2 * linux/net/sunrpc/xprtsock.c
4 * Client-side transport implementation for sockets.
6 * TCP callback races fixes (C) 1998 Red Hat Software <alan@redhat.com>
7 * TCP send fixes (C) 1998 Red Hat Software <alan@redhat.com>
8 * TCP NFS related read + write fixes
9 * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
11 * Rewrite of larges part of the code in order to stabilize TCP stuff.
12 * Fix behaviour when socket buffer is full.
13 * (C) 1999 Trond Myklebust <trond.myklebust@fys.uio.no>
15 * IP socket transport implementation, (C) 2005 Chuck Lever <cel@netapp.com>
18 #include <linux/types.h>
19 #include <linux/slab.h>
20 #include <linux/capability.h>
21 #include <linux/sched.h>
22 #include <linux/pagemap.h>
23 #include <linux/errno.h>
24 #include <linux/socket.h>
26 #include <linux/net.h>
28 #include <linux/udp.h>
29 #include <linux/tcp.h>
30 #include <linux/sunrpc/clnt.h>
31 #include <linux/sunrpc/sched.h>
32 #include <linux/file.h>
35 #include <net/checksum.h>
42 unsigned int xprt_udp_slot_table_entries = RPC_DEF_SLOT_TABLE;
43 unsigned int xprt_tcp_slot_table_entries = RPC_DEF_SLOT_TABLE;
45 unsigned int xprt_min_resvport = RPC_DEF_MIN_RESVPORT;
46 unsigned int xprt_max_resvport = RPC_DEF_MAX_RESVPORT;
49 * How many times to try sending a request on a socket before waiting
50 * for the socket buffer to clear.
52 #define XS_SENDMSG_RETRY (10U)
55 * Time out for an RPC UDP socket connect. UDP socket connects are
56 * synchronous, but we set a timeout anyway in case of resource
57 * exhaustion on the local host.
59 #define XS_UDP_CONN_TO (5U * HZ)
62 * Wait duration for an RPC TCP connection to be established. Solaris
63 * NFS over TCP uses 60 seconds, for example, which is in line with how
64 * long a server takes to reboot.
66 #define XS_TCP_CONN_TO (60U * HZ)
69 * Wait duration for a reply from the RPC portmapper.
71 #define XS_BIND_TO (60U * HZ)
74 * Delay if a UDP socket connect error occurs. This is most likely some
75 * kind of resource problem on the local host.
77 #define XS_UDP_REEST_TO (2U * HZ)
80 * The reestablish timeout allows clients to delay for a bit before attempting
81 * to reconnect to a server that just dropped our connection.
83 * We implement an exponential backoff when trying to reestablish a TCP
84 * transport connection with the server. Some servers like to drop a TCP
85 * connection when they are overworked, so we start with a short timeout and
86 * increase over time if the server is down or not responding.
88 #define XS_TCP_INIT_REEST_TO (3U * HZ)
89 #define XS_TCP_MAX_REEST_TO (5U * 60 * HZ)
92 * TCP idle timeout; client drops the transport socket if it is idle
93 * for this long. Note that we also timeout UDP sockets to prevent
94 * holding port numbers when there is no RPC traffic.
96 #define XS_IDLE_DISC_TO (5U * 60 * HZ)
99 # undef RPC_DEBUG_DATA
100 # define RPCDBG_FACILITY RPCDBG_TRANS
103 #ifdef RPC_DEBUG_DATA
104 static void xs_pktdump(char *msg, u32 *packet, unsigned int count)
106 u8 *buf = (u8 *) packet;
109 dprintk("RPC: %s\n", msg);
110 for (j = 0; j < count && j < 128; j += 4) {
114 dprintk("0x%04x ", j);
116 dprintk("%02x%02x%02x%02x ",
117 buf[j], buf[j+1], buf[j+2], buf[j+3]);
122 static inline void xs_pktdump(char *msg, u32 *packet, unsigned int count)
128 static void xs_format_peer_addresses(struct rpc_xprt *xprt)
130 struct sockaddr_in *addr = (struct sockaddr_in *) &xprt->addr;
133 buf = kzalloc(20, GFP_KERNEL);
135 snprintf(buf, 20, "%u.%u.%u.%u",
136 NIPQUAD(addr->sin_addr.s_addr));
138 xprt->address_strings[RPC_DISPLAY_ADDR] = buf;
140 buf = kzalloc(8, GFP_KERNEL);
142 snprintf(buf, 8, "%u",
143 ntohs(addr->sin_port));
145 xprt->address_strings[RPC_DISPLAY_PORT] = buf;
147 if (xprt->prot == IPPROTO_UDP)
148 xprt->address_strings[RPC_DISPLAY_PROTO] = "udp";
150 xprt->address_strings[RPC_DISPLAY_PROTO] = "tcp";
152 buf = kzalloc(48, GFP_KERNEL);
154 snprintf(buf, 48, "addr=%u.%u.%u.%u port=%u proto=%s",
155 NIPQUAD(addr->sin_addr.s_addr),
156 ntohs(addr->sin_port),
157 xprt->prot == IPPROTO_UDP ? "udp" : "tcp");
159 xprt->address_strings[RPC_DISPLAY_ALL] = buf;
162 static void xs_free_peer_addresses(struct rpc_xprt *xprt)
164 kfree(xprt->address_strings[RPC_DISPLAY_ADDR]);
165 kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
166 kfree(xprt->address_strings[RPC_DISPLAY_ALL]);
169 #define XS_SENDMSG_FLAGS (MSG_DONTWAIT | MSG_NOSIGNAL)
171 static inline int xs_send_head(struct socket *sock, struct sockaddr *addr, int addrlen, struct xdr_buf *xdr, unsigned int base, unsigned int len)
174 .iov_base = xdr->head[0].iov_base + base,
175 .iov_len = len - base,
177 struct msghdr msg = {
179 .msg_namelen = addrlen,
180 .msg_flags = XS_SENDMSG_FLAGS,
