Linux Audio

Check our new training course

In-person Linux kernel drivers training

Jun 16-20, 2025
Register
Loading...
   1/*
   2 * linux/net/sunrpc/svc_xprt.c
   3 *
   4 * Author: Tom Tucker <tom@opengridcomputing.com>
   5 */
   6
   7#include <linux/sched.h>
   8#include <linux/errno.h>
   9#include <linux/freezer.h>
  10#include <linux/kthread.h>
  11#include <linux/slab.h>
  12#include <net/sock.h>
  13#include <linux/sunrpc/addr.h>
  14#include <linux/sunrpc/stats.h>
  15#include <linux/sunrpc/svc_xprt.h>
  16#include <linux/sunrpc/svcsock.h>
  17#include <linux/sunrpc/xprt.h>
  18#include <linux/module.h>
  19#include <linux/netdevice.h>
  20#include <trace/events/sunrpc.h>
  21
  22#define RPCDBG_FACILITY	RPCDBG_SVCXPRT
  23
  24static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
  25static int svc_deferred_recv(struct svc_rqst *rqstp);
  26static struct cache_deferred_req *svc_defer(struct cache_req *req);
  27static void svc_age_temp_xprts(unsigned long closure);
  28static void svc_delete_xprt(struct svc_xprt *xprt);
  29
  30/* apparently the "standard" is that clients close
  31 * idle connections after 5 minutes, servers after
  32 * 6 minutes
  33 *   http://www.connectathon.org/talks96/nfstcp.pdf
  34 */
  35static int svc_conn_age_period = 6*60;
  36
  37/* List of registered transport classes */
  38static DEFINE_SPINLOCK(svc_xprt_class_lock);
  39static LIST_HEAD(svc_xprt_class_list);
  40
  41/* SMP locking strategy:
  42 *
  43 *	svc_pool->sp_lock protects most of the fields of that pool.
  44 *	svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
  45 *	when both need to be taken (rare), svc_serv->sv_lock is first.
  46 *	The "service mutex" protects svc_serv->sv_nrthread.
  47 *	svc_sock->sk_lock protects the svc_sock->sk_deferred list
  48 *             and the ->sk_info_authunix cache.
  49 *
  50 *	The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
  51 *	enqueued multiply. During normal transport processing this bit
  52 *	is set by svc_xprt_enqueue and cleared by svc_xprt_received.
  53 *	Providers should not manipulate this bit directly.
  54 *
  55 *	Some flags can be set to certain values at any time
  56 *	providing that certain rules are followed:
  57 *
  58 *	XPT_CONN, XPT_DATA:
  59 *		- Can be set or cleared at any time.
  60 *		- After a set, svc_xprt_enqueue must be called to enqueue
  61 *		  the transport for processing.
  62 *		- After a clear, the transport must be read/accepted.
  63 *		  If this succeeds, it must be set again.
  64 *	XPT_CLOSE:
  65 *		- Can set at any time. It is never cleared.
  66 *      XPT_DEAD:
  67 *		- Can only be set while XPT_BUSY is held which ensures
  68 *		  that no other thread will be using the transport or will
  69 *		  try to set XPT_DEAD.
  70 */
 
  71int svc_reg_xprt_class(struct svc_xprt_class *xcl)
  72{
  73	struct svc_xprt_class *cl;
  74	int res = -EEXIST;
  75
  76	dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
  77
  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)
  83			goto out;
  84	}
  85	list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
  86	res = 0;
  87out:
  88	spin_unlock(&svc_xprt_class_lock);
  89	return res;
  90}
  91EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
  92
  93void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
  94{
  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);
  99}
 100EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
 101
 102/*
 103 * Format the transport list for printing
 104 */
 105int svc_print_xprts(char *buf, int maxlen)
 106{
 107	struct svc_xprt_class *xcl;
 108	char tmpstr[80];
 109	int len = 0;
 110	buf[0] = '\0';
 111
 112	spin_lock(&svc_xprt_class_lock);
 113	list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
 114		int slen;
 115
 116		sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
 117		slen = strlen(tmpstr);
 118		if (len + slen > maxlen)
 119			break;
 120		len += slen;
 121		strcat(buf, tmpstr);
 122	}
 123	spin_unlock(&svc_xprt_class_lock);
 124
 125	return len;
 126}
 127
 128static void svc_xprt_free(struct kref *kref)
 129{
 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);
 140	module_put(owner);
 141}
 142
 143void svc_xprt_put(struct svc_xprt *xprt)
 144{
 145	kref_put(&xprt->xpt_ref, svc_xprt_free);
 146}
 147EXPORT_SYMBOL_GPL(svc_xprt_put);
 148
 149/*
 150 * Called by transport drivers to initialize the transport independent
 151 * portion of the transport instance.
 152 */
 153void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
 154		   struct svc_xprt *xprt, struct svc_serv *serv)
 155{
 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);
 170}
 171EXPORT_SYMBOL_GPL(svc_xprt_init);
 172
 173static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
 174					 struct svc_serv *serv,
 175					 struct net *net,
 176					 const int family,
 177					 const unsigned short port,
 178					 int flags)
 179{
 180	struct sockaddr_in sin = {
 181		.sin_family		= AF_INET,
 182		.sin_addr.s_addr	= htonl(INADDR_ANY),
 183		.sin_port		= htons(port),
 184	};
 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),
 190	};
 191#endif
 192	struct sockaddr *sap;
 193	size_t len;
 194
 195	switch (family) {
 196	case PF_INET:
 197		sap = (struct sockaddr *)&sin;
 198		len = sizeof(sin);
 199		break;
 200#if IS_ENABLED(CONFIG_IPV6)
 201	case PF_INET6:
 202		sap = (struct sockaddr *)&sin6;
 203		len = sizeof(sin6);
 204		break;
 205#endif
 206	default:
 207		return ERR_PTR(-EAFNOSUPPORT);
 208	}
 209
 210	return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
 211}
 212
 213/*
 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.
 217 *
 218 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
 219 * insufficient) data.
 220 */
 221static void svc_xprt_received(struct svc_xprt *xprt)
 222{
 223	if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
 224		WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
 225		return;
 226	}
 227
 228	/* As soon as we clear busy, the xprt could be closed and
 229	 * 'put', so we need a reference to call svc_enqueue_xprt with:
 230	 */
 231	svc_xprt_get(xprt);
 232	smp_mb__before_atomic();
 233	clear_bit(XPT_BUSY, &xprt->xpt_flags);
 234	xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
 235	svc_xprt_put(xprt);
 236}
 237
 238void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
 239{
 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);
 245}
 246
 247int 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)
 250{
 251	struct svc_xprt_class *xcl;
 252
 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;
 258
 259		if (strcmp(xprt_name, xcl->xcl_name))
 260			continue;
 261
 262		if (!try_module_get(xcl->xcl_owner))
 263			goto err;
 264
 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);
 270		}
 271		svc_add_new_perm_xprt(serv, newxprt);
 
 
 
 
 272		newport = svc_xprt_local_port(newxprt);
 
 273		return newport;
 274	}
 275 err:
 276	spin_unlock(&svc_xprt_class_lock);
 277	dprintk("svc: transport %s not found\n", xprt_name);
 278
 279	/* This errno is exposed to user space.  Provide a reasonable
 280	 * perror msg for a bad transport. */
 281	return -EPROTONOSUPPORT;
 282}
 283EXPORT_SYMBOL_GPL(svc_create_xprt);
 284
 285/*
 286 * Copy the local and remote xprt addresses to the rqstp structure
 287 */
 288void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
 289{
 
 
 290	memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
 291	rqstp->rq_addrlen = xprt->xpt_remotelen;
 292
 293	/*
 294	 * Destination address in request is needed for binding the
 295	 * source address in RPC replies/callbacks later.
 296	 */
 297	memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
 298	rqstp->rq_daddrlen = xprt->xpt_locallen;
 
 
 
 
 
 
 
 299}
 300EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
 301
 302/**
 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
 307 *
 308 */
 309char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
 310{
 311	return __svc_print_addr(svc_addr(rqstp), buf, len);
 312}
 313EXPORT_SYMBOL_GPL(svc_print_addr);
 314
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 315static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
 316{
 317	if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
 318		return true;
 319	if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED)))
 320		return xprt->xpt_ops->xpo_has_wspace(xprt);
 321	return false;
 322}
 323
 324void svc_xprt_do_enqueue(struct svc_xprt *xprt)
 
 
 
 
 
 325{
 
 326	struct svc_pool *pool;
 327	struct svc_rqst	*rqstp = NULL;
 328	int cpu;
 329	bool queued = false;
 330
 331	if (!svc_xprt_has_something_to_do(xprt))
 332		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 333
 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.
 338	 */
 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);
 342		goto out;
 343	}
 344
 345	cpu = get_cpu();
 346	pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
 347
 348	atomic_long_inc(&pool->sp_stats.packets);
 349
 350redo_search:
 351	/* find a thread for this xprt */
 352	rcu_read_lock();
 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))
 356			continue;
 357
 358		/*
 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
 362		 * do so.
 363		 */
 364		if (!queued) {
 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);
 369				continue;
 370			}
 371
 372			/* this one will do */
 373			rqstp->rq_xprt = xprt;
 374			svc_xprt_get(xprt);
 375			spin_unlock_bh(&rqstp->rq_lock);
 376		}
 377		rcu_read_unlock();
 378
 379		atomic_long_inc(&pool->sp_stats.threads_woken);
 380		wake_up_process(rqstp->rq_task);
 381		put_cpu();
 382		goto out;
 383	}
 384	rcu_read_unlock();
 385
 386	/*
 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.
 391	 */
 392	if (!queued) {
 393		queued = true;
 
 
 
 
 
 
 
 
 
