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
))
349 /* Mark transport as busy. It will remain in this state until
350 * the provider calls svc_xprt_received. We update XPT_BUSY
351 * atomically because it also guards against trying to enqueue
352 * the transport twice.
354 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
)) {
355 /* Don't enqueue transport while already enqueued */
356 dprintk("svc: transport %p busy, not enqueued\n", xprt
);
361 pool
= svc_pool_for_cpu(xprt
->xpt_server
, cpu
);
362 spin_lock_bh(&pool
->sp_lock
);
364 pool
->sp_stats
.packets
++;
366 if (!list_empty(&pool
->sp_threads
)) {
367 rqstp
= list_entry(pool
->sp_threads
.next
,
370 dprintk("svc: transport %p served by daemon %p\n",
372 svc_thread_dequeue(pool
, rqstp
);
375 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
376 rqstp
, rqstp
->rq_xprt
);
377 /* Note the order of the following 3 lines:
378 * We want to assign xprt to rqstp->rq_xprt only _after_
379 * we've woken up the process, so that we don't race with
380 * the lockless check in svc_get_next_xprt().
383 wake_up_process(rqstp
->rq_task
);
384 rqstp
->rq_xprt
= xprt
;
385 pool
->sp_stats
.threads_woken
++;
387 dprintk("svc: transport %p put into queue\n", xprt
);
388 list_add_tail(&xprt
->xpt_ready
, &pool
->sp_sockets
);
389 pool
->sp_stats
.sockets_queued
++;
392 spin_unlock_bh(&pool
->sp_lock
);
397 * Queue up a transport with data pending. If there are idle nfsd
398 * processes, wake 'em up.
401 void svc_xprt_enqueue(struct svc_xprt
*xprt
)
403 if (test_bit(XPT_BUSY
, &xprt
->xpt_flags
))
405 svc_xprt_do_enqueue(xprt
);
407 EXPORT_SYMBOL_GPL(svc_xprt_enqueue
);
410 * Dequeue the first transport. Must be called with the pool->sp_lock held.
412 static struct svc_xprt
*svc_xprt_dequeue(struct svc_pool
*pool
)
414 struct svc_xprt
*xprt
;
416 if (list_empty(&pool
->sp_sockets
))
419 xprt
= list_entry(pool
->sp_sockets
.next
,
420 struct svc_xprt
, xpt_ready
);
421 list_del_init(&xprt
->xpt_ready
);
423 dprintk("svc: transport %p dequeued, inuse=%d\n",
424 xprt
, atomic_read(&xprt
->xpt_ref
.refcount
));
430 * svc_reserve - change the space reserved for the reply to a request.
431 * @rqstp: The request in question
432 * @space: new max space to reserve
434 * Each request reserves some space on the output queue of the transport
435 * to make sure the reply fits. This function reduces that reserved
436 * space to be the amount of space used already, plus @space.
439 void svc_reserve(struct svc_rqst
*rqstp
, int space
)
441 space
+= rqstp
->rq_res
.head
[0].iov_len
;
443 if (space
< rqstp
->rq_reserved
) {
444 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
445 atomic_sub((rqstp
->rq_reserved
- space
), &xprt
->xpt_reserved
);
446 rqstp
->rq_reserved
= space
;
448 if (xprt
->xpt_ops
->xpo_adjust_wspace
)
449 xprt
->xpt_ops
->xpo_adjust_wspace(xprt
);
450 svc_xprt_enqueue(xprt
);
453 EXPORT_SYMBOL_GPL(svc_reserve
);
455 static void svc_xprt_release(struct svc_rqst
*rqstp
)
457 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
459 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
461 kfree(rqstp
->rq_deferred
);
462 rqstp
->rq_deferred
= NULL
;
464 svc_free_res_pages(rqstp
);
465 rqstp
->rq_res
.page_len
= 0;
466 rqstp
->rq_res
.page_base
= 0;
468 /* Reset response buffer and release
470 * But first, check that enough space was reserved
471 * for the reply, otherwise we have a bug!