184 msg.msg_flags |= MSG_MORE;
186 if (likely(iov.iov_len))
187 return kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
188 return kernel_sendmsg(sock, &msg, NULL, 0, 0);
191 static int xs_send_tail(struct socket *sock, struct xdr_buf *xdr, unsigned int base, unsigned int len)
194 .iov_base = xdr->tail[0].iov_base + base,
195 .iov_len = len - base,
197 struct msghdr msg = {
198 .msg_flags = XS_SENDMSG_FLAGS,
201 return kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
205 * xs_sendpages - write pages directly to a socket
206 * @sock: socket to send on
207 * @addr: UDP only -- address of destination
208 * @addrlen: UDP only -- length of destination address
209 * @xdr: buffer containing this request
210 * @base: starting position in the buffer
213 static inline int xs_sendpages(struct socket *sock, struct sockaddr *addr, int addrlen, struct xdr_buf *xdr, unsigned int base)
215 struct page **ppage = xdr->pages;
216 unsigned int len, pglen = xdr->page_len;
218 ssize_t (*sendpage)(struct socket *, struct page *, int, size_t, int);
223 clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
225 len = xdr->head[0].iov_len;
226 if (base < len || (addr != NULL && base == 0)) {
227 err = xs_send_head(sock, addr, addrlen, xdr, base, len);
232 if (err != (len - base))
238 if (unlikely(pglen == 0))
240 if (unlikely(base >= pglen)) {
244 if (base || xdr->page_base) {
246 base += xdr->page_base;
247 ppage += base >> PAGE_CACHE_SHIFT;
248 base &= ~PAGE_CACHE_MASK;
251 sendpage = sock->ops->sendpage ? : sock_no_sendpage;
253 int flags = XS_SENDMSG_FLAGS;
255 len = PAGE_CACHE_SIZE;
261 if (pglen != len || xdr->tail[0].iov_len != 0)
264 /* Hmm... We might be dealing with highmem pages */
265 if (PageHighMem(*ppage))
266 sendpage = sock_no_sendpage;
267 err = sendpage(sock, *ppage, base, len, flags);
276 } while ((pglen -= len) != 0);
278 len = xdr->tail[0].iov_len;
280 err = xs_send_tail(sock, xdr, base, len);
291 * xs_nospace - place task on wait queue if transmit was incomplete
292 * @task: task to put to sleep
295 static void xs_nospace(struct rpc_task *task)
297 struct rpc_rqst *req = task->tk_rqstp;
298 struct rpc_xprt *xprt = req->rq_xprt;
300 dprintk("RPC: %4d xmit incomplete (%u left of %u)\n",
301 task->tk_pid, req->rq_slen - req->rq_bytes_sent,
304 if (test_bit(SOCK_ASYNC_NOSPACE, &xprt->sock->flags)) {
305 /* Protect against races with write_space */
306 spin_lock_bh(&xprt->transport_lock);
308 /* Don't race with disconnect */
309 if (!xprt_connected(xprt))
310 task->tk_status = -ENOTCONN;
311 else if (test_bit(SOCK_NOSPACE, &xprt->sock->flags))
312 xprt_wait_for_buffer_space(task);
314 spin_unlock_bh(&xprt->transport_lock);
316 /* Keep holding the socket if it is blocked */
317 rpc_delay(task, HZ>>4);
321 * xs_udp_send_request - write an RPC request to a UDP socket
322 * @task: address of RPC task that manages the state of an RPC request
325 * 0: The request has been sent
326 * EAGAIN: The socket was blocked, please call again later to
327 * complete the request
328 * ENOTCONN: Caller needs to invoke connect logic then call again
329 * other: Some other error occured, the request was not sent
331 static int xs_udp_send_request(struct rpc_task *task)
333 struct rpc_rqst *req = task->tk_rqstp;
334 struct rpc_xprt *xprt = req->rq_xprt;
335 struct xdr_buf *xdr = &req->rq_snd_buf;
338 xs_pktdump("packet data:",
339 req->rq_svec->iov_base,
340 req->rq_svec->iov_len);
342 req->rq_xtime = jiffies;
343 status = xs_sendpages(xprt->sock, (struct sockaddr *) &xprt->addr,
344 xprt->addrlen, xdr, req->rq_bytes_sent);
346 dprintk("RPC: xs_udp_send_request(%u) = %d\n",
347 xdr->len - req->rq_bytes_sent, status);
349 if (likely(status >= (int) req->rq_slen))
352 /* Still some bytes left; set up for a retry later. */
360 /* When the server has died, an ICMP port unreachable message
361 * prompts ECONNREFUSED. */
367 dprintk("RPC: sendmsg returned unrecognized error %d\n",
375 static inline void xs_encode_tcp_record_marker(struct xdr_buf *buf)
377 u32 reclen = buf->len - sizeof(rpc_fraghdr);
378 rpc_fraghdr *base = buf->head[0].iov_base;
379 *base = htonl(RPC_LAST_STREAM_FRAGMENT | reclen);
383 * xs_tcp_send_request - write an RPC request to a TCP socket
384 * @task: address of RPC task that manages the state of an RPC request
387 * 0: The request has been sent
388 * EAGAIN: The socket was blocked, please call again later to
389 * complete the request
390 * ENOTCONN: Caller needs to invoke connect logic then call again
391 * other: Some other error occured, the request was not sent
393 * XXX: In the case of soft timeouts, should we eventually give up
394 * if sendmsg is not able to make progress?
