Merge branch 'upstream'
[linux-2.6] / 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/sched.h>
22 #include <linux/pagemap.h>
23 #include <linux/errno.h>
24 #include <linux/socket.h>
25 #include <linux/in.h>
26 #include <linux/net.h>
27 #include <linux/mm.h>
28 #include <linux/udp.h>
29 #include <linux/tcp.h>
30 #include <linux/sunrpc/clnt.h>
31 #include <linux/sunrpc/sched.h>
32 #include <linux/file.h>
33
34 #include <net/sock.h>
35 #include <net/checksum.h>
36 #include <net/udp.h>
37 #include <net/tcp.h>
38
39 /*
40  * xprtsock tunables
41  */
42 unsigned int xprt_udp_slot_table_entries = RPC_DEF_SLOT_TABLE;
43 unsigned int xprt_tcp_slot_table_entries = RPC_DEF_SLOT_TABLE;
44
45 unsigned int xprt_min_resvport = RPC_DEF_MIN_RESVPORT;
46 unsigned int xprt_max_resvport = RPC_DEF_MAX_RESVPORT;
47
48 /*
49  * How many times to try sending a request on a socket before waiting
50  * for the socket buffer to clear.
51  */
52 #define XS_SENDMSG_RETRY        (10U)
53
54 /*
55  * Time out for an RPC UDP socket connect.  UDP socket connects are
56  * synchronous, but we set a timeout anyway in case of resource
57  * exhaustion on the local host.
58  */
59 #define XS_UDP_CONN_TO          (5U * HZ)
60
61 /*
62  * Wait duration for an RPC TCP connection to be established.  Solaris
63  * NFS over TCP uses 60 seconds, for example, which is in line with how
64  * long a server takes to reboot.
65  */
66 #define XS_TCP_CONN_TO          (60U * HZ)
67
68 /*
69  * Wait duration for a reply from the RPC portmapper.
70  */
71 #define XS_BIND_TO              (60U * HZ)
72
73 /*
74  * Delay if a UDP socket connect error occurs.  This is most likely some
75  * kind of resource problem on the local host.
76  */
77 #define XS_UDP_REEST_TO         (2U * HZ)
78
79 /*
80  * The reestablish timeout allows clients to delay for a bit before attempting
81  * to reconnect to a server that just dropped our connection.
82  *
83  * We implement an exponential backoff when trying to reestablish a TCP
84  * transport connection with the server.  Some servers like to drop a TCP
85  * connection when they are overworked, so we start with a short timeout and
86  * increase over time if the server is down or not responding.
87  */
88 #define XS_TCP_INIT_REEST_TO    (3U * HZ)
89 #define XS_TCP_MAX_REEST_TO     (5U * 60 * HZ)
90
91 /*
92  * TCP idle timeout; client drops the transport socket if it is idle
93  * for this long.  Note that we also timeout UDP sockets to prevent
94  * holding port numbers when there is no RPC traffic.
95  */
96 #define XS_IDLE_DISC_TO         (5U * 60 * HZ)
97
98 #ifdef RPC_DEBUG
99 # undef  RPC_DEBUG_DATA
100 # define RPCDBG_FACILITY        RPCDBG_TRANS
101 #endif
102
103 #ifdef RPC_DEBUG_DATA
104 static void xs_pktdump(char *msg, u32 *packet, unsigned int count)
105 {
106         u8 *buf = (u8 *) packet;
107         int j;
108
109         dprintk("RPC:      %s\n", msg);
110         for (j = 0; j < count && j < 128; j += 4) {
111                 if (!(j & 31)) {
112                         if (j)
113                                 dprintk("\n");
114                         dprintk("0x%04x ", j);
115                 }
116                 dprintk("%02x%02x%02x%02x ",
117                         buf[j], buf[j+1], buf[j+2], buf[j+3]);
118         }
119         dprintk("\n");
120 }
121 #else
122 static inline void xs_pktdump(char *msg, u32 *packet, unsigned int count)
123 {
124         /* NOP */
125 }
126 #endif
127
128 #define XS_SENDMSG_FLAGS        (MSG_DONTWAIT | MSG_NOSIGNAL)
129
130 static inline int xs_send_head(struct socket *sock, struct sockaddr *addr, int addrlen, struct xdr_buf *xdr, unsigned int base, unsigned int len)
131 {
132         struct kvec iov = {
133                 .iov_base       = xdr->head[0].iov_base + base,
134                 .iov_len        = len - base,
135         };
136         struct msghdr msg = {
137                 .msg_name       = addr,
138                 .msg_namelen    = addrlen,
139                 .msg_flags      = XS_SENDMSG_FLAGS,
140         };
141
142         if (xdr->len > len)
143                 msg.msg_flags |= MSG_MORE;
144
145         if (likely(iov.iov_len))
146                 return kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
147         return kernel_sendmsg(sock, &msg, NULL, 0, 0);
148 }
149
150 static int xs_send_tail(struct socket *sock, struct xdr_buf *xdr, unsigned int base, unsigned int len)
151 {
152         struct kvec iov = {
153                 .iov_base       = xdr->tail[0].iov_base + base,
154                 .iov_len        = len - base,
155         };
156         struct msghdr msg = {
157                 .msg_flags      = XS_SENDMSG_FLAGS,
158         };
159
160         return kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
161 }
162
163 /**
164  * xs_sendpages - write pages directly to a socket
165  * @sock: socket to send on
166  * @addr: UDP only -- address of destination
167  * @addrlen: UDP only -- length of destination address
168  * @xdr: buffer containing this request
169  * @base: starting position in the buffer
170  *
171  */
172 static inline int xs_sendpages(struct socket *sock, struct sockaddr *addr, int addrlen, struct xdr_buf *xdr, unsigned int base)
173 {
174         struct page **ppage = xdr->pages;
175         unsigned int len, pglen = xdr->page_len;
176         int err, ret = 0;
177         ssize_t (*sendpage)(struct socket *, struct page *, int, size_t, int);
178
179         if (unlikely(!sock))
180                 return -ENOTCONN;
181
182         clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
183
184         len = xdr->head[0].iov_len;
185         if (base < len || (addr != NULL && base == 0)) {
186                 err = xs_send_head(sock, addr, addrlen, xdr, base, len);
187                 if (ret == 0)
188                         ret = err;
189                 else if (err > 0)
190                         ret += err;
191                 if (err != (len - base))
192                         goto out;
