xref: /openbmc/qemu/hw/misc/ivshmem.c (revision 8908eb1a)
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 "qapi/error.h"
21 #include "qemu/cutils.h"
22 #include "hw/hw.h"
23 #include "hw/i386/pc.h"
24 #include "hw/pci/pci.h"
25 #include "hw/pci/msi.h"
26 #include "hw/pci/msix.h"
27 #include "sysemu/kvm.h"
28 #include "migration/blocker.h"
29 #include "qemu/error-report.h"
30 #include "qemu/event_notifier.h"
31 #include "qom/object_interfaces.h"
32 #include "chardev/char-fe.h"
33 #include "sysemu/hostmem.h"
34 #include "sysemu/qtest.h"
35 #include "qapi/visitor.h"
36 
37 #include "hw/misc/ivshmem.h"
38 
39 #define PCI_VENDOR_ID_IVSHMEM   PCI_VENDOR_ID_REDHAT_QUMRANET
40 #define PCI_DEVICE_ID_IVSHMEM   0x1110
41 
42 #define IVSHMEM_MAX_PEERS UINT16_MAX
43 #define IVSHMEM_IOEVENTFD   0
44 #define IVSHMEM_MSI     1
45 
46 #define IVSHMEM_REG_BAR_SIZE 0x100
47 
48 #define IVSHMEM_DEBUG 0
49 #define IVSHMEM_DPRINTF(fmt, ...)                       \
50     do {                                                \
51         if (IVSHMEM_DEBUG) {                            \
52             printf("IVSHMEM: " fmt, ## __VA_ARGS__);    \
53         }                                               \
54     } while (0)
55 
56 #define TYPE_IVSHMEM_COMMON "ivshmem-common"
57 #define IVSHMEM_COMMON(obj) \
58     OBJECT_CHECK(IVShmemState, (obj), TYPE_IVSHMEM_COMMON)
59 
60 #define TYPE_IVSHMEM_PLAIN "ivshmem-plain"
61 #define IVSHMEM_PLAIN(obj) \
62     OBJECT_CHECK(IVShmemState, (obj), TYPE_IVSHMEM_PLAIN)
63 
64 #define TYPE_IVSHMEM_DOORBELL "ivshmem-doorbell"
65 #define IVSHMEM_DOORBELL(obj) \
66     OBJECT_CHECK(IVShmemState, (obj), TYPE_IVSHMEM_DOORBELL)
67 
68 #define TYPE_IVSHMEM "ivshmem"
69 #define IVSHMEM(obj) \
70     OBJECT_CHECK(IVShmemState, (obj), TYPE_IVSHMEM)
71 
72 typedef struct Peer {
73     int nb_eventfds;
74     EventNotifier *eventfds;
75 } Peer;
76 
77 typedef struct MSIVector {
78     PCIDevice *pdev;
79     int virq;
80 } MSIVector;
81 
82 typedef struct IVShmemState {
83     /*< private >*/
84     PCIDevice parent_obj;
85     /*< public >*/
86 
87     uint32_t features;
88 
89     /* exactly one of these two may be set */
90     HostMemoryBackend *hostmem; /* with interrupts */
91     CharBackend server_chr; /* without interrupts */
92 
93     /* registers */
94     uint32_t intrmask;
95     uint32_t intrstatus;
96     int vm_id;
97 
98     /* BARs */
99     MemoryRegion ivshmem_mmio;  /* BAR 0 (registers) */
100     MemoryRegion *ivshmem_bar2; /* BAR 2 (shared memory) */
101     MemoryRegion server_bar2;   /* used with server_chr */
102 
103     /* interrupt support */
104     Peer *peers;
105     int nb_peers;               /* space in @peers[] */
106     uint32_t vectors;
107     MSIVector *msi_vectors;
108     uint64_t msg_buf;           /* buffer for receiving server messages */
109     int msg_buffered_bytes;     /* #bytes in @msg_buf */
110 
111     /* migration stuff */
112     OnOffAuto master;
113     Error *migration_blocker;
114 
115     /* legacy cruft */
116     char *role;
117     char *shmobj;
118     char *sizearg;
119     size_t legacy_size;
120     uint32_t not_legacy_32bit;
121 } IVShmemState;
122 
123 /* registers for the Inter-VM shared memory device */
124 enum ivshmem_registers {
125     INTRMASK = 0,
126     INTRSTATUS = 4,
127     IVPOSITION = 8,
128     DOORBELL = 12,
129 };
130 
131 static inline uint32_t ivshmem_has_feature(IVShmemState *ivs,
132                                                     unsigned int feature) {
133     return (ivs->features & (1 << feature));
134 }
135 
136 static inline bool ivshmem_is_master(IVShmemState *s)
137 {
138     assert(s->master != ON_OFF_AUTO_AUTO);
139     return s->master == ON_OFF_AUTO_ON;
140 }
141 
142 static void ivshmem_update_irq(IVShmemState *s)
143 {
144     PCIDevice *d = PCI_DEVICE(s);
145     uint32_t isr = s->intrstatus & s->intrmask;
146 
147     /*
148      * Do nothing unless the device actually uses INTx.  Here's how
149      * the device variants signal interrupts, what they put in PCI
150      * config space:
151      * Device variant    Interrupt  Interrupt Pin  MSI-X cap.
