]>
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 | 34 | |
9c55fe94 LS |
35 | #include "block/aio-wait.h" |
36 | #include "qemu/coroutine.h" | |
37 | ||
7dce4e6f | 38 | #define TYPE_COLO_COMPARE "colo-compare" |
db1015e9 | 39 | typedef struct CompareState CompareState; |
8110fa1d EH |
40 | DECLARE_INSTANCE_CHECKER(CompareState, COLO_COMPARE, |
41 | TYPE_COLO_COMPARE) | |
7dce4e6f | 42 | |
0ffcece3 ZC |
43 | static QTAILQ_HEAD(, CompareState) net_compares = |
44 | QTAILQ_HEAD_INITIALIZER(net_compares); | |
45 | ||
dccd0313 ZC |
46 | static NotifierList colo_compare_notifiers = |
47 | NOTIFIER_LIST_INITIALIZER(colo_compare_notifiers); | |
48 | ||
0682e15b | 49 | #define COMPARE_READ_LEN_MAX NET_BUFSIZE |
b6540d40 ZC |
50 | #define MAX_QUEUE_SIZE 1024 |
51 | ||
f449c9e5 MZ |
52 | #define COLO_COMPARE_FREE_PRIMARY 0x01 |
53 | #define COLO_COMPARE_FREE_SECONDARY 0x02 | |
54 | ||
2f2fcff3 | 55 | #define REGULAR_PACKET_CHECK_MS 1000 |
9cc43c94 | 56 | #define DEFAULT_TIME_OUT_MS 3000 |
0682e15b | 57 | |
8c8ed038 RB |
58 | /* #define DEBUG_COLO_PACKETS */ |
59 | ||
45942b79 LS |
60 | static QemuMutex colo_compare_mutex; |
61 | static bool colo_compare_active; | |
0ffcece3 ZC |
62 | static QemuMutex event_mtx; |
63 | static QemuCond event_complete_cond; | |
64 | static int event_unhandled_count; | |
a2e5cb7a | 65 | static uint32_t max_queue_size; |
0ffcece3 | 66 | |
59509ec1 | 67 | /* |
61c5f469 ZC |
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 | * +--------+ +--------+ +--------+ +--------+ | |
88 | */ | |
9c55fe94 LS |
89 | |
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; | |
99 | ||
100 | typedef struct SendEntry { | |
101 | uint32_t size; | |
102 | uint32_t vnet_hdr_len; | |
103 | uint8_t *buf; | |
104 | } SendEntry; | |
105 | ||
db1015e9 | 106 | struct CompareState { |
7dce4e6f ZC |
107 | Object parent; |
108 | ||
109 | char *pri_indev; | |
110 | char *sec_indev; | |
111 | char *outdev; | |
cf6af766 | 112 | char *notify_dev; |
32a6ebec MAL |
113 | CharBackend chr_pri_in; |
114 | CharBackend chr_sec_in; | |
115 | CharBackend chr_out; | |
13025fee | 116 | CharBackend chr_notify_dev; |
7dce4e6f ZC |
117 | SocketReadState pri_rs; |
118 | SocketReadState sec_rs; | |
13025fee | 119 | SocketReadState notify_rs; |
9c55fe94 LS |
120 | SendCo out_sendco; |
121 | SendCo notify_sendco; | |
aa3a7032 | 122 | bool vnet_hdr; |
0c4266ef | 123 | uint64_t compare_timeout; |
cca35ac4 | 124 | uint32_t expired_scan_cycle; |
59509ec1 | 125 | |
61c5f469 ZC |
126 | /* |
127 | * Record the connection that through the NIC | |
128 | * Element type: Connection | |
b6540d40 ZC |
129 | */ |
130 | GQueue conn_list; | |
61c5f469 | 131 | /* Record the connection without repetition */ |
59509ec1 | 132 | GHashTable *connection_track_table; |
dfd917a9 | 133 | |
dd321ecf | 134 | IOThread *iothread; |
b43decb0 | 135 | GMainContext *worker_context; |
dd321ecf | 136 | QEMUTimer *packet_check_timer; |
0ffcece3 ZC |
137 | |
138 | QEMUBH *event_bh; | |
139 | enum colo_event event; | |
140 | ||
141 | QTAILQ_ENTRY(CompareState) next; | |
db1015e9 | 142 | }; |
7dce4e6f ZC |
143 | |
144 | typedef struct CompareClass { | |
145 | ObjectClass parent_class; | |
146 | } CompareClass; | |
147 | ||
59509ec1 ZC |
148 | enum { |
149 | PRIMARY_IN = 0, | |
150 | SECONDARY_IN, | |
151 | }; | |
152 | ||
bdadbb0f DS |
153 | static const char *colo_mode[] = { |
154 | [PRIMARY_IN] = "primary", | |
155 | [SECONDARY_IN] = "secondary", | |
156 | }; | |
24525e93 | 157 | |
3037e7a5 | 158 | static int compare_chr_send(CompareState *s, |
9c55fe94 | 159 | uint8_t *buf, |
aa3a7032 | 160 | uint32_t size, |
30685c00 | 161 | uint32_t vnet_hdr_len, |
9c55fe94 LS |
162 | bool notify_remote_frame, |
163 | bool zero_copy); | |
59509ec1 | 164 | |
f77bed14 ZC |
165 | static bool packet_matches_str(const char *str, |
166 | const uint8_t *buf, | |
167 | uint32_t packet_len) | |
168 | { | |
169 | if (packet_len != strlen(str)) { | |
170 | return false; | |
171 | } | |
172 | ||
173 | return !memcmp(str, buf, strlen(str)); | |
174 | } | |
175 | ||
1d09f700 ZC |
176 | static void notify_remote_frame(CompareState *s) |
177 | { | |
178 | char msg[] = "DO_CHECKPOINT"; | |
179 | int ret = 0; | |
180 | ||
9c55fe94 | 181 | ret = compare_chr_send(s, (uint8_t *)msg, strlen(msg), 0, true, false); |
1d09f700 ZC |
182 | if (ret < 0) { |
183 | error_report("Notify Xen COLO-frame failed"); | |
184 | } | |
185 | } | |
186 | ||
187 | static void colo_compare_inconsistency_notify(CompareState *s) | |
188 | { | |
189 | if (s->notify_dev) { | |
190 | notify_remote_frame(s); | |
191 | } else { | |
192 | notifier_list_notify(&colo_compare_notifiers, | |
193 | migrate_get_current()); | |
194 | } | |
195 | } | |
196 | ||
33609e95 | 197 | /* Use restricted to colo_insert_packet() */ |
a935cc31 ZC |
198 | static gint seq_sorter(Packet *a, Packet *b, gpointer data) |
199 | { | |
33609e95 | 200 | return a->tcp_seq - b->tcp_seq; |
a935cc31 ZC |
201 | } |
202 | ||
f449c9e5 MZ |
203 | static void fill_pkt_tcp_info(void *data, uint32_t *max_ack) |
204 | { | |
205 | Packet *pkt = data; | |
e05ae1d9 | 206 | struct tcp_hdr *tcphd; |
f449c9e5 | 207 | |
e05ae1d9 | 208 | tcphd = (struct tcp_hdr *)pkt->transport_header; |
f449c9e5 MZ |
209 | |
210 | pkt->tcp_seq = ntohl(tcphd->th_seq); | |
211 | pkt->tcp_ack = ntohl(tcphd->th_ack); | |
212 | *max_ack = *max_ack > pkt->tcp_ack ? *max_ack : pkt->tcp_ack; | |
213 | pkt->header_size = pkt->transport_header - (uint8_t *)pkt->data | |
214 | + (tcphd->th_off << 2) - pkt->vnet_hdr_len; | |
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; | |
218 | } | |
219 | ||
8850d4ca MZ |
220 | /* |
221 | * Return 1 on success, if return 0 means the | |
222 | * packet will be dropped | |
223 | */ | |
f449c9e5 | 224 | static int colo_insert_packet(GQueue *queue, Packet *pkt, uint32_t *max_ack) |
8850d4ca | 225 | { |
a2e5cb7a | 226 | if (g_queue_get_length(queue) <= max_queue_size) { |
8850d4ca | 227 | if (pkt->ip->ip_p == IPPROTO_TCP) { |
f449c9e5 | 228 | fill_pkt_tcp_info(pkt, max_ack); |
8850d4ca MZ |
229 | g_queue_insert_sorted(queue, |
230 | pkt, | |
231 | (GCompareDataFunc)seq_sorter, | |
232 | NULL); | |
233 | } else { | |
234 | g_queue_push_tail(queue, pkt); | |
235 | } | |
236 | return 1; | |
237 | } | |
238 | return 0; | |
239 | } | |
240 | ||
59509ec1 ZC |
241 | /* |
242 | * Return 0 on success, if return -1 means the pkt | |
243 | * is unsupported(arp and ipv6) and will be sent later | |
244 | */ | |
8ec14402 | 245 | static int packet_enqueue(CompareState *s, int mode, Connection **con) |
