vdpa: Add virtio-net mac address via CVQ at start
[qemu.git] / net / colo-compare.c
blob787c740f1425a5c16d5e72160bf24b5942e2a515
1 /*
2 * COarse-grain LOck-stepping Virtual Machines for Non-stop Service (COLO)
3 * (a.k.a. Fault Tolerance or Continuous Replication)
5 * Copyright (c) 2016 HUAWEI TECHNOLOGIES CO., LTD.
6 * Copyright (c) 2016 FUJITSU LIMITED
7 * Copyright (c) 2016 Intel Corporation
9 * Author: Zhang Chen <zhangchen.fnst@cn.fujitsu.com>
11 * This work is licensed under the terms of the GNU GPL, version 2 or
12 * later. See the COPYING file in the top-level directory.
15 #include "qemu/osdep.h"
16 #include "qemu/error-report.h"
17 #include "trace.h"
18 #include "qapi/error.h"
19 #include "net/net.h"
20 #include "net/eth.h"
21 #include "qom/object_interfaces.h"
22 #include "qemu/iov.h"
23 #include "qom/object.h"
24 #include "net/queue.h"
25 #include "chardev/char-fe.h"
26 #include "qemu/sockets.h"
27 #include "colo.h"
28 #include "sysemu/iothread.h"
29 #include "net/colo-compare.h"
30 #include "migration/colo.h"
31 #include "migration/migration.h"
32 #include "util.h"
34 #include "block/aio-wait.h"
35 #include "qemu/coroutine.h"
37 #define TYPE_COLO_COMPARE "colo-compare"
38 typedef struct CompareState CompareState;
39 DECLARE_INSTANCE_CHECKER(CompareState, COLO_COMPARE,
40 TYPE_COLO_COMPARE)
42 static QTAILQ_HEAD(, CompareState) net_compares =
43 QTAILQ_HEAD_INITIALIZER(net_compares);
45 static NotifierList colo_compare_notifiers =
46 NOTIFIER_LIST_INITIALIZER(colo_compare_notifiers);
48 #define COMPARE_READ_LEN_MAX NET_BUFSIZE
49 #define MAX_QUEUE_SIZE 1024
51 #define COLO_COMPARE_FREE_PRIMARY 0x01
52 #define COLO_COMPARE_FREE_SECONDARY 0x02
54 #define REGULAR_PACKET_CHECK_MS 1000
55 #define DEFAULT_TIME_OUT_MS 3000
57 /* #define DEBUG_COLO_PACKETS */
59 static QemuMutex colo_compare_mutex;
60 static bool colo_compare_active;
61 static QemuMutex event_mtx;
62 static QemuCond event_complete_cond;
63 static int event_unhandled_count;
64 static uint32_t max_queue_size;
67 * + CompareState ++
68 * | |
69 * +---------------+ +---------------+ +---------------+
70 * | conn list + - > conn + ------- > conn + -- > ......
71 * +---------------+ +---------------+ +---------------+
72 * | | | | | |
73 * +---------------+ +---v----+ +---v----+ +---v----+ +---v----+
74 * |primary | |secondary |primary | |secondary
75 * |packet | |packet + |packet | |packet +
76 * +--------+ +--------+ +--------+ +--------+
77 * | | | |
78 * +---v----+ +---v----+ +---v----+ +---v----+
79 * |primary | |secondary |primary | |secondary
80 * |packet | |packet + |packet | |packet +
81 * +--------+ +--------+ +--------+ +--------+
82 * | | | |
83 * +---v----+ +---v----+ +---v----+ +---v----+
84 * |primary | |secondary |primary | |secondary
85 * |packet | |packet + |packet | |packet +
86 * +--------+ +--------+ +--------+ +--------+
89 typedef struct SendCo {
90 Coroutine *co;
91 struct CompareState *s;
92 CharBackend *chr;
93 GQueue send_list;
94 bool notify_remote_frame;
95 bool done;
96 int ret;
97 } SendCo;
99 typedef struct SendEntry {
100 uint32_t size;
101 uint32_t vnet_hdr_len;
102 uint8_t *buf;
103 } SendEntry;
105 struct CompareState {
106 Object parent;
108 char *pri_indev;
109 char *sec_indev;
110 char *outdev;
111 char *notify_dev;
112 CharBackend chr_pri_in;
113 CharBackend chr_sec_in;
114 CharBackend chr_out;
115 CharBackend chr_notify_dev;
116 SocketReadState pri_rs;
117 SocketReadState sec_rs;
118 SocketReadState notify_rs;
119 SendCo out_sendco;
120 SendCo notify_sendco;
121 bool vnet_hdr;
122 uint64_t compare_timeout;
123 uint32_t expired_scan_cycle;
126 * Record the connection that through the NIC
127 * Element type: Connection
129 GQueue conn_list;
130 /* Record the connection without repetition */
131 GHashTable *connection_track_table;
133 IOThread *iothread;
134 GMainContext *worker_context;
135 QEMUTimer *packet_check_timer;
137 QEMUBH *event_bh;
138 enum colo_event event;
140 QTAILQ_ENTRY(CompareState) next;
143 typedef struct CompareClass {
144 ObjectClass parent_class;
145 } CompareClass;
147 enum {
148 PRIMARY_IN = 0,
149 SECONDARY_IN,
152 static const char *colo_mode[] = {
153 [PRIMARY_IN] = "primary",
154 [SECONDARY_IN] = "secondary",
157 static int compare_chr_send(CompareState *s,
158 uint8_t *buf,
159 uint32_t size,
160 uint32_t vnet_hdr_len,
161 bool notify_remote_frame,
162 bool zero_copy);
164 static bool packet_matches_str(const char *str,
165 const uint8_t *buf,
166 uint32_t packet_len)
168 if (packet_len != strlen(str)) {
169 return false;
172 return !memcmp(str, buf, packet_len);
175 static void notify_remote_frame(CompareState *s)
177 char msg[] = "DO_CHECKPOINT";
178 int ret = 0;
180 ret = compare_chr_send(s, (uint8_t *)msg, strlen(msg), 0, true, false);
181 if (ret < 0) {
182 error_report("Notify Xen COLO-frame failed");
186 static void colo_compare_inconsistency_notify(CompareState *s)
188 if (s->notify_dev) {
189 notify_remote_frame(s);
190 } else {
191 notifier_list_notify(&colo_compare_notifiers,
192 migrate_get_current());
196 /* Use restricted to colo_insert_packet() */
197 static gint seq_sorter(Packet *a, Packet *b, gpointer data)
199 return b->tcp_seq - a->tcp_seq;
202 static void fill_pkt_tcp_info(void *data, uint32_t *max_ack)
204 Packet *pkt = data;
205 struct tcp_hdr *tcphd;
207 tcphd = (struct tcp_hdr *)pkt->transport_header;
209 pkt->tcp_seq = ntohl(tcphd->th_seq);
210 pkt->tcp_ack = ntohl(tcphd->th_ack);
