hw/misc/ivshmem.c

   1 /*
   2  * Inter-VM Shared Memory PCI device.
   3  *
   4  * Author:
   5  *      Cam Macdonell <[email protected]>
   6  *
   7  * Based On: cirrus_vga.c
   8  *          Copyright (c) 2004 Fabrice Bellard
   9  *          Copyright (c) 2004 Makoto Suzuki (suzu)
  10  *
  11  *      and rtl8139.c
  12  *          Copyright (c) 2006 Igor Kovalenko
  13  *
  14  * This code is licensed under the GNU GPL v2.
  15  *
  16  * Contributions after 2012-01-13 are licensed under the terms of the
  17  * GNU GPL, version 2 or (at your option) any later version.
  18  */
  19 #include "qemu/osdep.h"
  20 #include "hw/hw.h"
  21 #include "hw/i386/pc.h"
  22 #include "hw/pci/pci.h"
  23 #include "hw/pci/msi.h"
  24 #include "hw/pci/msix.h"
  25 #include "sysemu/kvm.h"
  26 #include "migration/migration.h"
  27 #include "qemu/error-report.h"
  28 #include "qemu/event_notifier.h"
  29 #include "qom/object_interfaces.h"
  30 #include "sysemu/char.h"
  31 #include "sysemu/hostmem.h"
  32 #include "sysemu/qtest.h"
  33 #include "qapi/visitor.h"
  34 #include "exec/ram_addr.h"
  35
  36 #include "hw/misc/ivshmem.h"
  37
  38 #include <sys/mman.h>
  39
  40 #define PCI_VENDOR_ID_IVSHMEM   PCI_VENDOR_ID_REDHAT_QUMRANET
  41 #define PCI_DEVICE_ID_IVSHMEM   0x1110
  42
  43 #define IVSHMEM_MAX_PEERS UINT16_MAX
  44 #define IVSHMEM_IOEVENTFD   0
  45 #define IVSHMEM_MSI     1
  46
  47 #define IVSHMEM_REG_BAR_SIZE 0x100
  48
  49 #define IVSHMEM_DEBUG 0
  50 #define IVSHMEM_DPRINTF(fmt, ...)                       \
  51     do {                                                \
  52         if (IVSHMEM_DEBUG) {                            \
  53             printf("IVSHMEM: " fmt, ## __VA_ARGS__);    \
  54         }                                               \
  55     } while (0)
  56
  57 #define TYPE_IVSHMEM_COMMON "ivshmem-common"
  58 #define IVSHMEM_COMMON(obj) \
  59     OBJECT_CHECK(IVShmemState, (obj), TYPE_IVSHMEM_COMMON)
  60
  61 #define TYPE_IVSHMEM_PLAIN "ivshmem-plain"
  62 #define IVSHMEM_PLAIN(obj) \
  63     OBJECT_CHECK(IVShmemState, (obj), TYPE_IVSHMEM_PLAIN)
  64
  65 #define TYPE_IVSHMEM_DOORBELL "ivshmem-doorbell"
  66 #define IVSHMEM_DOORBELL(obj) \
  67     OBJECT_CHECK(IVShmemState, (obj), TYPE_IVSHMEM_DOORBELL)
  68
  69 #define TYPE_IVSHMEM "ivshmem"
  70 #define IVSHMEM(obj) \
  71     OBJECT_CHECK(IVShmemState, (obj), TYPE_IVSHMEM)
  72
  73 typedef struct Peer {
  74     int nb_eventfds;
  75     EventNotifier *eventfds;
  76 } Peer;
  77
  78 typedef struct MSIVector {
  79     PCIDevice *pdev;
  80     int virq;
  81 } MSIVector;
  82
  83 typedef struct IVShmemState {
  84     /*< private >*/
  85     PCIDevice parent_obj;
  86     /*< public >*/
  87
  88     uint32_t features;
  89
  90     /* exactly one of these two may be set */
  91     HostMemoryBackend *hostmem; /* with interrupts */
  92     CharDriverState *server_chr; /* without interrupts */
  93
  94     /* registers */
  95     uint32_t intrmask;
  96     uint32_t intrstatus;
  97     int vm_id;
  98
  99     /* BARs */
 100     MemoryRegion ivshmem_mmio;  /* BAR 0 (registers) */
 101     MemoryRegion *ivshmem_bar2; /* BAR 2 (shared memory) */
 102     MemoryRegion server_bar2;   /* used with server_chr */
 103
 104     /* interrupt support */
 105     Peer *peers;
 106     int nb_peers;               /* space in @peers[] */
 107     uint32_t vectors;
 108     MSIVector *msi_vectors;
 109     uint64_t msg_buf;           /* buffer for receiving server messages */
 110     int msg_buffered_bytes;     /* #bytes in @msg_buf */
 111
 112     /* migration stuff */
 113     OnOffAuto master;
 114     Error *migration_blocker;
 115
 116     /* legacy cruft */
 117     char *role;
 118     char *shmobj;
 119     char *sizearg;
 120     size_t legacy_size;
 121     uint32_t not_legacy_32bit;
 122 } IVShmemState;
 123
 124 /* registers for the Inter-VM shared memory device */
 125 enum ivshmem_registers {
 126     INTRMASK = 0,
 127     INTRSTATUS = 4,
 128     IVPOSITION = 8,
 129     DOORBELL = 12,
 130 };
 131
 132 static inline uint32_t ivshmem_has_feature(IVShmemState *ivs,
 133                                                     unsigned int feature) {
 134     return (ivs->features & (1 << feature));
 135 }
 136
 137 static inline bool ivshmem_is_master(IVShmemState *s)
 138 {
 139     assert(s->master != ON_OFF_AUTO_AUTO);
 140     return s->master == ON_OFF_AUTO_ON;
 141 }
 142
 143 static void ivshmem_update_irq(IVShmemState *s)
 144 {
 145     PCIDevice *d = PCI_DEVICE(s);
 146     uint32_t isr = s->intrstatus & s->intrmask;
 147
 148     /*
 149      * Do nothing unless the device actually uses INTx.  Here's how
 150      * the device variants signal interrupts, what they put in PCI
 151      * config space:
 152      * Device variant    Interrupt  Interrupt Pin  MSI-X cap.
