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 */
516 LASSERT (!route->ksnr_connecting);
518 route->ksnr_connecting = 1; /* scheduling conn for autoconnectd */
519 atomic_inc (&route->ksnr_refcount); /* extra ref for autoconnectd */
521 spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
523 list_add_tail (&route->ksnr_connect_list,
524 &ksocknal_data.ksnd_autoconnectd_routes);
525 cfs_waitq_signal (&ksocknal_data.ksnd_autoconnectd_waitq);
527 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
531 ksocknal_find_target_peer_locked (ksock_tx_t *tx, ptl_nid_t nid)
533 char ipbuf[PTL_NALFMT_SIZE];
534 ptl_nid_t target_nid;
536 ksock_peer_t *peer = ksocknal_find_peer_locked (nid);
542 CERROR ("Can't send packet to "LPX64
543 " %s: routed target is not a peer\n",
544 nid, portals_nid2str(SOCKNAL, nid, ipbuf));
548 rc = kpr_lookup (&ksocknal_data.ksnd_router, nid, tx->tx_nob,
551 CERROR ("Can't route to "LPX64" %s: router error %d\n",
552 nid, portals_nid2str(SOCKNAL, nid, ipbuf), rc);
556 peer = ksocknal_find_peer_locked (target_nid);
560 CERROR ("Can't send packet to "LPX64" %s: no peer entry\n",
561 target_nid, portals_nid2str(SOCKNAL, target_nid, ipbuf));
566 ksocknal_find_conn_locked (ksock_tx_t *tx, ksock_peer_t *peer)
568 struct list_head *tmp;
569 ksock_conn_t *typed = NULL;
571 ksock_conn_t *fallback = NULL;
575 list_for_each (tmp, &peer->ksnp_conns) {
576 ksock_conn_t *c = list_entry(tmp, ksock_conn_t, ksnc_list);
577 #if SOCKNAL_ROUND_ROBIN
580 int nob = atomic_read(&c->ksnc_tx_nob) +
581 SOCK_WMEM_QUEUED(c->ksnc_sock);
583 LASSERT (!c->ksnc_closing);
585 if (fallback == NULL || nob < fnob) {
590 if (!ksocknal_tunables.ksnd_typed_conns)
593 switch (c->ksnc_type) {
595 CERROR("ksnc_type bad: %u\n", c->ksnc_type);
597 case SOCKNAL_CONN_ANY:
599 case SOCKNAL_CONN_BULK_IN:
601 case SOCKNAL_CONN_BULK_OUT:
602 if (tx->tx_nob < ksocknal_tunables.ksnd_min_bulk)
605 case SOCKNAL_CONN_CONTROL:
606 if (tx->tx_nob >= ksocknal_tunables.ksnd_min_bulk)
611 if (typed == NULL || nob < tnob) {
617 /* prefer the typed selection */
618 conn = (typed != NULL) ? typed : fallback;
620 #if SOCKNAL_ROUND_ROBIN
622 /* round-robin all else being equal */
623 list_del (&conn->ksnc_list);
624 list_add_tail (&conn->ksnc_list, &peer->ksnp_conns);
631 ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn)
634 ksock_sched_t *sched = conn->ksnc_scheduler;
636 /* called holding global lock (read or irq-write) and caller may
637 * not have dropped this lock between finding conn and calling me,
638 * so we don't need the {get,put}connsock dance to deref
640 LASSERT(!conn->ksnc_closing);
641 LASSERT(tx->tx_resid == tx->tx_nob);
643 CDEBUG (D_NET, "Sending to "LPX64" ip %d.%d.%d.%d:%d\n",
644 conn->ksnc_peer->ksnp_nid,
645 HIPQUAD(conn->ksnc_ipaddr),
648 atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
652 zccd_init (&tx->tx_zccd, ksocknal_zc_callback);
653 /* NB this sets 1 ref on zccd, so the callback can only occur after
654 * I've released this ref. */
656 spin_lock_irqsave (&sched->kss_lock, flags);
658 if (list_empty(&conn->ksnc_tx_queue) &&
659 SOCK_WMEM_QUEUED(conn->ksnc_sock) == 0) {
660 /* First packet starts the timeout */
661 conn->ksnc_tx_deadline = cfs_time_shift(ksocknal_tunables.ksnd_io_timeout);
662 conn->ksnc_tx_bufnob = 0;
663 mb(); /* order with adding to tx_queue */
666 list_add_tail (&tx->tx_list, &conn->ksnc_tx_queue);
668 if (conn->ksnc_tx_ready && /* able to send */
669 !conn->ksnc_tx_scheduled) { /* not scheduled to send */
670 /* +1 ref for scheduler */
671 atomic_inc (&conn->ksnc_refcount);
672 list_add_tail (&conn->ksnc_tx_list,
673 &sched->kss_tx_conns);
674 conn->ksnc_tx_scheduled = 1;
675 cfs_waitq_signal (&sched->kss_waitq);
678 spin_unlock_irqrestore (&sched->kss_lock, flags);
682 ksocknal_find_connectable_route_locked (ksock_peer_t *peer)
684 struct list_head *tmp;
685 ksock_route_t *route;
688 list_for_each (tmp, &peer->ksnp_routes) {
689 route = list_entry (tmp, ksock_route_t, ksnr_list);
690 bits = route->ksnr_connected;
692 if (ksocknal_tunables.ksnd_typed_conns) {
693 /* All typed connections established? */
694 if ((bits & KSNR_TYPED_ROUTES) == KSNR_TYPED_ROUTES)
697 /* Untyped connection established? */
698 if ((bits & (1 << SOCKNAL_CONN_ANY)) != 0)
702 /* connection being established? */
703 if (route->ksnr_connecting)
706 /* too soon to retry this guy? */
707 if (!cfs_time_aftereq (cfs_time_current(), route->ksnr_timeout))
717 ksocknal_find_connecting_route_locked (ksock_peer_t *peer)
719 struct list_head *tmp;
720 ksock_route_t *route;
722 list_for_each (tmp, &peer->ksnp_routes) {
723 route = list_entry (tmp, ksock_route_t, ksnr_list);
725 if (route->ksnr_connecting)
733 ksocknal_launch_packet (ksock_tx_t *tx, ptl_nid_t nid)
738 ksock_route_t *route;
741 /* Ensure the frags we've been given EXACTLY match the number of
742 * bytes we want to send. Many TCP/IP stacks disregard any total
743 * size parameters passed to them and just look at the frags.
745 * We always expect at least 1 mapped fragment containing the
746 * complete portals header. */
747 LASSERT (lib_iov_nob (tx->tx_niov, tx->tx_iov) +
748 lib_kiov_nob (tx->tx_nkiov, tx->tx_kiov) == tx->tx_nob);
749 LASSERT (tx->tx_niov >= 1);
750 LASSERT (tx->tx_iov[0].iov_len >= sizeof (ptl_hdr_t));
752 CDEBUG (D_NET, "packet %p type %d, nob %d niov %d nkiov %d\n",
753 tx, ((ptl_hdr_t *)tx->tx_iov[0].iov_base)->type,
754 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
756 tx->tx_conn = NULL; /* only set when assigned a conn */
757 tx->tx_resid = tx->tx_nob;
758 tx->tx_hdr = (ptl_hdr_t *)tx->tx_iov[0].iov_base;
760 g_lock = &ksocknal_data.ksnd_global_lock;
761 #if !SOCKNAL_ROUND_ROBIN
764 peer = ksocknal_find_target_peer_locked (tx, nid);
766 read_unlock (g_lock);
767 return (-EHOSTUNREACH);
770 if (ksocknal_find_connectable_route_locked(peer) == NULL) {
771 conn = ksocknal_find_conn_locked (tx, peer);
773 /* I've got no autoconnect routes that need to be
774 * connecting and I do have an actual connection... */
775 ksocknal_queue_tx_locked (tx, conn);
776 read_unlock (g_lock);
781 /* I'll need a write lock... */
782 read_unlock (g_lock);
784 write_lock_irqsave(g_lock, flags);
786 peer = ksocknal_find_target_peer_locked (tx, nid);
788 write_unlock_irqrestore(g_lock, flags);
789 return (-EHOSTUNREACH);
793 /* launch any/all autoconnections that need it */
794 route = ksocknal_find_connectable_route_locked (peer);
798 ksocknal_launch_autoconnect_locked (route);
801 conn = ksocknal_find_conn_locked (tx, peer);
803 /* Connection exists; queue message on it */
804 ksocknal_queue_tx_locked (tx, conn);
805 write_unlock_irqrestore (g_lock, flags);
809 route = ksocknal_find_connecting_route_locked (peer);
811 /* At least 1 connection is being established; queue the
813 list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
814 write_unlock_irqrestore (g_lock, flags);
818 write_unlock_irqrestore (g_lock, flags);
819 return (-EHOSTUNREACH);
823 ksocknal_sendmsg(lib_nal_t *nal,
830 unsigned int payload_niov,
831 struct iovec *payload_iov,
832 ptl_kiov_t *payload_kiov,
833 size_t payload_offset,
840 /* NB 'private' is different depending on what we're sending.