211 /* Need to consider ACK will bigger than uint32_t MAX */
212 *max_ack = pkt->tcp_ack - *max_ack > 0 ? pkt->tcp_ack : *max_ack;
213 pkt->header_size = pkt->transport_header - (uint8_t *)pkt->data
214 + (tcphd->th_off << 2);
215 pkt->payload_size = pkt->size - pkt->header_size;
216 pkt->seq_end = pkt->tcp_seq + pkt->payload_size;
217 pkt->flags = tcphd->th_flags;
221 * Return 1 on success, if return 0 means the
222 * packet will be dropped
224 static int colo_insert_packet(GQueue *queue, Packet *pkt, uint32_t *max_ack)
226 if (g_queue_get_length(queue) <= max_queue_size) {
227 if (pkt->ip->ip_p == IPPROTO_TCP) {
228 fill_pkt_tcp_info(pkt, max_ack);
229 g_queue_insert_sorted(queue,
230 pkt,
231 (GCompareDataFunc)seq_sorter,
232 NULL);
233 } else {
234 g_queue_push_tail(queue, pkt);
236 return 1;
238 return 0;
242 * Return 0 on success, if return -1 means the pkt
243 * is unsupported(arp and ipv6) and will be sent later
245 static int packet_enqueue(CompareState *s, int mode, Connection **con)
247 ConnectionKey key;
248 Packet *pkt = NULL;
249 Connection *conn;
250 int ret;
252 if (mode == PRIMARY_IN) {
253 pkt = packet_new(s->pri_rs.buf,
254 s->pri_rs.packet_len,
255 s->pri_rs.vnet_hdr_len);
256 } else {
257 pkt = packet_new(s->sec_rs.buf,
258 s->sec_rs.packet_len,
259 s->sec_rs.vnet_hdr_len);
262 if (parse_packet_early(pkt)) {
263 packet_destroy(pkt, NULL);
264 pkt = NULL;
265 return -1;
267 fill_connection_key(pkt, &key, false);
269 conn = connection_get(s->connection_track_table,
270 &key,
271 &s->conn_list);
273 if (!conn->processing) {
274 g_queue_push_tail(&s->conn_list, conn);
275 conn->processing = true;
278 if (mode == PRIMARY_IN) {
279 ret = colo_insert_packet(&conn->primary_list, pkt, &conn->pack);
280 } else {
281 ret = colo_insert_packet(&conn->secondary_list, pkt, &conn->sack);
284 if (!ret) {
285 trace_colo_compare_drop_packet(colo_mode[mode],
286 "queue size too big, drop packet");
287 packet_destroy(pkt, NULL);
288 pkt = NULL;
291 *con = conn;
293 return 0;
296 static inline bool after(uint32_t seq1, uint32_t seq2)
298 return (int32_t)(seq1 - seq2) > 0;
301 static void colo_release_primary_pkt(CompareState *s, Packet *pkt)
303 int ret;
304 ret = compare_chr_send(s,
305 pkt->data,
306 pkt->size,
307 pkt->vnet_hdr_len,
308 false,
309 true);
310 if (ret < 0) {
311 error_report("colo send primary packet failed");
313 trace_colo_compare_main("packet same and release packet");
314 packet_destroy_partial(pkt, NULL);
318 * The IP packets sent by primary and secondary
319 * will be compared in here
320 * TODO support ip fragment, Out-Of-Order
321 * return: 0 means packet same
322 * > 0 || < 0 means packet different
324 static int colo_compare_packet_payload(Packet *ppkt,
325 Packet *spkt,
326 uint16_t poffset,
327 uint16_t soffset,
328 uint16_t len)
331 if (trace_event_get_state_backends(TRACE_COLO_COMPARE_IP_INFO)) {
332 char pri_ip_src[20], pri_ip_dst[20], sec_ip_src[20], sec_ip_dst[20];
334 strcpy(pri_ip_src, inet_ntoa(ppkt->ip->ip_src));
335 strcpy(pri_ip_dst, inet_ntoa(ppkt->ip->ip_dst));
336 strcpy(sec_ip_src, inet_ntoa(spkt->ip->ip_src));
337 strcpy(sec_ip_dst, inet_ntoa(spkt->ip->ip_dst));
339 trace_colo_compare_ip_info(ppkt->size, pri_ip_src,
340 pri_ip_dst, spkt->size,
341 sec_ip_src, sec_ip_dst);
344 return memcmp(ppkt->data + poffset, spkt->data + soffset, len);
348 * return true means that the payload is consist and
349 * need to make the next comparison, false means do
350 * the checkpoint
352 static bool colo_mark_tcp_pkt(Packet *ppkt, Packet *spkt,
353 int8_t *mark, uint32_t max_ack)
355 *mark = 0;
357 if (ppkt->tcp_seq == spkt->tcp_seq && ppkt->seq_end == spkt->seq_end) {
358 if (!colo_compare_packet_payload(ppkt, spkt,
359 ppkt->header_size, spkt->header_size,
360 ppkt->payload_size)) {
361 *mark = COLO_COMPARE_FREE_SECONDARY | COLO_COMPARE_FREE_PRIMARY;
362 return true;
366 /* one part of secondary packet payload still need to be compared */
367 if (!after(ppkt->seq_end, spkt->seq_end)) {
368 if (!colo_compare_packet_payload(ppkt, spkt,
369 ppkt->header_size + ppkt->offset,
370 spkt->header_size + spkt->offset,
371 ppkt->payload_size - ppkt->offset)) {
372 if (!after(ppkt->tcp_ack, max_ack)) {
373 *mark = COLO_COMPARE_FREE_PRIMARY;
374 spkt->offset += ppkt->payload_size - ppkt->offset;
375 return true;
376 } else {
377 /* secondary guest hasn't ack the data, don't send
378 * out this packet
380 return false;
383 } else {
384 /* primary packet is longer than secondary packet, compare
385 * the same part and mark the primary packet offset
387 if (!colo_compare_packet_payload(ppkt, spkt,
388 ppkt->header_size + ppkt->offset,
389 spkt->header_size + spkt->offset,
390 spkt->payload_size - spkt->offset)) {
391 *mark = COLO_COMPARE_FREE_SECONDARY;
392 ppkt->offset += spkt->payload_size - spkt->offset;
393 return true;
397 return false;
400 static void colo_compare_tcp(CompareState *s, Connection *conn)
402 Packet *ppkt = NULL, *spkt = NULL;
403 int8_t mark;
406 * If ppkt and spkt have the same payload, but ppkt's ACK
407 * is greater than spkt's ACK, in this case we can not
408 * send the ppkt because it will cause the secondary guest
409 * to miss sending some data in the next. Therefore, we
410 * record the maximum ACK in the current queue at both
411 * primary side and secondary side. Only when the ack is
412 * less than the smaller of the two maximum ack, then we
413 * can ensure that the packet's payload is acknowledged by
414 * primary and secondary.
