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.
40 ksocknal_lib_get_conn_addrs (ksock_conn_t *conn)
42 int rc = libcfs_sock_getaddr(conn->ksnc_sock, 1,
46 /* Didn't need the {get,put}connsock dance to deref ksnc_sock... */
47 LASSERT (!conn->ksnc_closing);
50 CERROR ("Error %d getting sock peer IP\n", rc);
54 rc = libcfs_sock_getaddr(conn->ksnc_sock, 0,
55 &conn->ksnc_myipaddr, NULL);
57 CERROR ("Error %d getting sock local IP\n", rc);
65 ksocknal_lib_zc_capable(ksock_conn_t *conn)
67 int caps = conn->ksnc_sock->sk->sk_route_caps;
69 if (conn->ksnc_proto == &ksocknal_protocol_v1x)
72 /* ZC if the socket supports scatter/gather and doesn't need software
74 return ((caps & NETIF_F_SG) != 0 && (caps & NETIF_F_ALL_CSUM) != 0);
78 ksocknal_lib_send_iov (ksock_conn_t *conn, ksock_tx_t *tx)
80 struct socket *sock = conn->ksnc_sock;
84 if (*ksocknal_tunables.ksnd_enable_csum && /* checksum enabled */
85 conn->ksnc_proto == &ksocknal_protocol_v2x && /* V2.x connection */
86 tx->tx_nob == tx->tx_resid && /* frist sending */
87 tx->tx_msg.ksm_csum == 0) /* not checksummed */
88 ksocknal_lib_csum_tx(tx);
90 /* NB we can't trust socket ops to either consume our iovs
91 * or leave them alone. */
94 #if SOCKNAL_SINGLE_FRAG_TX
96 struct iovec *scratchiov = &scratch;
97 unsigned int niov = 1;
99 struct iovec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
100 unsigned int niov = tx->tx_niov;
102 struct msghdr msg = {
105 .msg_iov = scratchiov,
109 .msg_flags = MSG_DONTWAIT
111 mm_segment_t oldmm = get_fs();
114 for (nob = i = 0; i < niov; i++) {
115 scratchiov[i] = tx->tx_iov[i];
116 nob += scratchiov[i].iov_len;
119 if (!list_empty(&conn->ksnc_tx_queue) ||
121 msg.msg_flags |= MSG_MORE;
124 rc = sock_sendmsg(sock, &msg, nob);
131 ksocknal_lib_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
133 struct socket *sock = conn->ksnc_sock;
134 lnet_kiov_t *kiov = tx->tx_kiov;
138 /* Not NOOP message */
139 LASSERT (tx->tx_lnetmsg != NULL);
141 /* NB we can't trust socket ops to either consume our iovs
142 * or leave them alone. */
143 if (tx->tx_msg.ksm_zc_cookies[0] != 0) {
144 /* Zero copy is enabled */
145 struct sock *sk = sock->sk;
146 struct page *page = kiov->kiov_page;
147 int offset = kiov->kiov_offset;
148 int fragsize = kiov->kiov_len;
149 int msgflg = MSG_DONTWAIT;
151 CDEBUG(D_NET, "page %p + offset %x for %d\n",
152 page, offset, kiov->kiov_len);
154 if (!list_empty(&conn->ksnc_tx_queue) ||
155 fragsize < tx->tx_resid)
158 if (sk->sk_prot->sendpage != NULL) {
159 rc = sk->sk_prot->sendpage(sk, page,
160 offset, fragsize, msgflg);
162 rc = cfs_tcp_sendpage(sk, page, offset, fragsize,
166 #if SOCKNAL_SINGLE_FRAG_TX || !SOCKNAL_RISK_KMAP_DEADLOCK
167 struct iovec scratch;
168 struct iovec *scratchiov = &scratch;
169 unsigned int niov = 1;
171 #ifdef CONFIG_HIGHMEM
172 #warning "XXX risk of kmap deadlock on multiple frags..."
