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/addr.h>
14#include <linux/sunrpc/stats.h>
15#include <linux/sunrpc/svc_xprt.h>
16#include <linux/sunrpc/svcsock.h>
17#include <linux/sunrpc/xprt.h>
18#include <linux/module.h>
19#include <linux/netdevice.h>
20#include <trace/events/sunrpc.h>
21
22#define RPCDBG_FACILITY RPCDBG_SVCXPRT
23
24static unsigned int svc_rpc_per_connection_limit __read_mostly;
25module_param(svc_rpc_per_connection_limit, uint, 0644);
26
27
28static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
29static int svc_deferred_recv(struct svc_rqst *rqstp);
30static struct cache_deferred_req *svc_defer(struct cache_req *req);
31static void svc_age_temp_xprts(struct timer_list *t);
32static void svc_delete_xprt(struct svc_xprt *xprt);
33
34/* apparently the "standard" is that clients close
35 * idle connections after 5 minutes, servers after
36 * 6 minutes
37 * http://www.connectathon.org/talks96/nfstcp.pdf
38 */
39static int svc_conn_age_period = 6*60;
40
41/* List of registered transport classes */
42static DEFINE_SPINLOCK(svc_xprt_class_lock);
43static LIST_HEAD(svc_xprt_class_list);
44
45/* SMP locking strategy:
46 *
47 * svc_pool->sp_lock protects most of the fields of that pool.
48 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
49 * when both need to be taken (rare), svc_serv->sv_lock is first.
50 * The "service mutex" protects svc_serv->sv_nrthread.
51 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
52 * and the ->sk_info_authunix cache.
53 *
54 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
55 * enqueued multiply. During normal transport processing this bit
56 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
57 * Providers should not manipulate this bit directly.
58 *
59 * Some flags can be set to certain values at any time
60 * providing that certain rules are followed:
61 *
62 * XPT_CONN, XPT_DATA:
63 * - Can be set or cleared at any time.
64 * - After a set, svc_xprt_enqueue must be called to enqueue
65 * the transport for processing.
66 * - After a clear, the transport must be read/accepted.
67 * If this succeeds, it must be set again.
68 * XPT_CLOSE:
69 * - Can set at any time. It is never cleared.
70 * XPT_DEAD:
71 * - Can only be set while XPT_BUSY is held which ensures
72 * that no other thread will be using the transport or will
73 * try to set XPT_DEAD.
74 */
75int svc_reg_xprt_class(struct svc_xprt_class *xcl)
76{
77 struct svc_xprt_class *cl;
78 int res = -EEXIST;
79
80 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
81
82 INIT_LIST_HEAD(&xcl->xcl_list);
83 spin_lock(&svc_xprt_class_lock);
84 /* Make sure there isn't already a class with the same name */
85 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
86 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
87 goto out;
88 }
89 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
90 res = 0;
91out:
92 spin_unlock(&svc_xprt_class_lock);
93 return res;
94}
95EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
96
97void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
98{
99 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
100 spin_lock(&svc_xprt_class_lock);
101 list_del_init(&xcl->xcl_list);
102 spin_unlock(&svc_xprt_class_lock);
103}
104EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
105
106/*
107 * Format the transport list for printing
108 */
109int svc_print_xprts(char *buf, int maxlen)
110{
111 struct svc_xprt_class *xcl;
112 char tmpstr[80];
113 int len = 0;
114 buf[0] = '\0';
115
116 spin_lock(&svc_xprt_class_lock);
117 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
118 int slen;
119
120 sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
121 slen = strlen(tmpstr);
122 if (len + slen > maxlen)
123 break;
124 len += slen;
125 strcat(buf, tmpstr);
126 }
127 spin_unlock(&svc_xprt_class_lock);
128
129 return len;
130}
131
132static void svc_xprt_free(struct kref *kref)
133{
134 struct svc_xprt *xprt =
135 container_of(kref, struct svc_xprt, xpt_ref);
136 struct module *owner = xprt->xpt_class->xcl_owner;
137 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
138 svcauth_unix_info_release(xprt);
139 put_net(xprt->xpt_net);
140 /* See comment on corresponding get in xs_setup_bc_tcp(): */
141 if (xprt->xpt_bc_xprt)
142 xprt_put(xprt->xpt_bc_xprt);
143 if (xprt->xpt_bc_xps)
144 xprt_switch_put(xprt->xpt_bc_xps);
145 xprt->xpt_ops->xpo_free(xprt);
146 module_put(owner);
147}
148
149void svc_xprt_put(struct svc_xprt *xprt)
150{
151 kref_put(&xprt->xpt_ref, svc_xprt_free);
152}
153EXPORT_SYMBOL_GPL(svc_xprt_put);
154
155/*
156 * Called by transport drivers to initialize the transport independent
157 * portion of the transport instance.
158 */
159void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
160 struct svc_xprt *xprt, struct svc_serv *serv)
161{
162 memset(xprt, 0, sizeof(*xprt));
163 xprt->xpt_class = xcl;
164 xprt->xpt_ops = xcl->xcl_ops;
165 kref_init(&xprt->xpt_ref);
166 xprt->xpt_server = serv;
167 INIT_LIST_HEAD(&xprt->xpt_list);
168 INIT_LIST_HEAD(&xprt->xpt_ready);
169 INIT_LIST_HEAD(&xprt->xpt_deferred);
170 INIT_LIST_HEAD(&xprt->xpt_users);
171 mutex_init(&xprt->xpt_mutex);
172 spin_lock_init(&xprt->xpt_lock);
173 set_bit(XPT_BUSY, &xprt->xpt_flags);
174 rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
175 xprt->xpt_net = get_net(net);
176 strcpy(xprt->xpt_remotebuf, "uninitialized");
177}
178EXPORT_SYMBOL_GPL(svc_xprt_init);
179
180static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
181 struct svc_serv *serv,
182 struct net *net,
183 const int family,
184 const unsigned short port,
185 int flags)
186{
187 struct sockaddr_in sin = {
188 .sin_family = AF_INET,
189 .sin_addr.s_addr = htonl(INADDR_ANY),
190 .sin_port = htons(port),
191 };
192#if IS_ENABLED(CONFIG_IPV6)
193 struct sockaddr_in6 sin6 = {
194 .sin6_family = AF_INET6,
195 .sin6_addr = IN6ADDR_ANY_INIT,
196 .sin6_port = htons(port),
197 };
198#endif
199 struct sockaddr *sap;
200 size_t len;
201
202 switch (family) {
203 case PF_INET:
204 sap = (struct sockaddr *)&sin;
205 len = sizeof(sin);
206 break;
207#if IS_ENABLED(CONFIG_IPV6)
208 case PF_INET6:
209 sap = (struct sockaddr *)&sin6;
210 len = sizeof(sin6);
211 break;
212#endif
213 default:
214 return ERR_PTR(-EAFNOSUPPORT);
215 }
216
217 return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
218}
219
220/*
221 * svc_xprt_received conditionally queues the transport for processing
222 * by another thread. The caller must hold the XPT_BUSY bit and must
223 * not thereafter touch transport data.
224 *
225 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
226 * insufficient) data.
227 */
228static void svc_xprt_received(struct svc_xprt *xprt)
229{
230 if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
231 WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
232 return;
233 }
234
235 /* As soon as we clear busy, the xprt could be closed and
236 * 'put', so we need a reference to call svc_enqueue_xprt with:
237 */
238 svc_xprt_get(xprt);
239 smp_mb__before_atomic();
240 clear_bit(XPT_BUSY, &xprt->xpt_flags);
241 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
242 svc_xprt_put(xprt);
243}
244
245void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
246{
247 clear_bit(XPT_TEMP, &new->xpt_flags);
248 spin_lock_bh(&serv->sv_lock);
249 list_add(&new->xpt_list, &serv->sv_permsocks);
250 spin_unlock_bh(&serv->sv_lock);
251 svc_xprt_received(new);
252}
253
254static int _svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
255 struct net *net, const int family,
256 const unsigned short port, int flags)
257{
258 struct svc_xprt_class *xcl;
259
260 spin_lock(&svc_xprt_class_lock);
261 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
262 struct svc_xprt *newxprt;
263 unsigned short newport;
264
265 if (strcmp(xprt_name, xcl->xcl_name))
266 continue;
267
268 if (!try_module_get(xcl->xcl_owner))
269 goto err;
270
271 spin_unlock(&svc_xprt_class_lock);
272 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
273 if (IS_ERR(newxprt)) {
274 module_put(xcl->xcl_owner);
275 return PTR_ERR(newxprt);
276 }
277 svc_add_new_perm_xprt(serv, newxprt);
278 newport = svc_xprt_local_port(newxprt);
279 return newport;
280 }
281 err:
282 spin_unlock(&svc_xprt_class_lock);
283 /* This errno is exposed to user space. Provide a reasonable
284 * perror msg for a bad transport. */
285 return -EPROTONOSUPPORT;
286}
287
288int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
289 struct net *net, const int family,
290 const unsigned short port, int flags)
291{
292 int err;
293
294 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
295 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags);
296 if (err == -EPROTONOSUPPORT) {
297 request_module("svc%s", xprt_name);
298 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags);
299 }
300 if (err)
301 dprintk("svc: transport %s not found, err %d\n",
302 xprt_name, err);
303 return err;
304}
305EXPORT_SYMBOL_GPL(svc_create_xprt);
306
307/*
308 * Copy the local and remote xprt addresses to the rqstp structure
309 */
310void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
311{
312 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
313 rqstp->rq_addrlen = xprt->xpt_remotelen;
314
315 /*
316 * Destination address in request is needed for binding the
317 * source address in RPC replies/callbacks later.
318 */
319 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
320 rqstp->rq_daddrlen = xprt->xpt_locallen;
321}
322EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
323
324/**
325 * svc_print_addr - Format rq_addr field for printing
326 * @rqstp: svc_rqst struct containing address to print
327 * @buf: target buffer for formatted address
328 * @len: length of target buffer
329 *
330 */
331char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
332{
333 return __svc_print_addr(svc_addr(rqstp), buf, len);
334}
335EXPORT_SYMBOL_GPL(svc_print_addr);
336
337static bool svc_xprt_slots_in_range(struct svc_xprt *xprt)
338{
339 unsigned int limit = svc_rpc_per_connection_limit;
340 int nrqsts = atomic_read(&xprt->xpt_nr_rqsts);
341
342 return limit == 0 || (nrqsts >= 0 && nrqsts < limit);
343}
344
345static bool svc_xprt_reserve_slot(struct svc_rqst *rqstp, struct svc_xprt *xprt)
346{
347 if (!test_bit(RQ_DATA, &rqstp->rq_flags)) {
348 if (!svc_xprt_slots_in_range(xprt))
349 return false;
350 atomic_inc(&xprt->xpt_nr_rqsts);
351 set_bit(RQ_DATA, &rqstp->rq_flags);
352 }
353 return true;
354}
355
356static void svc_xprt_release_slot(struct svc_rqst *rqstp)
357{
358 struct svc_xprt *xprt = rqstp->rq_xprt;
359 if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
360 atomic_dec(&xprt->xpt_nr_rqsts);
361 svc_xprt_enqueue(xprt);
362 }
363}
364
365static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
366{
367 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
368 return true;
369 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED))) {
370 if (xprt->xpt_ops->xpo_has_wspace(xprt) &&
371 svc_xprt_slots_in_range(xprt))
372 return true;
373 trace_svc_xprt_no_write_space(xprt);
374 return false;
375 }
376 return false;
377}
378
379void svc_xprt_do_enqueue(struct svc_xprt *xprt)
380{
381 struct svc_pool *pool;
382 struct svc_rqst *rqstp = NULL;
383 int cpu;
384
385 if (!svc_xprt_has_something_to_do(xprt))
386 return;
387
388 /* Mark transport as busy. It will remain in this state until
389 * the provider calls svc_xprt_received. We update XPT_BUSY
390 * atomically because it also guards against trying to enqueue
391 * the transport twice.
