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 *);
95 * Print the buffer contents (first 128 bytes only--just enough for
99 xprt_pktdump(char *msg, u32 *packet, unsigned int count)
101 u8 *buf = (u8 *) packet;
104 dprintk("RPC: %s\n", msg);
105 for (j = 0; j < count && j < 128; j += 4) {
109 dprintk("0x%04x ", j);
111 dprintk("%02x%02x%02x%02x ",
112 buf[j], buf[j+1], buf[j+2], buf[j+3]);
118 xprt_pktdump(char *msg, u32 *packet, unsigned int count)
125 * Look up RPC transport given an INET socket
127 static inline struct rpc_xprt *
128 xprt_from_sock(struct sock *sk)
130 return (struct rpc_xprt *) sk->sk_user_data;
134 * Serialize write access to sockets, in order to prevent different
135 * requests from interfering with each other.
136 * Also prevents TCP socket connects from colliding with writes.
139 __xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
141 struct rpc_rqst *req = task->tk_rqstp;
143 if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate)) {
144 if (task == xprt->snd_task)
150 if (xprt->nocong || __xprt_get_cong(xprt, task)) {
151 xprt->snd_task = task;
153 req->rq_bytes_sent = 0;
158 smp_mb__before_clear_bit();
159 clear_bit(XPRT_LOCKED, &xprt->sockstate);
160 smp_mb__after_clear_bit();
162 dprintk("RPC: %4d failed to lock socket %p\n", task->tk_pid, xprt);
163 task->tk_timeout = 0;
164 task->tk_status = -EAGAIN;
165 if (req && req->rq_ntrans)
166 rpc_sleep_on(&xprt->resend, task, NULL, NULL);
168 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
173 xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
177 spin_lock_bh(&xprt->sock_lock);
178 retval = __xprt_lock_write(xprt, task);
179 spin_unlock_bh(&xprt->sock_lock);
185 __xprt_lock_write_next(struct rpc_xprt *xprt)
187 struct rpc_task *task;
189 if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate))
191 if (!xprt->nocong && RPCXPRT_CONGESTED(xprt))
193 task = rpc_wake_up_next(&xprt->resend);
195 task = rpc_wake_up_next(&xprt->sending);
199 if (xprt->nocong || __xprt_get_cong(xprt, task)) {
200 struct rpc_rqst *req = task->tk_rqstp;
201 xprt->snd_task = task;
203 req->rq_bytes_sent = 0;
209 smp_mb__before_clear_bit();
210 clear_bit(XPRT_LOCKED, &xprt->sockstate);
211 smp_mb__after_clear_bit();
215 * Releases the socket for use by other requests.
218 __xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
220 if (xprt->snd_task == task) {
221 xprt->snd_task = NULL;
222 smp_mb__before_clear_bit();
223 clear_bit(XPRT_LOCKED, &xprt->sockstate);
224 smp_mb__after_clear_bit();
225 __xprt_lock_write_next(xprt);
230 xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
232 spin_lock_bh(&xprt->sock_lock);
233 __xprt_release_write(xprt, task);
234 spin_unlock_bh(&xprt->sock_lock);
238 * Write data to socket.
241 xprt_sendmsg(struct rpc_xprt *xprt, struct rpc_rqst *req)
243 struct socket *sock = xprt->sock;
244 struct xdr_buf *xdr = &req->rq_snd_buf;
245 struct sockaddr *addr = NULL;
253 xprt_pktdump("packet data:",
254 req->rq_svec->iov_base,
255 req->rq_svec->iov_len);
257 /* For UDP, we need to provide an address */
259 addr = (struct sockaddr *) &xprt->addr;
260 addrlen = sizeof(xprt->addr);
262 /* Dont repeat bytes */
263 skip = req->rq_bytes_sent;
265 clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
266 result = xdr_sendpages(sock, addr, addrlen, xdr, skip, MSG_DONTWAIT);
268 dprintk("RPC: xprt_sendmsg(%d) = %d\n", xdr->len - skip, result);
275 /* When the server has died, an ICMP port unreachable message
276 * prompts ECONNREFUSED.
283 /* connection broken */
288 printk(KERN_NOTICE "RPC: sendmsg returned error %d\n", -result);
294 * Van Jacobson congestion avoidance. Check if the congestion window
295 * overflowed. Put the task to sleep if this is the case.
298 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
300 struct rpc_rqst *req = task->tk_rqstp;
304 dprintk("RPC: %4d xprt_cwnd_limited cong = %ld cwnd = %ld\n",
305 task->tk_pid, xprt->cong, xprt->cwnd);
306 if (RPCXPRT_CONGESTED(xprt))
309 xprt->cong += RPC_CWNDSCALE;
314 * Adjust the congestion window, and wake up the next task
315 * that has been sleeping due to congestion
318 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
323 xprt->cong -= RPC_CWNDSCALE;
324 __xprt_lock_write_next(xprt);
328 * Adjust RPC congestion window
329 * We use a time-smoothed congestion estimator to avoid heavy oscillation.
