Loading...
1/*
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>
9#include <linux/freezer.h>
10#include <linux/kthread.h>
11#include <linux/slab.h>
12#include <net/sock.h>
13#include <linux/sunrpc/stats.h>
14#include <linux/sunrpc/svc_xprt.h>
15#include <linux/sunrpc/svcsock.h>
16#include <linux/sunrpc/xprt.h>
17#include <linux/module.h>
18
19#define RPCDBG_FACILITY RPCDBG_SVCXPRT
20
21static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
22static int svc_deferred_recv(struct svc_rqst *rqstp);
23static struct cache_deferred_req *svc_defer(struct cache_req *req);
24static void svc_age_temp_xprts(unsigned long closure);
25static void svc_delete_xprt(struct svc_xprt *xprt);
26
27/* apparently the "standard" is that clients close
28 * idle connections after 5 minutes, servers after
29 * 6 minutes
30 * http://www.connectathon.org/talks96/nfstcp.pdf
31 */
32static int svc_conn_age_period = 6*60;
33
34/* List of registered transport classes */
35static DEFINE_SPINLOCK(svc_xprt_class_lock);
36static LIST_HEAD(svc_xprt_class_list);
37
38/* SMP locking strategy:
39 *
40 * svc_pool->sp_lock protects most of the fields of that pool.
41 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
42 * when both need to be taken (rare), svc_serv->sv_lock is first.
43 * BKL protects svc_serv->sv_nrthread.
44 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
45 * and the ->sk_info_authunix cache.
46 *
47 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
48 * enqueued multiply. During normal transport processing this bit
49 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
50 * Providers should not manipulate this bit directly.
51 *
52 * Some flags can be set to certain values at any time
53 * providing that certain rules are followed:
54 *
55 * XPT_CONN, XPT_DATA:
56 * - Can be set or cleared at any time.
57 * - After a set, svc_xprt_enqueue must be called to enqueue
58 * the transport for processing.
59 * - After a clear, the transport must be read/accepted.
60 * If this succeeds, it must be set again.
61 * XPT_CLOSE:
62 * - Can set at any time. It is never cleared.
63 * XPT_DEAD:
64 * - Can only be set while XPT_BUSY is held which ensures
65 * that no other thread will be using the transport or will
66 * try to set XPT_DEAD.
67 */
68
69int svc_reg_xprt_class(struct svc_xprt_class *xcl)
70{
71 struct svc_xprt_class *cl;
72 int res = -EEXIST;
73
74 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
75
76 INIT_LIST_HEAD(&xcl->xcl_list);
77 spin_lock(&svc_xprt_class_lock);
78 /* Make sure there isn't already a class with the same name */
79 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
80 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
81 goto out;
82 }
83 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
84 res = 0;
85out:
86 spin_unlock(&svc_xprt_class_lock);
87 return res;
88}
89EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
90
91void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
92{
93 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
94 spin_lock(&svc_xprt_class_lock);
95 list_del_init(&xcl->xcl_list);
96 spin_unlock(&svc_xprt_class_lock);
97}
98EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
99
100/*
101 * Format the transport list for printing
102 */
103int svc_print_xprts(char *buf, int maxlen)
104{
105 struct svc_xprt_class *xcl;
106 char tmpstr[80];
107 int len = 0;
108 buf[0] = '\0';
109
110 spin_lock(&svc_xprt_class_lock);
111 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
112 int slen;
113
114 sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
115 slen = strlen(tmpstr);
116 if (len + slen > maxlen)
117 break;
118 len += slen;
119 strcat(buf, tmpstr);
120 }
121 spin_unlock(&svc_xprt_class_lock);
122
123 return len;
124}
125
126static void svc_xprt_free(struct kref *kref)
127{
128 struct svc_xprt *xprt =
129 container_of(kref, struct svc_xprt, xpt_ref);
130 struct module *owner = xprt->xpt_class->xcl_owner;
131 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
132 svcauth_unix_info_release(xprt);
133 put_net(xprt->xpt_net);
134 /* See comment on corresponding get in xs_setup_bc_tcp(): */
135 if (xprt->xpt_bc_xprt)
136 xprt_put(xprt->xpt_bc_xprt);
137 xprt->xpt_ops->xpo_free(xprt);
138 module_put(owner);
139}
140
141void svc_xprt_put(struct svc_xprt *xprt)
142{
143 kref_put(&xprt->xpt_ref, svc_xprt_free);
144}
145EXPORT_SYMBOL_GPL(svc_xprt_put);
146
147/*
148 * Called by transport drivers to initialize the transport independent
149 * portion of the transport instance.
