ceph: make page alignment explicit in osd interface
[linux-2.6.git] / net / ceph / messenger.c
1 #include <linux/ceph/ceph_debug.h>
2
3 #include <linux/crc32c.h>
4 #include <linux/ctype.h>
5 #include <linux/highmem.h>
6 #include <linux/inet.h>
7 #include <linux/kthread.h>
8 #include <linux/net.h>
9 #include <linux/slab.h>
10 #include <linux/socket.h>
11 #include <linux/string.h>
12 #include <linux/bio.h>
13 #include <linux/blkdev.h>
14 #include <net/tcp.h>
15
16 #include <linux/ceph/libceph.h>
17 #include <linux/ceph/messenger.h>
18 #include <linux/ceph/decode.h>
19 #include <linux/ceph/pagelist.h>
20
21 /*
22  * Ceph uses the messenger to exchange ceph_msg messages with other
23  * hosts in the system.  The messenger provides ordered and reliable
24  * delivery.  We tolerate TCP disconnects by reconnecting (with
25  * exponential backoff) in the case of a fault (disconnection, bad
26  * crc, protocol error).  Acks allow sent messages to be discarded by
27  * the sender.
28  */
29
30 /* static tag bytes (protocol control messages) */
31 static char tag_msg = CEPH_MSGR_TAG_MSG;
32 static char tag_ack = CEPH_MSGR_TAG_ACK;
33 static char tag_keepalive = CEPH_MSGR_TAG_KEEPALIVE;
34
35 #ifdef CONFIG_LOCKDEP
36 static struct lock_class_key socket_class;
37 #endif
38
39
40 static void queue_con(struct ceph_connection *con);
41 static void con_work(struct work_struct *);
42 static void ceph_fault(struct ceph_connection *con);
43
44 /*
45  * nicely render a sockaddr as a string.
46  */
47 #define MAX_ADDR_STR 20
48 #define MAX_ADDR_STR_LEN 60
49 static char addr_str[MAX_ADDR_STR][MAX_ADDR_STR_LEN];
50 static DEFINE_SPINLOCK(addr_str_lock);
51 static int last_addr_str;
52
53 const char *ceph_pr_addr(const struct sockaddr_storage *ss)
54 {
55         int i;
56         char *s;
57         struct sockaddr_in *in4 = (void *)ss;
58         struct sockaddr_in6 *in6 = (void *)ss;
59
60         spin_lock(&addr_str_lock);
61         i = last_addr_str++;
62         if (last_addr_str == MAX_ADDR_STR)
63                 last_addr_str = 0;
64         spin_unlock(&addr_str_lock);
65         s = addr_str[i];
66
67         switch (ss->ss_family) {
68         case AF_INET:
69                 snprintf(s, MAX_ADDR_STR_LEN, "%pI4:%u", &in4->sin_addr,
70                          (unsigned int)ntohs(in4->sin_port));
71                 break;
72
73         case AF_INET6:
74                 snprintf(s, MAX_ADDR_STR_LEN, "[%pI6c]:%u", &in6->sin6_addr,
75                          (unsigned int)ntohs(in6->sin6_port));
76                 break;
77
78         default:
79                 sprintf(s, "(unknown sockaddr family %d)", (int)ss->ss_family);
80         }
81
82         return s;
83 }
84 EXPORT_SYMBOL(ceph_pr_addr);
85
86 static void encode_my_addr(struct ceph_messenger *msgr)
87 {
88         memcpy(&msgr->my_enc_addr, &msgr->inst.addr, sizeof(msgr->my_enc_addr));
89         ceph_encode_addr(&msgr->my_enc_addr);
90 }
91
92 /*
93  * work queue for all reading and writing to/from the socket.
94  */
95 struct workqueue_struct *ceph_msgr_wq;
96
97 int ceph_msgr_init(void)
98 {
99         ceph_msgr_wq = create_workqueue("ceph-msgr");
100         if (IS_ERR(ceph_msgr_wq)) {
101                 int ret = PTR_ERR(ceph_msgr_wq);
102                 pr_err("msgr_init failed to create workqueue: %d\n", ret);
103                 ceph_msgr_wq = NULL;
104                 return ret;
105         }
106         return 0;
107 }
108 EXPORT_SYMBOL(ceph_msgr_init);
109
110 void ceph_msgr_exit(void)
111 {
112         destroy_workqueue(ceph_msgr_wq);
113 }
114 EXPORT_SYMBOL(ceph_msgr_exit);
115
116 void ceph_msgr_flush(void)
117 {
118         flush_workqueue(ceph_msgr_wq);
119 }
120 EXPORT_SYMBOL(ceph_msgr_flush);
121
122
123 /*
124  * socket callback functions
125  */
126
127 /* data available on socket, or listen socket received a connect */
128 static void ceph_data_ready(struct sock *sk, int count_unused)
129 {
130         struct ceph_connection *con =
131                 (struct ceph_connection *)sk->sk_user_data;
132         if (sk->sk_state != TCP_CLOSE_WAIT) {
133                 dout("ceph_data_ready on %p state = %lu, queueing work\n",
134                      con, con->state);
135                 queue_con(con);
136         }
137 }
138
139 /* socket has buffer space for writing */
140 static void ceph_write_space(struct sock *sk)
141 {
142         struct ceph_connection *con =
143                 (struct ceph_connection *)sk->sk_user_data;
144
145         /* only queue to workqueue if there is data we want to write. */
146         if (test_bit(WRITE_PENDING, &con->state)) {
147                 dout("ceph_write_space %p queueing write work\n", con);
148                 queue_con(con);
149         } else {
150                 dout("ceph_write_space %p nothing to write\n", con);
151         }
152
153         /* since we have our own write_space, clear the SOCK_NOSPACE flag */
154         clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
155 }
156
157 /* socket's state has changed */
158 static void ceph_state_change(struct sock *sk)
159 {
160         struct ceph_connection *con =
161                 (struct ceph_connection *)sk->sk_user_data;
162
163         dout("ceph_state_change %p state = %lu sk_state = %u\n",
164              con, con->state, sk->sk_state);
165
166         if (test_bit(CLOSED, &con->state))
167                 return;
168
169         switch (sk->sk_state) {
170         case TCP_CLOSE:
171                 dout("ceph_state_change TCP_CLOSE\n");
172         case TCP_CLOSE_WAIT:
173                 dout("ceph_state_change TCP_CLOSE_WAIT\n");
174                 if (test_and_set_bit(SOCK_CLOSED, &con->state) == 0) {
175                         if (test_bit(CONNECTING, &con->state))
176                                 con->error_msg = "connection failed";
177                         else
178                                 con->error_msg = "socket closed";
179                         queue_con(con);
180                 }
181                 break;
182         case TCP_ESTABLISHED:
183                 dout("ceph_state_change TCP_ESTABLISHED\n");
184                 queue_con(con);
185                 break;
186         }
187 }
188
189 /*
190  * set up socket callbacks
191  */
192 static void set_sock_callbacks(struct socket *sock,
193                                struct ceph_connection *con)
194 {
195         struct sock *sk = sock->sk;
196         sk->sk_user_data = (void *)con;
197         sk->sk_data_ready = ceph_data_ready;
198         sk->sk_write_space = ceph_write_space;
199         sk->sk_state_change = ceph_state_change;
200 }
201
202
203 /*
204  * socket helpers
205  */
206
207 /*
208  * initiate connection to a remote socket.
209  */
210 static struct socket *ceph_tcp_connect(struct ceph_connection *con)
211 {
212         struct sockaddr_storage *paddr = &con->peer_addr.in_addr;
213         struct socket *sock;
214         int ret;
215
216         BUG_ON(con->sock);
217         ret = sock_create_kern(con->peer_addr.in_addr.ss_family, SOCK_STREAM,
218                                IPPROTO_TCP, &sock);
219         if (ret)
220                 return ERR_PTR(ret);
221         con->sock = sock;
222         sock->sk->sk_allocation = GFP_NOFS;
223
224 #ifdef CONFIG_LOCKDEP
225         lockdep_set_class(&sock->sk->sk_lock, &socket_class);
226 #endif
227
228         set_sock_callbacks(sock, con);
229
230         dout("connect %s\n", ceph_pr_addr(&con->peer_addr.in_addr));
231
232         ret = sock->ops->connect(sock, (struct sockaddr *)paddr, sizeof(*paddr),
233                                  O_NONBLOCK);
234         if (ret == -EINPROGRESS) {
235                 dout("connect %s EINPROGRESS sk_state = %u\n",
236                      ceph_pr_addr(&con->peer_addr.in_addr),
237                      sock->sk->sk_state);
238                 ret = 0;
239         }
240         if (ret < 0) {
241                 pr_err("connect %s error %d\n",
242                        ceph_pr_addr(&con->peer_addr.in_addr), ret);
243                 sock_release(sock);
244                 con->sock = NULL;
245                 con->error_msg = "connect error";
246         }
247
248         if (ret < 0)
249                 return ERR_PTR(ret);
250         return sock;
251 }
252
253 static int ceph_tcp_recvmsg(struct socket *sock, void *buf, size_t len)
254 {
255         struct kvec iov = {buf, len};
256         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
257
258         return kernel_recvmsg(sock, &msg, &iov, 1, len, msg.msg_flags);
259 }
260
261 /*
262  * write something.  @more is true if caller will be sending more data
263  * shortly.
264  */
265 static int ceph_tcp_sendmsg(struct socket *sock, struct kvec *iov,
266                      size_t kvlen, size_t len, int more)
267 {
268         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
269
270         if (more)
271                 msg.msg_flags |= MSG_MORE;
272         else
273                 msg.msg_flags |= MSG_EOR;  /* superfluous, but what the hell */
274
275         return kernel_sendmsg(sock, &msg, iov, kvlen, len);
276 }
277
278
279 /*
280  * Shutdown/close the socket for the given connection.
281  */
282 static int con_close_socket(struct ceph_connection *con)
283 {
284         int rc;
285
286         dout("con_close_socket on %p sock %p\n", con, con->sock);
287         if (!con->sock)
288                 return 0;
289         set_bit(SOCK_CLOSED, &con->state);
290         rc = con->sock->ops->shutdown(con->sock, SHUT_RDWR);
291         sock_release(con->sock);
292         con->sock = NULL;
293         clear_bit(SOCK_CLOSED, &con->state);
294         return rc;
295 }
296
297 /*
298  * Reset a connection.  Discard all incoming and outgoing messages
299  * and clear *_seq state.
300  */
301 static void ceph_msg_remove(struct ceph_msg *msg)
302 {
303         list_del_init(&msg->list_head);
304         ceph_msg_put(msg);
305 }
306 static void ceph_msg_remove_list(struct list_head *head)
307 {
308         while (!list_empty(head)) {
309                 struct ceph_msg *msg = list_first_entry(head, struct ceph_msg,
310                                                         list_head);
311                 ceph_msg_remove(msg);
312         }
313 }
314
315 static void reset_connection(struct ceph_connection *con)
316 {
317         /* reset connection, out_queue, msg_ and connect_seq */
318         /* discard existing out_queue and msg_seq */
319         ceph_msg_remove_list(&con->out_queue);
320         ceph_msg_remove_list(&con->out_sent);
321
322         if (con->in_msg) {
323                 ceph_msg_put(con->in_msg);
324                 con->in_msg = NULL;
325         }
326
327         con->connect_seq = 0;
328         con->out_seq = 0;
329         if (con->out_msg) {
330                 ceph_msg_put(con->out_msg);
331                 con->out_msg = NULL;
332         }
333         con->out_keepalive_pending = false;
334         con->in_seq = 0;
335         con->in_seq_acked = 0;
336 }
337
338 /*
339  * mark a peer down.  drop any open connections.
340  */
341 void ceph_con_close(struct ceph_connection *con)
342 {
343         dout("con_close %p peer %s\n", con,
344              ceph_pr_addr(&con->peer_addr.in_addr));
345         set_bit(CLOSED, &con->state);  /* in case there's queued work */
346         clear_bit(STANDBY, &con->state);  /* avoid connect_seq bump */
347         clear_bit(LOSSYTX, &con->state);  /* so we retry next connect */
348         clear_bit(KEEPALIVE_PENDING, &con->state);
349         clear_bit(WRITE_PENDING, &con->state);
350         mutex_lock(&con->mutex);
351         reset_connection(con);
352         con->peer_global_seq = 0;
353         cancel_delayed_work(&con->work);
354         mutex_unlock(&con->mutex);
355         queue_con(con);
356 }
357 EXPORT_SYMBOL(ceph_con_close);
358
359 /*
360  * Reopen a closed connection, with a new peer address.
361  */
362 void ceph_con_open(struct ceph_connection *con, struct ceph_entity_addr *addr)
363 {
364         dout("con_open %p %s\n", con, ceph_pr_addr(&addr->in_addr));
365         set_bit(OPENING, &con->state);
366         clear_bit(CLOSED, &con->state);
367         memcpy(&con->peer_addr, addr, sizeof(*addr));
368         con->delay = 0;      /* reset backoff memory */
369         queue_con(con);
370 }
371 EXPORT_SYMBOL(ceph_con_open);
372
373 /*
374  * return true if this connection ever successfully opened
375  */
376 bool ceph_con_opened(struct ceph_connection *con)
377 {
378         return con->connect_seq > 0;
379 }
380
381 /*
382  * generic get/put
383  */
384 struct ceph_connection *ceph_con_get(struct ceph_connection *con)
385 {
386         dout("con_get %p nref = %d -> %d\n", con,
387              atomic_read(&con->nref), atomic_read(&con->nref) + 1);
388         if (atomic_inc_not_zero(&con->nref))
389                 return con;
390         return NULL;
391 }
392
393 void ceph_con_put(struct ceph_connection *con)
394 {
395         dout("con_put %p nref = %d -> %d\n", con,
396              atomic_read(&con->nref), atomic_read(&con->nref) - 1);
397         BUG_ON(atomic_read(&con->nref) == 0);
398         if (atomic_dec_and_test(&con->nref)) {
399                 BUG_ON(con->sock);
400                 kfree(con);
401         }
402 }
403
404 /*
405  * initialize a new connection.
406  */
407 void ceph_con_init(struct ceph_messenger *msgr, struct ceph_connection *con)
408 {
409         dout("con_init %p\n", con);
410         memset(con, 0, sizeof(*con));
411         atomic_set(&con->nref, 1);
412         con->msgr = msgr;
413         mutex_init(&con->mutex);
414         INIT_LIST_HEAD(&con->out_queue);
415         INIT_LIST_HEAD(&con->out_sent);
416         INIT_DELAYED_WORK(&con->work, con_work);
417 }
418 EXPORT_SYMBOL(ceph_con_init);
419
420
421 /*
422  * We maintain a global counter to order connection attempts.  Get
423  * a unique seq greater than @gt.
424  */
425 static u32 get_global_seq(struct ceph_messenger *msgr, u32 gt)
426 {
427         u32 ret;
428
429         spin_lock(&msgr->global_seq_lock);
430         if (msgr->global_seq < gt)
431                 msgr->global_seq = gt;
432         ret = ++msgr->global_seq;
433         spin_unlock(&msgr->global_seq_lock);
434         return ret;
435 }
436
437
438 /*
439  * Prepare footer for currently outgoing message, and finish things
440  * off.  Assumes out_kvec* are already valid.. we just add on to the end.
441  */
442 static void prepare_write_message_footer(struct ceph_connection *con, int v)
443 {
444         struct ceph_msg *m = con->out_msg;
445
446         dout("prepare_write_message_footer %p\n", con);
447         con->out_kvec_is_msg = true;
448         con->out_kvec[v].iov_base = &m->footer;
449         con->out_kvec[v].iov_len = sizeof(m->footer);
450         con->out_kvec_bytes += sizeof(m->footer);
451         con->out_kvec_left++;
452         con->out_more = m->more_to_follow;
453         con->out_msg_done = true;
454 }
455
456 /*
457  * Prepare headers for the next outgoing message.
458  */
459 static void prepare_write_message(struct ceph_connection *con)
460 {
461         struct ceph_msg *m;
462         int v = 0;
463
464         con->out_kvec_bytes = 0;
465         con->out_kvec_is_msg = true;
466         con->out_msg_done = false;
467
468         /* Sneak an ack in there first?  If we can get it into the same
469          * TCP packet that's a good thing. */
470         if (con->in_seq > con->in_seq_acked) {
471                 con->in_seq_acked = con->in_seq;
472                 con->out_kvec[v].iov_base = &tag_ack;
473                 con->out_kvec[v++].iov_len = 1;
474                 con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
475                 con->out_kvec[v].iov_base = &con->out_temp_ack;
476                 con->out_kvec[v++].iov_len = sizeof(con->out_temp_ack);
477                 con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
478         }
479
480         m = list_first_entry(&con->out_queue,
481                        struct ceph_msg, list_head);
482         con->out_msg = m;
483         if (test_bit(LOSSYTX, &con->state)) {
484                 list_del_init(&m->list_head);
485         } else {
486                 /* put message on sent list */
487                 ceph_msg_get(m);
488                 list_move_tail(&m->list_head, &con->out_sent);
489         }
490
491         /*
492          * only assign outgoing seq # if we haven't sent this message
493          * yet.  if it is requeued, resend with it's original seq.
494          */
495         if (m->needs_out_seq) {
496                 m->hdr.seq = cpu_to_le64(++con->out_seq);
497                 m->needs_out_seq = false;
498         }
499
500         dout("prepare_write_message %p seq %lld type %d len %d+%d+%d %d pgs\n",
501              m, con->out_seq, le16_to_cpu(m->hdr.type),
502              le32_to_cpu(m->hdr.front_len), le32_to_cpu(m->hdr.middle_len),
503              le32_to_cpu(m->hdr.data_len),
504              m->nr_pages);
505         BUG_ON(le32_to_cpu(m->hdr.front_len) != m->front.iov_len);
506
507         /* tag + hdr + front + middle */
508         con->out_kvec[v].iov_base = &tag_msg;
509         con->out_kvec[v++].iov_len = 1;
510         con->out_kvec[v].iov_base = &m->hdr;
511         con->out_kvec[v++].iov_len = sizeof(m->hdr);
512         con->out_kvec[v++] = m->front;
513         if (m->middle)
514                 con->out_kvec[v++] = m->middle->vec;
515         con->out_kvec_left = v;
516         con->out_kvec_bytes += 1 + sizeof(m->hdr) + m->front.iov_len +
517                 (m->middle ? m->middle->vec.iov_len : 0);
518         con->out_kvec_cur = con->out_kvec;
519
520         /* fill in crc (except data pages), footer */
521         con->out_msg->hdr.crc =
522                 cpu_to_le32(crc32c(0, (void *)&m->hdr,
523                                       sizeof(m->hdr) - sizeof(m->hdr.crc)));
524         con->out_msg->footer.flags = CEPH_MSG_FOOTER_COMPLETE;
525         con->out_msg->footer.front_crc =
526                 cpu_to_le32(crc32c(0, m->front.iov_base, m->front.iov_len));
527         if (m->middle)
528                 con->out_msg->footer.middle_crc =
529                         cpu_to_le32(crc32c(0, m->middle->vec.iov_base,
530                                            m->middle->vec.iov_len));
531         else
532                 con->out_msg->footer.middle_crc = 0;
533         con->out_msg->footer.data_crc = 0;
534         dout("prepare_write_message front_crc %u data_crc %u\n",
535              le32_to_cpu(con->out_msg->footer.front_crc),
536              le32_to_cpu(con->out_msg->footer.middle_crc));
537
538         /* is there a data payload? */
539         if (le32_to_cpu(m->hdr.data_len) > 0) {
540                 /* initialize page iterator */
541                 con->out_msg_pos.page = 0;
542                 if (m->pages)
543                         con->out_msg_pos.page_pos =
544                                 le16_to_cpu(m->hdr.data_off) & ~PAGE_MASK;
545                 else
546                         con->out_msg_pos.page_pos = 0;
547                 con->out_msg_pos.data_pos = 0;
548                 con->out_msg_pos.did_page_crc = 0;
549                 con->out_more = 1;  /* data + footer will follow */
550         } else {
551                 /* no, queue up footer too and be done */
552                 prepare_write_message_footer(con, v);
553         }
554
555         set_bit(WRITE_PENDING, &con->state);
556 }
557
558 /*
559  * Prepare an ack.
560  */
561 static void prepare_write_ack(struct ceph_connection *con)
562 {
563         dout("prepare_write_ack %p %llu -> %llu\n", con,
564              con->in_seq_acked, con->in_seq);
565         con->in_seq_acked = con->in_seq;
566
567         con->out_kvec[0].iov_base = &tag_ack;
568         con->out_kvec[0].iov_len = 1;
569         con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
570         con->out_kvec[1].iov_base = &con->out_temp_ack;
571         con->out_kvec[1].iov_len = sizeof(con->out_temp_ack);
572         con->out_kvec_left = 2;
573         con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
574         con->out_kvec_cur = con->out_kvec;
575         con->out_more = 1;  /* more will follow.. eventually.. */
576         set_bit(WRITE_PENDING, &con->state);
577 }
578
579 /*
580  * Prepare to write keepalive byte.
581  */
582 static void prepare_write_keepalive(struct ceph_connection *con)
583 {
584         dout("prepare_write_keepalive %p\n", con);
585         con->out_kvec[0].iov_base = &tag_keepalive;
586         con->out_kvec[0].iov_len = 1;
587         con->out_kvec_left = 1;
588         con->out_kvec_bytes = 1;
589         con->out_kvec_cur = con->out_kvec;
590         set_bit(WRITE_PENDING, &con->state);
591 }
592
593 /*
594  * Connection negotiation.
595  */
596
597 static void prepare_connect_authorizer(struct ceph_connection *con)
598 {
599         void *auth_buf;
600         int auth_len = 0;
601         int auth_protocol = 0;
602
603         mutex_unlock(&con->mutex);
604         if (con->ops->get_authorizer)
605                 con->ops->get_authorizer(con, &auth_buf, &auth_len,
606                                          &auth_protocol, &con->auth_reply_buf,
607                                          &con->auth_reply_buf_len,
608                                          con->auth_retry);
609         mutex_lock(&con->mutex);
610
611         con->out_connect.authorizer_protocol = cpu_to_le32(auth_protocol);
612         con->out_connect.authorizer_len = cpu_to_le32(auth_len);
613
614         con->out_kvec[con->out_kvec_left].iov_base = auth_buf;
615         con->out_kvec[con->out_kvec_left].iov_len = auth_len;
616         con->out_kvec_left++;
617         con->out_kvec_bytes += auth_len;
618 }
619
620 /*
621  * We connected to a peer and are saying hello.
622  */
623 static void prepare_write_banner(struct ceph_messenger *msgr,
624                                  struct ceph_connection *con)
625 {
626         int len = strlen(CEPH_BANNER);
627
628         con->out_kvec[0].iov_base = CEPH_BANNER;
629         con->out_kvec[0].iov_len = len;
630         con->out_kvec[1].iov_base = &msgr->my_enc_addr;
631         con->out_kvec[1].iov_len = sizeof(msgr->my_enc_addr);
632         con->out_kvec_left = 2;
633         con->out_kvec_bytes = len + sizeof(msgr->my_enc_addr);
634         con->out_kvec_cur = con->out_kvec;
635         con->out_more = 0;
636         set_bit(WRITE_PENDING, &con->state);
637 }
638
639 static void prepare_write_connect(struct ceph_messenger *msgr,
640                                   struct ceph_connection *con,
641                                   int after_banner)
642 {
643         unsigned global_seq = get_global_seq(con->msgr, 0);
644         int proto;
645
646         switch (con->peer_name.type) {
647         case CEPH_ENTITY_TYPE_MON:
648                 proto = CEPH_MONC_PROTOCOL;
649                 break;
650         case CEPH_ENTITY_TYPE_OSD:
651                 proto = CEPH_OSDC_PROTOCOL;
652                 break;
653         case CEPH_ENTITY_TYPE_MDS:
654                 proto = CEPH_MDSC_PROTOCOL;
655                 break;
656         default:
657                 BUG();
658         }
659
660         dout("prepare_write_connect %p cseq=%d gseq=%d proto=%d\n", con,
661              con->connect_seq, global_seq, proto);
662
663         con->out_connect.features = cpu_to_le64(msgr->supported_features);
664         con->out_connect.host_type = cpu_to_le32(CEPH_ENTITY_TYPE_CLIENT);
665         con->out_connect.connect_seq = cpu_to_le32(con->connect_seq);
666         con->out_connect.global_seq = cpu_to_le32(global_seq);
667         con->out_connect.protocol_version = cpu_to_le32(proto);
668         con->out_connect.flags = 0;
669
670         if (!after_banner) {
671                 con->out_kvec_left = 0;
672                 con->out_kvec_bytes = 0;
673         }
674         con->out_kvec[con->out_kvec_left].iov_base = &con->out_connect;
675         con->out_kvec[con->out_kvec_left].iov_len = sizeof(con->out_connect);
676         con->out_kvec_left++;
677         con->out_kvec_bytes += sizeof(con->out_connect);
678         con->out_kvec_cur = con->out_kvec;
679         con->out_more = 0;
680         set_bit(WRITE_PENDING, &con->state);
681
682         prepare_connect_authorizer(con);
683 }
684
685
686 /*
687  * write as much of pending kvecs to the socket as we can.
688  *  1 -> done
689  *  0 -> socket full, but more to do
690  * <0 -> error
691  */
692 static int write_partial_kvec(struct ceph_connection *con)
693 {
694         int ret;
695
696         dout("write_partial_kvec %p %d left\n", con, con->out_kvec_bytes);
697         while (con->out_kvec_bytes > 0) {
698                 ret = ceph_tcp_sendmsg(con->sock, con->out_kvec_cur,
699                                        con->out_kvec_left, con->out_kvec_bytes,
700                                        con->out_more);
701                 if (ret <= 0)
702                         goto out;
703                 con->out_kvec_bytes -= ret;
704                 if (con->out_kvec_bytes == 0)
705                         break;            /* done */
706                 while (ret > 0) {
707                         if (ret >= con->out_kvec_cur->iov_len) {
708                                 ret -= con->out_kvec_cur->iov_len;
709                                 con->out_kvec_cur++;
710                                 con->out_kvec_left--;
711                         } else {
712                                 con->out_kvec_cur->iov_len -= ret;
713                                 con->out_kvec_cur->iov_base += ret;
714                                 ret = 0;
715                                 break;
716                         }
717                 }
718         }
719         con->out_kvec_left = 0;
720         con->out_kvec_is_msg = false;
721         ret = 1;
722 out:
723         dout("write_partial_kvec %p %d left in %d kvecs ret = %d\n", con,
724              con->out_kvec_bytes, con->out_kvec_left, ret);
725         return ret;  /* done! */
726 }
727
728 #ifdef CONFIG_BLOCK
729 static void init_bio_iter(struct bio *bio, struct bio **iter, int *seg)
730 {
731         if (!bio) {
732                 *iter = NULL;
733                 *seg = 0;
734                 return;
735         }
736         *iter = bio;
737         *seg = bio->bi_idx;
738 }
739
740 static void iter_bio_next(struct bio **bio_iter, int *seg)
741 {
742         if (*bio_iter == NULL)
743                 return;
744
745         BUG_ON(*seg >= (*bio_iter)->bi_vcnt);
746
747         (*seg)++;
748         if (*seg == (*bio_iter)->bi_vcnt)
749                 init_bio_iter((*bio_iter)->bi_next, bio_iter, seg);
750 }
751 #endif
752
753 /*
754  * Write as much message data payload as we can.  If we finish, queue
755  * up the footer.
756  *  1 -> done, footer is now queued in out_kvec[].
757  *  0 -> socket full, but more to do
758  * <0 -> error
759  */
760 static int write_partial_msg_pages(struct ceph_connection *con)
761 {
762         struct ceph_msg *msg = con->out_msg;
763         unsigned data_len = le32_to_cpu(msg->hdr.data_len);
764         size_t len;
765         int crc = con->msgr->nocrc;
766         int ret;
767         int total_max_write;
768         int in_trail = 0;
769         size_t trail_len = (msg->trail ? msg->trail->length : 0);
770
771         dout("write_partial_msg_pages %p msg %p page %d/%d offset %d\n",
772              con, con->out_msg, con->out_msg_pos.page, con->out_msg->nr_pages,
773              con->out_msg_pos.page_pos);
774
775 #ifdef CONFIG_BLOCK
776         if (msg->bio && !msg->bio_iter)
777                 init_bio_iter(msg->bio, &msg->bio_iter, &msg->bio_seg);
778 #endif
779
780         while (data_len > con->out_msg_pos.data_pos) {
781                 struct page *page = NULL;
782                 void *kaddr = NULL;
783                 int max_write = PAGE_SIZE;
784                 int page_shift = 0;
785
786                 total_max_write = data_len - trail_len -
787                         con->out_msg_pos.data_pos;
788
789                 /*
790                  * if we are calculating the data crc (the default), we need
791                  * to map the page.  if our pages[] has been revoked, use the
792                  * zero page.
793                  */
794
795                 /* have we reached the trail part of the data? */
796                 if (con->out_msg_pos.data_pos >= data_len - trail_len) {
797                         in_trail = 1;
798
799                         total_max_write = data_len - con->out_msg_pos.data_pos;
800
801                         page = list_first_entry(&msg->trail->head,
802                                                 struct page, lru);
803                         if (crc)
804                                 kaddr = kmap(page);
805                         max_write = PAGE_SIZE;
806                 } else if (msg->pages) {
807                         page = msg->pages[con->out_msg_pos.page];
808                         if (crc)
809                                 kaddr = kmap(page);
810                 } else if (msg->pagelist) {
811                         page = list_first_entry(&msg->pagelist->head,
812                                                 struct page, lru);
813                         if (crc)
814                                 kaddr = kmap(page);
815 #ifdef CONFIG_BLOCK
816                 } else if (msg->bio) {
817                         struct bio_vec *bv;
818
819                         bv = bio_iovec_idx(msg->bio_iter, msg->bio_seg);
820                         page = bv->bv_page;
821                         page_shift = bv->bv_offset;
822                         if (crc)
823                                 kaddr = kmap(page) + page_shift;
824                         max_write = bv->bv_len;
825 #endif
826                 } else {
827                         page = con->msgr->zero_page;
828                         if (crc)
829                                 kaddr = page_address(con->msgr->zero_page);
830                 }
831                 len = min_t(int, max_write - con->out_msg_pos.page_pos,
832                             total_max_write);
833
834                 if (crc && !con->out_msg_pos.did_page_crc) {
835                         void *base = kaddr + con->out_msg_pos.page_pos;
836                         u32 tmpcrc = le32_to_cpu(con->out_msg->footer.data_crc);
837
838                         BUG_ON(kaddr == NULL);
839                         con->out_msg->footer.data_crc =
840                                 cpu_to_le32(crc32c(tmpcrc, base, len));
841                         con->out_msg_pos.did_page_crc = 1;
842                 }
843                 ret = kernel_sendpage(con->sock, page,
844                                       con->out_msg_pos.page_pos + page_shift,
845                                       len,
846                                       MSG_DONTWAIT | MSG_NOSIGNAL |
847                                       MSG_MORE);
848
849                 if (crc &&
850                     (msg->pages || msg->pagelist || msg->bio || in_trail))
851                         kunmap(page);
852
853                 if (ret <= 0)
854                         goto out;
855
856                 con->out_msg_pos.data_pos += ret;
857                 con->out_msg_pos.page_pos += ret;
858                 if (ret == len) {
859                         con->out_msg_pos.page_pos = 0;
860                         con->out_msg_pos.page++;
861                         con->out_msg_pos.did_page_crc = 0;
862                         if (in_trail)
863                                 list_move_tail(&page->lru,
864                                                &msg->trail->head);
865                         else if (msg->pagelist)
866                                 list_move_tail(&page->lru,
867                                                &msg->pagelist->head);
868 #ifdef CONFIG_BLOCK
869                         else if (msg->bio)
870                                 iter_bio_next(&msg->bio_iter, &msg->bio_seg);
871 #endif
872                 }
873         }
874
875         dout("write_partial_msg_pages %p msg %p done\n", con, msg);
876
877         /* prepare and queue up footer, too */
878         if (!crc)
879                 con->out_msg->footer.flags |= CEPH_MSG_FOOTER_NOCRC;
880         con->out_kvec_bytes = 0;
881         con->out_kvec_left = 0;
882         con->out_kvec_cur = con->out_kvec;
883         prepare_write_message_footer(con, 0);
884         ret = 1;
885 out:
886         return ret;
887 }
888
889 /*
890  * write some zeros
891  */
892 static int write_partial_skip(struct ceph_connection *con)
893 {
894         int ret;
895
896         while (con->out_skip > 0) {
897                 struct kvec iov = {
898                         .iov_base = page_address(con->msgr->zero_page),
899                         .iov_len = min(con->out_skip, (int)PAGE_CACHE_SIZE)
900                 };
901
902                 ret = ceph_tcp_sendmsg(con->sock, &iov, 1, iov.iov_len, 1);
903                 if (ret <= 0)
904                         goto out;
905                 con->out_skip -= ret;
906         }
907         ret = 1;
908 out:
909         return ret;
910 }
911
912 /*
913  * Prepare to read connection handshake, or an ack.
914  */
915 static void prepare_read_banner(struct ceph_connection *con)
916 {
917         dout("prepare_read_banner %p\n", con);
918         con->in_base_pos = 0;
919 }
920
921 static void prepare_read_connect(struct ceph_connection *con)
922 {
923         dout("prepare_read_connect %p\n", con);
924         con->in_base_pos = 0;
925 }
926
927 static void prepare_read_ack(struct ceph_connection *con)
928 {
929         dout("prepare_read_ack %p\n", con);
930         con->in_base_pos = 0;
931 }
932
933 static void prepare_read_tag(struct ceph_connection *con)
934 {
935         dout("prepare_read_tag %p\n", con);
936         con->in_base_pos = 0;
937         con->in_tag = CEPH_MSGR_TAG_READY;
938 }
939
940 /*
941  * Prepare to read a message.
942  */
943 static int prepare_read_message(struct ceph_connection *con)
944 {
945         dout("prepare_read_message %p\n", con);
946         BUG_ON(con->in_msg != NULL);
947         con->in_base_pos = 0;
948         con->in_front_crc = con->in_middle_crc = con->in_data_crc = 0;
949         return 0;
950 }
951
952
953 static int read_partial(struct ceph_connection *con,
954                         int *to, int size, void *object)
955 {
956         *to += size;
957         while (con->in_base_pos < *to) {
958                 int left = *to - con->in_base_pos;
959                 int have = size - left;
960                 int ret = ceph_tcp_recvmsg(con->sock, object + have, left);
961                 if (ret <= 0)
962                         return ret;
963                 con->in_base_pos += ret;
964         }
965         return 1;
966 }
967
968
969 /*
970  * Read all or part of the connect-side handshake on a new connection
971  */
972 static int read_partial_banner(struct ceph_connection *con)
973 {
974         int ret, to = 0;
975
976         dout("read_partial_banner %p at %d\n", con, con->in_base_pos);
977
978         /* peer's banner */
979         ret = read_partial(con, &to, strlen(CEPH_BANNER), con->in_banner);
980         if (ret <= 0)
981                 goto out;
982         ret = read_partial(con, &to, sizeof(con->actual_peer_addr),
983                            &con->actual_peer_addr);
984         if (ret <= 0)
985                 goto out;
986         ret = read_partial(con, &to, sizeof(con->peer_addr_for_me),
987                            &con->peer_addr_for_me);
988         if (ret <= 0)
989                 goto out;
990 out:
991         return ret;
992 }
993
994 static int read_partial_connect(struct ceph_connection *con)
995 {
996         int ret, to = 0;
997
998         dout("read_partial_connect %p at %d\n", con, con->in_base_pos);
999
1000         ret = read_partial(con, &to, sizeof(con->in_reply), &con->in_reply);
1001         if (ret <= 0)
1002                 goto out;
1003         ret = read_partial(con, &to, le32_to_cpu(con->in_reply.authorizer_len),
1004                            con->auth_reply_buf);
1005         if (ret <= 0)
1006                 goto out;
1007
1008         dout("read_partial_connect %p tag %d, con_seq = %u, g_seq = %u\n",
1009              con, (int)con->in_reply.tag,
1010              le32_to_cpu(con->in_reply.connect_seq),
1011              le32_to_cpu(con->in_reply.global_seq));
1012 out:
1013         return ret;
1014
1015 }
1016
1017 /*
1018  * Verify the hello banner looks okay.
1019  */
1020 static int verify_hello(struct ceph_connection *con)
1021 {
1022         if (memcmp(con->in_banner, CEPH_BANNER, strlen(CEPH_BANNER))) {
1023                 pr_err("connect to %s got bad banner\n",
1024                        ceph_pr_addr(&con->peer_addr.in_addr));
1025                 con->error_msg = "protocol error, bad banner";
1026                 return -1;
1027         }
1028         return 0;
1029 }
1030
1031 static bool addr_is_blank(struct sockaddr_storage *ss)
1032 {
1033         switch (ss->ss_family) {
1034         case AF_INET:
1035                 return ((struct sockaddr_in *)ss)->sin_addr.s_addr == 0;
1036         case AF_INET6:
1037                 return
1038                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[0] == 0 &&
1039                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[1] == 0 &&
1040                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[2] == 0 &&
1041                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[3] == 0;
1042         }
1043         return false;
1044 }
1045
1046 static int addr_port(struct sockaddr_storage *ss)
1047 {
1048         switch (ss->ss_family) {
1049         case AF_INET:
1050                 return ntohs(((struct sockaddr_in *)ss)->sin_port);
1051         case AF_INET6:
1052                 return ntohs(((struct sockaddr_in6 *)ss)->sin6_port);
1053         }
1054         return 0;
1055 }
1056
1057 static void addr_set_port(struct sockaddr_storage *ss, int p)
1058 {
1059         switch (ss->ss_family) {
1060         case AF_INET:
1061                 ((struct sockaddr_in *)ss)->sin_port = htons(p);
1062         case AF_INET6:
1063                 ((struct sockaddr_in6 *)ss)->sin6_port = htons(p);
1064         }
1065 }
1066
1067 /*
1068  * Parse an ip[:port] list into an addr array.  Use the default
1069  * monitor port if a port isn't specified.
1070  */
1071 int ceph_parse_ips(const char *c, const char *end,
1072                    struct ceph_entity_addr *addr,
1073                    int max_count, int *count)
1074 {
1075         int i;
1076         const char *p = c;
1077
1078         dout("parse_ips on '%.*s'\n", (int)(end-c), c);
1079         for (i = 0; i < max_count; i++) {
1080                 const char *ipend;
1081                 struct sockaddr_storage *ss = &addr[i].in_addr;
1082                 struct sockaddr_in *in4 = (void *)ss;
1083                 struct sockaddr_in6 *in6 = (void *)ss;
1084                 int port;
1085                 char delim = ',';
1086
1087                 if (*p == '[') {
1088                         delim = ']';
1089                         p++;
1090                 }
1091
1092                 memset(ss, 0, sizeof(*ss));
1093                 if (in4_pton(p, end - p, (u8 *)&in4->sin_addr.s_addr,
1094                              delim, &ipend))
1095                         ss->ss_family = AF_INET;
1096                 else if (in6_pton(p, end - p, (u8 *)&in6->sin6_addr.s6_addr,
1097                                   delim, &ipend))
1098                         ss->ss_family = AF_INET6;
1099                 else
1100                         goto bad;
1101                 p = ipend;
1102
1103                 if (delim == ']') {
1104                         if (*p != ']') {
1105                                 dout("missing matching ']'\n");
1106                                 goto bad;
1107                         }
1108                         p++;
1109                 }
1110
1111                 /* port? */
1112                 if (p < end && *p == ':') {
1113                         port = 0;
1114                         p++;
1115                         while (p < end && *p >= '0' && *p <= '9') {
1116                                 port = (port * 10) + (*p - '0');
1117                                 p++;
1118                         }
1119                         if (port > 65535 || port == 0)
1120                                 goto bad;
1121                 } else {
1122                         port = CEPH_MON_PORT;
1123                 }
1124
1125                 addr_set_port(ss, port);
1126
1127                 dout("parse_ips got %s\n", ceph_pr_addr(ss));
1128
1129                 if (p == end)
1130                         break;
1131                 if (*p != ',')
1132                         goto bad;
1133                 p++;
1134         }
1135
1136         if (p != end)
1137                 goto bad;
1138
1139         if (count)
1140                 *count = i + 1;
1141         return 0;
1142
1143 bad:
1144         pr_err("parse_ips bad ip '%.*s'\n", (int)(end - c), c);
1145         return -EINVAL;
1146 }
1147 EXPORT_SYMBOL(ceph_parse_ips);
1148
1149 static int process_banner(struct ceph_connection *con)
1150 {
1151         dout("process_banner on %p\n", con);
1152
1153         if (verify_hello(con) < 0)
1154                 return -1;
1155
1156         ceph_decode_addr(&con->actual_peer_addr);
1157         ceph_decode_addr(&con->peer_addr_for_me);
1158
1159         /*
1160          * Make sure the other end is who we wanted.  note that the other
1161          * end may not yet know their ip address, so if it's 0.0.0.0, give
1162          * them the benefit of the doubt.
1163          */
1164         if (memcmp(&con->peer_addr, &con->actual_peer_addr,
1165                    sizeof(con->peer_addr)) != 0 &&
1166             !(addr_is_blank(&con->actual_peer_addr.in_addr) &&
1167               con->actual_peer_addr.nonce == con->peer_addr.nonce)) {
1168                 pr_warning("wrong peer, want %s/%d, got %s/%d\n",
1169                            ceph_pr_addr(&con->peer_addr.in_addr),
1170                            (int)le32_to_cpu(con->peer_addr.nonce),
1171                            ceph_pr_addr(&con->actual_peer_addr.in_addr),
1172                            (int)le32_to_cpu(con->actual_peer_addr.nonce));
1173                 con->error_msg = "wrong peer at address";
1174                 return -1;
1175         }
1176
1177         /*
1178          * did we learn our address?
1179          */
1180         if (addr_is_blank(&con->msgr->inst.addr.in_addr)) {
1181                 int port = addr_port(&con->msgr->inst.addr.in_addr);
1182
1183                 memcpy(&con->msgr->inst.addr.in_addr,
1184                        &con->peer_addr_for_me.in_addr,
1185                        sizeof(con->peer_addr_for_me.in_addr));
1186                 addr_set_port(&con->msgr->inst.addr.in_addr, port);
1187                 encode_my_addr(con->msgr);
1188                 dout("process_banner learned my addr is %s\n",
1189                      ceph_pr_addr(&con->msgr->inst.addr.in_addr));
1190         }
1191
1192         set_bit(NEGOTIATING, &con->state);
1193         prepare_read_connect(con);
1194         return 0;
1195 }
1196
1197 static void fail_protocol(struct ceph_connection *con)
1198 {
1199         reset_connection(con);
1200         set_bit(CLOSED, &con->state);  /* in case there's queued work */
1201
1202         mutex_unlock(&con->mutex);
1203         if (con->ops->bad_proto)
1204                 con->ops->bad_proto(con);
1205         mutex_lock(&con->mutex);
1206 }
1207
1208 static int process_connect(struct ceph_connection *con)
1209 {
1210         u64 sup_feat = con->msgr->supported_features;
1211         u64 req_feat = con->msgr->required_features;
1212         u64 server_feat = le64_to_cpu(con->in_reply.features);
1213
1214         dout("process_connect on %p tag %d\n", con, (int)con->in_tag);
1215
1216         switch (con->in_reply.tag) {
1217         case CEPH_MSGR_TAG_FEATURES:
1218                 pr_err("%s%lld %s feature set mismatch,"
1219                        " my %llx < server's %llx, missing %llx\n",
1220                        ENTITY_NAME(con->peer_name),
1221                        ceph_pr_addr(&con->peer_addr.in_addr),
1222                        sup_feat, server_feat, server_feat & ~sup_feat);
1223                 con->error_msg = "missing required protocol features";
1224                 fail_protocol(con);
1225                 return -1;
1226
1227         case CEPH_MSGR_TAG_BADPROTOVER:
1228                 pr_err("%s%lld %s protocol version mismatch,"
1229                        " my %d != server's %d\n",
1230                        ENTITY_NAME(con->peer_name),
1231                        ceph_pr_addr(&con->peer_addr.in_addr),
1232                        le32_to_cpu(con->out_connect.protocol_version),
1233                        le32_to_cpu(con->in_reply.protocol_version));
1234                 con->error_msg = "protocol version mismatch";
1235                 fail_protocol(con);
1236                 return -1;
1237
1238         case CEPH_MSGR_TAG_BADAUTHORIZER:
1239                 con->auth_retry++;
1240                 dout("process_connect %p got BADAUTHORIZER attempt %d\n", con,
1241                      con->auth_retry);
1242                 if (con->auth_retry == 2) {
1243                         con->error_msg = "connect authorization failure";
1244                         reset_connection(con);
1245                         set_bit(CLOSED, &con->state);
1246                         return -1;
1247                 }
1248                 con->auth_retry = 1;
1249                 prepare_write_connect(con->msgr, con, 0);
1250                 prepare_read_connect(con);
1251                 break;
1252
1253         case CEPH_MSGR_TAG_RESETSESSION:
1254                 /*
1255                  * If we connected with a large connect_seq but the peer
1256                  * has no record of a session with us (no connection, or
1257                  * connect_seq == 0), they will send RESETSESION to indicate
1258                  * that they must have reset their session, and may have
1259                  * dropped messages.
1260                  */
1261                 dout("process_connect got RESET peer seq %u\n",
1262                      le32_to_cpu(con->in_connect.connect_seq));
1263                 pr_err("%s%lld %s connection reset\n",
1264                        ENTITY_NAME(con->peer_name),
1265                        ceph_pr_addr(&con->peer_addr.in_addr));
1266                 reset_connection(con);
1267                 prepare_write_connect(con->msgr, con, 0);
1268                 prepare_read_connect(con);
1269
1270                 /* Tell ceph about it. */
1271                 mutex_unlock(&con->mutex);
1272                 pr_info("reset on %s%lld\n", ENTITY_NAME(con->peer_name));
1273                 if (con->ops->peer_reset)
1274                         con->ops->peer_reset(con);
1275                 mutex_lock(&con->mutex);
1276                 break;
1277
1278         case CEPH_MSGR_TAG_RETRY_SESSION:
1279                 /*
1280                  * If we sent a smaller connect_seq than the peer has, try
1281                  * again with a larger value.
1282                  */
1283                 dout("process_connect got RETRY my seq = %u, peer_seq = %u\n",
1284                      le32_to_cpu(con->out_connect.connect_seq),
1285                      le32_to_cpu(con->in_connect.connect_seq));
1286                 con->connect_seq = le32_to_cpu(con->in_connect.connect_seq);
1287                 prepare_write_connect(con->msgr, con, 0);
1288                 prepare_read_connect(con);
1289                 break;
1290
1291         case CEPH_MSGR_TAG_RETRY_GLOBAL:
1292                 /*
1293                  * If we sent a smaller global_seq than the peer has, try
1294                  * again with a larger value.
1295                  */
1296                 dout("process_connect got RETRY_GLOBAL my %u peer_gseq %u\n",
1297                      con->peer_global_seq,
1298                      le32_to_cpu(con->in_connect.global_seq));
1299                 get_global_seq(con->msgr,
1300                                le32_to_cpu(con->in_connect.global_seq));
1301                 prepare_write_connect(con->msgr, con, 0);
1302                 prepare_read_connect(con);
1303                 break;
1304
1305         case CEPH_MSGR_TAG_READY:
1306                 if (req_feat & ~server_feat) {
1307                         pr_err("%s%lld %s protocol feature mismatch,"
1308                                " my required %llx > server's %llx, need %llx\n",
1309                                ENTITY_NAME(con->peer_name),
1310                                ceph_pr_addr(&con->peer_addr.in_addr),
1311                                req_feat, server_feat, req_feat & ~server_feat);
1312                         con->error_msg = "missing required protocol features";
1313                         fail_protocol(con);
1314                         return -1;
1315                 }
1316                 clear_bit(CONNECTING, &con->state);
1317                 con->peer_global_seq = le32_to_cpu(con->in_reply.global_seq);
1318                 con->connect_seq++;
1319                 con->peer_features = server_feat;
1320                 dout("process_connect got READY gseq %d cseq %d (%d)\n",
1321                      con->peer_global_seq,
1322                      le32_to_cpu(con->in_reply.connect_seq),
1323                      con->connect_seq);
1324                 WARN_ON(con->connect_seq !=
1325                         le32_to_cpu(con->in_reply.connect_seq));
1326
1327                 if (con->in_reply.flags & CEPH_MSG_CONNECT_LOSSY)
1328                         set_bit(LOSSYTX, &con->state);
1329
1330                 prepare_read_tag(con);
1331                 break;
1332
1333         case CEPH_MSGR_TAG_WAIT:
1334                 /*
1335                  * If there is a connection race (we are opening
1336                  * connections to each other), one of us may just have
1337                  * to WAIT.  This shouldn't happen if we are the
1338                  * client.
1339                  */
1340                 pr_err("process_connect peer connecting WAIT\n");
1341
1342         default:
1343                 pr_err("connect protocol error, will retry\n");
1344                 con->error_msg = "protocol error, garbage tag during connect";
1345                 return -1;
1346         }
1347         return 0;
1348 }
1349
1350
1351 /*
1352  * read (part of) an ack
1353  */
1354 static int read_partial_ack(struct ceph_connection *con)
1355 {
1356         int to = 0;
1357
1358         return read_partial(con, &to, sizeof(con->in_temp_ack),
1359                             &con->in_temp_ack);
1360 }
1361
1362
1363 /*
1364  * We can finally discard anything that's been acked.
1365  */
1366 static void process_ack(struct ceph_connection *con)
1367 {
1368         struct ceph_msg *m;
1369         u64 ack = le64_to_cpu(con->in_temp_ack);
1370         u64 seq;
1371
1372         while (!list_empty(&con->out_sent)) {
1373                 m = list_first_entry(&con->out_sent, struct ceph_msg,
1374                                      list_head);
1375                 seq = le64_to_cpu(m->hdr.seq);
1376                 if (seq > ack)
1377                         break;
1378                 dout("got ack for seq %llu type %d at %p\n", seq,
1379                      le16_to_cpu(m->hdr.type), m);
1380                 ceph_msg_remove(m);
1381         }
1382         prepare_read_tag(con);
1383 }
1384
1385
1386
1387
1388 static int read_partial_message_section(struct ceph_connection *con,
1389                                         struct kvec *section,
1390                                         unsigned int sec_len, u32 *crc)
1391 {
1392         int ret, left;
1393
1394         BUG_ON(!section);
1395
1396         while (section->iov_len < sec_len) {
1397                 BUG_ON(section->iov_base == NULL);
1398                 left = sec_len - section->iov_len;
1399                 ret = ceph_tcp_recvmsg(con->sock, (char *)section->iov_base +
1400                                        section->iov_len, left);
1401                 if (ret <= 0)
1402                         return ret;
1403                 section->iov_len += ret;
1404                 if (section->iov_len == sec_len)
1405                         *crc = crc32c(0, section->iov_base,
1406                                       section->iov_len);
1407         }
1408
1409         return 1;
1410 }
1411
1412 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
1413                                 struct ceph_msg_header *hdr,
1414                                 int *skip);
1415
1416
1417 static int read_partial_message_pages(struct ceph_connection *con,
1418                                       struct page **pages,
1419                                       unsigned data_len, int datacrc)
1420 {
1421         void *p;
1422         int ret;
1423         int left;
1424
1425         left = min((int)(data_len - con->in_msg_pos.data_pos),
1426                    (int)(PAGE_SIZE - con->in_msg_pos.page_pos));
1427         /* (page) data */
1428         BUG_ON(pages == NULL);
1429         p = kmap(pages[con->in_msg_pos.page]);
1430         ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1431                                left);
1432         if (ret > 0 && datacrc)
1433                 con->in_data_crc =
1434                         crc32c(con->in_data_crc,
1435                                   p + con->in_msg_pos.page_pos, ret);
1436         kunmap(pages[con->in_msg_pos.page]);
1437         if (ret <= 0)
1438                 return ret;
1439         con->in_msg_pos.data_pos += ret;
1440         con->in_msg_pos.page_pos += ret;
1441         if (con->in_msg_pos.page_pos == PAGE_SIZE) {
1442                 con->in_msg_pos.page_pos = 0;
1443                 con->in_msg_pos.page++;
1444         }
1445
1446         return ret;
1447 }
1448
1449 #ifdef CONFIG_BLOCK
1450 static int read_partial_message_bio(struct ceph_connection *con,
1451                                     struct bio **bio_iter, int *bio_seg,
1452                                     unsigned data_len, int datacrc)
1453 {
1454         struct bio_vec *bv = bio_iovec_idx(*bio_iter, *bio_seg);
1455         void *p;
1456         int ret, left;
1457
1458         if (IS_ERR(bv))
1459                 return PTR_ERR(bv);
1460
1461         left = min((int)(data_len - con->in_msg_pos.data_pos),
1462                    (int)(bv->bv_len - con->in_msg_pos.page_pos));
1463
1464         p = kmap(bv->bv_page) + bv->bv_offset;
1465
1466         ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1467                                left);
1468         if (ret > 0 && datacrc)
1469                 con->in_data_crc =
1470                         crc32c(con->in_data_crc,
1471                                   p + con->in_msg_pos.page_pos, ret);
1472         kunmap(bv->bv_page);
1473         if (ret <= 0)
1474                 return ret;
1475         con->in_msg_pos.data_pos += ret;
1476         con->in_msg_pos.page_pos += ret;
1477         if (con->in_msg_pos.page_pos == bv->bv_len) {
1478                 con->in_msg_pos.page_pos = 0;
1479                 iter_bio_next(bio_iter, bio_seg);
1480         }
1481
1482         return ret;
1483 }
1484 #endif
1485
1486 /*
1487  * read (part of) a message.
1488  */
1489 static int read_partial_message(struct ceph_connection *con)
1490 {
1491         struct ceph_msg *m = con->in_msg;
1492         int ret;
1493         int to, left;
1494         unsigned front_len, middle_len, data_len, data_off;
1495         int datacrc = con->msgr->nocrc;
1496         int skip;
1497         u64 seq;
1498
1499         dout("read_partial_message con %p msg %p\n", con, m);
1500
1501         /* header */
1502         while (con->in_base_pos < sizeof(con->in_hdr)) {
1503                 left = sizeof(con->in_hdr) - con->in_base_pos;
1504                 ret = ceph_tcp_recvmsg(con->sock,
1505                                        (char *)&con->in_hdr + con->in_base_pos,
1506                                        left);
1507                 if (ret <= 0)
1508                         return ret;
1509                 con->in_base_pos += ret;
1510                 if (con->in_base_pos == sizeof(con->in_hdr)) {
1511                         u32 crc = crc32c(0, (void *)&con->in_hdr,
1512                                  sizeof(con->in_hdr) - sizeof(con->in_hdr.crc));
1513                         if (crc != le32_to_cpu(con->in_hdr.crc)) {
1514                                 pr_err("read_partial_message bad hdr "
1515                                        " crc %u != expected %u\n",
1516                                        crc, con->in_hdr.crc);
1517                                 return -EBADMSG;
1518                         }
1519                 }
1520         }
1521         front_len = le32_to_cpu(con->in_hdr.front_len);
1522         if (front_len > CEPH_MSG_MAX_FRONT_LEN)
1523                 return -EIO;
1524         middle_len = le32_to_cpu(con->in_hdr.middle_len);
1525         if (middle_len > CEPH_MSG_MAX_DATA_LEN)
1526                 return -EIO;
1527         data_len = le32_to_cpu(con->in_hdr.data_len);
1528         if (data_len > CEPH_MSG_MAX_DATA_LEN)
1529                 return -EIO;
1530         data_off = le16_to_cpu(con->in_hdr.data_off);
1531
1532         /* verify seq# */
1533         seq = le64_to_cpu(con->in_hdr.seq);
1534         if ((s64)seq - (s64)con->in_seq < 1) {
1535                 pr_info("skipping %s%lld %s seq %lld expected %lld\n",
1536                         ENTITY_NAME(con->peer_name),
1537                         ceph_pr_addr(&con->peer_addr.in_addr),
1538                         seq, con->in_seq + 1);
1539                 con->in_base_pos = -front_len - middle_len - data_len -
1540                         sizeof(m->footer);
1541                 con->in_tag = CEPH_MSGR_TAG_READY;
1542                 return 0;
1543         } else if ((s64)seq - (s64)con->in_seq > 1) {
1544                 pr_err("read_partial_message bad seq %lld expected %lld\n",
1545                        seq, con->in_seq + 1);
1546                 con->error_msg = "bad message sequence # for incoming message";
1547                 return -EBADMSG;
1548         }
1549
1550         /* allocate message? */
1551         if (!con->in_msg) {
1552                 dout("got hdr type %d front %d data %d\n", con->in_hdr.type,
1553                      con->in_hdr.front_len, con->in_hdr.data_len);
1554                 skip = 0;
1555                 con->in_msg = ceph_alloc_msg(con, &con->in_hdr, &skip);
1556                 if (skip) {
1557                         /* skip this message */
1558                         dout("alloc_msg said skip message\n");
1559                         BUG_ON(con->in_msg);
1560                         con->in_base_pos = -front_len - middle_len - data_len -
1561                                 sizeof(m->footer);
1562                         con->in_tag = CEPH_MSGR_TAG_READY;
1563                         con->in_seq++;
1564                         return 0;
1565                 }
1566                 if (!con->in_msg) {
1567                         con->error_msg =
1568                                 "error allocating memory for incoming message";
1569                         return -ENOMEM;
1570                 }
1571                 m = con->in_msg;
1572                 m->front.iov_len = 0;    /* haven't read it yet */
1573                 if (m->middle)
1574                         m->middle->vec.iov_len = 0;
1575
1576                 con->in_msg_pos.page = 0;
1577                 if (m->pages)
1578                         con->in_msg_pos.page_pos = data_off & ~PAGE_MASK;
1579                 else
1580                         con->in_msg_pos.page_pos = 0;
1581                 con->in_msg_pos.data_pos = 0;
1582         }
1583
1584         /* front */
1585         ret = read_partial_message_section(con, &m->front, front_len,
1586                                            &con->in_front_crc);
1587         if (ret <= 0)
1588                 return ret;
1589
1590         /* middle */
1591         if (m->middle) {
1592                 ret = read_partial_message_section(con, &m->middle->vec,
1593                                                    middle_len,
1594                                                    &con->in_middle_crc);
1595                 if (ret <= 0)
1596                         return ret;
1597         }
1598 #ifdef CONFIG_BLOCK
1599         if (m->bio && !m->bio_iter)
1600                 init_bio_iter(m->bio, &m->bio_iter, &m->bio_seg);
1601 #endif
1602
1603         /* (page) data */
1604         while (con->in_msg_pos.data_pos < data_len) {
1605                 if (m->pages) {
1606                         ret = read_partial_message_pages(con, m->pages,
1607                                                  data_len, datacrc);
1608                         if (ret <= 0)
1609                                 return ret;
1610 #ifdef CONFIG_BLOCK
1611                 } else if (m->bio) {
1612
1613                         ret = read_partial_message_bio(con,
1614                                                  &m->bio_iter, &m->bio_seg,
1615                                                  data_len, datacrc);
1616                         if (ret <= 0)
1617                                 return ret;
1618 #endif
1619                 } else {
1620                         BUG_ON(1);
1621                 }
1622         }
1623
1624         /* footer */
1625         to = sizeof(m->hdr) + sizeof(m->footer);
1626         while (con->in_base_pos < to) {
1627                 left = to - con->in_base_pos;
1628                 ret = ceph_tcp_recvmsg(con->sock, (char *)&m->footer +
1629                                        (con->in_base_pos - sizeof(m->hdr)),
1630                                        left);
1631                 if (ret <= 0)
1632                         return ret;
1633                 con->in_base_pos += ret;
1634         }
1635         dout("read_partial_message got msg %p %d (%u) + %d (%u) + %d (%u)\n",
1636              m, front_len, m->footer.front_crc, middle_len,
1637              m->footer.middle_crc, data_len, m->footer.data_crc);
1638
1639         /* crc ok? */
1640         if (con->in_front_crc != le32_to_cpu(m->footer.front_crc)) {
1641                 pr_err("read_partial_message %p front crc %u != exp. %u\n",
1642                        m, con->in_front_crc, m->footer.front_crc);
1643                 return -EBADMSG;
1644         }
1645         if (con->in_middle_crc != le32_to_cpu(m->footer.middle_crc)) {
1646                 pr_err("read_partial_message %p middle crc %u != exp %u\n",
1647                        m, con->in_middle_crc, m->footer.middle_crc);
1648                 return -EBADMSG;
1649         }
1650         if (datacrc &&
1651             (m->footer.flags & CEPH_MSG_FOOTER_NOCRC) == 0 &&
1652             con->in_data_crc != le32_to_cpu(m->footer.data_crc)) {
1653                 pr_err("read_partial_message %p data crc %u != exp. %u\n", m,
1654                        con->in_data_crc, le32_to_cpu(m->footer.data_crc));
1655                 return -EBADMSG;
1656         }
1657
1658         return 1; /* done! */
1659 }
1660
1661 /*
1662  * Process message.  This happens in the worker thread.  The callback should
1663  * be careful not to do anything that waits on other incoming messages or it
1664  * may deadlock.
1665  */
1666 static void process_message(struct ceph_connection *con)
1667 {
1668         struct ceph_msg *msg;
1669
1670         msg = con->in_msg;
1671         con->in_msg = NULL;
1672
1673         /* if first message, set peer_name */
1674         if (con->peer_name.type == 0)
1675                 con->peer_name = msg->hdr.src;
1676
1677         con->in_seq++;
1678         mutex_unlock(&con->mutex);
1679
1680         dout("===== %p %llu from %s%lld %d=%s len %d+%d (%u %u %u) =====\n",
1681              msg, le64_to_cpu(msg->hdr.seq),
1682              ENTITY_NAME(msg->hdr.src),
1683              le16_to_cpu(msg->hdr.type),
1684              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1685              le32_to_cpu(msg->hdr.front_len),
1686              le32_to_cpu(msg->hdr.data_len),
1687              con->in_front_crc, con->in_middle_crc, con->in_data_crc);
1688         con->ops->dispatch(con, msg);
1689
1690         mutex_lock(&con->mutex);
1691         prepare_read_tag(con);
1692 }
1693
1694
1695 /*
1696  * Write something to the socket.  Called in a worker thread when the
1697  * socket appears to be writeable and we have something ready to send.
1698  */
1699 static int try_write(struct ceph_connection *con)
1700 {
1701         struct ceph_messenger *msgr = con->msgr;
1702         int ret = 1;
1703
1704         dout("try_write start %p state %lu nref %d\n", con, con->state,
1705              atomic_read(&con->nref));
1706
1707 more:
1708         dout("try_write out_kvec_bytes %d\n", con->out_kvec_bytes);
1709
1710         /* open the socket first? */
1711         if (con->sock == NULL) {
1712                 /*
1713                  * if we were STANDBY and are reconnecting _this_
1714                  * connection, bump connect_seq now.  Always bump
1715                  * global_seq.
1716                  */
1717                 if (test_and_clear_bit(STANDBY, &con->state))
1718                         con->connect_seq++;
1719
1720                 prepare_write_banner(msgr, con);
1721                 prepare_write_connect(msgr, con, 1);
1722                 prepare_read_banner(con);
1723                 set_bit(CONNECTING, &con->state);
1724                 clear_bit(NEGOTIATING, &con->state);
1725
1726                 BUG_ON(con->in_msg);
1727                 con->in_tag = CEPH_MSGR_TAG_READY;
1728                 dout("try_write initiating connect on %p new state %lu\n",
1729                      con, con->state);
1730                 con->sock = ceph_tcp_connect(con);
1731                 if (IS_ERR(con->sock)) {
1732                         con->sock = NULL;
1733                         con->error_msg = "connect error";
1734                         ret = -1;
1735                         goto out;
1736                 }
1737         }
1738
1739 more_kvec:
1740         /* kvec data queued? */
1741         if (con->out_skip) {
1742                 ret = write_partial_skip(con);
1743                 if (ret <= 0)
1744                         goto done;
1745                 if (ret < 0) {
1746                         dout("try_write write_partial_skip err %d\n", ret);
1747                         goto done;
1748                 }
1749         }
1750         if (con->out_kvec_left) {
1751                 ret = write_partial_kvec(con);
1752                 if (ret <= 0)
1753                         goto done;
1754         }
1755
1756         /* msg pages? */
1757         if (con->out_msg) {
1758                 if (con->out_msg_done) {
1759                         ceph_msg_put(con->out_msg);
1760                         con->out_msg = NULL;   /* we're done with this one */
1761                         goto do_next;
1762                 }
1763
1764                 ret = write_partial_msg_pages(con);
1765                 if (ret == 1)
1766                         goto more_kvec;  /* we need to send the footer, too! */
1767                 if (ret == 0)
1768                         goto done;
1769                 if (ret < 0) {
1770                         dout("try_write write_partial_msg_pages err %d\n",
1771                              ret);
1772                         goto done;
1773                 }
1774         }
1775
1776 do_next:
1777         if (!test_bit(CONNECTING, &con->state)) {
1778                 /* is anything else pending? */
1779                 if (!list_empty(&con->out_queue)) {
1780                         prepare_write_message(con);
1781                         goto more;
1782                 }
1783                 if (con->in_seq > con->in_seq_acked) {
1784                         prepare_write_ack(con);
1785                         goto more;
1786                 }
1787                 if (test_and_clear_bit(KEEPALIVE_PENDING, &con->state)) {
1788                         prepare_write_keepalive(con);
1789                         goto more;
1790                 }
1791         }
1792
1793         /* Nothing to do! */
1794         clear_bit(WRITE_PENDING, &con->state);
1795         dout("try_write nothing else to write.\n");
1796 done:
1797         ret = 0;
1798 out:
1799         dout("try_write done on %p\n", con);
1800         return ret;
1801 }
1802
1803
1804
1805 /*
1806  * Read what we can from the socket.
1807  */
1808 static int try_read(struct ceph_connection *con)
1809 {
1810         int ret = -1;
1811
1812         if (!con->sock)
1813                 return 0;
1814
1815         if (test_bit(STANDBY, &con->state))
1816                 return 0;
1817
1818         dout("try_read start on %p\n", con);
1819
1820 more:
1821         dout("try_read tag %d in_base_pos %d\n", (int)con->in_tag,
1822              con->in_base_pos);
1823         if (test_bit(CONNECTING, &con->state)) {
1824                 if (!test_bit(NEGOTIATING, &con->state)) {
1825                         dout("try_read connecting\n");
1826                         ret = read_partial_banner(con);
1827                         if (ret <= 0)
1828                                 goto done;
1829                         if (process_banner(con) < 0) {
1830                                 ret = -1;
1831                                 goto out;
1832                         }
1833                 }
1834                 ret = read_partial_connect(con);
1835                 if (ret <= 0)
1836                         goto done;
1837                 if (process_connect(con) < 0) {
1838                         ret = -1;
1839                         goto out;
1840                 }
1841                 goto more;
1842         }
1843
1844         if (con->in_base_pos < 0) {
1845                 /*
1846                  * skipping + discarding content.
