]> git.proxmox.com Git - qemu-server.git/blob - qmeventd/qmeventd.c
qmeventd: rework 'forced_cleanup' handling and set timeout to 60s
[qemu-server.git] / qmeventd / qmeventd.c
1 // SPDX-License-Identifier: AGPL-3.0-or-later
2 /*
3 Copyright (C) 2018 - 2021 Proxmox Server Solutions GmbH
4
5 Author: Dominik Csapak <d.csapak@proxmox.com>
6 Author: Stefan Reiter <s.reiter@proxmox.com>
7
8 Description:
9
10 qmeventd listens on a given socket, and waits for qemu processes to
11 connect. After accepting a connection qmeventd waits for shutdown events
12 followed by the closing of the socket. Once that happens `qm cleanup` will
13 be executed with following three arguments:
14 VMID <graceful> <guest>
15 Where `graceful` can be `1` or `0` depending if shutdown event was observed
16 before the socket got closed. The second parameter `guest` is also boolean
17 `1` or `0` depending if the shutdown was requested from the guest OS
18 (i.e., the "inside").
19 */
20
21 #ifndef _GNU_SOURCE
22 #define _GNU_SOURCE
23 #endif
24
25 #include <errno.h>
26 #include <fcntl.h>
27 #include <gmodule.h>
28 #include <json.h>
29 #include <signal.h>
30 #include <stdbool.h>
31 #include <stdio.h>
32 #include <stdlib.h>
33 #include <string.h>
34 #include <sys/epoll.h>
35 #include <sys/socket.h>
36 #include <sys/types.h>
37 #include <sys/un.h>
38 #include <sys/wait.h>
39 #include <unistd.h>
40 #include <time.h>
41
42 #include "qmeventd.h"
43
44 #define DEFAULT_KILL_TIMEOUT 60
45
46 static int verbose = 0;
47 static int kill_timeout = DEFAULT_KILL_TIMEOUT;
48 static int epoll_fd = 0;
49 static const char *progname;
50 GHashTable *vm_clients; // key=vmid (freed on remove), value=*Client (free manually)
51 GSList *forced_cleanups;
52 static int needs_cleanup = 0;
53
54 /*
55 * Helper functions
56 */
57
58 static void
59 usage()
60 {
61 fprintf(stderr, "Usage: %s [-f] [-v] PATH\n", progname);
62 fprintf(stderr, " -f run in foreground (default: false)\n");
63 fprintf(stderr, " -v verbose (default: false)\n");
64 fprintf(stderr, " -t <s> kill timeout (default: %ds)\n", DEFAULT_KILL_TIMEOUT);
65 fprintf(stderr, " PATH use PATH for socket\n");
66 }
67
68 static pid_t
69 get_pid_from_fd(int fd)
70 {
71 struct ucred credentials = { .pid = 0, .uid = 0, .gid = 0 };
72 socklen_t len = sizeof(struct ucred);
73 log_neg(getsockopt(fd, SOL_SOCKET, SO_PEERCRED, &credentials, &len), "getsockopt");
74 return credentials.pid;
75 }
76
77 /*
78 * reads the vmid from /proc/<pid>/cmdline
79 * after the '-id' argument
80 */
81 static unsigned long
82 get_vmid_from_pid(pid_t pid)
83 {
84 char filename[32] = { 0 };
85 int len = snprintf(filename, sizeof(filename), "/proc/%d/cmdline", pid);
86 if (len < 0) {
87 fprintf(stderr, "error during snprintf for %d: %s\n", pid,
88 strerror(errno));
89 return 0;
90 }
91 if ((size_t)len >= sizeof(filename)) {
92 fprintf(stderr, "error: pid %d too long\n", pid);
93 return 0;
94 }
95 FILE *fp = fopen(filename, "re");
96 if (fp == NULL) {
97 fprintf(stderr, "error opening %s: %s\n", filename, strerror(errno));
98 return 0;
99 }
100
101 unsigned long vmid = 0;
102 ssize_t rc = 0;
103 char *buf = NULL;
104 size_t buflen = 0;
105 while ((rc = getdelim(&buf, &buflen, '\0', fp)) >= 0) {
106 if (!strcmp(buf, "-id")) {
107 break;
108 }
