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.
27 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
28 # include <linux/syscalls.h>
32 * LIB functions follow
36 ksocknal_read(nal_cb_t *nal, void *private, void *dst_addr,
37 user_ptr src_addr, size_t len)
39 CDEBUG(D_NET, LPX64": reading %ld bytes from %p -> %p\n",
40 nal->ni.nid, (long)len, src_addr, dst_addr);
42 memcpy( dst_addr, src_addr, len );
47 ksocknal_write(nal_cb_t *nal, void *private, user_ptr dst_addr,
48 void *src_addr, size_t len)
50 CDEBUG(D_NET, LPX64": writing %ld bytes from %p -> %p\n",
51 nal->ni.nid, (long)len, src_addr, dst_addr);
53 memcpy( dst_addr, src_addr, len );
58 ksocknal_malloc(nal_cb_t *nal, size_t len)
62 PORTAL_ALLOC(buf, len);
71 ksocknal_free(nal_cb_t *nal, void *buf, size_t len)
73 PORTAL_FREE(buf, len);
77 ksocknal_printf(nal_cb_t *nal, const char *fmt, ...)
83 vsnprintf (msg, sizeof (msg), fmt, ap); /* sprint safely */
86 msg[sizeof (msg) - 1] = 0; /* ensure terminated */
88 CDEBUG (D_NET, "%s", msg);
92 ksocknal_cli(nal_cb_t *nal, unsigned long *flags)
94 ksock_nal_data_t *data = nal->nal_data;
96 /* OK to ignore 'flags'; we're only ever serialise threads and
97 * never need to lock out interrupts */
98 spin_lock(&data->ksnd_nal_cb_lock);
102 ksocknal_sti(nal_cb_t *nal, unsigned long *flags)
104 ksock_nal_data_t *data;
105 data = nal->nal_data;
107 /* OK to ignore 'flags'; we're only ever serialise threads and
108 * never need to lock out interrupts */
109 spin_unlock(&data->ksnd_nal_cb_lock);
113 ksocknal_callback(nal_cb_t *nal, void *private, lib_eq_t *eq, ptl_event_t *ev)
115 /* holding ksnd_nal_cb_lock */
117 if (eq->event_callback != NULL)
118 eq->event_callback(ev);
120 if (waitqueue_active(&ksocknal_data.ksnd_yield_waitq))
121 wake_up_all(&ksocknal_data.ksnd_yield_waitq);
125 ksocknal_dist(nal_cb_t *nal, ptl_nid_t nid, unsigned long *dist)
127 /* I would guess that if ksocknal_get_peer (nid) == NULL,
128 and we're not routing, then 'nid' is very distant :) */
129 if ( nal->ni.nid == nid ) {
139 ksocknal_free_ltx (ksock_ltx_t *ltx)
141 atomic_dec(&ksocknal_data.ksnd_nactive_ltxs);
142 PORTAL_FREE(ltx, ltx->ltx_desc_size);
145 #if (SOCKNAL_ZC && SOCKNAL_VADDR_ZC)
147 ksocknal_kvaddr_to_page (unsigned long vaddr)
151 if (vaddr >= VMALLOC_START &&
153 page = vmalloc_to_page ((void *)vaddr);
155 else if (vaddr >= PKMAP_BASE &&
156 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE))
157 page = vmalloc_to_page ((void *)vaddr);
158 /* in 2.4 ^ just walks the page tables */
161 page = virt_to_page (vaddr);
172 ksocknal_send_iov (ksock_conn_t *conn, ksock_tx_t *tx)
174 struct socket *sock = conn->ksnc_sock;
175 struct iovec *iov = tx->tx_iov;
176 int fragsize = iov->iov_len;
177 unsigned long vaddr = (unsigned long)iov->iov_base;
178 int more = (tx->tx_niov > 1) ||
179 (tx->tx_nkiov > 0) ||
180 (!list_empty (&conn->ksnc_tx_queue));
181 #if (SOCKNAL_ZC && SOCKNAL_VADDR_ZC)
182 int offset = vaddr & (PAGE_SIZE - 1);
183 int zcsize = MIN (fragsize, PAGE_SIZE - offset);
188 /* NB we can't trust socket ops to either consume our iovs
189 * or leave them alone, so we only send 1 frag at a time. */
190 LASSERT (fragsize <= tx->tx_resid);
191 LASSERT (tx->tx_niov > 0);
193 #if (SOCKNAL_ZC && SOCKNAL_VADDR_ZC)
194 if (zcsize >= ksocknal_data.ksnd_zc_min_frag &&
195 (sock->sk->route_caps & NETIF_F_SG) &&
196 (sock->sk->route_caps & (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM | NETIF_F_HW_CSUM)) &&
197 (page = ksocknal_kvaddr_to_page (vaddr)) != NULL) {
199 CDEBUG(D_NET, "vaddr %p, page %p->%p + offset %x for %d\n",
200 (void *)vaddr, page, page_address(page), offset, zcsize);
202 if (fragsize > zcsize) {
207 rc = tcp_sendpage_zccd(sock, page, offset, zcsize,
208 more ? (MSG_DONTWAIT | MSG_MORE) : MSG_DONTWAIT,
213 /* NB don't pass tx's iov; sendmsg may or may not update it */
214 struct iovec fragiov = { .iov_base = (void *)vaddr,
215 .iov_len = fragsize};
216 struct msghdr msg = {
223 .msg_flags = more ? (MSG_DONTWAIT | MSG_MORE) : MSG_DONTWAIT
225 mm_segment_t oldmm = get_fs();
228 rc = sock_sendmsg(sock, &msg, fragsize);
235 if (rc < iov->iov_len) {
236 /* didn't send whole iov entry... */
237 iov->iov_base = (void *)(vaddr + rc);
249 ksocknal_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
251 struct socket *sock = conn->ksnc_sock;
252 ptl_kiov_t *kiov = tx->tx_kiov;
253 int fragsize = kiov->kiov_len;
254 struct page *page = kiov->kiov_page;
255 int offset = kiov->kiov_offset;
256 int more = (tx->tx_nkiov > 1) ||
257 (!list_empty (&conn->ksnc_tx_queue));
260 /* NB we can't trust socket ops to either consume our iovs
261 * or leave them alone, so we only send 1 frag at a time. */
262 LASSERT (fragsize <= tx->tx_resid);
263 LASSERT (offset + fragsize <= PAGE_SIZE);
264 LASSERT (tx->tx_niov == 0);
265 LASSERT (tx->tx_nkiov > 0);
268 if (fragsize >= ksocknal_tunables.ksnd_zc_min_frag &&
269 (sock->sk->route_caps & NETIF_F_SG) &&
270 (sock->sk->route_caps & (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM | NETIF_F_HW_CSUM))) {
272 CDEBUG(D_NET, "page %p + offset %x for %d\n",
273 page, offset, fragsize);
275 rc = tcp_sendpage_zccd(sock, page, offset, fragsize,
276 more ? (MSG_DONTWAIT | MSG_MORE) : MSG_DONTWAIT,
281 char *addr = ((char *)kmap (page)) + offset;
282 struct iovec fragiov = {.iov_base = addr,
283 .iov_len = fragsize};
284 struct msghdr msg = {
291 .msg_flags = more ? (MSG_DONTWAIT | MSG_MORE) : MSG_DONTWAIT
293 mm_segment_t oldmm = get_fs();
296 rc = sock_sendmsg(sock, &msg, fragsize);
306 kiov->kiov_offset = offset + rc;
307 kiov->kiov_len = fragsize - rc;
318 ksocknal_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
322 if (ksocknal_data.ksnd_stall_tx != 0) {
323 set_current_state (TASK_UNINTERRUPTIBLE);
324 schedule_timeout (ksocknal_data.ksnd_stall_tx * HZ);
327 LASSERT (tx->tx_resid != 0);
329 rc = ksocknal_getconnsock (conn);
331 LASSERT (conn->ksnc_closing);
336 if (ksocknal_data.ksnd_enomem_tx > 0) {
338 ksocknal_data.ksnd_enomem_tx--;
340 } else if (tx->tx_niov != 0) {
341 rc = ksocknal_send_iov (conn, tx);
343 rc = ksocknal_send_kiov (conn, tx);
347 /* Didn't write anything.
349 * NB: rc == 0 and rc == -EAGAIN both mean try
350 * again later (linux stack returns -EAGAIN for
351 * this, but Adaptech TOE returns 0).
