Loading...
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
7 * TCP send fixes (C) 1998 Red Hat
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 * IPv6 support contributed by Gilles Quillard, Bull Open Source, 2005.
18 * <gilles.quillard@bull.net>
19 */
20
21#include <linux/types.h>
22#include <linux/string.h>
23#include <linux/slab.h>
24#include <linux/module.h>
25#include <linux/capability.h>
26#include <linux/pagemap.h>
27#include <linux/errno.h>
28#include <linux/socket.h>
29#include <linux/in.h>
30#include <linux/net.h>
31#include <linux/mm.h>
32#include <linux/un.h>
33#include <linux/udp.h>
34#include <linux/tcp.h>
35#include <linux/sunrpc/clnt.h>
36#include <linux/sunrpc/addr.h>
37#include <linux/sunrpc/sched.h>
38#include <linux/sunrpc/svcsock.h>
39#include <linux/sunrpc/xprtsock.h>
40#include <linux/file.h>
41#ifdef CONFIG_SUNRPC_BACKCHANNEL
42#include <linux/sunrpc/bc_xprt.h>
43#endif
44
45#include <net/sock.h>
46#include <net/checksum.h>
47#include <net/udp.h>
48#include <net/tcp.h>
49
50#include <trace/events/sunrpc.h>
51
52#include "sunrpc.h"
53
54static void xs_close(struct rpc_xprt *xprt);
55
56/*
57 * xprtsock tunables
58 */
59static unsigned int xprt_udp_slot_table_entries = RPC_DEF_SLOT_TABLE;
60static unsigned int xprt_tcp_slot_table_entries = RPC_MIN_SLOT_TABLE;
61static unsigned int xprt_max_tcp_slot_table_entries = RPC_MAX_SLOT_TABLE;
62
63static unsigned int xprt_min_resvport = RPC_DEF_MIN_RESVPORT;
64static unsigned int xprt_max_resvport = RPC_DEF_MAX_RESVPORT;
65
66#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
67
68#define XS_TCP_LINGER_TO (15U * HZ)
69static unsigned int xs_tcp_fin_timeout __read_mostly = XS_TCP_LINGER_TO;
70
71/*
72 * We can register our own files under /proc/sys/sunrpc by
73 * calling register_sysctl_table() again. The files in that
74 * directory become the union of all files registered there.
75 *
76 * We simply need to make sure that we don't collide with
77 * someone else's file names!
78 */
79
80static unsigned int min_slot_table_size = RPC_MIN_SLOT_TABLE;
81static unsigned int max_slot_table_size = RPC_MAX_SLOT_TABLE;
82static unsigned int max_tcp_slot_table_limit = RPC_MAX_SLOT_TABLE_LIMIT;
83static unsigned int xprt_min_resvport_limit = RPC_MIN_RESVPORT;
84static unsigned int xprt_max_resvport_limit = RPC_MAX_RESVPORT;
85
86static struct ctl_table_header *sunrpc_table_header;
87
88/*
89 * FIXME: changing the UDP slot table size should also resize the UDP
90 * socket buffers for existing UDP transports
91 */
92static struct ctl_table xs_tunables_table[] = {
93 {
94 .procname = "udp_slot_table_entries",
95 .data = &xprt_udp_slot_table_entries,
96 .maxlen = sizeof(unsigned int),
97 .mode = 0644,
98 .proc_handler = proc_dointvec_minmax,
99 .extra1 = &min_slot_table_size,
100 .extra2 = &max_slot_table_size
101 },
102 {
103 .procname = "tcp_slot_table_entries",
104 .data = &xprt_tcp_slot_table_entries,
105 .maxlen = sizeof(unsigned int),
106 .mode = 0644,
107 .proc_handler = proc_dointvec_minmax,
108 .extra1 = &min_slot_table_size,
109 .extra2 = &max_slot_table_size
110 },
111 {
112 .procname = "tcp_max_slot_table_entries",
113 .data = &xprt_max_tcp_slot_table_entries,
114 .maxlen = sizeof(unsigned int),
115 .mode = 0644,
116 .proc_handler = proc_dointvec_minmax,
117 .extra1 = &min_slot_table_size,
118 .extra2 = &max_tcp_slot_table_limit
119 },
120 {
121 .procname = "min_resvport",
122 .data = &xprt_min_resvport,
123 .maxlen = sizeof(unsigned int),
124 .mode = 0644,
125 .proc_handler = proc_dointvec_minmax,
126 .extra1 = &xprt_min_resvport_limit,
127 .extra2 = &xprt_max_resvport_limit
128 },
129 {
130 .procname = "max_resvport",
131 .data = &xprt_max_resvport,
132 .maxlen = sizeof(unsigned int),
133 .mode = 0644,
134 .proc_handler = proc_dointvec_minmax,
135 .extra1 = &xprt_min_resvport_limit,
136 .extra2 = &xprt_max_resvport_limit
137 },
138 {
139 .procname = "tcp_fin_timeout",
140 .data = &xs_tcp_fin_timeout,
141 .maxlen = sizeof(xs_tcp_fin_timeout),
142 .mode = 0644,
143 .proc_handler = proc_dointvec_jiffies,
144 },
145 { },
146};
147
148static struct ctl_table sunrpc_table[] = {
149 {
150 .procname = "sunrpc",
151 .mode = 0555,
152 .child = xs_tunables_table
153 },
154 { },
155};
156
157#endif
158
159/*
160 * Wait duration for a reply from the RPC portmapper.
161 */
162#define XS_BIND_TO (60U * HZ)
163
164/*
165 * Delay if a UDP socket connect error occurs. This is most likely some
166 * kind of resource problem on the local host.
167 */
168#define XS_UDP_REEST_TO (2U * HZ)
169
170/*
171 * The reestablish timeout allows clients to delay for a bit before attempting
172 * to reconnect to a server that just dropped our connection.
173 *
174 * We implement an exponential backoff when trying to reestablish a TCP
175 * transport connection with the server. Some servers like to drop a TCP
176 * connection when they are overworked, so we start with a short timeout and
177 * increase over time if the server is down or not responding.
178 */
179#define XS_TCP_INIT_REEST_TO (3U * HZ)
180#define XS_TCP_MAX_REEST_TO (5U * 60 * HZ)
181
182/*
183 * TCP idle timeout; client drops the transport socket if it is idle
184 * for this long. Note that we also timeout UDP sockets to prevent
185 * holding port numbers when there is no RPC traffic.
186 */
187#define XS_IDLE_DISC_TO (5U * 60 * HZ)
188
189#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
190# undef RPC_DEBUG_DATA
191# define RPCDBG_FACILITY RPCDBG_TRANS
192#endif
193
194#ifdef RPC_DEBUG_DATA
195static void xs_pktdump(char *msg, u32 *packet, unsigned int count)
196{
197 u8 *buf = (u8 *) packet;
198 int j;
199
200 dprintk("RPC: %s\n", msg);
201 for (j = 0; j < count && j < 128; j += 4) {
202 if (!(j & 31)) {
203 if (j)
204 dprintk("\n");
205 dprintk("0x%04x ", j);
206 }
207 dprintk("%02x%02x%02x%02x ",
208 buf[j], buf[j+1], buf[j+2], buf[j+3]);
209 }
210 dprintk("\n");
211}
212#else
213static inline void xs_pktdump(char *msg, u32 *packet, unsigned int count)
214{
215 /* NOP */
216}
217#endif
218
219static inline struct rpc_xprt *xprt_from_sock(struct sock *sk)
220{
221 return (struct rpc_xprt *) sk->sk_user_data;
222}
223
224static inline struct sockaddr *xs_addr(struct rpc_xprt *xprt)
225{
226 return (struct sockaddr *) &xprt->addr;
227}
228
229static inline struct sockaddr_un *xs_addr_un(struct rpc_xprt *xprt)
230{
231 return (struct sockaddr_un *) &xprt->addr;
232}
233
234static inline struct sockaddr_in *xs_addr_in(struct rpc_xprt *xprt)
235{
236 return (struct sockaddr_in *) &xprt->addr;
237}
238
239static inline struct sockaddr_in6 *xs_addr_in6(struct rpc_xprt *xprt)
240{
241 return (struct sockaddr_in6 *) &xprt->addr;
242}
243
244static void xs_format_common_peer_addresses(struct rpc_xprt *xprt)
245{
246 struct sockaddr *sap = xs_addr(xprt);
247 struct sockaddr_in6 *sin6;
248 struct sockaddr_in *sin;
249 struct sockaddr_un *sun;
250 char buf[128];
251
252 switch (sap->sa_family) {
253 case AF_LOCAL:
254 sun = xs_addr_un(xprt);
255 strlcpy(buf, sun->sun_path, sizeof(buf));
256 xprt->address_strings[RPC_DISPLAY_ADDR] =
257 kstrdup(buf, GFP_KERNEL);
258 break;
259 case AF_INET:
260 (void)rpc_ntop(sap, buf, sizeof(buf));
261 xprt->address_strings[RPC_DISPLAY_ADDR] =
262 kstrdup(buf, GFP_KERNEL);
263 sin = xs_addr_in(xprt);
264 snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
265 break;
266 case AF_INET6:
267 (void)rpc_ntop(sap, buf, sizeof(buf));
268 xprt->address_strings[RPC_DISPLAY_ADDR] =
269 kstrdup(buf, GFP_KERNEL);
270 sin6 = xs_addr_in6(xprt);
271 snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
272 break;
273 default:
274 BUG();
275 }
276
277 xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
278}
279
280static void xs_format_common_peer_ports(struct rpc_xprt *xprt)
281{
282 struct sockaddr *sap = xs_addr(xprt);
283 char buf[128];
284
285 snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
286 xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
287
288 snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
289 xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
290}
291
292static void xs_format_peer_addresses(struct rpc_xprt *xprt,
293 const char *protocol,
294 const char *netid)
295{
296 xprt->address_strings[RPC_DISPLAY_PROTO] = protocol;
297 xprt->address_strings[RPC_DISPLAY_NETID] = netid;
298 xs_format_common_peer_addresses(xprt);
299 xs_format_common_peer_ports(xprt);
300}
301
302static void xs_update_peer_port(struct rpc_xprt *xprt)
303{
304 kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]);
305 kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
306
307 xs_format_common_peer_ports(xprt);
308}
309
310static void xs_free_peer_addresses(struct rpc_xprt *xprt)
311{
312 unsigned int i;
313
314 for (i = 0; i < RPC_DISPLAY_MAX; i++)
315 switch (i) {
316 case RPC_DISPLAY_PROTO:
317 case RPC_DISPLAY_NETID:
318 continue;
319 default:
320 kfree(xprt->address_strings[i]);
321 }
322}
323
324#define XS_SENDMSG_FLAGS (MSG_DONTWAIT | MSG_NOSIGNAL)
325
326static int xs_send_kvec(struct socket *sock, struct sockaddr *addr, int addrlen, struct kvec *vec, unsigned int base, int more)
327{
328 struct msghdr msg = {
329 .msg_name = addr,
330 .msg_namelen = addrlen,
331 .msg_flags = XS_SENDMSG_FLAGS | (more ? MSG_MORE : 0),
332 };
333 struct kvec iov = {
334 .iov_base = vec->iov_base + base,
335 .iov_len = vec->iov_len - base,
336 };
337
338 if (iov.iov_len != 0)
339 return kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
340 return kernel_sendmsg(sock, &msg, NULL, 0, 0);
341}
342
343static int xs_send_pagedata(struct socket *sock, struct xdr_buf *xdr, unsigned int base, int more, bool zerocopy, int *sent_p)
344{
345 ssize_t (*do_sendpage)(struct socket *sock, struct page *page,
346 int offset, size_t size, int flags);
347 struct page **ppage;
348 unsigned int remainder;
349 int err;
350
351 remainder = xdr->page_len - base;
352 base += xdr->page_base;
353 ppage = xdr->pages + (base >> PAGE_SHIFT);
354 base &= ~PAGE_MASK;
355 do_sendpage = sock->ops->sendpage;
356 if (!zerocopy)
357 do_sendpage = sock_no_sendpage;
358 for(;;) {
359 unsigned int len = min_t(unsigned int, PAGE_SIZE - base, remainder);
360 int flags = XS_SENDMSG_FLAGS;
361
362 remainder -= len;
363 if (more)
364 flags |= MSG_MORE;
365 if (remainder != 0)
366 flags |= MSG_SENDPAGE_NOTLAST | MSG_MORE;
367 err = do_sendpage(sock, *ppage, base, len, flags);
368 if (remainder == 0 || err != len)
369 break;
370 *sent_p += err;
371 ppage++;
372 base = 0;
373 }
374 if (err > 0) {
375 *sent_p += err;
376 err = 0;
377 }
378 return err;
379}
380
381/**
382 * xs_sendpages - write pages directly to a socket
383 * @sock: socket to send on
384 * @addr: UDP only -- address of destination
385 * @addrlen: UDP only -- length of destination address
386 * @xdr: buffer containing this request
387 * @base: starting position in the buffer
388 * @zerocopy: true if it is safe to use sendpage()
389 * @sent_p: return the total number of bytes successfully queued for sending
390 *
391 */
392static int xs_sendpages(struct socket *sock, struct sockaddr *addr, int addrlen, struct xdr_buf *xdr, unsigned int base, bool zerocopy, int *sent_p)
393{
394 unsigned int remainder = xdr->len - base;
395 int err = 0;
396 int sent = 0;
397
398 if (unlikely(!sock))
399 return -ENOTSOCK;
400
401 if (base != 0) {
402 addr = NULL;
403 addrlen = 0;
404 }
405
406 if (base < xdr->head[0].iov_len || addr != NULL) {
407 unsigned int len = xdr->head[0].iov_len - base;
408 remainder -= len;
409 err = xs_send_kvec(sock, addr, addrlen, &xdr->head[0], base, remainder != 0);
410 if (remainder == 0 || err != len)
411 goto out;
412 *sent_p += err;
413 base = 0;
414 } else
415 base -= xdr->head[0].iov_len;
416
417 if (base < xdr->page_len) {
418 unsigned int len = xdr->page_len - base;
419 remainder -= len;
420 err = xs_send_pagedata(sock, xdr, base, remainder != 0, zerocopy, &sent);
421 *sent_p += sent;
422 if (remainder == 0 || sent != len)
423 goto out;
424 base = 0;
425 } else
426 base -= xdr->page_len;
427
428 if (base >= xdr->tail[0].iov_len)
429 return 0;
430 err = xs_send_kvec(sock, NULL, 0, &xdr->tail[0], base, 0);
431out:
432 if (err > 0) {
433 *sent_p += err;
434 err = 0;
435 }
436 return err;
437}
438
439static void xs_nospace_callback(struct rpc_task *task)
440{
441 struct sock_xprt *transport = container_of(task->tk_rqstp->rq_xprt, struct sock_xprt, xprt);
442
443 transport->inet->sk_write_pending--;
444}
445
446/**
447 * xs_nospace - place task on wait queue if transmit was incomplete
448 * @task: task to put to sleep
449 *
450 */
451static int xs_nospace(struct rpc_task *task)
452{
453 struct rpc_rqst *req = task->tk_rqstp;
454 struct rpc_xprt *xprt = req->rq_xprt;
455 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
456 struct sock *sk = transport->inet;
457 int ret = -EAGAIN;
458
459 dprintk("RPC: %5u xmit incomplete (%u left of %u)\n",
460 task->tk_pid, req->rq_slen - req->rq_bytes_sent,
461 req->rq_slen);
462
463 /* Protect against races with write_space */
464 spin_lock_bh(&xprt->transport_lock);
465
466 /* Don't race with disconnect */
467 if (xprt_connected(xprt)) {
468 /* wait for more buffer space */
469 sk->sk_write_pending++;
470 xprt_wait_for_buffer_space(task, xs_nospace_callback);
471 } else
472 ret = -ENOTCONN;
473
474 spin_unlock_bh(&xprt->transport_lock);
475
476 /* Race breaker in case memory is freed before above code is called */
477 sk->sk_write_space(sk);
478 return ret;
479}
480
481/*
482 * Construct a stream transport record marker in @buf.
483 */
484static inline void xs_encode_stream_record_marker(struct xdr_buf *buf)
485{
486 u32 reclen = buf->len - sizeof(rpc_fraghdr);
487 rpc_fraghdr *base = buf->head[0].iov_base;
488 *base = cpu_to_be32(RPC_LAST_STREAM_FRAGMENT | reclen);
489}
490
491/**
492 * xs_local_send_request - write an RPC request to an AF_LOCAL socket
493 * @task: RPC task that manages the state of an RPC request
494 *
495 * Return values:
496 * 0: The request has been sent
497 * EAGAIN: The socket was blocked, please call again later to
498 * complete the request
499 * ENOTCONN: Caller needs to invoke connect logic then call again
500 * other: Some other error occured, the request was not sent
501 */
502static int xs_local_send_request(struct rpc_task *task)
503{
504 struct rpc_rqst *req = task->tk_rqstp;
505 struct rpc_xprt *xprt = req->rq_xprt;
506 struct sock_xprt *transport =
507 container_of(xprt, struct sock_xprt, xprt);
508 struct xdr_buf *xdr = &req->rq_snd_buf;
509 int status;
510 int sent = 0;
511
512 xs_encode_stream_record_marker(&req->rq_snd_buf);
513
514 xs_pktdump("packet data:",
515 req->rq_svec->iov_base, req->rq_svec->iov_len);
516
517 status = xs_sendpages(transport->sock, NULL, 0, xdr, req->rq_bytes_sent,
518 true, &sent);
519 dprintk("RPC: %s(%u) = %d\n",
520 __func__, xdr->len - req->rq_bytes_sent, status);
521
522 if (status == -EAGAIN && sock_writeable(transport->inet))
523 status = -ENOBUFS;
524
525 if (likely(sent > 0) || status == 0) {
526 req->rq_bytes_sent += sent;
527 req->rq_xmit_bytes_sent += sent;
528 if (likely(req->rq_bytes_sent >= req->rq_slen)) {
529 req->rq_bytes_sent = 0;
530 return 0;
531 }
532 status = -EAGAIN;
533 }
534
535 switch (status) {
536 case -ENOBUFS:
537 break;
538 case -EAGAIN:
539 status = xs_nospace(task);
540 break;
541 default:
542 dprintk("RPC: sendmsg returned unrecognized error %d\n",
543 -status);
544 case -EPIPE:
545 xs_close(xprt);
546 status = -ENOTCONN;
547 }
548
549 return status;
550}
551
552/**
553 * xs_udp_send_request - write an RPC request to a UDP socket
554 * @task: address of RPC task that manages the state of an RPC request
555 *
556 * Return values:
557 * 0: The request has been sent
558 * EAGAIN: The socket was blocked, please call again later to
559 * complete the request
560 * ENOTCONN: Caller needs to invoke connect logic then call again
561 * other: Some other error occurred, the request was not sent
562 */
563static int xs_udp_send_request(struct rpc_task *task)
564{
565 struct rpc_rqst *req = task->tk_rqstp;
566 struct rpc_xprt *xprt = req->rq_xprt;
567 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
568 struct xdr_buf *xdr = &req->rq_snd_buf;
569 int sent = 0;
570 int status;
571
572 xs_pktdump("packet data:",
573 req->rq_svec->iov_base,
574 req->rq_svec->iov_len);
575
576 if (!xprt_bound(xprt))
577 return -ENOTCONN;
578 status = xs_sendpages(transport->sock, xs_addr(xprt), xprt->addrlen,
579 xdr, req->rq_bytes_sent, true, &sent);
580
581 dprintk("RPC: xs_udp_send_request(%u) = %d\n",
582 xdr->len - req->rq_bytes_sent, status);
583
584 /* firewall is blocking us, don't return -EAGAIN or we end up looping */
585 if (status == -EPERM)
586 goto process_status;
587
588 if (status == -EAGAIN && sock_writeable(transport->inet))
589 status = -ENOBUFS;
590
591 if (sent > 0 || status == 0) {
592 req->rq_xmit_bytes_sent += sent;
593 if (sent >= req->rq_slen)
594 return 0;
595 /* Still some bytes left; set up for a retry later. */
596 status = -EAGAIN;
597 }
598
599process_status:
600 switch (status) {
601 case -ENOTSOCK:
602 status = -ENOTCONN;
603 /* Should we call xs_close() here? */
604 break;
605 case -EAGAIN:
606 status = xs_nospace(task);
607 break;
608 case -ENETUNREACH:
609 case -ENOBUFS:
610 case -EPIPE:
611 case -ECONNREFUSED:
612 case -EPERM:
613 /* When the server has died, an ICMP port unreachable message
614 * prompts ECONNREFUSED. */
615 break;
616 default:
617 dprintk("RPC: sendmsg returned unrecognized error %d\n",
618 -status);
619 }
620
621 return status;
622}
623
624/**
625 * xs_tcp_send_request - write an RPC request to a TCP socket
626 * @task: address of RPC task that manages the state of an RPC request
627 *
628 * Return values:
629 * 0: The request has been sent
630 * EAGAIN: The socket was blocked, please call again later to
631 * complete the request
632 * ENOTCONN: Caller needs to invoke connect logic then call again
633 * other: Some other error occurred, the request was not sent
634 *
635 * XXX: In the case of soft timeouts, should we eventually give up
636 * if sendmsg is not able to make progress?
637 */
638static int xs_tcp_send_request(struct rpc_task *task)
639{
640 struct rpc_rqst *req = task->tk_rqstp;
641 struct rpc_xprt *xprt = req->rq_xprt;
642 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
643 struct xdr_buf *xdr = &req->rq_snd_buf;
644 bool zerocopy = true;
645 int status;
646 int sent;
647
648 xs_encode_stream_record_marker(&req->rq_snd_buf);
649
650 xs_pktdump("packet data:",
651 req->rq_svec->iov_base,
652 req->rq_svec->iov_len);
653 /* Don't use zero copy if this is a resend. If the RPC call
654 * completes while the socket holds a reference to the pages,
655 * then we may end up resending corrupted data.
656 */
657 if (task->tk_flags & RPC_TASK_SENT)
658 zerocopy = false;
659
660 /* Continue transmitting the packet/record. We must be careful
661 * to cope with writespace callbacks arriving _after_ we have
662 * called sendmsg(). */
663 while (1) {
664 sent = 0;
665 status = xs_sendpages(transport->sock, NULL, 0, xdr,
666 req->rq_bytes_sent, zerocopy, &sent);
667
668 dprintk("RPC: xs_tcp_send_request(%u) = %d\n",
669 xdr->len - req->rq_bytes_sent, status);
670
671 /* If we've sent the entire packet, immediately
672 * reset the count of bytes sent. */
673 req->rq_bytes_sent += sent;
674 req->rq_xmit_bytes_sent += sent;
675 if (likely(req->rq_bytes_sent >= req->rq_slen)) {
676 req->rq_bytes_sent = 0;
677 return 0;
678 }
679
680 if (status < 0)
681 break;
682 if (sent == 0) {
683 status = -EAGAIN;
684 break;
685 }
686 }
687 if (status == -EAGAIN && sk_stream_is_writeable(transport->inet))
688 status = -ENOBUFS;
689
690 switch (status) {
691 case -ENOTSOCK:
692 status = -ENOTCONN;
693 /* Should we call xs_close() here? */
694 break;
695 case -EAGAIN:
696 status = xs_nospace(task);
697 break;
698 case -ECONNRESET:
699 case -ECONNREFUSED:
700 case -ENOTCONN:
701 case -EADDRINUSE:
702 case -ENOBUFS:
703 case -EPIPE:
704 break;
705 default:
706 dprintk("RPC: sendmsg returned unrecognized error %d\n",
707 -status);
708 }
709
710 return status;
711}
712
713/**
714 * xs_tcp_release_xprt - clean up after a tcp transmission
715 * @xprt: transport
716 * @task: rpc task
717 *
718 * This cleans up if an error causes us to abort the transmission of a request.
719 * In this case, the socket may need to be reset in order to avoid confusing
720 * the server.
721 */
722static void xs_tcp_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
723{
724 struct rpc_rqst *req;
725
726 if (task != xprt->snd_task)
727 return;
728 if (task == NULL)
729 goto out_release;
730 req = task->tk_rqstp;
731 if (req == NULL)
732 goto out_release;
733 if (req->rq_bytes_sent == 0)
734 goto out_release;
735 if (req->rq_bytes_sent == req->rq_snd_buf.len)
736 goto out_release;
737 set_bit(XPRT_CLOSE_WAIT, &xprt->state);
738out_release:
739 xprt_release_xprt(xprt, task);
740}
741
742static void xs_save_old_callbacks(struct sock_xprt *transport, struct sock *sk)
743{
744 transport->old_data_ready = sk->sk_data_ready;
745 transport->old_state_change = sk->sk_state_change;
746 transport->old_write_space = sk->sk_write_space;
747 transport->old_error_report = sk->sk_error_report;
748}
749
750static void xs_restore_old_callbacks(struct sock_xprt *transport, struct sock *sk)
751{
752 sk->sk_data_ready = transport->old_data_ready;
753 sk->sk_state_change = transport->old_state_change;
754 sk->sk_write_space = transport->old_write_space;
755 sk->sk_error_report = transport->old_error_report;
756}
757
758static void xs_sock_reset_connection_flags(struct rpc_xprt *xprt)
759{
760 smp_mb__before_atomic();
761 clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
762 clear_bit(XPRT_CLOSING, &xprt->state);
763 smp_mb__after_atomic();
764}
765
766static void xs_sock_mark_closed(struct rpc_xprt *xprt)
767{
768 xs_sock_reset_connection_flags(xprt);
769 /* Mark transport as closed and wake up all pending tasks */
770 xprt_disconnect_done(xprt);
771}
772
773/**
774 * xs_error_report - callback to handle TCP socket state errors
775 * @sk: socket
776 *
777 * Note: we don't call sock_error() since there may be a rpc_task
778 * using the socket, and so we don't want to clear sk->sk_err.
779 */
780static void xs_error_report(struct sock *sk)
781{
782 struct rpc_xprt *xprt;
783 int err;
784
785 read_lock_bh(&sk->sk_callback_lock);
786 if (!(xprt = xprt_from_sock(sk)))
787 goto out;
788
789 err = -sk->sk_err;
790 if (err == 0)
791 goto out;
792 /* Is this a reset event? */
793 if (sk->sk_state == TCP_CLOSE)
794 xs_sock_mark_closed(xprt);
795 dprintk("RPC: xs_error_report client %p, error=%d...\n",
796 xprt, -err);
797 trace_rpc_socket_error(xprt, sk->sk_socket, err);
798 xprt_wake_pending_tasks(xprt, err);
799 out:
800 read_unlock_bh(&sk->sk_callback_lock);
801}
802
803static void xs_reset_transport(struct sock_xprt *transport)
804{
805 struct socket *sock = transport->sock;
806 struct sock *sk = transport->inet;
807 struct rpc_xprt *xprt = &transport->xprt;
808
809 if (sk == NULL)
810 return;
811
812 if (atomic_read(&transport->xprt.swapper))
813 sk_clear_memalloc(sk);
814
815 kernel_sock_shutdown(sock, SHUT_RDWR);
816
817 mutex_lock(&transport->recv_mutex);
818 write_lock_bh(&sk->sk_callback_lock);
819 transport->inet = NULL;
820 transport->sock = NULL;
821
822 sk->sk_user_data = NULL;
823
824 xs_restore_old_callbacks(transport, sk);
825 xprt_clear_connected(xprt);
826 write_unlock_bh(&sk->sk_callback_lock);
827 xs_sock_reset_connection_flags(xprt);
828 mutex_unlock(&transport->recv_mutex);
829
830 trace_rpc_socket_close(xprt, sock);
831 sock_release(sock);
832}
833
834/**
835 * xs_close - close a socket
836 * @xprt: transport
837 *
838 * This is used when all requests are complete; ie, no DRC state remains
839 * on the server we want to save.
840 *
841 * The caller _must_ be holding XPRT_LOCKED in order to avoid issues with
842 * xs_reset_transport() zeroing the socket from underneath a writer.
