2 * linux/net/sunrpc/xprt.c
4 * This is a generic RPC call interface supporting congestion avoidance,
5 * and asynchronous calls.
7 * The interface works like this:
9 * - When a process places a call, it allocates a request slot if
10 * one is available. Otherwise, it sleeps on the backlog queue
12 * - Next, the caller puts together the RPC message, stuffs it into
13 * the request struct, and calls xprt_call().
14 * - xprt_call transmits the message and installs the caller on the
15 * socket's wait list. At the same time, it installs a timer that
16 * is run after the packet's timeout has expired.
17 * - When a packet arrives, the data_ready handler walks the list of
18 * pending requests for that socket. If a matching XID is found, the
19 * caller is woken up, and the timer removed.
20 * - When no reply arrives within the timeout interval, the timer is
21 * fired by the kernel and runs xprt_timer(). It either adjusts the
22 * timeout values (minor timeout) or wakes up the caller with a status
24 * - When the caller receives a notification from RPC that a reply arrived,
25 * it should release the RPC slot, and process the reply.
26 * If the call timed out, it may choose to retry the operation by
27 * adjusting the initial timeout value, and simply calling rpc_call
30 * Support for async RPC is done through a set of RPC-specific scheduling
31 * primitives that `transparently' work for processes as well as async
32 * tasks that rely on callbacks.
34 * Copyright (C) 1995-1997, Olaf Kirch <okir@monad.swb.de>
36 * TCP callback races fixes (C) 1998 Red Hat Software <alan@redhat.com>
37 * TCP send fixes (C) 1998 Red Hat Software <alan@redhat.com>
38 * TCP NFS related read + write fixes
39 * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
41 * Rewrite of larges part of the code in order to stabilize TCP stuff.
42 * Fix behaviour when socket buffer is full.
43 * (C) 1999 Trond Myklebust <trond.myklebust@fys.uio.no>
46 #include <linux/types.h>
47 #include <linux/slab.h>
48 #include <linux/capability.h>
49 #include <linux/sched.h>
50 #include <linux/errno.h>
51 #include <linux/socket.h>
53 #include <linux/net.h>
55 #include <linux/udp.h>
56 #include <linux/tcp.h>
57 #include <linux/sunrpc/clnt.h>
58 #include <linux/file.h>
59 #include <linux/workqueue.h>
60 #include <linux/random.h>
63 #include <net/checksum.h>
72 # undef RPC_DEBUG_DATA
73 # define RPCDBG_FACILITY RPCDBG_XPRT
76 #define XPRT_MAX_BACKOFF (8)
77 #define XPRT_IDLE_TIMEOUT (5*60*HZ)
78 #define XPRT_MAX_RESVPORT (800)
83 static void xprt_request_init(struct rpc_task *, struct rpc_xprt *);
84 static inline void do_xprt_reserve(struct rpc_task *);
85 static void xprt_disconnect(struct rpc_xprt *);
86 static void xprt_connect_status(struct rpc_task *task);
87 static struct rpc_xprt * xprt_setup(int proto, struct sockaddr_in *ap,
88 struct rpc_timeout *to);
89 static struct socket *xprt_create_socket(struct rpc_xprt *, int, int);
90 static void xprt_bind_socket(struct rpc_xprt *, struct socket *);
91 static int __xprt_get_cong(struct rpc_xprt *, struct rpc_task *);
93 static int xprt_clear_backlog(struct rpc_xprt *xprt);
97 * Print the buffer contents (first 128 bytes only--just enough for
101 xprt_pktdump(char *msg, u32 *packet, unsigned int count)
103 u8 *buf = (u8 *) packet;
106 dprintk("RPC: %s\n", msg);
107 for (j = 0; j < count && j < 128; j += 4) {
111 dprintk("0x%04x ", j);
113 dprintk("%02x%02x%02x%02x ",
114 buf[j], buf[j+1], buf[j+2], buf[j+3]);
120 xprt_pktdump(char *msg, u32 *packet, unsigned int count)
127 * Look up RPC transport given an INET socket
129 static inline struct rpc_xprt *
130 xprt_from_sock(struct sock *sk)
132 return (struct rpc_xprt *) sk->sk_user_data;
136 * Serialize write access to sockets, in order to prevent different
137 * requests from interfering with each other.
138 * Also prevents TCP socket connects from colliding with writes.
141 __xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
143 struct rpc_rqst *req = task->tk_rqstp;
145 if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate)) {
146 if (task == xprt->snd_task)
152 if (xprt->nocong || __xprt_get_cong(xprt, task)) {
153 xprt->snd_task = task;
155 req->rq_bytes_sent = 0;
160 smp_mb__before_clear_bit();
161 clear_bit(XPRT_LOCKED, &xprt->sockstate);
162 smp_mb__after_clear_bit();
164 dprintk("RPC: %4d failed to lock socket %p\n", task->tk_pid, xprt);
165 task->tk_timeout = 0;
166 task->tk_status = -EAGAIN;
167 if (req && req->rq_ntrans)
168 rpc_sleep_on(&xprt->resend, task, NULL, NULL);
170 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
175 xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
179 spin_lock_bh(&xprt->sock_lock);
180 retval = __xprt_lock_write(xprt, task);
181 spin_unlock_bh(&xprt->sock_lock);
187 __xprt_lock_write_next(struct rpc_xprt *xprt)
189 struct rpc_task *task;
191 if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate))
193 if (!xprt->nocong && RPCXPRT_CONGESTED(xprt))
195 task = rpc_wake_up_next(&xprt->resend);
197 task = rpc_wake_up_next(&xprt->sending);
201 if (xprt->nocong || __xprt_get_cong(xprt, task)) {
202 struct rpc_rqst *req = task->tk_rqstp;
203 xprt->snd_task = task;
205 req->rq_bytes_sent = 0;
211 smp_mb__before_clear_bit();
212 clear_bit(XPRT_LOCKED, &xprt->sockstate);
213 smp_mb__after_clear_bit();
217 * Releases the socket for use by other requests.
