2 * Inter-VM Shared Memory PCI device.
7 * Based On: cirrus_vga.c
8 * Copyright (c) 2004 Fabrice Bellard
9 * Copyright (c) 2004 Makoto Suzuki (suzu)
12 * Copyright (c) 2006 Igor Kovalenko
14 * This code is licensed under the GNU GPL v2.
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.
24 #include "migration.h"
26 #include "event_notifier.h"
29 #include <sys/types.h>
31 #define IVSHMEM_IOEVENTFD 0
34 #define IVSHMEM_PEER 0
35 #define IVSHMEM_MASTER 1
37 #define IVSHMEM_REG_BAR_SIZE 0x100
39 //#define DEBUG_IVSHMEM
41 #define IVSHMEM_DPRINTF(fmt, ...) \
42 do {printf("IVSHMEM: " fmt, ## __VA_ARGS__); } while (0)
44 #define IVSHMEM_DPRINTF(fmt, ...)
49 EventNotifier *eventfds;
52 typedef struct EventfdEntry {
57 typedef struct IVShmemState {
63 CharDriverState **eventfd_chr;
64 CharDriverState *server_chr;
65 MemoryRegion ivshmem_mmio;
67 /* We might need to register the BAR before we actually have the memory.
68 * So prepare a container MemoryRegion for the BAR immediately and
69 * add a subregion when we have the memory.
73 uint64_t ivshmem_size; /* size of shared memory region */
74 int shm_fd; /* shared memory file descriptor */
77 int nb_peers; /* how many guests we have space for */
78 int max_peer; /* maximum numbered peer */
83 EventfdEntry *eventfd_table;
85 Error *migration_blocker;
90 int role_val; /* scalar to avoid multiple string comparisons */
93 /* registers for the Inter-VM shared memory device */
94 enum ivshmem_registers {
101 static inline uint32_t ivshmem_has_feature(IVShmemState *ivs,
102 unsigned int feature) {
103 return (ivs->features & (1 << feature));
106 static inline bool is_power_of_two(uint64_t x) {
107 return (x & (x - 1)) == 0;
110 /* accessing registers - based on rtl8139 */
111 static void ivshmem_update_irq(IVShmemState *s, int val)
114 isr = (s->intrstatus & s->intrmask) & 0xffffffff;
116 /* don't print ISR resets */
118 IVSHMEM_DPRINTF("Set IRQ to %d (%04x %04x)\n",
119 isr ? 1 : 0, s->intrstatus, s->intrmask);
122 qemu_set_irq(s->dev.irq[0], (isr != 0));
125 static void ivshmem_IntrMask_write(IVShmemState *s, uint32_t val)
127 IVSHMEM_DPRINTF("IntrMask write(w) val = 0x%04x\n", val);
131 ivshmem_update_irq(s, val);
134 static uint32_t ivshmem_IntrMask_read(IVShmemState *s)
136 uint32_t ret = s->intrmask;
138 IVSHMEM_DPRINTF("intrmask read(w) val = 0x%04x\n", ret);
143 static void ivshmem_IntrStatus_write(IVShmemState *s, uint32_t val)
145 IVSHMEM_DPRINTF("IntrStatus write(w) val = 0x%04x\n", val);
149 ivshmem_update_irq(s, val);
153 static uint32_t ivshmem_IntrStatus_read(IVShmemState *s)
155 uint32_t ret = s->intrstatus;
157 /* reading ISR clears all interrupts */
160 ivshmem_update_irq(s, 0);
165 static void ivshmem_io_write(void *opaque, target_phys_addr_t addr,
166 uint64_t val, unsigned size)
168 IVShmemState *s = opaque;
170 uint16_t dest = val >> 16;
171 uint16_t vector = val & 0xff;
175 IVSHMEM_DPRINTF("writing to addr " TARGET_FMT_plx "\n", addr);
179 ivshmem_IntrMask_write(s, val);
183 ivshmem_IntrStatus_write(s, val);
187 /* check that dest VM ID is reasonable */
188 if (dest > s->max_peer) {
189 IVSHMEM_DPRINTF("Invalid destination VM ID (%d)\n", dest);
193 /* check doorbell range */
194 if (vector < s->peers[dest].nb_eventfds) {
195 IVSHMEM_DPRINTF("Notifying VM %d on vector %d\n", dest, vector);
196 event_notifier_set(&s->peers[dest].eventfds[vector]);
200 IVSHMEM_DPRINTF("Invalid VM Doorbell VM %d\n", dest);
204 static uint64_t ivshmem_io_read(void *opaque, target_phys_addr_t addr,
208 IVShmemState *s = opaque;
214 ret = ivshmem_IntrMask_read(s);
