X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fptlrpc%2Fservice.c;h=c95f76e70c9de598e249a2f30c4a745bb4f38554;hb=582c110231cf06bcd7e5e0b3bdf4f2058e18ebe4;hp=d742a1792a62cffacc444355644290d9f841bc71;hpb=31cdf88e8c9de9ec198f6274cdccdc031af20fc1;p=fs%2Flustre-release.git diff --git a/lustre/ptlrpc/service.c b/lustre/ptlrpc/service.c index d742a17..c95f76e 100644 --- a/lustre/ptlrpc/service.c +++ b/lustre/ptlrpc/service.c @@ -1,26 +1,37 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: +/* + * GPL HEADER START * - * Copyright (C) 2002 Cluster File Systems, Inc. + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * - * This file is part of the Lustre file system, http://www.lustre.org - * Lustre is a trademark of Cluster File Systems, Inc. + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 only, + * as published by the Free Software Foundation. * - * You may have signed or agreed to another license before downloading - * this software. If so, you are bound by the terms and conditions - * of that agreement, and the following does not apply to you. See the - * LICENSE file included with this distribution for more information. + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * General Public License version 2 for more details (a copy is included + * in the LICENSE file that accompanied this code). * - * If you did not agree to a different license, then this copy of Lustre - * is open source software; you can redistribute it and/or modify it - * under the terms of version 2 of the GNU General Public License as - * published by the Free Software Foundation. + * You should have received a copy of the GNU General Public License + * version 2 along with this program; If not, see + * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf * - * In either case, Lustre is distributed in the hope that it will be - * useful, but WITHOUT ANY WARRANTY; without even the implied warranty - * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * license text for more details. + * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, + * CA 95054 USA or visit www.sun.com if you need additional information or + * have any questions. * + * GPL HEADER END + */ +/* + * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. + * + * Copyright (c) 2011, 2012, Whamcloud, Inc. + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. */ #define DEBUG_SUBSYSTEM S_RPC @@ -38,127 +49,117 @@ int test_req_buffer_pressure = 0; CFS_MODULE_PARM(test_req_buffer_pressure, "i", int, 0444, "set non-zero to put pressure on request buffer pools"); -unsigned int at_min = 0; CFS_MODULE_PARM(at_min, "i", int, 0644, "Adaptive timeout minimum (sec)"); -unsigned int at_max = 600; -EXPORT_SYMBOL(at_max); CFS_MODULE_PARM(at_max, "i", int, 0644, "Adaptive timeout maximum (sec)"); -unsigned int at_history = 600; CFS_MODULE_PARM(at_history, "i", int, 0644, "Adaptive timeouts remember the slowest event that took place " "within this period (sec)"); -static int at_early_margin = 5; CFS_MODULE_PARM(at_early_margin, "i", int, 0644, "How soon before an RPC deadline to send an early reply"); -static int at_extra = 30; CFS_MODULE_PARM(at_extra, "i", int, 0644, "How much extra time to give with each early reply"); /* forward ref */ -static int ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc); +static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt); +static void ptlrpc_hpreq_fini(struct ptlrpc_request *req); static CFS_LIST_HEAD(ptlrpc_all_services); -spinlock_t ptlrpc_all_services_lock; - -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); -} +cfs_spinlock_t ptlrpc_all_services_lock; struct ptlrpc_request_buffer_desc * -ptlrpc_alloc_rqbd (struct ptlrpc_service *svc) +ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt) { - struct ptlrpc_request_buffer_desc *rqbd; + struct ptlrpc_service *svc = svcpt->scp_service; + struct ptlrpc_request_buffer_desc *rqbd; - OBD_ALLOC_PTR(rqbd); - if (rqbd == NULL) - return (NULL); + OBD_ALLOC_PTR(rqbd); + if (rqbd == NULL) + return NULL; - rqbd->rqbd_service = svc; + rqbd->rqbd_svcpt = svcpt; rqbd->rqbd_refcount = 0; rqbd->rqbd_cbid.cbid_fn = request_in_callback; rqbd->rqbd_cbid.cbid_arg = rqbd; CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs); - rqbd->rqbd_buffer = ptlrpc_alloc_request_buffer(svc->srv_buf_size); + OBD_ALLOC_LARGE(rqbd->rqbd_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); + cfs_spin_lock(&svcpt->scp_lock); + cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle); + svcpt->scp_nrqbds_total++; + cfs_spin_unlock(&svcpt->scp_lock); - return (rqbd); + return rqbd; } void -ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd) +ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd) { - struct ptlrpc_service *svc = rqbd->rqbd_service; + struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt; - LASSERT (rqbd->rqbd_refcount == 0); - LASSERT (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); + cfs_spin_lock(&svcpt->scp_lock); + cfs_list_del(&rqbd->rqbd_list); + svcpt->scp_nrqbds_total--; + cfs_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) { + 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); for (i = 0; i < svc->srv_nbuf_per_group; i++) { - rqbd = ptlrpc_alloc_rqbd(svc); + /* NB: another thread might be doing this as well, we need to + * make sure that 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); - } + if (ptlrpc_server_post_idle_rqbds(svcpt) < 0) { + rc = -EAGAIN; + break; + } + } + + CDEBUG(D_RPCTRACE, + "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n", + svc->srv_name, i, svc->srv_buf_size, + svcpt->scp_nrqbds_posted, svcpt->scp_nrqbds_total, rc); - 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) +ptlrpc_save_lock(struct ptlrpc_request *req, + struct lustre_handle *lock, int mode, int no_ack) { struct ptlrpc_reply_state *rs = req->rq_reply_state; int idx; @@ -166,219 +167,493 @@ ptlrpc_save_lock (struct ptlrpc_request *req, LASSERT(rs != NULL); LASSERT(rs->rs_nlocks < RS_MAX_LOCKS); - idx = rs->rs_nlocks++; - rs->rs_locks[idx] = *lock; - rs->rs_modes[idx] = mode; - rs->rs_difficult = 1; + if (req->rq_export->exp_disconnected) { + ldlm_lock_decref(lock, mode); + } else { + idx = rs->rs_nlocks++; + rs->rs_locks[idx] = *lock; + rs->rs_modes[idx] = mode; + rs->rs_difficult = 1; + rs->rs_no_ack = !!no_ack; + } } -void -ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs) +#ifdef __KERNEL__ + +#define HRT_RUNNING 0 +#define HRT_STOPPING 1 + +struct ptlrpc_hr_thread { + cfs_spinlock_t hrt_lock; + unsigned long hrt_flags; + cfs_waitq_t hrt_wait; + cfs_list_t hrt_queue; + cfs_completion_t hrt_completion; +}; + +struct ptlrpc_hr_service { + int hr_index; + int hr_n_threads; + int hr_size; + struct ptlrpc_hr_thread hr_threads[0]; +}; + +struct rs_batch { + 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; + +/** + * maximum mumber of replies scheduled in one batch + */ +#define MAX_SCHEDULED 256 + +/** + * Initialize a reply batch. + * + * \param b batch + */ +static void rs_batch_init(struct rs_batch *b) { - struct ptlrpc_service *svc = rs->rs_service; + memset(b, 0, sizeof *b); + CFS_INIT_LIST_HEAD(&b->rsb_replies); +} -#ifdef CONFIG_SMP - LASSERT (spin_is_locked (&svc->srv_lock)); +/** + * Choose an hr thread to dispatch requests to. + */ +static unsigned int get_hr_thread_index(struct ptlrpc_hr_service *hr) +{ + unsigned int idx; + + /* Concurrent modification of hr_index w/o any spinlock + protection is harmless as long as the result fits + [0..