 153      * ivshmem-plain         none            0         no
 154      * ivshmem-doorbell     MSI-X            1        yes(1)
 155      * ivshmem,msi=off       INTx            1         no
 156      * ivshmem,msi=on       MSI-X            1(2)     yes(1)
 157      * (1) if guest enabled MSI-X
 158      * (2) the device lies
 159      * Leads to the condition for doing nothing:
 160      */
 161     if (ivshmem_has_feature(s, IVSHMEM_MSI)
 162         || !d->config[PCI_INTERRUPT_PIN]) {
 163         return;
 164     }
 165
 166     /* don't print ISR resets */
 167     if (isr) {
 168         IVSHMEM_DPRINTF("Set IRQ to %d (%04x %04x)\n",
 169                         isr ? 1 : 0, s->intrstatus, s->intrmask);
 170     }
 171
 172     pci_set_irq(d, isr != 0);
 173 }
 174
 175 static void ivshmem_IntrMask_write(IVShmemState *s, uint32_t val)
 176 {
 177     IVSHMEM_DPRINTF("IntrMask write(w) val = 0x%04x\n", val);
 178
 179     s->intrmask = val;
 180     ivshmem_update_irq(s);
 181 }
 182
 183 static uint32_t ivshmem_IntrMask_read(IVShmemState *s)
 184 {
 185     uint32_t ret = s->intrmask;
 186
 187     IVSHMEM_DPRINTF("intrmask read(w) val = 0x%04x\n", ret);
 188     return ret;
 189 }
 190
 191 static void ivshmem_IntrStatus_write(IVShmemState *s, uint32_t val)
 192 {
 193     IVSHMEM_DPRINTF("IntrStatus write(w) val = 0x%04x\n", val);
 194
 195     s->intrstatus = val;
 196     ivshmem_update_irq(s);
 197 }
 198
 199 static uint32_t ivshmem_IntrStatus_read(IVShmemState *s)
 200 {
 201     uint32_t ret = s->intrstatus;
 202
 203     /* reading ISR clears all interrupts */
 204     s->intrstatus = 0;
 205     ivshmem_update_irq(s);
 206     return ret;
 207 }
 208
 209 static void ivshmem_io_write(void *opaque, hwaddr addr,
 210                              uint64_t val, unsigned size)
 211 {
 212     IVShmemState *s = opaque;
 213
 214     uint16_t dest = val >> 16;
 215     uint16_t vector = val & 0xff;
 216
 217     addr &= 0xfc;
 218
 219     IVSHMEM_DPRINTF("writing to addr " TARGET_FMT_plx "\n", addr);
 220     switch (addr)
 221     {
 222         case INTRMASK:
 223             ivshmem_IntrMask_write(s, val);
 224             break;
 225
 226         case INTRSTATUS:
 227             ivshmem_IntrStatus_write(s, val);
 228             break;
 229
 230         case DOORBELL:
 231             /* check that dest VM ID is reasonable */
 232             if (dest >= s->nb_peers) {
 233                 IVSHMEM_DPRINTF("Invalid destination VM ID (%d)\n", dest);
 234                 break;
 235             }
 236
 237             /* check doorbell range */
 238             if (vector < s->peers[dest].nb_eventfds) {
 239                 IVSHMEM_DPRINTF("Notifying VM %d on vector %d\n", dest, vector);
 240                 event_notifier_set(&s->peers[dest].eventfds[vector]);
 241             } else {
 242                 IVSHMEM_DPRINTF("Invalid destination vector %d on VM %d\n",
 243                                 vector, dest);
 244             }
 245             break;
 246         default:
 247             IVSHMEM_DPRINTF("Unhandled write " TARGET_FMT_plx "\n", addr);
 248     }
 249 }
 250
 251 static uint64_t ivshmem_io_read(void *opaque, hwaddr addr,
 252                                 unsigned size)
 253 {
 254
 255     IVShmemState *s = opaque;
 256     uint32_t ret;
 257
 258     switch (addr)
 259     {
 260         case INTRMASK:
 261             ret = ivshmem_IntrMask_read(s);
 262             break;
 263
 264         case INTRSTATUS:
 265             ret = ivshmem_IntrStatus_read(s);
 266             break;
 267
 268         case IVPOSITION:
 269             ret = s->vm_id;
 270             break;
 271
 272         default:
 273             IVSHMEM_DPRINTF("why are we reading " TARGET_FMT_plx "\n", addr);
 274             ret = 0;
 275     }
 276
 277     return ret;
 278 }
 279
 280 static const MemoryRegionOps ivshmem_mmio_ops = {
 281     .read = ivshmem_io_read,
 282     .write = ivshmem_io_write,
 283     .endianness = DEVICE_NATIVE_ENDIAN,
 284     .impl = {
 285         .min_access_size = 4,
 286         .max_access_size = 4,
 287     },
 288 };
 289
 290 static void ivshmem_vector_notify(void *opaque)
 291 {
 292     MSIVector *entry = opaque;
 293     PCIDevice *pdev = entry->pdev;
 294     IVShmemState *s = IVSHMEM_COMMON(pdev);
 295     int vector = entry - s->msi_vectors;
 296     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 297
 298     if (!event_notifier_test_and_clear(n)) {
 299         return;
 300     }
 301
 302     IVSHMEM_DPRINTF("interrupt on vector %p %d\n", pdev, vector);
 303     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 304         if (msix_enabled(pdev)) {
 305             msix_notify(pdev, vector);
 306         }
 307     } else {
 308         ivshmem_IntrStatus_write(s, 1);
 309     }
 310 }
 311
 312 static int ivshmem_vector_unmask(PCIDevice *dev, unsigned vector,
 313                                  MSIMessage msg)
 314 {
 315     IVShmemState *s = IVSHMEM_COMMON(dev);
 316     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 317     MSIVector *v = &s->msi_vectors[vector];
 318     int ret;
 319
 320     IVSHMEM_DPRINTF("vector unmask %p %d\n", dev, vector);
 321
 322     ret = kvm_irqchip_update_msi_route(kvm_state, v->virq, msg, dev);
 323     if (ret < 0) {
 324         return ret;
 325     }
 326
 327     return kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, n, NULL, v->virq);
 328 }
 329
 330 static void ivshmem_vector_mask(PCIDevice *dev, unsigned vector)
 331 {
 332     IVShmemState *s = IVSHMEM_COMMON(dev);
 333     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 334     int ret;
 335
 336     IVSHMEM_DPRINTF("vector mask %p %d\n", dev, vector);
 337
