xref: /linux/net/ceph/mon_client.c (revision 0883c2c06fb5bcf5b9e008270827e63c09a88c1e)
1 #include <linux/ceph/ceph_debug.h>
2 
3 #include <linux/module.h>
4 #include <linux/types.h>
5 #include <linux/slab.h>
6 #include <linux/random.h>
7 #include <linux/sched.h>
8 
9 #include <linux/ceph/mon_client.h>
10 #include <linux/ceph/libceph.h>
11 #include <linux/ceph/debugfs.h>
12 #include <linux/ceph/decode.h>
13 #include <linux/ceph/auth.h>
14 
15 /*
16  * Interact with Ceph monitor cluster.  Handle requests for new map
17  * versions, and periodically resend as needed.  Also implement
18  * statfs() and umount().
19  *
20  * A small cluster of Ceph "monitors" are responsible for managing critical
21  * cluster configuration and state information.  An odd number (e.g., 3, 5)
22  * of cmon daemons use a modified version of the Paxos part-time parliament
23  * algorithm to manage the MDS map (mds cluster membership), OSD map, and
24  * list of clients who have mounted the file system.
25  *
26  * We maintain an open, active session with a monitor at all times in order to
27  * receive timely MDSMap updates.  We periodically send a keepalive byte on the
28  * TCP socket to ensure we detect a failure.  If the connection does break, we
29  * randomly hunt for a new monitor.  Once the connection is reestablished, we
30  * resend any outstanding requests.
31  */
32 
33 static const struct ceph_connection_operations mon_con_ops;
34 
35 static int __validate_auth(struct ceph_mon_client *monc);
36 
37 /*
38  * Decode a monmap blob (e.g., during mount).
39  */
40 struct ceph_monmap *ceph_monmap_decode(void *p, void *end)
41 {
42 	struct ceph_monmap *m = NULL;
43 	int i, err = -EINVAL;
44 	struct ceph_fsid fsid;
45 	u32 epoch, num_mon;
46 	u16 version;
47 	u32 len;
48 
49 	ceph_decode_32_safe(&p, end, len, bad);
50 	ceph_decode_need(&p, end, len, bad);
51 
52 	dout("monmap_decode %p %p len %d\n", p, end, (int)(end-p));
53 
54 	ceph_decode_16_safe(&p, end, version, bad);
55 
56 	ceph_decode_need(&p, end, sizeof(fsid) + 2*sizeof(u32), bad);
57 	ceph_decode_copy(&p, &fsid, sizeof(fsid));
58 	epoch = ceph_decode_32(&p);
59 
60 	num_mon = ceph_decode_32(&p);
61 	ceph_decode_need(&p, end, num_mon*sizeof(m->mon_inst[0]), bad);
62 
63 	if (num_mon >= CEPH_MAX_MON)
64 		goto bad;
65 	m = kmalloc(sizeof(*m) + sizeof(m->mon_inst[0])*num_mon, GFP_NOFS);
66 	if (m == NULL)
67 		return ERR_PTR(-ENOMEM);
68 	m->fsid = fsid;
69 	m->epoch = epoch;
70 	m->num_mon = num_mon;
71 	ceph_decode_copy(&p, m->mon_inst, num_mon*sizeof(m->mon_inst[0]));
72 	for (i = 0; i < num_mon; i++)
73 		ceph_decode_addr(&m->mon_inst[i].addr);
74 
75 	dout("monmap_decode epoch %d, num_mon %d\n", m->epoch,
76 	     m->num_mon);
77 	for (i = 0; i < m->num_mon; i++)
78 		dout("monmap_decode  mon%d is %s\n", i,
79 		     ceph_pr_addr(&m->mon_inst[i].addr.in_addr));
80 	return m;
81 
82 bad:
83 	dout("monmap_decode failed with %d\n", err);
84 	kfree(m);
85 	return ERR_PTR(err);
86 }
87 
88 /*
89  * return true if *addr is included in the monmap.
90  */
91 int ceph_monmap_contains(struct ceph_monmap *m, struct ceph_entity_addr *addr)
92 {
93 	int i;
94 
95 	for (i = 0; i < m->num_mon; i++)
96 		if (memcmp(addr, &m->mon_inst[i].addr, sizeof(*addr)) == 0)
97 			return 1;
98 	return 0;
99 }
100 
101 /*
102  * Send an auth request.
103  */
104 static void __send_prepared_auth_request(struct ceph_mon_client *monc, int len)
105 {
106 	monc->pending_auth = 1;
107 	monc->m_auth->front.iov_len = len;
108 	monc->m_auth->hdr.front_len = cpu_to_le32(len);
109 	ceph_msg_revoke(monc->m_auth);
110 	ceph_msg_get(monc->m_auth);  /* keep our ref */
111 	ceph_con_send(&monc->con, monc->m_auth);
112 }
113 
114 /*
115  * Close monitor session, if any.
116  */
117 static void __close_session(struct ceph_mon_client *monc)
118 {
119 	dout("__close_session closing mon%d\n", monc->cur_mon);
120 	ceph_msg_revoke(monc->m_auth);
121 	ceph_msg_revoke_incoming(monc->m_auth_reply);
122 	ceph_msg_revoke(monc->m_subscribe);
123 	ceph_msg_revoke_incoming(monc->m_subscribe_ack);
124 	ceph_con_close(&monc->con);
125 
126 	monc->pending_auth = 0;
127 	ceph_auth_reset(monc->auth);
128 }
129 
130 /*
131  * Pick a new monitor at random and set cur_mon.  If we are repicking
132  * (i.e. cur_mon is already set), be sure to pick a different one.