109 }
110
111 if (rc < 0) {
112 goto err;
113 }
114
115 if (getdelim(&buf, &buflen, '\0', fp) >= 0) {
116 if (buf[0] == '-' || buf[0] == '\0') {
117 fprintf(stderr, "invalid vmid %s\n", buf);
118 goto ret;
119 }
120
121 errno = 0;
122 char *endptr = NULL;
123 vmid = strtoul(buf, &endptr, 10);
124 if (errno != 0) {
125 vmid = 0;
126 goto err;
127 } else if (*endptr != '\0') {
128 fprintf(stderr, "invalid vmid %s\n", buf);
129 vmid = 0;
130 }
131
132 goto ret;
133 }
134
135 err:
136 fprintf(stderr, "error parsing vmid for %d: %s\n", pid, strerror(errno));
137
138 ret:
139 free(buf);
140 fclose(fp);
141 return vmid;
142 }
143
144 static bool
145 must_write(int fd, const char *buf, size_t len)
146 {
147 ssize_t wlen;
148 do {
149 wlen = write(fd, buf, len);
150 } while (wlen < 0 && errno == EINTR);
151
152 return (wlen == (ssize_t)len);
153 }
154
155 /*
156 * qmp handling functions
157 */
158
159 static void
160 send_qmp_cmd(struct Client *client, const char *buf, size_t len)
161 {
162 if (!must_write(client->fd, buf, len - 1)) {
163 fprintf(stderr, "%s: cannot send QMP message\n", client->qemu.vmid);
164 cleanup_client(client);
165 }
166 }
167
168 void
169 handle_qmp_handshake(struct Client *client)
170 {
171 VERBOSE_PRINT("pid%d: got QMP handshake, assuming QEMU client\n", client->pid);
172
173 // extract vmid from cmdline, now that we know it's a QEMU process
174 unsigned long vmid = get_vmid_from_pid(client->pid);
175 int res = snprintf(client->qemu.vmid, sizeof(client->qemu.vmid), "%lu", vmid);
176 if (vmid == 0 || res < 0 || res >= (int)sizeof(client->qemu.vmid)) {
177 fprintf(stderr, "could not get vmid from pid %d\n", client->pid);
178 cleanup_client(client);
179 return;
180 }
181
182 VERBOSE_PRINT("pid%d: assigned VMID: %s\n", client->pid, client->qemu.vmid);
183 client->type = CLIENT_QEMU;
184 if(!g_hash_table_insert(vm_clients, strdup(client->qemu.vmid), client)) {
185 // not fatal, just means backup handling won't work
186 fprintf(stderr, "%s: could not insert client into VMID->client table\n",
187 client->qemu.vmid);
188 }
189
190 static const char qmp_answer[] = "{\"execute\":\"qmp_capabilities\"}\n";
191 send_qmp_cmd(client, qmp_answer, sizeof(qmp_answer));
192 }
193
194 void
195 handle_qmp_event(struct Client *client, struct json_object *obj)
196 {
197 struct json_object *event;
198 if (!json_object_object_get_ex(obj, "event", &event)) {
199 return;
200 }
201 VERBOSE_PRINT("%s: got QMP event: %s\n", client->qemu.vmid, json_object_get_string(event));
202
203 if (client->state == STATE_TERMINATING) {
204 // QEMU sometimes sends a second SHUTDOWN after SIGTERM, ignore
205 VERBOSE_PRINT("%s: event was after termination, ignoring\n", client->qemu.vmid);
206 return;
207 }
208
209 // event, check if shutdown and get guest parameter
210 if (!strcmp(json_object_get_string(event), "SHUTDOWN")) {
211 client->qemu.graceful = 1;
212 struct json_object *data;
213 struct json_object *guest;
214 if (json_object_object_get_ex(obj, "data", &data) &&
215 json_object_object_get_ex(data, "guest", &guest))
216 {
217 client->qemu.guest = (unsigned short)json_object_get_boolean(guest);
218 }
219
220 // check if a backup is running and kill QEMU process if not
221 terminate_check(client);
222 }
223 }
224
225 void
