Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / drivers / staging / lustre / lnet / klnds / socklnd / socklnd_lib-linux.c
diff --git a/kernel/drivers/staging/lustre/lnet/klnds/socklnd/socklnd_lib-linux.c b/kernel/drivers/staging/lustre/lnet/klnds/socklnd/socklnd_lib-linux.c
new file mode 100644 (file)
index 0000000..f5e8ab0
--- /dev/null
@@ -0,0 +1,714 @@
+/*
+ * GPL HEADER START
+ *
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 only,
+ * as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License version 2 for more details (a copy is included
+ * in the LICENSE file that accompanied this code).
+ *
+ * You should have received a copy of the GNU General Public License
+ * version 2 along with this program; If not, see
+ * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
+ *
+ * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
+ *
+ * GPL HEADER END
+ */
+/*
+ * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Use is subject to license terms.
+ *
+ * Copyright (c) 2011, 2012, Intel Corporation.
+ */
+/*
+ * This file is part of Lustre, http://www.lustre.org/
+ * Lustre is a trademark of Sun Microsystems, Inc.
+ */
+
+#include "socklnd.h"
+
+int
+ksocknal_lib_get_conn_addrs(ksock_conn_t *conn)
+{
+       int rc = libcfs_sock_getaddr(conn->ksnc_sock, 1,
+                                    &conn->ksnc_ipaddr,
+                                    &conn->ksnc_port);
+
+       /* Didn't need the {get,put}connsock dance to deref ksnc_sock... */
+       LASSERT(!conn->ksnc_closing);
+
+       if (rc != 0) {
+               CERROR("Error %d getting sock peer IP\n", rc);
+               return rc;
+       }
+
+       rc = libcfs_sock_getaddr(conn->ksnc_sock, 0,
+                                &conn->ksnc_myipaddr, NULL);
+       if (rc != 0) {
+               CERROR("Error %d getting sock local IP\n", rc);
+               return rc;
+       }
+
+       return 0;
+}
+
+int
+ksocknal_lib_zc_capable(ksock_conn_t *conn)
+{
+       int  caps = conn->ksnc_sock->sk->sk_route_caps;
+
+       if (conn->ksnc_proto == &ksocknal_protocol_v1x)
+               return 0;
+
+       /* ZC if the socket supports scatter/gather and doesn't need software
+        * checksums */
+       return ((caps & NETIF_F_SG) != 0 && (caps & NETIF_F_ALL_CSUM) != 0);
+}
+
+int
+ksocknal_lib_send_iov(ksock_conn_t *conn, ksock_tx_t *tx)
+{
+       struct socket *sock = conn->ksnc_sock;
+       int         nob;
+       int         rc;
+
+       if (*ksocknal_tunables.ksnd_enable_csum && /* checksum enabled */
+           conn->ksnc_proto == &ksocknal_protocol_v2x && /* V2.x connection  */
+           tx->tx_nob == tx->tx_resid           && /* frist sending    */
+           tx->tx_msg.ksm_csum == 0)                /* not checksummed  */
+               ksocknal_lib_csum_tx(tx);
+
+       /* NB we can't trust socket ops to either consume our iovs
+        * or leave them alone. */
+
+       {
+#if SOCKNAL_SINGLE_FRAG_TX
+               struct kvec    scratch;
+               struct kvec   *scratchiov = &scratch;
+               unsigned int    niov = 1;
+#else
+               struct kvec   *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
+               unsigned int    niov = tx->tx_niov;
+#endif
+               struct msghdr msg = {.msg_flags = MSG_DONTWAIT};
+               int  i;
+
+               for (nob = i = 0; i < niov; i++) {
+                       scratchiov[i] = tx->tx_iov[i];
+                       nob += scratchiov[i].iov_len;
+               }
+
+               if (!list_empty(&conn->ksnc_tx_queue) ||
+                   nob < tx->tx_resid)
+                       msg.msg_flags |= MSG_MORE;
+
+               rc = kernel_sendmsg(sock, &msg, scratchiov, niov, nob);
+       }
+       return rc;
+}
+
+int
+ksocknal_lib_send_kiov(ksock_conn_t *conn, ksock_tx_t *tx)
+{
+       struct socket *sock = conn->ksnc_sock;
+       lnet_kiov_t   *kiov = tx->tx_kiov;
+       int         rc;
+       int         nob;
+
+       /* Not NOOP message */
+       LASSERT(tx->tx_lnetmsg != NULL);