332 xprt_adjust_cwnd(struct rpc_xprt *xprt, int result)
337 if (result >= 0 && cwnd <= xprt->cong) {
338 /* The (cwnd >> 1) term makes sure
339 * the result gets rounded properly. */
340 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
341 if (cwnd > RPC_MAXCWND(xprt))
342 cwnd = RPC_MAXCWND(xprt);
343 __xprt_lock_write_next(xprt);
344 } else if (result == -ETIMEDOUT) {
346 if (cwnd < RPC_CWNDSCALE)
347 cwnd = RPC_CWNDSCALE;
349 dprintk("RPC: cong %ld, cwnd was %ld, now %ld\n",
350 xprt->cong, xprt->cwnd, cwnd);
355 * Reset the major timeout value
357 static void xprt_reset_majortimeo(struct rpc_rqst *req)
359 struct rpc_timeout *to = &req->rq_xprt->timeout;
361 req->rq_majortimeo = req->rq_timeout;
362 if (to->to_exponential)
363 req->rq_majortimeo <<= to->to_retries;
365 req->rq_majortimeo += to->to_increment * to->to_retries;
366 if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
367 req->rq_majortimeo = to->to_maxval;
368 req->rq_majortimeo += jiffies;
372 * Adjust timeout values etc for next retransmit
374 int xprt_adjust_timeout(struct rpc_rqst *req)
376 struct rpc_xprt *xprt = req->rq_xprt;
377 struct rpc_timeout *to = &xprt->timeout;
380 if (time_before(jiffies, req->rq_majortimeo)) {
381 if (to->to_exponential)
382 req->rq_timeout <<= 1;
384 req->rq_timeout += to->to_increment;
385 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
386 req->rq_timeout = to->to_maxval;
388 pprintk("RPC: %lu retrans\n", jiffies);
390 req->rq_timeout = to->to_initval;
392 xprt_reset_majortimeo(req);
393 /* Reset the RTT counters == "slow start" */
394 spin_lock_bh(&xprt->sock_lock);
395 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
396 spin_unlock_bh(&xprt->sock_lock);
397 pprintk("RPC: %lu timeout\n", jiffies);
401 if (req->rq_timeout == 0) {
402 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
403 req->rq_timeout = 5 * HZ;
409 * Close down a transport socket
412 xprt_close(struct rpc_xprt *xprt)
414 struct socket *sock = xprt->sock;
415 struct sock *sk = xprt->inet;
420 write_lock_bh(&sk->sk_callback_lock);
424 sk->sk_user_data = NULL;
425 sk->sk_data_ready = xprt->old_data_ready;
426 sk->sk_state_change = xprt->old_state_change;
427 sk->sk_write_space = xprt->old_write_space;
428 write_unlock_bh(&sk->sk_callback_lock);
436 xprt_socket_autoclose(void *args)
438 struct rpc_xprt *xprt = (struct rpc_xprt *)args;
440 xprt_disconnect(xprt);
442 xprt_release_write(xprt, NULL);
446 * Mark a transport as disconnected
449 xprt_disconnect(struct rpc_xprt *xprt)
451 dprintk("RPC: disconnected transport %p\n", xprt);
452 spin_lock_bh(&xprt->sock_lock);
453 xprt_clear_connected(xprt);
454 rpc_wake_up_status(&xprt->pending, -ENOTCONN);
455 spin_unlock_bh(&xprt->sock_lock);
459 * Used to allow disconnection when we've been idle
462 xprt_init_autodisconnect(unsigned long data)
464 struct rpc_xprt *xprt = (struct rpc_xprt *)data;
466 spin_lock(&xprt->sock_lock);
467 if (!list_empty(&xprt->recv) || xprt->shutdown)
469 if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate))
471 spin_unlock(&xprt->sock_lock);
472 /* Let keventd close the socket */
473 if (test_bit(XPRT_CONNECTING, &xprt->sockstate) != 0)
474 xprt_release_write(xprt, NULL);
476 schedule_work(&xprt->task_cleanup);
479 spin_unlock(&xprt->sock_lock);
482 static void xprt_socket_connect(void *args)
484 struct rpc_xprt *xprt = (struct rpc_xprt *)args;
485 struct socket *sock = xprt->sock;
488 if (xprt->shutdown || xprt->addr.sin_port == 0)
492 * Start by resetting any existing state
495 sock = xprt_create_socket(xprt, xprt->prot, xprt->resvport);
497 /* couldn't create socket or bind to reserved port;
498 * this is likely a permanent error, so cause an abort */
501 xprt_bind_socket(xprt, sock);
502 xprt_sock_setbufsize(xprt);
509 * Tell the socket layer to start connecting...
511 status = sock->ops->connect(sock, (struct sockaddr *) &xprt->addr,
512 sizeof(xprt->addr), O_NONBLOCK);
513 dprintk("RPC: %p connect status %d connected %d sock state %d\n",
514 xprt, -status, xprt_connected(xprt), sock->sk->sk_state);
524 rpc_wake_up_status(&xprt->pending, status);
526 rpc_wake_up(&xprt->pending);
528 smp_mb__before_clear_bit();
529 clear_bit(XPRT_CONNECTING, &xprt->sockstate);
530 smp_mb__after_clear_bit();
534 * Attempt to connect a TCP socket.