416 uint32_t min_ack = conn->pack - conn->sack > 0 ?
417 conn->sack : conn->pack;
419 pri:
420 if (g_queue_is_empty(&conn->primary_list)) {
421 return;
423 ppkt = g_queue_pop_tail(&conn->primary_list);
424 sec:
425 if (g_queue_is_empty(&conn->secondary_list)) {
426 g_queue_push_tail(&conn->primary_list, ppkt);
427 return;
429 spkt = g_queue_pop_tail(&conn->secondary_list);
431 if (ppkt->tcp_seq == ppkt->seq_end) {
432 colo_release_primary_pkt(s, ppkt);
433 ppkt = NULL;
436 if (ppkt && conn->compare_seq && !after(ppkt->seq_end, conn->compare_seq)) {
437 trace_colo_compare_main("pri: this packet has compared");
438 colo_release_primary_pkt(s, ppkt);
439 ppkt = NULL;
442 if (spkt->tcp_seq == spkt->seq_end) {
443 packet_destroy(spkt, NULL);
444 if (!ppkt) {
445 goto pri;
446 } else {
447 goto sec;
449 } else {
450 if (conn->compare_seq && !after(spkt->seq_end, conn->compare_seq)) {
451 trace_colo_compare_main("sec: this packet has compared");
452 packet_destroy(spkt, NULL);
453 if (!ppkt) {
454 goto pri;
455 } else {
456 goto sec;
459 if (!ppkt) {
460 g_queue_push_tail(&conn->secondary_list, spkt);
461 goto pri;
465 if (colo_mark_tcp_pkt(ppkt, spkt, &mark, min_ack)) {
466 trace_colo_compare_tcp_info("pri",
467 ppkt->tcp_seq, ppkt->tcp_ack,
468 ppkt->header_size, ppkt->payload_size,
469 ppkt->offset, ppkt->flags);
471 trace_colo_compare_tcp_info("sec",
472 spkt->tcp_seq, spkt->tcp_ack,
473 spkt->header_size, spkt->payload_size,
474 spkt->offset, spkt->flags);
476 if (mark == COLO_COMPARE_FREE_PRIMARY) {
477 conn->compare_seq = ppkt->seq_end;
478 colo_release_primary_pkt(s, ppkt);
479 g_queue_push_tail(&conn->secondary_list, spkt);
480 goto pri;
481 } else if (mark == COLO_COMPARE_FREE_SECONDARY) {
482 conn->compare_seq = spkt->seq_end;
483 packet_destroy(spkt, NULL);
484 goto sec;
485 } else if (mark == (COLO_COMPARE_FREE_PRIMARY | COLO_COMPARE_FREE_SECONDARY)) {
486 conn->compare_seq = ppkt->seq_end;
487 colo_release_primary_pkt(s, ppkt);
488 packet_destroy(spkt, NULL);
489 goto pri;
491 } else {
492 g_queue_push_tail(&conn->primary_list, ppkt);
493 g_queue_push_tail(&conn->secondary_list, spkt);
495 #ifdef DEBUG_COLO_PACKETS
496 qemu_hexdump(stderr, "colo-compare ppkt", ppkt->data, ppkt->size);
497 qemu_hexdump(stderr, "colo-compare spkt", spkt->data, spkt->size);
498 #endif
500 colo_compare_inconsistency_notify(s);
506 * Called from the compare thread on the primary
507 * for compare udp packet
509 static int colo_packet_compare_udp(Packet *spkt, Packet *ppkt)
511 uint16_t network_header_length = ppkt->ip->ip_hl << 2;
512 uint16_t offset = network_header_length + ETH_HLEN + ppkt->vnet_hdr_len;
514 trace_colo_compare_main("compare udp");
517 * Because of ppkt and spkt are both in the same connection,
518 * The ppkt's src ip, dst ip, src port, dst port, ip_proto all are
519 * same with spkt. In addition, IP header's Identification is a random
520 * field, we can handle it in IP fragmentation function later.
521 * COLO just concern the response net packet payload from primary guest
522 * and secondary guest are same or not, So we ignored all IP header include
523 * other field like TOS,TTL,IP Checksum. we only need to compare
524 * the ip payload here.
