1 #include "ceph_debug.h"
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>
9 #include <linux/socket.h>
10 #include <linux/string.h>
14 #include "messenger.h"
19 * Ceph uses the messenger to exchange ceph_msg messages with other
20 * hosts in the system. The messenger provides ordered and reliable
21 * delivery. We tolerate TCP disconnects by reconnecting (with
22 * exponential backoff) in the case of a fault (disconnection, bad
23 * crc, protocol error). Acks allow sent messages to be discarded by
27 /* static tag bytes (protocol control messages) */
28 static char tag_msg = CEPH_MSGR_TAG_MSG;
29 static char tag_ack = CEPH_MSGR_TAG_ACK;
30 static char tag_keepalive = CEPH_MSGR_TAG_KEEPALIVE;
33 static void queue_con(struct ceph_connection *con);
34 static void con_work(struct work_struct *);
35 static void ceph_fault(struct ceph_connection *con);
37 const char *ceph_name_type_str(int t)
40 case CEPH_ENTITY_TYPE_MON: return "mon";
41 case CEPH_ENTITY_TYPE_MDS: return "mds";
42 case CEPH_ENTITY_TYPE_OSD: return "osd";
43 case CEPH_ENTITY_TYPE_CLIENT: return "client";
44 case CEPH_ENTITY_TYPE_ADMIN: return "admin";
45 default: return "???";
50 * nicely render a sockaddr as a string.
52 #define MAX_ADDR_STR 20
53 static char addr_str[MAX_ADDR_STR][40];
54 static DEFINE_SPINLOCK(addr_str_lock);
55 static int last_addr_str;
57 const char *pr_addr(const struct sockaddr_storage *ss)
61 struct sockaddr_in *in4 = (void *)ss;
62 unsigned char *quad = (void *)&in4->sin_addr.s_addr;
63 struct sockaddr_in6 *in6 = (void *)ss;
65 spin_lock(&addr_str_lock);
67 if (last_addr_str == MAX_ADDR_STR)
69 spin_unlock(&addr_str_lock);
72 switch (ss->ss_family) {
74 sprintf(s, "%u.%u.%u.%u:%u",
75 (unsigned int)quad[0],
76 (unsigned int)quad[1],
77 (unsigned int)quad[2],
78 (unsigned int)quad[3],
79 (unsigned int)ntohs(in4->sin_port));
83 sprintf(s, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x:%u",
84 in6->sin6_addr.s6_addr16[0],
85 in6->sin6_addr.s6_addr16[1],
86 in6->sin6_addr.s6_addr16[2],
87 in6->sin6_addr.s6_addr16[3],
88 in6->sin6_addr.s6_addr16[4],
89 in6->sin6_addr.s6_addr16[5],
90 in6->sin6_addr.s6_addr16[6],
91 in6->sin6_addr.s6_addr16[7],
92 (unsigned int)ntohs(in6->sin6_port));
96 sprintf(s, "(unknown sockaddr family %d)", (int)ss->ss_family);
102 static void encode_my_addr(struct ceph_messenger *msgr)
104 memcpy(&msgr->my_enc_addr, &msgr->inst.addr, sizeof(msgr->my_enc_addr));
105 ceph_encode_addr(&msgr->my_enc_addr);
109 * work queue for all reading and writing to/from the socket.
111 struct workqueue_struct *ceph_msgr_wq;
113 int __init ceph_msgr_init(void)
115 ceph_msgr_wq = create_workqueue("ceph-msgr");
116 if (IS_ERR(ceph_msgr_wq)) {
117 int ret = PTR_ERR(ceph_msgr_wq);
118 pr_err("msgr_init failed to create workqueue: %d\n", ret);
125 void ceph_msgr_exit(void)
127 destroy_workqueue(ceph_msgr_wq);
131 * socket callback functions
134 /* data available on socket, or listen socket received a connect */
135 static void ceph_data_ready(struct sock *sk, int count_unused)
137 struct ceph_connection *con =
138 (struct ceph_connection *)sk->sk_user_data;
139 if (sk->sk_state != TCP_CLOSE_WAIT) {
140 dout("ceph_data_ready on %p state = %lu, queueing work\n",
146 /* socket has buffer space for writing */
147 static void ceph_write_space(struct sock *sk)
149 struct ceph_connection *con =
150 (struct ceph_connection *)sk->sk_user_data;
152 /* only queue to workqueue if there is data we want to write. */
153 if (test_bit(WRITE_PENDING, &con->state)) {
154 dout("ceph_write_space %p queueing write work\n", con);
157 dout("ceph_write_space %p nothing to write\n", con);
160 /* since we have our own write_space, clear the SOCK_NOSPACE flag */
161 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
164 /* socket's state has changed */
165 static void ceph_state_change(struct sock *sk)
167 struct ceph_connection *con =
168 (struct ceph_connection *)sk->sk_user_data;
170 dout("ceph_state_change %p state = %lu sk_state = %u\n",
171 con, con->state, sk->sk_state);
173 if (test_bit(CLOSED, &con->state))
176 switch (sk->sk_state) {
178 dout("ceph_state_change TCP_CLOSE\n");
180 dout("ceph_state_change TCP_CLOSE_WAIT\n");
181 if (test_and_set_bit(SOCK_CLOSED, &con->state) == 0) {
182 if (test_bit(CONNECTING, &con->state))
183 con->error_msg = "connection failed";
185 con->error_msg = "socket closed";
189 case TCP_ESTABLISHED:
190 dout("ceph_state_change TCP_ESTABLISHED\n");
197 * set up socket callbacks
199 static void set_sock_callbacks(struct socket *sock,
200 struct ceph_connection *con)
202 struct sock *sk = sock->sk;
203 sk->sk_user_data = (void *)con;
204 sk->sk_data_ready = ceph_data_ready;
205 sk->sk_write_space = ceph_write_space;
206 sk->sk_state_change = ceph_state_change;
215 * initiate connection to a remote socket.
217 static struct socket *ceph_tcp_connect(struct ceph_connection *con)
219 struct sockaddr *paddr = (struct sockaddr *)&con->peer_addr.in_addr;
224 ret = sock_create_kern(AF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
228 sock->sk->sk_allocation = GFP_NOFS;
230 set_sock_callbacks(sock, con);
232 dout("connect %s\n", pr_addr(&con->peer_addr.in_addr));
234 ret = sock->ops->connect(sock, paddr, sizeof(*paddr), O_NONBLOCK);
235 if (ret == -EINPROGRESS) {
236 dout("connect %s EINPROGRESS sk_state = %u\n",
237 pr_addr(&con->peer_addr.in_addr),
242 pr_err("connect %s error %d\n",
243 pr_addr(&con->peer_addr.in_addr), ret);
246 con->error_msg = "connect error";
254 static int ceph_tcp_recvmsg(struct socket *sock, void *buf, size_t len)
256 struct kvec iov = {buf, len};
257 struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
259 return kernel_recvmsg(sock, &msg, &iov, 1, len, msg.msg_flags);
263 * write something. @more is true if caller will be sending more data
266 static int ceph_tcp_sendmsg(struct socket *sock, struct kvec *iov,
267 size_t kvlen, size_t len, int more)
269 struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
272 msg.msg_flags |= MSG_MORE;
274 msg.msg_flags |= MSG_EOR; /* superfluous, but what the hell */
276 return kernel_sendmsg(sock, &msg, iov, kvlen, len);
281 * Shutdown/close the socket for the given connection.
283 static int con_close_socket(struct ceph_connection *con)
287 dout("con_close_socket on %p sock %p\n", con, con->sock);
290 set_bit(SOCK_CLOSED, &con->state);
291 rc = con->sock->ops->shutdown(con->sock, SHUT_RDWR);
292 sock_release(con->sock);
294 clear_bit(SOCK_CLOSED, &con->state);
299 * Reset a connection. Discard all incoming and outgoing messages
300 * and clear *_seq state.