353 * Also, sends never fail with -ENOMEM, just
354 * -EAGAIN, but with the added bonus that we can't
355 * expect write_space() to call us back to tell us
356 * when to try sending again. We use the
357 * SOCK_NOSPACE flag to diagnose... */
359 LASSERT(rc != -ENOMEM);
361 if (rc == 0 || rc == -EAGAIN) {
362 if (test_bit(SOCK_NOSPACE,
363 &conn->ksnc_sock->flags)) {
369 if ((counter & (-counter)) == counter)
370 CWARN("%d ENOMEM tx %p\n",
380 /* Consider the connection alive since we managed to chuck
381 * more data into it. Really, we'd like to consider it
382 * alive only when the peer ACKs something, but
383 * write_space() only gets called back while SOCK_NOSPACE
384 * is set. Instead, we presume peer death has occurred if
385 * the socket doesn't drain within a timout */
386 conn->ksnc_tx_deadline = jiffies +
387 ksocknal_tunables.ksnd_io_timeout * HZ;
388 conn->ksnc_peer->ksnp_last_alive = jiffies;
390 } while (tx->tx_resid != 0);
392 ksocknal_putconnsock (conn);
397 ksocknal_eager_ack (ksock_conn_t *conn)
400 mm_segment_t oldmm = get_fs();
401 struct socket *sock = conn->ksnc_sock;
403 /* Remind the socket to ACK eagerly. If I don't, the socket might
404 * think I'm about to send something it could piggy-back the ACK
405 * on, introducing delay in completing zero-copy sends in my
409 sock->ops->setsockopt (sock, SOL_TCP, TCP_QUICKACK,
410 (char *)&opt, sizeof (opt));
415 ksocknal_recv_iov (ksock_conn_t *conn)
417 struct iovec *iov = conn->ksnc_rx_iov;
418 int fragsize = iov->iov_len;
419 unsigned long vaddr = (unsigned long)iov->iov_base;
420 struct iovec fragiov = { .iov_base = (void *)vaddr,
421 .iov_len = fragsize};
422 struct msghdr msg = {
431 mm_segment_t oldmm = get_fs();
434 /* NB we can't trust socket ops to either consume our iovs
435 * or leave them alone, so we only receive 1 frag at a time. */
436 LASSERT (conn->ksnc_rx_niov > 0);
437 LASSERT (fragsize <= conn->ksnc_rx_nob_wanted);
440 rc = sock_recvmsg (conn->ksnc_sock, &msg, fragsize, MSG_DONTWAIT);
441 /* NB this is just a boolean............................^ */
447 /* received something... */
448 conn->ksnc_peer->ksnp_last_alive = jiffies;
449 conn->ksnc_rx_deadline = jiffies +
450 ksocknal_tunables.ksnd_io_timeout * HZ;
451 mb(); /* order with setting rx_started */
452 conn->ksnc_rx_started = 1;
454 conn->ksnc_rx_nob_wanted -= rc;
455 conn->ksnc_rx_nob_left -= rc;
458 iov->iov_base = (void *)(vaddr + rc);
459 iov->iov_len = fragsize - rc;
464 conn->ksnc_rx_niov--;
469 ksocknal_recv_kiov (ksock_conn_t *conn)
471 ptl_kiov_t *kiov = conn->ksnc_rx_kiov;
472 struct page *page = kiov->kiov_page;
473 int offset = kiov->kiov_offset;
474 int fragsize = kiov->kiov_len;
475 unsigned long vaddr = ((unsigned long)kmap (page)) + offset;
476 struct iovec fragiov = { .iov_base = (void *)vaddr,
477 .iov_len = fragsize};
478 struct msghdr msg = {
487 mm_segment_t oldmm = get_fs();
490 /* NB we can't trust socket ops to either consume our iovs
491 * or leave them alone, so we only receive 1 frag at a time. */
492 LASSERT (fragsize <= conn->ksnc_rx_nob_wanted);
493 LASSERT (conn->ksnc_rx_nkiov > 0);
494 LASSERT (offset + fragsize <= PAGE_SIZE);
497 rc = sock_recvmsg (conn->ksnc_sock, &msg, fragsize, MSG_DONTWAIT);
498 /* NB this is just a boolean............................^ */
506 /* received something... */
507 conn->ksnc_peer->ksnp_last_alive = jiffies;
508 conn->ksnc_rx_deadline = jiffies +
509 ksocknal_tunables.ksnd_io_timeout * HZ;
510 mb(); /* order with setting rx_started */
511 conn->ksnc_rx_started = 1;
513 conn->ksnc_rx_nob_wanted -= rc;
514 conn->ksnc_rx_nob_left -= rc;
517 kiov->kiov_offset = offset + rc;
518 kiov->kiov_len = fragsize - rc;
522 conn->ksnc_rx_kiov++;
523 conn->ksnc_rx_nkiov--;
528 ksocknal_receive (ksock_conn_t *conn)
530 /* Return 1 on success, 0 on EOF, < 0 on error.
531 * Caller checks ksnc_rx_nob_wanted to determine
532 * progress/completion. */
536 if (ksocknal_data.ksnd_stall_rx != 0) {
537 set_current_state (TASK_UNINTERRUPTIBLE);
538 schedule_timeout (ksocknal_data.ksnd_stall_rx * HZ);
541 rc = ksocknal_getconnsock (conn);
543 LASSERT (conn->ksnc_closing);
548 if (conn->ksnc_rx_niov != 0)
549 rc = ksocknal_recv_iov (conn);
551 rc = ksocknal_recv_kiov (conn);
554 /* error/EOF or partial receive */
557 } else if (rc == 0 && conn->ksnc_rx_started) {
558 /* EOF in the middle of a message */
564 /* Completed a fragment */
566 if (conn->ksnc_rx_nob_wanted == 0) {
567 /* Completed a message segment (header or payload) */
568 if ((ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0 &&
569 (conn->ksnc_rx_state == SOCKNAL_RX_BODY ||
570 conn->ksnc_rx_state == SOCKNAL_RX_BODY_FWD)) {
571 /* Remind the socket to ack eagerly... */
572 ksocknal_eager_ack(conn);
579 ksocknal_putconnsock (conn);
585 ksocknal_zc_callback (zccd_t *zcd)
587 ksock_tx_t *tx = KSOCK_ZCCD_2_TX(zcd);
588 ksock_sched_t *sched = tx->tx_conn->ksnc_scheduler;
592 /* Schedule tx for cleanup (can't do it now due to lock conflicts) */
594 spin_lock_irqsave (&sched->kss_lock, flags);
596 list_add_tail (&tx->tx_list, &sched->kss_zctxdone_list);
597 wake_up (&sched->kss_waitq);
599 spin_unlock_irqrestore (&sched->kss_lock, flags);
605 ksocknal_tx_done (ksock_tx_t *tx, int asynch)
610 if (tx->tx_conn != NULL) {
611 /* This tx got queued on a conn; do the accounting... */
612 atomic_sub (tx->tx_nob, &tx->tx_conn->ksnc_tx_nob);
614 /* zero copy completion isn't always from
615 * process_transmit() so it needs to keep a ref on
618 ksocknal_put_conn (tx->tx_conn);
624 if (tx->tx_isfwd) { /* was a forwarded packet? */
625 kpr_fwd_done (&ksocknal_data.ksnd_router,
626 KSOCK_TX_2_KPR_FWD_DESC (tx),
627 (tx->tx_resid == 0) ? 0 : -ECONNABORTED);
633 ltx = KSOCK_TX_2_KSOCK_LTX (tx);
635 lib_finalize (&ksocknal_lib, ltx->ltx_private, ltx->ltx_cookie,
636 (tx->tx_resid == 0) ? PTL_OK : PTL_FAIL);
638 ksocknal_free_ltx (ltx);
643 ksocknal_tx_launched (ksock_tx_t *tx)
646 if (atomic_read (&tx->tx_zccd.zccd_count) != 1) {
647 ksock_conn_t *conn = tx->tx_conn;
649 /* zccd skbufs are still in-flight. First take a ref on
650 * conn, so it hangs about for ksocknal_tx_done... */
651 atomic_inc (&conn->ksnc_refcount);
653 /* ...then drop the initial ref on zccd, so the zero copy
654 * callback can occur */
655 zccd_put (&tx->tx_zccd);
659 /* Any zero-copy-ness (if any) has completed; I can complete the
660 * transmit now, avoiding an extra schedule */
661 ksocknal_tx_done (tx, 0);
665 ksocknal_process_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
670 rc = ksocknal_transmit (conn, tx);
672 CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc);
674 if (tx->tx_resid == 0) {
675 /* Sent everything OK */
678 ksocknal_tx_launched (tx);
686 /* Queue on ksnd_enomem_conns for retry after a timeout */
687 spin_lock_irqsave(&ksocknal_data.ksnd_reaper_lock, flags);
689 /* enomem list takes over scheduler's ref... */
690 LASSERT (conn->ksnc_tx_scheduled);
691 list_add_tail(&conn->ksnc_tx_list,
692 &ksocknal_data.ksnd_enomem_conns);
693 if (!time_after_eq(jiffies + SOCKNAL_ENOMEM_RETRY,
694 ksocknal_data.ksnd_reaper_waketime))
695 wake_up (&ksocknal_data.ksnd_reaper_waitq);
697 spin_unlock_irqrestore(&ksocknal_data.ksnd_reaper_lock, flags);
704 if (!conn->ksnc_closing)
705 CERROR("[%p] Error %d on write to "LPX64
706 " ip %d.%d.%d.%d:%d\n", conn, rc,
707 conn->ksnc_peer->ksnp_nid,
708 HIPQUAD(conn->ksnc_ipaddr),
711 ksocknal_close_conn_and_siblings (conn, rc);
712 ksocknal_tx_launched (tx);
718 ksocknal_launch_autoconnect_locked (ksock_route_t *route)
722 /* called holding write lock on ksnd_global_lock */
724 LASSERT (!route->ksnr_deleted);
725 LASSERT ((route->ksnr_connected & (1 << SOCKNAL_CONN_ANY)) == 0);
726 LASSERT ((route->ksnr_connected & KSNR_TYPED_ROUTES) != KSNR_TYPED_ROUTES);
727 LASSERT (!route->ksnr_connecting);
729 if (ksocknal_tunables.ksnd_typed_conns)
730 route->ksnr_connecting =
731 KSNR_TYPED_ROUTES & ~route->ksnr_connected;
733 route->ksnr_connecting = (1 << SOCKNAL_CONN_ANY);
735 atomic_inc (&route->ksnr_refcount); /* extra ref for asynchd */
737 spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
739 list_add_tail (&route->ksnr_connect_list,
740 &ksocknal_data.ksnd_autoconnectd_routes);
741 wake_up (&ksocknal_data.ksnd_autoconnectd_waitq);
743 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