526 if (ppkt->size != spkt->size) {
527 trace_colo_compare_main("UDP: payload size of packets are different");
528 return -1;
530 if (colo_compare_packet_payload(ppkt, spkt, offset, offset,
531 ppkt->size - offset)) {
532 trace_colo_compare_udp_miscompare("primary pkt size", ppkt->size);
533 trace_colo_compare_udp_miscompare("Secondary pkt size", spkt->size);
534 #ifdef DEBUG_COLO_PACKETS
535 qemu_hexdump(stderr, "colo-compare pri pkt", ppkt->data, ppkt->size);
536 qemu_hexdump(stderr, "colo-compare sec pkt", spkt->data, spkt->size);
537 #endif
538 return -1;
539 } else {
540 return 0;
545 * Called from the compare thread on the primary
546 * for compare icmp packet
548 static int colo_packet_compare_icmp(Packet *spkt, Packet *ppkt)
550 uint16_t network_header_length = ppkt->ip->ip_hl << 2;
551 uint16_t offset = network_header_length + ETH_HLEN + ppkt->vnet_hdr_len;
553 trace_colo_compare_main("compare icmp");
556 * Because of ppkt and spkt are both in the same connection,
557 * The ppkt's src ip, dst ip, src port, dst port, ip_proto all are
558 * same with spkt. In addition, IP header's Identification is a random
559 * field, we can handle it in IP fragmentation function later.
560 * COLO just concern the response net packet payload from primary guest
561 * and secondary guest are same or not, So we ignored all IP header include
562 * other field like TOS,TTL,IP Checksum. we only need to compare
563 * the ip payload here.
565 if (ppkt->size != spkt->size) {
566 trace_colo_compare_main("ICMP: payload size of packets are different");
567 return -1;
569 if (colo_compare_packet_payload(ppkt, spkt, offset, offset,
570 ppkt->size - offset)) {
571 trace_colo_compare_icmp_miscompare("primary pkt size",
572 ppkt->size);
573 trace_colo_compare_icmp_miscompare("Secondary pkt size",
574 spkt->size);
575 #ifdef DEBUG_COLO_PACKETS
576 qemu_hexdump(stderr, "colo-compare pri pkt", ppkt->data, ppkt->size);
577 qemu_hexdump(stderr, "colo-compare sec pkt", spkt->data, spkt->size);
578 #endif
579 return -1;
580 } else {
581 return 0;
586 * Called from the compare thread on the primary
587 * for compare other packet
589 static int colo_packet_compare_other(Packet *spkt, Packet *ppkt)
591 uint16_t offset = ppkt->vnet_hdr_len;
593 trace_colo_compare_main("compare other");
594 if (ppkt->size != spkt->size) {
595 trace_colo_compare_main("Other: payload size of packets are different");
596 return -1;
598 return colo_compare_packet_payload(ppkt, spkt, offset, offset,
599 ppkt->size - offset);
602 static int colo_old_packet_check_one(Packet *pkt, int64_t *check_time)
604 int64_t now = qemu_clock_get_ms(QEMU_CLOCK_HOST);
606 if ((now - pkt->creation_ms) > (*check_time)) {
607 trace_colo_old_packet_check_found(pkt->creation_ms);
608 return 0;
609 } else {
610 return 1;
614 void colo_compare_register_notifier(Notifier *notify)
616 notifier_list_add(&colo_compare_notifiers, notify);
619 void colo_compare_unregister_notifier(Notifier *notify)
621 notifier_remove(notify);
624 static int colo_old_packet_check_one_conn(Connection *conn,
625 CompareState *s)
627 if (!g_queue_is_empty(&conn->primary_list)) {
628 if (g_queue_find_custom(&conn->primary_list,
629 &s->compare_timeout,
630 (GCompareFunc)colo_old_packet_check_one))
631 goto out;
634 if (!g_queue_is_empty(&conn->secondary_list)) {
635 if (g_queue_find_custom(&conn->secondary_list,
636 &s->compare_timeout,
637 (GCompareFunc)colo_old_packet_check_one))
638 goto out;
641 return 1;
643 out:
644 /* Do checkpoint will flush old packet */
645 colo_compare_inconsistency_notify(s);
646 return 0;
650 * Look for old packets that the secondary hasn't matched,
651 * if we have some then we have to checkpoint to wake
652 * the secondary up.
654 static void colo_old_packet_check(void *opaque)
656 CompareState *s = opaque;
659 * If we find one old packet, stop finding job and notify
660 * COLO frame do checkpoint.
662 g_queue_find_custom(&s->conn_list, s,
663 (GCompareFunc)colo_old_packet_check_one_conn);
666 static void colo_compare_packet(CompareState *s, Connection *conn,
667 int (*HandlePacket)(Packet *spkt,
668 Packet *ppkt))
670 Packet *pkt = NULL;
671 GList *result = NULL;
673 while (!g_queue_is_empty(&conn->primary_list) &&
674 !g_queue_is_empty(&conn->secondary_list)) {
675 pkt = g_queue_pop_tail(&conn->primary_list);
676 result = g_queue_find_custom(&conn->secondary_list,
677 pkt, (GCompareFunc)HandlePacket);
679 if (result) {
680 colo_release_primary_pkt(s, pkt);
681 packet_destroy(result->data, NULL);
682 g_queue_delete_link(&conn->secondary_list, result);
683 } else {
685 * If one packet arrive late, the secondary_list or
686 * primary_list will be empty, so we can't compare it
687 * until next comparison. If the packets in the list are
688 * timeout, it will trigger a checkpoint request.
690 trace_colo_compare_main("packet different");
691 g_queue_push_tail(&conn->primary_list, pkt);
693 colo_compare_inconsistency_notify(s);
694 break;
700 * Called from the compare thread on the primary
701 * for compare packet with secondary list of the
702 * specified connection when a new packet was
703 * queued to it.
705 static void colo_compare_connection(void *opaque, void *user_data)
707 CompareState *s = user_data;
708 Connection *conn = opaque;
710 switch (conn->ip_proto) {
711 case IPPROTO_TCP:
712 colo_compare_tcp(s, conn);
713 break;
714 case IPPROTO_UDP:
715 colo_compare_packet(s, conn, colo_packet_compare_udp);
716 break;
717 case IPPROTO_ICMP:
718 colo_compare_packet(s, conn, colo_packet_compare_icmp);
719 break;
720 default:
721 colo_compare_packet(s, conn, colo_packet_compare_other);
722 break;
726 static void coroutine_fn _compare_chr_send(void *opaque)
728 SendCo *sendco = opaque;
729 CompareState *s = sendco->s;
730 int ret = 0;
732 while (!g_queue_is_empty(&sendco->send_list)) {
733 SendEntry *entry = g_queue_pop_tail(&sendco->send_list);
734 uint32_t len = htonl(entry->size);
736 ret = qemu_chr_fe_write_all(sendco->chr, (uint8_t *)&len, sizeof(len));
738 if (ret != sizeof(len)) {
739 g_free(entry->buf);
740 g_slice_free(SendEntry, entry);
741 goto err;
744 if (!sendco->notify_remote_frame && s->vnet_hdr) {
746 * We send vnet header len make other module(like filter-redirector)
747 * know how to parse net packet correctly.
