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/file.h>
34 #include <net/checksum.h>
39 * How many times to try sending a request on a socket before waiting
40 * for the socket buffer to clear.
42 #define XS_SENDMSG_RETRY (10U)
45 * Time out for an RPC UDP socket connect. UDP socket connects are
46 * synchronous, but we set a timeout anyway in case of resource
47 * exhaustion on the local host.
49 #define XS_UDP_CONN_TO (5U * HZ)
52 * Wait duration for an RPC TCP connection to be established. Solaris
53 * NFS over TCP uses 60 seconds, for example, which is in line with how
54 * long a server takes to reboot.
56 #define XS_TCP_CONN_TO (60U * HZ)
59 * Wait duration for a reply from the RPC portmapper.
61 #define XS_BIND_TO (60U * HZ)
64 * Delay if a UDP socket connect error occurs. This is most likely some
65 * kind of resource problem on the local host.
67 #define XS_UDP_REEST_TO (2U * HZ)
70 * The reestablish timeout allows clients to delay for a bit before attempting
71 * to reconnect to a server that just dropped our connection.
73 * We implement an exponential backoff when trying to reestablish a TCP
74 * transport connection with the server. Some servers like to drop a TCP
75 * connection when they are overworked, so we start with a short timeout and
76 * increase over time if the server is down or not responding.
78 #define XS_TCP_INIT_REEST_TO (3U * HZ)
79 #define XS_TCP_MAX_REEST_TO (5U * 60 * HZ)
82 * TCP idle timeout; client drops the transport socket if it is idle
83 * for this long. Note that we also timeout UDP sockets to prevent
84 * holding port numbers when there is no RPC traffic.
86 #define XS_IDLE_DISC_TO (5U * 60 * HZ)
89 # undef RPC_DEBUG_DATA
90 # define RPCDBG_FACILITY RPCDBG_TRANS
94 static void xs_pktdump(char *msg
, u32
*packet
, unsigned int count
)
96 u8
*buf
= (u8
*) packet
;
99 dprintk("RPC: %s\n", msg
);
100 for (j
= 0; j
< count
&& j
< 128; j
+= 4) {
104 dprintk("0x%04x ", j
);
106 dprintk("%02x%02x%02x%02x ",
107 buf
[j
], buf
[j
+1], buf
[j
+2], buf
[j
+3]);
112 static inline void xs_pktdump(char *msg
, u32
*packet
, unsigned int count
)
118 #define XS_SENDMSG_FLAGS (MSG_DONTWAIT | MSG_NOSIGNAL)
120 static inline int xs_send_head(struct socket
*sock
, struct sockaddr
*addr
, int addrlen
, struct xdr_buf
*xdr
, unsigned int base
, unsigned int len
)
123 .iov_base
= xdr
->head
[0].iov_base
+ base
,
124 .iov_len
= len
- base
,
126 struct msghdr msg
= {
128 .msg_namelen
= addrlen
,
129 .msg_flags
= XS_SENDMSG_FLAGS
,
133 msg
.msg_flags
|= MSG_MORE
;
135 if (likely(iov
.iov_len
))
136 return kernel_sendmsg(sock
, &msg
, &iov
, 1, iov
.iov_len
);
137 return kernel_sendmsg(sock
, &msg
, NULL
, 0, 0);
140 static int xs_send_tail(struct socket
*sock
, struct xdr_buf
*xdr
, unsigned int base
, unsigned int len
)
143 .iov_base
= xdr
->tail
[0].iov_base
+ base
,
144 .iov_len
= len
- base
,
146 struct msghdr msg
= {
147 .msg_flags
= XS_SENDMSG_FLAGS
,
150 return kernel_sendmsg(sock
, &msg
, &iov
, 1, iov
.iov_len
);
154 * xs_sendpages - write pages directly to a socket
155 * @sock: socket to send on
156 * @addr: UDP only -- address of destination
157 * @addrlen: UDP only -- length of destination address
158 * @xdr: buffer containing this request
159 * @base: starting position in the buffer
162 static inline int xs_sendpages(struct socket
*sock
, struct sockaddr
*addr
, int addrlen
, struct xdr_buf
*xdr
, unsigned int base
)
164 struct page
**ppage
= xdr
->pages
;
165 unsigned int len
, pglen
= xdr
->page_len
;
167 ssize_t (*sendpage
)(struct socket
*, struct page
*, int, size_t, int);
172 clear_bit(SOCK_ASYNC_NOSPACE
, &sock
->flags
);
174 len
= xdr
->head
