ceph: detect lossy state of connection
[safe/jmp/linux-2.6] / fs / ceph / messenger.c
1 #include "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/socket.h>
10 #include <linux/string.h>
11 #include <net/tcp.h>
12
13 #include "super.h"
14 #include "messenger.h"
15 #include "decode.h"
16
17 /*
18  * Ceph uses the messenger to exchange ceph_msg messages with other
19  * hosts in the system.  The messenger provides ordered and reliable
20  * delivery.  We tolerate TCP disconnects by reconnecting (with
21  * exponential backoff) in the case of a fault (disconnection, bad
22  * crc, protocol error).  Acks allow sent messages to be discarded by
23  * the sender.
24  */
25
26 /* static tag bytes (protocol control messages) */
27 static char tag_msg = CEPH_MSGR_TAG_MSG;
28 static char tag_ack = CEPH_MSGR_TAG_ACK;
29 static char tag_keepalive = CEPH_MSGR_TAG_KEEPALIVE;
30
31
32 static void queue_con(struct ceph_connection *con);
33 static void con_work(struct work_struct *);
34 static void ceph_fault(struct ceph_connection *con);
35
36 const char *ceph_name_type_str(int t)
37 {
38         switch (t) {
39         case CEPH_ENTITY_TYPE_MON: return "mon";
40         case CEPH_ENTITY_TYPE_MDS: return "mds";
41         case CEPH_ENTITY_TYPE_OSD: return "osd";
42         case CEPH_ENTITY_TYPE_CLIENT: return "client";
43         case CEPH_ENTITY_TYPE_ADMIN: return "admin";
44         default: return "???";
45         }
46 }
47
48 /*
49  * nicely render a sockaddr as a string.
50  */
51 #define MAX_ADDR_STR 20
52 static char addr_str[MAX_ADDR_STR][40];
53 static DEFINE_SPINLOCK(addr_str_lock);
54 static int last_addr_str;
55
56 const char *pr_addr(const struct sockaddr_storage *ss)
57 {
58         int i;
59         char *s;
60         struct sockaddr_in *in4 = (void *)ss;
61         unsigned char *quad = (void *)&in4->sin_addr.s_addr;
62         struct sockaddr_in6 *in6 = (void *)ss;
63
64         spin_lock(&addr_str_lock);
65         i = last_addr_str++;
66         if (last_addr_str == MAX_ADDR_STR)
67                 last_addr_str = 0;
68         spin_unlock(&addr_str_lock);
69         s = addr_str[i];
70
71         switch (ss->ss_family) {
72         case AF_INET:
73                 sprintf(s, "%u.%u.%u.%u:%u",
74                         (unsigned int)quad[0],
75                         (unsigned int)quad[1],
76                         (unsigned int)quad[2],
77                         (unsigned int)quad[3],
78                         (unsigned int)ntohs(in4->sin_port));
79                 break;
80
81         case AF_INET6:
82                 sprintf(s, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x:%u",
83                         in6->sin6_addr.s6_addr16[0],
84                         in6->sin6_addr.s6_addr16[1],
85                         in6->sin6_addr.s6_addr16[2],
86                         in6->sin6_addr.s6_addr16[3],
87                         in6->sin6_addr.s6_addr16[4],
88                         in6->sin6_addr.s6_addr16[5],
89                         in6->sin6_addr.s6_addr16[6],
90                         in6->sin6_addr.s6_addr16[7],
91                         (unsigned int)ntohs(in6->sin6_port));
92                 break;
93
94         default:
95                 sprintf(s, "(unknown sockaddr family %d)", (int)ss->ss_family);
96         }
97
98         return s;
99 }
100
101 static void encode_my_addr(struct ceph_messenger *msgr)
102 {
103         memcpy(&msgr->my_enc_addr, &msgr->inst.addr, sizeof(msgr->my_enc_addr));
104         ceph_encode_addr(&msgr->my_enc_addr);
105 }
106
107 /*
108  * work queue for all reading and writing to/from the socket.
109  */
110 struct workqueue_struct *ceph_msgr_wq;
111
112 int __init ceph_msgr_init(void)
113 {
114         ceph_msgr_wq = create_workqueue("ceph-msgr");
115         if (IS_ERR(ceph_msgr_wq)) {
116                 int ret = PTR_ERR(ceph_msgr_wq);
117                 pr_err("msgr_init failed to create workqueue: %d\n", ret);
118                 ceph_msgr_wq = NULL;
119                 return ret;
120         }
121         return 0;
122 }
123
124 void ceph_msgr_exit(void)
125 {
126         destroy_workqueue(ceph_msgr_wq);
127 }
128
129 /*
130  * socket callback functions
131  */
132
133 /* data available on socket, or listen socket received a connect */
134 static void ceph_data_ready(struct sock *sk, int count_unused)
135 {
136         struct ceph_connection *con =
137                 (struct ceph_connection *)sk->sk_user_data;
138         if (sk->sk_state != TCP_CLOSE_WAIT) {
139                 dout("ceph_data_ready on %p state = %lu, queueing work\n",
140                      con, con->state);
141                 queue_con(con);
142         }
143 }
144
145 /* socket has buffer space for writing */
146 static void ceph_write_space(struct sock *sk)
147 {
148         struct ceph_connection *con =
149                 (struct ceph_connection *)sk->sk_user_data;
150
151         /* only queue to workqueue if there is data we want to write. */
152         if (test_bit(WRITE_PENDING, &con->state)) {
153                 dout("ceph_write_space %p queueing write work\n", con);
154                 queue_con(con);
155         } else {
156                 dout("ceph_write_space %p nothing to write\n", con);
157         }
158
159         /* since we have our own write_space, clear the SOCK_NOSPACE flag */
160         clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
161 }
162
163 /* socket's state has changed */
164 static void ceph_state_change(struct sock *sk)
165 {
166         struct ceph_connection *con =
167                 (struct ceph_connection *)sk->sk_user_data;
168
169         dout("ceph_state_change %p state = %lu sk_state = %u\n",
170              con, con->state, sk->sk_state);
171
172         if (test_bit(CLOSED, &con->state))
173                 return;
174
175         switch (sk->sk_state) {
176         case TCP_CLOSE:
177                 dout("ceph_state_change TCP_CLOSE\n");
178         case TCP_CLOSE_WAIT:
179                 dout("ceph_state_change TCP_CLOSE_WAIT\n");
180                 if (test_and_set_bit(SOCK_CLOSED, &con->state) == 0) {
181                         if (test_bit(CONNECTING, &con->state))
182                                 con->error_msg = "connection failed";
183                         else
184                                 con->error_msg = "socket closed";
185                         queue_con(con);
186                 }
187                 break;
188         case TCP_ESTABLISHED:
189                 dout("ceph_state_change TCP_ESTABLISHED\n");
190                 queue_con(con);
191                 break;
192         }
193 }
194
195 /*
196  * set up socket callbacks
197  */
198 static void set_sock_callbacks(struct socket *sock,
199                                struct ceph_connection *con)
200 {
201         struct sock *sk = sock->sk;
202         sk->sk_user_data = (void *)con;
203         sk->sk_data_ready = ceph_data_ready;
204         sk->sk_write_space = ceph_write_space;
205         sk->sk_state_change = ceph_state_change;
206 }
207
208
209 /*
210  * socket helpers
211  */
212
213 /*
214  * initiate connection to a remote socket.
215  */
216 static struct socket *ceph_tcp_connect(struct ceph_connection *con)
217 {
218         struct sockaddr *paddr = (struct sockaddr *)&con->peer_addr.in_addr;
219         struct socket *sock;
220         int ret;
221
222         BUG_ON(con->sock);
223         ret = sock_create_kern(AF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
224         if (ret)
225                 return ERR_PTR(ret);
226         con->sock = sock;
227         sock->sk->sk_allocation = GFP_NOFS;
228
229         set_sock_callbacks(sock, con);
230
231         dout("connect %s\n", pr_addr(&con->peer_addr.in_addr));
232
233         ret = sock->ops->connect(sock, paddr, sizeof(*paddr), O_NONBLOCK);
234         if (ret == -EINPROGRESS) {
235                 dout("connect %s EINPROGRESS sk_state = %u\n",
236                      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                        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         mutex_lock(&con->out_mutex);
320         ceph_msg_remove_list(&con->out_queue);
321         ceph_msg_remove_list(&con->out_sent);
322
323         con->connect_seq = 0;
324         con->out_seq = 0;
325         if (con->out_msg) {
326                 ceph_msg_put(con->out_msg);
327                 con->out_msg = NULL;
328         }
329         con->in_seq = 0;
330         mutex_unlock(&con->out_mutex);
331 }
332
333 /*
334  * mark a peer down.  drop any open connections.
335  */
336 void ceph_con_close(struct ceph_connection *con)
337 {
338         dout("con_close %p peer %s\n", con, pr_addr(&con->peer_addr.in_addr));
339         set_bit(CLOSED, &con->state);  /* in case there's queued work */
340         clear_bit(STANDBY, &con->state);  /* avoid connect_seq bump */
341         reset_connection(con);
342         queue_con(con);
343 }
344
345 /*
346  * Reopen a closed connection, with a new peer address.
