blob: e46c825f49548923f79f1363d4a7712ef4ead513 [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
162int svc_create_xprt(struct svc_serv *serv, char *xprt_name, unsigned short port,
163 int flags)
164{
165 struct svc_xprt_class *xcl;
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 };
171 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
172 spin_lock(&svc_xprt_class_lock);
173 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
Tom Tucker4e5caaa2007-12-30 21:08:20 -0600174 struct svc_xprt *newxprt;
175
176 if (strcmp(xprt_name, xcl->xcl_name))
177 continue;
178
179 if (!try_module_get(xcl->xcl_owner))
180 goto err;
181
182 spin_unlock(&svc_xprt_class_lock);
183 newxprt = xcl->xcl_ops->
184 xpo_create(serv, (struct sockaddr *)&sin, sizeof(sin),
185 flags);
186 if (IS_ERR(newxprt)) {
187 module_put(xcl->xcl_owner);
188 return PTR_ERR(newxprt);
Tom Tuckerb700cbb2007-12-30 21:07:42 -0600189 }
Tom Tucker4e5caaa2007-12-30 21:08:20 -0600190
191 clear_bit(XPT_TEMP, &newxprt->xpt_flags);
192 spin_lock_bh(&serv->sv_lock);
193 list_add(&newxprt->xpt_list, &serv->sv_permsocks);
194 spin_unlock_bh(&serv->sv_lock);
195 clear_bit(XPT_BUSY, &newxprt->xpt_flags);
196 return svc_xprt_local_port(newxprt);
Tom Tuckerb700cbb2007-12-30 21:07:42 -0600197 }
Tom Tucker4e5caaa2007-12-30 21:08:20 -0600198 err:
Tom Tuckerb700cbb2007-12-30 21:07:42 -0600199 spin_unlock(&svc_xprt_class_lock);
200 dprintk("svc: transport %s not found\n", xprt_name);
Tom Tucker4e5caaa2007-12-30 21:08:20 -0600201 return -ENOENT;
Tom Tuckerb700cbb2007-12-30 21:07:42 -0600202}
203EXPORT_SYMBOL_GPL(svc_create_xprt);
Tom Tucker9dbc2402007-12-30 21:08:12 -0600204
205/*
206 * Copy the local and remote xprt addresses to the rqstp structure
207 */
208void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
209{
210 struct sockaddr *sin;
211
212 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
213 rqstp->rq_addrlen = xprt->xpt_remotelen;
214
215 /*
216 * Destination address in request is needed for binding the
217 * source address in RPC replies/callbacks later.
218 */
219 sin = (struct sockaddr *)&xprt->xpt_local;
220 switch (sin->sa_family) {
221 case AF_INET:
222 rqstp->rq_daddr.addr = ((struct sockaddr_in *)sin)->sin_addr;
223 break;
224 case AF_INET6:
225 rqstp->rq_daddr.addr6 = ((struct sockaddr_in6 *)sin)->sin6_addr;
226 break;
227 }
228}
229EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
230
Tom Tucker0f0257e2007-12-30 21:08:27 -0600231/**
232 * svc_print_addr - Format rq_addr field for printing
233 * @rqstp: svc_rqst struct containing address to print
234 * @buf: target buffer for formatted address
235 * @len: length of target buffer
236 *
237 */
238char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
239{
240 return __svc_print_addr(svc_addr(rqstp), buf, len);
241}
242EXPORT_SYMBOL_GPL(svc_print_addr);
243
244/*
245 * Queue up an idle server thread. Must have pool->sp_lock held.
246 * Note: this is really a stack rather than a queue, so that we only
247 * use as many different threads as we need, and the rest don't pollute
248 * the cache.
249 */
250static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
251{
252 list_add(&rqstp->rq_list, &pool->sp_threads);
253}
254
255/*
256 * Dequeue an nfsd thread. Must have pool->sp_lock held.
257 */
258static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
259{
260 list_del(&rqstp->rq_list);
261}
262
263/*
264 * Queue up a transport with data pending. If there are idle nfsd
265 * processes, wake 'em up.
