X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fptlrpc%2Fservice.c;h=ca2aa01675c3d13a8f674842aeaa8cd699868d2b;hp=76bcc246e70be67c36c2b32fd845e4f1f2c30d43;hb=9db3b77445d8b96f1f4a20ab03a6b8df3aec01e4;hpb=2e27af9b3bd15f1e9ffaa397375253ebf60c7d8a diff --git a/lustre/ptlrpc/service.c b/lustre/ptlrpc/service.c index 76bcc24..ca2aa01 100644 --- a/lustre/ptlrpc/service.c +++ b/lustre/ptlrpc/service.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -26,8 +24,10 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2010, 2013, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -63,104 +63,125 @@ CFS_MODULE_PARM(at_extra, "i", int, 0644, /* forward ref */ -static int ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc); - -static CFS_LIST_HEAD(ptlrpc_all_services); -spinlock_t ptlrpc_all_services_lock; +static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt); +static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req); +static void ptlrpc_at_remove_timed(struct ptlrpc_request *req); -static char * -ptlrpc_alloc_request_buffer (int size) -{ - char *ptr; - - if (size > SVC_BUF_VMALLOC_THRESHOLD) - OBD_VMALLOC(ptr, size); - else - OBD_ALLOC(ptr, size); - - return (ptr); -} - -static void -ptlrpc_free_request_buffer (char *ptr, int size) -{ - if (size > SVC_BUF_VMALLOC_THRESHOLD) - OBD_VFREE(ptr, size); - else - OBD_FREE(ptr, size); -} +/** 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 *svc) +ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt) { - struct ptlrpc_request_buffer_desc *rqbd; - - OBD_ALLOC_PTR(rqbd); - if (rqbd == NULL) - return (NULL); - - rqbd->rqbd_service = svc; - 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); - rqbd->rqbd_buffer = ptlrpc_alloc_request_buffer(svc->srv_buf_size); - - if (rqbd->rqbd_buffer == NULL) { - OBD_FREE_PTR(rqbd); - return (NULL); - } - - spin_lock(&svc->srv_lock); - list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds); - svc->srv_nbufs++; - spin_unlock(&svc->srv_lock); - - return (rqbd); + struct ptlrpc_service *svc = svcpt->scp_service; + struct ptlrpc_request_buffer_desc *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_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; + } + + spin_lock(&svcpt->scp_lock); + cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle); + svcpt->scp_nrqbds_total++; + spin_unlock(&svcpt->scp_lock); + + return rqbd; } void -ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd) +ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd) { - struct ptlrpc_service *svc = rqbd->rqbd_service; + struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt; - LASSERT (rqbd->rqbd_refcount == 0); - LASSERT (list_empty(&rqbd->rqbd_reqs)); + LASSERT(rqbd->rqbd_refcount == 0); + LASSERT(cfs_list_empty(&rqbd->rqbd_reqs)); - spin_lock(&svc->srv_lock); - list_del(&rqbd->rqbd_list); - svc->srv_nbufs--; - spin_unlock(&svc->srv_lock); + spin_lock(&svcpt->scp_lock); + cfs_list_del(&rqbd->rqbd_list); + svcpt->scp_nrqbds_total--; + spin_unlock(&svcpt->scp_lock); - ptlrpc_free_request_buffer (rqbd->rqbd_buffer, svc->srv_buf_size); - OBD_FREE_PTR(rqbd); + OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size); + OBD_FREE_PTR(rqbd); } int -ptlrpc_grow_req_bufs(struct ptlrpc_service *svc) +ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post) { + struct ptlrpc_service *svc = svcpt->scp_service; struct ptlrpc_request_buffer_desc *rqbd; + int rc = 0; int i; - CDEBUG(D_RPCTRACE, "%s: allocate %d new %d-byte reqbufs (%d/%d left)\n", - svc->srv_name, svc->srv_nbuf_per_group, svc->srv_buf_size, - svc->srv_nrqbd_receiving, svc->srv_nbufs); + 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++) { - rqbd = ptlrpc_alloc_rqbd(svc); + /* 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; + + rqbd = ptlrpc_alloc_rqbd(svcpt); if (rqbd == NULL) { - CERROR ("%s: Can't allocate request buffer\n", - svc->srv_name); - return (-ENOMEM); + CERROR("%s: Can't allocate request buffer\n", + svc->srv_name); + rc = -ENOMEM; + break; } + } - if (ptlrpc_server_post_idle_rqbds(svc) < 0) - return (-EAGAIN); - } + spin_lock(&svcpt->scp_lock); + + LASSERT(svcpt->scp_rqbd_allocating == 1); + svcpt->scp_rqbd_allocating--; + + spin_unlock(&svcpt->scp_lock); + + CDEBUG(D_RPCTRACE, + "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n", + svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted, + svcpt->scp_nrqbds_total, rc); + + try_post: + if (post && rc == 0) + rc = ptlrpc_server_post_idle_rqbds(svcpt); - return (0); + return rc; } +/** + * Part of Rep-Ack logic. + * Puts a lock and its mode into reply state assotiated to request reply. + */ void ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock, int mode, int no_ack) @@ -181,37 +202,58 @@ 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 { - spinlock_t hrt_lock; - unsigned long hrt_flags; - cfs_waitq_t hrt_wait; - struct list_head hrt_queue; - struct completion 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 { - struct list_head rsb_replies; - struct ptlrpc_service *rsb_svc; - unsigned int rsb_n_replies; + cfs_list_t rsb_replies; + unsigned int rsb_n_replies; + struct ptlrpc_service_part *rsb_svcpt; }; -/** - * 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 @@ -230,6 +272,31 @@ static void rs_batch_init(struct rs_batch *b) } /** + * Choose an hr thread to dispatch requests to. + */ +static struct ptlrpc_hr_thread * +ptlrpc_hr_select(struct ptlrpc_service_part *svcpt) +{ + 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); + + hrp = ptlrpc_hr.hr_partitions[rotor]; + } + + rotor = hrp->hrp_rotor++; + return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs]; +} + +/** * Dispatch all replies accumulated in the batch to one from * dedicated reply handling threads. * @@ -237,21 +304,18 @@ static void rs_batch_init(struct rs_batch *b) */ 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 = hr->hr_index++; - if (hr->hr_index >= hr->hr_n_threads) - hr->hr_index = 0; + hrt = ptlrpc_hr_select(b->rsb_svcpt); - spin_lock(&hr->hr_threads[idx].hrt_lock); - list_splice_init(&b->rsb_replies, - &hr->hr_threads[idx].hrt_queue); - 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; + } } /** @@ -263,24 +327,25 @@ static void rs_batch_dispatch(struct rs_batch *b) */ static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs) { - struct ptlrpc_service *svc = rs->rs_service; - - if (svc != b->rsb_svc || b->rsb_n_replies >= MAX_SCHEDULED) { - if (b->rsb_svc != NULL) { - rs_batch_dispatch(b); - spin_unlock(&b->rsb_svc->srv_lock); - } - spin_lock(&svc->srv_lock); - b->rsb_svc = svc; - } - spin_lock(&rs->rs_lock); - rs->rs_scheduled_ever = 1; - if (rs->rs_scheduled == 0) { - list_move(&rs->rs_list, &b->rsb_replies); - rs->rs_scheduled = 1; - b->rsb_n_replies++; - } - spin_unlock(&rs->rs_lock); + struct ptlrpc_service_part *svcpt = rs->rs_svcpt; + + if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) { + if (b->rsb_svcpt != NULL) { + rs_batch_dispatch(b); + spin_unlock(&b->rsb_svcpt->scp_rep_lock); + } + spin_lock(&svcpt->scp_rep_lock); + b->rsb_svcpt = svcpt; + } + 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); } /** @@ -292,10 +357,10 @@ static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs) */ static void rs_batch_fini(struct rs_batch *b) { - if (b->rsb_svc != 0) { - rs_batch_dispatch(b); - spin_unlock(&b->rsb_svc->srv_lock); - } + if (b->rsb_svcpt != NULL) { + rs_batch_dispatch(b); + spin_unlock(&b->rsb_svcpt->scp_rep_lock); + } } #define DECLARE_RS_BATCH(b) struct rs_batch b @@ -309,48 +374,52 @@ static void rs_batch_fini(struct rs_batch *b) #endif /* __KERNEL__ */ +/** + * Put reply state into a queue for processing because we received + * ACK from the client + */ 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(list_empty(&rs->rs_list)); + LASSERT(cfs_list_empty(&rs->rs_list)); - idx = hr->hr_index++; - if (hr->hr_index >= hr->hr_n_threads) - hr->hr_index = 0; - spin_lock(&hr->hr_threads[idx].hrt_lock); - list_add_tail(&rs->rs_list, &hr->hr_threads[idx].hrt_queue); - 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 - list_add_tail(&rs->rs_list, &rs->rs_service->srv_reply_queue); + cfs_list_add_tail(&rs->rs_list, &rs->rs_svcpt->scp_rep_queue); #endif } void -ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs) +ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs) { - ENTRY; - - LASSERT_SPIN_LOCKED(&rs->rs_service->srv_lock); - LASSERT_SPIN_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; - } - - rs->rs_scheduled = 1; - list_del_init(&rs->rs_list); - ptlrpc_dispatch_difficult_reply(rs); - EXIT; + ENTRY; + + 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; + } + + 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) { @@ -363,216 +432,426 @@ void ptlrpc_commit_replies(struct obd_export *exp) * to attend to complete them. */ /* CAVEAT EMPTOR: spinlock ordering!!! */ - spin_lock(&exp->exp_uncommitted_replies_lock); - list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies, - rs_obd_list) { + 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); /* VBR: per-export last_committed */ LASSERT(rs->rs_export); if (rs->rs_transno <= exp->exp_last_committed) { - list_del_init(&rs->rs_obd_list); + cfs_list_del_init(&rs->rs_obd_list); rs_batch_add(&batch, rs); } } - 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 *svc) +ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt) { - struct ptlrpc_request_buffer_desc *rqbd; - int rc; - int posted = 0; + struct ptlrpc_request_buffer_desc *rqbd; + int rc; + int posted = 0; - for (;;) { - spin_lock(&svc->srv_lock); + for (;;) { + spin_lock(&svcpt->scp_lock); - if (list_empty (&svc->srv_idle_rqbds)) { - spin_unlock(&svc->srv_lock); - return (posted); - } + if (cfs_list_empty(&svcpt->scp_rqbd_idle)) { + spin_unlock(&svcpt->scp_lock); + return posted; + } - rqbd = list_entry(svc->srv_idle_rqbds.next, - struct ptlrpc_request_buffer_desc, - rqbd_list); - list_del (&rqbd->rqbd_list); + rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next, + struct ptlrpc_request_buffer_desc, + rqbd_list); + cfs_list_del(&rqbd->rqbd_list); - /* assume we will post successfully */ - svc->srv_nrqbd_receiving++; - list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds); + /* assume we will post successfully */ + svcpt->scp_nrqbds_posted++; + cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted); - spin_unlock(&svc->srv_lock); + spin_unlock(&svcpt->scp_lock); - rc = ptlrpc_register_rqbd(rqbd); - if (rc != 0) - break; + rc = ptlrpc_register_rqbd(rqbd); + if (rc != 0) + break; - posted = 1; - } + posted = 1; + } - spin_lock(&svc->srv_lock); + spin_lock(&svcpt->scp_lock); - svc->srv_nrqbd_receiving--; - list_del(&rqbd->rqbd_list); - list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds); + svcpt->scp_nrqbds_posted--; + cfs_list_del(&rqbd->rqbd_list); + cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle); - /* Don't complain if no request buffers are posted right now; LNET - * won't drop requests because we set the portal lazy! */ + /* Don't complain if no request buffers are posted right now; LNET + * won't drop requests because we set the portal lazy! */ - spin_unlock(&svc->srv_lock); + spin_unlock(&svcpt->scp_lock); - return (-1); + return -1; } -struct ptlrpc_service *ptlrpc_init_svc_conf(struct ptlrpc_service_conf *c, - svc_handler_t h, char *name, - struct proc_dir_entry *proc_entry, - svcreq_printfn_t prntfn, - char *threadname) +static void ptlrpc_at_timer(unsigned long castmeharder) { - return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize, - c->psc_max_req_size, c->psc_max_reply_size, - c->psc_req_portal, c->psc_rep_portal, - c->psc_watchdog_factor, - h, name, proc_entry, - prntfn, c->psc_min_threads, c->psc_max_threads, - threadname, c->psc_ctx_tags, NULL); + struct ptlrpc_service_part *svcpt; + + svcpt = (struct ptlrpc_service_part *)castmeharder; + + svcpt->scp_at_check = 1; + svcpt->scp_at_checktime = cfs_time_current(); + wake_up(&svcpt->scp_waitq); } -EXPORT_SYMBOL(ptlrpc_init_svc_conf); -static void ptlrpc_at_timer(unsigned long castmeharder) +static void +ptlrpc_server_nthreads_check(struct ptlrpc_service *svc, + struct ptlrpc_service_conf *conf) { - struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder; - svc->srv_at_check = 1; - svc->srv_at_checktime = cfs_time_current(); - cfs_waitq_signal(&svc->srv_waitq); +#ifdef __KERNEL__ + struct ptlrpc_service_thr_conf *tc = &conf->psc_thr; + unsigned init; + unsigned total; + unsigned nthrs; + int