blob: 556721665335cf3daa7f4cf0366768e76f6a7f8a [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>
11#include <linux/ceph/decode.h>
12
13#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);
Sage Weil97069002010-05-21 12:31:49 -0700109 ceph_con_revoke(monc->con, monc->m_auth);
Sage Weil5ce6e9d2010-02-15 16:22:28 -0800110 ceph_msg_get(monc->m_auth); /* keep our ref */
111 ceph_con_send(monc->con, monc->m_auth);
112}
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);
120 ceph_con_revoke(monc->con, monc->m_auth);
121 ceph_con_close(monc->con);
122 monc->cur_mon = -1;
123 monc->pending_auth = 0;
124 ceph_auth_reset(monc->auth);
Sage Weilba75bb92009-10-06 11:31:11 -0700125}
126
127/*
128 * Open a session with a (new) monitor.
129 */
130static int __open_session(struct ceph_mon_client *monc)
131{
132 char r;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800133 int ret;
Sage Weilba75bb92009-10-06 11:31:11 -0700134
135 if (monc->cur_mon < 0) {
136 get_random_bytes(&r, 1);
137 monc->cur_mon = r % monc->monmap->num_mon;
138 dout("open_session num=%d r=%d -> mon%d\n",
139 monc->monmap->num_mon, r, monc->cur_mon);
140 monc->sub_sent = 0;
141 monc->sub_renew_after = jiffies; /* i.e., expired */
142 monc->want_next_osdmap = !!monc->want_next_osdmap;
143
144 dout("open_session mon%d opening\n", monc->cur_mon);
145 monc->con->peer_name.type = CEPH_ENTITY_TYPE_MON;
146 monc->con->peer_name.num = cpu_to_le64(monc->cur_mon);
147 ceph_con_open(monc->con,
148 &monc->monmap->mon_inst[monc->cur_mon].addr);
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800149
150 /* initiatiate authentication handshake */
151 ret = ceph_auth_build_hello(monc->auth,
152 monc->m_auth->front.iov_base,
153 monc->m_auth->front_max);
Sage Weil5ce6e9d2010-02-15 16:22:28 -0800154 __send_prepared_auth_request(monc, ret);
Sage Weilba75bb92009-10-06 11:31:11 -0700155 } else {
156 dout("open_session mon%d already open\n", monc->cur_mon);
157 }
158 return 0;
159}
160
161static bool __sub_expired(struct ceph_mon_client *monc)
162{
163 return time_after_eq(jiffies, monc->sub_renew_after);
164}
165
166/*
167 * Reschedule delayed work timer.
168 */
169static void __schedule_delayed(struct ceph_mon_client *monc)
170{
171 unsigned delay;
172
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800173 if (monc->cur_mon < 0 || __sub_expired(monc))
Sage Weilba75bb92009-10-06 11:31:11 -0700174 delay = 10 * HZ;
175 else
176 delay = 20 * HZ;
177 dout("__schedule_delayed after %u\n", delay);
178 schedule_delayed_work(&monc->delayed_work, delay);
179}
180
181/*
182 * Send subscribe request for mdsmap and/or osdmap.
183 */
184static void __send_subscribe(struct ceph_mon_client *monc)
185{
186 dout("__send_subscribe sub_sent=%u exp=%u want_osd=%d\n",
187 (unsigned)monc->sub_sent, __sub_expired(monc),
188 monc->want_next_osdmap);
189 if ((__sub_expired(monc) && !monc->sub_sent) ||
190 monc->want_next_osdmap == 1) {
Sage Weil240ed682010-05-21 14:57:25 -0700191 struct ceph_msg *msg = monc->m_subscribe;
Sage Weilba75bb92009-10-06 11:31:11 -0700192 struct ceph_mon_subscribe_item *i;
193 void *p, *end;
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700194 int num;
Sage Weilba75bb92009-10-06 11:31:11 -0700195
Sage Weilba75bb92009-10-06 11:31:11 -0700196 p = msg->front.iov_base;
Sage Weil240ed682010-05-21 14:57:25 -0700197 end = p + msg->front_max;
Sage Weilba75bb92009-10-06 11:31:11 -0700198
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700199 num = 1 + !!monc->want_next_osdmap + !!monc->want_mdsmap;
200 ceph_encode_32(&p, num);
201
Sage Weilba75bb92009-10-06 11:31:11 -0700202 if (monc->want_next_osdmap) {
203 dout("__send_subscribe to 'osdmap' %u\n",
204 (unsigned)monc->have_osdmap);
Sage Weilba75bb92009-10-06 11:31:11 -0700205 ceph_encode_string(&p, end, "osdmap", 6);
206 i = p;
207 i->have = cpu_to_le64(monc->have_osdmap);
208 i->onetime = 1;
209 p += sizeof(*i);
210 monc->want_next_osdmap = 2; /* requested */
