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_hdr(struct ksock_conn *conn, struct ksock_tx *tx,
75 struct kvec *scratchiov)
77 struct socket *sock = conn->ksnc_sock;
81 if (*ksocknal_tunables.ksnd_enable_csum && /* checksum enabled */
82 conn->ksnc_proto == &ksocknal_protocol_v2x && /* V2.x connection */
83 tx->tx_nob == tx->tx_resid && /* frist sending */
84 tx->tx_msg.ksm_csum == 0) /* not checksummed */
85 ksocknal_lib_csum_tx(tx);
87 /* NB we can't trust socket ops to either consume our iovs
88 * or leave them alone. */
91 #if SOCKNAL_SINGLE_FRAG_TX
93 struct kvec *scratchiov = &scratch;
94 unsigned int niov = 1;
96 unsigned int niov = tx->tx_niov;
98 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
101 scratchiov[0] = tx->tx_hdr;
102 nob += scratchiov[0].iov_len;
105 if (!list_empty(&conn->ksnc_tx_queue) ||
107 msg.msg_flags |= MSG_MORE;
109 rc = kernel_sendmsg(sock, &msg, scratchiov, niov, nob);
115 ksocknal_lib_send_kiov(struct ksock_conn *conn, struct ksock_tx *tx,
116 struct kvec *scratchiov)
118 struct socket *sock = conn->ksnc_sock;
119 struct bio_vec *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->bv_page;
132 int offset = kiov->bv_offset;
133 int fragsize = kiov->bv_len;
134 int msgflg = MSG_DONTWAIT;
136 CDEBUG(D_NET, "page %p + offset %x for %d\n",
137 page, offset, kiov->bv_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 = tcp_sendpage(sk, page, offset, fragsize, msgflg);
150 #if SOCKNAL_SINGLE_FRAG_TX || !SOCKNAL_RISK_KMAP_DEADLOCK
152 struct kvec *scratchiov = &scratch;
153 unsigned int niov = 1;
155 #ifdef CONFIG_HIGHMEM
156 #warning "XXX risk of kmap deadlock on multiple frags..."
158 unsigned int niov = tx->tx_nkiov;
160 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
163 for (nob = i = 0; i < niov; i++) {
164 scratchiov[i].iov_base = kmap(kiov[i].bv_page) +
166 nob += scratchiov[i].iov_len = kiov[i].bv_len;
169 if (!list_empty(&conn->ksnc_tx_queue) ||
171 msg.msg_flags |= MSG_MORE;
173 rc = kernel_sendmsg(sock, &msg, scratchiov, niov, nob);
175 for (i = 0; i < niov; i++)
176 kunmap(kiov[i].bv_page);
182 ksocknal_lib_eager_ack(struct ksock_conn *conn)
185 struct socket *sock = conn->ksnc_sock;
187 /* Remind the socket to ACK eagerly. If I don't, the socket might
188 * think I'm about to send something it could piggy-back the ACK
189 * on, introducing delay in completing zero-copy sends in my
192 kernel_setsockopt(sock, SOL_TCP, TCP_QUICKACK,
193 (char *)&opt, sizeof(opt));
197 ksocknal_lib_recv_iov(struct ksock_conn *conn, struct kvec *scratchiov)
199 #if SOCKNAL_SINGLE_FRAG_RX
201 struct kvec *scratchiov = &scratch;
202 unsigned int niov = 1;
204 unsigned int niov = conn->ksnc_rx_niov;
206 struct kvec *iov = conn->ksnc_rx_iov;
207 struct msghdr msg = {
217 /* NB we can't trust socket ops to either consume our iovs
218 * or leave them alone. */
221 for (nob = i = 0; i < niov; i++) {
222 scratchiov[i] = iov[i];
223 nob += scratchiov[i].iov_len;
225 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
227 rc = kernel_recvmsg(conn->ksnc_sock, &msg, scratchiov, niov, nob,
231 if (conn->ksnc_proto == &ksocknal_protocol_v2x) {
232 saved_csum = conn->ksnc_msg.ksm_csum;
233 conn->ksnc_msg.ksm_csum = 0;
236 if (saved_csum != 0) {
237 /* accumulate checksum */
