libceph: embed ceph messenger structure in ceph_client
[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 <linux/dns_resolver.h>
15 #include <net/tcp.h>
16
17 #include <linux/ceph/libceph.h>
18 #include <linux/ceph/messenger.h>
19 #include <linux/ceph/decode.h>
20 #include <linux/ceph/pagelist.h>
21 #include <linux/export.h>
22
23 /*
24  * Ceph uses the messenger to exchange ceph_msg messages with other
25  * hosts in the system.  The messenger provides ordered and reliable
26  * delivery.  We tolerate TCP disconnects by reconnecting (with
27  * exponential backoff) in the case of a fault (disconnection, bad
28  * crc, protocol error).  Acks allow sent messages to be discarded by
29  * the sender.
30  */
31
32 /* static tag bytes (protocol control messages) */
33 static char tag_msg = CEPH_MSGR_TAG_MSG;
34 static char tag_ack = CEPH_MSGR_TAG_ACK;
35 static char tag_keepalive = CEPH_MSGR_TAG_KEEPALIVE;
36
37 #ifdef CONFIG_LOCKDEP
38 static struct lock_class_key socket_class;
39 #endif
40
41 /*
42  * When skipping (ignoring) a block of input we read it into a "skip
43  * buffer," which is this many bytes in size.
44  */
45 #define SKIP_BUF_SIZE   1024
46
47 static void queue_con(struct ceph_connection *con);
48 static void con_work(struct work_struct *);
49 static void ceph_fault(struct ceph_connection *con);
50
51 /*
52  * Nicely render a sockaddr as a string.  An array of formatted
53  * strings is used, to approximate reentrancy.
54  */
55 #define ADDR_STR_COUNT_LOG      5       /* log2(# address strings in array) */
56 #define ADDR_STR_COUNT          (1 << ADDR_STR_COUNT_LOG)
57 #define ADDR_STR_COUNT_MASK     (ADDR_STR_COUNT - 1)
58 #define MAX_ADDR_STR_LEN        64      /* 54 is enough */
59
60 static char addr_str[ADDR_STR_COUNT][MAX_ADDR_STR_LEN];
61 static atomic_t addr_str_seq = ATOMIC_INIT(0);
62
63 static struct page *zero_page;          /* used in certain error cases */
64
65 const char *ceph_pr_addr(const struct sockaddr_storage *ss)
66 {
67         int i;
68         char *s;
69         struct sockaddr_in *in4 = (struct sockaddr_in *) ss;
70         struct sockaddr_in6 *in6 = (struct sockaddr_in6 *) ss;
71
72         i = atomic_inc_return(&addr_str_seq) & ADDR_STR_COUNT_MASK;
73         s = addr_str[i];
74
75         switch (ss->ss_family) {
76         case AF_INET:
77                 snprintf(s, MAX_ADDR_STR_LEN, "%pI4:%hu", &in4->sin_addr,
78                          ntohs(in4->sin_port));
79                 break;
80
81         case AF_INET6:
82                 snprintf(s, MAX_ADDR_STR_LEN, "[%pI6c]:%hu", &in6->sin6_addr,
83                          ntohs(in6->sin6_port));
84                 break;
85
86         default:
87                 snprintf(s, MAX_ADDR_STR_LEN, "(unknown sockaddr family %hu)",
88                          ss->ss_family);
89         }
90
91         return s;
92 }
93 EXPORT_SYMBOL(ceph_pr_addr);
94
95 static void encode_my_addr(struct ceph_messenger *msgr)
96 {
97         memcpy(&msgr->my_enc_addr, &msgr->inst.addr, sizeof(msgr->my_enc_addr));
98         ceph_encode_addr(&msgr->my_enc_addr);
99 }
100
101 /*
102  * work queue for all reading and writing to/from the socket.
103  */
104 static struct workqueue_struct *ceph_msgr_wq;
105
106 void _ceph_msgr_exit(void)
107 {
108         if (ceph_msgr_wq) {
109                 destroy_workqueue(ceph_msgr_wq);
110                 ceph_msgr_wq = NULL;
111         }
112
113         BUG_ON(zero_page == NULL);
114         kunmap(zero_page);
115         page_cache_release(zero_page);
116         zero_page = NULL;
117 }
118
119 int ceph_msgr_init(void)
120 {
121         BUG_ON(zero_page != NULL);
122         zero_page = ZERO_PAGE(0);
123         page_cache_get(zero_page);
124
125         ceph_msgr_wq = alloc_workqueue("ceph-msgr", WQ_NON_REENTRANT, 0);
126         if (ceph_msgr_wq)
127                 return 0;
128
129         pr_err("msgr_init failed to create workqueue\n");
130         _ceph_msgr_exit();
131
132         return -ENOMEM;
133 }
134 EXPORT_SYMBOL(ceph_msgr_init);
135
136 void ceph_msgr_exit(void)
137 {
138         BUG_ON(ceph_msgr_wq == NULL);
139
140         _ceph_msgr_exit();
141 }
142 EXPORT_SYMBOL(ceph_msgr_exit);
143
144 void ceph_msgr_flush(void)
145 {
146         flush_workqueue(ceph_msgr_wq);
147 }
148 EXPORT_SYMBOL(ceph_msgr_flush);
149
150
151 /*
152  * socket callback functions
153  */
154
155 /* data available on socket, or listen socket received a connect */
156 static void ceph_sock_data_ready(struct sock *sk, int count_unused)
157 {
158         struct ceph_connection *con = sk->sk_user_data;
159
160         if (sk->sk_state != TCP_CLOSE_WAIT) {
161                 dout("%s on %p state = %lu, queueing work\n", __func__,
162                      con, con->state);
163                 queue_con(con);
164         }
165 }
166
167 /* socket has buffer space for writing */
168 static void ceph_sock_write_space(struct sock *sk)
169 {
170         struct ceph_connection *con = sk->sk_user_data;
171
172         /* only queue to workqueue if there is data we want to write,
173          * and there is sufficient space in the socket buffer to accept
174          * more data.  clear SOCK_NOSPACE so that ceph_sock_write_space()
175          * doesn't get called again until try_write() fills the socket
176          * buffer. See net/ipv4/tcp_input.c:tcp_check_space()
177          * and net/core/stream.c:sk_stream_write_space().
178          */
179         if (test_bit(WRITE_PENDING, &con->state)) {
180                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
181                         dout("%s %p queueing write work\n", __func__, con);
182                         clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
183                         queue_con(con);
184                 }
185         } else {
186                 dout("%s %p nothing to write\n", __func__, con);
187         }
188 }
189
190 /* socket's state has changed */
191 static void ceph_sock_state_change(struct sock *sk)
192 {
193         struct ceph_connection *con = sk->sk_user_data;
194
195         dout("%s %p state = %lu sk_state = %u\n", __func__,
196              con, con->state, sk->sk_state);
197
198         if (test_bit(CLOSED, &con->state))
199                 return;
200
201         switch (sk->sk_state) {
202         case TCP_CLOSE:
203                 dout("%s TCP_CLOSE\n", __func__);
204         case TCP_CLOSE_WAIT:
205                 dout("%s TCP_CLOSE_WAIT\n", __func__);
206                 if (test_and_set_bit(SOCK_CLOSED, &con->state) == 0) {
207                         if (test_bit(CONNECTING, &con->state))
208                                 con->error_msg = "connection failed";
209                         else
210                                 con->error_msg = "socket closed";
211                         queue_con(con);
212                 }
213                 break;
214         case TCP_ESTABLISHED:
215                 dout("%s TCP_ESTABLISHED\n", __func__);
216                 queue_con(con);
217                 break;
218         default:        /* Everything else is uninteresting */
219                 break;
220         }
221 }
222
223 /*
224  * set up socket callbacks
225  */
226 static void set_sock_callbacks(struct socket *sock,
227                                struct ceph_connection *con)
228 {
229         struct sock *sk = sock->sk;
230         sk->sk_user_data = con;
231         sk->sk_data_ready = ceph_sock_data_ready;
232         sk->sk_write_space = ceph_sock_write_space;
233         sk->sk_state_change = ceph_sock_state_change;
234 }
235
236
237 /*
238  * socket helpers
239  */
240
241 /*
242  * initiate connection to a remote socket.
243  */
244 static int ceph_tcp_connect(struct ceph_connection *con)
245 {
246         struct sockaddr_storage *paddr = &con->peer_addr.in_addr;
247         struct socket *sock;
248         int ret;
249
250         BUG_ON(con->sock);
251         ret = sock_create_kern(con->peer_addr.in_addr.ss_family, SOCK_STREAM,
252                                IPPROTO_TCP, &sock);
253         if (ret)
254                 return ret;
255         sock->sk->sk_allocation = GFP_NOFS;
256
257 #ifdef CONFIG_LOCKDEP
258         lockdep_set_class(&sock->sk->sk_lock, &socket_class);
259 #endif
260
261         set_sock_callbacks(sock, con);
262
263         dout("connect %s\n", ceph_pr_addr(&con->peer_addr.in_addr));
264
265         ret = sock->ops->connect(sock, (struct sockaddr *)paddr, sizeof(*paddr),
266                                  O_NONBLOCK);
267         if (ret == -EINPROGRESS) {
268                 dout("connect %s EINPROGRESS sk_state = %u\n",
269                      ceph_pr_addr(&con->peer_addr.in_addr),
270                      sock->sk->sk_state);
271         } else if (ret < 0) {
272                 pr_err("connect %s error %d\n",
273                        ceph_pr_addr(&con->peer_addr.in_addr), ret);
274                 sock_release(sock);
275                 con->error_msg = "connect error";
276
277                 return ret;
278         }
279         con->sock = sock;
280
281         return 0;
282 }
283
284 static int ceph_tcp_recvmsg(struct socket *sock, void *buf, size_t len)
285 {
286         struct kvec iov = {buf, len};
287         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
288         int r;
289
290         r = kernel_recvmsg(sock, &msg, &iov, 1, len, msg.msg_flags);
291         if (r == -EAGAIN)
292                 r = 0;
293         return r;
294 }
295
296 /*
297  * write something.  @more is true if caller will be sending more data
298  * shortly.
299  */
300 static int ceph_tcp_sendmsg(struct socket *sock, struct kvec *iov,
301                      size_t kvlen, size_t len, int more)
302 {
303         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
304         int r;
305
306         if (more)
307                 msg.msg_flags |= MSG_MORE;
308         else
309                 msg.msg_flags |= MSG_EOR;  /* superfluous, but what the hell */
310
311         r = kernel_sendmsg(sock, &msg, iov, kvlen, len);
312         if (r == -EAGAIN)
313                 r = 0;
314         return r;
315 }
316
317 static int ceph_tcp_sendpage(struct socket *sock, struct page *page,
318                      int offset, size_t size, int more)
319 {
320         int flags = MSG_DONTWAIT | MSG_NOSIGNAL | (more ? MSG_MORE : MSG_EOR);
321         int ret;
322
323         ret = kernel_sendpage(sock, page, offset, size, flags);
324         if (ret == -EAGAIN)
325                 ret = 0;
326
327         return ret;
328 }
329
330
331 /*
332  * Shutdown/close the socket for the given connection.
333  */
334 static int con_close_socket(struct ceph_connection *con)
335 {
336         int rc;
337
338         dout("con_close_socket on %p sock %p\n", con, con->sock);
339         if (!con->sock)
340                 return 0;
341         set_bit(SOCK_CLOSED, &con->state);
342         rc = con->sock->ops->shutdown(con->sock, SHUT_RDWR);
343         sock_release(con->sock);
344         con->sock = NULL;
345         clear_bit(SOCK_CLOSED, &con->state);
346         return rc;
347 }
348
349 /*
350  * Reset a connection.  Discard all incoming and outgoing messages
351  * and clear *_seq state.
352  */
353 static void ceph_msg_remove(struct ceph_msg *msg)
354 {
355         list_del_init(&msg->list_head);
356         ceph_msg_put(msg);
357 }
358 static void ceph_msg_remove_list(struct list_head *head)
359 {
360         while (!list_empty(head)) {
361                 struct ceph_msg *msg = list_first_entry(head, struct ceph_msg,
362                                                         list_head);
363                 ceph_msg_remove(msg);
364         }
365 }
366
367 static void reset_connection(struct ceph_connection *con)
368 {
369         /* reset connection, out_queue, msg_ and connect_seq */
370         /* discard existing out_queue and msg_seq */
371         ceph_msg_remove_list(&con->out_queue);
372         ceph_msg_remove_list(&con->out_sent);
373
374         if (con->in_msg) {
375                 ceph_msg_put(con->in_msg);
376                 con->in_msg = NULL;
377         }
378
379         con->connect_seq = 0;
380         con->out_seq = 0;
381         if (con->out_msg) {
382                 ceph_msg_put(con->out_msg);
383                 con->out_msg = NULL;
384         }
385         con->in_seq = 0;
386         con->in_seq_acked = 0;
387 }
388
389 /*
390  * mark a peer down.  drop any open connections.
391  */
392 void ceph_con_close(struct ceph_connection *con)
393 {
394         dout("con_close %p peer %s\n", con,
395              ceph_pr_addr(&con->peer_addr.in_addr));
396         set_bit(CLOSED, &con->state);  /* in case there's queued work */
397         clear_bit(STANDBY, &con->state);  /* avoid connect_seq bump */
398         clear_bit(LOSSYTX, &con->state);  /* so we retry next connect */
399         clear_bit(KEEPALIVE_PENDING, &con->state);
400         clear_bit(WRITE_PENDING, &con->state);
401         mutex_lock(&con->mutex);
402         reset_connection(con);
403         con->peer_global_seq = 0;
404         cancel_delayed_work(&con->work);
405         mutex_unlock(&con->mutex);
406         queue_con(con);
407 }
408 EXPORT_SYMBOL(ceph_con_close);
409
410 /*
411  * Reopen a closed connection, with a new peer address.
412  */
413 void ceph_con_open(struct ceph_connection *con, struct ceph_entity_addr *addr)
414 {
415         dout("con_open %p %s\n", con, ceph_pr_addr(&addr->in_addr));
416         set_bit(OPENING, &con->state);
417         clear_bit(CLOSED, &con->state);
418         memcpy(&con->peer_addr, addr, sizeof(*addr));
419         con->delay = 0;      /* reset backoff memory */
420         queue_con(con);
421 }
422 EXPORT_SYMBOL(ceph_con_open);
423
424 /*
425  * return true if this connection ever successfully opened
426  */
427 bool ceph_con_opened(struct ceph_connection *con)
428 {
429         return con->connect_seq > 0;
430 }
431
432 /*
433  * generic get/put
434  */
435 struct ceph_connection *ceph_con_get(struct ceph_connection *con)
436 {
437         int nref = __atomic_add_unless(&con->nref, 1, 0);
438
439         dout("con_get %p nref = %d -> %d\n", con, nref, nref + 1);
440
441         return nref ? con : NULL;
442 }
443
444 void ceph_con_put(struct ceph_connection *con)
445 {
446         int nref = atomic_dec_return(&con->nref);
447
448         BUG_ON(nref < 0);
449         if (nref == 0) {
450                 BUG_ON(con->sock);
451                 kfree(con);
452         }
453         dout("con_put %p nref = %d -> %d\n", con, nref + 1, nref);
454 }
455
456 /*
457  * initialize a new connection.
458  */
459 void ceph_con_init(struct ceph_messenger *msgr, struct ceph_connection *con)
460 {
461         dout("con_init %p\n", con);
462         memset(con, 0, sizeof(*con));
463         atomic_set(&con->nref, 1);
464         con->msgr = msgr;
465         mutex_init(&con->mutex);
466         INIT_LIST_HEAD(&con->out_queue);
467         INIT_LIST_HEAD(&con->out_sent);
468         INIT_DELAYED_WORK(&con->work, con_work);
469 }
470 EXPORT_SYMBOL(ceph_con_init);
471
472
473 /*
474  * We maintain a global counter to order connection attempts.  Get
475  * a unique seq greater than @gt.
476  */
477 static u32 get_global_seq(struct ceph_messenger *msgr, u32 gt)
478 {
479         u32 ret;
480
481         spin_lock(&msgr->global_seq_lock);
482         if (msgr->global_seq < gt)
483                 msgr->global_seq = gt;
484         ret = ++msgr->global_seq;
485         spin_unlock(&msgr->global_seq_lock);
486         return ret;
487 }
488
489 static void con_out_kvec_reset(struct ceph_connection *con)
490 {
491         con->out_kvec_left = 0;
492         con->out_kvec_bytes = 0;
493         con->out_kvec_cur = &con->out_kvec[0];
494 }
495
496 static void con_out_kvec_add(struct ceph_connection *con,
497                                 size_t size, void *data)
498 {
499         int index;
500
501         index = con->out_kvec_left;
502         BUG_ON(index >= ARRAY_SIZE(con->out_kvec));
503
504         con->out_kvec[index].iov_len = size;
505         con->out_kvec[index].iov_base = data;
506         con->out_kvec_left++;
507         con->out_kvec_bytes += size;
508 }
509
510 /*
511  * Prepare footer for currently outgoing message, and finish things
512  * off.  Assumes out_kvec* are already valid.. we just add on to the end.
513  */
514 static void prepare_write_message_footer(struct ceph_connection *con)
515 {
516         struct ceph_msg *m = con->out_msg;
517         int v = con->out_kvec_left;
518
519         dout("prepare_write_message_footer %p\n", con);
520         con->out_kvec_is_msg = true;
521         con->out_kvec[v].iov_base = &m->footer;
522         con->out_kvec[v].iov_len = sizeof(m->footer);
523         con->out_kvec_bytes += sizeof(m->footer);
524         con->out_kvec_left++;
525         con->out_more = m->more_to_follow;
526         con->out_msg_done = true;
527 }
528
529 /*
530  * Prepare headers for the next outgoing message.
531  */
532 static void prepare_write_message(struct ceph_connection *con)
533 {
534         struct ceph_msg *m;
535         u32 crc;
536
537         con_out_kvec_reset(con);
538         con->out_kvec_is_msg = true;
539         con->out_msg_done = false;
540
541         /* Sneak an ack in there first?  If we can get it into the same
542          * TCP packet that's a good thing. */
543         if (con->in_seq > con->in_seq_acked) {
544                 con->in_seq_acked = con->in_seq;
545                 con_out_kvec_add(con, sizeof (tag_ack), &tag_ack);
546                 con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
547                 con_out_kvec_add(con, sizeof (con->out_temp_ack),
548                         &con->out_temp_ack);
549         }
550
551         m = list_first_entry(&con->out_queue, struct ceph_msg, list_head);
552         con->out_msg = m;
553
554         /* put message on sent list */
555         ceph_msg_get(m);
556         list_move_tail(&m->list_head, &con->out_sent);
557
558         /*
559          * only assign outgoing seq # if we haven't sent this message
560          * yet.  if it is requeued, resend with it's original seq.
561          */
562         if (m->needs_out_seq) {
563                 m->hdr.seq = cpu_to_le64(++con->out_seq);
564                 m->needs_out_seq = false;
565         }
566 #ifdef CONFIG_BLOCK
567         else
568                 m->bio_iter = NULL;
569 #endif
570
571         dout("prepare_write_message %p seq %lld type %d len %d+%d+%d %d pgs\n",
572              m, con->out_seq, le16_to_cpu(m->hdr.type),
573              le32_to_cpu(m->hdr.front_len), le32_to_cpu(m->hdr.middle_len),
574              le32_to_cpu(m->hdr.data_len),
575              m->nr_pages);
576         BUG_ON(le32_to_cpu(m->hdr.front_len) != m->front.iov_len);
577
578         /* tag + hdr + front + middle */
579         con_out_kvec_add(con, sizeof (tag_msg), &tag_msg);
580         con_out_kvec_add(con, sizeof (m->hdr), &m->hdr);
581         con_out_kvec_add(con, m->front.iov_len, m->front.iov_base);
582
583         if (m->middle)
584                 con_out_kvec_add(con, m->middle->vec.iov_len,
585                         m->middle->vec.iov_base);
586
587         /* fill in crc (except data pages), footer */
588         crc = crc32c(0, &m->hdr, offsetof(struct ceph_msg_header, crc));
589         con->out_msg->hdr.crc = cpu_to_le32(crc);
590         con->out_msg->footer.flags = CEPH_MSG_FOOTER_COMPLETE;
591
592         crc = crc32c(0, m->front.iov_base, m->front.iov_len);
593         con->out_msg->footer.front_crc = cpu_to_le32(crc);
594         if (m->middle) {
595                 crc = crc32c(0, m->middle->vec.iov_base,
596                                 m->middle->vec.iov_len);
597                 con->out_msg->footer.middle_crc = cpu_to_le32(crc);
598         } else
599                 con->out_msg->footer.middle_crc = 0;
600         con->out_msg->footer.data_crc = 0;
601         dout("prepare_write_message front_crc %u data_crc %u\n",
602              le32_to_cpu(con->out_msg->footer.front_crc),
603              le32_to_cpu(con->out_msg->footer.middle_crc));
604
605         /* is there a data payload? */
606         if (le32_to_cpu(m->hdr.data_len) > 0) {
607                 /* initialize page iterator */
608                 con->out_msg_pos.page = 0;
609                 if (m->pages)
610                         con->out_msg_pos.page_pos = m->page_alignment;
611                 else
612                         con->out_msg_pos.page_pos = 0;
613                 con->out_msg_pos.data_pos = 0;
614                 con->out_msg_pos.did_page_crc = false;
615                 con->out_more = 1;  /* data + footer will follow */
616         } else {
617                 /* no, queue up footer too and be done */
618                 prepare_write_message_footer(con);
619         }
620
621         set_bit(WRITE_PENDING, &con->state);
622 }
623
624 /*
625  * Prepare an ack.
626  */
627 static void prepare_write_ack(struct ceph_connection *con)
628 {
629         dout("prepare_write_ack %p %llu -> %llu\n", con,
630              con->in_seq_acked, con->in_seq);
631         con->in_seq_acked = con->in_seq;
632
633         con_out_kvec_reset(con);
634
635         con_out_kvec_add(con, sizeof (tag_ack), &tag_ack);
636
637         con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
638         con_out_kvec_add(con, sizeof (con->out_temp_ack),
639                                 &con->out_temp_ack);
640
641         con->out_more = 1;  /* more will follow.. eventually.. */
642         set_bit(WRITE_PENDING, &con->state);
643 }
644
645 /*
646  * Prepare to write keepalive byte.
647  */
648 static void prepare_write_keepalive(struct ceph_connection *con)
649 {
650         dout("prepare_write_keepalive %p\n", con);
651         con_out_kvec_reset(con);
652         con_out_kvec_add(con, sizeof (tag_keepalive), &tag_keepalive);
653         set_bit(WRITE_PENDING, &con->state);
654 }
655
656 /*
657  * Connection negotiation.
658  */
659
660 static struct ceph_auth_handshake *get_connect_authorizer(struct ceph_connection *con,
661                                                 int *auth_proto)
662 {
663         struct ceph_auth_handshake *auth;
664
665         if (!con->ops->get_authorizer) {
666                 con->out_connect.authorizer_protocol = CEPH_AUTH_UNKNOWN;
667                 con->out_connect.authorizer_len = 0;
668
669                 return NULL;
670         }
671
672         /* Can't hold the mutex while getting authorizer */
673
674         mutex_unlock(&con->mutex);
675
676         auth = con->ops->get_authorizer(con, auth_proto, con->auth_retry);
677
678         mutex_lock(&con->mutex);
679
680         if (IS_ERR(auth))
681                 return auth;
682         if (test_bit(CLOSED, &con->state) || test_bit(OPENING, &con->state))
683                 return ERR_PTR(-EAGAIN);
684
685         con->auth_reply_buf = auth->authorizer_reply_buf;
686         con->auth_reply_buf_len = auth->authorizer_reply_buf_len;
687
688
689         return auth;
690 }
691
692 /*
693  * We connected to a peer and are saying hello.
694  */
695 static void prepare_write_banner(struct ceph_connection *con)
696 {
697         con_out_kvec_add(con, strlen(CEPH_BANNER), CEPH_BANNER);
698         con_out_kvec_add(con, sizeof (con->msgr->my_enc_addr),
699                                         &con->msgr->my_enc_addr);
700
701         con->out_more = 0;
702         set_bit(WRITE_PENDING, &con->state);
703 }
704
705 static int prepare_write_connect(struct ceph_connection *con)
706 {
707         unsigned global_seq = get_global_seq(con->msgr, 0);
708         int proto;
709         int auth_proto;
710         struct ceph_auth_handshake *auth;
711
712         switch (con->peer_name.type) {
713         case CEPH_ENTITY_TYPE_MON:
714                 proto = CEPH_MONC_PROTOCOL;
715                 break;
716         case CEPH_ENTITY_TYPE_OSD:
717                 proto = CEPH_OSDC_PROTOCOL;
718                 break;
719         case CEPH_ENTITY_TYPE_MDS:
720                 proto = CEPH_MDSC_PROTOCOL;
721                 break;
722         default:
723                 BUG();
724         }
725
726         dout("prepare_write_connect %p cseq=%d gseq=%d proto=%d\n", con,
727              con->connect_seq, global_seq, proto);
728
729         con->out_connect.features = cpu_to_le64(con->msgr->supported_features);
730         con->out_connect.host_type = cpu_to_le32(CEPH_ENTITY_TYPE_CLIENT);
731         con->out_connect.connect_seq = cpu_to_le32(con->connect_seq);
732         con->out_connect.global_seq = cpu_to_le32(global_seq);
733         con->out_connect.protocol_version = cpu_to_le32(proto);
734         con->out_connect.flags = 0;
735
736         auth_proto = CEPH_AUTH_UNKNOWN;
737         auth = get_connect_authorizer(con, &auth_proto);
738         if (IS_ERR(auth))
739                 return PTR_ERR(auth);
740
741         con->out_connect.authorizer_protocol = cpu_to_le32(auth_proto);
742         con->out_connect.authorizer_len = auth ?
743                 cpu_to_le32(auth->authorizer_buf_len) : 0;
744
745         con_out_kvec_add(con, sizeof (con->out_connect),
746                                         &con->out_connect);
747         if (auth && auth->authorizer_buf_len)
748                 con_out_kvec_add(con, auth->authorizer_buf_len,
749                                         auth->authorizer_buf);
750
751         con->out_more = 0;
752         set_bit(WRITE_PENDING, &con->state);
753
754         return 0;
755 }
756
757 /*
758  * write as much of pending kvecs to the socket as we can.
759  *  1 -> done
760  *  0 -> socket full, but more to do
761  * <0 -> error
762  */
763 static int write_partial_kvec(struct ceph_connection *con)
764 {
765         int ret;
766
767         dout("write_partial_kvec %p %d left\n", con, con->out_kvec_bytes);
768         while (con->out_kvec_bytes > 0) {
769                 ret = ceph_tcp_sendmsg(con->sock, con->out_kvec_cur,
770                                        con->out_kvec_left, con->out_kvec_bytes,
771                                        con->out_more);
772                 if (ret <= 0)
773                         goto out;
774                 con->out_kvec_bytes -= ret;
775                 if (con->out_kvec_bytes == 0)
776                         break;            /* done */
777
778                 /* account for full iov entries consumed */
779                 while (ret >= con->out_kvec_cur->iov_len) {
780                         BUG_ON(!con->out_kvec_left);
781                         ret -= con->out_kvec_cur->iov_len;
782                         con->out_kvec_cur++;
783                         con->out_kvec_left--;
784                 }
785                 /* and for a partially-consumed entry */
786                 if (ret) {
787                         con->out_kvec_cur->iov_len -= ret;
788                         con->out_kvec_cur->iov_base += ret;
789                 }
790         }
791         con->out_kvec_left = 0;
792         con->out_kvec_is_msg = false;
793         ret = 1;
794 out:
795         dout("write_partial_kvec %p %d left in %d kvecs ret = %d\n", con,
796              con->out_kvec_bytes, con->out_kvec_left, ret);
797         return ret;  /* done! */
798 }
799
800 #ifdef CONFIG_BLOCK
801 static void init_bio_iter(struct bio *bio, struct bio **iter, int *seg)
802 {
803         if (!bio) {
804                 *iter = NULL;
805                 *seg = 0;
806                 return;
807         }
808         *iter = bio;
809         *seg = bio->bi_idx;
810 }
811
812 static void iter_bio_next(struct bio **bio_iter, int *seg)
813 {
814         if (*bio_iter == NULL)
815                 return;
816
817         BUG_ON(*seg >= (*bio_iter)->bi_vcnt);
818
819         (*seg)++;
820         if (*seg == (*bio_iter)->bi_vcnt)
821                 init_bio_iter((*bio_iter)->bi_next, bio_iter, seg);
822 }
823 #endif
824
825 /*
826  * Write as much message data payload as we can.  If we finish, queue
827  * up the footer.
828  *  1 -> done, footer is now queued in out_kvec[].
829  *  0 -> socket full, but more to do
830  * <0 -> error
831  */
832 static int write_partial_msg_pages(struct ceph_connection *con)
833 {
834         struct ceph_msg *msg = con->out_msg;
835         unsigned data_len = le32_to_cpu(msg->hdr.data_len);
836         size_t len;
837         bool do_datacrc = !con->msgr->nocrc;
838         int ret;
839         int total_max_write;
840         int in_trail = 0;
841         size_t trail_len = (msg->trail ? msg->trail->length : 0);
842
843         dout("write_partial_msg_pages %p msg %p page %d/%d offset %d\n",
844              con, con->out_msg, con->out_msg_pos.page, con->out_msg->nr_pages,
845              con->out_msg_pos.page_pos);
846
847 #ifdef CONFIG_BLOCK
848         if (msg->bio && !msg->bio_iter)
849                 init_bio_iter(msg->bio, &msg->bio_iter, &msg->bio_seg);
850 #endif
851
852         while (data_len > con->out_msg_pos.data_pos) {
853                 struct page *page = NULL;
854                 int max_write = PAGE_SIZE;
855                 int bio_offset = 0;
856
857                 total_max_write = data_len - trail_len -
858                         con->out_msg_pos.data_pos;
859
860                 /*
861                  * if we are calculating the data crc (the default), we need
862                  * to map the page.  if our pages[] has been revoked, use the
863                  * zero page.
864                  */
865
866                 /* have we reached the trail part of the data? */
867                 if (con->out_msg_pos.data_pos >= data_len - trail_len) {
868                         in_trail = 1;
869
870                         total_max_write = data_len - con->out_msg_pos.data_pos;
871
872                         page = list_first_entry(&msg->trail->head,
873                                                 struct page, lru);
874                         max_write = PAGE_SIZE;
875                 } else if (msg->pages) {
876                         page = msg->pages[con->out_msg_pos.page];
877                 } else if (msg->pagelist) {
878                         page = list_first_entry(&msg->pagelist->head,
879                                                 struct page, lru);
880 #ifdef CONFIG_BLOCK
881                 } else if (msg->bio) {
882                         struct bio_vec *bv;
883
884                         bv = bio_iovec_idx(msg->bio_iter, msg->bio_seg);
885                         page = bv->bv_page;
886                         bio_offset = bv->bv_offset;
887                         max_write = bv->bv_len;
888 #endif
889                 } else {
890                         page = zero_page;
891                 }
892                 len = min_t(int, max_write - con->out_msg_pos.page_pos,
893                             total_max_write);
894
895                 if (do_datacrc && !con->out_msg_pos.did_page_crc) {
896                         void *base;
897                         u32 crc;
898                         u32 tmpcrc = le32_to_cpu(con->out_msg->footer.data_crc);
899                         char *kaddr;
900
901                         kaddr = kmap(page);
902                         BUG_ON(kaddr == NULL);
903                         base = kaddr + con->out_msg_pos.page_pos + bio_offset;
904                         crc = crc32c(tmpcrc, base, len);
905                         con->out_msg->footer.data_crc = cpu_to_le32(crc);
906                         con->out_msg_pos.did_page_crc = true;
907                 }
908                 ret = ceph_tcp_sendpage(con->sock, page,
909                                       con->out_msg_pos.page_pos + bio_offset,
910                                       len, 1);
911
912                 if (do_datacrc)
913                         kunmap(page);
914
915                 if (ret <= 0)
916                         goto out;
917
918                 con->out_msg_pos.data_pos += ret;
919                 con->out_msg_pos.page_pos += ret;
920                 if (ret == len) {
921                         con->out_msg_pos.page_pos = 0;
922                         con->out_msg_pos.page++;
923                         con->out_msg_pos.did_page_crc = false;
924                         if (in_trail)
925                                 list_move_tail(&page->lru,
926                                                &msg->trail->head);
927                         else if (msg->pagelist)
928                                 list_move_tail(&page->lru,
929                                                &msg->pagelist->head);
930 #ifdef CONFIG_BLOCK
931                         else if (msg->bio)
932                                 iter_bio_next(&msg->bio_iter, &msg->bio_seg);
933 #endif
934                 }
935         }
936
937         dout("write_partial_msg_pages %p msg %p done\n", con, msg);
938
939         /* prepare and queue up footer, too */
940         if (!do_datacrc)
941                 con->out_msg->footer.flags |= CEPH_MSG_FOOTER_NOCRC;
942         con_out_kvec_reset(con);
943         prepare_write_message_footer(con);
944         ret = 1;
945 out:
946         return ret;
947 }
948
949 /*
950  * write some zeros
951  */
952 static int write_partial_skip(struct ceph_connection *con)
953 {
954         int ret;
955
956         while (con->out_skip > 0) {
957                 size_t size = min(con->out_skip, (int) PAGE_CACHE_SIZE);
958
959                 ret = ceph_tcp_sendpage(con->sock, zero_page, 0, size, 1);
960                 if (ret <= 0)
961                         goto out;
962                 con->out_skip -= ret;
963         }
964         ret = 1;
965 out:
966         return ret;
967 }
968
969 /*
970  * Prepare to read connection handshake, or an ack.
971  */
972 static void prepare_read_banner(struct ceph_connection *con)
973 {
974         dout("prepare_read_banner %p\n", con);
975         con->in_base_pos = 0;
976 }
977
978 static void prepare_read_connect(struct ceph_connection *con)
979 {
980         dout("prepare_read_connect %p\n", con);
981         con->in_base_pos = 0;
982 }
983
984 static void prepare_read_ack(struct ceph_connection *con)
985 {
986         dout("prepare_read_ack %p\n", con);
987         con->in_base_pos = 0;
988 }
989
990 static void prepare_read_tag(struct ceph_connection *con)
991 {
992         dout("prepare_read_tag %p\n", con);
993         con->in_base_pos = 0;
994         con->in_tag = CEPH_MSGR_TAG_READY;
995 }
996
997 /*
998  * Prepare to read a message.
999  */
1000 static int prepare_read_message(struct ceph_connection *con)
1001 {
1002         dout("prepare_read_message %p\n", con);
1003         BUG_ON(con->in_msg != NULL);
1004         con->in_base_pos = 0;
1005         con->in_front_crc = con->in_middle_crc = con->in_data_crc = 0;
1006         return 0;
1007 }
1008
1009
1010 static int read_partial(struct ceph_connection *con,
1011                         int end, int size, void *object)
1012 {
1013         while (con->in_base_pos < end) {
1014                 int left = end - con->in_base_pos;
1015                 int have = size - left;
1016                 int ret = ceph_tcp_recvmsg(con->sock, object + have, left);
1017                 if (ret <= 0)
1018                         return ret;
1019                 con->in_base_pos += ret;
1020         }
1021         return 1;
1022 }
1023
1024
1025 /*
1026  * Read all or part of the connect-side handshake on a new connection
1027  */
1028 static int read_partial_banner(struct ceph_connection *con)
1029 {
1030         int size;
1031         int end;
1032         int ret;
1033
1034         dout("read_partial_banner %p at %d\n", con, con->in_base_pos);
1035
1036         /* peer's banner */
1037         size = strlen(CEPH_BANNER);
1038         end = size;
1039         ret = read_partial(con, end, size, con->in_banner);
1040         if (ret <= 0)
1041                 goto out;
1042
1043         size = sizeof (con->actual_peer_addr);
1044         end += size;
1045         ret = read_partial(con, end, size, &con->actual_peer_addr);
1046         if (ret <= 0)
1047                 goto out;
1048
1049         size = sizeof (con->peer_addr_for_me);
1050         end += size;
1051         ret = read_partial(con, end, size, &con->peer_addr_for_me);
1052         if (ret <= 0)
1053                 goto out;
1054
1055 out:
1056         return ret;
1057 }
1058
1059 static int read_partial_connect(struct ceph_connection *con)
1060 {
1061         int size;
1062         int end;
1063         int ret;
1064
1065         dout("read_partial_connect %p at %d\n", con, con->in_base_pos);
1066
1067         size = sizeof (con->in_reply);
1068         end = size;
1069         ret = read_partial(con, end, size, &con->in_reply);
1070         if (ret <= 0)
1071                 goto out;
1072
1073         size = le32_to_cpu(con->in_reply.authorizer_len);
1074         end += size;
1075         ret = read_partial(con, end, size, con->auth_reply_buf);
1076         if (ret <= 0)
1077                 goto out;
1078
1079         dout("read_partial_connect %p tag %d, con_seq = %u, g_seq = %u\n",
1080              con, (int)con->in_reply.tag,
1081              le32_to_cpu(con->in_reply.connect_seq),
1082              le32_to_cpu(con->in_reply.global_seq));
1083 out:
1084         return ret;
1085
1086 }
1087
1088 /*
1089  * Verify the hello banner looks okay.
1090  */
1091 static int verify_hello(struct ceph_connection *con)
1092 {
1093         if (memcmp(con->in_banner, CEPH_BANNER, strlen(CEPH_BANNER))) {
1094                 pr_err("connect to %s got bad banner\n",
1095                        ceph_pr_addr(&con->peer_addr.in_addr));
1096                 con->error_msg = "protocol error, bad banner";
1097                 return -1;
1098         }
1099         return 0;
1100 }
1101
1102 static bool addr_is_blank(struct sockaddr_storage *ss)
1103 {
1104         switch (ss->ss_family) {
1105         case AF_INET:
1106                 return ((struct sockaddr_in *)ss)->sin_addr.s_addr == 0;
1107         case AF_INET6:
1108                 return
1109                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[0] == 0 &&
1110                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[1] == 0 &&
1111                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[2] == 0 &&
1112                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[3] == 0;
1113         }
1114         return false;
1115 }
1116
1117 static int addr_port(struct sockaddr_storage *ss)
1118 {
1119         switch (ss->ss_family) {
1120         case AF_INET:
1121                 return ntohs(((struct sockaddr_in *)ss)->sin_port);
1122         case AF_INET6:
1123                 return ntohs(((struct sockaddr_in6 *)ss)->sin6_port);
1124         }
1125         return 0;
1126 }
1127
1128 static void addr_set_port(struct sockaddr_storage *ss, int p)
1129 {
1130         switch (ss->ss_family) {
1131         case AF_INET:
1132                 ((struct sockaddr_in *)ss)->sin_port = htons(p);
1133                 break;
1134         case AF_INET6:
1135                 ((struct sockaddr_in6 *)ss)->sin6_port = htons(p);
1136                 break;
1137         }
1138 }
1139
1140 /*
1141  * Unlike other *_pton function semantics, zero indicates success.
1142  */
1143 static int ceph_pton(const char *str, size_t len, struct sockaddr_storage *ss,
1144                 char delim, const char **ipend)
1145 {
1146         struct sockaddr_in *in4 = (struct sockaddr_in *) ss;
1147         struct sockaddr_in6 *in6 = (struct sockaddr_in6 *) ss;
1148
1149         memset(ss, 0, sizeof(*ss));
1150
1151         if (in4_pton(str, len, (u8 *)&in4->sin_addr.s_addr, delim, ipend)) {
1152                 ss->ss_family = AF_INET;
1153                 return 0;
1154         }
1155
1156         if (in6_pton(str, len, (u8 *)&in6->sin6_addr.s6_addr, delim, ipend)) {
1157                 ss->ss_family = AF_INET6;
1158                 return 0;
1159         }
1160
1161         return -EINVAL;
1162 }
1163
1164 /*
1165  * Extract hostname string and resolve using kernel DNS facility.
1166  */
1167 #ifdef CONFIG_CEPH_LIB_USE_DNS_RESOLVER
1168 static int ceph_dns_resolve_name(const char *name, size_t namelen,
1169                 struct sockaddr_storage *ss, char delim, const char **ipend)
1170 {
1171         const char *end, *delim_p;
1172         char *colon_p, *ip_addr = NULL;
1173         int ip_len, ret;
1174
1175         /*
1176          * The end of the hostname occurs immediately preceding the delimiter or
1177          * the port marker (':') where the delimiter takes precedence.
1178          */
1179         delim_p = memchr(name, delim, namelen);
1180         colon_p = memchr(name, ':', namelen);
1181
1182         if (delim_p && colon_p)
1183                 end = delim_p < colon_p ? delim_p : colon_p;
1184         else if (!delim_p && colon_p)
1185                 end = colon_p;
1186         else {
1187                 end = delim_p;
1188                 if (!end) /* case: hostname:/ */
1189                         end = name + namelen;
1190         }
1191
1192         if (end <= name)
1193                 return -EINVAL;
1194
1195         /* do dns_resolve upcall */
1196         ip_len = dns_query(NULL, name, end - name, NULL, &ip_addr, NULL);
1197         if (ip_len > 0)
1198                 ret = ceph_pton(ip_addr, ip_len, ss, -1, NULL);
1199         else
1200                 ret = -ESRCH;
1201
1202         kfree(ip_addr);
1203
1204         *ipend = end;
1205
1206         pr_info("resolve '%.*s' (ret=%d): %s\n", (int)(end - name), name,
1207                         ret, ret ? "failed" : ceph_pr_addr(ss));
1208
1209         return ret;
1210 }
1211 #else
1212 static inline int ceph_dns_resolve_name(const char *name, size_t namelen,
1213                 struct sockaddr_storage *ss, char delim, const char **ipend)
1214 {
1215         return -EINVAL;
1216 }
1217 #endif
1218
1219 /*
1220  * Parse a server name (IP or hostname). If a valid IP address is not found
1221  * then try to extract a hostname to resolve using userspace DNS upcall.
1222  */
1223 static int ceph_parse_server_name(const char *name, size_t namelen,
1224                         struct sockaddr_storage *ss, char delim, const char **ipend)
1225 {
1226         int ret;
1227
1228         ret = ceph_pton(name, namelen, ss, delim, ipend);
1229         if (ret)
1230                 ret = ceph_dns_resolve_name(name, namelen, ss, delim, ipend);
1231
1232         return ret;
1233 }
1234
1235 /*
1236  * Parse an ip[:port] list into an addr array.  Use the default
1237  * monitor port if a port isn't specified.
1238  */
1239 int ceph_parse_ips(const char *c, const char *end,
1240                    struct ceph_entity_addr *addr,
1241                    int max_count, int *count)
1242 {
1243         int i, ret = -EINVAL;
1244         const char *p = c;
1245
1246         dout("parse_ips on '%.*s'\n", (int)(end-c), c);
1247         for (i = 0; i < max_count; i++) {
1248                 const char *ipend;
1249                 struct sockaddr_storage *ss = &addr[i].in_addr;
1250                 int port;
1251                 char delim = ',';
1252
1253                 if (*p == '[') {
1254                         delim = ']';
1255                         p++;
1256                 }
1257
1258                 ret = ceph_parse_server_name(p, end - p, ss, delim, &ipend);
1259                 if (ret)
1260                         goto bad;
1261                 ret = -EINVAL;
1262
1263                 p = ipend;
1264
1265                 if (delim == ']') {
1266                         if (*p != ']') {
1267                                 dout("missing matching ']'\n");
1268                                 goto bad;
1269                         }
1270                         p++;
1271                 }
1272
1273                 /* port? */
1274                 if (p < end && *p == ':') {
1275                         port = 0;
1276                         p++;
1277                         while (p < end && *p >= '0' && *p <= '9') {
1278                                 port = (port * 10) + (*p - '0');
1279                                 p++;
1280                         }
1281                         if (port > 65535 || port == 0)
1282                                 goto bad;
1283                 } else {
1284                         port = CEPH_MON_PORT;
1285                 }
1286
1287                 addr_set_port(ss, port);
1288
1289                 dout("parse_ips got %s\n", ceph_pr_addr(ss));
1290
1291                 if (p == end)
1292                         break;
1293                 if (*p != ',')
1294                         goto bad;
1295                 p++;
1296         }
1297
1298         if (p != end)
1299                 goto bad;
1300
1301         if (count)
1302                 *count = i + 1;
1303         return 0;
1304
1305 bad:
1306         pr_err("parse_ips bad ip '%.*s'\n", (int)(end - c), c);
1307         return ret;
1308 }
1309 EXPORT_SYMBOL(ceph_parse_ips);
1310
1311 static int process_banner(struct ceph_connection *con)
1312 {
1313         dout("process_banner on %p\n", con);
1314
1315         if (verify_hello(con) < 0)
1316                 return -1;
1317
1318         ceph_decode_addr(&con->actual_peer_addr);
1319         ceph_decode_addr(&con->peer_addr_for_me);
1320
1321         /*
1322          * Make sure the other end is who we wanted.  note that the other
1323          * end may not yet know their ip address, so if it's 0.0.0.0, give
1324          * them the benefit of the doubt.
1325          */
1326         if (memcmp(&con->peer_addr, &con->actual_peer_addr,
1327                    sizeof(con->peer_addr)) != 0 &&
1328             !(addr_is_blank(&con->actual_peer_addr.in_addr) &&
1329               con->actual_peer_addr.nonce == con->peer_addr.nonce)) {
1330                 pr_warning("wrong peer, want %s/%d, got %s/%d\n",
1331                            ceph_pr_addr(&con->peer_addr.in_addr),
1332                            (int)le32_to_cpu(con->peer_addr.nonce),
1333                            ceph_pr_addr(&con->actual_peer_addr.in_addr),
1334                            (int)le32_to_cpu(con->actual_peer_addr.nonce));
1335                 con->error_msg = "wrong peer at address";
1336                 return -1;
1337         }
1338
1339         /*
1340          * did we learn our address?
1341          */
1342         if (addr_is_blank(&con->msgr->inst.addr.in_addr)) {
1343                 int port = addr_port(&con->msgr->inst.addr.in_addr);
1344
1345                 memcpy(&con->msgr->inst.addr.in_addr,
1346                        &con->peer_addr_for_me.in_addr,
1347                        sizeof(con->peer_addr_for_me.in_addr));
1348                 addr_set_port(&con->msgr->inst.addr.in_addr, port);
1349                 encode_my_addr(con->msgr);
1350                 dout("process_banner learned my addr is %s\n",
1351                      ceph_pr_addr(&con->msgr->inst.addr.in_addr));
1352         }
1353
1354         set_bit(NEGOTIATING, &con->state);
1355         prepare_read_connect(con);
1356         return 0;
1357 }
1358
1359 static void fail_protocol(struct ceph_connection *con)
1360 {
1361         reset_connection(con);
1362         set_bit(CLOSED, &con->state);  /* in case there's queued work */
1363 }
1364
1365 static int process_connect(struct ceph_connection *con)
1366 {
1367         u64 sup_feat = con->msgr->supported_features;
1368         u64 req_feat = con->msgr->required_features;
1369         u64 server_feat = le64_to_cpu(con->in_reply.features);
1370         int ret;
1371
1372         dout("process_connect on %p tag %d\n", con, (int)con->in_tag);
1373
1374         switch (con->in_reply.tag) {
1375         case CEPH_MSGR_TAG_FEATURES:
1376                 pr_err("%s%lld %s feature set mismatch,"
1377                        " my %llx < server's %llx, missing %llx\n",
1378                        ENTITY_NAME(con->peer_name),
1379                        ceph_pr_addr(&con->peer_addr.in_addr),
1380                        sup_feat, server_feat, server_feat & ~sup_feat);
1381                 con->error_msg = "missing required protocol features";
1382                 fail_protocol(con);
1383                 return -1;
1384
1385         case CEPH_MSGR_TAG_BADPROTOVER:
1386                 pr_err("%s%lld %s protocol version mismatch,"
1387                        " my %d != server's %d\n",
1388                        ENTITY_NAME(con->peer_name),
1389                        ceph_pr_addr(&con->peer_addr.in_addr),
1390                        le32_to_cpu(con->out_connect.protocol_version),
1391                        le32_to_cpu(con->in_reply.protocol_version));
1392                 con->error_msg = "protocol version mismatch";
1393                 fail_protocol(con);
1394                 return -1;
1395
1396         case CEPH_MSGR_TAG_BADAUTHORIZER:
1397                 con->auth_retry++;
1398                 dout("process_connect %p got BADAUTHORIZER attempt %d\n", con,
1399                      con->auth_retry);
1400                 if (con->auth_retry == 2) {
1401                         con->error_msg = "connect authorization failure";
1402                         return -1;
1403                 }
1404                 con->auth_retry = 1;
1405                 con_out_kvec_reset(con);
1406                 ret = prepare_write_connect(con);
1407                 if (ret < 0)
1408                         return ret;
1409                 prepare_read_connect(con);
1410                 break;
1411
1412         case CEPH_MSGR_TAG_RESETSESSION:
1413                 /*
1414                  * If we connected with a large connect_seq but the peer
1415                  * has no record of a session with us (no connection, or
1416                  * connect_seq == 0), they will send RESETSESION to indicate
1417                  * that they must have reset their session, and may have
1418                  * dropped messages.
1419                  */
1420                 dout("process_connect got RESET peer seq %u\n",
1421                      le32_to_cpu(con->in_reply.connect_seq));
1422                 pr_err("%s%lld %s connection reset\n",
1423                        ENTITY_NAME(con->peer_name),
1424                        ceph_pr_addr(&con->peer_addr.in_addr));
1425                 reset_connection(con);
1426                 con_out_kvec_reset(con);
1427                 ret = prepare_write_connect(con);
1428                 if (ret < 0)
1429                         return ret;
1430                 prepare_read_connect(con);
1431
1432                 /* Tell ceph about it. */
1433                 mutex_unlock(&con->mutex);
1434                 pr_info("reset on %s%lld\n", ENTITY_NAME(con->peer_name));
1435                 if (con->ops->peer_reset)
1436                         con->ops->peer_reset(con);
1437                 mutex_lock(&con->mutex);
1438                 if (test_bit(CLOSED, &con->state) ||
1439                     test_bit(OPENING, &con->state))
1440                         return -EAGAIN;
1441                 break;
1442
1443         case CEPH_MSGR_TAG_RETRY_SESSION:
1444                 /*
1445                  * If we sent a smaller connect_seq than the peer has, try
1446                  * again with a larger value.
1447                  */
1448                 dout("process_connect got RETRY_SESSION my seq %u, peer %u\n",
1449                      le32_to_cpu(con->out_connect.connect_seq),
1450                      le32_to_cpu(con->in_reply.connect_seq));
1451                 con->connect_seq = le32_to_cpu(con->in_reply.connect_seq);
1452                 con_out_kvec_reset(con);
1453                 ret = prepare_write_connect(con);
1454                 if (ret < 0)
1455                         return ret;
1456                 prepare_read_connect(con);
1457                 break;
1458
1459         case CEPH_MSGR_TAG_RETRY_GLOBAL:
1460                 /*
1461                  * If we sent a smaller global_seq than the peer has, try
1462                  * again with a larger value.
1463                  */
1464                 dout("process_connect got RETRY_GLOBAL my %u peer_gseq %u\n",
1465                      con->peer_global_seq,
1466                      le32_to_cpu(con->in_reply.global_seq));
1467                 get_global_seq(con->msgr,
1468                                le32_to_cpu(con->in_reply.global_seq));
1469                 con_out_kvec_reset(con);
1470                 ret = prepare_write_connect(con);
1471                 if (ret < 0)
1472                         return ret;
1473                 prepare_read_connect(con);
1474                 break;
1475
1476         case CEPH_MSGR_TAG_READY:
1477                 if (req_feat & ~server_feat) {
1478                         pr_err("%s%lld %s protocol feature mismatch,"
1479                                " my required %llx > server's %llx, need %llx\n",
1480                                ENTITY_NAME(con->peer_name),
1481                                ceph_pr_addr(&con->peer_addr.in_addr),
1482                                req_feat, server_feat, req_feat & ~server_feat);
1483                         con->error_msg = "missing required protocol features";
1484                         fail_protocol(con);
1485                         return -1;
1486                 }
1487                 clear_bit(CONNECTING, &con->state);
1488                 con->peer_global_seq = le32_to_cpu(con->in_reply.global_seq);
1489                 con->connect_seq++;
1490                 con->peer_features = server_feat;
1491                 dout("process_connect got READY gseq %d cseq %d (%d)\n",
1492                      con->peer_global_seq,
1493                      le32_to_cpu(con->in_reply.connect_seq),
1494                      con->connect_seq);
1495                 WARN_ON(con->connect_seq !=
1496                         le32_to_cpu(con->in_reply.connect_seq));
1497
1498                 if (con->in_reply.flags & CEPH_MSG_CONNECT_LOSSY)
1499                         set_bit(LOSSYTX, &con->state);
1500
1501                 prepare_read_tag(con);
1502                 break;
1503
1504         case CEPH_MSGR_TAG_WAIT:
1505                 /*
1506                  * If there is a connection race (we are opening
1507                  * connections to each other), one of us may just have
1508                  * to WAIT.  This shouldn't happen if we are the
1509                  * client.
1510                  */
1511                 pr_err("process_connect got WAIT as client\n");
1512                 con->error_msg = "protocol error, got WAIT as client";
1513                 return -1;
1514
1515         default:
1516                 pr_err("connect protocol error, will retry\n");
1517                 con->error_msg = "protocol error, garbage tag during connect";
1518                 return -1;
1519         }
1520         return 0;
1521 }
1522
1523
1524 /*
1525  * read (part of) an ack
1526  */
1527 static int read_partial_ack(struct ceph_connection *con)
1528 {
1529         int size = sizeof (con->in_temp_ack);
1530         int end = size;
1531
1532         return read_partial(con, end, size, &con->in_temp_ack);
1533 }
1534
1535
1536 /*
1537  * We can finally discard anything that's been acked.
1538  */
1539 static void process_ack(struct ceph_connection *con)
1540 {
1541         struct ceph_msg *m;
1542         u64 ack = le64_to_cpu(con->in_temp_ack);
1543         u64 seq;
1544
1545         while (!list_empty(&con->out_sent)) {
1546                 m = list_first_entry(&con->out_sent, struct ceph_msg,
1547                                      list_head);
1548                 seq = le64_to_cpu(m->hdr.seq);
1549                 if (seq > ack)
1550                         break;
1551                 dout("got ack for seq %llu type %d at %p\n", seq,
1552                      le16_to_cpu(m->hdr.type), m);
1553                 m->ack_stamp = jiffies;
1554                 ceph_msg_remove(m);
1555         }
1556         prepare_read_tag(con);
1557 }
1558
1559
1560
1561
1562 static int read_partial_message_section(struct ceph_connection *con,
1563                                         struct kvec *section,
1564                                         unsigned int sec_len, u32 *crc)
1565 {
1566         int ret, left;
1567
1568         BUG_ON(!section);
1569
1570         while (section->iov_len < sec_len) {
1571                 BUG_ON(section->iov_base == NULL);
1572                 left = sec_len - section->iov_len;
1573                 ret = ceph_tcp_recvmsg(con->sock, (char *)section->iov_base +
1574                                        section->iov_len, left);
1575                 if (ret <= 0)
1576                         return ret;
1577                 section->iov_len += ret;
1578         }
1579         if (section->iov_len == sec_len)
1580                 *crc = crc32c(0, section->iov_base, section->iov_len);
1581
1582         return 1;
1583 }
1584
1585 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
1586                                 struct ceph_msg_header *hdr,
1587                                 int *skip);
1588
1589
1590 static int read_partial_message_pages(struct ceph_connection *con,
1591                                       struct page **pages,
1592                                       unsigned data_len, bool do_datacrc)
1593 {
1594         void *p;
1595         int ret;
1596         int left;
1597
1598         left = min((int)(data_len - con->in_msg_pos.data_pos),
1599                    (int)(PAGE_SIZE - con->in_msg_pos.page_pos));
1600         /* (page) data */
1601         BUG_ON(pages == NULL);
1602         p = kmap(pages[con->in_msg_pos.page]);
1603         ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1604                                left);
1605         if (ret > 0 && do_datacrc)
1606                 con->in_data_crc =
1607                         crc32c(con->in_data_crc,
1608                                   p + con->in_msg_pos.page_pos, ret);
1609         kunmap(pages[con->in_msg_pos.page]);
1610         if (ret <= 0)
1611                 return ret;
1612         con->in_msg_pos.data_pos += ret;
1613         con->in_msg_pos.page_pos += ret;
1614         if (con->in_msg_pos.page_pos == PAGE_SIZE) {
1615                 con->in_msg_pos.page_pos = 0;
1616                 con->in_msg_pos.page++;
1617         }
1618
1619         return ret;
1620 }
1621
1622 #ifdef CONFIG_BLOCK
1623 static int read_partial_message_bio(struct ceph_connection *con,
1624                                     struct bio **bio_iter, int *bio_seg,
1625                                     unsigned data_len, bool do_datacrc)
1626 {
1627         struct bio_vec *bv = bio_iovec_idx(*bio_iter, *bio_seg);
1628         void *p;
1629         int ret, left;
1630
1631         if (IS_ERR(bv))
1632                 return PTR_ERR(bv);
1633
1634         left = min((int)(data_len - con->in_msg_pos.data_pos),
1635                    (int)(bv->bv_len - con->in_msg_pos.page_pos));
1636
1637         p = kmap(bv->bv_page) + bv->bv_offset;
1638
1639         ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1640                                left);
1641         if (ret > 0 && do_datacrc)
1642                 con->in_data_crc =
1643                         crc32c(con->in_data_crc,
1644                                   p + con->in_msg_pos.page_pos, ret);
1645         kunmap(bv->bv_page);
1646         if (ret <= 0)
1647                 return ret;
1648         con->in_msg_pos.data_pos += ret;
1649         con->in_msg_pos.page_pos += ret;
1650         if (con->in_msg_pos.page_pos == bv->bv_len) {
1651                 con->in_msg_pos.page_pos = 0;
1652                 iter_bio_next(bio_iter, bio_seg);
1653         }
1654
1655         return ret;
1656 }
1657 #endif
1658
1659 /*
1660  * read (part of) a message.
1661  */
1662 static int read_partial_message(struct ceph_connection *con)
1663 {
1664         struct ceph_msg *m = con->in_msg;
1665         int size;
1666         int end;
1667         int ret;
1668         unsigned front_len, middle_len, data_len;
1669         bool do_datacrc = !con->msgr->nocrc;
1670         int skip;
1671         u64 seq;
1672         u32 crc;
1673
1674         dout("read_partial_message con %p msg %p\n", con, m);
1675
1676         /* header */
1677         size = sizeof (con->in_hdr);
1678         end = size;
1679         ret = read_partial(con, end, size, &con->in_hdr);
1680         if (ret <= 0)
1681                 return ret;
1682
1683         crc = crc32c(0, &con->in_hdr, offsetof(struct ceph_msg_header, crc));
1684         if (cpu_to_le32(crc) != con->in_hdr.crc) {
1685                 pr_err("read_partial_message bad hdr "
1686                        " crc %u != expected %u\n",
1687                        crc, con->in_hdr.crc);
1688                 return -EBADMSG;
1689         }
1690
1691         front_len = le32_to_cpu(con->in_hdr.front_len);
1692         if (front_len > CEPH_MSG_MAX_FRONT_LEN)
1693                 return -EIO;
1694         middle_len = le32_to_cpu(con->in_hdr.middle_len);
1695         if (middle_len > CEPH_MSG_MAX_DATA_LEN)
1696                 return -EIO;
1697         data_len = le32_to_cpu(con->in_hdr.data_len);
1698         if (data_len > CEPH_MSG_MAX_DATA_LEN)
1699                 return -EIO;
1700
1701         /* verify seq# */
1702         seq = le64_to_cpu(con->in_hdr.seq);
1703         if ((s64)seq - (s64)con->in_seq < 1) {
1704                 pr_info("skipping %s%lld %s seq %lld expected %lld\n",
1705                         ENTITY_NAME(con->peer_name),
1706                         ceph_pr_addr(&con->peer_addr.in_addr),
1707                         seq, con->in_seq + 1);
1708                 con->in_base_pos = -front_len - middle_len - data_len -
1709                         sizeof(m->footer);
1710                 con->in_tag = CEPH_MSGR_TAG_READY;
1711                 return 0;
1712         } else if ((s64)seq - (s64)con->in_seq > 1) {
1713                 pr_err("read_partial_message bad seq %lld expected %lld\n",
1714                        seq, con->in_seq + 1);
1715                 con->error_msg = "bad message sequence # for incoming message";
1716                 return -EBADMSG;
1717         }
1718
1719         /* allocate message? */
1720         if (!con->in_msg) {
1721                 dout("got hdr type %d front %d data %d\n", con->in_hdr.type,
1722                      con->in_hdr.front_len, con->in_hdr.data_len);
1723                 skip = 0;
1724                 con->in_msg = ceph_alloc_msg(con, &con->in_hdr, &skip);
1725                 if (skip) {
1726                         /* skip this message */
1727                         dout("alloc_msg said skip message\n");
1728                         BUG_ON(con->in_msg);
1729                         con->in_base_pos = -front_len - middle_len - data_len -
1730                                 sizeof(m->footer);
1731                         con->in_tag = CEPH_MSGR_TAG_READY;
1732                         con->in_seq++;
1733                         return 0;
1734                 }
1735                 if (!con->in_msg) {
1736                         con->error_msg =
1737                                 "error allocating memory for incoming message";
1738                         return -ENOMEM;
1739                 }
1740                 m = con->in_msg;
1741                 m->front.iov_len = 0;    /* haven't read it yet */
1742                 if (m->middle)
1743                         m->middle->vec.iov_len = 0;
1744
1745                 con->in_msg_pos.page = 0;
1746                 if (m->pages)
1747                         con->in_msg_pos.page_pos = m->page_alignment;
1748                 else
1749                         con->in_msg_pos.page_pos = 0;
1750                 con->in_msg_pos.data_pos = 0;
1751         }
1752
1753         /* front */
1754         ret = read_partial_message_section(con, &m->front, front_len,
1755                                            &con->in_front_crc);
1756         if (ret <= 0)
1757                 return ret;
1758
1759         /* middle */
1760         if (m->middle) {
1761                 ret = read_partial_message_section(con, &m->middle->vec,
1762                                                    middle_len,
1763                                                    &con->in_middle_crc);
1764                 if (ret <= 0)
1765                         return ret;
1766         }
1767 #ifdef CONFIG_BLOCK
1768         if (m->bio && !m->bio_iter)
1769                 init_bio_iter(m->bio, &m->bio_iter, &m->bio_seg);
1770 #endif
1771
1772         /* (page) data */
1773         while (con->in_msg_pos.data_pos < data_len) {
1774                 if (m->pages) {
1775                         ret = read_partial_message_pages(con, m->pages,
1776                                                  data_len, do_datacrc);
1777                         if (ret <= 0)
1778                                 return ret;
1779 #ifdef CONFIG_BLOCK
1780                 } else if (m->bio) {
1781
1782                         ret = read_partial_message_bio(con,
1783                                                  &m->bio_iter, &m->bio_seg,
1784                                                  data_len, do_datacrc);
1785                         if (ret <= 0)
1786                                 return ret;
1787 #endif
1788                 } else {
1789                         BUG_ON(1);
1790                 }
1791         }
1792
1793         /* footer */
1794         size = sizeof (m->footer);
1795         end += size;
1796         ret = read_partial(con, end, size, &m->footer);
1797         if (ret <= 0)
1798                 return ret;
1799
1800         dout("read_partial_message got msg %p %d (%u) + %d (%u) + %d (%u)\n",
1801              m, front_len, m->footer.front_crc, middle_len,
1802              m->footer.middle_crc, data_len, m->footer.data_crc);
1803
1804         /* crc ok? */
1805         if (con->in_front_crc != le32_to_cpu(m->footer.front_crc)) {
1806                 pr_err("read_partial_message %p front crc %u != exp. %u\n",
1807                        m, con->in_front_crc, m->footer.front_crc);
1808                 return -EBADMSG;
1809         }
1810         if (con->in_middle_crc != le32_to_cpu(m->footer.middle_crc)) {
1811                 pr_err("read_partial_message %p middle crc %u != exp %u\n",
1812                        m, con->in_middle_crc, m->footer.middle_crc);
1813                 return -EBADMSG;
1814         }
1815         if (do_datacrc &&
1816             (m->footer.flags & CEPH_MSG_FOOTER_NOCRC) == 0 &&
1817             con->in_data_crc != le32_to_cpu(m->footer.data_crc)) {
1818                 pr_err("read_partial_message %p data crc %u != exp. %u\n", m,
1819                        con->in_data_crc, le32_to_cpu(m->footer.data_crc));
1820                 return -EBADMSG;
1821         }
1822
1823         return 1; /* done! */
1824 }
1825
1826 /*
1827  * Process message.  This happens in the worker thread.  The callback should
1828  * be careful not to do anything that waits on other incoming messages or it
1829  * may deadlock.
1830  */
1831 static void process_message(struct ceph_connection *con)
1832 {
1833         struct ceph_msg *msg;
1834
1835         msg = con->in_msg;
1836         con->in_msg = NULL;
1837
1838         /* if first message, set peer_name */
1839         if (con->peer_name.type == 0)
1840                 con->peer_name = msg->hdr.src;
1841
1842         con->in_seq++;
1843         mutex_unlock(&con->mutex);
1844
1845         dout("===== %p %llu from %s%lld %d=%s len %d+%d (%u %u %u) =====\n",
1846              msg, le64_to_cpu(msg->hdr.seq),
1847              ENTITY_NAME(msg->hdr.src),
1848              le16_to_cpu(msg->hdr.type),
1849              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1850              le32_to_cpu(msg->hdr.front_len),
1851              le32_to_cpu(msg->hdr.data_len),
1852              con->in_front_crc, con->in_middle_crc, con->in_data_crc);
1853         con->ops->dispatch(con, msg);
1854
1855         mutex_lock(&con->mutex);
1856         prepare_read_tag(con);
1857 }
1858
1859
1860 /*
1861  * Write something to the socket.  Called in a worker thread when the
1862  * socket appears to be writeable and we have something ready to send.
1863  */
1864 static int try_write(struct ceph_connection *con)
1865 {
1866         int ret = 1;
1867
1868         dout("try_write start %p state %lu nref %d\n", con, con->state,
1869              atomic_read(&con->nref));
1870
1871 more:
1872         dout("try_write out_kvec_bytes %d\n", con->out_kvec_bytes);
1873
1874         /* open the socket first? */
1875         if (con->sock == NULL) {
1876                 con_out_kvec_reset(con);
1877                 prepare_write_banner(con);
1878                 ret = prepare_write_connect(con);
1879                 if (ret < 0)
1880                         goto out;
1881                 prepare_read_banner(con);
1882                 set_bit(CONNECTING, &con->state);
1883                 clear_bit(NEGOTIATING, &con->state);
1884
1885                 BUG_ON(con->in_msg);
1886                 con->in_tag = CEPH_MSGR_TAG_READY;
1887                 dout("try_write initiating connect on %p new state %lu\n",
1888                      con, con->state);
1889                 ret = ceph_tcp_connect(con);
1890                 if (ret < 0) {
1891                         con->error_msg = "connect error";
1892                         goto out;
1893                 }
1894         }
1895
1896 more_kvec:
1897         /* kvec data queued? */
1898         if (con->out_skip) {
1899                 ret = write_partial_skip(con);
1900                 if (ret <= 0)
1901                         goto out;
1902         }
1903         if (con->out_kvec_left) {
1904                 ret = write_partial_kvec(con);
1905                 if (ret <= 0)
1906                         goto out;
1907         }
1908
1909         /* msg pages? */
1910         if (con->out_msg) {
1911                 if (con->out_msg_done) {
1912                         ceph_msg_put(con->out_msg);
1913                         con->out_msg = NULL;   /* we're done with this one */
1914                         goto do_next;
1915                 }
1916
1917                 ret = write_partial_msg_pages(con);
1918                 if (ret == 1)
1919                         goto more_kvec;  /* we need to send the footer, too! */
1920                 if (ret == 0)
1921                         goto out;
1922                 if (ret < 0) {
1923                         dout("try_write write_partial_msg_pages err %d\n",
1924                              ret);
1925                         goto out;
1926                 }
1927         }
1928
1929 do_next:
1930         if (!test_bit(CONNECTING, &con->state)) {
1931                 /* is anything else pending? */
1932                 if (!list_empty(&con->out_queue)) {
1933                         prepare_write_message(con);
1934                         goto more;
1935                 }
1936                 if (con->in_seq > con->in_seq_acked) {
1937                         prepare_write_ack(con);
1938                         goto more;
1939                 }
1940                 if (test_and_clear_bit(KEEPALIVE_PENDING, &con->state)) {
1941                         prepare_write_keepalive(con);
1942                         goto more;
1943                 }
1944         }
1945
1946         /* Nothing to do! */
1947         clear_bit(WRITE_PENDING, &con->state);
1948         dout("try_write nothing else to write.\n");
1949         ret = 0;
1950 out:
1951         dout("try_write done on %p ret %d\n", con, ret);
1952         return ret;
1953 }
1954
1955
1956
1957 /*
1958  * Read what we can from the socket.
