35b36b86d762892a118b7e34e88b654979db30a1
[linux-2.6.git] / net / ceph / messenger.c
1 #include <linux/ceph/ceph_debug.h>
2
3 #include <linux/crc32c.h>
4 #include <linux/ctype.h>
5 #include <linux/highmem.h>
6 #include <linux/inet.h>
7 #include <linux/kthread.h>
8 #include <linux/net.h>
9 #include <linux/slab.h>
10 #include <linux/socket.h>
11 #include <linux/string.h>
12 #include <linux/bio.h>
13 #include <linux/blkdev.h>
14 #include <net/tcp.h>
15
16 #include <linux/ceph/libceph.h>
17 #include <linux/ceph/messenger.h>
18 #include <linux/ceph/decode.h>
19 #include <linux/ceph/pagelist.h>
20
21 /*
22  * Ceph uses the messenger to exchange ceph_msg messages with other
23  * hosts in the system.  The messenger provides ordered and reliable
24  * delivery.  We tolerate TCP disconnects by reconnecting (with
25  * exponential backoff) in the case of a fault (disconnection, bad
26  * crc, protocol error).  Acks allow sent messages to be discarded by
27  * the sender.
28  */
29
30 /* static tag bytes (protocol control messages) */
31 static char tag_msg = CEPH_MSGR_TAG_MSG;
32 static char tag_ack = CEPH_MSGR_TAG_ACK;
33 static char tag_keepalive = CEPH_MSGR_TAG_KEEPALIVE;
34
35 #ifdef CONFIG_LOCKDEP
36 static struct lock_class_key socket_class;
37 #endif
38
39
40 static void queue_con(struct ceph_connection *con);
41 static void con_work(struct work_struct *);
42 static void ceph_fault(struct ceph_connection *con);
43
44 /*
45  * nicely render a sockaddr as a string.
46  */
47 #define MAX_ADDR_STR 20
48 #define MAX_ADDR_STR_LEN 60
49 static char addr_str[MAX_ADDR_STR][MAX_ADDR_STR_LEN];
50 static DEFINE_SPINLOCK(addr_str_lock);
51 static int last_addr_str;
52
53 const char *ceph_pr_addr(const struct sockaddr_storage *ss)
54 {
55         int i;
56         char *s;
57         struct sockaddr_in *in4 = (void *)ss;
58         struct sockaddr_in6 *in6 = (void *)ss;
59
60         spin_lock(&addr_str_lock);
61         i = last_addr_str++;
62         if (last_addr_str == MAX_ADDR_STR)
63                 last_addr_str = 0;
64         spin_unlock(&addr_str_lock);
65         s = addr_str[i];
66
67         switch (ss->ss_family) {
68         case AF_INET:
69                 snprintf(s, MAX_ADDR_STR_LEN, "%pI4:%u", &in4->sin_addr,
70                          (unsigned int)ntohs(in4->sin_port));
71                 break;
72
73         case AF_INET6:
74                 snprintf(s, MAX_ADDR_STR_LEN, "[%pI6c]:%u", &in6->sin6_addr,
75                          (unsigned int)ntohs(in6->sin6_port));
76                 break;
77
78         default:
79                 sprintf(s, "(unknown sockaddr family %d)", (int)ss->ss_family);
80         }
81
82         return s;
83 }
84 EXPORT_SYMBOL(ceph_pr_addr);
85
86 static void encode_my_addr(struct ceph_messenger *msgr)
87 {
88         memcpy(&msgr->my_enc_addr, &msgr->inst.addr, sizeof(msgr->my_enc_addr));
89         ceph_encode_addr(&msgr->my_enc_addr);
90 }
91
92 /*
93  * work queue for all reading and writing to/from the socket.
94  */
95 struct workqueue_struct *ceph_msgr_wq;
96
97 int ceph_msgr_init(void)
98 {
99         ceph_msgr_wq = alloc_workqueue("ceph-msgr", WQ_NON_REENTRANT, 0);
100         if (!ceph_msgr_wq) {
101                 pr_err("msgr_init failed to create workqueue\n");
102                 return -ENOMEM;
103         }
104         return 0;
105 }
106 EXPORT_SYMBOL(ceph_msgr_init);
107
108 void ceph_msgr_exit(void)
109 {
110         destroy_workqueue(ceph_msgr_wq);
111 }
112 EXPORT_SYMBOL(ceph_msgr_exit);
113
114 void ceph_msgr_flush(void)
115 {
116         flush_workqueue(ceph_msgr_wq);
117 }
118 EXPORT_SYMBOL(ceph_msgr_flush);
119
120
121 /*
122  * socket callback functions
123  */
124
125 /* data available on socket, or listen socket received a connect */
126 static void ceph_data_ready(struct sock *sk, int count_unused)
127 {
128         struct ceph_connection *con =
129                 (struct ceph_connection *)sk->sk_user_data;
130         if (sk->sk_state != TCP_CLOSE_WAIT) {
131                 dout("ceph_data_ready on %p state = %lu, queueing work\n",
132                      con, con->state);
133                 queue_con(con);
134         }
135 }
136
137 /* socket has buffer space for writing */
138 static void ceph_write_space(struct sock *sk)
139 {
140         struct ceph_connection *con =
141                 (struct ceph_connection *)sk->sk_user_data;
142
143         /* only queue to workqueue if there is data we want to write. */
144         if (test_bit(WRITE_PENDING, &con->state)) {
145                 dout("ceph_write_space %p queueing write work\n", con);
146                 queue_con(con);
147         } else {
148                 dout("ceph_write_space %p nothing to write\n", con);
149         }
150
151         /* since we have our own write_space, clear the SOCK_NOSPACE flag */
152         clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
153 }
154
155 /* socket's state has changed */
156 static void ceph_state_change(struct sock *sk)
157 {
158         struct ceph_connection *con =
159                 (struct ceph_connection *)sk->sk_user_data;
160
161         dout("ceph_state_change %p state = %lu sk_state = %u\n",
162              con, con->state, sk->sk_state);
163
164         if (test_bit(CLOSED, &con->state))
165                 return;
166
167         switch (sk->sk_state) {
168         case TCP_CLOSE:
169                 dout("ceph_state_change TCP_CLOSE\n");
170         case TCP_CLOSE_WAIT:
171                 dout("ceph_state_change TCP_CLOSE_WAIT\n");
172                 if (test_and_set_bit(SOCK_CLOSED, &con->state) == 0) {
173                         if (test_bit(CONNECTING, &con->state))
174                                 con->error_msg = "connection failed";
175                         else
176                                 con->error_msg = "socket closed";
177                         queue_con(con);
178                 }
179                 break;
180         case TCP_ESTABLISHED:
181                 dout("ceph_state_change TCP_ESTABLISHED\n");
182                 queue_con(con);
183                 break;
184         }
185 }
186
187 /*
188  * set up socket callbacks
189  */
190 static void set_sock_callbacks(struct socket *sock,
191                                struct ceph_connection *con)
192 {
193         struct sock *sk = sock->sk;
194         sk->sk_user_data = (void *)con;
195         sk->sk_data_ready = ceph_data_ready;
196         sk->sk_write_space = ceph_write_space;
197         sk->sk_state_change = ceph_state_change;
198 }
199
200
201 /*
202  * socket helpers
203  */
204
205 /*
206  * initiate connection to a remote socket.
207  */
208 static struct socket *ceph_tcp_connect(struct ceph_connection *con)
209 {
210         struct sockaddr_storage *paddr = &con->peer_addr.in_addr;
211         struct socket *sock;
212         int ret;
213
214         BUG_ON(con->sock);
215         ret = sock_create_kern(con->peer_addr.in_addr.ss_family, SOCK_STREAM,
216                                IPPROTO_TCP, &sock);
217         if (ret)
218                 return ERR_PTR(ret);
219         con->sock = sock;
220         sock->sk->sk_allocation = GFP_NOFS;
221
222 #ifdef CONFIG_LOCKDEP
223         lockdep_set_class(&sock->sk->sk_lock, &socket_class);
224 #endif
225
226         set_sock_callbacks(sock, con);
227
228         dout("connect %s\n", ceph_pr_addr(&con->peer_addr.in_addr));
229
230         ret = sock->ops->connect(sock, (struct sockaddr *)paddr, sizeof(*paddr),
231                                  O_NONBLOCK);
232         if (ret == -EINPROGRESS) {
233                 dout("connect %s EINPROGRESS sk_state = %u\n",
234                      ceph_pr_addr(&con->peer_addr.in_addr),
235                      sock->sk->sk_state);
236                 ret = 0;
237         }
238         if (ret < 0) {
239                 pr_err("connect %s error %d\n",
240                        ceph_pr_addr(&con->peer_addr.in_addr), ret);
241                 sock_release(sock);
242                 con->sock = NULL;
243                 con->error_msg = "connect error";
244         }
245
246         if (ret < 0)
247                 return ERR_PTR(ret);
248         return sock;
249 }
250
251 static int ceph_tcp_recvmsg(struct socket *sock, void *buf, size_t len)
252 {
253         struct kvec iov = {buf, len};
254         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
255         int r;
256
257         r = kernel_recvmsg(sock, &msg, &iov, 1, len, msg.msg_flags);
258         if (r == -EAGAIN)
259                 r = 0;
260         return r;
261 }
262
263 /*
264  * write something.  @more is true if caller will be sending more data
265  * shortly.
266  */
267 static int ceph_tcp_sendmsg(struct socket *sock, struct kvec *iov,
268                      size_t kvlen, size_t len, int more)
269 {
270         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
271         int r;
272
273         if (more)
274                 msg.msg_flags |= MSG_MORE;
275         else
276                 msg.msg_flags |= MSG_EOR;  /* superfluous, but what the hell */
277
278         r = kernel_sendmsg(sock, &msg, iov, kvlen, len);
279         if (r == -EAGAIN)
280                 r = 0;
281         return r;
282 }
283
284
285 /*
286  * Shutdown/close the socket for the given connection.
287  */
288 static int con_close_socket(struct ceph_connection *con)
289 {
290         int rc;
291
292         dout("con_close_socket on %p sock %p\n", con, con->sock);
293         if (!con->sock)
294                 return 0;
295         set_bit(SOCK_CLOSED, &con->state);
296         rc = con->sock->ops->shutdown(con->sock, SHUT_RDWR);
297         sock_release(con->sock);
298         con->sock = NULL;
299         clear_bit(SOCK_CLOSED, &con->state);
300         return rc;
301 }
302
303 /*
304  * Reset a connection.  Discard all incoming and outgoing messages
305  * and clear *_seq state.
