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,
39 &conn->ksnc_peeraddr);
41 /* Didn't need the {get,put}connsock dance to deref ksnc_sock... */
42 LASSERT(!conn->ksnc_closing);
45 CERROR("Error %d getting sock peer_ni IP\n", rc);
49 rc = lnet_sock_getaddr(conn->ksnc_sock, false,
52 CERROR("Error %d getting sock local IP\n", rc);
60 ksocknal_lib_zc_capable(struct ksock_conn *conn)
62 int caps = conn->ksnc_sock->sk->sk_route_caps;
64 if (conn->ksnc_proto == &ksocknal_protocol_v1x)
67 /* ZC if the socket supports scatter/gather and doesn't need software
69 return ((caps & NETIF_F_SG) != 0 && (caps & NETIF_F_CSUM_MASK) != 0);
73 ksocknal_lib_send_hdr(struct ksock_conn *conn, struct ksock_tx *tx,
74 struct kvec *scratchiov)
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 unsigned int niov = tx->tx_niov;
97 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
100 scratchiov[0] = tx->tx_hdr;
101 nob += scratchiov[0].iov_len;
104 if (!list_empty(&conn->ksnc_tx_queue) ||
106 msg.msg_flags |= MSG_MORE;
108 rc = kernel_sendmsg(sock, &msg, scratchiov, niov, nob);
114 ksocknal_lib_send_kiov(struct ksock_conn *conn, struct ksock_tx *tx,
115 struct kvec *scratchiov)
117 struct socket *sock = conn->ksnc_sock;
118 struct bio_vec *kiov = tx->tx_kiov;
122 /* Not NOOP message */
123 LASSERT(tx->tx_lnetmsg != NULL);
125 /* NB we can't trust socket ops to either consume our iovs
126 * or leave them alone. */
127 if (tx->tx_msg.ksm_zc_cookies[0] != 0) {
128 /* Zero copy is enabled */
129 struct sock *sk = sock->sk;
130 struct page *page = kiov->bv_page;
131 int offset = kiov->bv_offset;
132 int fragsize = kiov->bv_len;
133 int msgflg = MSG_DONTWAIT;
135 CDEBUG(D_NET, "page %p + offset %x for %d\n",
136 page, offset, kiov->bv_len);
138 if (!list_empty(&conn->ksnc_tx_queue) ||
139 fragsize < tx->tx_resid)
142 rc = sk->sk_prot->sendpage(sk, page,
143 offset, fragsize, msgflg);
145 #if SOCKNAL_SINGLE_FRAG_TX || !SOCKNAL_RISK_KMAP_DEADLOCK
147 struct kvec *scratchiov = &scratch;
148 unsigned int niov = 1;
150 #ifdef CONFIG_HIGHMEM
151 #warning "XXX risk of kmap deadlock on multiple frags..."
153 unsigned int niov = tx->tx_nkiov;
155 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
158 for (nob = i = 0; i < niov; i++) {
159 scratchiov[i].iov_base = kmap(kiov[i].bv_page) +
161 nob += scratchiov[i].iov_len = kiov[i].bv_len;
164 if (!list_empty(&conn->ksnc_tx_queue) ||
166 msg.msg_flags |= MSG_MORE;
168 rc = kernel_sendmsg(sock, &msg, scratchiov, niov, nob);
170 for (i = 0; i < niov; i++)
171 kunmap(kiov[i].bv_page);
177 ksocknal_lib_eager_ack(struct ksock_conn *conn)
180 struct socket *sock = conn->ksnc_sock;
182 /* Remind the socket to ACK eagerly. If I don't, the socket might
183 * think I'm about to send something it could piggy-back the ACK
184 * on, introducing delay in completing zero-copy sends in my
187 kernel_setsockopt(sock, SOL_TCP, TCP_QUICKACK,