193                 base = 0;
194         } else
195                 base -= len;
196
197         if (unlikely(pglen == 0))
198                 goto copy_tail;
199         if (unlikely(base >= pglen)) {
200                 base -= pglen;
201                 goto copy_tail;
202         }
203         if (base || xdr->page_base) {
204                 pglen -= base;
205                 base += xdr->page_base;
206                 ppage += base >> PAGE_CACHE_SHIFT;
207                 base &= ~PAGE_CACHE_MASK;
208         }
209
210         sendpage = sock->ops->sendpage ? : sock_no_sendpage;
211         do {
212                 int flags = XS_SENDMSG_FLAGS;
213
214                 len = PAGE_CACHE_SIZE;
215                 if (base)
216                         len -= base;
217                 if (pglen < len)
218                         len = pglen;
219
220                 if (pglen != len || xdr->tail[0].iov_len != 0)
221                         flags |= MSG_MORE;
222
223                 /* Hmm... We might be dealing with highmem pages */
224                 if (PageHighMem(*ppage))
225                         sendpage = sock_no_sendpage;
226                 err = sendpage(sock, *ppage, base, len, flags);
227                 if (ret == 0)
228                         ret = err;
229                 else if (err > 0)
230                         ret += err;
231                 if (err != len)
232                         goto out;
233                 base = 0;
234                 ppage++;
235         } while ((pglen -= len) != 0);
236 copy_tail:
237         len = xdr->tail[0].iov_len;
238         if (base < len) {
239                 err = xs_send_tail(sock, xdr, base, len);
240                 if (ret == 0)
241                         ret = err;
242                 else if (err > 0)
243                         ret += err;
244         }
245 out:
246         return ret;
247 }
248
249 /**
250  * xs_nospace - place task on wait queue if transmit was incomplete
251  * @task: task to put to sleep
252  *
253  */
254 static void xs_nospace(struct rpc_task *task)
255 {
256         struct rpc_rqst *req = task->tk_rqstp;
257         struct rpc_xprt *xprt = req->rq_xprt;
258
259         dprintk("RPC: %4d xmit incomplete (%u left of %u)\n",
260                         task->tk_pid, req->rq_slen - req->rq_bytes_sent,
261                         req->rq_slen);
262
263         if (test_bit(SOCK_ASYNC_NOSPACE, &xprt->sock->flags)) {
264                 /* Protect against races with write_space */
265                 spin_lock_bh(&xprt->transport_lock);
266
267                 /* Don't race with disconnect */
268                 if (!xprt_connected(xprt))
269                         task->tk_status = -ENOTCONN;
270                 else if (test_bit(SOCK_NOSPACE, &xprt->sock->flags))
271                         xprt_wait_for_buffer_space(task);
272
273                 spin_unlock_bh(&xprt->transport_lock);
274         } else
275                 /* Keep holding the socket if it is blocked */
276                 rpc_delay(task, HZ>>4);
277 }
278
279 /**
280  * xs_udp_send_request - write an RPC request to a UDP socket
281  * @task: address of RPC task that manages the state of an RPC request
282  *
283  * Return values:
284  *        0:    The request has been sent
285  *   EAGAIN:    The socket was blocked, please call again later to
286  *              complete the request
287  * ENOTCONN:    Caller needs to invoke connect logic then call again
288  *    other:    Some other error occured, the request was not sent
289  */
290 static int xs_udp_send_request(struct rpc_task *task)
291 {
292         struct rpc_rqst *req = task->tk_rqstp;
293         struct rpc_xprt *xprt = req->rq_xprt;
294         struct xdr_buf *xdr = &req->rq_snd_buf;
295         int status;
296
297         xs_pktdump("packet data:",
298                                 req->rq_svec->iov_base,
299                                 req->rq_svec->iov_len);
300
301         req->rq_xtime = jiffies;
302         status = xs_sendpages(xprt->sock, (struct sockaddr *) &xprt->addr,
303                                 sizeof(xprt->addr), xdr, req->rq_bytes_sent);
304
305         dprintk("RPC:      xs_udp_send_request(%u) = %d\n",
306                         xdr->len - req->rq_bytes_sent, status);
307
308         if (likely(status >= (int) req->rq_slen))
309                 return 0;
310
311         /* Still some bytes left; set up for a retry later. */
312         if (status > 0)
313                 status = -EAGAIN;
314
315         switch (status) {
316         case -ENETUNREACH:
317         case -EPIPE:
318         case -ECONNREFUSED:
319                 /* When the server has died, an ICMP port unreachable message
320                  * prompts ECONNREFUSED. */
321                 break;
322         case -EAGAIN:
323                 xs_nospace(task);
324                 break;
325         default:
326                 dprintk("RPC:      sendmsg returned unrecognized error %d\n",
327                         -status);
328                 break;
329         }
330
331         return status;
332 }
333
334 static inline void xs_encode_tcp_record_marker(struct xdr_buf *buf)
335 {
336         u32 reclen = buf->len - sizeof(rpc_fraghdr);
337         rpc_fraghdr *base = buf->head[0].iov_base;
338         *base = htonl(RPC_LAST_STREAM_FRAGMENT | reclen);
339 }
340
341 /**
342  * xs_tcp_send_request - write an RPC request to a TCP socket
343  * @task: address of RPC task that manages the state of an RPC request
344  *
345  * Return values:
346  *        0:    The request has been sent
347  *   EAGAIN:    The socket was blocked, please call again later to
348  *              complete the request
349  * ENOTCONN:    Caller needs to invoke connect logic then call again
350  *    other:    Some other error occured, the request was not sent
351  *
352  * XXX: In the case of soft timeouts, should we eventually give up
353  *      if sendmsg is not able to make progress?
