X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fptlrpc%2Fservice.c;h=ca2aa01675c3d13a8f674842aeaa8cd699868d2b;hp=84f3474ded23409e066bd0fc435acc5e25369d36;hb=9db3b77445d8b96f1f4a20ab03a6b8df3aec01e4;hpb=07b8db220e48782369f48d86213c5d404a628ded diff --git a/lustre/ptlrpc/service.c b/lustre/ptlrpc/service.c index 84f3474..ca2aa01 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, 2013, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -64,10 +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) @@ -75,26 +78,26 @@ ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt) struct ptlrpc_service *svc = svcpt->scp_service; struct ptlrpc_request_buffer_desc *rqbd; - OBD_ALLOC_PTR(rqbd); + OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt); if (rqbd == NULL) return NULL; rqbd->rqbd_svcpt = svcpt; - rqbd->rqbd_refcount = 0; - rqbd->rqbd_cbid.cbid_fn = request_in_callback; - rqbd->rqbd_cbid.cbid_arg = rqbd; - CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs); - OBD_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_buf_size); - - if (rqbd->rqbd_buffer == NULL) { - OBD_FREE_PTR(rqbd); - return (NULL); - } + rqbd->rqbd_refcount = 0; + rqbd->rqbd_cbid.cbid_fn = request_in_callback; + rqbd->rqbd_cbid.cbid_arg = rqbd; + CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs); + OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable, + svcpt->scp_cpt, svc->srv_buf_size); + if (rqbd->rqbd_buffer == NULL) { + OBD_FREE_PTR(rqbd); + 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; } @@ -107,26 +110,42 @@ 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); } int -ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt) +ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post) { struct ptlrpc_service *svc = svcpt->scp_service; struct ptlrpc_request_buffer_desc *rqbd; 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; @@ -138,17 +157,23 @@ ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt) rc = -ENOMEM; break; } - - if (ptlrpc_server_post_idle_rqbds(svcpt) < 0) { - rc = -EAGAIN; - break; - } } + 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); + 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); return rc; } @@ -177,25 +202,48 @@ ptlrpc_save_lock(struct ptlrpc_request *req, rs->rs_no_ack = !!no_ack; } } +EXPORT_SYMBOL(ptlrpc_save_lock); #ifdef __KERNEL__ -#define HRT_RUNNING 0 -#define HRT_STOPPING 1 +struct ptlrpc_hr_partition; struct ptlrpc_hr_thread { - cfs_spinlock_t hrt_lock; - unsigned long hrt_flags; - cfs_waitq_t hrt_wait; - cfs_list_t hrt_queue; - cfs_completion_t hrt_completion; + int hrt_id; /* thread ID */ + spinlock_t hrt_lock; + wait_queue_head_t hrt_waitq; + cfs_list_t hrt_queue; /* RS queue */ + struct ptlrpc_hr_partition *hrt_partition; }; +struct ptlrpc_hr_partition { + /* # of started threads */ + cfs_atomic_t hrp_nstarted; + /* # of stopped threads */ + cfs_atomic_t hrp_nstopped; + /* cpu partition id */ + int hrp_cpt; + /* round-robin rotor for choosing thread */ + int hrp_rotor; + /* total number of threads on this partition */ + int hrp_nthrs; + /* threads table */ + struct ptlrpc_hr_thread *hrp_thrs; +}; + +#define HRT_RUNNING 0 +#define HRT_STOPPING 1 + struct ptlrpc_hr_service { - int hr_index; - int hr_n_threads; - int hr_size; - struct ptlrpc_hr_thread hr_threads[0]; + /* CPU partition table, it's just cfs_cpt_table for now */ + struct cfs_cpt_table *hr_cpt_table; + /** controller sleep waitq */ + wait_queue_head_t hr_waitq; + unsigned int hr_stopping; + /** roundrobin rotor for non-affinity service */ + unsigned int hr_rotor; + /* partition data */ + struct ptlrpc_hr_partition **hr_partitions; }; struct rs_batch { @@ -204,10 +252,8 @@ struct rs_batch { struct ptlrpc_service_part *rsb_svcpt; }; -/** - * A pointer to per-node reply handling service. - */ -static struct ptlrpc_hr_service *ptlrpc_hr = NULL; +/** reply handling service. */ +static struct ptlrpc_hr_service ptlrpc_hr; /** * maximum mumber of replies scheduled in one batch @@ -228,17 +274,26 @@ static void rs_batch_init(struct rs_batch *b) /** * Choose an hr thread to dispatch requests to. */ -static unsigned int get_hr_thread_index(struct ptlrpc_hr_service *hr) +static struct ptlrpc_hr_thread * +ptlrpc_hr_select(struct ptlrpc_service_part *svcpt) { - unsigned int idx; + struct ptlrpc_hr_partition *hrp; + unsigned int rotor; + + if (svcpt->scp_cpt >= 0 && + svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) { + /* directly match partition */ + hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt]; + + } else { + rotor = ptlrpc_hr.hr_rotor++; + rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table); - /* Concurrent modification of hr_index w/o any spinlock - protection is harmless as long as the result fits - [0..(hr_n_threads-1)] range and each thread gets near equal - load. */ - idx = hr->hr_index; - hr->hr_index = (idx >= hr->hr_n_threads - 1) ? 0 : idx + 1; - return idx; + hrp = ptlrpc_hr.hr_partitions[rotor]; + } + + rotor = hrp->hrp_rotor++; + return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs]; } /** @@ -249,19 +304,18 @@ static unsigned int get_hr_thread_index(struct ptlrpc_hr_service *hr) */ static void rs_batch_dispatch(struct rs_batch *b) { - if (b->rsb_n_replies != 0) { - struct ptlrpc_hr_service *hr = ptlrpc_hr; - int idx; + if (b->rsb_n_replies != 0) { + struct ptlrpc_hr_thread *hrt; - idx = get_hr_thread_index(hr); + hrt = ptlrpc_hr_select(b->rsb_svcpt); - cfs_spin_lock(&hr->hr_threads[idx].hrt_lock); - cfs_list_splice_init(&b->rsb_replies, - &hr->hr_threads[idx].hrt_queue); - cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock); - cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait); - b->rsb_n_replies = 0; - } + spin_lock(&hrt->hrt_lock); + cfs_list_splice_init(&b->rsb_replies, &hrt->hrt_queue); + spin_unlock(&hrt->hrt_lock); + + wake_up(&hrt->hrt_waitq); + b->rsb_n_replies = 0; + } } /** @@ -278,20 +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); } /** @@ -305,7 +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); } } @@ -327,18 +381,19 @@ static void rs_batch_fini(struct rs_batch *b) void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs) { #ifdef __KERNEL__ - struct ptlrpc_hr_service *hr = ptlrpc_hr; - int idx; - ENTRY; + struct ptlrpc_hr_thread *hrt; + ENTRY; - LASSERT(cfs_list_empty(&rs->rs_list)); + LASSERT(cfs_list_empty(&rs->rs_list)); - idx = get_hr_thread_index(hr); - cfs_spin_lock(&hr->hr_threads[idx].hrt_lock); - cfs_list_add_tail(&rs->rs_list, &hr->hr_threads[idx].hrt_queue); - cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock); - cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait); - EXIT; + hrt = ptlrpc_hr_select(rs->rs_svcpt); + + spin_lock(&hrt->hrt_lock); + cfs_list_add_tail(&rs->rs_list, &hrt->hrt_queue); + spin_unlock(&hrt->hrt_lock); + + wake_up(&hrt->hrt_waitq); + EXIT; #else cfs_list_add_tail(&rs->rs_list, &rs->rs_svcpt->scp_rep_queue); #endif @@ -349,21 +404,22 @@ ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs) { ENTRY; - 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 */ + LASSERT(spin_is_locked(&rs->rs_svcpt->scp_rep_lock)); + LASSERT(spin_is_locked(&rs->rs_lock)); + LASSERT (rs->rs_difficult); + rs->rs_scheduled_ever = 1; /* flag any notification attempt */ - if (rs->rs_scheduled) { /* being set up or already notified */ - EXIT; - return; - } + if (rs->rs_scheduled) { /* being set up or already notified */ + EXIT; + return; + } - rs->rs_scheduled = 1; - cfs_list_del_init(&rs->rs_list); - ptlrpc_dispatch_difficult_reply(rs); - EXIT; + rs->rs_scheduled = 1; + cfs_list_del_init(&rs->rs_list); + ptlrpc_dispatch_difficult_reply(rs); + EXIT; } +EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply); void ptlrpc_commit_replies(struct obd_export *exp) { @@ -376,7 +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); @@ -387,10 +443,11 @@ 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); static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt) @@ -400,10 +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; } @@ -416,7 +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) @@ -425,7 +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); @@ -434,7 +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; } @@ -447,53 +504,110 @@ static void ptlrpc_at_timer(unsigned long castmeharder) svcpt->scp_at_check = 1; svcpt->scp_at_checktime = cfs_time_current(); - cfs_waitq_signal(&svcpt->scp_waitq); + wake_up(&svcpt->scp_waitq); } static void -ptlrpc_server_nthreads_check(struct ptlrpc_service_conf *conf, - int *min_p, int *max_p) +ptlrpc_server_nthreads_check(struct ptlrpc_service *svc, + struct ptlrpc_service_conf *conf) { #ifdef __KERNEL__ struct ptlrpc_service_thr_conf *tc = &conf->psc_thr; - int nthrs_min; - int nthrs; + unsigned init; + unsigned total; + unsigned nthrs; + int weight; - nthrs_min = PTLRPC_NTHRS_MIN + (conf->psc_ops.so_hpreq_handler != NULL); - nthrs_min = max_t(int, nthrs_min, tc->tc_nthrs_min); + /* + * Common code for estimating & validating threads number. + * CPT affinity service could have percpt thread-pool instead + * of a global thread-pool, which means user might not always + * get the threads number they give it in conf::tc_nthrs_user + * even they did set. It's because we need to validate threads + * number for each CPT to guarantee each pool will have enough + * threads to keep the service healthy. + */ + init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL); + init = max_t(int, init, tc->tc_nthrs_init); + + /* NB: please see comments in lustre_lnet.h for definition + * details of these members */ + LASSERT(tc->tc_nthrs_max != 0); + + if (tc->tc_nthrs_user != 0) { + /* In case there is a reason to test a service with many + * threads, we give a less strict check here, it can + * be up to 8 * nthrs_max */ + total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user); + nthrs = total / svc->srv_ncpts; + init = max(init, nthrs); + goto out; + } - nthrs = tc->tc_nthrs_user; - if (nthrs != 0) { /* validate it */ - nthrs = min_t(int, nthrs, tc->tc_nthrs_max); - nthrs = max_t(int, nthrs, nthrs_min); - *min_p = *max_p = nthrs; - return; + total = tc->tc_nthrs_max; + if (tc->tc_nthrs_base == 0) { + /* don't care about base threads number per partition, + * this is most for non-affinity service */ + nthrs = total / svc->srv_ncpts; + goto out; } - /* - * NB: we will add some common code here for estimating, for example: - * add a new member ptlrpc_service_thr_conf::tc_factor, and estimate - * threads number based on: - * (online_cpus * conf::tc_factor) + conf::tc_nthrs_base. - * - * So we can remove code block like estimation in ost_setup, also, - * we might estimate MDS threads number as well instead of using - * absolute number, and have more threads on fat servers to improve - * availability of service. - * - * Also, we will need to validate threads number at here for - * CPT affinity service (CPU ParTion) in the future. - * A service can have percpt thread-pool instead of a global thread - * pool for each service, which means user might not always get the - * threads number they want even they set it in conf::tc_nthrs_user, - * because we need to adjust threads number for each CPT, instead of - * just use (conf::tc_nthrs_user / NCPTS), to make sure each pool - * will be healthy. - */ - *max_p = tc->tc_nthrs_max; - *min_p = nthrs_min; -#else /* __KERNEL__ */ - *max_p = *min_p = 1; /* whatever */ + nthrs = tc->tc_nthrs_base; + if (svc->srv_ncpts == 1) { + int i; + + /* NB: Increase the base number if it's single partition + * and total number of cores/HTs is larger or equal to 4. + * result will always < 2 * nthrs_base */ + weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY); + for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */ + (tc->tc_nthrs_base >> i) != 0; i++) + nthrs += tc->tc_nthrs_base >> i; + } + + if (tc->tc_thr_factor != 0) { + int factor = tc->tc_thr_factor; + const int fade = 4; + + /* + * User wants to increase number of threads with for + * each CPU core/HT, most likely the factor is larger then + * one thread/core because service threads are supposed to + * be blocked by lock or wait for IO. + */ + /* + * Amdahl's law says that adding processors wouldn't give + * a linear increasing of parallelism, so it's nonsense to + * have too many threads no matter how many cores/HTs + * there are. + */ + if (cfs_cpu_ht_nsiblings(0) > 1) { /* weight is # of HTs */ + /* depress thread factor for hyper-thread */ + factor = factor - (factor >> 1) + (factor >> 3); + } + + weight = cfs_cpt_weight(svc->srv_cptable, 0); + LASSERT(weight > 0); + + for (; factor > 0 && weight > 0; factor--, weight -= fade) + nthrs += min(weight, fade) * factor; + } + + if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) { + nthrs = max(tc->tc_nthrs_base, + tc->tc_nthrs_max / svc->srv_ncpts); + } + out: + nthrs = max(nthrs, tc->tc_nthrs_init); + svc->srv_nthrs_cpt_limit = nthrs; + svc->srv_nthrs_cpt_init = init; + + if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) { + CDEBUG(D_OTHER, "%s: This service may have more threads (%d) " + "than the given soft limit (%d)\n", + svc->srv_name, nthrs * svc->srv_ncpts, + tc->tc_nthrs_max); + } #endif } @@ -502,42 +616,41 @@ ptlrpc_server_nthreads_check(struct ptlrpc_service_conf *conf, */ static int ptlrpc_service_part_init(struct ptlrpc_service *svc, - struct ptlrpc_service_part *svcpt) + struct ptlrpc_service_part *svcpt, int cpt) { struct ptlrpc_at_array *array; int size; int index; int rc; + svcpt->scp_cpt = cpt; 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); - cfs_waitq_init(&svcpt->scp_waitq); + init_waitqueue_head(&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); + 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); #endif CFS_INIT_LIST_HEAD(&svcpt->scp_rep_idle); - cfs_waitq_init(&svcpt->scp_rep_waitq); + init_waitqueue_head(&svcpt->scp_rep_waitq); 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); @@ -546,14 +659,16 @@ ptlrpc_service_part_init(struct ptlrpc_service *svc, array->paa_deadline = -1; /* allocate memory for scp_at_array (ptlrpc_at_array) */ - OBD_ALLOC(array->paa_reqs_array, sizeof(cfs_list_t) * size); + OBD_CPT_ALLOC(array->paa_reqs_array, + svc->srv_cptable, cpt, 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); + OBD_CPT_ALLOC(array->paa_reqs_count, + svc->srv_cptable, cpt, sizeof(__u32) * size); if (array->paa_reqs_count == NULL) goto failed; @@ -565,7 +680,7 @@ ptlrpc_service_part_init(struct ptlrpc_service *svc, /* 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); + rc = ptlrpc_grow_req_bufs(svcpt, 0); /* 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) @@ -595,10 +710,17 @@ ptlrpc_service_part_init(struct ptlrpc_service *svc, */ struct ptlrpc_service * ptlrpc_register_service(struct ptlrpc_service_conf *conf, - cfs_proc_dir_entry_t *proc_entry) + struct proc_dir_entry *proc_entry) { + struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt; struct ptlrpc_service *service; + struct ptlrpc_service_part *svcpt; + struct cfs_cpt_table *cptable; + __u32 *cpts = NULL; + int ncpts; + int cpt; int rc; + int i; ENTRY; LASSERT(conf->psc_buf.bc_nbufs > 0); @@ -606,12 +728,56 @@ ptlrpc_register_service(struct ptlrpc_service_conf *conf, conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD); LASSERT(conf->psc_thr.tc_ctx_tags != 0); - OBD_ALLOC_PTR(service); - if (service == NULL) + cptable = cconf->cc_cptable; + if (cptable == NULL) + cptable = cfs_cpt_table; + + if (!conf->psc_thr.tc_cpu_affinity) { + ncpts = 1; + } else { + ncpts = cfs_cpt_number(cptable); + if (cconf->cc_pattern != NULL) { + struct cfs_expr_list *el; + + rc = cfs_expr_list_parse(cconf->cc_pattern, + strlen(cconf->cc_pattern), + 0, ncpts - 1, &el); + if (rc != 0) { + CERROR("%s: invalid CPT pattern string: %s", + conf->psc_name, cconf->cc_pattern); + RETURN(ERR_PTR(-EINVAL)); + } + + rc = cfs_expr_list_values(el, ncpts, &cpts); + cfs_expr_list_free(el); + if (rc <= 0) { + CERROR("%s: failed to parse CPT array %s: %d\n", + conf->psc_name, cconf->cc_pattern, rc); + if (cpts != NULL) + OBD_FREE(cpts, sizeof(*cpts) * ncpts); + RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL)); + } + ncpts = rc; + } + } + + OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts])); + if (service == NULL) { + if (cpts != NULL) + OBD_FREE(cpts, sizeof(*cpts) * ncpts); RETURN(ERR_PTR(-ENOMEM)); + } + + service->srv_cptable = cptable; + service->srv_cpts = cpts; + service->srv_ncpts = ncpts; + + service->srv_cpt_bits = 0; /* it's zero already, easy to read... */ + while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable)) + 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 */ @@ -631,35 +797,45 @@ ptlrpc_register_service(struct ptlrpc_service_conf *conf, 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); - service->srv_thread_name = conf->psc_thr.tc_thr_name; 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_ops = conf->psc_ops; - OBD_ALLOC_PTR(service->srv_part); - if (service->srv_part == NULL) - GOTO(failed, rc = -ENOMEM); + for (i = 0; i < ncpts; i++) { + if (!conf->psc_thr.tc_cpu_affinity) + cpt = CFS_CPT_ANY; + else + cpt = cpts != NULL ? cpts[i] : i; - rc = ptlrpc_service_part_init(service, service->srv_part); - if (rc != 0) - GOTO(failed, rc); + OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt)); + if (svcpt == NULL) + GOTO(failed, rc = -ENOMEM); + + service->srv_parts[i] = svcpt; + rc = ptlrpc_service_part_init(service, svcpt, cpt); + if (rc != 0) + GOTO(failed, rc); + } + + ptlrpc_server_nthreads_check(service, 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); @@ -675,6 +851,7 @@ failed: ptlrpc_unregister_service(service); RETURN(ERR_PTR(rc)); } +EXPORT_SYMBOL(ptlrpc_register_service); /** * to actually free the request, must be called without holding svc_lock. @@ -692,11 +869,11 @@ static void ptlrpc_server_free_request(struct ptlrpc_request *req) sptlrpc_svc_ctx_decref(req); if (req != &req->rq_rqbd->rqbd_req) { - /* NB request buffers use an embedded - * req if the incoming req unlinked the - * MD; this isn't one of them! */ - OBD_FREE(req, sizeof(*req)); - } + /* NB request buffers use an embedded + * req if the incoming req unlinked the + * MD; this isn't one of them! */ + ptlrpc_request_cache_free(req); + } } /** @@ -716,31 +893,23 @@ void ptlrpc_server_drop_request(struct ptlrpc_request *req) return; if (req->rq_at_linked) { - struct ptlrpc_at_array *array = &svcpt->scp_at_array; - __u32 index = req->rq_at_index; - - cfs_spin_lock(&svcpt->scp_at_lock); - - LASSERT(!cfs_list_empty(&req->rq_timed_list)); - cfs_list_del_init(&req->rq_timed_list); - cfs_spin_lock(&req->rq_lock); - req->rq_at_linked = 0; - cfs_spin_unlock(&req->rq_lock); - array->paa_reqs_count[index]--; - array->paa_count--; - - cfs_spin_unlock(&svcpt->scp_at_lock); - } else { - LASSERT(cfs_list_empty(&req->rq_timed_list)); + 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); + spin_unlock(&svcpt->scp_at_lock); } + LASSERT(cfs_list_empty(&req->rq_timed_list)); + /* finalize request */ if (req->rq_export) { class_export_put(req->rq_export); req->rq_export = NULL; } - cfs_spin_lock(&svcpt->scp_lock); + spin_lock(&svcpt->scp_lock); cfs_list_add(&req->rq_list, &rqbd->rqbd_reqs); @@ -754,7 +923,7 @@ void ptlrpc_server_drop_request(struct ptlrpc_request *req) /* 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) { + while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) { rqbd = cfs_list_entry(svcpt->scp_hist_rqbds.next, struct ptlrpc_request_buffer_desc, rqbd_list); @@ -776,7 +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, @@ -786,7 +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. @@ -797,36 +966,93 @@ 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); cfs_list_del_init(&req->rq_history_list); - cfs_spin_unlock(&svcpt->scp_lock); + /* Track the highest culled req seq */ + if (req->rq_history_seq > svcpt->scp_hist_seq_culled) + svcpt->scp_hist_seq_culled = req->rq_history_seq; + + spin_unlock(&svcpt->scp_lock); ptlrpc_server_free_request(req); } else { - cfs_spin_unlock(&svcpt->scp_lock); + spin_unlock(&svcpt->scp_lock); } } +/** Change request export and move hp request from old export to new */ +void ptlrpc_request_change_export(struct ptlrpc_request *req, + struct obd_export *export) +{ + if (req->rq_export != NULL) { + LASSERT(!list_empty(&req->rq_exp_list)); + /* remove rq_exp_list from last export */ + spin_lock_bh(&req->rq_export->exp_rpc_lock); + list_del_init(&req->rq_exp_list); + spin_unlock_bh(&req->rq_export->exp_rpc_lock); + /* export has one reference already, so it`s safe to + * add req to export queue here and get another + * reference for request later */ + spin_lock_bh(&export->exp_rpc_lock); + if (req->rq_ops != NULL) /* hp request */ + list_add(&req->rq_exp_list, &export->exp_hp_rpcs); + else + list_add(&req->rq_exp_list, &export->exp_reg_rpcs); + spin_unlock_bh(&export->exp_rpc_lock); + + class_export_rpc_dec(req->rq_export); + class_export_put(req->rq_export); + } + + /* request takes one export refcount */ + req->rq_export = class_export_get(export); + class_export_rpc_inc(export); + + return; +} + /** * to finish a request: stop sending more early replies, and release - * the request. should be called after we finished handling the request. + * the request. */ 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); + if (req->rq_session.lc_thread != NULL) { + lu_context_exit(&req->rq_session); + lu_context_fini(&req->rq_session); + } + + ptlrpc_server_drop_request(req); +} + +/** + * to finish a active request: stop sending more early replies, and release + * the request. should be called after we finished handling the request. + */ +static void ptlrpc_server_finish_active_request( + struct ptlrpc_service_part *svcpt, + struct ptlrpc_request *req) +{ + 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_server_drop_request(req); + ptlrpc_nrs_req_finalize(req); + + if (req->rq_export != NULL) + class_export_rpc_dec(req->rq_export); + + ptlrpc_server_finish_request(svcpt, req); } /** @@ -862,11 +1088,11 @@ 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; } @@ -876,7 +1102,7 @@ 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 */ @@ -921,14 +1147,14 @@ 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 @@ -998,7 +1224,7 @@ 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; @@ -1021,21 +1247,40 @@ 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; } +static void +ptlrpc_at_remove_timed(struct ptlrpc_request *req) +{ + struct ptlrpc_at_array *array; + + array = &req->rq_rqbd->rqbd_svcpt->scp_at_array; + + /* NB: must call with hold svcpt::scp_at_lock */ + LASSERT(!cfs_list_empty(&req->rq_timed_list)); + cfs_list_del_init(&req->rq_timed_list); + + spin_lock(&req->rq_lock); + req->rq_at_linked = 0; + spin_unlock(&req->rq_lock); + + array->paa_reqs_count[req->rq_at_index]--; + array->paa_count--; +} + static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req) { struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt; @@ -1104,14 +1349,12 @@ static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req) } newdl = cfs_time_current_sec() + at_get(&svcpt->scp_at_estimate); - OBD_ALLOC(reqcopy, sizeof *reqcopy); - if (reqcopy == NULL) - RETURN(-ENOMEM); - OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen); - if (!reqmsg) { - OBD_FREE(reqcopy, sizeof *reqcopy); - RETURN(-ENOMEM); - } + reqcopy = ptlrpc_request_cache_alloc(__GFP_IO); + if (reqcopy == NULL) + RETURN(-ENOMEM); + OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen); + if (!reqmsg) + GOTO(out_free, rc = -ENOMEM); *reqcopy = *req; reqcopy->rq_reply_state = NULL; @@ -1139,7 +1382,7 @@ static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req) GOTO(out, rc = -ENODEV); /* RPC ref */ - class_export_rpc_get(reqcopy->rq_export); + class_export_rpc_inc(reqcopy->rq_export); if (reqcopy->rq_export->exp_obd && reqcopy->rq_export->exp_obd->obd_fail) GOTO(out_put, rc = -ENODEV); @@ -1163,13 +1406,14 @@ static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req) ptlrpc_req_drop_rs(reqcopy); out_put: - class_export_rpc_put(reqcopy->rq_export); - class_export_put(reqcopy->rq_export); + class_export_rpc_dec(reqcopy->rq_export); + class_export_put(reqcopy->rq_export); out: - sptlrpc_svc_ctx_decref(reqcopy); - OBD_FREE_LARGE(reqmsg, req->rq_reqlen); - OBD_FREE(reqcopy, sizeof *reqcopy); - RETURN(rc); + sptlrpc_svc_ctx_decref(reqcopy); + OBD_FREE_LARGE(reqmsg, req->rq_reqlen); +out_free: + ptlrpc_request_cache_free(reqcopy); + RETURN(rc); } /* Send early