152      * ivshmem-plain         none            0         no
153      * ivshmem-doorbell     MSI-X            1        yes(1)
154      * ivshmem,msi=off       INTx            1         no
155      * ivshmem,msi=on       MSI-X            1(2)     yes(1)
156      * (1) if guest enabled MSI-X
157      * (2) the device lies
158      * Leads to the condition for doing nothing:
159      */
160     if (ivshmem_has_feature(s, IVSHMEM_MSI)
161         || !d->config[PCI_INTERRUPT_PIN]) {
162         return;
163     }
164 
165     /* don't print ISR resets */
166     if (isr) {
167         IVSHMEM_DPRINTF("Set IRQ to %d (%04x %04x)\n",
168                         isr ? 1 : 0, s->intrstatus, s->intrmask);
169     }
170 
171     pci_set_irq(d, isr != 0);
172 }
173 
174 static void ivshmem_IntrMask_write(IVShmemState *s, uint32_t val)
175 {
176     IVSHMEM_DPRINTF("IntrMask write(w) val = 0x%04x\n", val);
177 
178     s->intrmask = val;
179     ivshmem_update_irq(s);
180 }
181 
182 static uint32_t ivshmem_IntrMask_read(IVShmemState *s)
183 {
184     uint32_t ret = s->intrmask;
185 
186     IVSHMEM_DPRINTF("intrmask read(w) val = 0x%04x\n", ret);
187     return ret;
188 }
189 
190 static void ivshmem_IntrStatus_write(IVShmemState *s, uint32_t val)
191 {
192     IVSHMEM_DPRINTF("IntrStatus write(w) val = 0x%04x\n", val);
193 
194     s->intrstatus = val;
195     ivshmem_update_irq(s);
196 }
197 
198 static uint32_t ivshmem_IntrStatus_read(IVShmemState *s)
199 {
200     uint32_t ret = s->intrstatus;
201 
202     /* reading ISR clears all interrupts */
203     s->intrstatus = 0;
204     ivshmem_update_irq(s);
205     return ret;
206 }
207 
208 static void ivshmem_io_write(void *opaque, hwaddr addr,
209                              uint64_t val, unsigned size)
210 {
211     IVShmemState *s = opaque;
212 
213     uint16_t dest = val >> 16;
214     uint16_t vector = val & 0xff;
215 
216     addr &= 0xfc;
217 
218     IVSHMEM_DPRINTF("writing to addr " TARGET_FMT_plx "\n", addr);
219     switch (addr)
220     {
221         case INTRMASK:
222             ivshmem_IntrMask_write(s, val);
223             break;
224 
225         case INTRSTATUS:
226             ivshmem_IntrStatus_write(s, val);
227             break;
228 
229         case DOORBELL:
230             /* check that dest VM ID is reasonable */
231             if (dest >= s->nb_peers) {
232                 IVSHMEM_DPRINTF("Invalid destination VM ID (%d)\n", dest);
233                 break;
234             }
235 
236             /* check doorbell range */
237             if (vector < s->peers[dest].nb_eventfds) {
238                 IVSHMEM_DPRINTF("Notifying VM %d on vector %d\n", dest, vector);
239                 event_notifier_set(&s->peers[dest].eventfds[vector]);
240             } else {
241                 IVSHMEM_DPRINTF("Invalid destination vector %d on VM %d\n",
242                                 vector, dest);
243             }
244             break;
245         default:
246             IVSHMEM_DPRINTF("Unhandled write " TARGET_FMT_plx "\n", addr);
247     }
248 }
249 
250 static uint64_t ivshmem_io_read(void *opaque, hwaddr addr,
251                                 unsigned size)
252 {
253 
254     IVShmemState *s = opaque;
255     uint32_t ret;
256 
257     switch (addr)
258     {
259         case INTRMASK:
260             ret = ivshmem_IntrMask_read(s);
261             break;
262 
263         case INTRSTATUS:
264             ret = ivshmem_IntrStatus_read(s);
265             break;
266 
267         case IVPOSITION:
268             ret = s->vm_id;
269             break;
270 
271         default:
272             IVSHMEM_DPRINTF("why are we reading " TARGET_FMT_plx "\n", addr);
273             ret = 0;
274     }
275 
276     return ret;
277 }
278 
279 static const MemoryRegionOps ivshmem_mmio_ops = {
280     .read = ivshmem_io_read,
281     .write = ivshmem_io_write,
282     .endianness = DEVICE_NATIVE_ENDIAN,
283     .impl = {
284         .min_access_size = 4,
285         .max_access_size = 4,
286     },
287 };
288 
289 static void ivshmem_vector_notify(void *opaque)
290 {
291     MSIVector *entry = opaque;
292     PCIDevice *pdev = entry->pdev;
293     IVShmemState *s = IVSHMEM_COMMON(pdev);
294     int vector = entry - s->msi_vectors;
295     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
296 
297     if (!event_notifier_test_and_clear(n)) {
298         return;
299     }
300 
301     IVSHMEM_DPRINTF("interrupt on vector %p %d\n", pdev, vector);
