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>
19 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
21 static struct svc_deferred_req
*svc_deferred_dequeue(struct svc_xprt
*xprt
);
22 static int svc_deferred_recv(struct svc_rqst
*rqstp
);
23 static struct cache_deferred_req
*svc_defer(struct cache_req
*req
);
24 static void svc_age_temp_xprts(unsigned long closure
);
25 static void svc_delete_xprt(struct svc_xprt
*xprt
);
26 static void svc_xprt_do_enqueue(struct svc_xprt
*xprt
);
28 /* apparently the "standard" is that clients close
29 * idle connections after 5 minutes, servers after
31 * http://www.connectathon.org/talks96/nfstcp.pdf
33 static int svc_conn_age_period
= 6*60;
35 /* List of registered transport classes */
36 static DEFINE_SPINLOCK(svc_xprt_class_lock
);
37 static LIST_HEAD(svc_xprt_class_list
);
39 /* SMP locking strategy:
41 * svc_pool->sp_lock protects most of the fields of that pool.
42 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
43 * when both need to be taken (rare), svc_serv->sv_lock is first.
44 * BKL protects svc_serv->sv_nrthread.
45 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
46 * and the ->sk_info_authunix cache.
48 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
49 * enqueued multiply. During normal transport processing this bit
50 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
51 * Providers should not manipulate this bit directly.
53 * Some flags can be set to certain values at any time
54 * providing that certain rules are followed:
57 * - Can be set or cleared at any time.
58 * - After a set, svc_xprt_enqueue must be called to enqueue
59 * the transport for processing.
60 * - After a clear, the transport must be read/accepted.
61 * If this succeeds, it must be set again.
63 * - Can set at any time. It is never cleared.
65 * - Can only be set while XPT_BUSY is held which ensures
66 * that no other thread will be using the transport or will
67 * try to set XPT_DEAD.
70 int svc_reg_xprt_class(struct svc_xprt_class
*xcl
)
72 struct svc_xprt_class
*cl
;
75 dprintk("svc: Adding svc transport class '%s'\n", xcl
->xcl_name
);
77 INIT_LIST_HEAD(&xcl
->xcl_list
);
78 spin_lock(&svc_xprt_class_lock
);
79 /* Make sure there isn't already a class with the same name */
80 list_for_each_entry(cl
, &svc_xprt_class_list
, xcl_list
) {
81 if (strcmp(xcl
->xcl_name
, cl
->xcl_name
) == 0)
84 list_add_tail(&xcl
->xcl_list
, &svc_xprt_class_list
);
87 spin_unlock(&svc_xprt_class_lock
);
90 EXPORT_SYMBOL_GPL(svc_reg_xprt_class
);
92 void svc_unreg_xprt_class(struct svc_xprt_class
*xcl
)
94 dprintk("svc: Removing svc transport class '%s'\n", xcl
->xcl_name
);
95 spin_lock(&svc_xprt_class_lock
);
96 list_del_init(&xcl
->xcl_list
);
97 spin_unlock(&svc_xprt_class_lock
);
99 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class
);
102 * Format the transport list for printing
104 int svc_print_xprts(char *buf
, int maxlen
)
106 struct svc_xprt_class
*xcl
;
111 spin_lock(&svc_xprt_class_lock
);
112 list_for_each_entry(xcl
, &svc_xprt_class_list
, xcl_list
) {
115 sprintf(tmpstr
, "%s %d\n", xcl
->xcl_name
, xcl
->xcl_max_payload
);
116 slen
= strlen(tmpstr
);
117 if (len
+ slen
> maxlen
)
122 spin_unlock(&svc_xprt_class_lock
);
127 static void svc_xprt_free(struct kref
*kref
)
129 struct svc_xprt
*xprt
=
130 container_of(kref
, struct svc_xprt
, xpt_ref
);
131 struct module
*owner
= xprt
->xpt_class
->xcl_owner
;
132 if (test_bit(XPT_CACHE_AUTH
, &xprt
->xpt_flags
))
133 svcauth_unix_info_release(xprt
);
134 put_net(xprt
->xpt_net
);
135 /* See comment on corresponding get in xs_setup_bc_tcp(): */
136 if (xprt
->xpt_bc_xprt
)
137 xprt_put(xprt
->xpt_bc_xprt
);
138 xprt
->xpt_ops
->xpo_free(xprt
);
142 void svc_xprt_put(struct svc_xprt
*xprt
)
144 kref_put(&xprt
->xpt_ref
, svc_xprt_free
);
146 EXPORT_SYMBOL_GPL(svc_xprt_put
);
149 * Called by transport drivers to initialize the transport independent
150 * portion of the transport instance.