266 *
267 */
268void svc_xprt_enqueue(struct svc_xprt *xprt)
269{
270 struct svc_serv *serv = xprt->xpt_server;
271 struct svc_pool *pool;
272 struct svc_rqst *rqstp;
273 int cpu;
274
275 if (!(xprt->xpt_flags &
276 ((1<<XPT_CONN)|(1<<XPT_DATA)|(1<<XPT_CLOSE)|(1<<XPT_DEFERRED))))
277 return;
Tom Tucker0f0257e2007-12-30 21:08:27 -0600278
279 cpu = get_cpu();
280 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
281 put_cpu();
282
283 spin_lock_bh(&pool->sp_lock);
284
285 if (!list_empty(&pool->sp_threads) &&
286 !list_empty(&pool->sp_sockets))
287 printk(KERN_ERR
288 "svc_xprt_enqueue: "
289 "threads and transports both waiting??\n");
290
291 if (test_bit(XPT_DEAD, &xprt->xpt_flags)) {
292 /* Don't enqueue dead transports */
293 dprintk("svc: transport %p is dead, not enqueued\n", xprt);
294 goto out_unlock;
295 }
296
297 /* Mark transport as busy. It will remain in this state until
298 * the provider calls svc_xprt_received. We update XPT_BUSY
299 * atomically because it also guards against trying to enqueue
300 * the transport twice.
301 */
302 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
303 /* Don't enqueue transport while already enqueued */
304 dprintk("svc: transport %p busy, not enqueued\n", xprt);
305 goto out_unlock;
306 }
307 BUG_ON(xprt->xpt_pool != NULL);
308 xprt->xpt_pool = pool;
309
310 /* Handle pending connection */
311 if (test_bit(XPT_CONN, &xprt->xpt_flags))
312 goto process;
313
314 /* Handle close in-progress */
315 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
316 goto process;
317
318 /* Check if we have space to reply to a request */
319 if (!xprt->xpt_ops->xpo_has_wspace(xprt)) {
320 /* Don't enqueue while not enough space for reply */
321 dprintk("svc: no write space, transport %p not enqueued\n",
322 xprt);
323 xprt->xpt_pool = NULL;
324 clear_bit(XPT_BUSY, &xprt->xpt_flags);
325 goto out_unlock;
326 }
327
328 process:
329 if (!list_empty(&pool->sp_threads)) {
330 rqstp = list_entry(pool->sp_threads.next,
331 struct svc_rqst,
332 rq_list);
333 dprintk("svc: transport %p served by daemon %p\n",
334 xprt, rqstp);
335 svc_thread_dequeue(pool, rqstp);
336 if (rqstp->rq_xprt)
337 printk(KERN_ERR
338 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
339 rqstp, rqstp->rq_xprt);
340 rqstp->rq_xprt = xprt;
341 svc_xprt_get(xprt);
342 rqstp->rq_reserved = serv->sv_max_mesg;
343 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
344 BUG_ON(xprt->xpt_pool != pool);
345 wake_up(&rqstp->rq_wait);
346 } else {
347 dprintk("svc: transport %p put into queue\n", xprt);
348 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
349 BUG_ON(xprt->xpt_pool != pool);
350 }
351
352out_unlock:
353 spin_unlock_bh(&pool->sp_lock);
354}
355EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
356
357/*
358 * Dequeue the first transport. Must be called with the pool->sp_lock held.
359 */
360static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
361{
362 struct svc_xprt *xprt;
363
364 if (list_empty(&pool->sp_sockets))
365 return NULL;
366
367 xprt = list_entry(pool->sp_sockets.next,
368 struct svc_xprt, xpt_ready);
369 list_del_init(&xprt->xpt_ready);
370
371 dprintk("svc: transport %p dequeued, inuse=%d\n",
372 xprt, atomic_read(&xprt->xpt_ref.refcount));
373
374 return xprt;
375}
376
377/*
378 * svc_xprt_received conditionally queues the transport for processing
379 * by another thread. The caller must hold the XPT_BUSY bit and must
380 * not thereafter touch transport data.
381 *
382 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
383 * insufficient) data.
384 */
385void svc_xprt_received(struct svc_xprt *xprt)
386{
387 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
388 xprt->xpt_pool = NULL;
389 clear_bit(XPT_BUSY, &xprt->xpt_flags);
390 svc_xprt_enqueue(xprt);
391}
392EXPORT_SYMBOL_GPL(svc_xprt_received);
393
394/**
395 * svc_reserve - change the space reserved for the reply to a request.