392 */
393 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
394 return;
395
396 cpu = get_cpu();
397 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
398
399 atomic_long_inc(&pool->sp_stats.packets);
400
401 spin_lock_bh(&pool->sp_lock);
402 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
403 pool->sp_stats.sockets_queued++;
404 spin_unlock_bh(&pool->sp_lock);
405
406 /* find a thread for this xprt */
407 rcu_read_lock();
408 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
409 if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags))
410 continue;
411 atomic_long_inc(&pool->sp_stats.threads_woken);
412 rqstp->rq_qtime = ktime_get();
413 wake_up_process(rqstp->rq_task);
414 goto out_unlock;
415 }
416 set_bit(SP_CONGESTED, &pool->sp_flags);
417 rqstp = NULL;
418out_unlock:
419 rcu_read_unlock();
420 put_cpu();
421 trace_svc_xprt_do_enqueue(xprt, rqstp);
422}
423EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
424
425/*
426 * Queue up a transport with data pending. If there are idle nfsd
427 * processes, wake 'em up.
428 *
429 */
430void svc_xprt_enqueue(struct svc_xprt *xprt)
431{
432 if (test_bit(XPT_BUSY, &xprt->xpt_flags))
433 return;
434 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
435}
436EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
437
438/*
439 * Dequeue the first transport, if there is one.
440 */
441static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
442{
443 struct svc_xprt *xprt = NULL;
444
445 if (list_empty(&pool->sp_sockets))
446 goto out;
447
448 spin_lock_bh(&pool->sp_lock);
449 if (likely(!list_empty(&pool->sp_sockets))) {
450 xprt = list_first_entry(&pool->sp_sockets,
451 struct svc_xprt, xpt_ready);
452 list_del_init(&xprt->xpt_ready);
453 svc_xprt_get(xprt);
454 }
455 spin_unlock_bh(&pool->sp_lock);
456out:
457 return xprt;
458}
459
460/**
461 * svc_reserve - change the space reserved for the reply to a request.
462 * @rqstp: The request in question
463 * @space: new max space to reserve
464 *
465 * Each request reserves some space on the output queue of the transport
466 * to make sure the reply fits. This function reduces that reserved
467 * space to be the amount of space used already, plus @space.
468 *
469 */
470void svc_reserve(struct svc_rqst *rqstp, int space)
471{
472 space += rqstp->rq_res.head[0].iov_len;
473
474 if (space < rqstp->rq_reserved) {
475 struct svc_xprt *xprt = rqstp->rq_xprt;
476 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
477 rqstp->rq_reserved = space;
478
479 svc_xprt_enqueue(xprt);
480 }
481}
482EXPORT_SYMBOL_GPL(svc_reserve);
483
484static void svc_xprt_release(struct svc_rqst *rqstp)
485{
486 struct svc_xprt *xprt = rqstp->rq_xprt;
487
488 xprt->xpt_ops->xpo_release_rqst(rqstp);
489
490 kfree(rqstp->rq_deferred);
491 rqstp->rq_deferred = NULL;
492
493 svc_free_res_pages(rqstp);
494 rqstp->rq_res.page_len = 0;
495 rqstp->rq_res.page_base = 0;
496
497 /* Reset response buffer and release
498 * the reservation.
499 * But first, check that enough space was reserved
500 * for the reply, otherwise we have a bug!
501 */
502 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
503 printk(KERN_ERR "RPC request reserved %d but used %d\n",
504 rqstp->rq_reserved,
505 rqstp->rq_res.len);
506
507 rqstp->rq_res.head[0].iov_len = 0;
508 svc_reserve(rqstp, 0);
509 svc_xprt_release_slot(rqstp);
510 rqstp->rq_xprt = NULL;
511 svc_xprt_put(xprt);
512}
513
514/*
515 * Some svc_serv's will have occasional work to do, even when a xprt is not
516 * waiting to be serviced. This function is there to "kick" a task in one of
517 * those services so that it can wake up and do that work. Note that we only
518 * bother with pool 0 as we don't need to wake up more than one thread for
519 * this purpose.
520 */
521void svc_wake_up(struct svc_serv *serv)
522{
523 struct svc_rqst *rqstp;
524 struct svc_pool *pool;
525
526 pool = &serv->sv_pools[0];
527
528 rcu_read_lock();
529 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
530 /* skip any that aren't queued */
531 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
532 continue;
533 rcu_read_unlock();
534 wake_up_process(rqstp->rq_task);
535 trace_svc_wake_up(rqstp->rq_task->pid);
536 return;
537 }
538 rcu_read_unlock();
539
540 /* No free entries available */
541 set_bit(SP_TASK_PENDING, &pool->sp_flags);
542 smp_wmb();
543 trace_svc_wake_up(0);
544}
545EXPORT_SYMBOL_GPL(svc_wake_up);
546
547int svc_port_is_privileged(struct sockaddr *sin)
548{
549 switch (sin->sa_family) {
550 case AF_INET:
551 return ntohs(((struct sockaddr_in *)sin)->sin_port)
552 < PROT_SOCK;
553 case AF_INET6:
554 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
555 < PROT_SOCK;
556 default:
557 return 0;
558 }
559}
560
561/*
562 * Make sure that we don't have too many active connections. If we have,
563 * something must be dropped. It's not clear what will happen if we allow
564 * "too many" connections, but when dealing with network-facing software,
565 * we have to code defensively. Here we do that by imposing hard limits.
566 *
567 * There's no point in trying to do random drop here for DoS
568 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
569 * attacker can easily beat that.
570 *
571 * The only somewhat efficient mechanism would be if drop old
572 * connections from the same IP first. But right now we don't even
573 * record the client IP in svc_sock.
574 *
575 * single-threaded services that expect a lot of clients will probably
576 * need to set sv_maxconn to override the default value which is based
577 * on the number of threads
578 */
579static void svc_check_conn_limits(struct svc_serv *serv)
580{
581 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
582 (serv->sv_nrthreads+3) * 20;
583
584 if (serv->sv_tmpcnt > limit) {
585 struct svc_xprt *xprt = NULL;
586 spin_lock_bh(&serv->sv_lock);
587 if (!list_empty(&serv->sv_tempsocks)) {
588 /* Try to help the admin */
589 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
590 serv->sv_name, serv->sv_maxconn ?
591 "max number of connections" :
592 "number of threads");
593 /*
594 * Always select the oldest connection. It's not fair,
595 * but so is life
596 */
597 xprt = list_entry(serv->sv_tempsocks.prev,
598 struct svc_xprt,
599 xpt_list);
600 set_bit(XPT_CLOSE, &xprt->xpt_flags);
601 svc_xprt_get(xprt);
602 }
603 spin_unlock_bh(&serv->sv_lock);
604
605 if (xprt) {
606 svc_xprt_enqueue(xprt);
607 svc_xprt_put(xprt);
608 }
609 }
610}
611
612static int svc_alloc_arg(struct svc_rqst *rqstp)
613{
614 struct svc_serv *serv = rqstp->rq_server;
615 struct xdr_buf *arg;
616 int pages;
617 int i;
618
619 /* now allocate needed pages. If we get a failure, sleep briefly */
620 pages = (serv->sv_max_mesg + 2 * PAGE_SIZE) >> PAGE_SHIFT;
621 if (pages > RPCSVC_MAXPAGES) {
622 pr_warn_once("svc: warning: pages=%u > RPCSVC_MAXPAGES=%lu\n",
623 pages, RPCSVC_MAXPAGES);
624 /* use as many pages as possible */
625 pages = RPCSVC_MAXPAGES;
626 }
627 for (i = 0; i < pages ; i++)
628 while (rqstp->rq_pages[i] == NULL) {
629 struct page *p = alloc_page(GFP_KERNEL);
630 if (!p) {
631 set_current_state(TASK_INTERRUPTIBLE);
632 if (signalled() || kthread_should_stop()) {
633 set_current_state(TASK_RUNNING);
634 return -EINTR;
635 }
636 schedule_timeout(msecs_to_jiffies(500));
637 }
638 rqstp->rq_pages[i] = p;
639 }
640 rqstp->rq_page_end = &rqstp->rq_pages[i];
641 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
642
643 /* Make arg->head point to first page and arg->pages point to rest */
644 arg = &rqstp->rq_arg;
645 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
646 arg->head[0].iov_len = PAGE_SIZE;
647 arg->pages = rqstp->rq_pages + 1;
648 arg->page_base = 0;
649 /* save at least one page for response */
650 arg->page_len = (pages-2)*PAGE_SIZE;
651 arg->len = (pages-1)*PAGE_SIZE;
652 arg->tail[0].iov_len = 0;
653 return 0;
654}
655
656static bool
657rqst_should_sleep(struct svc_rqst *rqstp)
658{
659 struct svc_pool *pool = rqstp->rq_pool;
660
661 /* did someone call svc_wake_up? */
662 if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
663 return false;
664
665 /* was a socket queued? */
666 if (!list_empty(&pool->sp_sockets))
667 return false;
668
669 /* are we shutting down? */
670 if (signalled() || kthread_should_stop())
671 return false;
672
673 /* are we freezing? */
674 if (freezing(current))
675 return false;
676
677 return true;
678}
679
680static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
681{
682 struct svc_pool *pool = rqstp->rq_pool;
683 long time_left = 0;
684
685 /* rq_xprt should be clear on entry */
686 WARN_ON_ONCE(rqstp->rq_xprt);
687
688 rqstp->rq_xprt = svc_xprt_dequeue(pool);
689 if (rqstp->rq_xprt)
690 goto out_found;
691
692 /*
693 * We have to be able to interrupt this wait
694 * to bring down the daemons ...
695 */
696 set_current_state(TASK_INTERRUPTIBLE);
697 smp_mb__before_atomic();
698 clear_bit(SP_CONGESTED, &pool->sp_flags);
699 clear_bit(RQ_BUSY, &rqstp->rq_flags);
700 smp_mb__after_atomic();
701
702 if (likely(rqst_should_sleep(rqstp)))
703 time_left = schedule_timeout(timeout);
704 else
705 __set_current_state(TASK_RUNNING);
706
707 try_to_freeze();
708
709 set_bit(RQ_BUSY, &rqstp->rq_flags);
710 smp_mb__after_atomic();
711 rqstp->rq_xprt = svc_xprt_dequeue(pool);
712 if (rqstp->rq_xprt)
713 goto out_found;
714
715 if (!time_left)
716 atomic_long_inc(&pool->sp_stats.threads_timedout);
717
718 if (signalled() || kthread_should_stop())
719 return ERR_PTR(-EINTR);
720 return ERR_PTR(-EAGAIN);
721out_found:
722 /* Normally we will wait up to 5 seconds for any required
723 * cache information to be provided.