152 void svc_xprt_init(struct net
*net
, struct svc_xprt_class
*xcl
,
153 struct svc_xprt
*xprt
, struct svc_serv
*serv
)
155 memset(xprt
, 0, sizeof(*xprt
));
156 xprt
->xpt_class
= xcl
;
157 xprt
->xpt_ops
= xcl
->xcl_ops
;
158 kref_init(&xprt
->xpt_ref
);
159 xprt
->xpt_server
= serv
;
160 INIT_LIST_HEAD(&xprt
->xpt_list
);
161 INIT_LIST_HEAD(&xprt
->xpt_ready
);
162 INIT_LIST_HEAD(&xprt
->xpt_deferred
);
163 INIT_LIST_HEAD(&xprt
->xpt_users
);
164 mutex_init(&xprt
->xpt_mutex
);
165 spin_lock_init(&xprt
->xpt_lock
);
166 set_bit(XPT_BUSY
, &xprt
->xpt_flags
);
167 rpc_init_wait_queue(&xprt
->xpt_bc_pending
, "xpt_bc_pending");
168 xprt
->xpt_net
= get_net(net
);
170 EXPORT_SYMBOL_GPL(svc_xprt_init
);
172 static struct svc_xprt
*__svc_xpo_create(struct svc_xprt_class
*xcl
,
173 struct svc_serv
*serv
,
176 const unsigned short port
,
179 struct sockaddr_in sin
= {
180 .sin_family
= AF_INET
,
181 .sin_addr
.s_addr
= htonl(INADDR_ANY
),
182 .sin_port
= htons(port
),
184 #if IS_ENABLED(CONFIG_IPV6)
185 struct sockaddr_in6 sin6
= {
186 .sin6_family
= AF_INET6
,
187 .sin6_addr
= IN6ADDR_ANY_INIT
,
188 .sin6_port
= htons(port
),
191 struct sockaddr
*sap
;
196 sap
= (struct sockaddr
*)&sin
;
199 #if IS_ENABLED(CONFIG_IPV6)
201 sap
= (struct sockaddr
*)&sin6
;
206 return ERR_PTR(-EAFNOSUPPORT
);
209 return xcl
->xcl_ops
->xpo_create(serv
, net
, sap
, len
, flags
);
213 * svc_xprt_received conditionally queues the transport for processing
214 * by another thread. The caller must hold the XPT_BUSY bit and must
215 * not thereafter touch transport data.
217 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
218 * insufficient) data.
220 static void svc_xprt_received(struct svc_xprt
*xprt
)
222 WARN_ON_ONCE(!test_bit(XPT_BUSY
, &xprt
->xpt_flags
));
223 if (!test_bit(XPT_BUSY
, &xprt
->xpt_flags
))
225 /* As soon as we clear busy, the xprt could be closed and
226 * 'put', so we need a reference to call svc_xprt_do_enqueue with:
229 smp_mb__before_atomic();
230 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
231 svc_xprt_do_enqueue(xprt
);
235 void svc_add_new_perm_xprt(struct svc_serv
*serv
, struct svc_xprt
*new)
237 clear_bit(XPT_TEMP
, &new->xpt_flags
);
238 spin_lock_bh(&serv
->sv_lock
);
239 list_add(&new->xpt_list
, &serv
->sv_permsocks
);
240 spin_unlock_bh(&serv
->sv_lock
);
241 svc_xprt_received(new);
244 int svc_create_xprt(struct svc_serv
*serv
, const char *xprt_name
,
245 struct net
*net
, const int family
,
246 const unsigned short port
, int flags
)
248 struct svc_xprt_class
*xcl
;
250 dprintk("svc: creating transport %s[%d]\n", xprt_name
, port
);
251 spin_lock(&svc_xprt_class_lock
);
252 list_for_each_entry(xcl
, &svc_xprt_class_list
, xcl_list
) {
253 struct svc_xprt
*newxprt
;
254 unsigned short newport
;
256 if (strcmp(xprt_name
, xcl
->xcl_name
))
259 if (!try_module_get(xcl
->xcl_owner
))
262 spin_unlock(&svc_xprt_class_lock
);
263 newxprt
= __svc_xpo_create(xcl
, serv
, net
, family
, port
, flags
);
264 if (IS_ERR(newxprt
)) {
265 module_put(xcl
->xcl_owner
);
266 return PTR_ERR(newxprt
);
268 svc_add_new_perm_xprt(serv
, newxprt
);
269 newport
= svc_xprt_local_port(newxprt
);
273 spin_unlock(&svc_xprt_class_lock
);
274 dprintk("svc: transport %s not found\n", xprt_name
);
276 /* This errno is exposed to user space. Provide a reasonable
277 * perror msg for a bad transport. */
278 return -EPROTONOSUPPORT
;
280 EXPORT_SYMBOL_GPL(svc_create_xprt
);
283 * Copy the local and remote xprt addresses to the rqstp structure
285 void svc_xprt_copy_addrs(struct svc_rqst
*rqstp
, struct svc_xprt
*xprt
)
287 memcpy(&rqstp
->rq_addr
, &xprt
->xpt_remote
, xprt
->xpt_remotelen
);
288 rqstp
->rq_addrlen
= xprt
->xpt_remotelen
;
291 * Destination address in request is needed for binding the
292 * source address in RPC replies/callbacks later.
294 memcpy(&rqstp
->rq_daddr
, &xprt
->xpt_local
, xprt
->xpt_locallen
);
295 rqstp
->rq_daddrlen
= xprt
->xpt_locallen
;
297 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs
);
300 * svc_print_addr - Format rq_addr field for printing
301 * @rqstp: svc_rqst struct containing address to print
302 * @buf: target buffer for formatted address
303 * @len: length of target buffer
306 char *svc_print_addr(struct svc_rqst
*rqstp
, char *buf
, size_t len
)
308 return __svc_print_addr(svc_addr(rqstp
), buf
, len
);
310 EXPORT_SYMBOL_GPL(svc_print_addr
);
313 * Queue up an idle server thread. Must have pool->sp_lock held.
314 * Note: this is really a stack rather than a queue, so that we only
315 * use as many different threads as we need, and the rest don't pollute
318 static void svc_thread_enqueue(struct svc_pool
*pool
, struct svc_rqst
*rqstp
)
320 list_add(&rqstp
->rq_list
, &pool
->sp_threads
);
324 * Dequeue an nfsd thread. Must have pool->sp_lock held.