537 void xprt_connect(struct rpc_task *task)
539 struct rpc_xprt *xprt = task->tk_xprt;
541 dprintk("RPC: %4d xprt_connect xprt %p %s connected\n", task->tk_pid,
542 xprt, (xprt_connected(xprt) ? "is" : "is not"));
544 if (xprt->shutdown) {
545 task->tk_status = -EIO;
548 if (!xprt->addr.sin_port) {
549 task->tk_status = -EIO;
552 if (!xprt_lock_write(xprt, task))
554 if (xprt_connected(xprt))
558 task->tk_rqstp->rq_bytes_sent = 0;
560 task->tk_timeout = RPC_CONNECT_TIMEOUT;
561 rpc_sleep_on(&xprt->pending, task, xprt_connect_status, NULL);
562 if (!test_and_set_bit(XPRT_CONNECTING, &xprt->sockstate)) {
563 /* Note: if we are here due to a dropped connection
564 * we delay reconnecting by RPC_REESTABLISH_TIMEOUT/HZ
567 if (xprt->sock != NULL)
568 schedule_delayed_work(&xprt->sock_connect,
569 RPC_REESTABLISH_TIMEOUT);
571 schedule_work(&xprt->sock_connect);
575 xprt_release_write(xprt, task);
579 * We arrive here when awoken from waiting on connection establishment.
582 xprt_connect_status(struct rpc_task *task)
584 struct rpc_xprt *xprt = task->tk_xprt;
586 if (task->tk_status >= 0) {
587 dprintk("RPC: %4d xprt_connect_status: connection established\n",
592 /* if soft mounted, just cause this RPC to fail */
593 if (RPC_IS_SOFT(task))
594 task->tk_status = -EIO;
596 switch (task->tk_status) {
602 dprintk("RPC: %4d xprt_connect_status: timed out\n",
606 printk(KERN_ERR "RPC: error %d connecting to server %s\n",
607 -task->tk_status, task->tk_client->cl_server);
609 xprt_release_write(xprt, task);
613 * Look up the RPC request corresponding to a reply, and then lock it.
615 static inline struct rpc_rqst *
616 xprt_lookup_rqst(struct rpc_xprt *xprt, u32 xid)
618 struct list_head *pos;
619 struct rpc_rqst *req = NULL;
621 list_for_each(pos, &xprt->recv) {
622 struct rpc_rqst *entry = list_entry(pos, struct rpc_rqst, rq_list);
623 if (entry->rq_xid == xid) {
632 * Complete reply received.
633 * The TCP code relies on us to remove the request from xprt->pending.
636 xprt_complete_rqst(struct rpc_xprt *xprt, struct rpc_rqst *req, int copied)
638 struct rpc_task *task = req->rq_task;
639 struct rpc_clnt *clnt = task->tk_client;
641 /* Adjust congestion window */
643 unsigned timer = task->tk_msg.rpc_proc->p_timer;
644 xprt_adjust_cwnd(xprt, copied);
645 __xprt_put_cong(xprt, req);
647 if (req->rq_ntrans == 1)
648 rpc_update_rtt(clnt->cl_rtt, timer,
649 (long)jiffies - req->rq_xtime);
650 rpc_set_timeo(clnt->cl_rtt, timer, req->rq_ntrans - 1);
655 /* Profile only reads for now */
657 static unsigned long nextstat;
658 static unsigned long pkt_rtt, pkt_len, pkt_cnt;
661 pkt_len += req->rq_slen + copied;
662 pkt_rtt += jiffies - req->rq_xtime;
663 if (time_before(nextstat, jiffies)) {
664 printk("RPC: %lu %ld cwnd\n", jiffies, xprt->cwnd);
665 printk("RPC: %ld %ld %ld %ld stat\n",
666 jiffies, pkt_cnt, pkt_len, pkt_rtt);
667 pkt_rtt = pkt_len = pkt_cnt = 0;
668 nextstat = jiffies + 5 * HZ;
673 dprintk("RPC: %4d has input (%d bytes)\n", task->tk_pid, copied);
674 list_del_init(&req->rq_list);
675 req->rq_received = req->rq_private_buf.len = copied;
677 /* ... and wake up the process. */
678 rpc_wake_up_task(task);
683 skb_read_bits(skb_reader_t *desc, void *to, size_t len)
685 if (len > desc->count)
687 if (skb_copy_bits(desc->skb, desc->offset, to, len))
695 skb_read_and_csum_bits(skb_reader_t *desc, void *to, size_t len)
697 unsigned int csum2, pos;
699 if (len > desc->count)
702 csum2 = skb_copy_and_csum_bits(desc->skb, pos, to, len, 0);
703 desc->csum = csum_block_add(desc->csum, csum2, pos);
710 * We have set things up such that we perform the checksum of the UDP
711 * packet in parallel with the copies into the RPC client iovec. -DaveM
714 csum_partial_copy_to_xdr(struct xdr_buf *xdr, struct sk_buff *skb)
719 desc.offset = sizeof(struct udphdr);
720 desc.count = skb->len - desc.offset;
722 if (skb->ip_summed == CHECKSUM_UNNECESSARY)
725 desc.csum = csum_partial(skb->data, desc.offset, skb->csum);
726 xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_and_csum_bits);
727 if (desc.offset != skb->len) {
729 csum2 = skb_checksum(skb, desc.offset, skb->len - desc.offset, 0);
730 desc.csum = csum_block_add(desc.csum, csum2, desc.offset);
734 if ((unsigned short)csum_fold(desc.csum))
738 xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_bits);
745 * Input handler for RPC replies. Called from a bottom half and hence
749 udp_data_ready(struct sock *sk, int len)
751 struct rpc_task *task;
752 struct rpc_xprt *xprt;
753 struct rpc_rqst *rovr;
755 int err, repsize, copied;
758 read_lock(&sk->sk_callback_lock);
759 dprintk("RPC: udp_data_ready...\n");
760 if (!(xprt = xprt_from_sock(sk))) {
761 printk("RPC: udp_data_ready request not found!\n");
765 dprintk("RPC: udp_data_ready client %p\n", xprt);
767 if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
773 repsize = skb->len - sizeof(struct udphdr);
775 printk("RPC: impossible RPC reply size %d!\n", repsize);
779 /* Copy the XID from the skb... */
780 if (skb_copy_bits(skb, sizeof(struct udphdr), &xid, sizeof(xid)) < 0)
783 /* Look up and lock the request corresponding to the given XID */
784 spin_lock(&xprt->sock_lock);
785 rovr = xprt_lookup_rqst(xprt, xid);
788 task = rovr->rq_task;
790 dprintk("RPC: %4d received reply\n", task->tk_pid);
792 if ((copied = rovr->rq_private_buf.buflen) > repsize)
795 /* Suck it into the iovec, verify checksum if not done by hw. */
796 if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
799 /* Something worked... */
800 dst_confirm(skb->dst);
802 xprt_complete_rqst(xprt, rovr, copied);
805 spin_unlock(&xprt->sock_lock);
807 skb_free_datagram(sk, skb);
809 read_unlock(&sk->sk_callback_lock);
813 * Copy from an skb into memory and shrink the skb.
