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 = {.msg_flags = MSG_DONTWAIT};
105 for (nob = i = 0; i < niov; i++) {
106 scratchiov[i] = tx->tx_iov[i];
107 nob += scratchiov[i].iov_len;
110 if (!list_empty(&conn->ksnc_tx_queue) ||
112 msg.msg_flags |= MSG_MORE;
114 rc = kernel_sendmsg(sock, &msg, (struct kvec *)scratchiov, niov, nob);
120 ksocknal_lib_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
122 struct socket *sock = conn->ksnc_sock;
123 lnet_kiov_t *kiov = tx->tx_kiov;
127 /* Not NOOP message */
128 LASSERT (tx->tx_lnetmsg != NULL);
130 /* NB we can't trust socket ops to either consume our iovs
131 * or leave them alone. */
132 if (tx->tx_msg.ksm_zc_cookies[0] != 0) {
133 /* Zero copy is enabled */
134 struct sock *sk = sock->sk;
135 struct page *page = kiov->kiov_page;
136 int offset = kiov->kiov_offset;
137 int fragsize = kiov->kiov_len;
138 int msgflg = MSG_DONTWAIT;
140 CDEBUG(D_NET, "page %p + offset %x for %d\n",
141 page, offset, kiov->kiov_len);
143 if (!list_empty(&conn->ksnc_tx_queue) ||
144 fragsize < tx->tx_resid)
147 if (sk->sk_prot->sendpage != NULL) {
148 rc = sk->sk_prot->sendpage(sk, page,
149 offset, fragsize, msgflg);
151 rc = cfs_tcp_sendpage(sk, page, offset, fragsize,
155 #if SOCKNAL_SINGLE_FRAG_TX || !SOCKNAL_RISK_KMAP_DEADLOCK
156 struct iovec scratch;
157 struct iovec *scratchiov = &scratch;
158 unsigned int niov = 1;
160 #ifdef CONFIG_HIGHMEM
161 #warning "XXX risk of kmap deadlock on multiple frags..."
163 struct iovec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
164 unsigned int niov = tx->tx_nkiov;
166 struct msghdr msg = {.msg_flags = MSG_DONTWAIT};
169 for (nob = i = 0; i < niov; i++) {
170 scratchiov[i].iov_base = kmap(kiov[i].kiov_page) +
172 nob += scratchiov[i].iov_len = kiov[i].kiov_len;
175 if (!list_empty(&conn->ksnc_tx_queue) ||
177 msg.msg_flags |= MSG_MORE;
179 rc = kernel_sendmsg(sock, &msg, (struct kvec *)scratchiov, niov, nob);
181 for (i = 0; i < niov; i++)
182 kunmap(kiov[i].kiov_page);
188 ksocknal_lib_eager_ack (ksock_conn_t *conn)
191 mm_segment_t oldmm = get_fs();
192 struct socket *sock = conn->ksnc_sock;
194 /* Remind the socket to ACK eagerly. If I don't, the socket might
195 * think I'm about to send something it could piggy-back the ACK
196 * on, introducing delay in completing zero-copy sends in my
200 sock->ops->setsockopt (sock, SOL_TCP, TCP_QUICKACK,
201 (char *)&opt, sizeof (opt));
206 ksocknal_lib_recv_iov (ksock_conn_t *conn)
208 #if SOCKNAL_SINGLE_FRAG_RX
209 struct iovec scratch;
210 struct iovec *scratchiov = &scratch;
211 unsigned int niov = 1;
213 struct iovec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
214 unsigned int niov = conn->ksnc_rx_niov;
216 struct iovec *iov = conn->ksnc_rx_iov;
217 struct msghdr msg = {
227 /* NB we can't trust socket ops to either consume our iovs
228 * or leave them alone. */
231 for (nob = i = 0; i < niov; i++) {
232 scratchiov[i] = iov[i];
233 nob += scratchiov[i].iov_len;
235 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
237 rc = kernel_recvmsg(conn->ksnc_sock, &msg,
238 (struct kvec *)scratchiov, niov, nob, MSG_DONTWAIT);
241 if (conn->ksnc_proto == &ksocknal_protocol_v2x) {
