| 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" |
| 16 | #include "qemu/error-report.h" |
| 17 | #include "trace.h" |
| 18 | #include "qapi/error.h" |
| 19 | #include "net/net.h" |
| 20 | #include "net/eth.h" |
| 21 | #include "qom/object_interfaces.h" |
| 22 | #include "qemu/iov.h" |
| 23 | #include "qom/object.h" |
| 24 | #include "net/queue.h" |
| 25 | #include "chardev/char-fe.h" |
| 26 | #include "qemu/sockets.h" |
| 27 | #include "colo.h" |
| 28 | #include "system/iothread.h" |
| 29 | #include "net/colo-compare.h" |
| 30 | #include "migration/colo.h" |
| 31 | #include "util.h" |
| 32 | |
| 33 | #include "qemu/aio-wait.h" |
| 34 | #include "qemu/coroutine.h" |
| 35 | |
| 36 | #define TYPE_COLO_COMPARE "colo-compare" |
| 37 | typedef struct CompareState CompareState; |
| 38 | DECLARE_INSTANCE_CHECKER(CompareState, COLO_COMPARE, |
| 39 | TYPE_COLO_COMPARE) |
| 40 | |
| 41 | static QTAILQ_HEAD(, CompareState) net_compares = |
| 42 | QTAILQ_HEAD_INITIALIZER(net_compares); |
| 43 | |
| 44 | static NotifierList colo_compare_notifiers = |
| 45 | NOTIFIER_LIST_INITIALIZER(colo_compare_notifiers); |
| 46 | |
| 47 | #define COMPARE_READ_LEN_MAX NET_BUFSIZE |
| 48 | #define MAX_QUEUE_SIZE 1024 |
| 49 | |
| 50 | #define COLO_COMPARE_FREE_PRIMARY 0x01 |
| 51 | #define COLO_COMPARE_FREE_SECONDARY 0x02 |
| 52 | |
| 53 | #define REGULAR_PACKET_CHECK_MS 1000 |
| 54 | #define DEFAULT_TIME_OUT_MS 3000 |
| 55 | |
| 56 | /* #define DEBUG_COLO_PACKETS */ |
| 57 | |
| 58 | static QemuMutex colo_compare_mutex; |
| 59 | static bool colo_compare_active; |
| 60 | static QemuMutex event_mtx; |
| 61 | static QemuCond event_complete_cond; |
| 62 | static int event_unhandled_count; |
| 63 | static uint32_t max_queue_size; |
| 64 | |
| 65 | /* |
| 66 | * + CompareState ++ |
| 67 | * | | |
| 68 | * +---------------+ +---------------+ +---------------+ |
| 69 | * | conn list + - > conn + ------- > conn + -- > ...... |
| 70 | * +---------------+ +---------------+ +---------------+ |
| 71 | * | | | | | | |
| 72 | * +---------------+ +---v----+ +---v----+ +---v----+ +---v----+ |
| 73 | * |primary | |secondary |primary | |secondary |
| 74 | * |packet | |packet + |packet | |packet + |
| 75 | * +--------+ +--------+ +--------+ +--------+ |
| 76 | * | | | | |
| 77 | * +---v----+ +---v----+ +---v----+ +---v----+ |
| 78 | * |primary | |secondary |primary | |secondary |
| 79 | * |packet | |packet + |packet | |packet + |
| 80 | * +--------+ +--------+ +--------+ +--------+ |
| 81 | * | | | | |
| 82 | * +---v----+ +---v----+ +---v----+ +---v----+ |
| 83 | * |primary | |secondary |primary | |secondary |
| 84 | * |packet | |packet + |packet | |packet + |
| 85 | * +--------+ +--------+ +--------+ +--------+ |
| 86 | */ |
| 87 | |
| 88 | typedef struct SendCo { |
| 89 | Coroutine *co; |
| 90 | struct CompareState *s; |
| 91 | CharFrontend *chr; |
| 92 | GQueue send_list; |
| 93 | bool notify_remote_frame; |
| 94 | bool done; |
| 95 | int ret; |
| 96 | } SendCo; |
| 97 | |
| 98 | typedef struct SendEntry { |
| 99 | uint32_t size; |
| 100 | uint32_t vnet_hdr_len; |
| 101 | uint8_t *buf; |
| 102 | } SendEntry; |
| 103 | |
| 104 | struct CompareState { |
| 105 | Object parent; |
| 106 | |
| 107 | char *pri_indev; |
| 108 | char *sec_indev; |
| 109 | char *outdev; |
| 110 | char *notify_dev; |
| 111 | CharFrontend chr_pri_in; |
| 112 | CharFrontend chr_sec_in; |
| 113 | CharFrontend chr_out; |
| 114 | CharFrontend chr_notify_dev; |
| 115 | SocketReadState pri_rs; |
| 116 | SocketReadState sec_rs; |
| 117 | SocketReadState notify_rs; |
| 118 | SendCo out_sendco; |
| 119 | SendCo notify_sendco; |
| 120 | bool vnet_hdr; |
| 121 | uint64_t compare_timeout; |
| 122 | uint32_t expired_scan_cycle; |
| 123 | |
| 124 | /* |
| 125 | * Record the connection that through the NIC |
| 126 | * Element type: Connection |
| 127 | */ |
| 128 | GQueue conn_list; |
| 129 | /* Record the connection without repetition */ |
| 130 | GHashTable *connection_track_table; |
| 131 | |
| 132 | IOThread *iothread; |
| 133 | GMainContext *worker_context; |
| 134 | QEMUTimer *packet_check_timer; |
| 135 | |
| 136 | QEMUBH *event_bh; |
| 137 | enum colo_event event; |
| 138 | |
| 139 | QTAILQ_ENTRY(CompareState) next; |
| 140 | }; |
| 141 | |
| 142 | typedef struct CompareClass { |
| 143 | ObjectClass parent_class; |
| 144 | } CompareClass; |
| 145 | |
| 146 | enum { |
| 147 | PRIMARY_IN = 0, |
| 148 | SECONDARY_IN, |
| 149 | }; |
| 150 | |
| 151 | static const char *colo_mode[] = { |
| 152 | [PRIMARY_IN] = "primary", |
| 153 | [SECONDARY_IN] = "secondary", |
| 154 | }; |
| 155 | |
| 156 | static int compare_chr_send(CompareState *s, |
| 157 | uint8_t *buf, |
| 158 | uint32_t size, |
| 159 | uint32_t vnet_hdr_len, |
| 160 | bool notify_remote_frame, |
| 161 | bool zero_copy); |
| 162 | |
| 163 | static bool packet_matches_str(const char *str, |
| 164 | const uint8_t *buf, |
| 165 | uint32_t packet_len) |
| 166 | { |
| 167 | if (packet_len != strlen(str)) { |
| 168 | return false; |
| 169 | } |
| 170 | |
| 171 | return !memcmp(str, buf, packet_len); |
| 172 | } |
| 173 | |
| 174 | static void notify_remote_frame(CompareState *s) |
| 175 | { |
| 176 | char msg[] = "DO_CHECKPOINT"; |
| 177 | int ret = 0; |
| 178 | |
| 179 | ret = compare_chr_send(s, (uint8_t *)msg, strlen(msg), 0, true, false); |
| 180 | if (ret < 0) { |
| 181 | error_report("Notify Xen COLO-frame failed"); |
| 182 | } |
| 183 | } |
| 184 | |
| 185 | static void colo_compare_inconsistency_notify(CompareState *s) |
| 186 | { |
| 187 | if (s->notify_dev) { |
| 188 | notify_remote_frame(s); |
| 189 | } else { |
| 190 | notifier_list_notify(&colo_compare_notifiers, |
| 191 | NULL); |
| 192 | } |
| 193 | } |
| 194 | |
| 195 | /* Use restricted to colo_insert_packet() */ |
| 196 | static gint seq_sorter(Packet *a, Packet *b, gpointer data) |
| 197 | { |
| 198 | return b->tcp_seq - a->tcp_seq; |
| 199 | } |
| 200 | |
| 201 | static void fill_pkt_tcp_info(void *data, uint32_t *max_ack) |
| 202 | { |
| 203 | Packet *pkt = data; |
| 204 | struct tcp_hdr *tcphd; |
| 205 | |
| 206 | tcphd = (struct tcp_hdr *)pkt->transport_header; |
| 207 | |
| 208 | pkt->tcp_seq = ntohl(tcphd->th_seq); |
| 209 | pkt->tcp_ack = ntohl(tcphd->th_ack); |