226 terminate_check(struct Client *client)
227 {
228 if (client->state != STATE_IDLE) {
229 // if we're already in a request, queue this one until after
230 VERBOSE_PRINT("%s: terminate_check queued\n", client->qemu.vmid);
231 client->qemu.term_check_queued = true;
232 return;
233 }
234
235 client->qemu.term_check_queued = false;
236
237 VERBOSE_PRINT("%s: query-status\n", client->qemu.vmid);
238 client->state = STATE_EXPECT_STATUS_RESP;
239 static const char qmp_req[] = "{\"execute\":\"query-status\"}\n";
240 send_qmp_cmd(client, qmp_req, sizeof(qmp_req));
241 }
242
243 void
244 handle_qmp_return(struct Client *client, struct json_object *data, bool error)
245 {
246 if (error) {
247 const char *msg = "n/a";
248 struct json_object *desc;
249 if (json_object_object_get_ex(data, "desc", &desc)) {
250 msg = json_object_get_string(desc);
251 }
252 fprintf(stderr, "%s: received error from QMP: %s\n",
253 client->qemu.vmid, msg);
254 client->state = STATE_IDLE;
255 goto out;
256 }
257
258 struct json_object *status;
259 json_bool has_status = data &&
260 json_object_object_get_ex(data, "status", &status);
261
262 bool active = false;
263 if (has_status) {
264 const char *status_str = json_object_get_string(status);
265 active = status_str &&
266 (!strcmp(status_str, "running") || !strcmp(status_str, "paused"));
267 }
268
269 switch (client->state) {
270 case STATE_EXPECT_STATUS_RESP:
271 client->state = STATE_IDLE;
272 if (active) {
273 VERBOSE_PRINT("%s: got status: VM is active\n", client->qemu.vmid);
274 } else if (!client->qemu.backup) {
275 terminate_client(client);
276 } else {
277 // if we're in a backup, don't do anything, vzdump will notify
278 // us when the backup finishes
279 VERBOSE_PRINT("%s: not active, but running backup - keep alive\n",
280 client->qemu.vmid);
281 }
282 break;
283
284 // this means we received the empty return from our handshake answer
285 case STATE_HANDSHAKE:
286 client->state = STATE_IDLE;
287 VERBOSE_PRINT("%s: QMP handshake complete\n", client->qemu.vmid);
288 break;
289
290 case STATE_IDLE:
291 case STATE_TERMINATING:
292 VERBOSE_PRINT("%s: spurious return value received\n",
293 client->qemu.vmid);
294 break;
295 }
296
297 out:
298 if (client->qemu.term_check_queued) {
299 terminate_check(client);
300 }
301 }
302
303 /*
304 * VZDump specific client functions
305 */
306
307 void
308 handle_vzdump_handshake(struct Client *client, struct json_object *data)
309 {
310 client->state = STATE_IDLE;
311
312 struct json_object *vmid_obj;
313 json_bool has_vmid = data && json_object_object_get_ex(data, "vmid", &vmid_obj);
314
315 if (!has_vmid) {
316 VERBOSE_PRINT("pid%d: invalid vzdump handshake: no vmid\n", client->pid);
317 return;
318 }
319
320 const char *vmid_str = json_object_get_string(vmid_obj);
321
322 if (!vmid_str) {
323 VERBOSE_PRINT("pid%d: invalid vzdump handshake: vmid is not a string\n", client->pid);
324 return;
325 }
326
327 int res = snprintf(client->vzdump.vmid, sizeof(client->vzdump.vmid), "%s", vmid_str);
328 if (res < 0 || res >= (int)sizeof(client->vzdump.vmid)) {
329 VERBOSE_PRINT("pid%d: invalid vzdump handshake: vmid too long or invalid\n", client->pid);
330 return;
331 }
332
333 struct Client *vmc = (struct Client*) g_hash_table_lookup(vm_clients, client->vzdump.vmid);
