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 (%u allocated)\n",
462 counter, conn, atomic_read(&portal_kmemory));
464 /* Queue on ksnd_enomem_conns for retry after a timeout */
465 spin_lock_irqsave(&ksocknal_data.ksnd_reaper_lock, flags);
467 /* enomem list takes over scheduler's ref... */
468 LASSERT (conn->ksnc_tx_scheduled);
469 list_add_tail(&conn->ksnc_tx_list,
470 &ksocknal_data.ksnd_enomem_conns);
471 if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(),
472 SOCKNAL_ENOMEM_RETRY),
473 ksocknal_data.ksnd_reaper_waketime))
474 cfs_waitq_signal (&ksocknal_data.ksnd_reaper_waitq);
476 spin_unlock_irqrestore(&ksocknal_data.ksnd_reaper_lock, flags);
483 if (!conn->ksnc_closing) {
486 LCONSOLE_WARN("Host %u.%u.%u.%u reset our connection "
487 "while we were sending data; it may have "
489 HIPQUAD(conn->ksnc_ipaddr));
492 LCONSOLE_WARN("There was an unexpected network error "
493 "while writing to %u.%u.%u.%u: %d.\n",
494 HIPQUAD(conn->ksnc_ipaddr), rc);
497 CERROR("[%p] Error %d on write to "LPX64
498 " ip %d.%d.%d.%d:%d\n", conn, rc,
499 conn->ksnc_peer->ksnp_nid,
500 HIPQUAD(conn->ksnc_ipaddr),
504 ksocknal_close_conn_and_siblings (conn, rc);
505 ksocknal_tx_launched (tx);
511 ksocknal_launch_autoconnect_locked (ksock_route_t *route)
515 /* called holding write lock on ksnd_global_lock */
517 LASSERT (!route->ksnr_deleted);
518 LASSERT ((route->ksnr_connected & (1 << SOCKNAL_CONN_ANY)) == 0);
519 LASSERT ((route->ksnr_connected & KSNR_TYPED_ROUTES) != KSNR_TYPED_ROUTES);
520 LASSERT (route->ksnr_connecting == 0);
522 if (ksocknal_tunables.ksnd_typed_conns)
523 route->ksnr_connecting =
524 KSNR_TYPED_ROUTES & ~route->ksnr_connected;
526 route->ksnr_connecting = (1 << SOCKNAL_CONN_ANY);
528 atomic_inc (&route->ksnr_refcount); /* extra ref for asynchd */
530 spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
532 list_add_tail (&route->ksnr_connect_list,
533 &ksocknal_data.ksnd_autoconnectd_routes);
534 cfs_waitq_signal (&ksocknal_data.ksnd_autoconnectd_waitq);
536 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
540 ksocknal_find_target_peer_locked (ksock_tx_t *tx, ptl_nid_t nid)
542 char ipbuf[PTL_NALFMT_SIZE];
543 ptl_nid_t target_nid;
545 ksock_peer_t *peer = ksocknal_find_peer_locked (nid);
551 CERROR ("Can't send packet to "LPX64
552 " %s: routed target is not a peer\n",
553 nid, portals_nid2str(SOCKNAL, nid, ipbuf));
557 rc = kpr_lookup (&ksocknal_data.ksnd_router, nid, tx->tx_nob,
560 CERROR ("Can't route to "LPX64" %s: router error %d\n",
561 nid, portals_nid2str(SOCKNAL, nid, ipbuf), rc);
565 peer = ksocknal_find_peer_locked (target_nid);
569 CERROR ("Can't send packet to "LPX64" %s: no peer entry\n",
570 target_nid, portals_nid2str(SOCKNAL, target_nid, ipbuf));
575 ksocknal_find_conn_locked (ksock_tx_t *tx, ksock_peer_t *peer)
577 struct list_head *tmp;
578 ksock_conn_t *typed = NULL;
580 ksock_conn_t *fallback = NULL;
584 list_for_each (tmp, &peer->ksnp_conns) {
585 ksock_conn_t *c = list_entry(tmp, ksock_conn_t, ksnc_list);
586 #if SOCKNAL_ROUND_ROBIN
589 int nob = atomic_read(&c->ksnc_tx_nob) +
590 SOCK_WMEM_QUEUED(c->ksnc_sock);
592 LASSERT (!c->ksnc_closing);
594 if (fallback == NULL || nob < fnob) {
599 if (!ksocknal_tunables.ksnd_typed_conns)
602 switch (c->ksnc_type) {
604 CERROR("ksnc_type bad: %u\n", c->ksnc_type);
606 case SOCKNAL_CONN_ANY:
608 case SOCKNAL_CONN_BULK_IN:
610 case SOCKNAL_CONN_BULK_OUT:
611 if (tx->tx_nob < ksocknal_tunables.ksnd_min_bulk)
614 case SOCKNAL_CONN_CONTROL:
615 if (tx->tx_nob >= ksocknal_tunables.ksnd_min_bulk)
620 if (typed == NULL || nob < tnob) {
626 /* prefer the typed selection */
627 conn = (typed != NULL) ? typed : fallback;
629 #if SOCKNAL_ROUND_ROBIN
631 /* round-robin all else being equal */
632 list_del (&conn->ksnc_list);
633 list_add_tail (&conn->ksnc_list, &peer->ksnp_conns);
640 ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn)
643 ksock_sched_t *sched = conn->ksnc_scheduler;
645 /* called holding global lock (read or irq-write) and caller may
646 * not have dropped this lock between finding conn and calling me,
647 * so we don't need the {get,put}connsock dance to deref
649 LASSERT(!conn->ksnc_closing);
650 LASSERT(tx->tx_resid == tx->tx_nob);
652 CDEBUG (D_NET, "Sending to "LPX64" ip %d.%d.%d.%d:%d\n",
653 conn->ksnc_peer->ksnp_nid,
654 HIPQUAD(conn->ksnc_ipaddr),
657 atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
661 zccd_init (&tx->tx_zccd, ksocknal_zc_callback);
662 /* NB this sets 1 ref on zccd, so the callback can only occur after
663 * I've released this ref. */
665 spin_lock_irqsave (&sched->kss_lock, flags);
667 if (list_empty(&conn->ksnc_tx_queue) &&
668 SOCK_WMEM_QUEUED(conn->ksnc_sock) == 0) {
669 /* First packet starts the timeout */
670 conn->ksnc_tx_deadline = cfs_time_shift(ksocknal_tunables.ksnd_io_timeout);
671 conn->ksnc_tx_bufnob = 0;
672 mb(); /* order with adding to tx_queue */
675 list_add_tail (&tx->tx_list, &conn->ksnc_tx_queue);
677 if (conn->ksnc_tx_ready && /* able to send */
678 !conn->ksnc_tx_scheduled) { /* not scheduled to send */
679 /* +1 ref for scheduler */
680 atomic_inc (&conn->ksnc_refcount);
681 list_add_tail (&conn->ksnc_tx_list,
682 &sched->kss_tx_conns);
683 conn->ksnc_tx_scheduled = 1;
684 cfs_waitq_signal (&sched->kss_waitq);
687 spin_unlock_irqrestore (&sched->kss_lock, flags);
691 ksocknal_find_connectable_route_locked (ksock_peer_t *peer)
693 struct list_head *tmp;
694 ksock_route_t *route;
697 list_for_each (tmp, &peer->ksnp_routes) {
698 route = list_entry (tmp, ksock_route_t, ksnr_list);
699 bits = route->ksnr_connected;
701 /* All typed connections established? */
702 if ((bits & KSNR_TYPED_ROUTES) == KSNR_TYPED_ROUTES)
705 /* Untyped connection established? */
706 if ((bits & (1 << SOCKNAL_CONN_ANY)) != 0)
709 /* connection being established? */
710 if (route->ksnr_connecting != 0)
713 /* too soon to retry this guy? */
714 if (!cfs_time_aftereq (cfs_time_current(), route->ksnr_timeout))
724 ksocknal_find_connecting_route_locked (ksock_peer_t *peer)
726 struct list_head *tmp;
727 ksock_route_t *route;
729 list_for_each (tmp, &peer->ksnp_routes) {
730 route = list_entry (tmp, ksock_route_t, ksnr_list);
732 if (route->ksnr_connecting != 0)
740 ksocknal_launch_packet (ksock_tx_t *tx, ptl_nid_t nid)
745 ksock_route_t *route;
748 /* Ensure the frags we've been given EXACTLY match the number of
749 * bytes we want to send. Many TCP/IP stacks disregard any total
750 * size parameters passed to them and just look at the frags.