| 210 | /* Need to consider ACK will bigger than uint32_t MAX */ |
| 211 | *max_ack = pkt->tcp_ack - *max_ack > 0 ? pkt->tcp_ack : *max_ack; |
| 212 | pkt->header_size = pkt->transport_header - (uint8_t *)pkt->data |
| 213 | + (tcphd->th_off << 2); |
| 214 | pkt->payload_size = pkt->size - pkt->header_size; |
| 215 | pkt->seq_end = pkt->tcp_seq + pkt->payload_size; |
| 216 | pkt->flags = tcphd->th_flags; |
| 217 | } |
| 218 | |
| 219 | /* |
| 220 | * Return 1 on success, if return 0 means the |
| 221 | * packet will be dropped |
| 222 | */ |
| 223 | static int colo_insert_packet(GQueue *queue, Packet *pkt, uint32_t *max_ack) |
| 224 | { |
| 225 | if (g_queue_get_length(queue) <= max_queue_size) { |
| 226 | if (pkt->ip->ip_p == IPPROTO_TCP) { |
| 227 | fill_pkt_tcp_info(pkt, max_ack); |
| 228 | g_queue_insert_sorted(queue, |
| 229 | pkt, |
| 230 | (GCompareDataFunc)seq_sorter, |
| 231 | NULL); |
| 232 | } else { |
| 233 | g_queue_push_tail(queue, pkt); |
| 234 | } |
| 235 | return 1; |
| 236 | } |
| 237 | return 0; |
| 238 | } |
| 239 | |
| 240 | /* |
| 241 | * Return 0 on success, if return -1 means the pkt |
| 242 | * is unsupported(arp and ipv6) and will be sent later |
| 243 | */ |
| 244 | static int packet_enqueue(CompareState *s, int mode, Connection **con) |
| 245 | { |
| 246 | ConnectionKey key; |
| 247 | Packet *pkt = NULL; |
| 248 | Connection *conn; |
| 249 | int ret; |
| 250 | |
| 251 | if (mode == PRIMARY_IN) { |
| 252 | pkt = packet_new(s->pri_rs.buf, |
| 253 | s->pri_rs.packet_len, |
| 254 | s->pri_rs.vnet_hdr_len); |
| 255 | } else { |
| 256 | pkt = packet_new(s->sec_rs.buf, |
| 257 | s->sec_rs.packet_len, |
| 258 | s->sec_rs.vnet_hdr_len); |
| 259 | } |
| 260 | |
| 261 | if (parse_packet_early(pkt)) { |
| 262 | packet_destroy(pkt, NULL); |
| 263 | pkt = NULL; |
| 264 | return -1; |
| 265 | } |
| 266 | fill_connection_key(pkt, &key, false); |
| 267 | |
| 268 | conn = connection_get(s->connection_track_table, |
| 269 | &key, |
| 270 | &s->conn_list); |
| 271 | |
| 272 | if (!conn->processing) { |
| 273 | g_queue_push_tail(&s->conn_list, conn); |
| 274 | conn->processing = true; |
| 275 | } |
| 276 | |
| 277 | if (mode == PRIMARY_IN) { |
| 278 | ret = colo_insert_packet(&conn->primary_list, pkt, &conn->pack); |
| 279 | } else { |
| 280 | ret = colo_insert_packet(&conn->secondary_list, pkt, &conn->sack); |
| 281 | } |
| 282 | |
| 283 | if (!ret) { |
| 284 | trace_colo_compare_drop_packet(colo_mode[mode], |
| 285 | "queue size too big, drop packet"); |
| 286 | packet_destroy(pkt, NULL); |
| 287 | pkt = NULL; |
| 288 | } |
| 289 | |
| 290 | *con = conn; |
| 291 | |
| 292 | return 0; |
| 293 | } |
| 294 | |
| 295 | static inline bool after(uint32_t seq1, uint32_t seq2) |
| 296 | { |
| 297 | return (int32_t)(seq1 - seq2) > 0; |
| 298 | } |
| 299 | |
| 300 | static void colo_release_primary_pkt(CompareState *s, Packet *pkt) |
| 301 | { |
| 302 | int ret; |
| 303 | ret = compare_chr_send(s, |
| 304 | pkt->data, |
| 305 | pkt->size, |
| 306 | pkt->vnet_hdr_len, |
| 307 | false, |
| 308 | true); |
| 309 | if (ret < 0) { |
| 310 | error_report("colo send primary packet failed"); |
| 311 | } |
| 312 | trace_colo_compare_main("packet same and release packet"); |
| 313 | packet_destroy_partial(pkt, NULL); |
| 314 | } |
| 315 | |
| 316 | /* |
| 317 | * The IP packets sent by primary and secondary |
| 318 | * will be compared in here |
| 319 | * TODO support ip fragment, Out-Of-Order |
| 320 | * return: 0 means packet same |
| 321 | * > 0 || < 0 means packet different |
| 322 | */ |
| 323 | static int colo_compare_packet_payload(Packet *ppkt, |
| 324 | Packet *spkt, |
| 325 | uint16_t poffset, |
| 326 | uint16_t soffset, |
| 327 | uint16_t len) |
| 328 | |
| 329 | { |
| 330 | if (trace_event_get_state_backends(TRACE_COLO_COMPARE_IP_INFO)) { |
| 331 | char pri_ip_src[20], pri_ip_dst[20], sec_ip_src[20], sec_ip_dst[20]; |
| 332 | |
| 333 | strcpy(pri_ip_src, inet_ntoa(ppkt->ip->ip_src)); |
| 334 | strcpy(pri_ip_dst, inet_ntoa(ppkt->ip->ip_dst)); |
| 335 | strcpy(sec_ip_src, inet_ntoa(spkt->ip->ip_src)); |
| 336 | strcpy(sec_ip_dst, inet_ntoa(spkt->ip->ip_dst)); |
| 337 | |
| 338 | trace_colo_compare_ip_info(ppkt->size, pri_ip_src, |
| 339 | pri_ip_dst, spkt->size, |
| 340 | sec_ip_src, sec_ip_dst); |
| 341 | } |
| 342 | |
| 343 | return memcmp(ppkt->data + poffset, spkt->data + soffset, len); |
| 344 | } |
| 345 | |
| 346 | /* |
| 347 | * return true means that the payload is consist and |
| 348 | * need to make the next comparison, false means do |
| 349 | * the checkpoint |
| 350 | */ |
| 351 | static bool colo_mark_tcp_pkt(Packet *ppkt, Packet *spkt, |
| 352 | int8_t *mark, uint32_t max_ack) |
| 353 | { |
| 354 | *mark = 0; |
| 355 | |
| 356 | if (ppkt->tcp_seq == spkt->tcp_seq && ppkt->seq_end == spkt->seq_end) { |
| 357 | if (!colo_compare_packet_payload(ppkt, spkt, |
| 358 | ppkt->header_size, spkt->header_size, |
| 359 | ppkt->payload_size)) { |
| 360 | *mark = COLO_COMPARE_FREE_SECONDARY | COLO_COMPARE_FREE_PRIMARY; |
| 361 | return true; |
| 362 | } |
| 363 | } |
| 364 | |
| 365 | /* one part of secondary packet payload still need to be compared */ |
| 366 | if (!after(ppkt->seq_end, spkt->seq_end)) { |
| 367 | if (!colo_compare_packet_payload(ppkt, spkt, |
| 368 | ppkt->header_size + ppkt->offset, |
| 369 | spkt->header_size + spkt->offset, |
| 370 | ppkt->payload_size - ppkt->offset)) { |
| 371 | if (!after(ppkt->tcp_ack, max_ack)) { |
| 372 | *mark = COLO_COMPARE_FREE_PRIMARY; |
| 373 | spkt->offset += ppkt->payload_size - ppkt->offset; |
| 374 | return true; |
| 375 | } else { |
| 376 | /* secondary guest hasn't ack the data, don't send |
| 377 | * out this packet |
| 378 | */ |
| 379 | return false; |
| 380 | } |
| 381 | } |
| 382 | } else { |
| 383 | /* primary packet is longer than secondary packet, compare |
| 384 | * the same part and mark the primary packet offset |
| 385 | */ |
| 386 | if (!colo_compare_packet_payload(ppkt, spkt, |
| 387 | ppkt->header_size + ppkt->offset, |
| 388 | spkt->header_size + spkt->offset, |
| 389 | spkt->payload_size - spkt->offset)) { |
| 390 | *mark = COLO_COMPARE_FREE_SECONDARY; |
| 391 | ppkt->offset += spkt->payload_size - spkt->offset; |
| 392 | return true; |
| 393 | } |
| 394 | } |
