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 = lnet_sock_getaddr(conn->ksnc_sock, true,
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 = lnet_sock_getaddr(conn->ksnc_sock, false,
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_CSUM_MASK) != 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 kvec *scratchiov = &scratch;
97 unsigned int niov = 1;
99 struct kvec *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, 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
157 struct kvec *scratchiov = &scratch;
158 unsigned int niov = 1;
160 #ifdef CONFIG_HIGHMEM
161 #warning "XXX risk of kmap deadlock on multiple frags..."
163 struct kvec *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, 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 struct socket *sock = conn->ksnc_sock;
193 /* Remind the socket to ACK eagerly. If I don't, the socket might
194 * think I'm about to send something it could piggy-back the ACK
195 * on, introducing delay in completing zero-copy sends in my
198 kernel_setsockopt(sock, SOL_TCP, TCP_QUICKACK,
199 (char *)&opt, sizeof(opt));
203 ksocknal_lib_recv_iov (ksock_conn_t *conn)
205 #if SOCKNAL_SINGLE_FRAG_RX
207 struct kvec *scratchiov = &scratch;
208 unsigned int niov = 1;
210 struct kvec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
211 unsigned int niov = conn->ksnc_rx_niov;
213 struct kvec *iov = conn->ksnc_rx_iov;
214 struct msghdr msg = {
224 /* NB we can't trust socket ops to either consume our iovs
225 * or leave them alone. */
228 for (nob = i = 0; i < niov; i++) {
229 scratchiov[i] = iov[i];
230 nob += scratchiov[i].iov_len;
232 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
234 rc = kernel_recvmsg(conn->ksnc_sock, &msg, scratchiov, niov, nob,
238 if (conn->ksnc_proto == &ksocknal_protocol_v2x) {
239 saved_csum = conn->ksnc_msg.ksm_csum;
240 conn->ksnc_msg.ksm_csum = 0;
243 if (saved_csum != 0) {
244 /* accumulate checksum */
245 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
248 fragnob = iov[i].iov_len;
252 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
253 iov[i].iov_base, fragnob);
255 conn->ksnc_msg.ksm_csum = saved_csum;
262 ksocknal_lib_kiov_vunmap(void *addr)
271 ksocknal_lib_kiov_vmap(lnet_kiov_t *kiov, int niov,
272 struct kvec *iov, struct page **pages)
278 if (!*ksocknal_tunables.ksnd_zc_recv || pages == NULL)
281 LASSERT (niov <= LNET_MAX_IOV);
284 niov < *ksocknal_tunables.ksnd_zc_recv_min_nfrags)
287 for (nob = i = 0; i < niov; i++) {
288 if ((kiov[i].kiov_offset != 0 && i > 0) ||
289 (kiov[i].kiov_offset + kiov[i].kiov_len !=
290 PAGE_CACHE_SIZE && i < niov - 1))
293 pages[i] = kiov[i].kiov_page;
294 nob += kiov[i].kiov_len;
297 addr = vmap(pages, niov, VM_MAP, PAGE_KERNEL);
301 iov->iov_base = addr + kiov[0].kiov_offset;
308 ksocknal_lib_recv_kiov (ksock_conn_t *conn)
310 #if SOCKNAL_SINGLE_FRAG_RX || !SOCKNAL_RISK_KMAP_DEADLOCK
312 struct kvec *scratchiov = &scratch;
313 struct page **pages = NULL;
314 unsigned int niov = 1;
316 #ifdef CONFIG_HIGHMEM
317 #warning "XXX risk of kmap deadlock on multiple frags..."
319 struct kvec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
320 struct page **pages = conn->ksnc_scheduler->kss_rx_scratch_pgs;
321 unsigned int niov = conn->ksnc_rx_nkiov;
323 lnet_kiov_t *kiov = conn->ksnc_rx_kiov;
324 struct msghdr msg = {
336 /* NB we can't trust socket ops to either consume our iovs
337 * or leave them alone. */
338 if ((addr = ksocknal_lib_kiov_vmap(kiov, niov, scratchiov, pages)) != NULL) {
339 nob = scratchiov[0].iov_len;
343 for (nob = i = 0; i < niov; i++) {
344 nob += scratchiov[i].iov_len = kiov[i].kiov_len;
345 scratchiov[i].iov_base = kmap(kiov[i].kiov_page) +
351 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
353 rc = kernel_recvmsg(conn->ksnc_sock, &msg, scratchiov, n, nob,
356 if (conn->ksnc_msg.ksm_csum != 0) {
357 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
360 /* Dang! have to kmap again because I have nowhere to stash the