302 static void ceph_msg_remove(struct ceph_msg *msg)
304 list_del_init(&msg->list_head);
307 static void ceph_msg_remove_list(struct list_head *head)
309 while (!list_empty(head)) {
310 struct ceph_msg *msg = list_first_entry(head, struct ceph_msg,
312 ceph_msg_remove(msg);
316 static void reset_connection(struct ceph_connection *con)
318 /* reset connection, out_queue, msg_ and connect_seq */
319 /* discard existing out_queue and msg_seq */
320 ceph_msg_remove_list(&con->out_queue);
321 ceph_msg_remove_list(&con->out_sent);
324 ceph_msg_put(con->in_msg);
328 con->connect_seq = 0;
331 ceph_msg_put(con->out_msg);
338 * mark a peer down. drop any open connections.
340 void ceph_con_close(struct ceph_connection *con)
342 dout("con_close %p peer %s\n", con, pr_addr(&con->peer_addr.in_addr));
343 set_bit(CLOSED, &con->state); /* in case there's queued work */
344 clear_bit(STANDBY, &con->state); /* avoid connect_seq bump */
345 mutex_lock(&con->mutex);
346 reset_connection(con);
347 mutex_unlock(&con->mutex);
352 * Reopen a closed connection, with a new peer address.
354 void ceph_con_open(struct ceph_connection *con, struct ceph_entity_addr *addr)
356 dout("con_open %p %s\n", con, pr_addr(&addr->in_addr));
357 set_bit(OPENING, &con->state);
358 clear_bit(CLOSED, &con->state);
359 memcpy(&con->peer_addr, addr, sizeof(*addr));
360 con->delay = 0; /* reset backoff memory */
367 struct ceph_connection *ceph_con_get(struct ceph_connection *con)
369 dout("con_get %p nref = %d -> %d\n", con,
370 atomic_read(&con->nref), atomic_read(&con->nref) + 1);
371 if (atomic_inc_not_zero(&con->nref))
376 void ceph_con_put(struct ceph_connection *con)
378 dout("con_put %p nref = %d -> %d\n", con,
379 atomic_read(&con->nref), atomic_read(&con->nref) - 1);
380 BUG_ON(atomic_read(&con->nref) == 0);
381 if (atomic_dec_and_test(&con->nref)) {
388 * initialize a new connection.
390 void ceph_con_init(struct ceph_messenger *msgr, struct ceph_connection *con)
392 dout("con_init %p\n", con);
393 memset(con, 0, sizeof(*con));
394 atomic_set(&con->nref, 1);
396 mutex_init(&con->mutex);
397 INIT_LIST_HEAD(&con->out_queue);
398 INIT_LIST_HEAD(&con->out_sent);
399 INIT_DELAYED_WORK(&con->work, con_work);
404 * We maintain a global counter to order connection attempts. Get
405 * a unique seq greater than @gt.
407 static u32 get_global_seq(struct ceph_messenger *msgr, u32 gt)
411 spin_lock(&msgr->global_seq_lock);
412 if (msgr->global_seq < gt)
413 msgr->global_seq = gt;
414 ret = ++msgr->global_seq;
415 spin_unlock(&msgr->global_seq_lock);
421 * Prepare footer for currently outgoing message, and finish things
422 * off. Assumes out_kvec* are already valid.. we just add on to the end.
424 static void prepare_write_message_footer(struct ceph_connection *con, int v)
426 struct ceph_msg *m = con->out_msg;
428 dout("prepare_write_message_footer %p\n", con);
429 con->out_kvec_is_msg = true;
430 con->out_kvec[v].iov_base = &m->footer;
431 con->out_kvec[v].iov_len = sizeof(m->footer);
432 con->out_kvec_bytes += sizeof(m->footer);
433 con->out_kvec_left++;
434 con->out_more = m->more_to_follow;
435 con->out_msg_done = true;
439 * Prepare headers for the next outgoing message.
441 static void prepare_write_message(struct ceph_connection *con)
446 con->out_kvec_bytes = 0;
447 con->out_kvec_is_msg = true;
448 con->out_msg_done = false;
450 /* Sneak an ack in there first? If we can get it into the same
451 * TCP packet that's a good thing. */
452 if (con->in_seq > con->in_seq_acked) {
453 con->in_seq_acked = con->in_seq;
454 con->out_kvec[v].iov_base = &tag_ack;
455 con->out_kvec[v++].iov_len = 1;
456 con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
457 con->out_kvec[v].iov_base = &con->out_temp_ack;
458 con->out_kvec[v++].iov_len = sizeof(con->out_temp_ack);
459 con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
462 m = list_first_entry(&con->out_queue,
463 struct ceph_msg, list_head);
465 if (test_bit(LOSSYTX, &con->state)) {
466 /* put message on sent list */
468 list_move_tail(&m->list_head, &con->out_sent);
470 list_del_init(&m->list_head);
473 m->hdr.seq = cpu_to_le64(++con->out_seq);
475 dout("prepare_write_message %p seq %lld type %d len %d+%d+%d %d pgs\n",
476 m, con->out_seq, le16_to_cpu(m->hdr.type),
477 le32_to_cpu(m->hdr.front_len), le32_to_cpu(m->hdr.middle_len),
478 le32_to_cpu(m->hdr.data_len),
480 BUG_ON(le32_to_cpu(m->hdr.front_len) != m->front.iov_len);
482 /* tag + hdr + front + middle */
483 con->out_kvec[v].iov_base = &tag_msg;
484 con->out_kvec[v++].iov_len = 1;
485 con->out_kvec[v].iov_base = &m->hdr;
486 con->out_kvec[v++].iov_len = sizeof(m->hdr);
487 con->out_kvec[v++] = m->front;
489 con->out_kvec[v++] = m->middle->vec;
490 con->out_kvec_left = v;
491 con->out_kvec_bytes += 1 + sizeof(m->hdr) + m->front.iov_len +
492 (m->middle ? m->middle->vec.iov_len : 0);
493 con->out_kvec_cur = con->out_kvec;
495 /* fill in crc (except data pages), footer */
496 con->out_msg->hdr.crc =
497 cpu_to_le32(crc32c(0, (void *)&m->hdr,
498 sizeof(m->hdr) - sizeof(m->hdr.crc)));
499 con->out_msg->footer.flags = CEPH_MSG_FOOTER_COMPLETE;
500 con->out_msg->footer.front_crc =
501 cpu_to_le32(crc32c(0, m->front.iov_base, m->front.iov_len));
503 con->out_msg->footer.middle_crc =
504 cpu_to_le32(crc32c(0, m->middle->vec.iov_base,
505 m->middle->vec.iov_len));
507 con->out_msg->footer.middle_crc = 0;
508 con->out_msg->footer.data_crc = 0;
509 dout("prepare_write_message front_crc %u data_crc %u\n",
510 le32_to_cpu(con->out_msg->footer.front_crc),
511 le32_to_cpu(con->out_msg->footer.middle_crc));
513 /* is there a data payload? */
514 if (le32_to_cpu(m->hdr.data_len) > 0) {
515 /* initialize page iterator */
516 con->out_msg_pos.page = 0;
517 con->out_msg_pos.page_pos =
518 le16_to_cpu(m->hdr.data_off) & ~PAGE_MASK;
519 con->out_msg_pos.data_pos = 0;
520 con->out_msg_pos.did_page_crc = 0;
521 con->out_more = 1; /* data + footer will follow */
523 /* no, queue up footer too and be done */
524 prepare_write_message_footer(con, v);
527 set_bit(WRITE_PENDING, &con->state);
533 static void prepare_write_ack(struct ceph_connection *con)
535 dout("prepare_write_ack %p %llu -> %llu\n", con,
536 con->in_seq_acked, con->in_seq);
537 con->in_seq_acked = con->in_seq;
539 con->out_kvec[0].iov_base = &tag_ack;
540 con->out_kvec[0].iov_len = 1;
541 con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
542 con->out_kvec[1].iov_base = &con->out_temp_ack;
543 con->out_kvec[1].iov_len = sizeof(con->out_temp_ack);
544 con->out_kvec_left = 2;
545 con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
546 con->out_kvec_cur = con->out_kvec;
547 con->out_more = 1; /* more will follow.. eventually.. */
548 set_bit(WRITE_PENDING, &con->state);
552 * Prepare to write keepalive byte.