749 len = htonl(entry->vnet_hdr_len);
751 ret = qemu_chr_fe_write_all(sendco->chr,
752 (uint8_t *)&len,
753 sizeof(len));
755 if (ret != sizeof(len)) {
756 g_free(entry->buf);
757 g_slice_free(SendEntry, entry);
758 goto err;
762 ret = qemu_chr_fe_write_all(sendco->chr,
763 (uint8_t *)entry->buf,
764 entry->size);
766 if (ret != entry->size) {
767 g_free(entry->buf);
768 g_slice_free(SendEntry, entry);
769 goto err;
772 g_free(entry->buf);
773 g_slice_free(SendEntry, entry);
776 sendco->ret = 0;
777 goto out;
779 err:
780 while (!g_queue_is_empty(&sendco->send_list)) {
781 SendEntry *entry = g_queue_pop_tail(&sendco->send_list);
782 g_free(entry->buf);
783 g_slice_free(SendEntry, entry);
785 sendco->ret = ret < 0 ? ret : -EIO;
786 out:
787 sendco->co = NULL;
788 sendco->done = true;
789 aio_wait_kick();
792 static int compare_chr_send(CompareState *s,
793 uint8_t *buf,
794 uint32_t size,
795 uint32_t vnet_hdr_len,
796 bool notify_remote_frame,
797 bool zero_copy)
799 SendCo *sendco;
800 SendEntry *entry;
802 if (notify_remote_frame) {
803 sendco = &s->notify_sendco;
804 } else {
805 sendco = &s->out_sendco;
808 if (!size) {
809 return -1;
812 entry = g_slice_new(SendEntry);
813 entry->size = size;
814 entry->vnet_hdr_len = vnet_hdr_len;
815 if (zero_copy) {
816 entry->buf = buf;
817 } else {
818 entry->buf = g_malloc(size);
819 memcpy(entry->buf, buf, size);
821 g_queue_push_tail(&sendco->send_list, entry);
823 if (sendco->done) {
824 sendco->co = qemu_coroutine_create(_compare_chr_send, sendco);
825 sendco->done = false;
826 qemu_coroutine_enter(sendco->co);
827 if (sendco->done) {
828 /* report early errors */
829 return sendco->ret;
833 /* assume success */
834 return 0;
837 static int compare_chr_can_read(void *opaque)
839 return COMPARE_READ_LEN_MAX;
843 * Called from the main thread on the primary for packets
844 * arriving over the socket from the primary.
846 static void compare_pri_chr_in(void *opaque, const uint8_t *buf, int size)
848 CompareState *s = COLO_COMPARE(opaque);
849 int ret;
851 ret = net_fill_rstate(&s->pri_rs, buf, size);
852 if (ret == -1) {
853 qemu_chr_fe_set_handlers(&s->chr_pri_in, NULL, NULL, NULL, NULL,
854 NULL, NULL, true);
855 error_report("colo-compare primary_in error");
860 * Called from the main thread on the primary for packets
861 * arriving over the socket from the secondary.
863 static void compare_sec_chr_in(void *opaque, const uint8_t *buf, int size)
865 CompareState *s = COLO_COMPARE(opaque);
866 int ret;
868 ret = net_fill_rstate(&s->sec_rs, buf, size);
869 if (ret == -1) {
870 qemu_chr_fe_set_handlers(&s->chr_sec_in, NULL, NULL, NULL, NULL,
871 NULL, NULL, true);
872 error_report("colo-compare secondary_in error");
876 static void compare_notify_chr(void *opaque, const uint8_t *buf, int size)
878 CompareState *s = COLO_COMPARE(opaque);
879 int ret;
881 ret = net_fill_rstate(&s->notify_rs, buf, size);
882 if (ret == -1) {
883 qemu_chr_fe_set_handlers(&s->chr_notify_dev, NULL, NULL, NULL, NULL,
884 NULL, NULL, true);
885 error_report("colo-compare notify_dev error");
890 * Check old packet regularly so it can watch for any packets
891 * that the secondary hasn't produced equivalents of.