306  */
307 static void ceph_msg_remove(struct ceph_msg *msg)
308 {
309         list_del_init(&msg->list_head);
310         ceph_msg_put(msg);
311 }
312 static void ceph_msg_remove_list(struct list_head *head)
313 {
314         while (!list_empty(head)) {
315                 struct ceph_msg *msg = list_first_entry(head, struct ceph_msg,
316                                                         list_head);
317                 ceph_msg_remove(msg);
318         }
319 }
320
321 static void reset_connection(struct ceph_connection *con)
322 {
323         /* reset connection, out_queue, msg_ and connect_seq */
324         /* discard existing out_queue and msg_seq */
325         ceph_msg_remove_list(&con->out_queue);
326         ceph_msg_remove_list(&con->out_sent);
327
328         if (con->in_msg) {
329                 ceph_msg_put(con->in_msg);
330                 con->in_msg = NULL;
331         }
332
333         con->connect_seq = 0;
334         con->out_seq = 0;
335         if (con->out_msg) {
336                 ceph_msg_put(con->out_msg);
337                 con->out_msg = NULL;
338         }
339         con->out_keepalive_pending = false;
340         con->in_seq = 0;
341         con->in_seq_acked = 0;
342 }
343
344 /*
345  * mark a peer down.  drop any open connections.
346  */
347 void ceph_con_close(struct ceph_connection *con)
348 {
349         dout("con_close %p peer %s\n", con,
350              ceph_pr_addr(&con->peer_addr.in_addr));
351         set_bit(CLOSED, &con->state);  /* in case there's queued work */
352         clear_bit(STANDBY, &con->state);  /* avoid connect_seq bump */
353         clear_bit(LOSSYTX, &con->state);  /* so we retry next connect */
354         clear_bit(KEEPALIVE_PENDING, &con->state);
355         clear_bit(WRITE_PENDING, &con->state);
356         mutex_lock(&con->mutex);
357         reset_connection(con);
358         con->peer_global_seq = 0;
359         cancel_delayed_work(&con->work);
360         mutex_unlock(&con->mutex);
361         queue_con(con);
362 }
363 EXPORT_SYMBOL(ceph_con_close);
364
365 /*
366  * Reopen a closed connection, with a new peer address.
367  */
368 void ceph_con_open(struct ceph_connection *con, struct ceph_entity_addr *addr)
369 {
370         dout("con_open %p %s\n", con, ceph_pr_addr(&addr->in_addr));
371         set_bit(OPENING, &con->state);
372         clear_bit(CLOSED, &con->state);
373         memcpy(&con->peer_addr, addr, sizeof(*addr));
374         con->delay = 0;      /* reset backoff memory */
375         queue_con(con);
376 }
377 EXPORT_SYMBOL(ceph_con_open);
378
379 /*
380  * return true if this connection ever successfully opened
381  */
382 bool ceph_con_opened(struct ceph_connection *con)
383 {
384         return con->connect_seq > 0;
385 }
386
387 /*
388  * generic get/put
389  */
390 struct ceph_connection *ceph_con_get(struct ceph_connection *con)
391 {
392         dout("con_get %p nref = %d -> %d\n", con,
393              atomic_read(&con->nref), atomic_read(&con->nref) + 1);
394         if (atomic_inc_not_zero(&con->nref))
395                 return con;
396         return NULL;
397 }
398
399 void ceph_con_put(struct ceph_connection *con)
400 {
401         dout("con_put %p nref = %d -> %d\n", con,
402              atomic_read(&con->nref), atomic_read(&con->nref) - 1);
403         BUG_ON(atomic_read(&con->nref) == 0);
404         if (atomic_dec_and_test(&con->nref)) {
405                 BUG_ON(con->sock);
406                 kfree(con);
407         }
408 }
409
410 /*
411  * initialize a new connection.
412  */
413 void ceph_con_init(struct ceph_messenger *msgr, struct ceph_connection *con)
414 {
415         dout("con_init %p\n", con);
416         memset(con, 0, sizeof(*con));
417         atomic_set(&con->nref, 1);
418         con->msgr = msgr;
419         mutex_init(&con->mutex);
420         INIT_LIST_HEAD(&con->out_queue);
421         INIT_LIST_HEAD(&con->out_sent);
422         INIT_DELAYED_WORK(&con->work, con_work);
423 }
424 EXPORT_SYMBOL(ceph_con_init);
425
426
427 /*
428  * We maintain a global counter to order connection attempts.  Get
429  * a unique seq greater than @gt.
430  */
431 static u32 get_global_seq(struct ceph_messenger *msgr, u32 gt)
432 {
433         u32 ret;
434
435         spin_lock(&msgr->global_seq_lock);
436         if (msgr->global_seq < gt)
437                 msgr->global_seq = gt;
438         ret = ++msgr->global_seq;
439         spin_unlock(&msgr->global_seq_lock);
440         return ret;
441 }
442
443
444 /*
445  * Prepare footer for currently outgoing message, and finish things
446  * off.  Assumes out_kvec* are already valid.. we just add on to the end.
447  */
448 static void prepare_write_message_footer(struct ceph_connection *con, int v)
449 {
450         struct ceph_msg *m = con->out_msg;
451
452         dout("prepare_write_message_footer %p\n", con);
453         con->out_kvec_is_msg = true;
454         con->out_kvec[v].iov_base = &m->footer;
455         con->out_kvec[v].iov_len = sizeof(m->footer);
456         con->out_kvec_bytes += sizeof(m->footer);
457         con->out_kvec_left++;
458         con->out_more = m->more_to_follow;
459         con->out_msg_done = true;
460 }
461
462 /*
463  * Prepare headers for the next outgoing message.
464  */
465 static void prepare_write_message(struct ceph_connection *con)
466 {
467         struct ceph_msg *m;
468         int v = 0;
469
470         con->out_kvec_bytes = 0;
471         con->out_kvec_is_msg = true;
472         con->out_msg_done = false;
473
474         /* Sneak an ack in there first?  If we can get it into the same
475          * TCP packet that's a good thing. */
476         if (con->in_seq > con->in_seq_acked) {
477                 con->in_seq_acked = con->in_seq;
478                 con->out_kvec[v].iov_base = &tag_ack;
479                 con->out_kvec[v++].iov_len = 1;
480                 con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
481                 con->out_kvec[v].iov_base = &con->out_temp_ack;
482                 con->out_kvec[v++].iov_len = sizeof(con->out_temp_ack);
483                 con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
484         }
485
486         m = list_first_entry(&con->out_queue,
487                        struct ceph_msg, list_head);
488         con->out_msg = m;
489         if (test_bit(LOSSYTX, &con->state)) {
490                 list_del_init(&m->list_head);
491         } else {
492                 /* put message on sent list */
493                 ceph_msg_get(m);
494                 list_move_tail(&m->list_head, &con->out_sent);
495         }
496
497         /*
498          * only assign outgoing seq # if we haven't sent this message
499          * yet.  if it is requeued, resend with it's original seq.
500          */
501         if (m->needs_out_seq) {
502                 m->hdr.seq = cpu_to_le64(++con->out_seq);
503                 m->needs_out_seq = false;
504         }
505
506         dout("prepare_write_message %p seq %lld type %d len %d+%d+%d %d pgs\n",
507              m, con->out_seq, le16_to_cpu(m->hdr.type),
508              le32_to_cpu(m->hdr.front_len), le32_to_cpu(m->hdr.middle_len),
509              le32_to_cpu(m->hdr.data_len),
510              m->nr_pages);
511         BUG_ON(le32_to_cpu(m->hdr.front_len) != m->front.iov_len);
512
513         /* tag + hdr + front + middle */
514         con->out_kvec[v].iov_base = &tag_msg;
515         con->out_kvec[v++].iov_len = 1;
516         con->out_kvec[v].iov_base = &m->hdr;
517         con->out_kvec[v++].iov_len = sizeof(m->hdr);
518         con->out_kvec[v++] = m->front;
519         if (m->middle)
520                 con->out_kvec[v++] = m->middle->vec;
521         con->out_kvec_left = v;
522         con->out_kvec_bytes += 1 + sizeof(m->hdr) + m->front.iov_len +
523                 (m->middle ? m->middle->vec.iov_len : 0);
524         con->out_kvec_cur = con->out_kvec;
525
526         /* fill in crc (except data pages), footer */
527         con->out_msg->hdr.crc =
528                 cpu_to_le32(crc32c(0, (void *)&m->hdr,
529                                       sizeof(m->hdr) - sizeof(m->hdr.crc)));
530         con->out_msg->footer.flags = CEPH_MSG_FOOTER_COMPLETE;
531         con->out_msg->footer.front_crc =
532                 cpu_to_le32(crc32c(0, m->front.iov_base, m->front.iov_len));
533         if (m->middle)
534                 con->out_msg->footer.middle_crc =
535                         cpu_to_le32(crc32c(0, m->middle->vec.iov_base,
536                                            m->middle->vec.iov_len));
537         else
538                 con->out_msg->footer.middle_crc = 0;
539         con->out_msg->footer.data_crc = 0;
540         dout("prepare_write_message front_crc %u data_crc %u\n",
541              le32_to_cpu(con->out_msg->footer.front_crc),
542              le32_to_cpu(con->out_msg->footer.middle_crc));
543
544         /* is there a data payload? */
545         if (le32_to_cpu(m->hdr.data_len) > 0) {
546                 /* initialize page iterator */
547                 con->out_msg_pos.page = 0;
548                 if (m->pages)
549                         con->out_msg_pos.page_pos = m->page_alignment;
550                 else
551                         con->out_msg_pos.page_pos = 0;
552                 con->out_msg_pos.data_pos = 0;
553                 con->out_msg_pos.did_page_crc = 0;
554                 con->out_more = 1;  /* data + footer will follow */
555         } else {
556                 /* no, queue up footer too and be done */
557                 prepare_write_message_footer(con, v);
558         }
559
560         set_bit(WRITE_PENDING, &con->state);
561 }
562
563 /*
564  * Prepare an ack.
565  */
566 static void prepare_write_ack(struct ceph_connection *con)
567 {
568         dout("prepare_write_ack %p %llu -> %llu\n", con,
569              con->in_seq_acked, con->in_seq);
570         con->in_seq_acked = con->in_seq;
571
572         con->out_kvec[0].iov_base = &tag_ack;
573         con->out_kvec[0].iov_len = 1;
574         con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
575         con->out_kvec[1].iov_base = &con->out_temp_ack;
576         con->out_kvec[1].iov_len = sizeof(con->out_temp_ack);
577         con->out_kvec_left = 2;
578         con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
579         con->out_kvec_cur = con->out_kvec;
580         con->out_more = 1;  /* more will follow.. eventually.. */
581         set_bit(WRITE_PENDING, &con->state);
582 }
583
584 /*
585  * Prepare to write keepalive byte.
586  */
587 static void prepare_write_keepalive(struct ceph_connection *con)
588 {
589         dout("prepare_write_keepalive %p\n", con);
590         con->out_kvec[0].iov_base = &tag_keepalive;
591         con->out_kvec[0].iov_len = 1;
592         con->out_kvec_left = 1;
593         con->out_kvec_bytes = 1;
594         con->out_kvec_cur = con->out_kvec;
595         set_bit(WRITE_PENDING, &con->state);
596 }
597
598 /*
599  * Connection negotiation.
