blob: cf638c009cfabe9e95d114c1e187785db46f958c [file] [log] [blame]
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07001#include <linux/ceph/ceph_debug.h>
Sage Weilba75bb92009-10-06 11:31:11 -07002
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07003#include <linux/module.h>
Sage Weilba75bb92009-10-06 11:31:11 -07004#include <linux/types.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +09005#include <linux/slab.h>
Sage Weilba75bb92009-10-06 11:31:11 -07006#include <linux/random.h>
7#include <linux/sched.h>
8
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07009#include <linux/ceph/mon_client.h>
10#include <linux/ceph/libceph.h>
Sage Weilab434b62012-01-13 22:22:03 -080011#include <linux/ceph/debugfs.h>
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -070012#include <linux/ceph/decode.h>
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -070013#include <linux/ceph/auth.h>
Sage Weilba75bb92009-10-06 11:31:11 -070014
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
Tobias Klauser9e327892010-05-20 10:40:19 +020033static const struct ceph_connection_operations mon_con_ops;
Sage Weilba75bb92009-10-06 11:31:11 -070034
Sage Weil9bd2e6f2010-02-02 16:21:06 -080035static int __validate_auth(struct ceph_mon_client *monc);
36
Sage Weilba75bb92009-10-06 11:31:11 -070037/*
38 * Decode a monmap blob (e.g., during mount).
39 */
40struct 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;
Sage Weil4e7a5dc2009-11-18 16:19:57 -080047 u32 len;
48
49 ceph_decode_32_safe(&p, end, len, bad);
50 ceph_decode_need(&p, end, len, bad);
Sage Weilba75bb92009-10-06 11:31:11 -070051
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));
Sage Weilc89136e2009-10-14 09:59:09 -070058 epoch = ceph_decode_32(&p);
Sage Weilba75bb92009-10-06 11:31:11 -070059
Sage Weilc89136e2009-10-14 09:59:09 -070060 num_mon = ceph_decode_32(&p);
Sage Weilba75bb92009-10-06 11:31:11 -070061 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]));
Sage Weil63f2d212009-11-03 15:17:56 -080072 for (i = 0; i < num_mon; i++)
73 ceph_decode_addr(&m->mon_inst[i].addr);
Sage Weilba75bb92009-10-06 11:31:11 -070074
Sage Weilba75bb92009-10-06 11:31:11 -070075 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,
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -070079 ceph_pr_addr(&m->mon_inst[i].addr.in_addr));
Sage Weilba75bb92009-10-06 11:31:11 -070080 return m;
81
82bad:
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 */
91int 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++)
Sage Weil103e2d32010-01-07 16:12:36 -080096 if (memcmp(addr, &m->mon_inst[i].addr, sizeof(*addr)) == 0)
Sage Weilba75bb92009-10-06 11:31:11 -070097 return 1;
98 return 0;
99}
100
101/*
Sage Weil5ce6e9d2010-02-15 16:22:28 -0800102 * Send an auth request.
103 */
104static 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);
Alex Elder6740a842012-06-01 14:56:43 -0500109 ceph_msg_revoke(monc->m_auth);
Sage Weil5ce6e9d2010-02-15 16:22:28 -0800110 ceph_msg_get(monc->m_auth); /* keep our ref */
Alex Elder67130932012-05-26 23:26:43 -0500111 ceph_con_send(&monc->con, monc->m_auth);
Sage Weil5ce6e9d2010-02-15 16:22:28 -0800112}
113
114/*
Sage Weilba75bb92009-10-06 11:31:11 -0700115 * Close monitor session, if any.
116 */
117static void __close_session(struct ceph_mon_client *monc)
118{
Sage Weilf6a2f5b2011-08-09 09:27:44 -0700119 dout("__close_session closing mon%d\n", monc->cur_mon);
Alex Elder6740a842012-06-01 14:56:43 -0500120 ceph_msg_revoke(monc->m_auth);
Sage Weil4f471e42012-07-30 18:16:40 -0700121 ceph_msg_revoke_incoming(monc->m_auth_reply);
122 ceph_msg_revoke(monc->m_subscribe);
123 ceph_msg_revoke_incoming(monc->m_subscribe_ack);
Alex Elder67130932012-05-26 23:26:43 -0500124 ceph_con_close(&monc->con);
Ilya Dryomov0e04dc22016-01-20 17:50:31 +0100125
Sage Weilf6a2f5b2011-08-09 09:27:44 -0700126 monc->pending_auth = 0;
127 ceph_auth_reset(monc->auth);
Sage Weilba75bb92009-10-06 11:31:11 -0700128}
129
130/*
Ilya Dryomov0e04dc22016-01-20 17:50:31 +0100131 * 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.
Sage Weilba75bb92009-10-06 11:31:11 -0700133 */
Ilya Dryomov0e04dc22016-01-20 17:50:31 +0100134static void pick_new_mon(struct ceph_mon_client *monc)
Sage Weilba75bb92009-10-06 11:31:11 -0700135{
Ilya Dryomov0e04dc22016-01-20 17:50:31 +0100136 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 */
168static void __open_session(struct ceph_mon_client *monc)
169{
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800170 int ret;
Sage Weilba75bb92009-10-06 11:31:11 -0700171
Ilya Dryomov0e04dc22016-01-20 17:50:31 +0100172 pick_new_mon(monc);
Sage Weilba75bb92009-10-06 11:31:11 -0700173
Ilya Dryomov1752b502016-01-21 16:33:25 +0100174 monc->hunting = true;
Ilya Dryomov168b9092016-01-21 16:33:19 +0100175 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
Ilya Dryomov0e04dc22016-01-20 17:50:31 +0100181 monc->sub_renew_after = jiffies; /* i.e., expired */
182 monc->sub_renew_sent = 0;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800183
Ilya Dryomov0e04dc22016-01-20 17:50:31 +0100184 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);
Yan, Zheng8b9558a2015-09-01 17:19:38 +0800187
Ilya Dryomov0e04dc22016-01-20 17:50:31 +0100188 /*
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);
Sage Weilba75bb92009-10-06 11:31:11 -0700200}
201
Ilya Dryomov1752b502016-01-21 16:33:25 +0100202static 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
Sage Weilba75bb92009-10-06 11:31:11 -0700212/*
213 * Reschedule delayed work timer.
214 */
215static void __schedule_delayed(struct ceph_mon_client *monc)
216{
Yan, Zheng8b9558a2015-09-01 17:19:38 +0800217 unsigned long delay;
Sage Weilba75bb92009-10-06 11:31:11 -0700218
Ilya Dryomov168b9092016-01-21 16:33:19 +0100219 if (monc->hunting)
220 delay = CEPH_MONC_HUNT_INTERVAL * monc->hunt_mult;
221 else
Ilya Dryomov58d81b12016-01-21 16:33:15 +0100222 delay = CEPH_MONC_PING_INTERVAL;
Ilya Dryomov168b9092016-01-21 16:33:19 +0100223
Yan, Zheng8b9558a2015-09-01 17:19:38 +0800224 dout("__schedule_delayed after %lu\n", delay);
Ilya Dryomovbee3a372016-01-22 18:42:17 +0100225 mod_delayed_work(system_wq, &monc->delayed_work,
226 round_jiffies_relative(delay));
Sage Weilba75bb92009-10-06 11:31:11 -0700227}
228
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100229const char *ceph_sub_str[] = {
230 [CEPH_SUB_MDSMAP] = "mdsmap",
231 [CEPH_SUB_MONMAP] = "monmap",
232 [CEPH_SUB_OSDMAP] = "osdmap",
233};
234
Sage Weilba75bb92009-10-06 11:31:11 -0700235/*
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100236 * Send subscribe request for one or more maps, according to
237 * monc->subs.