1959  */
1960 static int try_read(struct ceph_connection *con)
1961 {
1962         int ret = -1;
1963
1964         if (!con->sock)
1965                 return 0;
1966
1967         if (test_bit(STANDBY, &con->state))
1968                 return 0;
1969
1970         dout("try_read start on %p\n", con);
1971
1972 more:
1973         dout("try_read tag %d in_base_pos %d\n", (int)con->in_tag,
1974              con->in_base_pos);
1975
1976         /*
1977          * process_connect and process_message drop and re-take
1978          * con->mutex.  make sure we handle a racing close or reopen.
1979          */
1980         if (test_bit(CLOSED, &con->state) ||
1981             test_bit(OPENING, &con->state)) {
1982                 ret = -EAGAIN;
1983                 goto out;
1984         }
1985
1986         if (test_bit(CONNECTING, &con->state)) {
1987                 if (!test_bit(NEGOTIATING, &con->state)) {
1988                         dout("try_read connecting\n");
1989                         ret = read_partial_banner(con);
1990                         if (ret <= 0)
1991                                 goto out;
1992                         ret = process_banner(con);
1993                         if (ret < 0)
1994                                 goto out;
1995                 }
1996                 ret = read_partial_connect(con);
1997                 if (ret <= 0)
1998                         goto out;
1999                 ret = process_connect(con);
2000                 if (ret < 0)
2001                         goto out;
2002                 goto more;
2003         }
2004
2005         if (con->in_base_pos < 0) {
2006                 /*
2007                  * skipping + discarding content.
2008                  *
2009                  * FIXME: there must be a better way to do this!
2010                  */
2011                 static char buf[SKIP_BUF_SIZE];
2012                 int skip = min((int) sizeof (buf), -con->in_base_pos);
2013
2014                 dout("skipping %d / %d bytes\n", skip, -con->in_base_pos);
2015                 ret = ceph_tcp_recvmsg(con->sock, buf, skip);
2016                 if (ret <= 0)
2017                         goto out;
2018                 con->in_base_pos += ret;
2019                 if (con->in_base_pos)
2020                         goto more;
2021         }
2022         if (con->in_tag == CEPH_MSGR_TAG_READY) {
2023                 /*
2024                  * what's next?
2025                  */
2026                 ret = ceph_tcp_recvmsg(con->sock, &con->in_tag, 1);
2027                 if (ret <= 0)
2028                         goto out;
2029                 dout("try_read got tag %d\n", (int)con->in_tag);
2030                 switch (con->in_tag) {
2031                 case CEPH_MSGR_TAG_MSG:
2032                         prepare_read_message(con);
2033                         break;
2034                 case CEPH_MSGR_TAG_ACK:
2035                         prepare_read_ack(con);
2036                         break;
2037                 case CEPH_MSGR_TAG_CLOSE:
2038                         set_bit(CLOSED, &con->state);   /* fixme */
2039                         goto out;
2040                 default:
2041                         goto bad_tag;
2042                 }
2043         }
2044         if (con->in_tag == CEPH_MSGR_TAG_MSG) {
2045                 ret = read_partial_message(con);
2046                 if (ret <= 0) {
2047                         switch (ret) {
2048                         case -EBADMSG:
2049                                 con->error_msg = "bad crc";
2050                                 ret = -EIO;
2051                                 break;
2052                         case -EIO:
2053                                 con->error_msg = "io error";
2054                                 break;
2055                         }
2056                         goto out;
2057                 }
2058                 if (con->in_tag == CEPH_MSGR_TAG_READY)
2059                         goto more;
2060                 process_message(con);
2061                 goto more;
2062         }
2063         if (con->in_tag == CEPH_MSGR_TAG_ACK) {
2064                 ret = read_partial_ack(con);
2065                 if (ret <= 0)
2066                         goto out;
2067                 process_ack(con);
2068                 goto more;
2069         }
2070
2071 out:
2072         dout("try_read done on %p ret %d\n", con, ret);
2073         return ret;
2074
2075 bad_tag:
2076         pr_err("try_read bad con->in_tag = %d\n", (int)con->in_tag);
2077         con->error_msg = "protocol error, garbage tag";
2078         ret = -1;
2079         goto out;
2080 }
2081
2082
2083 /*
2084  * Atomically queue work on a connection.  Bump @con reference to
2085  * avoid races with connection teardown.
2086  */
2087 static void queue_con(struct ceph_connection *con)
2088 {
2089         if (!con->ops->get(con)) {
2090                 dout("queue_con %p ref count 0\n", con);
2091                 return;
2092         }
2093
2094         if (!queue_delayed_work(ceph_msgr_wq, &con->work, 0)) {
2095                 dout("queue_con %p - already queued\n", con);
2096                 con->ops->put(con);
2097         } else {
2098                 dout("queue_con %p\n", con);
2099         }
2100 }
2101
2102 /*
2103  * Do some work on a connection.  Drop a connection ref when we're done.
2104  */
2105 static void con_work(struct work_struct *work)
2106 {
2107         struct ceph_connection *con = container_of(work, struct ceph_connection,
2108                                                    work.work);
2109         int ret;
2110
2111         mutex_lock(&con->mutex);
2112 restart:
2113         if (test_and_clear_bit(BACKOFF, &con->state)) {
2114                 dout("con_work %p backing off\n", con);
2115                 if (queue_delayed_work(ceph_msgr_wq, &con->work,
2116                                        round_jiffies_relative(con->delay))) {
2117                         dout("con_work %p backoff %lu\n", con, con->delay);
2118                         mutex_unlock(&con->mutex);
2119                         return;
2120                 } else {
2121                         con->ops->put(con);
2122                         dout("con_work %p FAILED to back off %lu\n", con,
2123                              con->delay);
2124                 }
2125         }
2126
2127         if (test_bit(STANDBY, &con->state)) {
2128                 dout("con_work %p STANDBY\n", con);
2129                 goto done;
2130         }
2131         if (test_bit(CLOSED, &con->state)) { /* e.g. if we are replaced */
2132                 dout("con_work CLOSED\n");
2133                 con_close_socket(con);
2134                 goto done;
2135         }
2136         if (test_and_clear_bit(OPENING, &con->state)) {
2137                 /* reopen w/ new peer */
2138                 dout("con_work OPENING\n");
2139                 con_close_socket(con);
2140         }
2141
2142         if (test_and_clear_bit(SOCK_CLOSED, &con->state))
2143                 goto fault;
2144
2145         ret = try_read(con);
2146         if (ret == -EAGAIN)
2147                 goto restart;
2148         if (ret < 0)
2149                 goto fault;
2150
2151         ret = try_write(con);
2152         if (ret == -EAGAIN)
2153                 goto restart;
2154         if (ret < 0)
2155                 goto fault;
2156
2157 done:
2158         mutex_unlock(&con->mutex);
2159 done_unlocked:
2160         con->ops->put(con);
2161         return;
2162
2163 fault:
2164         mutex_unlock(&con->mutex);
2165         ceph_fault(con);     /* error/fault path */
2166         goto done_unlocked;
2167 }
2168
2169
2170 /*
2171  * Generic error/fault handler.  A retry mechanism is used with
2172  * exponential backoff
2173  */
2174 static void ceph_fault(struct ceph_connection *con)
2175 {
2176         pr_err("%s%lld %s %s\n", ENTITY_NAME(con->peer_name),
2177                ceph_pr_addr(&con->peer_addr.in_addr), con->error_msg);
2178         dout("fault %p state %lu to peer %s\n",
2179              con, con->state, ceph_pr_addr(&con->peer_addr.in_addr));
2180
2181         if (test_bit(LOSSYTX, &con->state)) {
2182                 dout("fault on LOSSYTX channel\n");
2183                 goto out;
2184         }
2185
2186         mutex_lock(&con->mutex);
2187         if (test_bit(CLOSED, &con->state))
2188                 goto out_unlock;
2189
2190         con_close_socket(con);
2191
2192         if (con->in_msg) {
2193                 ceph_msg_put(con->in_msg);
2194                 con->in_msg = NULL;
2195         }
2196
2197         /* Requeue anything that hasn't been acked */
2198         list_splice_init(&con->out_sent, &con->out_queue);
2199
2200         /* If there are no messages queued or keepalive pending, place
2201          * the connection in a STANDBY state */
2202         if (list_empty(&con->out_queue) &&
2203             !test_bit(KEEPALIVE_PENDING, &con->state)) {
2204                 dout("fault %p setting STANDBY clearing WRITE_PENDING\n", con);
2205                 clear_bit(WRITE_PENDING, &con->state);
2206                 set_bit(STANDBY, &con->state);
2207         } else {
2208                 /* retry after a delay. */
2209                 if (con->delay == 0)
2210                         con->delay = BASE_DELAY_INTERVAL;
2211                 else if (con->delay < MAX_DELAY_INTERVAL)
2212                         con->delay *= 2;
2213                 con->ops->get(con);
2214                 if (queue_delayed_work(ceph_msgr_wq, &con->work,
2215                                        round_jiffies_relative(con->delay))) {
2216                         dout("fault queued %p delay %lu\n", con, con->delay);
2217                 } else {
2218                         con->ops->put(con);
2219                         dout("fault failed to queue %p delay %lu, backoff\n",
2220                              con, con->delay);
2221                         /*
2222                          * In many cases we see a socket state change
2223                          * while con_work is running and end up
2224                          * queuing (non-delayed) work, such that we
2225                          * can't backoff with a delay.  Set a flag so
2226                          * that when con_work restarts we schedule the
2227                          * delay then.
2228                          */
2229                         set_bit(BACKOFF, &con->state);
2230                 }
2231         }
2232
2233 out_unlock:
2234         mutex_unlock(&con->mutex);
2235 out:
2236         /*
2237          * in case we faulted due to authentication, invalidate our
2238          * current tickets so that we can get new ones.
2239          */
2240         if (con->auth_retry && con->ops->invalidate_authorizer) {
2241                 dout("calling invalidate_authorizer()\n");
2242                 con->ops->invalidate_authorizer(con);
2243         }
2244
2245         if (con->ops->fault)
2246                 con->ops->fault(con);
2247 }
2248
2249
2250
2251 /*
2252  * initialize a new messenger instance
2253  */
2254 void ceph_messenger_init(struct ceph_messenger *msgr,
2255                         struct ceph_entity_addr *myaddr,
2256                         u32 supported_features,
2257                         u32 required_features,
2258                         bool nocrc)
2259 {
2260         msgr->supported_features = supported_features;
2261         msgr->required_features = required_features;
2262
2263         spin_lock_init(&msgr->global_seq_lock);
2264
2265         if (myaddr)
2266                 msgr->inst.addr = *myaddr;
2267
2268         /* select a random nonce */
2269         msgr->inst.addr.type = 0;
2270         get_random_bytes(&msgr->inst.addr.nonce, sizeof(msgr->inst.addr.nonce));
2271         encode_my_addr(msgr);
2272         msgr->nocrc = nocrc;
2273
2274         dout("%s %p\n", __func__, msgr);
2275 }
2276 EXPORT_SYMBOL(ceph_messenger_init);
2277
2278 static void clear_standby(struct ceph_connection *con)
2279 {
2280         /* come back from STANDBY? */
2281         if (test_and_clear_bit(STANDBY, &con->state)) {
2282                 mutex_lock(&con->mutex);
2283                 dout("clear_standby %p and ++connect_seq\n", con);
2284                 con->connect_seq++;
2285                 WARN_ON(test_bit(WRITE_PENDING, &con->state));
2286                 WARN_ON(test_bit(KEEPALIVE_PENDING, &con->state));
2287                 mutex_unlock(&con->mutex);
2288         }
2289 }
2290
2291 /*
2292  * Queue up an outgoing message on the given connection.
2293  */
2294 void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg)
2295 {
2296         if (test_bit(CLOSED, &con->state)) {
2297                 dout("con_send %p closed, dropping %p\n", con, msg);
2298                 ceph_msg_put(msg);
2299                 return;
2300         }
2301
2302         /* set src+dst */
2303         msg->hdr.src = con->msgr->inst.name;
2304
2305         BUG_ON(msg->front.iov_len != le32_to_cpu(msg->hdr.front_len));
2306
2307         msg->needs_out_seq = true;
2308
2309         /* queue */
2310         mutex_lock(&con->mutex);
2311         BUG_ON(!list_empty(&msg->list_head));
2312         list_add_tail(&msg->list_head, &con->out_queue);
2313         dout("----- %p to %s%lld %d=%s len %d+%d+%d -----\n", msg,
2314              ENTITY_NAME(con->peer_name), le16_to_cpu(msg->hdr.type),
2315              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
2316              le32_to_cpu(msg->hdr.front_len),
2317              le32_to_cpu(msg->hdr.middle_len),
2318              le32_to_cpu(msg->hdr.data_len));
2319         mutex_unlock(&con->mutex);
2320
2321         /* if there wasn't anything waiting to send before, queue
2322          * new work */
2323         clear_standby(con);
2324         if (test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2325                 queue_con(con);
2326 }
2327 EXPORT_SYMBOL(ceph_con_send);
2328
2329 /*
2330  * Revoke a message that was previously queued for send
2331  */
2332 void ceph_con_revoke(struct ceph_connection *con, struct ceph_msg *msg)
2333 {
2334         mutex_lock(&con->mutex);
2335         if (!list_empty(&msg->list_head)) {
2336                 dout("con_revoke %p msg %p - was on queue\n", con, msg);
2337                 list_del_init(&msg->list_head);
2338                 ceph_msg_put(msg);
2339                 msg->hdr.seq = 0;
2340         }
2341         if (con->out_msg == msg) {
2342                 dout("con_revoke %p msg %p - was sending\n", con, msg);
2343                 con->out_msg = NULL;
2344                 if (con->out_kvec_is_msg) {
2345                         con->out_skip = con->out_kvec_bytes;
2346                         con->out_kvec_is_msg = false;
2347                 }
2348                 ceph_msg_put(msg);
2349                 msg->hdr.seq = 0;
2350         }
2351         mutex_unlock(&con->mutex);
2352 }
2353
2354 /*
2355  * Revoke a message that we may be reading data into
2356  */
2357 void ceph_con_revoke_message(struct ceph_connection *con, struct ceph_msg *msg)
2358 {
2359         mutex_lock(&con->mutex);
2360         if (con->in_msg && con->in_msg == msg) {
2361                 unsigned front_len = le32_to_cpu(con->in_hdr.front_len);
2362                 unsigned middle_len = le32_to_cpu(con->in_hdr.middle_len);
2363                 unsigned data_len = le32_to_cpu(con->in_hdr.data_len);
2364
2365                 /* skip rest of message */
2366                 dout("con_revoke_pages %p msg %p revoked\n", con, msg);
2367                         con->in_base_pos = con->in_base_pos -
2368                                 sizeof(struct ceph_msg_header) -
2369                                 front_len -
2370                                 middle_len -
2371                                 data_len -
2372                                 sizeof(struct ceph_msg_footer);
2373                 ceph_msg_put(con->in_msg);
2374                 con->in_msg = NULL;
2375                 con->in_tag = CEPH_MSGR_TAG_READY;
2376                 con->in_seq++;
2377         } else {
2378                 dout("con_revoke_pages %p msg %p pages %p no-op\n",
2379                      con, con->in_msg, msg);
2380         }
2381         mutex_unlock(&con->mutex);
2382 }
2383
2384 /*
2385  * Queue a keepalive byte to ensure the tcp connection is alive.
2386  */
2387 void ceph_con_keepalive(struct ceph_connection *con)
2388 {
2389         dout("con_keepalive %p\n", con);
2390         clear_standby(con);
2391         if (test_and_set_bit(KEEPALIVE_PENDING, &con->state) == 0 &&
2392             test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2393                 queue_con(con);
2394 }
2395 EXPORT_SYMBOL(ceph_con_keepalive);
2396
2397
2398 /*
2399  * construct a new message with given type, size
2400  * the new msg has a ref count of 1.
2401  */
2402 struct ceph_msg *ceph_msg_new(int type, int front_len, gfp_t flags,
2403                               bool can_fail)
2404 {
2405         struct ceph_msg *m;
2406
2407         m = kmalloc(sizeof(*m), flags);
2408         if (m == NULL)
2409                 goto out;
2410         kref_init(&m->kref);
2411         INIT_LIST_HEAD(&m->list_head);
2412
2413         m->hdr.tid = 0;
2414         m->hdr.type = cpu_to_le16(type);
2415         m->hdr.priority = cpu_to_le16(CEPH_MSG_PRIO_DEFAULT);
2416         m->hdr.version = 0;
2417         m->hdr.front_len = cpu_to_le32(front_len);
2418         m->hdr.middle_len = 0;
2419         m->hdr.data_len = 0;
2420         m->hdr.data_off = 0;
2421         m->hdr.reserved = 0;
2422         m->footer.front_crc = 0;
2423         m->footer.middle_crc = 0;
2424         m->footer.data_crc = 0;
2425         m->footer.flags = 0;
2426         m->front_max = front_len;
2427         m->front_is_vmalloc = false;
2428         m->more_to_follow = false;
2429         m->ack_stamp = 0;
2430         m->pool = NULL;
2431
2432         /* middle */
2433         m->middle = NULL;
2434
2435         /* data */
2436         m->nr_pages = 0;
2437         m->page_alignment = 0;
2438         m->pages = NULL;
2439         m->pagelist = NULL;
2440         m->bio = NULL;
2441         m->bio_iter = NULL;
2442         m->bio_seg = 0;
2443         m->trail = NULL;
2444
2445         /* front */
2446         if (front_len) {
2447                 if (front_len > PAGE_CACHE_SIZE) {
2448                         m->front.iov_base = __vmalloc(front_len, flags,
2449                                                       PAGE_KERNEL);
2450                         m->front_is_vmalloc = true;
2451                 } else {
2452                         m->front.iov_base = kmalloc(front_len, flags);
2453                 }
2454                 if (m->front.iov_base == NULL) {
2455                         dout("ceph_msg_new can't allocate %d bytes\n",
2456                              front_len);
2457                         goto out2;
2458                 }
2459         } else {
2460                 m->front.iov_base = NULL;
2461         }
2462         m->front.iov_len = front_len;
2463
2464         dout("ceph_msg_new %p front %d\n", m, front_len);
2465         return m;
2466
2467 out2:
2468         ceph_msg_put(m);
2469 out:
2470         if (!can_fail) {
2471                 pr_err("msg_new can't create type %d front %d\n", type,
2472                        front_len);
2473                 WARN_ON(1);
2474         } else {
2475                 dout("msg_new can't create type %d front %d\n", type,
2476                      front_len);
2477         }
2478         return NULL;
2479 }
2480 EXPORT_SYMBOL(ceph_msg_new);
2481
2482 /*
2483  * Allocate "middle" portion of a message, if it is needed and wasn't
2484  * allocated by alloc_msg.  This allows us to read a small fixed-size
2485  * per-type header in the front and then gracefully fail (i.e.,
2486  * propagate the error to the caller based on info in the front) when
2487  * the middle is too large.
2488  */
2489 static int ceph_alloc_middle(struct ceph_connection *con, struct ceph_msg *msg)
2490 {
2491         int type = le16_to_cpu(msg->hdr.type);
2492         int middle_len = le32_to_cpu(msg->hdr.middle_len);
2493
2494         dout("alloc_middle %p type %d %s middle_len %d\n", msg, type,
2495              ceph_msg_type_name(type), middle_len);
2496         BUG_ON(!middle_len);
2497         BUG_ON(msg->middle);
2498
2499         msg->middle = ceph_buffer_new(middle_len, GFP_NOFS);
2500         if (!msg->middle)
2501                 return -ENOMEM;
2502         return 0;
2503 }
2504
2505 /*
2506  * Generic message allocator, for incoming messages.
2507  */
2508 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
2509                                 struct ceph_msg_header *hdr,
2510                                 int *skip)
2511 {
2512         int type = le16_to_cpu(hdr->type);
2513         int front_len = le32_to_cpu(hdr->front_len);
2514         int middle_len = le32_to_cpu(hdr->middle_len);
2515         struct ceph_msg *msg = NULL;
2516         int ret;
2517
2518         if (con->ops->alloc_msg) {
2519                 mutex_unlock(&con->mutex);
2520                 msg = con->ops->alloc_msg(con, hdr, skip);
2521                 mutex_lock(&con->mutex);
2522                 if (!msg || *skip)
2523                         return NULL;
2524         }
2525         if (!msg) {
2526                 *skip = 0;
2527                 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
2528                 if (!msg) {
2529                         pr_err("unable to allocate msg type %d len %d\n",
2530                                type, front_len);
2531                         return NULL;
2532                 }
2533                 msg->page_alignment = le16_to_cpu(hdr->data_off);
2534         }
2535         memcpy(&msg->hdr, &con->in_hdr, sizeof(con->in_hdr));
2536
2537         if (middle_len && !msg->middle) {
2538                 ret = ceph_alloc_middle(con, msg);
2539                 if (ret < 0) {
2540                         ceph_msg_put(msg);
2541                         return NULL;
2542                 }
2543         }
2544
2545         return msg;
2546 }
2547
2548
2549 /*
2550  * Free a generically kmalloc'd message.
2551  */
2552 void ceph_msg_kfree(struct ceph_msg *m)
2553 {
2554         dout("msg_kfree %p\n", m);
2555         if (m->front_is_vmalloc)
2556                 vfree(m->front.iov_base);
2557         else
2558                 kfree(m->front.iov_base);
2559         kfree(m);
2560 }
2561
2562 /*
2563  * Drop a msg ref.  Destroy as needed.
2564  */
2565 void ceph_msg_last_put(struct kref *kref)
2566 {
2567         struct ceph_msg *m = container_of(kref, struct ceph_msg, kref);
2568
2569         dout("ceph_msg_put last one on %p\n", m);
2570         WARN_ON(!list_empty(&m->list_head));
2571
2572         /* drop middle, data, if any */
2573         if (m->middle) {
2574                 ceph_buffer_put(m->middle);
2575                 m->middle = NULL;
2576         }
2577         m->nr_pages = 0;
2578         m->pages = NULL;
2579
2580         if (m->pagelist) {
2581                 ceph_pagelist_release(m->pagelist);
2582                 kfree(m->pagelist);
2583                 m->pagelist = NULL;
2584         }
2585
2586         m->trail = NULL;
2587
2588         if (m->pool)
2589                 ceph_msgpool_put(m->pool, m);
2590         else
2591                 ceph_msg_kfree(m);
2592 }
2593 EXPORT_SYMBOL(ceph_msg_last_put);
2594
2595 void ceph_msg_dump(struct ceph_msg *msg)
2596 {
2597         pr_debug("msg_dump %p (front_max %d nr_pages %d)\n", msg,
2598                  msg->front_max, msg->nr_pages);
2599         print_hex_dump(KERN_DEBUG, "header: ",
2600                        DUMP_PREFIX_OFFSET, 16, 1,
2601                        &msg->hdr, sizeof(msg->hdr), true);
2602         print_hex_dump(KERN_DEBUG, " front: ",
2603                        DUMP_PREFIX_OFFSET, 16, 1,
2604                        msg->front.iov_base, msg->front.iov_len, true);
2605         if (msg->middle)
2606                 print_hex_dump(KERN_DEBUG, "middle: ",
2607                                DUMP_PREFIX_OFFSET, 16, 1,
2608                                msg->middle->vec.iov_base,
2609                                msg->middle->vec.iov_len, true);
2610         print_hex_dump(KERN_DEBUG, "footer: ",
2611                        DUMP_PREFIX_OFFSET, 16, 1,
2612                        &msg->footer, sizeof(msg->footer), true);
2613 }
2614 EXPORT_SYMBOL(ceph_msg_dump);