843 */
844static void xs_close(struct rpc_xprt *xprt)
845{
846 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
847
848 dprintk("RPC: xs_close xprt %p\n", xprt);
849
850 xs_reset_transport(transport);
851 xprt->reestablish_timeout = 0;
852
853 xprt_disconnect_done(xprt);
854}
855
856static void xs_inject_disconnect(struct rpc_xprt *xprt)
857{
858 dprintk("RPC: injecting transport disconnect on xprt=%p\n",
859 xprt);
860 xprt_disconnect_done(xprt);
861}
862
863static void xs_xprt_free(struct rpc_xprt *xprt)
864{
865 xs_free_peer_addresses(xprt);
866 xprt_free(xprt);
867}
868
869/**
870 * xs_destroy - prepare to shutdown a transport
871 * @xprt: doomed transport
872 *
873 */
874static void xs_destroy(struct rpc_xprt *xprt)
875{
876 struct sock_xprt *transport = container_of(xprt,
877 struct sock_xprt, xprt);
878 dprintk("RPC: xs_destroy xprt %p\n", xprt);
879
880 cancel_delayed_work_sync(&transport->connect_worker);
881 xs_close(xprt);
882 cancel_work_sync(&transport->recv_worker);
883 xs_xprt_free(xprt);
884 module_put(THIS_MODULE);
885}
886
887static int xs_local_copy_to_xdr(struct xdr_buf *xdr, struct sk_buff *skb)
888{
889 struct xdr_skb_reader desc = {
890 .skb = skb,
891 .offset = sizeof(rpc_fraghdr),
892 .count = skb->len - sizeof(rpc_fraghdr),
893 };
894
895 if (xdr_partial_copy_from_skb(xdr, 0, &desc, xdr_skb_read_bits) < 0)
896 return -1;
897 if (desc.count)
898 return -1;
899 return 0;
900}
901
902/**
903 * xs_local_data_read_skb
904 * @xprt: transport
905 * @sk: socket
906 * @skb: skbuff
907 *
908 * Currently this assumes we can read the whole reply in a single gulp.
909 */
910static void xs_local_data_read_skb(struct rpc_xprt *xprt,
911 struct sock *sk,
912 struct sk_buff *skb)
913{
914 struct rpc_task *task;
915 struct rpc_rqst *rovr;
916 int repsize, copied;
917 u32 _xid;
918 __be32 *xp;
919
920 repsize = skb->len - sizeof(rpc_fraghdr);
921 if (repsize < 4) {
922 dprintk("RPC: impossible RPC reply size %d\n", repsize);
923 return;
924 }
925
926 /* Copy the XID from the skb... */
927 xp = skb_header_pointer(skb, sizeof(rpc_fraghdr), sizeof(_xid), &_xid);
928 if (xp == NULL)
929 return;
930
931 /* Look up and lock the request corresponding to the given XID */
932 spin_lock_bh(&xprt->transport_lock);
933 rovr = xprt_lookup_rqst(xprt, *xp);
934 if (!rovr)
935 goto out_unlock;
936 task = rovr->rq_task;
937
938 copied = rovr->rq_private_buf.buflen;
939 if (copied > repsize)
940 copied = repsize;
941
942 if (xs_local_copy_to_xdr(&rovr->rq_private_buf, skb)) {
943 dprintk("RPC: sk_buff copy failed\n");
944 goto out_unlock;
945 }
946
947 xprt_complete_rqst(task, copied);
948
949 out_unlock:
950 spin_unlock_bh(&xprt->transport_lock);
951}
952
953static void xs_local_data_receive(struct sock_xprt *transport)
954{
955 struct sk_buff *skb;
956 struct sock *sk;
957 int err;
958
959 mutex_lock(&transport->recv_mutex);
960 sk = transport->inet;
961 if (sk == NULL)
962 goto out;
963 for (;;) {
964 skb = skb_recv_datagram(sk, 0, 1, &err);
965 if (skb == NULL)
966 break;
967 xs_local_data_read_skb(&transport->xprt, sk, skb);
968 skb_free_datagram(sk, skb);
969 }
970out:
971 mutex_unlock(&transport->recv_mutex);
972}
973
974static void xs_local_data_receive_workfn(struct work_struct *work)
975{
976 struct sock_xprt *transport =
977 container_of(work, struct sock_xprt, recv_worker);
978 xs_local_data_receive(transport);
979}
980
981/**
982 * xs_udp_data_read_skb - receive callback for UDP sockets
983 * @xprt: transport
984 * @sk: socket
985 * @skb: skbuff
986 *
987 */
988static void xs_udp_data_read_skb(struct rpc_xprt *xprt,
989 struct sock *sk,
990 struct sk_buff *skb)
991{
992 struct rpc_task *task;
993 struct rpc_rqst *rovr;
994 int repsize, copied;
995 u32 _xid;
996 __be32 *xp;
997
998 repsize = skb->len - sizeof(struct udphdr);
999 if (repsize < 4) {
1000 dprintk("RPC: impossible RPC reply size %d!\n", repsize);
1001 return;
1002 }
1003
1004 /* Copy the XID from the skb... */
1005 xp = skb_header_pointer(skb, sizeof(struct udphdr),
1006 sizeof(_xid), &_xid);
1007 if (xp == NULL)
1008 return;
1009
1010 /* Look up and lock the request corresponding to the given XID */
1011 spin_lock_bh(&xprt->transport_lock);
1012 rovr = xprt_lookup_rqst(xprt, *xp);
1013 if (!rovr)
1014 goto out_unlock;
1015 task = rovr->rq_task;
1016
1017 if ((copied = rovr->rq_private_buf.buflen) > repsize)
1018 copied = repsize;
1019
1020 /* Suck it into the iovec, verify checksum if not done by hw. */
1021 if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb)) {
1022 UDPX_INC_STATS_BH(sk, UDP_MIB_INERRORS);
1023 goto out_unlock;
1024 }
1025
1026 UDPX_INC_STATS_BH(sk, UDP_MIB_INDATAGRAMS);
1027
1028 xprt_adjust_cwnd(xprt, task, copied);
1029 xprt_complete_rqst(task, copied);
1030
1031 out_unlock:
1032 spin_unlock_bh(&xprt->transport_lock);
1033}
1034
1035static void xs_udp_data_receive(struct sock_xprt *transport)
1036{
1037 struct sk_buff *skb;
1038 struct sock *sk;
1039 int err;
1040
1041 mutex_lock(&transport->recv_mutex);
1042 sk = transport->inet;
1043 if (sk == NULL)
1044 goto out;
1045 for (;;) {
1046 skb = skb_recv_datagram(sk, 0, 1, &err);
1047 if (skb == NULL)
1048 break;
1049 xs_udp_data_read_skb(&transport->xprt, sk, skb);
1050 skb_free_datagram(sk, skb);
1051 }
1052out:
1053 mutex_unlock(&transport->recv_mutex);
1054}
1055
1056static void xs_udp_data_receive_workfn(struct work_struct *work)
1057{
1058 struct sock_xprt *transport =
1059 container_of(work, struct sock_xprt, recv_worker);
1060 xs_udp_data_receive(transport);
1061}
1062
1063/**
1064 * xs_data_ready - "data ready" callback for UDP sockets
1065 * @sk: socket with data to read
1066 *
1067 */
1068static void xs_data_ready(struct sock *sk)
1069{
1070 struct rpc_xprt *xprt;
1071
1072 read_lock_bh(&sk->sk_callback_lock);
1073 dprintk("RPC: xs_data_ready...\n");
1074 xprt = xprt_from_sock(sk);
1075 if (xprt != NULL) {
1076 struct sock_xprt *transport = container_of(xprt,
1077 struct sock_xprt, xprt);
1078 queue_work(rpciod_workqueue, &transport->recv_worker);
1079 }
1080 read_unlock_bh(&sk->sk_callback_lock);
1081}
1082
1083/*
1084 * Helper function to force a TCP close if the server is sending
1085 * junk and/or it has put us in CLOSE_WAIT
1086 */
1087static void xs_tcp_force_close(struct rpc_xprt *xprt)
1088{
1089 xprt_force_disconnect(xprt);
1090}
1091
1092static inline void xs_tcp_read_fraghdr(struct rpc_xprt *xprt, struct xdr_skb_reader *desc)
1093{
1094 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1095 size_t len, used;
1096 char *p;
1097
1098 p = ((char *) &transport->tcp_fraghdr) + transport->tcp_offset;
1099 len = sizeof(transport->tcp_fraghdr) - transport->tcp_offset;
1100 used = xdr_skb_read_bits(desc, p, len);
1101 transport->tcp_offset += used;
1102 if (used != len)
1103 return;
1104
1105 transport->tcp_reclen = ntohl(transport->tcp_fraghdr);
1106 if (transport->tcp_reclen & RPC_LAST_STREAM_FRAGMENT)
1107 transport->tcp_flags |= TCP_RCV_LAST_FRAG;
1108 else
1109 transport->tcp_flags &= ~TCP_RCV_LAST_FRAG;
1110 transport->tcp_reclen &= RPC_FRAGMENT_SIZE_MASK;
1111
1112 transport->tcp_flags &= ~TCP_RCV_COPY_FRAGHDR;
1113 transport->tcp_offset = 0;
1114
1115 /* Sanity check of the record length */
1116 if (unlikely(transport->tcp_reclen < 8)) {
1117 dprintk("RPC: invalid TCP record fragment length\n");
1118 xs_tcp_force_close(xprt);
1119 return;
1120 }
1121 dprintk("RPC: reading TCP record fragment of length %d\n",
1122 transport->tcp_reclen);
1123}
1124
1125static void xs_tcp_check_fraghdr(struct sock_xprt *transport)
1126{
1127 if (transport->tcp_offset == transport->tcp_reclen) {
1128 transport->tcp_flags |= TCP_RCV_COPY_FRAGHDR;
1129 transport->tcp_offset = 0;
1130 if (transport->tcp_flags & TCP_RCV_LAST_FRAG) {
1131 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
1132 transport->tcp_flags |= TCP_RCV_COPY_XID;
1133 transport->tcp_copied = 0;
1134 }
1135 }
1136}
1137
1138static inline void xs_tcp_read_xid(struct sock_xprt *transport, struct xdr_skb_reader *desc)
1139{
1140 size_t len, used;
1141 char *p;
1142
1143 len = sizeof(transport->tcp_xid) - transport->tcp_offset;
1144 dprintk("RPC: reading XID (%Zu bytes)\n", len);
1145 p = ((char *) &transport->tcp_xid) + transport->tcp_offset;
1146 used = xdr_skb_read_bits(desc, p, len);
1147 transport->tcp_offset += used;
1148 if (used != len)
1149 return;
1150 transport->tcp_flags &= ~TCP_RCV_COPY_XID;
1151 transport->tcp_flags |= TCP_RCV_READ_CALLDIR;
1152 transport->tcp_copied = 4;
1153 dprintk("RPC: reading %s XID %08x\n",
1154 (transport->tcp_flags & TCP_RPC_REPLY) ? "reply for"
1155 : "request with",
1156 ntohl(transport->tcp_xid));
1157 xs_tcp_check_fraghdr(transport);
1158}
1159
1160static inline void xs_tcp_read_calldir(struct sock_xprt *transport,
1161 struct xdr_skb_reader *desc)
1162{
1163 size_t len, used;
1164 u32 offset;
1165 char *p;
1166
1167 /*
1168 * We want transport->tcp_offset to be 8 at the end of this routine
1169 * (4 bytes for the xid and 4 bytes for the call/reply flag).
1170 * When this function is called for the first time,
1171 * transport->tcp_offset is 4 (after having already read the xid).
1172 */
1173 offset = transport->tcp_offset - sizeof(transport->tcp_xid);
1174 len = sizeof(transport->tcp_calldir) - offset;
1175 dprintk("RPC: reading CALL/REPLY flag (%Zu bytes)\n", len);
1176 p = ((char *) &transport->tcp_calldir) + offset;
1177 used = xdr_skb_read_bits(desc, p, len);
1178 transport->tcp_offset += used;
1179 if (used != len)
1180 return;
1181 transport->tcp_flags &= ~TCP_RCV_READ_CALLDIR;
1182 /*
1183 * We don't yet have the XDR buffer, so we will write the calldir
1184 * out after we get the buffer from the 'struct rpc_rqst'
1185 */
1186 switch (ntohl(transport->tcp_calldir)) {
1187 case RPC_REPLY:
1188 transport->tcp_flags |= TCP_RCV_COPY_CALLDIR;
1189 transport->tcp_flags |= TCP_RCV_COPY_DATA;
1190 transport->tcp_flags |= TCP_RPC_REPLY;
1191 break;
1192 case RPC_CALL:
1193 transport->tcp_flags |= TCP_RCV_COPY_CALLDIR;
1194 transport->tcp_flags |= TCP_RCV_COPY_DATA;
1195 transport->tcp_flags &= ~TCP_RPC_REPLY;
1196 break;
1197 default:
1198 dprintk("RPC: invalid request message type\n");
1199 xs_tcp_force_close(&transport->xprt);
1200 }
1201 xs_tcp_check_fraghdr(transport);
1202}
1203
1204static inline void xs_tcp_read_common(struct rpc_xprt *xprt,
1205 struct xdr_skb_reader *desc,
1206 struct rpc_rqst *req)
1207{
1208 struct sock_xprt *transport =
1209 container_of(xprt, struct sock_xprt, xprt);
1210 struct xdr_buf *rcvbuf;
1211 size_t len;
1212 ssize_t r;
1213
1214 rcvbuf = &req->rq_private_buf;
1215
1216 if (transport->tcp_flags & TCP_RCV_COPY_CALLDIR) {
1217 /*
1218 * Save the RPC direction in the XDR buffer
1219 */
1220 memcpy(rcvbuf->head[0].iov_base + transport->tcp_copied,
1221 &transport->tcp_calldir,
1222 sizeof(transport->tcp_calldir));
1223 transport->tcp_copied += sizeof(transport->tcp_calldir);
1224 transport->tcp_flags &= ~TCP_RCV_COPY_CALLDIR;
1225 }
1226
1227 len = desc->count;
1228 if (len > transport->tcp_reclen - transport->tcp_offset) {
1229 struct xdr_skb_reader my_desc;
1230
1231 len = transport->tcp_reclen - transport->tcp_offset;
1232 memcpy(&my_desc, desc, sizeof(my_desc));
1233 my_desc.count = len;
1234 r = xdr_partial_copy_from_skb(rcvbuf, transport->tcp_copied,
1235 &my_desc, xdr_skb_read_bits);
1236 desc->count -= r;
1237 desc->offset += r;
1238 } else
1239 r = xdr_partial_copy_from_skb(rcvbuf, transport->tcp_copied,
1240 desc, xdr_skb_read_bits);
1241
1242 if (r > 0) {
1243 transport->tcp_copied += r;
1244 transport->tcp_offset += r;
1245 }
1246 if (r != len) {
1247 /* Error when copying to the receive buffer,
1248 * usually because we weren't able to allocate
1249 * additional buffer pages. All we can do now
1250 * is turn off TCP_RCV_COPY_DATA, so the request
1251 * will not receive any additional updates,
1252 * and time out.
1253 * Any remaining data from this record will
1254 * be discarded.
1255 */
1256 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
1257 dprintk("RPC: XID %08x truncated request\n",
1258 ntohl(transport->tcp_xid));
1259 dprintk("RPC: xprt = %p, tcp_copied = %lu, "
1260 "tcp_offset = %u, tcp_reclen = %u\n",
1261 xprt, transport->tcp_copied,
1262 transport->tcp_offset, transport->tcp_reclen);
1263 return;
1264 }
1265
1266 dprintk("RPC: XID %08x read %Zd bytes\n",
1267 ntohl(transport->tcp_xid), r);
1268 dprintk("RPC: xprt = %p, tcp_copied = %lu, tcp_offset = %u, "
1269 "tcp_reclen = %u\n", xprt, transport->tcp_copied,
1270 transport->tcp_offset, transport->tcp_reclen);
1271
1272 if (transport->tcp_copied == req->rq_private_buf.buflen)
1273 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
1274 else if (transport->tcp_offset == transport->tcp_reclen) {
1275 if (transport->tcp_flags & TCP_RCV_LAST_FRAG)
1276 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
1277 }
1278}
1279
1280/*
1281 * Finds the request corresponding to the RPC xid and invokes the common
1282 * tcp read code to read the data.
1283 */
1284static inline int xs_tcp_read_reply(struct rpc_xprt *xprt,
1285 struct xdr_skb_reader *desc)
1286{
1287 struct sock_xprt *transport =
1288 container_of(xprt, struct sock_xprt, xprt);
1289 struct rpc_rqst *req;
1290
1291 dprintk("RPC: read reply XID %08x\n", ntohl(transport->tcp_xid));
1292
1293 /* Find and lock the request corresponding to this xid */
1294 spin_lock_bh(&xprt->transport_lock);
1295 req = xprt_lookup_rqst(xprt, transport->tcp_xid);
1296 if (!req) {
1297 dprintk("RPC: XID %08x request not found!\n",
1298 ntohl(transport->tcp_xid));
1299 spin_unlock_bh(&xprt->transport_lock);
1300 return -1;
1301 }
1302
1303 xs_tcp_read_common(xprt, desc, req);
1304
1305 if (!(transport->tcp_flags & TCP_RCV_COPY_DATA))
1306 xprt_complete_rqst(req->rq_task, transport->tcp_copied);
1307
1308 spin_unlock_bh(&xprt->transport_lock);
1309 return 0;
1310}
1311
1312#if defined(CONFIG_SUNRPC_BACKCHANNEL)
1313/*
1314 * Obtains an rpc_rqst previously allocated and invokes the common
1315 * tcp read code to read the data. The result is placed in the callback
1316 * queue.
1317 * If we're unable to obtain the rpc_rqst we schedule the closing of the
1318 * connection and return -1.
1319 */
1320static int xs_tcp_read_callback(struct rpc_xprt *xprt,
1321 struct xdr_skb_reader *desc)
1322{
1323 struct sock_xprt *transport =
1324 container_of(xprt, struct sock_xprt, xprt);
1325 struct rpc_rqst *req;
1326
1327 /* Look up and lock the request corresponding to the given XID */
1328 spin_lock_bh(&xprt->transport_lock);
1329 req = xprt_lookup_bc_request(xprt, transport->tcp_xid);
1330 if (req == NULL) {
1331 spin_unlock_bh(&xprt->transport_lock);
1332 printk(KERN_WARNING "Callback slot table overflowed\n");
1333 xprt_force_disconnect(xprt);
1334 return -1;
1335 }
1336
1337 dprintk("RPC: read callback XID %08x\n", ntohl(req->rq_xid));
1338 xs_tcp_read_common(xprt, desc, req);
1339
1340 if (!(transport->tcp_flags & TCP_RCV_COPY_DATA))
1341 xprt_complete_bc_request(req, transport->tcp_copied);
1342 spin_unlock_bh(&xprt->transport_lock);
1343
1344 return 0;
1345}
1346
1347static inline int _xs_tcp_read_data(struct rpc_xprt *xprt,
1348 struct xdr_skb_reader *desc)
1349{
1350 struct sock_xprt *transport =
1351 container_of(xprt, struct sock_xprt, xprt);
1352
1353 return (transport->tcp_flags & TCP_RPC_REPLY) ?
1354 xs_tcp_read_reply(xprt, desc) :
1355 xs_tcp_read_callback(xprt, desc);
1356}
1357
1358static int xs_tcp_bc_up(struct svc_serv *serv, struct net *net)
1359{
1360 int ret;
1361
1362 ret = svc_create_xprt(serv, "tcp-bc", net, PF_INET, 0,
1363 SVC_SOCK_ANONYMOUS);
1364 if (ret < 0)
1365 return ret;
1366 return 0;
1367}
1368#else
1369static inline int _xs_tcp_read_data(struct rpc_xprt *xprt,
1370 struct xdr_skb_reader *desc)
1371{
1372 return xs_tcp_read_reply(xprt, desc);
1373}
1374#endif /* CONFIG_SUNRPC_BACKCHANNEL */
1375
1376/*
1377 * Read data off the transport. This can be either an RPC_CALL or an
1378 * RPC_REPLY. Relay the processing to helper functions.
1379 */
1380static void xs_tcp_read_data(struct rpc_xprt *xprt,
1381 struct xdr_skb_reader *desc)
1382{
1383 struct sock_xprt *transport =
1384 container_of(xprt, struct sock_xprt, xprt);
1385
1386 if (_xs_tcp_read_data(xprt, desc) == 0)
1387 xs_tcp_check_fraghdr(transport);
1388 else {
1389 /*
1390 * The transport_lock protects the request handling.
1391 * There's no need to hold it to update the tcp_flags.
1392 */
1393 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
1394 }
1395}
1396
1397static inline void xs_tcp_read_discard(struct sock_xprt *transport, struct xdr_skb_reader *desc)
1398{
1399 size_t len;
1400
1401 len = transport->tcp_reclen - transport->tcp_offset;
1402 if (len > desc->count)
1403 len = desc->count;
1404 desc->count -= len;
1405 desc->offset += len;
1406 transport->tcp_offset += len;
1407 dprintk("RPC: discarded %Zu bytes\n", len);
1408 xs_tcp_check_fraghdr(transport);
1409}
1410
1411static int xs_tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, unsigned int offset, size_t len)
1412{
1413 struct rpc_xprt *xprt = rd_desc->arg.data;
1414 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1415 struct xdr_skb_reader desc = {
1416 .skb = skb,
1417 .offset = offset,
1418 .count = len,
1419 };
1420
1421 dprintk("RPC: xs_tcp_data_recv started\n");
1422 do {
1423 trace_xs_tcp_data_recv(transport);
1424 /* Read in a new fragment marker if necessary */
1425 /* Can we ever really expect to get completely empty fragments? */
1426 if (transport->tcp_flags & TCP_RCV_COPY_FRAGHDR) {
1427 xs_tcp_read_fraghdr(xprt, &desc);
1428 continue;
1429 }
1430 /* Read in the xid if necessary */
1431 if (transport->tcp_flags & TCP_RCV_COPY_XID) {
1432 xs_tcp_read_xid(transport, &desc);
1433 continue;
1434 }
1435 /* Read in the call/reply flag */
1436 if (transport->tcp_flags & TCP_RCV_READ_CALLDIR) {
1437 xs_tcp_read_calldir(transport, &desc);
1438 continue;
1439 }
1440 /* Read in the request data */
1441 if (transport->tcp_flags & TCP_RCV_COPY_DATA) {
1442 xs_tcp_read_data(xprt, &desc);
1443 continue;
1444 }
1445 /* Skip over any trailing bytes on short reads */
1446 xs_tcp_read_discard(transport, &desc);
1447 } while (desc.count);
1448 trace_xs_tcp_data_recv(transport);
1449 dprintk("RPC: xs_tcp_data_recv done\n");
1450 return len - desc.count;
1451}
1452
1453static void xs_tcp_data_receive(struct sock_xprt *transport)
1454{
1455 struct rpc_xprt *xprt = &transport->xprt;
1456 struct sock *sk;
1457 read_descriptor_t rd_desc = {
1458 .count = 2*1024*1024,
1459 .arg.data = xprt,
1460 };
1461 unsigned long total = 0;
1462 int read = 0;
1463
1464 mutex_lock(&transport->recv_mutex);
1465 sk = transport->inet;
1466 if (sk == NULL)
1467 goto out;
1468
1469 /* We use rd_desc to pass struct xprt to xs_tcp_data_recv */
1470 for (;;) {
1471 lock_sock(sk);
1472 read = tcp_read_sock(sk, &rd_desc, xs_tcp_data_recv);
1473 release_sock(sk);
1474 if (read <= 0)
1475 break;
1476 total += read;
1477 rd_desc.count = 65536;
1478 }
1479out:
1480 mutex_unlock(&transport->recv_mutex);
1481 trace_xs_tcp_data_ready(xprt, read, total);
1482}
1483
1484static void xs_tcp_data_receive_workfn(struct work_struct *work)
1485{
1486 struct sock_xprt *transport =
1487 container_of(work, struct sock_xprt, recv_worker);
1488 xs_tcp_data_receive(transport);
1489}
1490
1491/**
1492 * xs_tcp_data_ready - "data ready" callback for TCP sockets
1493 * @sk: socket with data to read
1494 *
1495 */
1496static void xs_tcp_data_ready(struct sock *sk)
1497{
1498 struct sock_xprt *transport;
1499 struct rpc_xprt *xprt;
1500
1501 dprintk("RPC: xs_tcp_data_ready...\n");
1502
1503 read_lock_bh(&sk->sk_callback_lock);
1504 if (!(xprt = xprt_from_sock(sk)))
1505 goto out;
1506 transport = container_of(xprt, struct sock_xprt, xprt);
1507
1508 /* Any data means we had a useful conversation, so
1509 * the we don't need to delay the next reconnect
1510 */
1511 if (xprt->reestablish_timeout)
1512 xprt->reestablish_timeout = 0;
1513 queue_work(rpciod_workqueue, &transport->recv_worker);
1514
1515out:
1516 read_unlock_bh(&sk->sk_callback_lock);
1517}
1518
1519/**
1520 * xs_tcp_state_change - callback to handle TCP socket state changes
1521 * @sk: socket whose state has changed
1522 *
1523 */
1524static void xs_tcp_state_change(struct sock *sk)
1525{
1526 struct rpc_xprt *xprt;
1527 struct sock_xprt *transport;
1528
1529 read_lock_bh(&sk->sk_callback_lock);
1530 if (!(xprt = xprt_from_sock(sk)))
1531 goto out;
1532 dprintk("RPC: xs_tcp_state_change client %p...\n", xprt);
1533 dprintk("RPC: state %x conn %d dead %d zapped %d sk_shutdown %d\n",
1534 sk->sk_state, xprt_connected(xprt),
1535 sock_flag(sk, SOCK_DEAD),
1536 sock_flag(sk, SOCK_ZAPPED),
1537 sk->sk_shutdown);
1538
1539 transport = container_of(xprt, struct sock_xprt, xprt);
1540 trace_rpc_socket_state_change(xprt, sk->sk_socket);
1541 switch (sk->sk_state) {
1542 case TCP_ESTABLISHED:
1543 spin_lock(&xprt->transport_lock);
1544 if (!xprt_test_and_set_connected(xprt)) {
1545
1546 /* Reset TCP record info */
1547 transport->tcp_offset = 0;
1548 transport->tcp_reclen = 0;
1549 transport->tcp_copied = 0;
1550 transport->tcp_flags =
1551 TCP_RCV_COPY_FRAGHDR | TCP_RCV_COPY_XID;
1552 xprt->connect_cookie++;
1553 clear_bit(XPRT_SOCK_CONNECTING, &transport->sock_state);
1554 xprt_clear_connecting(xprt);
1555
1556 xprt_wake_pending_tasks(xprt, -EAGAIN);
1557 }
1558 spin_unlock(&xprt->transport_lock);
1559 break;
1560 case TCP_FIN_WAIT1:
1561 /* The client initiated a shutdown of the socket */
1562 xprt->connect_cookie++;
1563 xprt->reestablish_timeout = 0;
1564 set_bit(XPRT_CLOSING, &xprt->state);
1565 smp_mb__before_atomic();
1566 clear_bit(XPRT_CONNECTED, &xprt->state);
1567 clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
1568 smp_mb__after_atomic();
1569 break;
1570 case TCP_CLOSE_WAIT:
1571 /* The server initiated a shutdown of the socket */
1572 xprt->connect_cookie++;
1573 clear_bit(XPRT_CONNECTED, &xprt->state);
1574 xs_tcp_force_close(xprt);
1575 case TCP_CLOSING:
1576 /*
1577 * If the server closed down the connection, make sure that
1578 * we back off before reconnecting
1579 */
1580 if (xprt->reestablish_timeout < XS_TCP_INIT_REEST_TO)
1581 xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
1582 break;
1583 case TCP_LAST_ACK:
1584 set_bit(XPRT_CLOSING, &xprt->state);
1585 smp_mb__before_atomic();
1586 clear_bit(XPRT_CONNECTED, &xprt->state);
1587 smp_mb__after_atomic();
1588 break;
1589 case TCP_CLOSE:
1590 if (test_and_clear_bit(XPRT_SOCK_CONNECTING,
1591 &transport->sock_state))
1592 xprt_clear_connecting(xprt);
1593 xs_sock_mark_closed(xprt);
1594 }
1595 out:
1596 read_unlock_bh(&sk->sk_callback_lock);
1597}
1598
1599static void xs_write_space(struct sock *sk)
1600{
1601 struct socket_wq *wq;
1602 struct rpc_xprt *xprt;
1603
1604 if (!sk->sk_socket)
1605 return;
1606 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1607
1608 if (unlikely(!(xprt = xprt_from_sock(sk))))
1609 return;
1610 rcu_read_lock();
1611 wq = rcu_dereference(sk->sk_wq);
1612 if (!wq || test_and_clear_bit(SOCKWQ_ASYNC_NOSPACE, &wq->flags) == 0)
1613 goto out;
1614
1615 xprt_write_space(xprt);
1616out:
1617 rcu_read_unlock();
1618}
1619
1620/**
1621 * xs_udp_write_space - callback invoked when socket buffer space
1622 * becomes available
1623 * @sk: socket whose state has changed
1624 *
1625 * Called when more output buffer space is available for this socket.