752 * We always expect at least 1 mapped fragment containing the
753 * complete portals header. */
754 LASSERT (lib_iov_nob (tx->tx_niov, tx->tx_iov) +
755 lib_kiov_nob (tx->tx_nkiov, tx->tx_kiov) == tx->tx_nob);
756 LASSERT (tx->tx_niov >= 1);
757 LASSERT (tx->tx_iov[0].iov_len >= sizeof (ptl_hdr_t));
759 CDEBUG (D_NET, "packet %p type %d, nob %d niov %d nkiov %d\n",
760 tx, ((ptl_hdr_t *)tx->tx_iov[0].iov_base)->type,
761 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
763 tx->tx_conn = NULL; /* only set when assigned a conn */
764 tx->tx_resid = tx->tx_nob;
765 tx->tx_hdr = (ptl_hdr_t *)tx->tx_iov[0].iov_base;
767 g_lock = &ksocknal_data.ksnd_global_lock;
768 #if !SOCKNAL_ROUND_ROBIN
771 peer = ksocknal_find_target_peer_locked (tx, nid);
773 read_unlock (g_lock);
774 return (-EHOSTUNREACH);
777 if (ksocknal_find_connectable_route_locked(peer) == NULL) {
778 conn = ksocknal_find_conn_locked (tx, peer);
780 /* I've got no autoconnect routes that need to be
781 * connecting and I do have an actual connection... */
782 ksocknal_queue_tx_locked (tx, conn);
783 read_unlock (g_lock);
788 /* I'll need a write lock... */
789 read_unlock (g_lock);
791 write_lock_irqsave(g_lock, flags);
793 peer = ksocknal_find_target_peer_locked (tx, nid);
795 write_unlock_irqrestore(g_lock, flags);
796 return (-EHOSTUNREACH);
800 /* launch any/all autoconnections that need it */
801 route = ksocknal_find_connectable_route_locked (peer);
805 ksocknal_launch_autoconnect_locked (route);
808 conn = ksocknal_find_conn_locked (tx, peer);
810 /* Connection exists; queue message on it */
811 ksocknal_queue_tx_locked (tx, conn);
812 write_unlock_irqrestore (g_lock, flags);
816 route = ksocknal_find_connecting_route_locked (peer);
818 /* At least 1 connection is being established; queue the
820 list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
821 write_unlock_irqrestore (g_lock, flags);
825 write_unlock_irqrestore (g_lock, flags);
826 return (-EHOSTUNREACH);
830 ksocknal_sendmsg(lib_nal_t *nal,
837 unsigned int payload_niov,
838 struct iovec *payload_iov,
839 ptl_kiov_t *payload_kiov,
840 size_t payload_offset,
847 /* NB 'private' is different depending on what we're sending.
848 * Just ignore it... */
850 CDEBUG(D_NET, "sending "LPSZ" bytes in %d frags to nid:"LPX64
851 " pid %d\n", payload_nob, payload_niov, nid , pid);
853 LASSERT (payload_nob == 0 || payload_niov > 0);
854 LASSERT (payload_niov <= PTL_MD_MAX_IOV);
856 /* It must be OK to kmap() if required */
857 LASSERT (payload_kiov == NULL || !in_interrupt ());
858 /* payload is either all vaddrs or all pages */
859 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
861 if (payload_iov != NULL)
862 desc_size = offsetof(ksock_ltx_t, ltx_iov[1 + payload_niov]);
864 desc_size = offsetof(ksock_ltx_t, ltx_kiov[payload_niov]);
866 if (in_interrupt() ||
867 type == PTL_MSG_ACK ||
868 type == PTL_MSG_REPLY) {
869 /* Can't block if in interrupt or responding to an incoming
871 PORTAL_ALLOC_ATOMIC(ltx, desc_size);
873 PORTAL_ALLOC(ltx, desc_size);
877 CERROR("Can't allocate tx desc type %d size %d %s\n",
878 type, desc_size, in_interrupt() ? "(intr)" : "");
879 return (PTL_NO_SPACE);
882 atomic_inc(&ksocknal_data.ksnd_nactive_ltxs);
884 ltx->ltx_desc_size = desc_size;
886 /* We always have 1 mapped frag for the header */
887 ltx->ltx_tx.tx_iov = ltx->ltx_iov;
888 ltx->ltx_iov[0].iov_base = <x->ltx_hdr;
889 ltx->ltx_iov[0].iov_len = sizeof(*hdr);
892 ltx->ltx_private = private;
893 ltx->ltx_cookie = cookie;
895 ltx->ltx_tx.tx_isfwd = 0;
896 ltx->ltx_tx.tx_nob = sizeof (*hdr) + payload_nob;
898 if (payload_iov != NULL) {
899 /* payload is all mapped */
900 ltx->ltx_tx.tx_kiov = NULL;
901 ltx->ltx_tx.tx_nkiov = 0;
903 ltx->ltx_tx.tx_niov =
904 1 + lib_extract_iov(payload_niov, <x->ltx_iov[1],
905 payload_niov, payload_iov,
906 payload_offset, payload_nob);
908 /* payload is all pages */
909 ltx->ltx_tx.tx_niov = 1;
911 ltx->ltx_tx.tx_kiov = ltx->ltx_kiov;
912 ltx->ltx_tx.tx_nkiov =
913 lib_extract_kiov(payload_niov, ltx->ltx_kiov,
914 payload_niov, payload_kiov,
915 payload_offset, payload_nob);
918 rc = ksocknal_launch_packet(<x->ltx_tx, nid);
922 ksocknal_free_ltx(ltx);
927 ksocknal_send (lib_nal_t *nal, void *private, lib_msg_t *cookie,
928 ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid,
929 unsigned int payload_niov, struct iovec *payload_iov,
930 size_t payload_offset, size_t payload_len)
932 return (ksocknal_sendmsg(nal, private, cookie,
934 payload_niov, payload_iov, NULL,
935 payload_offset, payload_len));
939 ksocknal_send_pages (lib_nal_t *nal, void *private, lib_msg_t *cookie,
940 ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid,
941 unsigned int payload_niov, ptl_kiov_t *payload_kiov,
942 size_t payload_offset, size_t payload_len)
944 return (ksocknal_sendmsg(nal, private, cookie,
946 payload_niov, NULL, payload_kiov,
947 payload_offset, payload_len));
951 ksocknal_fwd_packet (void *arg, kpr_fwd_desc_t *fwd)
953 ptl_nid_t nid = fwd->kprfd_gateway_nid;
954 ksock_ftx_t *ftx = (ksock_ftx_t *)&fwd->kprfd_scratch;
957 CDEBUG (D_NET, "Forwarding [%p] -> "LPX64" ("LPX64"))\n", fwd,
958 fwd->kprfd_gateway_nid, fwd->kprfd_target_nid);
960 /* I'm the gateway; must be the last hop */
961 if (nid == ksocknal_lib.libnal_ni.ni_pid.nid)
962 nid = fwd->kprfd_target_nid;
964 /* setup iov for hdr */
965 ftx->ftx_iov.iov_base = fwd->kprfd_hdr;
966 ftx->ftx_iov.iov_len = sizeof(ptl_hdr_t);
968 ftx->ftx_tx.tx_isfwd = 1; /* This is a forwarding packet */
969 ftx->ftx_tx.tx_nob = sizeof(ptl_hdr_t) + fwd->kprfd_nob;
970 ftx->ftx_tx.tx_niov = 1;
971 ftx->ftx_tx.tx_iov = &ftx->ftx_iov;
972 ftx->ftx_tx.tx_nkiov = fwd->kprfd_niov;
973 ftx->ftx_tx.tx_kiov = fwd->kprfd_kiov;
975 rc = ksocknal_launch_packet (&ftx->ftx_tx, nid);
977 kpr_fwd_done (&ksocknal_data.ksnd_router, fwd, rc);
981 ksocknal_thread_start (int (*fn)(void *arg), void *arg)
983 long pid = cfs_kernel_thread (fn, arg, 0);
989 write_lock_irqsave(&ksocknal_data.ksnd_global_lock, flags);
990 ksocknal_data.ksnd_nthreads++;
991 write_unlock_irqrestore(&ksocknal_data.ksnd_global_lock, flags);
996 ksocknal_thread_fini (void)
1000 write_lock_irqsave(&ksocknal_data.ksnd_global_lock, flags);