396 * @rqstp: The request in question
397 * @space: new max space to reserve
398 *
399 * Each request reserves some space on the output queue of the transport
400 * to make sure the reply fits. This function reduces that reserved
401 * space to be the amount of space used already, plus @space.
402 *
403 */
404void svc_reserve(struct svc_rqst *rqstp, int space)
405{
406 space += rqstp->rq_res.head[0].iov_len;
407
408 if (space < rqstp->rq_reserved) {
409 struct svc_xprt *xprt = rqstp->rq_xprt;
410 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
411 rqstp->rq_reserved = space;
412
413 svc_xprt_enqueue(xprt);
414 }
415}
Trond Myklebustd2f7e792007-07-14 15:39:58 -0400416EXPORT_SYMBOL(svc_reserve);
Tom Tucker0f0257e2007-12-30 21:08:27 -0600417
418static void svc_xprt_release(struct svc_rqst *rqstp)
419{
420 struct svc_xprt *xprt = rqstp->rq_xprt;
421
422 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
423
424 svc_free_res_pages(rqstp);
425 rqstp->rq_res.page_len = 0;
426 rqstp->rq_res.page_base = 0;
427
428 /* Reset response buffer and release
429 * the reservation.
430 * But first, check that enough space was reserved
431 * for the reply, otherwise we have a bug!
432 */
433 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
434 printk(KERN_ERR "RPC request reserved %d but used %d\n",
435 rqstp->rq_reserved,
436 rqstp->rq_res.len);
437
438 rqstp->rq_res.head[0].iov_len = 0;
439 svc_reserve(rqstp, 0);
440 rqstp->rq_xprt = NULL;
441
442 svc_xprt_put(xprt);
443}
444
445/*
446 * External function to wake up a server waiting for data
447 * This really only makes sense for services like lockd
448 * which have exactly one thread anyway.
449 */
450void svc_wake_up(struct svc_serv *serv)
451{
452 struct svc_rqst *rqstp;
453 unsigned int i;
454 struct svc_pool *pool;
455
456 for (i = 0; i < serv->sv_nrpools; i++) {
457 pool = &serv->sv_pools[i];
458
459 spin_lock_bh(&pool->sp_lock);
460 if (!list_empty(&pool->sp_threads)) {
461 rqstp = list_entry(pool->sp_threads.next,
462 struct svc_rqst,
463 rq_list);
464 dprintk("svc: daemon %p woken up.\n", rqstp);
465 /*
466 svc_thread_dequeue(pool, rqstp);
467 rqstp->rq_xprt = NULL;
468 */
469 wake_up(&rqstp->rq_wait);
470 }
471 spin_unlock_bh(&pool->sp_lock);
472 }
473}
Trond Myklebustd2f7e792007-07-14 15:39:58 -0400474EXPORT_SYMBOL(svc_wake_up);
Tom Tucker0f0257e2007-12-30 21:08:27 -0600475
476int svc_port_is_privileged(struct sockaddr *sin)
477{
478 switch (sin->sa_family) {
479 case AF_INET:
480 return ntohs(((struct sockaddr_in *)sin)->sin_port)
481 < PROT_SOCK;
482 case AF_INET6:
483 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
484 < PROT_SOCK;
485 default:
486 return 0;
487 }
488}
489
490/*
491 * Make sure that we don't have too many active connections. If we
492 * have, something must be dropped.
493 *
494 * There's no point in trying to do random drop here for DoS
495 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
496 * attacker can easily beat that.
497 *
498 * The only somewhat efficient mechanism would be if drop old
499 * connections from the same IP first. But right now we don't even
500 * record the client IP in svc_sock.