1626 * We try not to wake our writers until they can make "significant"
1627 * progress, otherwise we'll waste resources thrashing kernel_sendmsg
1628 * with a bunch of small requests.
1629 */
1630static void xs_udp_write_space(struct sock *sk)
1631{
1632 read_lock_bh(&sk->sk_callback_lock);
1633
1634 /* from net/core/sock.c:sock_def_write_space */
1635 if (sock_writeable(sk))
1636 xs_write_space(sk);
1637
1638 read_unlock_bh(&sk->sk_callback_lock);
1639}
1640
1641/**
1642 * xs_tcp_write_space - callback invoked when socket buffer space
1643 * becomes available
1644 * @sk: socket whose state has changed
1645 *
1646 * Called when more output buffer space is available for this socket.
1647 * We try not to wake our writers until they can make "significant"
1648 * progress, otherwise we'll waste resources thrashing kernel_sendmsg
1649 * with a bunch of small requests.
1650 */
1651static void xs_tcp_write_space(struct sock *sk)
1652{
1653 read_lock_bh(&sk->sk_callback_lock);
1654
1655 /* from net/core/stream.c:sk_stream_write_space */
1656 if (sk_stream_is_writeable(sk))
1657 xs_write_space(sk);
1658
1659 read_unlock_bh(&sk->sk_callback_lock);
1660}
1661
1662static void xs_udp_do_set_buffer_size(struct rpc_xprt *xprt)
1663{
1664 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1665 struct sock *sk = transport->inet;
1666
1667 if (transport->rcvsize) {
1668 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
1669 sk->sk_rcvbuf = transport->rcvsize * xprt->max_reqs * 2;
1670 }
1671 if (transport->sndsize) {
1672 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
1673 sk->sk_sndbuf = transport->sndsize * xprt->max_reqs * 2;
1674 sk->sk_write_space(sk);
1675 }
1676}
1677
1678/**
1679 * xs_udp_set_buffer_size - set send and receive limits
1680 * @xprt: generic transport
1681 * @sndsize: requested size of send buffer, in bytes
1682 * @rcvsize: requested size of receive buffer, in bytes
1683 *
1684 * Set socket send and receive buffer size limits.
1685 */
1686static void xs_udp_set_buffer_size(struct rpc_xprt *xprt, size_t sndsize, size_t rcvsize)
1687{
1688 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1689
1690 transport->sndsize = 0;
1691 if (sndsize)
1692 transport->sndsize = sndsize + 1024;
1693 transport->rcvsize = 0;
1694 if (rcvsize)
1695 transport->rcvsize = rcvsize + 1024;
1696
1697 xs_udp_do_set_buffer_size(xprt);
1698}
1699
1700/**
1701 * xs_udp_timer - called when a retransmit timeout occurs on a UDP transport
1702 * @task: task that timed out
1703 *
1704 * Adjust the congestion window after a retransmit timeout has occurred.
1705 */
1706static void xs_udp_timer(struct rpc_xprt *xprt, struct rpc_task *task)
1707{
1708 xprt_adjust_cwnd(xprt, task, -ETIMEDOUT);
1709}
1710
1711static unsigned short xs_get_random_port(void)
1712{
1713 unsigned short range = xprt_max_resvport - xprt_min_resvport;
1714 unsigned short rand = (unsigned short) prandom_u32() % range;
1715 return rand + xprt_min_resvport;
1716}
1717
1718/**
1719 * xs_set_reuseaddr_port - set the socket's port and address reuse options
1720 * @sock: socket
1721 *
1722 * Note that this function has to be called on all sockets that share the
1723 * same port, and it must be called before binding.
1724 */
1725static void xs_sock_set_reuseport(struct socket *sock)
1726{
1727 int opt = 1;
1728
1729 kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEPORT,
1730 (char *)&opt, sizeof(opt));
1731}
1732
1733static unsigned short xs_sock_getport(struct socket *sock)
1734{
1735 struct sockaddr_storage buf;
1736 int buflen;
1737 unsigned short port = 0;
1738
1739 if (kernel_getsockname(sock, (struct sockaddr *)&buf, &buflen) < 0)
1740 goto out;
1741 switch (buf.ss_family) {
1742 case AF_INET6:
1743 port = ntohs(((struct sockaddr_in6 *)&buf)->sin6_port);
1744 break;
1745 case AF_INET:
1746 port = ntohs(((struct sockaddr_in *)&buf)->sin_port);
1747 }
1748out:
1749 return port;
1750}
1751
1752/**
1753 * xs_set_port - reset the port number in the remote endpoint address
1754 * @xprt: generic transport
1755 * @port: new port number
1756 *
1757 */
1758static void xs_set_port(struct rpc_xprt *xprt, unsigned short port)
1759{
1760 dprintk("RPC: setting port for xprt %p to %u\n", xprt, port);
1761
1762 rpc_set_port(xs_addr(xprt), port);
1763 xs_update_peer_port(xprt);
1764}
1765
1766static void xs_set_srcport(struct sock_xprt *transport, struct socket *sock)
1767{
1768 if (transport->srcport == 0)
1769 transport->srcport = xs_sock_getport(sock);
1770}
1771
1772static unsigned short xs_get_srcport(struct sock_xprt *transport)
1773{
1774 unsigned short port = transport->srcport;
1775
1776 if (port == 0 && transport->xprt.resvport)
1777 port = xs_get_random_port();
1778 return port;
1779}
1780
1781static unsigned short xs_next_srcport(struct sock_xprt *transport, unsigned short port)
1782{
1783 if (transport->srcport != 0)
1784 transport->srcport = 0;
1785 if (!transport->xprt.resvport)
1786 return 0;
1787 if (port <= xprt_min_resvport || port > xprt_max_resvport)
1788 return xprt_max_resvport;
1789 return --port;
1790}
1791static int xs_bind(struct sock_xprt *transport, struct socket *sock)
1792{
1793 struct sockaddr_storage myaddr;
1794 int err, nloop = 0;
1795 unsigned short port = xs_get_srcport(transport);
1796 unsigned short last;
1797
1798 /*
1799 * If we are asking for any ephemeral port (i.e. port == 0 &&
1800 * transport->xprt.resvport == 0), don't bind. Let the local
1801 * port selection happen implicitly when the socket is used
1802 * (for example at connect time).
1803 *
1804 * This ensures that we can continue to establish TCP
1805 * connections even when all local ephemeral ports are already
1806 * a part of some TCP connection. This makes no difference
1807 * for UDP sockets, but also doens't harm them.
1808 *
1809 * If we're asking for any reserved port (i.e. port == 0 &&
1810 * transport->xprt.resvport == 1) xs_get_srcport above will
1811 * ensure that port is non-zero and we will bind as needed.
1812 */
1813 if (port == 0)
1814 return 0;
1815
1816 memcpy(&myaddr, &transport->srcaddr, transport->xprt.addrlen);
1817 do {
1818 rpc_set_port((struct sockaddr *)&myaddr, port);
1819 err = kernel_bind(sock, (struct sockaddr *)&myaddr,
1820 transport->xprt.addrlen);
1821 if (err == 0) {
1822 transport->srcport = port;
1823 break;
1824 }
1825 last = port;
1826 port = xs_next_srcport(transport, port);
1827 if (port > last)
1828 nloop++;
1829 } while (err == -EADDRINUSE && nloop != 2);
1830
1831 if (myaddr.ss_family == AF_INET)
1832 dprintk("RPC: %s %pI4:%u: %s (%d)\n", __func__,
1833 &((struct sockaddr_in *)&myaddr)->sin_addr,
1834 port, err ? "failed" : "ok", err);
1835 else
1836 dprintk("RPC: %s %pI6:%u: %s (%d)\n", __func__,
1837 &((struct sockaddr_in6 *)&myaddr)->sin6_addr,
1838 port, err ? "failed" : "ok", err);
1839 return err;
1840}
1841
1842/*
1843 * We don't support autobind on AF_LOCAL sockets
1844 */
1845static void xs_local_rpcbind(struct rpc_task *task)
1846{
1847 xprt_set_bound(task->tk_xprt);
1848}
1849
1850static void xs_local_set_port(struct rpc_xprt *xprt, unsigned short port)
1851{
1852}
1853
1854#ifdef CONFIG_DEBUG_LOCK_ALLOC
1855static struct lock_class_key xs_key[2];
1856static struct lock_class_key xs_slock_key[2];
1857
1858static inline void xs_reclassify_socketu(struct socket *sock)
1859{
1860 struct sock *sk = sock->sk;
1861
1862 sock_lock_init_class_and_name(sk, "slock-AF_LOCAL-RPC",
1863 &xs_slock_key[1], "sk_lock-AF_LOCAL-RPC", &xs_key[1]);
1864}
1865
1866static inline void xs_reclassify_socket4(struct socket *sock)
1867{
1868 struct sock *sk = sock->sk;
1869
1870 sock_lock_init_class_and_name(sk, "slock-AF_INET-RPC",
1871 &xs_slock_key[0], "sk_lock-AF_INET-RPC", &xs_key[0]);
1872}
1873
1874static inline void xs_reclassify_socket6(struct socket *sock)
1875{
1876 struct sock *sk = sock->sk;
1877
1878 sock_lock_init_class_and_name(sk, "slock-AF_INET6-RPC",
1879 &xs_slock_key[1], "sk_lock-AF_INET6-RPC", &xs_key[1]);
1880}
1881
1882static inline void xs_reclassify_socket(int family, struct socket *sock)
1883{
1884 WARN_ON_ONCE(sock_owned_by_user(sock->sk));
1885 if (sock_owned_by_user(sock->sk))
1886 return;
1887
1888 switch (family) {
1889 case AF_LOCAL:
1890 xs_reclassify_socketu(sock);
1891 break;
1892 case AF_INET:
1893 xs_reclassify_socket4(sock);
1894 break;
1895 case AF_INET6:
1896 xs_reclassify_socket6(sock);
1897 break;
1898 }
1899}
1900#else
1901static inline void xs_reclassify_socket(int family, struct socket *sock)
1902{
1903}
1904#endif
1905
1906static void xs_dummy_setup_socket(struct work_struct *work)
1907{
1908}
1909
1910static struct socket *xs_create_sock(struct rpc_xprt *xprt,
1911 struct sock_xprt *transport, int family, int type,
1912 int protocol, bool reuseport)
1913{
1914 struct socket *sock;
1915 int err;
1916
1917 err = __sock_create(xprt->xprt_net, family, type, protocol, &sock, 1);
1918 if (err < 0) {
1919 dprintk("RPC: can't create %d transport socket (%d).\n",
1920 protocol, -err);
1921 goto out;
1922 }
1923 xs_reclassify_socket(family, sock);
1924
1925 if (reuseport)
1926 xs_sock_set_reuseport(sock);
1927
1928 err = xs_bind(transport, sock);
1929 if (err) {
1930 sock_release(sock);
1931 goto out;
1932 }
1933
1934 return sock;
1935out:
1936 return ERR_PTR(err);
1937}
1938
1939static int xs_local_finish_connecting(struct rpc_xprt *xprt,
1940 struct socket *sock)
1941{
1942 struct sock_xprt *transport = container_of(xprt, struct sock_xprt,
1943 xprt);
1944
1945 if (!transport->inet) {
1946 struct sock *sk = sock->sk;
1947
1948 write_lock_bh(&sk->sk_callback_lock);
1949
1950 xs_save_old_callbacks(transport, sk);
1951
1952 sk->sk_user_data = xprt;
1953 sk->sk_data_ready = xs_data_ready;
1954 sk->sk_write_space = xs_udp_write_space;
1955 sk->sk_error_report = xs_error_report;
1956 sk->sk_allocation = GFP_NOIO;
1957
1958 xprt_clear_connected(xprt);
1959
1960 /* Reset to new socket */
1961 transport->sock = sock;
1962 transport->inet = sk;
1963
1964 write_unlock_bh(&sk->sk_callback_lock);
1965 }
1966
1967 /* Tell the socket layer to start connecting... */
1968 xprt->stat.connect_count++;
1969 xprt->stat.connect_start = jiffies;
1970 return kernel_connect(sock, xs_addr(xprt), xprt->addrlen, 0);
1971}
1972
1973/**
1974 * xs_local_setup_socket - create AF_LOCAL socket, connect to a local endpoint
1975 * @transport: socket transport to connect
1976 */
1977static int xs_local_setup_socket(struct sock_xprt *transport)
1978{
1979 struct rpc_xprt *xprt = &transport->xprt;
1980 struct socket *sock;
1981 int status = -EIO;
1982
1983 status = __sock_create(xprt->xprt_net, AF_LOCAL,
1984 SOCK_STREAM, 0, &sock, 1);
1985 if (status < 0) {
1986 dprintk("RPC: can't create AF_LOCAL "
1987 "transport socket (%d).\n", -status);
1988 goto out;
1989 }
1990 xs_reclassify_socket(AF_LOCAL, sock);
1991
1992 dprintk("RPC: worker connecting xprt %p via AF_LOCAL to %s\n",
1993 xprt, xprt->address_strings[RPC_DISPLAY_ADDR]);
1994
1995 status = xs_local_finish_connecting(xprt, sock);
1996 trace_rpc_socket_connect(xprt, sock, status);
1997 switch (status) {
1998 case 0:
1999 dprintk("RPC: xprt %p connected to %s\n",
2000 xprt, xprt->address_strings[RPC_DISPLAY_ADDR]);
2001 xprt_set_connected(xprt);
2002 case -ENOBUFS:
2003 break;
2004 case -ENOENT:
2005 dprintk("RPC: xprt %p: socket %s does not exist\n",
2006 xprt, xprt->address_strings[RPC_DISPLAY_ADDR]);
2007 break;
2008 case -ECONNREFUSED:
2009 dprintk("RPC: xprt %p: connection refused for %s\n",
2010 xprt, xprt->address_strings[RPC_DISPLAY_ADDR]);
2011 break;
2012 default:
2013 printk(KERN_ERR "%s: unhandled error (%d) connecting to %s\n",
2014 __func__, -status,
2015 xprt->address_strings[RPC_DISPLAY_ADDR]);
2016 }
2017
2018out:
2019 xprt_clear_connecting(xprt);
2020 xprt_wake_pending_tasks(xprt, status);
2021 return status;
2022}
2023
2024static void xs_local_connect(struct rpc_xprt *xprt, struct rpc_task *task)
2025{
2026 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
2027 int ret;
2028
2029 if (RPC_IS_ASYNC(task)) {
2030 /*
2031 * We want the AF_LOCAL connect to be resolved in the
2032 * filesystem namespace of the process making the rpc
2033 * call. Thus we connect synchronously.
2034 *
2035 * If we want to support asynchronous AF_LOCAL calls,
2036 * we'll need to figure out how to pass a namespace to
2037 * connect.
2038 */
2039 rpc_exit(task, -ENOTCONN);
2040 return;
2041 }
2042 ret = xs_local_setup_socket(transport);
2043 if (ret && !RPC_IS_SOFTCONN(task))
2044 msleep_interruptible(15000);
2045}
2046
2047#if IS_ENABLED(CONFIG_SUNRPC_SWAP)
2048/*
2049 * Note that this should be called with XPRT_LOCKED held (or when we otherwise
2050 * know that we have exclusive access to the socket), to guard against
2051 * races with xs_reset_transport.
2052 */
2053static void xs_set_memalloc(struct rpc_xprt *xprt)
2054{
2055 struct sock_xprt *transport = container_of(xprt, struct sock_xprt,
2056 xprt);
2057
2058 /*
2059 * If there's no sock, then we have nothing to set. The
2060 * reconnecting process will get it for us.
2061 */
2062 if (!transport->inet)
2063 return;
2064 if (atomic_read(&xprt->swapper))
2065 sk_set_memalloc(transport->inet);
2066}
2067
2068/**
2069 * xs_enable_swap - Tag this transport as being used for swap.
2070 * @xprt: transport to tag
2071 *
2072 * Take a reference to this transport on behalf of the rpc_clnt, and
2073 * optionally mark it for swapping if it wasn't already.
2074 */
2075static int
2076xs_enable_swap(struct rpc_xprt *xprt)
2077{
2078 struct sock_xprt *xs = container_of(xprt, struct sock_xprt, xprt);
2079
2080 if (atomic_inc_return(&xprt->swapper) != 1)
2081 return 0;
2082 if (wait_on_bit_lock(&xprt->state, XPRT_LOCKED, TASK_KILLABLE))
2083 return -ERESTARTSYS;
2084 if (xs->inet)
2085 sk_set_memalloc(xs->inet);
2086 xprt_release_xprt(xprt, NULL);
2087 return 0;
2088}
2089
2090/**
2091 * xs_disable_swap - Untag this transport as being used for swap.
2092 * @xprt: transport to tag
2093 *
2094 * Drop a "swapper" reference to this xprt on behalf of the rpc_clnt. If the
2095 * swapper refcount goes to 0, untag the socket as a memalloc socket.
2096 */
2097static void
2098xs_disable_swap(struct rpc_xprt *xprt)
2099{
2100 struct sock_xprt *xs = container_of(xprt, struct sock_xprt, xprt);
2101
2102 if (!atomic_dec_and_test(&xprt->swapper))
2103 return;
2104 if (wait_on_bit_lock(&xprt->state, XPRT_LOCKED, TASK_KILLABLE))
2105 return;
2106 if (xs->inet)
2107 sk_clear_memalloc(xs->inet);
2108 xprt_release_xprt(xprt, NULL);
2109}
2110#else
2111static void xs_set_memalloc(struct rpc_xprt *xprt)
2112{
2113}
2114
2115static int
2116xs_enable_swap(struct rpc_xprt *xprt)
2117{
2118 return -EINVAL;
2119}
2120
2121static void
2122xs_disable_swap(struct rpc_xprt *xprt)
2123{
2124}
2125#endif
2126
2127static void xs_udp_finish_connecting(struct rpc_xprt *xprt, struct socket *sock)
2128{
2129 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
2130
2131 if (!transport->inet) {
2132 struct sock *sk = sock->sk;
2133
2134 write_lock_bh(&sk->sk_callback_lock);
2135
2136 xs_save_old_callbacks(transport, sk);
2137
2138 sk->sk_user_data = xprt;
2139 sk->sk_data_ready = xs_data_ready;
2140 sk->sk_write_space = xs_udp_write_space;
2141 sk->sk_allocation = GFP_NOIO;
2142
2143 xprt_set_connected(xprt);
2144
2145 /* Reset to new socket */
2146 transport->sock = sock;
2147 transport->inet = sk;
2148
2149 xs_set_memalloc(xprt);
2150
2151 write_unlock_bh(&sk->sk_callback_lock);
2152 }
2153 xs_udp_do_set_buffer_size(xprt);
2154}
2155
2156static void xs_udp_setup_socket(struct work_struct *work)
2157{
2158 struct sock_xprt *transport =
2159 container_of(work, struct sock_xprt, connect_worker.work);
2160 struct rpc_xprt *xprt = &transport->xprt;
2161 struct socket *sock = transport->sock;
2162 int status = -EIO;
2163
2164 sock = xs_create_sock(xprt, transport,
2165 xs_addr(xprt)->sa_family, SOCK_DGRAM,
2166 IPPROTO_UDP, false);
2167 if (IS_ERR(sock))
2168 goto out;
2169
2170 dprintk("RPC: worker connecting xprt %p via %s to "
2171 "%s (port %s)\n", xprt,
2172 xprt->address_strings[RPC_DISPLAY_PROTO],
2173 xprt->address_strings[RPC_DISPLAY_ADDR],
2174 xprt->address_strings[RPC_DISPLAY_PORT]);
2175
2176 xs_udp_finish_connecting(xprt, sock);
2177 trace_rpc_socket_connect(xprt, sock, 0);
2178 status = 0;
2179out:
2180 xprt_unlock_connect(xprt, transport);
2181 xprt_clear_connecting(xprt);
2182 xprt_wake_pending_tasks(xprt, status);
2183}
2184
2185/**
2186 * xs_tcp_shutdown - gracefully shut down a TCP socket
2187 * @xprt: transport
2188 *
2189 * Initiates a graceful shutdown of the TCP socket by calling the
2190 * equivalent of shutdown(SHUT_RDWR);
2191 */
2192static void xs_tcp_shutdown(struct rpc_xprt *xprt)
2193{
2194 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
2195 struct socket *sock = transport->sock;
2196
2197 if (sock == NULL)
2198 return;
2199 if (xprt_connected(xprt)) {
2200 kernel_sock_shutdown(sock, SHUT_RDWR);
2201 trace_rpc_socket_shutdown(xprt, sock);
2202 } else
2203 xs_reset_transport(transport);
2204}
2205
2206static int xs_tcp_finish_connecting(struct rpc_xprt *xprt, struct socket *sock)
2207{
2208 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
2209 int ret = -ENOTCONN;
2210
2211 if (!transport->inet) {
2212 struct sock *sk = sock->sk;
2213 unsigned int keepidle = xprt->timeout->to_initval / HZ;
2214 unsigned int keepcnt = xprt->timeout->to_retries + 1;
2215 unsigned int opt_on = 1;
2216 unsigned int timeo;
2217
2218 /* TCP Keepalive options */
2219 kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
2220 (char *)&opt_on, sizeof(opt_on));
2221 kernel_setsockopt(sock, SOL_TCP, TCP_KEEPIDLE,
2222 (char *)&keepidle, sizeof(keepidle));
2223 kernel_setsockopt(sock, SOL_TCP, TCP_KEEPINTVL,
2224 (char *)&keepidle, sizeof(keepidle));
2225 kernel_setsockopt(sock, SOL_TCP, TCP_KEEPCNT,
2226 (char *)&keepcnt, sizeof(keepcnt));
2227
2228 /* TCP user timeout (see RFC5482) */
2229 timeo = jiffies_to_msecs(xprt->timeout->to_initval) *
2230 (xprt->timeout->to_retries + 1);
2231 kernel_setsockopt(sock, SOL_TCP, TCP_USER_TIMEOUT,
2232 (char *)&timeo, sizeof(timeo));
2233
2234 write_lock_bh(&sk->sk_callback_lock);
2235
2236 xs_save_old_callbacks(transport, sk);
2237
2238 sk->sk_user_data = xprt;
2239 sk->sk_data_ready = xs_tcp_data_ready;
2240 sk->sk_state_change = xs_tcp_state_change;
2241 sk->sk_write_space = xs_tcp_write_space;
2242 sk->sk_error_report = xs_error_report;
2243 sk->sk_allocation = GFP_NOIO;
2244
2245 /* socket options */
2246 sock_reset_flag(sk, SOCK_LINGER);
2247 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
2248
2249 xprt_clear_connected(xprt);
2250
2251 /* Reset to new socket */
2252 transport->sock = sock;
2253 transport->inet = sk;
2254
2255 write_unlock_bh(&sk->sk_callback_lock);
2256 }
2257
2258 if (!xprt_bound(xprt))
2259 goto out;
2260
2261 xs_set_memalloc(xprt);
2262
2263 /* Tell the socket layer to start connecting... */
2264 xprt->stat.connect_count++;
2265 xprt->stat.connect_start = jiffies;
2266 set_bit(XPRT_SOCK_CONNECTING, &transport->sock_state);
2267 ret = kernel_connect(sock, xs_addr(xprt), xprt->addrlen, O_NONBLOCK);
2268 switch (ret) {
2269 case 0:
2270 xs_set_srcport(transport, sock);
2271 case -EINPROGRESS:
2272 /* SYN_SENT! */
2273 if (xprt->reestablish_timeout < XS_TCP_INIT_REEST_TO)
2274 xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
2275 }
2276out:
2277 return ret;
2278}
2279
2280/**
2281 * xs_tcp_setup_socket - create a TCP socket and connect to a remote endpoint
2282 *
2283 * Invoked by a work queue tasklet.
2284 */
2285static void xs_tcp_setup_socket(struct work_struct *work)
2286{
2287 struct sock_xprt *transport =
2288 container_of(work, struct sock_xprt, connect_worker.work);
2289 struct socket *sock = transport->sock;
2290 struct rpc_xprt *xprt = &transport->xprt;
2291 int status = -EIO;
2292
2293 if (!sock) {
2294 sock = xs_create_sock(xprt, transport,
2295 xs_addr(xprt)->sa_family, SOCK_STREAM,
2296 IPPROTO_TCP, true);
2297 if (IS_ERR(sock)) {
2298 status = PTR_ERR(sock);
2299 goto out;
2300 }
2301 }
2302
2303 dprintk("RPC: worker connecting xprt %p via %s to "
2304 "%s (port %s)\n", xprt,
2305 xprt->address_strings[RPC_DISPLAY_PROTO],
2306 xprt->address_strings[RPC_DISPLAY_ADDR],
2307 xprt->address_strings[RPC_DISPLAY_PORT]);
2308
2309 status = xs_tcp_finish_connecting(xprt, sock);
2310 trace_rpc_socket_connect(xprt, sock, status);
2311 dprintk("RPC: %p connect status %d connected %d sock state %d\n",
2312 xprt, -status, xprt_connected(xprt),
2313 sock->sk->sk_state);
2314 switch (status) {
2315 default:
2316 printk("%s: connect returned unhandled error %d\n",
2317 __func__, status);
2318 case -EADDRNOTAVAIL:
2319 /* We're probably in TIME_WAIT. Get rid of existing socket,
2320 * and retry
2321 */
2322 xs_tcp_force_close(xprt);
2323 break;
2324 case 0:
2325 case -EINPROGRESS:
2326 case -EALREADY:
2327 xprt_unlock_connect(xprt, transport);
2328 return;
2329 case -EINVAL:
2330 /* Happens, for instance, if the user specified a link
2331 * local IPv6 address without a scope-id.
2332 */
2333 case -ECONNREFUSED:
2334 case -ECONNRESET:
2335 case -ENETUNREACH:
2336 case -EADDRINUSE:
2337 case -ENOBUFS:
2338 /* retry with existing socket, after a delay */
2339 xs_tcp_force_close(xprt);
2340 goto out;
2341 }
2342 status = -EAGAIN;
2343out:
2344 xprt_unlock_connect(xprt, transport);
2345 xprt_clear_connecting(xprt);
2346 xprt_wake_pending_tasks(xprt, status);
2347}
2348
2349/**
2350 * xs_connect - connect a socket to a remote endpoint
2351 * @xprt: pointer to transport structure
2352 * @task: address of RPC task that manages state of connect request
2353 *
2354 * TCP: If the remote end dropped the connection, delay reconnecting.
2355 *
2356 * UDP socket connects are synchronous, but we use a work queue anyway
2357 * to guarantee that even unprivileged user processes can set up a
2358 * socket on a privileged port.
2359 *
2360 * If a UDP socket connect fails, the delay behavior here prevents
2361 * retry floods (hard mounts).