133  */
134 static void pick_new_mon(struct ceph_mon_client *monc)
135 {
136 	int old_mon = monc->cur_mon;
137 
138 	BUG_ON(monc->monmap->num_mon < 1);
139 
140 	if (monc->monmap->num_mon == 1) {
141 		monc->cur_mon = 0;
142 	} else {
143 		int max = monc->monmap->num_mon;
144 		int o = -1;
145 		int n;
146 
147 		if (monc->cur_mon >= 0) {
148 			if (monc->cur_mon < monc->monmap->num_mon)
149 				o = monc->cur_mon;
150 			if (o >= 0)
151 				max--;
152 		}
153 
154 		n = prandom_u32() % max;
155 		if (o >= 0 && n >= o)
156 			n++;
157 
158 		monc->cur_mon = n;
159 	}
160 
161 	dout("%s mon%d -> mon%d out of %d mons\n", __func__, old_mon,
162 	     monc->cur_mon, monc->monmap->num_mon);
163 }
164 
165 /*
166  * Open a session with a new monitor.
167  */
168 static void __open_session(struct ceph_mon_client *monc)
169 {
170 	int ret;
171 
172 	pick_new_mon(monc);
173 
174 	monc->hunting = true;
175 	if (monc->had_a_connection) {
176 		monc->hunt_mult *= CEPH_MONC_HUNT_BACKOFF;
177 		if (monc->hunt_mult > CEPH_MONC_HUNT_MAX_MULT)
178 			monc->hunt_mult = CEPH_MONC_HUNT_MAX_MULT;
179 	}
180 
181 	monc->sub_renew_after = jiffies; /* i.e., expired */
182 	monc->sub_renew_sent = 0;
183 
184 	dout("%s opening mon%d\n", __func__, monc->cur_mon);
185 	ceph_con_open(&monc->con, CEPH_ENTITY_TYPE_MON, monc->cur_mon,
186 		      &monc->monmap->mon_inst[monc->cur_mon].addr);
187 
188 	/*
189 	 * send an initial keepalive to ensure our timestamp is valid
190 	 * by the time we are in an OPENED state
191 	 */
192 	ceph_con_keepalive(&monc->con);
193 
194 	/* initiate authentication handshake */
195 	ret = ceph_auth_build_hello(monc->auth,
196 				    monc->m_auth->front.iov_base,
197 				    monc->m_auth->front_alloc_len);
198 	BUG_ON(ret <= 0);
199 	__send_prepared_auth_request(monc, ret);
200 }
201 
202 static void reopen_session(struct ceph_mon_client *monc)
203 {
204 	if (!monc->hunting)
205 		pr_info("mon%d %s session lost, hunting for new mon\n",
206 		    monc->cur_mon, ceph_pr_addr(&monc->con.peer_addr.in_addr));
207 
208 	__close_session(monc);
209 	__open_session(monc);
210 }
211 
212 /*
213  * Reschedule delayed work timer.
214  */
215 static void __schedule_delayed(struct ceph_mon_client *monc)
216 {
217 	unsigned long delay;
218 
219 	if (monc->hunting)
220 		delay = CEPH_MONC_HUNT_INTERVAL * monc->hunt_mult;
221 	else
222 		delay = CEPH_MONC_PING_INTERVAL;
223 
224 	dout("__schedule_delayed after %lu\n", delay);
225 	mod_delayed_work(system_wq, &monc->delayed_work,
226 			 round_jiffies_relative(delay));
227 }
228 
229 const char *ceph_sub_str[] = {
230 	[CEPH_SUB_MDSMAP] = "mdsmap",
231 	[CEPH_SUB_MONMAP] = "monmap",
232 	[CEPH_SUB_OSDMAP] = "osdmap",
233 };
234 
235 /*
236  * Send subscribe request for one or more maps, according to
237  * monc->subs.
238  */
239 static void __send_subscribe(struct ceph_mon_client *monc)
240 {
241 	struct ceph_msg *msg = monc->m_subscribe;
242 	void *p = msg->front.iov_base;
243 	void *const end = p + msg->front_alloc_len;
244 	int num = 0;
245 	int i;
246 
247 	dout("%s sent %lu\n", __func__, monc->sub_renew_sent);
248 
249 	BUG_ON(monc->cur_mon < 0);
250 
251 	if (!monc->sub_renew_sent)
252 		monc->sub_renew_sent = jiffies | 1; /* never 0 */
253 
254 	msg->hdr.version = cpu_to_le16(2);
255 
256 	for (i = 0; i < ARRAY_SIZE(monc->subs); i++) {
257 		if (monc->subs[i].want)
258 			num++;
259 	}
260 	BUG_ON(num < 1); /* monmap sub is always there */
261 	ceph_encode_32(&p, num);
262 	for (i = 0; i < ARRAY_SIZE(monc->subs); i++) {
263 		char buf[32];
264 		int len;
265 
266 		if (!monc->subs[i].want)
267 			continue;
268 
269 		len = sprintf(buf, "%s", ceph_sub_str[i]);
270 		if (i == CEPH_SUB_MDSMAP &&
271 		    monc->fs_cluster_id != CEPH_FS_CLUSTER_ID_NONE)
272 			len += sprintf(buf + len, ".%d", monc->fs_cluster_id);
273 
274 		dout("%s %s start %llu flags 0x%x\n", __func__, buf,
275 		     le64_to_cpu(monc->subs[i].item.start),
276 		     monc->subs[i].item.flags);
277 		ceph_encode_string(&p, end, buf, len);
278 		memcpy(p, &monc->subs[i].item, sizeof(monc->subs[i].item));
279 		p += sizeof(monc->subs[i].item);
280 	}
281 
282 	BUG_ON(p > end);
283 	msg->front.iov_len = p - msg->front.iov_base;
