target/arm: Tidy sve_exception_el for CPACR_EL1 access
[qemu.git] / net / colo-compare.c
blobb966e7e514d031f67d7ddd9a482d7d1ee4cb4d1d
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-common.h"
17 #include "qemu/error-report.h"
18 #include "trace.h"
19 #include "qapi/error.h"
20 #include "net/net.h"
21 #include "net/eth.h"
22 #include "qom/object_interfaces.h"
23 #include "qemu/iov.h"
24 #include "qom/object.h"
25 #include "net/queue.h"
26 #include "chardev/char-fe.h"
27 #include "qemu/sockets.h"
28 #include "colo.h"
29 #include "sysemu/iothread.h"
30 #include "net/colo-compare.h"
31 #include "migration/colo.h"
32 #include "migration/migration.h"
33 #include "util.h"
35 #include "block/aio-wait.h"
36 #include "qemu/coroutine.h"
38 #define TYPE_COLO_COMPARE "colo-compare"
39 typedef struct CompareState CompareState;
40 DECLARE_INSTANCE_CHECKER(CompareState, COLO_COMPARE,
41 TYPE_COLO_COMPARE)
43 static QTAILQ_HEAD(, CompareState) net_compares =
44 QTAILQ_HEAD_INITIALIZER(net_compares);
46 static NotifierList colo_compare_notifiers =
47 NOTIFIER_LIST_INITIALIZER(colo_compare_notifiers);
49 #define COMPARE_READ_LEN_MAX NET_BUFSIZE
50 #define MAX_QUEUE_SIZE 1024
52 #define COLO_COMPARE_FREE_PRIMARY 0x01
53 #define COLO_COMPARE_FREE_SECONDARY 0x02
55 #define REGULAR_PACKET_CHECK_MS 1000
56 #define DEFAULT_TIME_OUT_MS 3000
58 /* #define DEBUG_COLO_PACKETS */
60 static QemuMutex colo_compare_mutex;
61 static bool colo_compare_active;
62 static QemuMutex event_mtx;
63 static QemuCond event_complete_cond;
64 static int event_unhandled_count;
65 static uint32_t max_queue_size;
68 * + CompareState ++
69 * | |
70 * +---------------+ +---------------+ +---------------+
71 * | conn list + - > conn + ------- > conn + -- > ......
72 * +---------------+ +---------------+ +---------------+
73 * | | | | | |
74 * +---------------+ +---v----+ +---v----+ +---v----+ +---v----+
75 * |primary | |secondary |primary | |secondary
76 * |packet | |packet + |packet | |packet +
77 * +--------+ +--------+ +--------+ +--------+
78 * | | | |
79 * +---v----+ +---v----+ +---v----+ +---v----+
80 * |primary | |secondary |primary | |secondary
81 * |packet | |packet + |packet | |packet +
82 * +--------+ +--------+ +--------+ +--------+
83 * | | | |
84 * +---v----+ +---v----+ +---v----+ +---v----+
85 * |primary | |secondary |primary | |secondary
86 * |packet | |packet + |packet | |packet +
87 * +--------+ +--------+ +--------+ +--------+
90 typedef struct SendCo {
91 Coroutine *co;
92 struct CompareState *s;
93 CharBackend *chr;
94 GQueue send_list;
95 bool notify_remote_frame;
96 bool done;
97 int ret;
98 } SendCo;
100 typedef struct SendEntry {
101 uint32_t size;
102 uint32_t vnet_hdr_len;
103 uint8_t *buf;
104 } SendEntry;
106 struct CompareState {
107 Object parent;
109 char *pri_indev;
110 char *sec_indev;
111 char *outdev;
112 char *notify_dev;
113 CharBackend chr_pri_in;
114 CharBackend chr_sec_in;
115 CharBackend chr_out;
116 CharBackend chr_notify_dev;
117 SocketReadState pri_rs;
118 SocketReadState sec_rs;
119 SocketReadState notify_rs;
120 SendCo out_sendco;
121 SendCo notify_sendco;
122 bool vnet_hdr;
123 uint64_t compare_timeout;
124 uint32_t expired_scan_cycle;
127 * Record the connection that through the NIC
128 * Element type: Connection
130 GQueue conn_list;
131 /* Record the connection without repetition */
132 GHashTable *connection_track_table;
134 IOThread *iothread;
135 GMainContext *worker_context;
136 QEMUTimer *packet_check_timer;
138 QEMUBH *event_bh;
139 enum colo_event event;
141 QTAILQ_ENTRY(CompareState) next;
144 typedef struct CompareClass {
145 ObjectClass parent_class;
146 } CompareClass;
148 enum {
149 PRIMARY_IN = 0,
150 SECONDARY_IN,
153 static const char *colo_mode[] = {
154 [PRIMARY_IN] = "primary",
155 [SECONDARY_IN] = "secondary",
158 static int compare_chr_send(CompareState *s,
159 uint8_t *buf,
160 uint32_t size,
161 uint32_t vnet_hdr_len,
162 bool notify_remote_frame,
163 bool zero_copy);
165 static bool packet_matches_str(const char *str,
166 const uint8_t *buf,
167 uint32_t packet_len)
169 if (packet_len != strlen(str)) {
170 return false;
173 return !memcmp(str, buf, packet_len);
176 static void notify_remote_frame(CompareState *s)
178 char msg[] = "DO_CHECKPOINT";
179 int ret = 0;
181 ret = compare_chr_send(s, (uint8_t *)msg, strlen(msg), 0, true, false);
182 if (ret < 0) {
183 error_report("Notify Xen COLO-frame failed");
187 static void colo_compare_inconsistency_notify(CompareState *s)
189 if (s->notify_dev) {
190 notify_remote_frame(s);
191 } else {
192 notifier_list_notify(&colo_compare_notifiers,
193 migrate_get_current());
197 /* Use restricted to colo_insert_packet() */
198 static gint seq_sorter(Packet *a, Packet *b, gpointer data)
200 return a->tcp_seq - b->tcp_seq;
203 static void fill_pkt_tcp_info(void *data, uint32_t *max_ack)
205 Packet *pkt = data;
206 struct tcp_hdr *tcphd;
208 tcphd = (struct tcp_hdr *)pkt->transport_header;
210 pkt->tcp_seq = ntohl(tcphd->th_seq);
211 pkt->tcp_ack = ntohl(tcphd->th_ack);
212 /* Need to consider ACK will bigger than uint32_t MAX */
213 *max_ack = pkt->tcp_ack - *max_ack > 0 ? pkt->tcp_ack : *max_ack;
214 pkt->header_size = pkt->transport_header - (uint8_t *)pkt->data
215 + (tcphd->th_off << 2);
