X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fptlrpc%2Fevents.c;h=660a05c71769442addc2bb64c39261e9987bccc2;hb=917655fc2938b90a9c246dd2d58408c42aa1658d;hp=d865e1f94a3efc12a81fc9e2972ed715073c6250;hpb=37113b4cd285399154f0d8a27df7d4090cc5fee4;p=fs%2Flustre-release.git diff --git a/lustre/ptlrpc/events.c b/lustre/ptlrpc/events.c index d865e1f..660a05c 100644 --- a/lustre/ptlrpc/events.c +++ b/lustre/ptlrpc/events.c @@ -1,182 +1,309 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: +/* + * GPL HEADER START * - * Copyright (c) 2002, 2003 Cluster File Systems, Inc. + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * - * This file is part of Lustre, http://www.lustre.org. + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 only, + * as published by the Free Software Foundation. * - * Lustre is free software; you can redistribute it and/or - * modify it under the terms of version 2 of the GNU General Public - * License as published by the Free Software Foundation. + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * General Public License version 2 for more details (a copy is included + * in the LICENSE file that accompanied this code). * - * Lustre is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. + * You should have received a copy of the GNU General Public License + * version 2 along with this program; If not, see + * http://www.gnu.org/licenses/gpl-2.0.html * - * You should have received a copy of the GNU General Public License - * along with Lustre; if not, write to the Free Software - * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * GPL HEADER END + */ +/* + * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. * + * Copyright (c) 2012, 2017, Intel Corporation. + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. */ #define DEBUG_SUBSYSTEM S_RPC -#ifdef __KERNEL__ -#include -#else -#include -#endif -#include -#include +#include +#include +#include +#include +#include +#include #include "ptlrpc_internal.h" -#if !defined(__KERNEL__) && defined(CRAY_PORTALS) -/* forward ref in events.c */ -static void cray_portals_callback(ptl_event_t *ev); -#endif +struct lnet_handle_eq ptlrpc_eq_h; - -struct ptlrpc_ni ptlrpc_interfaces[NAL_MAX_NR]; -int ptlrpc_ninterfaces; - -/* +/* * Client's outgoing request callback */ -void request_out_callback(ptl_event_t *ev) +void request_out_callback(struct lnet_event *ev) { - struct ptlrpc_cb_id *cbid = ev->mem_desc.user_ptr; - struct ptlrpc_request *req = cbid->cbid_arg; - unsigned long flags; - ENTRY; + struct ptlrpc_cb_id *cbid = ev->md.user_ptr; + struct ptlrpc_request *req = cbid->cbid_arg; + bool wakeup = false; + ENTRY; - LASSERT (ev->type == PTL_EVENT_SEND_END || - ev->type == PTL_EVENT_UNLINK); - LASSERT (ev->unlinked); + LASSERT(ev->type == LNET_EVENT_SEND || ev->type == LNET_EVENT_UNLINK); + LASSERT(ev->unlinked); - DEBUG_REQ((ev->ni_fail_type == PTL_NI_OK) ? D_NET : D_ERROR, req, - "type %d, status %d", ev->type, ev->ni_fail_type); + DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status); - if (ev->type == PTL_EVENT_UNLINK || - ev->ni_fail_type != PTL_NI_OK) { + sptlrpc_request_out_callback(req); - /* Failed send: make it seem like the reply timed out, just - * like failing sends in client.c does currently... */ + spin_lock(&req->rq_lock); + req->rq_real_sent = ktime_get_real_seconds(); + req->rq_req_unlinked = 1; + /* reply_in_callback happened before request_out_callback? */ + if (req->rq_reply_unlinked) + wakeup = true; - spin_lock_irqsave(&req->rq_lock, flags); - req->rq_net_err = 1; - spin_unlock_irqrestore(&req->rq_lock, flags); - - ptlrpc_wake_client_req(req); - } + if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) { + /* Failed send: make it seem like the reply timed out, just + * like failing sends in client.c does currently... */ + req->rq_net_err = 1; + wakeup = true; + } - /* this balances the atomic_inc in ptl_send_rpc() */ - ptlrpc_req_finished(req); - EXIT; + if (wakeup) + ptlrpc_client_wake_req(req); + + spin_unlock(&req->rq_lock); + + ptlrpc_req_finished(req); + EXIT; } /* * Client's incoming reply callback */ -void reply_in_callback(ptl_event_t *ev) +void reply_in_callback(struct lnet_event *ev) { - struct ptlrpc_cb_id *cbid = ev->mem_desc.user_ptr; + struct ptlrpc_cb_id *cbid = ev->md.user_ptr; struct ptlrpc_request *req = cbid->cbid_arg; - unsigned long flags; ENTRY; - LASSERT (ev->type == PTL_EVENT_PUT_END || - ev->type == PTL_EVENT_UNLINK); - LASSERT (ev->unlinked); - LASSERT (ev->mem_desc.start == req->rq_repmsg); - LASSERT (ev->offset == 0); - LASSERT (ev->mlength <= req->rq_replen); - - DEBUG_REQ((ev->ni_fail_type == PTL_NI_OK) ? D_NET : D_ERROR, req, - "type %d, status %d", ev->type, ev->ni_fail_type); + DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status); + + LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK); + LASSERT (ev->md.start == req->rq_repbuf); + LASSERT (ev->offset + ev->mlength <= req->rq_repbuf_len); + /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests + for adaptive timeouts' early reply. */ + LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0); + + spin_lock(&req->rq_lock); - spin_lock_irqsave (&req->rq_lock, flags); + req->rq_receiving_reply = 0; + req->rq_early = 0; + if (ev->unlinked) + req->rq_reply_unlinked = 1; - LASSERT (req->rq_receiving_reply); - req->rq_receiving_reply = 0; + if (ev->status) + goto out_wake; - if (ev->type == PTL_EVENT_PUT_END && - ev->ni_fail_type == PTL_NI_OK) { + if (ev->type == LNET_EVENT_UNLINK) { + LASSERT(ev->unlinked); + DEBUG_REQ(D_NET, req, "unlink"); + goto out_wake; + } + + if (ev->mlength < ev->rlength ) { + CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req, + req->rq_replen, ev->rlength, ev->offset); + req->rq_reply_truncated = 1; req->rq_replied = 1; + req->rq_status = -EOVERFLOW; + req->rq_nob_received = ev->rlength + ev->offset; + goto out_wake; + } + + if ((ev->offset == 0) && + ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) { + /* Early reply */ + DEBUG_REQ(D_ADAPTTO, req, + "Early reply received, mlen=%u offset=%d replen=%d replied=%d unlinked=%d", + ev->mlength, ev->offset, + req->rq_replen, req->rq_replied, ev->unlinked); + + req->rq_early_count++; /* number received, client side */ + + /* already got the real reply or buffers are already unlinked */ + if (req->rq_replied || + req->rq_reply_unlinked == 1) + goto out_wake; + + req->rq_early = 1; + req->rq_reply_off = ev->offset; req->rq_nob_received = ev->mlength; + /* And we're still receiving */ + req->rq_receiving_reply = 1; + } else { + /* Real reply */ + req->rq_rep_swab_mask = 0; + req->rq_replied = 1; + /* Got reply, no resend required */ + req->rq_resend = 0; + req->rq_reply_off = ev->offset; + req->rq_nob_received = ev->mlength; + /* LNetMDUnlink can't be called under the LNET_LOCK, + so we must unlink in ptlrpc_unregister_reply */ + DEBUG_REQ(D_INFO, req, + "reply in flags=%x mlen=%u offset=%d replen=%d", + lustre_msg_get_flags(req->rq_reqmsg), + ev->mlength, ev->offset, req->rq_replen); } - /* NB don't unlock till after wakeup; req can disappear under us - * since we don't have our own ref */ - ptlrpc_wake_client_req(req); + if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING) + req->rq_import->imp_last_reply_time = ktime_get_real_seconds(); - spin_unlock_irqrestore (&req->rq_lock, flags); - EXIT; +out_wake: + /* NB don't unlock till after wakeup; req can disappear under us + * since we don't have our own ref */ + ptlrpc_client_wake_req(req); + spin_unlock(&req->rq_lock); + EXIT; } -/* +/* * Client's bulk has been written/read */ -void client_bulk_callback (ptl_event_t *ev) +void client_bulk_callback(struct lnet_event *ev) { - struct ptlrpc_cb_id *cbid = ev->mem_desc.user_ptr; + struct ptlrpc_cb_id *cbid = ev->md.user_ptr; struct ptlrpc_bulk_desc *desc = cbid->cbid_arg; - unsigned long flags; + struct ptlrpc_request *req; ENTRY; - LASSERT ((desc->bd_type == BULK_PUT_SINK && - ev->type == PTL_EVENT_PUT_END) || - (desc->bd_type == BULK_GET_SOURCE && - ev->type == PTL_EVENT_GET_END) || - ev->type == PTL_EVENT_UNLINK); - LASSERT (ev->unlinked); - - CDEBUG((ev->ni_fail_type == PTL_NI_OK) ? D_NET : D_ERROR, - "event type %d, status %d, desc %p\n", - ev->type, ev->ni_fail_type, desc); - - spin_lock_irqsave (&desc->bd_lock, flags); + LASSERT((ptlrpc_is_bulk_put_sink(desc->bd_type) && + ev->type == LNET_EVENT_PUT) || + (ptlrpc_is_bulk_get_source(desc->bd_type) && + ev->type == LNET_EVENT_GET) || + ev->type == LNET_EVENT_UNLINK); + LASSERT(ev->unlinked); + + if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE)) + ev->status = -EIO; + + if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE)) + ev->status = -EIO; + + CDEBUG((ev->status == 0) ? D_NET : D_ERROR, + "event type %d, status %d, desc %p\n", + ev->type, ev->status, desc); + + spin_lock(&desc->bd_lock); + req = desc->bd_req; + LASSERT(desc->bd_md_count > 0); + desc->bd_md_count--; + + if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) { + desc->bd_nob_transferred += ev->mlength; + desc->bd_sender = ev->sender; + } else { + /* start reconnect and resend if network error hit */ + spin_lock(&req->rq_lock); + req->rq_net_err = 1; + spin_unlock(&req->rq_lock); + } + + if (ev->status != 0) + desc->bd_failure = 1; + + /* NB don't unlock till after wakeup; desc can disappear under us + * otherwise */ + if (desc->bd_md_count == 0) + ptlrpc_client_wake_req(desc->bd_req); + + spin_unlock(&desc->bd_lock); + EXIT; +} - LASSERT(desc->bd_network_rw); - desc->bd_network_rw = 0; +/* + * We will have percpt request history list for ptlrpc service in upcoming + * patches because we don't want to be serialized by current per-service + * history operations. So we require history ID can (somehow) show arriving + * order w/o grabbing global lock, and user can sort them in userspace. + * + * This is how we generate history ID for ptlrpc_request: + * ---------------------------------------------------- + * | 32 bits | 16 bits | (16 - X)bits | X bits | + * ---------------------------------------------------- + * | seconds | usec / 16 | sequence | CPT id | + * ---------------------------------------------------- + * + * it might not be precise but should be good enough. + */ - if (ev->type != PTL_EVENT_UNLINK && - ev->ni_fail_type == PTL_NI_OK) { - desc->bd_success = 1; - desc->bd_nob_transferred = ev->mlength; - } +#define REQS_CPT_BITS(svcpt) ((svcpt)->scp_service->srv_cpt_bits) - /* NB don't unlock till after wakeup; desc can disappear under us - * otherwise */ - ptlrpc_wake_client_req(desc->bd_req); +#define REQS_SEC_SHIFT 32 +#define REQS_USEC_SHIFT 16 +#define REQS_SEQ_SHIFT(svcpt) REQS_CPT_BITS(svcpt) - spin_unlock_irqrestore (&desc->bd_lock, flags); - EXIT; +static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt, + struct ptlrpc_request *req) +{ + u64 sec = req->rq_arrival_time.tv_sec; + u32 usec = req->rq_arrival_time.tv_nsec / NSEC_PER_USEC / 16; /* usec / 16 */ + u64 new_seq; + + /* set sequence ID for request and add it to history list, + * it must be called with hold svcpt::scp_lock */ + + new_seq = (sec << REQS_SEC_SHIFT) | + (usec << REQS_USEC_SHIFT) | + (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt); + + if (new_seq > svcpt->scp_hist_seq) { + /* This handles the initial case of scp_hist_seq == 0 or + * we just jumped into a new time window */ + svcpt->scp_hist_seq = new_seq; + } else { + LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT); + /* NB: increase sequence number in current usec bucket, + * however, it's possible that we used up all bits for + * sequence and jumped into the next usec bucket (future time), + * then we hope there will be less RPCs per bucket at some + * point, and sequence will catch up again */ + svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt)); + new_seq = svcpt->scp_hist_seq; + } + + req->rq_history_seq = new_seq; + + list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs); } -/* +/* * Server's incoming request