X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fptlrpc%2Fniobuf.c;h=c560e72011e8d9f7fff31c099fb8f82f843e106d;hp=ab6684a8b4c4bddf9727a4008bf5040dc4536f37;hb=c56f7675bfb17b3847b38044153e86550ced3c8e;hpb=0343ecb7de2dae4cf8016416bf1af5d34e9a746d diff --git a/lustre/ptlrpc/niobuf.c b/lustre/ptlrpc/niobuf.c index ab6684a..c560e72 100644 --- a/lustre/ptlrpc/niobuf.c +++ b/lustre/ptlrpc/niobuf.c @@ -1,762 +1,912 @@ -/* -*- 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.sun.com/software/products/lustre/docs/GPLv2.pdf * - * 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. + * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, + * CA 95054 USA or visit www.sun.com if you need additional information or + * have any questions. * + * GPL HEADER END + */ +/* + * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. + * + * Copyright (c) 2012, 2014, 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 -#ifndef __KERNEL__ -#include -#include -#endif -#include -#include -#include -#include +#include +#include +#include +#include +#include #include "ptlrpc_internal.h" -static int ptl_send_buf(struct ptlrpc_request *request, - struct ptlrpc_connection *conn, int portal) +/** + * Helper function. Sends \a len bytes from \a base at offset \a offset + * over \a conn connection to portal \a portal. + * Returns 0 on success or error code. + */ +static int ptl_send_buf (lnet_handle_md_t *mdh, void *base, int len, + lnet_ack_req_t ack, struct ptlrpc_cb_id *cbid, + struct ptlrpc_connection *conn, int portal, __u64 xid, + unsigned int offset) { - int rc; - int rc2; - ptl_process_id_t remote_id; - ptl_handle_md_t md_h; - ptl_ack_req_t ack_req; - char str[PTL_NALFMT_SIZE]; + int rc; + lnet_md_t md; + ENTRY; LASSERT (portal != 0); LASSERT (conn != NULL); - CDEBUG (D_INFO, "conn=%p ni %s nid "LPX64" (%s) on %s\n", - conn, conn->c_peer.peer_ni->pni_name, - conn->c_peer.peer_nid, - portals_nid2str(conn->c_peer.peer_ni->pni_number, - conn->c_peer.peer_nid, str), - conn->c_peer.peer_ni->pni_name); - - request->rq_req_md.user_ptr = request; - - switch (request->rq_type) { - case PTL_RPC_MSG_REQUEST: - request->rq_reqmsg->type = request->rq_type; - request->rq_req_md.start = request->rq_reqmsg; - request->rq_req_md.length = request->rq_reqlen; - request->rq_req_md.eventq = - conn->c_peer.peer_ni->pni_request_out_eq_h; - LASSERT (!request->rq_want_ack); - break; - case PTL_RPC_MSG_ERR: - case PTL_RPC_MSG_REPLY: - request->rq_repmsg->type = request->rq_type; - request->rq_req_md.start = request->rq_repmsg; - request->rq_req_md.length = request->rq_replen; - request->rq_req_md.eventq = - conn->c_peer.peer_ni->pni_reply_out_eq_h; - break; - default: - LBUG(); - return -1; /* notreached */ - } - if (request->rq_want_ack) { - request->rq_req_md.threshold = 2; /* SENT and ACK */ - ack_req = PTL_ACK_REQ; - } else { - request->rq_req_md.threshold = 1; - ack_req = PTL_NOACK_REQ; - } - request->rq_req_md.options = PTL_MD_OP_PUT; - request->rq_req_md.user_ptr = request; - - if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_ACK | OBD_FAIL_ONCE)) { - request->rq_req_md.options |= PTL_MD_ACK_DISABLE; - obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED; + CDEBUG (D_INFO, "conn=%p id %s\n", conn, libcfs_id2str(conn->c_peer)); + md.start = base; + md.length = len; + md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1; + md.options = PTLRPC_MD_OPTIONS; + md.user_ptr = cbid; + md.eq_handle = ptlrpc_eq_h; + + if (unlikely(ack == LNET_ACK_REQ && + OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){ + /* don't ask for the ack to simulate failing client */ + ack = LNET_NOACK_REQ; } - /* NB if the send fails, we back out of the send and return - * failure; it's down to the caller to handle missing callbacks */ - - rc = PtlMDBind(conn->c_peer.peer_ni->pni_ni_h, request->rq_req_md, - &md_h); - if (rc != PTL_OK) { - CERROR("PtlMDBind failed: %d\n", rc); - LASSERT (rc == PTL_NOSPACE); + rc = LNetMDBind (md, LNET_UNLINK, mdh); + if (unlikely(rc != 0)) { + CERROR ("LNetMDBind failed: %d\n", rc); + LASSERT (rc == -ENOMEM); RETURN (-ENOMEM); } - if (request->rq_type != PTL_RPC_MSG_REQUEST) - memcpy(&request->rq_reply_md_h, &md_h, sizeof(md_h)); - - remote_id.nid = conn->c_peer.peer_nid; - remote_id.pid = 0; - - CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64"\n", - request->rq_req_md.length, portal, request->rq_xid); - - rc = PtlPut(md_h, ack_req, remote_id, portal, 0, request->rq_xid, 0, 0); - if (rc != PTL_OK) { - CERROR("PtlPut("LPU64", %d, "LPD64") failed: %d\n", - remote_id.nid, portal, request->rq_xid, rc); - rc2 = PtlMDUnlink(md_h); - LASSERT (rc2 == PTL_OK); - RETURN ((rc == PTL_NOSPACE) ? -ENOMEM : -ECOMM); + + CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64", offset %u\n", + len, portal, xid, offset); + + rc = LNetPut (conn->c_self, *mdh, ack, + conn->c_peer, portal, xid, offset, 0); + if (unlikely(rc != 0)) { + int rc2; + /* We're going to get an UNLINK event when I unlink below, + * which will complete just like any other failed send, so + * I fall through and return success here! */ + CERROR("LNetPut(%s, %d, "LPD64") failed: %d\n", + libcfs_id2str(conn->c_peer), portal, xid, rc); + rc2 = LNetMDUnlink(*mdh); + LASSERTF(rc2 == 0, "rc2 = %d\n", rc2); } - return 0; + RETURN (0); } -static inline ptl_kiov_t * -ptlrpc_get_bulk_iov (struct ptlrpc_bulk_desc *desc) +static void mdunlink_iterate_helper(lnet_handle_md_t *bd_mds, int count) { - ptl_kiov_t *iov; - - if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (*iov)) - return (desc->bd_iov); + int i; - OBD_ALLOC (iov, desc->bd_page_count * sizeof (*iov)); - if (iov == NULL) - LBUG(); - - return (iov); + for (i = 0; i < count; i++) + LNetMDUnlink(bd_mds[i]); } -static inline void -ptlrpc_put_bulk_iov (struct ptlrpc_bulk_desc *desc, ptl_kiov_t *iov) +#ifdef HAVE_SERVER_SUPPORT +/** + * Prepare bulk descriptor for specified incoming request \a req that + * can fit \a npages * pages. \a type is bulk type. \a portal is where + * the bulk to be