841 * Just ignore it... */
843 CDEBUG(D_NET, "sending "LPSZ" bytes in %d frags to nid:"LPX64
844 " pid %d\n", payload_nob, payload_niov, nid , pid);
846 LASSERT (payload_nob == 0 || payload_niov > 0);
847 LASSERT (payload_niov <= PTL_MD_MAX_IOV);
849 /* It must be OK to kmap() if required */
850 LASSERT (payload_kiov == NULL || !in_interrupt ());
851 /* payload is either all vaddrs or all pages */
852 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
854 if (payload_iov != NULL)
855 desc_size = offsetof(ksock_ltx_t, ltx_iov[1 + payload_niov]);
857 desc_size = offsetof(ksock_ltx_t, ltx_kiov[payload_niov]);
859 if (in_interrupt() ||
860 type == PTL_MSG_ACK ||
861 type == PTL_MSG_REPLY) {
862 /* Can't block if in interrupt or responding to an incoming
864 PORTAL_ALLOC_ATOMIC(ltx, desc_size);
866 PORTAL_ALLOC(ltx, desc_size);
870 CERROR("Can't allocate tx desc type %d size %d %s\n",
871 type, desc_size, in_interrupt() ? "(intr)" : "");
872 return (PTL_NO_SPACE);
875 atomic_inc(&ksocknal_data.ksnd_nactive_ltxs);
877 ltx->ltx_desc_size = desc_size;
879 /* We always have 1 mapped frag for the header */
880 ltx->ltx_tx.tx_iov = ltx->ltx_iov;
881 ltx->ltx_iov[0].iov_base = <x->ltx_hdr;
882 ltx->ltx_iov[0].iov_len = sizeof(*hdr);
885 ltx->ltx_private = private;
886 ltx->ltx_cookie = cookie;
888 ltx->ltx_tx.tx_isfwd = 0;
889 ltx->ltx_tx.tx_nob = sizeof (*hdr) + payload_nob;
891 if (payload_iov != NULL) {
892 /* payload is all mapped */
893 ltx->ltx_tx.tx_kiov = NULL;
894 ltx->ltx_tx.tx_nkiov = 0;
896 ltx->ltx_tx.tx_niov =
897 1 + lib_extract_iov(payload_niov, <x->ltx_iov[1],
898 payload_niov, payload_iov,
899 payload_offset, payload_nob);
901 /* payload is all pages */
902 ltx->ltx_tx.tx_niov = 1;
904 ltx->ltx_tx.tx_kiov = ltx->ltx_kiov;
905 ltx->ltx_tx.tx_nkiov =
906 lib_extract_kiov(payload_niov, ltx->ltx_kiov,
907 payload_niov, payload_kiov,
908 payload_offset, payload_nob);
911 rc = ksocknal_launch_packet(<x->ltx_tx, nid);
915 ksocknal_free_ltx(ltx);
920 ksocknal_send (lib_nal_t *nal, void *private, lib_msg_t *cookie,
921 ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid,
922 unsigned int payload_niov, struct iovec *payload_iov,
923 size_t payload_offset, size_t payload_len)
925 return (ksocknal_sendmsg(nal, private, cookie,
927 payload_niov, payload_iov, NULL,
928 payload_offset, payload_len));
932 ksocknal_send_pages (lib_nal_t *nal, void *private, lib_msg_t *cookie,
933 ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid,
934 unsigned int payload_niov, ptl_kiov_t *payload_kiov,
935 size_t payload_offset, size_t payload_len)
937 return (ksocknal_sendmsg(nal, private, cookie,
939 payload_niov, NULL, payload_kiov,
940 payload_offset, payload_len));
944 ksocknal_fwd_packet (void *arg, kpr_fwd_desc_t *fwd)
946 ptl_nid_t nid = fwd->kprfd_gateway_nid;
947 ksock_ftx_t *ftx = (ksock_ftx_t *)&fwd->kprfd_scratch;
950 CDEBUG (D_NET, "Forwarding [%p] -> "LPX64" ("LPX64"))\n", fwd,
951 fwd->kprfd_gateway_nid, fwd->kprfd_target_nid);
953 /* I'm the gateway; must be the last hop */
954 if (nid == ksocknal_lib.libnal_ni.ni_pid.nid)
955 nid = fwd->kprfd_target_nid;
957 /* setup iov for hdr */
958 ftx->ftx_iov.iov_base = fwd->kprfd_hdr;
959 ftx->ftx_iov.iov_len = sizeof(ptl_hdr_t);
961 ftx->ftx_tx.tx_isfwd = 1; /* This is a forwarding packet */
962 ftx->ftx_tx.tx_nob = sizeof(ptl_hdr_t) + fwd->kprfd_nob;
963 ftx->ftx_tx.tx_niov = 1;
964 ftx->ftx_tx.tx_iov = &ftx->ftx_iov;
965 ftx->ftx_tx.tx_nkiov = fwd->kprfd_niov;
966 ftx->ftx_tx.tx_kiov = fwd->kprfd_kiov;
968 rc = ksocknal_launch_packet (&ftx->ftx_tx, nid);
970 kpr_fwd_done (&ksocknal_data.ksnd_router, fwd, rc);
974 ksocknal_thread_start (int (*fn)(void *arg), void *arg)
976 long pid = cfs_kernel_thread (fn, arg, 0);
982 write_lock_irqsave(&ksocknal_data.ksnd_global_lock, flags);
983 ksocknal_data.ksnd_nthreads++;
984 write_unlock_irqrestore(&ksocknal_data.ksnd_global_lock, flags);
989 ksocknal_thread_fini (void)
993 write_lock_irqsave(&ksocknal_data.ksnd_global_lock, flags);
994 ksocknal_data.ksnd_nthreads--;
995 write_unlock_irqrestore(&ksocknal_data.ksnd_global_lock, flags);
999 ksocknal_fmb_callback (void *arg, int error)
1001 ksock_fmb_t *fmb = (ksock_fmb_t *)arg;
1002 ksock_fmb_pool_t *fmp = fmb->fmb_pool;
1003 ptl_hdr_t *hdr = &fmb->fmb_hdr;
1004 ksock_conn_t *conn = NULL;
1005 ksock_sched_t *sched;
1006 unsigned long flags;
1007 char ipbuf[PTL_NALFMT_SIZE];
1008 char ipbuf2[PTL_NALFMT_SIZE];
1011 CERROR("Failed to route packet from "
1012 LPX64" %s to "LPX64" %s: %d\n",
1013 le64_to_cpu(hdr->src_nid),
1014 portals_nid2str(SOCKNAL, le64_to_cpu(hdr->src_nid), ipbuf),
1015 le64_to_cpu(hdr->dest_nid),
1016 portals_nid2str(SOCKNAL, le64_to_cpu(hdr->dest_nid), ipbuf2),
1019 CDEBUG (D_NET, "routed packet from "LPX64" to "LPX64": OK\n",
1020 le64_to_cpu(hdr->src_nid), le64_to_cpu(hdr->dest_nid));
1022 /* drop peer ref taken on init */
1023 ksocknal_put_peer (fmb->fmb_peer);
1025 spin_lock_irqsave (&fmp->fmp_lock, flags);
1027 list_add (&fmb->fmb_list, &fmp->fmp_idle_fmbs);
1028 fmp->fmp_nactive_fmbs--;
1030 if (!list_empty (&fmp->fmp_blocked_conns)) {
1031 conn = list_entry (fmb->fmb_pool->fmp_blocked_conns.next,
1032 ksock_conn_t, ksnc_rx_list);
1033 list_del (&conn->ksnc_rx_list);
1036 spin_unlock_irqrestore (&fmp->fmp_lock, flags);
1041 CDEBUG (D_NET, "Scheduling conn %p\n", conn);
1042 LASSERT (conn->ksnc_rx_scheduled);
1043 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_FMB_SLEEP);
1045 conn->ksnc_rx_state = SOCKNAL_RX_GET_FMB;
1047 sched = conn->ksnc_scheduler;
1049 spin_lock_irqsave (&sched->kss_lock, flags);
1051 list_add_tail (&conn->ksnc_rx_list, &sched->kss_rx_conns);
1052 cfs_waitq_signal (&sched->kss_waitq);
1054 spin_unlock_irqrestore (&sched->kss_lock, flags);
1058 ksocknal_get_idle_fmb (ksock_conn_t *conn)
1060 int payload_nob = conn->ksnc_rx_nob_left;
1061 unsigned long flags;
1062 ksock_fmb_pool_t *pool;
1065 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB);
1066 LASSERT (kpr_routing(&ksocknal_data.ksnd_router));
1068 if (payload_nob <= SOCKNAL_SMALL_FWD_PAGES * CFS_PAGE_SIZE)
1069 pool = &ksocknal_data.ksnd_small_fmp;
1071 pool = &ksocknal_data.ksnd_large_fmp;
1073 spin_lock_irqsave (&pool->fmp_lock, flags);
