X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Fklnds%2Fsocklnd%2Fsocklnd_cb.c;h=84396529992733be57f77010a441e79e6bee8ff5;hp=ac5088dc902e74122884e22a24fbe2839aa73c2d;hb=d720fbaadad9343a481863ae3b04ba580541328d;hpb=96bbc5b674d8c766e375c4266a3a7859a47b172e diff --git a/lnet/klnds/socklnd/socklnd_cb.c b/lnet/klnds/socklnd/socklnd_cb.c index ac5088d..8439652 100644 --- a/lnet/klnds/socklnd/socklnd_cb.c +++ b/lnet/klnds/socklnd/socklnd_cb.c @@ -1,14 +1,14 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: +/* + * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. * - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2011, 2017, Intel Corporation. * * Author: Zach Brown * Author: Peter J. Braam * Author: Phil Schwan * Author: Eric Barton * - * This file is part of Portals, http://www.sf.net/projects/sandiaportals/ + * This file is part of Lustre, https://wiki.whamcloud.com/ * * Portals is free software; you can redistribute it and/or * modify it under the terms of version 2 of the GNU General Public @@ -26,25 +26,25 @@ #include "socklnd.h" -ksock_tx_t * +struct ksock_tx * ksocknal_alloc_tx(int type, int size) { - ksock_tx_t *tx = NULL; + struct ksock_tx *tx = NULL; - if (type == KSOCK_MSG_NOOP) { - LASSERT (size == KSOCK_NOOP_TX_SIZE); + if (type == KSOCK_MSG_NOOP) { + LASSERT(size == KSOCK_NOOP_TX_SIZE); - /* searching for a noop tx in free list */ - cfs_spin_lock(&ksocknal_data.ksnd_tx_lock); + /* searching for a noop tx in free list */ + spin_lock(&ksocknal_data.ksnd_tx_lock); - if (!cfs_list_empty(&ksocknal_data.ksnd_idle_noop_txs)) { - tx = cfs_list_entry(ksocknal_data.ksnd_idle_noop_txs. \ - next, ksock_tx_t, tx_list); - LASSERT(tx->tx_desc_size == size); - cfs_list_del(&tx->tx_list); - } + if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) { + tx = list_entry(ksocknal_data.ksnd_idle_noop_txs.next, + struct ksock_tx, tx_list); + LASSERT(tx->tx_desc_size == size); + list_del(&tx->tx_list); + } - cfs_spin_unlock(&ksocknal_data.ksnd_tx_lock); + spin_unlock(&ksocknal_data.ksnd_tx_lock); } if (tx == NULL) @@ -53,20 +53,22 @@ ksocknal_alloc_tx(int type, int size) if (tx == NULL) return NULL; - cfs_atomic_set(&tx->tx_refcount, 1); - tx->tx_zc_capable = 0; - tx->tx_zc_checked = 0; - tx->tx_desc_size = size; + atomic_set(&tx->tx_refcount, 1); + tx->tx_zc_aborted = 0; + tx->tx_zc_capable = 0; + tx->tx_zc_checked = 0; + tx->tx_hstatus = LNET_MSG_STATUS_OK; + tx->tx_desc_size = size; - cfs_atomic_inc(&ksocknal_data.ksnd_nactive_txs); + atomic_inc(&ksocknal_data.ksnd_nactive_txs); - return tx; + return tx; } -ksock_tx_t * +struct ksock_tx * ksocknal_alloc_tx_noop(__u64 cookie, int nonblk) { - ksock_tx_t *tx; + struct ksock_tx *tx; tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE); if (tx == NULL) { @@ -82,7 +84,9 @@ ksocknal_alloc_tx_noop(__u64 cookie, int nonblk) tx->tx_niov = 1; tx->tx_nonblk = nonblk; - socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_NOOP); + tx->tx_msg.ksm_csum = 0; + tx->tx_msg.ksm_type = KSOCK_MSG_NOOP; + tx->tx_msg.ksm_zc_cookies[0] = 0; tx->tx_msg.ksm_zc_cookies[1] = cookie; return tx; @@ -90,355 +94,384 @@ ksocknal_alloc_tx_noop(__u64 cookie, int nonblk) void -ksocknal_free_tx (ksock_tx_t *tx) +ksocknal_free_tx(struct ksock_tx *tx) { - cfs_atomic_dec(&ksocknal_data.ksnd_nactive_txs); + atomic_dec(&ksocknal_data.ksnd_nactive_txs); - if (tx->tx_lnetmsg == NULL && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) { - /* it's a noop tx */ - cfs_spin_lock(&ksocknal_data.ksnd_tx_lock); + if (tx->tx_lnetmsg == NULL && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) { + /* it's a noop tx */ + spin_lock(&ksocknal_data.ksnd_tx_lock); - cfs_list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs); + list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs); - cfs_spin_unlock(&ksocknal_data.ksnd_tx_lock); - } else { - LIBCFS_FREE(tx, tx->tx_desc_size); - } + spin_unlock(&ksocknal_data.ksnd_tx_lock); + } else { + LIBCFS_FREE(tx, tx->tx_desc_size); + } } -int -ksocknal_send_iov (ksock_conn_t *conn, ksock_tx_t *tx) +static int +ksocknal_send_iov(struct ksock_conn *conn, struct ksock_tx *tx, + struct kvec *scratch_iov) { - struct iovec *iov = tx->tx_iov; - int nob; - int rc; + struct kvec *iov = tx->tx_iov; + int nob; + int rc; - LASSERT (tx->tx_niov > 0); + LASSERT(tx->tx_niov > 0); - /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */ - rc = ksocknal_lib_send_iov(conn, tx); + /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */ + rc = ksocknal_lib_send_iov(conn, tx, scratch_iov); - if (rc <= 0) /* sent nothing? */ - return (rc); + if (rc <= 0) /* sent nothing? */ + return rc; - nob = rc; - LASSERT (nob <= tx->tx_resid); - tx->tx_resid -= nob; + nob = rc; + LASSERT(nob <= tx->tx_resid); + tx->tx_resid -= nob; - /* "consume" iov */ - do { - LASSERT (tx->tx_niov > 0); + /* "consume" iov */ + do { + LASSERT(tx->tx_niov > 0); - if (nob < (int) iov->iov_len) { - iov->iov_base = (void *)((char *)iov->iov_base + nob); - iov->iov_len -= nob; - return (rc); - } + if (nob < (int) iov->iov_len) { + iov->iov_base += nob; + iov->iov_len -= nob; + return rc; + } - nob -= iov->iov_len; - tx->tx_iov = ++iov; - tx->tx_niov--; - } while (nob != 0); + nob -= iov->iov_len; + tx->tx_iov = ++iov; + tx->tx_niov--; + } while (nob != 0); - return (rc); + return rc; } -int -ksocknal_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx) +static int +ksocknal_send_kiov(struct ksock_conn *conn, struct ksock_tx *tx, + struct kvec *scratch_iov) { - lnet_kiov_t *kiov = tx->tx_kiov; - int nob; - int rc; + struct bio_vec *kiov = tx->tx_kiov; + int nob; + int rc; - LASSERT (tx->tx_niov == 0); - LASSERT (tx->tx_nkiov > 0); + LASSERT(tx->tx_niov == 0); + LASSERT(tx->tx_nkiov > 0); - /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */ - rc = ksocknal_lib_send_kiov(conn, tx); + /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */ + rc = ksocknal_lib_send_kiov(conn, tx, scratch_iov); - if (rc <= 0) /* sent nothing? */ - return (rc); + if (rc <= 0) /* sent nothing? */ + return rc; - nob = rc; - LASSERT (nob <= tx->tx_resid); - tx->tx_resid -= nob; + nob = rc; + LASSERT(nob <= tx->tx_resid); + tx->tx_resid -= nob; - /* "consume" kiov */ - do { - LASSERT(tx->tx_nkiov > 0); + /* "consume" kiov */ + do { + LASSERT(tx->tx_nkiov > 0); - if (nob < (int)kiov->kiov_len) { - kiov->kiov_offset += nob; - kiov->kiov_len -= nob; - return rc; - } + if (nob < (int)kiov->bv_len) { + kiov->bv_offset += nob; + kiov->bv_len -= nob; + return rc; + } - nob -= (int)kiov->kiov_len; - tx->tx_kiov = ++kiov; - tx->tx_nkiov--; - } while (nob != 0); + nob -= (int)kiov->bv_len; + tx->tx_kiov = ++kiov; + tx->tx_nkiov--; + } while (nob != 0); - return (rc); + return rc; } -int -ksocknal_transmit (ksock_conn_t *conn, ksock_tx_t *tx) +static int +ksocknal_transmit(struct ksock_conn *conn, struct ksock_tx *tx, + struct kvec *scratch_iov) { - int rc; - int bufnob; - - if (ksocknal_data.ksnd_stall_tx != 0) { - cfs_pause(cfs_time_seconds(ksocknal_data.ksnd_stall_tx)); - } - - LASSERT (tx->tx_resid != 0); - - rc = ksocknal_connsock_addref(conn); - if (rc != 0) { - LASSERT (conn->ksnc_closing); - return (-ESHUTDOWN); - } - - do { - if (ksocknal_data.ksnd_enomem_tx > 0) { - /* testing... */ - ksocknal_data.ksnd_enomem_tx--; - rc = -EAGAIN; - } else if (tx->tx_niov != 0) { - rc = ksocknal_send_iov (conn, tx); - } else { - rc = ksocknal_send_kiov (conn, tx); - } - - bufnob = libcfs_sock_wmem_queued(conn->ksnc_sock); - if (rc > 0) /* sent something? */ - conn->ksnc_tx_bufnob += rc; /* account it */ - - if (bufnob < conn->ksnc_tx_bufnob) { - /* allocated send buffer bytes < computed; infer - * something got ACKed */ - conn->ksnc_tx_deadline = - cfs_time_shift(*ksocknal_tunables.ksnd_timeout); - conn->ksnc_peer->ksnp_last_alive = cfs_time_current(); - conn->ksnc_tx_bufnob = bufnob; - cfs_mb(); - } - - if (rc <= 0) { /* Didn't write anything? */ - - if (rc == 0) /* some stacks return 0 instead of -EAGAIN */ - rc = -EAGAIN; - - /* Check if EAGAIN is due to memory pressure */ - if(rc == -EAGAIN && ksocknal_lib_memory_pressure(conn)) - rc = -ENOMEM; - - break; - } - - /* socket's wmem_queued now includes 'rc' bytes */ - cfs_atomic_sub (rc, &conn->ksnc_tx_nob); - rc = 0; - - } while (tx->tx_resid != 0); - - ksocknal_connsock_decref(conn); - return (rc); + int rc; + int bufnob; + + if (ksocknal_data.ksnd_stall_tx != 0) + schedule_timeout_uninterruptible( + cfs_time_seconds(ksocknal_data.ksnd_stall_tx)); + + LASSERT(tx->tx_resid != 0); + + rc = ksocknal_connsock_addref(conn); + if (rc != 0) { + LASSERT(conn->ksnc_closing); + return -ESHUTDOWN; + } + + do { + if (ksocknal_data.ksnd_enomem_tx > 0) { + /* testing... */ + ksocknal_data.ksnd_enomem_tx--; + rc = -EAGAIN; + } else if (tx->tx_niov != 0) { + rc = ksocknal_send_iov(conn, tx, scratch_iov); + } else { + rc = ksocknal_send_kiov(conn, tx, scratch_iov); + } + + bufnob = conn->ksnc_sock->sk->sk_wmem_queued; + if (rc > 0) /* sent something? */ + conn->ksnc_tx_bufnob += rc; /* account it */ + + if (bufnob < conn->ksnc_tx_bufnob) { + /* allocated send buffer bytes < computed; infer + * something got ACKed */ + conn->ksnc_tx_deadline = ktime_get_seconds() + + lnet_get_lnd_timeout(); + conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds(); + conn->ksnc_tx_bufnob = bufnob; + smp_mb(); + } + + if (rc <= 0) { /* Didn't write anything? */ + /* some stacks return 0 instead of -EAGAIN */ + if (rc == 0) + rc = -EAGAIN; + + /* Check if EAGAIN is due to memory pressure */ + if (rc == -EAGAIN && ksocknal_lib_memory_pressure(conn)) + rc = -ENOMEM; + + break; + } + + /* socket's wmem_queued now includes 'rc' bytes */ + atomic_sub (rc, &conn->ksnc_tx_nob); + rc = 0; + + } while (tx->tx_resid != 0); + + ksocknal_connsock_decref(conn); + return rc; } -int -ksocknal_recv_iov (ksock_conn_t *conn) +static int +ksocknal_recv_iov(struct ksock_conn *conn, struct kvec *scratchiov) { - struct iovec *iov = conn->ksnc_rx_iov; - int nob; - int rc; + struct kvec *iov = conn->ksnc_rx_iov; + int nob; + int rc; - LASSERT (conn->ksnc_rx_niov > 0); + LASSERT(conn->ksnc_rx_niov > 0); - /* Never touch conn->ksnc_rx_iov or change connection - * status inside ksocknal_lib_recv_iov */ - rc = ksocknal_lib_recv_iov(conn); + /* Never touch conn->ksnc_rx_iov or change connection + * status inside ksocknal_lib_recv_iov */ + rc = ksocknal_lib_recv_iov(conn, scratchiov); - if (rc <= 0) - return (rc); + if (rc <= 0) + return rc; - /* received something... */ - nob = rc; + /* received something... */ + nob = rc; - conn->ksnc_peer->ksnp_last_alive = cfs_time_current(); - conn->ksnc_rx_deadline = - cfs_time_shift(*ksocknal_tunables.ksnd_timeout); - cfs_mb(); /* order with setting rx_started */ - conn->ksnc_rx_started = 1; + conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds(); + conn->ksnc_rx_deadline = ktime_get_seconds() + + lnet_get_lnd_timeout(); + smp_mb(); /* order with setting rx_started */ + conn->ksnc_rx_started = 1; - conn->ksnc_rx_nob_wanted -= nob; - conn->ksnc_rx_nob_left -= nob; + conn->ksnc_rx_nob_wanted -= nob; + conn->ksnc_rx_nob_left -= nob; - do { - LASSERT (conn->ksnc_rx_niov > 0); + do { + LASSERT(conn->ksnc_rx_niov > 0); - if (nob < (int)iov->iov_len) { - iov->iov_len -= nob; - iov->iov_base = (void *)((char *)iov->iov_base + nob); - return (-EAGAIN); - } + if (nob < (int)iov->iov_len) { + iov->iov_len -= nob; + iov->iov_base += nob; + return -EAGAIN; + } - nob -= iov->iov_len; - conn->ksnc_rx_iov = ++iov; - conn->ksnc_rx_niov--; - } while (nob != 0); + nob -= iov->iov_len; + conn->ksnc_rx_iov = ++iov; + conn->ksnc_rx_niov--; + } while (nob != 0); - return (rc); + return rc; } -int -ksocknal_recv_kiov (ksock_conn_t *conn) +static int +ksocknal_recv_kiov(struct ksock_conn *conn, struct page **rx_scratch_pgs, + struct kvec *scratch_iov) { - lnet_kiov_t *kiov = conn->ksnc_rx_kiov; - int nob; - int rc; - LASSERT (conn->ksnc_rx_nkiov > 0); + struct bio_vec *kiov = conn->ksnc_rx_kiov; + int nob; + int rc; + LASSERT(conn->ksnc_rx_nkiov > 0); - /* Never touch conn->ksnc_rx_kiov or change connection - * status inside ksocknal_lib_recv_iov */ - rc = ksocknal_lib_recv_kiov(conn); + /* Never touch conn->ksnc_rx_kiov or change connection + * status inside ksocknal_lib_recv_iov */ + rc = ksocknal_lib_recv_kiov(conn, rx_scratch_pgs, scratch_iov); - if (rc <= 0) - return (rc); + if (rc <= 0) + return rc; - /* received something... */ - nob = rc; + /* received something... */ + nob = rc; - conn->ksnc_peer->ksnp_last_alive = cfs_time_current(); - conn->ksnc_rx_deadline = - cfs_time_shift(*ksocknal_tunables.ksnd_timeout); - cfs_mb(); /* order with setting rx_started */ - conn->ksnc_rx_started = 1; + conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds(); + conn->ksnc_rx_deadline = ktime_get_seconds() + + lnet_get_lnd_timeout(); + smp_mb(); /* order with setting rx_started */ + conn->ksnc_rx_started = 1; - conn->ksnc_rx_nob_wanted -= nob; - conn->ksnc_rx_nob_left -= nob; + conn->ksnc_rx_nob_wanted -= nob; + conn->ksnc_rx_nob_left -= nob; - do { - LASSERT (conn->ksnc_rx_nkiov > 0); + do { + LASSERT(conn->ksnc_rx_nkiov > 0); - if (nob < (int) kiov->kiov_len) { - kiov->kiov_offset += nob; - kiov->kiov_len -= nob; - return -EAGAIN; - } + if (nob < (int) kiov->bv_len) { + kiov->bv_offset += nob; + kiov->bv_len -= nob; + return -EAGAIN; + } - nob -= kiov->kiov_len; - conn->ksnc_rx_kiov = ++kiov; - conn->ksnc_rx_nkiov--; - } while (nob != 0); + nob -= kiov->bv_len; + conn->ksnc_rx_kiov = ++kiov; + conn->ksnc_rx_nkiov--; + } while (nob != 0); - return 