| 395 | |
| 396 | return false; |
| 397 | } |
| 398 | |
| 399 | static void colo_compare_tcp(CompareState *s, Connection *conn) |
| 400 | { |
| 401 | Packet *ppkt = NULL, *spkt = NULL; |
| 402 | int8_t mark; |
| 403 | |
| 404 | /* |
| 405 | * If ppkt and spkt have the same payload, but ppkt's ACK |
| 406 | * is greater than spkt's ACK, in this case we can not |
| 407 | * send the ppkt because it will cause the secondary guest |
| 408 | * to miss sending some data in the next. Therefore, we |
| 409 | * record the maximum ACK in the current queue at both |
| 410 | * primary side and secondary side. Only when the ack is |
| 411 | * less than the smaller of the two maximum ack, then we |
| 412 | * can ensure that the packet's payload is acknowledged by |
| 413 | * primary and secondary. |
| 414 | */ |
| 415 | uint32_t min_ack = MIN(conn->pack, conn->sack); |
| 416 | |
| 417 | pri: |
| 418 | if (g_queue_is_empty(&conn->primary_list)) { |
| 419 | return; |
| 420 | } |
| 421 | ppkt = g_queue_pop_tail(&conn->primary_list); |
| 422 | sec: |
| 423 | if (g_queue_is_empty(&conn->secondary_list)) { |
| 424 | g_queue_push_tail(&conn->primary_list, ppkt); |
| 425 | return; |
| 426 | } |
| 427 | spkt = g_queue_pop_tail(&conn->secondary_list); |
| 428 | |
| 429 | if (ppkt->tcp_seq == ppkt->seq_end) { |
| 430 | colo_release_primary_pkt(s, ppkt); |
| 431 | ppkt = NULL; |
| 432 | } |
| 433 | |
| 434 | if (ppkt && conn->compare_seq && !after(ppkt->seq_end, conn->compare_seq)) { |
| 435 | trace_colo_compare_main("pri: this packet has compared"); |
| 436 | colo_release_primary_pkt(s, ppkt); |
| 437 | ppkt = NULL; |
| 438 | } |
| 439 | |
| 440 | if (spkt->tcp_seq == spkt->seq_end) { |
| 441 | packet_destroy(spkt, NULL); |
| 442 | if (!ppkt) { |
| 443 | goto pri; |
| 444 | } else { |
| 445 | goto sec; |
| 446 | } |
| 447 | } else { |
| 448 | if (conn->compare_seq && !after(spkt->seq_end, conn->compare_seq)) { |
| 449 | trace_colo_compare_main("sec: this packet has compared"); |
| 450 | packet_destroy(spkt, NULL); |
| 451 | if (!ppkt) { |
| 452 | goto pri; |
| 453 | } else { |
| 454 | goto sec; |
| 455 | } |
| 456 | } |
| 457 | if (!ppkt) { |
| 458 | g_queue_push_tail(&conn->secondary_list, spkt); |
| 459 | goto pri; |
| 460 | } |
| 461 | } |
| 462 | |
| 463 | if (colo_mark_tcp_pkt(ppkt, spkt, &mark, min_ack)) { |
| 464 | trace_colo_compare_tcp_info("pri", |
| 465 | ppkt->tcp_seq, ppkt->tcp_ack, |
| 466 | ppkt->header_size, ppkt->payload_size, |
| 467 | ppkt->offset, ppkt->flags); |
| 468 | |
| 469 | trace_colo_compare_tcp_info("sec", |
| 470 | spkt->tcp_seq, spkt->tcp_ack, |
| 471 | spkt->header_size, spkt->payload_size, |
| 472 | spkt->offset, spkt->flags); |
| 473 | |
| 474 | if (mark == COLO_COMPARE_FREE_PRIMARY) { |
| 475 | conn->compare_seq = ppkt->seq_end; |
| 476 | colo_release_primary_pkt(s, ppkt); |
| 477 | g_queue_push_tail(&conn->secondary_list, spkt); |
| 478 | goto pri; |
| 479 | } else if (mark == COLO_COMPARE_FREE_SECONDARY) { |
| 480 | conn->compare_seq = spkt->seq_end; |
| 481 | packet_destroy(spkt, NULL); |
| 482 | goto sec; |
| 483 | } else if (mark == (COLO_COMPARE_FREE_PRIMARY | COLO_COMPARE_FREE_SECONDARY)) { |
| 484 | conn->compare_seq = ppkt->seq_end; |
| 485 | colo_release_primary_pkt(s, ppkt); |
| 486 | packet_destroy(spkt, NULL); |
| 487 | goto pri; |
| 488 | } |
| 489 | } else { |
| 490 | g_queue_push_tail(&conn->primary_list, ppkt); |
| 491 | g_queue_push_tail(&conn->secondary_list, spkt); |
| 492 | |
| 493 | #ifdef DEBUG_COLO_PACKETS |
| 494 | qemu_hexdump(stderr, "colo-compare ppkt", ppkt->data, ppkt->size); |
| 495 | qemu_hexdump(stderr, "colo-compare spkt", spkt->data, spkt->size); |
| 496 | #endif |
| 497 | |
| 498 | colo_compare_inconsistency_notify(s); |
| 499 | } |
| 500 | } |
| 501 | |
| 502 | |
| 503 | /* |
| 504 | * Called from the compare thread on the primary |
| 505 | * for compare udp packet |
| 506 | */ |
| 507 | static int colo_packet_compare_udp(Packet *spkt, Packet *ppkt) |
| 508 | { |
| 509 | uint16_t network_header_length = ppkt->ip->ip_hl << 2; |
| 510 | uint16_t offset = network_header_length + ETH_HLEN + ppkt->vnet_hdr_len; |
| 511 | |
| 512 | trace_colo_compare_main("compare udp"); |
| 513 | |
| 514 | /* |
| 515 | * Because of ppkt and spkt are both in the same connection, |
| 516 | * The ppkt's src ip, dst ip, src port, dst port, ip_proto all are |
| 517 | * same with spkt. In addition, IP header's Identification is a random |
| 518 | * field, we can handle it in IP fragmentation function later. |
| 519 | * COLO just concern the response net packet payload from primary guest |
| 520 | * and secondary guest are same or not, So we ignored all IP header include |
| 521 | * other field like TOS,TTL,IP Checksum. we only need to compare |
| 522 | * the ip payload here. |
| 523 | */ |
| 524 | if (ppkt->size != spkt->size) { |
| 525 | trace_colo_compare_main("UDP: payload size of packets are different"); |
| 526 | return -1; |
| 527 | } |
| 528 | if (colo_compare_packet_payload(ppkt, spkt, offset, offset, |
| 529 | ppkt->size - offset)) { |
| 530 | trace_colo_compare_udp_miscompare("primary pkt size", ppkt->size); |
| 531 | trace_colo_compare_udp_miscompare("Secondary pkt size", spkt->size); |
| 532 | #ifdef DEBUG_COLO_PACKETS |
| 533 | qemu_hexdump(stderr, "colo-compare pri pkt", ppkt->data, ppkt->size); |
| 534 | qemu_hexdump(stderr, "colo-compare sec pkt", spkt->data, spkt->size); |
| 535 | #endif |
| 536 | return -1; |
| 537 | } else { |
| 538 | return 0; |
| 539 | } |
| 540 | } |
| 541 | |
| 542 | /* |
| 543 | * Called from the compare thread on the primary |
| 544 | * for compare icmp packet |
| 545 | */ |
| 546 | static int colo_packet_compare_icmp(Packet *spkt, Packet *ppkt) |
| 547 | { |
| 548 | uint16_t network_header_length = ppkt->ip->ip_hl << 2; |
| 549 | uint16_t offset = network_header_length + ETH_HLEN + ppkt->vnet_hdr_len; |
| 550 | |
| 551 | trace_colo_compare_main("compare icmp"); |
| 552 | |
| 553 | /* |
| 554 | * Because of ppkt and spkt are both in the same connection, |
| 555 | * The ppkt's src ip, dst ip, src port, dst port, ip_proto all are |
| 556 | * same with spkt. In addition, IP header's Identification is a random |
| 557 | * field, we can handle it in IP fragmentation function later. |