1075 if (!list_empty (&pool->fmp_idle_fmbs)) {
1076 fmb = list_entry(pool->fmp_idle_fmbs.next,
1077 ksock_fmb_t, fmb_list);
1078 list_del (&fmb->fmb_list);
1079 pool->fmp_nactive_fmbs++;
1080 spin_unlock_irqrestore (&pool->fmp_lock, flags);
1085 /* deschedule until fmb free */
1087 conn->ksnc_rx_state = SOCKNAL_RX_FMB_SLEEP;
1089 list_add_tail (&conn->ksnc_rx_list,
1090 &pool->fmp_blocked_conns);
1092 spin_unlock_irqrestore (&pool->fmp_lock, flags);
1097 ksocknal_init_fmb (ksock_conn_t *conn, ksock_fmb_t *fmb)
1099 int payload_nob = conn->ksnc_rx_nob_left;
1100 ptl_nid_t dest_nid = le64_to_cpu(conn->ksnc_hdr.dest_nid);
1102 int nob = payload_nob;
1104 LASSERT (conn->ksnc_rx_scheduled);
1105 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB);
1106 LASSERT (conn->ksnc_rx_nob_wanted == conn->ksnc_rx_nob_left);
1107 LASSERT (payload_nob >= 0);
1108 LASSERT (payload_nob <= fmb->fmb_pool->fmp_buff_pages * CFS_PAGE_SIZE);
1109 LASSERT (sizeof (ptl_hdr_t) < CFS_PAGE_SIZE);
1110 LASSERT (fmb->fmb_kiov[0].kiov_offset == 0);
1112 /* Take a ref on the conn's peer to prevent module unload before
1113 * forwarding completes. */
1114 fmb->fmb_peer = conn->ksnc_peer;
1115 atomic_inc (&conn->ksnc_peer->ksnp_refcount);
1117 /* Copy the header we just read into the forwarding buffer. If
1118 * there's payload, start reading reading it into the buffer,
1119 * otherwise the forwarding buffer can be kicked off
1121 fmb->fmb_hdr = conn->ksnc_hdr;
1124 LASSERT (niov < fmb->fmb_pool->fmp_buff_pages);
1125 LASSERT (fmb->fmb_kiov[niov].kiov_offset == 0);
1126 fmb->fmb_kiov[niov].kiov_len = MIN (CFS_PAGE_SIZE, nob);
1127 nob -= CFS_PAGE_SIZE;
1131 kpr_fwd_init(&fmb->fmb_fwd, dest_nid, &fmb->fmb_hdr,
1132 payload_nob, niov, fmb->fmb_kiov,
1133 ksocknal_fmb_callback, fmb);
1135 if (payload_nob == 0) { /* got complete packet already */
1136 CDEBUG (D_NET, "%p "LPX64"->"LPX64" fwd_start (immediate)\n",
1137 conn, le64_to_cpu(conn->ksnc_hdr.src_nid), dest_nid);
1139 kpr_fwd_start (&ksocknal_data.ksnd_router, &fmb->fmb_fwd);
1141 ksocknal_new_packet (conn, 0); /* on to next packet */
1145 conn->ksnc_cookie = fmb; /* stash fmb for later */
1146 conn->ksnc_rx_state = SOCKNAL_RX_BODY_FWD; /* read in the payload */
1148 /* Set up conn->ksnc_rx_kiov to read the payload into fmb's kiov-ed
1150 LASSERT (niov <= sizeof(conn->ksnc_rx_iov_space)/sizeof(ptl_kiov_t));
1152 conn->ksnc_rx_niov = 0;
1153 conn->ksnc_rx_nkiov = niov;
1154 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1155 memcpy(conn->ksnc_rx_kiov, fmb->fmb_kiov, niov * sizeof(ptl_kiov_t));
1157 CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d reading body\n", conn,
1158 le64_to_cpu(conn->ksnc_hdr.src_nid), dest_nid, payload_nob);
1163 ksocknal_fwd_parse (ksock_conn_t *conn)
1166 ptl_nid_t dest_nid = le64_to_cpu(conn->ksnc_hdr.dest_nid);
1167 ptl_nid_t src_nid = le64_to_cpu(conn->ksnc_hdr.src_nid);
1168 int body_len = le32_to_cpu(conn->ksnc_hdr.payload_length);
1169 char str[PTL_NALFMT_SIZE];
1170 char str2[PTL_NALFMT_SIZE];
1172 CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d parsing header\n", conn,
1173 src_nid, dest_nid, conn->ksnc_rx_nob_left);
1175 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_HEADER);
1176 LASSERT (conn->ksnc_rx_scheduled);
1178 if (body_len < 0) { /* length corrupt (overflow) */
1179 CERROR("dropping packet from "LPX64" (%s) for "LPX64" (%s): "
1180 "packet size %d illegal\n",
1181 src_nid, portals_nid2str(TCPNAL, src_nid, str),
1182 dest_nid, portals_nid2str(TCPNAL, dest_nid, str2),
1185 ksocknal_new_packet (conn, 0); /* on to new packet */
1189 if (!kpr_routing(&ksocknal_data.ksnd_router)) { /* not forwarding */
1190 CERROR("dropping packet from "LPX64" (%s) for "LPX64
1191 " (%s): not forwarding\n",
1192 src_nid, portals_nid2str(TCPNAL, src_nid, str),
1193 dest_nid, portals_nid2str(TCPNAL, dest_nid, str2));
1194 /* on to new packet (skip this one's body) */
1195 ksocknal_new_packet (conn, body_len);
1199 if (body_len > PTL_MTU) { /* too big to forward */
1200 CERROR ("dropping packet from "LPX64" (%s) for "LPX64
1201 "(%s): packet size %d too big\n",
1202 src_nid, portals_nid2str(TCPNAL, src_nid, str),
1203 dest_nid, portals_nid2str(TCPNAL, dest_nid, str2),
1205 /* on to new packet (skip this one's body) */
1206 ksocknal_new_packet (conn, body_len);
1210 /* should have gone direct */
1211 peer = ksocknal_get_peer (conn->ksnc_hdr.dest_nid);
1213 CERROR ("dropping packet from "LPX64" (%s) for "LPX64
1214 "(%s): target is a peer\n",
1215 src_nid, portals_nid2str(TCPNAL, src_nid, str),
1216 dest_nid, portals_nid2str(TCPNAL, dest_nid, str2));
1217 ksocknal_put_peer (peer); /* drop ref from get above */
1219 /* on to next packet (skip this one's body) */
1220 ksocknal_new_packet (conn, body_len);
1224 conn->ksnc_rx_state = SOCKNAL_RX_GET_FMB; /* Getting FMB now */
1225 conn->ksnc_rx_nob_left = body_len; /* stash packet size */
1226 conn->ksnc_rx_nob_wanted = body_len; /* (no slop) */
1230 ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip)
1232 static char ksocknal_slop_buffer[4096];
1238 if (nob_to_skip == 0) { /* right at next packet boundary now */
1239 conn->ksnc_rx_started = 0;
1240 mb (); /* racing with timeout thread */
1242 conn->ksnc_rx_state = SOCKNAL_RX_HEADER;
1243 conn->ksnc_rx_nob_wanted = sizeof (ptl_hdr_t);
1244 conn->ksnc_rx_nob_left = sizeof (ptl_hdr_t);
1246 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1247 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_hdr;
1248 conn->ksnc_rx_iov[0].iov_len = sizeof (ptl_hdr_t);
1249 conn->ksnc_rx_niov = 1;
1251 conn->ksnc_rx_kiov = NULL;
1252 conn->ksnc_rx_nkiov = 0;
1256 /* Set up to skip as much a possible now. If there's more left
1257 * (ran out of iov entries) we'll get called again */
1259 conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1260 conn->ksnc_rx_nob_left = nob_to_skip;
1261 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1266 nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer));
1268 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1269 conn->ksnc_rx_iov[niov].iov_len = nob;
1274 } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */
1275 niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec));