 338     ret = kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state, n,
 339                                                 s->msi_vectors[vector].virq);
 340     if (ret != 0) {
 341         error_report("remove_irqfd_notifier_gsi failed");
 342     }
 343 }
 344
 345 static void ivshmem_vector_poll(PCIDevice *dev,
 346                                 unsigned int vector_start,
 347                                 unsigned int vector_end)
 348 {
 349     IVShmemState *s = IVSHMEM_COMMON(dev);
 350     unsigned int vector;
 351
 352     IVSHMEM_DPRINTF("vector poll %p %d-%d\n", dev, vector_start, vector_end);
 353
 354     vector_end = MIN(vector_end, s->vectors);
 355
 356     for (vector = vector_start; vector < vector_end; vector++) {
 357         EventNotifier *notifier = &s->peers[s->vm_id].eventfds[vector];
 358
 359         if (!msix_is_masked(dev, vector)) {
 360             continue;
 361         }
 362
 363         if (event_notifier_test_and_clear(notifier)) {
 364             msix_set_pending(dev, vector);
 365         }
 366     }
 367 }
 368
 369 static void watch_vector_notifier(IVShmemState *s, EventNotifier *n,
 370                                  int vector)
 371 {
 372     int eventfd = event_notifier_get_fd(n);
 373
 374     assert(!s->msi_vectors[vector].pdev);
 375     s->msi_vectors[vector].pdev = PCI_DEVICE(s);
 376
 377     qemu_set_fd_handler(eventfd, ivshmem_vector_notify,
 378                         NULL, &s->msi_vectors[vector]);
 379 }
 380
 381 static void ivshmem_add_eventfd(IVShmemState *s, int posn, int i)
 382 {
 383     memory_region_add_eventfd(&s->ivshmem_mmio,
 384                               DOORBELL,
 385                               4,
 386                               true,
 387                               (posn << 16) | i,
 388                               &s->peers[posn].eventfds[i]);
 389 }
 390
 391 static void ivshmem_del_eventfd(IVShmemState *s, int posn, int i)
 392 {
 393     memory_region_del_eventfd(&s->ivshmem_mmio,
 394                               DOORBELL,
 395                               4,
 396                               true,
 397                               (posn << 16) | i,
 398                               &s->peers[posn].eventfds[i]);
 399 }
 400
 401 static void close_peer_eventfds(IVShmemState *s, int posn)
 402 {
 403     int i, n;
 404
 405     assert(posn >= 0 && posn < s->nb_peers);
 406     n = s->peers[posn].nb_eventfds;
 407
 408     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
 409         memory_region_transaction_begin();
 410         for (i = 0; i < n; i++) {
 411             ivshmem_del_eventfd(s, posn, i);
 412         }
 413         memory_region_transaction_commit();
 414     }
 415
 416     for (i = 0; i < n; i++) {
 417         event_notifier_cleanup(&s->peers[posn].eventfds[i]);
 418     }
 419
 420     g_free(s->peers[posn].eventfds);
 421     s->peers[posn].nb_eventfds = 0;
 422 }
 423
 424 static void resize_peers(IVShmemState *s, int nb_peers)
 425 {
 426     int old_nb_peers = s->nb_peers;
 427     int i;
 428
 429     assert(nb_peers > old_nb_peers);
 430     IVSHMEM_DPRINTF("bumping storage to %d peers\n", nb_peers);
 431
 432     s->peers = g_realloc(s->peers, nb_peers * sizeof(Peer));
 433     s->nb_peers = nb_peers;
 434
 435     for (i = old_nb_peers; i < nb_peers; i++) {
 436         s->peers[i].eventfds = g_new0(EventNotifier, s->vectors);
 437         s->peers[i].nb_eventfds = 0;
 438     }
 439 }
 440
 441 static void ivshmem_add_kvm_msi_virq(IVShmemState *s, int vector,
 442                                      Error **errp)
 443 {
 444     PCIDevice *pdev = PCI_DEVICE(s);
 445     MSIMessage msg = msix_get_message(pdev, vector);
 446     int ret;
 447
 448     IVSHMEM_DPRINTF("ivshmem_add_kvm_msi_virq vector:%d\n", vector);
 449     assert(!s->msi_vectors[vector].pdev);
 450
 451     ret = kvm_irqchip_add_msi_route(kvm_state, msg, pdev);
 452     if (ret < 0) {
 453         error_setg(errp, "kvm_irqchip_add_msi_route failed");
 454         return;
 455     }
 456
 457     s->msi_vectors[vector].virq = ret;
 458     s->msi_vectors[vector].pdev = pdev;
 459 }
 460
 461 static void setup_interrupt(IVShmemState *s, int vector, Error **errp)
 462 {
 463     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
 464     bool with_irqfd = kvm_msi_via_irqfd_enabled() &&
 465         ivshmem_has_feature(s, IVSHMEM_MSI);
 466     PCIDevice *pdev = PCI_DEVICE(s);
 467     Error *err = NULL;
 468
 469     IVSHMEM_DPRINTF("setting up interrupt for vector: %d\n", vector);
 470
 471     if (!with_irqfd) {
 472         IVSHMEM_DPRINTF("with eventfd\n");
 473         watch_vector_notifier(s, n, vector);
 474     } else if (msix_enabled(pdev)) {
 475         IVSHMEM_DPRINTF("with irqfd\n");
 476         ivshmem_add_kvm_msi_virq(s, vector, &err);
 477         if (err) {
 478             error_propagate(errp, err);
 479             return;
 480         }
 481
 482         if (!msix_is_masked(pdev, vector)) {
 483             kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, n, NULL,
 484                                                s->msi_vectors[vector].virq);
 485             /* TODO handle error */
 486         }
 487     } else {
 488         /* it will be delayed until msix is enabled, in write_config */
 489         IVSHMEM_DPRINTF("with irqfd, delayed until msix enabled\n");
 490     }
 491 }
 492
 493 static void process_msg_shmem(IVShmemState *s, int fd, Error **errp)
 494 {
 495     struct stat buf;
 496     size_t size;
 497     void *ptr;
 498
 499     if (s->ivshmem_bar2) {
 500         error_setg(errp, "server sent unexpected shared memory message");
 501         close(fd);
 502         return;
 503     }
 504
 505     if (fstat(fd, &buf) < 0) {