1001 ksocknal_data.ksnd_nthreads--;
1002 write_unlock_irqrestore(&ksocknal_data.ksnd_global_lock, flags);
1006 ksocknal_fmb_callback (void *arg, int error)
1008 ksock_fmb_t *fmb = (ksock_fmb_t *)arg;
1009 ksock_fmb_pool_t *fmp = fmb->fmb_pool;
1010 ptl_hdr_t *hdr = &fmb->fmb_hdr;
1011 ksock_conn_t *conn = NULL;
1012 ksock_sched_t *sched;
1013 unsigned long flags;
1014 char ipbuf[PTL_NALFMT_SIZE];
1015 char ipbuf2[PTL_NALFMT_SIZE];
1018 CERROR("Failed to route packet from "
1019 LPX64" %s to "LPX64" %s: %d\n",
1020 le64_to_cpu(hdr->src_nid),
1021 portals_nid2str(SOCKNAL, le64_to_cpu(hdr->src_nid), ipbuf),
1022 le64_to_cpu(hdr->dest_nid),
1023 portals_nid2str(SOCKNAL, le64_to_cpu(hdr->dest_nid), ipbuf2),
1026 CDEBUG (D_NET, "routed packet from "LPX64" to "LPX64": OK\n",
1027 le64_to_cpu(hdr->src_nid), le64_to_cpu(hdr->dest_nid));
1029 /* drop peer ref taken on init */
1030 ksocknal_put_peer (fmb->fmb_peer);
1032 spin_lock_irqsave (&fmp->fmp_lock, flags);
1034 list_add (&fmb->fmb_list, &fmp->fmp_idle_fmbs);
1035 fmp->fmp_nactive_fmbs--;
1037 if (!list_empty (&fmp->fmp_blocked_conns)) {
1038 conn = list_entry (fmb->fmb_pool->fmp_blocked_conns.next,
1039 ksock_conn_t, ksnc_rx_list);
1040 list_del (&conn->ksnc_rx_list);
1043 spin_unlock_irqrestore (&fmp->fmp_lock, flags);
1048 CDEBUG (D_NET, "Scheduling conn %p\n", conn);
1049 LASSERT (conn->ksnc_rx_scheduled);
1050 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_FMB_SLEEP);
1052 conn->ksnc_rx_state = SOCKNAL_RX_GET_FMB;
1054 sched = conn->ksnc_scheduler;
1056 spin_lock_irqsave (&sched->kss_lock, flags);
1058 list_add_tail (&conn->ksnc_rx_list, &sched->kss_rx_conns);
1059 cfs_waitq_signal (&sched->kss_waitq);
1061 spin_unlock_irqrestore (&sched->kss_lock, flags);
1065 ksocknal_get_idle_fmb (ksock_conn_t *conn)
1067 int payload_nob = conn->ksnc_rx_nob_left;
1068 unsigned long flags;
1069 ksock_fmb_pool_t *pool;
1072 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB);
1073 LASSERT (kpr_routing(&ksocknal_data.ksnd_router));
1075 if (payload_nob <= SOCKNAL_SMALL_FWD_PAGES * CFS_PAGE_SIZE)
1076 pool = &ksocknal_data.ksnd_small_fmp;
1078 pool = &ksocknal_data.ksnd_large_fmp;
1080 spin_lock_irqsave (&pool->fmp_lock, flags);
1082 if (!list_empty (&pool->fmp_idle_fmbs)) {
1083 fmb = list_entry(pool->fmp_idle_fmbs.next,
1084 ksock_fmb_t, fmb_list);
1085 list_del (&fmb->fmb_list);
1086 pool->fmp_nactive_fmbs++;
1087 spin_unlock_irqrestore (&pool->fmp_lock, flags);
1092 /* deschedule until fmb free */
1094 conn->ksnc_rx_state = SOCKNAL_RX_FMB_SLEEP;
1096 list_add_tail (&conn->ksnc_rx_list,
1097 &pool->fmp_blocked_conns);
1099 spin_unlock_irqrestore (&pool->fmp_lock, flags);
1104 ksocknal_init_fmb (ksock_conn_t *conn, ksock_fmb_t *fmb)
1106 int payload_nob = conn->ksnc_rx_nob_left;
1107 ptl_nid_t dest_nid = le64_to_cpu(conn->ksnc_hdr.dest_nid);
1109 int nob = payload_nob;
1111 LASSERT (conn->ksnc_rx_scheduled);
1112 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB);
1113 LASSERT (conn->ksnc_rx_nob_wanted == conn->ksnc_rx_nob_left);
1114 LASSERT (payload_nob >= 0);
1115 LASSERT (payload_nob <= fmb->fmb_pool->fmp_buff_pages * CFS_PAGE_SIZE);
1116 LASSERT (sizeof (ptl_hdr_t) < CFS_PAGE_SIZE);
1117 LASSERT (fmb->fmb_kiov[0].kiov_offset == 0);
1119 /* Take a ref on the conn's peer to prevent module unload before
1120 * forwarding completes. */
1121 fmb->fmb_peer = conn->ksnc_peer;
1122 atomic_inc (&conn->ksnc_peer->ksnp_refcount);
1124 /* Copy the header we just read into the forwarding buffer. If
1125 * there's payload, start reading reading it into the buffer,
1126 * otherwise the forwarding buffer can be kicked off
1128 fmb->fmb_hdr = conn->ksnc_hdr;
1131 LASSERT (niov < fmb->fmb_pool->fmp_buff_pages);
1132 LASSERT (fmb->fmb_kiov[niov].kiov_offset == 0);
1133 fmb->fmb_kiov[niov].kiov_len = MIN (CFS_PAGE_SIZE, nob);
1134 nob -= CFS_PAGE_SIZE;
1138 kpr_fwd_init(&fmb->fmb_fwd, dest_nid, &fmb->fmb_hdr,
1139 payload_nob, niov, fmb->fmb_kiov,
1140 ksocknal_fmb_callback, fmb);
1142 if (payload_nob == 0) { /* got complete packet already */
1143 CDEBUG (D_NET, "%p "LPX64"->"LPX64" fwd_start (immediate)\n",
1144 conn, le64_to_cpu(conn->ksnc_hdr.src_nid), dest_nid);
1146 kpr_fwd_start (&ksocknal_data.ksnd_router, &fmb->fmb_fwd);
1148 ksocknal_new_packet (conn, 0); /* on to next packet */
1152 conn->ksnc_cookie = fmb; /* stash fmb for later */
1153 conn->ksnc_rx_state = SOCKNAL_RX_BODY_FWD; /* read in the payload */
1155 /* Set up conn->ksnc_rx_kiov to read the payload into fmb's kiov-ed
1157 LASSERT (niov <= sizeof(conn->ksnc_rx_iov_space)/sizeof(ptl_kiov_t));
1159 conn->ksnc_rx_niov = 0;
1160 conn->ksnc_rx_nkiov = niov;
1161 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1162 memcpy(conn->ksnc_rx_kiov, fmb->fmb_kiov, niov * sizeof(ptl_kiov_t));
1164 CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d reading body\n", conn,
1165 le64_to_cpu(conn->ksnc_hdr.src_nid), dest_nid, payload_nob);
1170 ksocknal_fwd_parse (ksock_conn_t *conn)
1173 ptl_nid_t dest_nid = le64_to_cpu(conn->ksnc_hdr.dest_nid);
1174 ptl_nid_t src_nid = le64_to_cpu(conn->ksnc_hdr.src_nid);
1175 int body_len = le32_to_cpu(conn->ksnc_hdr.payload_length);
1176 char str[PTL_NALFMT_SIZE];
1177 char str2[PTL_NALFMT_SIZE];
1179 CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d parsing header\n", conn,
1180 src_nid, dest_nid, conn->ksnc_rx_nob_left);
1182 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_HEADER);
1183 LASSERT (conn->ksnc_rx_scheduled);
1185 if (body_len < 0) { /* length corrupt (overflow) */
1186 CERROR("dropping packet from "LPX64" (%s) for "LPX64" (%s): "
1187 "packet size %d illegal\n",
1188 src_nid, portals_nid2str(TCPNAL, src_nid, str),
1189 dest_nid, portals_nid2str(TCPNAL, dest_nid, str2),
1192 ksocknal_new_packet (conn, 0); /* on to new packet */
1196 if (!kpr_routing(&ksocknal_data.ksnd_router)) { /* not forwarding */
1197 CERROR("dropping packet from "LPX64" (%s) for "LPX64
1198 " (%s): not forwarding\n",
1199 src_nid, portals_nid2str(TCPNAL, src_nid, str),
1200 dest_nid, portals_nid2str(TCPNAL, dest_nid, str2));
1201 /* on to new packet (skip this one's body) */