+
+       /* NB we can't trust socket ops to either consume our iovs
+        * or leave them alone. */
+       if (tx->tx_msg.ksm_zc_cookies[0] != 0) {
+               /* Zero copy is enabled */
+               struct sock   *sk = sock->sk;
+               struct page   *page = kiov->kiov_page;
+               int         offset = kiov->kiov_offset;
+               int         fragsize = kiov->kiov_len;
+               int         msgflg = MSG_DONTWAIT;
+
+               CDEBUG(D_NET, "page %p + offset %x for %d\n",
+                              page, offset, kiov->kiov_len);
+
+               if (!list_empty(&conn->ksnc_tx_queue) ||
+                   fragsize < tx->tx_resid)
+                       msgflg |= MSG_MORE;
+
+               if (sk->sk_prot->sendpage != NULL) {
+                       rc = sk->sk_prot->sendpage(sk, page,
+                                                  offset, fragsize, msgflg);
+               } else {
+                       rc = cfs_tcp_sendpage(sk, page, offset, fragsize,
+                                             msgflg);
+               }
+       } else {
+#if SOCKNAL_SINGLE_FRAG_TX || !SOCKNAL_RISK_KMAP_DEADLOCK
+               struct kvec  scratch;
+               struct kvec *scratchiov = &scratch;
+               unsigned int  niov = 1;
+#else
+#ifdef CONFIG_HIGHMEM
+#warning "XXX risk of kmap deadlock on multiple frags..."
+#endif
+               struct kvec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
+               unsigned int  niov = tx->tx_nkiov;
+#endif
+               struct msghdr msg = {.msg_flags = MSG_DONTWAIT};
+               int        i;
+
+               for (nob = i = 0; i < niov; i++) {
+                       scratchiov[i].iov_base = kmap(kiov[i].kiov_page) +
+                                                kiov[i].kiov_offset;
+                       nob += scratchiov[i].iov_len = kiov[i].kiov_len;
+               }
+
+               if (!list_empty(&conn->ksnc_tx_queue) ||
+                   nob < tx->tx_resid)
+                       msg.msg_flags |= MSG_MORE;
+
+               rc = kernel_sendmsg(sock, &msg, (struct kvec *)scratchiov, niov, nob);
+
+               for (i = 0; i < niov; i++)
+                       kunmap(kiov[i].kiov_page);
+       }
+       return rc;
+}
+
+void
+ksocknal_lib_eager_ack(ksock_conn_t *conn)
+{
+       int         opt = 1;
+       struct socket *sock = conn->ksnc_sock;
+
+       /* Remind the socket to ACK eagerly.  If I don't, the socket might
+        * think I'm about to send something it could piggy-back the ACK
+        * on, introducing delay in completing zero-copy sends in my
+        * peer. */
+
+       kernel_setsockopt(sock, SOL_TCP, TCP_QUICKACK,
+                              (char *)&opt, sizeof(opt));
+}
+
+int
+ksocknal_lib_recv_iov(ksock_conn_t *conn)
+{
+#if SOCKNAL_SINGLE_FRAG_RX
+       struct kvec  scratch;
+       struct kvec *scratchiov = &scratch;
+       unsigned int  niov = 1;
+#else
+       struct kvec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
+       unsigned int  niov = conn->ksnc_rx_niov;
+#endif
+       struct kvec *iov = conn->ksnc_rx_iov;
+       struct msghdr msg = {
+               .msg_flags      = 0
+       };
+       int       nob;
+       int       i;
+       int       rc;
+       int       fragnob;
+       int       sum;
+       __u32   saved_csum;
+
+       /* NB we can't trust socket ops to either consume our iovs
+        * or leave them alone. */
+       LASSERT(niov > 0);
+
+       for (nob = i = 0; i < niov; i++) {
+               scratchiov[i] = iov[i];
+               nob += scratchiov[i].iov_len;
+       }
+       LASSERT(nob <= conn->ksnc_rx_nob_wanted);
+
+       rc = kernel_recvmsg(conn->ksnc_sock, &msg,
+               scratchiov, niov, nob, MSG_DONTWAIT);
+
+       saved_csum = 0;
+       if (conn->ksnc_proto == &ksocknal_protocol_v2x) {
+               saved_csum = conn->ksnc_msg.ksm_csum;
+               conn->ksnc_msg.ksm_csum = 0;
+       }
+
+       if (saved_csum != 0) {
