2 * linux/net/sunrpc/svc_xprt.c
4 * Author: Tom Tucker <tom@opengridcomputing.com>
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>
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 #include <trace/events/sunrpc.h>
20 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
22 static struct svc_deferred_req
*svc_deferred_dequeue(struct svc_xprt
*xprt
);
23 static int svc_deferred_recv(struct svc_rqst
*rqstp
);
24 static struct cache_deferred_req
*svc_defer(struct cache_req
*req
);
25 static void svc_age_temp_xprts(unsigned long closure
);
26 static void svc_delete_xprt(struct svc_xprt
*xprt
);
27 static void svc_xprt_do_enqueue(struct svc_xprt
*xprt
);
29 /* apparently the "standard" is that clients close
30 * idle connections after 5 minutes, servers after
32 * http://www.connectathon.org/talks96/nfstcp.pdf
34 static int svc_conn_age_period
= 6*60;
36 /* List of registered transport classes */
37 static DEFINE_SPINLOCK(svc_xprt_class_lock
);
38 static LIST_HEAD(svc_xprt_class_list
);
40 /* SMP locking strategy:
42 * svc_pool->sp_lock protects most of the fields of that pool.
43 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
44 * when both need to be taken (rare), svc_serv->sv_lock is first.
45 * BKL protects svc_serv->sv_nrthread.
46 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
47 * and the ->sk_info_authunix cache.
49 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
50 * enqueued multiply. During normal transport processing this bit
51 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
52 * Providers should not manipulate this bit directly.
54 * Some flags can be set to certain values at any time
55 * providing that certain rules are followed:
58 * - Can be set or cleared at any time.
59 * - After a set, svc_xprt_enqueue must be called to enqueue
60 * the transport for processing.
61 * - After a clear, the transport must be read/accepted.
62 * If this succeeds, it must be set again.
64 * - Can set at any time. It is never cleared.
66 * - Can only be set while XPT_BUSY is held which ensures
67 * that no other thread will be using the transport or will
68 * try to set XPT_DEAD.
71 int svc_reg_xprt_class(struct svc_xprt_class
*xcl
)
73 struct svc_xprt_class
*cl
;
76 dprintk("svc: Adding svc transport class '%s'\n", xcl
->xcl_name
);
78 INIT_LIST_HEAD(&xcl
->xcl_list
);
79 spin_lock(&svc_xprt_class_lock
);
80 /* Make sure there isn't already a class with the same name */
81 list_for_each_entry(cl
, &svc_xprt_class_list
, xcl_list
) {
82 if (strcmp(xcl
->xcl_name
, cl
->xcl_name
) == 0)
85 list_add_tail(&xcl
->xcl_list
, &svc_xprt_class_list
);
88 spin_unlock(&svc_xprt_class_lock
);
91 EXPORT_SYMBOL_GPL(svc_reg_xprt_class
);
93 void svc_unreg_xprt_class(struct svc_xprt_class
*xcl
)
95 dprintk("svc: Removing svc transport class '%s'\n", xcl
->xcl_name
);
96 spin_lock(&svc_xprt_class_lock
);
97 list_del_init(&xcl
->xcl_list
);
98 spin_unlock(&svc_xprt_class_lock
);
100 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class
);
103 * Format the transport list for printing
105 int svc_print_xprts(char *buf
, int maxlen
)
107 struct svc_xprt_class
*xcl
;
112 spin_lock(&svc_xprt_class_lock
);
113 list_for_each_entry(xcl
, &svc_xprt_class_list
, xcl_list
) {
116 sprintf(tmpstr
, "%s %d\n", xcl
->xcl_name
, xcl
->xcl_max_payload
);
117 slen
= strlen(tmpstr
);
118 if (len
+ slen
> maxlen
)
123 spin_unlock(&svc_xprt_class_lock
);
128 static void svc_xprt_free(struct kref
*kref
)
130 struct svc_xprt
*xprt
=
131 container_of(kref
, struct svc_xprt
, xpt_ref
);
132 struct module
*owner
= xprt
->xpt_class
->xcl_owner
;
133 if (test_bit(XPT_CACHE_AUTH
, &xprt
->xpt_flags
))
134 svcauth_unix_info_release(xprt
);
135 put_net(xprt
->xpt_net
);
136 /* See comment on corresponding get in xs_setup_bc_tcp(): */
137 if (xprt
->xpt_bc_xprt
)
138 xprt_put(xprt
->xpt_bc_xprt
);
139 xprt
->xpt_ops
->xpo_free(xprt
);
143 void svc_xprt_put(struct svc_xprt
*xprt
)
145 kref_put(&xprt
->xpt_ref
, svc_xprt_free
);
147 EXPORT_SYMBOL_GPL(svc_xprt_put
);
150 * Called by transport drivers to initialize the transport independent
151 * portion of the transport instance.
