4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
39 # if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
41 static struct ctl_table ksocknal_ctl_table[] = {
44 .procname = "timeout",
45 .data = &ksocknal_tunables.ksnd_timeout,
46 .maxlen = sizeof (int),
48 .proc_handler = &proc_dointvec,
53 .procname = "credits",
54 .data = &ksocknal_tunables.ksnd_credits,
55 .maxlen = sizeof (int),
57 .proc_handler = &proc_dointvec,
62 .procname = "peer_credits",
63 .data = &ksocknal_tunables.ksnd_peertxcredits,
64 .maxlen = sizeof (int),
66 .proc_handler = &proc_dointvec,
71 .procname = "peer_buffer_credits",
72 .data = &ksocknal_tunables.ksnd_peerrtrcredits,
73 .maxlen = sizeof (int),
75 .proc_handler = &proc_dointvec,
80 .procname = "peer_timeout",
81 .data = &ksocknal_tunables.ksnd_peertimeout,
82 .maxlen = sizeof (int),
84 .proc_handler = &proc_dointvec
89 .procname = "nconnds",
90 .data = &ksocknal_tunables.ksnd_nconnds,
91 .maxlen = sizeof (int),
93 .proc_handler = &proc_dointvec,
98 .procname = "min_reconnectms",
99 .data = &ksocknal_tunables.ksnd_min_reconnectms,
100 .maxlen = sizeof (int),
102 .proc_handler = &proc_dointvec,
107 .procname = "max_reconnectms",
108 .data = &ksocknal_tunables.ksnd_max_reconnectms,
109 .maxlen = sizeof (int),
111 .proc_handler = &proc_dointvec,
116 .procname = "eager_ack",
117 .data = &ksocknal_tunables.ksnd_eager_ack,
118 .maxlen = sizeof (int),
120 .proc_handler = &proc_dointvec,
125 .procname = "zero_copy",
126 .data = &ksocknal_tunables.ksnd_zc_min_payload,
127 .maxlen = sizeof (int),
129 .proc_handler = &proc_dointvec,
134 .procname = "zero_copy_recv",
135 .data = &ksocknal_tunables.ksnd_zc_recv,
136 .maxlen = sizeof (int),
138 .proc_handler = &proc_dointvec,
143 .procname = "zero_copy_recv",
144 .data = &ksocknal_tunables.ksnd_zc_recv_min_nfrags,
145 .maxlen = sizeof (int),
147 .proc_handler = &proc_dointvec,
153 .data = &ksocknal_tunables.ksnd_typed_conns,
154 .maxlen = sizeof (int),
156 .proc_handler = &proc_dointvec,
161 .procname = "min_bulk",
162 .data = &ksocknal_tunables.ksnd_min_bulk,
163 .maxlen = sizeof (int),
165 .proc_handler = &proc_dointvec,
170 .procname = "rx_buffer_size",
171 .data = &ksocknal_tunables.ksnd_rx_buffer_size,
172 .maxlen = sizeof(int),
174 .proc_handler = &proc_dointvec,
179 .procname = "tx_buffer_size",
180 .data = &ksocknal_tunables.ksnd_tx_buffer_size,
181 .maxlen = sizeof(int),
183 .proc_handler = &proc_dointvec,
189 .data = &ksocknal_tunables.ksnd_nagle,
190 .maxlen = sizeof(int),
192 .proc_handler = &proc_dointvec,
198 .procname = "irq_affinity",
199 .data = &ksocknal_tunables.ksnd_irq_affinity,
200 .maxlen = sizeof(int),
202 .proc_handler = &proc_dointvec,
208 .procname = "round_robin",
209 .data = &ksocknal_tunables.ksnd_round_robin,
210 .maxlen = sizeof(int),
212 .proc_handler = &proc_dointvec,
217 .procname = "keepalive",
218 .data = &ksocknal_tunables.ksnd_keepalive,
219 .maxlen = sizeof(int),
221 .proc_handler = &proc_dointvec,
226 .procname = "keepalive_idle",
227 .data = &ksocknal_tunables.ksnd_keepalive_idle,
228 .maxlen = sizeof(int),
230 .proc_handler = &proc_dointvec,
235 .procname = "keepalive_count",
236 .data = &ksocknal_tunables.ksnd_keepalive_count,
237 .maxlen = sizeof(int),
239 .proc_handler = &proc_dointvec,
244 .procname = "keepalive_intvl",
245 .data = &ksocknal_tunables.ksnd_keepalive_intvl,