sent. Used on server-side after request was already + * received. + * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on + * error. + */ +struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req, + unsigned npages, unsigned max_brw, + unsigned type, unsigned portal) { - if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (*iov)) - return; + struct obd_export *exp = req->rq_export; + struct ptlrpc_bulk_desc *desc; - OBD_FREE (iov, desc->bd_page_count * sizeof (*iov)); -} + ENTRY; + LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK); -int ptlrpc_bulk_put(struct ptlrpc_bulk_desc *desc) -{ - int rc; - int rc2; - struct ptlrpc_peer *peer; - struct list_head *tmp, *next; - ptl_process_id_t remote_id; - ptl_kiov_t *iov; - __u64 xid; - ENTRY; + desc = ptlrpc_new_bulk(npages, max_brw, type, portal); + if (desc == NULL) + RETURN(NULL); - /* NB no locking required until desc is on the network */ - LASSERT (!desc->bd_network_rw); - LASSERT (desc->bd_type == BULK_PUT_SOURCE); - desc->bd_complete = 0; + desc->bd_export = class_export_get(exp); + desc->bd_req = req; - iov = ptlrpc_get_bulk_iov (desc); - if (iov == NULL) - RETURN (-ENOMEM); - - peer = &desc->bd_export->exp_connection->c_peer; - - desc->bd_md.start = iov; - desc->bd_md.niov = 0; - desc->bd_md.length = 0; - desc->bd_md.eventq = peer->peer_ni->pni_bulk_put_source_eq_h; - desc->bd_md.threshold = 2; /* SENT and ACK */ - desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_KIOV; - desc->bd_md.user_ptr = desc; - - desc->bd_callback_count = 2; - - list_for_each_safe(tmp, next, &desc->bd_page_list) { - struct ptlrpc_bulk_page *bulk; - bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link); - - LASSERT(desc->bd_md.niov < desc->bd_page_count); - - iov[desc->bd_md.niov].kiov_page = bulk->bp_page; - iov[desc->bd_md.niov].kiov_offset = bulk->bp_pageoffset; - iov[desc->bd_md.niov].kiov_len = bulk->bp_buflen; - - LASSERT (iov[desc->bd_md.niov].kiov_offset + - iov[desc->bd_md.niov].kiov_len <= PAGE_SIZE); - desc->bd_md.niov++; - desc->bd_md.length += bulk->bp_buflen; - } - - /* NB total length may be 0 for a read past EOF, so we send a 0 - * length bulk, since the client expects a bulk event. */ - LASSERT(desc->bd_md.niov == desc->bd_page_count); + desc->bd_cbid.cbid_fn = server_bulk_callback; + desc->bd_cbid.cbid_arg = desc; - rc = PtlMDBind(peer->peer_ni->pni_ni_h, desc->bd_md, - &desc->bd_md_h); + /* NB we don't assign rq_bulk here; server-side requests are + * re-used, and the handler frees the bulk desc explicitly. */ - ptlrpc_put_bulk_iov (desc, iov); /*move down to reduce latency to send*/ - - if (rc != PTL_OK) { - CERROR("PtlMDBind failed: %d\n", rc); - LASSERT (rc == PTL_NOSPACE); - RETURN(-ENOMEM); - } - - /* Client's bulk and reply matchbits are the same */ - xid = desc->bd_req->rq_xid; - remote_id.nid = peer->peer_nid; - remote_id.pid = 0; - - CDEBUG(D_NET, "Sending %u pages %u bytes to portal %d on %s " - "nid "LPX64" pid %d xid "LPX64"\n", - desc->bd_md.niov, desc->bd_md.length, - desc->bd_portal, peer->peer_ni->pni_name, - remote_id.nid, remote_id.pid, xid); - - desc->bd_network_rw = 1; - rc = PtlPut(desc->bd_md_h, PTL_ACK_REQ, remote_id, - desc->bd_portal, 0, xid, 0, 0); - if (rc != PTL_OK) { - desc->bd_network_rw = 0; - CERROR("PtlPut("LPU64", %d, "LPX64") failed: %d\n", - remote_id.nid, desc->bd_portal, xid, rc); - rc2 = PtlMDUnlink(desc->bd_md_h); - LASSERT (rc2 == PTL_OK); - RETURN((rc == PTL_NOSPACE) ? -ENOMEM : -ECOMM); - } - - RETURN(0); + return desc; } +EXPORT_SYMBOL(ptlrpc_prep_bulk_exp); -int ptlrpc_bulk_get(struct ptlrpc_bulk_desc *desc) +/** + * Starts bulk transfer for descriptor \a desc on the server. + * Returns 0 on success or error code. + */ +int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc) { - int rc; - int rc2; - struct ptlrpc_peer *peer; - struct list_head *tmp, *next; - ptl_process_id_t remote_id; - ptl_kiov_t *iov; - __u64 xid; - ENTRY; - - /* NB no locking required until desc is on the network */ - LASSERT (!desc->bd_network_rw); - LASSERT (desc->bd_type == BULK_GET_SINK); - desc->bd_complete = 0; - - iov = ptlrpc_get_bulk_iov (desc); - if (iov == NULL) - RETURN(-ENOMEM); - - peer = &desc->bd_export->exp_connection->c_peer; - - desc->bd_md.start = iov; - desc->bd_md.niov = 0; - desc->bd_md.length = 0; - desc->bd_md.eventq = peer->peer_ni->pni_bulk_get_sink_eq_h; - desc->bd_md.threshold = 2; /* SENT and REPLY */ - desc->bd_md.options = PTL_MD_OP_GET | PTL_MD_KIOV; - desc->bd_md.user_ptr = desc; - - desc->bd_callback_count = 2; - - list_for_each_safe(tmp, next, &desc->bd_page_list) { - struct ptlrpc_bulk_page *bulk; - bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link); - - LASSERT(desc->bd_md.niov < desc->bd_page_count); - - iov[desc->bd_md.niov].kiov_page = bulk->bp_page; - iov[desc->bd_md.niov].kiov_len = bulk->bp_buflen; - iov[desc->bd_md.niov].kiov_offset = bulk->bp_pageoffset; - - LASSERT (iov[desc->bd_md.niov].kiov_offset + - iov[desc->bd_md.niov].kiov_len <= PAGE_SIZE); - desc->bd_md.niov++; - desc->bd_md.length += bulk->bp_buflen; - } - - LASSERT(desc->bd_md.niov == desc->bd_page_count); - LASSERT(desc->bd_md.niov != 0); - - rc = PtlMDBind(peer->peer_ni->pni_ni_h, desc->bd_md, &desc->bd_md_h); - - ptlrpc_put_bulk_iov(desc, iov); /*move down to reduce latency to send*/ - - if (rc != PTL_OK) { - CERROR("PtlMDBind failed: %d\n", rc); - LASSERT (rc == PTL_NOSPACE); - RETURN(-ENOMEM); - } - - /* Client's bulk and reply matchbits are the same */ - xid = desc->bd_req->rq_xid; - remote_id.nid = desc->bd_export->exp_connection->c_peer.peer_nid; - remote_id.pid = 0; - - CDEBUG(D_NET, "Fetching %u pages %u bytes from portal %d on %s " - "nid "LPX64" pid %d xid "LPX64"\n", - desc->bd_md.niov, desc->bd_md.length, desc->bd_portal, - peer->peer_ni->pni_name, remote_id.nid, remote_id.pid, - xid); - - desc->bd_network_rw = 1; - rc = PtlGet(desc->bd_md_h, remote_id, desc->bd_portal, 0, - xid, 0); - if (rc != PTL_OK) { - desc->bd_network_rw = 0; - CERROR("PtlGet("LPU64", %d, "LPX64") failed: %d\n", - remote_id.nid, desc->bd_portal, xid, rc); - rc2 = PtlMDUnlink(desc->bd_md_h); - LASSERT (rc2 == PTL_OK); - RETURN((rc == PTL_NOSPACE) ? -ENOMEM : -ECOMM); - } - - RETURN(0); + struct obd_export *exp = desc->bd_export; + struct ptlrpc_connection *conn = exp->exp_connection; + int rc = 0; + __u64 xid; + int posted_md; + int total_md; + lnet_md_t md; + ENTRY; + + if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET)) + RETURN(0); + + /* NB no locking required until desc is on the network */ + LASSERT(desc->bd_md_count == 0); + LASSERT(desc->bd_type == BULK_PUT_SOURCE || + desc->bd_type == BULK_GET_SINK); + + LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback); + LASSERT(desc->bd_cbid.cbid_arg == desc); + + /* NB total length may be 0 for a read past EOF, so we send 0 + * length bulks, since the client expects bulk events. + * + * The client may not need all of the bulk XIDs for the RPC. The RPC + * used the XID of the highest bulk XID needed, and the server masks + * off high bits to get bulk count for this RPC. LU-1431 */ + xid = desc->bd_req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1); + total_md = desc->bd_req->rq_xid - xid + 1; + + desc->bd_md_count = total_md; + desc->bd_failure = 0; + + md.user_ptr = &desc->bd_cbid; + md.eq_handle = ptlrpc_eq_h; + md.threshold = 2; /* SENT and ACK/REPLY */ + + for (posted_md = 0; posted_md < total_md; xid++) { + md.options = PTLRPC_MD_OPTIONS; + + /* NB it's assumed that source and sink buffer frags are + * page-aligned. Otherwise we'd have to send client bulk + * sizes over and split server buffer accordingly */ + ptlrpc_fill_bulk_md(&md, desc, posted_md); + rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_mds[posted_md]); + if (rc != 0) { + CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n", + exp->exp_obd->obd_name, posted_md, rc); + LASSERT(rc == -ENOMEM); + if (posted_md == 0) { + desc->bd_md_count = 0; + RETURN(-ENOMEM); + } + break; + } + /* Network is about to get at the memory */ + if (desc->bd_type == BULK_PUT_SOURCE) + rc = LNetPut(conn->c_self, desc->bd_mds[posted_md], + LNET_ACK_REQ, conn->c_peer, + desc->bd_portal, xid, 0, 0); + else + rc = LNetGet(conn->c_self, desc->bd_mds[posted_md], + conn->c_peer, desc->bd_portal, xid, 0); + + posted_md++; + if (rc != 0) { + CERROR("%s: failed bulk transfer with %s:%u x"LPU64": " + "rc = %d\n", exp->exp_obd->obd_name, + libcfs_id2str(conn->c_peer), desc->bd_portal, + xid, rc); + break; + } + } + + if (rc != 0) { + /* Can't send, so we unlink the MD bound above. The UNLINK + * event this creates will signal completion with failure, + * so we return SUCCESS here! */ + spin_lock(&desc->bd_lock); + desc->bd_md_count -= total_md - posted_md; + spin_unlock(&desc->bd_lock); + LASSERT(desc->bd_md_count >= 0); + + mdunlink_iterate_helper(desc->bd_mds, posted_md); + RETURN(0); + } + + CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d " + "id %s xid "LPX64"-"LPX64"\n", desc->bd_iov_count, + desc->bd_nob, desc->bd_portal, libcfs_id2str(conn->c_peer), + xid - posted_md, xid - 1); + + RETURN(0); } -void ptlrpc_abort_bulk (struct ptlrpc_bulk_desc *desc) +/** + * Server side bulk abort. Idempotent. Not thread-safe (i.e. only + * serialises with completion callback) + */ +void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc) { - /* Server side bulk abort. Idempotent. Not thread-safe (i.e. only - * serialises with completion callback) */ - unsigned long flags; - struct l_wait_info lwi; - int callback_count; - int rc; - - LASSERT (!in_interrupt ()); /* might sleep */ - - /* NB. server-side bulk gets 2 events, so we have to keep trying to - * unlink the MD until all callbacks have happened, or - * PtlMDUnlink() returns OK or INVALID */ - again: - spin_lock_irqsave (&desc->bd_lock, flags); - if (!desc->bd_network_rw) { - /* completed or never even registered. NB holding bd_lock - * guarantees callback has completed if it ran. */ - spin_unlock_irqrestore (&desc->bd_lock, flags); - return; - } - - /* sample callback count while we have the lock */ - callback_count = desc->bd_callback_count; - spin_unlock_irqrestore (&desc->bd_lock, flags); - - rc = PtlMDUnlink (desc->bd_md_h); - switch (rc) { - default: - CERROR("PtlMDUnlink returned %d\n", rc); - LBUG (); - case PTL_OK: /* Won the race with the network */ - LASSERT (!desc->bd_complete); /* Not all callbacks ran */ - desc->bd_network_rw = 0; - return; - - case PTL_MD_INUSE: /* MD is being accessed right now */ - for (;;) { - /* Network access will complete in finite time but the - * timeout lets us CERROR for visibility */ - lwi = LWI_TIMEOUT (10 * HZ, NULL, NULL); - rc = l_wait_event(desc->bd_waitq, - desc->bd_callback_count != - callback_count, &lwi); - if (rc == -ETIMEDOUT) { - CERROR("Unexpectedly long timeout: desc %p\n", - desc); - continue; - } - LASSERT (rc == 0); - break; - } - /* go back and try again... */ - goto again; - - case PTL_INV_MD: /* Lost the race with completion */ - LASSERT (desc->bd_complete); /* Callbacks all ran */ - LASSERT (!desc->bd_network_rw); - return; - } + struct l_wait_info lwi; + int rc; + + LASSERT(!in_interrupt()); /* might sleep */ + + if (!ptlrpc_server_bulk_active(desc)) /* completed or */ + return; /* never started */ + + /* We used to poison the pages with 0xab here because we did not want to + * send any meaningful data over the wire for evicted clients (bug 9297) + * However, this is no longer safe now that we use the page cache on the + * OSS (bug 20560) */ + + /* The unlink ensures the callback happens ASAP and is the last + * one. If it fails, it must be because completion just happened, + * but we must still l_wait_event() in this case, to give liblustre + * a chance to run server_bulk_callback()*/ + mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_count); + + for (;;) { + /* Network access will complete in finite time but the HUGE + * timeout lets us CWARN for visibility of sluggish NALs */ + lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK), + cfs_time_seconds(1), NULL, NULL); + rc = l_wait_event(desc->bd_waitq, + !ptlrpc_server_bulk_active(desc), &lwi); + if (rc == 0) + return; + + LASSERT(rc == -ETIMEDOUT); + CWARN("Unexpectedly long timeout: desc %p\n", desc); + } } +#endif /* HAVE_SERVER_SUPPORT */ -int ptlrpc_register_bulk (struct ptlrpc_request *req) +/** + * Register bulk at the sender for later transfer. + * Returns 0 on success or error code. + */ +int ptlrpc_register_bulk(struct ptlrpc_request *req) { - struct ptlrpc_bulk_desc *desc = req->rq_bulk; - struct ptlrpc_peer *peer; - struct list_head *tmp, *next; - int rc; - int rc2; - ptl_kiov_t *iov; - ptl_process_id_t source_id; - ENTRY; - - /* NB no locking required until desc is on the network */ - LASSERT (!desc->bd_network_rw); - LASSERT (desc->bd_page_count <= PTL_MD_MAX_PAGES); - LASSERT (desc->bd_req != NULL); - LASSERT (desc->bd_type == BULK_PUT_SINK || - desc->bd_type == BULK_GET_SOURCE); - - desc->bd_complete = 0; - - iov = ptlrpc_get_bulk_iov (desc); - if (iov == NULL) - return (-ENOMEM); - - peer = &desc->bd_import->imp_connection->c_peer; - - desc->bd_md.start = iov; - desc->bd_md.niov = 0; - desc->bd_md.length = 0; - desc->bd_md.threshold = 1; - desc->bd_md.user_ptr = desc; - - if (desc->bd_type == BULK_GET_SOURCE) { - desc->bd_md.options = PTL_MD_OP_GET | PTL_MD_KIOV; - desc->bd_md.eventq = peer->peer_ni->pni_bulk_get_source_eq_h; - } else { - desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_KIOV; - desc->bd_md.eventq = peer->peer_ni->pni_bulk_put_sink_eq_h; - } - - list_for_each_safe(tmp, next, &desc->bd_page_list) { - struct ptlrpc_bulk_page *bulk; - bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link); - - LASSERT(desc->bd_md.niov < desc->bd_page_count); - - iov[desc->bd_md.niov].kiov_page = bulk->bp_page; - iov[desc->bd_md.niov].kiov_len = bulk->bp_buflen; - iov[desc->bd_md.niov].kiov_offset = bulk->bp_pageoffset; - - LASSERT (bulk->bp_pageoffset + bulk->bp_buflen <= PAGE_SIZE); - desc->bd_md.niov++; - desc->bd_md.length += bulk->bp_buflen; - } - - LASSERT(desc->bd_md.niov == desc->bd_page_count); - LASSERT(desc->bd_md.niov != 0); - - /* XXX Registering the same xid on retried bulk makes my head - * explode trying to understand how the original request's bulk - * might interfere with the retried request -eeb */ - LASSERT (!desc->bd_registered || req->rq_xid != desc->bd_last_xid); - desc->bd_registered = 1; - desc->bd_last_xid = desc->bd_last_xid; - - source_id.nid = desc->bd_import->imp_connection->c_peer.peer_nid; - source_id.pid = PTL_PID_ANY; - - rc = PtlMEAttach(peer->peer_ni->pni_ni_h, - desc->bd_portal, source_id, req->rq_xid, 0, - PTL_UNLINK, PTL_INS_AFTER, &desc->bd_me_h); + struct ptlrpc_bulk_desc *desc = req->rq_bulk; + lnet_process_id_t peer; + int rc = 0; + int rc2; + int posted_md; + int total_md; + __u64 xid; + lnet_handle_me_t me_h; + lnet_md_t md; + ENTRY; + + if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET)) + RETURN(0); + + /* NB no locking required until desc is on the network */ + LASSERT(desc->bd_nob > 0); + LASSERT(desc->bd_md_count == 0); + LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT); + LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES); + LASSERT(desc->bd_req != NULL); + LASSERT(desc->bd_type == BULK_PUT_SINK || + desc->bd_type == BULK_GET_SOURCE); + + /* cleanup the state of the bulk for it will be reused */ + if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY) + desc->bd_nob_transferred = 0; + else + LASSERT(desc->bd_nob_transferred == 0); + + desc->bd_failure = 0; + + peer = desc->bd_import->imp_connection->c_peer; + + LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback); + LASSERT(desc->bd_cbid.cbid_arg == desc); + + /* An XID is only used for a single request from the client. + * For retried bulk transfers, a new XID will be allocated in + * in ptlrpc_check_set() if it needs to be resent, so it is not + * using the same RDMA match bits after an error. + * + * For multi-bulk RPCs, rq_xid is the last XID needed for bulks. The + * first bulk XID is power-of-two aligned before rq_xid. LU-1431 */ + xid = req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1); + LASSERTF(!(desc->bd_registered && + req->rq_send_state != LUSTRE_IMP_REPLAY) || + xid != desc->bd_last_xid, + "registered: %d rq_xid: "LPU64" bd_last_xid: "LPU64"\n", + desc->bd_registered, xid, desc->bd_last_xid); + + total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV; + desc->bd_registered = 1; + desc->bd_last_xid = xid; + desc->bd_md_count = total_md; + md.user_ptr = &desc->bd_cbid; + md.eq_handle = ptlrpc_eq_h; + md.threshold = 1; /* PUT or GET */ + + for (posted_md = 0; posted_md < total_md; posted_md++, xid++) { + md.options = PTLRPC_MD_OPTIONS | + ((desc->bd_type == BULK_GET_SOURCE) ? + LNET_MD_OP_GET : LNET_MD_OP_PUT); + ptlrpc_fill_bulk_md(&md, desc, posted_md); + + rc = LNetMEAttach(desc->bd_portal, peer, xid, 0, + LNET_UNLINK, LNET_INS_AFTER, &me_h); + if (rc != 0) { + CERROR("%s: LNetMEAttach failed x"LPU64"/%d: rc = %d\n", + desc->bd_import->imp_obd->obd_name, xid, + posted_md, rc); + break; + } + + /* About to let the network at it... */ + rc = LNetMDAttach(me_h, md, LNET_UNLINK, + &desc->bd_mds[posted_md]); + if (rc != 0) { + CERROR("%s: LNetMDAttach failed x"LPU64"/%d: rc = %d\n", + desc->bd_import->imp_obd->obd_name, xid, + posted_md, rc); + rc2 = LNetMEUnlink(me_h); + LASSERT(rc2 == 0); + break; + } + } + + if (rc != 0) { + LASSERT(rc == -ENOMEM); + spin_lock(&desc->bd_lock); + desc->bd_md_count -= total_md - posted_md; + spin_unlock(&desc->bd_lock); + LASSERT(desc->bd_md_count >= 0); + mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw); + req->rq_status = -ENOMEM; + RETURN(-ENOMEM); + } + + /* Set rq_xid to matchbits of the final bulk so that server can + * infer the number of bulks that were prepared */ + req->rq_xid = --xid; + LASSERTF(desc->bd_last_xid == (req->rq_xid & PTLRPC_BULK_OPS_MASK), + "bd_last_xid = x"LPU64", rq_xid = x"LPU64"\n", + desc->bd_last_xid, req->rq_xid); + + spin_lock(&desc->bd_lock); + /* Holler