282efd447a61d535b23c6322aecc2bc756a73195
[linux-2.6.git] / net / sunrpc / xprtsock.c
1 /*
2  * linux/net/sunrpc/xprtsock.c
3  *
4  * Client-side transport implementation for sockets.
5  *
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>
10  *
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>
14  *
15  * IP socket transport implementation, (C) 2005 Chuck Lever <cel@netapp.com>
16  */
17
18 #include <linux/types.h>
19 #include <linux/slab.h>
20 #include <linux/capability.h>
21 #include <linux/pagemap.h>
22 #include <linux/errno.h>
23 #include <linux/socket.h>
24 #include <linux/in.h>
25 #include <linux/net.h>
26 #include <linux/mm.h>
27 #include <linux/udp.h>
28 #include <linux/tcp.h>
29 #include <linux/sunrpc/clnt.h>
30 #include <linux/sunrpc/sched.h>
31 #include <linux/file.h>
32
33 #include <net/sock.h>
34 #include <net/checksum.h>
35 #include <net/udp.h>
36 #include <net/tcp.h>
37
38 /*
39  * xprtsock tunables
40  */
41 unsigned int xprt_udp_slot_table_entries = RPC_DEF_SLOT_TABLE;
42 unsigned int xprt_tcp_slot_table_entries = RPC_DEF_SLOT_TABLE;
43
44 unsigned int xprt_min_resvport = RPC_DEF_MIN_RESVPORT;
45 unsigned int xprt_max_resvport = RPC_DEF_MAX_RESVPORT;
46
47 /*
48  * We can register our own files under /proc/sys/sunrpc by
49  * calling register_sysctl_table() again.  The files in that
50  * directory become the union of all files registered there.
51  *
52  * We simply need to make sure that we don't collide with
53  * someone else's file names!
54  */
55
56 #ifdef RPC_DEBUG
57
58 static unsigned int min_slot_table_size = RPC_MIN_SLOT_TABLE;
59 static unsigned int max_slot_table_size = RPC_MAX_SLOT_TABLE;
60 static unsigned int xprt_min_resvport_limit = RPC_MIN_RESVPORT;
61 static unsigned int xprt_max_resvport_limit = RPC_MAX_RESVPORT;
62
63 static struct ctl_table_header *sunrpc_table_header;
64
65 /*
66  * FIXME: changing the UDP slot table size should also resize the UDP
67  *        socket buffers for existing UDP transports
68  */
69 static ctl_table xs_tunables_table[] = {
70         {
71                 .ctl_name       = CTL_SLOTTABLE_UDP,
72                 .procname       = "udp_slot_table_entries",
73                 .data           = &xprt_udp_slot_table_entries,
74                 .maxlen         = sizeof(unsigned int),
75                 .mode           = 0644,
76                 .proc_handler   = &proc_dointvec_minmax,
77                 .strategy       = &sysctl_intvec,
78                 .extra1         = &min_slot_table_size,
79                 .extra2         = &max_slot_table_size
80         },
81         {
82                 .ctl_name       = CTL_SLOTTABLE_TCP,
83                 .procname       = "tcp_slot_table_entries",
84                 .data           = &xprt_tcp_slot_table_entries,
85                 .maxlen         = sizeof(unsigned int),
86                 .mode           = 0644,
87                 .proc_handler   = &proc_dointvec_minmax,
88                 .strategy       = &sysctl_intvec,
89                 .extra1         = &min_slot_table_size,
90                 .extra2         = &max_slot_table_size
91         },
92         {
93                 .ctl_name       = CTL_MIN_RESVPORT,
94                 .procname       = "min_resvport",
95                 .data           = &xprt_min_resvport,
96                 .maxlen         = sizeof(unsigned int),
97                 .mode           = 0644,
98                 .proc_handler   = &proc_dointvec_minmax,
99                 .strategy       = &sysctl_intvec,
100                 .extra1         = &xprt_min_resvport_limit,
101                 .extra2         = &xprt_max_resvport_limit
102         },
103         {
104                 .ctl_name       = CTL_MAX_RESVPORT,
105                 .procname       = "max_resvport",
106                 .data           = &xprt_max_resvport,
107                 .maxlen         = sizeof(unsigned int),
108                 .mode           = 0644,
109                 .proc_handler   = &proc_dointvec_minmax,
110                 .strategy       = &sysctl_intvec,
111                 .extra1         = &xprt_min_resvport_limit,
112                 .extra2         = &xprt_max_resvport_limit
113         },
114         {
115                 .ctl_name = 0,
116         },
117 };
118
119 static ctl_table sunrpc_table[] = {
120         {
121                 .ctl_name       = CTL_SUNRPC,
122                 .procname       = "sunrpc",
123                 .mode           = 0555,
124                 .child          = xs_tunables_table
125         },
126         {
127                 .ctl_name = 0,
128         },
129 };
130
131 #endif
132
133 /*
134  * How many times to try sending a request on a socket before waiting
135  * for the socket buffer to clear.
136  */
137 #define XS_SENDMSG_RETRY        (10U)
138
139 /*
140  * Time out for an RPC UDP socket connect.  UDP socket connects are
141  * synchronous, but we set a timeout anyway in case of resource
142  * exhaustion on the local host.
143  */
144 #define XS_UDP_CONN_TO          (5U * HZ)
145
146 /*
147  * Wait duration for an RPC TCP connection to be established.  Solaris
148  * NFS over TCP uses 60 seconds, for example, which is in line with how
149  * long a server takes to reboot.
150  */
151 #define XS_TCP_CONN_TO          (60U * HZ)
152
153 /*
154  * Wait duration for a reply from the RPC portmapper.
155  */
156 #define XS_BIND_TO              (60U * HZ)
157
158 /*
159  * Delay if a UDP socket connect error occurs.  This is most likely some
160  * kind of resource problem on the local host.
161  */
162 #define XS_UDP_REEST_TO         (2U * HZ)
163
164 /*
165  * The reestablish timeout allows clients to delay for a bit before attempting
166  * to reconnect to a server that just dropped our connection.
167  *
168  * We implement an exponential backoff when trying to reestablish a TCP
169  * transport connection with the server.  Some servers like to drop a TCP
170  * connection when they are overworked, so we start with a short timeout and
171  * increase over time if the server is down or not responding.
172  */
173 #define XS_TCP_INIT_REEST_TO    (3U * HZ)
174 #define XS_TCP_MAX_REEST_TO     (5U * 60 * HZ)
175
176 /*
177  * TCP idle timeout; client drops the transport socket if it is idle
178  * for this long.  Note that we also timeout UDP sockets to prevent
179  * holding port numbers when there is no RPC traffic.
