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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
36 ksocknal_lib_get_conn_addrs(struct ksock_conn *conn)
38 int rc = lnet_sock_getaddr(conn->ksnc_sock, true,
42 /* Didn't need the {get,put}connsock dance to deref ksnc_sock... */
43 LASSERT (!conn->ksnc_closing);
46 CERROR ("Error %d getting sock peer_ni IP\n", rc);
50 rc = lnet_sock_getaddr(conn->ksnc_sock, false,
51 &conn->ksnc_myipaddr, NULL);
53 CERROR ("Error %d getting sock local IP\n", rc);
61 ksocknal_lib_zc_capable(struct ksock_conn *conn)
63 int caps = conn->ksnc_sock->sk->sk_route_caps;
65 if (conn->ksnc_proto == &ksocknal_protocol_v1x)
68 /* ZC if the socket supports scatter/gather and doesn't need software
70 return ((caps & NETIF_F_SG) != 0 && (caps & NETIF_F_CSUM_MASK) != 0);
74 ksocknal_lib_send_iov(struct ksock_conn *conn, struct ksock_tx *tx)
76 struct socket *sock = conn->ksnc_sock;
80 if (*ksocknal_tunables.ksnd_enable_csum && /* checksum enabled */
81 conn->ksnc_proto == &ksocknal_protocol_v2x && /* V2.x connection */
82 tx->tx_nob == tx->tx_resid && /* frist sending */
83 tx->tx_msg.ksm_csum == 0) /* not checksummed */
84 ksocknal_lib_csum_tx(tx);
86 /* NB we can't trust socket ops to either consume our iovs
87 * or leave them alone. */
90 #if SOCKNAL_SINGLE_FRAG_TX
92 struct kvec *scratchiov = &scratch;
93 unsigned int niov = 1;
95 struct kvec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
96 unsigned int niov = tx->tx_niov;
98 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
101 for (nob = i = 0; i < niov; i++) {
102 scratchiov[i] = tx->tx_iov[i];
103 nob += scratchiov[i].iov_len;
106 if (!list_empty(&conn->ksnc_tx_queue) ||
108 msg.msg_flags |= MSG_MORE;
110 rc = kernel_sendmsg(sock, &msg, scratchiov, niov, nob);
116 ksocknal_lib_send_kiov(struct ksock_conn *conn, struct ksock_tx *tx)
118 struct socket *sock = conn->ksnc_sock;
119 lnet_kiov_t *kiov = tx->tx_kiov;
123 /* Not NOOP message */
124 LASSERT (tx->tx_lnetmsg != NULL);
126 /* NB we can't trust socket ops to either consume our iovs
127 * or leave them alone. */
128 if (tx->tx_msg.ksm_zc_cookies[0] != 0) {
129 /* Zero copy is enabled */
130 struct sock *sk = sock->sk;
131 struct page *page = kiov->kiov_page;
132 int offset = kiov->kiov_offset;
133 int fragsize = kiov->kiov_len;
134 int msgflg = MSG_DONTWAIT;
136 CDEBUG(D_NET, "page %p + offset %x for %d\n",
137 page, offset, kiov->kiov_len);
139 if (!list_empty(&conn->ksnc_tx_queue) ||
140 fragsize < tx->tx_resid)
143 if (sk->sk_prot->sendpage != NULL) {
144 rc = sk->sk_prot->sendpage(sk, page,
145 offset, fragsize, msgflg);
147 rc = cfs_tcp_sendpage(sk, page, offset, fragsize,
151 #if SOCKNAL_SINGLE_FRAG_TX || !SOCKNAL_RISK_KMAP_DEADLOCK
153 struct kvec *scratchiov = &scratch;
154 unsigned int niov = 1;
156 #ifdef CONFIG_HIGHMEM
157 #warning "XXX risk of kmap deadlock on multiple frags..."