 
 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);
 399		goto redo_search;
 400	}
 401	rqstp = NULL;
 402	put_cpu();
 403out:
 404	trace_svc_xprt_do_enqueue(xprt, rqstp);
 405}
 406EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
 407
 408/*
 409 * Queue up a transport with data pending. If there are idle nfsd
 410 * processes, wake 'em up.
 411 *
 412 */
 413void svc_xprt_enqueue(struct svc_xprt *xprt)
 414{
 415	if (test_bit(XPT_BUSY, &xprt->xpt_flags))
 416		return;
 417	xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
 418}
 419EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
 420
 421/*
 422 * Dequeue the first transport, if there is one.
 423 */
 424static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
 425{
 426	struct svc_xprt	*xprt = NULL;
 427
 428	if (list_empty(&pool->sp_sockets))
 429		goto out;
 430
 431	spin_lock_bh(&pool->sp_lock);
 432	if (likely(!list_empty(&pool->sp_sockets))) {
 433		xprt = list_first_entry(&pool->sp_sockets,
 434					struct svc_xprt, xpt_ready);
 435		list_del_init(&xprt->xpt_ready);
 436		svc_xprt_get(xprt);
 437
 438		dprintk("svc: transport %p dequeued, inuse=%d\n",
 439			xprt, atomic_read(&xprt->xpt_ref.refcount));
 440	}
 441	spin_unlock_bh(&pool->sp_lock);
 442out:
 443	trace_svc_xprt_dequeue(xprt);
 444	return xprt;
 445}
 446
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 447/**
 448 * svc_reserve - change the space reserved for the reply to a request.
 449 * @rqstp:  The request in question
 450 * @space: new max space to reserve
 451 *
 452 * Each request reserves some space on the output queue of the transport
 453 * to make sure the reply fits.  This function reduces that reserved
 454 * space to be the amount of space used already, plus @space.
 455 *
 456 */
 457void svc_reserve(struct svc_rqst *rqstp, int space)
 458{
 459	space += rqstp->rq_res.head[0].iov_len;
 460
 461	if (space < rqstp->rq_reserved) {
 462		struct svc_xprt *xprt = rqstp->rq_xprt;
 463		atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
 464		rqstp->rq_reserved = space;
 465
 466		if (xprt->xpt_ops->xpo_adjust_wspace)
 467			xprt->xpt_ops->xpo_adjust_wspace(xprt);
 468		svc_xprt_enqueue(xprt);
 469	}
 470}
 471EXPORT_SYMBOL_GPL(svc_reserve);
 472
 473static void svc_xprt_release(struct svc_rqst *rqstp)
 474{
 475	struct svc_xprt	*xprt = rqstp->rq_xprt;
 476
 477	rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
 478
 479	kfree(rqstp->rq_deferred);
 480	rqstp->rq_deferred = NULL;
 481
 482	svc_free_res_pages(rqstp);
 483	rqstp->rq_res.page_len = 0;
 484	rqstp->rq_res.page_base = 0;
 485
 486	/* Reset response buffer and release
 487	 * the reservation.
 488	 * But first, check that enough space was reserved
 489	 * for the reply, otherwise we have a bug!
 490	 */
 491	if ((rqstp->rq_res.len) >  rqstp->rq_reserved)
 492		printk(KERN_ERR "RPC request reserved %d but used %d\n",
 493		       rqstp->rq_reserved,
 494		       rqstp->rq_res.len);
 495
 496	rqstp->rq_res.head[0].iov_len = 0;
 497	svc_reserve(rqstp, 0);
 498	rqstp->rq_xprt = NULL;
 499
 500	svc_xprt_put(xprt);
 501}
 502
 503/*
 504 * Some svc_serv's will have occasional work to do, even when a xprt is not
 505 * waiting to be serviced. This function is there to "kick" a task in one of
 506 * those services so that it can wake up and do that work. Note that we only
 507 * bother with pool 0 as we don't need to wake up more than one thread for
 508 * this purpose.
 509 */
 510void svc_wake_up(struct svc_serv *serv)
 511{
 512	struct svc_rqst	*rqstp;
 
 513	struct svc_pool *pool;
 514
 515	pool = &serv->sv_pools[0];
 
 516
 517	rcu_read_lock();
 518	list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
 519		/* skip any that aren't queued */
 520		if (test_bit(RQ_BUSY, &rqstp->rq_flags))
 521			continue;
 522		rcu_read_unlock();
 523		dprintk("svc: daemon %p woken up.\n", rqstp);
 524		wake_up_process(rqstp->rq_task);
 525		trace_svc_wake_up(rqstp->rq_task->pid);
 526		return;
 
 
 
 527	}
 528	rcu_read_unlock();
 529
 530	/* No free entries available */
 531	set_bit(SP_TASK_PENDING, &pool->sp_flags);
 532	smp_wmb();
 533	trace_svc_wake_up(0);
 534}
 535EXPORT_SYMBOL_GPL(svc_wake_up);
 536
 537int svc_port_is_privileged(struct sockaddr *sin)
 538{
 539	switch (sin->sa_family) {
 540	case AF_INET:
 541		return ntohs(((struct sockaddr_in *)sin)->sin_port)
 542			< PROT_SOCK;
 543	case AF_INET6:
 544		return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
 545			< PROT_SOCK;
 546	default:
 547		return 0;
 548	}
 549}
 550
 551/*
 552 * Make sure that we don't have too many active connections. If we have,
 553 * something must be dropped. It's not clear what will happen if we allow
 554 * "too many" connections, but when dealing with network-facing software,
 555 * we have to code defensively. Here we do that by imposing hard limits.
 556 *
 557 * There's no point in trying to do random drop here for DoS
 558 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
 559 * attacker can easily beat that.
 560 *
 561 * The only somewhat efficient mechanism would be if drop old
 562 * connections from the same IP first. But right now we don't even
 563 * record the client IP in svc_sock.
 564 *
 565 * single-threaded services that expect a lot of clients will probably
 566 * need to set sv_maxconn to override the default value which is based
 567 * on the number of threads
 568 */
 569static void svc_check_conn_limits(struct svc_serv *serv)
 570{
 571	unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
 572				(serv->sv_nrthreads+3) * 20;
 573
 574	if (serv->sv_tmpcnt > limit) {
 575		struct svc_xprt *xprt = NULL;
 576		spin_lock_bh(&serv->sv_lock);
 577		if (!list_empty(&serv->sv_tempsocks)) {
 578			/* Try to help the admin */
 579			net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
 580					       serv->sv_name, serv->sv_maxconn ?
 581					       "max number of connections" :
 582					       "number of threads");
 
 
 
 583			/*
 584			 * Always select the oldest connection. It's not fair,
 585			 * but so is life
 586			 */
 587			xprt = list_entry(serv->sv_tempsocks.prev,
 588					  struct svc_xprt,
 589					  xpt_list);
 590			set_bit(XPT_CLOSE, &xprt->xpt_flags);
 591			svc_xprt_get(xprt);
 592		}
 593		spin_unlock_bh(&serv->sv_lock);
 594
 595		if (xprt) {
 596			svc_xprt_enqueue(xprt);
 597			svc_xprt_put(xprt);
 598		}
 599	}
 600}
 601
 602static int svc_alloc_arg(struct svc_rqst *rqstp)
 
 
 
 
 
 603{
 604	struct svc_serv *serv = rqstp->rq_server;
 605	struct xdr_buf *arg;
 606	int pages;
 607	int i;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 608
 609	/* now allocate needed pages.  If we get a failure, sleep briefly */
 610	pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
 611	WARN_ON_ONCE(pages >= RPCSVC_MAXPAGES);
 612	if (pages >= RPCSVC_MAXPAGES)
 613		/* use as many pages as possible */
 614		pages = RPCSVC_MAXPAGES - 1;
 615	for (i = 0; i < pages ; i++)
 616		while (rqstp->rq_pages[i] == NULL) {
 617			struct page *p = alloc_page(GFP_KERNEL);
 618			if (!p) {
 619				set_current_state(TASK_INTERRUPTIBLE);
 620				if (signalled() || kthread_should_stop()) {
 621					set_current_state(TASK_RUNNING);
 622					return -EINTR;
 623				}
 624				schedule_timeout(msecs_to_jiffies(500));
 625			}
 626			rqstp->rq_pages[i] = p;
 627		}
 628	rqstp->rq_page_end = &rqstp->rq_pages[i];
 629	rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
 
 630
 631	/* Make arg->head point to first page and arg->pages point to rest */
 632	arg = &rqstp->rq_arg;
 633	arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
 634	arg->head[0].iov_len = PAGE_SIZE;
 635	arg->pages = rqstp->rq_pages + 1;
 636	arg->page_base = 0;
 637	/* save at least one page for response */
 638	arg->page_len = (pages-2)*PAGE_SIZE;
 639	arg->len = (pages-1)*PAGE_SIZE;
 640	arg->tail[0].iov_len = 0;
 641	return 0;
 642}
 643
 644static bool
 645rqst_should_sleep(struct svc_rqst *rqstp)
 646{
 647	struct svc_pool		*pool = rqstp->rq_pool;
 648
 649	/* did someone call svc_wake_up? */
 650	if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
 651		return false;
 652
 653	/* was a socket queued? */
 654	if (!list_empty(&pool->sp_sockets))
 655		return false;
 656
 657	/* are we shutting down? */
 
 658	if (signalled() || kthread_should_stop())
 659		return false;
 660
 661	/* are we freezing? */
 662	if (freezing(current))
 663		return false;
 664
 665	return true;
 666}
 667
 668static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
 669{
 670	struct svc_xprt *xprt;
 671	struct svc_pool		*pool = rqstp->rq_pool;
 672	long			time_left = 0;
 673
 674	/* rq_xprt should be clear on entry */
 675	WARN_ON_ONCE(rqstp->rq_xprt);
 676
 677	/* Normally we will wait up to 5 seconds for any required
 678	 * cache information to be provided.
 679	 */
 680	rqstp->rq_chandle.thread_wait = 5*HZ;
 681
 
 682	xprt = svc_xprt_dequeue(pool);
 683	if (xprt) {
 684		rqstp->rq_xprt = xprt;
 
 
 
 685
 686		/* As there is a shortage of threads and this request
 687		 * had to be queued, don't allow the thread to wait so
 688		 * long for cache updates.
 689		 */
 690		rqstp->rq_chandle.thread_wait = 1*HZ;
 691		clear_bit(SP_TASK_PENDING, &pool->sp_flags);
 692		return xprt;
 693	}
 694
 695	/*
 696	 * We have to be able to interrupt this wait
 697	 * to bring down the daemons ...
 698	 */
 699	set_current_state(TASK_INTERRUPTIBLE);
 700	clear_bit(RQ_BUSY, &rqstp->rq_flags);
 701	smp_mb();
 702
 703	if (likely(rqst_should_sleep(rqstp)))
 704		time_left = schedule_timeout(timeout);
 705	else
 706		__set_current_state(TASK_RUNNING);
 707
 708	try_to_freeze();
 
 
 
 
 709
 710	spin_lock_bh(&rqstp->rq_lock);
 711	set_bit(RQ_BUSY, &rqstp->rq_flags);
 712	spin_unlock_bh(&rqstp->rq_lock);
 
 
 
 
 
 
 
 
 
 713
 714	xprt = rqstp->rq_xprt;
 715	if (xprt != NULL)
 716		return xprt;
 717
 718	if (!time_left)
 719		atomic_long_inc(&pool->sp_stats.threads_timedout);
 720
 721	if (signalled() || kthread_should_stop())
 722		return ERR_PTR(-EINTR);
 723	return ERR_PTR(-EAGAIN);
 724}
 725
 726static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
 727{
 728	spin_lock_bh(&serv->sv_lock);
 729	set_bit(XPT_TEMP, &newxpt->xpt_flags);
 730	list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
 731	serv->sv_tmpcnt++;
 732	if (serv->sv_temptimer.function == NULL) {
 733		/* setup timer to age temp transports */
 734		setup_timer(&serv->sv_temptimer, svc_age_temp_xprts,
 735			    (unsigned long)serv);
 736		mod_timer(&serv->sv_temptimer,
 737			  jiffies + svc_conn_age_period * HZ);
 
 
 