1; + return 1; } -int -ksocknal_receive (ksock_conn_t *conn) +static int +ksocknal_receive(struct ksock_conn *conn, struct page **rx_scratch_pgs, + struct kvec *scratch_iov) { - /* Return 1 on success, 0 on EOF, < 0 on error. - * Caller checks ksnc_rx_nob_wanted to determine - * progress/completion. */ - int rc; - ENTRY; - - if (ksocknal_data.ksnd_stall_rx != 0) { - cfs_pause(cfs_time_seconds (ksocknal_data.ksnd_stall_rx)); - } - - rc = ksocknal_connsock_addref(conn); - if (rc != 0) { - LASSERT (conn->ksnc_closing); - return (-ESHUTDOWN); - } - - for (;;) { - if (conn->ksnc_rx_niov != 0) - rc = ksocknal_recv_iov (conn); - else - rc = ksocknal_recv_kiov (conn); - - if (rc <= 0) { - /* error/EOF or partial receive */ - if (rc == -EAGAIN) { - rc = 1; - } else if (rc == 0 && conn->ksnc_rx_started) { - /* EOF in the middle of a message */ - rc = -EPROTO; - } - break; - } - - /* Completed a fragment */ - - if (conn->ksnc_rx_nob_wanted == 0) { - rc = 1; - break; - } - } - - ksocknal_connsock_decref(conn); - RETURN (rc); + /* Return 1 on success, 0 on EOF, < 0 on error. + * Caller checks ksnc_rx_nob_wanted to determine + * progress/completion. */ + int rc; + ENTRY; + + if (ksocknal_data.ksnd_stall_rx != 0) + schedule_timeout_uninterruptible( + cfs_time_seconds(ksocknal_data.ksnd_stall_rx)); + + rc = ksocknal_connsock_addref(conn); + if (rc != 0) { + LASSERT(conn->ksnc_closing); + return -ESHUTDOWN; + } + + for (;;) { + if (conn->ksnc_rx_niov != 0) + rc = ksocknal_recv_iov(conn, scratch_iov); + else + rc = ksocknal_recv_kiov(conn, rx_scratch_pgs, + scratch_iov); + + if (rc <= 0) { + /* error/EOF or partial receive */ + if (rc == -EAGAIN) { + rc = 1; + } else if (rc == 0 && conn->ksnc_rx_started) { + /* EOF in the middle of a message */ + rc = -EPROTO; + } + break; + } + + /* Completed a fragment */ + + if (conn->ksnc_rx_nob_wanted == 0) { + rc = 1; + break; + } + } + + ksocknal_connsock_decref(conn); + RETURN(rc); } void -ksocknal_tx_done (lnet_ni_t *ni, ksock_tx_t *tx) +ksocknal_tx_done(struct lnet_ni *ni, struct ksock_tx *tx, int rc) { - lnet_msg_t *lnetmsg = tx->tx_lnetmsg; - int rc = (tx->tx_resid == 0) ? 0 : -EIO; + struct lnet_msg *lnetmsg = tx->tx_lnetmsg; + enum lnet_msg_hstatus hstatus = tx->tx_hstatus; ENTRY; - LASSERT(ni != NULL || tx->tx_conn != NULL); + LASSERT(ni != NULL || tx->tx_conn != NULL); - if (tx->tx_conn != NULL) - ksocknal_conn_decref(tx->tx_conn); + if (!rc && (tx->tx_resid != 0 || tx->tx_zc_aborted)) { + rc = -EIO; + if (hstatus == LNET_MSG_STATUS_OK) + hstatus = LNET_MSG_STATUS_LOCAL_ERROR; + } - if (ni == NULL && tx->tx_conn != NULL) - ni = tx->tx_conn->ksnc_peer->ksnp_ni; + if (tx->tx_conn != NULL) + ksocknal_conn_decref(tx->tx_conn); - ksocknal_free_tx (tx); - if (lnetmsg != NULL) /* KSOCK_MSG_NOOP go without lnetmsg */ - lnet_finalize (ni, lnetmsg, rc); + ksocknal_free_tx(tx); + if (lnetmsg != NULL) { /* KSOCK_MSG_NOOP go without lnetmsg */ + lnetmsg->msg_health_status = hstatus; + lnet_finalize(lnetmsg, rc); + } - EXIT; + EXIT; } void -ksocknal_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist, int error) +ksocknal_txlist_done(struct lnet_ni *ni, struct list_head *txlist, int error) { - ksock_tx_t *tx; - - while (!cfs_list_empty (txlist)) { - tx = cfs_list_entry (txlist->next, ksock_tx_t, tx_list); - - if (error && tx->tx_lnetmsg != NULL) { - CDEBUG (D_NETERROR, "Deleting packet type %d len %d %s->%s\n", - le32_to_cpu (tx->tx_lnetmsg->msg_hdr.type), - le32_to_cpu (tx->tx_lnetmsg->msg_hdr.payload_length), - libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)), - libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid))); - } else if (error) { - CDEBUG (D_NETERROR, "Deleting noop packet\n"); - } - - cfs_list_del (&tx->tx_list); - - LASSERT (cfs_atomic_read(&tx->tx_refcount) == 1); - ksocknal_tx_done (ni, tx); - } + struct ksock_tx *tx; + + while (!list_empty(txlist)) { + tx = list_entry(txlist->next, struct ksock_tx, tx_list); + + if (error && tx->tx_lnetmsg != NULL) { + CNETERR("Deleting packet type %d len %d %s->%s\n", + le32_to_cpu(tx->tx_lnetmsg->msg_hdr.type), + le32_to_cpu(tx->tx_lnetmsg->msg_hdr.payload_length), + libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)), + libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid))); + } else if (error) { + CNETERR("Deleting noop packet\n"); + } + + list_del(&tx->tx_list); + + if (tx->tx_hstatus == LNET_MSG_STATUS_OK) { + if (error == -ETIMEDOUT) + tx->tx_hstatus = + LNET_MSG_STATUS_LOCAL_TIMEOUT; + else if (error == -ENETDOWN || + error == -EHOSTUNREACH || + error == -ENETUNREACH || + error == -ECONNREFUSED || + error == -ECONNRESET) + tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_DROPPED; + /* + * for all other errors we don't want to + * retransmit + */ + else if (error) + tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR; + } + + LASSERT(atomic_read(&tx->tx_refcount) == 1); + ksocknal_tx_done(ni, tx, error); + } } static void -ksocknal_check_zc_req(ksock_tx_t *tx) +ksocknal_check_zc_req(struct ksock_tx *tx) { - ksock_conn_t *conn = tx->tx_conn; - ksock_peer_t *peer = conn->ksnc_peer; + struct ksock_conn *conn = tx->tx_conn; + struct ksock_peer_ni *peer_ni = conn->ksnc_peer; /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx * to ksnp_zc_req_list if some fragment of this message should be sent - * zero-copy. Our peer will send an ACK containing this cookie when + * zero-copy. Our peer_ni will send an ACK containing this cookie when * she has received this message to tell us we can signal completion. * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on * ksnp_zc_req_list. */ @@ -456,132 +489,156 @@ ksocknal_check_zc_req(ksock_tx_t *tx) ksocknal_tx_addref(tx); - cfs_spin_lock(&peer->ksnp_lock); + spin_lock(&peer_ni->ksnp_lock); - /* ZC_REQ is going to be pinned to the peer */ - tx->tx_deadline = - cfs_time_shift(*ksocknal_tunables.ksnd_timeout); + /* ZC_REQ is going to be pinned to the peer_ni */ + tx->tx_deadline = ktime_get_seconds() + + lnet_get_lnd_timeout(); LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0); - tx->tx_msg.ksm_zc_cookies[0] = peer->ksnp_zc_next_cookie++; + tx->tx_msg.ksm_zc_cookies[0] = peer_ni->ksnp_zc_next_cookie++; - if (peer->ksnp_zc_next_cookie == 0) - peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1; + if (peer_ni->ksnp_zc_next_cookie == 0) + peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1; - cfs_list_add_tail(&tx->tx_zc_list, &peer->ksnp_zc_req_list); + list_add_tail(&tx->tx_zc_list, &peer_ni->ksnp_zc_req_list); - cfs_spin_unlock(&peer->ksnp_lock); + spin_unlock(&peer_ni->ksnp_lock); } static void -ksocknal_uncheck_zc_req(ksock_tx_t *tx) +ksocknal_uncheck_zc_req(struct ksock_tx *tx) { - ksock_peer_t *peer = tx->tx_conn->ksnc_peer; + struct ksock_peer_ni *peer_ni = tx->tx_conn->ksnc_peer; - LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP); - LASSERT (tx->tx_zc_capable); + LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP); + LASSERT(tx->tx_zc_capable); - tx->tx_zc_checked = 0; + tx->tx_zc_checked = 0; - cfs_spin_lock(&peer->ksnp_lock); + spin_lock(&peer_ni->ksnp_lock); - if (tx->tx_msg.ksm_zc_cookies[0] == 0) { - /* Not waiting for an ACK */ - cfs_spin_unlock(&peer->ksnp_lock); - return; - } + if (tx->tx_msg.ksm_zc_cookies[0] == 0) { + /* Not waiting for an ACK */ + spin_unlock(&peer_ni->ksnp_lock); + return; + } - tx->tx_msg.ksm_zc_cookies[0] = 0; - cfs_list_del(&tx->tx_zc_list); + tx->tx_msg.ksm_zc_cookies[0] = 0; + list_del(&tx->tx_zc_list); - cfs_spin_unlock(&peer->ksnp_lock); + spin_unlock(&peer_ni->ksnp_lock); - ksocknal_tx_decref(tx); + ksocknal_tx_decref(tx); } -int -ksocknal_process_transmit (ksock_conn_t *conn, ksock_tx_t *tx) +static int +ksocknal_process_transmit(struct ksock_conn *conn, struct ksock_tx *tx, + struct kvec *scratch_iov) { - int rc; - - if (tx->tx_zc_capable && !tx->tx_zc_checked) - ksocknal_check_zc_req(tx); - - rc = ksocknal_transmit (conn, tx); - - CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc); - - if (tx->tx_resid == 0) { - /* Sent everything OK */ - LASSERT (rc == 0); - - return (0); - } - - if (rc == -EAGAIN) - return (rc); - - if (rc == -ENOMEM) { - static int counter; - - counter++; /* exponential backoff warnings */ - if ((counter & (-counter)) == counter) - CWARN("%u ENOMEM tx %p (%u allocated)\n", - counter, conn, cfs_atomic_read(&libcfs_kmemory)); - - /* Queue on ksnd_enomem_conns for retry after a timeout */ - cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock); - - /* enomem list takes over scheduler's ref... */ - LASSERT (conn->ksnc_tx_scheduled); - cfs_list_add_tail(&conn->ksnc_tx_list, - &ksocknal_data.ksnd_enomem_conns); - if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(), - SOCKNAL_ENOMEM_RETRY), - ksocknal_data.ksnd_reaper_waketime)) - cfs_waitq_signal (&ksocknal_data.ksnd_reaper_waitq); - - cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock); - return (rc); - } - - /* Actual error */ - LASSERT (rc < 0); - - if (!conn->ksnc_closing) { - switch (rc) { - case -ECONNRESET: - LCONSOLE_WARN("Host %u.%u.%u.%u reset our connection " - "while we were sending data; it may have " - "rebooted.\n", - HIPQUAD(conn->ksnc_ipaddr)); - break; - default: - LCONSOLE_WARN("There was an unexpected network error " - "while writing to %u.%u.%u.%u: %d.\n", - HIPQUAD(conn->ksnc_ipaddr), rc); - break; - } - CDEBUG(D_NET, "[%p] Error %d on write to %s" - " ip %d.%d.%d.%d:%d\n", conn, rc, - libcfs_id2str(conn->ksnc_peer->ksnp_id), - HIPQUAD(conn->ksnc_ipaddr), - conn->ksnc_port); - } - - if (tx->tx_zc_checked) - ksocknal_uncheck_zc_req(tx); - - /* it's not an error if conn is being closed */ - ksocknal_close_conn_and_siblings (conn, - (conn->ksnc_closing) ? 0 : rc); - - return (rc); + int rc; + bool error_sim = false; + + if (lnet_send_error_simulation(tx->tx_lnetmsg, &tx->tx_hstatus)) { + error_sim = true; + rc = -EINVAL; + goto simulate_error; + } + + if (tx->tx_zc_capable && !tx->tx_zc_checked) + ksocknal_check_zc_req(tx); + + rc = ksocknal_transmit(conn, tx, scratch_iov); + + CDEBUG(D_NET, "send(%d) %d\n", tx->tx_resid, rc); + + if (tx->tx_resid == 0) { + /* Sent everything OK */ + LASSERT(rc == 0); + + return 0; + } + + if (rc == -EAGAIN) + return rc; + + if (rc == -ENOMEM) { + static int counter; + + counter++; /* exponential backoff warnings */ + if ((counter & (-counter)) == counter) + CWARN("%u ENOMEM tx %p (%u allocated)\n", + counter, conn, atomic_read(&libcfs_kmemory)); + + /* Queue on ksnd_enomem_conns for retry after a timeout */ + spin_lock_bh(&ksocknal_data.ksnd_reaper_lock); + + /* enomem list takes over scheduler's ref... */ + LASSERT(conn->ksnc_tx_scheduled); + list_add_tail(&conn->ksnc_tx_list, + &ksocknal_data.ksnd_enomem_conns); + if (ktime_get_seconds() + SOCKNAL_ENOMEM_RETRY < + ksocknal_data.ksnd_reaper_waketime) + wake_up(&ksocknal_data.ksnd_reaper_waitq); + + spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock); + + /* + * set the health status of the message which determines + * whether we should retry the transmit + */ + tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR; + return (rc); + } + +simulate_error: + + /* Actual error */ + LASSERT(rc < 0); + + if (!error_sim) { + /* + * set the health status of the message which determines + * whether we should retry the transmit + */ + if (rc == -ETIMEDOUT) + tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_TIMEOUT; + else + tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR; + } + + if (!conn->ksnc_closing) { + switch (rc) { + case -ECONNRESET: + LCONSOLE_WARN("Host %pI4h reset our connection " + "while we were sending data; it may have " + "rebooted.\n", + &conn->ksnc_ipaddr); + break; + default: + LCONSOLE_WARN("There was an unexpected network error " + "while writing to %pI4h: %d.\n", + &conn->ksnc_ipaddr, rc); + break; + } + CDEBUG(D_NET, "[%p] Error %d on write to %s ip %pI4h:%d\n", + conn, rc, libcfs_id2str(conn->ksnc_peer->ksnp_id), + &conn->ksnc_ipaddr, conn->ksnc_port); + } + + if (tx->tx_zc_checked) + ksocknal_uncheck_zc_req(tx); + + /* it's not an error if conn is being closed */ + ksocknal_close_conn_and_siblings(conn, + (conn->ksnc_closing) ? 