242 saved_csum = conn->ksnc_msg.ksm_csum;
243 conn->ksnc_msg.ksm_csum = 0;
246 if (saved_csum != 0) {
247 /* accumulate checksum */
248 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
251 fragnob = iov[i].iov_len;
255 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
256 iov[i].iov_base, fragnob);
258 conn->ksnc_msg.ksm_csum = saved_csum;
265 ksocknal_lib_kiov_vunmap(void *addr)
274 ksocknal_lib_kiov_vmap(lnet_kiov_t *kiov, int niov,
275 struct iovec *iov, struct page **pages)
281 if (!*ksocknal_tunables.ksnd_zc_recv || pages == NULL)
284 LASSERT (niov <= LNET_MAX_IOV);
287 niov < *ksocknal_tunables.ksnd_zc_recv_min_nfrags)
290 for (nob = i = 0; i < niov; i++) {
291 if ((kiov[i].kiov_offset != 0 && i > 0) ||
292 (kiov[i].kiov_offset + kiov[i].kiov_len != PAGE_CACHE_SIZE && i < niov - 1))
295 pages[i] = kiov[i].kiov_page;
296 nob += kiov[i].kiov_len;
299 addr = vmap(pages, niov, VM_MAP, PAGE_KERNEL);
303 iov->iov_base = addr + kiov[0].kiov_offset;
310 ksocknal_lib_recv_kiov (ksock_conn_t *conn)
312 #if SOCKNAL_SINGLE_FRAG_RX || !SOCKNAL_RISK_KMAP_DEADLOCK
313 struct iovec scratch;
314 struct iovec *scratchiov = &scratch;
315 struct page **pages = NULL;
316 unsigned int niov = 1;
318 #ifdef CONFIG_HIGHMEM
319 #warning "XXX risk of kmap deadlock on multiple frags..."
321 struct iovec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
322 struct page **pages = conn->ksnc_scheduler->kss_rx_scratch_pgs;
323 unsigned int niov = conn->ksnc_rx_nkiov;
325 lnet_kiov_t *kiov = conn->ksnc_rx_kiov;
326 struct msghdr msg = {
338 /* NB we can't trust socket ops to either consume our iovs
339 * or leave them alone. */
340 addr = ksocknal_lib_kiov_vmap(kiov, niov, scratchiov, pages);
342 nob = scratchiov[0].iov_len;
346 for (nob = i = 0; i < niov; i++) {
347 nob += scratchiov[i].iov_len = kiov[i].kiov_len;
348 scratchiov[i].iov_base = kmap(kiov[i].kiov_page) +
354 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
356 rc = kernel_recvmsg(conn->ksnc_sock, &msg,
357 (struct kvec *)scratchiov, n, nob, MSG_DONTWAIT);
359 if (conn->ksnc_msg.ksm_csum != 0) {
360 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
363 /* Dang! have to kmap again because I have nowhere to stash the
364 * mapped address. But by doing it while the page is still
365 * mapped, the kernel just bumps the map count and returns me
366 * the address it stashed. */
367 base = kmap(kiov[i].kiov_page) + kiov[i].kiov_offset;
368 fragnob = kiov[i].kiov_len;
372 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
375 kunmap(kiov[i].kiov_page);
380 ksocknal_lib_kiov_vunmap(addr);
382 for (i = 0; i < niov; i++)
383 kunmap(kiov[i].kiov_page);
390 ksocknal_lib_csum_tx(ksock_tx_t *tx)
396 LASSERT(tx->tx_iov[0].iov_base == (void *)&tx->tx_msg);
397 LASSERT(tx->tx_conn != NULL);
398 LASSERT(tx->tx_conn->ksnc_proto == &ksocknal_protocol_v2x);
400 tx->tx_msg.ksm_csum = 0;
402 csum = ksocknal_csum(~0, (void *)tx->tx_iov[0].iov_base,
403 tx->tx_iov[0].iov_len);
405 if (tx->tx_kiov != NULL) {
406 for (i = 0; i < tx->tx_nkiov; i++) {
407 base = kmap(tx->tx_kiov[i].kiov_page) +
408 tx->tx_kiov[i].kiov_offset;
410 csum = ksocknal_csum(csum, base, tx->tx_kiov[i].kiov_len);