220 __xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
222 if (xprt->snd_task == task) {
223 xprt->snd_task = NULL;
224 smp_mb__before_clear_bit();
225 clear_bit(XPRT_LOCKED, &xprt->sockstate);
226 smp_mb__after_clear_bit();
227 __xprt_lock_write_next(xprt);
232 xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
234 spin_lock_bh(&xprt->sock_lock);
235 __xprt_release_write(xprt, task);
236 spin_unlock_bh(&xprt->sock_lock);
240 * Write data to socket.
243 xprt_sendmsg(struct rpc_xprt *xprt, struct rpc_rqst *req)
245 struct socket *sock = xprt->sock;
246 struct xdr_buf *xdr = &req->rq_snd_buf;
247 struct sockaddr *addr = NULL;
255 xprt_pktdump("packet data:",
256 req->rq_svec->iov_base,
257 req->rq_svec->iov_len);
259 /* For UDP, we need to provide an address */
261 addr = (struct sockaddr *) &xprt->addr;
262 addrlen = sizeof(xprt->addr);
264 /* Dont repeat bytes */
265 skip = req->rq_bytes_sent;
267 clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
268 result = xdr_sendpages(sock, addr, addrlen, xdr, skip, MSG_DONTWAIT);
270 dprintk("RPC: xprt_sendmsg(%d) = %d\n", xdr->len - skip, result);
277 /* When the server has died, an ICMP port unreachable message
278 * prompts ECONNREFUSED.
285 /* connection broken */
290 printk(KERN_NOTICE "RPC: sendmsg returned error %d\n", -result);
296 * Van Jacobson congestion avoidance. Check if the congestion window
297 * overflowed. Put the task to sleep if this is the case.
300 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
302 struct rpc_rqst *req = task->tk_rqstp;
306 dprintk("RPC: %4d xprt_cwnd_limited cong = %ld cwnd = %ld\n",
307 task->tk_pid, xprt->cong, xprt->cwnd);
308 if (RPCXPRT_CONGESTED(xprt))
311 xprt->cong += RPC_CWNDSCALE;
316 * Adjust the congestion window, and wake up the next task
317 * that has been sleeping due to congestion
320 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
325 xprt->cong -= RPC_CWNDSCALE;
326 __xprt_lock_write_next(xprt);
330 * Adjust RPC congestion window
331 * We use a time-smoothed congestion estimator to avoid heavy oscillation.
334 xprt_adjust_cwnd(struct rpc_xprt *xprt, int result)
339 if (result >= 0 && cwnd <= xprt->cong) {
340 /* The (cwnd >> 1) term makes sure
341 * the result gets rounded properly. */
342 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
343 if (cwnd > RPC_MAXCWND(xprt))
344 cwnd = RPC_MAXCWND(xprt);
345 __xprt_lock_write_next(xprt);
346 } else if (result == -ETIMEDOUT) {
348 if (cwnd < RPC_CWNDSCALE)
349 cwnd = RPC_CWNDSCALE;
351 dprintk("RPC: cong %ld, cwnd was %ld, now %ld\n",
352 xprt->cong, xprt->cwnd, cwnd);
357 * Reset the major timeout value
359 static void xprt_reset_majortimeo(struct rpc_rqst *req)
361 struct rpc_timeout *to = &req->rq_xprt->timeout;
363 req->rq_majortimeo = req->rq_timeout;
364 if (to->to_exponential)
365 req->rq_majortimeo <<= to->to_retries;
367 req->rq_majortimeo += to->to_increment * to->to_retries;
368 if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
369 req->rq_majortimeo = to->to_maxval;
370 req->rq_majortimeo += jiffies;
374 * Adjust timeout values etc for next retransmit
376 int xprt_adjust_timeout(struct rpc_rqst *req)
378 struct rpc_xprt *xprt = req->rq_xprt;
379 struct rpc_timeout *to = &xprt->timeout;
382 if (time_before(jiffies, req->rq_majortimeo)) {
383 if (to->to_exponential)
384 req->rq_timeout <<= 1;
386 req->rq_timeout += to->to_increment;
387 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
388 req->rq_timeout = to->to_maxval;
390 pprintk("RPC: %lu retrans\n", jiffies);
392 req->rq_timeout = to->to_initval;
394 xprt_reset_majortimeo(req);
395 /* Reset the RTT counters == "slow start" */
396 spin_lock_bh(&xprt->sock_lock);
397 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
398 spin_unlock_bh(&xprt->sock_lock);
399 pprintk("RPC: %lu timeout\n", jiffies);
403 if (req->rq_timeout == 0) {
404 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
405 req->rq_timeout = 5 * HZ;
411 * Close down a transport socket
414 xprt_close(struct rpc_xprt *xprt)
416 struct socket *sock = xprt->sock;
417 struct sock *sk = xprt->inet;
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 xprt_socket_autoclose(void *args)
440 struct rpc_xprt *xprt = (struct rpc_xprt *)args;
442 xprt_disconnect(xprt);
444 xprt_release_write(xprt, NULL);
448 * Mark a transport as disconnected
451 xprt_disconnect(struct rpc_xprt *xprt)
453 dprintk("RPC: disconnected transport %p\n", xprt);
454 spin_lock_bh(&xprt->sock_lock);
455 xprt_clear_connected(xprt);
456 rpc_wake_up_status(&xprt->pending, -ENOTCONN);
457 spin_unlock_bh(&xprt->sock_lock);
461 * Used to allow disconnection when we've been idle
464 xprt_init_autodisconnect(unsigned long data)
466 struct rpc_xprt *xprt = (struct rpc_xprt *)data;
468 spin_lock(&xprt->sock_lock);
469 if (!list_empty(&xprt->recv) || xprt->shutdown)
471 if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate))
473 spin_unlock(&xprt->sock_lock);
474 /* Let keventd close the socket */
475 if (test_bit(XPRT_CONNECTING, &xprt->sockstate) != 0)
476 xprt_release_write(xprt, NULL);
478 schedule_work(&xprt->task_cleanup);
481 spin_unlock(&xprt->sock_lock);
484 static void xprt_socket_connect(void *args)
486 struct rpc_xprt *xprt = (struct rpc_xprt *)args;
487 struct socket *sock = xprt->sock;
490 if (xprt->shutdown || xprt->addr.sin_port == 0)
494 * Start by resetting any existing state
497 sock = xprt_create_socket(xprt, xprt->prot, xprt->resvport);
499 /* couldn't create socket or bind to reserved port;
500 * this is likely a permanent error, so cause an abort */
503 xprt_bind_socket(xprt, sock);
504 xprt_sock_setbufsize(xprt);
511 * Tell the socket layer to start connecting...
