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 DEBUG_IVSHMEM
  52 #ifdef DEBUG_IVSHMEM
  53 #define IVSHMEM_DPRINTF(fmt, ...)        \
  54     do {printf("IVSHMEM: " fmt, ## __VA_ARGS__); } while (0)
  55 #else
  56 #define IVSHMEM_DPRINTF(fmt, ...)
  57 #endif
  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 **eventfd_chr;
  83     CharDriverState *server_chr;
  84     Fifo8 incoming_fifo;
  85     MemoryRegion ivshmem_mmio;
  86
  87     /* We might need to register the BAR before we actually have the memory.
  88      * So prepare a container MemoryRegion for the BAR immediately and
  89      * add a subregion when we have the memory.
  90      */
  91     MemoryRegion bar;
  92     MemoryRegion ivshmem;
  93     uint64_t ivshmem_size; /* size of shared memory region */
  94     uint32_t ivshmem_64bit;
  95
  96     Peer *peers;
  97     int nb_peers; /* how many peers we have space for */
  98
  99     int vm_id;
 100     uint32_t vectors;
 101     uint32_t features;
 102     MSIVector *msi_vectors;
 103
 104     Error *migration_blocker;
 105
 106     char * shmobj;
 107     char * sizearg;
 108     char * role;
 109     int role_val;   /* scalar to avoid multiple string comparisons */
 110 } IVShmemState;
 111
 112 /* registers for the Inter-VM shared memory device */
 113 enum ivshmem_registers {
 114     INTRMASK = 0,
 115     INTRSTATUS = 4,
 116     IVPOSITION = 8,
 117     DOORBELL = 12,
 118 };
 119
 120 static inline uint32_t ivshmem_has_feature(IVShmemState *ivs,
 121                                                     unsigned int feature) {
 122     return (ivs->features & (1 << feature));
 123 }
 124
 125 /* accessing registers - based on rtl8139 */
 126 static void ivshmem_update_irq(IVShmemState *s)
 127 {
 128     PCIDevice *d = PCI_DEVICE(s);
 129     int isr;
 130     isr = (s->intrstatus & s->intrmask) & 0xffffffff;
 131
 132     /* don't print ISR resets */
 133     if (isr) {
 134         IVSHMEM_DPRINTF("Set IRQ to %d (%04x %04x)\n",
 135                         isr ? 1 : 0, s->intrstatus, s->intrmask);
 136     }
 137
 138     pci_set_irq(d, (isr != 0));
 139 }
 140
 141 static void ivshmem_IntrMask_write(IVShmemState *s, uint32_t val)
 142 {
 143     IVSHMEM_DPRINTF("IntrMask write(w) val = 0x%04x\n", val);
 144
 145     s->intrmask = val;
 146
 147     ivshmem_update_irq(s);
 148 }
 149
 150 static uint32_t ivshmem_IntrMask_read(IVShmemState *s)
 151 {
 152     uint32_t ret = s->intrmask;
 153
 154     IVSHMEM_DPRINTF("intrmask read(w) val = 0x%04x\n", ret);
 155
 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
 165     ivshmem_update_irq(s);
 166 }
 167
 168 static uint32_t ivshmem_IntrStatus_read(IVShmemState *s)
 169 {
 170     uint32_t ret = s->intrstatus;
 171
 172     /* reading ISR clears all interrupts */
 173     s->intrstatus = 0;
 174
 175     ivshmem_update_irq(s);
 176
 177     return ret;
 178 }
 179
 180 static void ivshmem_io_write(void *opaque, hwaddr addr,
 181                              uint64_t val, unsigned size)
 182 {
 183     IVShmemState *s = opaque;
 184
 185     uint16_t dest = val >> 16;
 186     uint16_t vector = val & 0xff;
 187
 188     addr &= 0xfc;
 189
 190     IVSHMEM_DPRINTF("writing to addr " TARGET_FMT_plx "\n", addr);
 191     switch (addr)
 192     {
 193         case INTRMASK:
 194             ivshmem_IntrMask_write(s, val);
 195             break;
 196
 197         case INTRSTATUS:
 198             ivshmem_IntrStatus_write(s, val);
 199             break;
 200
 201         case DOORBELL:
 202             /* check that dest VM ID is reasonable */
 203             if (dest >= s->nb_peers) {
 204                 IVSHMEM_DPRINTF("Invalid destination VM ID (%d)\n", dest);
 205                 break;
 206             }
 207
 208             /* check doorbell range */
 209             if (vector < s->peers[dest].nb_eventfds) {
 210                 IVSHMEM_DPRINTF("Notifying VM %d on vector %d\n", dest, vector);
 211                 event_notifier_set(&s->peers[dest].eventfds[vector]);
 212             } else {
 213                 IVSHMEM_DPRINTF("Invalid destination vector %d on VM %d\n",
 214                                 vector, dest);
 215             }
 216             break;
 217         default:
 218             IVSHMEM_DPRINTF("Unhandled write " TARGET_FMT_plx "\n", addr);
 219     }
 220 }
 221
 222 static uint64_t ivshmem_io_read(void *opaque, hwaddr addr,
 223                                 unsigned size)
 224 {
 225
 226     IVShmemState *s = opaque;
 227     uint32_t ret;
 228
 229     switch (addr)
 230     {
 231         case INTRMASK:
 232             ret = ivshmem_IntrMask_read(s);
 233             break;
 234
 235         case INTRSTATUS:
 236             ret = ivshmem_IntrStatus_read(s);
 237             break;
 238
 239         case IVPOSITION:
 240             /* return my VM ID if the memory is mapped */
 241             if (memory_region_is_mapped(&s->ivshmem)) {
 242                 ret = s->vm_id;
 243             } else {
 244                 ret = -1;
 245             }
 246             break;
 247
 248         default:
 249             IVSHMEM_DPRINTF("why are we reading " TARGET_FMT_plx "\n", addr);
 250             ret = 0;
 251     }
 252
 253     return ret;
 254 }
 255
 256 static const MemoryRegionOps ivshmem_mmio_ops = {
 257     .read = ivshmem_io_read,
 258     .write = ivshmem_io_write,
 259     .endianness = DEVICE_NATIVE_ENDIAN,
 260     .impl = {
 261         .min_access_size = 4,
 262         .max_access_size = 4,
 263     },
 264 };
 265
 266 static int ivshmem_can_receive(void * opaque)
 267 {
 268     return sizeof(int64_t);
 269 }
 270
 271 static void ivshmem_event(void *opaque, int event)
 272 {
 273     IVSHMEM_DPRINTF("ivshmem_event %d\n", event);
 274 }
 275
 276 static void ivshmem_vector_notify(void *opaque)
 277 {
 278     MSIVector *entry = opaque;
 279     PCIDevice *pdev = entry->pdev;
 280     IVShmemState *s = IVSHMEM(pdev);
 281     int vector = entry - s->msi_vectors;
 282     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 283
 284     if (!event_notifier_test_and_clear(n)) {
 285         return;
 286     }
 287
 288     IVSHMEM_DPRINTF("interrupt on vector %p %d\n", pdev, vector);
 289     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 290         msix_notify(pdev, vector);
 291     } else {
 292         ivshmem_IntrStatus_write(s, 1);
 293     }
 294 }
 295
 296 static int ivshmem_vector_unmask(PCIDevice *dev, unsigned vector,
 297                                  MSIMessage msg)
 298 {
 299     IVShmemState *s = IVSHMEM(dev);
 300     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 301     MSIVector *v = &s->msi_vectors[vector];
 302     int ret;
 303
 304     IVSHMEM_DPRINTF("vector unmask %p %d\n", dev, vector);
 305
 306     ret = kvm_irqchip_update_msi_route(kvm_state, v->virq, msg, dev);
 307     if (ret < 0) {
 308         return ret;
 309     }
 310
 311     return kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, n, NULL, v->virq);
 312 }
 313
 314 static void ivshmem_vector_mask(PCIDevice *dev, unsigned vector)
 315 {
 316     IVShmemState *s = IVSHMEM(dev);
 317     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 318     int ret;
 319
 320     IVSHMEM_DPRINTF("vector mask %p %d\n", dev, vector);
 321
 322     ret = kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state, n,
 323                                                 s->msi_vectors[vector].virq);
 324     if (ret != 0) {
 325         error_report("remove_irqfd_notifier_gsi failed");
 326     }
 327 }
 328
 329 static void ivshmem_vector_poll(PCIDevice *dev,
 330                                 unsigned int vector_start,
 331                                 unsigned int vector_end)
 332 {
 333     IVShmemState *s = IVSHMEM(dev);
 334     unsigned int vector;
 335
 336     IVSHMEM_DPRINTF("vector poll %p %d-%d\n", dev, vector_start, vector_end);
 337
 338     vector_end = MIN(vector_end, s->vectors);
 339
 340     for (vector = vector_start; vector < vector_end; vector++) {
 341         EventNotifier *notifier = &s->peers[s->vm_id].eventfds[vector];
 342
 343         if (!msix_is_masked(dev, vector)) {
 344             continue;
 345         }
 346
 347         if (event_notifier_test_and_clear(notifier)) {
 348             msix_set_pending(dev, vector);
 349         }
 350     }
 351 }
 352
 353 static void watch_vector_notifier(IVShmemState *s, EventNotifier *n,
 354                                  int vector)
 355 {
 356     int eventfd = event_notifier_get_fd(n);
 357
 358     /* if MSI is supported we need multiple interrupts */
 359     s->msi_vectors[vector].pdev = PCI_DEVICE(s);
 360
 361     qemu_set_fd_handler(eventfd, ivshmem_vector_notify,
 362                         NULL, &s->msi_vectors[vector]);
 363 }
 364
 365 static int check_shm_size(IVShmemState *s, int fd, Error **errp)
 366 {
 367     /* check that the guest isn't going to try and map more memory than the
 368      * the object has allocated return -1 to indicate error */
 369
 370     struct stat buf;
 371
 372     if (fstat(fd, &buf) < 0) {
 373         error_setg(errp, "exiting: fstat on fd %d failed: %s",
 374                    fd, strerror(errno));
 375         return -1;
