]>
Commit | Line | Data |
---|---|---|
7dce4e6f ZC |
1 | /* |
2 | * COarse-grain LOck-stepping Virtual Machines for Non-stop Service (COLO) | |
3 | * (a.k.a. Fault Tolerance or Continuous Replication) | |
4 | * | |
5 | * Copyright (c) 2016 HUAWEI TECHNOLOGIES CO., LTD. | |
6 | * Copyright (c) 2016 FUJITSU LIMITED | |
7 | * Copyright (c) 2016 Intel Corporation | |
8 | * | |
9 | * Author: Zhang Chen <zhangchen.fnst@cn.fujitsu.com> | |
10 | * | |
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. | |
13 | */ | |
14 | ||
15 | #include "qemu/osdep.h" | |
a8d25326 | 16 | #include "qemu-common.h" |
7dce4e6f | 17 | #include "qemu/error-report.h" |
59509ec1 | 18 | #include "trace.h" |
7dce4e6f ZC |
19 | #include "qapi/error.h" |
20 | #include "net/net.h" | |
f4b61836 | 21 | #include "net/eth.h" |
7dce4e6f ZC |
22 | #include "qom/object_interfaces.h" |
23 | #include "qemu/iov.h" | |
24 | #include "qom/object.h" | |
7dce4e6f | 25 | #include "net/queue.h" |
4d43a603 | 26 | #include "chardev/char-fe.h" |
7dce4e6f | 27 | #include "qemu/sockets.h" |
f27f01db | 28 | #include "colo.h" |
dd321ecf | 29 | #include "sysemu/iothread.h" |
0ffcece3 ZC |
30 | #include "net/colo-compare.h" |
31 | #include "migration/colo.h" | |
dccd0313 | 32 | #include "migration/migration.h" |
e05ae1d9 | 33 | #include "util.h" |
7dce4e6f ZC |
34 | |
35 | #define TYPE_COLO_COMPARE "colo-compare" | |
36 | #define COLO_COMPARE(obj) \ | |
37 | OBJECT_CHECK(CompareState, (obj), TYPE_COLO_COMPARE) | |
38 | ||
0ffcece3 ZC |
39 | static QTAILQ_HEAD(, CompareState) net_compares = |
40 | QTAILQ_HEAD_INITIALIZER(net_compares); | |
41 | ||
dccd0313 ZC |
42 | static NotifierList colo_compare_notifiers = |
43 | NOTIFIER_LIST_INITIALIZER(colo_compare_notifiers); | |
44 | ||
0682e15b | 45 | #define COMPARE_READ_LEN_MAX NET_BUFSIZE |
b6540d40 ZC |
46 | #define MAX_QUEUE_SIZE 1024 |
47 | ||
f449c9e5 MZ |
48 | #define COLO_COMPARE_FREE_PRIMARY 0x01 |
49 | #define COLO_COMPARE_FREE_SECONDARY 0x02 | |
50 | ||
0682e15b ZC |
51 | /* TODO: Should be configurable */ |
52 | #define REGULAR_PACKET_CHECK_MS 3000 | |
53 | ||
0ffcece3 ZC |
54 | static QemuMutex event_mtx; |
55 | static QemuCond event_complete_cond; | |
56 | static int event_unhandled_count; | |
57 | ||
59509ec1 | 58 | /* |
61c5f469 ZC |
59 | * + CompareState ++ |
60 | * | | | |
61 | * +---------------+ +---------------+ +---------------+ | |
62 | * | conn list + - > conn + ------- > conn + -- > ...... | |
63 | * +---------------+ +---------------+ +---------------+ | |
64 | * | | | | | | | |
65 | * +---------------+ +---v----+ +---v----+ +---v----+ +---v----+ | |
66 | * |primary | |secondary |primary | |secondary | |
67 | * |packet | |packet + |packet | |packet + | |
68 | * +--------+ +--------+ +--------+ +--------+ | |
69 | * | | | | | |
70 | * +---v----+ +---v----+ +---v----+ +---v----+ | |
71 | * |primary | |secondary |primary | |secondary | |
72 | * |packet | |packet + |packet | |packet + | |
73 | * +--------+ +--------+ +--------+ +--------+ | |
74 | * | | | | | |
75 | * +---v----+ +---v----+ +---v----+ +---v----+ | |
76 | * |primary | |secondary |primary | |secondary | |
77 | * |packet | |packet + |packet | |packet + | |
78 | * +--------+ +--------+ +--------+ +--------+ | |
79 | */ | |
7dce4e6f ZC |
80 | typedef struct CompareState { |
81 | Object parent; | |
82 | ||
83 | char *pri_indev; | |
84 | char *sec_indev; | |
85 | char *outdev; | |
cf6af766 | 86 | char *notify_dev; |
32a6ebec MAL |
87 | CharBackend chr_pri_in; |
88 | CharBackend chr_sec_in; | |
89 | CharBackend chr_out; | |
13025fee | 90 | CharBackend chr_notify_dev; |
7dce4e6f ZC |
91 | SocketReadState pri_rs; |
92 | SocketReadState sec_rs; | |
13025fee | 93 | SocketReadState notify_rs; |
aa3a7032 | 94 | bool vnet_hdr; |
59509ec1 | 95 | |
61c5f469 ZC |
96 | /* |
97 | * Record the connection that through the NIC | |
98 | * Element type: Connection | |
b6540d40 ZC |
99 | */ |
100 | GQueue conn_list; | |
61c5f469 | 101 | /* Record the connection without repetition */ |
59509ec1 | 102 | GHashTable *connection_track_table; |
dfd917a9 | 103 | |
dd321ecf | 104 | IOThread *iothread; |
b43decb0 | 105 | GMainContext *worker_context; |
dd321ecf | 106 | QEMUTimer *packet_check_timer; |
0ffcece3 ZC |
107 | |
108 | QEMUBH *event_bh; | |
109 | enum colo_event event; | |
110 | ||
111 | QTAILQ_ENTRY(CompareState) next; | |
7dce4e6f ZC |
112 | } CompareState; |
113 | ||
114 | typedef struct CompareClass { | |
115 | ObjectClass parent_class; | |
116 | } CompareClass; | |
117 | ||
59509ec1 ZC |
118 | enum { |
119 | PRIMARY_IN = 0, | |
120 | SECONDARY_IN, | |
121 | }; | |
122 | ||
24525e93 | 123 | |
3037e7a5 | 124 | static int compare_chr_send(CompareState *s, |
59509ec1 | 125 | const uint8_t *buf, |
aa3a7032 | 126 | uint32_t size, |
30685c00 ZC |
127 | uint32_t vnet_hdr_len, |
128 | bool notify_remote_frame); | |
59509ec1 | 129 | |
1d09f700 ZC |
130 | static void notify_remote_frame(CompareState *s) |
131 | { | |
132 | char msg[] = "DO_CHECKPOINT"; | |
133 | int ret = 0; | |
134 | ||
135 | ret = compare_chr_send(s, (uint8_t *)msg, strlen(msg), 0, true); | |
136 | if (ret < 0) { | |
137 | error_report("Notify Xen COLO-frame failed"); | |
138 | } | |
139 | } | |
140 | ||
141 | static void colo_compare_inconsistency_notify(CompareState *s) | |
142 | { | |
143 | if (s->notify_dev) { | |
144 | notify_remote_frame(s); | |
145 | } else { | |
146 | notifier_list_notify(&colo_compare_notifiers, | |
147 | migrate_get_current()); | |
148 | } | |
149 | } | |
150 | ||
a935cc31 ZC |
151 | static gint seq_sorter(Packet *a, Packet *b, gpointer data) |
152 | { | |
e05ae1d9 | 153 | struct tcp_hdr *atcp, *btcp; |
a935cc31 | 154 | |
e05ae1d9 MAL |
155 | atcp = (struct tcp_hdr *)(a->transport_header); |