Sage Weilba75bb92009-10-06 11:31:11 -0700238 */
239static void __send_subscribe(struct ceph_mon_client *monc)
240{
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100241 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;
Sage Weilba75bb92009-10-06 11:31:11 -0700246
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100247 dout("%s sent %lu\n", __func__, monc->sub_renew_sent);
Sage Weilba75bb92009-10-06 11:31:11 -0700248
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100249 BUG_ON(monc->cur_mon < 0);
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700250
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100251 if (!monc->sub_renew_sent)
252 monc->sub_renew_sent = jiffies | 1; /* never 0 */
Sage Weilba75bb92009-10-06 11:31:11 -0700253
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100254 msg->hdr.version = cpu_to_le16(2);
Sage Weilba75bb92009-10-06 11:31:11 -0700255
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100256 for (i = 0; i < ARRAY_SIZE(monc->subs); i++) {
257 if (monc->subs[i].want)
258 num++;
Sage Weilba75bb92009-10-06 11:31:11 -0700259 }
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100260 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 const char *s = ceph_sub_str[i];
264
265 if (!monc->subs[i].want)
266 continue;
267
268 dout("%s %s start %llu flags 0x%x\n", __func__, s,
269 le64_to_cpu(monc->subs[i].item.start),
270 monc->subs[i].item.flags);
271 ceph_encode_string(&p, end, s, strlen(s));
272 memcpy(p, &monc->subs[i].item, sizeof(monc->subs[i].item));
273 p += sizeof(monc->subs[i].item);
274 }
275
276 BUG_ON(p != (end - 35 - (ARRAY_SIZE(monc->subs) - num) * 19));
277 msg->front.iov_len = p - msg->front.iov_base;
278 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
279 ceph_msg_revoke(msg);
280 ceph_con_send(&monc->con, ceph_msg_get(msg));
Sage Weilba75bb92009-10-06 11:31:11 -0700281}
282
283static void handle_subscribe_ack(struct ceph_mon_client *monc,
284 struct ceph_msg *msg)
285{
Eric Dumazet95c96172012-04-15 05:58:06 +0000286 unsigned int seconds;
Sage Weil07bd10f2009-10-14 17:26:40 -0700287 struct ceph_mon_subscribe_ack *h = msg->front.iov_base;
Sage Weilba75bb92009-10-06 11:31:11 -0700288
Sage Weil07bd10f2009-10-14 17:26:40 -0700289 if (msg->front.iov_len < sizeof(*h))
290 goto bad;
291 seconds = le32_to_cpu(h->duration);
292
Sage Weilba75bb92009-10-06 11:31:11 -0700293 mutex_lock(&monc->mutex);
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100294 if (monc->sub_renew_sent) {
295 monc->sub_renew_after = monc->sub_renew_sent +
296 (seconds >> 1) * HZ - 1;
297 dout("%s sent %lu duration %d renew after %lu\n", __func__,
298 monc->sub_renew_sent, seconds, monc->sub_renew_after);
299 monc->sub_renew_sent = 0;
300 } else {
301 dout("%s sent %lu renew after %lu, ignoring\n", __func__,
302 monc->sub_renew_sent, monc->sub_renew_after);
303 }
Sage Weilba75bb92009-10-06 11:31:11 -0700304 mutex_unlock(&monc->mutex);
305 return;
306bad:
307 pr_err("got corrupt subscribe-ack msg\n");
Sage Weil9ec7cab2009-12-14 15:13:47 -0800308 ceph_msg_dump(msg);
Sage Weilba75bb92009-10-06 11:31:11 -0700309}
310
311/*
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100312 * Register interest in a map
313 *
314 * @sub: one of CEPH_SUB_*
315 * @epoch: X for "every map since X", or 0 for "just the latest"
Sage Weilba75bb92009-10-06 11:31:11 -0700316 */
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100317static bool __ceph_monc_want_map(struct ceph_mon_client *monc, int sub,
318 u32 epoch, bool continuous)
Sage Weilba75bb92009-10-06 11:31:11 -0700319{
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100320 __le64 start = cpu_to_le64(epoch);
321 u8 flags = !continuous ? CEPH_SUBSCRIBE_ONETIME : 0;
Sage Weilba75bb92009-10-06 11:31:11 -0700322
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100323 dout("%s %s epoch %u continuous %d\n", __func__, ceph_sub_str[sub],
324 epoch, continuous);
325
326 if (monc->subs[sub].want &&
327 monc->subs[sub].item.start == start &&
328 monc->subs[sub].item.flags == flags)
329 return false;
330
331 monc->subs[sub].item.start = start;
332 monc->subs[sub].item.flags = flags;
333 monc->subs[sub].want = true;
334
335 return true;
336}
337
338bool ceph_monc_want_map(struct ceph_mon_client *monc, int sub, u32 epoch,
339 bool continuous)
340{
341 bool need_request;
342
343 mutex_lock(&monc->mutex);
344 need_request = __ceph_monc_want_map(monc, sub, epoch, continuous);
345 mutex_unlock(&monc->mutex);
346
347 return need_request;
348}
349EXPORT_SYMBOL(ceph_monc_want_map);
350
351/*
352 * Keep track of which maps we have
353 *
354 * @sub: one of CEPH_SUB_*
355 */
356static void __ceph_monc_got_map(struct ceph_mon_client *monc, int sub,
357 u32 epoch)
358{
359 dout("%s %s epoch %u\n", __func__, ceph_sub_str[sub], epoch);
360
361 if (monc->subs[sub].want) {
362 if (monc->subs[sub].item.flags & CEPH_SUBSCRIBE_ONETIME)
363 monc->subs[sub].want = false;
364 else
365 monc->subs[sub].item.start = cpu_to_le64(epoch + 1);
366 }
367
368 monc->subs[sub].have = epoch;
369}
370
371void ceph_monc_got_map(struct ceph_mon_client *monc, int sub, u32 epoch)
Sage Weilba75bb92009-10-06 11:31:11 -0700372{
373 mutex_lock(&monc->mutex);
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100374 __ceph_monc_got_map(monc, sub, epoch);
Sage Weilba75bb92009-10-06 11:31:11 -0700375 mutex_unlock(&monc->mutex);
Sage Weilba75bb92009-10-06 11:31:11 -0700376}
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100377EXPORT_SYMBOL(ceph_monc_got_map);
Sage Weilba75bb92009-10-06 11:31:11 -0700378
379/*
380 * Register interest in the next osdmap
381 */
382void ceph_monc_request_next_osdmap(struct ceph_mon_client *monc)
383{
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100384 dout("%s have %u\n", __func__, monc->subs[CEPH_SUB_OSDMAP].have);
Sage Weilba75bb92009-10-06 11:31:11 -0700385 mutex_lock(&monc->mutex);
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100386 if (__ceph_monc_want_map(monc, CEPH_SUB_OSDMAP,
387 monc->subs[CEPH_SUB_OSDMAP].have + 1, false))
Sage Weilba75bb92009-10-06 11:31:11 -0700388 __send_subscribe(monc);
389 mutex_unlock(&monc->mutex);
390}
Ilya Dryomov6044cde2014-05-13 11:19:27 +0400391EXPORT_SYMBOL(ceph_monc_request_next_osdmap);
392
Ilya Dryomova319bf52015-05-15 12:02:17 +0300393/*
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 */
Ilya Dryomov6044cde2014-05-13 11:19:27 +0400399int ceph_monc_wait_osdmap(struct ceph_mon_client *monc, u32 epoch,
400 unsigned long timeout)
401{
402 unsigned long started = jiffies;
Ilya Dryomov216639d2015-05-19 12:03:33 +0300403 long ret;
Ilya Dryomov6044cde2014-05-13 11:19:27 +0400404
405 mutex_lock(&monc->mutex);
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100406 while (monc->subs[CEPH_SUB_OSDMAP].have < epoch) {
Ilya Dryomov6044cde2014-05-13 11:19:27 +0400407 mutex_unlock(&monc->mutex);
408
Ilya Dryomova319bf52015-05-15 12:02:17 +0300409 if (timeout && time_after_eq(jiffies, started + timeout))
Ilya Dryomov6044cde2014-05-13 11:19:27 +0400410 return -ETIMEDOUT;
411
412 ret = wait_event_interruptible_timeout(monc->client->auth_wq,
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100413 monc->subs[CEPH_SUB_OSDMAP].have >= epoch,