513 status = sock->ops->connect(sock, (struct sockaddr *) &xprt->addr,
514 sizeof(xprt->addr), O_NONBLOCK);
515 dprintk("RPC: %p connect status %d connected %d sock state %d\n",
516 xprt, -status, xprt_connected(xprt), sock->sk->sk_state);
526 rpc_wake_up_status(&xprt->pending, status);
528 rpc_wake_up(&xprt->pending);
530 smp_mb__before_clear_bit();
531 clear_bit(XPRT_CONNECTING, &xprt->sockstate);
532 smp_mb__after_clear_bit();
536 * Attempt to connect a TCP socket.
539 void xprt_connect(struct rpc_task *task)
541 struct rpc_xprt *xprt = task->tk_xprt;
543 dprintk("RPC: %4d xprt_connect xprt %p %s connected\n", task->tk_pid,
544 xprt, (xprt_connected(xprt) ? "is" : "is not"));
546 if (xprt->shutdown) {
547 task->tk_status = -EIO;
550 if (!xprt->addr.sin_port) {
551 task->tk_status = -EIO;
554 if (!xprt_lock_write(xprt, task))
556 if (xprt_connected(xprt))
560 task->tk_rqstp->rq_bytes_sent = 0;
562 task->tk_timeout = RPC_CONNECT_TIMEOUT;
563 rpc_sleep_on(&xprt->pending, task, xprt_connect_status, NULL);
564 if (!test_and_set_bit(XPRT_CONNECTING, &xprt->sockstate)) {
565 /* Note: if we are here due to a dropped connection
566 * we delay reconnecting by RPC_REESTABLISH_TIMEOUT/HZ
569 if (xprt->sock != NULL)
570 schedule_delayed_work(&xprt->sock_connect,
571 RPC_REESTABLISH_TIMEOUT);
573 schedule_work(&xprt->sock_connect);
577 xprt_release_write(xprt, task);
581 * We arrive here when awoken from waiting on connection establishment.
584 xprt_connect_status(struct rpc_task *task)
586 struct rpc_xprt *xprt = task->tk_xprt;
588 if (task->tk_status >= 0) {
589 dprintk("RPC: %4d xprt_connect_status: connection established\n",
594 /* if soft mounted, just cause this RPC to fail */
595 if (RPC_IS_SOFT(task))
596 task->tk_status = -EIO;
598 switch (task->tk_status) {
604 dprintk("RPC: %4d xprt_connect_status: timed out\n",
608 printk(KERN_ERR "RPC: error %d connecting to server %s\n",
609 -task->tk_status, task->tk_client->cl_server);
611 xprt_release_write(xprt, task);
615 * Look up the RPC request corresponding to a reply, and then lock it.
617 static inline struct rpc_rqst *
618 xprt_lookup_rqst(struct rpc_xprt *xprt, u32 xid)
620 struct list_head *pos;
621 struct rpc_rqst *req = NULL;
623 list_for_each(pos, &xprt->recv) {
624 struct rpc_rqst *entry = list_entry(pos, struct rpc_rqst, rq_list);
625 if (entry->rq_xid == xid) {
634 * Complete reply received.
635 * The TCP code relies on us to remove the request from xprt->pending.
638 xprt_complete_rqst(struct rpc_xprt *xprt, struct rpc_rqst *req, int copied)
640 struct rpc_task *task = req->rq_task;
641 struct rpc_clnt *clnt = task->tk_client;
643 /* Adjust congestion window */
645 unsigned timer = task->tk_msg.rpc_proc->p_timer;
646 xprt_adjust_cwnd(xprt, copied);
647 __xprt_put_cong(xprt, req);
649 if (req->rq_ntrans == 1)
650 rpc_update_rtt(clnt->cl_rtt, timer,
651 (long)jiffies - req->rq_xtime);
652 rpc_set_timeo(clnt->cl_rtt, timer, req->rq_ntrans - 1);
657 /* Profile only reads for now */
659 static unsigned long nextstat;
660 static unsigned long pkt_rtt, pkt_len, pkt_cnt;
663 pkt_len += req->rq_slen + copied;
664 pkt_rtt += jiffies - req->rq_xtime;
665 if (time_before(nextstat, jiffies)) {
666 printk("RPC: %lu %ld cwnd\n", jiffies, xprt->cwnd);
667 printk("RPC: %ld %ld %ld %ld stat\n",
668 jiffies, pkt_cnt, pkt_len, pkt_rtt);
669 pkt_rtt = pkt_len = pkt_cnt = 0;
670 nextstat = jiffies + 5 * HZ;
675 dprintk("RPC: %4d has input (%d bytes)\n", task->tk_pid, copied);
676 list_del_init(&req->rq_list);
677 req->rq_received = req->rq_private_buf.len = copied;
679 /* ... and wake up the process. */
680 rpc_wake_up_task(task);
685 skb_read_bits(skb_reader_t *desc, void *to, size_t len)
687 if (len > desc->count)
689 if (skb_copy_bits(desc->skb, desc->offset, to, len))
697 skb_read_and_csum_bits(skb_reader_t *desc, void *to, size_t len)
699 unsigned int csum2, pos;
701 if (len > desc->count)
704 csum2 = skb_copy_and_csum_bits(desc->skb, pos, to, len, 0);
705 desc->csum = csum_block_add(desc->csum, csum2, pos);
712 * We have set things up such that we perform the checksum of the UDP
713 * packet in parallel with the copies into the RPC client iovec. -DaveM
716 csum_partial_copy_to_xdr(struct xdr_buf *xdr, struct sk_buff *skb)
721 desc.offset = sizeof(struct udphdr);
722 desc.count = skb->len - desc.offset;
724 if (skb->ip_summed == CHECKSUM_UNNECESSARY)
727 desc.csum = csum_partial(skb->data, desc.offset, skb->csum);