361 * mapped address. But by doing it while the page is still
362 * mapped, the kernel just bumps the map count and returns me
363 * the address it stashed. */
364 base = kmap(kiov[i].kiov_page) + kiov[i].kiov_offset;
365 fragnob = kiov[i].kiov_len;
369 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
372 kunmap(kiov[i].kiov_page);
377 ksocknal_lib_kiov_vunmap(addr);
379 for (i = 0; i < niov; i++)
380 kunmap(kiov[i].kiov_page);
387 ksocknal_lib_csum_tx(ksock_tx_t *tx)
393 LASSERT(tx->tx_iov[0].iov_base == (void *)&tx->tx_msg);
394 LASSERT(tx->tx_conn != NULL);
395 LASSERT(tx->tx_conn->ksnc_proto == &ksocknal_protocol_v2x);
397 tx->tx_msg.ksm_csum = 0;
399 csum = ksocknal_csum(~0, (void *)tx->tx_iov[0].iov_base,
400 tx->tx_iov[0].iov_len);
402 if (tx->tx_kiov != NULL) {
403 for (i = 0; i < tx->tx_nkiov; i++) {
404 base = kmap(tx->tx_kiov[i].kiov_page) +
405 tx->tx_kiov[i].kiov_offset;
407 csum = ksocknal_csum(csum, base, tx->tx_kiov[i].kiov_len);
409 kunmap(tx->tx_kiov[i].kiov_page);
412 for (i = 1; i < tx->tx_niov; i++)
413 csum = ksocknal_csum(csum, tx->tx_iov[i].iov_base,
414 tx->tx_iov[i].iov_len);
417 if (*ksocknal_tunables.ksnd_inject_csum_error) {
419 *ksocknal_tunables.ksnd_inject_csum_error = 0;
422 tx->tx_msg.ksm_csum = csum;
426 ksocknal_lib_get_conn_tunables (ksock_conn_t *conn, int *txmem, int *rxmem, int *nagle)
428 struct socket *sock = conn->ksnc_sock;
432 rc = ksocknal_connsock_addref(conn);
434 LASSERT (conn->ksnc_closing);
435 *txmem = *rxmem = *nagle = 0;
439 rc = lnet_sock_getbuf(sock, txmem, rxmem);
441 len = sizeof(*nagle);
442 rc = kernel_getsockopt(sock, SOL_TCP, TCP_NODELAY,
443 (char *)nagle, &len);
446 ksocknal_connsock_decref(conn);
451 *txmem = *rxmem = *nagle = 0;
457 ksocknal_lib_setup_sock (struct socket *sock)
465 struct linger linger;
467 sock->sk->sk_allocation = GFP_NOFS;
469 /* Ensure this socket aborts active sends immediately when we close
475 rc = kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER,
476 (char *)&linger, sizeof(linger));
478 CERROR ("Can't set SO_LINGER: %d\n", rc);
483 rc = kernel_setsockopt(sock, SOL_TCP, TCP_LINGER2,
484 (char *)&option, sizeof(option));
486 CERROR ("Can't set SO_LINGER2: %d\n", rc);
490 if (!*ksocknal_tunables.ksnd_nagle) {
493 rc = kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY,
494 (char *)&option, sizeof(option));
496 CERROR ("Can't disable nagle: %d\n", rc);
501 rc = lnet_sock_setbuf(sock,
502 *ksocknal_tunables.ksnd_tx_buffer_size,
503 *ksocknal_tunables.ksnd_rx_buffer_size);
505 CERROR ("Can't set buffer tx %d, rx %d buffers: %d\n",
506 *ksocknal_tunables.ksnd_tx_buffer_size,
507 *ksocknal_tunables.ksnd_rx_buffer_size, rc);
511 /* TCP_BACKOFF_* sockopt tunables unsupported in stock kernels */
512 #ifdef SOCKNAL_BACKOFF
513 if (*ksocknal_tunables.ksnd_backoff_init > 0) {
514 option = *ksocknal_tunables.ksnd_backoff_init;
515 #ifdef SOCKNAL_BACKOFF_MS
519 rc = kernel_setsockopt(sock, SOL_TCP, TCP_BACKOFF_INIT,
520 (char *)&option, sizeof(option));
522 CERROR ("Can't set initial tcp backoff %d: %d\n",
528 if (*ksocknal_tunables.ksnd_backoff_max > 0) {
529 option = *ksocknal_tunables.ksnd_backoff_max;
530 #ifdef SOCKNAL_BACKOFF_MS
534 rc = kernel_setsockopt(sock, SOL_TCP, TCP_BACKOFF_MAX,
535 (char *)&option, sizeof(option));
537 CERROR ("Can't set maximum tcp backoff %d: %d\n",
544 /* snapshot tunables */
545 keep_idle = *ksocknal_tunables.ksnd_keepalive_idle;
546 keep_count = *ksocknal_tunables.ksnd_keepalive_count;
547 keep_intvl = *ksocknal_tunables.ksnd_keepalive_intvl;
549 do_keepalive = (keep_idle > 0 && keep_count > 0 && keep_intvl > 0);
551 option = (do_keepalive ? 1 : 0);
552 rc = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
553 (char *)&option, sizeof(option));
555 CERROR ("Can't set SO_KEEPALIVE: %d\n", rc);
562 rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPIDLE,
563 (char *)&keep_idle, sizeof(keep_idle));
565 CERROR ("Can't set TCP_KEEPIDLE: %d\n", rc);
569 rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPINTVL,
