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_read(nal_cb_t *nal, void *private, void *dst_addr,
34 user_ptr src_addr, size_t len)
36 CDEBUG(D_NET, LPX64": reading %ld bytes from %p -> %p\n",
37 nal->ni.nid, (long)len, src_addr, dst_addr);
39 memcpy( dst_addr, src_addr, len );
44 ksocknal_write(nal_cb_t *nal, void *private, user_ptr dst_addr,
45 void *src_addr, size_t len)
47 CDEBUG(D_NET, LPX64": writing %ld bytes from %p -> %p\n",
48 nal->ni.nid, (long)len, src_addr, dst_addr);
50 memcpy( dst_addr, src_addr, len );
55 ksocknal_callback (nal_cb_t * nal, void *private, lib_eq_t *eq,
58 CDEBUG(D_NET, LPX64": callback eq %p ev %p\n",
61 if (eq->event_callback != NULL)
62 eq->event_callback(ev);
68 ksocknal_malloc(nal_cb_t *nal, size_t len)
72 PORTAL_ALLOC(buf, len);
81 ksocknal_free(nal_cb_t *nal, void *buf, size_t len)
83 PORTAL_FREE(buf, len);
87 ksocknal_printf(nal_cb_t *nal, const char *fmt, ...)
93 vsnprintf (msg, sizeof (msg), fmt, ap); /* sprint safely */
96 msg[sizeof (msg) - 1] = 0; /* ensure terminated */
98 CDEBUG (D_NET, "%s", msg);
102 ksocknal_cli(nal_cb_t *nal, unsigned long *flags)
104 ksock_nal_data_t *data = nal->nal_data;
106 spin_lock(&data->ksnd_nal_cb_lock);
110 ksocknal_sti(nal_cb_t *nal, unsigned long *flags)
112 ksock_nal_data_t *data;
113 data = nal->nal_data;
115 spin_unlock(&data->ksnd_nal_cb_lock);
119 ksocknal_dist(nal_cb_t *nal, ptl_nid_t nid, unsigned long *dist)
121 /* I would guess that if ksocknal_get_peer (nid) == NULL,
122 and we're not routing, then 'nid' is very distant :) */
123 if ( nal->ni.nid == nid ) {
133 ksocknal_get_ltx (int may_block)
136 ksock_ltx_t *ltx = NULL;
139 spin_lock_irqsave (&ksocknal_data.ksnd_idle_ltx_lock, flags);
141 if (!list_empty (&ksocknal_data.ksnd_idle_ltx_list)) {
142 ltx = list_entry(ksocknal_data.ksnd_idle_ltx_list.next,
143 ksock_ltx_t, ltx_tx.tx_list);
144 list_del (<x->ltx_tx.tx_list);
145 ksocknal_data.ksnd_active_ltxs++;
150 if (!list_empty(&ksocknal_data.ksnd_idle_nblk_ltx_list)) {
151 ltx = list_entry(ksocknal_data.ksnd_idle_nblk_ltx_list.next,
152 ksock_ltx_t, ltx_tx.tx_list);
153 list_del (<x->ltx_tx.tx_list);
154 ksocknal_data.ksnd_active_ltxs++;
159 spin_unlock_irqrestore(&ksocknal_data.ksnd_idle_ltx_lock,
162 wait_event (ksocknal_data.ksnd_idle_ltx_waitq,
163 !list_empty (&ksocknal_data.ksnd_idle_ltx_list));
166 spin_unlock_irqrestore (&ksocknal_data.ksnd_idle_ltx_lock, flags);
172 ksocknal_put_ltx (ksock_ltx_t *ltx)
176 spin_lock_irqsave (&ksocknal_data.ksnd_idle_ltx_lock, flags);
178 ksocknal_data.ksnd_active_ltxs--;
179 list_add_tail (<x->ltx_tx.tx_list, ltx->ltx_idle);
181 /* normal tx desc => wakeup anyone blocking for one */
182 if (ltx->ltx_idle == &ksocknal_data.ksnd_idle_ltx_list)
183 wake_up (&ksocknal_data.ksnd_idle_ltx_waitq);
185 spin_unlock_irqrestore (&ksocknal_data.ksnd_idle_ltx_lock, flags);
190 ksocknal_kvaddr_to_page (unsigned long vaddr)
194 if (vaddr >= VMALLOC_START &&
196 page = vmalloc_to_page ((void *)vaddr);
198 else if (vaddr >= PKMAP_BASE &&
199 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE))
200 page = vmalloc_to_page ((void *)vaddr);
201 /* in 2.4 ^ just walks the page tables */
204 page = virt_to_page (vaddr);
215 ksocknal_send_iov (ksock_conn_t *conn, ksock_tx_t *tx)
217 struct socket *sock = conn->ksnc_sock;
218 struct iovec *iov = tx->tx_iov;
219 int fragsize = iov->iov_len;
220 unsigned long vaddr = (unsigned long)iov->iov_base;
221 int more = (!list_empty (&conn->ksnc_tx_queue)) |
225 int offset = vaddr & (PAGE_SIZE - 1);
226 int zcsize = MIN (fragsize, PAGE_SIZE - offset);
231 /* NB we can't trust socket ops to either consume our iovs
232 * or leave them alone, so we only send 1 frag at a time. */
233 LASSERT (fragsize <= tx->tx_resid);
234 LASSERT (tx->tx_niov > 0);
237 if (zcsize >= ksocknal_data.ksnd_zc_min_frag &&
238 (sock->sk->route_caps & NETIF_F_SG) &&
239 (sock->sk->route_caps & (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM | NETIF_F_HW_CSUM)) &&
240 (page = ksocknal_kvaddr_to_page (vaddr)) != NULL) {
242 CDEBUG(D_NET, "vaddr %p, page %p->%p + offset %x for %d\n",
243 (void *)vaddr, page, page_address(page), offset, zcsize);
245 if (fragsize > zcsize) {
250 rc = tcp_sendpage_zccd(sock, page, offset, zcsize,
251 more ? (MSG_DONTWAIT | MSG_MORE) : MSG_DONTWAIT,
256 /* NB don't pass tx's iov; sendmsg may or may not update it */
257 struct iovec fragiov = { .iov_base = (void *)vaddr,
258 .iov_len = fragsize};
259 struct msghdr msg = {
266 .msg_flags = more ? (MSG_DONTWAIT | MSG_MORE) : MSG_DONTWAIT
268 mm_segment_t oldmm = get_fs();
271 rc = sock_sendmsg(sock, &msg, fragsize);
280 if (rc < iov->iov_len) {
281 /* didn't send whole iov entry... */
282 iov->iov_base = (void *)(vaddr + rc);
284 /* ...but did we send everything we tried to send? */
285 return ((rc == fragsize) ? 1 : -EAGAIN);
294 ksocknal_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
296 struct socket *sock = conn->ksnc_sock;
297 ptl_kiov_t *kiov = tx->tx_kiov;
298 int fragsize = kiov->kiov_len;
299 struct page *page = kiov->kiov_page;
300 int offset = kiov->kiov_offset;
301 int more = (!list_empty (&conn->ksnc_tx_queue)) |
305 /* NB we can't trust socket ops to either consume our iovs
306 * or leave them alone, so we only send 1 frag at a time. */
307 LASSERT (fragsize <= tx->tx_resid);
308 LASSERT (offset + fragsize <= PAGE_SIZE);
309 LASSERT (tx->tx_niov == 0);
310 LASSERT (tx->tx_nkiov > 0);
313 if (fragsize >= ksocknal_data.ksnd_zc_min_frag &&
314 (sock->sk->route_caps & NETIF_F_SG) &&
315 (sock->sk->route_caps & (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM | NETIF_F_HW_CSUM))) {
317 CDEBUG(D_NET, "page %p + offset %x for %d\n",
318 page, offset, fragsize);
320 rc = tcp_sendpage_zccd(sock, page, offset, fragsize,
321 more ? (MSG_DONTWAIT | MSG_MORE) : MSG_DONTWAIT,
326 char *addr = ((char *)kmap (page)) + offset;
327 struct iovec fragiov = {.iov_base = addr,
328 .iov_len = fragsize};
329 struct msghdr msg = {
336 .msg_flags = more ? (MSG_DONTWAIT | MSG_MORE) : MSG_DONTWAIT
338 mm_segment_t oldmm = get_fs();
341 rc = sock_sendmsg(sock, &msg, fragsize);
353 /* didn't send whole frag */
354 kiov->kiov_offset = offset + rc;
355 kiov->kiov_len = fragsize - rc;
359 /* everything went */
360 LASSERT (rc == fragsize);
367 ksocknal_sendmsg (ksock_conn_t *conn, ksock_tx_t *tx)
369 /* Return 0 on success, < 0 on error.
