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/
35 ksocknal_lib_get_conn_addrs(struct ksock_conn *conn)
37 int rc = lnet_sock_getaddr(conn->ksnc_sock, true,
38 &conn->ksnc_peeraddr);
40 /* Didn't need the {get,put}connsock dance to deref ksnc_sock... */
41 LASSERT(!conn->ksnc_closing);
44 CERROR("Error %d getting sock peer_ni IP\n", rc);
48 rc = lnet_sock_getaddr(conn->ksnc_sock, false,
51 CERROR("Error %d getting sock local IP\n", rc);
59 ksocknal_lib_zc_capable(struct ksock_conn *conn)
61 int caps = conn->ksnc_sock->sk->sk_route_caps;
63 if (conn->ksnc_proto == &ksocknal_protocol_v1x)
66 /* ZC if the socket supports scatter/gather and doesn't need software
68 return ((caps & NETIF_F_SG) != 0 && (caps & NETIF_F_CSUM_MASK) != 0);
72 ksocknal_lib_send_hdr(struct ksock_conn *conn, struct ksock_tx *tx,
73 struct kvec *scratchiov)
75 struct socket *sock = conn->ksnc_sock;
79 if (*ksocknal_tunables.ksnd_enable_csum && /* checksum enabled */
80 conn->ksnc_proto == &ksocknal_protocol_v2x && /* V2.x connection */
81 tx->tx_nob == tx->tx_resid && /* frist sending */
82 tx->tx_msg.ksm_csum == 0) /* not checksummed */
83 ksocknal_lib_csum_tx(tx);
85 /* NB we can't trust socket ops to either consume our iovs
86 * or leave them alone. */
89 #if SOCKNAL_SINGLE_FRAG_TX
91 struct kvec *scratchiov = &scratch;
92 unsigned int niov = 1;
94 unsigned int niov = tx->tx_niov;
96 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
99 scratchiov[0] = tx->tx_hdr;
100 nob += scratchiov[0].iov_len;
103 if (!list_empty(&conn->ksnc_tx_queue) ||
105 msg.msg_flags |= MSG_MORE;
107 rc = kernel_sendmsg(sock, &msg, scratchiov, niov, nob);
113 ksocknal_lib_sendpage(struct socket *sock, struct bio_vec *kiov, int msgflg)
115 #ifdef MSG_SPLICE_PAGES
116 struct msghdr msg = {.msg_flags = msgflg | MSG_SPLICE_PAGES};
118 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, kiov, 1, kiov->bv_len);
120 return sock_sendmsg(sock, &msg);
122 struct sock *sk = sock->sk;
123 struct page *page = kiov->bv_page;
124 int offset = kiov->bv_offset;
125 int fragsize = kiov->bv_len;
127 return sk->sk_prot->sendpage(sk, page, offset, fragsize, msgflg);
132 ksocknal_lib_send_kiov(struct ksock_conn *conn, struct ksock_tx *tx,
133 struct kvec *scratchiov)
135 struct socket *sock = conn->ksnc_sock;
136 struct bio_vec *kiov = tx->tx_kiov;
140 /* Not NOOP message */
141 LASSERT(tx->tx_lnetmsg != NULL);
143 /* NB we can't trust socket ops to either consume our iovs
144 * or leave them alone. */
145 if (tx->tx_msg.ksm_zc_cookies[0] != 0) {
146 /* Zero copy is enabled */
147 int msgflg = MSG_DONTWAIT;
149 CDEBUG(D_NET, "page %p + offset %x for %d\n",
150 kiov->bv_page, kiov->bv_offset, kiov->bv_len);
152 if (!list_empty(&conn->ksnc_tx_queue) ||
153 kiov->bv_len < tx->tx_resid)
156 rc = ksocknal_lib_sendpage(sock, kiov, msgflg);
158 #if SOCKNAL_SINGLE_FRAG_TX || !SOCKNAL_RISK_KMAP_DEADLOCK
160 struct kvec *scratchiov = &scratch;
161 unsigned int niov = 1;
163 #ifdef CONFIG_HIGHMEM
164 #warning "XXX risk of kmap deadlock on multiple frags..."
166 unsigned int niov = tx->tx_nkiov;
168 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
171 for (nob = i = 0; i < niov; i++) {
172 scratchiov[i].iov_base = kmap(kiov[i].bv_page) +
174 nob += scratchiov[i].iov_len = kiov[i].bv_len;
177 if (!list_empty(&conn->ksnc_tx_queue) ||
179 msg.msg_flags |= MSG_MORE;
181 rc = kernel_sendmsg(sock, &msg, scratchiov, niov, nob);
183 for (i = 0; i < niov; i++)
184 kunmap(kiov[i].bv_page);
190 ksocknal_lib_eager_ack(struct ksock_conn *conn)
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 on,
196 * introducing delay in completing zero-copy sends in my peer_ni.