1277 conn->ksnc_rx_niov = niov;
1278 conn->ksnc_rx_kiov = NULL;
1279 conn->ksnc_rx_nkiov = 0;
1280 conn->ksnc_rx_nob_wanted = skipped;
1285 ksocknal_process_receive (ksock_conn_t *conn)
1290 LASSERT (atomic_read (&conn->ksnc_refcount) > 0);
1292 /* doesn't need a forwarding buffer */
1293 if (conn->ksnc_rx_state != SOCKNAL_RX_GET_FMB)
1297 fmb = ksocknal_get_idle_fmb (conn);
1299 /* conn descheduled waiting for idle fmb */
1303 if (ksocknal_init_fmb (conn, fmb)) {
1304 /* packet forwarded */
1309 /* NB: sched lock NOT held */
1310 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_HEADER ||
1311 conn->ksnc_rx_state == SOCKNAL_RX_BODY ||
1312 conn->ksnc_rx_state == SOCKNAL_RX_BODY_FWD ||
1313 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1315 LASSERT (conn->ksnc_rx_nob_wanted > 0);
1317 rc = ksocknal_receive(conn);
1320 LASSERT (rc != -EAGAIN);
1323 CWARN ("[%p] EOF from "LPX64" ip %d.%d.%d.%d:%d\n",
1324 conn, conn->ksnc_peer->ksnp_nid,
1325 HIPQUAD(conn->ksnc_ipaddr),
1327 else if (!conn->ksnc_closing)
1328 CERROR ("[%p] Error %d on read from "LPX64
1329 " ip %d.%d.%d.%d:%d\n",
1330 conn, rc, conn->ksnc_peer->ksnp_nid,
1331 HIPQUAD(conn->ksnc_ipaddr),
1334 ksocknal_close_conn_and_siblings (conn, rc);
1335 return (rc == 0 ? -ESHUTDOWN : rc);
1338 if (conn->ksnc_rx_nob_wanted != 0) {
1343 switch (conn->ksnc_rx_state) {
1344 case SOCKNAL_RX_HEADER:
1345 if (conn->ksnc_hdr.type != cpu_to_le32(PTL_MSG_HELLO) &&
1346 le64_to_cpu(conn->ksnc_hdr.dest_nid) !=
1347 ksocknal_lib.libnal_ni.ni_pid.nid) {
1348 /* This packet isn't for me */
1349 ksocknal_fwd_parse (conn);
1350 switch (conn->ksnc_rx_state) {
1351 case SOCKNAL_RX_HEADER: /* skipped (zero payload) */
1352 return (0); /* => come back later */
1353 case SOCKNAL_RX_SLOP: /* skipping packet's body */
1354 goto try_read; /* => go read it */
1355 case SOCKNAL_RX_GET_FMB: /* forwarding */
1356 goto get_fmb; /* => go get a fwd msg buffer */
1363 /* sets wanted_len, iovs etc */
1364 rc = lib_parse(&ksocknal_lib, &conn->ksnc_hdr, conn);
1367 /* I just received garbage: give up on this conn */
1368 ksocknal_close_conn_and_siblings (conn, rc);
1372 if (conn->ksnc_rx_nob_wanted != 0) { /* need to get payload? */
1373 conn->ksnc_rx_state = SOCKNAL_RX_BODY;
1374 goto try_read; /* go read the payload */
1376 /* Fall through (completed packet for me) */
1378 case SOCKNAL_RX_BODY:
1379 /* payload all received */
1380 lib_finalize(&ksocknal_lib, NULL, conn->ksnc_cookie, PTL_OK);
1383 case SOCKNAL_RX_SLOP:
1384 /* starting new packet? */
1385 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1386 return (0); /* come back later */
1387 goto try_read; /* try to finish reading slop now */
1389 case SOCKNAL_RX_BODY_FWD:
1390 /* payload all received */
1391 CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d fwd_start (got body)\n",
1392 conn, le64_to_cpu(conn->ksnc_hdr.src_nid),
1393 le64_to_cpu(conn->ksnc_hdr.dest_nid),
1394 conn->ksnc_rx_nob_left);
1396 /* forward the packet. NB ksocknal_init_fmb() put fmb into
1397 * conn->ksnc_cookie */
1398 fmb = (ksock_fmb_t *)conn->ksnc_cookie;
1399 kpr_fwd_start (&ksocknal_data.ksnd_router, &fmb->fmb_fwd);
1401 /* no slop in forwarded packets */
1402 LASSERT (conn->ksnc_rx_nob_left == 0);
1404 ksocknal_new_packet (conn, 0); /* on to next packet */
1405 return (0); /* (later) */
1413 return (-EINVAL); /* keep gcc happy */
1417 ksocknal_recv (lib_nal_t *nal, void *private, lib_msg_t *msg,
1418 unsigned int niov, struct iovec *iov,
1419 size_t offset, size_t mlen, size_t rlen)
1421 ksock_conn_t *conn = (ksock_conn_t *)private;
1423 LASSERT (mlen <= rlen);
1424 LASSERT (niov <= PTL_MD_MAX_IOV);
1426 conn->ksnc_cookie = msg;
1427 conn->ksnc_rx_nob_wanted = mlen;
1428 conn->ksnc_rx_nob_left = rlen;
1430 conn->ksnc_rx_nkiov = 0;
1431 conn->ksnc_rx_kiov = NULL;
1432 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1433 conn->ksnc_rx_niov =
1434 lib_extract_iov(PTL_MD_MAX_IOV, conn->ksnc_rx_iov,
1435 niov, iov, offset, mlen);
1438 lib_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1439 lib_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1445 ksocknal_recv_pages (lib_nal_t *nal, void *private, lib_msg_t *msg,
1446 unsigned int niov, ptl_kiov_t *kiov,
1447 size_t offset, size_t mlen, size_t rlen)
1449 ksock_conn_t *conn = (ksock_conn_t *)private;
1451 LASSERT (mlen <= rlen);
1452 LASSERT (niov <= PTL_MD_MAX_IOV);
1454 conn->ksnc_cookie = msg;
1455 conn->ksnc_rx_nob_wanted = mlen;
1456 conn->ksnc_rx_nob_left = rlen;
1458 conn->ksnc_rx_niov = 0;
1459 conn->ksnc_rx_iov = NULL;
1460 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1461 conn->ksnc_rx_nkiov =
1462 lib_extract_kiov(PTL_MD_MAX_IOV, conn->ksnc_rx_kiov,
1463 niov, kiov, offset, mlen);
1466 lib_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1467 lib_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1473 ksocknal_sched_cansleep(ksock_sched_t *sched)
1475 unsigned long flags;
1478 spin_lock_irqsave(&sched->kss_lock, flags);
1480 rc = (!ksocknal_data.ksnd_shuttingdown &&
1482 list_empty(&sched->kss_zctxdone_list) &&
1484 list_empty(&sched->kss_rx_conns) &&
1485 list_empty(&sched->kss_tx_conns));
1487 spin_unlock_irqrestore(&sched->kss_lock, flags);
1491 int ksocknal_scheduler (void *arg)
1493 ksock_sched_t *sched = (ksock_sched_t *)arg;
1496 unsigned long flags;
1499 int id = sched - ksocknal_data.ksnd_schedulers;
1502 snprintf (name, sizeof (name),"ksocknald_%02d", id);
1503 kportal_daemonize (name);
1504 kportal_blockallsigs ();
1506 #if (CONFIG_SMP && CPU_AFFINITY)
1507 id = ksocknal_sched2cpu(id);
1508 if (cpu_online(id)) {
1511 set_cpus_allowed(current, m);
1513 CERROR ("Can't set CPU affinity for %s to %d\n", name, id);
1515 #endif /* CONFIG_SMP && CPU_AFFINITY */
1517 spin_lock_irqsave (&sched->kss_lock, flags);
1519 while (!ksocknal_data.ksnd_shuttingdown) {
1520 int did_something = 0;
1522 /* Ensure I progress everything semi-fairly */
1524 if (!list_empty (&sched->kss_rx_conns)) {
1525 conn = list_entry(sched->kss_rx_conns.next,
1526 ksock_conn_t, ksnc_rx_list);
1527 list_del(&conn->ksnc_rx_list);
1529 LASSERT(conn->ksnc_rx_scheduled);
1530 LASSERT(conn->ksnc_rx_ready);
1532 /* clear rx_ready in case receive isn't complete.