replies to everybody expiring within at_early_margin @@ -1186,16 +1430,16 @@ 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); } @@ -1204,7 +1448,7 @@ 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); } @@ -1219,29 +1463,23 @@ static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt) cfs_list_for_each_entry_safe(rq, n, &array->paa_reqs_array[index], rq_timed_list) { - if (rq->rq_deadline <= now + at_early_margin) { - cfs_list_del_init(&rq->rq_timed_list); - /** - * ptlrpc_server_drop_request() may drop - * refcount to 0 already. Let's check this and - * don't add entry to work_list - */ - if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount))) - cfs_list_add(&rq->rq_timed_list, &work_list); - counter++; - array->paa_reqs_count[index]--; - array->paa_count--; - cfs_spin_lock(&rq->rq_lock); - rq->rq_at_linked = 0; - cfs_spin_unlock(&rq->rq_lock); - continue; - } - - /* update the earliest deadline */ - if (deadline == -1 || rq->rq_deadline < deadline) - deadline = rq->rq_deadline; + if (rq->rq_deadline > now + at_early_margin) { + /* update the earliest deadline */ + if (deadline == -1 || + rq->rq_deadline < deadline) + deadline = rq->rq_deadline; + break; + } - break; + ptlrpc_at_remove_timed(rq); + /** + * ptlrpc_server_drop_request() may drop + * refcount to 0 already. Let's check this and + * don't add entry to work_list + */ + if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount))) + cfs_list_add(&rq->rq_timed_list, &work_list); + counter++; } if (++index >= array->paa_size) @@ -1251,7 +1489,7 @@ 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); @@ -1284,134 +1522,161 @@ static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt) RETURN(1); /* return "did_something" for liblustre */ } -/** - * 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) +/* Check if we are already handling earlier incarnation of this request. + * Called under &req->rq_export->exp_rpc_lock locked */ +static int ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req) { - int rc = 0; - ENTRY; + struct ptlrpc_request *tmp = NULL; - 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 (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) || + (cfs_atomic_read(&req->rq_export->exp_rpc_count) == 0)) + return 0; - RETURN(rc); -} + /* bulk request are aborted upon reconnect, don't try to + * find a match */ + if (req->rq_bulk_write || req->rq_bulk_read) + return 0; -/** Remove the request from the export list. */ -static void ptlrpc_hpreq_fini(struct ptlrpc_request *req) -{ - ENTRY; - if (req->rq_export && req->rq_ops) { - /* refresh lock timeout again so that client has more - * room to send lock cancel RPC. */ - 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; + /* This list should not be longer than max_requests in + * flights on the client, so it is not all that long. + * Also we only hit this codepath in case of a resent + * request which makes it even more rarely hit */ + cfs_list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs, + rq_exp_list) { + /* Found duplicate one */ + if (tmp->rq_xid == req->rq_xid) + goto found; + } + cfs_list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs, + rq_exp_list) { + /* Found duplicate one */ + if (tmp->rq_xid == req->rq_xid) + goto found; + } + return 0; + +found: + DEBUG_REQ(D_HA, req, "Found duplicate req in processing\n"); + DEBUG_REQ(D_HA, tmp, "Request being processed\n"); + return -EBUSY; } /** - * 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(). + * Put the request to the export list if the request may become + * a high priority one. */ -static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service_part *svcpt, - struct ptlrpc_request *req) +static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt, + struct ptlrpc_request *req) { - ENTRY; + cfs_list_t *list; + int rc, hp = 0; - cfs_spin_lock(&req->rq_lock); - if (req->rq_hp == 0) { - int opc = lustre_msg_get_opc(req->rq_reqmsg); + ENTRY; - /* 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; + 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) { + if (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); + hp = rc; + } + list = &req->rq_export->exp_hp_rpcs; + } else { + list = &req->rq_export->exp_reg_rpcs; + } + + /* do search for duplicated xid and the adding to the list + * atomically */ + spin_lock_bh(&req->rq_export->exp_rpc_lock); + rc = ptlrpc_server_check_resend_in_progress(req); + if (rc < 0) { + spin_unlock_bh(&req->rq_export->exp_rpc_lock); + RETURN(rc); + } + cfs_list_add(&req->rq_exp_list, list); + spin_unlock_bh(&req->rq_export->exp_rpc_lock); + } + + ptlrpc_nrs_req_initialize(svcpt, req, !!hp); + + RETURN(hp); } -/** - * \see ptlrpc_hpreq_reorder_nolock - */ -void ptlrpc_hpreq_reorder(struct ptlrpc_request *req) +/** Remove the request from the export list. */ +static void ptlrpc_server_hpreq_fini(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); + if (req->rq_export) { + /* refresh lock timeout again so that client has more + * room to send lock cancel RPC. */ + if (req->rq_ops && req->rq_ops->hpreq_fini) + req->rq_ops->hpreq_fini(req); + + 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; } -/** Check if the request is a high priority one. */ -static int ptlrpc_server_hpreq_check(struct ptlrpc_service *svc, - struct ptlrpc_request *req) +static int ptlrpc_hpreq_check(struct ptlrpc_request *req) { - ENTRY; + return 1; +} + +static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = { + .hpreq_check = ptlrpc_hpreq_check, +}; + +/* Hi-Priority RPC check by RPC operation code. */ +int ptlrpc_hpreq_handler(struct ptlrpc_request *req) +{ + int opc = lustre_msg_get_opc(req->rq_reqmsg); - /* Check by request opc. */ - if (OBD_PING == lustre_msg_get_opc(req->rq_reqmsg)) - RETURN(1); + /* Check for export to let only reconnects for not yet evicted + * export to become a HP rpc. */ + if ((req->rq_export != NULL) && + (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT)) + req->rq_ops = &ptlrpc_hpreq_common; - RETURN(ptlrpc_hpreq_init(svc, req)); + return 0; } +EXPORT_SYMBOL(ptlrpc_hpreq_handler); -/** 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); } @@ -1421,24 +1686,40 @@ static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt, * 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_part *svcpt, - int force) +static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt, + bool force) { + int running = svcpt->scp_nthrs_running; + + if (!nrs_svcpt_has_hp(svcpt)) + return false; + if (force) - return 1; + return true; + + 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 >= svcpt->scp_nthrs_running - 1) - return 0; + if (svcpt->scp_nreqs_active >= running - 1) + return false; + + if (svcpt->scp_nhreqs_active == 0) + return true; - return cfs_list_empty(&svcpt->scp_req_pending) || + return !ptlrpc_nrs_req_pending_nolock(svcpt, false) || svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio; } -static int ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt, - int force) +static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt, + bool force) { return ptlrpc_server_allow_high(svcpt, force) && - !cfs_list_empty(&svcpt->scp_hreq_pending); + ptlrpc_nrs_req_pending_nolock(svcpt, true); } /** @@ -1450,29 +1731,37 @@ static int ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt, * 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_part *svcpt, - int force) +static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt, + bool force) { + int running = svcpt->scp_nthrs_running; #ifndef __KERNEL__ if (1) /* always allow to handle normal request for liblustre */ - return 1; + return true; #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) - return 1; + svcpt->scp_nreqs_active < running - 2) + return true; - if (svcpt->scp_nreqs_active >= svcpt->scp_nthrs_running - 1) - return 0; + if (svcpt->scp_nreqs_active >= running - 1) + return false; - return svcpt->scp_nhreqs_active > 0 || - svcpt->scp_service->srv_ops.so_hpreq_handler == NULL; + return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt); } -static int ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt, - int force) +static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt, + bool force) { return ptlrpc_server_allow_normal(svcpt, force) && - !cfs_list_empty(&svcpt->scp_req_pending); + ptlrpc_nrs_req_pending_nolock(svcpt, false); } /** @@ -1483,8 +1772,8 @@ static int ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt, * \see ptlrpc_server_allow_normal * \see ptlrpc_server_allow high */ -static inline int -ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, int force) +static inline bool +ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force) { return ptlrpc_server_high_pending(svcpt, force) || ptlrpc_server_normal_pending(svcpt, force); @@ -1496,25 +1785,50 @@ ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, int force) * Returns a pointer to fetched request. */ static struct ptlrpc_request * -ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, int force) +ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force) { - struct ptlrpc_request *req; + struct ptlrpc_request *req = NULL; ENTRY; + spin_lock(&svcpt->scp_req_lock); +#ifndef __KERNEL__ + /* !