724 */
725 if (!test_bit(SP_CONGESTED, &pool->sp_flags))
726 rqstp->rq_chandle.thread_wait = 5*HZ;
727 else
728 rqstp->rq_chandle.thread_wait = 1*HZ;
729 trace_svc_xprt_dequeue(rqstp);
730 return rqstp->rq_xprt;
731}
732
733static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
734{
735 spin_lock_bh(&serv->sv_lock);
736 set_bit(XPT_TEMP, &newxpt->xpt_flags);
737 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
738 serv->sv_tmpcnt++;
739 if (serv->sv_temptimer.function == NULL) {
740 /* setup timer to age temp transports */
741 serv->sv_temptimer.function = svc_age_temp_xprts;
742 mod_timer(&serv->sv_temptimer,
743 jiffies + svc_conn_age_period * HZ);
744 }
745 spin_unlock_bh(&serv->sv_lock);
746 svc_xprt_received(newxpt);
747}
748
749static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
750{
751 struct svc_serv *serv = rqstp->rq_server;
752 int len = 0;
753
754 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
755 dprintk("svc_recv: found XPT_CLOSE\n");
756 if (test_and_clear_bit(XPT_KILL_TEMP, &xprt->xpt_flags))
757 xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
758 svc_delete_xprt(xprt);
759 /* Leave XPT_BUSY set on the dead xprt: */
760 goto out;
761 }
762 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
763 struct svc_xprt *newxpt;
764 /*
765 * We know this module_get will succeed because the
766 * listener holds a reference too
767 */
768 __module_get(xprt->xpt_class->xcl_owner);
769 svc_check_conn_limits(xprt->xpt_server);
770 newxpt = xprt->xpt_ops->xpo_accept(xprt);
771 if (newxpt)
772 svc_add_new_temp_xprt(serv, newxpt);
773 else
774 module_put(xprt->xpt_class->xcl_owner);
775 } else if (svc_xprt_reserve_slot(rqstp, xprt)) {
776 /* XPT_DATA|XPT_DEFERRED case: */
777 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
778 rqstp, rqstp->rq_pool->sp_id, xprt,
779 kref_read(&xprt->xpt_ref));
780 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
781 if (rqstp->rq_deferred)
782 len = svc_deferred_recv(rqstp);
783 else
784 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
785 rqstp->rq_stime = ktime_get();
786 rqstp->rq_reserved = serv->sv_max_mesg;
787 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
788 }
789 /* clear XPT_BUSY: */
790 svc_xprt_received(xprt);
791out:
792 trace_svc_handle_xprt(xprt, len);
793 return len;
794}
795
796/*
797 * Receive the next request on any transport. This code is carefully
798 * organised not to touch any cachelines in the shared svc_serv
799 * structure, only cachelines in the local svc_pool.
800 */
801int svc_recv(struct svc_rqst *rqstp, long timeout)
802{
803 struct svc_xprt *xprt = NULL;
804 struct svc_serv *serv = rqstp->rq_server;
805 int len, err;
806
807 dprintk("svc: server %p waiting for data (to = %ld)\n",
808 rqstp, timeout);
809
810 if (rqstp->rq_xprt)
811 printk(KERN_ERR
812 "svc_recv: service %p, transport not NULL!\n",
813 rqstp);
814
815 err = svc_alloc_arg(rqstp);
816 if (err)
817 goto out;
818
819 try_to_freeze();
820 cond_resched();
821 err = -EINTR;
822 if (signalled() || kthread_should_stop())
823 goto out;
824
825 xprt = svc_get_next_xprt(rqstp, timeout);
826 if (IS_ERR(xprt)) {
827 err = PTR_ERR(xprt);
828 goto out;
829 }
830
831 len = svc_handle_xprt(rqstp, xprt);
832
833 /* No data, incomplete (TCP) read, or accept() */
834 err = -EAGAIN;
835 if (len <= 0)
836 goto out_release;
837
838 clear_bit(XPT_OLD, &xprt->xpt_flags);
839
840 xprt->xpt_ops->xpo_secure_port(rqstp);
841 rqstp->rq_chandle.defer = svc_defer;
842 rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
843
844 if (serv->sv_stats)
845 serv->sv_stats->netcnt++;
846 trace_svc_recv(rqstp, len);
847 return len;
848out_release:
849 rqstp->rq_res.len = 0;
850 svc_xprt_release(rqstp);
851out:
852 return err;
853}
854EXPORT_SYMBOL_GPL(svc_recv);
855
856/*
857 * Drop request
858 */
859void svc_drop(struct svc_rqst *rqstp)
860{
861 trace_svc_drop(rqstp);
862 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
863 svc_xprt_release(rqstp);
864}
865EXPORT_SYMBOL_GPL(svc_drop);
866
867/*
868 * Return reply to client.
869 */
870int svc_send(struct svc_rqst *rqstp)
871{
872 struct svc_xprt *xprt;
873 int len = -EFAULT;
874 struct xdr_buf *xb;
875
876 xprt = rqstp->rq_xprt;
877 if (!xprt)
878 goto out;
879
880 /* release the receive skb before sending the reply */
881 xprt->xpt_ops->xpo_release_rqst(rqstp);
882
883 /* calculate over-all length */
884 xb = &rqstp->rq_res;
885 xb->len = xb->head[0].iov_len +
886 xb->page_len +
887 xb->tail[0].iov_len;
888
889 /* Grab mutex to serialize outgoing data. */
890 mutex_lock(&xprt->xpt_mutex);
891 trace_svc_stats_latency(rqstp);
892 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
893 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
894 len = -ENOTCONN;
895 else
896 len = xprt->xpt_ops->xpo_sendto(rqstp);
897 mutex_unlock(&xprt->xpt_mutex);
898 rpc_wake_up(&xprt->xpt_bc_pending);
899 trace_svc_send(rqstp, len);
900 svc_xprt_release(rqstp);
901
902 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
903 len = 0;
904out:
905 return len;
906}
907
908/*
909 * Timer function to close old temporary transports, using
910 * a mark-and-sweep algorithm.
911 */
912static void svc_age_temp_xprts(struct timer_list *t)
913{
914 struct svc_serv *serv = from_timer(serv, t, sv_temptimer);
915 struct svc_xprt *xprt;
916 struct list_head *le, *next;
917
918 dprintk("svc_age_temp_xprts\n");
919
920 if (!spin_trylock_bh(&serv->sv_lock)) {
921 /* busy, try again 1 sec later */
922 dprintk("svc_age_temp_xprts: busy\n");
923 mod_timer(&serv->sv_temptimer, jiffies + HZ);
924 return;
925 }
926
927 list_for_each_safe(le, next, &serv->sv_tempsocks) {
928 xprt = list_entry(le, struct svc_xprt, xpt_list);
929
930 /* First time through, just mark it OLD. Second time
931 * through, close it. */
932 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
933 continue;
934 if (kref_read(&xprt->xpt_ref) > 1 ||
935 test_bit(XPT_BUSY, &xprt->xpt_flags))
936 continue;
937 list_del_init(le);
938 set_bit(XPT_CLOSE, &xprt->xpt_flags);
939 dprintk("queuing xprt %p for closing\n", xprt);
940
941 /* a thread will dequeue and close it soon */
942 svc_xprt_enqueue(xprt);
943 }
944 spin_unlock_bh(&serv->sv_lock);
945
946 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
947}
948
949/* Close temporary transports whose xpt_local matches server_addr immediately
950 * instead of waiting for them to be picked up by the timer.
951 *
952 * This is meant to be called from a notifier_block that runs when an ip
953 * address is deleted.
954 */
955void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
956{
957 struct svc_xprt *xprt;
958 struct list_head *le, *next;
959 LIST_HEAD(to_be_closed);
960
961 spin_lock_bh(&serv->sv_lock);
962 list_for_each_safe(le, next, &serv->sv_tempsocks) {
963 xprt = list_entry(le, struct svc_xprt, xpt_list);
964 if (rpc_cmp_addr(server_addr, (struct sockaddr *)
965 &xprt->xpt_local)) {
966 dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
967 list_move(le, &to_be_closed);
968 }
969 }
970 spin_unlock_bh(&serv->sv_lock);
971
972 while (!list_empty(&to_be_closed)) {
973 le = to_be_closed.next;
974 list_del_init(le);
975 xprt = list_entry(le, struct svc_xprt, xpt_list);
976 set_bit(XPT_CLOSE, &xprt->xpt_flags);
977 set_bit(XPT_KILL_TEMP, &xprt->xpt_flags);
978 dprintk("svc_age_temp_xprts_now: queuing xprt %p for closing\n",
979 xprt);
980 svc_xprt_enqueue(xprt);
981 }
982}
983EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
984
985static void call_xpt_users(struct svc_xprt *xprt)
986{
987 struct svc_xpt_user *u;
988
989 spin_lock(&xprt->xpt_lock);
990 while (!list_empty(&xprt->xpt_users)) {
991 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
992 list_del(&u->list);
993 u->callback(u);
994 }
995 spin_unlock(&xprt->xpt_lock);
996}
997
998/*
999 * Remove a dead transport
1000 */
1001static void svc_delete_xprt(struct svc_xprt *xprt)
1002{
1003 struct svc_serv *serv = xprt->xpt_server;
1004 struct svc_deferred_req *dr;
1005
1006 /* Only do this once */
1007 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1008 BUG();
1009
1010 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1011 xprt->xpt_ops->xpo_detach(xprt);
1012
1013 spin_lock_bh(&serv->sv_lock);
1014 list_del_init(&xprt->xpt_list);
1015 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
1016 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1017 serv->sv_tmpcnt--;
1018 spin_unlock_bh(&serv->sv_lock);
1019
1020 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1021 kfree(dr);
1022
1023 call_xpt_users(xprt);
1024 svc_xprt_put(xprt);
1025}
1026
1027void svc_close_xprt(struct svc_xprt *xprt)
1028{
1029 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1030 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1031 /* someone else will have to effect the close */
1032 return;
1033 /*
1034 * We expect svc_close_xprt() to work even when no threads are
1035 * running (e.g., while configuring the server before starting
1036 * any threads), so if the transport isn't busy, we delete
1037 * it ourself:
1038 */
1039 svc_delete_xprt(xprt);
1040}
1041EXPORT_SYMBOL_GPL(svc_close_xprt);
1042
1043static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1044{
1045 struct svc_xprt *xprt;
1046 int ret = 0;
1047
1048 spin_lock(&serv->sv_lock);
1049 list_for_each_entry(xprt, xprt_list, xpt_list) {
1050 if (xprt->xpt_net != net)
1051 continue;
1052 ret++;
1053 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1054 svc_xprt_enqueue(xprt);
1055 }
1056 spin_unlock(&serv->sv_lock);
1057 return ret;
1058}
1059
1060static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1061{
1062 struct svc_pool *pool;
1063 struct svc_xprt *xprt;
1064 struct svc_xprt *tmp;
1065 int i;
1066
1067 for (i = 0; i < serv->sv_nrpools; i++) {
1068 pool = &serv->sv_pools[i];
1069
1070 spin_lock_bh(&pool->sp_lock);
1071 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1072 if (xprt->xpt_net != net)
1073 continue;
1074 list_del_init(&xprt->xpt_ready);
1075 spin_unlock_bh(&pool->sp_lock);
1076 return xprt;
1077 }
1078 spin_unlock_bh(&pool->sp_lock);
1079 }
1080 return NULL;
1081}
1082
1083static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1084{
1085 struct svc_xprt *xprt;
1086
1087 while ((xprt = svc_dequeue_net(serv, net))) {
1088 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1089 svc_delete_xprt(xprt);
1090 }
1091}
1092
1093/*
1094 * Server threads may still be running (especially in the case where the
1095 * service is still running in other network namespaces).
1096 *
1097 * So we shut down sockets the same way we would on a running server, by
1098 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1099 * the close. In the case there are no such other threads,
1100 * threads running, svc_clean_up_xprts() does a simple version of a
1101 * server's main event loop, and in the case where there are other
1102 * threads, we may need to wait a little while and then check again to
1103 * see if they're done.