302     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
303         if (msix_enabled(pdev)) {
304             msix_notify(pdev, vector);
305         }
306     } else {
307         ivshmem_IntrStatus_write(s, 1);
308     }
309 }
310 
311 static int ivshmem_vector_unmask(PCIDevice *dev, unsigned vector,
312                                  MSIMessage msg)
313 {
314     IVShmemState *s = IVSHMEM_COMMON(dev);
315     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
316     MSIVector *v = &s->msi_vectors[vector];
317     int ret;
318 
319     IVSHMEM_DPRINTF("vector unmask %p %d\n", dev, vector);
320 
321     ret = kvm_irqchip_update_msi_route(kvm_state, v->virq, msg, dev);
322     if (ret < 0) {
323         return ret;
324     }
325     kvm_irqchip_commit_routes(kvm_state);
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     int ret;
446 
447     IVSHMEM_DPRINTF("ivshmem_add_kvm_msi_virq vector:%d\n", vector);
448     assert(!s->msi_vectors[vector].pdev);
449 
450     ret = kvm_irqchip_add_msi_route(kvm_state, vector, pdev);
451     if (ret < 0) {
452         error_setg(errp, "kvm_irqchip_add_msi_route failed");
453         return;
454     }
455 
456     s->msi_vectors[vector].virq = ret;
457     s->msi_vectors[vector].pdev = pdev;
458 }
459 
460 static void setup_interrupt(IVShmemState *s, int vector, Error **errp)
461 {
462     EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
463     bool with_irqfd = kvm_msi_via_irqfd_enabled() &&
464         ivshmem_has_feature(s, IVSHMEM_MSI);
465     PCIDevice *pdev = PCI_DEVICE(s);
466     Error *err = NULL;
467 
468     IVSHMEM_DPRINTF("setting up interrupt for vector: %d\n", vector);
469 
470     if (!with_irqfd) {
471         IVSHMEM_DPRINTF("with eventfd\n");
472         watch_vector_notifier(s, n, vector);
473     } else if (msix_enabled(pdev)) {
474         IVSHMEM_DPRINTF("with irqfd\n");
475         ivshmem_add_kvm_msi_virq(s, vector, &err);
476         if (err) {
477             error_propagate(errp, err);
478             return;
479         }
480 
481         if (!msix_is_masked(pdev, vector)) {
482             kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, n, NULL,
483                                                s->msi_vectors[vector].virq);
484             /* TODO handle error */
485         }
486     } else {
487         /* it will be delayed until msix is enabled, in write_config */
488         IVSHMEM_DPRINTF("with irqfd, delayed until msix enabled\n");
489     }
490 }
491 
492 static void process_msg_shmem(IVShmemState *s, int fd, Error **errp)
493 {
494     Error *local_err = NULL;
495     struct stat buf;
496     size_t size;
497 
498     if (s->ivshmem_bar2) {
499         error_setg(errp, "server sent unexpected shared memory message");
500         close(fd);
501         return;
502     }
503 
504     if (fstat(fd, &buf) < 0) {
505         error_setg_errno(errp, errno,
506             "can't determine size of shared memory sent by server");
507         close(fd);
508         return;
509     }
510 
511     size = buf.st_size;
512 
513     /* Legacy cruft */
514     if (s->legacy_size != SIZE_MAX) {
515         if (size < s->legacy_size) {
516             error_setg(errp, "server sent only %zd bytes of shared memory",
517                        (size_t)buf.st_size);
518             close(fd);
519             return;
520         }
521         size = s->legacy_size;
522     }
523 
524     /* mmap the region and map into the BAR2 */
525     memory_region_init_ram_from_fd(&s->server_bar2, OBJECT(s),
526                                    "ivshmem.bar2", size, true, fd, &local_err);
527     if (local_err) {
528         error_propagate(errp, local_err);
529         return;
530     }
531 
532     s->ivshmem_bar2 = &s->server_bar2;
533 }
534 
535 static void process_msg_disconnect(IVShmemState *s, uint16_t posn,
536                                    Error **errp)
537 {
538     IVSHMEM_DPRINTF("posn %d has gone away\n", posn);
539     if (posn >= s->nb_peers || posn == s->vm_id) {
540         error_setg(errp, "invalid peer %d", posn);
541         return;
542     }
543     close_peer_eventfds(s, posn);
544 }
545 
546 static void process_msg_connect(IVShmemState *s, uint16_t posn, int fd,
547                                 Error **errp)
548 {
549     Peer *peer = &s->peers[posn];
550     int vector;
551 
552     /*
553      * The N-th connect message for this peer comes with the file
554      * descriptor for vector N-1.  Count messages to find the vector.