501 */
502static void svc_check_conn_limits(struct svc_serv *serv)
503{
504 if (serv->sv_tmpcnt > (serv->sv_nrthreads+3)*20) {
505 struct svc_xprt *xprt = NULL;
506 spin_lock_bh(&serv->sv_lock);
507 if (!list_empty(&serv->sv_tempsocks)) {
508 if (net_ratelimit()) {
509 /* Try to help the admin */
510 printk(KERN_NOTICE "%s: too many open "
511 "connections, consider increasing the "
512 "number of nfsd threads\n",
513 serv->sv_name);
514 }
515 /*
516 * Always select the oldest connection. It's not fair,
517 * but so is life
518 */
519 xprt = list_entry(serv->sv_tempsocks.prev,
520 struct svc_xprt,
521 xpt_list);
522 set_bit(XPT_CLOSE, &xprt->xpt_flags);
523 svc_xprt_get(xprt);
524 }
525 spin_unlock_bh(&serv->sv_lock);
526
527 if (xprt) {
528 svc_xprt_enqueue(xprt);
529 svc_xprt_put(xprt);
530 }
531 }
532}
533
534/*
535 * Receive the next request on any transport. This code is carefully
536 * organised not to touch any cachelines in the shared svc_serv
537 * structure, only cachelines in the local svc_pool.
538 */
539int svc_recv(struct svc_rqst *rqstp, long timeout)
540{
541 struct svc_xprt *xprt = NULL;
542 struct svc_serv *serv = rqstp->rq_server;
543 struct svc_pool *pool = rqstp->rq_pool;
544 int len, i;
545 int pages;
546 struct xdr_buf *arg;
547 DECLARE_WAITQUEUE(wait, current);
548
549 dprintk("svc: server %p waiting for data (to = %ld)\n",
550 rqstp, timeout);
551
552 if (rqstp->rq_xprt)
553 printk(KERN_ERR
554 "svc_recv: service %p, transport not NULL!\n",
555 rqstp);
556 if (waitqueue_active(&rqstp->rq_wait))
557 printk(KERN_ERR
558 "svc_recv: service %p, wait queue active!\n",
559 rqstp);
560
561 /* now allocate needed pages. If we get a failure, sleep briefly */
562 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
563 for (i = 0; i < pages ; i++)
564 while (rqstp->rq_pages[i] == NULL) {
565 struct page *p = alloc_page(GFP_KERNEL);
566 if (!p) {
Jeff Layton7b54fe62008-02-12 11:47:24 -0500567 set_current_state(TASK_INTERRUPTIBLE);
568 if (signalled() || kthread_should_stop()) {
569 set_current_state(TASK_RUNNING);
Jeff Layton70867212008-02-07 16:34:54 -0500570 return -EINTR;
Jeff Layton7b54fe62008-02-12 11:47:24 -0500571 }
572 schedule_timeout(msecs_to_jiffies(500));
Tom Tucker0f0257e2007-12-30 21:08:27 -0600573 }
574 rqstp->rq_pages[i] = p;
575 }
576 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
577 BUG_ON(pages >= RPCSVC_MAXPAGES);
578
579 /* Make arg->head point to first page and arg->pages point to rest */
580 arg = &rqstp->rq_arg;
581 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
582 arg->head[0].iov_len = PAGE_SIZE;
583 arg->pages = rqstp->rq_pages + 1;
584 arg->page_base = 0;
585 /* save at least one page for response */
586 arg->page_len = (pages-2)*PAGE_SIZE;
587 arg->len = (pages-1)*PAGE_SIZE;
588 arg->tail[0].iov_len = 0;
589
590 try_to_freeze();
591 cond_resched();
Jeff Layton70867212008-02-07 16:34:54 -0500592 if (signalled() || kthread_should_stop())
Tom Tucker0f0257e2007-12-30 21:08:27 -0600593 return -EINTR;
594
595 spin_lock_bh(&pool->sp_lock);
596 xprt = svc_xprt_dequeue(pool);
597 if (xprt) {
598 rqstp->rq_xprt = xprt;
599 svc_xprt_get(xprt);
600 rqstp->rq_reserved = serv->sv_max_mesg;
601 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
602 } else {
603 /* No data pending. Go to sleep */
604 svc_thread_enqueue(pool, rqstp);
605
606 /*
607 * We have to be able to interrupt this wait
608 * to bring down the daemons ...