370 * caller checks tx_resid to determine progress/completion */
374 if (ksocknal_data.ksnd_stall_tx != 0) {
375 set_current_state (TASK_UNINTERRUPTIBLE);
376 schedule_timeout (ksocknal_data.ksnd_stall_tx * HZ);
379 rc = ksocknal_getconnsock (conn);
384 LASSERT (tx->tx_resid != 0);
386 if (conn->ksnc_closing) {
391 if (tx->tx_niov != 0)
392 rc = ksocknal_send_iov (conn, tx);
394 rc = ksocknal_send_kiov (conn, tx);
396 if (rc <= 0) { /* error or socket full? */
397 /* NB: rc == 0 and rc == -EAGAIN both mean try
398 * again later (linux stack returns -EAGAIN for
399 * this, but Adaptech TOE returns 0) */
405 /* Consider the connection alive since we managed to chuck
406 * more data into it. Really, we'd like to consider it
407 * alive only when the peer ACKs something, but
408 * write_space() only gets called back while SOCK_NOSPACE
409 * is set. Instead, we presume peer death has occurred if
410 * the socket doesn't drain within a timout */
411 conn->ksnc_tx_deadline = jiffies +
412 ksocknal_data.ksnd_io_timeout * HZ;
413 conn->ksnc_peer->ksnp_last_alive = jiffies;
415 if (tx->tx_resid == 0) { /* sent everything */
421 ksocknal_putconnsock (conn);
426 ksocknal_eager_ack (ksock_conn_t *conn)
429 mm_segment_t oldmm = get_fs();
430 struct socket *sock = conn->ksnc_sock;
432 /* Remind the socket to ACK eagerly. If I don't, the socket might
433 * think I'm about to send something it could piggy-back the ACK
434 * on, introducing delay in completing zero-copy sends in my
438 sock->ops->setsockopt (sock, SOL_TCP, TCP_QUICKACK,
439 (char *)&opt, sizeof (opt));
444 ksocknal_recv_iov (ksock_conn_t *conn)
446 struct iovec *iov = conn->ksnc_rx_iov;
447 int fragsize = iov->iov_len;
448 unsigned long vaddr = (unsigned long)iov->iov_base;
449 struct iovec fragiov = { .iov_base = (void *)vaddr,
450 .iov_len = fragsize};
451 struct msghdr msg = {
460 mm_segment_t oldmm = get_fs();
463 /* NB we can't trust socket ops to either consume our iovs
464 * or leave them alone, so we only receive 1 frag at a time. */
465 LASSERT (conn->ksnc_rx_niov > 0);
466 LASSERT (fragsize <= conn->ksnc_rx_nob_wanted);
469 rc = sock_recvmsg (conn->ksnc_sock, &msg, fragsize, MSG_DONTWAIT);
470 /* NB this is just a boolean............................^ */
476 /* received something... */
477 conn->ksnc_peer->ksnp_last_alive = jiffies;
478 conn->ksnc_rx_deadline = jiffies +
479 ksocknal_data.ksnd_io_timeout * HZ;
480 mb(); /* order with setting rx_started */
481 conn->ksnc_rx_started = 1;
483 conn->ksnc_rx_nob_wanted -= rc;
484 conn->ksnc_rx_nob_left -= rc;
487 iov->iov_base = (void *)(vaddr + rc);
488 iov->iov_len = fragsize - rc;
493 conn->ksnc_rx_niov--;
498 ksocknal_recv_kiov (ksock_conn_t *conn)
500 ptl_kiov_t *kiov = conn->ksnc_rx_kiov;
501 struct page *page = kiov->kiov_page;
502 int offset = kiov->kiov_offset;
503 int fragsize = kiov->kiov_len;
504 unsigned long vaddr = ((unsigned long)kmap (page)) + offset;
505 struct iovec fragiov = { .iov_base = (void *)vaddr,
506 .iov_len = fragsize};
507 struct msghdr msg = {
516 mm_segment_t oldmm = get_fs();
519 /* NB we can't trust socket ops to either consume our iovs
520 * or leave them alone, so we only receive 1 frag at a time. */
521 LASSERT (fragsize <= conn->ksnc_rx_nob_wanted);
522 LASSERT (conn->ksnc_rx_nkiov > 0);
523 LASSERT (offset + fragsize <= PAGE_SIZE);
526 rc = sock_recvmsg (conn->ksnc_sock, &msg, fragsize, MSG_DONTWAIT);
527 /* NB this is just a boolean............................^ */
535 /* received something... */
536 conn->ksnc_peer->ksnp_last_alive = jiffies;
537 conn->ksnc_rx_deadline = jiffies +
538 ksocknal_data.ksnd_io_timeout * HZ;
539 mb(); /* order with setting rx_started */
540 conn->ksnc_rx_started = 1;
542 conn->ksnc_rx_nob_wanted -= rc;
543 conn->ksnc_rx_nob_left -= rc;
546 kiov->kiov_offset = offset + rc;
547 kiov->kiov_len = fragsize - rc;
551 conn->ksnc_rx_kiov++;
552 conn->ksnc_rx_nkiov--;
557 ksocknal_recvmsg (ksock_conn_t *conn)
559 /* Return 1 on success, 0 on EOF, < 0 on error.
560 * Caller checks ksnc_rx_nob_wanted to determine
561 * progress/completion. */
565 if (ksocknal_data.ksnd_stall_rx != 0) {
566 set_current_state (TASK_UNINTERRUPTIBLE);
567 schedule_timeout (ksocknal_data.ksnd_stall_rx * HZ);
570 rc = ksocknal_getconnsock (conn);
575 if (conn->ksnc_closing) {
580 if (conn->ksnc_rx_niov != 0)
581 rc = ksocknal_recv_iov (conn);
583 rc = ksocknal_recv_kiov (conn);
586 /* error/EOF or partial receive */
589 } else if (rc == 0 && conn->ksnc_rx_started) {
590 /* EOF in the middle of a message */
596 /* Completed a fragment */
598 if (conn->ksnc_rx_nob_wanted == 0) {
599 /* Completed a message segment (header or payload) */
600 if (ksocknal_data.ksnd_eager_ack &&
601 (conn->ksnc_rx_state == SOCKNAL_RX_BODY ||
602 conn->ksnc_rx_state == SOCKNAL_RX_BODY_FWD)) {
603 /* Remind the socket to ack eagerly... */
604 ksocknal_eager_ack(conn);
611 ksocknal_putconnsock (conn);
617 ksocknal_zc_callback (zccd_t *zcd)
619 ksock_tx_t *tx = KSOCK_ZCCD_2_TX(zcd);
620 ksock_sched_t *sched = tx->tx_conn->ksnc_scheduler;
624 /* Schedule tx for cleanup (can't do it now due to lock conflicts) */
626 spin_lock_irqsave (&sched->kss_lock, flags);
628 list_add_tail (&tx->tx_list, &sched->kss_zctxdone_list);
629 wake_up (&sched->kss_waitq);
631 spin_unlock_irqrestore (&sched->kss_lock, flags);
637 ksocknal_tx_done (ksock_tx_t *tx, int asynch)
642 if (tx->tx_conn != NULL) {
643 /* This tx got queued on a conn; do the accounting... */
644 atomic_sub (tx->tx_nob, &tx->tx_conn->ksnc_tx_nob);
646 /* zero copy completion isn't always from
647 * process_transmit() so it needs to keep a ref on
650 ksocknal_put_conn (tx->tx_conn);
656 if (tx->tx_isfwd) { /* was a forwarded packet? */
657 kpr_fwd_done (&ksocknal_data.ksnd_router,
658 KSOCK_TX_2_KPR_FWD_DESC (tx), 0);
664 ltx = KSOCK_TX_2_KSOCK_LTX (tx);
666 lib_finalize (&ksocknal_lib, ltx->ltx_private, ltx->ltx_cookie);
668 ksocknal_put_ltx (ltx);
673 ksocknal_tx_launched (ksock_tx_t *tx)
676 if (atomic_read (&tx->tx_zccd.zccd_count) != 1) {
677 ksock_conn_t *conn = tx->tx_conn;
679 /* zccd skbufs are still in-flight. First take a ref on
680 * conn, so it hangs about for ksocknal_tx_done... */
681 atomic_inc (&conn->ksnc_refcount);
683 /* ...then drop the initial ref on zccd, so the zero copy
684 * callback can occur */
685 zccd_put (&tx->tx_zccd);
689 /* Any zero-copy-ness (if any) has completed; I can complete the
690 * transmit now, avoiding an extra schedule */
691 ksocknal_tx_done (tx, 0);
695 ksocknal_process_transmit (ksock_sched_t *sched, unsigned long *irq_flags)
701 LASSERT (!list_empty (&sched->kss_tx_conns));
702 conn = list_entry(sched->kss_tx_conns.next, ksock_conn_t, ksnc_tx_list);
703 list_del (&conn->ksnc_tx_list);
705 LASSERT (conn->ksnc_tx_scheduled);
706 LASSERT (conn->ksnc_tx_ready);
707 LASSERT (!list_empty (&conn->ksnc_tx_queue));
708 tx = list_entry (conn->ksnc_tx_queue.next, ksock_tx_t, tx_list);
709 /* assume transmit will complete now, so dequeue while I've got lock */
710 list_del (&tx->tx_list);
712 spin_unlock_irqrestore (&sched->kss_lock, *irq_flags);
714 LASSERT (tx->tx_resid > 0);
716 conn->ksnc_tx_ready = 0;/* write_space may race with me and set ready */
717 mb(); /* => clear BEFORE trying to write */
719 rc = ksocknal_sendmsg (conn, tx);
721 CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc);
724 if (ksocknal_close_conn_unlocked (conn, rc)) {
725 /* I'm the first to close */
726 CERROR ("[%p] Error %d on write to "LPX64" ip %08x:%d\n",
727 conn, rc, conn->ksnc_peer->ksnp_nid,
728 conn->ksnc_ipaddr, conn->ksnc_port);
730 ksocknal_tx_launched (tx);