354  */
355 static int xs_tcp_send_request(struct rpc_task *task)
356 {
357         struct rpc_rqst *req = task->tk_rqstp;
358         struct rpc_xprt *xprt = req->rq_xprt;
359         struct xdr_buf *xdr = &req->rq_snd_buf;
360         int status, retry = 0;
361
362         xs_encode_tcp_record_marker(&req->rq_snd_buf);
363
364         xs_pktdump("packet data:",
365                                 req->rq_svec->iov_base,
366                                 req->rq_svec->iov_len);
367
368         /* Continue transmitting the packet/record. We must be careful
369          * to cope with writespace callbacks arriving _after_ we have
370          * called sendmsg(). */
371         while (1) {
372                 req->rq_xtime = jiffies;
373                 status = xs_sendpages(xprt->sock, NULL, 0, xdr,
374                                                 req->rq_bytes_sent);
375
376                 dprintk("RPC:      xs_tcp_send_request(%u) = %d\n",
377                                 xdr->len - req->rq_bytes_sent, status);
378
379                 if (unlikely(status < 0))
380                         break;
381
382                 /* If we've sent the entire packet, immediately
383                  * reset the count of bytes sent. */
384                 req->rq_bytes_sent += status;
385                 if (likely(req->rq_bytes_sent >= req->rq_slen)) {
386                         req->rq_bytes_sent = 0;
387                         return 0;
388                 }
389
390                 status = -EAGAIN;
391                 if (retry++ > XS_SENDMSG_RETRY)
392                         break;
393         }
394
395         switch (status) {
396         case -EAGAIN:
397                 xs_nospace(task);
398                 break;
399         case -ECONNREFUSED:
400         case -ECONNRESET:
401         case -ENOTCONN:
402         case -EPIPE:
403                 status = -ENOTCONN;
404                 break;
405         default:
406                 dprintk("RPC:      sendmsg returned unrecognized error %d\n",
407                         -status);
408                 xprt_disconnect(xprt);
409                 break;
410         }
411
412         return status;
413 }
414
415 /**
416  * xs_close - close a socket
417  * @xprt: transport
418  *
419  * This is used when all requests are complete; ie, no DRC state remains
420  * on the server we want to save.
421  */
422 static void xs_close(struct rpc_xprt *xprt)
423 {
424         struct socket *sock = xprt->sock;
425         struct sock *sk = xprt->inet;
426
427         if (!sk)
428                 goto clear_close_wait;
429
430         dprintk("RPC:      xs_close xprt %p\n", xprt);
431
432         write_lock_bh(&sk->sk_callback_lock);
433         xprt->inet = NULL;
434         xprt->sock = NULL;
435
436         sk->sk_user_data = NULL;
437         sk->sk_data_ready = xprt->old_data_ready;
438         sk->sk_state_change = xprt->old_state_change;
439         sk->sk_write_space = xprt->old_write_space;
440         write_unlock_bh(&sk->sk_callback_lock);
441
442         sk->sk_no_check = 0;
443
444         sock_release(sock);
445 clear_close_wait:
446         smp_mb__before_clear_bit();
447         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
448         smp_mb__after_clear_bit();
449 }
450
451 /**
452  * xs_destroy - prepare to shutdown a transport
453  * @xprt: doomed transport
454  *
455  */
456 static void xs_destroy(struct rpc_xprt *xprt)
457 {
458         dprintk("RPC:      xs_destroy xprt %p\n", xprt);
459
460         cancel_delayed_work(&xprt->connect_worker);
461         flush_scheduled_work();
462
463         xprt_disconnect(xprt);
464         xs_close(xprt);
465         kfree(xprt->slot);
466 }
467
468 static inline struct rpc_xprt *xprt_from_sock(struct sock *sk)
469 {
470         return (struct rpc_xprt *) sk->sk_user_data;
471 }
472
473 /**
474  * xs_udp_data_ready - "data ready" callback for UDP sockets
475  * @sk: socket with data to read
476  * @len: how much data to read
477  *
478  */
479 static void xs_udp_data_ready(struct sock *sk, int len)
480 {
481         struct rpc_task *task;
482         struct rpc_xprt *xprt;
483         struct rpc_rqst *rovr;
484         struct sk_buff *skb;
485         int err, repsize, copied;
486         u32 _xid, *xp;
487
488         read_lock(&sk->sk_callback_lock);
489         dprintk("RPC:      xs_udp_data_ready...\n");
490         if (!(xprt = xprt_from_sock(sk)))
491                 goto out;
492
493         if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
494                 goto out;
495
496         if (xprt->shutdown)
497                 goto dropit;
498
499         repsize = skb->len - sizeof(struct udphdr);
500         if (repsize < 4) {
501                 dprintk("RPC:      impossible RPC reply size %d!\n", repsize);
502                 goto dropit;
503         }
504
505         /* Copy the XID from the skb... */
506         xp = skb_header_pointer(skb, sizeof(struct udphdr),
507                                 sizeof(_xid), &_xid);
508         if (xp == NULL)
509                 goto dropit;
510
511         /* Look up and lock the request corresponding to the given XID */
512         spin_lock(&xprt->transport_lock);
513         rovr = xprt_lookup_rqst(xprt, *xp);
514         if (!rovr)
515                 goto out_unlock;
516         task = rovr->rq_task;
517
518         if ((copied = rovr->rq_private_buf.buflen) > repsize)
519                 copied = repsize;
520
521         /* Suck it into the iovec, verify checksum if not done by hw. */
522         if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
523                 goto out_unlock;
524
525         /* Something worked... */
526         dst_confirm(skb->dst);