347  */
348 void ceph_con_open(struct ceph_connection *con, struct ceph_entity_addr *addr)
349 {
350         dout("con_open %p %s\n", con, pr_addr(&addr->in_addr));
351         set_bit(OPENING, &con->state);
352         clear_bit(CLOSED, &con->state);
353         memcpy(&con->peer_addr, addr, sizeof(*addr));
354         con->delay = 0;      /* reset backoff memory */
355         queue_con(con);
356 }
357
358 /*
359  * generic get/put
360  */
361 struct ceph_connection *ceph_con_get(struct ceph_connection *con)
362 {
363         dout("con_get %p nref = %d -> %d\n", con,
364              atomic_read(&con->nref), atomic_read(&con->nref) + 1);
365         if (atomic_inc_not_zero(&con->nref))
366                 return con;
367         return NULL;
368 }
369
370 void ceph_con_put(struct ceph_connection *con)
371 {
372         dout("con_put %p nref = %d -> %d\n", con,
373              atomic_read(&con->nref), atomic_read(&con->nref) - 1);
374         BUG_ON(atomic_read(&con->nref) == 0);
375         if (atomic_dec_and_test(&con->nref)) {
376                 BUG_ON(con->sock);
377                 kfree(con);
378         }
379 }
380
381 /*
382  * initialize a new connection.
383  */
384 void ceph_con_init(struct ceph_messenger *msgr, struct ceph_connection *con)
385 {
386         dout("con_init %p\n", con);
387         memset(con, 0, sizeof(*con));
388         atomic_set(&con->nref, 1);
389         con->msgr = msgr;
390         mutex_init(&con->out_mutex);
391         INIT_LIST_HEAD(&con->out_queue);
392         INIT_LIST_HEAD(&con->out_sent);
393         INIT_DELAYED_WORK(&con->work, con_work);
394 }
395
396
397 /*
398  * We maintain a global counter to order connection attempts.  Get
399  * a unique seq greater than @gt.
400  */
401 static u32 get_global_seq(struct ceph_messenger *msgr, u32 gt)
402 {
403         u32 ret;
404
405         spin_lock(&msgr->global_seq_lock);
406         if (msgr->global_seq < gt)
407                 msgr->global_seq = gt;
408         ret = ++msgr->global_seq;
409         spin_unlock(&msgr->global_seq_lock);
410         return ret;
411 }
412
413
414 /*
415  * Prepare footer for currently outgoing message, and finish things
416  * off.  Assumes out_kvec* are already valid.. we just add on to the end.
417  */
418 static void prepare_write_message_footer(struct ceph_connection *con, int v)
419 {
420         struct ceph_msg *m = con->out_msg;
421
422         dout("prepare_write_message_footer %p\n", con);
423         con->out_kvec_is_msg = true;
424         con->out_kvec[v].iov_base = &m->footer;
425         con->out_kvec[v].iov_len = sizeof(m->footer);
426         con->out_kvec_bytes += sizeof(m->footer);
427         con->out_kvec_left++;
428         con->out_more = m->more_to_follow;
429         con->out_msg_done = true;
430 }
431
432 /*
433  * Prepare headers for the next outgoing message.
434  */
435 static void prepare_write_message(struct ceph_connection *con)
436 {
437         struct ceph_msg *m;
438         int v = 0;
439
440         con->out_kvec_bytes = 0;
441         con->out_kvec_is_msg = true;
442         con->out_msg_done = false;
443
444         /* Sneak an ack in there first?  If we can get it into the same
445          * TCP packet that's a good thing. */
446         if (con->in_seq > con->in_seq_acked) {
447                 con->in_seq_acked = con->in_seq;
448                 con->out_kvec[v].iov_base = &tag_ack;
449                 con->out_kvec[v++].iov_len = 1;
450                 con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
451                 con->out_kvec[v].iov_base = &con->out_temp_ack;
452                 con->out_kvec[v++].iov_len = sizeof(con->out_temp_ack);
453                 con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
454         }
455
456         /* move message to sending/sent list */
457         m = list_first_entry(&con->out_queue,
458                        struct ceph_msg, list_head);
459         con->out_msg = m;
460         ceph_msg_get(m);
461         list_move_tail(&m->list_head, &con->out_sent);
462
463         m->hdr.seq = cpu_to_le64(++con->out_seq);
464
465         dout("prepare_write_message %p seq %lld type %d len %d+%d+%d %d pgs\n",
466              m, con->out_seq, le16_to_cpu(m->hdr.type),
467              le32_to_cpu(m->hdr.front_len), le32_to_cpu(m->hdr.middle_len),
468              le32_to_cpu(m->hdr.data_len),
469              m->nr_pages);
470         BUG_ON(le32_to_cpu(m->hdr.front_len) != m->front.iov_len);
471
472         /* tag + hdr + front + middle */
473         con->out_kvec[v].iov_base = &tag_msg;
474         con->out_kvec[v++].iov_len = 1;
475         con->out_kvec[v].iov_base = &m->hdr;
476         con->out_kvec[v++].iov_len = sizeof(m->hdr);
477         con->out_kvec[v++] = m->front;
478         if (m->middle)
479                 con->out_kvec[v++] = m->middle->vec;
480         con->out_kvec_left = v;
481         con->out_kvec_bytes += 1 + sizeof(m->hdr) + m->front.iov_len +
482                 (m->middle ? m->middle->vec.iov_len : 0);
483         con->out_kvec_cur = con->out_kvec;
484
485         /* fill in crc (except data pages), footer */
486         con->out_msg->hdr.crc =
487                 cpu_to_le32(crc32c(0, (void *)&m->hdr,
488                                       sizeof(m->hdr) - sizeof(m->hdr.crc)));
489         con->out_msg->footer.flags = CEPH_MSG_FOOTER_COMPLETE;
490         con->out_msg->footer.front_crc =
491                 cpu_to_le32(crc32c(0, m->front.iov_base, m->front.iov_len));
492         if (m->middle)
493                 con->out_msg->footer.middle_crc =
494                         cpu_to_le32(crc32c(0, m->middle->vec.iov_base,
495                                            m->middle->vec.iov_len));
496         else
497                 con->out_msg->footer.middle_crc = 0;
498         con->out_msg->footer.data_crc = 0;
499         dout("prepare_write_message front_crc %u data_crc %u\n",
500              le32_to_cpu(con->out_msg->footer.front_crc),
501              le32_to_cpu(con->out_msg->footer.middle_crc));
502
503         /* is there a data payload? */
504         if (le32_to_cpu(m->hdr.data_len) > 0) {
505                 /* initialize page iterator */
506                 con->out_msg_pos.page = 0;
507                 con->out_msg_pos.page_pos =
508                         le16_to_cpu(m->hdr.data_off) & ~PAGE_MASK;
509                 con->out_msg_pos.data_pos = 0;
510                 con->out_msg_pos.did_page_crc = 0;
511                 con->out_more = 1;  /* data + footer will follow */
512         } else {
513                 /* no, queue up footer too and be done */
514                 prepare_write_message_footer(con, v);
515         }
516
517         set_bit(WRITE_PENDING, &con->state);
518 }
519
520 /*
521  * Prepare an ack.
522  */
523 static void prepare_write_ack(struct ceph_connection *con)
524 {
525         dout("prepare_write_ack %p %llu -> %llu\n", con,
526              con->in_seq_acked, con->in_seq);
527         con->in_seq_acked = con->in_seq;
528
529         con->out_kvec[0].iov_base = &tag_ack;
530         con->out_kvec[0].iov_len = 1;
531         con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
532         con->out_kvec[1].iov_base = &con->out_temp_ack;
533         con->out_kvec[1].iov_len = sizeof(con->out_temp_ack);
534         con->out_kvec_left = 2;
535         con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
536         con->out_kvec_cur = con->out_kvec;
537         con->out_more = 1;  /* more will follow.. eventually.. */
538         set_bit(WRITE_PENDING, &con->state);
539 }
540
541 /*
542  * Prepare to write keepalive byte.
543  */
544 static void prepare_write_keepalive(struct ceph_connection *con)
545 {
546         dout("prepare_write_keepalive %p\n", con);
547         con->out_kvec[0].iov_base = &tag_keepalive;
548         con->out_kvec[0].iov_len = 1;
549         con->out_kvec_left = 1;
550         con->out_kvec_bytes = 1;
551         con->out_kvec_cur = con->out_kvec;
552         set_bit(WRITE_PENDING, &con->state);
553 }
554
555 /*
556  * Connection negotiation.
557  */
558
559 static void prepare_connect_authorizer(struct ceph_connection *con)
560 {
561         void *auth_buf;
562         int auth_len = 0;
563         int auth_protocol = 0;
564
565         if (con->ops->get_authorizer)
566                 con->ops->get_authorizer(con, &auth_buf, &auth_len,
567                                          &auth_protocol, &con->auth_reply_buf,
568                                          &con->auth_reply_buf_len,
569                                          con->auth_retry);
570
571         con->out_connect.authorizer_protocol = cpu_to_le32(auth_protocol);
572         con->out_connect.authorizer_len = cpu_to_le32(auth_len);
573
574         con->out_kvec[con->out_kvec_left].iov_base = auth_buf;
575         con->out_kvec[con->out_kvec_left].iov_len = auth_len;
576         con->out_kvec_left++;
577         con->out_kvec_bytes += auth_len;
578 }
579
580 /*
581  * We connected to a peer and are saying hello.