334 if (vmc) {
335 vmc->qemu.backup = true;
336
337 // only mark as VZDUMP once we have set everything up, otherwise 'cleanup'
338 // might try to access an invalid value
339 client->type = CLIENT_VZDUMP;
340 VERBOSE_PRINT("%s: vzdump backup started\n", client->vzdump.vmid);
341 } else {
342 VERBOSE_PRINT("%s: vzdump requested backup start for unregistered VM\n", client->vzdump.vmid);
343 }
344 }
345
346 /*
347 * client management functions
348 */
349
350 void
351 add_new_client(int client_fd)
352 {
353 struct Client *client = calloc(sizeof(struct Client), 1);
354 if (client == NULL) {
355 fprintf(stderr, "could not add new client - allocation failed!\n");
356 fflush(stderr);
357 return;
358 }
359 client->state = STATE_HANDSHAKE;
360 client->type = CLIENT_NONE;
361 client->fd = client_fd;
362 client->pid = get_pid_from_fd(client_fd);
363 if (client->pid == 0) {
364 fprintf(stderr, "could not get pid from client\n");
365 goto err;
366 }
367
368 struct epoll_event ev;
369 ev.events = EPOLLIN;
370 ev.data.ptr = client;
371 int res = epoll_ctl(epoll_fd, EPOLL_CTL_ADD, client_fd, &ev);
372 if (res < 0) {
373 perror("epoll_ctl client add");
374 goto err;
375 }
376
377 VERBOSE_PRINT("added new client, pid: %d\n", client->pid);
378
379 return;
380 err:
381 (void)close(client_fd);
382 free(client);
383 }
384
385 static void
386 cleanup_qemu_client(struct Client *client)
387 {
388 unsigned short graceful = client->qemu.graceful;
389 unsigned short guest = client->qemu.guest;
390 char vmid[sizeof(client->qemu.vmid)];
391 strncpy(vmid, client->qemu.vmid, sizeof(vmid));
392 g_hash_table_remove(vm_clients, &vmid); // frees key, ignore errors
393 VERBOSE_PRINT("%s: executing cleanup (graceful: %d, guest: %d)\n",
394 vmid, graceful, guest);
395
396 int pid = fork();
397 if (pid < 0) {
398 fprintf(stderr, "fork failed: %s\n", strerror(errno));
399 return;
400 }
401 if (pid == 0) {
402 char *script = "/usr/sbin/qm";
403
404 char *args[] = {
405 script,
406 "cleanup",
407 vmid,
408 graceful ? "1" : "0",
409 guest ? "1" : "0",
410 NULL
411 };
412
413 execvp(script, args);
414 perror("execvp");
415 _exit(1);
416 }
417 }
418
419 void
420 cleanup_client(struct Client *client)
421 {
422 log_neg(epoll_ctl(epoll_fd, EPOLL_CTL_DEL, client->fd, NULL), "epoll del");
423 (void)close(client->fd);
424
425 struct Client *vmc;
426 switch (client->type) {
427 case CLIENT_QEMU:
428 cleanup_qemu_client(client);
429 break;
430
431 case CLIENT_VZDUMP:
432 vmc = (struct Client*) g_hash_table_lookup(vm_clients, client->vzdump.vmid);
433 if (vmc) {
434 VERBOSE_PRINT("%s: backup ended\n", client->vzdump.vmid);
435 vmc->qemu.backup = false;
436 terminate_check(vmc);
437 }
438 break;
439
440 case CLIENT_NONE:
441 // do nothing, only close socket
442 break;
443 }
444
445 if (client->pidfd > 0) {
446 (void)close(client->pidfd);
447 }
448 VERBOSE_PRINT("removing %s from forced cleanups\n", client->qemu.vmid);
449 forced_cleanups = g_slist_remove(forced_cleanups, client);
450 free(client);
451 }
452
453 void
454 terminate_client(struct Client *client)
455 {
456 VERBOSE_PRINT("%s: terminating client (pid %d)\n", client->qemu.vmid, client->pid);
457
458 client->state = STATE_TERMINATING;
459
460 // open a pidfd before kill for later cleanup