153 void svc_xprt_init(struct net
*net
, struct svc_xprt_class
*xcl
,
154 struct svc_xprt
*xprt
, struct svc_serv
*serv
)
156 memset(xprt
, 0, sizeof(*xprt
));
157 xprt
->xpt_class
= xcl
;
158 xprt
->xpt_ops
= xcl
->xcl_ops
;
159 kref_init(&xprt
->xpt_ref
);
160 xprt
->xpt_server
= serv
;
161 INIT_LIST_HEAD(&xprt
->xpt_list
);
162 INIT_LIST_HEAD(&xprt
->xpt_ready
);
163 INIT_LIST_HEAD(&xprt
->xpt_deferred
);
164 INIT_LIST_HEAD(&xprt
->xpt_users
);
165 mutex_init(&xprt
->xpt_mutex
);
166 spin_lock_init(&xprt
->xpt_lock
);
167 set_bit(XPT_BUSY
, &xprt
->xpt_flags
);
168 rpc_init_wait_queue(&xprt
->xpt_bc_pending
, "xpt_bc_pending");
169 xprt
->xpt_net
= get_net(net
);
171 EXPORT_SYMBOL_GPL(svc_xprt_init
);
173 static struct svc_xprt
*__svc_xpo_create(struct svc_xprt_class
*xcl
,
174 struct svc_serv
*serv
,
177 const unsigned short port
,
180 struct sockaddr_in sin
= {
181 .sin_family
= AF_INET
,
182 .sin_addr
.s_addr
= htonl(INADDR_ANY
),
183 .sin_port
= htons(port
),
185 #if IS_ENABLED(CONFIG_IPV6)
186 struct sockaddr_in6 sin6
= {
187 .sin6_family
= AF_INET6
,
188 .sin6_addr
= IN6ADDR_ANY_INIT
,
189 .sin6_port
= htons(port
),
192 struct sockaddr
*sap
;
197 sap
= (struct sockaddr
*)&sin
;
200 #if IS_ENABLED(CONFIG_IPV6)
202 sap
= (struct sockaddr
*)&sin6
;
207 return ERR_PTR(-EAFNOSUPPORT
);
210 return xcl
->xcl_ops
->xpo_create(serv
, net
, sap
, len
, flags
);
214 * svc_xprt_received conditionally queues the transport for processing
215 * by another thread. The caller must hold the XPT_BUSY bit and must
216 * not thereafter touch transport data.
218 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
219 * insufficient) data.
221 static void svc_xprt_received(struct svc_xprt
*xprt
)
223 WARN_ON_ONCE(!test_bit(XPT_BUSY
, &xprt
->xpt_flags
));
224 if (!test_bit(XPT_BUSY
, &xprt
->xpt_flags
))
226 /* As soon as we clear busy, the xprt could be closed and
227 * 'put', so we need a reference to call svc_xprt_do_enqueue with:
230 smp_mb__before_atomic();
231 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
232 svc_xprt_do_enqueue(xprt
);
236 void svc_add_new_perm_xprt(struct svc_serv
*serv
, struct svc_xprt
*new)
238 clear_bit(XPT_TEMP
, &new->xpt_flags
);
239 spin_lock_bh(&serv
->sv_lock
);
240 list_add(&new->xpt_list
, &serv
->sv_permsocks
);
241 spin_unlock_bh(&serv
->sv_lock
);
242 svc_xprt_received(new);
245 int svc_create_xprt(struct svc_serv
*serv
, const char *xprt_name
,
246 struct net
*net
, const int family
,
247 const unsigned short port
, int flags
)
249 struct svc_xprt_class
*xcl
;
251 dprintk("svc: creating transport %s[%d]\n", xprt_name
, port
);
252 spin_lock(&svc_xprt_class_lock
);
253 list_for_each_entry(xcl
, &svc_xprt_class_list
, xcl_list
) {
254 struct svc_xprt
*newxprt
;
255 unsigned short newport
;
257 if (strcmp(xprt_name
, xcl
->xcl_name
))
260 if (!try_module_get(xcl
->xcl_owner
))
263 spin_unlock(&svc_xprt_class_lock
);
264 newxprt
= __svc_xpo_create(xcl
, serv
, net
, family
, port
, flags
);
265 if (IS_ERR(newxprt
)) {
266 module_put(xcl
->xcl_owner
);
267 return PTR_ERR(newxprt
);
269 svc_add_new_perm_xprt(serv
, newxprt
);
270 newport
= svc_xprt_local_port(newxprt
);
274 spin_unlock(&svc_xprt_class_lock
);
275 dprintk("svc: transport %s not found\n", xprt_name
);
277 /* This errno is exposed to user space. Provide a reasonable
278 * perror msg for a bad transport. */
279 return -EPROTONOSUPPORT
;
281 EXPORT_SYMBOL_GPL(svc_create_xprt
);
284 * Copy the local and remote xprt addresses to the rqstp structure
286 void svc_xprt_copy_addrs(struct svc_rqst
*rqstp
, struct svc_xprt
*xprt
)
288 memcpy(&rqstp
->rq_addr
, &xprt
->xpt_remote
, xprt
->xpt_remotelen
);
289 rqstp
->rq_addrlen
= xprt
->xpt_remotelen
;
292 * Destination address in request is needed for binding the
293 * source address in RPC replies/callbacks later.
295 memcpy(&rqstp
->rq_daddr
, &xprt
->xpt_local
, xprt
->xpt_locallen
);
296 rqstp
->rq_daddrlen
= xprt
->xpt_locallen
;
298 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs
);
301 * svc_print_addr - Format rq_addr field for printing
302 * @rqstp: svc_rqst struct containing address to print
303 * @buf: target buffer for formatted address
304 * @len: length of target buffer
307 char *svc_print_addr(struct svc_rqst
*rqstp
, char *buf
, size_t len
)
309 return __svc_print_addr(svc_addr(rqstp
), buf
, len
);
311 EXPORT_SYMBOL_GPL(svc_print_addr
);
314 * Queue up an idle server thread. Must have pool->sp_lock held.
315 * Note: this is really a stack rather than a queue, so that we only
316 * use as many different threads as we need, and the rest don't pollute
319 static void svc_thread_enqueue(struct svc_pool
*pool
, struct svc_rqst
*rqstp
)
321 list_add(&rqstp
->rq_list
, &pool
->sp_threads
);
325 * Dequeue an nfsd thread. Must have pool->sp_lock held.