246 .maxlen = sizeof(int),
248 .proc_handler = &proc_dointvec,
251 #ifdef SOCKNAL_BACKOFF
254 .procname = "backoff_init",
255 .data = &ksocknal_tunables.ksnd_backoff_init,
256 .maxlen = sizeof(int),
258 .proc_handler = &proc_dointvec,
263 .procname = "backoff_max",
264 .data = &ksocknal_tunables.ksnd_backoff_max,
265 .maxlen = sizeof(int),
267 .proc_handler = &proc_dointvec,
271 #if SOCKNAL_VERSION_DEBUG
274 .procname = "protocol",
275 .data = &ksocknal_tunables.ksnd_protocol,
276 .maxlen = sizeof(int),
278 .proc_handler = &proc_dointvec,
285 struct ctl_table ksocknal_top_ctl_table[] = {
288 .procname = "socknal",
292 .child = ksocknal_ctl_table
298 ksocknal_lib_tunables_init ()
300 if (!*ksocknal_tunables.ksnd_typed_conns) {
302 #if SOCKNAL_VERSION_DEBUG
303 if (*ksocknal_tunables.ksnd_protocol < 3)
307 CERROR("Protocol V3.x MUST have typed connections\n");
312 if (*ksocknal_tunables.ksnd_zc_recv_min_nfrags < 2)
313 *ksocknal_tunables.ksnd_zc_recv_min_nfrags = 2;
314 if (*ksocknal_tunables.ksnd_zc_recv_min_nfrags > LNET_MAX_IOV)
315 *ksocknal_tunables.ksnd_zc_recv_min_nfrags = LNET_MAX_IOV;
317 ksocknal_tunables.ksnd_sysctl =
318 register_sysctl_table(ksocknal_top_ctl_table);
320 if (ksocknal_tunables.ksnd_sysctl == NULL)
321 CWARN("Can't setup /proc tunables\n");
327 ksocknal_lib_tunables_fini ()
329 if (ksocknal_tunables.ksnd_sysctl != NULL)
330 unregister_sysctl_table(ksocknal_tunables.ksnd_sysctl);
334 ksocknal_lib_tunables_init ()
340 ksocknal_lib_tunables_fini ()
343 #endif /* # if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM */
346 ksocknal_lib_get_conn_addrs (ksock_conn_t *conn)
348 int rc = libcfs_sock_getaddr(conn->ksnc_sock, 1,
352 /* Didn't need the {get,put}connsock dance to deref ksnc_sock... */
353 LASSERT (!conn->ksnc_closing);
356 CERROR ("Error %d getting sock peer IP\n", rc);
360 rc = libcfs_sock_getaddr(conn->ksnc_sock, 0,
361 &conn->ksnc_myipaddr, NULL);
363 CERROR ("Error %d getting sock local IP\n", rc);
371 ksocknal_lib_zc_capable(ksock_conn_t *conn)
373 int caps = conn->ksnc_sock->sk->sk_route_caps;
375 if (conn->ksnc_proto == &ksocknal_protocol_v1x)
378 /* ZC if the socket supports scatter/gather and doesn't need software
380 return ((caps & NETIF_F_SG) != 0 && (caps & NETIF_F_ALL_CSUM) != 0);
384 ksocknal_lib_send_iov (ksock_conn_t *conn, ksock_tx_t *tx)
386 struct socket *sock = conn->ksnc_sock;
390 if (*ksocknal_tunables.ksnd_enable_csum && /* checksum enabled */
391 conn->ksnc_proto == &ksocknal_protocol_v2x && /* V2.x connection */
392 tx->tx_nob == tx->tx_resid && /* frist sending */
393 tx->tx_msg.ksm_csum == 0) /* not checksummed */
394 ksocknal_lib_csum_tx(tx);
396 /* NB we can't trust socket ops to either consume our iovs
397 * or leave them alone. */
400 #if SOCKNAL_SINGLE_FRAG_TX
401 struct iovec scratch;
402 struct iovec *scratchiov = &scratch;
403 unsigned int niov = 1;
405 struct iovec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
406 unsigned int niov = tx->tx_niov;
408 struct msghdr msg = {
411 .msg_iov = scratchiov,
415 .msg_flags = MSG_DONTWAIT
417 mm_segment_t oldmm = get_fs();
420 for (nob = i = 0; i < niov; i++) {
421 scratchiov[i] = tx->tx_iov[i];
422 nob += scratchiov[i].iov_len;
425 if (!cfs_list_empty(&conn->ksnc_tx_queue) ||
427 msg.msg_flags |= MSG_MORE;
430 rc = sock_sendmsg(sock, &msg, nob);
