861     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD) &&
 862         !ivshmem_has_feature(s, IVSHMEM_MSI)) {
 863         error_setg(errp, "ioeventfd/irqfd requires MSI");
 864         return;
 865     }
 866
 867     /* check that role is reasonable */
 868     if (s->role) {
 869         if (strncmp(s->role, "peer", 5) == 0) {
 870             s->role_val = IVSHMEM_PEER;
 871         } else if (strncmp(s->role, "master", 7) == 0) {
 872             s->role_val = IVSHMEM_MASTER;
 873         } else {
 874             error_setg(errp, "'role' must be 'peer' or 'master'");
 875             return;
 876         }
 877     } else {
 878         s->role_val = IVSHMEM_MASTER; /* default */
 879     }
 880
 881     pci_conf = dev->config;
 882     pci_conf[PCI_COMMAND] = PCI_COMMAND_IO | PCI_COMMAND_MEMORY;
 883
 884     pci_config_set_interrupt_pin(pci_conf, 1);
 885
 886     memory_region_init_io(&s->ivshmem_mmio, OBJECT(s), &ivshmem_mmio_ops, s,
 887                           "ivshmem-mmio", IVSHMEM_REG_BAR_SIZE);
 888
 889     /* region for registers*/
 890     pci_register_bar(dev, 0, PCI_BASE_ADDRESS_SPACE_MEMORY,
 891                      &s->ivshmem_mmio);
 892
 893     memory_region_init(&s->bar, OBJECT(s), "ivshmem-bar2-container", s->ivshmem_size);
 894     if (s->ivshmem_64bit) {
 895         attr |= PCI_BASE_ADDRESS_MEM_TYPE_64;
 896     }
 897
 898     if (s->hostmem != NULL) {
 899         MemoryRegion *mr;
 900
 901         IVSHMEM_DPRINTF("using hostmem\n");
 902
 903         mr = host_memory_backend_get_memory(MEMORY_BACKEND(s->hostmem),
 904                                             &error_abort);
 905         vmstate_register_ram(mr, DEVICE(s));
 906         memory_region_add_subregion(&s->bar, 0, mr);
 907         pci_register_bar(PCI_DEVICE(s), 2, attr, &s->bar);
 908     } else if (s->server_chr != NULL) {
 909         /* FIXME do not rely on what chr drivers put into filename */
 910         if (strncmp(s->server_chr->filename, "unix:", 5)) {
 911             error_setg(errp, "chardev is not a unix client socket");
 912             return;
 913         }
 914
 915         /* if we get a UNIX socket as the parameter we will talk
 916          * to the ivshmem server to receive the memory region */
 917
 918         IVSHMEM_DPRINTF("using shared memory server (socket = %s)\n",
 919                         s->server_chr->filename);
 920
 921         if (ivshmem_setup_interrupts(s) < 0) {
 922             error_setg(errp, "failed to initialize interrupts");
 923             return;
 924         }
 925
 926         /* we allocate enough space for 16 peers and grow as needed */
 927         resize_peers(s, 16);
 928         s->vm_id = -1;
 929
 930         pci_register_bar(dev, 2, attr, &s->bar);
 931
 932         qemu_chr_add_handlers(s->server_chr, ivshmem_can_receive,
 933                               ivshmem_check_version, NULL, s);
 934     } else {
 935         /* just map the file immediately, we're not using a server */
 936         int fd;
 937
 938         IVSHMEM_DPRINTF("using shm_open (shm object = %s)\n", s->shmobj);
 939
 940         /* try opening with O_EXCL and if it succeeds zero the memory
 941          * by truncating to 0 */
 942         if ((fd = shm_open(s->shmobj, O_CREAT|O_RDWR|O_EXCL,
 943                         S_IRWXU|S_IRWXG|S_IRWXO)) > 0) {
 944            /* truncate file to length PCI device's memory */
 945             if (ftruncate(fd, s->ivshmem_size) != 0) {
 946                 error_report("could not truncate shared file");
 947             }
 948
 949         } else if ((fd = shm_open(s->shmobj, O_CREAT|O_RDWR,