156 | btcp = (struct tcp_hdr *)(b->transport_header); | |
a935cc31 ZC |
157 | return ntohl(atcp->th_seq) - ntohl(btcp->th_seq); |
158 | } | |
159 | ||
f449c9e5 MZ |
160 | static void fill_pkt_tcp_info(void *data, uint32_t *max_ack) |
161 | { | |
162 | Packet *pkt = data; | |
e05ae1d9 | 163 | struct tcp_hdr *tcphd; |
f449c9e5 | 164 | |
e05ae1d9 | 165 | tcphd = (struct tcp_hdr *)pkt->transport_header; |
f449c9e5 MZ |
166 | |
167 | pkt->tcp_seq = ntohl(tcphd->th_seq); | |
168 | pkt->tcp_ack = ntohl(tcphd->th_ack); | |
169 | *max_ack = *max_ack > pkt->tcp_ack ? *max_ack : pkt->tcp_ack; | |
170 | pkt->header_size = pkt->transport_header - (uint8_t *)pkt->data | |
171 | + (tcphd->th_off << 2) - pkt->vnet_hdr_len; | |
172 | pkt->payload_size = pkt->size - pkt->header_size; | |
173 | pkt->seq_end = pkt->tcp_seq + pkt->payload_size; | |
174 | pkt->flags = tcphd->th_flags; | |
175 | } | |
176 | ||
8850d4ca MZ |
177 | /* |
178 | * Return 1 on success, if return 0 means the | |
179 | * packet will be dropped | |
180 | */ | |
f449c9e5 | 181 | static int colo_insert_packet(GQueue *queue, Packet *pkt, uint32_t *max_ack) |
8850d4ca MZ |
182 | { |
183 | if (g_queue_get_length(queue) <= MAX_QUEUE_SIZE) { | |
184 | if (pkt->ip->ip_p == IPPROTO_TCP) { | |
f449c9e5 | 185 | fill_pkt_tcp_info(pkt, max_ack); |
8850d4ca MZ |
186 | g_queue_insert_sorted(queue, |
187 | pkt, | |
188 | (GCompareDataFunc)seq_sorter, | |
189 | NULL); | |
190 | } else { | |
191 | g_queue_push_tail(queue, pkt); | |
192 | } | |
193 | return 1; | |
194 | } | |
195 | return 0; | |
196 | } | |
197 | ||
59509ec1 ZC |
198 | /* |
199 | * Return 0 on success, if return -1 means the pkt | |
200 | * is unsupported(arp and ipv6) and will be sent later | |
201 | */ | |
8ec14402 | 202 | static int packet_enqueue(CompareState *s, int mode, Connection **con) |
59509ec1 | 203 | { |
b6540d40 | 204 | ConnectionKey key; |
59509ec1 | 205 | Packet *pkt = NULL; |
b6540d40 | 206 | Connection *conn; |
59509ec1 ZC |
207 | |
208 | if (mode == PRIMARY_IN) { | |
ada1a33f ZC |
209 | pkt = packet_new(s->pri_rs.buf, |
210 | s->pri_rs.packet_len, | |
211 | s->pri_rs.vnet_hdr_len); | |
59509ec1 | 212 | } else { |
ada1a33f ZC |
213 | pkt = packet_new(s->sec_rs.buf, |
214 | s->sec_rs.packet_len, | |
215 | s->sec_rs.vnet_hdr_len); | |
59509ec1 ZC |
216 | } |
217 | ||
218 | if (parse_packet_early(pkt)) { | |
219 | packet_destroy(pkt, NULL); | |
220 | pkt = NULL; | |
221 | return -1; | |
222 | } | |
b6540d40 | 223 | fill_connection_key(pkt, &key); |
59509ec1 | 224 | |
b6540d40 ZC |
225 | conn = connection_get(s->connection_track_table, |
226 | &key, | |
227 | &s->conn_list); | |
59509ec1 | 228 | |
b6540d40 ZC |
229 | if (!conn->processing) { |
230 | g_queue_push_tail(&s->conn_list, conn); | |
231 | conn->processing = true; | |
232 | } | |
233 | ||
234 | if (mode == PRIMARY_IN) { | |
f449c9e5 | 235 | if (!colo_insert_packet(&conn->primary_list, pkt, &conn->pack)) { |
b6540d40 ZC |
236 | error_report("colo compare primary queue size too big," |
237 | "drop packet"); | |
238 | } | |
239 | } else { | |
f449c9e5 | 240 | if (!colo_insert_packet(&conn->secondary_list, pkt, &conn->sack)) { |
b6540d40 ZC |
241 | error_report("colo compare secondary queue size too big," |
242 | "drop packet"); | |
243 | } | |
244 | } | |
4d366235 | 245 | *con = conn; |
59509ec1 ZC |
246 | |
247 | return 0; | |
248 | } | |
249 | ||
f449c9e5 MZ |
250 | static inline bool after(uint32_t seq1, uint32_t seq2) |
251 | { | |
252 | return (int32_t)(seq1 - seq2) > 0; | |
253 | } | |
254 | ||
255 | static void colo_release_primary_pkt(CompareState *s, Packet *pkt) | |
256 | { | |
257 | int ret; | |
258 | ret = compare_chr_send(s, | |
259 | pkt->data, | |
260 | pkt->size, | |
30685c00 ZC |
261 | pkt->vnet_hdr_len, |
262 | false); | |
f449c9e5 MZ |
263 | if (ret < 0) { |
264 | error_report("colo send primary packet failed"); | |
265 | } | |
266 | trace_colo_compare_main("packet same and release packet"); | |
267 | packet_destroy(pkt, NULL); | |
268 | } | |
269 | ||
0682e15b ZC |
270 | /* |
271 | * The IP packets sent by primary and secondary | |
272 | * will be compared in here | |
273 | * TODO support ip fragment, Out-Of-Order | |
274 | * return: 0 means packet same | |
275 | * > 0 || < 0 means packet different | |
276 | */ | |
9394133f MZ |
277 | static int colo_compare_packet_payload(Packet *ppkt, |
278 | Packet *spkt, | |
279 | uint16_t poffset, | |
280 | uint16_t soffset, | |
281 | uint16_t len) | |
282 | ||
0682e15b | 283 | { |
d87aa138 | 284 | if (trace_event_get_state_backends(TRACE_COLO_COMPARE_MISCOMPARE)) { |
e630b2bf ZC |
285 | char pri_ip_src[20], pri_ip_dst[20], sec_ip_src[20], sec_ip_dst[20]; |
286 | ||
287 | strcpy(pri_ip_src, inet_ntoa(ppkt->ip->ip_src)); | |
288 | strcpy(pri_ip_dst, inet_ntoa(ppkt->ip->ip_dst)); | |
289 | strcpy(sec_ip_src, inet_ntoa(spkt->ip->ip_src)); | |
290 | strcpy(sec_ip_dst, inet_ntoa(spkt->ip->ip_dst)); | |
291 | ||
292 | trace_colo_compare_ip_info(ppkt->size, pri_ip_src, | |
293 | pri_ip_dst, spkt->size, | |
294 | sec_ip_src, sec_ip_dst); | |
295 | } | |
0682e15b | 296 | |
9394133f | 297 | return memcmp(ppkt->data + poffset, spkt->data + soffset, len); |
0682e15b ZC |
298 | } |
299 | ||
f4b61836 | 300 | /* |
f449c9e5 MZ |
301 | * return true means that the payload is consist and |
302 | * need to make the next comparison, false means do | |
303 | * the checkpoint | |
304 | */ | |
305 | static bool colo_mark_tcp_pkt(Packet *ppkt, Packet *spkt, | |
306 | int8_t *mark, uint32_t max_ack) | |
0682e15b | 307 | { |
f449c9e5 MZ |
308 | *mark = 0; |
309 | ||
f449c9e5 MZ |
310 | if (ppkt->tcp_seq == spkt->tcp_seq && ppkt->seq_end == spkt->seq_end) { |
311 | if (colo_compare_packet_payload(ppkt, spkt, | |
312 | ppkt->header_size, spkt->header_size, | |