+               /* accumulate checksum */
+               for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
+                       LASSERT(i < niov);
+
+                       fragnob = iov[i].iov_len;
+                       if (fragnob > sum)
+                               fragnob = sum;
+
+                       conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
+                                                          iov[i].iov_base, fragnob);
+               }
+               conn->ksnc_msg.ksm_csum = saved_csum;
+       }
+
+       return rc;
+}
+
+static void
+ksocknal_lib_kiov_vunmap(void *addr)
+{
+       if (addr == NULL)
+               return;
+
+       vunmap(addr);
+}
+
+static void *
+ksocknal_lib_kiov_vmap(lnet_kiov_t *kiov, int niov,
+                      struct kvec *iov, struct page **pages)
+{
+       void         *addr;
+       int            nob;
+       int            i;
+
+       if (!*ksocknal_tunables.ksnd_zc_recv || pages == NULL)
+               return NULL;
+
+       LASSERT(niov <= LNET_MAX_IOV);
+
+       if (niov < 2 ||
+           niov < *ksocknal_tunables.ksnd_zc_recv_min_nfrags)
+               return NULL;
+
+       for (nob = i = 0; i < niov; i++) {
+               if ((kiov[i].kiov_offset != 0 && i > 0) ||
+                   (kiov[i].kiov_offset + kiov[i].kiov_len != PAGE_CACHE_SIZE && i < niov - 1))
+                       return NULL;
+
+               pages[i] = kiov[i].kiov_page;
+               nob += kiov[i].kiov_len;
+       }
+
+       addr = vmap(pages, niov, VM_MAP, PAGE_KERNEL);
+       if (addr == NULL)
+               return NULL;
+
+       iov->iov_base = addr + kiov[0].kiov_offset;
+       iov->iov_len = nob;
+
+       return addr;
+}
+
+int
+ksocknal_lib_recv_kiov(ksock_conn_t *conn)
+{
+#if SOCKNAL_SINGLE_FRAG_RX || !SOCKNAL_RISK_KMAP_DEADLOCK
+       struct kvec   scratch;
+       struct kvec  *scratchiov = &scratch;
+       struct page  **pages      = NULL;
+       unsigned int   niov       = 1;
+#else
+#ifdef CONFIG_HIGHMEM
+#warning "XXX risk of kmap deadlock on multiple frags..."
+#endif
+       struct kvec  *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
+       struct page  **pages      = conn->ksnc_scheduler->kss_rx_scratch_pgs;
+       unsigned int   niov       = conn->ksnc_rx_nkiov;
+#endif
+       lnet_kiov_t   *kiov = conn->ksnc_rx_kiov;
+       struct msghdr msg = {
+               .msg_flags      = 0
+       };
+       int       nob;
+       int       i;
+       int       rc;
+       void    *base;
+       void    *addr;
+       int       sum;
+       int       fragnob;
+       int n;
+
+       /* NB we can't trust socket ops to either consume our iovs
+        * or leave them alone. */
+       addr = ksocknal_lib_kiov_vmap(kiov, niov, scratchiov, pages);
+       if (addr != NULL) {
+               nob = scratchiov[0].iov_len;
+               n = 1;
+
+       } else {
+               for (nob = i = 0; i < niov; i++) {
+                       nob += scratchiov[i].iov_len = kiov[i].kiov_len;
+                       scratchiov[i].iov_base = kmap(kiov[i].kiov_page) +
+                                                kiov[i].kiov_offset;
+               }
+               n = niov;
+       }
+
+       LASSERT(nob <= conn->ksnc_rx_nob_wanted);
+
+       rc = kernel_recvmsg(conn->ksnc_sock, &msg,
+                       (struct kvec *)scratchiov, n, nob, MSG_DONTWAIT);
+
+       if (conn->ksnc_msg.ksm_csum != 0) {
+               for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
+                       LASSERT(i < niov);
+
+                       /* Dang! have to kmap again because I have nowhere to stash the
+                        * mapped address.  But by doing it while the page is still
+                        * mapped, the kernel just bumps the map count and returns me
+                        * the address it stashed. */
+                       base = kmap(kiov[i].kiov_page) + kiov[i].kiov_offset;
+                       fragnob = kiov[i].kiov_len;