731 spin_lock_irqsave (&sched->kss_lock, *irq_flags);
733 } else if (tx->tx_resid == 0) {
734 /* everything went; assume more can go, and avoid
735 * write_space locking */
736 conn->ksnc_tx_ready = 1;
738 ksocknal_tx_launched (tx);
739 spin_lock_irqsave (&sched->kss_lock, *irq_flags);
741 spin_lock_irqsave (&sched->kss_lock, *irq_flags);
743 /* back onto HEAD of tx_queue */
744 list_add (&tx->tx_list, &conn->ksnc_tx_queue);
747 /* no space to write, or nothing to write? */
748 if (!conn->ksnc_tx_ready ||
749 list_empty (&conn->ksnc_tx_queue)) {
750 /* mark not scheduled */
751 conn->ksnc_tx_scheduled = 0;
752 /* drop scheduler's ref */
753 ksocknal_put_conn (conn);
756 list_add_tail (&conn->ksnc_tx_list, &sched->kss_tx_conns);
761 ksocknal_launch_autoconnect_locked (ksock_route_t *route)
765 /* called holding write lock on ksnd_global_lock */
767 LASSERT (route->ksnr_conn == NULL);
768 LASSERT (!route->ksnr_deleted && !route->ksnr_connecting);
770 route->ksnr_connecting = 1;
771 atomic_inc (&route->ksnr_refcount); /* extra ref for asynchd */
773 spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
775 list_add_tail (&route->ksnr_connect_list,
776 &ksocknal_data.ksnd_autoconnectd_routes);
777 wake_up (&ksocknal_data.ksnd_autoconnectd_waitq);
779 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
783 ksocknal_find_target_peer_locked (ksock_tx_t *tx, ptl_nid_t nid)
785 ptl_nid_t target_nid;
787 ksock_peer_t *peer = ksocknal_find_peer_locked (nid);
793 CERROR ("Can't send packet to "LPX64
794 ": routed target is not a peer\n", nid);
798 rc = kpr_lookup (&ksocknal_data.ksnd_router, nid, tx->tx_nob,
801 CERROR ("Can't route to "LPX64": router error %d\n", nid, rc);
805 peer = ksocknal_find_peer_locked (target_nid);
809 CERROR ("Can't send packet to "LPX64": no peer entry\n", target_nid);
814 ksocknal_find_conn_locked (ksock_tx_t *tx, ksock_peer_t *peer)
816 struct list_head *tmp;
817 ksock_conn_t *conn = NULL;
819 /* Find the conn with the shortest tx queue */
820 list_for_each (tmp, &peer->ksnp_conns) {
821 ksock_conn_t *c = list_entry (tmp, ksock_conn_t, ksnc_list);
823 LASSERT (!c->ksnc_closing);
826 atomic_read (&conn->ksnc_tx_nob) >
827 atomic_read (&c->ksnc_tx_nob))
835 ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn)
838 ksock_sched_t *sched = conn->ksnc_scheduler;
840 /* called holding global lock (read or irq-write) */
842 CDEBUG (D_NET, "Sending to "LPX64" on port %d\n",
843 conn->ksnc_peer->ksnp_nid, conn->ksnc_port);
845 atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
846 tx->tx_resid = tx->tx_nob;
850 zccd_init (&tx->tx_zccd, ksocknal_zc_callback);
851 /* NB this sets 1 ref on zccd, so the callback can only occur after
852 * I've released this ref. */
855 spin_lock_irqsave (&sched->kss_lock, flags);
857 list_add_tail (&tx->tx_list, &conn->ksnc_tx_queue);
859 if (conn->ksnc_tx_ready && /* able to send */
860 !conn->ksnc_tx_scheduled) { /* not scheduled to send */
861 /* +1 ref for scheduler */
862 atomic_inc (&conn->ksnc_refcount);
863 list_add_tail (&conn->ksnc_tx_list,
864 &sched->kss_tx_conns);
865 conn->ksnc_tx_scheduled = 1;
866 wake_up (&sched->kss_waitq);
869 spin_unlock_irqrestore (&sched->kss_lock, flags);
873 ksocknal_find_connectable_route_locked (ksock_peer_t *peer, int eager_only)
875 struct list_head *tmp;
876 ksock_route_t *route;
878 list_for_each (tmp, &peer->ksnp_routes) {
879 route = list_entry (tmp, ksock_route_t, ksnr_list);
881 if (route->ksnr_conn == NULL && /* not connected */
882 !route->ksnr_connecting && /* not connecting */
883 (!eager_only || route->ksnr_eager) && /* wants to connect */
884 time_after_eq (jiffies, route->ksnr_timeout)) /* OK to retry */
892 ksocknal_find_connecting_route_locked (ksock_peer_t *peer)
894 struct list_head *tmp;
895 ksock_route_t *route;
897 list_for_each (tmp, &peer->ksnp_routes) {
898 route = list_entry (tmp, ksock_route_t, ksnr_list);
900 if (route->ksnr_connecting)
908 ksocknal_launch_packet (ksock_tx_t *tx, ptl_nid_t nid)
913 ksock_route_t *route;
916 /* Ensure the frags we've been given EXACTLY match the number of
917 * bytes we want to send. Many TCP/IP stacks disregard any total
918 * size parameters passed to them and just look at the frags.
920 * We always expect at least 1 mapped fragment containing the
921 * complete portals header. */
922 LASSERT (lib_iov_nob (tx->tx_niov, tx->tx_iov) +
923 lib_kiov_nob (tx->tx_nkiov, tx->tx_kiov) == tx->tx_nob);
924 LASSERT (tx->tx_niov >= 1);
925 LASSERT (tx->tx_iov[0].iov_len >= sizeof (ptl_hdr_t));
927 CDEBUG (D_NET, "packet %p type %d, nob %d niov %d nkiov %d\n",
928 tx, ((ptl_hdr_t *)tx->tx_iov[0].iov_base)->type,
929 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
931 tx->tx_conn = NULL; /* only set when assigned a conn */
933 g_lock = &ksocknal_data.ksnd_global_lock;
936 peer = ksocknal_find_target_peer_locked (tx, nid);
938 read_unlock (g_lock);
942 if (ksocknal_find_connectable_route_locked(peer, 1) == NULL) {
943 conn = ksocknal_find_conn_locked (tx, peer);
945 /* I've got no unconnected autoconnect routes that
946 * need to be connected, and I do have an actual
948 ksocknal_queue_tx_locked (tx, conn);
949 read_unlock (g_lock);
954 /* Making one or more connections; I'll need a write lock... */
956 atomic_inc (&peer->ksnp_refcount); /* +1 ref for me while I unlock */
957 read_unlock (g_lock);
958 write_lock_irqsave (g_lock, flags);
960 if (peer->ksnp_closing) { /* peer deleted as I blocked! */
961 write_unlock_irqrestore (g_lock, flags);
962 ksocknal_put_peer (peer);
965 ksocknal_put_peer (peer); /* drop ref I got above */
969 /* launch all eager autoconnections */
970 route = ksocknal_find_connectable_route_locked (peer, 1);
974 ksocknal_launch_autoconnect_locked (route);
977 conn = ksocknal_find_conn_locked (tx, peer);
979 /* Connection exists; queue message on it */
980 ksocknal_queue_tx_locked (tx, conn);
981 write_unlock_irqrestore (g_lock, flags);
985 if (ksocknal_find_connecting_route_locked (peer) == NULL) {
986 /* no autoconnect routes actually connecting now. Scrape
987 * the barrel for non-eager autoconnects */
988 route = ksocknal_find_connectable_route_locked (peer, 0);
990 ksocknal_launch_autoconnect_locked (route);
992 write_unlock_irqrestore (g_lock, flags);
997 /* At least 1 connection is being established; queue the message... */
998 list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
1000 write_unlock_irqrestore (g_lock, flags);
1005 ksocknal_setup_hdr (nal_cb_t *nal, void *private, lib_msg_t *cookie,
1006 ptl_hdr_t *hdr, int type)
1010 /* I may not block for a transmit descriptor if I might block the
1011 * receiver, or an interrupt handler. */
1012 ltx = ksocknal_get_ltx (!(type == PTL_MSG_ACK ||
1013 type == PTL_MSG_REPLY ||
1016 CERROR ("Can't allocate tx desc\n");
1020 /* Init local send packet (storage for hdr, finalize() args) */
1021 ltx->ltx_hdr = *hdr;
1022 ltx->ltx_private = private;
1023 ltx->ltx_cookie = cookie;
1025 /* Init common ltx_tx */
1026 ltx->ltx_tx.tx_isfwd = 0;
1027 ltx->ltx_tx.tx_nob = sizeof (*hdr);
1029 /* We always have 1 mapped frag for the header */
1030 ltx->ltx_tx.tx_niov = 1;
1031 ltx->ltx_tx.tx_iov = <x->ltx_iov_space.hdr;
1032 ltx->ltx_tx.tx_iov[0].iov_base = <x->ltx_hdr;
1033 ltx->ltx_tx.tx_iov[0].iov_len = sizeof (ltx->ltx_hdr);
1035 ltx->ltx_tx.tx_kiov = NULL;
1036 ltx->ltx_tx.tx_nkiov = 0;
1042 ksocknal_send (nal_cb_t *nal, void *private, lib_msg_t *cookie,
1043 ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid,
1044 unsigned int payload_niov, struct iovec *payload_iov,
1050 /* NB 'private' is different depending on what we're sending.