326 static void svc_thread_dequeue(struct svc_pool
*pool
, struct svc_rqst
*rqstp
)
328 list_del(&rqstp
->rq_list
);
331 static bool svc_xprt_has_something_to_do(struct svc_xprt
*xprt
)
333 if (xprt
->xpt_flags
& ((1<<XPT_CONN
)|(1<<XPT_CLOSE
)))
335 if (xprt
->xpt_flags
& ((1<<XPT_DATA
)|(1<<XPT_DEFERRED
)))
336 return xprt
->xpt_ops
->xpo_has_wspace(xprt
);
340 static void svc_xprt_do_enqueue(struct svc_xprt
*xprt
)
342 struct svc_pool
*pool
;
343 struct svc_rqst
*rqstp
;
346 if (!svc_xprt_has_something_to_do(xprt
))
350 pool
= svc_pool_for_cpu(xprt
->xpt_server
, cpu
);
353 spin_lock_bh(&pool
->sp_lock
);
355 if (!list_empty(&pool
->sp_threads
) &&
356 !list_empty(&pool
->sp_sockets
))
359 "threads and transports both waiting??\n");
361 pool
->sp_stats
.packets
++;
363 /* Mark transport as busy. It will remain in this state until
364 * the provider calls svc_xprt_received. We update XPT_BUSY
365 * atomically because it also guards against trying to enqueue
366 * the transport twice.
368 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
)) {
369 /* Don't enqueue transport while already enqueued */
370 dprintk("svc: transport %p busy, not enqueued\n", xprt
);
374 if (!list_empty(&pool
->sp_threads
)) {
375 rqstp
= list_entry(pool
->sp_threads
.next
,
378 dprintk("svc: transport %p served by daemon %p\n",
380 svc_thread_dequeue(pool
, rqstp
);
383 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
384 rqstp
, rqstp
->rq_xprt
);
385 rqstp
->rq_xprt
= xprt
;
387 pool
->sp_stats
.threads_woken
++;
388 wake_up(&rqstp
->rq_wait
);
390 dprintk("svc: transport %p put into queue\n", xprt
);
391 list_add_tail(&xprt
->xpt_ready
, &pool
->sp_sockets
);
392 pool
->sp_stats
.sockets_queued
++;
396 spin_unlock_bh(&pool
->sp_lock
);
400 * Queue up a transport with data pending. If there are idle nfsd
401 * processes, wake 'em up.
404 void svc_xprt_enqueue(struct svc_xprt
*xprt
)
406 if (test_bit(XPT_BUSY
, &xprt
->xpt_flags
))
408 svc_xprt_do_enqueue(xprt
);
410 EXPORT_SYMBOL_GPL(svc_xprt_enqueue
);
413 * Dequeue the first transport. Must be called with the pool->sp_lock held.
415 static struct svc_xprt
*svc_xprt_dequeue(struct svc_pool
*pool
)
417 struct svc_xprt
*xprt
;
419 if (list_empty(&pool
->sp_sockets
))
422 xprt
= list_entry(pool
->sp_sockets
.next
,
423 struct svc_xprt
, xpt_ready
);
424 list_del_init(&xprt
->xpt_ready
);
426 dprintk("svc: transport %p dequeued, inuse=%d\n",
427 xprt
, atomic_read(&xprt
->xpt_ref
.refcount
));
433 * svc_reserve - change the space reserved for the reply to a request.
434 * @rqstp: The request in question
435 * @space: new max space to reserve
437 * Each request reserves some space on the output queue of the transport
438 * to make sure the reply fits. This function reduces that reserved
439 * space to be the amount of space used already, plus @space.
442 void svc_reserve(struct svc_rqst
*rqstp
, int space
)
444 space
+= rqstp
->rq_res
.head
[0].iov_len
;
446 if (space
< rqstp
->rq_reserved
) {
447 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
448 atomic_sub((rqstp
->rq_reserved
- space
), &xprt
->xpt_reserved
);
449 rqstp
->rq_reserved
= space
;
451 if (xprt
->xpt_ops
->xpo_adjust_wspace
)
452 xprt
->xpt_ops
->xpo_adjust_wspace(xprt
);
453 svc_xprt_enqueue(xprt
);
456 EXPORT_SYMBOL_GPL(svc_reserve
);
458 static void svc_xprt_release(struct svc_rqst
*rqstp
)
460 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
462 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
464 kfree(rqstp
->rq_deferred
);
465 rqstp
->rq_deferred
= NULL
;
467 svc_free_res_pages(rqstp
);
468 rqstp
->rq_res
.page_len
= 0;
469 rqstp
->rq_res
.page_base
= 0;
471 /* Reset response buffer and release
473 * But first, check that enough space was reserved
474 * for the reply, otherwise we have a bug!
476 if ((rqstp
->rq_res
.len
) > rqstp
->rq_reserved
)
477 printk(KERN_ERR
"RPC request reserved %d but used %d\n",
481 rqstp
->rq_res
.head
[0].iov_len
= 0;
482 svc_reserve(rqstp
, 0);
483 rqstp
->rq_xprt
= NULL
;
489 * External function to wake up a server waiting for data
490 * This really only makes sense for services like lockd
491 * which have exactly one thread anyway.