473 if ((rqstp
->rq_res
.len
) > rqstp
->rq_reserved
)
474 printk(KERN_ERR
"RPC request reserved %d but used %d\n",
478 rqstp
->rq_res
.head
[0].iov_len
= 0;
479 svc_reserve(rqstp
, 0);
480 rqstp
->rq_xprt
= NULL
;
486 * External function to wake up a server waiting for data
487 * This really only makes sense for services like lockd
488 * which have exactly one thread anyway.
490 void svc_wake_up(struct svc_serv
*serv
)
492 struct svc_rqst
*rqstp
;
494 struct svc_pool
*pool
;
496 for (i
= 0; i
< serv
->sv_nrpools
; i
++) {
497 pool
= &serv
->sv_pools
[i
];
499 spin_lock_bh(&pool
->sp_lock
);
500 if (!list_empty(&pool
->sp_threads
)) {
501 rqstp
= list_entry(pool
->sp_threads
.next
,
504 dprintk("svc: daemon %p woken up.\n", rqstp
);
506 svc_thread_dequeue(pool, rqstp);
507 rqstp->rq_xprt = NULL;
509 wake_up_process(rqstp
->rq_task
);
511 pool
->sp_task_pending
= 1;
512 spin_unlock_bh(&pool
->sp_lock
);
515 EXPORT_SYMBOL_GPL(svc_wake_up
);
517 int svc_port_is_privileged(struct sockaddr
*sin
)
519 switch (sin
->sa_family
) {
521 return ntohs(((struct sockaddr_in
*)sin
)->sin_port
)
524 return ntohs(((struct sockaddr_in6
*)sin
)->sin6_port
)
532 * Make sure that we don't have too many active connections. If we have,
533 * something must be dropped. It's not clear what will happen if we allow
534 * "too many" connections, but when dealing with network-facing software,
535 * we have to code defensively. Here we do that by imposing hard limits.
537 * There's no point in trying to do random drop here for DoS
538 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
539 * attacker can easily beat that.
541 * The only somewhat efficient mechanism would be if drop old
542 * connections from the same IP first. But right now we don't even
543 * record the client IP in svc_sock.
545 * single-threaded services that expect a lot of clients will probably
546 * need to set sv_maxconn to override the default value which is based
547 * on the number of threads
549 static void svc_check_conn_limits(struct svc_serv
*serv
)
551 unsigned int limit
= serv
->sv_maxconn
? serv
->sv_maxconn
:
552 (serv
->sv_nrthreads
+3) * 20;
554 if (serv
->sv_tmpcnt
> limit
) {
555 struct svc_xprt
*xprt
= NULL
;
556 spin_lock_bh(&serv
->sv_lock
);
557 if (!list_empty(&serv
->sv_tempsocks
)) {
558 /* Try to help the admin */
559 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
560 serv
->sv_name
, serv
->sv_maxconn
?