180  */
181 #define XS_IDLE_DISC_TO         (5U * 60 * HZ)
182
183 #ifdef RPC_DEBUG
184 # undef  RPC_DEBUG_DATA
185 # define RPCDBG_FACILITY        RPCDBG_TRANS
186 #endif
187
188 #ifdef RPC_DEBUG_DATA
189 static void xs_pktdump(char *msg, u32 *packet, unsigned int count)
190 {
191         u8 *buf = (u8 *) packet;
192         int j;
193
194         dprintk("RPC:       %s\n", msg);
195         for (j = 0; j < count && j < 128; j += 4) {
196                 if (!(j & 31)) {
197                         if (j)
198                                 dprintk("\n");
199                         dprintk("0x%04x ", j);
200                 }
201                 dprintk("%02x%02x%02x%02x ",
202                         buf[j], buf[j+1], buf[j+2], buf[j+3]);
203         }
204         dprintk("\n");
205 }
206 #else
207 static inline void xs_pktdump(char *msg, u32 *packet, unsigned int count)
208 {
209         /* NOP */
210 }
211 #endif
212
213 struct sock_xprt {
214         struct rpc_xprt         xprt;
215
216         /*
217          * Network layer
218          */
219         struct socket *         sock;
220         struct sock *           inet;
221
222         /*
223          * State of TCP reply receive
224          */
225         __be32                  tcp_fraghdr,
226                                 tcp_xid;
227
228         u32                     tcp_offset,
229                                 tcp_reclen;
230
231         unsigned long           tcp_copied,
232                                 tcp_flags;
233
234         /*
235          * Connection of transports
236          */
237         struct delayed_work     connect_worker;
238         struct sockaddr_storage addr;
239         unsigned short          port;
240
241         /*
242          * UDP socket buffer size parameters
243          */
244         size_t                  rcvsize,
245                                 sndsize;
246
247         /*
248          * Saved socket callback addresses
249          */
250         void                    (*old_data_ready)(struct sock *, int);
251         void                    (*old_state_change)(struct sock *);
252         void                    (*old_write_space)(struct sock *);
253 };
254
255 /*
256  * TCP receive state flags
257  */
258 #define TCP_RCV_LAST_FRAG       (1UL << 0)
259 #define TCP_RCV_COPY_FRAGHDR    (1UL << 1)
260 #define TCP_RCV_COPY_XID        (1UL << 2)
261 #define TCP_RCV_COPY_DATA       (1UL << 3)
262
263 static void xs_format_peer_addresses(struct rpc_xprt *xprt)
264 {
265         struct sockaddr_in *addr = (struct sockaddr_in *) &xprt->addr;
266         char *buf;
267
268         buf = kzalloc(20, GFP_KERNEL);
269         if (buf) {
270                 snprintf(buf, 20, "%u.%u.%u.%u",
271                                 NIPQUAD(addr->sin_addr.s_addr));
272         }
273         xprt->address_strings[RPC_DISPLAY_ADDR] = buf;
274
275         buf = kzalloc(8, GFP_KERNEL);
276         if (buf) {
277                 snprintf(buf, 8, "%u",
278                                 ntohs(addr->sin_port));
279         }
280         xprt->address_strings[RPC_DISPLAY_PORT] = buf;
281
282         if (xprt->prot == IPPROTO_UDP)
283                 xprt->address_strings[RPC_DISPLAY_PROTO] = "udp";
284         else
285                 xprt->address_strings[RPC_DISPLAY_PROTO] = "tcp";
286
287         buf = kzalloc(48, GFP_KERNEL);
288         if (buf) {
289                 snprintf(buf, 48, "addr=%u.%u.%u.%u port=%u proto=%s",
290                         NIPQUAD(addr->sin_addr.s_addr),
291                         ntohs(addr->sin_port),
292                         xprt->prot == IPPROTO_UDP ? "udp" : "tcp");
293         }
294         xprt->address_strings[RPC_DISPLAY_ALL] = buf;
295 }
296
297 static void xs_free_peer_addresses(struct rpc_xprt *xprt)
298 {
299         kfree(xprt->address_strings[RPC_DISPLAY_ADDR]);
300         kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
301         kfree(xprt->address_strings[RPC_DISPLAY_ALL]);
302 }
303
304 #define XS_SENDMSG_FLAGS        (MSG_DONTWAIT | MSG_NOSIGNAL)
305
306 static int xs_send_kvec(struct socket *sock, struct sockaddr *addr, int addrlen, struct kvec *vec, unsigned int base, int more)
307 {
308         struct msghdr msg = {
309                 .msg_name       = addr,
310                 .msg_namelen    = addrlen,
311                 .msg_flags      = XS_SENDMSG_FLAGS | (more ? MSG_MORE : 0),
312         };
313         struct kvec iov = {
314                 .iov_base       = vec->iov_base + base,
315                 .iov_len        = vec->iov_len - base,
316         };
317
318         if (iov.iov_len != 0)
319                 return kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
320         return kernel_sendmsg(sock, &msg, NULL, 0, 0);
321 }
322
323 static int xs_send_pagedata(struct socket *sock, struct xdr_buf *xdr, unsigned int base, int more)
324 {
325         struct page **ppage;
326         unsigned int remainder;
327         int err, sent = 0;
328
329         remainder = xdr->page_len - base;
330         base += xdr->page_base;
331         ppage = xdr->pages + (base >> PAGE_SHIFT);
332         base &= ~PAGE_MASK;
333         for(;;) {
334                 unsigned int len = min_t(unsigned int, PAGE_SIZE - base, remainder);
335                 int flags = XS_SENDMSG_FLAGS;
336
337                 remainder -= len;
338                 if (remainder != 0 || more)
339                         flags |= MSG_MORE;
340                 err = sock->ops->sendpage(sock, *ppage, base, len, flags);
341                 if (remainder == 0 || err != len)
342                         break;
343                 sent += err;
344                 ppage++;
345                 base = 0;
346         }
347         if (sent == 0)
348                 return err;
349         if (err > 0)
350                 sent += err;
351         return sent;
352 }
353
354 /**
355  * xs_sendpages - write pages directly to a socket
356  * @sock: socket to send on
357  * @addr: UDP only -- address of destination
358  * @addrlen: UDP only -- length of destination address
359  * @xdr: buffer containing this request
360  * @base: starting position in the buffer
361  *
362  */
363 static int xs_sendpages(struct socket *sock, struct sockaddr *addr, int addrlen, struct xdr_buf *xdr, unsigned int base)
364 {
365         unsigned int remainder = xdr->len - base;
366         int err, sent = 0;
367
368         if (unlikely(!sock))
369                 return -ENOTCONN;
370
371         clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
372         if (base != 0) {
373                 addr = NULL;
374                 addrlen = 0;
375         }
376
377         if (base < xdr->head[0].iov_len || addr != NULL) {
378                 unsigned int len = xdr->head[0].iov_len - base;
379                 remainder -= len;
380                 err = xs_send_kvec(sock, addr, addrlen, &xdr->head[0], base, remainder != 0);
381                 if (remainder == 0 || err != len)
382                         goto out;
383                 sent += err;
384                 base = 0;
385         } else
386                 base -= xdr->head[0].iov_len;
387
388         if (base < xdr->page_len) {
389                 unsigned int len = xdr->page_len - base;
390                 remainder -= len;
391                 err = xs_send_pagedata(sock, xdr, base, remainder != 0);
392                 if (remainder == 0 || err != len)
393                         goto out;
394                 sent += err;
395                 base = 0;
396         } else
397                 base -= xdr->page_len;
398
399         if (base >= xdr->tail[0].iov_len)
400                 return sent;
401         err = xs_send_kvec(sock, NULL, 0, &xdr->tail[0], base, 0);
402 out:
403         if (sent == 0)
404                 return err;