747 ksocknal_find_target_peer_locked (ksock_tx_t *tx, ptl_nid_t nid)
749 char ipbuf[PTL_NALFMT_SIZE];
750 ptl_nid_t target_nid;
752 ksock_peer_t *peer = ksocknal_find_peer_locked (nid);
758 CERROR ("Can't send packet to "LPX64
759 " %s: routed target is not a peer\n",
760 nid, portals_nid2str(SOCKNAL, nid, ipbuf));
764 rc = kpr_lookup (&ksocknal_data.ksnd_router, nid, tx->tx_nob,
767 CERROR ("Can't route to "LPX64" %s: router error %d\n",
768 nid, portals_nid2str(SOCKNAL, nid, ipbuf), rc);
772 peer = ksocknal_find_peer_locked (target_nid);
776 CERROR ("Can't send packet to "LPX64" %s: no peer entry\n",
777 target_nid, portals_nid2str(SOCKNAL, target_nid, ipbuf));
782 ksocknal_find_conn_locked (ksock_tx_t *tx, ksock_peer_t *peer)
784 struct list_head *tmp;
785 ksock_conn_t *typed = NULL;
787 ksock_conn_t *fallback = NULL;
790 /* Find the conn with the shortest tx queue */
791 list_for_each (tmp, &peer->ksnp_conns) {
792 ksock_conn_t *c = list_entry(tmp, ksock_conn_t, ksnc_list);
793 int nob = atomic_read(&c->ksnc_tx_nob) +
794 c->ksnc_sock->sk->sk_wmem_queued;
796 LASSERT (!c->ksnc_closing);
798 if (fallback == NULL || nob < fnob) {
803 if (!ksocknal_tunables.ksnd_typed_conns)
806 switch (c->ksnc_type) {
809 case SOCKNAL_CONN_ANY:
811 case SOCKNAL_CONN_BULK_IN:
813 case SOCKNAL_CONN_BULK_OUT:
814 if (tx->tx_nob < ksocknal_tunables.ksnd_min_bulk)
817 case SOCKNAL_CONN_CONTROL:
818 if (tx->tx_nob >= ksocknal_tunables.ksnd_min_bulk)
823 if (typed == NULL || nob < tnob) {
829 /* prefer the typed selection */
830 return ((typed != NULL) ? typed : fallback);
834 ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn)
837 ksock_sched_t *sched = conn->ksnc_scheduler;
839 /* called holding global lock (read or irq-write) and caller may
840 * not have dropped this lock between finding conn and calling me,
841 * so we don't need the {get,put}connsock dance to deref
843 LASSERT(!conn->ksnc_closing);
844 LASSERT(tx->tx_resid == tx->tx_nob);
846 CDEBUG (D_NET, "Sending to "LPX64" ip %d.%d.%d.%d:%d\n",
847 conn->ksnc_peer->ksnp_nid,
848 HIPQUAD(conn->ksnc_ipaddr),
851 atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
855 zccd_init (&tx->tx_zccd, ksocknal_zc_callback);
856 /* NB this sets 1 ref on zccd, so the callback can only occur after
857 * I've released this ref. */
859 spin_lock_irqsave (&sched->kss_lock, flags);
861 conn->ksnc_tx_deadline = jiffies +
862 ksocknal_tunables.ksnd_io_timeout * HZ;
863 mb(); /* order with list_add_tail */
865 list_add_tail (&tx->tx_list, &conn->ksnc_tx_queue);
867 if (conn->ksnc_tx_ready && /* able to send */
868 !conn->ksnc_tx_scheduled) { /* not scheduled to send */
869 /* +1 ref for scheduler */
870 atomic_inc (&conn->ksnc_refcount);
871 list_add_tail (&conn->ksnc_tx_list,
872 &sched->kss_tx_conns);
873 conn->ksnc_tx_scheduled = 1;
874 wake_up (&sched->kss_waitq);
877 spin_unlock_irqrestore (&sched->kss_lock, flags);
881 ksocknal_find_connectable_route_locked (ksock_peer_t *peer)
883 struct list_head *tmp;
884 ksock_route_t *route;
885 ksock_route_t *candidate = NULL;
889 list_for_each (tmp, &peer->ksnp_routes) {
890 route = list_entry (tmp, ksock_route_t, ksnr_list);
891 bits = route->ksnr_connected;
893 if ((bits & KSNR_TYPED_ROUTES) == KSNR_TYPED_ROUTES ||
894 (bits & (1 << SOCKNAL_CONN_ANY)) != 0 ||
895 route->ksnr_connecting != 0) {
896 /* All typed connections have been established, or
897 * an untyped connection has been established, or
898 * connections are currently being established */
903 /* too soon to retry this guy? */
904 if (!time_after_eq (jiffies, route->ksnr_timeout))
907 /* always do eager routes */
908 if (route->ksnr_eager)
911 if (candidate == NULL) {
912 /* If we don't find any other route that is fully
913 * connected or connecting, the first connectable
914 * route is returned. If it fails to connect, it
915 * will get placed at the end of the list */
920 return (found ? NULL : candidate);
924 ksocknal_find_connecting_route_locked (ksock_peer_t *peer)
926 struct list_head *tmp;
927 ksock_route_t *route;
929 list_for_each (tmp, &peer->ksnp_routes) {
930 route = list_entry (tmp, ksock_route_t, ksnr_list);
932 if (route->ksnr_connecting != 0)
940 ksocknal_launch_packet (ksock_tx_t *tx, ptl_nid_t nid)
945 ksock_route_t *route;
948 /* Ensure the frags we've been given EXACTLY match the number of
949 * bytes we want to send. Many TCP/IP stacks disregard any total
950 * size parameters passed to them and just look at the frags.
952 * We always expect at least 1 mapped fragment containing the
953 * complete portals header. */
954 LASSERT (lib_iov_nob (tx->tx_niov, tx->tx_iov) +
955 lib_kiov_nob (tx->tx_nkiov, tx->tx_kiov) == tx->tx_nob);
956 LASSERT (tx->tx_niov >= 1);
957 LASSERT (tx->tx_iov[0].iov_len >= sizeof (ptl_hdr_t));
959 CDEBUG (D_NET, "packet %p type %d, nob %d niov %d nkiov %d\n",
960 tx, ((ptl_hdr_t *)tx->tx_iov[0].iov_base)->type,
961 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
963 tx->tx_conn = NULL; /* only set when assigned a conn */
964 tx->tx_resid = tx->tx_nob;
965 tx->tx_hdr = (ptl_hdr_t *)tx->tx_iov[0].iov_base;
967 g_lock = &ksocknal_data.ksnd_global_lock;
970 peer = ksocknal_find_target_peer_locked (tx, nid);
972 read_unlock (g_lock);
973 return (-EHOSTUNREACH);
976 if (ksocknal_find_connectable_route_locked(peer) == NULL) {
977 conn = ksocknal_find_conn_locked (tx, peer);
979 /* I've got no autoconnect routes that need to be
980 * connecting and I do have an actual connection... */
981 ksocknal_queue_tx_locked (tx, conn);
982 read_unlock (g_lock);
987 /* Making one or more connections; I'll need a write lock... */
989 atomic_inc (&peer->ksnp_refcount); /* +1 ref for me while I unlock */
990 read_unlock (g_lock);
991 write_lock_irqsave (g_lock, flags);
993 if (peer->ksnp_closing) { /* peer deleted as I blocked! */
994 write_unlock_irqrestore (g_lock, flags);
995 ksocknal_put_peer (peer);
996 return (-EHOSTUNREACH);
998 ksocknal_put_peer (peer); /* drop ref I got above */
1001 /* launch any/all autoconnections that need it */
1002 route = ksocknal_find_connectable_route_locked (peer);
1006 ksocknal_launch_autoconnect_locked (route);
1009 conn = ksocknal_find_conn_locked (tx, peer);
1011 /* Connection exists; queue message on it */
1012 ksocknal_queue_tx_locked (tx, conn);
1013 write_unlock_irqrestore (g_lock, flags);
1017 route = ksocknal_find_connecting_route_locked (peer);
1018 if (route != NULL) {
1019 /* At least 1 connection is being established; queue the
1021 list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
1022 write_unlock_irqrestore (g_lock, flags);
1026 write_unlock_irqrestore (g_lock, flags);
1027 return (-EHOSTUNREACH);
1031 ksocknal_sendmsg(nal_cb_t *nal,
1038 unsigned int payload_niov,
1039 struct iovec *payload_iov,
1040 ptl_kiov_t *payload_kiov,
1041 size_t payload_offset,
1048 /* NB 'private' is different depending on what we're sending.
1049 * Just ignore it... */
1051 CDEBUG(D_NET, "sending "LPSZ" bytes in %d frags to nid:"LPX64
1052 " pid %d\n", payload_nob, payload_niov, nid , pid);
1054 LASSERT (payload_nob == 0 || payload_niov > 0);
1055 LASSERT (payload_niov <= PTL_MD_MAX_IOV);
1057 /* It must be OK to kmap() if required */
1058 LASSERT (payload_kiov == NULL || !in_interrupt ());
1059 /* payload is either all vaddrs or all pages */
1060 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1062 if (payload_iov != NULL)
1063 desc_size = offsetof(ksock_ltx_t, ltx_iov[1 + payload_niov]);
1065 desc_size = offsetof(ksock_ltx_t, ltx_kiov[payload_niov]);
1067 if (in_interrupt() ||
1068 type == PTL_MSG_ACK ||
1069 type == PTL_MSG_REPLY) {
1070 /* Can't block if in interrupt or responding to an incoming
1072 PORTAL_ALLOC_ATOMIC(ltx, desc_size);
1074 PORTAL_ALLOC(ltx, desc_size);
1078 CERROR("Can't allocate tx desc type %d size %d %s\n",
1079 type, desc_size, in_interrupt() ? "(intr)" : "");
1080 return (PTL_NO_SPACE);
1083 atomic_inc(&ksocknal_data.ksnd_nactive_ltxs);
1085 ltx->ltx_desc_size = desc_size;
1087 /* We always have 1 mapped frag for the header */
1088 ltx->ltx_tx.tx_iov = ltx->ltx_iov;
1089 ltx->ltx_iov[0].iov_base = <x->ltx_hdr;
1090 ltx->ltx_iov[0].iov_len = sizeof(*hdr);
1091 ltx->ltx_hdr = *hdr;
1093 ltx->ltx_private = private;
1094 ltx->ltx_cookie = cookie;
1096 ltx->ltx_tx.tx_isfwd = 0;