150 */
151void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
152 struct svc_xprt *xprt, struct svc_serv *serv)
153{
154 memset(xprt, 0, sizeof(*xprt));
155 xprt->xpt_class = xcl;
156 xprt->xpt_ops = xcl->xcl_ops;
157 kref_init(&xprt->xpt_ref);
158 xprt->xpt_server = serv;
159 INIT_LIST_HEAD(&xprt->xpt_list);
160 INIT_LIST_HEAD(&xprt->xpt_ready);
161 INIT_LIST_HEAD(&xprt->xpt_deferred);
162 INIT_LIST_HEAD(&xprt->xpt_users);
163 mutex_init(&xprt->xpt_mutex);
164 spin_lock_init(&xprt->xpt_lock);
165 set_bit(XPT_BUSY, &xprt->xpt_flags);
166 rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
167 xprt->xpt_net = get_net(net);
168}
169EXPORT_SYMBOL_GPL(svc_xprt_init);
170
171static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
172 struct svc_serv *serv,
173 struct net *net,
174 const int family,
175 const unsigned short port,
176 int flags)
177{
178 struct sockaddr_in sin = {
179 .sin_family = AF_INET,
180 .sin_addr.s_addr = htonl(INADDR_ANY),
181 .sin_port = htons(port),
182 };
183#if IS_ENABLED(CONFIG_IPV6)
184 struct sockaddr_in6 sin6 = {
185 .sin6_family = AF_INET6,
186 .sin6_addr = IN6ADDR_ANY_INIT,
187 .sin6_port = htons(port),
188 };
189#endif
190 struct sockaddr *sap;
191 size_t len;
192
193 switch (family) {
194 case PF_INET:
195 sap = (struct sockaddr *)&sin;
196 len = sizeof(sin);
197 break;
198#if IS_ENABLED(CONFIG_IPV6)
199 case PF_INET6:
200 sap = (struct sockaddr *)&sin6;
201 len = sizeof(sin6);
202 break;
203#endif
204 default:
205 return ERR_PTR(-EAFNOSUPPORT);
206 }
207
208 return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
209}
210
211int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
212 struct net *net, const int family,
213 const unsigned short port, int flags)
214{
215 struct svc_xprt_class *xcl;
216
217 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
218 spin_lock(&svc_xprt_class_lock);
219 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
220 struct svc_xprt *newxprt;
221 unsigned short newport;
222
223 if (strcmp(xprt_name, xcl->xcl_name))
224 continue;
225
226 if (!try_module_get(xcl->xcl_owner))
227 goto err;
228
229 spin_unlock(&svc_xprt_class_lock);
230 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
231 if (IS_ERR(newxprt)) {
232 module_put(xcl->xcl_owner);
233 return PTR_ERR(newxprt);
234 }
235
236 clear_bit(XPT_TEMP, &newxprt->xpt_flags);
237 spin_lock_bh(&serv->sv_lock);
238 list_add(&newxprt->xpt_list, &serv->sv_permsocks);
239 spin_unlock_bh(&serv->sv_lock);
240 newport = svc_xprt_local_port(newxprt);
241 clear_bit(XPT_BUSY, &newxprt->xpt_flags);
242 return newport;
243 }
244 err:
245 spin_unlock(&svc_xprt_class_lock);
246 dprintk("svc: transport %s not found\n", xprt_name);
247
248 /* This errno is exposed to user space. Provide a reasonable
249 * perror msg for a bad transport. */
250 return -EPROTONOSUPPORT;
251}
252EXPORT_SYMBOL_GPL(svc_create_xprt);
253
254/*
255 * Copy the local and remote xprt addresses to the rqstp structure
256 */
257void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
258{
259 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
260 rqstp->rq_addrlen = xprt->xpt_remotelen;
261
262 /*
263 * Destination address in request is needed for binding the
264 * source address in RPC replies/callbacks later.
265 */
266 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
267 rqstp->rq_daddrlen = xprt->xpt_locallen;
268}
269EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
270
271/**
272 * svc_print_addr - Format rq_addr field for printing
273 * @rqstp: svc_rqst struct containing address to print
274 * @buf: target buffer for formatted address
275 * @len: length of target buffer
276 *
277 */
278char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
279{
280 return __svc_print_addr(svc_addr(rqstp), buf, len);
281}
282EXPORT_SYMBOL_GPL(svc_print_addr);
283
284/*
285 * Queue up an idle server thread. Must have pool->sp_lock held.
286 * Note: this is really a stack rather than a queue, so that we only
287 * use as many different threads as we need, and the rest don't pollute
288 * the cache.
289 */
290static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
291{
292 list_add(&rqstp->rq_list, &pool->sp_threads);
293}
294
295/*
296 * Dequeue an nfsd thread. Must have pool->sp_lock held.
297 */
298static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
299{
300 list_del(&rqstp->rq_list);
301}
302
303static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
304{
305 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
306 return true;
307 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED)))
308 return xprt->xpt_ops->xpo_has_wspace(xprt);
309 return false;
310}
311
312/*
313 * Queue up a transport with data pending. If there are idle nfsd
314 * processes, wake 'em up.