405         if (err > 0)
406                 sent += err;
407         return sent;
408 }
409
410 /**
411  * xs_nospace - place task on wait queue if transmit was incomplete
412  * @task: task to put to sleep
413  *
414  */
415 static void xs_nospace(struct rpc_task *task)
416 {
417         struct rpc_rqst *req = task->tk_rqstp;
418         struct rpc_xprt *xprt = req->rq_xprt;
419         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
420
421         dprintk("RPC: %5u xmit incomplete (%u left of %u)\n",
422                         task->tk_pid, req->rq_slen - req->rq_bytes_sent,
423                         req->rq_slen);
424
425         if (test_bit(SOCK_ASYNC_NOSPACE, &transport->sock->flags)) {
426                 /* Protect against races with write_space */
427                 spin_lock_bh(&xprt->transport_lock);
428
429                 /* Don't race with disconnect */
430                 if (!xprt_connected(xprt))
431                         task->tk_status = -ENOTCONN;
432                 else if (test_bit(SOCK_NOSPACE, &transport->sock->flags))
433                         xprt_wait_for_buffer_space(task);
434
435                 spin_unlock_bh(&xprt->transport_lock);
436         } else
437                 /* Keep holding the socket if it is blocked */
438                 rpc_delay(task, HZ>>4);
439 }
440
441 /**
442  * xs_udp_send_request - write an RPC request to a UDP socket
443  * @task: address of RPC task that manages the state of an RPC request
444  *
445  * Return values:
446  *        0:    The request has been sent
447  *   EAGAIN:    The socket was blocked, please call again later to
448  *              complete the request
449  * ENOTCONN:    Caller needs to invoke connect logic then call again
450  *    other:    Some other error occured, the request was not sent
451  */
452 static int xs_udp_send_request(struct rpc_task *task)
453 {
454         struct rpc_rqst *req = task->tk_rqstp;
455         struct rpc_xprt *xprt = req->rq_xprt;
456         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
457         struct xdr_buf *xdr = &req->rq_snd_buf;
458         int status;
459
460         xs_pktdump("packet data:",
461                                 req->rq_svec->iov_base,
462                                 req->rq_svec->iov_len);
463
464         req->rq_xtime = jiffies;
465         status = xs_sendpages(transport->sock,
466                               (struct sockaddr *) &xprt->addr,
467                               xprt->addrlen, xdr,
468                               req->rq_bytes_sent);
469
470         dprintk("RPC:       xs_udp_send_request(%u) = %d\n",
471                         xdr->len - req->rq_bytes_sent, status);
472
473         if (likely(status >= (int) req->rq_slen))
474                 return 0;
475
476         /* Still some bytes left; set up for a retry later. */
477         if (status > 0)
478                 status = -EAGAIN;
479
480         switch (status) {
481         case -ENETUNREACH:
482         case -EPIPE:
483         case -ECONNREFUSED:
484                 /* When the server has died, an ICMP port unreachable message
485                  * prompts ECONNREFUSED. */
486                 break;
487         case -EAGAIN:
488                 xs_nospace(task);
489                 break;
490         default:
491                 dprintk("RPC:       sendmsg returned unrecognized error %d\n",
492                         -status);
493                 break;
494         }
495
496         return status;
497 }
498
499 static inline void xs_encode_tcp_record_marker(struct xdr_buf *buf)
500 {
501         u32 reclen = buf->len - sizeof(rpc_fraghdr);
502         rpc_fraghdr *base = buf->head[0].iov_base;
503         *base = htonl(RPC_LAST_STREAM_FRAGMENT | reclen);
504 }
505
506 /**
507  * xs_tcp_send_request - write an RPC request to a TCP socket
508  * @task: address of RPC task that manages the state of an RPC request
509  *
510  * Return values:
511  *        0:    The request has been sent
512  *   EAGAIN:    The socket was blocked, please call again later to
513  *              complete the request
514  * ENOTCONN:    Caller needs to invoke connect logic then call again
515  *    other:    Some other error occured, the request was not sent
516  *
517  * XXX: In the case of soft timeouts, should we eventually give up
518  *      if sendmsg is not able to make progress?
519  */
520 static int xs_tcp_send_request(struct rpc_task *task)
521 {
522         struct rpc_rqst *req = task->tk_rqstp;
523         struct rpc_xprt *xprt = req->rq_xprt;
524         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
525         struct xdr_buf *xdr = &req->rq_snd_buf;
526         int status, retry = 0;
527
528         xs_encode_tcp_record_marker(&req->rq_snd_buf);
529
530         xs_pktdump("packet data:",
531                                 req->rq_svec->iov_base,
532                                 req->rq_svec->iov_len);
533
534         /* Continue transmitting the packet/record. We must be careful
535          * to cope with writespace callbacks arriving _after_ we have
536          * called sendmsg(). */
537         while (1) {
538                 req->rq_xtime = jiffies;
539                 status = xs_sendpages(transport->sock,
540                                         NULL, 0, xdr, req->rq_bytes_sent);
541
542                 dprintk("RPC:       xs_tcp_send_request(%u) = %d\n",
543                                 xdr->len - req->rq_bytes_sent, status);
544
545                 if (unlikely(status < 0))
546                         break;
547
548                 /* If we've sent the entire packet, immediately
549                  * reset the count of bytes sent. */
550                 req->rq_bytes_sent += status;
551                 task->tk_bytes_sent += status;
552                 if (likely(req->rq_bytes_sent >= req->rq_slen)) {
553                         req->rq_bytes_sent = 0;
554                         return 0;
555                 }
556
557                 status = -EAGAIN;
558                 if (retry++ > XS_SENDMSG_RETRY)
559                         break;
560         }
561
562         switch (status) {
563         case -EAGAIN:
564                 xs_nospace(task);
565                 break;
566         case -ECONNREFUSED:
567         case -ECONNRESET:
568         case -ENOTCONN:
569         case -EPIPE:
570                 status = -ENOTCONN;
571                 break;
572         default:
573                 dprintk("RPC:       sendmsg returned unrecognized error %d\n",
574                         -status);
575                 xprt_disconnect(xprt);
576                 break;
577         }
578
579         return status;
580 }
581
582 /**
583  * xs_tcp_release_xprt - clean up after a tcp transmission
584  * @xprt: transport
585  * @task: rpc task
586  *
587  * This cleans up if an error causes us to abort the transmission of a request.
588  * In this case, the socket may need to be reset in order to avoid confusing
589  * the server.
590  */
591 static void xs_tcp_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
592 {
593         struct rpc_rqst *req;
594
595         if (task != xprt->snd_task)
596                 return;
597         if (task == NULL)
598                 goto out_release;
599         req = task->tk_rqstp;
600         if (req->rq_bytes_sent == 0)
601                 goto out_release;
602         if (req->rq_bytes_sent == req->rq_snd_buf.len)
603                 goto out_release;
604         set_bit(XPRT_CLOSE_WAIT, &task->tk_xprt->state);
605 out_release:
606         xprt_release_xprt(xprt, task);
607 }
608
609 /**
610  * xs_close - close a socket
611  * @xprt: transport
612  *
613  * This is used when all requests are complete; ie, no DRC state remains
614  * on the server we want to save.