327 static void svc_thread_dequeue(struct svc_pool
*pool
, struct svc_rqst
*rqstp
)
329 list_del(&rqstp
->rq_list
);
332 static bool svc_xprt_has_something_to_do(struct svc_xprt
*xprt
)
334 if (xprt
->xpt_flags
& ((1<<XPT_CONN
)|(1<<XPT_CLOSE
)))
336 if (xprt
->xpt_flags
& ((1<<XPT_DATA
)|(1<<XPT_DEFERRED
)))
337 return xprt
->xpt_ops
->xpo_has_wspace(xprt
);
341 static void svc_xprt_do_enqueue(struct svc_xprt
*xprt
)
343 struct svc_pool
*pool
;
344 struct svc_rqst
*rqstp
;
347 if (!svc_xprt_has_something_to_do(xprt
))
350 /* Mark transport as busy. It will remain in this state until
351 * the provider calls svc_xprt_received. We update XPT_BUSY
352 * atomically because it also guards against trying to enqueue
353 * the transport twice.
355 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
)) {
356 /* Don't enqueue transport while already enqueued */
357 dprintk("svc: transport %p busy, not enqueued\n", xprt
);
362 pool
= svc_pool_for_cpu(xprt
->xpt_server
, cpu
);
363 spin_lock_bh(&pool
->sp_lock
);
365 pool
->sp_stats
.packets
++;
367 if (!list_empty(&pool
->sp_threads
)) {
368 rqstp
= list_entry(pool
->sp_threads
.next
,
371 dprintk("svc: transport %p served by daemon %p\n",
373 svc_thread_dequeue(pool
, rqstp
);
376 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
377 rqstp
, rqstp
->rq_xprt
);
378 /* Note the order of the following 3 lines:
379 * We want to assign xprt to rqstp->rq_xprt only _after_
380 * we've woken up the process, so that we don't race with
381 * the lockless check in svc_get_next_xprt().
384 wake_up_process(rqstp
->rq_task
);
385 rqstp
->rq_xprt
= xprt
;
386 pool
->sp_stats
.threads_woken
++;
388 dprintk("svc: transport %p put into queue\n", xprt
);
389 list_add_tail(&xprt
->xpt_ready
, &pool
->sp_sockets
);
390 pool
->sp_stats
.sockets_queued
++;
393 spin_unlock_bh(&pool
->sp_lock
);
398 * Queue up a transport with data pending. If there are idle nfsd
399 * processes, wake 'em up.
402 void svc_xprt_enqueue(struct svc_xprt
*xprt
)
404 if (test_bit(XPT_BUSY
, &xprt
->xpt_flags
))
406 svc_xprt_do_enqueue(xprt
);
408 EXPORT_SYMBOL_GPL(svc_xprt_enqueue
);
411 * Dequeue the first transport. Must be called with the pool->sp_lock held.
413 static struct svc_xprt
*svc_xprt_dequeue(struct svc_pool
*pool
)
415 struct svc_xprt
*xprt
;
417 if (list_empty(&pool
->sp_sockets
))
420 xprt
= list_entry(pool
->sp_sockets
.next
,
421 struct svc_xprt
, xpt_ready
);
422 list_del_init(&xprt
->xpt_ready
);
424 dprintk("svc: transport %p dequeued, inuse=%d\n",
425 xprt
, atomic_read(&xprt
->xpt_ref
.refcount
));
431 * svc_reserve - change the space reserved for the reply to a request.
432 * @rqstp: The request in question
433 * @space: new max space to reserve
435 * Each request reserves some space on the output queue of the transport
436 * to make sure the reply fits. This function reduces that reserved
437 * space to be the amount of space used already, plus @space.
440 void svc_reserve(struct svc_rqst
*rqstp
, int space
)
442 space
+= rqstp
->rq_res
.head
[0].iov_len
;
444 if (space
< rqstp
->rq_reserved
) {
445 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
446 atomic_sub((rqstp
->rq_reserved
- space
), &xprt
->xpt_reserved
);
447 rqstp
->rq_reserved
= space
;
449 if (xprt
->xpt_ops
->xpo_adjust_wspace
)
450 xprt
->xpt_ops
->xpo_adjust_wspace(xprt
);
451 svc_xprt_enqueue(xprt
);
454 EXPORT_SYMBOL_GPL(svc_reserve
);
456 static void svc_xprt_release(struct svc_rqst
*rqstp
)
458 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
460 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
462 kfree(rqstp
->rq_deferred
);
463 rqstp
->rq_deferred
= NULL
;
465 svc_free_res_pages(rqstp
);
466 rqstp
->rq_res
.page_len
= 0;
467 rqstp
->rq_res
.page_base
= 0;
469 /* Reset response buffer and release
471 * But first, check that enough space was reserved
472 * for the reply, otherwise we have a bug!
474 if ((rqstp
->rq_res
.len
) > rqstp
->rq_reserved
)
475 printk(KERN_ERR
"RPC request reserved %d but used %d\n",
479 rqstp
->rq_res
.head
[0].iov_len
= 0;
480 svc_reserve(rqstp
, 0);
481 rqstp
->rq_xprt
= NULL
;
487 * External function to wake up a server waiting for data
488 * This really only makes sense for services like lockd
489 * which have exactly one thread anyway.