199 tcp_sock_set_quickack(sock->sk, 1);
203 ksocknal_lib_recv_iov(struct ksock_conn *conn, struct kvec *scratchiov)
205 #if SOCKNAL_SINGLE_FRAG_RX
207 struct kvec *scratchiov = &scratch;
208 unsigned int niov = 1;
210 unsigned int niov = conn->ksnc_rx_niov;
212 struct kvec *iov = conn->ksnc_rx_iov;
213 struct msghdr msg = {
223 /* NB we can't trust socket ops to either consume our iovs
224 * or leave them alone. */
227 for (nob = i = 0; i < niov; i++) {
228 scratchiov[i] = iov[i];
229 nob += scratchiov[i].iov_len;
231 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
233 rc = kernel_recvmsg(conn->ksnc_sock, &msg, scratchiov, niov, nob,
237 if (conn->ksnc_proto == &ksocknal_protocol_v2x) {
238 saved_csum = conn->ksnc_msg.ksm_csum;
239 conn->ksnc_msg.ksm_csum = 0;
242 if (saved_csum != 0) {
243 /* accumulate checksum */
244 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
247 fragnob = iov[i].iov_len;
251 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
252 iov[i].iov_base, fragnob);
254 conn->ksnc_msg.ksm_csum = saved_csum;
261 ksocknal_lib_kiov_vunmap(void *addr)
270 ksocknal_lib_kiov_vmap(struct bio_vec *kiov, int niov,
271 struct kvec *iov, struct page **pages)
277 if (!*ksocknal_tunables.ksnd_zc_recv || pages == NULL)
280 LASSERT (niov <= LNET_MAX_IOV);
283 niov < *ksocknal_tunables.ksnd_zc_recv_min_nfrags)
286 for (nob = i = 0; i < niov; i++) {
287 if ((kiov[i].bv_offset != 0 && i > 0) ||
288 (kiov[i].bv_offset + kiov[i].bv_len !=
289 PAGE_SIZE && i < niov - 1))
292 pages[i] = kiov[i].bv_page;
293 nob += kiov[i].bv_len;
296 addr = vmap(pages, niov, VM_MAP, PAGE_KERNEL);
300 iov->iov_base = addr + kiov[0].bv_offset;
307 ksocknal_lib_recv_kiov(struct ksock_conn *conn, struct page **pages,
308 struct kvec *scratchiov)
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 unsigned int niov = conn->ksnc_rx_nkiov;
321 struct bio_vec *kiov = conn->ksnc_rx_kiov;
322 struct msghdr msg = {
334 /* NB we can't trust socket ops to either consume our iovs
335 * or leave them alone. */
336 if ((addr = ksocknal_lib_kiov_vmap(kiov, niov, scratchiov, pages)) != NULL) {
337 nob = scratchiov[0].iov_len;
341 for (nob = i = 0; i < niov; i++) {
342 nob += scratchiov[i].iov_len = kiov[i].bv_len;
343 scratchiov[i].iov_base = kmap(kiov[i].bv_page) +
349 LASSERT (nob <= conn->ksnc_rx_nob_wanted);
351 rc = kernel_recvmsg(conn->ksnc_sock, &msg, scratchiov, n, nob,
354 if (conn->ksnc_msg.ksm_csum != 0) {
355 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
358 /* Dang! have to kmap again because I have nowhere to
359 * stash the mapped address. But by doing it while the
360 * page is still mapped, the kernel just bumps the map
361 * count and returns me the address it stashed.