1051 * Just ignore it until we can rely on it
1055 "sending "LPSZ" bytes in %d mapped frags to nid: "LPX64
1056 " pid %d\n", payload_len, payload_niov, nid, pid);
1058 ltx = ksocknal_setup_hdr (nal, private, cookie, hdr, type);
1062 /* append the payload_iovs to the one pointing at the header */
1063 LASSERT (ltx->ltx_tx.tx_niov == 1 && ltx->ltx_tx.tx_nkiov == 0);
1064 LASSERT (payload_niov <= PTL_MD_MAX_IOV);
1066 memcpy (ltx->ltx_tx.tx_iov + 1, payload_iov,
1067 payload_niov * sizeof (*payload_iov));
1068 ltx->ltx_tx.tx_niov = 1 + payload_niov;
1069 ltx->ltx_tx.tx_nob = sizeof (*hdr) + payload_len;
1071 rc = ksocknal_launch_packet (<x->ltx_tx, nid);
1073 ksocknal_put_ltx (ltx);
1079 ksocknal_send_pages (nal_cb_t *nal, void *private, lib_msg_t *cookie,
1080 ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid,
1081 unsigned int payload_niov, ptl_kiov_t *payload_iov, size_t payload_len)
1086 /* NB 'private' is different depending on what we're sending.
1087 * Just ignore it until we can rely on it */
1090 "sending "LPSZ" bytes in %d mapped frags to nid: "LPX64" pid %d\n",
1091 payload_len, payload_niov, nid, pid);
1093 ltx = ksocknal_setup_hdr (nal, private, cookie, hdr, type);
1097 LASSERT (ltx->ltx_tx.tx_niov == 1 && ltx->ltx_tx.tx_nkiov == 0);
1098 LASSERT (payload_niov <= PTL_MD_MAX_IOV);
1100 ltx->ltx_tx.tx_kiov = ltx->ltx_iov_space.payload.kiov;
1101 memcpy (ltx->ltx_tx.tx_kiov, payload_iov,
1102 payload_niov * sizeof (*payload_iov));
1103 ltx->ltx_tx.tx_nkiov = payload_niov;
1104 ltx->ltx_tx.tx_nob = sizeof (*hdr) + payload_len;
1106 rc = ksocknal_launch_packet (<x->ltx_tx, nid);
1108 ksocknal_put_ltx (ltx);
1114 ksocknal_fwd_packet (void *arg, kpr_fwd_desc_t *fwd)
1116 ptl_nid_t nid = fwd->kprfd_gateway_nid;
1117 ksock_tx_t *tx = (ksock_tx_t *)&fwd->kprfd_scratch;
1120 CDEBUG (D_NET, "Forwarding [%p] -> "LPX64" ("LPX64"))\n", fwd,
1121 fwd->kprfd_gateway_nid, fwd->kprfd_target_nid);
1123 /* I'm the gateway; must be the last hop */
1124 if (nid == ksocknal_lib.ni.nid)
1125 nid = fwd->kprfd_target_nid;
1127 tx->tx_isfwd = 1; /* This is a forwarding packet */
1128 tx->tx_nob = fwd->kprfd_nob;
1129 tx->tx_niov = fwd->kprfd_niov;
1130 tx->tx_iov = fwd->kprfd_iov;
1133 tx->tx_hdr = (ptl_hdr_t *)fwd->kprfd_iov[0].iov_base;
1135 rc = ksocknal_launch_packet (tx, nid);
1137 /* FIXME, could pass a better completion error */
1138 kpr_fwd_done (&ksocknal_data.ksnd_router, fwd, -EHOSTUNREACH);
1143 ksocknal_thread_start (int (*fn)(void *arg), void *arg)
1145 long pid = kernel_thread (fn, arg, 0);
1150 atomic_inc (&ksocknal_data.ksnd_nthreads);
1155 ksocknal_thread_fini (void)
1157 atomic_dec (&ksocknal_data.ksnd_nthreads);
1161 ksocknal_fmb_callback (void *arg, int error)
1163 ksock_fmb_t *fmb = (ksock_fmb_t *)arg;
1164 ksock_fmb_pool_t *fmp = fmb->fmb_pool;
1165 ptl_hdr_t *hdr = (ptl_hdr_t *) page_address(fmb->fmb_pages[0]);
1166 ksock_conn_t *conn = NULL;
1167 ksock_sched_t *sched;
1168 unsigned long flags;
1171 CERROR("Failed to route packet from "LPX64" to "LPX64": %d\n",
1172 NTOH__u64(hdr->src_nid), NTOH__u64(hdr->dest_nid),
1175 CDEBUG (D_NET, "routed packet from "LPX64" to "LPX64": OK\n",
1176 NTOH__u64 (hdr->src_nid), NTOH__u64 (hdr->dest_nid));
1178 /* drop peer ref taken on init */
1179 ksocknal_put_peer (fmb->fmb_peer);
1181 spin_lock_irqsave (&fmp->fmp_lock, flags);
1183 list_add (&fmb->fmb_list, &fmp->fmp_idle_fmbs);
1185 if (!list_empty (&fmp->fmp_blocked_conns)) {
1186 conn = list_entry (fmb->fmb_pool->fmp_blocked_conns.next,
1187 ksock_conn_t, ksnc_rx_list);
1188 list_del (&conn->ksnc_rx_list);
1191 spin_unlock_irqrestore (&fmp->fmp_lock, flags);
1196 CDEBUG (D_NET, "Scheduling conn %p\n", conn);
1197 LASSERT (conn->ksnc_rx_scheduled);
1198 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_FMB_SLEEP);
1200 conn->ksnc_rx_state = SOCKNAL_RX_GET_FMB;
1202 sched = conn->ksnc_scheduler;
1204 spin_lock_irqsave (&sched->kss_lock, flags);
1206 list_add_tail (&conn->ksnc_rx_list, &sched->kss_rx_conns);
1207 wake_up (&sched->kss_waitq);
1209 spin_unlock_irqrestore (&sched->kss_lock, flags);
1213 ksocknal_get_idle_fmb (ksock_conn_t *conn)
1215 int payload_nob = conn->ksnc_rx_nob_left;
1216 int packet_nob = sizeof (ptl_hdr_t) + payload_nob;
1217 unsigned long flags;
1218 ksock_fmb_pool_t *pool;
1221 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB);
1222 LASSERT (ksocknal_data.ksnd_fmbs != NULL);
1224 if (packet_nob <= SOCKNAL_SMALL_FWD_PAGES * PAGE_SIZE)
1225 pool = &ksocknal_data.ksnd_small_fmp;
1227 pool = &ksocknal_data.ksnd_large_fmp;
1229 spin_lock_irqsave (&pool->fmp_lock, flags);
1231 if (!list_empty (&pool->fmp_idle_fmbs)) {
1232 fmb = list_entry(pool->fmp_idle_fmbs.next,
1233 ksock_fmb_t, fmb_list);
1234 list_del (&fmb->fmb_list);
1235 spin_unlock_irqrestore (&pool->fmp_lock, flags);
1240 /* deschedule until fmb free */
1242 conn->ksnc_rx_state = SOCKNAL_RX_FMB_SLEEP;
1244 list_add_tail (&conn->ksnc_rx_list,
1245 &pool->fmp_blocked_conns);
1247 spin_unlock_irqrestore (&pool->fmp_lock, flags);
1252 ksocknal_init_fmb (ksock_conn_t *conn, ksock_fmb_t *fmb)
1254 int payload_nob = conn->ksnc_rx_nob_left;
1255 int packet_nob = sizeof (ptl_hdr_t) + payload_nob;
1256 ptl_nid_t dest_nid = NTOH__u64 (conn->ksnc_hdr.dest_nid);
1257 int niov; /* at least the header */
1260 LASSERT (conn->ksnc_rx_scheduled);
1261 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB);
1262 LASSERT (conn->ksnc_rx_nob_wanted == conn->ksnc_rx_nob_left);
1263 LASSERT (payload_nob >= 0);
1264 LASSERT (packet_nob <= fmb->fmb_npages * PAGE_SIZE);
1265 LASSERT (sizeof (ptl_hdr_t) < PAGE_SIZE);
1267 /* Got a forwarding buffer; copy the header we just read into the
1268 * forwarding buffer. If there's payload, start reading reading it
1269 * into the buffer, otherwise the forwarding buffer can be kicked
1272 * NB fmb->fmb_iov spans the WHOLE packet.
1273 * conn->ksnc_rx_iov spans just the payload.