527
528         xprt_adjust_cwnd(task, copied);
529         xprt_update_rtt(task);
530         xprt_complete_rqst(task, copied);
531
532  out_unlock:
533         spin_unlock(&xprt->transport_lock);
534  dropit:
535         skb_free_datagram(sk, skb);
536  out:
537         read_unlock(&sk->sk_callback_lock);
538 }
539
540 static inline size_t xs_tcp_copy_data(skb_reader_t *desc, void *p, size_t len)
541 {
542         if (len > desc->count)
543                 len = desc->count;
544         if (skb_copy_bits(desc->skb, desc->offset, p, len)) {
545                 dprintk("RPC:      failed to copy %zu bytes from skb. %zu bytes remain\n",
546                                 len, desc->count);
547                 return 0;
548         }
549         desc->offset += len;
550         desc->count -= len;
551         dprintk("RPC:      copied %zu bytes from skb. %zu bytes remain\n",
552                         len, desc->count);
553         return len;
554 }
555
556 static inline void xs_tcp_read_fraghdr(struct rpc_xprt *xprt, skb_reader_t *desc)
557 {
558         size_t len, used;
559         char *p;
560
561         p = ((char *) &xprt->tcp_recm) + xprt->tcp_offset;
562         len = sizeof(xprt->tcp_recm) - xprt->tcp_offset;
563         used = xs_tcp_copy_data(desc, p, len);
564         xprt->tcp_offset += used;
565         if (used != len)
566                 return;
567
568         xprt->tcp_reclen = ntohl(xprt->tcp_recm);
569         if (xprt->tcp_reclen & RPC_LAST_STREAM_FRAGMENT)
570                 xprt->tcp_flags |= XPRT_LAST_FRAG;
571         else
572                 xprt->tcp_flags &= ~XPRT_LAST_FRAG;
573         xprt->tcp_reclen &= RPC_FRAGMENT_SIZE_MASK;
574
575         xprt->tcp_flags &= ~XPRT_COPY_RECM;
576         xprt->tcp_offset = 0;
577
578         /* Sanity check of the record length */
579         if (unlikely(xprt->tcp_reclen < 4)) {
580                 dprintk("RPC:      invalid TCP record fragment length\n");
581                 xprt_disconnect(xprt);
582                 return;
583         }
584         dprintk("RPC:      reading TCP record fragment of length %d\n",
585                         xprt->tcp_reclen);
586 }
587
588 static void xs_tcp_check_recm(struct rpc_xprt *xprt)
589 {
590         dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u, tcp_flags = %lx\n",
591                         xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen, xprt->tcp_flags);
592         if (xprt->tcp_offset == xprt->tcp_reclen) {
593                 xprt->tcp_flags |= XPRT_COPY_RECM;
594                 xprt->tcp_offset = 0;
595                 if (xprt->tcp_flags & XPRT_LAST_FRAG) {
596                         xprt->tcp_flags &= ~XPRT_COPY_DATA;
597                         xprt->tcp_flags |= XPRT_COPY_XID;
598                         xprt->tcp_copied = 0;
599                 }
600         }
601 }
602
603 static inline void xs_tcp_read_xid(struct rpc_xprt *xprt, skb_reader_t *desc)
604 {
605         size_t len, used;
606         char *p;
607
608         len = sizeof(xprt->tcp_xid) - xprt->tcp_offset;
609         dprintk("RPC:      reading XID (%Zu bytes)\n", len);
610         p = ((char *) &xprt->tcp_xid) + xprt->tcp_offset;
611         used = xs_tcp_copy_data(desc, p, len);
612         xprt->tcp_offset += used;
613         if (used != len)
614                 return;
615         xprt->tcp_flags &= ~XPRT_COPY_XID;
616         xprt->tcp_flags |= XPRT_COPY_DATA;
617         xprt->tcp_copied = 4;
618         dprintk("RPC:      reading reply for XID %08x\n",
619                                                 ntohl(xprt->tcp_xid));
620         xs_tcp_check_recm(xprt);
621 }
622
623 static inline void xs_tcp_read_request(struct rpc_xprt *xprt, skb_reader_t *desc)
624 {
625         struct rpc_rqst *req;
626         struct xdr_buf *rcvbuf;
627         size_t len;
628         ssize_t r;
629
630         /* Find and lock the request corresponding to this xid */
631         spin_lock(&xprt->transport_lock);
632         req = xprt_lookup_rqst(xprt, xprt->tcp_xid);
633         if (!req) {
634                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
635                 dprintk("RPC:      XID %08x request not found!\n",
636                                 ntohl(xprt->tcp_xid));
637                 spin_unlock(&xprt->transport_lock);
638                 return;
639         }
640
641         rcvbuf = &req->rq_private_buf;
642         len = desc->count;
643         if (len > xprt->tcp_reclen - xprt->tcp_offset) {
644                 skb_reader_t my_desc;
645
646                 len = xprt->tcp_reclen - xprt->tcp_offset;
647                 memcpy(&my_desc, desc, sizeof(my_desc));
648                 my_desc.count = len;
649                 r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
650                                           &my_desc, xs_tcp_copy_data);
651                 desc->count -= r;
652                 desc->offset += r;
653         } else
654                 r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
655                                           desc, xs_tcp_copy_data);
656
657         if (r > 0) {
658                 xprt->tcp_copied += r;
659                 xprt->tcp_offset += r;
660         }
661         if (r != len) {
662                 /* Error when copying to the receive buffer,
663                  * usually because we weren't able to allocate
664                  * additional buffer pages. All we can do now
665                  * is turn off XPRT_COPY_DATA, so the request
666                  * will not receive any additional updates,
667                  * and time out.
668                  * Any remaining data from this record will
669                  * be discarded.