59509ec1 | 246 | { |
b6540d40 | 247 | ConnectionKey key; |
59509ec1 | 248 | Packet *pkt = NULL; |
b6540d40 | 249 | Connection *conn; |
bdadbb0f | 250 | int ret; |
59509ec1 ZC |
251 | |
252 | if (mode == PRIMARY_IN) { | |
ada1a33f ZC |
253 | pkt = packet_new(s->pri_rs.buf, |
254 | s->pri_rs.packet_len, | |
255 | s->pri_rs.vnet_hdr_len); | |
59509ec1 | 256 | } else { |
ada1a33f ZC |
257 | pkt = packet_new(s->sec_rs.buf, |
258 | s->sec_rs.packet_len, | |
259 | s->sec_rs.vnet_hdr_len); | |
59509ec1 ZC |
260 | } |
261 | ||
262 | if (parse_packet_early(pkt)) { | |
263 | packet_destroy(pkt, NULL); | |
264 | pkt = NULL; | |
265 | return -1; | |
266 | } | |
b6540d40 | 267 | fill_connection_key(pkt, &key); |
59509ec1 | 268 | |
b6540d40 ZC |
269 | conn = connection_get(s->connection_track_table, |
270 | &key, | |
271 | &s->conn_list); | |
59509ec1 | 272 | |
b6540d40 ZC |
273 | if (!conn->processing) { |
274 | g_queue_push_tail(&s->conn_list, conn); | |
275 | conn->processing = true; | |
276 | } | |
277 | ||
278 | if (mode == PRIMARY_IN) { | |
bdadbb0f | 279 | ret = colo_insert_packet(&conn->primary_list, pkt, &conn->pack); |
b6540d40 | 280 | } else { |
bdadbb0f | 281 | ret = colo_insert_packet(&conn->secondary_list, pkt, &conn->sack); |
b6540d40 | 282 | } |
bdadbb0f DS |
283 | |
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; | |
289 | } | |
290 | ||
4d366235 | 291 | *con = conn; |
59509ec1 ZC |
292 | |
293 | return 0; | |
294 | } | |
295 | ||
f449c9e5 MZ |
296 | static inline bool after(uint32_t seq1, uint32_t seq2) |
297 | { | |
298 | return (int32_t)(seq1 - seq2) > 0; | |
299 | } | |
300 | ||
301 | static void colo_release_primary_pkt(CompareState *s, Packet *pkt) | |
302 | { | |
303 | int ret; | |
304 | ret = compare_chr_send(s, | |
305 | pkt->data, | |
306 | pkt->size, | |
30685c00 | 307 | pkt->vnet_hdr_len, |
9c55fe94 LS |
308 | false, |
309 | true); | |
f449c9e5 MZ |
310 | if (ret < 0) { |
311 | error_report("colo send primary packet failed"); | |
312 | } | |
313 | trace_colo_compare_main("packet same and release packet"); | |
9c55fe94 | 314 | packet_destroy_partial(pkt, NULL); |
f449c9e5 MZ |
315 | } |
316 | ||
0682e15b ZC |
317 | /* |
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 | |
323 | */ | |
9394133f MZ |
324 | static int colo_compare_packet_payload(Packet *ppkt, |
325 | Packet *spkt, | |
326 | uint16_t poffset, | |
327 | uint16_t soffset, | |
328 | uint16_t len) | |
329 | ||
0682e15b | 330 | { |
8c8ed038 | 331 | if (trace_event_get_state_backends(TRACE_COLO_COMPARE_IP_INFO)) { |
e630b2bf ZC |
332 | char pri_ip_src[20], pri_ip_dst[20], sec_ip_src[20], sec_ip_dst[20]; |
333 | ||
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)); | |
338 | ||
339 | trace_colo_compare_ip_info(ppkt->size, pri_ip_src, | |
340 | pri_ip_dst, spkt->size, | |
341 | sec_ip_src, sec_ip_dst); | |
342 | } | |
0682e15b | 343 | |
9394133f | 344 | return memcmp(ppkt->data + poffset, spkt->data + soffset, len); |
0682e15b ZC |
345 | } |
346 | ||
f4b61836 | 347 | /* |
f449c9e5 MZ |
348 | * return true means that the payload is consist and |
349 | * need to make the next comparison, false means do | |
350 | * the checkpoint | |
351 | */ | |
352 | static bool colo_mark_tcp_pkt(Packet *ppkt, Packet *spkt, | |
353 | int8_t *mark, uint32_t max_ack) | |
0682e15b | 354 | { |
f449c9e5 MZ |
355 | *mark = 0; |
356 | ||
f449c9e5 | 357 | if (ppkt->tcp_seq == spkt->tcp_seq && ppkt->seq_end == spkt->seq_end) { |
1e907a32 | 358 | if (!colo_compare_packet_payload(ppkt, spkt, |
f449c9e5 MZ |
359 | ppkt->header_size, spkt->header_size, |
360 | ppkt->payload_size)) { | |
361 | *mark = COLO_COMPARE_FREE_SECONDARY | COLO_COMPARE_FREE_PRIMARY; | |
362 | return true; | |
363 | } | |
364 | } | |
f4b61836 | 365 | |
f449c9e5 MZ |
366 | /* one part of secondary packet payload still need to be compared */ |
367 | if (!after(ppkt->seq_end, spkt->seq_end)) { | |
1e907a32 | 368 | if (!colo_compare_packet_payload(ppkt, spkt, |
f449c9e5 MZ |
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 | |
379 | */ | |
380 | return false; | |
381 | } | |
382 | } | |
383 | } else { | |
384 | /* primary packet is longer than secondary packet, compare | |
385 | * the same part and mark the primary packet offset | |
386 | */ | |
1e907a32 | 387 | if (!colo_compare_packet_payload(ppkt, spkt, |
f449c9e5 MZ |
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; | |
394 | } | |
395 | } | |
396 | ||
397 | return false; | |
398 | } | |
2ad7ca4c | 399 | |
f449c9e5 MZ |
400 | static void colo_compare_tcp(CompareState *s, Connection *conn) |
401 | { | |
402 | Packet *ppkt = NULL, *spkt = NULL; | |
403 | int8_t mark; | |
f4b61836 ZC |
404 | |
405 | /* | |
f449c9e5 MZ |
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. | |
415 | */ | |
416 | uint32_t min_ack = conn->pack > conn->sack ? conn->sack : conn->pack; | |
417 | ||
418 | pri: | |
419 | if (g_queue_is_empty(&conn->primary_list)) { | |
420 | return; | |
421 | } | |
422 | ppkt = g_queue_pop_head(&conn->primary_list); | |
423 | sec: | |
424 | if (g_queue_is_empty(&conn->secondary_list)) { | |
425 | g_queue_push_head(&conn->primary_list, ppkt); | |
426 | return; | |
f4b61836 | 427 | } |
f449c9e5 | 428 | spkt = g_queue_pop_head(&conn->secondary_list); |
f4b61836 | 429 | |
f449c9e5 MZ |
430 | if (ppkt->tcp_seq == ppkt->seq_end) { |
431 | colo_release_primary_pkt(s, ppkt); | |
432 | ppkt = NULL; | |
433 | } | |
9394133f | 434 | |
f449c9e5 MZ |
435 | if (ppkt && conn->compare_seq && !after(ppkt->seq_end, conn->compare_seq)) { |
436 | trace_colo_compare_main("pri: this packet has compared"); | |
437 | colo_release_primary_pkt(s, ppkt); | |
438 | ppkt = NULL; | |
439 | } | |
9394133f | 440 | |
f449c9e5 MZ |
441 | if (spkt->tcp_seq == spkt->seq_end) { |
442 | packet_destroy(spkt, NULL); | |
443 | if (!ppkt) { | |
444 | goto pri; | |
445 | } else { | |
446 | goto sec; | |
447 | } | |
6efeb328 | 448 | } else { |
f449c9e5 MZ |
449 | if (conn->compare_seq && !after(spkt->seq_end, conn->compare_seq)) { |
450 | trace_colo_compare_main("sec: this packet has compared"); | |
451 | packet_destroy(spkt, NULL); | |
452 | if (!ppkt) { | |
453 | goto pri; | |
454 | } else { | |
455 | goto sec; | |
456 | } | |
457 | } | |
458 | if (!ppkt) { | |
459 | g_queue_push_head(&conn->secondary_list, spkt); | |
460 | goto pri; | |
461 | } | |
6efeb328 | 462 | } |
f4b61836 | 463 | |
f449c9e5 MZ |
464 | if (colo_mark_tcp_pkt(ppkt, spkt, &mark, min_ack)) { |
465 | trace_colo_compare_tcp_info("pri", | |