893 static void check_old_packet_regular(void *opaque)
895 CompareState *s = opaque;
897 /* if have old packet we will notify checkpoint */
898 colo_old_packet_check(s);
899 timer_mod(s->packet_check_timer, qemu_clock_get_ms(QEMU_CLOCK_HOST) +
900 s->expired_scan_cycle);
903 /* Public API, Used for COLO frame to notify compare event */
904 void colo_notify_compares_event(void *opaque, int event, Error **errp)
906 CompareState *s;
907 qemu_mutex_lock(&colo_compare_mutex);
909 if (!colo_compare_active) {
910 qemu_mutex_unlock(&colo_compare_mutex);
911 return;
914 qemu_mutex_lock(&event_mtx);
915 QTAILQ_FOREACH(s, &net_compares, next) {
916 s->event = event;
917 qemu_bh_schedule(s->event_bh);
918 event_unhandled_count++;
920 /* Wait all compare threads to finish handling this event */
921 while (event_unhandled_count > 0) {
922 qemu_cond_wait(&event_complete_cond, &event_mtx);
925 qemu_mutex_unlock(&event_mtx);
926 qemu_mutex_unlock(&colo_compare_mutex);
929 static void colo_compare_timer_init(CompareState *s)
931 AioContext *ctx = iothread_get_aio_context(s->iothread);
933 s->packet_check_timer = aio_timer_new(ctx, QEMU_CLOCK_HOST,
934 SCALE_MS, check_old_packet_regular,
936 timer_mod(s->packet_check_timer, qemu_clock_get_ms(QEMU_CLOCK_HOST) +
937 s->expired_scan_cycle);
940 static void colo_compare_timer_del(CompareState *s)
942 if (s->packet_check_timer) {
943 timer_free(s->packet_check_timer);
944 s->packet_check_timer = NULL;
948 static void colo_flush_packets(void *opaque, void *user_data);
950 static void colo_compare_handle_event(void *opaque)
952 CompareState *s = opaque;
954 switch (s->event) {
955 case COLO_EVENT_CHECKPOINT:
956 g_queue_foreach(&s->conn_list, colo_flush_packets, s);
957 break;
958 case COLO_EVENT_FAILOVER:
959 break;
960 default:
961 break;
964 qemu_mutex_lock(&event_mtx);
965 assert(event_unhandled_count > 0);
966 event_unhandled_count--;
967 qemu_cond_broadcast(&event_complete_cond);
968 qemu_mutex_unlock(&event_mtx);
971 static void colo_compare_iothread(CompareState *s)
973 AioContext *ctx = iothread_get_aio_context(s->iothread);
974 object_ref(OBJECT(s->iothread));
975 s->worker_context = iothread_get_g_main_context(s->iothread);
977 qemu_chr_fe_set_handlers(&s->chr_pri_in, compare_chr_can_read,
978 compare_pri_chr_in, NULL, NULL,
979 s, s->worker_context, true);
980 qemu_chr_fe_set_handlers(&s->chr_sec_in, compare_chr_can_read,
981 compare_sec_chr_in, NULL, NULL,
982 s, s->worker_context, true);
983 if (s->notify_dev) {
984 qemu_chr_fe_set_handlers(&s->chr_notify_dev, compare_chr_can_read,
985 compare_notify_chr, NULL, NULL,
986 s, s->worker_context, true);
989 colo_compare_timer_init(s);
990 s->event_bh = aio_bh_new(ctx, colo_compare_handle_event, s);
993 static char *compare_get_pri_indev(Object *obj, Error **errp)
995 CompareState *s = COLO_COMPARE(obj);
997 return g_strdup(s->pri_indev);
1000 static void compare_set_pri_indev(Object *obj, const char *value, Error **errp)
1002 CompareState *s = COLO_COMPARE(obj);
1004 g_free(s->pri_indev);
1005 s->pri_indev = g_strdup(value);
1008 static char *compare_get_sec_indev(Object *obj, Error **errp)
1010 CompareState *s = COLO_COMPARE(obj);
1012 return g_strdup(s->sec_indev);
1015 static void compare_set_sec_indev(Object *obj, const char *value, Error **errp)
1017 CompareState *s = COLO_COMPARE(obj);
1019 g_free(s->sec_indev);
1020 s->sec_indev = g_strdup(value);
1023 static char *compare_get_outdev(Object *obj, Error **errp)
1025 CompareState *s = COLO_COMPARE(obj);
1027 return g_strdup(s->outdev);
1030 static void compare_set_outdev(Object *obj, const char *value, Error **errp)
1032 CompareState *s = COLO_COMPARE(obj);
1034 g_free(s->outdev);
1035 s->outdev = g_strdup(value);
1038 static bool compare_get_vnet_hdr(Object *obj, Error **errp)
1040 CompareState *s = COLO_COMPARE(obj);
1042 return s->vnet_hdr;
1045 static void compare_set_vnet_hdr(Object *obj,
1046 bool value,
1047 Error **errp)
1049 CompareState *s = COLO_COMPARE(obj);
1051 s->vnet_hdr = value;
1054 static char *compare_get_notify_dev(Object *obj, Error **errp)
1056 CompareState *s = COLO_COMPARE(obj);
1058 return g_strdup(s->notify_dev);
1061 static void compare_set_notify_dev(Object *obj, const char *value, Error **errp)
1063 CompareState *s = COLO_COMPARE(obj);
1065 g_free(s->notify_dev);
1066 s->notify_dev = g_strdup(value);
1069 static void compare_get_timeout(Object *obj, Visitor *v,
1070 const char *name, void *opaque,
1071 Error **errp)
1073 CompareState *s = COLO_COMPARE(obj);
1074 uint64_t value = s->compare_timeout;
1076 visit_type_uint64(v, name, &value, errp);
1079 static void compare_set_timeout(Object *obj, Visitor *v,
1080 const char *name, void *opaque,
1081 Error **errp)
1083 CompareState *s = COLO_COMPARE(obj);
1084 uint32_t value;
1086 if (!visit_type_uint32(v, name, &value, errp)) {
1087 return;
1089 if (!value) {
1090 error_setg(errp, "Property '%s.%s' requires a positive value",
1091 object_get_typename(obj), name);
1092 return;
1094 s->compare_timeout = value;
1097 static void compare_get_expired_scan_cycle(Object *obj, Visitor *v,
1098 const char *name, void *opaque,
1099 Error **errp)
1101 CompareState *s = COLO_COMPARE(obj);
1102 uint32_t value = s->expired_scan_cycle;
1104 visit_type_uint32(v, name, &value, errp);
