blob: bf5b5cdafebfcc82866369bdb3d895c83379a4e3 [file] [log] [blame]
Tom Tucker1d8206b92007-12-30 21:07:15 -06001/*
2 * linux/net/sunrpc/svc_xprt.c
3 *
4 * Author: Tom Tucker <tom@opengridcomputing.com>
5 */
6
7#include <linux/sched.h>
8#include <linux/errno.h>
Tom Tucker1d8206b92007-12-30 21:07:15 -06009#include <linux/freezer.h>
Jeff Layton70867212008-02-07 16:34:54 -050010#include <linux/kthread.h>
Tom Tucker1d8206b92007-12-30 21:07:15 -060011#include <net/sock.h>
Tom Tucker1d8206b92007-12-30 21:07:15 -060012#include <linux/sunrpc/stats.h>
13#include <linux/sunrpc/svc_xprt.h>
14
15#define RPCDBG_FACILITY RPCDBG_SVCXPRT
16
Tom Tucker0f0257e2007-12-30 21:08:27 -060017static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
18static int svc_deferred_recv(struct svc_rqst *rqstp);
19static struct cache_deferred_req *svc_defer(struct cache_req *req);
20static void svc_age_temp_xprts(unsigned long closure);
21
22/* apparently the "standard" is that clients close
23 * idle connections after 5 minutes, servers after
24 * 6 minutes
25 * http://www.connectathon.org/talks96/nfstcp.pdf
26 */
27static int svc_conn_age_period = 6*60;
28
Tom Tucker1d8206b92007-12-30 21:07:15 -060029/* List of registered transport classes */
30static DEFINE_SPINLOCK(svc_xprt_class_lock);
31static LIST_HEAD(svc_xprt_class_list);
32
Tom Tucker0f0257e2007-12-30 21:08:27 -060033/* SMP locking strategy:
34 *
35 * svc_pool->sp_lock protects most of the fields of that pool.
36 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
37 * when both need to be taken (rare), svc_serv->sv_lock is first.
38 * BKL protects svc_serv->sv_nrthread.
39 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
40 * and the ->sk_info_authunix cache.
41 *
42 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
43 * enqueued multiply. During normal transport processing this bit
44 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
45 * Providers should not manipulate this bit directly.
46 *
47 * Some flags can be set to certain values at any time
48 * providing that certain rules are followed:
49 *
50 * XPT_CONN, XPT_DATA:
51 * - Can be set or cleared at any time.
52 * - After a set, svc_xprt_enqueue must be called to enqueue
53 * the transport for processing.
54 * - After a clear, the transport must be read/accepted.
55 * If this succeeds, it must be set again.
56 * XPT_CLOSE:
57 * - Can set at any time. It is never cleared.
58 * XPT_DEAD:
59 * - Can only be set while XPT_BUSY is held which ensures
60 * that no other thread will be using the transport or will
61 * try to set XPT_DEAD.
62 */
63
Tom Tucker1d8206b92007-12-30 21:07:15 -060064int svc_reg_xprt_class(struct svc_xprt_class *xcl)
65{
66 struct svc_xprt_class *cl;
67 int res = -EEXIST;
68
69 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
70
71 INIT_LIST_HEAD(&xcl->xcl_list);
72 spin_lock(&svc_xprt_class_lock);
73 /* Make sure there isn't already a class with the same name */
74 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
75 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
76 goto out;
77 }
78 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
79 res = 0;
80out:
81 spin_unlock(&svc_xprt_class_lock);
82 return res;
83}
84EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
85
86void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
87{
88 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
89 spin_lock(&svc_xprt_class_lock);
90 list_del_init(&xcl->xcl_list);
91 spin_unlock(&svc_xprt_class_lock);
92}
93EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
94
Tom Tuckerdc9a16e2007-12-30 21:08:31 -060095/*
96 * Format the transport list for printing
97 */
98int svc_print_xprts(char *buf, int maxlen)
99{
100 struct list_head *le;
101 char tmpstr[80];
102 int len = 0;
103 buf[0] = '\0';
104
105 spin_lock(&svc_xprt_class_lock);
106 list_for_each(le, &svc_xprt_class_list) {
107 int slen;
108 struct svc_xprt_class *xcl =
109 list_entry(le, struct svc_xprt_class, xcl_list);
110
111 sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
112 slen = strlen(tmpstr);
113 if (len + slen > maxlen)
114 break;
115 len += slen;
116 strcat(buf, tmpstr);
117 }
118 spin_unlock(&svc_xprt_class_lock);
119
120 return len;
121}
122
Tom Tuckere1b31572007-12-30 21:07:46 -0600123static void svc_xprt_free(struct kref *kref)
124{
125 struct svc_xprt *xprt =
126 container_of(kref, struct svc_xprt, xpt_ref);
127 struct module *owner = xprt->xpt_class->xcl_owner;
Tom Tuckerdef13d72007-12-30 21:08:08 -0600128 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags)
129 && xprt->xpt_auth_cache != NULL)
130 svcauth_unix_info_release(xprt->xpt_auth_cache);
Tom Tuckere1b31572007-12-30 21:07:46 -0600131 xprt->xpt_ops->xpo_free(xprt);
132 module_put(owner);
133}
134
135void svc_xprt_put(struct svc_xprt *xprt)
136{
137 kref_put(&xprt->xpt_ref, svc_xprt_free);
138}
139EXPORT_SYMBOL_GPL(svc_xprt_put);
140
Tom Tucker1d8206b92007-12-30 21:07:15 -0600141/*
142 * Called by transport drivers to initialize the transport independent
143 * portion of the transport instance.