1847                  *
1848                  * FIXME: there must be a better way to do this!
1849                  */
1850                 static char buf[1024];
1851                 int skip = min(1024, -con->in_base_pos);
1852                 dout("skipping %d / %d bytes\n", skip, -con->in_base_pos);
1853                 ret = ceph_tcp_recvmsg(con->sock, buf, skip);
1854                 if (ret <= 0)
1855                         goto done;
1856                 con->in_base_pos += ret;
1857                 if (con->in_base_pos)
1858                         goto more;
1859         }
1860         if (con->in_tag == CEPH_MSGR_TAG_READY) {
1861                 /*
1862                  * what's next?
1863                  */
1864                 ret = ceph_tcp_recvmsg(con->sock, &con->in_tag, 1);
1865                 if (ret <= 0)
1866                         goto done;
1867                 dout("try_read got tag %d\n", (int)con->in_tag);
1868                 switch (con->in_tag) {
1869                 case CEPH_MSGR_TAG_MSG:
1870                         prepare_read_message(con);
1871                         break;
1872                 case CEPH_MSGR_TAG_ACK:
1873                         prepare_read_ack(con);
1874                         break;
1875                 case CEPH_MSGR_TAG_CLOSE:
1876                         set_bit(CLOSED, &con->state);   /* fixme */
1877                         goto done;
1878                 default:
1879                         goto bad_tag;
1880                 }
1881         }
1882         if (con->in_tag == CEPH_MSGR_TAG_MSG) {
1883                 ret = read_partial_message(con);
1884                 if (ret <= 0) {
1885                         switch (ret) {
1886                         case -EBADMSG:
1887                                 con->error_msg = "bad crc";
1888                                 ret = -EIO;
1889                                 goto out;
1890                         case -EIO:
1891                                 con->error_msg = "io error";
1892                                 goto out;
1893                         default:
1894                                 goto done;
1895                         }
1896                 }
1897                 if (con->in_tag == CEPH_MSGR_TAG_READY)
1898                         goto more;
1899                 process_message(con);
1900                 goto more;
1901         }
1902         if (con->in_tag == CEPH_MSGR_TAG_ACK) {
1903                 ret = read_partial_ack(con);
1904                 if (ret <= 0)
1905                         goto done;
1906                 process_ack(con);
1907                 goto more;
1908         }
1909
1910 done:
1911         ret = 0;
1912 out:
1913         dout("try_read done on %p\n", con);
1914         return ret;
1915
1916 bad_tag:
1917         pr_err("try_read bad con->in_tag = %d\n", (int)con->in_tag);
1918         con->error_msg = "protocol error, garbage tag";
1919         ret = -1;
1920         goto out;
1921 }
1922
1923
1924 /*
1925  * Atomically queue work on a connection.  Bump @con reference to
1926  * avoid races with connection teardown.
1927  *
1928  * There is some trickery going on with QUEUED and BUSY because we
1929  * only want a _single_ thread operating on each connection at any
1930  * point in time, but we want to use all available CPUs.
1931  *
1932  * The worker thread only proceeds if it can atomically set BUSY.  It
1933  * clears QUEUED and does it's thing.  When it thinks it's done, it
1934  * clears BUSY, then rechecks QUEUED.. if it's set again, it loops
1935  * (tries again to set BUSY).
1936  *
1937  * To queue work, we first set QUEUED, _then_ if BUSY isn't set, we
1938  * try to queue work.  If that fails (work is already queued, or BUSY)
1939  * we give up (work also already being done or is queued) but leave QUEUED
1940  * set so that the worker thread will loop if necessary.
1941  */
1942 static void queue_con(struct ceph_connection *con)
1943 {
1944         if (test_bit(DEAD, &con->state)) {
1945                 dout("queue_con %p ignoring: DEAD\n",
1946                      con);
1947                 return;
1948         }
1949
1950         if (!con->ops->get(con)) {
1951                 dout("queue_con %p ref count 0\n", con);
1952                 return;
1953         }
1954
1955         set_bit(QUEUED, &con->state);
1956         if (test_bit(BUSY, &con->state)) {
1957                 dout("queue_con %p - already BUSY\n", con);
1958                 con->ops->put(con);
1959         } else if (!queue_work(ceph_msgr_wq, &con->work.work)) {
1960                 dout("queue_con %p - already queued\n", con);
1961                 con->ops->put(con);
1962         } else {
1963                 dout("queue_con %p\n", con);
1964         }
1965 }
1966
1967 /*
1968  * Do some work on a connection.  Drop a connection ref when we're done.
1969  */
1970 static void con_work(struct work_struct *work)
1971 {
1972         struct ceph_connection *con = container_of(work, struct ceph_connection,
1973                                                    work.work);
1974         int backoff = 0;
1975
1976 more:
1977         if (test_and_set_bit(BUSY, &con->state) != 0) {
1978                 dout("con_work %p BUSY already set\n", con);
1979                 goto out;
1980         }
1981         dout("con_work %p start, clearing QUEUED\n", con);
1982         clear_bit(QUEUED, &con->state);
1983
1984         mutex_lock(&con->mutex);
1985
1986         if (test_bit(CLOSED, &con->state)) { /* e.g. if we are replaced */
1987                 dout("con_work CLOSED\n");
1988                 con_close_socket(con);
1989                 goto done;
1990         }
1991         if (test_and_clear_bit(OPENING, &con->state)) {
1992                 /* reopen w/ new peer */
1993                 dout("con_work OPENING\n");
1994                 con_close_socket(con);
1995         }
1996
1997         if (test_and_clear_bit(SOCK_CLOSED, &con->state) ||
1998             try_read(con) < 0 ||
1999             try_write(con) < 0) {
2000                 mutex_unlock(&con->mutex);
2001                 backoff = 1;
2002                 ceph_fault(con);     /* error/fault path */
2003                 goto done_unlocked;
2004         }
2005
2006 done:
2007         mutex_unlock(&con->mutex);
2008
2009 done_unlocked:
2010         clear_bit(BUSY, &con->state);
2011         dout("con->state=%lu\n", con->state);
2012         if (test_bit(QUEUED, &con->state)) {
2013                 if (!backoff || test_bit(OPENING, &con->state)) {
2014                         dout("con_work %p QUEUED reset, looping\n", con);
2015                         goto more;
2016                 }
2017                 dout("con_work %p QUEUED reset, but just faulted\n", con);
2018                 clear_bit(QUEUED, &con->state);
2019         }
2020         dout("con_work %p done\n", con);
2021
2022 out:
2023         con->ops->put(con);
2024 }
2025
2026
2027 /*
2028  * Generic error/fault handler.  A retry mechanism is used with
2029  * exponential backoff
2030  */
2031 static void ceph_fault(struct ceph_connection *con)
2032 {
2033         pr_err("%s%lld %s %s\n", ENTITY_NAME(con->peer_name),
2034                ceph_pr_addr(&con->peer_addr.in_addr), con->error_msg);
2035         dout("fault %p state %lu to peer %s\n",
2036              con, con->state, ceph_pr_addr(&con->peer_addr.in_addr));
2037
2038         if (test_bit(LOSSYTX, &con->state)) {
2039                 dout("fault on LOSSYTX channel\n");
2040                 goto out;
2041         }
2042
2043         mutex_lock(&con->mutex);
2044         if (test_bit(CLOSED, &con->state))
2045                 goto out_unlock;
2046
2047         con_close_socket(con);
2048
2049         if (con->in_msg) {
2050                 ceph_msg_put(con->in_msg);
2051                 con->in_msg = NULL;
2052         }
2053
2054         /* Requeue anything that hasn't been acked */
2055         list_splice_init(&con->out_sent, &con->out_queue);
2056
2057         /* If there are no messages in the queue, place the connection
2058          * in a STANDBY state (i.e., don't try to reconnect just yet). */
2059         if (list_empty(&con->out_queue) && !con->out_keepalive_pending) {
2060                 dout("fault setting STANDBY\n");
2061                 set_bit(STANDBY, &con->state);
2062         } else {
2063                 /* retry after a delay. */
2064                 if (con->delay == 0)
2065                         con->delay = BASE_DELAY_INTERVAL;
2066                 else if (con->delay < MAX_DELAY_INTERVAL)
2067                         con->delay *= 2;
2068                 dout("fault queueing %p delay %lu\n", con, con->delay);
2069                 con->ops->get(con);
2070                 if (queue_delayed_work(ceph_msgr_wq, &con->work,
2071                                        round_jiffies_relative(con->delay)) == 0)
2072                         con->ops->put(con);
2073         }
2074
2075 out_unlock:
2076         mutex_unlock(&con->mutex);
2077 out:
2078         /*
2079          * in case we faulted due to authentication, invalidate our
2080          * current tickets so that we can get new ones.
2081          */
2082         if (con->auth_retry && con->ops->invalidate_authorizer) {
2083                 dout("calling invalidate_authorizer()\n");
2084                 con->ops->invalidate_authorizer(con);
2085         }
2086
2087         if (con->ops->fault)
2088                 con->ops->fault(con);
2089 }
2090
2091
2092
2093 /*
2094  * create a new messenger instance
2095  */
2096 struct ceph_messenger *ceph_messenger_create(struct ceph_entity_addr *myaddr,
2097                                              u32 supported_features,
2098                                              u32 required_features)
2099 {
2100         struct ceph_messenger *msgr;
2101
2102         msgr = kzalloc(sizeof(*msgr), GFP_KERNEL);
2103         if (msgr == NULL)
2104                 return ERR_PTR(-ENOMEM);
2105
2106         msgr->supported_features = supported_features;
2107         msgr->required_features = required_features;
2108
2109         spin_lock_init(&msgr->global_seq_lock);
2110
2111         /* the zero page is needed if a request is "canceled" while the message
2112          * is being written over the socket */
2113         msgr->zero_page = __page_cache_alloc(GFP_KERNEL | __GFP_ZERO);
2114         if (!msgr->zero_page) {
2115                 kfree(msgr);
2116                 return ERR_PTR(-ENOMEM);
2117         }
2118         kmap(msgr->zero_page);
2119
2120         if (myaddr)
2121                 msgr->inst.addr = *myaddr;
2122
2123         /* select a random nonce */
2124         msgr->inst.addr.type = 0;
2125         get_random_bytes(&msgr->inst.addr.nonce, sizeof(msgr->inst.addr.nonce));
2126         encode_my_addr(msgr);
2127
2128         dout("messenger_create %p\n", msgr);
2129         return msgr;
2130 }
2131 EXPORT_SYMBOL(ceph_messenger_create);
2132
2133 void ceph_messenger_destroy(struct ceph_messenger *msgr)
2134 {
2135         dout("destroy %p\n", msgr);
2136         kunmap(msgr->zero_page);
2137         __free_page(msgr->zero_page);
2138         kfree(msgr);
2139         dout("destroyed messenger %p\n", msgr);
2140 }
2141 EXPORT_SYMBOL(ceph_messenger_destroy);
2142
2143 /*
2144  * Queue up an outgoing message on the given connection.