216 pkt->payload_size = pkt->size - pkt->header_size;
217 pkt->seq_end = pkt->tcp_seq + pkt->payload_size;
218 pkt->flags = tcphd->th_flags;
222 * Return 1 on success, if return 0 means the
223 * packet will be dropped
225 static int colo_insert_packet(GQueue *queue, Packet *pkt, uint32_t *max_ack)
227 if (g_queue_get_length(queue) <= max_queue_size) {
228 if (pkt->ip->ip_p == IPPROTO_TCP) {
229 fill_pkt_tcp_info(pkt, max_ack);
230 g_queue_insert_sorted(queue,
231 pkt,
232 (GCompareDataFunc)seq_sorter,
233 NULL);
234 } else {
235 g_queue_push_tail(queue, pkt);
237 return 1;
239 return 0;
243 * Return 0 on success, if return -1 means the pkt
244 * is unsupported(arp and ipv6) and will be sent later
246 static int packet_enqueue(CompareState *s, int mode, Connection **con)
248 ConnectionKey key;
249 Packet *pkt = NULL;
250 Connection *conn;
251 int ret;
253 if (mode == PRIMARY_IN) {
254 pkt = packet_new(s->pri_rs.buf,
255 s->pri_rs.packet_len,
256 s->pri_rs.vnet_hdr_len);
257 } else {
258 pkt = packet_new(s->sec_rs.buf,
259 s->sec_rs.packet_len,
260 s->sec_rs.vnet_hdr_len);
263 if (parse_packet_early(pkt)) {
264 packet_destroy(pkt, NULL);
265 pkt = NULL;
266 return -1;
268 fill_connection_key(pkt, &key, false);
270 conn = connection_get(s->connection_track_table,
271 &key,
272 &s->conn_list);
274 if (!conn->processing) {
275 g_queue_push_tail(&s->conn_list, conn);
276 conn->processing = true;
279 if (mode == PRIMARY_IN) {
280 ret = colo_insert_packet(&conn->primary_list, pkt, &conn->pack);
281 } else {
282 ret = colo_insert_packet(&conn->secondary_list, pkt, &conn->sack);
285 if (!ret) {
286 trace_colo_compare_drop_packet(colo_mode[mode],
287 "queue size too big, drop packet");
288 packet_destroy(pkt, NULL);
289 pkt = NULL;
292 *con = conn;
294 return 0;
297 static inline bool after(uint32_t seq1, uint32_t seq2)
299 return (int32_t)(seq1 - seq2) > 0;
302 static void colo_release_primary_pkt(CompareState *s, Packet *pkt)
304 int ret;
305 ret = compare_chr_send(s,
306 pkt->data,
307 pkt->size,
308 pkt->vnet_hdr_len,
309 false,
310 true);
311 if (ret < 0) {
312 error_report("colo send primary packet failed");
314 trace_colo_compare_main("packet same and release packet");
315 packet_destroy_partial(pkt, NULL);
319 * The IP packets sent by primary and secondary
320 * will be compared in here
321 * TODO support ip fragment, Out-Of-Order
322 * return: 0 means packet same
323 * > 0 || < 0 means packet different
325 static int colo_compare_packet_payload(Packet *ppkt,
326 Packet *spkt,
327 uint16_t poffset,
328 uint16_t soffset,
329 uint16_t len)
332 if (trace_event_get_state_backends(TRACE_COLO_COMPARE_IP_INFO)) {
333 char pri_ip_src[20], pri_ip_dst[20], sec_ip_src[20], sec_ip_dst[20];
335 strcpy(pri_ip_src, inet_ntoa(ppkt->ip->ip_src));
336 strcpy(pri_ip_dst, inet_ntoa(ppkt->ip->ip_dst));
337 strcpy(sec_ip_src, inet_ntoa(spkt->ip->ip_src));
338 strcpy(sec_ip_dst, inet_ntoa(spkt->ip->ip_dst));
340 trace_colo_compare_ip_info(ppkt->size, pri_ip_src,
341 pri_ip_dst, spkt->size,
342 sec_ip_src, sec_ip_dst);
345 return memcmp(ppkt->data + poffset, spkt->data + soffset, len);
349 * return true means that the payload is consist and
350 * need to make the next comparison, false means do
351 * the checkpoint
353 static bool colo_mark_tcp_pkt(Packet *ppkt, Packet *spkt,
354 int8_t *mark, uint32_t max_ack)
356 *mark = 0;
358 if (ppkt->tcp_seq == spkt->tcp_seq && ppkt->seq_end == spkt->seq_end) {
359 if (!colo_compare_packet_payload(ppkt, spkt,
360 ppkt->header_size, spkt->header_size,
361 ppkt->payload_size)) {
362 *mark = COLO_COMPARE_FREE_SECONDARY | COLO_COMPARE_FREE_PRIMARY;
363 return true;
367 /* one part of secondary packet payload still need to be compared */
368 if (!after(ppkt->seq_end, spkt->seq_end)) {
369 if (!colo_compare_packet_payload(ppkt, spkt,
370 ppkt->header_size + ppkt->offset,
371 spkt->header_size + spkt->offset,
372 ppkt->payload_size - ppkt->offset)) {
373 if (!after(ppkt->tcp_ack, max_ack)) {
374 *mark = COLO_COMPARE_FREE_PRIMARY;
375 spkt->offset += ppkt->payload_size - ppkt->offset;
376 return true;
377 } else {
378 /* secondary guest hasn't ack the data, don't send
379 * out this packet
381 return false;
384 } else {
385 /* primary packet is longer than secondary packet, compare
386 * the same part and mark the primary packet offset
388 if (!colo_compare_packet_payload(ppkt, spkt,
389 ppkt->header_size + ppkt->offset,
390 spkt->header_size + spkt->offset,
391 spkt->payload_size - spkt->offset)) {
392 *mark = COLO_COMPARE_FREE_SECONDARY;
393 ppkt->offset += spkt->payload_size - spkt->offset;
394 return true;
398 return false;
401 static void colo_compare_tcp(CompareState *s, Connection *conn)
403 Packet *ppkt = NULL, *spkt = NULL;
404 int8_t mark;
407 * If ppkt and spkt have the same payload, but ppkt's ACK
408 * is greater than spkt's ACK, in this case we can not
409 * send the ppkt because it will cause the secondary guest
410 * to miss sending some data in the next. Therefore, we
411 * record the maximum ACK in the current queue at both
412 * primary side and secondary side. Only when the ack is
413 * less than the smaller of the two maximum ack, then we
414 * can ensure that the packet's payload is acknowledged by
415 * primary and secondary.