414 ceph_timeout_jiffies(timeout));
Ilya Dryomov6044cde2014-05-13 11:19:27 +0400415 if (ret < 0)
416 return ret;
417
418 mutex_lock(&monc->mutex);
419 }
420
421 mutex_unlock(&monc->mutex);
422 return 0;
423}
424EXPORT_SYMBOL(ceph_monc_wait_osdmap);
Sage Weilba75bb92009-10-06 11:31:11 -0700425
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800426/*
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100427 * Open a session with a random monitor. Request monmap and osdmap,
428 * which are waited upon in __ceph_open_session().
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800429 */
430int ceph_monc_open_session(struct ceph_mon_client *monc)
Sage Weilba75bb92009-10-06 11:31:11 -0700431{
Sage Weilba75bb92009-10-06 11:31:11 -0700432 mutex_lock(&monc->mutex);
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100433 __ceph_monc_want_map(monc, CEPH_SUB_MONMAP, 0, true);
434 __ceph_monc_want_map(monc, CEPH_SUB_OSDMAP, 0, false);
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800435 __open_session(monc);
Sage Weilba75bb92009-10-06 11:31:11 -0700436 __schedule_delayed(monc);
437 mutex_unlock(&monc->mutex);
438 return 0;
439}
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700440EXPORT_SYMBOL(ceph_monc_open_session);
Sage Weilba75bb92009-10-06 11:31:11 -0700441
Sage Weil07433042009-11-18 16:50:41 -0800442static void ceph_monc_handle_map(struct ceph_mon_client *monc,
443 struct ceph_msg *msg)
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800444{
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));
Sage Weild4a780c2009-12-11 08:55:23 -0800459 goto out;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800460 }
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800461
Sage Weil07433042009-11-18 16:50:41 -0800462 if (ceph_check_fsid(monc->client, &monmap->fsid) < 0) {
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800463 kfree(monmap);
Sage Weild4a780c2009-12-11 08:55:23 -0800464 goto out;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800465 }
466
467 client->monc.monmap = monmap;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800468 kfree(old);
469
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100470 __ceph_monc_got_map(monc, CEPH_SUB_MONMAP, monc->monmap->epoch);
Ilya Dryomov02ac9562016-01-06 12:56:21 +0300471 client->have_fsid = true;
Sage Weild1c338a2012-08-19 12:29:16 -0700472
Sage Weild4a780c2009-12-11 08:55:23 -0800473out:
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800474 mutex_unlock(&monc->mutex);
Yehuda Sadeh03066f22010-07-27 13:11:08 -0700475 wake_up_all(&client->auth_wq);
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800476}
Sage Weilba75bb92009-10-06 11:31:11 -0700477
Sage Weilba75bb92009-10-06 11:31:11 -0700478/*
Ilya Dryomov7a6fdeb2014-12-22 19:14:26 +0300479 * generic requests (currently statfs, mon_get_version)
Sage Weilba75bb92009-10-06 11:31:11 -0700480 */
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700481static struct ceph_mon_generic_request *__lookup_generic_req(
Sage Weil85ff03f2010-02-15 14:47:28 -0800482 struct ceph_mon_client *monc, u64 tid)
483{
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700484 struct ceph_mon_generic_request *req;
485 struct rb_node *n = monc->generic_request_tree.rb_node;
Sage Weil85ff03f2010-02-15 14:47:28 -0800486
487 while (n) {
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700488 req = rb_entry(n, struct ceph_mon_generic_request, node);
Sage Weil85ff03f2010-02-15 14:47:28 -0800489 if (tid < req->tid)
490 n = n->rb_left;
491 else if (tid > req->tid)
492 n = n->rb_right;
493 else
494 return req;
495 }
496 return NULL;
497}
498
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700499static void __insert_generic_request(struct ceph_mon_client *monc,
500 struct ceph_mon_generic_request *new)
Sage Weil85ff03f2010-02-15 14:47:28 -0800501{
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700502 struct rb_node **p = &monc->generic_request_tree.rb_node;
Sage Weil85ff03f2010-02-15 14:47:28 -0800503 struct rb_node *parent = NULL;
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700504 struct ceph_mon_generic_request *req = NULL;
Sage Weil85ff03f2010-02-15 14:47:28 -0800505
506 while (*p) {
507 parent = *p;
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700508 req = rb_entry(parent, struct ceph_mon_generic_request, node);
Sage Weil85ff03f2010-02-15 14:47:28 -0800509 if (new->tid < req->tid)
510 p = &(*p)->rb_left;
511 else if (new->tid > req->tid)
512 p = &(*p)->rb_right;
513 else
514 BUG();
515 }
516
517 rb_link_node(&new->node, parent, p);
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700518 rb_insert_color(&new->node, &monc->generic_request_tree);
Sage Weil85ff03f2010-02-15 14:47:28 -0800519}
520
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700521static void release_generic_request(struct kref *kref)
Sage Weil3143edd2010-03-24 21:43:33 -0700522{
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700523 struct ceph_mon_generic_request *req =
524 container_of(kref, struct ceph_mon_generic_request, kref);
Sage Weil3143edd2010-03-24 21:43:33 -0700525
526 if (req->reply)
527 ceph_msg_put(req->reply);
528 if (req->request)
529 ceph_msg_put(req->request);
Yehuda Sadeh20547562010-06-01 10:37:40 -0700530
531 kfree(req);
Sage Weil3143edd2010-03-24 21:43:33 -0700532}
533
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700534static void put_generic_request(struct ceph_mon_generic_request *req)
Sage Weil3143edd2010-03-24 21:43:33 -0700535{
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700536 kref_put(&req->kref, release_generic_request);
Sage Weil3143edd2010-03-24 21:43:33 -0700537}
538
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700539static void get_generic_request(struct ceph_mon_generic_request *req)
Sage Weil3143edd2010-03-24 21:43:33 -0700540{
541 kref_get(&req->kref);
542}
543
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700544static struct ceph_msg *get_generic_reply(struct ceph_connection *con,
Sage Weil3143edd2010-03-24 21:43:33 -0700545 struct ceph_msg_header *hdr,
546 int *skip)
547{
548 struct ceph_mon_client *monc = con->private;
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700549 struct ceph_mon_generic_request *req;
Sage Weil3143edd2010-03-24 21:43:33 -0700550 u64 tid = le64_to_cpu(hdr->tid);
551 struct ceph_msg *m;
552
553 mutex_lock(&monc->mutex);
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700554 req = __lookup_generic_req(monc, tid);
Sage Weil3143edd2010-03-24 21:43:33 -0700555 if (!req) {
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700556 dout("get_generic_reply %lld dne\n", tid);
Sage Weil3143edd2010-03-24 21:43:33 -0700557 *skip = 1;
558 m = NULL;
559 } else {
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700560 dout("get_generic_reply %lld got %p\n", tid, req->reply);
Alex Elder1c20f2d2012-06-04 14:43:32 -0500561 *skip = 0;
Sage Weil3143edd2010-03-24 21:43:33 -0700562 m = ceph_msg_get(req->reply);
563 /*
564 * we don't need to track the connection reading into
565 * this reply because we only have one open connection
566 * at a time, ever.