554 static void prepare_write_keepalive(struct ceph_connection *con)
556 dout("prepare_write_keepalive %p\n", con);
557 con->out_kvec[0].iov_base = &tag_keepalive;
558 con->out_kvec[0].iov_len = 1;
559 con->out_kvec_left = 1;
560 con->out_kvec_bytes = 1;
561 con->out_kvec_cur = con->out_kvec;
562 set_bit(WRITE_PENDING, &con->state);
566 * Connection negotiation.
569 static void prepare_connect_authorizer(struct ceph_connection *con)
573 int auth_protocol = 0;
575 mutex_unlock(&con->mutex);
576 if (con->ops->get_authorizer)
577 con->ops->get_authorizer(con, &auth_buf, &auth_len,
578 &auth_protocol, &con->auth_reply_buf,
579 &con->auth_reply_buf_len,
581 mutex_lock(&con->mutex);
583 con->out_connect.authorizer_protocol = cpu_to_le32(auth_protocol);
584 con->out_connect.authorizer_len = cpu_to_le32(auth_len);
586 con->out_kvec[con->out_kvec_left].iov_base = auth_buf;
587 con->out_kvec[con->out_kvec_left].iov_len = auth_len;
588 con->out_kvec_left++;
589 con->out_kvec_bytes += auth_len;
593 * We connected to a peer and are saying hello.
595 static void prepare_write_banner(struct ceph_messenger *msgr,
596 struct ceph_connection *con)
598 int len = strlen(CEPH_BANNER);
600 con->out_kvec[0].iov_base = CEPH_BANNER;
601 con->out_kvec[0].iov_len = len;
602 con->out_kvec[1].iov_base = &msgr->my_enc_addr;
603 con->out_kvec[1].iov_len = sizeof(msgr->my_enc_addr);
604 con->out_kvec_left = 2;
605 con->out_kvec_bytes = len + sizeof(msgr->my_enc_addr);
606 con->out_kvec_cur = con->out_kvec;
608 set_bit(WRITE_PENDING, &con->state);
611 static void prepare_write_connect(struct ceph_messenger *msgr,
612 struct ceph_connection *con,
615 unsigned global_seq = get_global_seq(con->msgr, 0);
618 switch (con->peer_name.type) {
619 case CEPH_ENTITY_TYPE_MON:
620 proto = CEPH_MONC_PROTOCOL;
622 case CEPH_ENTITY_TYPE_OSD:
623 proto = CEPH_OSDC_PROTOCOL;
625 case CEPH_ENTITY_TYPE_MDS:
626 proto = CEPH_MDSC_PROTOCOL;
632 dout("prepare_write_connect %p cseq=%d gseq=%d proto=%d\n", con,
633 con->connect_seq, global_seq, proto);
635 con->out_connect.features = CEPH_FEATURE_SUPPORTED;
636 con->out_connect.host_type = cpu_to_le32(CEPH_ENTITY_TYPE_CLIENT);
637 con->out_connect.connect_seq = cpu_to_le32(con->connect_seq);
638 con->out_connect.global_seq = cpu_to_le32(global_seq);
639 con->out_connect.protocol_version = cpu_to_le32(proto);
640 con->out_connect.flags = 0;
643 con->out_kvec_left = 0;
644 con->out_kvec_bytes = 0;
646 con->out_kvec[con->out_kvec_left].iov_base = &con->out_connect;
647 con->out_kvec[con->out_kvec_left].iov_len = sizeof(con->out_connect);
648 con->out_kvec_left++;
649 con->out_kvec_bytes += sizeof(con->out_connect);
650 con->out_kvec_cur = con->out_kvec;
652 set_bit(WRITE_PENDING, &con->state);
654 prepare_connect_authorizer(con);
659 * write as much of pending kvecs to the socket as we can.
661 * 0 -> socket full, but more to do
664 static int write_partial_kvec(struct ceph_connection *con)
668 dout("write_partial_kvec %p %d left\n", con, con->out_kvec_bytes);
669 while (con->out_kvec_bytes > 0) {
670 ret = ceph_tcp_sendmsg(con->sock, con->out_kvec_cur,
671 con->out_kvec_left, con->out_kvec_bytes,
675 con->out_kvec_bytes -= ret;
676 if (con->out_kvec_bytes == 0)
679 if (ret >= con->out_kvec_cur->iov_len) {
680 ret -= con->out_kvec_cur->iov_len;
682 con->out_kvec_left--;
684 con->out_kvec_cur->iov_len -= ret;
685 con->out_kvec_cur->iov_base += ret;
691 con->out_kvec_left = 0;
692 con->out_kvec_is_msg = false;
695 dout("write_partial_kvec %p %d left in %d kvecs ret = %d\n", con,
696 con->out_kvec_bytes, con->out_kvec_left, ret);
697 return ret; /* done! */
701 * Write as much message data payload as we can. If we finish, queue
703 * 1 -> done, footer is now queued in out_kvec[].
704 * 0 -> socket full, but more to do
707 static int write_partial_msg_pages(struct ceph_connection *con)
709 struct ceph_msg *msg = con->out_msg;
710 unsigned data_len = le32_to_cpu(msg->hdr.data_len);
712 int crc = con->msgr->nocrc;
715 dout("write_partial_msg_pages %p msg %p page %d/%d offset %d\n",
716 con, con->out_msg, con->out_msg_pos.page, con->out_msg->nr_pages,
717 con->out_msg_pos.page_pos);
719 while (con->out_msg_pos.page < con->out_msg->nr_pages) {
720 struct page *page = NULL;
724 * if we are calculating the data crc (the default), we need
725 * to map the page. if our pages[] has been revoked, use the
729 page = msg->pages[con->out_msg_pos.page];
732 } else if (msg->pagelist) {
733 page = list_first_entry(&msg->pagelist->head,
738 page = con->msgr->zero_page;
740 kaddr = page_address(con->msgr->zero_page);
742 len = min((int)(PAGE_SIZE - con->out_msg_pos.page_pos),
743 (int)(data_len - con->out_msg_pos.data_pos));
744 if (crc && !con->out_msg_pos.did_page_crc) {
745 void *base = kaddr + con->out_msg_pos.page_pos;
746 u32 tmpcrc = le32_to_cpu(con->out_msg->footer.data_crc);
748 BUG_ON(kaddr == NULL);
749 con->out_msg->footer.data_crc =
750 cpu_to_le32(crc32c(tmpcrc, base, len));
751 con->out_msg_pos.did_page_crc = 1;
754 ret = kernel_sendpage(con->sock, page,
755 con->out_msg_pos.page_pos, len,
756 MSG_DONTWAIT | MSG_NOSIGNAL |
759 if (crc && (msg->pages || msg->pagelist))
765 con->out_msg_pos.data_pos += ret;
766 con->out_msg_pos.page_pos += ret;
768 con->out_msg_pos.page_pos = 0;
769 con->out_msg_pos.page++;
770 con->out_msg_pos.did_page_crc = 0;
772 list_move_tail(&page->lru,
773 &msg->pagelist->head);
777 dout("write_partial_msg_pages %p msg %p done\n", con, msg);
779 /* prepare and queue up footer, too */
781 con->out_msg->footer.flags |= CEPH_MSG_FOOTER_NOCRC;
782 con->out_kvec_bytes = 0;
783 con->out_kvec_left = 0;
784 con->out_kvec_cur = con->out_kvec;
785 prepare_write_message_footer(con, 0);
794 static int write_partial_skip(struct ceph_connection *con)
798 while (con->out_skip > 0) {
800 .iov_base = page_address(con->msgr->zero_page),
801 .iov_len = min(con->out_skip, (int)PAGE_CACHE_SIZE)
804 ret = ceph_tcp_sendmsg(con->sock, &iov, 1, iov.iov_len, 1);
807 con->out_skip -= ret;
815 * Prepare to read connection handshake, or an ack.