1202 ksocknal_new_packet (conn, body_len);
1206 if (body_len > PTL_MTU) { /* too big to forward */
1207 CERROR ("dropping packet from "LPX64" (%s) for "LPX64
1208 "(%s): packet size %d too big\n",
1209 src_nid, portals_nid2str(TCPNAL, src_nid, str),
1210 dest_nid, portals_nid2str(TCPNAL, dest_nid, str2),
1212 /* on to new packet (skip this one's body) */
1213 ksocknal_new_packet (conn, body_len);
1217 /* should have gone direct */
1218 peer = ksocknal_get_peer (conn->ksnc_hdr.dest_nid);
1220 CERROR ("dropping packet from "LPX64" (%s) for "LPX64
1221 "(%s): target is a peer\n",
1222 src_nid, portals_nid2str(TCPNAL, src_nid, str),
1223 dest_nid, portals_nid2str(TCPNAL, dest_nid, str2));
1224 ksocknal_put_peer (peer); /* drop ref from get above */
1226 /* on to next packet (skip this one's body) */
1227 ksocknal_new_packet (conn, body_len);
1231 conn->ksnc_rx_state = SOCKNAL_RX_GET_FMB; /* Getting FMB now */
1232 conn->ksnc_rx_nob_left = body_len; /* stash packet size */
1233 conn->ksnc_rx_nob_wanted = body_len; /* (no slop) */
1237 ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip)
1239 static char ksocknal_slop_buffer[4096];
1245 if (nob_to_skip == 0) { /* right at next packet boundary now */
1246 conn->ksnc_rx_started = 0;
1247 mb (); /* racing with timeout thread */
1249 conn->ksnc_rx_state = SOCKNAL_RX_HEADER;
1250 conn->ksnc_rx_nob_wanted = sizeof (ptl_hdr_t);
1251 conn->ksnc_rx_nob_left = sizeof (ptl_hdr_t);
1253 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1254 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_hdr;
1255 conn->ksnc_rx_iov[0].iov_len = sizeof (ptl_hdr_t);
1256 conn->ksnc_rx_niov = 1;
1258 conn->ksnc_rx_kiov = NULL;
1259 conn->ksnc_rx_nkiov = 0;
1263 /* Set up to skip as much a possible now. If there's more left
1264 * (ran out of iov entries) we'll get called again */
1266 conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1267 conn->ksnc_rx_nob_left = nob_to_skip;
1268 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1273 nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer));
1275 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1276 conn->ksnc_rx_iov[niov].iov_len = nob;
1281 } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */
1282 niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec));
1284 conn->ksnc_rx_niov = niov;
1285 conn->ksnc_rx_kiov = NULL;
1286 conn->ksnc_rx_nkiov = 0;
1287 conn->ksnc_rx_nob_wanted = skipped;
1292 ksocknal_process_receive (ksock_conn_t *conn)
1297 LASSERT (atomic_read (&conn->ksnc_refcount) > 0);
1299 /* doesn't need a forwarding buffer */
1300 if (conn->ksnc_rx_state != SOCKNAL_RX_GET_FMB)
1304 fmb = ksocknal_get_idle_fmb (conn);
1306 /* conn descheduled waiting for idle fmb */
1310 if (ksocknal_init_fmb (conn, fmb)) {
1311 /* packet forwarded */
1316 /* NB: sched lock NOT held */
1317 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_HEADER ||
1318 conn->ksnc_rx_state == SOCKNAL_RX_BODY ||
1319 conn->ksnc_rx_state == SOCKNAL_RX_BODY_FWD ||
1320 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1322 LASSERT (conn->ksnc_rx_nob_wanted > 0);
1324 rc = ksocknal_receive(conn);
1327 LASSERT (rc != -EAGAIN);
1330 CWARN ("[%p] EOF from "LPX64" ip %d.%d.%d.%d:%d\n",
1331 conn, conn->ksnc_peer->ksnp_nid,
1332 HIPQUAD(conn->ksnc_ipaddr),
1334 else if (!conn->ksnc_closing)
1335 CERROR ("[%p] Error %d on read from "LPX64
1336 " ip %d.%d.%d.%d:%d\n",
1337 conn, rc, conn->ksnc_peer->ksnp_nid,
1338 HIPQUAD(conn->ksnc_ipaddr),
1341 ksocknal_close_conn_and_siblings (conn, rc);
1342 return (rc == 0 ? -ESHUTDOWN : rc);
1345 if (conn->ksnc_rx_nob_wanted != 0) {
1350 switch (conn->ksnc_rx_state) {
1351 case SOCKNAL_RX_HEADER:
1352 if (conn->ksnc_hdr.type != cpu_to_le32(PTL_MSG_HELLO) &&
1353 le64_to_cpu(conn->ksnc_hdr.dest_nid) !=
1354 ksocknal_lib.libnal_ni.ni_pid.nid) {
1355 /* This packet isn't for me */
1356 ksocknal_fwd_parse (conn);
1357 switch (conn->ksnc_rx_state) {
1358 case SOCKNAL_RX_HEADER: /* skipped (zero payload) */
1359 return (0); /* => come back later */
1360 case SOCKNAL_RX_SLOP: /* skipping packet's body */
1361 goto try_read; /* => go read it */
1362 case SOCKNAL_RX_GET_FMB: /* forwarding */
1363 goto get_fmb; /* => go get a fwd msg buffer */
1370 /* sets wanted_len, iovs etc */
1371 rc = lib_parse(&ksocknal_lib, &conn->ksnc_hdr, conn);
1374 /* I just received garbage: give up on this conn */
1375 ksocknal_close_conn_and_siblings (conn, rc);
1379 if (conn->ksnc_rx_nob_wanted != 0) { /* need to get payload? */
1380 conn->ksnc_rx_state = SOCKNAL_RX_BODY;
1381 goto try_read; /* go read the payload */
1383 /* Fall through (completed packet for me) */
1385 case SOCKNAL_RX_BODY:
1386 /* payload all received */
1387 lib_finalize(&ksocknal_lib, NULL, conn->ksnc_cookie, PTL_OK);
1390 case SOCKNAL_RX_SLOP:
1391 /* starting new packet? */
1392 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1393 return (0); /* come back later */
1394 goto try_read; /* try to finish reading slop now */
1396 case SOCKNAL_RX_BODY_FWD:
1397 /* payload all received */
1398 CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d fwd_start (got body)\n",
1399 conn, le64_to_cpu(conn->ksnc_hdr.src_nid),
1400 le64_to_cpu(conn->ksnc_hdr.dest_nid),
1401 conn->ksnc_rx_nob_left);
1403 /* forward the packet. NB ksocknal_init_fmb() put fmb into
1404 * conn->ksnc_cookie */
1405 fmb = (ksock_fmb_t *)conn->ksnc_cookie;
1406 kpr_fwd_start (&ksocknal_data.ksnd_router, &fmb->fmb_fwd);
1408 /* no slop in forwarded packets */
1409 LASSERT (conn->ksnc_rx_nob_left == 0);
1411 ksocknal_new_packet (conn, 0); /* on to next packet */
1412 return (0); /* (later) */
1420 return (-EINVAL); /* keep gcc happy */
1424 ksocknal_recv (lib_nal_t *nal, void *private, lib_msg_t *msg,
1425 unsigned int niov, struct iovec *iov,
1426 size_t offset, size_t mlen, size_t rlen)
1428 ksock_conn_t *conn = (ksock_conn_t *)private;
1430 LASSERT (mlen <= rlen);
1431 LASSERT (niov <= PTL_MD_MAX_IOV);
1433 conn->ksnc_cookie = msg;
1434 conn->ksnc_rx_nob_wanted = mlen;
1435 conn->ksnc_rx_nob_left = rlen;
1437 conn->ksnc_rx_nkiov = 0;
1438 conn->ksnc_rx_kiov = NULL;
1439 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1440 conn->ksnc_rx_niov =