670                  */
671                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
672                 dprintk("RPC:      XID %08x truncated request\n",
673                                 ntohl(xprt->tcp_xid));
674                 dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
675                                 xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen);
676                 goto out;
677         }
678
679         dprintk("RPC:      XID %08x read %Zd bytes\n",
680                         ntohl(xprt->tcp_xid), r);
681         dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
682                         xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen);
683
684         if (xprt->tcp_copied == req->rq_private_buf.buflen)
685                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
686         else if (xprt->tcp_offset == xprt->tcp_reclen) {
687                 if (xprt->tcp_flags & XPRT_LAST_FRAG)
688                         xprt->tcp_flags &= ~XPRT_COPY_DATA;
689         }
690
691 out:
692         if (!(xprt->tcp_flags & XPRT_COPY_DATA))
693                 xprt_complete_rqst(req->rq_task, xprt->tcp_copied);
694         spin_unlock(&xprt->transport_lock);
695         xs_tcp_check_recm(xprt);
696 }
697
698 static inline void xs_tcp_read_discard(struct rpc_xprt *xprt, skb_reader_t *desc)
699 {
700         size_t len;
701
702         len = xprt->tcp_reclen - xprt->tcp_offset;
703         if (len > desc->count)
704                 len = desc->count;
705         desc->count -= len;
706         desc->offset += len;
707         xprt->tcp_offset += len;
708         dprintk("RPC:      discarded %Zu bytes\n", len);
709         xs_tcp_check_recm(xprt);
710 }
711
712 static int xs_tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, unsigned int offset, size_t len)
713 {
714         struct rpc_xprt *xprt = rd_desc->arg.data;
715         skb_reader_t desc = {
716                 .skb    = skb,
717                 .offset = offset,
718                 .count  = len,
719                 .csum   = 0
720         };
721
722         dprintk("RPC:      xs_tcp_data_recv started\n");
723         do {
724                 /* Read in a new fragment marker if necessary */
725                 /* Can we ever really expect to get completely empty fragments? */
726                 if (xprt->tcp_flags & XPRT_COPY_RECM) {
727                         xs_tcp_read_fraghdr(xprt, &desc);
728                         continue;
729                 }
730                 /* Read in the xid if necessary */
731                 if (xprt->tcp_flags & XPRT_COPY_XID) {
732                         xs_tcp_read_xid(xprt, &desc);
733                         continue;
734                 }
735                 /* Read in the request data */
736                 if (xprt->tcp_flags & XPRT_COPY_DATA) {
737                         xs_tcp_read_request(xprt, &desc);
738                         continue;
739                 }
740                 /* Skip over any trailing bytes on short reads */
741                 xs_tcp_read_discard(xprt, &desc);
742         } while (desc.count);
743         dprintk("RPC:      xs_tcp_data_recv done\n");
744         return len - desc.count;
745 }
746
747 /**
748  * xs_tcp_data_ready - "data ready" callback for TCP sockets
749  * @sk: socket with data to read
750  * @bytes: how much data to read
751  *
752  */
753 static void xs_tcp_data_ready(struct sock *sk, int bytes)
754 {
755         struct rpc_xprt *xprt;
756         read_descriptor_t rd_desc;
757
758         read_lock(&sk->sk_callback_lock);
759         dprintk("RPC:      xs_tcp_data_ready...\n");
760         if (!(xprt = xprt_from_sock(sk)))
761                 goto out;
762         if (xprt->shutdown)
763                 goto out;
764
765         /* We use rd_desc to pass struct xprt to xs_tcp_data_recv */
766         rd_desc.arg.data = xprt;
767         rd_desc.count = 65536;
768         tcp_read_sock(sk, &rd_desc, xs_tcp_data_recv);
769 out:
770         read_unlock(&sk->sk_callback_lock);
771 }
772
773 /**
774  * xs_tcp_state_change - callback to handle TCP socket state changes
775  * @sk: socket whose state has changed
776  *
777  */
778 static void xs_tcp_state_change(struct sock *sk)
779 {
780         struct rpc_xprt *xprt;
781
782         read_lock(&sk->sk_callback_lock);
783         if (!(xprt = xprt_from_sock(sk)))
784                 goto out;
785         dprintk("RPC:      xs_tcp_state_change client %p...\n", xprt);
786         dprintk("RPC:      state %x conn %d dead %d zapped %d\n",
787                                 sk->sk_state, xprt_connected(xprt),
788                                 sock_flag(sk, SOCK_DEAD),
789                                 sock_flag(sk, SOCK_ZAPPED));
790
791         switch (sk->sk_state) {
792         case TCP_ESTABLISHED:
793                 spin_lock_bh(&xprt->transport_lock);
794                 if (!xprt_test_and_set_connected(xprt)) {
795                         /* Reset TCP record info */
796                         xprt->tcp_offset = 0;
797                         xprt->tcp_reclen = 0;
798                         xprt->tcp_copied = 0;
799                         xprt->tcp_flags = XPRT_COPY_RECM | XPRT_COPY_XID;
800                         xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
801                         xprt_wake_pending_tasks(xprt, 0);
802                 }
803                 spin_unlock_bh(&xprt->transport_lock);
804                 break;
805         case TCP_SYN_SENT:
806         case TCP_SYN_RECV:
807                 break;
808         case TCP_CLOSE_WAIT:
809                 /* Try to schedule an autoclose RPC calls */
810                 set_bit(XPRT_CLOSE_WAIT, &xprt->state);
811                 if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
812                         schedule_work(&xprt->task_cleanup);
813         default:
814                 xprt_disconnect(xprt);
815         }
816  out:
817         read_unlock(&sk->sk_callback_lock);
818 }
819
820 /**
821  * xs_udp_write_space - callback invoked when socket buffer space
822  *                             becomes available
823  * @sk: socket whose state has changed
824  *
825  * Called when more output buffer space is available for this socket.
826  * We try not to wake our writers until they can make "significant"
827  * progress, otherwise we'll waste resources thrashing kernel_sendmsg
828  * with a bunch of small requests.