| 558 | * COLO just concern the response net packet payload from primary guest |
| 559 | * and secondary guest are same or not, So we ignored all IP header include |
| 560 | * other field like TOS,TTL,IP Checksum. we only need to compare |
| 561 | * the ip payload here. |
| 562 | */ |
| 563 | if (ppkt->size != spkt->size) { |
| 564 | trace_colo_compare_main("ICMP: payload size of packets are different"); |
| 565 | return -1; |
| 566 | } |
| 567 | if (colo_compare_packet_payload(ppkt, spkt, offset, offset, |
| 568 | ppkt->size - offset)) { |
| 569 | trace_colo_compare_icmp_miscompare("primary pkt size", |
| 570 | ppkt->size); |
| 571 | trace_colo_compare_icmp_miscompare("Secondary pkt size", |
| 572 | spkt->size); |
| 573 | #ifdef DEBUG_COLO_PACKETS |
| 574 | qemu_hexdump(stderr, "colo-compare pri pkt", ppkt->data, ppkt->size); |
| 575 | qemu_hexdump(stderr, "colo-compare sec pkt", spkt->data, spkt->size); |
| 576 | #endif |
| 577 | return -1; |
| 578 | } else { |
| 579 | return 0; |
| 580 | } |
| 581 | } |
| 582 | |
| 583 | /* |
| 584 | * Called from the compare thread on the primary |
| 585 | * for compare other packet |
| 586 | */ |
| 587 | static int colo_packet_compare_other(Packet *spkt, Packet *ppkt) |
| 588 | { |
| 589 | uint16_t offset = ppkt->vnet_hdr_len; |
| 590 | |
| 591 | trace_colo_compare_main("compare other"); |
| 592 | if (ppkt->size != spkt->size) { |
| 593 | trace_colo_compare_main("Other: payload size of packets are different"); |
| 594 | return -1; |
| 595 | } |
| 596 | return colo_compare_packet_payload(ppkt, spkt, offset, offset, |
| 597 | ppkt->size - offset); |
| 598 | } |
| 599 | |
| 600 | static int colo_old_packet_check_one(Packet *pkt, int64_t *check_time) |
| 601 | { |
| 602 | int64_t now = qemu_clock_get_ms(QEMU_CLOCK_HOST); |
| 603 | |
| 604 | if ((now - pkt->creation_ms) > (*check_time)) { |
| 605 | trace_colo_old_packet_check_found(pkt->creation_ms); |
| 606 | return 0; |
| 607 | } else { |
| 608 | return 1; |
| 609 | } |
| 610 | } |
| 611 | |
| 612 | void colo_compare_register_notifier(Notifier *notify) |
| 613 | { |
| 614 | notifier_list_add(&colo_compare_notifiers, notify); |
| 615 | } |
| 616 | |
| 617 | void colo_compare_unregister_notifier(Notifier *notify) |
| 618 | { |
| 619 | notifier_remove(notify); |
| 620 | } |
| 621 | |
| 622 | static int colo_old_packet_check_one_conn(Connection *conn, |
| 623 | CompareState *s) |
| 624 | { |
| 625 | if (!g_queue_is_empty(&conn->primary_list)) { |
| 626 | if (g_queue_find_custom(&conn->primary_list, |
| 627 | &s->compare_timeout, |
| 628 | (GCompareFunc)colo_old_packet_check_one)) |
| 629 | goto out; |
| 630 | } |
| 631 | |
| 632 | if (!g_queue_is_empty(&conn->secondary_list)) { |
| 633 | if (g_queue_find_custom(&conn->secondary_list, |
| 634 | &s->compare_timeout, |
| 635 | (GCompareFunc)colo_old_packet_check_one)) |
| 636 | goto out; |
| 637 | } |
| 638 | |
| 639 | return 1; |
| 640 | |
| 641 | out: |
| 642 | /* Do checkpoint will flush old packet */ |
| 643 | colo_compare_inconsistency_notify(s); |
| 644 | return 0; |
| 645 | } |
| 646 | |
| 647 | /* |
| 648 | * Look for old packets that the secondary hasn't matched, |
| 649 | * if we have some then we have to checkpoint to wake |
| 650 | * the secondary up. |
| 651 | */ |
| 652 | static void colo_old_packet_check(void *opaque) |
| 653 | { |
| 654 | CompareState *s = opaque; |
| 655 | |
| 656 | /* |
| 657 | * If we find one old packet, stop finding job and notify |
| 658 | * COLO frame do checkpoint. |
| 659 | */ |
| 660 | g_queue_find_custom(&s->conn_list, s, |
| 661 | (GCompareFunc)colo_old_packet_check_one_conn); |
| 662 | } |
| 663 | |
| 664 | static void colo_compare_packet(CompareState *s, Connection *conn, |
| 665 | int (*HandlePacket)(Packet *spkt, |
| 666 | Packet *ppkt)) |
| 667 | { |
| 668 | Packet *pkt = NULL; |
| 669 | GList *result = NULL; |
| 670 | |
| 671 | while (!g_queue_is_empty(&conn->primary_list) && |
| 672 | !g_queue_is_empty(&conn->secondary_list)) { |
| 673 | pkt = g_queue_pop_tail(&conn->primary_list); |
| 674 | result = g_queue_find_custom(&conn->secondary_list, |
| 675 | pkt, (GCompareFunc)HandlePacket); |
| 676 | |
| 677 | if (result) { |
| 678 | colo_release_primary_pkt(s, pkt); |
| 679 | packet_destroy(result->data, NULL); |
| 680 | g_queue_delete_link(&conn->secondary_list, result); |
| 681 | } else { |
| 682 | /* |
| 683 | * If one packet arrive late, the secondary_list or |
| 684 | * primary_list will be empty, so we can't compare it |
| 685 | * until next comparison. If the packets in the list are |
| 686 | * timeout, it will trigger a checkpoint request. |
| 687 | */ |
| 688 | trace_colo_compare_main("packet different"); |
| 689 | g_queue_push_tail(&conn->primary_list, pkt); |
| 690 | |
| 691 | colo_compare_inconsistency_notify(s); |
| 692 | break; |
| 693 | } |
| 694 | } |
| 695 | } |
| 696 | |
| 697 | /* |
| 698 | * Called from the compare thread on the primary |
| 699 | * for compare packet with secondary list of the |
| 700 | * specified connection when a new packet was |
| 701 | * queued to it. |
| 702 | */ |
| 703 | static void colo_compare_connection(void *opaque, void *user_data) |
| 704 | { |
| 705 | CompareState *s = user_data; |
| 706 | Connection *conn = opaque; |
| 707 | |
| 708 | switch (conn->ip_proto) { |
| 709 | case IPPROTO_TCP: |
| 710 | colo_compare_tcp(s, conn); |
| 711 | break; |
| 712 | case IPPROTO_UDP: |
| 713 | colo_compare_packet(s, conn, colo_packet_compare_udp); |
| 714 | break; |
| 715 | case IPPROTO_ICMP: |
| 716 | colo_compare_packet(s, conn, colo_packet_compare_icmp); |
| 717 | break; |
| 718 | default: |
| 719 | colo_compare_packet(s, conn, colo_packet_compare_other); |
| 720 | break; |
| 721 | } |
| 722 | } |
| 723 | |
| 724 | static void coroutine_fn _compare_chr_send(void *opaque) |
| 725 | { |
| 726 | SendCo *sendco = opaque; |
| 727 | CompareState *s = sendco->s; |
| 728 | int ret = 0; |
| 729 | |
| 730 | while (!g_queue_is_empty(&sendco->send_list)) { |
| 731 | SendEntry *entry = g_queue_pop_tail(&sendco->send_list); |
| 732 | uint32_t len = htonl(entry->size); |
| 733 | |
| 734 | ret = qemu_chr_fe_write_all(sendco->chr, (uint8_t *)&len, sizeof(len)); |
| 735 | |
| 736 | if (ret != sizeof(len)) { |