567 */
568 }
569 mutex_unlock(&monc->mutex);
570 return m;
571}
572
Ilya Dryomov513a8242014-05-13 11:19:26 +0400573static int __do_generic_request(struct ceph_mon_client *monc, u64 tid,
574 struct ceph_mon_generic_request *req)
Yehuda Sadehe56fa102010-05-17 12:40:28 -0700575{
576 int err;
577
578 /* register request */
Ilya Dryomov513a8242014-05-13 11:19:26 +0400579 req->tid = tid != 0 ? tid : ++monc->last_tid;
Yehuda Sadehe56fa102010-05-17 12:40:28 -0700580 req->request->hdr.tid = cpu_to_le64(req->tid);
581 __insert_generic_request(monc, req);
582 monc->num_generic_requests++;
Alex Elder67130932012-05-26 23:26:43 -0500583 ceph_con_send(&monc->con, ceph_msg_get(req->request));
Yehuda Sadehe56fa102010-05-17 12:40:28 -0700584 mutex_unlock(&monc->mutex);
585
586 err = wait_for_completion_interruptible(&req->completion);
587
588 mutex_lock(&monc->mutex);
589 rb_erase(&req->node, &monc->generic_request_tree);
590 monc->num_generic_requests--;
Yehuda Sadehe56fa102010-05-17 12:40:28 -0700591
592 if (!err)
593 err = req->result;
594 return err;
595}
596
Ilya Dryomov513a8242014-05-13 11:19:26 +0400597static int do_generic_request(struct ceph_mon_client *monc,
598 struct ceph_mon_generic_request *req)
599{
600 int err;
601
602 mutex_lock(&monc->mutex);
603 err = __do_generic_request(monc, 0, req);
604 mutex_unlock(&monc->mutex);
605
606 return err;
607}
608
Yehuda Sadehe56fa102010-05-17 12:40:28 -0700609/*
610 * statfs
611 */
Sage Weilba75bb92009-10-06 11:31:11 -0700612static void handle_statfs_reply(struct ceph_mon_client *monc,
613 struct ceph_msg *msg)
614{
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700615 struct ceph_mon_generic_request *req;
Sage Weilba75bb92009-10-06 11:31:11 -0700616 struct ceph_mon_statfs_reply *reply = msg->front.iov_base;
Sage Weil3143edd2010-03-24 21:43:33 -0700617 u64 tid = le64_to_cpu(msg->hdr.tid);
Sage Weilba75bb92009-10-06 11:31:11 -0700618
619 if (msg->front.iov_len != sizeof(*reply))
620 goto bad;
Sage Weilba75bb92009-10-06 11:31:11 -0700621 dout("handle_statfs_reply %p tid %llu\n", msg, tid);
622
623 mutex_lock(&monc->mutex);
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700624 req = __lookup_generic_req(monc, tid);
Sage Weilba75bb92009-10-06 11:31:11 -0700625 if (req) {
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700626 *(struct ceph_statfs *)req->buf = reply->st;
Sage Weilba75bb92009-10-06 11:31:11 -0700627 req->result = 0;
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700628 get_generic_request(req);
Sage Weilba75bb92009-10-06 11:31:11 -0700629 }
630 mutex_unlock(&monc->mutex);
Sage Weil3143edd2010-03-24 21:43:33 -0700631 if (req) {
Yehuda Sadeh03066f22010-07-27 13:11:08 -0700632 complete_all(&req->completion);
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700633 put_generic_request(req);
Sage Weil3143edd2010-03-24 21:43:33 -0700634 }
Sage Weilba75bb92009-10-06 11:31:11 -0700635 return;
636
637bad:
Ilya Dryomov7a6fdeb2014-12-22 19:14:26 +0300638 pr_err("corrupt statfs reply, tid %llu\n", tid);
Sage Weil9ec7cab2009-12-14 15:13:47 -0800639 ceph_msg_dump(msg);
Sage Weilba75bb92009-10-06 11:31:11 -0700640}
641
642/*
Sage Weilba75bb92009-10-06 11:31:11 -0700643 * Do a synchronous statfs().
