X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fptlrpc%2Fniobuf.c;h=c560e72011e8d9f7fff31c099fb8f82f843e106d;hp=8be8f00c4bc83386490ac5bf5436705479338700;hb=8f1c8dbe2389e1ef1e1d3387e343fb9a1bb84198;hpb=6e3ec5812ebd1b5ecf7cae584f429b013ffe7431 diff --git a/lustre/ptlrpc/niobuf.c b/lustre/ptlrpc/niobuf.c index 8be8f00..c560e72 100644 --- a/lustre/ptlrpc/niobuf.c +++ b/lustre/ptlrpc/niobuf.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -26,8 +24,10 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * 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/ @@ -35,15 +35,18 @@ */ #define DEBUG_SUBSYSTEM S_RPC -#ifndef __KERNEL__ -#include -#endif #include #include #include #include +#include #include "ptlrpc_internal.h" +/** + * 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, @@ -95,216 +98,351 @@ static int ptl_send_buf (lnet_handle_md_t *mdh, void *base, int len, RETURN (0); } -int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc) +static void mdunlink_iterate_helper(lnet_handle_md_t *bd_mds, int count) { - struct ptlrpc_connection *conn = desc->bd_export->exp_connection; - int rc; - int rc2; - lnet_md_t md; - __u64 xid; - 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_network_rw); - LASSERT (desc->bd_type == BULK_PUT_SOURCE || - desc->bd_type == BULK_GET_SINK); - desc->bd_success = 0; + int i; - md.user_ptr = &desc->bd_cbid; - md.eq_handle = ptlrpc_eq_h; - md.threshold = 2; /* SENT and ACK/REPLY */ - md.options = PTLRPC_MD_OPTIONS; - ptlrpc_fill_bulk_md(&md, desc); + for (i = 0; i < count; i++) + LNetMDUnlink(bd_mds[i]); +} - LASSERT (desc->bd_cbid.cbid_fn == server_bulk_callback); - LASSERT (desc->bd_cbid.cbid_arg == desc); +#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) +{ + struct obd_export *exp = req->rq_export; + struct ptlrpc_bulk_desc *desc; - /* 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. */ + ENTRY; + LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK); - rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_md_h); - if (rc != 0) { - CERROR("LNetMDBind failed: %d\n", rc); - LASSERT (rc == -ENOMEM); - RETURN(-ENOMEM); - } + desc = ptlrpc_new_bulk(npages, max_brw, type, portal); + if (desc == NULL) + RETURN(NULL); - /* Client's bulk and reply matchbits are the same */ - xid = desc->bd_req->rq_xid; - CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d " - "id %s xid "LPX64"\n", desc->bd_iov_count, - desc->bd_nob, desc->bd_portal, - libcfs_id2str(conn->c_peer), xid); + desc->bd_export = class_export_get(exp); + desc->bd_req = req; - /* Network is about to get at the memory */ - desc->bd_network_rw = 1; + desc->bd_cbid.cbid_fn = server_bulk_callback; + desc->bd_cbid.cbid_arg = desc; - if (desc->bd_type == BULK_PUT_SOURCE) - rc = LNetPut (conn->c_self, desc->bd_md_h, LNET_ACK_REQ, - conn->c_peer, desc->bd_portal, xid, 0, 0); - else - rc = LNetGet (conn->c_self, desc->bd_md_h, - conn->c_peer, desc->bd_portal, xid, 0); + /* NB we don't assign rq_bulk here; server-side requests are + * re-used, and the handler frees the bulk desc explicitly. */ - 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! */ - CERROR("Transfer(%s, %d, "LPX64") failed: %d\n", - libcfs_id2str(conn->c_peer), desc->bd_portal, xid, rc); - rc2 = LNetMDUnlink(desc->bd_md_h); - LASSERT (rc2 == 0); - } + return desc; +} +EXPORT_SYMBOL(ptlrpc_prep_bulk_exp); - RETURN(0); +/** + * 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) +{ + 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); } -/* Server side bulk abort. Idempotent. Not thread-safe (i.e. only - * serialises with completion callback) */ +/** + * Server side bulk abort. Idempotent. Not thread-safe (i.e. only + * serialises with completion callback) + */ void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc) { - struct l_wait_info lwi; - int rc; - - LASSERT(!cfs_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()*/ - - LNetMDUnlink(desc->bd_md_h); - - 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); - } + 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 */ +/** + * 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; - lnet_process_id_t peer; - int rc; - int rc2; - lnet_handle_me_t me_h; - lnet_md_t md; - ENTRY; + 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_network_rw); - 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); - - desc->bd_success = 0; - - peer = desc->bd_import->imp_connection->c_peer; - - md.user_ptr = &desc->bd_cbid; - md.eq_handle = ptlrpc_eq_h; - md.threshold = 1; /* PUT or GET */ - 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); - - LASSERT (desc->bd_cbid.cbid_fn == client_bulk_callback); - LASSERT (desc->bd_cbid.cbid_arg == desc); - - /* 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 */ - LASSERTF (!desc->bd_registered || req->rq_xid != desc->bd_last_xid, - "registered: %d