[0].iov_len
;
175 if (base
< len
|| (addr
!= NULL
&& base
== 0)) {
176 err
= xs_send_head(sock
, addr
, addrlen
, xdr
, base
, len
);
181 if (err
!= (len
- base
))
187 if (unlikely(pglen
== 0))
189 if (unlikely(base
>= pglen
)) {
193 if (base
|| xdr
->page_base
) {
195 base
+= xdr
->page_base
;
196 ppage
+= base
>> PAGE_CACHE_SHIFT
;
197 base
&= ~PAGE_CACHE_MASK
;
200 sendpage
= sock
->ops
->sendpage
? : sock_no_sendpage
;
202 int flags
= XS_SENDMSG_FLAGS
;
204 len
= PAGE_CACHE_SIZE
;
210 if (pglen
!= len
|| xdr
->tail
[0].iov_len
!= 0)
213 /* Hmm... We might be dealing with highmem pages */
214 if (PageHighMem(*ppage
))
215 sendpage
= sock_no_sendpage
;
216 err
= sendpage(sock
, *ppage
, base
, len
, flags
);
225 } while ((pglen
-= len
) != 0);
227 len
= xdr
->tail
[0].iov_len
;
229 err
= xs_send_tail(sock
, xdr
, base
, len
);
240 * xs_nospace - place task on wait queue if transmit was incomplete
241 * @task: task to put to sleep
244 static void xs_nospace(struct rpc_task
*task
)
246 struct rpc_rqst
*req
= task
->tk_rqstp
;
247 struct rpc_xprt
*xprt
= req
->rq_xprt
;
249 dprintk("RPC: %4d xmit incomplete (%u left of %u)\n",
250 task
->tk_pid
, req
->rq_slen
- req
->rq_bytes_sent
,
253 if (test_bit(SOCK_ASYNC_NOSPACE
, &xprt
->sock
->flags
)) {
254 /* Protect against races with write_space */
255 spin_lock_bh(&xprt
->transport_lock
);
257 /* Don't race with disconnect */
258 if (!xprt_connected(xprt
))
259 task
->tk_status
= -ENOTCONN
;
260 else if (test_bit(SOCK_NOSPACE
, &xprt
->sock
->flags
))
261 xprt_wait_for_buffer_space(task
);
263 spin_unlock_bh(&xprt
->transport_lock
);
265 /* Keep holding the socket if it is blocked */
266 rpc_delay(task
, HZ
>>4);
270 * xs_udp_send_request - write an RPC request to a UDP socket
271 * @task: address of RPC task that manages the state of an RPC request
274 * 0: The request has been sent
275 * EAGAIN: The socket was blocked, please call again later to
276 * complete the request
277 * ENOTCONN: Caller needs to invoke connect logic then call again
278 * other: Some other error occured, the request was not sent
280 static int xs_udp_send_request(struct rpc_task
*task
)
282 struct rpc_rqst
*req
= task
->tk_rqstp
;
283 struct rpc_xprt
*xprt
= req
->rq_xprt
;
284 struct xdr_buf
*xdr
= &req
->rq_snd_buf
;
287 xs_pktdump("packet data:",
288 req
->rq_svec
->iov_base
,
289 req
->rq_svec
->iov_len
);
291 req
->rq_xtime
= jiffies
;
292 status
= xs_sendpages(xprt
->sock
, (struct sockaddr
*) &xprt
->addr
,
293 sizeof(xprt
->addr
), xdr
, req
->rq_bytes_sent
);
295 dprintk("RPC: xs_udp_send_request(%u) = %d\n",
296 xdr
->len
- req
->rq_bytes_sent
, status
);
298 if (likely(status
>= (int) req
->rq_slen
))
301 /* Still some bytes left; set up for a retry later. */
309 /* When the server has died, an ICMP port unreachable message
310 * prompts ECONNREFUSED. */
316 dprintk("RPC: sendmsg returned unrecognized error %d\n",
324 static inline void xs_encode_tcp_record_marker(struct xdr_buf
*buf
)
326 u32 reclen
= buf
->len
- sizeof(rpc_fraghdr
);
327 rpc_fraghdr
*base
= buf
->head
[0].iov_base
;
328 *base
= htonl(RPC_LAST_STREAM_FRAGMENT
| reclen
);
332 * xs_tcp_send_request - write an RPC request to a TCP socket
333 * @task: address of RPC task that manages the state of an RPC request
336 * 0: The request has been sent
337 * EAGAIN: The socket was blocked, please call again later to
338 * complete the request
339 * ENOTCONN: Caller needs to invoke connect logic then call again
340 * other: Some other error occured, the request was not sent
342 * XXX: In the case of soft timeouts, should we eventually give up
343 * if sendmsg is not able to make progress?