159 struct kvec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
160 unsigned int niov = tx->tx_nkiov;
162 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
165 for (nob = i = 0; i < niov; i++) {
166 scratchiov[i].iov_base = kmap(kiov[i].kiov_page) +
168 nob += scratchiov[i].iov_len = kiov[i].kiov_len;
171 if (!list_empty(&conn->ksnc_tx_queue) ||
173 msg.msg_flags |= MSG_MORE;
175 rc = kernel_sendmsg(sock, &msg, scratchiov, niov, nob);
177 for (i = 0; i < niov; i++)
178 kunmap(kiov[i].kiov_page);
184 ksocknal_lib_eager_ack(struct ksock_conn *conn)
187 struct socket *sock = conn->ksnc_sock;
189 /* Remind the socket to ACK eagerly. If I don't, the socket might
190 * think I'm about to send something it could piggy-back the ACK
191 * on, introducing delay in completing zero-copy sends in my
194 kernel_setsockopt(sock, SOL_TCP, TCP_QUICKACK,
195 (char *)&opt, sizeof(opt));
199 ksocknal_lib_recv_iov(struct ksock_conn *conn)
201 #if SOCKNAL_SINGLE_FRAG_RX
203 struct kvec *scratchiov = &scratch;
204 unsigned int niov = 1;
206 struct kvec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
207 unsigned int niov = conn->ksnc_rx_niov;
209 struct kvec *iov = conn->ksnc_rx_iov;
210 struct msghdr msg = {
220 /* NB we can't trust socket ops to either consume our iovs
221 * or leave them alone. */
224 for (nob = i = 0; i < niov; i++) {
225 scratchiov[i] = iov[i];
226 nob += scratchiov[i].iov_len;
228 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
230 rc = kernel_recvmsg(conn->ksnc_sock, &msg, scratchiov, niov, nob,
234 if (conn->ksnc_proto == &ksocknal_protocol_v2x) {
235 saved_csum = conn->ksnc_msg.ksm_csum;
236 conn->ksnc_msg.ksm_csum = 0;
239 if (saved_csum != 0) {
240 /* accumulate checksum */
241 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
244 fragnob = iov[i].iov_len;
248 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
249 iov[i].iov_base, fragnob);
251 conn->ksnc_msg.ksm_csum = saved_csum;
258 ksocknal_lib_kiov_vunmap(void *addr)
267 ksocknal_lib_kiov_vmap(lnet_kiov_t *kiov, int niov,
268 struct kvec *iov, struct page **pages)
274 if (!*ksocknal_tunables.ksnd_zc_recv || pages == NULL)
277 LASSERT (niov <= LNET_MAX_IOV);
280 niov < *ksocknal_tunables.ksnd_zc_recv_min_nfrags)
283 for (nob = i = 0; i < niov; i++) {
284 if ((kiov[i].kiov_offset != 0 && i > 0) ||
285 (kiov[i].kiov_offset + kiov[i].kiov_len !=
286 PAGE_SIZE && i < niov - 1))
289 pages[i] = kiov[i].kiov_page;
290 nob += kiov[i].kiov_len;
293 addr = vmap(pages, niov, VM_MAP, PAGE_KERNEL);
297 iov->iov_base = addr + kiov[0].kiov_offset;
304 ksocknal_lib_recv_kiov(struct ksock_conn *conn)
306 #if SOCKNAL_SINGLE_FRAG_RX || !SOCKNAL_RISK_KMAP_DEADLOCK
308 struct kvec *scratchiov = &scratch;
309 struct page **pages = NULL;
310 unsigned int niov = 1;
312 #ifdef CONFIG_HIGHMEM
313 #warning "XXX risk of kmap deadlock on multiple frags..."
315 struct kvec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
316 struct page **pages = conn->ksnc_scheduler->kss_rx_scratch_pgs;
317 unsigned int niov = conn->ksnc_rx_nkiov;
319 lnet_kiov_t *kiov = conn->ksnc_rx_kiov;
320 struct msghdr msg = {
332 /* NB we can't trust socket ops to either consume our iovs
333 * or leave them alone. */
334 if ((addr = ksocknal_lib_kiov_vmap(kiov, niov, scratchiov, pages)) != NULL) {
335 nob = scratchiov[0].iov_len;
339 for (nob = i = 0; i < niov; i++) {
340 nob += scratchiov[i].iov_len = kiov[i].kiov_len;
341 scratchiov[i].iov_base = kmap(kiov[i].kiov_page) +
347 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
349 rc = kernel_recvmsg(conn->ksnc_sock, &msg, scratchiov, n, nob,
352 if (conn->ksnc_msg.ksm_csum != 0) {
353 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
356 /* Dang! have to kmap again because I have nowhere to stash the
357 * mapped address. But by doing it while the page is still
358 * mapped, the kernel just bumps the map count and returns me
359 * the address it stashed. */
360 base = kmap(kiov[i].kiov_page) + kiov[i].kiov_offset;