 506         error_setg_errno(errp, errno,
 507             "can't determine size of shared memory sent by server");
 508         close(fd);
 509         return;
 510     }
 511
 512     size = buf.st_size;
 513
 514     /* Legacy cruft */
 515     if (s->legacy_size != SIZE_MAX) {
 516         if (size < s->legacy_size) {
 517             error_setg(errp, "server sent only %zd bytes of shared memory",
 518                        (size_t)buf.st_size);
 519             close(fd);
 520             return;
 521         }
 522         size = s->legacy_size;
 523     }
 524
 525     /* mmap the region and map into the BAR2 */
 526     ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 527     if (ptr == MAP_FAILED) {
 528         error_setg_errno(errp, errno, "Failed to mmap shared memory");
 529         close(fd);
 530         return;
 531     }
 532     memory_region_init_ram_ptr(&s->server_bar2, OBJECT(s),
 533                                "ivshmem.bar2", size, ptr);
 534     qemu_set_ram_fd(memory_region_get_ram_addr(&s->server_bar2), fd);
 535     s->ivshmem_bar2 = &s->server_bar2;
 536 }
 537
 538 static void process_msg_disconnect(IVShmemState *s, uint16_t posn,
 539                                    Error **errp)
 540 {
 541     IVSHMEM_DPRINTF("posn %d has gone away\n", posn);
 542     if (posn >= s->nb_peers || posn == s->vm_id) {
 543         error_setg(errp, "invalid peer %d", posn);
 544         return;
 545     }
 546     close_peer_eventfds(s, posn);
 547 }
 548
 549 static void process_msg_connect(IVShmemState *s, uint16_t posn, int fd,
 550                                 Error **errp)
 551 {
 552     Peer *peer = &s->peers[posn];
 553     int vector;
 554
 555     /*
 556      * The N-th connect message for this peer comes with the file
 557      * descriptor for vector N-1.  Count messages to find the vector.
 558      */
 559     if (peer->nb_eventfds >= s->vectors) {
 560         error_setg(errp, "Too many eventfd received, device has %d vectors",
 561                    s->vectors);
 562         close(fd);
 563         return;
 564     }
 565     vector = peer->nb_eventfds++;
 566
 567     IVSHMEM_DPRINTF("eventfds[%d][%d] = %d\n", posn, vector, fd);
 568     event_notifier_init_fd(&peer->eventfds[vector], fd);
 569     fcntl_setfl(fd, O_NONBLOCK); /* msix/irqfd poll non block */
 570
 571     if (posn == s->vm_id) {
 572         setup_interrupt(s, vector, errp);
 573         /* TODO do we need to handle the error? */
 574     }
 575
 576     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
 577         ivshmem_add_eventfd(s, posn, vector);
 578     }
 579 }
 580
 581 static void process_msg(IVShmemState *s, int64_t msg, int fd, Error **errp)
 582 {
 583     IVSHMEM_DPRINTF("posn is %" PRId64 ", fd is %d\n", msg, fd);
 584
 585     if (msg < -1 || msg > IVSHMEM_MAX_PEERS) {
 586         error_setg(errp, "server sent invalid message %" PRId64, msg);
 587         close(fd);
 588         return;
 589     }
 590
 591     if (msg == -1) {
 592         process_msg_shmem(s, fd, errp);
 593         return;
 594     }
 595
 596     if (msg >= s->nb_peers) {
 597         resize_peers(s, msg + 1);
 598     }
 599
 600     if (fd >= 0) {
 601         process_msg_connect(s, msg, fd, errp);
 602     } else {
 603         process_msg_disconnect(s, msg, errp);
 604     }
 605 }
 606
 607 static int ivshmem_can_receive(void *opaque)
 608 {
 609     IVShmemState *s = opaque;
 610
 611     assert(s->msg_buffered_bytes < sizeof(s->msg_buf));
 612     return sizeof(s->msg_buf) - s->msg_buffered_bytes;
 613 }
 614
 615 static void ivshmem_read(void *opaque, const uint8_t *buf, int size)
 616 {
 617     IVShmemState *s = opaque;
 618     Error *err = NULL;
 619     int fd;
 620     int64_t msg;
 621
 622     assert(size >= 0 && s->msg_buffered_bytes + size <= sizeof(s->msg_buf));
 623     memcpy((unsigned char *)&s->msg_buf + s->msg_buffered_bytes, buf, size);
 624     s->msg_buffered_bytes += size;
 625     if (s->msg_buffered_bytes < sizeof(s->msg_buf)) {
 626         return;
 627     }
 628     msg = le64_to_cpu(s->msg_buf);
 629     s->msg_buffered_bytes = 0;
 630
 631     fd = qemu_chr_fe_get_msgfd(s->server_chr);
 632     IVSHMEM_DPRINTF("posn is %" PRId64 ", fd is %d\n", msg, fd);
 633
 634     process_msg(s, msg, fd, &err);
 635     if (err) {
 636         error_report_err(err);
 637     }
 638 }
 639
 640 static int64_t ivshmem_recv_msg(IVShmemState *s, int *pfd, Error **errp)
 641 {
 642     int64_t msg;
 643     int n, ret;
 644
 645     n = 0;
 646     do {
 647         ret = qemu_chr_fe_read_all(s->server_chr, (uint8_t *)&msg + n,
 648                                  sizeof(msg) - n);
 649         if (ret < 0 && ret != -EINTR) {
 650             error_setg_errno(errp, -ret, "read from server failed");
 651             return INT64_MIN;
 652         }
 653         n += ret;
 654     } while (n < sizeof(msg));
 655
 656     *pfd = qemu_chr_fe_get_msgfd(s->server_chr);
 657     return msg;
 658 }
 659
 660 static void ivshmem_recv_setup(IVShmemState *s, Error **errp)
 661 {
 662     Error *err = NULL;
 663     int64_t msg;
 664     int fd;
 665
 666     msg = ivshmem_recv_msg(s, &fd, &err);
 667     if (err) {
 668         error_propagate(errp, err);
 669         return;
 670     }
 671     if (msg != IVSHMEM_PROTOCOL_VERSION) {
 672         error_setg(errp, "server sent version %" PRId64 ", expecting %d",
 673                    msg, IVSHMEM_PROTOCOL_VERSION);
 674         return;
 675     }
 676     if (fd != -1) {
 677         error_setg(errp, "server sent invalid version message");
 678         return;
 679     }
 680
 681     /*
 682      * ivshmem-server sends the remaining initial messages in a fixed
 683      * order, but the device has always accepted them in any order.