144 */
Tom Tuckerbb5cf162007-12-30 21:07:50 -0600145void svc_xprt_init(struct svc_xprt_class *xcl, struct svc_xprt *xprt,
146 struct svc_serv *serv)
Tom Tucker1d8206b92007-12-30 21:07:15 -0600147{
148 memset(xprt, 0, sizeof(*xprt));
149 xprt->xpt_class = xcl;
150 xprt->xpt_ops = xcl->xcl_ops;
Tom Tuckere1b31572007-12-30 21:07:46 -0600151 kref_init(&xprt->xpt_ref);
Tom Tuckerbb5cf162007-12-30 21:07:50 -0600152 xprt->xpt_server = serv;
Tom Tucker7a182082007-12-30 21:07:53 -0600153 INIT_LIST_HEAD(&xprt->xpt_list);
154 INIT_LIST_HEAD(&xprt->xpt_ready);
Tom Tucker8c7b0172007-12-30 21:08:10 -0600155 INIT_LIST_HEAD(&xprt->xpt_deferred);
Tom Tuckera50fea22007-12-30 21:07:59 -0600156 mutex_init(&xprt->xpt_mutex);
Tom Tuckerdef13d72007-12-30 21:08:08 -0600157 spin_lock_init(&xprt->xpt_lock);
Tom Tucker4e5caaa2007-12-30 21:08:20 -0600158 set_bit(XPT_BUSY, &xprt->xpt_flags);
Tom Tucker1d8206b92007-12-30 21:07:15 -0600159}
160EXPORT_SYMBOL_GPL(svc_xprt_init);
Tom Tuckerb700cbb2007-12-30 21:07:42 -0600161
Chuck Lever5dd248f2008-06-30 18:45:37 -0400162static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
163 struct svc_serv *serv,
164 unsigned short port, int flags)
Tom Tuckerb700cbb2007-12-30 21:07:42 -0600165{
Tom Tuckerb700cbb2007-12-30 21:07:42 -0600166 struct sockaddr_in sin = {
167 .sin_family = AF_INET,
Al Viroe6f1ceb2008-03-17 22:44:53 -0700168 .sin_addr.s_addr = htonl(INADDR_ANY),
Tom Tuckerb700cbb2007-12-30 21:07:42 -0600169 .sin_port = htons(port),
170 };
Chuck Lever5dd248f2008-06-30 18:45:37 -0400171 struct sockaddr_in6 sin6 = {
172 .sin6_family = AF_INET6,
173 .sin6_addr = IN6ADDR_ANY_INIT,
174 .sin6_port = htons(port),
175 };
176 struct sockaddr *sap;
177 size_t len;
178
179 switch (serv->sv_family) {
180 case AF_INET:
181 sap = (struct sockaddr *)&sin;
182 len = sizeof(sin);
183 break;
184 case AF_INET6:
185 sap = (struct sockaddr *)&sin6;
186 len = sizeof(sin6);
187 break;
188 default:
189 return ERR_PTR(-EAFNOSUPPORT);
190 }
191
192 return xcl->xcl_ops->xpo_create(serv, sap, len, flags);
193}
194
195int svc_create_xprt(struct svc_serv *serv, char *xprt_name, unsigned short port,
196 int flags)
197{
198 struct svc_xprt_class *xcl;
199
Tom Tuckerb700cbb2007-12-30 21:07:42 -0600200 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
201 spin_lock(&svc_xprt_class_lock);
202 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
Tom Tucker4e5caaa2007-12-30 21:08:20 -0600203 struct svc_xprt *newxprt;
204
205 if (strcmp(xprt_name, xcl->xcl_name))
206 continue;
207
208 if (!try_module_get(xcl->xcl_owner))
209 goto err;
210
211 spin_unlock(&svc_xprt_class_lock);
Chuck Lever5dd248f2008-06-30 18:45:37 -0400212 newxprt = __svc_xpo_create(xcl, serv, port, flags);
Tom Tucker4e5caaa2007-12-30 21:08:20 -0600213 if (IS_ERR(newxprt)) {
214 module_put(xcl->xcl_owner);
215 return PTR_ERR(newxprt);
Tom Tuckerb700cbb2007-12-30 21:07:42 -0600216 }
Tom Tucker4e5caaa2007-12-30 21:08:20 -0600217
218 clear_bit(XPT_TEMP, &newxprt->xpt_flags);
219 spin_lock_bh(&serv->sv_lock);
220 list_add(&newxprt->xpt_list, &serv->sv_permsocks);
221 spin_unlock_bh(&serv->sv_lock);
222 clear_bit(XPT_BUSY, &newxprt->xpt_flags);
223 return svc_xprt_local_port(newxprt);
Tom Tuckerb700cbb2007-12-30 21:07:42 -0600224 }
Tom Tucker4e5caaa2007-12-30 21:08:20 -0600225 err:
Tom Tuckerb700cbb2007-12-30 21:07:42 -0600226 spin_unlock(&svc_xprt_class_lock);
227 dprintk("svc: transport %s not found\n", xprt_name);
Tom Tucker4e5caaa2007-12-30 21:08:20 -0600228 return -ENOENT;
Tom Tuckerb700cbb2007-12-30 21:07:42 -0600229}
230EXPORT_SYMBOL_GPL(svc_create_xprt);
Tom Tucker9dbc2402007-12-30 21:08:12 -0600231
232/*
233 * Copy the local and remote xprt addresses to the rqstp structure
234 */
235void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
236{
237 struct sockaddr *sin;
238
239 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
240 rqstp->rq_addrlen = xprt->xpt_remotelen;
241
242 /*
243 * Destination address in request is needed for binding the
244 * source address in RPC replies/callbacks later.