363 base = kmap(kiov[i].bv_page) + kiov[i].bv_offset;
364 fragnob = kiov[i].bv_len;
368 conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
371 kunmap(kiov[i].bv_page);
376 ksocknal_lib_kiov_vunmap(addr);
378 for (i = 0; i < niov; i++)
379 kunmap(kiov[i].bv_page);
386 ksocknal_lib_csum_tx(struct ksock_tx *tx)
392 LASSERT(tx->tx_hdr.iov_base == (void *)&tx->tx_msg);
393 LASSERT(tx->tx_conn != NULL);
394 LASSERT(tx->tx_conn->ksnc_proto == &ksocknal_protocol_v2x);
396 tx->tx_msg.ksm_csum = 0;
398 csum = ksocknal_csum(~0, (void *)tx->tx_hdr.iov_base,
401 for (i = 0; i < tx->tx_nkiov; i++) {
402 base = kmap(tx->tx_kiov[i].bv_page) +
403 tx->tx_kiov[i].bv_offset;
405 csum = ksocknal_csum(csum, base, tx->tx_kiov[i].bv_len);
407 kunmap(tx->tx_kiov[i].bv_page);
410 if (*ksocknal_tunables.ksnd_inject_csum_error) {
412 *ksocknal_tunables.ksnd_inject_csum_error = 0;
415 tx->tx_msg.ksm_csum = csum;
419 ksocknal_lib_get_conn_tunables(struct ksock_conn *conn, int *txmem, int *rxmem, int *nagle)
421 struct socket *sock = conn->ksnc_sock;
422 struct tcp_sock *tp = tcp_sk(sock->sk);
424 if (ksocknal_connsock_addref(conn) < 0) {
425 LASSERT(conn->ksnc_closing);
432 lnet_sock_getbuf(sock, txmem, rxmem);
434 *nagle = !(tp->nonagle & TCP_NAGLE_OFF);
436 ksocknal_connsock_decref(conn);
443 ksocknal_lib_setup_sock(struct socket *sock, struct lnet_ni *ni)
450 struct tcp_sock *tp = tcp_sk(sock->sk);
451 struct lnet_ioctl_config_socklnd_tunables *lndtun;
453 sock->sk->sk_allocation = GFP_NOFS;
455 /* Ensure this socket aborts active sends immediately when closed. */
456 sock_reset_flag(sock->sk, SOCK_LINGER);
460 if (!*ksocknal_tunables.ksnd_nagle)
461 tcp_sock_set_nodelay(sock->sk);
463 lnet_sock_setbuf(sock,
464 *ksocknal_tunables.ksnd_tx_buffer_size,
465 *ksocknal_tunables.ksnd_rx_buffer_size);
467 /* TCP_BACKOFF_* sockopt tunables unsupported in stock kernels */
468 #ifdef SOCKNAL_BACKOFF
469 if (*ksocknal_tunables.ksnd_backoff_init > 0) {
470 int option = *ksocknal_tunables.ksnd_backoff_init;
471 #ifdef SOCKNAL_BACKOFF_MS
475 rc = kernel_setsockopt(sock, SOL_TCP, TCP_BACKOFF_INIT,
476 (char *)&option, sizeof(option));
478 CERROR("Can't set initial tcp backoff %d: %d\n",
484 if (*ksocknal_tunables.ksnd_backoff_max > 0) {
485 int option = *ksocknal_tunables.ksnd_backoff_max;
486 #ifdef SOCKNAL_BACKOFF_MS
490 rc = kernel_setsockopt(sock, SOL_TCP, TCP_BACKOFF_MAX,
491 (char *)&option, sizeof(option));
493 CERROR("Can't set maximum tcp backoff %d: %d\n",
500 /* snapshot tunables */
501 keep_idle = *ksocknal_tunables.ksnd_keepalive_idle;
502 keep_count = *ksocknal_tunables.ksnd_keepalive_count;
503 keep_intvl = *ksocknal_tunables.ksnd_keepalive_intvl;
505 do_keepalive = (keep_idle > 0 && keep_count > 0 && keep_intvl > 0);
507 #ifdef HAVE_KERNEL_SETSOCKOPT
508 /* open-coded version doesn't work in all kernels, and
509 * there is no helper function, so call kernel_setsockopt()
513 int option = (do_keepalive ? 1 : 0);
514 kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
515 (char *)&option, sizeof(option));
518 if (sock->sk->sk_prot->keepalive)
519 sock->sk->sk_prot->keepalive(sock->sk, do_keepalive);
521 sock_set_flag(sock->sk, SOCK_KEEPOPEN);
523 sock_reset_flag(sock->sk, SOCK_KEEPOPEN);
524 #endif /* HAVE_KERNEL_SETSOCKOPT */
529 rc = tcp_sock_set_keepidle(sock->sk, keep_idle);
531 CERROR("Can't set TCP_KEEPIDLE: %d\n", rc);
535 rc = tcp_sock_set_keepintvl(sock->sk, keep_intvl);
537 CERROR("Can't set TCP_KEEPINTVL: %d\n", rc);
541 rc = tcp_sock_set_keepcnt(sock->sk, keep_count);
543 CERROR("Can't set TCP_KEEPCNT: %d\n", rc);
547 lndtun = &ni->ni_lnd_tunables.lnd_tun_u.lnd_sock;
548 if (lndtun->lnd_tos >= 0)
549 ip_sock_set_tos(sock->sk, lndtun->lnd_tos);
555 ksocknal_lib_push_conn(struct ksock_conn *conn)