345 static int xs_tcp_send_request(struct rpc_task
*task
)
347 struct rpc_rqst
*req
= task
->tk_rqstp
;
348 struct rpc_xprt
*xprt
= req
->rq_xprt
;
349 struct xdr_buf
*xdr
= &req
->rq_snd_buf
;
350 int status
, retry
= 0;
352 xs_encode_tcp_record_marker(&req
->rq_snd_buf
);
354 xs_pktdump("packet data:",
355 req
->rq_svec
->iov_base
,
356 req
->rq_svec
->iov_len
);
358 /* Continue transmitting the packet/record. We must be careful
359 * to cope with writespace callbacks arriving _after_ we have
360 * called sendmsg(). */
362 req
->rq_xtime
= jiffies
;
363 status
= xs_sendpages(xprt
->sock
, NULL
, 0, xdr
,
366 dprintk("RPC: xs_tcp_send_request(%u) = %d\n",
367 xdr
->len
- req
->rq_bytes_sent
, status
);
369 if (unlikely(status
< 0))
372 /* If we've sent the entire packet, immediately
373 * reset the count of bytes sent. */
374 req
->rq_bytes_sent
+= status
;
375 if (likely(req
->rq_bytes_sent
>= req
->rq_slen
)) {
376 req
->rq_bytes_sent
= 0;
381 if (retry
++ > XS_SENDMSG_RETRY
)
396 dprintk("RPC: sendmsg returned unrecognized error %d\n",
398 xprt_disconnect(xprt
);
406 * xs_close - close a socket
409 * This is used when all requests are complete; ie, no DRC state remains
410 * on the server we want to save.
412 static void xs_close(struct rpc_xprt
*xprt
)
414 struct socket
*sock
= xprt
->sock
;
415 struct sock
*sk
= xprt
->inet
;
420 dprintk("RPC: xs_close xprt %p\n", xprt
);
422 write_lock_bh(&sk
->sk_callback_lock
);
426 sk
->sk_user_data
= NULL
;
427 sk
->sk_data_ready
= xprt
->old_data_ready
;
428 sk
->sk_state_change
= xprt
->old_state_change
;
429 sk
->sk_write_space
= xprt
->old_write_space
;
430 write_unlock_bh(&sk
->sk_callback_lock
);
438 * xs_destroy - prepare to shutdown a transport
439 * @xprt: doomed transport
442 static void xs_destroy(struct rpc_xprt
*xprt
)
444 dprintk("RPC: xs_destroy xprt %p\n", xprt
);
446 cancel_delayed_work(&xprt
->connect_worker
);
447 flush_scheduled_work();
449 xprt_disconnect(xprt
);
454 static inline struct rpc_xprt
*xprt_from_sock(struct sock
*sk
)
456 return (struct rpc_xprt
*) sk
->sk_user_data
;
460 * xs_udp_data_ready - "data ready" callback for UDP sockets
461 * @sk: socket with data to read
462 * @len: how much data to read
465 static void xs_udp_data_ready(struct sock
*sk
, int len
)
467 struct rpc_task
*task
;
468 struct rpc_xprt
*xprt
;
469 struct rpc_rqst
*rovr
;
471 int err
, repsize
, copied
;
474 read_lock(&sk
->sk_callback_lock
);
475 dprintk("RPC: xs_udp_data_ready...\n");
476 if (!(xprt
= xprt_from_sock(sk
)))
479 if ((skb
= skb_recv_datagram(sk
, 0, 1, &err
)) == NULL
)
485 repsize
= skb
->len
- sizeof(struct udphdr
);
487 dprintk("RPC: impossible RPC reply size %d!\n", repsize
);
491 /* Copy the XID from the skb... */
492 xp
= skb_header_pointer(skb
, sizeof(struct udphdr
),
493 sizeof(_xid
), &_xid
);
497 /* Look up and lock the request corresponding to the given XID */
498 spin_lock(&xprt
->transport_lock
);
499 rovr
= xprt_lookup_rqst(xprt
, *xp
);
502 task
= rovr
->rq_task
;
504 if ((copied
= rovr
->rq_private_buf
.buflen
) > repsize
)
507 /* Suck it into the iovec, verify checksum if not done by hw. */
508 if (csum_partial_copy_to_xdr(&rovr
->rq_private_buf
, skb
))
511 /* Something worked... */
512 dst_confirm(skb
->dst
);
514 xprt_adjust_cwnd(task
, copied
);
515 xprt_update_rtt(task
);
516 xprt_complete_rqst(task
, copied
);
519 spin_unlock(&xprt
->transport_lock
);
521 skb_free_datagram(sk
, skb
);
523 read_unlock(&sk
->sk_callback_lock
);
526 static inline size_t xs_tcp_copy_data(skb_reader_t
*desc
, void *p
, size_t len
)
528 if (len
> desc
->count
)
530 if (skb_copy_bits(desc
->skb
, desc
->offset
, p
, len
)) {