1104 */
1105void svc_close_net(struct svc_serv *serv, struct net *net)
1106{
1107 int delay = 0;
1108
1109 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1110 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1111
1112 svc_clean_up_xprts(serv, net);
1113 msleep(delay++);
1114 }
1115}
1116
1117/*
1118 * Handle defer and revisit of requests
1119 */
1120
1121static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1122{
1123 struct svc_deferred_req *dr =
1124 container_of(dreq, struct svc_deferred_req, handle);
1125 struct svc_xprt *xprt = dr->xprt;
1126
1127 spin_lock(&xprt->xpt_lock);
1128 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1129 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1130 spin_unlock(&xprt->xpt_lock);
1131 dprintk("revisit canceled\n");
1132 svc_xprt_put(xprt);
1133 trace_svc_drop_deferred(dr);
1134 kfree(dr);
1135 return;
1136 }
1137 dprintk("revisit queued\n");
1138 dr->xprt = NULL;
1139 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1140 spin_unlock(&xprt->xpt_lock);
1141 svc_xprt_enqueue(xprt);
1142 svc_xprt_put(xprt);
1143}
1144
1145/*
1146 * Save the request off for later processing. The request buffer looks
1147 * like this:
1148 *
1149 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1150 *
1151 * This code can only handle requests that consist of an xprt-header
1152 * and rpc-header.
1153 */
1154static struct cache_deferred_req *svc_defer(struct cache_req *req)
1155{
1156 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1157 struct svc_deferred_req *dr;
1158
1159 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1160 return NULL; /* if more than a page, give up FIXME */
1161 if (rqstp->rq_deferred) {
1162 dr = rqstp->rq_deferred;
1163 rqstp->rq_deferred = NULL;
1164 } else {
1165 size_t skip;
1166 size_t size;
1167 /* FIXME maybe discard if size too large */
1168 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1169 dr = kmalloc(size, GFP_KERNEL);
1170 if (dr == NULL)
1171 return NULL;
1172
1173 dr->handle.owner = rqstp->rq_server;
1174 dr->prot = rqstp->rq_prot;
1175 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1176 dr->addrlen = rqstp->rq_addrlen;
1177 dr->daddr = rqstp->rq_daddr;
1178 dr->argslen = rqstp->rq_arg.len >> 2;
1179 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1180
1181 /* back up head to the start of the buffer and copy */
1182 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1183 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1184 dr->argslen << 2);
1185 }
1186 svc_xprt_get(rqstp->rq_xprt);
1187 dr->xprt = rqstp->rq_xprt;
1188 set_bit(RQ_DROPME, &rqstp->rq_flags);
1189
1190 dr->handle.revisit = svc_revisit;
1191 trace_svc_defer(rqstp);
1192 return &dr->handle;
1193}
1194
1195/*
1196 * recv data from a deferred request into an active one
1197 */
1198static int svc_deferred_recv(struct svc_rqst *rqstp)
1199{
1200 struct svc_deferred_req *dr = rqstp->rq_deferred;
1201
1202 /* setup iov_base past transport header */
1203 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1204 /* The iov_len does not include the transport header bytes */
1205 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1206 rqstp->rq_arg.page_len = 0;
1207 /* The rq_arg.len includes the transport header bytes */
1208 rqstp->rq_arg.len = dr->argslen<<2;
1209 rqstp->rq_prot = dr->prot;
1210 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1211 rqstp->rq_addrlen = dr->addrlen;
1212 /* Save off transport header len in case we get deferred again */
1213 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1214 rqstp->rq_daddr = dr->daddr;
1215 rqstp->rq_respages = rqstp->rq_pages;
1216 return (dr->argslen<<2) - dr->xprt_hlen;
1217}
1218
1219
1220static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1221{
1222 struct svc_deferred_req *dr = NULL;
1223
1224 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1225 return NULL;
1226 spin_lock(&xprt->xpt_lock);
1227 if (!list_empty(&xprt->xpt_deferred)) {
1228 dr = list_entry(xprt->xpt_deferred.next,
1229 struct svc_deferred_req,
1230 handle.recent);
1231 list_del_init(&dr->handle.recent);
1232 trace_svc_revisit_deferred(dr);
1233 } else
1234 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1235 spin_unlock(&xprt->xpt_lock);
1236 return dr;
1237}
1238
1239/**
1240 * svc_find_xprt - find an RPC transport instance
1241 * @serv: pointer to svc_serv to search
1242 * @xcl_name: C string containing transport's class name
1243 * @net: owner net pointer
1244 * @af: Address family of transport's local address
1245 * @port: transport's IP port number
1246 *
1247 * Return the transport instance pointer for the endpoint accepting
1248 * connections/peer traffic from the specified transport class,
1249 * address family and port.
1250 *
1251 * Specifying 0 for the address family or port is effectively a
1252 * wild-card, and will result in matching the first transport in the
1253 * service's list that has a matching class name.
1254 */
1255struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1256 struct net *net, const sa_family_t af,
1257 const unsigned short port)
1258{
1259 struct svc_xprt *xprt;
1260 struct svc_xprt *found = NULL;
1261
1262 /* Sanity check the args */
1263 if (serv == NULL || xcl_name == NULL)
1264 return found;
1265
1266 spin_lock_bh(&serv->sv_lock);
1267 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1268 if (xprt->xpt_net != net)
1269 continue;
1270 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1271 continue;
1272 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1273 continue;
1274 if (port != 0 && port != svc_xprt_local_port(xprt))
1275 continue;
1276 found = xprt;
1277 svc_xprt_get(xprt);
1278 break;
1279 }
1280 spin_unlock_bh(&serv->sv_lock);
1281 return found;
1282}
1283EXPORT_SYMBOL_GPL(svc_find_xprt);
1284
1285static int svc_one_xprt_name(const struct svc_xprt *xprt,
1286 char *pos, int remaining)
1287{
1288 int len;
1289
1290 len = snprintf(pos, remaining, "%s %u\n",
1291 xprt->xpt_class->xcl_name,
1292 svc_xprt_local_port(xprt));
1293 if (len >= remaining)
1294 return -ENAMETOOLONG;
1295 return len;
1296}
1297
1298/**
1299 * svc_xprt_names - format a buffer with a list of transport names
1300 * @serv: pointer to an RPC service
1301 * @buf: pointer to a buffer to be filled in
1302 * @buflen: length of buffer to be filled in
1303 *
1304 * Fills in @buf with a string containing a list of transport names,
1305 * each name terminated with '\n'.
1306 *
1307 * Returns positive length of the filled-in string on success; otherwise
1308 * a negative errno value is returned if an error occurs.
1309 */
1310int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1311{
1312 struct svc_xprt *xprt;
1313 int len, totlen;
1314 char *pos;
1315
1316 /* Sanity check args */
1317 if (!serv)
1318 return 0;
1319
1320 spin_lock_bh(&serv->sv_lock);
1321
1322 pos = buf;
1323 totlen = 0;
1324 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1325 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1326 if (len < 0) {
1327 *buf = '\0';
1328 totlen = len;
1329 }
1330 if (len <= 0)
1331 break;
1332
1333 pos += len;
1334 totlen += len;
1335 }
1336
1337 spin_unlock_bh(&serv->sv_lock);
1338 return totlen;
1339}
1340EXPORT_SYMBOL_GPL(svc_xprt_names);
1341
1342
1343/*----------------------------------------------------------------------------*/
1344
1345static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1346{
1347 unsigned int pidx = (unsigned int)*pos;
1348 struct svc_serv *serv = m->private;
1349
1350 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1351
1352 if (!pidx)
1353 return SEQ_START_TOKEN;
1354 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1355}
1356
1357static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1358{
1359 struct svc_pool *pool = p;
1360 struct svc_serv *serv = m->private;
1361
1362 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1363
1364 if (p == SEQ_START_TOKEN) {
1365 pool = &serv->sv_pools[0];
1366 } else {
1367 unsigned int pidx = (pool - &serv->sv_pools[0]);
1368 if (pidx < serv->sv_nrpools-1)
1369 pool = &serv->sv_pools[pidx+1];
1370 else
1371 pool = NULL;
1372 }
1373 ++*pos;
1374 return pool;
1375}
1376
1377static void svc_pool_stats_stop(struct seq_file *m, void *p)
1378{
1379}
1380
1381static int svc_pool_stats_show(struct seq_file *m, void *p)
1382{
1383 struct svc_pool *pool = p;
1384
1385 if (p == SEQ_START_TOKEN) {
1386 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1387 return 0;
1388 }
1389
1390 seq_printf(m, "%u %lu %lu %lu %lu\n",
1391 pool->sp_id,
1392 (unsigned long)atomic_long_read(&pool->sp_stats.packets),
1393 pool->sp_stats.sockets_queued,
1394 (unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1395 (unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1396
1397 return 0;
1398}
1399
1400static const struct seq_operations svc_pool_stats_seq_ops = {
1401 .start = svc_pool_stats_start,
1402 .next = svc_pool_stats_next,
1403 .stop = svc_pool_stats_stop,
1404 .show = svc_pool_stats_show,
1405};
1406
1407int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1408{
1409 int err;
1410
1411 err = seq_open(file, &svc_pool_stats_seq_ops);
1412 if (!err)
1413 ((struct seq_file *) file->private_data)->private = serv;
1414 return err;
1415}
1416EXPORT_SYMBOL(svc_pool_stats_open);
1417
1418/*----------------------------------------------------------------------------*/
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/net/sunrpc/svc_xprt.c
4 *
5 * Author: Tom Tucker <tom@opengridcomputing.com>
6 */
7
8#include <linux/sched.h>
9#include <linux/sched/mm.h>
10#include <linux/errno.h>
11#include <linux/freezer.h>
12#include <linux/slab.h>
13#include <net/sock.h>
14#include <linux/sunrpc/addr.h>
15#include <linux/sunrpc/stats.h>
16#include <linux/sunrpc/svc_xprt.h>
17#include <linux/sunrpc/svcsock.h>
18#include <linux/sunrpc/xprt.h>
19#include <linux/sunrpc/bc_xprt.h>
20#include <linux/module.h>
21#include <linux/netdevice.h>
22#include <trace/events/sunrpc.h>
23
24#define RPCDBG_FACILITY RPCDBG_SVCXPRT
25
26static unsigned int svc_rpc_per_connection_limit __read_mostly;
27module_param(svc_rpc_per_connection_limit, uint, 0644);
28
29
30static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
31static int svc_deferred_recv(struct svc_rqst *rqstp);
32static struct cache_deferred_req *svc_defer(struct cache_req *req);
33static void svc_age_temp_xprts(struct timer_list *t);
34static void svc_delete_xprt(struct svc_xprt *xprt);
35
36/* apparently the "standard" is that clients close
37 * idle connections after 5 minutes, servers after
38 * 6 minutes
39 * http://nfsv4bat.org/Documents/ConnectAThon/1996/nfstcp.pdf
40 */
41static int svc_conn_age_period = 6*60;
42
43/* List of registered transport classes */
44static DEFINE_SPINLOCK(svc_xprt_class_lock);
45static LIST_HEAD(svc_xprt_class_list);
46
47/* SMP locking strategy:
48 *
49 * svc_pool->sp_lock protects most of the fields of that pool.
50 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
51 * when both need to be taken (rare), svc_serv->sv_lock is first.
52 * The "service mutex" protects svc_serv->sv_nrthread.
53 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
54 * and the ->sk_info_authunix cache.
55 *
56 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
57 * enqueued multiply. During normal transport processing this bit
58 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
59 * Providers should not manipulate this bit directly.
60 *
61 * Some flags can be set to certain values at any time
62 * providing that certain rules are followed:
63 *
64 * XPT_CONN, XPT_DATA:
65 * - Can be set or cleared at any time.