493 void svc_wake_up(struct svc_serv
*serv
)
495 struct svc_rqst
*rqstp
;
497 struct svc_pool
*pool
;
499 for (i
= 0; i
< serv
->sv_nrpools
; i
++) {
500 pool
= &serv
->sv_pools
[i
];
502 spin_lock_bh(&pool
->sp_lock
);
503 if (!list_empty(&pool
->sp_threads
)) {
504 rqstp
= list_entry(pool
->sp_threads
.next
,
507 dprintk("svc: daemon %p woken up.\n", rqstp
);
509 svc_thread_dequeue(pool, rqstp);
510 rqstp->rq_xprt = NULL;
512 wake_up(&rqstp
->rq_wait
);
514 pool
->sp_task_pending
= 1;
515 spin_unlock_bh(&pool
->sp_lock
);
518 EXPORT_SYMBOL_GPL(svc_wake_up
);
520 int svc_port_is_privileged(struct sockaddr
*sin
)
522 switch (sin
->sa_family
) {
524 return ntohs(((struct sockaddr_in
*)sin
)->sin_port
)
527 return ntohs(((struct sockaddr_in6
*)sin
)->sin6_port
)
535 * Make sure that we don't have too many active connections. If we have,
536 * something must be dropped. It's not clear what will happen if we allow
537 * "too many" connections, but when dealing with network-facing software,
538 * we have to code defensively. Here we do that by imposing hard limits.
540 * There's no point in trying to do random drop here for DoS
541 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
542 * attacker can easily beat that.
544 * The only somewhat efficient mechanism would be if drop old
545 * connections from the same IP first. But right now we don't even
546 * record the client IP in svc_sock.
548 * single-threaded services that expect a lot of clients will probably
549 * need to set sv_maxconn to override the default value which is based
550 * on the number of threads
552 static void svc_check_conn_limits(struct svc_serv
*serv
)
554 unsigned int limit
= serv
->sv_maxconn
? serv
->sv_maxconn
:
555 (serv
->sv_nrthreads
+3) * 20;
557 if (serv
->sv_tmpcnt
> limit
) {
558 struct svc_xprt
*xprt
= NULL
;
559 spin_lock_bh(&serv
->sv_lock
);
560 if (!list_empty(&serv
->sv_tempsocks
)) {
561 /* Try to help the admin */
562 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
563 serv
->sv_name
, serv
->sv_maxconn
?
564 "max number of connections" :
565 "number of threads");
567 * Always select the oldest connection. It's not fair,
570 xprt
= list_entry(serv
->sv_tempsocks
.prev
,
573 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
576 spin_unlock_bh(&serv
->sv_lock
);
579 svc_xprt_enqueue(xprt
);
585 static int svc_alloc_arg(struct svc_rqst
*rqstp
)
587 struct svc_serv
*serv
= rqstp
->rq_server
;
592 /* now allocate needed pages. If we get a failure, sleep briefly */
593 pages
= (serv
->sv_max_mesg
+ PAGE_SIZE
) / PAGE_SIZE
;
594 WARN_ON_ONCE(pages
>= RPCSVC_MAXPAGES
);
595 if (pages
>= RPCSVC_MAXPAGES
)
596 /* use as many pages as possible */
597 pages
= RPCSVC_MAXPAGES
- 1;
598 for (i
= 0; i
< pages
; i
++)
599 while (rqstp
->rq_pages
[i
] == NULL
) {
600 struct page
*p
= alloc_page(GFP_KERNEL
);
602 set_current_state(TASK_INTERRUPTIBLE
);
603 if (signalled() || kthread_should_stop()) {
604 set_current_state(TASK_RUNNING
);
607 schedule_timeout(msecs_to_jiffies(500));
609 rqstp
->rq_pages
[i
] = p
;
611 rqstp
->rq_page_end
= &rqstp
->rq_pages
[i
];
612 rqstp
->rq_pages
[i
++] = NULL
; /* this might be seen in nfs_read_actor */
614 /* Make arg->head point to first page and arg->pages point to rest */
615 arg
= &rqstp
->rq_arg
;
616 arg
->head
[0].iov_base
= page_address(rqstp
->rq_pages
[0]);
617 arg
->head
[0].iov_len
= PAGE_SIZE
;
618 arg
->pages
= rqstp
->rq_pages
+ 1;
620 /* save at least one page for response */
621 arg
->page_len
= (pages
-2)*PAGE_SIZE
;
622 arg
->len
= (pages
-1)*PAGE_SIZE
;
623 arg
->tail
[0].iov_len
= 0;
627 static struct svc_xprt
*svc_get_next_xprt(struct svc_rqst
*rqstp
, long timeout
)
629 struct svc_xprt
*xprt
;
630 struct svc_pool
*pool
= rqstp
->rq_pool
;
631 DECLARE_WAITQUEUE(wait
, current
);
634 /* Normally we will wait up to 5 seconds for any required
635 * cache information to be provided.
637 rqstp
->rq_chandle
.thread_wait
= 5*HZ
;
639 spin_lock_bh(&pool
->sp_lock
);
640 xprt
= svc_xprt_dequeue(pool
);
642 rqstp
->rq_xprt
= xprt
;
645 /* As there is a shortage of threads and this request
646 * had to be queued, don't allow the thread to wait so
647 * long for cache updates.
649 rqstp
->rq_chandle
.thread_wait
= 1*HZ
;
650 pool
->sp_task_pending
= 0;
652 if (pool
->sp_task_pending
) {
653 pool
->sp_task_pending
= 0;
654 spin_unlock_bh(&pool
->sp_lock
);
655 return ERR_PTR(-EAGAIN
);
657 /* No data pending. Go to sleep */
658 svc_thread_enqueue(pool
, rqstp
);
661 * We have to be able to interrupt this wait
662 * to bring down the daemons ...