1533 * Do it BEFORE we call process_recv, since
1534 * data_ready can set it any time after we release
1536 conn->ksnc_rx_ready = 0;
1537 spin_unlock_irqrestore(&sched->kss_lock, flags);
1539 rc = ksocknal_process_receive(conn);
1541 spin_lock_irqsave(&sched->kss_lock, flags);
1543 /* I'm the only one that can clear this flag */
1544 LASSERT(conn->ksnc_rx_scheduled);
1546 /* Did process_receive get everything it wanted? */
1548 conn->ksnc_rx_ready = 1;
1550 if (conn->ksnc_rx_state == SOCKNAL_RX_FMB_SLEEP ||
1551 conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB) {
1552 /* Conn blocked for a forwarding buffer.
1553 * It will get queued for my attention when
1554 * one becomes available (and it might just
1555 * already have been!). Meanwhile my ref
1556 * on it stays put. */
1557 } else if (conn->ksnc_rx_ready) {
1558 /* reschedule for rx */
1559 list_add_tail (&conn->ksnc_rx_list,
1560 &sched->kss_rx_conns);
1562 conn->ksnc_rx_scheduled = 0;
1564 ksocknal_put_conn(conn);
1570 if (!list_empty (&sched->kss_tx_conns)) {
1571 conn = list_entry(sched->kss_tx_conns.next,
1572 ksock_conn_t, ksnc_tx_list);
1573 list_del (&conn->ksnc_tx_list);
1575 LASSERT(conn->ksnc_tx_scheduled);
1576 LASSERT(conn->ksnc_tx_ready);
1577 LASSERT(!list_empty(&conn->ksnc_tx_queue));
1579 tx = list_entry(conn->ksnc_tx_queue.next,
1580 ksock_tx_t, tx_list);
1581 /* dequeue now so empty list => more to send */
1582 list_del(&tx->tx_list);
1584 /* Clear tx_ready in case send isn't complete. Do
1585 * it BEFORE we call process_transmit, since
1586 * write_space can set it any time after we release
1588 conn->ksnc_tx_ready = 0;
1589 spin_unlock_irqrestore (&sched->kss_lock, flags);
1591 rc = ksocknal_process_transmit(conn, tx);
1593 spin_lock_irqsave (&sched->kss_lock, flags);
1595 if (rc == -ENOMEM || rc == -EAGAIN) {
1596 /* Incomplete send: replace tx on HEAD of tx_queue */
1597 list_add (&tx->tx_list, &conn->ksnc_tx_queue);
1599 /* Complete send; assume space for more */
1600 conn->ksnc_tx_ready = 1;
1603 if (rc == -ENOMEM) {
1604 /* Do nothing; after a short timeout, this
1605 * conn will be reposted on kss_tx_conns. */
1606 } else if (conn->ksnc_tx_ready &&
1607 !list_empty (&conn->ksnc_tx_queue)) {
1608 /* reschedule for tx */
1609 list_add_tail (&conn->ksnc_tx_list,
1610 &sched->kss_tx_conns);
1612 conn->ksnc_tx_scheduled = 0;
1614 ksocknal_put_conn (conn);
1620 if (!list_empty (&sched->kss_zctxdone_list)) {
1622 list_entry(sched->kss_zctxdone_list.next,
1623 ksock_tx_t, tx_list);
1626 list_del (&tx->tx_list);
1627 spin_unlock_irqrestore (&sched->kss_lock, flags);
1629 ksocknal_tx_done (tx, 1);
1631 spin_lock_irqsave (&sched->kss_lock, flags);
1634 if (!did_something || /* nothing to do */
1635 ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1636 spin_unlock_irqrestore (&sched->kss_lock, flags);
1640 if (!did_something) { /* wait for something to do */
1641 rc = wait_event_interruptible (sched->kss_waitq,
1642 !ksocknal_sched_cansleep(sched));
1647 spin_lock_irqsave (&sched->kss_lock, flags);
1651 spin_unlock_irqrestore (&sched->kss_lock, flags);
1652 ksocknal_thread_fini ();
1657 * Add connection to kss_rx_conns of scheduler
1658 * and wakeup the scheduler.
1660 void ksocknal_read_callback (ksock_conn_t *conn)
1662 ksock_sched_t *sched;
1663 unsigned long flags;
1666 sched = conn->ksnc_scheduler;
1668 spin_lock_irqsave (&sched->kss_lock, flags);
1670 conn->ksnc_rx_ready = 1;
1672 if (!conn->ksnc_rx_scheduled) { /* not being progressed */
1673 list_add_tail(&conn->ksnc_rx_list,
1674 &sched->kss_rx_conns);
1675 conn->ksnc_rx_scheduled = 1;
1676 /* extra ref for scheduler */
1677 atomic_inc (&conn->ksnc_refcount);
1679 cfs_waitq_signal (&sched->kss_waitq);
1681 spin_unlock_irqrestore (&sched->kss_lock, flags);
1687 * Add connection to kss_tx_conns of scheduler
1688 * and wakeup the scheduler.