315 *
316 */
317void svc_xprt_enqueue(struct svc_xprt *xprt)
318{
319 struct svc_pool *pool;
320 struct svc_rqst *rqstp;
321 int cpu;
322
323 if (!svc_xprt_has_something_to_do(xprt))
324 return;
325
326 cpu = get_cpu();
327 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
328 put_cpu();
329
330 spin_lock_bh(&pool->sp_lock);
331
332 if (!list_empty(&pool->sp_threads) &&
333 !list_empty(&pool->sp_sockets))
334 printk(KERN_ERR
335 "svc_xprt_enqueue: "
336 "threads and transports both waiting??\n");
337
338 pool->sp_stats.packets++;
339
340 /* Mark transport as busy. It will remain in this state until
341 * the provider calls svc_xprt_received. We update XPT_BUSY
342 * atomically because it also guards against trying to enqueue
343 * the transport twice.
344 */
345 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
346 /* Don't enqueue transport while already enqueued */
347 dprintk("svc: transport %p busy, not enqueued\n", xprt);
348 goto out_unlock;
349 }
350
351 if (!list_empty(&pool->sp_threads)) {
352 rqstp = list_entry(pool->sp_threads.next,
353 struct svc_rqst,
354 rq_list);
355 dprintk("svc: transport %p served by daemon %p\n",
356 xprt, rqstp);
357 svc_thread_dequeue(pool, rqstp);
358 if (rqstp->rq_xprt)
359 printk(KERN_ERR
360 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
361 rqstp, rqstp->rq_xprt);
362 rqstp->rq_xprt = xprt;
363 svc_xprt_get(xprt);
364 pool->sp_stats.threads_woken++;
365 wake_up(&rqstp->rq_wait);
366 } else {
367 dprintk("svc: transport %p put into queue\n", xprt);
368 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
369 pool->sp_stats.sockets_queued++;
370 }
371
372out_unlock:
373 spin_unlock_bh(&pool->sp_lock);
374}
375EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
376
377/*
378 * Dequeue the first transport. Must be called with the pool->sp_lock held.
379 */
380static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
381{
382 struct svc_xprt *xprt;
383
384 if (list_empty(&pool->sp_sockets))
385 return NULL;
386
387 xprt = list_entry(pool->sp_sockets.next,
388 struct svc_xprt, xpt_ready);
389 list_del_init(&xprt->xpt_ready);
390
391 dprintk("svc: transport %p dequeued, inuse=%d\n",
392 xprt, atomic_read(&xprt->xpt_ref.refcount));
393
394 return xprt;
395}
396
397/*
398 * svc_xprt_received conditionally queues the transport for processing
399 * by another thread. The caller must hold the XPT_BUSY bit and must
400 * not thereafter touch transport data.
401 *
402 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
403 * insufficient) data.
404 */
405void svc_xprt_received(struct svc_xprt *xprt)
406{
407 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
408 /* As soon as we clear busy, the xprt could be closed and
409 * 'put', so we need a reference to call svc_xprt_enqueue with:
410 */
411 svc_xprt_get(xprt);
412 clear_bit(XPT_BUSY, &xprt->xpt_flags);
413 svc_xprt_enqueue(xprt);
414 svc_xprt_put(xprt);
415}
416EXPORT_SYMBOL_GPL(svc_xprt_received);
417
418/**
419 * svc_reserve - change the space reserved for the reply to a request.
420 * @rqstp: The request in question
421 * @space: new max space to reserve
422 *
423 * Each request reserves some space on the output queue of the transport
424 * to make sure the reply fits. This function reduces that reserved
425 * space to be the amount of space used already, plus @space.
426 *
427 */
428void svc_reserve(struct svc_rqst *rqstp, int space)
429{
430 space += rqstp->rq_res.head[0].iov_len;
431
432 if (space < rqstp->rq_reserved) {
433 struct svc_xprt *xprt = rqstp->rq_xprt;
434 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
435 rqstp->rq_reserved = space;
436
437 svc_xprt_enqueue(xprt);
438 }
439}
440EXPORT_SYMBOL_GPL(svc_reserve);
441
442static void svc_xprt_release(struct svc_rqst *rqstp)
443{
444 struct svc_xprt *xprt = rqstp->rq_xprt;
445
446 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
447
448 kfree(rqstp->rq_deferred);
449 rqstp->rq_deferred = NULL;
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}
501EXPORT_SYMBOL_GPL(svc_wake_up);
502
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 have,
519 * something must be dropped. It's not clear what will happen if we allow
520 * "too many" connections, but when dealing with network-facing software,
521 * we have to code defensively. Here we do that by imposing hard limits.
522 *
523 * There's no point in trying to do random drop here for DoS
524 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
525 * attacker can easily beat that.
526 *
527 * The only somewhat efficient mechanism would be if drop old
528 * connections from the same IP first. But right now we don't even
529 * record the client IP in svc_sock.
