From: Liang Zhen Date: Wed, 23 May 2012 06:34:18 +0000 (+0800) Subject: LU-56 ptlrpc: partition data for ptlrpc service X-Git-Tag: 2.2.58~17 X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=commitdiff_plain;h=24564b398f53009521aeda5d653e57fe8b525775 LU-56 ptlrpc: partition data for ptlrpc service We will have multiple partition data & threads for ptlrpc service, this patch is the first step work, we moved quite a lot members of ptlrpc_service to a new structure ptlrpc_service_part. Now we only create one instance of ptlrpc_service_part for each service, but we will have multiple instances for each service very soon (instance per CPT, CPU ParTion). Signed-off-by: Liang Zhen Change-Id: I63d816bdf44a22528c6097fe348060f57d862df3 Reviewed-on: http://review.whamcloud.com/2895 Tested-by: Hudson Reviewed-by: wangdi Tested-by: Maloo Reviewed-by: Andreas Dilger --- diff --git a/lustre/include/lustre_net.h b/lustre/include/lustre_net.h index ff325fb..64b5878 100644 --- a/lustre/include/lustre_net.h +++ b/lustre/include/lustre_net.h @@ -283,8 +283,6 @@ struct ptlrpc_request_set { set_interpreter_func set_interpret; /** opaq argument passed to completion \a set_interpret callback. */ void *set_arg; - /** rq_status of requests that have been freed already */ - int set_rc; /** * Lock for \a set_new_requests manipulations * locked so that any old caller can communicate requests to @@ -294,6 +292,8 @@ struct ptlrpc_request_set { /** List of new yet unsent requests. Only used with ptlrpcd now. */ cfs_list_t set_new_requests; + /** rq_status of requests that have been freed already */ + int set_rc; /** Additional fields used by the flow control extension */ /** Maximum number of RPCs in flight */ int set_max_inflight; @@ -316,6 +316,7 @@ struct ptlrpc_set_cbdata { }; struct ptlrpc_bulk_desc; +struct ptlrpc_service_part; /** * ptlrpc callback & work item stuff @@ -371,7 +372,7 @@ struct ptlrpc_reply_state { /** xid */ __u64 rs_xid; struct obd_export *rs_export; - struct ptlrpc_service *rs_service; + struct ptlrpc_service_part *rs_svcpt; /** Lnet metadata handle for the reply */ lnet_handle_md_t rs_md_h; cfs_atomic_t rs_refcount; @@ -994,9 +995,9 @@ struct ptlrpc_thread { /** * the svc this thread belonged to b=18582 */ - struct ptlrpc_service *t_svc; - cfs_waitq_t t_ctl_waitq; - struct lu_env *t_env; + struct ptlrpc_service_part *t_svcpt; + cfs_waitq_t t_ctl_waitq; + struct lu_env *t_env; }; static inline int thread_is_init(struct ptlrpc_thread *thread) @@ -1072,7 +1073,7 @@ struct ptlrpc_request_buffer_desc { /** History of requests for this buffer */ cfs_list_t rqbd_reqs; /** Back pointer to service for which this buffer is registered */ - struct ptlrpc_service *rqbd_service; + struct ptlrpc_service_part *rqbd_svcpt; /** LNet descriptor */ lnet_handle_md_t rqbd_md_h; int rqbd_refcount; @@ -1130,21 +1131,10 @@ struct ptlrpc_service_ops { * The service is listening on a particular portal (like tcp port) * and perform actions for a specific server like IO service for OST * or general metadata service for MDS. - * - * ptlrpc service has four locks: - * \a srv_lock - * serialize operations on rqbd and requests waiting for preprocess - * \a srv_rq_lock - * serialize operations active requests sent to this portal - * \a srv_at_lock - * serialize adaptive timeout stuff - * \a srv_rs_lock - * serialize operations on RS list (reply states) - * - * We don't have any use-case to take two or more locks at the same time - * for now, so there is no lock order issue. */ struct ptlrpc_service { + /** serialize /proc operations */ + cfs_spinlock_t srv_lock; /** most often accessed fields */ /** chain thru all services */ cfs_list_t srv_list; @@ -1160,13 +1150,6 @@ struct ptlrpc_service { int srv_threads_min; /** thread upper limit */ int srv_threads_max; - /** always increasing number */ - unsigned srv_threads_next_id; - /** # of starting threads */ - int srv_threads_starting; - /** # running threads */ - int srv_threads_running; - /** Root of /proc dir tree for this service */ cfs_proc_dir_entry_t *srv_procroot; /** Pointer to statistic data for this service */ @@ -1197,97 +1180,142 @@ struct ptlrpc_service { /** under unregister_service */ unsigned srv_is_stopping:1; - /** - * serialize the following fields, used for protecting - * rqbd list and incoming requests waiting for preprocess - */ - cfs_spinlock_t srv_lock __cfs_cacheline_aligned; - /** incoming reqs */ - cfs_list_t srv_req_in_queue; - /** total # req buffer descs allocated */ - int srv_nbufs; - /** # posted request buffers */ - int srv_nrqbd_receiving; - /** timeout before re-posting reqs, in tick */ - cfs_duration_t srv_rqbd_timeout; - /** request buffers to be reposted */ - cfs_list_t srv_idle_rqbds; - /** req buffers receiving */ - cfs_list_t srv_active_rqbds; - /** request buffer history */ - cfs_list_t srv_history_rqbds; - /** # request buffers in history */ - int srv_n_history_rqbds; - /** max # request buffers in history */ - int srv_max_history_rqbds; - /** request history */ - cfs_list_t srv_request_history; - /** next request sequence # */ - __u64 srv_request_seq; - /** highest seq culled from history */ - __u64 srv_request_max_cull_seq; - /** - * all threads sleep on this. This wait-queue is signalled when new - * incoming request arrives and when difficult reply has to be handled. - */ - cfs_waitq_t srv_waitq; + /** + * max # request buffers in history, it needs to be convert into + * per-partition value when we have multiple partitions + */ + int srv_max_history_rqbds; + /** + * partition data for ptlrpc service, only one instance so far, + * instance per CPT will come soon + */ + struct ptlrpc_service_part *srv_part; +}; - /** - * serialize the following fields, used for processing requests - * sent to this portal - */ - cfs_spinlock_t srv_rq_lock __cfs_cacheline_aligned; - /** # reqs in either of the queues below */ - /** reqs waiting for service */ - cfs_list_t srv_request_queue; - /** high priority queue */ - cfs_list_t srv_request_hpq; - /** # incoming reqs */ - int srv_n_queued_reqs; - /** # reqs being served */ - int srv_n_active_reqs; - /** # HPreqs being served */ - int srv_n_active_hpreq; - /** # hp requests handled */ - int srv_hpreq_count; - - /** AT stuff */ - /** @{ */ - /** - * serialize the following fields, used for changes on - * adaptive timeout - */ - cfs_spinlock_t srv_at_lock __cfs_cacheline_aligned; - /** estimated rpc service time */ - struct adaptive_timeout srv_at_estimate; - /** reqs waiting for replies */ - struct ptlrpc_at_array srv_at_array; - /** early reply timer */ - cfs_timer_t srv_at_timer; - /** check early replies */ - unsigned srv_at_check; - /** debug */ - cfs_time_t srv_at_checktime; - /** @} */ +/** + * Definition of PortalRPC service partition data. + * Although a service only has one instance of it right now, but we + * will have multiple instances very soon (instance per CPT). + * + * it has four locks: + * \a scp_lock + * serialize operations on rqbd and requests waiting for preprocess + * \a scp_req_lock + * serialize operations active requests sent to this portal + * \a scp_at_lock + * serialize adaptive timeout stuff + * \a scp_rep_lock + * serialize operations on RS list (reply states) + * + * We don't have any use-case to take two or more locks at the same time + * for now, so there is no lock order issue. + */ +struct ptlrpc_service_part { + /** back reference to owner */ + struct ptlrpc_service *scp_service __cfs_cacheline_aligned; + /* CPT id, reserved */ + int scp_cpt; + /** always increasing number */ + int scp_thr_nextid; + /** # of starting threads */ + int scp_nthrs_starting; + /** # of stopping threads, reserved for shrinking threads */ + int scp_nthrs_stopping; + /** # running threads */ + int scp_nthrs_running; + /** service threads list */ + cfs_list_t scp_threads; - /** - * serialize the following fields, used for processing - * replies for this portal - */ - cfs_spinlock_t srv_rs_lock __cfs_cacheline_aligned; - /** all the active replies */ - cfs_list_t srv_active_replies; + /** + * serialize the following fields, used for protecting + * rqbd list and incoming requests waiting for preprocess, + * threads starting & stopping are also protected by this lock. + */ + cfs_spinlock_t scp_lock __cfs_cacheline_aligned; + /** total # req buffer descs allocated */ + int scp_nrqbds_total; + /** # posted request buffers for receiving */ + int scp_nrqbds_posted; + /** # incoming reqs */ + int scp_nreqs_incoming; + /** request buffers to be reposted */ + cfs_list_t scp_rqbd_idle; + /** req buffers receiving */ + cfs_list_t scp_rqbd_posted; + /** incoming reqs */ + cfs_list_t scp_req_incoming; + /** timeout before re-posting reqs, in tick */ + cfs_duration_t scp_rqbd_timeout; + /** + * all threads sleep on this. This wait-queue is signalled when new + * incoming request arrives and when difficult reply has to be handled. + */ + cfs_waitq_t scp_waitq; + + /** request history */ + cfs_list_t scp_hist_reqs; + /** request buffer history */ + cfs_list_t scp_hist_rqbds; + /** # request buffers in history */ + int scp_hist_nrqbds; + /** sequence number for request */ + __u64 scp_hist_seq; + /** highest seq culled from history */ + __u64 scp_hist_seq_culled; + + /** + * serialize the following fields, used for processing requests + * sent to this portal + */ + cfs_spinlock_t scp_req_lock __cfs_cacheline_aligned; + /** # reqs in either of the queues below */ + /** reqs waiting for service */ + cfs_list_t scp_req_pending; + /** high priority queue */ + cfs_list_t scp_hreq_pending; + /** # reqs being served */ + int scp_nreqs_active; + /** # HPreqs being served */ + int scp_nhreqs_active; + /** # hp requests handled */ + int scp_hreq_count; + + /** AT stuff */ + /** @{ */ + /** + * serialize the following fields, used for changes on + * adaptive timeout + */ + cfs_spinlock_t scp_at_lock __cfs_cacheline_aligned; + /** estimated rpc service time */ + struct adaptive_timeout scp_at_estimate; + /** reqs waiting for replies */ + struct ptlrpc_at_array scp_at_array; + /** early reply timer */ + cfs_timer_t scp_at_timer; + /** debug */ + cfs_time_t scp_at_checktime; + /** check early replies */ + unsigned scp_at_check; + /** @} */ + + /** + * serialize the following fields, used for processing + * replies for this portal + */ + cfs_spinlock_t scp_rep_lock __cfs_cacheline_aligned; + /** all the active replies */ + cfs_list_t scp_rep_active; #ifndef __KERNEL__ - /** replies waiting for service */ - cfs_list_t srv_reply_queue; + /** replies waiting for service */ + cfs_list_t scp_rep_queue; #endif - /** List of free reply_states */ - cfs_list_t srv_free_rs_list; - /** waitq to run, when adding stuff to srv_free_rs_list */ - cfs_waitq_t srv_free_rs_waitq; - /** # 'difficult' replies */ - cfs_atomic_t srv_n_difficult_replies; - //struct ptlrpc_srv_ni srv_interfaces[0]; + /** List of free reply_states */ + cfs_list_t scp_rep_idle; + /** waitq to run, when adding stuff to srv_free_rs_list */ + cfs_waitq_t scp_rep_waitq; + /** # 'difficult' replies */ + cfs_atomic_t scp_nreps_difficult; }; /** @@ -1625,7 +1653,6 @@ struct ptlrpc_service *ptlrpc_register_service( void ptlrpc_stop_all_threads(struct ptlrpc_service *svc); int ptlrpc_start_threads(struct ptlrpc_service *svc); -int ptlrpc_start_thread(struct ptlrpc_service *svc); int ptlrpc_unregister_service(struct ptlrpc_service *service); int liblustre_check_services(void *arg); void ptlrpc_daemonize(char *name); @@ -1912,6 +1939,22 @@ static inline int ptlrpc_no_resend(struct ptlrpc_request *req) return req->rq_no_resend; } +static inline int +ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt) +{ + int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate); + + return svcpt->scp_service->srv_watchdog_factor * + max_t(int, at, obd_timeout); +} + +static inline struct ptlrpc_service * +ptlrpc_req2svc(struct ptlrpc_request *req) +{ + LASSERT(req->rq_rqbd != NULL); + return req->rq_rqbd->rqbd_svcpt->scp_service; +} + /* ldlm/ldlm_lib.c */ /** * Target client logic diff --git a/lustre/ldlm/ldlm_lib.c b/lustre/ldlm/ldlm_lib.c index c6ba490..9802ba0 100644 --- a/lustre/ldlm/ldlm_lib.c +++ b/lustre/ldlm/ldlm_lib.c @@ -595,18 +595,20 @@ int server_disconnect_export(struct obd_export *exp) struct ptlrpc_reply_state *rs = cfs_list_entry(exp->exp_outstanding_replies.next, struct ptlrpc_reply_state, rs_exp_list); - struct ptlrpc_service *svc = rs->rs_service; + struct ptlrpc_service_part *svcpt = rs->rs_svcpt; - cfs_spin_lock(&svc->srv_rs_lock); - cfs_list_del_init(&rs->rs_exp_list); - cfs_spin_lock(&rs->rs_lock); - ptlrpc_schedule_difficult_reply(rs); - cfs_spin_unlock(&rs->rs_lock); - cfs_spin_unlock(&svc->srv_rs_lock); - } - cfs_spin_unlock(&exp->exp_lock); + cfs_spin_lock(&svcpt->scp_rep_lock); - RETURN(rc); + cfs_list_del_init(&rs->rs_exp_list); + cfs_spin_lock(&rs->rs_lock); + ptlrpc_schedule_difficult_reply(rs); + cfs_spin_unlock(&rs->rs_lock); + + cfs_spin_unlock(&svcpt->scp_rep_lock); + } + cfs_spin_unlock(&exp->exp_lock); + + RETURN(rc); } /* -------------------------------------------------------------------------- @@ -1517,7 +1519,7 @@ check_and_start_recovery_timer(struct obd_device *obd, if (!new_client && service_time) /* Teach server about old server's estimates, as first guess * at how long new requests will take. */ - at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate, + at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate, service_time); target_start_recovery_timer(obd); @@ -1828,15 +1830,17 @@ static int handle_recovery_req(struct ptlrpc_thread *thread, * this client may come in recovery time */ if (!AT_OFF) { - struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service; - /* If the server sent early reply for this request, - * the client will recalculate the timeout according to - * current server estimate service time, so we will - * use the maxium timeout here for waiting the client - * sending the next req */ - to = max((int)at_est2timeout( - at_get(&svc->srv_at_estimate)), - (int)lustre_msg_get_timeout(req->rq_reqmsg)); + struct ptlrpc_service_part *svcpt; + + svcpt = req->rq_rqbd->rqbd_svcpt; + /* If the server sent early reply for this request, + * the client will recalculate the timeout according to + * current server estimate service time, so we will + * use the maxium timeout here for waiting the client + * sending the next req */ + to = max((int)at_est2timeout( + at_get(&svcpt->scp_at_estimate)), + (int)lustre_msg_get_timeout(req->rq_reqmsg)); /* Add net_latency (see ptlrpc_replay_req) */ to += lustre_msg_get_service_time(req->rq_reqmsg); } @@ -2314,10 +2318,10 @@ int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id) void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) { + struct ptlrpc_service_part *svcpt; int netrc; struct ptlrpc_reply_state *rs; struct obd_export *exp; - struct ptlrpc_service *svc; ENTRY; if (req->rq_no_reply) { @@ -2325,7 +2329,7 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) return; } - svc = req->rq_rqbd->rqbd_service; + svcpt = req->rq_rqbd->rqbd_svcpt; rs = req->rq_reply_state; if (rs == NULL || !rs->rs_difficult) { /* no notifiers */ @@ -2337,7 +2341,7 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) /* must be an export if locks saved */ LASSERT (req->rq_export != NULL); /* req/reply consistent */ - LASSERT (rs->rs_service == svc); + LASSERT(rs->rs_svcpt == svcpt); /* "fresh" reply */ LASSERT (!rs->rs_scheduled); @@ -2374,9 +2378,9 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) netrc = target_send_reply_msg (req, rc, fail_id); - cfs_spin_lock(&svc->srv_rs_lock); + cfs_spin_lock(&svcpt->scp_rep_lock); - cfs_atomic_inc(&svc->srv_n_difficult_replies); + cfs_atomic_inc(&svcpt->scp_nreps_difficult); if (netrc != 0) { /* error sending: reply is off the net. Also we need +1 @@ -2396,12 +2400,12 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) CDEBUG(D_HA, "Schedule reply immediately\n"); ptlrpc_dispatch_difficult_reply(rs); } else { - cfs_list_add (&rs->rs_list, &svc->srv_active_replies); - rs->rs_scheduled = 0; /* allow notifier to schedule */ - } - cfs_spin_unlock(&rs->rs_lock); - cfs_spin_unlock(&svc->srv_rs_lock); - EXIT; + cfs_list_add(&rs->rs_list, &svcpt->scp_rep_active); + rs->rs_scheduled = 0; /* allow notifier to schedule */ + } + cfs_spin_unlock(&rs->rs_lock); + cfs_spin_unlock(&svcpt->scp_rep_lock); + EXIT; } int target_handle_qc_callback(struct ptlrpc_request *req) diff --git a/lustre/ldlm/ldlm_lockd.c b/lustre/ldlm/ldlm_lockd.c index 4aad818..0883897 100644 --- a/lustre/ldlm/ldlm_lockd.c +++ b/lustre/ldlm/ldlm_lockd.c @@ -1058,9 +1058,8 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, LASSERT(req->rq_export); - if (req->rq_rqbd->rqbd_service->srv_stats) - ldlm_svc_get_eopc(dlm_req, - req->rq_rqbd->rqbd_service->srv_stats); + if (ptlrpc_req2svc(req)->srv_stats != NULL) + ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats); if (req->rq_export && req->rq_export->exp_nid_stats && req->rq_export->exp_nid_stats->nid_ldlm_stats) diff --git a/lustre/mdt/mdt_handler.c b/lustre/mdt/mdt_handler.c index e58ac8b..ee01e1c 100644 --- a/lustre/mdt/mdt_handler.c +++ b/lustre/mdt/mdt_handler.c @@ -336,20 +336,20 @@ static int mdt_getstatus(struct mdt_thread_info *info) static int mdt_statfs(struct mdt_thread_info *info) { - struct ptlrpc_request *req = mdt_info_req(info); - struct md_device *next = info->mti_mdt->mdt_child; - struct ptlrpc_service *svc; - struct obd_statfs *osfs; - int rc; + struct ptlrpc_request *req = mdt_info_req(info); + struct md_device *next = info->mti_mdt->mdt_child; + struct ptlrpc_service_part *svcpt; + struct obd_statfs *osfs; + int rc; - ENTRY; + ENTRY; - svc = info->mti_pill->rc_req->rq_rqbd->rqbd_service; + svcpt = info->mti_pill->rc_req->rq_rqbd->rqbd_svcpt; - /* This will trigger a watchdog timeout */ - OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP, - (MDT_SERVICE_WATCHDOG_FACTOR * - at_get(&svc->srv_at_estimate)) + 1); + /* This will trigger a watchdog timeout */ + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP, + (MDT_SERVICE_WATCHDOG_FACTOR * + at_get(&svcpt->scp_at_estimate)) + 1); rc = mdt_check_ucred(info); if (rc) diff --git a/lustre/mdt/mdt_recovery.c b/lustre/mdt/mdt_recovery.c index 43b41d4..08ddc03 100644 --- a/lustre/mdt/mdt_recovery.c +++ b/lustre/mdt/mdt_recovery.c @@ -589,10 +589,10 @@ void mdt_fs_cleanup(const struct lu_env *env, struct mdt_device *mdt) /* reconstruction code */ static void mdt_steal_ack_locks(struct ptlrpc_request *req) { + struct ptlrpc_service_part *svcpt; struct obd_export *exp = req->rq_export; cfs_list_t *tmp; struct ptlrpc_reply_state *oldrep; - struct ptlrpc_service *svc; int i; /* CAVEAT EMPTOR: spinlock order */ @@ -610,8 +610,8 @@ static void mdt_steal_ack_locks(struct ptlrpc_request *req) lustre_msg_get_opc(req->rq_reqmsg), oldrep->rs_opc); - svc = oldrep->rs_service; - cfs_spin_lock (&svc->srv_rs_lock); + svcpt = oldrep->rs_svcpt; + cfs_spin_lock(&svcpt->scp_rep_lock); cfs_list_del_init (&oldrep->rs_exp_list); @@ -631,7 +631,7 @@ static void mdt_steal_ack_locks(struct ptlrpc_request *req) ptlrpc_schedule_difficult_reply (oldrep); cfs_spin_unlock(&oldrep->rs_lock); - cfs_spin_unlock (&svc->srv_rs_lock); + cfs_spin_unlock(&svcpt->scp_rep_lock); break; } cfs_spin_unlock(&exp->exp_lock); diff --git a/lustre/ost/ost_handler.c b/lustre/ost/ost_handler.c index 05e51df..20cd111 100644 --- a/lustre/ost/ost_handler.c +++ b/lustre/ost/ost_handler.c @@ -1689,12 +1689,13 @@ struct ost_prolong_data { */ static inline int prolong_timeout(struct ptlrpc_request *req) { - struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service; + struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt; - if (AT_OFF) - return obd_timeout / 2; + if (AT_OFF) + return obd_timeout / 2; - return max(at_est2timeout(at_get(&svc->srv_at_estimate)), ldlm_timeout); + return max(at_est2timeout(at_get(&svcpt->scp_at_estimate)), + ldlm_timeout); } static void ost_prolong_lock_one(struct ost_prolong_data *opd, @@ -2189,11 +2190,11 @@ int ost_handle(struct ptlrpc_request *req) req_capsule_set(&req->rq_pill, &RQF_OST_BRW_WRITE); CDEBUG(D_INODE, "write\n"); /* req->rq_request_portal would be nice, if it was set */ - if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){ - CERROR("%s: deny write request from %s to portal %u\n", - req->rq_export->exp_obd->obd_name, - obd_export_nid2str(req->rq_export), - req->rq_rqbd->rqbd_service->srv_req_portal); + if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL) { + CERROR("%s: deny write request from %s to portal %u\n", + req->rq_export->exp_obd->obd_name, + obd_export_nid2str(req->rq_export), + ptlrpc_req2svc(req)->srv_req_portal); GOTO(out, rc = -EPROTO); } if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET)) @@ -2210,11 +2211,11 @@ int ost_handle(struct ptlrpc_request *req) req_capsule_set(&req->rq_pill, &RQF_OST_BRW_READ); CDEBUG(D_INODE, "read\n"); /* req->rq_request_portal would be nice, if it was set */ - if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){ - CERROR("%s: deny read request from %s to portal %u\n", - req->rq_export->exp_obd->obd_name, - obd_export_nid2str(req->rq_export), - req->rq_rqbd->rqbd_service->srv_req_portal); + if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL) { + CERROR("%s: deny read request from %s to portal %u\n", + req->rq_export->exp_obd->obd_name, + obd_export_nid2str(req->rq_export), + ptlrpc_req2svc(req)->srv_req_portal); GOTO(out, rc = -EPROTO); } if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET)) diff --git a/lustre/ptlrpc/events.c b/lustre/ptlrpc/events.c index ab6da13..5565049 100644 --- a/lustre/ptlrpc/events.c +++ b/lustre/ptlrpc/events.c @@ -221,13 +221,74 @@ void client_bulk_callback (lnet_event_t *ev) } /* + * 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. + */ +#define REQS_ALL_BITS(svcpt) ((int)(sizeof((svcpt)->scp_hist_seq) * 8)) +#define REQS_SEC_BITS 32 +#define REQS_USEC_BITS 16 +/* will be replaced by bits for total service partition number soon */ +#define REQS_CPT_BITS(svcpt) 0 +#define REQS_SEQ_BITS(svcpt) (REQS_ALL_BITS(svcpt) - REQS_CPT_BITS(svcpt) -\ + REQS_SEC_BITS - REQS_USEC_BITS) + +#define REQS_SEQ_SHIFT(svcpt) (REQS_CPT_BITS(svcpt)) +#define REQS_USEC_SHIFT(svcpt) (REQS_SEQ_SHIFT(svcpt) + REQS_SEQ_BITS(svcpt)) +#define REQS_SEC_SHIFT(svcpt) (REQS_USEC_SHIFT(svcpt) + REQS_USEC_BITS) + +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_usec >> 4; /* 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 */ + + LASSERT(REQS_SEQ_BITS(svcpt) > 0); + + new_seq = (sec << REQS_SEC_SHIFT(svcpt)) | + (usec << REQS_USEC_SHIFT(svcpt)) | 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 { + /* 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_CPT_BITS(svcpt)); + new_seq = svcpt->scp_hist_seq; + } + + req->rq_history_seq = new_seq; + + cfs_list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs); +} + +/* * Server's incoming request callback */ void request_in_callback(lnet_event_t *ev) { - struct ptlrpc_cb_id *cbid = ev->md.user_ptr; - struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg; - struct ptlrpc_service *service = rqbd->rqbd_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; ENTRY; @@ -289,21 +350,20 @@ void request_in_callback(lnet_event_t *ev) CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer)); - cfs_spin_lock(&service->srv_lock); + cfs_spin_lock(&svcpt->scp_lock); - req->rq_history_seq = service->srv_request_seq++; - cfs_list_add_tail(&req->rq_history_list, &service->srv_request_history); + ptlrpc_req_add_history(svcpt, req); - if (ev->unlinked) { - service->srv_nrqbd_receiving--; - CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n", - service->srv_nrqbd_receiving); - - /* 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 && - service->srv_nrqbd_receiving == 0) + 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); @@ -313,15 +373,15 @@ void request_in_callback(lnet_event_t *ev) rqbd->rqbd_refcount++; } - cfs_list_add_tail(&req->rq_list, &service->srv_req_in_queue); - service->srv_n_queued_reqs++; + cfs_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... */ - cfs_waitq_signal(&service->srv_waitq); + /* NB everything can disappear under us once the request + * has been queued and we unlock, so do the wake now... */ + cfs_waitq_signal(&svcpt->scp_waitq); - cfs_spin_unlock(&service->srv_lock); - EXIT; + cfs_spin_unlock(&svcpt->scp_lock); + EXIT; } /* @@ -329,9 +389,9 @@ void request_in_callback(lnet_event_t *ev) */ void reply_out_callback(lnet_event_t *ev) { - struct ptlrpc_cb_id *cbid = ev->md.user_ptr; - struct ptlrpc_reply_state *rs = cbid->cbid_arg; - struct ptlrpc_service *svc = rs->rs_service; + 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 == LNET_EVENT_SEND || @@ -352,17 +412,19 @@ void reply_out_callback(lnet_event_t *ev) if (ev->unlinked) { /* Last network callback. The net's ref on 'rs' stays put * until ptlrpc_handle_rs() is done with it */ - cfs_spin_lock(&svc->srv_rs_lock); - cfs_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) - ptlrpc_schedule_difficult_reply (rs); - cfs_spin_unlock(&rs->rs_lock); - cfs_spin_unlock(&svc->srv_rs_lock); - } - - EXIT; + cfs_spin_lock(&svcpt->scp_rep_lock); + cfs_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) + ptlrpc_schedule_difficult_reply(rs); + + cfs_spin_unlock(&rs->rs_lock); + cfs_spin_unlock(&svcpt->scp_rep_lock); + } + EXIT; } #ifdef HAVE_SERVER_SUPPORT diff --git a/lustre/ptlrpc/lproc_ptlrpc.c b/lustre/ptlrpc/lproc_ptlrpc.c index 5e29563..0d6ceec 100644 --- a/lustre/ptlrpc/lproc_ptlrpc.c +++ b/lustre/ptlrpc/lproc_ptlrpc.c @@ -248,12 +248,12 @@ void ptlrpc_lprocfs_register(struct proc_dir_entry *root, char *dir, static int ptlrpc_lprocfs_read_req_history_len(char *page, char **start, off_t off, - int count, int *eof, void *data) + int count, int *eof, void *data) { - struct ptlrpc_service *svc = data; + struct ptlrpc_service *svc = data; - *eof = 1; - return snprintf(page, count, "%d\n", svc->srv_n_history_rqbds); + *eof = 1; + return snprintf(page, count, "%d\n", svc->srv_part->scp_hist_nrqbds); } static int @@ -288,11 +288,11 @@ ptlrpc_lprocfs_write_req_history_max(struct file *file, const char *buffer, if (val > cfs_num_physpages/(2 * bufpages)) return -ERANGE; - cfs_spin_lock(&svc->srv_lock); - svc->srv_max_history_rqbds = val; - cfs_spin_unlock(&svc->srv_lock); + cfs_spin_lock(&svc->srv_lock); + svc->srv_max_history_rqbds = val; + cfs_spin_unlock(&svc->srv_lock); - return count; + return count; } static int @@ -318,23 +318,27 @@ ptlrpc_lprocfs_wr_threads_min(struct file *file, const char *buffer, if (val < 2) return -ERANGE; - if (val > svc->srv_threads_max) - return -ERANGE; + cfs_spin_lock(&svc->srv_lock); + if (val > svc->srv_threads_max) { + cfs_spin_unlock(&svc->srv_lock); + return -ERANGE; + } - cfs_spin_lock(&svc->srv_lock); - svc->srv_threads_min = val; - cfs_spin_unlock(&svc->srv_lock); + svc->srv_threads_min = val; + cfs_spin_unlock(&svc->srv_lock); - return count; + return count; } static int ptlrpc_lprocfs_rd_threads_started(char *page, char **start, off_t off, - int count, int *eof, void *data) + int count, int *eof, void *data) { - struct ptlrpc_service *svc = data; + struct ptlrpc_service *svc = data; - return snprintf(page, count, "%d\n", svc->srv_threads_running); + LASSERT(svc->srv_part != NULL); + return snprintf(page, count, "%d\n", + svc->srv_part->scp_nthrs_running); } static int @@ -360,14 +364,16 @@ ptlrpc_lprocfs_wr_threads_max(struct file *file, const char *buffer, if (val < 2) return -ERANGE; - if (val < svc->srv_threads_min) - return -ERANGE; + cfs_spin_lock(&svc->srv_lock); + if (val < svc->srv_threads_min) { + cfs_spin_unlock(&svc->srv_lock); + return -ERANGE; + } - cfs_spin_lock(&svc->srv_lock); - svc->srv_threads_max = val; - cfs_spin_unlock(&svc->srv_lock); + svc->srv_threads_max = val; + cfs_spin_unlock(&svc->srv_lock); - return count; + return count; } struct ptlrpc_srh_iterator { @@ -376,15 +382,15 @@ struct ptlrpc_srh_iterator { }; int -ptlrpc_lprocfs_svc_req_history_seek(struct ptlrpc_service *svc, - struct ptlrpc_srh_iterator *srhi, - __u64 seq) +ptlrpc_lprocfs_svc_req_history_seek(struct ptlrpc_service_part *svcpt, + struct ptlrpc_srh_iterator *srhi, + __u64 seq) { - cfs_list_t *e; - struct ptlrpc_request *req; + cfs_list_t *e; + struct ptlrpc_request *req; - if (srhi->srhi_req != NULL && - srhi->srhi_seq > svc->srv_request_max_cull_seq && + if (srhi->srhi_req != NULL && + srhi->srhi_seq > svcpt->scp_hist_seq_culled && srhi->srhi_seq <= seq) { /* If srhi_req was set previously, hasn't been culled and * we're searching for a seq on or after it (i.e. more @@ -393,14 +399,14 @@ ptlrpc_lprocfs_svc_req_history_seek(struct ptlrpc_service *svc, * be near the head), we shouldn't have to do long * re-scans */ LASSERT (srhi->srhi_seq == srhi->srhi_req->rq_history_seq); - LASSERT (!cfs_list_empty(&svc->srv_request_history)); - e = &srhi->srhi_req->rq_history_list; - } else { - /* search from start */ - e = svc->srv_request_history.next; - } - - while (e != &svc->srv_request_history) { + LASSERT(!cfs_list_empty(&svcpt->scp_hist_reqs)); + e = &srhi->srhi_req->rq_history_list; + } else { + /* search from start */ + e = svcpt->scp_hist_reqs.next; + } + + while (e != &svcpt->scp_hist_reqs) { req = cfs_list_entry(e, struct ptlrpc_request, rq_history_list); if (req->rq_history_seq >= seq) { @@ -428,9 +434,9 @@ ptlrpc_lprocfs_svc_req_history_start(struct seq_file *s, loff_t *pos) srhi->srhi_seq = 0; srhi->srhi_req = NULL; - cfs_spin_lock(&svc->srv_lock); - rc = ptlrpc_lprocfs_svc_req_history_seek(svc, srhi, *pos); - cfs_spin_unlock(&svc->srv_lock); + cfs_spin_lock(&svc->srv_part->scp_lock); + rc = ptlrpc_lprocfs_svc_req_history_seek(svc->srv_part, srhi, *pos); + cfs_spin_unlock(&svc->srv_part->scp_lock); if (rc == 0) { *pos = srhi->srhi_seq; @@ -452,15 +458,16 @@ ptlrpc_lprocfs_svc_req_history_stop(struct seq_file *s, void *iter) static void * ptlrpc_lprocfs_svc_req_history_next(struct seq_file *s, - void *iter, loff_t *pos) + void *iter, loff_t *pos) { - struct ptlrpc_service *svc = s->private; - struct ptlrpc_srh_iterator *srhi = iter; - int rc; + struct ptlrpc_service *svc = s->private; + struct ptlrpc_service_part *svcpt = svc->srv_part; + struct ptlrpc_srh_iterator *srhi = iter; + int rc; - cfs_spin_lock(&svc->srv_lock); - rc = ptlrpc_lprocfs_svc_req_history_seek(svc, srhi, *pos + 1); - cfs_spin_unlock(&svc->srv_lock); + cfs_spin_lock(&svcpt->scp_lock); + rc = ptlrpc_lprocfs_svc_req_history_seek(svcpt, srhi, *pos + 1); + cfs_spin_unlock(&svcpt->scp_lock); if (rc != 0) { OBD_FREE(srhi, sizeof(*srhi)); @@ -503,14 +510,15 @@ EXPORT_SYMBOL(target_print_req); static int ptlrpc_lprocfs_svc_req_history_show(struct seq_file *s, void *iter) { - struct ptlrpc_service *svc = s->private; - struct ptlrpc_srh_iterator *srhi = iter; - struct ptlrpc_request *req; - int rc; + struct ptlrpc_service *svc = s->private; + struct ptlrpc_service_part *svcpt = svc->srv_part; + struct ptlrpc_srh_iterator *srhi = iter; + struct ptlrpc_request *req; + int rc; - cfs_spin_lock(&svc->srv_lock); + cfs_spin_lock(&svcpt->scp_lock); - rc = ptlrpc_lprocfs_svc_req_history_seek(svc, srhi, srhi->srhi_seq); + rc = ptlrpc_lprocfs_svc_req_history_seek(svcpt, srhi, srhi->srhi_seq); if (rc == 0) { req = srhi->srhi_req; @@ -534,9 +542,8 @@ static int ptlrpc_lprocfs_svc_req_history_show(struct seq_file *s, void *iter) svc->srv_ops.so_req_printer(s, srhi->srhi_req); } - cfs_spin_unlock(&svc->srv_lock); - - return rc; + cfs_spin_unlock(&svcpt->scp_lock); + return rc; } static int @@ -566,19 +573,24 @@ ptlrpc_lprocfs_svc_req_history_open(struct inode *inode, struct file *file) /* See also lprocfs_rd_timeouts */ static int ptlrpc_lprocfs_rd_timeouts(char *page, char **start, off_t off, - int count, int *eof, void *data) -{ - struct ptlrpc_service *svc = data; - unsigned int cur, worst; - time_t worstt; - struct dhms ts; - int rc = 0; - - *eof = 1; - cur = at_get(&svc->srv_at_estimate); - worst = svc->srv_at_estimate.at_worst_ever; - worstt = svc->srv_at_estimate.at_worst_time; - s2dhms(&ts, cfs_time_current_sec() - worstt); + int count, int *eof, void *data) +{ + struct ptlrpc_service *svc = data; + struct ptlrpc_service_part *svcpt; + struct dhms ts; + time_t worstt; + unsigned int cur; + unsigned int worst; + int rc = 0; + + svcpt = svc->srv_part; + LASSERT(svcpt != NULL); + + *eof = 1; + cur = at_get(&svcpt->scp_at_estimate); + worst = svcpt->scp_at_estimate.at_worst_ever; + worstt = svcpt->scp_at_estimate.at_worst_time; + s2dhms(&ts, cfs_time_current_sec() - worstt); if (AT_OFF) rc += snprintf(page + rc, count - rc, "adaptive timeouts off, using obd_timeout %u\n", @@ -587,9 +599,8 @@ static int ptlrpc_lprocfs_rd_timeouts(char *page, char **start, off_t off, "%10s : cur %3u worst %3u (at %ld, "DHMS_FMT" ago) ", "service", cur, worst, worstt, DHMS_VARS(&ts)); - rc = lprocfs_at_hist_helper(page, count, rc, - &svc->srv_at_estimate); - return rc; + rc = lprocfs_at_hist_helper(page, count, rc, &svcpt->scp_at_estimate); + return rc; } static int ptlrpc_lprocfs_rd_hp_ratio(char *page, char **start, off_t off, @@ -612,10 +623,11 @@ static int ptlrpc_lprocfs_wr_hp_ratio(struct file *file, const char *buffer, if (val < 0) return -ERANGE; - cfs_spin_lock(&svc->srv_lock); - svc->srv_hpreq_ratio = val; - cfs_spin_unlock(&svc->srv_lock); - return count; + cfs_spin_lock(&svc->srv_lock); + svc->srv_hpreq_ratio = val; + cfs_spin_unlock(&svc->srv_lock); + + return count; } void ptlrpc_lprocfs_register_service(struct proc_dir_entry *entry, diff --git a/lustre/ptlrpc/niobuf.c b/lustre/ptlrpc/niobuf.c index c65ae1c..69c616f 100644 --- a/lustre/ptlrpc/niobuf.c +++ b/lustre/ptlrpc/niobuf.c @@ -395,7 +395,8 @@ 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); @@ -407,12 +408,14 @@ 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_measured(&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); @@ -424,7 +427,7 @@ static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags) lustre_msg_set_timeout(req->rq_repmsg, 0); else lustre_msg_set_timeout(req->rq_repmsg, - at_get(&svc->srv_at_estimate)); + at_get(&svcpt->scp_at_estimate)); if (req->rq_reqmsg && !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) { @@ -444,7 +447,6 @@ static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags) */ 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; @@ -518,7 +520,8 @@ 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)) @@ -760,9 +763,9 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) */ 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; diff --git a/lustre/ptlrpc/pack_generic.c b/lustre/ptlrpc/pack_generic.c index 545d767..edc835d 100644 --- a/lustre/ptlrpc/pack_generic.c +++ b/lustre/ptlrpc/pack_generic.c @@ -275,49 +275,51 @@ do { \ # define PTLRPC_RS_DEBUG_LRU_DEL(rs) do {} while(0) #endif -struct ptlrpc_reply_state *lustre_get_emerg_rs(struct ptlrpc_service *svc) -{ - struct ptlrpc_reply_state *rs = NULL; - - cfs_spin_lock(&svc->srv_rs_lock); - /* See if we have anything in a pool, and wait if nothing */ - while (cfs_list_empty(&svc->srv_free_rs_list)) { - struct l_wait_info lwi; - int rc; - cfs_spin_unlock(&svc->srv_rs_lock); - /* If we cannot get anything for some long time, we better - bail out instead of waiting infinitely */ - lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL); - rc = l_wait_event(svc->srv_free_rs_waitq, - !cfs_list_empty(&svc->srv_free_rs_list), - &lwi); - if (rc) - goto out; - cfs_spin_lock(&svc->srv_rs_lock); - } - - rs = cfs_list_entry(svc->srv_free_rs_list.next, - struct ptlrpc_reply_state, rs_list); - cfs_list_del(&rs->rs_list); - cfs_spin_unlock(&svc->srv_rs_lock); - LASSERT(rs); - memset(rs, 0, svc->srv_max_reply_size); - rs->rs_service = svc; - rs->rs_prealloc = 1; +struct ptlrpc_reply_state * +lustre_get_emerg_rs(struct ptlrpc_service_part *svcpt) +{ + struct ptlrpc_reply_state *rs = NULL; + + cfs_spin_lock(&svcpt->scp_rep_lock); + + /* See if we have anything in a pool, and wait if nothing */ + while (cfs_list_empty(&svcpt->scp_rep_idle)) { + struct l_wait_info lwi; + int rc; + + cfs_spin_unlock(&svcpt->scp_rep_lock); + /* If we cannot get anything for some long time, we better + * bail out instead of waiting infinitely */ + lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL); + rc = l_wait_event(svcpt->scp_rep_waitq, + !cfs_list_empty(&svcpt->scp_rep_idle), &lwi); + if (rc != 0) + goto out; + cfs_spin_lock(&svcpt->scp_rep_lock); + } + + rs = cfs_list_entry(svcpt->scp_rep_idle.next, + struct ptlrpc_reply_state, rs_list); + cfs_list_del(&rs->rs_list); + + cfs_spin_unlock(&svcpt->scp_rep_lock); + + LASSERT(rs != NULL); + memset(rs, 0, svcpt->scp_service->srv_max_reply_size); + rs->rs_svcpt = svcpt; + rs->rs_prealloc = 1; out: - return rs; + return rs; } void lustre_put_emerg_rs(struct ptlrpc_reply_state *rs) { - struct ptlrpc_service *svc = rs->rs_service; - - LASSERT(svc); + struct ptlrpc_service_part *svcpt = rs->rs_svcpt; - cfs_spin_lock(&svc->srv_rs_lock); - cfs_list_add(&rs->rs_list, &svc->srv_free_rs_list); - cfs_spin_unlock(&svc->srv_rs_lock); - cfs_waitq_signal(&svc->srv_free_rs_waitq); + cfs_spin_lock(&svcpt->scp_rep_lock); + cfs_list_add(&rs->rs_list, &svcpt->scp_rep_idle); + cfs_spin_unlock(&svcpt->scp_rep_lock); + cfs_waitq_signal(&svcpt->scp_rep_waitq); } int lustre_pack_reply_v2(struct ptlrpc_request *req, int count, @@ -344,7 +346,7 @@ int lustre_pack_reply_v2(struct ptlrpc_request *req, int count, cfs_atomic_set(&rs->rs_refcount, 1); /* 1 ref for rq_reply_state */ rs->rs_cb_id.cbid_fn = reply_out_callback; rs->rs_cb_id.cbid_arg = rs; - rs->rs_service = req->rq_rqbd->rqbd_service; + rs->rs_svcpt = req->rq_rqbd->rqbd_svcpt; CFS_INIT_LIST_HEAD(&rs->rs_exp_list); CFS_INIT_LIST_HEAD(&rs->rs_obd_list); CFS_INIT_LIST_HEAD(&rs->rs_list); diff --git a/lustre/ptlrpc/ptlrpc_internal.h b/lustre/ptlrpc/ptlrpc_internal.h index 7ce2dcc..2282bf8 100644 --- a/lustre/ptlrpc/ptlrpc_internal.h +++ b/lustre/ptlrpc/ptlrpc_internal.h @@ -47,6 +47,7 @@ struct ldlm_res_id; struct ptlrpc_request_set; extern int test_req_buffer_pressure; +int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt); /* ptlrpcd.c */ int ptlrpcd_start(int index, int max, const char *name, struct ptlrpcd_ctl *pc); @@ -93,7 +94,8 @@ void ptlrpc_add_bulk_page(struct ptlrpc_bulk_desc *desc, cfs_page_t *page, int pageoffset, int len); /* pack_generic.c */ -struct ptlrpc_reply_state *lustre_get_emerg_rs(struct ptlrpc_service *svc); +struct ptlrpc_reply_state * +lustre_get_emerg_rs(struct ptlrpc_service_part *svcpt); void lustre_put_emerg_rs(struct ptlrpc_reply_state *rs); /* pinger.c */ diff --git a/lustre/ptlrpc/ptlrpc_module.c b/lustre/ptlrpc/ptlrpc_module.c index cd06da6..f9487b2 100644 --- a/lustre/ptlrpc/ptlrpc_module.c +++ b/lustre/ptlrpc/ptlrpc_module.c @@ -231,7 +231,6 @@ EXPORT_SYMBOL(ptlrpc_commit_replies); EXPORT_SYMBOL(ptlrpc_register_service); EXPORT_SYMBOL(ptlrpc_stop_all_threads); EXPORT_SYMBOL(ptlrpc_start_threads); -EXPORT_SYMBOL(ptlrpc_start_thread); EXPORT_SYMBOL(ptlrpc_unregister_service); EXPORT_SYMBOL(ptlrpc_service_health_check); EXPORT_SYMBOL(ptlrpc_hpreq_reorder); diff --git a/lustre/ptlrpc/sec.c b/lustre/ptlrpc/sec.c index 66cf583..d876a22 100644 --- a/lustre/ptlrpc/sec.c +++ b/lustre/ptlrpc/sec.c @@ -2108,7 +2108,7 @@ int sptlrpc_svc_alloc_rs(struct ptlrpc_request *req, int msglen) rc = policy->sp_sops->alloc_rs(req, msglen); if (unlikely(rc == -ENOMEM)) { /* failed alloc, try emergency pool */ - rs = lustre_get_emerg_rs(req->rq_rqbd->rqbd_service); + rs = lustre_get_emerg_rs(req->rq_rqbd->rqbd_svcpt); if (rs == NULL) RETURN(-ENOMEM); diff --git a/lustre/ptlrpc/service.c b/lustre/ptlrpc/service.c index d807e49..c0960e3 100644 --- a/lustre/ptlrpc/service.c +++ b/lustre/ptlrpc/service.c @@ -63,22 +63,23 @@ CFS_MODULE_PARM(at_extra, "i", int, 0644, /* forward ref */ -static int ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc); +static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt); static void ptlrpc_hpreq_fini(struct ptlrpc_request *req); static CFS_LIST_HEAD(ptlrpc_all_services); cfs_spinlock_t ptlrpc_all_services_lock; struct ptlrpc_request_buffer_desc * -ptlrpc_alloc_rqbd (struct ptlrpc_service *svc) +ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt) { - struct ptlrpc_request_buffer_desc *rqbd; + struct ptlrpc_service *svc = svcpt->scp_service; + struct ptlrpc_request_buffer_desc *rqbd; - OBD_ALLOC_PTR(rqbd); - if (rqbd == NULL) - return (NULL); + OBD_ALLOC_PTR(rqbd); + if (rqbd == NULL) + return NULL; - rqbd->rqbd_service = svc; + rqbd->rqbd_svcpt = svcpt; rqbd->rqbd_refcount = 0; rqbd->rqbd_cbid.cbid_fn = request_in_callback; rqbd->rqbd_cbid.cbid_arg = rqbd; @@ -90,34 +91,35 @@ ptlrpc_alloc_rqbd (struct ptlrpc_service *svc) return (NULL); } - cfs_spin_lock(&svc->srv_lock); - cfs_list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds); - svc->srv_nbufs++; - cfs_spin_unlock(&svc->srv_lock); + cfs_spin_lock(&svcpt->scp_lock); + cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle); + svcpt->scp_nrqbds_total++; + cfs_spin_unlock(&svcpt->scp_lock); - return (rqbd); + return rqbd; } void -ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd) +ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd) { - struct ptlrpc_service *svc = rqbd->rqbd_service; + struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt; - LASSERT (rqbd->rqbd_refcount == 0); - LASSERT (cfs_list_empty(&rqbd->rqbd_reqs)); + LASSERT(rqbd->rqbd_refcount == 0); + LASSERT(cfs_list_empty(&rqbd->rqbd_reqs)); - cfs_spin_lock(&svc->srv_lock); - cfs_list_del(&rqbd->rqbd_list); - svc->srv_nbufs--; - cfs_spin_unlock(&svc->srv_lock); + cfs_spin_lock(&svcpt->scp_lock); + cfs_list_del(&rqbd->rqbd_list); + svcpt->scp_nrqbds_total--; + cfs_spin_unlock(&svcpt->scp_lock); - OBD_FREE_LARGE(rqbd->rqbd_buffer, svc->srv_buf_size); - OBD_FREE_PTR(rqbd); + OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size); + OBD_FREE_PTR(rqbd); } int -ptlrpc_grow_req_bufs(struct ptlrpc_service *svc) +ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt) { + struct ptlrpc_service *svc = svcpt->scp_service; struct ptlrpc_request_buffer_desc *rqbd; int rc = 0; int i; @@ -125,10 +127,10 @@ ptlrpc_grow_req_bufs(struct ptlrpc_service *svc) for (i = 0; i < svc->srv_nbuf_per_group; i++) { /* NB: another thread might be doing this as well, we need to * make sure that it wouldn't over-allocate, see LU-1212. */ - if (svc->srv_nrqbd_receiving >= svc->srv_nbuf_per_group) - break; + if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group) + break; - rqbd = ptlrpc_alloc_rqbd(svc); + rqbd = ptlrpc_alloc_rqbd(svcpt); if (rqbd == NULL) { CERROR("%s: Can't allocate request buffer\n", @@ -137,18 +139,18 @@ ptlrpc_grow_req_bufs(struct ptlrpc_service *svc) break; } - if (ptlrpc_server_post_idle_rqbds(svc) < 0) { - rc = -EAGAIN; - break; - } - } + if (ptlrpc_server_post_idle_rqbds(svcpt) < 0) { + rc = -EAGAIN; + break; + } + } - CDEBUG(D_RPCTRACE, - "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n", - svc->srv_name, i, svc->srv_buf_size, - svc->srv_nrqbd_receiving, svc->srv_nbufs, rc); + CDEBUG(D_RPCTRACE, + "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n", + svc->srv_name, i, svc->srv_buf_size, + svcpt->scp_nrqbds_posted, svcpt->scp_nrqbds_total, rc); - return