 738	}
 739	spin_unlock_bh(&serv->sv_lock);
 740	svc_xprt_received(newxpt);
 741}
 742
 743static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
 744{
 745	struct svc_serv *serv = rqstp->rq_server;
 746	int len = 0;
 747
 
 748	if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
 749		dprintk("svc_recv: found XPT_CLOSE\n");
 750		svc_delete_xprt(xprt);
 751		/* Leave XPT_BUSY set on the dead xprt: */
 752		goto out;
 753	}
 754	if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
 755		struct svc_xprt *newxpt;
 756		/*
 757		 * We know this module_get will succeed because the
 758		 * listener holds a reference too
 759		 */
 760		__module_get(xprt->xpt_class->xcl_owner);
 761		svc_check_conn_limits(xprt->xpt_server);
 762		newxpt = xprt->xpt_ops->xpo_accept(xprt);
 763		if (newxpt)
 764			svc_add_new_temp_xprt(serv, newxpt);
 765		else
 766			module_put(xprt->xpt_class->xcl_owner);
 767	} else {
 768		/* XPT_DATA|XPT_DEFERRED case: */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 769		dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
 770			rqstp, rqstp->rq_pool->sp_id, xprt,
 771			atomic_read(&xprt->xpt_ref.refcount));
 772		rqstp->rq_deferred = svc_deferred_dequeue(xprt);
 773		if (rqstp->rq_deferred)
 774			len = svc_deferred_recv(rqstp);
 775		else
 776			len = xprt->xpt_ops->xpo_recvfrom(rqstp);
 777		dprintk("svc: got len=%d\n", len);
 778		rqstp->rq_reserved = serv->sv_max_mesg;
 779		atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
 780	}
 781	/* clear XPT_BUSY: */
 782	svc_xprt_received(xprt);
 783out:
 784	trace_svc_handle_xprt(xprt, len);
 785	return len;
 786}
 787
 788/*
 789 * Receive the next request on any transport.  This code is carefully
 790 * organised not to touch any cachelines in the shared svc_serv
 791 * structure, only cachelines in the local svc_pool.
 792 */
 793int svc_recv(struct svc_rqst *rqstp, long timeout)
 794{
 795	struct svc_xprt		*xprt = NULL;
 796	struct svc_serv		*serv = rqstp->rq_server;
 797	int			len, err;
 798
 799	dprintk("svc: server %p waiting for data (to = %ld)\n",
 800		rqstp, timeout);
 801
 802	if (rqstp->rq_xprt)
 803		printk(KERN_ERR
 804			"svc_recv: service %p, transport not NULL!\n",
 805			 rqstp);
 806
 807	err = svc_alloc_arg(rqstp);
 808	if (err)
 809		goto out;
 810
 811	try_to_freeze();
 812	cond_resched();
 813	err = -EINTR;
 814	if (signalled() || kthread_should_stop())
 815		goto out;
 816
 817	xprt = svc_get_next_xprt(rqstp, timeout);
 818	if (IS_ERR(xprt)) {
 819		err = PTR_ERR(xprt);
 820		goto out;
 821	}
 822
 823	len = svc_handle_xprt(rqstp, xprt);
 824
 825	/* No data, incomplete (TCP) read, or accept() */
 826	err = -EAGAIN;
 827	if (len <= 0)
 828		goto out_release;
 829
 830	clear_bit(XPT_OLD, &xprt->xpt_flags);
 831
 832	if (xprt->xpt_ops->xpo_secure_port(rqstp))
 833		set_bit(RQ_SECURE, &rqstp->rq_flags);
 834	else
 835		clear_bit(RQ_SECURE, &rqstp->rq_flags);
 836	rqstp->rq_chandle.defer = svc_defer;
 837	rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
 838
 839	if (serv->sv_stats)
 840		serv->sv_stats->netcnt++;
 841	trace_svc_recv(rqstp, len);
 842	return len;
 843out_release:
 844	rqstp->rq_res.len = 0;
 845	svc_xprt_release(rqstp);
 846out:
 847	trace_svc_recv(rqstp, err);
 848	return err;
 849}
 850EXPORT_SYMBOL_GPL(svc_recv);
 851
 852/*
 853 * Drop request
 854 */
 855void svc_drop(struct svc_rqst *rqstp)
 856{
 857	dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
 858	svc_xprt_release(rqstp);
 859}
 860EXPORT_SYMBOL_GPL(svc_drop);
 861
 862/*
 863 * Return reply to client.
 864 */
 865int svc_send(struct svc_rqst *rqstp)
 866{
 867	struct svc_xprt	*xprt;
 868	int		len = -EFAULT;
 869	struct xdr_buf	*xb;
 870
 871	xprt = rqstp->rq_xprt;
 872	if (!xprt)
 873		goto out;
 874
 875	/* release the receive skb before sending the reply */
 876	rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
 877
 878	/* calculate over-all length */
 879	xb = &rqstp->rq_res;
 880	xb->len = xb->head[0].iov_len +
 881		xb->page_len +
 882		xb->tail[0].iov_len;
 883
 884	/* Grab mutex to serialize outgoing data. */
 885	mutex_lock(&xprt->xpt_mutex);
 886	if (test_bit(XPT_DEAD, &xprt->xpt_flags)
 887			|| test_bit(XPT_CLOSE, &xprt->xpt_flags))
 888		len = -ENOTCONN;
 889	else
 890		len = xprt->xpt_ops->xpo_sendto(rqstp);
 891	mutex_unlock(&xprt->xpt_mutex);
 892	rpc_wake_up(&xprt->xpt_bc_pending);
 893	svc_xprt_release(rqstp);
 894
 895	if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
 896		len = 0;
 897out:
 898	trace_svc_send(rqstp, len);
 899	return len;
 900}
 901
 902/*
 903 * Timer function to close old temporary transports, using
 904 * a mark-and-sweep algorithm.
 905 */
 906static void svc_age_temp_xprts(unsigned long closure)
 907{
 908	struct svc_serv *serv = (struct svc_serv *)closure;
 909	struct svc_xprt *xprt;
 910	struct list_head *le, *next;
 
 911
 912	dprintk("svc_age_temp_xprts\n");
 913
 914	if (!spin_trylock_bh(&serv->sv_lock)) {
 915		/* busy, try again 1 sec later */
 916		dprintk("svc_age_temp_xprts: busy\n");
 917		mod_timer(&serv->sv_temptimer, jiffies + HZ);
 918		return;
 919	}
 920
 921	list_for_each_safe(le, next, &serv->sv_tempsocks) {
 922		xprt = list_entry(le, struct svc_xprt, xpt_list);
 923
 924		/* First time through, just mark it OLD. Second time
 925		 * through, close it. */
 926		if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
 927			continue;
 928		if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
 929		    test_bit(XPT_BUSY, &xprt->xpt_flags))
 930			continue;
 931		list_del_init(le);
 
 932		set_bit(XPT_CLOSE, &xprt->xpt_flags);
 933		dprintk("queuing xprt %p for closing\n", xprt);
 934
 935		/* a thread will dequeue and close it soon */
 936		svc_xprt_enqueue(xprt);
 937	}
 938	spin_unlock_bh(&serv->sv_lock);
 939
 940	mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
 941}
 
 
 
 942
 943/* Close temporary transports whose xpt_local matches server_addr immediately
 944 * instead of waiting for them to be picked up by the timer.
 945 *
 946 * This is meant to be called from a notifier_block that runs when an ip
 947 * address is deleted.
 948 */
 949void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
 950{
 951	struct svc_xprt *xprt;
 952	struct svc_sock *svsk;
 953	struct socket *sock;
 954	struct list_head *le, *next;
 955	LIST_HEAD(to_be_closed);
 956	struct linger no_linger = {
 957		.l_onoff = 1,
 958		.l_linger = 0,
 959	};
 960
 961	spin_lock_bh(&serv->sv_lock);
 962	list_for_each_safe(le, next, &serv->sv_tempsocks) {
 963		xprt = list_entry(le, struct svc_xprt, xpt_list);
 964		if (rpc_cmp_addr(server_addr, (struct sockaddr *)
 965				&xprt->xpt_local)) {
 966			dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
 967			list_move(le, &to_be_closed);
 968		}
 969	}
 970	spin_unlock_bh(&serv->sv_lock);
 971
 972	while (!list_empty(&to_be_closed)) {
 973		le = to_be_closed.next;
 974		list_del_init(le);
 975		xprt = list_entry(le, struct svc_xprt, xpt_list);
 976		dprintk("svc_age_temp_xprts_now: closing %p\n", xprt);
 977		svsk = container_of(xprt, struct svc_sock, sk_xprt);
 978		sock = svsk->sk_sock;
 979		kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER,
 980				  (char *)&no_linger, sizeof(no_linger));
 981		svc_close_xprt(xprt);
 982	}
 983}
 984EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
 985
 986static void call_xpt_users(struct svc_xprt *xprt)
 987{
 988	struct svc_xpt_user *u;
 989
 990	spin_lock(&xprt->xpt_lock);
 991	while (!list_empty(&xprt->xpt_users)) {
 992		u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
 993		list_del(&u->list);
 994		u->callback(u);
 995	}
 996	spin_unlock(&xprt->xpt_lock);
 997}
 998
 999/*
1000 * Remove a dead transport
1001 */
1002static void svc_delete_xprt(struct svc_xprt *xprt)
1003{
1004	struct svc_serv	*serv = xprt->xpt_server;
1005	struct svc_deferred_req *dr;
1006
1007	/* Only do this once */
1008	if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1009		BUG();
1010
1011	dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1012	xprt->xpt_ops->xpo_detach(xprt);
1013
1014	spin_lock_bh(&serv->sv_lock);
1015	list_del_init(&xprt->xpt_list);
1016	WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
 
 
 
 
 
 
 
 
1017	if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1018		serv->sv_tmpcnt--;
1019	spin_unlock_bh(&serv->sv_lock);
1020
1021	while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1022		kfree(dr);
1023
1024	call_xpt_users(xprt);
1025	svc_xprt_put(xprt);
1026}
1027
1028void svc_close_xprt(struct svc_xprt *xprt)
1029{
1030	set_bit(XPT_CLOSE, &xprt->xpt_flags);
1031	if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1032		/* someone else will have to effect the close */
1033		return;
1034	/*
1035	 * We expect svc_close_xprt() to work even when no threads are
1036	 * running (e.g., while configuring the server before starting
1037	 * any threads), so if the transport isn't busy, we delete
1038	 * it ourself:
1039	 */
1040	svc_delete_xprt(xprt);
1041}
1042EXPORT_SYMBOL_GPL(svc_close_xprt);
1043
1044static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1045{
1046	struct svc_xprt *xprt;
1047	int ret = 0;
1048
1049	spin_lock(&serv->sv_lock);
1050	list_for_each_entry(xprt, xprt_list, xpt_list) {
1051		if (xprt->xpt_net != net)
1052			continue;
1053		ret++;
1054		set_bit(XPT_CLOSE, &xprt->xpt_flags);
1055		svc_xprt_enqueue(xprt);
1056	}
1057	spin_unlock(&serv->sv_lock);
1058	return ret;
1059}
1060
1061static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1062{
1063	struct svc_pool *pool;
1064	struct svc_xprt *xprt;
1065	struct svc_xprt *tmp;
1066	int i;
1067
1068	for (i = 0; i < serv->sv_nrpools; i++) {
1069		pool = &serv->sv_pools[i];
1070
1071		spin_lock_bh(&pool->sp_lock);
1072		list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1073			if (xprt->xpt_net != net)
1074				continue;
1075			list_del_init(&xprt->xpt_ready);
1076			spin_unlock_bh(&pool->sp_lock);
1077			return xprt;
1078		}
1079		spin_unlock_bh(&pool->sp_lock);
1080	}
1081	return NULL;
1082}
1083
1084static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1085{
1086	struct svc_xprt *xprt;
1087
1088	while ((xprt = svc_dequeue_net(serv, net))) {
 
