X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lnet%2Fklnds%2Fsocklnd%2Fsocklnd_cb.c;h=84396529992733be57f77010a441e79e6bee8ff5;hb=d720fbaadad9343a481863ae3b04ba580541328d;hp=bd26027871acbd66316517fcfd9900383e7e6aad;hpb=439addad84514e7ff6452710e6a7f15b80d7b589;p=fs%2Flustre-release.git diff --git a/lnet/klnds/socklnd/socklnd_cb.c b/lnet/klnds/socklnd/socklnd_cb.c index bd26027..8439652 100644 --- a/lnet/klnds/socklnd/socklnd_cb.c +++ b/lnet/klnds/socklnd/socklnd_cb.c @@ -1,13 +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 (c) 2011, 2017, Intel Corporation. * - * Copyright (C) 2001, 2002 Cluster File Systems, Inc. * 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 @@ -23,689 +24,843 @@ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. */ -#include "socknal.h" - -/* - * LIB functions follow - * - */ -int -ksocknal_dist(lib_nal_t *nal, ptl_nid_t nid, unsigned long *dist) -{ - /* I would guess that if ksocknal_get_peer (nid) == NULL, - and we're not routing, then 'nid' is very distant :) */ - if (nal->libnal_ni.ni_pid.nid == nid) { - *dist = 0; - } else { - *dist = 1; - } - - return 0; -} +#include "socklnd.h" -void -ksocknal_free_ltx (ksock_ltx_t *ltx) +struct ksock_tx * +ksocknal_alloc_tx(int type, int size) { - atomic_dec(&ksocknal_data.ksnd_nactive_ltxs); - PORTAL_FREE(ltx, ltx->ltx_desc_size); -} + struct ksock_tx *tx = NULL; -int -ksocknal_send_iov (ksock_conn_t *conn, ksock_tx_t *tx) -{ - struct iovec *iov = tx->tx_iov; - int nob; - int rc; + if (type == KSOCK_MSG_NOOP) { + LASSERT(size == KSOCK_NOOP_TX_SIZE); - LASSERT (tx->tx_niov > 0); + /* searching for a noop tx in free list */ + spin_lock(&ksocknal_data.ksnd_tx_lock); - /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */ - rc = ksocknal_lib_send_iov(conn, tx); + 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); + } - if (rc <= 0) /* sent nothing? */ - return (rc); + spin_unlock(&ksocknal_data.ksnd_tx_lock); + } - nob = rc; - LASSERT (nob <= tx->tx_resid); - tx->tx_resid -= nob; + if (tx == NULL) + LIBCFS_ALLOC(tx, size); - /* "consume" iov */ - do { - LASSERT (tx->tx_niov > 0); + if (tx == NULL) + return NULL; - if (nob < iov->iov_len) { - iov->iov_base = (void *)(((unsigned long)(iov->iov_base)) + nob); - iov->iov_len -= nob; - return (rc); - } + 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; - nob -= iov->iov_len; - tx->tx_iov = ++iov; - tx->tx_niov--; - } while (nob != 0); + atomic_inc(&ksocknal_data.ksnd_nactive_txs); - return (rc); + return tx; } -int -ksocknal_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx) -{ - ptl_kiov_t *kiov = tx->tx_kiov; - int nob; - int rc; - - 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); - - if (rc <= 0) /* sent nothing? */ - return (rc); - - nob = rc; - LASSERT (nob <= tx->tx_resid); - tx->tx_resid -= nob; +struct ksock_tx * +ksocknal_alloc_tx_noop(__u64 cookie, int nonblk) +{ + struct ksock_tx *tx; - /* "consume" kiov */ - do { - LASSERT(tx->tx_nkiov > 0); + tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE); + if (tx == NULL) { + CERROR("Can't allocate noop tx desc\n"); + return NULL; + } - if (nob < kiov->kiov_len) { - kiov->kiov_offset += nob; - kiov->kiov_len -= nob; - return rc; - } + tx->tx_conn = NULL; + tx->tx_lnetmsg = NULL; + tx->tx_kiov = NULL; + tx->tx_nkiov = 0; + tx->tx_iov = tx->tx_frags.virt.iov; + tx->tx_niov = 1; + tx->tx_nonblk = nonblk; - nob -= kiov->kiov_len; - tx->tx_kiov = ++kiov; - tx->tx_nkiov--; - } while (nob != 0); + 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 (rc); + return tx; } -int -ksocknal_transmit (ksock_conn_t *conn, ksock_tx_t *tx) -{ - int rc; - int bufnob; - - if (ksocknal_data.ksnd_stall_tx != 0) { - set_current_state (TASK_UNINTERRUPTIBLE); - schedule_timeout (cfs_time_seconds(ksocknal_data.ksnd_stall_tx)); - } - LASSERT (tx->tx_resid != 0); +void +ksocknal_free_tx(struct ksock_tx *tx) +{ + atomic_dec(&ksocknal_data.ksnd_nactive_txs); - rc = ksocknal_getconnsock (conn); - if (rc != 0) { - LASSERT (conn->ksnc_closing); - return (-ESHUTDOWN); - } + 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); - 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); - } + list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs); - bufnob = 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_io_timeout); - conn->ksnc_peer->ksnp_last_alive = cfs_time_current(); - conn->ksnc_tx_bufnob = bufnob; - mb(); - } + spin_unlock(&ksocknal_data.ksnd_tx_lock); + } else { + LIBCFS_FREE(tx, tx->tx_desc_size); + } +} - if (rc <= 0) { /* Didn't write anything? */ - unsigned long flags; - ksock_sched_t *sched; +static int +ksocknal_send_iov(struct ksock_conn *conn, struct ksock_tx *tx, + struct kvec *scratch_iov) +{ + struct kvec *iov = tx->tx_iov; + int nob; + int rc; - if (rc == 0) /* some stacks return 0 instead of -EAGAIN */ - rc = -EAGAIN; + LASSERT(tx->tx_niov > 0); - if (rc != -EAGAIN) - break; + /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */ + rc = ksocknal_lib_send_iov(conn, tx, scratch_iov); - /* Check if EAGAIN is due to memory pressure */ + if (rc <= 0) /* sent nothing? */ + return rc; - sched = conn->ksnc_scheduler; - spin_lock_irqsave(&sched->kss_lock, flags); - - if (!SOCK_TEST_NOSPACE(conn->ksnc_sock) && - !conn->ksnc_tx_ready) { - /* SOCK_NOSPACE is set when the socket fills - * and cleared in the write_space callback - * (which also sets ksnc_tx_ready). If - * SOCK_NOSPACE and ksnc_tx_ready are BOTH - * zero, I didn't fill the socket and - * write_space won't reschedule me, so I - * return -ENOMEM to get my caller to retry - * after a timeout */ - rc = -ENOMEM; - } + nob = rc; + LASSERT(nob <= tx->tx_resid); + tx->tx_resid -= nob; - spin_unlock_irqrestore(&sched->kss_lock, flags); - break; - } + /* "consume" iov */ + do { + LASSERT(tx->tx_niov > 0); - /* socket's wmem_queued now includes 'rc' bytes */ - atomic_sub (rc, &conn->ksnc_tx_nob); - rc = 0; + if (nob < (int) iov->iov_len) { + iov->iov_base += nob; + iov->iov_len -= nob; + return rc; + } - } while (tx->tx_resid != 0); + nob -= iov->iov_len; + tx->tx_iov = ++iov; + tx->tx_niov--; + } while (nob != 0); - ksocknal_putconnsock (conn); - return (rc); + return rc; } -int -ksocknal_recv_iov (ksock_conn_t *conn) -{ - struct iovec *iov = conn->ksnc_rx_iov; - int nob; - int rc; - - LASSERT (conn->ksnc_rx_niov > 0); +static int +ksocknal_send_kiov(struct ksock_conn *conn, struct ksock_tx *tx, + struct kvec *scratch_iov) +{ + struct bio_vec *kiov = tx->tx_kiov; + int nob; + int rc; - /* Never touch conn->ksnc_rx_iov or change connection - * status inside ksocknal_lib_recv_iov */ - rc = ksocknal_lib_recv_iov(conn); + LASSERT(tx->tx_niov == 0); + LASSERT(tx->tx_nkiov > 0); - if (rc <= 0) - return (rc); + /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */ + rc = ksocknal_lib_send_kiov(conn, tx, scratch_iov); - /* received something... */ - nob = rc; + if (rc <= 0) /* sent nothing? */ + return rc; - conn->ksnc_peer->ksnp_last_alive = cfs_time_current(); - conn->ksnc_rx_deadline = cfs_time_shift (ksocknal_tunables.ksnd_io_timeout); - mb(); /* order with setting rx_started */ - conn->ksnc_rx_started = 1; + nob = rc; + LASSERT(nob <= tx->tx_resid); + tx->tx_resid -= nob; - conn->ksnc_rx_nob_wanted -= nob; - conn->ksnc_rx_nob_left -= nob; + /* "consume" kiov */ + do { + LASSERT(tx->tx_nkiov > 0); - do { - LASSERT (conn->ksnc_rx_niov > 0); + if (nob < (int)kiov->bv_len) { + kiov->bv_offset += nob; + kiov->bv_len -= nob; + return rc; + } - if (nob < iov->iov_len) { - iov->iov_len -= nob; - iov->iov_base = (void *)(((unsigned long)iov->iov_base) + nob); - return (-EAGAIN); - } + nob -= (int)kiov->bv_len; + tx->tx_kiov = ++kiov; + tx->tx_nkiov--; + } while (nob != 0); - nob -= iov->iov_len; - conn->ksnc_rx_iov = ++iov; - conn->ksnc_rx_niov--; - } while (nob != 0); + return rc; +} - return (rc); +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) + 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_kiov (ksock_conn_t *conn) +static int +ksocknal_recv_iov(struct ksock_conn *conn, struct kvec *scratchiov) { - ptl_kiov_t *kiov = conn->ksnc_rx_kiov; - int nob; - int rc; - LASSERT (conn->ksnc_rx_nkiov > 0); + struct kvec *iov = conn->ksnc_rx_iov; + int nob; + int rc; - /* Never touch conn->ksnc_rx_kiov or change connection - * status inside ksocknal_lib_recv_iov */ - rc = ksocknal_lib_recv_kiov(conn); + LASSERT(conn->ksnc_rx_niov > 0); - if (rc <= 0) - return (rc); + /* Never touch conn->ksnc_rx_iov or change connection + * status inside ksocknal_lib_recv_iov */ + rc = ksocknal_lib_recv_iov(conn, scratchiov); - /* received something... */ - nob = rc; + if (rc <= 0) + return rc; - conn->ksnc_peer->ksnp_last_alive = cfs_time_current(); - conn->ksnc_rx_deadline = cfs_time_shift (ksocknal_tunables.ksnd_io_timeout); - mb(); /* order with setting rx_started */ - conn->ksnc_rx_started = 1; + /* received something... */ + nob = rc; - conn->ksnc_rx_nob_wanted -= nob; - conn->ksnc_rx_nob_left -= nob; + 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; - do { - LASSERT (conn->ksnc_rx_nkiov > 0); + conn->ksnc_rx_nob_wanted -= nob; + conn->ksnc_rx_nob_left -= nob; - if (nob < kiov->kiov_len) { - kiov->kiov_offset += nob; - kiov->kiov_len -= nob; - return -EAGAIN; - } + do { + LASSERT(conn->ksnc_rx_niov > 0); - nob -= kiov->kiov_len; - conn->ksnc_rx_kiov = ++kiov; - conn->ksnc_rx_nkiov--; - } while (nob != 0); + if (nob < (int)iov->iov_len) { + iov->iov_len -= nob; + iov->iov_base += nob; + return -EAGAIN; + } - return 1; + nob -= iov->iov_len; + conn->ksnc_rx_iov = ++iov; + conn->ksnc_rx_niov--; + } while (nob != 0); + + return rc; } -int -ksocknal_receive (ksock_conn_t *conn) +static int +ksocknal_recv_kiov(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; + struct bio_vec *kiov = conn->ksnc_rx_kiov; + int nob; + int rc; + LASSERT(conn->ksnc_rx_nkiov > 0); - if (ksocknal_data.ksnd_stall_rx != 0) { - set_current_state (TASK_UNINTERRUPTIBLE); - schedule_timeout(cfs_time_seconds (ksocknal_data.ksnd_stall_rx)); - } - - rc = ksocknal_getconnsock (conn); - if (rc != 0) { - LASSERT (conn->ksnc_closing); - return (-ESHUTDOWN); - } + /* 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); - 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; - } + if (rc <= 0) + return rc; - /* Completed a fragment */ + /* received something... */ + nob = rc; - if (conn->ksnc_rx_nob_wanted == 0) { - /* Completed a message segment (header or payload) */ - if ((ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0 && - (conn->ksnc_rx_state == SOCKNAL_RX_BODY || - conn->ksnc_rx_state == SOCKNAL_RX_BODY_FWD)) { - /* Remind the socket to ack eagerly... */ - ksocknal_lib_eager_ack(conn); - } - rc = 1; - break; - } - } + 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; - ksocknal_putconnsock (conn); - RETURN (rc); -} + conn->ksnc_rx_nob_wanted -= nob; + conn->ksnc_rx_nob_left -= nob; -#if SOCKNAL_ZC -void -ksocknal_zc_callback (zccd_t *zcd) -{ - ksock_tx_t *tx = KSOCK_ZCCD_2_TX(zcd); - ksock_sched_t *sched = tx->tx_conn->ksnc_scheduler; - unsigned long flags; - ENTRY; + do { + LASSERT(conn->ksnc_rx_nkiov > 0); - /* Schedule tx for cleanup (can't do it now due to lock conflicts) */ + if (nob < (int) kiov->bv_len) { + kiov->bv_offset += nob; + kiov->bv_len -= nob; + return -EAGAIN; + } - spin_lock_irqsave (&sched->kss_lock, flags); + nob -= kiov->bv_len; + conn->ksnc_rx_kiov = ++kiov; + conn->ksnc_rx_nkiov--; + } while (nob != 0); - list_add_tail (&tx->tx_list, &sched->kss_zctxdone_list); - cfs_waitq_signal (&sched->kss_waitq); + return 1; +} - spin_unlock_irqrestore (&sched->kss_lock, flags); - EXIT; +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) + 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); } -#endif void -ksocknal_tx_done (ksock_tx_t *tx, int asynch) +ksocknal_tx_done(struct lnet_ni *ni, struct ksock_tx *tx, int rc) { - ksock_ltx_t *ltx; + struct lnet_msg *lnetmsg = tx->tx_lnetmsg; + enum lnet_msg_hstatus hstatus = tx->tx_hstatus; ENTRY; - if (tx->tx_conn != NULL) { -#if SOCKNAL_ZC - /* zero copy completion isn't always from - * process_transmit() so it needs to keep a ref on - * tx_conn... */ - if (asynch) - ksocknal_put_conn (tx->tx_conn); -#else - LASSERT (!asynch); -#endif - } + LASSERT(ni != NULL || tx->tx_conn != NULL); - if (tx->tx_isfwd) { /* was a forwarded packet? */ - kpr_fwd_done (&ksocknal_data.ksnd_router, - KSOCK_TX_2_KPR_FWD_DESC (tx), - (tx->tx_resid == 0) ? 