530 *
531 * single-threaded services that expect a lot of clients will probably
532 * need to set sv_maxconn to override the default value which is based
533 * on the number of threads
534 */
535static void svc_check_conn_limits(struct svc_serv *serv)
536{
537 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
538 (serv->sv_nrthreads+3) * 20;
539
540 if (serv->sv_tmpcnt > limit) {
541 struct svc_xprt *xprt = NULL;
542 spin_lock_bh(&serv->sv_lock);
543 if (!list_empty(&serv->sv_tempsocks)) {
544 /* Try to help the admin */
545 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
546 serv->sv_name, serv->sv_maxconn ?
547 "max number of connections" :
548 "number of threads");
549 /*
550 * Always select the oldest connection. It's not fair,
551 * but so is life
552 */
553 xprt = list_entry(serv->sv_tempsocks.prev,
554 struct svc_xprt,
555 xpt_list);
556 set_bit(XPT_CLOSE, &xprt->xpt_flags);
557 svc_xprt_get(xprt);
558 }
559 spin_unlock_bh(&serv->sv_lock);
560
561 if (xprt) {
562 svc_xprt_enqueue(xprt);
563 svc_xprt_put(xprt);
564 }
565 }
566}
567
568/*
569 * Receive the next request on any transport. This code is carefully
570 * organised not to touch any cachelines in the shared svc_serv
571 * structure, only cachelines in the local svc_pool.
572 */
573int svc_recv(struct svc_rqst *rqstp, long timeout)
574{
575 struct svc_xprt *xprt = NULL;
576 struct svc_serv *serv = rqstp->rq_server;
577 struct svc_pool *pool = rqstp->rq_pool;
578 int len, i;
579 int pages;
580 struct xdr_buf *arg;
581 DECLARE_WAITQUEUE(wait, current);
582 long time_left;
583
584 dprintk("svc: server %p waiting for data (to = %ld)\n",
585 rqstp, timeout);
586
587 if (rqstp->rq_xprt)
588 printk(KERN_ERR
589 "svc_recv: service %p, transport not NULL!\n",
590 rqstp);
591 if (waitqueue_active(&rqstp->rq_wait))
592 printk(KERN_ERR
593 "svc_recv: service %p, wait queue active!\n",
594 rqstp);
595
596 /* now allocate needed pages. If we get a failure, sleep briefly */
597 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
598 BUG_ON(pages >= RPCSVC_MAXPAGES);
599 for (i = 0; i < pages ; i++)
600 while (rqstp->rq_pages[i] == NULL) {
601 struct page *p = alloc_page(GFP_KERNEL);
602 if (!p) {
603 set_current_state(TASK_INTERRUPTIBLE);
604 if (signalled() || kthread_should_stop()) {
605 set_current_state(TASK_RUNNING);
606 return -EINTR;
607 }
608 schedule_timeout(msecs_to_jiffies(500));
609 }
610 rqstp->rq_pages[i] = p;
611 }
612 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
613
614 /* Make arg->head point to first page and arg->pages point to rest */
615 arg = &rqstp->rq_arg;
616 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
617 arg->head[0].iov_len = PAGE_SIZE;
618 arg->pages = rqstp->rq_pages + 1;
619 arg->page_base = 0;
620 /* save at least one page for response */
621 arg->page_len = (pages-2)*PAGE_SIZE;
622 arg->len = (pages-1)*PAGE_SIZE;
623 arg->tail[0].iov_len = 0;
624
625 try_to_freeze();
626 cond_resched();
627 if (signalled() || kthread_should_stop())
628 return -EINTR;
629
630 /* Normally we will wait up to 5 seconds for any required
631 * cache information to be provided.
632 */
633 rqstp->rq_chandle.thread_wait = 5*HZ;
634
635 spin_lock_bh(&pool->sp_lock);
636 xprt = svc_xprt_dequeue(pool);
637 if (xprt) {
638 rqstp->rq_xprt = xprt;
639 svc_xprt_get(xprt);
640
641 /* As there is a shortage of threads and this request
642 * had to be queued, don't allow the thread to wait so
643 * long for cache updates.
644 */
645 rqstp->rq_chandle.thread_wait = 1*HZ;
646 } else {
647 /* No data pending. Go to sleep */
648 svc_thread_enqueue(pool, rqstp);
649
650 /*
651 * We have to be able to interrupt this wait
652 * to bring down the daemons ...