609 */
610 set_current_state(TASK_INTERRUPTIBLE);
Jeff Layton70867212008-02-07 16:34:54 -0500611
612 /*
613 * checking kthread_should_stop() here allows us to avoid
614 * locking and signalling when stopping kthreads that call
615 * svc_recv. If the thread has already been woken up, then
616 * we can exit here without sleeping. If not, then it
617 * it'll be woken up quickly during the schedule_timeout
618 */
619 if (kthread_should_stop()) {
620 set_current_state(TASK_RUNNING);
621 spin_unlock_bh(&pool->sp_lock);
622 return -EINTR;
623 }
624
Tom Tucker0f0257e2007-12-30 21:08:27 -0600625 add_wait_queue(&rqstp->rq_wait, &wait);
626 spin_unlock_bh(&pool->sp_lock);
627
628 schedule_timeout(timeout);
629
630 try_to_freeze();
631
632 spin_lock_bh(&pool->sp_lock);
633 remove_wait_queue(&rqstp->rq_wait, &wait);
634
635 xprt = rqstp->rq_xprt;
636 if (!xprt) {
637 svc_thread_dequeue(pool, rqstp);
638 spin_unlock_bh(&pool->sp_lock);
639 dprintk("svc: server %p, no data yet\n", rqstp);
Jeff Layton70867212008-02-07 16:34:54 -0500640 if (signalled() || kthread_should_stop())
641 return -EINTR;
642 else
643 return -EAGAIN;
Tom Tucker0f0257e2007-12-30 21:08:27 -0600644 }
645 }
646 spin_unlock_bh(&pool->sp_lock);
647
648 len = 0;
649 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
650 dprintk("svc_recv: found XPT_CLOSE\n");
651 svc_delete_xprt(xprt);
652 } else if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
653 struct svc_xprt *newxpt;
654 newxpt = xprt->xpt_ops->xpo_accept(xprt);
655 if (newxpt) {
656 /*
657 * We know this module_get will succeed because the
658 * listener holds a reference too
659 */
660 __module_get(newxpt->xpt_class->xcl_owner);
661 svc_check_conn_limits(xprt->xpt_server);
662 spin_lock_bh(&serv->sv_lock);
663 set_bit(XPT_TEMP, &newxpt->xpt_flags);
664 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
665 serv->sv_tmpcnt++;
666 if (serv->sv_temptimer.function == NULL) {
667 /* setup timer to age temp transports */
668 setup_timer(&serv->sv_temptimer,
669 svc_age_temp_xprts,
670 (unsigned long)serv);
671 mod_timer(&serv->sv_temptimer,
672 jiffies + svc_conn_age_period * HZ);
673 }
674 spin_unlock_bh(&serv->sv_lock);
675 svc_xprt_received(newxpt);
676 }
677 svc_xprt_received(xprt);
678 } else {
679 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
680 rqstp, pool->sp_id, xprt,
681 atomic_read(&xprt->xpt_ref.refcount));
682 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
683 if (rqstp->rq_deferred) {
684 svc_xprt_received(xprt);
685 len = svc_deferred_recv(rqstp);
686 } else
687 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
688 dprintk("svc: got len=%d\n", len);
689 }
690
691 /* No data, incomplete (TCP) read, or accept() */
692 if (len == 0 || len == -EAGAIN) {
693 rqstp->rq_res.len = 0;
694 svc_xprt_release(rqstp);
695 return -EAGAIN;
696 }
697 clear_bit(XPT_OLD, &xprt->xpt_flags);
698
699 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
700 rqstp->rq_chandle.defer = svc_defer;
701
702 if (serv->sv_stats)
703 serv->sv_stats->netcnt++;
704 return len;
705}
Trond Myklebustd2f7e792007-07-14 15:39:58 -0400706EXPORT_SYMBOL(svc_recv);
Tom Tucker0f0257e2007-12-30 21:08:27 -0600707
708/*
709 * Drop request
710 */
711void svc_drop(struct svc_rqst *rqstp)
712{
713 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
714 svc_xprt_release(rqstp);
715}
Trond Myklebustd2f7e792007-07-14 15:39:58 -0400716EXPORT_SYMBOL(svc_drop);
Tom Tucker0f0257e2007-12-30 21:08:27 -0600717
718/*
719 * Return reply to client.