188 (char *)&opt, sizeof(opt));
192 ksocknal_lib_recv_iov(struct ksock_conn *conn, struct kvec *scratchiov)
194 #if SOCKNAL_SINGLE_FRAG_RX
196 struct kvec *scratchiov = &scratch;
197 unsigned int niov = 1;
199 unsigned int niov = conn->ksnc_rx_niov;
201 struct kvec *iov = conn->ksnc_rx_iov;
202 struct msghdr msg = {
212 /* NB we can't trust socket ops to either consume our iovs
213 * or leave them alone. */
216 for (nob = i = 0; i < niov; i++) {
217 scratchiov[i] = iov[i];
218 nob += scratchiov[i].iov_len;
220 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
222 rc = kernel_recvmsg(conn->ksnc_sock, &msg, scratchiov, niov, nob,
226 if (conn->ksnc_proto == &ksocknal_protocol_v2x) {
227 saved_csum = conn->ksnc_msg.ksm_csum;
228 conn->ksnc_msg.ksm_csum = 0;
231 if (saved_csum != 0) {
232 /* accumulate checksum */
233 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
236 fragnob = iov[i].iov_len;
240 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
241 iov[i].iov_base, fragnob);
243 conn->ksnc_msg.ksm_csum = saved_csum;
250 ksocknal_lib_kiov_vunmap(void *addr)
259 ksocknal_lib_kiov_vmap(struct bio_vec *kiov, int niov,
260 struct kvec *iov, struct page **pages)
266 if (!*ksocknal_tunables.ksnd_zc_recv || pages == NULL)
269 LASSERT (niov <= LNET_MAX_IOV);
272 niov < *ksocknal_tunables.ksnd_zc_recv_min_nfrags)
275 for (nob = i = 0; i < niov; i++) {
276 if ((kiov[i].bv_offset != 0 && i > 0) ||
277 (kiov[i].bv_offset + kiov[i].bv_len !=
278 PAGE_SIZE && i < niov - 1))
281 pages[i] = kiov[i].bv_page;
282 nob += kiov[i].bv_len;
285 addr = vmap(pages, niov, VM_MAP, PAGE_KERNEL);
289 iov->iov_base = addr + kiov[0].bv_offset;
296 ksocknal_lib_recv_kiov(struct ksock_conn *conn, struct page **pages,
297 struct kvec *scratchiov)
299 #if SOCKNAL_SINGLE_FRAG_RX || !SOCKNAL_RISK_KMAP_DEADLOCK
301 struct kvec *scratchiov = &scratch;
302 struct page **pages = NULL;
303 unsigned int niov = 1;
305 #ifdef CONFIG_HIGHMEM
306 #warning "XXX risk of kmap deadlock on multiple frags..."
308 unsigned int niov = conn->ksnc_rx_nkiov;
310 struct bio_vec *kiov = conn->ksnc_rx_kiov;
311 struct msghdr msg = {
323 /* NB we can't trust socket ops to either consume our iovs
324 * or leave them alone. */
325 if ((addr = ksocknal_lib_kiov_vmap(kiov, niov, scratchiov, pages)) != NULL) {
326 nob = scratchiov[0].iov_len;
330 for (nob = i = 0; i < niov; i++) {
331 nob += scratchiov[i].iov_len = kiov[i].bv_len;
332 scratchiov[i].iov_base = kmap(kiov[i].bv_page) +
338 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
340 rc = kernel_recvmsg(conn->ksnc_sock, &msg, scratchiov, n, nob,
343 if (conn->ksnc_msg.ksm_csum != 0) {
344 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
347 /* Dang! have to kmap again because I have nowhere to
348 * stash the mapped address. But by doing it while the
349 * page is still mapped, the kernel just bumps the map
350 * count and returns me the address it stashed.