582  */
583 static void prepare_write_banner(struct ceph_messenger *msgr,
584                                  struct ceph_connection *con)
585 {
586         int len = strlen(CEPH_BANNER);
587
588         con->out_kvec[0].iov_base = CEPH_BANNER;
589         con->out_kvec[0].iov_len = len;
590         con->out_kvec[1].iov_base = &msgr->my_enc_addr;
591         con->out_kvec[1].iov_len = sizeof(msgr->my_enc_addr);
592         con->out_kvec_left = 2;
593         con->out_kvec_bytes = len + sizeof(msgr->my_enc_addr);
594         con->out_kvec_cur = con->out_kvec;
595         con->out_more = 0;
596         set_bit(WRITE_PENDING, &con->state);
597 }
598
599 static void prepare_write_connect(struct ceph_messenger *msgr,
600                                   struct ceph_connection *con,
601                                   int after_banner)
602 {
603         unsigned global_seq = get_global_seq(con->msgr, 0);
604         int proto;
605
606         switch (con->peer_name.type) {
607         case CEPH_ENTITY_TYPE_MON:
608                 proto = CEPH_MONC_PROTOCOL;
609                 break;
610         case CEPH_ENTITY_TYPE_OSD:
611                 proto = CEPH_OSDC_PROTOCOL;
612                 break;
613         case CEPH_ENTITY_TYPE_MDS:
614                 proto = CEPH_MDSC_PROTOCOL;
615                 break;
616         default:
617                 BUG();
618         }
619
620         dout("prepare_write_connect %p cseq=%d gseq=%d proto=%d\n", con,
621              con->connect_seq, global_seq, proto);
622
623         con->out_connect.host_type = cpu_to_le32(CEPH_ENTITY_TYPE_CLIENT);
624         con->out_connect.connect_seq = cpu_to_le32(con->connect_seq);
625         con->out_connect.global_seq = cpu_to_le32(global_seq);
626         con->out_connect.protocol_version = cpu_to_le32(proto);
627         con->out_connect.flags = 0;
628
629         if (!after_banner) {
630                 con->out_kvec_left = 0;
631                 con->out_kvec_bytes = 0;
632         }
633         con->out_kvec[con->out_kvec_left].iov_base = &con->out_connect;
634         con->out_kvec[con->out_kvec_left].iov_len = sizeof(con->out_connect);
635         con->out_kvec_left++;
636         con->out_kvec_bytes += sizeof(con->out_connect);
637         con->out_kvec_cur = con->out_kvec;
638         con->out_more = 0;
639         set_bit(WRITE_PENDING, &con->state);
640
641         prepare_connect_authorizer(con);
642 }
643
644
645 /*
646  * write as much of pending kvecs to the socket as we can.
647  *  1 -> done
648  *  0 -> socket full, but more to do
649  * <0 -> error
650  */
651 static int write_partial_kvec(struct ceph_connection *con)
652 {
653         int ret;
654
655         dout("write_partial_kvec %p %d left\n", con, con->out_kvec_bytes);
656         while (con->out_kvec_bytes > 0) {
657                 ret = ceph_tcp_sendmsg(con->sock, con->out_kvec_cur,
658                                        con->out_kvec_left, con->out_kvec_bytes,
659                                        con->out_more);
660                 if (ret <= 0)
661                         goto out;
662                 con->out_kvec_bytes -= ret;
663                 if (con->out_kvec_bytes == 0)
664                         break;            /* done */
665                 while (ret > 0) {
666                         if (ret >= con->out_kvec_cur->iov_len) {
667                                 ret -= con->out_kvec_cur->iov_len;
668                                 con->out_kvec_cur++;
669                                 con->out_kvec_left--;
670                         } else {
671                                 con->out_kvec_cur->iov_len -= ret;
672                                 con->out_kvec_cur->iov_base += ret;
673                                 ret = 0;
674                                 break;
675                         }
676                 }
677         }
678         con->out_kvec_left = 0;
679         con->out_kvec_is_msg = false;
680         ret = 1;
681 out:
682         dout("write_partial_kvec %p %d left in %d kvecs ret = %d\n", con,
683              con->out_kvec_bytes, con->out_kvec_left, ret);
684         return ret;  /* done! */
685 }
686
687 /*
688  * Write as much message data payload as we can.  If we finish, queue
689  * up the footer.
690  *  1 -> done, footer is now queued in out_kvec[].
691  *  0 -> socket full, but more to do
692  * <0 -> error
693  */
694 static int write_partial_msg_pages(struct ceph_connection *con)
695 {
696         struct ceph_msg *msg = con->out_msg;
697         unsigned data_len = le32_to_cpu(msg->hdr.data_len);
698         size_t len;
699         int crc = con->msgr->nocrc;
700         int ret;
701
702         dout("write_partial_msg_pages %p msg %p page %d/%d offset %d\n",
703              con, con->out_msg, con->out_msg_pos.page, con->out_msg->nr_pages,
704              con->out_msg_pos.page_pos);
705
706         while (con->out_msg_pos.page < con->out_msg->nr_pages) {
707                 struct page *page = NULL;
708                 void *kaddr = NULL;
709
710                 /*
711                  * if we are calculating the data crc (the default), we need
712                  * to map the page.  if our pages[] has been revoked, use the
713                  * zero page.
714                  */
715                 if (msg->pages) {
716                         page = msg->pages[con->out_msg_pos.page];
717                         if (crc)
718                                 kaddr = kmap(page);
719                 } else {
720                         page = con->msgr->zero_page;
721                         if (crc)
722                                 kaddr = page_address(con->msgr->zero_page);
723                 }
724                 len = min((int)(PAGE_SIZE - con->out_msg_pos.page_pos),
725                           (int)(data_len - con->out_msg_pos.data_pos));
726                 if (crc && !con->out_msg_pos.did_page_crc) {
727                         void *base = kaddr + con->out_msg_pos.page_pos;
728                         u32 tmpcrc = le32_to_cpu(con->out_msg->footer.data_crc);
729
730                         BUG_ON(kaddr == NULL);
731                         con->out_msg->footer.data_crc =
732                                 cpu_to_le32(crc32c(tmpcrc, base, len));
733                         con->out_msg_pos.did_page_crc = 1;
734                 }
735
736                 ret = kernel_sendpage(con->sock, page,
737                                       con->out_msg_pos.page_pos, len,
738                                       MSG_DONTWAIT | MSG_NOSIGNAL |
739                                       MSG_MORE);
740
741                 if (crc && msg->pages)
742                         kunmap(page);
743
744                 if (ret <= 0)
745                         goto out;
746
747                 con->out_msg_pos.data_pos += ret;
748                 con->out_msg_pos.page_pos += ret;
749                 if (ret == len) {
750                         con->out_msg_pos.page_pos = 0;
751                         con->out_msg_pos.page++;
752                         con->out_msg_pos.did_page_crc = 0;
753                 }
754         }
755
756         dout("write_partial_msg_pages %p msg %p done\n", con, msg);
757
758         /* prepare and queue up footer, too */
759         if (!crc)
760                 con->out_msg->footer.flags |= CEPH_MSG_FOOTER_NOCRC;
761         con->out_kvec_bytes = 0;
762         con->out_kvec_left = 0;
763         con->out_kvec_cur = con->out_kvec;
764         prepare_write_message_footer(con, 0);
765         ret = 1;
766 out:
767         return ret;
768 }
769
770 /*
771  * write some zeros
772  */
773 static int write_partial_skip(struct ceph_connection *con)
774 {
775         int ret;
776
777         while (con->out_skip > 0) {
778                 struct kvec iov = {
779                         .iov_base = page_address(con->msgr->zero_page),
780                         .iov_len = min(con->out_skip, (int)PAGE_CACHE_SIZE)
781                 };
782
783                 ret = ceph_tcp_sendmsg(con->sock, &iov, 1, iov.iov_len, 1);
784                 if (ret <= 0)
785                         goto out;
786                 con->out_skip -= ret;
787         }
788         ret = 1;
789 out:
790         return ret;
791 }
792
793 /*
794  * Prepare to read connection handshake, or an ack.
795  */
796 static void prepare_read_banner(struct ceph_connection *con)
797 {
798         dout("prepare_read_banner %p\n", con);
799         con->in_base_pos = 0;
800 }
801
802 static void prepare_read_connect(struct ceph_connection *con)
803 {
804         dout("prepare_read_connect %p\n", con);
805         con->in_base_pos = 0;
806 }
807
808 static void prepare_read_connect_retry(struct ceph_connection *con)
809 {
810         dout("prepare_read_connect_retry %p\n", con);
811         con->in_base_pos = strlen(CEPH_BANNER) + sizeof(con->actual_peer_addr)
812                 + sizeof(con->peer_addr_for_me);
813 }
814
815 static void prepare_read_ack(struct ceph_connection *con)
816 {
817         dout("prepare_read_ack %p\n", con);
818         con->in_base_pos = 0;
819 }
820
821 static void prepare_read_tag(struct ceph_connection *con)
822 {
823         dout("prepare_read_tag %p\n", con);
824         con->in_base_pos = 0;
825         con->in_tag = CEPH_MSGR_TAG_READY;
826 }
827
828 /*
829  * Prepare to read a message.
830  */
831 static int prepare_read_message(struct ceph_connection *con)
832 {
833         dout("prepare_read_message %p\n", con);
834         BUG_ON(con->in_msg != NULL);
835         con->in_base_pos = 0;
836         con->in_front_crc = con->in_middle_crc = con->in_data_crc = 0;
837         return 0;
838 }
839
840
841 static int read_partial(struct ceph_connection *con,
842                         int *to, int size, void *object)
843 {
844         *to += size;
845         while (con->in_base_pos < *to) {
846                 int left = *to - con->in_base_pos;
847                 int have = size - left;
848                 int ret = ceph_tcp_recvmsg(con->sock, object + have, left);
849                 if (ret <= 0)
850                         return ret;
851                 con->in_base_pos += ret;
852         }
853         return 1;
854 }
855
856
857 /*
858  * Read all or part of the connect-side handshake on a new connection
859  */
860 static int read_partial_banner(struct ceph_connection *con)
861 {
862         int ret, to = 0;
863
864         dout("read_partial_banner %p at %d\n", con, con->in_base_pos);
865
866         /* peer's banner */
867         ret = read_partial(con, &to, strlen(CEPH_BANNER), con->in_banner);
868         if (ret <= 0)
869                 goto out;
870         ret = read_partial(con, &to, sizeof(con->actual_peer_addr),
871                            &con->actual_peer_addr);
872         if (ret <= 0)
873                 goto out;
874         ret = read_partial(con, &to, sizeof(con->peer_addr_for_me),
875                            &con->peer_addr_for_me);
876         if (ret <= 0)
877                 goto out;
878 out:
879         return ret;
880 }
881
882 static int read_partial_connect(struct ceph_connection *con)
883 {
884         int ret, to = 0;
885
886         dout("read_partial_connect %p at %d\n", con, con->in_base_pos);
887
888         ret = read_partial(con, &to, sizeof(con->in_reply), &con->in_reply);
889         if (ret <= 0)
890                 goto out;
891         ret = read_partial(con, &to, le32_to_cpu(con->in_reply.authorizer_len),
892                            con->auth_reply_buf);
893         if (ret <= 0)
894                 goto out;
895
896         dout("read_partial_connect %p tag %d, con_seq = %u, g_seq = %u\n",
897              con, (int)con->in_reply.tag,
898              le32_to_cpu(con->in_reply.connect_seq),
899              le32_to_cpu(con->in_reply.global_seq));
900 out:
901         return ret;
902
903 }
904
905 /*
906  * Verify the hello banner looks okay.
907  */
908 static int verify_hello(struct ceph_connection *con)
909 {
910         if (memcmp(con->in_banner, CEPH_BANNER, strlen(CEPH_BANNER))) {
911                 pr_err("connect to %s got bad banner\n",
912                        pr_addr(&con->peer_addr.in_addr));
913                 con->error_msg = "protocol error, bad banner";
914                 return -1;
915         }
916         return 0;
917 }
918
919 static bool addr_is_blank(struct sockaddr_storage *ss)
920 {
921         switch (ss->ss_family) {
922         case AF_INET:
923                 return ((struct sockaddr_in *)ss)->sin_addr.s_addr == 0;
924         case AF_INET6:
925                 return
926                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[0] == 0 &&
927                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[1] == 0 &&
928                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[2] == 0 &&
929                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[3] == 0;
930         }
931         return false;
932 }
933
934 static int addr_port(struct sockaddr_storage *ss)
935 {
936         switch (ss->ss_family) {
937         case AF_INET:
938                 return ntohs(((struct sockaddr_in *)ss)->sin_port);
939         case AF_INET6:
940                 return ntohs(((struct sockaddr_in6 *)ss)->sin6_port);
941         }
942         return 0;
943 }
944
945 static void addr_set_port(struct sockaddr_storage *ss, int p)
946 {
947         switch (ss->ss_family) {
948         case AF_INET:
949                 ((struct sockaddr_in *)ss)->sin_port = htons(p);
950         case AF_INET6:
951                 ((struct sockaddr_in6 *)ss)->sin6_port = htons(p);
952         }
953 }
954
955 /*
956  * Parse an ip[:port] list into an addr array.  Use the default
957  * monitor port if a port isn't specified.
958  */
959 int ceph_parse_ips(const char *c, const char *end,
960                    struct ceph_entity_addr *addr,
961                    int max_count, int *count)
962 {
963         int i;
964         const char *p = c;
965
966         dout("parse_ips on '%.*s'\n", (int)(end-c), c);
967         for (i = 0; i < max_count; i++) {
968                 const char *ipend;
969                 struct sockaddr_storage *ss = &addr[i].in_addr;
970                 struct sockaddr_in *in4 = (void *)ss;
971                 struct sockaddr_in6 *in6 = (void *)ss;
972                 int port;
973
974                 memset(ss, 0, sizeof(*ss));
975                 if (in4_pton(p, end - p, (u8 *)&in4->sin_addr.s_addr,
976                              ',', &ipend)) {
977                         ss->ss_family = AF_INET;
978                 } else if (in6_pton(p, end - p, (u8 *)&in6->sin6_addr.s6_addr,
979                                     ',', &ipend)) {
980                         ss->ss_family = AF_INET6;
981                 } else {
982                         goto bad;
983                 }
984                 p = ipend;
985
986                 /* port? */
987                 if (p < end && *p == ':') {
988                         port = 0;
989                         p++;
990                         while (p < end && *p >= '0' && *p <= '9') {
991                                 port = (port * 10) + (*p - '0');
992                                 p++;
993                         }
994                         if (port > 65535 || port == 0)
995                                 goto bad;
996                 } else {
997                         port = CEPH_MON_PORT;
998                 }
999
1000                 addr_set_port(ss, port);
1001
1002                 dout("parse_ips got %s\n", pr_addr(ss));
1003
1004                 if (p == end)
1005                         break;
1006                 if (*p != ',')
1007                         goto bad;
1008                 p++;
1009         }
1010
1011         if (p != end)
1012                 goto bad;
1013
1014         if (count)
1015                 *count = i + 1;
1016         return 0;
1017
1018 bad:
1019         pr_err("parse_ips bad ip '%s'\n", c);
1020         return -EINVAL;
1021 }
1022
1023 static int process_banner(struct ceph_connection *con)
1024 {
1025         dout("process_banner on %p\n", con);
1026
1027         if (verify_hello(con) < 0)
1028                 return -1;
1029
1030         ceph_decode_addr(&con->actual_peer_addr);
1031         ceph_decode_addr(&con->peer_addr_for_me);
1032
1033         /*
1034          * Make sure the other end is who we wanted.  note that the other
1035          * end may not yet know their ip address, so if it's 0.0.0.0, give
1036          * them the benefit of the doubt.
1037          */
1038         if (!ceph_entity_addr_is_local(&con->peer_addr,
1039                                        &con->actual_peer_addr) &&
1040             !(addr_is_blank(&con->actual_peer_addr.in_addr) &&
1041               con->actual_peer_addr.nonce == con->peer_addr.nonce)) {
1042                 pr_err("wrong peer, want %s/%d, "
1043                        "got %s/%d, wtf\n",
1044                        pr_addr(&con->peer_addr.in_addr),
1045                        con->peer_addr.nonce,
1046                        pr_addr(&con->actual_peer_addr.in_addr),
1047                        con->actual_peer_addr.nonce);
1048                 con->error_msg = "protocol error, wrong peer";
1049                 return -1;
1050         }
1051
1052         /*
1053          * did we learn our address?
1054          */
1055         if (addr_is_blank(&con->msgr->inst.addr.in_addr)) {
1056                 int port = addr_port(&con->msgr->inst.addr.in_addr);
1057
1058                 memcpy(&con->msgr->inst.addr.in_addr,
1059                        &con->peer_addr_for_me.in_addr,
1060                        sizeof(con->peer_addr_for_me.in_addr));
1061                 addr_set_port(&con->msgr->inst.addr.in_addr, port);
1062                 encode_my_addr(con->msgr);
1063                 dout("process_banner learned my addr is %s\n",
1064                      pr_addr(&con->msgr->inst.addr.in_addr));
1065         }
1066
1067         set_bit(NEGOTIATING, &con->state);
1068         prepare_read_connect(con);
1069         return 0;
1070 }
1071
1072 static int process_connect(struct ceph_connection *con)
1073 {
1074         dout("process_connect on %p tag %d\n", con, (int)con->in_tag);
1075
1076         switch (con->in_reply.tag) {
1077         case CEPH_MSGR_TAG_BADPROTOVER:
1078                 dout("process_connect got BADPROTOVER my %d != their %d\n",
1079                      le32_to_cpu(con->out_connect.protocol_version),
1080                      le32_to_cpu(con->in_reply.protocol_version));
1081                 pr_err("%s%lld %s protocol version mismatch,"
1082                        " my %d != server's %d\n",
1083                        ENTITY_NAME(con->peer_name),
1084                        pr_addr(&con->peer_addr.in_addr),
1085                        le32_to_cpu(con->out_connect.protocol_version),
1086                        le32_to_cpu(con->in_reply.protocol_version));
1087                 con->error_msg = "protocol version mismatch";
1088                 if (con->ops->bad_proto)
1089                         con->ops->bad_proto(con);
1090                 reset_connection(con);
1091                 set_bit(CLOSED, &con->state);  /* in case there's queued work */
1092                 return -1;
1093
1094         case CEPH_MSGR_TAG_BADAUTHORIZER:
1095                 con->auth_retry++;
1096                 dout("process_connect %p got BADAUTHORIZER attempt %d\n", con,
1097                      con->auth_retry);
1098                 if (con->auth_retry == 2) {
1099                         con->error_msg = "connect authorization failure";
1100                         reset_connection(con);
1101                         set_bit(CLOSED, &con->state);
1102                         return -1;
1103                 }
1104                 con->auth_retry = 1;
1105                 prepare_write_connect(con->msgr, con, 0);
1106                 prepare_read_connect_retry(con);
1107                 break;
1108
1109         case CEPH_MSGR_TAG_RESETSESSION:
1110                 /*
1111                  * If we connected with a large connect_seq but the peer
1112                  * has no record of a session with us (no connection, or
1113                  * connect_seq == 0), they will send RESETSESION to indicate
1114                  * that they must have reset their session, and may have
1115                  * dropped messages.
1116                  */
1117                 dout("process_connect got RESET peer seq %u\n",
1118                      le32_to_cpu(con->in_connect.connect_seq));
1119                 pr_err("%s%lld %s connection reset\n",
1120                        ENTITY_NAME(con->peer_name),
1121                        pr_addr(&con->peer_addr.in_addr));
1122                 reset_connection(con);
1123                 prepare_write_connect(con->msgr, con, 0);
1124                 prepare_read_connect(con);
1125
1126                 /* Tell ceph about it. */
1127                 pr_info("reset on %s%lld\n", ENTITY_NAME(con->peer_name));
1128                 if (con->ops->peer_reset)
1129                         con->ops->peer_reset(con);
1130                 break;
1131
1132         case CEPH_MSGR_TAG_RETRY_SESSION:
1133                 /*
1134                  * If we sent a smaller connect_seq than the peer has, try
1135                  * again with a larger value.
1136                  */
1137                 dout("process_connect got RETRY my seq = %u, peer_seq = %u\n",
1138                      le32_to_cpu(con->out_connect.connect_seq),
1139                      le32_to_cpu(con->in_connect.connect_seq));
1140                 con->connect_seq = le32_to_cpu(con->in_connect.connect_seq);
1141                 prepare_write_connect(con->msgr, con, 0);
1142                 prepare_read_connect(con);
1143                 break;
1144
1145         case CEPH_MSGR_TAG_RETRY_GLOBAL:
1146                 /*
1147                  * If we sent a smaller global_seq than the peer has, try
1148                  * again with a larger value.
1149                  */
1150                 dout("process_connect got RETRY_GLOBAL my %u peer_gseq %u\n",
1151                      con->peer_global_seq,
1152                      le32_to_cpu(con->in_connect.global_seq));
1153                 get_global_seq(con->msgr,
1154                                le32_to_cpu(con->in_connect.global_seq));
1155                 prepare_write_connect(con->msgr, con, 0);
1156                 prepare_read_connect(con);
1157                 break;
1158
1159         case CEPH_MSGR_TAG_READY:
1160                 clear_bit(CONNECTING, &con->state);
1161                 con->peer_global_seq = le32_to_cpu(con->in_reply.global_seq);
1162                 con->connect_seq++;
1163                 dout("process_connect got READY gseq %d cseq %d (%d)\n",
1164                      con->peer_global_seq,
1165                      le32_to_cpu(con->in_reply.connect_seq),
1166                      con->connect_seq);
1167                 WARN_ON(con->connect_seq !=
1168                         le32_to_cpu(con->in_reply.connect_seq));
1169
1170                 if (con->in_reply.flags & CEPH_MSG_CONNECT_LOSSY)
1171                         set_bit(LOSSYTX, &con->state);
1172
1173                 prepare_read_tag(con);
1174                 break;
1175
1176         case CEPH_MSGR_TAG_WAIT:
1177                 /*
1178                  * If there is a connection race (we are opening
1179                  * connections to each other), one of us may just have
1180                  * to WAIT.  This shouldn't happen if we are the
1181                  * client.
1182                  */
1183                 pr_err("process_connect peer connecting WAIT\n");
1184
1185         default:
1186                 pr_err("connect protocol error, will retry\n");
1187                 con->error_msg = "protocol error, garbage tag during connect";
1188                 return -1;
1189         }
1190         return 0;
1191 }
1192
1193
1194 /*
1195  * read (part of) an ack
1196  */
1197 static int read_partial_ack(struct ceph_connection *con)
1198 {
1199         int to = 0;
1200
1201         return read_partial(con, &to, sizeof(con->in_temp_ack),
1202                             &con->in_temp_ack);
1203 }
1204
1205
1206 /*
1207  * We can finally discard anything that's been acked.
1208  */
1209 static void process_ack(struct ceph_connection *con)
1210 {
1211         struct ceph_msg *m;
1212         u64 ack = le64_to_cpu(con->in_temp_ack);
1213         u64 seq;
1214
1215         mutex_lock(&con->out_mutex);
1216         while (!list_empty(&con->out_sent)) {
1217                 m = list_first_entry(&con->out_sent, struct ceph_msg,
1218                                      list_head);
1219                 seq = le64_to_cpu(m->hdr.seq);
1220                 if (seq > ack)
1221                         break;
1222                 dout("got ack for seq %llu type %d at %p\n", seq,
1223                      le16_to_cpu(m->hdr.type), m);
1224                 ceph_msg_remove(m);
1225         }
1226         mutex_unlock(&con->out_mutex);
1227         prepare_read_tag(con);
1228 }
1229
1230
1231
1232
1233
1234
1235 /*
1236  * read (part of) a message.
1237  */
1238 static int read_partial_message(struct ceph_connection *con)
1239 {
1240         struct ceph_msg *m = con->in_msg;
1241         void *p;
1242         int ret;
1243         int to, want, left;
1244         unsigned front_len, middle_len, data_len, data_off;
1245         int datacrc = con->msgr->nocrc;
1246
1247         dout("read_partial_message con %p msg %p\n", con, m);
1248
1249         /* header */
1250         while (con->in_base_pos < sizeof(con->in_hdr)) {
1251                 left = sizeof(con->in_hdr) - con->in_base_pos;
1252                 ret = ceph_tcp_recvmsg(con->sock,
1253                                        (char *)&con->in_hdr + con->in_base_pos,
1254                                        left);
1255                 if (ret <= 0)
1256                         return ret;
1257                 con->in_base_pos += ret;
1258                 if (con->in_base_pos == sizeof(con->in_hdr)) {
1259                         u32 crc = crc32c(0, (void *)&con->in_hdr,
1260                                  sizeof(con->in_hdr) - sizeof(con->in_hdr.crc));
1261                         if (crc != le32_to_cpu(con->in_hdr.crc)) {
1262                                 pr_err("read_partial_message bad hdr "
1263                                        " crc %u != expected %u\n",
1264                                        crc, con->in_hdr.crc);
1265                                 return -EBADMSG;
1266                         }
1267                 }
1268         }
1269
1270         front_len = le32_to_cpu(con->in_hdr.front_len);
1271         if (front_len > CEPH_MSG_MAX_FRONT_LEN)
1272                 return -EIO;
1273         middle_len = le32_to_cpu(con->in_hdr.middle_len);
1274         if (middle_len > CEPH_MSG_MAX_DATA_LEN)
1275                 return -EIO;
1276         data_len = le32_to_cpu(con->in_hdr.data_len);
1277         if (data_len > CEPH_MSG_MAX_DATA_LEN)
1278                 return -EIO;
1279
1280         /* allocate message? */
1281         if (!con->in_msg) {
1282                 dout("got hdr type %d front %d data %d\n", con->in_hdr.type,
1283                      con->in_hdr.front_len, con->in_hdr.data_len);
1284                 con->in_msg = con->ops->alloc_msg(con, &con->in_hdr);
1285                 if (!con->in_msg) {
1286                         /* skip this message */
1287                         dout("alloc_msg returned NULL, skipping message\n");
1288                         con->in_base_pos = -front_len - middle_len - data_len -
1289                                 sizeof(m->footer);
1290                         con->in_tag = CEPH_MSGR_TAG_READY;
1291                         return 0;
1292                 }
1293                 if (IS_ERR(con->in_msg)) {
1294                         ret = PTR_ERR(con->in_msg);
1295                         con->in_msg = NULL;
1296                         con->error_msg = "out of memory for incoming message";
1297                         return ret;
1298                 }
1299                 m = con->in_msg;
1300                 m->front.iov_len = 0;    /* haven't read it yet */
1301                 memcpy(&m->hdr, &con->in_hdr, sizeof(con->in_hdr));
1302         }
1303
1304         /* front */
1305         while (m->front.iov_len < front_len) {
1306                 BUG_ON(m->front.iov_base == NULL);
1307                 left = front_len - m->front.iov_len;
1308                 ret = ceph_tcp_recvmsg(con->sock, (char *)m->front.iov_base +
1309                                        m->front.iov_len, left);
1310                 if (ret <= 0)
1311                         return ret;
1312                 m->front.iov_len += ret;
1313                 if (m->front.iov_len == front_len)
1314                         con->in_front_crc = crc32c(0, m->front.iov_base,
1315                                                       m->front.iov_len);
1316         }
1317
1318         /* middle */
1319         while (middle_len > 0 && (!m->middle ||
1320                                   m->middle->vec.iov_len < middle_len)) {
1321                 if (m->middle == NULL) {
1322                         ret = -EOPNOTSUPP;
1323                         if (con->ops->alloc_middle)
1324                                 ret = con->ops->alloc_middle(con, m);
1325                         if (ret < 0) {
1326                                 dout("alloc_middle failed, skipping payload\n");
1327                                 con->in_base_pos = -middle_len - data_len
1328                                         - sizeof(m->footer);
1329                                 ceph_msg_put(con->in_msg);
1330                                 con->in_msg = NULL;
1331                                 con->in_tag = CEPH_MSGR_TAG_READY;
1332                                 return 0;
1333                         }
1334                         m->middle->vec.iov_len = 0;
1335                 }
1336                 left = middle_len - m->middle->vec.iov_len;
1337                 ret = ceph_tcp_recvmsg(con->sock,
1338                                        (char *)m->middle->vec.iov_base +
1339                                        m->middle->vec.iov_len, left);
1340                 if (ret <= 0)
1341                         return ret;
1342                 m->middle->vec.iov_len += ret;
1343                 if (m->middle->vec.iov_len == middle_len)
1344                         con->in_middle_crc = crc32c(0, m->middle->vec.iov_base,
1345                                                       m->middle->vec.iov_len);
1346         }
1347
1348         /* (page) data */
1349         data_off = le16_to_cpu(m->hdr.data_off);
1350         if (data_len == 0)
1351                 goto no_data;
1352
1353         if (m->nr_pages == 0) {
1354                 con->in_msg_pos.page = 0;
1355                 con->in_msg_pos.page_pos = data_off & ~PAGE_MASK;
1356                 con->in_msg_pos.data_pos = 0;
1357                 /* find pages for data payload */
1358                 want = calc_pages_for(data_off & ~PAGE_MASK, data_len);
1359                 ret = -1;
1360                 if (con->ops->prepare_pages)
1361                         ret = con->ops->prepare_pages(con, m, want);
1362                 if (ret < 0) {
1363                         dout("%p prepare_pages failed, skipping payload\n", m);
1364                         con->in_base_pos = -data_len - sizeof(m->footer);
1365                         ceph_msg_put(con->in_msg);
1366                         con->in_msg = NULL;
1367                         con->in_tag = CEPH_MSGR_TAG_READY;
1368                         return 0;
1369                 }
1370                 BUG_ON(m->nr_pages < want);
1371         }
1372         while (con->in_msg_pos.data_pos < data_len) {
1373                 left = min((int)(data_len - con->in_msg_pos.data_pos),
1374                            (int)(PAGE_SIZE - con->in_msg_pos.page_pos));
1375                 BUG_ON(m->pages == NULL);
1376                 p = kmap(m->pages[con->in_msg_pos.page]);
1377                 ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1378                                        left);
1379                 if (ret > 0 && datacrc)
1380                         con->in_data_crc =
1381                                 crc32c(con->in_data_crc,
1382                                           p + con->in_msg_pos.page_pos, ret);
1383                 kunmap(m->pages[con->in_msg_pos.page]);
1384                 if (ret <= 0)
1385                         return ret;
1386                 con->in_msg_pos.data_pos += ret;
1387                 con->in_msg_pos.page_pos += ret;
1388                 if (con->in_msg_pos.page_pos == PAGE_SIZE) {
1389                         con->in_msg_pos.page_pos = 0;
1390                         con->in_msg_pos.page++;
1391                 }
1392         }
1393
1394 no_data:
1395         /* footer */
1396         to = sizeof(m->hdr) + sizeof(m->footer);
1397         while (con->in_base_pos < to) {
1398                 left = to - con->in_base_pos;
1399                 ret = ceph_tcp_recvmsg(con->sock, (char *)&m->footer +
1400                                        (con->in_base_pos - sizeof(m->hdr)),
1401                                        left);
1402                 if (ret <= 0)
1403                         return ret;
1404                 con->in_base_pos += ret;
1405         }
1406         dout("read_partial_message got msg %p %d (%u) + %d (%u) + %d (%u)\n",
1407              m, front_len, m->footer.front_crc, middle_len,
1408              m->footer.middle_crc, data_len, m->footer.data_crc);
1409
1410         /* crc ok? */
1411         if (con->in_front_crc != le32_to_cpu(m->footer.front_crc)) {
1412                 pr_err("read_partial_message %p front crc %u != exp. %u\n",
1413                        m, con->in_front_crc, m->footer.front_crc);
1414                 return -EBADMSG;
1415         }
1416         if (con->in_middle_crc != le32_to_cpu(m->footer.middle_crc)) {
1417                 pr_err("read_partial_message %p middle crc %u != exp %u\n",
1418                        m, con->in_middle_crc, m->footer.middle_crc);
1419                 return -EBADMSG;
1420         }
1421         if (datacrc &&
1422             (m->footer.flags & CEPH_MSG_FOOTER_NOCRC) == 0 &&
1423             con->in_data_crc != le32_to_cpu(m->footer.data_crc)) {
1424                 pr_err("read_partial_message %p data crc %u != exp. %u\n", m,
1425                        con->in_data_crc, le32_to_cpu(m->footer.data_crc));
1426                 return -EBADMSG;
1427         }
1428
1429         return 1; /* done! */
1430 }
1431
1432 /*
1433  * Process message.  This happens in the worker thread.  The callback should
1434  * be careful not to do anything that waits on other incoming messages or it
1435  * may deadlock.
1436  */
1437 static void process_message(struct ceph_connection *con)
1438 {
1439         struct ceph_msg *msg;
1440
1441         msg = con->in_msg;
1442         con->in_msg = NULL;
1443
1444         /* if first message, set peer_name */
1445         if (con->peer_name.type == 0)
1446                 con->peer_name = msg->hdr.src.name;
1447
1448         mutex_lock(&con->out_mutex);
1449         con->in_seq++;
1450         mutex_unlock(&con->out_mutex);
1451
1452         dout("===== %p %llu from %s%lld %d=%s len %d+%d (%u %u %u) =====\n",
1453              msg, le64_to_cpu(msg->hdr.seq),
1454              ENTITY_NAME(msg->hdr.src.name),
1455              le16_to_cpu(msg->hdr.type),
1456              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1457              le32_to_cpu(msg->hdr.front_len),
1458              le32_to_cpu(msg->hdr.data_len),
1459              con->in_front_crc, con->in_middle_crc, con->in_data_crc);
1460         con->ops->dispatch(con, msg);
1461         prepare_read_tag(con);
1462 }
1463
1464
1465 /*
1466  * Write something to the socket.  Called in a worker thread when the
1467  * socket appears to be writeable and we have something ready to send.
1468  */
1469 static int try_write(struct ceph_connection *con)
1470 {
1471         struct ceph_messenger *msgr = con->msgr;
1472         int ret = 1;
1473
1474         dout("try_write start %p state %lu nref %d\n", con, con->state,
1475              atomic_read(&con->nref));
1476
1477         mutex_lock(&con->out_mutex);
1478 more:
1479         dout("try_write out_kvec_bytes %d\n", con->out_kvec_bytes);
1480
1481         /* open the socket first? */
1482         if (con->sock == NULL) {
1483                 /*
1484                  * if we were STANDBY and are reconnecting _this_
1485                  * connection, bump connect_seq now.  Always bump
1486                  * global_seq.
1487                  */
1488                 if (test_and_clear_bit(STANDBY, &con->state))
1489                         con->connect_seq++;
1490
1491                 prepare_write_banner(msgr, con);
1492                 prepare_write_connect(msgr, con, 1);
1493                 prepare_read_banner(con);
1494                 set_bit(CONNECTING, &con->state);
1495                 clear_bit(NEGOTIATING, &con->state);
1496
1497                 con->in_tag = CEPH_MSGR_TAG_READY;
1498                 dout("try_write initiating connect on %p new state %lu\n",
1499                      con, con->state);
1500                 con->sock = ceph_tcp_connect(con);
1501                 if (IS_ERR(con->sock)) {
1502                         con->sock = NULL;
1503                         con->error_msg = "connect error";
1504                         ret = -1;
1505                         goto out;
1506                 }
1507         }
1508
1509 more_kvec:
1510         /* kvec data queued? */
1511         if (con->out_skip) {
1512                 ret = write_partial_skip(con);
1513                 if (ret <= 0)
1514                         goto done;
1515                 if (ret < 0) {
1516                         dout("try_write write_partial_skip err %d\n", ret);
1517                         goto done;
1518                 }
1519         }
1520         if (con->out_kvec_left) {
1521                 ret = write_partial_kvec(con);
1522                 if (ret <= 0)
1523                         goto done;
1524                 if (ret < 0) {
1525                         dout("try_write write_partial_kvec err %d\n", ret);
1526                         goto done;
1527                 }
1528         }
1529
1530         /* msg pages? */
1531         if (con->out_msg) {
1532                 if (con->out_msg_done) {
1533                         ceph_msg_put(con->out_msg);
1534                         con->out_msg = NULL;   /* we're done with this one */
1535                         goto do_next;
1536                 }
1537
1538                 ret = write_partial_msg_pages(con);
1539                 if (ret == 1)
1540                         goto more_kvec;  /* we need to send the footer, too! */
1541                 if (ret == 0)
1542                         goto done;
1543                 if (ret < 0) {
1544                         dout("try_write write_partial_msg_pages err %d\n",
1545                              ret);
1546                         goto done;
1547                 }
1548         }
1549
1550 do_next:
1551         if (!test_bit(CONNECTING, &con->state)) {
1552                 /* is anything else pending? */
1553                 if (!list_empty(&con->out_queue)) {
1554                         prepare_write_message(con);
1555                         goto more;
1556                 }
1557                 if (con->in_seq > con->in_seq_acked) {
1558                         prepare_write_ack(con);
1559                         goto more;
1560                 }
1561                 if (test_and_clear_bit(KEEPALIVE_PENDING, &con->state)) {
1562                         prepare_write_keepalive(con);
1563                         goto more;
1564                 }
1565         }
1566
1567         /* Nothing to do! */
1568         clear_bit(WRITE_PENDING, &con->state);
1569         dout("try_write nothing else to write.\n");
1570 done:
1571         ret = 0;
1572 out:
1573         mutex_unlock(&con->out_mutex);
1574         dout("try_write done on %p\n", con);
1575         return ret;
1576 }
1577
1578
1579
1580 /*
1581  * Read what we can from the socket.