 950                         S_IRWXU|S_IRWXG|S_IRWXO)) < 0) {
 951             error_setg(errp, "could not open shared file");
 952             return;
 953         }
 954
 955         if (check_shm_size(s, fd, errp) == -1) {
 956             return;
 957         }
 958
 959         create_shared_memory_BAR(s, fd, attr, &err);
 960         if (err) {
 961             error_propagate(errp, err);
 962             return;
 963         }
 964     }
 965
 966     fifo8_create(&s->incoming_fifo, sizeof(int64_t));
 967
 968     if (s->role_val == IVSHMEM_PEER) {
 969         error_setg(&s->migration_blocker,
 970                    "Migration is disabled when using feature 'peer mode' in device 'ivshmem'");
 971         migrate_add_blocker(s->migration_blocker);
 972     }
 973 }
 974
 975 static void pci_ivshmem_exit(PCIDevice *dev)
 976 {
 977     IVShmemState *s = IVSHMEM(dev);
 978     int i;
 979
 980     fifo8_destroy(&s->incoming_fifo);
 981
 982     if (s->migration_blocker) {
 983         migrate_del_blocker(s->migration_blocker);
 984         error_free(s->migration_blocker);
 985     }
 986
 987     if (memory_region_is_mapped(&s->ivshmem)) {
 988         if (!s->hostmem) {
 989             void *addr = memory_region_get_ram_ptr(&s->ivshmem);
 990             int fd;
 991
 992             if (munmap(addr, s->ivshmem_size) == -1) {
 993                 error_report("Failed to munmap shared memory %s",
 994                              strerror(errno));
 995             }
 996
 997             fd = qemu_get_ram_fd(memory_region_get_ram_addr(&s->ivshmem));
 998             if (fd != -1) {
 999                 close(fd);
1000             }
1001         }
1002
1003         vmstate_unregister_ram(&s->ivshmem, DEVICE(dev));
1004         memory_region_del_subregion(&s->bar, &s->ivshmem);
1005     }
1006
1007     if (s->peers) {
1008         for (i = 0; i < s->nb_peers; i++) {
1009             close_peer_eventfds(s, i);
1010         }
1011         g_free(s->peers);
1012     }
1013
1014     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
1015         msix_uninit_exclusive_bar(dev);
1016     }
1017
1018     g_free(s->msi_vectors);
1019 }
1020
1021 static bool test_msix(void *opaque, int version_id)
1022 {
1023     IVShmemState *s = opaque;
1024
1025     return ivshmem_has_feature(s, IVSHMEM_MSI);
1026 }
1027
1028 static bool test_no_msix(void *opaque, int version_id)
1029 {
1030     return !test_msix(opaque, version_id);
1031 }
1032
1033 static int ivshmem_pre_load(void *opaque)
1034 {
1035     IVShmemState *s = opaque;
1036
1037     if (s->role_val == IVSHMEM_PEER) {
1038         error_report("'peer' devices are not migratable");
1039         return -EINVAL;
1040     }
1041
1042     return 0;
1043 }
1044
1045 static int ivshmem_post_load(void *opaque, int version_id)
1046 {
1047     IVShmemState *s = opaque;
1048
1049     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
1050         ivshmem_use_msix(s);
1051     }
1052
1053     return 0;
1054 }
1055
1056 static int ivshmem_load_old(QEMUFile *f, void *opaque, int version_id)
1057 {
1058     IVShmemState *s = opaque;
1059     PCIDevice *pdev = PCI_DEVICE(s);
1060     int ret;
1061
1062     IVSHMEM_DPRINTF("ivshmem_load_old\n");
1063
1064     if (version_id != 0) {
1065         return -EINVAL;
1066     }
1067
1068     if (s->role_val == IVSHMEM_PEER) {
1069         error_report("'peer' devices are not migratable");
1070         return -EINVAL;
1071     }
1072
1073     ret = pci_device_load(pdev, f);
1074     if (ret) {
1075         return ret;
1076     }
1077
1078     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
1079         msix_load(pdev, f);