664 set_current_state(TASK_INTERRUPTIBLE
);
667 * checking kthread_should_stop() here allows us to avoid
668 * locking and signalling when stopping kthreads that call
669 * svc_recv. If the thread has already been woken up, then
670 * we can exit here without sleeping. If not, then it
671 * it'll be woken up quickly during the schedule_timeout
673 if (kthread_should_stop()) {
674 set_current_state(TASK_RUNNING
);
675 spin_unlock_bh(&pool
->sp_lock
);
676 return ERR_PTR(-EINTR
);
679 add_wait_queue(&rqstp
->rq_wait
, &wait
);
680 spin_unlock_bh(&pool
->sp_lock
);
682 time_left
= schedule_timeout(timeout
);
686 spin_lock_bh(&pool
->sp_lock
);
687 remove_wait_queue(&rqstp
->rq_wait
, &wait
);
689 pool
->sp_stats
.threads_timedout
++;
691 xprt
= rqstp
->rq_xprt
;
693 svc_thread_dequeue(pool
, rqstp
);
694 spin_unlock_bh(&pool
->sp_lock
);
695 dprintk("svc: server %p, no data yet\n", rqstp
);
696 if (signalled() || kthread_should_stop())
697 return ERR_PTR(-EINTR
);
699 return ERR_PTR(-EAGAIN
);
702 spin_unlock_bh(&pool
->sp_lock
);
706 static void svc_add_new_temp_xprt(struct svc_serv
*serv
, struct svc_xprt
*newxpt
)
708 spin_lock_bh(&serv
->sv_lock
);
709 set_bit(XPT_TEMP
, &newxpt
->xpt_flags
);
710 list_add(&newxpt
->xpt_list
, &serv
->sv_tempsocks
);
712 if (serv
->sv_temptimer
.function
== NULL
) {
713 /* setup timer to age temp transports */
714 setup_timer(&serv
->sv_temptimer
, svc_age_temp_xprts
,
715 (unsigned long)serv
);
716 mod_timer(&serv
->sv_temptimer
,
717 jiffies
+ svc_conn_age_period
* HZ
);
719 spin_unlock_bh(&serv
->sv_lock
);
720 svc_xprt_received(newxpt
);
723 static int svc_handle_xprt(struct svc_rqst
*rqstp
, struct svc_xprt
*xprt
)
725 struct svc_serv
*serv
= rqstp
->rq_server
;
728 if (test_bit(XPT_CLOSE
, &xprt
->xpt_flags
)) {
729 dprintk("svc_recv: found XPT_CLOSE\n");
730 svc_delete_xprt(xprt
);
731 /* Leave XPT_BUSY set on the dead xprt: */
734 if (test_bit(XPT_LISTENER
, &xprt
->xpt_flags
)) {
735 struct svc_xprt
*newxpt
;
737 * We know this module_get will succeed because the
738 * listener holds a reference too
740 __module_get(xprt
->xpt_class
->xcl_owner
);
741 svc_check_conn_limits(xprt
->xpt_server
);
742 newxpt
= xprt
->xpt_ops
->xpo_accept(xprt
);
744 svc_add_new_temp_xprt(serv
, newxpt
);
746 module_put(xprt
->xpt_class
->xcl_owner
);
747 } else if (xprt
->xpt_ops
->xpo_has_wspace(xprt
)) {
748 /* XPT_DATA|XPT_DEFERRED case: */
749 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
750 rqstp
, rqstp
->rq_pool
->sp_id
, xprt
,
751 atomic_read(&xprt
->xpt_ref
.refcount
));
752 rqstp
->rq_deferred
= svc_deferred_dequeue(xprt
);
753 if (rqstp
->rq_deferred
)
754 len
= svc_deferred_recv(rqstp
);
756 len
= xprt
->xpt_ops
->xpo_recvfrom(rqstp
);
757 dprintk("svc: got len=%d\n", len
);
758 rqstp
->rq_reserved
= serv
->sv_max_mesg
;
759 atomic_add(rqstp
->rq_reserved
, &xprt
->xpt_reserved
);
761 /* clear XPT_BUSY: */
762 svc_xprt_received(xprt
);
767 * Receive the next request on any transport. This code is carefully
768 * organised not to touch any cachelines in the shared svc_serv
769 * structure, only cachelines in the local svc_pool.
771 int svc_recv(struct svc_rqst
*rqstp
, long timeout
)
773 struct svc_xprt
*xprt
= NULL
;
774 struct svc_serv
*serv
= rqstp
->rq_server
;
777 dprintk("svc: server %p waiting for data (to = %ld)\n",
782 "svc_recv: service %p, transport not NULL!\n",
784 if (waitqueue_active(&rqstp
->rq_wait
))
786 "svc_recv: service %p, wait queue active!\n",
789 err
= svc_alloc_arg(rqstp
);
795 if (signalled() || kthread_should_stop())
798 xprt
= svc_get_next_xprt(rqstp
, timeout
);
800 return PTR_ERR(xprt
);
802 len
= svc_handle_xprt(rqstp
, xprt
);
804 /* No data, incomplete (TCP) read, or accept() */
808 clear_bit(XPT_OLD
, &xprt
->xpt_flags
);
810 rqstp
->rq_secure
= xprt
->xpt_ops
->xpo_secure_port(rqstp
);
811 rqstp
->rq_chandle
.defer
= svc_defer
;
814 serv
->sv_stats
->netcnt
++;
817 rqstp
->rq_res
.len
= 0;
818 svc_xprt_release(rqstp
);
821 EXPORT_SYMBOL_GPL(svc_recv
);
826 void svc_drop(struct svc_rqst
*rqstp
)
828 dprintk("svc: xprt %p dropped request\n", rqstp
->rq_xprt
);
829 svc_xprt_release(rqstp
);
831 EXPORT_SYMBOL_GPL(svc_drop
);
834 * Return reply to client.
