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 if (!test_bit(XPT_BUSY
, &xprt
->xpt_flags
)) {
224 WARN_ONCE(1, "xprt=0x%p already busy!", xprt
);
228 /* As soon as we clear busy, the xprt could be closed and
229 * 'put', so we need a reference to call svc_xprt_do_enqueue with:
232 smp_mb__before_atomic();
233 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
234 svc_xprt_do_enqueue(xprt
);
238 void svc_add_new_perm_xprt(struct svc_serv
*serv
, struct svc_xprt
*new)
240 clear_bit(XPT_TEMP
, &new->xpt_flags
);
241 spin_lock_bh(&serv
->sv_lock
);
242 list_add(&new->xpt_list
, &serv
->sv_permsocks
);
243 spin_unlock_bh(&serv
->sv_lock
);
244 svc_xprt_received(new);
247 int svc_create_xprt(struct svc_serv
*serv
, const char *xprt_name
,
248 struct net
*net
, const int family
,
249 const unsigned short port
, int flags
)
251 struct svc_xprt_class
*xcl
;
253 dprintk("svc: creating transport %s[%d]\n", xprt_name
, port
);
254 spin_lock(&svc_xprt_class_lock
);
255 list_for_each_entry(xcl
, &svc_xprt_class_list
, xcl_list
) {
256 struct svc_xprt
*newxprt
;
257 unsigned short newport
;
259 if (strcmp(xprt_name
, xcl
->xcl_name
))
262 if (!try_module_get(xcl
->xcl_owner
))
265 spin_unlock(&svc_xprt_class_lock
);
266 newxprt
= __svc_xpo_create(xcl
, serv
, net
, family
, port
, flags
);
267 if (IS_ERR(newxprt
)) {
268 module_put(xcl
->xcl_owner
);
269 return PTR_ERR(newxprt
);
271 svc_add_new_perm_xprt(serv
, newxprt
);
272 newport
= svc_xprt_local_port(newxprt
);
276 spin_unlock(&svc_xprt_class_lock
);
277 dprintk("svc: transport %s not found\n", xprt_name
);
279 /* This errno is exposed to user space. Provide a reasonable
280 * perror msg for a bad transport. */
281 return -EPROTONOSUPPORT
;
283 EXPORT_SYMBOL_GPL(svc_create_xprt
);
286 * Copy the local and remote xprt addresses to the rqstp structure
288 void svc_xprt_copy_addrs(struct svc_rqst
*rqstp
, struct svc_xprt
*xprt
)
290 memcpy(&rqstp
->rq_addr
, &xprt
->xpt_remote
, xprt
->xpt_remotelen
);
291 rqstp
->rq_addrlen
= xprt
->xpt_remotelen
;
294 * Destination address in request is needed for binding the
295 * source address in RPC replies/callbacks later.
297 memcpy(&rqstp
->rq_daddr
, &xprt
->xpt_local
, xprt
->xpt_locallen
);
298 rqstp
->rq_daddrlen
= xprt
->xpt_locallen
;
300 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs
);
303 * svc_print_addr - Format rq_addr field for printing
304 * @rqstp: svc_rqst struct containing address to print
305 * @buf: target buffer for formatted address
306 * @len: length of target buffer
309 char *svc_print_addr(struct svc_rqst
*rqstp
, char *buf
, size_t len
)
311 return __svc_print_addr(svc_addr(rqstp
), buf
, len
);
313 EXPORT_SYMBOL_GPL(svc_print_addr
);
315 static bool svc_xprt_has_something_to_do(struct svc_xprt
*xprt
)
317 if (xprt
->xpt_flags
& ((1<<XPT_CONN
)|(1<<XPT_CLOSE
)))
319 if (xprt
->xpt_flags
& ((1<<XPT_DATA
)|(1<<XPT_DEFERRED
)))
320 return xprt
->xpt_ops
->xpo_has_wspace(xprt
);
324 static void svc_xprt_do_enqueue(struct svc_xprt
*xprt
)
326 struct svc_pool
*pool
;
327 struct svc_rqst
*rqstp
= NULL
;
331 if (!svc_xprt_has_something_to_do(xprt
))
334 /* Mark transport as busy. It will remain in this state until
335 * the provider calls svc_xprt_received. We update XPT_BUSY
336 * atomically because it also guards against trying to enqueue
337 * the transport twice.