466 | ppkt->tcp_seq, ppkt->tcp_ack, | |
467 | ppkt->header_size, ppkt->payload_size, | |
468 | ppkt->offset, ppkt->flags); | |
469 | ||
470 | trace_colo_compare_tcp_info("sec", | |
471 | spkt->tcp_seq, spkt->tcp_ack, | |
472 | spkt->header_size, spkt->payload_size, | |
473 | spkt->offset, spkt->flags); | |
474 | ||
475 | if (mark == COLO_COMPARE_FREE_PRIMARY) { | |
476 | conn->compare_seq = ppkt->seq_end; | |
477 | colo_release_primary_pkt(s, ppkt); | |
478 | g_queue_push_head(&conn->secondary_list, spkt); | |
479 | goto pri; | |
45b9e8c3 | 480 | } else if (mark == COLO_COMPARE_FREE_SECONDARY) { |
f449c9e5 MZ |
481 | conn->compare_seq = spkt->seq_end; |
482 | packet_destroy(spkt, NULL); | |
483 | goto sec; | |
45b9e8c3 | 484 | } else if (mark == (COLO_COMPARE_FREE_PRIMARY | COLO_COMPARE_FREE_SECONDARY)) { |
f449c9e5 MZ |
485 | conn->compare_seq = ppkt->seq_end; |
486 | colo_release_primary_pkt(s, ppkt); | |
487 | packet_destroy(spkt, NULL); | |
488 | goto pri; | |
489 | } | |
490 | } else { | |
491 | g_queue_push_head(&conn->primary_list, ppkt); | |
492 | g_queue_push_head(&conn->secondary_list, spkt); | |
2061c14c | 493 | |
8c8ed038 | 494 | #ifdef DEBUG_COLO_PACKETS |
842038f5 PM |
495 | qemu_hexdump(stderr, "colo-compare ppkt", ppkt->data, ppkt->size); |
496 | qemu_hexdump(stderr, "colo-compare spkt", spkt->data, spkt->size); | |
8c8ed038 | 497 | #endif |
f4b61836 | 498 | |
1d09f700 | 499 | colo_compare_inconsistency_notify(s); |
f449c9e5 | 500 | } |
f4b61836 ZC |
501 | } |
502 | ||
f449c9e5 | 503 | |
f4b61836 ZC |
504 | /* |
505 | * Called from the compare thread on the primary | |
506 | * for compare udp packet | |
507 | */ | |
508 | static int colo_packet_compare_udp(Packet *spkt, Packet *ppkt) | |
509 | { | |
9394133f MZ |
510 | uint16_t network_header_length = ppkt->ip->ip_hl << 2; |
511 | uint16_t offset = network_header_length + ETH_HLEN + ppkt->vnet_hdr_len; | |
f4b61836 ZC |
512 | |
513 | trace_colo_compare_main("compare udp"); | |
2ad7ca4c | 514 | |
6efeb328 ZC |
515 | /* |
516 | * Because of ppkt and spkt are both in the same connection, | |
517 | * The ppkt's src ip, dst ip, src port, dst port, ip_proto all are | |
518 | * same with spkt. In addition, IP header's Identification is a random | |
519 | * field, we can handle it in IP fragmentation function later. | |
520 | * COLO just concern the response net packet payload from primary guest | |
521 | * and secondary guest are same or not, So we ignored all IP header include | |
522 | * other field like TOS,TTL,IP Checksum. we only need to compare | |
523 | * the ip payload here. | |
524 | */ | |
9394133f MZ |
525 | if (ppkt->size != spkt->size) { |
526 | trace_colo_compare_main("UDP: payload size of packets are different"); | |
527 | return -1; | |
528 | } | |
529 | if (colo_compare_packet_payload(ppkt, spkt, offset, offset, | |
530 | ppkt->size - offset)) { | |
f4b61836 | 531 | trace_colo_compare_udp_miscompare("primary pkt size", ppkt->size); |
f4b61836 | 532 | trace_colo_compare_udp_miscompare("Secondary pkt size", spkt->size); |
8c8ed038 | 533 | #ifdef DEBUG_COLO_PACKETS |
842038f5 PM |
534 | qemu_hexdump(stderr, "colo-compare pri pkt", ppkt->data, ppkt->size); |
535 | qemu_hexdump(stderr, "colo-compare sec pkt", spkt->data, spkt->size); | |
8c8ed038 | 536 | #endif |
9394133f MZ |
537 | return -1; |
538 | } else { | |
539 | return 0; | |
f4b61836 | 540 | } |
f4b61836 ZC |
541 | } |
542 | ||
543 | /* | |
544 | * Called from the compare thread on the primary | |
545 | * for compare icmp packet | |
546 | */ | |
547 | static int colo_packet_compare_icmp(Packet *spkt, Packet *ppkt) | |
548 | { | |
9394133f MZ |
549 | uint16_t network_header_length = ppkt->ip->ip_hl << 2; |
550 | uint16_t offset = network_header_length + ETH_HLEN + ppkt->vnet_hdr_len; | |
6efeb328 | 551 | |
f4b61836 | 552 | trace_colo_compare_main("compare icmp"); |
f4b61836 | 553 | |
6efeb328 ZC |
554 | /* |
555 | * Because of ppkt and spkt are both in the same connection, | |
556 | * The ppkt's src ip, dst ip, src port, dst port, ip_proto all are | |
557 | * same with spkt. In addition, IP header's Identification is a random | |
558 | * field, we can handle it in IP fragmentation function later. | |
559 | * COLO just concern the response net packet payload from primary guest | |
560 | * and secondary guest are same or not, So we ignored all IP header include | |
561 | * other field like TOS,TTL,IP Checksum. we only need to compare | |
562 | * the ip payload here. | |
563 | */ | |
9394133f MZ |
564 | if (ppkt->size != spkt->size) { |
565 | trace_colo_compare_main("ICMP: payload size of packets are different"); | |
566 | return -1; | |
567 | } | |
568 | if (colo_compare_packet_payload(ppkt, spkt, offset, offset, | |
569 | ppkt->size - offset)) { | |
f4b61836 ZC |
570 | trace_colo_compare_icmp_miscompare("primary pkt size", |
571 | ppkt->size); | |
f4b61836 ZC |
572 | trace_colo_compare_icmp_miscompare("Secondary pkt size", |
573 | spkt->size); | |
8c8ed038 | 574 | #ifdef DEBUG_COLO_PACKETS |
842038f5 PM |
575 | qemu_hexdump(stderr, "colo-compare pri pkt", ppkt->data, ppkt->size); |
576 | qemu_hexdump(stderr, "colo-compare sec pkt", spkt->data, spkt->size); | |
8c8ed038 | 577 | #endif |
f4b61836 ZC |
578 | return -1; |
579 | } else { | |
580 | return 0; | |
581 | } | |
582 | } | |
583 | ||
584 | /* | |
585 | * Called from the compare thread on the primary | |
586 | * for compare other packet | |
587 | */ | |
588 | static int colo_packet_compare_other(Packet *spkt, Packet *ppkt) | |
589 | { | |
9394133f MZ |
590 | uint16_t offset = ppkt->vnet_hdr_len; |
591 | ||
f4b61836 | 592 | trace_colo_compare_main("compare other"); |
8c8ed038 | 593 | if (trace_event_get_state_backends(TRACE_COLO_COMPARE_IP_INFO)) { |
e630b2bf ZC |
594 | char pri_ip_src[20], pri_ip_dst[20], sec_ip_src[20], sec_ip_dst[20]; |
595 | ||
596 | strcpy(pri_ip_src, inet_ntoa(ppkt->ip->ip_src)); | |
597 | strcpy(pri_ip_dst, inet_ntoa(ppkt->ip->ip_dst)); | |
598 | strcpy(sec_ip_src, inet_ntoa(spkt->ip->ip_src)); | |
599 | strcpy(sec_ip_dst, inet_ntoa(spkt->ip->ip_dst)); | |
600 | ||
601 | trace_colo_compare_ip_info(ppkt->size, pri_ip_src, | |
602 | pri_ip_dst, spkt->size, | |
603 | sec_ip_src, sec_ip_dst); | |
604 | } | |
605 | ||
9394133f MZ |
606 | if (ppkt->size != spkt->size) { |
607 | trace_colo_compare_main("Other: payload size of packets are different"); | |
608 | return -1; | |
609 | } | |
610 | return colo_compare_packet_payload(ppkt, spkt, offset, offset, | |
611 | ppkt->size - offset); | |