 
 
1089		set_bit(XPT_CLOSE, &xprt->xpt_flags);
1090		svc_delete_xprt(xprt);
1091	}
1092}
1093
1094/*
1095 * Server threads may still be running (especially in the case where the
1096 * service is still running in other network namespaces).
1097 *
1098 * So we shut down sockets the same way we would on a running server, by
1099 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1100 * the close.  In the case there are no such other threads,
1101 * threads running, svc_clean_up_xprts() does a simple version of a
1102 * server's main event loop, and in the case where there are other
1103 * threads, we may need to wait a little while and then check again to
1104 * see if they're done.
1105 */
1106void svc_close_net(struct svc_serv *serv, struct net *net)
1107{
1108	int delay = 0;
1109
1110	while (svc_close_list(serv, &serv->sv_permsocks, net) +
1111	       svc_close_list(serv, &serv->sv_tempsocks, net)) {
1112
1113		svc_clean_up_xprts(serv, net);
1114		msleep(delay++);
1115	}
1116}
1117
1118/*
1119 * Handle defer and revisit of requests
1120 */
1121
1122static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1123{
1124	struct svc_deferred_req *dr =
1125		container_of(dreq, struct svc_deferred_req, handle);
1126	struct svc_xprt *xprt = dr->xprt;
1127
1128	spin_lock(&xprt->xpt_lock);
1129	set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1130	if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1131		spin_unlock(&xprt->xpt_lock);
1132		dprintk("revisit canceled\n");
1133		svc_xprt_put(xprt);
1134		kfree(dr);
1135		return;
1136	}
1137	dprintk("revisit queued\n");
1138	dr->xprt = NULL;
1139	list_add(&dr->handle.recent, &xprt->xpt_deferred);
1140	spin_unlock(&xprt->xpt_lock);
1141	svc_xprt_enqueue(xprt);
1142	svc_xprt_put(xprt);
1143}
1144
1145/*
1146 * Save the request off for later processing. The request buffer looks
1147 * like this:
1148 *
1149 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1150 *
1151 * This code can only handle requests that consist of an xprt-header
1152 * and rpc-header.
1153 */
1154static struct cache_deferred_req *svc_defer(struct cache_req *req)
1155{
1156	struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1157	struct svc_deferred_req *dr;
1158
1159	if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1160		return NULL; /* if more than a page, give up FIXME */
1161	if (rqstp->rq_deferred) {
1162		dr = rqstp->rq_deferred;
1163		rqstp->rq_deferred = NULL;
1164	} else {
1165		size_t skip;
1166		size_t size;
1167		/* FIXME maybe discard if size too large */
1168		size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1169		dr = kmalloc(size, GFP_KERNEL);
1170		if (dr == NULL)
1171			return NULL;
1172
1173		dr->handle.owner = rqstp->rq_server;
1174		dr->prot = rqstp->rq_prot;
1175		memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1176		dr->addrlen = rqstp->rq_addrlen;
1177		dr->daddr = rqstp->rq_daddr;
1178		dr->argslen = rqstp->rq_arg.len >> 2;
1179		dr->xprt_hlen = rqstp->rq_xprt_hlen;
1180
1181		/* back up head to the start of the buffer and copy */
1182		skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1183		memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1184		       dr->argslen << 2);
1185	}
1186	svc_xprt_get(rqstp->rq_xprt);
1187	dr->xprt = rqstp->rq_xprt;
1188	set_bit(RQ_DROPME, &rqstp->rq_flags);
1189
1190	dr->handle.revisit = svc_revisit;
1191	return &dr->handle;
1192}
1193
1194/*
1195 * recv data from a deferred request into an active one
1196 */
1197static int svc_deferred_recv(struct svc_rqst *rqstp)
1198{
1199	struct svc_deferred_req *dr = rqstp->rq_deferred;
1200
1201	/* setup iov_base past transport header */
1202	rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1203	/* The iov_len does not include the transport header bytes */
1204	rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1205	rqstp->rq_arg.page_len = 0;
1206	/* The rq_arg.len includes the transport header bytes */
1207	rqstp->rq_arg.len     = dr->argslen<<2;
1208	rqstp->rq_prot        = dr->prot;
1209	memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1210	rqstp->rq_addrlen     = dr->addrlen;
1211	/* Save off transport header len in case we get deferred again */
1212	rqstp->rq_xprt_hlen   = dr->xprt_hlen;
1213	rqstp->rq_daddr       = dr->daddr;
1214	rqstp->rq_respages    = rqstp->rq_pages;
1215	return (dr->argslen<<2) - dr->xprt_hlen;
1216}
1217
1218
1219static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1220{
1221	struct svc_deferred_req *dr = NULL;
1222
1223	if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1224		return NULL;
1225	spin_lock(&xprt->xpt_lock);
1226	if (!list_empty(&xprt->xpt_deferred)) {
1227		dr = list_entry(xprt->xpt_deferred.next,
1228				struct svc_deferred_req,
1229				handle.recent);
1230		list_del_init(&dr->handle.recent);
1231	} else
1232		clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1233	spin_unlock(&xprt->xpt_lock);
1234	return dr;
1235}
1236
1237/**
1238 * svc_find_xprt - find an RPC transport instance
1239 * @serv: pointer to svc_serv to search
1240 * @xcl_name: C string containing transport's class name
1241 * @net: owner net pointer
1242 * @af: Address family of transport's local address
1243 * @port: transport's IP port number
1244 *
1245 * Return the transport instance pointer for the endpoint accepting
1246 * connections/peer traffic from the specified transport class,
1247 * address family and port.
1248 *
1249 * Specifying 0 for the address family or port is effectively a
1250 * wild-card, and will result in matching the first transport in the
1251 * service's list that has a matching class name.
1252 */
1253struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1254			       struct net *net, const sa_family_t af,
1255			       const unsigned short port)
1256{
1257	struct svc_xprt *xprt;
1258	struct svc_xprt *found = NULL;
1259
1260	/* Sanity check the args */
1261	if (serv == NULL || xcl_name == NULL)
1262		return found;
1263
1264	spin_lock_bh(&serv->sv_lock);
1265	list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1266		if (xprt->xpt_net != net)
1267			continue;
1268		if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1269			continue;
1270		if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1271			continue;
1272		if (port != 0 && port != svc_xprt_local_port(xprt))
1273			continue;
1274		found = xprt;
1275		svc_xprt_get(xprt);
1276		break;
1277	}
1278	spin_unlock_bh(&serv->sv_lock);
1279	return found;
1280}
1281EXPORT_SYMBOL_GPL(svc_find_xprt);
1282
1283static int svc_one_xprt_name(const struct svc_xprt *xprt,
1284			     char *pos, int remaining)
1285{
1286	int len;
1287
1288	len = snprintf(pos, remaining, "%s %u\n",
1289			xprt->xpt_class->xcl_name,
1290			svc_xprt_local_port(xprt));
1291	if (len >= remaining)
1292		return -ENAMETOOLONG;
1293	return len;
1294}
1295
1296/**
1297 * svc_xprt_names - format a buffer with a list of transport names
1298 * @serv: pointer to an RPC service
1299 * @buf: pointer to a buffer to be filled in
1300 * @buflen: length of buffer to be filled in
1301 *
1302 * Fills in @buf with a string containing a list of transport names,
1303 * each name terminated with '\n'.
1304 *
1305 * Returns positive length of the filled-in string on success; otherwise
1306 * a negative errno value is returned if an error occurs.
1307 */
1308int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1309{
1310	struct svc_xprt *xprt;
1311	int len, totlen;
1312	char *pos;
1313
1314	/* Sanity check args */
1315	if (!serv)
1316		return 0;
1317
1318	spin_lock_bh(&serv->sv_lock);
1319
1320	pos = buf;
1321	totlen = 0;
1322	list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1323		len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1324		if (len < 0) {
1325			*buf = '\0';
1326			totlen = len;
1327		}
1328		if (len <= 0)
1329			break;
1330
1331		pos += len;
1332		totlen += len;
1333	}
1334
1335	spin_unlock_bh(&serv->sv_lock);
1336	return totlen;
1337}
1338EXPORT_SYMBOL_GPL(svc_xprt_names);
1339
1340
1341/*----------------------------------------------------------------------------*/
1342
1343static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1344{
1345	unsigned int pidx = (unsigned int)*pos;
1346	struct svc_serv *serv = m->private;
1347
1348	dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1349
1350	if (!pidx)
1351		return SEQ_START_TOKEN;
1352	return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1353}
1354
1355static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1356{
1357	struct svc_pool *pool = p;
1358	struct svc_serv *serv = m->private;
1359
1360	dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1361
1362	if (p == SEQ_START_TOKEN) {
1363		pool = &serv->sv_pools[0];
1364	} else {
1365		unsigned int pidx = (pool - &serv->sv_pools[0]);
1366		if (pidx < serv->sv_nrpools-1)
1367			pool = &serv->sv_pools[pidx+1];
1368		else
1369			pool = NULL;
1370	}
1371	++*pos;
1372	return pool;
1373}
1374
1375static void svc_pool_stats_stop(struct seq_file *m, void *p)
1376{
1377}
1378
1379static int svc_pool_stats_show(struct seq_file *m, void *p)
1380{
1381	struct svc_pool *pool = p;
1382
1383	if (p == SEQ_START_TOKEN) {
1384		seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1385		return 0;
1386	}
1387
1388	seq_printf(m, "%u %lu %lu %lu %lu\n",
1389		pool->sp_id,
1390		(unsigned long)atomic_long_read(&pool->sp_stats.packets),
1391		pool->sp_stats.sockets_queued,
1392		(unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1393		(unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1394
1395	return 0;
1396}
1397
1398static const struct seq_operations svc_pool_stats_seq_ops = {
1399	.start	= svc_pool_stats_start,
1400	.next	= svc_pool_stats_next,
1401	.stop	= svc_pool_stats_stop,
1402	.show	= svc_pool_stats_show,
1403};
1404
1405int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1406{
1407	int err;
1408
1409	err = seq_open(file, &svc_pool_stats_seq_ops);
1410	if (!err)
1411		((struct seq_file *) file->private_data)->private = serv;
1412	return err;
1413}
1414EXPORT_SYMBOL(svc_pool_stats_open);
1415
1416/*----------------------------------------------------------------------------*/
   1/*
   2 * linux/net/sunrpc/svc_xprt.c
   3 *
   4 * Author: Tom Tucker <tom@opengridcomputing.com>
   5 */
   6
   7#include <linux/sched.h>
   8#include <linux/errno.h>
   9#include <linux/freezer.h>
  10#include <linux/kthread.h>
  11#include <linux/slab.h>
  12#include <net/sock.h>
 