1275 fmb->fmb_iov[0].iov_base = page_address (fmb->fmb_pages[0]);
1278 memcpy (fmb->fmb_iov[0].iov_base, &conn->ksnc_hdr, sizeof (ptl_hdr_t));
1280 /* Take a ref on the conn's peer to prevent module unload before
1281 * forwarding completes. NB we ref peer and not conn since because
1282 * all refs on conn after it has been closed must remove themselves
1284 fmb->fmb_peer = conn->ksnc_peer;
1285 atomic_inc (&conn->ksnc_peer->ksnp_refcount);
1287 if (payload_nob == 0) { /* got complete packet already */
1288 CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d fwd_start (immediate)\n",
1289 conn, NTOH__u64 (conn->ksnc_hdr.src_nid),
1290 dest_nid, packet_nob);
1292 fmb->fmb_iov[0].iov_len = sizeof (ptl_hdr_t);
1294 kpr_fwd_init (&fmb->fmb_fwd, dest_nid,
1295 packet_nob, 1, fmb->fmb_iov,
1296 ksocknal_fmb_callback, fmb);
1298 /* forward it now */
1299 kpr_fwd_start (&ksocknal_data.ksnd_router, &fmb->fmb_fwd);
1301 ksocknal_new_packet (conn, 0); /* on to next packet */
1306 if (packet_nob <= PAGE_SIZE) { /* whole packet fits in first page */
1307 fmb->fmb_iov[0].iov_len = packet_nob;
1309 fmb->fmb_iov[0].iov_len = PAGE_SIZE;
1310 nob = packet_nob - PAGE_SIZE;
1313 LASSERT (niov < fmb->fmb_npages);
1314 fmb->fmb_iov[niov].iov_base =
1315 page_address (fmb->fmb_pages[niov]);
1316 fmb->fmb_iov[niov].iov_len = MIN (PAGE_SIZE, nob);
1322 kpr_fwd_init (&fmb->fmb_fwd, dest_nid,
1323 packet_nob, niov, fmb->fmb_iov,
1324 ksocknal_fmb_callback, fmb);
1326 conn->ksnc_cookie = fmb; /* stash fmb for later */
1327 conn->ksnc_rx_state = SOCKNAL_RX_BODY_FWD; /* read in the payload */
1329 /* payload is desc's iov-ed buffer, but skipping the hdr */
1330 LASSERT (niov <= sizeof (conn->ksnc_rx_iov_space) /
1331 sizeof (struct iovec));
1333 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1334 conn->ksnc_rx_iov[0].iov_base =
1335 (void *)(((unsigned long)fmb->fmb_iov[0].iov_base) +
1336 sizeof (ptl_hdr_t));
1337 conn->ksnc_rx_iov[0].iov_len =
1338 fmb->fmb_iov[0].iov_len - sizeof (ptl_hdr_t);
1341 memcpy(&conn->ksnc_rx_iov[1], &fmb->fmb_iov[1],
1342 (niov - 1) * sizeof (struct iovec));
1344 conn->ksnc_rx_niov = niov;
1346 CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d reading body\n", conn,
1347 NTOH__u64 (conn->ksnc_hdr.src_nid), dest_nid, payload_nob);
1352 ksocknal_fwd_parse (ksock_conn_t *conn)
1355 ptl_nid_t dest_nid = NTOH__u64 (conn->ksnc_hdr.dest_nid);
1356 int body_len = NTOH__u32 (PTL_HDR_LENGTH(&conn->ksnc_hdr));
1358 CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d parsing header\n", conn,
1359 NTOH__u64 (conn->ksnc_hdr.src_nid),
1360 dest_nid, conn->ksnc_rx_nob_left);
1362 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_HEADER);
1363 LASSERT (conn->ksnc_rx_scheduled);
1365 if (body_len < 0) { /* length corrupt (overflow) */
1366 CERROR("dropping packet from "LPX64" for "LPX64": packet "
1367 "size %d illegal\n", NTOH__u64 (conn->ksnc_hdr.src_nid),
1368 dest_nid, body_len);
1370 ksocknal_new_packet (conn, 0); /* on to new packet */
1371 ksocknal_close_conn_unlocked (conn, -EINVAL); /* give up on conn */
1375 if (ksocknal_data.ksnd_fmbs == NULL) { /* not forwarding */
1376 CERROR("dropping packet from "LPX64" for "LPX64": not "
1377 "forwarding\n", conn->ksnc_hdr.src_nid,
1378 conn->ksnc_hdr.dest_nid);
1379 /* on to new packet (skip this one's body) */
1380 ksocknal_new_packet (conn, body_len);
1384 if (body_len > SOCKNAL_MAX_FWD_PAYLOAD) { /* too big to forward */
1385 CERROR ("dropping packet from "LPX64" for "LPX64
1386 ": packet size %d too big\n", conn->ksnc_hdr.src_nid,
1387 conn->ksnc_hdr.dest_nid, body_len);
1388 /* on to new packet (skip this one's body) */
1389 ksocknal_new_packet (conn, body_len);
1393 /* should have gone direct */
1394 peer = ksocknal_get_peer (conn->ksnc_hdr.dest_nid);
1396 CERROR ("dropping packet from "LPX64" for "LPX64
1397 ": target is a peer\n", conn->ksnc_hdr.src_nid,
1398 conn->ksnc_hdr.dest_nid);
1399 ksocknal_put_peer (peer); /* drop ref from get above */
1401 /* on to next packet (skip this one's body) */
1402 ksocknal_new_packet (conn, body_len);
1406 conn->ksnc_rx_state = SOCKNAL_RX_GET_FMB; /* Getting FMB now */
1407 conn->ksnc_rx_nob_left = body_len; /* stash packet size */
1408 conn->ksnc_rx_nob_wanted = body_len; /* (no slop) */
1412 ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip)
1414 static char ksocknal_slop_buffer[4096];
1420 if (nob_to_skip == 0) { /* right at next packet boundary now */
1421 conn->ksnc_rx_started = 0;
1422 mb (); /* racing with timeout thread */
1424 conn->ksnc_rx_state = SOCKNAL_RX_HEADER;
1425 conn->ksnc_rx_nob_wanted = sizeof (ptl_hdr_t);
1426 conn->ksnc_rx_nob_left = sizeof (ptl_hdr_t);
1428 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1429 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_hdr;
1430 conn->ksnc_rx_iov[0].iov_len = sizeof (ptl_hdr_t);
1431 conn->ksnc_rx_niov = 1;
1433 conn->ksnc_rx_kiov = NULL;
1434 conn->ksnc_rx_nkiov = 0;
1438 /* Set up to skip as much a possible now. If there's more left
1439 * (ran out of iov entries) we'll get called again */
1441 conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1442 conn->ksnc_rx_nob_left = nob_to_skip;
1443 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1448 nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer));
1450 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1451 conn->ksnc_rx_iov[niov].iov_len = nob;
1456 } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */
1457 niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec));
1459 conn->ksnc_rx_niov = niov;
1460 conn->ksnc_rx_kiov = NULL;
1461 conn->ksnc_rx_nkiov = 0;
1462 conn->ksnc_rx_nob_wanted = skipped;
1467 ksocknal_process_receive (ksock_sched_t *sched, unsigned long *irq_flags)
1473 /* NB: sched->ksnc_lock lock held */
1475 LASSERT (!list_empty (&sched->kss_rx_conns));
1476 conn = list_entry(sched->kss_rx_conns.next, ksock_conn_t, ksnc_rx_list);
1477 list_del (&conn->ksnc_rx_list);
1479 spin_unlock_irqrestore (&sched->kss_lock, *irq_flags);
1481 CDEBUG(D_NET, "sched %p conn %p\n", sched, conn);
1482 LASSERT (atomic_read (&conn->ksnc_refcount) > 0);
1483 LASSERT (conn->ksnc_rx_scheduled);
1484 LASSERT (conn->ksnc_rx_ready);
1486 /* doesn't need a forwarding buffer */
1487 if (conn->ksnc_rx_state != SOCKNAL_RX_GET_FMB)
1491 fmb = ksocknal_get_idle_fmb (conn);
1492 if (fmb == NULL) { /* conn descheduled waiting for idle fmb */
1493 spin_lock_irqsave (&sched->kss_lock, *irq_flags);
1497 if (ksocknal_init_fmb (conn, fmb)) /* packet forwarded ? */
1498 goto out; /* come back later for next packet */
1501 /* NB: sched lock NOT held */
1502 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_HEADER ||
1503 conn->ksnc_rx_state == SOCKNAL_RX_BODY ||
1504 conn->ksnc_rx_state == SOCKNAL_RX_BODY_FWD ||
1505 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1507 LASSERT (conn->ksnc_rx_nob_wanted > 0);
1509 conn->ksnc_rx_ready = 0;/* data ready may race with me and set ready */
1510 mb(); /* => clear BEFORE trying to read */
1512 rc = ksocknal_recvmsg(conn);
1515 if (ksocknal_close_conn_unlocked (conn, rc)) {
1516 /* I'm the first to close */
1518 CERROR ("[%p] Error %d on read from "LPX64" ip %08x:%d\n",
1519 conn, rc, conn->ksnc_peer->ksnp_nid,
1520 conn->ksnc_ipaddr, conn->ksnc_port);
1522 CWARN ("[%p] EOF from "LPX64" ip %08x:%d\n",
1523 conn, conn->ksnc_peer->ksnp_nid,
1524 conn->ksnc_ipaddr, conn->ksnc_port);
1530 if (conn->ksnc_rx_nob_wanted != 0) /* short read */