816 tcp_copy_data(skb_reader_t *desc, void *p, size_t len)
818 if (len > desc->count)
820 if (skb_copy_bits(desc->skb, desc->offset, p, len))
828 * TCP read fragment marker
831 tcp_read_fraghdr(struct rpc_xprt *xprt, skb_reader_t *desc)
836 p = ((char *) &xprt->tcp_recm) + xprt->tcp_offset;
837 len = sizeof(xprt->tcp_recm) - xprt->tcp_offset;
838 used = tcp_copy_data(desc, p, len);
839 xprt->tcp_offset += used;
842 xprt->tcp_reclen = ntohl(xprt->tcp_recm);
843 if (xprt->tcp_reclen & 0x80000000)
844 xprt->tcp_flags |= XPRT_LAST_FRAG;
846 xprt->tcp_flags &= ~XPRT_LAST_FRAG;
847 xprt->tcp_reclen &= 0x7fffffff;
848 xprt->tcp_flags &= ~XPRT_COPY_RECM;
849 xprt->tcp_offset = 0;
850 /* Sanity check of the record length */
851 if (xprt->tcp_reclen < 4) {
852 printk(KERN_ERR "RPC: Invalid TCP record fragment length\n");
853 xprt_disconnect(xprt);
855 dprintk("RPC: reading TCP record fragment of length %d\n",
860 tcp_check_recm(struct rpc_xprt *xprt)
862 if (xprt->tcp_offset == xprt->tcp_reclen) {
863 xprt->tcp_flags |= XPRT_COPY_RECM;
864 xprt->tcp_offset = 0;
865 if (xprt->tcp_flags & XPRT_LAST_FRAG) {
866 xprt->tcp_flags &= ~XPRT_COPY_DATA;
867 xprt->tcp_flags |= XPRT_COPY_XID;
868 xprt->tcp_copied = 0;
877 tcp_read_xid(struct rpc_xprt *xprt, skb_reader_t *desc)
882 len = sizeof(xprt->tcp_xid) - xprt->tcp_offset;
883 dprintk("RPC: reading XID (%Zu bytes)\n", len);
884 p = ((char *) &xprt->tcp_xid) + xprt->tcp_offset;
885 used = tcp_copy_data(desc, p, len);
886 xprt->tcp_offset += used;
889 xprt->tcp_flags &= ~XPRT_COPY_XID;
890 xprt->tcp_flags |= XPRT_COPY_DATA;
891 xprt->tcp_copied = 4;
892 dprintk("RPC: reading reply for XID %08x\n", xprt->tcp_xid);
893 tcp_check_recm(xprt);
897 * TCP read and complete request
900 tcp_read_request(struct rpc_xprt *xprt, skb_reader_t *desc)
902 struct rpc_rqst *req;
903 struct xdr_buf *rcvbuf;
906 /* Find and lock the request corresponding to this xid */
907 spin_lock(&xprt->sock_lock);
908 req = xprt_lookup_rqst(xprt, xprt->tcp_xid);
910 xprt->tcp_flags &= ~XPRT_COPY_DATA;
911 dprintk("RPC: XID %08x request not found!\n",
913 spin_unlock(&xprt->sock_lock);
917 rcvbuf = &req->rq_private_buf;
919 if (len > xprt->tcp_reclen - xprt->tcp_offset) {
920 skb_reader_t my_desc;
922 len = xprt->tcp_reclen - xprt->tcp_offset;
923 memcpy(&my_desc, desc, sizeof(my_desc));
925 xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
926 &my_desc, tcp_copy_data);
930 xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
931 desc, tcp_copy_data);
932 xprt->tcp_copied += len;
933 xprt->tcp_offset += len;
935 if (xprt->tcp_copied == req->rq_private_buf.buflen)
936 xprt->tcp_flags &= ~XPRT_COPY_DATA;
937 else if (xprt->tcp_offset == xprt->tcp_reclen) {
938 if (xprt->tcp_flags & XPRT_LAST_FRAG)
939 xprt->tcp_flags &= ~XPRT_COPY_DATA;
942 if (!(xprt->tcp_flags & XPRT_COPY_DATA)) {
943 dprintk("RPC: %4d received reply complete\n",
944 req->rq_task->tk_pid);
945 xprt_complete_rqst(xprt, req, xprt->tcp_copied);
947 spin_unlock(&xprt->sock_lock);
948 tcp_check_recm(xprt);
952 * TCP discard extra bytes from a short read
955 tcp_read_discard(struct rpc_xprt *xprt, skb_reader_t *desc)
959 len = xprt->tcp_reclen - xprt->tcp_offset;
960 if (len > desc->count)
964 xprt->tcp_offset += len;
965 tcp_check_recm(xprt);
969 * TCP record receive routine
970 * We first have to grab the record marker, then the XID, then the data.