2145  */
2146 void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg)
2147 {
2148         if (test_bit(CLOSED, &con->state)) {
2149                 dout("con_send %p closed, dropping %p\n", con, msg);
2150                 ceph_msg_put(msg);
2151                 return;
2152         }
2153
2154         /* set src+dst */
2155         msg->hdr.src = con->msgr->inst.name;
2156
2157         BUG_ON(msg->front.iov_len != le32_to_cpu(msg->hdr.front_len));
2158
2159         msg->needs_out_seq = true;
2160
2161         /* queue */
2162         mutex_lock(&con->mutex);
2163         BUG_ON(!list_empty(&msg->list_head));
2164         list_add_tail(&msg->list_head, &con->out_queue);
2165         dout("----- %p to %s%lld %d=%s len %d+%d+%d -----\n", msg,
2166              ENTITY_NAME(con->peer_name), le16_to_cpu(msg->hdr.type),
2167              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
2168              le32_to_cpu(msg->hdr.front_len),
2169              le32_to_cpu(msg->hdr.middle_len),
2170              le32_to_cpu(msg->hdr.data_len));
2171         mutex_unlock(&con->mutex);
2172
2173         /* if there wasn't anything waiting to send before, queue
2174          * new work */
2175         if (test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2176                 queue_con(con);
2177 }
2178 EXPORT_SYMBOL(ceph_con_send);
2179
2180 /*
2181  * Revoke a message that was previously queued for send
2182  */
2183 void ceph_con_revoke(struct ceph_connection *con, struct ceph_msg *msg)
2184 {
2185         mutex_lock(&con->mutex);
2186         if (!list_empty(&msg->list_head)) {
2187                 dout("con_revoke %p msg %p - was on queue\n", con, msg);
2188                 list_del_init(&msg->list_head);
2189                 ceph_msg_put(msg);
2190                 msg->hdr.seq = 0;
2191         }
2192         if (con->out_msg == msg) {
2193                 dout("con_revoke %p msg %p - was sending\n", con, msg);
2194                 con->out_msg = NULL;
2195                 if (con->out_kvec_is_msg) {
2196                         con->out_skip = con->out_kvec_bytes;
2197                         con->out_kvec_is_msg = false;
2198                 }
2199                 ceph_msg_put(msg);
2200                 msg->hdr.seq = 0;
2201         }
2202         mutex_unlock(&con->mutex);
2203 }
2204
2205 /*
2206  * Revoke a message that we may be reading data into
2207  */
2208 void ceph_con_revoke_message(struct ceph_connection *con, struct ceph_msg *msg)
2209 {
2210         mutex_lock(&con->mutex);
2211         if (con->in_msg && con->in_msg == msg) {
2212                 unsigned front_len = le32_to_cpu(con->in_hdr.front_len);
2213                 unsigned middle_len = le32_to_cpu(con->in_hdr.middle_len);
2214                 unsigned data_len = le32_to_cpu(con->in_hdr.data_len);
2215
2216                 /* skip rest of message */
2217                 dout("con_revoke_pages %p msg %p revoked\n", con, msg);
2218                         con->in_base_pos = con->in_base_pos -
2219                                 sizeof(struct ceph_msg_header) -
2220                                 front_len -
2221                                 middle_len -
2222                                 data_len -
2223                                 sizeof(struct ceph_msg_footer);
2224                 ceph_msg_put(con->in_msg);
2225                 con->in_msg = NULL;
2226                 con->in_tag = CEPH_MSGR_TAG_READY;
2227                 con->in_seq++;
2228         } else {
2229                 dout("con_revoke_pages %p msg %p pages %p no-op\n",
2230                      con, con->in_msg, msg);
2231         }
2232         mutex_unlock(&con->mutex);
2233 }
2234
2235 /*
2236  * Queue a keepalive byte to ensure the tcp connection is alive.
2237  */
2238 void ceph_con_keepalive(struct ceph_connection *con)
2239 {
2240         if (test_and_set_bit(KEEPALIVE_PENDING, &con->state) == 0 &&
2241             test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2242                 queue_con(con);
2243 }
2244 EXPORT_SYMBOL(ceph_con_keepalive);
2245
2246
2247 /*
2248  * construct a new message with given type, size
2249  * the new msg has a ref count of 1.
2250  */
2251 struct ceph_msg *ceph_msg_new(int type, int front_len, gfp_t flags)
2252 {
2253         struct ceph_msg *m;
2254
2255         m = kmalloc(sizeof(*m), flags);
2256         if (m == NULL)
2257                 goto out;
2258         kref_init(&m->kref);
2259         INIT_LIST_HEAD(&m->list_head);
2260
2261         m->hdr.tid = 0;
2262         m->hdr.type = cpu_to_le16(type);
2263         m->hdr.priority = cpu_to_le16(CEPH_MSG_PRIO_DEFAULT);
2264         m->hdr.version = 0;
2265         m->hdr.front_len = cpu_to_le32(front_len);
2266         m->hdr.middle_len = 0;
2267         m->hdr.data_len = 0;
2268         m->hdr.data_off = 0;
2269         m->hdr.reserved = 0;
2270         m->footer.front_crc = 0;
2271         m->footer.middle_crc = 0;
2272         m->footer.data_crc = 0;
2273         m->footer.flags = 0;
2274         m->front_max = front_len;
2275         m->front_is_vmalloc = false;
2276         m->more_to_follow = false;
2277         m->pool = NULL;
2278
2279         /* front */
2280         if (front_len) {
2281                 if (front_len > PAGE_CACHE_SIZE) {
2282                         m->front.iov_base = __vmalloc(front_len, flags,
2283                                                       PAGE_KERNEL);
2284                         m->front_is_vmalloc = true;
2285                 } else {
2286                         m->front.iov_base = kmalloc(front_len, flags);
2287                 }
2288                 if (m->front.iov_base == NULL) {
2289                         pr_err("msg_new can't allocate %d bytes\n",
2290                              front_len);
2291                         goto out2;
2292                 }
2293         } else {
2294                 m->front.iov_base = NULL;
2295         }
2296         m->front.iov_len = front_len;
2297
2298         /* middle */
2299         m->middle = NULL;
2300
2301         /* data */
2302         m->nr_pages = 0;
2303         m->pages = NULL;
2304         m->pagelist = NULL;
2305         m->bio = NULL;
2306         m->bio_iter = NULL;
2307         m->bio_seg = 0;
2308         m->trail = NULL;
2309
2310         dout("ceph_msg_new %p front %d\n", m, front_len);
2311         return m;
2312
2313 out2:
2314         ceph_msg_put(m);
2315 out:
2316         pr_err("msg_new can't create type %d front %d\n", type, front_len);
2317         return NULL;
2318 }
2319 EXPORT_SYMBOL(ceph_msg_new);
2320
2321 /*
2322  * Allocate "middle" portion of a message, if it is needed and wasn't
2323  * allocated by alloc_msg.  This allows us to read a small fixed-size
2324  * per-type header in the front and then gracefully fail (i.e.,
2325  * propagate the error to the caller based on info in the front) when
2326  * the middle is too large.
2327  */
2328 static int ceph_alloc_middle(struct ceph_connection *con, struct ceph_msg *msg)
2329 {
2330         int type = le16_to_cpu(msg->hdr.type);
2331         int middle_len = le32_to_cpu(msg->hdr.middle_len);
2332
2333         dout("alloc_middle %p type %d %s middle_len %d\n", msg, type,
2334              ceph_msg_type_name(type), middle_len);
2335         BUG_ON(!middle_len);
2336         BUG_ON(msg->middle);
2337
2338         msg->middle = ceph_buffer_new(middle_len, GFP_NOFS);
2339         if (!msg->middle)
2340                 return -ENOMEM;
2341         return 0;
2342 }
2343
2344 /*
2345  * Generic message allocator, for incoming messages.
2346  */
2347 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
2348                                 struct ceph_msg_header *hdr,
2349                                 int *skip)
2350 {
2351         int type = le16_to_cpu(hdr->type);
2352         int front_len = le32_to_cpu(hdr->front_len);
2353         int middle_len = le32_to_cpu(hdr->middle_len);
2354         struct ceph_msg *msg = NULL;
2355         int ret;
2356
2357         if (con->ops->alloc_msg) {
2358                 mutex_unlock(&con->mutex);
2359                 msg = con->ops->alloc_msg(con, hdr, skip);
2360                 mutex_lock(&con->mutex);
2361                 if (!msg || *skip)
2362                         return NULL;
2363         }
2364         if (!msg) {
2365                 *skip = 0;
2366                 msg = ceph_msg_new(type, front_len, GFP_NOFS);
2367                 if (!msg) {
2368                         pr_err("unable to allocate msg type %d len %d\n",
2369                                type, front_len);
2370                         return NULL;
2371                 }
2372         }
2373         memcpy(&msg->hdr, &con->in_hdr, sizeof(con->in_hdr));
2374
2375         if (middle_len && !msg->middle) {
2376                 ret = ceph_alloc_middle(con, msg);
2377                 if (ret < 0) {
2378                         ceph_msg_put(msg);
2379                         return NULL;
2380                 }
2381         }
2382
2383         return msg;
2384 }
2385
2386
2387 /*
2388  * Free a generically kmalloc'd message.
2389  */
2390 void ceph_msg_kfree(struct ceph_msg *m)
2391 {
2392         dout("msg_kfree %p\n", m);
2393         if (m->front_is_vmalloc)
2394                 vfree(m->front.iov_base);
2395         else
2396                 kfree(m->front.iov_base);
2397         kfree(m);
2398 }
2399
2400 /*
2401  * Drop a msg ref.  Destroy as needed.
2402  */
2403 void ceph_msg_last_put(struct kref *kref)
2404 {
2405         struct ceph_msg *m = container_of(kref, struct ceph_msg, kref);
2406
2407         dout("ceph_msg_put last one on %p\n", m);
2408         WARN_ON(!list_empty(&m->list_head));
2409
2410         /* drop middle, data, if any */
2411         if (m->middle) {
2412                 ceph_buffer_put(m->middle);
2413                 m->middle = NULL;
2414         }
2415         m->nr_pages = 0;
2416         m->pages = NULL;
2417
2418         if (m->pagelist) {
2419                 ceph_pagelist_release(m->pagelist);
2420                 kfree(m->pagelist);
2421                 m->pagelist = NULL;
2422         }
2423
2424         m->trail = NULL;
2425
2426         if (m->pool)
2427                 ceph_msgpool_put(m->pool, m);
2428         else
2429                 ceph_msg_kfree(m);
2430 }
2431 EXPORT_SYMBOL(ceph_msg_last_put);
2432
2433 void ceph_msg_dump(struct ceph_msg *msg)
2434 {
2435         pr_debug("msg_dump %p (front_max %d nr_pages %d)\n", msg,
2436                  msg->front_max, msg->nr_pages);
2437         print_hex_dump(KERN_DEBUG, "header: ",
2438                        DUMP_PREFIX_OFFSET, 16, 1,
2439                        &msg->hdr, sizeof(msg->hdr), true);
2440         print_hex_dump(KERN_DEBUG, " front: ",
2441                        DUMP_PREFIX_OFFSET, 16, 1,
2442                        msg->front.iov_base, msg->front.iov_len, true);
2443         if (msg->middle)
2444                 print_hex_dump(KERN_DEBUG, "middle: ",
2445                                DUMP_PREFIX_OFFSET, 16, 1,
2446                                msg->middle->vec.iov_base,
2447                                msg->middle->vec.iov_len, true);
2448         print_hex_dump(KERN_DEBUG, "footer: ",
2449                        DUMP_PREFIX_OFFSET, 16, 1,
2450                        &msg->footer, sizeof(msg->footer), true);
2451 }
2452 EXPORT_SYMBOL(ceph_msg_dump);