361 fragnob = kiov[i].kiov_len;
365 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
368 kunmap(kiov[i].kiov_page);
373 ksocknal_lib_kiov_vunmap(addr);
375 for (i = 0; i < niov; i++)
376 kunmap(kiov[i].kiov_page);
383 ksocknal_lib_csum_tx(struct ksock_tx *tx)
389 LASSERT(tx->tx_iov[0].iov_base == (void *)&tx->tx_msg);
390 LASSERT(tx->tx_conn != NULL);
391 LASSERT(tx->tx_conn->ksnc_proto == &ksocknal_protocol_v2x);
393 tx->tx_msg.ksm_csum = 0;
395 csum = ksocknal_csum(~0, (void *)tx->tx_iov[0].iov_base,
396 tx->tx_iov[0].iov_len);
398 if (tx->tx_kiov != NULL) {
399 for (i = 0; i < tx->tx_nkiov; i++) {
400 base = kmap(tx->tx_kiov[i].kiov_page) +
401 tx->tx_kiov[i].kiov_offset;
403 csum = ksocknal_csum(csum, base, tx->tx_kiov[i].kiov_len);
405 kunmap(tx->tx_kiov[i].kiov_page);
408 for (i = 1; i < tx->tx_niov; i++)
409 csum = ksocknal_csum(csum, tx->tx_iov[i].iov_base,
410 tx->tx_iov[i].iov_len);
413 if (*ksocknal_tunables.ksnd_inject_csum_error) {
415 *ksocknal_tunables.ksnd_inject_csum_error = 0;
418 tx->tx_msg.ksm_csum = csum;
422 ksocknal_lib_get_conn_tunables(struct ksock_conn *conn, int *txmem, int *rxmem, int *nagle)
424 struct socket *sock = conn->ksnc_sock;
428 rc = ksocknal_connsock_addref(conn);
430 LASSERT (conn->ksnc_closing);
431 *txmem = *rxmem = *nagle = 0;
435 rc = lnet_sock_getbuf(sock, txmem, rxmem);
437 len = sizeof(*nagle);
438 rc = kernel_getsockopt(sock, SOL_TCP, TCP_NODELAY,
439 (char *)nagle, &len);
442 ksocknal_connsock_decref(conn);
447 *txmem = *rxmem = *nagle = 0;
453 ksocknal_lib_setup_sock (struct socket *sock)
461 struct linger linger;
463 sock->sk->sk_allocation = GFP_NOFS;
465 /* Ensure this socket aborts active sends immediately when we close
471 rc = kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER,
472 (char *)&linger, sizeof(linger));
474 CERROR ("Can't set SO_LINGER: %d\n", rc);
479 rc = kernel_setsockopt(sock, SOL_TCP, TCP_LINGER2,
480 (char *)&option, sizeof(option));
482 CERROR ("Can't set SO_LINGER2: %d\n", rc);
486 if (!*ksocknal_tunables.ksnd_nagle) {
489 rc = kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY,
490 (char *)&option, sizeof(option));
492 CERROR ("Can't disable nagle: %d\n", rc);
497 rc = lnet_sock_setbuf(sock,
498 *ksocknal_tunables.ksnd_tx_buffer_size,
499 *ksocknal_tunables.ksnd_rx_buffer_size);
501 CERROR ("Can't set buffer tx %d, rx %d buffers: %d\n",
502 *ksocknal_tunables.ksnd_tx_buffer_size,
503 *ksocknal_tunables.ksnd_rx_buffer_size, rc);
507 /* TCP_BACKOFF_* sockopt tunables unsupported in stock kernels */
508 #ifdef SOCKNAL_BACKOFF
509 if (*ksocknal_tunables.ksnd_backoff_init > 0) {
510 option = *ksocknal_tunables.ksnd_backoff_init;
511 #ifdef SOCKNAL_BACKOFF_MS
515 rc = kernel_setsockopt(sock, SOL_TCP, TCP_BACKOFF_INIT,
516 (char *)&option, sizeof(option));
518 CERROR ("Can't set initial tcp backoff %d: %d\n",
524 if (*ksocknal_tunables.ksnd_backoff_max > 0) {
525 option = *ksocknal_tunables.ksnd_backoff_max;
526 #ifdef SOCKNAL_BACKOFF_MS
530 rc = kernel_setsockopt(sock, SOL_TCP, TCP_BACKOFF_MAX,
531 (char *)&option, sizeof(option));
533 CERROR ("Can't set maximum tcp backoff %d: %d\n",
540 /* snapshot tunables */
541 keep_idle = *ksocknal_tunables.ksnd_keepalive_idle;
542 keep_count = *ksocknal_tunables.ksnd_keepalive_count;
543 keep_intvl = *ksocknal_tunables.ksnd_keepalive_intvl;
545 do_keepalive = (keep_idle > 0 && keep_count > 0 && keep_intvl > 0);
547 option = (do_keepalive ? 1 : 0);
548 rc = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
549 (char *)&option, sizeof(option));
551 CERROR ("Can't set SO_KEEPALIVE: %d\n", rc);
558 rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPIDLE,
559 (char *)&keep_idle, sizeof(keep_idle));
561 CERROR ("Can't set TCP_KEEPIDLE: %d\n", rc);
565 rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPINTVL,
566 (char *)&keep_intvl, sizeof(keep_intvl));
568 CERROR ("Can't set TCP_KEEPINTVL: %d\n", rc);