0682e15b ZC |
612 | } |
613 | ||
614 | static int colo_old_packet_check_one(Packet *pkt, int64_t *check_time) | |
615 | { | |
616 | int64_t now = qemu_clock_get_ms(QEMU_CLOCK_HOST); | |
617 | ||
618 | if ((now - pkt->creation_ms) > (*check_time)) { | |
619 | trace_colo_old_packet_check_found(pkt->creation_ms); | |
620 | return 0; | |
621 | } else { | |
622 | return 1; | |
623 | } | |
624 | } | |
625 | ||
dccd0313 ZC |
626 | void colo_compare_register_notifier(Notifier *notify) |
627 | { | |
628 | notifier_list_add(&colo_compare_notifiers, notify); | |
629 | } | |
630 | ||
631 | void colo_compare_unregister_notifier(Notifier *notify) | |
632 | { | |
633 | notifier_remove(notify); | |
634 | } | |
635 | ||
d25a7dab | 636 | static int colo_old_packet_check_one_conn(Connection *conn, |
1d09f700 | 637 | CompareState *s) |
0682e15b | 638 | { |
17475df2 ZC |
639 | if (!g_queue_is_empty(&conn->primary_list)) { |
640 | if (g_queue_find_custom(&conn->primary_list, | |
641 | &s->compare_timeout, | |
642 | (GCompareFunc)colo_old_packet_check_one)) | |
643 | goto out; | |
644 | } | |
0682e15b | 645 | |
17475df2 ZC |
646 | if (!g_queue_is_empty(&conn->secondary_list)) { |
647 | if (g_queue_find_custom(&conn->secondary_list, | |
648 | &s->compare_timeout, | |
649 | (GCompareFunc)colo_old_packet_check_one)) | |
650 | goto out; | |
0682e15b | 651 | } |
d25a7dab ZC |
652 | |
653 | return 1; | |
17475df2 ZC |
654 | |
655 | out: | |
656 | /* Do checkpoint will flush old packet */ | |
657 | colo_compare_inconsistency_notify(s); | |
658 | return 0; | |
0682e15b ZC |
659 | } |
660 | ||
661 | /* | |
662 | * Look for old packets that the secondary hasn't matched, | |
663 | * if we have some then we have to checkpoint to wake | |
664 | * the secondary up. | |
665 | */ | |
666 | static void colo_old_packet_check(void *opaque) | |
667 | { | |
668 | CompareState *s = opaque; | |
669 | ||
d25a7dab ZC |
670 | /* |
671 | * If we find one old packet, stop finding job and notify | |
672 | * COLO frame do checkpoint. | |
673 | */ | |
1d09f700 | 674 | g_queue_find_custom(&s->conn_list, s, |
d25a7dab | 675 | (GCompareFunc)colo_old_packet_check_one_conn); |
0682e15b ZC |
676 | } |
677 | ||
f449c9e5 MZ |
678 | static void colo_compare_packet(CompareState *s, Connection *conn, |
679 | int (*HandlePacket)(Packet *spkt, | |
680 | Packet *ppkt)) | |
0682e15b | 681 | { |
0682e15b ZC |
682 | Packet *pkt = NULL; |
683 | GList *result = NULL; | |
0682e15b ZC |
684 | |
685 | while (!g_queue_is_empty(&conn->primary_list) && | |
686 | !g_queue_is_empty(&conn->secondary_list)) { | |
626bba98 | 687 | pkt = g_queue_pop_head(&conn->primary_list); |
f449c9e5 MZ |
688 | result = g_queue_find_custom(&conn->secondary_list, |
689 | pkt, (GCompareFunc)HandlePacket); | |
0682e15b ZC |
690 | |
691 | if (result) { | |
f449c9e5 | 692 | colo_release_primary_pkt(s, pkt); |
0682e15b | 693 | g_queue_remove(&conn->secondary_list, result->data); |
0682e15b ZC |
694 | } else { |
695 | /* | |
696 | * If one packet arrive late, the secondary_list or | |
697 | * primary_list will be empty, so we can't compare it | |
dccd0313 ZC |
698 | * until next comparison. If the packets in the list are |
699 | * timeout, it will trigger a checkpoint request. | |
0682e15b ZC |
700 | */ |
701 | trace_colo_compare_main("packet different"); | |
626bba98 | 702 | g_queue_push_head(&conn->primary_list, pkt); |
1d09f700 ZC |
703 | |
704 | colo_compare_inconsistency_notify(s); | |
0682e15b ZC |
705 | break; |
706 | } | |
707 | } | |
708 | } | |
709 | ||
f449c9e5 MZ |
710 | /* |
711 | * Called from the compare thread on the primary | |
712 | * for compare packet with secondary list of the | |
713 | * specified connection when a new packet was | |
714 | * queued to it. | |
715 | */ | |
716 | static void colo_compare_connection(void *opaque, void *user_data) | |
717 | { | |
718 | CompareState *s = user_data; | |
719 | Connection *conn = opaque; | |
720 | ||
721 | switch (conn->ip_proto) { | |
722 | case IPPROTO_TCP: | |
723 | colo_compare_tcp(s, conn); | |
724 | break; | |
725 | case IPPROTO_UDP: | |
726 | colo_compare_packet(s, conn, colo_packet_compare_udp); | |
727 | break; | |
728 | case IPPROTO_ICMP: | |
729 | colo_compare_packet(s, conn, colo_packet_compare_icmp); | |
730 | break; | |
731 | default: | |
732 | colo_compare_packet(s, conn, colo_packet_compare_other); | |
733 | break; | |
734 | } | |
735 | } | |
736 | ||
9c55fe94 | 737 | static void coroutine_fn _compare_chr_send(void *opaque) |
59509ec1 | 738 | { |
9c55fe94 LS |
739 | SendCo *sendco = opaque; |
740 | CompareState *s = sendco->s; | |
59509ec1 | 741 | int ret = 0; |
59509ec1 | 742 | |
9c55fe94 LS |
743 | while (!g_queue_is_empty(&sendco->send_list)) { |
744 | SendEntry *entry = g_queue_pop_tail(&sendco->send_list); | |
745 | uint32_t len = htonl(entry->size); | |
59509ec1 | 746 | |
9c55fe94 | 747 | ret = qemu_chr_fe_write_all(sendco->chr, (uint8_t *)&len, sizeof(len)); |
30685c00 | 748 | |
9c55fe94 LS |
749 | if (ret != sizeof(len)) { |
750 | g_free(entry->buf); | |
751 | g_slice_free(SendEntry, entry); | |
752 | goto err; | |
753 | } | |
59509ec1 | 754 | |
9c55fe94 LS |
755 | if (!sendco->notify_remote_frame && s->vnet_hdr) { |
756 | /* | |
757 | * We send vnet header len make other module(like filter-redirector) | |
758 | * know how to parse net packet correctly. | |
759 | */ | |
760 | len = htonl(entry->vnet_hdr_len); | |
30685c00 | 761 | |
9c55fe94 | 762 | ret = qemu_chr_fe_write_all(sendco->chr, |
30685c00 ZC |
763 | (uint8_t *)&len, |
764 | sizeof(len)); | |
9c55fe94 LS |
765 | |
766 | if (ret != sizeof(len)) { | |
767 | g_free(entry->buf); | |
768 | g_slice_free(SendEntry, entry); | |
769 | goto err; | |
770 | } | |
30685c00 ZC |
771 | } |
772 | ||
9c55fe94 LS |
773 | ret = qemu_chr_fe_write_all(sendco->chr, |
774 | (uint8_t *)entry->buf, | |
775 | entry->size); | |
776 | ||
777 | if (ret != entry->size) { | |
778 | g_free(entry->buf); | |
779 | g_slice_free(SendEntry, entry); | |
aa3a7032 ZC |
780 | goto err; |
781 | } | |
9c55fe94 LS |
782 | |
783 | g_free(entry->buf); | |
784 | g_slice_free(SendEntry, entry); | |
aa3a7032 ZC |
785 | } |
786 | ||
9c55fe94 LS |
787 | sendco->ret = 0; |
788 | goto out; | |
789 | ||
790 | err: | |
791 | while (!g_queue_is_empty(&sendco->send_list)) { | |
792 | SendEntry *entry = g_queue_pop_tail(&sendco->send_list); | |
793 | g_free(entry->buf); | |
794 | g_slice_free(SendEntry, entry); | |
795 | } | |
796 | sendco->ret = ret < 0 ? ret : -EIO; | |
797 | out: | |