  13#include <linux/sunrpc/stats.h>
  14#include <linux/sunrpc/svc_xprt.h>
  15#include <linux/sunrpc/svcsock.h>
  16#include <linux/sunrpc/xprt.h>
 
 
 
  17
  18#define RPCDBG_FACILITY	RPCDBG_SVCXPRT
  19
  20static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
  21static int svc_deferred_recv(struct svc_rqst *rqstp);
  22static struct cache_deferred_req *svc_defer(struct cache_req *req);
  23static void svc_age_temp_xprts(unsigned long closure);
 
  24
  25/* apparently the "standard" is that clients close
  26 * idle connections after 5 minutes, servers after
  27 * 6 minutes
  28 *   http://www.connectathon.org/talks96/nfstcp.pdf
  29 */
  30static int svc_conn_age_period = 6*60;
  31
  32/* List of registered transport classes */
  33static DEFINE_SPINLOCK(svc_xprt_class_lock);
  34static LIST_HEAD(svc_xprt_class_list);
  35
  36/* SMP locking strategy:
  37 *
  38 *	svc_pool->sp_lock protects most of the fields of that pool.
  39 *	svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
  40 *	when both need to be taken (rare), svc_serv->sv_lock is first.
  41 *	BKL protects svc_serv->sv_nrthread.
  42 *	svc_sock->sk_lock protects the svc_sock->sk_deferred list
  43 *             and the ->sk_info_authunix cache.
  44 *
  45 *	The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
  46 *	enqueued multiply. During normal transport processing this bit
  47 *	is set by svc_xprt_enqueue and cleared by svc_xprt_received.
  48 *	Providers should not manipulate this bit directly.
  49 *
  50 *	Some flags can be set to certain values at any time
  51 *	providing that certain rules are followed:
  52 *
  53 *	XPT_CONN, XPT_DATA:
  54 *		- Can be set or cleared at any time.
  55 *		- After a set, svc_xprt_enqueue must be called to enqueue
  56 *		  the transport for processing.
  57 *		- After a clear, the transport must be read/accepted.
  58 *		  If this succeeds, it must be set again.
  59 *	XPT_CLOSE:
  60 *		- Can set at any time. It is never cleared.
  61 *      XPT_DEAD:
  62 *		- Can only be set while XPT_BUSY is held which ensures
  63 *		  that no other thread will be using the transport or will
  64 *		  try to set XPT_DEAD.
  65 */
  66
  67int svc_reg_xprt_class(struct svc_xprt_class *xcl)
  68{
  69	struct svc_xprt_class *cl;
  70	int res = -EEXIST;
  71
  72	dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
  73
  74	INIT_LIST_HEAD(&xcl->xcl_list);
  75	spin_lock(&svc_xprt_class_lock);
  76	/* Make sure there isn't already a class with the same name */
  77	list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
  78		if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
  79			goto out;
  80	}
  81	list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
  82	res = 0;
  83out:
  84	spin_unlock(&svc_xprt_class_lock);
  85	return res;
  86}
  87EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
  88
  89void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
  90{
  91	dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
  92	spin_lock(&svc_xprt_class_lock);
  93	list_del_init(&xcl->xcl_list);
  94	spin_unlock(&svc_xprt_class_lock);
  95}
  96EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
  97
  98/*
  99 * Format the transport list for printing
 100 */
 101int svc_print_xprts(char *buf, int maxlen)
 102{
 103	struct svc_xprt_class *xcl;
 104	char tmpstr[80];
 105	int len = 0;
 106	buf[0] = '\0';
 107
 108	spin_lock(&svc_xprt_class_lock);
 109	list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
 110		int slen;
 111
 112		sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
 113		slen = strlen(tmpstr);
 114		if (len + slen > maxlen)
 115			break;
 116		len += slen;
 117		strcat(buf, tmpstr);
 118	}
 119	spin_unlock(&svc_xprt_class_lock);
 120
 121	return len;
 122}
 123
 124static void svc_xprt_free(struct kref *kref)
 125{
 126	struct svc_xprt *xprt =
 127		container_of(kref, struct svc_xprt, xpt_ref);
 128	struct module *owner = xprt->xpt_class->xcl_owner;
 129	if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
 130		svcauth_unix_info_release(xprt);
 131	put_net(xprt->xpt_net);
 132	/* See comment on corresponding get in xs_setup_bc_tcp(): */
 133	if (xprt->xpt_bc_xprt)
 134		xprt_put(xprt->xpt_bc_xprt);
 135	xprt->xpt_ops->xpo_free(xprt);
 136	module_put(owner);
 137}
 138
 139void svc_xprt_put(struct svc_xprt *xprt)
 140{
 141	kref_put(&xprt->xpt_ref, svc_xprt_free);
 142}
 143EXPORT_SYMBOL_GPL(svc_xprt_put);
 144
 145/*
 146 * Called by transport drivers to initialize the transport independent
 147 * portion of the transport instance.
 148 */
 149void svc_xprt_init(struct svc_xprt_class *xcl, struct svc_xprt *xprt,
 150		   struct svc_serv *serv)
 151{
 152	memset(xprt, 0, sizeof(*xprt));
 153	xprt->xpt_class = xcl;
 154	xprt->xpt_ops = xcl->xcl_ops;
 155	kref_init(&xprt->xpt_ref);
 156	xprt->xpt_server = serv;
 157	INIT_LIST_HEAD(&xprt->xpt_list);
 158	INIT_LIST_HEAD(&xprt->xpt_ready);
 159	INIT_LIST_HEAD(&xprt->xpt_deferred);
 160	INIT_LIST_HEAD(&xprt->xpt_users);
 161	mutex_init(&xprt->xpt_mutex);
 162	spin_lock_init(&xprt->xpt_lock);
 163	set_bit(XPT_BUSY, &xprt->xpt_flags);
 164	rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
 165	xprt->xpt_net = get_net(&init_net);
 166}
 167EXPORT_SYMBOL_GPL(svc_xprt_init);
 168
 169static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
 170					 struct svc_serv *serv,
 171					 struct net *net,
 172					 const int family,
 173					 const unsigned short port,
 174					 int flags)
 175{
 176	struct sockaddr_in sin = {
 177		.sin_family		= AF_INET,
 178		.sin_addr.s_addr	= htonl(INADDR_ANY),
 179		.sin_port		= htons(port),
 180	};
 181#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 182	struct sockaddr_in6 sin6 = {
 183		.sin6_family		= AF_INET6,
 184		.sin6_addr		= IN6ADDR_ANY_INIT,
 185		.sin6_port		= htons(port),
 186	};
 187#endif	/* defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) */
 188	struct sockaddr *sap;
 189	size_t len;
 190
 191	switch (family) {
 192	case PF_INET:
 193		sap = (struct sockaddr *)&sin;
 194		len = sizeof(sin);
 195		break;
 196#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 197	case PF_INET6:
 198		sap = (struct sockaddr *)&sin6;
 199		len = sizeof(sin6);
 200		break;
 201#endif	/* defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) */
 202	default:
 203		return ERR_PTR(-EAFNOSUPPORT);
 204	}
 205
 206	return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
 207}
 208
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 209int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
 210		    struct net *net, const int family,
 211		    const unsigned short port, int flags)
 212{
 213	struct svc_xprt_class *xcl;
 214
 215	dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
 216	spin_lock(&svc_xprt_class_lock);
 217	list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
 218		struct svc_xprt *newxprt;
 219		unsigned short newport;
 220
 221		if (strcmp(xprt_name, xcl->xcl_name))
 222			continue;
 223
 224		if (!try_module_get(xcl->xcl_owner))
 225			goto err;
 226
 227		spin_unlock(&svc_xprt_class_lock);
 228		newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
 229		if (IS_ERR(newxprt)) {
 230			module_put(xcl->xcl_owner);
 231			return PTR_ERR(newxprt);
 232		}
 233
 234		clear_bit(XPT_TEMP, &newxprt->xpt_flags);
 235		spin_lock_bh(&serv->sv_lock);
 236		list_add(&newxprt->xpt_list, &serv->sv_permsocks);
 237		spin_unlock_bh(&serv->sv_lock);
 238		newport = svc_xprt_local_port(newxprt);
 239		clear_bit(XPT_BUSY, &newxprt->xpt_flags);
 240		return newport;
 241	}
 242 err:
 243	spin_unlock(&svc_xprt_class_lock);
 244	dprintk("svc: transport %s not found\n", xprt_name);
 245
 246	/* This errno is exposed to user space.  Provide a reasonable
 247	 * perror msg for a bad transport. */
 248	return -EPROTONOSUPPORT;
 249}
 250EXPORT_SYMBOL_GPL(svc_create_xprt);
 251
 252/*
 253 * Copy the local and remote xprt addresses to the rqstp structure
 254 */
 255void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
 256{
 257	struct sockaddr *sin;
 258
 259	memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
 260	rqstp->rq_addrlen = xprt->xpt_remotelen;
 261
 262	/*
 263	 * Destination address in request is needed for binding the
 264	 * source address in RPC replies/callbacks later.
 265	 */
 266	sin = (struct sockaddr *)&xprt->xpt_local;
 267	switch (sin->sa_family) {
 268	case AF_INET:
 269		rqstp->rq_daddr.addr = ((struct sockaddr_in *)sin)->sin_addr;
 270		break;
 271	case AF_INET6:
 272		rqstp->rq_daddr.addr6 = ((struct sockaddr_in6 *)sin)->sin6_addr;
 273		break;
 274	}
 275}
 276EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
 277
 278/**
 279 * svc_print_addr - Format rq_addr field for printing
 280 * @rqstp: svc_rqst struct containing address to print
 281 * @buf: target buffer for formatted address
 282 * @len: length of target buffer
 283 *
 284 */
 285char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
 286{
 287	return __svc_print_addr(svc_addr(rqstp), buf, len);
 288}
 289EXPORT_SYMBOL_GPL(svc_print_addr);
 290
 291/*
 292 * Queue up an idle server thread.  Must have pool->sp_lock held.
 293 * Note: this is really a stack rather than a queue, so that we only
 294 * use as many different threads as we need, and the rest don't pollute
 295 * the cache.
 296 */
 297static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
 298{
 299	list_add(&rqstp->rq_list, &pool->sp_threads);
 300}
 301
 302/*
 303 * Dequeue an nfsd thread.  Must have pool->sp_lock held.
 304 */
 305static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
 306{
 307	list_del(&rqstp->rq_list);
 308}
 309
 310static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
 311{
 312	if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
 313		return true;
 314	if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED)))
 315		return xprt->xpt_ops->xpo_has_wspace(xprt);
 316	return false;
 317}
 318
 319/*
 320 * Queue up a transport with data pending. If there are idle nfsd
 321 * processes, wake 'em up.
 322 *
 323 */
 324void svc_xprt_enqueue(struct svc_xprt *xprt)
 325{
 326	struct svc_serv	*serv = xprt->xpt_server;
 327	struct svc_pool *pool;
 328	struct svc_rqst	*rqstp;
 329	int cpu;
 