555      */
556     if (peer->nb_eventfds >= s->vectors) {
557         error_setg(errp, "Too many eventfd received, device has %d vectors",
558                    s->vectors);
559         close(fd);
560         return;
561     }
562     vector = peer->nb_eventfds++;
563 
564     IVSHMEM_DPRINTF("eventfds[%d][%d] = %d\n", posn, vector, fd);
565     event_notifier_init_fd(&peer->eventfds[vector], fd);
566     fcntl_setfl(fd, O_NONBLOCK); /* msix/irqfd poll non block */
567 
568     if (posn == s->vm_id) {
569         setup_interrupt(s, vector, errp);
570         /* TODO do we need to handle the error? */
571     }
572 
573     if (ivshmem_has_feature(s, IVSHMEM_IOEVENTFD)) {
574         ivshmem_add_eventfd(s, posn, vector);
575     }
576 }
577 
578 static void process_msg(IVShmemState *s, int64_t msg, int fd, Error **errp)
579 {
580     IVSHMEM_DPRINTF("posn is %" PRId64 ", fd is %d\n", msg, fd);
581 
582     if (msg < -1 || msg > IVSHMEM_MAX_PEERS) {
583         error_setg(errp, "server sent invalid message %" PRId64, msg);
584         close(fd);
585         return;
586     }
587 
588     if (msg == -1) {
589         process_msg_shmem(s, fd, errp);
590         return;
591     }
592 
593     if (msg >= s->nb_peers) {
594         resize_peers(s, msg + 1);
595     }
596 
597     if (fd >= 0) {
598         process_msg_connect(s, msg, fd, errp);
599     } else {
600         process_msg_disconnect(s, msg, errp);
601     }
602 }
603 
604 static int ivshmem_can_receive(void *opaque)
605 {
606     IVShmemState *s = opaque;
607 
608     assert(s->msg_buffered_bytes < sizeof(s->msg_buf));
609     return sizeof(s->msg_buf) - s->msg_buffered_bytes;
610 }
611 
612 static void ivshmem_read(void *opaque, const uint8_t *buf, int size)
613 {
614     IVShmemState *s = opaque;
615     Error *err = NULL;
616     int fd;
617     int64_t msg;
618 
619     assert(size >= 0 && s->msg_buffered_bytes + size <= sizeof(s->msg_buf));
620     memcpy((unsigned char *)&s->msg_buf + s->msg_buffered_bytes, buf, size);
621     s->msg_buffered_bytes += size;
622     if (s->msg_buffered_bytes < sizeof(s->msg_buf)) {
623         return;
624     }
625     msg = le64_to_cpu(s->msg_buf);
626     s->msg_buffered_bytes = 0;
627 
628     fd = qemu_chr_fe_get_msgfd(&s->server_chr);
629 
630     process_msg(s, msg, fd, &err);
631     if (err) {
632         error_report_err(err);
633     }
634 }
635 
636 static int64_t ivshmem_recv_msg(IVShmemState *s, int *pfd, Error **errp)
637 {
638     int64_t msg;
639     int n, ret;
640 
641     n = 0;
642     do {
643         ret = qemu_chr_fe_read_all(&s->server_chr, (uint8_t *)&msg + n,
644                                    sizeof(msg) - n);
645         if (ret < 0) {
646             if (ret == -EINTR) {
647                 continue;
648             }
649             error_setg_errno(errp, -ret, "read from server failed");
650             return INT64_MIN;
651         }
652         n += ret;
653     } while (n < sizeof(msg));
654 
655     *pfd = qemu_chr_fe_get_msgfd(&s->server_chr);
656     return msg;
657 }
658 
659 static void ivshmem_recv_setup(IVShmemState *s, Error **errp)
660 {
661     Error *err = NULL;
662     int64_t msg;
663     int fd;
664 
665     msg = ivshmem_recv_msg(s, &fd, &err);
666     if (err) {
667         error_propagate(errp, err);
668         return;
669     }
670     if (msg != IVSHMEM_PROTOCOL_VERSION) {
671         error_setg(errp, "server sent version %" PRId64 ", expecting %d",
672                    msg, IVSHMEM_PROTOCOL_VERSION);
673         return;
674     }
675     if (fd != -1) {
676         error_setg(errp, "server sent invalid version message");
677         return;
678     }
679 
680     /*
681      * ivshmem-server sends the remaining initial messages in a fixed
682      * order, but the device has always accepted them in any order.
683      * Stay as compatible as practical, just in case people use
684      * servers that behave differently.
685      */
686 
687     /*
688      * ivshmem_device_spec.txt has always required the ID message
689      * right here, and ivshmem-server has always complied.  However,
690      * older versions of the device accepted it out of order, but
691      * broke when an interrupt setup message arrived before it.
692      */
693     msg = ivshmem_recv_msg(s, &fd, &err);
694     if (err) {
695         error_propagate(errp, err);
696         return;
697     }
698     if (fd != -1 || msg < 0 || msg > IVSHMEM_MAX_PEERS) {
699         error_setg(errp, "server sent invalid ID message");
700         return;
701     }
702     s->vm_id = msg;
703 
704     /*
705      * Receive more messages until we got shared memory.
706      */
707     do {
708         msg = ivshmem_recv_msg(s, &fd, &err);
709         if (err) {
710             error_propagate(errp, err);
711             return;
712         }
713         process_msg(s, msg, fd, &err);
714         if (err) {
715             error_propagate(errp, err);
716             return;
717         }
718     } while (msg != -1);
719 
720     /*
721      * This function must either map the shared memory or fail.  The
722      * loop above ensures that: it terminates normally only after it
723      * successfully processed the server's shared memory message.
724      * Assert that actually mapped the shared memory:
725      */
726     assert(s->ivshmem_bar2);
727 }
728 
729 /* Select the MSI-X vectors used by device.