weight; + + /* + * 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; + } + + 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; + } + + 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 } -/* @threadname should be 11 characters or less - 3 will be added on */ +/** + * Initialize percpt data for a service + */ +static int +ptlrpc_service_part_init(struct ptlrpc_service *svc, + 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 */ + 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); + 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 */ + spin_lock_init(&svcpt->scp_req_lock); + + /* reply states */ + 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); + init_waitqueue_head(&svcpt->scp_rep_waitq); + cfs_atomic_set(&svcpt->scp_nreps_difficult, 0); + + /* adaptive timeout */ + spin_lock_init(&svcpt->scp_at_lock); + array = &svcpt->scp_at_array; + + size = at_est2timeout(at_max); + array->paa_size = size; + array->paa_count = 0; + array->paa_deadline = -1; + + /* allocate memory for scp_at_array (ptlrpc_at_array) */ + OBD_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_CPT_ALLOC(array->paa_reqs_count, + svc->srv_cptable, cpt, sizeof(__u32) * size); + if (array->paa_reqs_count == NULL) + goto failed; + + cfs_timer_init(&svcpt->scp_at_timer, ptlrpc_at_timer, svcpt); + /* At SOW, service time should be quick; 10s seems generous. If client + * timeout is less than this, we'll be sending an early reply. */ + at_init(&svcpt->scp_at_estimate, 10, 0); + + /* assign this before call ptlrpc_grow_req_bufs */ + svcpt->scp_service = svc; + /* Now allocate the request buffers, but don't post them now */ + rc = ptlrpc_grow_req_bufs(svcpt, 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) + goto failed; + + return 0; + + failed: + if (array->paa_reqs_count != NULL) { + OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size); + array->paa_reqs_count = NULL; + } + + if (array->paa_reqs_array != NULL) { + OBD_FREE(array->paa_reqs_array, + sizeof(cfs_list_t) * array->paa_size); + array->paa_reqs_array = NULL; + } + + return -ENOMEM; +} + +/** + * Initialize service on a given portal. + * This includes starting serving threads , allocating and posting rqbds and + * so on. + */ struct ptlrpc_service * -ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size, - int req_portal, int rep_portal, int watchdog_factor, - svc_handler_t handler, char *name, - cfs_proc_dir_entry_t *proc_entry, - svcreq_printfn_t svcreq_printfn, - int min_threads, int max_threads, - char *threadname, __u32 ctx_tags, - svc_hpreq_handler_t hp_handler) -{ - int rc; - struct ptlrpc_at_array *array; - struct ptlrpc_service *service; - unsigned int size, index; - ENTRY; +ptlrpc_register_service(struct ptlrpc_service_conf *conf, + 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); + LASSERT(conf->psc_buf.bc_buf_size >= + conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD); + LASSERT(conf->psc_thr.tc_ctx_tags != 0); + + 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 */ + spin_lock_init(&service->srv_lock); + service->srv_name = conf->psc_name; + service->srv_watchdog_factor = conf->psc_watchdog_factor; + CFS_INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */ + + /* buffer configuration */ + service->srv_nbuf_per_group = test_req_buffer_pressure ? + 1 : conf->psc_buf.bc_nbufs; + service->srv_max_req_size = conf->psc_buf.bc_req_max_size + + SPTLRPC_MAX_PAYLOAD; + service->srv_buf_size = conf->psc_buf.bc_buf_size; + service->srv_rep_portal = conf->psc_buf.bc_rep_portal; + service->srv_req_portal = conf->psc_buf.bc_req_portal; + + /* Increase max reply size to next power of two */ + service->srv_max_reply_size = 1; + while (service->srv_max_reply_size < + conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD) + service->srv_max_reply_size <<= 1; + + service->srv_thread_name = conf->psc_thr.tc_thr_name; + service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags; + service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO; + service->srv_ops = conf->psc_ops; + + for (i = 0; i < ncpts; i++) { + if (!conf->psc_thr.tc_cpu_affinity) + cpt = CFS_CPT_ANY; + else + cpt = cpts != NULL ? cpts[i] : i; + + 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); + + 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); + + 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); - LASSERT (nbufs > 0); - LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD); - LASSERT (ctx_tags != 0); - - OBD_ALLOC_PTR(service); - if (service == NULL) - RETURN(NULL); - - /* First initialise enough for early teardown */ - - service->srv_name = name; - spin_lock_init(&service->srv_lock); - CFS_INIT_LIST_HEAD(&service->srv_threads); - cfs_waitq_init(&service->srv_waitq); - - service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs; - service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD; - service->srv_buf_size = bufsize; - service->srv_rep_portal = rep_portal; - service->srv_req_portal = req_portal; - service->srv_watchdog_factor = watchdog_factor; - service->srv_handler = handler; - service->srv_request_history_print_fn = svcreq_printfn; - service->srv_request_seq = 1; /* valid seq #s start at 1 */ - service->srv_request_max_cull_seq = 0; - service->srv_threads_min = min_threads; - service->srv_threads_max = max_threads; - service->srv_thread_name = threadname; - service->srv_ctx_tags = ctx_tags; - service->srv_hpreq_handler = hp_handler; - service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO; - service->srv_hpreq_count = 0; - service->srv_n_hpreq = 0; - - rc = LNetSetLazyPortal(service->srv_req_portal); - LASSERT (rc == 0); - - CFS_INIT_LIST_HEAD(&service->srv_request_queue); - CFS_INIT_LIST_HEAD(&service->srv_request_hpq); - CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds); - CFS_INIT_LIST_HEAD(&service->srv_active_rqbds); - CFS_INIT_LIST_HEAD(&service->srv_history_rqbds); - CFS_INIT_LIST_HEAD(&service->srv_request_history); - CFS_INIT_LIST_HEAD(&service->srv_active_replies); -#ifndef __KERNEL__ - CFS_INIT_LIST_HEAD(&service->srv_reply_queue); +#ifdef __KERNEL__ + rc = ptlrpc_start_threads(service); + if (rc != 0) { + CERROR("Failed to start threads for service %s: %d\n", + service->srv_name, rc); + GOTO(failed, rc); + } #endif - CFS_INIT_LIST_HEAD(&service->srv_free_rs_list); - cfs_waitq_init(&service->srv_free_rs_waitq); - atomic_set(&service->srv_n_difficult_replies, 0); - - spin_lock_init(&service->srv_at_lock); - CFS_INIT_LIST_HEAD(&service->srv_req_in_queue); - - array = &service->srv_at_array; - size = at_est2timeout(at_max); - array->paa_size = size; - array->paa_count = 0; - array->paa_deadline = -1; - - /* allocate memory for srv_at_array (ptlrpc_at_array) */ - OBD_ALLOC(array->paa_reqs_array, sizeof(struct list_head) * size); - if (array->paa_reqs_array == NULL) - GOTO(failed, NULL); - - for (index = 0; index < size; index++) - CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]); - - OBD_ALLOC(array->paa_reqs_count, sizeof(__u32) * size); - if (array->paa_reqs_count == NULL) - GOTO(failed, NULL); - - cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service); - /* At SOW, service time should be quick; 10s seems generous. If client - timeout is less than this, we'll be sending an early reply. */ - at_init(&service->srv_at_estimate, 10, 0); - - spin_lock (&ptlrpc_all_services_lock); - list_add (&service->srv_list, &ptlrpc_all_services); - spin_unlock (&ptlrpc_all_services_lock); - - /* Now allocate the request buffers */ - rc = ptlrpc_grow_req_bufs(service); - /* We shouldn't be under memory pressure at startup, so - * fail if we can't post all our buffers at this time. */ - if (rc != 0) - GOTO(failed, NULL); - - /* Now allocate pool of reply buffers */ - /* Increase max reply size to next power of two */ - service->srv_max_reply_size = 1; - while (service->srv_max_reply_size < - max_reply_size + SPTLRPC_MAX_PAYLOAD) - service->srv_max_reply_size <<= 1; - - if (proc_entry != NULL) - ptlrpc_lprocfs_register_service(proc_entry, service); - - CDEBUG(D_NET, "%s: Started, listening on portal %d\n", - service->srv_name, service->srv_req_portal); - - RETURN(service); + + RETURN(service); failed: - ptlrpc_unregister_service(service); - return NULL; + 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. @@ -580,8 +859,8 @@ failed: */ static void ptlrpc_server_free_request(struct ptlrpc_request *req) { - LASSERT(atomic_read(&req->rq_refcount) == 0); - LASSERT(list_empty(&req->rq_timed_list)); + LASSERT(cfs_atomic_read(&req->rq_refcount) == 0); + LASSERT(cfs_list_empty(&req->rq_timed_list)); /* DEBUG_REQ() assumes the reply state of a request with a valid * ref will not be destroyed until that reference is dropped. */ @@ -590,130 +869,197 @@ 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); + } } /** * drop a reference count of the request. if it reaches 0, we either * put it into history list, or free it immediately. */ -static void ptlrpc_server_drop_request(struct ptlrpc_request *req) +void ptlrpc_server_drop_request(struct ptlrpc_request *req) { - struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd; - struct ptlrpc_service *svc = rqbd->rqbd_service; + struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd; + struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt; + struct ptlrpc_service *svc = svcpt->scp_service; int refcount; - struct list_head *tmp; - struct list_head *nxt; + cfs_list_t *tmp; + cfs_list_t *nxt; - if (!atomic_dec_and_test(&req->rq_refcount)) + if (!cfs_atomic_dec_and_test(&req->rq_refcount)) return; - spin_lock(&svc->srv_lock); + if (req->rq_at_linked) { + 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); + } - svc->srv_n_active_reqs--; - list_add(&req->rq_list, &rqbd->rqbd_reqs); + LASSERT(cfs_list_empty(&req->rq_timed_list)); + + /* finalize request */ + if (req->rq_export) { + class_export_put(req->rq_export); + req->rq_export = NULL; + } + + spin_lock(&svcpt->scp_lock); + + cfs_list_add(&req->rq_list, &rqbd->rqbd_reqs); refcount = --(rqbd->rqbd_refcount); if (refcount == 0) { /* request buffer is now idle: add to history */ - list_del(&rqbd->rqbd_list); - list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds); - svc->srv_n_history_rqbds++; + cfs_list_del(&rqbd->rqbd_list); + + cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds); + svcpt->scp_hist_nrqbds++; - /* cull some history? - * I expect only about 1 or 2 rqbds need to be recycled here */ - while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) { - rqbd = list_entry(svc->srv_history_rqbds.next, - struct ptlrpc_request_buffer_desc, - rqbd_list); + /* cull some history? + * I expect only about 1 or 2 rqbds need to be recycled here */ + 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); - list_del(&rqbd->rqbd_list); - svc->srv_n_history_rqbds--; + cfs_list_del(&rqbd->rqbd_list); + svcpt->scp_hist_nrqbds--; /* remove rqbd's reqs from svc's req history while * I've got the service lock */ - list_for_each(tmp, &rqbd->rqbd_reqs) { - req = list_entry(tmp, struct ptlrpc_request, - rq_list); + cfs_list_for_each(tmp, &rqbd->rqbd_reqs) { + req = cfs_list_entry(tmp, struct ptlrpc_request, + rq_list); /* Track the highest culled req seq */ - if (req->rq_history_seq > - svc->srv_request_max_cull_seq) - svc->srv_request_max_cull_seq = - req->rq_history_seq; - list_del(&req->rq_history_list); - } - - spin_unlock(&svc->srv_lock); - - list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) { - req = list_entry(rqbd->rqbd_reqs.next, - struct ptlrpc_request, - rq_list); - list_del(&req->rq_list); + if (req->rq_history_seq > + svcpt->scp_hist_seq_culled) { + svcpt->scp_hist_seq_culled = + req->rq_history_seq; + } + cfs_list_del(&req->rq_history_list); + } + + spin_unlock(&svcpt->scp_lock); + + cfs_list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) { + req = cfs_list_entry(rqbd->rqbd_reqs.next, + struct ptlrpc_request, + rq_list); + cfs_list_del(&req->rq_list); ptlrpc_server_free_request(req); } - spin_lock(&svc->srv_lock); - /* - * now all reqs including the embedded req has been - * disposed, schedule request buffer for re-use. - */ - LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0); - list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds); - } - - spin_unlock(&svc->srv_lock); - } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) { - /* If we are low on memory, we are not interested in history */ - list_del(&req->rq_list); - list_del_init(&req->rq_history_list); - spin_unlock(&svc->srv_lock); + spin_lock(&svcpt->scp_lock); + /* + * now all reqs including the embedded req has been + * disposed, schedule request buffer for re-use. + */ + LASSERT(cfs_atomic_read(&rqbd->rqbd_req.rq_refcount) == + 0); + cfs_list_add_tail(&rqbd->rqbd_list, + &svcpt->scp_rqbd_idle); + } + + 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); + + /* 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 { + spin_unlock(&svcpt->scp_lock); + } +} - ptlrpc_server_free_request(req); - } else { - spin_unlock(&svc->srv_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_request *req) +static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt, + struct ptlrpc_request *req) { - struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service; + ptlrpc_server_hpreq_fini(req); - if (req->rq_export) { - class_export_put(req->rq_export); - req->rq_export = NULL; - } + if (req->rq_session.lc_thread != NULL) { + lu_context_exit(&req->rq_session); + lu_context_fini(&req->rq_session); + } - if (req->rq_phase != RQ_PHASE_NEW) /* incorrect message magic */ - DEBUG_REQ(D_INFO, req, "free req"); + ptlrpc_server_drop_request(req); +} - spin_lock(&svc->srv_at_lock); - req->rq_sent_final = 1; - list_del_init(&req->rq_timed_list); - if (req->rq_at_linked) { - struct ptlrpc_at_array *array = &svc->srv_at_array; - __u32 index = req->rq_at_index; +/** + * 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--; + spin_unlock(&svcpt->scp_req_lock); - req->rq_at_linked = 0; - array->paa_reqs_count[index]--; - array->paa_count--; - } - spin_unlock(&svc->srv_at_lock); + ptlrpc_nrs_req_finalize(req); - ptlrpc_server_drop_request(req); + if (req->rq_export != NULL) + class_export_rpc_dec(req->rq_export); + + ptlrpc_server_finish_request(svcpt, req); } -/* This function makes sure dead exports are evicted in a timely manner. - This function is only called when some export receives a message (i.e., - the network is up.) */ +/** + * This function makes sure dead exports are evicted in a timely manner. + * This function is only called when some export receives a message (i.e., + * the network is up.) + */ static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay) { struct obd_export *oldest_exp; @@ -742,21 +1088,21 @@ 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 */ - spin_lock(&exp->exp_obd->obd_dev_lock); + spin_lock(&exp->exp_obd->obd_dev_lock); - if (list_empty(&exp->exp_obd_chain_timed)) { - /* this one is not timed */ - 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; } - list_move_tail(&exp->exp_obd_chain_timed, - &exp->exp_obd->obd_exports_timed); + cfs_list_move_tail(&exp->exp_obd_chain_timed, + &exp->exp_obd->obd_exports_timed); - oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next, - struct obd_export, exp_obd_chain_timed); + 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; - 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 */ @@ -776,8 +1122,8 @@ static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay) exp->exp_obd->obd_eviction_timer = cfs_time_current_sec() + 3 * PING_INTERVAL; CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n", - exp->exp_obd->obd_name, obd_export_nid2str(exp), - oldest_time); + exp->exp_obd->obd_name, + obd_export_nid2str(oldest_exp), oldest_time); } } else { if (cfs_time_current_sec() > @@ -793,61 +1139,81 @@ static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay) EXIT; } +/** + * Sanity check request \a req. + * Return 0 if all is ok, error code otherwise. + */ static int ptlrpc_check_req(struct ptlrpc_request *req) { - 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; - } + int rc = 0; + + 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 error response instead. */ CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n", req, req->rq_export->exp_obd->obd_name); - req->rq_status = -ENODEV; + rc = -ENODEV; + } else if (lustre_msg_get_flags(req->rq_reqmsg) & + (MSG_REPLAY | MSG_REQ_REPLAY_DONE) && + !(req->rq_export->exp_obd->obd_recovering)) { + DEBUG_REQ(D_ERROR, req, + "Invalid replay without recovery"); + class_fail_export(req->rq_export); + rc = -ENODEV; + } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 && + !(req->rq_export->exp_obd->obd_recovering)) { + DEBUG_REQ(D_ERROR, req, "Invalid req with transno " + LPU64" without recovery", + lustre_msg_get_transno(req->rq_reqmsg)); + class_fail_export(req->rq_export); + rc = -ENODEV; + } + + if (unlikely(rc < 0)) { + req->rq_status = rc; ptlrpc_error(req); - return -ENODEV; } - - return 0; + return rc; } -static void ptlrpc_at_set_timer(struct ptlrpc_service *svc) +static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt) { - struct ptlrpc_at_array *array = &svc->srv_at_array; - __s32 next; - - spin_lock(&svc->srv_at_lock); - if (array->paa_count == 0) { - cfs_timer_disarm(&svc->srv_at_timer); - spin_unlock(&svc->srv_at_lock); - return; - } - - /* Set timer for closest deadline */ - next = (__s32)(array->paa_deadline - cfs_time_current_sec() - - at_early_margin); - if (next <= 0) - ptlrpc_at_timer((unsigned long)svc); - else - cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next)); - spin_unlock(&svc->srv_at_lock); - CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next); + struct ptlrpc_at_array *array = &svcpt->scp_at_array; + __s32 next; + + if (array->paa_count == 0) { + cfs_timer_disarm(&svcpt->scp_at_timer); + return; + } + + /* Set timer for closest deadline */ + next = (__s32)(array->paa_deadline - cfs_time_current_sec() - + at_early_margin); + if (next <= 0) { + ptlrpc_at_timer((unsigned long)svcpt); + } else { + cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next)); + CDEBUG(D_INFO, "armed %s at %+ds\n", + svcpt->scp_service->srv_name, next); + } } /* Add rpc to early reply check list */ static int ptlrpc_at_add_timed(struct ptlrpc_request *req) { - struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service; + struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt; + struct ptlrpc_at_array *array = &svcpt->scp_at_array; struct ptlrpc_request *rq = NULL; - struct ptlrpc_at_array *array = &svc->srv_at_array; __u32 index; - int found = 0; if (AT_OFF) return(0); @@ -858,53 +1224,66 @@ static int ptlrpc_at_add_timed(struct ptlrpc_request *req) if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0) return(-ENOSYS); - spin_lock(&svc->srv_at_lock); - - if (unlikely(req->rq_sent_final)) { - spin_unlock(&svc->srv_at_lock); - return 0; - } - - LASSERT(list_empty(&req->rq_timed_list)); + spin_lock(&svcpt->scp_at_lock); + LASSERT(cfs_list_empty(&req->rq_timed_list)); index = (unsigned long)req->rq_deadline % array->paa_size; if (array->paa_reqs_count[index] > 0) { /* latest rpcs will have the latest deadlines in the list, * so search backward. */ - list_for_each_entry_reverse(rq, &array->paa_reqs_array[index], - rq_timed_list) { + cfs_list_for_each_entry_reverse(rq, + &array->paa_reqs_array[index], + rq_timed_list) { if (req->rq_deadline >= rq->rq_deadline) { - list_add(&req->rq_timed_list, - &rq->rq_timed_list); + cfs_list_add(&req->rq_timed_list, + &rq->rq_timed_list); break; } } } /* Add the request at the head of the list */ - if (list_empty(&req->rq_timed_list)) - list_add(&req->rq_timed_list, &array->paa_reqs_array[index]); - - req->rq_at_linked = 1; - 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; - found = 1; - } - spin_unlock(&svc->srv_at_lock); + if (cfs_list_empty(&req->rq_timed_list)) + cfs_list_add(&req->rq_timed_list, + &array->paa_reqs_array[index]); + + 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); + } + 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; - if (found) - ptlrpc_at_set_timer(svc); + /* 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); - return 0; + 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, - int extra_time) +static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req) { - struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service; + struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt; struct ptlrpc_request *reqcopy; struct lustre_msg *reqmsg; cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec(); @@ -917,8 +1296,8 @@ static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req, DEBUG_REQ(D_ADAPTTO, req, "%ssending early reply (deadline %+lds, margin %+lds) for " "%d+%d", AT_OFF ? "AT off - not " : "", - olddl, olddl - at_get(&svc->srv_at_estimate), - at_get(&svc->srv_at_estimate), extra_time); + olddl, olddl - at_get(&svcpt->scp_at_estimate), + at_get(&svcpt->scp_at_estimate), at_extra); if (AT_OFF) RETURN(0); @@ -938,40 +1317,44 @@ static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req, RETURN(-ENOSYS); } - if (req->rq_export && req->rq_export->exp_in_recovery) { - /* don't increase server estimates during recovery, and give - clients the full recovery time. */ - newdl = cfs_time_current_sec() + - req->rq_export->exp_obd->obd_recovery_timeout; - } else { - if (extra_time) { - /* Fake our processing time into the future to ask the - clients for some extra amount of time */ - extra_time += cfs_time_current_sec() - - req->rq_arrival_time.tv_sec; - at_add(&svc->srv_at_estimate, extra_time); - } - newdl = req->rq_arrival_time.tv_sec + - at_get(&svc->srv_at_estimate); - } - if (req->rq_deadline >= newdl) { - /* We're not adding any time, no need to send an early reply - (e.g. maybe at adaptive_max) */ - DEBUG_REQ(D_WARNING, req, "Couldn't add any time (" - CFS_DURATION_T"/"CFS_DURATION_T"), " - "not sending early reply\n", olddl, - cfs_time_sub(newdl, cfs_time_current_sec())); - RETURN(-ETIMEDOUT); - } - - OBD_ALLOC(reqcopy, sizeof *reqcopy); - if (reqcopy == NULL) - RETURN(-ENOMEM); - OBD_ALLOC(reqmsg, req->rq_reqlen); - if (!reqmsg) { - OBD_FREE(reqcopy, sizeof *reqcopy); - RETURN(-ENOMEM); - } + if (req->rq_export && + lustre_msg_get_flags(req->rq_reqmsg) & + (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) { + /* During recovery, we don't want to send too many early + * replies, but on the other hand we want to make sure the + * client has enough time to resend if the rpc is lost. So + * during the recovery period send at least 4 early replies, + * spacing them every at_extra if we can. at_estimate should + * always equal this fixed value during recovery. */ + at_measured(&svcpt->scp_at_estimate, min(at_extra, + req->rq_export->exp_obd->obd_recovery_timeout / 4)); + } else { + /* Fake our processing time into the future to ask the clients + * for some extra amount of time */ + at_measured(&svcpt->scp_at_estimate, at_extra + + cfs_time_current_sec() - + req->rq_arrival_time.tv_sec); + + /* Check to see if we've actually increased the deadline - + * we may be past adaptive_max */ + if (req->rq_deadline >= req->rq_arrival_time.tv_sec + + at_get(&svcpt->scp_at_estimate)) { + DEBUG_REQ(D_WARNING, req, "Couldn't add any time " + "(%ld/%ld), not sending early reply\n", + olddl, req->rq_arrival_time.tv_sec + + at_get(&svcpt->scp_at_estimate) - + cfs_time_current_sec()); + RETURN(-ETIMEDOUT); + } + } + newdl = cfs_time_current_sec() + at_get(&svcpt->scp_at_estimate); + + 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; @@ -984,10 +1367,12 @@ static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req, reqcopy->rq_reqmsg = reqmsg; memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen); - if (req->rq_sent_final) { + LASSERT(cfs_atomic_read(&req->rq_refcount)); + /** if it is last refcount then early reply isn't needed */ + if (cfs_atomic_read(&req->rq_refcount) == 1) { DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, " "abort sending early reply\n"); - GOTO(out, rc = 0); + GOTO(out, rc = -EINVAL); } /* Connection ref */ @@ -997,7 +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); @@ -1021,22 +1406,23 @@ 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(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 asking for at_extra time */ -static int ptlrpc_at_check_timed(struct ptlrpc_service *svc) +static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt) { + struct ptlrpc_at_array *array = &svcpt->scp_at_array; struct ptlrpc_request *rq, *n; - struct list_head work_list; - struct ptlrpc_at_array *array = &svc->srv_at_array; + cfs_list_t work_list; __u32 index, count; time_t deadline; time_t now = cfs_time_current_sec(); @@ -1044,27 +1430,27 @@ static int ptlrpc_at_check_timed(struct ptlrpc_service *svc) int first, counter = 0; ENTRY; - spin_lock(&svc->srv_at_lock); - if (svc->srv_at_check == 