313 | ppkt->payload_size)) { | |
314 | *mark = COLO_COMPARE_FREE_SECONDARY | COLO_COMPARE_FREE_PRIMARY; | |
315 | return true; | |
316 | } | |
317 | } | |
f4b61836 | 318 | |
f449c9e5 MZ |
319 | /* one part of secondary packet payload still need to be compared */ |
320 | if (!after(ppkt->seq_end, spkt->seq_end)) { | |
321 | if (colo_compare_packet_payload(ppkt, spkt, | |
322 | ppkt->header_size + ppkt->offset, | |
323 | spkt->header_size + spkt->offset, | |
324 | ppkt->payload_size - ppkt->offset)) { | |
325 | if (!after(ppkt->tcp_ack, max_ack)) { | |
326 | *mark = COLO_COMPARE_FREE_PRIMARY; | |
327 | spkt->offset += ppkt->payload_size - ppkt->offset; | |
328 | return true; | |
329 | } else { | |
330 | /* secondary guest hasn't ack the data, don't send | |
331 | * out this packet | |
332 | */ | |
333 | return false; | |
334 | } | |
335 | } | |
336 | } else { | |
337 | /* primary packet is longer than secondary packet, compare | |
338 | * the same part and mark the primary packet offset | |
339 | */ | |
340 | if (colo_compare_packet_payload(ppkt, spkt, | |
341 | ppkt->header_size + ppkt->offset, | |
342 | spkt->header_size + spkt->offset, | |
343 | spkt->payload_size - spkt->offset)) { | |
344 | *mark = COLO_COMPARE_FREE_SECONDARY; | |
345 | ppkt->offset += spkt->payload_size - spkt->offset; | |
346 | return true; | |
347 | } | |
348 | } | |
349 | ||
350 | return false; | |
351 | } | |
2ad7ca4c | 352 | |
f449c9e5 MZ |
353 | static void colo_compare_tcp(CompareState *s, Connection *conn) |
354 | { | |
355 | Packet *ppkt = NULL, *spkt = NULL; | |
356 | int8_t mark; | |
f4b61836 ZC |
357 | |
358 | /* | |
f449c9e5 MZ |
359 | * If ppkt and spkt have the same payload, but ppkt's ACK |
360 | * is greater than spkt's ACK, in this case we can not | |
361 | * send the ppkt because it will cause the secondary guest | |
362 | * to miss sending some data in the next. Therefore, we | |
363 | * record the maximum ACK in the current queue at both | |
364 | * primary side and secondary side. Only when the ack is | |
365 | * less than the smaller of the two maximum ack, then we | |
366 | * can ensure that the packet's payload is acknowledged by | |
367 | * primary and secondary. | |
368 | */ | |
369 | uint32_t min_ack = conn->pack > conn->sack ? conn->sack : conn->pack; | |
370 | ||
371 | pri: | |
372 | if (g_queue_is_empty(&conn->primary_list)) { | |
373 | return; | |
374 | } | |
375 | ppkt = g_queue_pop_head(&conn->primary_list); | |
376 | sec: | |
377 | if (g_queue_is_empty(&conn->secondary_list)) { | |
378 | g_queue_push_head(&conn->primary_list, ppkt); | |
379 | return; | |
f4b61836 | 380 | } |
f449c9e5 | 381 | spkt = g_queue_pop_head(&conn->secondary_list); |
f4b61836 | 382 | |
f449c9e5 MZ |
383 | if (ppkt->tcp_seq == ppkt->seq_end) { |
384 | colo_release_primary_pkt(s, ppkt); | |
385 | ppkt = NULL; | |
386 | } | |
9394133f | 387 | |
f449c9e5 MZ |
388 | if (ppkt && conn->compare_seq && !after(ppkt->seq_end, conn->compare_seq)) { |
389 | trace_colo_compare_main("pri: this packet has compared"); | |
390 | colo_release_primary_pkt(s, ppkt); | |
391 | ppkt = NULL; | |
392 | } | |
9394133f | 393 | |
f449c9e5 MZ |
394 | if (spkt->tcp_seq == spkt->seq_end) { |
395 | packet_destroy(spkt, NULL); | |
396 | if (!ppkt) { | |
397 | goto pri; | |
398 | } else { | |
399 | goto sec; | |
400 | } | |
6efeb328 | 401 | } else { |
f449c9e5 MZ |
402 | if (conn->compare_seq && !after(spkt->seq_end, conn->compare_seq)) { |
403 | trace_colo_compare_main("sec: this packet has compared"); | |
404 | packet_destroy(spkt, NULL); | |
405 | if (!ppkt) { | |
406 | goto pri; | |
407 | } else { | |
408 | goto sec; | |
409 | } | |
410 | } | |
411 | if (!ppkt) { | |
412 | g_queue_push_head(&conn->secondary_list, spkt); | |
413 | goto pri; | |
414 | } | |
6efeb328 | 415 | } |
f4b61836 | 416 | |
f449c9e5 MZ |
417 | if (colo_mark_tcp_pkt(ppkt, spkt, &mark, min_ack)) { |
418 | trace_colo_compare_tcp_info("pri", | |
419 | ppkt->tcp_seq, ppkt->tcp_ack, | |
420 | ppkt->header_size, ppkt->payload_size, | |
421 | ppkt->offset, ppkt->flags); | |
422 | ||
423 | trace_colo_compare_tcp_info("sec", | |
424 | spkt->tcp_seq, spkt->tcp_ack, | |
425 | spkt->header_size, spkt->payload_size, | |
426 | spkt->offset, spkt->flags); | |
427 | ||
428 | if (mark == COLO_COMPARE_FREE_PRIMARY) { | |
429 | conn->compare_seq = ppkt->seq_end; | |
430 | colo_release_primary_pkt(s, ppkt); | |
431 | g_queue_push_head(&conn->secondary_list, spkt); | |
432 | goto pri; | |
433 | } | |
434 | if (mark == COLO_COMPARE_FREE_SECONDARY) { | |
435 | conn->compare_seq = spkt->seq_end; | |
436 | packet_destroy(spkt, NULL); | |
437 | goto sec; | |
438 | } | |
439 | if (mark == (COLO_COMPARE_FREE_PRIMARY | COLO_COMPARE_FREE_SECONDARY)) { | |
440 | conn->compare_seq = ppkt->seq_end; | |
441 | colo_release_primary_pkt(s, ppkt); | |
442 | packet_destroy(spkt, NULL); | |
443 | goto pri; | |
444 | } | |
445 | } else { | |
446 | g_queue_push_head(&conn->primary_list, ppkt); | |
447 | g_queue_push_head(&conn->secondary_list, spkt); | |
2061c14c ZC |
448 | |
449 | qemu_hexdump((char *)ppkt->data, stderr, | |
450 | "colo-compare ppkt", ppkt->size); | |
451 | qemu_hexdump((char *)spkt->data, stderr, | |
452 | "colo-compare spkt", spkt->size); | |
f4b61836 | 453 | |
1d09f700 | 454 | colo_compare_inconsistency_notify(s); |
f449c9e5 | 455 | } |
f4b61836 ZC |
456 | } |
457 | ||
f449c9e5 | 458 | |
f4b61836 ZC |
459 | /* |
460 | * Called from the compare thread on the primary | |
461 | * for compare udp packet | |
462 | */ | |
463 | static int colo_packet_compare_udp(Packet *spkt, Packet *ppkt) | |
464 | { | |
9394133f MZ |
465 | uint16_t network_header_length = ppkt->ip->ip_hl << 2; |
466 | uint16_t offset = network_header_length + ETH_HLEN + ppkt->vnet_hdr_len; | |
f4b61836 ZC |