615  */
616 static void xs_close(struct rpc_xprt *xprt)
617 {
618         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
619         struct socket *sock = transport->sock;
620         struct sock *sk = transport->inet;
621
622         if (!sk)
623                 goto clear_close_wait;
624
625         dprintk("RPC:       xs_close xprt %p\n", xprt);
626
627         write_lock_bh(&sk->sk_callback_lock);
628         transport->inet = NULL;
629         transport->sock = NULL;
630
631         sk->sk_user_data = NULL;
632         sk->sk_data_ready = transport->old_data_ready;
633         sk->sk_state_change = transport->old_state_change;
634         sk->sk_write_space = transport->old_write_space;
635         write_unlock_bh(&sk->sk_callback_lock);
636
637         sk->sk_no_check = 0;
638
639         sock_release(sock);
640 clear_close_wait:
641         smp_mb__before_clear_bit();
642         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
643         smp_mb__after_clear_bit();
644 }
645
646 /**
647  * xs_destroy - prepare to shutdown a transport
648  * @xprt: doomed transport
649  *
650  */
651 static void xs_destroy(struct rpc_xprt *xprt)
652 {
653         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
654
655         dprintk("RPC:       xs_destroy xprt %p\n", xprt);
656
657         cancel_rearming_delayed_work(&transport->connect_worker);
658
659         xprt_disconnect(xprt);
660         xs_close(xprt);
661         xs_free_peer_addresses(xprt);
662         kfree(xprt->slot);
663         kfree(xprt);
664 }
665
666 static inline struct rpc_xprt *xprt_from_sock(struct sock *sk)
667 {
668         return (struct rpc_xprt *) sk->sk_user_data;
669 }
670
671 /**
672  * xs_udp_data_ready - "data ready" callback for UDP sockets
673  * @sk: socket with data to read
674  * @len: how much data to read
675  *
676  */
677 static void xs_udp_data_ready(struct sock *sk, int len)
678 {
679         struct rpc_task *task;
680         struct rpc_xprt *xprt;
681         struct rpc_rqst *rovr;
682         struct sk_buff *skb;
683         int err, repsize, copied;
684         u32 _xid;
685         __be32 *xp;
686
687         read_lock(&sk->sk_callback_lock);
688         dprintk("RPC:       xs_udp_data_ready...\n");
689         if (!(xprt = xprt_from_sock(sk)))
690                 goto out;
691
692         if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
693                 goto out;
694
695         if (xprt->shutdown)
696                 goto dropit;
697
698         repsize = skb->len - sizeof(struct udphdr);
699         if (repsize < 4) {
700                 dprintk("RPC:       impossible RPC reply size %d!\n", repsize);
701                 goto dropit;
702         }
703
704         /* Copy the XID from the skb... */
705         xp = skb_header_pointer(skb, sizeof(struct udphdr),
706                                 sizeof(_xid), &_xid);
707         if (xp == NULL)
708                 goto dropit;
709
710         /* Look up and lock the request corresponding to the given XID */
711         spin_lock(&xprt->transport_lock);
712         rovr = xprt_lookup_rqst(xprt, *xp);
713         if (!rovr)
714                 goto out_unlock;
715         task = rovr->rq_task;
716
717         if ((copied = rovr->rq_private_buf.buflen) > repsize)
718                 copied = repsize;
719
720         /* Suck it into the iovec, verify checksum if not done by hw. */
721         if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
722                 goto out_unlock;
723
724         /* Something worked... */
725         dst_confirm(skb->dst);
726
727         xprt_adjust_cwnd(task, copied);
728         xprt_update_rtt(task);
729         xprt_complete_rqst(task, copied);
730
731  out_unlock:
732         spin_unlock(&xprt->transport_lock);
733  dropit:
734         skb_free_datagram(sk, skb);
735  out:
736         read_unlock(&sk->sk_callback_lock);
737 }
738
739 static inline void xs_tcp_read_fraghdr(struct rpc_xprt *xprt, struct xdr_skb_reader *desc)
740 {
741         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
742         size_t len, used;
743         char *p;
744
745         p = ((char *) &transport->tcp_fraghdr) + transport->tcp_offset;
746         len = sizeof(transport->tcp_fraghdr) - transport->tcp_offset;
747         used = xdr_skb_read_bits(desc, p, len);
748         transport->tcp_offset += used;
749         if (used != len)
750                 return;
751
752         transport->tcp_reclen = ntohl(transport->tcp_fraghdr);
753         if (transport->tcp_reclen & RPC_LAST_STREAM_FRAGMENT)
754                 transport->tcp_flags |= TCP_RCV_LAST_FRAG;
755         else
756                 transport->tcp_flags &= ~TCP_RCV_LAST_FRAG;
757         transport->tcp_reclen &= RPC_FRAGMENT_SIZE_MASK;
758
759         transport->tcp_flags &= ~TCP_RCV_COPY_FRAGHDR;
760         transport->tcp_offset = 0;
761
762         /* Sanity check of the record length */
763         if (unlikely(transport->tcp_reclen < 4)) {
764                 dprintk("RPC:       invalid TCP record fragment length\n");
765                 xprt_disconnect(xprt);
766                 return;
767         }
768         dprintk("RPC:       reading TCP record fragment of length %d\n",
769                         transport->tcp_reclen);
770 }
771
772 static void xs_tcp_check_fraghdr(struct sock_xprt *transport)
773 {
774         if (transport->tcp_offset == transport->tcp_reclen) {
775                 transport->tcp_flags |= TCP_RCV_COPY_FRAGHDR;
776                 transport->tcp_offset = 0;
777                 if (transport->tcp_flags & TCP_RCV_LAST_FRAG) {
778                         transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
779                         transport->tcp_flags |= TCP_RCV_COPY_XID;
780                         transport->tcp_copied = 0;
781                 }
782         }
783 }
784
785 static inline void xs_tcp_read_xid(struct sock_xprt *transport, struct xdr_skb_reader *desc)
786 {
787         size_t len, used;
788         char *p;
789
790         len = sizeof(transport->tcp_xid) - transport->tcp_offset;
791         dprintk("RPC:       reading XID (%Zu bytes)\n", len);
792         p = ((char *) &transport->tcp_xid) + transport->tcp_offset;
793         used = xdr_skb_read_bits(desc, p, len);
794         transport->tcp_offset += used;
795         if (used != len)
796                 return;
797         transport->tcp_flags &= ~TCP_RCV_COPY_XID;
798         transport->tcp_flags |= TCP_RCV_COPY_DATA;
799         transport->tcp_copied = 4;
800         dprintk("RPC:       reading reply for XID %08x\n",
801                         ntohl(transport->tcp_xid));
802         xs_tcp_check_fraghdr(transport);
803 }
804
805 static inline void xs_tcp_read_request(struct rpc_xprt *xprt, struct xdr_skb_reader *desc)
806 {
807         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
808         struct rpc_rqst *req;
809         struct xdr_buf *rcvbuf;
810         size_t len;
811         ssize_t r;
812
813         /* Find and lock the request corresponding to this xid */
814         spin_lock(&xprt->transport_lock);
815         req = xprt_lookup_rqst(xprt, transport->tcp_xid);
816         if (!req) {
817                 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
818                 dprintk("RPC:       XID %08x request not found!\n",
819                                 ntohl(transport->tcp_xid));
820                 spin_unlock(&xprt->transport_lock);
821                 return;
822         }
823
824         rcvbuf = &req->rq_private_buf;
825         len = desc->count;
826         if (len > transport->tcp_reclen - transport->tcp_offset) {
827                 struct xdr_skb_reader my_desc;
828
829                 len = transport->tcp_reclen - transport->tcp_offset;
830                 memcpy(&my_desc, desc, sizeof(my_desc));
831                 my_desc.count = len;
832                 r = xdr_partial_copy_from_skb(rcvbuf, transport->tcp_copied,
833                                           &my_desc, xdr_skb_read_bits);
834                 desc->count -= r;
835                 desc->offset += r;
836         } else
837                 r = xdr_partial_copy_from_skb(rcvbuf, transport->tcp_copied,
838                                           desc, xdr_skb_read_bits);
839
840         if (r > 0) {
841                 transport->tcp_copied += r;
842                 transport->tcp_offset += r;
843         }
844         if (r != len) {
845                 /* Error when copying to the receive buffer,
846                  * usually because we weren't able to allocate
847                  * additional buffer pages. All we can do now
848                  * is turn off TCP_RCV_COPY_DATA, so the request
849                  * will not receive any additional updates,
850                  * and time out.
851                  * Any remaining data from this record will
852                  * be discarded.