417 uint32_t min_ack = conn->pack - conn->sack > 0 ?
418 conn->sack : conn->pack;
420 pri:
421 if (g_queue_is_empty(&conn->primary_list)) {
422 return;
424 ppkt = g_queue_pop_head(&conn->primary_list);
425 sec:
426 if (g_queue_is_empty(&conn->secondary_list)) {
427 g_queue_push_head(&conn->primary_list, ppkt);
428 return;
430 spkt = g_queue_pop_head(&conn->secondary_list);
432 if (ppkt->tcp_seq == ppkt->seq_end) {
433 colo_release_primary_pkt(s, ppkt);
434 ppkt = NULL;
437 if (ppkt && conn->compare_seq && !after(ppkt->seq_end, conn->compare_seq)) {
438 trace_colo_compare_main("pri: this packet has compared");
439 colo_release_primary_pkt(s, ppkt);
440 ppkt = NULL;
443 if (spkt->tcp_seq == spkt->seq_end) {
444 packet_destroy(spkt, NULL);
445 if (!ppkt) {
446 goto pri;
447 } else {
448 goto sec;
450 } else {
451 if (conn->compare_seq && !after(spkt->seq_end, conn->compare_seq)) {
452 trace_colo_compare_main("sec: this packet has compared");
453 packet_destroy(spkt, NULL);
454 if (!ppkt) {
455 goto pri;
456 } else {
457 goto sec;
460 if (!ppkt) {
461 g_queue_push_head(&conn->secondary_list, spkt);
462 goto pri;
466 if (colo_mark_tcp_pkt(ppkt, spkt, &mark, min_ack)) {
467 trace_colo_compare_tcp_info("pri",
468 ppkt->tcp_seq, ppkt->tcp_ack,
469 ppkt->header_size, ppkt->payload_size,
470 ppkt->offset, ppkt->flags);
472 trace_colo_compare_tcp_info("sec",
473 spkt->tcp_seq, spkt->tcp_ack,
474 spkt->header_size, spkt->payload_size,
475 spkt->offset, spkt->flags);
477 if (mark == COLO_COMPARE_FREE_PRIMARY) {
478 conn->compare_seq = ppkt->seq_end;
479 colo_release_primary_pkt(s, ppkt);
480 g_queue_push_head(&conn->secondary_list, spkt);
481 goto pri;
482 } else if (mark == COLO_COMPARE_FREE_SECONDARY) {
483 conn->compare_seq = spkt->seq_end;
484 packet_destroy(spkt, NULL);
485 goto sec;
486 } else if (mark == (COLO_COMPARE_FREE_PRIMARY | COLO_COMPARE_FREE_SECONDARY)) {
487 conn->compare_seq = ppkt->seq_end;
488 colo_release_primary_pkt(s, ppkt);
489 packet_destroy(spkt, NULL);
490 goto pri;
492 } else {
493 g_queue_push_head(&conn->primary_list, ppkt);
494 g_queue_push_head(&conn->secondary_list, spkt);
496 #ifdef DEBUG_COLO_PACKETS
497 qemu_hexdump(stderr, "colo-compare ppkt", ppkt->data, ppkt->size);
498 qemu_hexdump(stderr, "colo-compare spkt", spkt->data, spkt->size);
499 #endif
501 colo_compare_inconsistency_notify(s);
507 * Called from the compare thread on the primary
508 * for compare udp packet
510 static int colo_packet_compare_udp(Packet *spkt, Packet *ppkt)
512 uint16_t network_header_length = ppkt->ip->ip_hl << 2;
513 uint16_t offset = network_header_length + ETH_HLEN + ppkt->vnet_hdr_len;
515 trace_colo_compare_main("compare udp");
518 * Because of ppkt and spkt are both in the same connection,
519 * The ppkt's src ip, dst ip, src port, dst port, ip_proto all are
520 * same with spkt. In addition, IP header's Identification is a random
521 * field, we can handle it in IP fragmentation function later.
522 * COLO just concern the response net packet payload from primary guest
523 * and secondary guest are same or not, So we ignored all IP header include
524 * other field like TOS,TTL,IP Checksum. we only need to compare
525 * the ip payload here.
527 if (ppkt->size != spkt->size) {
528 trace_colo_compare_main("UDP: payload size of packets are different");
529 return -1;
531 if (colo_compare_packet_payload(ppkt, spkt, offset, offset,
532 ppkt->size - offset)) {
533 trace_colo_compare_udp_miscompare("primary pkt size", ppkt->size);
534 trace_colo_compare_udp_miscompare("Secondary pkt size", spkt->size);
535 #ifdef DEBUG_COLO_PACKETS
536 qemu_hexdump(stderr, "colo-compare pri pkt", ppkt->data, ppkt->size);
537 qemu_hexdump(stderr, "colo-compare sec pkt", spkt->data, spkt->size);
538 #endif
539 return -1;
540 } else {
541 return 0;
546 * Called from the compare thread on the primary
547 * for compare icmp packet
549 static int colo_packet_compare_icmp(Packet *spkt, Packet *ppkt)
551 uint16_t network_header_length = ppkt->ip->ip_hl << 2;
552 uint16_t offset = network_header_length + ETH_HLEN + ppkt->vnet_hdr_len;
554 trace_colo_compare_main("compare icmp");
557 * Because of ppkt and spkt are both in the same connection,
558 * The ppkt's src ip, dst ip, src port, dst port, ip_proto all are
559 * same with spkt. In addition, IP header's Identification is a random
560 * field, we can handle it in IP fragmentation function later.