467 | |
468 | trace_colo_compare_main("compare udp"); | |
2ad7ca4c | 469 | |
6efeb328 ZC |
470 | /* |
471 | * Because of ppkt and spkt are both in the same connection, | |
472 | * The ppkt's src ip, dst ip, src port, dst port, ip_proto all are | |
473 | * same with spkt. In addition, IP header's Identification is a random | |
474 | * field, we can handle it in IP fragmentation function later. | |
475 | * COLO just concern the response net packet payload from primary guest | |
476 | * and secondary guest are same or not, So we ignored all IP header include | |
477 | * other field like TOS,TTL,IP Checksum. we only need to compare | |
478 | * the ip payload here. | |
479 | */ | |
9394133f MZ |
480 | if (ppkt->size != spkt->size) { |
481 | trace_colo_compare_main("UDP: payload size of packets are different"); | |
482 | return -1; | |
483 | } | |
484 | if (colo_compare_packet_payload(ppkt, spkt, offset, offset, | |
485 | ppkt->size - offset)) { | |
f4b61836 | 486 | trace_colo_compare_udp_miscompare("primary pkt size", ppkt->size); |
f4b61836 | 487 | trace_colo_compare_udp_miscompare("Secondary pkt size", spkt->size); |
d87aa138 | 488 | if (trace_event_get_state_backends(TRACE_COLO_COMPARE_MISCOMPARE)) { |
1723a7f7 ZC |
489 | qemu_hexdump((char *)ppkt->data, stderr, "colo-compare pri pkt", |
490 | ppkt->size); | |
491 | qemu_hexdump((char *)spkt->data, stderr, "colo-compare sec pkt", | |
492 | spkt->size); | |
493 | } | |
9394133f MZ |
494 | return -1; |
495 | } else { | |
496 | return 0; | |
f4b61836 | 497 | } |
f4b61836 ZC |
498 | } |
499 | ||
500 | /* | |
501 | * Called from the compare thread on the primary | |
502 | * for compare icmp packet | |
503 | */ | |
504 | static int colo_packet_compare_icmp(Packet *spkt, Packet *ppkt) | |
505 | { | |
9394133f MZ |
506 | uint16_t network_header_length = ppkt->ip->ip_hl << 2; |
507 | uint16_t offset = network_header_length + ETH_HLEN + ppkt->vnet_hdr_len; | |
6efeb328 | 508 | |
f4b61836 | 509 | trace_colo_compare_main("compare icmp"); |
f4b61836 | 510 | |
6efeb328 ZC |
511 | /* |
512 | * Because of ppkt and spkt are both in the same connection, | |
513 | * The ppkt's src ip, dst ip, src port, dst port, ip_proto all are | |
514 | * same with spkt. In addition, IP header's Identification is a random | |
515 | * field, we can handle it in IP fragmentation function later. | |
516 | * COLO just concern the response net packet payload from primary guest | |
517 | * and secondary guest are same or not, So we ignored all IP header include | |
518 | * other field like TOS,TTL,IP Checksum. we only need to compare | |
519 | * the ip payload here. | |
520 | */ | |
9394133f MZ |
521 | if (ppkt->size != spkt->size) { |
522 | trace_colo_compare_main("ICMP: payload size of packets are different"); | |
523 | return -1; | |
524 | } | |
525 | if (colo_compare_packet_payload(ppkt, spkt, offset, offset, | |
526 | ppkt->size - offset)) { | |
f4b61836 ZC |
527 | trace_colo_compare_icmp_miscompare("primary pkt size", |
528 | ppkt->size); | |
f4b61836 ZC |
529 | trace_colo_compare_icmp_miscompare("Secondary pkt size", |
530 | spkt->size); | |
d87aa138 | 531 | if (trace_event_get_state_backends(TRACE_COLO_COMPARE_MISCOMPARE)) { |
1723a7f7 ZC |
532 | qemu_hexdump((char *)ppkt->data, stderr, "colo-compare pri pkt", |
533 | ppkt->size); | |
534 | qemu_hexdump((char *)spkt->data, stderr, "colo-compare sec pkt", | |
535 | spkt->size); | |
536 | } | |
f4b61836 ZC |
537 | return -1; |
538 | } else { | |
539 | return 0; | |
540 | } | |
541 | } | |
542 | ||
543 | /* | |
544 | * Called from the compare thread on the primary | |
545 | * for compare other packet | |
546 | */ | |
547 | static int colo_packet_compare_other(Packet *spkt, Packet *ppkt) | |
548 | { | |
9394133f MZ |
549 | uint16_t offset = ppkt->vnet_hdr_len; |
550 | ||
f4b61836 | 551 | trace_colo_compare_main("compare other"); |
d87aa138 | 552 | if (trace_event_get_state_backends(TRACE_COLO_COMPARE_MISCOMPARE)) { |
e630b2bf ZC |
553 | char pri_ip_src[20], pri_ip_dst[20], sec_ip_src[20], sec_ip_dst[20]; |
554 | ||
555 | strcpy(pri_ip_src, inet_ntoa(ppkt->ip->ip_src)); | |
556 | strcpy(pri_ip_dst, inet_ntoa(ppkt->ip->ip_dst)); | |
557 | strcpy(sec_ip_src, inet_ntoa(spkt->ip->ip_src)); | |
558 | strcpy(sec_ip_dst, inet_ntoa(spkt->ip->ip_dst)); | |
559 | ||
560 | trace_colo_compare_ip_info(ppkt->size, pri_ip_src, | |
561 | pri_ip_dst, spkt->size, | |
562 | sec_ip_src, sec_ip_dst); | |
563 | } | |
564 | ||
9394133f MZ |
565 | if (ppkt->size != spkt->size) { |
566 | trace_colo_compare_main("Other: payload size of packets are different"); | |
567 | return -1; | |
568 | } | |
569 | return colo_compare_packet_payload(ppkt, spkt, offset, offset, | |
570 | ppkt->size - offset); | |
0682e15b ZC |
571 | } |
572 | ||
573 | static int colo_old_packet_check_one(Packet *pkt, int64_t *check_time) | |
574 | { | |
575 | int64_t now = qemu_clock_get_ms(QEMU_CLOCK_HOST); | |
576 | ||
577 | if ((now - pkt->creation_ms) > (*check_time)) { | |
578 | trace_colo_old_packet_check_found(pkt->creation_ms); | |
579 | return 0; | |
580 | } else { | |
581 | return 1; | |
582 | } | |
583 | } | |
584 | ||
dccd0313 ZC |
585 | void colo_compare_register_notifier(Notifier *notify) |
586 | { | |
587 | notifier_list_add(&colo_compare_notifiers, notify); | |
588 | } | |
589 | ||
590 | void colo_compare_unregister_notifier(Notifier *notify) | |
591 | { | |
592 | notifier_remove(notify); | |
593 | } | |
594 | ||
d25a7dab | 595 | static int colo_old_packet_check_one_conn(Connection *conn, |
1d09f700 | 596 | CompareState *s) |
0682e15b | 597 | { |
0682e15b ZC |
598 | GList *result = NULL; |
599 | int64_t check_time = REGULAR_PACKET_CHECK_MS; | |
600 | ||
601 | result = g_queue_find_custom(&conn->primary_list, | |