 684      * Stay as compatible as practical, just in case people use
 685      * servers that behave differently.
 686      */
 687
 688     /*
 689      * ivshmem_device_spec.txt has always required the ID message
 690      * right here, and ivshmem-server has always complied.  However,
 691      * older versions of the device accepted it out of order, but
 692      * broke when an interrupt setup message arrived before it.
 693      */
 694     msg = ivshmem_recv_msg(s, &fd, &err);
 695     if (err) {
 696         error_propagate(errp, err);
 697         return;
 698     }
 699     if (fd != -1 || msg < 0 || msg > IVSHMEM_MAX_PEERS) {
 700         error_setg(errp, "server sent invalid ID message");
 701         return;
 702     }
 703     s->vm_id = msg;
 704
 705     /*
 706      * Receive more messages until we got shared memory.
 707      */
 708     do {
 709         msg = ivshmem_recv_msg(s, &fd, &err);
 710         if (err) {
 711             error_propagate(errp, err);
 712             return;
 713         }
 714         process_msg(s, msg, fd, &err);
 715         if (err) {
 716             error_propagate(errp, err);
 717             return;
 718         }
 719     } while (msg != -1);
 720
 721     /*
 722      * This function must either map the shared memory or fail.  The
 723      * loop above ensures that: it terminates normally only after it
 724      * successfully processed the server's shared memory message.
 725      * Assert that actually mapped the shared memory:
 726      */
 727     assert(s->ivshmem_bar2);
 728 }
 729
 730 /* Select the MSI-X vectors used by device.
 731  * ivshmem maps events to vectors statically, so
 732  * we just enable all vectors on init and after reset. */
 733 static void ivshmem_msix_vector_use(IVShmemState *s)
 734 {
 735     PCIDevice *d = PCI_DEVICE(s);
 736     int i;
 737
 738     for (i = 0; i < s->vectors; i++) {
 739         msix_vector_use(d, i);
 740     }
 741 }
 742
 743 static void ivshmem_reset(DeviceState *d)
 744 {
 745     IVShmemState *s = IVSHMEM_COMMON(d);
 746
 747     s->intrstatus = 0;
 748     s->intrmask = 0;
 749     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 750         ivshmem_msix_vector_use(s);
 751     }
 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_msix_vector_use(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         Error *err = NULL;
 778
 779         ivshmem_add_kvm_msi_virq(s, i, &err);
 780         if (err) {
 781             error_report_err(err);
 782             /* TODO do we need to handle the error? */
 783         }
 784     }
 785
 786     if (msix_set_vector_notifiers(pdev,
 787                                   ivshmem_vector_unmask,
 788                                   ivshmem_vector_mask,
 789                                   ivshmem_vector_poll)) {
 790         error_report("ivshmem: msix_set_vector_notifiers failed");
 791     }
 792 }
 793
 794 static void ivshmem_remove_kvm_msi_virq(IVShmemState *s, int vector)
 795 {
 796     IVSHMEM_DPRINTF("ivshmem_remove_kvm_msi_virq vector:%d\n", vector);
 797
 798     if (s->msi_vectors[vector].pdev == NULL) {
 799         return;
 800     }
 801
 802     /* it was cleaned when masked in the frontend. */
 803     kvm_irqchip_release_virq(kvm_state, s->msi_vectors[vector].virq);
 804
 805     s->msi_vectors[vector].pdev = NULL;
 806 }
 807
 808 static void ivshmem_disable_irqfd(IVShmemState *s)
 809 {
 810     PCIDevice *pdev = PCI_DEVICE(s);
 811     int i;
 812
 813     for (i = 0; i < s->peers[s->vm_id].nb_eventfds; i++) {
 814         ivshmem_remove_kvm_msi_virq(s, i);
 815     }
 816
 817     msix_unset_vector_notifiers(pdev);
 818 }
 819
 820 static void ivshmem_write_config(PCIDevice *pdev, uint32_t address,
 821                                  uint32_t val, int len)
 822 {
 823     IVShmemState *s = IVSHMEM_COMMON(pdev);
 824     int is_enabled, was_enabled = msix_enabled(pdev);
 825
 826     pci_default_write_config(pdev, address, val, len);
 827     is_enabled = msix_enabled(pdev);
 828
 829     if (kvm_msi_via_irqfd_enabled()) {
 830         if (!was_enabled && is_enabled) {
 831             ivshmem_enable_irqfd(s);
 832         } else if (was_enabled && !is_enabled) {
 833             ivshmem_disable_irqfd(s);
 834         }
 835     }
 836 }
 837
 838 static void ivshmem_common_realize(PCIDevice *dev, Error **errp)
 839 {
 840     IVShmemState *s = IVSHMEM_COMMON(dev);
 841     Error *err = NULL;
 842     uint8_t *pci_conf;
 843     uint8_t attr = PCI_BASE_ADDRESS_SPACE_MEMORY |
 844         PCI_BASE_ADDRESS_MEM_PREFETCH;
 845
 846     /* IRQFD requires MSI */
 847     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD) &&
 848         !ivshmem_has_feature(s, IVSHMEM_MSI)) {
 849         error_setg(errp, "ioeventfd/irqfd requires MSI");
 850         return;
 851     }
 852
 853     pci_conf = dev->config;
 854     pci_conf[PCI_COMMAND] = PCI_COMMAND_IO | PCI_COMMAND_MEMORY;
 855
 856     memory_region_init_io(&s->ivshmem_mmio, OBJECT(s), &ivshmem_mmio_ops, s,
 857                           "ivshmem-mmio", IVSHMEM_REG_BAR_SIZE);
 858
 859     /* region for registers*/
 860     pci_register_bar(dev, 0, PCI_BASE_ADDRESS_SPACE_MEMORY,
 861                      &s->ivshmem_mmio);
 862
 863     if (!s->not_legacy_32bit) {
 864         attr |= PCI_BASE_ADDRESS_MEM_TYPE_64;
 865     }
 866
 867     if (s->hostmem != NULL) {
 868         IVSHMEM_DPRINTF("using hostmem\n");
 869
 870         s->ivshmem_bar2 = host_memory_backend_get_memory(s->hostmem,
 871                                                          &error_abort);
 872     } else {
 873         IVSHMEM_DPRINTF("using shared memory server (socket = %s)\n",
 874                         s->server_chr->filename);
 875
 876         /* we allocate enough space for 16 peers and grow as needed */
 877         resize_peers(s, 16);
 878
 879         /*
 880          * Receive setup messages from server synchronously.