+                       if (fragnob > sum)
+                               fragnob = sum;
+
+                       conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
+                                                          base, fragnob);
+
+                       kunmap(kiov[i].kiov_page);
+               }
+       }
+
+       if (addr != NULL) {
+               ksocknal_lib_kiov_vunmap(addr);
+       } else {
+               for (i = 0; i < niov; i++)
+                       kunmap(kiov[i].kiov_page);
+       }
+
+       return rc;
+}
+
+void
+ksocknal_lib_csum_tx(ksock_tx_t *tx)
+{
+       int       i;
+       __u32   csum;
+       void    *base;
+
+       LASSERT(tx->tx_iov[0].iov_base == &tx->tx_msg);
+       LASSERT(tx->tx_conn != NULL);
+       LASSERT(tx->tx_conn->ksnc_proto == &ksocknal_protocol_v2x);
+
+       tx->tx_msg.ksm_csum = 0;
+
+       csum = ksocknal_csum(~0, tx->tx_iov[0].iov_base,
+                            tx->tx_iov[0].iov_len);
+
+       if (tx->tx_kiov != NULL) {
+               for (i = 0; i < tx->tx_nkiov; i++) {
+                       base = kmap(tx->tx_kiov[i].kiov_page) +
+                              tx->tx_kiov[i].kiov_offset;
+
+                       csum = ksocknal_csum(csum, base, tx->tx_kiov[i].kiov_len);
+
+                       kunmap(tx->tx_kiov[i].kiov_page);
+               }
+       } else {
+               for (i = 1; i < tx->tx_niov; i++)
+                       csum = ksocknal_csum(csum, tx->tx_iov[i].iov_base,
+                                            tx->tx_iov[i].iov_len);
+       }
+
+       if (*ksocknal_tunables.ksnd_inject_csum_error) {
+               csum++;
+               *ksocknal_tunables.ksnd_inject_csum_error = 0;
+       }
+
+       tx->tx_msg.ksm_csum = csum;
+}
+
+int
+ksocknal_lib_get_conn_tunables(ksock_conn_t *conn, int *txmem, int *rxmem, int *nagle)
+{
+       struct socket *sock = conn->ksnc_sock;
+       int         len;
+       int         rc;
+
+       rc = ksocknal_connsock_addref(conn);
+       if (rc != 0) {
+               LASSERT(conn->ksnc_closing);
+               *txmem = *rxmem = *nagle = 0;
+               return -ESHUTDOWN;
+       }
+
+       rc = libcfs_sock_getbuf(sock, txmem, rxmem);
+       if (rc == 0) {
+               len = sizeof(*nagle);
+               rc = kernel_getsockopt(sock, SOL_TCP, TCP_NODELAY,
+                                          (char *)nagle, &len);
+       }
+
+       ksocknal_connsock_decref(conn);
+
+       if (rc == 0)
+               *nagle = !*nagle;
+       else
+               *txmem = *rxmem = *nagle = 0;
+
+       return rc;
+}
+
+int
+ksocknal_lib_setup_sock(struct socket *sock)
+{
+       int          rc;
+       int          option;
+       int          keep_idle;
+       int          keep_intvl;
+       int          keep_count;
+       int          do_keepalive;
+       struct linger   linger;
+
+       sock->sk->sk_allocation = GFP_NOFS;
+
+       /* Ensure this socket aborts active sends immediately when we close
+        * it. */
+
+       linger.l_onoff = 0;
+       linger.l_linger = 0;
+
+       rc = kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER,
+                             (char *)&linger, sizeof(linger));
+       if (rc != 0) {
+               CERROR("Can't set SO_LINGER: %d\n", rc);
+               return rc;
+       }
+
+       option = -1;
+       rc = kernel_setsockopt(sock, SOL_TCP, TCP_LINGER2,
+                                   (char *)&option, sizeof(option));
+       if (rc != 0) {
+               CERROR("Can't set SO_LINGER2: %d\n", rc);
+               return rc;
+       }
+
+       if (!*ksocknal_tunables.ksnd_nagle) {
+               option = 1;
+
+               rc = kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY,
+                                           (char *)&option, sizeof(option));
+               if (rc != 0) {
+                       CERROR("Can't disable nagle: %d\n", rc);
+                       return rc;
+               }
+       }
+
+       rc = libcfs_sock_setbuf(sock,