Sage Weilba75bb92009-10-06 11:31:11 -0700211 }
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700212 if (monc->want_mdsmap) {
213 dout("__send_subscribe to 'mdsmap' %u+\n",
214 (unsigned)monc->have_mdsmap);
215 ceph_encode_string(&p, end, "mdsmap", 6);
216 i = p;
217 i->have = cpu_to_le64(monc->have_mdsmap);
218 i->onetime = 0;
219 p += sizeof(*i);
220 }
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800221 ceph_encode_string(&p, end, "monmap", 6);
222 i = p;
223 i->have = 0;
224 i->onetime = 0;
225 p += sizeof(*i);
Sage Weilba75bb92009-10-06 11:31:11 -0700226
227 msg->front.iov_len = p - msg->front.iov_base;
228 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
Sage Weil240ed682010-05-21 14:57:25 -0700229 ceph_con_revoke(monc->con, msg);
230 ceph_con_send(monc->con, ceph_msg_get(msg));
Sage Weilba75bb92009-10-06 11:31:11 -0700231
232 monc->sub_sent = jiffies | 1; /* never 0 */
233 }
234}
235
236static void handle_subscribe_ack(struct ceph_mon_client *monc,
237 struct ceph_msg *msg)
238{
239 unsigned seconds;
Sage Weil07bd10f2009-10-14 17:26:40 -0700240 struct ceph_mon_subscribe_ack *h = msg->front.iov_base;
Sage Weilba75bb92009-10-06 11:31:11 -0700241
Sage Weil07bd10f2009-10-14 17:26:40 -0700242 if (msg->front.iov_len < sizeof(*h))
243 goto bad;
244 seconds = le32_to_cpu(h->duration);
245
Sage Weilba75bb92009-10-06 11:31:11 -0700246 mutex_lock(&monc->mutex);
247 if (monc->hunting) {
248 pr_info("mon%d %s session established\n",
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700249 monc->cur_mon,
250 ceph_pr_addr(&monc->con->peer_addr.in_addr));
Sage Weilba75bb92009-10-06 11:31:11 -0700251 monc->hunting = false;
252 }
253 dout("handle_subscribe_ack after %d seconds\n", seconds);
Sage Weil0656d112009-10-08 10:25:46 -0700254 monc->sub_renew_after = monc->sub_sent + (seconds >> 1)*HZ - 1;
Sage Weilba75bb92009-10-06 11:31:11 -0700255 monc->sub_sent = 0;
256 mutex_unlock(&monc->mutex);
257 return;
258bad:
259 pr_err("got corrupt subscribe-ack msg\n");
Sage Weil9ec7cab2009-12-14 15:13:47 -0800260 ceph_msg_dump(msg);
Sage Weilba75bb92009-10-06 11:31:11 -0700261}
262
263/*
264 * Keep track of which maps we have
265 */
266int ceph_monc_got_mdsmap(struct ceph_mon_client *monc, u32 got)
267{
268 mutex_lock(&monc->mutex);
269 monc->have_mdsmap = got;
270 mutex_unlock(&monc->mutex);
271 return 0;
272}
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700273EXPORT_SYMBOL(ceph_monc_got_mdsmap);
Sage Weilba75bb92009-10-06 11:31:11 -0700274
275int ceph_monc_got_osdmap(struct ceph_mon_client *monc, u32 got)
276{
277 mutex_lock(&monc->mutex);
278 monc->have_osdmap = got;
279 monc->want_next_osdmap = 0;
280 mutex_unlock(&monc->mutex);
281 return 0;
282}
283
284/*
285 * Register interest in the next osdmap
286 */
287void ceph_monc_request_next_osdmap(struct ceph_mon_client *monc)
288{
289 dout("request_next_osdmap have %u\n", monc->have_osdmap);
290 mutex_lock(&monc->mutex);
291 if (!monc->want_next_osdmap)
292 monc->want_next_osdmap = 1;
293 if (monc->want_next_osdmap < 2)
294 __send_subscribe(monc);
295 mutex_unlock(&monc->mutex);
296}
297
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800298/*
Sage Weil50b885b2009-12-01 14:12:07 -0800299 *
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800300 */
301int ceph_monc_open_session(struct ceph_mon_client *monc)
Sage Weilba75bb92009-10-06 11:31:11 -0700302{
Sage Weilba75bb92009-10-06 11:31:11 -0700303 mutex_lock(&monc->mutex);
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800304 __open_session(monc);
Sage Weilba75bb92009-10-06 11:31:11 -0700305 __schedule_delayed(monc);
306 mutex_unlock(&monc->mutex);
307 return 0;
308}
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700309EXPORT_SYMBOL(ceph_monc_open_session);
Sage Weilba75bb92009-10-06 11:31:11 -0700310
311/*
312 * The monitor responds with mount ack indicate mount success. The
313 * included client ticket allows the client to talk to MDSs and OSDs.