218 ret = ivshmem_IntrStatus_read(s);
222 /* return my VM ID if the memory is mapped */
231 IVSHMEM_DPRINTF("why are we reading " TARGET_FMT_plx "\n", addr);
238 static const MemoryRegionOps ivshmem_mmio_ops = {
239 .read = ivshmem_io_read,
240 .write = ivshmem_io_write,
241 .endianness = DEVICE_NATIVE_ENDIAN,
243 .min_access_size = 4,
244 .max_access_size = 4,
248 static void ivshmem_receive(void *opaque, const uint8_t *buf, int size)
250 IVShmemState *s = opaque;
252 ivshmem_IntrStatus_write(s, *buf);
254 IVSHMEM_DPRINTF("ivshmem_receive 0x%02x\n", *buf);
257 static int ivshmem_can_receive(void * opaque)
262 static void ivshmem_event(void *opaque, int event)
264 IVSHMEM_DPRINTF("ivshmem_event %d\n", event);
267 static void fake_irqfd(void *opaque, const uint8_t *buf, int size) {
269 EventfdEntry *entry = opaque;
270 PCIDevice *pdev = entry->pdev;
272 IVSHMEM_DPRINTF("interrupt on vector %p %d\n", pdev, entry->vector);
273 msix_notify(pdev, entry->vector);
276 static CharDriverState* create_eventfd_chr_device(void * opaque, EventNotifier *n,
279 /* create a event character device based on the passed eventfd */
280 IVShmemState *s = opaque;
281 CharDriverState * chr;
282 int eventfd = event_notifier_get_fd(n);
284 chr = qemu_chr_open_eventfd(eventfd);
287 fprintf(stderr, "creating eventfd for eventfd %d failed\n", eventfd);
291 /* if MSI is supported we need multiple interrupts */
292 if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
293 s->eventfd_table[vector].pdev = &s->dev;
294 s->eventfd_table[vector].vector = vector;
296 qemu_chr_add_handlers(chr, ivshmem_can_receive, fake_irqfd,
297 ivshmem_event, &s->eventfd_table[vector]);
299 qemu_chr_add_handlers(chr, ivshmem_can_receive, ivshmem_receive,
307 static int check_shm_size(IVShmemState *s, int fd) {
308 /* check that the guest isn't going to try and map more memory than the
309 * the object has allocated return -1 to indicate error */
315 if (s->ivshmem_size > buf.st_size) {
317 "IVSHMEM ERROR: Requested memory size greater"
318 " than shared object size (%" PRIu64 " > %" PRIu64")\n",
319 s->ivshmem_size, (uint64_t)buf.st_size);
326 /* create the shared memory BAR when we are not using the server, so we can
327 * create the BAR and map the memory immediately */
328 static void create_shared_memory_BAR(IVShmemState *s, int fd) {
334 ptr = mmap(0, s->ivshmem_size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
336 memory_region_init_ram_ptr(&s->ivshmem, "ivshmem.bar2",
337 s->ivshmem_size, ptr);
338 vmstate_register_ram(&s->ivshmem, &s->dev.qdev);
339 memory_region_add_subregion(&s->bar, 0, &s->ivshmem);
341 /* region for shared memory */
342 pci_register_bar(&s->dev, 2, PCI_BASE_ADDRESS_SPACE_MEMORY, &s->bar);
345 static void ivshmem_add_eventfd(IVShmemState *s, int posn, int i)
347 memory_region_add_eventfd(&s->ivshmem_mmio,
352 &s->peers[posn].eventfds[i]);
355 static void ivshmem_del_eventfd(IVShmemState *s, int posn, int i)
357 memory_region_del_eventfd(&s->ivshmem_mmio,
362 &s->peers[posn].eventfds[i]);
365 static void close_guest_eventfds(IVShmemState *s, int posn)
367 int i, guest_curr_max;
369 guest_curr_max = s->peers[posn].nb_eventfds;
371 memory_region_transaction_begin();
372 for (i = 0; i < guest_curr_max; i++) {
373 ivshmem_del_eventfd(s, posn, i);
375 memory_region_transaction_commit();
376 for (i = 0; i < guest_curr_max; i++) {
377 event_notifier_cleanup(&s->peers[posn].eventfds[i]);
380 g_free(s->peers[posn].eventfds);
381 s->peers[posn].nb_eventfds = 0;
384 static void setup_ioeventfds(IVShmemState *s) {