 376     }
 377
 378     if (s->ivshmem_size > buf.st_size) {
 379         error_setg(errp, "Requested memory size greater"
 380                    " than shared object size (%" PRIu64 " > %" PRIu64")",
 381                    s->ivshmem_size, (uint64_t)buf.st_size);
 382         return -1;
 383     } else {
 384         return 0;
 385     }
 386 }
 387
 388 /* create the shared memory BAR when we are not using the server, so we can
 389  * create the BAR and map the memory immediately */
 390 static int create_shared_memory_BAR(IVShmemState *s, int fd, uint8_t attr,
 391                                     Error **errp)
 392 {
 393     void * ptr;
 394
 395     ptr = mmap(0, s->ivshmem_size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
 396     if (ptr == MAP_FAILED) {
 397         error_setg_errno(errp, errno, "Failed to mmap shared memory");
 398         return -1;
 399     }
 400
 401     memory_region_init_ram_ptr(&s->ivshmem, OBJECT(s), "ivshmem.bar2",
 402                                s->ivshmem_size, ptr);
 403     qemu_set_ram_fd(memory_region_get_ram_addr(&s->ivshmem), fd);
 404     vmstate_register_ram(&s->ivshmem, DEVICE(s));
 405     memory_region_add_subregion(&s->bar, 0, &s->ivshmem);
 406
 407     /* region for shared memory */
 408     pci_register_bar(PCI_DEVICE(s), 2, attr, &s->bar);
 409
 410     return 0;
 411 }
 412
 413 static void ivshmem_add_eventfd(IVShmemState *s, int posn, int i)
 414 {
 415     memory_region_add_eventfd(&s->ivshmem_mmio,
 416                               DOORBELL,
 417                               4,
 418                               true,
 419                               (posn << 16) | i,
 420                               &s->peers[posn].eventfds[i]);
 421 }
 422
 423 static void ivshmem_del_eventfd(IVShmemState *s, int posn, int i)
 424 {
 425     memory_region_del_eventfd(&s->ivshmem_mmio,
 426                               DOORBELL,
 427                               4,
 428                               true,
 429                               (posn << 16) | i,
 430                               &s->peers[posn].eventfds[i]);
 431 }
 432
 433 static void close_peer_eventfds(IVShmemState *s, int posn)
 434 {
 435     int i, n;
 436
 437     if (!ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
 438         return;
 439     }
 440     if (posn < 0 || posn >= s->nb_peers) {
 441         error_report("invalid peer %d", posn);
 442         return;
 443     }
 444
 445     n = s->peers[posn].nb_eventfds;
 446
 447     memory_region_transaction_begin();
 448     for (i = 0; i < n; i++) {
 449         ivshmem_del_eventfd(s, posn, i);
 450     }
 451     memory_region_transaction_commit();
 452     for (i = 0; i < n; i++) {
 453         event_notifier_cleanup(&s->peers[posn].eventfds[i]);
 454     }
 455
 456     g_free(s->peers[posn].eventfds);
 457     s->peers[posn].nb_eventfds = 0;
 458 }
 459
 460 /* this function increase the dynamic storage need to store data about other
 461  * peers */
 462 static int resize_peers(IVShmemState *s, int new_min_size)
 463 {
 464
 465     int j, old_size;
 466
 467     /* limit number of max peers */
 468     if (new_min_size <= 0 || new_min_size > IVSHMEM_MAX_PEERS) {
 469         return -1;
 470     }
 471     if (new_min_size <= s->nb_peers) {
 472         return 0;
 473     }
 474
 475     old_size = s->nb_peers;
 476     s->nb_peers = new_min_size;
 477
 478     IVSHMEM_DPRINTF("bumping storage to %d peers\n", s->nb_peers);
 479
 480     s->peers = g_realloc(s->peers, s->nb_peers * sizeof(Peer));
 481
 482     for (j = old_size; j < s->nb_peers; j++) {
 483         s->peers[j].eventfds = g_new0(EventNotifier, s->vectors);
 484         s->peers[j].nb_eventfds = 0;
 485     }
 486
 487     return 0;
 488 }
 489
 490 static bool fifo_update_and_get(IVShmemState *s, const uint8_t *buf, int size,
 491                                 void *data, size_t len)
 492 {
 493     const uint8_t *p;
 494     uint32_t num;
 495
 496     assert(len <= sizeof(int64_t)); /* limitation of the fifo */
 497     if (fifo8_is_empty(&s->incoming_fifo) && size == len) {
 498         memcpy(data, buf, size);
 499         return true;
 500     }
 501
 502     IVSHMEM_DPRINTF("short read of %d bytes\n", size);
 503
 504     num = MIN(size, sizeof(int64_t) - fifo8_num_used(&s->incoming_fifo));
 505     fifo8_push_all(&s->incoming_fifo, buf, num);
 506
 507     if (fifo8_num_used(&s->incoming_fifo) < len) {
 508         assert(num == 0);
 509         return false;
 510     }
 511
 512     size -= num;
 513     buf += num;
 514     p = fifo8_pop_buf(&s->incoming_fifo, len, &num);