66 * - After a set, svc_xprt_enqueue must be called to enqueue
67 * the transport for processing.
68 * - After a clear, the transport must be read/accepted.
69 * If this succeeds, it must be set again.
70 * XPT_CLOSE:
71 * - Can set at any time. It is never cleared.
72 * XPT_DEAD:
73 * - Can only be set while XPT_BUSY is held which ensures
74 * that no other thread will be using the transport or will
75 * try to set XPT_DEAD.
76 */
77
78/**
79 * svc_reg_xprt_class - Register a server-side RPC transport class
80 * @xcl: New transport class to be registered
81 *
82 * Returns zero on success; otherwise a negative errno is returned.
83 */
84int svc_reg_xprt_class(struct svc_xprt_class *xcl)
85{
86 struct svc_xprt_class *cl;
87 int res = -EEXIST;
88
89 INIT_LIST_HEAD(&xcl->xcl_list);
90 spin_lock(&svc_xprt_class_lock);
91 /* Make sure there isn't already a class with the same name */
92 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
93 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
94 goto out;
95 }
96 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
97 res = 0;
98out:
99 spin_unlock(&svc_xprt_class_lock);
100 return res;
101}
102EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
103
104/**
105 * svc_unreg_xprt_class - Unregister a server-side RPC transport class
106 * @xcl: Transport class to be unregistered
107 *
108 */
109void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
110{
111 spin_lock(&svc_xprt_class_lock);
112 list_del_init(&xcl->xcl_list);
113 spin_unlock(&svc_xprt_class_lock);
114}
115EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
116
117/**
118 * svc_print_xprts - Format the transport list for printing
119 * @buf: target buffer for formatted address
120 * @maxlen: length of target buffer
121 *
122 * Fills in @buf with a string containing a list of transport names, each name
123 * terminated with '\n'. If the buffer is too small, some entries may be
124 * missing, but it is guaranteed that all lines in the output buffer are
125 * complete.
126 *
127 * Returns positive length of the filled-in string.
128 */
129int svc_print_xprts(char *buf, int maxlen)
130{
131 struct svc_xprt_class *xcl;
132 char tmpstr[80];
133 int len = 0;
134 buf[0] = '\0';
135
136 spin_lock(&svc_xprt_class_lock);
137 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
138 int slen;
139
140 slen = snprintf(tmpstr, sizeof(tmpstr), "%s %d\n",
141 xcl->xcl_name, xcl->xcl_max_payload);
142 if (slen >= sizeof(tmpstr) || len + slen >= maxlen)
143 break;
144 len += slen;
145 strcat(buf, tmpstr);
146 }
147 spin_unlock(&svc_xprt_class_lock);
148
149 return len;
150}
151
152/**
153 * svc_xprt_deferred_close - Close a transport
154 * @xprt: transport instance
155 *
156 * Used in contexts that need to defer the work of shutting down
157 * the transport to an nfsd thread.
158 */
159void svc_xprt_deferred_close(struct svc_xprt *xprt)
160{
161 if (!test_and_set_bit(XPT_CLOSE, &xprt->xpt_flags))
162 svc_xprt_enqueue(xprt);
163}
164EXPORT_SYMBOL_GPL(svc_xprt_deferred_close);
165
166static void svc_xprt_free(struct kref *kref)
167{
168 struct svc_xprt *xprt =
169 container_of(kref, struct svc_xprt, xpt_ref);
170 struct module *owner = xprt->xpt_class->xcl_owner;
171 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
172 svcauth_unix_info_release(xprt);
173 put_cred(xprt->xpt_cred);
174 put_net_track(xprt->xpt_net, &xprt->ns_tracker);
175 /* See comment on corresponding get in xs_setup_bc_tcp(): */
176 if (xprt->xpt_bc_xprt)
177 xprt_put(xprt->xpt_bc_xprt);
178 if (xprt->xpt_bc_xps)
179 xprt_switch_put(xprt->xpt_bc_xps);
180 trace_svc_xprt_free(xprt);
181 xprt->xpt_ops->xpo_free(xprt);
182 module_put(owner);
183}
184
185void svc_xprt_put(struct svc_xprt *xprt)
186{
187 kref_put(&xprt->xpt_ref, svc_xprt_free);
188}
189EXPORT_SYMBOL_GPL(svc_xprt_put);
190
191/*
192 * Called by transport drivers to initialize the transport independent
193 * portion of the transport instance.
194 */
195void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
196 struct svc_xprt *xprt, struct svc_serv *serv)
197{
198 memset(xprt, 0, sizeof(*xprt));
199 xprt->xpt_class = xcl;
200 xprt->xpt_ops = xcl->xcl_ops;
201 kref_init(&xprt->xpt_ref);
202 xprt->xpt_server = serv;
203 INIT_LIST_HEAD(&xprt->xpt_list);
204 INIT_LIST_HEAD(&xprt->xpt_deferred);
205 INIT_LIST_HEAD(&xprt->xpt_users);
206 mutex_init(&xprt->xpt_mutex);
207 spin_lock_init(&xprt->xpt_lock);
208 set_bit(XPT_BUSY, &xprt->xpt_flags);
209 xprt->xpt_net = get_net_track(net, &xprt->ns_tracker, GFP_ATOMIC);
210 strcpy(xprt->xpt_remotebuf, "uninitialized");
211}
212EXPORT_SYMBOL_GPL(svc_xprt_init);
213
214static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
215 struct svc_serv *serv,
216 struct net *net,
217 const int family,
218 const unsigned short port,
219 int flags)
220{
221 struct sockaddr_in sin = {
222 .sin_family = AF_INET,
223 .sin_addr.s_addr = htonl(INADDR_ANY),
224 .sin_port = htons(port),
225 };
226#if IS_ENABLED(CONFIG_IPV6)
227 struct sockaddr_in6 sin6 = {
228 .sin6_family = AF_INET6,
229 .sin6_addr = IN6ADDR_ANY_INIT,
230 .sin6_port = htons(port),
231 };
232#endif
233 struct svc_xprt *xprt;
234 struct sockaddr *sap;
235 size_t len;
236
237 switch (family) {
238 case PF_INET:
239 sap = (struct sockaddr *)&sin;
240 len = sizeof(sin);
241 break;
242#if IS_ENABLED(CONFIG_IPV6)
243 case PF_INET6:
244 sap = (struct sockaddr *)&sin6;
245 len = sizeof(sin6);
246 break;
247#endif
248 default:
249 return ERR_PTR(-EAFNOSUPPORT);
250 }
251
252 xprt = xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
253 if (IS_ERR(xprt))
254 trace_svc_xprt_create_err(serv->sv_program->pg_name,
255 xcl->xcl_name, sap, len, xprt);
256 return xprt;
257}
258
259/**
260 * svc_xprt_received - start next receiver thread
261 * @xprt: controlling transport
262 *
263 * The caller must hold the XPT_BUSY bit and must
264 * not thereafter touch transport data.
265 *
266 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
267 * insufficient) data.
268 */
269void svc_xprt_received(struct svc_xprt *xprt)
270{
271 if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
272 WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
273 return;
274 }
275
276 /* As soon as we clear busy, the xprt could be closed and
277 * 'put', so we need a reference to call svc_xprt_enqueue with:
278 */
279 svc_xprt_get(xprt);
280 smp_mb__before_atomic();
281 clear_bit(XPT_BUSY, &xprt->xpt_flags);
282 svc_xprt_enqueue(xprt);
283 svc_xprt_put(xprt);
284}
285EXPORT_SYMBOL_GPL(svc_xprt_received);
286
287void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
288{
289 clear_bit(XPT_TEMP, &new->xpt_flags);
290 spin_lock_bh(&serv->sv_lock);
291 list_add(&new->xpt_list, &serv->sv_permsocks);
292 spin_unlock_bh(&serv->sv_lock);
293 svc_xprt_received(new);
294}
295
296static int _svc_xprt_create(struct svc_serv *serv, const char *xprt_name,
297 struct net *net, const int family,
298 const unsigned short port, int flags,
299 const struct cred *cred)
300{
301 struct svc_xprt_class *xcl;
302
303 spin_lock(&svc_xprt_class_lock);
304 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
305 struct svc_xprt *newxprt;
306 unsigned short newport;
307
308 if (strcmp(xprt_name, xcl->xcl_name))
309 continue;
310
311 if (!try_module_get(xcl->xcl_owner))
312 goto err;
313
314 spin_unlock(&svc_xprt_class_lock);
315 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
316 if (IS_ERR(newxprt)) {
317 module_put(xcl->xcl_owner);
318 return PTR_ERR(newxprt);
319 }
320 newxprt->xpt_cred = get_cred(cred);
321 svc_add_new_perm_xprt(serv, newxprt);
322 newport = svc_xprt_local_port(newxprt);
323 return newport;
324 }
325 err:
326 spin_unlock(&svc_xprt_class_lock);
327 /* This errno is exposed to user space. Provide a reasonable
328 * perror msg for a bad transport. */
329 return -EPROTONOSUPPORT;
330}
331
332/**
333 * svc_xprt_create - Add a new listener to @serv
334 * @serv: target RPC service
335 * @xprt_name: transport class name
336 * @net: network namespace
337 * @family: network address family
338 * @port: listener port
339 * @flags: SVC_SOCK flags
340 * @cred: credential to bind to this transport
341 *
342 * Return values:
343 * %0: New listener added successfully
344 * %-EPROTONOSUPPORT: Requested transport type not supported
345 */
346int svc_xprt_create(struct svc_serv *serv, const char *xprt_name,
347 struct net *net, const int family,
348 const unsigned short port, int flags,
349 const struct cred *cred)
350{
351 int err;
352
353 err = _svc_xprt_create(serv, xprt_name, net, family, port, flags, cred);
354 if (err == -EPROTONOSUPPORT) {
355 request_module("svc%s", xprt_name);
356 err = _svc_xprt_create(serv, xprt_name, net, family, port, flags, cred);
357 }
358 return err;
359}
360EXPORT_SYMBOL_GPL(svc_xprt_create);
361
362/*
363 * Copy the local and remote xprt addresses to the rqstp structure
364 */
365void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
366{
367 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
368 rqstp->rq_addrlen = xprt->xpt_remotelen;
369
370 /*
371 * Destination address in request is needed for binding the
372 * source address in RPC replies/callbacks later.