798 | sendco->co = NULL; | |
799 | sendco->done = true; | |
800 | aio_wait_kick(); | |
801 | } | |
802 | ||
803 | static int compare_chr_send(CompareState *s, | |
804 | uint8_t *buf, | |
805 | uint32_t size, | |
806 | uint32_t vnet_hdr_len, | |
807 | bool notify_remote_frame, | |
808 | bool zero_copy) | |
809 | { | |
810 | SendCo *sendco; | |
811 | SendEntry *entry; | |
812 | ||
30685c00 | 813 | if (notify_remote_frame) { |
9c55fe94 | 814 | sendco = &s->notify_sendco; |
30685c00 | 815 | } else { |
9c55fe94 | 816 | sendco = &s->out_sendco; |
30685c00 ZC |
817 | } |
818 | ||
9c55fe94 LS |
819 | if (!size) { |
820 | return 0; | |
59509ec1 ZC |
821 | } |
822 | ||
9c55fe94 LS |
823 | entry = g_slice_new(SendEntry); |
824 | entry->size = size; | |
825 | entry->vnet_hdr_len = vnet_hdr_len; | |
826 | if (zero_copy) { | |
827 | entry->buf = buf; | |
828 | } else { | |
829 | entry->buf = g_malloc(size); | |
830 | memcpy(entry->buf, buf, size); | |
831 | } | |
832 | g_queue_push_head(&sendco->send_list, entry); | |
833 | ||
834 | if (sendco->done) { | |
835 | sendco->co = qemu_coroutine_create(_compare_chr_send, sendco); | |
836 | sendco->done = false; | |
837 | qemu_coroutine_enter(sendco->co); | |
838 | if (sendco->done) { | |
839 | /* report early errors */ | |
840 | return sendco->ret; | |
841 | } | |
842 | } | |
59509ec1 | 843 | |
9c55fe94 LS |
844 | /* assume success */ |
845 | return 0; | |
59509ec1 ZC |
846 | } |
847 | ||
0682e15b ZC |
848 | static int compare_chr_can_read(void *opaque) |
849 | { | |
850 | return COMPARE_READ_LEN_MAX; | |
851 | } | |
852 | ||
853 | /* | |
854 | * Called from the main thread on the primary for packets | |
855 | * arriving over the socket from the primary. | |
856 | */ | |
857 | static void compare_pri_chr_in(void *opaque, const uint8_t *buf, int size) | |
858 | { | |
859 | CompareState *s = COLO_COMPARE(opaque); | |
860 | int ret; | |
861 | ||
862 | ret = net_fill_rstate(&s->pri_rs, buf, size); | |
863 | if (ret == -1) { | |
81517ba3 | 864 | qemu_chr_fe_set_handlers(&s->chr_pri_in, NULL, NULL, NULL, NULL, |
39ab61c6 | 865 | NULL, NULL, true); |
0682e15b ZC |
866 | error_report("colo-compare primary_in error"); |
867 | } | |
868 | } | |
869 | ||
870 | /* | |
871 | * Called from the main thread on the primary for packets | |
872 | * arriving over the socket from the secondary. | |
873 | */ | |
874 | static void compare_sec_chr_in(void *opaque, const uint8_t *buf, int size) | |
875 | { | |
876 | CompareState *s = COLO_COMPARE(opaque); | |
877 | int ret; | |
878 | ||
879 | ret = net_fill_rstate(&s->sec_rs, buf, size); | |
880 | if (ret == -1) { | |
81517ba3 | 881 | qemu_chr_fe_set_handlers(&s->chr_sec_in, NULL, NULL, NULL, NULL, |
39ab61c6 | 882 | NULL, NULL, true); |
0682e15b ZC |
883 | error_report("colo-compare secondary_in error"); |
884 | } | |
885 | } | |
886 | ||
13025fee ZC |
887 | static void compare_notify_chr(void *opaque, const uint8_t *buf, int size) |
888 | { | |
889 | CompareState *s = COLO_COMPARE(opaque); | |
890 | int ret; | |
891 | ||
892 | ret = net_fill_rstate(&s->notify_rs, buf, size); | |
893 | if (ret == -1) { | |
894 | qemu_chr_fe_set_handlers(&s->chr_notify_dev, NULL, NULL, NULL, NULL, | |
895 | NULL, NULL, true); | |
896 | error_report("colo-compare notify_dev error"); | |
897 | } | |
898 | } | |
899 | ||
66d2a242 HZ |
900 | /* |
901 | * Check old packet regularly so it can watch for any packets | |
902 | * that the secondary hasn't produced equivalents of. | |
903 | */ | |
dd321ecf | 904 | static void check_old_packet_regular(void *opaque) |
66d2a242 HZ |
905 | { |
906 | CompareState *s = opaque; | |
907 | ||
908 | /* if have old packet we will notify checkpoint */ | |
909 | colo_old_packet_check(s); | |
ec081984 | 910 | timer_mod(s->packet_check_timer, qemu_clock_get_ms(QEMU_CLOCK_HOST) + |
cca35ac4 | 911 | s->expired_scan_cycle); |
dd321ecf WY |
912 | } |
913 | ||
0ffcece3 ZC |
914 | /* Public API, Used for COLO frame to notify compare event */ |
915 | void colo_notify_compares_event(void *opaque, int event, Error **errp) | |
916 | { | |
917 | CompareState *s; | |
45942b79 LS |
918 | qemu_mutex_lock(&colo_compare_mutex); |
919 | ||
920 | if (!colo_compare_active) { | |
921 | qemu_mutex_unlock(&colo_compare_mutex); | |
922 | return; | |
923 | } | |
0ffcece3 ZC |
924 | |
925 | qemu_mutex_lock(&event_mtx); | |
926 | QTAILQ_FOREACH(s, &net_compares, next) { | |
927 | s->event = event; | |
928 | qemu_bh_schedule(s->event_bh); | |
929 | event_unhandled_count++; | |
930 | } | |
931 | /* Wait all compare threads to finish handling this event */ | |
932 | while (event_unhandled_count > 0) { | |
933 | qemu_cond_wait(&event_complete_cond, &event_mtx); | |
934 | } | |
935 | ||
936 | qemu_mutex_unlock(&event_mtx); | |
45942b79 | 937 | qemu_mutex_unlock(&colo_compare_mutex); |
0ffcece3 ZC |
938 | } |
939 | ||
dd321ecf WY |
940 | static void colo_compare_timer_init(CompareState *s) |
941 | { | |
942 | AioContext *ctx = iothread_get_aio_context(s->iothread); | |
66d2a242 | 943 | |
ec081984 | 944 | s->packet_check_timer = aio_timer_new(ctx, QEMU_CLOCK_HOST, |
dd321ecf WY |
945 | SCALE_MS, check_old_packet_regular, |
946 | s); | |
ec081984 | 947 | timer_mod(s->packet_check_timer, qemu_clock_get_ms(QEMU_CLOCK_HOST) + |
cca35ac4 | 948 | s->expired_scan_cycle); |
66d2a242 HZ |
949 | } |
950 | ||
dd321ecf | 951 | static void colo_compare_timer_del(CompareState *s) |
0682e15b | 952 | { |
dd321ecf WY |
953 | if (s->packet_check_timer) { |
954 | timer_del(s->packet_check_timer); | |
955 | timer_free(s->packet_check_timer); | |
956 | s->packet_check_timer = NULL; | |
957 | } | |
958 | } | |
0682e15b | 959 | |
0ffcece3 ZC |
960 | static void colo_flush_packets(void *opaque, void *user_data); |
961 | ||
962 | static void colo_compare_handle_event(void *opaque) | |
963 | { | |
964 | CompareState *s = opaque; | |
965 | ||
966 | switch (s->event) { | |
967 | case COLO_EVENT_CHECKPOINT: | |
968 | g_queue_foreach(&s->conn_list, colo_flush_packets, s); | |
969 | break; | |
970 | case COLO_EVENT_FAILOVER: | |
971 | break; | |
972 | default: | |
973 | break; | |
974 | } | |
975 | ||
0ffcece3 | 976 | qemu_mutex_lock(&event_mtx); |
78e4f446 | 977 | assert(event_unhandled_count > 0); |
0ffcece3 ZC |
978 | event_unhandled_count--; |
979 | qemu_cond_broadcast(&event_complete_cond); | |
980 | qemu_mutex_unlock(&event_mtx); | |
981 | } | |
982 | ||
dd321ecf WY |
983 | static void colo_compare_iothread(CompareState *s) |
984 | { | |
5893c738 | 985 | AioContext *ctx = iothread_get_aio_context(s->iothread); |