rc; + return rc; } /** @@ -197,9 +199,9 @@ struct ptlrpc_hr_service { }; struct rs_batch { - cfs_list_t rsb_replies; - struct ptlrpc_service *rsb_svc; - unsigned int rsb_n_replies; + cfs_list_t rsb_replies; + unsigned int rsb_n_replies; + struct ptlrpc_service_part *rsb_svcpt; }; /** @@ -271,15 +273,15 @@ static void rs_batch_dispatch(struct rs_batch *b) */ static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs) { - struct ptlrpc_service *svc = rs->rs_service; + struct ptlrpc_service_part *svcpt = rs->rs_svcpt; - if (svc != b->rsb_svc || b->rsb_n_replies >= MAX_SCHEDULED) { - if (b->rsb_svc != NULL) { - rs_batch_dispatch(b); - cfs_spin_unlock(&b->rsb_svc->srv_rs_lock); - } - cfs_spin_lock(&svc->srv_rs_lock); - b->rsb_svc = svc; + if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) { + if (b->rsb_svcpt != NULL) { + rs_batch_dispatch(b); + cfs_spin_unlock(&b->rsb_svcpt->scp_rep_lock); + } + cfs_spin_lock(&svcpt->scp_rep_lock); + b->rsb_svcpt = svcpt; } cfs_spin_lock(&rs->rs_lock); rs->rs_scheduled_ever = 1; @@ -301,10 +303,10 @@ static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs) */ static void rs_batch_fini(struct rs_batch *b) { - if (b->rsb_svc != 0) { - rs_batch_dispatch(b); - cfs_spin_unlock(&b->rsb_svc->srv_rs_lock); - } + if (b->rsb_svcpt != NULL) { + rs_batch_dispatch(b); + cfs_spin_unlock(&b->rsb_svcpt->scp_rep_lock); + } } #define DECLARE_RS_BATCH(b) struct rs_batch b @@ -338,16 +340,16 @@ void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs) cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait); EXIT; #else - cfs_list_add_tail(&rs->rs_list, &rs->rs_service->srv_reply_queue); + cfs_list_add_tail(&rs->rs_list, &rs->rs_svcpt->scp_rep_queue); #endif } void -ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs) +ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs) { - ENTRY; + ENTRY; - LASSERT_SPIN_LOCKED(&rs->rs_service->srv_rs_lock); + LASSERT_SPIN_LOCKED(&rs->rs_svcpt->scp_rep_lock); LASSERT_SPIN_LOCKED(&rs->rs_lock); LASSERT (rs->rs_difficult); rs->rs_scheduled_ever = 1; /* flag any notification attempt */ @@ -391,58 +393,61 @@ void ptlrpc_commit_replies(struct obd_export *exp) } static int -ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc) +ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt) { - struct ptlrpc_request_buffer_desc *rqbd; - int rc; - int posted = 0; + struct ptlrpc_request_buffer_desc *rqbd; + int rc; + int posted = 0; - for (;;) { - cfs_spin_lock(&svc->srv_lock); + for (;;) { + cfs_spin_lock(&svcpt->scp_lock); - if (cfs_list_empty (&svc->srv_idle_rqbds)) { - cfs_spin_unlock(&svc->srv_lock); - return (posted); - } + if (cfs_list_empty(&svcpt->scp_rqbd_idle)) { + cfs_spin_unlock(&svcpt->scp_lock); + return posted; + } - rqbd = cfs_list_entry(svc->srv_idle_rqbds.next, - struct ptlrpc_request_buffer_desc, - rqbd_list); - cfs_list_del (&rqbd->rqbd_list); + rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next, + struct ptlrpc_request_buffer_desc, + rqbd_list); + cfs_list_del(&rqbd->rqbd_list); - /* assume we will post successfully */ - svc->srv_nrqbd_receiving++; - cfs_list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds); + /* assume we will post successfully */ + svcpt->scp_nrqbds_posted++; + cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted); - cfs_spin_unlock(&svc->srv_lock); + cfs_spin_unlock(&svcpt->scp_lock); - rc = ptlrpc_register_rqbd(rqbd); - if (rc != 0) - break; + rc = ptlrpc_register_rqbd(rqbd); + if (rc != 0) + break; - posted = 1; - } + posted = 1; + } - cfs_spin_lock(&svc->srv_lock); + cfs_spin_lock(&svcpt->scp_lock); - svc->srv_nrqbd_receiving--; - cfs_list_del(&rqbd->rqbd_list); - cfs_list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds); + svcpt->scp_nrqbds_posted--; + cfs_list_del(&rqbd->rqbd_list); + cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle); - /* Don't complain if no request buffers are posted right now; LNET - * won't drop requests because we set the portal lazy! */ + /* Don't complain if no request buffers are posted right now; LNET + * won't drop requests because we set the portal lazy! */ - cfs_spin_unlock(&svc->srv_lock); + cfs_spin_unlock(&svcpt->scp_lock); - return (-1); + return -1; } static void ptlrpc_at_timer(unsigned long castmeharder) { - struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder; - svc->srv_at_check = 1; - svc->srv_at_checktime = cfs_time_current(); - cfs_waitq_signal(&svc->srv_waitq); + struct ptlrpc_service_part *svcpt; + + svcpt = (struct ptlrpc_service_part *)castmeharder; + + svcpt->scp_at_check = 1; + svcpt->scp_at_checktime = cfs_time_current(); + cfs_waitq_signal(&svcpt->scp_waitq); } static void @@ -493,6 +498,97 @@ ptlrpc_server_nthreads_check(struct ptlrpc_service_conf *conf, } /** + * Initialize percpt data for a service + */ +static int +ptlrpc_service_part_init(struct ptlrpc_service *svc, + struct ptlrpc_service_part *svcpt) +{ + struct ptlrpc_at_array *array; + int size; + int index; + int rc; + + CFS_INIT_LIST_HEAD(&svcpt->scp_threads); + + /* rqbd and incoming request queue */ + cfs_spin_lock_init(&svcpt->scp_lock); + CFS_INIT_LIST_HEAD(&svcpt->scp_rqbd_idle); + CFS_INIT_LIST_HEAD(&svcpt->scp_rqbd_posted); + CFS_INIT_LIST_HEAD(&svcpt->scp_req_incoming); + cfs_waitq_init(&svcpt->scp_waitq); + /* history request & rqbd list */ + CFS_INIT_LIST_HEAD(&svcpt->scp_hist_reqs); + CFS_INIT_LIST_HEAD(&svcpt->scp_hist_rqbds); + + /* acitve requests and hp requests */ + cfs_spin_lock_init(&svcpt->scp_req_lock); + CFS_INIT_LIST_HEAD(&svcpt->scp_req_pending); + CFS_INIT_LIST_HEAD(&svcpt->scp_hreq_pending); + + /* reply states */ + cfs_spin_lock_init(&svcpt->scp_rep_lock); + CFS_INIT_LIST_HEAD(&svcpt->scp_rep_active); +#ifndef __KERNEL__ + CFS_INIT_LIST_HEAD(&svcpt->scp_rep_queue); +#endif + CFS_INIT_LIST_HEAD(&svcpt->scp_rep_idle); + cfs_waitq_init(&svcpt->scp_rep_waitq); + cfs_atomic_set(&svcpt->scp_nreps_difficult, 0); + + /* adaptive timeout */ + cfs_spin_lock_init(&svcpt->scp_at_lock); + array = &svcpt->scp_at_array; + + size = at_est2timeout(at_max); + array->paa_size = size; + array->paa_count = 0; + array->paa_deadline = -1; + + /* allocate memory for scp_at_array (ptlrpc_at_array) */ + OBD_ALLOC(array->paa_reqs_array, sizeof(cfs_list_t) * size); + if (array->paa_reqs_array == NULL) + return -ENOMEM; + + for (index = 0; index < size; index++) + CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]); + + OBD_ALLOC(array->paa_reqs_count, sizeof(__u32) * size); + if (array->paa_reqs_count == NULL) + goto failed; + + cfs_timer_init(&svcpt->scp_at_timer, ptlrpc_at_timer, svcpt); + /* At SOW, service time should be quick; 10s seems generous. If client + * timeout is less than this, we'll be sending an early reply. */ + at_init(&svcpt->scp_at_estimate, 10, 0); + + /* assign this before call ptlrpc_grow_req_bufs */ + svcpt->scp_service = svc; + /* Now allocate the request buffers, but don't post them now */ + rc = ptlrpc_grow_req_bufs(svcpt); + /* We shouldn't be under memory pressure at startup, so + * fail if we can't allocate all our buffers at this time. */ + if (rc != 0) + goto failed; + + return 0; + + failed: + if (array->paa_reqs_count != NULL) { + OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size); + array->paa_reqs_count = NULL; + } + + if (array->paa_reqs_array != NULL) { + OBD_FREE(array->paa_reqs_array, + sizeof(cfs_list_t) * array->paa_size); + array->paa_reqs_array = NULL; + } + + return -ENOMEM; +} + +/** * Initialize service on a given portal. * This includes starting serving threads , allocating and posting rqbds and * so on. @@ -501,11 +597,8 @@ struct ptlrpc_service * ptlrpc_register_service(struct ptlrpc_service_conf *conf, cfs_proc_dir_entry_t *proc_entry) { - struct ptlrpc_service *service; - struct ptlrpc_at_array *array; - unsigned int index; - unsigned int size; - int rc; + struct ptlrpc_service *service; + int rc; ENTRY; LASSERT(conf->psc_buf.bc_nbufs > 0); @@ -517,16 +610,13 @@ ptlrpc_register_service(struct ptlrpc_service_conf *conf, if (service == NULL) RETURN(ERR_PTR(-ENOMEM)); - /* First initialise enough for early teardown */ - - cfs_spin_lock_init(&service->srv_lock); - cfs_spin_lock_init(&service->srv_rq_lock); - cfs_spin_lock_init(&service->srv_rs_lock); - CFS_INIT_LIST_HEAD(&service->srv_threads); - cfs_waitq_init(&service->srv_waitq); - + /* public members */ + cfs_spin_lock_init(&service->srv_lock); service->srv_name = conf->psc_name; service->srv_watchdog_factor = conf->psc_watchdog_factor; + CFS_INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */ + + /* buffer configuration */ service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : conf->psc_buf.bc_nbufs; service->srv_max_req_size = conf->psc_buf.bc_req_max_size + @@ -534,8 +624,12 @@ ptlrpc_register_service(struct ptlrpc_service_conf *conf, service->srv_buf_size = conf->psc_buf.bc_buf_size; service->srv_rep_portal = conf->psc_buf.bc_rep_portal; service->srv_req_portal = conf->psc_buf.bc_req_portal; - service->srv_request_seq = 1; /* valid seq #s start at 1 */ - service->srv_request_max_cull_seq = 0; + + /* Increase max reply size to next power of two */ + service->srv_max_reply_size = 1; + while (service->srv_max_reply_size < + conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD) + service->srv_max_reply_size <<= 1; ptlrpc_server_nthreads_check(conf, &service->srv_threads_min, &service->srv_threads_max); @@ -544,71 +638,23 @@ ptlrpc_register_service(struct ptlrpc_service_conf *conf, service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags; service->srv_cpu_affinity = !!conf->psc_thr.tc_cpu_affinity; service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO; - service->srv_hpreq_count = 0; - service->srv_n_active_hpreq = 0; service->srv_ops = conf->psc_ops; - rc = LNetSetLazyPortal(service->srv_req_portal); - LASSERT (rc == 0); - - CFS_INIT_LIST_HEAD(&service->srv_request_queue); - CFS_INIT_LIST_HEAD(&service->srv_request_hpq); - CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds); - CFS_INIT_LIST_HEAD(&service->srv_active_rqbds); - CFS_INIT_LIST_HEAD(&service->srv_history_rqbds); - CFS_INIT_LIST_HEAD(&service->srv_request_history); - CFS_INIT_LIST_HEAD(&service->srv_active_replies); -#ifndef __KERNEL__ - CFS_INIT_LIST_HEAD(&service->srv_reply_queue); -#endif - CFS_INIT_LIST_HEAD(&service->srv_free_rs_list); - cfs_waitq_init(&service->srv_free_rs_waitq); - cfs_atomic_set(&service->srv_n_difficult_replies, 0); - - cfs_spin_lock_init(&service->srv_at_lock); - CFS_INIT_LIST_HEAD(&service->srv_req_in_queue); - - array = &service->srv_at_array; - size = at_est2timeout(at_max); - array->paa_size = size; - array->paa_count = 0; - array->paa_deadline = -1; - - /* allocate memory for srv_at_array (ptlrpc_at_array) */ - OBD_ALLOC(array->paa_reqs_array, sizeof(cfs_list_t) * size); - if (array->paa_reqs_array == NULL) + OBD_ALLOC_PTR(service->srv_part); + if (service->srv_part == NULL) GOTO(failed, rc = -ENOMEM); - for (index = 0; index < size; index++) - CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]); - - OBD_ALLOC(array->paa_reqs_count, sizeof(__u32) * size); - if (array->paa_reqs_count == NULL) - GOTO(failed, rc = -ENOMEM); + rc = ptlrpc_service_part_init(service, service->srv_part); + if (rc != 0) + GOTO(failed, rc); - cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service); - /* At SOW, service time should be quick; 10s seems generous. If client - timeout is less than this, we'll be sending an early reply. */ - at_init(&service->srv_at_estimate, 10, 0); + rc = LNetSetLazyPortal(service->srv_req_portal); + LASSERT(rc == 0); cfs_spin_lock (&ptlrpc_all_services_lock); cfs_list_add (&service->srv_list, &ptlrpc_all_services); cfs_spin_unlock (&ptlrpc_all_services_lock); - /* Now allocate the request buffers */ - rc = ptlrpc_grow_req_bufs(service); - /* We shouldn't be under memory pressure at startup, so - * fail if we can't post all our buffers at this time. */ - if (rc != 0) - GOTO(failed, rc = -ENOMEM); - - /* Now allocate pool of reply buffers */ - /* Increase max reply size to next power of two */ - service->srv_max_reply_size = 1; - while (service->srv_max_reply_size < - conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD) - service->srv_max_reply_size <<= 1; - if (proc_entry != NULL) ptlrpc_lprocfs_register_service(proc_entry, service); @@ -659,8 +705,9 @@ static void ptlrpc_server_free_request(struct ptlrpc_request *req) */ void ptlrpc_server_drop_request(struct ptlrpc_request *req) { - struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd; - struct ptlrpc_service *svc = rqbd->rqbd_service; + struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd; + struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt; + struct ptlrpc_service *svc = svcpt->scp_service; int refcount; cfs_list_t *tmp; cfs_list_t *nxt; @@ -668,9 +715,9 @@ void ptlrpc_server_drop_request(struct ptlrpc_request *req) if (!cfs_atomic_dec_and_test(&req->rq_refcount)) return; - cfs_spin_lock(&svc->srv_at_lock); - if (req->rq_at_linked) { - struct ptlrpc_at_array *array = &svc->srv_at_array; + cfs_spin_lock(&svcpt->scp_at_lock); + if (req->rq_at_linked) { + struct ptlrpc_at_array *array = &svcpt->scp_at_array; __u32 index = req->rq_at_index; LASSERT(!cfs_list_empty(&req->rq_timed_list)); @@ -682,7 +729,8 @@ void ptlrpc_server_drop_request(struct ptlrpc_request *req) array->paa_count--; } else LASSERT(cfs_list_empty(&req->rq_timed_list)); - cfs_spin_unlock(&svc->srv_at_lock); + + cfs_spin_unlock(&svcpt->scp_at_lock); /* finalize request */ if (req->rq_export) { @@ -690,7 +738,7 @@ void ptlrpc_server_drop_request(struct ptlrpc_request *req) req->rq_export = NULL; } - cfs_spin_lock(&svc->srv_lock); + cfs_spin_lock(&svcpt->scp_lock); cfs_list_add(&req->rq_list, &rqbd->rqbd_reqs); @@ -698,18 +746,19 @@ void ptlrpc_server_drop_request(struct ptlrpc_request *req) if (refcount == 0) { /* request buffer is now idle: add to history */ cfs_list_del(&rqbd->rqbd_list); - cfs_list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds); - svc->srv_n_history_rqbds++; - /* cull some history? - * I expect only about 1 or 2 rqbds need to be recycled here */ - while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) { - rqbd = cfs_list_entry(svc->srv_history_rqbds.next, - struct ptlrpc_request_buffer_desc, - rqbd_list); + cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds); + svcpt->scp_hist_nrqbds++; + + /* cull