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 = lnet_sock_getaddr(conn->ksnc_sock, true,
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 = lnet_sock_getaddr(conn->ksnc_sock, false,
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 = { .msg_flags = MSG_DONTWAIT };
411 for (nob = i = 0; i < niov; i++) {
412 scratchiov[i] = tx->tx_iov[i];
413 nob += scratchiov[i].iov_len;
416 if (!list_empty(&conn->ksnc_tx_queue) ||
418 msg.msg_flags |= MSG_MORE;
420 rc = kernel_sendmsg(sock, &msg, (struct kvec *)scratchiov,
427 ksocknal_lib_send_kiov(ksock_conn_t *conn, ksock_tx_t *tx)
429 struct socket *sock = conn->ksnc_sock;
430 lnet_kiov_t *kiov = tx->tx_kiov;
434 /* Not NOOP message */
435 LASSERT (tx->tx_lnetmsg != NULL);
437 /* NB we can't trust socket ops to either consume our iovs
438 * or leave them alone. */
439 if (tx->tx_msg.ksm_zc_cookies[0] != 0) {
440 /* Zero copy is enabled */
441 struct sock *sk = sock->sk;
442 struct page *page = kiov->kiov_page;
443 int offset = kiov->kiov_offset;
444 int fragsize = kiov->kiov_len;
445 int msgflg = MSG_DONTWAIT;
447 CDEBUG(D_NET, "page %p + offset %x for %d\n",
448 page, offset, kiov->kiov_len);
450 if (!list_empty(&conn->ksnc_tx_queue) ||
451 fragsize < tx->tx_resid)
454 if (sk->sk_prot->sendpage != NULL) {
455 rc = sk->sk_prot->sendpage(sk, page,
456 offset, fragsize, msgflg);
458 rc = cfs_tcp_sendpage(sk, page, offset, fragsize,
462 #if SOCKNAL_SINGLE_FRAG_TX || !SOCKNAL_RISK_KMAP_DEADLOCK
463 struct iovec scratch;
464 struct iovec *scratchiov = &scratch;
465 unsigned int niov = 1;
467 #ifdef CONFIG_HIGHMEM
468 #warning "XXX risk of kmap deadlock on multiple frags..."
470 struct iovec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
471 unsigned int niov = tx->tx_nkiov;
473 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
476 for (nob = i = 0; i < niov; i++) {
477 scratchiov[i].iov_base = kmap(kiov[i].kiov_page) +
479 nob += scratchiov[i].iov_len = kiov[i].kiov_len;
482 if (!list_empty(&conn->ksnc_tx_queue) ||
484 msg.msg_flags |= MSG_MORE;
486 rc = kernel_sendmsg(sock, &msg, (struct kvec *)scratchiov, niov, nob);
488 for (i = 0; i < niov; i++)
489 kunmap(kiov[i].kiov_page);
495 ksocknal_lib_eager_ack (ksock_conn_t *conn)
498 mm_segment_t oldmm = get_fs();
499 struct socket *sock = conn->ksnc_sock;
501 /* Remind the socket to ACK eagerly. If I don't, the socket might
502 * think I'm about to send something it could piggy-back the ACK
503 * on, introducing delay in completing zero-copy sends in my
507 sock->ops->setsockopt (sock, SOL_TCP, TCP_QUICKACK,
508 (char *)&opt, sizeof (opt));
513 ksocknal_lib_recv_iov (ksock_conn_t *conn)
515 #if SOCKNAL_SINGLE_FRAG_RX
516 struct iovec scratch;
517 struct iovec *scratchiov = &scratch;
518 unsigned int niov = 1;
520 struct iovec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
521 unsigned int niov = conn->ksnc_rx_niov;
523 struct iovec *iov = conn->ksnc_rx_iov;
524 struct msghdr msg = {
534 /* NB we can't trust socket ops to either consume our iovs
535 * or leave them alone. */
538 for (nob = i = 0; i < niov; i++) {
539 scratchiov[i] = iov[i];
540 nob += scratchiov[i].iov_len;
542 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
544 rc = kernel_recvmsg(conn->ksnc_sock, &msg,
545 (struct kvec *)scratchiov, niov, nob, MSG_DONTWAIT);
548 if (conn->ksnc_proto == &ksocknal_protocol_v2x) {
549 saved_csum = conn->ksnc_msg.ksm_csum;
550 conn->ksnc_msg.ksm_csum = 0;