1441 lib_extract_iov(PTL_MD_MAX_IOV, conn->ksnc_rx_iov,
1442 niov, iov, offset, mlen);
1445 lib_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1446 lib_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1452 ksocknal_recv_pages (lib_nal_t *nal, void *private, lib_msg_t *msg,
1453 unsigned int niov, ptl_kiov_t *kiov,
1454 size_t offset, size_t mlen, size_t rlen)
1456 ksock_conn_t *conn = (ksock_conn_t *)private;
1458 LASSERT (mlen <= rlen);
1459 LASSERT (niov <= PTL_MD_MAX_IOV);
1461 conn->ksnc_cookie = msg;
1462 conn->ksnc_rx_nob_wanted = mlen;
1463 conn->ksnc_rx_nob_left = rlen;
1465 conn->ksnc_rx_niov = 0;
1466 conn->ksnc_rx_iov = NULL;
1467 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1468 conn->ksnc_rx_nkiov =
1469 lib_extract_kiov(PTL_MD_MAX_IOV, conn->ksnc_rx_kiov,
1470 niov, kiov, offset, mlen);
1473 lib_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1474 lib_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1480 ksocknal_sched_cansleep(ksock_sched_t *sched)
1482 unsigned long flags;
1485 spin_lock_irqsave(&sched->kss_lock, flags);
1487 rc = (!ksocknal_data.ksnd_shuttingdown &&
1489 list_empty(&sched->kss_zctxdone_list) &&
1491 list_empty(&sched->kss_rx_conns) &&
1492 list_empty(&sched->kss_tx_conns));
1494 spin_unlock_irqrestore(&sched->kss_lock, flags);
1498 int ksocknal_scheduler (void *arg)
1500 ksock_sched_t *sched = (ksock_sched_t *)arg;
1503 unsigned long flags;
1506 int id = sched - ksocknal_data.ksnd_schedulers;
1509 snprintf (name, sizeof (name),"ksocknald_%02d", id);
1510 kportal_daemonize (name);
1511 kportal_blockallsigs ();
1513 #if (CONFIG_SMP && CPU_AFFINITY)
1514 id = ksocknal_sched2cpu(id);
1515 if (cpu_online(id)) {
1518 set_cpus_allowed(current, m);
1520 CERROR ("Can't set CPU affinity for %s to %d\n", name, id);
1522 #endif /* CONFIG_SMP && CPU_AFFINITY */
1524 spin_lock_irqsave (&sched->kss_lock, flags);
1526 while (!ksocknal_data.ksnd_shuttingdown) {
1527 int did_something = 0;
1529 /* Ensure I progress everything semi-fairly */
1531 if (!list_empty (&sched->kss_rx_conns)) {
1532 conn = list_entry(sched->kss_rx_conns.next,
1533 ksock_conn_t, ksnc_rx_list);
1534 list_del(&conn->ksnc_rx_list);
1536 LASSERT(conn->ksnc_rx_scheduled);
1537 LASSERT(conn->ksnc_rx_ready);
1539 /* clear rx_ready in case receive isn't complete.
1540 * Do it BEFORE we call process_recv, since
1541 * data_ready can set it any time after we release
1543 conn->ksnc_rx_ready = 0;
1544 spin_unlock_irqrestore(&sched->kss_lock, flags);
1546 rc = ksocknal_process_receive(conn);
1548 spin_lock_irqsave(&sched->kss_lock, flags);
1550 /* I'm the only one that can clear this flag */
1551 LASSERT(conn->ksnc_rx_scheduled);
1553 /* Did process_receive get everything it wanted? */
1555 conn->ksnc_rx_ready = 1;
1557 if (conn->ksnc_rx_state == SOCKNAL_RX_FMB_SLEEP ||
1558 conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB) {
1559 /* Conn blocked for a forwarding buffer.
1560 * It will get queued for my attention when
1561 * one becomes available (and it might just
1562 * already have been!). Meanwhile my ref
1563 * on it stays put. */
1564 } else if (conn->ksnc_rx_ready) {
1565 /* reschedule for rx */
1566 list_add_tail (&conn->ksnc_rx_list,
1567 &sched->kss_rx_conns);
1569 conn->ksnc_rx_scheduled = 0;
1571 ksocknal_put_conn(conn);
1577 if (!list_empty (&sched->kss_tx_conns)) {
1578 conn = list_entry(sched->kss_tx_conns.next,
1579 ksock_conn_t, ksnc_tx_list);
1580 list_del (&conn->ksnc_tx_list);
1582 LASSERT(conn->ksnc_tx_scheduled);
1583 LASSERT(conn->ksnc_tx_ready);
1584 LASSERT(!list_empty(&conn->ksnc_tx_queue));
1586 tx = list_entry(conn->ksnc_tx_queue.next,
1587 ksock_tx_t, tx_list);
1588 /* dequeue now so empty list => more to send */
1589 list_del(&tx->tx_list);
1591 /* Clear tx_ready in case send isn't complete. Do
1592 * it BEFORE we call process_transmit, since
1593 * write_space can set it any time after we release
1595 conn->ksnc_tx_ready = 0;
1596 spin_unlock_irqrestore (&sched->kss_lock, flags);
1598 rc = ksocknal_process_transmit(conn, tx);
1600 spin_lock_irqsave (&sched->kss_lock, flags);
1602 if (rc == -ENOMEM || rc == -EAGAIN) {
1603 /* Incomplete send: replace tx on HEAD of tx_queue */
1604 list_add (&tx->tx_list, &conn->ksnc_tx_queue);
1606 /* Complete send; assume space for more */
1607 conn->ksnc_tx_ready = 1;
1610 if (rc == -ENOMEM) {
1611 /* Do nothing; after a short timeout, this
1612 * conn will be reposted on kss_tx_conns. */
1613 } else if (conn->ksnc_tx_ready &&
1614 !list_empty (&conn->ksnc_tx_queue)) {
1615 /* reschedule for tx */
1616 list_add_tail (&conn->ksnc_tx_list,
1617 &sched->kss_tx_conns);
1619 conn->ksnc_tx_scheduled = 0;
1621 ksocknal_put_conn (conn);
1627 if (!list_empty (&sched->kss_zctxdone_list)) {
1629 list_entry(sched->kss_zctxdone_list.next,
1630 ksock_tx_t, tx_list);
1633 list_del (&tx->tx_list);
1634 spin_unlock_irqrestore (&sched->kss_lock, flags);
1636 ksocknal_tx_done (tx, 1);
1638 spin_lock_irqsave (&sched->kss_lock, flags);
1641 if (!did_something || /* nothing to do */
1642 ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1643 spin_unlock_irqrestore (&sched->kss_lock, flags);
1647 if (!did_something) { /* wait for something to do */
1648 rc = wait_event_interruptible (sched->kss_waitq,
1649 !ksocknal_sched_cansleep(sched));
1654 spin_lock_irqsave (&sched->kss_lock, flags);
1658 spin_unlock_irqrestore (&sched->kss_lock, flags);
1659 ksocknal_thread_fini ();
1664 * Add connection to kss_rx_conns of scheduler
1665 * and wakeup the scheduler.
1667 void ksocknal_read_callback (ksock_conn_t *conn)
1669 ksock_sched_t *sched;
1670 unsigned long flags;
1673 sched = conn->ksnc_scheduler;
1675 spin_lock_irqsave (&sched->kss_lock, flags);
1677 conn->ksnc_rx_ready = 1;
1679 if (!conn->ksnc_rx_scheduled) { /* not being progressed */
1680 list_add_tail(&conn->ksnc_rx_list,
1681 &sched->kss_rx_conns);
1682 conn->ksnc_rx_scheduled = 1;
1683 /* extra ref for scheduler */
1684 atomic_inc (&conn->ksnc_refcount);
1686 cfs_waitq_signal (&sched->kss_waitq);
1688 spin_unlock_irqrestore (&sched->kss_lock, flags);
1694 * Add connection to kss_tx_conns of scheduler
1695 * and wakeup the scheduler.