+                               *ksocknal_tunables.ksnd_tx_buffer_size,
+                               *ksocknal_tunables.ksnd_rx_buffer_size);
+       if (rc != 0) {
+               CERROR("Can't set buffer tx %d, rx %d buffers: %d\n",
+                       *ksocknal_tunables.ksnd_tx_buffer_size,
+                       *ksocknal_tunables.ksnd_rx_buffer_size, rc);
+               return rc;
+       }
+
+/* TCP_BACKOFF_* sockopt tunables unsupported in stock kernels */
+
+       /* snapshot tunables */
+       keep_idle  = *ksocknal_tunables.ksnd_keepalive_idle;
+       keep_count = *ksocknal_tunables.ksnd_keepalive_count;
+       keep_intvl = *ksocknal_tunables.ksnd_keepalive_intvl;
+
+       do_keepalive = (keep_idle > 0 && keep_count > 0 && keep_intvl > 0);
+
+       option = (do_keepalive ? 1 : 0);
+       rc = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
+                             (char *)&option, sizeof(option));
+       if (rc != 0) {
+               CERROR("Can't set SO_KEEPALIVE: %d\n", rc);
+               return rc;
+       }
+
+       if (!do_keepalive)
+               return 0;
+
+       rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPIDLE,
+                                   (char *)&keep_idle, sizeof(keep_idle));
+       if (rc != 0) {
+               CERROR("Can't set TCP_KEEPIDLE: %d\n", rc);
+               return rc;
+       }
+
+       rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPINTVL,
+                                   (char *)&keep_intvl, sizeof(keep_intvl));
+       if (rc != 0) {
+               CERROR("Can't set TCP_KEEPINTVL: %d\n", rc);
+               return rc;
+       }
+
+       rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPCNT,
+                                   (char *)&keep_count, sizeof(keep_count));
+       if (rc != 0) {
+               CERROR("Can't set TCP_KEEPCNT: %d\n", rc);
+               return rc;
+       }
+
+       return 0;
+}
+
+void
+ksocknal_lib_push_conn(ksock_conn_t *conn)
+{
+       struct sock    *sk;
+       struct tcp_sock *tp;
+       int          nonagle;
+       int          val = 1;
+       int          rc;
+
+       rc = ksocknal_connsock_addref(conn);
+       if (rc != 0)                        /* being shut down */
+               return;
+
+       sk = conn->ksnc_sock->sk;
+       tp = tcp_sk(sk);
+
+       lock_sock(sk);
+       nonagle = tp->nonagle;
+       tp->nonagle = 1;
+       release_sock(sk);
+
+       rc = kernel_setsockopt(conn->ksnc_sock, SOL_TCP, TCP_NODELAY,
+                                     (char *)&val, sizeof(val));
+       LASSERT(rc == 0);
+
+       lock_sock(sk);
+       tp->nonagle = nonagle;
+       release_sock(sk);
+
+       ksocknal_connsock_decref(conn);
+}
+
+extern void ksocknal_read_callback(ksock_conn_t *conn);
+extern void ksocknal_write_callback(ksock_conn_t *conn);
+/*
+ * socket call back in Linux
+ */
+static void
+ksocknal_data_ready(struct sock *sk)
+{
+       ksock_conn_t  *conn;
+
+       /* interleave correctly with closing sockets... */
+       LASSERT(!in_irq());
+       read_lock(&ksocknal_data.ksnd_global_lock);
+
+       conn = sk->sk_user_data;
+       if (conn == NULL) {          /* raced with ksocknal_terminate_conn */
+               LASSERT(sk->sk_data_ready != &ksocknal_data_ready);
+               sk->sk_data_ready(sk);
+       } else
+               ksocknal_read_callback(conn);
+
+       read_unlock(&ksocknal_data.ksnd_global_lock);
+}
+
+static void
+ksocknal_write_space(struct sock *sk)
+{
+       ksock_conn_t  *conn;
+       int         wspace;
+       int         min_wpace;
+
+       /* interleave correctly with closing sockets... */
+       LASSERT(!in_irq());
+       read_lock(&ksocknal_data.ksnd_global_lock);
+
+       conn = sk->sk_user_data;
+       wspace = SOCKNAL_WSPACE(sk);
+       min_wpace = SOCKNAL_MIN_WSPACE(sk);
+
+       CDEBUG(D_NET, "sk %p wspace %d low water %d conn %p%s%s%s\n",
+              sk, wspace, min_wpace, conn,
+              (conn == NULL) ? "" : (conn->ksnc_tx_ready ?