2362 */
2363static void xs_connect(struct rpc_xprt *xprt, struct rpc_task *task)
2364{
2365 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
2366
2367 WARN_ON_ONCE(!xprt_lock_connect(xprt, task, transport));
2368
2369 if (transport->sock != NULL) {
2370 dprintk("RPC: xs_connect delayed xprt %p for %lu "
2371 "seconds\n",
2372 xprt, xprt->reestablish_timeout / HZ);
2373
2374 /* Start by resetting any existing state */
2375 xs_reset_transport(transport);
2376
2377 queue_delayed_work(rpciod_workqueue,
2378 &transport->connect_worker,
2379 xprt->reestablish_timeout);
2380 xprt->reestablish_timeout <<= 1;
2381 if (xprt->reestablish_timeout < XS_TCP_INIT_REEST_TO)
2382 xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
2383 if (xprt->reestablish_timeout > XS_TCP_MAX_REEST_TO)
2384 xprt->reestablish_timeout = XS_TCP_MAX_REEST_TO;
2385 } else {
2386 dprintk("RPC: xs_connect scheduled xprt %p\n", xprt);
2387 queue_delayed_work(rpciod_workqueue,
2388 &transport->connect_worker, 0);
2389 }
2390}
2391
2392/**
2393 * xs_local_print_stats - display AF_LOCAL socket-specifc stats
2394 * @xprt: rpc_xprt struct containing statistics
2395 * @seq: output file
2396 *
2397 */
2398static void xs_local_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
2399{
2400 long idle_time = 0;
2401
2402 if (xprt_connected(xprt))
2403 idle_time = (long)(jiffies - xprt->last_used) / HZ;
2404
2405 seq_printf(seq, "\txprt:\tlocal %lu %lu %lu %ld %lu %lu %lu "
2406 "%llu %llu %lu %llu %llu\n",
2407 xprt->stat.bind_count,
2408 xprt->stat.connect_count,
2409 xprt->stat.connect_time,
2410 idle_time,
2411 xprt->stat.sends,
2412 xprt->stat.recvs,
2413 xprt->stat.bad_xids,
2414 xprt->stat.req_u,
2415 xprt->stat.bklog_u,
2416 xprt->stat.max_slots,
2417 xprt->stat.sending_u,
2418 xprt->stat.pending_u);
2419}
2420
2421/**
2422 * xs_udp_print_stats - display UDP socket-specifc stats
2423 * @xprt: rpc_xprt struct containing statistics
2424 * @seq: output file
2425 *
2426 */
2427static void xs_udp_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
2428{
2429 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
2430
2431 seq_printf(seq, "\txprt:\tudp %u %lu %lu %lu %lu %llu %llu "
2432 "%lu %llu %llu\n",
2433 transport->srcport,
2434 xprt->stat.bind_count,
2435 xprt->stat.sends,
2436 xprt->stat.recvs,
2437 xprt->stat.bad_xids,
2438 xprt->stat.req_u,
2439 xprt->stat.bklog_u,
2440 xprt->stat.max_slots,
2441 xprt->stat.sending_u,
2442 xprt->stat.pending_u);
2443}
2444
2445/**
2446 * xs_tcp_print_stats - display TCP socket-specifc stats
2447 * @xprt: rpc_xprt struct containing statistics
2448 * @seq: output file
2449 *
2450 */
2451static void xs_tcp_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
2452{
2453 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
2454 long idle_time = 0;
2455
2456 if (xprt_connected(xprt))
2457 idle_time = (long)(jiffies - xprt->last_used) / HZ;
2458
2459 seq_printf(seq, "\txprt:\ttcp %u %lu %lu %lu %ld %lu %lu %lu "
2460 "%llu %llu %lu %llu %llu\n",
2461 transport->srcport,
2462 xprt->stat.bind_count,
2463 xprt->stat.connect_count,
2464 xprt->stat.connect_time,
2465 idle_time,
2466 xprt->stat.sends,
2467 xprt->stat.recvs,
2468 xprt->stat.bad_xids,
2469 xprt->stat.req_u,
2470 xprt->stat.bklog_u,
2471 xprt->stat.max_slots,
2472 xprt->stat.sending_u,
2473 xprt->stat.pending_u);
2474}
2475
2476/*
2477 * Allocate a bunch of pages for a scratch buffer for the rpc code. The reason
2478 * we allocate pages instead doing a kmalloc like rpc_malloc is because we want
2479 * to use the server side send routines.
2480 */
2481static void *bc_malloc(struct rpc_task *task, size_t size)
2482{
2483 struct page *page;
2484 struct rpc_buffer *buf;
2485
2486 WARN_ON_ONCE(size > PAGE_SIZE - sizeof(struct rpc_buffer));
2487 if (size > PAGE_SIZE - sizeof(struct rpc_buffer))
2488 return NULL;
2489
2490 page = alloc_page(GFP_KERNEL);
2491 if (!page)
2492 return NULL;
2493
2494 buf = page_address(page);
2495 buf->len = PAGE_SIZE;
2496
2497 return buf->data;
2498}
2499
2500/*
2501 * Free the space allocated in the bc_alloc routine
2502 */
2503static void bc_free(void *buffer)
2504{
2505 struct rpc_buffer *buf;
2506
2507 if (!buffer)
2508 return;
2509
2510 buf = container_of(buffer, struct rpc_buffer, data);
2511 free_page((unsigned long)buf);
2512}
2513
2514/*
2515 * Use the svc_sock to send the callback. Must be called with svsk->sk_mutex
2516 * held. Borrows heavily from svc_tcp_sendto and xs_tcp_send_request.
2517 */
2518static int bc_sendto(struct rpc_rqst *req)
2519{
2520 int len;
2521 struct xdr_buf *xbufp = &req->rq_snd_buf;
2522 struct rpc_xprt *xprt = req->rq_xprt;
2523 struct sock_xprt *transport =
2524 container_of(xprt, struct sock_xprt, xprt);
2525 struct socket *sock = transport->sock;
2526 unsigned long headoff;
2527 unsigned long tailoff;
2528
2529 xs_encode_stream_record_marker(xbufp);
2530
2531 tailoff = (unsigned long)xbufp->tail[0].iov_base & ~PAGE_MASK;
2532 headoff = (unsigned long)xbufp->head[0].iov_base & ~PAGE_MASK;
2533 len = svc_send_common(sock, xbufp,
2534 virt_to_page(xbufp->head[0].iov_base), headoff,
2535 xbufp->tail[0].iov_base, tailoff);
2536
2537 if (len != xbufp->len) {
2538 printk(KERN_NOTICE "Error sending entire callback!\n");
2539 len = -EAGAIN;
2540 }
2541
2542 return len;
2543}
2544
2545/*
2546 * The send routine. Borrows from svc_send
2547 */
2548static int bc_send_request(struct rpc_task *task)
2549{
2550 struct rpc_rqst *req = task->tk_rqstp;
2551 struct svc_xprt *xprt;
2552 int len;
2553
2554 dprintk("sending request with xid: %08x\n", ntohl(req->rq_xid));
2555 /*
2556 * Get the server socket associated with this callback xprt
2557 */
2558 xprt = req->rq_xprt->bc_xprt;
2559
2560 /*
2561 * Grab the mutex to serialize data as the connection is shared
2562 * with the fore channel
2563 */
2564 if (!mutex_trylock(&xprt->xpt_mutex)) {
2565 rpc_sleep_on(&xprt->xpt_bc_pending, task, NULL);
2566 if (!mutex_trylock(&xprt->xpt_mutex))
2567 return -EAGAIN;
2568 rpc_wake_up_queued_task(&xprt->xpt_bc_pending, task);
2569 }
2570 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
2571 len = -ENOTCONN;
2572 else
2573 len = bc_sendto(req);
2574 mutex_unlock(&xprt->xpt_mutex);
2575
2576 if (len > 0)
2577 len = 0;
2578
2579 return len;
2580}
2581
2582/*
2583 * The close routine. Since this is client initiated, we do nothing
2584 */
2585
2586static void bc_close(struct rpc_xprt *xprt)
2587{
2588}
2589
2590/*
2591 * The xprt destroy routine. Again, because this connection is client
2592 * initiated, we do nothing
2593 */
2594
2595static void bc_destroy(struct rpc_xprt *xprt)
2596{
2597 dprintk("RPC: bc_destroy xprt %p\n", xprt);
2598
2599 xs_xprt_free(xprt);
2600 module_put(THIS_MODULE);
2601}
2602
2603static struct rpc_xprt_ops xs_local_ops = {
2604 .reserve_xprt = xprt_reserve_xprt,
2605 .release_xprt = xs_tcp_release_xprt,
2606 .alloc_slot = xprt_alloc_slot,
2607 .rpcbind = xs_local_rpcbind,
2608 .set_port = xs_local_set_port,
2609 .connect = xs_local_connect,
2610 .buf_alloc = rpc_malloc,
2611 .buf_free = rpc_free,
2612 .send_request = xs_local_send_request,
2613 .set_retrans_timeout = xprt_set_retrans_timeout_def,
2614 .close = xs_close,
2615 .destroy = xs_destroy,
2616 .print_stats = xs_local_print_stats,
2617 .enable_swap = xs_enable_swap,
2618 .disable_swap = xs_disable_swap,
2619};
2620
2621static struct rpc_xprt_ops xs_udp_ops = {
2622 .set_buffer_size = xs_udp_set_buffer_size,
2623 .reserve_xprt = xprt_reserve_xprt_cong,
2624 .release_xprt = xprt_release_xprt_cong,
2625 .alloc_slot = xprt_alloc_slot,
2626 .rpcbind = rpcb_getport_async,
2627 .set_port = xs_set_port,
2628 .connect = xs_connect,
2629 .buf_alloc = rpc_malloc,
2630 .buf_free = rpc_free,
2631 .send_request = xs_udp_send_request,
2632 .set_retrans_timeout = xprt_set_retrans_timeout_rtt,
2633 .timer = xs_udp_timer,
2634 .release_request = xprt_release_rqst_cong,
2635 .close = xs_close,
2636 .destroy = xs_destroy,
2637 .print_stats = xs_udp_print_stats,
2638 .enable_swap = xs_enable_swap,
2639 .disable_swap = xs_disable_swap,
2640 .inject_disconnect = xs_inject_disconnect,
2641};
2642
2643static struct rpc_xprt_ops xs_tcp_ops = {
2644 .reserve_xprt = xprt_reserve_xprt,
2645 .release_xprt = xs_tcp_release_xprt,
2646 .alloc_slot = xprt_lock_and_alloc_slot,
2647 .rpcbind = rpcb_getport_async,
2648 .set_port = xs_set_port,
2649 .connect = xs_connect,
2650 .buf_alloc = rpc_malloc,
2651 .buf_free = rpc_free,
2652 .send_request = xs_tcp_send_request,
2653 .set_retrans_timeout = xprt_set_retrans_timeout_def,
2654 .close = xs_tcp_shutdown,
2655 .destroy = xs_destroy,
2656 .print_stats = xs_tcp_print_stats,
2657 .enable_swap = xs_enable_swap,
2658 .disable_swap = xs_disable_swap,
2659 .inject_disconnect = xs_inject_disconnect,
2660#ifdef CONFIG_SUNRPC_BACKCHANNEL
2661 .bc_setup = xprt_setup_bc,
2662 .bc_up = xs_tcp_bc_up,
2663 .bc_free_rqst = xprt_free_bc_rqst,
2664 .bc_destroy = xprt_destroy_bc,
2665#endif
2666};
2667
2668/*
2669 * The rpc_xprt_ops for the server backchannel
2670 */
2671
2672static struct rpc_xprt_ops bc_tcp_ops = {
2673 .reserve_xprt = xprt_reserve_xprt,
2674 .release_xprt = xprt_release_xprt,
2675 .alloc_slot = xprt_alloc_slot,
2676 .buf_alloc = bc_malloc,
2677 .buf_free = bc_free,
2678 .send_request = bc_send_request,
2679 .set_retrans_timeout = xprt_set_retrans_timeout_def,
2680 .close = bc_close,
2681 .destroy = bc_destroy,
2682 .print_stats = xs_tcp_print_stats,
2683 .enable_swap = xs_enable_swap,
2684 .disable_swap = xs_disable_swap,
2685 .inject_disconnect = xs_inject_disconnect,
2686};
2687
2688static int xs_init_anyaddr(const int family, struct sockaddr *sap)
2689{
2690 static const struct sockaddr_in sin = {
2691 .sin_family = AF_INET,
2692 .sin_addr.s_addr = htonl(INADDR_ANY),
2693 };
2694 static const struct sockaddr_in6 sin6 = {
2695 .sin6_family = AF_INET6,
2696 .sin6_addr = IN6ADDR_ANY_INIT,
2697 };
2698
2699 switch (family) {
2700 case AF_LOCAL:
2701 break;
2702 case AF_INET:
2703 memcpy(sap, &sin, sizeof(sin));
2704 break;
2705 case AF_INET6:
2706 memcpy(sap, &sin6, sizeof(sin6));
2707 break;
2708 default:
2709 dprintk("RPC: %s: Bad address family\n", __func__);
2710 return -EAFNOSUPPORT;
2711 }
2712 return 0;
2713}
2714
2715static struct rpc_xprt *xs_setup_xprt(struct xprt_create *args,
2716 unsigned int slot_table_size,
2717 unsigned int max_slot_table_size)
2718{
2719 struct rpc_xprt *xprt;
2720 struct sock_xprt *new;
2721
2722 if (args->addrlen > sizeof(xprt->addr)) {
2723 dprintk("RPC: xs_setup_xprt: address too large\n");
2724 return ERR_PTR(-EBADF);
2725 }
2726
2727 xprt = xprt_alloc(args->net, sizeof(*new), slot_table_size,
2728 max_slot_table_size);
2729 if (xprt == NULL) {
2730 dprintk("RPC: xs_setup_xprt: couldn't allocate "
2731 "rpc_xprt\n");
2732 return ERR_PTR(-ENOMEM);
2733 }
2734
2735 new = container_of(xprt, struct sock_xprt, xprt);
2736 mutex_init(&new->recv_mutex);
2737 memcpy(&xprt->addr, args->dstaddr, args->addrlen);
2738 xprt->addrlen = args->addrlen;
2739 if (args->srcaddr)
2740 memcpy(&new->srcaddr, args->srcaddr, args->addrlen);
2741 else {
2742 int err;
2743 err = xs_init_anyaddr(args->dstaddr->sa_family,
2744 (struct sockaddr *)&new->srcaddr);
2745 if (err != 0) {
2746 xprt_free(xprt);
2747 return ERR_PTR(err);
2748 }
2749 }
2750
2751 return xprt;
2752}
2753
2754static const struct rpc_timeout xs_local_default_timeout = {
2755 .to_initval = 10 * HZ,
2756 .to_maxval = 10 * HZ,
2757 .to_retries = 2,
2758};
2759
2760/**
2761 * xs_setup_local - Set up transport to use an AF_LOCAL socket
2762 * @args: rpc transport creation arguments
2763 *
2764 * AF_LOCAL is a "tpi_cots_ord" transport, just like TCP
2765 */
2766static struct rpc_xprt *xs_setup_local(struct xprt_create *args)
2767{
2768 struct sockaddr_un *sun = (struct sockaddr_un *)args->dstaddr;
2769 struct sock_xprt *transport;
2770 struct rpc_xprt *xprt;
2771 struct rpc_xprt *ret;
2772
2773 xprt = xs_setup_xprt(args, xprt_tcp_slot_table_entries,
2774 xprt_max_tcp_slot_table_entries);
2775 if (IS_ERR(xprt))
2776 return xprt;
2777 transport = container_of(xprt, struct sock_xprt, xprt);
2778
2779 xprt->prot = 0;
2780 xprt->tsh_size = sizeof(rpc_fraghdr) / sizeof(u32);
2781 xprt->max_payload = RPC_MAX_FRAGMENT_SIZE;
2782
2783 xprt->bind_timeout = XS_BIND_TO;
2784 xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
2785 xprt->idle_timeout = XS_IDLE_DISC_TO;
2786
2787 xprt->ops = &xs_local_ops;
2788 xprt->timeout = &xs_local_default_timeout;
2789
2790 INIT_WORK(&transport->recv_worker, xs_local_data_receive_workfn);
2791 INIT_DELAYED_WORK(&transport->connect_worker,
2792 xs_dummy_setup_socket);
2793
2794 switch (sun->sun_family) {
2795 case AF_LOCAL:
2796 if (sun->sun_path[0] != '/') {
2797 dprintk("RPC: bad AF_LOCAL address: %s\n",
2798 sun->sun_path);
2799 ret = ERR_PTR(-EINVAL);
2800 goto out_err;
2801 }
2802 xprt_set_bound(xprt);
2803 xs_format_peer_addresses(xprt, "local", RPCBIND_NETID_LOCAL);
2804 ret = ERR_PTR(xs_local_setup_socket(transport));
2805 if (ret)
2806 goto out_err;
2807 break;
2808 default:
2809 ret = ERR_PTR(-EAFNOSUPPORT);
2810 goto out_err;
2811 }
2812
2813 dprintk("RPC: set up xprt to %s via AF_LOCAL\n",
2814 xprt->address_strings[RPC_DISPLAY_ADDR]);
2815
2816 if (try_module_get(THIS_MODULE))
2817 return xprt;
2818 ret = ERR_PTR(-EINVAL);
2819out_err:
2820 xs_xprt_free(xprt);
2821 return ret;
2822}
2823
2824static const struct rpc_timeout xs_udp_default_timeout = {
2825 .to_initval = 5 * HZ,
2826 .to_maxval = 30 * HZ,
2827 .to_increment = 5 * HZ,
2828 .to_retries = 5,
2829};
2830
2831/**
2832 * xs_setup_udp - Set up transport to use a UDP socket
2833 * @args: rpc transport creation arguments
2834 *
2835 */
2836static struct rpc_xprt *xs_setup_udp(struct xprt_create *args)
2837{
2838 struct sockaddr *addr = args->dstaddr;
2839 struct rpc_xprt *xprt;
2840 struct sock_xprt *transport;
2841 struct rpc_xprt *ret;
2842
2843 xprt = xs_setup_xprt(args, xprt_udp_slot_table_entries,
2844 xprt_udp_slot_table_entries);
2845 if (IS_ERR(xprt))
2846 return xprt;
2847 transport = container_of(xprt, struct sock_xprt, xprt);
2848
2849 xprt->prot = IPPROTO_UDP;
2850 xprt->tsh_size = 0;
2851 /* XXX: header size can vary due to auth type, IPv6, etc. */
2852 xprt->max_payload = (1U << 16) - (MAX_HEADER << 3);
2853
2854 xprt->bind_timeout = XS_BIND_TO;
2855 xprt->reestablish_timeout = XS_UDP_REEST_TO;
2856 xprt->idle_timeout = XS_IDLE_DISC_TO;
2857
2858 xprt->ops = &xs_udp_ops;
2859
2860 xprt->timeout = &xs_udp_default_timeout;
2861
2862 INIT_WORK(&transport->recv_worker, xs_udp_data_receive_workfn);
2863 INIT_DELAYED_WORK(&transport->connect_worker, xs_udp_setup_socket);
2864
2865 switch (addr->sa_family) {
2866 case AF_INET:
2867 if (((struct sockaddr_in *)addr)->sin_port != htons(0))
2868 xprt_set_bound(xprt);
2869
2870 xs_format_peer_addresses(xprt, "udp", RPCBIND_NETID_UDP);
2871 break;
2872 case AF_INET6:
2873 if (((struct sockaddr_in6 *)addr)->sin6_port != htons(0))
2874 xprt_set_bound(xprt);
2875
2876 xs_format_peer_addresses(xprt, "udp", RPCBIND_NETID_UDP6);
2877 break;
2878 default:
2879 ret = ERR_PTR(-EAFNOSUPPORT);
2880 goto out_err;
2881 }
2882
2883 if (xprt_bound(xprt))
2884 dprintk("RPC: set up xprt to %s (port %s) via %s\n",
2885 xprt->address_strings[RPC_DISPLAY_ADDR],
2886 xprt->address_strings[RPC_DISPLAY_PORT],
2887 xprt->address_strings[RPC_DISPLAY_PROTO]);
2888 else
2889 dprintk("RPC: set up xprt to %s (autobind) via %s\n",
2890 xprt->address_strings[RPC_DISPLAY_ADDR],
2891 xprt->address_strings[RPC_DISPLAY_PROTO]);
2892
2893 if (try_module_get(THIS_MODULE))
2894 return xprt;
2895 ret = ERR_PTR(-EINVAL);
2896out_err:
2897 xs_xprt_free(xprt);
2898 return ret;
2899}
2900
2901static const struct rpc_timeout xs_tcp_default_timeout = {
2902 .to_initval = 60 * HZ,
2903 .to_maxval = 60 * HZ,
2904 .to_retries = 2,
2905};
2906
2907/**
2908 * xs_setup_tcp - Set up transport to use a TCP socket
2909 * @args: rpc transport creation arguments
2910 *
2911 */
2912static struct rpc_xprt *xs_setup_tcp(struct xprt_create *args)
2913{
2914 struct sockaddr *addr = args->dstaddr;
2915 struct rpc_xprt *xprt;
2916 struct sock_xprt *transport;
2917 struct rpc_xprt *ret;
2918 unsigned int max_slot_table_size = xprt_max_tcp_slot_table_entries;
2919
2920 if (args->flags & XPRT_CREATE_INFINITE_SLOTS)
2921 max_slot_table_size = RPC_MAX_SLOT_TABLE_LIMIT;
2922
2923 xprt = xs_setup_xprt(args, xprt_tcp_slot_table_entries,
2924 max_slot_table_size);
2925 if (IS_ERR(xprt))
2926 return xprt;
2927 transport = container_of(xprt, struct sock_xprt, xprt);
2928
2929 xprt->prot = IPPROTO_TCP;
2930 xprt->tsh_size = sizeof(rpc_fraghdr) / sizeof(u32);
2931 xprt->max_payload = RPC_MAX_FRAGMENT_SIZE;
2932
2933 xprt->bind_timeout = XS_BIND_TO;
2934 xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
2935 xprt->idle_timeout = XS_IDLE_DISC_TO;
2936
2937 xprt->ops = &xs_tcp_ops;
2938 xprt->timeout = &xs_tcp_default_timeout;
2939
2940 INIT_WORK(&transport->recv_worker, xs_tcp_data_receive_workfn);
2941 INIT_DELAYED_WORK(&transport->connect_worker, xs_tcp_setup_socket);
2942
2943 switch (addr->sa_family) {
2944 case AF_INET:
2945 if (((struct sockaddr_in *)addr)->sin_port != htons(0))
2946 xprt_set_bound(xprt);
2947
2948 xs_format_peer_addresses(xprt, "tcp", RPCBIND_NETID_TCP);
2949 break;
2950 case AF_INET6:
2951 if (((struct sockaddr_in6 *)addr)->sin6_port != htons(0))
2952 xprt_set_bound(xprt);
2953
2954 xs_format_peer_addresses(xprt, "tcp", RPCBIND_NETID_TCP6);
2955 break;
2956 default:
2957 ret = ERR_PTR(-EAFNOSUPPORT);
2958 goto out_err;
2959 }
2960
2961 if (xprt_bound(xprt))
2962 dprintk("RPC: set up xprt to %s (port %s) via %s\n",
2963 xprt->address_strings[RPC_DISPLAY_ADDR],
2964 xprt->address_strings[RPC_DISPLAY_PORT],
2965 xprt->address_strings[RPC_DISPLAY_PROTO]);
2966 else
2967 dprintk("RPC: set up xprt to %s (autobind) via %s\n",
2968 xprt->address_strings[RPC_DISPLAY_ADDR],
2969 xprt->address_strings[RPC_DISPLAY_PROTO]);
2970
2971 if (try_module_get(THIS_MODULE))
2972 return xprt;
2973 ret = ERR_PTR(-EINVAL);
2974out_err:
2975 xs_xprt_free(xprt);
2976 return ret;
2977}
2978
2979/**
2980 * xs_setup_bc_tcp - Set up transport to use a TCP backchannel socket
2981 * @args: rpc transport creation arguments
2982 *
2983 */
2984static struct rpc_xprt *xs_setup_bc_tcp(struct xprt_create *args)
2985{
2986 struct sockaddr *addr = args->dstaddr;
2987 struct rpc_xprt *xprt;
2988 struct sock_xprt *transport;
2989 struct svc_sock *bc_sock;
2990 struct rpc_xprt *ret;
2991
2992 xprt = xs_setup_xprt(args, xprt_tcp_slot_table_entries,
2993 xprt_tcp_slot_table_entries);
2994 if (IS_ERR(xprt))
2995 return xprt;
2996 transport = container_of(xprt, struct sock_xprt, xprt);
2997
2998 xprt->prot = IPPROTO_TCP;
2999 xprt->tsh_size = sizeof(rpc_fraghdr) / sizeof(u32);
3000 xprt->max_payload = RPC_MAX_FRAGMENT_SIZE;
3001 xprt->timeout = &xs_tcp_default_timeout;
3002
3003 /* backchannel */
3004 xprt_set_bound(xprt);
3005 xprt->bind_timeout = 0;
3006 xprt->reestablish_timeout = 0;
3007 xprt->idle_timeout = 0;
3008
3009 xprt->ops = &bc_tcp_ops;
3010
3011 switch (addr->sa_family) {
3012 case AF_INET:
3013 xs_format_peer_addresses(xprt, "tcp",
3014 RPCBIND_NETID_TCP);
3015 break;
3016 case AF_INET6:
3017 xs_format_peer_addresses(xprt, "tcp",
3018 RPCBIND_NETID_TCP6);
3019 break;
3020 default:
3021 ret = ERR_PTR(-EAFNOSUPPORT);
3022 goto out_err;
3023 }
3024
3025 dprintk("RPC: set up xprt to %s (port %s) via %s\n",
3026 xprt->address_strings[RPC_DISPLAY_ADDR],
3027 xprt->address_strings[RPC_DISPLAY_PORT],
3028 xprt->address_strings[RPC_DISPLAY_PROTO]);
3029
3030 /*
3031 * Once we've associated a backchannel xprt with a connection,
3032 * we want to keep it around as long as the connection lasts,
3033 * in case we need to start using it for a backchannel again;
3034 * this reference won't be dropped until bc_xprt is destroyed.