491 void svc_wake_up(struct svc_serv
*serv
)
493 struct svc_rqst
*rqstp
;
495 struct svc_pool
*pool
;
497 for (i
= 0; i
< serv
->sv_nrpools
; i
++) {
498 pool
= &serv
->sv_pools
[i
];
500 spin_lock_bh(&pool
->sp_lock
);
501 if (!list_empty(&pool
->sp_threads
)) {
502 rqstp
= list_entry(pool
->sp_threads
.next
,
505 dprintk("svc: daemon %p woken up.\n", rqstp
);
507 svc_thread_dequeue(pool, rqstp);
508 rqstp->rq_xprt = NULL;
510 wake_up_process(rqstp
->rq_task
);
512 pool
->sp_task_pending
= 1;
513 spin_unlock_bh(&pool
->sp_lock
);
516 EXPORT_SYMBOL_GPL(svc_wake_up
);
518 int svc_port_is_privileged(struct sockaddr
*sin
)
520 switch (sin
->sa_family
) {
522 return ntohs(((struct sockaddr_in
*)sin
)->sin_port
)
525 return ntohs(((struct sockaddr_in6
*)sin
)->sin6_port
)
533 * Make sure that we don't have too many active connections. If we have,
534 * something must be dropped. It's not clear what will happen if we allow
535 * "too many" connections, but when dealing with network-facing software,
536 * we have to code defensively. Here we do that by imposing hard limits.
538 * There's no point in trying to do random drop here for DoS
539 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
540 * attacker can easily beat that.
542 * The only somewhat efficient mechanism would be if drop old
543 * connections from the same IP first. But right now we don't even
544 * record the client IP in svc_sock.
546 * single-threaded services that expect a lot of clients will probably
547 * need to set sv_maxconn to override the default value which is based
548 * on the number of threads
550 static void svc_check_conn_limits(struct svc_serv
*serv
)
552 unsigned int limit
= serv
->sv_maxconn
? serv
->sv_maxconn
:
553 (serv
->sv_nrthreads
+3) * 20;
555 if (serv
->sv_tmpcnt
> limit
) {
556 struct svc_xprt
*xprt
= NULL
;
557 spin_lock_bh(&serv
->sv_lock
);
558 if (!list_empty(&serv
->sv_tempsocks
)) {
559 /* Try to help the admin */
560 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
561 serv
->sv_name
, serv
->sv_maxconn
?
562 "max number of connections" :
563 "number of threads");
565 * Always select the oldest connection. It's not fair,
568 xprt
= list_entry(serv
->sv_tempsocks
.prev
,
571 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
574 spin_unlock_bh(&serv
->sv_lock
);
577 svc_xprt_enqueue(xprt
);
583 static int svc_alloc_arg(struct svc_rqst
*rqstp
)
585 struct svc_serv
*serv
= rqstp
->rq_server
;
590 /* now allocate needed pages. If we get a failure, sleep briefly */
591 pages
= (serv
->sv_max_mesg
+ PAGE_SIZE
) / PAGE_SIZE
;
592 WARN_ON_ONCE(pages
>= RPCSVC_MAXPAGES
);
593 if (pages
>= RPCSVC_MAXPAGES
)
594 /* use as many pages as possible */
595 pages
= RPCSVC_MAXPAGES
- 1;
596 for (i
= 0; i
< pages
; i
++)
597 while (rqstp
->rq_pages
[i
] == NULL
) {
598 struct page
*p
= alloc_page(GFP_KERNEL
);
600 set_current_state(TASK_INTERRUPTIBLE
);
601 if (signalled() || kthread_should_stop()) {
602 set_current_state(TASK_RUNNING
);
605 schedule_timeout(msecs_to_jiffies(500));
607 rqstp
->rq_pages
[i
] = p
;
609 rqstp
->rq_page_end
= &rqstp
->rq_pages
[i
];
610 rqstp
->rq_pages
[i
++] = NULL
; /* this might be seen in nfs_read_actor */
612 /* Make arg->head point to first page and arg->pages point to rest */
613 arg
= &rqstp
->rq_arg
;
614 arg
->head
[0].iov_base
= page_address(rqstp
->rq_pages
[0]);
615 arg
->head
[0].iov_len
= PAGE_SIZE
;
616 arg
->pages
= rqstp
->rq_pages
+ 1;
618 /* save at least one page for response */
619 arg
->page_len
= (pages
-2)*PAGE_SIZE
;
620 arg
->len
= (pages
-1)*PAGE_SIZE
;
621 arg
->tail
[0].iov_len
= 0;
625 static struct svc_xprt
*svc_get_next_xprt(struct svc_rqst
*rqstp
, long timeout
)
627 struct svc_xprt
*xprt
;
628 struct svc_pool
*pool
= rqstp
->rq_pool
;
631 /* Normally we will wait up to 5 seconds for any required
632 * cache information to be provided.
634 rqstp
->rq_chandle
.thread_wait
= 5*HZ
;
636 spin_lock_bh(&pool
->sp_lock
);
637 xprt
= svc_xprt_dequeue(pool
);
639 rqstp
->rq_xprt
= xprt
;
642 /* As there is a shortage of threads and this request
643 * had to be queued, don't allow the thread to wait so
644 * long for cache updates.
646 rqstp
->rq_chandle
.thread_wait
= 1*HZ
;
647 pool
->sp_task_pending
= 0;
649 if (pool
->sp_task_pending
) {
650 pool
->sp_task_pending
= 0;
651 xprt
= ERR_PTR(-EAGAIN
);
655 * We have to be able to interrupt this wait
656 * to bring down the daemons ...
658 set_current_state(TASK_INTERRUPTIBLE
);
660 /* No data pending. Go to sleep */
661 svc_thread_enqueue(pool
, rqstp
);
662 spin_unlock_bh(&pool
->sp_lock
);
664 if (!(signalled() || kthread_should_stop())) {
665 time_left
= schedule_timeout(timeout
);
666 __set_current_state(TASK_RUNNING
);
670 xprt
= rqstp
->rq_xprt
;
674 __set_current_state(TASK_RUNNING
);
676 spin_lock_bh(&pool
->sp_lock
);
678 pool
->sp_stats
.threads_timedout
++;
680 xprt
= rqstp
->rq_xprt
;
682 svc_thread_dequeue(pool
, rqstp
);
683 spin_unlock_bh(&pool
->sp_lock
);
684 dprintk("svc: server %p, no data yet\n", rqstp
);
685 if (signalled() || kthread_should_stop())
686 return ERR_PTR(-EINTR
);
688 return ERR_PTR(-EAGAIN
);
692 spin_unlock_bh(&pool
->sp_lock
);
696 static void svc_add_new_temp_xprt(struct svc_serv
*serv
, struct svc_xprt
*newxpt
)
698 spin_lock_bh(&serv
->sv_lock
);
699 set_bit(XPT_TEMP
, &newxpt
->xpt_flags
);
700 list_add(&newxpt
->xpt_list
, &serv
->sv_tempsocks
);
702 if (serv
->sv_temptimer
.function
== NULL
) {
703 /* setup timer to age temp transports */
704 setup_timer(&serv
->sv_temptimer
, svc_age_temp_xprts
,
705 (unsigned long)serv
);
706 mod_timer(&serv
->sv_temptimer
,
707 jiffies
+ svc_conn_age_period
* HZ
);
709 spin_unlock_bh(&serv
->sv_lock
);
710 svc_xprt_received(newxpt
);
713 static int svc_handle_xprt(struct svc_rqst
*rqstp
, struct svc_xprt
*xprt
)
715 struct svc_serv
*serv
= rqstp
->rq_server
;
718 if (test_bit(XPT_CLOSE
, &xprt
->xpt_flags
)) {
719 dprintk("svc_recv: found XPT_CLOSE\n");
720 svc_delete_xprt(xprt
);
721 /* Leave XPT_BUSY set on the dead xprt: */
724 if (test_bit(XPT_LISTENER
, &xprt
->xpt_flags
)) {
725 struct svc_xprt
*newxpt
;
727 * We know this module_get will succeed because the
728 * listener holds a reference too
730 __module_get(xprt
->xpt_class
->xcl_owner
);
731 svc_check_conn_limits(xprt
->xpt_server
);
732 newxpt
= xprt
->xpt_ops
->xpo_accept(xprt
);
734 svc_add_new_temp_xprt(serv
, newxpt
);
736 module_put(xprt
->xpt_class
->xcl_owner
);
738 /* XPT_DATA|XPT_DEFERRED case: */
739 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
740 rqstp
, rqstp
->rq_pool
->sp_id
, xprt
,
741 atomic_read(&xprt
->xpt_ref
.refcount
));
742 rqstp
->rq_deferred
= svc_deferred_dequeue(xprt
);
743 if (rqstp
->rq_deferred
)
744 len
= svc_deferred_recv(rqstp
);
746 len
= xprt
->xpt_ops
->xpo_recvfrom(rqstp
);
747 dprintk("svc: got len=%d\n", len
);
748 rqstp
->rq_reserved
= serv
->sv_max_mesg
;
749 atomic_add(rqstp
->rq_reserved
, &xprt
->xpt_reserved
);
751 /* clear XPT_BUSY: */
752 svc_xprt_received(xprt
);
757 * Receive the next request on any transport. This code is carefully
758 * organised not to touch any cachelines in the shared svc_serv
759 * structure, only cachelines in the local svc_pool.
761 int svc_recv(struct svc_rqst
*rqstp
, long timeout
)
763 struct svc_xprt
*xprt
= NULL
;
764 struct svc_serv
*serv
= rqstp
->rq_server
;
767 dprintk("svc: server %p waiting for data (to = %ld)\n",
772 "svc_recv: service %p, transport not NULL!\n",
775 err
= svc_alloc_arg(rqstp
);
782 if (signalled() || kthread_should_stop())
785 xprt
= svc_get_next_xprt(rqstp
, timeout
);
791 len
= svc_handle_xprt(rqstp
, xprt
);
793 /* No data, incomplete (TCP) read, or accept() */
798 clear_bit(XPT_OLD
, &xprt
->xpt_flags
);
800 rqstp
->rq_secure
= xprt
->xpt_ops
->xpo_secure_port(rqstp
);
801 rqstp
->rq_chandle
.defer
= svc_defer
;
802 rqstp
->rq_xid
= svc_getu32(&rqstp
->rq_arg
.head
[0]);
805 serv
->sv_stats
->netcnt
++;
806 trace_svc_recv(rqstp
, len
);
809 rqstp
->rq_res
.len
= 0;
810 svc_xprt_release(rqstp
);
812 trace_svc_recv(rqstp
, err
);
815 EXPORT_SYMBOL_GPL(svc_recv
);
820 void svc_drop(struct svc_rqst
*rqstp
)
822 dprintk("svc: xprt %p dropped request\n", rqstp
->rq_xprt
);
823 svc_xprt_release(rqstp
);
825 EXPORT_SYMBOL_GPL(svc_drop
);
828 * Return reply to client.