| 737 | g_free(entry->buf); |
| 738 | g_slice_free(SendEntry, entry); |
| 739 | goto err; |
| 740 | } |
| 741 | |
| 742 | if (!sendco->notify_remote_frame && s->vnet_hdr) { |
| 743 | /* |
| 744 | * We send vnet header len make other module(like filter-redirector) |
| 745 | * know how to parse net packet correctly. |
| 746 | */ |
| 747 | len = htonl(entry->vnet_hdr_len); |
| 748 | |
| 749 | ret = qemu_chr_fe_write_all(sendco->chr, |
| 750 | (uint8_t *)&len, |
| 751 | sizeof(len)); |
| 752 | |
| 753 | if (ret != sizeof(len)) { |
| 754 | g_free(entry->buf); |
| 755 | g_slice_free(SendEntry, entry); |
| 756 | goto err; |
| 757 | } |
| 758 | } |
| 759 | |
| 760 | ret = qemu_chr_fe_write_all(sendco->chr, |
| 761 | (uint8_t *)entry->buf, |
| 762 | entry->size); |
| 763 | |
| 764 | if (ret != entry->size) { |
| 765 | g_free(entry->buf); |
| 766 | g_slice_free(SendEntry, entry); |
| 767 | goto err; |
| 768 | } |
| 769 | |
| 770 | g_free(entry->buf); |
| 771 | g_slice_free(SendEntry, entry); |
| 772 | } |
| 773 | |
| 774 | sendco->ret = 0; |
| 775 | goto out; |
| 776 | |
| 777 | err: |
| 778 | while (!g_queue_is_empty(&sendco->send_list)) { |
| 779 | SendEntry *entry = g_queue_pop_tail(&sendco->send_list); |
| 780 | g_free(entry->buf); |
| 781 | g_slice_free(SendEntry, entry); |
| 782 | } |
| 783 | sendco->ret = ret < 0 ? ret : -EIO; |
| 784 | out: |
| 785 | sendco->co = NULL; |
| 786 | sendco->done = true; |
| 787 | aio_wait_kick(); |
| 788 | } |
| 789 | |
| 790 | static int compare_chr_send(CompareState *s, |
| 791 | uint8_t *buf, |
| 792 | uint32_t size, |
| 793 | uint32_t vnet_hdr_len, |
| 794 | bool notify_remote_frame, |
| 795 | bool zero_copy) |
| 796 | { |
| 797 | SendCo *sendco; |
| 798 | SendEntry *entry; |
| 799 | |
| 800 | if (notify_remote_frame) { |
| 801 | sendco = &s->notify_sendco; |
| 802 | } else { |
| 803 | sendco = &s->out_sendco; |
| 804 | } |
| 805 | |
| 806 | if (!size) { |
| 807 | return -1; |
| 808 | } |
| 809 | |
| 810 | entry = g_slice_new(SendEntry); |
| 811 | entry->size = size; |
| 812 | entry->vnet_hdr_len = vnet_hdr_len; |
| 813 | if (zero_copy) { |
| 814 | entry->buf = buf; |
| 815 | } else { |
| 816 | entry->buf = g_malloc(size); |
| 817 | memcpy(entry->buf, buf, size); |
| 818 | } |
| 819 | g_queue_push_tail(&sendco->send_list, entry); |
| 820 | |
| 821 | if (sendco->done) { |
| 822 | sendco->co = qemu_coroutine_create(_compare_chr_send, sendco); |
| 823 | sendco->done = false; |
| 824 | qemu_coroutine_enter(sendco->co); |
| 825 | if (sendco->done) { |
| 826 | /* report early errors */ |
| 827 | return sendco->ret; |
| 828 | } |
| 829 | } |
| 830 | |
| 831 | /* assume success */ |
| 832 | return 0; |
| 833 | } |
| 834 | |
| 835 | static int compare_chr_can_read(void *opaque) |
| 836 | { |
| 837 | return COMPARE_READ_LEN_MAX; |
| 838 | } |
| 839 | |
| 840 | /* |
| 841 | * Called from the main thread on the primary for packets |
| 842 | * arriving over the socket from the primary. |
| 843 | */ |
| 844 | static void compare_pri_chr_in(void *opaque, const uint8_t *buf, int size) |
| 845 | { |
| 846 | CompareState *s = COLO_COMPARE(opaque); |
| 847 | int ret; |
| 848 | |
| 849 | ret = net_fill_rstate(&s->pri_rs, buf, size); |
| 850 | if (ret == -1) { |
| 851 | qemu_chr_fe_set_handlers(&s->chr_pri_in, NULL, NULL, NULL, NULL, |
| 852 | NULL, NULL, true); |
| 853 | error_report("colo-compare primary_in error"); |
| 854 | } |
| 855 | } |
| 856 | |
| 857 | /* |
| 858 | * Called from the main thread on the primary for packets |
| 859 | * arriving over the socket from the secondary. |
| 860 | */ |
| 861 | static void compare_sec_chr_in(void *opaque, const uint8_t *buf, int size) |
| 862 | { |
| 863 | CompareState *s = COLO_COMPARE(opaque); |
| 864 | int ret; |
| 865 | |
| 866 | ret = net_fill_rstate(&s->sec_rs, buf, size); |
| 867 | if (ret == -1) { |
| 868 | qemu_chr_fe_set_handlers(&s->chr_sec_in, NULL, NULL, NULL, NULL, |
| 869 | NULL, NULL, true); |
| 870 | error_report("colo-compare secondary_in error"); |
| 871 | } |
| 872 | } |
| 873 | |
| 874 | static void compare_notify_chr(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->notify_rs, buf, size); |
| 880 | if (ret == -1) { |
| 881 | qemu_chr_fe_set_handlers(&s->chr_notify_dev, NULL, NULL, NULL, NULL, |
| 882 | NULL, NULL, true); |
| 883 | error_report("colo-compare notify_dev error"); |
| 884 | } |
| 885 | } |
| 886 | |
| 887 | /* |
| 888 | * Check old packet regularly so it can watch for any packets |
| 889 | * that the secondary hasn't produced equivalents of. |
| 890 | */ |
| 891 | static void check_old_packet_regular(void *opaque) |
| 892 | { |
| 893 | CompareState *s = opaque; |
| 894 | |
| 895 | /* if have old packet we will notify checkpoint */ |
| 896 | colo_old_packet_check(s); |
| 897 | timer_mod(s->packet_check_timer, qemu_clock_get_ms(QEMU_CLOCK_HOST) + |
| 898 | s->expired_scan_cycle); |
| 899 | } |
| 900 | |
| 901 | /* Public API, Used for COLO frame to notify compare event */ |
| 902 | void colo_notify_compares_event(void *opaque, int event, Error **errp) |
| 903 | { |
| 904 | CompareState *s; |
| 905 | QEMU_LOCK_GUARD(&colo_compare_mutex); |
| 906 | |
| 907 | if (!colo_compare_active) { |
| 908 | return; |
| 909 | } |
| 910 | |
| 911 | QEMU_LOCK_GUARD(&event_mtx); |
| 912 | QTAILQ_FOREACH(s, &net_compares, next) { |
| 913 | s->event = event; |
| 914 | qemu_bh_schedule(s->event_bh); |
| 915 | event_unhandled_count++; |
| 916 | } |
| 917 | /* Wait all compare threads to finish handling this event */ |
| 918 | while (event_unhandled_count > 0) { |
| 919 | qemu_cond_wait(&event_complete_cond, &event_mtx); |
| 920 | } |
| 921 | } |
| 922 | |
| 923 | static void colo_compare_timer_init(CompareState *s) |
| 924 | { |
| 925 | AioContext *ctx = iothread_get_aio_context(s->iothread); |
| 926 | |
| 927 | s->packet_check_timer = aio_timer_new(ctx, QEMU_CLOCK_HOST, |
| 928 | SCALE_MS, check_old_packet_regular, |
| 929 | s); |
| 930 | timer_mod(s->packet_check_timer, qemu_clock_get_ms(QEMU_CLOCK_HOST) + |
| 931 | s->expired_scan_cycle); |
| 932 | } |
| 933 | |
| 934 | static void colo_compare_timer_del(CompareState *s) |
| 935 | { |
| 936 | if (s->packet_check_timer) { |
| 937 | timer_free(s->packet_check_timer); |
| 938 | s->packet_check_timer = NULL; |