973 tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
974 unsigned int offset, size_t len)
976 struct rpc_xprt *xprt = (struct rpc_xprt *)rd_desc->buf;
977 skb_reader_t desc = {
984 dprintk("RPC: tcp_data_recv\n");
986 /* Read in a new fragment marker if necessary */
987 /* Can we ever really expect to get completely empty fragments? */
988 if (xprt->tcp_flags & XPRT_COPY_RECM) {
989 tcp_read_fraghdr(xprt, &desc);
992 /* Read in the xid if necessary */
993 if (xprt->tcp_flags & XPRT_COPY_XID) {
994 tcp_read_xid(xprt, &desc);
997 /* Read in the request data */
998 if (xprt->tcp_flags & XPRT_COPY_DATA) {
999 tcp_read_request(xprt, &desc);
1002 /* Skip over any trailing bytes on short reads */
1003 tcp_read_discard(xprt, &desc);
1004 } while (desc.count);
1005 dprintk("RPC: tcp_data_recv done\n");
1006 return len - desc.count;
1009 static void tcp_data_ready(struct sock *sk, int bytes)
1011 struct rpc_xprt *xprt;
1012 read_descriptor_t rd_desc;
1014 read_lock(&sk->sk_callback_lock);
1015 dprintk("RPC: tcp_data_ready...\n");
1016 if (!(xprt = xprt_from_sock(sk))) {
1017 printk("RPC: tcp_data_ready socket info not found!\n");
1023 /* We use rd_desc to pass struct xprt to tcp_data_recv */
1024 rd_desc.buf = (char *)xprt;
1025 rd_desc.count = 65536;
1026 tcp_read_sock(sk, &rd_desc, tcp_data_recv);
1028 read_unlock(&sk->sk_callback_lock);
1032 tcp_state_change(struct sock *sk)
1034 struct rpc_xprt *xprt;
1036 read_lock(&sk->sk_callback_lock);
1037 if (!(xprt = xprt_from_sock(sk)))
1039 dprintk("RPC: tcp_state_change client %p...\n", xprt);
1040 dprintk("RPC: state %x conn %d dead %d zapped %d\n",
1041 sk->sk_state, xprt_connected(xprt),
1042 sock_flag(sk, SOCK_DEAD), sk->sk_zapped);
1044 switch (sk->sk_state) {
1045 case TCP_ESTABLISHED:
1046 spin_lock_bh(&xprt->sock_lock);
1047 if (!xprt_test_and_set_connected(xprt)) {
1048 /* Reset TCP record info */
1049 xprt->tcp_offset = 0;
1050 xprt->tcp_reclen = 0;
1051 xprt->tcp_copied = 0;
1052 xprt->tcp_flags = XPRT_COPY_RECM | XPRT_COPY_XID;
1053 rpc_wake_up(&xprt->pending);
1055 spin_unlock_bh(&xprt->sock_lock);
1061 if (xprt_test_and_clear_connected(xprt))
1062 rpc_wake_up_status(&xprt->pending, -ENOTCONN);
1066 read_unlock(&sk->sk_callback_lock);
1070 * Called when more output buffer space is available for this socket.
1071 * We try not to wake our writers until they can make "significant"
1072 * progress, otherwise we'll waste resources thrashing sock_sendmsg
1073 * with a bunch of small requests.
1076 xprt_write_space(struct sock *sk)
1078 struct rpc_xprt *xprt;
1079 struct socket *sock;
1081 read_lock(&sk->sk_callback_lock);
1082 if (!(xprt = xprt_from_sock(sk)) || !(sock = sk->sk_socket))
1087 /* Wait until we have enough socket memory */
1089 /* from net/ipv4/tcp.c:tcp_write_space */
1090 if (tcp_wspace(sk) < tcp_min_write_space(sk))
1093 /* from net/core/sock.c:sock_def_write_space */
1094 if (!sock_writeable(sk))
1098 if (!test_and_clear_bit(SOCK_NOSPACE, &sock->flags))
1101 spin_lock_bh(&xprt->sock_lock);
1102 if (xprt->snd_task && xprt->snd_task->tk_rpcwait == &xprt->pending)
1103 rpc_wake_up_task(xprt->snd_task);
1104 spin_unlock_bh(&xprt->sock_lock);
1106 read_unlock(&sk->sk_callback_lock);
1110 * RPC receive timeout handler.