some history? + * I expect only about 1 or 2 rqbds need to be recycled here */ + while (svcpt->scp_hist_nrqbds > svc->srv_max_history_rqbds) { + rqbd = cfs_list_entry(svcpt->scp_hist_rqbds.next, + struct ptlrpc_request_buffer_desc, + rqbd_list); - cfs_list_del(&rqbd->rqbd_list); - svc->srv_n_history_rqbds--; + cfs_list_del(&rqbd->rqbd_list); + svcpt->scp_hist_nrqbds--; /* remove rqbd's reqs from svc's req history while * I've got the service lock */ @@ -717,14 +766,15 @@ void ptlrpc_server_drop_request(struct ptlrpc_request *req) req = cfs_list_entry(tmp, struct ptlrpc_request, rq_list); /* Track the highest culled req seq */ - if (req->rq_history_seq > - svc->srv_request_max_cull_seq) - svc->srv_request_max_cull_seq = - req->rq_history_seq; - cfs_list_del(&req->rq_history_list); - } + if (req->rq_history_seq > + svcpt->scp_hist_seq_culled) { + svcpt->scp_hist_seq_culled = + req->rq_history_seq; + } + cfs_list_del(&req->rq_history_list); + } - cfs_spin_unlock(&svc->srv_lock); + cfs_spin_unlock(&svcpt->scp_lock); cfs_list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) { req = cfs_list_entry(rqbd->rqbd_reqs.next, @@ -734,46 +784,47 @@ void ptlrpc_server_drop_request(struct ptlrpc_request *req) ptlrpc_server_free_request(req); } - cfs_spin_lock(&svc->srv_lock); - /* - * now all reqs including the embedded req has been - * disposed, schedule request buffer for re-use. - */ - LASSERT(cfs_atomic_read(&rqbd->rqbd_req.rq_refcount) == - 0); - cfs_list_add_tail(&rqbd->rqbd_list, - &svc->srv_idle_rqbds); - } - - cfs_spin_unlock(&svc->srv_lock); - } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) { - /* If we are low on memory, we are not interested in history */ - cfs_list_del(&req->rq_list); - cfs_list_del_init(&req->rq_history_list); - cfs_spin_unlock(&svc->srv_lock); - - ptlrpc_server_free_request(req); - } else { - cfs_spin_unlock(&svc->srv_lock); - } + cfs_spin_lock(&svcpt->scp_lock); + /* + * now all reqs including the embedded req has been + * disposed, schedule request buffer for re-use. + */ + LASSERT(cfs_atomic_read(&rqbd->rqbd_req.rq_refcount) == + 0); + cfs_list_add_tail(&rqbd->rqbd_list, + &svcpt->scp_rqbd_idle); + } + + cfs_spin_unlock(&svcpt->scp_lock); + } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) { + /* If we are low on memory, we are not interested in history */ + cfs_list_del(&req->rq_list); + cfs_list_del_init(&req->rq_history_list); + + cfs_spin_unlock(&svcpt->scp_lock); + + ptlrpc_server_free_request(req); + } else { + cfs_spin_unlock(&svcpt->scp_lock); + } } /** * to finish a request: stop sending more early replies, and release * the request. should be called after we finished handling the request. */ -static void ptlrpc_server_finish_request(struct ptlrpc_service *svc, - struct ptlrpc_request *req) +static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt, + struct ptlrpc_request *req) { - ptlrpc_hpreq_fini(req); + ptlrpc_hpreq_fini(req); - cfs_spin_lock(&svc->srv_rq_lock); - svc->srv_n_active_reqs--; - if (req->rq_hp) - svc->srv_n_active_hpreq--; - cfs_spin_unlock(&svc->srv_rq_lock); + cfs_spin_lock(&svcpt->scp_req_lock); + svcpt->scp_nreqs_active--; + if (req->rq_hp) + svcpt->scp_nhreqs_active--; + cfs_spin_unlock(&svcpt->scp_req_lock); - ptlrpc_server_drop_request(req); + ptlrpc_server_drop_request(req); } /** @@ -906,35 +957,37 @@ static int ptlrpc_check_req(struct ptlrpc_request *req) return rc; } -static void ptlrpc_at_set_timer(struct ptlrpc_service *svc) +static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt) { - struct ptlrpc_at_array *array = &svc->srv_at_array; - __s32 next; + struct ptlrpc_at_array *array = &svcpt->scp_at_array; + __s32 next; - cfs_spin_lock(&svc->srv_at_lock); - if (array->paa_count == 0) { - cfs_timer_disarm(&svc->srv_at_timer); - cfs_spin_unlock(&svc->srv_at_lock); - return; - } + cfs_spin_lock(&svcpt->scp_at_lock); + if (array->paa_count == 0) { + cfs_timer_disarm(&svcpt->scp_at_timer); + cfs_spin_unlock(&svcpt->scp_at_lock); + return; + } - /* Set timer for closest deadline */ - next = (__s32)(array->paa_deadline - cfs_time_current_sec() - - at_early_margin); - if (next <= 0) - ptlrpc_at_timer((unsigned long)svc); - else - cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next)); - cfs_spin_unlock(&svc->srv_at_lock); - CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next); + /* Set timer for closest deadline */ + next = (__s32)(array->paa_deadline - cfs_time_current_sec() - + at_early_margin); + if (next <= 0) { + ptlrpc_at_timer((unsigned long)svcpt); + } else { + cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next)); + CDEBUG(D_INFO, "armed %s at %+ds\n", + svcpt->scp_service->srv_name, next); + } + cfs_spin_unlock(&svcpt->scp_at_lock); } /* Add rpc to early reply check list */ static int ptlrpc_at_add_timed(struct ptlrpc_request *req) { - struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service; + struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt; + struct ptlrpc_at_array *array = &svcpt->scp_at_array; struct ptlrpc_request *rq = NULL; - struct ptlrpc_at_array *array = &svc->srv_at_array; __u32 index; int found = 0; @@ -947,7 +1000,7 @@ static int ptlrpc_at_add_timed(struct ptlrpc_request *req) if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0) return(-ENOSYS); - cfs_spin_lock(&svc->srv_at_lock); + cfs_spin_lock(&svcpt->scp_at_lock); LASSERT(cfs_list_empty(&req->rq_timed_list)); index = (unsigned long)req->rq_deadline % array->paa_size; @@ -980,17 +1033,17 @@ static int ptlrpc_at_add_timed(struct ptlrpc_request *req) array->paa_deadline = req->rq_deadline; found = 1; } - cfs_spin_unlock(&svc->srv_at_lock); + cfs_spin_unlock(&svcpt->scp_at_lock); - if (found) - ptlrpc_at_set_timer(svc); + if (found) + ptlrpc_at_set_timer(svcpt); - return 0; + return 0; } static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req) { - struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service; + struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt; struct ptlrpc_request *reqcopy; struct lustre_msg *reqmsg; cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec(); @@ -1003,8 +1056,8 @@ static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req) DEBUG_REQ(D_ADAPTTO, req, "%ssending early reply (deadline %+lds, margin %+lds) for " "%d+%d", AT_OFF ? "AT off - not " : "", - olddl, olddl - at_get(&svc->srv_at_estimate), - at_get(&svc->srv_at_estimate), at_extra); + olddl, olddl - at_get(&svcpt->scp_at_estimate), + at_get(&svcpt->scp_at_estimate), at_extra); if (AT_OFF) RETURN(0); @@ -1033,28 +1086,28 @@ static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req) * during the recovery period send at least 4 early replies, * spacing them every at_extra if we can. at_estimate should * always equal this fixed value during recovery. */ - at_measured(&svc->srv_at_estimate, min(at_extra, - req->rq_export->exp_obd->obd_recovery_timeout / 4)); - } else { - /* Fake our processing time into the future to ask the clients - * for some extra amount of time */ - at_measured(&svc->srv_at_estimate, at_extra + - cfs_time_current_sec() - - req->rq_arrival_time.tv_sec); - - /* Check to see if we've actually increased the deadline - - * we may be past adaptive_max */ - if (req->rq_deadline >= req->rq_arrival_time.tv_sec + - at_get(&svc->srv_at_estimate)) { - DEBUG_REQ(D_WARNING, req, "Couldn't add any time " - "(%ld/%ld), not sending early reply\n", - olddl, req->rq_arrival_time.tv_sec + - at_get(&svc->srv_at_estimate) - - cfs_time_current_sec()); - RETURN(-ETIMEDOUT); - } - } - newdl = cfs_time_current_sec() + at_get(&svc->srv_at_estimate); + at_measured(&svcpt->scp_at_estimate, min(at_extra, + req->rq_export->exp_obd->obd_recovery_timeout / 4)); + } else { + /* Fake our processing time into the future to ask the clients + * for some extra amount of time */ + at_measured(&svcpt->scp_at_estimate, at_extra + + cfs_time_current_sec() - + req->rq_arrival_time.tv_sec); + + /* Check to see if we've actually increased the deadline - + * we may be past adaptive_max */ + if (req->rq_deadline >= req->rq_arrival_time.tv_sec + + at_get(&svcpt->scp_at_estimate)) { + DEBUG_REQ(D_WARNING, req, "Couldn't add any time " + "(%ld/%ld), not sending early reply\n", + olddl, req->rq_arrival_time.tv_sec + + at_get(&svcpt->scp_at_estimate) - + cfs_time_current_sec()); + RETURN(-ETIMEDOUT); + } + } + newdl = cfs_time_current_sec() + at_get(&svcpt->scp_at_estimate); OBD_ALLOC(reqcopy, sizeof *reqcopy); if (reqcopy == NULL) @@ -1126,11 +1179,11 @@ out: /* Send early replies to everybody expiring within at_early_margin asking for at_extra time */ -static int ptlrpc_at_check_timed(struct ptlrpc_service *svc) +static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt) { + struct ptlrpc_at_array *array = &svcpt->scp_at_array; struct ptlrpc_request *rq, *n; cfs_list_t work_list; - struct ptlrpc_at_array *array = &svc->srv_at_array; __u32 index, count; time_t deadline; time_t now = cfs_time_current_sec(); @@ -1138,25 +1191,25 @@ static int ptlrpc_at_check_timed(struct ptlrpc_service *svc) int first, counter = 0; ENTRY; - cfs_spin_lock(&svc->srv_at_lock); - if (svc->srv_at_check == 0) { - cfs_spin_unlock(&svc->srv_at_lock); - RETURN(0); - } - delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime); - svc->srv_at_check = 0; + cfs_spin_lock(&svcpt->scp_at_lock); + if (svcpt->scp_at_check == 0) { + cfs_spin_unlock(&svcpt->scp_at_lock); + RETURN(0); + } + delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime); + svcpt->scp_at_check = 0; - if (array->paa_count == 0) { - cfs_spin_unlock(&svc->srv_at_lock); - RETURN(0); - } + if (array->paa_count == 0) { + cfs_spin_unlock(&svcpt->scp_at_lock); + RETURN(0); + } - /* The timer went off, but maybe the nearest rpc already completed. */ - first = array->paa_deadline - now; - if (first > at_early_margin) { - /* We've still got plenty of time. Reset the timer. */ - cfs_spin_unlock(&svc->srv_at_lock); - ptlrpc_at_set_timer(svc); + /* The timer went off, but maybe the nearest rpc already completed. */ + first = array->paa_deadline - now; + if (first > at_early_margin) { + /* We've still got plenty of time. Reset the timer. */ + cfs_spin_unlock(&svcpt->scp_at_lock); + ptlrpc_at_set_timer(svcpt); RETURN(0); } @@ -1200,10 +1253,10 @@ static int ptlrpc_at_check_timed(struct ptlrpc_service *svc) index = 0; } array->paa_deadline = deadline; - cfs_spin_unlock(&svc->srv_at_lock); + cfs_spin_unlock(&svcpt->scp_at_lock); - /* we have a new earliest deadline, restart the timer */ - ptlrpc_at_set_timer(svc); + /* we have a new earliest deadline, restart the timer */ + ptlrpc_at_set_timer(svcpt); CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early " "replies\n", first, at_extra, counter); @@ -1211,11 +1264,13 @@ static int ptlrpc_at_check_timed(struct ptlrpc_service *svc) /* We're already past request deadlines before we even get a chance to send early replies */ LCONSOLE_WARN("%s: This server is not able to keep up with " - "request traffic (cpu-bound).\n", svc->srv_name); - CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, " - "delay="CFS_DURATION_T"(jiff)\n", - counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs, - at_get(&svc->srv_at_estimate), delay); + "request traffic (cpu-bound).\n", + svcpt->scp_service->srv_name); + CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, " + "delay="CFS_DURATION_T"(jiff)\n", + counter, svcpt->scp_nreqs_incoming, + svcpt->scp_nreqs_active, + at_get(&svcpt->scp_at_estimate), delay); } /* we took additional refcount so entries can't be deleted from list, no @@ -1231,7 +1286,7 @@ static int ptlrpc_at_check_timed(struct ptlrpc_service *svc) ptlrpc_server_drop_request(rq); } - RETURN(0); + RETURN(1); /* return "did_something" for liblustre */ } /** @@ -1286,23 +1341,23 @@ static void ptlrpc_hpreq_fini(struct ptlrpc_request *req) * Make the request a high priority one. * * All the high priority requests are queued in a separate FIFO - * ptlrpc_service::srv_request_hpq list which is parallel to - * ptlrpc_service::srv_request_queue list but has a higher priority + * ptlrpc_service_part::scp_hpreq_pending list which is parallel to + * ptlrpc_service_part::scp_req_pending list but has a higher priority * for handling. * * \see ptlrpc_server_handle_request(). */ -static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc, +static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service_part *svcpt, struct ptlrpc_request *req) { ENTRY; - LASSERT(svc != NULL); + cfs_spin_lock(&req->rq_lock); if (req->rq_hp == 0) { int opc = lustre_msg_get_opc(req->rq_reqmsg); /* Add to the high priority queue. */ - cfs_list_move_tail(&req->rq_list, &svc->srv_request_hpq); + cfs_list_move_tail(&req->rq_list, &svcpt->scp_hreq_pending); req->rq_hp = 1; if (opc != OBD_PING) DEBUG_REQ(D_RPCTRACE, req, "high priority req"); @@ -1316,18 +1371,18 @@ static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc, */ void ptlrpc_hpreq_reorder(struct ptlrpc_request *req) { - struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service; - ENTRY; + struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt; + ENTRY; - cfs_spin_lock(&svc->srv_rq_lock); - /* It may happen that the request is already taken for the processing - * but still in the export list, or the request is not in the request - * queue but in the export list already, do not add it into the - * HP list. */ - if (!cfs_list_empty(&req->rq_list)) - ptlrpc_hpreq_reorder_nolock(svc, req); - cfs_spin_unlock(&svc->srv_rq_lock); - EXIT; + cfs_spin_lock(&svcpt->scp_req_lock); + /* It may happen that the request is already taken for the processing + * but still in the export list, or the request is not in the request + * queue but in the export list already, do not add it into the + * HP list. */ + if (!cfs_list_empty(&req->rq_list)) + ptlrpc_hpreq_reorder_nolock(svcpt, req); + cfs_spin_unlock(&svcpt->scp_req_lock); + EXIT; } /** Check if the request is a high priority one. */ @@ -1344,50 +1399,51 @@ static int ptlrpc_server_hpreq_check(struct ptlrpc_service *svc, } /** Check if a request is a high priority one. */ -static int ptlrpc_server_request_add(struct ptlrpc_service *svc, - struct ptlrpc_request *req) +static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt, + struct ptlrpc_request *req) { - int rc; - ENTRY; + int rc; + ENTRY; - rc = ptlrpc_server_hpreq_check(svc, req); - if (rc < 0) - RETURN(rc); + rc = ptlrpc_server_hpreq_check(svcpt->scp_service, req); + if (rc < 0) + RETURN(rc); - cfs_spin_lock(&svc->srv_rq_lock); + cfs_spin_lock(&svcpt->scp_req_lock); - if (rc) - ptlrpc_hpreq_reorder_nolock(svc, req); - else - cfs_list_add_tail(&req->rq_list, - &svc->srv_request_queue); + if (rc) + ptlrpc_hpreq_reorder_nolock(svcpt, req); + else + cfs_list_add_tail(&req->rq_list, &svcpt->scp_req_pending); - cfs_spin_unlock(&svc->srv_rq_lock); + cfs_spin_unlock(&svcpt->scp_req_lock); - RETURN(0); + RETURN(0); } /** * Allow to handle high priority request - * User can call it w/o any lock but need to hold ptlrpc_service::srv_rq_lock - * to get reliable result + * User can call it w/o any lock but need to hold + * ptlrpc_service_part::scp_req_lock to get reliable result */ -static int ptlrpc_server_allow_high(struct ptlrpc_service *svc, int force) +static int ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt, + int force) { - if (force) - return 1; + if (force) + return 1; - if (svc->srv_n_active_reqs >= svc->srv_threads_running - 1) - return 0; + if (svcpt->scp_nreqs_active >= svcpt->scp_nthrs_running - 1) + return 0; - return cfs_list_empty(&svc->srv_request_queue) || - svc->srv_hpreq_count < svc->srv_hpreq_ratio; + return cfs_list_empty(&svcpt->scp_req_pending) || + svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio; } -static int ptlrpc_server_high_pending(struct ptlrpc_service *svc, int force) +static int ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt, + int force) { - return ptlrpc_server_allow_high(svc, force) && - !cfs_list_empty(&svc->srv_request_hpq); + return ptlrpc_server_allow_high(svcpt, force) && + !cfs_list_empty(&svcpt->scp_hreq_pending); } /** @@ -1396,45 +1452,47 @@ static int ptlrpc_server_high_pending(struct ptlrpc_service *svc, int force) * already being processed (i.e. those threads can service more high-priority * requests), or if there are enough idle threads that a later thread can do * a high priority request. - * User can call it w/o any lock but need to hold ptlrpc_service::srv_rq_lock - * to get reliable result + * User can call it w/o any lock but need to hold + * ptlrpc_service_part::scp_req_lock to get reliable result */ -static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force) +static int ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt, + int force) { #ifndef __KERNEL__ - if (1) /* always allow to handle normal request for liblustre */ - return 1; + if (1) /* always allow to handle normal request for liblustre */ + return 1; #endif - if (force || - svc->srv_n_active_reqs < svc->srv_threads_running - 2) - return 1; + if (force || + svcpt->scp_nreqs_active < svcpt->scp_nthrs_running - 2) + return 1; - if (svc->srv_n_active_reqs >= svc->srv_threads_running - 1) - return 0; + if (svcpt->scp_nreqs_active >= svcpt->scp_nthrs_running - 1) + return 0; - return svc->srv_n_active_hpreq > 0 || - svc->srv_ops.so_hpreq_handler == NULL; + return svcpt->scp_nhreqs_active > 0 || + svcpt->scp_service->srv_ops.so_hpreq_handler == NULL; } -static int ptlrpc_server_normal_pending(struct ptlrpc_service *svc, int force) +static int ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt, + int force) { - return ptlrpc_server_allow_normal(svc, force) && - !cfs_list_empty(&svc->srv_request_queue); + return ptlrpc_server_allow_normal(svcpt, force) && + !cfs_list_empty(&svcpt->scp_req_pending); } /** * Returns true if there are requests available in incoming * request queue for processing and it is allowed to fetch them. - * User can call it w/o any lock but need to hold ptlrpc_service::srv_rq_lock + * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock * to get reliable result * \see ptlrpc_server_allow_normal * \see ptlrpc_server_allow high */ static inline int -ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force) +ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, int force) { - return ptlrpc_server_high_pending(svc, force) || - ptlrpc_server_normal_pending(svc, force); + return ptlrpc_server_high_pending(svcpt, force) || + ptlrpc_server_normal_pending(svcpt, force); } /** @@ -1443,26 +1501,25 @@ ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force) * Returns a pointer to fetched request. */ static struct ptlrpc_request * -ptlrpc_server_request_get(struct ptlrpc_service *svc, int force) +ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, int force) { - struct ptlrpc_request *req; - ENTRY; - - if (ptlrpc_server_high_pending(svc, force)) { - req = cfs_list_entry(svc->srv_request_hpq.next, - struct ptlrpc_request, rq_list); - svc->srv_hpreq_count++; - RETURN(req); + struct ptlrpc_request *req; + ENTRY; - } + if (ptlrpc_server_high_pending(svcpt, force)) { + req = cfs_list_entry(svcpt->scp_hreq_pending.next, + struct ptlrpc_request, rq_list); + svcpt->scp_hreq_count++; + RETURN(req); + } - if (ptlrpc_server_normal_pending(svc, force)) { - req = cfs_list_entry(svc->srv_request_queue.next, - struct ptlrpc_request, rq_list); - svc->srv_hpreq_count = 0; - RETURN(req); - } - RETURN(NULL); + if (ptlrpc_server_normal_pending(svcpt, force)) { + req = cfs_list_entry(svcpt->scp_req_pending.next, + struct ptlrpc_request, rq_list); + svcpt->scp_hreq_count = 0; + RETURN(req); + } + RETURN(NULL); } /** @@ -1472,28 +1529,27 @@ ptlrpc_server_request_get(struct ptlrpc_service *svc, int force) * ptlrpc_server_handle_req later on. */ static int -ptlrpc_server_handle_req_in(struct ptlrpc_service *svc) +ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt) { - struct ptlrpc_request *req; - __u32 deadline; - int rc; - ENTRY; - - LASSERT(svc); + struct ptlrpc_service *svc = svcpt->scp_service; + struct ptlrpc_request *req; + __u32 deadline; + int rc; + ENTRY; - cfs_spin_lock(&svc->srv_lock); - if (cfs_list_empty(&svc->srv_req_in_queue)) { - cfs_spin_unlock(&svc->srv_lock); - RETURN(0); - } + cfs_spin_lock(&svcpt->scp_lock); + if (cfs_list_empty(&svcpt->scp_req_incoming)) { + cfs_spin_unlock(&svcpt->scp_lock); + RETURN(0); + } - req = cfs_list_entry(svc->srv_req_in_queue.next, - struct ptlrpc_request, rq_list); - cfs_list_del_init (&req->rq_list); - svc->srv_n_queued_reqs--; - /* Consider this still a "queued" request as far as stats are - concerned */ - cfs_spin_unlock(&svc->srv_lock); + req = cfs_list_entry(svcpt->scp_req_incoming.next, + struct ptlrpc_request, rq_list); + cfs_list_del_init(&req->rq_list); + svcpt->scp_nreqs_incoming--; + /* Consider this still a "queued" request as far as stats are + * concerned */ + cfs_spin_unlock(&svcpt->scp_lock); /* go through security check/transform */ rc = sptlrpc_svc_unwrap_request(req); @@ -1596,21 +1652,21 @@ ptlrpc_server_handle_req_in(struct ptlrpc_service *svc) ptlrpc_at_add_timed(req); /* Move it over to the request processing queue */ - rc = ptlrpc_server_request_add(svc, req); - if (rc) { - ptlrpc_hpreq_fini(req); - GOTO(err_req, rc); - } - cfs_waitq_signal(&svc->srv_waitq); - RETURN(1); + rc = ptlrpc_server_request_add(svcpt, req); + if (rc) { + ptlrpc_hpreq_fini(req); + GOTO(err_req, rc); + } + cfs_waitq_signal(&svcpt->scp_waitq); + RETURN(1); err_req: - cfs_spin_lock(&svc->srv_rq_lock); - svc->srv_n_active_reqs++; - cfs_spin_unlock(&svc->srv_rq_lock); - ptlrpc_server_finish_request(svc, req); + cfs_spin_lock(&svcpt->scp_req_lock); + svcpt->scp_nreqs_active++; + cfs_spin_unlock(&svcpt->scp_req_lock); + ptlrpc_server_finish_request(svcpt, req); - RETURN(1); + RETURN(1); } /** @@ -1618,9 +1674,10 @@ err_req: * Calls handler function from service to do actual processing. */ static int -ptlrpc_server_handle_request(struct ptlrpc_service *svc, - struct ptlrpc_thread *thread) +ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt, + struct ptlrpc_thread *thread) { + struct ptlrpc_service *svc = svcpt->scp_service; struct obd_export *export = NULL; struct ptlrpc_request *request; struct timeval work_start; @@ -1630,19 +1687,17 @@ ptlrpc_server_handle_request(struct ptlrpc_service *svc, int fail_opc = 0; ENTRY; - LASSERT(svc); - - cfs_spin_lock(&svc->srv_rq_lock); + cfs_spin_lock(&svcpt->scp_req_lock); #ifndef __KERNEL__ - /* !@%$# liblustre only has 1 thread */ - if (cfs_atomic_read(&svc->srv_n_difficult_replies) != 0) { - cfs_spin_unlock(&svc->srv_rq_lock); - RETURN(0); - } + /* !@%$# liblustre only has 1 thread */ + if (cfs_atomic_read(&svcpt->scp_nreps_difficult) != 0) { + cfs_spin_unlock(&svcpt->scp_req_lock); + RETURN(0); + } #endif - request = ptlrpc_server_request_get(svc, 0); - if (request == NULL) { - cfs_spin_unlock(&svc->srv_rq_lock); + request = ptlrpc_server_request_get(svcpt, 0); + if (request == NULL) { + cfs_spin_unlock(&svcpt->scp_req_lock); RETURN(0); } @@ -1653,23 +1708,25 @@ ptlrpc_server_handle_request(struct ptlrpc_service *svc, if (unlikely(fail_opc)) { if (request->rq_export && request->rq_ops) { - cfs_spin_unlock(&svc->srv_rq_lock); - OBD_FAIL_TIMEOUT(fail_opc, 4); - cfs_spin_lock(&svc->srv_rq_lock); - request = ptlrpc_server_request_get(svc, 0); - if (request == NULL) { - cfs_spin_unlock(&svc->srv_rq_lock); - RETURN(0); - } - } - } + cfs_spin_unlock(&svcpt->scp_req_lock); + + OBD_FAIL_TIMEOUT(fail_opc, 4); - cfs_list_del_init(&request->rq_list); - svc->srv_n_active_reqs++; - if (request->rq_hp) - svc->srv_n_active_hpreq++; + cfs_spin_lock(&svcpt->scp_req_lock); + request = ptlrpc_server_request_get(svcpt, 0); + if (request == NULL) { + cfs_spin_unlock(&svcpt->scp_req_lock); + RETURN(0); + } + } + } + + cfs_list_del_init(&request->rq_list); + svcpt->scp_nreqs_active++; + if (request->rq_hp) + svcpt->scp_nhreqs_active++; - cfs_spin_unlock(&svc->srv_rq_lock); + cfs_spin_unlock(&svcpt->scp_req_lock); ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET); @@ -1682,11 +1739,11 @@ ptlrpc_server_handle_request(struct ptlrpc_service *svc, lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR, timediff); lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR, - svc->srv_n_queued_reqs); - lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR, - svc->srv_n_active_reqs); - lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT, - at_get(&svc->srv_at_estimate)); + svcpt->scp_nreqs_incoming); + lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR, + svcpt->scp_nreqs_active); + lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT, + at_get(&svcpt->scp_at_estimate)); } rc = lu_context_init(&request->rq_session, @@ -1801,18 +1858,19 @@ put_conn: } out_req: - ptlrpc_server_finish_request(svc, request); + ptlrpc_server_finish_request(svcpt, request); - RETURN(1); + RETURN(1); } /** * An internal function to process a single reply state object. */ static int -ptlrpc_handle_rs (struct ptlrpc_reply_state *rs) +ptlrpc_handle_rs(struct ptlrpc_reply_state *rs) { - struct ptlrpc_service *svc = rs->rs_service; + struct ptlrpc_service_part *svcpt = rs->rs_svcpt; + struct ptlrpc_service *svc = svcpt->scp_service; struct obd_export *exp; int nlocks; int been_handled; @@ -1900,15 +1958,15 @@ ptlrpc_handle_rs (struct ptlrpc_reply_state *rs) class_export_put (exp); rs->rs_export = NULL; ptlrpc_rs_decref (rs); - if (cfs_atomic_dec_and_test(&svc->srv_n_difficult_replies) && - svc->srv_is_stopping) - cfs_waitq_broadcast(&svc->srv_waitq); - RETURN(1); - } + if (cfs_atomic_dec_and_test(&svcpt->scp_nreps_difficult) && + svc->srv_is_stopping) + cfs_waitq_broadcast(&svcpt->scp_waitq); + RETURN(1); + } - /* still on the net; callback will schedule */ - cfs_spin_unlock(&rs->rs_lock); - RETURN(1); + /* still on the net; callback will schedule */ + cfs_spin_unlock(&rs->rs_lock); + RETURN(1); } #ifndef __KERNEL__ @@ -1922,22 +1980,22 @@ ptlrpc_handle_rs (struct ptlrpc_reply_state *rs) * \retval 1 one reply processed */ static int -ptlrpc_server_handle_reply(struct ptlrpc_service *svc) +ptlrpc_server_handle_reply(struct ptlrpc_service_part *svcpt) { - struct ptlrpc_reply_state *rs = NULL; - ENTRY; + struct ptlrpc_reply_state *rs = NULL; + ENTRY; - cfs_spin_lock(&svc->srv_rs_lock); - if (!cfs_list_empty(&svc->srv_reply_queue)) { - rs = cfs_list_entry(svc->srv_reply_queue.prev, - struct ptlrpc_reply_state, - rs_list); - cfs_list_del_init(&rs->rs_list); - } - cfs_spin_unlock(&svc->srv_rs_lock); - if (rs != NULL) - ptlrpc_handle_rs(rs); - RETURN(rs != NULL); + cfs_spin_lock(&svcpt->scp_rep_lock); + if (!cfs_list_empty(&svcpt->scp_rep_queue)) { + rs = cfs_list_entry(svcpt->scp_rep_queue.prev, + struct ptlrpc_reply_state, + rs_list); + cfs_list_del_init(&rs->rs_list); + } + cfs_spin_unlock(&svcpt->scp_rep_lock); + if (rs != NULL) + ptlrpc_handle_rs(rs); + RETURN(rs != NULL); } /* FIXME make use of timeout later */ @@ -1954,41 +2012,44 @@ liblustre_check_services (void *arg) cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) { struct ptlrpc_service *svc = cfs_list_entry (tmp, struct ptlrpc_service, srv_list); + struct ptlrpc_service_part *svcpt; - if (svc->srv_threads_running != 0) /* I've recursed */ - continue; + svcpt = svc->srv_part; - /* service threads can block for bulk, so this limits us - * (arbitrarily) to recursing 1 stack frame per service. - * Note that the problem with recursion is that we have to - * unwind completely before our caller can resume. */ + if (svcpt->scp_nthrs_running != 0) /* I've recursed */ + continue; - svc->srv_threads_running++; + /* service threads can block for bulk, so this limits us + * (arbitrarily) to recursing 1 stack frame per service. + * Note that the problem with recursion is that we have to + * unwind completely before our caller can resume. */ - do { - rc = ptlrpc_server_handle_req_in(svc); - rc |= ptlrpc_server_handle_reply(svc); - rc |= ptlrpc_at_check_timed(svc); - rc |= ptlrpc_server_handle_request(svc, NULL); - rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0); - did_something |= rc; - } while (rc); + svcpt->scp_nthrs_running++; - svc->srv_threads_running--; - } + do { + rc = ptlrpc_server_handle_req_in(svcpt); + rc |= ptlrpc_server_handle_reply(svcpt); + rc |= ptlrpc_at_check_timed(svcpt); + rc |= ptlrpc_server_handle_request(svcpt, NULL); + rc |= (ptlrpc_server_post_idle_rqbds(svcpt) > 0); + did_something |= rc; + } while (rc); - RETURN(did_something); + svcpt->scp_nthrs_running--; + } + + RETURN(did_something); } #define ptlrpc_stop_all_threads(s) do {} while (0) #else /* __KERNEL__ */ static void -ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc) +ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt) { - int avail = svc->srv_nrqbd_receiving; - int low_water = test_req_buffer_pressure ? 