socklnd_lib-linux.c

非常经典的一个分布式系统

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 iovec    scratch;
                struct iovec   *scratchiov = &scratch;
                unsigned int    niov = 1;
#else
                struct iovec   *scratchiov = conn->ksnc_tx_scratch_iov;
                unsigned int    niov = tx->tx_niov;
#endif
                struct msghdr msg = {
                        .msg_name       = NULL,
                        .msg_namelen    = 0,
                        .msg_iov        = scratchiov,
                        .msg_iovlen     = niov,
                        .msg_control    = NULL,
                        .msg_controllen = 0,
                        .msg_flags      = MSG_DONTWAIT
                };
                mm_segment_t oldmm = get_fs();
                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;

                set_fs (KERNEL_DS);
                rc = sock_sendmsg(sock, &msg, nob);
                set_fs (oldmm);
        }
        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;

        /* NB we can't trust socket ops to either consume our iovs
         * or leave them alone. */

        if (kiov->kiov_len >= *ksocknal_tunables.ksnd_zc_min_frag &&
            tx->tx_msg.ksm_zc_req_cookie != 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 = tcp_sendpage(sock, page, offset, fragsize, msgflg);
                }
        } else {
#if SOCKNAL_SINGLE_FRAG_TX || !SOCKNAL_RISK_KMAP_DEADLOCK
                struct iovec  scratch;
                struct iovec *scratchiov = &scratch;
                unsigned int  niov = 1;
#else
#ifdef CONFIG_HIGHMEM
#warning "XXX risk of kmap deadlock on multiple frags..."
#endif
                struct iovec *scratchiov = conn->ksnc_tx_scratch_iov;
                unsigned int  niov = tx->tx_nkiov;
#endif
                struct msghdr msg = {
                        .msg_name       = NULL,
                        .msg_namelen    = 0,
                        .msg_iov        = scratchiov,
                        .msg_iovlen     = niov,
                        .msg_control    = NULL,
                        .msg_controllen = 0,
                        .msg_flags      = MSG_DONTWAIT
                };
                mm_segment_t  oldmm = get_fs();
                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;

                set_fs (KERNEL_DS);
                rc = sock_sendmsg(sock, &msg, nob);
                set_fs (oldmm);

                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;
        mm_segment_t   oldmm = get_fs();
        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. */

        set_fs(KERNEL_DS);
        sock->ops->setsockopt (sock, SOL_TCP, TCP_QUICKACK,
                               (char *)&opt, sizeof (opt));
        set_fs(oldmm);
}

int
ksocknal_lib_recv_iov (ksock_conn_t *conn)
{
#if SOCKNAL_SINGLE_FRAG_RX
        struct iovec  scratch;
        struct iovec *scratchiov = &scratch;
        unsigned int  niov = 1;
#else
        struct iovec *scratchiov = conn->ksnc_rx_scratch_iov;
        unsigned int  niov = conn->ksnc_rx_niov;
#endif
        struct iovec *iov = conn->ksnc_rx_iov;
        struct msghdr msg = {
                .msg_name       = NULL,
                .msg_namelen    = 0,
                .msg_iov        = scratchiov,
                .msg_iovlen     = niov,
                .msg_control    = NULL,
                .msg_controllen = 0,
                .msg_flags      = 0
        };
        mm_segment_t oldmm = get_fs();
        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);

        set_fs (KERNEL_DS);
        rc = sock_recvmsg (conn->ksnc_sock, &msg, nob, MSG_DONTWAIT);
        /* NB this is just a boolean..........................^ */
        set_fs (oldmm);

        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;
}

int
ksocknal_lib_recv_kiov (ksock_conn_t *conn)
{
#if SOCKNAL_SINGLE_FRAG_RX || !SOCKNAL_RISK_KMAP_DEADLOCK
        struct iovec  scratch;
        struct iovec *scratchiov = &scratch;
        unsigned int  niov = 1;
#else
#ifdef CONFIG_HIGHMEM
#warning "XXX risk of kmap deadlock on multiple frags..."
#endif
        struct iovec *scratchiov = conn->ksnc_rx_scratch_iov;
        unsigned int  niov = conn->ksnc_rx_nkiov;
#endif
        lnet_kiov_t   *kiov = conn->ksnc_rx_kiov;
        struct msghdr msg = {
                .msg_name       = NULL,
                .msg_namelen    = 0,
                .msg_iov        = scratchiov,
                .msg_iovlen     = niov,
                .msg_control    = NULL,
                .msg_controllen = 0,
                .msg_flags      = 0
        };
        mm_segment_t oldmm = get_fs();
        int          nob;
        int          i;
        int          rc;
        void        *base;
        int          sum;
        int          fragnob;

        /* NB we can't trust socket ops to either consume our iovs
         * or leave them alone. */
        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;
        }
        LASSERT (nob <= conn->ksnc_rx_nob_wanted);

        set_fs (KERNEL_DS);
        rc = sock_recvmsg (conn->ksnc_sock, &msg, nob, MSG_DONTWAIT);
        /* NB this is just a boolean.......................^ */
        set_fs (oldmm);

        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);
                }
        }
        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 == (void *)&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, (void *)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)
{
        mm_segment_t   oldmm = get_fs ();
        struct socket *sock = conn->ksnc_sock;
        int            len;
        int            rc;