 881          * Older versions did it asynchronously, but that creates a
 882          * number of entertaining race conditions.
 883          */
 884         ivshmem_recv_setup(s, &err);
 885         if (err) {
 886             error_propagate(errp, err);
 887             return;
 888         }
 889
 890         if (s->master == ON_OFF_AUTO_ON && s->vm_id != 0) {
 891             error_setg(errp,
 892                        "master must connect to the server before any peers");
 893             return;
 894         }
 895
 896         qemu_chr_add_handlers(s->server_chr, ivshmem_can_receive,
 897                               ivshmem_read, NULL, s);
 898
 899         if (ivshmem_setup_interrupts(s) < 0) {
 900             error_setg(errp, "failed to initialize interrupts");
 901             return;
 902         }
 903     }
 904
 905     vmstate_register_ram(s->ivshmem_bar2, DEVICE(s));
 906     pci_register_bar(PCI_DEVICE(s), 2, attr, s->ivshmem_bar2);
 907
 908     if (s->master == ON_OFF_AUTO_AUTO) {
 909         s->master = s->vm_id == 0 ? ON_OFF_AUTO_ON : ON_OFF_AUTO_OFF;
 910     }
 911
 912     if (!ivshmem_is_master(s)) {
 913         error_setg(&s->migration_blocker,
 914                    "Migration is disabled when using feature 'peer mode' in device 'ivshmem'");
 915         migrate_add_blocker(s->migration_blocker);
 916     }
 917 }
 918
 919 static void ivshmem_exit(PCIDevice *dev)
 920 {
 921     IVShmemState *s = IVSHMEM_COMMON(dev);
 922     int i;
 923
 924     if (s->migration_blocker) {
 925         migrate_del_blocker(s->migration_blocker);
 926         error_free(s->migration_blocker);
 927     }
 928
 929     if (memory_region_is_mapped(s->ivshmem_bar2)) {
 930         if (!s->hostmem) {
 931             void *addr = memory_region_get_ram_ptr(s->ivshmem_bar2);
 932             int fd;
 933
 934             if (munmap(addr, memory_region_size(s->ivshmem_bar2) == -1)) {
 935                 error_report("Failed to munmap shared memory %s",
 936                              strerror(errno));
 937             }
 938
 939             fd = qemu_get_ram_fd(memory_region_get_ram_addr(s->ivshmem_bar2));
 940             close(fd);
 941         }
 942
 943         vmstate_unregister_ram(s->ivshmem_bar2, DEVICE(dev));
 944     }
 945
 946     if (s->peers) {
 947         for (i = 0; i < s->nb_peers; i++) {
 948             close_peer_eventfds(s, i);
 949         }
 950         g_free(s->peers);
 951     }
 952
 953     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 954         msix_uninit_exclusive_bar(dev);
 955     }
 956
 957     g_free(s->msi_vectors);
 958 }
 959
 960 static int ivshmem_pre_load(void *opaque)
 961 {
 962     IVShmemState *s = opaque;
 963
 964     if (!ivshmem_is_master(s)) {
 965         error_report("'peer' devices are not migratable");
 966         return -EINVAL;
 967     }
 968
 969     return 0;
 970 }
 971
 972 static int ivshmem_post_load(void *opaque, int version_id)
 973 {
 974     IVShmemState *s = opaque;
 975
 976     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
 977         ivshmem_msix_vector_use(s);
 978     }
 979     return 0;
 980 }
 981
 982 static void ivshmem_common_class_init(ObjectClass *klass, void *data)
 983 {
 984     DeviceClass *dc = DEVICE_CLASS(klass);
 985     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
 986
 987     k->realize = ivshmem_common_realize;
 988     k->exit = ivshmem_exit;
 989     k->config_write = ivshmem_write_config;
 990     k->vendor_id = PCI_VENDOR_ID_IVSHMEM;
 991     k->device_id = PCI_DEVICE_ID_IVSHMEM;
 992     k->class_id = PCI_CLASS_MEMORY_RAM;
 993     k->revision = 1;
 994     dc->reset = ivshmem_reset;
 995     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
 996     dc->desc = "Inter-VM shared memory";
 997 }
 998
 999 static const TypeInfo ivshmem_common_info = {
1000     .name          = TYPE_IVSHMEM_COMMON,
1001     .parent        = TYPE_PCI_DEVICE,
1002     .instance_size = sizeof(IVShmemState),
1003     .abstract      = true,
1004     .class_init    = ivshmem_common_class_init,
1005 };
1006
1007 static void ivshmem_check_memdev_is_busy(Object *obj, const char *name,
1008                                          Object *val, Error **errp)
1009 {
1010     MemoryRegion *mr;
1011
1012     mr = host_memory_backend_get_memory(MEMORY_BACKEND(val), &error_abort);
1013     if (memory_region_is_mapped(mr)) {
1014         char *path = object_get_canonical_path_component(val);
1015         error_setg(errp, "can't use already busy memdev: %s", path);
1016         g_free(path);
1017     } else {
1018         qdev_prop_allow_set_link_before_realize(obj, name, val, errp);