830 int svc_send(struct svc_rqst
*rqstp
)
832 struct svc_xprt
*xprt
;
836 xprt
= rqstp
->rq_xprt
;
840 /* release the receive skb before sending the reply */
841 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
843 /* calculate over-all length */
845 xb
->len
= xb
->head
[0].iov_len
+
849 /* Grab mutex to serialize outgoing data. */
850 mutex_lock(&xprt
->xpt_mutex
);
851 if (test_bit(XPT_DEAD
, &xprt
->xpt_flags
)
852 || test_bit(XPT_CLOSE
, &xprt
->xpt_flags
))
855 len
= xprt
->xpt_ops
->xpo_sendto(rqstp
);
856 mutex_unlock(&xprt
->xpt_mutex
);
857 rpc_wake_up(&xprt
->xpt_bc_pending
);
858 svc_xprt_release(rqstp
);
860 if (len
== -ECONNREFUSED
|| len
== -ENOTCONN
|| len
== -EAGAIN
)
863 trace_svc_send(rqstp
, len
);
868 * Timer function to close old temporary transports, using
869 * a mark-and-sweep algorithm.
871 static void svc_age_temp_xprts(unsigned long closure
)
873 struct svc_serv
*serv
= (struct svc_serv
*)closure
;
874 struct svc_xprt
*xprt
;
875 struct list_head
*le
, *next
;
877 dprintk("svc_age_temp_xprts\n");
879 if (!spin_trylock_bh(&serv
->sv_lock
)) {
880 /* busy, try again 1 sec later */
881 dprintk("svc_age_temp_xprts: busy\n");
882 mod_timer(&serv
->sv_temptimer
, jiffies
+ HZ
);
886 list_for_each_safe(le
, next
, &serv
->sv_tempsocks
) {
887 xprt
= list_entry(le
, struct svc_xprt
, xpt_list
);
889 /* First time through, just mark it OLD. Second time
890 * through, close it. */
891 if (!test_and_set_bit(XPT_OLD
, &xprt
->xpt_flags
))
893 if (atomic_read(&xprt
->xpt_ref
.refcount
) > 1 ||
894 test_bit(XPT_BUSY
, &xprt
->xpt_flags
))
897 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
898 set_bit(XPT_DETACHED
, &xprt
->xpt_flags
);
899 dprintk("queuing xprt %p for closing\n", xprt
);
901 /* a thread will dequeue and close it soon */
902 svc_xprt_enqueue(xprt
);
904 spin_unlock_bh(&serv
->sv_lock
);
906 mod_timer(&serv
->sv_temptimer
, jiffies
+ svc_conn_age_period
* HZ
);
909 static void call_xpt_users(struct svc_xprt
*xprt
)
911 struct svc_xpt_user
*u
;
913 spin_lock(&xprt
->xpt_lock
);
914 while (!list_empty(&xprt
->xpt_users
)) {
915 u
= list_first_entry(&xprt
->xpt_users
, struct svc_xpt_user
, list
);
919 spin_unlock(&xprt
->xpt_lock
);
923 * Remove a dead transport
925 static void svc_delete_xprt(struct svc_xprt
*xprt
)
927 struct svc_serv
*serv
= xprt
->xpt_server
;
928 struct svc_deferred_req
*dr
;
930 /* Only do this once */
931 if (test_and_set_bit(XPT_DEAD
, &xprt
->xpt_flags
))
934 dprintk("svc: svc_delete_xprt(%p)\n", xprt
);
935 xprt
->xpt_ops
->xpo_detach(xprt
);
937 spin_lock_bh(&serv
->sv_lock
);
938 if (!test_and_set_bit(XPT_DETACHED
, &xprt
->xpt_flags
))
939 list_del_init(&xprt
->xpt_list
);
940 WARN_ON_ONCE(!list_empty(&xprt
->xpt_ready
));
941 if (test_bit(XPT_TEMP
, &xprt
->xpt_flags
))
943 spin_unlock_bh(&serv
->sv_lock
);
945 while ((dr
= svc_deferred_dequeue(xprt
)) != NULL
)
948 call_xpt_users(xprt
);
952 void svc_close_xprt(struct svc_xprt
*xprt
)
954 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
955 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
))
956 /* someone else will have to effect the close */
959 * We expect svc_close_xprt() to work even when no threads are
960 * running (e.g., while configuring the server before starting
961 * any threads), so if the transport isn't busy, we delete
964 svc_delete_xprt(xprt
);
966 EXPORT_SYMBOL_GPL(svc_close_xprt
);
968 static int svc_close_list(struct svc_serv
*serv
, struct list_head
*xprt_list
, struct net
*net
)
970 struct svc_xprt
*xprt
;
973 spin_lock(&serv
->sv_lock
);
974 list_for_each_entry(xprt
, xprt_list
, xpt_list
) {
975 if (xprt
->xpt_net
!= net
)
978 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
979 svc_xprt_enqueue(xprt
);
981 spin_unlock(&serv
->sv_lock
);
985 static struct svc_xprt
*svc_dequeue_net(struct svc_serv
*serv
, struct net
*net
)
987 struct svc_pool
*pool
;
988 struct svc_xprt
*xprt
;
989 struct svc_xprt
*tmp
;
992 for (i
= 0; i
< serv
->sv_nrpools
; i
++) {
993 pool
= &serv
->sv_pools
[i
];
995 spin_lock_bh(&pool
->sp_lock
);
996 list_for_each_entry_safe(xprt
, tmp
, &pool
->sp_sockets
, xpt_ready
) {
997 if (xprt
->xpt_net
!= net
)
999 list_del_init(&xprt
->xpt_ready
);
1000 spin_unlock_bh(&pool
->sp_lock
);
1003 spin_unlock_bh(&pool
->sp_lock
);
1008 static void svc_clean_up_xprts(struct svc_serv
*serv
, struct net
*net
)
1010 struct svc_xprt
*xprt
;
1012 while ((xprt
= svc_dequeue_net(serv
, net
))) {
1013 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
1014 svc_delete_xprt(xprt
);
1019 * Server threads may still be running (especially in the case where the
1020 * service is still running in other network namespaces).