339 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
)) {
340 /* Don't enqueue transport while already enqueued */
341 dprintk("svc: transport %p busy, not enqueued\n", xprt
);
346 pool
= svc_pool_for_cpu(xprt
->xpt_server
, cpu
);
348 atomic_long_inc(&pool
->sp_stats
.packets
);
351 /* find a thread for this xprt */
353 list_for_each_entry_rcu(rqstp
, &pool
->sp_all_threads
, rq_all
) {
354 /* Do a lockless check first */
355 if (test_bit(RQ_BUSY
, &rqstp
->rq_flags
))
359 * Once the xprt has been queued, it can only be dequeued by
360 * the task that intends to service it. All we can do at that
361 * point is to try to wake this thread back up so that it can
365 spin_lock_bh(&rqstp
->rq_lock
);
366 if (test_and_set_bit(RQ_BUSY
, &rqstp
->rq_flags
)) {
367 /* already busy, move on... */
368 spin_unlock_bh(&rqstp
->rq_lock
);
372 /* this one will do */
373 rqstp
->rq_xprt
= xprt
;
375 spin_unlock_bh(&rqstp
->rq_lock
);
379 atomic_long_inc(&pool
->sp_stats
.threads_woken
);
380 wake_up_process(rqstp
->rq_task
);
387 * We didn't find an idle thread to use, so we need to queue the xprt.
388 * Do so and then search again. If we find one, we can't hook this one
389 * up to it directly but we can wake the thread up in the hopes that it
390 * will pick it up once it searches for a xprt to service.
394 dprintk("svc: transport %p put into queue\n", xprt
);
395 spin_lock_bh(&pool
->sp_lock
);
396 list_add_tail(&xprt
->xpt_ready
, &pool
->sp_sockets
);
397 pool
->sp_stats
.sockets_queued
++;
398 spin_unlock_bh(&pool
->sp_lock
);
404 trace_svc_xprt_do_enqueue(xprt
, rqstp
);
408 * Queue up a transport with data pending. If there are idle nfsd
409 * processes, wake 'em up.
412 void svc_xprt_enqueue(struct svc_xprt
*xprt
)
414 if (test_bit(XPT_BUSY
, &xprt
->xpt_flags
))
416 svc_xprt_do_enqueue(xprt
);
418 EXPORT_SYMBOL_GPL(svc_xprt_enqueue
);
421 * Dequeue the first transport, if there is one.
423 static struct svc_xprt
*svc_xprt_dequeue(struct svc_pool
*pool
)
425 struct svc_xprt
*xprt
= NULL
;
427 if (list_empty(&pool
->sp_sockets
))
430 spin_lock_bh(&pool
->sp_lock
);
431 if (likely(!list_empty(&pool
->sp_sockets
))) {
432 xprt
= list_first_entry(&pool
->sp_sockets
,
433 struct svc_xprt
, xpt_ready
);
434 list_del_init(&xprt
->xpt_ready
);
437 dprintk("svc: transport %p dequeued, inuse=%d\n",
438 xprt
, atomic_read(&xprt
->xpt_ref
.refcount
));
440 spin_unlock_bh(&pool
->sp_lock
);
442 trace_svc_xprt_dequeue(xprt
);
447 * svc_reserve - change the space reserved for the reply to a request.
448 * @rqstp: The request in question
449 * @space: new max space to reserve
451 * Each request reserves some space on the output queue of the transport
452 * to make sure the reply fits. This function reduces that reserved
453 * space to be the amount of space used already, plus @space.
456 void svc_reserve(struct svc_rqst
*rqstp
, int space
)
458 space
+= rqstp
->rq_res
.head
[0].iov_len
;
460 if (space
< rqstp
->rq_reserved
) {
461 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
462 atomic_sub((rqstp
->rq_reserved
- space
), &xprt
->xpt_reserved
);
463 rqstp
->rq_reserved
= space
;
465 if (xprt
->xpt_ops
->xpo_adjust_wspace
)
466 xprt
->xpt_ops
->xpo_adjust_wspace(xprt
);
467 svc_xprt_enqueue(xprt
);
470 EXPORT_SYMBOL_GPL(svc_reserve
);
472 static void svc_xprt_release(struct svc_rqst
*rqstp
)
474 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
476 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
478 kfree(rqstp
->rq_deferred
);
479 rqstp
->rq_deferred
= NULL
;
481 svc_free_res_pages(rqstp
);
482 rqstp
->rq_res
.page_len
= 0;
483 rqstp
->rq_res
.page_base
= 0;
485 /* Reset response buffer and release
487 * But first, check that enough space was reserved
488 * for the reply, otherwise we have a bug!