1019     }
1020 }
1021
1022 static const VMStateDescription ivshmem_plain_vmsd = {
1023     .name = TYPE_IVSHMEM_PLAIN,
1024     .version_id = 0,
1025     .minimum_version_id = 0,
1026     .pre_load = ivshmem_pre_load,
1027     .post_load = ivshmem_post_load,
1028     .fields = (VMStateField[]) {
1029         VMSTATE_PCI_DEVICE(parent_obj, IVShmemState),
1030         VMSTATE_UINT32(intrstatus, IVShmemState),
1031         VMSTATE_UINT32(intrmask, IVShmemState),
1032         VMSTATE_END_OF_LIST()
1033     },
1034 };
1035
1036 static Property ivshmem_plain_properties[] = {
1037     DEFINE_PROP_ON_OFF_AUTO("master", IVShmemState, master, ON_OFF_AUTO_OFF),
1038     DEFINE_PROP_END_OF_LIST(),
1039 };
1040
1041 static void ivshmem_plain_init(Object *obj)
1042 {
1043     IVShmemState *s = IVSHMEM_PLAIN(obj);
1044
1045     object_property_add_link(obj, "memdev", TYPE_MEMORY_BACKEND,
1046                              (Object **)&s->hostmem,
1047                              ivshmem_check_memdev_is_busy,
1048                              OBJ_PROP_LINK_UNREF_ON_RELEASE,
1049                              &error_abort);
1050 }
1051
1052 static void ivshmem_plain_class_init(ObjectClass *klass, void *data)
1053 {
1054     DeviceClass *dc = DEVICE_CLASS(klass);
1055
1056     dc->props = ivshmem_plain_properties;
1057     dc->vmsd = &ivshmem_plain_vmsd;
1058 }
1059
1060 static const TypeInfo ivshmem_plain_info = {
1061     .name          = TYPE_IVSHMEM_PLAIN,
1062     .parent        = TYPE_IVSHMEM_COMMON,
1063     .instance_size = sizeof(IVShmemState),
1064     .instance_init = ivshmem_plain_init,
1065     .class_init    = ivshmem_plain_class_init,
1066 };
1067
1068 static const VMStateDescription ivshmem_doorbell_vmsd = {
1069     .name = TYPE_IVSHMEM_DOORBELL,
1070     .version_id = 0,
1071     .minimum_version_id = 0,
1072     .pre_load = ivshmem_pre_load,
1073     .post_load = ivshmem_post_load,
1074     .fields = (VMStateField[]) {
1075         VMSTATE_PCI_DEVICE(parent_obj, IVShmemState),
1076         VMSTATE_MSIX(parent_obj, IVShmemState),
1077         VMSTATE_UINT32(intrstatus, IVShmemState),
1078         VMSTATE_UINT32(intrmask, IVShmemState),
1079         VMSTATE_END_OF_LIST()
1080     },
1081 };
1082
1083 static Property ivshmem_doorbell_properties[] = {
1084     DEFINE_PROP_CHR("chardev", IVShmemState, server_chr),
1085     DEFINE_PROP_UINT32("vectors", IVShmemState, vectors, 1),
1086     DEFINE_PROP_BIT("ioeventfd", IVShmemState, features, IVSHMEM_IOEVENTFD,
1087                     true),
1088     DEFINE_PROP_ON_OFF_AUTO("master", IVShmemState, master, ON_OFF_AUTO_OFF),
1089     DEFINE_PROP_END_OF_LIST(),
1090 };
1091
1092 static void ivshmem_doorbell_init(Object *obj)
1093 {
1094     IVShmemState *s = IVSHMEM_DOORBELL(obj);
1095
1096     s->features |= (1 << IVSHMEM_MSI);
1097     s->legacy_size = SIZE_MAX;  /* whatever the server sends */
1098 }
1099
1100 static void ivshmem_doorbell_class_init(ObjectClass *klass, void *data)
1101 {
1102     DeviceClass *dc = DEVICE_CLASS(klass);
1103
1104     dc->props = ivshmem_doorbell_properties;
1105     dc->vmsd = &ivshmem_doorbell_vmsd;
1106 }
1107
1108 static const TypeInfo ivshmem_doorbell_info = {
1109     .name          = TYPE_IVSHMEM_DOORBELL,
1110     .parent        = TYPE_IVSHMEM_COMMON,
1111     .instance_size = sizeof(IVShmemState),
1112     .instance_init = ivshmem_doorbell_init,
1113     .class_init    = ivshmem_doorbell_class_init,
1114 };
1115
1116 static int ivshmem_load_old(QEMUFile *f, void *opaque, int version_id)
1117 {
1118     IVShmemState *s = opaque;
1119     PCIDevice *pdev = PCI_DEVICE(s);
1120     int ret;
1121
1122     IVSHMEM_DPRINTF("ivshmem_load_old\n");
1123
1124     if (version_id != 0) {
1125         return -EINVAL;
1126     }
1127
1128     ret = ivshmem_pre_load(s);
1129     if (ret) {
1130         return ret;
1131     }
1132
1133     ret = pci_device_load(pdev, f);
1134     if (ret) {
1135         return ret;
1136     }
1137
1138     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
1139         msix_load(pdev, f);
1140         ivshmem_msix_vector_use(s);
1141     } else {
1142         s->intrstatus = qemu_get_be32(f);
1143         s->intrmask = qemu_get_be32(f);
1144     }
1145
1146     return 0;
1147 }
1148
1149 static bool test_msix(void *opaque, int version_id)
1150 {
1151     IVShmemState *s = opaque;
1152
1153     return ivshmem_has_feature(s, IVSHMEM_MSI);
1154 }
1155
1156 static bool test_no_msix(void *opaque, int version_id)
1157 {
1158     return !test_msix(opaque, version_id);
1159 }
1160
1161 static const VMStateDescription ivshmem_vmsd = {
1162     .name = "ivshmem",