238 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
241 fragnob = iov[i].iov_len;
245 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
246 iov[i].iov_base, fragnob);
248 conn->ksnc_msg.ksm_csum = saved_csum;
255 ksocknal_lib_kiov_vunmap(void *addr)
264 ksocknal_lib_kiov_vmap(struct bio_vec *kiov, int niov,
265 struct kvec *iov, struct page **pages)
271 if (!*ksocknal_tunables.ksnd_zc_recv || pages == NULL)
274 LASSERT (niov <= LNET_MAX_IOV);
277 niov < *ksocknal_tunables.ksnd_zc_recv_min_nfrags)
280 for (nob = i = 0; i < niov; i++) {
281 if ((kiov[i].bv_offset != 0 && i > 0) ||
282 (kiov[i].bv_offset + kiov[i].bv_len !=
283 PAGE_SIZE && i < niov - 1))
286 pages[i] = kiov[i].bv_page;
287 nob += kiov[i].bv_len;
290 addr = vmap(pages, niov, VM_MAP, PAGE_KERNEL);
294 iov->iov_base = addr + kiov[0].bv_offset;
301 ksocknal_lib_recv_kiov(struct ksock_conn *conn, struct page **pages,
302 struct kvec *scratchiov)
304 #if SOCKNAL_SINGLE_FRAG_RX || !SOCKNAL_RISK_KMAP_DEADLOCK
306 struct kvec *scratchiov = &scratch;
307 struct page **pages = NULL;
308 unsigned int niov = 1;
310 #ifdef CONFIG_HIGHMEM
311 #warning "XXX risk of kmap deadlock on multiple frags..."
313 unsigned int niov = conn->ksnc_rx_nkiov;
315 struct bio_vec *kiov = conn->ksnc_rx_kiov;
316 struct msghdr msg = {
328 /* NB we can't trust socket ops to either consume our iovs
329 * or leave them alone. */
330 if ((addr = ksocknal_lib_kiov_vmap(kiov, niov, scratchiov, pages)) != NULL) {
331 nob = scratchiov[0].iov_len;
335 for (nob = i = 0; i < niov; i++) {
336 nob += scratchiov[i].iov_len = kiov[i].bv_len;
337 scratchiov[i].iov_base = kmap(kiov[i].bv_page) +
343 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
345 rc = kernel_recvmsg(conn->ksnc_sock, &msg, scratchiov, n, nob,
348 if (conn->ksnc_msg.ksm_csum != 0) {
349 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
352 /* Dang! have to kmap again because I have nowhere to
353 * stash the mapped address. But by doing it while the
354 * page is still mapped, the kernel just bumps the map
355 * count and returns me the address it stashed.
357 base = kmap(kiov[i].bv_page) + kiov[i].bv_offset;
358 fragnob = kiov[i].bv_len;
362 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
365 kunmap(kiov[i].bv_page);
370 ksocknal_lib_kiov_vunmap(addr);
372 for (i = 0; i < niov; i++)
373 kunmap(kiov[i].bv_page);
380 ksocknal_lib_csum_tx(struct ksock_tx *tx)
386 LASSERT(tx->tx_hdr.iov_base == (void *)&tx->tx_msg);
387 LASSERT(tx->tx_conn != NULL);
388 LASSERT(tx->tx_conn->ksnc_proto == &ksocknal_protocol_v2x);
390 tx->tx_msg.ksm_csum = 0;
392 csum = ksocknal_csum(~0, (void *)tx->tx_hdr.iov_base,
395 for (i = 0; i < tx->tx_nkiov; i++) {
396 base = kmap(tx->tx_kiov[i].bv_page) +
397 tx->tx_kiov[i].bv_offset;
399 csum = ksocknal_csum(csum, base, tx->tx_kiov[i].bv_len);
401 kunmap(tx->tx_kiov[i].bv_page);
404 if (*ksocknal_tunables.ksnd_inject_csum_error) {
406 *ksocknal_tunables.ksnd_inject_csum_error = 0;
409 tx->tx_msg.ksm_csum = csum;
413 ksocknal_lib_get_conn_tunables(struct ksock_conn *conn, int *txmem, int *rxmem, int *nagle)
415 struct socket *sock = conn->ksnc_sock;
419 rc = ksocknal_connsock_addref(conn);
421 LASSERT (conn->ksnc_closing);
422 *txmem = *rxmem = *nagle = 0;
426 rc = lnet_sock_getbuf(sock, txmem, rxmem);