if peer manages to touch buffers before he knows the xid */ + if (desc->bd_md_count != total_md) + CWARN("%s: Peer %s touched %d buffers while I registered\n", + desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer), + total_md - desc->bd_md_count); + spin_unlock(&desc->bd_lock); + + CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, " + "xid x"LPX64"-"LPX64", portal %u\n", desc->bd_md_count, + desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink", + desc->bd_iov_count, desc->bd_nob, + desc->bd_last_xid, req->rq_xid, desc->bd_portal); + + RETURN(0); +} - if (rc != PTL_OK) { - CERROR("PtlMEAttach failed: %d\n", rc); - LASSERT (rc == PTL_NOSPACE); - GOTO(out, rc = -ENOMEM); - } +/** + * Disconnect a bulk desc from the network. Idempotent. Not + * thread-safe (i.e. only interlocks with completion callback). + * Returns 1 on success or 0 if network unregistration failed for whatever + * reason. + */ +int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async) +{ + struct ptlrpc_bulk_desc *desc = req->rq_bulk; + struct l_wait_info lwi; + int rc; + ENTRY; + + LASSERT(!in_interrupt()); /* might sleep */ + + /* Let's setup deadline for reply unlink. */ + if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) && + async && req->rq_bulk_deadline == 0) + req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK; + + if (ptlrpc_client_bulk_active(req) == 0) /* completed or */ + RETURN(1); /* never registered */ + + LASSERT(desc->bd_req == req); /* bd_req NULL until registered */ + + /* the unlink ensures the callback happens ASAP and is the last + * one. If it fails, it must be because completion just happened, + * but we must still l_wait_event() in this case to give liblustre + * a chance to run client_bulk_callback() */ + mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw); + + if (ptlrpc_client_bulk_active(req) == 0) /* completed or */ + RETURN(1); /* never registered */ + + /* Move to "Unregistering" phase as bulk was not unlinked yet. */ + ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING); + + /* Do not wait for unlink to finish. */ + if (async) + RETURN(0); + + for (;;) { + /* The wq argument is ignored by user-space wait_event macros */ + wait_queue_head_t *wq = (req->rq_set != NULL) ? + &req->rq_set->set_waitq : + &req->rq_reply_waitq; + /* Network access will complete in finite time but the HUGE + * timeout lets us CWARN for visibility of sluggish NALs */ + lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK), + cfs_time_seconds(1), NULL, NULL); + rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi); + if (rc == 0) { + ptlrpc_rqphase_move(req, req->rq_next_phase); + RETURN(1); + } - /* About to let the network at it... */ - desc->bd_network_rw = 1; - rc = PtlMDAttach(desc->bd_me_h, desc->bd_md, PTL_UNLINK, - &desc->bd_md_h); - if (rc != PTL_OK) { - CERROR("PtlMDAttach failed: %d\n", rc); - LASSERT (rc == PTL_NOSPACE); - desc->bd_network_rw = 0; - rc2 = PtlMEUnlink (desc->bd_me_h); - LASSERT (rc2 == PTL_OK); - GOTO(out, rc = -ENOMEM); + LASSERT(rc == -ETIMEDOUT); + DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p", + desc); } - rc = 0; - - CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPX64", " - "portal %u on %s\n", - desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink", - desc->bd_md.niov, desc->bd_md.length, - req->rq_xid, desc->bd_portal, peer->peer_ni->pni_name); - - out: - ptlrpc_put_bulk_iov (desc, iov); - RETURN(rc); + RETURN(0); } -void ptlrpc_unregister_bulk (struct ptlrpc_request *req) +static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags) { - /* Disconnect a bulk desc from the network. Idempotent. Not - * thread-safe (i.e. only interlocks with completion callback). */ - struct ptlrpc_bulk_desc *desc = req->rq_bulk; - wait_queue_head_t *wq; - unsigned long flags; - struct l_wait_info lwi; - int rc; - - LASSERT (!in_interrupt ()); /* might sleep */ - - spin_lock_irqsave (&desc->bd_lock, flags); - if (!desc->bd_network_rw) { /* completed or never even registered */ - spin_unlock_irqrestore (&desc->bd_lock, flags); - return; - } - spin_unlock_irqrestore (&desc->bd_lock, flags); - - LASSERT (desc->bd_req == req); /* NB bd_req NULL until registered */ - - /* NB... - * 1. If the MD unlink is successful, the ME gets unlinked too. - * 2. Since client-side bulk only gets a single event and a - * .. threshold of 1. If the MD was inuse at the first link - * .. attempt, the callback is due any minute, and the MD/ME will - * .. unlink themselves. - */ - rc = PtlMDUnlink (desc->bd_md_h); - switch (rc) { - default: - CERROR("PtlMDUnlink returned %d\n", rc); - LBUG (); - case PTL_OK: /* Won the race with completion */ - LASSERT (!desc->bd_complete); /* Callback hasn't happened */ - desc->bd_network_rw = 0; - return; - case PTL_MD_INUSE: /* MD is being accessed right now */ - for (;;) { - /* Network access will complete in finite time but the - * timeout lets us CERROR for visibility */ - if (desc->bd_req->rq_set != NULL) - wq = &req->rq_set->set_waitq; - else - wq = &req->rq_reply_waitq; - lwi = LWI_TIMEOUT (10 * HZ, NULL, NULL); - rc = l_wait_event(*wq, ptlrpc_bulk_complete(desc), &lwi); - LASSERT (rc == 0 || rc == -ETIMEDOUT); - if (rc == 0) - break; - CERROR ("Unexpectedly long timeout: desc %p\n", desc); - LBUG(); - } - /* Fall through */ - case PTL_INV_MD: /* Lost the race with completion */ - LASSERT (desc->bd_complete);/* Callback has run to completion */ - LASSERT (!desc->bd_network_rw); - return; + struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt; + struct ptlrpc_service *svc = svcpt->scp_service; + int service_time = max_t(int, cfs_time_current_sec() - + req->rq_arrival_time.tv_sec, 1); + + if (!(flags & PTLRPC_REPLY_EARLY) && + (req->rq_type != PTL_RPC_MSG_ERR) && + (req->rq_reqmsg != NULL) && + !