561 * COLO just concern the response net packet payload from primary guest
562 * and secondary guest are same or not, So we ignored all IP header include
563 * other field like TOS,TTL,IP Checksum. we only need to compare
564 * the ip payload here.
566 if (ppkt->size != spkt->size) {
567 trace_colo_compare_main("ICMP: payload size of packets are different");
568 return -1;
570 if (colo_compare_packet_payload(ppkt, spkt, offset, offset,
571 ppkt->size - offset)) {
572 trace_colo_compare_icmp_miscompare("primary pkt size",
573 ppkt->size);
574 trace_colo_compare_icmp_miscompare("Secondary pkt size",
575 spkt->size);
576 #ifdef DEBUG_COLO_PACKETS
577 qemu_hexdump(stderr, "colo-compare pri pkt", ppkt->data, ppkt->size);
578 qemu_hexdump(stderr, "colo-compare sec pkt", spkt->data, spkt->size);
579 #endif
580 return -1;
581 } else {
582 return 0;
587 * Called from the compare thread on the primary
588 * for compare other packet
590 static int colo_packet_compare_other(Packet *spkt, Packet *ppkt)
592 uint16_t offset = ppkt->vnet_hdr_len;
594 trace_colo_compare_main("compare other");
595 if (ppkt->size != spkt->size) {
596 trace_colo_compare_main("Other: payload size of packets are different");
597 return -1;
599 return colo_compare_packet_payload(ppkt, spkt, offset, offset,
600 ppkt->size - offset);
603 static int colo_old_packet_check_one(Packet *pkt, int64_t *check_time)
605 int64_t now = qemu_clock_get_ms(QEMU_CLOCK_HOST);
607 if ((now - pkt->creation_ms) > (*check_time)) {
608 trace_colo_old_packet_check_found(pkt->creation_ms);
609 return 0;
610 } else {
611 return 1;
615 void colo_compare_register_notifier(Notifier *notify)
617 notifier_list_add(&colo_compare_notifiers, notify);
620 void colo_compare_unregister_notifier(Notifier *notify)
622 notifier_remove(notify);
625 static int colo_old_packet_check_one_conn(Connection *conn,
626 CompareState *s)
628 if (!g_queue_is_empty(&conn->primary_list)) {
629 if (g_queue_find_custom(&conn->primary_list,
630 &s->compare_timeout,
631 (GCompareFunc)colo_old_packet_check_one))
632 goto out;
635 if (!g_queue_is_empty(&conn->secondary_list)) {
636 if (g_queue_find_custom(&conn->secondary_list,
637 &s->compare_timeout,
638 (GCompareFunc)colo_old_packet_check_one))
639 goto out;
642 return 1;
644 out:
645 /* Do checkpoint will flush old packet */
646 colo_compare_inconsistency_notify(s);
647 return 0;
651 * Look for old packets that the secondary hasn't matched,
652 * if we have some then we have to checkpoint to wake
653 * the secondary up.
655 static void colo_old_packet_check(void *opaque)
657 CompareState *s = opaque;
660 * If we find one old packet, stop finding job and notify
661 * COLO frame do checkpoint.
663 g_queue_find_custom(&s->conn_list, s,
664 (GCompareFunc)colo_old_packet_check_one_conn);
667 static void colo_compare_packet(CompareState *s, Connection *conn,
668 int (*HandlePacket)(Packet *spkt,
669 Packet *ppkt))
671 Packet *pkt = NULL;
672 GList *result = NULL;
674 while (!g_queue_is_empty(&conn->primary_list) &&
675 !g_queue_is_empty(&conn->secondary_list)) {
676 pkt = g_queue_pop_head(&conn->primary_list);
677 result = g_queue_find_custom(&conn->secondary_list,
678 pkt, (GCompareFunc)HandlePacket);
680 if (result) {
681 colo_release_primary_pkt(s, pkt);
682 packet_destroy(result->data, NULL);
683 g_queue_delete_link(&conn->secondary_list, result);
684 } else {
686 * If one packet arrive late, the secondary_list or
687 * primary_list will be empty, so we can't compare it
688 * until next comparison. If the packets in the list are
689 * timeout, it will trigger a checkpoint request.
691 trace_colo_compare_main("packet different");
692 g_queue_push_head(&conn->primary_list, pkt);
694 colo_compare_inconsistency_notify(s);
695 break;
701 * Called from the compare thread on the primary
702 * for compare packet with secondary list of the
703 * specified connection when a new packet was
704 * queued to it.
706 static void colo_compare_connection(void *opaque, void *user_data)
708 CompareState *s = user_data;
709 Connection *conn = opaque;
711 switch (conn->ip_proto) {
712 case IPPROTO_TCP:
713 colo_compare_tcp(s, conn);
714 break;
715 case IPPROTO_UDP:
716 colo_compare_packet(s, conn, colo_packet_compare_udp);
717 break;
718 case IPPROTO_ICMP:
719 colo_compare_packet(s, conn, colo_packet_compare_icmp);
720 break;
721 default:
722 colo_compare_packet(s, conn, colo_packet_compare_other);
723 break;
727 static void coroutine_fn _compare_chr_send(void *opaque)
729 SendCo *sendco = opaque;
730 CompareState *s = sendco->s;
731 int ret = 0;
733 while (!g_queue_is_empty(&sendco->send_list)) {
734 SendEntry *entry = g_queue_pop_tail(&sendco->send_list);
735 uint32_t len = htonl(entry->size);
737 ret = qemu_chr_fe_write_all(sendco->chr, (uint8_t *)&len, sizeof(len));
739 if (ret != sizeof(len)) {
740 g_free(entry->buf);
741 g_slice_free(SendEntry, entry);
742 goto err;
745 if (!sendco->notify_remote_frame && s->vnet_hdr) {
747 * We send vnet header len make other module(like filter-redirector)
748 * know how to parse net packet correctly.