561 "max number of connections" :
562 "number of threads");
564 * Always select the oldest connection. It's not fair,
567 xprt
= list_entry(serv
->sv_tempsocks
.prev
,
570 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
573 spin_unlock_bh(&serv
->sv_lock
);
576 svc_xprt_enqueue(xprt
);
582 static int svc_alloc_arg(struct svc_rqst
*rqstp
)
584 struct svc_serv
*serv
= rqstp
->rq_server
;
589 /* now allocate needed pages. If we get a failure, sleep briefly */
590 pages
= (serv
->sv_max_mesg
+ PAGE_SIZE
) / PAGE_SIZE
;
591 WARN_ON_ONCE(pages
>= RPCSVC_MAXPAGES
);
592 if (pages
>= RPCSVC_MAXPAGES
)
593 /* use as many pages as possible */
594 pages
= RPCSVC_MAXPAGES
- 1;
595 for (i
= 0; i
< pages
; i
++)
596 while (rqstp
->rq_pages
[i
] == NULL
) {
597 struct page
*p
= alloc_page(GFP_KERNEL
);
599 set_current_state(TASK_INTERRUPTIBLE
);
600 if (signalled() || kthread_should_stop()) {
601 set_current_state(TASK_RUNNING
);
604 schedule_timeout(msecs_to_jiffies(500));
606 rqstp
->rq_pages
[i
] = p
;
608 rqstp
->rq_page_end
= &rqstp
->rq_pages
[i
];
609 rqstp
->rq_pages
[i
++] = NULL
; /* this might be seen in nfs_read_actor */
611 /* Make arg->head point to first page and arg->pages point to rest */
612 arg
= &rqstp
->rq_arg
;
613 arg
->head
[0].iov_base
= page_address(rqstp
->rq_pages
[0]);
614 arg
->head
[0].iov_len
= PAGE_SIZE
;
615 arg
->pages
= rqstp
->rq_pages
+ 1;
617 /* save at least one page for response */
618 arg
->page_len
= (pages
-2)*PAGE_SIZE
;
619 arg
->len
= (pages
-1)*PAGE_SIZE
;
620 arg
->tail
[0].iov_len
= 0;
624 static struct svc_xprt
*svc_get_next_xprt(struct svc_rqst
*rqstp
, long timeout
)
626 struct svc_xprt
*xprt
;
627 struct svc_pool
*pool
= rqstp
->rq_pool
;
630 /* Normally we will wait up to 5 seconds for any required
631 * cache information to be provided.
633 rqstp
->rq_chandle
.thread_wait
= 5*HZ
;
635 spin_lock_bh(&pool
->sp_lock
);
636 xprt
= svc_xprt_dequeue(pool
);
638 rqstp
->rq_xprt
= xprt
;
641 /* As there is a shortage of threads and this request
642 * had to be queued, don't allow the thread to wait so
643 * long for cache updates.
645 rqstp
->rq_chandle
.thread_wait
= 1*HZ
;
646 pool
->sp_task_pending
= 0;
648 if (pool
->sp_task_pending
) {
649 pool
->sp_task_pending
= 0;
650 xprt
= ERR_PTR(-EAGAIN
);
654 * We have to be able to interrupt this wait
655 * to bring down the daemons ...
657 set_current_state(TASK_INTERRUPTIBLE
);
659 /* No data pending. Go to sleep */
660 svc_thread_enqueue(pool
, rqstp
);
661 spin_unlock_bh(&pool
->sp_lock
);
663 if (!(signalled() || kthread_should_stop())) {
664 time_left
= schedule_timeout(timeout
);
665 __set_current_state(TASK_RUNNING
);
669 xprt
= rqstp
->rq_xprt
;
673 __set_current_state(TASK_RUNNING
);
675 spin_lock_bh(&pool
->sp_lock
);
677 pool
->sp_stats
.threads_timedout
++;
679 xprt
= rqstp
->rq_xprt
;
681 svc_thread_dequeue(pool
, rqstp
);
682 spin_unlock_bh(&pool
->sp_lock
);
683 dprintk("svc: server %p, no data yet\n", rqstp
);
684 if (signalled() || kthread_should_stop())
685 return ERR_PTR(-EINTR
);
687 return ERR_PTR(-EAGAIN
);
691 spin_unlock_bh(&pool
->sp_lock
);
695 static void svc_add_new_temp_xprt(struct svc_serv
*serv
, struct svc_xprt
*newxpt
)
697 spin_lock_bh(&serv
->sv_lock
);
698 set_bit(XPT_TEMP
, &newxpt
->xpt_flags
);
699 list_add(&newxpt
->xpt_list
, &serv
->sv_tempsocks
);
701 if (serv
->sv_temptimer
.function
== NULL
) {
702 /* setup timer to age temp transports */
703 setup_timer(&serv