853                  */
854                 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
855                 dprintk("RPC:       XID %08x truncated request\n",
856                                 ntohl(transport->tcp_xid));
857                 dprintk("RPC:       xprt = %p, tcp_copied = %lu, "
858                                 "tcp_offset = %u, tcp_reclen = %u\n",
859                                 xprt, transport->tcp_copied,
860                                 transport->tcp_offset, transport->tcp_reclen);
861                 goto out;
862         }
863
864         dprintk("RPC:       XID %08x read %Zd bytes\n",
865                         ntohl(transport->tcp_xid), r);
866         dprintk("RPC:       xprt = %p, tcp_copied = %lu, tcp_offset = %u, "
867                         "tcp_reclen = %u\n", xprt, transport->tcp_copied,
868                         transport->tcp_offset, transport->tcp_reclen);
869
870         if (transport->tcp_copied == req->rq_private_buf.buflen)
871                 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
872         else if (transport->tcp_offset == transport->tcp_reclen) {
873                 if (transport->tcp_flags & TCP_RCV_LAST_FRAG)
874                         transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
875         }
876
877 out:
878         if (!(transport->tcp_flags & TCP_RCV_COPY_DATA))
879                 xprt_complete_rqst(req->rq_task, transport->tcp_copied);
880         spin_unlock(&xprt->transport_lock);
881         xs_tcp_check_fraghdr(transport);
882 }
883
884 static inline void xs_tcp_read_discard(struct sock_xprt *transport, struct xdr_skb_reader *desc)
885 {
886         size_t len;
887
888         len = transport->tcp_reclen - transport->tcp_offset;
889         if (len > desc->count)
890                 len = desc->count;
891         desc->count -= len;
892         desc->offset += len;
893         transport->tcp_offset += len;
894         dprintk("RPC:       discarded %Zu bytes\n", len);
895         xs_tcp_check_fraghdr(transport);
896 }
897
898 static int xs_tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, unsigned int offset, size_t len)
899 {
900         struct rpc_xprt *xprt = rd_desc->arg.data;
901         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
902         struct xdr_skb_reader desc = {
903                 .skb    = skb,
904                 .offset = offset,
905                 .count  = len,
906         };
907
908         dprintk("RPC:       xs_tcp_data_recv started\n");
909         do {
910                 /* Read in a new fragment marker if necessary */
911                 /* Can we ever really expect to get completely empty fragments? */
912                 if (transport->tcp_flags & TCP_RCV_COPY_FRAGHDR) {
913                         xs_tcp_read_fraghdr(xprt, &desc);
914                         continue;
915                 }
916                 /* Read in the xid if necessary */
917                 if (transport->tcp_flags & TCP_RCV_COPY_XID) {
918                         xs_tcp_read_xid(transport, &desc);
919                         continue;
920                 }
921                 /* Read in the request data */
922                 if (transport->tcp_flags & TCP_RCV_COPY_DATA) {
923                         xs_tcp_read_request(xprt, &desc);
924                         continue;
925                 }
926                 /* Skip over any trailing bytes on short reads */
927                 xs_tcp_read_discard(transport, &desc);
928         } while (desc.count);
929         dprintk("RPC:       xs_tcp_data_recv done\n");
930         return len - desc.count;
931 }
932
933 /**
934  * xs_tcp_data_ready - "data ready" callback for TCP sockets
935  * @sk: socket with data to read
936  * @bytes: how much data to read
937  *
938  */
939 static void xs_tcp_data_ready(struct sock *sk, int bytes)
940 {
941         struct rpc_xprt *xprt;
942         read_descriptor_t rd_desc;
943
944         dprintk("RPC:       xs_tcp_data_ready...\n");
945
946         read_lock(&sk->sk_callback_lock);
947         if (!(xprt = xprt_from_sock(sk)))
948                 goto out;
949         if (xprt->shutdown)
950                 goto out;
951
952         /* We use rd_desc to pass struct xprt to xs_tcp_data_recv */
953         rd_desc.arg.data = xprt;
954         rd_desc.count = 65536;
955         tcp_read_sock(sk, &rd_desc, xs_tcp_data_recv);
956 out:
957         read_unlock(&sk->sk_callback_lock);
958 }
959
960 /**
961  * xs_tcp_state_change - callback to handle TCP socket state changes
962  * @sk: socket whose state has changed
963  *
964  */
965 static void xs_tcp_state_change(struct sock *sk)
966 {
967         struct rpc_xprt *xprt;
968
969         read_lock(&sk->sk_callback_lock);
970         if (!(xprt = xprt_from_sock(sk)))
971                 goto out;
972         dprintk("RPC:       xs_tcp_state_change client %p...\n", xprt);
973         dprintk("RPC:       state %x conn %d dead %d zapped %d\n",
974                         sk->sk_state, xprt_connected(xprt),
975                         sock_flag(sk, SOCK_DEAD),
976                         sock_flag(sk, SOCK_ZAPPED));
977
978         switch (sk->sk_state) {
979         case TCP_ESTABLISHED:
980                 spin_lock_bh(&xprt->transport_lock);
981                 if (!xprt_test_and_set_connected(xprt)) {
982                         struct sock_xprt *transport = container_of(xprt,
983                                         struct sock_xprt, xprt);
984
985                         /* Reset TCP record info */
986                         transport->tcp_offset = 0;
987                         transport->tcp_reclen = 0;
988                         transport->tcp_copied = 0;
989                         transport->tcp_flags =
990                                 TCP_RCV_COPY_FRAGHDR | TCP_RCV_COPY_XID;
991
992                         xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
993                         xprt_wake_pending_tasks(xprt, 0);
994                 }
995                 spin_unlock_bh(&xprt->transport_lock);
996                 break;
997         case TCP_SYN_SENT:
998         case TCP_SYN_RECV:
999                 break;
1000         case TCP_CLOSE_WAIT:
1001                 /* Try to schedule an autoclose RPC calls */
1002                 set_bit(XPRT_CLOSE_WAIT, &xprt->state);
1003                 if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
1004                         queue_work(rpciod_workqueue, &xprt->task_cleanup);
1005         default:
1006                 xprt_disconnect(xprt);
1007         }
1008  out:
1009         read_unlock(&sk->sk_callback_lock);
1010 }
1011
1012 /**
1013  * xs_udp_write_space - callback invoked when socket buffer space
1014  *                             becomes available
1015  * @sk: socket whose state has changed
1016  *
1017  * Called when more output buffer space is available for this socket.
1018  * We try not to wake our writers until they can make "significant"
1019  * progress, otherwise we'll waste resources thrashing kernel_sendmsg
1020  * with a bunch of small requests.
1021  */
1022 static void xs_udp_write_space(struct sock *sk)
1023 {
1024         read_lock(&sk->sk_callback_lock);
1025
1026         /* from net/core/sock.c:sock_def_write_space */
1027         if (sock_writeable(sk)) {
1028                 struct socket *sock;
1029                 struct rpc_xprt *xprt;
1030
1031                 if (unlikely(!(sock = sk->sk_socket)))
1032                         goto out;
1033                 if (unlikely(!(xprt = xprt_from_sock(sk))))
1034                         goto out;
1035                 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE, &sock->flags)))
1036                         goto out;
1037
1038                 xprt_write_space(xprt);
1039         }
1040
1041  out:
1042         read_unlock(&sk->sk_callback_lock);
1043 }
1044
1045 /**
1046  * xs_tcp_write_space - callback invoked when socket buffer space
1047  *                             becomes available
1048  * @sk: socket whose state has changed
1049  *
1050  * Called when more output buffer space is available for this socket.
1051  * We try not to wake our writers until they can make "significant"
1052  * progress, otherwise we'll waste resources thrashing kernel_sendmsg
1053  * with a bunch of small requests.