1097 ltx->ltx_tx.tx_nob = sizeof (*hdr) + payload_nob;
1099 if (payload_iov != NULL) {
1100 /* payload is all mapped */
1101 ltx->ltx_tx.tx_kiov = NULL;
1102 ltx->ltx_tx.tx_nkiov = 0;
1104 ltx->ltx_tx.tx_niov =
1105 1 + lib_extract_iov(payload_niov, <x->ltx_iov[1],
1106 payload_niov, payload_iov,
1107 payload_offset, payload_nob);
1109 /* payload is all pages */
1110 ltx->ltx_tx.tx_niov = 1;
1112 ltx->ltx_tx.tx_kiov = ltx->ltx_kiov;
1113 ltx->ltx_tx.tx_nkiov =
1114 lib_extract_kiov(payload_niov, ltx->ltx_kiov,
1115 payload_niov, payload_kiov,
1116 payload_offset, payload_nob);
1119 rc = ksocknal_launch_packet(<x->ltx_tx, nid);
1123 ksocknal_free_ltx(ltx);
1128 ksocknal_send (nal_cb_t *nal, void *private, lib_msg_t *cookie,
1129 ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid,
1130 unsigned int payload_niov, struct iovec *payload_iov,
1131 size_t payload_offset, size_t payload_len)
1133 return (ksocknal_sendmsg(nal, private, cookie,
1134 hdr, type, nid, pid,
1135 payload_niov, payload_iov, NULL,
1136 payload_offset, payload_len));
1140 ksocknal_send_pages (nal_cb_t *nal, void *private, lib_msg_t *cookie,
1141 ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid,
1142 unsigned int payload_niov, ptl_kiov_t *payload_kiov,
1143 size_t payload_offset, size_t payload_len)
1145 return (ksocknal_sendmsg(nal, private, cookie,
1146 hdr, type, nid, pid,
1147 payload_niov, NULL, payload_kiov,
1148 payload_offset, payload_len));
1152 ksocknal_fwd_packet (void *arg, kpr_fwd_desc_t *fwd)
1154 ptl_nid_t nid = fwd->kprfd_gateway_nid;
1155 ksock_ftx_t *ftx = (ksock_ftx_t *)&fwd->kprfd_scratch;
1158 CDEBUG (D_NET, "Forwarding [%p] -> "LPX64" ("LPX64"))\n", fwd,
1159 fwd->kprfd_gateway_nid, fwd->kprfd_target_nid);
1161 /* I'm the gateway; must be the last hop */
1162 if (nid == ksocknal_lib.ni.nid)
1163 nid = fwd->kprfd_target_nid;
1165 /* setup iov for hdr */
1166 ftx->ftx_iov.iov_base = fwd->kprfd_hdr;
1167 ftx->ftx_iov.iov_len = sizeof(ptl_hdr_t);
1169 ftx->ftx_tx.tx_isfwd = 1; /* This is a forwarding packet */
1170 ftx->ftx_tx.tx_nob = sizeof(ptl_hdr_t) + fwd->kprfd_nob;
1171 ftx->ftx_tx.tx_niov = 1;
1172 ftx->ftx_tx.tx_iov = &ftx->ftx_iov;
1173 ftx->ftx_tx.tx_nkiov = fwd->kprfd_niov;
1174 ftx->ftx_tx.tx_kiov = fwd->kprfd_kiov;
1176 rc = ksocknal_launch_packet (&ftx->ftx_tx, nid);
1178 kpr_fwd_done (&ksocknal_data.ksnd_router, fwd, rc);
1182 ksocknal_thread_start (int (*fn)(void *arg), void *arg)
1184 long pid = kernel_thread (fn, arg, 0);
1189 atomic_inc (&ksocknal_data.ksnd_nthreads);
1194 ksocknal_thread_fini (void)
1196 atomic_dec (&ksocknal_data.ksnd_nthreads);
1200 ksocknal_fmb_callback (void *arg, int error)
1202 ksock_fmb_t *fmb = (ksock_fmb_t *)arg;
1203 ksock_fmb_pool_t *fmp = fmb->fmb_pool;
1204 ptl_hdr_t *hdr = (ptl_hdr_t *)page_address(fmb->fmb_kiov[0].kiov_page);
1205 ksock_conn_t *conn = NULL;
1206 ksock_sched_t *sched;
1207 unsigned long flags;
1208 char ipbuf[PTL_NALFMT_SIZE];
1209 char ipbuf2[PTL_NALFMT_SIZE];
1212 CERROR("Failed to route packet from "
1213 LPX64" %s to "LPX64" %s: %d\n",
1214 NTOH__u64(hdr->src_nid),
1215 portals_nid2str(SOCKNAL, NTOH__u64(hdr->src_nid), ipbuf),
1216 NTOH__u64(hdr->dest_nid),
1217 portals_nid2str(SOCKNAL, NTOH__u64(hdr->dest_nid), ipbuf2),
1220 CDEBUG (D_NET, "routed packet from "LPX64" to "LPX64": OK\n",
1221 NTOH__u64 (hdr->src_nid), NTOH__u64 (hdr->dest_nid));
1223 /* drop peer ref taken on init */
1224 ksocknal_put_peer (fmb->fmb_peer);
1226 spin_lock_irqsave (&fmp->fmp_lock, flags);
1228 list_add (&fmb->fmb_list, &fmp->fmp_idle_fmbs);
1229 fmp->fmp_nactive_fmbs--;
1231 if (!list_empty (&fmp->fmp_blocked_conns)) {
1232 conn = list_entry (fmb->fmb_pool->fmp_blocked_conns.next,
1233 ksock_conn_t, ksnc_rx_list);
1234 list_del (&conn->ksnc_rx_list);
1237 spin_unlock_irqrestore (&fmp->fmp_lock, flags);
1242 CDEBUG (D_NET, "Scheduling conn %p\n", conn);
1243 LASSERT (conn->ksnc_rx_scheduled);
1244 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_FMB_SLEEP);
1246 conn->ksnc_rx_state = SOCKNAL_RX_GET_FMB;
1248 sched = conn->ksnc_scheduler;
1250 spin_lock_irqsave (&sched->kss_lock, flags);
1252 list_add_tail (&conn->ksnc_rx_list, &sched->kss_rx_conns);
1253 wake_up (&sched->kss_waitq);
1255 spin_unlock_irqrestore (&sched->kss_lock, flags);
1259 ksocknal_get_idle_fmb (ksock_conn_t *conn)
1261 int payload_nob = conn->ksnc_rx_nob_left;
1262 unsigned long flags;
1263 ksock_fmb_pool_t *pool;
1266 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB);
1267 LASSERT (kpr_routing(&ksocknal_data.ksnd_router));
1269 if (payload_nob <= SOCKNAL_SMALL_FWD_PAGES * PAGE_SIZE)
1270 pool = &ksocknal_data.ksnd_small_fmp;
1272 pool = &ksocknal_data.ksnd_large_fmp;
1274 spin_lock_irqsave (&pool->fmp_lock, flags);
1276 if (!list_empty (&pool->fmp_idle_fmbs)) {
1277 fmb = list_entry(pool->fmp_idle_fmbs.next,
1278 ksock_fmb_t, fmb_list);
1279 list_del (&fmb->fmb_list);
1280 pool->fmp_nactive_fmbs++;
1281 spin_unlock_irqrestore (&pool->fmp_lock, flags);
1286 /* deschedule until fmb free */
1288 conn->ksnc_rx_state = SOCKNAL_RX_FMB_SLEEP;
1290 list_add_tail (&conn->ksnc_rx_list,
1291 &pool->fmp_blocked_conns);
1293 spin_unlock_irqrestore (&pool->fmp_lock, flags);
1298 ksocknal_init_fmb (ksock_conn_t *conn, ksock_fmb_t *fmb)
1300 int payload_nob = conn->ksnc_rx_nob_left;
1301 ptl_nid_t dest_nid = NTOH__u64 (conn->ksnc_hdr.dest_nid);
1303 int nob = payload_nob;
1305 LASSERT (conn->ksnc_rx_scheduled);
1306 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB);
1307 LASSERT (conn->ksnc_rx_nob_wanted == conn->ksnc_rx_nob_left);
1308 LASSERT (payload_nob >= 0);
1309 LASSERT (payload_nob <= fmb->fmb_pool->fmp_buff_pages * PAGE_SIZE);
1310 LASSERT (sizeof (ptl_hdr_t) < PAGE_SIZE);
1311 LASSERT (fmb->fmb_kiov[0].kiov_offset == 0);
1313 /* Take a ref on the conn's peer to prevent module unload before
1314 * forwarding completes. */
1315 fmb->fmb_peer = conn->ksnc_peer;
1316 atomic_inc (&conn->ksnc_peer->ksnp_refcount);
1318 /* Copy the header we just read into the forwarding buffer. If
1319 * there's payload, start reading reading it into the buffer,
1320 * otherwise the forwarding buffer can be kicked off
1322 fmb->fmb_hdr = conn->ksnc_hdr;
1325 LASSERT (niov < fmb->fmb_pool->fmp_buff_pages);
1326 LASSERT (fmb->fmb_kiov[niov].kiov_offset == 0);
1327 fmb->fmb_kiov[niov].kiov_len = MIN (PAGE_SIZE, nob);
1332 kpr_fwd_init(&fmb->fmb_fwd, dest_nid, &fmb->fmb_hdr,
1333 payload_nob, niov, fmb->fmb_kiov,
1334 ksocknal_fmb_callback, fmb);
1336 if (payload_nob == 0) { /* got complete packet already */
1337 CDEBUG (D_NET, "%p "LPX64"->"LPX64" fwd_start (immediate)\n",
1338 conn, NTOH__u64 (conn->ksnc_hdr.src_nid), dest_nid);
1340 kpr_fwd_start (&ksocknal_data.ksnd_router, &fmb->fmb_fwd);
1342 ksocknal_new_packet (conn, 0); /* on to next packet */
1346 conn->ksnc_cookie = fmb; /* stash fmb for later */
1347 conn->ksnc_rx_state = SOCKNAL_RX_BODY_FWD; /* read in the payload */
1349 /* Set up conn->ksnc_rx_kiov to read the payload into fmb's kiov-ed
1351 LASSERT (niov <= sizeof(conn->ksnc_rx_iov_space)/sizeof(ptl_kiov_t));
1353 conn->ksnc_rx_niov = 0;
1354 conn->ksnc_rx_nkiov = niov;
1355 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1356 memcpy(conn->ksnc_rx_kiov, fmb->fmb_kiov, niov * sizeof(ptl_kiov_t));
1358 CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d reading body\n", conn,
1359 NTOH__u64 (conn->ksnc_hdr.src_nid), dest_nid, payload_nob);
1364 ksocknal_fwd_parse (ksock_conn_t *conn)
1367 ptl_nid_t dest_nid = NTOH__u64 (conn->ksnc_hdr.dest_nid);
1368 ptl_nid_t src_nid = NTOH__u64 (conn->ksnc_hdr.src_nid);
1369 int body_len = NTOH__u32 (conn->ksnc_hdr.payload_length);
1370 char str[PTL_NALFMT_SIZE];
1371 char str2[PTL_NALFMT_SIZE];
1373 CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d parsing header\n", conn,
1374 src_nid, dest_nid, conn->ksnc_rx_nob_left);
1376 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_HEADER);
1377 LASSERT (conn->ksnc_rx_scheduled);
1379 if (body_len < 0) { /* length corrupt (overflow) */
1380 CERROR("dropping packet from "LPX64" (%s) for "LPX64" (%s): "
1381 "packet size %d illegal\n",
1382 src_nid, portals_nid2str(TCPNAL, src_nid, str),
1383 dest_nid, portals_nid2str(TCPNAL, dest_nid, str2),
1386 ksocknal_new_packet (conn, 0); /* on to new packet */
1390 if (!kpr_routing(&ksocknal_data.ksnd_router)) { /* not forwarding */
