linux/net/sunrpc/svc_xprt.c
<<
>>
Prefs
   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 <net/sock.h>
  12#include <linux/sunrpc/stats.h>
  13#include <linux/sunrpc/svc_xprt.h>
  14
  15#define RPCDBG_FACILITY RPCDBG_SVCXPRT
  16
  17#define SVC_MAX_WAKING 5
  18
  19static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
  20static int svc_deferred_recv(struct svc_rqst *rqstp);
  21static struct cache_deferred_req *svc_defer(struct cache_req *req);
  22static void svc_age_temp_xprts(unsigned long closure);
  23
  24/* apparently the "standard" is that clients close
  25 * idle connections after 5 minutes, servers after
  26 * 6 minutes
  27 *   http://www.connectathon.org/talks96/nfstcp.pdf
  28 */
  29static int svc_conn_age_period = 6*60;
  30
  31/* List of registered transport classes */
  32static DEFINE_SPINLOCK(svc_xprt_class_lock);
  33static LIST_HEAD(svc_xprt_class_list);
  34
  35/* SMP locking strategy:
  36 *
  37 *      svc_pool->sp_lock protects most of the fields of that pool.
  38 *      svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
  39 *      when both need to be taken (rare), svc_serv->sv_lock is first.
  40 *      BKL protects svc_serv->sv_nrthread.
  41 *      svc_sock->sk_lock protects the svc_sock->sk_deferred list
  42 *             and the ->sk_info_authunix cache.
  43 *
  44 *      The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
  45 *      enqueued multiply. During normal transport processing this bit
  46 *      is set by svc_xprt_enqueue and cleared by svc_xprt_received.
  47 *      Providers should not manipulate this bit directly.
  48 *
  49 *      Some flags can be set to certain values at any time
  50 *      providing that certain rules are followed:
  51 *
  52 *      XPT_CONN, XPT_DATA:
  53 *              - Can be set or cleared at any time.
  54 *              - After a set, svc_xprt_enqueue must be called to enqueue
  55 *                the transport for processing.
  56 *              - After a clear, the transport must be read/accepted.
  57 *                If this succeeds, it must be set again.
  58 *      XPT_CLOSE:
  59 *              - Can set at any time. It is never cleared.
  60 *      XPT_DEAD:
  61 *              - Can only be set while XPT_BUSY is held which ensures
  62 *                that no other thread will be using the transport or will
  63 *                try to set XPT_DEAD.
  64 */
  65
  66int svc_reg_xprt_class(struct svc_xprt_class *xcl)
  67{
  68        struct svc_xprt_class *cl;
  69        int res = -EEXIST;
  70
  71        dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
  72
  73        INIT_LIST_HEAD(&xcl->xcl_list);
  74        spin_lock(&svc_xprt_class_lock);
  75        /* Make sure there isn't already a class with the same name */
  76        list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
  77                if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
  78                        goto out;
  79        }
  80        list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
  81        res = 0;
  82out:
  83        spin_unlock(&svc_xprt_class_lock);
  84        return res;
  85}
  86EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
  87
  88void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
  89{
  90        dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
  91        spin_lock(&svc_xprt_class_lock);
  92        list_del_init(&xcl->xcl_list);
  93        spin_unlock(&svc_xprt_class_lock);
  94}
  95EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
  96
  97/*
  98 * Format the transport list for printing
  99 */
 100int svc_print_xprts(char *buf, int maxlen)
 101{
 102        struct list_head *le;
 103        char tmpstr[80];
 104        int len = 0;
 105        buf[0] = '\0';
 106
 107        spin_lock(&svc_xprt_class_lock);
 108        list_for_each(le, &svc_xprt_class_list) {
 109                int slen;
 110                struct svc_xprt_class *xcl =
 111                        list_entry(le, struct svc_xprt_class, xcl_list);
 112
 113                sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
 114                slen = strlen(tmpstr);
 115                if (len + slen > maxlen)
 116                        break;
 117                len += slen;
 118                strcat(buf, tmpstr);
 119        }
 120        spin_unlock(&svc_xprt_class_lock);
 121
 122        return len;
 123}
 124
 125static void svc_xprt_free(struct kref *kref)
 126{
 127        struct svc_xprt *xprt =
 128                container_of(kref, struct svc_xprt, xpt_ref);
 129        struct module *owner = xprt->xpt_class->xcl_owner;
 130        if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags)
 131            && xprt->xpt_auth_cache != NULL)
 132                svcauth_unix_info_release(xprt->xpt_auth_cache);
 133        xprt->xpt_ops->xpo_free(xprt);
 134        module_put(owner);
 135}
 136
 137void svc_xprt_put(struct svc_xprt *xprt)
 138{
 139        kref_put(&xprt->xpt_ref, svc_xprt_free);
 140}
 141EXPORT_SYMBOL_GPL(svc_xprt_put);
 142
 143/*
 144 * Called by transport drivers to initialize the transport independent
 145 * portion of the transport instance.