750 len = htonl(entry->vnet_hdr_len);
752 ret = qemu_chr_fe_write_all(sendco->chr,
753 (uint8_t *)&len,
754 sizeof(len));
756 if (ret != sizeof(len)) {
757 g_free(entry->buf);
758 g_slice_free(SendEntry, entry);
759 goto err;
763 ret = qemu_chr_fe_write_all(sendco->chr,
764 (uint8_t *)entry->buf,
765 entry->size);
767 if (ret != entry->size) {
768 g_free(entry->buf);
769 g_slice_free(SendEntry, entry);
770 goto err;
773 g_free(entry->buf);
774 g_slice_free(SendEntry, entry);
777 sendco->ret = 0;
778 goto out;
780 err:
781 while (!g_queue_is_empty(&sendco->send_list)) {
782 SendEntry *entry = g_queue_pop_tail(&sendco->send_list);
783 g_free(entry->buf);
784 g_slice_free(SendEntry, entry);
786 sendco->ret = ret < 0 ? ret : -EIO;
787 out:
788 sendco->co = NULL;
789 sendco->done = true;
790 aio_wait_kick();
793 static int compare_chr_send(CompareState *s,
794 uint8_t *buf,
795 uint32_t size,
796 uint32_t vnet_hdr_len,
797 bool notify_remote_frame,
798 bool zero_copy)
800 SendCo *sendco;
801 SendEntry *entry;
803 if (notify_remote_frame) {
804 sendco = &s->notify_sendco;
805 } else {
806 sendco = &s->out_sendco;
809 if (!size) {
810 return -1;
813 entry = g_slice_new(SendEntry);
814 entry->size = size;
815 entry->vnet_hdr_len = vnet_hdr_len;
816 if (zero_copy) {
817 entry->buf = buf;
818 } else {
819 entry->buf = g_malloc(size);
820 memcpy(entry->buf, buf, size);
822 g_queue_push_head(&sendco->send_list, entry);
824 if (sendco->done) {
825 sendco->co = qemu_coroutine_create(_compare_chr_send, sendco);
826 sendco->done = false;
827 qemu_coroutine_enter(sendco->co);
828 if (sendco->done) {
829 /* report early errors */
830 return sendco->ret;
834 /* assume success */
835 return 0;
838 static int compare_chr_can_read(void *opaque)
840 return COMPARE_READ_LEN_MAX;
844 * Called from the main thread on the primary for packets
845 * arriving over the socket from the primary.
847 static void compare_pri_chr_in(void *opaque, const uint8_t *buf, int size)
849 CompareState *s = COLO_COMPARE(opaque);
850 int ret;
852 ret = net_fill_rstate(&s->pri_rs, buf, size);
853 if (ret == -1) {
854 qemu_chr_fe_set_handlers(&s->chr_pri_in, NULL, NULL, NULL, NULL,
855 NULL, NULL, true);
856 error_report("colo-compare primary_in error");
861 * Called from the main thread on the primary for packets
862 * arriving over the socket from the secondary.
864 static void compare_sec_chr_in(void *opaque, const uint8_t *buf, int size)
866 CompareState *s = COLO_COMPARE(opaque);
867 int ret;
869 ret = net_fill_rstate(&s->sec_rs, buf, size);
870 if (ret == -1) {
871 qemu_chr_fe_set_handlers(&s->chr_sec_in, NULL, NULL, NULL, NULL,
872 NULL, NULL, true);
873 error_report("colo-compare secondary_in error");
877 static void compare_notify_chr(void *opaque, const uint8_t *buf, int size)
879 CompareState *s = COLO_COMPARE(opaque);
880 int ret;
882 ret = net_fill_rstate(&s->notify_rs, buf, size);
883 if (ret == -1) {
884 qemu_chr_fe_set_handlers(&s->chr_notify_dev, NULL, NULL, NULL, NULL,
885 NULL, NULL, true);
886 error_report("colo-compare notify_dev error");
891 * Check old packet regularly so it can watch for any packets
892 * that the secondary hasn't produced equivalents of.
894 static void check_old_packet_regular(void *opaque)
896 CompareState *s = opaque;
898 /* if have old packet we will notify checkpoint */
899 colo_old_packet_check(s);
900 timer_mod(s->packet_check_timer, qemu_clock_get_ms(QEMU_CLOCK_HOST) +
901 s->expired_scan_cycle);
904 /* Public API, Used for COLO frame to notify compare event */
905 void colo_notify_compares_event(void *opaque, int event, Error **errp)
907 CompareState *s;
908 qemu_mutex_lock(&colo_compare_mutex);
910 if (!colo_compare_active) {
911 qemu_mutex_unlock(&colo_compare_mutex);
912 return;
915 qemu_mutex_lock(&event_mtx);
916 QTAILQ_FOREACH(s, &net_compares, next) {
917 s->event = event;
918 qemu_bh_schedule(s->event_bh);
919 event_unhandled_count++;
921 /* Wait all compare threads to finish handling this event */
922 while (event_unhandled_count > 0) {
923 qemu_cond_wait(&event_complete_cond, &event_mtx);
926 qemu_mutex_unlock(&event_mtx);
927 qemu_mutex_unlock(&colo_compare_mutex);
930 static void colo_compare_timer_init(CompareState *s)
932 AioContext *ctx = iothread_get_aio_context(s->iothread);
934 s->packet_check_timer = aio_timer_new(ctx, QEMU_CLOCK_HOST,
935 SCALE_MS, check_old_packet_regular,
937 timer_mod(s->packet_check_timer, qemu_clock_get_ms(QEMU_CLOCK_HOST) +
938 s->expired_scan_cycle);
941 static void colo_compare_timer_del(CompareState *s)
943 if (s->packet_check_timer) {
944 timer_free(s->packet_check_timer);
945 s->packet_check_timer = NULL;
949 static void colo_flush_packets(void *opaque, void *user_data);
951 static void colo_compare_handle_event(void *opaque)
