X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fptlrpc%2Fservice.c;h=6278f7b5d37bd5b69326683738132738b45b85b0;hp=c614eaa70a68081176ac8625066983efb0b6f4bf;hb=468c7759dd73594065dc029eeb73c514d589db3a;hpb=1ba7eb520d6f50908985b45cb1aecad83bdf9c63 diff --git a/lustre/ptlrpc/service.c b/lustre/ptlrpc/service.c index c614eaa..6278f7b 100644 --- a/lustre/ptlrpc/service.c +++ b/lustre/ptlrpc/service.c @@ -27,7 +27,7 @@ * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2012, Whamcloud, Inc. + * Copyright (c) 2010, 2012, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -64,11 +64,13 @@ CFS_MODULE_PARM(at_extra, "i", int, 0644, /* forward ref */ static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt); -static void ptlrpc_hpreq_fini(struct ptlrpc_request *req); +static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req); static void ptlrpc_at_remove_timed(struct ptlrpc_request *req); -static CFS_LIST_HEAD(ptlrpc_all_services); -cfs_spinlock_t ptlrpc_all_services_lock; +/** Holds a list of all PTLRPC services */ +CFS_LIST_HEAD(ptlrpc_all_services); +/** Used to protect the \e ptlrpc_all_services list */ +struct mutex ptlrpc_all_services_mutex; struct ptlrpc_request_buffer_desc * ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt) @@ -92,10 +94,10 @@ ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt) return NULL; } - cfs_spin_lock(&svcpt->scp_lock); + 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); + spin_unlock(&svcpt->scp_lock); return rqbd; } @@ -108,10 +110,10 @@ ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd) LASSERT(rqbd->rqbd_refcount == 0); LASSERT(cfs_list_empty(&rqbd->rqbd_reqs)); - cfs_spin_lock(&svcpt->scp_lock); + spin_lock(&svcpt->scp_lock); cfs_list_del(&rqbd->rqbd_list); svcpt->scp_nrqbds_total--; - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size); OBD_FREE_PTR(rqbd); @@ -125,9 +127,25 @@ ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post) int rc = 0; int i; + if (svcpt->scp_rqbd_allocating) + goto try_post; + + spin_lock(&svcpt->scp_lock); + /* check again with lock */ + if (svcpt->scp_rqbd_allocating) { + /* NB: we might allow more than one thread in the future */ + LASSERT(svcpt->scp_rqbd_allocating == 1); + spin_unlock(&svcpt->scp_lock); + goto try_post; + } + + svcpt->scp_rqbd_allocating++; + spin_unlock(&svcpt->scp_lock); + + 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. */ + /* NB: another thread might have recycled enough rqbds, we + * need to make sure it wouldn't over-allocate, see LU-1212. */ if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group) break; @@ -141,11 +159,19 @@ ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post) } } + spin_lock(&svcpt->scp_lock); + + LASSERT(svcpt->scp_rqbd_allocating == 1); + svcpt->scp_rqbd_allocating--; + + spin_unlock(&svcpt->scp_lock); + 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); + try_post: if (post && rc == 0) rc = ptlrpc_server_post_idle_rqbds(svcpt); @@ -184,7 +210,7 @@ struct ptlrpc_hr_partition; struct ptlrpc_hr_thread { int hrt_id; /* thread ID */ - cfs_spinlock_t hrt_lock; + spinlock_t hrt_lock; cfs_waitq_t hrt_waitq; cfs_list_t hrt_queue; /* RS queue */ struct ptlrpc_hr_partition *hrt_partition; @@ -283,9 +309,9 @@ static void rs_batch_dispatch(struct rs_batch *b) hrt = ptlrpc_hr_select(b->rsb_svcpt); - cfs_spin_lock(&hrt->hrt_lock); + spin_lock(&hrt->hrt_lock); cfs_list_splice_init(&b->rsb_replies, &hrt->hrt_queue); - cfs_spin_unlock(&hrt->hrt_lock); + spin_unlock(&hrt->hrt_lock); cfs_waitq_signal(&hrt->hrt_waitq); b->rsb_n_replies = 0; @@ -306,20 +332,20 @@ static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs) 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); + spin_unlock(&b->rsb_svcpt->scp_rep_lock); } - cfs_spin_lock(&svcpt->scp_rep_lock); + spin_lock(&svcpt->scp_rep_lock); b->rsb_svcpt = svcpt; - } - cfs_spin_lock(&rs->rs_lock); - rs->rs_scheduled_ever = 1; - if (rs->rs_scheduled == 0) { - cfs_list_move(&rs->rs_list, &b->rsb_replies); - rs->rs_scheduled = 1; - b->rsb_n_replies++; - } - rs->rs_committed = 1; - cfs_spin_unlock(&rs->rs_lock); + } + spin_lock(&rs->rs_lock); + rs->rs_scheduled_ever = 1; + if (rs->rs_scheduled == 0) { + cfs_list_move(&rs->rs_list, &b->rsb_replies); + rs->rs_scheduled = 1; + b->rsb_n_replies++; + } + rs->rs_committed = 1; + spin_unlock(&rs->rs_lock); } /** @@ -333,7 +359,7 @@ static void rs_batch_fini(struct rs_batch *b) { if (b->rsb_svcpt != NULL) { rs_batch_dispatch(b); - cfs_spin_unlock(&b->rsb_svcpt->scp_rep_lock); + spin_unlock(&b->rsb_svcpt->scp_rep_lock); } } @@ -362,9 +388,9 @@ void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs) hrt = ptlrpc_hr_select(rs->rs_svcpt); - cfs_spin_lock(&hrt->hrt_lock); + spin_lock(&hrt->hrt_lock); cfs_list_add_tail(&rs->rs_list, &hrt->hrt_queue); - cfs_spin_unlock(&hrt->hrt_lock); + spin_unlock(&hrt->hrt_lock); cfs_waitq_signal(&hrt->hrt_waitq); EXIT; @@ -406,7 +432,7 @@ void