245 */
246 sin = (struct sockaddr *)&xprt->xpt_local;
247 switch (sin->sa_family) {
248 case AF_INET:
249 rqstp->rq_daddr.addr = ((struct sockaddr_in *)sin)->sin_addr;
250 break;
251 case AF_INET6:
252 rqstp->rq_daddr.addr6 = ((struct sockaddr_in6 *)sin)->sin6_addr;
253 break;
254 }
255}
256EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
257
Tom Tucker0f0257e2007-12-30 21:08:27 -0600258/**
259 * svc_print_addr - Format rq_addr field for printing
260 * @rqstp: svc_rqst struct containing address to print
261 * @buf: target buffer for formatted address
262 * @len: length of target buffer
263 *
264 */
265char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
266{
267 return __svc_print_addr(svc_addr(rqstp), buf, len);
268}
269EXPORT_SYMBOL_GPL(svc_print_addr);
270
271/*
272 * Queue up an idle server thread. Must have pool->sp_lock held.
273 * Note: this is really a stack rather than a queue, so that we only
274 * use as many different threads as we need, and the rest don't pollute
275 * the cache.
276 */
277static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
278{
279 list_add(&rqstp->rq_list, &pool->sp_threads);
280}
281
282/*
283 * Dequeue an nfsd thread. Must have pool->sp_lock held.
284 */
285static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
286{
287 list_del(&rqstp->rq_list);
288}
289
290/*
291 * Queue up a transport with data pending. If there are idle nfsd
292 * processes, wake 'em up.
293 *
294 */
295void svc_xprt_enqueue(struct svc_xprt *xprt)
296{
297 struct svc_serv *serv = xprt->xpt_server;
298 struct svc_pool *pool;
299 struct svc_rqst *rqstp;
300 int cpu;
301
302 if (!(xprt->xpt_flags &
303 ((1<<XPT_CONN)|(1<<XPT_DATA)|(1<<XPT_CLOSE)|(1<<XPT_DEFERRED))))
304 return;
Tom Tucker0f0257e2007-12-30 21:08:27 -0600305
306 cpu = get_cpu();
307 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
308 put_cpu();
309
310 spin_lock_bh(&pool->sp_lock);
311
312 if (!list_empty(&pool->sp_threads) &&
313 !list_empty(&pool->sp_sockets))
314 printk(KERN_ERR
315 "svc_xprt_enqueue: "
316 "threads and transports both waiting??\n");
317
318 if (test_bit(XPT_DEAD, &xprt->xpt_flags)) {
319 /* Don't enqueue dead transports */
320 dprintk("svc: transport %p is dead, not enqueued\n", xprt);
321 goto out_unlock;
322 }
323
324 /* Mark transport as busy. It will remain in this state until
325 * the provider calls svc_xprt_received. We update XPT_BUSY
326 * atomically because it also guards against trying to enqueue
327 * the transport twice.
328 */
329 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
330 /* Don't enqueue transport while already enqueued */
331 dprintk("svc: transport %p busy, not enqueued\n", xprt);
332 goto out_unlock;
333 }
334 BUG_ON(xprt->xpt_pool != NULL);
335 xprt->xpt_pool = pool;
336
337 /* Handle pending connection */
338 if (test_bit(XPT_CONN, &xprt->xpt_flags))
339 goto process;
340
341 /* Handle close in-progress */
342 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
343 goto process;
344
345 /* Check if we have space to reply to a request */
346 if (!xprt->xpt_ops->xpo_has_wspace(xprt)) {
347 /* Don't enqueue while not enough space for reply */
348 dprintk("svc: no write space, transport %p not enqueued\n",
349 xprt);
350 xprt->xpt_pool = NULL;
351 clear_bit(XPT_BUSY, &xprt->xpt_flags);
352 goto out_unlock;
353 }
354
355 process:
356 if (!list_empty(&pool->sp_threads)) {
357 rqstp = list_entry(pool->sp_threads.next,
358 struct svc_rqst,
359 rq_list);
360 dprintk("svc: transport %p served by daemon %p\n",
361 xprt, rqstp);
362 svc_thread_dequeue(pool, rqstp);
363 if (rqstp->rq_xprt)
364 printk(KERN_ERR
365 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
366 rqstp, rqstp->rq_xprt);
367 rqstp->rq_xprt = xprt;
368 svc_xprt_get(xprt);
369 rqstp->rq_reserved = serv->sv_max_mesg;
370 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
371 BUG_ON(xprt->xpt_pool != pool);
372 wake_up(&rqstp->rq_wait);
373 } else {
374 dprintk("svc: transport %p put into queue\n", xprt);
375 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
376 BUG_ON(xprt->xpt_pool != pool);
377 }
378
379out_unlock:
380 spin_unlock_bh(&pool->sp_lock);
381}
382EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
383
384/*
385 * Dequeue the first transport. Must be called with the pool->sp_lock held.