0 : -ECONNABORTED); - EXIT; - return; - } + if (!rc && (tx->tx_resid != 0 || tx->tx_zc_aborted)) { + rc = -EIO; + if (hstatus == LNET_MSG_STATUS_OK) + hstatus = LNET_MSG_STATUS_LOCAL_ERROR; + } - /* local send */ - ltx = KSOCK_TX_2_KSOCK_LTX (tx); + if (tx->tx_conn != NULL) + ksocknal_conn_decref(tx->tx_conn); - lib_finalize (&ksocknal_lib, ltx->ltx_private, ltx->ltx_cookie, - (tx->tx_resid == 0) ? PTL_OK : PTL_FAIL); + ksocknal_free_tx(tx); + if (lnetmsg != NULL) { /* KSOCK_MSG_NOOP go without lnetmsg */ + lnetmsg->msg_health_status = hstatus; + lnet_finalize(lnetmsg, rc); + } - ksocknal_free_ltx (ltx); - EXIT; + EXIT; } void -ksocknal_tx_launched (ksock_tx_t *tx) +ksocknal_txlist_done(struct lnet_ni *ni, struct list_head *txlist, int error) { -#if SOCKNAL_ZC - if (atomic_read (&tx->tx_zccd.zccd_count) != 1) { - ksock_conn_t *conn = tx->tx_conn; - - /* zccd skbufs are still in-flight. First take a ref on - * conn, so it hangs about for ksocknal_tx_done... */ - atomic_inc (&conn->ksnc_refcount); - - /* ...then drop the initial ref on zccd, so the zero copy - * callback can occur */ - zccd_put (&tx->tx_zccd); - return; - } -#endif - /* Any zero-copy-ness (if any) has completed; I can complete the - * transmit now, avoiding an extra schedule */ - ksocknal_tx_done (tx, 0); + 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); + } } -int -ksocknal_process_transmit (ksock_conn_t *conn, ksock_tx_t *tx) +static void +ksocknal_check_zc_req(struct ksock_tx *tx) { - unsigned long flags; - int rc; + 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_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. */ + LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP); + LASSERT (tx->tx_zc_capable); + + tx->tx_zc_checked = 1; + + if (conn->ksnc_proto == &ksocknal_protocol_v1x || + !conn->ksnc_zc_capable) + return; - rc = ksocknal_transmit (conn, tx); + /* assign cookie and queue tx to pending list, it will be released when + * a matching ack is received. See ksocknal_handle_zcack() */ - CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc); + ksocknal_tx_addref(tx); - if (tx->tx_resid == 0) { - /* Sent everything OK */ - LASSERT (rc == 0); + spin_lock(&peer_ni->ksnp_lock); - ksocknal_tx_launched (tx); - return (0); - } + /* ZC_REQ is going to be pinned to the peer_ni */ + tx->tx_deadline = ktime_get_seconds() + + lnet_get_lnd_timeout(); - if (rc == -EAGAIN) - return (rc); + LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0); - if (rc == -ENOMEM) { - static int counter; + tx->tx_msg.ksm_zc_cookies[0] = peer_ni->ksnp_zc_next_cookie++; - counter++; /* exponential backoff warnings */ - if ((counter & (-counter)) == counter) - CWARN("%d ENOMEM tx %p (%u allocated)\n", - counter, conn, atomic_read(&portal_kmemory)); + if (peer_ni->ksnp_zc_next_cookie == 0) + peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1; - /* Queue on ksnd_enomem_conns for retry after a timeout */ - spin_lock_irqsave(&ksocknal_data.ksnd_reaper_lock, flags); + list_add_tail(&tx->tx_zc_list, &peer_ni->ksnp_zc_req_list); - /* 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 (!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); + spin_unlock(&peer_ni->ksnp_lock); +} - spin_unlock_irqrestore(&ksocknal_data.ksnd_reaper_lock, flags); - return (rc); - } +static void +ksocknal_uncheck_zc_req(struct ksock_tx *tx) +{ + struct ksock_peer_ni *peer_ni = tx->tx_conn->ksnc_peer; - /* Actual error */ - LASSERT (rc < 0); + LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP); + LASSERT(tx->tx_zc_capable); - 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_HA, "[%p] Error %d on write to "LPX64 - " ip %d.%d.%d.%d:%d\n", conn, rc, - conn->ksnc_peer->ksnp_nid, - HIPQUAD(conn->ksnc_ipaddr), - conn->ksnc_port); - } + tx->tx_zc_checked = 0; + + spin_lock(&peer_ni->ksnp_lock); + + 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; + list_del(&tx->tx_zc_list); - ksocknal_close_conn_and_siblings (conn, rc); - ksocknal_tx_launched (tx); + spin_unlock(&peer_ni->ksnp_lock); - return (rc); + ksocknal_tx_decref(tx); } -void -ksocknal_launch_autoconnect_locked (ksock_route_t *route) +static int +ksocknal_process_transmit(struct ksock_conn *conn, struct ksock_tx *tx, + struct kvec *scratch_iov) +{ + 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; +} + +static void +ksocknal_launch_connection_locked(struct ksock_route *route) { - unsigned long flags; /* called holding write lock on ksnd_global_lock */ + + LASSERT (!route->ksnr_scheduled); LASSERT (!route->ksnr_connecting); + LASSERT ((ksocknal_route_mask() & ~route->ksnr_connected) != 0); - route->ksnr_connecting = 1; /* scheduling conn for autoconnectd */ - atomic_inc (&route->ksnr_refcount); /* extra ref for autoconnectd */ + route->ksnr_scheduled = 1; /* scheduling conn for connd */ + ksocknal_route_addref(route); /* extra ref for connd */ - spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags); + spin_lock_bh(&ksocknal_data.ksnd_connd_lock); - list_add_tail (&route->ksnr_connect_list, - &ksocknal_data.ksnd_autoconnectd_routes); - cfs_waitq_signal (&ksocknal_data.ksnd_autoconnectd_waitq); + list_add_tail(&route->ksnr_connd_list, + &ksocknal_data.ksnd_connd_routes); + wake_up(&ksocknal_data.ksnd_connd_waitq); - spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags); + spin_unlock_bh(&ksocknal_data.ksnd_connd_lock); } -ksock_peer_t * -ksocknal_find_target_peer_locked (ksock_tx_t *tx, ptl_nid_t nid) +void +ksocknal_launch_all_connections_locked(struct ksock_peer_ni *peer_ni) { - char ipbuf[PTL_NALFMT_SIZE]; - ptl_nid_t target_nid; - int rc; - ksock_peer_t *peer = ksocknal_find_peer_locked (nid); - - if (peer != NULL) - return (peer); - - if (tx->tx_isfwd) { - CERROR ("Can't send packet to "LPX64 - " %s: routed target is not a peer\n", - nid, portals_nid2str(SOCKNAL, nid, ipbuf)); - return (NULL); - } - - rc = kpr_lookup (&ksocknal_data.ksnd_router, nid, tx->tx_nob, - &target_nid); - if (rc != 0) { - CERROR ("Can't route to "LPX64" %s: router error %d\n", - nid, portals_nid2str(SOCKNAL, nid, ipbuf), rc); - return (NULL); - } + struct ksock_route *route; - peer = ksocknal_find_peer_locked (target_nid); - if (peer != NULL) - return (peer); + /* called holding write lock on ksnd_global_lock */ + for (;;) { + /* launch any/all connections that need it */ + route = ksocknal_find_connectable_route_locked(peer_ni); + if (route == NULL) + return; - CERROR ("Can't send packet to "LPX64" %s: no peer entry\n", - target_nid, portals_nid2str(SOCKNAL, target_nid, ipbuf)); - return (NULL); + ksocknal_launch_connection_locked(route); + } } -ksock_conn_t * -ksocknal_find_conn_locked (ksock_tx_t *tx, ksock_peer_t *peer) +struct ksock_conn * +ksocknal_find_conn_locked(struct ksock_peer_ni *peer_ni, struct ksock_tx *tx, int nonblk) { - struct list_head *tmp; - ksock_conn_t *typed = NULL; - int tnob = 0; - ksock_conn_t *fallback = NULL; - int fnob = 0; - ksock_conn_t *conn; - - list_for_each (tmp, &peer->ksnp_conns) { - ksock_conn_t *c = list_entry(tmp, ksock_conn_t, ksnc_list); -#if SOCKNAL_ROUND_ROBIN - const int nob = 0; -#else - int nob = atomic_read(&c->ksnc_tx_nob) + - SOCK_WMEM_QUEUED(c->ksnc_sock); -#endif - LASSERT (!c->ksnc_closing); + 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; - if (fallback == NULL || nob < fnob) { - fallback = c; - fnob = nob; - } + LASSERT (!c->ksnc_closing); + LASSERT (c->ksnc_proto != NULL && + c->ksnc_proto->pro_match_tx != NULL); - if (!ksocknal_tunables.ksnd_typed_conns) - continue; + rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk); - switch (c->ksnc_type) { + switch (rc) { default: - CERROR("ksnc_type bad: %u\n", c->ksnc_type); LBUG(); - case SOCKNAL_CONN_ANY: - break; - case SOCKNAL_CONN_BULK_IN: + case SOCKNAL_MATCH_NO: /* protocol rejected the tx */ continue; - case SOCKNAL_CONN_BULK_OUT: - if (tx->tx_nob < ksocknal_tunables.ksnd_min_bulk) - continue; - break; - case SOCKNAL_CONN_CONTROL: - if (tx->tx_nob >= ksocknal_tunables.ksnd_min_bulk) - continue; + + case SOCKNAL_MATCH_YES: /* typed connection */ + if (typed == NULL || tnob > nob || + (tnob == nob && *ksocknal_tunables.ksnd_round_robin && + typed->ksnc_tx_last_post > c->ksnc_tx_last_post)) { + typed = c; + tnob = nob; + } break; - } - if (typed == NULL || nob < tnob) { - typed = c; - tnob = nob; + case SOCKNAL_MATCH_MAY: /* fallback connection */ + if (fallback == NULL || fnob > nob || + (fnob == nob && *ksocknal_tunables.ksnd_round_robin && + fallback->ksnc_tx_last_post > c->ksnc_tx_last_post)) { + fallback = c; + fnob = nob; + } + break; } } /* prefer the typed selection */ conn = (typed != NULL) ? typed : fallback; -#if SOCKNAL_ROUND_ROBIN - if (conn != NULL) { - /* round-robin all else being equal */ - list_del (&conn->ksnc_list); - list_add_tail (&conn->ksnc_list, &peer->ksnp_conns); - } -#endif + if (conn != NULL) + conn->ksnc_tx_last_post = ktime_get_seconds(); + return conn; } void -ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn) +ksocknal_tx_prep(struct ksock_conn *conn, struct ksock_tx *tx) +{ + conn->ksnc_proto->pro_pack(tx); + + 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(struct ksock_tx *tx, struct ksock_conn *conn) { - unsigned long flags; - ksock_sched_t *sched = conn->ksnc_scheduler; + 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, * so we don't need the {get,put}connsock dance to deref * ksnc_sock... */ LASSERT(!conn->ksnc_closing); - LASSERT(tx->tx_resid == tx->tx_nob); - CDEBUG (D_NET, "Sending to "LPX64" ip %d.%d.%d.