callback */ -void request_in_callback(ptl_event_t *ev) +void request_in_callback(struct lnet_event *ev) { - struct ptlrpc_cb_id *cbid = ev->mem_desc.user_ptr; - struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg; - struct ptlrpc_srv_ni *srv_ni = rqbd->rqbd_srv_ni; - struct ptlrpc_service *service = srv_ni->sni_service; + struct ptlrpc_cb_id *cbid = ev->md.user_ptr; + struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg; + struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt; + struct ptlrpc_service *service = svcpt->scp_service; struct ptlrpc_request *req; - char str[PTL_NALFMT_SIZE]; - unsigned long flags; ENTRY; - LASSERT (ev->type == PTL_EVENT_PUT_END || - ev->type == PTL_EVENT_UNLINK); - LASSERT ((char *)ev->mem_desc.start >= rqbd->rqbd_buffer); - LASSERT ((char *)ev->mem_desc.start + ev->offset + ev->mlength <= + LASSERT (ev->type == LNET_EVENT_PUT || + ev->type == LNET_EVENT_UNLINK); + LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer); + LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <= rqbd->rqbd_buffer + service->srv_buf_size); - CDEBUG((ev->ni_fail_type == PTL_OK) ? D_NET : D_ERROR, - "event type %d, status %d, service %s\n", - ev->type, ev->ni_fail_type, service->srv_name); + CDEBUG((ev->status == 0) ? D_NET : D_ERROR, + "event type %d, status %d, service %s\n", + ev->type, ev->status, service->srv_name); if (ev->unlinked) { /* If this is the last request message to fit in the @@ -187,87 +314,96 @@ void request_in_callback(ptl_event_t *ev) req = &rqbd->rqbd_req; memset(req, 0, sizeof (*req)); } else { - LASSERT (ev->type == PTL_EVENT_PUT_END); - if (ev->ni_fail_type != PTL_NI_OK) { + LASSERT (ev->type == LNET_EVENT_PUT); + if (ev->status != 0) { /* We moaned above already... */ return; } - OBD_ALLOC_GFP(req, sizeof(*req), GFP_ATOMIC); + req = ptlrpc_request_cache_alloc(GFP_ATOMIC); if (req == NULL) { CERROR("Can't allocate incoming request descriptor: " "Dropping %s RPC from %s\n", - service->srv_name, - portals_nid2str(srv_ni->sni_ni->pni_number, - ev->initiator.nid, str)); + service->srv_name, + libcfs_id2str(ev->initiator)); return; } } - /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL, - * flags are reset and scalars are zero. We only set the message - * size to non-zero if this was a successful receive. */ - req->rq_xid = ev->match_bits; - req->rq_reqmsg = ev->mem_desc.start + ev->offset; - if (ev->type == PTL_EVENT_PUT_END && - ev->ni_fail_type == PTL_NI_OK) - req->rq_reqlen = ev->mlength; - do_gettimeofday(&req->rq_arrival_time); - req->rq_peer.peer_nid = ev->initiator.nid; - req->rq_peer.peer_ni = rqbd->rqbd_srv_ni->sni_ni; - req->rq_rqbd = rqbd; - - spin_lock_irqsave (&service->srv_lock, flags); + ptlrpc_srv_req_init(req); + /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL, + * flags are reset and scalars are zero. We only set the message + * size to non-zero if this was a successful receive. */ + req->rq_xid = ev->match_bits; + req->rq_reqbuf = ev->md.start + ev->offset; + if (ev->type == LNET_EVENT_PUT && ev->status == 0) + req->rq_reqdata_len = ev->mlength; + ktime_get_real_ts64(&req->rq_arrival_time); + /* Multi-Rail: keep track of both initiator and source NID. */ + req->rq_peer = ev->initiator; + req->rq_source = ev->source; + req->rq_self = ev->target.nid; + req->rq_rqbd = rqbd; + req->rq_phase = RQ_PHASE_NEW; + if (ev->type == LNET_EVENT_PUT) + CDEBUG(D_INFO, "incoming req@%p x%llu msgsize %u\n", + req, req->rq_xid, ev->mlength); + + CDEBUG(D_RPCTRACE, "peer: %s (source: %s)\n", + libcfs_id2str(req->rq_peer), libcfs_id2str(req->rq_source)); + + spin_lock(&svcpt->scp_lock); + + ptlrpc_req_add_history(svcpt, req); + + if (ev->unlinked) { + svcpt->scp_nrqbds_posted--; + CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n", + svcpt->scp_nrqbds_posted); + + /* Normally, don't complain about 0 buffers posted; LNET won't + * drop incoming reqs since we set the portal lazy */ + if (test_req_buffer_pressure && + ev->type != LNET_EVENT_UNLINK && + svcpt->scp_nrqbds_posted == 0) + CWARN("All %s request buffers busy\n", + service->srv_name); - if (ev->unlinked) { - srv_ni->sni_nrqbd_receiving--; - if (ev->type != PTL_EVENT_UNLINK && - srv_ni->sni_nrqbd_receiving == 0) { - /* This service is off-air on this interface because - * all its request buffers are busy. Portals will - * start dropping incoming requests until more buffers - * get posted. NB don't moan if it's because we're - * tearing down the service. */ - CWARN("All %s %s request buffers busy\n", - service->srv_name, srv_ni->sni_ni->pni_name); - } /* req takes over the network's ref on rqbd */ } else { /* req takes a ref on rqbd */ rqbd->rqbd_refcount++; } - list_add_tail(&req->rq_list, &service->srv_request_queue); - service->srv_n_queued_reqs++; + list_add_tail(&req->rq_list, &svcpt->scp_req_incoming); + svcpt->scp_nreqs_incoming++; - /* NB everything can disappear under us once the request - * has been queued and we unlock, so do the wake now... */ - wake_up(&service->srv_waitq); + /* NB everything can disappear under us once the request + * has been queued and we unlock, so do the wake now... */ + wake_up(&svcpt->scp_waitq); - spin_unlock_irqrestore(&service->srv_lock, flags); - EXIT; + spin_unlock(&svcpt->scp_lock); + EXIT; } -/* +/* * Server's outgoing reply callback */ -void reply_out_callback(ptl_event_t *ev) +void reply_out_callback(struct lnet_event *ev) { - struct ptlrpc_cb_id *cbid = ev->mem_desc.user_ptr; - struct ptlrpc_reply_state *rs = cbid->cbid_arg; - struct ptlrpc_srv_ni *sni = rs->rs_srv_ni; - struct ptlrpc_service *svc = sni->sni_service; - unsigned long flags; + struct ptlrpc_cb_id *cbid = ev->md.user_ptr; + struct ptlrpc_reply_state *rs = cbid->cbid_arg; + struct ptlrpc_service_part *svcpt = rs->rs_svcpt; ENTRY; - LASSERT (ev->type == PTL_EVENT_SEND_END || - ev->type == PTL_EVENT_ACK || - ev->type == PTL_EVENT_UNLINK); + LASSERT (ev->type == LNET_EVENT_SEND || + ev->type == LNET_EVENT_ACK || + ev->type == LNET_EVENT_UNLINK); if (!rs->rs_difficult) { - /* I'm totally responsible for freeing "easy" replies */ + /* 'Easy' replies have no further processing so I drop the + * net's ref on 'rs' */ LASSERT (ev->unlinked); - lustre_free_reply_state (rs); - atomic_dec (&svc->srv_outstanding_replies); + ptlrpc_rs_decref(rs); EXIT; return; } @@ -275,63 +411,78 @@ void reply_out_callback(ptl_event_t *ev) LASSERT (rs->rs_on_net); if (ev->unlinked) { - /* Last network callback */ - spin_lock_irqsave (&svc->srv_lock, flags); - rs->rs_on_net = 0; - ptlrpc_schedule_difficult_reply (rs); - spin_unlock_irqrestore (&svc->srv_lock, flags); - } - - EXIT; + /* Last network callback. The net's ref on 'rs' stays put + * until ptlrpc_handle_rs() is done with it */ + spin_lock(&svcpt->scp_rep_lock); + spin_lock(&rs->rs_lock); + + rs->rs_on_net = 0; + if (!rs->rs_no_ack || + rs->rs_transno <= + rs->rs_export->exp_obd->obd_last_committed || + list_empty(&rs->rs_obd_list)) + ptlrpc_schedule_difficult_reply(rs); + + spin_unlock(&rs->rs_lock); + spin_unlock(&svcpt->scp_rep_lock); + } + EXIT; } +#ifdef HAVE_SERVER_SUPPORT /* * Server's bulk completion callback */ -void server_bulk_callback (ptl_event_t *ev) +void server_bulk_callback(struct lnet_event *ev) { - struct ptlrpc_cb_id *cbid = ev->mem_desc.user_ptr; - struct ptlrpc_bulk_desc *desc = cbid->cbid_arg; - unsigned long flags; - ENTRY; - - LASSERT (ev->type == PTL_EVENT_SEND_END || - ev->type == PTL_EVENT_UNLINK || - (desc->bd_type == BULK_PUT_SOURCE && - ev->type == PTL_EVENT_ACK) || - (desc->bd_type == BULK_GET_SINK && - ev->type == PTL_EVENT_REPLY_END)); - - CDEBUG((ev->ni_fail_type == PTL_NI_OK) ? D_NET : D_ERROR, - "event type %d, status %d, desc %p\n", - ev->type, ev->ni_fail_type, desc); - - spin_lock_irqsave (&desc->bd_lock, flags); - - if ((ev->type == PTL_EVENT_ACK || - ev->type == PTL_EVENT_REPLY_END) && - ev->ni_fail_type == PTL_NI_OK) { - /* We heard back from the peer, so even if we get this - * before the SENT event (oh