314 */
Sage Weil07433042009-11-18 16:50:41 -0800315static void ceph_monc_handle_map(struct ceph_mon_client *monc,
316 struct ceph_msg *msg)
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800317{
318 struct ceph_client *client = monc->client;
319 struct ceph_monmap *monmap = NULL, *old = monc->monmap;
320 void *p, *end;
321
322 mutex_lock(&monc->mutex);
323
324 dout("handle_monmap\n");
325 p = msg->front.iov_base;
326 end = p + msg->front.iov_len;
327
328 monmap = ceph_monmap_decode(p, end);
329 if (IS_ERR(monmap)) {
330 pr_err("problem decoding monmap, %d\n",
331 (int)PTR_ERR(monmap));
Sage Weild4a780c2009-12-11 08:55:23 -0800332 goto out;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800333 }
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800334
Sage Weil07433042009-11-18 16:50:41 -0800335 if (ceph_check_fsid(monc->client, &monmap->fsid) < 0) {
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800336 kfree(monmap);
Sage Weild4a780c2009-12-11 08:55:23 -0800337 goto out;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800338 }
339
340 client->monc.monmap = monmap;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800341 kfree(old);
342
Sage Weild4a780c2009-12-11 08:55:23 -0800343out:
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800344 mutex_unlock(&monc->mutex);
Yehuda Sadeh03066f22010-07-27 13:11:08 -0700345 wake_up_all(&client->auth_wq);
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800346}
Sage Weilba75bb92009-10-06 11:31:11 -0700347
Sage Weilba75bb92009-10-06 11:31:11 -0700348/*
Yehuda Sadehe56fa102010-05-17 12:40:28 -0700349 * generic requests (e.g., statfs, poolop)
Sage Weilba75bb92009-10-06 11:31:11 -0700350 */
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700351static struct ceph_mon_generic_request *__lookup_generic_req(
Sage Weil85ff03f2010-02-15 14:47:28 -0800352 struct ceph_mon_client *monc, u64 tid)
353{
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700354 struct ceph_mon_generic_request *req;
355 struct rb_node *n = monc->generic_request_tree.rb_node;
Sage Weil85ff03f2010-02-15 14:47:28 -0800356
357 while (n) {
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700358 req = rb_entry(n, struct ceph_mon_generic_request, node);
Sage Weil85ff03f2010-02-15 14:47:28 -0800359 if (tid < req->tid)
360 n = n->rb_left;
361 else if (tid > req->tid)
362 n = n->rb_right;
363 else
364 return req;
365 }
366 return NULL;
367}
368
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700369static void __insert_generic_request(struct ceph_mon_client *monc,
370 struct ceph_mon_generic_request *new)
Sage Weil85ff03f2010-02-15 14:47:28 -0800371{
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700372 struct rb_node **p = &monc->generic_request_tree.rb_node;
Sage Weil85ff03f2010-02-15 14:47:28 -0800373 struct rb_node *parent = NULL;
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700374 struct ceph_mon_generic_request *req = NULL;
Sage Weil85ff03f2010-02-15 14:47:28 -0800375
376 while (*p) {
377 parent = *p;
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700378 req = rb_entry(parent, struct ceph_mon_generic_request, node);
Sage Weil85ff03f2010-02-15 14:47:28 -0800379 if (new->tid < req->tid)
380 p = &(*p)->rb_left;
381 else if (new->tid > req->tid)
382 p = &(*p)->rb_right;
383 else
384 BUG();
385 }
386
387 rb_link_node(&new->node, parent, p);
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700388 rb_insert_color(&new->node, &monc->generic_request_tree);
Sage Weil85ff03f2010-02-15 14:47:28 -0800389}
390
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700391static void release_generic_request(struct kref *kref)
Sage Weil3143edd2010-03-24 21:43:33 -0700392{
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700393 struct ceph_mon_generic_request *req =
394 container_of(kref, struct ceph_mon_generic_request, kref);
Sage Weil3143edd2010-03-24 21:43:33 -0700395
396 if (req->reply)
397 ceph_msg_put(req->reply);
398 if (req->request)
399 ceph_msg_put(req->request);
Yehuda Sadeh20547562010-06-01 10:37:40 -0700400
401 kfree(req);
Sage Weil3143edd2010-03-24 21:43:33 -0700402}
403
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700404static void put_generic_request(struct ceph_mon_generic_request *req)
Sage Weil3143edd2010-03-24 21:43:33 -0700405{
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700406 kref_put(&req->kref, release_generic_request);
Sage Weil3143edd2010-03-24 21:43:33 -0700407}
408
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700409static void get_generic_request(struct ceph_mon_generic_request *req)
Sage Weil3143edd2010-03-24 21:43:33 -0700410{
411 kref_get(&req->kref);
412}
413
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700414static struct ceph_msg *get_generic_reply(struct ceph_connection *con,
Sage Weil3143edd2010-03-24 21:43:33 -0700415 struct ceph_msg_header *hdr,
416 int *skip)
417{
418 struct ceph_mon_client *monc = con->private;
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700419 struct ceph_mon_generic_request *req;
Sage Weil3143edd2010-03-24 21:43:33 -0700420 u64 tid = le64_to_cpu(hdr->tid);
421 struct ceph_msg *m;
422
423 mutex_lock(&monc->mutex);
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700424 req = __lookup_generic_req(monc, tid);
Sage Weil3143edd2010-03-24 21:43:33 -0700425 if (!req) {
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700426 dout("get_generic_reply %lld dne\n", tid);
Sage Weil3143edd2010-03-24 21:43:33 -0700427 *skip = 1;
428 m = NULL;
429 } else {
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700430 dout("get_generic_reply %lld got %p\n", tid, req->reply);
Sage Weil3143edd2010-03-24 21:43:33 -0700431 m = ceph_msg_get(req->reply);
432 /*
433 * we don't need to track the connection reading into
434 * this reply because we only have one open connection
435 * at a time, ever.