->sv_temptimer
, svc_age_temp_xprts
,
704 (unsigned long)serv
);
705 mod_timer(&serv
->sv_temptimer
,
706 jiffies
+ svc_conn_age_period
* HZ
);
708 spin_unlock_bh(&serv
->sv_lock
);
709 svc_xprt_received(newxpt
);
712 static int svc_handle_xprt(struct svc_rqst
*rqstp
, struct svc_xprt
*xprt
)
714 struct svc_serv
*serv
= rqstp
->rq_server
;
717 if (test_bit(XPT_CLOSE
, &xprt
->xpt_flags
)) {
718 dprintk("svc_recv: found XPT_CLOSE\n");
719 svc_delete_xprt(xprt
);
720 /* Leave XPT_BUSY set on the dead xprt: */
723 if (test_bit(XPT_LISTENER
, &xprt
->xpt_flags
)) {
724 struct svc_xprt
*newxpt
;
726 * We know this module_get will succeed because the
727 * listener holds a reference too
729 __module_get(xprt
->xpt_class
->xcl_owner
);
730 svc_check_conn_limits(xprt
->xpt_server
);
731 newxpt
= xprt
->xpt_ops
->xpo_accept(xprt
);
733 svc_add_new_temp_xprt(serv
, newxpt
);
735 module_put(xprt
->xpt_class
->xcl_owner
);
737 /* XPT_DATA|XPT_DEFERRED case: */
738 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
739 rqstp
, rqstp
->rq_pool
->sp_id
, xprt
,
740 atomic_read(&xprt
->xpt_ref
.refcount
));
741 rqstp
->rq_deferred
= svc_deferred_dequeue(xprt
);
742 if (rqstp
->rq_deferred
)
743 len
= svc_deferred_recv(rqstp
);
745 len
= xprt
->xpt_ops
->xpo_recvfrom(rqstp
);
746 dprintk("svc: got len=%d\n", len
);
747 rqstp
->rq_reserved
= serv
->sv_max_mesg
;
748 atomic_add(rqstp
->rq_reserved
, &xprt
->xpt_reserved
);
750 /* clear XPT_BUSY: */
751 svc_xprt_received(xprt
);
756 * Receive the next request on any transport. This code is carefully
757 * organised not to touch any cachelines in the shared svc_serv
758 * structure, only cachelines in the local svc_pool.
760 int svc_recv(struct svc_rqst
*rqstp
, long timeout
)
762 struct svc_xprt
*xprt
= NULL
;
763 struct svc_serv
*serv
= rqstp
->rq_server
;
766 dprintk("svc: server %p waiting for data (to = %ld)\n",
771 "svc_recv: service %p, transport not NULL!\n",
774 err
= svc_alloc_arg(rqstp
);
780 if (signalled() || kthread_should_stop())
783 xprt
= svc_get_next_xprt(rqstp
, timeout
);
785 return PTR_ERR(xprt
);
787 len
= svc_handle_xprt(rqstp
, xprt
);
789 /* No data, incomplete (TCP) read, or accept() */
793 clear_bit(XPT_OLD
, &xprt
->xpt_flags
);
795 rqstp
->rq_secure
= xprt
->xpt_ops
->xpo_secure_port(rqstp
);
796 rqstp
->rq_chandle
.defer
= svc_defer
;
799 serv
->sv_stats
->netcnt
++;
802 rqstp
->rq_res
.len
= 0;
803 svc_xprt_release(rqstp
);
806 EXPORT_SYMBOL_GPL(svc_recv
);
811 void svc_drop(struct svc_rqst
*rqstp
)
813 dprintk("svc: xprt %p dropped request\n", rqstp
->rq_xprt
);
814 svc_xprt_release(rqstp
);
816 EXPORT_SYMBOL_GPL(svc_drop
);
819 * Return reply to client.
821 int svc_send(struct svc_rqst
*rqstp
)
823 struct svc_xprt
*xprt
;
827 xprt
= rqstp
->rq_xprt
;
831 /* release the receive skb before sending the reply */
832 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
834 /* calculate over-all length */
836 xb
->len
= xb
->head
[0].iov_len
+
840 /* Grab mutex to serialize outgoing data. */
841 mutex_lock(&xprt
->xpt_mutex
);
842 if (test_bit(XPT_DEAD
, &xprt
->xpt_flags
)
843 || test_bit(XPT_CLOSE
, &xprt
->xpt_flags
))
846 len
= xprt
->xpt_ops
->xpo_sendto(rqstp
);
847 mutex_unlock(&xprt
->xpt_mutex
);
848 rpc_wake_up(&xprt
->xpt_bc_pending
);
849 svc_xprt_release(rqstp
);
851 if (len
== -ECONNREFUSED
|| len
== -ENOTCONN
|| len
== -EAGAIN
)
857 * Timer function to close old temporary transports, using
858 * a mark-and-sweep algorithm.