| 939 | } |
| 940 | } |
| 941 | |
| 942 | static void colo_flush_packets(void *opaque, void *user_data); |
| 943 | |
| 944 | static void colo_compare_handle_event(void *opaque) |
| 945 | { |
| 946 | CompareState *s = opaque; |
| 947 | |
| 948 | switch (s->event) { |
| 949 | case COLO_EVENT_CHECKPOINT: |
| 950 | g_queue_foreach(&s->conn_list, colo_flush_packets, s); |
| 951 | break; |
| 952 | case COLO_EVENT_FAILOVER: |
| 953 | break; |
| 954 | default: |
| 955 | break; |
| 956 | } |
| 957 | |
| 958 | qemu_mutex_lock(&event_mtx); |
| 959 | assert(event_unhandled_count > 0); |
| 960 | event_unhandled_count--; |
| 961 | qemu_cond_broadcast(&event_complete_cond); |
| 962 | qemu_mutex_unlock(&event_mtx); |
| 963 | } |
| 964 | |
| 965 | static void colo_compare_iothread(CompareState *s) |
| 966 | { |
| 967 | AioContext *ctx = iothread_get_aio_context(s->iothread); |
| 968 | object_ref(OBJECT(s->iothread)); |
| 969 | s->worker_context = iothread_get_g_main_context(s->iothread); |
| 970 | |
| 971 | qemu_chr_fe_set_handlers(&s->chr_pri_in, compare_chr_can_read, |
| 972 | compare_pri_chr_in, NULL, NULL, |
| 973 | s, s->worker_context, true); |
| 974 | qemu_chr_fe_set_handlers(&s->chr_sec_in, compare_chr_can_read, |
| 975 | compare_sec_chr_in, NULL, NULL, |
| 976 | s, s->worker_context, true); |
| 977 | if (s->notify_dev) { |
| 978 | qemu_chr_fe_set_handlers(&s->chr_notify_dev, compare_chr_can_read, |
| 979 | compare_notify_chr, NULL, NULL, |
| 980 | s, s->worker_context, true); |
| 981 | } |
| 982 | |
| 983 | colo_compare_timer_init(s); |
| 984 | s->event_bh = aio_bh_new(ctx, colo_compare_handle_event, s); |
| 985 | } |
| 986 | |
| 987 | static char *compare_get_pri_indev(Object *obj, Error **errp) |
| 988 | { |
| 989 | CompareState *s = COLO_COMPARE(obj); |
| 990 | |
| 991 | return g_strdup(s->pri_indev); |
| 992 | } |
| 993 | |
| 994 | static void compare_set_pri_indev(Object *obj, const char *value, Error **errp) |
| 995 | { |
| 996 | CompareState *s = COLO_COMPARE(obj); |
| 997 | |
| 998 | g_free(s->pri_indev); |
| 999 | s->pri_indev = g_strdup(value); |
| 1000 | } |
| 1001 | |
| 1002 | static char *compare_get_sec_indev(Object *obj, Error **errp) |
| 1003 | { |
| 1004 | CompareState *s = COLO_COMPARE(obj); |
| 1005 | |
| 1006 | return g_strdup(s->sec_indev); |
| 1007 | } |
| 1008 | |
| 1009 | static void compare_set_sec_indev(Object *obj, const char *value, Error **errp) |
| 1010 | { |
| 1011 | CompareState *s = COLO_COMPARE(obj); |
| 1012 | |
| 1013 | g_free(s->sec_indev); |
| 1014 | s->sec_indev = g_strdup(value); |
| 1015 | } |
| 1016 | |
| 1017 | static char *compare_get_outdev(Object *obj, Error **errp) |
| 1018 | { |
| 1019 | CompareState *s = COLO_COMPARE(obj); |
| 1020 | |
| 1021 | return g_strdup(s->outdev); |
| 1022 | } |
| 1023 | |
| 1024 | static void compare_set_outdev(Object *obj, const char *value, Error **errp) |
| 1025 | { |
| 1026 | CompareState *s = COLO_COMPARE(obj); |
| 1027 | |
| 1028 | g_free(s->outdev); |
| 1029 | s->outdev = g_strdup(value); |
| 1030 | } |
| 1031 | |
| 1032 | static bool compare_get_vnet_hdr(Object *obj, Error **errp) |
| 1033 | { |
| 1034 | CompareState *s = COLO_COMPARE(obj); |
| 1035 | |
| 1036 | return s->vnet_hdr; |
| 1037 | } |
| 1038 | |
| 1039 | static void compare_set_vnet_hdr(Object *obj, |
| 1040 | bool value, |
| 1041 | Error **errp) |
| 1042 | { |
| 1043 | CompareState *s = COLO_COMPARE(obj); |
| 1044 | |
| 1045 | s->vnet_hdr = value; |
| 1046 | } |
| 1047 | |
| 1048 | static char *compare_get_notify_dev(Object *obj, Error **errp) |
| 1049 | { |
| 1050 | CompareState *s = COLO_COMPARE(obj); |
| 1051 | |
| 1052 | return g_strdup(s->notify_dev); |
| 1053 | } |
| 1054 | |
| 1055 | static void compare_set_notify_dev(Object *obj, const char *value, Error **errp) |
| 1056 | { |
| 1057 | CompareState *s = COLO_COMPARE(obj); |
| 1058 | |
| 1059 | g_free(s->notify_dev); |
| 1060 | s->notify_dev = g_strdup(value); |
| 1061 | } |
| 1062 | |
| 1063 | static void compare_get_timeout(Object *obj, Visitor *v, |
| 1064 | const char *name, void *opaque, |
| 1065 | Error **errp) |
| 1066 | { |
| 1067 | CompareState *s = COLO_COMPARE(obj); |
| 1068 | uint64_t value = s->compare_timeout; |
| 1069 | |
| 1070 | visit_type_uint64(v, name, &value, errp); |
| 1071 | } |
| 1072 | |
| 1073 | static void compare_set_timeout(Object *obj, Visitor *v, |
| 1074 | const char *name, void *opaque, |
| 1075 | Error **errp) |
| 1076 | { |
| 1077 | CompareState *s = COLO_COMPARE(obj); |
| 1078 | uint32_t value; |
| 1079 | |
| 1080 | if (!visit_type_uint32(v, name, &value, errp)) { |
| 1081 | return; |
| 1082 | } |
| 1083 | if (!value) { |
| 1084 | error_setg(errp, "Property '%s.%s' requires a positive value", |
| 1085 | object_get_typename(obj), name); |
| 1086 | return; |
| 1087 | } |
| 1088 | s->compare_timeout = value; |
| 1089 | } |
| 1090 | |
| 1091 | static void compare_get_expired_scan_cycle(Object *obj, Visitor *v, |
| 1092 | const char *name, void *opaque, |
| 1093 | Error **errp) |
| 1094 | { |
| 1095 | CompareState *s = COLO_COMPARE(obj); |
| 1096 | uint32_t value = s->expired_scan_cycle; |
| 1097 | |
| 1098 | visit_type_uint32(v, name, &value, errp); |
| 1099 | } |
| 1100 | |
| 1101 | static void compare_set_expired_scan_cycle(Object *obj, Visitor *v, |
| 1102 | const char *name, void *opaque, |
| 1103 | Error **errp) |
| 1104 | { |
| 1105 | CompareState *s = COLO_COMPARE(obj); |
| 1106 | uint32_t value; |
| 1107 | |
| 1108 | if (!visit_type_uint32(v, name, &value, errp)) { |
| 1109 | return; |
| 1110 | } |
| 1111 | if (!value) { |
| 1112 | error_setg(errp, "Property '%s.%s' requires a positive value", |
| 1113 | object_get_typename(obj), name); |
| 1114 | return; |
| 1115 | } |
| 1116 | s->expired_scan_cycle = value; |
| 1117 | } |
| 1118 | |
| 1119 | static void get_max_queue_size(Object *obj, Visitor *v, |
| 1120 | const char *name, void *opaque, |
| 1121 | Error **errp) |
| 1122 | { |
| 1123 | uint32_t value = max_queue_size; |
| 1124 | |
| 1125 | visit_type_uint32(v, name, &value, errp); |
| 1126 | } |
| 1127 | |
| 1128 | static void set_max_queue_size(Object *obj, Visitor *v, |
| 1129 | const char *name, void *opaque, |
| 1130 | Error **errp) |
| 1131 | { |
| 1132 | uint64_t value; |
| 1133 | |