829  */
830 static void xs_udp_write_space(struct sock *sk)
831 {
832         read_lock(&sk->sk_callback_lock);
833
834         /* from net/core/sock.c:sock_def_write_space */
835         if (sock_writeable(sk)) {
836                 struct socket *sock;
837                 struct rpc_xprt *xprt;
838
839                 if (unlikely(!(sock = sk->sk_socket)))
840                         goto out;
841                 if (unlikely(!(xprt = xprt_from_sock(sk))))
842                         goto out;
843                 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE, &sock->flags)))
844                         goto out;
845
846                 xprt_write_space(xprt);
847         }
848
849  out:
850         read_unlock(&sk->sk_callback_lock);
851 }
852
853 /**
854  * xs_tcp_write_space - callback invoked when socket buffer space
855  *                             becomes available
856  * @sk: socket whose state has changed
857  *
858  * Called when more output buffer space is available for this socket.
859  * We try not to wake our writers until they can make "significant"
860  * progress, otherwise we'll waste resources thrashing kernel_sendmsg
861  * with a bunch of small requests.
862  */
863 static void xs_tcp_write_space(struct sock *sk)
864 {
865         read_lock(&sk->sk_callback_lock);
866
867         /* from net/core/stream.c:sk_stream_write_space */
868         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
869                 struct socket *sock;
870                 struct rpc_xprt *xprt;
871
872                 if (unlikely(!(sock = sk->sk_socket)))
873                         goto out;
874                 if (unlikely(!(xprt = xprt_from_sock(sk))))
875                         goto out;
876                 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE, &sock->flags)))
877                         goto out;
878
879                 xprt_write_space(xprt);
880         }
881
882  out:
883         read_unlock(&sk->sk_callback_lock);
884 }
885
886 static void xs_udp_do_set_buffer_size(struct rpc_xprt *xprt)
887 {
888         struct sock *sk = xprt->inet;
889
890         if (xprt->rcvsize) {
891                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
892                 sk->sk_rcvbuf = xprt->rcvsize * xprt->max_reqs *  2;
893         }
894         if (xprt->sndsize) {
895                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
896                 sk->sk_sndbuf = xprt->sndsize * xprt->max_reqs * 2;
897                 sk->sk_write_space(sk);
898         }
899 }
900
901 /**
902  * xs_udp_set_buffer_size - set send and receive limits
903  * @xprt: generic transport
904  * @sndsize: requested size of send buffer, in bytes
905  * @rcvsize: requested size of receive buffer, in bytes
906  *
907  * Set socket send and receive buffer size limits.
908  */
909 static void xs_udp_set_buffer_size(struct rpc_xprt *xprt, size_t sndsize, size_t rcvsize)
910 {
911         xprt->sndsize = 0;
912         if (sndsize)
913                 xprt->sndsize = sndsize + 1024;
914         xprt->rcvsize = 0;
915         if (rcvsize)
916                 xprt->rcvsize = rcvsize + 1024;
917
918         xs_udp_do_set_buffer_size(xprt);
919 }
920
921 /**
922  * xs_udp_timer - called when a retransmit timeout occurs on a UDP transport
923  * @task: task that timed out
924  *
925  * Adjust the congestion window after a retransmit timeout has occurred.
926  */
927 static void xs_udp_timer(struct rpc_task *task)
928 {
929         xprt_adjust_cwnd(task, -ETIMEDOUT);
930 }
931
932 /**
933  * xs_set_port - reset the port number in the remote endpoint address
934  * @xprt: generic transport
935  * @port: new port number
936  *
937  */
938 static void xs_set_port(struct rpc_xprt *xprt, unsigned short port)
939 {
940         dprintk("RPC:      setting port for xprt %p to %u\n", xprt, port);
941         xprt->addr.sin_port = htons(port);
942 }
943
944 static int xs_bindresvport(struct rpc_xprt *xprt, struct socket *sock)
945 {
946         struct sockaddr_in myaddr = {
947                 .sin_family = AF_INET,
948         };
949         int err;
950         unsigned short port = xprt->port;
951
952         do {
953                 myaddr.sin_port = htons(port);
954                 err = sock->ops->bind(sock, (struct sockaddr *) &myaddr,
955                                                 sizeof(myaddr));
956                 if (err == 0) {
957                         xprt->port = port;
958                         dprintk("RPC:      xs_bindresvport bound to port %u\n",
959                                         port);
960                         return 0;
961                 }
962                 if (port <= xprt_min_resvport)
963                         port = xprt_max_resvport;
964                 else
965                         port--;
966         } while (err == -EADDRINUSE && port != xprt->port);
967
968         dprintk("RPC:      can't bind to reserved port (%d).\n", -err);
969         return err;
970 }
971
972 /**
973  * xs_udp_connect_worker - set up a UDP socket
974  * @args: RPC transport to connect
975  *
976  * Invoked by a work queue tasklet.
977  */
978 static void xs_udp_connect_worker(void *args)
979 {
980         struct rpc_xprt *xprt = (struct rpc_xprt *) args;
981         struct socket *sock = xprt->sock;
982         int err, status = -EIO;
983
984         if (xprt->shutdown || xprt->addr.sin_port == 0)
985                 goto out;
986
987         dprintk("RPC:      xs_udp_connect_worker for xprt %p\n", xprt);
988
989         /* Start by resetting any existing state */
990         xs_close(xprt);
991
992         if ((err = sock_create_kern(PF_INET, SOCK_DGRAM, IPPROTO_UDP, &sock)) < 0) {
993                 dprintk("RPC:      can't create UDP transport socket (%d).\n", -err);
994                 goto out;
995         }
996
997         if (xprt->resvport && xs_bindresvport(xprt, sock) < 0) {
998                 sock_release(sock);
999                 goto out;
1000         }
1001
1002         if (!xprt->inet) {
1003                 struct sock *sk = sock->sk;
1004
1005                 write_lock_bh(&sk->sk_callback_lock);
1006
1007                 sk->sk_user_data = xprt;
1008                 xprt->old_data_ready = sk->sk_data_ready;
1009                 xprt->old_state_change = sk->sk_state_change;
1010                 xprt->old_write_space = sk->sk_write_space;
1011                 sk->sk_data_ready = xs_udp_data_ready;
1012                 sk->sk_write_space = xs_udp_write_space;
1013                 sk->sk_no_check = UDP_CSUM_NORCV;
1014                 sk->sk_allocation = GFP_ATOMIC;
1015
1016                 xprt_set_connected(xprt);
1017
1018                 /* Reset to new socket */
1019                 xprt->sock = sock;
1020                 xprt->inet = sk;
1021
1022                 write_unlock_bh(&sk->sk_callback_lock);
1023         }
1024         xs_udp_do_set_buffer_size(xprt);
1025         status = 0;
1026 out:
1027         xprt_wake_pending_tasks(xprt, status);
1028         xprt_clear_connecting(xprt);
1029 }
1030
1031 /*
1032  * We need to preserve the port number so the reply cache on the server can
1033  * find our cached RPC replies when we get around to reconnecting.