ptlrpc_commit_replies(struct obd_export *exp) * to attend to complete them. */ /* CAVEAT EMPTOR: spinlock ordering!!! */ - cfs_spin_lock(&exp->exp_uncommitted_replies_lock); + spin_lock(&exp->exp_uncommitted_replies_lock); cfs_list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies, rs_obd_list) { LASSERT (rs->rs_difficult); @@ -417,9 +443,9 @@ void ptlrpc_commit_replies(struct obd_export *exp) rs_batch_add(&batch, rs); } } - cfs_spin_unlock(&exp->exp_uncommitted_replies_lock); - rs_batch_fini(&batch); - EXIT; + spin_unlock(&exp->exp_uncommitted_replies_lock); + rs_batch_fini(&batch); + EXIT; } EXPORT_SYMBOL(ptlrpc_commit_replies); @@ -431,10 +457,10 @@ ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt) int posted = 0; for (;;) { - cfs_spin_lock(&svcpt->scp_lock); + spin_lock(&svcpt->scp_lock); if (cfs_list_empty(&svcpt->scp_rqbd_idle)) { - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); return posted; } @@ -447,7 +473,7 @@ ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt) svcpt->scp_nrqbds_posted++; cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted); - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); rc = ptlrpc_register_rqbd(rqbd); if (rc != 0) @@ -456,7 +482,7 @@ ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt) posted = 1; } - cfs_spin_lock(&svcpt->scp_lock); + spin_lock(&svcpt->scp_lock); svcpt->scp_nrqbds_posted--; cfs_list_del(&rqbd->rqbd_list); @@ -465,7 +491,7 @@ ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt) /* 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(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); return -1; } @@ -601,7 +627,7 @@ ptlrpc_service_part_init(struct ptlrpc_service *svc, CFS_INIT_LIST_HEAD(&svcpt->scp_threads); /* rqbd and incoming request queue */ - cfs_spin_lock_init(&svcpt->scp_lock); + 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); @@ -611,12 +637,10 @@ ptlrpc_service_part_init(struct ptlrpc_service *svc, 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); + spin_lock_init(&svcpt->scp_req_lock); /* reply states */ - cfs_spin_lock_init(&svcpt->scp_rep_lock); + 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); @@ -626,7 +650,7 @@ ptlrpc_service_part_init(struct ptlrpc_service *svc, cfs_atomic_set(&svcpt->scp_nreps_difficult, 0); /* adaptive timeout */ - cfs_spin_lock_init(&svcpt->scp_at_lock); + spin_lock_init(&svcpt->scp_at_lock); array = &svcpt->scp_at_array; size = at_est2timeout(at_max); @@ -753,15 +777,14 @@ ptlrpc_register_service(struct ptlrpc_service_conf *conf, service->srv_cpt_bits++; /* public members */ - cfs_spin_lock_init(&service->srv_lock); + 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 : - max(conf->psc_buf.bc_nbufs / - service->srv_ncpts, 1U); + 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 + SPTLRPC_MAX_PAYLOAD; service->srv_buf_size = conf->psc_buf.bc_buf_size; @@ -800,15 +823,19 @@ ptlrpc_register_service(struct ptlrpc_service_conf *conf, 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); + mutex_lock(&ptlrpc_all_services_mutex); + cfs_list_add (&service->srv_list, &ptlrpc_all_services); + mutex_unlock(&ptlrpc_all_services_mutex); + + if (proc_entry != NULL) + ptlrpc_lprocfs_register_service(proc_entry, service); - if (proc_entry != NULL) - ptlrpc_lprocfs_register_service(proc_entry, service); + rc = ptlrpc_service_nrs_setup(service); + if (rc != 0) + GOTO(failed, rc); - CDEBUG(D_NET, "%s: Started, listening on portal %d\n", - service->srv_name, service->srv_req_portal); + CDEBUG(D_NET, "%s: Started, listening on portal %d\n", + service->srv_name, service->srv_req_portal); #ifdef __KERNEL__ rc = ptlrpc_start_threads(service); @@ -866,12 +893,12 @@ void ptlrpc_server_drop_request(struct ptlrpc_request *req) return; if (req->rq_at_linked) { - cfs_spin_lock(&svcpt->scp_at_lock); + spin_lock(&svcpt->scp_at_lock); /* recheck with lock, in case it's unlinked by * ptlrpc_at_check_timed() */ if (likely(req->rq_at_linked)) ptlrpc_at_remove_timed(req); - cfs_spin_unlock(&svcpt->scp_at_lock); + spin_unlock(&svcpt->scp_at_lock); } LASSERT(cfs_list_empty(&req->rq_timed_list)); @@ -882,7 +909,7 @@ void ptlrpc_server_drop_request(struct ptlrpc_request *req) req->rq_export = NULL; } - cfs_spin_lock(&svcpt->scp_lock); + spin_lock(&svcpt->scp_lock); cfs_list_add(&req->rq_list, &rqbd->rqbd_reqs); @@ -918,7 +945,7 @@ void ptlrpc_server_drop_request(struct ptlrpc_request *req) cfs_list_del(&req->rq_history_list); } - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); cfs_list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) { req = cfs_list_entry(rqbd->rqbd_reqs.next, @@ -928,7 +955,7 @@ void ptlrpc_server_drop_request(struct ptlrpc_request *req) ptlrpc_server_free_request(req); } - cfs_spin_lock(&svcpt->scp_lock); + spin_lock(&svcpt->scp_lock); /* * now all reqs including the embedded req has been * disposed, schedule request buffer for re-use. @@ -939,7 +966,7 @@ void ptlrpc_server_drop_request(struct