553 if (saved_csum != 0) {
554 /* accumulate checksum */
555 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
558 fragnob = iov[i].iov_len;
562 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
563 iov[i].iov_base, fragnob);
565 conn->ksnc_msg.ksm_csum = saved_csum;
572 ksocknal_lib_kiov_vunmap(void *addr)
581 ksocknal_lib_kiov_vmap(lnet_kiov_t *kiov, int niov,
582 struct iovec *iov, struct page **pages)
588 if (!*ksocknal_tunables.ksnd_zc_recv || pages == NULL)
591 LASSERT (niov <= LNET_MAX_IOV);
594 niov < *ksocknal_tunables.ksnd_zc_recv_min_nfrags)
597 for (nob = i = 0; i < niov; i++) {
598 if ((kiov[i].kiov_offset != 0 && i > 0) ||
599 (kiov[i].kiov_offset + kiov[i].kiov_len !=
600 PAGE_CACHE_SIZE && i < niov - 1))
603 pages[i] = kiov[i].kiov_page;
604 nob += kiov[i].kiov_len;
607 addr = vmap(pages, niov, VM_MAP, PAGE_KERNEL);
611 iov->iov_base = addr + kiov[0].kiov_offset;
618 ksocknal_lib_recv_kiov (ksock_conn_t *conn)
620 #if SOCKNAL_SINGLE_FRAG_RX || !SOCKNAL_RISK_KMAP_DEADLOCK
621 struct iovec scratch;
622 struct iovec *scratchiov = &scratch;
623 struct page **pages = NULL;
624 unsigned int niov = 1;
626 #ifdef CONFIG_HIGHMEM
627 #warning "XXX risk of kmap deadlock on multiple frags..."
629 struct iovec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
630 struct page **pages = conn->ksnc_scheduler->kss_rx_scratch_pgs;
631 unsigned int niov = conn->ksnc_rx_nkiov;
633 lnet_kiov_t *kiov = conn->ksnc_rx_kiov;
634 struct msghdr msg = {
646 /* NB we can't trust socket ops to either consume our iovs
647 * or leave them alone. */
648 if ((addr = ksocknal_lib_kiov_vmap(kiov, niov, scratchiov, pages)) != NULL) {
649 nob = scratchiov[0].iov_len;
653 for (nob = i = 0; i < niov; i++) {
654 nob += scratchiov[i].iov_len = kiov[i].kiov_len;
655 scratchiov[i].iov_base = kmap(kiov[i].kiov_page) +
661 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
663 rc = kernel_recvmsg(conn->ksnc_sock, &msg,
664 (struct kvec *)scratchiov, n, nob, MSG_DONTWAIT);
666 if (conn->ksnc_msg.ksm_csum != 0) {
667 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
670 /* Dang! have to kmap again because I have nowhere to stash the
671 * mapped address. But by doing it while the page is still
672 * mapped, the kernel just bumps the map count and returns me
673 * the address it stashed. */
674 base = kmap(kiov[i].kiov_page) + kiov[i].kiov_offset;
675 fragnob = kiov[i].kiov_len;
679 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
682 kunmap(kiov[i].kiov_page);
687 ksocknal_lib_kiov_vunmap(addr);
689 for (i = 0; i < niov; i++)
690 kunmap(kiov[i].kiov_page);
697 ksocknal_lib_csum_tx(ksock_tx_t *tx)
703 LASSERT(tx->tx_iov[0].iov_base == (void *)&tx->tx_msg);
704 LASSERT(tx->tx_conn != NULL);
705 LASSERT(tx->tx_conn->ksnc_proto == &ksocknal_protocol_v2x);
707 tx->tx_msg.ksm_csum = 0;
709 csum = ksocknal_csum(~0, (void *)tx->tx_iov[0].iov_base,
710 tx->tx_iov[0].iov_len);
712 if (tx->tx_kiov != NULL) {
713 for (i = 0; i < tx->tx_nkiov; i++) {
714 base = kmap(tx->tx_kiov[i].kiov_page) +
715 tx->tx_kiov[i].kiov_offset;
717 csum = ksocknal_csum(csum, base, tx->tx_kiov[i].kiov_len);
719 kunmap(tx->tx_kiov[i].kiov_page);
722 for (i = 1; i < tx->tx_niov; i++)
723 csum = ksocknal_csum(csum, tx->tx_iov[i].iov_base,
724 tx->tx_iov[i].iov_len);
727 if (*ksocknal_tunables.ksnd_inject_csum_error) {
729 *ksocknal_tunables.ksnd_inject_csum_error = 0;
732 tx->tx_msg.ksm_csum = csum;