1034  */
1035 static void xs_tcp_reuse_connection(struct rpc_xprt *xprt)
1036 {
1037         int result;
1038         struct socket *sock = xprt->sock;
1039         struct sockaddr any;
1040
1041         dprintk("RPC:      disconnecting xprt %p to reuse port\n", xprt);
1042
1043         /*
1044          * Disconnect the transport socket by doing a connect operation
1045          * with AF_UNSPEC.  This should return immediately...
1046          */
1047         memset(&any, 0, sizeof(any));
1048         any.sa_family = AF_UNSPEC;
1049         result = sock->ops->connect(sock, &any, sizeof(any), 0);
1050         if (result)
1051                 dprintk("RPC:      AF_UNSPEC connect return code %d\n",
1052                                 result);
1053 }
1054
1055 /**
1056  * xs_tcp_connect_worker - connect a TCP socket to a remote endpoint
1057  * @args: RPC transport to connect
1058  *
1059  * Invoked by a work queue tasklet.
1060  */
1061 static void xs_tcp_connect_worker(void *args)
1062 {
1063         struct rpc_xprt *xprt = (struct rpc_xprt *)args;
1064         struct socket *sock = xprt->sock;
1065         int err, status = -EIO;
1066
1067         if (xprt->shutdown || xprt->addr.sin_port == 0)
1068                 goto out;
1069
1070         dprintk("RPC:      xs_tcp_connect_worker for xprt %p\n", xprt);
1071
1072         if (!xprt->sock) {
1073                 /* start from scratch */
1074                 if ((err = sock_create_kern(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock)) < 0) {
1075                         dprintk("RPC:      can't create TCP transport socket (%d).\n", -err);
1076                         goto out;
1077                 }
1078
1079                 if (xprt->resvport && xs_bindresvport(xprt, sock) < 0) {
1080                         sock_release(sock);
1081                         goto out;
1082                 }
1083         } else
1084                 /* "close" the socket, preserving the local port */
1085                 xs_tcp_reuse_connection(xprt);
1086
1087         if (!xprt->inet) {
1088                 struct sock *sk = sock->sk;
1089
1090                 write_lock_bh(&sk->sk_callback_lock);
1091
1092                 sk->sk_user_data = xprt;
1093                 xprt->old_data_ready = sk->sk_data_ready;
1094                 xprt->old_state_change = sk->sk_state_change;
1095                 xprt->old_write_space = sk->sk_write_space;
1096                 sk->sk_data_ready = xs_tcp_data_ready;
1097                 sk->sk_state_change = xs_tcp_state_change;
1098                 sk->sk_write_space = xs_tcp_write_space;
1099                 sk->sk_allocation = GFP_ATOMIC;
1100
1101                 /* socket options */
1102                 sk->sk_userlocks |= SOCK_BINDPORT_LOCK;
1103                 sock_reset_flag(sk, SOCK_LINGER);
1104                 tcp_sk(sk)->linger2 = 0;
1105                 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
1106
1107                 xprt_clear_connected(xprt);
1108
1109                 /* Reset to new socket */
1110                 xprt->sock = sock;
1111                 xprt->inet = sk;
1112
1113                 write_unlock_bh(&sk->sk_callback_lock);
1114         }
1115
1116         /* Tell the socket layer to start connecting... */
1117         status = sock->ops->connect(sock, (struct sockaddr *) &xprt->addr,
1118                         sizeof(xprt->addr), O_NONBLOCK);
1119         dprintk("RPC: %p  connect status %d connected %d sock state %d\n",
1120                         xprt, -status, xprt_connected(xprt), sock->sk->sk_state);
1121         if (status < 0) {
1122                 switch (status) {
1123                         case -EINPROGRESS:
1124                         case -EALREADY:
1125                                 goto out_clear;
1126                         case -ECONNREFUSED:
1127                         case -ECONNRESET:
1128                                 /* retry with existing socket, after a delay */
1129                                 break;
1130                         default:
1131                                 /* get rid of existing socket, and retry */
1132                                 xs_close(xprt);
1133                                 break;
1134                 }
1135         }
1136 out:
1137         xprt_wake_pending_tasks(xprt, status);
1138 out_clear:
1139         xprt_clear_connecting(xprt);
1140 }
1141
1142 /**
1143  * xs_connect - connect a socket to a remote endpoint
1144  * @task: address of RPC task that manages state of connect request
1145  *
1146  * TCP: If the remote end dropped the connection, delay reconnecting.
1147  *
1148  * UDP socket connects are synchronous, but we use a work queue anyway
1149  * to guarantee that even unprivileged user processes can set up a
1150  * socket on a privileged port.
1151  *
1152  * If a UDP socket connect fails, the delay behavior here prevents
1153  * retry floods (hard mounts).