396 static int xs_tcp_send_request(struct rpc_task *task)
398 struct rpc_rqst *req = task->tk_rqstp;
399 struct rpc_xprt *xprt = req->rq_xprt;
400 struct xdr_buf *xdr = &req->rq_snd_buf;
401 int status, retry = 0;
403 xs_encode_tcp_record_marker(&req->rq_snd_buf);
405 xs_pktdump("packet data:",
406 req->rq_svec->iov_base,
407 req->rq_svec->iov_len);
409 /* Continue transmitting the packet/record. We must be careful
410 * to cope with writespace callbacks arriving _after_ we have
411 * called sendmsg(). */
413 req->rq_xtime = jiffies;
414 status = xs_sendpages(xprt->sock, NULL, 0, xdr,
417 dprintk("RPC: xs_tcp_send_request(%u) = %d\n",
418 xdr->len - req->rq_bytes_sent, status);
420 if (unlikely(status < 0))
423 /* If we've sent the entire packet, immediately
424 * reset the count of bytes sent. */
425 req->rq_bytes_sent += status;
426 task->tk_bytes_sent += status;
427 if (likely(req->rq_bytes_sent >= req->rq_slen)) {
428 req->rq_bytes_sent = 0;
433 if (retry++ > XS_SENDMSG_RETRY)
448 dprintk("RPC: sendmsg returned unrecognized error %d\n",
450 xprt_disconnect(xprt);
458 * xs_tcp_release_xprt - clean up after a tcp transmission
462 * This cleans up if an error causes us to abort the transmission of a request.
463 * In this case, the socket may need to be reset in order to avoid confusing
466 static void xs_tcp_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
468 struct rpc_rqst *req;
470 if (task != xprt->snd_task)
474 req = task->tk_rqstp;
475 if (req->rq_bytes_sent == 0)
477 if (req->rq_bytes_sent == req->rq_snd_buf.len)
479 set_bit(XPRT_CLOSE_WAIT, &task->tk_xprt->state);
481 xprt_release_xprt(xprt, task);
485 * xs_close - close a socket
488 * This is used when all requests are complete; ie, no DRC state remains
489 * on the server we want to save.
491 static void xs_close(struct rpc_xprt *xprt)
493 struct socket *sock = xprt->sock;
494 struct sock *sk = xprt->inet;
497 goto clear_close_wait;
499 dprintk("RPC: xs_close xprt %p\n", xprt);
501 write_lock_bh(&sk->sk_callback_lock);
505 sk->sk_user_data = NULL;
506 sk->sk_data_ready = xprt->old_data_ready;
507 sk->sk_state_change = xprt->old_state_change;
508 sk->sk_write_space = xprt->old_write_space;
509 write_unlock_bh(&sk->sk_callback_lock);
515 smp_mb__before_clear_bit();
516 clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
517 smp_mb__after_clear_bit();
521 * xs_destroy - prepare to shutdown a transport
522 * @xprt: doomed transport
525 static void xs_destroy(struct rpc_xprt *xprt)
527 dprintk("RPC: xs_destroy xprt %p\n", xprt);
529 cancel_delayed_work(&xprt->connect_worker);
530 flush_scheduled_work();
532 xprt_disconnect(xprt);
534 xs_free_peer_addresses(xprt);
538 static inline struct rpc_xprt *xprt_from_sock(struct sock *sk)
540 return (struct rpc_xprt *) sk->sk_user_data;
544 * xs_udp_data_ready - "data ready" callback for UDP sockets
545 * @sk: socket with data to read
546 * @len: how much data to read
549 static void xs_udp_data_ready(struct sock *sk, int len)
551 struct rpc_task *task;
552 struct rpc_xprt *xprt;
553 struct rpc_rqst *rovr;
555 int err, repsize, copied;
558 read_lock(&sk->sk_callback_lock);
559 dprintk("RPC: xs_udp_data_ready...\n");
560 if (!(xprt = xprt_from_sock(sk)))
563 if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
569 repsize = skb->len - sizeof(struct udphdr);
571 dprintk("RPC: impossible RPC reply size %d!\n", repsize);
575 /* Copy the XID from the skb... */
576 xp = skb_header_pointer(skb, sizeof(struct udphdr),
577 sizeof(_xid), &_xid);
581 /* Look up and lock the request corresponding to the given XID */
582 spin_lock(&xprt->transport_lock);
583 rovr = xprt_lookup_rqst(xprt, *xp);
586 task = rovr->rq_task;
588 if ((copied = rovr->rq_private_buf.buflen) > repsize)
591 /* Suck it into the iovec, verify checksum if not done by hw. */
592 if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
595 /* Something worked... */
596 dst_confirm(skb->dst);
598 xprt_adjust_cwnd(task, copied);
599 xprt_update_rtt(task);
600 xprt_complete_rqst(task, copied);
603 spin_unlock(&xprt->transport_lock);
605 skb_free_datagram(sk, skb);
607 read_unlock(&sk->sk_callback_lock);