736 ksocknal_lib_get_conn_tunables (ksock_conn_t *conn, int *txmem, int *rxmem, int *nagle)
738 mm_segment_t oldmm = get_fs ();
739 struct socket *sock = conn->ksnc_sock;
743 rc = ksocknal_connsock_addref(conn);
745 LASSERT (conn->ksnc_closing);
746 *txmem = *rxmem = *nagle = 0;
750 rc = lnet_sock_getbuf(sock, txmem, rxmem);
752 len = sizeof(*nagle);
754 rc = sock->ops->getsockopt(sock, SOL_TCP, TCP_NODELAY,
755 (char *)nagle, &len);
759 ksocknal_connsock_decref(conn);
764 *txmem = *rxmem = *nagle = 0;
770 ksocknal_lib_setup_sock (struct socket *sock)
772 mm_segment_t oldmm = get_fs ();
779 struct linger linger;
781 sock->sk->sk_allocation = GFP_NOFS;
783 /* Ensure this socket aborts active sends immediately when we close
790 rc = sock_setsockopt (sock, SOL_SOCKET, SO_LINGER,
791 (char *)&linger, sizeof (linger));
794 CERROR ("Can't set SO_LINGER: %d\n", rc);
800 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_LINGER2,
801 (char *)&option, sizeof (option));
804 CERROR ("Can't set SO_LINGER2: %d\n", rc);
808 if (!*ksocknal_tunables.ksnd_nagle) {
812 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_NODELAY,
813 (char *)&option, sizeof (option));
816 CERROR ("Can't disable nagle: %d\n", rc);
821 rc = lnet_sock_setbuf(sock,
822 *ksocknal_tunables.ksnd_tx_buffer_size,
823 *ksocknal_tunables.ksnd_rx_buffer_size);
825 CERROR ("Can't set buffer tx %d, rx %d buffers: %d\n",
826 *ksocknal_tunables.ksnd_tx_buffer_size,
827 *ksocknal_tunables.ksnd_rx_buffer_size, rc);
831 /* TCP_BACKOFF_* sockopt tunables unsupported in stock kernels */
832 #ifdef SOCKNAL_BACKOFF
833 if (*ksocknal_tunables.ksnd_backoff_init > 0) {
834 option = *ksocknal_tunables.ksnd_backoff_init;
835 #ifdef SOCKNAL_BACKOFF_MS
840 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_BACKOFF_INIT,
841 (char *)&option, sizeof (option));
844 CERROR ("Can't set initial tcp backoff %d: %d\n",
850 if (*ksocknal_tunables.ksnd_backoff_max > 0) {
851 option = *ksocknal_tunables.ksnd_backoff_max;
852 #ifdef SOCKNAL_BACKOFF_MS
857 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_BACKOFF_MAX,
858 (char *)&option, sizeof (option));
861 CERROR ("Can't set maximum tcp backoff %d: %d\n",
868 /* snapshot tunables */
869 keep_idle = *ksocknal_tunables.ksnd_keepalive_idle;
870 keep_count = *ksocknal_tunables.ksnd_keepalive_count;
871 keep_intvl = *ksocknal_tunables.ksnd_keepalive_intvl;
873 do_keepalive = (keep_idle > 0 && keep_count > 0 && keep_intvl > 0);
875 option = (do_keepalive ? 1 : 0);
877 rc = sock_setsockopt (sock, SOL_SOCKET, SO_KEEPALIVE,
878 (char *)&option, sizeof (option));
881 CERROR ("Can't set SO_KEEPALIVE: %d\n", rc);
889 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPIDLE,
890 (char *)&keep_idle, sizeof (keep_idle));
893 CERROR ("Can't set TCP_KEEPIDLE: %d\n", rc);
898 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPINTVL,
899 (char *)&keep_intvl, sizeof (keep_intvl));
902 CERROR ("Can't set TCP_KEEPINTVL: %d\n", rc);
907 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPCNT,
908 (char *)&keep_count, sizeof (keep_count));
911 CERROR ("Can't set TCP_KEEPCNT: %d\n", rc);
919 ksocknal_lib_push_conn (ksock_conn_t *conn)
928 rc = ksocknal_connsock_addref(conn);
929 if (rc != 0) /* being shut down */
932 sk = conn->ksnc_sock->sk;
936 nonagle = tp->nonagle;
943 rc = sk->sk_prot->setsockopt (sk, SOL_TCP, TCP_NODELAY,
944 (char *)&val, sizeof (val));
950 tp->nonagle = nonagle;
953 ksocknal_connsock_decref(conn);