436 */
437 }
438 mutex_unlock(&monc->mutex);
439 return m;
440}
441
Yehuda Sadehe56fa102010-05-17 12:40:28 -0700442static int do_generic_request(struct ceph_mon_client *monc,
443 struct ceph_mon_generic_request *req)
444{
445 int err;
446
447 /* register request */
448 mutex_lock(&monc->mutex);
449 req->tid = ++monc->last_tid;
450 req->request->hdr.tid = cpu_to_le64(req->tid);
451 __insert_generic_request(monc, req);
452 monc->num_generic_requests++;
453 ceph_con_send(monc->con, ceph_msg_get(req->request));
454 mutex_unlock(&monc->mutex);
455
456 err = wait_for_completion_interruptible(&req->completion);
457
458 mutex_lock(&monc->mutex);
459 rb_erase(&req->node, &monc->generic_request_tree);
460 monc->num_generic_requests--;
461 mutex_unlock(&monc->mutex);
462
463 if (!err)
464 err = req->result;
465 return err;
466}
467
468/*
469 * statfs
470 */
Sage Weilba75bb92009-10-06 11:31:11 -0700471static void handle_statfs_reply(struct ceph_mon_client *monc,
472 struct ceph_msg *msg)
473{
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700474 struct ceph_mon_generic_request *req;
Sage Weilba75bb92009-10-06 11:31:11 -0700475 struct ceph_mon_statfs_reply *reply = msg->front.iov_base;
Sage Weil3143edd2010-03-24 21:43:33 -0700476 u64 tid = le64_to_cpu(msg->hdr.tid);
Sage Weilba75bb92009-10-06 11:31:11 -0700477
478 if (msg->front.iov_len != sizeof(*reply))
479 goto bad;
Sage Weilba75bb92009-10-06 11:31:11 -0700480 dout("handle_statfs_reply %p tid %llu\n", msg, tid);
481
482 mutex_lock(&monc->mutex);
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700483 req = __lookup_generic_req(monc, tid);
Sage Weilba75bb92009-10-06 11:31:11 -0700484 if (req) {
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700485 *(struct ceph_statfs *)req->buf = reply->st;
Sage Weilba75bb92009-10-06 11:31:11 -0700486 req->result = 0;
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700487 get_generic_request(req);
Sage Weilba75bb92009-10-06 11:31:11 -0700488 }
489 mutex_unlock(&monc->mutex);
Sage Weil3143edd2010-03-24 21:43:33 -0700490 if (req) {
Yehuda Sadeh03066f22010-07-27 13:11:08 -0700491 complete_all(&req->completion);
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700492 put_generic_request(req);
Sage Weil3143edd2010-03-24 21:43:33 -0700493 }
Sage Weilba75bb92009-10-06 11:31:11 -0700494 return;
495
496bad:
Yehuda Sadehe56fa102010-05-17 12:40:28 -0700497 pr_err("corrupt generic reply, tid %llu\n", tid);
Sage Weil9ec7cab2009-12-14 15:13:47 -0800498 ceph_msg_dump(msg);
Sage Weilba75bb92009-10-06 11:31:11 -0700499}
500
501/*
Sage Weilba75bb92009-10-06 11:31:11 -0700502 * Do a synchronous statfs().
503 */
504int ceph_monc_do_statfs(struct ceph_mon_client *monc, struct ceph_statfs *buf)
505{
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700506 struct ceph_mon_generic_request *req;
Sage Weil3143edd2010-03-24 21:43:33 -0700507 struct ceph_mon_statfs *h;
Sage Weilba75bb92009-10-06 11:31:11 -0700508 int err;
509
Julia Lawallcffe7b62010-05-13 22:07:29 +0200510 req = kzalloc(sizeof(*req), GFP_NOFS);
Sage Weil3143edd2010-03-24 21:43:33 -0700511 if (!req)
512 return -ENOMEM;
Sage Weilba75bb92009-10-06 11:31:11 -0700513
Sage Weil3143edd2010-03-24 21:43:33 -0700514 kref_init(&req->kref);
515 req->buf = buf;
Yehuda Sadehe56fa102010-05-17 12:40:28 -0700516 req->buf_len = sizeof(*buf);
Sage Weil3143edd2010-03-24 21:43:33 -0700517 init_completion(&req->completion);
518
Sage Weila79832f2010-04-01 16:06:19 -0700519 err = -ENOMEM;
Yehuda Sadeh34d23762010-04-06 14:33:58 -0700520 req->request = ceph_msg_new(CEPH_MSG_STATFS, sizeof(*h), GFP_NOFS);
Sage Weila79832f2010-04-01 16:06:19 -0700521 if (!req->request)
Sage Weil3143edd2010-03-24 21:43:33 -0700522 goto out;
Yehuda Sadeh34d23762010-04-06 14:33:58 -0700523 req->reply = ceph_msg_new(CEPH_MSG_STATFS_REPLY, 1024, GFP_NOFS);
Sage Weila79832f2010-04-01 16:06:19 -0700524 if (!req->reply)
Sage Weil3143edd2010-03-24 21:43:33 -0700525 goto out;
Sage Weil3143edd2010-03-24 21:43:33 -0700526
527 /* fill out request */
528 h = req->request->front.iov_base;
529 h->monhdr.have_version = 0;
530 h->monhdr.session_mon = cpu_to_le16(-1);
531 h->monhdr.session_mon_tid = 0;
532 h->fsid = monc->monmap->fsid;
Sage Weilba75bb92009-10-06 11:31:11 -0700533
Yehuda Sadehe56fa102010-05-17 12:40:28 -0700534 err = do_generic_request(monc, req);
Sage Weil3143edd2010-03-24 21:43:33 -0700535
536out:
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700537 kref_put(&req->kref, release_generic_request);
Sage Weilba75bb92009-10-06 11:31:11 -0700538 return err;
539}
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700540EXPORT_SYMBOL(ceph_monc_do_statfs);
Sage Weilba75bb92009-10-06 11:31:11 -0700541
542/*
Yehuda Sadehe56fa102010-05-17 12:40:28 -0700543 * pool ops
544 */
545static int get_poolop_reply_buf(const char *src, size_t src_len,
546 char *dst, size_t dst_len)
547{
548 u32 buf_len;
549
550 if (src_len != sizeof(u32) + dst_len)
551 return -EINVAL;
552
553 buf_len = le32_to_cpu(*(u32 *)src);
554 if (buf_len != dst_len)
555 return -EINVAL;
556
557 memcpy(dst, src + sizeof(u32), dst_len);
558 return 0;
559}
560
561static void handle_poolop_reply(struct ceph_mon_client *monc,
562 struct ceph_msg *msg)
563{
564 struct ceph_mon_generic_request *req;
565 struct ceph_mon_poolop_reply *reply = msg->front.iov_base;
566 u64 tid = le64_to_cpu(msg->hdr.tid);
567
568 if (msg->front.iov_len < sizeof(*reply))
569 goto bad;
570 dout("handle_poolop_reply %p tid %llu\n", msg, tid);
571
572 mutex_lock(&monc->mutex);
573 req = __lookup_generic_req(monc, tid);
574 if (req) {
575 if (req->buf_len &&
576 get_poolop_reply_buf(msg->front.iov_base + sizeof(*reply),
577 msg->front.iov_len - sizeof(*reply),
578 req->buf, req->buf_len) < 0) {
579 mutex_unlock(&monc->mutex);
580 goto bad;
581 }
582 req->result = le32_to_cpu(reply->reply_code);
583 get_generic_request(req);
584 }
585 mutex_unlock(&monc->mutex);
586 if (req) {
587 complete(&req->completion);
588 put_generic_request(req);
589 }
590 return;
591
592bad:
593 pr_err("corrupt generic reply, tid %llu\n", tid);
594 ceph_msg_dump(msg);
595}
596
597/*
598 * Do a synchronous pool op.