610 static inline size_t xs_tcp_copy_data(skb_reader_t *desc, void *p, size_t len)
612 if (len > desc->count)
614 if (skb_copy_bits(desc->skb, desc->offset, p, len)) {
615 dprintk("RPC: failed to copy %zu bytes from skb. %zu bytes remain\n",
621 dprintk("RPC: copied %zu bytes from skb. %zu bytes remain\n",
626 static inline void xs_tcp_read_fraghdr(struct rpc_xprt *xprt, skb_reader_t *desc)
631 p = ((char *) &xprt->tcp_recm) + xprt->tcp_offset;
632 len = sizeof(xprt->tcp_recm) - xprt->tcp_offset;
633 used = xs_tcp_copy_data(desc, p, len);
634 xprt->tcp_offset += used;
638 xprt->tcp_reclen = ntohl(xprt->tcp_recm);
639 if (xprt->tcp_reclen & RPC_LAST_STREAM_FRAGMENT)
640 xprt->tcp_flags |= XPRT_LAST_FRAG;
642 xprt->tcp_flags &= ~XPRT_LAST_FRAG;
643 xprt->tcp_reclen &= RPC_FRAGMENT_SIZE_MASK;
645 xprt->tcp_flags &= ~XPRT_COPY_RECM;
646 xprt->tcp_offset = 0;
648 /* Sanity check of the record length */
649 if (unlikely(xprt->tcp_reclen < 4)) {
650 dprintk("RPC: invalid TCP record fragment length\n");
651 xprt_disconnect(xprt);
654 dprintk("RPC: reading TCP record fragment of length %d\n",
658 static void xs_tcp_check_recm(struct rpc_xprt *xprt)
660 dprintk("RPC: xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u, tcp_flags = %lx\n",
661 xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen, xprt->tcp_flags);
662 if (xprt->tcp_offset == xprt->tcp_reclen) {
663 xprt->tcp_flags |= XPRT_COPY_RECM;
664 xprt->tcp_offset = 0;
665 if (xprt->tcp_flags & XPRT_LAST_FRAG) {
666 xprt->tcp_flags &= ~XPRT_COPY_DATA;
667 xprt->tcp_flags |= XPRT_COPY_XID;
668 xprt->tcp_copied = 0;
673 static inline void xs_tcp_read_xid(struct rpc_xprt *xprt, skb_reader_t *desc)
678 len = sizeof(xprt->tcp_xid) - xprt->tcp_offset;
679 dprintk("RPC: reading XID (%Zu bytes)\n", len);
680 p = ((char *) &xprt->tcp_xid) + xprt->tcp_offset;
681 used = xs_tcp_copy_data(desc, p, len);
682 xprt->tcp_offset += used;
685 xprt->tcp_flags &= ~XPRT_COPY_XID;
686 xprt->tcp_flags |= XPRT_COPY_DATA;
687 xprt->tcp_copied = 4;
688 dprintk("RPC: reading reply for XID %08x\n",
689 ntohl(xprt->tcp_xid));
690 xs_tcp_check_recm(xprt);
693 static inline void xs_tcp_read_request(struct rpc_xprt *xprt, skb_reader_t *desc)
695 struct rpc_rqst *req;
696 struct xdr_buf *rcvbuf;
700 /* Find and lock the request corresponding to this xid */
701 spin_lock(&xprt->transport_lock);
702 req = xprt_lookup_rqst(xprt, xprt->tcp_xid);
704 xprt->tcp_flags &= ~XPRT_COPY_DATA;
705 dprintk("RPC: XID %08x request not found!\n",
706 ntohl(xprt->tcp_xid));
707 spin_unlock(&xprt->transport_lock);
711 rcvbuf = &req->rq_private_buf;
713 if (len > xprt->tcp_reclen - xprt->tcp_offset) {
714 skb_reader_t my_desc;
716 len = xprt->tcp_reclen - xprt->tcp_offset;
717 memcpy(&my_desc, desc, sizeof(my_desc));
719 r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
720 &my_desc, xs_tcp_copy_data);
724 r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
725 desc, xs_tcp_copy_data);
728 xprt->tcp_copied += r;
729 xprt->tcp_offset += r;
732 /* Error when copying to the receive buffer,
733 * usually because we weren't able to allocate
734 * additional buffer pages. All we can do now
735 * is turn off XPRT_COPY_DATA, so the request
736 * will not receive any additional updates,
738 * Any remaining data from this record will
741 xprt->tcp_flags &= ~XPRT_COPY_DATA;
742 dprintk("RPC: XID %08x truncated request\n",
743 ntohl(xprt->tcp_xid));
744 dprintk("RPC: xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
745 xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen);
749 dprintk("RPC: XID %08x read %Zd bytes\n",
750 ntohl(xprt->tcp_xid), r);
751 dprintk("RPC: xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
752 xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen);
754 if (xprt->tcp_copied == req->rq_private_buf.buflen)
755 xprt->tcp_flags &= ~XPRT_COPY_DATA;
756 else if (xprt->tcp_offset == xprt->tcp_reclen) {
757 if (xprt->tcp_flags & XPRT_LAST_FRAG)
758 xprt->tcp_flags &= ~XPRT_COPY_DATA;
762 if (!(xprt->tcp_flags & XPRT_COPY_DATA))
763 xprt_complete_rqst(req->rq_task, xprt->tcp_copied);