860 static void svc_age_temp_xprts(unsigned long closure
)
862 struct svc_serv
*serv
= (struct svc_serv
*)closure
;
863 struct svc_xprt
*xprt
;
864 struct list_head
*le
, *next
;
866 dprintk("svc_age_temp_xprts\n");
868 if (!spin_trylock_bh(&serv
->sv_lock
)) {
869 /* busy, try again 1 sec later */
870 dprintk("svc_age_temp_xprts: busy\n");
871 mod_timer(&serv
->sv_temptimer
, jiffies
+ HZ
);
875 list_for_each_safe(le
, next
, &serv
->sv_tempsocks
) {
876 xprt
= list_entry(le
, struct svc_xprt
, xpt_list
);
878 /* First time through, just mark it OLD. Second time
879 * through, close it. */
880 if (!test_and_set_bit(XPT_OLD
, &xprt
->xpt_flags
))
882 if (atomic_read(&xprt
->xpt_ref
.refcount
) > 1 ||
883 test_bit(XPT_BUSY
, &xprt
->xpt_flags
))
886 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
887 set_bit(XPT_DETACHED
, &xprt
->xpt_flags
);
888 dprintk("queuing xprt %p for closing\n", xprt
);
890 /* a thread will dequeue and close it soon */
891 svc_xprt_enqueue(xprt
);
893 spin_unlock_bh(&serv
->sv_lock
);
895 mod_timer(&serv
->sv_temptimer
, jiffies
+ svc_conn_age_period
* HZ
);
898 static void call_xpt_users(struct svc_xprt
*xprt
)
900 struct svc_xpt_user
*u
;
902 spin_lock(&xprt
->xpt_lock
);
903 while (!list_empty(&xprt
->xpt_users
)) {
904 u
= list_first_entry(&xprt
->xpt_users
, struct svc_xpt_user
, list
);
908 spin_unlock(&xprt
->xpt_lock
);
912 * Remove a dead transport
914 static void svc_delete_xprt(struct svc_xprt
*xprt
)
916 struct svc_serv
*serv
= xprt
->xpt_server
;
917 struct svc_deferred_req
*dr
;
919 /* Only do this once */
920 if (test_and_set_bit(XPT_DEAD
, &xprt
->xpt_flags
))
923 dprintk("svc: svc_delete_xprt(%p)\n", xprt
);
924 xprt
->xpt_ops
->xpo_detach(xprt
);
926 spin_lock_bh(&serv
->sv_lock
);
927 if (!test_and_set_bit(XPT_DETACHED
, &xprt
->xpt_flags
))
928 list_del_init(&xprt
->xpt_list
);
929 WARN_ON_ONCE(!list_empty(&xprt
->xpt_ready
));
930 if (test_bit(XPT_TEMP
, &xprt
->xpt_flags
))
932 spin_unlock_bh(&serv
->sv_lock
);
934 while ((dr
= svc_deferred_dequeue(xprt
)) != NULL
)
937 call_xpt_users(xprt
);
941 void svc_close_xprt(struct svc_xprt
*xprt
)
943 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
944 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
))
945 /* someone else will have to effect the close */
948 * We expect svc_close_xprt() to work even when no threads are
949 * running (e.g., while configuring the server before starting
950 * any threads), so if the transport isn't busy, we delete
953 svc_delete_xprt(xprt
);
955 EXPORT_SYMBOL_GPL(svc_close_xprt
);
957 static int svc_close_list(struct svc_serv
*serv
, struct list_head
*xprt_list
, struct net
*net
)
959 struct svc_xprt
*xprt
;
962 spin_lock(&serv
->sv_lock
);
963 list_for_each_entry(xprt
, xprt_list
, xpt_list
) {
964 if (xprt
->xpt_net
!= net
)
967 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
968 svc_xprt_enqueue(xprt
);