437 ksocknal_lib_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
439 struct socket *sock = conn->ksnc_sock;
440 lnet_kiov_t *kiov = tx->tx_kiov;
444 /* Not NOOP message */
445 LASSERT (tx->tx_lnetmsg != NULL);
447 /* NB we can't trust socket ops to either consume our iovs
448 * or leave them alone. */
449 if (tx->tx_msg.ksm_zc_cookies[0] != 0) {
450 /* Zero copy is enabled */
451 struct sock *sk = sock->sk;
452 struct page *page = kiov->kiov_page;
453 int offset = kiov->kiov_offset;
454 int fragsize = kiov->kiov_len;
455 int msgflg = MSG_DONTWAIT;
457 CDEBUG(D_NET, "page %p + offset %x for %d\n",
458 page, offset, kiov->kiov_len);
460 if (!cfs_list_empty(&conn->ksnc_tx_queue) ||
461 fragsize < tx->tx_resid)
464 if (sk->sk_prot->sendpage != NULL) {
465 rc = sk->sk_prot->sendpage(sk, page,
466 offset, fragsize, msgflg);
468 rc = cfs_tcp_sendpage(sk, page, offset, fragsize,
472 #if SOCKNAL_SINGLE_FRAG_TX || !SOCKNAL_RISK_KMAP_DEADLOCK
473 struct iovec scratch;
474 struct iovec *scratchiov = &scratch;
475 unsigned int niov = 1;
477 #ifdef CONFIG_HIGHMEM
478 #warning "XXX risk of kmap deadlock on multiple frags..."
480 struct iovec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
481 unsigned int niov = tx->tx_nkiov;
483 struct msghdr msg = {
486 .msg_iov = scratchiov,
490 .msg_flags = MSG_DONTWAIT
492 mm_segment_t oldmm = get_fs();
495 for (nob = i = 0; i < niov; i++) {
496 scratchiov[i].iov_base = kmap(kiov[i].kiov_page) +
498 nob += scratchiov[i].iov_len = kiov[i].kiov_len;
501 if (!cfs_list_empty(&conn->ksnc_tx_queue) ||
503 msg.msg_flags |= MSG_MORE;
506 rc = sock_sendmsg(sock, &msg, nob);
509 for (i = 0; i < niov; i++)
510 kunmap(kiov[i].kiov_page);
516 ksocknal_lib_eager_ack (ksock_conn_t *conn)
519 mm_segment_t oldmm = get_fs();
520 struct socket *sock = conn->ksnc_sock;
522 /* Remind the socket to ACK eagerly. If I don't, the socket might
523 * think I'm about to send something it could piggy-back the ACK
524 * on, introducing delay in completing zero-copy sends in my
528 sock->ops->setsockopt (sock, SOL_TCP, TCP_QUICKACK,
529 (char *)&opt, sizeof (opt));
534 ksocknal_lib_recv_iov (ksock_conn_t *conn)
536 #if SOCKNAL_SINGLE_FRAG_RX
537 struct iovec scratch;
538 struct iovec *scratchiov = &scratch;
539 unsigned int niov = 1;
541 struct iovec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
542 unsigned int niov = conn->ksnc_rx_niov;
544 struct iovec *iov = conn->ksnc_rx_iov;
545 struct msghdr msg = {
548 .msg_iov = scratchiov,
554 mm_segment_t oldmm = get_fs();
562 /* NB we can't trust socket ops to either consume our iovs
563 * or leave them alone. */
566 for (nob = i = 0; i < niov; i++) {
567 scratchiov[i] = iov[i];
568 nob += scratchiov[i].iov_len;
570 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
573 rc = sock_recvmsg (conn->ksnc_sock, &msg, nob, MSG_DONTWAIT);
574 /* NB this is just a boolean..........................^ */
578 if (conn->ksnc_proto == &ksocknal_protocol_v2x) {
579 saved_csum = conn->ksnc_msg.ksm_csum;
580 conn->ksnc_msg.ksm_csum = 0;
583 if (saved_csum != 0) {
584 /* accumulate checksum */
585 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
588 fragnob = iov[i].iov_len;
592 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
593 iov[i].iov_base, fragnob);
595 conn->ksnc_msg.ksm_csum = saved_csum;
602 ksocknal_lib_kiov_vunmap(void *addr)