 515     assert(num == len);
 516
 517     memcpy(data, p, len);
 518
 519     if (size > 0) {
 520         fifo8_push_all(&s->incoming_fifo, buf, size);
 521     }
 522
 523     return true;
 524 }
 525
 526 static bool fifo_update_and_get_i64(IVShmemState *s,
 527                                     const uint8_t *buf, int size, int64_t *i64)
 528 {
 529     if (fifo_update_and_get(s, buf, size, i64, sizeof(*i64))) {
 530         *i64 = GINT64_FROM_LE(*i64);
 531         return true;
 532     }
 533
 534     return false;
 535 }
 536
 537 static int ivshmem_add_kvm_msi_virq(IVShmemState *s, int vector)
 538 {
 539     PCIDevice *pdev = PCI_DEVICE(s);
 540     MSIMessage msg = msix_get_message(pdev, vector);
 541     int ret;
 542
 543     IVSHMEM_DPRINTF("ivshmem_add_kvm_msi_virq vector:%d\n", vector);
 544
 545     if (s->msi_vectors[vector].pdev != NULL) {
 546         return 0;
 547     }
 548
 549     ret = kvm_irqchip_add_msi_route(kvm_state, msg, pdev);
 550     if (ret < 0) {
 551         error_report("ivshmem: kvm_irqchip_add_msi_route failed");
 552         return -1;
 553     }
 554
 555     s->msi_vectors[vector].virq = ret;
 556     s->msi_vectors[vector].pdev = pdev;
 557
 558     return 0;
 559 }
 560
 561 static void setup_interrupt(IVShmemState *s, int vector)
 562 {
 563     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 564     bool with_irqfd = kvm_msi_via_irqfd_enabled() &&
 565         ivshmem_has_feature(s, IVSHMEM_MSI);
 566     PCIDevice *pdev = PCI_DEVICE(s);
 567
 568     IVSHMEM_DPRINTF("setting up interrupt for vector: %d\n", vector);
 569
 570     if (!with_irqfd) {
 571         IVSHMEM_DPRINTF("with eventfd");
 572         watch_vector_notifier(s, n, vector);
 573     } else if (msix_enabled(pdev)) {
 574         IVSHMEM_DPRINTF("with irqfd");
 575         if (ivshmem_add_kvm_msi_virq(s, vector) < 0) {
 576             return;
 577         }
 578
 579         if (!msix_is_masked(pdev, vector)) {
 580             kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, n, NULL,
 581                                                s->msi_vectors[vector].virq);
 582         }
 583     } else {
 584         /* it will be delayed until msix is enabled, in write_config */
 585         IVSHMEM_DPRINTF("with irqfd, delayed until msix enabled");
 586     }
 587 }
 588
 589 static void ivshmem_read(void *opaque, const uint8_t *buf, int size)
 590 {
 591     IVShmemState *s = opaque;
 592     int incoming_fd;
 593     int new_eventfd;
 594     int64_t incoming_posn;
 595     Error *err = NULL;
 596     Peer *peer;
 597
 598     if (!fifo_update_and_get_i64(s, buf, size, &incoming_posn)) {
 599         return;
 600     }
 601
 602     if (incoming_posn < -1) {
 603         IVSHMEM_DPRINTF("invalid incoming_posn %" PRId64 "\n", incoming_posn);
 604         return;
 605     }
 606
 607     /* pick off s->server_chr->msgfd and store it, posn should accompany msg */
 608     incoming_fd = qemu_chr_fe_get_msgfd(s->server_chr);
 609     IVSHMEM_DPRINTF("posn is %" PRId64 ", fd is %d\n",
 610                     incoming_posn, incoming_fd);
 611
 612     /* make sure we have enough space for this peer */
 613     if (incoming_posn >= s->nb_peers) {
 614         if (resize_peers(s, incoming_posn + 1) < 0) {
 615             error_report("failed to resize peers array");
 616             if (incoming_fd != -1) {
 617                 close(incoming_fd);
 618             }
 619             return;
 620         }
 621     }
 622
 623     peer = &s->peers[incoming_posn];
 624
 625     if (incoming_fd == -1) {
 626         /* if posn is positive and unseen before then this is our posn*/
 627         if (incoming_posn >= 0 && s->vm_id == -1) {
 628             /* receive our posn */
 629             s->vm_id = incoming_posn;
 630         } else {
 631             /* otherwise an fd == -1 means an existing peer has gone away */
 632             IVSHMEM_DPRINTF("posn %" PRId64 " has gone away\n", incoming_posn);
 633             close_peer_eventfds(s, incoming_posn);
 634         }
 635         return;
 636     }
 637
 638     /* if the position is -1, then it's shared memory region fd */
 639     if (incoming_posn == -1) {
 640         void * map_ptr;
 641
 642         if (memory_region_is_mapped(&s->ivshmem)) {
 643             error_report("shm already initialized");
 644             close(incoming_fd);
 645             return;
 646         }
 647
 648         if (check_shm_size(s, incoming_fd, &err) == -1) {
 649             error_report_err(err);
 650             close(incoming_fd);
 651             return;
 652         }
 653