@%$# liblustre only has 1 thread */ + if (cfs_atomic_read(&svcpt->scp_nreps_difficult) != 0) { + spin_unlock(&svcpt->scp_req_lock); + RETURN(NULL); + } +#endif + 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); + req = ptlrpc_nrs_req_get_nolock(svcpt, true, force); + if (req != NULL) { + svcpt->scp_hreq_count++; + goto got_request; + } } 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); + req = ptlrpc_nrs_req_get_nolock(svcpt, false, force); + if (req != NULL) { + svcpt->scp_hreq_count = 0; + goto got_request; + } } + + spin_unlock(&svcpt->scp_req_lock); RETURN(NULL); + +got_request: + svcpt->scp_nreqs_active++; + if (req->rq_hp) + svcpt->scp_nhreqs_active++; + + spin_unlock(&svcpt->scp_req_lock); + + if (likely(req->rq_export)) + class_export_rpc_inc(req->rq_export); + + RETURN(req); } /** @@ -1524,7 +1838,8 @@ ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, int force) * ptlrpc_server_handle_req later on. */ static int -ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt) +ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt, + struct ptlrpc_thread *thread) { struct ptlrpc_service *svc = svcpt->scp_service; struct ptlrpc_request *req; @@ -1532,9 +1847,9 @@ 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); } @@ -1544,7 +1859,7 @@ 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); @@ -1644,21 +1959,33 @@ ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt) goto err_req; } - ptlrpc_at_add_timed(req); + req->rq_svc_thread = thread; + if (thread != NULL) { + /* initialize request session, it is needed for request + * processing by target */ + rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION | + LCT_NOREF); + if (rc) { + CERROR("%s: failure to initialize session: rc = %d\n", + thread->t_name, rc); + goto err_req; + } + req->rq_session.lc_thread = thread; + lu_context_enter(&req->rq_session); + req->rq_svc_thread->t_env->le_ses = &req->rq_session; + } + + ptlrpc_at_add_timed(req); - /* Move it over to the request processing queue */ + /* Move it over to the request processing queue */ rc = ptlrpc_server_request_add(svcpt, req); - if (rc) { - ptlrpc_hpreq_fini(req); + if (rc) GOTO(err_req, rc); - } - cfs_waitq_signal(&svcpt->scp_waitq); + + wake_up(&svcpt->scp_waitq); RETURN(1); err_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); @@ -1672,29 +1999,18 @@ static int 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; - struct timeval work_end; - long timediff; - int rc; - int fail_opc = 0; - ENTRY; + struct ptlrpc_service *svc = svcpt->scp_service; + struct ptlrpc_request *request; + struct timeval work_start; + struct timeval work_end; + long timediff; + int fail_opc = 0; - cfs_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); + ENTRY; + + request = ptlrpc_server_request_get(svcpt, false); + if (request == NULL) RETURN(0); - } -#endif - request = ptlrpc_server_request_get(svcpt, 0); - if (request == NULL) { - cfs_spin_unlock(&svcpt->scp_req_lock); - RETURN(0); - } if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT)) fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT; @@ -1702,35 +2018,18 @@ ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt, fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT; if (unlikely(fail_opc)) { - if (request->rq_export && request->rq_ops) { - cfs_spin_unlock(&svcpt->scp_req_lock); - + if (request->rq_export && request->rq_ops) OBD_FAIL_TIMEOUT(fail_opc, 4); - - 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(&svcpt->scp_req_lock); - ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET); - if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG)) - libcfs_debug_dumplog(); + if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG)) + libcfs_debug_dumplog(); - cfs_gettimeofday(&work_start); - timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL); - if (likely(svc->srv_stats != NULL)) { + do_gettimeofday(&work_start); + timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL); + if (likely(svc->srv_stats != NULL)) { lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR, timediff); lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR, @@ -1741,26 +2040,10 @@ ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt, at_get(&svcpt->scp_at_estimate)); } - rc = lu_context_init(&request->rq_session, LCT_SESSION | LCT_NOREF); - if (rc) { - CERROR("Failure to initialize session: %d\n", rc); - goto out_req; - } - request->rq_session.lc_thread = thread; - request->rq_session.lc_cookie = 0x5; - lu_context_enter(&request->rq_session); - - CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid); - - request->rq_svc_thread = thread; - if (thread) - 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 (likely(request->rq_export)) { + 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. @@ -1773,51 +2056,55 @@ 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 " - "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(), - (request->rq_export ? - (char *)request->rq_export->exp_client_uuid.uuid : "0"), - (request->rq_export ? - cfs_atomic_read(&request->rq_export->exp_refcount) : -99), - lustre_msg_get_status(request->rq_reqmsg), request->rq_xid, - libcfs_id2str(request->rq_peer), - lustre_msg_get_opc(request->rq_reqmsg)); + CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc " + "%s:%s+%d:%d:x"LPU64":%s:%d\n", current_comm(), + (request->rq_export ? + (char *)request->rq_export->exp_client_uuid.uuid : "0"), + (request->rq_export ? + cfs_atomic_read(&request->rq_export->exp_refcount) : -99), + lustre_msg_get_status(request->rq_reqmsg), request->rq_xid, + libcfs_id2str(request->rq_peer), + lustre_msg_get_opc(request->rq_reqmsg)); if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING) CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val); - rc = svc->srv_ops.so_req_handler(request); + CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid); + + /* re-assign request and sesson thread to the current one */ + request->rq_svc_thread = thread; + if (thread != NULL) { + LASSERT(request->rq_session.lc_thread != NULL); + request->rq_session.lc_thread = thread; + request->rq_session.lc_cookie = 0x55; + thread->t_env->le_ses = &request->rq_session; + } + svc->srv_ops.so_req_handler(request); - ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE); + 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 " - "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in " - "%ldus (%ldus total) trans "LPU64" rc %d/%d\n", - cfs_curproc_comm(), - (request->rq_export ? - (char *)request->rq_export->exp_client_uuid.uuid : "0"), + do_gettimeofday(&work_end); + timediff = cfs_timeval_sub(&work_end, &work_start, NULL); + 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", + current_comm(), + (request->rq_export ? + (char *)request->rq_export->exp_client_uuid.uuid : "0"), (request->rq_export ? cfs_atomic_read(&request->rq_export->exp_refcount) : -99), lustre_msg_get_status(request->rq_reqmsg), @@ -1851,8 +2138,7 @@ put_conn: request->rq_arrival_time.tv_sec)); } -out_req: - ptlrpc_server_finish_request(svcpt, request); + ptlrpc_server_finish_active_request(svcpt, request); RETURN(1); } @@ -1876,10 +2162,10 @@ 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 @@ -1903,13 +2189,13 @@ 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; @@ -1920,46 +2206,45 @@ 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; ptlrpc_rs_decref (rs); if (cfs_atomic_dec_and_test(&svcpt->scp_nreps_difficult) && svc->srv_is_stopping) - cfs_waitq_broadcast(&svcpt->scp_waitq); + wake_up_all(&svcpt->scp_waitq); RETURN(1); } /* still on the net; callback will schedule */ - cfs_spin_unlock(&rs->rs_lock); + spin_unlock(&rs->rs_lock); RETURN(1); } @@ -1979,14 +2264,14 @@ 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); @@ -2008,7 +2293,8 @@ liblustre_check_services (void *arg) cfs_list_entry (tmp, struct ptlrpc_service, srv_list); struct ptlrpc_service_part *svcpt; - svcpt = svc->srv_part; + LASSERT(svc->srv_ncpts == 1); + svcpt = svc->srv_parts[0]; if (svcpt->scp_nthrs_running != 0) /* I've recursed */ continue; @@ -2021,7 +2307,7 @@ liblustre_check_services (void *arg) svcpt->scp_nthrs_running++; do { - rc = ptlrpc_server_handle_req_in(svcpt); + rc = ptlrpc_server_handle_req_in(svcpt, NULL); rc |= ptlrpc_server_handle_reply(svcpt); rc |= ptlrpc_at_check_timed(svcpt); rc |= ptlrpc_server_handle_request(svcpt, NULL); @@ -2053,7 +2339,7 @@ ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt) * space. */ if (avail <= low_water) - ptlrpc_grow_req_bufs(svcpt); + ptlrpc_grow_req_bufs(svcpt, 1); if (svcpt->scp_service->srv_stats) { lprocfs_counter_add(svcpt->scp_service->srv_stats, @@ -2087,7 +2373,8 @@ static inline int ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt) { return svcpt->scp_nthrs_running + - svcpt->scp_nthrs_starting < svcpt->scp_service->srv_threads_max; + svcpt->scp_nthrs_starting < + svcpt->scp_service->srv_nthrs_cpt_limit; } /** @@ -2141,12 +2428,12 @@ ptlrpc_wait_event(struct ptlrpc_service_part *svcpt, lc_watchdog_disable(thread->t_watchdog); - cfs_cond_resched(); + cond_resched(); l_wait_event_exclusive_head(svcpt->scp_waitq, ptlrpc_thread_stopping(thread) || ptlrpc_server_request_incoming(svcpt) || - ptlrpc_server_request_pending(svcpt, 0) || + ptlrpc_server_request_pending(svcpt, false) || ptlrpc_rqbd_pending(svcpt) || ptlrpc_at_check(svcpt), &lwi); @@ -2171,43 +2458,33 @@ static int ptlrpc_main(void *arg) struct ptlrpc_service *svc = svcpt->scp_service; struct ptlrpc_reply_state *rs; #ifdef WITH_GROUP_INFO - cfs_group_info_t *ginfo = NULL; + struct group_info *ginfo = NULL; #endif - struct lu_env *env; - int counter = 0, rc = 0; - ENTRY; - - thread->t_pid = cfs_curproc_pid(); - cfs_daemonize_ctxt(thread->t_name); + struct lu_env *env; + int counter = 0, rc = 0; + ENTRY; -#if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA) - /* we need to do this before any per-thread allocation is done so that - * we get the per-thread allocations on local node. bug 7342 */ - if (svc->srv_cpu_affinity) { - int cpu, num_cpu; + thread->t_pid = current_pid(); + unshare_fs_struct(); - for (cpu = 0, num_cpu = 0; cpu < cfs_num_possible_cpus(); - cpu++) { - if (!cpu_online(cpu)) - continue; - if (num_cpu == thread->t_id % cfs_num_online_cpus()) - break; - num_cpu++; - } - cfs_set_cpus_allowed(cfs_current(), - node_to_cpumask(cpu_to_node(cpu))); - } -#endif + /* NB: we will call cfs_cpt_bind() for all threads, because we + * might want to run lustre server only on a subset of system CPUs, + * in that case ->scp_cpt is CFS_CPT_ANY */ + rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt); + if (rc != 0) { + CWARN("%s: failed to bind %s on CPT %d\n", + svc->srv_name, thread->t_name, svcpt->scp_cpt); + } #ifdef WITH_GROUP_INFO - ginfo = cfs_groups_alloc(0); - if (!ginfo) { - rc = -ENOMEM; - goto out; - } + ginfo = groups_alloc(0); + if (!ginfo) { + rc = -ENOMEM; + goto out; + } - cfs_set_current_groups(ginfo); - cfs_put_group_info(ginfo); + set_current_groups(ginfo); + put_group_info(ginfo); #endif if (svc->srv_ops.so_thr_init != NULL) { @@ -2231,6 +2508,16 @@ static int ptlrpc_main(void *arg) env->le_ctx.lc_thread = thread; env->le_ctx.lc_cookie = 0x6; + while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) { + rc = ptlrpc_server_post_idle_rqbds(svcpt); + if (rc >= 0) + continue; + + CERROR("Failed to post rqbd for %s on CPT %d: %d\n", + svc->srv_name, svcpt->scp_cpt, rc); + goto out_srv_fini; + } + /* Alloc reply state structure for this one */ OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size); if (!rs) { @@ -2238,7 +2525,7 @@ 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); @@ -2252,18 +2539,18 @@ 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); + wake_up(&thread->t_ctl_waitq); 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); + wake_up(&svcpt->scp_rep_waitq); + spin_unlock(&svcpt->scp_rep_lock); CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id, svcpt->scp_nthrs_running); @@ -2282,7 +2569,11 @@ static int ptlrpc_main(void *arg) /* Process all incoming reqs before handling any */ if (ptlrpc_server_request_incoming(svcpt)) { - ptlrpc_server_handle_req_in(svcpt); + lu_context_enter(&env->le_ctx); + env->le_ses = NULL; + ptlrpc_server_handle_req_in(svcpt, thread); + lu_context_exit(&env->le_ctx); + /* but limit ourselves in case of flood */ if (counter++ < 100) continue; @@ -2292,7 +2583,7 @@ static int ptlrpc_main(void *arg) if (ptlrpc_at_check(svcpt)) ptlrpc_at_check_timed(svcpt); - if (ptlrpc_server_request_pending(svcpt, 0)) { + if (ptlrpc_server_request_pending(svcpt, false)) { lu_context_enter(&env->le_ctx); ptlrpc_server_handle_request(svcpt, thread); lu_context_exit(&env->le_ctx); @@ -2327,7 +2618,7 @@ 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--; @@ -2339,29 +2630,24 @@ out: thread->t_id = rc; thread_add_flags(thread, SVC_STOPPED); - cfs_waitq_signal(&thread->t_ctl_waitq); - cfs_spin_unlock(&svcpt->scp_lock); + wake_up(&thread->t_ctl_waitq); + spin_unlock(&svcpt->scp_lock); return rc; } -struct ptlrpc_hr_args { - int thread_index; - int cpu_index; - struct ptlrpc_hr_service *hrs; -}; - -static int hrt_dont_sleep(struct ptlrpc_hr_thread *t, - cfs_list_t *replies) +static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt, + cfs_list_t *replies) { - int result; + int result; - cfs_spin_lock(&t->hrt_lock); - cfs_list_splice_init(&t->hrt_queue, replies); - result = cfs_test_bit(HRT_STOPPING, &t->hrt_flags) || - !cfs_list_empty(replies); - cfs_spin_unlock(&t->hrt_lock); - return result; + spin_lock(&hrt->hrt_lock); + + cfs_list_splice_init(&hrt->hrt_queue, replies); + result = ptlrpc_hr.hr_stopping || !cfs_list_empty(replies); + + spin_unlock(&hrt->hrt_lock); + return result; } /** @@ -2370,26 +2656,28 @@ static int hrt_dont_sleep(struct ptlrpc_hr_thread *t, */ static int ptlrpc_hr_main(void *arg) { - struct ptlrpc_hr_args * hr_args = arg; - struct ptlrpc_hr_service *hr = hr_args->hrs; - struct ptlrpc_hr_thread *t = &hr->hr_threads[hr_args->thread_index]; - char threadname[20]; - CFS_LIST_HEAD(replies); + struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg; + struct ptlrpc_hr_partition *hrp = hrt->hrt_partition; + CFS_LIST_HEAD (replies); + char threadname[20]; + int rc; - snprintf(threadname, sizeof(threadname), - "ptlrpc_hr_%d", hr_args->thread_index); + snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d", + hrp->hrp_cpt, hrt->hrt_id); + unshare_fs_struct(); - cfs_daemonize_ctxt(threadname); -#if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK) - cfs_set_cpus_allowed(cfs_current(), - node_to_cpumask(cpu_to_node(hr_args->cpu_index))); -#endif - cfs_set_bit(HRT_RUNNING, &t->hrt_flags); - cfs_waitq_signal(&t->hrt_wait); + rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt); + if (rc != 0) { + CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n", + threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc); + } - while (!cfs_test_bit(HRT_STOPPING, &t->hrt_flags)) { + cfs_atomic_inc(&hrp->hrp_nstarted); + wake_up(&ptlrpc_hr.hr_waitq); + + while (!ptlrpc_hr.hr_stopping) { + l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies)); - l_wait_condition(t->hrt_wait, hrt_dont_sleep(t, &replies)); while (!cfs_list_empty(&replies)) { struct ptlrpc_reply_state *rs; @@ -2401,89 +2689,67 @@ static int ptlrpc_hr_main(void *arg) } } - cfs_clear_bit(HRT_RUNNING, &t->hrt_flags); - cfs_complete(&t->hrt_completion); - - return 0; -} - -static int ptlrpc_start_hr_thread(struct ptlrpc_hr_service *hr, int n, int cpu) -{ - struct ptlrpc_hr_thread *t = &hr->hr_threads[n]; - struct ptlrpc_hr_args args; - int rc; - ENTRY; + cfs_atomic_inc(&hrp->hrp_nstopped); + wake_up(&ptlrpc_hr.hr_waitq); - args.thread_index = n; - args.cpu_index = cpu; - args.hrs = hr; - - rc = cfs_create_thread(ptlrpc_hr_main, (void*)&args, CFS_DAEMON_FLAGS); - if (rc < 0) { - cfs_complete(&t->hrt_completion); - GOTO(out, rc); - } - l_wait_condition(t->hrt_wait, cfs_test_bit(HRT_RUNNING, &t->hrt_flags)); - RETURN(0); - out: - return rc; + return 0; } -static void ptlrpc_stop_hr_thread(struct ptlrpc_hr_thread *t) +static void ptlrpc_stop_hr_threads(void) { - ENTRY; + struct ptlrpc_hr_partition *hrp; + int i; + int j; - cfs_set_bit(HRT_STOPPING, &t->hrt_flags); - cfs_waitq_signal(&t->hrt_wait); - cfs_wait_for_completion(&t->hrt_completion); + ptlrpc_hr.hr_stopping = 1; - EXIT; -} - -static void ptlrpc_stop_hr_threads(struct ptlrpc_hr_service *hrs) -{ - int n; - ENTRY; - - for (n = 0; n < hrs->hr_n_threads; n++) - ptlrpc_stop_hr_thread(&hrs->hr_threads[n]); + cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) { + if (hrp->hrp_thrs == NULL) + continue; /* uninitialized */ + for (j = 0; j < hrp->hrp_nthrs; j++) + wake_up_all(&hrp->hrp_thrs[j].hrt_waitq); + } - EXIT; + cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) { + if (hrp->hrp_thrs == NULL) + continue; /* uninitialized */ + wait_event(ptlrpc_hr.hr_waitq, + cfs_atomic_read(&hrp->hrp_nstopped) == + cfs_atomic_read(&hrp->hrp_nstarted)); + } } -static int ptlrpc_start_hr_threads(struct ptlrpc_hr_service *hr) +static int ptlrpc_start_hr_threads(void) { - int rc = -ENOMEM; - int n, cpu, threads_started = 0; - ENTRY; + struct ptlrpc_hr_partition *hrp; + int i; + int j; + ENTRY; - LASSERT(hr != NULL); - LASSERT(hr->hr_n_threads > 0); + cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) { + int rc = 0; + + for (j = 0; j < hrp->hrp_nthrs; j++) { + struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j]; + rc = PTR_ERR(kthread_run(ptlrpc_hr_main, + &hrp->hrp_thrs[j], + "ptlrpc_hr%02d_%03d", + hrp->hrp_cpt, + hrt->hrt_id)); + if (IS_ERR_VALUE(rc)) + break; + } + wait_event(ptlrpc_hr.hr_waitq, + cfs_atomic_read(&hrp->hrp_nstarted) == j); + if (!IS_ERR_VALUE(rc)) + continue; - for (n = 0, cpu = 0; n < hr->hr_n_threads; n++) { -#if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK) - while (!cpu_online(cpu)) { - cpu++; - if (cpu >= cfs_num_possible_cpus()) - cpu = 0; - } -#endif - rc = ptlrpc_start_hr_thread(hr, n, cpu); - if (rc != 0) - break; - threads_started++; - cpu++; - } - if (threads_started == 0) { - CERROR("No reply handling threads started\n"); - RETURN(-ESRCH); - } - if (threads_started < hr->hr_n_threads) { - CWARN("Started only %d reply handling threads from %d\n", - threads_started, hr->hr_n_threads); - hr->hr_n_threads = threads_started; - } - RETURN(0); + CERROR("Reply handling thread %d:%d Failed on starting: " + "rc = %d\n", i, j, rc); + ptlrpc_stop_hr_threads(); + RETURN(rc); + } + RETURN(0); } static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt) @@ -2497,7 +2763,7 @@ 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", @@ -2505,7 +2771,7 @@ static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt) thread_add_flags(thread, SVC_STOPPING); } - cfs_waitq_broadcast(&svcpt->scp_waitq); + wake_up_all(&svcpt->scp_waitq); while (!cfs_list_empty(&svcpt->scp_threads)) { thread = cfs_list_entry(svcpt->scp_threads.next, @@ -2515,17 +2781,17 @@ 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, @@ -2541,51 +2807,66 @@ static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt) */ void ptlrpc_stop_all_threads(struct ptlrpc_service *svc) { + struct ptlrpc_service_part *svcpt; + int i; ENTRY; - if (svc != NULL && svc->srv_part != NULL) - ptlrpc_svcpt_stop_threads(svc->srv_part); + ptlrpc_service_for_each_part(svcpt, i, svc) { + if (svcpt->scp_service != NULL) + ptlrpc_svcpt_stop_threads(svcpt); + } + EXIT; } +EXPORT_SYMBOL(ptlrpc_stop_all_threads); int ptlrpc_start_threads(struct ptlrpc_service *svc) { - int i, rc = 0; - ENTRY; + int rc = 0; + int i; + int j; + ENTRY; - /* We require 2 threads min - see note in - ptlrpc_server_handle_request */ - LASSERT(svc->srv_threads_min >= 2); - for (i = 0; i < svc->srv_threads_min; i++) { - rc = ptlrpc_start_thread(svc->srv_part, 1); - /* We have enough threads, don't start more. b=15759 */ - if (rc == -EMFILE) { - rc = 0; - break; - } - if (rc) { - CERROR("cannot start %s thread #%d: rc %d\n", - svc->srv_thread_name, i, rc); - ptlrpc_stop_all_threads(svc); - break; - } - } - RETURN(rc); + /* We require 2 threads min, see note in ptlrpc_server_handle_request */ + LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT); + + for (i = 0; i < svc->srv_ncpts; i++) { + for (j = 0; j < svc->srv_nthrs_cpt_init; j++) { + rc = ptlrpc_start_thread(svc->srv_parts[i], 1); + if (rc == 0) + continue; + + if (rc != -EMFILE) + goto failed; + /* We have enough threads, don't start more. b=15759 */ + break; + } + } + + RETURN(0); + failed: + CERROR("cannot start %s thread #%d_%d: rc %d\n", + svc->srv_thread_name, i, j, rc); + ptlrpc_stop_all_threads(svc); + RETURN(rc); } +EXPORT_SYMBOL(ptlrpc_start_threads); 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); - 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); + 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); again: if (unlikely(svc->srv_is_stopping)) @@ -2593,17 +2874,17 @@ int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait) if (!ptlrpc_threads_increasable(svcpt) || (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) && - svcpt->scp_nthrs_running == svc->srv_threads_min - 1)) + svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1)) RETURN(-EMFILE); - OBD_ALLOC_PTR(thread); + OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt); if (thread == NULL) RETURN(-ENOMEM); - cfs_waitq_init(&thread->t_ctl_waitq); + init_waitqueue_head(&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); } @@ -2612,12 +2893,12 @@ 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", svc->srv_thread_name, svcpt->scp_thr_nextid); - cfs_schedule(); + schedule(); goto again; } @@ -2632,26 +2913,35 @@ 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); - snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, - "%s_%02d", svc->srv_thread_name, thread->t_id); + if (svcpt->scp_cpt >= 0) { + snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d", + svc->srv_thread_name, svcpt->scp_cpt, thread->t_id); + } else { + snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d", + svc->srv_thread_name, thread->t_id); + } CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name); - /* - * CLONE_VM and CLONE_FILES just avoid a needless copy, because we - * just drop the VM and FILES in cfs_daemonize_ctxt() right away. - */ - rc = cfs_create_thread(ptlrpc_main, thread, CFS_DAEMON_FLAGS); - if (rc < 0) { + rc = PTR_ERR(kthread_run(ptlrpc_main, thread, thread->t_name)); + if (IS_ERR_VALUE(rc)) { CERROR("cannot start thread '%s': rc %d\n", thread->t_name, rc); - cfs_spin_lock(&svcpt->scp_lock); - cfs_list_del(&thread->t_link); + spin_lock(&svcpt->scp_lock); --svcpt->scp_nthrs_starting; - cfs_spin_unlock(&svcpt->scp_lock); - - OBD_FREE(thread, sizeof(*thread)); + if (thread_is_stopping(thread)) { + /* this ptlrpc_thread is being hanled + * by ptlrpc_svcpt_stop_threads now + */ + thread_add_flags(thread, SVC_STOPPED); + wake_up(&thread->t_ctl_waitq); + spin_unlock(&svcpt->scp_lock); + } else { + cfs_list_del(&thread->t_link); + spin_unlock(&svcpt->scp_lock); + OBD_FREE_PTR(thread); + } RETURN(rc); } @@ -2668,46 +2958,75 @@ int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait) int ptlrpc_hr_init(void) { - int i; - int n_cpus = cfs_num_online_cpus(); - struct ptlrpc_hr_service *hr; - int size; - int rc; - ENTRY; + struct ptlrpc_hr_partition *hrp; + struct ptlrpc_hr_thread *hrt; + int rc; + int i; + int j; + ENTRY; - LASSERT(ptlrpc_hr == NULL); + memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr)); + ptlrpc_hr.hr_cpt_table = cfs_cpt_table; - size = offsetof(struct ptlrpc_hr_service, hr_threads[n_cpus]); - OBD_ALLOC(hr, size); - if (hr == NULL) - RETURN(-ENOMEM); - for (i = 0; i < n_cpus; i++) { - struct ptlrpc_hr_thread *t = &hr->hr_threads[i]; + ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table, + sizeof(*hrp)); + if (ptlrpc_hr.hr_partitions == NULL) + RETURN(-ENOMEM); - cfs_spin_lock_init(&t->hrt_lock); - cfs_waitq_init(&t->hrt_wait); - CFS_INIT_LIST_HEAD(&t->hrt_queue); - cfs_init_completion(&t->hrt_completion); - } - hr->hr_n_threads = n_cpus; - hr->hr_size = size; - ptlrpc_hr = hr; + init_waitqueue_head(&ptlrpc_hr.hr_waitq); - rc = ptlrpc_start_hr_threads(hr); - if (rc) { - OBD_FREE(hr, hr->hr_size); - ptlrpc_hr = NULL; - } - RETURN(rc); + cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) { + hrp->hrp_cpt = i; + + cfs_atomic_set(&hrp->hrp_nstarted, 0); + cfs_atomic_set(&hrp->hrp_nstopped, 0); + + hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i); + hrp->hrp_nthrs /= cfs_cpu_ht_nsiblings(0); + + LASSERT(hrp->hrp_nthrs > 0); + OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i, + hrp->hrp_nthrs * sizeof(*hrt)); + if (hrp->hrp_thrs == NULL) + GOTO(out, rc = -ENOMEM); + + for (j = 0; j < hrp->hrp_nthrs; j++) { + hrt = &hrp->hrp_thrs[j]; + + hrt->hrt_id = j; + hrt->hrt_partition = hrp; + init_waitqueue_head(&hrt->hrt_waitq); + spin_lock_init(&hrt->hrt_lock); + CFS_INIT_LIST_HEAD(&hrt->hrt_queue); + } + } + + rc = ptlrpc_start_hr_threads(); +out: + if (rc != 0) + ptlrpc_hr_fini(); + RETURN(rc); } void ptlrpc_hr_fini(void) { - if (ptlrpc_hr != NULL) { - ptlrpc_stop_hr_threads(ptlrpc_hr); - OBD_FREE(ptlrpc_hr, ptlrpc_hr->hr_size); - ptlrpc_hr = NULL; - } + struct ptlrpc_hr_partition *hrp; + int i; + + if (ptlrpc_hr.hr_partitions == NULL) + return; + + ptlrpc_stop_hr_threads(); + + cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) { + if (hrp->hrp_thrs != NULL) { + OBD_FREE(hrp->hrp_thrs, + hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0])); + } + } + + cfs_percpt_free(ptlrpc_hr.hr_partitions); + ptlrpc_hr.hr_partitions = NULL; } #endif /* __KERNEL__ */ @@ -2734,16 +3053,14 @@ static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt) static void ptlrpc_service_del_atimer(struct ptlrpc_service *svc) { - struct ptlrpc_service_part *svcpt; + struct ptlrpc_service_part *svcpt; + int i; /* early disarm AT timer... */ - do { /* iterrate over multiple partitions in the future */ - svcpt = svc->srv_part; - if (svcpt == NULL || svcpt->scp_service == NULL) - break; - - cfs_timer_disarm(&svcpt->scp_at_timer); - } while (0); + ptlrpc_service_for_each_part(svcpt, i, svc) { + if (svcpt->scp_service != NULL) + cfs_timer_disarm(&svcpt->scp_at_timer); + } } static void @@ -2753,17 +3070,17 @@ ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc) struct ptlrpc_request_buffer_desc *rqbd; struct l_wait_info lwi; int rc; + int i; - /* All history will be culled when the next request buffer is + /* All history will be culled when the next request buffer is * freed in ptlrpc_service_purge_all() */ - svc->srv_max_history_rqbds = 0; + svc->srv_hist_nrqbds_cpt_max = 0; rc = LNetClearLazyPortal(svc->srv_req_portal); LASSERT(rc == 0); - do { /* iterrate over multiple partitions in the future */ - svcpt = svc->srv_part; - if (svcpt == NULL || svcpt->scp_service == NULL) + ptlrpc_service_for_each_part(svcpt, i, svc) { + if (svcpt->scp_service == NULL) break; /* Unlink all the request buffers. This forces a 'final' @@ -2773,18 +3090,17 @@ ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc) rc = LNetMDUnlink(rqbd->rqbd_md_h); LASSERT(rc == 0 || rc == -ENOENT); } - } while (0); + } - do { /* iterrate over multiple partitions in the future */ - svcpt = svc->srv_part; - if (svcpt == NULL || svcpt->scp_service == NULL) + ptlrpc_service_for_each_part(svcpt, i, svc) { + if (svcpt->scp_service == NULL) break; /* 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 */ @@ -2798,10 +3114,10 @@ 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); - } while (0); + spin_unlock(&svcpt->scp_lock); + } } static void @@ -2811,22 +3127,21 @@ ptlrpc_service_purge_all(struct ptlrpc_service *svc) struct ptlrpc_request_buffer_desc *rqbd; struct ptlrpc_request *req; struct ptlrpc_reply_state *rs; + int i; - do { /* iterrate over multiple partitions in the future */ - /* schedule all outstanding replies to terminate them */ - svcpt = svc->srv_part; - if (svcpt == NULL || svcpt->scp_service == NULL) + ptlrpc_service_for_each_part(svcpt, i, 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 @@ -2837,16 +3152,12 @@ ptlrpc_service_purge_all(struct ptlrpc_service *svc) 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)) { - req = ptlrpc_server_request_get(svcpt, 1); - cfs_list_del(&req->rq_list); - svcpt->scp_nreqs_active++; - ptlrpc_hpreq_fini(req); - ptlrpc_server_finish_request(svcpt, req); + while (ptlrpc_server_request_pending(svcpt, true)) { + req = ptlrpc_server_request_get(svcpt, true); + ptlrpc_server_finish_active_request(svcpt, req); } LASSERT(cfs_list_empty(&svcpt->scp_rqbd_posted)); @@ -2874,7 +3185,7 @@ ptlrpc_service_purge_all(struct ptlrpc_service *svc) cfs_list_del(&rs->rs_list); OBD_FREE_LARGE(rs, svc->srv_max_reply_size); } - } while (0); + } } static void @@ -2882,10 +3193,10 @@ ptlrpc_service_free(struct ptlrpc_service *svc) { struct ptlrpc_service_part *svcpt; struct ptlrpc_at_array *array; + int i; - do { /* iterrate over multiple partitions in the future */ - svcpt = svc->srv_part; - if (svcpt == NULL || svcpt->scp_service == NULL) + ptlrpc_service_for_each_part(svcpt, i, svc) { + if (svcpt->scp_service == NULL) break; /* In case somebody rearmed this in the meantime */ @@ -2903,16 +3214,16 @@ ptlrpc_service_free(struct ptlrpc_service *svc) sizeof(__u32) * array->paa_size); array->paa_reqs_count = NULL; } - svcpt->scp_service = NULL; - } while (0); + } + + ptlrpc_service_for_each_part(svcpt, i, svc) + OBD_FREE_PTR(svcpt); - do { /* iterrate over multiple partitions in the future */ - svcpt = svc->srv_part; - if (svcpt != NULL) - OBD_FREE_PTR(svcpt); - } while (0); + if (svc->srv_cpts != NULL) + cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts); - OBD_FREE_PTR(svc); + OBD_FREE(svc, offsetof(struct ptlrpc_service, + srv_parts[svc->srv_ncpts])); } int ptlrpc_unregister_service(struct ptlrpc_service *service) @@ -2923,21 +3234,24 @@ 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); } +EXPORT_SYMBOL(ptlrpc_unregister_service); /** * Returns 0 if the service is healthy. @@ -2945,36 +3259,28 @@ int ptlrpc_unregister_service(struct ptlrpc_service *service) * Right now, it just checks to make sure that requests aren't languishing * in the queue. We'll use this health check to govern whether a node needs * to be shot, so it's intentionally non-aggressive. */ -int ptlrpc_service_health_check(struct ptlrpc_service *svc) +int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt) { - struct ptlrpc_service_part *svcpt; - struct ptlrpc_request *request; + struct ptlrpc_request *request = NULL; struct timeval right_now; long timediff; - if (svc == NULL || svc->srv_part == NULL) - return 0; + do_gettimeofday(&right_now); - cfs_gettimeofday(&right_now); + spin_lock(&svcpt->scp_req_lock); + /* How long has the next entry been waiting? */ + if (ptlrpc_server_high_pending(svcpt, true)) + request = ptlrpc_nrs_req_peek_nolock(svcpt, true); + else if (ptlrpc_server_normal_pending(svcpt, true)) + request = ptlrpc_nrs_req_peek_nolock(svcpt, false); - 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); + if (request == NULL) { + spin_unlock(&svcpt->scp_req_lock); return 0; } - /* 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); - } - 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)) { @@ -2985,3 +3291,22 @@ int ptlrpc_service_health_check(struct ptlrpc_service *svc) return 0; } + +int +ptlrpc_service_health_check(struct ptlrpc_service *svc) +{ + struct ptlrpc_service_part *svcpt; + int i; + + if (svc == NULL) + return 0; + + ptlrpc_service_for_each_part(svcpt, i, svc) { + int rc = ptlrpc_svcpt_health_check(svcpt); + + if (rc != 0) + return rc; + } + return 0; +} +EXPORT_SYMBOL(ptlrpc_service_health_check);