1107 static void compare_set_expired_scan_cycle(Object *obj, Visitor *v,
1108 const char *name, void *opaque,
1109 Error **errp)
1111 CompareState *s = COLO_COMPARE(obj);
1112 uint32_t value;
1114 if (!visit_type_uint32(v, name, &value, errp)) {
1115 return;
1117 if (!value) {
1118 error_setg(errp, "Property '%s.%s' requires a positive value",
1119 object_get_typename(obj), name);
1120 return;
1122 s->expired_scan_cycle = value;
1125 static void get_max_queue_size(Object *obj, Visitor *v,
1126 const char *name, void *opaque,
1127 Error **errp)
1129 uint32_t value = max_queue_size;
1131 visit_type_uint32(v, name, &value, errp);
1134 static void set_max_queue_size(Object *obj, Visitor *v,
1135 const char *name, void *opaque,
1136 Error **errp)
1138 Error *local_err = NULL;
1139 uint64_t value;
1141 visit_type_uint64(v, name, &value, &local_err);
1142 if (local_err) {
1143 goto out;
1145 if (!value) {
1146 error_setg(&local_err, "Property '%s.%s' requires a positive value",
1147 object_get_typename(obj), name);
1148 goto out;
1150 max_queue_size = value;
1152 out:
1153 error_propagate(errp, local_err);
1156 static void compare_pri_rs_finalize(SocketReadState *pri_rs)
1158 CompareState *s = container_of(pri_rs, CompareState, pri_rs);
1159 Connection *conn = NULL;
1161 if (packet_enqueue(s, PRIMARY_IN, &conn)) {
1162 trace_colo_compare_main("primary: unsupported packet in");
1163 compare_chr_send(s,
1164 pri_rs->buf,
1165 pri_rs->packet_len,
1166 pri_rs->vnet_hdr_len,
1167 false,
1168 false);
1169 } else {
1170 /* compare packet in the specified connection */
1171 colo_compare_connection(conn, s);
1175 static void compare_sec_rs_finalize(SocketReadState *sec_rs)
1177 CompareState *s = container_of(sec_rs, CompareState, sec_rs);
1178 Connection *conn = NULL;
1180 if (packet_enqueue(s, SECONDARY_IN, &conn)) {
1181 trace_colo_compare_main("secondary: unsupported packet in");
1182 } else {
1183 /* compare packet in the specified connection */
1184 colo_compare_connection(conn, s);
1188 static void compare_notify_rs_finalize(SocketReadState *notify_rs)
1190 CompareState *s = container_of(notify_rs, CompareState, notify_rs);
1192 const char msg[] = "COLO_COMPARE_GET_XEN_INIT";
1193 int ret;
1195 if (packet_matches_str("COLO_USERSPACE_PROXY_INIT",
1196 notify_rs->buf,
1197 notify_rs->packet_len)) {
1198 ret = compare_chr_send(s, (uint8_t *)msg, strlen(msg), 0, true, false);
1199 if (ret < 0) {
1200 error_report("Notify Xen COLO-frame INIT failed");
1202 } else if (packet_matches_str("COLO_CHECKPOINT",
1203 notify_rs->buf,
1204 notify_rs->packet_len)) {
1205 /* colo-compare do checkpoint, flush pri packet and remove sec packet */
1206 g_queue_foreach(&s->conn_list, colo_flush_packets, s);
1207 } else {
1208 error_report("COLO compare got unsupported instruction");
1213 * Return 0 is success.
1214 * Return 1 is failed.
1216 static int find_and_check_chardev(Chardev **chr,
1217 char *chr_name,
1218 Error **errp)
1220 *chr = qemu_chr_find(chr_name);
1221 if (*chr == NULL) {
1222 error_setg(errp, "Device '%s' not found",
1223 chr_name);
1224 return 1;
1227 if (!qemu_chr_has_feature(*chr, QEMU_CHAR_FEATURE_RECONNECTABLE)) {
1228 error_setg(errp, "chardev \"%s\" is not reconnectable",
1229 chr_name);
1230 return 1;
1233 if (!qemu_chr_has_feature(*chr, QEMU_CHAR_FEATURE_GCONTEXT)) {
1234 error_setg(errp, "chardev \"%s\" cannot switch context",
1235 chr_name);
1236 return 1;
1239 return 0;
1243 * Called from the main thread on the primary
1244 * to setup colo-compare.
1246 static void colo_compare_complete(UserCreatable *uc, Error **errp)
1248 CompareState *s = COLO_COMPARE(uc);
1249 Chardev *chr;
1251 if (!s->pri_indev || !s->sec_indev || !s->outdev || !s->iothread) {
1252 error_setg(errp, "colo compare needs 'primary_in' ,"
1253 "'secondary_in','outdev','iothread' property set");
1254 return;
1255 } else if (!strcmp(s->pri_indev, s->outdev) ||
1256 !strcmp(s->sec_indev, s->outdev) ||
1257 !strcmp(s->pri_indev, s->sec_indev)) {
1258 error_setg(errp, "'indev' and 'outdev' could not be same "
1259 "for compare module");
1260 return;
1263 if (!s->compare_timeout) {
1264 /* Set default value to 3000 MS */
1265 s->compare_timeout = DEFAULT_TIME_OUT_MS;
1268 if (!s->expired_scan_cycle) {
1269 /* Set default value to 1000 MS */
1270 s->expired_scan_cycle = REGULAR_PACKET_CHECK_MS;
1273 if (!max_queue_size) {
1274 /* Set default queue size to 1024 */
1275 max_queue_size = MAX_QUEUE_SIZE;
1278 if (find_and_check_chardev(&chr, s->pri_indev, errp) ||
1279 !qemu_chr_fe_init(&s->chr_pri_in, chr, errp)) {
1280 return;
1283 if (find_and_check_chardev(&chr, s->sec_indev, errp) ||
1284 !qemu_chr_fe_init(&s->chr_sec_in, chr, errp)) {
1285 return;
1288 if (find_and_check_chardev(&chr, s->outdev, errp) ||
1289 !qemu_chr_fe_init(&s->chr_out, chr, errp)) {
1290 return;
1293 net_socket_rs_init(&s->pri_rs, compare_pri_rs_finalize, s->vnet_hdr);
1294 net_socket_rs_init(&s->sec_rs, compare_sec_rs_finalize, s->vnet_hdr);
1296 /* Try to enable remote notify chardev, currently just for Xen COLO */
1297 if (s->notify_dev) {
1298 if (find_and_check_chardev(&chr, s->notify_dev, errp) ||
1299 !qemu_chr_fe_init(&s->chr_notify_dev, chr, errp)) {
1300 return;
1303 net_socket_rs_init(&s->notify_rs, compare_notify_rs_finalize,
1304 s->vnet_hdr);