388 for (i = 0; i <= s->max_peer; i++) {
389 for (j = 0; j < s->peers[i].nb_eventfds; j++) {
390 ivshmem_add_eventfd(s, i, j);
395 /* this function increase the dynamic storage need to store data about other
397 static void increase_dynamic_storage(IVShmemState *s, int new_min_size) {
401 old_nb_alloc = s->nb_peers;
403 while (new_min_size >= s->nb_peers)
404 s->nb_peers = s->nb_peers * 2;
406 IVSHMEM_DPRINTF("bumping storage to %d guests\n", s->nb_peers);
407 s->peers = g_realloc(s->peers, s->nb_peers * sizeof(Peer));
409 /* zero out new pointers */
410 for (j = old_nb_alloc; j < s->nb_peers; j++) {
411 s->peers[j].eventfds = NULL;
412 s->peers[j].nb_eventfds = 0;
416 static void ivshmem_read(void *opaque, const uint8_t * buf, int flags)
418 IVShmemState *s = opaque;
419 int incoming_fd, tmp_fd;
420 int guest_max_eventfd;
423 memcpy(&incoming_posn, buf, sizeof(long));
424 /* pick off s->server_chr->msgfd and store it, posn should accompany msg */
425 tmp_fd = qemu_chr_fe_get_msgfd(s->server_chr);
426 IVSHMEM_DPRINTF("posn is %ld, fd is %d\n", incoming_posn, tmp_fd);
428 /* make sure we have enough space for this guest */
429 if (incoming_posn >= s->nb_peers) {
430 increase_dynamic_storage(s, incoming_posn);
434 /* if posn is positive and unseen before then this is our posn*/
435 if ((incoming_posn >= 0) &&
436 (s->peers[incoming_posn].eventfds == NULL)) {
437 /* receive our posn */
438 s->vm_id = incoming_posn;
441 /* otherwise an fd == -1 means an existing guest has gone away */
442 IVSHMEM_DPRINTF("posn %ld has gone away\n", incoming_posn);
443 close_guest_eventfds(s, incoming_posn);
448 /* because of the implementation of get_msgfd, we need a dup */
449 incoming_fd = dup(tmp_fd);
451 if (incoming_fd == -1) {
452 fprintf(stderr, "could not allocate file descriptor %s\n",
457 /* if the position is -1, then it's shared memory region fd */
458 if (incoming_posn == -1) {
464 if (check_shm_size(s, incoming_fd) == -1) {
468 /* mmap the region and map into the BAR2 */
469 map_ptr = mmap(0, s->ivshmem_size, PROT_READ|PROT_WRITE, MAP_SHARED,
471 memory_region_init_ram_ptr(&s->ivshmem,
472 "ivshmem.bar2", s->ivshmem_size, map_ptr);
473 vmstate_register_ram(&s->ivshmem, &s->dev.qdev);
475 IVSHMEM_DPRINTF("guest h/w addr = %" PRIu64 ", size = %" PRIu64 "\n",
476 s->ivshmem_offset, s->ivshmem_size);
478 memory_region_add_subregion(&s->bar, 0, &s->ivshmem);
480 /* only store the fd if it is successfully mapped */
481 s->shm_fd = incoming_fd;
486 /* each guest has an array of eventfds, and we keep track of how many
487 * guests for each VM */
488 guest_max_eventfd = s->peers[incoming_posn].nb_eventfds;
490 if (guest_max_eventfd == 0) {
491 /* one eventfd per MSI vector */
492 s->peers[incoming_posn].eventfds = g_new(EventNotifier, s->vectors);
495 /* this is an eventfd for a particular guest VM */
496 IVSHMEM_DPRINTF("eventfds[%ld][%d] = %d\n", incoming_posn,
497 guest_max_eventfd, incoming_fd);
498 event_notifier_init_fd(&s->peers[incoming_posn].eventfds[guest_max_eventfd],
501 /* increment count for particular guest */
502 s->peers[incoming_posn].nb_eventfds++;
504 /* keep track of the maximum VM ID */
505 if (incoming_posn > s->max_peer) {
506 s->max_peer = incoming_posn;
509 if (incoming_posn == s->vm_id) {
510 s->eventfd_chr[guest_max_eventfd] = create_eventfd_chr_device(s,
511 &s->peers[s->vm_id].eventfds[guest_max_eventfd],
515 if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
516 ivshmem_add_eventfd(s, incoming_posn, guest_max_eventfd);
522 /* Select the MSI-X vectors used by device.