720 */
721int svc_send(struct svc_rqst *rqstp)
722{
723 struct svc_xprt *xprt;
724 int len;
725 struct xdr_buf *xb;
726
727 xprt = rqstp->rq_xprt;
728 if (!xprt)
729 return -EFAULT;
730
731 /* release the receive skb before sending the reply */
732 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
733
734 /* calculate over-all length */
735 xb = &rqstp->rq_res;
736 xb->len = xb->head[0].iov_len +
737 xb->page_len +
738 xb->tail[0].iov_len;
739
740 /* Grab mutex to serialize outgoing data. */
741 mutex_lock(&xprt->xpt_mutex);
742 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
743 len = -ENOTCONN;
744 else
745 len = xprt->xpt_ops->xpo_sendto(rqstp);
746 mutex_unlock(&xprt->xpt_mutex);
747 svc_xprt_release(rqstp);
748
749 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
750 return 0;
751 return len;
752}
753
754/*
755 * Timer function to close old temporary transports, using
756 * a mark-and-sweep algorithm.
757 */
758static void svc_age_temp_xprts(unsigned long closure)
759{
760 struct svc_serv *serv = (struct svc_serv *)closure;
761 struct svc_xprt *xprt;
762 struct list_head *le, *next;
763 LIST_HEAD(to_be_aged);
764
765 dprintk("svc_age_temp_xprts\n");
766
767 if (!spin_trylock_bh(&serv->sv_lock)) {
768 /* busy, try again 1 sec later */
769 dprintk("svc_age_temp_xprts: busy\n");
770 mod_timer(&serv->sv_temptimer, jiffies + HZ);
771 return;
772 }
773
774 list_for_each_safe(le, next, &serv->sv_tempsocks) {
775 xprt = list_entry(le, struct svc_xprt, xpt_list);
776
777 /* First time through, just mark it OLD. Second time
778 * through, close it. */
779 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
780 continue;
781 if (atomic_read(&xprt->xpt_ref.refcount) > 1
782 || test_bit(XPT_BUSY, &xprt->xpt_flags))
783 continue;
784 svc_xprt_get(xprt);
785 list_move(le, &to_be_aged);
786 set_bit(XPT_CLOSE, &xprt->xpt_flags);
787 set_bit(XPT_DETACHED, &xprt->xpt_flags);
788 }
789 spin_unlock_bh(&serv->sv_lock);
790
791 while (!list_empty(&to_be_aged)) {
792 le = to_be_aged.next;
793 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
794 list_del_init(le);
795 xprt = list_entry(le, struct svc_xprt, xpt_list);
796
797 dprintk("queuing xprt %p for closing\n", xprt);
798
799 /* a thread will dequeue and close it soon */
800 svc_xprt_enqueue(xprt);
801 svc_xprt_put(xprt);
802 }
803
804 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
805}
806
807/*
808 * Remove a dead transport
809 */
810void svc_delete_xprt(struct svc_xprt *xprt)
811{
812 struct svc_serv *serv = xprt->xpt_server;
813
814 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
815 xprt->xpt_ops->xpo_detach(xprt);
816
817 spin_lock_bh(&serv->sv_lock);
818 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
819 list_del_init(&xprt->xpt_list);
820 /*
821 * We used to delete the transport from whichever list
822 * it's sk_xprt.xpt_ready node was on, but we don't actually
823 * need to. This is because the only time we're called
824 * while still attached to a queue, the queue itself
825 * is about to be destroyed (in svc_destroy).