653 */
654 set_current_state(TASK_INTERRUPTIBLE);
655
656 /*
657 * checking kthread_should_stop() here allows us to avoid
658 * locking and signalling when stopping kthreads that call
659 * svc_recv. If the thread has already been woken up, then
660 * we can exit here without sleeping. If not, then it
661 * it'll be woken up quickly during the schedule_timeout
662 */
663 if (kthread_should_stop()) {
664 set_current_state(TASK_RUNNING);
665 spin_unlock_bh(&pool->sp_lock);
666 return -EINTR;
667 }
668
669 add_wait_queue(&rqstp->rq_wait, &wait);
670 spin_unlock_bh(&pool->sp_lock);
671
672 time_left = schedule_timeout(timeout);
673
674 try_to_freeze();
675
676 spin_lock_bh(&pool->sp_lock);
677 remove_wait_queue(&rqstp->rq_wait, &wait);
678 if (!time_left)
679 pool->sp_stats.threads_timedout++;
680
681 xprt = rqstp->rq_xprt;
682 if (!xprt) {
683 svc_thread_dequeue(pool, rqstp);
684 spin_unlock_bh(&pool->sp_lock);
685 dprintk("svc: server %p, no data yet\n", rqstp);
686 if (signalled() || kthread_should_stop())
687 return -EINTR;
688 else
689 return -EAGAIN;
690 }
691 }
692 spin_unlock_bh(&pool->sp_lock);
693
694 len = 0;
695 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
696 dprintk("svc_recv: found XPT_CLOSE\n");
697 svc_delete_xprt(xprt);
698 /* Leave XPT_BUSY set on the dead xprt: */
699 goto out;
700 }
701 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
702 struct svc_xprt *newxpt;
703 newxpt = xprt->xpt_ops->xpo_accept(xprt);
704 if (newxpt) {
705 /*
706 * We know this module_get will succeed because the
707 * listener holds a reference too
708 */
709 __module_get(newxpt->xpt_class->xcl_owner);
710 svc_check_conn_limits(xprt->xpt_server);
711 spin_lock_bh(&serv->sv_lock);
712 set_bit(XPT_TEMP, &newxpt->xpt_flags);
713 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
714 serv->sv_tmpcnt++;
715 if (serv->sv_temptimer.function == NULL) {
716 /* setup timer to age temp transports */
717 setup_timer(&serv->sv_temptimer,
718 svc_age_temp_xprts,
719 (unsigned long)serv);
720 mod_timer(&serv->sv_temptimer,
721 jiffies + svc_conn_age_period * HZ);
722 }
723 spin_unlock_bh(&serv->sv_lock);
724 svc_xprt_received(newxpt);
725 }
726 } else if (xprt->xpt_ops->xpo_has_wspace(xprt)) {
727 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
728 rqstp, pool->sp_id, xprt,
729 atomic_read(&xprt->xpt_ref.refcount));
730 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
731 if (rqstp->rq_deferred)
732 len = svc_deferred_recv(rqstp);
733 else
734 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
735 dprintk("svc: got len=%d\n", len);
736 rqstp->rq_reserved = serv->sv_max_mesg;
737 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
738 }
739 svc_xprt_received(xprt);
740
741 /* No data, incomplete (TCP) read, or accept() */
742 if (len == 0 || len == -EAGAIN)
743 goto out;
744
745 clear_bit(XPT_OLD, &xprt->xpt_flags);
746
747 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
748 rqstp->rq_chandle.defer = svc_defer;
749
750 if (serv->sv_stats)
751 serv->sv_stats->netcnt++;
752 return len;
753out:
754 rqstp->rq_res.len = 0;
755 svc_xprt_release(rqstp);
756 return -EAGAIN;
757}
758EXPORT_SYMBOL_GPL(svc_recv);
759
760/*
761 * Drop request
762 */
763void svc_drop(struct svc_rqst *rqstp)
764{
765 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
766 svc_xprt_release(rqstp);
767}
768EXPORT_SYMBOL_GPL(svc_drop);
769
770/*
771 * Return reply to client.
772 */
773int svc_send(struct svc_rqst *rqstp)
774{
775 struct svc_xprt *xprt;
776 int len;
777 struct xdr_buf *xb;
778
779 xprt = rqstp->rq_xprt;
780 if (!xprt)
781 return -EFAULT;
782
783 /* release the receive skb before sending the reply */
784 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
785
786 /* calculate over-all length */
787 xb = &rqstp->rq_res;
788 xb->len = xb->head[0].iov_len +
789 xb->page_len +
790 xb->tail[0].iov_len;
791
792 /* Grab mutex to serialize outgoing data. */
793 mutex_lock(&xprt->xpt_mutex);
794 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
795 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
796 len = -ENOTCONN;
797 else
798 len = xprt->xpt_ops->xpo_sendto(rqstp);
799 mutex_unlock(&xprt->xpt_mutex);
800 rpc_wake_up(&xprt->xpt_bc_pending);
801 svc_xprt_release(rqstp);
802
803 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
804 return 0;
805 return len;
806}
807
808/*
809 * Timer function to close old temporary transports, using
810 * a mark-and-sweep algorithm.