490 if ((rqstp
->rq_res
.len
) > rqstp
->rq_reserved
)
491 printk(KERN_ERR
"RPC request reserved %d but used %d\n",
495 rqstp
->rq_res
.head
[0].iov_len
= 0;
496 svc_reserve(rqstp
, 0);
497 rqstp
->rq_xprt
= NULL
;
503 * Some svc_serv's will have occasional work to do, even when a xprt is not
504 * waiting to be serviced. This function is there to "kick" a task in one of
505 * those services so that it can wake up and do that work. Note that we only
506 * bother with pool 0 as we don't need to wake up more than one thread for
509 void svc_wake_up(struct svc_serv
*serv
)
511 struct svc_rqst
*rqstp
;
512 struct svc_pool
*pool
;
514 pool
= &serv
->sv_pools
[0];
517 list_for_each_entry_rcu(rqstp
, &pool
->sp_all_threads
, rq_all
) {
518 /* skip any that aren't queued */
519 if (test_bit(RQ_BUSY
, &rqstp
->rq_flags
))
522 dprintk("svc: daemon %p woken up.\n", rqstp
);
523 wake_up_process(rqstp
->rq_task
);
524 trace_svc_wake_up(rqstp
->rq_task
->pid
);
529 /* No free entries available */
530 set_bit(SP_TASK_PENDING
, &pool
->sp_flags
);
532 trace_svc_wake_up(0);
534 EXPORT_SYMBOL_GPL(svc_wake_up
);
536 int svc_port_is_privileged(struct sockaddr
*sin
)
538 switch (sin
->sa_family
) {
540 return ntohs(((struct sockaddr_in
*)sin
)->sin_port
)
543 return ntohs(((struct sockaddr_in6
*)sin
)->sin6_port
)
551 * Make sure that we don't have too many active connections. If we have,
552 * something must be dropped. It's not clear what will happen if we allow
553 * "too many" connections, but when dealing with network-facing software,
554 * we have to code defensively. Here we do that by imposing hard limits.
556 * There's no point in trying to do random drop here for DoS
557 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
558 * attacker can easily beat that.
560 * The only somewhat efficient mechanism would be if drop old
561 * connections from the same IP first. But right now we don't even
562 * record the client IP in svc_sock.
564 * single-threaded services that expect a lot of clients will probably
565 * need to set sv_maxconn to override the default value which is based
566 * on the number of threads
568 static void svc_check_conn_limits(struct svc_serv
*serv
)
570 unsigned int limit
= serv
->sv_maxconn
? serv
->sv_maxconn
:
571 (serv
->sv_nrthreads
+3) * 20;
573 if (serv
->sv_tmpcnt
> limit
) {
574 struct svc_xprt
*xprt
= NULL
;
575 spin_lock_bh(&serv
->sv_lock
);
576 if (!list_empty(&serv
->sv_tempsocks
)) {
577 /* Try to help the admin */
578 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
579 serv
->sv_name
, serv
->sv_maxconn
?
580 "max number of connections" :
581 "number of threads");
583 * Always select the oldest connection. It's not fair,
586 xprt
= list_entry(serv
->sv_tempsocks
.prev
,
589 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
592 spin_unlock_bh(&serv
->sv_lock
);
595 svc_xprt_enqueue(xprt
);
601 static int svc_alloc_arg(struct svc_rqst
*rqstp
)
603 struct svc_serv
*serv
= rqstp
->rq_server
;
608 /* now allocate needed pages. If we get a failure, sleep briefly */
609 pages
= (serv
->sv_max_mesg
+ PAGE_SIZE
) / PAGE_SIZE
;
610 WARN_ON_ONCE(pages
>= RPCSVC_MAXPAGES
);
611 if (pages
>= RPCSVC_MAXPAGES
)
612 /* use as many pages as possible */
613 pages
= RPCSVC_MAXPAGES
- 1;
614 for (i
= 0; i
< pages
; i
++)
615 while (rqstp
->rq_pages
[i
] == NULL
) {
616 struct page
*p
= alloc_page(GFP_KERNEL
);
618 set_current_state(TASK_INTERRUPTIBLE
);
619 if (signalled() || kthread_should_stop()) {
620 set_current_state(TASK_RUNNING
);
623 schedule_timeout(msecs_to_jiffies(500));
625 rqstp
->rq_pages
[i
] = p
;
627 rqstp
->rq_page_end
= &rqstp
->rq_pages
[i
];
628 rqstp
->rq_pages
[i
++] = NULL
; /* this might be seen in nfs_read_actor */
630 /* Make arg->head point to first page and arg->pages point to rest */
631 arg
= &rqstp
->rq_arg
;
632 arg
->head
[0].iov_base
= page_address(rqstp
->rq_pages
[0]);
633 arg
->head
[0].iov_len
= PAGE_SIZE
;
634 arg
->pages
= rqstp
->rq_pages
+ 1;
636 /* save at least one page for response */
637 arg
->page_len
= (pages
-2)*PAGE_SIZE
;
638 arg
->len
= (pages
-1)*PAGE_SIZE
;
639 arg
->tail
[0].iov_len
= 0;
644 rqst_should_sleep(struct svc_rqst
*rqstp
)
646 struct svc_pool
*pool
= rqstp
->rq_pool
;
648 /* did someone call svc_wake_up? */
649 if (test_and_clear_bit(SP_TASK_PENDING
, &pool
->sp_flags
))
652 /* was a socket queued? */
653 if (!list_empty(&pool
->sp_sockets
))
656 /* are we shutting down? */
657 if (signalled() || kthread_should_stop())
660 /* are we freezing? */
661 if (freezing(current
))
667 static struct svc_xprt
*svc_get_next_xprt(struct svc_rqst
*rqstp
, long timeout
)
669 struct svc_xprt
*xprt
;
670 struct svc_pool
*pool
= rqstp
->rq_pool
;
673 /* rq_xprt should be clear on entry */
674 WARN_ON_ONCE(rqstp
->rq_xprt
);
676 /* Normally we will wait up to 5 seconds for any required
677 * cache information to be provided.