599 */
600int ceph_monc_do_poolop(struct ceph_mon_client *monc, u32 op,
601 u32 pool, u64 snapid,
602 char *buf, int len)
603{
604 struct ceph_mon_generic_request *req;
605 struct ceph_mon_poolop *h;
606 int err;
607
608 req = kzalloc(sizeof(*req), GFP_NOFS);
609 if (!req)
610 return -ENOMEM;
611
612 kref_init(&req->kref);
613 req->buf = buf;
614 req->buf_len = len;
615 init_completion(&req->completion);
616
617 err = -ENOMEM;
618 req->request = ceph_msg_new(CEPH_MSG_POOLOP, sizeof(*h), GFP_NOFS);
619 if (!req->request)
620 goto out;
621 req->reply = ceph_msg_new(CEPH_MSG_POOLOP_REPLY, 1024, GFP_NOFS);
622 if (!req->reply)
623 goto out;
624
625 /* fill out request */
626 req->request->hdr.version = cpu_to_le16(2);
627 h = req->request->front.iov_base;
628 h->monhdr.have_version = 0;
629 h->monhdr.session_mon = cpu_to_le16(-1);
630 h->monhdr.session_mon_tid = 0;
631 h->fsid = monc->monmap->fsid;
632 h->pool = cpu_to_le32(pool);
633 h->op = cpu_to_le32(op);
634 h->auid = 0;
635 h->snapid = cpu_to_le64(snapid);
636 h->name_len = 0;
637
638 err = do_generic_request(monc, req);
639
640out:
641 kref_put(&req->kref, release_generic_request);
642 return err;
643}
644
645int ceph_monc_create_snapid(struct ceph_mon_client *monc,
646 u32 pool, u64 *snapid)
647{
648 return ceph_monc_do_poolop(monc, POOL_OP_CREATE_UNMANAGED_SNAP,
649 pool, 0, (char *)snapid, sizeof(*snapid));
650
651}
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700652EXPORT_SYMBOL(ceph_monc_create_snapid);
Yehuda Sadehe56fa102010-05-17 12:40:28 -0700653
654int ceph_monc_delete_snapid(struct ceph_mon_client *monc,
655 u32 pool, u64 snapid)
656{
657 return ceph_monc_do_poolop(monc, POOL_OP_CREATE_UNMANAGED_SNAP,
658 pool, snapid, 0, 0);
659
660}
661
662/*
663 * Resend pending generic requests.
Sage Weilba75bb92009-10-06 11:31:11 -0700664 */
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700665static void __resend_generic_request(struct ceph_mon_client *monc)
Sage Weilba75bb92009-10-06 11:31:11 -0700666{
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700667 struct ceph_mon_generic_request *req;
Sage Weil85ff03f2010-02-15 14:47:28 -0800668 struct rb_node *p;
Sage Weilba75bb92009-10-06 11:31:11 -0700669
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700670 for (p = rb_first(&monc->generic_request_tree); p; p = rb_next(p)) {
671 req = rb_entry(p, struct ceph_mon_generic_request, node);
Sage Weil97069002010-05-21 12:31:49 -0700672 ceph_con_revoke(monc->con, req->request);
Sage Weil3143edd2010-03-24 21:43:33 -0700673 ceph_con_send(monc->con, ceph_msg_get(req->request));
Sage Weilba75bb92009-10-06 11:31:11 -0700674 }
675}
676
677/*
678 * Delayed work. If we haven't mounted yet, retry. Otherwise,
679 * renew/retry subscription as needed (in case it is timing out, or we
680 * got an ENOMEM). And keep the monitor connection alive.
681 */
682static void delayed_work(struct work_struct *work)
683{
684 struct ceph_mon_client *monc =
685 container_of(work, struct ceph_mon_client, delayed_work.work);
686
687 dout("monc delayed_work\n");
688 mutex_lock(&monc->mutex);
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800689 if (monc->hunting) {
690 __close_session(monc);
691 __open_session(monc); /* continue hunting */
Sage Weilba75bb92009-10-06 11:31:11 -0700692 } else {
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800693 ceph_con_keepalive(monc->con);
Sage Weil9bd2e6f2010-02-02 16:21:06 -0800694
695 __validate_auth(monc);
696
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800697 if (monc->auth->ops->is_authenticated(monc->auth))
698 __send_subscribe(monc);
Sage Weilba75bb92009-10-06 11:31:11 -0700699 }
Sage Weilba75bb92009-10-06 11:31:11 -0700700 __schedule_delayed(monc);
701 mutex_unlock(&monc->mutex);
702}
703
Sage Weil6b805182009-10-27 11:50:50 -0700704/*
705 * On startup, we build a temporary monmap populated with the IPs
706 * provided by mount(2).