570 (char *)&keep_intvl, sizeof(keep_intvl));
572 CERROR ("Can't set TCP_KEEPINTVL: %d\n", rc);
576 rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPCNT,
577 (char *)&keep_count, sizeof(keep_count));
579 CERROR ("Can't set TCP_KEEPCNT: %d\n", rc);
587 ksocknal_lib_push_conn (ksock_conn_t *conn)
595 rc = ksocknal_connsock_addref(conn);
596 if (rc != 0) /* being shut down */
599 sk = conn->ksnc_sock->sk;
603 nonagle = tp->nonagle;
607 rc = kernel_setsockopt(conn->ksnc_sock, SOL_TCP, TCP_NODELAY,
608 (char *)&val, sizeof(val));
612 tp->nonagle = nonagle;
615 ksocknal_connsock_decref(conn);
618 extern void ksocknal_read_callback (ksock_conn_t *conn);
619 extern void ksocknal_write_callback (ksock_conn_t *conn);
621 * socket call back in Linux
624 #ifdef HAVE_SK_DATA_READY_ONE_ARG
625 ksocknal_data_ready(struct sock *sk)
627 ksocknal_data_ready(struct sock *sk, int n)
633 /* interleave correctly with closing sockets... */
635 read_lock(&ksocknal_data.ksnd_global_lock);
637 conn = sk->sk_user_data;
638 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
639 LASSERT(sk->sk_data_ready != &ksocknal_data_ready);
640 #ifdef HAVE_SK_DATA_READY_ONE_ARG
641 sk->sk_data_ready(sk);
643 sk->sk_data_ready(sk, n);
646 ksocknal_read_callback(conn);
648 read_unlock(&ksocknal_data.ksnd_global_lock);
654 ksocknal_write_space (struct sock *sk)
660 /* interleave correctly with closing sockets... */
662 read_lock(&ksocknal_data.ksnd_global_lock);
664 conn = sk->sk_user_data;
665 wspace = sk_stream_wspace(sk);
666 min_wpace = sk_stream_min_wspace(sk);
668 CDEBUG(D_NET, "sk %p wspace %d low water %d conn %p%s%s%s\n",
669 sk, wspace, min_wpace, conn,
670 (conn == NULL) ? "" : (conn->ksnc_tx_ready ?
671 " ready" : " blocked"),
672 (conn == NULL) ? "" : (conn->ksnc_tx_scheduled ?
673 " scheduled" : " idle"),
674 (conn == NULL) ? "" : (list_empty(&conn->ksnc_tx_queue) ?
675 " empty" : " queued"));
677 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
678 LASSERT (sk->sk_write_space != &ksocknal_write_space);
679 sk->sk_write_space (sk);
681 read_unlock(&ksocknal_data.ksnd_global_lock);
685 if (wspace >= min_wpace) { /* got enough space */
686 ksocknal_write_callback(conn);
688 /* Clear SOCK_NOSPACE _after_ ksocknal_write_callback so the
689 * ENOMEM check in ksocknal_transmit is race-free (think about
692 clear_bit (SOCK_NOSPACE, &sk->sk_socket->flags);
695 read_unlock(&ksocknal_data.ksnd_global_lock);
699 ksocknal_lib_save_callback(struct socket *sock, ksock_conn_t *conn)
701 conn->ksnc_saved_data_ready = sock->sk->sk_data_ready;
702 conn->ksnc_saved_write_space = sock->sk->sk_write_space;
706 ksocknal_lib_set_callback(struct socket *sock, ksock_conn_t *conn)
708 sock->sk->sk_user_data = conn;
709 sock->sk->sk_data_ready = ksocknal_data_ready;
710 sock->sk->sk_write_space = ksocknal_write_space;
715 ksocknal_lib_reset_callback(struct socket *sock, ksock_conn_t *conn)
717 /* Remove conn's network callbacks.
718 * NB I _have_ to restore the callback, rather than storing a noop,
719 * since the socket could survive past this module being unloaded!! */
720 sock->sk->sk_data_ready = conn->ksnc_saved_data_ready;
721 sock->sk->sk_write_space = conn->ksnc_saved_write_space;
723 /* A callback could be in progress already; they hold a read lock
724 * on ksnd_global_lock (to serialise with me) and NOOP if
725 * sk_user_data is NULL. */
726 sock->sk->sk_user_data = NULL;
732 ksocknal_lib_memory_pressure(ksock_conn_t *conn)
735 ksock_sched_t *sched;
737 sched = conn->ksnc_scheduler;
738 spin_lock_bh(&sched->kss_lock);
740 if (!test_bit(SOCK_NOSPACE, &conn->ksnc_sock->flags) &&
741 !conn->ksnc_tx_ready) {
742 /* SOCK_NOSPACE is set when the socket fills
743 * and cleared in the write_space callback
744 * (which also sets ksnc_tx_ready). If
745 * SOCK_NOSPACE and ksnc_tx_ready are BOTH
746 * zero, I didn't fill the socket and
747 * write_space won't reschedule me, so I
748 * return -ENOMEM to get my caller to retry
753 spin_unlock_bh(&sched->kss_lock);