428 len = sizeof(*nagle);
429 rc = kernel_getsockopt(sock, SOL_TCP, TCP_NODELAY,
430 (char *)nagle, &len);
433 ksocknal_connsock_decref(conn);
438 *txmem = *rxmem = *nagle = 0;
444 ksocknal_lib_setup_sock (struct socket *sock)
452 struct linger linger;
454 sock->sk->sk_allocation = GFP_NOFS;
456 /* Ensure this socket aborts active sends immediately when we close
462 rc = kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER,
463 (char *)&linger, sizeof(linger));
465 CERROR ("Can't set SO_LINGER: %d\n", rc);
470 rc = kernel_setsockopt(sock, SOL_TCP, TCP_LINGER2,
471 (char *)&option, sizeof(option));
473 CERROR ("Can't set SO_LINGER2: %d\n", rc);
477 if (!*ksocknal_tunables.ksnd_nagle) {
480 rc = kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY,
481 (char *)&option, sizeof(option));
483 CERROR ("Can't disable nagle: %d\n", rc);
488 rc = lnet_sock_setbuf(sock,
489 *ksocknal_tunables.ksnd_tx_buffer_size,
490 *ksocknal_tunables.ksnd_rx_buffer_size);
492 CERROR ("Can't set buffer tx %d, rx %d buffers: %d\n",
493 *ksocknal_tunables.ksnd_tx_buffer_size,
494 *ksocknal_tunables.ksnd_rx_buffer_size, rc);
498 /* TCP_BACKOFF_* sockopt tunables unsupported in stock kernels */
499 #ifdef SOCKNAL_BACKOFF
500 if (*ksocknal_tunables.ksnd_backoff_init > 0) {
501 option = *ksocknal_tunables.ksnd_backoff_init;
502 #ifdef SOCKNAL_BACKOFF_MS
506 rc = kernel_setsockopt(sock, SOL_TCP, TCP_BACKOFF_INIT,
507 (char *)&option, sizeof(option));
509 CERROR ("Can't set initial tcp backoff %d: %d\n",
515 if (*ksocknal_tunables.ksnd_backoff_max > 0) {
516 option = *ksocknal_tunables.ksnd_backoff_max;
517 #ifdef SOCKNAL_BACKOFF_MS
521 rc = kernel_setsockopt(sock, SOL_TCP, TCP_BACKOFF_MAX,
522 (char *)&option, sizeof(option));
524 CERROR ("Can't set maximum tcp backoff %d: %d\n",
531 /* snapshot tunables */
532 keep_idle = *ksocknal_tunables.ksnd_keepalive_idle;
533 keep_count = *ksocknal_tunables.ksnd_keepalive_count;
534 keep_intvl = *ksocknal_tunables.ksnd_keepalive_intvl;
536 do_keepalive = (keep_idle > 0 && keep_count > 0 && keep_intvl > 0);
538 option = (do_keepalive ? 1 : 0);
539 rc = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
540 (char *)&option, sizeof(option));
542 CERROR ("Can't set SO_KEEPALIVE: %d\n", rc);
549 rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPIDLE,
550 (char *)&keep_idle, sizeof(keep_idle));
552 CERROR ("Can't set TCP_KEEPIDLE: %d\n", rc);
556 rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPINTVL,
557 (char *)&keep_intvl, sizeof(keep_intvl));
559 CERROR ("Can't set TCP_KEEPINTVL: %d\n", rc);
563 rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPCNT,
564 (char *)&keep_count, sizeof(keep_count));
566 CERROR ("Can't set TCP_KEEPCNT: %d\n", rc);
574 ksocknal_lib_push_conn(struct ksock_conn *conn)
582 rc = ksocknal_connsock_addref(conn);
583 if (rc != 0) /* being shut down */
586 sk = conn->ksnc_sock->sk;
590 nonagle = tp->nonagle;
594 rc = kernel_setsockopt(conn->ksnc_sock, SOL_TCP, TCP_NODELAY,
595 (char *)&val, sizeof(val));
599 tp->nonagle = nonagle;
602 ksocknal_connsock_decref(conn);
605 void ksocknal_read_callback(struct ksock_conn *conn);
606 void ksocknal_write_callback(struct ksock_conn *conn);
608 * socket call back in Linux
611 #ifdef HAVE_SK_DATA_READY_ONE_ARG