817 static void prepare_read_banner(struct ceph_connection *con)
819 dout("prepare_read_banner %p\n", con);
820 con->in_base_pos = 0;
823 static void prepare_read_connect(struct ceph_connection *con)
825 dout("prepare_read_connect %p\n", con);
826 con->in_base_pos = 0;
829 static void prepare_read_connect_retry(struct ceph_connection *con)
831 dout("prepare_read_connect_retry %p\n", con);
832 con->in_base_pos = strlen(CEPH_BANNER) + sizeof(con->actual_peer_addr)
833 + sizeof(con->peer_addr_for_me);
836 static void prepare_read_ack(struct ceph_connection *con)
838 dout("prepare_read_ack %p\n", con);
839 con->in_base_pos = 0;
842 static void prepare_read_tag(struct ceph_connection *con)
844 dout("prepare_read_tag %p\n", con);
845 con->in_base_pos = 0;
846 con->in_tag = CEPH_MSGR_TAG_READY;
850 * Prepare to read a message.
852 static int prepare_read_message(struct ceph_connection *con)
854 dout("prepare_read_message %p\n", con);
855 BUG_ON(con->in_msg != NULL);
856 con->in_base_pos = 0;
857 con->in_front_crc = con->in_middle_crc = con->in_data_crc = 0;
862 static int read_partial(struct ceph_connection *con,
863 int *to, int size, void *object)
866 while (con->in_base_pos < *to) {
867 int left = *to - con->in_base_pos;
868 int have = size - left;
869 int ret = ceph_tcp_recvmsg(con->sock, object + have, left);
872 con->in_base_pos += ret;
879 * Read all or part of the connect-side handshake on a new connection
881 static int read_partial_banner(struct ceph_connection *con)
885 dout("read_partial_banner %p at %d\n", con, con->in_base_pos);
888 ret = read_partial(con, &to, strlen(CEPH_BANNER), con->in_banner);
891 ret = read_partial(con, &to, sizeof(con->actual_peer_addr),
892 &con->actual_peer_addr);
895 ret = read_partial(con, &to, sizeof(con->peer_addr_for_me),
896 &con->peer_addr_for_me);
903 static int read_partial_connect(struct ceph_connection *con)
907 dout("read_partial_connect %p at %d\n", con, con->in_base_pos);
909 ret = read_partial(con, &to, sizeof(con->in_reply), &con->in_reply);
912 ret = read_partial(con, &to, le32_to_cpu(con->in_reply.authorizer_len),
913 con->auth_reply_buf);
917 dout("read_partial_connect %p tag %d, con_seq = %u, g_seq = %u\n",
918 con, (int)con->in_reply.tag,
919 le32_to_cpu(con->in_reply.connect_seq),
920 le32_to_cpu(con->in_reply.global_seq));
927 * Verify the hello banner looks okay.
929 static int verify_hello(struct ceph_connection *con)
931 if (memcmp(con->in_banner, CEPH_BANNER, strlen(CEPH_BANNER))) {
932 pr_err("connect to %s got bad banner\n",
933 pr_addr(&con->peer_addr.in_addr));
934 con->error_msg = "protocol error, bad banner";
940 static bool addr_is_blank(struct sockaddr_storage *ss)
942 switch (ss->ss_family) {
944 return ((struct sockaddr_in *)ss)->sin_addr.s_addr == 0;
947 ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[0] == 0 &&
948 ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[1] == 0 &&
949 ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[2] == 0 &&
950 ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[3] == 0;
955 static int addr_port(struct sockaddr_storage *ss)
957 switch (ss->ss_family) {
959 return ntohs(((struct sockaddr_in *)ss)->sin_port);
961 return ntohs(((struct sockaddr_in6 *)ss)->sin6_port);
966 static void addr_set_port(struct sockaddr_storage *ss, int p)
968 switch (ss->ss_family) {
970 ((struct sockaddr_in *)ss)->sin_port = htons(p);
972 ((struct sockaddr_in6 *)ss)->sin6_port = htons(p);
977 * Parse an ip[:port] list into an addr array. Use the default
978 * monitor port if a port isn't specified.
980 int ceph_parse_ips(const char *c, const char *end,
981 struct ceph_entity_addr *addr,
982 int max_count, int *count)
987 dout("parse_ips on '%.*s'\n", (int)(end-c), c);
988 for (i = 0; i < max_count; i++) {
990 struct sockaddr_storage *ss = &addr[i].in_addr;
991 struct sockaddr_in *in4 = (void *)ss;
992 struct sockaddr_in6 *in6 = (void *)ss;
995 memset(ss, 0, sizeof(*ss));
996 if (in4_pton(p, end - p, (u8 *)&in4->sin_addr.s_addr,
998 ss->ss_family = AF_INET;
999 } else if (in6_pton(p, end - p, (u8 *)&in6->sin6_addr.s6_addr,
1001 ss->ss_family = AF_INET6;
1008 if (p < end && *p == ':') {
1011 while (p < end && *p >= '0' && *p <= '9') {
1012 port = (port * 10) + (*p - '0');
1015 if (port > 65535 || port == 0)
1018 port = CEPH_MON_PORT;
1021 addr_set_port(ss, port);
1023 dout("parse_ips got %s\n", pr_addr(ss));
1040 pr_err("parse_ips bad ip '%s'\n", c);
1044 static int process_banner(struct ceph_connection *con)
1046 dout("process_banner on %p\n", con);
1048 if (verify_hello(con) < 0)
1051 ceph_decode_addr(&con->actual_peer_addr);
1052 ceph_decode_addr(&con->peer_addr_for_me);
1055 * Make sure the other end is who we wanted. note that the other
1056 * end may not yet know their ip address, so if it's 0.0.0.0, give
1057 * them the benefit of the doubt.
1059 if (memcmp(&con->peer_addr, &con->actual_peer_addr,
1060 sizeof(con->peer_addr)) != 0 &&
1061 !(addr_is_blank(&con->actual_peer_addr.in_addr) &&
1062 con->actual_peer_addr.nonce == con->peer_addr.nonce)) {
1063 pr_warning("wrong peer, want %s/%lld, got %s/%lld\n",
1064 pr_addr(&con->peer_addr.in_addr),
1065 le64_to_cpu(con->peer_addr.nonce),
1066 pr_addr(&con->actual_peer_addr.in_addr),
1067 le64_to_cpu(con->actual_peer_addr.nonce));
1068 con->error_msg = "wrong peer at address";
1073 * did we learn our address?