ptlrpc_request *req) &svcpt->scp_rqbd_idle); } - cfs_spin_unlock(&svcpt->scp_lock); + 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); @@ -949,11 +976,11 @@ void ptlrpc_server_drop_request(struct ptlrpc_request *req) if (req->rq_history_seq > svcpt->scp_hist_seq_culled) svcpt->scp_hist_seq_culled = req->rq_history_seq; - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); ptlrpc_server_free_request(req); } else { - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); } } @@ -964,13 +991,16 @@ void ptlrpc_server_drop_request(struct ptlrpc_request *req) static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt, struct ptlrpc_request *req) { - ptlrpc_hpreq_fini(req); + ptlrpc_server_hpreq_fini(req); - cfs_spin_lock(&svcpt->scp_req_lock); + spin_lock(&svcpt->scp_req_lock); + ptlrpc_nrs_req_stop_nolock(req); svcpt->scp_nreqs_active--; if (req->rq_hp) svcpt->scp_nhreqs_active--; - cfs_spin_unlock(&svcpt->scp_req_lock); + spin_unlock(&svcpt->scp_req_lock); + + ptlrpc_nrs_req_finalize(req); ptlrpc_server_drop_request(req); } @@ -1008,11 +1038,11 @@ static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay) /* exports may get disconnected from the chain even though the export has references, so we must keep the spin lock while manipulating the lists */ - cfs_spin_lock(&exp->exp_obd->obd_dev_lock); + spin_lock(&exp->exp_obd->obd_dev_lock); - if (cfs_list_empty(&exp->exp_obd_chain_timed)) { - /* this one is not timed */ - cfs_spin_unlock(&exp->exp_obd->obd_dev_lock); + if (cfs_list_empty(&exp->exp_obd_chain_timed)) { + /* this one is not timed */ + spin_unlock(&exp->exp_obd->obd_dev_lock); RETURN_EXIT; } @@ -1022,7 +1052,7 @@ static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay) oldest_exp = cfs_list_entry(exp->exp_obd->obd_exports_timed.next, struct obd_export, exp_obd_chain_timed); oldest_time = oldest_exp->exp_last_request_time; - cfs_spin_unlock(&exp->exp_obd->obd_dev_lock); + spin_unlock(&exp->exp_obd->obd_dev_lock); if (exp->exp_obd->obd_recovering) { /* be nice to everyone during recovery */ @@ -1067,14 +1097,14 @@ static int ptlrpc_check_req(struct ptlrpc_request *req) { int rc = 0; - if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) < - req->rq_export->exp_conn_cnt)) { - DEBUG_REQ(D_ERROR, req, - "DROPPING req from old connection %d < %d", - lustre_msg_get_conn_cnt(req->rq_reqmsg), - req->rq_export->exp_conn_cnt); - return -EEXIST; - } + if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) < + req->rq_export->exp_conn_cnt)) { + DEBUG_REQ(D_RPCTRACE, req, + "DROPPING req from old connection %d < %d", + lustre_msg_get_conn_cnt(req->rq_reqmsg), + req->rq_export->exp_conn_cnt); + return -EEXIST; + } if (unlikely(req->rq_export->exp_obd && req->rq_export->exp_obd->obd_fail)) { /* Failing over, don't handle any more reqs, send @@ -1144,7 +1174,7 @@ 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(&svcpt->scp_at_lock); + spin_lock(&svcpt->scp_at_lock); LASSERT(cfs_list_empty(&req->rq_timed_list)); index = (unsigned long)req->rq_deadline % array->paa_size; @@ -1167,17 +1197,17 @@ static int ptlrpc_at_add_timed(struct ptlrpc_request *req) cfs_list_add(&req->rq_timed_list, &array->paa_reqs_array[index]); - cfs_spin_lock(&req->rq_lock); - req->rq_at_linked = 1; - cfs_spin_unlock(&req->rq_lock); - req->rq_at_index = index; - array->paa_reqs_count[index]++; - array->paa_count++; - if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) { - array->paa_deadline = req->rq_deadline; + spin_lock(&req->rq_lock); + req->rq_at_linked = 1; + spin_unlock(&req->rq_lock); + req->rq_at_index = index; + array->paa_reqs_count[index]++; + array->paa_count++; + if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) { + array->paa_deadline = req->rq_deadline; ptlrpc_at_set_timer(svcpt); } - cfs_spin_unlock(&svcpt->scp_at_lock); + spin_unlock(&svcpt->scp_at_lock); return 0; } @@ -1193,9 +1223,9 @@ ptlrpc_at_remove_timed(struct ptlrpc_request *req) LASSERT(!cfs_list_empty(&req->rq_timed_list)); cfs_list_del_init(&req->rq_timed_list); - cfs_spin_lock(&req->rq_lock); + spin_lock(&req->rq_lock); req->rq_at_linked = 0; - cfs_spin_unlock(&req->rq_lock); + spin_unlock(&req->rq_lock); array->paa_reqs_count[req->rq_at_index]--; array->paa_count--; @@ -1351,16 +1381,16 @@ static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt) int first, counter = 0; ENTRY; - cfs_spin_lock(&svcpt->scp_at_lock); + spin_lock(&svcpt->scp_at_lock); if (svcpt->scp_at_check == 0) { - cfs_spin_unlock(&svcpt->scp_at_lock); + 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(&svcpt->scp_at_lock); + spin_unlock(&svcpt->scp_at_lock); RETURN(0); } @@ -1369,7 +1399,7 @@ static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt) if (first > at_early_margin) { /* We've still got plenty of time. Reset the timer. */ ptlrpc_at_set_timer(svcpt); - cfs_spin_unlock(&svcpt->scp_at_lock); + spin_unlock(&svcpt->scp_at_lock); RETURN(0); } @@ -1410,7 +1440,7 @@ static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt) /* we have a new earliest deadline, restart the timer */ ptlrpc_at_set_timer(svcpt); - cfs_spin_unlock(&svcpt->scp_at_lock); + spin_unlock(&svcpt->scp_at_lock); CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early " "replies\n", first, at_extra, counter); @@ -1447,35 +1477,53 @@ static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt) * Put the request to the export list if the request may become * a high priority one. */ -static int ptlrpc_hpreq_init(struct ptlrpc_service *svc, - struct ptlrpc_request *req) +static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt, + struct ptlrpc_request *req) { - int rc = 0; - ENTRY; + int rc = 0; + ENTRY; - if (svc->srv_ops.so_hpreq_handler) { - rc = svc->srv_ops.so_hpreq_handler(req); - if (rc) - RETURN(rc); - } - if (req->rq_export && req->rq_ops) { - /* Perform request specific check. We should do this check - * before the request is added into exp_hp_rpcs list otherwise - * it may hit swab race at LU-1044. */ - if (req->rq_ops->hpreq_check) - rc = req->rq_ops->hpreq_check(req); - - cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock); - cfs_list_add(&req->rq_exp_list, - &req->rq_export->exp_hp_rpcs); - cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock); - } + if (svcpt->scp_service->srv_ops.so_hpreq_handler) { + rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req); + if (rc < 0) + RETURN(rc); + LASSERT(rc == 0); + } + if (req->rq_export && req->rq_ops) { + /* Perform request specific check. We should do this check + * before the request is added into exp_hp_rpcs list otherwise + * it may hit swab race at LU-1044. */ + if (req->rq_ops->hpreq_check) { + rc = req->rq_ops->hpreq_check(req); + /** + * XXX: Out of all current + * ptlrpc_hpreq_ops::hpreq_check(), only + * ldlm_cancel_hpreq_check() can return an error code; + * other functions assert in similar places, which seems + * odd. What also does not seem right is that handlers + * for those RPCs do not assert on the same checks, but + * rather handle the error cases. e.g. see + * ost_rw_hpreq_check(), and ost_brw_read(), + * ost_brw_write(). + */ + if (rc < 0) + RETURN(rc); + LASSERT(rc == 0 || rc == 1); + } - RETURN(rc); + spin_lock_bh(&req->rq_export->exp_rpc_lock); + cfs_list_add(&req->rq_exp_list, + &req->rq_export->exp_hp_rpcs); + spin_unlock_bh(&req->rq_export->exp_rpc_lock); + } + + ptlrpc_nrs_req_initialize(svcpt, req, rc); + + RETURN(rc); } /** Remove the request from the export list. */ -static void ptlrpc_hpreq_fini(struct ptlrpc_request *req) +static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req) { ENTRY; if (req->rq_export && req->rq_ops) { @@ -1484,11 +1532,11 @@ static void ptlrpc_hpreq_fini(struct ptlrpc_request *req) if (req->rq_ops->hpreq_fini) req->rq_ops->hpreq_fini(req); - cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock); - cfs_list_del_init(&req->rq_exp_list); - cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock); - } - EXIT; + spin_lock_bh(&req->rq_export->exp_rpc_lock); + cfs_list_del_init(&req->rq_exp_list); + spin_unlock_bh(&req->rq_export->exp_rpc_lock); + } + EXIT; } static int ptlrpc_hpreq_check(struct ptlrpc_request *req) @@ -1497,9 +1545,7 @@ static int ptlrpc_hpreq_check(struct ptlrpc_request *req) } static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = { - .hpreq_lock_match = NULL, .hpreq_check = ptlrpc_hpreq_check, - .hpreq_fini = NULL, }; /* Hi-Priority RPC check by RPC operation code. */ @@ -1517,81 +1563,17 @@ int ptlrpc_hpreq_handler(struct ptlrpc_request *req) } EXPORT_SYMBOL(ptlrpc_hpreq_handler); -/** - * Make the request a high priority one. - * - * All the high priority requests are queued in a separate FIFO - * 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_part *svcpt, - struct ptlrpc_request *req) -{ - ENTRY; - - 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, &svcpt->scp_hreq_pending); - req->rq_hp = 1; - if (opc != OBD_PING) - DEBUG_REQ(D_RPCTRACE, req, "high priority req"); - } - cfs_spin_unlock(&req->rq_lock); - EXIT; -} - -/** - * \see ptlrpc_hpreq_reorder_nolock - */ -void ptlrpc_hpreq_reorder(struct ptlrpc_request *req) -{ - struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt; - ENTRY; - - 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; -} -EXPORT_SYMBOL(ptlrpc_hpreq_reorder); - -/** Check if the request is a high priority one. */ -static int ptlrpc_server_hpreq_check(struct ptlrpc_service *svc, - struct ptlrpc_request *req) -{ - return ptlrpc_hpreq_init(svc, req); -} - -/** Check if a request is a high priority one. */ static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt, struct ptlrpc_request *req) { int rc; ENTRY; - rc = ptlrpc_server_hpreq_check(svcpt->scp_service, req); + rc = ptlrpc_server_hpreq_init(svcpt, req); if (rc < 0) RETURN(rc); - cfs_spin_lock(&svcpt->scp_req_lock); - - if (rc) - ptlrpc_hpreq_reorder_nolock(svcpt, req); - else - cfs_list_add_tail(&req->rq_list, &svcpt->scp_req_pending); - - cfs_spin_unlock(&svcpt->scp_req_lock); + ptlrpc_nrs_req_add(svcpt, req, !!rc); RETURN(0); } @@ -1604,13 +1586,29 @@ static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt, static int ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt, int force) { + int running = svcpt->scp_nthrs_running; + + if (!nrs_svcpt_has_hp(svcpt)) + return 0; + if (force) return 1; - if (svcpt->scp_nreqs_active >= svcpt->scp_nthrs_running - 1) + if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL && + CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) { + /* leave just 1 thread for normal RPCs */ + running = PTLRPC_NTHRS_INIT; + if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) + running += 1; + } + + if (svcpt->scp_nreqs_active >= running - 1) return 0; - return cfs_list_empty(&svcpt->scp_req_pending) || + if (svcpt->scp_nhreqs_active == 0) + return 1; + + return !ptlrpc_nrs_req_pending_nolock(svcpt, false) || svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio; } @@ -1618,7 +1616,7 @@ static int ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt, int force) { return ptlrpc_server_allow_high(svcpt, force) && - !cfs_list_empty(&svcpt->scp_hreq_pending); + ptlrpc_nrs_req_pending_nolock(svcpt, true); } /** @@ -1633,26 +1631,35 @@ static int ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt, static int ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt, int force) { + int running = svcpt->scp_nthrs_running; #ifndef __KERNEL__ if (1) /* always allow to handle normal request for liblustre */ return 1; #endif + if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL && + CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) { + /* leave just 1 thread for normal RPCs */ + running = PTLRPC_NTHRS_INIT; + if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) + running += 1; + } + if (force || - svcpt->scp_nreqs_active < svcpt->scp_nthrs_running - 2) + svcpt->scp_nreqs_active < running - 2) return 1; - if (svcpt->scp_nreqs_active >= svcpt->scp_nthrs_running - 1) + if (svcpt->scp_nreqs_active >= running - 1) return 0; return svcpt->scp_nhreqs_active > 0 || - svcpt->scp_service->srv_ops.so_hpreq_handler == NULL; + !nrs_svcpt_has_hp(svcpt); } static int ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt, int force) { return ptlrpc_server_allow_normal(svcpt, force) && - !cfs_list_empty(&svcpt->scp_req_pending); + ptlrpc_nrs_req_pending_nolock(svcpt, false); } /** @@ -1682,15 +1689,13 @@ ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, int force) ENTRY; if (ptlrpc_server_high_pending(svcpt, force)) { - req = cfs_list_entry(svcpt->scp_hreq_pending.next, - struct ptlrpc_request, rq_list); + req = ptlrpc_nrs_req_poll_nolock(svcpt, true); svcpt->scp_hreq_count++; RETURN(req); } if (ptlrpc_server_normal_pending(svcpt, force)) { - req = cfs_list_entry(svcpt->scp_req_pending.next, - struct ptlrpc_request, rq_list); + req = ptlrpc_nrs_req_poll_nolock(svcpt, false); svcpt->scp_hreq_count = 0; RETURN(req); } @@ -1712,9 +1717,9 @@ ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt) int rc; ENTRY; - cfs_spin_lock(&svcpt->scp_lock); + spin_lock(&svcpt->scp_lock); if (cfs_list_empty(&svcpt->scp_req_incoming)) { - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); RETURN(0); } @@ -1724,7 +1729,7 @@ ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt) svcpt->scp_nreqs_incoming--; /* Consider this still a "queued" request as far as stats are * concerned */ - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); /* go through security check/transform */ rc = sptlrpc_svc_unwrap_request(req); @@ -1794,6 +1799,7 @@ ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt) req->rq_export = class_conn2export( lustre_msg_get_handle(req->rq_reqmsg)); if (req->rq_export) { + class_export_rpc_get(req->rq_export); rc = ptlrpc_check_req(req); if (rc == 0) { rc = sptlrpc_target_export_check(req->rq_export, req); @@ -1829,16 +1835,18 @@ ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt) /* Move it over to the request processing queue */ rc = ptlrpc_server_request_add(svcpt, req); if (rc) { - ptlrpc_hpreq_fini(req); + ptlrpc_server_hpreq_fini(req); GOTO(err_req, rc); } cfs_waitq_signal(&svcpt->scp_waitq); RETURN(1); err_req: - cfs_spin_lock(&svcpt->scp_req_lock); + if (req->rq_export) + class_export_rpc_put(req->rq_export); + spin_lock(&svcpt->scp_req_lock); svcpt->scp_nreqs_active++; - cfs_spin_unlock(&svcpt->scp_req_lock); + spin_unlock(&svcpt->scp_req_lock); ptlrpc_server_finish_request(svcpt, req); RETURN(1); @@ -1862,17 +1870,17 @@ ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt, int fail_opc = 0; ENTRY; - cfs_spin_lock(&svcpt->scp_req_lock); + spin_lock(&svcpt->scp_req_lock); #ifndef __KERNEL__ /* !