3035 */
3036 xprt_get(xprt);
3037 args->bc_xprt->xpt_bc_xprt = xprt;
3038 xprt->bc_xprt = args->bc_xprt;
3039 bc_sock = container_of(args->bc_xprt, struct svc_sock, sk_xprt);
3040 transport->sock = bc_sock->sk_sock;
3041 transport->inet = bc_sock->sk_sk;
3042
3043 /*
3044 * Since we don't want connections for the backchannel, we set
3045 * the xprt status to connected
3046 */
3047 xprt_set_connected(xprt);
3048
3049 if (try_module_get(THIS_MODULE))
3050 return xprt;
3051
3052 args->bc_xprt->xpt_bc_xprt = NULL;
3053 xprt_put(xprt);
3054 ret = ERR_PTR(-EINVAL);
3055out_err:
3056 xs_xprt_free(xprt);
3057 return ret;
3058}
3059
3060static struct xprt_class xs_local_transport = {
3061 .list = LIST_HEAD_INIT(xs_local_transport.list),
3062 .name = "named UNIX socket",
3063 .owner = THIS_MODULE,
3064 .ident = XPRT_TRANSPORT_LOCAL,
3065 .setup = xs_setup_local,
3066};
3067
3068static struct xprt_class xs_udp_transport = {
3069 .list = LIST_HEAD_INIT(xs_udp_transport.list),
3070 .name = "udp",
3071 .owner = THIS_MODULE,
3072 .ident = XPRT_TRANSPORT_UDP,
3073 .setup = xs_setup_udp,
3074};
3075
3076static struct xprt_class xs_tcp_transport = {
3077 .list = LIST_HEAD_INIT(xs_tcp_transport.list),
3078 .name = "tcp",
3079 .owner = THIS_MODULE,
3080 .ident = XPRT_TRANSPORT_TCP,
3081 .setup = xs_setup_tcp,
3082};
3083
3084static struct xprt_class xs_bc_tcp_transport = {
3085 .list = LIST_HEAD_INIT(xs_bc_tcp_transport.list),
3086 .name = "tcp NFSv4.1 backchannel",
3087 .owner = THIS_MODULE,
3088 .ident = XPRT_TRANSPORT_BC_TCP,
3089 .setup = xs_setup_bc_tcp,
3090};
3091
3092/**
3093 * init_socket_xprt - set up xprtsock's sysctls, register with RPC client
3094 *
3095 */
3096int init_socket_xprt(void)
3097{
3098#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
3099 if (!sunrpc_table_header)
3100 sunrpc_table_header = register_sysctl_table(sunrpc_table);
3101#endif
3102
3103 xprt_register_transport(&xs_local_transport);
3104 xprt_register_transport(&xs_udp_transport);
3105 xprt_register_transport(&xs_tcp_transport);
3106 xprt_register_transport(&xs_bc_tcp_transport);
3107
3108 return 0;
3109}
3110
3111/**
3112 * cleanup_socket_xprt - remove xprtsock's sysctls, unregister
3113 *
3114 */
3115void cleanup_socket_xprt(void)
3116{
3117#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
3118 if (sunrpc_table_header) {
3119 unregister_sysctl_table(sunrpc_table_header);
3120 sunrpc_table_header = NULL;
3121 }
3122#endif
3123
3124 xprt_unregister_transport(&xs_local_transport);
3125 xprt_unregister_transport(&xs_udp_transport);
3126 xprt_unregister_transport(&xs_tcp_transport);
3127 xprt_unregister_transport(&xs_bc_tcp_transport);
3128}
3129
3130static int param_set_uint_minmax(const char *val,
3131 const struct kernel_param *kp,
3132 unsigned int min, unsigned int max)
3133{
3134 unsigned int num;
3135 int ret;
3136
3137 if (!val)
3138 return -EINVAL;
3139 ret = kstrtouint(val, 0, &num);
3140 if (ret == -EINVAL || num < min || num > max)
3141 return -EINVAL;
3142 *((unsigned int *)kp->arg) = num;
3143 return 0;
3144}
3145
3146static int param_set_portnr(const char *val, const struct kernel_param *kp)
3147{
3148 return param_set_uint_minmax(val, kp,
3149 RPC_MIN_RESVPORT,
3150 RPC_MAX_RESVPORT);
3151}
3152
3153static const struct kernel_param_ops param_ops_portnr = {
3154 .set = param_set_portnr,
3155 .get = param_get_uint,
3156};
3157
3158#define param_check_portnr(name, p) \
3159 __param_check(name, p, unsigned int);
3160
3161module_param_named(min_resvport, xprt_min_resvport, portnr, 0644);
3162module_param_named(max_resvport, xprt_max_resvport, portnr, 0644);
3163
3164static int param_set_slot_table_size(const char *val,
3165 const struct kernel_param *kp)
3166{
3167 return param_set_uint_minmax(val, kp,
3168 RPC_MIN_SLOT_TABLE,
3169 RPC_MAX_SLOT_TABLE);
3170}
3171
3172static const struct kernel_param_ops param_ops_slot_table_size = {
3173 .set = param_set_slot_table_size,
3174 .get = param_get_uint,
3175};
3176
3177#define param_check_slot_table_size(name, p) \
3178 __param_check(name, p, unsigned int);
3179
3180static int param_set_max_slot_table_size(const char *val,
3181 const struct kernel_param *kp)
3182{
3183 return param_set_uint_minmax(val, kp,
3184 RPC_MIN_SLOT_TABLE,
3185 RPC_MAX_SLOT_TABLE_LIMIT);
3186}
3187
3188static const struct kernel_param_ops param_ops_max_slot_table_size = {
3189 .set = param_set_max_slot_table_size,
3190 .get = param_get_uint,
3191};
3192
3193#define param_check_max_slot_table_size(name, p) \
3194 __param_check(name, p, unsigned int);
3195
3196module_param_named(tcp_slot_table_entries, xprt_tcp_slot_table_entries,
3197 slot_table_size, 0644);
3198module_param_named(tcp_max_slot_table_entries, xprt_max_tcp_slot_table_entries,
3199 max_slot_table_size, 0644);
3200module_param_named(udp_slot_table_entries, xprt_udp_slot_table_entries,
3201 slot_table_size, 0644);
3202
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
7 * TCP send fixes (C) 1998 Red Hat
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 * IPv6 support contributed by Gilles Quillard, Bull Open Source, 2005.
18 * <gilles.quillard@bull.net>
19 */
20
21#include <linux/types.h>
22#include <linux/string.h>
23#include <linux/slab.h>
24#include <linux/module.h>
25#include <linux/capability.h>
26#include <linux/pagemap.h>
27#include <linux/errno.h>
28#include <linux/socket.h>
29#include <linux/in.h>
30#include <linux/net.h>
31#include <linux/mm.h>
32#include <linux/un.h>
33#include <linux/udp.h>
34#include <linux/tcp.h>
35#include <linux/sunrpc/clnt.h>
36#include <linux/sunrpc/sched.h>
37#include <linux/sunrpc/svcsock.h>
38#include <linux/sunrpc/xprtsock.h>
39#include <linux/file.h>
40#ifdef CONFIG_SUNRPC_BACKCHANNEL
41#include <linux/sunrpc/bc_xprt.h>
42#endif
43
44#include <net/sock.h>
45#include <net/checksum.h>
46#include <net/udp.h>
47#include <net/tcp.h>
48
49#include "sunrpc.h"
50
51static void xs_close(struct rpc_xprt *xprt);
52
53/*
54 * xprtsock tunables
55 */
56static unsigned int xprt_udp_slot_table_entries = RPC_DEF_SLOT_TABLE;
57static unsigned int xprt_tcp_slot_table_entries = RPC_MIN_SLOT_TABLE;
58static unsigned int xprt_max_tcp_slot_table_entries = RPC_MAX_SLOT_TABLE;
59
60static unsigned int xprt_min_resvport = RPC_DEF_MIN_RESVPORT;
61static unsigned int xprt_max_resvport = RPC_DEF_MAX_RESVPORT;
62
63#define XS_TCP_LINGER_TO (15U * HZ)
64static unsigned int xs_tcp_fin_timeout __read_mostly = XS_TCP_LINGER_TO;
65
66/*
67 * We can register our own files under /proc/sys/sunrpc by
68 * calling register_sysctl_table() again. The files in that
69 * directory become the union of all files registered there.
70 *
71 * We simply need to make sure that we don't collide with
72 * someone else's file names!
73 */
74
75#ifdef RPC_DEBUG
76
77static unsigned int min_slot_table_size = RPC_MIN_SLOT_TABLE;
78static unsigned int max_slot_table_size = RPC_MAX_SLOT_TABLE;
79static unsigned int max_tcp_slot_table_limit = RPC_MAX_SLOT_TABLE_LIMIT;
80static unsigned int xprt_min_resvport_limit = RPC_MIN_RESVPORT;
81static unsigned int xprt_max_resvport_limit = RPC_MAX_RESVPORT;
82
83static struct ctl_table_header *sunrpc_table_header;
84
85/*
86 * FIXME: changing the UDP slot table size should also resize the UDP
87 * socket buffers for existing UDP transports
88 */
89static ctl_table xs_tunables_table[] = {
90 {
91 .procname = "udp_slot_table_entries",
92 .data = &xprt_udp_slot_table_entries,
93 .maxlen = sizeof(unsigned int),
94 .mode = 0644,
95 .proc_handler = proc_dointvec_minmax,
96 .extra1 = &min_slot_table_size,
97 .extra2 = &max_slot_table_size
98 },
99 {
100 .procname = "tcp_slot_table_entries",
101 .data = &xprt_tcp_slot_table_entries,
102 .maxlen = sizeof(unsigned int),
103 .mode = 0644,
104 .proc_handler = proc_dointvec_minmax,
105 .extra1 = &min_slot_table_size,
106 .extra2 = &max_slot_table_size
107 },
108 {
109 .procname = "tcp_max_slot_table_entries",
110 .data = &xprt_max_tcp_slot_table_entries,
111 .maxlen = sizeof(unsigned int),
112 .mode = 0644,
113 .proc_handler = proc_dointvec_minmax,
114 .extra1 = &min_slot_table_size,
115 .extra2 = &max_tcp_slot_table_limit
116 },
117 {
118 .procname = "min_resvport",
119 .data = &xprt_min_resvport,
120 .maxlen = sizeof(unsigned int),
121 .mode = 0644,
122 .proc_handler = proc_dointvec_minmax,
123 .extra1 = &xprt_min_resvport_limit,
124 .extra2 = &xprt_max_resvport_limit
125 },
126 {
127 .procname = "max_resvport",
128 .data = &xprt_max_resvport,
129 .maxlen = sizeof(unsigned int),
130 .mode = 0644,
131 .proc_handler = proc_dointvec_minmax,
132 .extra1 = &xprt_min_resvport_limit,
133 .extra2 = &xprt_max_resvport_limit
134 },
135 {
136 .procname = "tcp_fin_timeout",
137 .data = &xs_tcp_fin_timeout,
138 .maxlen = sizeof(xs_tcp_fin_timeout),
139 .mode = 0644,
140 .proc_handler = proc_dointvec_jiffies,
141 },
142 { },
143};
144
145static ctl_table sunrpc_table[] = {
146 {
147 .procname = "sunrpc",
148 .mode = 0555,
149 .child = xs_tunables_table
150 },
151 { },
152};
153
154#endif
155
156/*
157 * Wait duration for a reply from the RPC portmapper.
158 */
159#define XS_BIND_TO (60U * HZ)
160
161/*
162 * Delay if a UDP socket connect error occurs. This is most likely some
163 * kind of resource problem on the local host.
164 */
165#define XS_UDP_REEST_TO (2U * HZ)
166
167/*
168 * The reestablish timeout allows clients to delay for a bit before attempting
169 * to reconnect to a server that just dropped our connection.
170 *
171 * We implement an exponential backoff when trying to reestablish a TCP
172 * transport connection with the server. Some servers like to drop a TCP
173 * connection when they are overworked, so we start with a short timeout and
174 * increase over time if the server is down or not responding.
175 */
176#define XS_TCP_INIT_REEST_TO (3U * HZ)
177#define XS_TCP_MAX_REEST_TO (5U * 60 * HZ)
178
179/*
180 * TCP idle timeout; client drops the transport socket if it is idle
181 * for this long. Note that we also timeout UDP sockets to prevent
182 * holding port numbers when there is no RPC traffic.
183 */
184#define XS_IDLE_DISC_TO (5U * 60 * HZ)
185
186#ifdef RPC_DEBUG
187# undef RPC_DEBUG_DATA
188# define RPCDBG_FACILITY RPCDBG_TRANS
189#endif
190
191#ifdef RPC_DEBUG_DATA
192static void xs_pktdump(char *msg, u32 *packet, unsigned int count)
193{
194 u8 *buf = (u8 *) packet;
195 int j;
196
197 dprintk("RPC: %s\n", msg);
198 for (j = 0; j < count && j < 128; j += 4) {
199 if (!(j & 31)) {
200 if (j)
201 dprintk("\n");
202 dprintk("0x%04x ", j);
203 }
204 dprintk("%02x%02x%02x%02x ",
205 buf[j], buf[j+1], buf[j+2], buf[j+3]);
206 }
207 dprintk("\n");
208}
209#else
210static inline void xs_pktdump(char *msg, u32 *packet, unsigned int count)
211{
212 /* NOP */
213}
214#endif
215
216struct sock_xprt {
217 struct rpc_xprt xprt;
218
219 /*
220 * Network layer
221 */
222 struct socket * sock;
223 struct sock * inet;
224
225 /*
226 * State of TCP reply receive
227 */
228 __be32 tcp_fraghdr,
229 tcp_xid,
230 tcp_calldir;
231
232 u32 tcp_offset,
233 tcp_reclen;
234
235 unsigned long tcp_copied,
236 tcp_flags;
237
238 /*
239 * Connection of transports
240 */
241 struct delayed_work connect_worker;
242 struct sockaddr_storage srcaddr;
243 unsigned short srcport;
244
245 /*
246 * UDP socket buffer size parameters
247 */
248 size_t rcvsize,
249 sndsize;
250
251 /*
252 * Saved socket callback addresses
253 */
254 void (*old_data_ready)(struct sock *, int);
255 void (*old_state_change)(struct sock *);
256 void (*old_write_space)(struct sock *);
257 void (*old_error_report)(struct sock *);
258};
259
260/*
261 * TCP receive state flags
262 */
263#define TCP_RCV_LAST_FRAG (1UL << 0)
264#define TCP_RCV_COPY_FRAGHDR (1UL << 1)
265#define TCP_RCV_COPY_XID (1UL << 2)
266#define TCP_RCV_COPY_DATA (1UL << 3)
267#define TCP_RCV_READ_CALLDIR (1UL << 4)
268#define TCP_RCV_COPY_CALLDIR (1UL << 5)
269
270/*
271 * TCP RPC flags
272 */
273#define TCP_RPC_REPLY (1UL << 6)
274
275static inline struct sockaddr *xs_addr(struct rpc_xprt *xprt)
276{
277 return (struct sockaddr *) &xprt->addr;
278}
279
280static inline struct sockaddr_un *xs_addr_un(struct rpc_xprt *xprt)
281{
282 return (struct sockaddr_un *) &xprt->addr;
283}
284
285static inline struct sockaddr_in *xs_addr_in(struct rpc_xprt *xprt)
286{
287 return (struct sockaddr_in *) &xprt->addr;
288}
289
290static inline struct sockaddr_in6 *xs_addr_in6(struct rpc_xprt *xprt)
291{
292 return (struct sockaddr_in6 *) &xprt->addr;
293}
294
295static void xs_format_common_peer_addresses(struct rpc_xprt *xprt)
296{
297 struct sockaddr *sap = xs_addr(xprt);
298 struct sockaddr_in6 *sin6;
299 struct sockaddr_in *sin;
300 struct sockaddr_un *sun;
301 char buf[128];
302
303 switch (sap->sa_family) {
304 case AF_LOCAL:
305 sun = xs_addr_un(xprt);
306 strlcpy(buf, sun->sun_path, sizeof(buf));
307 xprt->address_strings[RPC_DISPLAY_ADDR] =
308 kstrdup(buf, GFP_KERNEL);
309 break;
310 case AF_INET:
311 (void)rpc_ntop(sap, buf, sizeof(buf));
312 xprt->address_strings[RPC_DISPLAY_ADDR] =
313 kstrdup(buf, GFP_KERNEL);
314 sin = xs_addr_in(xprt);
315 snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
316 break;
317 case AF_INET6:
318 (void)rpc_ntop(sap, buf, sizeof(buf));
319 xprt->address_strings[RPC_DISPLAY_ADDR] =
320 kstrdup(buf, GFP_KERNEL);
321 sin6 = xs_addr_in6(xprt);
322 snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
323 break;
324 default:
325 BUG();
326 }
327
328 xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
329}
330
331static void xs_format_common_peer_ports(struct rpc_xprt *xprt)
332{
333 struct sockaddr *sap = xs_addr(xprt);
334 char buf[128];
335
336 snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
337 xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
338
339 snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
340 xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
341}
342
343static void xs_format_peer_addresses(struct rpc_xprt *xprt,
344 const char *protocol,
345 const char *netid)
346{
347 xprt->address_strings[RPC_DISPLAY_PROTO] = protocol;
348 xprt->address_strings[RPC_DISPLAY_NETID] = netid;
349 xs_format_common_peer_addresses(xprt);
350 xs_format_common_peer_ports(xprt);
351}
352
353static void xs_update_peer_port(struct rpc_xprt *xprt)
354{
355 kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]);
356 kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
357
358 xs_format_common_peer_ports(xprt);
359}
360
361static void xs_free_peer_addresses(struct rpc_xprt *xprt)
362{
363 unsigned int i;
364
365 for (i = 0; i < RPC_DISPLAY_MAX; i++)
366 switch (i) {
367 case RPC_DISPLAY_PROTO:
368 case RPC_DISPLAY_NETID:
369 continue;
370 default:
371 kfree(xprt->address_strings[i]);
372 }
373}
374
375#define XS_SENDMSG_FLAGS (MSG_DONTWAIT | MSG_NOSIGNAL)
376
377static int xs_send_kvec(struct socket *sock, struct sockaddr *addr, int addrlen, struct kvec *vec, unsigned int base, int more)
378{
379 struct msghdr msg = {
380 .msg_name = addr,
381 .msg_namelen = addrlen,
382 .msg_flags = XS_SENDMSG_FLAGS | (more ? MSG_MORE : 0),
383 };
384 struct kvec iov = {
385 .iov_base = vec->iov_base + base,
386 .iov_len = vec->iov_len - base,
387 };
388
389 if (iov.iov_len != 0)
390 return kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
391 return kernel_sendmsg(sock, &msg, NULL, 0, 0);
392}
393
394static int xs_send_pagedata(struct socket *sock, struct xdr_buf *xdr, unsigned int base, int more)
395{
396 struct page **ppage;
397 unsigned int remainder;
398 int err, sent = 0;
399
400 remainder = xdr->page_len - base;
401 base += xdr->page_base;
402 ppage = xdr->pages + (base >> PAGE_SHIFT);
403 base &= ~PAGE_MASK;
404 for(;;) {
405 unsigned int len = min_t(unsigned int, PAGE_SIZE - base, remainder);
406 int flags = XS_SENDMSG_FLAGS;
407
408 remainder -= len;
409 if (remainder != 0 || more)
410 flags |= MSG_MORE;
411 err = sock->ops->sendpage(sock, *ppage, base, len, flags);
412 if (remainder == 0 || err != len)
413 break;
414 sent += err;
415 ppage++;
416 base = 0;
417 }
418 if (sent == 0)
419 return err;
420 if (err > 0)
421 sent += err;
422 return sent;
423}
424
425/**
426 * xs_sendpages - write pages directly to a socket
427 * @sock: socket to send on
428 * @addr: UDP only -- address of destination
429 * @addrlen: UDP only -- length of destination address
430 * @xdr: buffer containing this request
431 * @base: starting position in the buffer
432 *
433 */
434static int xs_sendpages(struct socket *sock, struct sockaddr *addr, int addrlen, struct xdr_buf *xdr, unsigned int base)
435{
436 unsigned int remainder = xdr->len - base;
437 int err, sent = 0;
438
439 if (unlikely(!sock))
440 return -ENOTSOCK;
441
442 clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
443 if (base != 0) {
444 addr = NULL;
445 addrlen = 0;
446 }
447
448 if (base < xdr->head[0].iov_len || addr != NULL) {
449 unsigned int len = xdr->head[0].iov_len - base;
450 remainder -= len;
451 err = xs_send_kvec(sock, addr, addrlen, &xdr->head[0], base, remainder != 0);
452 if (remainder == 0 || err != len)
453 goto out;
454 sent += err;
455 base = 0;
456 } else
457 base -= xdr->head[0].iov_len;
458
459 if (base < xdr->page_len) {
460 unsigned int len = xdr->page_len - base;
461 remainder -= len;
462 err = xs_send_pagedata(sock, xdr, base, remainder != 0);
463 if (remainder == 0 || err != len)
464 goto out;
465 sent += err;
466 base = 0;
467 } else
468 base -= xdr->page_len;
469
470 if (base >= xdr->tail[0].iov_len)
471 return sent;
472 err = xs_send_kvec(sock, NULL, 0, &xdr->tail[0], base, 0);
473out:
474 if (sent == 0)
475 return err;
476 if (err > 0)
477 sent += err;
478 return sent;
479}
480
481static void xs_nospace_callback(struct rpc_task *task)
482{
483 struct sock_xprt *transport = container_of(task->tk_rqstp->rq_xprt, struct sock_xprt, xprt);
484
485 transport->inet->sk_write_pending--;
486 clear_bit(SOCK_ASYNC_NOSPACE, &transport->sock->flags);
487}
488
489/**
490 * xs_nospace - place task on wait queue if transmit was incomplete
491 * @task: task to put to sleep
492 *
493 */
494static int xs_nospace(struct rpc_task *task)
495{
496 struct rpc_rqst *req = task->tk_rqstp;
497 struct rpc_xprt *xprt = req->rq_xprt;
498 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
499 int ret = -EAGAIN;
500
501 dprintk("RPC: %5u xmit incomplete (%u left of %u)\n",
502 task->tk_pid, req->rq_slen - req->rq_bytes_sent,
503 req->rq_slen);
504
505 /* Protect against races with write_space */
506 spin_lock_bh(&xprt->transport_lock);
507
508 /* Don't race with disconnect */
509 if (xprt_connected(xprt)) {
510 if (test_bit(SOCK_ASYNC_NOSPACE, &transport->sock->flags)) {
511 /*
512 * Notify TCP that we're limited by the application
513 * window size
514 */
515 set_bit(SOCK_NOSPACE, &transport->sock->flags);
516 transport->inet->sk_write_pending++;
517 /* ...and wait for more buffer space */
518 xprt_wait_for_buffer_space(task, xs_nospace_callback);
519 }
520 } else {
521 clear_bit(SOCK_ASYNC_NOSPACE, &transport->sock->flags);
522 ret = -ENOTCONN;
523 }
524
525 spin_unlock_bh(&xprt->transport_lock);
526 return ret;
527}
528
529/*
530 * Construct a stream transport record marker in @buf.
531 */
532static inline void xs_encode_stream_record_marker(struct xdr_buf *buf)
533{
534 u32 reclen = buf->len - sizeof(rpc_fraghdr);
535 rpc_fraghdr *base = buf->head[0].iov_base;
536 *base = cpu_to_be32(RPC_LAST_STREAM_FRAGMENT | reclen);
537}
538
539/**
540 * xs_local_send_request - write an RPC request to an AF_LOCAL socket
541 * @task: RPC task that manages the state of an RPC request
542 *
543 * Return values:
544 * 0: The request has been sent
545 * EAGAIN: The socket was blocked, please call again later to
546 * complete the request
547 * ENOTCONN: Caller needs to invoke connect logic then call again
548 * other: Some other error occured, the request was not sent
549 */
550static int xs_local_send_request(struct rpc_task *task)
551{
552 struct rpc_rqst *req = task->tk_rqstp;
553 struct rpc_xprt *xprt = req->rq_xprt;
554 struct sock_xprt *transport =
555 container_of(xprt, struct sock_xprt, xprt);
556 struct xdr_buf *xdr = &req->rq_snd_buf;
557 int status;
558
559 xs_encode_stream_record_marker(&req->rq_snd_buf);
560
561 xs_pktdump("packet data:",
562 req->rq_svec->iov_base, req->rq_svec->iov_len);
563
564 status = xs_sendpages(transport->sock, NULL, 0,
565 xdr, req->rq_bytes_sent);
566 dprintk("RPC: %s(%u) = %d\n",
567 __func__, xdr->len - req->rq_bytes_sent, status);
568 if (likely(status >= 0)) {
569 req->rq_bytes_sent += status;
570 req->rq_xmit_bytes_sent += status;
571 if (likely(req->rq_bytes_sent >= req->rq_slen)) {
572 req->rq_bytes_sent = 0;
573 return 0;
574 }
575 status = -EAGAIN;
576 }
577
578 switch (status) {
579 case -EAGAIN:
580 status = xs_nospace(task);
581 break;
582 default:
583 dprintk("RPC: sendmsg returned unrecognized error %d\n",
584 -status);
585 case -EPIPE:
586 xs_close(xprt);
587 status = -ENOTCONN;
588 }
589
590 return status;
591}
592
593/**
594 * xs_udp_send_request - write an RPC request to a UDP socket
595 * @task: address of RPC task that manages the state of an RPC request
596 *
597 * Return values:
598 * 0: The request has been sent
599 * EAGAIN: The socket was blocked, please call again later to
600 * complete the request
601 * ENOTCONN: Caller needs to invoke connect logic then call again
602 * other: Some other error occurred, the request was not sent
603 */
604static int xs_udp_send_request(struct rpc_task *task)
605{
606 struct rpc_rqst *req = task->tk_rqstp;
607 struct rpc_xprt *xprt = req->rq_xprt;
608 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
609 struct xdr_buf *xdr = &req->rq_snd_buf;
610 int status;
611
612 xs_pktdump("packet data:",
613 req->rq_svec->iov_base,
614 req->rq_svec->iov_len);
615
616 if (!xprt_bound(xprt))
617 return -ENOTCONN;
618 status = xs_sendpages(transport->sock,
619 xs_addr(xprt),
620 xprt->addrlen, xdr,
621 req->rq_bytes_sent);
622
623 dprintk("RPC: xs_udp_send_request(%u) = %d\n",
624 xdr->len - req->rq_bytes_sent, status);
625
626 if (status >= 0) {
627 req->rq_xmit_bytes_sent += status;
628 if (status >= req->rq_slen)
629 return 0;
630 /* Still some bytes left; set up for a retry later. */
631 status = -EAGAIN;
632 }
633
634 switch (status) {
635 case -ENOTSOCK:
636 status = -ENOTCONN;
637 /* Should we call xs_close() here? */
638 break;
639 case -EAGAIN:
640 status = xs_nospace(task);
641 break;
642 default:
643 dprintk("RPC: sendmsg returned unrecognized error %d\n",
644 -status);
645 case -ENETUNREACH:
646 case -EPIPE:
647 case -ECONNREFUSED:
648 /* When the server has died, an ICMP port unreachable message
649 * prompts ECONNREFUSED. */
650 clear_bit(SOCK_ASYNC_NOSPACE, &transport->sock->flags);
651 }
652
653 return status;
654}
655
656/**
657 * xs_tcp_shutdown - gracefully shut down a TCP socket
658 * @xprt: transport
659 *
660 * Initiates a graceful shutdown of the TCP socket by calling the
661 * equivalent of shutdown(SHUT_WR);
662 */
663static void xs_tcp_shutdown(struct rpc_xprt *xprt)
664{
665 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
666 struct socket *sock = transport->sock;
667
668 if (sock != NULL)
669 kernel_sock_shutdown(sock, SHUT_WR);
670}
671
672/**
673 * xs_tcp_send_request - write an RPC request to a TCP socket
674 * @task: address of RPC task that manages the state of an RPC request
675 *
676 * Return values:
677 * 0: The request has been sent
678 * EAGAIN: The socket was blocked, please call again later to
679 * complete the request
680 * ENOTCONN: Caller needs to invoke connect logic then call again
681 * other: Some other error occurred, the request was not sent
682 *
683 * XXX: In the case of soft timeouts, should we eventually give up
684 * if sendmsg is not able to make progress?
685 */
686static int xs_tcp_send_request(struct rpc_task *task)
687{
688 struct rpc_rqst *req = task->tk_rqstp;
689 struct rpc_xprt *xprt = req->rq_xprt;
690 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
691 struct xdr_buf *xdr = &req->rq_snd_buf;
692 int status;
693
694 xs_encode_stream_record_marker(&req->rq_snd_buf);
695
696 xs_pktdump("packet data:",
697 req->rq_svec->iov_base,
698 req->rq_svec->iov_len);
699
700 /* Continue transmitting the packet/record. We must be careful
701 * to cope with writespace callbacks arriving _after_ we have
702 * called sendmsg(). */
703 while (1) {
704 status = xs_sendpages(transport->sock,
705 NULL, 0, xdr, req->rq_bytes_sent);
706
707 dprintk("RPC: xs_tcp_send_request(%u) = %d\n",
708 xdr->len - req->rq_bytes_sent, status);
709
710 if (unlikely(status < 0))
711 break;
712
713 /* If we've sent the entire packet, immediately
714 * reset the count of bytes sent. */
715 req->rq_bytes_sent += status;
716 req->rq_xmit_bytes_sent += status;
717 if (likely(req->rq_bytes_sent >= req->rq_slen)) {
718 req->rq_bytes_sent = 0;
719 return 0;
720 }
721
722 if (status != 0)
723 continue;
724 status = -EAGAIN;
725 break;
726 }
727
728 switch (status) {
729 case -ENOTSOCK:
730 status = -ENOTCONN;
731 /* Should we call xs_close() here? */
732 break;
733 case -EAGAIN:
734 status = xs_nospace(task);
735 break;
736 default:
737 dprintk("RPC: sendmsg returned unrecognized error %d\n",
738 -status);
739 case -ECONNRESET:
740 case -EPIPE:
741 xs_tcp_shutdown(xprt);
742 case -ECONNREFUSED:
743 case -ENOTCONN:
744 clear_bit(SOCK_ASYNC_NOSPACE, &transport->sock->flags);
745 }
746
747 return status;
748}
749
750/**
751 * xs_tcp_release_xprt - clean up after a tcp transmission
752 * @xprt: transport
753 * @task: rpc task
754 *
755 * This cleans up if an error causes us to abort the transmission of a request.