970 spin_unlock(&serv
->sv_lock
);
974 static struct svc_xprt
*svc_dequeue_net(struct svc_serv
*serv
, struct net
*net
)
976 struct svc_pool
*pool
;
977 struct svc_xprt
*xprt
;
978 struct svc_xprt
*tmp
;
981 for (i
= 0; i
< serv
->sv_nrpools
; i
++) {
982 pool
= &serv
->sv_pools
[i
];
984 spin_lock_bh(&pool
->sp_lock
);
985 list_for_each_entry_safe(xprt
, tmp
, &pool
->sp_sockets
, xpt_ready
) {
986 if (xprt
->xpt_net
!= net
)
988 list_del_init(&xprt
->xpt_ready
);
989 spin_unlock_bh(&pool
->sp_lock
);
992 spin_unlock_bh(&pool
->sp_lock
);
997 static void svc_clean_up_xprts(struct svc_serv
*serv
, struct net
*net
)
999 struct svc_xprt
*xprt
;
1001 while ((xprt
= svc_dequeue_net(serv
, net
))) {
1002 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
1003 svc_delete_xprt(xprt
);
1008 * Server threads may still be running (especially in the case where the
1009 * service is still running in other network namespaces).
1011 * So we shut down sockets the same way we would on a running server, by
1012 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1013 * the close. In the case there are no such other threads,
1014 * threads running, svc_clean_up_xprts() does a simple version of a
1015 * server's main event loop, and in the case where there are other
1016 * threads, we may need to wait a little while and then check again to
1017 * see if they're done.
1019 void svc_close_net(struct svc_serv
*serv
, struct net
*net
)
1023 while (svc_close_list(serv
, &serv
->sv_permsocks
, net
) +
1024 svc_close_list(serv
, &serv
->sv_tempsocks
, net
)) {
1026 svc_clean_up_xprts(serv
, net
);
1032 * Handle defer and revisit of requests
1035 static void svc_revisit(struct cache_deferred_req
*dreq
, int too_many
)
1037 struct svc_deferred_req
*dr
=
1038 container_of(dreq
, struct svc_deferred_req
, handle
);
1039 struct svc_xprt
*xprt
= dr
->xprt
;
1041 spin_lock(&xprt
->xpt_lock
);
1042 set_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1043 if (too_many
|| test_bit(XPT_DEAD
, &xprt
->xpt_flags
)) {
1044 spin_unlock(&xprt
->xpt_lock
);
1045 dprintk("revisit canceled\n");
1050 dprintk("revisit queued\n");
1052 list_add(&dr
->handle
.recent
, &xprt
->xpt_deferred
);
1053 spin_unlock(&xprt
->xpt_lock
);
1054 svc_xprt_enqueue(xprt
);
1059 * Save the request off for later processing. The request buffer looks
1062 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1064 * This code can only handle requests that consist of an xprt-header
1067 static struct cache_deferred_req
*svc_defer(struct cache_req
*req
)
1069 struct svc_rqst
*rqstp
= container_of(req
, struct svc_rqst
, rq_chandle
);
1070 struct svc_deferred_req
*dr
;
1072 if (rqstp
->rq_arg
.page_len
|| !rqstp
->rq_usedeferral
)
1073 return NULL
; /* if more than a page, give up FIXME */
1074 if (rqstp
->rq_deferred
) {
1075 dr
= rqstp
->rq_deferred
;
1076 rqstp
->rq_deferred
= NULL
;