764 spin_unlock(&xprt->transport_lock);
765 xs_tcp_check_recm(xprt);
768 static inline void xs_tcp_read_discard(struct rpc_xprt *xprt, skb_reader_t *desc)
772 len = xprt->tcp_reclen - xprt->tcp_offset;
773 if (len > desc->count)
777 xprt->tcp_offset += len;
778 dprintk("RPC: discarded %Zu bytes\n", len);
779 xs_tcp_check_recm(xprt);
782 static int xs_tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, unsigned int offset, size_t len)
784 struct rpc_xprt *xprt = rd_desc->arg.data;
785 skb_reader_t desc = {
792 dprintk("RPC: xs_tcp_data_recv started\n");
794 /* Read in a new fragment marker if necessary */
795 /* Can we ever really expect to get completely empty fragments? */
796 if (xprt->tcp_flags & XPRT_COPY_RECM) {
797 xs_tcp_read_fraghdr(xprt, &desc);
800 /* Read in the xid if necessary */
801 if (xprt->tcp_flags & XPRT_COPY_XID) {
802 xs_tcp_read_xid(xprt, &desc);
805 /* Read in the request data */
806 if (xprt->tcp_flags & XPRT_COPY_DATA) {
807 xs_tcp_read_request(xprt, &desc);
810 /* Skip over any trailing bytes on short reads */
811 xs_tcp_read_discard(xprt, &desc);
812 } while (desc.count);
813 dprintk("RPC: xs_tcp_data_recv done\n");
814 return len - desc.count;
818 * xs_tcp_data_ready - "data ready" callback for TCP sockets
819 * @sk: socket with data to read
820 * @bytes: how much data to read
823 static void xs_tcp_data_ready(struct sock *sk, int bytes)
825 struct rpc_xprt *xprt;
826 read_descriptor_t rd_desc;
828 read_lock(&sk->sk_callback_lock);
829 dprintk("RPC: xs_tcp_data_ready...\n");
830 if (!(xprt = xprt_from_sock(sk)))
835 /* We use rd_desc to pass struct xprt to xs_tcp_data_recv */
836 rd_desc.arg.data = xprt;
837 rd_desc.count = 65536;
838 tcp_read_sock(sk, &rd_desc, xs_tcp_data_recv);
840 read_unlock(&sk->sk_callback_lock);
844 * xs_tcp_state_change - callback to handle TCP socket state changes
845 * @sk: socket whose state has changed
848 static void xs_tcp_state_change(struct sock *sk)
850 struct rpc_xprt *xprt;
852 read_lock(&sk->sk_callback_lock);
853 if (!(xprt = xprt_from_sock(sk)))
855 dprintk("RPC: xs_tcp_state_change client %p...\n", xprt);
856 dprintk("RPC: state %x conn %d dead %d zapped %d\n",
857 sk->sk_state, xprt_connected(xprt),
858 sock_flag(sk, SOCK_DEAD),
859 sock_flag(sk, SOCK_ZAPPED));
861 switch (sk->sk_state) {
862 case TCP_ESTABLISHED:
863 spin_lock_bh(&xprt->transport_lock);
864 if (!xprt_test_and_set_connected(xprt)) {
865 /* Reset TCP record info */
866 xprt->tcp_offset = 0;
867 xprt->tcp_reclen = 0;
868 xprt->tcp_copied = 0;
869 xprt->tcp_flags = XPRT_COPY_RECM | XPRT_COPY_XID;
870 xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
871 xprt_wake_pending_tasks(xprt, 0);
873 spin_unlock_bh(&xprt->transport_lock);
879 /* Try to schedule an autoclose RPC calls */
880 set_bit(XPRT_CLOSE_WAIT, &xprt->state);
881 if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
882 schedule_work(&xprt->task_cleanup);
884 xprt_disconnect(xprt);
887 read_unlock(&sk->sk_callback_lock);
891 * xs_udp_write_space - callback invoked when socket buffer space
893 * @sk: socket whose state has changed
895 * Called when more output buffer space is available for this socket.
896 * We try not to wake our writers until they can make "significant"
897 * progress, otherwise we'll waste resources thrashing kernel_sendmsg
898 * with a bunch of small requests.
900 static void xs_udp_write_space(struct sock *sk)
902 read_lock(&sk->sk_callback_lock);
904 /* from net/core/sock.c:sock_def_write_space */
905 if (sock_writeable(sk)) {
907 struct rpc_xprt *xprt;
909 if (unlikely(!(sock = sk->sk_socket)))
911 if (unlikely(!(xprt = xprt_from_sock(sk))))
913 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE, &sock->flags)))
916 xprt_write_space(xprt);
920 read_unlock(&sk->sk_callback_lock);
924 * xs_tcp_write_space - callback invoked when socket buffer space
926 * @sk: socket whose state has changed
928 * Called when more output buffer space is available for this socket.
929 * We try not to wake our writers until they can make "significant"
930 * progress, otherwise we'll waste resources thrashing kernel_sendmsg
931 * with a bunch of small requests.