0 : rc); + + return rc; } -void -ksocknal_launch_connection_locked (ksock_route_t *route) +static void +ksocknal_launch_connection_locked(struct ksock_route *route) { /* called holding write lock on ksnd_global_lock */ @@ -593,24 +650,24 @@ ksocknal_launch_connection_locked (ksock_route_t *route) route->ksnr_scheduled = 1; /* scheduling conn for connd */ ksocknal_route_addref(route); /* extra ref for connd */ - cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock); + spin_lock_bh(&ksocknal_data.ksnd_connd_lock); - cfs_list_add_tail (&route->ksnr_connd_list, - &ksocknal_data.ksnd_connd_routes); - cfs_waitq_signal (&ksocknal_data.ksnd_connd_waitq); + list_add_tail(&route->ksnr_connd_list, + &ksocknal_data.ksnd_connd_routes); + wake_up(&ksocknal_data.ksnd_connd_waitq); - cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock); + spin_unlock_bh(&ksocknal_data.ksnd_connd_lock); } void -ksocknal_launch_all_connections_locked (ksock_peer_t *peer) +ksocknal_launch_all_connections_locked(struct ksock_peer_ni *peer_ni) { - ksock_route_t *route; + struct ksock_route *route; /* called holding write lock on ksnd_global_lock */ for (;;) { /* launch any/all connections that need it */ - route = ksocknal_find_connectable_route_locked(peer); + route = ksocknal_find_connectable_route_locked(peer_ni); if (route == NULL) return; @@ -618,21 +675,22 @@ ksocknal_launch_all_connections_locked (ksock_peer_t *peer) } } -ksock_conn_t * -ksocknal_find_conn_locked(ksock_peer_t *peer, ksock_tx_t *tx, int nonblk) +struct ksock_conn * +ksocknal_find_conn_locked(struct ksock_peer_ni *peer_ni, struct ksock_tx *tx, int nonblk) { - cfs_list_t *tmp; - ksock_conn_t *conn; - ksock_conn_t *typed = NULL; - ksock_conn_t *fallback = NULL; - int tnob = 0; - int fnob = 0; - - cfs_list_for_each (tmp, &peer->ksnp_conns) { - ksock_conn_t *c = cfs_list_entry(tmp, ksock_conn_t, ksnc_list); - int nob = cfs_atomic_read(&c->ksnc_tx_nob) + - libcfs_sock_wmem_queued(c->ksnc_sock); - int rc; + struct list_head *tmp; + struct ksock_conn *conn; + struct ksock_conn *typed = NULL; + struct ksock_conn *fallback = NULL; + int tnob = 0; + int fnob = 0; + + list_for_each(tmp, &peer_ni->ksnp_conns) { + struct ksock_conn *c = list_entry(tmp, struct ksock_conn, + ksnc_list); + int nob = atomic_read(&c->ksnc_tx_nob) + + c->ksnc_sock->sk->sk_wmem_queued; + int rc; LASSERT (!c->ksnc_closing); LASSERT (c->ksnc_proto != NULL && @@ -649,7 +707,7 @@ ksocknal_find_conn_locked(ksock_peer_t *peer, ksock_tx_t *tx, int nonblk) case SOCKNAL_MATCH_YES: /* typed connection */ if (typed == NULL || tnob > nob || (tnob == nob && *ksocknal_tunables.ksnd_round_robin && - cfs_time_after(typed->ksnc_tx_last_post, c->ksnc_tx_last_post))) { + typed->ksnc_tx_last_post > c->ksnc_tx_last_post)) { typed = c; tnob = nob; } @@ -658,7 +716,7 @@ ksocknal_find_conn_locked(ksock_peer_t *peer, ksock_tx_t *tx, int nonblk) case SOCKNAL_MATCH_MAY: /* fallback connection */ if (fallback == NULL || fnob > nob || (fnob == nob && *ksocknal_tunables.ksnd_round_robin && - cfs_time_after(fallback->ksnc_tx_last_post, c->ksnc_tx_last_post))) { + fallback->ksnc_tx_last_post > c->ksnc_tx_last_post)) { fallback = c; fnob = nob; } @@ -670,28 +728,28 @@ ksocknal_find_conn_locked(ksock_peer_t *peer, ksock_tx_t *tx, int nonblk) conn = (typed != NULL) ? typed : fallback; if (conn != NULL) - conn->ksnc_tx_last_post = cfs_time_current(); + conn->ksnc_tx_last_post = ktime_get_seconds(); return conn; } void -ksocknal_tx_prep(ksock_conn_t *conn, ksock_tx_t *tx) +ksocknal_tx_prep(struct ksock_conn *conn, struct ksock_tx *tx) { conn->ksnc_proto->pro_pack(tx); - cfs_atomic_add (tx->tx_nob, &conn->ksnc_tx_nob); + atomic_add (tx->tx_nob, &conn->ksnc_tx_nob); ksocknal_conn_addref(conn); /* +1 ref for tx */ tx->tx_conn = conn; } void -ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn) +ksocknal_queue_tx_locked(struct ksock_tx *tx, struct ksock_conn *conn) { - ksock_sched_t *sched = conn->ksnc_scheduler; - ksock_msg_t *msg = &tx->tx_msg; - ksock_tx_t *ztx = NULL; - int bufnob = 0; + struct ksock_sched *sched = conn->ksnc_scheduler; + struct ksock_msg *msg = &tx->tx_msg; + struct ksock_tx *ztx = NULL; + int bufnob = 0; /* called holding global lock (read or irq-write) and caller may * not have dropped this lock between finding conn and calling me, @@ -699,16 +757,15 @@ ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn) * ksnc_sock... */ LASSERT(!conn->ksnc_closing); - CDEBUG (D_NET, "Sending to %s ip %d.%d.%d.%d:%d\n", - libcfs_id2str(conn->ksnc_peer->ksnp_id), - HIPQUAD(conn->ksnc_ipaddr), - conn->ksnc_port); + CDEBUG(D_NET, "Sending to %s ip %pI4h:%d\n", + libcfs_id2str(conn->ksnc_peer->ksnp_id), + &conn->ksnc_ipaddr, conn->ksnc_port); ksocknal_tx_prep(conn, tx); /* Ensure the frags we've been given EXACTLY match the number of * bytes we want to send. Many TCP/IP stacks disregard any total - * size parameters passed to them and just look at the frags. + * size parameters passed to them and just look at the frags. * * We always expect at least 1 mapped fragment containing the * complete ksocknal message header. */ @@ -723,26 +780,22 @@ ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn) KSOCK_MSG_NOOP, tx->tx_nob, tx->tx_niov, tx->tx_nkiov); - /* - * FIXME: SOCK_WMEM_QUEUED and SOCK_ERROR could block in __DARWIN8__ - * but they're used inside spinlocks a lot. - */ - bufnob = libcfs_sock_wmem_queued(conn->ksnc_sock); - cfs_spin_lock_bh (&sched->kss_lock); - - if (cfs_list_empty(&conn->ksnc_tx_queue) && bufnob == 0) { - /* First packet starts the timeout */ - conn->ksnc_tx_deadline = - cfs_time_shift(*ksocknal_tunables.ksnd_timeout); - if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */ - conn->ksnc_peer->ksnp_last_alive = cfs_time_current(); - conn->ksnc_tx_bufnob = 0; - cfs_mb(); /* order with adding to tx_queue */ - } - - if (msg->ksm_type == KSOCK_MSG_NOOP) { - /* The packet is noop ZC ACK, try to piggyback the ack_cookie - * on a normal packet so I don't need to send it */ + bufnob = conn->ksnc_sock->sk->sk_wmem_queued; + spin_lock_bh(&sched->kss_lock); + + if (list_empty(&conn->ksnc_tx_queue) && bufnob == 0) { + /* First packet starts the timeout */ + conn->ksnc_tx_deadline = ktime_get_seconds() + + lnet_get_lnd_timeout(); + if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */ + conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds(); + conn->ksnc_tx_bufnob = 0; + smp_mb(); /* order with adding to tx_queue */ + } + + if (msg->ksm_type == KSOCK_MSG_NOOP) { + /* The packet is noop ZC ACK, try to piggyback the ack_cookie + * on a normal packet so I don't need to send it */ LASSERT (msg->ksm_zc_cookies[1] != 0); LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL); @@ -760,32 +813,33 @@ ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn) } if (ztx != NULL) { - cfs_atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob); - cfs_list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs); + atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob); + list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs); } - if (conn->ksnc_tx_ready && /* able to send */ - !conn->ksnc_tx_scheduled) { /* not scheduled to send */ - /* +1 ref for scheduler */ - ksocknal_conn_addref(conn); - cfs_list_add_tail (&conn->ksnc_tx_list, - &sched->kss_tx_conns); - conn->ksnc_tx_scheduled = 1; - cfs_waitq_signal (&sched->kss_waitq); - } - - cfs_spin_unlock_bh (&sched->kss_lock); + if (conn->ksnc_tx_ready && /* able to send */ + !conn->ksnc_tx_scheduled) { /* not scheduled to send */ + /* +1 ref for scheduler */ + ksocknal_conn_addref(conn); + list_add_tail(&conn->ksnc_tx_list, + &sched->kss_tx_conns); + conn->ksnc_tx_scheduled = 1; + wake_up(&sched->kss_waitq); + } + + spin_unlock_bh(&sched->kss_lock); } -ksock_route_t * -ksocknal_find_connectable_route_locked (ksock_peer_t *peer) +struct ksock_route * +ksocknal_find_connectable_route_locked(struct ksock_peer_ni *peer_ni) { - cfs_list_t *tmp; - ksock_route_t *route; + time64_t now = ktime_get_seconds(); + struct list_head *tmp; + struct ksock_route *route; - cfs_list_for_each (tmp, &peer->ksnp_routes) { - route = cfs_list_entry (tmp, ksock_route_t, ksnr_list); + list_for_each(tmp, &peer_ni->ksnp_routes) { + route = list_entry(tmp, struct ksock_route, ksnr_list); LASSERT (!route->ksnr_connecting || route->ksnr_scheduled); @@ -796,11 +850,17 @@ ksocknal_find_connectable_route_locked (ksock_peer_t *peer) if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0) continue; - /* too soon to retry this guy? */ if (!(route->ksnr_retry_interval == 0 || /* first attempt */ - cfs_time_aftereq (cfs_time_current(), - route->ksnr_timeout))) + now >= route->ksnr_timeout)) { + CDEBUG(D_NET, + "Too soon to retry route %pI4h " + "(cnted %d, interval %lld, %lld secs later)\n", + &route->ksnr_ipaddr, + route->ksnr_connected, + route->ksnr_retry_interval, + route->ksnr_timeout - now); continue; + } return (route); } @@ -808,14 +868,14 @@ ksocknal_find_connectable_route_locked (ksock_peer_t *peer) return (NULL); } -ksock_route_t * -ksocknal_find_connecting_route_locked (ksock_peer_t *peer) +struct ksock_route * +ksocknal_find_connecting_route_locked(struct ksock_peer_ni *peer_ni) { - cfs_list_t *tmp; - ksock_route_t *route; + struct list_head *tmp; + struct ksock_route *route; - cfs_list_for_each (tmp, &peer->ksnp_routes) { - route = cfs_list_entry (tmp, ksock_route_t, ksnr_list); + list_for_each(tmp, &peer_ni->ksnp_routes) { + route = list_entry(tmp, struct ksock_route, ksnr_list); LASSERT (!route->ksnr_connecting || route->ksnr_scheduled); @@ -827,45 +887,46 @@ ksocknal_find_connecting_route_locked (ksock_peer_t *peer) } int -ksocknal_launch_packet (lnet_ni_t *ni, ksock_tx_t *tx, lnet_process_id_t id) +ksocknal_launch_packet(struct lnet_ni *ni, struct ksock_tx *tx, + struct lnet_process_id id) { - ksock_peer_t *peer; - ksock_conn_t *conn; - cfs_rwlock_t *g_lock; - int retry; - int rc; + struct ksock_peer_ni *peer_ni; + struct ksock_conn *conn; + rwlock_t *g_lock; + int retry; + int rc; LASSERT (tx->tx_conn == NULL); g_lock = &ksocknal_data.ksnd_global_lock; for (retry = 0;; retry = 1) { - cfs_read_lock (g_lock); - peer = ksocknal_find_peer_locked(ni, id); - if (peer != NULL) { - if (ksocknal_find_connectable_route_locked(peer) == NULL) { - conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk); + read_lock(g_lock); + peer_ni = ksocknal_find_peer_locked(ni, id); + if (peer_ni != NULL) { + if (ksocknal_find_connectable_route_locked(peer_ni) == NULL) { + conn = ksocknal_find_conn_locked(peer_ni, tx, tx->tx_nonblk); if (conn != NULL) { /* I've got no routes that need to be * connecting and I do have an actual * connection... */ ksocknal_queue_tx_locked (tx, conn); - cfs_read_unlock (g_lock); + read_unlock(g_lock); return (0); } } } /* I'll need a write lock... */ - cfs_read_unlock (g_lock); + read_unlock(g_lock); - cfs_write_lock_bh (g_lock); + write_lock_bh(g_lock); - peer = ksocknal_find_peer_locked(ni, id); - if (peer != NULL) + peer_ni = ksocknal_find_peer_locked(ni, id); + if (peer_ni != NULL) break; - cfs_write_unlock_bh (g_lock); + write_unlock_bh(g_lock); if ((id.pid & LNET_PID_USERFLAG) != 0) { CERROR("Refusing to create a connection to " @@ -874,7 +935,7 @@ ksocknal_launch_packet (lnet_ni_t *ni, ksock_tx_t *tx, lnet_process_id_t id) } if (retry) { - CERROR("Can't find peer %s\n", libcfs_id2str(id)); + CERROR("Can't find peer_ni %s\n", libcfs_id2str(id)); return -EHOSTUNREACH; } @@ -882,54 +943,56 @@ ksocknal_launch_packet (lnet_ni_t *ni, ksock_tx_t *tx, lnet_process_id_t id) LNET_NIDADDR(id.nid), lnet_acceptor_port()); if (rc != 0) { - CERROR("Can't add peer %s: %d\n", + CERROR("Can't add peer_ni %s: %d\n", libcfs_id2str(id), rc); return rc; } } - ksocknal_launch_all_connections_locked(peer); + ksocknal_launch_all_connections_locked(peer_ni); - conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk); + conn = ksocknal_find_conn_locked(peer_ni, tx, tx->tx_nonblk); if (conn != NULL) { /* Connection exists; queue message on it */ ksocknal_queue_tx_locked (tx, conn); - cfs_write_unlock_bh (g_lock); + write_unlock_bh(g_lock); return (0); } - if (peer->ksnp_accepting > 0 || - ksocknal_find_connecting_route_locked (peer) != NULL) { - /* the message is going to be pinned to the peer */ - tx->tx_deadline = - cfs_time_shift(*ksocknal_tunables.ksnd_timeout); + if (peer_ni->ksnp_accepting > 0 || + ksocknal_find_connecting_route_locked (peer_ni) != NULL) { + /* the message is going to be pinned to the peer_ni */ + tx->tx_deadline = ktime_get_seconds() + + lnet_get_lnd_timeout(); /* Queue the message until a connection is established */ - cfs_list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue); - cfs_write_unlock_bh (g_lock); + list_add_tail(&tx->tx_list, &peer_ni->ksnp_tx_queue); + write_unlock_bh(g_lock); return 0; } - cfs_write_unlock_bh (g_lock); + write_unlock_bh(g_lock); /* NB Routes may be ignored if connections to them failed recently */ - CDEBUG(D_NETERROR, "No usable routes to %s\n", libcfs_id2str(id)); + CNETERR("No usable routes to %s\n", libcfs_id2str(id)); + tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_ERROR; return (-EHOSTUNREACH); } int -ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) +ksocknal_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg) { - int type = lntmsg->msg_type; - lnet_process_id_t target = lntmsg->msg_target; - unsigned int payload_niov = lntmsg->msg_niov; - struct iovec *payload_iov = lntmsg->msg_iov; - lnet_kiov_t *payload_kiov = lntmsg->msg_kiov; - unsigned int payload_offset = lntmsg->msg_offset; - unsigned int payload_nob = lntmsg->msg_len; - ksock_tx_t *tx; - int desc_size; - int rc; + int mpflag = 1; + int type = lntmsg->msg_type; + struct lnet_process_id target = lntmsg->msg_target; + unsigned int payload_niov = lntmsg->msg_niov; + struct kvec *payload_iov = lntmsg->msg_iov; + struct bio_vec *payload_kiov = lntmsg->msg_kiov; + unsigned int payload_offset = lntmsg->msg_offset; + unsigned int payload_nob = lntmsg->msg_len; + struct ksock_tx *tx; + int desc_size; + int rc; /* NB 'private' is different depending on what we're sending. * Just ignore it... */ @@ -937,23 +1000,27 @@ ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n", payload_nob, payload_niov, libcfs_id2str(target)); - LASSERT (payload_nob == 0 || payload_niov > 0); - LASSERT (payload_niov <= LNET_MAX_IOV); - /* payload is either all vaddrs or all pages */ - LASSERT (!(payload_kiov != NULL && payload_iov != NULL)); - LASSERT (!cfs_in_interrupt ()); - - if (payload_iov != NULL) - desc_size = offsetof(ksock_tx_t, - tx_frags.virt.iov[1 + payload_niov]); - else - desc_size = offsetof(ksock_tx_t, - tx_frags.paged.kiov[payload_niov]); - + LASSERT (payload_nob == 0 || payload_niov > 0); + LASSERT (payload_niov <= LNET_MAX_IOV); + /* payload is either all vaddrs or all pages */ + LASSERT (!