| 1134 | if (!visit_type_uint64(v, name, &value, errp)) { |
| 1135 | return; |
| 1136 | } |
| 1137 | if (!value) { |
| 1138 | error_setg(errp, "Property '%s.%s' requires a positive value", |
| 1139 | object_get_typename(obj), name); |
| 1140 | return; |
| 1141 | } |
| 1142 | max_queue_size = value; |
| 1143 | } |
| 1144 | |
| 1145 | static void compare_pri_rs_finalize(SocketReadState *pri_rs) |
| 1146 | { |
| 1147 | CompareState *s = container_of(pri_rs, CompareState, pri_rs); |
| 1148 | Connection *conn = NULL; |
| 1149 | |
| 1150 | if (packet_enqueue(s, PRIMARY_IN, &conn)) { |
| 1151 | trace_colo_compare_main("primary: unsupported packet in"); |
| 1152 | compare_chr_send(s, |
| 1153 | pri_rs->buf, |
| 1154 | pri_rs->packet_len, |
| 1155 | pri_rs->vnet_hdr_len, |
| 1156 | false, |
| 1157 | false); |
| 1158 | } else { |
| 1159 | /* compare packet in the specified connection */ |
| 1160 | colo_compare_connection(conn, s); |
| 1161 | } |
| 1162 | } |
| 1163 | |
| 1164 | static void compare_sec_rs_finalize(SocketReadState *sec_rs) |
| 1165 | { |
| 1166 | CompareState *s = container_of(sec_rs, CompareState, sec_rs); |
| 1167 | Connection *conn = NULL; |
| 1168 | |
| 1169 | if (packet_enqueue(s, SECONDARY_IN, &conn)) { |
| 1170 | trace_colo_compare_main("secondary: unsupported packet in"); |
| 1171 | } else { |
| 1172 | /* compare packet in the specified connection */ |
| 1173 | colo_compare_connection(conn, s); |
| 1174 | } |
| 1175 | } |
| 1176 | |
| 1177 | static void compare_notify_rs_finalize(SocketReadState *notify_rs) |
| 1178 | { |
| 1179 | CompareState *s = container_of(notify_rs, CompareState, notify_rs); |
| 1180 | |
| 1181 | const char msg[] = "COLO_COMPARE_GET_XEN_INIT"; |
| 1182 | int ret; |
| 1183 | |
| 1184 | if (packet_matches_str("COLO_USERSPACE_PROXY_INIT", |
| 1185 | notify_rs->buf, |
| 1186 | notify_rs->packet_len)) { |
| 1187 | ret = compare_chr_send(s, (uint8_t *)msg, strlen(msg), 0, true, false); |
| 1188 | if (ret < 0) { |
| 1189 | error_report("Notify Xen COLO-frame INIT failed"); |
| 1190 | } |
| 1191 | } else if (packet_matches_str("COLO_CHECKPOINT", |
| 1192 | notify_rs->buf, |
| 1193 | notify_rs->packet_len)) { |
| 1194 | /* colo-compare do checkpoint, flush pri packet and remove sec packet */ |
| 1195 | g_queue_foreach(&s->conn_list, colo_flush_packets, s); |
| 1196 | } else { |
| 1197 | error_report("COLO compare got unsupported instruction"); |
| 1198 | } |
| 1199 | } |
| 1200 | |
| 1201 | /* |
| 1202 | * Return 0 is success. |
| 1203 | * Return 1 is failed. |
| 1204 | */ |
| 1205 | static int find_and_check_chardev(Chardev **chr, |
| 1206 | char *chr_name, |
| 1207 | Error **errp) |
| 1208 | { |
| 1209 | *chr = qemu_chr_find(chr_name); |
| 1210 | if (*chr == NULL) { |
| 1211 | error_setg(errp, "Device '%s' not found", |
| 1212 | chr_name); |
| 1213 | return 1; |
| 1214 | } |
| 1215 | |
| 1216 | if (!qemu_chr_has_feature(*chr, QEMU_CHAR_FEATURE_RECONNECTABLE)) { |
| 1217 | error_setg(errp, "chardev \"%s\" is not reconnectable", |
| 1218 | chr_name); |
| 1219 | return 1; |
| 1220 | } |
| 1221 | |
| 1222 | if (!qemu_chr_has_feature(*chr, QEMU_CHAR_FEATURE_GCONTEXT)) { |
| 1223 | error_setg(errp, "chardev \"%s\" cannot switch context", |
| 1224 | chr_name); |
| 1225 | return 1; |
| 1226 | } |
| 1227 | |
| 1228 | return 0; |
| 1229 | } |
| 1230 | |
| 1231 | /* |
| 1232 | * Called from the main thread on the primary |
| 1233 | * to setup colo-compare. |
| 1234 | */ |
| 1235 | static void colo_compare_complete(UserCreatable *uc, Error **errp) |
| 1236 | { |
| 1237 | CompareState *s = COLO_COMPARE(uc); |
| 1238 | Chardev *chr; |
| 1239 | |
| 1240 | if (!s->pri_indev || !s->sec_indev || !s->outdev || !s->iothread) { |
| 1241 | error_setg(errp, "colo compare needs 'primary_in' ," |
| 1242 | "'secondary_in','outdev','iothread' property set"); |
| 1243 | return; |
| 1244 | } else if (!strcmp(s->pri_indev, s->outdev) || |
| 1245 | !strcmp(s->sec_indev, s->outdev) || |
| 1246 | !strcmp(s->pri_indev, s->sec_indev)) { |
| 1247 | error_setg(errp, "'indev' and 'outdev' could not be same " |
| 1248 | "for compare module"); |
| 1249 | return; |
| 1250 | } |
| 1251 | |
| 1252 | if (!s->compare_timeout) { |
| 1253 | /* Set default value to 3000 MS */ |
| 1254 | s->compare_timeout = DEFAULT_TIME_OUT_MS; |
| 1255 | } |
| 1256 | |
| 1257 | if (!s->expired_scan_cycle) { |
| 1258 | /* Set default value to 1000 MS */ |
| 1259 | s->expired_scan_cycle = REGULAR_PACKET_CHECK_MS; |
| 1260 | } |
| 1261 | |
| 1262 | if (!max_queue_size) { |
| 1263 | /* Set default queue size to 1024 */ |
| 1264 | max_queue_size = MAX_QUEUE_SIZE; |
| 1265 | } |
| 1266 | |
| 1267 | if (find_and_check_chardev(&chr, s->pri_indev, errp) || |
| 1268 | !qemu_chr_fe_init(&s->chr_pri_in, chr, errp)) { |
| 1269 | return; |
| 1270 | } |
| 1271 | |
| 1272 | if (find_and_check_chardev(&chr, s->sec_indev, errp) || |
| 1273 | !qemu_chr_fe_init(&s->chr_sec_in, chr, errp)) { |
| 1274 | return; |
| 1275 | } |
| 1276 | |
| 1277 | if (find_and_check_chardev(&chr, s->outdev, errp) || |
| 1278 | !qemu_chr_fe_init(&s->chr_out, chr, errp)) { |
| 1279 | return; |
| 1280 | } |
| 1281 | |
| 1282 | net_socket_rs_init(&s->pri_rs, compare_pri_rs_finalize, s->vnet_hdr); |
| 1283 | net_socket_rs_init(&s->sec_rs, compare_sec_rs_finalize, s->vnet_hdr); |
| 1284 | |
| 1285 | /* Try to enable remote notify chardev, currently just for Xen COLO */ |
| 1286 | if (s->notify_dev) { |
| 1287 | if (find_and_check_chardev(&chr, s->notify_dev, errp) || |
| 1288 | !qemu_chr_fe_init(&s->chr_notify_dev, chr, errp)) { |
| 1289 | return; |
| 1290 | } |
| 1291 | |
| 1292 | net_socket_rs_init(&s->notify_rs, compare_notify_rs_finalize, |
| 1293 | s->vnet_hdr); |
| 1294 | } |
| 1295 | |
| 1296 | s->out_sendco.s = s; |
| 1297 | s->out_sendco.chr = &s->chr_out; |
| 1298 | s->out_sendco.notify_remote_frame = false; |
| 1299 | s->out_sendco.done = true; |
| 1300 | g_queue_init(&s->out_sendco.send_list); |
| 1301 | |
| 1302 | if (s->notify_dev) { |
| 1303 | s->notify_sendco.s = s; |
| 1304 | s->notify_sendco.chr = &s->chr_notify_dev; |
| 1305 | s->notify_sendco.notify_remote_frame = true; |
| 1306 | s->notify_sendco.done = true; |
| 1307 | g_queue_init(&s->notify_sendco.send_list); |
| 1308 | } |
| 1309 | |