 654         /* mmap the region and map into the BAR2 */
 655         map_ptr = mmap(0, s->ivshmem_size, PROT_READ|PROT_WRITE, MAP_SHARED,
 656                                                             incoming_fd, 0);
 657         if (map_ptr == MAP_FAILED) {
 658             error_report("Failed to mmap shared memory %s", strerror(errno));
 659             close(incoming_fd);
 660             return;
 661         }
 662         memory_region_init_ram_ptr(&s->ivshmem, OBJECT(s),
 663                                    "ivshmem.bar2", s->ivshmem_size, map_ptr);
 664         qemu_set_ram_fd(memory_region_get_ram_addr(&s->ivshmem),
 665                         incoming_fd);
 666         vmstate_register_ram(&s->ivshmem, DEVICE(s));
 667
 668         IVSHMEM_DPRINTF("guest h/w addr = %p, size = %" PRIu64 "\n",
 669                         map_ptr, s->ivshmem_size);
 670
 671         memory_region_add_subregion(&s->bar, 0, &s->ivshmem);
 672
 673         return;
 674     }
 675
 676     /* each peer has an associated array of eventfds, and we keep
 677      * track of how many eventfds received so far */
 678     /* get a new eventfd: */
 679     if (peer->nb_eventfds >= s->vectors) {
 680         error_report("Too many eventfd received, device has %d vectors",
 681                      s->vectors);
 682         close(incoming_fd);
 683         return;
 684     }
 685
 686     new_eventfd = peer->nb_eventfds++;
 687
 688     /* this is an eventfd for a particular peer VM */
 689     IVSHMEM_DPRINTF("eventfds[%" PRId64 "][%d] = %d\n", incoming_posn,
 690                     new_eventfd, incoming_fd);
 691     event_notifier_init_fd(&peer->eventfds[new_eventfd], incoming_fd);
 692     fcntl_setfl(incoming_fd, O_NONBLOCK); /* msix/irqfd poll non block */
 693
 694     if (incoming_posn == s->vm_id) {
 695         setup_interrupt(s, new_eventfd);
 696     }
 697
 698     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
 699         ivshmem_add_eventfd(s, incoming_posn, new_eventfd);
 700     }
 701 }
 702
 703 static void ivshmem_check_version(void *opaque, const uint8_t * buf, int size)
 704 {
 705     IVShmemState *s = opaque;
 706     int tmp;
 707     int64_t version;
 708
 709     if (!fifo_update_and_get_i64(s, buf, size, &version)) {
 710         return;
 711     }
 712
 713     tmp = qemu_chr_fe_get_msgfd(s->server_chr);
 714     if (tmp != -1 || version != IVSHMEM_PROTOCOL_VERSION) {
 715         fprintf(stderr, "incompatible version, you are connecting to a ivshmem-"
 716                 "server using a different protocol please check your setup\n");
 717         qemu_chr_delete(s->server_chr);
 718         s->server_chr = NULL;
 719         return;
 720     }
 721
 722     IVSHMEM_DPRINTF("version check ok, switch to real chardev handler\n");
 723     qemu_chr_add_handlers(s->server_chr, ivshmem_can_receive, ivshmem_read,
 724                           ivshmem_event, s);
 725 }
 726
 727 /* Select the MSI-X vectors used by device.
 728  * ivshmem maps events to vectors statically, so
 729  * we just enable all vectors on init and after reset. */
 730 static void ivshmem_use_msix(IVShmemState * s)
 731 {
 732     PCIDevice *d = PCI_DEVICE(s);
 733     int i;
 734
 735     IVSHMEM_DPRINTF("%s, msix present: %d\n", __func__, msix_present(d));
 736     if (!msix_present(d)) {
 737         return;
 738     }
 739
 740     for (i = 0; i < s->vectors; i++) {
 741         msix_vector_use(d, i);
 742     }
 743 }
 744
 745 static void ivshmem_reset(DeviceState *d)
 746 {
 747     IVShmemState *s = IVSHMEM(d);
 748
 749     s->intrstatus = 0;
 750     s->intrmask = 0;
 751     ivshmem_use_msix(s);
 752 }
 753
 754 static int ivshmem_setup_interrupts(IVShmemState *s)
 755 {
 756     /* allocate QEMU callback data for receiving interrupts */
 757     s->msi_vectors = g_malloc0(s->vectors * sizeof(MSIVector));
 758
 759     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 760         if (msix_init_exclusive_bar(PCI_DEVICE(s), s->vectors, 1)) {
 761             return -1;
 762         }
 763
 764         IVSHMEM_DPRINTF("msix initialized (%d vectors)\n", s->vectors);
 765         ivshmem_use_msix(s);
 766     }
 767
 768     return 0;
 769 }
 770
 771 static void ivshmem_enable_irqfd(IVShmemState *s)
 772 {
 773     PCIDevice *pdev = PCI_DEVICE(s);
 774     int i;
 775
 776     for (i = 0; i < s->peers[s->vm_id].nb_eventfds; i++) {
 777         ivshmem_add_kvm_msi_virq(s, i);
 778     }
 779
 780     if (msix_set_vector_notifiers(pdev,
 781                                   ivshmem_vector_unmask,
 782                                   ivshmem_vector_mask,