(payload_kiov != NULL && payload_iov != NULL)); + LASSERT (!in_interrupt ()); + + if (payload_iov != NULL) + desc_size = offsetof(struct ksock_tx, + tx_frags.virt.iov[1 + payload_niov]); + else + desc_size = offsetof(struct ksock_tx, + tx_frags.paged.kiov[payload_niov]); + + if (lntmsg->msg_vmflush) + mpflag = cfs_memory_pressure_get_and_set(); tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size); if (tx == NULL) { CERROR("Can't allocate tx desc type %d size %d\n", type, desc_size); + if (lntmsg->msg_vmflush) + cfs_memory_pressure_restore(mpflag); return (-ENOMEM); } @@ -980,47 +1047,54 @@ ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) tx->tx_zc_capable = 1; } - socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_LNET); + tx->tx_msg.ksm_csum = 0; + tx->tx_msg.ksm_type = KSOCK_MSG_LNET; + tx->tx_msg.ksm_zc_cookies[0] = 0; + tx->tx_msg.ksm_zc_cookies[1] = 0; /* The first fragment will be set later in pro_pack */ rc = ksocknal_launch_packet(ni, tx, target); + if (!mpflag) + cfs_memory_pressure_restore(mpflag); + if (rc == 0) return (0); + lntmsg->msg_health_status = tx->tx_hstatus; ksocknal_free_tx(tx); return (-EIO); } int -ksocknal_thread_start (int (*fn)(void *arg), void *arg) +ksocknal_thread_start(int (*fn)(void *arg), void *arg, char *name) { - long pid = cfs_kernel_thread (fn, arg, 0); + struct task_struct *task = kthread_run(fn, arg, name); - if (pid < 0) - return ((int)pid); + if (IS_ERR(task)) + return PTR_ERR(task); - cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock); - ksocknal_data.ksnd_nthreads++; - cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock); - return (0); + write_lock_bh(&ksocknal_data.ksnd_global_lock); + ksocknal_data.ksnd_nthreads++; + write_unlock_bh(&ksocknal_data.ksnd_global_lock); + return 0; } void ksocknal_thread_fini (void) { - cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock); - ksocknal_data.ksnd_nthreads--; - cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock); + write_lock_bh(&ksocknal_data.ksnd_global_lock); + if (--ksocknal_data.ksnd_nthreads == 0) + wake_up_var(&ksocknal_data.ksnd_nthreads); + write_unlock_bh(&ksocknal_data.ksnd_global_lock); } int -ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip) +ksocknal_new_packet(struct ksock_conn *conn, int nob_to_skip) { static char ksocknal_slop_buffer[4096]; - - int nob; - unsigned int niov; - int skipped; + int nob; + unsigned int niov; + int skipped; LASSERT(conn->ksnc_proto != NULL); @@ -1029,31 +1103,31 @@ ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip) ksocknal_lib_eager_ack(conn); } - if (nob_to_skip == 0) { /* right at next packet boundary now */ - conn->ksnc_rx_started = 0; - cfs_mb(); /* racing with timeout thread */ + if (nob_to_skip == 0) { /* right at next packet boundary now */ + conn->ksnc_rx_started = 0; + smp_mb(); /* racing with timeout thread */ - switch (conn->ksnc_proto->pro_version) { - case KSOCK_PROTO_V2: - case KSOCK_PROTO_V3: + switch (conn->ksnc_proto->pro_version) { + case KSOCK_PROTO_V2: + case KSOCK_PROTO_V3: conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER; - conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space; + conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space; conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg; - conn->ksnc_rx_nob_wanted = offsetof(ksock_msg_t, ksm_u); - conn->ksnc_rx_nob_left = offsetof(ksock_msg_t, ksm_u); - conn->ksnc_rx_iov[0].iov_len = offsetof(ksock_msg_t, ksm_u); + conn->ksnc_rx_nob_wanted = offsetof(struct ksock_msg, ksm_u); + conn->ksnc_rx_nob_left = offsetof(struct ksock_msg, ksm_u); + conn->ksnc_rx_iov[0].iov_len = offsetof(struct ksock_msg, ksm_u); break; - case KSOCK_PROTO_V1: - /* Receiving bare lnet_hdr_t */ - conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER; - conn->ksnc_rx_nob_wanted = sizeof(lnet_hdr_t); - conn->ksnc_rx_nob_left = sizeof(lnet_hdr_t); + case KSOCK_PROTO_V1: + /* Receiving bare struct lnet_hdr */ + conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER; + conn->ksnc_rx_nob_wanted = sizeof(struct lnet_hdr); + conn->ksnc_rx_nob_left = sizeof(struct lnet_hdr); - conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space; - conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg; - conn->ksnc_rx_iov[0].iov_len = sizeof (lnet_hdr_t); + conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space; + conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg; + conn->ksnc_rx_iov[0].iov_len = sizeof(struct lnet_hdr); break; default: @@ -1070,23 +1144,23 @@ ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip) /* Set up to skip as much as possible now. If there's more left * (ran out of iov entries) we'll get called again */ - conn->ksnc_rx_state = SOCKNAL_RX_SLOP; - conn->ksnc_rx_nob_left = nob_to_skip; - conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space; - skipped = 0; - niov = 0; + conn->ksnc_rx_state = SOCKNAL_RX_SLOP; + conn->ksnc_rx_nob_left = nob_to_skip; + conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space; + skipped = 0; + niov = 0; - do { - nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer)); + do { + nob = min_t(int, nob_to_skip, sizeof(ksocknal_slop_buffer)); - conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer; - conn->ksnc_rx_iov[niov].iov_len = nob; - niov++; - skipped += nob; - nob_to_skip -=nob; + conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer; + conn->ksnc_rx_iov[niov].iov_len = nob; + niov++; + skipped += nob; + nob_to_skip -= nob; - } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */ - niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec)); + } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */ + niov < sizeof(conn->ksnc_rx_iov_space) / sizeof(struct kvec)); conn->ksnc_rx_niov = niov; conn->ksnc_rx_kiov = NULL; @@ -1095,41 +1169,46 @@ ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip) return (0); } -int -ksocknal_process_receive (ksock_conn_t *conn) +static int +ksocknal_process_receive(struct ksock_conn *conn, + struct page **rx_scratch_pgs, + struct kvec *scratch_iov) { - lnet_hdr_t *lhdr; - lnet_process_id_t *id; - int rc; - - LASSERT (cfs_atomic_read(&conn->ksnc_conn_refcount) > 0); - - /* NB: sched lock NOT held */ - /* SOCKNAL_RX_LNET_HEADER is here for backward compatability */ - LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER || - conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD || - conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER || - conn->ksnc_rx_state == SOCKNAL_RX_SLOP); + struct lnet_hdr *lhdr; + struct lnet_process_id *id; + int rc; + + LASSERT (atomic_read(&conn->ksnc_conn_refcount) > 0); + + /* NB: sched lock NOT held */ + /* SOCKNAL_RX_LNET_HEADER is here for backward compatibility */ + LASSERT(conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER || + conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD || + conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER || + conn->ksnc_rx_state == SOCKNAL_RX_SLOP); again: - if (conn->ksnc_rx_nob_wanted != 0) { - rc = ksocknal_receive(conn); + if (conn->ksnc_rx_nob_wanted != 0) { + rc = ksocknal_receive(conn, rx_scratch_pgs, + scratch_iov); - if (rc <= 0) { - LASSERT (rc != -EAGAIN); + if (rc <= 0) { + struct lnet_process_id ksnp_id; - if (rc == 0) - CDEBUG (D_NET, "[%p] EOF from %s" - " ip %d.%d.%d.%d:%d\n", conn, - libcfs_id2str(conn->ksnc_peer->ksnp_id), - HIPQUAD(conn->ksnc_ipaddr), - conn->ksnc_port); - else if (!conn->ksnc_closing) - CERROR ("[%p] Error %d on read from %s" - " ip %d.%d.%d.%d:%d\n", - conn, rc, - libcfs_id2str(conn->ksnc_peer->ksnp_id), - HIPQUAD(conn->ksnc_ipaddr), + ksnp_id = conn->ksnc_peer->ksnp_id; + + LASSERT(rc != -EAGAIN); + if (rc == 0) + CDEBUG(D_NET, "[%p] EOF from %s " + "ip %pI4h:%d\n", conn, + libcfs_id2str(ksnp_id), + &conn->ksnc_ipaddr, conn->ksnc_port); + else if (!conn->ksnc_closing) + CERROR("[%p] Error %d on read from %s " + "ip %pI4h:%d\n", conn, rc, + libcfs_id2str(ksnp_id), + &conn->ksnc_ipaddr, + conn->ksnc_port); /* it's not an error if conn is being closed */ ksocknal_close_conn_and_siblings (conn, @@ -1185,7 +1264,7 @@ ksocknal_process_receive (ksock_conn_t *conn) conn->ksnc_msg.ksm_zc_cookies[1]); if (rc != 0) { - CERROR("%s: Unknown ZC-ACK cookie: "LPU64", "LPU64"\n", + CERROR("%s: Unknown ZC-ACK cookie: %llu, %llu\n", libcfs_id2str(conn->ksnc_peer->ksnp_id), cookie, conn->ksnc_msg.ksm_zc_cookies[1]); ksocknal_new_packet(conn, 0); @@ -1200,12 +1279,12 @@ ksocknal_process_receive (ksock_conn_t *conn) } conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER; - conn->ksnc_rx_nob_wanted = sizeof(ksock_lnet_msg_t); - conn->ksnc_rx_nob_left = sizeof(ksock_lnet_msg_t); + conn->ksnc_rx_nob_wanted = sizeof(struct ksock_lnet_msg); + conn->ksnc_rx_nob_left = sizeof(struct ksock_lnet_msg); - conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space; + conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space; conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg; - conn->ksnc_rx_iov[0].iov_len = sizeof(ksock_lnet_msg_t); + conn->ksnc_rx_iov[0].iov_len = sizeof(struct ksock_lnet_msg); conn->ksnc_rx_niov = 1; conn->ksnc_rx_kiov = NULL; @@ -1218,7 +1297,7 @@ ksocknal_process_receive (ksock_conn_t *conn) conn->ksnc_proto->pro_unpack(&conn->ksnc_msg); if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) { - /* Userspace peer */ + /* Userspace peer_ni */ lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr; id = &conn->ksnc_peer->ksnp_id; @@ -1276,7 +1355,10 @@ ksocknal_process_receive (ksock_conn_t *conn) le64_to_cpu(lhdr->src_nid) != id->nid); } - lnet_finalize(conn->ksnc_peer->ksnp_ni, conn->ksnc_cookie, rc); + if (rc && conn->ksnc_lnet_msg) + conn->ksnc_lnet_msg->msg_health_status = + LNET_MSG_STATUS_REMOTE_ERROR; + lnet_finalize(conn->ksnc_lnet_msg, rc); if (rc != 0) { ksocknal_new_packet(conn, 0); @@ -1301,19 +1383,20 @@ ksocknal_process_receive (ksock_conn_t *conn) } int -ksocknal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed, - unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov, - unsigned int offset, unsigned int mlen, unsigned int rlen) +ksocknal_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg, + int delayed, unsigned int niov, struct kvec *iov, + struct bio_vec *kiov, unsigned int offset, unsigned int mlen, + unsigned int rlen) { - ksock_conn_t *conn = (ksock_conn_t *)private; - ksock_sched_t *sched = conn->ksnc_scheduler; + struct ksock_conn *conn = private; + struct ksock_sched *sched = conn->ksnc_scheduler; LASSERT (mlen <= rlen); LASSERT (niov <= LNET_MAX_IOV); - conn->ksnc_cookie = msg; - conn->ksnc_rx_nob_wanted = mlen; - conn->ksnc_rx_nob_left = rlen; + conn->ksnc_lnet_msg = msg; + conn->ksnc_rx_nob_wanted = mlen; + conn->ksnc_rx_nob_left = rlen; if (mlen == 0 || iov != NULL) { conn->ksnc_rx_nkiov = 0; @@ -1337,14 +1420,14 @@ ksocknal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed, LASSERT (conn->ksnc_rx_scheduled); - cfs_spin_lock_bh (&sched->kss_lock); + spin_lock_bh(&sched->kss_lock); - switch (conn->ksnc_rx_state) { - case SOCKNAL_RX_PARSE_WAIT: - cfs_list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns); - cfs_waitq_signal (&sched->kss_waitq); - LASSERT (conn->ksnc_rx_ready); - break; + switch (conn->ksnc_rx_state) { + case SOCKNAL_RX_PARSE_WAIT: + list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns); + wake_up(&sched->kss_waitq); + LASSERT(conn->ksnc_rx_ready); + break; case SOCKNAL_RX_PARSE: /* scheduler hasn't noticed I'm parsing yet */ @@ -1353,251 +1436,265 @@ ksocknal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed, conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD; - cfs_spin_unlock_bh (&sched->kss_lock); - ksocknal_conn_decref(conn); - return (0); + spin_unlock_bh(&sched->kss_lock); + ksocknal_conn_decref(conn); + return 0; } static inline int -ksocknal_sched_cansleep(ksock_sched_t *sched) +ksocknal_sched_cansleep(struct ksock_sched *sched) { - int rc; + int rc; - cfs_spin_lock_bh (&sched->kss_lock); + spin_lock_bh(&sched->kss_lock); - rc = (!ksocknal_data.ksnd_shuttingdown && - cfs_list_empty(&sched->kss_rx_conns) && - cfs_list_empty(&sched->kss_tx_conns)); + rc = (!ksocknal_data.ksnd_shuttingdown && + list_empty(&sched->kss_rx_conns) && + list_empty(&sched->kss_tx_conns)); - cfs_spin_unlock_bh (&sched->kss_lock); - return (rc); + spin_unlock_bh(&sched->kss_lock); + return rc; } -int ksocknal_scheduler (void *arg) +int ksocknal_scheduler(void *arg) { - ksock_sched_t *sched = (ksock_sched_t *)arg; - ksock_conn_t *conn; - ksock_tx_t *tx; - int rc; - int nloops = 0; - int id = (int)(sched - ksocknal_data.ksnd_schedulers); - char name[16]; - - snprintf (name, sizeof (name),"socknal_sd%02d", id); - cfs_daemonize (name); - cfs_block_allsigs (); - - if (ksocknal_lib_bind_thread_to_cpu(id)) - CERROR ("Can't set CPU affinity for %s to %d\n", name, id); - - cfs_spin_lock_bh (&sched->kss_lock); - - while (!ksocknal_data.ksnd_shuttingdown) { - int did_something = 0; - - /* Ensure I progress everything semi-fairly */ - - if (!cfs_list_empty (&sched->kss_rx_conns)) { - conn = cfs_list_entry(sched->kss_rx_conns.next, - ksock_conn_t, ksnc_rx_list); - cfs_list_del(&conn->ksnc_rx_list); - - LASSERT(conn->ksnc_rx_scheduled); - LASSERT(conn->ksnc_rx_ready); - - /* clear rx_ready in case receive isn't complete. - * Do it BEFORE we call process_recv, since - * data_ready can set it any time after we release - * kss_lock. */ - conn->ksnc_rx_ready = 0; - cfs_spin_unlock_bh (&sched->kss_lock); - - rc = ksocknal_process_receive(conn); - - cfs_spin_lock_bh (&sched->kss_lock); - - /* I'm the