933 static void xs_tcp_write_space(struct sock *sk)
935 read_lock(&sk->sk_callback_lock);
937 /* from net/core/stream.c:sk_stream_write_space */
938 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
940 struct rpc_xprt *xprt;
942 if (unlikely(!(sock = sk->sk_socket)))
944 if (unlikely(!(xprt = xprt_from_sock(sk))))
946 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE, &sock->flags)))
949 xprt_write_space(xprt);
953 read_unlock(&sk->sk_callback_lock);
956 static void xs_udp_do_set_buffer_size(struct rpc_xprt *xprt)
958 struct sock *sk = xprt->inet;
961 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
962 sk->sk_rcvbuf = xprt->rcvsize * xprt->max_reqs * 2;
965 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
966 sk->sk_sndbuf = xprt->sndsize * xprt->max_reqs * 2;
967 sk->sk_write_space(sk);
972 * xs_udp_set_buffer_size - set send and receive limits
973 * @xprt: generic transport
974 * @sndsize: requested size of send buffer, in bytes
975 * @rcvsize: requested size of receive buffer, in bytes
977 * Set socket send and receive buffer size limits.
979 static void xs_udp_set_buffer_size(struct rpc_xprt *xprt, size_t sndsize, size_t rcvsize)
983 xprt->sndsize = sndsize + 1024;
986 xprt->rcvsize = rcvsize + 1024;
988 xs_udp_do_set_buffer_size(xprt);
992 * xs_udp_timer - called when a retransmit timeout occurs on a UDP transport
993 * @task: task that timed out
995 * Adjust the congestion window after a retransmit timeout has occurred.
997 static void xs_udp_timer(struct rpc_task *task)
999 xprt_adjust_cwnd(task, -ETIMEDOUT);
1002 static unsigned short xs_get_random_port(void)
1004 unsigned short range = xprt_max_resvport - xprt_min_resvport;
1005 unsigned short rand = (unsigned short) net_random() % range;
1006 return rand + xprt_min_resvport;
1010 * xs_print_peer_address - format an IPv4 address for printing
1011 * @xprt: generic transport
1012 * @format: flags field indicating which parts of the address to render
1014 static char *xs_print_peer_address(struct rpc_xprt *xprt, enum rpc_display_format_t format)
1016 if (xprt->address_strings[format] != NULL)
1017 return xprt->address_strings[format];
1019 return "unprintable";
1023 * xs_set_port - reset the port number in the remote endpoint address
1024 * @xprt: generic transport
1025 * @port: new port number
1028 static void xs_set_port(struct rpc_xprt *xprt, unsigned short port)
1030 struct sockaddr_in *sap = (struct sockaddr_in *) &xprt->addr;
1032 dprintk("RPC: setting port for xprt %p to %u\n", xprt, port);
1034 sap->sin_port = htons(port);
1037 static int xs_bindresvport(struct rpc_xprt *xprt, struct socket *sock)
1039 struct sockaddr_in myaddr = {
1040 .sin_family = AF_INET,
1043 unsigned short port = xprt->port;
1046 myaddr.sin_port = htons(port);
1047 err = sock->ops->bind(sock, (struct sockaddr *) &myaddr,
1051 dprintk("RPC: xs_bindresvport bound to port %u\n",
1055 if (port <= xprt_min_resvport)
1056 port = xprt_max_resvport;
1059 } while (err == -EADDRINUSE && port != xprt->port);
1061 dprintk("RPC: can't bind to reserved port (%d).\n", -err);
1066 * xs_udp_connect_worker - set up a UDP socket
1067 * @args: RPC transport to connect
1069 * Invoked by a work queue tasklet.
1071 static void xs_udp_connect_worker(void *args)
1073 struct rpc_xprt *xprt = (struct rpc_xprt *) args;
1074 struct socket *sock = xprt->sock;
1075 int err, status = -EIO;
1077 if (xprt->shutdown || !xprt_bound(xprt))
1080 /* Start by resetting any existing state */
1083 if ((err = sock_create_kern(PF_INET, SOCK_DGRAM, IPPROTO_UDP, &sock)) < 0) {
1084 dprintk("RPC: can't create UDP transport socket (%d).\n", -err);
1088 if (xprt->resvport && xs_bindresvport(xprt, sock) < 0) {
1093 dprintk("RPC: worker connecting xprt %p to address: %s\n",
1094 xprt, xs_print_peer_address(xprt, RPC_DISPLAY_ALL));
1097 struct sock *sk = sock->sk;
1099 write_lock_bh(&sk->sk_callback_lock);
1101 sk->sk_user_data = xprt;
1102 xprt->old_data_ready = sk->sk_data_ready;
1103 xprt->old_state_change = sk->sk_state_change;
1104 xprt->old_write_space = sk->sk_write_space;
1105 sk->sk_data_ready = xs_udp_data_ready;
1106 sk->sk_write_space = xs_udp_write_space;
1107 sk->sk_no_check = UDP_CSUM_NORCV;
1108 sk->sk_allocation = GFP_ATOMIC;
1110 xprt_set_connected(xprt);
1112 /* Reset to new socket */
1116 write_unlock_bh(&sk->sk_callback_lock);
1118 xs_udp_do_set_buffer_size(xprt);
1121 xprt_wake_pending_tasks(xprt, status);
1122 xprt_clear_connecting(xprt);
1126 * We need to preserve the port number so the reply cache on the server can
1127 * find our cached RPC replies when we get around to reconnecting.