1022 * So we shut down sockets the same way we would on a running server, by
1023 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1024 * the close. In the case there are no such other threads,
1025 * threads running, svc_clean_up_xprts() does a simple version of a
1026 * server's main event loop, and in the case where there are other
1027 * threads, we may need to wait a little while and then check again to
1028 * see if they're done.
1030 void svc_close_net(struct svc_serv
*serv
, struct net
*net
)
1034 while (svc_close_list(serv
, &serv
->sv_permsocks
, net
) +
1035 svc_close_list(serv
, &serv
->sv_tempsocks
, net
)) {
1037 svc_clean_up_xprts(serv
, net
);
1043 * Handle defer and revisit of requests
1046 static void svc_revisit(struct cache_deferred_req
*dreq
, int too_many
)
1048 struct svc_deferred_req
*dr
=
1049 container_of(dreq
, struct svc_deferred_req
, handle
);
1050 struct svc_xprt
*xprt
= dr
->xprt
;
1052 spin_lock(&xprt
->xpt_lock
);
1053 set_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1054 if (too_many
|| test_bit(XPT_DEAD
, &xprt
->xpt_flags
)) {
1055 spin_unlock(&xprt
->xpt_lock
);
1056 dprintk("revisit canceled\n");
1061 dprintk("revisit queued\n");
1063 list_add(&dr
->handle
.recent
, &xprt
->xpt_deferred
);
1064 spin_unlock(&xprt
->xpt_lock
);
1065 svc_xprt_enqueue(xprt
);
1070 * Save the request off for later processing. The request buffer looks
1073 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1075 * This code can only handle requests that consist of an xprt-header
1078 static struct cache_deferred_req
*svc_defer(struct cache_req
*req
)
1080 struct svc_rqst
*rqstp
= container_of(req
, struct svc_rqst
, rq_chandle
);
1081 struct svc_deferred_req
*dr
;
1083 if (rqstp
->rq_arg
.page_len
|| !rqstp
->rq_usedeferral
)
1084 return NULL
; /* if more than a page, give up FIXME */
1085 if (rqstp
->rq_deferred
) {
1086 dr
= rqstp
->rq_deferred
;
1087 rqstp
->rq_deferred
= NULL
;
1091 /* FIXME maybe discard if size too large */
1092 size
= sizeof(struct svc_deferred_req
) + rqstp
->rq_arg
.len
;
1093 dr
= kmalloc(size
, GFP_KERNEL
);
1097 dr
->handle
.owner
= rqstp
->rq_server
;
1098 dr
->prot
= rqstp
->rq_prot
;
1099 memcpy(&dr
->addr
, &rqstp
->rq_addr
, rqstp
->rq_addrlen
);
1100 dr
->addrlen
= rqstp
->rq_addrlen
;
1101 dr
->daddr
= rqstp
->rq_daddr
;
1102 dr
->argslen
= rqstp
->rq_arg
.len
>> 2;
1103 dr
->xprt_hlen
= rqstp
->rq_xprt_hlen
;
1105 /* back up head to the start of the buffer and copy */
1106 skip
= rqstp
->rq_arg
.len
- rqstp
->rq_arg
.head
[0].iov_len
;
1107 memcpy(dr
->args
, rqstp
->rq_arg
.head
[0].iov_base
- skip
,
1110 svc_xprt_get(rqstp
->rq_xprt
);
1111 dr
->xprt
= rqstp
->rq_xprt
;
1112 rqstp
->rq_dropme
= true;
1114 dr
->handle
.revisit
= svc_revisit
;
1119 * recv data from a deferred request into an active one
1121 static int svc_deferred_recv(struct svc_rqst
*rqstp
)
1123 struct svc_deferred_req
*dr
= rqstp
->rq_deferred
;
1125 /* setup iov_base past transport header */
1126 rqstp
->rq_arg
.head
[0].iov_base
= dr
->args
+ (dr
->xprt_hlen
>>2);
1127 /* The iov_len does not include the transport header bytes */
1128 rqstp
->rq_arg
.head
[0].iov_len
= (dr
->argslen
<<2) - dr
->xprt_hlen
;
1129 rqstp
->rq_arg
.page_len
= 0;
1130 /* The rq_arg.len includes the transport header bytes */
1131 rqstp
->rq_arg
.len
= dr
->argslen
<<2;
1132 rqstp
->rq_prot
= dr
->prot
;
1133 memcpy(&rqstp
->rq_addr
, &dr
->addr
, dr
->addrlen
);
1134 rqstp
->rq_addrlen
= dr
->addrlen
;
1135 /* Save off transport header len in case we get deferred again */
1136 rqstp
->rq_xprt_hlen
= dr
->xprt_hlen
;
1137 rqstp
->rq_daddr
= dr
->daddr
;
1138 rqstp
->rq_respages
= rqstp
->rq_pages
;
1139 return (dr
->argslen
<<2) - dr
->xprt_hlen
;
1143 static struct svc_deferred_req
*svc_deferred_dequeue(struct svc_xprt
*xprt
)
1145 struct svc_deferred_req
*dr
= NULL
;
1147 if (!test_bit(XPT_DEFERRED
, &xprt
->xpt_flags
))
1149 spin_lock(&xprt
->xpt_lock
);
1150 if (!list_empty(&xprt
->xpt_deferred
)) {
1151 dr
= list_entry(xprt
->xpt_deferred
.next
,
1152 struct svc_deferred_req
,
1154 list_del_init(&dr
->handle
.recent
);
1156 clear_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1157 spin_unlock(&xprt
->xpt_lock
);
1162 * svc_find_xprt - find an RPC transport instance
1163 * @serv: pointer to svc_serv to search
1164 * @xcl_name: C string containing transport's class name
1165 * @net: owner net pointer
1166 * @af: Address family of transport's local address
1167 * @port: transport's IP port number
1169 * Return the transport instance pointer for the endpoint accepting
1170 * connections/peer traffic from the specified transport class,
1171 * address family and port.