dd321ecf WY |
986 | object_ref(OBJECT(s->iothread)); |
987 | s->worker_context = iothread_get_g_main_context(s->iothread); | |
0682e15b | 988 | |
5345fdb4 | 989 | qemu_chr_fe_set_handlers(&s->chr_pri_in, compare_chr_can_read, |
81517ba3 AN |
990 | compare_pri_chr_in, NULL, NULL, |
991 | s, s->worker_context, true); | |
5345fdb4 | 992 | qemu_chr_fe_set_handlers(&s->chr_sec_in, compare_chr_can_read, |
81517ba3 AN |
993 | compare_sec_chr_in, NULL, NULL, |
994 | s, s->worker_context, true); | |
13025fee ZC |
995 | if (s->notify_dev) { |
996 | qemu_chr_fe_set_handlers(&s->chr_notify_dev, compare_chr_can_read, | |
997 | compare_notify_chr, NULL, NULL, | |
998 | s, s->worker_context, true); | |
999 | } | |
0682e15b | 1000 | |
dd321ecf | 1001 | colo_compare_timer_init(s); |
5893c738 | 1002 | s->event_bh = aio_bh_new(ctx, colo_compare_handle_event, s); |
0682e15b ZC |
1003 | } |
1004 | ||
7dce4e6f ZC |
1005 | static char *compare_get_pri_indev(Object *obj, Error **errp) |
1006 | { | |
1007 | CompareState *s = COLO_COMPARE(obj); | |
1008 | ||
1009 | return g_strdup(s->pri_indev); | |
1010 | } | |
1011 | ||
1012 | static void compare_set_pri_indev(Object *obj, const char *value, Error **errp) | |
1013 | { | |
1014 | CompareState *s = COLO_COMPARE(obj); | |
1015 | ||
1016 | g_free(s->pri_indev); | |
1017 | s->pri_indev = g_strdup(value); | |
1018 | } | |
1019 | ||
1020 | static char *compare_get_sec_indev(Object *obj, Error **errp) | |
1021 | { | |
1022 | CompareState *s = COLO_COMPARE(obj); | |
1023 | ||
1024 | return g_strdup(s->sec_indev); | |
1025 | } | |
1026 | ||
1027 | static void compare_set_sec_indev(Object *obj, const char *value, Error **errp) | |
1028 | { | |
1029 | CompareState *s = COLO_COMPARE(obj); | |
1030 | ||
1031 | g_free(s->sec_indev); | |
1032 | s->sec_indev = g_strdup(value); | |
1033 | } | |
1034 | ||
1035 | static char *compare_get_outdev(Object *obj, Error **errp) | |
1036 | { | |
1037 | CompareState *s = COLO_COMPARE(obj); | |
1038 | ||
1039 | return g_strdup(s->outdev); | |
1040 | } | |
1041 | ||
1042 | static void compare_set_outdev(Object *obj, const char *value, Error **errp) | |
1043 | { | |
1044 | CompareState *s = COLO_COMPARE(obj); | |
1045 | ||
1046 | g_free(s->outdev); | |
1047 | s->outdev = g_strdup(value); | |
1048 | } | |
1049 | ||
aa3a7032 ZC |
1050 | static bool compare_get_vnet_hdr(Object *obj, Error **errp) |
1051 | { | |
1052 | CompareState *s = COLO_COMPARE(obj); | |
1053 | ||
1054 | return s->vnet_hdr; | |
1055 | } | |
1056 | ||
1057 | static void compare_set_vnet_hdr(Object *obj, | |
1058 | bool value, | |
1059 | Error **errp) | |
1060 | { | |
1061 | CompareState *s = COLO_COMPARE(obj); | |
1062 | ||
1063 | s->vnet_hdr = value; | |
1064 | } | |
1065 | ||
cf6af766 ZC |
1066 | static char *compare_get_notify_dev(Object *obj, Error **errp) |
1067 | { | |
1068 | CompareState *s = COLO_COMPARE(obj); | |
1069 | ||
1070 | return g_strdup(s->notify_dev); | |
1071 | } | |
1072 | ||
1073 | static void compare_set_notify_dev(Object *obj, const char *value, Error **errp) | |
1074 | { | |
1075 | CompareState *s = COLO_COMPARE(obj); | |
1076 | ||
1077 | g_free(s->notify_dev); | |
1078 | s->notify_dev = g_strdup(value); | |
1079 | } | |
1080 | ||
9cc43c94 ZC |
1081 | static void compare_get_timeout(Object *obj, Visitor *v, |
1082 | const char *name, void *opaque, | |
1083 | Error **errp) | |
1084 | { | |
1085 | CompareState *s = COLO_COMPARE(obj); | |
0c4266ef | 1086 | uint64_t value = s->compare_timeout; |
9cc43c94 | 1087 | |
0c4266ef | 1088 | visit_type_uint64(v, name, &value, errp); |
9cc43c94 ZC |
1089 | } |
1090 | ||
1091 | static void compare_set_timeout(Object *obj, Visitor *v, | |
1092 | const char *name, void *opaque, | |
1093 | Error **errp) | |
1094 | { | |
1095 | CompareState *s = COLO_COMPARE(obj); | |
9cc43c94 ZC |
1096 | uint32_t value; |
1097 | ||
668f62ec | 1098 | if (!visit_type_uint32(v, name, &value, errp)) { |
dcfe4805 | 1099 | return; |
9cc43c94 ZC |
1100 | } |
1101 | if (!value) { | |
dcfe4805 | 1102 | error_setg(errp, "Property '%s.%s' requires a positive value", |
9cc43c94 | 1103 | object_get_typename(obj), name); |
dcfe4805 | 1104 | return; |
9cc43c94 ZC |
1105 | } |
1106 | s->compare_timeout = value; | |
9cc43c94 ZC |
1107 | } |
1108 | ||
cca35ac4 ZC |
1109 | static void compare_get_expired_scan_cycle(Object *obj, Visitor *v, |
1110 | const char *name, void *opaque, | |
1111 | Error **errp) | |
1112 | { | |
1113 | CompareState *s = COLO_COMPARE(obj); | |
1114 | uint32_t value = s->expired_scan_cycle; | |
1115 | ||
1116 | visit_type_uint32(v, name, &value, errp); | |
1117 | } | |
1118 | ||
1119 | static void compare_set_expired_scan_cycle(Object *obj, Visitor *v, | |
1120 | const char *name, void *opaque, | |
1121 | Error **errp) | |
1122 | { | |
1123 | CompareState *s = COLO_COMPARE(obj); | |
cca35ac4 ZC |
1124 | uint32_t value; |
1125 | ||
668f62ec | 1126 | if (!visit_type_uint32(v, name, &value, errp)) { |
dcfe4805 | 1127 | return; |
cca35ac4 ZC |
1128 | } |
1129 | if (!value) { | |
dcfe4805 | 1130 | error_setg(errp, "Property '%s.%s' requires a positive value", |
cca35ac4 | 1131 | object_get_typename(obj), name); |
dcfe4805 | 1132 | return; |
cca35ac4 ZC |
1133 | } |
1134 | s->expired_scan_cycle = value; | |
cca35ac4 ZC |
1135 | } |
1136 | ||
a2e5cb7a ZC |
1137 | static void get_max_queue_size(Object *obj, Visitor *v, |
1138 | const char *name, void *opaque, | |
1139 | Error **errp) | |
1140 | { | |
1141 | uint32_t value = max_queue_size; | |
1142 | ||
1143 | visit_type_uint32(v, name, &value, errp); | |
1144 | } | |
1145 | ||
1146 | static void set_max_queue_size(Object *obj, Visitor *v, | |
1147 | const char *name, void *opaque, | |
1148 | Error **errp) | |
1149 | { | |
1150 | Error *local_err = NULL; | |
0c4266ef | 1151 | uint64_t value; |
a2e5cb7a | 1152 | |
0c4266ef | 1153 | visit_type_uint64(v, name, &value, &local_err); |
a2e5cb7a ZC |
1154 | if (local_err) { |
1155 | goto out; | |
1156 | } | |
1157 | if (!value) { | |
1158 | error_setg(&local_err, "Property '%s.%s' requires a positive value", | |
1159 | object_get_typename(obj), name); | |
1160 | goto out; | |
1161 | } | |
1162 | max_queue_size = value; | |
1163 | ||
1164 | out: | |
1165 | error_propagate(errp, local_err); | |
1166 | } | |
1167 | ||
7dce4e6f ZC |
1168 | static void compare_pri_rs_finalize(SocketReadState *pri_rs) |
1169 | { | |
59509ec1 | 1170 | CompareState *s = container_of(pri_rs, CompareState, pri_rs); |