352 base = kmap(kiov[i].bv_page) + kiov[i].bv_offset;
353 fragnob = kiov[i].bv_len;
357 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
360 kunmap(kiov[i].bv_page);
365 ksocknal_lib_kiov_vunmap(addr);
367 for (i = 0; i < niov; i++)
368 kunmap(kiov[i].bv_page);
375 ksocknal_lib_csum_tx(struct ksock_tx *tx)
381 LASSERT(tx->tx_hdr.iov_base == (void *)&tx->tx_msg);
382 LASSERT(tx->tx_conn != NULL);
383 LASSERT(tx->tx_conn->ksnc_proto == &ksocknal_protocol_v2x);
385 tx->tx_msg.ksm_csum = 0;
387 csum = ksocknal_csum(~0, (void *)tx->tx_hdr.iov_base,
390 for (i = 0; i < tx->tx_nkiov; i++) {
391 base = kmap(tx->tx_kiov[i].bv_page) +
392 tx->tx_kiov[i].bv_offset;
394 csum = ksocknal_csum(csum, base, tx->tx_kiov[i].bv_len);
396 kunmap(tx->tx_kiov[i].bv_page);
399 if (*ksocknal_tunables.ksnd_inject_csum_error) {
401 *ksocknal_tunables.ksnd_inject_csum_error = 0;
404 tx->tx_msg.ksm_csum = csum;
408 ksocknal_lib_get_conn_tunables(struct ksock_conn *conn, int *txmem, int *rxmem, int *nagle)
410 struct socket *sock = conn->ksnc_sock;
414 rc = ksocknal_connsock_addref(conn);
416 LASSERT (conn->ksnc_closing);
417 *txmem = *rxmem = *nagle = 0;
421 rc = lnet_sock_getbuf(sock, txmem, rxmem);
423 len = sizeof(*nagle);
424 rc = kernel_getsockopt(sock, SOL_TCP, TCP_NODELAY,
425 (char *)nagle, &len);
428 ksocknal_connsock_decref(conn);
433 *txmem = *rxmem = *nagle = 0;
439 ksocknal_lib_setup_sock (struct socket *sock)
447 struct linger linger;
449 sock->sk->sk_allocation = GFP_NOFS;
451 /* Ensure this socket aborts active sends immediately when we close
457 rc = kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER,
458 (char *)&linger, sizeof(linger));
460 CERROR ("Can't set SO_LINGER: %d\n", rc);
465 rc = kernel_setsockopt(sock, SOL_TCP, TCP_LINGER2,
466 (char *)&option, sizeof(option));
468 CERROR ("Can't set SO_LINGER2: %d\n", rc);
472 if (!*ksocknal_tunables.ksnd_nagle) {
475 rc = kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY,
476 (char *)&option, sizeof(option));
478 CERROR ("Can't disable nagle: %d\n", rc);
483 rc = lnet_sock_setbuf(sock,
484 *ksocknal_tunables.ksnd_tx_buffer_size,
485 *ksocknal_tunables.ksnd_rx_buffer_size);
487 CERROR ("Can't set buffer tx %d, rx %d buffers: %d\n",
488 *ksocknal_tunables.ksnd_tx_buffer_size,
489 *ksocknal_tunables.ksnd_rx_buffer_size, rc);
493 /* TCP_BACKOFF_* sockopt tunables unsupported in stock kernels */
494 #ifdef SOCKNAL_BACKOFF
495 if (*ksocknal_tunables.ksnd_backoff_init > 0) {
496 option = *ksocknal_tunables.ksnd_backoff_init;
497 #ifdef SOCKNAL_BACKOFF_MS
501 rc = kernel_setsockopt(sock, SOL_TCP, TCP_BACKOFF_INIT,
502 (char *)&option, sizeof(option));
504 CERROR ("Can't set initial tcp backoff %d: %d\n",
510 if (*ksocknal_tunables.ksnd_backoff_max > 0) {
511 option = *ksocknal_tunables.ksnd_backoff_max;
512 #ifdef SOCKNAL_BACKOFF_MS
516 rc = kernel_setsockopt(sock, SOL_TCP, TCP_BACKOFF_MAX,
517 (char *)&option, sizeof(option));
519 CERROR ("Can't set maximum tcp backoff %d: %d\n",
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);
534 rc = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
535 (char *)&option, sizeof(option));
537 CERROR ("Can't set SO_KEEPALIVE: %d\n", rc);
544 rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPIDLE,
545 (char *)&keep_idle, sizeof(keep_idle));
547 CERROR ("Can't set TCP_KEEPIDLE: %d\n", rc);
551 rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPINTVL,
552 (char *)&keep_intvl, sizeof(keep_intvl));
554 CERROR ("Can't set TCP_KEEPINTVL: %d\n", rc);
558 rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPCNT,
559 (char *)&keep_count, sizeof(keep_count));
561 CERROR ("Can't set TCP_KEEPCNT: %d\n", rc);
569 ksocknal_lib_push_conn(struct ksock_conn *conn)
577 rc = ksocknal_connsock_addref(conn);