0) { - spin_unlock(&svc->srv_at_lock); - RETURN(0); - } - delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime); - svc->srv_at_check = 0; - - if (array->paa_count == 0) { - spin_unlock(&svc->srv_at_lock); - RETURN(0); - } - - /* The timer went off, but maybe the nearest rpc already completed. */ - first = array->paa_deadline - now; - if (first > at_early_margin) { - /* We've still got plenty of time. Reset the timer. */ - spin_unlock(&svc->srv_at_lock); - ptlrpc_at_set_timer(svc); - RETURN(0); - } + spin_lock(&svcpt->scp_at_lock); + if (svcpt->scp_at_check == 0) { + 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) { + spin_unlock(&svcpt->scp_at_lock); + RETURN(0); + } + + /* The timer went off, but maybe the nearest rpc already completed. */ + first = array->paa_deadline - now; + if (first > at_early_margin) { + /* We've still got plenty of time. Reset the timer. */ + ptlrpc_at_set_timer(svcpt); + spin_unlock(&svcpt->scp_at_lock); + RETURN(0); + } /* We're close to a timeout, and we don't know how much longer the server will take. Send early replies to everyone expiring soon. */ @@ -1074,32 +1460,36 @@ static int ptlrpc_at_check_timed(struct ptlrpc_service *svc) count = array->paa_count; while (count > 0) { count -= array->paa_reqs_count[index]; - list_for_each_entry_safe(rq, n, &array->paa_reqs_array[index], - rq_timed_list) { - if (rq->rq_deadline <= now + at_early_margin) { - list_move(&rq->rq_timed_list, &work_list); - counter++; - array->paa_reqs_count[index]--; - array->paa_count--; - rq->rq_at_linked = 0; - continue; - } - - /* update the earliest deadline */ - if (deadline == -1 || rq->rq_deadline < deadline) - deadline = rq->rq_deadline; - - break; + cfs_list_for_each_entry_safe(rq, n, + &array->paa_reqs_array[index], + rq_timed_list) { + if (rq->rq_deadline > now + at_early_margin) { + /* update the earliest deadline */ + if (deadline == -1 || + rq->rq_deadline < deadline) + deadline = rq->rq_deadline; + 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) index = 0; } array->paa_deadline = deadline; - spin_unlock(&svc->srv_at_lock); + /* we have a new earliest deadline, restart the timer */ + ptlrpc_at_set_timer(svcpt); - /* we have a new earliest deadline, restart the timer */ - ptlrpc_at_set_timer(svc); + 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); @@ -1107,221 +1497,369 @@ static int ptlrpc_at_check_timed(struct ptlrpc_service *svc) /* We're already past request deadlines before we even get a chance to send early replies */ LCONSOLE_WARN("%s: This server is not able to keep up with " - "request traffic (cpu-bound).\n", svc->srv_name); - CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, " - "delay="CFS_DURATION_T"(jiff)\n", - counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs, - at_get(&svc->srv_at_estimate), delay); - } - - /* ptlrpc_server_finish_request may delete an entry out of - * the work list */ - spin_lock(&svc->srv_at_lock); - while (!list_empty(&work_list)) { - rq = list_entry(work_list.next, struct ptlrpc_request, - rq_timed_list); - list_del_init(&rq->rq_timed_list); - /* if the entry is still in the worklist, it hasn't been - deleted, and is safe to take a ref to keep the req around */ - atomic_inc(&rq->rq_refcount); - spin_unlock(&svc->srv_at_lock); - - if (ptlrpc_at_send_early_reply(rq, at_extra) == 0) + "request traffic (cpu-bound).\n", + svcpt->scp_service->srv_name); + CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, " + "delay="CFS_DURATION_T"(jiff)\n", + counter, svcpt->scp_nreqs_incoming, + svcpt->scp_nreqs_active, + at_get(&svcpt->scp_at_estimate), delay); + } + + /* we took additional refcount so entries can't be deleted from list, no + * locking is needed */ + while (!cfs_list_empty(&work_list)) { + rq = cfs_list_entry(work_list.next, struct ptlrpc_request, + rq_timed_list); + cfs_list_del_init(&rq->rq_timed_list); + + if (ptlrpc_at_send_early_reply(rq) == 0) ptlrpc_at_add_timed(rq); ptlrpc_server_drop_request(rq); - spin_lock(&svc->srv_at_lock); } - spin_unlock(&svc->srv_at_lock); - RETURN(0); + RETURN(1); /* return "did_something" for liblustre */ +} + +/* 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) +{ + struct ptlrpc_request *tmp = NULL; + + if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) || + (cfs_atomic_read(&req->rq_export->exp_rpc_count) == 0)) + return 0; + + /* bulk request are aborted upon reconnect, don't try to + * find a match */ + if (req->rq_bulk_write || req->rq_bulk_read) + return 0; + + /* 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; } /** * Put the request to the export list if the request may become * a high priority one. */ -static int ptlrpc_hpreq_init(struct ptlrpc_service *svc, - struct ptlrpc_request *req) +static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt, + struct ptlrpc_request *req) { - int rc; - ENTRY; - - if (svc->srv_hpreq_handler) { - rc = svc->srv_hpreq_handler(req); - if (rc) - RETURN(rc); - } - if (req->rq_export && req->rq_ops) { - spin_lock(&req->rq_export->exp_lock); - list_add(&req->rq_exp_list, &req->rq_export->exp_queued_rpc); - spin_unlock(&req->rq_export->exp_lock); - } - - RETURN(0); + cfs_list_t *list; + int rc, hp = 0; + + ENTRY; + + 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); } /** Remove the request from the export list. */ -static void ptlrpc_hpreq_fini(struct ptlrpc_request *req) +static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req) { - ENTRY; - if (req->rq_export && req->rq_ops) { - spin_lock(&req->rq_export->exp_lock); - list_del_init(&req->rq_exp_list); - spin_unlock(&req->rq_export->exp_lock); - } - EXIT; + ENTRY; + 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; } -/** - * Make the request a high priority one. - * - * All the high priority requests are queued in a separate FIFO - * ptlrpc_service::srv_request_hpq list which is parallel to - * ptlrpc_service::srv_request_queue list but has a higher priority - * for handling. - * - * \see ptlrpc_server_handle_request(). - */ -static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc, - struct ptlrpc_request *req) +static int ptlrpc_hpreq_check(struct ptlrpc_request *req) { - ENTRY; - LASSERT(svc != NULL); - spin_lock(&req->rq_lock); - if (req->rq_hp == 0) { - int opc = lustre_msg_get_opc(req->rq_reqmsg); - - /* Add to the high priority queue. */ - list_move_tail(&req->rq_list, &svc->srv_request_hpq); - req->rq_hp = 1; - if (opc != OBD_PING) - DEBUG_REQ(D_NET, req, "high priority req"); - } - spin_unlock(&req->rq_lock); - EXIT; + return 1; } -void ptlrpc_hpreq_reorder(struct ptlrpc_request *req) +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) { - struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service; - ENTRY; + int opc = lustre_msg_get_opc(req->rq_reqmsg); - spin_lock(&svc->srv_lock); - /* It may happen that the request is already taken for the processing - * but still in the export list, do not re-add it into the HP list. */ - if (req->rq_phase == RQ_PHASE_NEW) - ptlrpc_hpreq_reorder_nolock(svc, req); - spin_unlock(&svc->srv_lock); - EXIT; + /* 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 0; } +EXPORT_SYMBOL(ptlrpc_hpreq_handler); -/** Check if the request if a high priority one. */ -static int ptlrpc_server_hpreq_check(struct ptlrpc_request *req) +static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt, + struct ptlrpc_request *req) { - int opc, rc = 0; - ENTRY; + int rc; + ENTRY; - /* Check by request opc. */ - opc = lustre_msg_get_opc(req->rq_reqmsg); - if (opc == OBD_PING) - RETURN(1); + rc = ptlrpc_server_hpreq_init(svcpt, req); + if (rc < 0) + RETURN(rc); - /* Perform request specific check. */ - if (req->rq_ops && req->rq_ops->hpreq_check) - rc = req->rq_ops->hpreq_check(req); - RETURN(rc); + ptlrpc_nrs_req_add(svcpt, req, !!rc); + + RETURN(0); } -/** Check if a request is a high priority one. */ -static int ptlrpc_server_request_add(struct ptlrpc_service *svc, - struct ptlrpc_request *req) +/** + * Allow to handle high priority request + * User can call it w/o any lock but need to hold + * ptlrpc_service_part::scp_req_lock to get reliable result + */ +static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt, + bool force) { - int rc; - ENTRY; + int running = svcpt->scp_nthrs_running; - rc = ptlrpc_server_hpreq_check(req); - if (rc < 0) - RETURN(rc); + if (!nrs_svcpt_has_hp(svcpt)) + return false; - spin_lock(&svc->srv_lock); - /* Before inserting the request into the queue, check if it is not - * inserted yet, or even already handled -- it may happen due to - * a racing ldlm_server_blocking_ast(). */ - if (req->rq_phase == RQ_PHASE_NEW && list_empty(&req->rq_list)) { - if (rc) - ptlrpc_hpreq_reorder_nolock(svc, req); - else - list_add_tail(&req->rq_list, &svc->srv_request_queue); - } - spin_unlock(&svc->srv_lock); + if (force) + return true; - RETURN(0); + if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL && + CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) { + /* leave just 1 thread for normal RPCs */ + running = PTLRPC_NTHRS_INIT; + if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) + running += 1; + } + + if (svcpt->scp_nreqs_active >= running - 1) + return false; + + if (svcpt->scp_nhreqs_active == 0) + return true; + + return !ptlrpc_nrs_req_pending_nolock(svcpt, false) || + svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio; +} + +static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt, + bool force) +{ + return ptlrpc_server_allow_high(svcpt, force) && + ptlrpc_nrs_req_pending_nolock(svcpt, true); } -/* Only allow normal priority requests on a service that has a high-priority +/** + * Only allow normal priority requests on a service that has a high-priority * queue if forced (i.e. cleanup), if there are other high priority requests * already being processed (i.e. those threads can service more high-priority * requests), or if there are enough idle threads that a later thread can do - * a high priority request. */ -static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force) + * a high priority request. + * User can call it w/o any lock but need to hold + * ptlrpc_service_part::scp_req_lock to get reliable result + */ +static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt, + bool force) { - return force || !svc->srv_hpreq_handler || svc->srv_n_hpreq > 0 || - svc->srv_n_active_reqs < svc->srv_threads_running - 2; + int running = svcpt->scp_nthrs_running; +#ifndef __KERNEL__ + if (1) /* always allow to handle normal request for liblustre */ + 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 < running - 2) + return true; + + if (svcpt->scp_nreqs_active >= running - 1) + return false; + + return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt); } -static struct ptlrpc_request * -ptlrpc_server_request_get(struct ptlrpc_service *svc, int force) +static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt, + bool force) { - struct ptlrpc_request *req = NULL; - ENTRY; - - if (ptlrpc_server_allow_normal(svc, force) && - !list_empty(&svc->srv_request_queue) && - (list_empty(&svc->srv_request_hpq) || - svc->srv_hpreq_count >= svc->srv_hpreq_ratio)) { - req = list_entry(svc->srv_request_queue.next, - struct ptlrpc_request, rq_list); - svc->srv_hpreq_count = 0; - } else if (!list_empty(&svc->srv_request_hpq)) { - req = list_entry(svc->srv_request_hpq.next, - struct ptlrpc_request, rq_list); - svc->srv_hpreq_count++; - } - RETURN(req); + return ptlrpc_server_allow_normal(svcpt, force) && + ptlrpc_nrs_req_pending_nolock(svcpt, false); } -static int ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force) +/** + * Returns true if there are requests available in incoming + * request queue for processing and it is allowed to fetch them. + * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock + * to get reliable result + * \see ptlrpc_server_allow_normal + * \see ptlrpc_server_allow high + */ +static inline bool +ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force) { - return ((ptlrpc_server_allow_normal(svc, force) && - !list_empty(&svc->srv_request_queue)) || - !list_empty(&svc->srv_request_hpq)); + return ptlrpc_server_high_pending(svcpt, force) || + ptlrpc_server_normal_pending(svcpt, force); } -/* Handle freshly incoming reqs, add to timed early reply list, - pass on to regular request queue */ -static int -ptlrpc_server_handle_req_in(struct ptlrpc_service *svc) +/** + * Fetch a request for processing from queue of unprocessed requests. + * Favors high-priority