8ec14402 | 1171 | Connection *conn = NULL; |
59509ec1 | 1172 | |
8ec14402 | 1173 | if (packet_enqueue(s, PRIMARY_IN, &conn)) { |
59509ec1 | 1174 | trace_colo_compare_main("primary: unsupported packet in"); |
aa3a7032 ZC |
1175 | compare_chr_send(s, |
1176 | pri_rs->buf, | |
1177 | pri_rs->packet_len, | |
30685c00 | 1178 | pri_rs->vnet_hdr_len, |
9c55fe94 | 1179 | false, |
30685c00 | 1180 | false); |
0682e15b | 1181 | } else { |
3463218c | 1182 | /* compare packet in the specified connection */ |
8ec14402 | 1183 | colo_compare_connection(conn, s); |
59509ec1 | 1184 | } |
7dce4e6f ZC |
1185 | } |
1186 | ||
1187 | static void compare_sec_rs_finalize(SocketReadState *sec_rs) | |
1188 | { | |
59509ec1 | 1189 | CompareState *s = container_of(sec_rs, CompareState, sec_rs); |
8ec14402 | 1190 | Connection *conn = NULL; |
59509ec1 | 1191 | |
8ec14402 | 1192 | if (packet_enqueue(s, SECONDARY_IN, &conn)) { |
59509ec1 | 1193 | trace_colo_compare_main("secondary: unsupported packet in"); |
0682e15b | 1194 | } else { |
3463218c | 1195 | /* compare packet in the specified connection */ |
8ec14402 | 1196 | colo_compare_connection(conn, s); |
59509ec1 | 1197 | } |
7dce4e6f ZC |
1198 | } |
1199 | ||
13025fee ZC |
1200 | static void compare_notify_rs_finalize(SocketReadState *notify_rs) |
1201 | { | |
1d09f700 ZC |
1202 | CompareState *s = container_of(notify_rs, CompareState, notify_rs); |
1203 | ||
f77bed14 | 1204 | const char msg[] = "COLO_COMPARE_GET_XEN_INIT"; |
1d09f700 ZC |
1205 | int ret; |
1206 | ||
f77bed14 ZC |
1207 | if (packet_matches_str("COLO_USERSPACE_PROXY_INIT", |
1208 | notify_rs->buf, | |
1209 | notify_rs->packet_len)) { | |
9c55fe94 | 1210 | ret = compare_chr_send(s, (uint8_t *)msg, strlen(msg), 0, true, false); |
1d09f700 ZC |
1211 | if (ret < 0) { |
1212 | error_report("Notify Xen COLO-frame INIT failed"); | |
1213 | } | |
f77bed14 ZC |
1214 | } else if (packet_matches_str("COLO_CHECKPOINT", |
1215 | notify_rs->buf, | |
1216 | notify_rs->packet_len)) { | |
1d09f700 ZC |
1217 | /* colo-compare do checkpoint, flush pri packet and remove sec packet */ |
1218 | g_queue_foreach(&s->conn_list, colo_flush_packets, s); | |
f77bed14 ZC |
1219 | } else { |
1220 | error_report("COLO compare got unsupported instruction"); | |
1d09f700 | 1221 | } |
13025fee | 1222 | } |
7dce4e6f ZC |
1223 | |
1224 | /* | |
1225 | * Return 0 is success. | |
1226 | * Return 1 is failed. | |
1227 | */ | |
0ec7b3e7 | 1228 | static int find_and_check_chardev(Chardev **chr, |
7dce4e6f ZC |
1229 | char *chr_name, |
1230 | Error **errp) | |
1231 | { | |
7dce4e6f ZC |
1232 | *chr = qemu_chr_find(chr_name); |
1233 | if (*chr == NULL) { | |
1234 | error_setg(errp, "Device '%s' not found", | |
1235 | chr_name); | |
1236 | return 1; | |
1237 | } | |
1238 | ||
0a73336d DB |
1239 | if (!qemu_chr_has_feature(*chr, QEMU_CHAR_FEATURE_RECONNECTABLE)) { |
1240 | error_setg(errp, "chardev \"%s\" is not reconnectable", | |
7dce4e6f ZC |
1241 | chr_name); |
1242 | return 1; | |
1243 | } | |
fbf3cc3a | 1244 | |
269d25cd MAL |
1245 | if (!qemu_chr_has_feature(*chr, QEMU_CHAR_FEATURE_GCONTEXT)) { |
1246 | error_setg(errp, "chardev \"%s\" cannot switch context", | |
1247 | chr_name); | |
1248 | return 1; | |
1249 | } | |
1250 | ||
7dce4e6f ZC |
1251 | return 0; |
1252 | } | |
1253 | ||
1254 | /* | |
1255 | * Called from the main thread on the primary | |
1256 | * to setup colo-compare. | |
1257 | */ | |
1258 | static void colo_compare_complete(UserCreatable *uc, Error **errp) | |
1259 | { | |
1260 | CompareState *s = COLO_COMPARE(uc); | |
0ec7b3e7 | 1261 | Chardev *chr; |
7dce4e6f | 1262 | |
dd321ecf | 1263 | if (!s->pri_indev || !s->sec_indev || !s->outdev || !s->iothread) { |
7dce4e6f | 1264 | error_setg(errp, "colo compare needs 'primary_in' ," |
dd321ecf | 1265 | "'secondary_in','outdev','iothread' property set"); |
7dce4e6f ZC |
1266 | return; |
1267 | } else if (!strcmp(s->pri_indev, s->outdev) || | |
1268 | !strcmp(s->sec_indev, s->outdev) || | |
1269 | !strcmp(s->pri_indev, s->sec_indev)) { | |
1270 | error_setg(errp, "'indev' and 'outdev' could not be same " | |
1271 | "for compare module"); | |
1272 | return; | |
1273 | } | |
1274 | ||
9cc43c94 ZC |
1275 | if (!s->compare_timeout) { |
1276 | /* Set default value to 3000 MS */ | |
1277 | s->compare_timeout = DEFAULT_TIME_OUT_MS; | |
1278 | } | |
1279 | ||
cca35ac4 ZC |
1280 | if (!s->expired_scan_cycle) { |
1281 | /* Set default value to 3000 MS */ | |
1282 | s->expired_scan_cycle = REGULAR_PACKET_CHECK_MS; | |
1283 | } | |
1284 | ||
a2e5cb7a ZC |
1285 | if (!max_queue_size) { |
1286 | /* Set default queue size to 1024 */ | |
1287 | max_queue_size = MAX_QUEUE_SIZE; | |
1288 | } | |
1289 | ||
5345fdb4 MAL |
1290 | if (find_and_check_chardev(&chr, s->pri_indev, errp) || |
1291 | !qemu_chr_fe_init(&s->chr_pri_in, chr, errp)) { | |
7dce4e6f ZC |
1292 | return; |
1293 | } | |
1294 | ||
5345fdb4 MAL |
1295 | if (find_and_check_chardev(&chr, s->sec_indev, errp) || |
1296 | !qemu_chr_fe_init(&s->chr_sec_in, chr, errp)) { | |
7dce4e6f ZC |
1297 | return; |
1298 | } | |
1299 | ||
5345fdb4 MAL |
1300 | if (find_and_check_chardev(&chr, s->outdev, errp) || |
1301 | !qemu_chr_fe_init(&s->chr_out, chr, errp)) { | |
7dce4e6f ZC |
1302 | return; |
1303 | } | |
1304 | ||
aa3a7032 ZC |
1305 | net_socket_rs_init(&s->pri_rs, compare_pri_rs_finalize, s->vnet_hdr); |
1306 | net_socket_rs_init(&s->sec_rs, compare_sec_rs_finalize, s->vnet_hdr); | |
7dce4e6f | 1307 | |
13025fee ZC |
1308 | /* Try to enable remote notify chardev, currently just for Xen COLO */ |
1309 | if (s->notify_dev) { | |
1310 | if (find_and_check_chardev(&chr, s->notify_dev, errp) || | |
1311 | !qemu_chr_fe_init(&s->chr_notify_dev, chr, errp)) { | |
1312 | return; | |
1313 | } | |
1314 | ||
1315 | net_socket_rs_init(&s->notify_rs, compare_notify_rs_finalize, | |
1316 | s->vnet_hdr); | |
1317 | } | |
1318 | ||
9c55fe94 LS |
1319 | s->out_sendco.s = s; |
1320 | s->out_sendco.chr = &s->chr_out; | |
1321 | s->out_sendco.notify_remote_frame = false; | |
1322 | s->out_sendco.done = true; | |
1323 | g_queue_init(&s->out_sendco.send_list); | |
1324 | ||
1325 | if (s->notify_dev) { | |
1326 | s->notify_sendco.s = s; | |
1327 | s->notify_sendco.chr = &s->chr_notify_dev; | |
1328 | s->notify_sendco.notify_remote_frame = true; | |
1329 | s->notify_sendco.done = true; | |
1330 | g_queue_init(&s->notify_sendco.send_list); | |
1331 | } | |
1332 | ||
b6540d40 ZC |
1333 | g_queue_init(&s->conn_list); |
1334 | ||
1335 | s->connection_track_table = g_hash_table_new_full(connection_key_hash, | |