1307 s->out_sendco.s = s;
1308 s->out_sendco.chr = &s->chr_out;
1309 s->out_sendco.notify_remote_frame = false;
1310 s->out_sendco.done = true;
1311 g_queue_init(&s->out_sendco.send_list);
1313 if (s->notify_dev) {
1314 s->notify_sendco.s = s;
1315 s->notify_sendco.chr = &s->chr_notify_dev;
1316 s->notify_sendco.notify_remote_frame = true;
1317 s->notify_sendco.done = true;
1318 g_queue_init(&s->notify_sendco.send_list);
1321 g_queue_init(&s->conn_list);
1323 s->connection_track_table = g_hash_table_new_full(connection_key_hash,
1324 connection_key_equal,
1325 g_free,
1326 NULL);
1328 colo_compare_iothread(s);
1330 qemu_mutex_lock(&colo_compare_mutex);
1331 if (!colo_compare_active) {
1332 qemu_mutex_init(&event_mtx);
1333 qemu_cond_init(&event_complete_cond);
1334 colo_compare_active = true;
1336 QTAILQ_INSERT_TAIL(&net_compares, s, next);
1337 qemu_mutex_unlock(&colo_compare_mutex);
1339 return;
1342 static void colo_flush_packets(void *opaque, void *user_data)
1344 CompareState *s = user_data;
1345 Connection *conn = opaque;
1346 Packet *pkt = NULL;
1348 while (!g_queue_is_empty(&conn->primary_list)) {
1349 pkt = g_queue_pop_tail(&conn->primary_list);
1350 compare_chr_send(s,
1351 pkt->data,
1352 pkt->size,
1353 pkt->vnet_hdr_len,
1354 false,
1355 true);
1356 packet_destroy_partial(pkt, NULL);
1358 while (!g_queue_is_empty(&conn->secondary_list)) {
1359 pkt = g_queue_pop_tail(&conn->secondary_list);
1360 packet_destroy(pkt, NULL);
1364 static void colo_compare_class_init(ObjectClass *oc, void *data)
1366 UserCreatableClass *ucc = USER_CREATABLE_CLASS(oc);
1368 ucc->complete = colo_compare_complete;
1371 static void colo_compare_init(Object *obj)
1373 CompareState *s = COLO_COMPARE(obj);
1375 object_property_add_str(obj, "primary_in",
1376 compare_get_pri_indev, compare_set_pri_indev);
1377 object_property_add_str(obj, "secondary_in",
1378 compare_get_sec_indev, compare_set_sec_indev);
1379 object_property_add_str(obj, "outdev",
1380 compare_get_outdev, compare_set_outdev);
1381 object_property_add_link(obj, "iothread", TYPE_IOTHREAD,
1382 (Object **)&s->iothread,
1383 object_property_allow_set_link,
1384 OBJ_PROP_LINK_STRONG);
1385 /* This parameter just for Xen COLO */
1386 object_property_add_str(obj, "notify_dev",
1387 compare_get_notify_dev, compare_set_notify_dev);
1389 object_property_add(obj, "compare_timeout", "uint64",
1390 compare_get_timeout,
1391 compare_set_timeout, NULL, NULL);
1393 object_property_add(obj, "expired_scan_cycle", "uint32",
1394 compare_get_expired_scan_cycle,
1395 compare_set_expired_scan_cycle, NULL, NULL);
1397 object_property_add(obj, "max_queue_size", "uint32",
1398 get_max_queue_size,
1399 set_max_queue_size, NULL, NULL);
1401 s->vnet_hdr = false;
1402 object_property_add_bool(obj, "vnet_hdr_support", compare_get_vnet_hdr,
1403 compare_set_vnet_hdr);
1406 void colo_compare_cleanup(void)
1408 CompareState *tmp = NULL;
1409 CompareState *n = NULL;
1411 QTAILQ_FOREACH_SAFE(tmp, &net_compares, next, n) {
1412 object_unparent(OBJECT(tmp));
1416 static void colo_compare_finalize(Object *obj)
1418 CompareState *s = COLO_COMPARE(obj);
1419 CompareState *tmp = NULL;
1421 qemu_mutex_lock(&colo_compare_mutex);
1422 QTAILQ_FOREACH(tmp, &net_compares, next) {
1423 if (tmp == s) {
1424 QTAILQ_REMOVE(&net_compares, s, next);
1425 break;
1428 if (QTAILQ_EMPTY(&net_compares)) {
1429 colo_compare_active = false;
1430 qemu_mutex_destroy(&event_mtx);
1431 qemu_cond_destroy(&event_complete_cond);
1433 qemu_mutex_unlock(&colo_compare_mutex);
1435 qemu_chr_fe_deinit(&s->chr_pri_in, false);
1436 qemu_chr_fe_deinit(&s->chr_sec_in, false);
1437 qemu_chr_fe_deinit(&s->chr_out, false);
1438 if (s->notify_dev) {
1439 qemu_chr_fe_deinit(&s->chr_notify_dev, false);
1442 colo_compare_timer_del(s);
1444 qemu_bh_delete(s->event_bh);
1446 AioContext *ctx = iothread_get_aio_context(s->iothread);
1447 aio_context_acquire(ctx);
1448 AIO_WAIT_WHILE(ctx, !s->out_sendco.done);
1449 if (s->notify_dev) {
1450 AIO_WAIT_WHILE(ctx, !s->notify_sendco.done);
1452 aio_context_release(ctx);
1454 /* Release all unhandled packets after compare thead exited */
1455 g_queue_foreach(&s->conn_list, colo_flush_packets, s);
1456 AIO_WAIT_WHILE(NULL, !s->out_sendco.done);
1458 g_queue_clear(&s->conn_list);
1459 g_queue_clear(&s->out_sendco.send_list);
1460 if (s->notify_dev) {
1461 g_queue_clear(&s->notify_sendco.send_list);
1464 if (s->connection_track_table) {
1465 g_hash_table_destroy(s->connection_track_table);
1468 object_unref(OBJECT(s->iothread));
1470 g_free(s->pri_indev);
1471 g_free(s->sec_indev);
1472 g_free(s->outdev);
1473 g_free(s->notify_dev);
1476 static void __attribute__((__constructor__)) colo_compare_init_globals(void)
1478 colo_compare_active = false;
1479 qemu_mutex_init(&colo_compare_mutex);
1482 static const TypeInfo colo_compare_info = {
1483 .name = TYPE_COLO_COMPARE,
1484 .parent = TYPE_OBJECT,
1485 .instance_size = sizeof(CompareState),
1486 .instance_init = colo_compare_init,
1487 .instance_finalize = colo_compare_finalize,
1488 .class_size = sizeof(CompareClass),
1489 .class_init = colo_compare_class_init,
1490 .interfaces = (InterfaceInfo[]) {
1491 { TYPE_USER_CREATABLE },
1496 static void register_types(void)
1498 type_register_static(&colo_compare_info);
1501 type_init(register_types);