728 if (xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_and_csum_bits) < 0)
730 if (desc.offset != skb->len) {
732 csum2 = skb_checksum(skb, desc.offset, skb->len - desc.offset, 0);
733 desc.csum = csum_block_add(desc.csum, csum2, desc.offset);
737 if ((unsigned short)csum_fold(desc.csum))
741 if (xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_bits) < 0)
749 * Input handler for RPC replies. Called from a bottom half and hence
753 udp_data_ready(struct sock *sk, int len)
755 struct rpc_task *task;
756 struct rpc_xprt *xprt;
757 struct rpc_rqst *rovr;
759 int err, repsize, copied;
762 read_lock(&sk->sk_callback_lock);
763 dprintk("RPC: udp_data_ready...\n");
764 if (!(xprt = xprt_from_sock(sk))) {
765 printk("RPC: udp_data_ready request not found!\n");
769 dprintk("RPC: udp_data_ready client %p\n", xprt);
771 if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
777 repsize = skb->len - sizeof(struct udphdr);
779 printk("RPC: impossible RPC reply size %d!\n", repsize);
783 /* Copy the XID from the skb... */
784 xp = skb_header_pointer(skb, sizeof(struct udphdr),
785 sizeof(_xid), &_xid);
789 /* Look up and lock the request corresponding to the given XID */
790 spin_lock(&xprt->sock_lock);
791 rovr = xprt_lookup_rqst(xprt, *xp);
794 task = rovr->rq_task;
796 dprintk("RPC: %4d received reply\n", task->tk_pid);
798 if ((copied = rovr->rq_private_buf.buflen) > repsize)
801 /* Suck it into the iovec, verify checksum if not done by hw. */
802 if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
805 /* Something worked... */
806 dst_confirm(skb->dst);
808 xprt_complete_rqst(xprt, rovr, copied);
811 spin_unlock(&xprt->sock_lock);
813 skb_free_datagram(sk, skb);
815 read_unlock(&sk->sk_callback_lock);
819 * Copy from an skb into memory and shrink the skb.
822 tcp_copy_data(skb_reader_t *desc, void *p, size_t len)
824 if (len > desc->count)
826 if (skb_copy_bits(desc->skb, desc->offset, p, len)) {
827 dprintk("RPC: failed to copy %zu bytes from skb. %zu bytes remain\n",
833 dprintk("RPC: copied %zu bytes from skb. %zu bytes remain\n",
839 * TCP read fragment marker
842 tcp_read_fraghdr(struct rpc_xprt *xprt, skb_reader_t *desc)
847 p = ((char *) &xprt->tcp_recm) + xprt->tcp_offset;
848 len = sizeof(xprt->tcp_recm) - xprt->tcp_offset;
849 used = tcp_copy_data(desc, p, len);
850 xprt->tcp_offset += used;
853 xprt->tcp_reclen = ntohl(xprt->tcp_recm);
854 if (xprt->tcp_reclen & 0x80000000)
855 xprt->tcp_flags |= XPRT_LAST_FRAG;
857 xprt->tcp_flags &= ~XPRT_LAST_FRAG;
858 xprt->tcp_reclen &= 0x7fffffff;
859 xprt->tcp_flags &= ~XPRT_COPY_RECM;
860 xprt->tcp_offset = 0;
861 /* Sanity check of the record length */
862 if (xprt->tcp_reclen < 4) {
863 printk(KERN_ERR "RPC: Invalid TCP record fragment length\n");
864 xprt_disconnect(xprt);
866 dprintk("RPC: reading TCP record fragment of length %d\n",
871 tcp_check_recm(struct rpc_xprt *xprt)
873 dprintk("RPC: xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u, tcp_flags = %lx\n",
874 xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen, xprt->tcp_flags);
875 if (xprt->tcp_offset == xprt->tcp_reclen) {
876 xprt->tcp_flags |= XPRT_COPY_RECM;
877 xprt->tcp_offset = 0;
878 if (xprt->tcp_flags & XPRT_LAST_FRAG) {
879 xprt->tcp_flags &= ~XPRT_COPY_DATA;
880 xprt->tcp_flags |= XPRT_COPY_XID;
881 xprt->tcp_copied = 0;
890 tcp_read_xid(struct rpc_xprt *xprt, skb_reader_t *desc)
895 len = sizeof(xprt->tcp_xid) - xprt->tcp_offset;
896 dprintk("RPC: reading XID (%Zu bytes)\n", len);
897 p = ((char *) &xprt->tcp_xid) + xprt->tcp_offset;
898 used = tcp_copy_data(desc, p, len);
899 xprt->tcp_offset += used;
902 xprt->tcp_flags &= ~XPRT_COPY_XID;
903 xprt->tcp_flags |= XPRT_COPY_DATA;
904 xprt->tcp_copied = 4;
905 dprintk("RPC: reading reply for XID %08x\n",
906 ntohl(xprt->tcp_xid));
907 tcp_check_recm(xprt);
911 * TCP read and complete request
914 tcp_read_request(struct rpc_xprt *xprt, skb_reader_t *desc)
916 struct rpc_rqst *req;
917 struct xdr_buf *rcvbuf;
921 /* Find and lock the request corresponding to this xid */
922 spin_lock(&xprt->sock_lock);
923 req = xprt_lookup_rqst(xprt, xprt->tcp_xid);
925 xprt->tcp_flags &= ~XPRT_COPY_DATA;
926 dprintk("RPC: XID %08x request not found!\n",
927 ntohl(xprt->tcp_xid));
928 spin_unlock(&xprt->sock_lock);
932 rcvbuf = &req->rq_private_buf;
934 if (len > xprt->tcp_reclen - xprt->tcp_offset) {
935 skb_reader_t my_desc;