600  */
601
602 static void prepare_connect_authorizer(struct ceph_connection *con)
603 {
604         void *auth_buf;
605         int auth_len = 0;
606         int auth_protocol = 0;
607
608         mutex_unlock(&con->mutex);
609         if (con->ops->get_authorizer)
610                 con->ops->get_authorizer(con, &auth_buf, &auth_len,
611                                          &auth_protocol, &con->auth_reply_buf,
612                                          &con->auth_reply_buf_len,
613                                          con->auth_retry);
614         mutex_lock(&con->mutex);
615
616         con->out_connect.authorizer_protocol = cpu_to_le32(auth_protocol);
617         con->out_connect.authorizer_len = cpu_to_le32(auth_len);
618
619         con->out_kvec[con->out_kvec_left].iov_base = auth_buf;
620         con->out_kvec[con->out_kvec_left].iov_len = auth_len;
621         con->out_kvec_left++;
622         con->out_kvec_bytes += auth_len;
623 }
624
625 /*
626  * We connected to a peer and are saying hello.
627  */
628 static void prepare_write_banner(struct ceph_messenger *msgr,
629                                  struct ceph_connection *con)
630 {
631         int len = strlen(CEPH_BANNER);
632
633         con->out_kvec[0].iov_base = CEPH_BANNER;
634         con->out_kvec[0].iov_len = len;
635         con->out_kvec[1].iov_base = &msgr->my_enc_addr;
636         con->out_kvec[1].iov_len = sizeof(msgr->my_enc_addr);
637         con->out_kvec_left = 2;
638         con->out_kvec_bytes = len + sizeof(msgr->my_enc_addr);
639         con->out_kvec_cur = con->out_kvec;
640         con->out_more = 0;
641         set_bit(WRITE_PENDING, &con->state);
642 }
643
644 static void prepare_write_connect(struct ceph_messenger *msgr,
645                                   struct ceph_connection *con,
646                                   int after_banner)
647 {
648         unsigned global_seq = get_global_seq(con->msgr, 0);
649         int proto;
650
651         switch (con->peer_name.type) {
652         case CEPH_ENTITY_TYPE_MON:
653                 proto = CEPH_MONC_PROTOCOL;
654                 break;
655         case CEPH_ENTITY_TYPE_OSD:
656                 proto = CEPH_OSDC_PROTOCOL;
657                 break;
658         case CEPH_ENTITY_TYPE_MDS:
659                 proto = CEPH_MDSC_PROTOCOL;
660                 break;
661         default:
662                 BUG();
663         }
664
665         dout("prepare_write_connect %p cseq=%d gseq=%d proto=%d\n", con,
666              con->connect_seq, global_seq, proto);
667
668         con->out_connect.features = cpu_to_le64(msgr->supported_features);
669         con->out_connect.host_type = cpu_to_le32(CEPH_ENTITY_TYPE_CLIENT);
670         con->out_connect.connect_seq = cpu_to_le32(con->connect_seq);
671         con->out_connect.global_seq = cpu_to_le32(global_seq);
672         con->out_connect.protocol_version = cpu_to_le32(proto);
673         con->out_connect.flags = 0;
674
675         if (!after_banner) {
676                 con->out_kvec_left = 0;
677                 con->out_kvec_bytes = 0;
678         }
679         con->out_kvec[con->out_kvec_left].iov_base = &con->out_connect;
680         con->out_kvec[con->out_kvec_left].iov_len = sizeof(con->out_connect);
681         con->out_kvec_left++;
682         con->out_kvec_bytes += sizeof(con->out_connect);
683         con->out_kvec_cur = con->out_kvec;
684         con->out_more = 0;
685         set_bit(WRITE_PENDING, &con->state);
686
687         prepare_connect_authorizer(con);
688 }
689
690
691 /*
692  * write as much of pending kvecs to the socket as we can.
693  *  1 -> done
694  *  0 -> socket full, but more to do
695  * <0 -> error
696  */
697 static int write_partial_kvec(struct ceph_connection *con)
698 {
699         int ret;
700
701         dout("write_partial_kvec %p %d left\n", con, con->out_kvec_bytes);
702         while (con->out_kvec_bytes > 0) {
703                 ret = ceph_tcp_sendmsg(con->sock, con->out_kvec_cur,
704                                        con->out_kvec_left, con->out_kvec_bytes,
705                                        con->out_more);
706                 if (ret <= 0)
707                         goto out;
708                 con->out_kvec_bytes -= ret;
709                 if (con->out_kvec_bytes == 0)
710                         break;            /* done */
711                 while (ret > 0) {
712                         if (ret >= con->out_kvec_cur->iov_len) {
713                                 ret -= con->out_kvec_cur->iov_len;
714                                 con->out_kvec_cur++;
715                                 con->out_kvec_left--;
716                         } else {
717                                 con->out_kvec_cur->iov_len -= ret;
718                                 con->out_kvec_cur->iov_base += ret;
719                                 ret = 0;
720                                 break;
721                         }
722                 }
723         }
724         con->out_kvec_left = 0;
725         con->out_kvec_is_msg = false;
726         ret = 1;
727 out:
728         dout("write_partial_kvec %p %d left in %d kvecs ret = %d\n", con,
729              con->out_kvec_bytes, con->out_kvec_left, ret);
730         return ret;  /* done! */
731 }
732
733 #ifdef CONFIG_BLOCK
734 static void init_bio_iter(struct bio *bio, struct bio **iter, int *seg)
735 {
736         if (!bio) {
737                 *iter = NULL;
738                 *seg = 0;
739                 return;
740         }
741         *iter = bio;
742         *seg = bio->bi_idx;
743 }
744
745 static void iter_bio_next(struct bio **bio_iter, int *seg)
746 {
747         if (*bio_iter == NULL)
748                 return;
749
750         BUG_ON(*seg >= (*bio_iter)->bi_vcnt);
751
752         (*seg)++;
753         if (*seg == (*bio_iter)->bi_vcnt)
754                 init_bio_iter((*bio_iter)->bi_next, bio_iter, seg);
755 }
756 #endif
757
758 /*
759  * Write as much message data payload as we can.  If we finish, queue
760  * up the footer.
761  *  1 -> done, footer is now queued in out_kvec[].
762  *  0 -> socket full, but more to do
763  * <0 -> error
764  */
765 static int write_partial_msg_pages(struct ceph_connection *con)
766 {
767         struct ceph_msg *msg = con->out_msg;
768         unsigned data_len = le32_to_cpu(msg->hdr.data_len);
769         size_t len;
770         int crc = con->msgr->nocrc;
771         int ret;
772         int total_max_write;
773         int in_trail = 0;
774         size_t trail_len = (msg->trail ? msg->trail->length : 0);
775
776         dout("write_partial_msg_pages %p msg %p page %d/%d offset %d\n",
777              con, con->out_msg, con->out_msg_pos.page, con->out_msg->nr_pages,
778              con->out_msg_pos.page_pos);
779
780 #ifdef CONFIG_BLOCK
781         if (msg->bio && !msg->bio_iter)
782                 init_bio_iter(msg->bio, &msg->bio_iter, &msg->bio_seg);
783 #endif
784
785         while (data_len > con->out_msg_pos.data_pos) {
786                 struct page *page = NULL;
787                 void *kaddr = NULL;
788                 int max_write = PAGE_SIZE;
789                 int page_shift = 0;
790
791                 total_max_write = data_len - trail_len -
792                         con->out_msg_pos.data_pos;
793
794                 /*
795                  * if we are calculating the data crc (the default), we need
796                  * to map the page.  if our pages[] has been revoked, use the
797                  * zero page.
798                  */
799
800                 /* have we reached the trail part of the data? */
801                 if (con->out_msg_pos.data_pos >= data_len - trail_len) {
802                         in_trail = 1;
803
804                         total_max_write = data_len - con->out_msg_pos.data_pos;
805
806                         page = list_first_entry(&msg->trail->head,
807                                                 struct page, lru);
808                         if (crc)
809                                 kaddr = kmap(page);
810                         max_write = PAGE_SIZE;
811                 } else if (msg->pages) {
812                         page = msg->pages[con->out_msg_pos.page];
813                         if (crc)
814                                 kaddr = kmap(page);
815                 } else if (msg->pagelist) {
816                         page = list_first_entry(&msg->pagelist->head,
817                                                 struct page, lru);
818                         if (crc)
819                                 kaddr = kmap(page);
820 #ifdef CONFIG_BLOCK
821                 } else if (msg->bio) {
822                         struct bio_vec *bv;
823
824                         bv = bio_iovec_idx(msg->bio_iter, msg->bio_seg);
825                         page = bv->bv_page;
826                         page_shift = bv->bv_offset;
827                         if (crc)
828                                 kaddr = kmap(page) + page_shift;
829                         max_write = bv->bv_len;
830 #endif
831                 } else {
832                         page = con->msgr->zero_page;
833                         if (crc)
834                                 kaddr = page_address(con->msgr->zero_page);
835                 }
836                 len = min_t(int, max_write - con->out_msg_pos.page_pos,
837                             total_max_write);
838
839                 if (crc && !con->out_msg_pos.did_page_crc) {
840                         void *base = kaddr + con->out_msg_pos.page_pos;
841                         u32 tmpcrc = le32_to_cpu(con->out_msg->footer.data_crc);
842
843                         BUG_ON(kaddr == NULL);
844                         con->out_msg->footer.data_crc =
845                                 cpu_to_le32(crc32c(tmpcrc, base, len));
846                         con->out_msg_pos.did_page_crc = 1;
847                 }
848                 ret = kernel_sendpage(con->sock, page,
849                                       con->out_msg_pos.page_pos + page_shift,
850                                       len,
851                                       MSG_DONTWAIT | MSG_NOSIGNAL |
852                                       MSG_MORE);
853
854                 if (crc &&
855                     (msg->pages || msg->pagelist || msg->bio || in_trail))
856                         kunmap(page);
857
858                 if (ret == -EAGAIN)
859                         ret = 0;
860                 if (ret <= 0)
861                         goto out;
862
863                 con->out_msg_pos.data_pos += ret;
864                 con->out_msg_pos.page_pos += ret;
865                 if (ret == len) {
866                         con->out_msg_pos.page_pos = 0;
867                         con->out_msg_pos.page++;
868                         con->out_msg_pos.did_page_crc = 0;
869                         if (in_trail)
870                                 list_move_tail(&page->lru,
871                                                &msg->trail->head);
872                         else if (msg->pagelist)
873                                 list_move_tail(&page->lru,
874                                                &msg->pagelist->head);
875 #ifdef CONFIG_BLOCK
876                         else if (msg->bio)
877                                 iter_bio_next(&msg->bio_iter, &msg->bio_seg);
878 #endif
879                 }
880         }
881
882         dout("write_partial_msg_pages %p msg %p done\n", con, msg);
883
884         /* prepare and queue up footer, too */
885         if (!crc)
886                 con->out_msg->footer.flags |= CEPH_MSG_FOOTER_NOCRC;
887         con->out_kvec_bytes = 0;
888         con->out_kvec_left = 0;
889         con->out_kvec_cur = con->out_kvec;
890         prepare_write_message_footer(con, 0);
891         ret = 1;
892 out:
893         return ret;
894 }
895
896 /*
897  * write some zeros
898  */
899 static int write_partial_skip(struct ceph_connection *con)
900 {
901         int ret;
902
903         while (con->out_skip > 0) {
904                 struct kvec iov = {
905                         .iov_base = page_address(con->msgr->zero_page),
906                         .iov_len = min(con->out_skip, (int)PAGE_CACHE_SIZE)
907                 };
908
909                 ret = ceph_tcp_sendmsg(con->sock, &iov, 1, iov.iov_len, 1);
910                 if (ret <= 0)
911                         goto out;
912                 con->out_skip -= ret;
913         }
914         ret = 1;
915 out:
916         return ret;
917 }
918
919 /*
920  * Prepare to read connection handshake, or an ack.