(lustre_msg_get_flags(req->rq_reqmsg) & + (MSG_RESENT | MSG_REPLAY | + MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) { + /* early replies, errors and recovery requests don't count + * toward our service time estimate */ + int oldse = at_measured(&svcpt->scp_at_estimate, service_time); + + if (oldse != 0) { + DEBUG_REQ(D_ADAPTTO, req, + "svc %s changed estimate from %d to %d", + svc->srv_name, oldse, + at_get(&svcpt->scp_at_estimate)); + } } + /* Report actual service time for client latency calc */ + lustre_msg_set_service_time(req->rq_repmsg, service_time); + /* Report service time estimate for future client reqs, but report 0 + * (to be ignored by client) if it's an error reply during recovery. + * (bz15815) */ + if (req->rq_type == PTL_RPC_MSG_ERR && + (req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering)) + lustre_msg_set_timeout(req->rq_repmsg, 0); + else + lustre_msg_set_timeout(req->rq_repmsg, + at_get(&svcpt->scp_at_estimate)); + + if (req->rq_reqmsg && + !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) { + CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x " + "req_flags=%#x magic=%x/%x len=%d\n", + flags, lustre_msg_get_flags(req->rq_reqmsg), + lustre_msg_get_magic(req->rq_reqmsg), + lustre_msg_get_magic(req->rq_repmsg), req->rq_replen); + } } -int ptlrpc_reply(struct ptlrpc_request *req) +/** + * Send request reply from request \a req reply buffer. + * \a flags defines reply types + * Returns 0 on success or error code + */ +int ptlrpc_send_reply(struct ptlrpc_request *req, int flags) { - struct ptlrpc_connection *conn; - unsigned long flags; - int rc; + struct ptlrpc_reply_state *rs = req->rq_reply_state; + struct ptlrpc_connection *conn; + int rc; /* We must already have a reply buffer (only ptlrpc_error() may be - * called without one). We must also have a request buffer which - * is either the actual (swabbed) incoming request, or a saved copy - * if this is a req saved in target_queue_final_reply(). */ + * called without one). The reply generated by sptlrpc layer (e.g. + * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must + * have a request buffer which is either the actual (swabbed) incoming + * request, or a saved copy if this is a req saved in + * target_queue_final_reply(). + */ + LASSERT (req->rq_no_reply == 0); + LASSERT (req->rq_reqbuf != NULL); + LASSERT (rs != NULL); + LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult); LASSERT (req->rq_repmsg != NULL); - LASSERT (req->rq_reqmsg != NULL); + LASSERT (req->rq_repmsg == rs->rs_msg); + LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback); + LASSERT (rs->rs_cb_id.cbid_arg == rs); + + /* There may be no rq_export during failover */ + + if (unlikely(req->rq_export && req->rq_export->exp_obd && + req->rq_export->exp_obd->obd_fail)) { + /* Failed obd's only send ENODEV */ + req->rq_type = PTL_RPC_MSG_ERR; + req->rq_status = -ENODEV; + CDEBUG(D_HA, "sending ENODEV from failed obd %d\n", + req->rq_export->exp_obd->obd_minor); + } + + /* In order to keep interoprability with the client (< 2.3) which + * doesn't have pb_jobid in ptlrpc_body, We have to shrink the + * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the + * reply buffer on client will be overflow. + * + * XXX Remove this whenver we drop the interoprability with such client. + */ + req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0, + sizeof(struct ptlrpc_body_v2), 1); - /* FIXME: we need to increment the count of handled events */ if (req->rq_type != PTL_RPC_MSG_ERR) req->rq_type = PTL_RPC_MSG_REPLY; - req->rq_repmsg->status = req->rq_status; - req->rq_repmsg->opc = req->rq_reqmsg->opc; + lustre_msg_set_type(req->rq_repmsg, req->rq_type); + lustre_msg_set_status(req->rq_repmsg, + ptlrpc_status_hton(req->rq_status)); + lustre_msg_set_opc(req->rq_repmsg, + req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0); - if (req->rq_export == NULL) - conn = ptlrpc_get_connection(&req->rq_peer, NULL); + target_pack_pool_reply(req); + + ptlrpc_at_set_reply(req, flags); + + if (req->rq_export == NULL || req->rq_export->exp_connection == NULL) + conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL); else conn = ptlrpc_connection_addref(req->rq_export->exp_connection); - init_waitqueue_head(&req->rq_reply_waitq); - rc = ptl_send_buf(req, conn, - req->rq_svc->srv_rep_portal); - if (rc != 0) { - /* Do what the callback handler would have done */ - OBD_FREE (req->rq_repmsg, req->rq_replen); - - spin_lock_irqsave (&req->rq_lock, flags); - req->rq_want_ack = 0; - spin_unlock_irqrestore (&req->rq_lock, flags); + if (unlikely(conn == NULL)) { + CERROR("not replying on NULL connection\n"); /* bug 9635 */ + return -ENOTCONN; } - ptlrpc_put_connection(conn); + ptlrpc_rs_addref(rs); /* +1 ref for the network */ + + rc = sptlrpc_svc_wrap_reply(req); + if (unlikely(rc)) + goto out; + + req->rq_sent = cfs_time_current_sec(); + + rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len, + (rs->rs_difficult && !rs->rs_no_ack) ? + LNET_ACK_REQ : LNET_NOACK_REQ, + &rs->rs_cb_id, conn, + ptlrpc_req2svc(req)->srv_rep_portal, + req->rq_xid, req->rq_reply_off); +out: + if (unlikely(rc != 0)) + ptlrpc_req_drop_rs(req); + ptlrpc_connection_put(conn); return rc; } -int ptlrpc_error(struct ptlrpc_request *req) +int ptlrpc_reply (struct ptlrpc_request *req) +{ + if (req->rq_no_reply) + return 0; + else + return (ptlrpc_send_reply(req, 0)); +} + +/** + * For request \a req send an error reply back. Create empty + * reply buffers if necessary. + */ +int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult) { int rc; ENTRY; + if (req->rq_no_reply) + RETURN(0); + if (!req->rq_repmsg) { - rc = lustre_pack_reply(req, 0, NULL, NULL); + rc = lustre_pack_reply(req, 1, NULL, NULL); if (rc) RETURN(rc); } + if (req->rq_status != -ENOSPC && req->rq_status != -EACCES && + req->rq_status != -EPERM && req->rq_status != -ENOENT && + req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT) + req->rq_type = PTL_RPC_MSG_ERR; - req->rq_type = PTL_RPC_MSG_ERR; - - rc = ptlrpc_reply(req); + rc = ptlrpc_send_reply(req, may_be_difficult); RETURN(rc); } -int ptl_send_rpc(struct ptlrpc_request *request) +int ptlrpc_error(struct ptlrpc_request *req) +{ + return