1391 CERROR("dropping packet from "LPX64" (%s) for "LPX64
1392 " (%s): not forwarding\n",
1393 src_nid, portals_nid2str(TCPNAL, src_nid, str),
1394 dest_nid, portals_nid2str(TCPNAL, dest_nid, str2));
1395 /* on to new packet (skip this one's body) */
1396 ksocknal_new_packet (conn, body_len);
1400 if (body_len > PTL_MTU) { /* too big to forward */
1401 CERROR ("dropping packet from "LPX64" (%s) for "LPX64
1402 "(%s): packet size %d too big\n",
1403 src_nid, portals_nid2str(TCPNAL, src_nid, str),
1404 dest_nid, portals_nid2str(TCPNAL, dest_nid, str2),
1406 /* on to new packet (skip this one's body) */
1407 ksocknal_new_packet (conn, body_len);
1411 /* should have gone direct */
1412 peer = ksocknal_get_peer (conn->ksnc_hdr.dest_nid);
1414 CERROR ("dropping packet from "LPX64" (%s) for "LPX64
1415 "(%s): target is a peer\n",
1416 src_nid, portals_nid2str(TCPNAL, src_nid, str),
1417 dest_nid, portals_nid2str(TCPNAL, dest_nid, str2));
1418 ksocknal_put_peer (peer); /* drop ref from get above */
1420 /* on to next packet (skip this one's body) */
1421 ksocknal_new_packet (conn, body_len);
1425 conn->ksnc_rx_state = SOCKNAL_RX_GET_FMB; /* Getting FMB now */
1426 conn->ksnc_rx_nob_left = body_len; /* stash packet size */
1427 conn->ksnc_rx_nob_wanted = body_len; /* (no slop) */
1431 ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip)
1433 static char ksocknal_slop_buffer[4096];
1439 if (nob_to_skip == 0) { /* right at next packet boundary now */
1440 conn->ksnc_rx_started = 0;
1441 mb (); /* racing with timeout thread */
1443 conn->ksnc_rx_state = SOCKNAL_RX_HEADER;
1444 conn->ksnc_rx_nob_wanted = sizeof (ptl_hdr_t);
1445 conn->ksnc_rx_nob_left = sizeof (ptl_hdr_t);
1447 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1448 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_hdr;
1449 conn->ksnc_rx_iov[0].iov_len = sizeof (ptl_hdr_t);
1450 conn->ksnc_rx_niov = 1;
1452 conn->ksnc_rx_kiov = NULL;
1453 conn->ksnc_rx_nkiov = 0;
1457 /* Set up to skip as much a possible now. If there's more left
1458 * (ran out of iov entries) we'll get called again */
1460 conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1461 conn->ksnc_rx_nob_left = nob_to_skip;
1462 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1467 nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer));
1469 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1470 conn->ksnc_rx_iov[niov].iov_len = nob;
1475 } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */
1476 niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec));
1478 conn->ksnc_rx_niov = niov;
1479 conn->ksnc_rx_kiov = NULL;
1480 conn->ksnc_rx_nkiov = 0;
1481 conn->ksnc_rx_nob_wanted = skipped;
1486 ksocknal_process_receive (ksock_conn_t *conn)
1491 LASSERT (atomic_read (&conn->ksnc_refcount) > 0);
1493 /* doesn't need a forwarding buffer */
1494 if (conn->ksnc_rx_state != SOCKNAL_RX_GET_FMB)
1498 fmb = ksocknal_get_idle_fmb (conn);
1500 /* conn descheduled waiting for idle fmb */
1504 if (ksocknal_init_fmb (conn, fmb)) {
1505 /* packet forwarded */
1510 /* NB: sched lock NOT held */
1511 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_HEADER ||
1512 conn->ksnc_rx_state == SOCKNAL_RX_BODY ||
1513 conn->ksnc_rx_state == SOCKNAL_RX_BODY_FWD ||
1514 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1516 LASSERT (conn->ksnc_rx_nob_wanted > 0);
1518 rc = ksocknal_receive(conn);
1521 LASSERT (rc != -EAGAIN);
1524 CWARN ("[%p] EOF from "LPX64" ip %d.%d.%d.%d:%d\n",
1525 conn, conn->ksnc_peer->ksnp_nid,
1526 HIPQUAD(conn->ksnc_ipaddr),
1528 else if (!conn->ksnc_closing)
1529 CERROR ("[%p] Error %d on read from "LPX64
1530 " ip %d.%d.%d.%d:%d\n",
1531 conn, rc, conn->ksnc_peer->ksnp_nid,
1532 HIPQUAD(conn->ksnc_ipaddr),
1535 ksocknal_close_conn_and_siblings (conn, rc);
1536 return (rc == 0 ? -ESHUTDOWN : rc);
1539 if (conn->ksnc_rx_nob_wanted != 0) {
1544 switch (conn->ksnc_rx_state) {
1545 case SOCKNAL_RX_HEADER:
1546 if (conn->ksnc_hdr.type != HTON__u32(PTL_MSG_HELLO) &&
1547 NTOH__u64(conn->ksnc_hdr.dest_nid) != ksocknal_lib.ni.nid) {
1548 /* This packet isn't for me */
1549 ksocknal_fwd_parse (conn);
1550 switch (conn->ksnc_rx_state) {
1551 case SOCKNAL_RX_HEADER: /* skipped (zero payload) */
1552 return (0); /* => come back later */
1553 case SOCKNAL_RX_SLOP: /* skipping packet's body */
1554 goto try_read; /* => go read it */
1555 case SOCKNAL_RX_GET_FMB: /* forwarding */
1556 goto get_fmb; /* => go get a fwd msg buffer */
1563 /* sets wanted_len, iovs etc */
1564 lib_parse(&ksocknal_lib, &conn->ksnc_hdr, conn);
1566 if (conn->ksnc_rx_nob_wanted != 0) { /* need to get payload? */
1567 conn->ksnc_rx_state = SOCKNAL_RX_BODY;
1568 goto try_read; /* go read the payload */
1570 /* Fall through (completed packet for me) */
1572 case SOCKNAL_RX_BODY:
1573 /* payload all received */
1574 lib_finalize(&ksocknal_lib, NULL, conn->ksnc_cookie, PTL_OK);
1577 case SOCKNAL_RX_SLOP:
1578 /* starting new packet? */
1579 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1580 return (0); /* come back later */
1581 goto try_read; /* try to finish reading slop now */
1583 case SOCKNAL_RX_BODY_FWD:
1584 /* payload all received */
1585 CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d fwd_start (got body)\n",
1586 conn, NTOH__u64 (conn->ksnc_hdr.src_nid),
1587 NTOH__u64 (conn->ksnc_hdr.dest_nid),
1588 conn->ksnc_rx_nob_left);
1590 /* forward the packet. NB ksocknal_init_fmb() put fmb into
1591 * conn->ksnc_cookie */
1592 fmb = (ksock_fmb_t *)conn->ksnc_cookie;
1593 kpr_fwd_start (&ksocknal_data.ksnd_router, &fmb->fmb_fwd);
1595 /* no slop in forwarded packets */
1596 LASSERT (conn->ksnc_rx_nob_left == 0);
1598 ksocknal_new_packet (conn, 0); /* on to next packet */
1599 return (0); /* (later) */
1607 return (-EINVAL); /* keep gcc happy */
1611 ksocknal_recv (nal_cb_t *nal, void *private, lib_msg_t *msg,
1612 unsigned int niov, struct iovec *iov,
1613 size_t offset, size_t mlen, size_t rlen)
1615 ksock_conn_t *conn = (ksock_conn_t *)private;
1617 LASSERT (mlen <= rlen);
1618 LASSERT (niov <= PTL_MD_MAX_IOV);
1620 conn->ksnc_cookie = msg;
1621 conn->ksnc_rx_nob_wanted = mlen;
1622 conn->ksnc_rx_nob_left = rlen;
1624 conn->ksnc_rx_nkiov = 0;
1625 conn->ksnc_rx_kiov = NULL;
1626 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1627 conn->ksnc_rx_niov =
1628 lib_extract_iov(PTL_MD_MAX_IOV, conn->ksnc_rx_iov,
1629 niov, iov, offset, mlen);
1632 lib_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1633 lib_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1639 ksocknal_recv_pages (nal_cb_t *nal, void *private, lib_msg_t *msg,
1640 unsigned int niov, ptl_kiov_t *kiov,
1641 size_t offset, size_t mlen, size_t rlen)
1643 ksock_conn_t *conn = (ksock_conn_t *)private;
1645 LASSERT (mlen <= rlen);
1646 LASSERT (niov <= PTL_MD_MAX_IOV);
1648 conn->ksnc_cookie = msg;
1649 conn->ksnc_rx_nob_wanted = mlen;
1650 conn->ksnc_rx_nob_left = rlen;
1652 conn->ksnc_rx_niov = 0;
1653 conn->ksnc_rx_iov = NULL;
1654 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1655 conn->ksnc_rx_nkiov =
1656 lib_extract_kiov(PTL_MD_MAX_IOV, conn->ksnc_rx_kiov,
1657 niov, kiov, offset, mlen);
1660 lib_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1661 lib_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1666 int ksocknal_scheduler (void *arg)
1668 ksock_sched_t *sched = (ksock_sched_t *)arg;
1671 unsigned long flags;
1674 int id = sched - ksocknal_data.ksnd_schedulers;
1677 snprintf (name, sizeof (name),"ksocknald_%02d", id);
1678 kportal_daemonize (name);
1679 kportal_blockallsigs ();
1681 #if (CONFIG_SMP && CPU_AFFINITY)
1682 if ((cpu_online_map & (1 << id)) != 0) {
1684 current->cpus_allowed = (1 << id);
1686 set_cpus_allowed (current, 1<<id);
1689 CERROR ("Can't set CPU affinity for %s\n", name);
1691 #endif /* CONFIG_SMP && CPU_AFFINITY */
1693 spin_lock_irqsave (&sched->kss_lock, flags);
1695 while (!ksocknal_data.ksnd_shuttingdown) {
1696 int did_something = 0;
1698 /* Ensure I progress everything semi-fairly */
1700 if (!list_empty (&sched->kss_rx_conns)) {
1701 conn = list_entry(sched->kss_rx_conns.next,
1702 ksock_conn_t, ksnc_rx_list);
1703 list_del(&conn->ksnc_rx_list);
1705 LASSERT(conn->ksnc_rx_scheduled);
1706 LASSERT(conn->ksnc_rx_ready);
1708 /* clear rx_ready in case receive isn't complete.