679 rqstp
->rq_chandle
.thread_wait
= 5*HZ
;
681 xprt
= svc_xprt_dequeue(pool
);
683 rqstp
->rq_xprt
= xprt
;
685 /* As there is a shortage of threads and this request
686 * had to be queued, don't allow the thread to wait so
687 * long for cache updates.
689 rqstp
->rq_chandle
.thread_wait
= 1*HZ
;
690 clear_bit(SP_TASK_PENDING
, &pool
->sp_flags
);
695 * We have to be able to interrupt this wait
696 * to bring down the daemons ...
698 set_current_state(TASK_INTERRUPTIBLE
);
699 clear_bit(RQ_BUSY
, &rqstp
->rq_flags
);
702 if (likely(rqst_should_sleep(rqstp
)))
703 time_left
= schedule_timeout(timeout
);
705 __set_current_state(TASK_RUNNING
);
709 spin_lock_bh(&rqstp
->rq_lock
);
710 set_bit(RQ_BUSY
, &rqstp
->rq_flags
);
711 spin_unlock_bh(&rqstp
->rq_lock
);
713 xprt
= rqstp
->rq_xprt
;
718 atomic_long_inc(&pool
->sp_stats
.threads_timedout
);
720 if (signalled() || kthread_should_stop())
721 return ERR_PTR(-EINTR
);
722 return ERR_PTR(-EAGAIN
);
725 static void svc_add_new_temp_xprt(struct svc_serv
*serv
, struct svc_xprt
*newxpt
)
727 spin_lock_bh(&serv
->sv_lock
);
728 set_bit(XPT_TEMP
, &newxpt
->xpt_flags
);
729 list_add(&newxpt
->xpt_list
, &serv
->sv_tempsocks
);
731 if (serv
->sv_temptimer
.function
== NULL
) {
732 /* setup timer to age temp transports */
733 setup_timer(&serv
->sv_temptimer
, svc_age_temp_xprts
,
734 (unsigned long)serv
);
735 mod_timer(&serv
->sv_temptimer
,
736 jiffies
+ svc_conn_age_period
* HZ
);
738 spin_unlock_bh(&serv
->sv_lock
);
739 svc_xprt_received(newxpt
);
742 static int svc_handle_xprt(struct svc_rqst
*rqstp
, struct svc_xprt
*xprt
)
744 struct svc_serv
*serv
= rqstp
->rq_server
;
747 if (test_bit(XPT_CLOSE
, &xprt
->xpt_flags
)) {
748 dprintk("svc_recv: found XPT_CLOSE\n");
749 svc_delete_xprt(xprt
);
750 /* Leave XPT_BUSY set on the dead xprt: */
753 if (test_bit(XPT_LISTENER
, &xprt
->xpt_flags
)) {
754 struct svc_xprt
*newxpt
;
756 * We know this module_get will succeed because the
757 * listener holds a reference too
759 __module_get(xprt
->xpt_class
->xcl_owner
);
760 svc_check_conn_limits(xprt
->xpt_server
);
761 newxpt
= xprt
->xpt_ops
->xpo_accept(xprt
);
763 svc_add_new_temp_xprt(serv
, newxpt
);
765 module_put(xprt
->xpt_class
->xcl_owner
);
767 /* XPT_DATA|XPT_DEFERRED case: */
768 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
769 rqstp
, rqstp
->rq_pool
->sp_id
, xprt
,
770 atomic_read(&xprt
->xpt_ref
.refcount
));
771 rqstp
->rq_deferred
= svc_deferred_dequeue(xprt
);
772 if (rqstp
->rq_deferred
)
773 len
= svc_deferred_recv(rqstp
);
775 len
= xprt
->xpt_ops
->xpo_recvfrom(rqstp
);
776 dprintk("svc: got len=%d\n", len
);
777 rqstp
->rq_reserved
= serv
->sv_max_mesg
;
778 atomic_add(rqstp
->rq_reserved
, &xprt
->xpt_reserved
);
780 /* clear XPT_BUSY: */
781 svc_xprt_received(xprt
);
783 trace_svc_handle_xprt(xprt
, len
);
788 * Receive the next request on any transport. This code is carefully
789 * organised not to touch any cachelines in the shared svc_serv
790 * structure, only cachelines in the local svc_pool.
792 int svc_recv(struct svc_rqst
*rqstp
, long timeout
)
794 struct svc_xprt
*xprt
= NULL
;
795 struct svc_serv
*serv
= rqstp
->rq_server
;
798 dprintk("svc: server %p waiting for data (to = %ld)\n",
803 "svc_recv: service %p, transport not NULL!\n",
806 err
= svc_alloc_arg(rqstp
);
813 if (signalled() || kthread_should_stop())
816 xprt
= svc_get_next_xprt(rqstp
, timeout
);
822 len
= svc_handle_xprt(rqstp
, xprt
);
824 /* No data, incomplete (TCP) read, or accept() */
829 clear_bit(XPT_OLD
, &xprt
->xpt_flags
);
831 if (xprt
->xpt_ops
->xpo_secure_port(rqstp
))
832 set_bit(RQ_SECURE
, &rqstp
->rq_flags
);
834 clear_bit(RQ_SECURE
, &rqstp
->rq_flags
);
835 rqstp
->rq_chandle
.defer
= svc_defer
;
836 rqstp
->rq_xid
= svc_getu32(&rqstp
->rq_arg
.head
[0]);
839 serv
->sv_stats
->netcnt
++;
840 trace_svc_recv(rqstp
, len
);
843 rqstp
->rq_res
.len
= 0;
844 svc_xprt_release(rqstp
);
846 trace_svc_recv(rqstp
, err
);
849 EXPORT_SYMBOL_GPL(svc_recv
);
854 void svc_drop(struct svc_rqst
*rqstp
)
856 dprintk("svc: xprt %p dropped request\n", rqstp
->rq_xprt
);
857 svc_xprt_release(rqstp
);
859 EXPORT_SYMBOL_GPL(svc_drop
);
862 * Return reply to client.
864 int svc_send(struct svc_rqst
*rqstp
)
866 struct svc_xprt
*xprt
;
870 xprt
= rqstp
->rq_xprt
;
874 /* release the receive skb before sending the reply */
875 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
877 /* calculate over-all length */
879 xb
->len
= xb
->head
[0].iov_len
+
883 /* Grab mutex to serialize outgoing data. */
884 mutex_lock(&xprt
->xpt_mutex
);
885 if (test_bit(XPT_DEAD
, &xprt
->xpt_flags
)
886 || test_bit(XPT_CLOSE
, &xprt
->xpt_flags
))
889 len
= xprt
->xpt_ops
->xpo_sendto(rqstp
);
890 mutex_unlock(&xprt
->xpt_mutex
);
891 rpc_wake_up(&xprt
->xpt_bc_pending
);
892 svc_xprt_release(rqstp
);
894 if (len
== -ECONNREFUSED
|| len
== -ENOTCONN
|| len
== -EAGAIN
)
897 trace_svc_send(rqstp
, len
);
902 * Timer function to close old temporary transports, using
903 * a mark-and-sweep algorithm.
905 static void svc_age_temp_xprts(unsigned long closure
)
907 struct svc_serv
*serv
= (struct svc_serv
*)closure
;
908 struct svc_xprt
*xprt
;
909 struct list_head
*le
, *next
;
911 dprintk("svc_age_temp_xprts\n");
913 if (!spin_trylock_bh(&serv
->sv_lock
)) {
914 /* busy, try again 1 sec later */
915 dprintk("svc_age_temp_xprts: busy\n");
916 mod_timer(&serv
->sv_temptimer
, jiffies
+ HZ
);
920 list_for_each_safe(le
, next
, &serv
->sv_tempsocks
) {
921 xprt
= list_entry(le
, struct svc_xprt
, xpt_list
);
923 /* First time through, just mark it OLD. Second time
924 * through, close it. */
925 if (!test_and_set_bit(XPT_OLD
, &xprt
->xpt_flags
))
927 if (atomic_read(&xprt
->xpt_ref
.refcount
) > 1 ||
928 test_bit(XPT_BUSY
, &xprt
->xpt_flags
))
931 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
932 dprintk("queuing xprt %p for closing\n", xprt
);
934 /* a thread will dequeue and close it soon */
935 svc_xprt_enqueue(xprt
);
937 spin_unlock_bh(&serv
->sv_lock
);
939 mod_timer(&serv
->sv_temptimer
, jiffies
+ svc_conn_age_period
* HZ
);
942 static void call_xpt_users(struct svc_xprt