ptlrpc_send_error(req, 0); +} + +/** + * Send request \a request. + * if \a noreply is set, don't expect any reply back and don't set up + * reply buffers. + * Returns 0 on success or error code. + */ +int ptl_send_rpc(struct ptlrpc_request *request, int noreply) { int rc; int rc2; + int mpflag = 0; struct ptlrpc_connection *connection; - unsigned long flags; - ptl_process_id_t source_id; - ptl_handle_me_t reply_me_h; + lnet_handle_me_t reply_me_h; + lnet_md_t reply_md; + struct obd_import *imp = request->rq_import; + struct obd_device *obd = imp->imp_obd; ENTRY; - LASSERT (request->rq_type == PTL_RPC_MSG_REQUEST); + if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC)) + RETURN(0); + + LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST); + LASSERT(request->rq_wait_ctx == 0); /* If this is a re-transmit, we're required to have disengaged * cleanly from the previous attempt */ - LASSERT (!request->rq_receiving_reply); - - connection = request->rq_import->imp_connection; + LASSERT(!request->rq_receiving_reply); + LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) && + (imp->imp_state == LUSTRE_IMP_FULL))); + + if (unlikely(obd != NULL && obd->obd_fail)) { + CDEBUG(D_HA, "muting rpc for failed imp obd %s\n", + obd->obd_name); + /* this prevents us from waiting in ptlrpc_queue_wait */ + spin_lock(&request->rq_lock); + request->rq_err = 1; + spin_unlock(&request->rq_lock); + request->rq_status = -ENODEV; + RETURN(-ENODEV); + } + connection = imp->imp_connection; + + lustre_msg_set_handle(request->rq_reqmsg, + &imp->imp_remote_handle); + lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST); + lustre_msg_set_conn_cnt(request->rq_reqmsg, + imp->imp_conn_cnt); + lustre_msghdr_set_flags(request->rq_reqmsg, + imp->imp_msghdr_flags); + + /** For enabled AT all request should have AT_SUPPORT in the + * FULL import state when OBD_CONNECT_AT is set */ + LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL || + (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) || + !(imp->imp_connect_data.ocd_connect_flags & + OBD_CONNECT_AT)); + + if (request->rq_resend) { + lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT); + if (request->rq_resend_cb != NULL) + request->rq_resend_cb(request, &request->rq_async_args); + } + if (request->rq_memalloc) + mpflag = cfs_memory_pressure_get_and_set(); + + rc = sptlrpc_cli_wrap_request(request); + if (rc) + GOTO(out, rc); + + /* bulk register should be done after wrap_request() */ if (request->rq_bulk != NULL) { rc = ptlrpc_register_bulk (request); if (rc != 0) - RETURN(rc); + GOTO(out, rc); } - request->rq_reqmsg->handle = request->rq_import->imp_remote_handle; - request->rq_reqmsg->conn_cnt = request->rq_import->imp_conn_cnt; - - source_id.nid = connection->c_peer.peer_nid; - source_id.pid = PTL_PID_ANY; - - LASSERT (request->rq_replen != 0); - if (request->rq_repmsg == NULL) - OBD_ALLOC(request->rq_repmsg, request->rq_replen); - if (request->rq_repmsg == NULL) { - LBUG(); - RETURN(-ENOMEM); - } - - rc = PtlMEAttach(connection->c_peer.peer_ni->pni_ni_h, - request->rq_reply_portal, /* XXX FIXME bug 249 */ - source_id, request->rq_xid, 0, PTL_UNLINK, - PTL_INS_AFTER, &reply_me_h); - if (rc != PTL_OK) { - CERROR("PtlMEAttach failed: %d\n", rc); - LASSERT (rc == PTL_NOSPACE); - LBUG(); - GOTO(cleanup, rc = -ENOMEM); - } + if (!noreply) { + LASSERT (request->rq_replen != 0); + if (request->rq_repbuf == NULL) { + LASSERT(request->rq_repdata == NULL); + LASSERT(request->rq_repmsg == NULL); + rc = sptlrpc_cli_alloc_repbuf(request, + request->rq_replen); + if (rc) { + /* this prevents us from looping in + * ptlrpc_queue_wait */ + spin_lock(&request->rq_lock); + request->rq_err = 1; + spin_unlock(&request->rq_lock); + request->rq_status = rc; + GOTO(cleanup_bulk, rc); + } + } else { + request->rq_repdata = NULL; + request->rq_repmsg = NULL; + } - request->rq_reply_md.start = request->rq_repmsg; - request->rq_reply_md.length = request->rq_replen; - request->rq_reply_md.threshold = 1; - request->rq_reply_md.options = PTL_MD_OP_PUT; - request->rq_reply_md.user_ptr = request; - request->rq_reply_md.eventq = - connection->c_peer.peer_ni->pni_reply_in_eq_h; - - rc = PtlMDAttach(reply_me_h, request->rq_reply_md, - PTL_UNLINK, &request->rq_reply_md_h); - if (rc != PTL_OK) { - CERROR("PtlMDAttach failed: %d\n", rc); - LASSERT (rc == PTL_NOSPACE); - LBUG(); - GOTO(cleanup2, rc -ENOMEM); + rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/ + connection->c_peer, request->rq_xid, 0, + LNET_UNLINK, LNET_INS_AFTER, &reply_me_h); + if (rc != 0) { + CERROR("LNetMEAttach failed: %d\n", rc); + LASSERT (rc == -ENOMEM); + GOTO(cleanup_bulk, rc = -ENOMEM); + } } - CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64 - ", portal %u on %s\n", - request->rq_replen, request->rq_xid, - request->rq_reply_portal, - connection->c_peer.peer_ni->pni_name); - - ptlrpc_request_addref(request); /* 1 ref for the SENT callback */ - - spin_lock_irqsave (&request->rq_lock, flags); - request->rq_receiving_reply = 1; - /* Clear any flags that may be present from previous sends. */ + spin_lock(&request->rq_lock); + /* We are responsible for unlinking the reply buffer */ + request->rq_reply_unlinked = noreply; + request->rq_receiving_reply = !noreply; + /* Clear any flags that may be present from previous sends. */ + request->rq_req_unlinked = 0; request->rq_replied = 0; request->rq_err = 0; request->rq_timedout = 0; + request->rq_net_err = 0; request->rq_resend = 0; request->rq_restart = 0; - spin_unlock_irqrestore (&request->rq_lock, flags); - - request->rq_sent = LTIME_S(CURRENT_TIME); - ptlrpc_pinger_sending_on_import(request->rq_import); - rc = ptl_send_buf(request, connection, request->rq_request_portal); - if (rc == 0) { - ptlrpc_lprocfs_rpc_sent(request); - RETURN(rc); + request->rq_reply_truncated = 0; + spin_unlock(&request->rq_lock); + + if (!noreply) { + reply_md.start = request->rq_repbuf; + reply_md.length = request->rq_repbuf_len; + /* Allow multiple early replies */ + reply_md.threshold = LNET_MD_THRESH_INF; + /* Manage remote for early replies */ + reply_md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | + LNET_MD_MANAGE_REMOTE | + LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */; + reply_md.user_ptr = &request->rq_reply_cbid; + reply_md.eq_handle = ptlrpc_eq_h; + + /* We must see the unlink callback to set rq_reply_unlinked, + * so we can't auto-unlink */ + rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN, + &request->rq_reply_md_h); + if (rc != 0) { + CERROR("LNetMDAttach failed: %d\n", rc); + LASSERT (rc == -ENOMEM); + spin_lock(&request->rq_lock); + /* ...but the MD attach didn't succeed... */ + request->rq_receiving_reply = 0; + spin_unlock(&request->rq_lock); + GOTO(cleanup_me, rc = -ENOMEM); + } + + CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64 + ", portal %u\n", + request->rq_repbuf_len, request->rq_xid, + request->rq_reply_portal); } - spin_lock_irqsave (&request->rq_lock, flags); - request->rq_receiving_reply = 0; - spin_unlock_irqrestore (&request->rq_lock, flags); - ptlrpc_req_finished (request); /* drop callback ref */ - cleanup2: + /* add references on request for request_out_callback */ + ptlrpc_request_addref(request); + if (obd != NULL && obd->obd_svc_stats != NULL) + lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR, + atomic_read(&imp->imp_inflight)); + + OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5); + + do_gettimeofday(&request->rq_sent_tv); + request->rq_sent = cfs_time_current_sec(); + /* We give the server rq_timeout secs to process the req, and + add the network latency for our local timeout. */ + request->rq_deadline = request->rq_sent + request->rq_timeout + + ptlrpc_at_get_net_latency(request); + + ptlrpc_pinger_sending_on_import(imp); + + DEBUG_REQ(D_INFO, request, "send flg=%x", + lustre_msg_get_flags(request->rq_reqmsg)); + rc = ptl_send_buf(&request->rq_req_md_h, + request->rq_reqbuf, request->rq_reqdata_len, + LNET_NOACK_REQ, &request->rq_req_cbid, + connection, + request->rq_request_portal, + request->rq_xid, 0); + if (likely(rc == 0)) + GOTO(out, rc); + + request->rq_req_unlinked = 1; + ptlrpc_req_finished(request); + if (noreply) + GOTO(out, rc); + + cleanup_me: /* MEUnlink is safe; the PUT didn't even get off the ground, and * nobody apart from the PUT's target has the right nid+XID to * access the reply buffer. */ - rc2 = PtlMEUnlink(reply_me_h); - LASSERT (rc2 == PTL_OK); - cleanup: - OBD_FREE(request->rq_repmsg, request->rq_replen); - request->rq_repmsg = NULL; + rc2 = LNetMEUnlink(reply_me_h); + LASSERT (rc2 == 0); + /* UNLINKED callback called synchronously */ + LASSERT(!request->rq_receiving_reply); + + cleanup_bulk: + /* We do sync unlink here as there was no real transfer here so + * the chance to have long unlink to sluggish net is smaller here. */ + ptlrpc_unregister_bulk(request, 0); + out: + if (request->rq_memalloc) + cfs_memory_pressure_restore(mpflag); return rc; } +EXPORT_SYMBOL(ptl_send_rpc); -void ptlrpc_link_svc_me(struct ptlrpc_request_buffer_desc *rqbd) +/** + * Register request buffer descriptor for request receiving. + */ +int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd) { - struct ptlrpc_srv_ni *srv_ni = rqbd->rqbd_srv_ni; - struct ptlrpc_service *service = srv_ni->sni_service; - static ptl_process_id_t match_id = {PTL_NID_ANY, PTL_PID_ANY}; - int rc; - ptl_md_t dummy; - ptl_handle_md_t md_h; - - LASSERT(atomic_read(&rqbd->rqbd_refcount) == 0); - - CDEBUG(D_NET, "PtlMEAttach: portal %d on %s h %lx."LPX64"\n", - service->srv_req_portal, srv_ni->sni_ni->pni_name, - srv_ni->sni_ni->pni_ni_h.nal_idx, - srv_ni->sni_ni->pni_ni_h.cookie); - - /* Attach the leading ME on which we build the ring */ - rc = PtlMEAttach(srv_ni->sni_ni->pni_ni_h, service->srv_req_portal, - match_id, 0, ~0, - PTL_UNLINK, PTL_INS_AFTER, &rqbd->rqbd_me_h); - if (rc != PTL_OK) { - CERROR("PtlMEAttach failed: %d\n", rc); - /* BUG 1191 */ - LBUG(); - } + struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service; + static lnet_process_id_t match_id = {LNET_NID_ANY, LNET_PID_ANY}; + int rc; + lnet_md_t md; + lnet_handle_me_t me_h; + + CDEBUG(D_NET, "LNetMEAttach: portal %d\n", + service->srv_req_portal); - dummy.start = rqbd->rqbd_buffer; - dummy.length = service->srv_buf_size; - dummy.max_size = service->srv_max_req_size; - dummy.threshold = PTL_MD_THRESH_INF; - dummy.options = PTL_MD_OP_PUT | PTL_MD_MAX_SIZE | PTL_MD_AUTO_UNLINK; - dummy.user_ptr = rqbd; - dummy.eventq = srv_ni->sni_eq_h; - - atomic_inc(&srv_ni->sni_nrqbds_receiving); - atomic_set(&rqbd->rqbd_refcount, 1); /* 1 ref for portals */ - - rc = PtlMDAttach(rqbd->rqbd_me_h, dummy, PTL_UNLINK, &md_h); - if (rc != PTL_OK) { - CERROR("PtlMDAttach failed: %d\n", rc); - LASSERT (rc == PTL_NOSPACE); - LBUG(); - /* BUG 1191 */ - PtlMEUnlink (rqbd->rqbd_me_h); - atomic_set(&rqbd->rqbd_refcount, 0); - atomic_dec(&srv_ni->sni_nrqbds_receiving); + if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD)) + return (-ENOMEM); + + /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL, + * which means buffer can only be attached on local CPT, and LND + * threads can find it by grabbing a local lock */ + rc = LNetMEAttach(service->srv_req_portal, + match_id, 0, ~0, LNET_UNLINK, + rqbd->rqbd_svcpt->scp_cpt >= 0 ? + LNET_INS_LOCAL : LNET_INS_AFTER, &me_h); + if (rc != 0) { + CERROR("LNetMEAttach failed: %d\n", rc); + return (-ENOMEM); } + + LASSERT(rqbd->rqbd_refcount == 0); + rqbd->rqbd_refcount = 1; + + md.start = rqbd->rqbd_buffer; + md.length = service->srv_buf_size; + md.max_size = service->srv_max_req_size; + md.threshold = LNET_MD_THRESH_INF; + md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE; + md.user_ptr = &rqbd->rqbd_cbid; + md.eq_handle = ptlrpc_eq_h; + + rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h); + if (rc == 0) + return (0); + + CERROR("LNetMDAttach failed: %d; \n", rc); + LASSERT (rc == -ENOMEM); + rc = LNetMEUnlink (me_h); + LASSERT (rc == 0); + rqbd->rqbd_refcount = 0; + + return (-ENOMEM); }