+                                     " ready" : " blocked"),
+              (conn == NULL) ? "" : (conn->ksnc_tx_scheduled ?
+                                     " scheduled" : " idle"),
+              (conn == NULL) ? "" : (list_empty(&conn->ksnc_tx_queue) ?
+                                     " empty" : " queued"));
+
+       if (conn == NULL) {          /* raced with ksocknal_terminate_conn */
+               LASSERT(sk->sk_write_space != &ksocknal_write_space);
+               sk->sk_write_space(sk);
+
+               read_unlock(&ksocknal_data.ksnd_global_lock);
+               return;
+       }
+
+       if (wspace >= min_wpace) {            /* got enough space */
+               ksocknal_write_callback(conn);
+
+               /* Clear SOCK_NOSPACE _after_ ksocknal_write_callback so the
+                * ENOMEM check in ksocknal_transmit is race-free (think about
+                * it). */
+
+               clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
+       }
+
+       read_unlock(&ksocknal_data.ksnd_global_lock);
+}
+
+void
+ksocknal_lib_save_callback(struct socket *sock, ksock_conn_t *conn)
+{
+       conn->ksnc_saved_data_ready = sock->sk->sk_data_ready;
+       conn->ksnc_saved_write_space = sock->sk->sk_write_space;
+}
+
+void
+ksocknal_lib_set_callback(struct socket *sock,  ksock_conn_t *conn)
+{
+       sock->sk->sk_user_data = conn;
+       sock->sk->sk_data_ready = ksocknal_data_ready;
+       sock->sk->sk_write_space = ksocknal_write_space;
+       return;
+}
+
+void
+ksocknal_lib_reset_callback(struct socket *sock, ksock_conn_t *conn)
+{
+       /* Remove conn's network callbacks.
+        * NB I _have_ to restore the callback, rather than storing a noop,
+        * since the socket could survive past this module being unloaded!! */
+       sock->sk->sk_data_ready = conn->ksnc_saved_data_ready;
+       sock->sk->sk_write_space = conn->ksnc_saved_write_space;
+
+       /* A callback could be in progress already; they hold a read lock
+        * on ksnd_global_lock (to serialise with me) and NOOP if
+        * sk_user_data is NULL. */
+       sock->sk->sk_user_data = NULL;
+
+       return ;
+}
+
+int
+ksocknal_lib_memory_pressure(ksock_conn_t *conn)
+{
+       int         rc = 0;
+       ksock_sched_t *sched;
+
+       sched = conn->ksnc_scheduler;
+       spin_lock_bh(&sched->kss_lock);
+
+       if (!test_bit(SOCK_NOSPACE, &conn->ksnc_sock->flags) &&
+           !conn->ksnc_tx_ready) {
+               /* SOCK_NOSPACE is set when the socket fills
+                * and cleared in the write_space callback
+                * (which also sets ksnc_tx_ready).  If
+                * SOCK_NOSPACE and ksnc_tx_ready are BOTH
+                * zero, I didn't fill the socket and
+                * write_space won't reschedule me, so I
+                * return -ENOMEM to get my caller to retry
+                * after a timeout */
+               rc = -ENOMEM;
+       }
+
+       spin_unlock_bh(&sched->kss_lock);
+
+       return rc;
+}