461 int pidfd = pidfd_open(client->pid, 0);
462 if (pidfd < 0) {
463 switch (errno) {
464 case ESRCH:
465 // process already dead for some reason, cleanup done
466 VERBOSE_PRINT("%s: failed to open pidfd, process already dead (pid %d)\n",
467 client->qemu.vmid, client->pid);
468 return;
469
470 // otherwise fall back to just using the PID directly, but don't
471 // print if we only failed because we're running on an older kernel
472 case ENOSYS:
473 break;
474 default:
475 perror("failed to open QEMU pidfd for cleanup");
476 break;
477 }
478 }
479
480 int err = kill(client->pid, SIGTERM);
481 log_neg(err, "kill");
482
483 time_t timeout = time(NULL) + kill_timeout;
484
485 client->pidfd = pidfd;
486 client->timeout = timeout;
487
488 forced_cleanups = g_slist_prepend(forced_cleanups, (void *)client);
489 needs_cleanup = 1;
490 }
491
492 void
493 handle_client(struct Client *client)
494 {
495 VERBOSE_PRINT("pid%d: entering handle\n", client->pid);
496 ssize_t len;
497 do {
498 len = read(client->fd, (client->buf+client->buflen),
499 sizeof(client->buf) - client->buflen);
500 } while (len < 0 && errno == EINTR);
501
502 if (len < 0) {
503 if (!(errno == EAGAIN || errno == EWOULDBLOCK)) {
504 log_neg((int)len, "read");
505 cleanup_client(client);
506 }
507 return;
508 } else if (len == 0) {
509 VERBOSE_PRINT("pid%d: got EOF\n", client->pid);
510 cleanup_client(client);
511 return;
512 }
513
514 VERBOSE_PRINT("pid%d: read %ld bytes\n", client->pid, len);
515 client->buflen += len;
516
517 struct json_tokener *tok = json_tokener_new();
518 struct json_object *jobj = NULL;
519 enum json_tokener_error jerr = json_tokener_success;
520 while (jerr == json_tokener_success && client->buflen != 0) {
521 jobj = json_tokener_parse_ex(tok, client->buf, (int)client->buflen);
522 jerr = json_tokener_get_error(tok);
523 unsigned int offset = (unsigned int)tok->char_offset;
524 switch (jerr) {
525 case json_tokener_success:
526 // move rest from buffer to front
527 memmove(client->buf, client->buf + offset, client->buflen - offset);
528 client->buflen -= offset;
529 if (json_object_is_type(jobj, json_type_object)) {
530 struct json_object *obj;
531 if (json_object_object_get_ex(jobj, "QMP", &obj)) {
532 handle_qmp_handshake(client);
533 } else if (json_object_object_get_ex(jobj, "event", &obj)) {
534 handle_qmp_event(client, jobj);
535 } else if (json_object_object_get_ex(jobj, "return", &obj)) {
536 handle_qmp_return(client, obj, false);
537 } else if (json_object_object_get_ex(jobj, "error", &obj)) {
538 handle_qmp_return(client, obj, true);
539 } else if (json_object_object_get_ex(jobj, "vzdump", &obj)) {
540 handle_vzdump_handshake(client, obj);
541 } // else ignore message
542 }
543 break;
544 case json_tokener_continue:
545 if (client->buflen >= sizeof(client->buf)) {
546 VERBOSE_PRINT("pid%d: msg too large, discarding buffer\n", client->pid);
547 memset(client->buf, 0, sizeof(client->buf));
548 client->buflen = 0;
549 } // else we have enough space try again after next read
550 break;
551 default:
552 VERBOSE_PRINT("pid%d: parse error: %d, discarding buffer\n", client->pid, jerr);
553 memset(client->buf, 0, client->buflen);
554 client->buflen = 0;
555 break;
556 }
557 json_object_put(jobj);
558 }