953 CompareState *s = opaque;
955 switch (s->event) {
956 case COLO_EVENT_CHECKPOINT:
957 g_queue_foreach(&s->conn_list, colo_flush_packets, s);
958 break;
959 case COLO_EVENT_FAILOVER:
960 break;
961 default:
962 break;
965 qemu_mutex_lock(&event_mtx);
966 assert(event_unhandled_count > 0);
967 event_unhandled_count--;
968 qemu_cond_broadcast(&event_complete_cond);
969 qemu_mutex_unlock(&event_mtx);
972 static void colo_compare_iothread(CompareState *s)
974 AioContext *ctx = iothread_get_aio_context(s->iothread);
975 object_ref(OBJECT(s->iothread));
976 s->worker_context = iothread_get_g_main_context(s->iothread);
978 qemu_chr_fe_set_handlers(&s->chr_pri_in, compare_chr_can_read,
979 compare_pri_chr_in, NULL, NULL,
980 s, s->worker_context, true);
981 qemu_chr_fe_set_handlers(&s->chr_sec_in, compare_chr_can_read,
982 compare_sec_chr_in, NULL, NULL,
983 s, s->worker_context, true);
984 if (s->notify_dev) {
985 qemu_chr_fe_set_handlers(&s->chr_notify_dev, compare_chr_can_read,
986 compare_notify_chr, NULL, NULL,
987 s, s->worker_context, true);
990 colo_compare_timer_init(s);
991 s->event_bh = aio_bh_new(ctx, colo_compare_handle_event, s);
994 static char *compare_get_pri_indev(Object *obj, Error **errp)
996 CompareState *s = COLO_COMPARE(obj);
998 return g_strdup(s->pri_indev);
1001 static void compare_set_pri_indev(Object *obj, const char *value, Error **errp)
1003 CompareState *s = COLO_COMPARE(obj);
1005 g_free(s->pri_indev);
1006 s->pri_indev = g_strdup(value);
1009 static char *compare_get_sec_indev(Object *obj, Error **errp)
1011 CompareState *s = COLO_COMPARE(obj);
1013 return g_strdup(s->sec_indev);
1016 static void compare_set_sec_indev(Object *obj, const char *value, Error **errp)
1018 CompareState *s = COLO_COMPARE(obj);
1020 g_free(s->sec_indev);
1021 s->sec_indev = g_strdup(value);
1024 static char *compare_get_outdev(Object *obj, Error **errp)
1026 CompareState *s = COLO_COMPARE(obj);
1028 return g_strdup(s->outdev);
1031 static void compare_set_outdev(Object *obj, const char *value, Error **errp)
1033 CompareState *s = COLO_COMPARE(obj);
1035 g_free(s->outdev);
1036 s->outdev = g_strdup(value);
1039 static bool compare_get_vnet_hdr(Object *obj, Error **errp)
1041 CompareState *s = COLO_COMPARE(obj);
1043 return s->vnet_hdr;
1046 static void compare_set_vnet_hdr(Object *obj,
1047 bool value,
1048 Error **errp)
1050 CompareState *s = COLO_COMPARE(obj);
1052 s->vnet_hdr = value;
1055 static char *compare_get_notify_dev(Object *obj, Error **errp)
1057 CompareState *s = COLO_COMPARE(obj);
1059 return g_strdup(s->notify_dev);
1062 static void compare_set_notify_dev(Object *obj, const char *value, Error **errp)
1064 CompareState *s = COLO_COMPARE(obj);
1066 g_free(s->notify_dev);
1067 s->notify_dev = g_strdup(value);
1070 static void compare_get_timeout(Object *obj, Visitor *v,
1071 const char *name, void *opaque,
1072 Error **errp)
1074 CompareState *s = COLO_COMPARE(obj);
1075 uint64_t value = s->compare_timeout;
1077 visit_type_uint64(v, name, &value, errp);
1080 static void compare_set_timeout(Object *obj, Visitor *v,
1081 const char *name, void *opaque,
1082 Error **errp)
1084 CompareState *s = COLO_COMPARE(obj);
1085 uint32_t value;
1087 if (!visit_type_uint32(v, name, &value, errp)) {
1088 return;
1090 if (!value) {
1091 error_setg(errp, "Property '%s.%s' requires a positive value",
1092 object_get_typename(obj), name);
1093 return;
1095 s->compare_timeout = value;
1098 static void compare_get_expired_scan_cycle(Object *obj, Visitor *v,
1099 const char *name, void *opaque,
1100 Error **errp)
1102 CompareState *s = COLO_COMPARE(obj);
1103 uint32_t value = s->expired_scan_cycle;
1105 visit_type_uint32(v, name, &value, errp);
1108 static void compare_set_expired_scan_cycle(Object *obj, Visitor *v,
1109 const char *name, void *opaque,
1110 Error **errp)
1112 CompareState *s = COLO_COMPARE(obj);
1113 uint32_t value;
1115 if (!visit_type_uint32(v, name, &value, errp)) {
1116 return;
1118 if (!value) {
1119 error_setg(errp, "Property '%s.%s' requires a positive value",
1120 object_get_typename(obj), name);
1121 return;
1123 s->expired_scan_cycle = value;
1126 static void get_max_queue_size(Object *obj, Visitor *v,
1127 const char *name, void *opaque,
1128 Error **errp)
1130 uint32_t value = max_queue_size;
1132 visit_type_uint32(v, name, &value, errp);
1135 static void set_max_queue_size(Object *obj, Visitor *v,
1136 const char *name, void *opaque,
1137 Error **errp)
1139 Error *local_err = NULL;
1140 uint64_t value;
1142 visit_type_uint64(v, name, &value, &local_err);
1143 if (local_err) {
1144 goto out;
1146 if (!value) {
1147 error_setg(&local_err, "Property '%s.%s' requires a positive value",
1148 object_get_typename(obj), name);
1149 goto out;
1151 max_queue_size = value;