531 dprintk("RPC: failed to copy %zu bytes from skb. %zu bytes remain\n",
537 dprintk("RPC: copied %zu bytes from skb. %zu bytes remain\n",
542 static inline void xs_tcp_read_fraghdr(struct rpc_xprt
*xprt
, skb_reader_t
*desc
)
547 p
= ((char *) &xprt
->tcp_recm
) + xprt
->tcp_offset
;
548 len
= sizeof(xprt
->tcp_recm
) - xprt
->tcp_offset
;
549 used
= xs_tcp_copy_data(desc
, p
, len
);
550 xprt
->tcp_offset
+= used
;
554 xprt
->tcp_reclen
= ntohl(xprt
->tcp_recm
);
555 if (xprt
->tcp_reclen
& RPC_LAST_STREAM_FRAGMENT
)
556 xprt
->tcp_flags
|= XPRT_LAST_FRAG
;
558 xprt
->tcp_flags
&= ~XPRT_LAST_FRAG
;
559 xprt
->tcp_reclen
&= RPC_FRAGMENT_SIZE_MASK
;
561 xprt
->tcp_flags
&= ~XPRT_COPY_RECM
;
562 xprt
->tcp_offset
= 0;
564 /* Sanity check of the record length */
565 if (unlikely(xprt
->tcp_reclen
< 4)) {
566 dprintk("RPC: invalid TCP record fragment length\n");
567 xprt_disconnect(xprt
);
570 dprintk("RPC: reading TCP record fragment of length %d\n",
574 static void xs_tcp_check_recm(struct rpc_xprt
*xprt
)
576 dprintk("RPC: xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u, tcp_flags = %lx\n",
577 xprt
, xprt
->tcp_copied
, xprt
->tcp_offset
, xprt
->tcp_reclen
, xprt
->tcp_flags
);
578 if (xprt
->tcp_offset
== xprt
->tcp_reclen
) {
579 xprt
->tcp_flags
|= XPRT_COPY_RECM
;
580 xprt
->tcp_offset
= 0;
581 if (xprt
->tcp_flags
& XPRT_LAST_FRAG
) {
582 xprt
->tcp_flags
&= ~XPRT_COPY_DATA
;
583 xprt
->tcp_flags
|= XPRT_COPY_XID
;
584 xprt
->tcp_copied
= 0;
589 static inline void xs_tcp_read_xid(struct rpc_xprt
*xprt
, skb_reader_t
*desc
)
594 len
= sizeof(xprt
->tcp_xid
) - xprt
->tcp_offset
;
595 dprintk("RPC: reading XID (%Zu bytes)\n", len
);
596 p
= ((char *) &xprt
->tcp_xid
) + xprt
->tcp_offset
;
597 used
= xs_tcp_copy_data(desc
, p
, len
);
598 xprt
->tcp_offset
+= used
;
601 xprt
->tcp_flags
&= ~XPRT_COPY_XID
;
602 xprt
->tcp_flags
|= XPRT_COPY_DATA
;
603 xprt
->tcp_copied
= 4;
604 dprintk("RPC: reading reply for XID %08x\n",
605 ntohl(xprt
->tcp_xid
));
606 xs_tcp_check_recm(xprt
);
609 static inline void xs_tcp_read_request(struct rpc_xprt
*xprt
, skb_reader_t
*desc
)
611 struct rpc_rqst
*req
;
612 struct xdr_buf
*rcvbuf
;
616 /* Find and lock the request corresponding to this xid */
617 spin_lock(&xprt
->transport_lock
);
618 req
= xprt_lookup_rqst(xprt
, xprt
->tcp_xid
);
620 xprt
->tcp_flags
&= ~XPRT_COPY_DATA
;
621 dprintk("RPC: XID %08x request not found!\n",
622 ntohl(xprt
->tcp_xid
));
623 spin_unlock(&xprt
->transport_lock
);
627 rcvbuf
= &req
->rq_private_buf
;
629 if (len
> xprt
->tcp_reclen
- xprt
->tcp_offset
) {
630 skb_reader_t my_desc
;
632 len
= xprt
->tcp_reclen
- xprt
->tcp_offset
;
633 memcpy(&my_desc
, desc
, sizeof(my_desc
));
635 r
= xdr_partial_copy_from_skb(rcvbuf
, xprt
->tcp_copied
,
636 &my_desc
, xs_tcp_copy_data
);
640 r
= xdr_partial_copy_from_skb(rcvbuf
, xprt
->tcp_copied
,
641 desc
, xs_tcp_copy_data
);
644 xprt
->tcp_copied
+= r
;
645 xprt
->tcp_offset
+= r
;
648 /* Error when copying to the receive buffer,
649 * usually because we weren't able to allocate
650 * additional buffer pages. All we can do now
651 * is turn off XPRT_COPY_DATA, so the request
652 * will not receive any additional updates,
654 * Any remaining data from this record will
657 xprt
->tcp_flags
&= ~XPRT_COPY_DATA
;
658 dprintk("RPC: XID %08x truncated request\n",
659 ntohl(xprt
->tcp_xid
));
660 dprintk("RPC: xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
661 xprt
, xprt
->tcp_copied
, xprt
->tcp_offset
, xprt
->tcp_reclen
);
665 dprintk("RPC: XID %08x read %Zd bytes\n",
666 ntohl(xprt
->tcp_xid
), r
);
667 dprintk("RPC: xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