1690 void ksocknal_write_callback (ksock_conn_t *conn)
1692 ksock_sched_t *sched;
1693 unsigned long flags;
1696 sched = conn->ksnc_scheduler;
1698 spin_lock_irqsave (&sched->kss_lock, flags);
1700 conn->ksnc_tx_ready = 1;
1702 if (!conn->ksnc_tx_scheduled && // not being progressed
1703 !list_empty(&conn->ksnc_tx_queue)){//packets to send
1704 list_add_tail (&conn->ksnc_tx_list,
1705 &sched->kss_tx_conns);
1706 conn->ksnc_tx_scheduled = 1;
1707 /* extra ref for scheduler */
1708 atomic_inc (&conn->ksnc_refcount);
1710 cfs_waitq_signal (&sched->kss_waitq);
1713 spin_unlock_irqrestore (&sched->kss_lock, flags);
1719 ksocknal_sock_write (struct socket *sock, void *buffer, int nob)
1721 return ksocknal_lib_sock_write(sock, buffer, nob);
1725 ksocknal_sock_read (struct socket *sock, void *buffer, int nob)
1727 return ksocknal_lib_sock_read(sock, buffer, nob);
1731 ksocknal_send_hello (ksock_conn_t *conn, __u32 *ipaddrs, int nipaddrs)
1733 /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1734 struct socket *sock = conn->ksnc_sock;
1736 ptl_magicversion_t *hmv = (ptl_magicversion_t *)&hdr.dest_nid;
1740 LASSERT (conn->ksnc_type != SOCKNAL_CONN_NONE);
1741 LASSERT (nipaddrs <= SOCKNAL_MAX_INTERFACES);
1743 /* No need for getconnsock/putconnsock */
1744 LASSERT (!conn->ksnc_closing);
1746 LASSERT (sizeof (*hmv) == sizeof (hdr.dest_nid));
1747 hmv->magic = cpu_to_le32 (PORTALS_PROTO_MAGIC);
1748 hmv->version_major = cpu_to_le16 (PORTALS_PROTO_VERSION_MAJOR);
1749 hmv->version_minor = cpu_to_le16 (PORTALS_PROTO_VERSION_MINOR);
1751 hdr.src_nid = cpu_to_le64 (ksocknal_lib.libnal_ni.ni_pid.nid);
1752 hdr.type = cpu_to_le32 (PTL_MSG_HELLO);
1753 hdr.payload_length = cpu_to_le32 (nipaddrs * sizeof(*ipaddrs));
1755 hdr.msg.hello.type = cpu_to_le32 (conn->ksnc_type);
1756 hdr.msg.hello.incarnation =
1757 cpu_to_le64 (ksocknal_data.ksnd_incarnation);
1759 /* Receiver is eager */
1760 rc = ksocknal_sock_write (sock, &hdr, sizeof(hdr));
1762 CERROR ("Error %d sending HELLO hdr to %u.%u.%u.%u/%d\n",
1763 rc, HIPQUAD(conn->ksnc_ipaddr), conn->ksnc_port);
1770 for (i = 0; i < nipaddrs; i++) {
1771 ipaddrs[i] = __cpu_to_le32 (ipaddrs[i]);
1774 rc = ksocknal_sock_write (sock, ipaddrs, nipaddrs * sizeof(*ipaddrs));
1776 CERROR ("Error %d sending HELLO payload (%d)"
1777 " to %u.%u.%u.%u/%d\n", rc, nipaddrs,
1778 HIPQUAD(conn->ksnc_ipaddr), conn->ksnc_port);
1783 ksocknal_invert_type(int type)
1787 case SOCKNAL_CONN_ANY:
1788 case SOCKNAL_CONN_CONTROL:
1790 case SOCKNAL_CONN_BULK_IN:
1791 return SOCKNAL_CONN_BULK_OUT;
1792 case SOCKNAL_CONN_BULK_OUT:
1793 return SOCKNAL_CONN_BULK_IN;
1795 return (SOCKNAL_CONN_NONE);
1800 ksocknal_recv_hello (ksock_conn_t *conn, ptl_nid_t *nid,
1801 __u64 *incarnation, __u32 *ipaddrs)
1803 struct socket *sock = conn->ksnc_sock;
1809 ptl_magicversion_t *hmv;
1810 char ipbuf[PTL_NALFMT_SIZE];
1812 hmv = (ptl_magicversion_t *)&hdr.dest_nid;
1813 LASSERT (sizeof (*hmv) == sizeof (hdr.dest_nid));
1815 rc = ksocknal_sock_read (sock, hmv, sizeof (*hmv));
1817 CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n",
1818 rc, HIPQUAD(conn->ksnc_ipaddr));
1822 if (hmv->magic != le32_to_cpu (PORTALS_PROTO_MAGIC)) {
1823 CERROR ("Bad magic %#08x (%#08x expected) from %u.%u.%u.%u\n",
1824 __cpu_to_le32 (hmv->magic), PORTALS_PROTO_MAGIC,
1825 HIPQUAD(conn->ksnc_ipaddr));
1829 if (hmv->version_major != cpu_to_le16 (PORTALS_PROTO_VERSION_MAJOR) ||
1830 hmv->version_minor != cpu_to_le16 (PORTALS_PROTO_VERSION_MINOR)) {
1831 CERROR ("Incompatible protocol version %d.%d (%d.%d expected)"
1832 " from %u.%u.%u.%u\n",
1833 le16_to_cpu (hmv->version_major),
1834 le16_to_cpu (hmv->version_minor),
1835 PORTALS_PROTO_VERSION_MAJOR,
1836 PORTALS_PROTO_VERSION_MINOR,
1837 HIPQUAD(conn->ksnc_ipaddr));
1841 #if (PORTALS_PROTO_VERSION_MAJOR != 1)
1842 # error "This code only understands protocol version 1.x"
1844 /* version 1 sends magic/version as the dest_nid of a 'hello'
1845 * header, followed by payload full of interface IP addresses.
1846 * Read the rest of it in now... */
1848 rc = ksocknal_sock_read (sock, hmv + 1, sizeof (hdr) - sizeof (*hmv));
1850 CERROR ("Error %d reading rest of HELLO hdr from %u.%u.%u.%u\n",
1851 rc, HIPQUAD(conn->ksnc_ipaddr));
1855 /* ...and check we got what we expected */
1856 if (hdr.type != cpu_to_le32 (PTL_MSG_HELLO)) {
1857 CERROR ("Expecting a HELLO hdr,"
1858 " but got type %d from %u.%u.%u.%u\n",
1859 le32_to_cpu (hdr.type),
1860 HIPQUAD(conn->ksnc_ipaddr));
1864 if (le64_to_cpu(hdr.src_nid) == PTL_NID_ANY) {
1865 CERROR("Expecting a HELLO hdr with a NID, but got PTL_NID_ANY"
1866 "from %u.%u.%u.%u\n", HIPQUAD(conn->ksnc_ipaddr));
1870 if (*nid == PTL_NID_ANY) { /* don't know peer's nid yet */
1871 *nid = le64_to_cpu(hdr.src_nid);
1872 } else if (*nid != le64_to_cpu (hdr.src_nid)) {
1873 LCONSOLE_ERROR("Connected successfully to nid "LPX64" on host "
1874 "%u.%u.%u.%u, but they claimed they were nid "
1875 LPX64" (%s); please check your Lustre "
1877 *nid, HIPQUAD(conn->ksnc_ipaddr),
1878 le64_to_cpu(hdr.src_nid),
1879 portals_nid2str(SOCKNAL,
1880 le64_to_cpu(hdr.src_nid),
1883 CERROR ("Connected to nid "LPX64"@%u.%u.%u.%u "
1884 "but expecting "LPX64"\n",
1885 le64_to_cpu (hdr.src_nid),
1886 HIPQUAD(conn->ksnc_ipaddr), *nid);
1890 type = __le32_to_cpu(hdr.msg.hello.type);
1892 if (conn->ksnc_type == SOCKNAL_CONN_NONE) {
1893 /* I've accepted this connection; peer determines type */
1894 conn->ksnc_type = ksocknal_invert_type(type);
1895 if (conn->ksnc_type == SOCKNAL_CONN_NONE) {
1896 CERROR ("Unexpected type %d from "LPX64"@%u.%u.%u.%u\n",
1897 type, *nid, HIPQUAD(conn->ksnc_ipaddr));
1900 } else if (ksocknal_invert_type(type) != conn->ksnc_type) {
1901 CERROR ("Mismatched types: me %d, "LPX64"@%u.%u.%u.%u %d\n",
1902 conn->ksnc_type, *nid, HIPQUAD(conn->ksnc_ipaddr),
1903 le32_to_cpu(hdr.msg.hello.type));
1907 *incarnation = le64_to_cpu(hdr.msg.hello.incarnation);
1909 nips = __le32_to_cpu (hdr.payload_length) / sizeof (__u32);
1911 if (nips > SOCKNAL_MAX_INTERFACES ||
1912 nips * sizeof(__u32) != __le32_to_cpu (hdr.payload_length)) {
1913 CERROR("Bad payload length %d from "LPX64"@%u.%u.%u.%u\n",
1914 __le32_to_cpu (hdr.payload_length),
1915 *nid, HIPQUAD(conn->ksnc_ipaddr));
1921 rc = ksocknal_sock_read (sock, ipaddrs, nips * sizeof(*ipaddrs));
1923 CERROR ("Error %d reading IPs from "LPX64"@%u.%u.%u.%u\n",
1924 rc, *nid, HIPQUAD(conn->ksnc_ipaddr));
1928 for (i = 0; i < nips; i++) {
1929 ipaddrs[i] = __le32_to_cpu(ipaddrs[i]);