1080 /* FIXME maybe discard if size too large */
1081 size
= sizeof(struct svc_deferred_req
) + rqstp
->rq_arg
.len
;
1082 dr
= kmalloc(size
, GFP_KERNEL
);
1086 dr
->handle
.owner
= rqstp
->rq_server
;
1087 dr
->prot
= rqstp
->rq_prot
;
1088 memcpy(&dr
->addr
, &rqstp
->rq_addr
, rqstp
->rq_addrlen
);
1089 dr
->addrlen
= rqstp
->rq_addrlen
;
1090 dr
->daddr
= rqstp
->rq_daddr
;
1091 dr
->argslen
= rqstp
->rq_arg
.len
>> 2;
1092 dr
->xprt_hlen
= rqstp
->rq_xprt_hlen
;
1094 /* back up head to the start of the buffer and copy */
1095 skip
= rqstp
->rq_arg
.len
- rqstp
->rq_arg
.head
[0].iov_len
;
1096 memcpy(dr
->args
, rqstp
->rq_arg
.head
[0].iov_base
- skip
,
1099 svc_xprt_get(rqstp
->rq_xprt
);
1100 dr
->xprt
= rqstp
->rq_xprt
;
1101 rqstp
->rq_dropme
= true;
1103 dr
->handle
.revisit
= svc_revisit
;
1108 * recv data from a deferred request into an active one
1110 static int svc_deferred_recv(struct svc_rqst
*rqstp
)
1112 struct svc_deferred_req
*dr
= rqstp
->rq_deferred
;
1114 /* setup iov_base past transport header */
1115 rqstp
->rq_arg
.head
[0].iov_base
= dr
->args
+ (dr
->xprt_hlen
>>2);
1116 /* The iov_len does not include the transport header bytes */
1117 rqstp
->rq_arg
.head
[0].iov_len
= (dr
->argslen
<<2) - dr
->xprt_hlen
;
1118 rqstp
->rq_arg
.page_len
= 0;
1119 /* The rq_arg.len includes the transport header bytes */
1120 rqstp
->rq_arg
.len
= dr
->argslen
<<2;
1121 rqstp
->rq_prot
= dr
->prot
;
1122 memcpy(&rqstp
->rq_addr
, &dr
->addr
, dr
->addrlen
);
1123 rqstp
->rq_addrlen
= dr
->addrlen
;
1124 /* Save off transport header len in case we get deferred again */
1125 rqstp
->rq_xprt_hlen
= dr
->xprt_hlen
;
1126 rqstp
->rq_daddr
= dr
->daddr
;
1127 rqstp
->rq_respages
= rqstp
->rq_pages
;
1128 return (dr
->argslen
<<2) - dr
->xprt_hlen
;
1132 static struct svc_deferred_req
*svc_deferred_dequeue(struct svc_xprt
*xprt
)
1134 struct svc_deferred_req
*dr
= NULL
;
1136 if (!test_bit(XPT_DEFERRED
, &xprt
->xpt_flags
))
1138 spin_lock(&xprt
->xpt_lock
);
1139 if (!list_empty(&xprt
->xpt_deferred
)) {
1140 dr
= list_entry(xprt
->xpt_deferred
.next
,
1141 struct svc_deferred_req
,
1143 list_del_init(&dr
->handle
.recent
);
1145 clear_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1146 spin_unlock(&xprt
->xpt_lock
);
1151 * svc_find_xprt - find an RPC transport instance
1152 * @serv: pointer to svc_serv to search
1153 * @xcl_name: C string containing transport's class name
1154 * @net: owner net pointer
1155 * @af: Address family of transport's local address
1156 * @port: transport's IP port number
1158 * Return the transport instance pointer for the endpoint accepting
1159 * connections/peer traffic from the specified transport class,
1160 * address family and port.
1162 * Specifying 0 for the address family or port is effectively a
1163 * wild-card, and will result in matching the first transport in the
1164 * service's list that has a matching class name.