1113 xprt_timer(struct rpc_task *task)
1115 struct rpc_rqst *req = task->tk_rqstp;
1116 struct rpc_xprt *xprt = req->rq_xprt;
1118 spin_lock(&xprt->sock_lock);
1119 if (req->rq_received)
1122 xprt_adjust_cwnd(req->rq_xprt, -ETIMEDOUT);
1123 __xprt_put_cong(xprt, req);
1125 dprintk("RPC: %4d xprt_timer (%s request)\n",
1126 task->tk_pid, req ? "pending" : "backlogged");
1128 task->tk_status = -ETIMEDOUT;
1130 task->tk_timeout = 0;
1131 rpc_wake_up_task(task);
1132 spin_unlock(&xprt->sock_lock);
1136 * Place the actual RPC call.
1137 * We have to copy the iovec because sendmsg fiddles with its contents.
1140 xprt_prepare_transmit(struct rpc_task *task)
1142 struct rpc_rqst *req = task->tk_rqstp;
1143 struct rpc_xprt *xprt = req->rq_xprt;
1146 dprintk("RPC: %4d xprt_prepare_transmit\n", task->tk_pid);
1151 spin_lock_bh(&xprt->sock_lock);
1152 if (req->rq_received && !req->rq_bytes_sent) {
1153 err = req->rq_received;
1156 if (!__xprt_lock_write(xprt, task)) {
1161 if (!xprt_connected(xprt)) {
1166 spin_unlock_bh(&xprt->sock_lock);
1171 xprt_transmit(struct rpc_task *task)
1173 struct rpc_clnt *clnt = task->tk_client;
1174 struct rpc_rqst *req = task->tk_rqstp;
1175 struct rpc_xprt *xprt = req->rq_xprt;
1176 int status, retry = 0;
1179 dprintk("RPC: %4d xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
1181 /* set up everything as needed. */
1182 /* Write the record marker */
1184 u32 *marker = req->rq_svec[0].iov_base;
1186 *marker = htonl(0x80000000|(req->rq_slen-sizeof(*marker)));
1190 if (!req->rq_received) {
1191 if (list_empty(&req->rq_list)) {
1192 spin_lock_bh(&xprt->sock_lock);
1193 /* Update the softirq receive buffer */
1194 memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
1195 sizeof(req->rq_private_buf));
1196 /* Add request to the receive list */
1197 list_add_tail(&req->rq_list, &xprt->recv);
1198 spin_unlock_bh(&xprt->sock_lock);
1199 xprt_reset_majortimeo(req);
1201 } else if (!req->rq_bytes_sent)
1204 /* Continue transmitting the packet/record. We must be careful
1205 * to cope with writespace callbacks arriving _after_ we have
1206 * called xprt_sendmsg().
1209 req->rq_xtime = jiffies;
1210 status = xprt_sendmsg(xprt, req);
1216 req->rq_bytes_sent += status;
1218 /* If we've sent the entire packet, immediately
1219 * reset the count of bytes sent. */
1220 if (req->rq_bytes_sent >= req->rq_slen) {
1221 req->rq_bytes_sent = 0;
1225 if (status >= req->rq_slen)
1231 dprintk("RPC: %4d xmit incomplete (%d left of %d)\n",
1232 task->tk_pid, req->rq_slen - req->rq_bytes_sent,
1240 /* Note: at this point, task->tk_sleeping has not yet been set,
1241 * hence there is no danger of the waking up task being put on
1242 * schedq, and being picked up by a parallel run of rpciod().
1244 task->tk_status = status;
1248 if (test_bit(SOCK_ASYNC_NOSPACE, &xprt->sock->flags)) {
1249 /* Protect against races with xprt_write_space */
1250 spin_lock_bh(&xprt->sock_lock);
1251 /* Don't race with disconnect */
1252 if (!xprt_connected(xprt))
1253 task->tk_status = -ENOTCONN;
1254 else if (test_bit(SOCK_NOSPACE, &xprt->sock->flags)) {
1255 task->tk_timeout = req->rq_timeout;
1256 rpc_sleep_on(&xprt->pending, task, NULL, NULL);
1258 spin_unlock_bh(&xprt->sock_lock);
1261 /* Keep holding the socket if it is blocked */
1262 rpc_delay(task, HZ>>4);
1265 task->tk_timeout = RPC_REESTABLISH_TIMEOUT;
1266 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
1271 xprt_disconnect(xprt);
1273 xprt_release_write(xprt, task);
1276 dprintk("RPC: %4d xmit complete\n", task->tk_pid);
1277 /* Set the task's receive timeout value */
1278 spin_lock_bh(&xprt->sock_lock);
1279 if (!xprt->nocong) {
1280 int timer = task->tk_msg.rpc_proc->p_timer;
1281 task->tk_timeout = rpc_calc_rto(clnt->cl_rtt, timer);
1282 task->tk_timeout <<= rpc_ntimeo(clnt->cl_rtt, timer) + req->rq_retries;
1283 if (task->tk_timeout > xprt->timeout.to_maxval || task->tk_timeout == 0)
1284 task->tk_timeout = xprt->timeout.to_maxval;
1286 task->tk_timeout = req->rq_timeout;
1287 /* Don't race with disconnect */
1288 if (!xprt_connected(xprt))
1289 task->tk_status = -ENOTCONN;
1290 else if (!req->rq_received)
1291 rpc_sleep_on(&xprt->pending, task, NULL, xprt_timer);
1292 __xprt_release_write(xprt, task);
1293 spin_unlock_bh(&xprt->sock_lock);
1297 * Reserve an RPC call slot.