1931 if (ipaddrs[i] == 0) {
1932 CERROR("Zero IP[%d] from "LPX64"@%u.%u.%u.%u\n",
1933 i, *nid, HIPQUAD(conn->ksnc_ipaddr));
1942 ksocknal_connect_peer (ksock_route_t *route, int type)
1944 struct socket *sock;
1949 /* Iterate through reserved ports. When typed connections are
1950 * used, we will need to bind to multiple ports, but we only know
1951 * this at connect time. But, by that time we've already called
1952 * bind() so we need a new socket. */
1954 for (port = 1023; port > 512; --port) {
1956 rc = ksocknal_lib_connect_sock(&sock, &may_retry, route, port);
1959 rc = ksocknal_create_conn(route, sock, type);
1960 cfs_put_file(KSN_SOCK2FILE(sock));
1968 CERROR("Out of ports trying to bind to a reserved port\n");
1969 return (-EADDRINUSE);
1973 ksocknal_autoconnect (ksock_route_t *route)
1975 CFS_LIST_HEAD (zombies);
1978 unsigned long flags;
1983 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
1986 if (!ksocknal_tunables.ksnd_typed_conns) {
1987 if ((route->ksnr_connected & (1<<SOCKNAL_CONN_ANY)) == 0)
1988 type = SOCKNAL_CONN_ANY;
1990 break; /* got connected while route queued */
1992 if ((route->ksnr_connected & (1<<SOCKNAL_CONN_CONTROL)) == 0)
1993 type = SOCKNAL_CONN_CONTROL;
1994 else if ((route->ksnr_connected & (1<<SOCKNAL_CONN_BULK_IN)) == 0)
1995 type = SOCKNAL_CONN_BULK_IN;
1996 else if ((route->ksnr_connected & (1<<SOCKNAL_CONN_BULK_OUT)) == 0)
1997 type = SOCKNAL_CONN_BULK_OUT;
1999 break; /* got connected while route queued */
2002 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
2004 rc = ksocknal_connect_peer (route, type);
2008 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
2011 LASSERT (route->ksnr_connecting);
2012 route->ksnr_connecting = 0;
2013 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
2018 /* "normal" errors */
2020 LCONSOLE_ERROR("Connection was refused by host %u.%u.%u.%u on "
2021 "port %d; check that Lustre is running on that "
2023 HIPQUAD(route->ksnr_ipaddr), route->ksnr_port);
2027 LCONSOLE_ERROR("Host %u.%u.%u.%u was unreachable; the network "
2028 "or that node may be down, or Lustre may be "
2030 HIPQUAD(route->ksnr_ipaddr));
2033 LCONSOLE_ERROR("Connecting to host %u.%u.%u.%u on port %d took "
2034 "too long; that node may be hung or "
2035 "experiencing high load.\n",
2036 HIPQUAD(route->ksnr_ipaddr), route->ksnr_port);
2038 /* errors that should be rare */
2040 LCONSOLE_ERROR("Protocol error connecting to host %u.%u.%u.%u "
2041 "on port %d: Is it running a compatible version"
2043 HIPQUAD(route->ksnr_ipaddr), route->ksnr_port);
2046 LCONSOLE_ERROR("No privileged ports available to connect to "
2047 "host %u.%u.%u.%u on port %d\n",
2048 HIPQUAD(route->ksnr_ipaddr), route->ksnr_port);
2051 LCONSOLE_ERROR("Unexpected error %d connecting to "
2052 "host %u.%u.%u.%u on port %d\n", rc,
2053 HIPQUAD(route->ksnr_ipaddr), route->ksnr_port);
2057 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
2059 peer = route->ksnr_peer;
2060 LASSERT (route->ksnr_connecting);
2061 route->ksnr_connecting = 0;
2063 /* This is a retry rather than a new connection */
2064 LASSERT (route->ksnr_retry_interval != 0);
2065 route->ksnr_timeout = cfs_time_add(cfs_time_current(),
2066 route->ksnr_retry_interval);
2067 route->ksnr_retry_interval = MIN (route->ksnr_retry_interval * 2,
2068 SOCKNAL_MAX_RECONNECT_INTERVAL);
2070 if (!list_empty (&peer->ksnp_tx_queue) &&
2071 ksocknal_find_connecting_route_locked (peer) == NULL) {
2072 LASSERT (list_empty (&peer->ksnp_conns));
2074 /* None of the connections that the blocked packets are
2075 * waiting for have been successful. Complete them now... */
2077 tx = list_entry (peer->ksnp_tx_queue.next,
2078 ksock_tx_t, tx_list);
2079 list_del (&tx->tx_list);
2080 list_add_tail (&tx->tx_list, &zombies);
2081 } while (!list_empty (&peer->ksnp_tx_queue));
2084 #if 0 /* irrelevent with only eager routes */
2085 if (!route->ksnr_deleted) {
2086 /* make this route least-favourite for re-selection */
2087 list_del(&route->ksnr_list);
2088 list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
2091 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
2093 while (!list_empty (&zombies)) {
2094 char ipbuf[PTL_NALFMT_SIZE];
2095 char ipbuf2[PTL_NALFMT_SIZE];
2096 tx = list_entry (zombies.next, ksock_tx_t, tx_list);
2098 CERROR ("Deleting packet type %d len %d ("LPX64" %s->"LPX64" %s)\n",
2099 le32_to_cpu (tx->tx_hdr->type),
2100 le32_to_cpu (tx->tx_hdr->payload_length),
2101 le64_to_cpu (tx->tx_hdr->src_nid),
2102 portals_nid2str(SOCKNAL,
2103 le64_to_cpu(tx->tx_hdr->src_nid),
2105 le64_to_cpu (tx->tx_hdr->dest_nid),
2106 portals_nid2str(SOCKNAL,
2107 le64_to_cpu(tx->tx_hdr->src_nid),
2110 list_del (&tx->tx_list);
2112 ksocknal_tx_done (tx, 0);
2117 ksocknal_autoconnectd (void *arg)
2119 long id = (long)arg;
2121 unsigned long flags;
2122 ksock_route_t *route;
2125 snprintf (name, sizeof (name), "ksocknal_ad%02ld", id);
2126 kportal_daemonize (name);
2127 kportal_blockallsigs ();
2129 spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2131 while (!ksocknal_data.ksnd_shuttingdown) {
2133 if (!list_empty (&ksocknal_data.ksnd_autoconnectd_routes)) {
2134 route = list_entry (ksocknal_data.ksnd_autoconnectd_routes.next,
2135 ksock_route_t, ksnr_connect_list);
2137 list_del (&route->ksnr_connect_list);
2138 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2140 ksocknal_autoconnect (route);
2141 ksocknal_put_route (route);
2143 spin_lock_irqsave(&ksocknal_data.ksnd_autoconnectd_lock,
2148 spin_unlock_irqrestore(&ksocknal_data.ksnd_autoconnectd_lock,
2151 rc = wait_event_interruptible(ksocknal_data.ksnd_autoconnectd_waitq,
2152 ksocknal_data.ksnd_shuttingdown ||
2153 !list_empty(&ksocknal_data.ksnd_autoconnectd_routes));
2155 spin_lock_irqsave(&ksocknal_data.ksnd_autoconnectd_lock, flags);
2158 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2160 ksocknal_thread_fini ();
2165 ksocknal_find_timed_out_conn (ksock_peer_t *peer)
2167 /* We're called with a shared lock on ksnd_global_lock */
2169 struct list_head *ctmp;
2171 list_for_each (ctmp, &peer->ksnp_conns) {
2172 conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
2174 /* Don't need the {get,put}connsock dance to deref ksnc_sock... */
2175 LASSERT (!conn->ksnc_closing);
2177 if (SOCK_ERROR(conn->ksnc_sock) != 0) {
2178 atomic_inc (&conn->ksnc_refcount);
2180 switch (SOCK_ERROR(conn->ksnc_sock)) {
2182 LCONSOLE_WARN("A connection with %u.%u.%u.%u "
2183 "was reset; they may have "
2185 HIPQUAD(conn->ksnc_ipaddr));