826 */
827 if (!test_and_set_bit(XPT_DEAD, &xprt->xpt_flags)) {
828 BUG_ON(atomic_read(&xprt->xpt_ref.refcount) < 2);
829 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
830 serv->sv_tmpcnt--;
831 svc_xprt_put(xprt);
832 }
833 spin_unlock_bh(&serv->sv_lock);
834}
835
836void svc_close_xprt(struct svc_xprt *xprt)
837{
838 set_bit(XPT_CLOSE, &xprt->xpt_flags);
839 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
840 /* someone else will have to effect the close */
841 return;
842
843 svc_xprt_get(xprt);
844 svc_delete_xprt(xprt);
845 clear_bit(XPT_BUSY, &xprt->xpt_flags);
846 svc_xprt_put(xprt);
847}
Tom Tuckera2178132007-12-30 21:08:35 -0600848EXPORT_SYMBOL_GPL(svc_close_xprt);
Tom Tucker0f0257e2007-12-30 21:08:27 -0600849
850void svc_close_all(struct list_head *xprt_list)
851{
852 struct svc_xprt *xprt;
853 struct svc_xprt *tmp;
854
855 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
856 set_bit(XPT_CLOSE, &xprt->xpt_flags);
857 if (test_bit(XPT_BUSY, &xprt->xpt_flags)) {
858 /* Waiting to be processed, but no threads left,
859 * So just remove it from the waiting list
860 */
861 list_del_init(&xprt->xpt_ready);
862 clear_bit(XPT_BUSY, &xprt->xpt_flags);
863 }
864 svc_close_xprt(xprt);
865 }
866}
867
868/*
869 * Handle defer and revisit of requests
870 */
871
872static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
873{
874 struct svc_deferred_req *dr =
875 container_of(dreq, struct svc_deferred_req, handle);
876 struct svc_xprt *xprt = dr->xprt;
877
878 if (too_many) {
879 svc_xprt_put(xprt);
880 kfree(dr);
881 return;
882 }
883 dprintk("revisit queued\n");
884 dr->xprt = NULL;
885 spin_lock(&xprt->xpt_lock);
886 list_add(&dr->handle.recent, &xprt->xpt_deferred);
887 spin_unlock(&xprt->xpt_lock);
888 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
889 svc_xprt_enqueue(xprt);
890 svc_xprt_put(xprt);
891}
892
Tom Tucker260c1d12007-12-30 21:08:29 -0600893/*
894 * Save the request off for later processing. The request buffer looks
895 * like this:
896 *
897 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
898 *
899 * This code can only handle requests that consist of an xprt-header
900 * and rpc-header.
901 */
Tom Tucker0f0257e2007-12-30 21:08:27 -0600902static struct cache_deferred_req *svc_defer(struct cache_req *req)
903{
904 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
Tom Tucker0f0257e2007-12-30 21:08:27 -0600905 struct svc_deferred_req *dr;
906
907 if (rqstp->rq_arg.page_len)
908 return NULL; /* if more than a page, give up FIXME */
909 if (rqstp->rq_deferred) {
910 dr = rqstp->rq_deferred;
911 rqstp->rq_deferred = NULL;
912 } else {
Tom Tucker260c1d12007-12-30 21:08:29 -0600913 size_t skip;
914 size_t size;
Tom Tucker0f0257e2007-12-30 21:08:27 -0600915 /* FIXME maybe discard if size too large */
Tom Tucker260c1d12007-12-30 21:08:29 -0600916 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
Tom Tucker0f0257e2007-12-30 21:08:27 -0600917 dr = kmalloc(size, GFP_KERNEL);
918 if (dr == NULL)
919 return NULL;
920
921 dr->handle.owner = rqstp->rq_server;
922 dr->prot = rqstp->rq_prot;
923 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
924 dr->addrlen = rqstp->rq_addrlen;
925 dr->daddr = rqstp->rq_daddr;
926 dr->argslen = rqstp->rq_arg.len >> 2;
Tom Tucker260c1d12007-12-30 21:08:29 -0600927 dr->xprt_hlen = rqstp->rq_xprt_hlen;
928
929 /* back up head to the start of the buffer and copy */
930 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
931 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
932 dr->argslen << 2);
Tom Tucker0f0257e2007-12-30 21:08:27 -0600933 }
934 svc_xprt_get(rqstp->rq_xprt);
935 dr->xprt = rqstp->rq_xprt;
936
937 dr->handle.revisit = svc_revisit;
938 return &dr->handle;
939}
940
941/*
942 * recv data from a deferred request into an active one
943 */
944static int svc_deferred_recv(struct svc_rqst *rqstp)
945{
946 struct svc_deferred_req *dr = rqstp->rq_deferred;
947