730  * ivshmem maps events to vectors statically, so
731  * we just enable all vectors on init and after reset. */
732 static void ivshmem_msix_vector_use(IVShmemState *s)
733 {
734     PCIDevice *d = PCI_DEVICE(s);
735     int i;
736 
737     for (i = 0; i < s->vectors; i++) {
738         msix_vector_use(d, i);
739     }
740 }
741 
742 static void ivshmem_reset(DeviceState *d)
743 {
744     IVShmemState *s = IVSHMEM_COMMON(d);
745 
746     s->intrstatus = 0;
747     s->intrmask = 0;
748     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
749         ivshmem_msix_vector_use(s);
750     }
751 }
752 
753 static int ivshmem_setup_interrupts(IVShmemState *s, Error **errp)
754 {
755     /* allocate QEMU callback data for receiving interrupts */
756     s->msi_vectors = g_malloc0(s->vectors * sizeof(MSIVector));
757 
758     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
759         if (msix_init_exclusive_bar(PCI_DEVICE(s), s->vectors, 1, errp)) {
760             return -1;
761         }
762 
763         IVSHMEM_DPRINTF("msix initialized (%d vectors)\n", s->vectors);
764         ivshmem_msix_vector_use(s);
765     }
766 
767     return 0;
768 }
769 
770 static void ivshmem_enable_irqfd(IVShmemState *s)
771 {
772     PCIDevice *pdev = PCI_DEVICE(s);
773     int i;
774 
775     for (i = 0; i < s->peers[s->vm_id].nb_eventfds; i++) {
776         Error *err = NULL;
777 
778         ivshmem_add_kvm_msi_virq(s, i, &err);
779         if (err) {
780             error_report_err(err);
781             /* TODO do we need to handle the error? */
782         }
783     }
784 
785     if (msix_set_vector_notifiers(pdev,
786                                   ivshmem_vector_unmask,
787                                   ivshmem_vector_mask,
788                                   ivshmem_vector_poll)) {
789         error_report("ivshmem: msix_set_vector_notifiers failed");
790     }
791 }
792 
793 static void ivshmem_remove_kvm_msi_virq(IVShmemState *s, int vector)
794 {
795     IVSHMEM_DPRINTF("ivshmem_remove_kvm_msi_virq vector:%d\n", vector);
796 
797     if (s->msi_vectors[vector].pdev == NULL) {
798         return;
799     }
800 
801     /* it was cleaned when masked in the frontend. */
802     kvm_irqchip_release_virq(kvm_state, s->msi_vectors[vector].virq);
803 
804     s->msi_vectors[vector].pdev = NULL;
805 }
806 
807 static void ivshmem_disable_irqfd(IVShmemState *s)
808 {
809     PCIDevice *pdev = PCI_DEVICE(s);
810     int i;
811 
812     for (i = 0; i < s->peers[s->vm_id].nb_eventfds; i++) {
813         ivshmem_remove_kvm_msi_virq(s, i);
814     }
815 
816     msix_unset_vector_notifiers(pdev);
817 }
818 
819 static void ivshmem_write_config(PCIDevice *pdev, uint32_t address,
820                                  uint32_t val, int len)
821 {
822     IVShmemState *s = IVSHMEM_COMMON(pdev);
823     int is_enabled, was_enabled = msix_enabled(pdev);
824 
825     pci_default_write_config(pdev, address, val, len);
826     is_enabled = msix_enabled(pdev);
827 
828     if (kvm_msi_via_irqfd_enabled()) {
829         if (!was_enabled && is_enabled) {
830             ivshmem_enable_irqfd(s);
831         } else if (was_enabled && !is_enabled) {
832             ivshmem_disable_irqfd(s);
833         }
834     }
835 }
836 
837 static void ivshmem_common_realize(PCIDevice *dev, Error **errp)
838 {
839     IVShmemState *s = IVSHMEM_COMMON(dev);
840     Error *err = NULL;
841     uint8_t *pci_conf;
842     uint8_t attr = PCI_BASE_ADDRESS_SPACE_MEMORY |
843         PCI_BASE_ADDRESS_MEM_PREFETCH;
844     Error *local_err = NULL;
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         Chardev *chr = qemu_chr_fe_get_driver(&s->server_chr);
874         assert(chr);
875 
876         IVSHMEM_DPRINTF("using shared memory server (socket = %s)\n",
877                         chr->filename);
878 
879         /* we allocate enough space for 16 peers and grow as needed */
880         resize_peers(s, 16);
881 
882         /*
883          * Receive setup messages from server synchronously.
884          * Older versions did it asynchronously, but that creates a
885          * number of entertaining race conditions.