644 */
645int ceph_monc_do_statfs(struct ceph_mon_client *monc, struct ceph_statfs *buf)
646{
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700647 struct ceph_mon_generic_request *req;
Sage Weil3143edd2010-03-24 21:43:33 -0700648 struct ceph_mon_statfs *h;
Sage Weilba75bb92009-10-06 11:31:11 -0700649 int err;
650
Julia Lawallcffe7b62010-05-13 22:07:29 +0200651 req = kzalloc(sizeof(*req), GFP_NOFS);
Sage Weil3143edd2010-03-24 21:43:33 -0700652 if (!req)
653 return -ENOMEM;
Sage Weilba75bb92009-10-06 11:31:11 -0700654
Sage Weil3143edd2010-03-24 21:43:33 -0700655 kref_init(&req->kref);
656 req->buf = buf;
657 init_completion(&req->completion);
658
Sage Weila79832f2010-04-01 16:06:19 -0700659 err = -ENOMEM;
Sage Weilb61c2762011-08-09 15:03:46 -0700660 req->request = ceph_msg_new(CEPH_MSG_STATFS, sizeof(*h), GFP_NOFS,
661 true);
Sage Weila79832f2010-04-01 16:06:19 -0700662 if (!req->request)
Sage Weil3143edd2010-03-24 21:43:33 -0700663 goto out;
Sage Weilb61c2762011-08-09 15:03:46 -0700664 req->reply = ceph_msg_new(CEPH_MSG_STATFS_REPLY, 1024, GFP_NOFS,
665 true);
Sage Weila79832f2010-04-01 16:06:19 -0700666 if (!req->reply)
Sage Weil3143edd2010-03-24 21:43:33 -0700667 goto out;
Sage Weil3143edd2010-03-24 21:43:33 -0700668
669 /* fill out request */
670 h = req->request->front.iov_base;
671 h->monhdr.have_version = 0;
672 h->monhdr.session_mon = cpu_to_le16(-1);
673 h->monhdr.session_mon_tid = 0;
674 h->fsid = monc->monmap->fsid;
Sage Weilba75bb92009-10-06 11:31:11 -0700675
Yehuda Sadehe56fa102010-05-17 12:40:28 -0700676 err = do_generic_request(monc, req);
Sage Weil3143edd2010-03-24 21:43:33 -0700677
678out:
Ilya Dryomovf6469122014-12-22 19:35:17 +0300679 put_generic_request(req);
Sage Weilba75bb92009-10-06 11:31:11 -0700680 return err;
681}
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700682EXPORT_SYMBOL(ceph_monc_do_statfs);
Sage Weilba75bb92009-10-06 11:31:11 -0700683
Ilya Dryomov513a8242014-05-13 11:19:26 +0400684static void handle_get_version_reply(struct ceph_mon_client *monc,
685 struct ceph_msg *msg)
686{
687 struct ceph_mon_generic_request *req;
688 u64 tid = le64_to_cpu(msg->hdr.tid);
689 void *p = msg->front.iov_base;
690 void *end = p + msg->front_alloc_len;
691 u64 handle;
692
693 dout("%s %p tid %llu\n", __func__, msg, tid);
694
695 ceph_decode_need(&p, end, 2*sizeof(u64), bad);
696 handle = ceph_decode_64(&p);
697 if (tid != 0 && tid != handle)
698 goto bad;
699
700 mutex_lock(&monc->mutex);
701 req = __lookup_generic_req(monc, handle);
702 if (req) {
703 *(u64 *)req->buf = ceph_decode_64(&p);
704 req->result = 0;
705 get_generic_request(req);
706 }
707 mutex_unlock(&monc->mutex);
708 if (req) {
709 complete_all(&req->completion);
710 put_generic_request(req);
711 }
712
713 return;
714bad:
Ilya Dryomov7a6fdeb2014-12-22 19:14:26 +0300715 pr_err("corrupt mon_get_version reply, tid %llu\n", tid);
Ilya Dryomov513a8242014-05-13 11:19:26 +0400716 ceph_msg_dump(msg);
717}
718
719/*
720 * Send MMonGetVersion and wait for the reply.
721 *
722 * @what: one of "mdsmap", "osdmap" or "monmap"
723 */
724int ceph_monc_do_get_version(struct ceph_mon_client *monc, const char *what,
725 u64 *newest)
726{
727 struct ceph_mon_generic_request *req;
728 void *p, *end;
729 u64 tid;
730 int err;
731
732 req = kzalloc(sizeof(*req), GFP_NOFS);
733 if (!req)
734 return -ENOMEM;
735
736 kref_init(&req->kref);
737 req->buf = newest;
Ilya Dryomov513a8242014-05-13 11:19:26 +0400738 init_completion(&req->completion);
739
740 req->request = ceph_msg_new(CEPH_MSG_MON_GET_VERSION,
741 sizeof(u64) + sizeof(u32) + strlen(what),
742 GFP_NOFS, true);
743 if (!req->request) {
744 err = -ENOMEM;
745 goto out;
746 }
747
748 req->reply = ceph_msg_new(CEPH_MSG_MON_GET_VERSION_REPLY, 1024,
749 GFP_NOFS, true);
750 if (!req->reply) {
751 err = -ENOMEM;
752 goto out;
753 }
754
755 p = req->request->front.iov_base;
756 end = p + req->request->front_alloc_len;
757
758 /* fill out request */
759 mutex_lock(&monc->mutex);
760 tid = ++monc->last_tid;
761 ceph_encode_64(&p, tid); /* handle */
762 ceph_encode_string(&p, end, what, strlen(what));
763
764 err = __do_generic_request(monc, tid, req);
765
766 mutex_unlock(&monc->mutex);
767out:
Ilya Dryomovf6469122014-12-22 19:35:17 +0300768 put_generic_request(req);
Ilya Dryomov513a8242014-05-13 11:19:26 +0400769 return err;
770}
771EXPORT_SYMBOL(ceph_monc_do_get_version);
772
Sage Weilba75bb92009-10-06 11:31:11 -0700773/*
Yehuda Sadehe56fa102010-05-17 12:40:28 -0700774 * Resend pending generic requests.
Sage Weilba75bb92009-10-06 11:31:11 -0700775 */
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700776static void __resend_generic_request(struct ceph_mon_client *monc)
Sage Weilba75bb92009-10-06 11:31:11 -0700777{
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700778 struct ceph_mon_generic_request *req;
Sage Weil85ff03f2010-02-15 14:47:28 -0800779 struct rb_node *p;
Sage Weilba75bb92009-10-06 11:31:11 -0700780
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700781 for (p = rb_first(&monc->generic_request_tree); p; p = rb_next(p)) {
782 req = rb_entry(p, struct ceph_mon_generic_request, node);
Alex Elder6740a842012-06-01 14:56:43 -0500783 ceph_msg_revoke(req->request);
Sage Weil4f471e42012-07-30 18:16:40 -0700784 ceph_msg_revoke_incoming(req->reply);
Alex Elder67130932012-05-26 23:26:43 -0500785 ceph_con_send(&monc->con, ceph_msg_get(req->request));
Sage Weilba75bb92009-10-06 11:31:11 -0700786 }
787}
788
789/*
790 * Delayed work. If we haven't mounted yet, retry. Otherwise,
791 * renew/retry subscription as needed (in case it is timing out, or we
792 * got an ENOMEM). And keep the monitor connection alive.
793 */
794static void delayed_work(struct work_struct *work)
795{
796 struct ceph_mon_client *monc =
797 container_of(work, struct ceph_mon_client, delayed_work.work);
798
799 dout("monc delayed_work\n");
800 mutex_lock(&monc->mutex);
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800801 if (monc->hunting) {
Ilya Dryomov1752b502016-01-21 16:33:25 +0100802 dout("%s continuing hunt\n", __func__);
803 reopen_session(monc);
Sage Weilba75bb92009-10-06 11:31:11 -0700804 } else {
Yan, Zheng8b9558a2015-09-01 17:19:38 +0800805 int is_auth = ceph_auth_is_authenticated(monc->auth);
806 if (ceph_con_keepalive_expired(&monc->con,
Ilya Dryomov58d81b12016-01-21 16:33:15 +0100807 CEPH_MONC_PING_TIMEOUT)) {
Yan, Zheng8b9558a2015-09-01 17:19:38 +0800808 dout("monc keepalive timeout\n");
809 is_auth = 0;
Ilya Dryomov1752b502016-01-21 16:33:25 +0100810 reopen_session(monc);
Yan, Zheng8b9558a2015-09-01 17:19:38 +0800811 }
Sage Weil9bd2e6f2010-02-02 16:21:06 -0800812
Yan, Zheng8b9558a2015-09-01 17:19:38 +0800813 if (!monc->hunting) {
814 ceph_con_keepalive(&monc->con);
815 __validate_auth(monc);
816 }
Sage Weil9bd2e6f2010-02-02 16:21:06 -0800817
Ilya Dryomov82dcaba2016-01-19 16:19:06 +0100818 if (is_auth) {
819 unsigned long now = jiffies;
820
821 dout("%s renew subs? now %lu renew after %lu\n",
822 __func__, now, monc->sub_renew_after);
823 if (time_after_eq(now, monc->sub_renew_after))
824 __send_subscribe(monc);
825 }
Sage Weilba75bb92009-10-06 11:31:11 -0700826 }
Sage Weilba75bb92009-10-06 11:31:11 -0700827 __schedule_delayed(monc);
828 mutex_unlock(&monc->mutex);
829}
830
Sage Weil6b805182009-10-27 11:50:50 -0700831/*
832 * On startup, we build a temporary monmap populated with the IPs
833 * provided by mount(2).