572 rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPCNT,
573 (char *)&keep_count, sizeof(keep_count));
575 CERROR ("Can't set TCP_KEEPCNT: %d\n", rc);
583 ksocknal_lib_push_conn(struct ksock_conn *conn)
591 rc = ksocknal_connsock_addref(conn);
592 if (rc != 0) /* being shut down */
595 sk = conn->ksnc_sock->sk;
599 nonagle = tp->nonagle;
603 rc = kernel_setsockopt(conn->ksnc_sock, SOL_TCP, TCP_NODELAY,
604 (char *)&val, sizeof(val));
608 tp->nonagle = nonagle;
611 ksocknal_connsock_decref(conn);
614 void ksocknal_read_callback(struct ksock_conn *conn);
615 void ksocknal_write_callback(struct ksock_conn *conn);
617 * socket call back in Linux
620 #ifdef HAVE_SK_DATA_READY_ONE_ARG
621 ksocknal_data_ready(struct sock *sk)
623 ksocknal_data_ready(struct sock *sk, int n)
626 struct ksock_conn *conn;
629 /* interleave correctly with closing sockets... */
631 read_lock(&ksocknal_data.ksnd_global_lock);
633 conn = sk->sk_user_data;
634 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
635 LASSERT(sk->sk_data_ready != &ksocknal_data_ready);
636 #ifdef HAVE_SK_DATA_READY_ONE_ARG
637 sk->sk_data_ready(sk);
639 sk->sk_data_ready(sk, n);
642 ksocknal_read_callback(conn);
644 read_unlock(&ksocknal_data.ksnd_global_lock);
650 ksocknal_write_space (struct sock *sk)
652 struct ksock_conn *conn;
656 /* interleave correctly with closing sockets... */
658 read_lock(&ksocknal_data.ksnd_global_lock);
660 conn = sk->sk_user_data;
661 wspace = sk_stream_wspace(sk);
662 min_wpace = sk_stream_min_wspace(sk);
664 CDEBUG(D_NET, "sk %p wspace %d low water %d conn %p%s%s%s\n",
665 sk, wspace, min_wpace, conn,
666 (conn == NULL) ? "" : (conn->ksnc_tx_ready ?
667 " ready" : " blocked"),
668 (conn == NULL) ? "" : (conn->ksnc_tx_scheduled ?
669 " scheduled" : " idle"),
670 (conn == NULL) ? "" : (list_empty(&conn->ksnc_tx_queue) ?
671 " empty" : " queued"));
673 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
674 LASSERT (sk->sk_write_space != &ksocknal_write_space);
675 sk->sk_write_space (sk);
677 read_unlock(&ksocknal_data.ksnd_global_lock);
681 if (wspace >= min_wpace) { /* got enough space */
682 ksocknal_write_callback(conn);
684 /* Clear SOCK_NOSPACE _after_ ksocknal_write_callback so the
685 * ENOMEM check in ksocknal_transmit is race-free (think about
688 clear_bit (SOCK_NOSPACE, &sk->sk_socket->flags);
691 read_unlock(&ksocknal_data.ksnd_global_lock);
695 ksocknal_lib_save_callback(struct socket *sock, struct ksock_conn *conn)
697 conn->ksnc_saved_data_ready = sock->sk->sk_data_ready;
698 conn->ksnc_saved_write_space = sock->sk->sk_write_space;
702 ksocknal_lib_set_callback(struct socket *sock, struct ksock_conn *conn)
704 sock->sk->sk_user_data = conn;
705 sock->sk->sk_data_ready = ksocknal_data_ready;
706 sock->sk->sk_write_space = ksocknal_write_space;
711 ksocknal_lib_reset_callback(struct socket *sock, struct ksock_conn *conn)
713 /* Remove conn's network callbacks.
714 * NB I _have_ to restore the callback, rather than storing a noop,
715 * since the socket could survive past this module being unloaded!! */
716 sock->sk->sk_data_ready = conn->ksnc_saved_data_ready;
717 sock->sk->sk_write_space = conn->ksnc_saved_write_space;
719 /* A callback could be in progress already; they hold a read lock
720 * on ksnd_global_lock (to serialise with me) and NOOP if
721 * sk_user_data is NULL. */
722 sock->sk->sk_user_data = NULL;
728 ksocknal_lib_memory_pressure(struct ksock_conn *conn)
731 struct ksock_sched *sched;
733 sched = conn->ksnc_scheduler;
734 spin_lock_bh(&sched->kss_lock);
736 if (!test_bit(SOCK_NOSPACE, &conn->ksnc_sock->flags) &&
737 !conn->ksnc_tx_ready) {
738 /* SOCK_NOSPACE is set when the socket fills
739 * and cleared in the write_space callback
740 * (which also sets ksnc_tx_ready). If
741 * SOCK_NOSPACE and ksnc_tx_ready are BOTH
742 * zero, I didn't fill the socket and
743 * write_space won't reschedule me, so I
744 * return -ENOMEM to get my caller to retry
749 spin_unlock_bh(&sched->kss_lock);