1054  */
1055 static void xs_tcp_write_space(struct sock *sk)
1056 {
1057         read_lock(&sk->sk_callback_lock);
1058
1059         /* from net/core/stream.c:sk_stream_write_space */
1060         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
1061                 struct socket *sock;
1062                 struct rpc_xprt *xprt;
1063
1064                 if (unlikely(!(sock = sk->sk_socket)))
1065                         goto out;
1066                 if (unlikely(!(xprt = xprt_from_sock(sk))))
1067                         goto out;
1068                 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE, &sock->flags)))
1069                         goto out;
1070
1071                 xprt_write_space(xprt);
1072         }
1073
1074  out:
1075         read_unlock(&sk->sk_callback_lock);
1076 }
1077
1078 static void xs_udp_do_set_buffer_size(struct rpc_xprt *xprt)
1079 {
1080         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1081         struct sock *sk = transport->inet;
1082
1083         if (transport->rcvsize) {
1084                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
1085                 sk->sk_rcvbuf = transport->rcvsize * xprt->max_reqs * 2;
1086         }
1087         if (transport->sndsize) {
1088                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
1089                 sk->sk_sndbuf = transport->sndsize * xprt->max_reqs * 2;
1090                 sk->sk_write_space(sk);
1091         }
1092 }
1093
1094 /**
1095  * xs_udp_set_buffer_size - set send and receive limits
1096  * @xprt: generic transport
1097  * @sndsize: requested size of send buffer, in bytes
1098  * @rcvsize: requested size of receive buffer, in bytes
1099  *
1100  * Set socket send and receive buffer size limits.
1101  */
1102 static void xs_udp_set_buffer_size(struct rpc_xprt *xprt, size_t sndsize, size_t rcvsize)
1103 {
1104         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1105
1106         transport->sndsize = 0;
1107         if (sndsize)
1108                 transport->sndsize = sndsize + 1024;
1109         transport->rcvsize = 0;
1110         if (rcvsize)
1111                 transport->rcvsize = rcvsize + 1024;
1112
1113         xs_udp_do_set_buffer_size(xprt);
1114 }
1115
1116 /**
1117  * xs_udp_timer - called when a retransmit timeout occurs on a UDP transport
1118  * @task: task that timed out
1119  *
1120  * Adjust the congestion window after a retransmit timeout has occurred.
1121  */
1122 static void xs_udp_timer(struct rpc_task *task)
1123 {
1124         xprt_adjust_cwnd(task, -ETIMEDOUT);
1125 }
1126
1127 static unsigned short xs_get_random_port(void)
1128 {
1129         unsigned short range = xprt_max_resvport - xprt_min_resvport;
1130         unsigned short rand = (unsigned short) net_random() % range;
1131         return rand + xprt_min_resvport;
1132 }
1133
1134 /**
1135  * xs_set_port - reset the port number in the remote endpoint address
1136  * @xprt: generic transport
1137  * @port: new port number
1138  *
1139  */
1140 static void xs_set_port(struct rpc_xprt *xprt, unsigned short port)
1141 {
1142         struct sockaddr_in *sap = (struct sockaddr_in *) &xprt->addr;
1143
1144         dprintk("RPC:       setting port for xprt %p to %u\n", xprt, port);
1145
1146         sap->sin_port = htons(port);
1147 }
1148
1149 static int xs_bind(struct sock_xprt *transport, struct socket *sock)
1150 {
1151         struct sockaddr_in myaddr = {
1152                 .sin_family = AF_INET,
1153         };
1154         struct sockaddr_in *sa;
1155         int err;
1156         unsigned short port = transport->port;
1157
1158         if (!transport->xprt.resvport)
1159                 port = 0;
1160         sa = (struct sockaddr_in *)&transport->addr;
1161         myaddr.sin_addr = sa->sin_addr;
1162         do {
1163                 myaddr.sin_port = htons(port);
1164                 err = kernel_bind(sock, (struct sockaddr *) &myaddr,
1165                                                 sizeof(myaddr));
1166                 if (!transport->xprt.resvport)
1167                         break;
1168                 if (err == 0) {
1169                         transport->port = port;
1170                         break;
1171                 }
1172                 if (port <= xprt_min_resvport)
1173                         port = xprt_max_resvport;
1174                 else
1175                         port--;
1176         } while (err == -EADDRINUSE && port != transport->port);
1177         dprintk("RPC:       xs_bind "NIPQUAD_FMT":%u: %s (%d)\n",
1178                 NIPQUAD(myaddr.sin_addr), port, err ? "failed" : "ok", err);
1179         return err;
1180 }
1181
1182 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1183 static struct lock_class_key xs_key[2];
1184 static struct lock_class_key xs_slock_key[2];
1185
1186 static inline void xs_reclassify_socket(struct socket *sock)
1187 {
1188         struct sock *sk = sock->sk;
1189         BUG_ON(sock_owned_by_user(sk));
1190         switch (sk->sk_family) {
1191         case AF_INET:
1192                 sock_lock_init_class_and_name(sk, "slock-AF_INET-NFS",
1193                         &xs_slock_key[0], "sk_lock-AF_INET-NFS", &xs_key[0]);
1194                 break;
1195
1196         case AF_INET6:
1197                 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NFS",
1198                         &xs_slock_key[1], "sk_lock-AF_INET6-NFS", &xs_key[1]);
1199                 break;
1200
1201         default:
1202                 BUG();
1203         }
1204 }
1205 #else
1206 static inline void xs_reclassify_socket(struct socket *sock)
1207 {
1208 }
1209 #endif
1210
1211 /**
1212  * xs_udp_connect_worker - set up a UDP socket
1213  * @work: RPC transport to connect
1214  *
1215  * Invoked by a work queue tasklet.
1216  */
1217 static void xs_udp_connect_worker(struct work_struct *work)
1218 {
1219         struct sock_xprt *transport =
1220                 container_of(work, struct sock_xprt, connect_worker.work);
1221         struct rpc_xprt *xprt = &transport->xprt;
1222         struct socket *sock = transport->sock;
1223         int err, status = -EIO;
1224
1225         if (xprt->shutdown || !xprt_bound(xprt))
1226                 goto out;
1227
1228         /* Start by resetting any existing state */
1229         xs_close(xprt);
1230
1231         if ((err = sock_create_kern(PF_INET, SOCK_DGRAM, IPPROTO_UDP, &sock)) < 0) {
1232                 dprintk("RPC:       can't create UDP transport socket (%d).\n", -err);
1233                 goto out;
1234         }
1235         xs_reclassify_socket(sock);
1236
1237         if (xs_bind(transport, sock)) {
1238                 sock_release(sock);
1239                 goto out;
1240         }
1241
1242         dprintk("RPC:       worker connecting xprt %p to address: %s\n",
1243                         xprt, xprt->address_strings[RPC_DISPLAY_ALL]);
1244
1245         if (!transport->inet) {
1246                 struct sock *sk = sock->sk;
1247
1248                 write_lock_bh(&sk->sk_callback_lock);
1249
1250                 sk->sk_user_data = xprt;
1251                 transport->old_data_ready = sk->sk_data_ready;
1252                 transport->old_state_change = sk->sk_state_change;
1253                 transport->old_write_space = sk->sk_write_space;
1254                 sk->sk_data_ready = xs_udp_data_ready;
1255                 sk->sk_write_space = xs_udp_write_space;
1256                 sk->sk_no_check = UDP_CSUM_NORCV;
1257                 sk->sk_allocation = GFP_ATOMIC;
1258
1259                 xprt_set_connected(xprt);
1260
1261                 /* Reset to new socket */
1262                 transport->sock = sock;
1263                 transport->inet = sk;
1264
1265                 write_unlock_bh(&sk->sk_callback_lock);
1266         }
1267         xs_udp_do_set_buffer_size(xprt);
1268         status = 0;
1269 out:
1270         xprt_wake_pending_tasks(xprt, status);
1271         xprt_clear_connecting(xprt);
1272 }
1273
1274 /*
1275  * We need to preserve the port number so the reply cache on the server can
1276  * find our cached RPC replies when we get around to reconnecting.