%d:%d\n", - conn->ksnc_peer->ksnp_nid, - 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); - atomic_add (tx->tx_nob, &conn->ksnc_tx_nob); - tx->tx_conn = conn; + ksocknal_tx_prep(conn, tx); -#if SOCKNAL_ZC - zccd_init (&tx->tx_zccd, ksocknal_zc_callback); - /* NB this sets 1 ref on zccd, so the callback can only occur after - * I've released this ref. */ -#endif - spin_lock_irqsave (&sched->kss_lock, flags); - - if (list_empty(&conn->ksnc_tx_queue) && - SOCK_WMEM_QUEUED(conn->ksnc_sock) == 0) { - /* First packet starts the timeout */ - conn->ksnc_tx_deadline = cfs_time_shift(ksocknal_tunables.ksnd_io_timeout); - conn->ksnc_tx_bufnob = 0; - mb(); /* order with adding to tx_queue */ - } + /* 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. + * + * We always expect at least 1 mapped fragment containing the + * complete ksocknal message header. */ + LASSERT (lnet_iov_nob (tx->tx_niov, tx->tx_iov) + + lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) == + (unsigned int)tx->tx_nob); + LASSERT (tx->tx_niov >= 1); + LASSERT (tx->tx_resid == tx->tx_nob); - list_add_tail (&tx->tx_list, &conn->ksnc_tx_queue); + CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n", + tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type: + KSOCK_MSG_NOOP, + tx->tx_nob, tx->tx_niov, tx->tx_nkiov); + + 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); + + if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0)) + ztx = tx; /* ZC ACK piggybacked on ztx release tx later */ + + } else { + /* It's a normal packet - can it piggback a noop zc-ack that + * has been queued already? */ + LASSERT (msg->ksm_zc_cookies[1] == 0); + LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL); + + ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx); + /* ztx will be released later */ + } - if (conn->ksnc_tx_ready && /* able to send */ - !conn->ksnc_tx_scheduled) { /* not scheduled to send */ - /* +1 ref for scheduler */ - atomic_inc (&conn->ksnc_refcount); - list_add_tail (&conn->ksnc_tx_list, - &sched->kss_tx_conns); - conn->ksnc_tx_scheduled = 1; - cfs_waitq_signal (&sched->kss_waitq); + if (ztx != NULL) { + atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob); + list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs); } - spin_unlock_irqrestore (&sched->kss_lock, flags); + 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) { - struct list_head *tmp; - ksock_route_t *route; - int bits; + time64_t now = ktime_get_seconds(); + struct list_head *tmp; + struct ksock_route *route; - list_for_each (tmp, &peer->ksnp_routes) { - route = list_entry (tmp, ksock_route_t, ksnr_list); - bits = route->ksnr_connected; + list_for_each(tmp, &peer_ni->ksnp_routes) { + route = list_entry(tmp, struct ksock_route, ksnr_list); - if (ksocknal_tunables.ksnd_typed_conns) { - /* All typed connections established? */ - if ((bits & KSNR_TYPED_ROUTES) == KSNR_TYPED_ROUTES) - continue; - } else { - /* Untyped connection established? */ - if ((bits & (1 << SOCKNAL_CONN_ANY)) != 0) - continue; - } + LASSERT (!route->ksnr_connecting || route->ksnr_scheduled); + + if (route->ksnr_scheduled) /* connections being established */ + continue; - /* connection being established? */ - if (route->ksnr_connecting) + /* all route types connected ? */ + if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0) continue; - /* too soon to retry this guy? */ - if (!cfs_time_aftereq (cfs_time_current(), route->ksnr_timeout)) + if (!(route->ksnr_retry_interval == 0 || /* first attempt */ + 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); } @@ -713,16 +868,18 @@ 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) { - struct list_head *tmp; - ksock_route_t *route; + struct list_head *tmp; + struct ksock_route *route; - list_for_each (tmp, &peer->ksnp_routes) { - route = 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); - if (route->ksnr_connecting) + LASSERT (!route->ksnr_connecting || route->ksnr_scheduled); + + if (route->ksnr_scheduled) return (route); } @@ -730,679 +887,491 @@ ksocknal_find_connecting_route_locked (ksock_peer_t *peer) } int -ksocknal_launch_packet (ksock_tx_t *tx, ptl_nid_t nid) +ksocknal_launch_packet(struct lnet_ni *ni, struct ksock_tx *tx, + struct lnet_process_id id) { - unsigned long flags; - ksock_peer_t *peer; - ksock_conn_t *conn; - ksock_route_t *route; - rwlock_t *g_lock; - - /* 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. - * - * We always expect at least 1 mapped fragment containing the - * complete portals header. */ - LASSERT (lib_iov_nob (tx->tx_niov, tx->tx_iov) + - lib_kiov_nob (tx->tx_nkiov, tx->tx_kiov) == tx->tx_nob); - LASSERT (tx->tx_niov >= 1); - LASSERT (tx->tx_iov[0].iov_len >= sizeof (ptl_hdr_t)); - - CDEBUG (D_NET, "packet %p type %d, nob %d niov %d nkiov %d\n", - tx, ((ptl_hdr_t *)tx->tx_iov[0].iov_base)->type, - tx->tx_nob, tx->tx_niov, tx->tx_nkiov); + struct ksock_peer_ni *peer_ni; + struct ksock_conn *conn; + rwlock_t *g_lock; + int retry; + int rc; - tx->tx_conn = NULL; /* only set when assigned a conn */ - tx->tx_resid = tx->tx_nob; - tx->tx_hdr = (ptl_hdr_t *)tx->tx_iov[0].iov_base; + LASSERT (tx->tx_conn == NULL); g_lock = &ksocknal_data.ksnd_global_lock; -#if !SOCKNAL_ROUND_ROBIN - read_lock (g_lock); - - peer = ksocknal_find_target_peer_locked (tx, nid); - if (peer == NULL) { - read_unlock (g_lock); - return (-EHOSTUNREACH); - } - if (ksocknal_find_connectable_route_locked(peer) == NULL) { - conn = ksocknal_find_conn_locked (tx, peer); - if (conn != NULL) { - /* I've got no autoconnect routes that need to be - * connecting and I do have an actual connection... */ - ksocknal_queue_tx_locked (tx, conn); - read_unlock (g_lock); - return (0); + for (retry = 0;; retry = 1) { + 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); + read_unlock(g_lock); + return (0); + } + } } - } - /* I'll need a write lock... */ - read_unlock (g_lock); -#endif - write_lock_irqsave(g_lock, flags); + /* I'll need a write lock... */ + read_unlock(g_lock); - peer = ksocknal_find_target_peer_locked (tx, nid); - if (peer == NULL) { - write_unlock_irqrestore(g_lock, flags); - return (-EHOSTUNREACH); - } + write_lock_bh(g_lock); - for (;;) { - /* launch any/all autoconnections that need it */ - route = ksocknal_find_connectable_route_locked (peer); - if (route == NULL) + peer_ni = ksocknal_find_peer_locked(ni, id); + if (peer_ni != NULL) break; - ksocknal_launch_autoconnect_locked (route); + write_unlock_bh(g_lock); + + if ((id.pid & LNET_PID_USERFLAG) != 0) { + CERROR("Refusing to create a connection to " + "userspace process %s\n", libcfs_id2str(id)); + return -EHOSTUNREACH; + } + + if (retry) { + CERROR("Can't find peer_ni %s\n", libcfs_id2str(id)); + return -EHOSTUNREACH; + } + + rc = ksocknal_add_peer(ni, id, + LNET_NIDADDR(id.nid), + lnet_acceptor_port()); + if (rc != 0) { + CERROR("Can't add peer_ni %s: %d\n", + libcfs_id2str(id), rc); + return rc; + } } - conn = ksocknal_find_conn_locked (tx, peer); + ksocknal_launch_all_connections_locked(peer_ni); + + 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); - write_unlock_irqrestore (g_lock, flags); + write_unlock_bh(g_lock); return (0); } - route = ksocknal_find_connecting_route_locked (peer); - if (route != NULL) { - /* At least 1 connection is being established; queue the - * message... */ - list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue); - write_unlock_irqrestore (g_lock, flags); - return (0); + 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 */ + list_add_tail(&tx->tx_list, &peer_ni->ksnp_tx_queue); + write_unlock_bh(g_lock); + return 0; } - write_unlock_irqrestore (g_lock, flags); + write_unlock_bh(g_lock); + + /* NB Routes may be ignored if connections to them failed recently */ + CNETERR("No usable routes to %s\n", libcfs_id2str(id)); + tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_ERROR; return (-EHOSTUNREACH); } -ptl_err_t -ksocknal_sendmsg(lib_nal_t *nal, - void *private, - lib_msg_t *cookie, - ptl_hdr_t *hdr, - int type, - ptl_nid_t nid, - ptl_pid_t pid, - unsigned int payload_niov, - struct iovec *payload_iov, - ptl_kiov_t *payload_kiov, - size_t payload_offset, - size_t payload_nob) +int +ksocknal_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg) { - ksock_ltx_t *ltx; - 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... */ - CDEBUG(D_NET, "sending "LPSZ" bytes in %d frags to nid:"LPX64 - " pid %d\n", payload_nob, payload_niov, nid , pid); - - LASSERT (payload_nob == 0 || payload_niov > 0); - LASSERT (payload_niov <= PTL_MD_MAX_IOV); + CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n", + payload_nob, payload_niov, libcfs_id2str(target)); - /* It must be OK to kmap() if required */ - LASSERT (payload_kiov == NULL || !in_interrupt ()); - /* payload is either all vaddrs or all pages */ - LASSERT (!(payload_kiov != NULL && payload_iov != NULL)); + 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(ksock_ltx_t, ltx_iov[1 + payload_niov]); - else - desc_size = offsetof(ksock_ltx_t, ltx_kiov[payload_niov]); + 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 (in_interrupt() || - type == PTL_MSG_ACK || - type == PTL_MSG_REPLY) { - /* Can't block if in interrupt or responding to an incoming - * message */ - PORTAL_ALLOC_ATOMIC(ltx, desc_size); - } else { - PORTAL_ALLOC(ltx, desc_size); - } - - if (ltx == NULL) { - CERROR("Can't allocate tx desc type %d size %d %s\n", - type, desc_size, in_interrupt() ? "(intr)" : ""); - return (PTL_NO_SPACE); + 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); } - atomic_inc(&ksocknal_data.ksnd_nactive_ltxs); - - ltx->ltx_desc_size = desc_size; - - /* We always have 1 mapped frag for the header */ - ltx->ltx_tx.tx_iov = ltx->ltx_iov; - ltx->ltx_iov[0].iov_base = <x->ltx_hdr; - ltx->ltx_iov[0].iov_len = sizeof(*hdr); - ltx->ltx_hdr = *hdr; - - ltx->ltx_private = private; - ltx->ltx_cookie = cookie; - - ltx->ltx_tx.tx_isfwd = 0; - ltx->ltx_tx.tx_nob = sizeof (*hdr) + payload_nob; + tx->tx_conn = NULL; /* set when assigned a conn */ + tx->tx_lnetmsg = lntmsg; if (payload_iov != NULL) { - /* payload is all mapped */ - ltx->ltx_tx.tx_kiov = NULL; - ltx->ltx_tx.tx_nkiov = 0; - - ltx->ltx_tx.tx_niov = - 1 + lib_extract_iov(payload_niov, <x->ltx_iov[1], - payload_niov, payload_iov, - payload_offset, payload_nob); + tx->tx_kiov = NULL; + tx->tx_nkiov = 0; + tx->tx_iov = tx->tx_frags.virt.iov; + tx->tx_niov = 1 + + lnet_extract_iov(payload_niov, &tx->tx_iov[1], + payload_niov, payload_iov, + payload_offset, payload_nob); } else { - /* payload is all pages */ - ltx->ltx_tx.tx_niov = 1; - - ltx->ltx_tx.tx_kiov = ltx->ltx_kiov; - ltx->ltx_tx.tx_nkiov = - lib_extract_kiov(payload_niov, ltx->ltx_kiov, - payload_niov, payload_kiov, - payload_offset, payload_nob); - } + tx->tx_niov = 1; + tx->tx_iov = &tx->tx_frags.paged.iov; + tx->tx_kiov = tx->tx_frags.paged.kiov; + tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov, + payload_niov, payload_kiov, + payload_offset, payload_nob); - rc = ksocknal_launch_packet(<x->ltx_tx, nid); - if (rc == 0) - return (PTL_OK); + if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload) + tx->tx_zc_capable = 1; + } - ksocknal_free_ltx(ltx); - return (PTL_FAIL); -} + 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; -ptl_err_t -ksocknal_send (lib_nal_t *nal, void *private, lib_msg_t *cookie, - ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid, - unsigned int payload_niov, struct iovec *payload_iov, - size_t payload_offset, size_t payload_len) -{ - return (ksocknal_sendmsg(nal, private, cookie, - hdr, type, nid, pid, - payload_niov, payload_iov, NULL, - payload_offset, payload_len)); -} + /* The first fragment will be set later in pro_pack */ + rc = ksocknal_launch_packet(ni, tx, target); + if (!mpflag) + cfs_memory_pressure_restore(mpflag); -ptl_err_t -ksocknal_send_pages (lib_nal_t *nal, void *private, lib_msg_t *cookie, - ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid, - unsigned int payload_niov, ptl_kiov_t *payload_kiov, - size_t payload_offset, size_t payload_len) -{ - return (ksocknal_sendmsg(nal, private, cookie, - hdr, type, nid, pid, - payload_niov, NULL, payload_kiov, - payload_offset, payload_len)); -} + if (rc == 0) + return (0); -void -ksocknal_fwd_packet (void *arg, kpr_fwd_desc_t *fwd) -{ - ptl_nid_t nid = fwd->kprfd_gateway_nid; - ksock_ftx_t *ftx = (ksock_ftx_t *)&fwd->kprfd_scratch; - int rc; - - CDEBUG (D_NET, "Forwarding [%p] -> "LPX64" ("LPX64"))\n", fwd, - fwd->kprfd_gateway_nid, fwd->kprfd_target_nid); - - /* I'm the gateway; must be the last hop */ - if (nid == ksocknal_lib.libnal_ni.ni_pid.nid) - nid = fwd->kprfd_target_nid; - - /* setup iov for hdr */ - ftx->ftx_iov.iov_base = fwd->kprfd_hdr; - ftx->ftx_iov.iov_len = sizeof(ptl_hdr_t); - - ftx->ftx_tx.tx_isfwd = 1; /* This is a forwarding packet */ - ftx->ftx_tx.tx_nob = sizeof(ptl_hdr_t) + fwd->kprfd_nob; - ftx->ftx_tx.tx_niov = 1; - ftx->ftx_tx.tx_iov = &ftx->ftx_iov; - ftx->ftx_tx.tx_nkiov = fwd->kprfd_niov; - ftx->ftx_tx.tx_kiov = fwd->kprfd_kiov; - - rc = ksocknal_launch_packet (&ftx->ftx_tx, nid); - if (rc != 0) - kpr_fwd_done (&ksocknal_data.ksnd_router, fwd, rc); + 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); - unsigned long flags; + struct task_struct *task = kthread_run(fn, arg, name); - if (pid < 