1300 xprt_reserve(struct rpc_task *task)
1302 struct rpc_xprt *xprt = task->tk_xprt;
1304 task->tk_status = -EIO;
1305 if (!xprt->shutdown) {
1306 spin_lock(&xprt->xprt_lock);
1307 do_xprt_reserve(task);
1308 spin_unlock(&xprt->xprt_lock);
1310 del_timer_sync(&xprt->timer);
1315 do_xprt_reserve(struct rpc_task *task)
1317 struct rpc_xprt *xprt = task->tk_xprt;
1319 task->tk_status = 0;
1322 if (!list_empty(&xprt->free)) {
1323 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
1324 list_del_init(&req->rq_list);
1325 task->tk_rqstp = req;
1326 xprt_request_init(task, xprt);
1329 dprintk("RPC: waiting for request slot\n");
1330 task->tk_status = -EAGAIN;
1331 task->tk_timeout = 0;
1332 rpc_sleep_on(&xprt->backlog, task, NULL, NULL);
1336 * Allocate a 'unique' XID
1338 static inline u32 xprt_alloc_xid(struct rpc_xprt *xprt)
1343 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1345 get_random_bytes(&xprt->xid, sizeof(xprt->xid));
1349 * Initialize RPC request
1352 xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1354 struct rpc_rqst *req = task->tk_rqstp;
1356 req->rq_timeout = xprt->timeout.to_initval;
1357 req->rq_task = task;
1358 req->rq_xprt = xprt;
1359 req->rq_xid = xprt_alloc_xid(xprt);
1360 dprintk("RPC: %4d reserved req %p xid %08x\n", task->tk_pid,
1365 * Release an RPC call slot
1368 xprt_release(struct rpc_task *task)
1370 struct rpc_xprt *xprt = task->tk_xprt;
1371 struct rpc_rqst *req;
1373 if (!(req = task->tk_rqstp))
1375 spin_lock_bh(&xprt->sock_lock);
1376 __xprt_release_write(xprt, task);
1377 __xprt_put_cong(xprt, req);
1378 if (!list_empty(&req->rq_list))
1379 list_del(&req->rq_list);
1380 xprt->last_used = jiffies;
1381 if (list_empty(&xprt->recv) && !xprt->shutdown)
1382 mod_timer(&xprt->timer, xprt->last_used + XPRT_IDLE_TIMEOUT);
1383 spin_unlock_bh(&xprt->sock_lock);
1384 task->tk_rqstp = NULL;
1385 memset(req, 0, sizeof(*req)); /* mark unused */
1387 dprintk("RPC: %4d release request %p\n", task->tk_pid, req);
1389 spin_lock(&xprt->xprt_lock);
1390 list_add(&req->rq_list, &xprt->free);
1391 xprt_clear_backlog(xprt);
1392 spin_unlock(&xprt->xprt_lock);
1396 * Set default timeout parameters
1399 xprt_default_timeout(struct rpc_timeout *to, int proto)
1401 if (proto == IPPROTO_UDP)
1402 xprt_set_timeout(to, 5, 5 * HZ);
1404 xprt_set_timeout(to, 5, 60 * HZ);
1408 * Set constant timeout
1411 xprt_set_timeout(struct rpc_timeout *to, unsigned int retr, unsigned long incr)
1414 to->to_increment = incr;
1415 to->to_maxval = incr * retr;
1416 to->to_retries = retr;
1417 to->to_exponential = 0;
1420 unsigned int xprt_udp_slot_table_entries = RPC_DEF_SLOT_TABLE;
1421 unsigned int xprt_tcp_slot_table_entries = RPC_DEF_SLOT_TABLE;
1424 * Initialize an RPC client
1426 static struct rpc_xprt *
1427 xprt_setup(int proto, struct sockaddr_in *ap, struct rpc_timeout *to)
1429 struct rpc_xprt *xprt;
1430 unsigned int entries;
1431 size_t slot_table_size;
1432 struct rpc_rqst *req;
1434 dprintk("RPC: setting up %s transport...\n",
1435 proto == IPPROTO_UDP? "UDP" : "TCP");
1437 entries = (proto == IPPROTO_TCP)?