 330
 331	if (!svc_xprt_has_something_to_do(xprt))
 332		return;
 333
 334	cpu = get_cpu();
 335	pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
 336	put_cpu();
 337
 338	spin_lock_bh(&pool->sp_lock);
 339
 340	if (!list_empty(&pool->sp_threads) &&
 341	    !list_empty(&pool->sp_sockets))
 342		printk(KERN_ERR
 343		       "svc_xprt_enqueue: "
 344		       "threads and transports both waiting??\n");
 345
 346	pool->sp_stats.packets++;
 347
 348	/* Mark transport as busy. It will remain in this state until
 349	 * the provider calls svc_xprt_received. We update XPT_BUSY
 350	 * atomically because it also guards against trying to enqueue
 351	 * the transport twice.
 352	 */
 353	if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
 354		/* Don't enqueue transport while already enqueued */
 355		dprintk("svc: transport %p busy, not enqueued\n", xprt);
 356		goto out_unlock;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 357	}
 
 358
 359	if (!list_empty(&pool->sp_threads)) {
 360		rqstp = list_entry(pool->sp_threads.next,
 361				   struct svc_rqst,
 362				   rq_list);
 363		dprintk("svc: transport %p served by daemon %p\n",
 364			xprt, rqstp);
 365		svc_thread_dequeue(pool, rqstp);
 366		if (rqstp->rq_xprt)
 367			printk(KERN_ERR
 368				"svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
 369				rqstp, rqstp->rq_xprt);
 370		rqstp->rq_xprt = xprt;
 371		svc_xprt_get(xprt);
 372		rqstp->rq_reserved = serv->sv_max_mesg;
 373		atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
 374		pool->sp_stats.threads_woken++;
 375		wake_up(&rqstp->rq_wait);
 376	} else {
 377		dprintk("svc: transport %p put into queue\n", xprt);
 
 378		list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
 379		pool->sp_stats.sockets_queued++;
 
 
 380	}
 
 
 
 
 
 
 381
 382out_unlock:
 383	spin_unlock_bh(&pool->sp_lock);
 
 
 
 
 
 
 
 
 384}
 385EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
 386
 387/*
 388 * Dequeue the first transport.  Must be called with the pool->sp_lock held.
 389 */
 390static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
 391{
 392	struct svc_xprt	*xprt;
 393
 394	if (list_empty(&pool->sp_sockets))
 395		return NULL;
 396
 397	xprt = list_entry(pool->sp_sockets.next,
 398			  struct svc_xprt, xpt_ready);
 399	list_del_init(&xprt->xpt_ready);
 400
 401	dprintk("svc: transport %p dequeued, inuse=%d\n",
 402		xprt, atomic_read(&xprt->xpt_ref.refcount));
 403
 
 
 
 
 
 
 404	return xprt;
 405}
 406
 407/*
 408 * svc_xprt_received conditionally queues the transport for processing
 409 * by another thread. The caller must hold the XPT_BUSY bit and must
 410 * not thereafter touch transport data.
 411 *
 412 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
 413 * insufficient) data.
 414 */
 415void svc_xprt_received(struct svc_xprt *xprt)
 416{
 417	BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
 418	/* As soon as we clear busy, the xprt could be closed and
 419	 * 'put', so we need a reference to call svc_xprt_enqueue with:
 420	 */
 421	svc_xprt_get(xprt);
 422	clear_bit(XPT_BUSY, &xprt->xpt_flags);
 423	svc_xprt_enqueue(xprt);
 424	svc_xprt_put(xprt);
 425}
 426EXPORT_SYMBOL_GPL(svc_xprt_received);
 427
 428/**
 429 * svc_reserve - change the space reserved for the reply to a request.
 430 * @rqstp:  The request in question
 431 * @space: new max space to reserve
 432 *
 433 * Each request reserves some space on the output queue of the transport
 434 * to make sure the reply fits.  This function reduces that reserved
 435 * space to be the amount of space used already, plus @space.
 436 *
 437 */
 438void svc_reserve(struct svc_rqst *rqstp, int space)
 439{
 440	space += rqstp->rq_res.head[0].iov_len;
 441
 442	if (space < rqstp->rq_reserved) {
 443		struct svc_xprt *xprt = rqstp->rq_xprt;
 444		atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
 445		rqstp->rq_reserved = space;
 446
 
 
 447		svc_xprt_enqueue(xprt);
 448	}
 449}
 450EXPORT_SYMBOL_GPL(svc_reserve);
 451
 452static void svc_xprt_release(struct svc_rqst *rqstp)
 453{
 454	struct svc_xprt	*xprt = rqstp->rq_xprt;
 455
 456	rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
 457
 458	kfree(rqstp->rq_deferred);
 459	rqstp->rq_deferred = NULL;
 460
 461	svc_free_res_pages(rqstp);
 462	rqstp->rq_res.page_len = 0;
 463	rqstp->rq_res.page_base = 0;
 464
 465	/* Reset response buffer and release
 466	 * the reservation.
 467	 * But first, check that enough space was reserved
 468	 * for the reply, otherwise we have a bug!
 469	 */
 470	if ((rqstp->rq_res.len) >  rqstp->rq_reserved)
 471		printk(KERN_ERR "RPC request reserved %d but used %d\n",
 472		       rqstp->rq_reserved,
 473		       rqstp->rq_res.len);
 474
 475	rqstp->rq_res.head[0].iov_len = 0;
 476	svc_reserve(rqstp, 0);
 477	rqstp->rq_xprt = NULL;
 478
 479	svc_xprt_put(xprt);
 480}
 481
 482/*
 483 * External function to wake up a server waiting for data
 484 * This really only makes sense for services like lockd
 485 * which have exactly one thread anyway.
 
 
 486 */
 487void svc_wake_up(struct svc_serv *serv)
 488{
 489	struct svc_rqst	*rqstp;
 490	unsigned int i;
 491	struct svc_pool *pool;
 492
 493	for (i = 0; i < serv->sv_nrpools; i++) {
 494		pool = &serv->sv_pools[i];
 495
 496		spin_lock_bh(&pool->sp_lock);
 497		if (!list_empty(&pool->sp_threads)) {
 498			rqstp = list_entry(pool->sp_threads.next,
 499					   struct svc_rqst,
 500					   rq_list);
 501			dprintk("svc: daemon %p woken up.\n", rqstp);
 502			/*
 503			svc_thread_dequeue(pool, rqstp);
 504			rqstp->rq_xprt = NULL;
 505			 */
 506			wake_up(&rqstp->rq_wait);
 507		}
 508		spin_unlock_bh(&pool->sp_lock);
 509	}
 
 
 
 
 
 
 510}
 511EXPORT_SYMBOL_GPL(svc_wake_up);
 512
 513int svc_port_is_privileged(struct sockaddr *sin)
 514{
 515	switch (sin->sa_family) {
 516	case AF_INET:
 517		return ntohs(((struct sockaddr_in *)sin)->sin_port)
 518			< PROT_SOCK;
 519	case AF_INET6:
 520		return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
 521			< PROT_SOCK;
 522	default:
 523		return 0;
 524	}
 525}
 526
 527/*
 528 * Make sure that we don't have too many active connections. If we have,
 529 * something must be dropped. It's not clear what will happen if we allow
 530 * "too many" connections, but when dealing with network-facing software,
 531 * we have to code defensively. Here we do that by imposing hard limits.
 532 *
 533 * There's no point in trying to do random drop here for DoS
 534 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
 535 * attacker can easily beat that.
 536 *
 537 * The only somewhat efficient mechanism would be if drop old
 538 * connections from the same IP first. But right now we don't even
 539 * record the client IP in svc_sock.
 540 *
 541 * single-threaded services that expect a lot of clients will probably
 542 * need to set sv_maxconn to override the default value which is based
 543 * on the number of threads
 544 */
 545static void svc_check_conn_limits(struct svc_serv *serv)
 546{
 547	unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
 548				(serv->sv_nrthreads+3) * 20;
 549
 550	if (serv->sv_tmpcnt > limit) {
 551		struct svc_xprt *xprt = NULL;
 552		spin_lock_bh(&serv->sv_lock);
 553		if (!list_empty(&serv->sv_tempsocks)) {
 554			if (net_ratelimit()) {
 555				/* Try to help the admin */
 556				printk(KERN_NOTICE "%s: too many open  "
 557				       "connections, consider increasing %s\n",
 558				       serv->sv_name, serv->sv_maxconn ?
 559				       "the max number of connections." :
 560				       "the number of threads.");
 561			}
 562			/*
 563			 * Always select the oldest connection. It's not fair,
 564			 * but so is life
 565			 */
 566			xprt = list_entry(serv->sv_tempsocks.prev,
 567					  struct svc_xprt,
 568					  xpt_list);
 569			set_bit(XPT_CLOSE, &xprt->xpt_flags);
 570			svc_xprt_get(xprt);
 571		}
 572		spin_unlock_bh(&serv->sv_lock);
 573
 574		if (xprt) {
 575			svc_xprt_enqueue(xprt);
 576			svc_xprt_put(xprt);
 577		}
 578	}
 579}
 580
 581/*
 582 * Receive the next request on any transport.  This code is carefully
 583 * organised not to touch any cachelines in the shared svc_serv
 584 * structure, only cachelines in the local svc_pool.
 585 */
 586int svc_recv(struct svc_rqst *rqstp, long timeout)
 587{
 588	struct svc_xprt		*xprt = NULL;
 589	struct svc_serv		*serv = rqstp->rq_server;
 590	struct svc_pool		*pool = rqstp->rq_pool;
 591	int			len, i;
 592	int			pages;
 593	struct xdr_buf		*arg;
 594	DECLARE_WAITQUEUE(wait, current);
 595	long			time_left;
 596
 597	dprintk("svc: server %p waiting for data (to = %ld)\n",
 598		rqstp, timeout);
 599
 600	if (rqstp->rq_xprt)
 601		printk(KERN_ERR
 602			"svc_recv: service %p, transport not NULL!\n",
 603			 rqstp);
 604	if (waitqueue_active(&rqstp->rq_wait))
 605		printk(KERN_ERR
 606			"svc_recv: service %p, wait queue active!\n",
 607			 rqstp);
 608
 609	/* now allocate needed pages.  If we get a failure, sleep briefly */
 610	pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
 
 
 