559 json_tokener_free(tok);
560 }
561
562 static void
563 sigkill(void *ptr, void *time_ptr)
564 {
565 struct Client *data = ptr;
566 int err;
567
568 if (data->timeout != 0 && data->timeout > *(time_t *)time_ptr) {
569 return;
570 }
571
572 if (data->pidfd > 0) {
573 err = pidfd_send_signal(data->pidfd, SIGKILL, NULL, 0);
574 (void)close(data->pidfd);
575 data->pidfd = -1;
576 } else {
577 err = kill(data->pid, SIGKILL);
578 }
579
580 if (err < 0) {
581 if (errno != ESRCH) {
582 fprintf(stderr, "SIGKILL cleanup of pid '%d' failed - %s\n",
583 data->pid, strerror(errno));
584 }
585 } else {
586 fprintf(stderr, "cleanup failed, terminating pid '%d' with SIGKILL\n",
587 data->pid);
588 }
589
590 data->timeout = 0;
591
592 // remove ourselves from the list
593 forced_cleanups = g_slist_remove(forced_cleanups, ptr);
594 }
595
596 static void
597 handle_forced_cleanup()
598 {
599 if (g_slist_length(forced_cleanups) > 0) {
600 VERBOSE_PRINT("clearing forced cleanup backlog\n");
601 time_t cur_time = time(NULL);
602 g_slist_foreach(forced_cleanups, sigkill, &cur_time);
603 }
604 needs_cleanup = g_slist_length(forced_cleanups) > 0;
605 }
606
607 int
608 main(int argc, char *argv[])
609 {
610 int opt;
611 int daemonize = 1;
612 char *socket_path = NULL;
613 progname = argv[0];
614
615 while ((opt = getopt(argc, argv, "hfvt:")) != -1) {
616 switch (opt) {
617 case 'f':
618 daemonize = 0;
619 break;
620 case 'v':
621 verbose = 1;
622 break;
623 case 't':
624 errno = 0;
625 char *endptr = NULL;
626 kill_timeout = strtoul(optarg, &endptr, 10);
627 if (errno != 0 || *endptr != '\0' || kill_timeout == 0) {
628 usage();
629 exit(EXIT_FAILURE);
630 }
631 break;
632 case 'h':
633 usage();
634 exit(EXIT_SUCCESS);
635 break;
636 default:
637 usage();
638 exit(EXIT_FAILURE);
639 }
640 }
641
642 if (optind >= argc) {
643 usage();
644 exit(EXIT_FAILURE);
645 }
646
647 signal(SIGCHLD, SIG_IGN);
648
649 socket_path = argv[optind];
650
651 int sock = socket(AF_UNIX, SOCK_STREAM, 0);
652 bail_neg(sock, "socket");
653
654 struct sockaddr_un addr;
655 memset(&addr, 0, sizeof(addr));
656 addr.sun_family = AF_UNIX;
657 strncpy(addr.sun_path, socket_path, sizeof(addr.sun_path) - 1);
658
659 unlink(socket_path);
660 bail_neg(bind(sock, (struct sockaddr*)&addr, sizeof(addr)), "bind");
661
662 struct epoll_event ev, events[1];
663 epoll_fd = epoll_create1(EPOLL_CLOEXEC);
664 bail_neg(epoll_fd, "epoll_create1");
665
666 ev.events = EPOLLIN;
667 ev.data.fd = sock;
668 bail_neg(epoll_ctl(epoll_fd, EPOLL_CTL_ADD, sock, &ev), "epoll_ctl");
669
670 bail_neg(listen(sock, 10), "listen");
671
672 if (daemonize) {
673 bail_neg(daemon(0, 1), "daemon");
674 }
675
676 vm_clients = g_hash_table_new_full(g_str_hash, g_str_equal, free, NULL);
677
678 int nevents;
679
680 for(;;) {
681 nevents = epoll_wait(epoll_fd, events, 1, needs_cleanup ? 10*1000 : -1);
682 bail_neg(nevents, "epoll_wait");
683
684 for (int n = 0; n < nevents; n++) {
685 if (events[n].data.fd == sock) {
686
687 int conn_sock = accept4(sock, NULL, NULL, SOCK_NONBLOCK | SOCK_CLOEXEC);
688 log_neg(conn_sock, "accept");
689 if (conn_sock > -1) {
690 add_new_client(conn_sock);
691 }
692 } else {
693 handle_client((struct Client *)events[n].data.ptr);
694 }
695 }
696 handle_forced_cleanup();
697 }
698 }