284 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
285 	ceph_msg_revoke(msg);
286 	ceph_con_send(&monc->con, ceph_msg_get(msg));
287 }
288 
289 static void handle_subscribe_ack(struct ceph_mon_client *monc,
290 				 struct ceph_msg *msg)
291 {
292 	unsigned int seconds;
293 	struct ceph_mon_subscribe_ack *h = msg->front.iov_base;
294 
295 	if (msg->front.iov_len < sizeof(*h))
296 		goto bad;
297 	seconds = le32_to_cpu(h->duration);
298 
299 	mutex_lock(&monc->mutex);
300 	if (monc->sub_renew_sent) {
301 		monc->sub_renew_after = monc->sub_renew_sent +
302 					    (seconds >> 1) * HZ - 1;
303 		dout("%s sent %lu duration %d renew after %lu\n", __func__,
304 		     monc->sub_renew_sent, seconds, monc->sub_renew_after);
305 		monc->sub_renew_sent = 0;
306 	} else {
307 		dout("%s sent %lu renew after %lu, ignoring\n", __func__,
308 		     monc->sub_renew_sent, monc->sub_renew_after);
309 	}
310 	mutex_unlock(&monc->mutex);
311 	return;
312 bad:
313 	pr_err("got corrupt subscribe-ack msg\n");
314 	ceph_msg_dump(msg);
315 }
316 
317 /*
318  * Register interest in a map
319  *
320  * @sub: one of CEPH_SUB_*
321  * @epoch: X for "every map since X", or 0 for "just the latest"
322  */
323 static bool __ceph_monc_want_map(struct ceph_mon_client *monc, int sub,
324 				 u32 epoch, bool continuous)
325 {
326 	__le64 start = cpu_to_le64(epoch);
327 	u8 flags = !continuous ? CEPH_SUBSCRIBE_ONETIME : 0;
328 
329 	dout("%s %s epoch %u continuous %d\n", __func__, ceph_sub_str[sub],
330 	     epoch, continuous);
331 
332 	if (monc->subs[sub].want &&
333 	    monc->subs[sub].item.start == start &&
334 	    monc->subs[sub].item.flags == flags)
335 		return false;
336 
337 	monc->subs[sub].item.start = start;
338 	monc->subs[sub].item.flags = flags;
339 	monc->subs[sub].want = true;
340 
341 	return true;
342 }
343 
344 bool ceph_monc_want_map(struct ceph_mon_client *monc, int sub, u32 epoch,
345 			bool continuous)
346 {
347 	bool need_request;
348 
349 	mutex_lock(&monc->mutex);
350 	need_request = __ceph_monc_want_map(monc, sub, epoch, continuous);
351 	mutex_unlock(&monc->mutex);
352 
353 	return need_request;
354 }
355 EXPORT_SYMBOL(ceph_monc_want_map);
356 
357 /*
358  * Keep track of which maps we have
359  *
360  * @sub: one of CEPH_SUB_*
361  */
362 static void __ceph_monc_got_map(struct ceph_mon_client *monc, int sub,
363 				u32 epoch)
364 {
365 	dout("%s %s epoch %u\n", __func__, ceph_sub_str[sub], epoch);
366 
367 	if (monc->subs[sub].want) {
368 		if (monc->subs[sub].item.flags & CEPH_SUBSCRIBE_ONETIME)
369 			monc->subs[sub].want = false;
370 		else
371 			monc->subs[sub].item.start = cpu_to_le64(epoch + 1);
372 	}
373 
374 	monc->subs[sub].have = epoch;
375 }
376 
377 void ceph_monc_got_map(struct ceph_mon_client *monc, int sub, u32 epoch)
378 {
379 	mutex_lock(&monc->mutex);
380 	__ceph_monc_got_map(monc, sub, epoch);
381 	mutex_unlock(&monc->mutex);
382 }
383 EXPORT_SYMBOL(ceph_monc_got_map);
384 
385 void ceph_monc_renew_subs(struct ceph_mon_client *monc)
386 {
387 	mutex_lock(&monc->mutex);
388 	__send_subscribe(monc);
389 	mutex_unlock(&monc->mutex);
390 }
391 EXPORT_SYMBOL(ceph_monc_renew_subs);
392 
393 /*
394  * Wait for an osdmap with a given epoch.
395  *
396  * @epoch: epoch to wait for
397  * @timeout: in jiffies, 0 means "wait forever"
398  */
399 int ceph_monc_wait_osdmap(struct ceph_mon_client *monc, u32 epoch,
400 			  unsigned long timeout)
401 {
402 	unsigned long started = jiffies;
403 	long ret;
404 
405 	mutex_lock(&monc->mutex);
406 	while (monc->subs[CEPH_SUB_OSDMAP].have < epoch) {
407 		mutex_unlock(&monc->mutex);
408 
409 		if (timeout && time_after_eq(jiffies, started + timeout))
410 			return -ETIMEDOUT;
411 
412 		ret = wait_event_interruptible_timeout(monc->client->auth_wq,
413 				     monc->subs[CEPH_SUB_OSDMAP].have >= epoch,
414 				     ceph_timeout_jiffies(timeout));
415 		if (ret < 0)
416 			return ret;
417 
418 		mutex_lock(&monc->mutex);
419 	}
420 
421 	mutex_unlock(&monc->mutex);
422 	return 0;
423 }
424 EXPORT_SYMBOL(ceph_monc_wait_osdmap);
425 
426 /*
427  * Open a session with a random monitor.  Request monmap and osdmap,
428  * which are waited upon in __ceph_open_session().