1582  */
1583 static int try_read(struct ceph_connection *con)
1584 {
1585         struct ceph_messenger *msgr;
1586         int ret = -1;
1587
1588         if (!con->sock)
1589                 return 0;
1590
1591         if (test_bit(STANDBY, &con->state))
1592                 return 0;
1593
1594         dout("try_read start on %p\n", con);
1595         msgr = con->msgr;
1596
1597 more:
1598         dout("try_read tag %d in_base_pos %d\n", (int)con->in_tag,
1599              con->in_base_pos);
1600         if (test_bit(CONNECTING, &con->state)) {
1601                 if (!test_bit(NEGOTIATING, &con->state)) {
1602                         dout("try_read connecting\n");
1603                         ret = read_partial_banner(con);
1604                         if (ret <= 0)
1605                                 goto done;
1606                         if (process_banner(con) < 0) {
1607                                 ret = -1;
1608                                 goto out;
1609                         }
1610                 }
1611                 ret = read_partial_connect(con);
1612                 if (ret <= 0)
1613                         goto done;
1614                 if (process_connect(con) < 0) {
1615                         ret = -1;
1616                         goto out;
1617                 }
1618                 goto more;
1619         }
1620
1621         if (con->in_base_pos < 0) {
1622                 /*
1623                  * skipping + discarding content.
1624                  *
1625                  * FIXME: there must be a better way to do this!
1626                  */
1627                 static char buf[1024];
1628                 int skip = min(1024, -con->in_base_pos);
1629                 dout("skipping %d / %d bytes\n", skip, -con->in_base_pos);
1630                 ret = ceph_tcp_recvmsg(con->sock, buf, skip);
1631                 if (ret <= 0)
1632                         goto done;
1633                 con->in_base_pos += ret;
1634                 if (con->in_base_pos)
1635                         goto more;
1636         }
1637         if (con->in_tag == CEPH_MSGR_TAG_READY) {
1638                 /*
1639                  * what's next?
1640                  */
1641                 ret = ceph_tcp_recvmsg(con->sock, &con->in_tag, 1);
1642                 if (ret <= 0)
1643                         goto done;
1644                 dout("try_read got tag %d\n", (int)con->in_tag);
1645                 switch (con->in_tag) {
1646                 case CEPH_MSGR_TAG_MSG:
1647                         prepare_read_message(con);
1648                         break;
1649                 case CEPH_MSGR_TAG_ACK:
1650                         prepare_read_ack(con);
1651                         break;
1652                 case CEPH_MSGR_TAG_CLOSE:
1653                         set_bit(CLOSED, &con->state);   /* fixme */
1654                         goto done;
1655                 default:
1656                         goto bad_tag;
1657                 }
1658         }
1659         if (con->in_tag == CEPH_MSGR_TAG_MSG) {
1660                 ret = read_partial_message(con);
1661                 if (ret <= 0) {
1662                         switch (ret) {
1663                         case -EBADMSG:
1664                                 con->error_msg = "bad crc";
1665                                 ret = -EIO;
1666                                 goto out;
1667                         case -EIO:
1668                                 con->error_msg = "io error";
1669                                 goto out;
1670                         default:
1671                                 goto done;
1672                         }
1673                 }
1674                 if (con->in_tag == CEPH_MSGR_TAG_READY)
1675                         goto more;
1676                 process_message(con);
1677                 goto more;
1678         }
1679         if (con->in_tag == CEPH_MSGR_TAG_ACK) {
1680                 ret = read_partial_ack(con);
1681                 if (ret <= 0)
1682                         goto done;
1683                 process_ack(con);
1684                 goto more;
1685         }
1686
1687 done:
1688         ret = 0;
1689 out:
1690         dout("try_read done on %p\n", con);
1691         return ret;
1692
1693 bad_tag:
1694         pr_err("try_read bad con->in_tag = %d\n", (int)con->in_tag);
1695         con->error_msg = "protocol error, garbage tag";
1696         ret = -1;
1697         goto out;
1698 }
1699
1700
1701 /*
1702  * Atomically queue work on a connection.  Bump @con reference to
1703  * avoid races with connection teardown.
1704  *
1705  * There is some trickery going on with QUEUED and BUSY because we
1706  * only want a _single_ thread operating on each connection at any
1707  * point in time, but we want to use all available CPUs.
1708  *
1709  * The worker thread only proceeds if it can atomically set BUSY.  It
1710  * clears QUEUED and does it's thing.  When it thinks it's done, it
1711  * clears BUSY, then rechecks QUEUED.. if it's set again, it loops
1712  * (tries again to set BUSY).
1713  *
1714  * To queue work, we first set QUEUED, _then_ if BUSY isn't set, we
1715  * try to queue work.  If that fails (work is already queued, or BUSY)
1716  * we give up (work also already being done or is queued) but leave QUEUED
1717  * set so that the worker thread will loop if necessary.
1718  */
1719 static void queue_con(struct ceph_connection *con)
1720 {
1721         if (test_bit(DEAD, &con->state)) {
1722                 dout("queue_con %p ignoring: DEAD\n",
1723                      con);
1724                 return;
1725         }
1726
1727         if (!con->ops->get(con)) {
1728                 dout("queue_con %p ref count 0\n", con);
1729                 return;
1730         }
1731
1732         set_bit(QUEUED, &con->state);
1733         if (test_bit(BUSY, &con->state)) {
1734                 dout("queue_con %p - already BUSY\n", con);
1735                 con->ops->put(con);
1736         } else if (!queue_work(ceph_msgr_wq, &con->work.work)) {
1737                 dout("queue_con %p - already queued\n", con);
1738                 con->ops->put(con);
1739         } else {
1740                 dout("queue_con %p\n", con);
1741         }
1742 }
1743
1744 /*
1745  * Do some work on a connection.  Drop a connection ref when we're done.
1746  */
1747 static void con_work(struct work_struct *work)
1748 {
1749         struct ceph_connection *con = container_of(work, struct ceph_connection,
1750                                                    work.work);
1751         int backoff = 0;
1752
1753 more:
1754         if (test_and_set_bit(BUSY, &con->state) != 0) {
1755                 dout("con_work %p BUSY already set\n", con);
1756                 goto out;
1757         }
1758         dout("con_work %p start, clearing QUEUED\n", con);
1759         clear_bit(QUEUED, &con->state);
1760
1761         if (test_bit(CLOSED, &con->state)) { /* e.g. if we are replaced */
1762                 dout("con_work CLOSED\n");
1763                 con_close_socket(con);
1764                 goto done;
1765         }
1766         if (test_and_clear_bit(OPENING, &con->state)) {
1767                 /* reopen w/ new peer */
1768                 dout("con_work OPENING\n");
1769                 con_close_socket(con);
1770         }
1771
1772         if (test_and_clear_bit(SOCK_CLOSED, &con->state) ||
1773             try_read(con) < 0 ||
1774             try_write(con) < 0) {
1775                 backoff = 1;
1776                 ceph_fault(con);     /* error/fault path */
1777         }
1778
1779 done:
1780         clear_bit(BUSY, &con->state);
1781         dout("con->state=%lu\n", con->state);
1782         if (test_bit(QUEUED, &con->state)) {
1783                 if (!backoff) {
1784                         dout("con_work %p QUEUED reset, looping\n", con);
1785                         goto more;
1786                 }
1787                 dout("con_work %p QUEUED reset, but just faulted\n", con);
1788                 clear_bit(QUEUED, &con->state);
1789         }
1790         dout("con_work %p done\n", con);
1791
1792 out:
1793         con->ops->put(con);
1794 }
1795
1796
1797 /*
1798  * Generic error/fault handler.  A retry mechanism is used with
1799  * exponential backoff
1800  */
1801 static void ceph_fault(struct ceph_connection *con)
1802 {
1803         pr_err("%s%lld %s %s\n", ENTITY_NAME(con->peer_name),
1804                pr_addr(&con->peer_addr.in_addr), con->error_msg);
1805         dout("fault %p state %lu to peer %s\n",
1806              con, con->state, pr_addr(&con->peer_addr.in_addr));
1807
1808         if (test_bit(LOSSYTX, &con->state)) {
1809                 dout("fault on LOSSYTX channel\n");
1810                 goto out;
1811         }
1812
1813         clear_bit(BUSY, &con->state);  /* to avoid an improbable race */
1814
1815         con_close_socket(con);
1816
1817         if (con->in_msg) {
1818                 ceph_msg_put(con->in_msg);
1819                 con->in_msg = NULL;
1820         }
1821
1822         /* If there are no messages in the queue, place the connection
1823          * in a STANDBY state (i.e., don't try to reconnect just yet). */
1824         mutex_lock(&con->out_mutex);
1825         if (list_empty(&con->out_queue) && !con->out_keepalive_pending) {
1826                 dout("fault setting STANDBY\n");
1827                 set_bit(STANDBY, &con->state);
1828                 mutex_unlock(&con->out_mutex);
1829                 goto out;
1830         }
1831
1832         /* Requeue anything that hasn't been acked, and retry after a
1833          * delay. */
1834         list_splice_init(&con->out_sent, &con->out_queue);
1835         mutex_unlock(&con->out_mutex);
1836
1837         if (con->delay == 0)
1838                 con->delay = BASE_DELAY_INTERVAL;
1839         else if (con->delay < MAX_DELAY_INTERVAL)
1840                 con->delay *= 2;
1841
1842         /* explicitly schedule work to try to reconnect again later. */
1843         dout("fault queueing %p delay %lu\n", con, con->delay);
1844         con->ops->get(con);
1845         if (queue_delayed_work(ceph_msgr_wq, &con->work,
1846                                round_jiffies_relative(con->delay)) == 0)
1847                 con->ops->put(con);
1848
1849 out:
1850         if (con->ops->fault)
1851                 con->ops->fault(con);
1852 }
1853
1854
1855
1856 /*
1857  * create a new messenger instance
1858  */
1859 struct ceph_messenger *ceph_messenger_create(struct ceph_entity_addr *myaddr)
1860 {
1861         struct ceph_messenger *msgr;
1862
1863         msgr = kzalloc(sizeof(*msgr), GFP_KERNEL);
1864         if (msgr == NULL)
1865                 return ERR_PTR(-ENOMEM);
1866
1867         spin_lock_init(&msgr->global_seq_lock);
1868
1869         /* the zero page is needed if a request is "canceled" while the message
1870          * is being written over the socket */
1871         msgr->zero_page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1872         if (!msgr->zero_page) {
1873                 kfree(msgr);
1874                 return ERR_PTR(-ENOMEM);
1875         }
1876         kmap(msgr->zero_page);
1877
1878         if (myaddr)
1879                 msgr->inst.addr = *myaddr;
1880
1881         /* select a random nonce */
1882         get_random_bytes(&msgr->inst.addr.nonce,
1883                          sizeof(msgr->inst.addr.nonce));
1884         encode_my_addr(msgr);
1885
1886         dout("messenger_create %p\n", msgr);
1887         return msgr;
1888 }
1889
1890 void ceph_messenger_destroy(struct ceph_messenger *msgr)
1891 {
1892         dout("destroy %p\n", msgr);
1893         kunmap(msgr->zero_page);
1894         __free_page(msgr->zero_page);
1895         kfree(msgr);
1896         dout("destroyed messenger %p\n", msgr);
1897 }
1898
1899 /*
1900  * Queue up an outgoing message on the given connection.