only one that can clear this flag */ - LASSERT(conn->ksnc_rx_scheduled); - - /* Did process_receive get everything it wanted? */ - if (rc == 0) - conn->ksnc_rx_ready = 1; - - if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) { - /* Conn blocked waiting for ksocknal_recv() - * I change its state (under lock) to signal - * it can be rescheduled */ - conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT; - } else if (conn->ksnc_rx_ready) { - /* reschedule for rx */ - cfs_list_add_tail (&conn->ksnc_rx_list, - &sched->kss_rx_conns); - } else { - conn->ksnc_rx_scheduled = 0; - /* drop my ref */ - ksocknal_conn_decref(conn); - } - - did_something = 1; - } - - if (!cfs_list_empty (&sched->kss_tx_conns)) { - CFS_LIST_HEAD (zlist); - - if (!cfs_list_empty(&sched->kss_zombie_noop_txs)) { - cfs_list_add(&zlist, - &sched->kss_zombie_noop_txs); - cfs_list_del_init(&sched->kss_zombie_noop_txs); - } - - conn = cfs_list_entry(sched->kss_tx_conns.next, - ksock_conn_t, ksnc_tx_list); - cfs_list_del (&conn->ksnc_tx_list); - - LASSERT(conn->ksnc_tx_scheduled); - LASSERT(conn->ksnc_tx_ready); - LASSERT(!cfs_list_empty(&conn->ksnc_tx_queue)); - - tx = cfs_list_entry(conn->ksnc_tx_queue.next, - ksock_tx_t, tx_list); - - if (conn->ksnc_tx_carrier == tx) - ksocknal_next_tx_carrier(conn); - - /* dequeue now so empty list => more to send */ - cfs_list_del(&tx->tx_list); - - /* Clear tx_ready in case send isn't complete. Do - * it BEFORE we call process_transmit, since - * write_space can set it any time after we release - * kss_lock. */ - conn->ksnc_tx_ready = 0; - cfs_spin_unlock_bh (&sched->kss_lock); - - if (!cfs_list_empty(&zlist)) { - /* free zombie noop txs, it's fast because - * noop txs are just put in freelist */ - ksocknal_txlist_done(NULL, &zlist, 0); - } - - rc = ksocknal_process_transmit(conn, tx); - - if (rc == -ENOMEM || rc == -EAGAIN) { - /* Incomplete send: replace tx on HEAD of tx_queue */ - cfs_spin_lock_bh (&sched->kss_lock); - cfs_list_add (&tx->tx_list, - &conn->ksnc_tx_queue); - } else { - /* Complete send; tx -ref */ - ksocknal_tx_decref (tx); - - cfs_spin_lock_bh (&sched->kss_lock); - /* assume space for more */ - conn->ksnc_tx_ready = 1; - } - - if (rc == -ENOMEM) { - /* Do nothing; after a short timeout, this - * conn will be reposted on kss_tx_conns. */ - } else if (conn->ksnc_tx_ready && - !cfs_list_empty (&conn->ksnc_tx_queue)) { - /* reschedule for tx */ - cfs_list_add_tail (&conn->ksnc_tx_list, - &sched->kss_tx_conns); - } else { - conn->ksnc_tx_scheduled = 0; - /* drop my ref */ - ksocknal_conn_decref(conn); - } - - did_something = 1; - } - if (!did_something || /* nothing to do */ - ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */ - cfs_spin_unlock_bh (&sched->kss_lock); - - nloops = 0; - - if (!did_something) { /* wait for something to do */ - cfs_wait_event_interruptible_exclusive( - sched->kss_waitq, - !ksocknal_sched_cansleep(sched), rc); - LASSERT (rc == 0); - } else { - cfs_cond_resched(); - } - - cfs_spin_lock_bh (&sched->kss_lock); - } - } - - cfs_spin_unlock_bh (&sched->kss_lock); - ksocknal_thread_fini (); - return (0); + struct ksock_sched *sched; + struct ksock_conn *conn; + struct ksock_tx *tx; + int rc; + int nloops = 0; + long id = (long)arg; + struct page **rx_scratch_pgs; + struct kvec *scratch_iov; + + sched = ksocknal_data.ksnd_schedulers[KSOCK_THREAD_CPT(id)]; + + LIBCFS_CPT_ALLOC(rx_scratch_pgs, lnet_cpt_table(), sched->kss_cpt, + sizeof(*rx_scratch_pgs) * LNET_MAX_IOV); + if (!rx_scratch_pgs) { + CERROR("Unable to allocate scratch pages\n"); + return -ENOMEM; + } + + LIBCFS_CPT_ALLOC(scratch_iov, lnet_cpt_table(), sched->kss_cpt, + sizeof(*scratch_iov) * LNET_MAX_IOV); + if (!scratch_iov) { + CERROR("Unable to allocate scratch iov\n"); + return -ENOMEM; + } + + rc = cfs_cpt_bind(lnet_cpt_table(), sched->kss_cpt); + if (rc != 0) { + CWARN("Can't set CPU partition affinity to %d: %d\n", + sched->kss_cpt, rc); + } + + spin_lock_bh(&sched->kss_lock); + + while (!ksocknal_data.ksnd_shuttingdown) { + int did_something = 0; + + /* Ensure I progress everything semi-fairly */ + + if (!list_empty(&sched->kss_rx_conns)) { + conn = list_entry(sched->kss_rx_conns.next, + struct ksock_conn, ksnc_rx_list); + list_del(&conn->ksnc_rx_list); + + LASSERT(conn->ksnc_rx_scheduled); + LASSERT(conn->ksnc_rx_ready); + + /* clear rx_ready in case receive isn't complete. + * Do it BEFORE we call process_recv, since + * data_ready can set it any time after we release + * kss_lock. */ + conn->ksnc_rx_ready = 0; + spin_unlock_bh(&sched->kss_lock); + + rc = ksocknal_process_receive(conn, rx_scratch_pgs, + scratch_iov); + + spin_lock_bh(&sched->kss_lock); + + /* I'm the only one that can clear this flag */ + LASSERT(conn->ksnc_rx_scheduled); + + /* Did process_receive get everything it wanted? */ + if (rc == 0) + conn->ksnc_rx_ready = 1; + + if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) { + /* Conn blocked waiting for ksocknal_recv() + * I change its state (under lock) to signal + * it can be rescheduled */ + conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT; + } else if (conn->ksnc_rx_ready) { + /* reschedule for rx */ + list_add_tail(&conn->ksnc_rx_list, + &sched->kss_rx_conns); + } else { + conn->ksnc_rx_scheduled = 0; + /* drop my ref */ + ksocknal_conn_decref(conn); + } + + did_something = 1; + } + + if (!list_empty(&sched->kss_tx_conns)) { + LIST_HEAD(zlist); + + list_splice_init(&sched->kss_zombie_noop_txs, &zlist); + + conn = list_entry(sched->kss_tx_conns.next, + struct ksock_conn, ksnc_tx_list); + list_del(&conn->ksnc_tx_list); + + LASSERT(conn->ksnc_tx_scheduled); + LASSERT(conn->ksnc_tx_ready); + LASSERT(!list_empty(&conn->ksnc_tx_queue)); + + tx = list_entry(conn->ksnc_tx_queue.next, + struct ksock_tx, tx_list); + + if (conn->ksnc_tx_carrier == tx) + ksocknal_next_tx_carrier(conn); + + /* dequeue now so empty list => more to send */ + list_del(&tx->tx_list); + + /* Clear tx_ready in case send isn't complete. Do + * it BEFORE we call process_transmit, since + * write_space can set it any time after we release + * kss_lock. */ + conn->ksnc_tx_ready = 0; + spin_unlock_bh(&sched->kss_lock); + + if (!list_empty(&zlist)) { + /* free zombie noop txs, it's fast because + * noop txs are just put in freelist */ + ksocknal_txlist_done(NULL, &zlist, 0); + } + + rc = ksocknal_process_transmit(conn, tx, scratch_iov); + + if (rc == -ENOMEM || rc == -EAGAIN) { + /* Incomplete send: replace tx on HEAD of tx_queue */ + spin_lock_bh(&sched->kss_lock); + list_add(&tx->tx_list, + &conn->ksnc_tx_queue); + } else { + /* Complete send; tx -ref */ + ksocknal_tx_decref(tx); + + spin_lock_bh(&sched->kss_lock); + /* assume space for more */ + conn->ksnc_tx_ready = 1; + } + + if (rc == -ENOMEM) { + /* Do nothing; after a short timeout, this + * conn will be reposted on kss_tx_conns. */ + } else if (conn->ksnc_tx_ready && + !list_empty(&conn->ksnc_tx_queue)) { + /* reschedule for tx */ + list_add_tail(&conn->ksnc_tx_list, + &sched->kss_tx_conns); + } else { + conn->ksnc_tx_scheduled = 0; + /* drop my ref */ + ksocknal_conn_decref(conn); + } + + did_something = 1; + } + if (!did_something || /* nothing to do */ + ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */ + spin_unlock_bh(&sched->kss_lock); + + nloops = 0; + + if (!did_something) { /* wait for something to do */ + rc = wait_event_interruptible_exclusive( + sched->kss_waitq, + !ksocknal_sched_cansleep(sched)); + LASSERT (rc == 0); + } else { + cond_resched(); + } + + spin_lock_bh(&sched->kss_lock); + } + } + + spin_unlock_bh(&sched->kss_lock); + CFS_FREE_PTR_ARRAY(rx_scratch_pgs, LNET_MAX_IOV); + CFS_FREE_PTR_ARRAY(scratch_iov, LNET_MAX_IOV); + ksocknal_thread_fini(); + return 0; } /* * Add connection to kss_rx_conns of scheduler * and wakeup the scheduler. */ -void ksocknal_read_callback (ksock_conn_t *conn) +void ksocknal_read_callback(struct ksock_conn *conn) { - ksock_sched_t *sched; - ENTRY; + struct ksock_sched *sched; + ENTRY; - sched = conn->ksnc_scheduler; + sched = conn->ksnc_scheduler; - cfs_spin_lock_bh (&sched->kss_lock); + spin_lock_bh(&sched->kss_lock); - conn->ksnc_rx_ready = 1; + conn->ksnc_rx_ready = 1; - if (!conn->ksnc_rx_scheduled) { /* not being progressed */ - cfs_list_add_tail(&conn->ksnc_rx_list, - &sched->kss_rx_conns); - conn->ksnc_rx_scheduled = 1; - /* extra ref for scheduler */ - ksocknal_conn_addref(conn); + if (!conn->ksnc_rx_scheduled) { /* not being progressed */ + list_add_tail(&conn->ksnc_rx_list, + &sched->kss_rx_conns); + conn->ksnc_rx_scheduled = 1; + /* extra ref for scheduler */ + ksocknal_conn_addref(conn); - cfs_waitq_signal (&sched->kss_waitq); - } - cfs_spin_unlock_bh (&sched->kss_lock); + wake_up (&sched->kss_waitq); + } + spin_unlock_bh(&sched->kss_lock); - EXIT; + EXIT; } /* * Add connection to kss_tx_conns of scheduler * and wakeup the scheduler. */ -void ksocknal_write_callback (ksock_conn_t *conn) +void ksocknal_write_callback(struct ksock_conn *conn) { - ksock_sched_t *sched; - ENTRY; + struct ksock_sched *sched; + ENTRY; - sched = conn->ksnc_scheduler; + sched = conn->ksnc_scheduler; - cfs_spin_lock_bh (&sched->kss_lock); + spin_lock_bh(&sched->kss_lock); - conn->ksnc_tx_ready = 1; + conn->ksnc_tx_ready = 1; - if (!conn->ksnc_tx_scheduled && // not being progressed - !cfs_list_empty(&conn->ksnc_tx_queue)){//packets to send - cfs_list_add_tail (&conn->ksnc_tx_list, - &sched->kss_tx_conns); - conn->ksnc_tx_scheduled = 1; - /* extra ref for scheduler */ - ksocknal_conn_addref(conn); + if (!conn->ksnc_tx_scheduled && /* not being progressed */ + !list_empty(&conn->ksnc_tx_queue)) { /* packets to send */ + list_add_tail(&conn->ksnc_tx_list, &sched->kss_tx_conns); + conn->ksnc_tx_scheduled = 1; + /* extra ref for scheduler */ + ksocknal_conn_addref(conn); - cfs_waitq_signal (&sched->kss_waitq); - } + wake_up(&sched->kss_waitq); + } - cfs_spin_unlock_bh (&sched->kss_lock); + spin_unlock_bh(&sched->kss_lock); - EXIT; + EXIT; } -ksock_proto_t * -ksocknal_parse_proto_version (ksock_hello_msg_t *hello) +static const struct ksock_proto * +ksocknal_parse_proto_version(struct ksock_hello_msg *hello) { __u32 version = 0; @@ -1625,10 +1722,12 @@ ksocknal_parse_proto_version (ksock_hello_msg_t *hello) } if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) { - lnet_magicversion_t *hmv = (lnet_magicversion_t *)hello; + struct lnet_magicversion *hmv; + + BUILD_BUG_ON(sizeof(struct lnet_magicversion) != + offsetof(struct ksock_hello_msg, kshm_src_nid)); - CLASSERT (sizeof (lnet_magicversion_t) == - offsetof (ksock_hello_msg_t, kshm_src_nid)); + hmv = (struct lnet_magicversion *)hello; if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) && hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR)) @@ -1639,28 +1738,28 @@ ksocknal_parse_proto_version (ksock_hello_msg_t *hello) } int -ksocknal_send_hello (lnet_ni_t *ni, ksock_conn_t *conn, - lnet_nid_t peer_nid, ksock_hello_msg_t *hello) +ksocknal_send_hello(struct lnet_ni *ni, struct ksock_conn *conn, + lnet_nid_t peer_nid, struct ksock_hello_msg *hello) { - /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */ - ksock_net_t *net = (ksock_net_t *)ni->ni_data; + /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */ + struct ksock_net *net = (struct ksock_net *)ni->ni_data; - LASSERT (hello->kshm_nips <= LNET_MAX_INTERFACES); + LASSERT(hello->kshm_nips <= LNET_INTERFACES_NUM); - /* rely on caller to hold a ref on socket so it wouldn't disappear */ - LASSERT (conn->ksnc_proto != NULL); + /* rely on caller to hold a ref on socket so it wouldn't disappear */ + LASSERT(conn->ksnc_proto != NULL); - hello->kshm_src_nid = ni->ni_nid; - hello->kshm_dst_nid = peer_nid; - hello->kshm_src_pid = the_lnet.ln_pid; + hello->kshm_src_nid = ni->ni_nid; + hello->kshm_dst_nid = peer_nid; + hello->kshm_src_pid = the_lnet.ln_pid; - hello->kshm_src_incarnation = net->ksnn_incarnation; - hello->kshm_ctype = conn->ksnc_type; + hello->kshm_src_incarnation = net->ksnn_incarnation; + hello->kshm_ctype = conn->ksnc_type; - return conn->ksnc_proto->pro_send_hello(conn, hello); + return conn->ksnc_proto->pro_send_hello(conn, hello); } -int +static int ksocknal_invert_type(int type) { switch (type) @@ -1678,33 +1777,35 @@ ksocknal_invert_type(int type) } int -ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn, - ksock_hello_msg_t *hello, lnet_process_id_t *peerid, - __u64 *incarnation) +ksocknal_recv_hello(struct lnet_ni *ni, struct ksock_conn *conn, + struct ksock_hello_msg *hello, + struct lnet_process_id *peerid, + __u64 *incarnation) { /* Return < 0 fatal error * 0 success * EALREADY lost connection race * EPROTO protocol version mismatch */ - cfs_socket_t *sock = conn->ksnc_sock; + struct socket *sock = conn->ksnc_sock; int active = (conn->ksnc_proto != NULL); int timeout; int proto_match; int rc; - ksock_proto_t *proto; - lnet_process_id_t recv_id; + const struct ksock_proto *proto; + struct lnet_process_id recv_id; - /* socket type set on active connections - not set on passive */ - LASSERT (!active == !(conn->ksnc_type != SOCKLND_CONN_NONE)); + /* socket type set on active connections - not set on passive */ + LASSERT(!active == !