886          */
887         ivshmem_recv_setup(s, &err);
888         if (err) {
889             error_propagate(errp, err);
890             return;
891         }
892 
893         if (s->master == ON_OFF_AUTO_ON && s->vm_id != 0) {
894             error_setg(errp,
895                        "master must connect to the server before any peers");
896             return;
897         }
898 
899         qemu_chr_fe_set_handlers(&s->server_chr, ivshmem_can_receive,
900                                  ivshmem_read, NULL, NULL, s, NULL, true);
901 
902         if (ivshmem_setup_interrupts(s, errp) < 0) {
903             error_prepend(errp, "Failed to initialize interrupts: ");
904             return;
905         }
906     }
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, &local_err);
916         if (local_err) {
917             error_propagate(errp, local_err);
918             error_free(s->migration_blocker);
919             return;
920         }
921     }
922 
923     vmstate_register_ram(s->ivshmem_bar2, DEVICE(s));
924     pci_register_bar(PCI_DEVICE(s), 2, attr, s->ivshmem_bar2);
925 }
926 
927 static void ivshmem_exit(PCIDevice *dev)
928 {
929     IVShmemState *s = IVSHMEM_COMMON(dev);
930     int i;
931 
932     if (s->migration_blocker) {
933         migrate_del_blocker(s->migration_blocker);
934         error_free(s->migration_blocker);
935     }
936 
937     if (memory_region_is_mapped(s->ivshmem_bar2)) {
938         if (!s->hostmem) {
939             void *addr = memory_region_get_ram_ptr(s->ivshmem_bar2);
940             int fd;
941 
942             if (munmap(addr, memory_region_size(s->ivshmem_bar2) == -1)) {
943                 error_report("Failed to munmap shared memory %s",
944                              strerror(errno));
945             }
946 
947             fd = memory_region_get_fd(s->ivshmem_bar2);
948             close(fd);
949         }
950 
951         vmstate_unregister_ram(s->ivshmem_bar2, DEVICE(dev));
952     }
953 
954     if (s->peers) {
955         for (i = 0; i < s->nb_peers; i++) {
956             close_peer_eventfds(s, i);
957         }
958         g_free(s->peers);
959     }
960 
961     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
962         msix_uninit_exclusive_bar(dev);
963     }
964 
965     g_free(s->msi_vectors);
966 }
967 
968 static int ivshmem_pre_load(void *opaque)
969 {
970     IVShmemState *s = opaque;
971 
972     if (!ivshmem_is_master(s)) {
973         error_report("'peer' devices are not migratable");
974         return -EINVAL;
975     }
976 
977     return 0;
978 }
979 
980 static int ivshmem_post_load(void *opaque, int version_id)
981 {
982     IVShmemState *s = opaque;
983 
984     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
985         ivshmem_msix_vector_use(s);
986     }
987     return 0;
988 }
989 
990 static void ivshmem_common_class_init(ObjectClass *klass, void *data)
991 {
992     DeviceClass *dc = DEVICE_CLASS(klass);
993     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
994 
995     k->realize = ivshmem_common_realize;
996     k->exit = ivshmem_exit;
997     k->config_write = ivshmem_write_config;
998     k->vendor_id = PCI_VENDOR_ID_IVSHMEM;
999     k->device_id = PCI_DEVICE_ID_IVSHMEM;
1000     k->class_id = PCI_CLASS_MEMORY_RAM;
1001     k->revision = 1;
1002     dc->reset = ivshmem_reset;
1003     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
1004     dc->desc = "Inter-VM shared memory";
1005 }
1006 
1007 static const TypeInfo ivshmem_common_info = {
1008     .name          = TYPE_IVSHMEM_COMMON,
1009     .parent        = TYPE_PCI_DEVICE,
1010     .instance_size = sizeof(IVShmemState),
1011     .abstract      = true,
1012     .class_init    = ivshmem_common_class_init,
1013 };
1014 
1015 static const VMStateDescription ivshmem_plain_vmsd = {
1016     .name = TYPE_IVSHMEM_PLAIN,
1017     .version_id = 0,
1018     .minimum_version_id = 0,
1019     .pre_load = ivshmem_pre_load,
1020     .post_load = ivshmem_post_load,
1021     .fields = (VMStateField[]) {
1022         VMSTATE_PCI_DEVICE(parent_obj, IVShmemState),
1023         VMSTATE_UINT32(intrstatus, IVShmemState),
1024         VMSTATE_UINT32(intrmask, IVShmemState),
1025         VMSTATE_END_OF_LIST()
1026     },
1027 };
1028 
1029 static Property ivshmem_plain_properties[] = {
1030     DEFINE_PROP_ON_OFF_AUTO("master", IVShmemState, master, ON_OFF_AUTO_OFF),
1031     DEFINE_PROP_LINK("memdev", IVShmemState, hostmem, TYPE_MEMORY_BACKEND,
1032                      HostMemoryBackend *),
1033     DEFINE_PROP_END_OF_LIST(),
1034 };
1035 
1036 static void ivshmem_plain_init(Object *obj)
1037 {
1038     IVShmemState *s = IVSHMEM_PLAIN(obj);
1039 
1040     s->not_legacy_32bit = 1;
1041 }
1042 
1043 static void ivshmem_plain_realize(PCIDevice *dev, Error **errp)
1044 {
1045     IVShmemState *s = IVSHMEM_COMMON(dev);
1046 
1047     if (!s->hostmem) {
1048         error_setg(errp, "You must specify a 'memdev'");
1049         return;
1050     } else if (host_memory_backend_is_mapped(s->hostmem)) {
1051         char *path = object_get_canonical_path_component(OBJECT(s->hostmem));
1052         error_setg(errp, "can't use already busy memdev: %s", path);
1053         g_free(path);
1054         return;
1055     }
1056 
1057     ivshmem_common_realize(dev, errp);
1058     host_memory_backend_set_mapped(s->hostmem, true);
1059 }
1060 
1061 static void ivshmem_plain_exit(PCIDevice *pci_dev)
1062 {
1063     IVShmemState *s = IVSHMEM_COMMON(pci_dev);
1064 
1065     host_memory_backend_set_mapped(s->hostmem, false);
1066 }
1067 
1068 static void ivshmem_plain_class_init(ObjectClass *klass, void *data)
1069 {
1070     DeviceClass *dc = DEVICE_CLASS(klass);
1071     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
1072 
1073     k->realize = ivshmem_plain_realize;
1074     k->exit = ivshmem_plain_exit;
1075     dc->props = ivshmem_plain_properties;
1076     dc->vmsd = &ivshmem_plain_vmsd;
1077 }
1078 
1079 static const TypeInfo ivshmem_plain_info = {
1080     .name          = TYPE_IVSHMEM_PLAIN,
1081     .parent        = TYPE_IVSHMEM_COMMON,
1082     .instance_size = sizeof(IVShmemState),
1083     .instance_init = ivshmem_plain_init,
1084     .class_init    = ivshmem_plain_class_init,
1085 };
1086 
1087 static const VMStateDescription ivshmem_doorbell_vmsd = {
1088     .name = TYPE_IVSHMEM_DOORBELL,
1089     .version_id = 0,
1090     .minimum_version_id = 0,
1091     .pre_load = ivshmem_pre_load,
1092     .post_load = ivshmem_post_load,
1093     .fields = (VMStateField[]) {
1094         VMSTATE_PCI_DEVICE(parent_obj, IVShmemState),
1095         VMSTATE_MSIX(parent_obj, IVShmemState),
1096         VMSTATE_UINT32(intrstatus, IVShmemState),
1097         VMSTATE_UINT32(intrmask, IVShmemState),
1098         VMSTATE_END_OF_LIST()
1099     },
1100 };
1101 
1102 static Property ivshmem_doorbell_properties[] = {
1103     DEFINE_PROP_CHR("chardev", IVShmemState, server_chr),
1104     DEFINE_PROP_UINT32("vectors", IVShmemState, vectors, 1),
1105     DEFINE_PROP_BIT("ioeventfd", IVShmemState, features, IVSHMEM_IOEVENTFD,
1106                     true),
1107     DEFINE_PROP_ON_OFF_AUTO("master", IVShmemState, master, ON_OFF_AUTO_OFF),
1108     DEFINE_PROP_END_OF_LIST(),
1109 };
1110 
1111 static void ivshmem_doorbell_init(Object *obj)
1112 {
1113     IVShmemState *s = IVSHMEM_DOORBELL(obj);
1114 
1115     s->features |= (1 << IVSHMEM_MSI);
1116     s->legacy_size = SIZE_MAX;  /* whatever the server sends */
1117     s->not_legacy_32bit = 1;
1118 }
1119 
1120 static void ivshmem_doorbell_realize(PCIDevice *dev, Error **errp)
1121 {
1122     IVShmemState *s = IVSHMEM_COMMON(dev);
1123 
1124     if (!qemu_chr_fe_backend_connected(&s->server_chr)) {
1125         error_setg(errp, "You must specify a 'chardev'");
1126         return;
1127     }
1128 
1129     ivshmem_common_realize(dev, errp);
1130 }
1131 
1132 static void ivshmem_doorbell_class_init(ObjectClass *klass, void *data)
1133 {
1134     DeviceClass *dc = DEVICE_CLASS(klass);
1135     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
1136 
1137     k->realize = ivshmem_doorbell_realize;
1138     dc->props = ivshmem_doorbell_properties;
1139     dc->vmsd = &ivshmem_doorbell_vmsd;
1140 }
1141 
1142 static const TypeInfo ivshmem_doorbell_info = {
1143     .name          = TYPE_IVSHMEM_DOORBELL,
1144     .parent        = TYPE_IVSHMEM_COMMON,
1145     .instance_size = sizeof(IVShmemState),
1146     .instance_init = ivshmem_doorbell_init,
1147     .class_init    = ivshmem_doorbell_class_init,
1148 };
1149 
1150 static int ivshmem_load_old(QEMUFile *f, void *opaque, int version_id)
1151 {
1152     IVShmemState *s = opaque;
1153     PCIDevice *pdev = PCI_DEVICE(s);
1154     int ret;
1155 
1156     IVSHMEM_DPRINTF("ivshmem_load_old\n");
1157 
1158     if (version_id != 0) {
1159         return -EINVAL;
1160     }
1161 
1162     ret = ivshmem_pre_load(s);
1163     if (ret) {
1164         return ret;
1165     }
1166 
1167     ret = pci_device_load(pdev, f);
1168     if (ret) {
1169         return ret;
1170     }
1171 
1172     if (ivshmem_has_feature(s, IVSHMEM_MSI)) {
1173         msix_load(pdev, f);
1174         ivshmem_msix_vector_use(s);
1175     } else {
1176         s->intrstatus = qemu_get_be32(f);
1177         s->intrmask = qemu_get_be32(f);
1178     }
1179 
1180     return 0;
1181 }
1182 
1183 static bool test_msix(void *opaque, int version_id)
1184 {
1185     IVShmemState *s = opaque;
1186 
1187     return ivshmem_has_feature(s, IVSHMEM_MSI);
1188 }
1189 
1190 static bool test_no_msix(void *opaque, int version_id)
1191 {
1192     return !test_msix(opaque, version_id);