836 int svc_send(struct svc_rqst
*rqstp
)
838 struct svc_xprt
*xprt
;
842 xprt
= rqstp
->rq_xprt
;
846 /* release the receive skb before sending the reply */
847 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
849 /* calculate over-all length */
851 xb
->len
= xb
->head
[0].iov_len
+
855 /* Grab mutex to serialize outgoing data. */
856 mutex_lock(&xprt
->xpt_mutex
);
857 if (test_bit(XPT_DEAD
, &xprt
->xpt_flags
)
858 || test_bit(XPT_CLOSE
, &xprt
->xpt_flags
))
861 len
= xprt
->xpt_ops
->xpo_sendto(rqstp
);
862 mutex_unlock(&xprt
->xpt_mutex
);
863 rpc_wake_up(&xprt
->xpt_bc_pending
);
864 svc_xprt_release(rqstp
);
866 if (len
== -ECONNREFUSED
|| len
== -ENOTCONN
|| len
== -EAGAIN
)
872 * Timer function to close old temporary transports, using
873 * a mark-and-sweep algorithm.
875 static void svc_age_temp_xprts(unsigned long closure
)
877 struct svc_serv
*serv
= (struct svc_serv
*)closure
;
878 struct svc_xprt
*xprt
;
879 struct list_head
*le
, *next
;
881 dprintk("svc_age_temp_xprts\n");
883 if (!spin_trylock_bh(&serv
->sv_lock
)) {
884 /* busy, try again 1 sec later */
885 dprintk("svc_age_temp_xprts: busy\n");
886 mod_timer(&serv
->sv_temptimer
, jiffies
+ HZ
);
890 list_for_each_safe(le
, next
, &serv
->sv_tempsocks
) {
891 xprt
= list_entry(le
, struct svc_xprt
, xpt_list
);
893 /* First time through, just mark it OLD. Second time
894 * through, close it. */
895 if (!test_and_set_bit(XPT_OLD
, &xprt
->xpt_flags
))
897 if (atomic_read(&xprt
->xpt_ref
.refcount
) > 1 ||
898 test_bit(XPT_BUSY
, &xprt
->xpt_flags
))
901 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
902 set_bit(XPT_DETACHED
, &xprt
->xpt_flags
);
903 dprintk("queuing xprt %p for closing\n", xprt
);
905 /* a thread will dequeue and close it soon */
906 svc_xprt_enqueue(xprt
);
908 spin_unlock_bh(&serv
->sv_lock
);
910 mod_timer(&serv
->sv_temptimer
, jiffies
+ svc_conn_age_period
* HZ
);
913 static void call_xpt_users(struct svc_xprt
*xprt
)
915 struct svc_xpt_user
*u
;
917 spin_lock(&xprt
->xpt_lock
);
918 while (!list_empty(&xprt
->xpt_users
)) {
919 u
= list_first_entry(&xprt
->xpt_users
, struct svc_xpt_user
, list
);
923 spin_unlock(&xprt
->xpt_lock
);
927 * Remove a dead transport
929 static void svc_delete_xprt(struct svc_xprt
*xprt
)
931 struct svc_serv
*serv
= xprt
->xpt_server
;
932 struct svc_deferred_req
*dr
;
934 /* Only do this once */
935 if (test_and_set_bit(XPT_DEAD
, &xprt
->xpt_flags
))
938 dprintk("svc: svc_delete_xprt(%p)\n", xprt
);
939 xprt
->xpt_ops
->xpo_detach(xprt
);
941 spin_lock_bh(&serv
->sv_lock
);
942 if (!test_and_set_bit(XPT_DETACHED
, &xprt
->xpt_flags
))
943 list_del_init(&xprt
->xpt_list
);
944 WARN_ON_ONCE(!list_empty(&xprt
->xpt_ready
));
945 if (test_bit(XPT_TEMP
, &xprt
->xpt_flags
))
947 spin_unlock_bh(&serv
->sv_lock
);
949 while ((dr
= svc_deferred_dequeue(xprt
)) != NULL
)
952 call_xpt_users(xprt
);
956 void svc_close_xprt(struct svc_xprt
*xprt
)
958 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
959 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
))
960 /* someone else will have to effect the close */
963 * We expect svc_close_xprt() to work even when no threads are
964 * running (e.g., while configuring the server before starting
965 * any threads), so if the transport isn't busy, we delete
968 svc_delete_xprt(xprt
);
970 EXPORT_SYMBOL_GPL(svc_close_xprt
);
972 static int svc_close_list(struct svc_serv
*serv
, struct list_head
*xprt_list
, struct net
*net
)
974 struct svc_xprt
*xprt
;
977 spin_lock(&serv
->sv_lock
);
978 list_for_each_entry(xprt
, xprt_list
, xpt_list
) {
979 if (xprt
->xpt_net
!= net
)
982 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
983 svc_xprt_enqueue(xprt
);
985 spin_unlock(&serv
->sv_lock
);
989 static struct svc_xprt
*svc_dequeue_net(struct svc_serv
*serv
, struct net
*net
)
991 struct svc_pool
*pool
;
992 struct svc_xprt
*xprt
;
993 struct svc_xprt
*tmp
;
996 for (i
= 0; i
< serv
->sv_nrpools
; i
++) {
997 pool
= &serv
->sv_pools
[i
];
999 spin_lock_bh(&pool
->sp_lock
);
1000 list_for_each_entry_safe(xprt
, tmp
, &pool
->sp_sockets
, xpt_ready
) {
1001 if (xprt
->xpt_net
!= net
)
1003 list_del_init(&xprt
->xpt_ready
);
1004 spin_unlock_bh(&pool
->sp_lock
);
1007 spin_unlock_bh(&pool
->sp_lock
);
1012 static void svc_clean_up_xprts(struct svc_serv
*serv
, struct net
*net
)
1014 struct svc_xprt
*xprt
;
1016 while ((xprt
= svc_dequeue_net(serv
, net
))) {
1017 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
1018 svc_delete_xprt(xprt
);
1023 * Server threads may still be running (especially in the case where the
1024 * service is still running in other network namespaces).