412 kunmap(tx->tx_kiov[i].kiov_page);
415 for (i = 1; i < tx->tx_niov; i++)
416 csum = ksocknal_csum(csum, tx->tx_iov[i].iov_base,
417 tx->tx_iov[i].iov_len);
420 if (*ksocknal_tunables.ksnd_inject_csum_error) {
422 *ksocknal_tunables.ksnd_inject_csum_error = 0;
425 tx->tx_msg.ksm_csum = csum;
429 ksocknal_lib_get_conn_tunables (ksock_conn_t *conn, int *txmem, int *rxmem, int *nagle)
431 mm_segment_t oldmm = get_fs ();
432 struct socket *sock = conn->ksnc_sock;
436 rc = ksocknal_connsock_addref(conn);
438 LASSERT (conn->ksnc_closing);
439 *txmem = *rxmem = *nagle = 0;
443 rc = libcfs_sock_getbuf(sock, txmem, rxmem);
445 len = sizeof(*nagle);
447 rc = sock->ops->getsockopt(sock, SOL_TCP, TCP_NODELAY,
448 (char *)nagle, &len);
452 ksocknal_connsock_decref(conn);
457 *txmem = *rxmem = *nagle = 0;
463 ksocknal_lib_setup_sock (struct socket *sock)
465 mm_segment_t oldmm = get_fs ();
472 struct linger linger;
474 sock->sk->sk_allocation = GFP_NOFS;
476 /* Ensure this socket aborts active sends immediately when we close
483 rc = sock_setsockopt (sock, SOL_SOCKET, SO_LINGER,
484 (char *)&linger, sizeof (linger));
487 CERROR ("Can't set SO_LINGER: %d\n", rc);
493 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_LINGER2,
494 (char *)&option, sizeof (option));
497 CERROR ("Can't set SO_LINGER2: %d\n", rc);
501 if (!*ksocknal_tunables.ksnd_nagle) {
505 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_NODELAY,
506 (char *)&option, sizeof (option));
509 CERROR ("Can't disable nagle: %d\n", rc);
514 rc = libcfs_sock_setbuf(sock,
515 *ksocknal_tunables.ksnd_tx_buffer_size,
516 *ksocknal_tunables.ksnd_rx_buffer_size);
518 CERROR ("Can't set buffer tx %d, rx %d buffers: %d\n",
519 *ksocknal_tunables.ksnd_tx_buffer_size,
520 *ksocknal_tunables.ksnd_rx_buffer_size, rc);
524 /* TCP_BACKOFF_* sockopt tunables unsupported in stock kernels */
526 /* snapshot tunables */
527 keep_idle = *ksocknal_tunables.ksnd_keepalive_idle;
528 keep_count = *ksocknal_tunables.ksnd_keepalive_count;
529 keep_intvl = *ksocknal_tunables.ksnd_keepalive_intvl;
531 do_keepalive = (keep_idle > 0 && keep_count > 0 && keep_intvl > 0);
533 option = (do_keepalive ? 1 : 0);
535 rc = sock_setsockopt (sock, SOL_SOCKET, SO_KEEPALIVE,
536 (char *)&option, sizeof (option));
539 CERROR ("Can't set SO_KEEPALIVE: %d\n", rc);
547 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPIDLE,
548 (char *)&keep_idle, sizeof (keep_idle));
551 CERROR ("Can't set TCP_KEEPIDLE: %d\n", rc);
556 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPINTVL,
557 (char *)&keep_intvl, sizeof (keep_intvl));
560 CERROR ("Can't set TCP_KEEPINTVL: %d\n", rc);
565 rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPCNT,
566 (char *)&keep_count, sizeof (keep_count));
569 CERROR ("Can't set TCP_KEEPCNT: %d\n", rc);
577 ksocknal_lib_push_conn (ksock_conn_t *conn)
586 rc = ksocknal_connsock_addref(conn);
587 if (rc != 0) /* being shut down */
590 sk = conn->ksnc_sock->sk;
594 nonagle = tp->nonagle;
601 rc = sk->sk_prot->setsockopt (sk, SOL_TCP, TCP_NODELAY,
602 (char *)&val, sizeof (val));
608 tp->nonagle = nonagle;
611 ksocknal_connsock_decref(conn);
614 extern void ksocknal_read_callback (ksock_conn_t *conn);