562 rc = ksocknal_connsock_addref(conn);
563 if (rc != 0) /* being shut down */
566 sk = conn->ksnc_sock->sk;
570 nonagle = tp->nonagle;
571 tp->nonagle = TCP_NAGLE_OFF;
574 tcp_sock_set_nodelay(conn->ksnc_sock->sk);
577 tp->nonagle = nonagle;
580 ksocknal_connsock_decref(conn);
583 void ksocknal_read_callback(struct ksock_conn *conn);
584 void ksocknal_write_callback(struct ksock_conn *conn);
586 * socket call back in Linux
589 #ifdef HAVE_SK_DATA_READY_ONE_ARG
590 ksocknal_data_ready(struct sock *sk)
592 ksocknal_data_ready(struct sock *sk, int n)
595 struct ksock_conn *conn;
597 /* interleave correctly with closing sockets... */
599 read_lock_bh(&ksocknal_data.ksnd_global_lock);
601 conn = sk->sk_user_data;
602 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
603 LASSERT(sk->sk_data_ready != &ksocknal_data_ready);
604 #ifdef HAVE_SK_DATA_READY_ONE_ARG
605 sk->sk_data_ready(sk);
607 sk->sk_data_ready(sk, n);
610 ksocknal_read_callback(conn);
612 read_unlock_bh(&ksocknal_data.ksnd_global_lock);
616 ksocknal_write_space (struct sock *sk)
618 struct ksock_conn *conn;
622 /* interleave correctly with closing sockets... */
624 read_lock(&ksocknal_data.ksnd_global_lock);
626 conn = sk->sk_user_data;
627 wspace = sk_stream_wspace(sk);
628 min_wpace = sk_stream_min_wspace(sk);
630 CDEBUG(D_NET, "sk %p wspace %d low water %d conn %p%s%s%s\n",
631 sk, wspace, min_wpace, conn,
632 (conn == NULL) ? "" : (conn->ksnc_tx_ready ?
633 " ready" : " blocked"),
634 (conn == NULL) ? "" : (conn->ksnc_tx_scheduled ?
635 " scheduled" : " idle"),
636 (conn == NULL) ? "" : (list_empty(&conn->ksnc_tx_queue) ?
637 " empty" : " queued"));
639 if (conn == NULL) { /* raced with ksocknal_terminate_conn */
640 LASSERT (sk->sk_write_space != &ksocknal_write_space);
641 sk->sk_write_space (sk);
643 read_unlock(&ksocknal_data.ksnd_global_lock);
647 if (wspace >= min_wpace) { /* got enough space */
648 ksocknal_write_callback(conn);
650 /* Clear SOCK_NOSPACE _after_ ksocknal_write_callback so the
651 * ENOMEM check in ksocknal_transmit is race-free (think about
654 clear_bit (SOCK_NOSPACE, &sk->sk_socket->flags);
657 read_unlock(&ksocknal_data.ksnd_global_lock);
661 ksocknal_lib_save_callback(struct socket *sock, struct ksock_conn *conn)
663 conn->ksnc_saved_data_ready = sock->sk->sk_data_ready;
664 conn->ksnc_saved_write_space = sock->sk->sk_write_space;
668 ksocknal_lib_set_callback(struct socket *sock, struct ksock_conn *conn)
670 sock->sk->sk_user_data = conn;
671 sock->sk->sk_data_ready = ksocknal_data_ready;
672 sock->sk->sk_write_space = ksocknal_write_space;
676 ksocknal_lib_reset_callback(struct socket *sock, struct ksock_conn *conn)
678 /* Remove conn's network callbacks.
679 * NB I _have_ to restore the callback, rather than storing a noop,
680 * since the socket could survive past this module being unloaded!! */
681 sock->sk->sk_data_ready = conn->ksnc_saved_data_ready;
682 sock->sk->sk_write_space = conn->ksnc_saved_write_space;
684 /* A callback could be in progress already; they hold a read lock
685 * on ksnd_global_lock (to serialise with me) and NOOP if
686 * sk_user_data is NULL. */
687 sock->sk->sk_user_data = NULL;
693 ksocknal_lib_memory_pressure(struct ksock_conn *conn)
696 struct ksock_sched *sched;
698 sched = conn->ksnc_scheduler;
699 spin_lock_bh(&sched->kss_lock);
701 if (!test_bit(SOCK_NOSPACE, &conn->ksnc_sock->flags) &&
702 !conn->ksnc_tx_ready) {
703 /* SOCK_NOSPACE is set when the socket fills
704 * and cleared in the write_space callback
705 * (which also sets ksnc_tx_ready). If
706 * SOCK_NOSPACE and ksnc_tx_ready are BOTH
707 * zero, I didn't fill the socket and
708 * write_space won't reschedule me, so I
709 * return -ENOMEM to get my caller to retry
714 spin_unlock_bh(&sched->kss_lock);