1154  */
1155 static void xs_connect(struct rpc_task *task)
1156 {
1157         struct rpc_xprt *xprt = task->tk_xprt;
1158
1159         if (xprt_test_and_set_connecting(xprt))
1160                 return;
1161
1162         if (xprt->sock != NULL) {
1163                 dprintk("RPC:      xs_connect delayed xprt %p for %lu seconds\n",
1164                                 xprt, xprt->reestablish_timeout / HZ);
1165                 schedule_delayed_work(&xprt->connect_worker,
1166                                         xprt->reestablish_timeout);
1167                 xprt->reestablish_timeout <<= 1;
1168                 if (xprt->reestablish_timeout > XS_TCP_MAX_REEST_TO)
1169                         xprt->reestablish_timeout = XS_TCP_MAX_REEST_TO;
1170         } else {
1171                 dprintk("RPC:      xs_connect scheduled xprt %p\n", xprt);
1172                 schedule_work(&xprt->connect_worker);
1173
1174                 /* flush_scheduled_work can sleep... */
1175                 if (!RPC_IS_ASYNC(task))
1176                         flush_scheduled_work();
1177         }
1178 }
1179
1180 static struct rpc_xprt_ops xs_udp_ops = {
1181         .set_buffer_size        = xs_udp_set_buffer_size,
1182         .reserve_xprt           = xprt_reserve_xprt_cong,
1183         .release_xprt           = xprt_release_xprt_cong,
1184         .set_port               = xs_set_port,
1185         .connect                = xs_connect,
1186         .buf_alloc              = rpc_malloc,
1187         .buf_free               = rpc_free,
1188         .send_request           = xs_udp_send_request,
1189         .set_retrans_timeout    = xprt_set_retrans_timeout_rtt,
1190         .timer                  = xs_udp_timer,
1191         .release_request        = xprt_release_rqst_cong,
1192         .close                  = xs_close,
1193         .destroy                = xs_destroy,
1194 };
1195
1196 static struct rpc_xprt_ops xs_tcp_ops = {
1197         .reserve_xprt           = xprt_reserve_xprt,
1198         .release_xprt           = xprt_release_xprt,
1199         .set_port               = xs_set_port,
1200         .connect                = xs_connect,
1201         .buf_alloc              = rpc_malloc,
1202         .buf_free               = rpc_free,
1203         .send_request           = xs_tcp_send_request,
1204         .set_retrans_timeout    = xprt_set_retrans_timeout_def,
1205         .close                  = xs_close,
1206         .destroy                = xs_destroy,
1207 };
1208
1209 /**
1210  * xs_setup_udp - Set up transport to use a UDP socket
1211  * @xprt: transport to set up
1212  * @to:   timeout parameters
1213  *
1214  */
1215 int xs_setup_udp(struct rpc_xprt *xprt, struct rpc_timeout *to)
1216 {
1217         size_t slot_table_size;
1218
1219         dprintk("RPC:      setting up udp-ipv4 transport...\n");
1220
1221         xprt->max_reqs = xprt_udp_slot_table_entries;
1222         slot_table_size = xprt->max_reqs * sizeof(xprt->slot[0]);
1223         xprt->slot = kmalloc(slot_table_size, GFP_KERNEL);
1224         if (xprt->slot == NULL)
1225                 return -ENOMEM;
1226         memset(xprt->slot, 0, slot_table_size);
1227
1228         xprt->prot = IPPROTO_UDP;
1229         xprt->port = xprt_max_resvport;
1230         xprt->tsh_size = 0;
1231         xprt->resvport = capable(CAP_NET_BIND_SERVICE) ? 1 : 0;
1232         /* XXX: header size can vary due to auth type, IPv6, etc. */
1233         xprt->max_payload = (1U << 16) - (MAX_HEADER << 3);
1234
1235         INIT_WORK(&xprt->connect_worker, xs_udp_connect_worker, xprt);
1236         xprt->bind_timeout = XS_BIND_TO;
1237         xprt->connect_timeout = XS_UDP_CONN_TO;
1238         xprt->reestablish_timeout = XS_UDP_REEST_TO;
1239         xprt->idle_timeout = XS_IDLE_DISC_TO;
1240
1241         xprt->ops = &xs_udp_ops;
1242
1243         if (to)
1244                 xprt->timeout = *to;
1245         else
1246                 xprt_set_timeout(&xprt->timeout, 5, 5 * HZ);
1247
1248         return 0;
1249 }
1250
1251 /**
1252  * xs_setup_tcp - Set up transport to use a TCP socket
1253  * @xprt: transport to set up
1254  * @to: timeout parameters
1255  *
1256  */
1257 int xs_setup_tcp(struct rpc_xprt *xprt, struct rpc_timeout *to)
1258 {
1259         size_t slot_table_size;
1260
1261         dprintk("RPC:      setting up tcp-ipv4 transport...\n");
1262
1263         xprt->max_reqs = xprt_tcp_slot_table_entries;
1264         slot_table_size = xprt->max_reqs * sizeof(xprt->slot[0]);
1265         xprt->slot = kmalloc(slot_table_size, GFP_KERNEL);
1266         if (xprt->slot == NULL)
1267                 return -ENOMEM;
1268         memset(xprt->slot, 0, slot_table_size);
1269
1270         xprt->prot = IPPROTO_TCP;
1271         xprt->port = xprt_max_resvport;
1272         xprt->tsh_size = sizeof(rpc_fraghdr) / sizeof(u32);
1273         xprt->resvport = capable(CAP_NET_BIND_SERVICE) ? 1 : 0;
1274         xprt->max_payload = RPC_MAX_FRAGMENT_SIZE;
1275
1276         INIT_WORK(&xprt->connect_worker, xs_tcp_connect_worker, xprt);
1277         xprt->bind_timeout = XS_BIND_TO;
1278         xprt->connect_timeout = XS_TCP_CONN_TO;
1279         xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
1280         xprt->idle_timeout = XS_IDLE_DISC_TO;
1281
1282         xprt->ops = &xs_tcp_ops;
1283
1284         if (to)
1285                 xprt->timeout = *to;
1286         else
1287                 xprt_set_timeout(&xprt->timeout, 2, 60 * HZ);
1288
1289         return 0;
1290 }