1075 if (addr_is_blank(&con->msgr->inst.addr.in_addr)) {
1076 int port = addr_port(&con->msgr->inst.addr.in_addr);
1078 memcpy(&con->msgr->inst.addr.in_addr,
1079 &con->peer_addr_for_me.in_addr,
1080 sizeof(con->peer_addr_for_me.in_addr));
1081 addr_set_port(&con->msgr->inst.addr.in_addr, port);
1082 encode_my_addr(con->msgr);
1083 dout("process_banner learned my addr is %s\n",
1084 pr_addr(&con->msgr->inst.addr.in_addr));
1087 set_bit(NEGOTIATING, &con->state);
1088 prepare_read_connect(con);
1092 static void fail_protocol(struct ceph_connection *con)
1094 reset_connection(con);
1095 set_bit(CLOSED, &con->state); /* in case there's queued work */
1097 mutex_unlock(&con->mutex);
1098 if (con->ops->bad_proto)
1099 con->ops->bad_proto(con);
1100 mutex_lock(&con->mutex);
1103 static int process_connect(struct ceph_connection *con)
1105 u64 sup_feat = CEPH_FEATURE_SUPPORTED;
1106 u64 req_feat = CEPH_FEATURE_REQUIRED;
1107 u64 server_feat = le64_to_cpu(con->in_reply.features);
1109 dout("process_connect on %p tag %d\n", con, (int)con->in_tag);
1111 switch (con->in_reply.tag) {
1112 case CEPH_MSGR_TAG_FEATURES:
1113 pr_err("%s%lld %s feature set mismatch,"
1114 " my %llx < server's %llx, missing %llx\n",
1115 ENTITY_NAME(con->peer_name),
1116 pr_addr(&con->peer_addr.in_addr),
1117 sup_feat, server_feat, server_feat & ~sup_feat);
1118 con->error_msg = "missing required protocol features";
1122 case CEPH_MSGR_TAG_BADPROTOVER:
1123 pr_err("%s%lld %s protocol version mismatch,"
1124 " my %d != server's %d\n",
1125 ENTITY_NAME(con->peer_name),
1126 pr_addr(&con->peer_addr.in_addr),
1127 le32_to_cpu(con->out_connect.protocol_version),
1128 le32_to_cpu(con->in_reply.protocol_version));
1129 con->error_msg = "protocol version mismatch";
1133 case CEPH_MSGR_TAG_BADAUTHORIZER:
1135 dout("process_connect %p got BADAUTHORIZER attempt %d\n", con,
1137 if (con->auth_retry == 2) {
1138 con->error_msg = "connect authorization failure";
1139 reset_connection(con);
1140 set_bit(CLOSED, &con->state);
1143 con->auth_retry = 1;
1144 prepare_write_connect(con->msgr, con, 0);
1145 prepare_read_connect_retry(con);
1148 case CEPH_MSGR_TAG_RESETSESSION:
1150 * If we connected with a large connect_seq but the peer
1151 * has no record of a session with us (no connection, or
1152 * connect_seq == 0), they will send RESETSESION to indicate
1153 * that they must have reset their session, and may have
1156 dout("process_connect got RESET peer seq %u\n",
1157 le32_to_cpu(con->in_connect.connect_seq));
1158 pr_err("%s%lld %s connection reset\n",
1159 ENTITY_NAME(con->peer_name),
1160 pr_addr(&con->peer_addr.in_addr));
1161 reset_connection(con);
1162 prepare_write_connect(con->msgr, con, 0);
1163 prepare_read_connect(con);
1165 /* Tell ceph about it. */
1166 mutex_unlock(&con->mutex);
1167 pr_info("reset on %s%lld\n", ENTITY_NAME(con->peer_name));
1168 if (con->ops->peer_reset)
1169 con->ops->peer_reset(con);
1170 mutex_lock(&con->mutex);
1173 case CEPH_MSGR_TAG_RETRY_SESSION:
1175 * If we sent a smaller connect_seq than the peer has, try
1176 * again with a larger value.
1178 dout("process_connect got RETRY my seq = %u, peer_seq = %u\n",
1179 le32_to_cpu(con->out_connect.connect_seq),
1180 le32_to_cpu(con->in_connect.connect_seq));
1181 con->connect_seq = le32_to_cpu(con->in_connect.connect_seq);
1182 prepare_write_connect(con->msgr, con, 0);
1183 prepare_read_connect(con);
1186 case CEPH_MSGR_TAG_RETRY_GLOBAL:
1188 * If we sent a smaller global_seq than the peer has, try
1189 * again with a larger value.
1191 dout("process_connect got RETRY_GLOBAL my %u peer_gseq %u\n",
1192 con->peer_global_seq,
1193 le32_to_cpu(con->in_connect.global_seq));
1194 get_global_seq(con->msgr,
1195 le32_to_cpu(con->in_connect.global_seq));
1196 prepare_write_connect(con->msgr, con, 0);
1197 prepare_read_connect(con);
1200 case CEPH_MSGR_TAG_READY:
1201 if (req_feat & ~server_feat) {
1202 pr_err("%s%lld %s protocol feature mismatch,"
1203 " my required %llx > server's %llx, need %llx\n",
1204 ENTITY_NAME(con->peer_name),
1205 pr_addr(&con->peer_addr.in_addr),
1206 req_feat, server_feat, req_feat & ~server_feat);
1207 con->error_msg = "missing required protocol features";
1211 clear_bit(CONNECTING, &con->state);
1212 con->peer_global_seq = le32_to_cpu(con->in_reply.global_seq);
1214 dout("process_connect got READY gseq %d cseq %d (%d)\n",
1215 con->peer_global_seq,
1216 le32_to_cpu(con->in_reply.connect_seq),
1218 WARN_ON(con->connect_seq !=
1219 le32_to_cpu(con->in_reply.connect_seq));
1221 if (con->in_reply.flags & CEPH_MSG_CONNECT_LOSSY)
1222 set_bit(LOSSYTX, &con->state);
1224 prepare_read_tag(con);
1227 case CEPH_MSGR_TAG_WAIT:
1229 * If there is a connection race (we are opening
1230 * connections to each other), one of us may just have
1231 * to WAIT. This shouldn't happen if we are the
1234 pr_err("process_connect peer connecting WAIT\n");
1237 pr_err("connect protocol error, will retry\n");
1238 con->error_msg = "protocol error, garbage tag during connect";
1246 * read (part of) an ack
1248 static int read_partial_ack(struct ceph_connection *con)
1252 return read_partial(con, &to, sizeof(con->in_temp_ack),
1258 * We can finally discard anything that's been acked.
1260 static void process_ack(struct ceph_connection *con)
1263 u64 ack = le64_to_cpu(con->in_temp_ack);
1266 while (!list_empty(&con->out_sent)) {
1267 m = list_first_entry(&con->out_sent, struct ceph_msg,
1269 seq = le64_to_cpu(m->hdr.seq);
1272 dout("got ack for seq %llu type %d at %p\n", seq,
1273 le16_to_cpu(m->hdr.type), m);
1276 prepare_read_tag(con);
1282 static int read_partial_message_section(struct ceph_connection *con,
1283 struct kvec *section, unsigned int sec_len,
1291 while (section->iov_len < sec_len) {
1292 BUG_ON(section->iov_base == NULL);
1293 left = sec_len - section->iov_len;
1294 ret = ceph_tcp_recvmsg(con->sock, (char *)section->iov_base +
1295 section->iov_len, left);
1298 section->iov_len += ret;
1299 if (section->iov_len == sec_len)
1300 *crc = crc32c(0, section->iov_base,
1307 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
1308 struct ceph_msg_header *hdr,
1311 * read (part of) a message.