707 */
708static int build_initial_monmap(struct ceph_mon_client *monc)
709{
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700710 struct ceph_options *opt = monc->client->options;
711 struct ceph_entity_addr *mon_addr = opt->mon_addr;
712 int num_mon = opt->num_mon;
Sage Weil6b805182009-10-27 11:50:50 -0700713 int i;
714
715 /* build initial monmap */
716 monc->monmap = kzalloc(sizeof(*monc->monmap) +
717 num_mon*sizeof(monc->monmap->mon_inst[0]),
718 GFP_KERNEL);
719 if (!monc->monmap)
720 return -ENOMEM;
721 for (i = 0; i < num_mon; i++) {
722 monc->monmap->mon_inst[i].addr = mon_addr[i];
Sage Weil6b805182009-10-27 11:50:50 -0700723 monc->monmap->mon_inst[i].addr.nonce = 0;
724 monc->monmap->mon_inst[i].name.type =
725 CEPH_ENTITY_TYPE_MON;
726 monc->monmap->mon_inst[i].name.num = cpu_to_le64(i);
727 }
728 monc->monmap->num_mon = num_mon;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800729 monc->have_fsid = false;
Sage Weil6b805182009-10-27 11:50:50 -0700730 return 0;
731}
732
Sage Weilba75bb92009-10-06 11:31:11 -0700733int ceph_monc_init(struct ceph_mon_client *monc, struct ceph_client *cl)
734{
735 int err = 0;
736
737 dout("init\n");
738 memset(monc, 0, sizeof(*monc));
739 monc->client = cl;
740 monc->monmap = NULL;
741 mutex_init(&monc->mutex);
742
Sage Weil6b805182009-10-27 11:50:50 -0700743 err = build_initial_monmap(monc);
744 if (err)
745 goto out;
746
Sage Weilf6a2f5b2011-08-09 09:27:44 -0700747 /* connection */
748 monc->con = kmalloc(sizeof(*monc->con), GFP_KERNEL);
749 if (!monc->con)
750 goto out_monmap;
751 ceph_con_init(monc->client->msgr, monc->con);
752 monc->con->private = monc;
753 monc->con->ops = &mon_con_ops;
Sage Weilba75bb92009-10-06 11:31:11 -0700754
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800755 /* authentication */
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700756 monc->auth = ceph_auth_init(cl->options->name,
Tommi Virtanen8323c3a2011-03-25 16:32:57 -0700757 cl->options->key);
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800758 if (IS_ERR(monc->auth))
759 return PTR_ERR(monc->auth);
760 monc->auth->want_keys =
761 CEPH_ENTITY_TYPE_AUTH | CEPH_ENTITY_TYPE_MON |
762 CEPH_ENTITY_TYPE_OSD | CEPH_ENTITY_TYPE_MDS;
763
Sage Weil240ed682010-05-21 14:57:25 -0700764 /* msgs */
Sage Weila79832f2010-04-01 16:06:19 -0700765 err = -ENOMEM;
Sage Weil7c315c52010-03-24 21:52:30 -0700766 monc->m_subscribe_ack = ceph_msg_new(CEPH_MSG_MON_SUBSCRIBE_ACK,
Yehuda Sadeh34d23762010-04-06 14:33:58 -0700767 sizeof(struct ceph_mon_subscribe_ack),
768 GFP_NOFS);
Sage Weila79832f2010-04-01 16:06:19 -0700769 if (!monc->m_subscribe_ack)
Sage Weilf6a2f5b2011-08-09 09:27:44 -0700770 goto out_con;
Sage Weil6694d6b2010-03-24 21:48:05 -0700771
Sage Weil240ed682010-05-21 14:57:25 -0700772 monc->m_subscribe = ceph_msg_new(CEPH_MSG_MON_SUBSCRIBE, 96, GFP_NOFS);
773 if (!monc->m_subscribe)
774 goto out_subscribe_ack;
775
Yehuda Sadeh34d23762010-04-06 14:33:58 -0700776 monc->m_auth_reply = ceph_msg_new(CEPH_MSG_AUTH_REPLY, 4096, GFP_NOFS);
Sage Weila79832f2010-04-01 16:06:19 -0700777 if (!monc->m_auth_reply)
Sage Weil240ed682010-05-21 14:57:25 -0700778 goto out_subscribe;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800779
Yehuda Sadeh34d23762010-04-06 14:33:58 -0700780 monc->m_auth = ceph_msg_new(CEPH_MSG_AUTH, 4096, GFP_NOFS);
Sage Weil9bd2e6f2010-02-02 16:21:06 -0800781 monc->pending_auth = 0;
Sage Weila79832f2010-04-01 16:06:19 -0700782 if (!monc->m_auth)
Sage Weil6694d6b2010-03-24 21:48:05 -0700783 goto out_auth_reply;
Sage Weilba75bb92009-10-06 11:31:11 -0700784
785 monc->cur_mon = -1;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800786 monc->hunting = true;
Sage Weilba75bb92009-10-06 11:31:11 -0700787 monc->sub_renew_after = jiffies;
788 monc->sub_sent = 0;
789
790 INIT_DELAYED_WORK(&monc->delayed_work, delayed_work);
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700791 monc->generic_request_tree = RB_ROOT;
792 monc->num_generic_requests = 0;
Sage Weilba75bb92009-10-06 11:31:11 -0700793 monc->last_tid = 0;
794
795 monc->have_mdsmap = 0;
796 monc->have_osdmap = 0;
797 monc->want_next_osdmap = 1;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800798 return 0;
799
Sage Weil6694d6b2010-03-24 21:48:05 -0700800out_auth_reply:
801 ceph_msg_put(monc->m_auth_reply);
Sage Weil240ed682010-05-21 14:57:25 -0700802out_subscribe:
803 ceph_msg_put(monc->m_subscribe);
Sage Weil7c315c52010-03-24 21:52:30 -0700804out_subscribe_ack:
805 ceph_msg_put(monc->m_subscribe_ack);
Sage Weilf6a2f5b2011-08-09 09:27:44 -0700806out_con:
807 monc->con->ops->put(monc->con);
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800808out_monmap:
809 kfree(monc->monmap);
Sage Weilba75bb92009-10-06 11:31:11 -0700810out:
811 return err;
812}
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700813EXPORT_SYMBOL(ceph_monc_init);
Sage Weilba75bb92009-10-06 11:31:11 -0700814
815void ceph_monc_stop(struct ceph_mon_client *monc)
816{
817 dout("stop\n");
818 cancel_delayed_work_sync(&monc->delayed_work);
819
820 mutex_lock(&monc->mutex);
821 __close_session(monc);
Sage Weilf6a2f5b2011-08-09 09:27:44 -0700822
823 monc->con->private = NULL;
824 monc->con->ops->put(monc->con);
825 monc->con = NULL;
826
Sage Weilba75bb92009-10-06 11:31:11 -0700827 mutex_unlock(&monc->mutex);
828
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800829 ceph_auth_destroy(monc->auth);
830
831 ceph_msg_put(monc->m_auth);
Sage Weil6694d6b2010-03-24 21:48:05 -0700832 ceph_msg_put(monc->m_auth_reply);
Sage Weil240ed682010-05-21 14:57:25 -0700833 ceph_msg_put(monc->m_subscribe);
Sage Weil7c315c52010-03-24 21:52:30 -0700834 ceph_msg_put(monc->m_subscribe_ack);
Sage Weilba75bb92009-10-06 11:31:11 -0700835
836 kfree(monc->monmap);
837}
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700838EXPORT_SYMBOL(ceph_monc_stop);
Sage Weilba75bb92009-10-06 11:31:11 -0700839
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800840static void handle_auth_reply(struct ceph_mon_client *monc,
841 struct ceph_msg *msg)
842{
843 int ret;
Sage Weil09c4d6a2010-05-25 15:38:06 -0700844 int was_auth = 0;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800845
846 mutex_lock(&monc->mutex);
Sage Weil09c4d6a2010-05-25 15:38:06 -0700847 if (monc->auth->ops)
848 was_auth = monc->auth->ops->is_authenticated(monc->auth);
Sage Weil9bd2e6f2010-02-02 16:21:06 -0800849 monc->pending_auth = 0;
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800850 ret = ceph_handle_auth_reply(monc->auth, msg->front.iov_base,
851 msg->front.iov_len,
852 monc->m_auth->front.iov_base,
853 monc->m_auth->front_max);
854 if (ret < 0) {
Sage Weil9bd2e6f2010-02-02 16:21:06 -0800855 monc->client->auth_err = ret;
Yehuda Sadeh03066f22010-07-27 13:11:08 -0700856 wake_up_all(&monc->client->auth_wq);
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800857 } else if (ret > 0) {
Sage Weil9bd2e6f2010-02-02 16:21:06 -0800858 __send_prepared_auth_request(monc, ret);
Sage Weil09c4d6a2010-05-25 15:38:06 -0700859 } else if (!was_auth && monc->auth->ops->is_authenticated(monc->auth)) {
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800860 dout("authenticated, starting session\n");
Sage Weil07433042009-11-18 16:50:41 -0800861
862 monc->client->msgr->inst.name.type = CEPH_ENTITY_TYPE_CLIENT;
Yehuda Sadeh0cf55372010-06-11 15:57:06 -0700863 monc->client->msgr->inst.name.num =
864 cpu_to_le64(monc->auth->global_id);
Sage Weil07433042009-11-18 16:50:41 -0800865
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800866 __send_subscribe(monc);
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700867 __resend_generic_request(monc);
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800868 }
869 mutex_unlock(&monc->mutex);
870}
871
Sage Weil9bd2e6f2010-02-02 16:21:06 -0800872static int __validate_auth(struct ceph_mon_client *monc)
873{
874 int ret;
875
876 if (monc->pending_auth)
877 return 0;
878
879 ret = ceph_build_auth(monc->auth, monc->m_auth->front.iov_base,
880 monc->m_auth->front_max);
881 if (ret <= 0)
882 return ret; /* either an error, or no need to authenticate */
883 __send_prepared_auth_request(monc, ret);
884 return 0;
885}
886
887int ceph_monc_validate_auth(struct ceph_mon_client *monc)
888{
889 int ret;
890
891 mutex_lock(&monc->mutex);
892 ret = __validate_auth(monc);
893 mutex_unlock(&monc->mutex);
894 return ret;
895}
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700896EXPORT_SYMBOL(ceph_monc_validate_auth);
Sage Weil9bd2e6f2010-02-02 16:21:06 -0800897
Sage Weilba75bb92009-10-06 11:31:11 -0700898/*
899 * handle incoming message
900 */
901static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
902{
903 struct ceph_mon_client *monc = con->private;
904 int type = le16_to_cpu(msg->hdr.type);
905
906 if (!monc)
907 return;
908
909 switch (type) {
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800910 case CEPH_MSG_AUTH_REPLY:
911 handle_auth_reply(monc, msg);
Sage Weilba75bb92009-10-06 11:31:11 -0700912 break;
913
914 case CEPH_MSG_MON_SUBSCRIBE_ACK:
915 handle_subscribe_ack(monc, msg);
916 break;
917
918 case CEPH_MSG_STATFS_REPLY:
919 handle_statfs_reply(monc, msg);
920 break;
921
Yehuda Sadehe56fa102010-05-17 12:40:28 -0700922 case CEPH_MSG_POOLOP_REPLY:
923 handle_poolop_reply(monc, msg);
924 break;
925
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800926 case CEPH_MSG_MON_MAP:
927 ceph_monc_handle_map(monc, msg);
928 break;
929
Sage Weilba75bb92009-10-06 11:31:11 -0700930 case CEPH_MSG_OSD_MAP:
931 ceph_osdc_handle_map(&monc->client->osdc, msg);
932 break;
933
934 default:
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700935 /* can the chained handler handle it? */
936 if (monc->client->extra_mon_dispatch &&
937 monc->client->extra_mon_dispatch(monc->client, msg) == 0)
938 break;
939
Sage Weilba75bb92009-10-06 11:31:11 -0700940 pr_err("received unknown message type %d %s\n", type,
941 ceph_msg_type_name(type));
942 }
943 ceph_msg_put(msg);
944}
945
946/*
947 * Allocate memory for incoming message
948 */
949static struct ceph_msg *mon_alloc_msg(struct ceph_connection *con,
Yehuda Sadeh24504182010-01-08 13:58:34 -0800950 struct ceph_msg_header *hdr,
951 int *skip)
Sage Weilba75bb92009-10-06 11:31:11 -0700952{
953 struct ceph_mon_client *monc = con->private;
954 int type = le16_to_cpu(hdr->type);
Yehuda Sadeh24504182010-01-08 13:58:34 -0800955 int front_len = le32_to_cpu(hdr->front_len);
Sage Weil5b3a4db2010-02-19 21:43:23 -0800956 struct ceph_msg *m = NULL;
Sage Weilba75bb92009-10-06 11:31:11 -0700957
Yehuda Sadeh24504182010-01-08 13:58:34 -0800958 *skip = 0;
Yehuda Sadeh0547a9b2010-01-11 14:47:13 -0800959
Sage Weilba75bb92009-10-06 11:31:11 -0700960 switch (type) {
Sage Weilba75bb92009-10-06 11:31:11 -0700961 case CEPH_MSG_MON_SUBSCRIBE_ACK:
Sage Weil7c315c52010-03-24 21:52:30 -0700962 m = ceph_msg_get(monc->m_subscribe_ack);
Yehuda Sadeh24504182010-01-08 13:58:34 -0800963 break;
Yehuda Sadehe56fa102010-05-17 12:40:28 -0700964 case CEPH_MSG_POOLOP_REPLY:
Sage Weilba75bb92009-10-06 11:31:11 -0700965 case CEPH_MSG_STATFS_REPLY:
Yehuda Sadehf8c76f62010-04-22 15:40:37 -0700966 return get_generic_reply(con, hdr, skip);
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800967 case CEPH_MSG_AUTH_REPLY:
Sage Weil6694d6b2010-03-24 21:48:05 -0700968 m = ceph_msg_get(monc->m_auth_reply);
Yehuda Sadeh24504182010-01-08 13:58:34 -0800969 break;
Sage Weil5b3a4db2010-02-19 21:43:23 -0800970 case CEPH_MSG_MON_MAP:
971 case CEPH_MSG_MDS_MAP:
972 case CEPH_MSG_OSD_MAP:
Yehuda Sadeh34d23762010-04-06 14:33:58 -0700973 m = ceph_msg_new(type, front_len, GFP_NOFS);
Sage Weil5b3a4db2010-02-19 21:43:23 -0800974 break;
Sage Weilba75bb92009-10-06 11:31:11 -0700975 }
Yehuda Sadeh24504182010-01-08 13:58:34 -0800976
Sage Weil5b3a4db2010-02-19 21:43:23 -0800977 if (!m) {
978 pr_info("alloc_msg unknown type %d\n", type);
Yehuda Sadeh24504182010-01-08 13:58:34 -0800979 *skip = 1;
Sage Weil5b3a4db2010-02-19 21:43:23 -0800980 }
Yehuda Sadeh24504182010-01-08 13:58:34 -0800981 return m;
Sage Weilba75bb92009-10-06 11:31:11 -0700982}
983
984/*
985 * If the monitor connection resets, pick a new monitor and resubmit
986 * any pending requests.
987 */
988static void mon_fault(struct ceph_connection *con)
989{
990 struct ceph_mon_client *monc = con->private;
991
992 if (!monc)
993 return;
994
995 dout("mon_fault\n");
996 mutex_lock(&monc->mutex);
997 if (!con->private)
998 goto out;
999
Sage Weilf6a2f5b2011-08-09 09:27:44 -07001000 if (!monc->hunting)
Sage Weilba75bb92009-10-06 11:31:11 -07001001 pr_info("mon%d %s session lost, "
1002 "hunting for new mon\n", monc->cur_mon,
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07001003 ceph_pr_addr(&monc->con->peer_addr.in_addr));
Sage Weilba75bb92009-10-06 11:31:11 -07001004
1005 __close_session(monc);
1006 if (!monc->hunting) {
1007 /* start hunting */
1008 monc->hunting = true;
Sage Weil4e7a5dc2009-11-18 16:19:57 -08001009 __open_session(monc);
Sage Weilba75bb92009-10-06 11:31:11 -07001010 } else {
1011 /* already hunting, let's wait a bit */
1012 __schedule_delayed(monc);
1013 }
1014out:
1015 mutex_unlock(&monc->mutex);
1016}
1017
Tobias Klauser9e327892010-05-20 10:40:19 +02001018static const struct ceph_connection_operations mon_con_ops = {
Sage Weilba75bb92009-10-06 11:31:11 -07001019 .get = ceph_con_get,
1020 .put = ceph_con_put,
1021 .dispatch = dispatch,
1022 .fault = mon_fault,
1023 .alloc_msg = mon_alloc_msg,
Sage Weilba75bb92009-10-06 11:31:11 -07001024};