1129 static void xs_tcp_reuse_connection(struct rpc_xprt *xprt)
1132 struct socket *sock = xprt->sock;
1133 struct sockaddr any;
1135 dprintk("RPC: disconnecting xprt %p to reuse port\n", xprt);
1138 * Disconnect the transport socket by doing a connect operation
1139 * with AF_UNSPEC. This should return immediately...
1141 memset(&any, 0, sizeof(any));
1142 any.sa_family = AF_UNSPEC;
1143 result = sock->ops->connect(sock, &any, sizeof(any), 0);
1145 dprintk("RPC: AF_UNSPEC connect return code %d\n",
1150 * xs_tcp_connect_worker - connect a TCP socket to a remote endpoint
1151 * @args: RPC transport to connect
1153 * Invoked by a work queue tasklet.
1155 static void xs_tcp_connect_worker(void *args)
1157 struct rpc_xprt *xprt = (struct rpc_xprt *)args;
1158 struct socket *sock = xprt->sock;
1159 int err, status = -EIO;
1161 if (xprt->shutdown || !xprt_bound(xprt))
1165 /* start from scratch */
1166 if ((err = sock_create_kern(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock)) < 0) {
1167 dprintk("RPC: can't create TCP transport socket (%d).\n", -err);
1171 if (xprt->resvport && xs_bindresvport(xprt, sock) < 0) {
1176 /* "close" the socket, preserving the local port */
1177 xs_tcp_reuse_connection(xprt);
1179 dprintk("RPC: worker connecting xprt %p to address: %s\n",
1180 xprt, xs_print_peer_address(xprt, RPC_DISPLAY_ALL));
1183 struct sock *sk = sock->sk;
1185 write_lock_bh(&sk->sk_callback_lock);
1187 sk->sk_user_data = xprt;
1188 xprt->old_data_ready = sk->sk_data_ready;
1189 xprt->old_state_change = sk->sk_state_change;
1190 xprt->old_write_space = sk->sk_write_space;
1191 sk->sk_data_ready = xs_tcp_data_ready;
1192 sk->sk_state_change = xs_tcp_state_change;
1193 sk->sk_write_space = xs_tcp_write_space;
1194 sk->sk_allocation = GFP_ATOMIC;
1196 /* socket options */
1197 sk->sk_userlocks |= SOCK_BINDPORT_LOCK;
1198 sock_reset_flag(sk, SOCK_LINGER);
1199 tcp_sk(sk)->linger2 = 0;
1200 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
1202 xprt_clear_connected(xprt);
1204 /* Reset to new socket */
1208 write_unlock_bh(&sk->sk_callback_lock);
1211 /* Tell the socket layer to start connecting... */
1212 xprt->stat.connect_count++;
1213 xprt->stat.connect_start = jiffies;
1214 status = sock->ops->connect(sock, (struct sockaddr *) &xprt->addr,
1215 xprt->addrlen, O_NONBLOCK);
1216 dprintk("RPC: %p connect status %d connected %d sock state %d\n",
1217 xprt, -status, xprt_connected(xprt), sock->sk->sk_state);
1225 /* retry with existing socket, after a delay */
1228 /* get rid of existing socket, and retry */
1234 xprt_wake_pending_tasks(xprt, status);
1236 xprt_clear_connecting(xprt);
1240 * xs_connect - connect a socket to a remote endpoint
1241 * @task: address of RPC task that manages state of connect request
1243 * TCP: If the remote end dropped the connection, delay reconnecting.
1245 * UDP socket connects are synchronous, but we use a work queue anyway
1246 * to guarantee that even unprivileged user processes can set up a
1247 * socket on a privileged port.
1249 * If a UDP socket connect fails, the delay behavior here prevents
1250 * retry floods (hard mounts).
1252 static void xs_connect(struct rpc_task *task)
1254 struct rpc_xprt *xprt = task->tk_xprt;
1256 if (xprt_test_and_set_connecting(xprt))
1259 if (xprt->sock != NULL) {
1260 dprintk("RPC: xs_connect delayed xprt %p for %lu seconds\n",
1261 xprt, xprt->reestablish_timeout / HZ);
1262 schedule_delayed_work(&xprt->connect_worker,
1263 xprt->reestablish_timeout);
1264 xprt->reestablish_timeout <<= 1;
1265 if (xprt->reestablish_timeout > XS_TCP_MAX_REEST_TO)
1266 xprt->reestablish_timeout = XS_TCP_MAX_REEST_TO;
1268 dprintk("RPC: xs_connect scheduled xprt %p\n", xprt);
1269 schedule_work(&xprt->connect_worker);
1271 /* flush_scheduled_work can sleep... */
1272 if (!RPC_IS_ASYNC(task))
1273 flush_scheduled_work();
1278 * xs_udp_print_stats - display UDP socket-specifc stats
1279 * @xprt: rpc_xprt struct containing statistics
1283 static void xs_udp_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
1285 seq_printf(seq, "\txprt:\tudp %u %lu %lu %lu %lu %Lu %Lu\n",
1287 xprt->stat.bind_count,
1290 xprt->stat.bad_xids,
1292 xprt->stat.bklog_u);