1277  */
1278 static void xs_tcp_reuse_connection(struct rpc_xprt *xprt)
1279 {
1280         int result;
1281         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1282         struct sockaddr any;
1283
1284         dprintk("RPC:       disconnecting xprt %p to reuse port\n", xprt);
1285
1286         /*
1287          * Disconnect the transport socket by doing a connect operation
1288          * with AF_UNSPEC.  This should return immediately...
1289          */
1290         memset(&any, 0, sizeof(any));
1291         any.sa_family = AF_UNSPEC;
1292         result = kernel_connect(transport->sock, &any, sizeof(any), 0);
1293         if (result)
1294                 dprintk("RPC:       AF_UNSPEC connect return code %d\n",
1295                                 result);
1296 }
1297
1298 /**
1299  * xs_tcp_connect_worker - connect a TCP socket to a remote endpoint
1300  * @work: RPC transport to connect
1301  *
1302  * Invoked by a work queue tasklet.
1303  */
1304 static void xs_tcp_connect_worker(struct work_struct *work)
1305 {
1306         struct sock_xprt *transport =
1307                 container_of(work, struct sock_xprt, connect_worker.work);
1308         struct rpc_xprt *xprt = &transport->xprt;
1309         struct socket *sock = transport->sock;
1310         int err, status = -EIO;
1311
1312         if (xprt->shutdown || !xprt_bound(xprt))
1313                 goto out;
1314
1315         if (!sock) {
1316                 /* start from scratch */
1317                 if ((err = sock_create_kern(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock)) < 0) {
1318                         dprintk("RPC:       can't create TCP transport "
1319                                         "socket (%d).\n", -err);
1320                         goto out;
1321                 }
1322                 xs_reclassify_socket(sock);
1323
1324                 if (xs_bind(transport, sock)) {
1325                         sock_release(sock);
1326                         goto out;
1327                 }
1328         } else
1329                 /* "close" the socket, preserving the local port */
1330                 xs_tcp_reuse_connection(xprt);
1331
1332         dprintk("RPC:       worker connecting xprt %p to address: %s\n",
1333                         xprt, xprt->address_strings[RPC_DISPLAY_ALL]);
1334
1335         if (!transport->inet) {
1336                 struct sock *sk = sock->sk;
1337
1338                 write_lock_bh(&sk->sk_callback_lock);
1339
1340                 sk->sk_user_data = xprt;
1341                 transport->old_data_ready = sk->sk_data_ready;
1342                 transport->old_state_change = sk->sk_state_change;
1343                 transport->old_write_space = sk->sk_write_space;
1344                 sk->sk_data_ready = xs_tcp_data_ready;
1345                 sk->sk_state_change = xs_tcp_state_change;
1346                 sk->sk_write_space = xs_tcp_write_space;
1347                 sk->sk_allocation = GFP_ATOMIC;
1348
1349                 /* socket options */
1350                 sk->sk_userlocks |= SOCK_BINDPORT_LOCK;
1351                 sock_reset_flag(sk, SOCK_LINGER);
1352                 tcp_sk(sk)->linger2 = 0;
1353                 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
1354
1355                 xprt_clear_connected(xprt);
1356
1357                 /* Reset to new socket */
1358                 transport->sock = sock;
1359                 transport->inet = sk;
1360
1361                 write_unlock_bh(&sk->sk_callback_lock);
1362         }
1363
1364         /* Tell the socket layer to start connecting... */
1365         xprt->stat.connect_count++;
1366         xprt->stat.connect_start = jiffies;
1367         status = kernel_connect(sock, (struct sockaddr *) &xprt->addr,
1368                         xprt->addrlen, O_NONBLOCK);
1369         dprintk("RPC:       %p connect status %d connected %d sock state %d\n",
1370                         xprt, -status, xprt_connected(xprt),
1371                         sock->sk->sk_state);
1372         if (status < 0) {
1373                 switch (status) {
1374                         case -EINPROGRESS:
1375                         case -EALREADY:
1376                                 goto out_clear;
1377                         case -ECONNREFUSED:
1378                         case -ECONNRESET:
1379                                 /* retry with existing socket, after a delay */
1380                                 break;
1381                         default:
1382                                 /* get rid of existing socket, and retry */
1383                                 xs_close(xprt);
1384                                 break;
1385                 }
1386         }
1387 out:
1388         xprt_wake_pending_tasks(xprt, status);
1389 out_clear:
1390         xprt_clear_connecting(xprt);
1391 }
1392
1393 /**
1394  * xs_connect - connect a socket to a remote endpoint
1395  * @task: address of RPC task that manages state of connect request
1396  *
1397  * TCP: If the remote end dropped the connection, delay reconnecting.
1398  *
1399  * UDP socket connects are synchronous, but we use a work queue anyway
1400  * to guarantee that even unprivileged user processes can set up a
1401  * socket on a privileged port.
1402  *
1403  * If a UDP socket connect fails, the delay behavior here prevents
1404  * retry floods (hard mounts).
1405  */
1406 static void xs_connect(struct rpc_task *task)
1407 {
1408         struct rpc_xprt *xprt = task->tk_xprt;
1409         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1410
1411         if (xprt_test_and_set_connecting(xprt))
1412                 return;
1413
1414         if (transport->sock != NULL) {
1415                 dprintk("RPC:       xs_connect delayed xprt %p for %lu "
1416                                 "seconds\n",
1417                                 xprt, xprt->reestablish_timeout / HZ);
1418                 queue_delayed_work(rpciod_workqueue,
1419                                    &transport->connect_worker,
1420                                    xprt->reestablish_timeout);
1421                 xprt->reestablish_timeout <<= 1;
1422                 if (xprt->reestablish_timeout > XS_TCP_MAX_REEST_TO)
1423                         xprt->reestablish_timeout = XS_TCP_MAX_REEST_TO;
1424         } else {
1425                 dprintk("RPC:       xs_connect scheduled xprt %p\n", xprt);
1426                 queue_delayed_work(rpciod_workqueue,
1427                                    &transport->connect_worker, 0);
1428         }
1429 }
1430
1431 /**
1432  * xs_udp_print_stats - display UDP socket-specifc stats
1433  * @xprt: rpc_xprt struct containing statistics
1434  * @seq: output file
1435  *
1436  */
1437 static void xs_udp_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
1438 {
1439         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1440
1441         seq_printf(seq, "\txprt:\tudp %u %lu %lu %lu %lu %Lu %Lu\n",
1442                         transport->port,
1443                         xprt->stat.bind_count,
1444                         xprt->stat.sends,
1445                         xprt->stat.recvs,
1446                         xprt->stat.bad_xids,
1447                         xprt->stat.req_u,
1448                         xprt->stat.bklog_u);
1449 }
1450
1451 /**
1452  * xs_tcp_print_stats - display TCP socket-specifc stats
1453  * @xprt: rpc_xprt struct containing statistics
1454  * @seq: output file
1455  *
1456  */
1457 static void xs_tcp_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
1458 {