0 : - svc->srv_nbuf_per_group / 2; + int avail = svcpt->scp_nrqbds_posted; + int low_water = test_req_buffer_pressure ? 0 : + svcpt->scp_service->srv_nbuf_per_group / 2; /* NB I'm not locking; just looking. */ @@ -1998,107 +2059,109 @@ ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc) * space. */ if (avail <= low_water) - ptlrpc_grow_req_bufs(svc); + ptlrpc_grow_req_bufs(svcpt); - if (svc->srv_stats) - lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR, - avail); + if (svcpt->scp_service->srv_stats) { + lprocfs_counter_add(svcpt->scp_service->srv_stats, + PTLRPC_REQBUF_AVAIL_CNTR, avail); + } } static int ptlrpc_retry_rqbds(void *arg) { - struct ptlrpc_service *svc = (struct ptlrpc_service *)arg; + struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg; - svc->srv_rqbd_timeout = 0; - return (-ETIMEDOUT); + svcpt->scp_rqbd_timeout = 0; + return -ETIMEDOUT; } static inline int -ptlrpc_threads_enough(struct ptlrpc_service *svc) +ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt) { - return svc->srv_n_active_reqs < - svc->srv_threads_running - 1 - - (svc->srv_ops.so_hpreq_handler != NULL); + return svcpt->scp_nreqs_active < + svcpt->scp_nthrs_running - 1 - + (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL); } /** * allowed to create more threads - * user can call it w/o any lock but need to hold ptlrpc_service::srv_lock to - * get reliable result + * user can call it w/o any lock but need to hold + * ptlrpc_service_part::scp_lock to get reliable result */ static inline int -ptlrpc_threads_increasable(struct ptlrpc_service *svc) +ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt) { - return svc->srv_threads_running + - svc->srv_threads_starting < svc->srv_threads_max; + return svcpt->scp_nthrs_running + + svcpt->scp_nthrs_starting < svcpt->scp_service->srv_threads_max; } /** * too many requests and allowed to create more threads */ static inline int -ptlrpc_threads_need_create(struct ptlrpc_service *svc) +ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt) { - return !ptlrpc_threads_enough(svc) && ptlrpc_threads_increasable(svc); + return !ptlrpc_threads_enough(svcpt) && + ptlrpc_threads_increasable(svcpt); } static inline int ptlrpc_thread_stopping(struct ptlrpc_thread *thread) { - return thread_is_stopping(thread) || - thread->t_svc->srv_is_stopping; + return thread_is_stopping(thread) || + thread->t_svcpt->scp_service->srv_is_stopping; } static inline int -ptlrpc_rqbd_pending(struct ptlrpc_service *svc) +ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt) { - return !cfs_list_empty(&svc->srv_idle_rqbds) && - svc->srv_rqbd_timeout == 0; + return !cfs_list_empty(&svcpt->scp_rqbd_idle) && + svcpt->scp_rqbd_timeout == 0; } static inline int -ptlrpc_at_check(struct ptlrpc_service *svc) +ptlrpc_at_check(struct ptlrpc_service_part *svcpt) { - return svc->srv_at_check; + return svcpt->scp_at_check; } /** * requests wait on preprocessing - * user can call it w/o any lock but need to hold ptlrpc_service::srv_lock to - * get reliable result + * user can call it w/o any lock but need to hold + * ptlrpc_service_part::scp_lock to get reliable result */ static inline int -ptlrpc_server_request_waiting(struct ptlrpc_service *svc) +ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt) { - return !cfs_list_empty(&svc->srv_req_in_queue); + return !cfs_list_empty(&svcpt->scp_req_incoming); } static __attribute__((__noinline__)) int -ptlrpc_wait_event(struct ptlrpc_service *svc, - struct ptlrpc_thread *thread) +ptlrpc_wait_event(struct ptlrpc_service_part *svcpt, + struct ptlrpc_thread *thread) { - /* Don't exit while there are replies to be handled */ - struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout, - ptlrpc_retry_rqbds, svc); + /* Don't exit while there are replies to be handled */ + struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout, + ptlrpc_retry_rqbds, svcpt); - lc_watchdog_disable(thread->t_watchdog); + lc_watchdog_disable(thread->t_watchdog); - cfs_cond_resched(); + cfs_cond_resched(); - l_wait_event_exclusive_head(svc->srv_waitq, - ptlrpc_thread_stopping(thread) || - ptlrpc_server_request_waiting(svc) || - ptlrpc_server_request_pending(svc, 0) || - ptlrpc_rqbd_pending(svc) || - ptlrpc_at_check(svc), &lwi); + l_wait_event_exclusive_head(svcpt->scp_waitq, + ptlrpc_thread_stopping(thread) || + ptlrpc_server_request_incoming(svcpt) || + ptlrpc_server_request_pending(svcpt, 0) || + ptlrpc_rqbd_pending(svcpt) || + ptlrpc_at_check(svcpt), &lwi); - if (ptlrpc_thread_stopping(thread)) - return -EINTR; + if (ptlrpc_thread_stopping(thread)) + return -EINTR; - lc_watchdog_touch(thread->t_watchdog, CFS_GET_TIMEOUT(svc)); - - return 0; + lc_watchdog_touch(thread->t_watchdog, + ptlrpc_server_get_timeout(svcpt)); + return 0; } /** @@ -2109,10 +2172,11 @@ ptlrpc_wait_event(struct ptlrpc_service *svc, */ static int ptlrpc_main(void *arg) { - struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg; - struct ptlrpc_service *svc = data->svc; - struct ptlrpc_thread *thread = data->thread; - struct ptlrpc_reply_state *rs; + struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg; + struct ptlrpc_thread *thread = data->thread; + struct ptlrpc_service_part *svcpt = thread->t_svcpt; + struct ptlrpc_service *svc = svcpt->scp_service; + struct ptlrpc_reply_state *rs; #ifdef WITH_GROUP_INFO cfs_group_info_t *ginfo = NULL; #endif @@ -2181,74 +2245,76 @@ static int ptlrpc_main(void *arg) goto out_srv_fini; } - cfs_spin_lock(&svc->srv_lock); + cfs_spin_lock(&svcpt->scp_lock); - LASSERT(thread_is_starting(thread)); - thread_clear_flags(thread, SVC_STARTING); - svc->srv_threads_starting--; + LASSERT(thread_is_starting(thread)); + thread_clear_flags(thread, SVC_STARTING); - /* SVC_STOPPING may already be set here if someone else is trying - * to stop the service while this new thread has been dynamically - * forked. We still set SVC_RUNNING to let our creator know that - * we are now running, however we will exit as soon as possible */ - thread_add_flags(thread, SVC_RUNNING); - svc->srv_threads_running++; - cfs_spin_unlock(&svc->srv_lock); + svcpt->scp_nthrs_starting--; - /* - * wake up our creator. Note: @data is invalid after this point, - * because it's allocated on ptlrpc_start_thread() stack. - */ - cfs_waitq_signal(&thread->t_ctl_waitq); + /* SVC_STOPPING may already be set here if someone else is trying + * to stop the service while this new thread has been dynamically + * forked. We still set SVC_RUNNING to let our creator know that + * we are now running, however we will exit as soon as possible */ + thread_add_flags(thread, SVC_RUNNING); + svcpt->scp_nthrs_running++; + cfs_spin_unlock(&svcpt->scp_lock); - thread->t_watchdog = lc_watchdog_add(CFS_GET_TIMEOUT(svc), NULL, NULL); + /* + * wake up our creator. Note: @data is invalid after this point, + * because it's allocated on ptlrpc_start_thread() stack. + */ + cfs_waitq_signal(&thread->t_ctl_waitq); - cfs_spin_lock(&svc->srv_rs_lock); - cfs_list_add(&rs->rs_list, &svc->srv_free_rs_list); - cfs_waitq_signal(&svc->srv_free_rs_waitq); - cfs_spin_unlock(&svc->srv_rs_lock); + thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt), + NULL, NULL); - CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id, - svc->srv_threads_running); + cfs_spin_lock(&svcpt->scp_rep_lock); + cfs_list_add(&rs->rs_list, &svcpt->scp_rep_idle); + cfs_waitq_signal(&svcpt->scp_rep_waitq); + cfs_spin_unlock(&svcpt->scp_rep_lock); - /* XXX maintain a list of all managed devices: insert here */ - while (!ptlrpc_thread_stopping(thread)) { - if (ptlrpc_wait_event(svc, thread)) - break; + CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id, + svcpt->scp_nthrs_running); - ptlrpc_check_rqbd_pool(svc); + /* XXX maintain a list of all managed devices: insert here */ + while (!ptlrpc_thread_stopping(thread)) { + if (ptlrpc_wait_event(svcpt, thread)) + break; - if (ptlrpc_threads_need_create(svc)) { - /* Ignore return code - we tried... */ - ptlrpc_start_thread(svc); - } + ptlrpc_check_rqbd_pool(svcpt); - /* Process all incoming reqs before handling any */ - if (ptlrpc_server_request_waiting(svc)) { - ptlrpc_server_handle_req_in(svc); - /* but limit ourselves in case of flood */ - if (counter++ < 100) - continue; - counter = 0; + if (ptlrpc_threads_need_create(svcpt)) { + /* Ignore return code - we tried... */ + ptlrpc_start_thread(svcpt); } - if (ptlrpc_at_check(svc)) - ptlrpc_at_check_timed(svc); - - if (ptlrpc_server_request_pending(svc, 0)) { - lu_context_enter(&env->le_ctx); - ptlrpc_server_handle_request(svc, thread); - lu_context_exit(&env->le_ctx); + /* Process all incoming reqs before handling any */ + if (ptlrpc_server_request_incoming(svcpt)) { + ptlrpc_server_handle_req_in(svcpt); + /* but limit ourselves in case of flood */ + if (counter++ < 100) + continue; + counter = 0; + } + + if (ptlrpc_at_check(svcpt)) + ptlrpc_at_check_timed(svcpt); + + if (ptlrpc_server_request_pending(svcpt, 0)) { + lu_context_enter(&env->le_ctx); + ptlrpc_server_handle_request(svcpt, thread); + lu_context_exit(&env->le_ctx); } - if (ptlrpc_rqbd_pending(svc) && - ptlrpc_server_post_idle_rqbds(svc) < 0) { - /* I just failed to repost request buffers. - * Wait for a timeout (unless something else - * happens) before I try again */ - svc->srv_rqbd_timeout = cfs_time_seconds(1)/10; - CDEBUG(D_RPCTRACE,"Posted buffers: %d\n", - svc->srv_nrqbd_receiving); + if (ptlrpc_rqbd_pending(svcpt) && + ptlrpc_server_post_idle_rqbds(svcpt) < 0) { + /* I just failed to repost request buffers. + * Wait for a timeout (unless something else + * happens) before I try again */ + svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10; + CDEBUG(D_RPCTRACE, "Posted buffers: %d\n", + svcpt->scp_nrqbds_posted); } } @@ -2270,21 +2336,22 @@ out: CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n", thread, thread->t_pid, thread->t_id, rc); - cfs_spin_lock(&svc->srv_lock); - if (thread_test_and_clear_flags(thread, SVC_STARTING)) - svc->srv_threads_starting--; + cfs_spin_lock(&svcpt->scp_lock); + if (thread_test_and_clear_flags(thread, SVC_STARTING)) + svcpt->scp_nthrs_starting--; - if (thread_test_and_clear_flags(thread, SVC_RUNNING)) - /* must know immediately */ - svc->srv_threads_running--; + if (thread_test_and_clear_flags(thread, SVC_RUNNING)) { + /* must know immediately */ + svcpt->scp_nthrs_running--; + } - thread->t_id = rc; - thread_add_flags(thread, SVC_STOPPED); + thread->t_id = rc; + thread_add_flags(thread, SVC_STOPPED); - cfs_waitq_signal(&thread->t_ctl_waitq); - cfs_spin_unlock(&svc->srv_lock); + cfs_waitq_signal(&thread->t_ctl_waitq); + cfs_spin_unlock(&svcpt->scp_lock); - return rc; + return rc; } struct ptlrpc_hr_args { @@ -2428,30 +2495,30 @@ static int ptlrpc_start_hr_threads(struct ptlrpc_hr_service *hr) RETURN(0); } -static void ptlrpc_stop_thread(struct ptlrpc_service *svc, - struct ptlrpc_thread *thread) +static void ptlrpc_stop_thread(struct ptlrpc_service_part *svcpt, + struct ptlrpc_thread *thread) { - struct l_wait_info lwi = { 0 }; - ENTRY; + struct l_wait_info lwi = { 0 }; + ENTRY; - CDEBUG(D_RPCTRACE, "Stopping thread [ %p : %u ]\n", - thread, thread->t_pid); + CDEBUG(D_RPCTRACE, "Stopping thread [ %p : %u ]\n", + thread, thread->t_pid); - cfs_spin_lock(&svc->srv_lock); - /* let the thread know that we would like it to stop asap */ - thread_add_flags(thread, SVC_STOPPING); - cfs_spin_unlock(&svc->srv_lock); + cfs_spin_lock(&svcpt->scp_lock); + /* let the thread know that we would like it to stop asap */ + thread_add_flags(thread, SVC_STOPPING); + cfs_spin_unlock(&svcpt->scp_lock); - cfs_waitq_broadcast(&svc->srv_waitq); - l_wait_event(thread->t_ctl_waitq, - thread_is_stopped(thread), &lwi); + cfs_waitq_broadcast(&svcpt->scp_waitq); + l_wait_event(thread->t_ctl_waitq, + thread_is_stopped(thread), &lwi); - cfs_spin_lock(&svc->srv_lock); - cfs_list_del(&thread->t_link); - cfs_spin_unlock(&svc->srv_lock); + cfs_spin_lock(&svcpt->scp_lock); + cfs_list_del(&thread->t_link); + cfs_spin_unlock(&svcpt->scp_lock); - OBD_FREE_PTR(thread); - EXIT; + OBD_FREE_PTR(thread); + EXIT; } /** @@ -2459,21 +2526,24 @@ static void ptlrpc_stop_thread(struct ptlrpc_service *svc, */ void ptlrpc_stop_all_threads(struct ptlrpc_service *svc) { - struct ptlrpc_thread *thread; - ENTRY; + struct ptlrpc_service_part *svcpt = svc->srv_part; + struct ptlrpc_thread *thread; + ENTRY; - cfs_spin_lock(&svc->srv_lock); - while (!cfs_list_empty(&svc->srv_threads)) { - thread = cfs_list_entry(svc->srv_threads.next, - struct ptlrpc_thread, t_link); + LASSERT(svcpt != NULL); - cfs_spin_unlock(&svc->srv_lock); - ptlrpc_stop_thread(svc, thread); - cfs_spin_lock(&svc->srv_lock); - } + cfs_spin_lock(&svcpt->scp_lock); + while (!cfs_list_empty(&svcpt->scp_threads)) { + thread = cfs_list_entry(svcpt->scp_threads.next, + struct ptlrpc_thread, t_link); - cfs_spin_unlock(&svc->srv_lock); - EXIT; + cfs_spin_unlock(&svcpt->scp_lock); + ptlrpc_stop_thread(svcpt, thread); + cfs_spin_lock(&svcpt->scp_lock); + } + + cfs_spin_unlock(&svcpt->scp_lock); + EXIT; } int ptlrpc_start_threads(struct ptlrpc_service *svc) @@ -2485,7 +2555,7 @@ int ptlrpc_start_threads(struct ptlrpc_service *svc) ptlrpc_server_handle_request */ LASSERT(svc->srv_threads_min >= 2); for (i = 0; i < svc->srv_threads_min; i++) { - rc = ptlrpc_start_thread(svc); + rc = ptlrpc_start_thread(svc->srv_part); /* We have enough threads, don't start more. b=15759 */ if (rc == -EMFILE) { rc = 0; @@ -2501,46 +2571,49 @@ int ptlrpc_start_threads(struct ptlrpc_service *svc) RETURN(rc); } -int ptlrpc_start_thread(struct ptlrpc_service *svc) +int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt) { - struct l_wait_info lwi = { 0 }; - struct ptlrpc_svc_data d; - struct ptlrpc_thread *thread; - char name[32]; - int rc; - ENTRY; + struct l_wait_info lwi = { 0 }; + struct ptlrpc_svc_data d; + struct ptlrpc_thread *thread; + struct ptlrpc_service *svc = svcpt->scp_service; + char name[32]; + int rc; + ENTRY; - CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n", - svc->srv_name, svc->srv_threads_running, svc->srv_threads_min, - svc->srv_threads_max, svc->srv_threads_running); + LASSERT(svcpt != NULL); - if (unlikely(svc->srv_is_stopping)) - RETURN(-ESRCH); + CDEBUG(D_RPCTRACE, "%s started %d min %d max %d\n", + svc->srv_name, svcpt->scp_nthrs_running, + svc->srv_threads_min, svc->srv_threads_max); - if (!ptlrpc_threads_increasable(svc) || - (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) && - svc->srv_threads_running == svc->srv_threads_min - 1)) - RETURN(-EMFILE); + if (unlikely(svc->srv_is_stopping)) + RETURN(-ESRCH); - OBD_ALLOC_PTR(thread); - if (thread == NULL) - RETURN(-ENOMEM); - cfs_waitq_init(&thread->t_ctl_waitq); + if (!ptlrpc_threads_increasable(svcpt) || + (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) && + svcpt->scp_nthrs_running == svc->srv_threads_min - 1)) + RETURN(-EMFILE); - cfs_spin_lock(&svc->srv_lock); - if (!ptlrpc_threads_increasable(svc)) { - cfs_spin_unlock(&svc->srv_lock); - OBD_FREE_PTR(thread); - RETURN(-EMFILE); - } + OBD_ALLOC_PTR(thread); + if (thread == NULL) + RETURN(-ENOMEM); + cfs_waitq_init(&thread->t_ctl_waitq); + + cfs_spin_lock(&svcpt->scp_lock); + if (!ptlrpc_threads_increasable(svcpt)) { + cfs_spin_unlock(&svcpt->scp_lock); + OBD_FREE_PTR(thread); + RETURN(-EMFILE); + } - svc->srv_threads_starting++; - thread->t_id = svc->srv_threads_next_id++; - thread_add_flags(thread, SVC_STARTING); - thread->t_svc = svc; + svcpt->scp_nthrs_starting++; + thread->t_id = svcpt->scp_thr_nextid++; + thread_add_flags(thread, SVC_STARTING); + thread->t_svcpt = svcpt; - cfs_list_add(&thread->t_link, &svc->srv_threads); - cfs_spin_unlock(&svc->srv_lock); + cfs_list_add(&thread->t_link, &svcpt->scp_threads); + cfs_spin_unlock(&svcpt->scp_lock); sprintf(name, "%s_%02d", svc->srv_thread_name, thread->t_id); d.svc = svc; @@ -2556,10 +2629,10 @@ int ptlrpc_start_thread(struct ptlrpc_service *svc) if (rc < 0) { CERROR("cannot start thread '%s': rc %d\n", name, rc); - cfs_spin_lock(&svc->srv_lock); - cfs_list_del(&thread->t_link); - --svc->srv_threads_starting; - cfs_spin_unlock(&svc->srv_lock); + cfs_spin_lock(&svcpt->scp_lock); + cfs_list_del(&thread->t_link); + --svcpt->scp_nthrs_starting; + cfs_spin_unlock(&svcpt->scp_lock); OBD_FREE(thread, sizeof(*thread)); RETURN(rc); @@ -2572,7 +2645,6 @@ int ptlrpc_start_thread(struct ptlrpc_service *svc) RETURN(rc); } - int ptlrpc_hr_init(void) { int i; @@ -2622,35 +2694,43 @@ void ptlrpc_hr_fini(void) /** * Wait until all already scheduled replies are processed. */ -static void ptlrpc_wait_replies(struct ptlrpc_service *svc) -{ - while (1) { - int rc; - struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), - NULL, NULL); - rc = l_wait_event(svc->srv_waitq, cfs_atomic_read(&svc-> \ - srv_n_difficult_replies) == 0, - &lwi); - if (rc == 0) - break; - CWARN("Unexpectedly long timeout %p\n", svc); - } +static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt) +{ + while (1) { + int rc; + struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), + NULL, NULL); + + rc = l_wait_event(svcpt->scp_waitq, + cfs_atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi); + if (rc == 0) + break; + CWARN("Unexpectedly long timeout %s %p\n", + svcpt->scp_service->srv_name, svcpt->scp_service); + } } int ptlrpc_unregister_service(struct ptlrpc_service *service) { - int rc; - struct l_wait_info lwi; - cfs_list_t *tmp; - struct ptlrpc_reply_state *rs, *t; - struct ptlrpc_at_array *array = &service->srv_at_array; - ENTRY; + struct l_wait_info lwi; + struct ptlrpc_service_part *svcpt; + struct ptlrpc_reply_state *rs; + struct ptlrpc_reply_state *t; + struct ptlrpc_at_array *array; + cfs_list_t *tmp; + int rc; + ENTRY; + + service->srv_is_stopping = 1; + svcpt = service->srv_part; + + if (svcpt == NULL || /* no instance of ptlrpc_service_part */ + svcpt->scp_service == NULL) /* it's not fully initailzed */ + GOTO(out, rc = 0); - service->srv_is_stopping = 1; - cfs_timer_disarm(&service->srv_at_timer); + cfs_timer_disarm(&svcpt->scp_at_timer); - ptlrpc_stop_all_threads(service); - LASSERT(cfs_list_empty(&service->srv_threads)); + ptlrpc_stop_all_threads(service); cfs_spin_lock (&ptlrpc_all_services_lock); cfs_list_del_init (&service->srv_list); @@ -2667,9 +2747,9 @@ int ptlrpc_unregister_service(struct ptlrpc_service *service) rc = LNetClearLazyPortal(service->srv_req_portal); LASSERT (rc == 0); - /* Unlink all the request buffers. This forces a 'final' event with - * its 'unlink' flag set for each posted rqbd */ - cfs_list_for_each(tmp, &service->srv_active_rqbds) { + /* Unlink all the request buffers. This forces a 'final' event with + * its 'unlink' flag set for each posted rqbd */ + cfs_list_for_each(tmp, &svcpt->scp_rqbd_posted) { struct ptlrpc_request_buffer_desc *rqbd = cfs_list_entry(tmp, struct ptlrpc_request_buffer_desc, rqbd_list); @@ -2681,9 +2761,9 @@ int ptlrpc_unregister_service(struct ptlrpc_service *service) /* Wait for the network to release any buffers it's currently * filling */ for (;;) { - cfs_spin_lock(&service->srv_lock); - rc = service->srv_nrqbd_receiving; - cfs_spin_unlock(&service->srv_lock); + cfs_spin_lock(&svcpt->scp_lock); + rc = svcpt->scp_nrqbds_posted; + cfs_spin_unlock(&svcpt->scp_lock); if (rc == 0) break; @@ -2692,75 +2772,74 @@ int ptlrpc_unregister_service(struct ptlrpc_service *service) * 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(service->srv_waitq, - service->srv_nrqbd_receiving == 0, - &lwi); - if (rc == -ETIMEDOUT) - CWARN("Service %s waiting for request buffers\n", - service->srv_name); - } - - /* schedule all outstanding replies to terminate them */ - cfs_spin_lock(&service->srv_rs_lock); - while (!cfs_list_empty(&service->srv_active_replies)) { - struct ptlrpc_reply_state *rs = - cfs_list_entry(service->srv_active_replies.next, - struct ptlrpc_reply_state, rs_list); - cfs_spin_lock(&rs->rs_lock); - ptlrpc_schedule_difficult_reply(rs); - cfs_spin_unlock(&rs->rs_lock); - } - cfs_spin_unlock(&service->srv_rs_lock); - - /* purge the request queue. NB No new replies (rqbds all unlinked) - * and no service threads, so I'm the only thread noodling the - * request queue now */ - while (!cfs_list_empty(&service->srv_req_in_queue)) { - struct ptlrpc_request *req = - cfs_list_entry(service->srv_req_in_queue.next, - struct ptlrpc_request, - rq_list); - - cfs_list_del(&req->rq_list); - service->srv_n_queued_reqs--; - service->srv_n_active_reqs++; - ptlrpc_server_finish_request(service, req); - } - while (ptlrpc_server_request_pending(service, 1)) { - struct ptlrpc_request *req; - - req = ptlrpc_server_request_get(service, 1); - cfs_list_del(&req->rq_list); - service->srv_n_active_reqs++; - ptlrpc_server_finish_request(service, req); - } - LASSERT(service->srv_n_queued_reqs == 0); - LASSERT(service->srv_n_active_reqs == 0); - LASSERT(service->srv_n_history_rqbds == 0); - LASSERT(cfs_list_empty(&service->srv_active_rqbds)); - - /* Now free all the request buffers since nothing references them - * any more... */ - while (!cfs_list_empty(&service->srv_idle_rqbds)) { - struct ptlrpc_request_buffer_desc *rqbd = - cfs_list_entry(service->srv_idle_rqbds.next, - struct ptlrpc_request_buffer_desc, - rqbd_list); + rc = l_wait_event(svcpt->scp_waitq, + svcpt->scp_nrqbds_posted == 0, &lwi); + if (rc == -ETIMEDOUT) + CWARN("Service %s waiting for request buffers\n", + service->srv_name); + } - ptlrpc_free_rqbd(rqbd); - } + /* schedule all outstanding replies to terminate them */ + cfs_spin_lock(&svcpt->scp_rep_lock); + while (!cfs_list_empty(&svcpt->scp_rep_active)) { + struct ptlrpc_reply_state *rs = + cfs_list_entry(svcpt->scp_rep_active.next, + struct ptlrpc_reply_state, rs_list); + cfs_spin_lock(&rs->rs_lock); + ptlrpc_schedule_difficult_reply(rs); + cfs_spin_unlock(&rs->rs_lock); + } + cfs_spin_unlock(&svcpt->scp_rep_lock); + + /* purge the request queue. NB No new replies (rqbds all unlinked) + * and no service threads, so I'm the only thread noodling the + * request queue now */ + while (!cfs_list_empty(&svcpt->scp_req_incoming)) { + struct ptlrpc_request *req = + cfs_list_entry(svcpt->scp_req_incoming.next, + struct ptlrpc_request, + rq_list); + + cfs_list_del(&req->rq_list); + svcpt->scp_nreqs_incoming--; + svcpt->scp_nreqs_active++; + ptlrpc_server_finish_request(svcpt, req); + } + while (ptlrpc_server_request_pending(svcpt, 1)) { + struct ptlrpc_request *req; - ptlrpc_wait_replies(service); + req = ptlrpc_server_request_get(svcpt, 1); + cfs_list_del(&req->rq_list); + svcpt->scp_nreqs_active++; + ptlrpc_server_finish_request(svcpt, req); + } + LASSERT(svcpt->scp_nreqs_incoming == 0); + LASSERT(svcpt->scp_nreqs_active == 0); + LASSERT(svcpt->scp_hist_nrqbds == 0); + LASSERT(cfs_list_empty(&svcpt->scp_rqbd_posted)); + + /* Now free all the request buffers since nothing references them + * any more... */ + while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) { + struct ptlrpc_request_buffer_desc *rqbd = + cfs_list_entry(svcpt->scp_rqbd_idle.next, + struct ptlrpc_request_buffer_desc, + rqbd_list); + + ptlrpc_free_rqbd(rqbd); + } - cfs_list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, - rs_list) { - cfs_list_del(&rs->rs_list); - OBD_FREE_LARGE(rs, service->srv_max_reply_size); - } + ptlrpc_wait_replies(svcpt); + + cfs_list_for_each_entry_safe(rs, t, &svcpt->scp_rep_idle, rs_list) { + cfs_list_del(&rs->rs_list); + OBD_FREE_LARGE(rs, service->srv_max_reply_size); + } - /* In case somebody rearmed this in the meantime */ - cfs_timer_disarm(&service->srv_at_timer); + /* In case somebody rearmed this in the meantime */ + cfs_timer_disarm(&svcpt->scp_at_timer); + array = &svcpt->scp_at_array; if (array->paa_reqs_array != NULL) { OBD_FREE(array->paa_reqs_array, sizeof(cfs_list_t) * array->paa_size); @@ -2773,8 +2852,10 @@ int ptlrpc_unregister_service(struct ptlrpc_service *service) array->paa_reqs_count= NULL; } - OBD_FREE_PTR(service); - RETURN(0); + OBD_FREE_PTR(svcpt); + out: + OBD_FREE_PTR(service); + RETURN(0); } /** @@ -2785,37 +2866,41 @@ int ptlrpc_unregister_service(struct ptlrpc_service *service) * to be shot, so it's intentionally non-aggressive. */ int ptlrpc_service_health_check(struct ptlrpc_service *svc) { - struct ptlrpc_request *request; - struct timeval right_now; - long timediff; + struct ptlrpc_service_part *svcpt; + struct ptlrpc_request *request; + struct timeval right_now; + long timediff; - if (svc == NULL) - return 0; + if (svc == NULL || svc->srv_part == NULL) + return 0; - cfs_gettimeofday(&right_now); + cfs_gettimeofday(&right_now); - cfs_spin_lock(&svc->srv_rq_lock); - if (!ptlrpc_server_request_pending(svc, 1)) { - cfs_spin_unlock(&svc->srv_rq_lock); - return 0; - } + svcpt = svc->srv_part; + cfs_spin_lock(&svcpt->scp_req_lock); + if (!ptlrpc_server_request_pending(svcpt, 1)) { + cfs_spin_unlock(&svcpt->scp_req_lock); + return 0; + } - /* How long has the next entry been waiting? */ - if (cfs_list_empty(&svc->srv_request_queue)) - request = cfs_list_entry(svc->srv_request_hpq.next, - struct ptlrpc_request, rq_list); - else - request = cfs_list_entry(svc->srv_request_queue.next, - struct ptlrpc_request, rq_list); - timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL); - cfs_spin_unlock(&svc->srv_rq_lock); + /* How long has the next entry been waiting? */ + if (cfs_list_empty(&svcpt->scp_req_pending)) { + request = cfs_list_entry(svcpt->scp_hreq_pending.next, + struct ptlrpc_request, rq_list); + } else { + request = cfs_list_entry(svcpt->scp_req_pending.next, + struct ptlrpc_request, rq_list); + } - if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 : - at_max)) { - CERROR("%s: unhealthy - request has been waiting %lds\n", - svc->srv_name, timediff / ONE_MILLION); - return (-1); - } + timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL); + cfs_spin_unlock(&svcpt->scp_req_lock); - return 0; + if ((timediff / ONE_MILLION) > + (AT_OFF ? obd_timeout * 3 / 2 : at_max)) { + CERROR("%s: unhealthy - request has been waiting %lds\n", + svcpt->scp_service->srv_name, timediff / ONE_MILLION); + return -1; + } + + return 0; } diff --git a/lustre/quota/quota_context.c b/lustre/quota/quota_context.c index 7d92270..b6905f4 100644 --- a/lustre/quota/quota_context.c +++ b/lustre/quota/quota_context.c @@ -1128,7 +1128,8 @@ schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt, CDEBUG(D_QUOTA, "wake up when quota master is back\n"); if (oti->oti_thread->t_watchdog) lc_watchdog_touch(oti->oti_thread->t_watchdog, - CFS_GET_TIMEOUT(oti->oti_thread->t_svc)); + ptlrpc_server_get_timeout(\ + oti->oti_thread->t_svcpt)); } else { cfs_spin_unlock(&qctxt->lqc_lock); } diff --git a/lustre/quota/quota_interface.c b/lustre/quota/quota_interface.c index 1e68ade..77b10ec 100644 --- a/lustre/quota/quota_interface.c +++ b/lustre/quota/quota_interface.c @@ -452,23 +452,27 @@ static int quota_chk_acq_common(struct obd_device *obd, struct obd_export *exp, while ((rc = quota_check_common(obd, id, pending, count, cycle, isblk, inode, frags)) & QUOTA_RET_ACQUOTA) { - - cfs_spin_lock(&qctxt->lqc_lock); - if (!qctxt->lqc_import && oti) { - cfs_spin_unlock(&qctxt->lqc_lock); - LASSERT(oti->oti_thread); - /* The recovery thread doesn't have watchdog - * attached. LU-369 */ - if (oti->oti_thread->t_watchdog) - lc_watchdog_disable(oti->oti_thread->\ - t_watchdog); - CDEBUG(D_QUOTA, "sleep for quota master\n"); - l_wait_event(qctxt->lqc_wait_for_qmaster, check_qm(qctxt), - &lwi); - CDEBUG(D_QUOTA, "wake up when quota master is back\n"); - if (oti->oti_thread->t_watchdog) - lc_watchdog_touch(oti->oti_thread->t_watchdog, - CFS_GET_TIMEOUT(oti->oti_thread->t_svc)); + struct ptlrpc_thread *thr = oti != NULL ? + oti->oti_thread : NULL; + + cfs_spin_lock(&qctxt->lqc_lock); + if (!qctxt->lqc_import && oti != NULL) { + cfs_spin_unlock(&qctxt->lqc_lock); + + LASSERT(thr != NULL); + /* The recovery thread doesn't have watchdog + * attached. LU-369 */ + if (thr->t_watchdog != NULL) + lc_watchdog_disable(thr->t_watchdog); + CDEBUG(D_QUOTA, "sleep for quota master\n"); + l_wait_event(qctxt->lqc_wait_for_qmaster, + check_qm(qctxt), &lwi); + + CDEBUG(D_QUOTA, "wake up when quota master is back\n"); + if (thr->t_watchdog != NULL) { + lc_watchdog_touch(thr->t_watchdog, + ptlrpc_server_get_timeout(thr->t_svcpt)); + } } else { cfs_spin_unlock(&qctxt->lqc_lock); } @@ -510,9 +514,9 @@ static int quota_chk_acq_common(struct obd_device *obd, struct obd_export *exp, cfs_waitq_t waitq; struct l_wait_info lwi; - if (oti && oti->oti_thread && oti->oti_thread->t_watchdog) - lc_watchdog_touch(oti->oti_thread->t_watchdog, - CFS_GET_TIMEOUT(oti->oti_thread->t_svc)); + if (thr != NULL && thr->t_watchdog != NULL) + lc_watchdog_touch(thr->t_watchdog, + ptlrpc_server_get_timeout(thr->t_svcpt)); CDEBUG(D_QUOTA, "rc: %d, count_err: %d\n", rc, count_err++);