1697 void ksocknal_write_callback (ksock_conn_t *conn)
1699 ksock_sched_t *sched;
1700 unsigned long flags;
1703 sched = conn->ksnc_scheduler;
1705 spin_lock_irqsave (&sched->kss_lock, flags);
1707 conn->ksnc_tx_ready = 1;
1709 if (!conn->ksnc_tx_scheduled && // not being progressed
1710 !list_empty(&conn->ksnc_tx_queue)){//packets to send
1711 list_add_tail (&conn->ksnc_tx_list,
1712 &sched->kss_tx_conns);
1713 conn->ksnc_tx_scheduled = 1;
1714 /* extra ref for scheduler */
1715 atomic_inc (&conn->ksnc_refcount);
1717 cfs_waitq_signal (&sched->kss_waitq);
1720 spin_unlock_irqrestore (&sched->kss_lock, flags);
1726 ksocknal_sock_write (struct socket *sock, void *buffer, int nob)
1728 return ksocknal_lib_sock_write(sock, buffer, nob);
1732 ksocknal_sock_read (struct socket *sock, void *buffer, int nob)
1734 return ksocknal_lib_sock_read(sock, buffer, nob);
1738 ksocknal_send_hello (ksock_conn_t *conn, __u32 *ipaddrs, int nipaddrs)
1740 /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1741 struct socket *sock = conn->ksnc_sock;
1743 ptl_magicversion_t *hmv = (ptl_magicversion_t *)&hdr.dest_nid;
1747 LASSERT (conn->ksnc_type != SOCKNAL_CONN_NONE);
1748 LASSERT (nipaddrs <= SOCKNAL_MAX_INTERFACES);
1750 /* No need for getconnsock/putconnsock */
1751 LASSERT (!conn->ksnc_closing);
1753 LASSERT (sizeof (*hmv) == sizeof (hdr.dest_nid));
1754 hmv->magic = cpu_to_le32 (PORTALS_PROTO_MAGIC);
1755 hmv->version_major = cpu_to_le16 (PORTALS_PROTO_VERSION_MAJOR);
1756 hmv->version_minor = cpu_to_le16 (PORTALS_PROTO_VERSION_MINOR);
1758 hdr.src_nid = cpu_to_le64 (ksocknal_lib.libnal_ni.ni_pid.nid);
1759 hdr.type = cpu_to_le32 (PTL_MSG_HELLO);
1760 hdr.payload_length = cpu_to_le32 (nipaddrs * sizeof(*ipaddrs));
1762 hdr.msg.hello.type = cpu_to_le32 (conn->ksnc_type);
1763 hdr.msg.hello.incarnation =
1764 cpu_to_le64 (ksocknal_data.ksnd_incarnation);
1766 /* Receiver is eager */
1767 rc = ksocknal_sock_write (sock, &hdr, sizeof(hdr));
1769 CERROR ("Error %d sending HELLO hdr to %u.%u.%u.%u/%d\n",
1770 rc, HIPQUAD(conn->ksnc_ipaddr), conn->ksnc_port);
1777 for (i = 0; i < nipaddrs; i++) {
1778 ipaddrs[i] = __cpu_to_le32 (ipaddrs[i]);
1781 rc = ksocknal_sock_write (sock, ipaddrs, nipaddrs * sizeof(*ipaddrs));
1783 CERROR ("Error %d sending HELLO payload (%d)"
1784 " to %u.%u.%u.%u/%d\n", rc, nipaddrs,
1785 HIPQUAD(conn->ksnc_ipaddr), conn->ksnc_port);
1790 ksocknal_invert_type(int type)
1794 case SOCKNAL_CONN_ANY:
1795 case SOCKNAL_CONN_CONTROL:
1797 case SOCKNAL_CONN_BULK_IN:
1798 return SOCKNAL_CONN_BULK_OUT;
1799 case SOCKNAL_CONN_BULK_OUT:
1800 return SOCKNAL_CONN_BULK_IN;
1802 return (SOCKNAL_CONN_NONE);
1807 ksocknal_recv_hello (ksock_conn_t *conn, ptl_nid_t *nid,
1808 __u64 *incarnation, __u32 *ipaddrs)
1810 struct socket *sock = conn->ksnc_sock;
1816 ptl_magicversion_t *hmv;
1817 char ipbuf[PTL_NALFMT_SIZE];
1819 hmv = (ptl_magicversion_t *)&hdr.dest_nid;
1820 LASSERT (sizeof (*hmv) == sizeof (hdr.dest_nid));
1822 rc = ksocknal_sock_read (sock, hmv, sizeof (*hmv));
1824 CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n",
1825 rc, HIPQUAD(conn->ksnc_ipaddr));
1829 if (hmv->magic != le32_to_cpu (PORTALS_PROTO_MAGIC)) {
1830 CERROR ("Bad magic %#08x (%#08x expected) from %u.%u.%u.%u\n",
1831 __cpu_to_le32 (hmv->magic), PORTALS_PROTO_MAGIC,
1832 HIPQUAD(conn->ksnc_ipaddr));
1836 if (hmv->version_major != cpu_to_le16 (PORTALS_PROTO_VERSION_MAJOR) ||
1837 hmv->version_minor != cpu_to_le16 (PORTALS_PROTO_VERSION_MINOR)) {
1838 CERROR ("Incompatible protocol version %d.%d (%d.%d expected)"
1839 " from %u.%u.%u.%u\n",
1840 le16_to_cpu (hmv->version_major),
1841 le16_to_cpu (hmv->version_minor),
1842 PORTALS_PROTO_VERSION_MAJOR,
1843 PORTALS_PROTO_VERSION_MINOR,
1844 HIPQUAD(conn->ksnc_ipaddr));
1848 #if (PORTALS_PROTO_VERSION_MAJOR != 1)
1849 # error "This code only understands protocol version 1.x"
1851 /* version 1 sends magic/version as the dest_nid of a 'hello'
1852 * header, followed by payload full of interface IP addresses.
1853 * Read the rest of it in now... */
1855 rc = ksocknal_sock_read (sock, hmv + 1, sizeof (hdr) - sizeof (*hmv));
1857 CERROR ("Error %d reading rest of HELLO hdr from %u.%u.%u.%u\n",
1858 rc, HIPQUAD(conn->ksnc_ipaddr));
1862 /* ...and check we got what we expected */
1863 if (hdr.type != cpu_to_le32 (PTL_MSG_HELLO)) {
1864 CERROR ("Expecting a HELLO hdr,"
1865 " but got type %d from %u.%u.%u.%u\n",
1866 le32_to_cpu (hdr.type),
1867 HIPQUAD(conn->ksnc_ipaddr));
1871 if (le64_to_cpu(hdr.src_nid) == PTL_NID_ANY) {
1872 CERROR("Expecting a HELLO hdr with a NID, but got PTL_NID_ANY"
1873 "from %u.%u.%u.%u\n", HIPQUAD(conn->ksnc_ipaddr));
1877 if (*nid == PTL_NID_ANY) { /* don't know peer's nid yet */
1878 *nid = le64_to_cpu(hdr.src_nid);
1879 } else if (*nid != le64_to_cpu (hdr.src_nid)) {
1880 LCONSOLE_ERROR("Connected successfully to nid "LPX64" on host "
1881 "%u.%u.%u.%u, but they claimed they were nid "
1882 LPX64" (%s); please check your Lustre "
1884 *nid, HIPQUAD(conn->ksnc_ipaddr),
1885 le64_to_cpu(hdr.src_nid),
1886 portals_nid2str(SOCKNAL,
1887 le64_to_cpu(hdr.src_nid),
1890 CERROR ("Connected to nid "LPX64"@%u.%u.%u.%u "
1891 "but expecting "LPX64"\n",
1892 le64_to_cpu (hdr.src_nid),
1893 HIPQUAD(conn->ksnc_ipaddr), *nid);
1897 type = __le32_to_cpu(hdr.msg.hello.type);
1899 if (conn->ksnc_type == SOCKNAL_CONN_NONE) {
1900 /* I've accepted this connection; peer determines type */
1901 conn->ksnc_type = ksocknal_invert_type(type);
1902 if (conn->ksnc_type == SOCKNAL_CONN_NONE) {
1903 CERROR ("Unexpected type %d from "LPX64"@%u.%u.%u.%u\n",
1904 type, *nid, HIPQUAD(conn->ksnc_ipaddr));
1907 } else if (ksocknal_invert_type(type) != conn->ksnc_type) {
1908 CERROR ("Mismatched types: me %d, "LPX64"@%u.%u.%u.%u %d\n",
1909 conn->ksnc_type, *nid, HIPQUAD(conn->ksnc_ipaddr),
1910 le32_to_cpu(hdr.msg.hello.type));
1914 *incarnation = le64_to_cpu(hdr.msg.hello.incarnation);
1916 nips = __le32_to_cpu (hdr.payload_length) / sizeof (__u32);
1918 if (nips > SOCKNAL_MAX_INTERFACES ||
1919 nips * sizeof(__u32) != __le32_to_cpu (hdr.payload_length)) {
1920 CERROR("Bad payload length %d from "LPX64"@%u.%u.%u.%u\n",
1921 __le32_to_cpu (hdr.payload_length),
1922 *nid, HIPQUAD(conn->ksnc_ipaddr));
1928 rc = ksocknal_sock_read (sock, ipaddrs, nips * sizeof(*ipaddrs));
1930 CERROR ("Error %d reading IPs from "LPX64"@%u.%u.%u.%u\n",