602 | &check_time, | |
603 | (GCompareFunc)colo_old_packet_check_one); | |
604 | ||
605 | if (result) { | |
61c5f469 | 606 | /* Do checkpoint will flush old packet */ |
1d09f700 | 607 | colo_compare_inconsistency_notify(s); |
d25a7dab | 608 | return 0; |
0682e15b | 609 | } |
d25a7dab ZC |
610 | |
611 | return 1; | |
0682e15b ZC |
612 | } |
613 | ||
614 | /* | |
615 | * Look for old packets that the secondary hasn't matched, | |
616 | * if we have some then we have to checkpoint to wake | |
617 | * the secondary up. | |
618 | */ | |
619 | static void colo_old_packet_check(void *opaque) | |
620 | { | |
621 | CompareState *s = opaque; | |
622 | ||
d25a7dab ZC |
623 | /* |
624 | * If we find one old packet, stop finding job and notify | |
625 | * COLO frame do checkpoint. | |
626 | */ | |
1d09f700 | 627 | g_queue_find_custom(&s->conn_list, s, |
d25a7dab | 628 | (GCompareFunc)colo_old_packet_check_one_conn); |
0682e15b ZC |
629 | } |
630 | ||
f449c9e5 MZ |
631 | static void colo_compare_packet(CompareState *s, Connection *conn, |
632 | int (*HandlePacket)(Packet *spkt, | |
633 | Packet *ppkt)) | |
0682e15b | 634 | { |
0682e15b ZC |
635 | Packet *pkt = NULL; |
636 | GList *result = NULL; | |
0682e15b ZC |
637 | |
638 | while (!g_queue_is_empty(&conn->primary_list) && | |
639 | !g_queue_is_empty(&conn->secondary_list)) { | |
626bba98 | 640 | pkt = g_queue_pop_head(&conn->primary_list); |
f449c9e5 MZ |
641 | result = g_queue_find_custom(&conn->secondary_list, |
642 | pkt, (GCompareFunc)HandlePacket); | |
0682e15b ZC |
643 | |
644 | if (result) { | |
f449c9e5 | 645 | colo_release_primary_pkt(s, pkt); |
0682e15b | 646 | g_queue_remove(&conn->secondary_list, result->data); |
0682e15b ZC |
647 | } else { |
648 | /* | |
649 | * If one packet arrive late, the secondary_list or | |
650 | * primary_list will be empty, so we can't compare it | |
dccd0313 ZC |
651 | * until next comparison. If the packets in the list are |
652 | * timeout, it will trigger a checkpoint request. | |
0682e15b ZC |
653 | */ |
654 | trace_colo_compare_main("packet different"); | |
626bba98 | 655 | g_queue_push_head(&conn->primary_list, pkt); |
1d09f700 ZC |
656 | |
657 | colo_compare_inconsistency_notify(s); | |
0682e15b ZC |
658 | break; |
659 | } | |
660 | } | |
661 | } | |
662 | ||
f449c9e5 MZ |
663 | /* |
664 | * Called from the compare thread on the primary | |
665 | * for compare packet with secondary list of the | |
666 | * specified connection when a new packet was | |
667 | * queued to it. | |
668 | */ | |
669 | static void colo_compare_connection(void *opaque, void *user_data) | |
670 | { | |
671 | CompareState *s = user_data; | |
672 | Connection *conn = opaque; | |
673 | ||
674 | switch (conn->ip_proto) { | |
675 | case IPPROTO_TCP: | |
676 | colo_compare_tcp(s, conn); | |
677 | break; | |
678 | case IPPROTO_UDP: | |
679 | colo_compare_packet(s, conn, colo_packet_compare_udp); | |
680 | break; | |
681 | case IPPROTO_ICMP: | |
682 | colo_compare_packet(s, conn, colo_packet_compare_icmp); | |
683 | break; | |
684 | default: | |
685 | colo_compare_packet(s, conn, colo_packet_compare_other); | |
686 | break; | |
687 | } | |
688 | } | |
689 | ||
3037e7a5 | 690 | static int compare_chr_send(CompareState *s, |
59509ec1 | 691 | const uint8_t *buf, |
aa3a7032 | 692 | uint32_t size, |
30685c00 ZC |
693 | uint32_t vnet_hdr_len, |
694 | bool notify_remote_frame) | |
59509ec1 ZC |
695 | { |
696 | int ret = 0; | |
697 | uint32_t len = htonl(size); | |
698 | ||
699 | if (!size) { | |
700 | return 0; | |
701 | } | |
702 | ||
30685c00 ZC |
703 | if (notify_remote_frame) { |
704 | ret = qemu_chr_fe_write_all(&s->chr_notify_dev, | |
705 | (uint8_t *)&len, | |
706 | sizeof(len)); | |
707 | } else { | |
708 | ret = qemu_chr_fe_write_all(&s->chr_out, (uint8_t *)&len, sizeof(len)); | |
709 | } | |
710 | ||
59509ec1 ZC |
711 | if (ret != sizeof(len)) { |
712 | goto err; | |
713 | } | |
714 | ||
aa3a7032 ZC |
715 | if (s->vnet_hdr) { |
716 | /* | |
717 | * We send vnet header len make other module(like filter-redirector) | |
718 | * know how to parse net packet correctly. | |
719 | */ | |
720 | len = htonl(vnet_hdr_len); | |
30685c00 ZC |
721 | |
722 | if (!notify_remote_frame) { | |
723 | ret = qemu_chr_fe_write_all(&s->chr_out, | |
724 | (uint8_t *)&len, | |
725 | sizeof(len)); | |
726 | } | |
727 | ||
aa3a7032 ZC |
728 | if (ret != sizeof(len)) { |
729 | goto err; | |
730 | } | |
731 | } | |
732 | ||
30685c00 ZC |
733 | if (notify_remote_frame) { |
734 | ret = qemu_chr_fe_write_all(&s->chr_notify_dev, | |
735 | (uint8_t *)buf, | |
736 | size); | |
737 | } else { | |
738 | ret = qemu_chr_fe_write_all(&s->chr_out, (uint8_t *)buf, size); | |
739 | } | |
740 | ||
59509ec1 ZC |
741 | if (ret != size) { |
742 | goto err; | |
743 | } | |
744 | ||
745 | return 0; | |
746 | ||
747 | err: | |
748 | return ret < 0 ? ret : -EIO; | |
749 | } | |
750 | ||
0682e15b ZC |
751 | static int compare_chr_can_read(void *opaque) |
752 | { | |
753 | return COMPARE_READ_LEN_MAX; | |
754 | } | |
755 | ||
756 | /* | |
757 | * Called from the main thread on the primary for packets | |
758 | * arriving over the socket from the primary. | |
759 | */ | |
760 | static void compare_pri_chr_in(void *opaque, const uint8_t *buf, int size) | |
761 | { | |
762 | CompareState *s = COLO_COMPARE(opaque); | |
763 | int ret; | |
764 | ||
765 | ret = net_fill_rstate(&s->pri_rs, buf, size); | |
766 | if (ret == -1) { | |
81517ba3 | 767 | qemu_chr_fe_set_handlers(&s->chr_pri_in, NULL, NULL, NULL, NULL, |
39ab61c6 | 768 | NULL, NULL, true); |
0682e15b ZC |
769 | error_report("colo-compare primary_in error"); |
770 | } | |
771 | } | |
772 | ||
773 | /* | |
774 | * Called from the main thread on the primary for packets | |
775 | * arriving over the socket from the secondary. | |