1163     .version_id = 1,
1164     .minimum_version_id = 1,
1165     .pre_load = ivshmem_pre_load,
1166     .post_load = ivshmem_post_load,
1167     .fields = (VMStateField[]) {
1168         VMSTATE_PCI_DEVICE(parent_obj, IVShmemState),
1169
1170         VMSTATE_MSIX_TEST(parent_obj, IVShmemState, test_msix),
1171         VMSTATE_UINT32_TEST(intrstatus, IVShmemState, test_no_msix),
1172         VMSTATE_UINT32_TEST(intrmask, IVShmemState, test_no_msix),
1173
1174         VMSTATE_END_OF_LIST()
1175     },
1176     .load_state_old = ivshmem_load_old,
1177     .minimum_version_id_old = 0
1178 };
1179
1180 static Property ivshmem_properties[] = {
1181     DEFINE_PROP_CHR("chardev", IVShmemState, server_chr),
1182     DEFINE_PROP_STRING("size", IVShmemState, sizearg),
1183     DEFINE_PROP_UINT32("vectors", IVShmemState, vectors, 1),
1184     DEFINE_PROP_BIT("ioeventfd", IVShmemState, features, IVSHMEM_IOEVENTFD,
1185                     false),
1186     DEFINE_PROP_BIT("msi", IVShmemState, features, IVSHMEM_MSI, true),
1187     DEFINE_PROP_STRING("shm", IVShmemState, shmobj),
1188     DEFINE_PROP_STRING("role", IVShmemState, role),
1189     DEFINE_PROP_UINT32("use64", IVShmemState, not_legacy_32bit, 1),
1190     DEFINE_PROP_END_OF_LIST(),
1191 };
1192
1193 static void desugar_shm(IVShmemState *s)
1194 {
1195     Object *obj;
1196     char *path;
1197
1198     obj = object_new("memory-backend-file");
1199     path = g_strdup_printf("/dev/shm/%s", s->shmobj);
1200     object_property_set_str(obj, path, "mem-path", &error_abort);
1201     g_free(path);
1202     object_property_set_int(obj, s->legacy_size, "size", &error_abort);
1203     object_property_set_bool(obj, true, "share", &error_abort);
1204     object_property_add_child(OBJECT(s), "internal-shm-backend", obj,
1205                               &error_abort);
1206     user_creatable_complete(obj, &error_abort);
1207     s->hostmem = MEMORY_BACKEND(obj);
1208 }
1209
1210 static void ivshmem_realize(PCIDevice *dev, Error **errp)
1211 {
1212     IVShmemState *s = IVSHMEM_COMMON(dev);
1213
1214     if (!qtest_enabled()) {
1215         error_report("ivshmem is deprecated, please use ivshmem-plain"
1216                      " or ivshmem-doorbell instead");
1217     }
1218
1219     if (!!s->server_chr + !!s->shmobj != 1) {
1220         error_setg(errp, "You must specify either 'shm' or 'chardev'");
1221         return;
1222     }
1223
1224     if (s->sizearg == NULL) {
1225         s->legacy_size = 4 << 20; /* 4 MB default */
1226     } else {
1227         char *end;
1228         int64_t size = qemu_strtosz(s->sizearg, &end);
1229         if (size < 0 || (size_t)size != size || *end != '\0'
1230             || !is_power_of_2(size)) {
1231             error_setg(errp, "Invalid size %s", s->sizearg);
1232             return;
1233         }
1234         s->legacy_size = size;
1235     }
1236
1237     /* check that role is reasonable */
1238     if (s->role) {
1239         if (strncmp(s->role, "peer", 5) == 0) {
1240             s->master = ON_OFF_AUTO_OFF;
1241         } else if (strncmp(s->role, "master", 7) == 0) {
1242             s->master = ON_OFF_AUTO_ON;
1243         } else {
1244             error_setg(errp, "'role' must be 'peer' or 'master'");
1245             return;
1246         }
1247     } else {
1248         s->master = ON_OFF_AUTO_AUTO;
1249     }
1250
1251     if (s->shmobj) {
1252         desugar_shm(s);
1253     }
1254
1255     /*
1256      * Note: we don't use INTx with IVSHMEM_MSI at all, so this is a
1257      * bald-faced lie then.  But it's a backwards compatible lie.
1258      */
1259     pci_config_set_interrupt_pin(dev->config, 1);
1260
1261     ivshmem_common_realize(dev, errp);
1262 }
1263
1264 static void ivshmem_class_init(ObjectClass *klass, void *data)
1265 {
1266     DeviceClass *dc = DEVICE_CLASS(klass);
1267     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
1268
1269     k->realize = ivshmem_realize;
1270     k->revision = 0;
1271     dc->desc = "Inter-VM shared memory (legacy)";
1272     dc->props = ivshmem_properties;
1273     dc->vmsd = &ivshmem_vmsd;
1274 }
1275
1276 static const TypeInfo ivshmem_info = {
1277     .name          = TYPE_IVSHMEM,
1278     .parent        = TYPE_IVSHMEM_COMMON,
1279     .instance_size = sizeof(IVShmemState),
1280     .class_init    = ivshmem_class_init,
1281 };
1282
1283 static void ivshmem_register_types(void)
1284 {
1285     type_register_static(&ivshmem_common_info);
1286     type_register_static(&ivshmem_plain_info);
1287     type_register_static(&ivshmem_doorbell_info);
1288     type_register_static(&ivshmem_info);
1289 }
1290
1291 type_init(ivshmem_register_types)