611 ksocknal_lib_kiov_vmap(lnet_kiov_t *kiov, int niov,
612 struct iovec *iov, struct page **pages)
618 if (!*ksocknal_tunables.ksnd_zc_recv || pages == NULL)
621 LASSERT (niov <= LNET_MAX_IOV);
624 niov < *ksocknal_tunables.ksnd_zc_recv_min_nfrags)
627 for (nob = i = 0; i < niov; i++) {
628 if ((kiov[i].kiov_offset != 0 && i > 0) ||
629 (kiov[i].kiov_offset + kiov[i].kiov_len !=
630 PAGE_CACHE_SIZE && i < niov - 1))
633 pages[i] = kiov[i].kiov_page;
634 nob += kiov[i].kiov_len;
637 addr = vmap(pages, niov, VM_MAP, PAGE_KERNEL);
641 iov->iov_base = addr + kiov[0].kiov_offset;
648 ksocknal_lib_recv_kiov (ksock_conn_t *conn)
650 #if SOCKNAL_SINGLE_FRAG_RX || !SOCKNAL_RISK_KMAP_DEADLOCK
651 struct iovec scratch;
652 struct iovec *scratchiov = &scratch;
653 struct page **pages = NULL;
654 unsigned int niov = 1;
656 #ifdef CONFIG_HIGHMEM
657 #warning "XXX risk of kmap deadlock on multiple frags..."
659 struct iovec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
660 struct page **pages = conn->ksnc_scheduler->kss_rx_scratch_pgs;
661 unsigned int niov = conn->ksnc_rx_nkiov;
663 lnet_kiov_t *kiov = conn->ksnc_rx_kiov;
664 struct msghdr msg = {
667 .msg_iov = scratchiov,
672 mm_segment_t oldmm = get_fs();
681 /* NB we can't trust socket ops to either consume our iovs
682 * or leave them alone. */
683 if ((addr = ksocknal_lib_kiov_vmap(kiov, niov, scratchiov, pages)) != NULL) {
684 nob = scratchiov[0].iov_len;
688 for (nob = i = 0; i < niov; i++) {
689 nob += scratchiov[i].iov_len = kiov[i].kiov_len;
690 scratchiov[i].iov_base = kmap(kiov[i].kiov_page) +
693 msg.msg_iovlen = niov;
696 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
699 rc = sock_recvmsg (conn->ksnc_sock, &msg, nob, MSG_DONTWAIT);
700 /* NB this is just a boolean.......................^ */
703 if (conn->ksnc_msg.ksm_csum != 0) {
704 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
707 /* Dang! have to kmap again because I have nowhere to stash the
708 * mapped address. But by doing it while the page is still
709 * mapped, the kernel just bumps the map count and returns me
710 * the address it stashed. */
711 base = kmap(kiov[i].kiov_page) + kiov[i].kiov_offset;
712 fragnob = kiov[i].kiov_len;
716 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
719 kunmap(kiov[i].kiov_page);
724 ksocknal_lib_kiov_vunmap(addr);
726 for (i = 0; i < niov; i++)
727 kunmap(kiov[i].kiov_page);
734 ksocknal_lib_csum_tx(ksock_tx_t *tx)
740 LASSERT(tx->tx_iov[0].iov_base == (void *)&tx->tx_msg);
741 LASSERT(tx->tx_conn != NULL);
742 LASSERT(tx->tx_conn->ksnc_proto == &ksocknal_protocol_v2x);
744 tx->tx_msg.ksm_csum = 0;
746 csum = ksocknal_csum(~0, (void *)tx->tx_iov[0].iov_base,
747 tx->tx_iov[0].iov_len);
749 if (tx->tx_kiov != NULL) {
750 for (i = 0; i < tx->tx_nkiov; i++) {
751 base = kmap(tx->tx_kiov[i].kiov_page) +
752 tx->tx_kiov[i].kiov_offset;
754 csum = ksocknal_csum(csum, base, tx->tx_kiov[i].kiov_len);
756 kunmap(tx->tx_kiov[i].kiov_page);
759 for (i = 1; i < tx->tx_niov; i++)
760 csum = ksocknal_csum(csum, tx->tx_iov[i].iov_base,
761 tx->tx_iov[i].iov_len);
764 if (*ksocknal_tunables.ksnd_inject_csum_error) {
766 *ksocknal_tunables.ksnd_inject_csum_error = 0;
769 tx->tx_msg.ksm_csum = csum;
773 ksocknal_lib_get_conn_tunables (ksock_conn_t *conn, int *txmem, int *rxmem, int *nagle)
775 mm_segment_t oldmm = get_fs ();
776 struct socket *sock = conn->ksnc_sock;
780 rc = ksocknal_connsock_addref(conn);
782 LASSERT (conn->ksnc_closing);
783 *txmem = *rxmem = *nagle = 0;
787 rc = libcfs_sock_getbuf(sock, txmem, rxmem);
789 len = sizeof(*nagle);
791 rc = sock->ops->getsockopt(sock, SOL_TCP, TCP_NODELAY,
792 (char *)nagle, &len);
796 ksocknal_connsock_decref(conn);
801 *txmem = *rxmem = *nagle = 0;
807 ksocknal_lib_setup_sock (struct socket *sock)
809 mm_segment_t oldmm = get_fs ();
816 struct linger linger;
818 sock->sk->sk_allocation = GFP_NOFS;
820 /* Ensure this socket aborts active sends immediately when we close
827 rc = sock_setsockopt (sock, SOL_SOCKET, SO_LINGER,
828 (char *)&linger, sizeof (linger));
831 CERROR ("Can't set SO_LINGER: %d\n", rc);
837 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_LINGER2,
838 (char *)&option, sizeof (option));
841 CERROR ("Can't set SO_LINGER2: %d\n", rc);
845 if (!*ksocknal_tunables.ksnd_nagle) {
849 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_NODELAY,
850 (char *)&option, sizeof (option));
853 CERROR ("Can't disable nagle: %d\n", rc);
858 rc = libcfs_sock_setbuf(sock,
859 *ksocknal_tunables.ksnd_tx_buffer_size,
860 *ksocknal_tunables.ksnd_rx_buffer_size);
862 CERROR ("Can't set buffer tx %d, rx %d buffers: %d\n",
863 *ksocknal_tunables.ksnd_tx_buffer_size,
864 *ksocknal_tunables.ksnd_rx_buffer_size, rc);
868 /* TCP_BACKOFF_* sockopt tunables unsupported in stock kernels */
869 #ifdef SOCKNAL_BACKOFF
870 if (*ksocknal_tunables.ksnd_backoff_init > 0) {
871 option = *ksocknal_tunables.ksnd_backoff_init;
872 #ifdef SOCKNAL_BACKOFF_MS
877 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_BACKOFF_INIT,
878 (char *)&option, sizeof (option));
881 CERROR ("Can't set initial tcp backoff %d: %d\n",
887 if (*ksocknal_tunables.ksnd_backoff_max > 0) {
888 option = *ksocknal_tunables.ksnd_backoff_max;
889 #ifdef SOCKNAL_BACKOFF_MS
894 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_BACKOFF_MAX,
895 (char *)&option, sizeof (option));
898 CERROR ("Can't set maximum tcp backoff %d: %d\n",
905 /* snapshot tunables */
906 keep_idle = *ksocknal_tunables.ksnd_keepalive_idle;
907 keep_count = *ksocknal_tunables.ksnd_keepalive_count;
908 keep_intvl = *ksocknal_tunables.ksnd_keepalive_intvl;
910 do_keepalive = (keep_idle > 0 && keep_count > 0 && keep_intvl > 0);
912 option = (do_keepalive ? 1 : 0);
914 rc = sock_setsockopt (sock, SOL_SOCKET, SO_KEEPALIVE,
915 (char *)&option, sizeof (option));
918 CERROR ("Can't set SO_KEEPALIVE: %d\n", rc);
926 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPIDLE,
927 (char *)&keep_idle, sizeof (keep_idle));
930 CERROR ("Can't set TCP_KEEPIDLE: %d\n", rc);
935 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPINTVL,
936 (char *)&keep_intvl, sizeof (keep_intvl));
939 CERROR ("Can't set TCP_KEEPINTVL: %d\n", rc);
944 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPCNT,
945 (char *)&keep_count, sizeof (keep_count));
948 CERROR ("Can't set TCP_KEEPCNT: %d\n", rc);
956 ksocknal_lib_push_conn (ksock_conn_t *conn)
965 rc = ksocknal_connsock_addref(conn);
966 if (rc != 0) /* being shut down */
969 sk = conn->ksnc_sock->sk;
973 nonagle = tp->nonagle;
980 rc = sk->sk_prot->setsockopt (sk, SOL_TCP, TCP_NODELAY,
981 (char *)&val, sizeof (val));
987 tp->nonagle = nonagle;
990 ksocknal_connsock_decref(conn);
993 extern void ksocknal_read_callback (ksock_conn_t *conn);
994 extern void ksocknal_write_callback (ksock_conn_t *conn);
996 * socket call back in Linux
999 ksocknal_data_ready (struct sock *sk, int n)
1004 /* interleave correctly with closing sockets... */
1006 read_lock(&ksocknal_data.ksnd_global_lock);
1008 conn = sk->sk_user_data;
1009 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
1010 LASSERT (sk->sk_data_ready != &ksocknal_data_ready);
1011 sk->sk_data_ready (sk, n);
1013 ksocknal_read_callback(conn);
1015 read_unlock(&ksocknal_data.ksnd_global_lock);
1021 ksocknal_write_space (struct sock *sk)
1027 /* interleave correctly with closing sockets... */
1029 read_lock(&ksocknal_data.ksnd_global_lock);
1031 conn = sk->sk_user_data;
1032 wspace = SOCKNAL_WSPACE(sk);
1033 min_wpace = SOCKNAL_MIN_WSPACE(sk);
1035 CDEBUG(D_NET, "sk %p wspace %d low water %d conn %p%s%s%s\n",
1036 sk, wspace, min_wpace, conn,
1037 (conn == NULL) ? "" : (conn->ksnc_tx_ready ?
1038 " ready" : " blocked"),
1039 (conn == NULL) ? "" : (conn->ksnc_tx_scheduled ?
1040 " scheduled" : " idle"),
1041 (conn == NULL) ? "" : (cfs_list_empty (&conn->ksnc_tx_queue) ?
1042 " empty" : " queued"));
1044 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
1045 LASSERT (sk->sk_write_space != &ksocknal_write_space);
1046 sk->sk_write_space (sk);
1048 read_unlock(&ksocknal_data.ksnd_global_lock);
1052 if (wspace >= min_wpace) { /* got enough space */
1053 ksocknal_write_callback(conn);
1055 /* Clear SOCK_NOSPACE _after_ ksocknal_write_callback so the
1056 * ENOMEM check in ksocknal_transmit is race-free (think about
1059 clear_bit (SOCK_NOSPACE, &sk->sk_socket->flags);
1062 read_unlock(&ksocknal_data.ksnd_global_lock);
1066 ksocknal_lib_save_callback(struct socket *sock, ksock_conn_t *conn)
1068 conn->ksnc_saved_data_ready = sock->sk->sk_data_ready;
1069 conn->ksnc_saved_write_space = sock->sk->sk_write_space;
1073 ksocknal_lib_set_callback(struct socket *sock, ksock_conn_t *conn)
1075 sock->sk->sk_user_data = conn;
1076 sock->sk->sk_data_ready = ksocknal_data_ready;
1077 sock->sk->sk_write_space = ksocknal_write_space;
1082 ksocknal_lib_reset_callback(struct socket *sock, ksock_conn_t *conn)
1084 /* Remove conn's network callbacks.
1085 * NB I _have_ to restore the callback, rather than storing a noop,
1086 * since the socket could survive past this module being unloaded!! */
1087 sock->sk->sk_data_ready = conn->ksnc_saved_data_ready;
1088 sock->sk->sk_write_space = conn->ksnc_saved_write_space;
1090 /* A callback could be in progress already; they hold a read lock
1091 * on ksnd_global_lock (to serialise with me) and NOOP if
1092 * sk_user_data is NULL. */
1093 sock->sk->sk_user_data = NULL;
1099 ksocknal_lib_memory_pressure(ksock_conn_t *conn)
1102 ksock_sched_t *sched;
1104 sched = conn->ksnc_scheduler;
1105 spin_lock_bh(&sched->kss_lock);
1107 if (!SOCK_TEST_NOSPACE(conn->ksnc_sock) &&
1108 !conn->ksnc_tx_ready) {
1109 /* SOCK_NOSPACE is set when the socket fills
1110 * and cleared in the write_space callback
1111 * (which also sets ksnc_tx_ready). If
1112 * SOCK_NOSPACE and ksnc_tx_ready are BOTH
1113 * zero, I didn't fill the socket and
1114 * write_space won't reschedule me, so I
1115 * return -ENOMEM to get my caller to retry
1116 * after a timeout */
1120 spin_unlock_bh(&sched->kss_lock);