requests. + * Returns a pointer to fetched request. + */ +static struct ptlrpc_request * +ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force) { - struct ptlrpc_request *req; - __u32 deadline; - int rc; - ENTRY; + 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 - LASSERT(svc); + if (ptlrpc_server_high_pending(svcpt, force)) { + req = ptlrpc_nrs_req_get_nolock(svcpt, true, force); + if (req != NULL) { + svcpt->scp_hreq_count++; + goto got_request; + } + } - spin_lock(&svc->srv_lock); - if (list_empty(&svc->srv_req_in_queue)) { - spin_unlock(&svc->srv_lock); - RETURN(0); - } + if (ptlrpc_server_normal_pending(svcpt, force)) { + 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); - req = list_entry(svc->srv_req_in_queue.next, - struct ptlrpc_request, rq_list); - list_del_init (&req->rq_list); - /* Consider this still a "queued" request as far as stats are - concerned */ - spin_unlock(&svc->srv_lock); + if (likely(req->rq_export)) + class_export_rpc_inc(req->rq_export); + + RETURN(req); +} + +/** + * Handle freshly incoming reqs, add to timed early reply list, + * pass on to regular request queue. + * All incoming requests pass through here before getting into + * ptlrpc_server_handle_req later on. + */ +static int +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; + __u32 deadline; + int rc; + ENTRY; + + spin_lock(&svcpt->scp_lock); + if (cfs_list_empty(&svcpt->scp_req_incoming)) { + spin_unlock(&svcpt->scp_lock); + RETURN(0); + } + + req = cfs_list_entry(svcpt->scp_req_incoming.next, + struct ptlrpc_request, rq_list); + cfs_list_del_init(&req->rq_list); + svcpt->scp_nreqs_incoming--; + /* Consider this still a "queued" request as far as stats are + * concerned */ + spin_unlock(&svcpt->scp_lock); /* go through security check/transform */ rc = sptlrpc_svc_unwrap_request(req); @@ -1337,15 +1875,18 @@ ptlrpc_server_handle_req_in(struct ptlrpc_service *svc) LBUG(); } - /* Clear request swab mask; this is a new request */ - req->rq_req_swab_mask = 0; - - rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen); - if (rc != 0) { - CERROR("error unpacking request: ptl %d from %s x"LPU64"\n", - svc->srv_req_portal, libcfs_id2str(req->rq_peer), - req->rq_xid); - goto err_req; + /* + * for null-flavored rpc, msg has been unpacked by sptlrpc, although + * redo it wouldn't be harmful. + */ + if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) { + rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen); + if (rc != 0) { + CERROR("error unpacking request: ptl %d from %s " + "x"LPU64"\n", svc->srv_req_portal, + libcfs_id2str(req->rq_peer), req->rq_xid); + goto err_req; + } } rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF); @@ -1357,9 +1898,9 @@ ptlrpc_server_handle_req_in(struct ptlrpc_service *svc) } if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) && - lustre_msg_get_opc(req->rq_reqmsg) == obd_fail_val) { + lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) { CERROR("drop incoming rpc opc %u, x"LPU64"\n", - obd_fail_val, req->rq_xid); + cfs_fail_val, req->rq_xid); goto err_req; } @@ -1378,11 +1919,12 @@ ptlrpc_server_handle_req_in(struct ptlrpc_service *svc) break; case MDS_READPAGE: case OST_READ: + case MGS_CONFIG_READ: req->rq_bulk_read = 1; break; } - CDEBUG(D_NET, "got req "LPU64"\n", req->rq_xid); + CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid); req->rq_export = class_conn2export( lustre_msg_get_handle(req->rq_reqmsg)); @@ -1417,135 +1959,91 @@ ptlrpc_server_handle_req_in(struct ptlrpc_service *svc) goto err_req; } - ptlrpc_at_add_timed(req); - rc = ptlrpc_hpreq_init(svc, req); - if (rc) - GOTO(err_req, rc); - - /* Move it over to the request processing queue */ - rc = ptlrpc_server_request_add(svc, req); - if (rc) - GOTO(err_req, rc); - cfs_waitq_signal(&svc->srv_waitq); - RETURN(1); + 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 */ + rc = ptlrpc_server_request_add(svcpt, req); + if (rc) + GOTO(err_req, rc); + + wake_up(&svcpt->scp_waitq); + RETURN(1); err_req: - spin_lock(&svc->srv_lock); - svc->srv_n_queued_reqs--; - svc->srv_n_active_reqs++; - spin_unlock(&svc->srv_lock); - ptlrpc_server_finish_request(req); + ptlrpc_server_finish_request(svcpt, req); - RETURN(1); + RETURN(1); } +/** + * Main incoming request handling logic. + * Calls handler function from service to do actual processing. + */ static int -ptlrpc_server_handle_request(struct ptlrpc_service *svc, - struct ptlrpc_thread *thread) -{ - struct obd_export *export = NULL; - struct ptlrpc_request *request; - struct timeval work_start; - struct timeval work_end; - long timediff; - int opc, rc; - int fail_opc = 0; - ENTRY; +ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt, + struct ptlrpc_thread *thread) +{ + 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; - LASSERT(svc); + ENTRY; - spin_lock(&svc->srv_lock); - if (unlikely(!ptlrpc_server_request_pending(svc, 0) || - ( -#ifndef __KERNEL__ - /* !@%$# liblustre only has 1 thread */ - atomic_read(&svc->srv_n_difficult_replies) != 0 && -#endif - svc->srv_n_active_reqs >= (svc->srv_threads_running - 1)))) { - /* Don't handle regular requests in the last thread, in order * re - * to handle difficult replies (which might block other threads) - * as well as handle any incoming reqs, early replies, etc. - * That means we always need at least 2 service threads. */ - spin_unlock(&svc->srv_lock); - RETURN(0); - } + request = ptlrpc_server_request_get(svcpt, false); + if (request == NULL) + RETURN(0); - request = ptlrpc_server_request_get(svc, 0); - if (request == NULL) { - spin_unlock(&svc->srv_lock); - RETURN(0); - } - - opc = lustre_msg_get_opc(request->rq_reqmsg); if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT)) fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT; else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT)) fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT; if (unlikely(fail_opc)) { - if (request->rq_export && request->rq_ops) { - spin_unlock(&svc->srv_lock); - OBD_FAIL_TIMEOUT(fail_opc, 4); - spin_lock(&svc->srv_lock); - request = ptlrpc_server_request_get(svc, 0); - if (request == NULL) { - spin_unlock(&svc->srv_lock); - RETURN(0); - } - LASSERT(ptlrpc_server_request_pending(svc, 0)); - } - } - - list_del_init(&request->rq_list); - svc->srv_n_queued_reqs--; - svc->srv_n_active_reqs++; - if (request->rq_hp) - svc->srv_n_hpreq++; + if (request->rq_export && request->rq_ops) + OBD_FAIL_TIMEOUT(fail_opc, 4); + } - /* The phase is changed under the lock here because we need to know - * the request is under processing (see ptlrpc_hpreq_reorder()). */ ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET); - spin_unlock(&svc->srv_lock); - ptlrpc_hpreq_fini(request); + if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG)) + libcfs_debug_dumplog(); - if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG)) - libcfs_debug_dumplog(); - - do_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, - svc->srv_n_queued_reqs); - lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR, - svc->srv_n_active_reqs); - lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT, - at_get(&svc->srv_at_estimate)); - } - - rc = lu_context_init(&request->rq_session, - LCT_SESSION|LCT_REMEMBER|LCT_NOREF); - if (rc) { - CERROR("Failure to initialize session: %d\n", rc); - goto out_req; + svcpt->scp_nreqs_incoming); + lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR, + svcpt->scp_nreqs_active); + lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT, + at_get(&svcpt->scp_at_estimate)); } - 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. @@ -1558,64 +2056,69 @@ ptlrpc_server_handle_request(struct ptlrpc_service *svc, 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 ? - 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) - OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val); + CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val); - rc = svc->srv_handler(request); + CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid); - ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE); + /* 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); - CDEBUG(D_RPCTRACE, "Handled 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 ? - 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)); + 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)); - } - - do_gettimeofday(&work_end); - timediff = cfs_timeval_sub(&work_end, &work_start, NULL); - CDEBUG(D_RPCTRACE, "request x"LPU64" opc %u from %s processed in " - "%ldus (%ldus total) trans "LPU64" rc %d/%d\n", - request->rq_xid, lustre_msg_get_opc(request->rq_reqmsg), - libcfs_id2str(request->rq_peer), timediff, - cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL), - request->rq_repmsg ? lustre_msg_get_transno(request->rq_repmsg) : - request->rq_transno, request->rq_status, - request->rq_repmsg ? lustre_msg_get_status(request->rq_repmsg): - -999); + 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)); + } + + 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), + request->rq_xid, + libcfs_id2str(request->rq_peer), + lustre_msg_get_opc(request->rq_reqmsg), + timediff, + cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL), + (request->rq_repmsg ? + lustre_msg_get_transno(request->rq_repmsg) : + request->rq_transno), + request->rq_status, + (request->rq_repmsg ? + lustre_msg_get_status(request->rq_repmsg) : -999)); if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) { __u32 op = lustre_msg_get_opc(request->rq_reqmsg); int opc = opcode_offset(op); @@ -1635,50 +2138,64 @@ put_conn: request->rq_arrival_time.tv_sec)); } -out_req: - spin_lock(&svc->srv_lock); - if (request->rq_hp) - svc->srv_n_hpreq--; - spin_unlock(&svc->srv_lock); - ptlrpc_server_finish_request(request); + ptlrpc_server_finish_active_request(svcpt, request); - RETURN(1); + RETURN(1); } /** * An internal function to process a single reply state object. */ static int -ptlrpc_handle_rs (struct ptlrpc_reply_state *rs) +ptlrpc_handle_rs(struct ptlrpc_reply_state *rs) { - struct ptlrpc_service *svc = rs->rs_service; + struct ptlrpc_service_part *svcpt = rs->rs_svcpt; + struct ptlrpc_service *svc = svcpt->scp_service; struct obd_export *exp; - struct obd_device *obd; int nlocks; int been_handled; ENTRY; exp = rs->rs_export; - obd = exp->exp_obd; LASSERT (rs->rs_difficult); LASSERT (rs->rs_scheduled); - LASSERT (list_empty(&rs->rs_list)); - - spin_lock (&exp->exp_lock); - /* Noop if removed already */ - list_del_init (&rs->rs_exp_list); - spin_unlock (&exp->exp_lock); - - /* Avoid exp_uncommitted_replies_lock contention if we 100% sure that - * rs has been removed from the list already */ - if (!list_empty_careful(&rs->rs_obd_list)) { - spin_lock(&exp->exp_uncommitted_replies_lock); - list_del_init(&rs->rs_obd_list); - spin_unlock(&exp->exp_uncommitted_replies_lock); - } + LASSERT (cfs_list_empty(&rs->rs_list)); + + 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 + * exp_uncommitted_replies. It then drops this lock and schedules the + * replies it found for handling here. + * + * We can avoid contention for exp_uncommitted_replies_lock between the + * HRT threads and further commit callbacks by checking rs_committed + * which is set in the commit callback while it holds both + * rs_lock and exp_uncommitted_reples. + * + * If we see rs_committed clear, the commit callback _may_ not have + * handled this reply yet and we race with it to grab + * exp_uncommitted_replies_lock before removing the reply from + * exp_uncommitted_replies. Note that if we lose the race and the + * reply has already been removed, list_del_init() is a noop. + * + * If we see rs_committed set, we know the commit callback is handling, + * or has handled this reply since store reordering might allow us to + * see rs_committed set out of sequence. But since this is done + * 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) { + spin_lock(&exp->exp_uncommitted_replies_lock); + cfs_list_del_init(&rs->rs_obd_list); + spin_unlock(&exp->exp_uncommitted_replies_lock); + } - spin_lock(&rs->rs_lock); + spin_lock(&rs->rs_lock); been_handled = rs->rs_handled; rs->rs_handled = 1; @@ -1689,48 +2206,46 @@ 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) { - 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]); - 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 */ - 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); - atomic_dec (&svc->srv_outstanding_replies); - if (atomic_dec_and_test(&svc->srv_n_difficult_replies) && - svc->srv_is_stopping) - cfs_waitq_broadcast(&svc->srv_waitq); - RETURN(1); - } - - /* still on the net; callback will schedule */ - spin_unlock(&rs->rs_lock); - RETURN(1); + if (cfs_atomic_dec_and_test(&svcpt->scp_nreps_difficult) && + svc->srv_is_stopping) + wake_up_all(&svcpt->scp_waitq); + RETURN(1); + } + + /* still on the net; callback will schedule */ + spin_unlock(&rs->rs_lock); + RETURN(1); } #ifndef __KERNEL__ @@ -1744,22 +2259,22 @@ ptlrpc_handle_rs (struct ptlrpc_reply_state *rs) * \retval 1 one reply processed */ static int -ptlrpc_server_handle_reply(struct ptlrpc_service *svc) +ptlrpc_server_handle_reply(struct ptlrpc_service_part *svcpt) { - struct ptlrpc_reply_state *rs = NULL; - ENTRY; - - spin_lock(&svc->srv_lock); - if (!list_empty(&svc->srv_reply_queue)) { - rs = list_entry(svc->srv_reply_queue.prev, - struct ptlrpc_reply_state, - rs_list); - list_del_init(&rs->rs_list); - } - spin_unlock(&svc->srv_lock); - if (rs != NULL) - ptlrpc_handle_rs(rs); - RETURN(rs != NULL); + struct ptlrpc_reply_state *rs = NULL; + ENTRY; + + 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); + } + spin_unlock(&svcpt->scp_rep_lock); + if (rs != NULL) + ptlrpc_handle_rs(rs); + RETURN(rs != NULL); } /* FIXME make use of timeout later */ @@ -1768,49 +2283,53 @@ liblustre_check_services (void *arg) { int did_something = 0; int rc; - struct list_head *tmp, *nxt; + cfs_list_t *tmp, *nxt; ENTRY; /* I'm relying on being single threaded, not to have to lock * ptlrpc_all_services etc */ - list_for_each_safe (tmp, nxt, &ptlrpc_all_services) { + cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) { struct ptlrpc_service *svc = - list_entry (tmp, struct ptlrpc_service, srv_list); + cfs_list_entry (tmp, struct ptlrpc_service, srv_list); + struct ptlrpc_service_part *svcpt; - if (svc->srv_threads_running != 0) /* I've recursed */ - continue; + LASSERT(svc->srv_ncpts == 1); + svcpt = svc->srv_parts[0]; - /* service threads can block for bulk, so this limits us - * (arbitrarily) to recursing 1 stack frame per service. - * Note that the problem with recursion is that we have to - * unwind completely before our caller can resume. */ + if (svcpt->scp_nthrs_running != 0) /* I've recursed */ + continue; - svc->srv_threads_running++; + /* service threads can block for bulk, so this limits us + * (arbitrarily) to recursing 1 stack frame per service. + * Note that the problem with recursion is that we have to + * unwind completely before our caller can resume. */ - do { - rc = ptlrpc_server_handle_req_in(svc); - rc |= ptlrpc_server_handle_reply(svc); - rc |= ptlrpc_at_check_timed(svc); - rc |= ptlrpc_server_handle_request(svc, NULL); - rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0); - did_something |= rc; - } while (rc); + svcpt->scp_nthrs_running++; - svc->srv_threads_running--; - } + do { + rc = ptlrpc_server_handle_req_in(svcpt, NULL); + rc |= ptlrpc_server_handle_reply(svcpt); + rc |= ptlrpc_at_check_timed(svcpt); + rc |= ptlrpc_server_handle_request(svcpt, NULL); + rc |= (ptlrpc_server_post_idle_rqbds(svcpt) > 0); + did_something |= rc; + } while (rc); - RETURN(did_something); + svcpt->scp_nthrs_running--; + } + + RETURN(did_something); } #define ptlrpc_stop_all_threads(s) do {} while (0) #else /* __KERNEL__ */ static void -ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc) +ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt) { - int avail = svc->srv_nrqbd_receiving; - int low_water = test_req_buffer_pressure ? 0 : - svc->srv_nbuf_per_group/2; + int avail = svcpt->scp_nrqbds_posted; + int low_water = test_req_buffer_pressure ? 0 : + svcpt->scp_service->srv_nbuf_per_group / 2; /* NB I'm not locking; just looking. */ @@ -1820,167 +2339,264 @@ ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc) * space. */ if (avail <= low_water) - ptlrpc_grow_req_bufs(svc); + ptlrpc_grow_req_bufs(svcpt, 1); - if (svc->srv_stats) - lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR, - avail); + if (svcpt->scp_service->srv_stats) { + lprocfs_counter_add(svcpt->scp_service->srv_stats, + PTLRPC_REQBUF_AVAIL_CNTR, avail); + } } static int ptlrpc_retry_rqbds(void *arg) { - struct ptlrpc_service *svc = (struct ptlrpc_service *)arg; + struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg; - svc->srv_rqbd_timeout = 0; - return (-ETIMEDOUT); + svcpt->scp_rqbd_timeout = 0; + return -ETIMEDOUT; } -static int ptlrpc_main(void *arg) +static inline int +ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt) { - struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg; - struct ptlrpc_service *svc = data->svc; - struct ptlrpc_thread *thread = data->thread; - struct obd_device *dev = data->dev; - struct ptlrpc_reply_state *rs; -#ifdef WITH_GROUP_INFO - struct group_info *ginfo = NULL; -#endif - struct lu_env env; - int counter = 0, rc = 0; - ENTRY; + return svcpt->scp_nreqs_active < + svcpt->scp_nthrs_running - 1 - + (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL); +} - cfs_daemonize_ctxt(data->name); +/** + * allowed to create more threads + * user can call it w/o any lock but need to hold + * ptlrpc_service_part::scp_lock to get reliable result + */ +static inline int +ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt) +{ + return svcpt->scp_nthrs_running + + svcpt->scp_nthrs_starting < + svcpt->scp_service->srv_nthrs_cpt_limit; +} -#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; +/** + * too many requests and allowed to create more threads + */ +static inline int +ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt) +{ + return !ptlrpc_threads_enough(svcpt) && + ptlrpc_threads_increasable(svcpt); +} - for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) { - if (!cpu_online(cpu)) - continue; - if (num_cpu == thread->t_id % num_online_cpus()) - break; - num_cpu++; - } - set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu))); - } -#endif +static inline int +ptlrpc_thread_stopping(struct ptlrpc_thread *thread) +{ + return thread_is_stopping(thread) || + thread->t_svcpt->scp_service->srv_is_stopping; +} + +static inline int +ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt) +{ + return !cfs_list_empty(&svcpt->scp_rqbd_idle) && + svcpt->scp_rqbd_timeout == 0; +} + +static inline int +ptlrpc_at_check(struct ptlrpc_service_part *svcpt) +{ + return svcpt->scp_at_check; +} + +/** + * requests wait on preprocessing + * user can call it w/o any lock but need to hold + * ptlrpc_service_part::scp_lock to get reliable result + */ +static inline int +ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt) +{ + return !cfs_list_empty(&svcpt->scp_req_incoming); +} + +static __attribute__((__noinline__)) int +ptlrpc_wait_event(struct ptlrpc_service_part *svcpt, + struct ptlrpc_thread *thread) +{ + /* Don't exit while there are replies to be handled */ + struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout, + ptlrpc_retry_rqbds, svcpt); + + lc_watchdog_disable(thread->t_watchdog); + + cond_resched(); + + l_wait_event_exclusive_head(svcpt->scp_waitq, + ptlrpc_thread_stopping(thread) || + ptlrpc_server_request_incoming(svcpt) || + ptlrpc_server_request_pending(svcpt, false) || + ptlrpc_rqbd_pending(svcpt) || + ptlrpc_at_check(svcpt), &lwi); + + if (ptlrpc_thread_stopping(thread)) + return -EINTR; + + lc_watchdog_touch(thread->t_watchdog, + ptlrpc_server_get_timeout(svcpt)); + return 0; +} +/** + * Main thread body for service threads. + * Waits in a loop waiting for new requests to process to appear. + * Every time an incoming requests is added to its queue, a waitq + * is woken up and one of the threads will handle it. + */ +static int ptlrpc_main(void *arg) +{ + struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg; + struct ptlrpc_service_part *svcpt = thread->t_svcpt; + struct ptlrpc_service *svc = svcpt->scp_service; + struct ptlrpc_reply_state *rs; #ifdef WITH_GROUP_INFO - ginfo = groups_alloc(0); - if (!ginfo) { - rc = -ENOMEM; - goto out; - } + struct group_info *ginfo = NULL; +#endif + struct lu_env *env; + int counter = 0, rc = 0; + ENTRY; + + thread->t_pid = current_pid(); + unshare_fs_struct(); + + /* 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); + } - set_current_groups(ginfo); - put_group_info(ginfo); +#ifdef WITH_GROUP_INFO + ginfo = groups_alloc(0); + if (!ginfo) { + rc = -ENOMEM; + goto out; + } + + set_current_groups(ginfo); + put_group_info(ginfo); #endif - if (svc->srv_init != NULL) { - rc = svc->srv_init(thread); + if (svc->srv_ops.so_thr_init != NULL) { + rc = svc->srv_ops.so_thr_init(thread); if (rc) goto out; } - rc = lu_context_init(&env.le_ctx, + OBD_ALLOC_PTR(env); + if (env == NULL) { + rc = -ENOMEM; + goto out_srv_fini; + } + + rc = lu_context_init(&env->le_ctx, svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF); if (rc) goto out_srv_fini; - thread->t_env = &env; - env.le_ctx.lc_thread = thread; - env.le_ctx.lc_cookie = 0x6; + thread->t_env = env; + 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_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD); + OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size); if (!rs) { rc = -ENOMEM; goto out_srv_fini; } - /* Record that the thread is running */ - thread->t_flags = SVC_RUNNING; - /* - * wake up our creator. Note: @data is invalid after this point, - * because it's allocated on ptlrpc_start_thread() stack. - */ - cfs_waitq_signal(&thread->t_ctl_waitq); - - thread->t_watchdog = lc_watchdog_add(GET_TIMEOUT(svc), NULL, NULL); + spin_lock(&svcpt->scp_lock); - spin_lock(&svc->srv_lock); - svc->srv_threads_running++; - list_add(&rs->rs_list, &svc->srv_free_rs_list); - spin_unlock(&svc->srv_lock); - cfs_waitq_signal(&svc->srv_free_rs_waitq); + LASSERT(thread_is_starting(thread)); + thread_clear_flags(thread, SVC_STARTING); - CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id, - svc->srv_threads_running); + LASSERT(svcpt->scp_nthrs_starting == 1); + svcpt->scp_nthrs_starting--; - /* XXX maintain a list of all managed devices: insert here */ + /* SVC_STOPPING may already be set here if someone else is trying + * to stop the service while this new thread has been dynamically + * forked. We still set SVC_RUNNING to let our creator know that + * we are now running, however we will exit as soon as possible */ + thread_add_flags(thread, SVC_RUNNING); + svcpt->scp_nthrs_running++; + spin_unlock(&svcpt->scp_lock); - while ((thread->t_flags & SVC_STOPPING) == 0) { - /* Don't exit while there are replies to be handled */ - struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout, - ptlrpc_retry_rqbds, svc); + /* wake up our creator in case he's still waiting. */ + wake_up(&thread->t_ctl_waitq); - lc_watchdog_disable(thread->t_watchdog); + thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt), + NULL, NULL); - cond_resched(); + spin_lock(&svcpt->scp_rep_lock); + cfs_list_add(&rs->rs_list, &svcpt->scp_rep_idle); + wake_up(&svcpt->scp_rep_waitq); + spin_unlock(&svcpt->scp_rep_lock); - l_wait_event_exclusive (svc->srv_waitq, - ((thread->t_flags & SVC_STOPPING) != 0) || - (!list_empty(&svc->srv_idle_rqbds) && - svc->srv_rqbd_timeout == 0) || - !list_empty(&svc->srv_req_in_queue) || - (ptlrpc_server_request_pending(svc, 0) && - (svc->srv_n_active_reqs < - (svc->srv_threads_running - 1))) || - svc->srv_at_check, - &lwi); + CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id, + svcpt->scp_nthrs_running); - lc_watchdog_touch(thread->t_watchdog, GET_TIMEOUT(svc)); + /* XXX maintain a list of all managed devices: insert here */ + while (!ptlrpc_thread_stopping(thread)) { + if (ptlrpc_wait_event(svcpt, thread)) + break; - ptlrpc_check_rqbd_pool(svc); + ptlrpc_check_rqbd_pool(svcpt); - if ((svc->srv_threads_started < svc->srv_threads_max) && - (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){ - /* Ignore return code - we tried... */ - ptlrpc_start_thread(dev, svc); + if (ptlrpc_threads_need_create(svcpt)) { + /* Ignore return code - we tried... */ + ptlrpc_start_thread(svcpt, 0); } - if (!list_empty(&svc->srv_req_in_queue)) { - /* Process all incoming reqs before handling any */ - ptlrpc_server_handle_req_in(svc); - /* but limit ourselves in case of flood */ - if (counter++ < 1000) - continue; - counter = 0; + /* Process all incoming reqs before handling any */ + if (ptlrpc_server_request_incoming(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; + counter = 0; + } + + if (ptlrpc_at_check(svcpt)) + ptlrpc_at_check_timed(svcpt); + + 