(conn->ksnc_type != SOCKLND_CONN_NONE)); - timeout = active ? *ksocknal_tunables.ksnd_timeout : - lnet_acceptor_timeout(); + timeout = active ? lnet_get_lnd_timeout() : + lnet_acceptor_timeout(); - rc = libcfs_sock_read(sock, &hello->kshm_magic, sizeof (hello->kshm_magic), timeout); + rc = lnet_sock_read(sock, &hello->kshm_magic, + sizeof(hello->kshm_magic), timeout); if (rc != 0) { - CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n", - rc, HIPQUAD(conn->ksnc_ipaddr)); + CERROR("Error %d reading HELLO from %pI4h\n", + rc, &conn->ksnc_ipaddr); LASSERT (rc < 0); return rc; } @@ -1714,25 +1815,24 @@ ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn, hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) { /* Unexpected magic! */ CERROR ("Bad magic(1) %#08x (%#08x expected) from " - "%u.%u.%u.%u\n", __cpu_to_le32 (hello->kshm_magic), - LNET_PROTO_TCP_MAGIC, - HIPQUAD(conn->ksnc_ipaddr)); + "%pI4h\n", __cpu_to_le32 (hello->kshm_magic), + LNET_PROTO_TCP_MAGIC, &conn->ksnc_ipaddr); return -EPROTO; } - rc = libcfs_sock_read(sock, &hello->kshm_version, - sizeof(hello->kshm_version), timeout); + rc = lnet_sock_read(sock, &hello->kshm_version, + sizeof(hello->kshm_version), timeout); if (rc != 0) { - CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n", - rc, HIPQUAD(conn->ksnc_ipaddr)); - LASSERT (rc < 0); + CERROR("Error %d reading HELLO from %pI4h\n", + rc, &conn->ksnc_ipaddr); + LASSERT(rc < 0); return rc; } proto = ksocknal_parse_proto_version(hello); if (proto == NULL) { if (!active) { - /* unknown protocol from peer, tell peer my protocol */ + /* unknown protocol from peer_ni, tell peer_ni my protocol */ conn->ksnc_proto = &ksocknal_protocol_v3x; #if SOCKNAL_VERSION_DEBUG if (*ksocknal_tunables.ksnd_protocol == 2) @@ -1744,10 +1844,8 @@ ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn, ksocknal_send_hello(ni, conn, ni->ni_nid, hello); } - CERROR ("Unknown protocol version (%d.x expected)" - " from %u.%u.%u.%u\n", - conn->ksnc_proto->pro_version, - HIPQUAD(conn->ksnc_ipaddr)); + CERROR("Unknown protocol version (%d.x expected) from %pI4h\n", + conn->ksnc_proto->pro_version, &conn->ksnc_ipaddr); return -EPROTO; } @@ -1758,23 +1856,23 @@ ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn, /* receive the rest of hello message anyway */ rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout); if (rc != 0) { - CERROR("Error %d reading or checking hello from from %u.%u.%u.%u\n", - rc, HIPQUAD(conn->ksnc_ipaddr)); + CERROR("Error %d reading or checking hello from from %pI4h\n", + rc, &conn->ksnc_ipaddr); LASSERT (rc < 0); return rc; } *incarnation = hello->kshm_src_incarnation; - if (hello->kshm_src_nid == LNET_NID_ANY) { - CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY" - "from %u.%u.%u.%u\n", HIPQUAD(conn->ksnc_ipaddr)); - return -EPROTO; - } + if (hello->kshm_src_nid == LNET_NID_ANY) { + CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY from %pI4h\n", + &conn->ksnc_ipaddr); + return -EPROTO; + } if (!active && conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) { - /* Userspace NAL assigns peer process ID from socket */ + /* Userspace NAL assigns peer_ni process ID from socket */ recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG; recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr); } else { @@ -1785,26 +1883,25 @@ ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn, if (!active) { *peerid = recv_id; - /* peer determines type */ - conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype); - if (conn->ksnc_type == SOCKLND_CONN_NONE) { - CERROR ("Unexpected type %d from %s ip %u.%u.%u.%u\n", - hello->kshm_ctype, libcfs_id2str(*peerid), - HIPQUAD(conn->ksnc_ipaddr)); - return -EPROTO; - } - - return 0; - } + /* peer_ni determines type */ + conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype); + if (conn->ksnc_type == SOCKLND_CONN_NONE) { + CERROR("Unexpected type %d from %s ip %pI4h\n", + hello->kshm_ctype, libcfs_id2str(*peerid), + &conn->ksnc_ipaddr); + return -EPROTO; + } + return 0; + } if (peerid->pid != recv_id.pid || peerid->nid != recv_id.nid) { LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host" - " %u.%u.%u.%u, but they claimed they were " + " %pI4h, but they claimed they were " "%s; please check your Lustre " "configuration.\n", libcfs_id2str(*peerid), - HIPQUAD(conn->ksnc_ipaddr), + &conn->ksnc_ipaddr, libcfs_id2str(recv_id)); return -EPROTO; } @@ -1814,33 +1911,31 @@ ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn, return proto_match ? EALREADY : EPROTO; } - if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) { - CERROR ("Mismatched types: me %d, %s ip %u.%u.%u.%u %d\n", - conn->ksnc_type, libcfs_id2str(*peerid), - HIPQUAD(conn->ksnc_ipaddr), - hello->kshm_ctype); - return -EPROTO; - } - - return 0; + if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) { + CERROR("Mismatched types: me %d, %s ip %pI4h %d\n", + conn->ksnc_type, libcfs_id2str(*peerid), + &conn->ksnc_ipaddr, + hello->kshm_ctype); + return -EPROTO; + } + return 0; } -int -ksocknal_connect (ksock_route_t *route) +static int +ksocknal_connect(struct ksock_route *route) { - CFS_LIST_HEAD (zombies); - ksock_peer_t *peer = route->ksnr_peer; + LIST_HEAD(zombies); + struct ksock_peer_ni *peer_ni = route->ksnr_peer; int type; int wanted; - cfs_socket_t *sock; - cfs_time_t deadline; + struct socket *sock; + time64_t deadline; int retry_later = 0; int rc = 0; - deadline = cfs_time_add(cfs_time_current(), - cfs_time_seconds(*ksocknal_tunables.ksnd_timeout)); + deadline = ktime_get_seconds() + lnet_get_lnd_timeout(); - cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock); + write_lock_bh(&ksocknal_data.ksnd_global_lock); LASSERT (route->ksnr_scheduled); LASSERT (!route->ksnr_connecting); @@ -1850,19 +1945,19 @@ ksocknal_connect (ksock_route_t *route) for (;;) { wanted = ksocknal_route_mask() & ~route->ksnr_connected; - /* stop connecting if peer/route got closed under me, or + /* stop connecting if peer_ni/route got closed under me, or * route got connected while queued */ - if (peer->ksnp_closing || route->ksnr_deleted || + if (peer_ni->ksnp_closing || route->ksnr_deleted || wanted == 0) { retry_later = 0; break; } - /* reschedule if peer is connecting to me */ - if (peer->ksnp_accepting > 0) { + /* reschedule if peer_ni is connecting to me */ + if (peer_ni->ksnp_accepting > 0) { CDEBUG(D_NET, - "peer %s(%d) already connecting to me, retry later.\n", - libcfs_nid2str(peer->ksnp_id.nid), peer->ksnp_accepting); + "peer_ni %s(%d) already connecting to me, retry later.\n", + libcfs_nid2str(peer_ni->ksnp_id.nid), peer_ni->ksnp_accepting); retry_later = 1; } @@ -1880,25 +1975,28 @@ ksocknal_connect (ksock_route_t *route) type = SOCKLND_CONN_BULK_OUT; } - cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock); + write_unlock_bh(&ksocknal_data.ksnd_global_lock); - if (cfs_time_aftereq(cfs_time_current(), deadline)) { + if (ktime_get_seconds() >= deadline) { rc = -ETIMEDOUT; - lnet_connect_console_error(rc, peer->ksnp_id.nid, + lnet_connect_console_error(rc, peer_ni->ksnp_id.nid, route->ksnr_ipaddr, route->ksnr_port); goto failed; } - rc = lnet_connect(&sock, peer->ksnp_id.nid, - route->ksnr_myipaddr, - route->ksnr_ipaddr, route->ksnr_port); - if (rc != 0) - goto failed; + sock = lnet_connect(peer_ni->ksnp_id.nid, + route->ksnr_myiface, + route->ksnr_ipaddr, route->ksnr_port, + peer_ni->ksnp_ni->ni_net_ns); + if (IS_ERR(sock)) { + rc = PTR_ERR(sock); + goto failed; + } - rc = ksocknal_create_conn(peer->ksnp_ni, route, sock, type); + rc = ksocknal_create_conn(peer_ni->ksnp_ni, route, sock, type); if (rc < 0) { - lnet_connect_console_error(rc, peer->ksnp_id.nid, + lnet_connect_console_error(rc, peer_ni->ksnp_id.nid, route->ksnr_ipaddr, route->ksnr_port); goto failed; @@ -1908,10 +2006,10 @@ ksocknal_connect (ksock_route_t *route) * race or I have to renegotiate protocol version */ retry_later = (rc != 0); if (retry_later) - CDEBUG(D_NET, "peer %s: conn race, retry later.\n", - libcfs_nid2str(peer->ksnp_id.nid)); + CDEBUG(D_NET, "peer_ni %s: conn race, retry later.\n", + libcfs_nid2str(peer_ni->ksnp_id.nid)); - cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock); + write_lock_bh(&ksocknal_data.ksnd_global_lock); } route->ksnr_scheduled = 0; @@ -1919,169 +2017,265 @@ ksocknal_connect (ksock_route_t *route) if (retry_later) { /* re-queue for attention; this frees me up to handle - * the peer's incoming connection request */ + * the peer_ni's incoming connection request */ if (rc == EALREADY || - (rc == 0 && peer->ksnp_accepting > 0)) { + (rc == 0 && peer_ni->ksnp_accepting > 0)) { /* We want to introduce a delay before next * attempt to connect if we lost conn race, * but the race is resolved quickly usually, * so min_reconnectms should be good heuristic */ - route->ksnr_retry_interval = - cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000; - route->ksnr_timeout = cfs_time_add(cfs_time_current(), - route->ksnr_retry_interval); + route->ksnr_retry_interval = *ksocknal_tunables.ksnd_min_reconnectms / 1000; + route->ksnr_timeout = ktime_get_seconds() + + route->ksnr_retry_interval; } ksocknal_launch_connection_locked(route); } - cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock); + write_unlock_bh(&ksocknal_data.ksnd_global_lock); return retry_later; failed: - cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock); + write_lock_bh(&ksocknal_data.ksnd_global_lock); route->ksnr_scheduled = 0; route->ksnr_connecting = 0; - /* This is a retry rather than a new connection */ - route->ksnr_retry_interval *= 2; - route->ksnr_retry_interval = - MAX(route->ksnr_retry_interval, - cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000); - route->ksnr_retry_interval = - MIN(route->ksnr_retry_interval, - cfs_time_seconds(*ksocknal_tunables.ksnd_max_reconnectms)/1000); + /* This is a retry rather than a new connection */ + route->ksnr_retry_interval *= 2; + route->ksnr_retry_interval = + max_t(time64_t, route->ksnr_retry_interval, + *ksocknal_tunables.ksnd_min_reconnectms / 1000); + route->ksnr_retry_interval = + min_t(time64_t, route->ksnr_retry_interval, + *ksocknal_tunables.ksnd_max_reconnectms / 1000); - LASSERT (route->ksnr_retry_interval != 0); - route->ksnr_timeout = cfs_time_add(cfs_time_current(), - route->ksnr_retry_interval); + LASSERT(route->ksnr_retry_interval); + route->ksnr_timeout = ktime_get_seconds() + route->ksnr_retry_interval; - if (!cfs_list_empty(&peer->ksnp_tx_queue) && - peer->ksnp_accepting == 0 && - ksocknal_find_connecting_route_locked(peer) == NULL) { - ksock_conn_t *conn; + if (!list_empty(&peer_ni->ksnp_tx_queue) && + peer_ni->ksnp_accepting == 0 && + ksocknal_find_connecting_route_locked(peer_ni) == NULL) { + struct ksock_conn *conn; /* ksnp_tx_queue is queued on a conn on successful * connection for V1.x and V2.x */ - if (!cfs_list_empty (&peer->ksnp_conns)) { - conn = cfs_list_entry(peer->ksnp_conns.next, - ksock_conn_t, ksnc_list); + if (!list_empty(&peer_ni->ksnp_conns)) { + conn = list_entry(peer_ni->ksnp_conns.next, + struct ksock_conn, ksnc_list); LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x); } /* take all the blocked packets while I've got the lock and * complete below... */ - cfs_list_splice_init(&peer->ksnp_tx_queue, &zombies); + list_splice_init(&peer_ni->ksnp_tx_queue, &zombies); } -#if 0 /* irrelevent with only eager routes */ - if (!route->ksnr_deleted) { - /* make this route least-favourite for re-selection */ - cfs_list_del(&route->ksnr_list); - cfs_list_add_tail(&route->ksnr_list, &peer->ksnp_routes); - } -#endif - cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock); + write_unlock_bh(&ksocknal_data.ksnd_global_lock); - ksocknal_peer_failed(peer); - ksocknal_txlist_done(peer->ksnp_ni, &zombies, 1); - return 0; + ksocknal_peer_failed(peer_ni); + ksocknal_txlist_done(peer_ni->ksnp_ni, &zombies, rc); + return 0; } -/* Go through connd_routes queue looking for a route that - we can process right now */ -static ksock_route_t * -ksocknal_connd_get_route_locked(signed long *timeout_p) +/* + * check whether we need to create more connds. + * It will try to create new thread if it's necessary, @timeout can + * be updated if failed to create, so caller wouldn't keep try while + * running out of resource. + */ +static int +ksocknal_connd_check_start(time64_t sec, long *timeout) { - ksock_route_t *route; - cfs_time_t now; + char name[16]; + int rc; + int total = ksocknal_data.ksnd_connd_starting + + ksocknal_data.ksnd_connd_running; - /* Only handle an outgoing connection request if there - * is someone left to handle incoming connections */ - if ((ksocknal_data.ksnd_connd_connecting + 1) >= - *ksocknal_tunables.ksnd_nconnds) - return NULL; + if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) { + /* still in initializing */ + return 0; + } - now = cfs_time_current(); + if (total >= *ksocknal_tunables.ksnd_nconnds_max || + total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) { + /* can't create more connd, or still have enough + * threads to handle more connecting */ + return 0; + } - /* connd_routes can contain both pending and ordinary routes */ - cfs_list_for_each_entry (route, &ksocknal_data.ksnd_connd_routes, - ksnr_connd_list) { + if (list_empty(&ksocknal_data.ksnd_connd_routes)) { + /* no pending connecting request */ + return 0; + } - if (route->ksnr_retry_interval == 0 || - cfs_time_aftereq(now, route->ksnr_timeout)) - return route; + if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) { + /* may run out of resource, retry later */ + *timeout = cfs_time_seconds(1); + return 0; + } - if (*timeout_p == CFS_MAX_SCHEDULE_TIMEOUT || - (int)*timeout_p > (int)(route->ksnr_timeout - now)) - *timeout_p = (int)(route->ksnr_timeout - now); + if (ksocknal_data.ksnd_connd_starting > 0) { + /* serialize starting to avoid flood */ + return 0; } - return