1931 rc, *nid, HIPQUAD(conn->ksnc_ipaddr));
1935 for (i = 0; i < nips; i++) {
1936 ipaddrs[i] = __le32_to_cpu(ipaddrs[i]);
1938 if (ipaddrs[i] == 0) {
1939 CERROR("Zero IP[%d] from "LPX64"@%u.%u.%u.%u\n",
1940 i, *nid, HIPQUAD(conn->ksnc_ipaddr));
1949 ksocknal_get_conn_tunables (ksock_conn_t *conn, int *txmem, int *rxmem, int *nagle)
1951 return ksocknal_lib_get_conn_tunables(conn, txmem, rxmem, nagle);
1955 ksocknal_connect_peer (ksock_route_t *route, int type)
1957 struct socket *sock;
1962 /* Iterate through reserved ports. When typed connections are
1963 * used, we will need to bind to multiple ports, but we only know
1964 * this at connect time. But, by that time we've already called
1965 * bind() so we need a new socket. */
1967 for (port = 1023; port > 512; --port) {
1969 rc = ksocknal_lib_connect_sock(&sock, &may_retry, route, port);
1972 rc = ksocknal_create_conn(route, sock, type);
1973 cfs_put_file(KSN_SOCK2FILE(sock));
1981 CERROR("Out of ports trying to bind to a reserved port\n");
1982 return (-EADDRINUSE);
1986 ksocknal_autoconnect (ksock_route_t *route)
1988 CFS_LIST_HEAD (zombies);
1991 unsigned long flags;
1994 char *err_msg = NULL;
1997 for (type = 0; type < SOCKNAL_CONN_NTYPES; type++)
1998 if ((route->ksnr_connecting & (1 << type)) != 0)
2000 LASSERT (type < SOCKNAL_CONN_NTYPES);
2002 rc = ksocknal_connect_peer (route, type);
2006 /* successfully autoconnected: create_conn did the
2007 * route/conn binding and scheduled any blocked packets */
2009 if (route->ksnr_connecting == 0) {
2010 /* No more connections required */
2016 /* "normal" errors */
2018 LCONSOLE_ERROR("Connection was refused by host %u.%u.%u.%u on "
2019 "port %d; check that Lustre is running on that "
2021 HIPQUAD(route->ksnr_ipaddr),
2026 LCONSOLE_ERROR("Host %u.%u.%u.%u was unreachable; the network "
2027 "or that node may be down, or Lustre may be "
2029 HIPQUAD(route->ksnr_ipaddr));
2032 LCONSOLE_ERROR("Connecting to host %u.%u.%u.%u on port %d took "
2033 "too long; that node may be hung or "
2034 "experiencing high load.\n",
2035 HIPQUAD(route->ksnr_ipaddr),
2038 /* errors that should be rare */
2040 err_msg = "Portals could not negotiate a connection";
2044 /* -EAGAIN is out of ports, but we specify the ports
2045 * manually. we really should never get this */
2046 err_msg = "no privileged ports were available";
2049 err_msg = "unknown error";
2054 LCONSOLE_ERROR("There was an unexpected error connecting to host "
2055 "%u.%u.%u.%u on port %d: %s (error code %d).\n",
2056 HIPQUAD(route->ksnr_ipaddr),
2061 /* Connection attempt failed */
2063 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
2065 peer = route->ksnr_peer;
2066 route->ksnr_connecting = 0;
2068 /* This is a retry rather than a new connection */
2069 LASSERT (route->ksnr_retry_interval != 0);
2070 route->ksnr_timeout = cfs_time_add(cfs_time_current(),
2071 route->ksnr_retry_interval);
2072 route->ksnr_retry_interval = MIN (route->ksnr_retry_interval * 2,
2073 SOCKNAL_MAX_RECONNECT_INTERVAL);
2075 if (!list_empty (&peer->ksnp_tx_queue) &&
2076 ksocknal_find_connecting_route_locked (peer) == NULL) {
2077 LASSERT (list_empty (&peer->ksnp_conns));
2079 /* None of the connections that the blocked packets are
2080 * waiting for have been successful. Complete them now... */
2082 tx = list_entry (peer->ksnp_tx_queue.next,
2083 ksock_tx_t, tx_list);
2084 list_del (&tx->tx_list);
2085 list_add_tail (&tx->tx_list, &zombies);
2086 } while (!list_empty (&peer->ksnp_tx_queue));
2089 #if 0 /* irrelevent with only eager routes */
2090 if (!route->ksnr_deleted) {
2091 /* make this route least-favourite for re-selection */
2092 list_del(&route->ksnr_list);
2093 list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
2096 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
2098 while (!list_empty (&zombies)) {
2099 char ipbuf[PTL_NALFMT_SIZE];
2100 char ipbuf2[PTL_NALFMT_SIZE];
2101 tx = list_entry (zombies.next, ksock_tx_t, tx_list);
2103 CERROR ("Deleting packet type %d len %d ("LPX64" %s->"LPX64" %s)\n",
2104 le32_to_cpu (tx->tx_hdr->type),
2105 le32_to_cpu (tx->tx_hdr->payload_length),
2106 le64_to_cpu (tx->tx_hdr->src_nid),
2107 portals_nid2str(SOCKNAL,
2108 le64_to_cpu(tx->tx_hdr->src_nid),
2110 le64_to_cpu (tx->tx_hdr->dest_nid),
2111 portals_nid2str(SOCKNAL,
2112 le64_to_cpu(tx->tx_hdr->src_nid),
2115 list_del (&tx->tx_list);
2117 ksocknal_tx_done (tx, 0);
2122 ksocknal_autoconnectd (void *arg)
2124 long id = (long)arg;
2126 unsigned long flags;
2127 ksock_route_t *route;
2130 snprintf (name, sizeof (name), "ksocknal_ad%02ld", id);
2131 kportal_daemonize (name);
2132 kportal_blockallsigs ();
2134 spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2136 while (!ksocknal_data.ksnd_shuttingdown) {
2138 if (!list_empty (&ksocknal_data.ksnd_autoconnectd_routes)) {
2139 route = list_entry (ksocknal_data.ksnd_autoconnectd_routes.next,
2140 ksock_route_t, ksnr_connect_list);
2142 list_del (&route->ksnr_connect_list);
2143 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2145 ksocknal_autoconnect (route);
2146 ksocknal_put_route (route);
2148 spin_lock_irqsave(&ksocknal_data.ksnd_autoconnectd_lock,
2153 spin_unlock_irqrestore(&ksocknal_data.ksnd_autoconnectd_lock,
2156 rc = wait_event_interruptible(ksocknal_data.ksnd_autoconnectd_waitq,
2157 ksocknal_data.ksnd_shuttingdown ||
2158 !list_empty(&ksocknal_data.ksnd_autoconnectd_routes));
2160 spin_lock_irqsave(&ksocknal_data.ksnd_autoconnectd_lock, flags);
2163 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2165 ksocknal_thread_fini ();
2170 ksocknal_find_timed_out_conn (ksock_peer_t *peer)
2172 /* We're called with a shared lock on ksnd_global_lock */
2174 struct list_head *ctmp;
2176 list_for_each (ctmp, &peer->ksnp_conns) {
2177 conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
2179 /* Don't need the {get,put}connsock dance to deref ksnc_sock... */
2180 LASSERT (!conn->ksnc_closing);
2182 if (SOCK_ERROR(conn->ksnc_sock) != 0) {
2183 atomic_inc (&conn->ksnc_refcount);
2185 switch (SOCK_ERROR(conn->ksnc_sock)) {
2187 LCONSOLE_WARN("A connection with %u.%u.%u.%u "
2188 "was reset; they may have "
2190 HIPQUAD(conn->ksnc_ipaddr));
2193 LCONSOLE_WARN("A connection with %u.%u.%u.%u "
2194 "timed out; the network or that "
2195 "node may be down.\n",
2196 HIPQUAD(conn->ksnc_ipaddr));