776 | */ | |
777 | static void compare_sec_chr_in(void *opaque, const uint8_t *buf, int size) | |
778 | { | |
779 | CompareState *s = COLO_COMPARE(opaque); | |
780 | int ret; | |
781 | ||
782 | ret = net_fill_rstate(&s->sec_rs, buf, size); | |
783 | if (ret == -1) { | |
81517ba3 | 784 | qemu_chr_fe_set_handlers(&s->chr_sec_in, NULL, NULL, NULL, NULL, |
39ab61c6 | 785 | NULL, NULL, true); |
0682e15b ZC |
786 | error_report("colo-compare secondary_in error"); |
787 | } | |
788 | } | |
789 | ||
13025fee ZC |
790 | static void compare_notify_chr(void *opaque, const uint8_t *buf, int size) |
791 | { | |
792 | CompareState *s = COLO_COMPARE(opaque); | |
793 | int ret; | |
794 | ||
795 | ret = net_fill_rstate(&s->notify_rs, buf, size); | |
796 | if (ret == -1) { | |
797 | qemu_chr_fe_set_handlers(&s->chr_notify_dev, NULL, NULL, NULL, NULL, | |
798 | NULL, NULL, true); | |
799 | error_report("colo-compare notify_dev error"); | |
800 | } | |
801 | } | |
802 | ||
66d2a242 HZ |
803 | /* |
804 | * Check old packet regularly so it can watch for any packets | |
805 | * that the secondary hasn't produced equivalents of. | |
806 | */ | |
dd321ecf | 807 | static void check_old_packet_regular(void *opaque) |
66d2a242 HZ |
808 | { |
809 | CompareState *s = opaque; | |
810 | ||
811 | /* if have old packet we will notify checkpoint */ | |
812 | colo_old_packet_check(s); | |
dd321ecf WY |
813 | timer_mod(s->packet_check_timer, qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + |
814 | REGULAR_PACKET_CHECK_MS); | |
815 | } | |
816 | ||
0ffcece3 ZC |
817 | /* Public API, Used for COLO frame to notify compare event */ |
818 | void colo_notify_compares_event(void *opaque, int event, Error **errp) | |
819 | { | |
820 | CompareState *s; | |
821 | ||
822 | qemu_mutex_lock(&event_mtx); | |
823 | QTAILQ_FOREACH(s, &net_compares, next) { | |
824 | s->event = event; | |
825 | qemu_bh_schedule(s->event_bh); | |
826 | event_unhandled_count++; | |
827 | } | |
828 | /* Wait all compare threads to finish handling this event */ | |
829 | while (event_unhandled_count > 0) { | |
830 | qemu_cond_wait(&event_complete_cond, &event_mtx); | |
831 | } | |
832 | ||
833 | qemu_mutex_unlock(&event_mtx); | |
834 | } | |
835 | ||
dd321ecf WY |
836 | static void colo_compare_timer_init(CompareState *s) |
837 | { | |
838 | AioContext *ctx = iothread_get_aio_context(s->iothread); | |
66d2a242 | 839 | |
dd321ecf WY |
840 | s->packet_check_timer = aio_timer_new(ctx, QEMU_CLOCK_VIRTUAL, |
841 | SCALE_MS, check_old_packet_regular, | |
842 | s); | |
843 | timer_mod(s->packet_check_timer, qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + | |
844 | REGULAR_PACKET_CHECK_MS); | |
66d2a242 HZ |
845 | } |
846 | ||
dd321ecf | 847 | static void colo_compare_timer_del(CompareState *s) |
0682e15b | 848 | { |
dd321ecf WY |
849 | if (s->packet_check_timer) { |
850 | timer_del(s->packet_check_timer); | |
851 | timer_free(s->packet_check_timer); | |
852 | s->packet_check_timer = NULL; | |
853 | } | |
854 | } | |
0682e15b | 855 | |
0ffcece3 ZC |
856 | static void colo_flush_packets(void *opaque, void *user_data); |
857 | ||
858 | static void colo_compare_handle_event(void *opaque) | |
859 | { | |
860 | CompareState *s = opaque; | |
861 | ||
862 | switch (s->event) { | |
863 | case COLO_EVENT_CHECKPOINT: | |
864 | g_queue_foreach(&s->conn_list, colo_flush_packets, s); | |
865 | break; | |
866 | case COLO_EVENT_FAILOVER: | |
867 | break; | |
868 | default: | |
869 | break; | |
870 | } | |
871 | ||
0ffcece3 | 872 | qemu_mutex_lock(&event_mtx); |
78e4f446 | 873 | assert(event_unhandled_count > 0); |
0ffcece3 ZC |
874 | event_unhandled_count--; |
875 | qemu_cond_broadcast(&event_complete_cond); | |
876 | qemu_mutex_unlock(&event_mtx); | |
877 | } | |
878 | ||
dd321ecf WY |
879 | static void colo_compare_iothread(CompareState *s) |
880 | { | |
881 | object_ref(OBJECT(s->iothread)); | |
882 | s->worker_context = iothread_get_g_main_context(s->iothread); | |
0682e15b | 883 | |
5345fdb4 | 884 | qemu_chr_fe_set_handlers(&s->chr_pri_in, compare_chr_can_read, |
81517ba3 AN |
885 | compare_pri_chr_in, NULL, NULL, |
886 | s, s->worker_context, true); | |
5345fdb4 | 887 | qemu_chr_fe_set_handlers(&s->chr_sec_in, compare_chr_can_read, |
81517ba3 AN |
888 | compare_sec_chr_in, NULL, NULL, |
889 | s, s->worker_context, true); | |
13025fee ZC |
890 | if (s->notify_dev) { |
891 | qemu_chr_fe_set_handlers(&s->chr_notify_dev, compare_chr_can_read, | |
892 | compare_notify_chr, NULL, NULL, | |
893 | s, s->worker_context, true); | |
894 | } | |
0682e15b | 895 | |
dd321ecf | 896 | colo_compare_timer_init(s); |
0ffcece3 | 897 | s->event_bh = qemu_bh_new(colo_compare_handle_event, s); |
0682e15b ZC |
898 | } |
899 | ||
7dce4e6f ZC |
900 | static char *compare_get_pri_indev(Object *obj, Error **errp) |
901 | { | |
902 | CompareState *s = COLO_COMPARE(obj); | |
903 | ||
904 | return g_strdup(s->pri_indev); | |
905 | } | |
906 | ||
907 | static void compare_set_pri_indev(Object *obj, const char *value, Error **errp) | |
908 | { | |
909 | CompareState *s = COLO_COMPARE(obj); | |
910 | ||
911 | g_free(s->pri_indev); | |
912 | s->pri_indev = g_strdup(value); | |
913 | } | |
914 | ||
915 | static char *compare_get_sec_indev(Object *obj, Error **errp) | |
916 | { | |
917 | CompareState *s = COLO_COMPARE(obj); | |
918 | ||
919 | return g_strdup(s->sec_indev); | |
920 | } | |
921 | ||
922 | static void compare_set_sec_indev(Object *obj, const char *value, Error **errp) | |
923 | { | |
924 | CompareState *s = COLO_COMPARE(obj); | |
925 | ||
926 | g_free(s->sec_indev); | |
927 | s->sec_indev = g_strdup(value); | |
928 | } | |
929 | ||
930 | static char *compare_get_outdev(Object *obj, Error **errp) | |
931 | { | |