1438 xprt_tcp_slot_table_entries : xprt_udp_slot_table_entries;
1440 if ((xprt = kmalloc(sizeof(struct rpc_xprt), GFP_KERNEL)) == NULL)
1441 return ERR_PTR(-ENOMEM);
1442 memset(xprt, 0, sizeof(*xprt)); /* Nnnngh! */
1443 xprt->max_reqs = entries;
1444 slot_table_size = entries * sizeof(xprt->slot[0]);
1445 xprt->slot = kmalloc(slot_table_size, GFP_KERNEL);
1446 if (xprt->slot == NULL) {
1448 return ERR_PTR(-ENOMEM);
1450 memset(xprt->slot, 0, slot_table_size);
1454 xprt->stream = (proto == IPPROTO_TCP)? 1 : 0;
1456 xprt->cwnd = RPC_MAXCWND(xprt);
1459 xprt->cwnd = RPC_INITCWND;
1460 spin_lock_init(&xprt->sock_lock);
1461 spin_lock_init(&xprt->xprt_lock);
1462 init_waitqueue_head(&xprt->cong_wait);
1464 INIT_LIST_HEAD(&xprt->free);
1465 INIT_LIST_HEAD(&xprt->recv);
1466 INIT_WORK(&xprt->sock_connect, xprt_socket_connect, xprt);
1467 INIT_WORK(&xprt->task_cleanup, xprt_socket_autoclose, xprt);
1468 init_timer(&xprt->timer);
1469 xprt->timer.function = xprt_init_autodisconnect;
1470 xprt->timer.data = (unsigned long) xprt;
1471 xprt->last_used = jiffies;
1472 xprt->port = XPRT_MAX_RESVPORT;
1474 /* Set timeout parameters */
1476 xprt->timeout = *to;
1478 xprt_default_timeout(&xprt->timeout, xprt->prot);
1480 rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1481 rpc_init_wait_queue(&xprt->sending, "xprt_sending");
1482 rpc_init_wait_queue(&xprt->resend, "xprt_resend");
1483 rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1485 /* initialize free list */
1486 for (req = &xprt->slot[entries-1]; req >= &xprt->slot[0]; req--)
1487 list_add(&req->rq_list, &xprt->free);
1489 xprt_init_xid(xprt);
1491 /* Check whether we want to use a reserved port */
1492 xprt->resvport = capable(CAP_NET_BIND_SERVICE) ? 1 : 0;
1494 dprintk("RPC: created transport %p with %u slots\n", xprt,
1501 * Bind to a reserved port
1503 static inline int xprt_bindresvport(struct rpc_xprt *xprt, struct socket *sock)
1505 struct sockaddr_in myaddr = {
1506 .sin_family = AF_INET,
1510 /* Were we already bound to a given port? Try to reuse it */
1513 myaddr.sin_port = htons(port);
1514 err = sock->ops->bind(sock, (struct sockaddr *) &myaddr,
1521 port = XPRT_MAX_RESVPORT;
1522 } while (err == -EADDRINUSE && port != xprt->port);
1524 printk("RPC: Can't bind to reserved port (%d).\n", -err);
1529 xprt_bind_socket(struct rpc_xprt *xprt, struct socket *sock)
1531 struct sock *sk = sock->sk;
1536 write_lock_bh(&sk->sk_callback_lock);
1537 sk->sk_user_data = xprt;
1538 xprt->old_data_ready = sk->sk_data_ready;
1539 xprt->old_state_change = sk->sk_state_change;
1540 xprt->old_write_space = sk->sk_write_space;
1541 if (xprt->prot == IPPROTO_UDP) {
1542 sk->sk_data_ready = udp_data_ready;
1543 sk->sk_no_check = UDP_CSUM_NORCV;
1544 xprt_set_connected(xprt);
1546 struct tcp_opt *tp = tcp_sk(sk);
1547 tp->nonagle = 1; /* disable Nagle's algorithm */
1548 sk->sk_data_ready = tcp_data_ready;
1549 sk->sk_state_change = tcp_state_change;
1550 xprt_clear_connected(xprt);
1552 sk->sk_write_space = xprt_write_space;
1554 /* Reset to new socket */
1557 write_unlock_bh(&sk->sk_callback_lock);
1563 * Set socket buffer length
1566 xprt_sock_setbufsize(struct rpc_xprt *xprt)
1568 struct sock *sk = xprt->inet;
1572 if (xprt->rcvsize) {
1573 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
1574 sk->sk_rcvbuf = xprt->rcvsize * xprt->max_reqs * 2;
1576 if (xprt->sndsize) {
1577 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
1578 sk->sk_sndbuf = xprt->sndsize * xprt->max_reqs * 2;
1579 sk->sk_write_space(sk);
1584 * Datastream sockets are created here, but xprt_connect will create
1585 * and connect stream sockets.
1587 static struct socket * xprt_create_socket(struct rpc_xprt *xprt, int proto, int resvport)
1589 struct socket *sock;
1592 dprintk("RPC: xprt_create_socket(%s %d)\n",
1593 (proto == IPPROTO_UDP)? "udp" : "tcp", proto);
1595 type = (proto == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1597 if ((err = sock_create_kern(PF_INET, type, proto, &sock)) < 0) {
1598 printk("RPC: can't create socket (%d).\n", -err);
1602 /* If the caller has the capability, bind to a reserved port */
1603 if (resvport && xprt_bindresvport(xprt, sock) < 0) {
1604 printk("RPC: can't bind to reserved port.\n");
1616 * Create an RPC client transport given the protocol and peer address.
1619 xprt_create_proto(int proto, struct sockaddr_in *sap, struct rpc_timeout *to)
1621 struct rpc_xprt *xprt;
1623 xprt = xprt_setup(proto, sap, to);
1625 dprintk("RPC: xprt_create_proto failed\n");
1627 dprintk("RPC: xprt_create_proto created xprt %p\n", xprt);
1632 * Prepare for transport shutdown.
1635 xprt_shutdown(struct rpc_xprt *xprt)
1638 rpc_wake_up(&xprt->sending);
1639 rpc_wake_up(&xprt->resend);
1640 rpc_wake_up(&xprt->pending);
1641 rpc_wake_up(&xprt->backlog);
1642 wake_up(&xprt->cong_wait);
1643 del_timer_sync(&xprt->timer);
1647 * Clear the xprt backlog queue
1650 xprt_clear_backlog(struct rpc_xprt *xprt) {
1651 rpc_wake_up_next(&xprt->backlog);
1652 wake_up(&xprt->cong_wait);
1657 * Destroy an RPC transport, killing off all requests.
1660 xprt_destroy(struct rpc_xprt *xprt)
1662 dprintk("RPC: destroying transport %p\n", xprt);
1663 xprt_shutdown(xprt);
1664 xprt_disconnect(xprt);