1153 out:
1154 error_propagate(errp, local_err);
1157 static void compare_pri_rs_finalize(SocketReadState *pri_rs)
1159 CompareState *s = container_of(pri_rs, CompareState, pri_rs);
1160 Connection *conn = NULL;
1162 if (packet_enqueue(s, PRIMARY_IN, &conn)) {
1163 trace_colo_compare_main("primary: unsupported packet in");
1164 compare_chr_send(s,
1165 pri_rs->buf,
1166 pri_rs->packet_len,
1167 pri_rs->vnet_hdr_len,
1168 false,
1169 false);
1170 } else {
1171 /* compare packet in the specified connection */
1172 colo_compare_connection(conn, s);
1176 static void compare_sec_rs_finalize(SocketReadState *sec_rs)
1178 CompareState *s = container_of(sec_rs, CompareState, sec_rs);
1179 Connection *conn = NULL;
1181 if (packet_enqueue(s, SECONDARY_IN, &conn)) {
1182 trace_colo_compare_main("secondary: unsupported packet in");
1183 } else {
1184 /* compare packet in the specified connection */
1185 colo_compare_connection(conn, s);
1189 static void compare_notify_rs_finalize(SocketReadState *notify_rs)
1191 CompareState *s = container_of(notify_rs, CompareState, notify_rs);
1193 const char msg[] = "COLO_COMPARE_GET_XEN_INIT";
1194 int ret;
1196 if (packet_matches_str("COLO_USERSPACE_PROXY_INIT",
1197 notify_rs->buf,
1198 notify_rs->packet_len)) {
1199 ret = compare_chr_send(s, (uint8_t *)msg, strlen(msg), 0, true, false);
1200 if (ret < 0) {
1201 error_report("Notify Xen COLO-frame INIT failed");
1203 } else if (packet_matches_str("COLO_CHECKPOINT",
1204 notify_rs->buf,
1205 notify_rs->packet_len)) {
1206 /* colo-compare do checkpoint, flush pri packet and remove sec packet */
1207 g_queue_foreach(&s->conn_list, colo_flush_packets, s);
1208 } else {
1209 error_report("COLO compare got unsupported instruction");
1214 * Return 0 is success.
1215 * Return 1 is failed.
1217 static int find_and_check_chardev(Chardev **chr,
1218 char *chr_name,
1219 Error **errp)
1221 *chr = qemu_chr_find(chr_name);
1222 if (*chr == NULL) {
1223 error_setg(errp, "Device '%s' not found",
1224 chr_name);
1225 return 1;
1228 if (!qemu_chr_has_feature(*chr, QEMU_CHAR_FEATURE_RECONNECTABLE)) {
1229 error_setg(errp, "chardev \"%s\" is not reconnectable",
1230 chr_name);
1231 return 1;
1234 if (!qemu_chr_has_feature(*chr, QEMU_CHAR_FEATURE_GCONTEXT)) {
1235 error_setg(errp, "chardev \"%s\" cannot switch context",
1236 chr_name);
1237 return 1;
1240 return 0;
1244 * Called from the main thread on the primary
1245 * to setup colo-compare.
1247 static void colo_compare_complete(UserCreatable *uc, Error **errp)
1249 CompareState *s = COLO_COMPARE(uc);
1250 Chardev *chr;
1252 if (!s->pri_indev || !s->sec_indev || !s->outdev || !s->iothread) {
1253 error_setg(errp, "colo compare needs 'primary_in' ,"
1254 "'secondary_in','outdev','iothread' property set");
1255 return;
1256 } else if (!strcmp(s->pri_indev, s->outdev) ||
1257 !strcmp(s->sec_indev, s->outdev) ||
1258 !strcmp(s->pri_indev, s->sec_indev)) {
1259 error_setg(errp, "'indev' and 'outdev' could not be same "
1260 "for compare module");
1261 return;
1264 if (!s->compare_timeout) {
1265 /* Set default value to 3000 MS */
1266 s->compare_timeout = DEFAULT_TIME_OUT_MS;
1269 if (!s->expired_scan_cycle) {
1270 /* Set default value to 3000 MS */
1271 s->expired_scan_cycle = REGULAR_PACKET_CHECK_MS;
1274 if (!max_queue_size) {
1275 /* Set default queue size to 1024 */
1276 max_queue_size = MAX_QUEUE_SIZE;
1279 if (find_and_check_chardev(&chr, s->pri_indev, errp) ||
1280 !qemu_chr_fe_init(&s->chr_pri_in, chr, errp)) {
1281 return;
1284 if (find_and_check_chardev(&chr, s->sec_indev, errp) ||
1285 !qemu_chr_fe_init(&s->chr_sec_in, chr, errp)) {
1286 return;
1289 if (find_and_check_chardev(&chr, s->outdev, errp) ||
1290 !qemu_chr_fe_init(&s->chr_out, chr, errp)) {
1291 return;
1294 net_socket_rs_init(&s->pri_rs, compare_pri_rs_finalize, s->vnet_hdr);
1295 net_socket_rs_init(&s->sec_rs, compare_sec_rs_finalize, s->vnet_hdr);
1297 /* Try to enable remote notify chardev, currently just for Xen COLO */
1298 if (s->notify_dev) {
1299 if (find_and_check_chardev(&chr, s->notify_dev, errp) ||
1300 !qemu_chr_fe_init(&s->chr_notify_dev, chr, errp)) {
1301 return;
1304 net_socket_rs_init(&s->notify_rs, compare_notify_rs_finalize,
1305 s->vnet_hdr);
1308 s->out_sendco.s = s;
1309 s->out_sendco.chr = &s->chr_out;
1310 s->out_sendco.notify_remote_frame = false;
1311 s->out_sendco.done = true;
1312 g_queue_init(&s->out_sendco.send_list);
1314 if (s->notify_dev) {
1315 s->notify_sendco.s = s;
1316 s->notify_sendco.chr = &s->chr_notify_dev;
1317 s->notify_sendco.notify_remote_frame = true;
1318 s->notify_sendco.done = true;
1319 g_queue_init(&s->notify_sendco.send_list);
1322 g_queue_init(&s->conn_list);
1324 s->connection_track_table = g_hash_table_new_full(connection_key_hash,