668 xprt
, xprt
->tcp_copied
, xprt
->tcp_offset
, xprt
->tcp_reclen
);
670 if (xprt
->tcp_copied
== req
->rq_private_buf
.buflen
)
671 xprt
->tcp_flags
&= ~XPRT_COPY_DATA
;
672 else if (xprt
->tcp_offset
== xprt
->tcp_reclen
) {
673 if (xprt
->tcp_flags
& XPRT_LAST_FRAG
)
674 xprt
->tcp_flags
&= ~XPRT_COPY_DATA
;
678 if (!(xprt
->tcp_flags
& XPRT_COPY_DATA
))
679 xprt_complete_rqst(req
->rq_task
, xprt
->tcp_copied
);
680 spin_unlock(&xprt
->transport_lock
);
681 xs_tcp_check_recm(xprt
);
684 static inline void xs_tcp_read_discard(struct rpc_xprt
*xprt
, skb_reader_t
*desc
)
688 len
= xprt
->tcp_reclen
- xprt
->tcp_offset
;
689 if (len
> desc
->count
)
693 xprt
->tcp_offset
+= len
;
694 dprintk("RPC: discarded %Zu bytes\n", len
);
695 xs_tcp_check_recm(xprt
);
698 static int xs_tcp_data_recv(read_descriptor_t
*rd_desc
, struct sk_buff
*skb
, unsigned int offset
, size_t len
)
700 struct rpc_xprt
*xprt
= rd_desc
->arg
.data
;
701 skb_reader_t desc
= {
708 dprintk("RPC: xs_tcp_data_recv started\n");
710 /* Read in a new fragment marker if necessary */
711 /* Can we ever really expect to get completely empty fragments? */
712 if (xprt
->tcp_flags
& XPRT_COPY_RECM
) {
713 xs_tcp_read_fraghdr(xprt
, &desc
);
716 /* Read in the xid if necessary */
717 if (xprt
->tcp_flags
& XPRT_COPY_XID
) {
718 xs_tcp_read_xid(xprt
, &desc
);
721 /* Read in the request data */
722 if (xprt
->tcp_flags
& XPRT_COPY_DATA
) {
723 xs_tcp_read_request(xprt
, &desc
);
726 /* Skip over any trailing bytes on short reads */
727 xs_tcp_read_discard(xprt
, &desc
);
728 } while (desc
.count
);
729 dprintk("RPC: xs_tcp_data_recv done\n");
730 return len
- desc
.count
;
734 * xs_tcp_data_ready - "data ready" callback for TCP sockets
735 * @sk: socket with data to read
736 * @bytes: how much data to read
739 static void xs_tcp_data_ready(struct sock
*sk
, int bytes
)
741 struct rpc_xprt
*xprt
;
742 read_descriptor_t rd_desc
;
744 read_lock(&sk
->sk_callback_lock
);
745 dprintk("RPC: xs_tcp_data_ready...\n");
746 if (!(xprt
= xprt_from_sock(sk
)))
751 /* We use rd_desc to pass struct xprt to xs_tcp_data_recv */
752 rd_desc
.arg
.data
= xprt
;
753 rd_desc
.count
= 65536;
754 tcp_read_sock(sk
, &rd_desc
, xs_tcp_data_recv
);
756 read_unlock(&sk
->sk_callback_lock
);
760 * xs_tcp_state_change - callback to handle TCP socket state changes
761 * @sk: socket whose state has changed
764 static void xs_tcp_state_change(struct sock
*sk
)
766 struct rpc_xprt
*xprt
;
768 read_lock(&sk
->sk_callback_lock
);
769 if (!(xprt
= xprt_from_sock(sk
)))
771 dprintk("RPC: xs_tcp_state_change client %p...\n", xprt
);
772 dprintk("RPC: state %x conn %d dead %d zapped %d\n",
773 sk
->sk_state
, xprt_connected(xprt
),
774 sock_flag(sk
, SOCK_DEAD
),
775 sock_flag(sk
, SOCK_ZAPPED
));
777 switch (sk
->sk_state
) {
778 case TCP_ESTABLISHED
:
779 spin_lock_bh(&xprt
->transport_lock
);
780 if (!xprt_test_and_set_connected(xprt
)) {
781 /* Reset TCP record info */
782 xprt
->tcp_offset
= 0;
783 xprt
->tcp_reclen
= 0;
784 xprt
->tcp_copied
= 0;
785 xprt
->tcp_flags
= XPRT_COPY_RECM
| XPRT_COPY_XID
;
786 xprt
->reestablish_timeout
= XS_TCP_INIT_REEST_TO
;
787 xprt_wake_pending_tasks(xprt
, 0);
789 spin_unlock_bh(&xprt
->transport_lock
);
795 xprt_disconnect(xprt
);
799 read_unlock(&sk
->sk_callback_lock
);
803 * xs_udp_write_space - callback invoked when socket buffer space
805 * @sk: socket whose state has changed
807 * Called when more output buffer space is available for this socket.
808 * We try not to wake our writers until they can make "significant"
809 * progress, otherwise we'll waste resources thrashing kernel_sendmsg
810 * with a bunch of small requests.