756 * In this case, the socket may need to be reset in order to avoid confusing
757 * the server.
758 */
759static void xs_tcp_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
760{
761 struct rpc_rqst *req;
762
763 if (task != xprt->snd_task)
764 return;
765 if (task == NULL)
766 goto out_release;
767 req = task->tk_rqstp;
768 if (req == NULL)
769 goto out_release;
770 if (req->rq_bytes_sent == 0)
771 goto out_release;
772 if (req->rq_bytes_sent == req->rq_snd_buf.len)
773 goto out_release;
774 set_bit(XPRT_CLOSE_WAIT, &task->tk_xprt->state);
775out_release:
776 xprt_release_xprt(xprt, task);
777}
778
779static void xs_save_old_callbacks(struct sock_xprt *transport, struct sock *sk)
780{
781 transport->old_data_ready = sk->sk_data_ready;
782 transport->old_state_change = sk->sk_state_change;
783 transport->old_write_space = sk->sk_write_space;
784 transport->old_error_report = sk->sk_error_report;
785}
786
787static void xs_restore_old_callbacks(struct sock_xprt *transport, struct sock *sk)
788{
789 sk->sk_data_ready = transport->old_data_ready;
790 sk->sk_state_change = transport->old_state_change;
791 sk->sk_write_space = transport->old_write_space;
792 sk->sk_error_report = transport->old_error_report;
793}
794
795static void xs_reset_transport(struct sock_xprt *transport)
796{
797 struct socket *sock = transport->sock;
798 struct sock *sk = transport->inet;
799
800 if (sk == NULL)
801 return;
802
803 transport->srcport = 0;
804
805 write_lock_bh(&sk->sk_callback_lock);
806 transport->inet = NULL;
807 transport->sock = NULL;
808
809 sk->sk_user_data = NULL;
810
811 xs_restore_old_callbacks(transport, sk);
812 write_unlock_bh(&sk->sk_callback_lock);
813
814 sk->sk_no_check = 0;
815
816 sock_release(sock);
817}
818
819/**
820 * xs_close - close a socket
821 * @xprt: transport
822 *
823 * This is used when all requests are complete; ie, no DRC state remains
824 * on the server we want to save.
825 *
826 * The caller _must_ be holding XPRT_LOCKED in order to avoid issues with
827 * xs_reset_transport() zeroing the socket from underneath a writer.
828 */
829static void xs_close(struct rpc_xprt *xprt)
830{
831 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
832
833 dprintk("RPC: xs_close xprt %p\n", xprt);
834
835 xs_reset_transport(transport);
836 xprt->reestablish_timeout = 0;
837
838 smp_mb__before_clear_bit();
839 clear_bit(XPRT_CONNECTION_ABORT, &xprt->state);
840 clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
841 clear_bit(XPRT_CLOSING, &xprt->state);
842 smp_mb__after_clear_bit();
843 xprt_disconnect_done(xprt);
844}
845
846static void xs_tcp_close(struct rpc_xprt *xprt)
847{
848 if (test_and_clear_bit(XPRT_CONNECTION_CLOSE, &xprt->state))
849 xs_close(xprt);
850 else
851 xs_tcp_shutdown(xprt);
852}
853
854/**
855 * xs_destroy - prepare to shutdown a transport
856 * @xprt: doomed transport
857 *
858 */
859static void xs_destroy(struct rpc_xprt *xprt)
860{
861 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
862
863 dprintk("RPC: xs_destroy xprt %p\n", xprt);
864
865 cancel_delayed_work_sync(&transport->connect_worker);
866
867 xs_close(xprt);
868 xs_free_peer_addresses(xprt);
869 xprt_free(xprt);
870 module_put(THIS_MODULE);
871}
872
873static inline struct rpc_xprt *xprt_from_sock(struct sock *sk)
874{
875 return (struct rpc_xprt *) sk->sk_user_data;
876}
877
878static int xs_local_copy_to_xdr(struct xdr_buf *xdr, struct sk_buff *skb)
879{
880 struct xdr_skb_reader desc = {
881 .skb = skb,
882 .offset = sizeof(rpc_fraghdr),
883 .count = skb->len - sizeof(rpc_fraghdr),
884 };
885
886 if (xdr_partial_copy_from_skb(xdr, 0, &desc, xdr_skb_read_bits) < 0)
887 return -1;
888 if (desc.count)
889 return -1;
890 return 0;
891}
892
893/**
894 * xs_local_data_ready - "data ready" callback for AF_LOCAL sockets
895 * @sk: socket with data to read
896 * @len: how much data to read
897 *
898 * Currently this assumes we can read the whole reply in a single gulp.
899 */
900static void xs_local_data_ready(struct sock *sk, int len)
901{
902 struct rpc_task *task;
903 struct rpc_xprt *xprt;
904 struct rpc_rqst *rovr;
905 struct sk_buff *skb;
906 int err, repsize, copied;
907 u32 _xid;
908 __be32 *xp;
909
910 read_lock_bh(&sk->sk_callback_lock);
911 dprintk("RPC: %s...\n", __func__);
912 xprt = xprt_from_sock(sk);
913 if (xprt == NULL)
914 goto out;
915
916 skb = skb_recv_datagram(sk, 0, 1, &err);
917 if (skb == NULL)
918 goto out;
919
920 if (xprt->shutdown)
921 goto dropit;
922
923 repsize = skb->len - sizeof(rpc_fraghdr);
924 if (repsize < 4) {
925 dprintk("RPC: impossible RPC reply size %d\n", repsize);
926 goto dropit;
927 }
928
929 /* Copy the XID from the skb... */
930 xp = skb_header_pointer(skb, sizeof(rpc_fraghdr), sizeof(_xid), &_xid);
931 if (xp == NULL)
932 goto dropit;
933
934 /* Look up and lock the request corresponding to the given XID */
935 spin_lock(&xprt->transport_lock);
936 rovr = xprt_lookup_rqst(xprt, *xp);
937 if (!rovr)
938 goto out_unlock;
939 task = rovr->rq_task;
940
941 copied = rovr->rq_private_buf.buflen;
942 if (copied > repsize)
943 copied = repsize;
944
945 if (xs_local_copy_to_xdr(&rovr->rq_private_buf, skb)) {
946 dprintk("RPC: sk_buff copy failed\n");
947 goto out_unlock;
948 }
949
950 xprt_complete_rqst(task, copied);
951
952 out_unlock:
953 spin_unlock(&xprt->transport_lock);
954 dropit:
955 skb_free_datagram(sk, skb);
956 out:
957 read_unlock_bh(&sk->sk_callback_lock);
958}
959
960/**
961 * xs_udp_data_ready - "data ready" callback for UDP sockets
962 * @sk: socket with data to read
963 * @len: how much data to read
964 *
965 */
966static void xs_udp_data_ready(struct sock *sk, int len)
967{
968 struct rpc_task *task;
969 struct rpc_xprt *xprt;
970 struct rpc_rqst *rovr;
971 struct sk_buff *skb;
972 int err, repsize, copied;
973 u32 _xid;
974 __be32 *xp;
975
976 read_lock_bh(&sk->sk_callback_lock);
977 dprintk("RPC: xs_udp_data_ready...\n");
978 if (!(xprt = xprt_from_sock(sk)))
979 goto out;
980
981 if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
982 goto out;
983
984 if (xprt->shutdown)
985 goto dropit;
986
987 repsize = skb->len - sizeof(struct udphdr);
988 if (repsize < 4) {
989 dprintk("RPC: impossible RPC reply size %d!\n", repsize);
990 goto dropit;
991 }
992
993 /* Copy the XID from the skb... */
994 xp = skb_header_pointer(skb, sizeof(struct udphdr),
995 sizeof(_xid), &_xid);
996 if (xp == NULL)
997 goto dropit;
998
999 /* Look up and lock the request corresponding to the given XID */
1000 spin_lock(&xprt->transport_lock);
1001 rovr = xprt_lookup_rqst(xprt, *xp);
1002 if (!rovr)
1003 goto out_unlock;
1004 task = rovr->rq_task;
1005
1006 if ((copied = rovr->rq_private_buf.buflen) > repsize)
1007 copied = repsize;
1008
1009 /* Suck it into the iovec, verify checksum if not done by hw. */
1010 if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb)) {
1011 UDPX_INC_STATS_BH(sk, UDP_MIB_INERRORS);
1012 goto out_unlock;
1013 }
1014
1015 UDPX_INC_STATS_BH(sk, UDP_MIB_INDATAGRAMS);
1016
1017 /* Something worked... */
1018 dst_confirm(skb_dst(skb));
1019
1020 xprt_adjust_cwnd(task, copied);
1021 xprt_complete_rqst(task, copied);
1022
1023 out_unlock:
1024 spin_unlock(&xprt->transport_lock);
1025 dropit:
1026 skb_free_datagram(sk, skb);
1027 out:
1028 read_unlock_bh(&sk->sk_callback_lock);
1029}
1030
1031static inline void xs_tcp_read_fraghdr(struct rpc_xprt *xprt, struct xdr_skb_reader *desc)
1032{
1033 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1034 size_t len, used;
1035 char *p;
1036
1037 p = ((char *) &transport->tcp_fraghdr) + transport->tcp_offset;
1038 len = sizeof(transport->tcp_fraghdr) - transport->tcp_offset;
1039 used = xdr_skb_read_bits(desc, p, len);
1040 transport->tcp_offset += used;
1041 if (used != len)
1042 return;
1043
1044 transport->tcp_reclen = ntohl(transport->tcp_fraghdr);
1045 if (transport->tcp_reclen & RPC_LAST_STREAM_FRAGMENT)
1046 transport->tcp_flags |= TCP_RCV_LAST_FRAG;
1047 else
1048 transport->tcp_flags &= ~TCP_RCV_LAST_FRAG;
1049 transport->tcp_reclen &= RPC_FRAGMENT_SIZE_MASK;
1050
1051 transport->tcp_flags &= ~TCP_RCV_COPY_FRAGHDR;
1052 transport->tcp_offset = 0;
1053
1054 /* Sanity check of the record length */
1055 if (unlikely(transport->tcp_reclen < 8)) {
1056 dprintk("RPC: invalid TCP record fragment length\n");
1057 xprt_force_disconnect(xprt);
1058 return;
1059 }
1060 dprintk("RPC: reading TCP record fragment of length %d\n",
1061 transport->tcp_reclen);
1062}
1063
1064static void xs_tcp_check_fraghdr(struct sock_xprt *transport)
1065{
1066 if (transport->tcp_offset == transport->tcp_reclen) {
1067 transport->tcp_flags |= TCP_RCV_COPY_FRAGHDR;
1068 transport->tcp_offset = 0;
1069 if (transport->tcp_flags & TCP_RCV_LAST_FRAG) {
1070 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
1071 transport->tcp_flags |= TCP_RCV_COPY_XID;
1072 transport->tcp_copied = 0;
1073 }
1074 }
1075}
1076
1077static inline void xs_tcp_read_xid(struct sock_xprt *transport, struct xdr_skb_reader *desc)
1078{
1079 size_t len, used;
1080 char *p;
1081
1082 len = sizeof(transport->tcp_xid) - transport->tcp_offset;
1083 dprintk("RPC: reading XID (%Zu bytes)\n", len);
1084 p = ((char *) &transport->tcp_xid) + transport->tcp_offset;
1085 used = xdr_skb_read_bits(desc, p, len);
1086 transport->tcp_offset += used;
1087 if (used != len)
1088 return;
1089 transport->tcp_flags &= ~TCP_RCV_COPY_XID;
1090 transport->tcp_flags |= TCP_RCV_READ_CALLDIR;
1091 transport->tcp_copied = 4;
1092 dprintk("RPC: reading %s XID %08x\n",
1093 (transport->tcp_flags & TCP_RPC_REPLY) ? "reply for"
1094 : "request with",
1095 ntohl(transport->tcp_xid));
1096 xs_tcp_check_fraghdr(transport);
1097}
1098
1099static inline void xs_tcp_read_calldir(struct sock_xprt *transport,
1100 struct xdr_skb_reader *desc)
1101{
1102 size_t len, used;
1103 u32 offset;
1104 char *p;
1105
1106 /*
1107 * We want transport->tcp_offset to be 8 at the end of this routine
1108 * (4 bytes for the xid and 4 bytes for the call/reply flag).
1109 * When this function is called for the first time,
1110 * transport->tcp_offset is 4 (after having already read the xid).
1111 */
1112 offset = transport->tcp_offset - sizeof(transport->tcp_xid);
1113 len = sizeof(transport->tcp_calldir) - offset;
1114 dprintk("RPC: reading CALL/REPLY flag (%Zu bytes)\n", len);
1115 p = ((char *) &transport->tcp_calldir) + offset;
1116 used = xdr_skb_read_bits(desc, p, len);
1117 transport->tcp_offset += used;
1118 if (used != len)
1119 return;
1120 transport->tcp_flags &= ~TCP_RCV_READ_CALLDIR;
1121 /*
1122 * We don't yet have the XDR buffer, so we will write the calldir
1123 * out after we get the buffer from the 'struct rpc_rqst'
1124 */
1125 switch (ntohl(transport->tcp_calldir)) {
1126 case RPC_REPLY:
1127 transport->tcp_flags |= TCP_RCV_COPY_CALLDIR;
1128 transport->tcp_flags |= TCP_RCV_COPY_DATA;
1129 transport->tcp_flags |= TCP_RPC_REPLY;
1130 break;
1131 case RPC_CALL:
1132 transport->tcp_flags |= TCP_RCV_COPY_CALLDIR;
1133 transport->tcp_flags |= TCP_RCV_COPY_DATA;
1134 transport->tcp_flags &= ~TCP_RPC_REPLY;
1135 break;
1136 default:
1137 dprintk("RPC: invalid request message type\n");
1138 xprt_force_disconnect(&transport->xprt);
1139 }
1140 xs_tcp_check_fraghdr(transport);
1141}
1142
1143static inline void xs_tcp_read_common(struct rpc_xprt *xprt,
1144 struct xdr_skb_reader *desc,
1145 struct rpc_rqst *req)
1146{
1147 struct sock_xprt *transport =
1148 container_of(xprt, struct sock_xprt, xprt);
1149 struct xdr_buf *rcvbuf;
1150 size_t len;
1151 ssize_t r;
1152
1153 rcvbuf = &req->rq_private_buf;
1154
1155 if (transport->tcp_flags & TCP_RCV_COPY_CALLDIR) {
1156 /*
1157 * Save the RPC direction in the XDR buffer
1158 */
1159 memcpy(rcvbuf->head[0].iov_base + transport->tcp_copied,
1160 &transport->tcp_calldir,
1161 sizeof(transport->tcp_calldir));
1162 transport->tcp_copied += sizeof(transport->tcp_calldir);
1163 transport->tcp_flags &= ~TCP_RCV_COPY_CALLDIR;
1164 }
1165
1166 len = desc->count;
1167 if (len > transport->tcp_reclen - transport->tcp_offset) {
1168 struct xdr_skb_reader my_desc;
1169
1170 len = transport->tcp_reclen - transport->tcp_offset;
1171 memcpy(&my_desc, desc, sizeof(my_desc));
1172 my_desc.count = len;
1173 r = xdr_partial_copy_from_skb(rcvbuf, transport->tcp_copied,
1174 &my_desc, xdr_skb_read_bits);
1175 desc->count -= r;
1176 desc->offset += r;
1177 } else
1178 r = xdr_partial_copy_from_skb(rcvbuf, transport->tcp_copied,
1179 desc, xdr_skb_read_bits);
1180
1181 if (r > 0) {
1182 transport->tcp_copied += r;
1183 transport->tcp_offset += r;
1184 }
1185 if (r != len) {
1186 /* Error when copying to the receive buffer,
1187 * usually because we weren't able to allocate
1188 * additional buffer pages. All we can do now
1189 * is turn off TCP_RCV_COPY_DATA, so the request
1190 * will not receive any additional updates,
1191 * and time out.
1192 * Any remaining data from this record will
1193 * be discarded.
1194 */
1195 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
1196 dprintk("RPC: XID %08x truncated request\n",
1197 ntohl(transport->tcp_xid));
1198 dprintk("RPC: xprt = %p, tcp_copied = %lu, "
1199 "tcp_offset = %u, tcp_reclen = %u\n",
1200 xprt, transport->tcp_copied,
1201 transport->tcp_offset, transport->tcp_reclen);
1202 return;
1203 }
1204
1205 dprintk("RPC: XID %08x read %Zd bytes\n",
1206 ntohl(transport->tcp_xid), r);
1207 dprintk("RPC: xprt = %p, tcp_copied = %lu, tcp_offset = %u, "
1208 "tcp_reclen = %u\n", xprt, transport->tcp_copied,
1209 transport->tcp_offset, transport->tcp_reclen);
1210
1211 if (transport->tcp_copied == req->rq_private_buf.buflen)
1212 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
1213 else if (transport->tcp_offset == transport->tcp_reclen) {
1214 if (transport->tcp_flags & TCP_RCV_LAST_FRAG)
1215 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
1216 }
1217}
1218
1219/*
1220 * Finds the request corresponding to the RPC xid and invokes the common
1221 * tcp read code to read the data.
1222 */
1223static inline int xs_tcp_read_reply(struct rpc_xprt *xprt,
1224 struct xdr_skb_reader *desc)
1225{
1226 struct sock_xprt *transport =
1227 container_of(xprt, struct sock_xprt, xprt);
1228 struct rpc_rqst *req;
1229
1230 dprintk("RPC: read reply XID %08x\n", ntohl(transport->tcp_xid));
1231
1232 /* Find and lock the request corresponding to this xid */
1233 spin_lock(&xprt->transport_lock);
1234 req = xprt_lookup_rqst(xprt, transport->tcp_xid);
1235 if (!req) {
1236 dprintk("RPC: XID %08x request not found!\n",
1237 ntohl(transport->tcp_xid));
1238 spin_unlock(&xprt->transport_lock);
1239 return -1;
1240 }
1241
1242 xs_tcp_read_common(xprt, desc, req);
1243
1244 if (!(transport->tcp_flags & TCP_RCV_COPY_DATA))
1245 xprt_complete_rqst(req->rq_task, transport->tcp_copied);
1246
1247 spin_unlock(&xprt->transport_lock);
1248 return 0;
1249}
1250
1251#if defined(CONFIG_SUNRPC_BACKCHANNEL)
1252/*
1253 * Obtains an rpc_rqst previously allocated and invokes the common
1254 * tcp read code to read the data. The result is placed in the callback
1255 * queue.
1256 * If we're unable to obtain the rpc_rqst we schedule the closing of the
1257 * connection and return -1.
1258 */
1259static inline int xs_tcp_read_callback(struct rpc_xprt *xprt,
1260 struct xdr_skb_reader *desc)
1261{
1262 struct sock_xprt *transport =
1263 container_of(xprt, struct sock_xprt, xprt);
1264 struct rpc_rqst *req;
1265
1266 req = xprt_alloc_bc_request(xprt);
1267 if (req == NULL) {
1268 printk(KERN_WARNING "Callback slot table overflowed\n");
1269 xprt_force_disconnect(xprt);
1270 return -1;
1271 }
1272
1273 req->rq_xid = transport->tcp_xid;
1274 dprintk("RPC: read callback XID %08x\n", ntohl(req->rq_xid));
1275 xs_tcp_read_common(xprt, desc, req);
1276
1277 if (!(transport->tcp_flags & TCP_RCV_COPY_DATA)) {
1278 struct svc_serv *bc_serv = xprt->bc_serv;
1279
1280 /*
1281 * Add callback request to callback list. The callback
1282 * service sleeps on the sv_cb_waitq waiting for new
1283 * requests. Wake it up after adding enqueing the
1284 * request.
1285 */
1286 dprintk("RPC: add callback request to list\n");
1287 spin_lock(&bc_serv->sv_cb_lock);
1288 list_add(&req->rq_bc_list, &bc_serv->sv_cb_list);
1289 spin_unlock(&bc_serv->sv_cb_lock);
1290 wake_up(&bc_serv->sv_cb_waitq);
1291 }
1292
1293 req->rq_private_buf.len = transport->tcp_copied;
1294
1295 return 0;
1296}
1297
1298static inline int _xs_tcp_read_data(struct rpc_xprt *xprt,
1299 struct xdr_skb_reader *desc)
1300{
1301 struct sock_xprt *transport =
1302 container_of(xprt, struct sock_xprt, xprt);
1303
1304 return (transport->tcp_flags & TCP_RPC_REPLY) ?
1305 xs_tcp_read_reply(xprt, desc) :
1306 xs_tcp_read_callback(xprt, desc);
1307}
1308#else
1309static inline int _xs_tcp_read_data(struct rpc_xprt *xprt,
1310 struct xdr_skb_reader *desc)
1311{
1312 return xs_tcp_read_reply(xprt, desc);
1313}
1314#endif /* CONFIG_SUNRPC_BACKCHANNEL */
1315
1316/*
1317 * Read data off the transport. This can be either an RPC_CALL or an
1318 * RPC_REPLY. Relay the processing to helper functions.
1319 */
1320static void xs_tcp_read_data(struct rpc_xprt *xprt,
1321 struct xdr_skb_reader *desc)
1322{
1323 struct sock_xprt *transport =
1324 container_of(xprt, struct sock_xprt, xprt);
1325
1326 if (_xs_tcp_read_data(xprt, desc) == 0)
1327 xs_tcp_check_fraghdr(transport);
1328 else {
1329 /*
1330 * The transport_lock protects the request handling.
1331 * There's no need to hold it to update the tcp_flags.
1332 */
1333 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
1334 }
1335}
1336
1337static inline void xs_tcp_read_discard(struct sock_xprt *transport, struct xdr_skb_reader *desc)
1338{
1339 size_t len;
1340
1341 len = transport->tcp_reclen - transport->tcp_offset;
1342 if (len > desc->count)
1343 len = desc->count;
1344 desc->count -= len;
1345 desc->offset += len;
1346 transport->tcp_offset += len;
1347 dprintk("RPC: discarded %Zu bytes\n", len);
1348 xs_tcp_check_fraghdr(transport);
1349}
1350
1351static int xs_tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, unsigned int offset, size_t len)
1352{
1353 struct rpc_xprt *xprt = rd_desc->arg.data;
1354 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1355 struct xdr_skb_reader desc = {
1356 .skb = skb,
1357 .offset = offset,
1358 .count = len,
1359 };
1360
1361 dprintk("RPC: xs_tcp_data_recv started\n");
1362 do {
1363 /* Read in a new fragment marker if necessary */
1364 /* Can we ever really expect to get completely empty fragments? */
1365 if (transport->tcp_flags & TCP_RCV_COPY_FRAGHDR) {
1366 xs_tcp_read_fraghdr(xprt, &desc);
1367 continue;
1368 }
1369 /* Read in the xid if necessary */
1370 if (transport->tcp_flags & TCP_RCV_COPY_XID) {
1371 xs_tcp_read_xid(transport, &desc);
1372 continue;
1373 }
1374 /* Read in the call/reply flag */
1375 if (transport->tcp_flags & TCP_RCV_READ_CALLDIR) {
1376 xs_tcp_read_calldir(transport, &desc);
1377 continue;
1378 }
1379 /* Read in the request data */
1380 if (transport->tcp_flags & TCP_RCV_COPY_DATA) {
1381 xs_tcp_read_data(xprt, &desc);
1382 continue;
1383 }
1384 /* Skip over any trailing bytes on short reads */
1385 xs_tcp_read_discard(transport, &desc);
1386 } while (desc.count);
1387 dprintk("RPC: xs_tcp_data_recv done\n");
1388 return len - desc.count;
1389}
1390
1391/**
1392 * xs_tcp_data_ready - "data ready" callback for TCP sockets
1393 * @sk: socket with data to read
1394 * @bytes: how much data to read
1395 *
1396 */
1397static void xs_tcp_data_ready(struct sock *sk, int bytes)
1398{
1399 struct rpc_xprt *xprt;
1400 read_descriptor_t rd_desc;
1401 int read;
1402
1403 dprintk("RPC: xs_tcp_data_ready...\n");
1404
1405 read_lock_bh(&sk->sk_callback_lock);
1406 if (!(xprt = xprt_from_sock(sk)))
1407 goto out;
1408 if (xprt->shutdown)
1409 goto out;
1410
1411 /* Any data means we had a useful conversation, so
1412 * the we don't need to delay the next reconnect
1413 */
1414 if (xprt->reestablish_timeout)
1415 xprt->reestablish_timeout = 0;
1416
1417 /* We use rd_desc to pass struct xprt to xs_tcp_data_recv */
1418 rd_desc.arg.data = xprt;
1419 do {
1420 rd_desc.count = 65536;
1421 read = tcp_read_sock(sk, &rd_desc, xs_tcp_data_recv);
1422 } while (read > 0);
1423out:
1424 read_unlock_bh(&sk->sk_callback_lock);
1425}
1426
1427/*
1428 * Do the equivalent of linger/linger2 handling for dealing with
1429 * broken servers that don't close the socket in a timely
1430 * fashion
1431 */
1432static void xs_tcp_schedule_linger_timeout(struct rpc_xprt *xprt,
1433 unsigned long timeout)
1434{
1435 struct sock_xprt *transport;
1436
1437 if (xprt_test_and_set_connecting(xprt))
1438 return;
1439 set_bit(XPRT_CONNECTION_ABORT, &xprt->state);
1440 transport = container_of(xprt, struct sock_xprt, xprt);
1441 queue_delayed_work(rpciod_workqueue, &transport->connect_worker,
1442 timeout);
1443}
1444
1445static void xs_tcp_cancel_linger_timeout(struct rpc_xprt *xprt)
1446{
1447 struct sock_xprt *transport;
1448
1449 transport = container_of(xprt, struct sock_xprt, xprt);
1450
1451 if (!test_bit(XPRT_CONNECTION_ABORT, &xprt->state) ||
1452 !cancel_delayed_work(&transport->connect_worker))
1453 return;
1454 clear_bit(XPRT_CONNECTION_ABORT, &xprt->state);
1455 xprt_clear_connecting(xprt);
1456}
1457
1458static void xs_sock_mark_closed(struct rpc_xprt *xprt)
1459{
1460 smp_mb__before_clear_bit();
1461 clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
1462 clear_bit(XPRT_CLOSING, &xprt->state);
1463 smp_mb__after_clear_bit();
1464 /* Mark transport as closed and wake up all pending tasks */
1465 xprt_disconnect_done(xprt);
1466}
1467
1468/**
1469 * xs_tcp_state_change - callback to handle TCP socket state changes
1470 * @sk: socket whose state has changed
1471 *
1472 */
1473static void xs_tcp_state_change(struct sock *sk)
1474{
1475 struct rpc_xprt *xprt;
1476
1477 read_lock_bh(&sk->sk_callback_lock);
1478 if (!(xprt = xprt_from_sock(sk)))
1479 goto out;
1480 dprintk("RPC: xs_tcp_state_change client %p...\n", xprt);
1481 dprintk("RPC: state %x conn %d dead %d zapped %d sk_shutdown %d\n",
1482 sk->sk_state, xprt_connected(xprt),
1483 sock_flag(sk, SOCK_DEAD),
1484 sock_flag(sk, SOCK_ZAPPED),
1485 sk->sk_shutdown);
1486
1487 switch (sk->sk_state) {
1488 case TCP_ESTABLISHED:
1489 spin_lock(&xprt->transport_lock);
1490 if (!xprt_test_and_set_connected(xprt)) {
1491 struct sock_xprt *transport = container_of(xprt,
1492 struct sock_xprt, xprt);
1493
1494 /* Reset TCP record info */
1495 transport->tcp_offset = 0;
1496 transport->tcp_reclen = 0;
1497 transport->tcp_copied = 0;
1498 transport->tcp_flags =
1499 TCP_RCV_COPY_FRAGHDR | TCP_RCV_COPY_XID;
1500
1501 xprt_wake_pending_tasks(xprt, -EAGAIN);
1502 }
1503 spin_unlock(&xprt->transport_lock);
1504 break;
1505 case TCP_FIN_WAIT1:
1506 /* The client initiated a shutdown of the socket */
1507 xprt->connect_cookie++;
1508 xprt->reestablish_timeout = 0;
1509 set_bit(XPRT_CLOSING, &xprt->state);
1510 smp_mb__before_clear_bit();
1511 clear_bit(XPRT_CONNECTED, &xprt->state);
1512 clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
1513 smp_mb__after_clear_bit();
1514 xs_tcp_schedule_linger_timeout(xprt, xs_tcp_fin_timeout);
1515 break;
1516 case TCP_CLOSE_WAIT:
1517 /* The server initiated a shutdown of the socket */
1518 xprt_force_disconnect(xprt);
1519 xprt->connect_cookie++;
1520 case TCP_CLOSING:
1521 /*
1522 * If the server closed down the connection, make sure that
1523 * we back off before reconnecting
1524 */
1525 if (xprt->reestablish_timeout < XS_TCP_INIT_REEST_TO)
1526 xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
1527 break;
1528 case TCP_LAST_ACK:
1529 set_bit(XPRT_CLOSING, &xprt->state);
1530 xs_tcp_schedule_linger_timeout(xprt, xs_tcp_fin_timeout);
1531 smp_mb__before_clear_bit();
1532 clear_bit(XPRT_CONNECTED, &xprt->state);
1533 smp_mb__after_clear_bit();
1534 break;
1535 case TCP_CLOSE:
1536 xs_tcp_cancel_linger_timeout(xprt);
1537 xs_sock_mark_closed(xprt);
1538 }
1539 out:
1540 read_unlock_bh(&sk->sk_callback_lock);
1541}
1542
1543/**
1544 * xs_error_report - callback mainly for catching socket errors
1545 * @sk: socket
1546 */
1547static void xs_error_report(struct sock *sk)
1548{
1549 struct rpc_xprt *xprt;
1550
1551 read_lock_bh(&sk->sk_callback_lock);
1552 if (!(xprt = xprt_from_sock(sk)))
1553 goto out;
1554 dprintk("RPC: %s client %p...\n"
1555 "RPC: error %d\n",
1556 __func__, xprt, sk->sk_err);
1557 xprt_wake_pending_tasks(xprt, -EAGAIN);
1558out:
1559 read_unlock_bh(&sk->sk_callback_lock);
1560}
1561
1562static void xs_write_space(struct sock *sk)
1563{
1564 struct socket *sock;
1565 struct rpc_xprt *xprt;
1566
1567 if (unlikely(!(sock = sk->sk_socket)))
1568 return;
1569 clear_bit(SOCK_NOSPACE, &sock->flags);
1570
1571 if (unlikely(!(xprt = xprt_from_sock(sk))))
1572 return;
1573 if (test_and_clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags) == 0)
1574 return;
1575
1576 xprt_write_space(xprt);
1577}
1578
1579/**
1580 * xs_udp_write_space - callback invoked when socket buffer space
1581 * becomes available
1582 * @sk: socket whose state has changed
1583 *
1584 * Called when more output buffer space is available for this socket.