921  */
922 static void prepare_read_banner(struct ceph_connection *con)
923 {
924         dout("prepare_read_banner %p\n", con);
925         con->in_base_pos = 0;
926 }
927
928 static void prepare_read_connect(struct ceph_connection *con)
929 {
930         dout("prepare_read_connect %p\n", con);
931         con->in_base_pos = 0;
932 }
933
934 static void prepare_read_ack(struct ceph_connection *con)
935 {
936         dout("prepare_read_ack %p\n", con);
937         con->in_base_pos = 0;
938 }
939
940 static void prepare_read_tag(struct ceph_connection *con)
941 {
942         dout("prepare_read_tag %p\n", con);
943         con->in_base_pos = 0;
944         con->in_tag = CEPH_MSGR_TAG_READY;
945 }
946
947 /*
948  * Prepare to read a message.
949  */
950 static int prepare_read_message(struct ceph_connection *con)
951 {
952         dout("prepare_read_message %p\n", con);
953         BUG_ON(con->in_msg != NULL);
954         con->in_base_pos = 0;
955         con->in_front_crc = con->in_middle_crc = con->in_data_crc = 0;
956         return 0;
957 }
958
959
960 static int read_partial(struct ceph_connection *con,
961                         int *to, int size, void *object)
962 {
963         *to += size;
964         while (con->in_base_pos < *to) {
965                 int left = *to - con->in_base_pos;
966                 int have = size - left;
967                 int ret = ceph_tcp_recvmsg(con->sock, object + have, left);
968                 if (ret <= 0)
969                         return ret;
970                 con->in_base_pos += ret;
971         }
972         return 1;
973 }
974
975
976 /*
977  * Read all or part of the connect-side handshake on a new connection
978  */
979 static int read_partial_banner(struct ceph_connection *con)
980 {
981         int ret, to = 0;
982
983         dout("read_partial_banner %p at %d\n", con, con->in_base_pos);
984
985         /* peer's banner */
986         ret = read_partial(con, &to, strlen(CEPH_BANNER), con->in_banner);
987         if (ret <= 0)
988                 goto out;
989         ret = read_partial(con, &to, sizeof(con->actual_peer_addr),
990                            &con->actual_peer_addr);
991         if (ret <= 0)
992                 goto out;
993         ret = read_partial(con, &to, sizeof(con->peer_addr_for_me),
994                            &con->peer_addr_for_me);
995         if (ret <= 0)
996                 goto out;
997 out:
998         return ret;
999 }
1000
1001 static int read_partial_connect(struct ceph_connection *con)
1002 {
1003         int ret, to = 0;
1004
1005         dout("read_partial_connect %p at %d\n", con, con->in_base_pos);
1006
1007         ret = read_partial(con, &to, sizeof(con->in_reply), &con->in_reply);
1008         if (ret <= 0)
1009                 goto out;
1010         ret = read_partial(con, &to, le32_to_cpu(con->in_reply.authorizer_len),
1011                            con->auth_reply_buf);
1012         if (ret <= 0)
1013                 goto out;
1014
1015         dout("read_partial_connect %p tag %d, con_seq = %u, g_seq = %u\n",
1016              con, (int)con->in_reply.tag,
1017              le32_to_cpu(con->in_reply.connect_seq),
1018              le32_to_cpu(con->in_reply.global_seq));
1019 out:
1020         return ret;
1021
1022 }
1023
1024 /*
1025  * Verify the hello banner looks okay.
1026  */
1027 static int verify_hello(struct ceph_connection *con)
1028 {
1029         if (memcmp(con->in_banner, CEPH_BANNER, strlen(CEPH_BANNER))) {
1030                 pr_err("connect to %s got bad banner\n",
1031                        ceph_pr_addr(&con->peer_addr.in_addr));
1032                 con->error_msg = "protocol error, bad banner";
1033                 return -1;
1034         }
1035         return 0;
1036 }
1037
1038 static bool addr_is_blank(struct sockaddr_storage *ss)
1039 {
1040         switch (ss->ss_family) {
1041         case AF_INET:
1042                 return ((struct sockaddr_in *)ss)->sin_addr.s_addr == 0;
1043         case AF_INET6:
1044                 return
1045                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[0] == 0 &&
1046                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[1] == 0 &&
1047                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[2] == 0 &&
1048                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[3] == 0;
1049         }
1050         return false;
1051 }
1052
1053 static int addr_port(struct sockaddr_storage *ss)
1054 {
1055         switch (ss->ss_family) {
1056         case AF_INET:
1057                 return ntohs(((struct sockaddr_in *)ss)->sin_port);
1058         case AF_INET6:
1059                 return ntohs(((struct sockaddr_in6 *)ss)->sin6_port);
1060         }
1061         return 0;
1062 }
1063
1064 static void addr_set_port(struct sockaddr_storage *ss, int p)
1065 {
1066         switch (ss->ss_family) {
1067         case AF_INET:
1068                 ((struct sockaddr_in *)ss)->sin_port = htons(p);
1069         case AF_INET6:
1070                 ((struct sockaddr_in6 *)ss)->sin6_port = htons(p);
1071         }
1072 }
1073
1074 /*
1075  * Parse an ip[:port] list into an addr array.  Use the default
1076  * monitor port if a port isn't specified.
1077  */
1078 int ceph_parse_ips(const char *c, const char *end,
1079                    struct ceph_entity_addr *addr,
1080                    int max_count, int *count)
1081 {
1082         int i;
1083         const char *p = c;
1084
1085         dout("parse_ips on '%.*s'\n", (int)(end-c), c);
1086         for (i = 0; i < max_count; i++) {
1087                 const char *ipend;
1088                 struct sockaddr_storage *ss = &addr[i].in_addr;
1089                 struct sockaddr_in *in4 = (void *)ss;
1090                 struct sockaddr_in6 *in6 = (void *)ss;
1091                 int port;
1092                 char delim = ',';
1093
1094                 if (*p == '[') {
1095                         delim = ']';
1096                         p++;
1097                 }
1098
1099                 memset(ss, 0, sizeof(*ss));
1100                 if (in4_pton(p, end - p, (u8 *)&in4->sin_addr.s_addr,
1101                              delim, &ipend))
1102                         ss->ss_family = AF_INET;
1103                 else if (in6_pton(p, end - p, (u8 *)&in6->sin6_addr.s6_addr,
1104                                   delim, &ipend))
1105                         ss->ss_family = AF_INET6;
1106                 else
1107                         goto bad;
1108                 p = ipend;
1109
1110                 if (delim == ']') {
1111                         if (*p != ']') {
1112                                 dout("missing matching ']'\n");
1113                                 goto bad;
1114                         }
1115                         p++;
1116                 }
1117
1118                 /* port? */
1119                 if (p < end && *p == ':') {
1120                         port = 0;
1121                         p++;
1122                         while (p < end && *p >= '0' && *p <= '9') {
1123                                 port = (port * 10) + (*p - '0');
1124                                 p++;
1125                         }
1126                         if (port > 65535 || port == 0)
1127                                 goto bad;
1128                 } else {
1129                         port = CEPH_MON_PORT;
1130                 }
1131
1132                 addr_set_port(ss, port);
1133
1134                 dout("parse_ips got %s\n", ceph_pr_addr(ss));
1135
1136                 if (p == end)
1137                         break;
1138                 if (*p != ',')
1139                         goto bad;
1140                 p++;
1141         }
1142
1143         if (p != end)
1144                 goto bad;
1145
1146         if (count)
1147                 *count = i + 1;
1148         return 0;
1149
1150 bad:
1151         pr_err("parse_ips bad ip '%.*s'\n", (int)(end - c), c);
1152         return -EINVAL;
1153 }
1154 EXPORT_SYMBOL(ceph_parse_ips);
1155
1156 static int process_banner(struct ceph_connection *con)
1157 {
1158         dout("process_banner on %p\n", con);
1159
1160         if (verify_hello(con) < 0)
1161                 return -1;
1162
1163         ceph_decode_addr(&con->actual_peer_addr);
1164         ceph_decode_addr(&con->peer_addr_for_me);
1165
1166         /*
1167          * Make sure the other end is who we wanted.  note that the other
1168          * end may not yet know their ip address, so if it's 0.0.0.0, give
1169          * them the benefit of the doubt.
1170          */
1171         if (memcmp(&con->peer_addr, &con->actual_peer_addr,
1172                    sizeof(con->peer_addr)) != 0 &&
1173             !(addr_is_blank(&con->actual_peer_addr.in_addr) &&
1174               con->actual_peer_addr.nonce == con->peer_addr.nonce)) {
1175                 pr_warning("wrong peer, want %s/%d, got %s/%d\n",
1176                            ceph_pr_addr(&con->peer_addr.in_addr),
1177                            (int)le32_to_cpu(con->peer_addr.nonce),
1178                            ceph_pr_addr(&con->actual_peer_addr.in_addr),
1179                            (int)le32_to_cpu(con->actual_peer_addr.nonce));
1180                 con->error_msg = "wrong peer at address";
1181                 return -1;
1182         }
1183
1184         /*
1185          * did we learn our address?
1186          */
1187         if (addr_is_blank(&con->msgr->inst.addr.in_addr)) {
1188                 int port = addr_port(&con->msgr->inst.addr.in_addr);
1189
1190                 memcpy(&con->msgr->inst.addr.in_addr,
1191                        &con->peer_addr_for_me.in_addr,
1192                        sizeof(con->peer_addr_for_me.in_addr));
1193                 addr_set_port(&con->msgr->inst.addr.in_addr, port);
1194                 encode_my_addr(con->msgr);
1195                 dout("process_banner learned my addr is %s\n",
1196                      ceph_pr_addr(&con->msgr->inst.addr.in_addr));
1197         }
1198
1199         set_bit(NEGOTIATING, &con->state);
1200         prepare_read_connect(con);
1201         return 0;
1202 }
1203
1204 static void fail_protocol(struct ceph_connection *con)
1205 {
1206         reset_connection(con);
1207         set_bit(CLOSED, &con->state);  /* in case there's queued work */
1208
1209         mutex_unlock(&con->mutex);
1210         if (con->ops->bad_proto)
1211                 con->ops->bad_proto(con);
1212         mutex_lock(&con->mutex);
1213 }
1214
1215 static int process_connect(struct ceph_connection *con)
1216 {
1217         u64 sup_feat = con->msgr->supported_features;
1218         u64 req_feat = con->msgr->required_features;
1219         u64 server_feat = le64_to_cpu(con->in_reply.features);
1220
1221         dout("process_connect on %p tag %d\n", con, (int)con->in_tag);
1222
1223         switch (con->in_reply.tag) {
1224         case CEPH_MSGR_TAG_FEATURES:
1225                 pr_err("%s%lld %s feature set mismatch,"
1226                        " my %llx < server's %llx, missing %llx\n",
1227                        ENTITY_NAME(con->peer_name),
1228                        ceph_pr_addr(&con->peer_addr.in_addr),
1229                        sup_feat, server_feat, server_feat & ~sup_feat);
1230                 con->error_msg = "missing required protocol features";
1231                 fail_protocol(con);
1232                 return -1;
1233
1234         case CEPH_MSGR_TAG_BADPROTOVER:
1235                 pr_err("%s%lld %s protocol version mismatch,"
1236                        " my %d != server's %d\n",
1237                        ENTITY_NAME(con->peer_name),
1238                        ceph_pr_addr(&con->peer_addr.in_addr),
1239                        le32_to_cpu(con->out_connect.protocol_version),
1240                        le32_to_cpu(con->in_reply.protocol_version));
1241                 con->error_msg = "protocol version mismatch";
1242                 fail_protocol(con);
1243                 return -1;
1244
1245         case CEPH_MSGR_TAG_BADAUTHORIZER:
1246                 con->auth_retry++;
1247                 dout("process_connect %p got BADAUTHORIZER attempt %d\n", con,
1248                      con->auth_retry);
1249                 if (con->auth_retry == 2) {
1250                         con->error_msg = "connect authorization failure";
1251                         reset_connection(con);
1252                         set_bit(CLOSED, &con->state);
1253                         return -1;
1254                 }
1255                 con->auth_retry = 1;
1256                 prepare_write_connect(con->msgr, con, 0);
1257                 prepare_read_connect(con);
1258                 break;
1259
1260         case CEPH_MSGR_TAG_RESETSESSION:
1261                 /*
1262                  * If we connected with a large connect_seq but the peer
1263                  * has no record of a session with us (no connection, or
1264                  * connect_seq == 0), they will send RESETSESION to indicate
1265                  * that they must have reset their session, and may have
1266                  * dropped messages.
