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