1585 * We try not to wake our writers until they can make "significant"
1586 * progress, otherwise we'll waste resources thrashing kernel_sendmsg
1587 * with a bunch of small requests.
1588 */
1589static void xs_udp_write_space(struct sock *sk)
1590{
1591 read_lock_bh(&sk->sk_callback_lock);
1592
1593 /* from net/core/sock.c:sock_def_write_space */
1594 if (sock_writeable(sk))
1595 xs_write_space(sk);
1596
1597 read_unlock_bh(&sk->sk_callback_lock);
1598}
1599
1600/**
1601 * xs_tcp_write_space - callback invoked when socket buffer space
1602 * becomes available
1603 * @sk: socket whose state has changed
1604 *
1605 * Called when more output buffer space is available for this socket.
1606 * We try not to wake our writers until they can make "significant"
1607 * progress, otherwise we'll waste resources thrashing kernel_sendmsg
1608 * with a bunch of small requests.
1609 */
1610static void xs_tcp_write_space(struct sock *sk)
1611{
1612 read_lock_bh(&sk->sk_callback_lock);
1613
1614 /* from net/core/stream.c:sk_stream_write_space */
1615 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
1616 xs_write_space(sk);
1617
1618 read_unlock_bh(&sk->sk_callback_lock);
1619}
1620
1621static void xs_udp_do_set_buffer_size(struct rpc_xprt *xprt)
1622{
1623 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1624 struct sock *sk = transport->inet;
1625
1626 if (transport->rcvsize) {
1627 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
1628 sk->sk_rcvbuf = transport->rcvsize * xprt->max_reqs * 2;
1629 }
1630 if (transport->sndsize) {
1631 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
1632 sk->sk_sndbuf = transport->sndsize * xprt->max_reqs * 2;
1633 sk->sk_write_space(sk);
1634 }
1635}
1636
1637/**
1638 * xs_udp_set_buffer_size - set send and receive limits
1639 * @xprt: generic transport
1640 * @sndsize: requested size of send buffer, in bytes
1641 * @rcvsize: requested size of receive buffer, in bytes
1642 *
1643 * Set socket send and receive buffer size limits.
1644 */
1645static void xs_udp_set_buffer_size(struct rpc_xprt *xprt, size_t sndsize, size_t rcvsize)
1646{
1647 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1648
1649 transport->sndsize = 0;
1650 if (sndsize)
1651 transport->sndsize = sndsize + 1024;
1652 transport->rcvsize = 0;
1653 if (rcvsize)
1654 transport->rcvsize = rcvsize + 1024;
1655
1656 xs_udp_do_set_buffer_size(xprt);
1657}
1658
1659/**
1660 * xs_udp_timer - called when a retransmit timeout occurs on a UDP transport
1661 * @task: task that timed out
1662 *
1663 * Adjust the congestion window after a retransmit timeout has occurred.
1664 */
1665static void xs_udp_timer(struct rpc_task *task)
1666{
1667 xprt_adjust_cwnd(task, -ETIMEDOUT);
1668}
1669
1670static unsigned short xs_get_random_port(void)
1671{
1672 unsigned short range = xprt_max_resvport - xprt_min_resvport;
1673 unsigned short rand = (unsigned short) net_random() % range;
1674 return rand + xprt_min_resvport;
1675}
1676
1677/**
1678 * xs_set_port - reset the port number in the remote endpoint address
1679 * @xprt: generic transport
1680 * @port: new port number
1681 *
1682 */
1683static void xs_set_port(struct rpc_xprt *xprt, unsigned short port)
1684{
1685 dprintk("RPC: setting port for xprt %p to %u\n", xprt, port);
1686
1687 rpc_set_port(xs_addr(xprt), port);
1688 xs_update_peer_port(xprt);
1689}
1690
1691static unsigned short xs_get_srcport(struct sock_xprt *transport)
1692{
1693 unsigned short port = transport->srcport;
1694
1695 if (port == 0 && transport->xprt.resvport)
1696 port = xs_get_random_port();
1697 return port;
1698}
1699
1700static unsigned short xs_next_srcport(struct sock_xprt *transport, unsigned short port)
1701{
1702 if (transport->srcport != 0)
1703 transport->srcport = 0;
1704 if (!transport->xprt.resvport)
1705 return 0;
1706 if (port <= xprt_min_resvport || port > xprt_max_resvport)
1707 return xprt_max_resvport;
1708 return --port;
1709}
1710static int xs_bind(struct sock_xprt *transport, struct socket *sock)
1711{
1712 struct sockaddr_storage myaddr;
1713 int err, nloop = 0;
1714 unsigned short port = xs_get_srcport(transport);
1715 unsigned short last;
1716
1717 memcpy(&myaddr, &transport->srcaddr, transport->xprt.addrlen);
1718 do {
1719 rpc_set_port((struct sockaddr *)&myaddr, port);
1720 err = kernel_bind(sock, (struct sockaddr *)&myaddr,
1721 transport->xprt.addrlen);
1722 if (port == 0)
1723 break;
1724 if (err == 0) {
1725 transport->srcport = port;
1726 break;
1727 }
1728 last = port;
1729 port = xs_next_srcport(transport, port);
1730 if (port > last)
1731 nloop++;
1732 } while (err == -EADDRINUSE && nloop != 2);
1733
1734 if (myaddr.ss_family == AF_INET)
1735 dprintk("RPC: %s %pI4:%u: %s (%d)\n", __func__,
1736 &((struct sockaddr_in *)&myaddr)->sin_addr,
1737 port, err ? "failed" : "ok", err);
1738 else
1739 dprintk("RPC: %s %pI6:%u: %s (%d)\n", __func__,
1740 &((struct sockaddr_in6 *)&myaddr)->sin6_addr,
1741 port, err ? "failed" : "ok", err);
1742 return err;
1743}
1744
1745/*
1746 * We don't support autobind on AF_LOCAL sockets
1747 */
1748static void xs_local_rpcbind(struct rpc_task *task)
1749{
1750 xprt_set_bound(task->tk_xprt);
1751}
1752
1753static void xs_local_set_port(struct rpc_xprt *xprt, unsigned short port)
1754{
1755}
1756
1757#ifdef CONFIG_DEBUG_LOCK_ALLOC
1758static struct lock_class_key xs_key[2];
1759static struct lock_class_key xs_slock_key[2];
1760
1761static inline void xs_reclassify_socketu(struct socket *sock)
1762{
1763 struct sock *sk = sock->sk;
1764
1765 BUG_ON(sock_owned_by_user(sk));
1766 sock_lock_init_class_and_name(sk, "slock-AF_LOCAL-RPC",
1767 &xs_slock_key[1], "sk_lock-AF_LOCAL-RPC", &xs_key[1]);
1768}
1769
1770static inline void xs_reclassify_socket4(struct socket *sock)
1771{
1772 struct sock *sk = sock->sk;
1773
1774 BUG_ON(sock_owned_by_user(sk));
1775 sock_lock_init_class_and_name(sk, "slock-AF_INET-RPC",
1776 &xs_slock_key[0], "sk_lock-AF_INET-RPC", &xs_key[0]);
1777}
1778
1779static inline void xs_reclassify_socket6(struct socket *sock)
1780{
1781 struct sock *sk = sock->sk;
1782
1783 BUG_ON(sock_owned_by_user(sk));
1784 sock_lock_init_class_and_name(sk, "slock-AF_INET6-RPC",
1785 &xs_slock_key[1], "sk_lock-AF_INET6-RPC", &xs_key[1]);
1786}
1787
1788static inline void xs_reclassify_socket(int family, struct socket *sock)
1789{
1790 switch (family) {
1791 case AF_LOCAL:
1792 xs_reclassify_socketu(sock);
1793 break;
1794 case AF_INET:
1795 xs_reclassify_socket4(sock);
1796 break;
1797 case AF_INET6:
1798 xs_reclassify_socket6(sock);
1799 break;
1800 }
1801}
1802#else
1803static inline void xs_reclassify_socketu(struct socket *sock)
1804{
1805}
1806
1807static inline void xs_reclassify_socket4(struct socket *sock)
1808{
1809}
1810
1811static inline void xs_reclassify_socket6(struct socket *sock)
1812{
1813}
1814
1815static inline void xs_reclassify_socket(int family, struct socket *sock)
1816{
1817}
1818#endif
1819
1820static struct socket *xs_create_sock(struct rpc_xprt *xprt,
1821 struct sock_xprt *transport, int family, int type, int protocol)
1822{
1823 struct socket *sock;
1824 int err;
1825
1826 err = __sock_create(xprt->xprt_net, family, type, protocol, &sock, 1);
1827 if (err < 0) {
1828 dprintk("RPC: can't create %d transport socket (%d).\n",
1829 protocol, -err);
1830 goto out;
1831 }
1832 xs_reclassify_socket(family, sock);
1833
1834 err = xs_bind(transport, sock);
1835 if (err) {
1836 sock_release(sock);
1837 goto out;
1838 }
1839
1840 return sock;
1841out:
1842 return ERR_PTR(err);
1843}
1844
1845static int xs_local_finish_connecting(struct rpc_xprt *xprt,
1846 struct socket *sock)
1847{
1848 struct sock_xprt *transport = container_of(xprt, struct sock_xprt,
1849 xprt);
1850
1851 if (!transport->inet) {
1852 struct sock *sk = sock->sk;
1853
1854 write_lock_bh(&sk->sk_callback_lock);
1855
1856 xs_save_old_callbacks(transport, sk);
1857
1858 sk->sk_user_data = xprt;
1859 sk->sk_data_ready = xs_local_data_ready;
1860 sk->sk_write_space = xs_udp_write_space;
1861 sk->sk_error_report = xs_error_report;
1862 sk->sk_allocation = GFP_ATOMIC;
1863
1864 xprt_clear_connected(xprt);
1865
1866 /* Reset to new socket */
1867 transport->sock = sock;
1868 transport->inet = sk;
1869
1870 write_unlock_bh(&sk->sk_callback_lock);
1871 }
1872
1873 /* Tell the socket layer to start connecting... */
1874 xprt->stat.connect_count++;
1875 xprt->stat.connect_start = jiffies;
1876 return kernel_connect(sock, xs_addr(xprt), xprt->addrlen, 0);
1877}
1878
1879/**
1880 * xs_local_setup_socket - create AF_LOCAL socket, connect to a local endpoint
1881 * @xprt: RPC transport to connect
1882 * @transport: socket transport to connect
1883 * @create_sock: function to create a socket of the correct type
1884 *
1885 * Invoked by a work queue tasklet.
1886 */
1887static void xs_local_setup_socket(struct work_struct *work)
1888{
1889 struct sock_xprt *transport =
1890 container_of(work, struct sock_xprt, connect_worker.work);
1891 struct rpc_xprt *xprt = &transport->xprt;
1892 struct socket *sock;
1893 int status = -EIO;
1894
1895 if (xprt->shutdown)
1896 goto out;
1897
1898 current->flags |= PF_FSTRANS;
1899
1900 clear_bit(XPRT_CONNECTION_ABORT, &xprt->state);
1901 status = __sock_create(xprt->xprt_net, AF_LOCAL,
1902 SOCK_STREAM, 0, &sock, 1);
1903 if (status < 0) {
1904 dprintk("RPC: can't create AF_LOCAL "
1905 "transport socket (%d).\n", -status);
1906 goto out;
1907 }
1908 xs_reclassify_socketu(sock);
1909
1910 dprintk("RPC: worker connecting xprt %p via AF_LOCAL to %s\n",
1911 xprt, xprt->address_strings[RPC_DISPLAY_ADDR]);
1912
1913 status = xs_local_finish_connecting(xprt, sock);
1914 switch (status) {
1915 case 0:
1916 dprintk("RPC: xprt %p connected to %s\n",
1917 xprt, xprt->address_strings[RPC_DISPLAY_ADDR]);
1918 xprt_set_connected(xprt);
1919 break;
1920 case -ENOENT:
1921 dprintk("RPC: xprt %p: socket %s does not exist\n",
1922 xprt, xprt->address_strings[RPC_DISPLAY_ADDR]);
1923 break;
1924 default:
1925 printk(KERN_ERR "%s: unhandled error (%d) connecting to %s\n",
1926 __func__, -status,
1927 xprt->address_strings[RPC_DISPLAY_ADDR]);
1928 }
1929
1930out:
1931 xprt_clear_connecting(xprt);
1932 xprt_wake_pending_tasks(xprt, status);
1933 current->flags &= ~PF_FSTRANS;
1934}
1935
1936static void xs_udp_finish_connecting(struct rpc_xprt *xprt, struct socket *sock)
1937{
1938 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1939
1940 if (!transport->inet) {
1941 struct sock *sk = sock->sk;
1942
1943 write_lock_bh(&sk->sk_callback_lock);
1944
1945 xs_save_old_callbacks(transport, sk);
1946
1947 sk->sk_user_data = xprt;
1948 sk->sk_data_ready = xs_udp_data_ready;
1949 sk->sk_write_space = xs_udp_write_space;
1950 sk->sk_error_report = xs_error_report;
1951 sk->sk_no_check = UDP_CSUM_NORCV;
1952 sk->sk_allocation = GFP_ATOMIC;
1953
1954 xprt_set_connected(xprt);
1955
1956 /* Reset to new socket */
1957 transport->sock = sock;
1958 transport->inet = sk;
1959
1960 write_unlock_bh(&sk->sk_callback_lock);
1961 }
1962 xs_udp_do_set_buffer_size(xprt);
1963}
1964
1965static void xs_udp_setup_socket(struct work_struct *work)
1966{
1967 struct sock_xprt *transport =
1968 container_of(work, struct sock_xprt, connect_worker.work);
1969 struct rpc_xprt *xprt = &transport->xprt;
1970 struct socket *sock = transport->sock;
1971 int status = -EIO;
1972
1973 if (xprt->shutdown)
1974 goto out;
1975
1976 current->flags |= PF_FSTRANS;
1977
1978 /* Start by resetting any existing state */
1979 xs_reset_transport(transport);
1980 sock = xs_create_sock(xprt, transport,
1981 xs_addr(xprt)->sa_family, SOCK_DGRAM, IPPROTO_UDP);
1982 if (IS_ERR(sock))
1983 goto out;
1984
1985 dprintk("RPC: worker connecting xprt %p via %s to "
1986 "%s (port %s)\n", xprt,
1987 xprt->address_strings[RPC_DISPLAY_PROTO],
1988 xprt->address_strings[RPC_DISPLAY_ADDR],
1989 xprt->address_strings[RPC_DISPLAY_PORT]);
1990
1991 xs_udp_finish_connecting(xprt, sock);
1992 status = 0;
1993out:
1994 xprt_clear_connecting(xprt);
1995 xprt_wake_pending_tasks(xprt, status);
1996 current->flags &= ~PF_FSTRANS;
1997}
1998
1999/*
2000 * We need to preserve the port number so the reply cache on the server can
2001 * find our cached RPC replies when we get around to reconnecting.
2002 */
2003static void xs_abort_connection(struct sock_xprt *transport)
2004{
2005 int result;
2006 struct sockaddr any;
2007
2008 dprintk("RPC: disconnecting xprt %p to reuse port\n", transport);
2009
2010 /*
2011 * Disconnect the transport socket by doing a connect operation
2012 * with AF_UNSPEC. This should return immediately...
2013 */
2014 memset(&any, 0, sizeof(any));
2015 any.sa_family = AF_UNSPEC;
2016 result = kernel_connect(transport->sock, &any, sizeof(any), 0);
2017 if (!result)
2018 xs_sock_mark_closed(&transport->xprt);
2019 else
2020 dprintk("RPC: AF_UNSPEC connect return code %d\n",
2021 result);
2022}
2023
2024static void xs_tcp_reuse_connection(struct sock_xprt *transport)
2025{
2026 unsigned int state = transport->inet->sk_state;
2027
2028 if (state == TCP_CLOSE && transport->sock->state == SS_UNCONNECTED) {
2029 /* we don't need to abort the connection if the socket
2030 * hasn't undergone a shutdown
2031 */
2032 if (transport->inet->sk_shutdown == 0)
2033 return;
2034 dprintk("RPC: %s: TCP_CLOSEd and sk_shutdown set to %d\n",
2035 __func__, transport->inet->sk_shutdown);
2036 }
2037 if ((1 << state) & (TCPF_ESTABLISHED|TCPF_SYN_SENT)) {
2038 /* we don't need to abort the connection if the socket
2039 * hasn't undergone a shutdown
2040 */
2041 if (transport->inet->sk_shutdown == 0)
2042 return;
2043 dprintk("RPC: %s: ESTABLISHED/SYN_SENT "
2044 "sk_shutdown set to %d\n",
2045 __func__, transport->inet->sk_shutdown);
2046 }
2047 xs_abort_connection(transport);
2048}
2049
2050static int xs_tcp_finish_connecting(struct rpc_xprt *xprt, struct socket *sock)
2051{
2052 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
2053 int ret = -ENOTCONN;
2054
2055 if (!transport->inet) {
2056 struct sock *sk = sock->sk;
2057
2058 write_lock_bh(&sk->sk_callback_lock);
2059
2060 xs_save_old_callbacks(transport, sk);
2061
2062 sk->sk_user_data = xprt;
2063 sk->sk_data_ready = xs_tcp_data_ready;
2064 sk->sk_state_change = xs_tcp_state_change;
2065 sk->sk_write_space = xs_tcp_write_space;
2066 sk->sk_error_report = xs_error_report;
2067 sk->sk_allocation = GFP_ATOMIC;
2068
2069 /* socket options */
2070 sk->sk_userlocks |= SOCK_BINDPORT_LOCK;
2071 sock_reset_flag(sk, SOCK_LINGER);
2072 tcp_sk(sk)->linger2 = 0;
2073 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
2074
2075 xprt_clear_connected(xprt);
2076
2077 /* Reset to new socket */
2078 transport->sock = sock;
2079 transport->inet = sk;
2080
2081 write_unlock_bh(&sk->sk_callback_lock);
2082 }
2083
2084 if (!xprt_bound(xprt))
2085 goto out;
2086
2087 /* Tell the socket layer to start connecting... */
2088 xprt->stat.connect_count++;
2089 xprt->stat.connect_start = jiffies;
2090 ret = kernel_connect(sock, xs_addr(xprt), xprt->addrlen, O_NONBLOCK);
2091 switch (ret) {
2092 case 0:
2093 case -EINPROGRESS:
2094 /* SYN_SENT! */
2095 xprt->connect_cookie++;
2096 if (xprt->reestablish_timeout < XS_TCP_INIT_REEST_TO)
2097 xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
2098 }
2099out:
2100 return ret;
2101}
2102
2103/**
2104 * xs_tcp_setup_socket - create a TCP socket and connect to a remote endpoint
2105 * @xprt: RPC transport to connect
2106 * @transport: socket transport to connect
2107 * @create_sock: function to create a socket of the correct type
2108 *
2109 * Invoked by a work queue tasklet.
2110 */
2111static void xs_tcp_setup_socket(struct work_struct *work)
2112{
2113 struct sock_xprt *transport =
2114 container_of(work, struct sock_xprt, connect_worker.work);
2115 struct socket *sock = transport->sock;
2116 struct rpc_xprt *xprt = &transport->xprt;
2117 int status = -EIO;
2118
2119 if (xprt->shutdown)
2120 goto out;
2121
2122 current->flags |= PF_FSTRANS;
2123
2124 if (!sock) {
2125 clear_bit(XPRT_CONNECTION_ABORT, &xprt->state);
2126 sock = xs_create_sock(xprt, transport,
2127 xs_addr(xprt)->sa_family, SOCK_STREAM, IPPROTO_TCP);
2128 if (IS_ERR(sock)) {
2129 status = PTR_ERR(sock);
2130 goto out;
2131 }
2132 } else {
2133 int abort_and_exit;
2134
2135 abort_and_exit = test_and_clear_bit(XPRT_CONNECTION_ABORT,
2136 &xprt->state);
2137 /* "close" the socket, preserving the local port */
2138 xs_tcp_reuse_connection(transport);
2139
2140 if (abort_and_exit)
2141 goto out_eagain;
2142 }
2143
2144 dprintk("RPC: worker connecting xprt %p via %s to "
2145 "%s (port %s)\n", xprt,
2146 xprt->address_strings[RPC_DISPLAY_PROTO],
2147 xprt->address_strings[RPC_DISPLAY_ADDR],
2148 xprt->address_strings[RPC_DISPLAY_PORT]);
2149
2150 status = xs_tcp_finish_connecting(xprt, sock);
2151 dprintk("RPC: %p connect status %d connected %d sock state %d\n",
2152 xprt, -status, xprt_connected(xprt),
2153 sock->sk->sk_state);
2154 switch (status) {
2155 default:
2156 printk("%s: connect returned unhandled error %d\n",
2157 __func__, status);
2158 case -EADDRNOTAVAIL:
2159 /* We're probably in TIME_WAIT. Get rid of existing socket,
2160 * and retry
2161 */
2162 set_bit(XPRT_CONNECTION_CLOSE, &xprt->state);
2163 xprt_force_disconnect(xprt);
2164 break;
2165 case -ECONNREFUSED:
2166 case -ECONNRESET:
2167 case -ENETUNREACH:
2168 /* retry with existing socket, after a delay */
2169 case 0:
2170 case -EINPROGRESS:
2171 case -EALREADY:
2172 xprt_clear_connecting(xprt);
2173 current->flags &= ~PF_FSTRANS;
2174 return;
2175 case -EINVAL:
2176 /* Happens, for instance, if the user specified a link
2177 * local IPv6 address without a scope-id.
2178 */
2179 goto out;
2180 }
2181out_eagain:
2182 status = -EAGAIN;
2183out:
2184 xprt_clear_connecting(xprt);
2185 xprt_wake_pending_tasks(xprt, status);
2186 current->flags &= ~PF_FSTRANS;
2187}
2188
2189/**
2190 * xs_connect - connect a socket to a remote endpoint
2191 * @task: address of RPC task that manages state of connect request
2192 *
2193 * TCP: If the remote end dropped the connection, delay reconnecting.
2194 *
2195 * UDP socket connects are synchronous, but we use a work queue anyway
2196 * to guarantee that even unprivileged user processes can set up a
2197 * socket on a privileged port.
2198 *
2199 * If a UDP socket connect fails, the delay behavior here prevents
2200 * retry floods (hard mounts).
2201 */
2202static void xs_connect(struct rpc_task *task)
2203{
2204 struct rpc_xprt *xprt = task->tk_xprt;
2205 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
2206
2207 if (transport->sock != NULL && !RPC_IS_SOFTCONN(task)) {
2208 dprintk("RPC: xs_connect delayed xprt %p for %lu "
2209 "seconds\n",
2210 xprt, xprt->reestablish_timeout / HZ);
2211 queue_delayed_work(rpciod_workqueue,
2212 &transport->connect_worker,
2213 xprt->reestablish_timeout);
2214 xprt->reestablish_timeout <<= 1;
2215 if (xprt->reestablish_timeout < XS_TCP_INIT_REEST_TO)
2216 xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
2217 if (xprt->reestablish_timeout > XS_TCP_MAX_REEST_TO)
2218 xprt->reestablish_timeout = XS_TCP_MAX_REEST_TO;
2219 } else {
2220 dprintk("RPC: xs_connect scheduled xprt %p\n", xprt);
2221 queue_delayed_work(rpciod_workqueue,
2222 &transport->connect_worker, 0);
2223 }
2224}
2225
2226/**
2227 * xs_local_print_stats - display AF_LOCAL socket-specifc stats
2228 * @xprt: rpc_xprt struct containing statistics
2229 * @seq: output file
2230 *
2231 */
2232static void xs_local_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
2233{
2234 long idle_time = 0;
2235
2236 if (xprt_connected(xprt))
2237 idle_time = (long)(jiffies - xprt->last_used) / HZ;
2238
2239 seq_printf(seq, "\txprt:\tlocal %lu %lu %lu %ld %lu %lu %lu "
2240 "%llu %llu %lu %llu %llu\n",
2241 xprt->stat.bind_count,
2242 xprt->stat.connect_count,
2243 xprt->stat.connect_time,
2244 idle_time,
2245 xprt->stat.sends,
2246 xprt->stat.recvs,
2247 xprt->stat.bad_xids,
2248 xprt->stat.req_u,
2249 xprt->stat.bklog_u,
2250 xprt->stat.max_slots,
2251 xprt->stat.sending_u,
2252 xprt->stat.pending_u);
2253}
2254
2255/**
2256 * xs_udp_print_stats - display UDP socket-specifc stats
2257 * @xprt: rpc_xprt struct containing statistics
2258 * @seq: output file
2259 *
2260 */
2261static void xs_udp_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
2262{
2263 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
2264
2265 seq_printf(seq, "\txprt:\tudp %u %lu %lu %lu %lu %llu %llu "
2266 "%lu %llu %llu\n",
2267 transport->srcport,
2268 xprt->stat.bind_count,
2269 xprt->stat.sends,
2270 xprt->stat.recvs,
2271 xprt->stat.bad_xids,
2272 xprt->stat.req_u,
2273 xprt->stat.bklog_u,
2274 xprt->stat.max_slots,
2275 xprt->stat.sending_u,
2276 xprt->stat.pending_u);
2277}
2278
2279/**
2280 * xs_tcp_print_stats - display TCP socket-specifc stats
2281 * @xprt: rpc_xprt struct containing statistics
2282 * @seq: output file
2283 *
2284 */
2285static void xs_tcp_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
2286{
2287 struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
2288 long idle_time = 0;
2289
2290 if (xprt_connected(xprt))
2291 idle_time = (long)(jiffies - xprt->last_used) / HZ;
2292
2293 seq_printf(seq, "\txprt:\ttcp %u %lu %lu %lu %ld %lu %lu %lu "
2294 "%llu %llu %lu %llu %llu\n",
2295 transport->srcport,
2296 xprt->stat.bind_count,
2297 xprt->stat.connect_count,
2298 xprt->stat.connect_time,
2299 idle_time,
2300 xprt->stat.sends,
2301 xprt->stat.recvs,
2302 xprt->stat.bad_xids,
2303 xprt->stat.req_u,
2304 xprt->stat.bklog_u,
2305 xprt->stat.max_slots,
2306 xprt->stat.sending_u,
2307 xprt->stat.pending_u);
2308}
2309
2310/*
2311 * Allocate a bunch of pages for a scratch buffer for the rpc code. The reason
2312 * we allocate pages instead doing a kmalloc like rpc_malloc is because we want
2313 * to use the server side send routines.