1193 }
1194 
1195 static const VMStateDescription ivshmem_vmsd = {
1196     .name = "ivshmem",
1197     .version_id = 1,
1198     .minimum_version_id = 1,
1199     .pre_load = ivshmem_pre_load,
1200     .post_load = ivshmem_post_load,
1201     .fields = (VMStateField[]) {
1202         VMSTATE_PCI_DEVICE(parent_obj, IVShmemState),
1203 
1204         VMSTATE_MSIX_TEST(parent_obj, IVShmemState, test_msix),
1205         VMSTATE_UINT32_TEST(intrstatus, IVShmemState, test_no_msix),
1206         VMSTATE_UINT32_TEST(intrmask, IVShmemState, test_no_msix),
1207 
1208         VMSTATE_END_OF_LIST()
1209     },
1210     .load_state_old = ivshmem_load_old,
1211     .minimum_version_id_old = 0
1212 };
1213 
1214 static Property ivshmem_properties[] = {
1215     DEFINE_PROP_CHR("chardev", IVShmemState, server_chr),
1216     DEFINE_PROP_STRING("size", IVShmemState, sizearg),
1217     DEFINE_PROP_UINT32("vectors", IVShmemState, vectors, 1),
1218     DEFINE_PROP_BIT("ioeventfd", IVShmemState, features, IVSHMEM_IOEVENTFD,
1219                     false),
1220     DEFINE_PROP_BIT("msi", IVShmemState, features, IVSHMEM_MSI, true),
1221     DEFINE_PROP_STRING("shm", IVShmemState, shmobj),
1222     DEFINE_PROP_STRING("role", IVShmemState, role),
1223     DEFINE_PROP_UINT32("use64", IVShmemState, not_legacy_32bit, 1),
1224     DEFINE_PROP_END_OF_LIST(),
1225 };
1226 
1227 static void desugar_shm(IVShmemState *s)
1228 {
1229     Object *obj;
1230     char *path;
1231 
1232     obj = object_new("memory-backend-file");
1233     path = g_strdup_printf("/dev/shm/%s", s->shmobj);
1234     object_property_set_str(obj, path, "mem-path", &error_abort);
1235     g_free(path);
1236     object_property_set_int(obj, s->legacy_size, "size", &error_abort);
1237     object_property_set_bool(obj, true, "share", &error_abort);
1238     object_property_add_child(OBJECT(s), "internal-shm-backend", obj,
1239                               &error_abort);
1240     user_creatable_complete(obj, &error_abort);
1241     s->hostmem = MEMORY_BACKEND(obj);
1242 }
1243 
1244 static void ivshmem_realize(PCIDevice *dev, Error **errp)
1245 {
1246     IVShmemState *s = IVSHMEM_COMMON(dev);
1247 
1248     if (!qtest_enabled()) {
1249         error_report("ivshmem is deprecated, please use ivshmem-plain"
1250                      " or ivshmem-doorbell instead");
1251     }
1252 
1253     if (qemu_chr_fe_backend_connected(&s->server_chr) + !!s->shmobj != 1) {
1254         error_setg(errp, "You must specify either 'shm' or 'chardev'");
1255         return;
1256     }
1257 
1258     if (s->sizearg == NULL) {
1259         s->legacy_size = 4 << 20; /* 4 MB default */
1260     } else {
1261         int ret;
1262         uint64_t size;
1263 
1264         ret = qemu_strtosz_MiB(s->sizearg, NULL, &size);
1265         if (ret < 0 || (size_t)size != size || !is_power_of_2(size)) {
1266             error_setg(errp, "Invalid size %s", s->sizearg);
1267             return;
1268         }
1269         s->legacy_size = size;
1270     }
1271 
1272     /* check that role is reasonable */
1273     if (s->role) {
1274         if (strncmp(s->role, "peer", 5) == 0) {
1275             s->master = ON_OFF_AUTO_OFF;
1276         } else if (strncmp(s->role, "master", 7) == 0) {
1277             s->master = ON_OFF_AUTO_ON;
1278         } else {
1279             error_setg(errp, "'role' must be 'peer' or 'master'");
1280             return;
1281         }
1282     } else {
1283         s->master = ON_OFF_AUTO_AUTO;
1284     }
1285 
1286     if (s->shmobj) {
1287         desugar_shm(s);
1288     }
1289 
1290     /*
1291      * Note: we don't use INTx with IVSHMEM_MSI at all, so this is a
1292      * bald-faced lie then.  But it's a backwards compatible lie.
1293      */
1294     pci_config_set_interrupt_pin(dev->config, 1);
1295 
1296     ivshmem_common_realize(dev, errp);
1297 }
1298 
1299 static void ivshmem_class_init(ObjectClass *klass, void *data)
1300 {
1301     DeviceClass *dc = DEVICE_CLASS(klass);
1302     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
1303 
1304     k->realize = ivshmem_realize;
1305     k->revision = 0;
1306     dc->desc = "Inter-VM shared memory (legacy)";
1307     dc->props = ivshmem_properties;
1308     dc->vmsd = &ivshmem_vmsd;
1309 }
1310 
1311 static const TypeInfo ivshmem_info = {
1312     .name          = TYPE_IVSHMEM,
1313     .parent        = TYPE_IVSHMEM_COMMON,
1314     .instance_size = sizeof(IVShmemState),
1315     .class_init    = ivshmem_class_init,
1316 };
1317 
1318 static void ivshmem_register_types(void)
1319 {
1320     type_register_static(&ivshmem_common_info);
1321     type_register_static(&ivshmem_plain_info);
1322     type_register_static(&ivshmem_doorbell_info);
1323     type_register_static(&ivshmem_info);
1324 }
1325 
1326 type_init(ivshmem_register_types)
1327