174 struct iovec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
175 unsigned int niov = tx->tx_nkiov;
177 struct msghdr msg = {
180 .msg_iov = scratchiov,
184 .msg_flags = MSG_DONTWAIT
186 mm_segment_t oldmm = get_fs();
189 for (nob = i = 0; i < niov; i++) {
190 scratchiov[i].iov_base = kmap(kiov[i].kiov_page) +
192 nob += scratchiov[i].iov_len = kiov[i].kiov_len;
195 if (!list_empty(&conn->ksnc_tx_queue) ||
197 msg.msg_flags |= MSG_MORE;
200 rc = sock_sendmsg(sock, &msg, nob);
203 for (i = 0; i < niov; i++)
204 kunmap(kiov[i].kiov_page);
210 ksocknal_lib_eager_ack (ksock_conn_t *conn)
213 mm_segment_t oldmm = get_fs();
214 struct socket *sock = conn->ksnc_sock;
216 /* Remind the socket to ACK eagerly. If I don't, the socket might
217 * think I'm about to send something it could piggy-back the ACK
218 * on, introducing delay in completing zero-copy sends in my
222 sock->ops->setsockopt (sock, SOL_TCP, TCP_QUICKACK,
223 (char *)&opt, sizeof (opt));
228 ksocknal_lib_recv_iov (ksock_conn_t *conn)
230 #if SOCKNAL_SINGLE_FRAG_RX
231 struct iovec scratch;
232 struct iovec *scratchiov = &scratch;
233 unsigned int niov = 1;
235 struct iovec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
236 unsigned int niov = conn->ksnc_rx_niov;
238 struct iovec *iov = conn->ksnc_rx_iov;
239 struct msghdr msg = {
242 .msg_iov = scratchiov,
248 mm_segment_t oldmm = get_fs();
256 /* NB we can't trust socket ops to either consume our iovs
257 * or leave them alone. */
260 for (nob = i = 0; i < niov; i++) {
261 scratchiov[i] = iov[i];
262 nob += scratchiov[i].iov_len;
264 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
267 rc = sock_recvmsg (conn->ksnc_sock, &msg, nob, MSG_DONTWAIT);
268 /* NB this is just a boolean..........................^ */
272 if (conn->ksnc_proto == &ksocknal_protocol_v2x) {
273 saved_csum = conn->ksnc_msg.ksm_csum;
274 conn->ksnc_msg.ksm_csum = 0;
277 if (saved_csum != 0) {
278 /* accumulate checksum */
279 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
282 fragnob = iov[i].iov_len;
286 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
287 iov[i].iov_base, fragnob);
289 conn->ksnc_msg.ksm_csum = saved_csum;
296 ksocknal_lib_kiov_vunmap(void *addr)
305 ksocknal_lib_kiov_vmap(lnet_kiov_t *kiov, int niov,
306 struct iovec *iov, struct page **pages)
312 if (!*ksocknal_tunables.ksnd_zc_recv || pages == NULL)
315 LASSERT (niov <= LNET_MAX_IOV);
318 niov < *ksocknal_tunables.ksnd_zc_recv_min_nfrags)
321 for (nob = i = 0; i < niov; i++) {
322 if ((kiov[i].kiov_offset != 0 && i > 0) ||
323 (kiov[i].kiov_offset + kiov[i].kiov_len != PAGE_CACHE_SIZE && i < niov - 1))
326 pages[i] = kiov[i].kiov_page;
327 nob += kiov[i].kiov_len;
330 addr = vmap(pages, niov, VM_MAP, PAGE_KERNEL);
334 iov->iov_base = addr + kiov[0].kiov_offset;
341 ksocknal_lib_recv_kiov (ksock_conn_t *conn)
343 #if SOCKNAL_SINGLE_FRAG_RX || !SOCKNAL_RISK_KMAP_DEADLOCK
344 struct iovec scratch;
345 struct iovec *scratchiov = &scratch;
346 struct page **pages = NULL;
347 unsigned int niov = 1;
349 #ifdef CONFIG_HIGHMEM
350 #warning "XXX risk of kmap deadlock on multiple frags..."
352 struct iovec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
353 struct page **pages = conn->ksnc_scheduler->kss_rx_scratch_pgs;
354 unsigned int niov = conn->ksnc_rx_nkiov;
356 lnet_kiov_t *kiov = conn->ksnc_rx_kiov;
357 struct msghdr msg = {
360 .msg_iov = scratchiov,
365 mm_segment_t oldmm = get_fs();
374 /* NB we can't trust socket ops to either consume our iovs
375 * or leave them alone. */
376 if ((addr = ksocknal_lib_kiov_vmap(kiov, niov, scratchiov, pages)) != NULL) {
377 nob = scratchiov[0].iov_len;
381 for (nob = i = 0; i < niov; i++) {
382 nob += scratchiov[i].iov_len = kiov[i].kiov_len;
383 scratchiov[i].iov_base = kmap(kiov[i].kiov_page) +
386 msg.msg_iovlen = niov;