Tom Tucker260c1d12007-12-30 21:08:29 -0600948 /* setup iov_base past transport header */
949 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
950 /* The iov_len does not include the transport header bytes */
951 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
Tom Tucker0f0257e2007-12-30 21:08:27 -0600952 rqstp->rq_arg.page_len = 0;
Tom Tucker260c1d12007-12-30 21:08:29 -0600953 /* The rq_arg.len includes the transport header bytes */
954 rqstp->rq_arg.len = dr->argslen<<2;
Tom Tucker0f0257e2007-12-30 21:08:27 -0600955 rqstp->rq_prot = dr->prot;
956 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
957 rqstp->rq_addrlen = dr->addrlen;
Tom Tucker260c1d12007-12-30 21:08:29 -0600958 /* Save off transport header len in case we get deferred again */
959 rqstp->rq_xprt_hlen = dr->xprt_hlen;
Tom Tucker0f0257e2007-12-30 21:08:27 -0600960 rqstp->rq_daddr = dr->daddr;
961 rqstp->rq_respages = rqstp->rq_pages;
Tom Tucker260c1d12007-12-30 21:08:29 -0600962 return (dr->argslen<<2) - dr->xprt_hlen;
Tom Tucker0f0257e2007-12-30 21:08:27 -0600963}
964
965
966static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
967{
968 struct svc_deferred_req *dr = NULL;
969
970 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
971 return NULL;
972 spin_lock(&xprt->xpt_lock);
973 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
974 if (!list_empty(&xprt->xpt_deferred)) {
975 dr = list_entry(xprt->xpt_deferred.next,
976 struct svc_deferred_req,
977 handle.recent);
978 list_del_init(&dr->handle.recent);
979 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
980 }
981 spin_unlock(&xprt->xpt_lock);
982 return dr;
983}
Tom Tucker7fcb98d2007-12-30 21:08:33 -0600984
985/*
986 * Return the transport instance pointer for the endpoint accepting
987 * connections/peer traffic from the specified transport class,
988 * address family and port.
989 *
990 * Specifying 0 for the address family or port is effectively a
991 * wild-card, and will result in matching the first transport in the
992 * service's list that has a matching class name.
993 */
994struct svc_xprt *svc_find_xprt(struct svc_serv *serv, char *xcl_name,
995 int af, int port)
996{
997 struct svc_xprt *xprt;
998 struct svc_xprt *found = NULL;
999
1000 /* Sanity check the args */
1001 if (!serv || !xcl_name)
1002 return found;
1003
1004 spin_lock_bh(&serv->sv_lock);
1005 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1006 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1007 continue;
1008 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1009 continue;
1010 if (port && port != svc_xprt_local_port(xprt))
1011 continue;
1012 found = xprt;
Tom Tuckera2178132007-12-30 21:08:35 -06001013 svc_xprt_get(xprt);
Tom Tucker7fcb98d2007-12-30 21:08:33 -06001014 break;
1015 }
1016 spin_unlock_bh(&serv->sv_lock);
1017 return found;
1018}
1019EXPORT_SYMBOL_GPL(svc_find_xprt);
Tom Tucker9571af12007-12-30 21:08:37 -06001020
1021/*
1022 * Format a buffer with a list of the active transports. A zero for
1023 * the buflen parameter disables target buffer overflow checking.
1024 */
1025int svc_xprt_names(struct svc_serv *serv, char *buf, int buflen)
1026{
1027 struct svc_xprt *xprt;
1028 char xprt_str[64];
1029 int totlen = 0;
1030 int len;
1031
1032 /* Sanity check args */
1033 if (!serv)
1034 return 0;
1035
1036 spin_lock_bh(&serv->sv_lock);
1037 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1038 len = snprintf(xprt_str, sizeof(xprt_str),
1039 "%s %d\n", xprt->xpt_class->xcl_name,
1040 svc_xprt_local_port(xprt));
1041 /* If the string was truncated, replace with error string */
1042 if (len >= sizeof(xprt_str))
1043 strcpy(xprt_str, "name-too-long\n");
1044 /* Don't overflow buffer */
1045 len = strlen(xprt_str);
1046 if (buflen && (len + totlen >= buflen))
1047 break;
1048 strcpy(buf+totlen, xprt_str);
1049 totlen += len;
1050 }
1051 spin_unlock_bh(&serv->sv_lock);
1052 return totlen;
1053}
1054EXPORT_SYMBOL_GPL(svc_xprt_names);