1313 static int read_partial_message(struct ceph_connection *con)
1315 struct ceph_msg *m = con->in_msg;
1319 unsigned front_len, middle_len, data_len, data_off;
1320 int datacrc = con->msgr->nocrc;
1323 dout("read_partial_message con %p msg %p\n", con, m);
1326 while (con->in_base_pos < sizeof(con->in_hdr)) {
1327 left = sizeof(con->in_hdr) - con->in_base_pos;
1328 ret = ceph_tcp_recvmsg(con->sock,
1329 (char *)&con->in_hdr + con->in_base_pos,
1333 con->in_base_pos += ret;
1334 if (con->in_base_pos == sizeof(con->in_hdr)) {
1335 u32 crc = crc32c(0, (void *)&con->in_hdr,
1336 sizeof(con->in_hdr) - sizeof(con->in_hdr.crc));
1337 if (crc != le32_to_cpu(con->in_hdr.crc)) {
1338 pr_err("read_partial_message bad hdr "
1339 " crc %u != expected %u\n",
1340 crc, con->in_hdr.crc);
1345 front_len = le32_to_cpu(con->in_hdr.front_len);
1346 if (front_len > CEPH_MSG_MAX_FRONT_LEN)
1348 middle_len = le32_to_cpu(con->in_hdr.middle_len);
1349 if (middle_len > CEPH_MSG_MAX_DATA_LEN)
1351 data_len = le32_to_cpu(con->in_hdr.data_len);
1352 if (data_len > CEPH_MSG_MAX_DATA_LEN)
1354 data_off = le16_to_cpu(con->in_hdr.data_off);
1356 /* allocate message? */
1358 dout("got hdr type %d front %d data %d\n", con->in_hdr.type,
1359 con->in_hdr.front_len, con->in_hdr.data_len);
1360 con->in_msg = ceph_alloc_msg(con, &con->in_hdr, &skip);
1362 /* skip this message */
1363 pr_err("alloc_msg returned NULL, skipping message\n");
1364 con->in_base_pos = -front_len - middle_len - data_len -
1366 con->in_tag = CEPH_MSGR_TAG_READY;
1369 if (IS_ERR(con->in_msg)) {
1370 ret = PTR_ERR(con->in_msg);
1372 con->error_msg = "error allocating memory for incoming message";
1376 m->front.iov_len = 0; /* haven't read it yet */
1378 m->middle->vec.iov_len = 0;
1380 con->in_msg_pos.page = 0;
1381 con->in_msg_pos.page_pos = data_off & ~PAGE_MASK;
1382 con->in_msg_pos.data_pos = 0;
1386 ret = read_partial_message_section(con, &m->front, front_len,
1387 &con->in_front_crc);
1393 ret = read_partial_message_section(con, &m->middle->vec, middle_len,
1394 &con->in_middle_crc);
1400 while (con->in_msg_pos.data_pos < data_len) {
1401 left = min((int)(data_len - con->in_msg_pos.data_pos),
1402 (int)(PAGE_SIZE - con->in_msg_pos.page_pos));
1403 BUG_ON(m->pages == NULL);
1404 p = kmap(m->pages[con->in_msg_pos.page]);
1405 ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1407 if (ret > 0 && datacrc)
1409 crc32c(con->in_data_crc,
1410 p + con->in_msg_pos.page_pos, ret);
1411 kunmap(m->pages[con->in_msg_pos.page]);
1414 con->in_msg_pos.data_pos += ret;
1415 con->in_msg_pos.page_pos += ret;
1416 if (con->in_msg_pos.page_pos == PAGE_SIZE) {
1417 con->in_msg_pos.page_pos = 0;
1418 con->in_msg_pos.page++;
1423 to = sizeof(m->hdr) + sizeof(m->footer);
1424 while (con->in_base_pos < to) {
1425 left = to - con->in_base_pos;
1426 ret = ceph_tcp_recvmsg(con->sock, (char *)&m->footer +
1427 (con->in_base_pos - sizeof(m->hdr)),
1431 con->in_base_pos += ret;
1433 dout("read_partial_message got msg %p %d (%u) + %d (%u) + %d (%u)\n",
1434 m, front_len, m->footer.front_crc, middle_len,
1435 m->footer.middle_crc, data_len, m->footer.data_crc);
1438 if (con->in_front_crc != le32_to_cpu(m->footer.front_crc)) {
1439 pr_err("read_partial_message %p front crc %u != exp. %u\n",
1440 m, con->in_front_crc, m->footer.front_crc);
1443 if (con->in_middle_crc != le32_to_cpu(m->footer.middle_crc)) {
1444 pr_err("read_partial_message %p middle crc %u != exp %u\n",
1445 m, con->in_middle_crc, m->footer.middle_crc);
1449 (m->footer.flags & CEPH_MSG_FOOTER_NOCRC) == 0 &&
1450 con->in_data_crc != le32_to_cpu(m->footer.data_crc)) {
1451 pr_err("read_partial_message %p data crc %u != exp. %u\n", m,
1452 con->in_data_crc, le32_to_cpu(m->footer.data_crc));
1456 return 1; /* done! */
1460 * Process message. This happens in the worker thread. The callback should
1461 * be careful not to do anything that waits on other incoming messages or it
1464 static void process_message(struct ceph_connection *con)
1466 struct ceph_msg *msg;
1471 /* if first message, set peer_name */
1472 if (con->peer_name.type == 0)
1473 con->peer_name = msg->hdr.src.name;
1476 mutex_unlock(&con->mutex);
1478 dout("===== %p %llu from %s%lld %d=%s len %d+%d (%u %u %u) =====\n",
1479 msg, le64_to_cpu(msg->hdr.seq),
1480 ENTITY_NAME(msg->hdr.src.name),
1481 le16_to_cpu(msg->hdr.type),
1482 ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1483 le32_to_cpu(msg->hdr.front_len),
1484 le32_to_cpu(msg->hdr.data_len),
1485 con->in_front_crc, con->in_middle_crc, con->in_data_crc);
1486 con->ops->dispatch(con, msg);
1488 mutex_lock(&con->mutex);
1489 prepare_read_tag(con);
1494 * Write something to the socket. Called in a worker thread when the
1495 * socket appears to be writeable and we have something ready to send.
1497 static int try_write(struct ceph_connection *con)
1499 struct ceph_messenger *msgr = con->msgr;
1502 dout("try_write start %p state %lu nref %d\n", con, con->state,
1503 atomic_read(&con->nref));
1505 mutex_lock(&con->mutex);
1507 dout("try_write out_kvec_bytes %d\n", con->out_kvec_bytes);
1509 /* open the socket first? */
1510 if (con->sock == NULL) {
1512 * if we were STANDBY and are reconnecting _this_
1513 * connection, bump connect_seq now. Always bump
1516 if (test_and_clear_bit(STANDBY, &con->state))
1519 prepare_write_banner(msgr, con);
1520 prepare_write_connect(msgr, con, 1);
1521 prepare_read_banner(con);
1522 set_bit(CONNECTING, &con->state);
1523 clear_bit(NEGOTIATING, &con->state);
1525 BUG_ON(con->in_msg);
1526 con->in_tag = CEPH_MSGR_TAG_READY;
1527 dout("try_write initiating connect on %p new state %lu\n",
1529 con->sock = ceph_tcp_connect(con);
1530 if (IS_ERR(con->sock)) {
1532 con->error_msg = "connect error";
1539 /* kvec data queued? */
1540 if (con->out_skip) {
1541 ret = write_partial_skip(con);
1545 dout("try_write write_partial_skip err %d\n", ret);
1549 if (con->out_kvec_left) {
1550 ret = write_partial_kvec(con);
1557 if (con->out_msg_done) {
1558 ceph_msg_put(con->out_msg);
1559 con->out_msg = NULL; /* we're done with this one */
1563 ret = write_partial_msg_pages(con);
1565 goto more_kvec; /* we need to send the footer, too! */
1569 dout("try_write write_partial_msg_pages err %d\n",
1576 if (!test_bit(CONNECTING, &con->state)) {
1577 /* is anything else pending? */
1578 if (!list_empty(&con->out_queue)) {
1579 prepare_write_message(con);
1582 if (con->in_seq > con->in_seq_acked) {
1583 prepare_write_ack(con);
1586 if (test_and_clear_bit(KEEPALIVE_PENDING, &con->state)) {
1587 prepare_write_keepalive(con);
1592 /* Nothing to do! */
1593 clear_bit(WRITE_PENDING, &con->state);
1594 dout("try_write nothing else to write.\n");
1598 mutex_unlock(&con->mutex);
1599 dout("try_write done on %p\n", con);
1606 * Read what we can from the socket.
1608 static int try_read(struct ceph_connection *con)
1610 struct ceph_messenger *msgr;
1616 if (test_bit(STANDBY, &con->state))
1619 dout("try_read start on %p\n", con);
1622 mutex_lock(&con->mutex);
1625 dout("try_read tag %d in_base_pos %d\n", (int)con->in_tag,
1627 if (test_bit(CONNECTING, &con->state)) {
1628 if (!test_bit(NEGOTIATING, &con->state)) {
1629 dout("try_read connecting\n");
1630 ret = read_partial_banner(con);
1633 if (process_banner(con) < 0) {
1638 ret = read_partial_connect(con);
1641 if (process_connect(con) < 0) {
1648 if (con->in_base_pos < 0) {
1650 * skipping + discarding content.
1652 * FIXME: there must be a better way to do this!