812 static void xs_udp_write_space(struct sock
*sk
)
814 read_lock(&sk
->sk_callback_lock
);
816 /* from net/core/sock.c:sock_def_write_space */
817 if (sock_writeable(sk
)) {
819 struct rpc_xprt
*xprt
;
821 if (unlikely(!(sock
= sk
->sk_socket
)))
823 if (unlikely(!(xprt
= xprt_from_sock(sk
))))
825 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE
, &sock
->flags
)))
828 xprt_write_space(xprt
);
832 read_unlock(&sk
->sk_callback_lock
);
836 * xs_tcp_write_space - callback invoked when socket buffer space
838 * @sk: socket whose state has changed
840 * Called when more output buffer space is available for this socket.
841 * We try not to wake our writers until they can make "significant"
842 * progress, otherwise we'll waste resources thrashing kernel_sendmsg
843 * with a bunch of small requests.
845 static void xs_tcp_write_space(struct sock
*sk
)
847 read_lock(&sk
->sk_callback_lock
);
849 /* from net/core/stream.c:sk_stream_write_space */
850 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
)) {
852 struct rpc_xprt
*xprt
;
854 if (unlikely(!(sock
= sk
->sk_socket
)))
856 if (unlikely(!(xprt
= xprt_from_sock(sk
))))
858 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE
, &sock
->flags
)))
861 xprt_write_space(xprt
);
865 read_unlock(&sk
->sk_callback_lock
);
868 static void xs_udp_do_set_buffer_size(struct rpc_xprt
*xprt
)
870 struct sock
*sk
= xprt
->inet
;
873 sk
->sk_userlocks
|= SOCK_RCVBUF_LOCK
;
874 sk
->sk_rcvbuf
= xprt
->rcvsize
* xprt
->max_reqs
* 2;
877 sk
->sk_userlocks
|= SOCK_SNDBUF_LOCK
;
878 sk
->sk_sndbuf
= xprt
->sndsize
* xprt
->max_reqs
* 2;
879 sk
->sk_write_space(sk
);
884 * xs_udp_set_buffer_size - set send and receive limits
885 * @xprt: generic transport
886 * @sndsize: requested size of send buffer, in bytes
887 * @rcvsize: requested size of receive buffer, in bytes
889 * Set socket send and receive buffer size limits.
891 static void xs_udp_set_buffer_size(struct rpc_xprt
*xprt
, size_t sndsize
, size_t rcvsize
)
895 xprt
->sndsize
= sndsize
+ 1024;
898 xprt
->rcvsize
= rcvsize
+ 1024;
900 xs_udp_do_set_buffer_size(xprt
);
904 * xs_udp_timer - called when a retransmit timeout occurs on a UDP transport
905 * @task: task that timed out
907 * Adjust the congestion window after a retransmit timeout has occurred.
909 static void xs_udp_timer(struct rpc_task
*task
)
911 xprt_adjust_cwnd(task
, -ETIMEDOUT
);
914 static int xs_bindresvport(struct rpc_xprt
*xprt
, struct socket
*sock
)
916 struct sockaddr_in myaddr
= {
917 .sin_family
= AF_INET
,
920 unsigned short port
= xprt
->port
;
923 myaddr
.sin_port
= htons(port
);
924 err
= sock
->ops
->bind(sock
, (struct sockaddr
*) &myaddr
,
928 dprintk("RPC: xs_bindresvport bound to port %u\n",
932 if (port
<= xprt_min_resvport
)
933 port
= xprt_max_resvport
;
936 } while (err
== -EADDRINUSE
&& port
!= xprt
->port
);
938 dprintk("RPC: can't bind to reserved port (%d).\n", -err
);
943 * xs_udp_connect_worker - set up a UDP socket
944 * @args: RPC transport to connect
946 * Invoked by a work queue tasklet.
948 static void xs_udp_connect_worker(void *args
)
950 struct rpc_xprt
*xprt
= (struct rpc_xprt
*) args
;
951 struct socket
*sock
= xprt
->sock
;
952 int err
, status
= -EIO
;
954 if (xprt
->shutdown
|| xprt
->addr
.sin_port
== 0)
957 dprintk("RPC: xs_udp_connect_worker for xprt %p\n", xprt
);
959 /* Start by resetting any existing state */
962 if ((err
= sock_create_kern(PF_INET
, SOCK_DGRAM
, IPPROTO_UDP
, &sock
)) < 0) {
963 dprintk("RPC: can't create UDP transport socket (%d).\n", -err
);
967 if (xprt
->resvport
&& xs_bindresvport(xprt
, sock
) < 0) {
973 struct sock
*sk
= sock
->sk
;
975 write_lock_bh(&sk
->sk_callback_lock
);
977 sk
->sk_user_data
= xprt
;
978 xprt
->old_data_ready
= sk
->sk_data_ready
;
979 xprt
->old_state_change
= sk
->sk_state_change
;
980 xprt
->old_write_space
= sk
->sk_write_space
;
981 sk
->sk_data_ready
= xs_udp_data_ready
;
982 sk
->sk_write_space
= xs_udp_write_space
;
983 sk
->sk_no_check
= UDP_CSUM_NORCV
;
985 xprt_set_connected(xprt
);
987 /* Reset to new socket */
991 write_unlock_bh(&sk
->sk_callback_lock
);
993 xs_udp_do_set_buffer_size(xprt
);
996 xprt_wake_pending_tasks(xprt
, status
);
997 xprt_clear_connecting(xprt
);
1001 * We need to preserve the port number so the reply cache on the server can
1002 * find our cached RPC replies when we get around to reconnecting.