2314 */
2315static void *bc_malloc(struct rpc_task *task, size_t size)
2316{
2317 struct page *page;
2318 struct rpc_buffer *buf;
2319
2320 BUG_ON(size > PAGE_SIZE - sizeof(struct rpc_buffer));
2321 page = alloc_page(GFP_KERNEL);
2322
2323 if (!page)
2324 return NULL;
2325
2326 buf = page_address(page);
2327 buf->len = PAGE_SIZE;
2328
2329 return buf->data;
2330}
2331
2332/*
2333 * Free the space allocated in the bc_alloc routine
2334 */
2335static void bc_free(void *buffer)
2336{
2337 struct rpc_buffer *buf;
2338
2339 if (!buffer)
2340 return;
2341
2342 buf = container_of(buffer, struct rpc_buffer, data);
2343 free_page((unsigned long)buf);
2344}
2345
2346/*
2347 * Use the svc_sock to send the callback. Must be called with svsk->sk_mutex
2348 * held. Borrows heavily from svc_tcp_sendto and xs_tcp_send_request.
2349 */
2350static int bc_sendto(struct rpc_rqst *req)
2351{
2352 int len;
2353 struct xdr_buf *xbufp = &req->rq_snd_buf;
2354 struct rpc_xprt *xprt = req->rq_xprt;
2355 struct sock_xprt *transport =
2356 container_of(xprt, struct sock_xprt, xprt);
2357 struct socket *sock = transport->sock;
2358 unsigned long headoff;
2359 unsigned long tailoff;
2360
2361 xs_encode_stream_record_marker(xbufp);
2362
2363 tailoff = (unsigned long)xbufp->tail[0].iov_base & ~PAGE_MASK;
2364 headoff = (unsigned long)xbufp->head[0].iov_base & ~PAGE_MASK;
2365 len = svc_send_common(sock, xbufp,
2366 virt_to_page(xbufp->head[0].iov_base), headoff,
2367 xbufp->tail[0].iov_base, tailoff);
2368
2369 if (len != xbufp->len) {
2370 printk(KERN_NOTICE "Error sending entire callback!\n");
2371 len = -EAGAIN;
2372 }
2373
2374 return len;
2375}
2376
2377/*
2378 * The send routine. Borrows from svc_send
2379 */
2380static int bc_send_request(struct rpc_task *task)
2381{
2382 struct rpc_rqst *req = task->tk_rqstp;
2383 struct svc_xprt *xprt;
2384 struct svc_sock *svsk;
2385 u32 len;
2386
2387 dprintk("sending request with xid: %08x\n", ntohl(req->rq_xid));
2388 /*
2389 * Get the server socket associated with this callback xprt
2390 */
2391 xprt = req->rq_xprt->bc_xprt;
2392 svsk = container_of(xprt, struct svc_sock, sk_xprt);
2393
2394 /*
2395 * Grab the mutex to serialize data as the connection is shared
2396 * with the fore channel
2397 */
2398 if (!mutex_trylock(&xprt->xpt_mutex)) {
2399 rpc_sleep_on(&xprt->xpt_bc_pending, task, NULL);
2400 if (!mutex_trylock(&xprt->xpt_mutex))
2401 return -EAGAIN;
2402 rpc_wake_up_queued_task(&xprt->xpt_bc_pending, task);
2403 }
2404 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
2405 len = -ENOTCONN;
2406 else
2407 len = bc_sendto(req);
2408 mutex_unlock(&xprt->xpt_mutex);
2409
2410 if (len > 0)
2411 len = 0;
2412
2413 return len;
2414}
2415
2416/*
2417 * The close routine. Since this is client initiated, we do nothing
2418 */
2419
2420static void bc_close(struct rpc_xprt *xprt)
2421{
2422}
2423
2424/*
2425 * The xprt destroy routine. Again, because this connection is client
2426 * initiated, we do nothing
2427 */
2428
2429static void bc_destroy(struct rpc_xprt *xprt)
2430{
2431}
2432
2433static struct rpc_xprt_ops xs_local_ops = {
2434 .reserve_xprt = xprt_reserve_xprt,
2435 .release_xprt = xs_tcp_release_xprt,
2436 .rpcbind = xs_local_rpcbind,
2437 .set_port = xs_local_set_port,
2438 .connect = xs_connect,
2439 .buf_alloc = rpc_malloc,
2440 .buf_free = rpc_free,
2441 .send_request = xs_local_send_request,
2442 .set_retrans_timeout = xprt_set_retrans_timeout_def,
2443 .close = xs_close,
2444 .destroy = xs_destroy,
2445 .print_stats = xs_local_print_stats,
2446};
2447
2448static struct rpc_xprt_ops xs_udp_ops = {
2449 .set_buffer_size = xs_udp_set_buffer_size,
2450 .reserve_xprt = xprt_reserve_xprt_cong,
2451 .release_xprt = xprt_release_xprt_cong,
2452 .rpcbind = rpcb_getport_async,
2453 .set_port = xs_set_port,
2454 .connect = xs_connect,
2455 .buf_alloc = rpc_malloc,
2456 .buf_free = rpc_free,
2457 .send_request = xs_udp_send_request,
2458 .set_retrans_timeout = xprt_set_retrans_timeout_rtt,
2459 .timer = xs_udp_timer,
2460 .release_request = xprt_release_rqst_cong,
2461 .close = xs_close,
2462 .destroy = xs_destroy,
2463 .print_stats = xs_udp_print_stats,
2464};
2465
2466static struct rpc_xprt_ops xs_tcp_ops = {
2467 .reserve_xprt = xprt_reserve_xprt,
2468 .release_xprt = xs_tcp_release_xprt,
2469 .rpcbind = rpcb_getport_async,
2470 .set_port = xs_set_port,
2471 .connect = xs_connect,
2472 .buf_alloc = rpc_malloc,
2473 .buf_free = rpc_free,
2474 .send_request = xs_tcp_send_request,
2475 .set_retrans_timeout = xprt_set_retrans_timeout_def,
2476 .close = xs_tcp_close,
2477 .destroy = xs_destroy,
2478 .print_stats = xs_tcp_print_stats,
2479};
2480
2481/*
2482 * The rpc_xprt_ops for the server backchannel
2483 */
2484
2485static struct rpc_xprt_ops bc_tcp_ops = {
2486 .reserve_xprt = xprt_reserve_xprt,
2487 .release_xprt = xprt_release_xprt,
2488 .rpcbind = xs_local_rpcbind,
2489 .buf_alloc = bc_malloc,
2490 .buf_free = bc_free,
2491 .send_request = bc_send_request,
2492 .set_retrans_timeout = xprt_set_retrans_timeout_def,
2493 .close = bc_close,
2494 .destroy = bc_destroy,
2495 .print_stats = xs_tcp_print_stats,
2496};
2497
2498static int xs_init_anyaddr(const int family, struct sockaddr *sap)
2499{
2500 static const struct sockaddr_in sin = {
2501 .sin_family = AF_INET,
2502 .sin_addr.s_addr = htonl(INADDR_ANY),
2503 };
2504 static const struct sockaddr_in6 sin6 = {
2505 .sin6_family = AF_INET6,
2506 .sin6_addr = IN6ADDR_ANY_INIT,
2507 };
2508
2509 switch (family) {
2510 case AF_LOCAL:
2511 break;
2512 case AF_INET:
2513 memcpy(sap, &sin, sizeof(sin));
2514 break;
2515 case AF_INET6:
2516 memcpy(sap, &sin6, sizeof(sin6));
2517 break;
2518 default:
2519 dprintk("RPC: %s: Bad address family\n", __func__);
2520 return -EAFNOSUPPORT;
2521 }
2522 return 0;
2523}
2524
2525static struct rpc_xprt *xs_setup_xprt(struct xprt_create *args,
2526 unsigned int slot_table_size,
2527 unsigned int max_slot_table_size)
2528{
2529 struct rpc_xprt *xprt;
2530 struct sock_xprt *new;
2531
2532 if (args->addrlen > sizeof(xprt->addr)) {
2533 dprintk("RPC: xs_setup_xprt: address too large\n");
2534 return ERR_PTR(-EBADF);
2535 }
2536
2537 xprt = xprt_alloc(args->net, sizeof(*new), slot_table_size,
2538 max_slot_table_size);
2539 if (xprt == NULL) {
2540 dprintk("RPC: xs_setup_xprt: couldn't allocate "
2541 "rpc_xprt\n");
2542 return ERR_PTR(-ENOMEM);
2543 }
2544
2545 new = container_of(xprt, struct sock_xprt, xprt);
2546 memcpy(&xprt->addr, args->dstaddr, args->addrlen);
2547 xprt->addrlen = args->addrlen;
2548 if (args->srcaddr)
2549 memcpy(&new->srcaddr, args->srcaddr, args->addrlen);
2550 else {
2551 int err;
2552 err = xs_init_anyaddr(args->dstaddr->sa_family,
2553 (struct sockaddr *)&new->srcaddr);
2554 if (err != 0) {
2555 xprt_free(xprt);
2556 return ERR_PTR(err);
2557 }
2558 }
2559
2560 return xprt;
2561}
2562
2563static const struct rpc_timeout xs_local_default_timeout = {
2564 .to_initval = 10 * HZ,
2565 .to_maxval = 10 * HZ,
2566 .to_retries = 2,
2567};
2568
2569/**
2570 * xs_setup_local - Set up transport to use an AF_LOCAL socket
2571 * @args: rpc transport creation arguments
2572 *
2573 * AF_LOCAL is a "tpi_cots_ord" transport, just like TCP
2574 */
2575static struct rpc_xprt *xs_setup_local(struct xprt_create *args)
2576{
2577 struct sockaddr_un *sun = (struct sockaddr_un *)args->dstaddr;
2578 struct sock_xprt *transport;
2579 struct rpc_xprt *xprt;
2580 struct rpc_xprt *ret;
2581
2582 xprt = xs_setup_xprt(args, xprt_tcp_slot_table_entries,
2583 xprt_max_tcp_slot_table_entries);
2584 if (IS_ERR(xprt))
2585 return xprt;
2586 transport = container_of(xprt, struct sock_xprt, xprt);
2587
2588 xprt->prot = 0;
2589 xprt->tsh_size = sizeof(rpc_fraghdr) / sizeof(u32);
2590 xprt->max_payload = RPC_MAX_FRAGMENT_SIZE;
2591
2592 xprt->bind_timeout = XS_BIND_TO;
2593 xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
2594 xprt->idle_timeout = XS_IDLE_DISC_TO;
2595
2596 xprt->ops = &xs_local_ops;
2597 xprt->timeout = &xs_local_default_timeout;
2598
2599 switch (sun->sun_family) {
2600 case AF_LOCAL:
2601 if (sun->sun_path[0] != '/') {
2602 dprintk("RPC: bad AF_LOCAL address: %s\n",
2603 sun->sun_path);
2604 ret = ERR_PTR(-EINVAL);
2605 goto out_err;
2606 }
2607 xprt_set_bound(xprt);
2608 INIT_DELAYED_WORK(&transport->connect_worker,
2609 xs_local_setup_socket);
2610 xs_format_peer_addresses(xprt, "local", RPCBIND_NETID_LOCAL);
2611 break;
2612 default:
2613 ret = ERR_PTR(-EAFNOSUPPORT);
2614 goto out_err;
2615 }
2616
2617 dprintk("RPC: set up xprt to %s via AF_LOCAL\n",
2618 xprt->address_strings[RPC_DISPLAY_ADDR]);
2619
2620 if (try_module_get(THIS_MODULE))
2621 return xprt;
2622 ret = ERR_PTR(-EINVAL);
2623out_err:
2624 xprt_free(xprt);
2625 return ret;
2626}
2627
2628static const struct rpc_timeout xs_udp_default_timeout = {
2629 .to_initval = 5 * HZ,
2630 .to_maxval = 30 * HZ,
2631 .to_increment = 5 * HZ,
2632 .to_retries = 5,
2633};
2634
2635/**
2636 * xs_setup_udp - Set up transport to use a UDP socket
2637 * @args: rpc transport creation arguments
2638 *
2639 */
2640static struct rpc_xprt *xs_setup_udp(struct xprt_create *args)
2641{
2642 struct sockaddr *addr = args->dstaddr;
2643 struct rpc_xprt *xprt;
2644 struct sock_xprt *transport;
2645 struct rpc_xprt *ret;
2646
2647 xprt = xs_setup_xprt(args, xprt_udp_slot_table_entries,
2648 xprt_udp_slot_table_entries);
2649 if (IS_ERR(xprt))
2650 return xprt;
2651 transport = container_of(xprt, struct sock_xprt, xprt);
2652
2653 xprt->prot = IPPROTO_UDP;
2654 xprt->tsh_size = 0;
2655 /* XXX: header size can vary due to auth type, IPv6, etc. */
2656 xprt->max_payload = (1U << 16) - (MAX_HEADER << 3);
2657
2658 xprt->bind_timeout = XS_BIND_TO;
2659 xprt->reestablish_timeout = XS_UDP_REEST_TO;
2660 xprt->idle_timeout = XS_IDLE_DISC_TO;
2661
2662 xprt->ops = &xs_udp_ops;
2663
2664 xprt->timeout = &xs_udp_default_timeout;
2665
2666 switch (addr->sa_family) {
2667 case AF_INET:
2668 if (((struct sockaddr_in *)addr)->sin_port != htons(0))
2669 xprt_set_bound(xprt);
2670
2671 INIT_DELAYED_WORK(&transport->connect_worker,
2672 xs_udp_setup_socket);
2673 xs_format_peer_addresses(xprt, "udp", RPCBIND_NETID_UDP);
2674 break;
2675 case AF_INET6:
2676 if (((struct sockaddr_in6 *)addr)->sin6_port != htons(0))
2677 xprt_set_bound(xprt);
2678
2679 INIT_DELAYED_WORK(&transport->connect_worker,
2680 xs_udp_setup_socket);
2681 xs_format_peer_addresses(xprt, "udp", RPCBIND_NETID_UDP6);
2682 break;
2683 default:
2684 ret = ERR_PTR(-EAFNOSUPPORT);
2685 goto out_err;
2686 }
2687
2688 if (xprt_bound(xprt))
2689 dprintk("RPC: set up xprt to %s (port %s) via %s\n",
2690 xprt->address_strings[RPC_DISPLAY_ADDR],
2691 xprt->address_strings[RPC_DISPLAY_PORT],
2692 xprt->address_strings[RPC_DISPLAY_PROTO]);
2693 else
2694 dprintk("RPC: set up xprt to %s (autobind) via %s\n",
2695 xprt->address_strings[RPC_DISPLAY_ADDR],
2696 xprt->address_strings[RPC_DISPLAY_PROTO]);
2697
2698 if (try_module_get(THIS_MODULE))
2699 return xprt;
2700 ret = ERR_PTR(-EINVAL);
2701out_err:
2702 xprt_free(xprt);
2703 return ret;
2704}
2705
2706static const struct rpc_timeout xs_tcp_default_timeout = {
2707 .to_initval = 60 * HZ,
2708 .to_maxval = 60 * HZ,
2709 .to_retries = 2,
2710};
2711
2712/**
2713 * xs_setup_tcp - Set up transport to use a TCP socket
2714 * @args: rpc transport creation arguments
2715 *
2716 */
2717static struct rpc_xprt *xs_setup_tcp(struct xprt_create *args)
2718{
2719 struct sockaddr *addr = args->dstaddr;
2720 struct rpc_xprt *xprt;
2721 struct sock_xprt *transport;
2722 struct rpc_xprt *ret;
2723
2724 xprt = xs_setup_xprt(args, xprt_tcp_slot_table_entries,
2725 xprt_max_tcp_slot_table_entries);
2726 if (IS_ERR(xprt))
2727 return xprt;
2728 transport = container_of(xprt, struct sock_xprt, xprt);
2729
2730 xprt->prot = IPPROTO_TCP;
2731 xprt->tsh_size = sizeof(rpc_fraghdr) / sizeof(u32);
2732 xprt->max_payload = RPC_MAX_FRAGMENT_SIZE;
2733
2734 xprt->bind_timeout = XS_BIND_TO;
2735 xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
2736 xprt->idle_timeout = XS_IDLE_DISC_TO;
2737
2738 xprt->ops = &xs_tcp_ops;
2739 xprt->timeout = &xs_tcp_default_timeout;
2740
2741 switch (addr->sa_family) {
2742 case AF_INET:
2743 if (((struct sockaddr_in *)addr)->sin_port != htons(0))
2744 xprt_set_bound(xprt);
2745
2746 INIT_DELAYED_WORK(&transport->connect_worker,
2747 xs_tcp_setup_socket);
2748 xs_format_peer_addresses(xprt, "tcp", RPCBIND_NETID_TCP);
2749 break;
2750 case AF_INET6:
2751 if (((struct sockaddr_in6 *)addr)->sin6_port != htons(0))
2752 xprt_set_bound(xprt);
2753
2754 INIT_DELAYED_WORK(&transport->connect_worker,
2755 xs_tcp_setup_socket);
2756 xs_format_peer_addresses(xprt, "tcp", RPCBIND_NETID_TCP6);
2757 break;
2758 default:
2759 ret = ERR_PTR(-EAFNOSUPPORT);
2760 goto out_err;
2761 }
2762
2763 if (xprt_bound(xprt))
2764 dprintk("RPC: set up xprt to %s (port %s) via %s\n",
2765 xprt->address_strings[RPC_DISPLAY_ADDR],
2766 xprt->address_strings[RPC_DISPLAY_PORT],
2767 xprt->address_strings[RPC_DISPLAY_PROTO]);
2768 else
2769 dprintk("RPC: set up xprt to %s (autobind) via %s\n",
2770 xprt->address_strings[RPC_DISPLAY_ADDR],
2771 xprt->address_strings[RPC_DISPLAY_PROTO]);
2772
2773
2774 if (try_module_get(THIS_MODULE))
2775 return xprt;
2776 ret = ERR_PTR(-EINVAL);
2777out_err:
2778 xprt_free(xprt);
2779 return ret;
2780}
2781
2782/**
2783 * xs_setup_bc_tcp - Set up transport to use a TCP backchannel socket
2784 * @args: rpc transport creation arguments
2785 *
2786 */
2787static struct rpc_xprt *xs_setup_bc_tcp(struct xprt_create *args)
2788{
2789 struct sockaddr *addr = args->dstaddr;
2790 struct rpc_xprt *xprt;
2791 struct sock_xprt *transport;
2792 struct svc_sock *bc_sock;
2793 struct rpc_xprt *ret;
2794
2795 if (args->bc_xprt->xpt_bc_xprt) {
2796 /*
2797 * This server connection already has a backchannel
2798 * export; we can't create a new one, as we wouldn't be
2799 * able to match replies based on xid any more. So,
2800 * reuse the already-existing one:
2801 */
2802 return args->bc_xprt->xpt_bc_xprt;
2803 }
2804 xprt = xs_setup_xprt(args, xprt_tcp_slot_table_entries,
2805 xprt_tcp_slot_table_entries);
2806 if (IS_ERR(xprt))
2807 return xprt;
2808 transport = container_of(xprt, struct sock_xprt, xprt);
2809
2810 xprt->prot = IPPROTO_TCP;
2811 xprt->tsh_size = sizeof(rpc_fraghdr) / sizeof(u32);
2812 xprt->max_payload = RPC_MAX_FRAGMENT_SIZE;
2813 xprt->timeout = &xs_tcp_default_timeout;
2814
2815 /* backchannel */
2816 xprt_set_bound(xprt);
2817 xprt->bind_timeout = 0;
2818 xprt->reestablish_timeout = 0;
2819 xprt->idle_timeout = 0;
2820
2821 xprt->ops = &bc_tcp_ops;
2822
2823 switch (addr->sa_family) {
2824 case AF_INET:
2825 xs_format_peer_addresses(xprt, "tcp",
2826 RPCBIND_NETID_TCP);
2827 break;
2828 case AF_INET6:
2829 xs_format_peer_addresses(xprt, "tcp",
2830 RPCBIND_NETID_TCP6);
2831 break;
2832 default:
2833 ret = ERR_PTR(-EAFNOSUPPORT);
2834 goto out_err;
2835 }
2836
2837 dprintk("RPC: set up xprt to %s (port %s) via %s\n",
2838 xprt->address_strings[RPC_DISPLAY_ADDR],
2839 xprt->address_strings[RPC_DISPLAY_PORT],
2840 xprt->address_strings[RPC_DISPLAY_PROTO]);
2841
2842 /*
2843 * Once we've associated a backchannel xprt with a connection,
2844 * we want to keep it around as long as long as the connection
2845 * lasts, in case we need to start using it for a backchannel
2846 * again; this reference won't be dropped until bc_xprt is
2847 * destroyed.
2848 */
2849 xprt_get(xprt);
2850 args->bc_xprt->xpt_bc_xprt = xprt;
2851 xprt->bc_xprt = args->bc_xprt;
2852 bc_sock = container_of(args->bc_xprt, struct svc_sock, sk_xprt);
2853 transport->sock = bc_sock->sk_sock;
2854 transport->inet = bc_sock->sk_sk;
2855
2856 /*
2857 * Since we don't want connections for the backchannel, we set
2858 * the xprt status to connected
2859 */
2860 xprt_set_connected(xprt);
2861
2862
2863 if (try_module_get(THIS_MODULE))
2864 return xprt;
2865 xprt_put(xprt);
2866 ret = ERR_PTR(-EINVAL);
2867out_err:
2868 xprt_free(xprt);
2869 return ret;
2870}
2871
2872static struct xprt_class xs_local_transport = {
2873 .list = LIST_HEAD_INIT(xs_local_transport.list),
2874 .name = "named UNIX socket",
2875 .owner = THIS_MODULE,
2876 .ident = XPRT_TRANSPORT_LOCAL,
2877 .setup = xs_setup_local,
2878};
2879
2880static struct xprt_class xs_udp_transport = {
2881 .list = LIST_HEAD_INIT(xs_udp_transport.list),
2882 .name = "udp",
2883 .owner = THIS_MODULE,
2884 .ident = XPRT_TRANSPORT_UDP,
2885 .setup = xs_setup_udp,
2886};
2887
2888static struct xprt_class xs_tcp_transport = {
2889 .list = LIST_HEAD_INIT(xs_tcp_transport.list),
2890 .name = "tcp",
2891 .owner = THIS_MODULE,
2892 .ident = XPRT_TRANSPORT_TCP,
2893 .setup = xs_setup_tcp,
2894};
2895
2896static struct xprt_class xs_bc_tcp_transport = {
2897 .list = LIST_HEAD_INIT(xs_bc_tcp_transport.list),
2898 .name = "tcp NFSv4.1 backchannel",
2899 .owner = THIS_MODULE,
2900 .ident = XPRT_TRANSPORT_BC_TCP,
2901 .setup = xs_setup_bc_tcp,
2902};
2903
2904/**
2905 * init_socket_xprt - set up xprtsock's sysctls, register with RPC client
2906 *
2907 */
2908int init_socket_xprt(void)
2909{
2910#ifdef RPC_DEBUG
2911 if (!sunrpc_table_header)
2912 sunrpc_table_header = register_sysctl_table(sunrpc_table);
2913#endif
2914
2915 xprt_register_transport(&xs_local_transport);
2916 xprt_register_transport(&xs_udp_transport);
2917 xprt_register_transport(&xs_tcp_transport);
2918 xprt_register_transport(&xs_bc_tcp_transport);
2919
2920 return 0;
2921}
2922
2923/**
2924 * cleanup_socket_xprt - remove xprtsock's sysctls, unregister
2925 *
2926 */
2927void cleanup_socket_xprt(void)
2928{
2929#ifdef RPC_DEBUG
2930 if (sunrpc_table_header) {
2931 unregister_sysctl_table(sunrpc_table_header);
2932 sunrpc_table_header = NULL;
2933 }
2934#endif
2935
2936 xprt_unregister_transport(&xs_local_transport);
2937 xprt_unregister_transport(&xs_udp_transport);
2938 xprt_unregister_transport(&xs_tcp_transport);
2939 xprt_unregister_transport(&xs_bc_tcp_transport);
2940}
2941
2942static int param_set_uint_minmax(const char *val,
2943 const struct kernel_param *kp,
2944 unsigned int min, unsigned int max)
2945{
2946 unsigned long num;
2947 int ret;
2948
2949 if (!val)
2950 return -EINVAL;
2951 ret = strict_strtoul(val, 0, &num);
2952 if (ret == -EINVAL || num < min || num > max)
2953 return -EINVAL;
2954 *((unsigned int *)kp->arg) = num;
2955 return 0;
2956}
2957
2958static int param_set_portnr(const char *val, const struct kernel_param *kp)
2959{
2960 return param_set_uint_minmax(val, kp,
2961 RPC_MIN_RESVPORT,
2962 RPC_MAX_RESVPORT);
2963}
2964
2965static struct kernel_param_ops param_ops_portnr = {
2966 .set = param_set_portnr,
2967 .get = param_get_uint,
2968};
2969
2970#define param_check_portnr(name, p) \
2971 __param_check(name, p, unsigned int);
2972
2973module_param_named(min_resvport, xprt_min_resvport, portnr, 0644);
2974module_param_named(max_resvport, xprt_max_resvport, portnr, 0644);
2975
2976static int param_set_slot_table_size(const char *val,
2977 const struct kernel_param *kp)
2978{
2979 return param_set_uint_minmax(val, kp,
2980 RPC_MIN_SLOT_TABLE,
2981 RPC_MAX_SLOT_TABLE);
2982}
2983
2984static struct kernel_param_ops param_ops_slot_table_size = {
2985 .set = param_set_slot_table_size,
2986 .get = param_get_uint,
2987};
2988
2989#define param_check_slot_table_size(name, p) \
2990 __param_check(name, p, unsigned int);
2991
2992static int param_set_max_slot_table_size(const char *val,
2993 const struct kernel_param *kp)
2994{
2995 return param_set_uint_minmax(val, kp,
2996 RPC_MIN_SLOT_TABLE,
2997 RPC_MAX_SLOT_TABLE_LIMIT);
2998}
2999
3000static struct kernel_param_ops param_ops_max_slot_table_size = {
3001 .set = param_set_max_slot_table_size,
3002 .get = param_get_uint,
3003};
3004
3005#define param_check_max_slot_table_size(name, p) \
3006 __param_check(name, p, unsigned int);
3007
3008module_param_named(tcp_slot_table_entries, xprt_tcp_slot_table_entries,
3009 slot_table_size, 0644);
3010module_param_named(tcp_max_slot_table_entries, xprt_max_tcp_slot_table_entries,
3011 max_slot_table_size, 0644);
3012module_param_named(udp_slot_table_entries, xprt_udp_slot_table_entries,
3013 slot_table_size, 0644);
3014