 783                                   ivshmem_vector_poll)) {
 784         error_report("ivshmem: msix_set_vector_notifiers failed");
 785     }
 786 }
 787
 788 static void ivshmem_remove_kvm_msi_virq(IVShmemState *s, int vector)
 789 {
 790     IVSHMEM_DPRINTF("ivshmem_remove_kvm_msi_virq vector:%d\n", vector);
 791
 792     if (s->msi_vectors[vector].pdev == NULL) {
 793         return;
 794     }
 795
 796     /* it was cleaned when masked in the frontend. */
 797     kvm_irqchip_release_virq(kvm_state, s->msi_vectors[vector].virq);
 798
 799     s->msi_vectors[vector].pdev = NULL;
 800 }
 801
 802 static void ivshmem_disable_irqfd(IVShmemState *s)
 803 {
 804     PCIDevice *pdev = PCI_DEVICE(s);
 805     int i;
 806
 807     for (i = 0; i < s->peers[s->vm_id].nb_eventfds; i++) {
 808         ivshmem_remove_kvm_msi_virq(s, i);
 809     }
 810
 811     msix_unset_vector_notifiers(pdev);
 812 }
 813
 814 static void ivshmem_write_config(PCIDevice *pdev, uint32_t address,
 815                                  uint32_t val, int len)
 816 {
 817     IVShmemState *s = IVSHMEM(pdev);
 818     int is_enabled, was_enabled = msix_enabled(pdev);
 819
 820     pci_default_write_config(pdev, address, val, len);
 821     is_enabled = msix_enabled(pdev);
 822
 823     if (kvm_msi_via_irqfd_enabled() && s->vm_id != -1) {
 824         if (!was_enabled && is_enabled) {
 825             ivshmem_enable_irqfd(s);
 826         } else if (was_enabled && !is_enabled) {
 827             ivshmem_disable_irqfd(s);
 828         }
 829     }
 830 }
 831
 832 static void pci_ivshmem_realize(PCIDevice *dev, Error **errp)
 833 {
 834     IVShmemState *s = IVSHMEM(dev);
 835     uint8_t *pci_conf;
 836     uint8_t attr = PCI_BASE_ADDRESS_SPACE_MEMORY |
 837         PCI_BASE_ADDRESS_MEM_PREFETCH;
 838
 839     if (!!s->server_chr + !!s->shmobj + !!s->hostmem != 1) {
 840         error_setg(errp,
 841                    "You must specify either 'shm', 'chardev' or 'x-memdev'");
 842         return;
 843     }
 844
 845     if (s->hostmem) {
 846         MemoryRegion *mr;
 847
 848         if (s->sizearg) {
 849             g_warning("size argument ignored with hostmem");
 850         }
 851
 852         mr = host_memory_backend_get_memory(s->hostmem, errp);
 853         s->ivshmem_size = memory_region_size(mr);
 854     } else if (s->sizearg == NULL) {
 855         s->ivshmem_size = 4 << 20; /* 4 MB default */
 856     } else {
 857         char *end;
 858         int64_t size = qemu_strtosz(s->sizearg, &end);
 859         if (size < 0 || *end != '\0' || !is_power_of_2(size)) {
 860             error_setg(errp, "Invalid size %s", s->sizearg);
 861             return;
 862         }
 863         s->ivshmem_size = size;
 864     }
 865
 866     fifo8_create(&s->incoming_fifo, sizeof(int64_t));
 867
 868     /* IRQFD requires MSI */
 869     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD) &&
 870         !ivshmem_has_feature(s, IVSHMEM_MSI)) {
 871         error_setg(errp, "ioeventfd/irqfd requires MSI");
 872         return;
 873     }
 874
 875     /* check that role is reasonable */
 876     if (s->role) {
 877         if (strncmp(s->role, "peer", 5) == 0) {
 878             s->role_val = IVSHMEM_PEER;
 879         } else if (strncmp(s->role, "master", 7) == 0) {
 880             s->role_val = IVSHMEM_MASTER;
 881         } else {
 882             error_setg(errp, "'role' must be 'peer' or 'master'");
 883             return;
 884         }
 885     } else {
 886         s->role_val = IVSHMEM_MASTER; /* default */
 887     }
 888
 889     if (s->role_val == IVSHMEM_PEER) {
 890         error_setg(&s->migration_blocker,
 891                    "Migration is disabled when using feature 'peer mode' in device 'ivshmem'");
 892         migrate_add_blocker(s->migration_blocker);
 893     }
 894
 895     pci_conf = dev->config;
 896     pci_conf[PCI_COMMAND] = PCI_COMMAND_IO | PCI_COMMAND_MEMORY;
 897
 898     pci_config_set_interrupt_pin(pci_conf, 1);
 899
 900     memory_region_init_io(&s->ivshmem_mmio, OBJECT(s), &ivshmem_mmio_ops, s,
 901                           "ivshmem-mmio", IVSHMEM_REG_BAR_SIZE);
 902
 903     /* region for registers*/
 904     pci_register_bar(dev, 0, PCI_BASE_ADDRESS_SPACE_MEMORY,
 905                      &s->ivshmem_mmio);
 906
 907     memory_region_init(&s->bar, OBJECT(s), "ivshmem-bar2-container", s->ivshmem_size);
 908     if (s->ivshmem_64bit) {
 909         attr |= PCI_BASE_ADDRESS_MEM_TYPE_64;
 910     }
 911
 912     if (s->hostmem != NULL) {
 913         MemoryRegion *mr;
 914
 915         IVSHMEM_DPRINTF("using hostmem\n");
 916