*xprt
)
944 struct svc_xpt_user
*u
;
946 spin_lock(&xprt
->xpt_lock
);
947 while (!list_empty(&xprt
->xpt_users
)) {
948 u
= list_first_entry(&xprt
->xpt_users
, struct svc_xpt_user
, list
);
952 spin_unlock(&xprt
->xpt_lock
);
956 * Remove a dead transport
958 static void svc_delete_xprt(struct svc_xprt
*xprt
)
960 struct svc_serv
*serv
= xprt
->xpt_server
;
961 struct svc_deferred_req
*dr
;
963 /* Only do this once */
964 if (test_and_set_bit(XPT_DEAD
, &xprt
->xpt_flags
))
967 dprintk("svc: svc_delete_xprt(%p)\n", xprt
);
968 xprt
->xpt_ops
->xpo_detach(xprt
);
970 spin_lock_bh(&serv
->sv_lock
);
971 list_del_init(&xprt
->xpt_list
);
972 WARN_ON_ONCE(!list_empty(&xprt
->xpt_ready
));
973 if (test_bit(XPT_TEMP
, &xprt
->xpt_flags
))
975 spin_unlock_bh(&serv
->sv_lock
);
977 while ((dr
= svc_deferred_dequeue(xprt
)) != NULL
)
980 call_xpt_users(xprt
);
984 void svc_close_xprt(struct svc_xprt
*xprt
)
986 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
987 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
))
988 /* someone else will have to effect the close */
991 * We expect svc_close_xprt() to work even when no threads are
992 * running (e.g., while configuring the server before starting
993 * any threads), so if the transport isn't busy, we delete
996 svc_delete_xprt(xprt
);
998 EXPORT_SYMBOL_GPL(svc_close_xprt
);
1000 static int svc_close_list(struct svc_serv
*serv
, struct list_head
*xprt_list
, struct net
*net
)
1002 struct svc_xprt
*xprt
;
1005 spin_lock(&serv
->sv_lock
);
1006 list_for_each_entry(xprt
, xprt_list
, xpt_list
) {
1007 if (xprt
->xpt_net
!= net
)
1010 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
1011 svc_xprt_enqueue(xprt
);
1013 spin_unlock(&serv
->sv_lock
);
1017 static struct svc_xprt
*svc_dequeue_net(struct svc_serv
*serv
, struct net
*net
)
1019 struct svc_pool
*pool
;
1020 struct svc_xprt
*xprt
;
1021 struct svc_xprt
*tmp
;
1024 for (i
= 0; i
< serv
->sv_nrpools
; i
++) {
1025 pool
= &serv
->sv_pools
[i
];
1027 spin_lock_bh(&pool
->sp_lock
);
1028 list_for_each_entry_safe(xprt
, tmp
, &pool
->sp_sockets
, xpt_ready
) {
1029 if (xprt
->xpt_net
!= net
)
1031 list_del_init(&xprt
->xpt_ready
);
1032 spin_unlock_bh(&pool
->sp_lock
);
1035 spin_unlock_bh(&pool
->sp_lock
);
1040 static void svc_clean_up_xprts(struct svc_serv
*serv
, struct net
*net
)
1042 struct svc_xprt
*xprt
;
1044 while ((xprt
= svc_dequeue_net(serv
, net
))) {
1045 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
1046 svc_delete_xprt(xprt
);
1051 * Server threads may still be running (especially in the case where the
1052 * service is still running in other network namespaces).
1054 * So we shut down sockets the same way we would on a running server, by
1055 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1056 * the close. In the case there are no such other threads,
1057 * threads running, svc_clean_up_xprts() does a simple version of a
1058 * server's main event loop, and in the case where there are other
1059 * threads, we may need to wait a little while and then check again to
1060 * see if they're done.