yes we can), we know we - * read/wrote the peer buffer and how much... */ - desc->bd_success = 1; - desc->bd_nob_transferred = ev->mlength; - } - - if (ev->unlinked) { - /* This is the last callback no matter what... */ - desc->bd_network_rw = 0; - wake_up(&desc->bd_waitq); - } - - spin_unlock_irqrestore (&desc->bd_lock, flags); - EXIT; + struct ptlrpc_cb_id *cbid = ev->md.user_ptr; + struct ptlrpc_bulk_desc *desc = cbid->cbid_arg; + ENTRY; + + LASSERT(ev->type == LNET_EVENT_SEND || + ev->type == LNET_EVENT_UNLINK || + (ptlrpc_is_bulk_put_source(desc->bd_type) && + ev->type == LNET_EVENT_ACK) || + (ptlrpc_is_bulk_get_sink(desc->bd_type) && + ev->type == LNET_EVENT_REPLY)); + + CDEBUG((ev->status == 0) ? D_NET : D_ERROR, + "event type %d, status %d, desc %p\n", + ev->type, ev->status, desc); + + spin_lock(&desc->bd_lock); + + LASSERT(desc->bd_md_count > 0); + + if ((ev->type == LNET_EVENT_ACK || + ev->type == LNET_EVENT_REPLY) && + ev->status == 0) { + /* We heard back from the peer, so even if we get this + * before the SENT event (oh yes we can), we know we + * read/wrote the peer buffer and how much... */ + desc->bd_nob_transferred += ev->mlength; + desc->bd_sender = ev->sender; + } + + if (ev->status != 0) + desc->bd_failure = 1; + + if (ev->unlinked) { + desc->bd_md_count--; + /* This is the last callback no matter what... */ + if (desc->bd_md_count == 0) + wake_up(&desc->bd_waitq); + } + + spin_unlock(&desc->bd_lock); + EXIT; } +#endif -static void ptlrpc_master_callback(ptl_event_t *ev) +static void ptlrpc_master_callback(struct lnet_event *ev) { - struct ptlrpc_cb_id *cbid = ev->mem_desc.user_ptr; - void (*callback)(ptl_event_t *ev) = cbid->cbid_fn; + struct ptlrpc_cb_id *cbid = ev->md.user_ptr; + void (*callback)(struct lnet_event *ev) = cbid->cbid_fn; /* Honestly, it's best to find out early. */ LASSERT (cbid->cbid_arg != LP_POISON); @@ -339,302 +490,149 @@ static void ptlrpc_master_callback(ptl_event_t *ev) callback == reply_in_callback || callback == client_bulk_callback || callback == request_in_callback || - callback == reply_out_callback || - callback == server_bulk_callback); - - callback (ev); -} - -int ptlrpc_uuid_to_peer (struct obd_uuid *uuid, struct ptlrpc_peer *peer) -{ - struct ptlrpc_ni *pni; - __u32 peer_nal; - ptl_nid_t peer_nid; - int i; - char str[PTL_NALFMT_SIZE]; - int rc = lustre_uuid_to_peer(uuid->uuid, - &peer_nal, &peer_nid); - if (rc != 0) - RETURN (rc); - - for (i = 0; i < ptlrpc_ninterfaces; i++) { - pni = &ptlrpc_interfaces[i]; - - if (pni->pni_number == peer_nal) { - peer->peer_nid = peer_nid; - peer->peer_ni = pni; - return (0); - } - } - - CERROR("Can't find ptlrpc interface for NAL %d, NID %s\n", - peer_nal, portals_nid2str(peer_nal, peer_nid, str)); - return (-ENOENT); -} - -void ptlrpc_ni_fini(struct ptlrpc_ni *pni) -{ - wait_queue_head_t waitq; - struct l_wait_info lwi; - int rc; - int retries; - - /* Wait for the event queue to become idle since there may still be - * messages in flight with pending events (i.e. the fire-and-forget - * messages == client requests and "non-difficult" server - * replies */ - - for (retries = 0;; retries++) { - rc = PtlEQFree(pni->pni_eq_h); - switch (rc) { - default: - LBUG(); - - case PTL_OK: - PtlNIFini(pni->pni_ni_h); - return; - - case PTL_EQ_IN_USE: - if (retries != 0) - CWARN("Event queue for %s still busy\n", - pni->pni_name); - - /* Wait for a bit */ - init_waitqueue_head(&waitq); - lwi = LWI_TIMEOUT(2*HZ, NULL, NULL); - l_wait_event(waitq, 0, &lwi); - break; - } - } - /* notreached */ -} - -int ptlrpc_ni_init(int number, char *name, struct ptlrpc_ni *pni) -{ - int rc; - char str[20]; - ptl_handle_ni_t nih; - - /* We're not passing any limits yet... */ - rc = PtlNIInit(number, 0, NULL, NULL, &nih); - if (rc != PTL_OK && rc != PTL_IFACE_DUP) { - CDEBUG (D_NET, "Can't init network interface %s: %d\n", - name, rc); - return (-ENOENT); - } - - PtlSnprintHandle(str, sizeof(str), nih); - CDEBUG (D_NET, "init %d %s: %s\n", number, name, str); - - pni->pni_name = name; - pni->pni_number = number; - pni->pni_ni_h = nih; - - pni->pni_eq_h = PTL_INVALID_HANDLE; - - /* CAVEAT EMPTOR: how we process portals events is _radically_ - * different depending on... */ -#ifdef __KERNEL__ - /* kernel portals calls our master callback when events are added to - * the event queue. In fact lustre never pulls events off this queue, - * so it's only sized for some debug history. */ - rc = PtlEQAlloc(pni->pni_ni_h, 1024, ptlrpc_master_callback, - &pni->pni_eq_h); -#else - /* liblustre calls the master callback when it removes events from the - * event queue. The event queue has to be big enough not to drop - * anything */ -# if CRAY_PORTALS - /* cray portals implements a non-standard callback to notify us there - * are buffered events even when the app is not doing a filesystem - * call. */ - rc = PtlEQAlloc(pni->pni_ni_h, 10240, cray_portals_callback, - &pni->pni_eq_h); -# else - rc = PtlEQAlloc(pni->pni_ni_h, 10240, PTL_EQ_HANDLER_NONE, - &pni->pni_eq_h); -# endif + callback == reply_out_callback +#ifdef HAVE_SERVER_SUPPORT + || callback == server_bulk_callback #endif - if (rc != PTL_OK) - GOTO (fail, rc = -ENOMEM); - - return (0); - fail: - CERROR ("Failed to initialise network interface %s: %d\n", - name, rc); + ); - /* OK to do complete teardown since we invalidated the handles above */ - ptlrpc_ni_fini (pni); - return (rc); + callback (ev); } -#ifndef __KERNEL__ -LIST_HEAD(liblustre_wait_callbacks); -void *liblustre_services_callback; - -void * -liblustre_register_wait_callback (int (*fn)(void *arg), void *arg) +int ptlrpc_uuid_to_peer(struct obd_uuid *uuid, + struct lnet_process_id *peer, lnet_nid_t *self) { - struct liblustre_wait_callback *llwc; - - OBD_ALLOC(llwc, sizeof(*llwc)); - LASSERT (llwc != NULL); - - llwc->llwc_fn = fn; - llwc->llwc_arg = arg; - list_add_tail(&llwc->llwc_list, &liblustre_wait_callbacks); - - return (llwc); + int best_dist = 0; + __u32 best_order = 0; + int count = 0; + int rc = -ENOENT; + int dist; + __u32 order; + lnet_nid_t dst_nid; + lnet_nid_t src_nid; + + peer->pid = LNET_PID_LUSTRE; + + /* Choose the matching UUID that's closest */ + while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) { + if (peer->nid != LNET_NID_ANY && LNET_NIDADDR(peer->nid) == 0 && + LNET_NIDNET(dst_nid) != LNET_NIDNET(peer->nid)) + continue; + + dist = LNetDist(dst_nid, &src_nid, &order); + if (dist < 0) + continue; + + if (dist == 0) { /* local! use loopback LND */ + peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0); + rc = 0; + break; + } + + if (rc < 0 || + dist < best_dist || + (dist == best_dist && order < best_order)) { + best_dist = dist; + best_order = order; + + peer->nid = dst_nid; + *self = src_nid; + rc = 0; + } + } + + CDEBUG(D_NET, "%s->%s\n", uuid->uuid, libcfs_id2str(*peer)); + return rc; } -void -liblustre_deregister_wait_callback (void *opaque) +void ptlrpc_ni_fini(void) { - struct liblustre_wait_callback *llwc = opaque; - - list_del(&llwc->llwc_list); - OBD_FREE(llwc, sizeof(*llwc)); + int rc; + int retries; + + /* Wait for the event queue to become idle since there may still be + * messages in flight with pending events (i.e. the fire-and-forget + * messages == client requests and "non-difficult" server + * replies */ + + for (retries = 0;; retries++) { + rc = LNetEQFree(ptlrpc_eq_h); + switch (rc) { + default: + LBUG(); + + case 0: + LNetNIFini(); + return; + + case -EBUSY: + if (retries != 0) + CWARN("Event queue still busy\n"); + + /* Wait for a bit */ + ssleep(2); + break; + } + } + /* notreached */ } -int -liblustre_check_events (int timeout) +lnet_pid_t ptl_get_pid(void) { - ptl_event_t ev; - int rc; - int i; - ENTRY; - - rc = PtlEQPoll(&ptlrpc_interfaces[0].pni_eq_h, 1, timeout * 1000, - &ev, &i); - if (rc == PTL_EQ_EMPTY) - RETURN(0); - - LASSERT (rc == PTL_EQ_DROPPED || rc == PTL_OK); - - /* liblustre: no asynch callback so we can't affort to miss any - * events... */ - if (rc == PTL_EQ_DROPPED) { - CERROR ("Dropped an event!!!