0) - return ((int)pid); + if (IS_ERR(task)) + return PTR_ERR(task); - write_lock_irqsave(&ksocknal_data.ksnd_global_lock, flags); - ksocknal_data.ksnd_nthreads++; - write_unlock_irqrestore(&ksocknal_data.ksnd_global_lock, flags); - 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) { - unsigned long flags; - - write_lock_irqsave(&ksocknal_data.ksnd_global_lock, flags); - ksocknal_data.ksnd_nthreads--; - write_unlock_irqrestore(&ksocknal_data.ksnd_global_lock, flags); + 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); } -void -ksocknal_fmb_callback (void *arg, int error) +int +ksocknal_new_packet(struct ksock_conn *conn, int nob_to_skip) { - ksock_fmb_t *fmb = (ksock_fmb_t *)arg; - ksock_fmb_pool_t *fmp = fmb->fmb_pool; - ptl_hdr_t *hdr = &fmb->fmb_hdr; - ksock_conn_t *conn = NULL; - ksock_sched_t *sched; - unsigned long flags; - char ipbuf[PTL_NALFMT_SIZE]; - char ipbuf2[PTL_NALFMT_SIZE]; - - if (error != 0) - CERROR("Failed to route packet from " - LPX64" %s to "LPX64" %s: %d\n", - le64_to_cpu(hdr->src_nid), - portals_nid2str(SOCKNAL, le64_to_cpu(hdr->src_nid), ipbuf), - le64_to_cpu(hdr->dest_nid), - portals_nid2str(SOCKNAL, le64_to_cpu(hdr->dest_nid), ipbuf2), - error); - else - CDEBUG (D_NET, "routed packet from "LPX64" to "LPX64": OK\n", - le64_to_cpu(hdr->src_nid), le64_to_cpu(hdr->dest_nid)); - - /* drop peer ref taken on init */ - ksocknal_put_peer (fmb->fmb_peer); - - spin_lock_irqsave (&fmp->fmp_lock, flags); - - list_add (&fmb->fmb_list, &fmp->fmp_idle_fmbs); - fmp->fmp_nactive_fmbs--; - - if (!list_empty (&fmp->fmp_blocked_conns)) { - conn = list_entry (fmb->fmb_pool->fmp_blocked_conns.next, - ksock_conn_t, ksnc_rx_list); - list_del (&conn->ksnc_rx_list); - } - - spin_unlock_irqrestore (&fmp->fmp_lock, flags); - - if (conn == NULL) - return; - - CDEBUG (D_NET, "Scheduling conn %p\n", conn); - LASSERT (conn->ksnc_rx_scheduled); - LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_FMB_SLEEP); + static char ksocknal_slop_buffer[4096]; + int nob; + unsigned int niov; + int skipped; - conn->ksnc_rx_state = SOCKNAL_RX_GET_FMB; + LASSERT(conn->ksnc_proto != NULL); - sched = conn->ksnc_scheduler; + if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) { + /* Remind the socket to ack eagerly... */ + ksocknal_lib_eager_ack(conn); + } - spin_lock_irqsave (&sched->kss_lock, flags); + if (nob_to_skip == 0) { /* right at next packet boundary now */ + conn->ksnc_rx_started = 0; + smp_mb(); /* racing with timeout thread */ - list_add_tail (&conn->ksnc_rx_list, &sched->kss_rx_conns); - cfs_waitq_signal (&sched->kss_waitq); + 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 kvec *)&conn->ksnc_rx_iov_space; + conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg; - spin_unlock_irqrestore (&sched->kss_lock, flags); -} + 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; -ksock_fmb_t * -ksocknal_get_idle_fmb (ksock_conn_t *conn) -{ - int payload_nob = conn->ksnc_rx_nob_left; - unsigned long flags; - ksock_fmb_pool_t *pool; - ksock_fmb_t *fmb; + 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); - LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB); - LASSERT (kpr_routing(&ksocknal_data.ksnd_router)); + 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; - if (payload_nob <= SOCKNAL_SMALL_FWD_PAGES * CFS_PAGE_SIZE) - pool = &ksocknal_data.ksnd_small_fmp; - else - pool = &ksocknal_data.ksnd_large_fmp; + default: + LBUG (); + } + conn->ksnc_rx_niov = 1; - spin_lock_irqsave (&pool->fmp_lock, flags); + conn->ksnc_rx_kiov = NULL; + conn->ksnc_rx_nkiov = 0; + conn->ksnc_rx_csum = ~0; + return (1); + } - if (!list_empty (&pool->fmp_idle_fmbs)) { - fmb = list_entry(pool->fmp_idle_fmbs.next, - ksock_fmb_t, fmb_list); - list_del (&fmb->fmb_list); - pool->fmp_nactive_fmbs++; - spin_unlock_irqrestore (&pool->fmp_lock, flags); + /* Set up to skip as much as possible now. If there's more left + * (ran out of iov entries) we'll get called again */ - return (fmb); - } + 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; - /* deschedule until fmb free */ + do { + nob = min_t(int, nob_to_skip, sizeof(ksocknal_slop_buffer)); - conn->ksnc_rx_state = SOCKNAL_RX_FMB_SLEEP; + 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; - list_add_tail (&conn->ksnc_rx_list, - &pool->fmp_blocked_conns); + } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */ + niov < sizeof(conn->ksnc_rx_iov_space) / sizeof(struct kvec)); - spin_unlock_irqrestore (&pool->fmp_lock, flags); - return (NULL); + conn->ksnc_rx_niov = niov; + conn->ksnc_rx_kiov = NULL; + conn->ksnc_rx_nkiov = 0; + conn->ksnc_rx_nob_wanted = skipped; + return (0); } -int -ksocknal_init_fmb (ksock_conn_t *conn, ksock_fmb_t *fmb) +static int +ksocknal_process_receive(struct ksock_conn *conn, + struct page **rx_scratch_pgs, + struct kvec *scratch_iov) { - int payload_nob = conn->ksnc_rx_nob_left; - ptl_nid_t dest_nid = le64_to_cpu(conn->ksnc_hdr.dest_nid); - int niov = 0; - int nob = payload_nob; + 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, rx_scratch_pgs, + scratch_iov); + + if (rc <= 0) { + struct lnet_process_id ksnp_id; + + 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, + (conn->ksnc_closing) ? 0 : rc); + return (rc == 0 ? -ESHUTDOWN : rc); + } - LASSERT (conn->ksnc_rx_scheduled); - LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB); - LASSERT (conn->ksnc_rx_nob_wanted == conn->ksnc_rx_nob_left); - LASSERT (payload_nob >= 0); - LASSERT (payload_nob <= fmb->fmb_pool->fmp_buff_pages * CFS_PAGE_SIZE); - LASSERT (sizeof (ptl_hdr_t) < CFS_PAGE_SIZE); - LASSERT (fmb->fmb_kiov[0].kiov_offset == 0); - - /* Take a ref on the conn's peer to prevent module unload before - * forwarding completes. */ - fmb->fmb_peer = conn->ksnc_peer; - atomic_inc (&conn->ksnc_peer->ksnp_refcount); - - /* Copy the header we just read into the forwarding buffer. If - * there's payload, start reading reading it into the buffer, - * otherwise the forwarding buffer can be kicked off - * immediately. */ - fmb->fmb_hdr = conn->ksnc_hdr; - - while (nob > 0) { - LASSERT (niov < fmb->fmb_pool->fmp_buff_pages); - LASSERT (fmb->fmb_kiov[niov].kiov_offset == 0); - fmb->fmb_kiov[niov].kiov_len = MIN (CFS_PAGE_SIZE, nob); - nob -= CFS_PAGE_SIZE; - niov++; + if (conn->ksnc_rx_nob_wanted != 0) { + /* short read */ + return (-EAGAIN); + } } + switch (conn->ksnc_rx_state) { + case SOCKNAL_RX_KSM_HEADER: + if (conn->ksnc_flip) { + __swab32s(&conn->ksnc_msg.ksm_type); + __swab32s(&conn->ksnc_msg.ksm_csum); + __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]); + __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]); + } - kpr_fwd_init(&fmb->fmb_fwd, dest_nid, &fmb->fmb_hdr, - payload_nob, niov, fmb->fmb_kiov, - ksocknal_fmb_callback, fmb); + if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP && + conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) { + CERROR("%s: Unknown message type: %x\n", + libcfs_id2str(conn->ksnc_peer->ksnp_id), + conn->ksnc_msg.ksm_type); + ksocknal_new_packet(conn, 0); + ksocknal_close_conn_and_siblings(conn, -EPROTO); + return (-EPROTO); + } - if (payload_nob == 0) { /* got complete packet already */ - CDEBUG (D_NET, "%p "LPX64"->"LPX64" fwd_start (immediate)\n", - conn, le64_to_cpu(conn->ksnc_hdr.src_nid), dest_nid); + if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP && + conn->ksnc_msg.ksm_csum != 0 && /* has checksum */ + conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) { + /* NOOP Checksum error */ + CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n", + libcfs_id2str(conn->ksnc_peer->ksnp_id), + conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum); + ksocknal_new_packet(conn, 0); + ksocknal_close_conn_and_siblings(conn, -EPROTO); + return (-EIO); + } - kpr_fwd_start (&ksocknal_data.ksnd_router, &fmb->fmb_fwd); + if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) { + __u64 cookie = 0; - ksocknal_new_packet (conn, 0); /* on to next packet */ - return (1); - } + LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x); - conn->ksnc_cookie = fmb; /* stash fmb for later */ - conn->ksnc_rx_state = SOCKNAL_RX_BODY_FWD; /* read in the payload */ + if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) + cookie = conn->ksnc_msg.ksm_zc_cookies[0]; - /* Set up conn->ksnc_rx_kiov to read the payload into fmb's kiov-ed - * buffer */ - LASSERT (niov <= sizeof(conn->ksnc_rx_iov_space)/sizeof(ptl_kiov_t)); + rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie, + conn->ksnc_msg.ksm_zc_cookies[1]); - conn->ksnc_rx_niov = 0; - conn->ksnc_rx_nkiov = niov; - conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov; - memcpy(conn->ksnc_rx_kiov, fmb->fmb_kiov, niov * sizeof(ptl_kiov_t)); + if (rc != 0) { + 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); + ksocknal_close_conn_and_siblings(conn, -EPROTO); + return (rc); + } + } - CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d reading body\n", conn, - le64_to_cpu(conn->ksnc_hdr.src_nid), dest_nid, payload_nob); - return (0); -} - -void -ksocknal_fwd_parse (ksock_conn_t *conn) -{ - ksock_peer_t *peer; - ptl_nid_t dest_nid = le64_to_cpu(conn->ksnc_hdr.dest_nid); - ptl_nid_t src_nid = le64_to_cpu(conn->ksnc_hdr.src_nid); - int body_len = le32_to_cpu(conn->ksnc_hdr.payload_length); - char str[PTL_NALFMT_SIZE]; - char str2[PTL_NALFMT_SIZE]; - - CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d parsing header\n", conn, - src_nid, dest_nid, conn->ksnc_rx_nob_left); - - LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_HEADER); - LASSERT (conn->ksnc_rx_scheduled); - - if (body_len < 0) { /* length corrupt (overflow) */ - CERROR("dropping packet from "LPX64" (%s) for "LPX64" (%s): " - "packet size %d illegal\n", - src_nid, portals_nid2str(TCPNAL, src_nid, str), - dest_nid, portals_nid2str(TCPNAL, dest_nid, str2), - body_len); - - ksocknal_new_packet (conn, 0); /* on to new packet */ - return; - } - - if (!kpr_routing(&ksocknal_data.ksnd_router)) { /* not forwarding */ - CERROR("dropping packet from "LPX64" (%s) for "LPX64 - " (%s): not forwarding\n", - src_nid, portals_nid2str(TCPNAL, src_nid, str), - dest_nid, portals_nid2str(TCPNAL, dest_nid, str2)); - /* on to new packet (skip this one's body) */ - ksocknal_new_packet (conn, body_len); - return; - } - - if (body_len > PTL_MTU) { /* too big to forward */ - CERROR ("dropping packet from "LPX64" (%s) for "LPX64 - "(%s): packet size %d too big\n", - src_nid, portals_nid2str(TCPNAL, src_nid, str), - dest_nid, portals_nid2str(TCPNAL, dest_nid, str2), - body_len); - /* on to new packet (skip this one's body) */ - ksocknal_new_packet (conn, body_len); - return; - } - - /* should have gone direct */ - peer = ksocknal_get_peer (conn->ksnc_hdr.dest_nid); - if (peer != NULL) { - CERROR ("dropping packet from "LPX64" (%s) for "LPX64 - "(%s): target is a peer\n", - src_nid, portals_nid2str(TCPNAL, src_nid, str), - dest_nid, portals_nid2str(TCPNAL, dest_nid, str2)); - ksocknal_put_peer (peer); /* drop ref from get above */ - - /* on to next packet (skip this one's body) */ - ksocknal_new_packet (conn, body_len); - return; - } - - conn->ksnc_rx_state = SOCKNAL_RX_GET_FMB; /* Getting FMB now */ - conn->ksnc_rx_nob_left = body_len; /* stash packet size */ - conn->ksnc_rx_nob_wanted = body_len; /* (no slop) */ -} - -int -ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip) -{ - static char ksocknal_slop_buffer[4096]; - - int nob; - int niov; - int skipped; + if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) { + ksocknal_new_packet (conn, 0); + return 0; /* NOOP is done and just return */ + } - if (nob_to_skip == 0) { /* right at next packet boundary now */ - conn->ksnc_rx_started = 0; - mb (); /* racing with timeout thread */ + conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER; + conn->ksnc_rx_nob_wanted = sizeof(struct ksock_lnet_msg); + conn->ksnc_rx_nob_left = sizeof(struct ksock_lnet_msg); - conn->ksnc_rx_state = SOCKNAL_RX_HEADER; - conn->ksnc_rx_nob_wanted = sizeof (ptl_hdr_t); - conn->ksnc_rx_nob_left = sizeof (ptl_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 ksock_lnet_msg); - conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space; - conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_hdr; - conn->ksnc_rx_iov[0].iov_len = sizeof (ptl_hdr_t); conn->ksnc_rx_niov = 1; - conn->ksnc_rx_kiov = NULL; conn->ksnc_rx_nkiov = 0; - return (1); - } - /* Set up to skip as much a possible now. If there's more left - * (ran out of iov entries) we'll get called again */ + goto again; /* read lnet header now */ - 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; + case SOCKNAL_RX_LNET_HEADER: + /* unpack message header */ + conn->ksnc_proto->pro_unpack(&conn->ksnc_msg); - do { - nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer)); + if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) { + /* Userspace peer_ni */ + lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr; + id = &conn->ksnc_peer->ksnp_id; - 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; + /* Substitute process ID assigned at connection time */ + lhdr->src_pid = cpu_to_le32(id->pid); + lhdr->src_nid = cpu_to_le64(id->nid); + } - } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */ - niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec)); + conn->ksnc_rx_state = SOCKNAL_RX_PARSE; + ksocknal_conn_addref(conn); /* ++ref while parsing */ - conn->ksnc_rx_niov = niov; - conn->ksnc_rx_kiov = NULL; - conn->ksnc_rx_nkiov = 0; - conn->ksnc_rx_nob_wanted = skipped; - return (0); -} + rc = lnet_parse(conn->ksnc_peer->ksnp_ni, + &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr, + conn->ksnc_peer->ksnp_id.nid, conn, 0); + if (rc < 0) { + /* I just received garbage: give up on this conn */ + ksocknal_new_packet(conn, 0); + ksocknal_close_conn_and_siblings (conn, rc); + ksocknal_conn_decref(conn); + return (-EPROTO); + } -int -ksocknal_process_receive (ksock_conn_t *conn) -{ - ksock_fmb_t *fmb; - int rc; + /* I'm racing with ksocknal_recv() */ + LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE || + conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD); - LASSERT (atomic_read (&conn->ksnc_refcount) > 0); + if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD) + return 0; - /* doesn't need a forwarding buffer */ - if (conn->ksnc_rx_state != SOCKNAL_RX_GET_FMB) - goto try_read; + /* ksocknal_recv() got called */ + goto again; - get_fmb: - fmb = ksocknal_get_idle_fmb (conn); - if (fmb == NULL) { - /* conn descheduled waiting for idle fmb */ - return (0); - } + case SOCKNAL_RX_LNET_PAYLOAD: + /* payload all received */ + rc = 0; - if (ksocknal_init_fmb (conn, fmb)) { - /* packet forwarded */ - return (0); - } + if (conn->ksnc_rx_nob_left == 0 && /* not truncating */ + conn->ksnc_msg.ksm_csum != 0 && /* has checksum */ + conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) { + CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n", + libcfs_id2str(conn->ksnc_peer->ksnp_id), + conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum); + rc = -EIO; + } - try_read: - /* NB: sched lock NOT held */ - LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_HEADER || - conn->ksnc_rx_state == SOCKNAL_RX_BODY || - conn->ksnc_rx_state == SOCKNAL_RX_BODY_FWD || - conn->ksnc_rx_state == SOCKNAL_RX_SLOP); - - LASSERT (conn->ksnc_rx_nob_wanted > 0); - - rc = ksocknal_receive(conn); - - if (rc <= 0) { - LASSERT (rc != -EAGAIN); - - if (rc == 0) - CWARN ("[%p] EOF from "LPX64" ip %d.%d.%d.%d:%d\n", - conn, conn->ksnc_peer->ksnp_nid, - HIPQUAD(conn->ksnc_ipaddr), - conn->ksnc_port); - else if (!conn->ksnc_closing) - CERROR ("[%p] Error %d on read from "LPX64 - " ip %d.%d.%d.%d:%d\n", - conn, rc, conn->ksnc_peer->ksnp_nid, - HIPQUAD(conn->ksnc_ipaddr), - conn->ksnc_port); - - ksocknal_close_conn_and_siblings (conn, rc); - return (rc == 0 ? -ESHUTDOWN : rc); - } + if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) { + LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x); - if (conn->ksnc_rx_nob_wanted != 0) { - /* short read */ - return (-EAGAIN); - } + lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr; + id = &conn->ksnc_peer->ksnp_id; - switch (conn->ksnc_rx_state) { - case SOCKNAL_RX_HEADER: - if (conn->ksnc_hdr.type != cpu_to_le32(PTL_MSG_HELLO) && - le64_to_cpu(conn->ksnc_hdr.dest_nid) != - ksocknal_lib.libnal_ni.ni_pid.nid) { - /* This packet isn't for me */ - ksocknal_fwd_parse (conn); - switch (conn->ksnc_rx_state) { - case SOCKNAL_RX_HEADER: /* skipped (zero payload) */ - return (0); /* => come back later */ - case SOCKNAL_RX_SLOP: /* skipping packet's body */ - goto try_read; /* => go read it */ - case SOCKNAL_RX_GET_FMB: /* forwarding */ - goto get_fmb; /* => go get a fwd msg buffer */ - default: - LBUG (); - } - /* Not Reached */ + rc = conn->ksnc_proto->pro_handle_zcreq(conn, + conn->ksnc_msg.ksm_zc_cookies[0], + *ksocknal_tunables.ksnd_nonblk_zcack || + le64_to_cpu(lhdr->src_nid) != id->nid); } - /* sets wanted_len, iovs etc */ - rc = lib_parse(&ksocknal_lib, &conn->ksnc_hdr, conn); + 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 != PTL_OK) { - /* I just received garbage: give up on this conn */ + if (rc != 0) { + ksocknal_new_packet(conn, 0); ksocknal_close_conn_and_siblings (conn, rc); return (-EPROTO); } - - if (conn->ksnc_rx_nob_wanted != 0) { /* need to get payload? */ - conn->ksnc_rx_state = SOCKNAL_RX_BODY; - goto try_read; /* go read the payload */ - } - /* Fall through (completed packet for me) */ - - case SOCKNAL_RX_BODY: - /* payload all received */ - lib_finalize(&ksocknal_lib, NULL, conn->ksnc_cookie, PTL_OK); /* Fall through */ case SOCKNAL_RX_SLOP: /* starting new packet? */ if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left)) - return (0); /* come back later */ - goto try_read; /* try to finish reading slop now */ - - case SOCKNAL_RX_BODY_FWD: - /* payload all received */ - CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d fwd_start (got body)\n", - conn, le64_to_cpu(conn->ksnc_hdr.src_nid), - le64_to_cpu(conn->ksnc_hdr.dest_nid), - conn->ksnc_rx_nob_left); - - /* forward the packet. NB ksocknal_init_fmb() put fmb into - * conn->ksnc_cookie */ - fmb = (ksock_fmb_t *)conn->ksnc_cookie; - kpr_fwd_start (&ksocknal_data.ksnd_router, &fmb->fmb_fwd); - - /* no slop in forwarded packets */ - LASSERT (conn->ksnc_rx_nob_left == 0); - - ksocknal_new_packet (conn, 0); /* on to next packet */ - return (0); /* (later) */ + return 0; /* come back later */ + goto again; /* try to finish reading slop now */ default: break; @@ -1413,824 +1382,1019 @@ ksocknal_process_receive (ksock_conn_t *conn) return (-EINVAL); /* keep gcc happy */ } -ptl_err_t -ksocknal_recv (lib_nal_t *nal, void *private, lib_msg_t *msg, - unsigned int niov, struct iovec *iov, - size_t offset, size_t mlen, size_t rlen) +int +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; + struct ksock_conn *conn = private; + struct ksock_sched *sched = conn->ksnc_scheduler; LASSERT (mlen <= rlen); - LASSERT (niov <= PTL_MD_MAX_IOV); + 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; - conn->ksnc_rx_nkiov = 0; - conn->ksnc_rx_kiov = NULL; - conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov; - conn->ksnc_rx_niov = - lib_extract_iov(PTL_MD_MAX_IOV, conn->ksnc_rx_iov, - niov, iov, offset, mlen); - - LASSERT (mlen == - lib_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) + - lib_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov)); + if (mlen == 0 || iov != NULL) { + conn->ksnc_rx_nkiov = 0; + conn->ksnc_rx_kiov = NULL; + conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov; + conn->ksnc_rx_niov = + lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov, + niov, iov, offset, mlen); + } else { + conn->ksnc_rx_niov = 0; + conn->ksnc_rx_iov = NULL; + conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov; + conn->ksnc_rx_nkiov = + lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov, + niov, kiov, offset, mlen); + } - return (PTL_OK); -} + LASSERT (mlen == + lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) + + lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov)); -ptl_err_t -ksocknal_recv_pages (lib_nal_t *nal, void *private, lib_msg_t *msg, - unsigned int niov, ptl_kiov_t *kiov, - size_t offset, size_t mlen, size_t rlen) -{ - ksock_conn_t *conn = (ksock_conn_t *)private; + LASSERT (conn->ksnc_rx_scheduled); - LASSERT (mlen <= rlen); - LASSERT (niov <= PTL_MD_MAX_IOV); + spin_lock_bh(&sched->kss_lock); - conn->ksnc_cookie = msg; - conn->ksnc_rx_nob_wanted = mlen; - conn->ksnc_rx_nob_left = rlen; + 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; - conn->ksnc_rx_niov = 0; - conn->ksnc_rx_iov = NULL; - conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov; - conn->ksnc_rx_nkiov = - lib_extract_kiov(PTL_MD_MAX_IOV, conn->ksnc_rx_kiov, - niov, kiov, offset, mlen); + case SOCKNAL_RX_PARSE: + /* scheduler hasn't noticed I'm parsing yet */ + break; + } - LASSERT (mlen == - lib_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) + - lib_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov)); + conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD; - return (PTL_OK); + 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) { - unsigned long flags; - int rc; + int rc; - spin_lock_irqsave(&sched->kss_lock, flags); + spin_lock_bh(&sched->kss_lock); - rc = (!ksocknal_data.ksnd_shuttingdown && -#if SOCKNAL_ZC - list_empty(&sched->kss_zctxdone_list) && -#endif - list_empty(&sched->kss_rx_conns) && - list_empty(&sched->kss_tx_conns)); + rc = (!ksocknal_data.ksnd_shuttingdown && + list_empty(&sched->kss_rx_conns) && + list_empty(&sched->kss_tx_conns)); - spin_unlock_irqrestore(&sched->kss_lock, flags); - 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; - unsigned long flags; - int rc; - int nloops = 0; - int id = sched - ksocknal_data.ksnd_schedulers; - char name[16]; - - snprintf (name, sizeof (name),"ksocknald_%02d", id); - kportal_daemonize (name); - kportal_blockallsigs (); - -#if (CONFIG_SMP && CPU_AFFINITY) - id = ksocknal_sched2cpu(id); - if (cpu_online(id)) { - cpumask_t m; - cpu_set(id, m); - set_cpus_allowed(current, m); - } else { - CERROR ("Can't set CPU affinity for %s to %d\n", name, id); - } -#endif /* CONFIG_SMP && CPU_AFFINITY */ - - spin_lock_irqsave (&sched->kss_lock, flags); - - 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, - ksock_conn_t, 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_irqrestore(&sched->kss_lock, flags); - - rc = ksocknal_process_receive(conn); - - spin_lock_irqsave(&sched->kss_lock, flags); - - /* 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_FMB_SLEEP || - conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB) { - /* Conn blocked for a forwarding buffer. - * It will get queued for my attention when - * one becomes available (and it might just - * already have been!). Meanwhile my ref - * on it stays put. */ - } 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_put_conn(conn); - } - - did_something = 1; - } - - if (!list_empty (&sched->kss_tx_conns)) { - conn = list_entry(sched->kss_tx_conns.next, - ksock_conn_t, ksnc_tx_list); - list_del (&conn->ksnc_tx_list); + 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; +} - LASSERT(conn->ksnc_tx_scheduled); - LASSERT(conn->ksnc_tx_ready); - LASSERT(!list_empty(&conn->ksnc_tx_queue)); +/* + * Add connection to kss_rx_conns of scheduler + * and wakeup the scheduler. + */ +void ksocknal_read_callback(struct ksock_conn *conn) +{ + struct ksock_sched *sched; + ENTRY; - tx = list_entry(conn->ksnc_tx_queue.next, - ksock_tx_t, tx_list); - /* dequeue now so empty list => more to send */ - list_del(&tx->tx_list); + sched = conn->ksnc_scheduler; - /* 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_irqrestore (&sched->kss_lock, flags); + spin_lock_bh(&sched->kss_lock); - rc = ksocknal_process_transmit(conn, tx); + conn->ksnc_rx_ready = 1; - spin_lock_irqsave (&sched->kss_lock, flags); + 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); - if (rc == -ENOMEM || rc == -EAGAIN) { - /* Incomplete send: replace tx on HEAD of tx_queue */ - list_add (&tx->tx_list, &conn->ksnc_tx_queue); - } else { - /* Complete send; assume space for more */ - conn->ksnc_tx_ready = 1; - } + wake_up (&sched->kss_waitq); + } + spin_unlock_bh(&sched->kss_lock); - 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_put_conn (conn); - } + EXIT; +} - did_something = 1; - } -#if SOCKNAL_ZC - if (!list_empty (&sched->kss_zctxdone_list)) { - ksock_tx_t *tx = - list_entry(sched->kss_zctxdone_list.next, - ksock_tx_t, tx_list); - did_something = 1; +/* + * Add connection to kss_tx_conns of scheduler + * and wakeup the scheduler. + */ +void ksocknal_write_callback(struct ksock_conn *conn) +{ + struct ksock_sched *sched; + ENTRY; - list_del (&tx->tx_list); - spin_unlock_irqrestore (&sched->kss_lock, flags); + sched = conn->ksnc_scheduler; - ksocknal_tx_done (tx, 1); + spin_lock_bh(&sched->kss_lock); - spin_lock_irqsave (&sched->kss_lock, flags); - } -#endif - if (!did_something || /* nothing to do */ - ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */ - spin_unlock_irqrestore (&sched->kss_lock, flags); + conn->ksnc_tx_ready = 1; - nloops = 0; + 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); - if (!did_something) { /* wait for something to do */ - rc = wait_event_interruptible (sched->kss_waitq, - !ksocknal_sched_cansleep(sched)); - LASSERT (rc == 0); - } else - our_cond_resched(); + wake_up(&sched->kss_waitq); + } - spin_lock_irqsave (&sched->kss_lock, flags); - } - } + spin_unlock_bh(&sched->kss_lock); - spin_unlock_irqrestore (&sched->kss_lock, flags); - ksocknal_thread_fini (); - return (0); + EXIT; } -/* - * Add connection to kss_rx_conns of scheduler - * and wakeup the scheduler. - */ -void ksocknal_read_callback (ksock_conn_t *conn) +static const struct ksock_proto * +ksocknal_parse_proto_version(struct ksock_hello_msg *hello) { - ksock_sched_t *sched; - unsigned long flags; - ENTRY; - - sched = conn->ksnc_scheduler; - - spin_lock_irqsave (&sched->kss_lock, flags); + __u32 version = 0; - conn->ksnc_rx_ready = 1; + if (hello->kshm_magic == LNET_PROTO_MAGIC) + version = hello->kshm_version; + else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC)) + version = __swab32(hello->kshm_version); - 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 */ - atomic_inc (&conn->ksnc_refcount); + if (version != 0) { +#if SOCKNAL_VERSION_DEBUG + if (*ksocknal_tunables.ksnd_protocol == 1) + return NULL; - cfs_waitq_signal (&sched->kss_waitq); - } - spin_unlock_irqrestore (&sched->kss_lock, flags); - - EXIT; -} - -/* - * Add connection to kss_tx_conns of scheduler - * and wakeup the scheduler. - */ -void ksocknal_write_callback (ksock_conn_t *conn) -{ - ksock_sched_t *sched; - unsigned long flags; - ENTRY; + if (*ksocknal_tunables.ksnd_protocol == 2 && + version == KSOCK_PROTO_V3) + return NULL; +#endif + if (version == KSOCK_PROTO_V2) + return &ksocknal_protocol_v2x; - sched = conn->ksnc_scheduler; + if (version == KSOCK_PROTO_V3) + return &ksocknal_protocol_v3x; - spin_lock_irqsave (&sched->kss_lock, flags); + return NULL; + } - conn->ksnc_tx_ready = 1; + if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) { + struct lnet_magicversion *hmv; - 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 */ - atomic_inc (&conn->ksnc_refcount); + BUILD_BUG_ON(sizeof(struct lnet_magicversion) != + offsetof(struct ksock_hello_msg, kshm_src_nid)); - cfs_waitq_signal (&sched->kss_waitq); - } + hmv = (struct lnet_magicversion *)hello; - spin_unlock_irqrestore (&sched->kss_lock, flags); + if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) && + hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR)) + return &ksocknal_protocol_v1x; + } - EXIT; + return NULL; } int -ksocknal_sock_write (struct socket *sock, void *buffer, int nob) +ksocknal_send_hello(struct lnet_ni *ni, struct ksock_conn *conn, + lnet_nid_t peer_nid, struct ksock_hello_msg *hello) { - return ksocknal_lib_sock_write(sock, buffer, nob); -} + /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */ + struct ksock_net *net = (struct ksock_net *)ni->ni_data; -int -ksocknal_sock_read (struct socket *sock, void *buffer, int nob) -{ - return ksocknal_lib_sock_read(sock, buffer, nob); -} + LASSERT(hello->kshm_nips <= LNET_INTERFACES_NUM); -int -ksocknal_send_hello (ksock_conn_t *conn, __u32 *ipaddrs, int nipaddrs) -{ - /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */ - struct socket *sock = conn->ksnc_sock; - ptl_hdr_t hdr; - ptl_magicversion_t *hmv = (ptl_magicversion_t *)&hdr.dest_nid; - int i; - int rc; - - LASSERT (conn->ksnc_type != SOCKNAL_CONN_NONE); - LASSERT (nipaddrs <= SOCKNAL_MAX_INTERFACES); - - /* No need for getconnsock/putconnsock */ - LASSERT (!conn->ksnc_closing); - - LASSERT (sizeof (*hmv) == sizeof (hdr.dest_nid)); - hmv->magic = cpu_to_le32 (PORTALS_PROTO_MAGIC); - hmv->version_major = cpu_to_le16 (PORTALS_PROTO_VERSION_MAJOR); - hmv->version_minor = cpu_to_le16 (PORTALS_PROTO_VERSION_MINOR); - - hdr.src_nid = cpu_to_le64 (ksocknal_lib.libnal_ni.ni_pid.nid); - hdr.type = cpu_to_le32 (PTL_MSG_HELLO); - hdr.payload_length = cpu_to_le32 (nipaddrs * sizeof(*ipaddrs)); - - hdr.msg.hello.type = cpu_to_le32 (conn->ksnc_type); - hdr.msg.hello.incarnation = - cpu_to_le64 (ksocknal_data.ksnd_incarnation); - - /* Receiver is eager */ - rc = ksocknal_sock_write (sock, &hdr, sizeof(hdr)); - if (rc != 0) { - CERROR ("Error %d sending HELLO hdr to %u.%u.%u.%u/%d\n", - rc, HIPQUAD(conn->ksnc_ipaddr), conn->ksnc_port); - return (rc); - } + /* rely on caller to hold a ref on socket so it wouldn't disappear */ + LASSERT(conn->ksnc_proto != NULL); - if (nipaddrs == 0) - return (0); + hello->kshm_src_nid = ni->ni_nid; + hello->kshm_dst_nid = peer_nid; + hello->kshm_src_pid = the_lnet.ln_pid; - for (i = 0; i < nipaddrs; i++) { - ipaddrs[i] = __cpu_to_le32 (ipaddrs[i]); - } + hello->kshm_src_incarnation = net->ksnn_incarnation; + hello->kshm_ctype = conn->ksnc_type; - rc = ksocknal_sock_write (sock, ipaddrs, nipaddrs * sizeof(*ipaddrs)); - if (rc != 0) - CERROR ("Error %d sending HELLO payload (%d)" - " to %u.%u.%u.%u/%d\n", rc, nipaddrs, - HIPQUAD(conn->ksnc_ipaddr), conn->ksnc_port); - return (rc); + return conn->ksnc_proto->pro_send_hello(conn, hello); } -int +static int ksocknal_invert_type(int type) { switch (type) { - case SOCKNAL_CONN_ANY: - case SOCKNAL_CONN_CONTROL: + case SOCKLND_CONN_ANY: + case SOCKLND_CONN_CONTROL: return (type); - case SOCKNAL_CONN_BULK_IN: - return SOCKNAL_CONN_BULK_OUT; - case SOCKNAL_CONN_BULK_OUT: - return SOCKNAL_CONN_BULK_IN; + case SOCKLND_CONN_BULK_IN: + return SOCKLND_CONN_BULK_OUT; + case SOCKLND_CONN_BULK_OUT: + return SOCKLND_CONN_BULK_IN; default: - return (SOCKNAL_CONN_NONE); + return (SOCKLND_CONN_NONE); } } int -ksocknal_recv_hello (ksock_conn_t *conn, ptl_nid_t *nid, - __u64 *incarnation, __u32 *ipaddrs) +ksocknal_recv_hello(struct lnet_ni *ni, struct ksock_conn *conn, + struct ksock_hello_msg *hello, + struct lnet_process_id *peerid, + __u64 *incarnation) { - struct socket *sock = conn->ksnc_sock; - int rc; - int nips; - int i; - int type; - ptl_hdr_t hdr; - ptl_magicversion_t *hmv; - char ipbuf[PTL_NALFMT_SIZE]; - - hmv = (ptl_magicversion_t *)&hdr.dest_nid; - LASSERT (sizeof (*hmv) == sizeof (hdr.dest_nid)); - - rc = ksocknal_sock_read (sock, hmv, sizeof (*hmv)); + /* Return < 0 fatal error + * 0 success + * EALREADY lost connection race + * EPROTO protocol version mismatch + */ + struct socket *sock = conn->ksnc_sock; + int active = (conn->ksnc_proto != NULL); + int timeout; + int proto_match; + int rc; + 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)); + + timeout = active ? lnet_get_lnd_timeout() : + lnet_acceptor_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)); - return (rc); + CERROR("Error %d reading HELLO from %pI4h\n", + rc, &conn->ksnc_ipaddr); + LASSERT (rc < 0); + return rc; } - if (hmv->magic != le32_to_cpu (PORTALS_PROTO_MAGIC)) { - CERROR ("Bad magic %#08x (%#08x expected) from %u.%u.%u.%u\n", - __cpu_to_le32 (hmv->magic), PORTALS_PROTO_MAGIC, - HIPQUAD(conn->ksnc_ipaddr)); - return (-EPROTO); + if (hello->kshm_magic != LNET_PROTO_MAGIC && + hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) && + hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) { + /* Unexpected magic! */ + CERROR ("Bad magic(1) %#08x (%#08x expected) from " + "%pI4h\n", __cpu_to_le32 (hello->kshm_magic), + LNET_PROTO_TCP_MAGIC, &conn->ksnc_ipaddr); + return -EPROTO; } - if (hmv->version_major != cpu_to_le16 (PORTALS_PROTO_VERSION_MAJOR) || - hmv->version_minor != cpu_to_le16 (PORTALS_PROTO_VERSION_MINOR)) { - CERROR ("Incompatible protocol version %d.%d (%d.%d expected)" - " from %u.%u.%u.%u\n", - le16_to_cpu (hmv->version_major), - le16_to_cpu (hmv->version_minor), - PORTALS_PROTO_VERSION_MAJOR, - PORTALS_PROTO_VERSION_MINOR, - HIPQUAD(conn->ksnc_ipaddr)); - return (-EPROTO); + rc = lnet_sock_read(sock, &hello->kshm_version, + sizeof(hello->kshm_version), timeout); + if (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_ni, tell peer_ni my protocol */ + conn->ksnc_proto = &ksocknal_protocol_v3x; +#if SOCKNAL_VERSION_DEBUG + if (*ksocknal_tunables.ksnd_protocol == 2) + conn->ksnc_proto = &ksocknal_protocol_v2x; + else if (*ksocknal_tunables.ksnd_protocol == 1) + conn->ksnc_proto = &ksocknal_protocol_v1x; +#endif + hello->kshm_nips = 0; + ksocknal_send_hello(ni, conn, ni->ni_nid, hello); + } + + CERROR("Unknown protocol version (%d.x expected) from %pI4h\n", + conn->ksnc_proto->pro_version, &conn->ksnc_ipaddr); + + return -EPROTO; } -#if (PORTALS_PROTO_VERSION_MAJOR != 1) -# error "This code only understands protocol version 1.x" -#endif - /* version 1 sends magic/version as the dest_nid of a 'hello' - * header, followed by payload full of interface IP addresses. - * Read the rest of it in now... */ + proto_match = (conn->ksnc_proto == proto); + conn->ksnc_proto = proto; - rc = ksocknal_sock_read (sock, hmv + 1, sizeof (hdr) - sizeof (*hmv)); + /* receive the rest of hello message anyway */ + rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout); if (rc != 0) { - CERROR ("Error %d reading rest of HELLO hdr from %u.%u.%u.%u\n", - rc, HIPQUAD(conn->ksnc_ipaddr)); - return (rc); + CERROR("Error %d reading or checking hello from from %pI4h\n", + rc, &conn->ksnc_ipaddr); + LASSERT (rc < 0); + return rc; } - /* ...and check we got what we expected */ - if (hdr.type != cpu_to_le32 (PTL_MSG_HELLO)) { - CERROR ("Expecting a HELLO hdr," - " but got type %d from %u.%u.%u.%u\n", - le32_to_cpu (hdr.type), - HIPQUAD(conn->ksnc_ipaddr)); - return (-EPROTO); - } + *incarnation = hello->kshm_src_incarnation; - if (le64_to_cpu(hdr.src_nid) == PTL_NID_ANY) { - CERROR("Expecting a HELLO hdr with a NID, but got PTL_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 (*nid == PTL_NID_ANY) { /* don't know peer's nid yet */ - *nid = le64_to_cpu(hdr.src_nid); - } else if (*nid != le64_to_cpu (hdr.src_nid)) { - LCONSOLE_ERROR("Connected successfully to nid "LPX64" on host " - "%u.%u.%u.%u, but they claimed they were nid " - LPX64" (%s); please check your Lustre " - "configuration.