1267                  */
1268                 dout("process_connect got RESET peer seq %u\n",
1269                      le32_to_cpu(con->in_connect.connect_seq));
1270                 pr_err("%s%lld %s connection reset\n",
1271                        ENTITY_NAME(con->peer_name),
1272                        ceph_pr_addr(&con->peer_addr.in_addr));
1273                 reset_connection(con);
1274                 prepare_write_connect(con->msgr, con, 0);
1275                 prepare_read_connect(con);
1276
1277                 /* Tell ceph about it. */
1278                 mutex_unlock(&con->mutex);
1279                 pr_info("reset on %s%lld\n", ENTITY_NAME(con->peer_name));
1280                 if (con->ops->peer_reset)
1281                         con->ops->peer_reset(con);
1282                 mutex_lock(&con->mutex);
1283                 break;
1284
1285         case CEPH_MSGR_TAG_RETRY_SESSION:
1286                 /*
1287                  * If we sent a smaller connect_seq than the peer has, try
1288                  * again with a larger value.
1289                  */
1290                 dout("process_connect got RETRY my seq = %u, peer_seq = %u\n",
1291                      le32_to_cpu(con->out_connect.connect_seq),
1292                      le32_to_cpu(con->in_connect.connect_seq));
1293                 con->connect_seq = le32_to_cpu(con->in_connect.connect_seq);
1294                 prepare_write_connect(con->msgr, con, 0);
1295                 prepare_read_connect(con);
1296                 break;
1297
1298         case CEPH_MSGR_TAG_RETRY_GLOBAL:
1299                 /*
1300                  * If we sent a smaller global_seq than the peer has, try
1301                  * again with a larger value.
1302                  */
1303                 dout("process_connect got RETRY_GLOBAL my %u peer_gseq %u\n",
1304                      con->peer_global_seq,
1305                      le32_to_cpu(con->in_connect.global_seq));
1306                 get_global_seq(con->msgr,
1307                                le32_to_cpu(con->in_connect.global_seq));
1308                 prepare_write_connect(con->msgr, con, 0);
1309                 prepare_read_connect(con);
1310                 break;
1311
1312         case CEPH_MSGR_TAG_READY:
1313                 if (req_feat & ~server_feat) {
1314                         pr_err("%s%lld %s protocol feature mismatch,"
1315                                " my required %llx > server's %llx, need %llx\n",
1316                                ENTITY_NAME(con->peer_name),
1317                                ceph_pr_addr(&con->peer_addr.in_addr),
1318                                req_feat, server_feat, req_feat & ~server_feat);
1319                         con->error_msg = "missing required protocol features";
1320                         fail_protocol(con);
1321                         return -1;
1322                 }
1323                 clear_bit(CONNECTING, &con->state);
1324                 con->peer_global_seq = le32_to_cpu(con->in_reply.global_seq);
1325                 con->connect_seq++;
1326                 con->peer_features = server_feat;
1327                 dout("process_connect got READY gseq %d cseq %d (%d)\n",
1328                      con->peer_global_seq,
1329                      le32_to_cpu(con->in_reply.connect_seq),
1330                      con->connect_seq);
1331                 WARN_ON(con->connect_seq !=
1332                         le32_to_cpu(con->in_reply.connect_seq));
1333
1334                 if (con->in_reply.flags & CEPH_MSG_CONNECT_LOSSY)
1335                         set_bit(LOSSYTX, &con->state);
1336
1337                 prepare_read_tag(con);
1338                 break;
1339
1340         case CEPH_MSGR_TAG_WAIT:
1341                 /*
1342                  * If there is a connection race (we are opening
1343                  * connections to each other), one of us may just have
1344                  * to WAIT.  This shouldn't happen if we are the
1345                  * client.
1346                  */
1347                 pr_err("process_connect peer connecting WAIT\n");
1348
1349         default:
1350                 pr_err("connect protocol error, will retry\n");
1351                 con->error_msg = "protocol error, garbage tag during connect";
1352                 return -1;
1353         }
1354         return 0;
1355 }
1356
1357
1358 /*
1359  * read (part of) an ack
1360  */
1361 static int read_partial_ack(struct ceph_connection *con)
1362 {
1363         int to = 0;
1364
1365         return read_partial(con, &to, sizeof(con->in_temp_ack),
1366                             &con->in_temp_ack);
1367 }
1368
1369
1370 /*
1371  * We can finally discard anything that's been acked.
1372  */
1373 static void process_ack(struct ceph_connection *con)
1374 {
1375         struct ceph_msg *m;
1376         u64 ack = le64_to_cpu(con->in_temp_ack);
1377         u64 seq;
1378
1379         while (!list_empty(&con->out_sent)) {
1380                 m = list_first_entry(&con->out_sent, struct ceph_msg,
1381                                      list_head);
1382                 seq = le64_to_cpu(m->hdr.seq);
1383                 if (seq > ack)
1384                         break;
1385                 dout("got ack for seq %llu type %d at %p\n", seq,
1386                      le16_to_cpu(m->hdr.type), m);
1387                 ceph_msg_remove(m);
1388         }
1389         prepare_read_tag(con);
1390 }
1391
1392
1393
1394
1395 static int read_partial_message_section(struct ceph_connection *con,
1396                                         struct kvec *section,
1397                                         unsigned int sec_len, u32 *crc)
1398 {
1399         int ret, left;
1400
1401         BUG_ON(!section);
1402
1403         while (section->iov_len < sec_len) {
1404                 BUG_ON(section->iov_base == NULL);
1405                 left = sec_len - section->iov_len;
1406                 ret = ceph_tcp_recvmsg(con->sock, (char *)section->iov_base +
1407                                        section->iov_len, left);
1408                 if (ret <= 0)
1409                         return ret;
1410                 section->iov_len += ret;
1411                 if (section->iov_len == sec_len)
1412                         *crc = crc32c(0, section->iov_base,
1413                                       section->iov_len);
1414         }
1415
1416         return 1;
1417 }
1418
1419 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
1420                                 struct ceph_msg_header *hdr,
1421                                 int *skip);
1422
1423
1424 static int read_partial_message_pages(struct ceph_connection *con,
1425                                       struct page **pages,
1426                                       unsigned data_len, int datacrc)
1427 {
1428         void *p;
1429         int ret;
1430         int left;
1431
1432         left = min((int)(data_len - con->in_msg_pos.data_pos),
1433                    (int)(PAGE_SIZE - con->in_msg_pos.page_pos));
1434         /* (page) data */
1435         BUG_ON(pages == NULL);
1436         p = kmap(pages[con->in_msg_pos.page]);
1437         ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1438                                left);
1439         if (ret > 0 && datacrc)
1440                 con->in_data_crc =
1441                         crc32c(con->in_data_crc,
1442                                   p + con->in_msg_pos.page_pos, ret);
1443         kunmap(pages[con->in_msg_pos.page]);
1444         if (ret <= 0)
1445                 return ret;
1446         con->in_msg_pos.data_pos += ret;
1447         con->in_msg_pos.page_pos += ret;
1448         if (con->in_msg_pos.page_pos == PAGE_SIZE) {
1449                 con->in_msg_pos.page_pos = 0;
1450                 con->in_msg_pos.page++;
1451         }
1452
1453         return ret;
1454 }
1455
1456 #ifdef CONFIG_BLOCK
1457 static int read_partial_message_bio(struct ceph_connection *con,
1458                                     struct bio **bio_iter, int *bio_seg,
1459                                     unsigned data_len, int datacrc)
1460 {
1461         struct bio_vec *bv = bio_iovec_idx(*bio_iter, *bio_seg);
1462         void *p;
1463         int ret, left;
1464
1465         if (IS_ERR(bv))
1466                 return PTR_ERR(bv);
1467
1468         left = min((int)(data_len - con->in_msg_pos.data_pos),
1469                    (int)(bv->bv_len - con->in_msg_pos.page_pos));
1470
1471         p = kmap(bv->bv_page) + bv->bv_offset;
1472
1473         ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1474                                left);
1475         if (ret > 0 && datacrc)
1476                 con->in_data_crc =
1477                         crc32c(con->in_data_crc,
1478                                   p + con->in_msg_pos.page_pos, ret);
1479         kunmap(bv->bv_page);
1480         if (ret <= 0)
1481                 return ret;
1482         con->in_msg_pos.data_pos += ret;
1483         con->in_msg_pos.page_pos += ret;
1484         if (con->in_msg_pos.page_pos == bv->bv_len) {
1485                 con->in_msg_pos.page_pos = 0;
1486                 iter_bio_next(bio_iter, bio_seg);
1487         }
1488
1489         return ret;
1490 }
1491 #endif
1492
1493 /*
1494  * read (part of) a message.
1495  */
1496 static int read_partial_message(struct ceph_connection *con)
1497 {
1498         struct ceph_msg *m = con->in_msg;
1499         int ret;
1500         int to, left;
1501         unsigned front_len, middle_len, data_len;
1502         int datacrc = con->msgr->nocrc;
1503         int skip;
1504         u64 seq;
1505
1506         dout("read_partial_message con %p msg %p\n", con, m);
1507
1508         /* header */
1509         while (con->in_base_pos < sizeof(con->in_hdr)) {
1510                 left = sizeof(con->in_hdr) - con->in_base_pos;
1511                 ret = ceph_tcp_recvmsg(con->sock,
1512                                        (char *)&con->in_hdr + con->in_base_pos,
1513                                        left);
1514                 if (ret <= 0)
1515                         return ret;
1516                 con->in_base_pos += ret;
1517                 if (con->in_base_pos == sizeof(con->in_hdr)) {
1518                         u32 crc = crc32c(0, (void *)&con->in_hdr,
1519                                  sizeof(con->in_hdr) - sizeof(con->in_hdr.crc));
1520                         if (crc != le32_to_cpu(con->in_hdr.crc)) {
1521                                 pr_err("read_partial_message bad hdr "
1522                                        " crc %u != expected %u\n",
1523                                        crc, con->in_hdr.crc);
1524                                 return -EBADMSG;
1525                         }
1526                 }
1527         }
1528         front_len = le32_to_cpu(con->in_hdr.front_len);
1529         if (front_len > CEPH_MSG_MAX_FRONT_LEN)
1530                 return -EIO;
1531         middle_len = le32_to_cpu(con->in_hdr.middle_len);
1532         if (middle_len > CEPH_MSG_MAX_DATA_LEN)
1533                 return -EIO;
1534         data_len = le32_to_cpu(con->in_hdr.data_len);
1535         if (data_len > CEPH_MSG_MAX_DATA_LEN)
1536                 return -EIO;
1537
1538         /* verify seq# */
1539         seq = le64_to_cpu(con->in_hdr.seq);
1540         if ((s64)seq - (s64)con->in_seq < 1) {
1541                 pr_info("skipping %s%lld %s seq %lld expected %lld\n",
1542                         ENTITY_NAME(con->peer_name),
1543                         ceph_pr_addr(&con->peer_addr.in_addr),
1544                         seq, con->in_seq + 1);
1545                 con->in_base_pos = -front_len - middle_len - data_len -
1546                         sizeof(m->footer);
1547                 con->in_tag = CEPH_MSGR_TAG_READY;
1548                 return 0;
1549         } else if ((s64)seq - (s64)con->in_seq > 1) {
1550                 pr_err("read_partial_message bad seq %lld expected %lld\n",
1551                        seq, con->in_seq + 1);
1552                 con->error_msg = "bad message sequence # for incoming message";
1553                 return -EBADMSG;
1554         }
1555
1556         /* allocate message? */
1557         if (!con->in_msg) {
1558                 dout("got hdr type %d front %d data %d\n", con->in_hdr.type,
1559                      con->in_hdr.front_len, con->in_hdr.data_len);
1560                 skip = 0;
1561                 con->in_msg = ceph_alloc_msg(con, &con->in_hdr, &skip);
1562                 if (skip) {
1563                         /* skip this message */
1564                         dout("alloc_msg said skip message\n");
1565                         BUG_ON(con->in_msg);
1566                         con->in_base_pos = -front_len - middle_len - data_len -
1567                                 sizeof(m->footer);
1568                         con->in_tag = CEPH_MSGR_TAG_READY;
1569                         con->in_seq++;
1570                         return 0;
1571                 }
1572                 if (!con->in_msg) {
1573                         con->error_msg =
1574                                 "error allocating memory for incoming message";
1575                         return -ENOMEM;
1576                 }
1577                 m = con->in_msg;
1578                 m->front.iov_len = 0;    /* haven't read it yet */
1579                 if (m->middle)
1580                         m->middle->vec.iov_len = 0;
1581
1582                 con->in_msg_pos.page = 0;
1583                 if (m->pages)
1584                         con->in_msg_pos.page_pos = m->page_alignment;
1585                 else
1586                         con->in_msg_pos.page_pos = 0;
1587                 con->in_msg_pos.data_pos = 0;
1588         }
1589
1590         /* front */
1591         ret = read_partial_message_section(con, &m->front, front_len,
1592                                            &con->in_front_crc);
1593         if (ret <= 0)
1594                 return ret;
1595
1596         /* middle */
1597         if (m->middle) {
1598                 ret = read_partial_message_section(con, &m->middle->vec,
1599                                                    middle_len,
1600                                                    &con->in_middle_crc);
1601                 if (ret <= 0)
1602                         return ret;
1603         }
1604 #ifdef CONFIG_BLOCK
1605         if (m->bio && !m->bio_iter)
1606                 init_bio_iter(m->bio, &m->bio_iter, &m->bio_seg);
1607 #endif
1608
1609         /* (page) data */
1610         while (con->in_msg_pos.data_pos < data_len) {
1611                 if (m->pages) {
1612                         ret = read_partial_message_pages(con, m->pages,
1613                                                  data_len, datacrc);
1614                         if (ret <= 0)
1615                                 return ret;
1616 #ifdef CONFIG_BLOCK
1617                 } else if (m->bio) {
1618
1619                         ret = read_partial_message_bio(con,
1620                                                  &m->bio_iter, &m->bio_seg,
1621                                                  data_len, datacrc);
1622                         if (ret <= 0)
1623                                 return ret;
1624 #endif
1625                 } else {
1626                         BUG_ON(1);
1627                 }
1628         }
1629
1630         /* footer */
1631         to = sizeof(m->hdr) + sizeof(m->footer);
1632         while (con->in_base_pos < to) {
1633                 left = to - con->in_base_pos;
1634                 ret = ceph_tcp_recvmsg(con->sock, (char *)&m->footer +
1635                                        (con->in_base_pos - sizeof(m->hdr)),
1636                                        left);
1637                 if (ret <= 0)
1638                         return ret;
1639                 con->in_base_pos += ret;
1640         }
1641         dout("read_partial_message got msg %p %d (%u) + %d (%u) + %d (%u)\n",
1642              m, front_len, m->footer.front_crc, middle_len,
1643              m->footer.middle_crc, data_len, m->footer.data_crc);
1644
1645         /* crc ok? */
1646         if (con->in_front_crc != le32_to_cpu(m->footer.front_crc)) {
1647                 pr_err("read_partial_message %p front crc %u != exp. %u\n",
1648                        m, con->in_front_crc, m->footer.front_crc);
1649                 return -EBADMSG;
1650         }
1651         if (con->in_middle_crc != le32_to_cpu(m->footer.middle_crc)) {
1652                 pr_err("read_partial_message %p middle crc %u != exp %u\n",
1653                        m, con->in_middle_crc, m->footer.middle_crc);
1654                 return -EBADMSG;
1655         }
1656         if (datacrc &&
1657             (m->footer.flags & CEPH_MSG_FOOTER_NOCRC) == 0 &&
1658             con->in_data_crc != le32_to_cpu(m->footer.data_crc)) {
1659                 pr_err("read_partial_message %p data crc %u != exp. %u\n", m,
1660                        con->in_data_crc, le32_to_cpu(m->footer.data_crc));
1661                 return -EBADMSG;
1662         }
1663
1664         return 1; /* done! */
1665 }
1666
1667 /*
1668  * Process message.  This happens in the worker thread.  The callback should
1669  * be careful not to do anything that waits on other incoming messages or it
1670  * may deadlock.
1671  */
1672 static void process_message(struct ceph_connection *con)
1673 {
1674         struct ceph_msg *msg;
1675
1676         msg = con->in_msg;
1677         con->in_msg = NULL;
1678
1679         /* if first message, set peer_name */
1680         if (con->peer_name.type == 0)
1681                 con->peer_name = msg->hdr.src;
1682
1683         con->in_seq++;
1684         mutex_unlock(&con->mutex);
1685
1686         dout("===== %p %llu from %s%lld %d=%s len %d+%d (%u %u %u) =====\n",
1687              msg, le64_to_cpu(msg->hdr.seq),
1688              ENTITY_NAME(msg->hdr.src),
1689              le16_to_cpu(msg->hdr.type),
1690              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1691              le32_to_cpu(msg->hdr.front_len),
1692              le32_to_cpu(msg->hdr.data_len),
1693              con->in_front_crc, con->in_middle_crc, con->in_data_crc);
1694         con->ops->dispatch(con, msg);
1695
1696         mutex_lock(&con->mutex);
1697         prepare_read_tag(con);
1698 }
1699
1700
1701 /*
1702  * Write something to the socket.  Called in a worker thread when the
1703  * socket appears to be writeable and we have something ready to send.
1704  */
1705 static int try_write(struct ceph_connection *con)
1706 {
1707         struct ceph_messenger *msgr = con->msgr;
1708         int ret = 1;
1709
1710         dout("try_write start %p state %lu nref %d\n", con, con->state,
1711              atomic_read(&con->nref));
1712
1713 more:
1714         dout("try_write out_kvec_bytes %d\n", con->out_kvec_bytes);
1715
1716         /* open the socket first? */
1717         if (con->sock == NULL) {
1718                 /*
1719                  * if we were STANDBY and are reconnecting _this_
1720                  * connection, bump connect_seq now.  Always bump
1721                  * global_seq.
1722                  */
1723                 if (test_and_clear_bit(STANDBY, &con->state))
1724                         con->connect_seq++;
1725
1726                 prepare_write_banner(msgr, con);
1727                 prepare_write_connect(msgr, con, 1);
1728                 prepare_read_banner(con);
1729                 set_bit(CONNECTING, &con->state);
1730                 clear_bit(NEGOTIATING, &con->state);
1731
1732                 BUG_ON(con->in_msg);
1733                 con->in_tag = CEPH_MSGR_TAG_READY;
1734                 dout("try_write initiating connect on %p new state %lu\n",
1735                      con, con->state);
1736                 con->sock = ceph_tcp_connect(con);
1737                 if (IS_ERR(con->sock)) {
1738                         con->sock = NULL;
1739                         con->error_msg = "connect error";
1740                         ret = -1;
1741                         goto out;
1742                 }
1743         }
1744
1745 more_kvec:
1746         /* kvec data queued? */
1747         if (con->out_skip) {
1748                 ret = write_partial_skip(con);
1749                 if (ret <= 0)
1750                         goto out;
1751         }
1752         if (con->out_kvec_left) {
1753                 ret = write_partial_kvec(con);
1754                 if (ret <= 0)
1755                         goto out;
1756         }
1757
1758         /* msg pages? */
1759         if (con->out_msg) {
1760                 if (con->out_msg_done) {
1761                         ceph_msg_put(con->out_msg);
1762                         con->out_msg = NULL;   /* we're done with this one */
1763                         goto do_next;
1764                 }
1765
1766                 ret = write_partial_msg_pages(con);
1767                 if (ret == 1)
1768                         goto more_kvec;  /* we need to send the footer, too! */
1769                 if (ret == 0)
1770                         goto out;
1771                 if (ret < 0) {
1772                         dout("try_write write_partial_msg_pages err %d\n",
1773                              ret);
1774                         goto out;
1775                 }
1776         }
1777
1778 do_next:
1779         if (!test_bit(CONNECTING, &con->state)) {
1780                 /* is anything else pending? */
1781                 if (!list_empty(&con->out_queue)) {
1782                         prepare_write_message(con);
1783                         goto more;
1784                 }
1785                 if (con->in_seq > con->in_seq_acked) {
1786                         prepare_write_ack(con);
1787                         goto more;
1788                 }
1789                 if (test_and_clear_bit(KEEPALIVE_PENDING, &con->state)) {
1790                         prepare_write_keepalive(con);
1791                         goto more;
1792                 }
1793         }
1794
1795         /* Nothing to do! */
1796         clear_bit(WRITE_PENDING, &con->state);
1797         dout("try_write nothing else to write.\n");
1798         ret = 0;
1799 out:
1800         dout("try_write done on %p ret %d\n", con, ret);
1801         return ret;
1802 }
1803
1804
1805
1806 /*
1807  * Read what we can from the socket.
1808  */
1809 static int try_read(struct ceph_connection *con)
1810 {
1811         int ret = -1;
1812
1813         if (!con->sock)
1814                 return 0;
1815
1816         if (test_bit(STANDBY, &con->state))
1817                 return 0;
1818
1819         dout("try_read start on %p\n", con);
1820
1821 more:
1822         dout("try_read tag %d in_base_pos %d\n", (int)con->in_tag,
1823              con->in_base_pos);
1824         if (test_bit(CONNECTING, &con->state)) {
1825                 if (!test_bit(NEGOTIATING, &con->state)) {
1826                         dout("try_read connecting\n");
1827                         ret = read_partial_banner(con);
1828                         if (ret <= 0)
1829                                 goto out;
1830                         ret = process_banner(con);
1831                         if (ret < 0)
1832                                 goto out;
1833                 }
1834                 ret = read_partial_connect(con);
1835                 if (ret <= 0)
1836                         goto out;
1837                 ret = process_connect(con);
1838                 if (ret < 0)
1839                         goto out;
1840                 goto more;
1841         }
1842
1843         if (con->in_base_pos < 0) {
1844                 /*
1845                  * skipping + discarding content.
1846                  *
1847                  * FIXME: there must be a better way to do this!
1848                  */
1849                 static char buf[1024];
1850                 int skip = min(1024, -con->in_base_pos);
1851                 dout("skipping %d / %d bytes\n", skip, -con->in_base_pos);
1852                 ret = ceph_tcp_recvmsg(con->sock, buf, skip);
1853                 if (ret <= 0)
1854                         goto out;
1855                 con->in_base_pos += ret;
1856                 if (con->in_base_pos)
1857                         goto more;
1858         }
1859         if (con->in_tag == CEPH_MSGR_TAG_READY) {
1860                 /*
1861                  * what's next?
1862                  */
1863                 ret = ceph_tcp_recvmsg(con->sock, &con->in_tag, 1);
1864                 if (ret <= 0)
1865                         goto out;
1866                 dout("try_read got tag %d\n", (int)con->in_tag);
1867                 switch (con->in_tag) {
1868                 case CEPH_MSGR_TAG_MSG:
1869                         prepare_read_message(con);
1870                         break;
1871                 case CEPH_MSGR_TAG_ACK:
1872                         prepare_read_ack(con);
1873                         break;
1874                 case CEPH_MSGR_TAG_CLOSE:
1875                         set_bit(CLOSED, &con->state);   /* fixme */
1876                         goto out;
1877                 default:
1878                         goto bad_tag;
1879                 }
1880         }
1881         if (con->in_tag == CEPH_MSGR_TAG_MSG) {
1882                 ret = read_partial_message(con);
1883                 if (ret <= 0) {
1884                         switch (ret) {
1885                         case -EBADMSG:
1886                                 con->error_msg = "bad crc";
1887                                 ret = -EIO;
1888                                 break;
1889                         case -EIO:
1890                                 con->error_msg = "io error";
1891                                 break;
1892                         }
1893                         goto out;
1894                 }
1895                 if (con->in_tag == CEPH_MSGR_TAG_READY)
1896                         goto more;
1897                 process_message(con);
1898                 goto more;
1899         }
1900         if (con->in_tag == CEPH_MSGR_TAG_ACK) {
1901                 ret = read_partial_ack(con);
1902                 if (ret <= 0)
1903                         goto out;
1904                 process_ack(con);
1905                 goto more;
1906         }
1907
1908 out:
1909         dout("try_read done on %p ret %d\n", con, ret);
1910         return ret;
1911
1912 bad_tag:
1913         pr_err("try_read bad con->in_tag = %d\n", (int)con->in_tag);
1914         con->error_msg = "protocol error, garbage tag";
1915         ret = -1;
1916         goto out;
1917 }
1918
1919
1920 /*
1921  * Atomically queue work on a connection.  Bump @con reference to
1922  * avoid races with connection teardown.
1923  */
1924 static void queue_con(struct ceph_connection *con)
1925 {
1926         if (test_bit(DEAD, &con->state)) {
1927                 dout("queue_con %p ignoring: DEAD\n",
1928                      con);
1929                 return;
1930         }
1931
1932         if (!con->ops->get(con)) {
1933                 dout("queue_con %p ref count 0\n", con);
1934                 return;
1935         }
1936
1937         if (!queue_delayed_work(ceph_msgr_wq, &con->work, 0)) {
1938                 dout("queue_con %p - already queued\n", con);
1939                 con->ops->put(con);
1940         } else {
1941                 dout("queue_con %p\n", con);
1942         }
1943 }
1944
1945 /*
1946  * Do some work on a connection.  Drop a connection ref when we're done.
1947  */
1948 static void con_work(struct work_struct *work)
1949 {
1950         struct ceph_connection *con = container_of(work, struct ceph_connection,
1951                                                    work.work);
1952
1953         mutex_lock(&con->mutex);
1954
1955         if (test_bit(CLOSED, &con->state)) { /* e.g. if we are replaced */
1956                 dout("con_work CLOSED\n");
1957                 con_close_socket(con);
1958                 goto done;
1959         }
1960         if (test_and_clear_bit(OPENING, &con->state)) {
1961                 /* reopen w/ new peer */
1962                 dout("con_work OPENING\n");
1963                 con_close_socket(con);
1964         }
1965
1966         if (test_and_clear_bit(SOCK_CLOSED, &con->state) ||
1967             try_read(con) < 0 ||
1968             try_write(con) < 0) {
1969                 mutex_unlock(&con->mutex);
1970                 ceph_fault(con);     /* error/fault path */
1971                 goto done_unlocked;
1972         }
1973
1974 done:
1975         mutex_unlock(&con->mutex);
1976 done_unlocked:
1977         con->ops->put(con);
1978 }
1979
1980
1981 /*
1982  * Generic error/fault handler.  A retry mechanism is used with
1983  * exponential backoff
1984  */
1985 static void ceph_fault(struct ceph_connection *con)
1986 {
1987         pr_err("%s%lld %s %s\n", ENTITY_NAME(con->peer_name),
1988                ceph_pr_addr(&con->peer_addr.in_addr), con->error_msg);
1989         dout("fault %p state %lu to peer %s\n",
1990              con, con->state, ceph_pr_addr(&con->peer_addr.in_addr));
1991
1992         if (test_bit(LOSSYTX, &con->state)) {
1993                 dout("fault on LOSSYTX channel\n");
1994                 goto out;
1995         }
1996
1997         mutex_lock(&con->mutex);
1998         if (test_bit(CLOSED, &con->state))
1999                 goto out_unlock;
2000
2001         con_close_socket(con);
2002
2003         if (con->in_msg) {
2004                 ceph_msg_put(con->in_msg);
2005                 con->in_msg = NULL;
2006         }
2007
2008         /* Requeue anything that hasn't been acked */
2009         list_splice_init(&con->out_sent, &con->out_queue);
2010
2011         /* If there are no messages in the queue, place the connection
2012          * in a STANDBY state (i.e., don't try to reconnect just yet). */
2013         if (list_empty(&con->out_queue) && !con->out_keepalive_pending) {
2014                 dout("fault setting STANDBY\n");
2015                 set_bit(STANDBY, &con->state);
2016         } else {
2017                 /* retry after a delay. */
2018                 if (con->delay == 0)
2019                         con->delay = BASE_DELAY_INTERVAL;
2020                 else if (con->delay < MAX_DELAY_INTERVAL)
2021                         con->delay *= 2;
2022                 dout("fault queueing %p delay %lu\n", con, con->delay);
2023                 con->ops->get(con);
2024                 if (queue_delayed_work(ceph_msgr_wq, &con->work,
2025                                        round_jiffies_relative(con->delay)) == 0)
2026                         con->ops->put(con);
2027         }
2028
2029 out_unlock:
2030         mutex_unlock(&con->mutex);
2031 out:
2032         /*
2033          * in case we faulted due to authentication, invalidate our
2034          * current tickets so that we can get new ones.
2035          */
2036         if (con->auth_retry && con->ops->invalidate_authorizer) {
2037                 dout("calling invalidate_authorizer()\n");
2038                 con->ops->invalidate_authorizer(con);
2039         }
2040
2041         if (con->ops->fault)
2042                 con->ops->fault(con);
2043 }
2044
2045
2046
2047 /*
2048  * create a new messenger instance
2049  */
2050 struct ceph_messenger *ceph_messenger_create(struct ceph_entity_addr *myaddr,
2051                                              u32 supported_features,
2052                                              u32 required_features)
2053 {
2054         struct ceph_messenger *msgr;
2055
2056         msgr = kzalloc(sizeof(*msgr), GFP_KERNEL);
2057         if (msgr == NULL)
2058                 return ERR_PTR(-ENOMEM);
2059
2060         msgr->supported_features = supported_features;
2061         msgr->required_features = required_features;
2062
2063         spin_lock_init(&msgr->global_seq_lock);
2064
2065         /* the zero page is needed if a request is "canceled" while the message
2066          * is being written over the socket */
2067         msgr->zero_page = __page_cache_alloc(GFP_KERNEL | __GFP_ZERO);
2068         if (!msgr->zero_page) {
2069                 kfree(msgr);
2070                 return ERR_PTR(-ENOMEM);
2071         }
2072         kmap(msgr->zero_page);
2073
2074         if (myaddr)
2075                 msgr->inst.addr = *myaddr;
2076
2077         /* select a random nonce */
2078         msgr->inst.addr.type = 0;
2079         get_random_bytes(&msgr->inst.addr.nonce, sizeof(msgr->inst.addr.nonce));
2080         encode_my_addr(msgr);
2081
2082         dout("messenger_create %p\n", msgr);
2083         return msgr;
2084 }
2085 EXPORT_SYMBOL(ceph_messenger_create);
2086
2087 void ceph_messenger_destroy(struct ceph_messenger *msgr)
2088 {
2089         dout("destroy %p\n", msgr);
2090         kunmap(msgr->zero_page);
2091         __free_page(msgr->zero_page);
2092         kfree(msgr);
2093         dout("destroyed messenger %p\n", msgr);
2094 }
2095 EXPORT_SYMBOL(ceph_messenger_destroy);
2096
2097 /*
2098  * Queue up an outgoing message on the given connection.