1173 * Specifying 0 for the address family or port is effectively a
1174 * wild-card, and will result in matching the first transport in the
1175 * service's list that has a matching class name.
1177 struct svc_xprt
*svc_find_xprt(struct svc_serv
*serv
, const char *xcl_name
,
1178 struct net
*net
, const sa_family_t af
,
1179 const unsigned short port
)
1181 struct svc_xprt
*xprt
;
1182 struct svc_xprt
*found
= NULL
;
1184 /* Sanity check the args */
1185 if (serv
== NULL
|| xcl_name
== NULL
)
1188 spin_lock_bh(&serv
->sv_lock
);
1189 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1190 if (xprt
->xpt_net
!= net
)
1192 if (strcmp(xprt
->xpt_class
->xcl_name
, xcl_name
))
1194 if (af
!= AF_UNSPEC
&& af
!= xprt
->xpt_local
.ss_family
)
1196 if (port
!= 0 && port
!= svc_xprt_local_port(xprt
))
1202 spin_unlock_bh(&serv
->sv_lock
);
1205 EXPORT_SYMBOL_GPL(svc_find_xprt
);
1207 static int svc_one_xprt_name(const struct svc_xprt
*xprt
,
1208 char *pos
, int remaining
)
1212 len
= snprintf(pos
, remaining
, "%s %u\n",
1213 xprt
->xpt_class
->xcl_name
,
1214 svc_xprt_local_port(xprt
));
1215 if (len
>= remaining
)
1216 return -ENAMETOOLONG
;
1221 * svc_xprt_names - format a buffer with a list of transport names
1222 * @serv: pointer to an RPC service
1223 * @buf: pointer to a buffer to be filled in
1224 * @buflen: length of buffer to be filled in
1226 * Fills in @buf with a string containing a list of transport names,
1227 * each name terminated with '\n'.
1229 * Returns positive length of the filled-in string on success; otherwise
1230 * a negative errno value is returned if an error occurs.
1232 int svc_xprt_names(struct svc_serv
*serv
, char *buf
, const int buflen
)
1234 struct svc_xprt
*xprt
;
1238 /* Sanity check args */
1242 spin_lock_bh(&serv
->sv_lock
);
1246 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1247 len
= svc_one_xprt_name(xprt
, pos
, buflen
- totlen
);
1259 spin_unlock_bh(&serv
->sv_lock
);
1262 EXPORT_SYMBOL_GPL(svc_xprt_names
);
1265 /*----------------------------------------------------------------------------*/
1267 static void *svc_pool_stats_start(struct seq_file
*m
, loff_t
*pos
)
1269 unsigned int pidx
= (unsigned int)*pos
;
1270 struct svc_serv
*serv
= m
->private;
1272 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx
);
1275 return SEQ_START_TOKEN
;
1276 return (pidx
> serv
->sv_nrpools
? NULL
: &serv
->sv_pools
[pidx
-1]);
1279 static void *svc_pool_stats_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
1281 struct svc_pool
*pool
= p
;
1282 struct svc_serv
*serv
= m
->private;
1284 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos
);
1286 if (p
== SEQ_START_TOKEN
) {
1287 pool
= &serv
->sv_pools
[0];
1289 unsigned int pidx
= (pool
- &serv
->sv_pools
[0]);
1290 if (pidx
< serv
->sv_nrpools
-1)
1291 pool
= &serv
->sv_pools
[pidx
+1];
1299 static void svc_pool_stats_stop(struct seq_file
*m
, void *p
)
1303 static int svc_pool_stats_show(struct seq_file
*m
, void *p
)
1305 struct svc_pool
*pool
= p
;
1307 if (p
== SEQ_START_TOKEN
) {
1308 seq_puts(m
, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1312 seq_printf(m
, "%u %lu %lu %lu %lu\n",
1314 pool
->sp_stats
.packets
,
1315 pool
->sp_stats
.sockets_queued
,
1316 pool
->sp_stats
.threads_woken
,
1317 pool
->sp_stats
.threads_timedout
);
1322 static const struct seq_operations svc_pool_stats_seq_ops
= {
1323 .start
= svc_pool_stats_start
,
1324 .next
= svc_pool_stats_next
,
1325 .stop
= svc_pool_stats_stop
,
1326 .show
= svc_pool_stats_show
,
1329 int svc_pool_stats_open(struct svc_serv
*serv
, struct file
*file
)
1333 err
= seq_open(file
, &svc_pool_stats_seq_ops
);
1335 ((struct seq_file
*) file
->private_data
)->private = serv
;
1338 EXPORT_SYMBOL(svc_pool_stats_open
);
1340 /*----------------------------------------------------------------------------*/