429  */
430 int ceph_monc_open_session(struct ceph_mon_client *monc)
431 {
432 	mutex_lock(&monc->mutex);
433 	__ceph_monc_want_map(monc, CEPH_SUB_MONMAP, 0, true);
434 	__ceph_monc_want_map(monc, CEPH_SUB_OSDMAP, 0, false);
435 	__open_session(monc);
436 	__schedule_delayed(monc);
437 	mutex_unlock(&monc->mutex);
438 	return 0;
439 }
440 EXPORT_SYMBOL(ceph_monc_open_session);
441 
442 static void ceph_monc_handle_map(struct ceph_mon_client *monc,
443 				 struct ceph_msg *msg)
444 {
445 	struct ceph_client *client = monc->client;
446 	struct ceph_monmap *monmap = NULL, *old = monc->monmap;
447 	void *p, *end;
448 
449 	mutex_lock(&monc->mutex);
450 
451 	dout("handle_monmap\n");
452 	p = msg->front.iov_base;
453 	end = p + msg->front.iov_len;
454 
455 	monmap = ceph_monmap_decode(p, end);
456 	if (IS_ERR(monmap)) {
457 		pr_err("problem decoding monmap, %d\n",
458 		       (int)PTR_ERR(monmap));
459 		goto out;
460 	}
461 
462 	if (ceph_check_fsid(monc->client, &monmap->fsid) < 0) {
463 		kfree(monmap);
464 		goto out;
465 	}
466 
467 	client->monc.monmap = monmap;
468 	kfree(old);
469 
470 	__ceph_monc_got_map(monc, CEPH_SUB_MONMAP, monc->monmap->epoch);
471 	client->have_fsid = true;
472 
473 out:
474 	mutex_unlock(&monc->mutex);
475 	wake_up_all(&client->auth_wq);
476 }
477 
478 /*
479  * generic requests (currently statfs, mon_get_version)
480  */
481 DEFINE_RB_FUNCS(generic_request, struct ceph_mon_generic_request, tid, node)
482 
483 static void release_generic_request(struct kref *kref)
484 {
485 	struct ceph_mon_generic_request *req =
486 		container_of(kref, struct ceph_mon_generic_request, kref);
487 
488 	dout("%s greq %p request %p reply %p\n", __func__, req, req->request,
489 	     req->reply);
490 	WARN_ON(!RB_EMPTY_NODE(&req->node));
491 
492 	if (req->reply)
493 		ceph_msg_put(req->reply);
494 	if (req->request)
495 		ceph_msg_put(req->request);
496 
497 	kfree(req);
498 }
499 
500 static void put_generic_request(struct ceph_mon_generic_request *req)
501 {
502 	if (req)
503 		kref_put(&req->kref, release_generic_request);
504 }
505 
506 static void get_generic_request(struct ceph_mon_generic_request *req)
507 {
508 	kref_get(&req->kref);
509 }
510 
511 static struct ceph_mon_generic_request *
512 alloc_generic_request(struct ceph_mon_client *monc, gfp_t gfp)
513 {
514 	struct ceph_mon_generic_request *req;
515 
516 	req = kzalloc(sizeof(*req), gfp);
517 	if (!req)
518 		return NULL;
519 
520 	req->monc = monc;
521 	kref_init(&req->kref);
522 	RB_CLEAR_NODE(&req->node);
523 	init_completion(&req->completion);
524 
525 	dout("%s greq %p\n", __func__, req);
526 	return req;
527 }
528 
529 static void register_generic_request(struct ceph_mon_generic_request *req)
530 {
531 	struct ceph_mon_client *monc = req->monc;
532 
533 	WARN_ON(req->tid);
534 
535 	get_generic_request(req);
536 	req->tid = ++monc->last_tid;
537 	insert_generic_request(&monc->generic_request_tree, req);
538 }
539 
540 static void send_generic_request(struct ceph_mon_client *monc,
541 				 struct ceph_mon_generic_request *req)
542 {
543 	WARN_ON(!req->tid);
544 
545 	dout("%s greq %p tid %llu\n", __func__, req, req->tid);
546 	req->request->hdr.tid = cpu_to_le64(req->tid);
547 	ceph_con_send(&monc->con, ceph_msg_get(req->request));
548 }
549 
550 static void __finish_generic_request(struct ceph_mon_generic_request *req)
551 {
552 	struct ceph_mon_client *monc = req->monc;
553 
554 	dout("%s greq %p tid %llu\n", __func__, req, req->tid);
555 	erase_generic_request(&monc->generic_request_tree, req);
556 
557 	ceph_msg_revoke(req->request);
558 	ceph_msg_revoke_incoming(req->reply);
559 }
560 
561 static void finish_generic_request(struct ceph_mon_generic_request *req)
562 {
563 	__finish_generic_request(req);
564 	put_generic_request(req);
565 }
566 
567 static void complete_generic_request(struct ceph_mon_generic_request *req)
568 {
569 	if (req->complete_cb)
570 		req->complete_cb(req);
571 	else
572 		complete_all(&req->completion);
573 	put_generic_request(req);
574 }
575 
576 void cancel_generic_request(struct ceph_mon_generic_request *req)
577 {
578 	struct ceph_mon_client *monc = req->monc;
579 	struct ceph_mon_generic_request *lookup_req;
580 
581 	dout("%s greq %p tid %llu\n", __func__, req, req->tid);
582 
583 	mutex_lock(&monc->mutex);
584 	lookup_req = lookup_generic_request(&monc->generic_request_tree,
585 					    req->tid);
586 	if (lookup_req) {
587 		WARN_ON(lookup_req != req);
588 		finish_generic_request(req);
589 	}
590 
591 	mutex_unlock(&monc->mutex);
592 }
593 
594 static int wait_generic_request(struct ceph_mon_generic_request *req)
595 {
596 	int ret;
597 
598 	dout("%s greq %p tid %llu\n", __func__, req, req->tid);
599 	ret = wait_for_completion_interruptible(&req->completion);
600 	if (ret)
601 		cancel_generic_request(req);
602 	else
603 		ret = req->result; /* completed */
604 
605 	return ret;
606 }
607 
608 static struct ceph_msg *get_generic_reply(struct ceph_connection *con,
609 					 struct ceph_msg_header *hdr,
610 					 int *skip)
611 {
612 	struct ceph_mon_client *monc = con->private;
613 	struct ceph_mon_generic_request *req;
614 	u64 tid = le64_to_cpu(hdr->tid);
615 	struct ceph_msg *m;
616 
617 	mutex_lock(&monc->mutex);
618 	req = lookup_generic_request(&monc->generic_request_tree, tid);
619 	if (!req) {
620 		dout("get_generic_reply %lld dne\n", tid);
621 		*skip = 1;
622 		m = NULL;
623 	} else {
624 		dout("get_generic_reply %lld got %p\n", tid, req->reply);
625 		*skip = 0;
626 		m = ceph_msg_get(req->reply);
627 		/*
628 		 * we don't need to track the connection reading into
629 		 * this reply because we only have one open connection
630 		 * at a time, ever.
631 		 */
632 	}
633 	mutex_unlock(&monc->mutex);
634 	return m;
635 }
636 
637 /*
638  * statfs
639  */
640 static void handle_statfs_reply(struct ceph_mon_client *monc,
641 				struct ceph_msg *msg)
642 {
643 	struct ceph_mon_generic_request *req;
644 	struct ceph_mon_statfs_reply *reply = msg->front.iov_base;
645 	u64 tid = le64_to_cpu(msg->hdr.tid);
646 
647 	dout("%s msg %p tid %llu\n", __func__, msg, tid);
648 
649 	if (msg->front.iov_len != sizeof(*reply))
650 		goto bad;
651 
652 	mutex_lock(&monc->mutex);
653 	req = lookup_generic_request(&monc->generic_request_tree, tid);
654 	if (!req) {
655 		mutex_unlock(&monc->mutex);
656 		return;
657 	}
658 
659 	req->result = 0;
660 	*req->u.st = reply->st; /* struct */
661 	__finish_generic_request(req);
662 	mutex_unlock(&monc->mutex);
663 
664 	complete_generic_request(req);
665 	return;
666 
667 bad:
668 	pr_err("corrupt statfs reply, tid %llu\n", tid);
669 	ceph_msg_dump(msg);
670 }
671 
672 /*
673  * Do a synchronous statfs().
674  */
675 int ceph_monc_do_statfs(struct ceph_mon_client *monc, struct ceph_statfs *buf)
676 {
677 	struct ceph_mon_generic_request *req;
678 	struct ceph_mon_statfs *h;
679 	int ret = -ENOMEM;
680 
681 	req = alloc_generic_request(monc, GFP_NOFS);
682 	if (!req)
683 		goto out;
684 
685 	req->request = ceph_msg_new(CEPH_MSG_STATFS, sizeof(*h), GFP_NOFS,
686 				    true);
687 	if (!req->request)
688 		goto out;
689 
690 	req->reply = ceph_msg_new(CEPH_MSG_STATFS_REPLY, 64, GFP_NOFS, true);
691 	if (!req->reply)
692 		goto out;
693 
694 	req->u.st = buf;
695 
696 	mutex_lock(&monc->mutex);
697 	register_generic_request(req);
698 	/* fill out request */
699 	h = req->request->front.iov_base;
700 	h->monhdr.have_version = 0;
701 	h->monhdr.session_mon = cpu_to_le16(-1);
702 	h->monhdr.session_mon_tid = 0;
703 	h->fsid = monc->monmap->fsid;
704 	send_generic_request(monc, req);
705 	mutex_unlock(&monc->mutex);
706 
707 	ret = wait_generic_request(req);
708 out:
709 	put_generic_request(req);
710 	return ret;
711 }
712 EXPORT_SYMBOL(ceph_monc_do_statfs);
713 
714 static void handle_get_version_reply(struct ceph_mon_client *monc,
715 				     struct ceph_msg *msg)
716 {
717 	struct ceph_mon_generic_request *req;
718 	u64 tid = le64_to_cpu(msg->hdr.tid);
719 	void *p = msg->front.iov_base;
720 	void *end = p + msg->front_alloc_len;
721 	u64 handle;
722 
723 	dout("%s msg %p tid %llu\n", __func__, msg, tid);
724 
725 	ceph_decode_need(&p, end, 2*sizeof(u64), bad);
726 	handle = ceph_decode_64(&p);
727 	if (tid != 0 && tid != handle)
728 		goto bad;
729 
730 	mutex_lock(&monc->mutex);
731 	req = lookup_generic_request(&monc->generic_request_tree, handle);
732 	if (!req) {
733 		mutex_unlock(&monc->mutex);
734 		return;
735 	}
736 
737 	req->result = 0;
738 	req->u.newest = ceph_decode_64(&p);
739 	__finish_generic_request(req);
740 	mutex_unlock(&monc->mutex);
741 
742 	complete_generic_request(req);
743 	return;
744 
745 bad:
746 	pr_err("corrupt mon_get_version reply, tid %llu\n", tid);
747 	ceph_msg_dump(msg);
748 }
749 
750 static struct ceph_mon_generic_request *
751 __ceph_monc_get_version(struct ceph_mon_client *monc, const char *what,
752 			ceph_monc_callback_t cb, u64 private_data)
753 {
754 	struct ceph_mon_generic_request *req;
755 
756 	req = alloc_generic_request(monc, GFP_NOIO);
757 	if (!req)
758 		goto err_put_req;
759 
760 	req->request = ceph_msg_new(CEPH_MSG_MON_GET_VERSION,
761 				    sizeof(u64) + sizeof(u32) + strlen(what),
762 				    GFP_NOIO, true);
763 	if (!req->request)
764 		goto err_put_req;
765 
766 	req->reply = ceph_msg_new(CEPH_MSG_MON_GET_VERSION_REPLY, 32, GFP_NOIO,
767 				  true);
768 	if (!req->reply)
769 		goto err_put_req;
770 
771 	req->complete_cb = cb;
772 	req->private_data = private_data;
773 
774 	mutex_lock(&monc->mutex);
775 	register_generic_request(req);
776 	{
777 		void *p = req->request->front.iov_base;
778 		void *const end = p + req->request->front_alloc_len;
779 
780 		ceph_encode_64(&p, req->tid); /* handle */
781 		ceph_encode_string(&p, end, what, strlen(what));
782 		WARN_ON(p != end);
783 	}
784 	send_generic_request(monc, req);
785 	mutex_unlock(&monc->mutex);
786 
787 	return req;
788 
789 err_put_req:
790 	put_generic_request(req);
791 	return ERR_PTR(-ENOMEM);
792 }
793 
794 /*
795  * Send MMonGetVersion and wait for the reply.