1026 * So we shut down sockets the same way we would on a running server, by
1027 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1028 * the close. In the case there are no such other threads,
1029 * threads running, svc_clean_up_xprts() does a simple version of a
1030 * server's main event loop, and in the case where there are other
1031 * threads, we may need to wait a little while and then check again to
1032 * see if they're done.
1034 void svc_close_net(struct svc_serv
*serv
, struct net
*net
)
1038 while (svc_close_list(serv
, &serv
->sv_permsocks
, net
) +
1039 svc_close_list(serv
, &serv
->sv_tempsocks
, net
)) {
1041 svc_clean_up_xprts(serv
, net
);
1047 * Handle defer and revisit of requests
1050 static void svc_revisit(struct cache_deferred_req
*dreq
, int too_many
)
1052 struct svc_deferred_req
*dr
=
1053 container_of(dreq
, struct svc_deferred_req
, handle
);
1054 struct svc_xprt
*xprt
= dr
->xprt
;
1056 spin_lock(&xprt
->xpt_lock
);
1057 set_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1058 if (too_many
|| test_bit(XPT_DEAD
, &xprt
->xpt_flags
)) {
1059 spin_unlock(&xprt
->xpt_lock
);
1060 dprintk("revisit canceled\n");
1065 dprintk("revisit queued\n");
1067 list_add(&dr
->handle
.recent
, &xprt
->xpt_deferred
);
1068 spin_unlock(&xprt
->xpt_lock
);
1069 svc_xprt_enqueue(xprt
);
1074 * Save the request off for later processing. The request buffer looks
1077 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1079 * This code can only handle requests that consist of an xprt-header
1082 static struct cache_deferred_req
*svc_defer(struct cache_req
*req
)
1084 struct svc_rqst
*rqstp
= container_of(req
, struct svc_rqst
, rq_chandle
);
1085 struct svc_deferred_req
*dr
;
1087 if (rqstp
->rq_arg
.page_len
|| !rqstp
->rq_usedeferral
)
1088 return NULL
; /* if more than a page, give up FIXME */
1089 if (rqstp
->rq_deferred
) {
1090 dr
= rqstp
->rq_deferred
;
1091 rqstp
->rq_deferred
= NULL
;
1095 /* FIXME maybe discard if size too large */
1096 size
= sizeof(struct svc_deferred_req
) + rqstp
->rq_arg
.len
;
1097 dr
= kmalloc(size
, GFP_KERNEL
);
1101 dr
->handle
.owner
= rqstp
->rq_server
;
1102 dr
->prot
= rqstp
->rq_prot
;
1103 memcpy(&dr
->addr
, &rqstp
->rq_addr
, rqstp
->rq_addrlen
);
1104 dr
->addrlen
= rqstp
->rq_addrlen
;
1105 dr
->daddr
= rqstp
->rq_daddr
;
1106 dr
->argslen
= rqstp
->rq_arg
.len
>> 2;
1107 dr
->xprt_hlen
= rqstp
->rq_xprt_hlen
;
1109 /* back up head to the start of the buffer and copy */
1110 skip
= rqstp
->rq_arg
.len
- rqstp
->rq_arg
.head
[0].iov_len
;
1111 memcpy(dr
->args
, rqstp
->rq_arg
.head
[0].iov_base
- skip
,
1114 svc_xprt_get(rqstp
->rq_xprt
);
1115 dr
->xprt
= rqstp
->rq_xprt
;
1116 rqstp
->rq_dropme
= true;
1118 dr
->handle
.revisit
= svc_revisit
;
1123 * recv data from a deferred request into an active one
1125 static int svc_deferred_recv(struct svc_rqst
*rqstp
)
1127 struct svc_deferred_req
*dr
= rqstp
->rq_deferred
;
1129 /* setup iov_base past transport header */
1130 rqstp
->rq_arg
.head
[0].iov_base
= dr
->args
+ (dr
->xprt_hlen
>>2);
1131 /* The iov_len does not include the transport header bytes */
1132 rqstp
->rq_arg
.head
[0].iov_len
= (dr
->argslen
<<2) - dr
->xprt_hlen
;
1133 rqstp
->rq_arg
.page_len
= 0;
1134 /* The rq_arg.len includes the transport header bytes */
1135 rqstp
->rq_arg
.len
= dr
->argslen
<<2;
1136 rqstp
->rq_prot
= dr
->prot
;
1137 memcpy(&rqstp
->rq_addr
, &dr
->addr
, dr
->addrlen
);
1138 rqstp
->rq_addrlen
= dr
->addrlen
;
1139 /* Save off transport header len in case we get deferred again */
1140 rqstp
->rq_xprt_hlen
= dr
->xprt_hlen
;
1141 rqstp
->rq_daddr
= dr
->daddr
;
1142 rqstp
->rq_respages
= rqstp
->rq_pages
;
1143 return (dr
->argslen
<<2) - dr
->xprt_hlen
;
1147 static struct svc_deferred_req
*svc_deferred_dequeue(struct svc_xprt
*xprt
)
1149 struct svc_deferred_req
*dr
= NULL
;
1151 if (!test_bit(XPT_DEFERRED
, &xprt
->xpt_flags
))
1153 spin_lock(&xprt
->xpt_lock
);
1154 if (!list_empty(&xprt
->xpt_deferred
)) {
1155 dr
= list_entry(xprt
->xpt_deferred
.next
,
1156 struct svc_deferred_req
,
1158 list_del_init(&dr
->handle
.recent
);
1160 clear_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1161 spin_unlock(&xprt
->xpt_lock
);
1166 * svc_find_xprt - find an RPC transport instance
1167 * @serv: pointer to svc_serv to search
1168 * @xcl_name: C string containing transport's class name
1169 * @net: owner net pointer
1170 * @af: Address family of transport's local address
1171 * @port: transport's IP port number
1173 * Return the transport instance pointer for the endpoint accepting
1174 * connections/peer traffic from the specified transport class,
1175 * address family and port.