1166 struct svc_xprt
*svc_find_xprt(struct svc_serv
*serv
, const char *xcl_name
,
1167 struct net
*net
, const sa_family_t af
,
1168 const unsigned short port
)
1170 struct svc_xprt
*xprt
;
1171 struct svc_xprt
*found
= NULL
;
1173 /* Sanity check the args */
1174 if (serv
== NULL
|| xcl_name
== NULL
)
1177 spin_lock_bh(&serv
->sv_lock
);
1178 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1179 if (xprt
->xpt_net
!= net
)
1181 if (strcmp(xprt
->xpt_class
->xcl_name
, xcl_name
))
1183 if (af
!= AF_UNSPEC
&& af
!= xprt
->xpt_local
.ss_family
)
1185 if (port
!= 0 && port
!= svc_xprt_local_port(xprt
))
1191 spin_unlock_bh(&serv
->sv_lock
);
1194 EXPORT_SYMBOL_GPL(svc_find_xprt
);
1196 static int svc_one_xprt_name(const struct svc_xprt
*xprt
,
1197 char *pos
, int remaining
)
1201 len
= snprintf(pos
, remaining
, "%s %u\n",
1202 xprt
->xpt_class
->xcl_name
,
1203 svc_xprt_local_port(xprt
));
1204 if (len
>= remaining
)
1205 return -ENAMETOOLONG
;
1210 * svc_xprt_names - format a buffer with a list of transport names
1211 * @serv: pointer to an RPC service
1212 * @buf: pointer to a buffer to be filled in
1213 * @buflen: length of buffer to be filled in
1215 * Fills in @buf with a string containing a list of transport names,
1216 * each name terminated with '\n'.
1218 * Returns positive length of the filled-in string on success; otherwise
1219 * a negative errno value is returned if an error occurs.
1221 int svc_xprt_names(struct svc_serv
*serv
, char *buf
, const int buflen
)
1223 struct svc_xprt
*xprt
;
1227 /* Sanity check args */
1231 spin_lock_bh(&serv
->sv_lock
);
1235 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1236 len
= svc_one_xprt_name(xprt
, pos
, buflen
- totlen
);
1248 spin_unlock_bh(&serv
->sv_lock
);
1251 EXPORT_SYMBOL_GPL(svc_xprt_names
);
1254 /*----------------------------------------------------------------------------*/
1256 static void *svc_pool_stats_start(struct seq_file
*m
, loff_t
*pos
)
1258 unsigned int pidx
= (unsigned int)*pos
;
1259 struct svc_serv
*serv
= m
->private;
1261 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx
);
1264 return SEQ_START_TOKEN
;
1265 return (pidx
> serv
->sv_nrpools
? NULL
: &serv
->sv_pools
[pidx
-1]);
1268 static void *svc_pool_stats_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
1270 struct svc_pool
*pool
= p
;
1271 struct svc_serv
*serv
= m
->private;
1273 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos
);
1275 if (p
== SEQ_START_TOKEN
) {
1276 pool
= &serv
->sv_pools
[0];
1278 unsigned int pidx
= (pool
- &serv
->sv_pools
[0]);
1279 if (pidx
< serv
->sv_nrpools
-1)
1280 pool
= &serv
->sv_pools
[pidx
+1];
1288 static void svc_pool_stats_stop(struct seq_file
*m
, void *p
)
1292 static int svc_pool_stats_show(struct seq_file
*m
, void *p
)
1294 struct svc_pool
*pool
= p
;
1296 if (p
== SEQ_START_TOKEN
) {
1297 seq_puts(m
, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1301 seq_printf(m
, "%u %lu %lu %lu %lu\n",
1303 pool
->sp_stats
.packets
,
1304 pool
->sp_stats
.sockets_queued
,
1305 pool
->sp_stats
.threads_woken
,
1306 pool
->sp_stats
.threads_timedout
);
1311 static const struct seq_operations svc_pool_stats_seq_ops
= {
1312 .start
= svc_pool_stats_start
,
1313 .next
= svc_pool_stats_next
,
1314 .stop
= svc_pool_stats_stop
,
1315 .show
= svc_pool_stats_show
,
1318 int svc_pool_stats_open(struct svc_serv
*serv
, struct file
*file
)
1322 err
= seq_open(file
, &svc_pool_stats_seq_ops
);
1324 ((struct seq_file
*) file
->private_data
)->private = serv
;
1327 EXPORT_SYMBOL(svc_pool_stats_open
);
1329 /*----------------------------------------------------------------------------*/