1062 void svc_close_net(struct svc_serv
*serv
, struct net
*net
)
1066 while (svc_close_list(serv
, &serv
->sv_permsocks
, net
) +
1067 svc_close_list(serv
, &serv
->sv_tempsocks
, net
)) {
1069 svc_clean_up_xprts(serv
, net
);
1075 * Handle defer and revisit of requests
1078 static void svc_revisit(struct cache_deferred_req
*dreq
, int too_many
)
1080 struct svc_deferred_req
*dr
=
1081 container_of(dreq
, struct svc_deferred_req
, handle
);
1082 struct svc_xprt
*xprt
= dr
->xprt
;
1084 spin_lock(&xprt
->xpt_lock
);
1085 set_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1086 if (too_many
|| test_bit(XPT_DEAD
, &xprt
->xpt_flags
)) {
1087 spin_unlock(&xprt
->xpt_lock
);
1088 dprintk("revisit canceled\n");
1093 dprintk("revisit queued\n");
1095 list_add(&dr
->handle
.recent
, &xprt
->xpt_deferred
);
1096 spin_unlock(&xprt
->xpt_lock
);
1097 svc_xprt_enqueue(xprt
);
1102 * Save the request off for later processing. The request buffer looks
1105 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1107 * This code can only handle requests that consist of an xprt-header
1110 static struct cache_deferred_req
*svc_defer(struct cache_req
*req
)
1112 struct svc_rqst
*rqstp
= container_of(req
, struct svc_rqst
, rq_chandle
);
1113 struct svc_deferred_req
*dr
;
1115 if (rqstp
->rq_arg
.page_len
|| !test_bit(RQ_USEDEFERRAL
, &rqstp
->rq_flags
))
1116 return NULL
; /* if more than a page, give up FIXME */
1117 if (rqstp
->rq_deferred
) {
1118 dr
= rqstp
->rq_deferred
;
1119 rqstp
->rq_deferred
= NULL
;
1123 /* FIXME maybe discard if size too large */
1124 size
= sizeof(struct svc_deferred_req
) + rqstp
->rq_arg
.len
;
1125 dr
= kmalloc(size
, GFP_KERNEL
);
1129 dr
->handle
.owner
= rqstp
->rq_server
;
1130 dr
->prot
= rqstp
->rq_prot
;
1131 memcpy(&dr
->addr
, &rqstp
->rq_addr
, rqstp
->rq_addrlen
);
1132 dr
->addrlen
= rqstp
->rq_addrlen
;
1133 dr
->daddr
= rqstp
->rq_daddr
;
1134 dr
->argslen
= rqstp
->rq_arg
.len
>> 2;
1135 dr
->xprt_hlen
= rqstp
->rq_xprt_hlen
;
1137 /* back up head to the start of the buffer and copy */
1138 skip
= rqstp
->rq_arg
.len
- rqstp
->rq_arg
.head
[0].iov_len
;
1139 memcpy(dr
->args
, rqstp
->rq_arg
.head
[0].iov_base
- skip
,
1142 svc_xprt_get(rqstp
->rq_xprt
);
1143 dr
->xprt
= rqstp
->rq_xprt
;
1144 set_bit(RQ_DROPME
, &rqstp
->rq_flags
);
1146 dr
->handle
.revisit
= svc_revisit
;
1151 * recv data from a deferred request into an active one
1153 static int svc_deferred_recv(struct svc_rqst
*rqstp
)
1155 struct svc_deferred_req
*dr
= rqstp
->rq_deferred
;
1157 /* setup iov_base past transport header */
1158 rqstp
->rq_arg
.head
[0].iov_base
= dr
->args
+ (dr
->xprt_hlen
>>2);
1159 /* The iov_len does not include the transport header bytes */
1160 rqstp
->rq_arg
.head
[0].iov_len
= (dr
->argslen
<<2) - dr
->xprt_hlen
;
1161 rqstp
->rq_arg
.page_len
= 0;
1162 /* The rq_arg.len includes the transport header bytes */
1163 rqstp
->rq_arg
.len
= dr
->argslen
<<2;
1164 rqstp
->rq_prot
= dr
->prot
;
1165 memcpy(&rqstp
->rq_addr
, &dr
->addr
, dr
->addrlen
);
1166 rqstp
->rq_addrlen
= dr
->addrlen
;
1167 /* Save off transport header len in case we get deferred again */
1168 rqstp
->rq_xprt_hlen
= dr
->xprt_hlen
;
1169 rqstp
->rq_daddr
= dr
->daddr
;
1170 rqstp
->rq_respages
= rqstp
->rq_pages
;
1171 return (dr
->argslen
<<2) - dr
->xprt_hlen
;
1175 static struct svc_deferred_req
*svc_deferred_dequeue(struct svc_xprt
*xprt
)
1177 struct svc_deferred_req
*dr
= NULL
;
1179 if (!test_bit(XPT_DEFERRED
, &xprt
->xpt_flags
))
1181 spin_lock(&xprt
->xpt_lock
);
1182 if (!list_empty(&xprt
->xpt_deferred
)) {
1183 dr
= list_entry(xprt
->xpt_deferred
.next
,
1184 struct svc_deferred_req
,
1186 list_del_init(&dr
->handle
.recent
);
1188 clear_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1189 spin_unlock(&xprt
->xpt_lock
);
1194 * svc_find_xprt - find an RPC transport instance
1195 * @serv: pointer to svc_serv to search
1196 * @xcl_name: C string containing transport's class name
1197 * @net: owner net pointer
1198 * @af: Address family of transport's local address
1199 * @port: transport's IP port number
1201 * Return the transport instance pointer for the endpoint accepting
1202 * connections/peer traffic from the specified transport class,
1203 * address family and port.