612 ksocknal_data_ready(struct sock *sk)
614 ksocknal_data_ready(struct sock *sk, int n)
617 struct ksock_conn *conn;
620 /* interleave correctly with closing sockets... */
622 read_lock(&ksocknal_data.ksnd_global_lock);
624 conn = sk->sk_user_data;
625 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
626 LASSERT(sk->sk_data_ready != &ksocknal_data_ready);
627 #ifdef HAVE_SK_DATA_READY_ONE_ARG
628 sk->sk_data_ready(sk);
630 sk->sk_data_ready(sk, n);
633 ksocknal_read_callback(conn);
635 read_unlock(&ksocknal_data.ksnd_global_lock);
641 ksocknal_write_space (struct sock *sk)
643 struct ksock_conn *conn;
647 /* interleave correctly with closing sockets... */
649 read_lock(&ksocknal_data.ksnd_global_lock);
651 conn = sk->sk_user_data;
652 wspace = sk_stream_wspace(sk);
653 min_wpace = sk_stream_min_wspace(sk);
655 CDEBUG(D_NET, "sk %p wspace %d low water %d conn %p%s%s%s\n",
656 sk, wspace, min_wpace, conn,
657 (conn == NULL) ? "" : (conn->ksnc_tx_ready ?
658 " ready" : " blocked"),
659 (conn == NULL) ? "" : (conn->ksnc_tx_scheduled ?
660 " scheduled" : " idle"),
661 (conn == NULL) ? "" : (list_empty(&conn->ksnc_tx_queue) ?
662 " empty" : " queued"));
664 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
665 LASSERT (sk->sk_write_space != &ksocknal_write_space);
666 sk->sk_write_space (sk);
668 read_unlock(&ksocknal_data.ksnd_global_lock);
672 if (wspace >= min_wpace) { /* got enough space */
673 ksocknal_write_callback(conn);
675 /* Clear SOCK_NOSPACE _after_ ksocknal_write_callback so the
676 * ENOMEM check in ksocknal_transmit is race-free (think about
679 clear_bit (SOCK_NOSPACE, &sk->sk_socket->flags);
682 read_unlock(&ksocknal_data.ksnd_global_lock);
686 ksocknal_lib_save_callback(struct socket *sock, struct ksock_conn *conn)
688 conn->ksnc_saved_data_ready = sock->sk->sk_data_ready;
689 conn->ksnc_saved_write_space = sock->sk->sk_write_space;
693 ksocknal_lib_set_callback(struct socket *sock, struct ksock_conn *conn)
695 sock->sk->sk_user_data = conn;
696 sock->sk->sk_data_ready = ksocknal_data_ready;
697 sock->sk->sk_write_space = ksocknal_write_space;
701 ksocknal_lib_reset_callback(struct socket *sock, struct ksock_conn *conn)
703 /* Remove conn's network callbacks.
704 * NB I _have_ to restore the callback, rather than storing a noop,
705 * since the socket could survive past this module being unloaded!! */
706 sock->sk->sk_data_ready = conn->ksnc_saved_data_ready;
707 sock->sk->sk_write_space = conn->ksnc_saved_write_space;
709 /* A callback could be in progress already; they hold a read lock
710 * on ksnd_global_lock (to serialise with me) and NOOP if
711 * sk_user_data is NULL. */
712 sock->sk->sk_user_data = NULL;
718 ksocknal_lib_memory_pressure(struct ksock_conn *conn)
721 struct ksock_sched *sched;
723 sched = conn->ksnc_scheduler;
724 spin_lock_bh(&sched->kss_lock);
726 if (!test_bit(SOCK_NOSPACE, &conn->ksnc_sock->flags) &&
727 !conn->ksnc_tx_ready) {
728 /* SOCK_NOSPACE is set when the socket fills
729 * and cleared in the write_space callback
730 * (which also sets ksnc_tx_ready). If
731 * SOCK_NOSPACE and ksnc_tx_ready are BOTH
732 * zero, I didn't fill the socket and
733 * write_space won't reschedule me, so I
734 * return -ENOMEM to get my caller to retry
739 spin_unlock_bh(&sched->kss_lock);