 
 611	for (i = 0; i < pages ; i++)
 612		while (rqstp->rq_pages[i] == NULL) {
 613			struct page *p = alloc_page(GFP_KERNEL);
 614			if (!p) {
 615				set_current_state(TASK_INTERRUPTIBLE);
 616				if (signalled() || kthread_should_stop()) {
 617					set_current_state(TASK_RUNNING);
 618					return -EINTR;
 619				}
 620				schedule_timeout(msecs_to_jiffies(500));
 621			}
 622			rqstp->rq_pages[i] = p;
 623		}
 
 624	rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
 625	BUG_ON(pages >= RPCSVC_MAXPAGES);
 626
 627	/* Make arg->head point to first page and arg->pages point to rest */
 628	arg = &rqstp->rq_arg;
 629	arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
 630	arg->head[0].iov_len = PAGE_SIZE;
 631	arg->pages = rqstp->rq_pages + 1;
 632	arg->page_base = 0;
 633	/* save at least one page for response */
 634	arg->page_len = (pages-2)*PAGE_SIZE;
 635	arg->len = (pages-1)*PAGE_SIZE;
 636	arg->tail[0].iov_len = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 637
 638	try_to_freeze();
 639	cond_resched();
 640	if (signalled() || kthread_should_stop())
 641		return -EINTR;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 642
 643	/* Normally we will wait up to 5 seconds for any required
 644	 * cache information to be provided.
 645	 */
 646	rqstp->rq_chandle.thread_wait = 5*HZ;
 647
 648	spin_lock_bh(&pool->sp_lock);
 649	xprt = svc_xprt_dequeue(pool);
 650	if (xprt) {
 651		rqstp->rq_xprt = xprt;
 652		svc_xprt_get(xprt);
 653		rqstp->rq_reserved = serv->sv_max_mesg;
 654		atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
 655
 656		/* As there is a shortage of threads and this request
 657		 * had to be queued, don't allow the thread to wait so
 658		 * long for cache updates.
 659		 */
 660		rqstp->rq_chandle.thread_wait = 1*HZ;
 661	} else {
 662		/* No data pending. Go to sleep */
 663		svc_thread_enqueue(pool, rqstp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 664
 665		/*
 666		 * We have to be able to interrupt this wait
 667		 * to bring down the daemons ...
 668		 */
 669		set_current_state(TASK_INTERRUPTIBLE);
 670
 671		/*
 672		 * checking kthread_should_stop() here allows us to avoid
 673		 * locking and signalling when stopping kthreads that call
 674		 * svc_recv. If the thread has already been woken up, then
 675		 * we can exit here without sleeping. If not, then it
 676		 * it'll be woken up quickly during the schedule_timeout
 677		 */
 678		if (kthread_should_stop()) {
 679			set_current_state(TASK_RUNNING);
 680			spin_unlock_bh(&pool->sp_lock);
 681			return -EINTR;
 682		}
 683
 684		add_wait_queue(&rqstp->rq_wait, &wait);
 685		spin_unlock_bh(&pool->sp_lock);
 
 686
 687		time_left = schedule_timeout(timeout);
 
 688
 689		try_to_freeze();
 
 
 
 690
 691		spin_lock_bh(&pool->sp_lock);
 692		remove_wait_queue(&rqstp->rq_wait, &wait);
 693		if (!time_left)
 694			pool->sp_stats.threads_timedout++;
 695
 696		xprt = rqstp->rq_xprt;
 697		if (!xprt) {
 698			svc_thread_dequeue(pool, rqstp);
 699			spin_unlock_bh(&pool->sp_lock);
 700			dprintk("svc: server %p, no data yet\n", rqstp);
 701			if (signalled() || kthread_should_stop())
 702				return -EINTR;
 703			else
 704				return -EAGAIN;
 705		}
 706	}
 707	spin_unlock_bh(&pool->sp_lock);
 
 
 
 
 
 
 
 708
 709	len = 0;
 710	if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
 711		dprintk("svc_recv: found XPT_CLOSE\n");
 712		svc_delete_xprt(xprt);
 713		/* Leave XPT_BUSY set on the dead xprt: */
 714		goto out;
 715	}
 716	if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
 717		struct svc_xprt *newxpt;
 
 
 
 
 
 
 718		newxpt = xprt->xpt_ops->xpo_accept(xprt);
 719		if (newxpt) {
 720			/*
 721			 * We know this module_get will succeed because the
 722			 * listener holds a reference too
 723			 */
 724			__module_get(newxpt->xpt_class->xcl_owner);
 725			svc_check_conn_limits(xprt->xpt_server);
 726			spin_lock_bh(&serv->sv_lock);
 727			set_bit(XPT_TEMP, &newxpt->xpt_flags);
 728			list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
 729			serv->sv_tmpcnt++;
 730			if (serv->sv_temptimer.function == NULL) {
 731				/* setup timer to age temp transports */
 732				setup_timer(&serv->sv_temptimer,
 733					    svc_age_temp_xprts,
 734					    (unsigned long)serv);
 735				mod_timer(&serv->sv_temptimer,
 736					  jiffies + svc_conn_age_period * HZ);
 737			}
 738			spin_unlock_bh(&serv->sv_lock);
 739			svc_xprt_received(newxpt);
 740		}
 741	} else if (xprt->xpt_ops->xpo_has_wspace(xprt)) {
 742		dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
 743			rqstp, pool->sp_id, xprt,
 744			atomic_read(&xprt->xpt_ref.refcount));
 745		rqstp->rq_deferred = svc_deferred_dequeue(xprt);
 746		if (rqstp->rq_deferred)
 747			len = svc_deferred_recv(rqstp);
 748		else
 749			len = xprt->xpt_ops->xpo_recvfrom(rqstp);
 750		dprintk("svc: got len=%d\n", len);
 
 
 751	}
 
 752	svc_xprt_received(xprt);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 753
 754	/* No data, incomplete (TCP) read, or accept() */
 755	if (len == 0 || len == -EAGAIN)
 756		goto out;
 
 757
 758	clear_bit(XPT_OLD, &xprt->xpt_flags);
 759
 760	rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
 
 
 
 761	rqstp->rq_chandle.defer = svc_defer;
 
 762
 763	if (serv->sv_stats)
 764		serv->sv_stats->netcnt++;
 
 765	return len;
 766out:
 767	rqstp->rq_res.len = 0;
 768	svc_xprt_release(rqstp);
 769	return -EAGAIN;
 
 
 770}
 771EXPORT_SYMBOL_GPL(svc_recv);
 772
 773/*
 774 * Drop request
 775 */
 776void svc_drop(struct svc_rqst *rqstp)
 777{
 778	dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
 779	svc_xprt_release(rqstp);
 780}
 781EXPORT_SYMBOL_GPL(svc_drop);
 782
 783/*
 784 * Return reply to client.
 785 */
 786int svc_send(struct svc_rqst *rqstp)
 787{
 788	struct svc_xprt	*xprt;
 789	int		len;
 790	struct xdr_buf	*xb;
 791
 792	xprt = rqstp->rq_xprt;
 793	if (!xprt)
 794		return -EFAULT;
 795
 796	/* release the receive skb before sending the reply */
 797	rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
 798
 799	/* calculate over-all length */
 800	xb = &rqstp->rq_res;
 801	xb->len = xb->head[0].iov_len +
 802		xb->page_len +
 803		xb->tail[0].iov_len;
 804
 805	/* Grab mutex to serialize outgoing data. */
 806	mutex_lock(&xprt->xpt_mutex);
 807	if (test_bit(XPT_DEAD, &xprt->xpt_flags))
 
 808		len = -ENOTCONN;
 809	else
 810		len = xprt->xpt_ops->xpo_sendto(rqstp);
 811	mutex_unlock(&xprt->xpt_mutex);
 812	rpc_wake_up(&xprt->xpt_bc_pending);
 813	svc_xprt_release(rqstp);
 814
 815	if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
 816		return 0;
 
 
 817	return len;
 818}
 819
 820/*
 821 * Timer function to close old temporary transports, using
 822 * a mark-and-sweep algorithm.
 823 */
 824static void svc_age_temp_xprts(unsigned long closure)
 825{
 826	struct svc_serv *serv = (struct svc_serv *)closure;
 827	struct svc_xprt *xprt;
 828	struct list_head *le, *next;
 829	LIST_HEAD(to_be_aged);
 830
 831	dprintk("svc_age_temp_xprts\n");
 832
 833	if (!spin_trylock_bh(&serv->sv_lock)) {
 834		/* busy, try again 1 sec later */
 835		dprintk("svc_age_temp_xprts: busy\n");
 836		mod_timer(&serv->sv_temptimer, jiffies + HZ);
 837		return;
 838	}
 839
 840	list_for_each_safe(le, next, &serv->sv_tempsocks) {
 841		xprt = list_entry(le, struct svc_xprt, xpt_list);
 842
 843		/* First time through, just mark it OLD. Second time
 844		 * through, close it. */
 845		if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
 846			continue;
 847		if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
 848		    test_bit(XPT_BUSY, &xprt->xpt_flags))
 849			continue;
 850		svc_xprt_get(xprt);
 851		list_move(le, &to_be_aged);
 852		set_bit(XPT_CLOSE, &xprt->xpt_flags);
 853		set_bit(XPT_DETACHED, &xprt->xpt_flags);
 
 
 
 854	}
 855	spin_unlock_bh(&serv->sv_lock);
 856
 857	while (!list_empty(&to_be_aged)) {
 858		le = to_be_aged.next;
 859		/* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
 860		list_del_init(le);
 861		xprt = list_entry(le, struct svc_xprt, xpt_list);
 862
 863		dprintk("queuing xprt %p for closing\n", xprt);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 864
 865		/* a thread will dequeue and close it soon */
 866		svc_xprt_enqueue(xprt);
 867		svc_xprt_put(xprt);
 
 
 
 
 
 868	}
 
 869
 870	mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
 
 
 
 
 
 
 
 
 
 
 871}
 
 872
 873static void call_xpt_users(struct svc_xprt *xprt)
 874{
 875	struct svc_xpt_user *u;
 876
 877	spin_lock(&xprt->xpt_lock);
 878	while (!list_empty(&xprt->xpt_users)) {
 879		u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
 880		list_del(&u->list);
 881		u->callback(u);
 882	}
 883	spin_unlock(&xprt->xpt_lock);
 884}
 885
 886/*
 887 * Remove a dead transport
 888 */
 889void svc_delete_xprt(struct svc_xprt *xprt)
 890{
 891	struct svc_serv	*serv = xprt->xpt_server;
 892	struct svc_deferred_req *dr;
 893
 894	/* Only do this once */
 895	if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
 896		BUG();
 897
 898	dprintk("svc: svc_delete_xprt(%p)\n", xprt);
 899	xprt->xpt_ops->xpo_detach(xprt);
 900
 901	spin_lock_bh(&serv->sv_lock);
 902	if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
 903		list_del_init(&xprt->xpt_list);
 904	/*
 905	 * The only time we're called while xpt_ready is still on a list
 906	 * is while the list itself is about to be destroyed (in
 907	 * svc_destroy).  BUT svc_xprt_enqueue could still be attempting
 908	 * to add new entries to the sp_sockets list, so we can't leave
 909	 * a freed xprt on it.
 910	 */
 911	list_del_init(&xprt->xpt_ready);
 912	if (test_bit(XPT_TEMP, &xprt->xpt_flags))
 913		serv->sv_tmpcnt--;
 914	spin_unlock_bh(&serv->sv_lock);
 915
 916	while ((dr = svc_deferred_dequeue(xprt)) != NULL)
 917		kfree(dr);
 918
 919	call_xpt_users(xprt);
 920	svc_xprt_put(xprt);
 921}
 922
 923void svc_close_xprt(struct svc_xprt *xprt)
 924{
 925	set_bit(XPT_CLOSE, &xprt->xpt_flags);
 926	if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
 927		/* someone else will have to effect the close */
 928		return;
 929	/*
 930	 * We expect svc_close_xprt() to work even when no threads are
 931	 * running (e.g., while configuring the server before starting
 932	 * any threads), so if the transport isn't busy, we delete
 933	 * it ourself:
 934	 */
 935	svc_delete_xprt(xprt);
 936}
 937EXPORT_SYMBOL_GPL(svc_close_xprt);
 938
 939void svc_close_all(struct list_head *xprt_list)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 940{
 