834 */
835static int build_initial_monmap(struct ceph_mon_client *monc)
836{
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700837 struct ceph_options *opt = monc->client->options;
838 struct ceph_entity_addr *mon_addr = opt->mon_addr;
839 int num_mon = opt->num_mon;
Sage Weil6b805182009-10-27 11:50:50 -0700840 int i;
841
842 /* build initial monmap */
843 monc->monmap = kzalloc(sizeof(*monc->monmap) +
844 num_mon*sizeof(monc->monmap->mon_inst[0]),
845 GFP_KERNEL);
846 if (!monc->monmap)
847 return -ENOMEM;
848 for (i = 0; i < num_mon; i++) {
849 monc->monmap->mon_inst[i].addr = mon_addr[i];
Sage Weil6b805182009-10-27 11:50:50 -0700850 monc->monmap->mon_inst[i].addr.nonce = 0;
851 monc->monmap->mon_inst[i].name.type =
852 CEPH_ENTITY_TYPE_MON;
853 monc->monmap->mon_inst[i].name.num = cpu_to_le64(i);
854 }
855 monc->monmap->num_mon = num_mon;
Sage Weil6b805182009-10-27 11:50:50 -0700856 return 0;
857}
858
Sage Weilba75bb92009-10-06 11:31:11 -0700859int ceph_monc_init(struct ceph_mon_client *monc, struct ceph_client *cl)
860{
861 int err = 0;
862
863 dout("init\n");
864 memset(monc, 0, sizeof(*monc));
865 monc->client = cl;
866 monc->monmap = NULL;
867 mutex_init(&monc->mutex);
868
Sage Weil6b805182009-10-27 11:50:50 -0700869 err = build_initial_monmap(monc);
870 if (err)
871 goto out;
872
Sage Weilf6a2f5b2011-08-09 09:27:44 -0700873 /* connection */
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800874 /* authentication */
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700875 monc->auth = ceph_auth_init(cl->options->name,
Tommi Virtanen8323c3a2011-03-25 16:32:57 -0700876 cl->options->key);
Noah Watkins49d92242011-09-12 14:51:58 -0700877 if (IS_ERR(monc->auth)) {
878 err = PTR_ERR(monc->auth);
Alex Elder67130932012-05-26 23:26:43 -0500879 goto out_monmap;
Noah Watkins49d92242011-09-12 14:51:58 -0700880 }
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800881 monc->auth->want_keys =
882 CEPH_ENTITY_TYPE_AUTH | CEPH_ENTITY_TYPE_MON |
883 CEPH_ENTITY_TYPE_OSD | CEPH_ENTITY_TYPE_MDS;
884
Sage Weil240ed682010-05-21 14:57:25 -0700885 /* msgs */
Sage Weila79832f2010-04-01 16:06:19 -0700886 err = -ENOMEM;
Sage Weil7c315c52010-03-24 21:52:30 -0700887 monc->m_subscribe_ack = ceph_msg_new(CEPH_MSG_MON_SUBSCRIBE_ACK,
Yehuda Sadeh34d23762010-04-06 14:33:58 -0700888 sizeof(struct ceph_mon_subscribe_ack),
Sage Weilb61c2762011-08-09 15:03:46 -0700889 GFP_NOFS, true);
Sage Weila79832f2010-04-01 16:06:19 -0700890 if (!monc->m_subscribe_ack)
Noah Watkins49d92242011-09-12 14:51:58 -0700891 goto out_auth;
Sage Weil6694d6b2010-03-24 21:48:05 -0700892
Sage Weilb61c2762011-08-09 15:03:46 -0700893 monc->m_subscribe = ceph_msg_new(CEPH_MSG_MON_SUBSCRIBE, 96, GFP_NOFS,
894 true);
Sage Weil240ed682010-05-21 14:57:25 -0700895 if (!monc->m_subscribe)
896 goto out_subscribe_ack;
897
Sage Weilb61c2762011-08-09 15:03:46 -0700898 monc->m_auth_reply = ceph_msg_new(CEPH_MSG_AUTH_REPLY, 4096, GFP_NOFS,
899 true);
Sage Weila79832f2010-04-01 16:06:19 -0700900 if (!monc->m_auth_reply)
Sage Weil240ed682010-05-21 14:57:25 -0700901 goto out_subscribe;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800902
Sage Weilb61c2762011-08-09 15:03:46 -0700903 monc->m_auth = ceph_msg_new(CEPH_MSG_AUTH, 4096, GFP_NOFS, true);
Sage Weil9bd2e6f2010-02-02 16:21:06 -0800904 monc->pending_auth = 0;
Sage Weila79832f2010-04-01 16:06:19 -0700905 if (!monc->m_auth)
Sage Weil6694d6b2010-03-24 21:48:05 -0700906 goto out_auth_reply;
Sage Weilba75bb92009-10-06 11:31:11 -0700907
Sage Weil735a72e2012-06-27 12:24:34 -0700908 ceph_con_init(&monc->con, monc, &mon_con_ops,
909 &monc->client->msgr);
910
Sage Weilba75bb92009-10-06 11:31:11 -0700911 monc->cur_mon = -1;
Ilya Dryomov168b9092016-01-21 16:33:19 +0100912 monc->had_a_connection = false;
913 monc->hunt_mult = 1;
Sage Weilba75bb92009-10-06 11:31:11 -0700914
915 INIT_DELAYED_WORK(&monc->delayed_work, delayed_work);
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700916 monc->generic_request_tree = RB_ROOT;
917 monc->num_generic_requests = 0;
Sage Weilba75bb92009-10-06 11:31:11 -0700918 monc->last_tid = 0;
919
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800920 return 0;
921
Sage Weil6694d6b2010-03-24 21:48:05 -0700922out_auth_reply:
923 ceph_msg_put(monc->m_auth_reply);
Sage Weil240ed682010-05-21 14:57:25 -0700924out_subscribe:
925 ceph_msg_put(monc->m_subscribe);
Sage Weil7c315c52010-03-24 21:52:30 -0700926out_subscribe_ack:
927 ceph_msg_put(monc->m_subscribe_ack);
Noah Watkins49d92242011-09-12 14:51:58 -0700928out_auth:
929 ceph_auth_destroy(monc->auth);
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800930out_monmap:
931 kfree(monc->monmap);
Sage Weilba75bb92009-10-06 11:31:11 -0700932out:
933 return err;
934}
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700935EXPORT_SYMBOL(ceph_monc_init);
Sage Weilba75bb92009-10-06 11:31:11 -0700936
937void ceph_monc_stop(struct ceph_mon_client *monc)
938{
939 dout("stop\n");
940 cancel_delayed_work_sync(&monc->delayed_work);
941
942 mutex_lock(&monc->mutex);
943 __close_session(monc);
Ilya Dryomov0e04dc22016-01-20 17:50:31 +0100944 monc->cur_mon = -1;
Sage Weilba75bb92009-10-06 11:31:11 -0700945 mutex_unlock(&monc->mutex);
946
Sage Weilf3dea7e2012-06-10 20:43:56 -0700947 /*
948 * flush msgr queue before we destroy ourselves to ensure that:
949 * - any work that references our embedded con is finished.