1336 | connection_key_equal, | |
1337 | g_free, | |
1338 | connection_destroy); | |
59509ec1 | 1339 | |
dd321ecf | 1340 | colo_compare_iothread(s); |
5bd57eba LS |
1341 | |
1342 | qemu_mutex_lock(&colo_compare_mutex); | |
1343 | if (!colo_compare_active) { | |
1344 | qemu_mutex_init(&event_mtx); | |
1345 | qemu_cond_init(&event_complete_cond); | |
1346 | colo_compare_active = true; | |
1347 | } | |
1348 | QTAILQ_INSERT_TAIL(&net_compares, s, next); | |
1349 | qemu_mutex_unlock(&colo_compare_mutex); | |
1350 | ||
7dce4e6f ZC |
1351 | return; |
1352 | } | |
1353 | ||
dfd917a9 HZ |
1354 | static void colo_flush_packets(void *opaque, void *user_data) |
1355 | { | |
1356 | CompareState *s = user_data; | |
1357 | Connection *conn = opaque; | |
1358 | Packet *pkt = NULL; | |
1359 | ||
1360 | while (!g_queue_is_empty(&conn->primary_list)) { | |
1361 | pkt = g_queue_pop_head(&conn->primary_list); | |
aa3a7032 ZC |
1362 | compare_chr_send(s, |
1363 | pkt->data, | |
1364 | pkt->size, | |
30685c00 | 1365 | pkt->vnet_hdr_len, |
9c55fe94 LS |
1366 | false, |
1367 | true); | |
1368 | packet_destroy_partial(pkt, NULL); | |
dfd917a9 HZ |
1369 | } |
1370 | while (!g_queue_is_empty(&conn->secondary_list)) { | |
1371 | pkt = g_queue_pop_head(&conn->secondary_list); | |
1372 | packet_destroy(pkt, NULL); | |
1373 | } | |
1374 | } | |
1375 | ||
7dce4e6f ZC |
1376 | static void colo_compare_class_init(ObjectClass *oc, void *data) |
1377 | { | |
1378 | UserCreatableClass *ucc = USER_CREATABLE_CLASS(oc); | |
1379 | ||
1380 | ucc->complete = colo_compare_complete; | |
1381 | } | |
1382 | ||
1383 | static void colo_compare_init(Object *obj) | |
1384 | { | |
aa3a7032 ZC |
1385 | CompareState *s = COLO_COMPARE(obj); |
1386 | ||
7dce4e6f | 1387 | object_property_add_str(obj, "primary_in", |
d2623129 | 1388 | compare_get_pri_indev, compare_set_pri_indev); |
7dce4e6f | 1389 | object_property_add_str(obj, "secondary_in", |
d2623129 | 1390 | compare_get_sec_indev, compare_set_sec_indev); |
7dce4e6f | 1391 | object_property_add_str(obj, "outdev", |
d2623129 | 1392 | compare_get_outdev, compare_set_outdev); |
dd321ecf WY |
1393 | object_property_add_link(obj, "iothread", TYPE_IOTHREAD, |
1394 | (Object **)&s->iothread, | |
1395 | object_property_allow_set_link, | |
d2623129 | 1396 | OBJ_PROP_LINK_STRONG); |
cf6af766 ZC |
1397 | /* This parameter just for Xen COLO */ |
1398 | object_property_add_str(obj, "notify_dev", | |
d2623129 | 1399 | compare_get_notify_dev, compare_set_notify_dev); |
aa3a7032 | 1400 | |
0c4266ef | 1401 | object_property_add(obj, "compare_timeout", "uint64", |
9cc43c94 | 1402 | compare_get_timeout, |
d2623129 | 1403 | compare_set_timeout, NULL, NULL); |
9cc43c94 | 1404 | |
cca35ac4 ZC |
1405 | object_property_add(obj, "expired_scan_cycle", "uint32", |
1406 | compare_get_expired_scan_cycle, | |
d2623129 | 1407 | compare_set_expired_scan_cycle, NULL, NULL); |
cca35ac4 | 1408 | |
a2e5cb7a ZC |
1409 | object_property_add(obj, "max_queue_size", "uint32", |
1410 | get_max_queue_size, | |
1411 | set_max_queue_size, NULL, NULL); | |
1412 | ||
aa3a7032 ZC |
1413 | s->vnet_hdr = false; |
1414 | object_property_add_bool(obj, "vnet_hdr_support", compare_get_vnet_hdr, | |
d2623129 | 1415 | compare_set_vnet_hdr); |
7dce4e6f ZC |
1416 | } |
1417 | ||
1418 | static void colo_compare_finalize(Object *obj) | |
1419 | { | |
1420 | CompareState *s = COLO_COMPARE(obj); | |
0ffcece3 | 1421 | CompareState *tmp = NULL; |
7dce4e6f | 1422 | |
45942b79 | 1423 | qemu_mutex_lock(&colo_compare_mutex); |
0ffcece3 ZC |
1424 | QTAILQ_FOREACH(tmp, &net_compares, next) { |
1425 | if (tmp == s) { | |
1426 | QTAILQ_REMOVE(&net_compares, s, next); | |
1427 | break; | |
1428 | } | |
1429 | } | |
45942b79 LS |
1430 | if (QTAILQ_EMPTY(&net_compares)) { |
1431 | colo_compare_active = false; | |
1432 | qemu_mutex_destroy(&event_mtx); | |
1433 | qemu_cond_destroy(&event_complete_cond); | |
1434 | } | |
1435 | qemu_mutex_unlock(&colo_compare_mutex); | |
0ffcece3 | 1436 | |
5bd57eba LS |
1437 | qemu_chr_fe_deinit(&s->chr_pri_in, false); |
1438 | qemu_chr_fe_deinit(&s->chr_sec_in, false); | |
1439 | qemu_chr_fe_deinit(&s->chr_out, false); | |
1440 | if (s->notify_dev) { | |
1441 | qemu_chr_fe_deinit(&s->chr_notify_dev, false); | |
1442 | } | |
1443 | ||
f81cddfe | 1444 | colo_compare_timer_del(s); |
5bd57eba LS |
1445 | |
1446 | qemu_bh_delete(s->event_bh); | |
1447 | ||
9c55fe94 LS |
1448 | AioContext *ctx = iothread_get_aio_context(s->iothread); |
1449 | aio_context_acquire(ctx); | |
1450 | AIO_WAIT_WHILE(ctx, !s->out_sendco.done); | |
1451 | if (s->notify_dev) { | |
1452 | AIO_WAIT_WHILE(ctx, !s->notify_sendco.done); | |
1453 | } | |
1454 | aio_context_release(ctx); | |
1455 | ||
dfd917a9 HZ |
1456 | /* Release all unhandled packets after compare thead exited */ |
1457 | g_queue_foreach(&s->conn_list, colo_flush_packets, s); | |
9c55fe94 | 1458 | AIO_WAIT_WHILE(NULL, !s->out_sendco.done); |
dfd917a9 | 1459 | |
727c2d76 | 1460 | g_queue_clear(&s->conn_list); |
9c55fe94 LS |
1461 | g_queue_clear(&s->out_sendco.send_list); |
1462 | if (s->notify_dev) { | |
1463 | g_queue_clear(&s->notify_sendco.send_list); | |
1464 | } | |
0682e15b | 1465 | |
dd321ecf WY |
1466 | if (s->connection_track_table) { |
1467 | g_hash_table_destroy(s->connection_track_table); | |
1468 | } | |
1469 | ||
f81cddfe | 1470 | object_unref(OBJECT(s->iothread)); |
0ffcece3 | 1471 | |
7dce4e6f ZC |
1472 | g_free(s->pri_indev); |
1473 | g_free(s->sec_indev); | |
1474 | g_free(s->outdev); | |
cf6af766 | 1475 | g_free(s->notify_dev); |
7dce4e6f ZC |
1476 | } |
1477 | ||
45942b79 LS |
1478 | static void __attribute__((__constructor__)) colo_compare_init_globals(void) |
1479 | { | |
1480 | colo_compare_active = false; | |
1481 | qemu_mutex_init(&colo_compare_mutex); | |
1482 | } | |
1483 | ||
7dce4e6f ZC |
1484 | static const TypeInfo colo_compare_info = { |
1485 | .name = TYPE_COLO_COMPARE, | |
1486 | .parent = TYPE_OBJECT, | |
1487 | .instance_size = sizeof(CompareState), | |
1488 | .instance_init = colo_compare_init, | |
1489 | .instance_finalize = colo_compare_finalize, | |
1490 | .class_size = sizeof(CompareClass), | |
1491 | .class_init = colo_compare_class_init, | |
1492 | .interfaces = (InterfaceInfo[]) { | |
1493 | { TYPE_USER_CREATABLE }, | |
1494 | { } | |
1495 | } | |
1496 | }; | |
1497 | ||
1498 | static void register_types(void) | |
1499 | { | |
1500 | type_register_static(&colo_compare_info); | |
1501 | } | |
1502 | ||
1503 | type_init(register_types); |