1654 static char buf[1024];
1655 int skip = min(1024, -con->in_base_pos);
1656 dout("skipping %d / %d bytes\n", skip, -con->in_base_pos);
1657 ret = ceph_tcp_recvmsg(con->sock, buf, skip);
1660 con->in_base_pos += ret;
1661 if (con->in_base_pos)
1664 if (con->in_tag == CEPH_MSGR_TAG_READY) {
1668 ret = ceph_tcp_recvmsg(con->sock, &con->in_tag, 1);
1671 dout("try_read got tag %d\n", (int)con->in_tag);
1672 switch (con->in_tag) {
1673 case CEPH_MSGR_TAG_MSG:
1674 prepare_read_message(con);
1676 case CEPH_MSGR_TAG_ACK:
1677 prepare_read_ack(con);
1679 case CEPH_MSGR_TAG_CLOSE:
1680 set_bit(CLOSED, &con->state); /* fixme */
1686 if (con->in_tag == CEPH_MSGR_TAG_MSG) {
1687 ret = read_partial_message(con);
1691 con->error_msg = "bad crc";
1695 con->error_msg = "io error";
1701 if (con->in_tag == CEPH_MSGR_TAG_READY)
1703 process_message(con);
1706 if (con->in_tag == CEPH_MSGR_TAG_ACK) {
1707 ret = read_partial_ack(con);
1717 mutex_unlock(&con->mutex);
1718 dout("try_read done on %p\n", con);
1722 pr_err("try_read bad con->in_tag = %d\n", (int)con->in_tag);
1723 con->error_msg = "protocol error, garbage tag";
1730 * Atomically queue work on a connection. Bump @con reference to
1731 * avoid races with connection teardown.
1733 * There is some trickery going on with QUEUED and BUSY because we
1734 * only want a _single_ thread operating on each connection at any
1735 * point in time, but we want to use all available CPUs.
1737 * The worker thread only proceeds if it can atomically set BUSY. It
1738 * clears QUEUED and does it's thing. When it thinks it's done, it
1739 * clears BUSY, then rechecks QUEUED.. if it's set again, it loops
1740 * (tries again to set BUSY).
1742 * To queue work, we first set QUEUED, _then_ if BUSY isn't set, we
1743 * try to queue work. If that fails (work is already queued, or BUSY)
1744 * we give up (work also already being done or is queued) but leave QUEUED
1745 * set so that the worker thread will loop if necessary.
1747 static void queue_con(struct ceph_connection *con)
1749 if (test_bit(DEAD, &con->state)) {
1750 dout("queue_con %p ignoring: DEAD\n",
1755 if (!con->ops->get(con)) {
1756 dout("queue_con %p ref count 0\n", con);
1760 set_bit(QUEUED, &con->state);
1761 if (test_bit(BUSY, &con->state)) {
1762 dout("queue_con %p - already BUSY\n", con);
1764 } else if (!queue_work(ceph_msgr_wq, &con->work.work)) {
1765 dout("queue_con %p - already queued\n", con);
1768 dout("queue_con %p\n", con);
1773 * Do some work on a connection. Drop a connection ref when we're done.
1775 static void con_work(struct work_struct *work)
1777 struct ceph_connection *con = container_of(work, struct ceph_connection,
1782 if (test_and_set_bit(BUSY, &con->state) != 0) {
1783 dout("con_work %p BUSY already set\n", con);
1786 dout("con_work %p start, clearing QUEUED\n", con);
1787 clear_bit(QUEUED, &con->state);
1789 if (test_bit(CLOSED, &con->state)) { /* e.g. if we are replaced */
1790 dout("con_work CLOSED\n");
1791 con_close_socket(con);
1794 if (test_and_clear_bit(OPENING, &con->state)) {
1795 /* reopen w/ new peer */
1796 dout("con_work OPENING\n");
1797 con_close_socket(con);
1800 if (test_and_clear_bit(SOCK_CLOSED, &con->state) ||
1801 try_read(con) < 0 ||
1802 try_write(con) < 0) {
1804 ceph_fault(con); /* error/fault path */
1808 clear_bit(BUSY, &con->state);
1809 dout("con->state=%lu\n", con->state);
1810 if (test_bit(QUEUED, &con->state)) {
1812 dout("con_work %p QUEUED reset, looping\n", con);
1815 dout("con_work %p QUEUED reset, but just faulted\n", con);
1816 clear_bit(QUEUED, &con->state);
1818 dout("con_work %p done\n", con);
1826 * Generic error/fault handler. A retry mechanism is used with
1827 * exponential backoff
1829 static void ceph_fault(struct ceph_connection *con)
1831 pr_err("%s%lld %s %s\n", ENTITY_NAME(con->peer_name),
1832 pr_addr(&con->peer_addr.in_addr), con->error_msg);
1833 dout("fault %p state %lu to peer %s\n",
1834 con, con->state, pr_addr(&con->peer_addr.in_addr));
1836 if (test_bit(LOSSYTX, &con->state)) {
1837 dout("fault on LOSSYTX channel\n");
1841 clear_bit(BUSY, &con->state); /* to avoid an improbable race */
1843 mutex_lock(&con->mutex);
1845 con_close_socket(con);
1848 ceph_msg_put(con->in_msg);
1852 /* If there are no messages in the queue, place the connection
1853 * in a STANDBY state (i.e., don't try to reconnect just yet). */
1854 if (list_empty(&con->out_queue) && !con->out_keepalive_pending) {
1855 dout("fault setting STANDBY\n");
1856 set_bit(STANDBY, &con->state);
1857 mutex_unlock(&con->mutex);
1861 /* Requeue anything that hasn't been acked, and retry after a
1863 list_splice_init(&con->out_sent, &con->out_queue);
1865 if (con->delay == 0)
1866 con->delay = BASE_DELAY_INTERVAL;
1867 else if (con->delay < MAX_DELAY_INTERVAL)
1870 mutex_unlock(&con->mutex);
1872 /* explicitly schedule work to try to reconnect again later. */
1873 dout("fault queueing %p delay %lu\n", con, con->delay);
1875 if (queue_delayed_work(ceph_msgr_wq, &con->work,
1876 round_jiffies_relative(con->delay)) == 0)
1880 if (con->ops->fault)
1881 con->ops->fault(con);
1887 * create a new messenger instance
1889 struct ceph_messenger *ceph_messenger_create(struct ceph_entity_addr *myaddr)
1891 struct ceph_messenger *msgr;
1893 msgr = kzalloc(sizeof(*msgr), GFP_KERNEL);
1895 return ERR_PTR(-ENOMEM);
1897 spin_lock_init(&msgr->global_seq_lock);
1899 /* the zero page is needed if a request is "canceled" while the message
1900 * is being written over the socket */
1901 msgr->zero_page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1902 if (!msgr->zero_page) {
1904 return ERR_PTR(-ENOMEM);
1906 kmap(msgr->zero_page);
1909 msgr->inst.addr = *myaddr;
1911 /* select a random nonce */
1912 msgr->inst.addr.type = 0;
1913 get_random_bytes(&msgr->inst.addr.nonce, sizeof(msgr->inst.addr.nonce));
1914 encode_my_addr(msgr);
1916 dout("messenger_create %p\n", msgr);
1920 void ceph_messenger_destroy(struct ceph_messenger *msgr)
1922 dout("destroy %p\n", msgr);
1923 kunmap(msgr->zero_page);
1924 __free_page(msgr->zero_page);
1926 dout("destroyed messenger %p\n", msgr);
1930 * Queue up an outgoing message on the given connection.
1932 void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg)
1934 if (test_bit(CLOSED, &con->state)) {
1935 dout("con_send %p closed, dropping %p\n", con, msg);
1941 msg->hdr.src.name = con->msgr->inst.name;
1942 msg->hdr.src.addr = con->msgr->my_enc_addr;
1943 msg->hdr.orig_src = msg->hdr.src;
1946 mutex_lock(&con->mutex);
1947 BUG_ON(!list_empty(&msg->list_head));
1948 list_add_tail(&msg->list_head, &con->out_queue);
1949 dout("----- %p to %s%lld %d=%s len %d+%d+%d -----\n", msg,
1950 ENTITY_NAME(con->peer_name), le16_to_cpu(msg->hdr.type),
1951 ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1952 le32_to_cpu(msg->hdr.front_len),
1953 le32_to_cpu(msg->hdr.middle_len),
1954 le32_to_cpu(msg->hdr.data_len));
1955 mutex_unlock(&con->mutex);
1957 /* if there wasn't anything waiting to send before, queue
1959 if (test_and_set_bit(WRITE_PENDING, &con->state) == 0)
1964 * Revoke a message that was previously queued for send
1966 void ceph_con_revoke(struct ceph_connection *con, struct ceph_msg *msg)
1968 mutex_lock(&con->mutex);
1969 if (!list_empty(&msg->list_head)) {
1970 dout("con_revoke %p msg %p\n", con, msg);
1971 list_del_init(&msg->list_head);
1974 if (con->out_msg == msg) {
1975 ceph_msg_put(con->out_msg);
1976 con->out_msg = NULL;
1978 if (con->out_kvec_is_msg) {
1979 con->out_skip = con->out_kvec_bytes;
1980 con->out_kvec_is_msg = false;
1983 dout("con_revoke %p msg %p - not queued (sent?)\n", con, msg);
1985 mutex_unlock(&con->mutex);
1989 * Revoke a message that we may be reading data into
1991 void ceph_con_revoke_message(struct ceph_connection *con, struct ceph_msg *msg)
1993 mutex_lock(&con->mutex);
1994 if (con->in_msg && con->in_msg == msg) {
1995 unsigned front_len = le32_to_cpu(con->in_hdr.front_len);
1996 unsigned middle_len = le32_to_cpu(con->in_hdr.middle_len);
1997 unsigned data_len = le32_to_cpu(con->in_hdr.data_len);
1999 /* skip rest of message */
2000 dout("con_revoke_pages %p msg %p revoked\n", con, msg);
2001 con->in_base_pos = con->in_base_pos -
2002 sizeof(struct ceph_msg_header) -
2006 sizeof(struct ceph_msg_footer);
2007 ceph_msg_put(con->in_msg);
2009 con->in_tag = CEPH_MSGR_TAG_READY;
2011 dout("con_revoke_pages %p msg %p pages %p no-op\n",
2012 con, con->in_msg, msg);
2014 mutex_unlock(&con->mutex);
2018 * Queue a keepalive byte to ensure the tcp connection is alive.
2020 void ceph_con_keepalive(struct ceph_connection *con)
2022 if (test_and_set_bit(KEEPALIVE_PENDING, &con->state) == 0 &&
2023 test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2029 * construct a new message with given type, size
2030 * the new msg has a ref count of 1.
2032 struct ceph_msg *ceph_msg_new(int type, int front_len,
2033 int page_len, int page_off, struct page **pages)
2037 m = kmalloc(sizeof(*m), GFP_NOFS);
2040 kref_init(&m->kref);
2041 INIT_LIST_HEAD(&m->list_head);
2043 m->hdr.type = cpu_to_le16(type);
2044 m->hdr.front_len = cpu_to_le32(front_len);
2045 m->hdr.middle_len = 0;
2046 m->hdr.data_len = cpu_to_le32(page_len);
2047 m->hdr.data_off = cpu_to_le16(page_off);
2048 m->hdr.priority = cpu_to_le16(CEPH_MSG_PRIO_DEFAULT);
2049 m->footer.front_crc = 0;
2050 m->footer.middle_crc = 0;
2051 m->footer.data_crc = 0;
2052 m->front_max = front_len;
2053 m->front_is_vmalloc = false;
2054 m->more_to_follow = false;
2059 if (front_len > PAGE_CACHE_SIZE) {
2060 m->front.iov_base = __vmalloc(front_len, GFP_NOFS,
2062 m->front_is_vmalloc = true;
2064 m->front.iov_base = kmalloc(front_len, GFP_NOFS);
2066 if (m->front.iov_base == NULL) {
2067 pr_err("msg_new can't allocate %d bytes\n",
2072 m->front.iov_base = NULL;
2074 m->front.iov_len = front_len;
2080 m->nr_pages = calc_pages_for(page_off, page_len);
2084 dout("ceph_msg_new %p page %d~%d -> %d\n", m, page_off, page_len,
2091 pr_err("msg_new can't create type %d len %d\n", type, front_len);
2092 return ERR_PTR(-ENOMEM);
2096 * Allocate "middle" portion of a message, if it is needed and wasn't
2097 * allocated by alloc_msg. This allows us to read a small fixed-size
2098 * per-type header in the front and then gracefully fail (i.e.,
2099 * propagate the error to the caller based on info in the front) when
2100 * the middle is too large.
2102 static int ceph_alloc_middle(struct ceph_connection *con, struct ceph_msg *msg)
2104 int type = le16_to_cpu(msg->hdr.type);
2105 int middle_len = le32_to_cpu(msg->hdr.middle_len);
2107 dout("alloc_middle %p type %d %s middle_len %d\n", msg, type,
2108 ceph_msg_type_name(type), middle_len);
2109 BUG_ON(!middle_len);
2110 BUG_ON(msg->middle);
2112 msg->middle = ceph_buffer_new(middle_len, GFP_NOFS);
2119 * Generic message allocator, for incoming messages.
2121 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
2122 struct ceph_msg_header *hdr,
2125 int type = le16_to_cpu(hdr->type);
2126 int front_len = le32_to_cpu(hdr->front_len);
2127 int middle_len = le32_to_cpu(hdr->middle_len);
2128 struct ceph_msg *msg = NULL;
2131 if (con->ops->alloc_msg) {
2132 mutex_unlock(&con->mutex);
2133 msg = con->ops->alloc_msg(con, hdr, skip);
2134 mutex_lock(&con->mutex);
2143 msg = ceph_msg_new(type, front_len, 0, 0, NULL);
2145 pr_err("unable to allocate msg type %d len %d\n",
2147 return ERR_PTR(-ENOMEM);
2150 memcpy(&msg->hdr, &con->in_hdr, sizeof(con->in_hdr));
2153 ret = ceph_alloc_middle(con, msg);
2166 * Free a generically kmalloc'd message.
2168 void ceph_msg_kfree(struct ceph_msg *m)
2170 dout("msg_kfree %p\n", m);
2171 if (m->front_is_vmalloc)
2172 vfree(m->front.iov_base);
2174 kfree(m->front.iov_base);
2179 * Drop a msg ref. Destroy as needed.
2181 void ceph_msg_last_put(struct kref *kref)
2183 struct ceph_msg *m = container_of(kref, struct ceph_msg, kref);
2185 dout("ceph_msg_put last one on %p\n", m);
2186 WARN_ON(!list_empty(&m->list_head));
2188 /* drop middle, data, if any */
2190 ceph_buffer_put(m->middle);
2197 ceph_pagelist_release(m->pagelist);
2203 ceph_msgpool_put(m->pool, m);
2208 void ceph_msg_dump(struct ceph_msg *msg)
2210 pr_debug("msg_dump %p (front_max %d nr_pages %d)\n", msg,
2211 msg->front_max, msg->nr_pages);
2212 print_hex_dump(KERN_DEBUG, "header: ",
2213 DUMP_PREFIX_OFFSET, 16, 1,
2214 &msg->hdr, sizeof(msg->hdr), true);
2215 print_hex_dump(KERN_DEBUG, " front: ",
2216 DUMP_PREFIX_OFFSET, 16, 1,
2217 msg->front.iov_base, msg->front.iov_len, true);
2219 print_hex_dump(KERN_DEBUG, "middle: ",
2220 DUMP_PREFIX_OFFSET, 16, 1,
2221 msg->middle->vec.iov_base,
2222 msg->middle->vec.iov_len, true);
2223 print_hex_dump(KERN_DEBUG, "footer: ",
2224 DUMP_PREFIX_OFFSET, 16, 1,
2225 &msg->footer, sizeof(msg->footer), true);