2099  */
2100 void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg)
2101 {
2102         if (test_bit(CLOSED, &con->state)) {
2103                 dout("con_send %p closed, dropping %p\n", con, msg);
2104                 ceph_msg_put(msg);
2105                 return;
2106         }
2107
2108         /* set src+dst */
2109         msg->hdr.src = con->msgr->inst.name;
2110
2111         BUG_ON(msg->front.iov_len != le32_to_cpu(msg->hdr.front_len));
2112
2113         msg->needs_out_seq = true;
2114
2115         /* queue */
2116         mutex_lock(&con->mutex);
2117         BUG_ON(!list_empty(&msg->list_head));
2118         list_add_tail(&msg->list_head, &con->out_queue);
2119         dout("----- %p to %s%lld %d=%s len %d+%d+%d -----\n", msg,
2120              ENTITY_NAME(con->peer_name), le16_to_cpu(msg->hdr.type),
2121              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
2122              le32_to_cpu(msg->hdr.front_len),
2123              le32_to_cpu(msg->hdr.middle_len),
2124              le32_to_cpu(msg->hdr.data_len));
2125         mutex_unlock(&con->mutex);
2126
2127         /* if there wasn't anything waiting to send before, queue
2128          * new work */
2129         if (test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2130                 queue_con(con);
2131 }
2132 EXPORT_SYMBOL(ceph_con_send);
2133
2134 /*
2135  * Revoke a message that was previously queued for send
2136  */
2137 void ceph_con_revoke(struct ceph_connection *con, struct ceph_msg *msg)
2138 {
2139         mutex_lock(&con->mutex);
2140         if (!list_empty(&msg->list_head)) {
2141                 dout("con_revoke %p msg %p - was on queue\n", con, msg);
2142                 list_del_init(&msg->list_head);
2143                 ceph_msg_put(msg);
2144                 msg->hdr.seq = 0;
2145         }
2146         if (con->out_msg == msg) {
2147                 dout("con_revoke %p msg %p - was sending\n", con, msg);
2148                 con->out_msg = NULL;
2149                 if (con->out_kvec_is_msg) {
2150                         con->out_skip = con->out_kvec_bytes;
2151                         con->out_kvec_is_msg = false;
2152                 }
2153                 ceph_msg_put(msg);
2154                 msg->hdr.seq = 0;
2155         }
2156         mutex_unlock(&con->mutex);
2157 }
2158
2159 /*
2160  * Revoke a message that we may be reading data into
2161  */
2162 void ceph_con_revoke_message(struct ceph_connection *con, struct ceph_msg *msg)
2163 {
2164         mutex_lock(&con->mutex);
2165         if (con->in_msg && con->in_msg == msg) {
2166                 unsigned front_len = le32_to_cpu(con->in_hdr.front_len);
2167                 unsigned middle_len = le32_to_cpu(con->in_hdr.middle_len);
2168                 unsigned data_len = le32_to_cpu(con->in_hdr.data_len);
2169
2170                 /* skip rest of message */
2171                 dout("con_revoke_pages %p msg %p revoked\n", con, msg);
2172                         con->in_base_pos = con->in_base_pos -
2173                                 sizeof(struct ceph_msg_header) -
2174                                 front_len -
2175                                 middle_len -
2176                                 data_len -
2177                                 sizeof(struct ceph_msg_footer);
2178                 ceph_msg_put(con->in_msg);
2179                 con->in_msg = NULL;
2180                 con->in_tag = CEPH_MSGR_TAG_READY;
2181                 con->in_seq++;
2182         } else {
2183                 dout("con_revoke_pages %p msg %p pages %p no-op\n",
2184                      con, con->in_msg, msg);
2185         }
2186         mutex_unlock(&con->mutex);
2187 }
2188
2189 /*
2190  * Queue a keepalive byte to ensure the tcp connection is alive.
2191  */
2192 void ceph_con_keepalive(struct ceph_connection *con)
2193 {
2194         if (test_and_set_bit(KEEPALIVE_PENDING, &con->state) == 0 &&
2195             test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2196                 queue_con(con);
2197 }
2198 EXPORT_SYMBOL(ceph_con_keepalive);
2199
2200
2201 /*
2202  * construct a new message with given type, size
2203  * the new msg has a ref count of 1.
2204  */
2205 struct ceph_msg *ceph_msg_new(int type, int front_len, gfp_t flags)
2206 {
2207         struct ceph_msg *m;
2208
2209         m = kmalloc(sizeof(*m), flags);
2210         if (m == NULL)
2211                 goto out;
2212         kref_init(&m->kref);
2213         INIT_LIST_HEAD(&m->list_head);
2214
2215         m->hdr.tid = 0;
2216         m->hdr.type = cpu_to_le16(type);
2217         m->hdr.priority = cpu_to_le16(CEPH_MSG_PRIO_DEFAULT);
2218         m->hdr.version = 0;
2219         m->hdr.front_len = cpu_to_le32(front_len);
2220         m->hdr.middle_len = 0;
2221         m->hdr.data_len = 0;
2222         m->hdr.data_off = 0;
2223         m->hdr.reserved = 0;
2224         m->footer.front_crc = 0;
2225         m->footer.middle_crc = 0;
2226         m->footer.data_crc = 0;
2227         m->footer.flags = 0;
2228         m->front_max = front_len;
2229         m->front_is_vmalloc = false;
2230         m->more_to_follow = false;
2231         m->pool = NULL;
2232
2233         /* front */
2234         if (front_len) {
2235                 if (front_len > PAGE_CACHE_SIZE) {
2236                         m->front.iov_base = __vmalloc(front_len, flags,
2237                                                       PAGE_KERNEL);
2238                         m->front_is_vmalloc = true;
2239                 } else {
2240                         m->front.iov_base = kmalloc(front_len, flags);
2241                 }
2242                 if (m->front.iov_base == NULL) {
2243                         pr_err("msg_new can't allocate %d bytes\n",
2244                              front_len);
2245                         goto out2;
2246                 }
2247         } else {
2248                 m->front.iov_base = NULL;
2249         }
2250         m->front.iov_len = front_len;
2251
2252         /* middle */
2253         m->middle = NULL;
2254
2255         /* data */
2256         m->nr_pages = 0;
2257         m->page_alignment = 0;
2258         m->pages = NULL;
2259         m->pagelist = NULL;
2260         m->bio = NULL;
2261         m->bio_iter = NULL;
2262         m->bio_seg = 0;
2263         m->trail = NULL;
2264
2265         dout("ceph_msg_new %p front %d\n", m, front_len);
2266         return m;
2267
2268 out2:
2269         ceph_msg_put(m);
2270 out:
2271         pr_err("msg_new can't create type %d front %d\n", type, front_len);
2272         return NULL;
2273 }
2274 EXPORT_SYMBOL(ceph_msg_new);
2275
2276 /*
2277  * Allocate "middle" portion of a message, if it is needed and wasn't
2278  * allocated by alloc_msg.  This allows us to read a small fixed-size
2279  * per-type header in the front and then gracefully fail (i.e.,
2280  * propagate the error to the caller based on info in the front) when
2281  * the middle is too large.
2282  */
2283 static int ceph_alloc_middle(struct ceph_connection *con, struct ceph_msg *msg)
2284 {
2285         int type = le16_to_cpu(msg->hdr.type);
2286         int middle_len = le32_to_cpu(msg->hdr.middle_len);
2287
2288         dout("alloc_middle %p type %d %s middle_len %d\n", msg, type,
2289              ceph_msg_type_name(type), middle_len);
2290         BUG_ON(!middle_len);
2291         BUG_ON(msg->middle);
2292
2293         msg->middle = ceph_buffer_new(middle_len, GFP_NOFS);
2294         if (!msg->middle)
2295                 return -ENOMEM;
2296         return 0;
2297 }
2298
2299 /*
2300  * Generic message allocator, for incoming messages.
2301  */
2302 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
2303                                 struct ceph_msg_header *hdr,
2304                                 int *skip)
2305 {
2306         int type = le16_to_cpu(hdr->type);
2307         int front_len = le32_to_cpu(hdr->front_len);
2308         int middle_len = le32_to_cpu(hdr->middle_len);
2309         struct ceph_msg *msg = NULL;
2310         int ret;
2311
2312         if (con->ops->alloc_msg) {
2313                 mutex_unlock(&con->mutex);
2314                 msg = con->ops->alloc_msg(con, hdr, skip);
2315                 mutex_lock(&con->mutex);
2316                 if (!msg || *skip)
2317                         return NULL;
2318         }
2319         if (!msg) {
2320                 *skip = 0;
2321                 msg = ceph_msg_new(type, front_len, GFP_NOFS);
2322                 if (!msg) {
2323                         pr_err("unable to allocate msg type %d len %d\n",
2324                                type, front_len);
2325                         return NULL;
2326                 }
2327                 msg->page_alignment = le16_to_cpu(hdr->data_off);
2328         }
2329         memcpy(&msg->hdr, &con->in_hdr, sizeof(con->in_hdr));
2330
2331         if (middle_len && !msg->middle) {
2332                 ret = ceph_alloc_middle(con, msg);
2333                 if (ret < 0) {
2334                         ceph_msg_put(msg);
2335                         return NULL;
2336                 }
2337         }
2338
2339         return msg;
2340 }
2341
2342
2343 /*
2344  * Free a generically kmalloc'd message.
2345  */
2346 void ceph_msg_kfree(struct ceph_msg *m)
2347 {
2348         dout("msg_kfree %p\n", m);
2349         if (m->front_is_vmalloc)
2350                 vfree(m->front.iov_base);
2351         else
2352                 kfree(m->front.iov_base);
2353         kfree(m);
2354 }
2355
2356 /*
2357  * Drop a msg ref.  Destroy as needed.
2358  */
2359 void ceph_msg_last_put(struct kref *kref)
2360 {
2361         struct ceph_msg *m = container_of(kref, struct ceph_msg, kref);
2362
2363         dout("ceph_msg_put last one on %p\n", m);
2364         WARN_ON(!list_empty(&m->list_head));
2365
2366         /* drop middle, data, if any */
2367         if (m->middle) {
2368                 ceph_buffer_put(m->middle);
2369                 m->middle = NULL;
2370         }
2371         m->nr_pages = 0;
2372         m->pages = NULL;
2373
2374         if (m->pagelist) {
2375                 ceph_pagelist_release(m->pagelist);
2376                 kfree(m->pagelist);
2377                 m->pagelist = NULL;
2378         }
2379
2380         m->trail = NULL;
2381
2382         if (m->pool)
2383                 ceph_msgpool_put(m->pool, m);
2384         else
2385                 ceph_msg_kfree(m);
2386 }
2387 EXPORT_SYMBOL(ceph_msg_last_put);
2388
2389 void ceph_msg_dump(struct ceph_msg *msg)
2390 {
2391         pr_debug("msg_dump %p (front_max %d nr_pages %d)\n", msg,
2392                  msg->front_max, msg->nr_pages);
2393         print_hex_dump(KERN_DEBUG, "header: ",
2394                        DUMP_PREFIX_OFFSET, 16, 1,
2395                        &msg->hdr, sizeof(msg->hdr), true);
2396         print_hex_dump(KERN_DEBUG, " front: ",
2397                        DUMP_PREFIX_OFFSET, 16, 1,
2398                        msg->front.iov_base, msg->front.iov_len, true);
2399         if (msg->middle)
2400                 print_hex_dump(KERN_DEBUG, "middle: ",
2401                                DUMP_PREFIX_OFFSET, 16, 1,
2402                                msg->middle->vec.iov_base,
2403                                msg->middle->vec.iov_len, true);
2404         print_hex_dump(KERN_DEBUG, "footer: ",
2405                        DUMP_PREFIX_OFFSET, 16, 1,
2406                        &msg->footer, sizeof(msg->footer), true);
2407 }
2408 EXPORT_SYMBOL(ceph_msg_dump);