| 1310 | g_queue_init(&s->conn_list); |
| 1311 | |
| 1312 | s->connection_track_table = g_hash_table_new_full(connection_key_hash, |
| 1313 | connection_key_equal, |
| 1314 | g_free, |
| 1315 | NULL); |
| 1316 | |
| 1317 | colo_compare_iothread(s); |
| 1318 | |
| 1319 | qemu_mutex_lock(&colo_compare_mutex); |
| 1320 | if (!colo_compare_active) { |
| 1321 | qemu_mutex_init(&event_mtx); |
| 1322 | qemu_cond_init(&event_complete_cond); |
| 1323 | colo_compare_active = true; |
| 1324 | } |
| 1325 | QTAILQ_INSERT_TAIL(&net_compares, s, next); |
| 1326 | qemu_mutex_unlock(&colo_compare_mutex); |
| 1327 | } |
| 1328 | |
| 1329 | static void colo_flush_packets(void *opaque, void *user_data) |
| 1330 | { |
| 1331 | CompareState *s = user_data; |
| 1332 | Connection *conn = opaque; |
| 1333 | Packet *pkt = NULL; |
| 1334 | |
| 1335 | while (!g_queue_is_empty(&conn->primary_list)) { |
| 1336 | pkt = g_queue_pop_tail(&conn->primary_list); |
| 1337 | compare_chr_send(s, |
| 1338 | pkt->data, |
| 1339 | pkt->size, |
| 1340 | pkt->vnet_hdr_len, |
| 1341 | false, |
| 1342 | true); |
| 1343 | packet_destroy_partial(pkt, NULL); |
| 1344 | } |
| 1345 | while (!g_queue_is_empty(&conn->secondary_list)) { |
| 1346 | pkt = g_queue_pop_tail(&conn->secondary_list); |
| 1347 | packet_destroy(pkt, NULL); |
| 1348 | } |
| 1349 | } |
| 1350 | |
| 1351 | static void colo_compare_class_init(ObjectClass *oc, const void *data) |
| 1352 | { |
| 1353 | UserCreatableClass *ucc = USER_CREATABLE_CLASS(oc); |
| 1354 | |
| 1355 | ucc->complete = colo_compare_complete; |
| 1356 | } |
| 1357 | |
| 1358 | static void colo_compare_init(Object *obj) |
| 1359 | { |
| 1360 | CompareState *s = COLO_COMPARE(obj); |
| 1361 | |
| 1362 | object_property_add_str(obj, "primary_in", |
| 1363 | compare_get_pri_indev, compare_set_pri_indev); |
| 1364 | object_property_add_str(obj, "secondary_in", |
| 1365 | compare_get_sec_indev, compare_set_sec_indev); |
| 1366 | object_property_add_str(obj, "outdev", |
| 1367 | compare_get_outdev, compare_set_outdev); |
| 1368 | object_property_add_link(obj, "iothread", TYPE_IOTHREAD, |
| 1369 | (Object **)&s->iothread, |
| 1370 | object_property_allow_set_link, |
| 1371 | OBJ_PROP_LINK_STRONG); |
| 1372 | /* This parameter just for Xen COLO */ |
| 1373 | object_property_add_str(obj, "notify_dev", |
| 1374 | compare_get_notify_dev, compare_set_notify_dev); |
| 1375 | |
| 1376 | object_property_add(obj, "compare_timeout", "uint64", |
| 1377 | compare_get_timeout, |
| 1378 | compare_set_timeout, NULL, NULL); |
| 1379 | |
| 1380 | object_property_add(obj, "expired_scan_cycle", "uint32", |
| 1381 | compare_get_expired_scan_cycle, |
| 1382 | compare_set_expired_scan_cycle, NULL, NULL); |
| 1383 | |
| 1384 | object_property_add(obj, "max_queue_size", "uint32", |
| 1385 | get_max_queue_size, |
| 1386 | set_max_queue_size, NULL, NULL); |
| 1387 | |
| 1388 | s->vnet_hdr = false; |
| 1389 | object_property_add_bool(obj, "vnet_hdr_support", compare_get_vnet_hdr, |
| 1390 | compare_set_vnet_hdr); |
| 1391 | } |
| 1392 | |
| 1393 | void colo_compare_cleanup(void) |
| 1394 | { |
| 1395 | CompareState *tmp; |
| 1396 | CompareState *n = NULL; |
| 1397 | |
| 1398 | QTAILQ_FOREACH_SAFE(tmp, &net_compares, next, n) { |
| 1399 | object_unparent(OBJECT(tmp)); |
| 1400 | } |
| 1401 | } |
| 1402 | |
| 1403 | static void colo_compare_finalize(Object *obj) |
| 1404 | { |
| 1405 | CompareState *s = COLO_COMPARE(obj); |
| 1406 | CompareState *tmp; |
| 1407 | |
| 1408 | qemu_mutex_lock(&colo_compare_mutex); |
| 1409 | QTAILQ_FOREACH(tmp, &net_compares, next) { |
| 1410 | if (tmp == s) { |
| 1411 | QTAILQ_REMOVE(&net_compares, s, next); |
| 1412 | break; |
| 1413 | } |
| 1414 | } |
| 1415 | if (colo_compare_active && QTAILQ_EMPTY(&net_compares)) { |
| 1416 | colo_compare_active = false; |
| 1417 | qemu_mutex_destroy(&event_mtx); |
| 1418 | qemu_cond_destroy(&event_complete_cond); |
| 1419 | } |
| 1420 | qemu_mutex_unlock(&colo_compare_mutex); |
| 1421 | |
| 1422 | qemu_chr_fe_deinit(&s->chr_pri_in, false); |
| 1423 | qemu_chr_fe_deinit(&s->chr_sec_in, false); |
| 1424 | qemu_chr_fe_deinit(&s->chr_out, false); |
| 1425 | if (s->notify_dev) { |
| 1426 | qemu_chr_fe_deinit(&s->chr_notify_dev, false); |
| 1427 | } |
| 1428 | |
| 1429 | colo_compare_timer_del(s); |
| 1430 | g_clear_pointer(&s->event_bh, qemu_bh_delete); |
| 1431 | |
| 1432 | if (s->iothread) { |
| 1433 | AioContext *ctx = iothread_get_aio_context(s->iothread); |
| 1434 | |
| 1435 | AIO_WAIT_WHILE(ctx, !s->out_sendco.done); |
| 1436 | if (s->notify_dev) { |
| 1437 | AIO_WAIT_WHILE(ctx, !s->notify_sendco.done); |
| 1438 | } |
| 1439 | |
| 1440 | /* Release all unhandled packets after compare thread exited */ |
| 1441 | g_queue_foreach(&s->conn_list, colo_flush_packets, s); |
| 1442 | AIO_WAIT_WHILE(NULL, !s->out_sendco.done); |
| 1443 | |
| 1444 | object_unref(OBJECT(s->iothread)); |
| 1445 | } |
| 1446 | |
| 1447 | g_queue_clear(&s->conn_list); |
| 1448 | g_queue_clear(&s->out_sendco.send_list); |
| 1449 | if (s->notify_dev) { |
| 1450 | g_queue_clear(&s->notify_sendco.send_list); |
| 1451 | } |
| 1452 | g_clear_pointer(&s->connection_track_table, g_hash_table_destroy); |
| 1453 | |
| 1454 | g_free(s->pri_indev); |
| 1455 | g_free(s->sec_indev); |
| 1456 | g_free(s->outdev); |
| 1457 | g_free(s->notify_dev); |
| 1458 | } |
| 1459 | |
| 1460 | static void __attribute__((__constructor__)) colo_compare_init_globals(void) |
| 1461 | { |
| 1462 | colo_compare_active = false; |
| 1463 | qemu_mutex_init(&colo_compare_mutex); |
| 1464 | } |
| 1465 | |
| 1466 | static const TypeInfo colo_compare_info = { |
| 1467 | .name = TYPE_COLO_COMPARE, |
| 1468 | .parent = TYPE_OBJECT, |
| 1469 | .instance_size = sizeof(CompareState), |
| 1470 | .instance_init = colo_compare_init, |
| 1471 | .instance_finalize = colo_compare_finalize, |
| 1472 | .class_size = sizeof(CompareClass), |
| 1473 | .class_init = colo_compare_class_init, |
| 1474 | .interfaces = (const InterfaceInfo[]) { |
| 1475 | { TYPE_USER_CREATABLE }, |
| 1476 | { } |
| 1477 | } |
| 1478 | }; |
| 1479 | |
| 1480 | static void register_types(void) |
| 1481 | { |
| 1482 | type_register_static(&colo_compare_info); |
| 1483 | } |
| 1484 | |
| 1485 | type_init(register_types); |