 146 */
 147void svc_xprt_init(struct svc_xprt_class *xcl, struct svc_xprt *xprt,
 148                   struct svc_serv *serv)
 149{
 150        memset(xprt, 0, sizeof(*xprt));
 151        xprt->xpt_class = xcl;
 152        xprt->xpt_ops = xcl->xcl_ops;
 153        kref_init(&xprt->xpt_ref);
 154        xprt->xpt_server = serv;
 155        INIT_LIST_HEAD(&xprt->xpt_list);
 156        INIT_LIST_HEAD(&xprt->xpt_ready);
 157        INIT_LIST_HEAD(&xprt->xpt_deferred);
 158        mutex_init(&xprt->xpt_mutex);
 159        spin_lock_init(&xprt->xpt_lock);
 160        set_bit(XPT_BUSY, &xprt->xpt_flags);
 161}
 162EXPORT_SYMBOL_GPL(svc_xprt_init);
 163
 164static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
 165                                         struct svc_serv *serv,
 166                                         const int family,
 167                                         const unsigned short port,
 168                                         int flags)
 169{
 170        struct sockaddr_in sin = {
 171                .sin_family             = AF_INET,
 172                .sin_addr.s_addr        = htonl(INADDR_ANY),
 173                .sin_port               = htons(port),
 174        };
 175        struct sockaddr_in6 sin6 = {
 176                .sin6_family            = AF_INET6,
 177                .sin6_addr              = IN6ADDR_ANY_INIT,
 178                .sin6_port              = htons(port),
 179        };
 180        struct sockaddr *sap;
 181        size_t len;
 182
 183        switch (family) {
 184        case PF_INET:
 185                sap = (struct sockaddr *)&sin;
 186                len = sizeof(sin);
 187                break;
 188        case PF_INET6:
 189                sap = (struct sockaddr *)&sin6;
 190                len = sizeof(sin6);
 191                break;
 192        default:
 193                return ERR_PTR(-EAFNOSUPPORT);
 194        }
 195
 196        return xcl->xcl_ops->xpo_create(serv, sap, len, flags);
 197}
 198
 199int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
 200                    const int family, const unsigned short port,
 201                    int flags)
 202{
 203        struct svc_xprt_class *xcl;
 204
 205        dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
 206        spin_lock(&svc_xprt_class_lock);
 207        list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
 208                struct svc_xprt *newxprt;
 209
 210                if (strcmp(xprt_name, xcl->xcl_name))
 211                        continue;
 212
 213                if (!try_module_get(xcl->xcl_owner))
 214                        goto err;
 215
 216                spin_unlock(&svc_xprt_class_lock);
 217                newxprt = __svc_xpo_create(xcl, serv, family, port, flags);
 218                if (IS_ERR(newxprt)) {
 219                        module_put(xcl->xcl_owner);
 220                        return PTR_ERR(newxprt);
 221                }
 222
 223                clear_bit(XPT_TEMP, &newxprt->xpt_flags);
 224                spin_lock_bh(&serv->sv_lock);
 225                list_add(&newxprt->xpt_list, &serv->sv_permsocks);
 226                spin_unlock_bh(&serv->sv_lock);
 227                clear_bit(XPT_BUSY, &newxprt->xpt_flags);
 228                return svc_xprt_local_port(newxprt);
 229        }
 230 err:
 231        spin_unlock(&svc_xprt_class_lock);
 232        dprintk("svc: transport %s not found\n", xprt_name);
 233        return -ENOENT;
 234}
 235EXPORT_SYMBOL_GPL(svc_create_xprt);
 236
 237/*
 238 * Copy the local and remote xprt addresses to the rqstp structure
 239 */
 240void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
 241{
 242        struct sockaddr *sin;
 243
 244        memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
 245        rqstp->rq_addrlen = xprt->xpt_remotelen;
 246
 247        /*
 248         * Destination address in request is needed for binding the
 249         * source address in RPC replies/callbacks later.
 250         */
 251        sin = (struct sockaddr *)&xprt->xpt_local;
 252        switch (sin->sa_family) {
 253        case AF_INET:
 254                rqstp->rq_daddr.addr = ((struct sockaddr_in *)sin)->sin_addr;
 255                break;
 256        case AF_INET6:
 257                rqstp->rq_daddr.addr6 = ((struct sockaddr_in6 *)sin)->sin6_addr;
 258                break;
 259        }
 260}
 261EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
 262
 263/**
 264 * svc_print_addr - Format rq_addr field for printing
 265 * @rqstp: svc_rqst struct containing address to print
 266 * @buf: target buffer for formatted address
 267 * @len: length of target buffer
 268 *
 269 */
 270char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
 271{
 272        return __svc_print_addr(svc_addr(rqstp), buf, len);
 273}
 274EXPORT_SYMBOL_GPL(svc_print_addr);
 275
 276/*
 277 * Queue up an idle server thread.  Must have pool->sp_lock held.
 278 * Note: this is really a stack rather than a queue, so that we only
 279 * use as many different threads as we need, and the rest don't pollute
 280 * the cache.
 281 */
 282static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
 283{
 284        list_add(&rqstp->rq_list, &pool->sp_threads);
 285}
 286
 287/*
 288 * Dequeue an nfsd thread.  Must have pool->sp_lock held.
 289 */
 290static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
 291{
 292        list_del(&rqstp->rq_list);
 293}
 294
 295/*
 296 * Queue up a transport with data pending. If there are idle nfsd
 297 * processes, wake 'em up.
 298 *
 299 */
 300void svc_xprt_enqueue(struct svc_xprt *xprt)
 301{
 302        struct svc_serv *serv = xprt->xpt_server;
 303        struct svc_pool *pool;
 304        struct svc_rqst *rqstp;
 305        int cpu;
 306        int thread_avail;
 307
 308        if (!(xprt->xpt_flags &
 309              ((1<<XPT_CONN)|(1<<XPT_DATA)|(1<<XPT_CLOSE)|(1<<XPT_DEFERRED))))
 310                return;
 311
 312        cpu = get_cpu();
 313        pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
 314        put_cpu();
 315
 316        spin_lock_bh(&pool->sp_lock);
 317
 318        if (test_bit(XPT_DEAD, &xprt->xpt_flags)) {
 319                /* Don't enqueue dead transports */
 320                dprintk("svc: transport %p is dead, not enqueued\n", xprt);
 321                goto out_unlock;
 322        }
 323
 324        pool->sp_stats.packets++;
 325
 326        /* Mark transport as busy. It will remain in this state until
 327         * the provider calls svc_xprt_received. We update XPT_BUSY
 328         * atomically because it also guards against trying to enqueue
 329         * the transport twice.
 330         */
 331        if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
 332                /* Don't enqueue transport while already enqueued */
 333                dprintk("svc: transport %p busy, not enqueued\n", xprt);
 334                goto out_unlock;
 335        }
 336        BUG_ON(xprt->xpt_pool != NULL);
 337        xprt->xpt_pool = pool;
 338
 339        /* Handle pending connection */
 340        if (test_bit(XPT_CONN, &xprt->xpt_flags))
 341                goto process;
 342
 343        /* Handle close in-progress */
 344        if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
 345                goto process;
 346
 347        /* Check if we have space to reply to a request */
 348        if (!xprt->xpt_ops->xpo_has_wspace(xprt)) {
 349                /* Don't enqueue while not enough space for reply */
 350                dprintk("svc: no write space, transport %p  not enqueued\n",
 351                        xprt);
 352                xprt->xpt_pool = NULL;
 353                clear_bit(XPT_BUSY, &xprt->xpt_flags);
 354                goto out_unlock;
 355        }
 356
 357 process:
 358        /* Work out whether threads are available */
 359        thread_avail = !list_empty(&pool->sp_threads);  /* threads are asleep */
 360        if (pool->sp_nwaking >= SVC_MAX_WAKING) {
 361                /* too many threads are runnable and trying to wake up */
 362                thread_avail = 0;
 363                pool->sp_stats.overloads_avoided++;
 364        }
 365
 366        if (thread_avail) {
 367                rqstp = list_entry(pool->sp_threads.next,
 368                                   struct svc_rqst,
 369                                   rq_list);
 370                dprintk("svc: transport %p served by daemon %p\n",
 371                        xprt, rqstp);
 372                svc_thread_dequeue(pool, rqstp);
 373                if (rqstp->rq_xprt)
 374                        printk(KERN_ERR
 375                                "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
 376                                rqstp, rqstp->rq_xprt);
 377                rqstp->rq_xprt = xprt;
 378                svc_xprt_get(xprt);
 379                rqstp->rq_reserved = serv->sv_max_mesg;
 380                atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
 381                rqstp->rq_waking = 1;
 382                pool->sp_nwaking++;
 383                pool->sp_stats.threads_woken++;
 384                BUG_ON(xprt->xpt_pool != pool);
 385                wake_up(&rqstp->rq_wait);
 386        } else {
 387                dprintk("svc: transport %p put into queue\n", xprt);
 388                list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
 389                pool->sp_stats.sockets_queued++;
 390                BUG_ON(xprt->xpt_pool != pool);
 391        }
 392
 393out_unlock:
 394        spin_unlock_bh(&pool->sp_lock);
 395}
 396EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
 397
 398/*
 399 * Dequeue the first transport.  Must be called with the pool->sp_lock held.
 400 */
 401static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
 402{
 403        struct svc_xprt *xprt;
 404
 405        if (list_empty(&pool->sp_sockets))
 406                return NULL;
 407
 408        xprt = list_entry(pool->sp_sockets.next,
 409                          struct svc_xprt, xpt_ready);
 410        list_del_init(&xprt->xpt_ready);
 411
 412        dprintk("svc: transport %p dequeued, inuse=%d\n",
 413                xprt, atomic_read(&xprt->xpt_ref.refcount));
 414
 415        return xprt;
 416}
 417
 418/*
 419 * svc_xprt_received conditionally queues the transport for processing
 420 * by another thread. The caller must hold the XPT_BUSY bit and must
 421 * not thereafter touch transport data.
 422 *
 423 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
 424 * insufficient) data.
 425 */
 426void svc_xprt_received(struct svc_xprt *xprt)
 427{
 428        BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
 429        xprt->xpt_pool = NULL;
 430        clear_bit(XPT_BUSY, &xprt->xpt_flags);
 431        svc_xprt_enqueue(xprt);
 432}
 433EXPORT_SYMBOL_GPL(svc_xprt_received);
 434
 435/**
 436 * svc_reserve - change the space reserved for the reply to a request.
 437 * @rqstp:  The request in question
 438 * @space: new max space to reserve
 439 *
 440 * Each request reserves some space on the output queue of the transport
 441 * to make sure the reply fits.  This function reduces that reserved
 442 * space to be the amount of space used already, plus @space.
 443 *
 444 */
 445void svc_reserve(struct svc_rqst *rqstp, int space)
 446{
 447        space += rqstp->rq_res.head[0].iov_len;
 448
 449        if (space < rqstp->rq_reserved) {
 450                struct svc_xprt *xprt = rqstp->rq_xprt;
 451                atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
 452                rqstp->rq_reserved = space;
 453
 454                svc_xprt_enqueue(xprt);
 455        }
 456}
 457EXPORT_SYMBOL_GPL(svc_reserve);
 458
 459static void svc_xprt_release(struct svc_rqst *rqstp)
 460{
 461        struct svc_xprt *xprt = rqstp->rq_xprt;
 462
 463        rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
 464
 465        kfree(rqstp->rq_deferred);
 466        rqstp->rq_deferred = NULL;
 467
 468        svc_free_res_pages(rqstp);
 469        rqstp->rq_res.page_len = 0;
 470        rqstp->rq_res.page_base = 0;
 471
 472        /* Reset response buffer and release
 473         * the reservation.
 474         * But first, check that enough space was reserved
 475         * for the reply, otherwise we have a bug!
 476         */
 477        if ((rqstp->rq_res.len) >  rqstp->rq_reserved)
 478                printk(KERN_ERR "RPC request reserved %d but used %d\n",
 479                       rqstp->rq_reserved,
 480                       rqstp->rq_res.len);
 481
 482        rqstp->rq_res.head[0].iov_len = 0;
 483        svc_reserve(rqstp, 0);
 484        rqstp->rq_xprt = NULL;
 485
 486        svc_xprt_put(xprt);
 487}
 488
 489/*
 490 * External function to wake up a server waiting for data
 491 * This really only makes sense for services like lockd
 492 * which have exactly one thread anyway.
 493 */
 494void svc_wake_up(struct svc_serv *serv)
 495{
 496        struct svc_rqst *rqstp;
 497        unsigned int i;
 498        struct svc_pool *pool;
 499
 500        for (i = 0; i < serv->sv_nrpools; i++) {
 501                pool = &serv->sv_pools[i];
 502
 503                spin_lock_bh(&pool->sp_lock);
 504                if (!list_empty(&pool->sp_threads)) {
 505                        rqstp = list_entry(pool->sp_threads.next,
 506                                           struct svc_rqst,
 507                                           rq_list);
 508                        dprintk("svc: daemon %p woken up.\n", rqstp);
 509                        /*
 510                        svc_thread_dequeue(pool, rqstp);
 511                        rqstp->rq_xprt = NULL;
 512                         */
 513                        wake_up(&rqstp->rq_wait);
 514                }
 515                spin_unlock_bh(&pool->sp_lock);
 516        }
 517}
 518EXPORT_SYMBOL_GPL(svc_wake_up);
 519
 520int svc_port_is_privileged(struct sockaddr *sin)
 521{
 522        switch (sin->sa_family) {
 523        case AF_INET:
 524                return ntohs(((struct sockaddr_in *)sin)->sin_port)
 525                        < PROT_SOCK;
 526        case AF_INET6:
 527                return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
 528                        < PROT_SOCK;
 529        default:
 530                return 0;
 531        }
 532}
 533
 534/*
 535 * Make sure that we don't have too many active connections. If we have,
 536 * something must be dropped. It's not clear what will happen if we allow
 537 * "too many" connections, but when dealing with network-facing software,
 538 * we have to code defensively. Here we do that by imposing hard limits.
 539 *
 540 * There's no point in trying to do random drop here for DoS
 541 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
 542 * attacker can easily beat that.
 543 *
 544 * The only somewhat efficient mechanism would be if drop old
 545 * connections from the same IP first. But right now we don't even
 546 * record the client IP in svc_sock.
 547 *
 548 * single-threaded services that expect a lot of clients will probably
 549 * need to set sv_maxconn to override the default value which is based
 550 * on the number of threads
 551 */
 552static void svc_check_conn_limits(struct svc_serv *serv)
 553{
 554        unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
 555                                (serv->sv_nrthreads+3) * 20;
 556
 557        if (serv->sv_tmpcnt > limit) {
 558                struct svc_xprt *xprt = NULL;
 559                spin_lock_bh(&serv->sv_lock);
 560                if (!list_empty(&serv->sv_tempsocks)) {
 561                        if (net_ratelimit()) {
 562                                /* Try to help the admin */
 563                                printk(KERN_NOTICE "%s: too many open  "
 564                                       "connections, consider increasing %s\n",
 565                                       serv->sv_name, serv->sv_maxconn ?
 566                                       "the max number of connections." :
 567                                       "the number of threads.");
 568                        }
 569                        /*
 570                         * Always select the oldest connection. It's not fair,
 571                         * but so is life
 572                         */
 573                        xprt = list_entry(serv->sv_tempsocks.prev,
 574                                          struct svc_xprt,
 575                                          xpt_list);
 576                        set_bit(XPT_CLOSE, &xprt->xpt_flags);
 577                        svc_xprt_get(xprt);
 578                }
 579                spin_unlock_bh(&serv->sv_lock);
 580
 581                if (xprt) {
 582                        svc_xprt_enqueue(xprt);
 583                        svc_xprt_put(xprt);
 584                }
 585        }
 586}
 587
 588/*
 589 * Receive the next request on any transport.  This code is carefully
 590 * organised not to touch any cachelines in the shared svc_serv
 591 * structure, only cachelines in the local svc_pool.
 592 */
 593int svc_recv(struct svc_rqst *rqstp, long timeout)
 594{
 595        struct svc_xprt         *xprt = NULL;
 596        struct svc_serv         *serv = rqstp->rq_server;
 597        struct svc_pool         *pool = rqstp->rq_pool;
 598        int                     len, i;
 599        int                     pages;
 600        struct xdr_buf          *arg;
 601        DECLARE_WAITQUEUE(wait, current);
 602        long                    time_left;
 603
 604        dprintk("svc: server %p waiting for data (to = %ld)\n",
 605                rqstp, timeout);
 606
 607        if (rqstp->rq_xprt)
 608                printk(KERN_ERR
 609                        "svc_recv: service %p, transport not NULL!\n",
 610                         rqstp);
 611        if (waitqueue_active(&rqstp->rq_wait))
 612                printk(KERN_ERR
 613                        "svc_recv: service %p, wait queue active!\n",
 614                         rqstp);
 615
 616        /* now allocate needed pages.  If we get a failure, sleep briefly */
 617        pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
 618        for (i = 0; i < pages ; i++)
 619                while (rqstp->rq_pages[i] == NULL) {
 620                        struct page *p = alloc_page(GFP_KERNEL);
 621                        if (!p) {
 622                                set_current_state(TASK_INTERRUPTIBLE);
 623                                if (signalled() || kthread_should_stop()) {
 624                                        set_current_state(TASK_RUNNING);
 625                                        return -EINTR;
 626                                }
 627                                schedule_timeout(msecs_to_jiffies(500));
 628                        }
 629                        rqstp->rq_pages[i] = p;
 630                }
 631        rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
 632        BUG_ON(pages >= RPCSVC_MAXPAGES);
 633
 634        /* Make arg->head point to first page and arg->pages point to rest */
 635        arg = &rqstp->rq_arg;
 636        arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
 637        arg->head[0].iov_len = PAGE_SIZE;
 638        arg->pages = rqstp->rq_pages + 1;
 639        arg->page_base = 0;
 640        /* save at least one page for response */
 641        arg->page_len = (pages-2)*PAGE_SIZE;
 642        arg->len = (pages-1)*PAGE_SIZE;
 643        arg->tail[0].iov_len = 0;
 644
 645        try_to_freeze();
 646        cond_resched();
 647        if (signalled() || kthread_should_stop())
 648                return -EINTR;
 649
 650        spin_lock_bh(&pool->sp_lock);
 651        if (rqstp->rq_waking) {
 652                rqstp->rq_waking = 0;
 653                pool->sp_nwaking--;
 654                BUG_ON(pool->sp_nwaking < 0);
 655        }
 656        xprt = svc_xprt_dequeue(pool);
 657        if (xprt) {
 658                rqstp->rq_xprt = xprt;
 659                svc_xprt_get(xprt);
 660                rqstp->rq_reserved = serv->sv_max_mesg;
 661                atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
 662        } else {
 663                /* No data pending. Go to sleep */
 664                svc_thread_enqueue(pool, rqstp);
 665
 666                /*
 667                 * We have to be able to interrupt this wait
 668                 * to bring down the daemons ...
 669                 */
 670                set_current_state(TASK_INTERRUPTIBLE);
 671
 672                /*
 673                 * checking kthread_should_stop() here allows us to avoid
 674                 * locking and signalling when stopping kthreads that call
 675                 * svc_recv. If the thread has already been woken up, then
 676                 * we can exit here without sleeping. If not, then it
 677                 * it'll be woken up quickly during the schedule_timeout
 678                 */
 679                if (kthread_should_stop()) {
 680                        set_current_state(TASK_RUNNING);
 681                        spin_unlock_bh(&pool->sp_lock);
 682                        return -EINTR;
 683                }
 684
 685                add_wait_queue(&rqstp->rq_wait, &wait);
 686                spin_unlock_bh(&pool->sp_lock);
 687
 688                time_left = schedule_timeout(timeout);
 689
 690                try_to_freeze();
 691
 692                spin_lock_bh(&pool->sp_lock);
 693                remove_wait_queue(&rqstp->rq_wait, &wait);
 694                if (!time_left)
 695                        pool->sp_stats.threads_timedout++;
 696
 697                xprt = rqstp->rq_xprt;
 698                if (!xprt) {
 699                        svc_thread_dequeue(pool, rqstp);
 700                        spin_unlock_bh(&pool->sp_lock);
 701                        dprintk("svc: server %p, no data yet\n", rqstp);
 702                        if (signalled() || kthread_should_stop())
 703                                return -EINTR;
 704                        else
 705                                return -EAGAIN;
 706                }
 707        }
 708        spin_unlock_bh(&pool->sp_lock);
 709
 710        len = 0;
 711        if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
 712                dprintk("svc_recv: found XPT_CLOSE\n");
 713                svc_delete_xprt(xprt);
 714        } else 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 {
 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        /* No data, incomplete (TCP) read, or accept() */
 754        if (len == 0 || len == -EAGAIN) {
 755                rqstp->rq_res.len = 0;
 756                svc_xprt_release(rqstp);
 757                return -EAGAIN;
 758        }
 759        clear_bit(XPT_OLD, &xprt->xpt_flags);
 760
 761        rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
 762        rqstp->rq_chandle.defer = svc_defer;
 763
 764        if (serv->sv_stats)
 765                serv->sv_stats->netcnt++;
 766        return len;
 767}
 768EXPORT_SYMBOL_GPL(svc_recv);
 769
 770/*
 771 * Drop request
 772 */
 773void svc_drop(struct svc_rqst *rqstp)
 774{
 775        dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
 776        svc_xprt_release(rqstp);
 777}
 778EXPORT_SYMBOL_GPL(svc_drop);
 779
 780/*
 781 * Return reply to client.
 782 */
 783int svc_send(struct svc_rqst *rqstp)
 784{
 785        struct svc_xprt *xprt;
 786        int             len;
 787        struct xdr_buf  *xb;
 788
 789        xprt = rqstp->rq_xprt;
 790        if (!xprt)
 791                return -EFAULT;
 792
 793        /* release the receive skb before sending the reply */
 794        rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
 795
 796        /* calculate over-all length */
 797        xb = &rqstp->rq_res;
 798        xb->len = xb->head[0].iov_len +
 799                xb->page_len +
 800                xb->tail[0].iov_len;
 801
 802        /* Grab mutex to serialize outgoing data. */
 803        mutex_lock(&xprt->xpt_mutex);
 804        if (test_bit(XPT_DEAD, &xprt->xpt_flags))
 805                len = -ENOTCONN;
 806        else
 807                len = xprt->xpt_ops->xpo_sendto(rqstp);
 808        mutex_unlock(&xprt->xpt_mutex);
 809        svc_xprt_release(rqstp);
 810
 811        if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
 812                return 0;
 813        return len;
 814}
 815
 816/*
 817 * Timer function to close old temporary transports, using
 818 * a mark-and-sweep algorithm.
 819 */
 820static void svc_age_temp_xprts(unsigned long closure)
 821{
 822        struct svc_serv *serv = (struct svc_serv *)closure;
 823        struct svc_xprt *xprt;
 824        struct list_head *le, *next;
 825        LIST_HEAD(to_be_aged);
 826
 827        dprintk("svc_age_temp_xprts\n");
 828
 829        if (!spin_trylock_bh(&serv->sv_lock)) {
 830                /* busy, try again 1 sec later */
 831                dprintk("svc_age_temp_xprts: busy\n");
 832                mod_timer(&serv->sv_temptimer, jiffies + HZ);
 833                return;
 834        }
 835
 836        list_for_each_safe(le, next, &serv->sv_tempsocks) {
 837                xprt = list_entry(le, struct svc_xprt, xpt_list);
 838
 839                /* First time through, just mark it OLD. Second time
 840                 * through, close it. */
 841                if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
 842                        continue;
 843                if (atomic_read(&xprt->xpt_ref.refcount) > 1
 844                    || test_bit(XPT_BUSY, &xprt->xpt_flags))
 845                        continue;
 846                svc_xprt_get(xprt);
 847                list_move(le, &to_be_aged);
 848                set_bit(XPT_CLOSE, &xprt->xpt_flags);
 849                set_bit(XPT_DETACHED, &xprt->xpt_flags);
 850        }
 851        spin_unlock_bh(&serv->sv_lock);
 852
 853        while (!list_empty(&to_be_aged)) {
 854                le = to_be_aged.next;
 855                /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
 856                list_del_init(le);
 857                xprt = list_entry(le, struct svc_xprt, xpt_list);
 858
 859                dprintk("queuing xprt %p for closing\n", xprt);
 860
 861                /* a thread will dequeue and close it soon */
 862                svc_xprt_enqueue(xprt);
 863                svc_xprt_put(xprt);
 864        }
 865
 866        mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
 867}
 868
 869/*
 870 * Remove a dead transport
 871 */
 872void svc_delete_xprt(struct svc_xprt *xprt)
 873{
 874        struct svc_serv *serv = xprt->xpt_server;
 875        struct svc_deferred_req *dr;
 876
 877        /* Only do this once */
 878        if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
 879                return;
 880
 881        dprintk("svc: svc_delete_xprt(%p)\n", xprt);
 882        xprt->xpt_ops->xpo_detach(xprt);
 883
 884        spin_lock_bh(&serv->sv_lock);
 885        if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
 886                list_del_init(&xprt->xpt_list);
 887        /*
 888         * We used to delete the transport from whichever list
 889         * it's sk_xprt.xpt_ready node was on, but we don't actually
 890         * need to.  This is because the only time we're called
 891         * while still attached to a queue, the queue itself
 892         * is about to be destroyed (in svc_destroy).