811 */
812static void svc_age_temp_xprts(unsigned long closure)
813{
814 struct svc_serv *serv = (struct svc_serv *)closure;
815 struct svc_xprt *xprt;
816 struct list_head *le, *next;
817 LIST_HEAD(to_be_aged);
818
819 dprintk("svc_age_temp_xprts\n");
820
821 if (!spin_trylock_bh(&serv->sv_lock)) {
822 /* busy, try again 1 sec later */
823 dprintk("svc_age_temp_xprts: busy\n");
824 mod_timer(&serv->sv_temptimer, jiffies + HZ);
825 return;
826 }
827
828 list_for_each_safe(le, next, &serv->sv_tempsocks) {
829 xprt = list_entry(le, struct svc_xprt, xpt_list);
830
831 /* First time through, just mark it OLD. Second time
832 * through, close it. */
833 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
834 continue;
835 if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
836 test_bit(XPT_BUSY, &xprt->xpt_flags))
837 continue;
838 svc_xprt_get(xprt);
839 list_move(le, &to_be_aged);
840 set_bit(XPT_CLOSE, &xprt->xpt_flags);
841 set_bit(XPT_DETACHED, &xprt->xpt_flags);
842 }
843 spin_unlock_bh(&serv->sv_lock);
844
845 while (!list_empty(&to_be_aged)) {
846 le = to_be_aged.next;
847 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
848 list_del_init(le);
849 xprt = list_entry(le, struct svc_xprt, xpt_list);
850
851 dprintk("queuing xprt %p for closing\n", xprt);
852
853 /* a thread will dequeue and close it soon */
854 svc_xprt_enqueue(xprt);
855 svc_xprt_put(xprt);
856 }
857
858 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
859}
860
861static void call_xpt_users(struct svc_xprt *xprt)
862{
863 struct svc_xpt_user *u;
864
865 spin_lock(&xprt->xpt_lock);
866 while (!list_empty(&xprt->xpt_users)) {
867 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
868 list_del(&u->list);
869 u->callback(u);
870 }
871 spin_unlock(&xprt->xpt_lock);
872}
873
874/*
875 * Remove a dead transport
876 */
877static void svc_delete_xprt(struct svc_xprt *xprt)
878{
879 struct svc_serv *serv = xprt->xpt_server;
880 struct svc_deferred_req *dr;
881
882 /* Only do this once */
883 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
884 BUG();
885
886 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
887 xprt->xpt_ops->xpo_detach(xprt);
888
889 spin_lock_bh(&serv->sv_lock);
890 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
891 list_del_init(&xprt->xpt_list);
892 BUG_ON(!list_empty(&xprt->xpt_ready));
893 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
894 serv->sv_tmpcnt--;
895 spin_unlock_bh(&serv->sv_lock);
896
897 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
898 kfree(dr);
899
900 call_xpt_users(xprt);
901 svc_xprt_put(xprt);
902}
903
904void svc_close_xprt(struct svc_xprt *xprt)
905{
906 set_bit(XPT_CLOSE, &xprt->xpt_flags);
907 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
908 /* someone else will have to effect the close */
909 return;
910 /*
911 * We expect svc_close_xprt() to work even when no threads are
912 * running (e.g., while configuring the server before starting
913 * any threads), so if the transport isn't busy, we delete
914 * it ourself:
915 */
916 svc_delete_xprt(xprt);
917}
918EXPORT_SYMBOL_GPL(svc_close_xprt);
919
920static void svc_close_list(struct list_head *xprt_list, struct net *net)
921{
922 struct svc_xprt *xprt;
923
924 list_for_each_entry(xprt, xprt_list, xpt_list) {
925 if (xprt->xpt_net != net)
926 continue;
927 set_bit(XPT_CLOSE, &xprt->xpt_flags);
928 set_bit(XPT_BUSY, &xprt->xpt_flags);
929 }
930}
931
932static void svc_clear_pools(struct svc_serv *serv, struct net *net)
933{
934 struct svc_pool *pool;
935 struct svc_xprt *xprt;
936 struct svc_xprt *tmp;
937 int i;
938
939 for (i = 0; i < serv->sv_nrpools; i++) {
940 pool = &serv->sv_pools[i];
941
942 spin_lock_bh(&pool->sp_lock);
943 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
944 if (xprt->xpt_net != net)
945 continue;
946 list_del_init(&xprt->xpt_ready);
947 }
948 spin_unlock_bh(&pool->sp_lock);
949 }
950}
951
952static void svc_clear_list(struct list_head *xprt_list, struct net *net)
953{
954 struct svc_xprt *xprt;
955 struct svc_xprt *tmp;
956
957 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
958 if (xprt->xpt_net != net)
959 continue;
960 svc_delete_xprt(xprt);
961 }
962 list_for_each_entry(xprt, xprt_list, xpt_list)
963 BUG_ON(xprt->xpt_net == net);
964}
965
966void svc_close_net(struct svc_serv *serv, struct net *net)
967{
968 svc_close_list(&serv->sv_tempsocks, net);
969 svc_close_list(&serv->sv_permsocks, net);
970
971 svc_clear_pools(serv, net);
972 /*
973 * At this point the sp_sockets lists will stay empty, since
974 * svc_xprt_enqueue will not add new entries without taking the
975 * sp_lock and checking XPT_BUSY.