523 * ivshmem maps events to vectors statically, so
524 * we just enable all vectors on init and after reset. */
525 static void ivshmem_use_msix(IVShmemState * s)
529 if (!msix_present(&s->dev)) {
533 for (i = 0; i < s->vectors; i++) {
534 msix_vector_use(&s->dev, i);
538 static void ivshmem_reset(DeviceState *d)
540 IVShmemState *s = DO_UPCAST(IVShmemState, dev.qdev, d);
547 static uint64_t ivshmem_get_size(IVShmemState * s) {
552 value = strtoull(s->sizearg, &ptr, 10);
554 case 0: case 'M': case 'm':
561 fprintf(stderr, "qemu: invalid ram size: %s\n", s->sizearg);
565 /* BARs must be a power of 2 */
566 if (!is_power_of_two(value)) {
567 fprintf(stderr, "ivshmem: size must be power of 2\n");
574 static void ivshmem_setup_msi(IVShmemState * s)
576 if (msix_init_exclusive_bar(&s->dev, s->vectors, 1)) {
577 IVSHMEM_DPRINTF("msix initialization failed\n");
581 IVSHMEM_DPRINTF("msix initialized (%d vectors)\n", s->vectors);
583 /* allocate QEMU char devices for receiving interrupts */
584 s->eventfd_table = g_malloc0(s->vectors * sizeof(EventfdEntry));
589 static void ivshmem_save(QEMUFile* f, void *opaque)
591 IVShmemState *proxy = opaque;
593 IVSHMEM_DPRINTF("ivshmem_save\n");
594 pci_device_save(&proxy->dev, f);
596 if (ivshmem_has_feature(proxy, IVSHMEM_MSI)) {
597 msix_save(&proxy->dev, f);
599 qemu_put_be32(f, proxy->intrstatus);
600 qemu_put_be32(f, proxy->intrmask);
605 static int ivshmem_load(QEMUFile* f, void *opaque, int version_id)
607 IVSHMEM_DPRINTF("ivshmem_load\n");
609 IVShmemState *proxy = opaque;
612 if (version_id > 0) {
616 if (proxy->role_val == IVSHMEM_PEER) {
617 fprintf(stderr, "ivshmem: 'peer' devices are not migratable\n");
621 ret = pci_device_load(&proxy->dev, f);
626 if (ivshmem_has_feature(proxy, IVSHMEM_MSI)) {
627 msix_load(&proxy->dev, f);
628 ivshmem_use_msix(proxy);
630 proxy->intrstatus = qemu_get_be32(f);
631 proxy->intrmask = qemu_get_be32(f);
637 static void ivshmem_write_config(PCIDevice *pci_dev, uint32_t address,
638 uint32_t val, int len)
640 pci_default_write_config(pci_dev, address, val, len);
641 msix_write_config(pci_dev, address, val, len);
644 static int pci_ivshmem_init(PCIDevice *dev)
646 IVShmemState *s = DO_UPCAST(IVShmemState, dev, dev);
649 if (s->sizearg == NULL)
650 s->ivshmem_size = 4 << 20; /* 4 MB default */
652 s->ivshmem_size = ivshmem_get_size(s);
655 register_savevm(&s->dev.qdev, "ivshmem", 0, 0, ivshmem_save, ivshmem_load,
658 /* IRQFD requires MSI */
659 if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD) &&
660 !ivshmem_has_feature(s, IVSHMEM_MSI)) {
661 fprintf(stderr, "ivshmem: ioeventfd/irqfd requires MSI\n");
665 /* check that role is reasonable */
667 if (strncmp(s->role, "peer", 5) == 0) {
668 s->role_val = IVSHMEM_PEER;
669 } else if (strncmp(s->role, "master", 7) == 0) {
670 s->role_val = IVSHMEM_MASTER;
672 fprintf(stderr, "ivshmem: 'role' must be 'peer' or 'master'\n");
676 s->role_val = IVSHMEM_MASTER; /* default */
679 if (s->role_val == IVSHMEM_PEER) {
680 error_set(&s->migration_blocker, QERR_DEVICE_FEATURE_BLOCKS_MIGRATION, "ivshmem", "peer mode");
681 migrate_add_blocker(s->migration_blocker);
684 pci_conf = s->dev.config;
685 pci_conf[PCI_COMMAND] = PCI_COMMAND_IO | PCI_COMMAND_MEMORY;
687 pci_config_set_interrupt_pin(pci_conf, 1);
691 memory_region_init_io(&s->ivshmem_mmio, &ivshmem_mmio_ops, s,
692 "ivshmem-mmio", IVSHMEM_REG_BAR_SIZE);
694 if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
698 /* region for registers*/
699 pci_register_bar(&s->dev, 0, PCI_BASE_ADDRESS_SPACE_MEMORY,