1177 * Specifying 0 for the address family or port is effectively a
1178 * wild-card, and will result in matching the first transport in the
1179 * service's list that has a matching class name.
1181 struct svc_xprt
*svc_find_xprt(struct svc_serv
*serv
, const char *xcl_name
,
1182 struct net
*net
, const sa_family_t af
,
1183 const unsigned short port
)
1185 struct svc_xprt
*xprt
;
1186 struct svc_xprt
*found
= NULL
;
1188 /* Sanity check the args */
1189 if (serv
== NULL
|| xcl_name
== NULL
)
1192 spin_lock_bh(&serv
->sv_lock
);
1193 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1194 if (xprt
->xpt_net
!= net
)
1196 if (strcmp(xprt
->xpt_class
->xcl_name
, xcl_name
))
1198 if (af
!= AF_UNSPEC
&& af
!= xprt
->xpt_local
.ss_family
)
1200 if (port
!= 0 && port
!= svc_xprt_local_port(xprt
))
1206 spin_unlock_bh(&serv
->sv_lock
);
1209 EXPORT_SYMBOL_GPL(svc_find_xprt
);
1211 static int svc_one_xprt_name(const struct svc_xprt
*xprt
,
1212 char *pos
, int remaining
)
1216 len
= snprintf(pos
, remaining
, "%s %u\n",
1217 xprt
->xpt_class
->xcl_name
,
1218 svc_xprt_local_port(xprt
));
1219 if (len
>= remaining
)
1220 return -ENAMETOOLONG
;
1225 * svc_xprt_names - format a buffer with a list of transport names
1226 * @serv: pointer to an RPC service
1227 * @buf: pointer to a buffer to be filled in
1228 * @buflen: length of buffer to be filled in
1230 * Fills in @buf with a string containing a list of transport names,
1231 * each name terminated with '\n'.
1233 * Returns positive length of the filled-in string on success; otherwise
1234 * a negative errno value is returned if an error occurs.
1236 int svc_xprt_names(struct svc_serv
*serv
, char *buf
, const int buflen
)
1238 struct svc_xprt
*xprt
;
1242 /* Sanity check args */
1246 spin_lock_bh(&serv
->sv_lock
);
1250 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1251 len
= svc_one_xprt_name(xprt
, pos
, buflen
- totlen
);
1263 spin_unlock_bh(&serv
->sv_lock
);
1266 EXPORT_SYMBOL_GPL(svc_xprt_names
);
1269 /*----------------------------------------------------------------------------*/
1271 static void *svc_pool_stats_start(struct seq_file
*m
, loff_t
*pos
)
1273 unsigned int pidx
= (unsigned int)*pos
;
1274 struct svc_serv
*serv
= m
->private;
1276 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx
);
1279 return SEQ_START_TOKEN
;
1280 return (pidx
> serv
->sv_nrpools
? NULL
: &serv
->sv_pools
[pidx
-1]);
1283 static void *svc_pool_stats_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
1285 struct svc_pool
*pool
= p
;
1286 struct svc_serv
*serv
= m
->private;
1288 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos
);
1290 if (p
== SEQ_START_TOKEN
) {
1291 pool
= &serv
->sv_pools
[0];
1293 unsigned int pidx
= (pool
- &serv
->sv_pools
[0]);
1294 if (pidx
< serv
->sv_nrpools
-1)
1295 pool
= &serv
->sv_pools
[pidx
+1];
1303 static void svc_pool_stats_stop(struct seq_file
*m
, void *p
)
1307 static int svc_pool_stats_show(struct seq_file
*m
, void *p
)
1309 struct svc_pool
*pool
= p
;
1311 if (p
== SEQ_START_TOKEN
) {
1312 seq_puts(m
, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1316 seq_printf(m
, "%u %lu %lu %lu %lu\n",
1318 pool
->sp_stats
.packets
,
1319 pool
->sp_stats
.sockets_queued
,
1320 pool
->sp_stats
.threads_woken
,
1321 pool
->sp_stats
.threads_timedout
);
1326 static const struct seq_operations svc_pool_stats_seq_ops
= {
1327 .start
= svc_pool_stats_start
,
1328 .next
= svc_pool_stats_next
,
1329 .stop
= svc_pool_stats_stop
,
1330 .show
= svc_pool_stats_show
,
1333 int svc_pool_stats_open(struct svc_serv
*serv
, struct file
*file
)
1337 err
= seq_open(file
, &svc_pool_stats_seq_ops
);
1339 ((struct seq_file
*) file
->private_data
)->private = serv
;
1342 EXPORT_SYMBOL(svc_pool_stats_open
);
1344 /*----------------------------------------------------------------------------*/