976 */
977 svc_clear_list(&serv->sv_tempsocks, net);
978 svc_clear_list(&serv->sv_permsocks, net);
979}
980
981/*
982 * Handle defer and revisit of requests
983 */
984
985static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
986{
987 struct svc_deferred_req *dr =
988 container_of(dreq, struct svc_deferred_req, handle);
989 struct svc_xprt *xprt = dr->xprt;
990
991 spin_lock(&xprt->xpt_lock);
992 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
993 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
994 spin_unlock(&xprt->xpt_lock);
995 dprintk("revisit canceled\n");
996 svc_xprt_put(xprt);
997 kfree(dr);
998 return;
999 }
1000 dprintk("revisit queued\n");
1001 dr->xprt = NULL;
1002 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1003 spin_unlock(&xprt->xpt_lock);
1004 svc_xprt_enqueue(xprt);
1005 svc_xprt_put(xprt);
1006}
1007
1008/*
1009 * Save the request off for later processing. The request buffer looks
1010 * like this:
1011 *
1012 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1013 *
1014 * This code can only handle requests that consist of an xprt-header
1015 * and rpc-header.
1016 */
1017static struct cache_deferred_req *svc_defer(struct cache_req *req)
1018{
1019 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1020 struct svc_deferred_req *dr;
1021
1022 if (rqstp->rq_arg.page_len || !rqstp->rq_usedeferral)
1023 return NULL; /* if more than a page, give up FIXME */
1024 if (rqstp->rq_deferred) {
1025 dr = rqstp->rq_deferred;
1026 rqstp->rq_deferred = NULL;
1027 } else {
1028 size_t skip;
1029 size_t size;
1030 /* FIXME maybe discard if size too large */
1031 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1032 dr = kmalloc(size, GFP_KERNEL);
1033 if (dr == NULL)
1034 return NULL;
1035
1036 dr->handle.owner = rqstp->rq_server;
1037 dr->prot = rqstp->rq_prot;
1038 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1039 dr->addrlen = rqstp->rq_addrlen;
1040 dr->daddr = rqstp->rq_daddr;
1041 dr->argslen = rqstp->rq_arg.len >> 2;
1042 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1043
1044 /* back up head to the start of the buffer and copy */
1045 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1046 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1047 dr->argslen << 2);
1048 }
1049 svc_xprt_get(rqstp->rq_xprt);
1050 dr->xprt = rqstp->rq_xprt;
1051 rqstp->rq_dropme = true;
1052
1053 dr->handle.revisit = svc_revisit;
1054 return &dr->handle;
1055}
1056
1057/*
1058 * recv data from a deferred request into an active one
1059 */
1060static int svc_deferred_recv(struct svc_rqst *rqstp)
1061{
1062 struct svc_deferred_req *dr = rqstp->rq_deferred;
1063
1064 /* setup iov_base past transport header */
1065 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1066 /* The iov_len does not include the transport header bytes */
1067 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1068 rqstp->rq_arg.page_len = 0;
1069 /* The rq_arg.len includes the transport header bytes */
1070 rqstp->rq_arg.len = dr->argslen<<2;
1071 rqstp->rq_prot = dr->prot;
1072 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1073 rqstp->rq_addrlen = dr->addrlen;
1074 /* Save off transport header len in case we get deferred again */
1075 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1076 rqstp->rq_daddr = dr->daddr;
1077 rqstp->rq_respages = rqstp->rq_pages;
1078 return (dr->argslen<<2) - dr->xprt_hlen;
1079}
1080
1081
1082static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1083{
1084 struct svc_deferred_req *dr = NULL;
1085
1086 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1087 return NULL;
1088 spin_lock(&xprt->xpt_lock);
1089 if (!list_empty(&xprt->xpt_deferred)) {
1090 dr = list_entry(xprt->xpt_deferred.next,
1091 struct svc_deferred_req,
1092 handle.recent);
1093 list_del_init(&dr->handle.recent);
1094 } else
1095 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1096 spin_unlock(&xprt->xpt_lock);
1097 return dr;
1098}
1099
1100/**
1101 * svc_find_xprt - find an RPC transport instance
1102 * @serv: pointer to svc_serv to search
1103 * @xcl_name: C string containing transport's class name
1104 * @net: owner net pointer
1105 * @af: Address family of transport's local address
1106 * @port: transport's IP port number
1107 *
1108 * Return the transport instance pointer for the endpoint accepting
1109 * connections/peer traffic from the specified transport class,
1110 * address family and port.