386 */
387static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
388{
389 struct svc_xprt *xprt;
390
391 if (list_empty(&pool->sp_sockets))
392 return NULL;
393
394 xprt = list_entry(pool->sp_sockets.next,
395 struct svc_xprt, xpt_ready);
396 list_del_init(&xprt->xpt_ready);
397
398 dprintk("svc: transport %p dequeued, inuse=%d\n",
399 xprt, atomic_read(&xprt->xpt_ref.refcount));
400
401 return xprt;
402}
403
404/*
405 * svc_xprt_received conditionally queues the transport for processing
406 * by another thread. The caller must hold the XPT_BUSY bit and must
407 * not thereafter touch transport data.
408 *
409 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
410 * insufficient) data.
411 */
412void svc_xprt_received(struct svc_xprt *xprt)
413{
414 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
415 xprt->xpt_pool = NULL;
416 clear_bit(XPT_BUSY, &xprt->xpt_flags);
417 svc_xprt_enqueue(xprt);
418}
419EXPORT_SYMBOL_GPL(svc_xprt_received);
420
421/**
422 * svc_reserve - change the space reserved for the reply to a request.
423 * @rqstp: The request in question
424 * @space: new max space to reserve
425 *
426 * Each request reserves some space on the output queue of the transport
427 * to make sure the reply fits. This function reduces that reserved
428 * space to be the amount of space used already, plus @space.
429 *
430 */
431void svc_reserve(struct svc_rqst *rqstp, int space)
432{
433 space += rqstp->rq_res.head[0].iov_len;
434
435 if (space < rqstp->rq_reserved) {
436 struct svc_xprt *xprt = rqstp->rq_xprt;
437 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
438 rqstp->rq_reserved = space;
439
440 svc_xprt_enqueue(xprt);
441 }
442}
Trond Myklebustd2f7e792007-07-14 15:39:58 -0400443EXPORT_SYMBOL(svc_reserve);
Tom Tucker0f0257e2007-12-30 21:08:27 -0600444
445static void svc_xprt_release(struct svc_rqst *rqstp)
446{
447 struct svc_xprt *xprt = rqstp->rq_xprt;
448
449 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
450
451 svc_free_res_pages(rqstp);
452 rqstp->rq_res.page_len = 0;
453 rqstp->rq_res.page_base = 0;
454
455 /* Reset response buffer and release
456 * the reservation.
457 * But first, check that enough space was reserved
458 * for the reply, otherwise we have a bug!
459 */
460 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
461 printk(KERN_ERR "RPC request reserved %d but used %d\n",
462 rqstp->rq_reserved,
463 rqstp->rq_res.len);
464
465 rqstp->rq_res.head[0].iov_len = 0;
466 svc_reserve(rqstp, 0);
467 rqstp->rq_xprt = NULL;
468
469 svc_xprt_put(xprt);
470}
471
472/*
473 * External function to wake up a server waiting for data
474 * This really only makes sense for services like lockd
475 * which have exactly one thread anyway.
476 */
477void svc_wake_up(struct svc_serv *serv)
478{
479 struct svc_rqst *rqstp;
480 unsigned int i;
481 struct svc_pool *pool;
482
483 for (i = 0; i < serv->sv_nrpools; i++) {
484 pool = &serv->sv_pools[i];
485
486 spin_lock_bh(&pool->sp_lock);
487 if (!list_empty(&pool->sp_threads)) {
488 rqstp = list_entry(pool->sp_threads.next,
489 struct svc_rqst,
490 rq_list);
491 dprintk("svc: daemon %p woken up.\n", rqstp);
492 /*
493 svc_thread_dequeue(pool, rqstp);
494 rqstp->rq_xprt = NULL;
495 */
496 wake_up(&rqstp->rq_wait);
497 }
498 spin_unlock_bh(&pool->sp_lock);
499 }
500}
Trond Myklebustd2f7e792007-07-14 15:39:58 -0400501EXPORT_SYMBOL(svc_wake_up);
Tom Tucker0f0257e2007-12-30 21:08:27 -0600502
503int svc_port_is_privileged(struct sockaddr *sin)
504{
505 switch (sin->sa_family) {
506 case AF_INET:
507 return ntohs(((struct sockaddr_in *)sin)->sin_port)
508 < PROT_SOCK;
509 case AF_INET6:
510 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
511 < PROT_SOCK;
512 default:
513 return 0;
514 }
515}
516
517/*
518 * Make sure that we don't have too many active connections. If we
519 * have, something must be dropped.
520 *
521 * There's no point in trying to do random drop here for DoS
522 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
523 * attacker can easily beat that.
524 *
525 * The only somewhat efficient mechanism would be if drop old
526 * connections from the same IP first. But right now we don't even
527 * record the client IP in svc_sock.