796  *
797  * @what: one of "mdsmap", "osdmap" or "monmap"
798  */
799 int ceph_monc_get_version(struct ceph_mon_client *monc, const char *what,
800 			  u64 *newest)
801 {
802 	struct ceph_mon_generic_request *req;
803 	int ret;
804 
805 	req = __ceph_monc_get_version(monc, what, NULL, 0);
806 	if (IS_ERR(req))
807 		return PTR_ERR(req);
808 
809 	ret = wait_generic_request(req);
810 	if (!ret)
811 		*newest = req->u.newest;
812 
813 	put_generic_request(req);
814 	return ret;
815 }
816 EXPORT_SYMBOL(ceph_monc_get_version);
817 
818 /*
819  * Send MMonGetVersion,
820  *
821  * @what: one of "mdsmap", "osdmap" or "monmap"
822  */
823 int ceph_monc_get_version_async(struct ceph_mon_client *monc, const char *what,
824 				ceph_monc_callback_t cb, u64 private_data)
825 {
826 	struct ceph_mon_generic_request *req;
827 
828 	req = __ceph_monc_get_version(monc, what, cb, private_data);
829 	if (IS_ERR(req))
830 		return PTR_ERR(req);
831 
832 	put_generic_request(req);
833 	return 0;
834 }
835 EXPORT_SYMBOL(ceph_monc_get_version_async);
836 
837 /*
838  * Resend pending generic requests.
839  */
840 static void __resend_generic_request(struct ceph_mon_client *monc)
841 {
842 	struct ceph_mon_generic_request *req;
843 	struct rb_node *p;
844 
845 	for (p = rb_first(&monc->generic_request_tree); p; p = rb_next(p)) {
846 		req = rb_entry(p, struct ceph_mon_generic_request, node);
847 		ceph_msg_revoke(req->request);
848 		ceph_msg_revoke_incoming(req->reply);
849 		ceph_con_send(&monc->con, ceph_msg_get(req->request));
850 	}
851 }
852 
853 /*
854  * Delayed work.  If we haven't mounted yet, retry.  Otherwise,
855  * renew/retry subscription as needed (in case it is timing out, or we
856  * got an ENOMEM).  And keep the monitor connection alive.
857  */
858 static void delayed_work(struct work_struct *work)
859 {
860 	struct ceph_mon_client *monc =
861 		container_of(work, struct ceph_mon_client, delayed_work.work);
862 
863 	dout("monc delayed_work\n");
864 	mutex_lock(&monc->mutex);
865 	if (monc->hunting) {
866 		dout("%s continuing hunt\n", __func__);
867 		reopen_session(monc);
868 	} else {
869 		int is_auth = ceph_auth_is_authenticated(monc->auth);
870 		if (ceph_con_keepalive_expired(&monc->con,
871 					       CEPH_MONC_PING_TIMEOUT)) {
872 			dout("monc keepalive timeout\n");
873 			is_auth = 0;
874 			reopen_session(monc);
875 		}
876 
877 		if (!monc->hunting) {
878 			ceph_con_keepalive(&monc->con);
879 			__validate_auth(monc);
880 		}
881 
882 		if (is_auth) {
883 			unsigned long now = jiffies;
884 
885 			dout("%s renew subs? now %lu renew after %lu\n",
886 			     __func__, now, monc->sub_renew_after);
887 			if (time_after_eq(now, monc->sub_renew_after))
888 				__send_subscribe(monc);
889 		}
890 	}
891 	__schedule_delayed(monc);
892 	mutex_unlock(&monc->mutex);
893 }
894 
895 /*
896  * On startup, we build a temporary monmap populated with the IPs
897  * provided by mount(2).
898  */
899 static int build_initial_monmap(struct ceph_mon_client *monc)
900 {
901 	struct ceph_options *opt = monc->client->options;
902 	struct ceph_entity_addr *mon_addr = opt->mon_addr;
903 	int num_mon = opt->num_mon;
904 	int i;
905 
906 	/* build initial monmap */
907 	monc->monmap = kzalloc(sizeof(*monc->monmap) +
908 			       num_mon*sizeof(monc->monmap->mon_inst[0]),
909 			       GFP_KERNEL);
910 	if (!monc->monmap)
911 		return -ENOMEM;
912 	for (i = 0; i < num_mon; i++) {
913 		monc->monmap->mon_inst[i].addr = mon_addr[i];
914 		monc->monmap->mon_inst[i].addr.nonce = 0;
915 		monc->monmap->mon_inst[i].name.type =
916 			CEPH_ENTITY_TYPE_MON;
917 		monc->monmap->mon_inst[i].name.num = cpu_to_le64(i);
918 	}
919 	monc->monmap->num_mon = num_mon;
920 	return 0;
921 }
922 
923 int ceph_monc_init(struct ceph_mon_client *monc, struct ceph_client *cl)
924 {
925 	int err = 0;
926 
927 	dout("init\n");
928 	memset(monc, 0, sizeof(*monc));
929 	monc->client = cl;
930 	monc->monmap = NULL;
931 	mutex_init(&monc->mutex);
932 
933 	err = build_initial_monmap(monc);
934 	if (err)
935 		goto out;
936 
937 	/* connection */
938 	/* authentication */
939 	monc->auth = ceph_auth_init(cl->options->name,
940 				    cl->options->key);
941 	if (IS_ERR(monc->auth)) {
942 		err = PTR_ERR(monc->auth);
943 		goto out_monmap;
944 	}
945 	monc->auth->want_keys =
946 		CEPH_ENTITY_TYPE_AUTH | CEPH_ENTITY_TYPE_MON |
947 		CEPH_ENTITY_TYPE_OSD | CEPH_ENTITY_TYPE_MDS;
948 
949 	/* msgs */