950 * - any osd_client or other work that may reference an authorizer
951 * finishes before we shut down the auth subsystem.
952 */
953 ceph_msgr_flush();
954
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800955 ceph_auth_destroy(monc->auth);
956
957 ceph_msg_put(monc->m_auth);
Sage Weil6694d6b2010-03-24 21:48:05 -0700958 ceph_msg_put(monc->m_auth_reply);
Sage Weil240ed682010-05-21 14:57:25 -0700959 ceph_msg_put(monc->m_subscribe);
Sage Weil7c315c52010-03-24 21:52:30 -0700960 ceph_msg_put(monc->m_subscribe_ack);
Sage Weilba75bb92009-10-06 11:31:11 -0700961
962 kfree(monc->monmap);
963}
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700964EXPORT_SYMBOL(ceph_monc_stop);
Sage Weilba75bb92009-10-06 11:31:11 -0700965
Ilya Dryomov0f9af162016-01-08 21:17:22 +0300966static void finish_hunting(struct ceph_mon_client *monc)
967{
968 if (monc->hunting) {
969 dout("%s found mon%d\n", __func__, monc->cur_mon);
970 monc->hunting = false;
Ilya Dryomov168b9092016-01-21 16:33:19 +0100971 monc->had_a_connection = true;
972 monc->hunt_mult /= 2; /* reduce by 50% */
973 if (monc->hunt_mult < 1)
974 monc->hunt_mult = 1;
Ilya Dryomov0f9af162016-01-08 21:17:22 +0300975 }
976}
977
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800978static void handle_auth_reply(struct ceph_mon_client *monc,
979 struct ceph_msg *msg)
980{
981 int ret;
Sage Weil09c4d6a2010-05-25 15:38:06 -0700982 int was_auth = 0;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800983
984 mutex_lock(&monc->mutex);
Sage Weil27859f92013-03-25 10:26:14 -0700985 was_auth = ceph_auth_is_authenticated(monc->auth);
Sage Weil9bd2e6f2010-02-02 16:21:06 -0800986 monc->pending_auth = 0;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800987 ret = ceph_handle_auth_reply(monc->auth, msg->front.iov_base,
988 msg->front.iov_len,
989 monc->m_auth->front.iov_base,
Ilya Dryomov3cea4c32014-01-09 20:08:21 +0200990 monc->m_auth->front_alloc_len);
Ilya Dryomov0f9af162016-01-08 21:17:22 +0300991 if (ret > 0) {
992 __send_prepared_auth_request(monc, ret);
993 goto out;
994 }
995
996 finish_hunting(monc);
997
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800998 if (ret < 0) {
Sage Weil9bd2e6f2010-02-02 16:21:06 -0800999 monc->client->auth_err = ret;
Sage Weil27859f92013-03-25 10:26:14 -07001000 } else if (!was_auth && ceph_auth_is_authenticated(monc->auth)) {
Sage Weil4e7a5dc2009-11-18 16:19:57 -08001001 dout("authenticated, starting session\n");
Sage Weil07433042009-11-18 16:50:41 -08001002
Alex Elder15d98822012-05-26 23:26:43 -05001003 monc->client->msgr.inst.name.type = CEPH_ENTITY_TYPE_CLIENT;
1004 monc->client->msgr.inst.name.num =
Yehuda Sadeh0cf55372010-06-11 15:57:06 -07001005 cpu_to_le64(monc->auth->global_id);
Sage Weil07433042009-11-18 16:50:41 -08001006
Sage Weil4e7a5dc2009-11-18 16:19:57 -08001007 __send_subscribe(monc);
Yehuda Sadehf8c76f62010-04-22 15:40:37 -07001008 __resend_generic_request(monc);
Ilya Dryomov0f9af162016-01-08 21:17:22 +03001009
1010 pr_info("mon%d %s session established\n", monc->cur_mon,
1011 ceph_pr_addr(&monc->con.peer_addr.in_addr));
Sage Weil4e7a5dc2009-11-18 16:19:57 -08001012 }
Ilya Dryomov0f9af162016-01-08 21:17:22 +03001013
1014out:
Sage Weil4e7a5dc2009-11-18 16:19:57 -08001015 mutex_unlock(&monc->mutex);
Ilya Dryomov0f9af162016-01-08 21:17:22 +03001016 if (monc->client->auth_err < 0)
1017 wake_up_all(&monc->client->auth_wq);
Sage Weil4e7a5dc2009-11-18 16:19:57 -08001018}
1019
Sage Weil9bd2e6f2010-02-02 16:21:06 -08001020static int __validate_auth(struct ceph_mon_client *monc)
1021{
1022 int ret;
1023
1024 if (monc->pending_auth)
1025 return 0;
1026
1027 ret = ceph_build_auth(monc->auth, monc->m_auth->front.iov_base,
Ilya Dryomov3cea4c32014-01-09 20:08:21 +02001028 monc->m_auth->front_alloc_len);
Sage Weil9bd2e6f2010-02-02 16:21:06 -08001029 if (ret <= 0)
1030 return ret; /* either an error, or no need to authenticate */
1031 __send_prepared_auth_request(monc, ret);
1032 return 0;
1033}
1034
1035int ceph_monc_validate_auth(struct ceph_mon_client *monc)
1036{
1037 int ret;
1038
1039 mutex_lock(&monc->mutex);
1040 ret = __validate_auth(monc);
1041 mutex_unlock(&monc->mutex);
1042 return ret;
1043}
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07001044EXPORT_SYMBOL(ceph_monc_validate_auth);
Sage Weil9bd2e6f2010-02-02 16:21:06 -08001045
Sage Weilba75bb92009-10-06 11:31:11 -07001046/*
1047 * handle incoming message
1048 */
1049static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1050{
1051 struct ceph_mon_client *monc = con->private;
1052 int type = le16_to_cpu(msg->hdr.type);
1053
1054 if (!monc)
1055 return;
1056
1057 switch (type) {
Sage Weil4e7a5dc2009-11-18 16:19:57 -08001058 case CEPH_MSG_AUTH_REPLY:
1059 handle_auth_reply(monc, msg);
Sage Weilba75bb92009-10-06 11:31:11 -07001060 break;
1061
1062 case CEPH_MSG_MON_SUBSCRIBE_ACK:
1063 handle_subscribe_ack(monc, msg);
1064 break;
1065
1066 case CEPH_MSG_STATFS_REPLY:
1067 handle_statfs_reply(monc, msg);
1068 break;
1069
Ilya Dryomov513a8242014-05-13 11:19:26 +04001070 case CEPH_MSG_MON_GET_VERSION_REPLY:
1071 handle_get_version_reply(monc, msg);
1072 break;
1073
Sage Weil4e7a5dc2009-11-18 16:19:57 -08001074 case CEPH_MSG_MON_MAP:
1075 ceph_monc_handle_map(monc, msg);
1076 break;
1077
Sage Weilba75bb92009-10-06 11:31:11 -07001078 case CEPH_MSG_OSD_MAP:
1079 ceph_osdc_handle_map(&monc->client->osdc, msg);
1080 break;
1081
1082 default:
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07001083 /* can the chained handler handle it? */
1084 if (monc->client->extra_mon_dispatch &&
1085 monc->client->extra_mon_dispatch(monc->client, msg) == 0)
1086 break;
1087
Sage Weilba75bb92009-10-06 11:31:11 -07001088 pr_err("received unknown message type %d %s\n", type,
1089 ceph_msg_type_name(type));
1090 }
1091 ceph_msg_put(msg);
1092}
1093
1094/*
1095 * Allocate memory for incoming message
1096 */
1097static struct ceph_msg *mon_alloc_msg(struct ceph_connection *con,
Yehuda Sadeh24504182010-01-08 13:58:34 -08001098 struct ceph_msg_header *hdr,
1099 int *skip)
Sage Weilba75bb92009-10-06 11:31:11 -07001100{
1101 struct ceph_mon_client *monc = con->private;
1102 int type = le16_to_cpu(hdr->type);
Yehuda Sadeh24504182010-01-08 13:58:34 -08001103 int front_len = le32_to_cpu(hdr->front_len);
Sage Weil5b3a4db2010-02-19 21:43:23 -08001104 struct ceph_msg *m = NULL;
Sage Weilba75bb92009-10-06 11:31:11 -07001105
Yehuda Sadeh24504182010-01-08 13:58:34 -08001106 *skip = 0;
Yehuda Sadeh0547a9b2010-01-11 14:47:13 -08001107
Sage Weilba75bb92009-10-06 11:31:11 -07001108 switch (type) {
Sage Weilba75bb92009-10-06 11:31:11 -07001109 case CEPH_MSG_MON_SUBSCRIBE_ACK:
Sage Weil7c315c52010-03-24 21:52:30 -07001110 m = ceph_msg_get(monc->m_subscribe_ack);
Yehuda Sadeh24504182010-01-08 13:58:34 -08001111 break;
Sage Weilba75bb92009-10-06 11:31:11 -07001112 case CEPH_MSG_STATFS_REPLY:
Yehuda Sadehf8c76f62010-04-22 15:40:37 -07001113 return get_generic_reply(con, hdr, skip);
Sage Weil4e7a5dc2009-11-18 16:19:57 -08001114 case CEPH_MSG_AUTH_REPLY:
Sage Weil6694d6b2010-03-24 21:48:05 -07001115 m = ceph_msg_get(monc->m_auth_reply);
Yehuda Sadeh24504182010-01-08 13:58:34 -08001116 break;
Ilya Dryomov513a8242014-05-13 11:19:26 +04001117 case CEPH_MSG_MON_GET_VERSION_REPLY:
1118 if (le64_to_cpu(hdr->tid) != 0)
1119 return get_generic_reply(con, hdr, skip);
1120
1121 /*
1122 * Older OSDs don't set reply tid even if the orignal
1123 * request had a non-zero tid. Workaround this weirdness
1124 * by falling through to the allocate case.
1125 */
Sage Weil5b3a4db2010-02-19 21:43:23 -08001126 case CEPH_MSG_MON_MAP:
1127 case CEPH_MSG_MDS_MAP:
1128 case CEPH_MSG_OSD_MAP:
Sage Weilb61c2762011-08-09 15:03:46 -07001129 m = ceph_msg_new(type, front_len, GFP_NOFS, false);
Alex Elder1c20f2d2012-06-04 14:43:32 -05001130 if (!m)
1131 return NULL; /* ENOMEM--return skip == 0 */
Sage Weil5b3a4db2010-02-19 21:43:23 -08001132 break;
Sage Weilba75bb92009-10-06 11:31:11 -07001133 }
Yehuda Sadeh24504182010-01-08 13:58:34 -08001134
Sage Weil5b3a4db2010-02-19 21:43:23 -08001135 if (!m) {
1136 pr_info("alloc_msg unknown type %d\n", type);
Yehuda Sadeh24504182010-01-08 13:58:34 -08001137 *skip = 1;
Sage Weil73c3d482014-08-04 07:01:54 -07001138 } else if (front_len > m->front_alloc_len) {
Joe Perchesb9a67892014-09-09 21:17:29 -07001139 pr_warn("mon_alloc_msg front %d > prealloc %d (%u#%llu)\n",
1140 front_len, m->front_alloc_len,
1141 (unsigned int)con->peer_name.type,
1142 le64_to_cpu(con->peer_name.num));
Sage Weil73c3d482014-08-04 07:01:54 -07001143 ceph_msg_put(m);
1144 m = ceph_msg_new(type, front_len, GFP_NOFS, false);
Sage Weil5b3a4db2010-02-19 21:43:23 -08001145 }
Sage Weil73c3d482014-08-04 07:01:54 -07001146
Yehuda Sadeh24504182010-01-08 13:58:34 -08001147 return m;
Sage Weilba75bb92009-10-06 11:31:11 -07001148}
1149
1150/*
1151 * If the monitor connection resets, pick a new monitor and resubmit
1152 * any pending requests.
1153 */
1154static void mon_fault(struct ceph_connection *con)
1155{
1156 struct ceph_mon_client *monc = con->private;
1157
Sage Weilba75bb92009-10-06 11:31:11 -07001158 mutex_lock(&monc->mutex);
Ilya Dryomovb5d91702016-01-23 15:57:51 +01001159 dout("%s mon%d\n", __func__, monc->cur_mon);
1160 if (monc->cur_mon >= 0) {
1161 if (!monc->hunting) {
1162 dout("%s hunting for new mon\n", __func__);
1163 reopen_session(monc);
1164 __schedule_delayed(monc);
1165 } else {
1166 dout("%s already hunting\n", __func__);
1167 }
Sage Weilba75bb92009-10-06 11:31:11 -07001168 }
Sage Weilba75bb92009-10-06 11:31:11 -07001169 mutex_unlock(&monc->mutex);
1170}
1171
Sage Weilec87ef42012-05-31 20:27:50 -07001172/*
1173 * We can ignore refcounting on the connection struct, as all references
1174 * will come from the messenger workqueue, which is drained prior to
1175 * mon_client destruction.
1176 */
1177static struct ceph_connection *con_get(struct ceph_connection *con)
1178{
1179 return con;
1180}
1181
1182static void con_put(struct ceph_connection *con)
1183{
1184}
1185
Tobias Klauser9e327892010-05-20 10:40:19 +02001186static const struct ceph_connection_operations mon_con_ops = {
Sage Weilec87ef42012-05-31 20:27:50 -07001187 .get = con_get,
1188 .put = con_put,
Sage Weilba75bb92009-10-06 11:31:11 -07001189 .dispatch = dispatch,
1190 .fault = mon_fault,
1191 .alloc_msg = mon_alloc_msg,
Sage Weilba75bb92009-10-06 11:31:11 -07001192};