528 */
529static void svc_check_conn_limits(struct svc_serv *serv)
530{
531 if (serv->sv_tmpcnt > (serv->sv_nrthreads+3)*20) {
532 struct svc_xprt *xprt = NULL;
533 spin_lock_bh(&serv->sv_lock);
534 if (!list_empty(&serv->sv_tempsocks)) {
535 if (net_ratelimit()) {
536 /* Try to help the admin */
537 printk(KERN_NOTICE "%s: too many open "
538 "connections, consider increasing the "
539 "number of nfsd threads\n",
540 serv->sv_name);
541 }
542 /*
543 * Always select the oldest connection. It's not fair,
544 * but so is life
545 */
546 xprt = list_entry(serv->sv_tempsocks.prev,
547 struct svc_xprt,
548 xpt_list);
549 set_bit(XPT_CLOSE, &xprt->xpt_flags);
550 svc_xprt_get(xprt);
551 }
552 spin_unlock_bh(&serv->sv_lock);
553
554 if (xprt) {
555 svc_xprt_enqueue(xprt);
556 svc_xprt_put(xprt);
557 }
558 }
559}
560
561/*
562 * Receive the next request on any transport. This code is carefully
563 * organised not to touch any cachelines in the shared svc_serv
564 * structure, only cachelines in the local svc_pool.
565 */
566int svc_recv(struct svc_rqst *rqstp, long timeout)
567{
568 struct svc_xprt *xprt = NULL;
569 struct svc_serv *serv = rqstp->rq_server;
570 struct svc_pool *pool = rqstp->rq_pool;
571 int len, i;
572 int pages;
573 struct xdr_buf *arg;
574 DECLARE_WAITQUEUE(wait, current);
575
576 dprintk("svc: server %p waiting for data (to = %ld)\n",
577 rqstp, timeout);
578
579 if (rqstp->rq_xprt)
580 printk(KERN_ERR
581 "svc_recv: service %p, transport not NULL!\n",
582 rqstp);
583 if (waitqueue_active(&rqstp->rq_wait))
584 printk(KERN_ERR
585 "svc_recv: service %p, wait queue active!\n",
586 rqstp);
587
588 /* now allocate needed pages. If we get a failure, sleep briefly */
589 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
590 for (i = 0; i < pages ; i++)
591 while (rqstp->rq_pages[i] == NULL) {
592 struct page *p = alloc_page(GFP_KERNEL);
593 if (!p) {
Jeff Layton7b54fe62008-02-12 11:47:24 -0500594 set_current_state(TASK_INTERRUPTIBLE);
595 if (signalled() || kthread_should_stop()) {
596 set_current_state(TASK_RUNNING);
Jeff Layton70867212008-02-07 16:34:54 -0500597 return -EINTR;
Jeff Layton7b54fe62008-02-12 11:47:24 -0500598 }
599 schedule_timeout(msecs_to_jiffies(500));
Tom Tucker0f0257e2007-12-30 21:08:27 -0600600 }
601 rqstp->rq_pages[i] = p;
602 }
603 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
604 BUG_ON(pages >= RPCSVC_MAXPAGES);
605
606 /* Make arg->head point to first page and arg->pages point to rest */
607 arg = &rqstp->rq_arg;
608 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
609 arg->head[0].iov_len = PAGE_SIZE;
610 arg->pages = rqstp->rq_pages + 1;
611 arg->page_base = 0;
612 /* save at least one page for response */
613 arg->page_len = (pages-2)*PAGE_SIZE;
614 arg->len = (pages-1)*PAGE_SIZE;
615 arg->tail[0].iov_len = 0;
616
617 try_to_freeze();
618 cond_resched();
Jeff Layton70867212008-02-07 16:34:54 -0500619 if (signalled() || kthread_should_stop())
Tom Tucker0f0257e2007-12-30 21:08:27 -0600620 return -EINTR;
621
622 spin_lock_bh(&pool->sp_lock);
623 xprt = svc_xprt_dequeue(pool);
624 if (xprt) {
625 rqstp->rq_xprt = xprt;
626 svc_xprt_get(xprt);
627 rqstp->rq_reserved = serv->sv_max_mesg;
628 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
629 } else {
630 /* No data pending. Go to sleep */
631 svc_thread_enqueue(pool, rqstp);
632
633 /*
634 * We have to be able to interrupt this wait
635 * to bring down the daemons ...