1901  */
1902 void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg)
1903 {
1904         if (test_bit(CLOSED, &con->state)) {
1905                 dout("con_send %p closed, dropping %p\n", con, msg);
1906                 ceph_msg_put(msg);
1907                 return;
1908         }
1909
1910         /* set src+dst */
1911         msg->hdr.src.name = con->msgr->inst.name;
1912         msg->hdr.src.addr = con->msgr->my_enc_addr;
1913         msg->hdr.orig_src = msg->hdr.src;
1914         msg->hdr.dst_erank = con->peer_addr.erank;
1915
1916         /* queue */
1917         mutex_lock(&con->out_mutex);
1918         BUG_ON(!list_empty(&msg->list_head));
1919         list_add_tail(&msg->list_head, &con->out_queue);
1920         dout("----- %p to %s%lld %d=%s len %d+%d+%d -----\n", msg,
1921              ENTITY_NAME(con->peer_name), le16_to_cpu(msg->hdr.type),
1922              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1923              le32_to_cpu(msg->hdr.front_len),
1924              le32_to_cpu(msg->hdr.middle_len),
1925              le32_to_cpu(msg->hdr.data_len));
1926         mutex_unlock(&con->out_mutex);
1927
1928         /* if there wasn't anything waiting to send before, queue
1929          * new work */
1930         if (test_and_set_bit(WRITE_PENDING, &con->state) == 0)
1931                 queue_con(con);
1932 }
1933
1934 /*
1935  * Revoke a message that was previously queued for send
1936  */
1937 void ceph_con_revoke(struct ceph_connection *con, struct ceph_msg *msg)
1938 {
1939         mutex_lock(&con->out_mutex);
1940         if (!list_empty(&msg->list_head)) {
1941                 dout("con_revoke %p msg %p\n", con, msg);
1942                 list_del_init(&msg->list_head);
1943                 ceph_msg_put(msg);
1944                 msg->hdr.seq = 0;
1945                 if (con->out_msg == msg) {
1946                         ceph_msg_put(con->out_msg);
1947                         con->out_msg = NULL;
1948                 }
1949                 if (con->out_kvec_is_msg) {
1950                         con->out_skip = con->out_kvec_bytes;
1951                         con->out_kvec_is_msg = false;
1952                 }
1953         } else {
1954                 dout("con_revoke %p msg %p - not queued (sent?)\n", con, msg);
1955         }
1956         mutex_unlock(&con->out_mutex);
1957 }
1958
1959 /*
1960  * Queue a keepalive byte to ensure the tcp connection is alive.
1961  */
1962 void ceph_con_keepalive(struct ceph_connection *con)
1963 {
1964         if (test_and_set_bit(KEEPALIVE_PENDING, &con->state) == 0 &&
1965             test_and_set_bit(WRITE_PENDING, &con->state) == 0)
1966                 queue_con(con);
1967 }
1968
1969
1970 /*
1971  * construct a new message with given type, size
1972  * the new msg has a ref count of 1.
1973  */
1974 struct ceph_msg *ceph_msg_new(int type, int front_len,
1975                               int page_len, int page_off, struct page **pages)
1976 {
1977         struct ceph_msg *m;
1978
1979         m = kmalloc(sizeof(*m), GFP_NOFS);
1980         if (m == NULL)
1981                 goto out;
1982         kref_init(&m->kref);
1983         INIT_LIST_HEAD(&m->list_head);
1984
1985         m->hdr.type = cpu_to_le16(type);
1986         m->hdr.front_len = cpu_to_le32(front_len);
1987         m->hdr.middle_len = 0;
1988         m->hdr.data_len = cpu_to_le32(page_len);
1989         m->hdr.data_off = cpu_to_le16(page_off);
1990         m->hdr.priority = cpu_to_le16(CEPH_MSG_PRIO_DEFAULT);
1991         m->footer.front_crc = 0;
1992         m->footer.middle_crc = 0;
1993         m->footer.data_crc = 0;
1994         m->front_max = front_len;
1995         m->front_is_vmalloc = false;
1996         m->more_to_follow = false;
1997         m->pool = NULL;
1998
1999         /* front */
2000         if (front_len) {
2001                 if (front_len > PAGE_CACHE_SIZE) {
2002                         m->front.iov_base = __vmalloc(front_len, GFP_NOFS,
2003                                                       PAGE_KERNEL);
2004                         m->front_is_vmalloc = true;
2005                 } else {
2006                         m->front.iov_base = kmalloc(front_len, GFP_NOFS);
2007                 }
2008                 if (m->front.iov_base == NULL) {
2009                         pr_err("msg_new can't allocate %d bytes\n",
2010                              front_len);
2011                         goto out2;
2012                 }
2013         } else {
2014                 m->front.iov_base = NULL;
2015         }
2016         m->front.iov_len = front_len;
2017
2018         /* middle */
2019         m->middle = NULL;
2020
2021         /* data */
2022         m->nr_pages = calc_pages_for(page_off, page_len);
2023         m->pages = pages;
2024
2025         dout("ceph_msg_new %p page %d~%d -> %d\n", m, page_off, page_len,
2026              m->nr_pages);
2027         return m;
2028
2029 out2:
2030         ceph_msg_put(m);
2031 out:
2032         pr_err("msg_new can't create type %d len %d\n", type, front_len);
2033         return ERR_PTR(-ENOMEM);
2034 }
2035
2036 /*
2037  * Generic message allocator, for incoming messages.
2038  */
2039 struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
2040                                 struct ceph_msg_header *hdr)
2041 {
2042         int type = le16_to_cpu(hdr->type);
2043         int front_len = le32_to_cpu(hdr->front_len);
2044         struct ceph_msg *msg = ceph_msg_new(type, front_len, 0, 0, NULL);
2045
2046         if (!msg) {
2047                 pr_err("unable to allocate msg type %d len %d\n",
2048                        type, front_len);
2049                 return ERR_PTR(-ENOMEM);
2050         }
2051         return msg;
2052 }
2053
2054 /*
2055  * Allocate "middle" portion of a message, if it is needed and wasn't
2056  * allocated by alloc_msg.  This allows us to read a small fixed-size
2057  * per-type header in the front and then gracefully fail (i.e.,
2058  * propagate the error to the caller based on info in the front) when
2059  * the middle is too large.
2060  */
2061 int ceph_alloc_middle(struct ceph_connection *con, struct ceph_msg *msg)
2062 {
2063         int type = le16_to_cpu(msg->hdr.type);
2064         int middle_len = le32_to_cpu(msg->hdr.middle_len);
2065
2066         dout("alloc_middle %p type %d %s middle_len %d\n", msg, type,
2067              ceph_msg_type_name(type), middle_len);
2068         BUG_ON(!middle_len);
2069         BUG_ON(msg->middle);
2070
2071         msg->middle = ceph_buffer_new(middle_len, GFP_NOFS);
2072         if (!msg->middle)
2073                 return -ENOMEM;
2074         return 0;
2075 }
2076
2077
2078 /*
2079  * Free a generically kmalloc'd message.
2080  */
2081 void ceph_msg_kfree(struct ceph_msg *m)
2082 {
2083         dout("msg_kfree %p\n", m);
2084         if (m->front_is_vmalloc)
2085                 vfree(m->front.iov_base);
2086         else
2087                 kfree(m->front.iov_base);
2088         kfree(m);
2089 }
2090
2091 /*
2092  * Drop a msg ref.  Destroy as needed.
2093  */
2094 void ceph_msg_last_put(struct kref *kref)
2095 {
2096         struct ceph_msg *m = container_of(kref, struct ceph_msg, kref);
2097
2098         dout("ceph_msg_put last one on %p\n", m);
2099         WARN_ON(!list_empty(&m->list_head));
2100
2101         /* drop middle, data, if any */
2102         if (m->middle) {
2103                 ceph_buffer_put(m->middle);
2104                 m->middle = NULL;
2105         }
2106         m->nr_pages = 0;
2107         m->pages = NULL;
2108
2109         if (m->pool)
2110                 ceph_msgpool_put(m->pool, m);
2111         else
2112                 ceph_msg_kfree(m);
2113 }