1004 static void xs_tcp_reuse_connection(struct rpc_xprt
*xprt
)
1007 struct socket
*sock
= xprt
->sock
;
1008 struct sockaddr any
;
1010 dprintk("RPC: disconnecting xprt %p to reuse port\n", xprt
);
1013 * Disconnect the transport socket by doing a connect operation
1014 * with AF_UNSPEC. This should return immediately...
1016 memset(&any
, 0, sizeof(any
));
1017 any
.sa_family
= AF_UNSPEC
;
1018 result
= sock
->ops
->connect(sock
, &any
, sizeof(any
), 0);
1020 dprintk("RPC: AF_UNSPEC connect return code %d\n",
1025 * xs_tcp_connect_worker - connect a TCP socket to a remote endpoint
1026 * @args: RPC transport to connect
1028 * Invoked by a work queue tasklet.
1030 static void xs_tcp_connect_worker(void *args
)
1032 struct rpc_xprt
*xprt
= (struct rpc_xprt
*)args
;
1033 struct socket
*sock
= xprt
->sock
;
1034 int err
, status
= -EIO
;
1036 if (xprt
->shutdown
|| xprt
->addr
.sin_port
== 0)
1039 dprintk("RPC: xs_tcp_connect_worker for xprt %p\n", xprt
);
1042 /* start from scratch */
1043 if ((err
= sock_create_kern(PF_INET
, SOCK_STREAM
, IPPROTO_TCP
, &sock
)) < 0) {
1044 dprintk("RPC: can't create TCP transport socket (%d).\n", -err
);
1048 if (xprt
->resvport
&& xs_bindresvport(xprt
, sock
) < 0) {
1053 /* "close" the socket, preserving the local port */
1054 xs_tcp_reuse_connection(xprt
);
1057 struct sock
*sk
= sock
->sk
;
1059 write_lock_bh(&sk
->sk_callback_lock
);
1061 sk
->sk_user_data
= xprt
;
1062 xprt
->old_data_ready
= sk
->sk_data_ready
;
1063 xprt
->old_state_change
= sk
->sk_state_change
;
1064 xprt
->old_write_space
= sk
->sk_write_space
;
1065 sk
->sk_data_ready
= xs_tcp_data_ready
;
1066 sk
->sk_state_change
= xs_tcp_state_change
;
1067 sk
->sk_write_space
= xs_tcp_write_space
;
1069 /* socket options */
1070 sk
->sk_userlocks
|= SOCK_BINDPORT_LOCK
;
1071 sock_reset_flag(sk
, SOCK_LINGER
);
1072 tcp_sk(sk
)->linger2
= 0;
1073 tcp_sk(sk
)->nonagle
|= TCP_NAGLE_OFF
;
1075 xprt_clear_connected(xprt
);
1077 /* Reset to new socket */
1081 write_unlock_bh(&sk
->sk_callback_lock
);
1084 /* Tell the socket layer to start connecting... */
1085 status
= sock
->ops
->connect(sock
, (struct sockaddr
*) &xprt
->addr
,
1086 sizeof(xprt
->addr
), O_NONBLOCK
);
1087 dprintk("RPC: %p connect status %d connected %d sock state %d\n",
1088 xprt
, -status
, xprt_connected(xprt
), sock
->sk
->sk_state
);
1096 /* retry with existing socket, after a delay */
1099 /* get rid of existing socket, and retry */
1105 xprt_wake_pending_tasks(xprt
, status
);
1107 xprt_clear_connecting(xprt
);
1111 * xs_connect - connect a socket to a remote endpoint
1112 * @task: address of RPC task that manages state of connect request
1114 * TCP: If the remote end dropped the connection, delay reconnecting.
1116 * UDP socket connects are synchronous, but we use a work queue anyway
1117 * to guarantee that even unprivileged user processes can set up a
1118 * socket on a privileged port.
1120 * If a UDP socket connect fails, the delay behavior here prevents
1121 * retry floods (hard mounts).