636 */
637 set_current_state(TASK_INTERRUPTIBLE);
Jeff Layton70867212008-02-07 16:34:54 -0500638
639 /*
640 * checking kthread_should_stop() here allows us to avoid
641 * locking and signalling when stopping kthreads that call
642 * svc_recv. If the thread has already been woken up, then
643 * we can exit here without sleeping. If not, then it
644 * it'll be woken up quickly during the schedule_timeout
645 */
646 if (kthread_should_stop()) {
647 set_current_state(TASK_RUNNING);
648 spin_unlock_bh(&pool->sp_lock);
649 return -EINTR;
650 }
651
Tom Tucker0f0257e2007-12-30 21:08:27 -0600652 add_wait_queue(&rqstp->rq_wait, &wait);
653 spin_unlock_bh(&pool->sp_lock);
654
655 schedule_timeout(timeout);
656
657 try_to_freeze();
658
659 spin_lock_bh(&pool->sp_lock);
660 remove_wait_queue(&rqstp->rq_wait, &wait);
661
662 xprt = rqstp->rq_xprt;
663 if (!xprt) {
664 svc_thread_dequeue(pool, rqstp);
665 spin_unlock_bh(&pool->sp_lock);
666 dprintk("svc: server %p, no data yet\n", rqstp);
Jeff Layton70867212008-02-07 16:34:54 -0500667 if (signalled() || kthread_should_stop())
668 return -EINTR;
669 else
670 return -EAGAIN;
Tom Tucker0f0257e2007-12-30 21:08:27 -0600671 }
672 }
673 spin_unlock_bh(&pool->sp_lock);
674
675 len = 0;
676 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
677 dprintk("svc_recv: found XPT_CLOSE\n");
678 svc_delete_xprt(xprt);
679 } else if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
680 struct svc_xprt *newxpt;
681 newxpt = xprt->xpt_ops->xpo_accept(xprt);
682 if (newxpt) {
683 /*
684 * We know this module_get will succeed because the
685 * listener holds a reference too
686 */
687 __module_get(newxpt->xpt_class->xcl_owner);
688 svc_check_conn_limits(xprt->xpt_server);
689 spin_lock_bh(&serv->sv_lock);
690 set_bit(XPT_TEMP, &newxpt->xpt_flags);
691 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
692 serv->sv_tmpcnt++;
693 if (serv->sv_temptimer.function == NULL) {
694 /* setup timer to age temp transports */
695 setup_timer(&serv->sv_temptimer,
696 svc_age_temp_xprts,
697 (unsigned long)serv);
698 mod_timer(&serv->sv_temptimer,
699 jiffies + svc_conn_age_period * HZ);
700 }
701 spin_unlock_bh(&serv->sv_lock);
702 svc_xprt_received(newxpt);
703 }
704 svc_xprt_received(xprt);
705 } else {
706 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
707 rqstp, pool->sp_id, xprt,
708 atomic_read(&xprt->xpt_ref.refcount));
709 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
710 if (rqstp->rq_deferred) {
711 svc_xprt_received(xprt);
712 len = svc_deferred_recv(rqstp);
713 } else
714 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
715 dprintk("svc: got len=%d\n", len);
716 }
717
718 /* No data, incomplete (TCP) read, or accept() */
719 if (len == 0 || len == -EAGAIN) {
720 rqstp->rq_res.len = 0;
721 svc_xprt_release(rqstp);
722 return -EAGAIN;
723 }
724 clear_bit(XPT_OLD, &xprt->xpt_flags);
725
726 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
727 rqstp->rq_chandle.defer = svc_defer;
728
729 if (serv->sv_stats)
730 serv->sv_stats->netcnt++;
731 return len;
732}
Trond Myklebustd2f7e792007-07-14 15:39:58 -0400733EXPORT_SYMBOL(svc_recv);
Tom Tucker0f0257e2007-12-30 21:08:27 -0600734
735/*
736 * Drop request
737 */
738void svc_drop(struct svc_rqst *rqstp)
739{
740 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
741 svc_xprt_release(rqstp);
742}
Trond Myklebustd2f7e792007-07-14 15:39:58 -0400743EXPORT_SYMBOL(svc_drop);
Tom Tucker0f0257e2007-12-30 21:08:27 -0600744
745/*
746 * Return reply to client.
747 */
748int svc_send(struct svc_rqst *rqstp)
749{
750 struct svc_xprt *xprt;
751 int len;
752 struct xdr_buf *xb;
753
754 xprt = rqstp->rq_xprt;
755 if (!xprt)
756 return -EFAULT;
757
758 /* release the receive skb before sending the reply */
759 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
760
761 /* calculate over-all length */
762 xb = &rqstp->rq_res;
763 xb->len = xb->head[0].iov_len +
764 xb->page_len +
765 xb->tail[0].iov_len;
766
767 /* Grab mutex to serialize outgoing data. */
768 mutex_lock(&xprt->xpt_mutex);
769 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
770 len = -ENOTCONN;
771 else
772 len = xprt->xpt_ops->xpo_sendto(rqstp);
773 mutex_unlock(&xprt->xpt_mutex);
774 svc_xprt_release(rqstp);
775
776 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
777 return 0;
778 return len;
779}
780
781/*
782 * Timer function to close old temporary transports, using
783 * a mark-and-sweep algorithm.