1325 connection_key_equal,
1326 g_free,
1327 connection_destroy);
1329 colo_compare_iothread(s);
1331 qemu_mutex_lock(&colo_compare_mutex);
1332 if (!colo_compare_active) {
1333 qemu_mutex_init(&event_mtx);
1334 qemu_cond_init(&event_complete_cond);
1335 colo_compare_active = true;
1337 QTAILQ_INSERT_TAIL(&net_compares, s, next);
1338 qemu_mutex_unlock(&colo_compare_mutex);
1340 return;
1343 static void colo_flush_packets(void *opaque, void *user_data)
1345 CompareState *s = user_data;
1346 Connection *conn = opaque;
1347 Packet *pkt = NULL;
1349 while (!g_queue_is_empty(&conn->primary_list)) {
1350 pkt = g_queue_pop_head(&conn->primary_list);
1351 compare_chr_send(s,
1352 pkt->data,
1353 pkt->size,
1354 pkt->vnet_hdr_len,
1355 false,
1356 true);
1357 packet_destroy_partial(pkt, NULL);
1359 while (!g_queue_is_empty(&conn->secondary_list)) {
1360 pkt = g_queue_pop_head(&conn->secondary_list);
1361 packet_destroy(pkt, NULL);
1365 static void colo_compare_class_init(ObjectClass *oc, void *data)
1367 UserCreatableClass *ucc = USER_CREATABLE_CLASS(oc);
1369 ucc->complete = colo_compare_complete;
1372 static void colo_compare_init(Object *obj)
1374 CompareState *s = COLO_COMPARE(obj);
1376 object_property_add_str(obj, "primary_in",
1377 compare_get_pri_indev, compare_set_pri_indev);
1378 object_property_add_str(obj, "secondary_in",
1379 compare_get_sec_indev, compare_set_sec_indev);
1380 object_property_add_str(obj, "outdev",
1381 compare_get_outdev, compare_set_outdev);
1382 object_property_add_link(obj, "iothread", TYPE_IOTHREAD,
1383 (Object **)&s->iothread,
1384 object_property_allow_set_link,
1385 OBJ_PROP_LINK_STRONG);
1386 /* This parameter just for Xen COLO */
1387 object_property_add_str(obj, "notify_dev",
1388 compare_get_notify_dev, compare_set_notify_dev);
1390 object_property_add(obj, "compare_timeout", "uint64",
1391 compare_get_timeout,
1392 compare_set_timeout, NULL, NULL);
1394 object_property_add(obj, "expired_scan_cycle", "uint32",
1395 compare_get_expired_scan_cycle,
1396 compare_set_expired_scan_cycle, NULL, NULL);
1398 object_property_add(obj, "max_queue_size", "uint32",
1399 get_max_queue_size,
1400 set_max_queue_size, NULL, NULL);
1402 s->vnet_hdr = false;
1403 object_property_add_bool(obj, "vnet_hdr_support", compare_get_vnet_hdr,
1404 compare_set_vnet_hdr);
1407 void colo_compare_cleanup(void)
1409 CompareState *tmp = NULL;
1410 CompareState *n = NULL;
1412 QTAILQ_FOREACH_SAFE(tmp, &net_compares, next, n) {
1413 object_unparent(OBJECT(tmp));
1417 static void colo_compare_finalize(Object *obj)
1419 CompareState *s = COLO_COMPARE(obj);
1420 CompareState *tmp = NULL;
1422 qemu_mutex_lock(&colo_compare_mutex);
1423 QTAILQ_FOREACH(tmp, &net_compares, next) {
1424 if (tmp == s) {
1425 QTAILQ_REMOVE(&net_compares, s, next);
1426 break;
1429 if (QTAILQ_EMPTY(&net_compares)) {
1430 colo_compare_active = false;
1431 qemu_mutex_destroy(&event_mtx);
1432 qemu_cond_destroy(&event_complete_cond);
1434 qemu_mutex_unlock(&colo_compare_mutex);
1436 qemu_chr_fe_deinit(&s->chr_pri_in, false);
1437 qemu_chr_fe_deinit(&s->chr_sec_in, false);
1438 qemu_chr_fe_deinit(&s->chr_out, false);
1439 if (s->notify_dev) {
1440 qemu_chr_fe_deinit(&s->chr_notify_dev, false);
1443 colo_compare_timer_del(s);
1445 qemu_bh_delete(s->event_bh);
1447 AioContext *ctx = iothread_get_aio_context(s->iothread);
1448 aio_context_acquire(ctx);
1449 AIO_WAIT_WHILE(ctx, !s->out_sendco.done);
1450 if (s->notify_dev) {
1451 AIO_WAIT_WHILE(ctx, !s->notify_sendco.done);
1453 aio_context_release(ctx);
1455 /* Release all unhandled packets after compare thead exited */
1456 g_queue_foreach(&s->conn_list, colo_flush_packets, s);
1457 AIO_WAIT_WHILE(NULL, !s->out_sendco.done);
1459 g_queue_clear(&s->conn_list);
1460 g_queue_clear(&s->out_sendco.send_list);
1461 if (s->notify_dev) {
1462 g_queue_clear(&s->notify_sendco.send_list);
1465 if (s->connection_track_table) {
1466 g_hash_table_destroy(s->connection_track_table);
1469 object_unref(OBJECT(s->iothread));
1471 g_free(s->pri_indev);
1472 g_free(s->sec_indev);
1473 g_free(s->outdev);
1474 g_free(s->notify_dev);
1477 static void __attribute__((__constructor__)) colo_compare_init_globals(void)
1479 colo_compare_active = false;
1480 qemu_mutex_init(&colo_compare_mutex);
1483 static const TypeInfo colo_compare_info = {
1484 .name = TYPE_COLO_COMPARE,
1485 .parent = TYPE_OBJECT,
1486 .instance_size = sizeof(CompareState),
1487 .instance_init = colo_compare_init,
1488 .instance_finalize = colo_compare_finalize,
1489 .class_size = sizeof(CompareClass),
1490 .class_init = colo_compare_class_init,
1491 .interfaces = (InterfaceInfo[]) {
1492 { TYPE_USER_CREATABLE },
1497 static void register_types(void)
1499 type_register_static(&colo_compare_info);
1502 type_init(register_types);