1205 * Specifying 0 for the address family or port is effectively a
1206 * wild-card, and will result in matching the first transport in the
1207 * service's list that has a matching class name.
1209 struct svc_xprt
*svc_find_xprt(struct svc_serv
*serv
, const char *xcl_name
,
1210 struct net
*net
, const sa_family_t af
,
1211 const unsigned short port
)
1213 struct svc_xprt
*xprt
;
1214 struct svc_xprt
*found
= NULL
;
1216 /* Sanity check the args */
1217 if (serv
== NULL
|| xcl_name
== NULL
)
1220 spin_lock_bh(&serv
->sv_lock
);
1221 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1222 if (xprt
->xpt_net
!= net
)
1224 if (strcmp(xprt
->xpt_class
->xcl_name
, xcl_name
))
1226 if (af
!= AF_UNSPEC
&& af
!= xprt
->xpt_local
.ss_family
)
1228 if (port
!= 0 && port
!= svc_xprt_local_port(xprt
))
1234 spin_unlock_bh(&serv
->sv_lock
);
1237 EXPORT_SYMBOL_GPL(svc_find_xprt
);
1239 static int svc_one_xprt_name(const struct svc_xprt
*xprt
,
1240 char *pos
, int remaining
)
1244 len
= snprintf(pos
, remaining
, "%s %u\n",
1245 xprt
->xpt_class
->xcl_name
,
1246 svc_xprt_local_port(xprt
));
1247 if (len
>= remaining
)
1248 return -ENAMETOOLONG
;
1253 * svc_xprt_names - format a buffer with a list of transport names
1254 * @serv: pointer to an RPC service
1255 * @buf: pointer to a buffer to be filled in
1256 * @buflen: length of buffer to be filled in
1258 * Fills in @buf with a string containing a list of transport names,
1259 * each name terminated with '\n'.
1261 * Returns positive length of the filled-in string on success; otherwise
1262 * a negative errno value is returned if an error occurs.
1264 int svc_xprt_names(struct svc_serv
*serv
, char *buf
, const int buflen
)
1266 struct svc_xprt
*xprt
;
1270 /* Sanity check args */
1274 spin_lock_bh(&serv
->sv_lock
);
1278 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1279 len
= svc_one_xprt_name(xprt
, pos
, buflen
- totlen
);
1291 spin_unlock_bh(&serv
->sv_lock
);
1294 EXPORT_SYMBOL_GPL(svc_xprt_names
);
1297 /*----------------------------------------------------------------------------*/
1299 static void *svc_pool_stats_start(struct seq_file
*m
, loff_t
*pos
)
1301 unsigned int pidx
= (unsigned int)*pos
;
1302 struct svc_serv
*serv
= m
->private;
1304 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx
);
1307 return SEQ_START_TOKEN
;
1308 return (pidx
> serv
->sv_nrpools
? NULL
: &serv
->sv_pools
[pidx
-1]);
1311 static void *svc_pool_stats_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
1313 struct svc_pool
*pool
= p
;
1314 struct svc_serv
*serv
= m
->private;
1316 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos
);
1318 if (p
== SEQ_START_TOKEN
) {
1319 pool
= &serv
->sv_pools
[0];
1321 unsigned int pidx
= (pool
- &serv
->sv_pools
[0]);
1322 if (pidx
< serv
->sv_nrpools
-1)
1323 pool
= &serv
->sv_pools
[pidx
+1];
1331 static void svc_pool_stats_stop(struct seq_file
*m
, void *p
)
1335 static int svc_pool_stats_show(struct seq_file
*m
, void *p
)
1337 struct svc_pool
*pool
= p
;
1339 if (p
== SEQ_START_TOKEN
) {
1340 seq_puts(m
, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1344 seq_printf(m
, "%u %lu %lu %lu %lu\n",
1346 (unsigned long)atomic_long_read(&pool
->sp_stats
.packets
),
1347 pool
->sp_stats
.sockets_queued
,
1348 (unsigned long)atomic_long_read(&pool
->sp_stats
.threads_woken
),
1349 (unsigned long)atomic_long_read(&pool
->sp_stats
.threads_timedout
));
1354 static const struct seq_operations svc_pool_stats_seq_ops
= {
1355 .start
= svc_pool_stats_start
,
1356 .next
= svc_pool_stats_next
,
1357 .stop
= svc_pool_stats_stop
,
1358 .show
= svc_pool_stats_show
,
1361 int svc_pool_stats_open(struct svc_serv
*serv
, struct file
*file
)
1365 err
= seq_open(file
, &svc_pool_stats_seq_ops
);
1367 ((struct seq_file
*) file
->private_data
)->private = serv
;
1370 EXPORT_SYMBOL(svc_pool_stats_open
);
1372 /*----------------------------------------------------------------------------*/