1531 goto out; /* try again later */
1533 /* got all I wanted, assume there's more - prevent data_ready locking */
1534 conn->ksnc_rx_ready = 1;
1536 switch (conn->ksnc_rx_state) {
1537 case SOCKNAL_RX_HEADER:
1538 if (conn->ksnc_hdr.type != HTON__u32(PTL_MSG_HELLO) &&
1539 NTOH__u64(conn->ksnc_hdr.dest_nid) != ksocknal_lib.ni.nid) {
1540 /* This packet isn't for me */
1541 ksocknal_fwd_parse (conn);
1542 switch (conn->ksnc_rx_state) {
1543 case SOCKNAL_RX_HEADER: /* skipped (zero payload) */
1544 goto out; /* => come back later */
1545 case SOCKNAL_RX_SLOP: /* skipping packet's body */
1546 goto try_read; /* => go read it */
1547 case SOCKNAL_RX_GET_FMB: /* forwarding */
1548 goto get_fmb; /* => go get a fwd msg buffer */
1555 /* sets wanted_len, iovs etc */
1556 lib_parse(&ksocknal_lib, &conn->ksnc_hdr, conn);
1558 if (conn->ksnc_rx_nob_wanted != 0) { /* need to get payload? */
1559 conn->ksnc_rx_state = SOCKNAL_RX_BODY;
1560 goto try_read; /* go read the payload */
1562 /* Fall through (completed packet for me) */
1564 case SOCKNAL_RX_BODY:
1565 /* payload all received */
1566 lib_finalize(&ksocknal_lib, NULL, conn->ksnc_cookie);
1569 case SOCKNAL_RX_SLOP:
1570 /* starting new packet? */
1571 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1572 goto out; /* come back later */
1573 goto try_read; /* try to finish reading slop now */
1575 case SOCKNAL_RX_BODY_FWD:
1576 /* payload all received */
1577 CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d fwd_start (got body)\n",
1578 conn, NTOH__u64 (conn->ksnc_hdr.src_nid),
1579 NTOH__u64 (conn->ksnc_hdr.dest_nid),
1580 conn->ksnc_rx_nob_left);
1582 /* forward the packet. NB ksocknal_init_fmb() put fmb into
1583 * conn->ksnc_cookie */
1584 fmb = (ksock_fmb_t *)conn->ksnc_cookie;
1585 kpr_fwd_start (&ksocknal_data.ksnd_router, &fmb->fmb_fwd);
1587 /* no slop in forwarded packets */
1588 LASSERT (conn->ksnc_rx_nob_left == 0);
1590 ksocknal_new_packet (conn, 0); /* on to next packet */
1591 goto out; /* (later) */
1601 spin_lock_irqsave (&sched->kss_lock, *irq_flags);
1603 /* no data there to read? */
1604 if (!conn->ksnc_rx_ready) {
1605 /* let socket callback schedule again */
1606 conn->ksnc_rx_scheduled = 0;
1607 /* drop scheduler's ref */
1608 ksocknal_put_conn (conn);
1610 /* stay scheduled */
1611 list_add_tail (&conn->ksnc_rx_list, &sched->kss_rx_conns);
1616 ksocknal_recv (nal_cb_t *nal, void *private, lib_msg_t *msg,
1617 unsigned int niov, struct iovec *iov, size_t mlen, size_t rlen)
1619 ksock_conn_t *conn = (ksock_conn_t *)private;
1621 LASSERT (mlen <= rlen);
1622 LASSERT (niov <= PTL_MD_MAX_IOV);
1624 conn->ksnc_cookie = msg;
1625 conn->ksnc_rx_nob_wanted = mlen;
1626 conn->ksnc_rx_nob_left = rlen;
1628 conn->ksnc_rx_nkiov = 0;
1629 conn->ksnc_rx_kiov = NULL;
1630 conn->ksnc_rx_niov = niov;
1631 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1632 memcpy (conn->ksnc_rx_iov, iov, niov * sizeof (*iov));
1635 lib_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1636 lib_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1642 ksocknal_recv_pages (nal_cb_t *nal, void *private, lib_msg_t *msg,
1643 unsigned int niov, ptl_kiov_t *kiov, size_t mlen, size_t rlen)
1645 ksock_conn_t *conn = (ksock_conn_t *)private;
1647 LASSERT (mlen <= rlen);
1648 LASSERT (niov <= PTL_MD_MAX_IOV);
1650 conn->ksnc_cookie = msg;
1651 conn->ksnc_rx_nob_wanted = mlen;
1652 conn->ksnc_rx_nob_left = rlen;
1654 conn->ksnc_rx_niov = 0;
1655 conn->ksnc_rx_iov = NULL;
1656 conn->ksnc_rx_nkiov = niov;
1657 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1658 memcpy (conn->ksnc_rx_kiov, kiov, niov * sizeof (*kiov));
1661 lib_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1662 lib_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1667 int ksocknal_scheduler (void *arg)
1669 ksock_sched_t *sched = (ksock_sched_t *)arg;
1670 unsigned long flags;
1673 int id = sched - ksocknal_data.ksnd_schedulers;
1676 snprintf (name, sizeof (name),"ksocknald_%02d", id);
1677 kportal_daemonize (name);
1678 kportal_blockallsigs ();
1680 #if (CONFIG_SMP && CPU_AFFINITY)
1681 if ((cpu_online_map & (1 << id)) != 0) {
1683 current->cpus_allowed = (1 << id);
1685 set_cpus_allowed (current, 1<<id);
1688 CERROR ("Can't set CPU affinity for %s\n", name);
1690 #endif /* CONFIG_SMP && CPU_AFFINITY */
1692 spin_lock_irqsave (&sched->kss_lock, flags);
1694 while (!ksocknal_data.ksnd_shuttingdown) {
1695 int did_something = 0;
1697 /* Ensure I progress everything semi-fairly */
1699 if (!list_empty (&sched->kss_rx_conns)) {
1701 /* drops & regains kss_lock */
1702 ksocknal_process_receive (sched, &flags);
1705 if (!list_empty (&sched->kss_tx_conns)) {
1707 /* drops and regains kss_lock */
1708 ksocknal_process_transmit (sched, &flags);
1711 if (!list_empty (&sched->kss_zctxdone_list)) {
1713 list_entry(sched->kss_zctxdone_list.next,
1714 ksock_tx_t, tx_list);
1717 list_del (&tx->tx_list);
1718 spin_unlock_irqrestore (&sched->kss_lock, flags);
1720 ksocknal_tx_done (tx, 1);
1722 spin_lock_irqsave (&sched->kss_lock, flags);
1725 if (!did_something || /* nothing to do */
1726 ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1727 spin_unlock_irqrestore (&sched->kss_lock, flags);
1731 if (!did_something) { /* wait for something to do */
1733 rc = wait_event_interruptible (sched->kss_waitq,
1734 ksocknal_data.ksnd_shuttingdown ||
1735 !list_empty(&sched->kss_rx_conns) ||
1736 !list_empty(&sched->kss_tx_conns) ||
1737 !list_empty(&sched->kss_zctxdone_list));
1739 rc = wait_event_interruptible (sched->kss_waitq,
1740 ksocknal_data.ksnd_shuttingdown ||
1741 !list_empty(&sched->kss_rx_conns) ||
1742 !list_empty(&sched->kss_tx_conns));
1748 spin_lock_irqsave (&sched->kss_lock, flags);
1752 spin_unlock_irqrestore (&sched->kss_lock, flags);
1753 ksocknal_thread_fini ();
1758 ksocknal_data_ready (struct sock *sk, int n)
1760 unsigned long flags;
1762 ksock_sched_t *sched;
1765 /* interleave correctly with closing sockets... */
1766 read_lock (&ksocknal_data.ksnd_global_lock);
1768 conn = sk->sk_user_data;
1769 if (conn == NULL) { /* raced with ksocknal_close_sock */
1770 LASSERT (sk->sk_data_ready != &ksocknal_data_ready);
1771 sk->sk_data_ready (sk, n);
1775 if (!conn->ksnc_rx_ready) { /* new news */
1776 /* Set ASAP in case of concurrent calls to me */
1777 conn->ksnc_rx_ready = 1;
1779 sched = conn->ksnc_scheduler;
1781 spin_lock_irqsave (&sched->kss_lock, flags);
1783 /* Set again (process_receive may have cleared while I blocked for the lock) */
1784 conn->ksnc_rx_ready = 1;
1786 if (!conn->ksnc_rx_scheduled) { /* not being progressed */
1787 list_add_tail(&conn->ksnc_rx_list,
1788 &sched->kss_rx_conns);
1789 conn->ksnc_rx_scheduled = 1;
1790 /* extra ref for scheduler */
1791 atomic_inc (&conn->ksnc_refcount);
1793 wake_up (&sched->kss_waitq);
1796 spin_unlock_irqrestore (&sched->kss_lock, flags);
1800 read_unlock (&ksocknal_data.ksnd_global_lock);
1806 ksocknal_write_space (struct sock *sk)
1808 unsigned long flags;
1810 ksock_sched_t *sched;
1812 /* interleave correctly with closing sockets... */
1813 read_lock (&ksocknal_data.ksnd_global_lock);
1815 conn = sk->sk_user_data;
1817 CDEBUG(D_NET, "sk %p wspace %d low water %d conn %p%s%s%s\n",
1818 sk, tcp_wspace(sk), SOCKNAL_TX_LOW_WATER(sk), conn,
1819 (conn == NULL) ? "" : (conn->ksnc_tx_ready ?
1820 " ready" : " blocked"),
1821 (conn == NULL) ? "" : (conn->ksnc_tx_scheduled ?
1822 " scheduled" : " idle"),
1823 (conn == NULL) ? "" : (list_empty (&conn->ksnc_tx_queue) ?