 917         mr = host_memory_backend_get_memory(MEMORY_BACKEND(s->hostmem), errp);
 918         vmstate_register_ram(mr, DEVICE(s));
 919         memory_region_add_subregion(&s->bar, 0, mr);
 920         pci_register_bar(PCI_DEVICE(s), 2, attr, &s->bar);
 921     } else if (s->server_chr != NULL) {
 922         /* FIXME do not rely on what chr drivers put into filename */
 923         if (strncmp(s->server_chr->filename, "unix:", 5)) {
 924             error_setg(errp, "chardev is not a unix client socket");
 925             return;
 926         }
 927
 928         /* if we get a UNIX socket as the parameter we will talk
 929          * to the ivshmem server to receive the memory region */
 930
 931         IVSHMEM_DPRINTF("using shared memory server (socket = %s)\n",
 932                         s->server_chr->filename);
 933
 934         if (ivshmem_setup_interrupts(s) < 0) {
 935             error_setg(errp, "failed to initialize interrupts");
 936             return;
 937         }
 938
 939         /* we allocate enough space for 16 peers and grow as needed */
 940         resize_peers(s, 16);
 941         s->vm_id = -1;
 942
 943         pci_register_bar(dev, 2, attr, &s->bar);
 944
 945         s->eventfd_chr = g_malloc0(s->vectors * sizeof(CharDriverState *));
 946
 947         qemu_chr_add_handlers(s->server_chr, ivshmem_can_receive,
 948                               ivshmem_check_version, ivshmem_event, s);
 949     } else {
 950         /* just map the file immediately, we're not using a server */
 951         int fd;
 952
 953         IVSHMEM_DPRINTF("using shm_open (shm object = %s)\n", s->shmobj);
 954
 955         /* try opening with O_EXCL and if it succeeds zero the memory
 956          * by truncating to 0 */
 957         if ((fd = shm_open(s->shmobj, O_CREAT|O_RDWR|O_EXCL,
 958                         S_IRWXU|S_IRWXG|S_IRWXO)) > 0) {
 959            /* truncate file to length PCI device's memory */
 960             if (ftruncate(fd, s->ivshmem_size) != 0) {
 961                 error_report("could not truncate shared file");
 962             }
 963
 964         } else if ((fd = shm_open(s->shmobj, O_CREAT|O_RDWR,
 965                         S_IRWXU|S_IRWXG|S_IRWXO)) < 0) {
 966             error_setg(errp, "could not open shared file");
 967             return;
 968         }
 969
 970         if (check_shm_size(s, fd, errp) == -1) {
 971             return;
 972         }
 973
 974         create_shared_memory_BAR(s, fd, attr, errp);
 975     }
 976 }
 977
 978 static void pci_ivshmem_exit(PCIDevice *dev)
 979 {
 980     IVShmemState *s = IVSHMEM(dev);
 981     int i;
 982
 983     fifo8_destroy(&s->incoming_fifo);
 984
 985     if (s->migration_blocker) {
 986         migrate_del_blocker(s->migration_blocker);
 987         error_free(s->migration_blocker);
 988     }
 989
 990     if (memory_region_is_mapped(&s->ivshmem)) {
 991         if (!s->hostmem) {
 992             void *addr = memory_region_get_ram_ptr(&s->ivshmem);
 993             int fd;
 994
 995             if (munmap(addr, s->ivshmem_size) == -1) {
 996                 error_report("Failed to munmap shared memory %s",
 997                              strerror(errno));
 998             }
 999
1000             fd = qemu_get_ram_fd(memory_region_get_ram_addr(&s->ivshmem));
1001             if (fd != -1) {
1002                 close(fd);
1003             }
1004         }
1005
1006         vmstate_unregister_ram(&s->ivshmem, DEVICE(dev));
1007         memory_region_del_subregion(&s->bar, &s->ivshmem);
1008     }
1009
1010     if (s->eventfd_chr) {
1011         for (i = 0; i < s->vectors; i++) {
1012             if (s->eventfd_chr[i]) {
1013                 qemu_chr_free(s->eventfd_chr[i]);
1014             }
1015         }
1016         g_free(s->eventfd_chr);
1017     }
1018
1019     if (s->peers) {
1020         for (i = 0; i < s->nb_peers; i++) {
1021             close_peer_eventfds(s, i);
1022         }
1023         g_free(s->peers);
1024     }
1025
1026     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
1027         msix_uninit_exclusive_bar(dev);
1028     }
1029
1030     g_free(s->msi_vectors);
1031 }
1032
1033 static bool test_msix(void *opaque, int version_id)
1034 {
1035     IVShmemState *s = opaque;
1036
1037     return ivshmem_has_feature(s, IVSHMEM_MSI);
1038 }
1039
1040 static bool test_no_msix(void *opaque, int version_id)
1041 {
1042     return !test_msix(opaque, version_id);
1043 }
1044
1045 static int ivshmem_pre_load(void *opaque)
1046 {
1047     IVShmemState *s = opaque;
1048
1049     if (s->role_val == IVSHMEM_PEER) {
1050         error_report("'peer' devices are not migratable");