2199 LCONSOLE_WARN("An unexpected network error "
2200 "occurred with %u.%u.%u.%u: %d.\n",
2201 HIPQUAD(conn->ksnc_ipaddr),
2202 SOCK_ERROR(conn->ksnc_sock));
2206 /* Something (e.g. failed keepalive) set the socket error */
2207 CERROR ("Socket error %d: "LPX64" %p %d.%d.%d.%d\n",
2208 SOCK_ERROR(conn->ksnc_sock), peer->ksnp_nid,
2209 conn, HIPQUAD(conn->ksnc_ipaddr));
2214 if (conn->ksnc_rx_started &&
2215 cfs_time_aftereq (cfs_time_current(),
2216 conn->ksnc_rx_deadline)) {
2217 /* Timed out incomplete incoming message */
2218 atomic_inc (&conn->ksnc_refcount);
2219 LCONSOLE_ERROR("A timeout occurred receiving data from "
2220 "%u.%u.%u.%u; the network or that node "
2222 HIPQUAD(conn->ksnc_ipaddr));
2223 CERROR ("Timed out RX from "LPX64" %p %d.%d.%d.%d\n",
2224 peer->ksnp_nid,conn,HIPQUAD(conn->ksnc_ipaddr));
2228 if ((!list_empty (&conn->ksnc_tx_queue) ||
2229 SOCK_WMEM_QUEUED(conn->ksnc_sock) != 0) &&
2230 cfs_time_aftereq (cfs_time_current(),
2231 conn->ksnc_tx_deadline)) {
2232 /* Timed out messages queued for sending or
2233 * buffered in the socket's send buffer */
2234 atomic_inc (&conn->ksnc_refcount);
2235 LCONSOLE_ERROR("A timeout occurred sending data to "
2236 "%u.%u.%u.%u; the network or that node "
2238 HIPQUAD(conn->ksnc_ipaddr));
2239 CERROR ("Timed out TX to "LPX64" %s%d %p %d.%d.%d.%d\n",
2241 list_empty (&conn->ksnc_tx_queue) ? "" : "Q ",
2242 SOCK_WMEM_QUEUED(conn->ksnc_sock), conn,
2243 HIPQUAD(conn->ksnc_ipaddr));
2252 ksocknal_check_peer_timeouts (int idx)
2254 struct list_head *peers = &ksocknal_data.ksnd_peers[idx];
2255 struct list_head *ptmp;
2260 /* NB. We expect to have a look at all the peers and not find any
2261 * connections to time out, so we just use a shared lock while we
2263 read_lock (&ksocknal_data.ksnd_global_lock);
2265 list_for_each (ptmp, peers) {
2266 peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
2267 conn = ksocknal_find_timed_out_conn (peer);
2270 read_unlock (&ksocknal_data.ksnd_global_lock);
2272 CERROR ("Timeout out conn->"LPX64" ip %d.%d.%d.%d:%d\n",
2274 HIPQUAD(conn->ksnc_ipaddr),
2276 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2278 /* NB we won't find this one again, but we can't
2279 * just proceed with the next peer, since we dropped
2280 * ksnd_global_lock and it might be dead already! */
2281 ksocknal_put_conn (conn);
2286 read_unlock (&ksocknal_data.ksnd_global_lock);
2290 ksocknal_reaper (void *arg)
2292 cfs_waitlink_t wait;
2293 unsigned long flags;
2295 ksock_sched_t *sched;
2296 struct list_head enomem_conns;
2298 cfs_duration_t timeout;
2301 cfs_time_t deadline = cfs_time_current();
2303 kportal_daemonize ("ksocknal_reaper");
2304 kportal_blockallsigs ();
2306 CFS_INIT_LIST_HEAD(&enomem_conns);
2307 cfs_waitlink_init (&wait);
2309 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2311 while (!ksocknal_data.ksnd_shuttingdown) {
2313 if (!list_empty (&ksocknal_data.ksnd_deathrow_conns)) {
2314 conn = list_entry (ksocknal_data.ksnd_deathrow_conns.next,
2315 ksock_conn_t, ksnc_list);
2316 list_del (&conn->ksnc_list);
2318 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2320 ksocknal_terminate_conn (conn);
2321 ksocknal_put_conn (conn);
2323 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2327 if (!list_empty (&ksocknal_data.ksnd_zombie_conns)) {
2328 conn = list_entry (ksocknal_data.ksnd_zombie_conns.next,
2329 ksock_conn_t, ksnc_list);
2330 list_del (&conn->ksnc_list);
2332 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2334 ksocknal_destroy_conn (conn);
2336 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2340 if (!list_empty (&ksocknal_data.ksnd_enomem_conns)) {
2341 list_add(&enomem_conns, &ksocknal_data.ksnd_enomem_conns);
2342 list_del_init(&ksocknal_data.ksnd_enomem_conns);
2345 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2347 /* reschedule all the connections that stalled with ENOMEM... */
2349 while (!list_empty (&enomem_conns)) {
2350 conn = list_entry (enomem_conns.next,
2351 ksock_conn_t, ksnc_tx_list);
2352 list_del (&conn->ksnc_tx_list);
2354 sched = conn->ksnc_scheduler;
2356 spin_lock_irqsave (&sched->kss_lock, flags);
2358 LASSERT (conn->ksnc_tx_scheduled);
2359 conn->ksnc_tx_ready = 1;
2360 list_add_tail (&conn->ksnc_tx_list, &sched->kss_tx_conns);
2361 cfs_waitq_signal (&sched->kss_waitq);
2363 spin_unlock_irqrestore (&sched->kss_lock, flags);
2367 /* careful with the jiffy wrap... */
2368 while ((timeout = cfs_time_sub(deadline,
2369 cfs_time_current())) <= 0) {
2372 int chunk = ksocknal_data.ksnd_peer_hash_size;
2374 /* Time to check for timeouts on a few more peers: I do
2375 * checks every 'p' seconds on a proportion of the peer
2376 * table and I need to check every connection 'n' times
2377 * within a timeout interval, to ensure I detect a
2378 * timeout on any connection within (n+1)/n times the
2379 * timeout interval. */
2381 if (ksocknal_tunables.ksnd_io_timeout > n * p)
2382 chunk = (chunk * n * p) /
2383 ksocknal_tunables.ksnd_io_timeout;
2387 for (i = 0; i < chunk; i++) {
2388 ksocknal_check_peer_timeouts (peer_index);
2389 peer_index = (peer_index + 1) %
2390 ksocknal_data.ksnd_peer_hash_size;
2393 deadline = cfs_time_add(deadline, cfs_time_seconds(p));
2396 if (nenomem_conns != 0) {
2397 /* Reduce my timeout if I rescheduled ENOMEM conns.
2398 * This also prevents me getting woken immediately
2399 * if any go back on my enomem list. */
2400 timeout = SOCKNAL_ENOMEM_RETRY;
2402 ksocknal_data.ksnd_reaper_waketime =
2403 cfs_time_add(cfs_time_current(), timeout);
2405 set_current_state (TASK_INTERRUPTIBLE);
2406 cfs_waitq_add (&ksocknal_data.ksnd_reaper_waitq, &wait);
2408 if (!ksocknal_data.ksnd_shuttingdown &&
2409 list_empty (&ksocknal_data.ksnd_deathrow_conns) &&
2410 list_empty (&ksocknal_data.ksnd_zombie_conns))
2411 cfs_waitq_timedwait (&wait, timeout);
2413 set_current_state (TASK_RUNNING);
2414 cfs_waitq_del (&ksocknal_data.ksnd_reaper_waitq, &wait);
2416 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2419 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2421 ksocknal_thread_fini ();
2425 lib_nal_t ksocknal_lib = {
2426 libnal_data: &ksocknal_data, /* NAL private data */
2427 libnal_send: ksocknal_send,
2428 libnal_send_pages: ksocknal_send_pages,
2429 libnal_recv: ksocknal_recv,
2430 libnal_recv_pages: ksocknal_recv_pages,
2431 libnal_dist: ksocknal_dist