1459         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1460         long idle_time = 0;
1461
1462         if (xprt_connected(xprt))
1463                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
1464
1465         seq_printf(seq, "\txprt:\ttcp %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu\n",
1466                         transport->port,
1467                         xprt->stat.bind_count,
1468                         xprt->stat.connect_count,
1469                         xprt->stat.connect_time,
1470                         idle_time,
1471                         xprt->stat.sends,
1472                         xprt->stat.recvs,
1473                         xprt->stat.bad_xids,
1474                         xprt->stat.req_u,
1475                         xprt->stat.bklog_u);
1476 }
1477
1478 static struct rpc_xprt_ops xs_udp_ops = {
1479         .set_buffer_size        = xs_udp_set_buffer_size,
1480         .reserve_xprt           = xprt_reserve_xprt_cong,
1481         .release_xprt           = xprt_release_xprt_cong,
1482         .rpcbind                = rpcb_getport_async,
1483         .set_port               = xs_set_port,
1484         .connect                = xs_connect,
1485         .buf_alloc              = rpc_malloc,
1486         .buf_free               = rpc_free,
1487         .send_request           = xs_udp_send_request,
1488         .set_retrans_timeout    = xprt_set_retrans_timeout_rtt,
1489         .timer                  = xs_udp_timer,
1490         .release_request        = xprt_release_rqst_cong,
1491         .close                  = xs_close,
1492         .destroy                = xs_destroy,
1493         .print_stats            = xs_udp_print_stats,
1494 };
1495
1496 static struct rpc_xprt_ops xs_tcp_ops = {
1497         .reserve_xprt           = xprt_reserve_xprt,
1498         .release_xprt           = xs_tcp_release_xprt,
1499         .rpcbind                = rpcb_getport_async,
1500         .set_port               = xs_set_port,
1501         .connect                = xs_connect,
1502         .buf_alloc              = rpc_malloc,
1503         .buf_free               = rpc_free,
1504         .send_request           = xs_tcp_send_request,
1505         .set_retrans_timeout    = xprt_set_retrans_timeout_def,
1506         .close                  = xs_close,
1507         .destroy                = xs_destroy,
1508         .print_stats            = xs_tcp_print_stats,
1509 };
1510
1511 static struct rpc_xprt *xs_setup_xprt(struct rpc_xprtsock_create *args, unsigned int slot_table_size)
1512 {
1513         struct rpc_xprt *xprt;
1514         struct sock_xprt *new;
1515
1516         if (args->addrlen > sizeof(xprt->addr)) {
1517                 dprintk("RPC:       xs_setup_xprt: address too large\n");
1518                 return ERR_PTR(-EBADF);
1519         }
1520
1521         new = kzalloc(sizeof(*new), GFP_KERNEL);
1522         if (new == NULL) {
1523                 dprintk("RPC:       xs_setup_xprt: couldn't allocate "
1524                                 "rpc_xprt\n");
1525                 return ERR_PTR(-ENOMEM);
1526         }
1527         xprt = &new->xprt;
1528
1529         xprt->max_reqs = slot_table_size;
1530         xprt->slot = kcalloc(xprt->max_reqs, sizeof(struct rpc_rqst), GFP_KERNEL);
1531         if (xprt->slot == NULL) {
1532                 kfree(xprt);
1533                 dprintk("RPC:       xs_setup_xprt: couldn't allocate slot "
1534                                 "table\n");
1535                 return ERR_PTR(-ENOMEM);
1536         }
1537
1538         memcpy(&xprt->addr, args->dstaddr, args->addrlen);
1539         xprt->addrlen = args->addrlen;
1540         if (args->srcaddr)
1541                 memcpy(&new->addr, args->srcaddr, args->addrlen);
1542         new->port = xs_get_random_port();
1543
1544         return xprt;
1545 }
1546
1547 /**
1548  * xs_setup_udp - Set up transport to use a UDP socket
1549  * @args: rpc transport creation arguments
1550  *
1551  */
1552 struct rpc_xprt *xs_setup_udp(struct rpc_xprtsock_create *args)
1553 {
1554         struct rpc_xprt *xprt;
1555         struct sock_xprt *transport;
1556
1557         xprt = xs_setup_xprt(args, xprt_udp_slot_table_entries);
1558         if (IS_ERR(xprt))
1559                 return xprt;
1560         transport = container_of(xprt, struct sock_xprt, xprt);
1561
1562         if (ntohs(((struct sockaddr_in *)args->dstaddr)->sin_port) != 0)
1563                 xprt_set_bound(xprt);
1564
1565         xprt->prot = IPPROTO_UDP;
1566         xprt->tsh_size = 0;
1567         /* XXX: header size can vary due to auth type, IPv6, etc. */
1568         xprt->max_payload = (1U << 16) - (MAX_HEADER << 3);
1569
1570         INIT_DELAYED_WORK(&transport->connect_worker, xs_udp_connect_worker);
1571         xprt->bind_timeout = XS_BIND_TO;
1572         xprt->connect_timeout = XS_UDP_CONN_TO;
1573         xprt->reestablish_timeout = XS_UDP_REEST_TO;
1574         xprt->idle_timeout = XS_IDLE_DISC_TO;
1575
1576         xprt->ops = &xs_udp_ops;
1577
1578         if (args->timeout)
1579                 xprt->timeout = *args->timeout;
1580         else
1581                 xprt_set_timeout(&xprt->timeout, 5, 5 * HZ);
1582
1583         xs_format_peer_addresses(xprt);
1584         dprintk("RPC:       set up transport to address %s\n",
1585                         xprt->address_strings[RPC_DISPLAY_ALL]);
1586
1587         return xprt;
1588 }
1589
1590 /**
1591  * xs_setup_tcp - Set up transport to use a TCP socket
1592  * @args: rpc transport creation arguments
1593  *
1594  */
1595 struct rpc_xprt *xs_setup_tcp(struct rpc_xprtsock_create *args)
1596 {
1597         struct rpc_xprt *xprt;
1598         struct sock_xprt *transport;
1599
1600         xprt = xs_setup_xprt(args, xprt_tcp_slot_table_entries);
1601         if (IS_ERR(xprt))
1602                 return xprt;
1603         transport = container_of(xprt, struct sock_xprt, xprt);
1604
1605         if (ntohs(((struct sockaddr_in *)args->dstaddr)->sin_port) != 0)
1606                 xprt_set_bound(xprt);
1607
1608         xprt->prot = IPPROTO_TCP;
1609         xprt->tsh_size = sizeof(rpc_fraghdr) / sizeof(u32);
1610         xprt->max_payload = RPC_MAX_FRAGMENT_SIZE;
1611
1612         INIT_DELAYED_WORK(&transport->connect_worker, xs_tcp_connect_worker);
1613         xprt->bind_timeout = XS_BIND_TO;
1614         xprt->connect_timeout = XS_TCP_CONN_TO;
1615         xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
1616         xprt->idle_timeout = XS_IDLE_DISC_TO;
1617
1618         xprt->ops = &xs_tcp_ops;
1619
1620         if (args->timeout)
1621                 xprt->timeout = *args->timeout;
1622         else
1623                 xprt_set_timeout(&xprt->timeout, 2, 60 * HZ);
1624
1625         xs_format_peer_addresses(xprt);
1626         dprintk("RPC:       set up transport to address %s\n",
1627                         xprt->address_strings[RPC_DISPLAY_ALL]);
1628
1629         return xprt;
1630 }
1631
1632 /**
1633  * init_socket_xprt - set up xprtsock's sysctls
1634  *
1635  */
1636 int init_socket_xprt(void)
1637 {
1638 #ifdef RPC_DEBUG
1639         if (!sunrpc_table_header)
1640                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
1641 #endif
1642
1643         return 0;
1644 }
1645
1646 /**
1647  * cleanup_socket_xprt - remove xprtsock's sysctls
1648  *
1649  */
1650 void cleanup_socket_xprt(void)
1651 {
1652 #ifdef RPC_DEBUG
1653         if (sunrpc_table_header) {
1654                 unregister_sysctl_table(sunrpc_table_header);
1655                 sunrpc_table_header = NULL;
1656         }
1657 #endif
1658 }