956 extern void ksocknal_read_callback (ksock_conn_t *conn);
957 extern void ksocknal_write_callback (ksock_conn_t *conn);
959 * socket call back in Linux
962 #ifdef HAVE_SK_DATA_READY_ONE_ARG
963 ksocknal_data_ready(struct sock *sk)
965 ksocknal_data_ready(struct sock *sk, int n)
971 /* interleave correctly with closing sockets... */
973 read_lock(&ksocknal_data.ksnd_global_lock);
975 conn = sk->sk_user_data;
976 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
977 LASSERT(sk->sk_data_ready != &ksocknal_data_ready);
978 #ifdef HAVE_SK_DATA_READY_ONE_ARG
979 sk->sk_data_ready(sk);
981 sk->sk_data_ready(sk, n);
984 ksocknal_read_callback(conn);
986 read_unlock(&ksocknal_data.ksnd_global_lock);
992 ksocknal_write_space (struct sock *sk)
998 /* interleave correctly with closing sockets... */
1000 read_lock(&ksocknal_data.ksnd_global_lock);
1002 conn = sk->sk_user_data;
1003 wspace = sk_stream_wspace(sk);
1004 min_wpace = sk_stream_min_wspace(sk);
1006 CDEBUG(D_NET, "sk %p wspace %d low water %d conn %p%s%s%s\n",
1007 sk, wspace, min_wpace, conn,
1008 (conn == NULL) ? "" : (conn->ksnc_tx_ready ?
1009 " ready" : " blocked"),
1010 (conn == NULL) ? "" : (conn->ksnc_tx_scheduled ?
1011 " scheduled" : " idle"),
1012 (conn == NULL) ? "" : (list_empty(&conn->ksnc_tx_queue) ?
1013 " empty" : " queued"));
1015 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
1016 LASSERT (sk->sk_write_space != &ksocknal_write_space);
1017 sk->sk_write_space (sk);
1019 read_unlock(&ksocknal_data.ksnd_global_lock);
1023 if (wspace >= min_wpace) { /* got enough space */
1024 ksocknal_write_callback(conn);
1026 /* Clear SOCK_NOSPACE _after_ ksocknal_write_callback so the
1027 * ENOMEM check in ksocknal_transmit is race-free (think about
1030 clear_bit (SOCK_NOSPACE, &sk->sk_socket->flags);
1033 read_unlock(&ksocknal_data.ksnd_global_lock);
1037 ksocknal_lib_save_callback(struct socket *sock, ksock_conn_t *conn)
1039 conn->ksnc_saved_data_ready = sock->sk->sk_data_ready;
1040 conn->ksnc_saved_write_space = sock->sk->sk_write_space;
1044 ksocknal_lib_set_callback(struct socket *sock, ksock_conn_t *conn)
1046 sock->sk->sk_user_data = conn;
1047 sock->sk->sk_data_ready = ksocknal_data_ready;
1048 sock->sk->sk_write_space = ksocknal_write_space;
1053 ksocknal_lib_reset_callback(struct socket *sock, ksock_conn_t *conn)
1055 /* Remove conn's network callbacks.
1056 * NB I _have_ to restore the callback, rather than storing a noop,
1057 * since the socket could survive past this module being unloaded!! */
1058 sock->sk->sk_data_ready = conn->ksnc_saved_data_ready;
1059 sock->sk->sk_write_space = conn->ksnc_saved_write_space;
1061 /* A callback could be in progress already; they hold a read lock
1062 * on ksnd_global_lock (to serialise with me) and NOOP if
1063 * sk_user_data is NULL. */
1064 sock->sk->sk_user_data = NULL;
1070 ksocknal_lib_memory_pressure(ksock_conn_t *conn)
1073 ksock_sched_t *sched;
1075 sched = conn->ksnc_scheduler;
1076 spin_lock_bh(&sched->kss_lock);
1078 if (!test_bit(SOCK_NOSPACE, &conn->ksnc_sock->flags) &&
1079 !conn->ksnc_tx_ready) {
1080 /* SOCK_NOSPACE is set when the socket fills
1081 * and cleared in the write_space callback
1082 * (which also sets ksnc_tx_ready). If
1083 * SOCK_NOSPACE and ksnc_tx_ready are BOTH
1084 * zero, I didn't fill the socket and
1085 * write_space won't reschedule me, so I
1086 * return -ENOMEM to get my caller to retry
1087 * after a timeout */
1091 spin_unlock_bh(&sched->kss_lock);