578 if (rc != 0) /* being shut down */
581 sk = conn->ksnc_sock->sk;
585 nonagle = tp->nonagle;
589 rc = kernel_setsockopt(conn->ksnc_sock, SOL_TCP, TCP_NODELAY,
590 (char *)&val, sizeof(val));
594 tp->nonagle = nonagle;
597 ksocknal_connsock_decref(conn);
600 void ksocknal_read_callback(struct ksock_conn *conn);
601 void ksocknal_write_callback(struct ksock_conn *conn);
603 * socket call back in Linux
606 #ifdef HAVE_SK_DATA_READY_ONE_ARG
607 ksocknal_data_ready(struct sock *sk)
609 ksocknal_data_ready(struct sock *sk, int n)
612 struct ksock_conn *conn;
615 /* interleave correctly with closing sockets... */
617 read_lock(&ksocknal_data.ksnd_global_lock);
619 conn = sk->sk_user_data;
620 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
621 LASSERT(sk->sk_data_ready != &ksocknal_data_ready);
622 #ifdef HAVE_SK_DATA_READY_ONE_ARG
623 sk->sk_data_ready(sk);
625 sk->sk_data_ready(sk, n);
628 ksocknal_read_callback(conn);
630 read_unlock(&ksocknal_data.ksnd_global_lock);
636 ksocknal_write_space (struct sock *sk)
638 struct ksock_conn *conn;
642 /* interleave correctly with closing sockets... */
644 read_lock(&ksocknal_data.ksnd_global_lock);
646 conn = sk->sk_user_data;
647 wspace = sk_stream_wspace(sk);
648 min_wpace = sk_stream_min_wspace(sk);
650 CDEBUG(D_NET, "sk %p wspace %d low water %d conn %p%s%s%s\n",
651 sk, wspace, min_wpace, conn,
652 (conn == NULL) ? "" : (conn->ksnc_tx_ready ?
653 " ready" : " blocked"),
654 (conn == NULL) ? "" : (conn->ksnc_tx_scheduled ?
655 " scheduled" : " idle"),
656 (conn == NULL) ? "" : (list_empty(&conn->ksnc_tx_queue) ?
657 " empty" : " queued"));
659 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
660 LASSERT (sk->sk_write_space != &ksocknal_write_space);
661 sk->sk_write_space (sk);
663 read_unlock(&ksocknal_data.ksnd_global_lock);
667 if (wspace >= min_wpace) { /* got enough space */
668 ksocknal_write_callback(conn);
670 /* Clear SOCK_NOSPACE _after_ ksocknal_write_callback so the
671 * ENOMEM check in ksocknal_transmit is race-free (think about
674 clear_bit (SOCK_NOSPACE, &sk->sk_socket->flags);
677 read_unlock(&ksocknal_data.ksnd_global_lock);
681 ksocknal_lib_save_callback(struct socket *sock, struct ksock_conn *conn)
683 conn->ksnc_saved_data_ready = sock->sk->sk_data_ready;
684 conn->ksnc_saved_write_space = sock->sk->sk_write_space;
688 ksocknal_lib_set_callback(struct socket *sock, struct ksock_conn *conn)
690 sock->sk->sk_user_data = conn;
691 sock->sk->sk_data_ready = ksocknal_data_ready;
692 sock->sk->sk_write_space = ksocknal_write_space;
696 ksocknal_lib_reset_callback(struct socket *sock, struct ksock_conn *conn)
698 /* Remove conn's network callbacks.
699 * NB I _have_ to restore the callback, rather than storing a noop,
700 * since the socket could survive past this module being unloaded!! */
701 sock->sk->sk_data_ready = conn->ksnc_saved_data_ready;
702 sock->sk->sk_write_space = conn->ksnc_saved_write_space;
704 /* A callback could be in progress already; they hold a read lock
705 * on ksnd_global_lock (to serialise with me) and NOOP if
706 * sk_user_data is NULL. */
707 sock->sk->sk_user_data = NULL;
713 ksocknal_lib_memory_pressure(struct ksock_conn *conn)
716 struct ksock_sched *sched;
718 sched = conn->ksnc_scheduler;
719 spin_lock_bh(&sched->kss_lock);
721 if (!test_bit(SOCK_NOSPACE, &conn->ksnc_sock->flags) &&
722 !conn->ksnc_tx_ready) {
723 /* SOCK_NOSPACE is set when the socket fills
724 * and cleared in the write_space callback
725 * (which also sets ksnc_tx_ready). If
726 * SOCK_NOSPACE and ksnc_tx_ready are BOTH
727 * zero, I didn't fill the socket and
728 * write_space won't reschedule me, so I
729 * return -ENOMEM to get my caller to retry
734 spin_unlock_bh(&sched->kss_lock);