389 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
392 rc = sock_recvmsg (conn->ksnc_sock, &msg, nob, MSG_DONTWAIT);
393 /* NB this is just a boolean.......................^ */
396 if (conn->ksnc_msg.ksm_csum != 0) {
397 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
400 /* Dang! have to kmap again because I have nowhere to stash the
401 * mapped address. But by doing it while the page is still
402 * mapped, the kernel just bumps the map count and returns me
403 * the address it stashed. */
404 base = kmap(kiov[i].kiov_page) + kiov[i].kiov_offset;
405 fragnob = kiov[i].kiov_len;
409 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
412 kunmap(kiov[i].kiov_page);
417 ksocknal_lib_kiov_vunmap(addr);
419 for (i = 0; i < niov; i++)
420 kunmap(kiov[i].kiov_page);
427 ksocknal_lib_csum_tx(ksock_tx_t *tx)
433 LASSERT(tx->tx_iov[0].iov_base == (void *)&tx->tx_msg);
434 LASSERT(tx->tx_conn != NULL);
435 LASSERT(tx->tx_conn->ksnc_proto == &ksocknal_protocol_v2x);
437 tx->tx_msg.ksm_csum = 0;
439 csum = ksocknal_csum(~0, (void *)tx->tx_iov[0].iov_base,
440 tx->tx_iov[0].iov_len);
442 if (tx->tx_kiov != NULL) {
443 for (i = 0; i < tx->tx_nkiov; i++) {
444 base = kmap(tx->tx_kiov[i].kiov_page) +
445 tx->tx_kiov[i].kiov_offset;
447 csum = ksocknal_csum(csum, base, tx->tx_kiov[i].kiov_len);
449 kunmap(tx->tx_kiov[i].kiov_page);
452 for (i = 1; i < tx->tx_niov; i++)
453 csum = ksocknal_csum(csum, tx->tx_iov[i].iov_base,
454 tx->tx_iov[i].iov_len);
457 if (*ksocknal_tunables.ksnd_inject_csum_error) {
459 *ksocknal_tunables.ksnd_inject_csum_error = 0;
462 tx->tx_msg.ksm_csum = csum;
466 ksocknal_lib_get_conn_tunables (ksock_conn_t *conn, int *txmem, int *rxmem, int *nagle)
468 mm_segment_t oldmm = get_fs ();
469 struct socket *sock = conn->ksnc_sock;
473 rc = ksocknal_connsock_addref(conn);
475 LASSERT (conn->ksnc_closing);
476 *txmem = *rxmem = *nagle = 0;
480 rc = libcfs_sock_getbuf(sock, txmem, rxmem);
482 len = sizeof(*nagle);
484 rc = sock->ops->getsockopt(sock, SOL_TCP, TCP_NODELAY,
485 (char *)nagle, &len);
489 ksocknal_connsock_decref(conn);
494 *txmem = *rxmem = *nagle = 0;
500 ksocknal_lib_setup_sock (struct socket *sock)
502 mm_segment_t oldmm = get_fs ();
509 struct linger linger;
511 sock->sk->sk_allocation = GFP_NOFS;
513 /* Ensure this socket aborts active sends immediately when we close
520 rc = sock_setsockopt (sock, SOL_SOCKET, SO_LINGER,
521 (char *)&linger, sizeof (linger));
524 CERROR ("Can't set SO_LINGER: %d\n", rc);
530 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_LINGER2,
531 (char *)&option, sizeof (option));
534 CERROR ("Can't set SO_LINGER2: %d\n", rc);
538 if (!*ksocknal_tunables.ksnd_nagle) {
542 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_NODELAY,
543 (char *)&option, sizeof (option));
546 CERROR ("Can't disable nagle: %d\n", rc);
551 rc = libcfs_sock_setbuf(sock,
552 *ksocknal_tunables.ksnd_tx_buffer_size,
553 *ksocknal_tunables.ksnd_rx_buffer_size);
555 CERROR ("Can't set buffer tx %d, rx %d buffers: %d\n",
556 *ksocknal_tunables.ksnd_tx_buffer_size,
557 *ksocknal_tunables.ksnd_rx_buffer_size, rc);
561 /* TCP_BACKOFF_* sockopt tunables unsupported in stock kernels */
563 /* snapshot tunables */
564 keep_idle = *ksocknal_tunables.ksnd_keepalive_idle;
565 keep_count = *ksocknal_tunables.ksnd_keepalive_count;
566 keep_intvl = *ksocknal_tunables.ksnd_keepalive_intvl;
568 do_keepalive = (keep_idle > 0 && keep_count > 0 && keep_intvl > 0);
570 option = (do_keepalive ? 1 : 0);
572 rc = sock_setsockopt (sock, SOL_SOCKET, SO_KEEPALIVE,
573 (char *)&option, sizeof (option));
576 CERROR ("Can't set SO_KEEPALIVE: %d\n", rc);
584 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPIDLE,
585 (char *)&keep_idle, sizeof (keep_idle));
588 CERROR ("Can't set TCP_KEEPIDLE: %d\n", rc);
593 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPINTVL,