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); } - if (svc->srv_at_check) - ptlrpc_at_check_timed(svc); - - /* don't handle requests in the last thread */ - if (ptlrpc_server_request_pending(svc, 0) && - (svc->srv_n_active_reqs < (svc->srv_threads_running - 1))) { - lu_context_enter(&env.le_ctx); - ptlrpc_server_handle_request(svc, thread); - lu_context_exit(&env.le_ctx); - } - - if (!list_empty(&svc->srv_idle_rqbds) && - ptlrpc_server_post_idle_rqbds(svc) < 0) { - /* I just failed to repost request buffers. Wait - * for a timeout (unless something else happens) - * before I try again */ - svc->srv_rqbd_timeout = cfs_time_seconds(1)/10; - CDEBUG(D_RPCTRACE,"Posted buffers: %d\n", - svc->srv_nrqbd_receiving); + if (ptlrpc_rqbd_pending(svcpt) && + ptlrpc_server_post_idle_rqbds(svcpt) < 0) { + /* I just failed to repost request buffers. + * Wait for a timeout (unless something else + * happens) before I try again */ + svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10; + CDEBUG(D_RPCTRACE, "Posted buffers: %d\n", + svcpt->scp_nrqbds_posted); } } @@ -1991,333 +2607,426 @@ out_srv_fini: /* * deconstruct service specific state created by ptlrpc_start_thread() */ - if (svc->srv_done != NULL) - svc->srv_done(thread); + if (svc->srv_ops.so_thr_done != NULL) + svc->srv_ops.so_thr_done(thread); - lu_context_fini(&env.le_ctx); + if (env != NULL) { + lu_context_fini(&env->le_ctx); + OBD_FREE_PTR(env); + } out: - CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc); + CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n", + thread, thread->t_pid, thread->t_id, rc); - spin_lock(&svc->srv_lock); - svc->srv_threads_running--; /* must know immediately */ - thread->t_id = rc; - thread->t_flags = SVC_STOPPED; + spin_lock(&svcpt->scp_lock); + if (thread_test_and_clear_flags(thread, SVC_STARTING)) + svcpt->scp_nthrs_starting--; - cfs_waitq_signal(&thread->t_ctl_waitq); - spin_unlock(&svc->srv_lock); + if (thread_test_and_clear_flags(thread, SVC_RUNNING)) { + /* must know immediately */ + svcpt->scp_nthrs_running--; + } - return rc; -} + thread->t_id = rc; + thread_add_flags(thread, SVC_STOPPED); -struct ptlrpc_hr_args { - int thread_index; - int cpu_index; - struct ptlrpc_hr_service *hrs; -}; + wake_up(&thread->t_ctl_waitq); + spin_unlock(&svcpt->scp_lock); + + return rc; +} -static int hrt_dont_sleep(struct ptlrpc_hr_thread *t, - struct list_head *replies) +static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt, + cfs_list_t *replies) { - int result; + int result; - spin_lock(&t->hrt_lock); - list_splice_init(&t->hrt_queue, replies); - result = test_bit(HRT_STOPPING, &t->hrt_flags) || - !list_empty(replies); - 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; } +/** + * Main body of "handle reply" function. + * It processes acked reply states + */ 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(HAVE_NODE_TO_CPUMASK) - set_cpus_allowed(cfs_current(), - node_to_cpumask(cpu_to_node(hr_args->cpu_index))); -#endif - 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 (!test_bit(HRT_STOPPING, &t->hrt_flags)) { + cfs_atomic_inc(&hrp->hrp_nstarted); + wake_up(&ptlrpc_hr.hr_waitq); - cfs_wait_event(t->hrt_wait, hrt_dont_sleep(t, &replies)); - while (!list_empty(&replies)) { + while (!ptlrpc_hr.hr_stopping) { + l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies)); + + while (!cfs_list_empty(&replies)) { struct ptlrpc_reply_state *rs; - rs = list_entry(replies.prev, - struct ptlrpc_reply_state, - rs_list); - list_del_init(&rs->rs_list); + rs = cfs_list_entry(replies.prev, + struct ptlrpc_reply_state, + rs_list); + cfs_list_del_init(&rs->rs_list); ptlrpc_handle_rs(rs); } } - clear_bit(HRT_RUNNING, &t->hrt_flags); - complete(&t->hrt_completion); + cfs_atomic_inc(&hrp->hrp_nstopped); + wake_up(&ptlrpc_hr.hr_waitq); - return 0; + return 0; } -static int ptlrpc_start_hr_thread(struct ptlrpc_hr_service *hr, int n, int cpu) +static void ptlrpc_stop_hr_threads(void) { - struct ptlrpc_hr_thread *t = &hr->hr_threads[n]; - struct ptlrpc_hr_args args; - int rc; - ENTRY; - - args.thread_index = n; - args.cpu_index = cpu; - args.hrs = hr; - - rc = cfs_kernel_thread(ptlrpc_hr_main, (void*)&args, - CLONE_VM|CLONE_FILES); - if (rc < 0) { - complete(&t->hrt_completion); - GOTO(out, rc); - } - cfs_wait_event(t->hrt_wait, test_bit(HRT_RUNNING, &t->hrt_flags)); - RETURN(0); - out: - return rc; -} - -static void ptlrpc_stop_hr_thread(struct ptlrpc_hr_thread *t) -{ - ENTRY; - - set_bit(HRT_STOPPING, &t->hrt_flags); - cfs_waitq_signal(&t->hrt_wait); - wait_for_completion(&t->hrt_completion); - - EXIT; + struct ptlrpc_hr_partition *hrp; + int i; + int j; + + ptlrpc_hr.hr_stopping = 1; + + 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); + } + + 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 void ptlrpc_stop_hr_threads(struct ptlrpc_hr_service *hrs) +static int ptlrpc_start_hr_threads(void) { - int n; - ENTRY; - - for (n = 0; n < hrs->hr_n_threads; n++) - ptlrpc_stop_hr_thread(&hrs->hr_threads[n]); - - EXIT; + struct ptlrpc_hr_partition *hrp; + int i; + int j; + ENTRY; + + 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; + + CERROR("Reply handling thread %d:%d Failed on starting: " + "rc = %d\n", i, j, rc); + ptlrpc_stop_hr_threads(); + RETURN(rc); + } + RETURN(0); } -static int ptlrpc_start_hr_threads(struct ptlrpc_hr_service *hr) +static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt) { - int rc = -ENOMEM; - int n, cpu, threads_started = 0; - ENTRY; - - LASSERT(hr != NULL); - LASSERT(hr->hr_n_threads > 0); - - for (n = 0, cpu = 0; n < hr->hr_n_threads; n++) { -#if defined(HAVE_NODE_TO_CPUMASK) - while(!cpu_online(cpu)) { - cpu++; - if (cpu >= 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); -} - -static void ptlrpc_stop_thread(struct ptlrpc_service *svc, - struct ptlrpc_thread *thread) -{ - struct l_wait_info lwi = { 0 }; - ENTRY; - - CDEBUG(D_RPCTRACE, "Stopping thread %p\n", thread); - spin_lock(&svc->srv_lock); - thread->t_flags = SVC_STOPPING; - spin_unlock(&svc->srv_lock); - - cfs_waitq_broadcast(&svc->srv_waitq); - l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED), - &lwi); - - spin_lock(&svc->srv_lock); - list_del(&thread->t_link); - spin_unlock(&svc->srv_lock); - - OBD_FREE_PTR(thread); - EXIT; + struct l_wait_info lwi = { 0 }; + struct ptlrpc_thread *thread; + CFS_LIST_HEAD (zombie); + + ENTRY; + + CDEBUG(D_INFO, "Stopping threads for service %s\n", + svcpt->scp_service->srv_name); + + 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", + svcpt->scp_service->srv_thread_name, thread->t_id); + thread_add_flags(thread, SVC_STOPPING); + } + + wake_up_all(&svcpt->scp_waitq); + + while (!cfs_list_empty(&svcpt->scp_threads)) { + thread = cfs_list_entry(svcpt->scp_threads.next, + struct ptlrpc_thread, t_link); + if (thread_is_stopped(thread)) { + cfs_list_del(&thread->t_link); + cfs_list_add(&thread->t_link, &zombie); + continue; + } + 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); + + spin_lock(&svcpt->scp_lock); + } + + spin_unlock(&svcpt->scp_lock); + + while (!cfs_list_empty(&zombie)) { + thread = cfs_list_entry(zombie.next, + struct ptlrpc_thread, t_link); + cfs_list_del(&thread->t_link); + OBD_FREE_PTR(thread); + } + EXIT; } +/** + * Stops all threads of a particular service \a svc + */ void ptlrpc_stop_all_threads(struct ptlrpc_service *svc) { - struct ptlrpc_thread *thread; - ENTRY; - - spin_lock(&svc->srv_lock); - while (!list_empty(&svc->srv_threads)) { - thread = list_entry(svc->srv_threads.next, - struct ptlrpc_thread, t_link); + struct ptlrpc_service_part *svcpt; + int i; + ENTRY; - spin_unlock(&svc->srv_lock); - ptlrpc_stop_thread(svc, thread); - spin_lock(&svc->srv_lock); - } + ptlrpc_service_for_each_part(svcpt, i, svc) { + if (svcpt->scp_service != NULL) + ptlrpc_svcpt_stop_threads(svcpt); + } - spin_unlock(&svc->srv_lock); - EXIT; + EXIT; } +EXPORT_SYMBOL(ptlrpc_stop_all_threads); -int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc) +int ptlrpc_start_threads(struct ptlrpc_service *svc) { - int i, rc = 0; - 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(dev, svc); - /* We have enough threads, don't start more. b=15759 */ - if (rc == -EMFILE) - break; - if (rc) { - CERROR("cannot start %s thread #%d: rc %d\n", - svc->srv_thread_name, i, rc); - ptlrpc_stop_all_threads(svc); - } - } - RETURN(rc); + int rc = 0; + int i; + int j; + ENTRY; + + /* 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 obd_device *dev, struct ptlrpc_service *svc) +int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait) { - struct l_wait_info lwi = { 0 }; - struct ptlrpc_svc_data d; - struct ptlrpc_thread *thread; - char name[32]; - int id, rc; - ENTRY; - - CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n", - svc->srv_name, svc->srv_threads_started, svc->srv_threads_min, - svc->srv_threads_max, svc->srv_threads_running); - if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) || - (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) && - svc->srv_threads_started == svc->srv_threads_min - 1)) - RETURN(-EMFILE); - - OBD_ALLOC_PTR(thread); - if (thread == NULL) - RETURN(-ENOMEM); - cfs_waitq_init(&thread->t_ctl_waitq); - - spin_lock(&svc->srv_lock); - if (svc->srv_threads_started >= svc->srv_threads_max) { - spin_unlock(&svc->srv_lock); - OBD_FREE_PTR(thread); - RETURN(-EMFILE); + struct l_wait_info lwi = { 0 }; + struct ptlrpc_thread *thread; + struct ptlrpc_service *svc; + int rc; + ENTRY; + + LASSERT(svcpt != NULL); + + svc = svcpt->scp_service; + + CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n", + svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running, + svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit); + + again: + if (unlikely(svc->srv_is_stopping)) + RETURN(-ESRCH); + + if (!ptlrpc_threads_increasable(svcpt) || + (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) && + svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1)) + RETURN(-EMFILE); + + OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt); + if (thread == NULL) + RETURN(-ENOMEM); + init_waitqueue_head(&thread->t_ctl_waitq); + + spin_lock(&svcpt->scp_lock); + if (!ptlrpc_threads_increasable(svcpt)) { + spin_unlock(&svcpt->scp_lock); + OBD_FREE_PTR(thread); + RETURN(-EMFILE); + } + + if (svcpt->scp_nthrs_starting != 0) { + /* serialize starting because some modules (obdfilter) + * might require unique and contiguous t_id */ + LASSERT(svcpt->scp_nthrs_starting == 1); + 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); + schedule(); + goto again; + } + + CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n", + svc->srv_thread_name, svcpt->scp_thr_nextid); + RETURN(-EAGAIN); + } + + svcpt->scp_nthrs_starting++; + thread->t_id = svcpt->scp_thr_nextid++; + thread_add_flags(thread, SVC_STARTING); + thread->t_svcpt = svcpt; + + cfs_list_add(&thread->t_link, &svcpt->scp_threads); + spin_unlock(&svcpt->scp_lock); + + 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); + 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); + spin_lock(&svcpt->scp_lock); + --svcpt->scp_nthrs_starting; + 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); } - list_add(&thread->t_link, &svc->srv_threads); - id = svc->srv_threads_started++; - spin_unlock(&svc->srv_lock); - - thread->t_svc = svc; - thread->t_id = id; - sprintf(name, "%s_%02d", svc->srv_thread_name, id); - d.dev = dev; - d.svc = svc; - d.name = name; - d.thread = thread; - - CDEBUG(D_RPCTRACE, "starting thread '%s'\n", 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_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES); - if (rc < 0) { - CERROR("cannot start thread '%s': rc %d\n", name, rc); - spin_lock(&svc->srv_lock); - list_del(&thread->t_link); - --svc->srv_threads_started; - spin_unlock(&svc->srv_lock); + if (!wait) + RETURN(0); - OBD_FREE(thread, sizeof(*thread)); - RETURN(rc); - } l_wait_event(thread->t_ctl_waitq, - thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi); + thread_is_running(thread) || thread_is_stopped(thread), + &lwi); - rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0; + rc = thread_is_stopped(thread) ? thread->t_id : 0; RETURN(rc); } - int ptlrpc_hr_init(void) { - int i; - int n_cpus = 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; + + memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr)); + ptlrpc_hr.hr_cpt_table = cfs_cpt_table; + + ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table, + sizeof(*hrp)); + if (ptlrpc_hr.hr_partitions == NULL) + RETURN(-ENOMEM); + + init_waitqueue_head(&ptlrpc_hr.hr_waitq); + + 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); +} - LASSERT(ptlrpc_hr == NULL); +void ptlrpc_hr_fini(void) +{ + struct ptlrpc_hr_partition *hrp; + int i; - 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]; + if (ptlrpc_hr.hr_partitions == NULL) + return; - spin_lock_init(&t->hrt_lock); - cfs_waitq_init(&t->hrt_wait); - CFS_INIT_LIST_HEAD(&t->hrt_queue); - init_completion(&t->hrt_completion); - } - hr->hr_n_threads = n_cpus; - hr->hr_size = size; - ptlrpc_hr = hr; + ptlrpc_stop_hr_threads(); - 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) { + if (hrp->hrp_thrs != NULL) { + OBD_FREE(hrp->hrp_thrs, + hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0])); + } + } -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; - } + cfs_percpt_free(ptlrpc_hr.hr_partitions); + ptlrpc_hr.hr_partitions = NULL; } #endif /* __KERNEL__ */ @@ -2325,200 +3034,279 @@ void ptlrpc_hr_fini(void) /** * Wait until all already scheduled replies are processed. */ -static void ptlrpc_wait_replies(struct ptlrpc_service *svc) -{ - while (1) { - int rc; - struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), - NULL, NULL); - rc = l_wait_event(svc->srv_waitq, - atomic_read(&svc->srv_n_difficult_replies) == 0, - &lwi); - if (rc == 0) - break; - CWARN("Unexpectedly long timeout %p\n", svc); - } +static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt) +{ + while (1) { + int rc; + struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), + NULL, NULL); + + rc = l_wait_event(svcpt->scp_waitq, + cfs_atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi); + if (rc == 0) + break; + CWARN("Unexpectedly long timeout %s %p\n", + svcpt->scp_service->srv_name, svcpt->scp_service); + } } -int ptlrpc_unregister_service(struct ptlrpc_service *service) +static void +ptlrpc_service_del_atimer(struct ptlrpc_service *svc) { - int rc; - struct l_wait_info lwi; - struct list_head *tmp; - struct ptlrpc_reply_state *rs, *t; - struct ptlrpc_at_array *array = &service->srv_at_array; - ENTRY; - - service->srv_is_stopping = 1; - cfs_timer_disarm(&service->srv_at_timer); - - ptlrpc_stop_all_threads(service); - LASSERT(list_empty(&service->srv_threads)); - - spin_lock (&ptlrpc_all_services_lock); - list_del_init (&service->srv_list); - spin_unlock (&ptlrpc_all_services_lock); - - ptlrpc_lprocfs_unregister_service(service); - - /* All history will be culled when the next request buffer is - * freed */ - service->srv_max_history_rqbds = 0; - - CDEBUG(D_NET, "%s: tearing down\n", service->srv_name); - - rc = LNetClearLazyPortal(service->srv_req_portal); - LASSERT (rc == 0); + struct ptlrpc_service_part *svcpt; + int i; + + /* early disarm AT timer... */ + ptlrpc_service_for_each_part(svcpt, i, svc) { + if (svcpt->scp_service != NULL) + cfs_timer_disarm(&svcpt->scp_at_timer); + } +} - /* Unlink all the request buffers. This forces a 'final' event with - * its 'unlink' flag set for each posted rqbd */ - list_for_each(tmp, &service->srv_active_rqbds) { - struct ptlrpc_request_buffer_desc *rqbd = - list_entry(tmp, struct ptlrpc_request_buffer_desc, - rqbd_list); +static void +ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc) +{ + struct ptlrpc_service_part *svcpt; + 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 + * freed in ptlrpc_service_purge_all() */ + svc->srv_hist_nrqbds_cpt_max = 0; + + rc = LNetClearLazyPortal(svc->srv_req_portal); + LASSERT(rc == 0); + + ptlrpc_service_for_each_part(svcpt, i, svc) { + if (svcpt->scp_service == NULL) + break; + + /* Unlink all the request buffers. This forces a 'final' + * event with its 'unlink' flag set for each posted rqbd */ + cfs_list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted, + rqbd_list) { + rc = LNetMDUnlink(rqbd->rqbd_md_h); + LASSERT(rc == 0 || rc == -ENOENT); + } + } + + 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 */ + spin_lock(&svcpt->scp_lock); + while (svcpt->scp_nrqbds_posted != 0) { + spin_unlock(&svcpt->scp_lock); + /* Network access will complete in finite time but + * the HUGE timeout lets us CWARN for visibility + * of sluggish NALs */ + lwi = LWI_TIMEOUT_INTERVAL( + cfs_time_seconds(LONG_UNLINK), + cfs_time_seconds(1), NULL, NULL); + rc = l_wait_event(svcpt->scp_waitq, + svcpt->scp_nrqbds_posted == 0, &lwi); + if (rc == -ETIMEDOUT) { + CWARN("Service %s waiting for " + "request buffers\n", + svcpt->scp_service->srv_name); + } + spin_lock(&svcpt->scp_lock); + } + spin_unlock(&svcpt->scp_lock); + } +} - rc = LNetMDUnlink(rqbd->rqbd_md_h); - LASSERT (rc == 0 || rc == -ENOENT); - } +static void +ptlrpc_service_purge_all(struct ptlrpc_service *svc) +{ + struct ptlrpc_service_part *svcpt; + struct ptlrpc_request_buffer_desc *rqbd; + struct ptlrpc_request *req; + struct ptlrpc_reply_state *rs; + int i; + + ptlrpc_service_for_each_part(svcpt, i, svc) { + if (svcpt->scp_service == NULL) + break; + + 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); + spin_lock(&rs->rs_lock); + ptlrpc_schedule_difficult_reply(rs); + spin_unlock(&rs->rs_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 + * thread noodling the request queue now */ + while (!cfs_list_empty(&svcpt->scp_req_incoming)) { + req = cfs_list_entry(svcpt->scp_req_incoming.next, + struct ptlrpc_request, rq_list); + + cfs_list_del(&req->rq_list); + svcpt->scp_nreqs_incoming--; + 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)); + LASSERT(svcpt->scp_nreqs_incoming == 0); + LASSERT(svcpt->scp_nreqs_active == 0); + /* history should have been culled by + * ptlrpc_server_finish_request */ + LASSERT(svcpt->scp_hist_nrqbds == 0); + + /* Now free all the request buffers since nothing + * references them any more... */ + + while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) { + rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next, + struct ptlrpc_request_buffer_desc, + rqbd_list); + ptlrpc_free_rqbd(rqbd); + } + ptlrpc_wait_replies(svcpt); + + while (!cfs_list_empty(&svcpt->scp_rep_idle)) { + rs = cfs_list_entry(svcpt->scp_rep_idle.next, + struct ptlrpc_reply_state, + rs_list); + cfs_list_del(&rs->rs_list); + OBD_FREE_LARGE(rs, svc->srv_max_reply_size); + } + } +} - /* Wait for the network to release any buffers it's currently - * filling */ - for (;;) { - spin_lock(&service->srv_lock); - rc = service->srv_nrqbd_receiving; - spin_unlock(&service->srv_lock); +static void +ptlrpc_service_free(struct ptlrpc_service *svc) +{ + struct ptlrpc_service_part *svcpt; + struct ptlrpc_at_array *array; + int i; + + ptlrpc_service_for_each_part(svcpt, i, svc) { + if (svcpt->scp_service == NULL) + break; + + /* In case somebody rearmed this in the meantime */ + cfs_timer_disarm(&svcpt->scp_at_timer); + array = &svcpt->scp_at_array; + + if (array->paa_reqs_array != NULL) { + OBD_FREE(array->paa_reqs_array, + sizeof(cfs_list_t) * array->paa_size); + array->paa_reqs_array = NULL; + } + + if (array->paa_reqs_count != NULL) { + OBD_FREE(array->paa_reqs_count, + sizeof(__u32) * array->paa_size); + array->paa_reqs_count = NULL; + } + } + + ptlrpc_service_for_each_part(svcpt, i, svc) + OBD_FREE_PTR(svcpt); + + if (svc->srv_cpts != NULL) + cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts); + + OBD_FREE(svc, offsetof(struct ptlrpc_service, + srv_parts[svc->srv_ncpts])); +} - if (rc == 0) - break; +int ptlrpc_unregister_service(struct ptlrpc_service *service) +{ + ENTRY; - /* Network access will complete in finite time but the HUGE - * timeout lets us CWARN for visibility of sluggish NALs */ - lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK), - cfs_time_seconds(1), NULL, NULL); - rc = l_wait_event(service->srv_waitq, - service->srv_nrqbd_receiving == 0, - &lwi); - if (rc == -ETIMEDOUT) - CWARN("Service %s waiting for request buffers\n", - service->srv_name); - } + CDEBUG(D_NET, "%s: tearing down\n", service->srv_name); - /* schedule all outstanding replies to terminate them */ - spin_lock(&service->srv_lock); - while (!list_empty(&service->srv_active_replies)) { - struct ptlrpc_reply_state *rs = - list_entry(service->srv_active_replies.next, - struct ptlrpc_reply_state, rs_list); - spin_lock(&rs->rs_lock); - ptlrpc_schedule_difficult_reply(rs); - spin_unlock(&rs->rs_lock); - } - spin_unlock(&service->srv_lock); - - /* purge the request queue. NB No new replies (rqbds all unlinked) - * and no service threads, so I'm the only thread noodling the - * request queue now */ - while (!list_empty(&service->srv_req_in_queue)) { - struct ptlrpc_request *req = - list_entry(service->srv_req_in_queue.next, - struct ptlrpc_request, - rq_list); - - list_del(&req->rq_list); - service->srv_n_queued_reqs--; - service->srv_n_active_reqs++; - ptlrpc_server_finish_request(req); - } - while (ptlrpc_server_request_pending(service, 1)) { - struct ptlrpc_request *req; - - req = ptlrpc_server_request_get(service, 1); - list_del(&req->rq_list); - service->srv_n_queued_reqs--; - service->srv_n_active_reqs++; - ptlrpc_hpreq_fini(req); - ptlrpc_server_finish_request(req); - } - LASSERT(service->srv_n_queued_reqs == 0); - LASSERT(service->srv_n_active_reqs == 0); - LASSERT(service->srv_n_history_rqbds == 0); - LASSERT(list_empty(&service->srv_active_rqbds)); - - /* Now free all the request buffers since nothing references them - * any more... */ - while (!list_empty(&service->srv_idle_rqbds)) { - struct ptlrpc_request_buffer_desc *rqbd = - list_entry(service->srv_idle_rqbds.next, - struct ptlrpc_request_buffer_desc, - rqbd_list); - - ptlrpc_free_rqbd(rqbd); - } + service->srv_is_stopping = 1; - ptlrpc_wait_replies(service); + mutex_lock(&ptlrpc_all_services_mutex); + cfs_list_del_init(&service->srv_list); + mutex_unlock(&ptlrpc_all_services_mutex); - list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) { - list_del(&rs->rs_list); - OBD_FREE(rs, service->srv_max_reply_size); - } + ptlrpc_service_del_atimer(service); + ptlrpc_stop_all_threads(service); - /* In case somebody rearmed this in the meantime */ - cfs_timer_disarm(&service->srv_at_timer); + ptlrpc_service_unlink_rqbd(service); + ptlrpc_service_purge_all(service); + ptlrpc_service_nrs_cleanup(service); - if (array->paa_reqs_array != NULL) { - OBD_FREE(array->paa_reqs_array, - sizeof(struct list_head) * array->paa_size); - array->paa_reqs_array = NULL; - } + ptlrpc_lprocfs_unregister_service(service); - if (array->paa_reqs_count != NULL) { - OBD_FREE(array->paa_reqs_count, - sizeof(__u32) * array->paa_size); - array->paa_reqs_count= NULL; - } + ptlrpc_service_free(service); - OBD_FREE_PTR(service); - RETURN(0); + RETURN(0); } +EXPORT_SYMBOL(ptlrpc_unregister_service); -/* Returns 0 if the service is healthy. +/** + * Returns 0 if the service is healthy. * * 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_request *request; - struct timeval right_now; - long timediff; - - if (svc == NULL) - return 0; + struct ptlrpc_request *request = NULL; + struct timeval right_now; + long timediff; + + do_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); + + if (request == NULL) { + spin_unlock(&svcpt->scp_req_lock); + return 0; + } + + timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL); + spin_unlock(&svcpt->scp_req_lock); + + if ((timediff / ONE_MILLION) > + (AT_OFF ? obd_timeout * 3 / 2 : at_max)) { + CERROR("%s: unhealthy - request has been waiting %lds\n", + svcpt->scp_service->srv_name, timediff / ONE_MILLION); + return -1; + } + + return 0; +} - do_gettimeofday(&right_now); +int +ptlrpc_service_health_check(struct ptlrpc_service *svc) +{ + struct ptlrpc_service_part *svcpt; + int i; - spin_lock(&svc->srv_lock); - if (!ptlrpc_server_request_pending(svc, 1)) { - spin_unlock(&svc->srv_lock); - return 0; - } + if (svc == NULL) + return 0; - /* How long has the next entry been waiting? */ - if (list_empty(&svc->srv_request_queue)) - request = list_entry(svc->srv_request_hpq.next, - struct ptlrpc_request, rq_list); - else - request = list_entry(svc->srv_request_queue.next, - struct ptlrpc_request, rq_list); - timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL); - spin_unlock(&svc->srv_lock); - - if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 : - at_max)) { - CERROR("%s: unhealthy - request has been waiting %lds\n", - svc->srv_name, timediff / ONE_MILLION); - return (-1); - } + ptlrpc_service_for_each_part(svcpt, i, svc) { + int rc = ptlrpc_svcpt_health_check(svcpt); - return 0; + if (rc != 0) + return rc; + } + return 0; } +EXPORT_SYMBOL(ptlrpc_service_health_check);