NULL; + ksocknal_data.ksnd_connd_starting_stamp = sec; + ksocknal_data.ksnd_connd_starting++; + spin_unlock_bh(&ksocknal_data.ksnd_connd_lock); + + /* NB: total is the next id */ + snprintf(name, sizeof(name), "socknal_cd%02d", total); + rc = ksocknal_thread_start(ksocknal_connd, NULL, name); + + spin_lock_bh(&ksocknal_data.ksnd_connd_lock); + if (rc == 0) + return 1; + + /* we tried ... */ + LASSERT(ksocknal_data.ksnd_connd_starting > 0); + ksocknal_data.ksnd_connd_starting--; + ksocknal_data.ksnd_connd_failed_stamp = ktime_get_real_seconds(); + + return 1; } -int -ksocknal_connd (void *arg) +/* + * check whether current thread can exit, it will return 1 if there are too + * many threads and no creating in past 120 seconds. + * Also, this function may update @timeout to make caller come back + * again to recheck these conditions. + */ +static int +ksocknal_connd_check_stop(time64_t sec, long *timeout) { - long id = (long)(long_ptr_t)arg; - char name[16]; - ksock_connreq_t *cr; - ksock_route_t *route; - cfs_waitlink_t wait; - signed long timeout; - int nloops = 0; - int cons_retry = 0; - int dropped_lock; + int val; + + if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) { + /* still in initializing */ + return 0; + } + + if (ksocknal_data.ksnd_connd_starting > 0) { + /* in progress of starting new thread */ + return 0; + } - snprintf (name, sizeof (name), "socknal_cd%02ld", id); - cfs_daemonize (name); - cfs_block_allsigs (); + if (ksocknal_data.ksnd_connd_running <= + *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */ + return 0; + } - cfs_waitlink_init (&wait); + /* created thread in past 120 seconds? */ + val = (int)(ksocknal_data.ksnd_connd_starting_stamp + + SOCKNAL_CONND_TIMEOUT - sec); - cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock); + *timeout = (val > 0) ? cfs_time_seconds(val) : + cfs_time_seconds(SOCKNAL_CONND_TIMEOUT); + if (val > 0) + return 0; - while (!ksocknal_data.ksnd_shuttingdown) { + /* no creating in past 120 seconds */ - dropped_lock = 0; + return ksocknal_data.ksnd_connd_running > + ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV; +} - if (!cfs_list_empty(&ksocknal_data.ksnd_connd_connreqs)) { - /* Connection accepted by the listener */ - cr = cfs_list_entry(ksocknal_data.ksnd_connd_connreqs. \ - next, ksock_connreq_t, ksncr_list); +/* Go through connd_routes queue looking for a route that we can process + * right now, @timeout_p can be updated if we need to come back later */ +static struct ksock_route * +ksocknal_connd_get_route_locked(signed long *timeout_p) +{ + time64_t now = ktime_get_seconds(); + struct ksock_route *route; - cfs_list_del(&cr->ksncr_list); - cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock); - dropped_lock = 1; + /* connd_routes can contain both pending and ordinary routes */ + list_for_each_entry(route, &ksocknal_data.ksnd_connd_routes, + ksnr_connd_list) { - ksocknal_create_conn(cr->ksncr_ni, NULL, - cr->ksncr_sock, SOCKLND_CONN_NONE); - lnet_ni_decref(cr->ksncr_ni); - LIBCFS_FREE(cr, sizeof(*cr)); + if (route->ksnr_retry_interval == 0 || + now >= route->ksnr_timeout) + return route; - cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock); + if (*timeout_p == MAX_SCHEDULE_TIMEOUT || + *timeout_p > cfs_time_seconds(route->ksnr_timeout - now)) + *timeout_p = cfs_time_seconds(route->ksnr_timeout - now); + } + + return NULL; +} + +int +ksocknal_connd(void *arg) +{ + spinlock_t *connd_lock = &ksocknal_data.ksnd_connd_lock; + struct ksock_connreq *cr; + wait_queue_entry_t wait; + int nloops = 0; + int cons_retry = 0; + + init_waitqueue_entry(&wait, current); + + spin_lock_bh(connd_lock); + + LASSERT(ksocknal_data.ksnd_connd_starting > 0); + ksocknal_data.ksnd_connd_starting--; + ksocknal_data.ksnd_connd_running++; + + while (!ksocknal_data.ksnd_shuttingdown) { + struct ksock_route *route = NULL; + time64_t sec = ktime_get_real_seconds(); + long timeout = MAX_SCHEDULE_TIMEOUT; + int dropped_lock = 0; + + if (ksocknal_connd_check_stop(sec, &timeout)) { + /* wakeup another one to check stop */ + wake_up(&ksocknal_data.ksnd_connd_waitq); + break; + } + + if (ksocknal_connd_check_start(sec, &timeout)) { + /* created new thread */ + dropped_lock = 1; } - /* Sleep till explicit wake_up if no pending routes present */ - timeout = CFS_MAX_SCHEDULE_TIMEOUT; + if (!list_empty(&ksocknal_data.ksnd_connd_connreqs)) { + /* Connection accepted by the listener */ + cr = list_entry(ksocknal_data.ksnd_connd_connreqs.next, + struct ksock_connreq, ksncr_list); + + list_del(&cr->ksncr_list); + spin_unlock_bh(connd_lock); + dropped_lock = 1; + + ksocknal_create_conn(cr->ksncr_ni, NULL, + cr->ksncr_sock, SOCKLND_CONN_NONE); + lnet_ni_decref(cr->ksncr_ni); + LIBCFS_FREE(cr, sizeof(*cr)); - /* Connection request */ - route = ksocknal_connd_get_route_locked(&timeout); + spin_lock_bh(connd_lock); + } + /* Only handle an outgoing connection request if there + * is a thread left to handle incoming connections and + * create new connd */ + if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV < + ksocknal_data.ksnd_connd_running) { + route = ksocknal_connd_get_route_locked(&timeout); + } if (route != NULL) { - cfs_list_del (&route->ksnr_connd_list); + list_del(&route->ksnr_connd_list); ksocknal_data.ksnd_connd_connecting++; - cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock); + spin_unlock_bh(connd_lock); dropped_lock = 1; if (ksocknal_connect(route)) { /* consecutive retry */ if (cons_retry++ > SOCKNAL_INSANITY_RECONN) { CWARN("massive consecutive " - "re-connecting to %u.%u.%u.%u\n", - HIPQUAD(route->ksnr_ipaddr)); + "re-connecting to %pI4h\n", + &route->ksnr_ipaddr); cons_retry = 0; } } else { @@ -2090,84 +2284,82 @@ ksocknal_connd (void *arg) ksocknal_route_decref(route); - cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock); - ksocknal_data.ksnd_connd_connecting--; - } - - if (dropped_lock) { - if (++nloops < SOCKNAL_RESCHED) - continue; - cfs_spin_unlock_bh(&ksocknal_data.ksnd_connd_lock); - nloops = 0; - cfs_cond_resched(); - cfs_spin_lock_bh(&ksocknal_data.ksnd_connd_lock); - continue; - } - - /* Nothing to do for 'timeout' */ - cfs_set_current_state (CFS_TASK_INTERRUPTIBLE); - cfs_waitq_add_exclusive (&ksocknal_data.ksnd_connd_waitq, - &wait); - cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock); - - nloops = 0; - cfs_waitq_timedwait (&wait, CFS_TASK_INTERRUPTIBLE, timeout); - - cfs_set_current_state (CFS_TASK_RUNNING); - cfs_waitq_del (&ksocknal_data.ksnd_connd_waitq, &wait); - cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock); - } - - cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock); - - ksocknal_thread_fini (); - return (0); + spin_lock_bh(connd_lock); + ksocknal_data.ksnd_connd_connecting--; + } + + if (dropped_lock) { + if (++nloops < SOCKNAL_RESCHED) + continue; + spin_unlock_bh(connd_lock); + nloops = 0; + cond_resched(); + spin_lock_bh(connd_lock); + continue; + } + + /* Nothing to do for 'timeout' */ + set_current_state(TASK_INTERRUPTIBLE); + add_wait_queue_exclusive(&ksocknal_data.ksnd_connd_waitq, &wait); + spin_unlock_bh(connd_lock); + + nloops = 0; + schedule_timeout(timeout); + + remove_wait_queue(&ksocknal_data.ksnd_connd_waitq, &wait); + spin_lock_bh(connd_lock); + } + ksocknal_data.ksnd_connd_running--; + spin_unlock_bh(connd_lock); + + ksocknal_thread_fini(); + return 0; } -ksock_conn_t * -ksocknal_find_timed_out_conn (ksock_peer_t *peer) +static struct ksock_conn * +ksocknal_find_timed_out_conn(struct ksock_peer_ni *peer_ni) { /* We're called with a shared lock on ksnd_global_lock */ - ksock_conn_t *conn; - cfs_list_t *ctmp; + struct ksock_conn *conn; + struct list_head *ctmp; + struct ksock_tx *tx; - cfs_list_for_each (ctmp, &peer->ksnp_conns) { - int error; - conn = cfs_list_entry (ctmp, ksock_conn_t, ksnc_list); + list_for_each(ctmp, &peer_ni->ksnp_conns) { + int error; + + conn = list_entry(ctmp, struct ksock_conn, ksnc_list); /* Don't need the {get,put}connsock dance to deref ksnc_sock */ LASSERT (!conn->ksnc_closing); - /* SOCK_ERROR will reset error code of socket in - * some platform (like Darwin8.x) */ - error = libcfs_sock_error(conn->ksnc_sock); + error = conn->ksnc_sock->sk->sk_err; if (error != 0) { ksocknal_conn_addref(conn); switch (error) { case ECONNRESET: - CDEBUG(D_NETERROR, "A connection with %s " - "(%u.%u.%u.%u:%d) was reset; " - "it may have rebooted.\n", - libcfs_id2str(peer->ksnp_id), - HIPQUAD(conn->ksnc_ipaddr), - conn->ksnc_port); + CNETERR("A connection with %s " + "(%pI4h:%d) was reset; " + "it may have rebooted.\n", + libcfs_id2str(peer_ni->ksnp_id), + &conn->ksnc_ipaddr, + conn->ksnc_port); break; case ETIMEDOUT: - CDEBUG(D_NETERROR, "A connection with %s " - "(%u.%u.%u.%u:%d) timed out; the " - "network or node may be down.\n", - libcfs_id2str(peer->ksnp_id), - HIPQUAD(conn->ksnc_ipaddr), - conn->ksnc_port); + CNETERR("A connection with %s " + "(%pI4h:%d) timed out; the " + "network or node may be down.\n", + libcfs_id2str(peer_ni->ksnp_id), + &conn->ksnc_ipaddr, + conn->ksnc_port); break; default: - CDEBUG(D_NETERROR, "An unexpected network error %d " - "occurred with %s " - "(%u.%u.%u.%u:%d\n", error, - libcfs_id2str(peer->ksnp_id), - HIPQUAD(conn->ksnc_ipaddr), - conn->ksnc_port); + CNETERR("An unexpected network error %d " + "occurred with %s " + "(%pI4h:%d\n", error, + libcfs_id2str(peer_ni->ksnp_id), + &conn->ksnc_ipaddr, + conn->ksnc_port); break; } @@ -2175,34 +2367,34 @@ ksocknal_find_timed_out_conn (ksock_peer_t *peer) } if (conn->ksnc_rx_started && - cfs_time_aftereq(cfs_time_current(), - conn->ksnc_rx_deadline)) { + ktime_get_seconds() >= conn->ksnc_rx_deadline) { /* Timed out incomplete incoming message */ ksocknal_conn_addref(conn); - CDEBUG(D_NETERROR, "Timeout receiving from %s " - "(%u.%u.%u.%u:%d), state %d wanted %d left %d\n", - libcfs_id2str(peer->ksnp_id), - HIPQUAD(conn->ksnc_ipaddr), - conn->ksnc_port, - conn->ksnc_rx_state, - conn->ksnc_rx_nob_wanted, - conn->ksnc_rx_nob_left); + CNETERR("Timeout receiving from %s (%pI4h:%d), " + "state %d wanted %d left %d\n", + libcfs_id2str(peer_ni->ksnp_id), + &conn->ksnc_ipaddr, + conn->ksnc_port, + conn->ksnc_rx_state, + conn->ksnc_rx_nob_wanted, + conn->ksnc_rx_nob_left); return (conn); } - if ((!cfs_list_empty(&conn->ksnc_tx_queue) || - libcfs_sock_wmem_queued(conn->ksnc_sock) != 0) && - cfs_time_aftereq(cfs_time_current(), - conn->ksnc_tx_deadline)) { + if ((!list_empty(&conn->ksnc_tx_queue) || + conn->ksnc_sock->sk->sk_wmem_queued != 0) && + ktime_get_seconds() >= conn->ksnc_tx_deadline) { /* Timed out messages queued for sending or * buffered in the socket's send buffer */ ksocknal_conn_addref(conn); - CDEBUG(D_NETERROR, "Timeout sending data to %s " - "(%u.%u.%u.%u:%d) the network or that " - "node may be down.\n", - libcfs_id2str(peer->ksnp_id), - HIPQUAD(conn->ksnc_ipaddr), - conn->ksnc_port); + list_for_each_entry(tx, &conn->ksnc_tx_queue, + tx_list) + tx->tx_hstatus = + LNET_MSG_STATUS_LOCAL_TIMEOUT; + CNETERR("Timeout sending data to %s (%pI4h:%d) " + "the network or that node may be down.\n", + libcfs_id2str(peer_ni->ksnp_id), + &conn->ksnc_ipaddr, conn->ksnc_port); return (conn); } } @@ -2211,303 +2403,295 @@ ksocknal_find_timed_out_conn (ksock_peer_t *peer) } static inline void -ksocknal_flush_stale_txs(ksock_peer_t *peer) +ksocknal_flush_stale_txs(struct ksock_peer_ni *peer_ni) { - ksock_tx_t *tx; - CFS_LIST_HEAD (stale_txs); + struct ksock_tx *tx; + LIST_HEAD(stale_txs); - cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock); + write_lock_bh(&ksocknal_data.ksnd_global_lock); - while (!cfs_list_empty (&peer->ksnp_tx_queue)) { - tx = cfs_list_entry (peer->ksnp_tx_queue.next, - ksock_tx_t, tx_list); + while (!list_empty(&peer_ni->ksnp_tx_queue)) { + tx = list_entry(peer_ni->ksnp_tx_queue.next, + struct ksock_tx, tx_list); - if (!cfs_time_aftereq(cfs_time_current(), - tx->tx_deadline)) - break; + if (ktime_get_seconds() < tx->tx_deadline) + break; - cfs_list_del (&tx->tx_list); - cfs_list_add_tail (&tx->tx_list, &stale_txs); - } + tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT; + + list_move_tail(&tx->tx_list, &stale_txs); + } - cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock); + write_unlock_bh(&ksocknal_data.ksnd_global_lock); - ksocknal_txlist_done(peer->ksnp_ni, &stale_txs, 1); + ksocknal_txlist_done(peer_ni->ksnp_ni, &stale_txs, -ETIMEDOUT); } -int -ksocknal_send_keepalive_locked(ksock_peer_t *peer) +static int +ksocknal_send_keepalive_locked(struct ksock_peer_ni *peer_ni) +__must_hold(&ksocknal_data.ksnd_global_lock) { - ksock_sched_t *sched; - ksock_conn_t *conn; - ksock_tx_t *tx; + struct ksock_sched *sched; + struct ksock_conn *conn; + struct ksock_tx *tx; - if (cfs_list_empty(&peer->ksnp_conns)) /* last_alive will be updated by create_conn */ + /* last_alive will be updated by create_conn */ + if (list_empty(&peer_ni->ksnp_conns)) return 0; - if (peer->ksnp_proto != &ksocknal_protocol_v3x) + if (peer_ni->ksnp_proto != &ksocknal_protocol_v3x) return 0; if (*ksocknal_tunables.ksnd_keepalive <= 0 || - cfs_time_before(cfs_time_current(), - cfs_time_add(peer->ksnp_last_alive, - cfs_time_seconds(*ksocknal_tunables.ksnd_keepalive)))) + ktime_get_seconds() < peer_ni->ksnp_last_alive + + *ksocknal_tunables.ksnd_keepalive) return 0; - if (cfs_time_before(cfs_time_current(), - peer->ksnp_send_keepalive)) + if (ktime_get_seconds() < peer_ni->ksnp_send_keepalive) return 0; /* retry 10 secs later, so we wouldn't put pressure - * on this peer if we failed to send keepalive this time */ - peer->ksnp_send_keepalive = cfs_time_shift(10); + * on this peer_ni if we failed to send keepalive this time */ + peer_ni->ksnp_send_keepalive = ktime_get_seconds() + 10; - conn = ksocknal_find_conn_locked(peer, NULL, 1); + conn = ksocknal_find_conn_locked(peer_ni, NULL, 1); if (conn != NULL) { sched = conn->ksnc_scheduler; - cfs_spin_lock_bh (&sched->kss_lock); - if (!cfs_list_empty(&conn->ksnc_tx_queue)) { - cfs_spin_unlock_bh(&sched->kss_lock); - /* there is an queued ACK, don't need keepalive */ - return 0; - } + spin_lock_bh(&sched->kss_lock); + if (!list_empty(&conn->ksnc_tx_queue)) { + spin_unlock_bh(&sched->kss_lock); + /* there is an queued ACK, don't need keepalive */ + return 0; + } - cfs_spin_unlock_bh(&sched->kss_lock); - } + spin_unlock_bh(&sched->kss_lock); + } - cfs_read_unlock(&ksocknal_data.ksnd_global_lock); + read_unlock(&ksocknal_data.ksnd_global_lock); - /* cookie = 1 is reserved for keepalive PING */ - tx = ksocknal_alloc_tx_noop(1, 1); - if (tx == NULL) { - cfs_read_lock(&ksocknal_data.ksnd_global_lock); - return -ENOMEM; - } + /* cookie = 1 is reserved for keepalive PING */ + tx = ksocknal_alloc_tx_noop(1, 1); + if (tx == NULL) { + read_lock(&ksocknal_data.ksnd_global_lock); + return -ENOMEM; + } - if (ksocknal_launch_packet(peer->ksnp_ni, tx, peer->ksnp_id) == 0) { - cfs_read_lock(&ksocknal_data.ksnd_global_lock); - return 1; - } + if (ksocknal_launch_packet(peer_ni->ksnp_ni, tx, peer_ni->ksnp_id) == 0) { + read_lock(&ksocknal_data.ksnd_global_lock); + return 1; + } - ksocknal_free_tx(tx); - cfs_read_lock(&ksocknal_data.ksnd_global_lock); + ksocknal_free_tx(tx); + read_lock(&ksocknal_data.ksnd_global_lock); - return -EIO; + return -EIO; } -void -ksocknal_check_peer_timeouts (int idx) +static void +ksocknal_check_peer_timeouts(int idx) { - cfs_list_t *peers = &ksocknal_data.ksnd_peers[idx]; - ksock_peer_t *peer; - ksock_conn_t *conn; - ksock_tx_t *tx; + struct hlist_head *peers = &ksocknal_data.ksnd_peers[idx]; + struct ksock_peer_ni *peer_ni; + struct ksock_conn *conn; + struct ksock_tx *tx; again: - /* NB. We expect to have a look at all the peers and not find any - * connections to time out, so we just use a shared lock while we - * take a look... */ - cfs_read_lock (&ksocknal_data.ksnd_global_lock); - - cfs_list_for_each_entry_typed(peer, peers, ksock_peer_t, ksnp_list) { - cfs_time_t deadline = 0; - int resid = 0; - int n = 0; - - if (ksocknal_send_keepalive_locked(peer) != 0) { - cfs_read_unlock (&ksocknal_data.ksnd_global_lock); - goto again; - } - - conn = ksocknal_find_timed_out_conn (peer); - - if (conn != NULL) { - cfs_read_unlock (&ksocknal_data.ksnd_global_lock); - - ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT); - - /* NB we won't find this one again, but we can't - * just proceed with the next peer, since we dropped - * ksnd_global_lock and it might be dead already! */ - ksocknal_conn_decref(conn); - goto again; - } - - /* we can't process stale txs right here because we're - * holding only shared lock */ - if (!cfs_list_empty (&peer->ksnp_tx_queue)) { - ksock_tx_t *tx = - cfs_list_entry (peer->ksnp_tx_queue.next, - ksock_tx_t, tx_list); - - if (cfs_time_aftereq(cfs_time_current(), - tx->tx_deadline)) { - - ksocknal_peer_addref(peer); - cfs_read_unlock (&ksocknal_data.ksnd_global_lock); - - ksocknal_flush_stale_txs(peer); - - ksocknal_peer_decref(peer); - goto again; - } - } - - if (cfs_list_empty(&peer->ksnp_zc_req_list)) - continue; - - cfs_spin_lock(&peer->ksnp_lock); - cfs_list_for_each_entry_typed(tx, &peer->ksnp_zc_req_list, - ksock_tx_t, tx_zc_list) { - if (!cfs_time_aftereq(cfs_time_current(), - tx->tx_deadline)) + /* NB. We expect to have a look at all the peers and not find any + * connections to time out, so we just use a shared lock while we + * take a look... + */ + read_lock(&ksocknal_data.ksnd_global_lock); + + hlist_for_each_entry(peer_ni, peers, ksnp_list) { + struct ksock_tx *tx_stale; + time64_t deadline = 0; + int resid = 0; + int n = 0; + + if (ksocknal_send_keepalive_locked(peer_ni) != 0) { + read_unlock(&ksocknal_data.ksnd_global_lock); + goto again; + } + + conn = ksocknal_find_timed_out_conn(peer_ni); + + if (conn != NULL) { + read_unlock(&ksocknal_data.ksnd_global_lock); + + ksocknal_close_conn_and_siblings(conn, -ETIMEDOUT); + + /* NB we won't find this one again, but we can't + * just proceed with the next peer_ni, since we dropped + * ksnd_global_lock and it might be dead already! + */ + ksocknal_conn_decref(conn); + goto again; + } + + /* we can't process stale txs right here because we're + * holding only shared lock + */ + if (!list_empty(&peer_ni->ksnp_tx_queue)) { + struct ksock_tx *tx; + + tx = list_entry(peer_ni->ksnp_tx_queue.next, + struct ksock_tx, tx_list); + if (ktime_get_seconds() >= tx->tx_deadline) { + ksocknal_peer_addref(peer_ni); + read_unlock(&ksocknal_data.ksnd_global_lock); + + ksocknal_flush_stale_txs(peer_ni); + + ksocknal_peer_decref(peer_ni); + goto again; + } + } + + if (list_empty(&peer_ni->ksnp_zc_req_list)) + continue; + + tx_stale = NULL; + spin_lock(&peer_ni->ksnp_lock); + list_for_each_entry(tx, &peer_ni->ksnp_zc_req_list, tx_zc_list) { + if (ktime_get_seconds() < tx->tx_deadline) break; /* ignore the TX if connection is being closed */ if (tx->tx_conn->ksnc_closing) continue; n++; + if (tx_stale == NULL) + tx_stale = tx; } - if (n == 0) { - cfs_spin_unlock(&peer->ksnp_lock); - continue; - } + if (tx_stale == NULL) { + spin_unlock(&peer_ni->ksnp_lock); + continue; + } - tx = cfs_list_entry(peer->ksnp_zc_req_list.next, - ksock_tx_t, tx_zc_list); - deadline = tx->tx_deadline; - resid = tx->tx_resid; - conn = tx->tx_conn; - ksocknal_conn_addref(conn); + deadline = tx_stale->tx_deadline; + resid = tx_stale->tx_resid; + conn = tx_stale->tx_conn; + ksocknal_conn_addref(conn); - cfs_spin_unlock(&peer->ksnp_lock); - cfs_read_unlock (&ksocknal_data.ksnd_global_lock); + spin_unlock(&peer_ni->ksnp_lock); + read_unlock(&ksocknal_data.ksnd_global_lock); - CERROR("Total %d stale ZC_REQs for peer %s detected; the " - "oldest(%p) timed out %ld secs ago, " - "resid: %d, wmem: %d\n", - n, libcfs_nid2str(peer->ksnp_id.nid), tx, - cfs_duration_sec(cfs_time_current() - deadline), - resid, libcfs_sock_wmem_queued(conn->ksnc_sock)); + CERROR("Total %d stale ZC_REQs for peer_ni %s detected; the " + "oldest(%p) timed out %lld secs ago, " + "resid: %d, wmem: %d\n", + n, libcfs_nid2str(peer_ni->ksnp_id.nid), tx_stale, + ktime_get_seconds() - deadline, + resid, conn->ksnc_sock->sk->sk_wmem_queued); ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT); ksocknal_conn_decref(conn); goto again; } - cfs_read_unlock (&ksocknal_data.ksnd_global_lock); + read_unlock(&ksocknal_data.ksnd_global_lock); } -int -ksocknal_reaper (void *arg) +int ksocknal_reaper(void *arg) { - cfs_waitlink_t wait; - ksock_conn_t *conn; - ksock_sched_t *sched; - cfs_list_t enomem_conns; - int nenomem_conns; - cfs_duration_t timeout; - int i; - int peer_index = 0; - cfs_time_t deadline = cfs_time_current(); + wait_queue_entry_t wait; + struct ksock_conn *conn; + struct ksock_sched *sched; + LIST_HEAD(enomem_conns); + int nenomem_conns; + time64_t timeout; + int i; + int peer_index = 0; + time64_t deadline = ktime_get_seconds(); - cfs_daemonize ("socknal_reaper"); - cfs_block_allsigs (); + init_waitqueue_entry(&wait, current); - CFS_INIT_LIST_HEAD(&enomem_conns); - cfs_waitlink_init (&wait); - - cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock); + spin_lock_bh(&ksocknal_data.ksnd_reaper_lock); while (!ksocknal_data.ksnd_shuttingdown) { + if (!list_empty(&ksocknal_data.ksnd_deathrow_conns)) { + conn = list_entry(ksocknal_data.ksnd_deathrow_conns.next, + struct ksock_conn, ksnc_list); + list_del(&conn->ksnc_list); - if (!cfs_list_empty (&ksocknal_data.ksnd_deathrow_conns)) { - conn = cfs_list_entry (ksocknal_data. \ - ksnd_deathrow_conns.next, - ksock_conn_t, ksnc_list); - cfs_list_del (&conn->ksnc_list); - - cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock); + spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock); - ksocknal_terminate_conn (conn); - ksocknal_conn_decref(conn); + ksocknal_terminate_conn(conn); + ksocknal_conn_decref(conn); - cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock); + spin_lock_bh(&ksocknal_data.ksnd_reaper_lock); continue; } - if (!cfs_list_empty (&ksocknal_data.ksnd_zombie_conns)) { - conn = cfs_list_entry (ksocknal_data.ksnd_zombie_conns.\ - next, ksock_conn_t, ksnc_list); - cfs_list_del (&conn->ksnc_list); + if (!list_empty(&ksocknal_data.ksnd_zombie_conns)) { + conn = list_entry(ksocknal_data.ksnd_zombie_conns.next, + struct ksock_conn, ksnc_list); + list_del(&conn->ksnc_list); - cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock); + spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock); - ksocknal_destroy_conn (conn); + ksocknal_destroy_conn(conn); - cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock); - continue; - } + spin_lock_bh(&ksocknal_data.ksnd_reaper_lock); + continue; + } - if (!cfs_list_empty (&ksocknal_data.ksnd_enomem_conns)) { - cfs_list_add(&enomem_conns, - &ksocknal_data.ksnd_enomem_conns); - cfs_list_del_init(&ksocknal_data.ksnd_enomem_conns); - } + list_splice_init(&ksocknal_data.ksnd_enomem_conns, + &enomem_conns); - cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock); + spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock); /* reschedule all the connections that stalled with ENOMEM... */ nenomem_conns = 0; - while (!cfs_list_empty (&enomem_conns)) { - conn = cfs_list_entry (enomem_conns.next, - ksock_conn_t, ksnc_tx_list); - cfs_list_del (&conn->ksnc_tx_list); + while (!list_empty(&enomem_conns)) { + conn = list_entry(enomem_conns.next, + struct ksock_conn, ksnc_tx_list); + list_del(&conn->ksnc_tx_list); sched = conn->ksnc_scheduler; - cfs_spin_lock_bh (&sched->kss_lock); + spin_lock_bh(&sched->kss_lock); - LASSERT (conn->ksnc_tx_scheduled); - conn->ksnc_tx_ready = 1; - cfs_list_add_tail(&conn->ksnc_tx_list, - &sched->kss_tx_conns); - cfs_waitq_signal (&sched->kss_waitq); + LASSERT(conn->ksnc_tx_scheduled); + conn->ksnc_tx_ready = 1; + list_add_tail(&conn->ksnc_tx_list, + &sched->kss_tx_conns); + wake_up(&sched->kss_waitq); - cfs_spin_unlock_bh (&sched->kss_lock); + spin_unlock_bh(&sched->kss_lock); nenomem_conns++; } - /* careful with the jiffy wrap... */ - while ((timeout = cfs_time_sub(deadline, - cfs_time_current())) <= 0) { - const int n = 4; - const int p = 1; - int chunk = ksocknal_data.ksnd_peer_hash_size; - - /* Time to check for timeouts on a few more peers: I do - * checks every 'p' seconds on a proportion of the peer - * table and I need to check every connection 'n' times - * within a timeout interval, to ensure I detect a - * timeout on any connection within (n+1)/n times the - * timeout interval. */ - - if (*ksocknal_tunables.ksnd_timeout > n * p) - chunk = (chunk * n * p) / - *ksocknal_tunables.ksnd_timeout; - if (chunk == 0) - chunk = 1; - - for (i = 0; i < chunk; i++) { - ksocknal_check_peer_timeouts (peer_index); - peer_index = (peer_index + 1) % - ksocknal_data.ksnd_peer_hash_size; - } - - deadline = cfs_time_add(deadline, cfs_time_seconds(p)); - } + /* careful with the jiffy wrap... */ + while ((timeout = deadline - ktime_get_seconds()) <= 0) { + const int n = 4; + const int p = 1; + int chunk = HASH_SIZE(ksocknal_data.ksnd_peers); + unsigned int lnd_timeout; + + /* Time to check for timeouts on a few more peers: I + * do checks every 'p' seconds on a proportion of the + * peer_ni table and I need to check every connection + * 'n' times within a timeout interval, to ensure I + * detect a timeout on any connection within (n+1)/n + * times the timeout interval. + */ + + lnd_timeout = lnet_get_lnd_timeout(); + if (lnd_timeout > n * p) + chunk = (chunk * n * p) / lnd_timeout; + if (chunk == 0) + chunk = 1; + + for (i = 0; i < chunk; i++) { + ksocknal_check_peer_timeouts(peer_index); + peer_index = (peer_index + 1) % + HASH_SIZE(ksocknal_data.ksnd_peers); + } + + deadline += p; + } if (nenomem_conns != 0) { /* Reduce my timeout if I rescheduled ENOMEM conns. @@ -2515,26 +2699,25 @@ ksocknal_reaper (void *arg) * if any go back on my enomem list. */ timeout = SOCKNAL_ENOMEM_RETRY; } - ksocknal_data.ksnd_reaper_waketime = - cfs_time_add(cfs_time_current(), timeout); + ksocknal_data.ksnd_reaper_waketime = ktime_get_seconds() + + timeout; - cfs_set_current_state (CFS_TASK_INTERRUPTIBLE); - cfs_waitq_add (&ksocknal_data.ksnd_reaper_waitq, &wait); + set_current_state(TASK_INTERRUPTIBLE); + add_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait); - if (!ksocknal_data.ksnd_shuttingdown && - cfs_list_empty (&ksocknal_data.ksnd_deathrow_conns) && - cfs_list_empty (&ksocknal_data.ksnd_zombie_conns)) - cfs_waitq_timedwait (&wait, CFS_TASK_INTERRUPTIBLE, - timeout); + if (!ksocknal_data.ksnd_shuttingdown && + list_empty(&ksocknal_data.ksnd_deathrow_conns) && + list_empty(&ksocknal_data.ksnd_zombie_conns)) + schedule_timeout(cfs_time_seconds(timeout)); - cfs_set_current_state (CFS_TASK_RUNNING); - cfs_waitq_del (&ksocknal_data.ksnd_reaper_waitq, &wait); + set_current_state(TASK_RUNNING); + remove_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait); - cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock); - } + spin_lock_bh(&ksocknal_data.ksnd_reaper_lock); + } - cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock); + spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock); - ksocknal_thread_fini (); - return (0); + ksocknal_thread_fini(); + return 0; }