594 (char *)&keep_intvl, sizeof (keep_intvl));
597 CERROR ("Can't set TCP_KEEPINTVL: %d\n", rc);
602 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPCNT,
603 (char *)&keep_count, sizeof (keep_count));
606 CERROR ("Can't set TCP_KEEPCNT: %d\n", rc);
614 ksocknal_lib_push_conn (ksock_conn_t *conn)
623 rc = ksocknal_connsock_addref(conn);
624 if (rc != 0) /* being shut down */
627 sk = conn->ksnc_sock->sk;
631 nonagle = tp->nonagle;
638 rc = sk->sk_prot->setsockopt (sk, SOL_TCP, TCP_NODELAY,
639 (char *)&val, sizeof (val));
645 tp->nonagle = nonagle;
648 ksocknal_connsock_decref(conn);
651 extern void ksocknal_read_callback (ksock_conn_t *conn);
652 extern void ksocknal_write_callback (ksock_conn_t *conn);
654 * socket call back in Linux
657 ksocknal_data_ready (struct sock *sk, int n)
661 /* interleave correctly with closing sockets... */
663 read_lock(&ksocknal_data.ksnd_global_lock);
665 conn = sk->sk_user_data;
666 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
667 LASSERT (sk->sk_data_ready != &ksocknal_data_ready);
668 sk->sk_data_ready (sk, n);
670 ksocknal_read_callback(conn);
672 read_unlock(&ksocknal_data.ksnd_global_lock);
676 ksocknal_write_space (struct sock *sk)
682 /* interleave correctly with closing sockets... */
684 read_lock(&ksocknal_data.ksnd_global_lock);
686 conn = sk->sk_user_data;
687 wspace = SOCKNAL_WSPACE(sk);
688 min_wpace = SOCKNAL_MIN_WSPACE(sk);
690 CDEBUG(D_NET, "sk %p wspace %d low water %d conn %p%s%s%s\n",
691 sk, wspace, min_wpace, conn,
692 (conn == NULL) ? "" : (conn->ksnc_tx_ready ?
693 " ready" : " blocked"),
694 (conn == NULL) ? "" : (conn->ksnc_tx_scheduled ?
695 " scheduled" : " idle"),
696 (conn == NULL) ? "" : (list_empty (&conn->ksnc_tx_queue) ?
697 " empty" : " queued"));
699 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
700 LASSERT (sk->sk_write_space != &ksocknal_write_space);
701 sk->sk_write_space (sk);
703 read_unlock(&ksocknal_data.ksnd_global_lock);
707 if (wspace >= min_wpace) { /* got enough space */
708 ksocknal_write_callback(conn);
710 /* Clear SOCK_NOSPACE _after_ ksocknal_write_callback so the
711 * ENOMEM check in ksocknal_transmit is race-free (think about
714 clear_bit (SOCK_NOSPACE, &sk->sk_socket->flags);
717 read_unlock(&ksocknal_data.ksnd_global_lock);
721 ksocknal_lib_save_callback(struct socket *sock, ksock_conn_t *conn)
723 conn->ksnc_saved_data_ready = sock->sk->sk_data_ready;
724 conn->ksnc_saved_write_space = sock->sk->sk_write_space;
728 ksocknal_lib_set_callback(struct socket *sock, ksock_conn_t *conn)
730 sock->sk->sk_user_data = conn;
731 sock->sk->sk_data_ready = ksocknal_data_ready;
732 sock->sk->sk_write_space = ksocknal_write_space;
737 ksocknal_lib_reset_callback(struct socket *sock, ksock_conn_t *conn)
739 /* Remove conn's network callbacks.
740 * NB I _have_ to restore the callback, rather than storing a noop,
741 * since the socket could survive past this module being unloaded!! */
742 sock->sk->sk_data_ready = conn->ksnc_saved_data_ready;
743 sock->sk->sk_write_space = conn->ksnc_saved_write_space;
745 /* A callback could be in progress already; they hold a read lock
746 * on ksnd_global_lock (to serialise with me) and NOOP if
747 * sk_user_data is NULL. */
748 sock->sk->sk_user_data = NULL;
754 ksocknal_lib_memory_pressure(ksock_conn_t *conn)
757 ksock_sched_t *sched;
759 sched = conn->ksnc_scheduler;
760 spin_lock_bh(&sched->kss_lock);
762 if (!SOCK_TEST_NOSPACE(conn->ksnc_sock) &&
763 !conn->ksnc_tx_ready) {
764 /* SOCK_NOSPACE is set when the socket fills
765 * and cleared in the write_space callback
766 * (which also sets ksnc_tx_ready). If
767 * SOCK_NOSPACE and ksnc_tx_ready are BOTH
768 * zero, I didn't fill the socket and
769 * write_space won't reschedule me, so I
770 * return -ENOMEM to get my caller to retry
775 spin_unlock_bh(&sched->kss_lock);