1709 * Do it BEFORE we call process_recv, since
1710 * data_ready can set it any time after we release
1712 conn->ksnc_rx_ready = 0;
1713 spin_unlock_irqrestore(&sched->kss_lock, flags);
1715 rc = ksocknal_process_receive(conn);
1717 spin_lock_irqsave(&sched->kss_lock, flags);
1719 /* I'm the only one that can clear this flag */
1720 LASSERT(conn->ksnc_rx_scheduled);
1722 /* Did process_receive get everything it wanted? */
1724 conn->ksnc_rx_ready = 1;
1726 if (conn->ksnc_rx_state == SOCKNAL_RX_FMB_SLEEP ||
1727 conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB) {
1728 /* Conn blocked for a forwarding buffer.
1729 * It will get queued for my attention when
1730 * one becomes available (and it might just
1731 * already have been!). Meanwhile my ref
1732 * on it stays put. */
1733 } else if (conn->ksnc_rx_ready) {
1734 /* reschedule for rx */
1735 list_add_tail (&conn->ksnc_rx_list,
1736 &sched->kss_rx_conns);
1738 conn->ksnc_rx_scheduled = 0;
1740 ksocknal_put_conn(conn);
1746 if (!list_empty (&sched->kss_tx_conns)) {
1747 conn = list_entry(sched->kss_tx_conns.next,
1748 ksock_conn_t, ksnc_tx_list);
1749 list_del (&conn->ksnc_tx_list);
1751 LASSERT(conn->ksnc_tx_scheduled);
1752 LASSERT(conn->ksnc_tx_ready);
1753 LASSERT(!list_empty(&conn->ksnc_tx_queue));
1755 tx = list_entry(conn->ksnc_tx_queue.next,
1756 ksock_tx_t, tx_list);
1757 /* dequeue now so empty list => more to send */
1758 list_del(&tx->tx_list);
1760 /* Clear tx_ready in case send isn't complete. Do
1761 * it BEFORE we call process_transmit, since
1762 * write_space can set it any time after we release
1764 conn->ksnc_tx_ready = 0;
1765 spin_unlock_irqrestore (&sched->kss_lock, flags);
1767 rc = ksocknal_process_transmit(conn, tx);
1769 spin_lock_irqsave (&sched->kss_lock, flags);
1771 if (rc == -ENOMEM || rc == -EAGAIN) {
1772 /* Incomplete send: replace tx on HEAD of tx_queue */
1773 list_add (&tx->tx_list, &conn->ksnc_tx_queue);
1775 /* Complete send; assume space for more */
1776 conn->ksnc_tx_ready = 1;
1779 if (rc == -ENOMEM) {
1780 /* Do nothing; after a short timeout, this
1781 * conn will be reposted on kss_tx_conns. */
1782 } else if (conn->ksnc_tx_ready &&
1783 !list_empty (&conn->ksnc_tx_queue)) {
1784 /* reschedule for tx */
1785 list_add_tail (&conn->ksnc_tx_list,
1786 &sched->kss_tx_conns);
1788 conn->ksnc_tx_scheduled = 0;
1790 ksocknal_put_conn (conn);
1796 if (!list_empty (&sched->kss_zctxdone_list)) {
1798 list_entry(sched->kss_zctxdone_list.next,
1799 ksock_tx_t, tx_list);
1802 list_del (&tx->tx_list);
1803 spin_unlock_irqrestore (&sched->kss_lock, flags);
1805 ksocknal_tx_done (tx, 1);
1807 spin_lock_irqsave (&sched->kss_lock, flags);
1810 if (!did_something || /* nothing to do */
1811 ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1812 spin_unlock_irqrestore (&sched->kss_lock, flags);
1816 if (!did_something) { /* wait for something to do */
1818 rc = wait_event_interruptible (sched->kss_waitq,
1819 ksocknal_data.ksnd_shuttingdown ||
1820 !list_empty(&sched->kss_rx_conns) ||
1821 !list_empty(&sched->kss_tx_conns) ||
1822 !list_empty(&sched->kss_zctxdone_list));
1824 rc = wait_event_interruptible (sched->kss_waitq,
1825 ksocknal_data.ksnd_shuttingdown ||
1826 !list_empty(&sched->kss_rx_conns) ||
1827 !list_empty(&sched->kss_tx_conns));
1833 spin_lock_irqsave (&sched->kss_lock, flags);
1837 spin_unlock_irqrestore (&sched->kss_lock, flags);
1838 ksocknal_thread_fini ();
1843 ksocknal_data_ready (struct sock *sk, int n)
1845 unsigned long flags;
1847 ksock_sched_t *sched;
1850 /* interleave correctly with closing sockets... */
1851 read_lock (&ksocknal_data.ksnd_global_lock);
1853 conn = sk->sk_user_data;
1854 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
1855 LASSERT (sk->sk_data_ready != &ksocknal_data_ready);
1856 sk->sk_data_ready (sk, n);
1858 sched = conn->ksnc_scheduler;
1860 spin_lock_irqsave (&sched->kss_lock, flags);
1862 conn->ksnc_rx_ready = 1;
1864 if (!conn->ksnc_rx_scheduled) { /* not being progressed */
1865 list_add_tail(&conn->ksnc_rx_list,
1866 &sched->kss_rx_conns);
1867 conn->ksnc_rx_scheduled = 1;
1868 /* extra ref for scheduler */
1869 atomic_inc (&conn->ksnc_refcount);
1871 wake_up (&sched->kss_waitq);
1874 spin_unlock_irqrestore (&sched->kss_lock, flags);
1877 read_unlock (&ksocknal_data.ksnd_global_lock);
1883 ksocknal_write_space (struct sock *sk)
1885 unsigned long flags;
1887 ksock_sched_t *sched;
1889 /* interleave correctly with closing sockets... */
1890 read_lock (&ksocknal_data.ksnd_global_lock);
1892 conn = sk->sk_user_data;
1894 CDEBUG(D_NET, "sk %p wspace %d low water %d conn %p%s%s%s\n",
1895 sk, tcp_wspace(sk), SOCKNAL_TX_LOW_WATER(sk), conn,
1896 (conn == NULL) ? "" : (conn->ksnc_tx_ready ?
1897 " ready" : " blocked"),
1898 (conn == NULL) ? "" : (conn->ksnc_tx_scheduled ?
1899 " scheduled" : " idle"),
1900 (conn == NULL) ? "" : (list_empty (&conn->ksnc_tx_queue) ?
1901 " empty" : " queued"));
1903 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
1904 LASSERT (sk->sk_write_space != &ksocknal_write_space);
1905 sk->sk_write_space (sk);
1907 read_unlock (&ksocknal_data.ksnd_global_lock);
1911 if (tcp_wspace(sk) >= SOCKNAL_TX_LOW_WATER(sk)) { /* got enough space */
1912 clear_bit (SOCK_NOSPACE, &sk->sk_socket->flags);
1914 sched = conn->ksnc_scheduler;
1916 spin_lock_irqsave (&sched->kss_lock, flags);
1918 conn->ksnc_tx_ready = 1;
1920 if (!conn->ksnc_tx_scheduled && // not being progressed
1921 !list_empty(&conn->ksnc_tx_queue)){//packets to send
1922 list_add_tail (&conn->ksnc_tx_list,
1923 &sched->kss_tx_conns);
1924 conn->ksnc_tx_scheduled = 1;
1925 /* extra ref for scheduler */
1926 atomic_inc (&conn->ksnc_refcount);
1928 wake_up (&sched->kss_waitq);
1931 spin_unlock_irqrestore (&sched->kss_lock, flags);
1934 read_unlock (&ksocknal_data.ksnd_global_lock);
1938 ksocknal_sock_write (struct socket *sock, void *buffer, int nob)
1941 mm_segment_t oldmm = get_fs();
1944 struct iovec iov = {
1948 struct msghdr msg = {
1953 .msg_control = NULL,
1954 .msg_controllen = 0,
1959 rc = sock_sendmsg (sock, &msg, iov.iov_len);
1966 CERROR ("Unexpected zero rc\n");
1967 return (-ECONNABORTED);
1970 buffer = ((char *)buffer) + rc;
1978 ksocknal_sock_read (struct socket *sock, void *buffer, int nob)
1981 mm_segment_t oldmm = get_fs();
1984 struct iovec iov = {
1988 struct msghdr msg = {
1993 .msg_control = NULL,
1994 .msg_controllen = 0,
1999 rc = sock_recvmsg (sock, &msg, iov.iov_len, 0);
2006 return (-ECONNABORTED);
2008 buffer = ((char *)buffer) + rc;
2016 ksocknal_hello (struct socket *sock, ptl_nid_t *nid, int *type,
2021 ptl_magicversion_t *hmv = (ptl_magicversion_t *)&hdr.dest_nid;
2022 char ipbuf[PTL_NALFMT_SIZE];
2023 char ipbuf2[PTL_NALFMT_SIZE];
2025 LASSERT (sizeof (*hmv) == sizeof (hdr.dest_nid));
2027 memset (&hdr, 0, sizeof (hdr));
2028 hmv->magic = __cpu_to_le32 (PORTALS_PROTO_MAGIC);
2029 hmv->version_major = __cpu_to_le16 (PORTALS_PROTO_VERSION_MAJOR);
2030 hmv->version_minor = __cpu_to_le16 (PORTALS_PROTO_VERSION_MINOR);
2032 hdr.src_nid = __cpu_to_le64 (ksocknal_lib.ni.nid);
2033 hdr.type = __cpu_to_le32 (PTL_MSG_HELLO);
2035 hdr.msg.hello.type = __cpu_to_le32 (*type);
2036 hdr.msg.hello.incarnation =
2037 __cpu_to_le64 (ksocknal_data.ksnd_incarnation);
2039 /* Assume sufficient socket buffering for this message */
2040 rc = ksocknal_sock_write (sock, &hdr, sizeof (hdr));
2042 CERROR ("Error %d sending HELLO to "LPX64" %s\n",
2043 rc, *nid, portals_nid2str(SOCKNAL, *nid, ipbuf));
2047 rc = ksocknal_sock_read (sock, hmv, sizeof (*hmv));
2049 CERROR ("Error %d reading HELLO from "LPX64" %s\n",
2050 rc, *nid, portals_nid2str(SOCKNAL, *nid, ipbuf));
2054 if (hmv->magic != __le32_to_cpu (PORTALS_PROTO_MAGIC)) {
2055 CERROR ("Bad magic %#08x (%#08x expected) from "LPX64" %s\n",
2056 __cpu_to_le32 (hmv->magic), PORTALS_PROTO_MAGIC, *nid,
2057 portals_nid2str(SOCKNAL, *nid, ipbuf));
2061 if (hmv->version_major != __cpu_to_le16 (PORTALS_PROTO_VERSION_MAJOR) ||
2062 hmv->version_minor != __cpu_to_le16 (PORTALS_PROTO_VERSION_MINOR)) {
2063 CERROR ("Incompatible protocol version %d.%d (%d.%d expected)"
2064 " from "LPX64" %s\n",
2065 __le16_to_cpu (hmv->version_major),
2066 __le16_to_cpu (hmv->version_minor),
2067 PORTALS_PROTO_VERSION_MAJOR,
2068 PORTALS_PROTO_VERSION_MINOR,
2069 *nid, portals_nid2str(SOCKNAL, *nid, ipbuf));
2073 #if (PORTALS_PROTO_VERSION_MAJOR != 0)
2074 # error "This code only understands protocol version 0.x"
2076 /* version 0 sends magic/version as the dest_nid of a 'hello' header,
2077 * so read the rest of it in now... */
2079 rc = ksocknal_sock_read (sock, hmv + 1, sizeof (hdr) - sizeof (*hmv));
2081 CERROR ("Error %d reading rest of HELLO hdr from "LPX64" %s\n",
2082 rc, *nid, portals_nid2str(SOCKNAL, *nid, ipbuf));
2086 /* ...and check we got what we expected */
2087 if (hdr.type != __cpu_to_le32 (PTL_MSG_HELLO) ||
2088 hdr.payload_length != __cpu_to_le32 (0)) {
2089 CERROR ("Expecting a HELLO hdr with 0 payload,"
2090 " but got type %d with %d payload from "LPX64" %s\n",
2091 __le32_to_cpu (hdr.type),
2092 __le32_to_cpu (hdr.payload_length), *nid,
2093 portals_nid2str(SOCKNAL, *nid, ipbuf));
2097 if (__le64_to_cpu(hdr.src_nid) == PTL_NID_ANY) {
2098 CERROR("Expecting a HELLO hdr with a NID, but got PTL_NID_ANY\n");
2102 if (*nid == PTL_NID_ANY) { /* don't know peer's nid yet */
2103 *nid = __le64_to_cpu(hdr.src_nid);
2104 } else if (*nid != __le64_to_cpu (hdr.src_nid)) {
2105 CERROR ("Connected to nid "LPX64" %s, but expecting "LPX64" %s\n",
2106 __le64_to_cpu (hdr.src_nid),
2107 portals_nid2str(SOCKNAL,
2108 __le64_to_cpu(hdr.src_nid),
2110 *nid, portals_nid2str(SOCKNAL, *nid, ipbuf2));
2114 if (*type == SOCKNAL_CONN_NONE) {
2115 /* I've accepted this connection; peer determines type */
2116 *type = __le32_to_cpu(hdr.msg.hello.type);
2118 case SOCKNAL_CONN_ANY:
2119 case SOCKNAL_CONN_CONTROL:
2121 case SOCKNAL_CONN_BULK_IN:
2122 *type = SOCKNAL_CONN_BULK_OUT;
2124 case SOCKNAL_CONN_BULK_OUT:
2125 *type = SOCKNAL_CONN_BULK_IN;
2128 CERROR ("Unexpected type %d from "LPX64" %s\n",
2130 portals_nid2str(SOCKNAL, *nid, ipbuf));
2133 } else if (__le32_to_cpu(hdr.msg.hello.type) != SOCKNAL_CONN_NONE) {
2134 CERROR ("Mismatched types: me %d "LPX64" %s %d\n",
2135 *type, *nid, portals_nid2str(SOCKNAL, *nid, ipbuf),
2136 __le32_to_cpu(hdr.msg.hello.type));
2140 *incarnation = __le64_to_cpu(hdr.msg.hello.incarnation);
2146 ksocknal_setup_sock (struct socket *sock)
2148 mm_segment_t oldmm = get_fs ();
2151 struct linger linger;
2153 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
2154 sock->sk->sk_allocation = GFP_NOFS;
2156 sock->sk->allocation = GFP_NOFS;
2159 /* Ensure this socket aborts active sends immediately when we close
2163 linger.l_linger = 0;
2166 rc = sock_setsockopt (sock, SOL_SOCKET, SO_LINGER,
2167 (char *)&linger, sizeof (linger));
2170 CERROR ("Can't set SO_LINGER: %d\n", rc);
2176 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_LINGER2,
2177 (char *)&option, sizeof (option));
2180 CERROR ("Can't set SO_LINGER2: %d\n", rc);
2184 #if SOCKNAL_USE_KEEPALIVES
2185 /* Keepalives: If 3/4 of the timeout elapses, start probing every
2186 * second until the timeout elapses. */
2188 option = (ksocknal_tunables.ksnd_io_timeout * 3) / 4;
2190 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPIDLE,
2191 (char *)&option, sizeof (option));
2194 CERROR ("Can't set TCP_KEEPIDLE: %d\n", rc);
2200 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPINTVL,
2201 (char *)&option, sizeof (option));
2204 CERROR ("Can't set TCP_KEEPINTVL: %d\n", rc);
2208 option = ksocknal_tunables.ksnd_io_timeout / 4;
2210 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPCNT,
2211 (char *)&option, sizeof (option));
2214 CERROR ("Can't set TCP_KEEPINTVL: %d\n", rc);
2220 rc = sock_setsockopt (sock, SOL_SOCKET, SO_KEEPALIVE,
2221 (char *)&option, sizeof (option));
2224 CERROR ("Can't set SO_KEEPALIVE: %d\n", rc);
2232 ksocknal_connect_peer (ksock_route_t *route, int type)
2234 struct sockaddr_in peer_addr;
2235 mm_segment_t oldmm = get_fs();
2238 struct socket *sock;
2240 char ipbuf[PTL_NALFMT_SIZE];
2242 rc = sock_create (PF_INET, SOCK_STREAM, 0, &sock);
2244 CERROR ("Can't create autoconnect socket: %d\n", rc);
2248 /* Ugh; have to map_fd for compatibility with sockets passed in
2249 * from userspace. And we actually need the sock->file refcounting
2250 * that this gives you :) */
2252 fd = sock_map_fd (sock);
2254 sock_release (sock);
2255 CERROR ("sock_map_fd error %d\n", fd);
2259 /* NB the fd now owns the ref on sock->file */
2260 LASSERT (sock->file != NULL);
2261 LASSERT (file_count(sock->file) == 1);
2263 /* Set the socket timeouts, so our connection attempt completes in
2265 tv.tv_sec = ksocknal_tunables.ksnd_io_timeout;
2269 rc = sock_setsockopt (sock, SOL_SOCKET, SO_SNDTIMEO,
2270 (char *)&tv, sizeof (tv));
2273 CERROR ("Can't set send timeout %d: %d\n",
2274 ksocknal_tunables.ksnd_io_timeout, rc);
2279 rc = sock_setsockopt (sock, SOL_SOCKET, SO_RCVTIMEO,
2280 (char *)&tv, sizeof (tv));
2283 CERROR ("Can't set receive timeout %d: %d\n",
2284 ksocknal_tunables.ksnd_io_timeout, rc);
2292 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_NODELAY,
2293 (char *)&option, sizeof (option));
2296 CERROR ("Can't disable nagle: %d\n", rc);
2301 if (route->ksnr_buffer_size != 0) {
2302 int option = route->ksnr_buffer_size;
2305 rc = sock_setsockopt (sock, SOL_SOCKET, SO_SNDBUF,
2306 (char *)&option, sizeof (option));
2309 CERROR ("Can't set send buffer %d: %d\n",
2310 route->ksnr_buffer_size, rc);
2315 rc = sock_setsockopt (sock, SOL_SOCKET, SO_RCVBUF,
2316 (char *)&option, sizeof (option));
2319 CERROR ("Can't set receive buffer %d: %d\n",
2320 route->ksnr_buffer_size, rc);
2325 memset (&peer_addr, 0, sizeof (peer_addr));
2326 peer_addr.sin_family = AF_INET;
2327 peer_addr.sin_port = htons (route->ksnr_port);
2328 peer_addr.sin_addr.s_addr = htonl (route->ksnr_ipaddr);
2330 rc = sock->ops->connect (sock, (struct sockaddr *)&peer_addr,
2331 sizeof (peer_addr), sock->file->f_flags);
2333 CERROR ("Error %d connecting to "LPX64" %s\n", rc,
2334 route->ksnr_peer->ksnp_nid,
2335 portals_nid2str(SOCKNAL,
2336 route->ksnr_peer->ksnp_nid,
2341 rc = ksocknal_create_conn (route, sock, route->ksnr_irq_affinity, type);
2343 /* Take an extra ref on sock->file to compensate for the
2344 * upcoming close which will lose fd's ref on it. */
2345 get_file (sock->file);
2354 ksocknal_autoconnect (ksock_route_t *route)
2356 LIST_HEAD (zombies);
2359 unsigned long flags;
2364 for (type = 0; type < SOCKNAL_CONN_NTYPES; type++)
2365 if ((route->ksnr_connecting & (1 << type)) != 0)
2367 LASSERT (type < SOCKNAL_CONN_NTYPES);
2369 rc = ksocknal_connect_peer (route, type);
2374 /* successfully autoconnected: create_conn did the
2375 * route/conn binding and scheduled any blocked packets */
2377 if (route->ksnr_connecting == 0) {
2378 /* No more connections required */
2383 /* Connection attempt failed */
2385 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
2387 peer = route->ksnr_peer;
2388 route->ksnr_connecting = 0;
2390 /* This is a retry rather than a new connection */
2391 LASSERT (route->ksnr_retry_interval != 0);
2392 route->ksnr_timeout = jiffies + route->ksnr_retry_interval;
2393 route->ksnr_retry_interval = MIN (route->ksnr_retry_interval * 2,
2394 SOCKNAL_MAX_RECONNECT_INTERVAL);
2396 if (!list_empty (&peer->ksnp_tx_queue) &&
2397 ksocknal_find_connecting_route_locked (peer) == NULL) {
2398 LASSERT (list_empty (&peer->ksnp_conns));
2400 /* None of the connections that the blocked packets are
2401 * waiting for have been successful. Complete them now... */
2403 tx = list_entry (peer->ksnp_tx_queue.next,
2404 ksock_tx_t, tx_list);
2405 list_del (&tx->tx_list);
2406 list_add_tail (&tx->tx_list, &zombies);
2407 } while (!list_empty (&peer->ksnp_tx_queue));
2410 /* make this route least-favourite for re-selection */
2411 if (!route->ksnr_deleted) {
2412 list_del(&route->ksnr_list);
2413 list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
2416 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
2418 while (!list_empty (&zombies)) {
2419 char ipbuf[PTL_NALFMT_SIZE];
2420 char ipbuf2[PTL_NALFMT_SIZE];
2421 tx = list_entry (zombies.next, ksock_tx_t, tx_list);
2423 CERROR ("Deleting packet type %d len %d ("LPX64" %s->"LPX64" %s)\n",
2424 NTOH__u32 (tx->tx_hdr->type),
2425 NTOH__u32 (tx->tx_hdr->payload_length),
2426 NTOH__u64 (tx->tx_hdr->src_nid),
2427 portals_nid2str(SOCKNAL,
2428 NTOH__u64(tx->tx_hdr->src_nid),
2430 NTOH__u64 (tx->tx_hdr->dest_nid),
2431 portals_nid2str(SOCKNAL,
2432 NTOH__u64(tx->tx_hdr->src_nid),
2435 list_del (&tx->tx_list);
2437 ksocknal_tx_done (tx, 0);
2442 ksocknal_autoconnectd (void *arg)
2444 long id = (long)arg;
2446 unsigned long flags;
2447 ksock_route_t *route;
2450 snprintf (name, sizeof (name), "ksocknal_ad%02ld", id);
2451 kportal_daemonize (name);
2452 kportal_blockallsigs ();
2454 spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2456 while (!ksocknal_data.ksnd_shuttingdown) {
2458 if (!list_empty (&ksocknal_data.ksnd_autoconnectd_routes)) {
2459 route = list_entry (ksocknal_data.ksnd_autoconnectd_routes.next,
2460 ksock_route_t, ksnr_connect_list);
2462 list_del (&route->ksnr_connect_list);
2463 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2465 ksocknal_autoconnect (route);
2466 ksocknal_put_route (route);
2468 spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2472 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2474 rc = wait_event_interruptible (ksocknal_data.ksnd_autoconnectd_waitq,
2475 ksocknal_data.ksnd_shuttingdown ||
2476 !list_empty (&ksocknal_data.ksnd_autoconnectd_routes));
2478 spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2481 spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
2483 ksocknal_thread_fini ();
2488 ksocknal_find_timed_out_conn (ksock_peer_t *peer)
2490 /* We're called with a shared lock on ksnd_global_lock */
2492 struct list_head *ctmp;
2493 ksock_sched_t *sched;
2495 list_for_each (ctmp, &peer->ksnp_conns) {
2496 conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
2497 sched = conn->ksnc_scheduler;
2499 /* Don't need the {get,put}connsock dance to deref ksnc_sock... */
2500 LASSERT (!conn->ksnc_closing);
2502 if (conn->ksnc_rx_started &&
2503 time_after_eq (jiffies, conn->ksnc_rx_deadline)) {
2504 /* Timed out incomplete incoming message */
2505 atomic_inc (&conn->ksnc_refcount);
2506 CERROR ("Timed out RX from "LPX64" %p %d.%d.%d.%d\n",
2507 peer->ksnp_nid, conn, HIPQUAD(conn->ksnc_ipaddr));
2511 if ((!list_empty (&conn->ksnc_tx_queue) ||
2512 conn->ksnc_sock->sk->sk_wmem_queued != 0) &&
2513 time_after_eq (jiffies, conn->ksnc_tx_deadline)) {
2514 /* Timed out messages queued for sending, or
2515 * messages buffered in the socket's send buffer */
2516 atomic_inc (&conn->ksnc_refcount);
2517 CERROR ("Timed out TX to "LPX64" %s%d %p %d.%d.%d.%d\n",
2519 list_empty (&conn->ksnc_tx_queue) ? "" : "Q ",
2520 conn->ksnc_sock->sk->sk_wmem_queued, conn,
2521 HIPQUAD(conn->ksnc_ipaddr));
2530 ksocknal_check_peer_timeouts (int idx)
2532 struct list_head *peers = &ksocknal_data.ksnd_peers[idx];
2533 struct list_head *ptmp;
2538 /* NB. We expect to have a look at all the peers and not find any
2539 * connections to time out, so we just use a shared lock while we
2541 read_lock (&ksocknal_data.ksnd_global_lock);
2543 list_for_each (ptmp, peers) {
2544 peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
2545 conn = ksocknal_find_timed_out_conn (peer);
2548 read_unlock (&ksocknal_data.ksnd_global_lock);
2550 CERROR ("Timeout out conn->"LPX64" ip %d.%d.%d.%d:%d\n",
2552 HIPQUAD(conn->ksnc_ipaddr),
2554 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2556 /* NB we won't find this one again, but we can't
2557 * just proceed with the next peer, since we dropped
2558 * ksnd_global_lock and it might be dead already! */
2559 ksocknal_put_conn (conn);
2564 read_unlock (&ksocknal_data.ksnd_global_lock);
2568 ksocknal_reaper (void *arg)
2571 unsigned long flags;
2573 ksock_sched_t *sched;
2574 struct list_head enomem_conns;
2579 unsigned long deadline = jiffies;
2581 kportal_daemonize ("ksocknal_reaper");
2582 kportal_blockallsigs ();
2584 INIT_LIST_HEAD(&enomem_conns);
2585 init_waitqueue_entry (&wait, current);
2587 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2589 while (!ksocknal_data.ksnd_shuttingdown) {
2591 if (!list_empty (&ksocknal_data.ksnd_deathrow_conns)) {
2592 conn = list_entry (ksocknal_data.ksnd_deathrow_conns.next,
2593 ksock_conn_t, ksnc_list);
2594 list_del (&conn->ksnc_list);
2596 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2598 ksocknal_terminate_conn (conn);
2599 ksocknal_put_conn (conn);
2601 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2605 if (!list_empty (&ksocknal_data.ksnd_zombie_conns)) {
2606 conn = list_entry (ksocknal_data.ksnd_zombie_conns.next,
2607 ksock_conn_t, ksnc_list);
2608 list_del (&conn->ksnc_list);
2610 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2612 ksocknal_destroy_conn (conn);
2614 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2618 if (!list_empty (&ksocknal_data.ksnd_enomem_conns)) {
2619 list_add(&enomem_conns, &ksocknal_data.ksnd_enomem_conns);
2620 list_del_init(&ksocknal_data.ksnd_enomem_conns);
2623 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2625 /* reschedule all the connections that stalled with ENOMEM... */
2627 while (!list_empty (&enomem_conns)) {
2628 conn = list_entry (enomem_conns.next,
2629 ksock_conn_t, ksnc_tx_list);
2630 list_del (&conn->ksnc_tx_list);
2632 sched = conn->ksnc_scheduler;
2634 spin_lock_irqsave (&sched->kss_lock, flags);
2636 LASSERT (conn->ksnc_tx_scheduled);
2637 conn->ksnc_tx_ready = 1;
2638 list_add_tail (&conn->ksnc_tx_list, &sched->kss_tx_conns);
2639 wake_up (&sched->kss_waitq);
2641 spin_unlock_irqrestore (&sched->kss_lock, flags);
2645 /* careful with the jiffy wrap... */
2646 while ((timeout = (int)(deadline - jiffies)) <= 0) {
2649 int chunk = ksocknal_data.ksnd_peer_hash_size;
2651 /* Time to check for timeouts on a few more peers: I do
2652 * checks every 'p' seconds on a proportion of the peer
2653 * table and I need to check every connection 'n' times
2654 * within a timeout interval, to ensure I detect a
2655 * timeout on any connection within (n+1)/n times the
2656 * timeout interval. */
2658 if (ksocknal_tunables.ksnd_io_timeout > n * p)
2659 chunk = (chunk * n * p) /
2660 ksocknal_tunables.ksnd_io_timeout;
2664 for (i = 0; i < chunk; i++) {
2665 ksocknal_check_peer_timeouts (peer_index);
2666 peer_index = (peer_index + 1) %
2667 ksocknal_data.ksnd_peer_hash_size;
2673 if (nenomem_conns != 0) {
2674 /* Reduce my timeout if I rescheduled ENOMEM conns.
2675 * This also prevents me getting woken immediately
2676 * if any go back on my enomem list. */
2677 timeout = SOCKNAL_ENOMEM_RETRY;
2679 ksocknal_data.ksnd_reaper_waketime = jiffies + timeout;
2681 set_current_state (TASK_INTERRUPTIBLE);
2682 add_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait);
2684 if (!ksocknal_data.ksnd_shuttingdown &&
2685 list_empty (&ksocknal_data.ksnd_deathrow_conns) &&
2686 list_empty (&ksocknal_data.ksnd_zombie_conns))
2687 schedule_timeout (timeout);
2689 set_current_state (TASK_RUNNING);
2690 remove_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait);
2692 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
2695 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
2697 ksocknal_thread_fini ();
2701 nal_cb_t ksocknal_lib = {
2702 nal_data: &ksocknal_data, /* NAL private data */
2703 cb_send: ksocknal_send,
2704 cb_send_pages: ksocknal_send_pages,
2705 cb_recv: ksocknal_recv,
2706 cb_recv_pages: ksocknal_recv_pages,
2707 cb_read: ksocknal_read,
2708 cb_write: ksocknal_write,
2709 cb_malloc: ksocknal_malloc,
2710 cb_free: ksocknal_free,
2711 cb_printf: ksocknal_printf,
2712 cb_cli: ksocknal_cli,
2713 cb_sti: ksocknal_sti,
2714 cb_callback: ksocknal_callback,
2715 cb_dist: ksocknal_dist