1824 " empty" : " queued"));
1826 if (conn == NULL) { /* raced with ksocknal_close_sock */
1827 LASSERT (sk->sk_write_space != &ksocknal_write_space);
1828 sk->sk_write_space (sk);
1830 read_unlock (&ksocknal_data.ksnd_global_lock);
1834 if (tcp_wspace(sk) >= SOCKNAL_TX_LOW_WATER(sk)) { /* got enough space */
1835 clear_bit (SOCK_NOSPACE, &sk->sk_socket->flags);
1837 if (!conn->ksnc_tx_ready) { /* new news */
1838 /* Set ASAP in case of concurrent calls to me */
1839 conn->ksnc_tx_ready = 1;
1841 sched = conn->ksnc_scheduler;
1843 spin_lock_irqsave (&sched->kss_lock, flags);
1845 /* Set again (process_transmit may have
1846 cleared while I blocked for the lock) */
1847 conn->ksnc_tx_ready = 1;
1849 if (!conn->ksnc_tx_scheduled && // not being progressed
1850 !list_empty(&conn->ksnc_tx_queue)){//packets to send
1851 list_add_tail (&conn->ksnc_tx_list,
1852 &sched->kss_tx_conns);
1853 conn->ksnc_tx_scheduled = 1;
1854 /* extra ref for scheduler */
1855 atomic_inc (&conn->ksnc_refcount);
1857 wake_up (&sched->kss_waitq);
1860 spin_unlock_irqrestore (&sched->kss_lock, flags);
1864 read_unlock (&ksocknal_data.ksnd_global_lock);
1868 ksocknal_sock_write (struct socket *sock, void *buffer, int nob)
1871 mm_segment_t oldmm = get_fs();
1874 struct iovec iov = {
1878 struct msghdr msg = {
1883 .msg_control = NULL,
1884 .msg_controllen = 0,
1889 rc = sock_sendmsg (sock, &msg, iov.iov_len);
1896 CERROR ("Unexpected zero rc\n");
1897 return (-ECONNABORTED);
1900 buffer = ((char *)buffer) + rc;
1908 ksocknal_sock_read (struct socket *sock, void *buffer, int nob)
1911 mm_segment_t oldmm = get_fs();
1914 struct iovec iov = {
1918 struct msghdr msg = {
1923 .msg_control = NULL,
1924 .msg_controllen = 0,
1929 rc = sock_recvmsg (sock, &msg, iov.iov_len, 0);
1936 return (-ECONNABORTED);
1938 buffer = ((char *)buffer) + rc;
1946 ksocknal_exchange_nids (struct socket *sock, ptl_nid_t nid)
1950 ptl_magicversion_t *hmv = (ptl_magicversion_t *)&hdr.dest_nid;
1952 LASSERT (sizeof (*hmv) == sizeof (hdr.dest_nid));
1954 memset (&hdr, 0, sizeof (hdr));
1955 hmv->magic = __cpu_to_le32 (PORTALS_PROTO_MAGIC);
1956 hmv->version_major = __cpu_to_le32 (PORTALS_PROTO_VERSION_MAJOR);
1957 hmv->version_minor = __cpu_to_le32 (PORTALS_PROTO_VERSION_MINOR);
1959 hdr.src_nid = __cpu_to_le64 (ksocknal_lib.ni.nid);
1960 hdr.type = __cpu_to_le32 (PTL_MSG_HELLO);
1962 /* Assume sufficient socket buffering for this message */
1963 rc = ksocknal_sock_write (sock, &hdr, sizeof (hdr));
1965 CERROR ("Error %d sending HELLO to "LPX64"\n", rc, nid);
1969 rc = ksocknal_sock_read (sock, hmv, sizeof (*hmv));
1971 CERROR ("Error %d reading HELLO from "LPX64"\n", rc, nid);
1975 if (hmv->magic != __le32_to_cpu (PORTALS_PROTO_MAGIC)) {
1976 CERROR ("Bad magic %#08x (%#08x expected) from "LPX64"\n",
1977 __cpu_to_le32 (hmv->magic), PORTALS_PROTO_MAGIC, nid);
1981 if (hmv->version_major != __cpu_to_le16 (PORTALS_PROTO_VERSION_MAJOR) ||
1982 hmv->version_minor != __cpu_to_le16 (PORTALS_PROTO_VERSION_MINOR)) {
1983 CERROR ("Incompatible protocol version %d.%d (%d.%d expected)"
1985 __le16_to_cpu (hmv->version_major),
1986 __le16_to_cpu (hmv->version_minor),
1987 PORTALS_PROTO_VERSION_MAJOR,
1988 PORTALS_PROTO_VERSION_MINOR,
1993 LASSERT (PORTALS_PROTO_VERSION_MAJOR == 0);
1994 /* version 0 sends magic/version as the dest_nid of a 'hello' header,
1995 * so read the rest of it in now... */
1997 rc = ksocknal_sock_read (sock, hmv + 1, sizeof (hdr) - sizeof (*hmv));
1999 CERROR ("Error %d reading rest of HELLO hdr from "LPX64"\n",
2004 /* ...and check we got what we expected */
2005 if (hdr.type != __cpu_to_le32 (PTL_MSG_HELLO) ||
2006 PTL_HDR_LENGTH (&hdr) != __cpu_to_le32 (0)) {
2007 CERROR ("Expecting a HELLO hdr with 0 payload,"
2008 " but got type %d with %d payload from "LPX64"\n",
2009 __le32_to_cpu (hdr.type),
2010 __le32_to_cpu (PTL_HDR_LENGTH (&hdr)), nid);
2014 if (__le64_to_cpu (hdr.src_nid) != nid) {
2015 CERROR ("Connected to nid "LPX64", but expecting "LPX64"\n",
2016 __le64_to_cpu (hdr.src_nid), nid);
2024 ksocknal_setup_sock (struct socket *sock)
2026 mm_segment_t oldmm = get_fs ();
2029 struct linger linger;
2031 /* Ensure this socket aborts active sends immediately when we close
2035 linger.l_linger = 0;
2038 rc = sock_setsockopt (sock, SOL_SOCKET, SO_LINGER,
2039 (char *)&linger, sizeof (linger));
2042 CERROR ("Can't set SO_LINGER: %d\n", rc);
2048 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_LINGER2,
2049 (char *)&option, sizeof (option));
2052 CERROR ("Can't set SO_LINGER2: %d\n", rc);
2056 #if SOCKNAL_USE_KEEPALIVES
2057 /* Keepalives: If 3/4 of the timeout elapses, start probing every
2058 * second until the timeout elapses. */
2060 option = (ksocknal_data.ksnd_io_timeout * 3) / 4;
2062 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPIDLE,
2063 (char *)&option, sizeof (option));
2066 CERROR ("Can't set TCP_KEEPIDLE: %d\n", rc);
2072 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPINTVL,
2073 (char *)&option, sizeof (option));
2076 CERROR ("Can't set TCP_KEEPINTVL: %d\n", rc);
2080 option = ksocknal_data.ksnd_io_timeout / 4;
2082 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPCNT,
2083 (char *)&option, sizeof (option));
2086 CERROR ("Can't set TCP_KEEPINTVL: %d\n", rc);
2092 rc = sock_setsockopt (sock, SOL_SOCKET, SO_KEEPALIVE,
2093 (char *)&option, sizeof (option));
2096 CERROR ("Can't set SO_KEEPALIVE: %d\n", rc);
2104 ksocknal_connect_peer (ksock_route_t *route)
2106 struct sockaddr_in peer_addr;
2107 mm_segment_t oldmm = get_fs();
2110 struct socket *sock;
2113 rc = sock_create (PF_INET, SOCK_STREAM, 0, &sock);
2115 CERROR ("Can't create autoconnect socket: %d\n", rc);
2119 /* Ugh; have to map_fd for compatibility with sockets passed in
2120 * from userspace. And we actually need the sock->file refcounting
2121 * that this gives you :) */
2123 fd = sock_map_fd (sock);
2125 sock_release (sock);
2126 CERROR ("sock_map_fd error %d\n", fd);
2130 /* NB the fd now owns the ref on sock->file */
2131 LASSERT (sock->file != NULL);
2132 LASSERT (file_count(sock->file) == 1);
2134 /* Set the socket timeouts, so our connection attempt completes in
2136 tv.tv_sec = ksocknal_data.ksnd_io_timeout;
2140 rc = sock_setsockopt (sock, SOL_SOCKET, SO_SNDTIMEO,
2141 (char *)&tv, sizeof (tv));
2144 CERROR ("Can't set send timeout %d: %d\n",
2145 ksocknal_data.ksnd_io_timeout, rc);
2150 rc = sock_setsockopt (sock, SOL_SOCKET, SO_RCVTIMEO,
2151 (char *)&tv, sizeof (tv));
2154 CERROR ("Can't set receive timeout %d: %d\n",
2155 ksocknal_data.ksnd_io_timeout, rc);
2159 if (route->ksnr_nonagel) {
2163 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_NODELAY,
2164 (char *)&option, sizeof (option));
2167 CERROR ("Can't disable nagel: %d\n", rc);
2172 if (route->ksnr_buffer_size != 0) {
2173 int option = route->ksnr_buffer_size;
2176 rc = sock_setsockopt (sock, SOL_SOCKET, SO_SNDBUF,
2177 (char *)&option, sizeof (option));
2180 CERROR ("Can't set send buffer %d: %d\n",
2181 route->ksnr_buffer_size, rc);
2186 rc = sock_setsockopt (sock, SOL_SOCKET, SO_RCVBUF,
2187 (char *)&option, sizeof (option));
2190 CERROR ("Can't set receive buffer %d: %d\n",
2191 route->ksnr_buffer_size, rc);
2196 memset (&peer_addr, 0, sizeof (peer_addr));
2197 peer_addr.sin_family = AF_INET;
2198 peer_addr.sin_port = htons (route->ksnr_port);
2199 peer_addr.sin_addr.s_addr = htonl (route->ksnr_ipaddr);
2201 rc = sock->ops->connect (sock, (struct sockaddr *)&peer_addr,
2202 sizeof (peer_addr), sock->file->f_flags);
2204 CERROR ("Error %d connecting to "LPX64"\n", rc,
2205 route->ksnr_peer->ksnp_nid);
2209 if (route->ksnr_xchange_nids) {
2210 rc = ksocknal_exchange_nids (sock, route->ksnr_peer->ksnp_nid);
2215 rc = ksocknal_create_conn (route->ksnr_peer->ksnp_nid,
2216 route, sock, route->ksnr_irq_affinity);