1080         ivshmem_use_msix(s);
1081     } else {
1082         s->intrstatus = qemu_get_be32(f);
1083         s->intrmask = qemu_get_be32(f);
1084     }
1085
1086     return 0;
1087 }
1088
1089 static const VMStateDescription ivshmem_vmsd = {
1090     .name = "ivshmem",
1091     .version_id = 1,
1092     .minimum_version_id = 1,
1093     .pre_load = ivshmem_pre_load,
1094     .post_load = ivshmem_post_load,
1095     .fields = (VMStateField[]) {
1096         VMSTATE_PCI_DEVICE(parent_obj, IVShmemState),
1097
1098         VMSTATE_MSIX_TEST(parent_obj, IVShmemState, test_msix),
1099         VMSTATE_UINT32_TEST(intrstatus, IVShmemState, test_no_msix),
1100         VMSTATE_UINT32_TEST(intrmask, IVShmemState, test_no_msix),
1101
1102         VMSTATE_END_OF_LIST()
1103     },
1104     .load_state_old = ivshmem_load_old,
1105     .minimum_version_id_old = 0
1106 };
1107
1108 static Property ivshmem_properties[] = {
1109     DEFINE_PROP_CHR("chardev", IVShmemState, server_chr),
1110     DEFINE_PROP_STRING("size", IVShmemState, sizearg),
1111     DEFINE_PROP_UINT32("vectors", IVShmemState, vectors, 1),
1112     DEFINE_PROP_BIT("ioeventfd", IVShmemState, features, IVSHMEM_IOEVENTFD, false),
1113     DEFINE_PROP_BIT("msi", IVShmemState, features, IVSHMEM_MSI, true),
1114     DEFINE_PROP_STRING("shm", IVShmemState, shmobj),
1115     DEFINE_PROP_STRING("role", IVShmemState, role),
1116     DEFINE_PROP_UINT32("use64", IVShmemState, ivshmem_64bit, 1),
1117     DEFINE_PROP_END_OF_LIST(),
1118 };
1119
1120 static void ivshmem_class_init(ObjectClass *klass, void *data)
1121 {
1122     DeviceClass *dc = DEVICE_CLASS(klass);
1123     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
1124
1125     k->realize = pci_ivshmem_realize;
1126     k->exit = pci_ivshmem_exit;
1127     k->config_write = ivshmem_write_config;
1128     k->vendor_id = PCI_VENDOR_ID_IVSHMEM;
1129     k->device_id = PCI_DEVICE_ID_IVSHMEM;
1130     k->class_id = PCI_CLASS_MEMORY_RAM;
1131     dc->reset = ivshmem_reset;
1132     dc->props = ivshmem_properties;
1133     dc->vmsd = &ivshmem_vmsd;
1134     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
1135     dc->desc = "Inter-VM shared memory";
1136 }
1137
1138 static void ivshmem_check_memdev_is_busy(Object *obj, const char *name,
1139                                          Object *val, Error **errp)
1140 {
1141     MemoryRegion *mr;
1142
1143     mr = host_memory_backend_get_memory(MEMORY_BACKEND(val), &error_abort);
1144     if (memory_region_is_mapped(mr)) {
1145         char *path = object_get_canonical_path_component(val);
1146         error_setg(errp, "can't use already busy memdev: %s", path);
1147         g_free(path);
1148     } else {
1149         qdev_prop_allow_set_link_before_realize(obj, name, val, errp);
1150     }
1151 }
1152
1153 static void ivshmem_init(Object *obj)
1154 {
1155     IVShmemState *s = IVSHMEM(obj);
1156
1157     object_property_add_link(obj, "x-memdev", TYPE_MEMORY_BACKEND,
1158                              (Object **)&s->hostmem,
1159                              ivshmem_check_memdev_is_busy,
1160                              OBJ_PROP_LINK_UNREF_ON_RELEASE,
1161                              &error_abort);
1162 }
1163
1164 static const TypeInfo ivshmem_info = {
1165     .name          = TYPE_IVSHMEM,
1166     .parent        = TYPE_PCI_DEVICE,
1167     .instance_size = sizeof(IVShmemState),
1168     .instance_init = ivshmem_init,
1169     .class_init    = ivshmem_class_init,
1170 };
1171
1172 static void ivshmem_register_types(void)
1173 {
1174     type_register_static(&ivshmem_info);
1175 }
1176
1177 type_init(ivshmem_register_types)