937 len = xprt->tcp_reclen - xprt->tcp_offset;
938 memcpy(&my_desc, desc, sizeof(my_desc));
940 r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
941 &my_desc, tcp_copy_data);
945 r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
946 desc, tcp_copy_data);
949 xprt->tcp_copied += r;
950 xprt->tcp_offset += r;
953 /* Error when copying to the receive buffer,
954 * usually because we weren't able to allocate
955 * additional buffer pages. All we can do now
956 * is turn off XPRT_COPY_DATA, so the request
957 * will not receive any additional updates,
959 * Any remaining data from this record will
962 xprt->tcp_flags &= ~XPRT_COPY_DATA;
963 dprintk("RPC: XID %08x truncated request\n",
964 ntohl(xprt->tcp_xid));
965 dprintk("RPC: xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
966 xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen);
970 dprintk("RPC: XID %08x read %u bytes\n",
971 ntohl(xprt->tcp_xid), r);
972 dprintk("RPC: xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
973 xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen);
975 if (xprt->tcp_copied == req->rq_private_buf.buflen)
976 xprt->tcp_flags &= ~XPRT_COPY_DATA;
977 else if (xprt->tcp_offset == xprt->tcp_reclen) {
978 if (xprt->tcp_flags & XPRT_LAST_FRAG)
979 xprt->tcp_flags &= ~XPRT_COPY_DATA;
983 if (!(xprt->tcp_flags & XPRT_COPY_DATA)) {
984 dprintk("RPC: %4d received reply complete\n",
985 req->rq_task->tk_pid);
986 xprt_complete_rqst(xprt, req, xprt->tcp_copied);
988 spin_unlock(&xprt->sock_lock);
989 tcp_check_recm(xprt);
993 * TCP discard extra bytes from a short read
996 tcp_read_discard(struct rpc_xprt *xprt, skb_reader_t *desc)
1000 len = xprt->tcp_reclen - xprt->tcp_offset;
1001 if (len > desc->count)
1004 desc->offset += len;
1005 xprt->tcp_offset += len;
1006 dprintk("RPC: discarded %u bytes\n", len);
1007 tcp_check_recm(xprt);
1011 * TCP record receive routine
1012 * We first have to grab the record marker, then the XID, then the data.
1015 tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
1016 unsigned int offset, size_t len)
1018 struct rpc_xprt *xprt = rd_desc->arg.data;
1019 skb_reader_t desc = {
1026 dprintk("RPC: tcp_data_recv\n");
1028 /* Read in a new fragment marker if necessary */
1029 /* Can we ever really expect to get completely empty fragments? */
1030 if (xprt->tcp_flags & XPRT_COPY_RECM) {
1031 tcp_read_fraghdr(xprt, &desc);
1034 /* Read in the xid if necessary */
1035 if (xprt->tcp_flags & XPRT_COPY_XID) {
1036 tcp_read_xid(xprt, &desc);
1039 /* Read in the request data */
1040 if (xprt->tcp_flags & XPRT_COPY_DATA) {
1041 tcp_read_request(xprt, &desc);
1044 /* Skip over any trailing bytes on short reads */
1045 tcp_read_discard(xprt, &desc);
1046 } while (desc.count);
1047 dprintk("RPC: tcp_data_recv done\n");
1048 return len - desc.count;
1051 static void tcp_data_ready(struct sock *sk, int bytes)
1053 struct rpc_xprt *xprt;
1054 read_descriptor_t rd_desc;
1056 read_lock(&sk->sk_callback_lock);
1057 dprintk("RPC: tcp_data_ready...\n");
1058 if (!(xprt = xprt_from_sock(sk))) {
1059 printk("RPC: tcp_data_ready socket info not found!\n");
1065 /* We use rd_desc to pass struct xprt to tcp_data_recv */
1066 rd_desc.arg.data = xprt;
1067 rd_desc.count = 65536;
1068 tcp_read_sock(sk, &rd_desc, tcp_data_recv);
1070 read_unlock(&sk->sk_callback_lock);
1074 tcp_state_change(struct sock *sk)
1076 struct rpc_xprt *xprt;
1078 read_lock(&sk->sk_callback_lock);
1079 if (!(xprt = xprt_from_sock(sk)))
1081 dprintk("RPC: tcp_state_change client %p...\n", xprt);
1082 dprintk("RPC: state %x conn %d dead %d zapped %d\n",
1083 sk->sk_state, xprt_connected(xprt),
1084 sock_flag(sk, SOCK_DEAD),
1085 sock_flag(sk, SOCK_ZAPPED));
1087 switch (sk->sk_state) {
1088 case TCP_ESTABLISHED:
1089 spin_lock_bh(&xprt->sock_lock);
1090 if (!xprt_test_and_set_connected(xprt)) {
1091 /* Reset TCP record info */
1092 xprt->tcp_offset = 0;
1093 xprt->tcp_reclen = 0;
1094 xprt->tcp_copied = 0;
1095 xprt->tcp_flags = XPRT_COPY_RECM | XPRT_COPY_XID;
1096 rpc_wake_up(&xprt->pending);
1098 spin_unlock_bh(&xprt->sock_lock);
1104 xprt_disconnect(xprt);
1108 read_unlock(&sk->sk_callback_lock);
1112 * Called when more output buffer space is available for this socket.
1113 * We try not to wake our writers until they can make "significant"
1114 * progress, otherwise we'll waste resources thrashing sock_sendmsg
1115 * with a bunch of small requests.