784 */
785static void svc_age_temp_xprts(unsigned long closure)
786{
787 struct svc_serv *serv = (struct svc_serv *)closure;
788 struct svc_xprt *xprt;
789 struct list_head *le, *next;
790 LIST_HEAD(to_be_aged);
791
792 dprintk("svc_age_temp_xprts\n");
793
794 if (!spin_trylock_bh(&serv->sv_lock)) {
795 /* busy, try again 1 sec later */
796 dprintk("svc_age_temp_xprts: busy\n");
797 mod_timer(&serv->sv_temptimer, jiffies + HZ);
798 return;
799 }
800
801 list_for_each_safe(le, next, &serv->sv_tempsocks) {
802 xprt = list_entry(le, struct svc_xprt, xpt_list);
803
804 /* First time through, just mark it OLD. Second time
805 * through, close it. */
806 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
807 continue;
808 if (atomic_read(&xprt->xpt_ref.refcount) > 1
809 || test_bit(XPT_BUSY, &xprt->xpt_flags))
810 continue;
811 svc_xprt_get(xprt);
812 list_move(le, &to_be_aged);
813 set_bit(XPT_CLOSE, &xprt->xpt_flags);
814 set_bit(XPT_DETACHED, &xprt->xpt_flags);
815 }
816 spin_unlock_bh(&serv->sv_lock);
817
818 while (!list_empty(&to_be_aged)) {
819 le = to_be_aged.next;
820 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
821 list_del_init(le);
822 xprt = list_entry(le, struct svc_xprt, xpt_list);
823
824 dprintk("queuing xprt %p for closing\n", xprt);
825
826 /* a thread will dequeue and close it soon */
827 svc_xprt_enqueue(xprt);
828 svc_xprt_put(xprt);
829 }
830
831 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
832}
833
834/*
835 * Remove a dead transport
836 */
837void svc_delete_xprt(struct svc_xprt *xprt)
838{
839 struct svc_serv *serv = xprt->xpt_server;
840
841 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
842 xprt->xpt_ops->xpo_detach(xprt);
843
844 spin_lock_bh(&serv->sv_lock);
845 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
846 list_del_init(&xprt->xpt_list);
847 /*
848 * We used to delete the transport from whichever list
849 * it's sk_xprt.xpt_ready node was on, but we don't actually
850 * need to. This is because the only time we're called
851 * while still attached to a queue, the queue itself
852 * is about to be destroyed (in svc_destroy).
853 */
854 if (!test_and_set_bit(XPT_DEAD, &xprt->xpt_flags)) {
855 BUG_ON(atomic_read(&xprt->xpt_ref.refcount) < 2);
856 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
857 serv->sv_tmpcnt--;
858 svc_xprt_put(xprt);
859 }
860 spin_unlock_bh(&serv->sv_lock);
861}
862
863void svc_close_xprt(struct svc_xprt *xprt)
864{
865 set_bit(XPT_CLOSE, &xprt->xpt_flags);
866 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
867 /* someone else will have to effect the close */
868 return;
869
870 svc_xprt_get(xprt);
871 svc_delete_xprt(xprt);
872 clear_bit(XPT_BUSY, &xprt->xpt_flags);
873 svc_xprt_put(xprt);
874}
Tom Tuckera2178132007-12-30 21:08:35 -0600875EXPORT_SYMBOL_GPL(svc_close_xprt);
Tom Tucker0f0257e2007-12-30 21:08:27 -0600876
877void svc_close_all(struct list_head *xprt_list)
878{
879 struct svc_xprt *xprt;
880 struct svc_xprt *tmp;
881
882 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
883 set_bit(XPT_CLOSE, &xprt->xpt_flags);
884 if (test_bit(XPT_BUSY, &xprt->xpt_flags)) {
885 /* Waiting to be processed, but no threads left,
886 * So just remove it from the waiting list
887 */
888 list_del_init(&xprt->xpt_ready);
889 clear_bit(XPT_BUSY, &xprt->xpt_flags);
890 }
891 svc_close_xprt(xprt);
892 }
893}
894
895/*
896 * Handle defer and revisit of requests
897 */
898
899static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
900{
901 struct svc_deferred_req *dr =
902 container_of(dreq, struct svc_deferred_req, handle);
903 struct svc_xprt *xprt = dr->xprt;
904
905 if (too_many) {
906 svc_xprt_put(xprt);
907 kfree(dr);
908 return;
909 }
910 dprintk("revisit queued\n");
911 dr->xprt = NULL;
912 spin_lock(&xprt->xpt_lock);
913 list_add(&dr->handle.recent, &xprt->xpt_deferred);
914 spin_unlock(&xprt->xpt_lock);
915 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
916 svc_xprt_enqueue(xprt);
917 svc_xprt_put(xprt);
918}
919
Tom Tucker260c1d12007-12-30 21:08:29 -0600920/*
921 * Save the request off for later processing. The request buffer looks
922 * like this:
923 *
924 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
925 *
926 * This code can only handle requests that consist of an xprt-header
927 * and rpc-header.