1296 * xs_tcp_print_stats - display TCP socket-specifc stats
1297 * @xprt: rpc_xprt struct containing statistics
1301 static void xs_tcp_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
1305 if (xprt_connected(xprt))
1306 idle_time = (long)(jiffies - xprt->last_used) / HZ;
1308 seq_printf(seq, "\txprt:\ttcp %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu\n",
1310 xprt->stat.bind_count,
1311 xprt->stat.connect_count,
1312 xprt->stat.connect_time,
1316 xprt->stat.bad_xids,
1318 xprt->stat.bklog_u);
1321 static struct rpc_xprt_ops xs_udp_ops = {
1322 .set_buffer_size = xs_udp_set_buffer_size,
1323 .print_addr = xs_print_peer_address,
1324 .reserve_xprt = xprt_reserve_xprt_cong,
1325 .release_xprt = xprt_release_xprt_cong,
1326 .rpcbind = rpc_getport,
1327 .set_port = xs_set_port,
1328 .connect = xs_connect,
1329 .buf_alloc = rpc_malloc,
1330 .buf_free = rpc_free,
1331 .send_request = xs_udp_send_request,
1332 .set_retrans_timeout = xprt_set_retrans_timeout_rtt,
1333 .timer = xs_udp_timer,
1334 .release_request = xprt_release_rqst_cong,
1336 .destroy = xs_destroy,
1337 .print_stats = xs_udp_print_stats,
1340 static struct rpc_xprt_ops xs_tcp_ops = {
1341 .print_addr = xs_print_peer_address,
1342 .reserve_xprt = xprt_reserve_xprt,
1343 .release_xprt = xs_tcp_release_xprt,
1344 .rpcbind = rpc_getport,
1345 .set_port = xs_set_port,
1346 .connect = xs_connect,
1347 .buf_alloc = rpc_malloc,
1348 .buf_free = rpc_free,
1349 .send_request = xs_tcp_send_request,
1350 .set_retrans_timeout = xprt_set_retrans_timeout_def,
1352 .destroy = xs_destroy,
1353 .print_stats = xs_tcp_print_stats,
1357 * xs_setup_udp - Set up transport to use a UDP socket
1358 * @xprt: transport to set up
1359 * @to: timeout parameters
1362 int xs_setup_udp(struct rpc_xprt *xprt, struct rpc_timeout *to)
1364 size_t slot_table_size;
1365 struct sockaddr_in *addr = (struct sockaddr_in *) &xprt->addr;
1367 xprt->max_reqs = xprt_udp_slot_table_entries;
1368 slot_table_size = xprt->max_reqs * sizeof(xprt->slot[0]);
1369 xprt->slot = kzalloc(slot_table_size, GFP_KERNEL);
1370 if (xprt->slot == NULL)
1373 if (ntohs(addr->sin_port != 0))
1374 xprt_set_bound(xprt);
1375 xprt->port = xs_get_random_port();
1377 xprt->prot = IPPROTO_UDP;
1379 xprt->resvport = capable(CAP_NET_BIND_SERVICE) ? 1 : 0;
1380 /* XXX: header size can vary due to auth type, IPv6, etc. */
1381 xprt->max_payload = (1U << 16) - (MAX_HEADER << 3);
1383 INIT_WORK(&xprt->connect_worker, xs_udp_connect_worker, xprt);
1384 xprt->bind_timeout = XS_BIND_TO;
1385 xprt->connect_timeout = XS_UDP_CONN_TO;
1386 xprt->reestablish_timeout = XS_UDP_REEST_TO;
1387 xprt->idle_timeout = XS_IDLE_DISC_TO;
1389 xprt->ops = &xs_udp_ops;
1392 xprt->timeout = *to;
1394 xprt_set_timeout(&xprt->timeout, 5, 5 * HZ);
1396 xs_format_peer_addresses(xprt);
1397 dprintk("RPC: set up transport to address %s\n",
1398 xs_print_peer_address(xprt, RPC_DISPLAY_ALL));
1404 * xs_setup_tcp - Set up transport to use a TCP socket
1405 * @xprt: transport to set up
1406 * @to: timeout parameters
1409 int xs_setup_tcp(struct rpc_xprt *xprt, struct rpc_timeout *to)
1411 size_t slot_table_size;
1412 struct sockaddr_in *addr = (struct sockaddr_in *) &xprt->addr;
1414 xprt->max_reqs = xprt_tcp_slot_table_entries;
1415 slot_table_size = xprt->max_reqs * sizeof(xprt->slot[0]);
1416 xprt->slot = kzalloc(slot_table_size, GFP_KERNEL);
1417 if (xprt->slot == NULL)
1420 if (ntohs(addr->sin_port) != 0)
1421 xprt_set_bound(xprt);
1422 xprt->port = xs_get_random_port();
1424 xprt->prot = IPPROTO_TCP;
1425 xprt->tsh_size = sizeof(rpc_fraghdr) / sizeof(u32);
1426 xprt->resvport = capable(CAP_NET_BIND_SERVICE) ? 1 : 0;
1427 xprt->max_payload = RPC_MAX_FRAGMENT_SIZE;
1429 INIT_WORK(&xprt->connect_worker, xs_tcp_connect_worker, xprt);
1430 xprt->bind_timeout = XS_BIND_TO;
1431 xprt->connect_timeout = XS_TCP_CONN_TO;
1432 xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
1433 xprt->idle_timeout = XS_IDLE_DISC_TO;
1435 xprt->ops = &xs_tcp_ops;
1438 xprt->timeout = *to;
1440 xprt_set_timeout(&xprt->timeout, 2, 60 * HZ);
1442 xs_format_peer_addresses(xprt);
1443 dprintk("RPC: set up transport to address %s\n",
1444 xs_print_peer_address(xprt, RPC_DISPLAY_ALL));