932 | CompareState *s = COLO_COMPARE(obj); | |
933 | ||
934 | return g_strdup(s->outdev); | |
935 | } | |
936 | ||
937 | static void compare_set_outdev(Object *obj, const char *value, Error **errp) | |
938 | { | |
939 | CompareState *s = COLO_COMPARE(obj); | |
940 | ||
941 | g_free(s->outdev); | |
942 | s->outdev = g_strdup(value); | |
943 | } | |
944 | ||
aa3a7032 ZC |
945 | static bool compare_get_vnet_hdr(Object *obj, Error **errp) |
946 | { | |
947 | CompareState *s = COLO_COMPARE(obj); | |
948 | ||
949 | return s->vnet_hdr; | |
950 | } | |
951 | ||
952 | static void compare_set_vnet_hdr(Object *obj, | |
953 | bool value, | |
954 | Error **errp) | |
955 | { | |
956 | CompareState *s = COLO_COMPARE(obj); | |
957 | ||
958 | s->vnet_hdr = value; | |
959 | } | |
960 | ||
cf6af766 ZC |
961 | static char *compare_get_notify_dev(Object *obj, Error **errp) |
962 | { | |
963 | CompareState *s = COLO_COMPARE(obj); | |
964 | ||
965 | return g_strdup(s->notify_dev); | |
966 | } | |
967 | ||
968 | static void compare_set_notify_dev(Object *obj, const char *value, Error **errp) | |
969 | { | |
970 | CompareState *s = COLO_COMPARE(obj); | |
971 | ||
972 | g_free(s->notify_dev); | |
973 | s->notify_dev = g_strdup(value); | |
974 | } | |
975 | ||
7dce4e6f ZC |
976 | static void compare_pri_rs_finalize(SocketReadState *pri_rs) |
977 | { | |
59509ec1 | 978 | CompareState *s = container_of(pri_rs, CompareState, pri_rs); |
8ec14402 | 979 | Connection *conn = NULL; |
59509ec1 | 980 | |
8ec14402 | 981 | if (packet_enqueue(s, PRIMARY_IN, &conn)) { |
59509ec1 | 982 | trace_colo_compare_main("primary: unsupported packet in"); |
aa3a7032 ZC |
983 | compare_chr_send(s, |
984 | pri_rs->buf, | |
985 | pri_rs->packet_len, | |
30685c00 ZC |
986 | pri_rs->vnet_hdr_len, |
987 | false); | |
0682e15b | 988 | } else { |
3463218c | 989 | /* compare packet in the specified connection */ |
8ec14402 | 990 | colo_compare_connection(conn, s); |
59509ec1 | 991 | } |
7dce4e6f ZC |
992 | } |
993 | ||
994 | static void compare_sec_rs_finalize(SocketReadState *sec_rs) | |
995 | { | |
59509ec1 | 996 | CompareState *s = container_of(sec_rs, CompareState, sec_rs); |
8ec14402 | 997 | Connection *conn = NULL; |
59509ec1 | 998 | |
8ec14402 | 999 | if (packet_enqueue(s, SECONDARY_IN, &conn)) { |
59509ec1 | 1000 | trace_colo_compare_main("secondary: unsupported packet in"); |
0682e15b | 1001 | } else { |
3463218c | 1002 | /* compare packet in the specified connection */ |
8ec14402 | 1003 | colo_compare_connection(conn, s); |
59509ec1 | 1004 | } |
7dce4e6f ZC |
1005 | } |
1006 | ||
13025fee ZC |
1007 | static void compare_notify_rs_finalize(SocketReadState *notify_rs) |
1008 | { | |
1d09f700 ZC |
1009 | CompareState *s = container_of(notify_rs, CompareState, notify_rs); |
1010 | ||
13025fee | 1011 | /* Get Xen colo-frame's notify and handle the message */ |
1d09f700 ZC |
1012 | char *data = g_memdup(notify_rs->buf, notify_rs->packet_len); |
1013 | char msg[] = "COLO_COMPARE_GET_XEN_INIT"; | |
1014 | int ret; | |
1015 | ||
1016 | if (!strcmp(data, "COLO_USERSPACE_PROXY_INIT")) { | |
1017 | ret = compare_chr_send(s, (uint8_t *)msg, strlen(msg), 0, true); | |
1018 | if (ret < 0) { | |
1019 | error_report("Notify Xen COLO-frame INIT failed"); | |
1020 | } | |
1021 | } | |
1022 | ||
1023 | if (!strcmp(data, "COLO_CHECKPOINT")) { | |
1024 | /* colo-compare do checkpoint, flush pri packet and remove sec packet */ | |
1025 | g_queue_foreach(&s->conn_list, colo_flush_packets, s); | |
1026 | } | |
13025fee | 1027 | } |
7dce4e6f ZC |
1028 | |
1029 | /* | |
1030 | * Return 0 is success. | |
1031 | * Return 1 is failed. | |
1032 | */ | |
0ec7b3e7 | 1033 | static int find_and_check_chardev(Chardev **chr, |
7dce4e6f ZC |
1034 | char *chr_name, |
1035 | Error **errp) | |
1036 | { | |
7dce4e6f ZC |
1037 | *chr = qemu_chr_find(chr_name); |
1038 | if (*chr == NULL) { | |
1039 | error_setg(errp, "Device '%s' not found", | |
1040 | chr_name); | |
1041 | return 1; | |
1042 | } | |
1043 | ||
0a73336d DB |
1044 | if (!qemu_chr_has_feature(*chr, QEMU_CHAR_FEATURE_RECONNECTABLE)) { |
1045 | error_setg(errp, "chardev \"%s\" is not reconnectable", | |
7dce4e6f ZC |
1046 | chr_name); |
1047 | return 1; | |
1048 | } | |
fbf3cc3a | 1049 | |
269d25cd MAL |
1050 | if (!qemu_chr_has_feature(*chr, QEMU_CHAR_FEATURE_GCONTEXT)) { |
1051 | error_setg(errp, "chardev \"%s\" cannot switch context", | |
1052 | chr_name); | |
1053 | return 1; | |
1054 | } | |
1055 | ||
7dce4e6f ZC |
1056 | return 0; |
1057 | } | |
1058 | ||
1059 | /* | |
1060 | * Called from the main thread on the primary | |
1061 | * to setup colo-compare. | |
1062 | */ | |
1063 | static void colo_compare_complete(UserCreatable *uc, Error **errp) | |
1064 | { | |
1065 | CompareState *s = COLO_COMPARE(uc); | |
0ec7b3e7 | 1066 | Chardev *chr; |
7dce4e6f | 1067 | |
dd321ecf | 1068 | if (!s->pri_indev || !s->sec_indev || !s->outdev || !s->iothread) { |
7dce4e6f | 1069 | error_setg(errp, "colo compare needs 'primary_in' ," |
dd321ecf | 1070 | "'secondary_in','outdev','iothread' property set"); |
7dce4e6f ZC |
1071 | return; |
1072 | } else if (!strcmp(s->pri_indev, s->outdev) || | |
1073 | !strcmp(s->sec_indev, s->outdev) || | |
1074 | !strcmp(s->pri_indev, s->sec_indev)) { | |
1075 | error_setg(errp, "'indev' and 'outdev' could not be same " | |
1076 | "for compare module"); | |
1077 | return; | |
1078 | } | |
1079 | ||
5345fdb4 MAL |
1080 | if (find_and_check_chardev(&chr, s->pri_indev, errp) || |
1081 | !qemu_chr_fe_init(&s->chr_pri_in, chr, errp)) { | |
7dce4e6f ZC |
1082 | return; |
1083 | } | |
1084 | ||
5345fdb4 MAL |
1085 | if (find_and_check_chardev(&chr, s->sec_indev, errp) || |
1086 | !qemu_chr_fe_init(&s->chr_sec_in, chr, errp)) { | |
7dce4e6f ZC |
1087 | return; |
1088 | } | |
1089 | ||