1123 static void xs_connect(struct rpc_task
*task
)
1125 struct rpc_xprt
*xprt
= task
->tk_xprt
;
1127 if (xprt_test_and_set_connecting(xprt
))
1130 if (xprt
->sock
!= NULL
) {
1131 dprintk("RPC: xs_connect delayed xprt %p for %lu seconds\n",
1132 xprt
, xprt
->reestablish_timeout
/ HZ
);
1133 schedule_delayed_work(&xprt
->connect_worker
,
1134 xprt
->reestablish_timeout
);
1135 xprt
->reestablish_timeout
<<= 1;
1136 if (xprt
->reestablish_timeout
> XS_TCP_MAX_REEST_TO
)
1137 xprt
->reestablish_timeout
= XS_TCP_MAX_REEST_TO
;
1139 dprintk("RPC: xs_connect scheduled xprt %p\n", xprt
);
1140 schedule_work(&xprt
->connect_worker
);
1142 /* flush_scheduled_work can sleep... */
1143 if (!RPC_IS_ASYNC(task
))
1144 flush_scheduled_work();
1148 static struct rpc_xprt_ops xs_udp_ops
= {
1149 .set_buffer_size
= xs_udp_set_buffer_size
,
1150 .reserve_xprt
= xprt_reserve_xprt_cong
,
1151 .release_xprt
= xprt_release_xprt_cong
,
1152 .connect
= xs_connect
,
1153 .send_request
= xs_udp_send_request
,
1154 .set_retrans_timeout
= xprt_set_retrans_timeout_rtt
,
1155 .timer
= xs_udp_timer
,
1156 .release_request
= xprt_release_rqst_cong
,
1158 .destroy
= xs_destroy
,
1161 static struct rpc_xprt_ops xs_tcp_ops
= {
1162 .reserve_xprt
= xprt_reserve_xprt
,
1163 .release_xprt
= xprt_release_xprt
,
1164 .connect
= xs_connect
,
1165 .send_request
= xs_tcp_send_request
,
1166 .set_retrans_timeout
= xprt_set_retrans_timeout_def
,
1168 .destroy
= xs_destroy
,
1172 * xs_setup_udp - Set up transport to use a UDP socket
1173 * @xprt: transport to set up
1174 * @to: timeout parameters
1177 int xs_setup_udp(struct rpc_xprt
*xprt
, struct rpc_timeout
*to
)
1179 size_t slot_table_size
;
1181 dprintk("RPC: setting up udp-ipv4 transport...\n");
1183 xprt
->max_reqs
= xprt_udp_slot_table_entries
;
1184 slot_table_size
= xprt
->max_reqs
* sizeof(xprt
->slot
[0]);
1185 xprt
->slot
= kmalloc(slot_table_size
, GFP_KERNEL
);
1186 if (xprt
->slot
== NULL
)
1188 memset(xprt
->slot
, 0, slot_table_size
);
1190 xprt
->prot
= IPPROTO_UDP
;
1191 xprt
->port
= xprt_max_resvport
;
1193 xprt
->resvport
= capable(CAP_NET_BIND_SERVICE
) ? 1 : 0;
1194 /* XXX: header size can vary due to auth type, IPv6, etc. */
1195 xprt
->max_payload
= (1U << 16) - (MAX_HEADER
<< 3);
1197 INIT_WORK(&xprt
->connect_worker
, xs_udp_connect_worker
, xprt
);
1198 xprt
->bind_timeout
= XS_BIND_TO
;
1199 xprt
->connect_timeout
= XS_UDP_CONN_TO
;
1200 xprt
->reestablish_timeout
= XS_UDP_REEST_TO
;
1201 xprt
->idle_timeout
= XS_IDLE_DISC_TO
;
1203 xprt
->ops
= &xs_udp_ops
;
1206 xprt
->timeout
= *to
;
1208 xprt_set_timeout(&xprt
->timeout
, 5, 5 * HZ
);
1214 * xs_setup_tcp - Set up transport to use a TCP socket
1215 * @xprt: transport to set up
1216 * @to: timeout parameters
1219 int xs_setup_tcp(struct rpc_xprt
*xprt
, struct rpc_timeout
*to
)
1221 size_t slot_table_size
;
1223 dprintk("RPC: setting up tcp-ipv4 transport...\n");
1225 xprt
->max_reqs
= xprt_tcp_slot_table_entries
;
1226 slot_table_size
= xprt
->max_reqs
* sizeof(xprt
->slot
[0]);
1227 xprt
->slot
= kmalloc(slot_table_size
, GFP_KERNEL
);
1228 if (xprt
->slot
== NULL
)
1230 memset(xprt
->slot
, 0, slot_table_size
);
1232 xprt
->prot
= IPPROTO_TCP
;
1233 xprt
->port
= xprt_max_resvport
;
1234 xprt
->tsh_size
= sizeof(rpc_fraghdr
) / sizeof(u32
);
1235 xprt
->resvport
= capable(CAP_NET_BIND_SERVICE
) ? 1 : 0;
1236 xprt
->max_payload
= RPC_MAX_FRAGMENT_SIZE
;
1238 INIT_WORK(&xprt
->connect_worker
, xs_tcp_connect_worker
, xprt
);
1239 xprt
->bind_timeout
= XS_BIND_TO
;
1240 xprt
->connect_timeout
= XS_TCP_CONN_TO
;
1241 xprt
->reestablish_timeout
= XS_TCP_INIT_REEST_TO
;
1242 xprt
->idle_timeout
= XS_IDLE_DISC_TO
;
1244 xprt
->ops
= &xs_tcp_ops
;
1247 xprt
->timeout
= *to
;
1249 xprt_set_timeout(&xprt
->timeout
, 2, 60 * HZ
);