1118 xprt_write_space(struct sock *sk)
1120 struct rpc_xprt *xprt;
1121 struct socket *sock;
1123 read_lock(&sk->sk_callback_lock);
1124 if (!(xprt = xprt_from_sock(sk)) || !(sock = sk->sk_socket))
1129 /* Wait until we have enough socket memory */
1131 /* from net/core/stream.c:sk_stream_write_space */
1132 if (sk_stream_wspace(sk) < sk_stream_min_wspace(sk))
1135 /* from net/core/sock.c:sock_def_write_space */
1136 if (!sock_writeable(sk))
1140 if (!test_and_clear_bit(SOCK_NOSPACE, &sock->flags))
1143 spin_lock_bh(&xprt->sock_lock);
1145 rpc_wake_up_task(xprt->snd_task);
1146 spin_unlock_bh(&xprt->sock_lock);
1148 read_unlock(&sk->sk_callback_lock);
1152 * RPC receive timeout handler.
1155 xprt_timer(struct rpc_task *task)
1157 struct rpc_rqst *req = task->tk_rqstp;
1158 struct rpc_xprt *xprt = req->rq_xprt;
1160 spin_lock(&xprt->sock_lock);
1161 if (req->rq_received)
1164 xprt_adjust_cwnd(req->rq_xprt, -ETIMEDOUT);
1165 __xprt_put_cong(xprt, req);
1167 dprintk("RPC: %4d xprt_timer (%s request)\n",
1168 task->tk_pid, req ? "pending" : "backlogged");
1170 task->tk_status = -ETIMEDOUT;
1172 task->tk_timeout = 0;
1173 rpc_wake_up_task(task);
1174 spin_unlock(&xprt->sock_lock);
1178 * Place the actual RPC call.
1179 * We have to copy the iovec because sendmsg fiddles with its contents.
1182 xprt_prepare_transmit(struct rpc_task *task)
1184 struct rpc_rqst *req = task->tk_rqstp;
1185 struct rpc_xprt *xprt = req->rq_xprt;
1188 dprintk("RPC: %4d xprt_prepare_transmit\n", task->tk_pid);
1193 spin_lock_bh(&xprt->sock_lock);
1194 if (req->rq_received && !req->rq_bytes_sent) {
1195 err = req->rq_received;
1198 if (!__xprt_lock_write(xprt, task)) {
1203 if (!xprt_connected(xprt)) {
1208 spin_unlock_bh(&xprt->sock_lock);
1213 xprt_transmit(struct rpc_task *task)
1215 struct rpc_clnt *clnt = task->tk_client;
1216 struct rpc_rqst *req = task->tk_rqstp;
1217 struct rpc_xprt *xprt = req->rq_xprt;
1218 int status, retry = 0;
1221 dprintk("RPC: %4d xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
1223 /* set up everything as needed. */
1224 /* Write the record marker */
1226 u32 *marker = req->rq_svec[0].iov_base;
1228 *marker = htonl(0x80000000|(req->rq_slen-sizeof(*marker)));
1232 if (!req->rq_received) {
1233 if (list_empty(&req->rq_list)) {
1234 spin_lock_bh(&xprt->sock_lock);
1235 /* Update the softirq receive buffer */
1236 memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
1237 sizeof(req->rq_private_buf));
1238 /* Add request to the receive list */
1239 list_add_tail(&req->rq_list, &xprt->recv);
1240 spin_unlock_bh(&xprt->sock_lock);
1241 xprt_reset_majortimeo(req);
1242 /* Turn off autodisconnect */
1243 del_singleshot_timer_sync(&xprt->timer);
1245 } else if (!req->rq_bytes_sent)
1248 /* Continue transmitting the packet/record. We must be careful
1249 * to cope with writespace callbacks arriving _after_ we have
1250 * called xprt_sendmsg().
1253 req->rq_xtime = jiffies;
1254 status = xprt_sendmsg(xprt, req);
1260 req->rq_bytes_sent += status;
1262 /* If we've sent the entire packet, immediately
1263 * reset the count of bytes sent. */
1264 if (req->rq_bytes_sent >= req->rq_slen) {
1265 req->rq_bytes_sent = 0;
1269 if (status >= req->rq_slen)
1275 dprintk("RPC: %4d xmit incomplete (%d left of %d)\n",
1276 task->tk_pid, req->rq_slen - req->rq_bytes_sent,
1284 /* Note: at this point, task->tk_sleeping has not yet been set,
1285 * hence there is no danger of the waking up task being put on
1286 * schedq, and being picked up by a parallel run of rpciod().
1288 task->tk_status = status;
1292 if (test_bit(SOCK_ASYNC_NOSPACE, &xprt->sock->flags)) {
1293 /* Protect against races with xprt_write_space */
1294 spin_lock_bh(&xprt->sock_lock);
1295 /* Don't race with disconnect */
1296 if (!xprt_connected(xprt))
1297 task->tk_status = -ENOTCONN;
1298 else if (test_bit(SOCK_NOSPACE, &xprt->sock->flags)) {
1299 task->tk_timeout = req->rq_timeout;
1300 rpc_sleep_on(&xprt->pending, task, NULL, NULL);
1302 spin_unlock_bh(&xprt->sock_lock);
1305 /* Keep holding the socket if it is blocked */
1306 rpc_delay(task, HZ>>4);
1309 task->tk_timeout = RPC_REESTABLISH_TIMEOUT;
1310 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
1315 xprt_disconnect(xprt);
1317 xprt_release_write(xprt, task);
1320 dprintk("RPC: %4d xmit complete\n", task->tk_pid);
1321 /* Set the task's receive timeout value */
1322 spin_lock_bh(&xprt->sock_lock);
1323 if (!xprt->nocong) {
1324 int timer = task->tk_msg.rpc_proc->p_timer;
1325 task->tk_timeout = rpc_calc_rto(clnt->cl_rtt, timer);
1326 task->tk_timeout <<= rpc_ntimeo(clnt->cl_rtt, timer) + req->rq_retries;
1327 if (task->tk_timeout > xprt->timeout.to_maxval || task->tk_timeout == 0)
1328 task->tk_timeout = xprt->timeout.to_maxval;
1330 task->tk_timeout = req->rq_timeout;
1331 /* Don't race with disconnect */
1332 if (!xprt_connected(xprt))
1333 task->tk_status = -ENOTCONN;
1334 else if (!req->rq_received)
1335 rpc_sleep_on(&xprt->pending, task, NULL, xprt_timer);
1336 __xprt_release_write(xprt, task);
1337 spin_unlock_bh(&xprt->sock_lock);
1341 * Reserve an RPC call slot.