5345fdb4 MAL |
1090 | if (find_and_check_chardev(&chr, s->outdev, errp) || |
1091 | !qemu_chr_fe_init(&s->chr_out, chr, errp)) { | |
7dce4e6f ZC |
1092 | return; |
1093 | } | |
1094 | ||
aa3a7032 ZC |
1095 | net_socket_rs_init(&s->pri_rs, compare_pri_rs_finalize, s->vnet_hdr); |
1096 | net_socket_rs_init(&s->sec_rs, compare_sec_rs_finalize, s->vnet_hdr); | |
7dce4e6f | 1097 | |
13025fee ZC |
1098 | /* Try to enable remote notify chardev, currently just for Xen COLO */ |
1099 | if (s->notify_dev) { | |
1100 | if (find_and_check_chardev(&chr, s->notify_dev, errp) || | |
1101 | !qemu_chr_fe_init(&s->chr_notify_dev, chr, errp)) { | |
1102 | return; | |
1103 | } | |
1104 | ||
1105 | net_socket_rs_init(&s->notify_rs, compare_notify_rs_finalize, | |
1106 | s->vnet_hdr); | |
1107 | } | |
1108 | ||
0ffcece3 ZC |
1109 | QTAILQ_INSERT_TAIL(&net_compares, s, next); |
1110 | ||
b6540d40 ZC |
1111 | g_queue_init(&s->conn_list); |
1112 | ||
0ffcece3 ZC |
1113 | qemu_mutex_init(&event_mtx); |
1114 | qemu_cond_init(&event_complete_cond); | |
1115 | ||
b6540d40 ZC |
1116 | s->connection_track_table = g_hash_table_new_full(connection_key_hash, |
1117 | connection_key_equal, | |
1118 | g_free, | |
1119 | connection_destroy); | |
59509ec1 | 1120 | |
dd321ecf | 1121 | colo_compare_iothread(s); |
7dce4e6f ZC |
1122 | return; |
1123 | } | |
1124 | ||
dfd917a9 HZ |
1125 | static void colo_flush_packets(void *opaque, void *user_data) |
1126 | { | |
1127 | CompareState *s = user_data; | |
1128 | Connection *conn = opaque; | |
1129 | Packet *pkt = NULL; | |
1130 | ||
1131 | while (!g_queue_is_empty(&conn->primary_list)) { | |
1132 | pkt = g_queue_pop_head(&conn->primary_list); | |
aa3a7032 ZC |
1133 | compare_chr_send(s, |
1134 | pkt->data, | |
1135 | pkt->size, | |
30685c00 ZC |
1136 | pkt->vnet_hdr_len, |
1137 | false); | |
dfd917a9 HZ |
1138 | packet_destroy(pkt, NULL); |
1139 | } | |
1140 | while (!g_queue_is_empty(&conn->secondary_list)) { | |
1141 | pkt = g_queue_pop_head(&conn->secondary_list); | |
1142 | packet_destroy(pkt, NULL); | |
1143 | } | |
1144 | } | |
1145 | ||
7dce4e6f ZC |
1146 | static void colo_compare_class_init(ObjectClass *oc, void *data) |
1147 | { | |
1148 | UserCreatableClass *ucc = USER_CREATABLE_CLASS(oc); | |
1149 | ||
1150 | ucc->complete = colo_compare_complete; | |
1151 | } | |
1152 | ||
1153 | static void colo_compare_init(Object *obj) | |
1154 | { | |
aa3a7032 ZC |
1155 | CompareState *s = COLO_COMPARE(obj); |
1156 | ||
7dce4e6f ZC |
1157 | object_property_add_str(obj, "primary_in", |
1158 | compare_get_pri_indev, compare_set_pri_indev, | |
1159 | NULL); | |
1160 | object_property_add_str(obj, "secondary_in", | |
1161 | compare_get_sec_indev, compare_set_sec_indev, | |
1162 | NULL); | |
1163 | object_property_add_str(obj, "outdev", | |
1164 | compare_get_outdev, compare_set_outdev, | |
1165 | NULL); | |
dd321ecf WY |
1166 | object_property_add_link(obj, "iothread", TYPE_IOTHREAD, |
1167 | (Object **)&s->iothread, | |
1168 | object_property_allow_set_link, | |
265b578c | 1169 | OBJ_PROP_LINK_STRONG, NULL); |
cf6af766 ZC |
1170 | /* This parameter just for Xen COLO */ |
1171 | object_property_add_str(obj, "notify_dev", | |
1172 | compare_get_notify_dev, compare_set_notify_dev, | |
1173 | NULL); | |
aa3a7032 ZC |
1174 | |
1175 | s->vnet_hdr = false; | |
1176 | object_property_add_bool(obj, "vnet_hdr_support", compare_get_vnet_hdr, | |
1177 | compare_set_vnet_hdr, NULL); | |
7dce4e6f ZC |
1178 | } |
1179 | ||
1180 | static void colo_compare_finalize(Object *obj) | |
1181 | { | |
1182 | CompareState *s = COLO_COMPARE(obj); | |
0ffcece3 | 1183 | CompareState *tmp = NULL; |
7dce4e6f | 1184 | |
1ce2610c MAL |
1185 | qemu_chr_fe_deinit(&s->chr_pri_in, false); |
1186 | qemu_chr_fe_deinit(&s->chr_sec_in, false); | |
1187 | qemu_chr_fe_deinit(&s->chr_out, false); | |
13025fee ZC |
1188 | if (s->notify_dev) { |
1189 | qemu_chr_fe_deinit(&s->chr_notify_dev, false); | |
1190 | } | |
1191 | ||
dd321ecf WY |
1192 | if (s->iothread) { |
1193 | colo_compare_timer_del(s); | |
1194 | } | |
0ffcece3 ZC |
1195 | |
1196 | qemu_bh_delete(s->event_bh); | |
1197 | ||
1198 | QTAILQ_FOREACH(tmp, &net_compares, next) { | |
1199 | if (tmp == s) { | |
1200 | QTAILQ_REMOVE(&net_compares, s, next); | |
1201 | break; | |
1202 | } | |
1203 | } | |
1204 | ||
dfd917a9 HZ |
1205 | /* Release all unhandled packets after compare thead exited */ |
1206 | g_queue_foreach(&s->conn_list, colo_flush_packets, s); | |
1207 | ||
727c2d76 | 1208 | g_queue_clear(&s->conn_list); |
0682e15b | 1209 | |
dd321ecf WY |
1210 | if (s->connection_track_table) { |
1211 | g_hash_table_destroy(s->connection_track_table); | |
1212 | } | |
1213 | ||
1214 | if (s->iothread) { | |
1215 | object_unref(OBJECT(s->iothread)); | |
1216 | } | |
0ffcece3 ZC |
1217 | |
1218 | qemu_mutex_destroy(&event_mtx); | |
1219 | qemu_cond_destroy(&event_complete_cond); | |
1220 | ||
7dce4e6f ZC |
1221 | g_free(s->pri_indev); |
1222 | g_free(s->sec_indev); | |
1223 | g_free(s->outdev); | |
cf6af766 | 1224 | g_free(s->notify_dev); |
7dce4e6f ZC |
1225 | } |
1226 | ||
1227 | static const TypeInfo colo_compare_info = { | |
1228 | .name = TYPE_COLO_COMPARE, | |
1229 | .parent = TYPE_OBJECT, | |
1230 | .instance_size = sizeof(CompareState), | |
1231 | .instance_init = colo_compare_init, | |
1232 | .instance_finalize = colo_compare_finalize, | |
1233 | .class_size = sizeof(CompareClass), | |
1234 | .class_init = colo_compare_class_init, | |
1235 | .interfaces = (InterfaceInfo[]) { | |
1236 | { TYPE_USER_CREATABLE }, | |
1237 | { } | |
1238 | } | |
1239 | }; | |
1240 | ||
1241 | static void register_types(void) | |
1242 | { | |
1243 | type_register_static(&colo_compare_info); | |
1244 | } | |
1245 | ||
1246 | type_init(register_types); |