2188 LCONSOLE_WARN("A connection with %u.%u.%u.%u "
2189 "timed out; the network or that "
2190 "node may be down.\n",
2191 HIPQUAD(conn->ksnc_ipaddr));
2194 LCONSOLE_WARN("An unexpected network error "
2195 "occurred with %u.%u.%u.%u: %d.\n",
2196 HIPQUAD(conn->ksnc_ipaddr),
2197 SOCK_ERROR(conn->ksnc_sock));
2201 /* Something (e.g. failed keepalive) set the socket error */
2202 CERROR ("Socket error %d: "LPX64" %p %d.%d.%d.%d\n",
2203 SOCK_ERROR(conn->ksnc_sock), peer->ksnp_nid,
2204 conn, HIPQUAD(conn->ksnc_ipaddr));
2209 if (conn->ksnc_rx_started &&
2210 cfs_time_aftereq (cfs_time_current(),
2211 conn->ksnc_rx_deadline)) {
2212 /* Timed out incomplete incoming message */
2213 atomic_inc (&conn->ksnc_refcount);
2214 LCONSOLE_ERROR("A timeout occurred receiving data from "
2215 "%u.%u.%u.%u; the network or that node "
2217 HIPQUAD(conn->ksnc_ipaddr));
2218 CERROR ("Timed out RX from "LPX64" %p %d.%d.%d.%d\n",
2219 peer->ksnp_nid,conn,HIPQUAD(conn->ksnc_ipaddr));
2223 if ((!list_empty (&conn->ksnc_tx_queue) ||
2224 SOCK_WMEM_QUEUED(conn->ksnc_sock) != 0) &&
2225 cfs_time_aftereq (cfs_time_current(),
2226 conn->ksnc_tx_deadline)) {
2227 /* Timed out messages queued for sending or
2228 * buffered in the socket's send buffer */
2229 atomic_inc (&conn->ksnc_refcount);
2230 LCONSOLE_ERROR("A timeout occurred sending data to "
2231 "%u.%u.%u.%u; the network or that node "
2233 HIPQUAD(conn->ksnc_ipaddr));
2234 CERROR ("Timed out TX to "LPX64" %s%d %p %d.%d.%d.%d\n",
2236 list_empty (&conn->ksnc_tx_queue) ? "" : "Q ",
2237 SOCK_WMEM_QUEUED(conn->ksnc_sock), conn,
2238 HIPQUAD(conn->ksnc_ipaddr));
2247 ksocknal_check_peer_timeouts (int idx)
2249 struct list_head *peers = &ksocknal_data.ksnd_peers[idx];
2250 struct list_head *ptmp;
2255 /* NB. We expect to have a look at all the peers and not find any
2256 * connections to time out, so we just use a shared lock while we
2258 read_lock (&ksocknal_data.ksnd_global_lock);
2260 list_for_each (ptmp, peers) {
2261 peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
2262 conn = ksocknal_find_timed_out_conn (peer);
2265 read_unlock (&ksocknal_data.ksnd_global_lock);
2267 CERROR ("Timeout out conn->"LPX64" ip %d.%d.%d.%d:%d\n",
2269 HIPQUAD(conn->ksnc_ipaddr),
2271 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2273 /* NB we won't find this one again, but we can't
2274 * just proceed with the next peer, since we dropped
2275 * ksnd_global_lock and it might be dead already! */
2276 ksocknal_put_conn (conn);
2281 read_unlock (&ksocknal_data.ksnd_global_lock);
2285 ksocknal_reaper (void *arg)
2287 cfs_waitlink_t wait;
2288 unsigned long flags;
2290 ksock_sched_t *sched;
2291 struct list_head enomem_conns;
2293 cfs_duration_t timeout;
2296 cfs_time_t deadline = cfs_time_current();
2298 kportal_daemonize ("ksocknal_reaper");
2299 kportal_blockallsigs ();
2301 CFS_INIT_LIST_HEAD(&enomem_conns);
2302 cfs_waitlink_init (&wait);
2304 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2306 while (!ksocknal_data.ksnd_shuttingdown) {
2308 if (!list_empty (&ksocknal_data.ksnd_deathrow_conns)) {
2309 conn = list_entry (ksocknal_data.ksnd_deathrow_conns.next,
2310 ksock_conn_t, ksnc_list);
2311 list_del (&conn->ksnc_list);
2313 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2315 ksocknal_terminate_conn (conn);
2316 ksocknal_put_conn (conn);
2318 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2322 if (!list_empty (&ksocknal_data.ksnd_zombie_conns)) {
2323 conn = list_entry (ksocknal_data.ksnd_zombie_conns.next,
2324 ksock_conn_t, ksnc_list);
2325 list_del (&conn->ksnc_list);
2327 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2329 ksocknal_destroy_conn (conn);
2331 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2335 if (!list_empty (&ksocknal_data.ksnd_enomem_conns)) {
2336 list_add(&enomem_conns, &ksocknal_data.ksnd_enomem_conns);
2337 list_del_init(&ksocknal_data.ksnd_enomem_conns);
2340 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2342 /* reschedule all the connections that stalled with ENOMEM... */
2344 while (!list_empty (&enomem_conns)) {
2345 conn = list_entry (enomem_conns.next,
2346 ksock_conn_t, ksnc_tx_list);
2347 list_del (&conn->ksnc_tx_list);
2349 sched = conn->ksnc_scheduler;
2351 spin_lock_irqsave (&sched->kss_lock, flags);
2353 LASSERT (conn->ksnc_tx_scheduled);
2354 conn->ksnc_tx_ready = 1;
2355 list_add_tail (&conn->ksnc_tx_list, &sched->kss_tx_conns);
2356 cfs_waitq_signal (&sched->kss_waitq);
2358 spin_unlock_irqrestore (&sched->kss_lock, flags);
2362 /* careful with the jiffy wrap... */
2363 while ((timeout = cfs_time_sub(deadline,
2364 cfs_time_current())) <= 0) {
2367 int chunk = ksocknal_data.ksnd_peer_hash_size;
2369 /* Time to check for timeouts on a few more peers: I do
2370 * checks every 'p' seconds on a proportion of the peer
2371 * table and I need to check every connection 'n' times
2372 * within a timeout interval, to ensure I detect a
2373 * timeout on any connection within (n+1)/n times the
2374 * timeout interval. */
2376 if (ksocknal_tunables.ksnd_io_timeout > n * p)
2377 chunk = (chunk * n * p) /
2378 ksocknal_tunables.ksnd_io_timeout;
2382 for (i = 0; i < chunk; i++) {
2383 ksocknal_check_peer_timeouts (peer_index);
2384 peer_index = (peer_index + 1) %
2385 ksocknal_data.ksnd_peer_hash_size;
2388 deadline = cfs_time_add(deadline, cfs_time_seconds(p));
2391 if (nenomem_conns != 0) {
2392 /* Reduce my timeout if I rescheduled ENOMEM conns.
2393 * This also prevents me getting woken immediately
2394 * if any go back on my enomem list. */
2395 timeout = SOCKNAL_ENOMEM_RETRY;
2397 ksocknal_data.ksnd_reaper_waketime =
2398 cfs_time_add(cfs_time_current(), timeout);
2400 set_current_state (TASK_INTERRUPTIBLE);
2401 cfs_waitq_add (&ksocknal_data.ksnd_reaper_waitq, &wait);
2403 if (!ksocknal_data.ksnd_shuttingdown &&
2404 list_empty (&ksocknal_data.ksnd_deathrow_conns) &&
2405 list_empty (&ksocknal_data.ksnd_zombie_conns))
2406 cfs_waitq_timedwait (&wait, timeout);
2408 set_current_state (TASK_RUNNING);
2409 cfs_waitq_del (&ksocknal_data.ksnd_reaper_waitq, &wait);
2411 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2414 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2416 ksocknal_thread_fini ();
2420 lib_nal_t ksocknal_lib = {
2421 libnal_data: &ksocknal_data, /* NAL private data */
2422 libnal_send: ksocknal_send,
2423 libnal_send_pages: ksocknal_send_pages,
2424 libnal_recv: ksocknal_recv,
2425 libnal_recv_pages: ksocknal_recv_pages,
2426 libnal_dist: ksocknal_dist