2218 /* Take an extra ref on sock->file to compensate for the
2219 * upcoming close which will lose fd's ref on it. */
2220 get_file (sock->file);
2229 ksocknal_autoconnect (ksock_route_t *route)
2231 LIST_HEAD (zombies);
2234 unsigned long flags;
2237 rc = ksocknal_connect_peer (route);
2239 /* successfully autoconnected: create_conn did the
2240 * route/conn binding and scheduled any blocked packets,
2241 * so there's nothing left to do now. */
2245 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
2247 peer = route->ksnr_peer;
2248 route->ksnr_connecting = 0;
2250 LASSERT (route->ksnr_retry_interval != 0);
2251 route->ksnr_timeout = jiffies + route->ksnr_retry_interval;
2252 route->ksnr_retry_interval = MIN (route->ksnr_retry_interval * 2,
2253 SOCKNAL_MAX_RECONNECT_INTERVAL);
2255 if (!list_empty (&peer->ksnp_tx_queue) &&
2256 ksocknal_find_connecting_route_locked (peer) == NULL) {
2257 LASSERT (list_empty (&peer->ksnp_conns));
2259 /* None of the connections that the blocked packets are
2260 * waiting for have been successful. Complete them now... */
2262 tx = list_entry (peer->ksnp_tx_queue.next,
2263 ksock_tx_t, tx_list);
2264 list_del (&tx->tx_list);
2265 list_add_tail (&tx->tx_list, &zombies);
2266 } while (!list_empty (&peer->ksnp_tx_queue));
2269 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
2271 while (!list_empty (&zombies)) {
2272 tx = list_entry (zombies.next, ksock_tx_t, tx_list);
2274 CERROR ("Deleting packet type %d len %d ("LPX64"->"LPX64")\n",
2275 NTOH__u32 (tx->tx_hdr->type),
2276 NTOH__u32 (PTL_HDR_LENGTH(tx->tx_hdr)),
2277 NTOH__u64 (tx->tx_hdr->src_nid),
2278 NTOH__u64 (tx->tx_hdr->dest_nid));
2280 list_del (&tx->tx_list);
2282 ksocknal_tx_done (tx, 0);
2287 ksocknal_autoconnectd (void *arg)
2289 long id = (long)arg;
2291 unsigned long flags;
2292 ksock_route_t *route;
2295 snprintf (name, sizeof (name), "ksocknal_ad%02ld", id);
2296 kportal_daemonize (name);
2297 kportal_blockallsigs ();
2299 spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2301 while (!ksocknal_data.ksnd_shuttingdown) {
2303 if (!list_empty (&ksocknal_data.ksnd_autoconnectd_routes)) {
2304 route = list_entry (ksocknal_data.ksnd_autoconnectd_routes.next,
2305 ksock_route_t, ksnr_connect_list);
2307 list_del (&route->ksnr_connect_list);
2308 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2310 ksocknal_autoconnect (route);
2311 ksocknal_put_route (route);
2313 spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2317 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2319 rc = wait_event_interruptible (ksocknal_data.ksnd_autoconnectd_waitq,
2320 ksocknal_data.ksnd_shuttingdown ||
2321 !list_empty (&ksocknal_data.ksnd_autoconnectd_routes));
2323 spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2326 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2328 ksocknal_thread_fini ();
2333 ksocknal_find_timed_out_conn (ksock_peer_t *peer)
2335 /* We're called with a shared lock on ksnd_global_lock */
2337 struct list_head *ctmp;
2338 ksock_sched_t *sched;
2340 list_for_each (ctmp, &peer->ksnp_conns) {
2341 conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
2342 sched = conn->ksnc_scheduler;
2344 /* Don't need the {get,put}connsock dance to deref ksnc_sock... */
2345 LASSERT (!conn->ksnc_closing);
2347 if (conn->ksnc_rx_started &&
2348 time_after_eq (jiffies, conn->ksnc_rx_deadline)) {
2349 /* Timed out incomplete incoming message */
2350 atomic_inc (&conn->ksnc_refcount);
2351 CERROR ("Timed out RX from "LPX64" %p\n",
2352 peer->ksnp_nid, conn);
2356 if ((!list_empty (&conn->ksnc_tx_queue) ||
2357 conn->ksnc_sock->sk->wmem_queued != 0) &&
2358 time_after_eq (jiffies, conn->ksnc_tx_deadline)) {
2359 /* Timed out messages queued for sending, or
2360 * messages buffered in the socket's send buffer */
2361 atomic_inc (&conn->ksnc_refcount);
2362 CERROR ("Timed out TX to "LPX64" %s%d %p\n",
2364 list_empty (&conn->ksnc_tx_queue) ? "" : "Q ",
2365 conn->ksnc_sock->sk->wmem_queued, conn);
2374 ksocknal_check_peer_timeouts (int idx)
2376 struct list_head *peers = &ksocknal_data.ksnd_peers[idx];
2377 struct list_head *ptmp;
2382 /* NB. We expect to have a look at all the peers and not find any
2383 * connections to time out, so we just use a shared lock while we
2385 read_lock (&ksocknal_data.ksnd_global_lock);
2387 list_for_each (ptmp, peers) {
2388 peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
2389 conn = ksocknal_find_timed_out_conn (peer);
2392 read_unlock (&ksocknal_data.ksnd_global_lock);
2394 if (ksocknal_close_conn_unlocked (conn, -ETIMEDOUT)) {
2395 /* I actually closed... */
2396 CERROR ("Timeout out conn->"LPX64" ip %x:%d\n",
2397 peer->ksnp_nid, conn->ksnc_ipaddr,
2401 /* NB we won't find this one again, but we can't
2402 * just proceed with the next peer, since we dropped
2403 * ksnd_global_lock and it might be dead already! */
2404 ksocknal_put_conn (conn);
2409 read_unlock (&ksocknal_data.ksnd_global_lock);
2413 ksocknal_reaper (void *arg)
2416 unsigned long flags;
2421 unsigned long deadline = jiffies;
2423 kportal_daemonize ("ksocknal_reaper");
2424 kportal_blockallsigs ();
2426 init_waitqueue_entry (&wait, current);
2428 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2430 while (!ksocknal_data.ksnd_shuttingdown) {
2432 if (!list_empty (&ksocknal_data.ksnd_deathrow_conns)) {
2433 conn = list_entry (ksocknal_data.ksnd_deathrow_conns.next,
2434 ksock_conn_t, ksnc_list);
2435 list_del (&conn->ksnc_list);
2437 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2439 ksocknal_terminate_conn (conn);
2440 ksocknal_put_conn (conn);
2442 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2446 if (!list_empty (&ksocknal_data.ksnd_zombie_conns)) {
2447 conn = list_entry (ksocknal_data.ksnd_zombie_conns.next,
2448 ksock_conn_t, ksnc_list);
2449 list_del (&conn->ksnc_list);
2451 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2453 ksocknal_destroy_conn (conn);
2455 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2459 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2461 /* careful with the jiffy wrap... */
2462 while ((timeout = ((int)deadline - (int)jiffies)) <= 0) {
2465 int chunk = ksocknal_data.ksnd_peer_hash_size;
2467 /* Time to check for timeouts on a few more peers: I do
2468 * checks every 'p' seconds on a proportion of the peer
2469 * table and I need to check every connection 'n' times
2470 * within a timeout interval, to ensure I detect a
2471 * timeout on any connection within (n+1)/n times the
2472 * timeout interval. */
2474 if (ksocknal_data.ksnd_io_timeout > n * p)
2475 chunk = (chunk * n * p) /
2476 ksocknal_data.ksnd_io_timeout;
2480 for (i = 0; i < chunk; i++) {
2481 ksocknal_check_peer_timeouts (peer_index);
2482 peer_index = (peer_index + 1) %
2483 ksocknal_data.ksnd_peer_hash_size;
2489 add_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait);
2490 set_current_state (TASK_INTERRUPTIBLE);
2492 if (!ksocknal_data.ksnd_shuttingdown &&
2493 list_empty (&ksocknal_data.ksnd_deathrow_conns) &&
2494 list_empty (&ksocknal_data.ksnd_zombie_conns))
2495 schedule_timeout (timeout);
2497 set_current_state (TASK_RUNNING);
2498 remove_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait);
2500 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2503 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2505 ksocknal_thread_fini ();
2509 nal_cb_t ksocknal_lib = {
2510 nal_data: &ksocknal_data, /* NAL private data */
2511 cb_send: ksocknal_send,
2512 cb_send_pages: ksocknal_send_pages,
2513 cb_recv: ksocknal_recv,
2514 cb_recv_pages: ksocknal_recv_pages,
2515 cb_read: ksocknal_read,
2516 cb_write: ksocknal_write,
2517 cb_callback: ksocknal_callback,
2518 cb_malloc: ksocknal_malloc,
2519 cb_free: ksocknal_free,
2520 cb_printf: ksocknal_printf,
2521 cb_cli: ksocknal_cli,
2522 cb_sti: ksocknal_sti,
2523 cb_dist: ksocknal_dist