@%$# liblustre only has 1 thread */ if (cfs_atomic_read(&svcpt->scp_nreps_difficult) != 0) { - cfs_spin_unlock(&svcpt->scp_req_lock); + spin_unlock(&svcpt->scp_req_lock); RETURN(0); } #endif request = ptlrpc_server_request_get(svcpt, 0); if (request == NULL) { - cfs_spin_unlock(&svcpt->scp_req_lock); + spin_unlock(&svcpt->scp_req_lock); RETURN(0); } @@ -1883,25 +1891,25 @@ ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt, if (unlikely(fail_opc)) { if (request->rq_export && request->rq_ops) { - cfs_spin_unlock(&svcpt->scp_req_lock); + spin_unlock(&svcpt->scp_req_lock); OBD_FAIL_TIMEOUT(fail_opc, 4); - cfs_spin_lock(&svcpt->scp_req_lock); + spin_lock(&svcpt->scp_req_lock); request = ptlrpc_server_request_get(svcpt, 0); if (request == NULL) { - cfs_spin_unlock(&svcpt->scp_req_lock); + spin_unlock(&svcpt->scp_req_lock); RETURN(0); } } } - - cfs_list_del_init(&request->rq_list); + ptlrpc_nrs_req_del_nolock(request); svcpt->scp_nreqs_active++; if (request->rq_hp) svcpt->scp_nhreqs_active++; - cfs_spin_unlock(&svcpt->scp_req_lock); + ptlrpc_nrs_req_start_nolock(request); + spin_unlock(&svcpt->scp_req_lock); ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET); @@ -1921,6 +1929,7 @@ ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt, at_get(&svcpt->scp_at_estimate)); } + export = request->rq_export; rc = lu_context_init(&request->rq_session, LCT_SESSION | LCT_NOREF); if (rc) { CERROR("Failure to initialize session: %d\n", rc); @@ -1937,10 +1946,9 @@ ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt, request->rq_svc_thread->t_env->le_ses = &request->rq_session; if (likely(request->rq_export)) { - if (unlikely(ptlrpc_check_req(request))) - goto put_conn; + if (unlikely(ptlrpc_check_req(request))) + goto put_conn; ptlrpc_update_export_timer(request->rq_export, timediff >> 19); - export = class_export_rpc_get(request->rq_export); } /* Discard requests queued for longer than the deadline. @@ -1953,7 +1961,7 @@ ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt, request->rq_arrival_time.tv_sec), cfs_time_sub(cfs_time_current_sec(), request->rq_deadline)); - goto put_rpc_export; + goto put_conn; } CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc " @@ -1973,26 +1981,23 @@ ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt, ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE); -put_rpc_export: - if (export != NULL) - class_export_rpc_put(export); put_conn: lu_context_exit(&request->rq_session); lu_context_fini(&request->rq_session); - if (unlikely(cfs_time_current_sec() > request->rq_deadline)) { - DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer " - "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);" - " client may timeout.", - request->rq_xid, cfs_time_sub(request->rq_deadline, - request->rq_arrival_time.tv_sec), - cfs_time_sub(cfs_time_current_sec(), - request->rq_deadline)); - } + if (unlikely(cfs_time_current_sec() > request->rq_deadline)) { + DEBUG_REQ(D_WARNING, request, "Request took longer " + "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);" + " client may timeout.", + cfs_time_sub(request->rq_deadline, + request->rq_arrival_time.tv_sec), + cfs_time_sub(cfs_time_current_sec(), + request->rq_deadline)); + } cfs_gettimeofday(&work_end); timediff = cfs_timeval_sub(&work_end, &work_start, NULL); - CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc " + CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc " "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in " "%ldus (%ldus total) trans "LPU64" rc %d/%d\n", cfs_curproc_comm(), @@ -2032,6 +2037,8 @@ put_conn: } out_req: + if (export != NULL) + class_export_rpc_put(export); ptlrpc_server_finish_request(svcpt, request); RETURN(1); @@ -2056,10 +2063,10 @@ ptlrpc_handle_rs(struct ptlrpc_reply_state *rs) LASSERT (rs->rs_scheduled); LASSERT (cfs_list_empty(&rs->rs_list)); - cfs_spin_lock (&exp->exp_lock); - /* Noop if removed already */ - cfs_list_del_init (&rs->rs_exp_list); - cfs_spin_unlock (&exp->exp_lock); + spin_lock(&exp->exp_lock); + /* Noop if removed already */ + cfs_list_del_init (&rs->rs_exp_list); + spin_unlock(&exp->exp_lock); /* The disk commit callback holds exp_uncommitted_replies_lock while it * iterates over newly committed replies, removing them from @@ -2083,13 +2090,13 @@ ptlrpc_handle_rs(struct ptlrpc_reply_state *rs) * holding rs_lock, we can be sure it has all completed once we hold * rs_lock, which we do right next. */ - if (!rs->rs_committed) { - cfs_spin_lock(&exp->exp_uncommitted_replies_lock); - cfs_list_del_init(&rs->rs_obd_list); - cfs_spin_unlock(&exp->exp_uncommitted_replies_lock); - } + if (!rs->rs_committed) { + spin_lock(&exp->exp_uncommitted_replies_lock); + cfs_list_del_init(&rs->rs_obd_list); + spin_unlock(&exp->exp_uncommitted_replies_lock); + } - cfs_spin_lock(&rs->rs_lock); + spin_lock(&rs->rs_lock); been_handled = rs->rs_handled; rs->rs_handled = 1; @@ -2100,34 +2107,33 @@ ptlrpc_handle_rs(struct ptlrpc_reply_state *rs) if (nlocks == 0 && !been_handled) { /* If we see this, we should already have seen the warning * in mds_steal_ack_locks() */ - CWARN("All locks stolen from rs %p x"LPD64".t"LPD64 - " o%d NID %s\n", - rs, - rs->rs_xid, rs->rs_transno, rs->rs_opc, - libcfs_nid2str(exp->exp_connection->c_peer.nid)); + CDEBUG(D_HA, "All locks stolen from rs %p x"LPD64".t"LPD64 + " o%d NID %s\n", + rs, + rs->rs_xid, rs->rs_transno, rs->rs_opc, + libcfs_nid2str(exp->exp_connection->c_peer.nid)); } if ((!been_handled && rs->rs_on_net) || nlocks > 0) { - cfs_spin_unlock(&rs->rs_lock); + spin_unlock(&rs->rs_lock); - if (!been_handled && rs->rs_on_net) { - LNetMDUnlink(rs->rs_md_h); - /* Ignore return code; we're racing with - * completion... */ - } + if (!been_handled && rs->rs_on_net) { + LNetMDUnlink(rs->rs_md_h); + /* Ignore return code; we're racing with completion */ + } - while (nlocks-- > 0) - ldlm_lock_decref(&rs->rs_locks[nlocks], - rs->rs_modes[nlocks]); + while (nlocks-- > 0) + ldlm_lock_decref(&rs->rs_locks[nlocks], + rs->rs_modes[nlocks]); - cfs_spin_lock(&rs->rs_lock); - } + spin_lock(&rs->rs_lock); + } - rs->rs_scheduled = 0; + rs->rs_scheduled = 0; - if (!rs->rs_on_net) { - /* Off the net */ - cfs_spin_unlock(&rs->rs_lock); + if (!rs->rs_on_net) { + /* Off the net */ + spin_unlock(&rs->rs_lock); class_export_put (exp); rs->rs_export = NULL; @@ -2139,7 +2145,7 @@ ptlrpc_handle_rs(struct ptlrpc_reply_state *rs) } /* still on the net; callback will schedule */ - cfs_spin_unlock(&rs->rs_lock); + spin_unlock(&rs->rs_lock); RETURN(1); } @@ -2159,14 +2165,14 @@ ptlrpc_server_handle_reply(struct ptlrpc_service_part *svcpt) struct ptlrpc_reply_state *rs = NULL; ENTRY; - cfs_spin_lock(&svcpt->scp_rep_lock); + 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); + spin_unlock(&svcpt->scp_rep_lock); if (rs != NULL) ptlrpc_handle_rs(rs); RETURN(rs != NULL); @@ -2420,7 +2426,7 @@ static int ptlrpc_main(void *arg) goto out_srv_fini; } - cfs_spin_lock(&svcpt->scp_lock); + spin_lock(&svcpt->scp_lock); LASSERT(thread_is_starting(thread)); thread_clear_flags(thread, SVC_STARTING); @@ -2434,7 +2440,7 @@ static int ptlrpc_main(void *arg) * 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); + spin_unlock(&svcpt->scp_lock); /* wake up our creator in case he's still waiting. */ cfs_waitq_signal(&thread->t_ctl_waitq); @@ -2442,10 +2448,10 @@ static int ptlrpc_main(void *arg) thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt), NULL, NULL); - cfs_spin_lock(&svcpt->scp_rep_lock); + 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); + spin_unlock(&svcpt->scp_rep_lock); CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id, svcpt->scp_nthrs_running); @@ -2509,7 +2515,7 @@ out: CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n", thread, thread->t_pid, thread->t_id, rc); - cfs_spin_lock(&svcpt->scp_lock); + spin_lock(&svcpt->scp_lock); if (thread_test_and_clear_flags(thread, SVC_STARTING)) svcpt->scp_nthrs_starting--; @@ -2522,7 +2528,7 @@ out: thread_add_flags(thread, SVC_STOPPED); cfs_waitq_signal(&thread->t_ctl_waitq); - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); return rc; } @@ -2532,12 +2538,12 @@ static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt, { int result; - cfs_spin_lock(&hrt->hrt_lock); + spin_lock(&hrt->hrt_lock); cfs_list_splice_init(&hrt->hrt_queue, replies); result = ptlrpc_hr.hr_stopping || !cfs_list_empty(replies); - cfs_spin_unlock(&hrt->hrt_lock); + spin_unlock(&hrt->hrt_lock); return result; } @@ -2651,7 +2657,7 @@ static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt) CDEBUG(D_INFO, "Stopping threads for service %s\n", svcpt->scp_service->srv_name); - cfs_spin_lock(&svcpt->scp_lock); + spin_lock(&svcpt->scp_lock); /* let the thread know that we would like it to stop asap */ list_for_each_entry(thread, &svcpt->scp_threads, t_link) { CDEBUG(D_INFO, "Stopping thread %s #%u\n", @@ -2669,17 +2675,17 @@ static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt) cfs_list_add(&thread->t_link, &zombie); continue; } - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n", svcpt->scp_service->srv_thread_name, thread->t_id); l_wait_event(thread->t_ctl_waitq, thread_is_stopped(thread), &lwi); - cfs_spin_lock(&svcpt->scp_lock); + spin_lock(&svcpt->scp_lock); } - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); while (!cfs_list_empty(&zombie)) { thread = cfs_list_entry(zombie.next, @@ -2744,12 +2750,14 @@ int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait) { struct l_wait_info lwi = { 0 }; struct ptlrpc_thread *thread; - struct ptlrpc_service *svc = svcpt->scp_service; + struct ptlrpc_service *svc; int rc; ENTRY; LASSERT(svcpt != NULL); + svc = svcpt->scp_service; + CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n", svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running, svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit); @@ -2768,9 +2776,9 @@ int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait) RETURN(-ENOMEM); cfs_waitq_init(&thread->t_ctl_waitq); - cfs_spin_lock(&svcpt->scp_lock); + spin_lock(&svcpt->scp_lock); if (!ptlrpc_threads_increasable(svcpt)) { - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); OBD_FREE_PTR(thread); RETURN(-EMFILE); } @@ -2779,7 +2787,7 @@ int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait) /* serialize starting because some modules (obdfilter) * might require unique and contiguous t_id */ LASSERT(svcpt->scp_nthrs_starting == 1); - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); OBD_FREE_PTR(thread); if (wait) { CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n", @@ -2799,7 +2807,7 @@ int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait) thread->t_svcpt = svcpt; cfs_list_add(&thread->t_link, &svcpt->scp_threads); - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); if (svcpt->scp_cpt >= 0) { snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d", @@ -2818,10 +2826,10 @@ int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait) if (rc < 0) { CERROR("cannot start thread '%s': rc %d\n", thread->t_name, rc); - cfs_spin_lock(&svcpt->scp_lock); + spin_lock(&svcpt->scp_lock); cfs_list_del(&thread->t_link); --svcpt->scp_nthrs_starting; - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); OBD_FREE(thread, sizeof(*thread)); RETURN(rc); @@ -2878,7 +2886,7 @@ int ptlrpc_hr_init(void) hrt->hrt_id = j; hrt->hrt_partition = hrp; cfs_waitq_init(&hrt->hrt_waitq); - cfs_spin_lock_init(&hrt->hrt_lock); + spin_lock_init(&hrt->hrt_lock); CFS_INIT_LIST_HEAD(&hrt->hrt_queue); } } @@ -2980,9 +2988,9 @@ ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc) /* Wait for the network to release any buffers * it's currently filling */ - cfs_spin_lock(&svcpt->scp_lock); + spin_lock(&svcpt->scp_lock); while (svcpt->scp_nrqbds_posted != 0) { - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); /* Network access will complete in finite time but * the HUGE timeout lets us CWARN for visibility * of sluggish NALs */ @@ -2996,9 +3004,9 @@ ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc) "request buffers\n", svcpt->scp_service->srv_name); } - cfs_spin_lock(&svcpt->scp_lock); + spin_lock(&svcpt->scp_lock); } - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); } } @@ -3015,15 +3023,15 @@ ptlrpc_service_purge_all(struct ptlrpc_service *svc) if (svcpt->scp_service == NULL) break; - cfs_spin_lock(&svcpt->scp_rep_lock); + spin_lock(&svcpt->scp_rep_lock); while (!cfs_list_empty(&svcpt->scp_rep_active)) { rs = cfs_list_entry(svcpt->scp_rep_active.next, struct ptlrpc_reply_state, rs_list); - cfs_spin_lock(&rs->rs_lock); + spin_lock(&rs->rs_lock); ptlrpc_schedule_difficult_reply(rs); - cfs_spin_unlock(&rs->rs_lock); + spin_unlock(&rs->rs_lock); } - cfs_spin_unlock(&svcpt->scp_rep_lock); + 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 @@ -3040,9 +3048,12 @@ ptlrpc_service_purge_all(struct ptlrpc_service *svc) while (ptlrpc_server_request_pending(svcpt, 1)) { req = ptlrpc_server_request_get(svcpt, 1); - cfs_list_del(&req->rq_list); + ptlrpc_nrs_req_del_nolock(req); svcpt->scp_nreqs_active++; - ptlrpc_hpreq_fini(req); + ptlrpc_server_hpreq_fini(req); + + if (req->rq_export != NULL) + class_export_rpc_put(req->rq_export); ptlrpc_server_finish_request(svcpt, req); } @@ -3120,17 +3131,19 @@ int ptlrpc_unregister_service(struct ptlrpc_service *service) service->srv_is_stopping = 1; - cfs_spin_lock(&ptlrpc_all_services_lock); + mutex_lock(&ptlrpc_all_services_mutex); cfs_list_del_init(&service->srv_list); - cfs_spin_unlock(&ptlrpc_all_services_lock); - - ptlrpc_lprocfs_unregister_service(service); + mutex_unlock(&ptlrpc_all_services_mutex); ptlrpc_service_del_atimer(service); ptlrpc_stop_all_threads(service); ptlrpc_service_unlink_rqbd(service); ptlrpc_service_purge_all(service); + ptlrpc_service_nrs_cleanup(service); + + ptlrpc_lprocfs_unregister_service(service); + ptlrpc_service_free(service); RETURN(0); @@ -3145,29 +3158,26 @@ EXPORT_SYMBOL(ptlrpc_unregister_service); * to be shot, so it's intentionally non-aggressive. */ int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt) { - struct ptlrpc_request *request; + struct ptlrpc_request *request = NULL; struct timeval right_now; long timediff; cfs_gettimeofday(&right_now); - cfs_spin_lock(&svcpt->scp_req_lock); - if (!ptlrpc_server_request_pending(svcpt, 1)) { - cfs_spin_unlock(&svcpt->scp_req_lock); - return 0; - } + spin_lock(&svcpt->scp_req_lock); + /* How long has the next entry been waiting? */ + if (ptlrpc_server_high_pending(svcpt, 1)) + request = ptlrpc_nrs_req_poll_nolock(svcpt, true); + else if (ptlrpc_server_normal_pending(svcpt, 1)) + request = ptlrpc_nrs_req_poll_nolock(svcpt, false); - /* 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 (request == NULL) { + spin_unlock(&svcpt->scp_req_lock); + return 0; } timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL); - cfs_spin_unlock(&svcpt->scp_req_lock); + spin_unlock(&svcpt->scp_req_lock); if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {