1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2001, 2002 Cluster File Systems, Inc.
5 * Author: Zach Brown <zab@zabbo.net>
6 * Author: Peter J. Braam <braam@clusterfs.com>
7 * Author: Phil Schwan <phil@clusterfs.com>
8 * Author: Eric Barton <eric@bartonsoftware.com>
10 * This file is part of Portals, http://www.sf.net/projects/sandiaportals/
12 * Portals is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Portals is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Portals; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
29 * LIB functions follow
33 ksocknal_dist(lib_nal_t *nal, ptl_nid_t nid, unsigned long *dist)
35 /* I would guess that if ksocknal_get_peer (nid) == NULL,
36 and we're not routing, then 'nid' is very distant :) */
37 if (nal->libnal_ni.ni_pid.nid == nid) {
47 ksocknal_free_ltx (ksock_ltx_t *ltx)
49 atomic_dec(&ksocknal_data.ksnd_nactive_ltxs);
50 PORTAL_FREE(ltx, ltx->ltx_desc_size);
54 ksocknal_send_iov (ksock_conn_t *conn, ksock_tx_t *tx)
56 struct iovec *iov = tx->tx_iov;
60 LASSERT (tx->tx_niov > 0);
62 /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */
63 rc = ksocknal_lib_send_iov(conn, tx);
65 if (rc <= 0) /* sent nothing? */
69 LASSERT (nob <= tx->tx_resid);
74 LASSERT (tx->tx_niov > 0);
76 if (nob < iov->iov_len) {
77 iov->iov_base = (void *)(((unsigned long)(iov->iov_base)) + nob);
91 ksocknal_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
93 ptl_kiov_t *kiov = tx->tx_kiov;
97 LASSERT (tx->tx_niov == 0);
98 LASSERT (tx->tx_nkiov > 0);
100 /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
101 rc = ksocknal_lib_send_kiov(conn, tx);
103 if (rc <= 0) /* sent nothing? */
107 LASSERT (nob <= tx->tx_resid);
112 LASSERT(tx->tx_nkiov > 0);
114 if (nob < kiov->kiov_len) {
115 kiov->kiov_offset += nob;
116 kiov->kiov_len -= nob;
120 nob -= kiov->kiov_len;
121 tx->tx_kiov = ++kiov;
129 ksocknal_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
134 if (ksocknal_data.ksnd_stall_tx != 0) {
135 set_current_state (TASK_UNINTERRUPTIBLE);
136 schedule_timeout (cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
139 LASSERT (tx->tx_resid != 0);
141 rc = ksocknal_getconnsock (conn);
143 LASSERT (conn->ksnc_closing);
148 if (ksocknal_data.ksnd_enomem_tx > 0) {
150 ksocknal_data.ksnd_enomem_tx--;
152 } else if (tx->tx_niov != 0) {
153 rc = ksocknal_send_iov (conn, tx);
155 rc = ksocknal_send_kiov (conn, tx);
158 bufnob = SOCK_WMEM_QUEUED(conn->ksnc_sock);
159 if (rc > 0) /* sent something? */
160 conn->ksnc_tx_bufnob += rc; /* account it */
162 if (bufnob < conn->ksnc_tx_bufnob) {
163 /* allocated send buffer bytes < computed; infer
164 * something got ACKed */
165 conn->ksnc_tx_deadline = cfs_time_shift(ksocknal_tunables.ksnd_io_timeout);
166 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
167 conn->ksnc_tx_bufnob = bufnob;
171 if (rc <= 0) { /* Didn't write anything? */
173 ksock_sched_t *sched;
175 if (rc == 0) /* some stacks return 0 instead of -EAGAIN */
181 /* Check if EAGAIN is due to memory pressure */
183 sched = conn->ksnc_scheduler;
184 spin_lock_irqsave(&sched->kss_lock, flags);
186 if (!SOCK_TEST_NOSPACE(conn->ksnc_sock) &&
187 !conn->ksnc_tx_ready) {
188 /* SOCK_NOSPACE is set when the socket fills
189 * and cleared in the write_space callback
190 * (which also sets ksnc_tx_ready). If
191 * SOCK_NOSPACE and ksnc_tx_ready are BOTH
192 * zero, I didn't fill the socket and
193 * write_space won't reschedule me, so I
194 * return -ENOMEM to get my caller to retry
199 spin_unlock_irqrestore(&sched->kss_lock, flags);
203 /* socket's wmem_queued now includes 'rc' bytes */
204 atomic_sub (rc, &conn->ksnc_tx_nob);
207 } while (tx->tx_resid != 0);
209 ksocknal_putconnsock (conn);
214 ksocknal_recv_iov (ksock_conn_t *conn)
216 struct iovec *iov = conn->ksnc_rx_iov;
220 LASSERT (conn->ksnc_rx_niov > 0);
222 /* Never touch conn->ksnc_rx_iov or change connection
223 * status inside ksocknal_lib_recv_iov */
224 rc = ksocknal_lib_recv_iov(conn);
229 /* received something... */
232 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
233 conn->ksnc_rx_deadline = cfs_time_shift (ksocknal_tunables.ksnd_io_timeout);
234 mb(); /* order with setting rx_started */
235 conn->ksnc_rx_started = 1;
237 conn->ksnc_rx_nob_wanted -= nob;
238 conn->ksnc_rx_nob_left -= nob;
241 LASSERT (conn->ksnc_rx_niov > 0);
243 if (nob < iov->iov_len) {
245 iov->iov_base = (void *)(((unsigned long)iov->iov_base) + nob);
250 conn->ksnc_rx_iov = ++iov;
251 conn->ksnc_rx_niov--;
258 ksocknal_recv_kiov (ksock_conn_t *conn)
260 ptl_kiov_t *kiov = conn->ksnc_rx_kiov;
263 LASSERT (conn->ksnc_rx_nkiov > 0);
265 /* Never touch conn->ksnc_rx_kiov or change connection
266 * status inside ksocknal_lib_recv_iov */
267 rc = ksocknal_lib_recv_kiov(conn);
272 /* received something... */
275 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
276 conn->ksnc_rx_deadline = cfs_time_shift (ksocknal_tunables.ksnd_io_timeout);
277 mb(); /* order with setting rx_started */
278 conn->ksnc_rx_started = 1;
280 conn->ksnc_rx_nob_wanted -= nob;
281 conn->ksnc_rx_nob_left -= nob;
284 LASSERT (conn->ksnc_rx_nkiov > 0);
286 if (nob < kiov->kiov_len) {
287 kiov->kiov_offset += nob;
288 kiov->kiov_len -= nob;
292 nob -= kiov->kiov_len;
293 conn->ksnc_rx_kiov = ++kiov;
294 conn->ksnc_rx_nkiov--;
301 ksocknal_receive (ksock_conn_t *conn)
303 /* Return 1 on success, 0 on EOF, < 0 on error.
304 * Caller checks ksnc_rx_nob_wanted to determine
305 * progress/completion. */
309 if (ksocknal_data.ksnd_stall_rx != 0) {
310 set_current_state (TASK_UNINTERRUPTIBLE);
311 schedule_timeout(cfs_time_seconds (ksocknal_data.ksnd_stall_rx));
314 rc = ksocknal_getconnsock (conn);
316 LASSERT (conn->ksnc_closing);
321 if (conn->ksnc_rx_niov != 0)
322 rc = ksocknal_recv_iov (conn);
324 rc = ksocknal_recv_kiov (conn);
327 /* error/EOF or partial receive */
330 } else if (rc == 0 && conn->ksnc_rx_started) {
331 /* EOF in the middle of a message */
337 /* Completed a fragment */
339 if (conn->ksnc_rx_nob_wanted == 0) {
340 /* Completed a message segment (header or payload) */
341 if ((ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0 &&
342 (conn->ksnc_rx_state == SOCKNAL_RX_BODY ||
343 conn->ksnc_rx_state == SOCKNAL_RX_BODY_FWD)) {
344 /* Remind the socket to ack eagerly... */
345 ksocknal_lib_eager_ack(conn);
352 ksocknal_putconnsock (conn);
358 ksocknal_zc_callback (zccd_t *zcd)
360 ksock_tx_t *tx = KSOCK_ZCCD_2_TX(zcd);
361 ksock_sched_t *sched = tx->tx_conn->ksnc_scheduler;
365 /* Schedule tx for cleanup (can't do it now due to lock conflicts) */
367 spin_lock_irqsave (&sched->kss_lock, flags);
369 list_add_tail (&tx->tx_list, &sched->kss_zctxdone_list);
370 cfs_waitq_signal (&sched->kss_waitq);
372 spin_unlock_irqrestore (&sched->kss_lock, flags);
378 ksocknal_tx_done (ksock_tx_t *tx, int asynch)
383 if (tx->tx_conn != NULL) {
385 /* zero copy completion isn't always from
386 * process_transmit() so it needs to keep a ref on
389 ksocknal_put_conn (tx->tx_conn);
395 if (tx->tx_isfwd) { /* was a forwarded packet? */
396 kpr_fwd_done (&ksocknal_data.ksnd_router,
397 KSOCK_TX_2_KPR_FWD_DESC (tx),
398 (tx->tx_resid == 0) ? 0 : -ECONNABORTED);
404 ltx = KSOCK_TX_2_KSOCK_LTX (tx);
406 lib_finalize (&ksocknal_lib, ltx->ltx_private, ltx->ltx_cookie,
407 (tx->tx_resid == 0) ? PTL_OK : PTL_FAIL);
409 ksocknal_free_ltx (ltx);
414 ksocknal_tx_launched (ksock_tx_t *tx)
417 if (atomic_read (&tx->tx_zccd.zccd_count) != 1) {
418 ksock_conn_t *conn = tx->tx_conn;
420 /* zccd skbufs are still in-flight. First take a ref on
421 * conn, so it hangs about for ksocknal_tx_done... */
422 atomic_inc (&conn->ksnc_refcount);
424 /* ...then drop the initial ref on zccd, so the zero copy
425 * callback can occur */
426 zccd_put (&tx->tx_zccd);
430 /* Any zero-copy-ness (if any) has completed; I can complete the
431 * transmit now, avoiding an extra schedule */
432 ksocknal_tx_done (tx, 0);
436 ksocknal_process_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
441 rc = ksocknal_transmit (conn, tx);
443 CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc);
445 if (tx->tx_resid == 0) {
446 /* Sent everything OK */
449 ksocknal_tx_launched (tx);
459 counter++; /* exponential backoff warnings */
460 if ((counter & (-counter)) == counter)
461 CWARN("%d ENOMEM tx %p\n", counter, conn);
463 /* Queue on ksnd_enomem_conns for retry after a timeout */
464 spin_lock_irqsave(&ksocknal_data.ksnd_reaper_lock, flags);
466 /* enomem list takes over scheduler's ref... */
467 LASSERT (conn->ksnc_tx_scheduled);
468 list_add_tail(&conn->ksnc_tx_list,
469 &ksocknal_data.ksnd_enomem_conns);
470 if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(),
471 SOCKNAL_ENOMEM_RETRY),
472 ksocknal_data.ksnd_reaper_waketime))
473 cfs_waitq_signal (&ksocknal_data.ksnd_reaper_waitq);
475 spin_unlock_irqrestore(&ksocknal_data.ksnd_reaper_lock, flags);
482 if (!conn->ksnc_closing)
483 CERROR("[%p] Error %d on write to "LPX64
484 " ip %d.%d.%d.%d:%d\n", conn, rc,
485 conn->ksnc_peer->ksnp_nid,
486 HIPQUAD(conn->ksnc_ipaddr),
489 ksocknal_close_conn_and_siblings (conn, rc);
490 ksocknal_tx_launched (tx);
496 ksocknal_launch_autoconnect_locked (ksock_route_t *route)
500 /* called holding write lock on ksnd_global_lock */
502 LASSERT (!route->ksnr_deleted);
503 LASSERT ((route->ksnr_connected & (1 << SOCKNAL_CONN_ANY)) == 0);
504 LASSERT ((route->ksnr_connected & KSNR_TYPED_ROUTES) != KSNR_TYPED_ROUTES);
505 LASSERT (route->ksnr_connecting == 0);
507 if (ksocknal_tunables.ksnd_typed_conns)
508 route->ksnr_connecting =
509 KSNR_TYPED_ROUTES & ~route->ksnr_connected;
511 route->ksnr_connecting = (1 << SOCKNAL_CONN_ANY);
513 atomic_inc (&route->ksnr_refcount); /* extra ref for asynchd */
515 spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
517 list_add_tail (&route->ksnr_connect_list,
518 &ksocknal_data.ksnd_autoconnectd_routes);
519 cfs_waitq_signal (&ksocknal_data.ksnd_autoconnectd_waitq);
521 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
525 ksocknal_find_target_peer_locked (ksock_tx_t *tx, ptl_nid_t nid)
527 char ipbuf[PTL_NALFMT_SIZE];
528 ptl_nid_t target_nid;
530 ksock_peer_t *peer = ksocknal_find_peer_locked (nid);
536 CERROR ("Can't send packet to "LPX64
537 " %s: routed target is not a peer\n",
538 nid, portals_nid2str(SOCKNAL, nid, ipbuf));
542 rc = kpr_lookup (&ksocknal_data.ksnd_router, nid, tx->tx_nob,
545 CERROR ("Can't route to "LPX64" %s: router error %d\n",
546 nid, portals_nid2str(SOCKNAL, nid, ipbuf), rc);
550 peer = ksocknal_find_peer_locked (target_nid);
554 CERROR ("Can't send packet to "LPX64" %s: no peer entry\n",
555 target_nid, portals_nid2str(SOCKNAL, target_nid, ipbuf));
560 ksocknal_find_conn_locked (ksock_tx_t *tx, ksock_peer_t *peer)
562 struct list_head *tmp;
563 ksock_conn_t *typed = NULL;
565 ksock_conn_t *fallback = NULL;
569 list_for_each (tmp, &peer->ksnp_conns) {
570 ksock_conn_t *c = list_entry(tmp, ksock_conn_t, ksnc_list);
571 #if SOCKNAL_ROUND_ROBIN
574 int nob = atomic_read(&c->ksnc_tx_nob) +
575 SOCK_WMEM_QUEUED(c->ksnc_sock);
577 LASSERT (!c->ksnc_closing);
579 if (fallback == NULL || nob < fnob) {
584 if (!ksocknal_tunables.ksnd_typed_conns)
587 switch (c->ksnc_type) {
590 case SOCKNAL_CONN_ANY:
592 case SOCKNAL_CONN_BULK_IN:
594 case SOCKNAL_CONN_BULK_OUT:
595 if (tx->tx_nob < ksocknal_tunables.ksnd_min_bulk)
598 case SOCKNAL_CONN_CONTROL:
599 if (tx->tx_nob >= ksocknal_tunables.ksnd_min_bulk)
604 if (typed == NULL || nob < tnob) {
610 /* prefer the typed selection */
611 conn = (typed != NULL) ? typed : fallback;
613 #if SOCKNAL_ROUND_ROBIN
615 /* round-robin all else being equal */
616 list_del (&conn->ksnc_list);
617 list_add_tail (&conn->ksnc_list, &peer->ksnp_conns);
624 ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn)
627 ksock_sched_t *sched = conn->ksnc_scheduler;
629 /* called holding global lock (read or irq-write) and caller may
630 * not have dropped this lock between finding conn and calling me,
631 * so we don't need the {get,put}connsock dance to deref
633 LASSERT(!conn->ksnc_closing);
634 LASSERT(tx->tx_resid == tx->tx_nob);
636 CDEBUG (D_NET, "Sending to "LPX64" ip %d.%d.%d.%d:%d\n",
637 conn->ksnc_peer->ksnp_nid,
638 HIPQUAD(conn->ksnc_ipaddr),
641 atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
645 zccd_init (&tx->tx_zccd, ksocknal_zc_callback);
646 /* NB this sets 1 ref on zccd, so the callback can only occur after
647 * I've released this ref. */
649 spin_lock_irqsave (&sched->kss_lock, flags);
651 if (list_empty(&conn->ksnc_tx_queue) &&
652 SOCK_WMEM_QUEUED(conn->ksnc_sock) == 0) {
653 /* First packet starts the timeout */
654 conn->ksnc_tx_deadline = cfs_time_shift(ksocknal_tunables.ksnd_io_timeout);
655 conn->ksnc_tx_bufnob = 0;
656 mb(); /* order with adding to tx_queue */
659 list_add_tail (&tx->tx_list, &conn->ksnc_tx_queue);
661 if (conn->ksnc_tx_ready && /* able to send */
662 !conn->ksnc_tx_scheduled) { /* not scheduled to send */
663 /* +1 ref for scheduler */
664 atomic_inc (&conn->ksnc_refcount);
665 list_add_tail (&conn->ksnc_tx_list,
666 &sched->kss_tx_conns);
667 conn->ksnc_tx_scheduled = 1;
668 cfs_waitq_signal (&sched->kss_waitq);
671 spin_unlock_irqrestore (&sched->kss_lock, flags);
675 ksocknal_find_connectable_route_locked (ksock_peer_t *peer)
677 struct list_head *tmp;
678 ksock_route_t *route;
681 list_for_each (tmp, &peer->ksnp_routes) {
682 route = list_entry (tmp, ksock_route_t, ksnr_list);
683 bits = route->ksnr_connected;
685 /* All typed connections established? */
686 if ((bits & KSNR_TYPED_ROUTES) == KSNR_TYPED_ROUTES)
689 /* Untyped connection established? */
690 if ((bits & (1 << SOCKNAL_CONN_ANY)) != 0)
693 /* connection being established? */
694 if (route->ksnr_connecting != 0)
697 /* too soon to retry this guy? */
698 if (!cfs_time_aftereq (cfs_time_current(), route->ksnr_timeout))
708 ksocknal_find_connecting_route_locked (ksock_peer_t *peer)
710 struct list_head *tmp;
711 ksock_route_t *route;
713 list_for_each (tmp, &peer->ksnp_routes) {
714 route = list_entry (tmp, ksock_route_t, ksnr_list);
716 if (route->ksnr_connecting != 0)
724 ksocknal_launch_packet (ksock_tx_t *tx, ptl_nid_t nid)
729 ksock_route_t *route;
732 /* Ensure the frags we've been given EXACTLY match the number of
733 * bytes we want to send. Many TCP/IP stacks disregard any total
734 * size parameters passed to them and just look at the frags.
736 * We always expect at least 1 mapped fragment containing the
737 * complete portals header. */
738 LASSERT (lib_iov_nob (tx->tx_niov, tx->tx_iov) +
739 lib_kiov_nob (tx->tx_nkiov, tx->tx_kiov) == tx->tx_nob);
740 LASSERT (tx->tx_niov >= 1);
741 LASSERT (tx->tx_iov[0].iov_len >= sizeof (ptl_hdr_t));
743 CDEBUG (D_NET, "packet %p type %d, nob %d niov %d nkiov %d\n",
744 tx, ((ptl_hdr_t *)tx->tx_iov[0].iov_base)->type,
745 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
747 tx->tx_conn = NULL; /* only set when assigned a conn */
748 tx->tx_resid = tx->tx_nob;
749 tx->tx_hdr = (ptl_hdr_t *)tx->tx_iov[0].iov_base;
751 g_lock = &ksocknal_data.ksnd_global_lock;
752 #if !SOCKNAL_ROUND_ROBIN
755 peer = ksocknal_find_target_peer_locked (tx, nid);
757 read_unlock (g_lock);
758 return (-EHOSTUNREACH);
761 if (ksocknal_find_connectable_route_locked(peer) == NULL) {
762 conn = ksocknal_find_conn_locked (tx, peer);
764 /* I've got no autoconnect routes that need to be
765 * connecting and I do have an actual connection... */
766 ksocknal_queue_tx_locked (tx, conn);
767 read_unlock (g_lock);
772 /* I'll need a write lock... */
773 read_unlock (g_lock);
775 write_lock_irqsave(g_lock, flags);
777 peer = ksocknal_find_target_peer_locked (tx, nid);
779 write_unlock_irqrestore(g_lock, flags);
780 return (-EHOSTUNREACH);
784 /* launch any/all autoconnections that need it */
785 route = ksocknal_find_connectable_route_locked (peer);
789 ksocknal_launch_autoconnect_locked (route);
792 conn = ksocknal_find_conn_locked (tx, peer);
794 /* Connection exists; queue message on it */
795 ksocknal_queue_tx_locked (tx, conn);
796 write_unlock_irqrestore (g_lock, flags);
800 route = ksocknal_find_connecting_route_locked (peer);
802 /* At least 1 connection is being established; queue the
804 list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
805 write_unlock_irqrestore (g_lock, flags);
809 write_unlock_irqrestore (g_lock, flags);
810 return (-EHOSTUNREACH);
814 ksocknal_sendmsg(lib_nal_t *nal,
821 unsigned int payload_niov,
822 struct iovec *payload_iov,
823 ptl_kiov_t *payload_kiov,
824 size_t payload_offset,
831 /* NB 'private' is different depending on what we're sending.
832 * Just ignore it... */
834 CDEBUG(D_NET, "sending "LPSZ" bytes in %d frags to nid:"LPX64
835 " pid %d\n", payload_nob, payload_niov, nid , pid);
837 LASSERT (payload_nob == 0 || payload_niov > 0);
838 LASSERT (payload_niov <= PTL_MD_MAX_IOV);
840 /* It must be OK to kmap() if required */
841 LASSERT (payload_kiov == NULL || !in_interrupt ());
842 /* payload is either all vaddrs or all pages */
843 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
845 if (payload_iov != NULL)
846 desc_size = offsetof(ksock_ltx_t, ltx_iov[1 + payload_niov]);
848 desc_size = offsetof(ksock_ltx_t, ltx_kiov[payload_niov]);
850 if (in_interrupt() ||
851 type == PTL_MSG_ACK ||
852 type == PTL_MSG_REPLY) {
853 /* Can't block if in interrupt or responding to an incoming
855 PORTAL_ALLOC_ATOMIC(ltx, desc_size);
857 PORTAL_ALLOC(ltx, desc_size);
861 CERROR("Can't allocate tx desc type %d size %d %s\n",
862 type, desc_size, in_interrupt() ? "(intr)" : "");
863 return (PTL_NO_SPACE);
866 atomic_inc(&ksocknal_data.ksnd_nactive_ltxs);
868 ltx->ltx_desc_size = desc_size;
870 /* We always have 1 mapped frag for the header */
871 ltx->ltx_tx.tx_iov = ltx->ltx_iov;
872 ltx->ltx_iov[0].iov_base = <x->ltx_hdr;
873 ltx->ltx_iov[0].iov_len = sizeof(*hdr);
876 ltx->ltx_private = private;
877 ltx->ltx_cookie = cookie;
879 ltx->ltx_tx.tx_isfwd = 0;
880 ltx->ltx_tx.tx_nob = sizeof (*hdr) + payload_nob;
882 if (payload_iov != NULL) {
883 /* payload is all mapped */
884 ltx->ltx_tx.tx_kiov = NULL;
885 ltx->ltx_tx.tx_nkiov = 0;
887 ltx->ltx_tx.tx_niov =
888 1 + lib_extract_iov(payload_niov, <x->ltx_iov[1],
889 payload_niov, payload_iov,
890 payload_offset, payload_nob);
892 /* payload is all pages */
893 ltx->ltx_tx.tx_niov = 1;
895 ltx->ltx_tx.tx_kiov = ltx->ltx_kiov;
896 ltx->ltx_tx.tx_nkiov =
897 lib_extract_kiov(payload_niov, ltx->ltx_kiov,
898 payload_niov, payload_kiov,
899 payload_offset, payload_nob);
902 rc = ksocknal_launch_packet(<x->ltx_tx, nid);
906 ksocknal_free_ltx(ltx);
911 ksocknal_send (lib_nal_t *nal, void *private, lib_msg_t *cookie,
912 ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid,
913 unsigned int payload_niov, struct iovec *payload_iov,
914 size_t payload_offset, size_t payload_len)
916 return (ksocknal_sendmsg(nal, private, cookie,
918 payload_niov, payload_iov, NULL,
919 payload_offset, payload_len));
923 ksocknal_send_pages (lib_nal_t *nal, void *private, lib_msg_t *cookie,
924 ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid,
925 unsigned int payload_niov, ptl_kiov_t *payload_kiov,
926 size_t payload_offset, size_t payload_len)
928 return (ksocknal_sendmsg(nal, private, cookie,
930 payload_niov, NULL, payload_kiov,
931 payload_offset, payload_len));
935 ksocknal_fwd_packet (void *arg, kpr_fwd_desc_t *fwd)
937 ptl_nid_t nid = fwd->kprfd_gateway_nid;
938 ksock_ftx_t *ftx = (ksock_ftx_t *)&fwd->kprfd_scratch;
941 CDEBUG (D_NET, "Forwarding [%p] -> "LPX64" ("LPX64"))\n", fwd,
942 fwd->kprfd_gateway_nid, fwd->kprfd_target_nid);
944 /* I'm the gateway; must be the last hop */
945 if (nid == ksocknal_lib.libnal_ni.ni_pid.nid)
946 nid = fwd->kprfd_target_nid;
948 /* setup iov for hdr */
949 ftx->ftx_iov.iov_base = fwd->kprfd_hdr;
950 ftx->ftx_iov.iov_len = sizeof(ptl_hdr_t);
952 ftx->ftx_tx.tx_isfwd = 1; /* This is a forwarding packet */
953 ftx->ftx_tx.tx_nob = sizeof(ptl_hdr_t) + fwd->kprfd_nob;
954 ftx->ftx_tx.tx_niov = 1;
955 ftx->ftx_tx.tx_iov = &ftx->ftx_iov;
956 ftx->ftx_tx.tx_nkiov = fwd->kprfd_niov;
957 ftx->ftx_tx.tx_kiov = fwd->kprfd_kiov;
959 rc = ksocknal_launch_packet (&ftx->ftx_tx, nid);
961 kpr_fwd_done (&ksocknal_data.ksnd_router, fwd, rc);
965 ksocknal_thread_start (int (*fn)(void *arg), void *arg)
967 long pid = cfs_kernel_thread (fn, arg, 0);
973 write_lock_irqsave(&ksocknal_data.ksnd_global_lock, flags);
974 ksocknal_data.ksnd_nthreads++;
975 write_unlock_irqrestore(&ksocknal_data.ksnd_global_lock, flags);
980 ksocknal_thread_fini (void)
984 write_lock_irqsave(&ksocknal_data.ksnd_global_lock, flags);
985 ksocknal_data.ksnd_nthreads--;
986 write_unlock_irqrestore(&ksocknal_data.ksnd_global_lock, flags);
990 ksocknal_fmb_callback (void *arg, int error)
992 ksock_fmb_t *fmb = (ksock_fmb_t *)arg;
993 ksock_fmb_pool_t *fmp = fmb->fmb_pool;
994 ptl_hdr_t *hdr = &fmb->fmb_hdr;
995 ksock_conn_t *conn = NULL;
996 ksock_sched_t *sched;
998 char ipbuf[PTL_NALFMT_SIZE];
999 char ipbuf2[PTL_NALFMT_SIZE];
1002 CERROR("Failed to route packet from "
1003 LPX64" %s to "LPX64" %s: %d\n",
1004 le64_to_cpu(hdr->src_nid),
1005 portals_nid2str(SOCKNAL, le64_to_cpu(hdr->src_nid), ipbuf),
1006 le64_to_cpu(hdr->dest_nid),
1007 portals_nid2str(SOCKNAL, le64_to_cpu(hdr->dest_nid), ipbuf2),
1010 CDEBUG (D_NET, "routed packet from "LPX64" to "LPX64": OK\n",
1011 le64_to_cpu(hdr->src_nid), le64_to_cpu(hdr->dest_nid));
1013 /* drop peer ref taken on init */
1014 ksocknal_put_peer (fmb->fmb_peer);
1016 spin_lock_irqsave (&fmp->fmp_lock, flags);
1018 list_add (&fmb->fmb_list, &fmp->fmp_idle_fmbs);
1019 fmp->fmp_nactive_fmbs--;
1021 if (!list_empty (&fmp->fmp_blocked_conns)) {
1022 conn = list_entry (fmb->fmb_pool->fmp_blocked_conns.next,
1023 ksock_conn_t, ksnc_rx_list);
1024 list_del (&conn->ksnc_rx_list);
1027 spin_unlock_irqrestore (&fmp->fmp_lock, flags);
1032 CDEBUG (D_NET, "Scheduling conn %p\n", conn);
1033 LASSERT (conn->ksnc_rx_scheduled);
1034 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_FMB_SLEEP);
1036 conn->ksnc_rx_state = SOCKNAL_RX_GET_FMB;
1038 sched = conn->ksnc_scheduler;
1040 spin_lock_irqsave (&sched->kss_lock, flags);
1042 list_add_tail (&conn->ksnc_rx_list, &sched->kss_rx_conns);
1043 cfs_waitq_signal (&sched->kss_waitq);
1045 spin_unlock_irqrestore (&sched->kss_lock, flags);
1049 ksocknal_get_idle_fmb (ksock_conn_t *conn)
1051 int payload_nob = conn->ksnc_rx_nob_left;
1052 unsigned long flags;
1053 ksock_fmb_pool_t *pool;
1056 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB);
1057 LASSERT (kpr_routing(&ksocknal_data.ksnd_router));
1059 if (payload_nob <= SOCKNAL_SMALL_FWD_PAGES * CFS_PAGE_SIZE)
1060 pool = &ksocknal_data.ksnd_small_fmp;
1062 pool = &ksocknal_data.ksnd_large_fmp;
1064 spin_lock_irqsave (&pool->fmp_lock, flags);
1066 if (!list_empty (&pool->fmp_idle_fmbs)) {
1067 fmb = list_entry(pool->fmp_idle_fmbs.next,
1068 ksock_fmb_t, fmb_list);
1069 list_del (&fmb->fmb_list);
1070 pool->fmp_nactive_fmbs++;
1071 spin_unlock_irqrestore (&pool->fmp_lock, flags);
1076 /* deschedule until fmb free */
1078 conn->ksnc_rx_state = SOCKNAL_RX_FMB_SLEEP;
1080 list_add_tail (&conn->ksnc_rx_list,
1081 &pool->fmp_blocked_conns);
1083 spin_unlock_irqrestore (&pool->fmp_lock, flags);
1088 ksocknal_init_fmb (ksock_conn_t *conn, ksock_fmb_t *fmb)
1090 int payload_nob = conn->ksnc_rx_nob_left;
1091 ptl_nid_t dest_nid = le64_to_cpu(conn->ksnc_hdr.dest_nid);
1093 int nob = payload_nob;
1095 LASSERT (conn->ksnc_rx_scheduled);
1096 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB);
1097 LASSERT (conn->ksnc_rx_nob_wanted == conn->ksnc_rx_nob_left);
1098 LASSERT (payload_nob >= 0);
1099 LASSERT (payload_nob <= fmb->fmb_pool->fmp_buff_pages * CFS_PAGE_SIZE);
1100 LASSERT (sizeof (ptl_hdr_t) < CFS_PAGE_SIZE);
1101 LASSERT (fmb->fmb_kiov[0].kiov_offset == 0);
1103 /* Take a ref on the conn's peer to prevent module unload before
1104 * forwarding completes. */
1105 fmb->fmb_peer = conn->ksnc_peer;
1106 atomic_inc (&conn->ksnc_peer->ksnp_refcount);
1108 /* Copy the header we just read into the forwarding buffer. If
1109 * there's payload, start reading reading it into the buffer,
1110 * otherwise the forwarding buffer can be kicked off
1112 fmb->fmb_hdr = conn->ksnc_hdr;
1115 LASSERT (niov < fmb->fmb_pool->fmp_buff_pages);
1116 LASSERT (fmb->fmb_kiov[niov].kiov_offset == 0);
1117 fmb->fmb_kiov[niov].kiov_len = MIN (CFS_PAGE_SIZE, nob);
1118 nob -= CFS_PAGE_SIZE;
1122 kpr_fwd_init(&fmb->fmb_fwd, dest_nid, &fmb->fmb_hdr,
1123 payload_nob, niov, fmb->fmb_kiov,
1124 ksocknal_fmb_callback, fmb);
1126 if (payload_nob == 0) { /* got complete packet already */
1127 CDEBUG (D_NET, "%p "LPX64"->"LPX64" fwd_start (immediate)\n",
1128 conn, le64_to_cpu(conn->ksnc_hdr.src_nid), dest_nid);
1130 kpr_fwd_start (&ksocknal_data.ksnd_router, &fmb->fmb_fwd);
1132 ksocknal_new_packet (conn, 0); /* on to next packet */
1136 conn->ksnc_cookie = fmb; /* stash fmb for later */
1137 conn->ksnc_rx_state = SOCKNAL_RX_BODY_FWD; /* read in the payload */
1139 /* Set up conn->ksnc_rx_kiov to read the payload into fmb's kiov-ed
1141 LASSERT (niov <= sizeof(conn->ksnc_rx_iov_space)/sizeof(ptl_kiov_t));
1143 conn->ksnc_rx_niov = 0;
1144 conn->ksnc_rx_nkiov = niov;
1145 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1146 memcpy(conn->ksnc_rx_kiov, fmb->fmb_kiov, niov * sizeof(ptl_kiov_t));
1148 CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d reading body\n", conn,
1149 le64_to_cpu(conn->ksnc_hdr.src_nid), dest_nid, payload_nob);
1154 ksocknal_fwd_parse (ksock_conn_t *conn)
1157 ptl_nid_t dest_nid = le64_to_cpu(conn->ksnc_hdr.dest_nid);
1158 ptl_nid_t src_nid = le64_to_cpu(conn->ksnc_hdr.src_nid);
1159 int body_len = le32_to_cpu(conn->ksnc_hdr.payload_length);
1160 char str[PTL_NALFMT_SIZE];
1161 char str2[PTL_NALFMT_SIZE];
1163 CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d parsing header\n", conn,
1164 src_nid, dest_nid, conn->ksnc_rx_nob_left);
1166 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_HEADER);
1167 LASSERT (conn->ksnc_rx_scheduled);
1169 if (body_len < 0) { /* length corrupt (overflow) */
1170 CERROR("dropping packet from "LPX64" (%s) for "LPX64" (%s): "
1171 "packet size %d illegal\n",
1172 src_nid, portals_nid2str(TCPNAL, src_nid, str),
1173 dest_nid, portals_nid2str(TCPNAL, dest_nid, str2),
1176 ksocknal_new_packet (conn, 0); /* on to new packet */
1180 if (!kpr_routing(&ksocknal_data.ksnd_router)) { /* not forwarding */
1181 CERROR("dropping packet from "LPX64" (%s) for "LPX64
1182 " (%s): not forwarding\n",
1183 src_nid, portals_nid2str(TCPNAL, src_nid, str),
1184 dest_nid, portals_nid2str(TCPNAL, dest_nid, str2));
1185 /* on to new packet (skip this one's body) */
1186 ksocknal_new_packet (conn, body_len);
1190 if (body_len > PTL_MTU) { /* too big to forward */
1191 CERROR ("dropping packet from "LPX64" (%s) for "LPX64
1192 "(%s): packet size %d too big\n",
1193 src_nid, portals_nid2str(TCPNAL, src_nid, str),
1194 dest_nid, portals_nid2str(TCPNAL, dest_nid, str2),
1196 /* on to new packet (skip this one's body) */
1197 ksocknal_new_packet (conn, body_len);
1201 /* should have gone direct */
1202 peer = ksocknal_get_peer (conn->ksnc_hdr.dest_nid);
1204 CERROR ("dropping packet from "LPX64" (%s) for "LPX64
1205 "(%s): target is a peer\n",
1206 src_nid, portals_nid2str(TCPNAL, src_nid, str),
1207 dest_nid, portals_nid2str(TCPNAL, dest_nid, str2));
1208 ksocknal_put_peer (peer); /* drop ref from get above */
1210 /* on to next packet (skip this one's body) */
1211 ksocknal_new_packet (conn, body_len);
1215 conn->ksnc_rx_state = SOCKNAL_RX_GET_FMB; /* Getting FMB now */
1216 conn->ksnc_rx_nob_left = body_len; /* stash packet size */
1217 conn->ksnc_rx_nob_wanted = body_len; /* (no slop) */
1221 ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip)
1223 static char ksocknal_slop_buffer[4096];
1229 if (nob_to_skip == 0) { /* right at next packet boundary now */
1230 conn->ksnc_rx_started = 0;
1231 mb (); /* racing with timeout thread */
1233 conn->ksnc_rx_state = SOCKNAL_RX_HEADER;
1234 conn->ksnc_rx_nob_wanted = sizeof (ptl_hdr_t);
1235 conn->ksnc_rx_nob_left = sizeof (ptl_hdr_t);
1237 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1238 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_hdr;
1239 conn->ksnc_rx_iov[0].iov_len = sizeof (ptl_hdr_t);
1240 conn->ksnc_rx_niov = 1;
1242 conn->ksnc_rx_kiov = NULL;
1243 conn->ksnc_rx_nkiov = 0;
1247 /* Set up to skip as much a possible now. If there's more left
1248 * (ran out of iov entries) we'll get called again */
1250 conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1251 conn->ksnc_rx_nob_left = nob_to_skip;
1252 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1257 nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer));
1259 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1260 conn->ksnc_rx_iov[niov].iov_len = nob;
1265 } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */
1266 niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec));
1268 conn->ksnc_rx_niov = niov;
1269 conn->ksnc_rx_kiov = NULL;
1270 conn->ksnc_rx_nkiov = 0;
1271 conn->ksnc_rx_nob_wanted = skipped;
1276 ksocknal_process_receive (ksock_conn_t *conn)
1281 LASSERT (atomic_read (&conn->ksnc_refcount) > 0);
1283 /* doesn't need a forwarding buffer */
1284 if (conn->ksnc_rx_state != SOCKNAL_RX_GET_FMB)
1288 fmb = ksocknal_get_idle_fmb (conn);
1290 /* conn descheduled waiting for idle fmb */
1294 if (ksocknal_init_fmb (conn, fmb)) {
1295 /* packet forwarded */
1300 /* NB: sched lock NOT held */
1301 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_HEADER ||
1302 conn->ksnc_rx_state == SOCKNAL_RX_BODY ||
1303 conn->ksnc_rx_state == SOCKNAL_RX_BODY_FWD ||
1304 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1306 LASSERT (conn->ksnc_rx_nob_wanted > 0);
1308 rc = ksocknal_receive(conn);
1311 LASSERT (rc != -EAGAIN);
1314 CWARN ("[%p] EOF from "LPX64" ip %d.%d.%d.%d:%d\n",
1315 conn, conn->ksnc_peer->ksnp_nid,
1316 HIPQUAD(conn->ksnc_ipaddr),
1318 else if (!conn->ksnc_closing)
1319 CERROR ("[%p] Error %d on read from "LPX64
1320 " ip %d.%d.%d.%d:%d\n",
1321 conn, rc, conn->ksnc_peer->ksnp_nid,
1322 HIPQUAD(conn->ksnc_ipaddr),
1325 ksocknal_close_conn_and_siblings (conn, rc);
1326 return (rc == 0 ? -ESHUTDOWN : rc);
1329 if (conn->ksnc_rx_nob_wanted != 0) {
1334 switch (conn->ksnc_rx_state) {
1335 case SOCKNAL_RX_HEADER:
1336 if (conn->ksnc_hdr.type != cpu_to_le32(PTL_MSG_HELLO) &&
1337 le64_to_cpu(conn->ksnc_hdr.dest_nid) !=
1338 ksocknal_lib.libnal_ni.ni_pid.nid) {
1339 /* This packet isn't for me */
1340 ksocknal_fwd_parse (conn);
1341 switch (conn->ksnc_rx_state) {
1342 case SOCKNAL_RX_HEADER: /* skipped (zero payload) */
1343 return (0); /* => come back later */
1344 case SOCKNAL_RX_SLOP: /* skipping packet's body */
1345 goto try_read; /* => go read it */
1346 case SOCKNAL_RX_GET_FMB: /* forwarding */
1347 goto get_fmb; /* => go get a fwd msg buffer */
1354 /* sets wanted_len, iovs etc */
1355 rc = lib_parse(&ksocknal_lib, &conn->ksnc_hdr, conn);
1358 /* I just received garbage: give up on this conn */
1359 ksocknal_close_conn_and_siblings (conn, rc);
1363 if (conn->ksnc_rx_nob_wanted != 0) { /* need to get payload? */
1364 conn->ksnc_rx_state = SOCKNAL_RX_BODY;
1365 goto try_read; /* go read the payload */
1367 /* Fall through (completed packet for me) */
1369 case SOCKNAL_RX_BODY:
1370 /* payload all received */
1371 lib_finalize(&ksocknal_lib, NULL, conn->ksnc_cookie, PTL_OK);
1374 case SOCKNAL_RX_SLOP:
1375 /* starting new packet? */
1376 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1377 return (0); /* come back later */
1378 goto try_read; /* try to finish reading slop now */
1380 case SOCKNAL_RX_BODY_FWD:
1381 /* payload all received */
1382 CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d fwd_start (got body)\n",
1383 conn, le64_to_cpu(conn->ksnc_hdr.src_nid),
1384 le64_to_cpu(conn->ksnc_hdr.dest_nid),
1385 conn->ksnc_rx_nob_left);
1387 /* forward the packet. NB ksocknal_init_fmb() put fmb into
1388 * conn->ksnc_cookie */
1389 fmb = (ksock_fmb_t *)conn->ksnc_cookie;
1390 kpr_fwd_start (&ksocknal_data.ksnd_router, &fmb->fmb_fwd);
1392 /* no slop in forwarded packets */
1393 LASSERT (conn->ksnc_rx_nob_left == 0);
1395 ksocknal_new_packet (conn, 0); /* on to next packet */
1396 return (0); /* (later) */
1404 return (-EINVAL); /* keep gcc happy */
1408 ksocknal_recv (lib_nal_t *nal, void *private, lib_msg_t *msg,
1409 unsigned int niov, struct iovec *iov,
1410 size_t offset, size_t mlen, size_t rlen)
1412 ksock_conn_t *conn = (ksock_conn_t *)private;
1414 LASSERT (mlen <= rlen);
1415 LASSERT (niov <= PTL_MD_MAX_IOV);
1417 conn->ksnc_cookie = msg;
1418 conn->ksnc_rx_nob_wanted = mlen;
1419 conn->ksnc_rx_nob_left = rlen;
1421 conn->ksnc_rx_nkiov = 0;
1422 conn->ksnc_rx_kiov = NULL;
1423 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1424 conn->ksnc_rx_niov =
1425 lib_extract_iov(PTL_MD_MAX_IOV, conn->ksnc_rx_iov,
1426 niov, iov, offset, mlen);
1429 lib_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1430 lib_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1436 ksocknal_recv_pages (lib_nal_t *nal, void *private, lib_msg_t *msg,
1437 unsigned int niov, ptl_kiov_t *kiov,
1438 size_t offset, size_t mlen, size_t rlen)
1440 ksock_conn_t *conn = (ksock_conn_t *)private;
1442 LASSERT (mlen <= rlen);
1443 LASSERT (niov <= PTL_MD_MAX_IOV);
1445 conn->ksnc_cookie = msg;
1446 conn->ksnc_rx_nob_wanted = mlen;
1447 conn->ksnc_rx_nob_left = rlen;
1449 conn->ksnc_rx_niov = 0;
1450 conn->ksnc_rx_iov = NULL;
1451 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1452 conn->ksnc_rx_nkiov =
1453 lib_extract_kiov(PTL_MD_MAX_IOV, conn->ksnc_rx_kiov,
1454 niov, kiov, offset, mlen);
1457 lib_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1458 lib_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1464 ksocknal_sched_cansleep(ksock_sched_t *sched)
1466 unsigned long flags;
1469 spin_lock_irqsave(&sched->kss_lock, flags);
1471 rc = (!ksocknal_data.ksnd_shuttingdown &&
1473 list_empty(&sched->kss_zctxdone_list) &&
1475 list_empty(&sched->kss_rx_conns) &&
1476 list_empty(&sched->kss_tx_conns));
1478 spin_unlock_irqrestore(&sched->kss_lock, flags);
1482 int ksocknal_scheduler (void *arg)
1484 ksock_sched_t *sched = (ksock_sched_t *)arg;
1487 unsigned long flags;
1490 int id = sched - ksocknal_data.ksnd_schedulers;
1493 snprintf (name, sizeof (name),"ksocknald_%02d", id);
1494 kportal_daemonize (name);
1495 kportal_blockallsigs ();
1497 #if (CONFIG_SMP && CPU_AFFINITY)
1498 id = ksocknal_sched2cpu(id);
1499 if (cpu_online(id)) {
1502 set_cpus_allowed(current, m);
1504 CERROR ("Can't set CPU affinity for %s to %d\n", name, id);
1506 #endif /* CONFIG_SMP && CPU_AFFINITY */
1508 spin_lock_irqsave (&sched->kss_lock, flags);
1510 while (!ksocknal_data.ksnd_shuttingdown) {
1511 int did_something = 0;
1513 /* Ensure I progress everything semi-fairly */
1515 if (!list_empty (&sched->kss_rx_conns)) {
1516 conn = list_entry(sched->kss_rx_conns.next,
1517 ksock_conn_t, ksnc_rx_list);
1518 list_del(&conn->ksnc_rx_list);
1520 LASSERT(conn->ksnc_rx_scheduled);
1521 LASSERT(conn->ksnc_rx_ready);
1523 /* clear rx_ready in case receive isn't complete.
1524 * Do it BEFORE we call process_recv, since
1525 * data_ready can set it any time after we release
1527 conn->ksnc_rx_ready = 0;
1528 spin_unlock_irqrestore(&sched->kss_lock, flags);
1530 rc = ksocknal_process_receive(conn);
1532 spin_lock_irqsave(&sched->kss_lock, flags);
1534 /* I'm the only one that can clear this flag */
1535 LASSERT(conn->ksnc_rx_scheduled);
1537 /* Did process_receive get everything it wanted? */
1539 conn->ksnc_rx_ready = 1;
1541 if (conn->ksnc_rx_state == SOCKNAL_RX_FMB_SLEEP ||
1542 conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB) {
1543 /* Conn blocked for a forwarding buffer.
1544 * It will get queued for my attention when
1545 * one becomes available (and it might just
1546 * already have been!). Meanwhile my ref
1547 * on it stays put. */
1548 } else if (conn->ksnc_rx_ready) {
1549 /* reschedule for rx */
1550 list_add_tail (&conn->ksnc_rx_list,
1551 &sched->kss_rx_conns);
1553 conn->ksnc_rx_scheduled = 0;
1555 ksocknal_put_conn(conn);
1561 if (!list_empty (&sched->kss_tx_conns)) {
1562 conn = list_entry(sched->kss_tx_conns.next,
1563 ksock_conn_t, ksnc_tx_list);
1564 list_del (&conn->ksnc_tx_list);
1566 LASSERT(conn->ksnc_tx_scheduled);
1567 LASSERT(conn->ksnc_tx_ready);
1568 LASSERT(!list_empty(&conn->ksnc_tx_queue));
1570 tx = list_entry(conn->ksnc_tx_queue.next,
1571 ksock_tx_t, tx_list);
1572 /* dequeue now so empty list => more to send */
1573 list_del(&tx->tx_list);
1575 /* Clear tx_ready in case send isn't complete. Do
1576 * it BEFORE we call process_transmit, since
1577 * write_space can set it any time after we release
1579 conn->ksnc_tx_ready = 0;
1580 spin_unlock_irqrestore (&sched->kss_lock, flags);
1582 rc = ksocknal_process_transmit(conn, tx);
1584 spin_lock_irqsave (&sched->kss_lock, flags);
1586 if (rc == -ENOMEM || rc == -EAGAIN) {
1587 /* Incomplete send: replace tx on HEAD of tx_queue */
1588 list_add (&tx->tx_list, &conn->ksnc_tx_queue);
1590 /* Complete send; assume space for more */
1591 conn->ksnc_tx_ready = 1;
1594 if (rc == -ENOMEM) {
1595 /* Do nothing; after a short timeout, this
1596 * conn will be reposted on kss_tx_conns. */
1597 } else if (conn->ksnc_tx_ready &&
1598 !list_empty (&conn->ksnc_tx_queue)) {
1599 /* reschedule for tx */
1600 list_add_tail (&conn->ksnc_tx_list,
1601 &sched->kss_tx_conns);
1603 conn->ksnc_tx_scheduled = 0;
1605 ksocknal_put_conn (conn);
1611 if (!list_empty (&sched->kss_zctxdone_list)) {
1613 list_entry(sched->kss_zctxdone_list.next,
1614 ksock_tx_t, tx_list);
1617 list_del (&tx->tx_list);
1618 spin_unlock_irqrestore (&sched->kss_lock, flags);
1620 ksocknal_tx_done (tx, 1);
1622 spin_lock_irqsave (&sched->kss_lock, flags);
1625 if (!did_something || /* nothing to do */
1626 ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1627 spin_unlock_irqrestore (&sched->kss_lock, flags);
1631 if (!did_something) { /* wait for something to do */
1632 rc = wait_event_interruptible (sched->kss_waitq,
1633 !ksocknal_sched_cansleep(sched));
1638 spin_lock_irqsave (&sched->kss_lock, flags);
1642 spin_unlock_irqrestore (&sched->kss_lock, flags);
1643 ksocknal_thread_fini ();
1648 * Add connection to kss_rx_conns of scheduler
1649 * and wakeup the scheduler.
1651 void ksocknal_read_callback (ksock_conn_t *conn)
1653 ksock_sched_t *sched;
1654 unsigned long flags;
1657 sched = conn->ksnc_scheduler;
1659 spin_lock_irqsave (&sched->kss_lock, flags);
1661 conn->ksnc_rx_ready = 1;
1663 if (!conn->ksnc_rx_scheduled) { /* not being progressed */
1664 list_add_tail(&conn->ksnc_rx_list,
1665 &sched->kss_rx_conns);
1666 conn->ksnc_rx_scheduled = 1;
1667 /* extra ref for scheduler */
1668 atomic_inc (&conn->ksnc_refcount);
1670 cfs_waitq_signal (&sched->kss_waitq);
1672 spin_unlock_irqrestore (&sched->kss_lock, flags);
1678 * Add connection to kss_tx_conns of scheduler
1679 * and wakeup the scheduler.
1681 void ksocknal_write_callback (ksock_conn_t *conn)
1683 ksock_sched_t *sched;
1684 unsigned long flags;
1687 sched = conn->ksnc_scheduler;
1689 spin_lock_irqsave (&sched->kss_lock, flags);
1691 conn->ksnc_tx_ready = 1;
1693 if (!conn->ksnc_tx_scheduled && // not being progressed
1694 !list_empty(&conn->ksnc_tx_queue)){//packets to send
1695 list_add_tail (&conn->ksnc_tx_list,
1696 &sched->kss_tx_conns);
1697 conn->ksnc_tx_scheduled = 1;
1698 /* extra ref for scheduler */
1699 atomic_inc (&conn->ksnc_refcount);
1701 cfs_waitq_signal (&sched->kss_waitq);
1704 spin_unlock_irqrestore (&sched->kss_lock, flags);
1710 ksocknal_sock_write (struct socket *sock, void *buffer, int nob)
1712 return ksocknal_lib_sock_write(sock, buffer, nob);
1716 ksocknal_sock_read (struct socket *sock, void *buffer, int nob)
1718 return ksocknal_lib_sock_read(sock, buffer, nob);
1722 ksocknal_send_hello (ksock_conn_t *conn, __u32 *ipaddrs, int nipaddrs)
1724 /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1725 struct socket *sock = conn->ksnc_sock;
1727 ptl_magicversion_t *hmv = (ptl_magicversion_t *)&hdr.dest_nid;
1731 LASSERT (conn->ksnc_type != SOCKNAL_CONN_NONE);
1732 LASSERT (nipaddrs <= SOCKNAL_MAX_INTERFACES);
1734 /* No need for getconnsock/putconnsock */
1735 LASSERT (!conn->ksnc_closing);
1737 LASSERT (sizeof (*hmv) == sizeof (hdr.dest_nid));
1738 hmv->magic = cpu_to_le32 (PORTALS_PROTO_MAGIC);
1739 hmv->version_major = cpu_to_le16 (PORTALS_PROTO_VERSION_MAJOR);
1740 hmv->version_minor = cpu_to_le16 (PORTALS_PROTO_VERSION_MINOR);
1742 hdr.src_nid = cpu_to_le64 (ksocknal_lib.libnal_ni.ni_pid.nid);
1743 hdr.type = cpu_to_le32 (PTL_MSG_HELLO);
1744 hdr.payload_length = cpu_to_le32 (nipaddrs * sizeof(*ipaddrs));
1746 hdr.msg.hello.type = cpu_to_le32 (conn->ksnc_type);
1747 hdr.msg.hello.incarnation =
1748 cpu_to_le64 (ksocknal_data.ksnd_incarnation);
1750 /* Receiver is eager */
1751 rc = ksocknal_sock_write (sock, &hdr, sizeof(hdr));
1753 CERROR ("Error %d sending HELLO hdr to %u.%u.%u.%u/%d\n",
1754 rc, HIPQUAD(conn->ksnc_ipaddr), conn->ksnc_port);
1761 for (i = 0; i < nipaddrs; i++) {
1762 ipaddrs[i] = __cpu_to_le32 (ipaddrs[i]);
1765 rc = ksocknal_sock_write (sock, ipaddrs, nipaddrs * sizeof(*ipaddrs));
1767 CERROR ("Error %d sending HELLO payload (%d)"
1768 " to %u.%u.%u.%u/%d\n", rc, nipaddrs,
1769 HIPQUAD(conn->ksnc_ipaddr), conn->ksnc_port);
1774 ksocknal_invert_type(int type)
1778 case SOCKNAL_CONN_ANY:
1779 case SOCKNAL_CONN_CONTROL:
1781 case SOCKNAL_CONN_BULK_IN:
1782 return SOCKNAL_CONN_BULK_OUT;
1783 case SOCKNAL_CONN_BULK_OUT:
1784 return SOCKNAL_CONN_BULK_IN;
1786 return (SOCKNAL_CONN_NONE);
1791 ksocknal_recv_hello (ksock_conn_t *conn, ptl_nid_t *nid,
1792 __u64 *incarnation, __u32 *ipaddrs)
1794 struct socket *sock = conn->ksnc_sock;
1800 ptl_magicversion_t *hmv;
1802 hmv = (ptl_magicversion_t *)&hdr.dest_nid;
1803 LASSERT (sizeof (*hmv) == sizeof (hdr.dest_nid));
1805 rc = ksocknal_sock_read (sock, hmv, sizeof (*hmv));
1807 CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n",
1808 rc, HIPQUAD(conn->ksnc_ipaddr));
1812 if (hmv->magic != le32_to_cpu (PORTALS_PROTO_MAGIC)) {
1813 CERROR ("Bad magic %#08x (%#08x expected) from %u.%u.%u.%u\n",
1814 __cpu_to_le32 (hmv->magic), PORTALS_PROTO_MAGIC,
1815 HIPQUAD(conn->ksnc_ipaddr));
1819 if (hmv->version_major != cpu_to_le16 (PORTALS_PROTO_VERSION_MAJOR) ||
1820 hmv->version_minor != cpu_to_le16 (PORTALS_PROTO_VERSION_MINOR)) {
1821 CERROR ("Incompatible protocol version %d.%d (%d.%d expected)"
1822 " from %u.%u.%u.%u\n",
1823 le16_to_cpu (hmv->version_major),
1824 le16_to_cpu (hmv->version_minor),
1825 PORTALS_PROTO_VERSION_MAJOR,
1826 PORTALS_PROTO_VERSION_MINOR,
1827 HIPQUAD(conn->ksnc_ipaddr));
1831 #if (PORTALS_PROTO_VERSION_MAJOR != 1)
1832 # error "This code only understands protocol version 1.x"
1834 /* version 1 sends magic/version as the dest_nid of a 'hello'
1835 * header, followed by payload full of interface IP addresses.
1836 * Read the rest of it in now... */
1838 rc = ksocknal_sock_read (sock, hmv + 1, sizeof (hdr) - sizeof (*hmv));
1840 CERROR ("Error %d reading rest of HELLO hdr from %u.%u.%u.%u\n",
1841 rc, HIPQUAD(conn->ksnc_ipaddr));
1845 /* ...and check we got what we expected */
1846 if (hdr.type != cpu_to_le32 (PTL_MSG_HELLO)) {
1847 CERROR ("Expecting a HELLO hdr,"
1848 " but got type %d from %u.%u.%u.%u\n",
1849 le32_to_cpu (hdr.type),
1850 HIPQUAD(conn->ksnc_ipaddr));
1854 if (le64_to_cpu(hdr.src_nid) == PTL_NID_ANY) {
1855 CERROR("Expecting a HELLO hdr with a NID, but got PTL_NID_ANY"
1856 "from %u.%u.%u.%u\n", HIPQUAD(conn->ksnc_ipaddr));
1860 if (*nid == PTL_NID_ANY) { /* don't know peer's nid yet */
1861 *nid = le64_to_cpu(hdr.src_nid);
1862 } else if (*nid != le64_to_cpu (hdr.src_nid)) {
1863 CERROR ("Connected to nid "LPX64"@%u.%u.%u.%u "
1864 "but expecting "LPX64"\n",
1865 le64_to_cpu (hdr.src_nid),
1866 HIPQUAD(conn->ksnc_ipaddr), *nid);
1870 type = __le32_to_cpu(hdr.msg.hello.type);
1872 if (conn->ksnc_type == SOCKNAL_CONN_NONE) {
1873 /* I've accepted this connection; peer determines type */
1874 conn->ksnc_type = ksocknal_invert_type(type);
1875 if (conn->ksnc_type == SOCKNAL_CONN_NONE) {
1876 CERROR ("Unexpected type %d from "LPX64"@%u.%u.%u.%u\n",
1877 type, *nid, HIPQUAD(conn->ksnc_ipaddr));
1880 } else if (ksocknal_invert_type(type) != conn->ksnc_type) {
1881 CERROR ("Mismatched types: me %d, "LPX64"@%u.%u.%u.%u %d\n",
1882 conn->ksnc_type, *nid, HIPQUAD(conn->ksnc_ipaddr),
1883 le32_to_cpu(hdr.msg.hello.type));
1887 *incarnation = le64_to_cpu(hdr.msg.hello.incarnation);
1889 nips = __le32_to_cpu (hdr.payload_length) / sizeof (__u32);
1891 if (nips > SOCKNAL_MAX_INTERFACES ||
1892 nips * sizeof(__u32) != __le32_to_cpu (hdr.payload_length)) {
1893 CERROR("Bad payload length %d from "LPX64"@%u.%u.%u.%u\n",
1894 __le32_to_cpu (hdr.payload_length),
1895 *nid, HIPQUAD(conn->ksnc_ipaddr));
1901 rc = ksocknal_sock_read (sock, ipaddrs, nips * sizeof(*ipaddrs));
1903 CERROR ("Error %d reading IPs from "LPX64"@%u.%u.%u.%u\n",
1904 rc, *nid, HIPQUAD(conn->ksnc_ipaddr));
1908 for (i = 0; i < nips; i++) {
1909 ipaddrs[i] = __le32_to_cpu(ipaddrs[i]);
1911 if (ipaddrs[i] == 0) {
1912 CERROR("Zero IP[%d] from "LPX64"@%u.%u.%u.%u\n",
1913 i, *nid, HIPQUAD(conn->ksnc_ipaddr));
1922 ksocknal_get_conn_tunables (ksock_conn_t *conn, int *txmem, int *rxmem, int *nagle)
1924 return ksocknal_lib_get_conn_tunables(conn, txmem, rxmem, nagle);
1928 ksocknal_connect_peer (ksock_route_t *route, int type)
1930 struct socket *sock;
1935 /* Iterate through reserved ports. When typed connections are
1936 * used, we will need to bind to multiple ports, but we only know
1937 * this at connect time. But, by that time we've already called
1938 * bind() so we need a new socket. */
1940 for (port = 1023; port > 512; --port) {
1942 rc = ksocknal_lib_connect_sock(&sock, &may_retry, route, port);
1945 rc = ksocknal_create_conn(route, sock, type);
1946 cfs_put_file(KSN_SOCK2FILE(sock));
1954 CERROR("Out of ports trying to bind to a reserved port\n");
1955 return (-EADDRINUSE);
1959 ksocknal_autoconnect (ksock_route_t *route)
1961 CFS_LIST_HEAD (zombies);
1964 unsigned long flags;
1969 for (type = 0; type < SOCKNAL_CONN_NTYPES; type++)
1970 if ((route->ksnr_connecting & (1 << type)) != 0)
1972 LASSERT (type < SOCKNAL_CONN_NTYPES);
1974 rc = ksocknal_connect_peer (route, type);
1978 /* successfully autoconnected: create_conn did the
1979 * route/conn binding and scheduled any blocked packets */
1981 if (route->ksnr_connecting == 0) {
1982 /* No more connections required */
1987 /* Connection attempt failed */
1989 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
1991 peer = route->ksnr_peer;
1992 route->ksnr_connecting = 0;
1994 /* This is a retry rather than a new connection */
1995 LASSERT (route->ksnr_retry_interval != 0);
1996 route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1997 route->ksnr_retry_interval);
1998 route->ksnr_retry_interval = MIN (route->ksnr_retry_interval * 2,
1999 SOCKNAL_MAX_RECONNECT_INTERVAL);
2001 if (!list_empty (&peer->ksnp_tx_queue) &&
2002 ksocknal_find_connecting_route_locked (peer) == NULL) {
2003 LASSERT (list_empty (&peer->ksnp_conns));
2005 /* None of the connections that the blocked packets are
2006 * waiting for have been successful. Complete them now... */
2008 tx = list_entry (peer->ksnp_tx_queue.next,
2009 ksock_tx_t, tx_list);
2010 list_del (&tx->tx_list);
2011 list_add_tail (&tx->tx_list, &zombies);
2012 } while (!list_empty (&peer->ksnp_tx_queue));
2015 #if 0 /* irrelevent with only eager routes */
2016 if (!route->ksnr_deleted) {
2017 /* make this route least-favourite for re-selection */
2018 list_del(&route->ksnr_list);
2019 list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
2022 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
2024 while (!list_empty (&zombies)) {
2025 char ipbuf[PTL_NALFMT_SIZE];
2026 char ipbuf2[PTL_NALFMT_SIZE];
2027 tx = list_entry (zombies.next, ksock_tx_t, tx_list);
2029 CERROR ("Deleting packet type %d len %d ("LPX64" %s->"LPX64" %s)\n",
2030 le32_to_cpu (tx->tx_hdr->type),
2031 le32_to_cpu (tx->tx_hdr->payload_length),
2032 le64_to_cpu (tx->tx_hdr->src_nid),
2033 portals_nid2str(SOCKNAL,
2034 le64_to_cpu(tx->tx_hdr->src_nid),
2036 le64_to_cpu (tx->tx_hdr->dest_nid),
2037 portals_nid2str(SOCKNAL,
2038 le64_to_cpu(tx->tx_hdr->src_nid),
2041 list_del (&tx->tx_list);
2043 ksocknal_tx_done (tx, 0);
2048 ksocknal_autoconnectd (void *arg)
2050 long id = (long)arg;
2052 unsigned long flags;
2053 ksock_route_t *route;
2056 snprintf (name, sizeof (name), "ksocknal_ad%02ld", id);
2057 kportal_daemonize (name);
2058 kportal_blockallsigs ();
2060 spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2062 while (!ksocknal_data.ksnd_shuttingdown) {
2064 if (!list_empty (&ksocknal_data.ksnd_autoconnectd_routes)) {
2065 route = list_entry (ksocknal_data.ksnd_autoconnectd_routes.next,
2066 ksock_route_t, ksnr_connect_list);
2068 list_del (&route->ksnr_connect_list);
2069 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2071 ksocknal_autoconnect (route);
2072 ksocknal_put_route (route);
2074 spin_lock_irqsave(&ksocknal_data.ksnd_autoconnectd_lock,
2079 spin_unlock_irqrestore(&ksocknal_data.ksnd_autoconnectd_lock,
2082 rc = wait_event_interruptible(ksocknal_data.ksnd_autoconnectd_waitq,
2083 ksocknal_data.ksnd_shuttingdown ||
2084 !list_empty(&ksocknal_data.ksnd_autoconnectd_routes));
2086 spin_lock_irqsave(&ksocknal_data.ksnd_autoconnectd_lock, flags);
2089 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2091 ksocknal_thread_fini ();
2096 ksocknal_find_timed_out_conn (ksock_peer_t *peer)
2098 /* We're called with a shared lock on ksnd_global_lock */
2100 struct list_head *ctmp;
2102 list_for_each (ctmp, &peer->ksnp_conns) {
2103 conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
2105 /* Don't need the {get,put}connsock dance to deref ksnc_sock... */
2106 LASSERT (!conn->ksnc_closing);
2108 if (SOCK_ERROR(conn->ksnc_sock) != 0) {
2109 /* Something (e.g. failed keepalive) set the socket error */
2110 atomic_inc (&conn->ksnc_refcount);
2111 CERROR ("Socket error %d: "LPX64" %p %d.%d.%d.%d\n",
2112 SOCK_ERROR(conn->ksnc_sock), peer->ksnp_nid,
2113 conn, HIPQUAD(conn->ksnc_ipaddr));
2117 if (conn->ksnc_rx_started &&
2118 cfs_time_aftereq (cfs_time_current(),
2119 conn->ksnc_rx_deadline)) {
2120 /* Timed out incomplete incoming message */
2121 atomic_inc (&conn->ksnc_refcount);
2122 CERROR ("Timed out RX from "LPX64" %p %d.%d.%d.%d\n",
2123 peer->ksnp_nid,conn,HIPQUAD(conn->ksnc_ipaddr));
2127 if ((!list_empty (&conn->ksnc_tx_queue) ||
2128 SOCK_WMEM_QUEUED(conn->ksnc_sock) != 0) &&
2129 cfs_time_aftereq (cfs_time_current(),
2130 conn->ksnc_tx_deadline)) {
2131 /* Timed out messages queued for sending or
2132 * buffered in the socket's send buffer */
2133 atomic_inc (&conn->ksnc_refcount);
2134 CERROR ("Timed out TX to "LPX64" %s%d %p %d.%d.%d.%d\n",
2136 list_empty (&conn->ksnc_tx_queue) ? "" : "Q ",
2137 SOCK_WMEM_QUEUED(conn->ksnc_sock), conn,
2138 HIPQUAD(conn->ksnc_ipaddr));
2147 ksocknal_check_peer_timeouts (int idx)
2149 struct list_head *peers = &ksocknal_data.ksnd_peers[idx];
2150 struct list_head *ptmp;
2155 /* NB. We expect to have a look at all the peers and not find any
2156 * connections to time out, so we just use a shared lock while we
2158 read_lock (&ksocknal_data.ksnd_global_lock);
2160 list_for_each (ptmp, peers) {
2161 peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
2162 conn = ksocknal_find_timed_out_conn (peer);
2165 read_unlock (&ksocknal_data.ksnd_global_lock);
2167 CERROR ("Timeout out conn->"LPX64" ip %d.%d.%d.%d:%d\n",
2169 HIPQUAD(conn->ksnc_ipaddr),
2171 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2173 /* NB we won't find this one again, but we can't
2174 * just proceed with the next peer, since we dropped
2175 * ksnd_global_lock and it might be dead already! */
2176 ksocknal_put_conn (conn);
2181 read_unlock (&ksocknal_data.ksnd_global_lock);
2185 ksocknal_reaper (void *arg)
2187 cfs_waitlink_t wait;
2188 unsigned long flags;
2190 ksock_sched_t *sched;
2191 struct list_head enomem_conns;
2193 cfs_duration_t timeout;
2196 cfs_time_t deadline = cfs_time_current();
2198 kportal_daemonize ("ksocknal_reaper");
2199 kportal_blockallsigs ();
2201 CFS_INIT_LIST_HEAD(&enomem_conns);
2202 cfs_waitlink_init (&wait);
2204 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2206 while (!ksocknal_data.ksnd_shuttingdown) {
2208 if (!list_empty (&ksocknal_data.ksnd_deathrow_conns)) {
2209 conn = list_entry (ksocknal_data.ksnd_deathrow_conns.next,
2210 ksock_conn_t, ksnc_list);
2211 list_del (&conn->ksnc_list);
2213 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2215 ksocknal_terminate_conn (conn);
2216 ksocknal_put_conn (conn);
2218 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2222 if (!list_empty (&ksocknal_data.ksnd_zombie_conns)) {
2223 conn = list_entry (ksocknal_data.ksnd_zombie_conns.next,
2224 ksock_conn_t, ksnc_list);
2225 list_del (&conn->ksnc_list);
2227 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2229 ksocknal_destroy_conn (conn);
2231 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2235 if (!list_empty (&ksocknal_data.ksnd_enomem_conns)) {
2236 list_add(&enomem_conns, &ksocknal_data.ksnd_enomem_conns);
2237 list_del_init(&ksocknal_data.ksnd_enomem_conns);
2240 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2242 /* reschedule all the connections that stalled with ENOMEM... */
2244 while (!list_empty (&enomem_conns)) {
2245 conn = list_entry (enomem_conns.next,
2246 ksock_conn_t, ksnc_tx_list);
2247 list_del (&conn->ksnc_tx_list);
2249 sched = conn->ksnc_scheduler;
2251 spin_lock_irqsave (&sched->kss_lock, flags);
2253 LASSERT (conn->ksnc_tx_scheduled);
2254 conn->ksnc_tx_ready = 1;
2255 list_add_tail (&conn->ksnc_tx_list, &sched->kss_tx_conns);
2256 cfs_waitq_signal (&sched->kss_waitq);
2258 spin_unlock_irqrestore (&sched->kss_lock, flags);
2262 /* careful with the jiffy wrap... */
2263 while ((timeout = cfs_time_sub(deadline,
2264 cfs_time_current())) <= 0) {
2267 int chunk = ksocknal_data.ksnd_peer_hash_size;
2269 /* Time to check for timeouts on a few more peers: I do
2270 * checks every 'p' seconds on a proportion of the peer
2271 * table and I need to check every connection 'n' times
2272 * within a timeout interval, to ensure I detect a
2273 * timeout on any connection within (n+1)/n times the
2274 * timeout interval. */
2276 if (ksocknal_tunables.ksnd_io_timeout > n * p)
2277 chunk = (chunk * n * p) /
2278 ksocknal_tunables.ksnd_io_timeout;
2282 for (i = 0; i < chunk; i++) {
2283 ksocknal_check_peer_timeouts (peer_index);
2284 peer_index = (peer_index + 1) %
2285 ksocknal_data.ksnd_peer_hash_size;
2288 deadline = cfs_time_add(deadline, cfs_time_seconds(p));
2291 if (nenomem_conns != 0) {
2292 /* Reduce my timeout if I rescheduled ENOMEM conns.
2293 * This also prevents me getting woken immediately
2294 * if any go back on my enomem list. */
2295 timeout = SOCKNAL_ENOMEM_RETRY;
2297 ksocknal_data.ksnd_reaper_waketime =
2298 cfs_time_add(cfs_time_current(), timeout);
2300 set_current_state (TASK_INTERRUPTIBLE);
2301 cfs_waitq_add (&ksocknal_data.ksnd_reaper_waitq, &wait);
2303 if (!ksocknal_data.ksnd_shuttingdown &&
2304 list_empty (&ksocknal_data.ksnd_deathrow_conns) &&
2305 list_empty (&ksocknal_data.ksnd_zombie_conns))
2306 cfs_waitq_timedwait (&wait, timeout);
2308 set_current_state (TASK_RUNNING);
2309 cfs_waitq_del (&ksocknal_data.ksnd_reaper_waitq, &wait);
2311 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2314 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2316 ksocknal_thread_fini ();
2320 lib_nal_t ksocknal_lib = {
2321 libnal_data: &ksocknal_data, /* NAL private data */
2322 libnal_send: ksocknal_send,
2323 libnal_send_pages: ksocknal_send_pages,
2324 libnal_recv: ksocknal_recv,
2325 libnal_recv_pages: ksocknal_recv_pages,
2326 libnal_dist: ksocknal_dist