 941	struct svc_xprt *xprt;
 942	struct svc_xprt *tmp;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 943
 944	/*
 945	 * The server is shutting down, and no more threads are running.
 946	 * svc_xprt_enqueue() might still be running, but at worst it
 947	 * will re-add the xprt to sp_sockets, which will soon get
 948	 * freed.  So we don't bother with any more locking, and don't
 949	 * leave the close to the (nonexistent) server threads:
 950	 */
 951	list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
 952		set_bit(XPT_CLOSE, &xprt->xpt_flags);
 953		svc_delete_xprt(xprt);
 954	}
 955}
 956
 957/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 958 * Handle defer and revisit of requests
 959 */
 960
 961static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
 962{
 963	struct svc_deferred_req *dr =
 964		container_of(dreq, struct svc_deferred_req, handle);
 965	struct svc_xprt *xprt = dr->xprt;
 966
 967	spin_lock(&xprt->xpt_lock);
 968	set_bit(XPT_DEFERRED, &xprt->xpt_flags);
 969	if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
 970		spin_unlock(&xprt->xpt_lock);
 971		dprintk("revisit canceled\n");
 972		svc_xprt_put(xprt);
 973		kfree(dr);
 974		return;
 975	}
 976	dprintk("revisit queued\n");
 977	dr->xprt = NULL;
 978	list_add(&dr->handle.recent, &xprt->xpt_deferred);
 979	spin_unlock(&xprt->xpt_lock);
 980	svc_xprt_enqueue(xprt);
 981	svc_xprt_put(xprt);
 982}
 983
 984/*
 985 * Save the request off for later processing. The request buffer looks
 986 * like this:
 987 *
 988 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
 989 *
 990 * This code can only handle requests that consist of an xprt-header
 991 * and rpc-header.
 992 */
 993static struct cache_deferred_req *svc_defer(struct cache_req *req)
 994{
 995	struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
 996	struct svc_deferred_req *dr;
 997
 998	if (rqstp->rq_arg.page_len || !rqstp->rq_usedeferral)
 999		return NULL; /* if more than a page, give up FIXME */
1000	if (rqstp->rq_deferred) {
1001		dr = rqstp->rq_deferred;
1002		rqstp->rq_deferred = NULL;
1003	} else {
1004		size_t skip;
1005		size_t size;
1006		/* FIXME maybe discard if size too large */
1007		size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1008		dr = kmalloc(size, GFP_KERNEL);
1009		if (dr == NULL)
1010			return NULL;
1011
1012		dr->handle.owner = rqstp->rq_server;
1013		dr->prot = rqstp->rq_prot;
1014		memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1015		dr->addrlen = rqstp->rq_addrlen;
1016		dr->daddr = rqstp->rq_daddr;
1017		dr->argslen = rqstp->rq_arg.len >> 2;
1018		dr->xprt_hlen = rqstp->rq_xprt_hlen;
1019
1020		/* back up head to the start of the buffer and copy */
1021		skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1022		memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1023		       dr->argslen << 2);
1024	}
1025	svc_xprt_get(rqstp->rq_xprt);
1026	dr->xprt = rqstp->rq_xprt;
1027	rqstp->rq_dropme = true;
1028
1029	dr->handle.revisit = svc_revisit;
1030	return &dr->handle;
1031}
1032
1033/*
1034 * recv data from a deferred request into an active one
1035 */
1036static int svc_deferred_recv(struct svc_rqst *rqstp)
1037{
1038	struct svc_deferred_req *dr = rqstp->rq_deferred;
1039
1040	/* setup iov_base past transport header */
1041	rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1042	/* The iov_len does not include the transport header bytes */
1043	rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1044	rqstp->rq_arg.page_len = 0;
1045	/* The rq_arg.len includes the transport header bytes */
1046	rqstp->rq_arg.len     = dr->argslen<<2;
1047	rqstp->rq_prot        = dr->prot;
1048	memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1049	rqstp->rq_addrlen     = dr->addrlen;
1050	/* Save off transport header len in case we get deferred again */
1051	rqstp->rq_xprt_hlen   = dr->xprt_hlen;
1052	rqstp->rq_daddr       = dr->daddr;
1053	rqstp->rq_respages    = rqstp->rq_pages;
1054	return (dr->argslen<<2) - dr->xprt_hlen;
1055}
1056
1057
1058static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1059{
1060	struct svc_deferred_req *dr = NULL;
1061
1062	if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1063		return NULL;
1064	spin_lock(&xprt->xpt_lock);
1065	if (!list_empty(&xprt->xpt_deferred)) {
1066		dr = list_entry(xprt->xpt_deferred.next,
1067				struct svc_deferred_req,
1068				handle.recent);
1069		list_del_init(&dr->handle.recent);
1070	} else
1071		clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1072	spin_unlock(&xprt->xpt_lock);
1073	return dr;
1074}
1075
1076/**
1077 * svc_find_xprt - find an RPC transport instance
1078 * @serv: pointer to svc_serv to search
1079 * @xcl_name: C string containing transport's class name
 
1080 * @af: Address family of transport's local address
1081 * @port: transport's IP port number
1082 *
1083 * Return the transport instance pointer for the endpoint accepting
1084 * connections/peer traffic from the specified transport class,
1085 * address family and port.
1086 *
1087 * Specifying 0 for the address family or port is effectively a
1088 * wild-card, and will result in matching the first transport in the
1089 * service's list that has a matching class name.
1090 */
1091struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1092			       const sa_family_t af, const unsigned short port)
 
1093{
1094	struct svc_xprt *xprt;
1095	struct svc_xprt *found = NULL;
1096
1097	/* Sanity check the args */
1098	if (serv == NULL || xcl_name == NULL)
1099		return found;
1100
1101	spin_lock_bh(&serv->sv_lock);
1102	list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
 
 
1103		if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1104			continue;
1105		if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1106			continue;
1107		if (port != 0 && port != svc_xprt_local_port(xprt))
1108			continue;
1109		found = xprt;
1110		svc_xprt_get(xprt);
1111		break;
1112	}
1113	spin_unlock_bh(&serv->sv_lock);
1114	return found;
1115}
1116EXPORT_SYMBOL_GPL(svc_find_xprt);
1117
1118static int svc_one_xprt_name(const struct svc_xprt *xprt,
1119			     char *pos, int remaining)
1120{
1121	int len;
1122
1123	len = snprintf(pos, remaining, "%s %u\n",
1124			xprt->xpt_class->xcl_name,
1125			svc_xprt_local_port(xprt));
1126	if (len >= remaining)
1127		return -ENAMETOOLONG;
1128	return len;
1129}
1130
1131/**
1132 * svc_xprt_names - format a buffer with a list of transport names
1133 * @serv: pointer to an RPC service
1134 * @buf: pointer to a buffer to be filled in
1135 * @buflen: length of buffer to be filled in
1136 *
1137 * Fills in @buf with a string containing a list of transport names,
1138 * each name terminated with '\n'.
1139 *
1140 * Returns positive length of the filled-in string on success; otherwise
1141 * a negative errno value is returned if an error occurs.
1142 */
1143int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1144{
1145	struct svc_xprt *xprt;
1146	int len, totlen;
1147	char *pos;
1148
1149	/* Sanity check args */
1150	if (!serv)
1151		return 0;
1152
1153	spin_lock_bh(&serv->sv_lock);
1154
1155	pos = buf;
1156	totlen = 0;
1157	list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1158		len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1159		if (len < 0) {
1160			*buf = '\0';
1161			totlen = len;
1162		}
1163		if (len <= 0)
1164			break;
1165
1166		pos += len;
1167		totlen += len;
1168	}
1169
1170	spin_unlock_bh(&serv->sv_lock);
1171	return totlen;
1172}
1173EXPORT_SYMBOL_GPL(svc_xprt_names);
1174
1175
1176/*----------------------------------------------------------------------------*/
1177
1178static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1179{
1180	unsigned int pidx = (unsigned int)*pos;
1181	struct svc_serv *serv = m->private;
1182
1183	dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1184
1185	if (!pidx)
1186		return SEQ_START_TOKEN;
1187	return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1188}
1189
1190static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1191{
1192	struct svc_pool *pool = p;
1193	struct svc_serv *serv = m->private;
1194
1195	dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1196
1197	if (p == SEQ_START_TOKEN) {
1198		pool = &serv->sv_pools[0];
1199	} else {
1200		unsigned int pidx = (pool - &serv->sv_pools[0]);
1201		if (pidx < serv->sv_nrpools-1)
1202			pool = &serv->sv_pools[pidx+1];
1203		else
1204			pool = NULL;
1205	}
1206	++*pos;
1207	return pool;
1208}
1209
1210static void svc_pool_stats_stop(struct seq_file *m, void *p)
1211{
1212}
1213
1214static int svc_pool_stats_show(struct seq_file *m, void *p)
1215{
1216	struct svc_pool *pool = p;
1217
1218	if (p == SEQ_START_TOKEN) {
1219		seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1220		return 0;
1221	}
1222
1223	seq_printf(m, "%u %lu %lu %lu %lu\n",
1224		pool->sp_id,
1225		pool->sp_stats.packets,
1226		pool->sp_stats.sockets_queued,
1227		pool->sp_stats.threads_woken,
1228		pool->sp_stats.threads_timedout);
1229
1230	return 0;
1231}
1232
1233static const struct seq_operations svc_pool_stats_seq_ops = {
1234	.start	= svc_pool_stats_start,
1235	.next	= svc_pool_stats_next,
1236	.stop	= svc_pool_stats_stop,
1237	.show	= svc_pool_stats_show,
1238};
1239
1240int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1241{
1242	int err;
1243
1244	err = seq_open(file, &svc_pool_stats_seq_ops);
1245	if (!err)
1246		((struct seq_file *) file->private_data)->private = serv;
1247	return err;
1248}
1249EXPORT_SYMBOL(svc_pool_stats_open);
1250
1251/*----------------------------------------------------------------------------*/