 893         */
 894        if (test_bit(XPT_TEMP, &xprt->xpt_flags))
 895                serv->sv_tmpcnt--;
 896
 897        for (dr = svc_deferred_dequeue(xprt); dr;
 898             dr = svc_deferred_dequeue(xprt)) {
 899                svc_xprt_put(xprt);
 900                kfree(dr);
 901        }
 902
 903        svc_xprt_put(xprt);
 904        spin_unlock_bh(&serv->sv_lock);
 905}
 906
 907void svc_close_xprt(struct svc_xprt *xprt)
 908{
 909        set_bit(XPT_CLOSE, &xprt->xpt_flags);
 910        if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
 911                /* someone else will have to effect the close */
 912                return;
 913
 914        svc_xprt_get(xprt);
 915        svc_delete_xprt(xprt);
 916        clear_bit(XPT_BUSY, &xprt->xpt_flags);
 917        svc_xprt_put(xprt);
 918}
 919EXPORT_SYMBOL_GPL(svc_close_xprt);
 920
 921void svc_close_all(struct list_head *xprt_list)
 922{
 923        struct svc_xprt *xprt;
 924        struct svc_xprt *tmp;
 925
 926        list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
 927                set_bit(XPT_CLOSE, &xprt->xpt_flags);
 928                if (test_bit(XPT_BUSY, &xprt->xpt_flags)) {
 929                        /* Waiting to be processed, but no threads left,
 930                         * So just remove it from the waiting list
 931                         */
 932                        list_del_init(&xprt->xpt_ready);
 933                        clear_bit(XPT_BUSY, &xprt->xpt_flags);
 934                }
 935                svc_close_xprt(xprt);
 936        }
 937}
 938
 939/*
 940 * Handle defer and revisit of requests
 941 */
 942
 943static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
 944{
 945        struct svc_deferred_req *dr =
 946                container_of(dreq, struct svc_deferred_req, handle);
 947        struct svc_xprt *xprt = dr->xprt;
 948
 949        spin_lock(&xprt->xpt_lock);
 950        set_bit(XPT_DEFERRED, &xprt->xpt_flags);
 951        if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
 952                spin_unlock(&xprt->xpt_lock);
 953                dprintk("revisit canceled\n");
 954                svc_xprt_put(xprt);
 955                kfree(dr);
 956                return;
 957        }
 958        dprintk("revisit queued\n");
 959        dr->xprt = NULL;
 960        list_add(&dr->handle.recent, &xprt->xpt_deferred);
 961        spin_unlock(&xprt->xpt_lock);
 962        svc_xprt_enqueue(xprt);
 963        svc_xprt_put(xprt);
 964}
 965
 966/*
 967 * Save the request off for later processing. The request buffer looks
 968 * like this:
 969 *
 970 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
 971 *
 972 * This code can only handle requests that consist of an xprt-header
 973 * and rpc-header.
 974 */
 975static struct cache_deferred_req *svc_defer(struct cache_req *req)
 976{
 977        struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
 978        struct svc_deferred_req *dr;
 979
 980        if (rqstp->rq_arg.page_len || !rqstp->rq_usedeferral)
 981                return NULL; /* if more than a page, give up FIXME */
 982        if (rqstp->rq_deferred) {
 983                dr = rqstp->rq_deferred;
 984                rqstp->rq_deferred = NULL;
 985        } else {
 986                size_t skip;
 987                size_t size;
 988                /* FIXME maybe discard if size too large */
 989                size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
 990                dr = kmalloc(size, GFP_KERNEL);
 991                if (dr == NULL)
 992                        return NULL;
 993
 994                dr->handle.owner = rqstp->rq_server;
 995                dr->prot = rqstp->rq_prot;
 996                memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
 997                dr->addrlen = rqstp->rq_addrlen;
 998                dr->daddr = rqstp->rq_daddr;
 999                dr->argslen = rqstp->rq_arg.len >> 2;
1000                dr->xprt_hlen = rqstp->rq_xprt_hlen;
1001
1002                /* back up head to the start of the buffer and copy */
1003                skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1004                memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1005                       dr->argslen << 2);
1006        }
1007        svc_xprt_get(rqstp->rq_xprt);
1008        dr->xprt = rqstp->rq_xprt;
1009
1010        dr->handle.revisit = svc_revisit;
1011        return &dr->handle;
1012}
1013
1014/*
1015 * recv data from a deferred request into an active one
1016 */
1017static int svc_deferred_recv(struct svc_rqst *rqstp)
1018{
1019        struct svc_deferred_req *dr = rqstp->rq_deferred;
1020
1021        /* setup iov_base past transport header */
1022        rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1023        /* The iov_len does not include the transport header bytes */
1024        rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1025        rqstp->rq_arg.page_len = 0;
1026        /* The rq_arg.len includes the transport header bytes */
1027        rqstp->rq_arg.len     = dr->argslen<<2;
1028        rqstp->rq_prot        = dr->prot;
1029        memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1030        rqstp->rq_addrlen     = dr->addrlen;
1031        /* Save off transport header len in case we get deferred again */
1032        rqstp->rq_xprt_hlen   = dr->xprt_hlen;
1033        rqstp->rq_daddr       = dr->daddr;
1034        rqstp->rq_respages    = rqstp->rq_pages;
1035        return (dr->argslen<<2) - dr->xprt_hlen;
1036}
1037
1038
1039static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1040{
1041        struct svc_deferred_req *dr = NULL;
1042
1043        if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1044                return NULL;
1045        spin_lock(&xprt->xpt_lock);
1046        clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1047        if (!list_empty(&xprt->xpt_deferred)) {
1048                dr = list_entry(xprt->xpt_deferred.next,
1049                                struct svc_deferred_req,
1050                                handle.recent);
1051                list_del_init(&dr->handle.recent);
1052                set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1053        }
1054        spin_unlock(&xprt->xpt_lock);
1055        return dr;
1056}
1057
1058/**
1059 * svc_find_xprt - find an RPC transport instance
1060 * @serv: pointer to svc_serv to search
1061 * @xcl_name: C string containing transport's class name
1062 * @af: Address family of transport's local address
1063 * @port: transport's IP port number
1064 *
1065 * Return the transport instance pointer for the endpoint accepting
1066 * connections/peer traffic from the specified transport class,
1067 * address family and port.
1068 *
1069 * Specifying 0 for the address family or port is effectively a
1070 * wild-card, and will result in matching the first transport in the
1071 * service's list that has a matching class name.
1072 */
1073struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1074                               const sa_family_t af, const unsigned short port)
1075{
1076        struct svc_xprt *xprt;
1077        struct svc_xprt *found = NULL;
1078
1079        /* Sanity check the args */
1080        if (serv == NULL || xcl_name == NULL)
1081                return found;
1082
1083        spin_lock_bh(&serv->sv_lock);
1084        list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1085                if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1086                        continue;
1087                if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1088                        continue;
1089                if (port != 0 && port != svc_xprt_local_port(xprt))
1090                        continue;
1091                found = xprt;
1092                svc_xprt_get(xprt);
1093                break;
1094        }
1095        spin_unlock_bh(&serv->sv_lock);
1096        return found;
1097}
1098EXPORT_SYMBOL_GPL(svc_find_xprt);
1099
1100/*
1101 * Format a buffer with a list of the active transports. A zero for
1102 * the buflen parameter disables target buffer overflow checking.
1103 */
1104int svc_xprt_names(struct svc_serv *serv, char *buf, int buflen)
1105{
1106        struct svc_xprt *xprt;
1107        char xprt_str[64];
1108        int totlen = 0;
1109        int len;
1110
1111        /* Sanity check args */
1112        if (!serv)
1113                return 0;
1114
1115        spin_lock_bh(&serv->sv_lock);
1116        list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1117                len = snprintf(xprt_str, sizeof(xprt_str),
1118                               "%s %d\n", xprt->xpt_class->xcl_name,
1119                               svc_xprt_local_port(xprt));
1120                /* If the string was truncated, replace with error string */
1121                if (len >= sizeof(xprt_str))
1122                        strcpy(xprt_str, "name-too-long\n");
1123                /* Don't overflow buffer */
1124                len = strlen(xprt_str);
1125                if (buflen && (len + totlen >= buflen))
1126                        break;
1127                strcpy(buf+totlen, xprt_str);
1128                totlen += len;
1129        }
1130        spin_unlock_bh(&serv->sv_lock);
1131        return totlen;
1132}
1133EXPORT_SYMBOL_GPL(svc_xprt_names);
1134
1135
1136/*----------------------------------------------------------------------------*/
1137
1138static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1139{
1140        unsigned int pidx = (unsigned int)*pos;
1141        struct svc_serv *serv = m->private;
1142
1143        dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1144
1145        lock_kernel();
1146        /* bump up the pseudo refcount while traversing */
1147        svc_get(serv);
1148        unlock_kernel();
1149
1150        if (!pidx)
1151                return SEQ_START_TOKEN;
1152        return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1153}
1154
1155static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1156{
1157        struct svc_pool *pool = p;
1158        struct svc_serv *serv = m->private;
1159
1160        dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1161
1162        if (p == SEQ_START_TOKEN) {
1163                pool = &serv->sv_pools[0];
1164        } else {
1165                unsigned int pidx = (pool - &serv->sv_pools[0]);
1166                if (pidx < serv->sv_nrpools-1)
1167                        pool = &serv->sv_pools[pidx+1];
1168                else
1169                        pool = NULL;
1170        }
1171        ++*pos;
1172        return pool;
1173}
1174
1175static void svc_pool_stats_stop(struct seq_file *m, void *p)
1176{
1177        struct svc_serv *serv = m->private;
1178
1179        lock_kernel();
1180        /* this function really, really should have been called svc_put() */
1181        svc_destroy(serv);
1182        unlock_kernel();
1183}
1184
1185static int svc_pool_stats_show(struct seq_file *m, void *p)
1186{
1187        struct svc_pool *pool = p;
1188
1189        if (p == SEQ_START_TOKEN) {
1190                seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken overloads-avoided threads-timedout\n");
1191                return 0;
1192        }
1193
1194        seq_printf(m, "%u %lu %lu %lu %lu %lu\n",
1195                pool->sp_id,
1196                pool->sp_stats.packets,
1197                pool->sp_stats.sockets_queued,
1198                pool->sp_stats.threads_woken,
1199                pool->sp_stats.overloads_avoided,
1200                pool->sp_stats.threads_timedout);
1201
1202        return 0;
1203}
1204
1205static const struct seq_operations svc_pool_stats_seq_ops = {
1206        .start  = svc_pool_stats_start,
1207        .next   = svc_pool_stats_next,
1208        .stop   = svc_pool_stats_stop,
1209        .show   = svc_pool_stats_show,
1210};
1211
1212int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1213{
1214        int err;
1215
1216        err = seq_open(file, &svc_pool_stats_seq_ops);
1217        if (!err)
1218                ((struct seq_file *) file->private_data)->private = serv;
1219        return err;
1220}
1221EXPORT_SYMBOL(svc_pool_stats_open);
1222
1223/*----------------------------------------------------------------------------*/
1224