1051         return -EINVAL;
1052     }
1053
1054     return 0;
1055 }
1056
1057 static int ivshmem_post_load(void *opaque, int version_id)
1058 {
1059     IVShmemState *s = opaque;
1060
1061     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
1062         ivshmem_use_msix(s);
1063     }
1064
1065     return 0;
1066 }
1067
1068 static int ivshmem_load_old(QEMUFile *f, void *opaque, int version_id)
1069 {
1070     IVShmemState *s = opaque;
1071     PCIDevice *pdev = PCI_DEVICE(s);
1072     int ret;
1073
1074     IVSHMEM_DPRINTF("ivshmem_load_old\n");
1075
1076     if (version_id != 0) {
1077         return -EINVAL;
1078     }
1079
1080     if (s->role_val == IVSHMEM_PEER) {
1081         error_report("'peer' devices are not migratable");
1082         return -EINVAL;
1083     }
1084
1085     ret = pci_device_load(pdev, f);
1086     if (ret) {
1087         return ret;
1088     }
1089
1090     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
1091         msix_load(pdev, f);
1092         ivshmem_use_msix(s);
1093     } else {
1094         s->intrstatus = qemu_get_be32(f);
1095         s->intrmask = qemu_get_be32(f);
1096     }
1097
1098     return 0;
1099 }
1100
1101 static const VMStateDescription ivshmem_vmsd = {
1102     .name = "ivshmem",
1103     .version_id = 1,
1104     .minimum_version_id = 1,
1105     .pre_load = ivshmem_pre_load,
1106     .post_load = ivshmem_post_load,
1107     .fields = (VMStateField[]) {
1108         VMSTATE_PCI_DEVICE(parent_obj, IVShmemState),
1109
1110         VMSTATE_MSIX_TEST(parent_obj, IVShmemState, test_msix),
1111         VMSTATE_UINT32_TEST(intrstatus, IVShmemState, test_no_msix),
1112         VMSTATE_UINT32_TEST(intrmask, IVShmemState, test_no_msix),
1113
1114         VMSTATE_END_OF_LIST()
1115     },
1116     .load_state_old = ivshmem_load_old,
1117     .minimum_version_id_old = 0
1118 };
1119
1120 static Property ivshmem_properties[] = {
1121     DEFINE_PROP_CHR("chardev", IVShmemState, server_chr),
1122     DEFINE_PROP_STRING("size", IVShmemState, sizearg),
1123     DEFINE_PROP_UINT32("vectors", IVShmemState, vectors, 1),
1124     DEFINE_PROP_BIT("ioeventfd", IVShmemState, features, IVSHMEM_IOEVENTFD, false),
1125     DEFINE_PROP_BIT("msi", IVShmemState, features, IVSHMEM_MSI, true),
1126     DEFINE_PROP_STRING("shm", IVShmemState, shmobj),
1127     DEFINE_PROP_STRING("role", IVShmemState, role),
1128     DEFINE_PROP_UINT32("use64", IVShmemState, ivshmem_64bit, 1),
1129     DEFINE_PROP_END_OF_LIST(),
1130 };
1131
1132 static void ivshmem_class_init(ObjectClass *klass, void *data)
1133 {
1134     DeviceClass *dc = DEVICE_CLASS(klass);
1135     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
1136
1137     k->realize = pci_ivshmem_realize;
1138     k->exit = pci_ivshmem_exit;
1139     k->config_write = ivshmem_write_config;
1140     k->vendor_id = PCI_VENDOR_ID_IVSHMEM;
1141     k->device_id = PCI_DEVICE_ID_IVSHMEM;
1142     k->class_id = PCI_CLASS_MEMORY_RAM;
1143     dc->reset = ivshmem_reset;
1144     dc->props = ivshmem_properties;
1145     dc->vmsd = &ivshmem_vmsd;
1146     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
1147     dc->desc = "Inter-VM shared memory";
1148 }
1149
1150 static void ivshmem_check_memdev_is_busy(Object *obj, const char *name,
1151                                          Object *val, Error **errp)
1152 {
1153     MemoryRegion *mr;
1154
1155     mr = host_memory_backend_get_memory(MEMORY_BACKEND(val), errp);
1156     if (memory_region_is_mapped(mr)) {
1157         char *path = object_get_canonical_path_component(val);
1158         error_setg(errp, "can't use already busy memdev: %s", path);
1159         g_free(path);
1160     } else {
1161         qdev_prop_allow_set_link_before_realize(obj, name, val, errp);
1162     }
1163 }
1164
1165 static void ivshmem_init(Object *obj)
1166 {
1167     IVShmemState *s = IVSHMEM(obj);
1168
1169     object_property_add_link(obj, "x-memdev", TYPE_MEMORY_BACKEND,
1170                              (Object **)&s->hostmem,
1171                              ivshmem_check_memdev_is_busy,
1172                              OBJ_PROP_LINK_UNREF_ON_RELEASE,
1173                              &error_abort);
1174 }
1175
1176 static const TypeInfo ivshmem_info = {
1177     .name          = TYPE_IVSHMEM,
1178     .parent        = TYPE_PCI_DEVICE,
1179     .instance_size = sizeof(IVShmemState),
1180     .instance_init = ivshmem_init,
1181     .class_init    = ivshmem_class_init,
1182 };
1183
1184 static void ivshmem_register_types(void)
1185 {
1186     type_register_static(&ivshmem_info);
1187 }
1188
1189 type_init(ivshmem_register_types)