702 memory_region_init(&s->bar, "ivshmem-bar2-container", s->ivshmem_size);
704 if ((s->server_chr != NULL) &&
705 (strncmp(s->server_chr->filename, "unix:", 5) == 0)) {
706 /* if we get a UNIX socket as the parameter we will talk
707 * to the ivshmem server to receive the memory region */
709 if (s->shmobj != NULL) {
710 fprintf(stderr, "WARNING: do not specify both 'chardev' "
711 "and 'shm' with ivshmem\n");
714 IVSHMEM_DPRINTF("using shared memory server (socket = %s)\n",
715 s->server_chr->filename);
717 if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
718 ivshmem_setup_msi(s);
721 /* we allocate enough space for 16 guests and grow as needed */
725 /* allocate/initialize space for interrupt handling */
726 s->peers = g_malloc0(s->nb_peers * sizeof(Peer));
728 pci_register_bar(&s->dev, 2,
729 PCI_BASE_ADDRESS_SPACE_MEMORY, &s->bar);
731 s->eventfd_chr = g_malloc0(s->vectors * sizeof(CharDriverState *));
733 qemu_chr_add_handlers(s->server_chr, ivshmem_can_receive, ivshmem_read,
736 /* just map the file immediately, we're not using a server */
739 if (s->shmobj == NULL) {
740 fprintf(stderr, "Must specify 'chardev' or 'shm' to ivshmem\n");
743 IVSHMEM_DPRINTF("using shm_open (shm object = %s)\n", s->shmobj);
745 /* try opening with O_EXCL and if it succeeds zero the memory
746 * by truncating to 0 */
747 if ((fd = shm_open(s->shmobj, O_CREAT|O_RDWR|O_EXCL,
748 S_IRWXU|S_IRWXG|S_IRWXO)) > 0) {
749 /* truncate file to length PCI device's memory */
750 if (ftruncate(fd, s->ivshmem_size) != 0) {
751 fprintf(stderr, "ivshmem: could not truncate shared file\n");
754 } else if ((fd = shm_open(s->shmobj, O_CREAT|O_RDWR,
755 S_IRWXU|S_IRWXG|S_IRWXO)) < 0) {
756 fprintf(stderr, "ivshmem: could not open shared file\n");
761 if (check_shm_size(s, fd) == -1) {
765 create_shared_memory_BAR(s, fd);
769 s->dev.config_write = ivshmem_write_config;
774 static void pci_ivshmem_uninit(PCIDevice *dev)
776 IVShmemState *s = DO_UPCAST(IVShmemState, dev, dev);
778 if (s->migration_blocker) {
779 migrate_del_blocker(s->migration_blocker);
780 error_free(s->migration_blocker);
783 memory_region_destroy(&s->ivshmem_mmio);
784 memory_region_del_subregion(&s->bar, &s->ivshmem);
785 vmstate_unregister_ram(&s->ivshmem, &s->dev.qdev);
786 memory_region_destroy(&s->ivshmem);
787 memory_region_destroy(&s->bar);
788 unregister_savevm(&dev->qdev, "ivshmem", s);
791 static Property ivshmem_properties[] = {
792 DEFINE_PROP_CHR("chardev", IVShmemState, server_chr),
793 DEFINE_PROP_STRING("size", IVShmemState, sizearg),
794 DEFINE_PROP_UINT32("vectors", IVShmemState, vectors, 1),
795 DEFINE_PROP_BIT("ioeventfd", IVShmemState, features, IVSHMEM_IOEVENTFD, false),
796 DEFINE_PROP_BIT("msi", IVShmemState, features, IVSHMEM_MSI, true),
797 DEFINE_PROP_STRING("shm", IVShmemState, shmobj),
798 DEFINE_PROP_STRING("role", IVShmemState, role),
799 DEFINE_PROP_END_OF_LIST(),
802 static void ivshmem_class_init(ObjectClass *klass, void *data)
804 DeviceClass *dc = DEVICE_CLASS(klass);
805 PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
807 k->init = pci_ivshmem_init;
808 k->exit = pci_ivshmem_uninit;
809 k->vendor_id = PCI_VENDOR_ID_REDHAT_QUMRANET;
810 k->device_id = 0x1110;
811 k->class_id = PCI_CLASS_MEMORY_RAM;
812 dc->reset = ivshmem_reset;
813 dc->props = ivshmem_properties;
816 static TypeInfo ivshmem_info = {
818 .parent = TYPE_PCI_DEVICE,
819 .instance_size = sizeof(IVShmemState),
820 .class_init = ivshmem_class_init,
823 static void ivshmem_register_types(void)
825 type_register_static(&ivshmem_info);
828 type_init(ivshmem_register_types)