373 */
374 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
375 rqstp->rq_daddrlen = xprt->xpt_locallen;
376}
377EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
378
379/**
380 * svc_print_addr - Format rq_addr field for printing
381 * @rqstp: svc_rqst struct containing address to print
382 * @buf: target buffer for formatted address
383 * @len: length of target buffer
384 *
385 */
386char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
387{
388 return __svc_print_addr(svc_addr(rqstp), buf, len);
389}
390EXPORT_SYMBOL_GPL(svc_print_addr);
391
392static bool svc_xprt_slots_in_range(struct svc_xprt *xprt)
393{
394 unsigned int limit = svc_rpc_per_connection_limit;
395 int nrqsts = atomic_read(&xprt->xpt_nr_rqsts);
396
397 return limit == 0 || (nrqsts >= 0 && nrqsts < limit);
398}
399
400static bool svc_xprt_reserve_slot(struct svc_rqst *rqstp, struct svc_xprt *xprt)
401{
402 if (!test_bit(RQ_DATA, &rqstp->rq_flags)) {
403 if (!svc_xprt_slots_in_range(xprt))
404 return false;
405 atomic_inc(&xprt->xpt_nr_rqsts);
406 set_bit(RQ_DATA, &rqstp->rq_flags);
407 }
408 return true;
409}
410
411static void svc_xprt_release_slot(struct svc_rqst *rqstp)
412{
413 struct svc_xprt *xprt = rqstp->rq_xprt;
414 if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
415 atomic_dec(&xprt->xpt_nr_rqsts);
416 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
417 svc_xprt_enqueue(xprt);
418 }
419}
420
421static bool svc_xprt_ready(struct svc_xprt *xprt)
422{
423 unsigned long xpt_flags;
424
425 /*
426 * If another cpu has recently updated xpt_flags,
427 * sk_sock->flags, xpt_reserved, or xpt_nr_rqsts, we need to
428 * know about it; otherwise it's possible that both that cpu and
429 * this one could call svc_xprt_enqueue() without either
430 * svc_xprt_enqueue() recognizing that the conditions below
431 * are satisfied, and we could stall indefinitely:
432 */
433 smp_rmb();
434 xpt_flags = READ_ONCE(xprt->xpt_flags);
435
436 trace_svc_xprt_enqueue(xprt, xpt_flags);
437 if (xpt_flags & BIT(XPT_BUSY))
438 return false;
439 if (xpt_flags & (BIT(XPT_CONN) | BIT(XPT_CLOSE) | BIT(XPT_HANDSHAKE)))
440 return true;
441 if (xpt_flags & (BIT(XPT_DATA) | BIT(XPT_DEFERRED))) {
442 if (xprt->xpt_ops->xpo_has_wspace(xprt) &&
443 svc_xprt_slots_in_range(xprt))
444 return true;
445 trace_svc_xprt_no_write_space(xprt);
446 return false;
447 }
448 return false;
449}
450
451/**
452 * svc_xprt_enqueue - Queue a transport on an idle nfsd thread
453 * @xprt: transport with data pending
454 *
455 */
456void svc_xprt_enqueue(struct svc_xprt *xprt)
457{
458 struct svc_pool *pool;
459
460 if (!svc_xprt_ready(xprt))
461 return;
462
463 /* Mark transport as busy. It will remain in this state until
464 * the provider calls svc_xprt_received. We update XPT_BUSY
465 * atomically because it also guards against trying to enqueue
466 * the transport twice.
467 */
468 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
469 return;
470
471 pool = svc_pool_for_cpu(xprt->xpt_server);
472
473 percpu_counter_inc(&pool->sp_sockets_queued);
474 lwq_enqueue(&xprt->xpt_ready, &pool->sp_xprts);
475
476 svc_pool_wake_idle_thread(pool);
477}
478EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
479
480/*
481 * Dequeue the first transport, if there is one.
482 */
483static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
484{
485 struct svc_xprt *xprt = NULL;
486
487 xprt = lwq_dequeue(&pool->sp_xprts, struct svc_xprt, xpt_ready);
488 if (xprt)
489 svc_xprt_get(xprt);
490 return xprt;
491}
492
493/**
494 * svc_reserve - change the space reserved for the reply to a request.
495 * @rqstp: The request in question
496 * @space: new max space to reserve
497 *
498 * Each request reserves some space on the output queue of the transport
499 * to make sure the reply fits. This function reduces that reserved
500 * space to be the amount of space used already, plus @space.
501 *
502 */
503void svc_reserve(struct svc_rqst *rqstp, int space)
504{
505 struct svc_xprt *xprt = rqstp->rq_xprt;
506
507 space += rqstp->rq_res.head[0].iov_len;
508
509 if (xprt && space < rqstp->rq_reserved) {
510 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
511 rqstp->rq_reserved = space;
512 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
513 svc_xprt_enqueue(xprt);
514 }
515}
516EXPORT_SYMBOL_GPL(svc_reserve);
517
518static void free_deferred(struct svc_xprt *xprt, struct svc_deferred_req *dr)
519{
520 if (!dr)
521 return;
522
523 xprt->xpt_ops->xpo_release_ctxt(xprt, dr->xprt_ctxt);
524 kfree(dr);
525}
526
527static void svc_xprt_release(struct svc_rqst *rqstp)
528{
529 struct svc_xprt *xprt = rqstp->rq_xprt;
530
531 xprt->xpt_ops->xpo_release_ctxt(xprt, rqstp->rq_xprt_ctxt);
532 rqstp->rq_xprt_ctxt = NULL;
533
534 free_deferred(xprt, rqstp->rq_deferred);
535 rqstp->rq_deferred = NULL;
536
537 svc_rqst_release_pages(rqstp);
538 rqstp->rq_res.page_len = 0;
539 rqstp->rq_res.page_base = 0;
540
541 /* Reset response buffer and release
542 * the reservation.
543 * But first, check that enough space was reserved
544 * for the reply, otherwise we have a bug!
545 */
546 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
547 printk(KERN_ERR "RPC request reserved %d but used %d\n",
548 rqstp->rq_reserved,
549 rqstp->rq_res.len);
550
551 rqstp->rq_res.head[0].iov_len = 0;
552 svc_reserve(rqstp, 0);
553 svc_xprt_release_slot(rqstp);
554 rqstp->rq_xprt = NULL;
555 svc_xprt_put(xprt);
556}
557
558/**
559 * svc_wake_up - Wake up a service thread for non-transport work
560 * @serv: RPC service
561 *
562 * Some svc_serv's will have occasional work to do, even when a xprt is not
563 * waiting to be serviced. This function is there to "kick" a task in one of
564 * those services so that it can wake up and do that work. Note that we only
565 * bother with pool 0 as we don't need to wake up more than one thread for
566 * this purpose.
567 */
568void svc_wake_up(struct svc_serv *serv)
569{
570 struct svc_pool *pool = &serv->sv_pools[0];
571
572 set_bit(SP_TASK_PENDING, &pool->sp_flags);
573 svc_pool_wake_idle_thread(pool);
574}
575EXPORT_SYMBOL_GPL(svc_wake_up);
576
577int svc_port_is_privileged(struct sockaddr *sin)
578{
579 switch (sin->sa_family) {
580 case AF_INET:
581 return ntohs(((struct sockaddr_in *)sin)->sin_port)
582 < PROT_SOCK;
583 case AF_INET6:
584 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
585 < PROT_SOCK;
586 default:
587 return 0;
588 }
589}
590
591/*
592 * Make sure that we don't have too many active connections. If we have,
593 * something must be dropped. It's not clear what will happen if we allow
594 * "too many" connections, but when dealing with network-facing software,
595 * we have to code defensively. Here we do that by imposing hard limits.
596 *
597 * There's no point in trying to do random drop here for DoS
598 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
599 * attacker can easily beat that.
600 *
601 * The only somewhat efficient mechanism would be if drop old
602 * connections from the same IP first. But right now we don't even
603 * record the client IP in svc_sock.
604 *
605 * single-threaded services that expect a lot of clients will probably
606 * need to set sv_maxconn to override the default value which is based
607 * on the number of threads
608 */
609static void svc_check_conn_limits(struct svc_serv *serv)
610{
611 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
612 (serv->sv_nrthreads+3) * 20;
613
614 if (serv->sv_tmpcnt > limit) {
615 struct svc_xprt *xprt = NULL;
616 spin_lock_bh(&serv->sv_lock);
617 if (!list_empty(&serv->sv_tempsocks)) {
618 /* Try to help the admin */
619 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
620 serv->sv_name, serv->sv_maxconn ?
621 "max number of connections" :
622 "number of threads");
623 /*
624 * Always select the oldest connection. It's not fair,
625 * but so is life
626 */
627 xprt = list_entry(serv->sv_tempsocks.prev,
628 struct svc_xprt,
629 xpt_list);
630 set_bit(XPT_CLOSE, &xprt->xpt_flags);
631 svc_xprt_get(xprt);
632 }
633 spin_unlock_bh(&serv->sv_lock);
634
635 if (xprt) {
636 svc_xprt_enqueue(xprt);
637 svc_xprt_put(xprt);
638 }
639 }
640}
641
642static bool svc_alloc_arg(struct svc_rqst *rqstp)
643{
644 struct svc_serv *serv = rqstp->rq_server;
645 struct xdr_buf *arg = &rqstp->rq_arg;
646 unsigned long pages, filled, ret;
647
648 pages = (serv->sv_max_mesg + 2 * PAGE_SIZE) >> PAGE_SHIFT;
649 if (pages > RPCSVC_MAXPAGES) {
650 pr_warn_once("svc: warning: pages=%lu > RPCSVC_MAXPAGES=%lu\n",
651 pages, RPCSVC_MAXPAGES);
652 /* use as many pages as possible */
653 pages = RPCSVC_MAXPAGES;
654 }
655
656 for (filled = 0; filled < pages; filled = ret) {
657 ret = alloc_pages_bulk_array(GFP_KERNEL, pages,
658 rqstp->rq_pages);
659 if (ret > filled)
660 /* Made progress, don't sleep yet */
661 continue;
662
663 set_current_state(TASK_IDLE);
664 if (svc_thread_should_stop(rqstp)) {
665 set_current_state(TASK_RUNNING);
666 return false;
667 }
668 trace_svc_alloc_arg_err(pages, ret);
669 memalloc_retry_wait(GFP_KERNEL);
670 }
671 rqstp->rq_page_end = &rqstp->rq_pages[pages];
672 rqstp->rq_pages[pages] = NULL; /* this might be seen in nfsd_splice_actor() */
673
674 /* Make arg->head point to first page and arg->pages point to rest */
675 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
676 arg->head[0].iov_len = PAGE_SIZE;
677 arg->pages = rqstp->rq_pages + 1;
678 arg->page_base = 0;
679 /* save at least one page for response */
680 arg->page_len = (pages-2)*PAGE_SIZE;
681 arg->len = (pages-1)*PAGE_SIZE;
682 arg->tail[0].iov_len = 0;
683
684 rqstp->rq_xid = xdr_zero;
685 return true;
686}
687
688static bool
689svc_thread_should_sleep(struct svc_rqst *rqstp)
690{
691 struct svc_pool *pool = rqstp->rq_pool;
692
693 /* did someone call svc_wake_up? */
694 if (test_bit(SP_TASK_PENDING, &pool->sp_flags))
695 return false;
696
697 /* was a socket queued? */
698 if (!lwq_empty(&pool->sp_xprts))
699 return false;
700
701 /* are we shutting down? */
702 if (svc_thread_should_stop(rqstp))
703 return false;
704
705#if defined(CONFIG_SUNRPC_BACKCHANNEL)
706 if (svc_is_backchannel(rqstp)) {
707 if (!lwq_empty(&rqstp->rq_server->sv_cb_list))
708 return false;
709 }
710#endif
711
712 return true;
713}
714
715static void svc_thread_wait_for_work(struct svc_rqst *rqstp)
716{
717 struct svc_pool *pool = rqstp->rq_pool;
718
719 if (svc_thread_should_sleep(rqstp)) {
720 set_current_state(TASK_IDLE | TASK_FREEZABLE);
721 llist_add(&rqstp->rq_idle, &pool->sp_idle_threads);
722 if (likely(svc_thread_should_sleep(rqstp)))
723 schedule();
724
725 while (!llist_del_first_this(&pool->sp_idle_threads,
726 &rqstp->rq_idle)) {
727 /* Work just became available. This thread can only
728 * handle it after removing rqstp from the idle
729 * list. If that attempt failed, some other thread
730 * must have queued itself after finding no
731 * work to do, so that thread has taken responsibly
732 * for this new work. This thread can safely sleep
733 * until woken again.
734 */
735 schedule();
736 set_current_state(TASK_IDLE | TASK_FREEZABLE);
737 }
738 __set_current_state(TASK_RUNNING);
739 } else {
740 cond_resched();
741 }
742 try_to_freeze();
743}
744
745static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
746{
747 spin_lock_bh(&serv->sv_lock);
748 set_bit(XPT_TEMP, &newxpt->xpt_flags);
749 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
750 serv->sv_tmpcnt++;
751 if (serv->sv_temptimer.function == NULL) {
752 /* setup timer to age temp transports */
753 serv->sv_temptimer.function = svc_age_temp_xprts;
754 mod_timer(&serv->sv_temptimer,
755 jiffies + svc_conn_age_period * HZ);
756 }
757 spin_unlock_bh(&serv->sv_lock);
758 svc_xprt_received(newxpt);
759}
760
761static void svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
762{
763 struct svc_serv *serv = rqstp->rq_server;
764 int len = 0;
765
766 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
767 if (test_and_clear_bit(XPT_KILL_TEMP, &xprt->xpt_flags))
768 xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
769 svc_delete_xprt(xprt);
770 /* Leave XPT_BUSY set on the dead xprt: */
771 goto out;
772 }
773 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
774 struct svc_xprt *newxpt;
775 /*
776 * We know this module_get will succeed because the
777 * listener holds a reference too
778 */
779 __module_get(xprt->xpt_class->xcl_owner);
780 svc_check_conn_limits(xprt->xpt_server);
781 newxpt = xprt->xpt_ops->xpo_accept(xprt);
782 if (newxpt) {
783 newxpt->xpt_cred = get_cred(xprt->xpt_cred);
784 svc_add_new_temp_xprt(serv, newxpt);
785 trace_svc_xprt_accept(newxpt, serv->sv_name);
786 } else {
787 module_put(xprt->xpt_class->xcl_owner);
788 }
789 svc_xprt_received(xprt);
790 } else if (test_bit(XPT_HANDSHAKE, &xprt->xpt_flags)) {
791 xprt->xpt_ops->xpo_handshake(xprt);
792 svc_xprt_received(xprt);
793 } else if (svc_xprt_reserve_slot(rqstp, xprt)) {
794 /* XPT_DATA|XPT_DEFERRED case: */
795 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
796 if (rqstp->rq_deferred)
797 len = svc_deferred_recv(rqstp);
798 else
799 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
800 rqstp->rq_reserved = serv->sv_max_mesg;
801 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
802 if (len <= 0)
803 goto out;
804
805 trace_svc_xdr_recvfrom(&rqstp->rq_arg);
806
807 clear_bit(XPT_OLD, &xprt->xpt_flags);
808
809 rqstp->rq_chandle.defer = svc_defer;
810
811 if (serv->sv_stats)
812 serv->sv_stats->netcnt++;
813 percpu_counter_inc(&rqstp->rq_pool->sp_messages_arrived);
814 rqstp->rq_stime = ktime_get();
815 svc_process(rqstp);
816 } else
817 svc_xprt_received(xprt);
818
819out:
820 rqstp->rq_res.len = 0;
821 svc_xprt_release(rqstp);
822}
823
824static void svc_thread_wake_next(struct svc_rqst *rqstp)
825{
826 if (!svc_thread_should_sleep(rqstp))
827 /* More work pending after I dequeued some,
828 * wake another worker
829 */
830 svc_pool_wake_idle_thread(rqstp->rq_pool);
831}
832
833/**
834 * svc_recv - Receive and process the next request on any transport
835 * @rqstp: an idle RPC service thread
836 *
837 * This code is carefully organised not to touch any cachelines in
838 * the shared svc_serv structure, only cachelines in the local
839 * svc_pool.
840 */
841void svc_recv(struct svc_rqst *rqstp)
842{
843 struct svc_pool *pool = rqstp->rq_pool;
844
845 if (!svc_alloc_arg(rqstp))
846 return;
847
848 svc_thread_wait_for_work(rqstp);
849
850 clear_bit(SP_TASK_PENDING, &pool->sp_flags);
851
852 if (svc_thread_should_stop(rqstp)) {
853 svc_thread_wake_next(rqstp);
854 return;
855 }
856
857 rqstp->rq_xprt = svc_xprt_dequeue(pool);
858 if (rqstp->rq_xprt) {
859 struct svc_xprt *xprt = rqstp->rq_xprt;
860
861 svc_thread_wake_next(rqstp);
862 /* Normally we will wait up to 5 seconds for any required
863 * cache information to be provided. When there are no
864 * idle threads, we reduce the wait time.
865 */
866 if (pool->sp_idle_threads.first)
867 rqstp->rq_chandle.thread_wait = 5 * HZ;
868 else
869 rqstp->rq_chandle.thread_wait = 1 * HZ;
870
871 trace_svc_xprt_dequeue(rqstp);
872 svc_handle_xprt(rqstp, xprt);
873 }
874
875#if defined(CONFIG_SUNRPC_BACKCHANNEL)
876 if (svc_is_backchannel(rqstp)) {
877 struct svc_serv *serv = rqstp->rq_server;
878 struct rpc_rqst *req;
879
880 req = lwq_dequeue(&serv->sv_cb_list,
881 struct rpc_rqst, rq_bc_list);
882 if (req) {
883 svc_thread_wake_next(rqstp);
884 svc_process_bc(req, rqstp);
885 }
886 }
887#endif
888}
889EXPORT_SYMBOL_GPL(svc_recv);
890
891/*
892 * Drop request
893 */
894void svc_drop(struct svc_rqst *rqstp)
895{
896 trace_svc_drop(rqstp);
897}
898EXPORT_SYMBOL_GPL(svc_drop);
899
900/**
901 * svc_send - Return reply to client
902 * @rqstp: RPC transaction context
903 *
904 */
905void svc_send(struct svc_rqst *rqstp)
906{
907 struct svc_xprt *xprt;
908 struct xdr_buf *xb;
909 int status;
910
911 xprt = rqstp->rq_xprt;
912
913 /* calculate over-all length */
914 xb = &rqstp->rq_res;
915 xb->len = xb->head[0].iov_len +
916 xb->page_len +
917 xb->tail[0].iov_len;
918 trace_svc_xdr_sendto(rqstp->rq_xid, xb);
919 trace_svc_stats_latency(rqstp);
920
921 status = xprt->xpt_ops->xpo_sendto(rqstp);
922
923 trace_svc_send(rqstp, status);
924}
925
926/*
927 * Timer function to close old temporary transports, using
928 * a mark-and-sweep algorithm.
929 */
930static void svc_age_temp_xprts(struct timer_list *t)
931{
932 struct svc_serv *serv = from_timer(serv, t, sv_temptimer);
933 struct svc_xprt *xprt;
934 struct list_head *le, *next;
935
936 dprintk("svc_age_temp_xprts\n");
937
938 if (!spin_trylock_bh(&serv->sv_lock)) {
939 /* busy, try again 1 sec later */
940 dprintk("svc_age_temp_xprts: busy\n");
941 mod_timer(&serv->sv_temptimer, jiffies + HZ);
942 return;
943 }
944
945 list_for_each_safe(le, next, &serv->sv_tempsocks) {
946 xprt = list_entry(le, struct svc_xprt, xpt_list);
947
948 /* First time through, just mark it OLD. Second time
949 * through, close it. */
950 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
951 continue;
952 if (kref_read(&xprt->xpt_ref) > 1 ||
953 test_bit(XPT_BUSY, &xprt->xpt_flags))
954 continue;
955 list_del_init(le);
956 set_bit(XPT_CLOSE, &xprt->xpt_flags);
957 dprintk("queuing xprt %p for closing\n", xprt);
958
959 /* a thread will dequeue and close it soon */
960 svc_xprt_enqueue(xprt);
961 }
962 spin_unlock_bh(&serv->sv_lock);
963
964 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
965}
966
967/* Close temporary transports whose xpt_local matches server_addr immediately
968 * instead of waiting for them to be picked up by the timer.
969 *
970 * This is meant to be called from a notifier_block that runs when an ip
971 * address is deleted.
972 */
973void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
974{
975 struct svc_xprt *xprt;
976 struct list_head *le, *next;
977 LIST_HEAD(to_be_closed);
978
979 spin_lock_bh(&serv->sv_lock);
980 list_for_each_safe(le, next, &serv->sv_tempsocks) {
981 xprt = list_entry(le, struct svc_xprt, xpt_list);
982 if (rpc_cmp_addr(server_addr, (struct sockaddr *)
983 &xprt->xpt_local)) {
984 dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
985 list_move(le, &to_be_closed);
986 }
987 }
988 spin_unlock_bh(&serv->sv_lock);
989
990 while (!list_empty(&to_be_closed)) {
991 le = to_be_closed.next;
992 list_del_init(le);
993 xprt = list_entry(le, struct svc_xprt, xpt_list);
994 set_bit(XPT_CLOSE, &xprt->xpt_flags);
995 set_bit(XPT_KILL_TEMP, &xprt->xpt_flags);
996 dprintk("svc_age_temp_xprts_now: queuing xprt %p for closing\n",
997 xprt);
998 svc_xprt_enqueue(xprt);
999 }
1000}
1001EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
1002
1003static void call_xpt_users(struct svc_xprt *xprt)
1004{
1005 struct svc_xpt_user *u;
1006
1007 spin_lock(&xprt->xpt_lock);
1008 while (!list_empty(&xprt->xpt_users)) {
1009 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
1010 list_del_init(&u->list);
1011 u->callback(u);
1012 }
1013 spin_unlock(&xprt->xpt_lock);
1014}
1015
1016/*
1017 * Remove a dead transport
1018 */
1019static void svc_delete_xprt(struct svc_xprt *xprt)
1020{
1021 struct svc_serv *serv = xprt->xpt_server;
1022 struct svc_deferred_req *dr;
1023
1024 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1025 return;
1026
1027 trace_svc_xprt_detach(xprt);
1028 xprt->xpt_ops->xpo_detach(xprt);
1029 if (xprt->xpt_bc_xprt)
1030 xprt->xpt_bc_xprt->ops->close(xprt->xpt_bc_xprt);
1031
1032 spin_lock_bh(&serv->sv_lock);
1033 list_del_init(&xprt->xpt_list);
1034 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1035 serv->sv_tmpcnt--;
1036 spin_unlock_bh(&serv->sv_lock);
1037
1038 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1039 free_deferred(xprt, dr);
1040
1041 call_xpt_users(xprt);
1042 svc_xprt_put(xprt);
1043}
1044
1045/**
1046 * svc_xprt_close - Close a client connection
1047 * @xprt: transport to disconnect
1048 *
1049 */
1050void svc_xprt_close(struct svc_xprt *xprt)
1051{
1052 trace_svc_xprt_close(xprt);
1053 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1054 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1055 /* someone else will have to effect the close */
1056 return;
1057 /*
1058 * We expect svc_close_xprt() to work even when no threads are
1059 * running (e.g., while configuring the server before starting
1060 * any threads), so if the transport isn't busy, we delete
1061 * it ourself:
1062 */
1063 svc_delete_xprt(xprt);
1064}
1065EXPORT_SYMBOL_GPL(svc_xprt_close);
1066
1067static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1068{
1069 struct svc_xprt *xprt;
1070 int ret = 0;
1071
1072 spin_lock_bh(&serv->sv_lock);
1073 list_for_each_entry(xprt, xprt_list, xpt_list) {
1074 if (xprt->xpt_net != net)
1075 continue;
1076 ret++;
1077 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1078 svc_xprt_enqueue(xprt);
1079 }
1080 spin_unlock_bh(&serv->sv_lock);
1081 return ret;
1082}
1083
1084static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1085{
1086 struct svc_xprt *xprt;
1087 int i;
1088
1089 for (i = 0; i < serv->sv_nrpools; i++) {
1090 struct svc_pool *pool = &serv->sv_pools[i];
1091 struct llist_node *q, **t1, *t2;
1092
1093 q = lwq_dequeue_all(&pool->sp_xprts);
1094 lwq_for_each_safe(xprt, t1, t2, &q, xpt_ready) {
1095 if (xprt->xpt_net == net) {
1096 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1097 svc_delete_xprt(xprt);
1098 xprt = NULL;
1099 }
1100 }
1101
1102 if (q)
1103 lwq_enqueue_batch(q, &pool->sp_xprts);
1104 }
1105}
1106
1107/**
1108 * svc_xprt_destroy_all - Destroy transports associated with @serv
1109 * @serv: RPC service to be shut down
1110 * @net: target network namespace
1111 *
1112 * Server threads may still be running (especially in the case where the
1113 * service is still running in other network namespaces).
1114 *
1115 * So we shut down sockets the same way we would on a running server, by
1116 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1117 * the close. In the case there are no such other threads,
1118 * threads running, svc_clean_up_xprts() does a simple version of a
1119 * server's main event loop, and in the case where there are other
1120 * threads, we may need to wait a little while and then check again to
1121 * see if they're done.
1122 */
1123void svc_xprt_destroy_all(struct svc_serv *serv, struct net *net)
1124{
1125 int delay = 0;
1126
1127 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1128 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1129
1130 svc_clean_up_xprts(serv, net);
1131 msleep(delay++);
1132 }
1133}
1134EXPORT_SYMBOL_GPL(svc_xprt_destroy_all);
1135
1136/*
1137 * Handle defer and revisit of requests
1138 */
1139
1140static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1141{
1142 struct svc_deferred_req *dr =
1143 container_of(dreq, struct svc_deferred_req, handle);
1144 struct svc_xprt *xprt = dr->xprt;
1145
1146 spin_lock(&xprt->xpt_lock);
1147 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1148 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1149 spin_unlock(&xprt->xpt_lock);
1150 trace_svc_defer_drop(dr);
1151 free_deferred(xprt, dr);
1152 svc_xprt_put(xprt);
1153 return;
1154 }
1155 dr->xprt = NULL;
1156 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1157 spin_unlock(&xprt->xpt_lock);
1158 trace_svc_defer_queue(dr);
1159 svc_xprt_enqueue(xprt);
1160 svc_xprt_put(xprt);
1161}
1162
1163/*
1164 * Save the request off for later processing. The request buffer looks
1165 * like this:
1166 *
1167 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1168 *
1169 * This code can only handle requests that consist of an xprt-header
1170 * and rpc-header.
1171 */
1172static struct cache_deferred_req *svc_defer(struct cache_req *req)
1173{
1174 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1175 struct svc_deferred_req *dr;
1176
1177 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1178 return NULL; /* if more than a page, give up FIXME */
1179 if (rqstp->rq_deferred) {
1180 dr = rqstp->rq_deferred;
1181 rqstp->rq_deferred = NULL;
1182 } else {
1183 size_t skip;
1184 size_t size;
1185 /* FIXME maybe discard if size too large */
1186 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1187 dr = kmalloc(size, GFP_KERNEL);
1188 if (dr == NULL)
1189 return NULL;
1190
1191 dr->handle.owner = rqstp->rq_server;
1192 dr->prot = rqstp->rq_prot;
1193 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1194 dr->addrlen = rqstp->rq_addrlen;
1195 dr->daddr = rqstp->rq_daddr;
1196 dr->argslen = rqstp->rq_arg.len >> 2;
1197
1198 /* back up head to the start of the buffer and copy */
1199 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1200 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1201 dr->argslen << 2);
1202 }
1203 dr->xprt_ctxt = rqstp->rq_xprt_ctxt;
1204 rqstp->rq_xprt_ctxt = NULL;
1205 trace_svc_defer(rqstp);
1206 svc_xprt_get(rqstp->rq_xprt);
1207 dr->xprt = rqstp->rq_xprt;
1208 set_bit(RQ_DROPME, &rqstp->rq_flags);
1209
1210 dr->handle.revisit = svc_revisit;
1211 return &dr->handle;
1212}
1213
1214/*
1215 * recv data from a deferred request into an active one
1216 */
1217static noinline int svc_deferred_recv(struct svc_rqst *rqstp)
1218{
1219 struct svc_deferred_req *dr = rqstp->rq_deferred;
1220
1221 trace_svc_defer_recv(dr);
1222
1223 /* setup iov_base past transport header */
1224 rqstp->rq_arg.head[0].iov_base = dr->args;
1225 /* The iov_len does not include the transport header bytes */
1226 rqstp->rq_arg.head[0].iov_len = dr->argslen << 2;
1227 rqstp->rq_arg.page_len = 0;
1228 /* The rq_arg.len includes the transport header bytes */
1229 rqstp->rq_arg.len = dr->argslen << 2;
1230 rqstp->rq_prot = dr->prot;
1231 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1232 rqstp->rq_addrlen = dr->addrlen;
1233 /* Save off transport header len in case we get deferred again */
1234 rqstp->rq_daddr = dr->daddr;
1235 rqstp->rq_respages = rqstp->rq_pages;
1236 rqstp->rq_xprt_ctxt = dr->xprt_ctxt;
1237
1238 dr->xprt_ctxt = NULL;
1239 svc_xprt_received(rqstp->rq_xprt);
1240 return dr->argslen << 2;
1241}
1242
1243
1244static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1245{
1246 struct svc_deferred_req *dr = NULL;
1247
1248 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1249 return NULL;
1250 spin_lock(&xprt->xpt_lock);
1251 if (!list_empty(&xprt->xpt_deferred)) {
1252 dr = list_entry(xprt->xpt_deferred.next,
1253 struct svc_deferred_req,
1254 handle.recent);
1255 list_del_init(&dr->handle.recent);
1256 } else
1257 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1258 spin_unlock(&xprt->xpt_lock);
1259 return dr;
1260}
1261
1262/**
1263 * svc_find_xprt - find an RPC transport instance
1264 * @serv: pointer to svc_serv to search
1265 * @xcl_name: C string containing transport's class name
1266 * @net: owner net pointer
1267 * @af: Address family of transport's local address
1268 * @port: transport's IP port number
1269 *
1270 * Return the transport instance pointer for the endpoint accepting
1271 * connections/peer traffic from the specified transport class,
1272 * address family and port.
1273 *
1274 * Specifying 0 for the address family or port is effectively a
1275 * wild-card, and will result in matching the first transport in the
1276 * service's list that has a matching class name.
1277 */
1278struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1279 struct net *net, const sa_family_t af,
1280 const unsigned short port)
1281{
1282 struct svc_xprt *xprt;
1283 struct svc_xprt *found = NULL;
1284
1285 /* Sanity check the args */
1286 if (serv == NULL || xcl_name == NULL)
1287 return found;
1288
1289 spin_lock_bh(&serv->sv_lock);
1290 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1291 if (xprt->xpt_net != net)
1292 continue;
1293 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1294 continue;
1295 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1296 continue;
1297 if (port != 0 && port != svc_xprt_local_port(xprt))
1298 continue;
1299 found = xprt;
1300 svc_xprt_get(xprt);
1301 break;
1302 }
1303 spin_unlock_bh(&serv->sv_lock);
1304 return found;
1305}
1306EXPORT_SYMBOL_GPL(svc_find_xprt);
1307
1308static int svc_one_xprt_name(const struct svc_xprt *xprt,
1309 char *pos, int remaining)
1310{
1311 int len;
1312
1313 len = snprintf(pos, remaining, "%s %u\n",
1314 xprt->xpt_class->xcl_name,
1315 svc_xprt_local_port(xprt));
1316 if (len >= remaining)
1317 return -ENAMETOOLONG;
1318 return len;
1319}
1320
1321/**
1322 * svc_xprt_names - format a buffer with a list of transport names
1323 * @serv: pointer to an RPC service
1324 * @buf: pointer to a buffer to be filled in
1325 * @buflen: length of buffer to be filled in
1326 *
1327 * Fills in @buf with a string containing a list of transport names,
1328 * each name terminated with '\n'.
1329 *
1330 * Returns positive length of the filled-in string on success; otherwise
1331 * a negative errno value is returned if an error occurs.
1332 */
1333int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1334{
1335 struct svc_xprt *xprt;
1336 int len, totlen;
1337 char *pos;
1338
1339 /* Sanity check args */
1340 if (!serv)
1341 return 0;
1342
1343 spin_lock_bh(&serv->sv_lock);
1344
1345 pos = buf;
1346 totlen = 0;
1347 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1348 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1349 if (len < 0) {
1350 *buf = '\0';
1351 totlen = len;
1352 }
1353 if (len <= 0)
1354 break;
1355
1356 pos += len;
1357 totlen += len;
1358 }
1359
1360 spin_unlock_bh(&serv->sv_lock);
1361 return totlen;
1362}
1363EXPORT_SYMBOL_GPL(svc_xprt_names);
1364
1365/*----------------------------------------------------------------------------*/
1366
1367static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1368{
1369 unsigned int pidx = (unsigned int)*pos;
1370 struct svc_info *si = m->private;
1371
1372 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1373
1374 mutex_lock(si->mutex);
1375
1376 if (!pidx)
1377 return SEQ_START_TOKEN;
1378 if (!si->serv)
1379 return NULL;
1380 return pidx > si->serv->sv_nrpools ? NULL
1381 : &si->serv->sv_pools[pidx - 1];
1382}
1383
1384static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1385{
1386 struct svc_pool *pool = p;
1387 struct svc_info *si = m->private;
1388 struct svc_serv *serv = si->serv;
1389
1390 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1391
1392 if (!serv) {
1393 pool = NULL;
1394 } else if (p == SEQ_START_TOKEN) {
1395 pool = &serv->sv_pools[0];
1396 } else {
1397 unsigned int pidx = (pool - &serv->sv_pools[0]);
1398 if (pidx < serv->sv_nrpools-1)
1399 pool = &serv->sv_pools[pidx+1];
1400 else
1401 pool = NULL;
1402 }
1403 ++*pos;
1404 return pool;
1405}
1406
1407static void svc_pool_stats_stop(struct seq_file *m, void *p)
1408{
1409 struct svc_info *si = m->private;
1410
1411 mutex_unlock(si->mutex);
1412}
1413
1414static int svc_pool_stats_show(struct seq_file *m, void *p)
1415{
1416 struct svc_pool *pool = p;
1417
1418 if (p == SEQ_START_TOKEN) {
1419 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1420 return 0;
1421 }
1422
1423 seq_printf(m, "%u %llu %llu %llu 0\n",
1424 pool->sp_id,
1425 percpu_counter_sum_positive(&pool->sp_messages_arrived),
1426 percpu_counter_sum_positive(&pool->sp_sockets_queued),
1427 percpu_counter_sum_positive(&pool->sp_threads_woken));
1428
1429 return 0;
1430}
1431
1432static const struct seq_operations svc_pool_stats_seq_ops = {
1433 .start = svc_pool_stats_start,
1434 .next = svc_pool_stats_next,
1435 .stop = svc_pool_stats_stop,
1436 .show = svc_pool_stats_show,
1437};
1438
1439int svc_pool_stats_open(struct svc_info *info, struct file *file)
1440{
1441 struct seq_file *seq;
1442 int err;
1443
1444 err = seq_open(file, &svc_pool_stats_seq_ops);
1445 if (err)
1446 return err;
1447 seq = file->private_data;
1448 seq->private = info;
1449
1450 return 0;
1451}
1452EXPORT_SYMBOL(svc_pool_stats_open);
1453
1454/*----------------------------------------------------------------------------*/