1344 do_xprt_reserve(struct rpc_task *task)
1346 struct rpc_xprt *xprt = task->tk_xprt;
1348 task->tk_status = 0;
1351 if (!list_empty(&xprt->free)) {
1352 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
1353 list_del_init(&req->rq_list);
1354 task->tk_rqstp = req;
1355 xprt_request_init(task, xprt);
1358 dprintk("RPC: waiting for request slot\n");
1359 task->tk_status = -EAGAIN;
1360 task->tk_timeout = 0;
1361 rpc_sleep_on(&xprt->backlog, task, NULL, NULL);
1365 xprt_reserve(struct rpc_task *task)
1367 struct rpc_xprt *xprt = task->tk_xprt;
1369 task->tk_status = -EIO;
1370 if (!xprt->shutdown) {
1371 spin_lock(&xprt->xprt_lock);
1372 do_xprt_reserve(task);
1373 spin_unlock(&xprt->xprt_lock);
1378 * Allocate a 'unique' XID
1380 static inline u32 xprt_alloc_xid(struct rpc_xprt *xprt)
1385 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1387 get_random_bytes(&xprt->xid, sizeof(xprt->xid));
1391 * Initialize RPC request
1394 xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1396 struct rpc_rqst *req = task->tk_rqstp;
1398 req->rq_timeout = xprt->timeout.to_initval;
1399 req->rq_task = task;
1400 req->rq_xprt = xprt;
1401 req->rq_xid = xprt_alloc_xid(xprt);
1402 dprintk("RPC: %4d reserved req %p xid %08x\n", task->tk_pid,
1403 req, ntohl(req->rq_xid));
1407 * Release an RPC call slot
1410 xprt_release(struct rpc_task *task)
1412 struct rpc_xprt *xprt = task->tk_xprt;
1413 struct rpc_rqst *req;
1415 if (!(req = task->tk_rqstp))
1417 spin_lock_bh(&xprt->sock_lock);
1418 __xprt_release_write(xprt, task);
1419 __xprt_put_cong(xprt, req);
1420 if (!list_empty(&req->rq_list))
1421 list_del(&req->rq_list);
1422 xprt->last_used = jiffies;
1423 if (list_empty(&xprt->recv) && !xprt->shutdown)
1424 mod_timer(&xprt->timer, xprt->last_used + XPRT_IDLE_TIMEOUT);
1425 spin_unlock_bh(&xprt->sock_lock);
1426 task->tk_rqstp = NULL;
1427 memset(req, 0, sizeof(*req)); /* mark unused */
1429 dprintk("RPC: %4d release request %p\n", task->tk_pid, req);
1431 spin_lock(&xprt->xprt_lock);
1432 list_add(&req->rq_list, &xprt->free);
1433 xprt_clear_backlog(xprt);
1434 spin_unlock(&xprt->xprt_lock);
1438 * Set default timeout parameters
1441 xprt_default_timeout(struct rpc_timeout *to, int proto)
1443 if (proto == IPPROTO_UDP)
1444 xprt_set_timeout(to, 5, 5 * HZ);
1446 xprt_set_timeout(to, 5, 60 * HZ);
1450 * Set constant timeout
1453 xprt_set_timeout(struct rpc_timeout *to, unsigned int retr, unsigned long incr)
1456 to->to_increment = incr;
1457 to->to_maxval = incr * retr;
1458 to->to_retries = retr;
1459 to->to_exponential = 0;
1462 unsigned int xprt_udp_slot_table_entries = RPC_DEF_SLOT_TABLE;
1463 unsigned int xprt_tcp_slot_table_entries = RPC_DEF_SLOT_TABLE;
1466 * Initialize an RPC client
1468 static struct rpc_xprt *
1469 xprt_setup(int proto, struct sockaddr_in *ap, struct rpc_timeout *to)
1471 struct rpc_xprt *xprt;
1472 unsigned int entries;
1473 size_t slot_table_size;
1474 struct rpc_rqst *req;
1476 dprintk("RPC: setting up %s transport...\n",
1477 proto == IPPROTO_UDP? "UDP" : "TCP");
1479 entries = (proto == IPPROTO_TCP)?