\n", - *nid, HIPQUAD(conn->ksnc_ipaddr), - le64_to_cpu(hdr.src_nid), - portals_nid2str(SOCKNAL, - le64_to_cpu(hdr.src_nid), - ipbuf)); - - return (-EPROTO); - } + if (!active && + conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) { + /* 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 { + recv_id.nid = hello->kshm_src_nid; + recv_id.pid = hello->kshm_src_pid; + } + + if (!active) { + *peerid = recv_id; + + /* 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" + " %pI4h, but they claimed they were " + "%s; please check your Lustre " + "configuration.\n", + libcfs_id2str(*peerid), + &conn->ksnc_ipaddr, + libcfs_id2str(recv_id)); + return -EPROTO; + } + + if (hello->kshm_ctype == SOCKLND_CONN_NONE) { + /* Possible protocol mismatch or I lost the connection race */ + return proto_match ? EALREADY : EPROTO; + } + + 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; +} - type = __le32_to_cpu(hdr.msg.hello.type); +static int +ksocknal_connect(struct ksock_route *route) +{ + LIST_HEAD(zombies); + struct ksock_peer_ni *peer_ni = route->ksnr_peer; + int type; + int wanted; + struct socket *sock; + time64_t deadline; + int retry_later = 0; + int rc = 0; - if (conn->ksnc_type == SOCKNAL_CONN_NONE) { - /* I've accepted this connection; peer determines type */ - conn->ksnc_type = ksocknal_invert_type(type); - if (conn->ksnc_type == SOCKNAL_CONN_NONE) { - CERROR ("Unexpected type %d from "LPX64"@%u.%u.%u.%u\n", - type, *nid, HIPQUAD(conn->ksnc_ipaddr)); - return (-EPROTO); - } - } else if (ksocknal_invert_type(type) != conn->ksnc_type) { - CERROR ("Mismatched types: me %d, "LPX64"@%u.%u.%u.%u %d\n", - conn->ksnc_type, *nid, HIPQUAD(conn->ksnc_ipaddr), - le32_to_cpu(hdr.msg.hello.type)); - return (-EPROTO); - } + deadline = ktime_get_seconds() + lnet_get_lnd_timeout(); - *incarnation = le64_to_cpu(hdr.msg.hello.incarnation); + write_lock_bh(&ksocknal_data.ksnd_global_lock); - nips = __le32_to_cpu (hdr.payload_length) / sizeof (__u32); + LASSERT (route->ksnr_scheduled); + LASSERT (!route->ksnr_connecting); - if (nips > SOCKNAL_MAX_INTERFACES || - nips * sizeof(__u32) != __le32_to_cpu (hdr.payload_length)) { - CERROR("Bad payload length %d from "LPX64"@%u.%u.%u.%u\n", - __le32_to_cpu (hdr.payload_length), - *nid, HIPQUAD(conn->ksnc_ipaddr)); - } + route->ksnr_connecting = 1; - if (nips == 0) - return (0); + for (;;) { + wanted = ksocknal_route_mask() & ~route->ksnr_connected; - rc = ksocknal_sock_read (sock, ipaddrs, nips * sizeof(*ipaddrs)); - if (rc != 0) { - CERROR ("Error %d reading IPs from "LPX64"@%u.%u.%u.%u\n", - rc, *nid, HIPQUAD(conn->ksnc_ipaddr)); - return (rc); - } + /* stop connecting if peer_ni/route got closed under me, or + * route got connected while queued */ + if (peer_ni->ksnp_closing || route->ksnr_deleted || + wanted == 0) { + retry_later = 0; + break; + } - for (i = 0; i < nips; i++) { - ipaddrs[i] = __le32_to_cpu(ipaddrs[i]); + /* reschedule if peer_ni is connecting to me */ + if (peer_ni->ksnp_accepting > 0) { + CDEBUG(D_NET, + "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; + } - if (ipaddrs[i] == 0) { - CERROR("Zero IP[%d] from "LPX64"@%u.%u.%u.%u\n", - i, *nid, HIPQUAD(conn->ksnc_ipaddr)); - return (-EPROTO); + if (retry_later) /* needs reschedule */ + break; + + if ((wanted & (1 << SOCKLND_CONN_ANY)) != 0) { + type = SOCKLND_CONN_ANY; + } else if ((wanted & (1 << SOCKLND_CONN_CONTROL)) != 0) { + type = SOCKLND_CONN_CONTROL; + } else if ((wanted & (1 << SOCKLND_CONN_BULK_IN)) != 0) { + type = SOCKLND_CONN_BULK_IN; + } else { + LASSERT ((wanted & (1 << SOCKLND_CONN_BULK_OUT)) != 0); + type = SOCKLND_CONN_BULK_OUT; + } + + write_unlock_bh(&ksocknal_data.ksnd_global_lock); + + if (ktime_get_seconds() >= deadline) { + rc = -ETIMEDOUT; + lnet_connect_console_error(rc, peer_ni->ksnp_id.nid, + route->ksnr_ipaddr, + route->ksnr_port); + 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_ni->ksnp_ni, route, sock, type); + if (rc < 0) { + lnet_connect_console_error(rc, peer_ni->ksnp_id.nid, + route->ksnr_ipaddr, + route->ksnr_port); + goto failed; + } + + /* A +ve RC means I have to retry because I lost the connection + * race or I have to renegotiate protocol version */ + retry_later = (rc != 0); + if (retry_later) + CDEBUG(D_NET, "peer_ni %s: conn race, retry later.\n", + libcfs_nid2str(peer_ni->ksnp_id.nid)); + + write_lock_bh(&ksocknal_data.ksnd_global_lock); } - return (nips); -} + route->ksnr_scheduled = 0; + route->ksnr_connecting = 0; -int -ksocknal_connect_peer (ksock_route_t *route, int type) -{ - struct socket *sock; - int rc; - int port; - int may_retry; + if (retry_later) { + /* re-queue for attention; this frees me up to handle + * the peer_ni's incoming connection request */ + + if (rc == EALREADY || + (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 = *ksocknal_tunables.ksnd_min_reconnectms / 1000; + route->ksnr_timeout = ktime_get_seconds() + + route->ksnr_retry_interval; + } - /* Iterate through reserved ports. When typed connections are - * used, we will need to bind to multiple ports, but we only know - * this at connect time. But, by that time we've already called - * bind() so we need a new socket. */ + ksocknal_launch_connection_locked(route); + } - for (port = 1023; port > 512; --port) { + write_unlock_bh(&ksocknal_data.ksnd_global_lock); + return retry_later; - rc = ksocknal_lib_connect_sock(&sock, &may_retry, route, port); + failed: + write_lock_bh(&ksocknal_data.ksnd_global_lock); - if (rc == 0) { - rc = ksocknal_create_conn(route, sock, type); - cfs_put_file(KSN_SOCK2FILE(sock)); - return rc; + 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_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); + route->ksnr_timeout = ktime_get_seconds() + route->ksnr_retry_interval; + + 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 (!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); } - if (!may_retry) - return rc; + /* take all the blocked packets while I've got the lock and + * complete below... */ + list_splice_init(&peer_ni->ksnp_tx_queue, &zombies); } - CERROR("Out of ports trying to bind to a reserved port\n"); - return (-EADDRINUSE); + write_unlock_bh(&ksocknal_data.ksnd_global_lock); + + ksocknal_peer_failed(peer_ni); + ksocknal_txlist_done(peer_ni->ksnp_ni, &zombies, rc); + return 0; } -void -ksocknal_autoconnect (ksock_route_t *route) +/* + * 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) { - CFS_LIST_HEAD (zombies); - ksock_tx_t *tx; - ksock_peer_t *peer; - unsigned long flags; - int type; - int rc = 0; + char name[16]; + int rc; + int total = ksocknal_data.ksnd_connd_starting + + ksocknal_data.ksnd_connd_running; - write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags); + if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) { + /* still in initializing */ + return 0; + } - for (;;) { - if (!ksocknal_tunables.ksnd_typed_conns) { - if ((route->ksnr_connected & (1<ksnr_connected & (1<ksnr_connected & (1<ksnr_connected & (1<= *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; + } - write_unlock_irqrestore(&ksocknal_data.ksnd_global_lock, flags); + if (list_empty(&ksocknal_data.ksnd_connd_routes)) { + /* no pending connecting request */ + return 0; + } - rc = ksocknal_connect_peer (route, type); - if (rc != 0) - goto failed; + if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) { + /* may run out of resource, retry later */ + *timeout = cfs_time_seconds(1); + return 0; + } - write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags); + if (ksocknal_data.ksnd_connd_starting > 0) { + /* serialize starting to avoid flood */ + return 0; } - LASSERT (route->ksnr_connecting); - route->ksnr_connecting = 0; - write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags); - return; + ksocknal_data.ksnd_connd_starting_stamp = sec; + ksocknal_data.ksnd_connd_starting++; + spin_unlock_bh(&ksocknal_data.ksnd_connd_lock); - failed: - switch (rc) { - /* "normal" errors */ - case -ECONNREFUSED: - LCONSOLE_ERROR("Connection was refused by host %u.%u.%u.%u on " - "port %d; check that Lustre is running on that " - "node.\n", - HIPQUAD(route->ksnr_ipaddr), route->ksnr_port); - break; - case -EHOSTUNREACH: - case -ENETUNREACH: - LCONSOLE_ERROR("Host %u.%u.%u.%u was unreachable; the network " - "or that node may be down, or Lustre may be " - "misconfigured.\n", - HIPQUAD(route->ksnr_ipaddr)); - break; - case -ETIMEDOUT: - LCONSOLE_ERROR("Connecting to host %u.%u.%u.%u on port %d took " - "too long; that node may be hung or " - "experiencing high load.\n", - HIPQUAD(route->ksnr_ipaddr), route->ksnr_port); - break; - /* errors that should be rare */ - case -EPROTO: - LCONSOLE_ERROR("Protocol error connecting to host %u.%u.%u.%u " - "on port %d: Is it running a compatible version" - " of Lustre?\n", - HIPQUAD(route->ksnr_ipaddr), route->ksnr_port); - break; - case -EADDRINUSE: - LCONSOLE_ERROR("No privileged ports available to connect to " - "host %u.%u.%u.%u on port %d\n", - HIPQUAD(route->ksnr_ipaddr), route->ksnr_port); - break; - default: - LCONSOLE_ERROR("Unexpected error %d connecting to " - "host %u.%u.%u.%u on port %d\n", rc, - HIPQUAD(route->ksnr_ipaddr), route->ksnr_port); - break; - } + /* NB: total is the next id */ + snprintf(name, sizeof(name), "socknal_cd%02d", total); + rc = ksocknal_thread_start(ksocknal_connd, NULL, name); - write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags); + spin_lock_bh(&ksocknal_data.ksnd_connd_lock); + if (rc == 0) + return 1; - peer = route->ksnr_peer; - LASSERT (route->ksnr_connecting); - route->ksnr_connecting = 0; + /* 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; +} + +/* + * 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) +{ + int val; - /* This is a retry rather than a new connection */ - LASSERT (route->ksnr_retry_interval != 0); - route->ksnr_timeout = cfs_time_add(cfs_time_current(), - route->ksnr_retry_interval); - route->ksnr_retry_interval = MIN (route->ksnr_retry_interval * 2, - SOCKNAL_MAX_RECONNECT_INTERVAL); - - if (!list_empty (&peer->ksnp_tx_queue) && - ksocknal_find_connecting_route_locked (peer) == NULL) { - LASSERT (list_empty (&peer->ksnp_conns)); - - /* None of the connections that the blocked packets are - * waiting for have been successful. Complete them now... */ - do { - tx = list_entry (peer->ksnp_tx_queue.next, - ksock_tx_t, tx_list); - list_del (&tx->tx_list); - list_add_tail (&tx->tx_list, &zombies); - } while (!list_empty (&peer->ksnp_tx_queue)); + if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) { + /* still in initializing */ + return 0; } -#if 0 /* irrelevent with only eager routes */ - if (!route->ksnr_deleted) { - /* make this route least-favourite for re-selection */ - list_del(&route->ksnr_list); - list_add_tail(&route->ksnr_list, &peer->ksnp_routes); + if (ksocknal_data.ksnd_connd_starting > 0) { + /* in progress of starting new thread */ + return 0; } -#endif - write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags); - - while (!list_empty (&zombies)) { - char ipbuf[PTL_NALFMT_SIZE]; - char ipbuf2[PTL_NALFMT_SIZE]; - tx = list_entry (zombies.next, ksock_tx_t, tx_list); - - CERROR ("Deleting packet type %d len %d ("LPX64" %s->"LPX64" %s)\n", - le32_to_cpu (tx->tx_hdr->type), - le32_to_cpu (tx->tx_hdr->payload_length), - le64_to_cpu (tx->tx_hdr->src_nid), - portals_nid2str(SOCKNAL, - le64_to_cpu(tx->tx_hdr->src_nid), - ipbuf), - le64_to_cpu (tx->tx_hdr->dest_nid), - portals_nid2str(SOCKNAL, - le64_to_cpu(tx->tx_hdr->src_nid), - ipbuf2)); - - list_del (&tx->tx_list); - /* complete now */ - ksocknal_tx_done (tx, 0); + + if (ksocknal_data.ksnd_connd_running <= + *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */ + return 0; } -} -int -ksocknal_autoconnectd (void *arg) -{ - long id = (long)arg; - char name[16]; - unsigned long flags; - ksock_route_t *route; - int rc; + /* created thread in past 120 seconds? */ + val = (int)(ksocknal_data.ksnd_connd_starting_stamp + + SOCKNAL_CONND_TIMEOUT - sec); - snprintf (name, sizeof (name), "ksocknal_ad%02ld", id); - kportal_daemonize (name); - kportal_blockallsigs (); + *timeout = (val > 0) ? cfs_time_seconds(val) : + cfs_time_seconds(SOCKNAL_CONND_TIMEOUT); + if (val > 0) + return 0; - spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags); + /* no creating in past 120 seconds */ - while (!ksocknal_data.ksnd_shuttingdown) { + return ksocknal_data.ksnd_connd_running > + ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV; +} - if (!list_empty (&ksocknal_data.ksnd_autoconnectd_routes)) { - route = list_entry (ksocknal_data.ksnd_autoconnectd_routes.next, - ksock_route_t, ksnr_connect_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; - list_del (&route->ksnr_connect_list); - spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags); + /* connd_routes can contain both pending and ordinary routes */ + list_for_each_entry(route, &ksocknal_data.ksnd_connd_routes, + ksnr_connd_list) { - ksocknal_autoconnect (route); - ksocknal_put_route (route); + if (route->ksnr_retry_interval == 0 || + now >= route->ksnr_timeout) + return route; - spin_lock_irqsave(&ksocknal_data.ksnd_autoconnectd_lock, - flags); - continue; + 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; } - spin_unlock_irqrestore(&ksocknal_data.ksnd_autoconnectd_lock, - flags); + 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); - rc = wait_event_interruptible(ksocknal_data.ksnd_autoconnectd_waitq, - ksocknal_data.ksnd_shuttingdown || - !list_empty(&ksocknal_data.ksnd_autoconnectd_routes)); + list_del(&cr->ksncr_list); + spin_unlock_bh(connd_lock); + dropped_lock = 1; - spin_lock_irqsave(&ksocknal_data.ksnd_autoconnectd_lock, flags); - } + ksocknal_create_conn(cr->ksncr_ni, NULL, + cr->ksncr_sock, SOCKLND_CONN_NONE); + lnet_ni_decref(cr->ksncr_ni); + LIBCFS_FREE(cr, sizeof(*cr)); - spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags); + spin_lock_bh(connd_lock); + } - ksocknal_thread_fini (); - return (0); + /* 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) { + list_del(&route->ksnr_connd_list); + ksocknal_data.ksnd_connd_connecting++; + 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 %pI4h\n", + &route->ksnr_ipaddr); + cons_retry = 0; + } + } else { + cons_retry = 0; + } + + ksocknal_route_decref(route); + + 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; - struct list_head *ctmp; + struct ksock_conn *conn; + struct list_head *ctmp; + struct ksock_tx *tx; - list_for_each (ctmp, &peer->ksnp_conns) { - conn = list_entry (ctmp, ksock_conn_t, ksnc_list); + list_for_each(ctmp, &peer_ni->ksnp_conns) { + int error; - /* Don't need the {get,put}connsock dance to deref ksnc_sock... */ + 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); - if (SOCK_ERROR(conn->ksnc_sock) != 0) { - atomic_inc (&conn->ksnc_refcount); + error = conn->ksnc_sock->sk->sk_err; + if (error != 0) { + ksocknal_conn_addref(conn); - switch (SOCK_ERROR(conn->ksnc_sock)) { + switch (error) { case ECONNRESET: - LCONSOLE_WARN("A connection with %u.%u.%u.%u " - "was reset; they may have " - "rebooted.\n", - HIPQUAD(conn->ksnc_ipaddr)); + 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: - LCONSOLE_WARN("A connection with %u.%u.%u.%u " - "timed out; the network or that " - "node may be down.\n", - HIPQUAD(conn->ksnc_ipaddr)); + 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: - LCONSOLE_WARN("An unexpected network error " - "occurred with %u.%u.%u.%u: %d\n", - HIPQUAD(conn->ksnc_ipaddr), - SOCK_ERROR(conn->ksnc_sock)); + 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; } - /* Something (e.g. failed keepalive) set the socket error */ - CDEBUG(D_HA,"Socket error %d: "LPX64" %p %d.%d.%d.%d\n", - SOCK_ERROR(conn->ksnc_sock), peer->ksnp_nid, - conn, HIPQUAD(conn->ksnc_ipaddr)); - return (conn); } 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 */ - atomic_inc (&conn->ksnc_refcount); - LCONSOLE_ERROR("A timeout occurred receiving data from " - "%u.%u.%u.%u; the network or that node " - "may be down.\n", - HIPQUAD(conn->ksnc_ipaddr)); - CERROR ("Timed out RX from "LPX64" %p %d.%d.%d.%d\n", - peer->ksnp_nid,conn,HIPQUAD(conn->ksnc_ipaddr)); + ksocknal_conn_addref(conn); + 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 ((!list_empty (&conn->ksnc_tx_queue) || - 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 */ - atomic_inc (&conn->ksnc_refcount); - LCONSOLE_ERROR("A timeout occurred sending data to " - "%u.%u.%u.%u; the network or that node " - "may be down.\n", - HIPQUAD(conn->ksnc_ipaddr)); - CERROR ("Timed out TX to "LPX64" %s%d %p %d.%d.%d.%d\n", - peer->ksnp_nid, - list_empty (&conn->ksnc_tx_queue) ? "" : "Q ", - SOCK_WMEM_QUEUED(conn->ksnc_sock), conn, - HIPQUAD(conn->ksnc_ipaddr)); + ksocknal_conn_addref(conn); + 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); } } @@ -2238,149 +2402,296 @@ ksocknal_find_timed_out_conn (ksock_peer_t *peer) return (NULL); } -void -ksocknal_check_peer_timeouts (int idx) +static inline void +ksocknal_flush_stale_txs(struct ksock_peer_ni *peer_ni) +{ + struct ksock_tx *tx; + LIST_HEAD(stale_txs); + + write_lock_bh(&ksocknal_data.ksnd_global_lock); + + while (!list_empty(&peer_ni->ksnp_tx_queue)) { + tx = list_entry(peer_ni->ksnp_tx_queue.next, + struct ksock_tx, tx_list); + + if (ktime_get_seconds() < tx->tx_deadline) + break; + + tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT; + + list_move_tail(&tx->tx_list, &stale_txs); + } + + write_unlock_bh(&ksocknal_data.ksnd_global_lock); + + ksocknal_txlist_done(peer_ni->ksnp_ni, &stale_txs, -ETIMEDOUT); +} + +static int +ksocknal_send_keepalive_locked(struct ksock_peer_ni *peer_ni) +__must_hold(&ksocknal_data.ksnd_global_lock) { - struct list_head *peers = &ksocknal_data.ksnd_peers[idx]; - struct list_head *ptmp; - ksock_peer_t *peer; - ksock_conn_t *conn; + struct ksock_sched *sched; + struct ksock_conn *conn; + struct ksock_tx *tx; + + /* last_alive will be updated by create_conn */ + if (list_empty(&peer_ni->ksnp_conns)) + return 0; + + if (peer_ni->ksnp_proto != &ksocknal_protocol_v3x) + return 0; + + if (*ksocknal_tunables.ksnd_keepalive <= 0 || + ktime_get_seconds() < peer_ni->ksnp_last_alive + + *ksocknal_tunables.ksnd_keepalive) + return 0; + + if (ktime_get_seconds() < peer_ni->ksnp_send_keepalive) + return 0; + + /* retry 10 secs later, so we wouldn't put pressure + * 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_ni, NULL, 1); + if (conn != NULL) { + sched = conn->ksnc_scheduler; + + 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; + } + + spin_unlock_bh(&sched->kss_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) { + read_lock(&ksocknal_data.ksnd_global_lock); + return -ENOMEM; + } + + 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); + read_lock(&ksocknal_data.ksnd_global_lock); + + return -EIO; +} + + +static void +ksocknal_check_peer_timeouts(int idx) +{ + 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... */ - read_lock (&ksocknal_data.ksnd_global_lock); - - list_for_each (ptmp, peers) { - peer = list_entry (ptmp, ksock_peer_t, ksnp_list); - conn = ksocknal_find_timed_out_conn (peer); - - if (conn != NULL) { - read_unlock (&ksocknal_data.ksnd_global_lock); - - CERROR("Timeout out conn->"LPX64" ip %d.%d.%d.%d:%d\n", - peer->ksnp_nid, HIPQUAD(conn->ksnc_ipaddr), - conn->ksnc_port); - 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_put_conn (conn); - goto 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... + */ + 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 (tx_stale == NULL) { + spin_unlock(&peer_ni->ksnp_lock); + continue; + } + + deadline = tx_stale->tx_deadline; + resid = tx_stale->tx_resid; + conn = tx_stale->tx_conn; + ksocknal_conn_addref(conn); + + spin_unlock(&peer_ni->ksnp_lock); + read_unlock(&ksocknal_data.ksnd_global_lock); + + 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; } - 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; - unsigned long flags; - ksock_conn_t *conn; - ksock_sched_t *sched; - struct list_head 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(); - kportal_daemonize ("ksocknal_reaper"); - kportal_blockallsigs (); + init_waitqueue_entry(&wait, current); - CFS_INIT_LIST_HEAD(&enomem_conns); - cfs_waitlink_init (&wait); - - spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags); + 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 (!list_empty (&ksocknal_data.ksnd_deathrow_conns)) { - conn = list_entry (ksocknal_data.ksnd_deathrow_conns.next, - ksock_conn_t, ksnc_list); - list_del (&conn->ksnc_list); - - spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags); + spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock); - ksocknal_terminate_conn (conn); - ksocknal_put_conn (conn); + ksocknal_terminate_conn(conn); + ksocknal_conn_decref(conn); - spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags); + spin_lock_bh(&ksocknal_data.ksnd_reaper_lock); continue; } - if (!list_empty (&ksocknal_data.ksnd_zombie_conns)) { - conn = list_entry (ksocknal_data.ksnd_zombie_conns.next, - ksock_conn_t, ksnc_list); - 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); - spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags); + spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock); - ksocknal_destroy_conn (conn); + ksocknal_destroy_conn(conn); - spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags); - continue; - } + spin_lock_bh(&ksocknal_data.ksnd_reaper_lock); + continue; + } - if (!list_empty (&ksocknal_data.ksnd_enomem_conns)) { - list_add(&enomem_conns, &ksocknal_data.ksnd_enomem_conns); - list_del_init(&ksocknal_data.ksnd_enomem_conns); - } + list_splice_init(&ksocknal_data.ksnd_enomem_conns, + &enomem_conns); - spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags); + spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock); /* reschedule all the connections that stalled with ENOMEM... */ nenomem_conns = 0; - while (!list_empty (&enomem_conns)) { - conn = list_entry (enomem_conns.next, - ksock_conn_t, ksnc_tx_list); - 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; - spin_lock_irqsave (&sched->kss_lock, flags); + spin_lock_bh(&sched->kss_lock); - LASSERT (conn->ksnc_tx_scheduled); - conn->ksnc_tx_ready = 1; - 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); - spin_unlock_irqrestore (&sched->kss_lock, flags); + 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_io_timeout > n * p) - chunk = (chunk * n * p) / - ksocknal_tunables.ksnd_io_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. @@ -2388,34 +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; - set_current_state (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 && - list_empty (&ksocknal_data.ksnd_deathrow_conns) && - list_empty (&ksocknal_data.ksnd_zombie_conns)) - cfs_waitq_timedwait (&wait, 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)); - set_current_state (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); - spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags); - } + spin_lock_bh(&ksocknal_data.ksnd_reaper_lock); + } - spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags); + spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock); - ksocknal_thread_fini (); - return (0); + ksocknal_thread_fini(); + return 0; } - -lib_nal_t ksocknal_lib = { - libnal_data: &ksocknal_data, /* NAL private data */ - libnal_send: ksocknal_send, - libnal_send_pages: ksocknal_send_pages, - libnal_recv: ksocknal_recv, - libnal_recv_pages: ksocknal_recv_pages, - libnal_dist: ksocknal_dist -};