\n"); - abort(); - } - - ptlrpc_master_callback (&ev); - RETURN(1); + return LNET_PID_LUSTRE; } -int liblustre_waiting = 0; - -int -liblustre_wait_event (int timeout) +int ptlrpc_ni_init(void) { - struct list_head *tmp; - struct liblustre_wait_callback *llwc; - int found_something = 0; - - /* single threaded recursion check... */ - liblustre_waiting = 1; - - for (;;) { - /* Deal with all pending events */ - while (liblustre_check_events(0)) - found_something = 1; - - /* Give all registered callbacks a bite at the cherry */ - list_for_each(tmp, &liblustre_wait_callbacks) { - llwc = list_entry(tmp, struct liblustre_wait_callback, - llwc_list); - - if (llwc->llwc_fn(llwc->llwc_arg)) - found_something = 1; - } + int rc; + lnet_pid_t pid; - if (found_something || timeout == 0) - break; + pid = ptl_get_pid(); + CDEBUG(D_NET, "My pid is: %x\n", pid); - /* Nothing so far, but I'm allowed to block... */ - found_something = liblustre_check_events(timeout); - if (!found_something) /* still nothing */ - break; /* I timed out */ + /* We're not passing any limits yet... */ + rc = LNetNIInit(pid); + if (rc < 0) { + CDEBUG (D_NET, "Can't init network interface: %d\n", rc); + return rc; } - liblustre_waiting = 0; - - return found_something; -} - -#ifdef CRAY_PORTALS -static void cray_portals_callback(ptl_event_t *ev) -{ - /* We get a callback from the client Cray portals implementation - * whenever anyone calls PtlEQPoll(), and an event queue with a - * callback handler has outstanding events. - * - * If it's not liblustre calling PtlEQPoll(), this lets us know we - * have outstanding events which we handle with - * liblustre_wait_event(). - * - * Otherwise, we're already eagerly consuming events and we'd - * handle events out of order if we recursed. */ - if (!liblustre_waiting) - liblustre_wait_event(0); + /* CAVEAT EMPTOR: how we process portals events is _radically_ + * different depending on... */ + /* kernel LNet calls our master callback when there are new event, + * because we are guaranteed to get every event via callback, + * so we just set EQ size to 0 to avoid overhread of serializing + * enqueue/dequeue operations in LNet. */ + rc = LNetEQAlloc(0, ptlrpc_master_callback, &ptlrpc_eq_h); + if (rc == 0) + return 0; + + CERROR ("Failed to allocate event queue: %d\n", rc); + LNetNIFini(); + + return rc; } -#endif -#endif /* __KERNEL__ */ - -int ptlrpc_default_nal(void) -{ - if (ptlrpc_ninterfaces == 0) - return (-ENOENT); - return (ptlrpc_interfaces[0].pni_number); -} int ptlrpc_init_portals(void) { - /* Add new portals network interfaces here. - * Order is irrelevent! */ - static struct { - int number; - char *name; - } ptl_nis[] = { - {QSWNAL, "qswnal"}, - {SOCKNAL, "socknal"}, - {GMNAL, "gmnal"}, - {IBNAL, "ibnal"}, - {TCPNAL, "tcpnal"}, - {CRAY_KB_ERNAL, "cray_kb_ernal"}}; - int rc; - int i; - - LASSERT(ptlrpc_ninterfaces == 0); - - for (i = 0; i < sizeof (ptl_nis) / sizeof (ptl_nis[0]); i++) { - LASSERT(ptlrpc_ninterfaces < (sizeof(ptlrpc_interfaces) / - sizeof(ptlrpc_interfaces[0]))); - - rc = ptlrpc_ni_init(ptl_nis[i].number, ptl_nis[i].name, - &ptlrpc_interfaces[ptlrpc_ninterfaces]); - if (rc == 0) - ptlrpc_ninterfaces++; - } + int rc = ptlrpc_ni_init(); - if (ptlrpc_ninterfaces == 0) { - CERROR("network initialisation failed: is a NAL module " - "loaded?\n"); - return -EIO; + if (rc != 0) { + CERROR("network initialisation failed\n"); + return rc; } -#ifndef __KERNEL__ - liblustre_services_callback = - liblustre_register_wait_callback(&liblustre_check_services, NULL); -#endif - return 0; + rc = ptlrpcd_addref(); + if (rc == 0) + return 0; + + CERROR("rpcd initialisation failed\n"); + ptlrpc_ni_fini(); + return rc; } void ptlrpc_exit_portals(void) { -#ifndef __KERNEL__ - liblustre_deregister_wait_callback(liblustre_services_callback); -#endif - while (ptlrpc_ninterfaces > 0) - ptlrpc_ni_fini (&ptlrpc_interfaces[--ptlrpc_ninterfaces]); + ptlrpcd_decref(); + ptlrpc_ni_fini(); }