1480 xprt_tcp_slot_table_entries : xprt_udp_slot_table_entries;
1482 if ((xprt = kmalloc(sizeof(struct rpc_xprt), GFP_KERNEL)) == NULL)
1483 return ERR_PTR(-ENOMEM);
1484 memset(xprt, 0, sizeof(*xprt)); /* Nnnngh! */
1485 xprt->max_reqs = entries;
1486 slot_table_size = entries * sizeof(xprt->slot[0]);
1487 xprt->slot = kmalloc(slot_table_size, GFP_KERNEL);
1488 if (xprt->slot == NULL) {
1490 return ERR_PTR(-ENOMEM);
1492 memset(xprt->slot, 0, slot_table_size);
1496 xprt->stream = (proto == IPPROTO_TCP)? 1 : 0;
1498 xprt->cwnd = RPC_MAXCWND(xprt);
1500 xprt->max_payload = (1U << 31) - 1;
1502 xprt->cwnd = RPC_INITCWND;
1503 xprt->max_payload = (1U << 16) - (MAX_HEADER << 3);
1505 spin_lock_init(&xprt->sock_lock);
1506 spin_lock_init(&xprt->xprt_lock);
1507 init_waitqueue_head(&xprt->cong_wait);
1509 INIT_LIST_HEAD(&xprt->free);
1510 INIT_LIST_HEAD(&xprt->recv);
1511 INIT_WORK(&xprt->sock_connect, xprt_socket_connect, xprt);
1512 INIT_WORK(&xprt->task_cleanup, xprt_socket_autoclose, xprt);
1513 init_timer(&xprt->timer);
1514 xprt->timer.function = xprt_init_autodisconnect;
1515 xprt->timer.data = (unsigned long) xprt;
1516 xprt->last_used = jiffies;
1517 xprt->port = XPRT_MAX_RESVPORT;
1519 /* Set timeout parameters */
1521 xprt->timeout = *to;
1523 xprt_default_timeout(&xprt->timeout, xprt->prot);
1525 rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1526 rpc_init_wait_queue(&xprt->sending, "xprt_sending");
1527 rpc_init_wait_queue(&xprt->resend, "xprt_resend");
1528 rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1530 /* initialize free list */
1531 for (req = &xprt->slot[entries-1]; req >= &xprt->slot[0]; req--)
1532 list_add(&req->rq_list, &xprt->free);
1534 xprt_init_xid(xprt);
1536 /* Check whether we want to use a reserved port */
1537 xprt->resvport = capable(CAP_NET_BIND_SERVICE) ? 1 : 0;
1539 dprintk("RPC: created transport %p with %u slots\n", xprt,
1546 * Bind to a reserved port
1548 static inline int xprt_bindresvport(struct rpc_xprt *xprt, struct socket *sock)
1550 struct sockaddr_in myaddr = {
1551 .sin_family = AF_INET,
1555 /* Were we already bound to a given port? Try to reuse it */
1558 myaddr.sin_port = htons(port);
1559 err = sock->ops->bind(sock, (struct sockaddr *) &myaddr,
1566 port = XPRT_MAX_RESVPORT;
1567 } while (err == -EADDRINUSE && port != xprt->port);
1569 printk("RPC: Can't bind to reserved port (%d).\n", -err);
1574 xprt_bind_socket(struct rpc_xprt *xprt, struct socket *sock)
1576 struct sock *sk = sock->sk;
1581 write_lock_bh(&sk->sk_callback_lock);
1582 sk->sk_user_data = xprt;
1583 xprt->old_data_ready = sk->sk_data_ready;
1584 xprt->old_state_change = sk->sk_state_change;
1585 xprt->old_write_space = sk->sk_write_space;
1586 if (xprt->prot == IPPROTO_UDP) {
1587 sk->sk_data_ready = udp_data_ready;
1588 sk->sk_no_check = UDP_CSUM_NORCV;
1589 xprt_set_connected(xprt);
1591 tcp_sk(sk)->nonagle = 1; /* disable Nagle's algorithm */
1592 sk->sk_data_ready = tcp_data_ready;
1593 sk->sk_state_change = tcp_state_change;
1594 xprt_clear_connected(xprt);
1596 sk->sk_write_space = xprt_write_space;
1598 /* Reset to new socket */
1601 write_unlock_bh(&sk->sk_callback_lock);
1607 * Set socket buffer length
1610 xprt_sock_setbufsize(struct rpc_xprt *xprt)
1612 struct sock *sk = xprt->inet;
1616 if (xprt->rcvsize) {
1617 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
1618 sk->sk_rcvbuf = xprt->rcvsize * xprt->max_reqs * 2;
1620 if (xprt->sndsize) {
1621 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
1622 sk->sk_sndbuf = xprt->sndsize * xprt->max_reqs * 2;
1623 sk->sk_write_space(sk);
1628 * Datastream sockets are created here, but xprt_connect will create
1629 * and connect stream sockets.
1631 static struct socket * xprt_create_socket(struct rpc_xprt *xprt, int proto, int resvport)
1633 struct socket *sock;
1636 dprintk("RPC: xprt_create_socket(%s %d)\n",
1637 (proto == IPPROTO_UDP)? "udp" : "tcp", proto);
1639 type = (proto == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1641 if ((err = sock_create_kern(PF_INET, type, proto, &sock)) < 0) {
1642 printk("RPC: can't create socket (%d).\n", -err);
1646 /* If the caller has the capability, bind to a reserved port */
1647 if (resvport && xprt_bindresvport(xprt, sock) < 0) {
1648 printk("RPC: can't bind to reserved port.\n");
1660 * Create an RPC client transport given the protocol and peer address.
1663 xprt_create_proto(int proto, struct sockaddr_in *sap, struct rpc_timeout *to)
1665 struct rpc_xprt *xprt;
1667 xprt = xprt_setup(proto, sap, to);
1669 dprintk("RPC: xprt_create_proto failed\n");
1671 dprintk("RPC: xprt_create_proto created xprt %p\n", xprt);
1676 * Prepare for transport shutdown.
1679 xprt_shutdown(struct rpc_xprt *xprt)
1682 rpc_wake_up(&xprt->sending);
1683 rpc_wake_up(&xprt->resend);
1684 rpc_wake_up(&xprt->pending);
1685 rpc_wake_up(&xprt->backlog);
1686 wake_up(&xprt->cong_wait);
1687 del_timer_sync(&xprt->timer);
1691 * Clear the xprt backlog queue
1694 xprt_clear_backlog(struct rpc_xprt *xprt) {
1695 rpc_wake_up_next(&xprt->backlog);
1696 wake_up(&xprt->cong_wait);
1701 * Destroy an RPC transport, killing off all requests.
1704 xprt_destroy(struct rpc_xprt *xprt)
1706 dprintk("RPC: destroying transport %p\n", xprt);
1707 xprt_shutdown(xprt);
1708 xprt_disconnect(xprt);