950 	err = -ENOMEM;
951 	monc->m_subscribe_ack = ceph_msg_new(CEPH_MSG_MON_SUBSCRIBE_ACK,
952 				     sizeof(struct ceph_mon_subscribe_ack),
953 				     GFP_NOFS, true);
954 	if (!monc->m_subscribe_ack)
955 		goto out_auth;
956 
957 	monc->m_subscribe = ceph_msg_new(CEPH_MSG_MON_SUBSCRIBE, 128, GFP_NOFS,
958 					 true);
959 	if (!monc->m_subscribe)
960 		goto out_subscribe_ack;
961 
962 	monc->m_auth_reply = ceph_msg_new(CEPH_MSG_AUTH_REPLY, 4096, GFP_NOFS,
963 					  true);
964 	if (!monc->m_auth_reply)
965 		goto out_subscribe;
966 
967 	monc->m_auth = ceph_msg_new(CEPH_MSG_AUTH, 4096, GFP_NOFS, true);
968 	monc->pending_auth = 0;
969 	if (!monc->m_auth)
970 		goto out_auth_reply;
971 
972 	ceph_con_init(&monc->con, monc, &mon_con_ops,
973 		      &monc->client->msgr);
974 
975 	monc->cur_mon = -1;
976 	monc->had_a_connection = false;
977 	monc->hunt_mult = 1;
978 
979 	INIT_DELAYED_WORK(&monc->delayed_work, delayed_work);
980 	monc->generic_request_tree = RB_ROOT;
981 	monc->last_tid = 0;
982 
983 	monc->fs_cluster_id = CEPH_FS_CLUSTER_ID_NONE;
984 
985 	return 0;
986 
987 out_auth_reply:
988 	ceph_msg_put(monc->m_auth_reply);
989 out_subscribe:
990 	ceph_msg_put(monc->m_subscribe);
991 out_subscribe_ack:
992 	ceph_msg_put(monc->m_subscribe_ack);
993 out_auth:
994 	ceph_auth_destroy(monc->auth);
995 out_monmap:
996 	kfree(monc->monmap);
997 out:
998 	return err;
999 }
1000 EXPORT_SYMBOL(ceph_monc_init);
1001 
1002 void ceph_monc_stop(struct ceph_mon_client *monc)
1003 {
1004 	dout("stop\n");
1005 	cancel_delayed_work_sync(&monc->delayed_work);
1006 
1007 	mutex_lock(&monc->mutex);
1008 	__close_session(monc);
1009 	monc->cur_mon = -1;
1010 	mutex_unlock(&monc->mutex);
1011 
1012 	/*
1013 	 * flush msgr queue before we destroy ourselves to ensure that:
1014 	 *  - any work that references our embedded con is finished.
1015 	 *  - any osd_client or other work that may reference an authorizer
1016 	 *    finishes before we shut down the auth subsystem.
1017 	 */
1018 	ceph_msgr_flush();
1019 
1020 	ceph_auth_destroy(monc->auth);
1021 
1022 	WARN_ON(!RB_EMPTY_ROOT(&monc->generic_request_tree));
1023 
1024 	ceph_msg_put(monc->m_auth);
1025 	ceph_msg_put(monc->m_auth_reply);
1026 	ceph_msg_put(monc->m_subscribe);
1027 	ceph_msg_put(monc->m_subscribe_ack);
1028 
1029 	kfree(monc->monmap);
1030 }
1031 EXPORT_SYMBOL(ceph_monc_stop);
1032 
1033 static void finish_hunting(struct ceph_mon_client *monc)
1034 {
1035 	if (monc->hunting) {
1036 		dout("%s found mon%d\n", __func__, monc->cur_mon);
1037 		monc->hunting = false;
1038 		monc->had_a_connection = true;
1039 		monc->hunt_mult /= 2; /* reduce by 50% */
1040 		if (monc->hunt_mult < 1)
1041 			monc->hunt_mult = 1;
1042 	}
1043 }
1044 
1045 static void handle_auth_reply(struct ceph_mon_client *monc,
1046 			      struct ceph_msg *msg)
1047 {
1048 	int ret;
1049 	int was_auth = 0;
1050 
1051 	mutex_lock(&monc->mutex);
1052 	was_auth = ceph_auth_is_authenticated(monc->auth);
1053 	monc->pending_auth = 0;
1054 	ret = ceph_handle_auth_reply(monc->auth, msg->front.iov_base,
1055 				     msg->front.iov_len,
1056 				     monc->m_auth->front.iov_base,
1057 				     monc->m_auth->front_alloc_len);
1058 	if (ret > 0) {
1059 		__send_prepared_auth_request(monc, ret);
1060 		goto out;
1061 	}
1062 
1063 	finish_hunting(monc);
1064 
1065 	if (ret < 0) {
1066 		monc->client->auth_err = ret;
1067 	} else if (!was_auth && ceph_auth_is_authenticated(monc->auth)) {
1068 		dout("authenticated, starting session\n");
1069 
1070 		monc->client->msgr.inst.name.type = CEPH_ENTITY_TYPE_CLIENT;
1071 		monc->client->msgr.inst.name.num =
1072 					cpu_to_le64(monc->auth->global_id);
1073 
1074 		__send_subscribe(monc);
1075 		__resend_generic_request(monc);
1076 
1077 		pr_info("mon%d %s session established\n", monc->cur_mon,
1078 			ceph_pr_addr(&monc->con.peer_addr.in_addr));
1079 	}
1080 
1081 out:
1082 	mutex_unlock(&monc->mutex);
1083 	if (monc->client->auth_err < 0)
1084 		wake_up_all(&monc->client->auth_wq);
1085 }
1086 
1087 static int __validate_auth(struct ceph_mon_client *monc)
1088 {
1089 	int ret;
1090 
1091 	if (monc->pending_auth)
1092 		return 0;
1093 
1094 	ret = ceph_build_auth(monc->auth, monc->m_auth->front.iov_base,
1095 			      monc->m_auth->front_alloc_len);
1096 	if (ret <= 0)
1097 		return ret; /* either an error, or no need to authenticate */
1098 	__send_prepared_auth_request(monc, ret);
1099 	return 0;
1100 }
1101 