rq_xid: "LPU64" bd_last_xid: "LPU64"\n", - desc->bd_registered, req->rq_xid, desc->bd_last_xid); - desc->bd_registered = 1; - desc->bd_last_xid = req->rq_xid; - - rc = LNetMEAttach(desc->bd_portal, peer, - req->rq_xid, 0, LNET_UNLINK, LNET_INS_AFTER, &me_h); - if (rc != 0) { - CERROR("LNetMEAttach failed: %d\n", rc); - LASSERT (rc == -ENOMEM); - RETURN (-ENOMEM); - } - - /* About to let the network at it... */ - desc->bd_network_rw = 1; - rc = LNetMDAttach(me_h, md, LNET_UNLINK, &desc->bd_md_h); - if (rc != 0) { - CERROR("LNetMDAttach failed: %d\n", rc); - LASSERT (rc == -ENOMEM); - desc->bd_network_rw = 0; - rc2 = LNetMEUnlink (me_h); - LASSERT (rc2 == 0); - RETURN (-ENOMEM); - } - - CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPU64", " - "portal %u\n", - desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink", - desc->bd_iov_count, desc->bd_nob, - req->rq_xid, desc->bd_portal); - 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); } -/* Disconnect a bulk desc from the network. Idempotent. Not - * thread-safe (i.e. only interlocks with completion callback). */ +/** + * 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; - cfs_waitq_t *wq; - struct l_wait_info lwi; - int rc; - ENTRY; + struct ptlrpc_bulk_desc *desc = req->rq_bulk; + struct l_wait_info lwi; + int rc; + ENTRY; - LASSERT(!cfs_in_interrupt()); /* might sleep */ + 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; + /* 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)) /* completed or */ - RETURN(1); /* never registered */ + if (ptlrpc_client_bulk_active(req) == 0) /* completed or */ + RETURN(1); /* never registered */ - LASSERT(desc->bd_req == req); /* bd_req NULL until 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() */ + /* 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); - LNetMDUnlink(desc->bd_md_h); - - if (!ptlrpc_client_bulk_active(req)) /* completed or */ - RETURN(1); /* never registered */ + 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); @@ -313,12 +451,11 @@ int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async) if (async) RETURN(0); - if (req->rq_set != NULL) - wq = &req->rq_set->set_waitq; - else - wq = &req->rq_reply_waitq; - 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), @@ -338,7 +475,8 @@ int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async) static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags) { - struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service; + 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); @@ -350,39 +488,44 @@ static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags) 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_add(&svc->srv_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(&svc->srv_at_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 a error reply during recovery. + * (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(&svc->srv_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=%d:%x/%x len=%d\n", - flags, lustre_msg_get_flags(req->rq_reqmsg), - lustre_msg_is_v1(req->rq_reqmsg), - lustre_msg_get_magic(req->rq_reqmsg), - lustre_msg_get_magic(req->rq_repmsg), req->rq_replen); - } + 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); + } } +/** + * 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_service *svc = req->rq_rqbd->rqbd_service; struct ptlrpc_reply_state *rs = req->rq_reply_state; struct ptlrpc_connection *conn; int rc; @@ -414,11 +557,22 @@ int ptlrpc_send_reply(struct ptlrpc_request *req, int flags) 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); + if (req->rq_type != PTL_RPC_MSG_ERR) req->rq_type = PTL_RPC_MSG_REPLY; lustre_msg_set_type(req->rq_repmsg, req->rq_type); - lustre_msg_set_status(req->rq_repmsg, req->rq_status); + 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); @@ -435,7 +589,6 @@ int ptlrpc_send_reply(struct ptlrpc_request *req, int flags) CERROR("not replying on NULL connection\n"); /* bug 9635 */ return -ENOTCONN; } - cfs_atomic_inc (&svc->srv_outstanding_replies); ptlrpc_rs_addref(rs); /* +1 ref for the network */ rc = sptlrpc_svc_wrap_reply(req); @@ -447,13 +600,12 @@ int ptlrpc_send_reply(struct ptlrpc_request *req, int flags) 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, svc->srv_rep_portal, + &rs->rs_cb_id, conn, + ptlrpc_req2svc(req)->srv_rep_portal, req->rq_xid, req->rq_reply_off); out: - if (unlikely(rc != 0)) { - cfs_atomic_dec (&svc->srv_outstanding_replies); + if (unlikely(rc != 0)) ptlrpc_req_drop_rs(req); - } ptlrpc_connection_put(conn); return rc; } @@ -466,6 +618,10 @@ int ptlrpc_reply (struct ptlrpc_request *req) 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; @@ -480,7 +636,10 @@ int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult) RETURN(rc); } - req->rq_type = PTL_RPC_MSG_ERR; + 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; rc = ptlrpc_send_reply(req, may_be_difficult); RETURN(rc); @@ -491,14 +650,22 @@ 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; lnet_handle_me_t reply_me_h; lnet_md_t reply_md; - struct obd_device *obd = request->rq_import->imp_obd; + struct obd_import *imp = request->rq_import; + struct obd_device *obd = imp->imp_obd; ENTRY; if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC)) @@ -510,39 +677,54 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) /* If this is a re-transmit, we're required to have disengaged * cleanly from the previous attempt */ LASSERT(!request->rq_receiving_reply); - - if (request->rq_import->imp_obd && - request->rq_import->imp_obd->obd_fail) { - CDEBUG(D_HA, "muting rpc for failed imp obd %s\n", - request->rq_import->imp_obd->obd_name); - /* this prevents us from waiting in ptlrpc_queue_wait */ - request->rq_err = 1; + 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 = request->rq_import->imp_connection; - - lustre_msg_set_handle(request->rq_reqmsg, - &request->rq_import->imp_remote_handle); - lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST); - lustre_msg_set_conn_cnt(request->rq_reqmsg, - request->rq_import->imp_conn_cnt); - lustre_msghdr_set_flags(request->rq_reqmsg, - request->rq_import->imp_msghdr_flags); - - if (request->rq_resend) - lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT); + 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) - RETURN(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); } if (!noreply) { @@ -555,7 +737,9 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) if (rc) { /* this prevents us from looping in * ptlrpc_queue_wait */ - request->rq_err = 1; + spin_lock(&request->rq_lock); + request->rq_err = 1; + spin_unlock(&request->rq_lock); request->rq_status = rc; GOTO(cleanup_bulk, rc); } @@ -574,20 +758,20 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) } } - cfs_spin_lock(&request->rq_lock); - /* If the MD attach succeeds, there _will_ be a reply_in callback */ - request->rq_receiving_reply = !noreply; - /* We are responsible for unlinking the reply buffer */ - request->rq_must_unlink = !noreply; - /* 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; - request->rq_reply_truncate = 0; - cfs_spin_unlock(&request->rq_lock); + request->rq_reply_truncated = 0; + spin_unlock(&request->rq_lock); if (!noreply) { reply_md.start = request->rq_repbuf; @@ -601,17 +785,17 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) reply_md.user_ptr = &request->rq_reply_cbid; reply_md.eq_handle = ptlrpc_eq_h; - /* We must see the unlink callback to unset rq_must_unlink, - so we can't auto-unlink */ + /* 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); - cfs_spin_lock(&request->rq_lock); - /* ...but the MD attach didn't succeed... */ - request->rq_receiving_reply = 0; - cfs_spin_unlock(&request->rq_lock); + 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); } @@ -623,20 +807,20 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) /* add references on request for request_out_callback */ ptlrpc_request_addref(request); - if (obd->obd_svc_stats != NULL) - lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR, - cfs_atomic_read(&request->rq_import->imp_inflight)); + 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); + OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5); - cfs_gettimeofday(&request->rq_arrival_time); - 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. */ + 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(request->rq_import); + ptlrpc_pinger_sending_on_import(imp); DEBUG_REQ(D_INFO, request, "send flg=%x", lustre_msg_get_flags(request->rq_reqmsg)); @@ -646,12 +830,13 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) connection, request->rq_request_portal, request->rq_xid, 0); - if (rc == 0) - RETURN(rc); + if (likely(rc == 0)) + GOTO(out, rc); + request->rq_req_unlinked = 1; ptlrpc_req_finished(request); if (noreply) - RETURN(rc); + GOTO(out, rc); cleanup_me: /* MEUnlink is safe; the PUT didn't even get off the ground, and @@ -666,14 +851,21 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) /* 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); +/** + * Register request buffer descriptor for request receiving. + */ int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd) { - struct ptlrpc_service *service = rqbd->rqbd_service; - static lnet_process_id_t match_id = {LNET_NID_ANY, LNET_PID_ANY}; - int rc; + 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; @@ -683,8 +875,13 @@ int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd) if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD)) return (-ENOMEM); - rc = LNetMEAttach(service->srv_req_portal, - match_id, 0, ~0, LNET_UNLINK, LNET_INS_AFTER, &me_h); + /* 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);