1111 *
1112 * Specifying 0 for the address family or port is effectively a
1113 * wild-card, and will result in matching the first transport in the
1114 * service's list that has a matching class name.
1115 */
1116struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1117 struct net *net, const sa_family_t af,
1118 const unsigned short port)
1119{
1120 struct svc_xprt *xprt;
1121 struct svc_xprt *found = NULL;
1122
1123 /* Sanity check the args */
1124 if (serv == NULL || xcl_name == NULL)
1125 return found;
1126
1127 spin_lock_bh(&serv->sv_lock);
1128 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1129 if (xprt->xpt_net != net)
1130 continue;
1131 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1132 continue;
1133 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1134 continue;
1135 if (port != 0 && port != svc_xprt_local_port(xprt))
1136 continue;
1137 found = xprt;
1138 svc_xprt_get(xprt);
1139 break;
1140 }
1141 spin_unlock_bh(&serv->sv_lock);
1142 return found;
1143}
1144EXPORT_SYMBOL_GPL(svc_find_xprt);
1145
1146static int svc_one_xprt_name(const struct svc_xprt *xprt,
1147 char *pos, int remaining)
1148{
1149 int len;
1150
1151 len = snprintf(pos, remaining, "%s %u\n",
1152 xprt->xpt_class->xcl_name,
1153 svc_xprt_local_port(xprt));
1154 if (len >= remaining)
1155 return -ENAMETOOLONG;
1156 return len;
1157}
1158
1159/**
1160 * svc_xprt_names - format a buffer with a list of transport names
1161 * @serv: pointer to an RPC service
1162 * @buf: pointer to a buffer to be filled in
1163 * @buflen: length of buffer to be filled in
1164 *
1165 * Fills in @buf with a string containing a list of transport names,
1166 * each name terminated with '\n'.
1167 *
1168 * Returns positive length of the filled-in string on success; otherwise
1169 * a negative errno value is returned if an error occurs.
1170 */
1171int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1172{
1173 struct svc_xprt *xprt;
1174 int len, totlen;
1175 char *pos;
1176
1177 /* Sanity check args */
1178 if (!serv)
1179 return 0;
1180
1181 spin_lock_bh(&serv->sv_lock);
1182
1183 pos = buf;
1184 totlen = 0;
1185 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1186 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1187 if (len < 0) {
1188 *buf = '\0';
1189 totlen = len;
1190 }
1191 if (len <= 0)
1192 break;
1193
1194 pos += len;
1195 totlen += len;
1196 }
1197
1198 spin_unlock_bh(&serv->sv_lock);
1199 return totlen;
1200}
1201EXPORT_SYMBOL_GPL(svc_xprt_names);
1202
1203
1204/*----------------------------------------------------------------------------*/
1205
1206static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1207{
1208 unsigned int pidx = (unsigned int)*pos;
1209 struct svc_serv *serv = m->private;
1210
1211 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1212
1213 if (!pidx)
1214 return SEQ_START_TOKEN;
1215 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1216}
1217
1218static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1219{
1220 struct svc_pool *pool = p;
1221 struct svc_serv *serv = m->private;
1222
1223 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1224
1225 if (p == SEQ_START_TOKEN) {
1226 pool = &serv->sv_pools[0];
1227 } else {
1228 unsigned int pidx = (pool - &serv->sv_pools[0]);
1229 if (pidx < serv->sv_nrpools-1)
1230 pool = &serv->sv_pools[pidx+1];
1231 else
1232 pool = NULL;
1233 }
1234 ++*pos;
1235 return pool;
1236}
1237
1238static void svc_pool_stats_stop(struct seq_file *m, void *p)
1239{
1240}
1241
1242static int svc_pool_stats_show(struct seq_file *m, void *p)
1243{
1244 struct svc_pool *pool = p;
1245
1246 if (p == SEQ_START_TOKEN) {
1247 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1248 return 0;
1249 }
1250
1251 seq_printf(m, "%u %lu %lu %lu %lu\n",
1252 pool->sp_id,
1253 pool->sp_stats.packets,
1254 pool->sp_stats.sockets_queued,
1255 pool->sp_stats.threads_woken,
1256 pool->sp_stats.threads_timedout);
1257
1258 return 0;
1259}
1260
1261static const struct seq_operations svc_pool_stats_seq_ops = {
1262 .start = svc_pool_stats_start,
1263 .next = svc_pool_stats_next,
1264 .stop = svc_pool_stats_stop,
1265 .show = svc_pool_stats_show,
1266};
1267
1268int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1269{
1270 int err;
1271
1272 err = seq_open(file, &svc_pool_stats_seq_ops);
1273 if (!err)
1274 ((struct seq_file *) file->private_data)->private = serv;
1275 return err;
1276}
1277EXPORT_SYMBOL(svc_pool_stats_open);
1278
1279/*----------------------------------------------------------------------------*/