928 */
Tom Tucker0f0257e2007-12-30 21:08:27 -0600929static struct cache_deferred_req *svc_defer(struct cache_req *req)
930{
931 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
Tom Tucker0f0257e2007-12-30 21:08:27 -0600932 struct svc_deferred_req *dr;
933
934 if (rqstp->rq_arg.page_len)
935 return NULL; /* if more than a page, give up FIXME */
936 if (rqstp->rq_deferred) {
937 dr = rqstp->rq_deferred;
938 rqstp->rq_deferred = NULL;
939 } else {
Tom Tucker260c1d12007-12-30 21:08:29 -0600940 size_t skip;
941 size_t size;
Tom Tucker0f0257e2007-12-30 21:08:27 -0600942 /* FIXME maybe discard if size too large */
Tom Tucker260c1d12007-12-30 21:08:29 -0600943 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
Tom Tucker0f0257e2007-12-30 21:08:27 -0600944 dr = kmalloc(size, GFP_KERNEL);
945 if (dr == NULL)
946 return NULL;
947
948 dr->handle.owner = rqstp->rq_server;
949 dr->prot = rqstp->rq_prot;
950 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
951 dr->addrlen = rqstp->rq_addrlen;
952 dr->daddr = rqstp->rq_daddr;
953 dr->argslen = rqstp->rq_arg.len >> 2;
Tom Tucker260c1d12007-12-30 21:08:29 -0600954 dr->xprt_hlen = rqstp->rq_xprt_hlen;
955
956 /* back up head to the start of the buffer and copy */
957 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
958 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
959 dr->argslen << 2);
Tom Tucker0f0257e2007-12-30 21:08:27 -0600960 }
961 svc_xprt_get(rqstp->rq_xprt);
962 dr->xprt = rqstp->rq_xprt;
963
964 dr->handle.revisit = svc_revisit;
965 return &dr->handle;
966}
967
968/*
969 * recv data from a deferred request into an active one
970 */
971static int svc_deferred_recv(struct svc_rqst *rqstp)
972{
973 struct svc_deferred_req *dr = rqstp->rq_deferred;
974
Tom Tucker260c1d12007-12-30 21:08:29 -0600975 /* setup iov_base past transport header */
976 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
977 /* The iov_len does not include the transport header bytes */
978 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
Tom Tucker0f0257e2007-12-30 21:08:27 -0600979 rqstp->rq_arg.page_len = 0;
Tom Tucker260c1d12007-12-30 21:08:29 -0600980 /* The rq_arg.len includes the transport header bytes */
981 rqstp->rq_arg.len = dr->argslen<<2;
Tom Tucker0f0257e2007-12-30 21:08:27 -0600982 rqstp->rq_prot = dr->prot;
983 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
984 rqstp->rq_addrlen = dr->addrlen;
Tom Tucker260c1d12007-12-30 21:08:29 -0600985 /* Save off transport header len in case we get deferred again */
986 rqstp->rq_xprt_hlen = dr->xprt_hlen;
Tom Tucker0f0257e2007-12-30 21:08:27 -0600987 rqstp->rq_daddr = dr->daddr;
988 rqstp->rq_respages = rqstp->rq_pages;
Tom Tucker260c1d12007-12-30 21:08:29 -0600989 return (dr->argslen<<2) - dr->xprt_hlen;
Tom Tucker0f0257e2007-12-30 21:08:27 -0600990}
991
992
993static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
994{
995 struct svc_deferred_req *dr = NULL;
996
997 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
998 return NULL;
999 spin_lock(&xprt->xpt_lock);
1000 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1001 if (!list_empty(&xprt->xpt_deferred)) {
1002 dr = list_entry(xprt->xpt_deferred.next,
1003 struct svc_deferred_req,
1004 handle.recent);
1005 list_del_init(&dr->handle.recent);
1006 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1007 }
1008 spin_unlock(&xprt->xpt_lock);
1009 return dr;
1010}
Tom Tucker7fcb98d2007-12-30 21:08:33 -06001011
1012/*
1013 * Return the transport instance pointer for the endpoint accepting
1014 * connections/peer traffic from the specified transport class,
1015 * address family and port.
1016 *
1017 * Specifying 0 for the address family or port is effectively a
1018 * wild-card, and will result in matching the first transport in the
1019 * service's list that has a matching class name.
1020 */
1021struct svc_xprt *svc_find_xprt(struct svc_serv *serv, char *xcl_name,
1022 int af, int port)
1023{
1024 struct svc_xprt *xprt;
1025 struct svc_xprt *found = NULL;
1026
1027 /* Sanity check the args */
1028 if (!serv || !xcl_name)
1029 return found;
1030
1031 spin_lock_bh(&serv->sv_lock);
1032 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1033 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1034 continue;
1035 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1036 continue;
1037 if (port && port != svc_xprt_local_port(xprt))
1038 continue;
1039 found = xprt;
Tom Tuckera2178132007-12-30 21:08:35 -06001040 svc_xprt_get(xprt);
Tom Tucker7fcb98d2007-12-30 21:08:33 -06001041 break;
1042 }
1043 spin_unlock_bh(&serv->sv_lock);
1044 return found;
1045}
1046EXPORT_SYMBOL_GPL(svc_find_xprt);
Tom Tucker9571af12007-12-30 21:08:37 -06001047
1048/*
1049 * Format a buffer with a list of the active transports. A zero for
1050 * the buflen parameter disables target buffer overflow checking.
1051 */
1052int svc_xprt_names(struct svc_serv *serv, char *buf, int buflen)
1053{
1054 struct svc_xprt *xprt;
1055 char xprt_str[64];
1056 int totlen = 0;
1057 int len;
1058
1059 /* Sanity check args */
1060 if (!serv)
1061 return 0;
1062
1063 spin_lock_bh(&serv->sv_lock);
1064 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1065 len = snprintf(xprt_str, sizeof(xprt_str),
1066 "%s %d\n", xprt->xpt_class->xcl_name,
1067 svc_xprt_local_port(xprt));
1068 /* If the string was truncated, replace with error string */
1069 if (len >= sizeof(xprt_str))
1070 strcpy(xprt_str, "name-too-long\n");
1071 /* Don't overflow buffer */
1072 len = strlen(xprt_str);
1073 if (buflen && (len + totlen >= buflen))
1074 break;
1075 strcpy(buf+totlen, xprt_str);
1076 totlen += len;
1077 }
1078 spin_unlock_bh(&serv->sv_lock);
1079 return totlen;
1080}
1081EXPORT_SYMBOL_GPL(svc_xprt_names);