1102 int ceph_monc_validate_auth(struct ceph_mon_client *monc)
1103 {
1104 	int ret;
1105 
1106 	mutex_lock(&monc->mutex);
1107 	ret = __validate_auth(monc);
1108 	mutex_unlock(&monc->mutex);
1109 	return ret;
1110 }
1111 EXPORT_SYMBOL(ceph_monc_validate_auth);
1112 
1113 /*
1114  * handle incoming message
1115  */
1116 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1117 {
1118 	struct ceph_mon_client *monc = con->private;
1119 	int type = le16_to_cpu(msg->hdr.type);
1120 
1121 	if (!monc)
1122 		return;
1123 
1124 	switch (type) {
1125 	case CEPH_MSG_AUTH_REPLY:
1126 		handle_auth_reply(monc, msg);
1127 		break;
1128 
1129 	case CEPH_MSG_MON_SUBSCRIBE_ACK:
1130 		handle_subscribe_ack(monc, msg);
1131 		break;
1132 
1133 	case CEPH_MSG_STATFS_REPLY:
1134 		handle_statfs_reply(monc, msg);
1135 		break;
1136 
1137 	case CEPH_MSG_MON_GET_VERSION_REPLY:
1138 		handle_get_version_reply(monc, msg);
1139 		break;
1140 
1141 	case CEPH_MSG_MON_MAP:
1142 		ceph_monc_handle_map(monc, msg);
1143 		break;
1144 
1145 	case CEPH_MSG_OSD_MAP:
1146 		ceph_osdc_handle_map(&monc->client->osdc, msg);
1147 		break;
1148 
1149 	default:
1150 		/* can the chained handler handle it? */
1151 		if (monc->client->extra_mon_dispatch &&
1152 		    monc->client->extra_mon_dispatch(monc->client, msg) == 0)
1153 			break;
1154 
1155 		pr_err("received unknown message type %d %s\n", type,
1156 		       ceph_msg_type_name(type));
1157 	}
1158 	ceph_msg_put(msg);
1159 }
1160 
1161 /*
1162  * Allocate memory for incoming message
1163  */
1164 static struct ceph_msg *mon_alloc_msg(struct ceph_connection *con,
1165 				      struct ceph_msg_header *hdr,
1166 				      int *skip)
1167 {
1168 	struct ceph_mon_client *monc = con->private;
1169 	int type = le16_to_cpu(hdr->type);
1170 	int front_len = le32_to_cpu(hdr->front_len);
1171 	struct ceph_msg *m = NULL;
1172 
1173 	*skip = 0;
1174 
1175 	switch (type) {
1176 	case CEPH_MSG_MON_SUBSCRIBE_ACK:
1177 		m = ceph_msg_get(monc->m_subscribe_ack);
1178 		break;
1179 	case CEPH_MSG_STATFS_REPLY:
1180 		return get_generic_reply(con, hdr, skip);
1181 	case CEPH_MSG_AUTH_REPLY:
1182 		m = ceph_msg_get(monc->m_auth_reply);
1183 		break;
1184 	case CEPH_MSG_MON_GET_VERSION_REPLY:
1185 		if (le64_to_cpu(hdr->tid) != 0)
1186 			return get_generic_reply(con, hdr, skip);
1187 
1188 		/*
1189 		 * Older OSDs don't set reply tid even if the orignal
1190 		 * request had a non-zero tid.  Workaround this weirdness
1191 		 * by falling through to the allocate case.
1192 		 */
1193 	case CEPH_MSG_MON_MAP:
1194 	case CEPH_MSG_MDS_MAP:
1195 	case CEPH_MSG_OSD_MAP:
1196 		m = ceph_msg_new(type, front_len, GFP_NOFS, false);
1197 		if (!m)
1198 			return NULL;	/* ENOMEM--return skip == 0 */
1199 		break;
1200 	}
1201 
1202 	if (!m) {
1203 		pr_info("alloc_msg unknown type %d\n", type);
1204 		*skip = 1;
1205 	} else if (front_len > m->front_alloc_len) {
1206 		pr_warn("mon_alloc_msg front %d > prealloc %d (%u#%llu)\n",
1207 			front_len, m->front_alloc_len,
1208 			(unsigned int)con->peer_name.type,
1209 			le64_to_cpu(con->peer_name.num));
1210 		ceph_msg_put(m);
1211 		m = ceph_msg_new(type, front_len, GFP_NOFS, false);
1212 	}
1213 
1214 	return m;
1215 }
1216 
1217 /*
1218  * If the monitor connection resets, pick a new monitor and resubmit
1219  * any pending requests.
1220  */
1221 static void mon_fault(struct ceph_connection *con)
1222 {
1223 	struct ceph_mon_client *monc = con->private;
1224 
1225 	mutex_lock(&monc->mutex);
1226 	dout("%s mon%d\n", __func__, monc->cur_mon);
1227 	if (monc->cur_mon >= 0) {
1228 		if (!monc->hunting) {
1229 			dout("%s hunting for new mon\n", __func__);
1230 			reopen_session(monc);
1231 			__schedule_delayed(monc);
1232 		} else {
1233 			dout("%s already hunting\n", __func__);
1234 		}
1235 	}
1236 	mutex_unlock(&monc->mutex);
1237 }
1238 
1239 /*
1240  * We can ignore refcounting on the connection struct, as all references
1241  * will come from the messenger workqueue, which is drained prior to
1242  * mon_client destruction.
1243  */
1244 static struct ceph_connection *con_get(struct ceph_connection *con)
1245 {
1246 	return con;
1247 }
1248 
1249 static void con_put(struct ceph_connection *con)
1250 {
1251 }
1252 
1253 static const struct ceph_connection_operations mon_con_ops = {
1254 	.get = con_get,
1255 	.put = con_put,
1256 	.dispatch = dispatch,
1257 	.fault = mon_fault,
1258 	.alloc_msg = mon_alloc_msg,
1259 };
1260