615 extern void ksocknal_write_callback (ksock_conn_t *conn);
617 * socket call back in Linux
620 ksocknal_data_ready (struct sock *sk, int n)
624 /* interleave correctly with closing sockets... */
626 read_lock(&ksocknal_data.ksnd_global_lock);
628 conn = sk->sk_user_data;
629 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
630 LASSERT (sk->sk_data_ready != &ksocknal_data_ready);
631 sk->sk_data_ready (sk, n);
633 ksocknal_read_callback(conn);
635 read_unlock(&ksocknal_data.ksnd_global_lock);
639 ksocknal_write_space (struct sock *sk)
645 /* interleave correctly with closing sockets... */
647 read_lock(&ksocknal_data.ksnd_global_lock);
649 conn = sk->sk_user_data;
650 wspace = SOCKNAL_WSPACE(sk);
651 min_wpace = SOCKNAL_MIN_WSPACE(sk);
653 CDEBUG(D_NET, "sk %p wspace %d low water %d conn %p%s%s%s\n",
654 sk, wspace, min_wpace, conn,
655 (conn == NULL) ? "" : (conn->ksnc_tx_ready ?
656 " ready" : " blocked"),
657 (conn == NULL) ? "" : (conn->ksnc_tx_scheduled ?
658 " scheduled" : " idle"),
659 (conn == NULL) ? "" : (list_empty (&conn->ksnc_tx_queue) ?
660 " empty" : " queued"));
662 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
663 LASSERT (sk->sk_write_space != &ksocknal_write_space);
664 sk->sk_write_space (sk);
666 read_unlock(&ksocknal_data.ksnd_global_lock);
670 if (wspace >= min_wpace) { /* got enough space */
671 ksocknal_write_callback(conn);
673 /* Clear SOCK_NOSPACE _after_ ksocknal_write_callback so the
674 * ENOMEM check in ksocknal_transmit is race-free (think about
677 clear_bit (SOCK_NOSPACE, &sk->sk_socket->flags);
680 read_unlock(&ksocknal_data.ksnd_global_lock);
684 ksocknal_lib_save_callback(struct socket *sock, ksock_conn_t *conn)
686 conn->ksnc_saved_data_ready = sock->sk->sk_data_ready;
687 conn->ksnc_saved_write_space = sock->sk->sk_write_space;
691 ksocknal_lib_set_callback(struct socket *sock, ksock_conn_t *conn)
693 sock->sk->sk_user_data = conn;
694 sock->sk->sk_data_ready = ksocknal_data_ready;
695 sock->sk->sk_write_space = ksocknal_write_space;
700 ksocknal_lib_reset_callback(struct socket *sock, ksock_conn_t *conn)
702 /* Remove conn's network callbacks.
703 * NB I _have_ to restore the callback, rather than storing a noop,
704 * since the socket could survive past this module being unloaded!! */
705 sock->sk->sk_data_ready = conn->ksnc_saved_data_ready;
706 sock->sk->sk_write_space = conn->ksnc_saved_write_space;
708 /* A callback could be in progress already; they hold a read lock
709 * on ksnd_global_lock (to serialise with me) and NOOP if
710 * sk_user_data is NULL. */
711 sock->sk->sk_user_data = NULL;
717 ksocknal_lib_memory_pressure(ksock_conn_t *conn)
720 ksock_sched_t *sched;
722 sched = conn->ksnc_scheduler;
723 spin_lock_bh(&sched->kss_lock);
725 if (!SOCK_TEST_NOSPACE(conn->ksnc_sock) &&
726 !conn->ksnc_tx_ready) {
727 /* SOCK_NOSPACE is set when the socket fills
728 * and cleared in the write_space callback
729 * (which also sets ksnc_tx_ready). If
730 * SOCK_NOSPACE and ksnc_tx_ready are BOTH
731 * zero, I didn't fill the socket and
732 * write_space won't reschedule me, so I
733 * return -ENOMEM to get my caller to retry
738 spin_unlock_bh(&sched->kss_lock);