(hr_n_threads-1)] range and each thread gets near equal + load. */ + idx = hr->hr_index; + hr->hr_index = (idx >= hr->hr_n_threads - 1) ? 0 : idx + 1; + return idx; +} + +/** + * Dispatch all replies accumulated in the batch to one from + * dedicated reply handling threads. + * + * \param b batch + */ +static void rs_batch_dispatch(struct rs_batch *b) +{ + if (b->rsb_n_replies != 0) { + struct ptlrpc_hr_service *hr = ptlrpc_hr; + int idx; + + idx = get_hr_thread_index(hr); + + cfs_spin_lock(&hr->hr_threads[idx].hrt_lock); + cfs_list_splice_init(&b->rsb_replies, + &hr->hr_threads[idx].hrt_queue); + cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock); + cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait); + b->rsb_n_replies = 0; + } +} + +/** + * Add a reply to a batch. + * Add one reply object to a batch, schedule batched replies if overload. + * + * \param b batch + * \param rs reply + */ +static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs) +{ + 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); + cfs_spin_unlock(&b->rsb_svcpt->scp_rep_lock); + } + cfs_spin_lock(&svcpt->scp_rep_lock); + b->rsb_svcpt = svcpt; + } + cfs_spin_lock(&rs->rs_lock); + rs->rs_scheduled_ever = 1; + if (rs->rs_scheduled == 0) { + cfs_list_move(&rs->rs_list, &b->rsb_replies); + rs->rs_scheduled = 1; + b->rsb_n_replies++; + } + rs->rs_committed = 1; + cfs_spin_unlock(&rs->rs_lock); +} + +/** + * Reply batch finalization. + * Dispatch remaining replies from the batch + * and release remaining spinlock. + * + * \param b batch + */ +static void rs_batch_fini(struct rs_batch *b) +{ + if (b->rsb_svcpt != NULL) { + rs_batch_dispatch(b); + cfs_spin_unlock(&b->rsb_svcpt->scp_rep_lock); + } +} + +#define DECLARE_RS_BATCH(b) struct rs_batch b + +#else /* __KERNEL__ */ + +#define rs_batch_init(b) do{}while(0) +#define rs_batch_fini(b) do{}while(0) +#define rs_batch_add(b, r) ptlrpc_schedule_difficult_reply(r) +#define DECLARE_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; + + LASSERT(cfs_list_empty(&rs->rs_list)); + + idx = get_hr_thread_index(hr); + cfs_spin_lock(&hr->hr_threads[idx].hrt_lock); + cfs_list_add_tail(&rs->rs_list, &hr->hr_threads[idx].hrt_queue); + cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock); + cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait); + EXIT; +#else + cfs_list_add_tail(&rs->rs_list, &rs->rs_svcpt->scp_rep_queue); #endif +} + +void +ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs) +{ + ENTRY; + + LASSERT_SPIN_LOCKED(&rs->rs_svcpt->scp_rep_lock); + LASSERT_SPIN_LOCKED(&rs->rs_lock); LASSERT (rs->rs_difficult); - rs->rs_scheduled_ever = 1; /* flag any notification attempt */ + rs->rs_scheduled_ever = 1; /* flag any notification attempt */ - if (rs->rs_scheduled) /* being set up or already notified */ + if (rs->rs_scheduled) { /* being set up or already notified */ + EXIT; return; + } rs->rs_scheduled = 1; - list_del (&rs->rs_list); - list_add (&rs->rs_list, &svc->srv_reply_queue); - cfs_waitq_signal (&svc->srv_waitq); + cfs_list_del_init(&rs->rs_list); + ptlrpc_dispatch_difficult_reply(rs); + EXIT; } -void -ptlrpc_commit_replies (struct obd_device *obd) +void ptlrpc_commit_replies(struct obd_export *exp) { - struct list_head *tmp; - struct list_head *nxt; + struct ptlrpc_reply_state *rs, *nxt; + DECLARE_RS_BATCH(batch); + ENTRY; + rs_batch_init(&batch); /* Find any replies that have been committed and get their service * to attend to complete them. */ /* CAVEAT EMPTOR: spinlock ordering!!! */ - spin_lock(&obd->obd_uncommitted_replies_lock); - - list_for_each_safe (tmp, nxt, &obd->obd_uncommitted_replies) { - struct ptlrpc_reply_state *rs = - list_entry(tmp, struct ptlrpc_reply_state, rs_obd_list); - + cfs_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); - - if (rs->rs_transno <= obd->obd_last_committed) { - struct ptlrpc_service *svc = rs->rs_service; - - spin_lock (&svc->srv_lock); - list_del_init (&rs->rs_obd_list); - ptlrpc_schedule_difficult_reply (rs); - spin_unlock (&svc->srv_lock); + /* VBR: per-export last_committed */ + LASSERT(rs->rs_export); + if (rs->rs_transno <= exp->exp_last_committed) { + cfs_list_del_init(&rs->rs_obd_list); + rs_batch_add(&batch, rs); } } - - spin_unlock(&obd->obd_uncommitted_replies_lock); + cfs_spin_unlock(&exp->exp_uncommitted_replies_lock); + rs_batch_fini(&batch); + EXIT; } 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 (;;) { + cfs_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)) { + cfs_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); + cfs_spin_unlock(&svcpt->scp_lock); - rc = ptlrpc_register_rqbd(rqbd); - if (rc != 0) - break; + rc = ptlrpc_register_rqbd(rqbd); + if (rc != 0) + break; - posted = 1; - } + posted = 1; + } - spin_lock(&svc->srv_lock); + cfs_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); + cfs_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); + struct ptlrpc_service_part *svcpt; + + svcpt = (struct ptlrpc_service_part *)castmeharder; + + svcpt->scp_at_check = 1; + svcpt->scp_at_checktime = cfs_time_current(); + cfs_waitq_signal(&svcpt->scp_waitq); } -EXPORT_SYMBOL(ptlrpc_init_svc_conf); -static void ptlrpc_at_timer(unsigned long castmeharder) +static void +ptlrpc_server_nthreads_check(struct ptlrpc_service_conf *conf, + int *min_p, int *max_p) { - struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder; - CDEBUG(D_INFO, "at timer %s hit at %ld%s\n", - svc->srv_name, cfs_time_current_sec(), - list_empty(&svc->srv_at_list) ? ", empty" : ""); - 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; + int nthrs_min; + int nthrs; + + nthrs_min = PTLRPC_NTHRS_MIN + (conf->psc_ops.so_hpreq_handler != NULL); + nthrs_min = max_t(int, nthrs_min, tc->tc_nthrs_min); + + nthrs = tc->tc_nthrs_user; + if (nthrs != 0) { /* validate it */ + nthrs = min_t(int, nthrs, tc->tc_nthrs_max); + nthrs = max_t(int, nthrs, nthrs_min); + *min_p = *max_p = nthrs; + return; + } + + /* + * NB: we will add some common code here for estimating, for example: + * add a new member ptlrpc_service_thr_conf::tc_factor, and estimate + * threads number based on: + * (online_cpus * conf::tc_factor) + conf::tc_nthrs_base. + * + * So we can remove code block like estimation in ost_setup, also, + * we might estimate MDS threads number as well instead of using + * absolute number, and have more threads on fat servers to improve + * availability of service. + * + * Also, we will need to validate threads number at here for + * CPT affinity service (CPU ParTion) in the future. + * A service can have percpt thread-pool instead of a global thread + * pool for each service, which means user might not always get the + * threads number they want even they set it in conf::tc_nthrs_user, + * because we need to adjust threads number for each CPT, instead of + * just use (conf::tc_nthrs_user / NCPTS), to make sure each pool + * will be healthy. + */ + *max_p = tc->tc_nthrs_max; + *min_p = nthrs_min; +#else /* __KERNEL__ */ + *max_p = *min_p = 1; /* whatever */ +#endif } -/* @threadname should be 11 characters or less - 3 will be added 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) +/** + * Initialize percpt data for a service + */ +static int +ptlrpc_service_part_init(struct ptlrpc_service *svc, + struct ptlrpc_service_part *svcpt) { - int rc; - struct ptlrpc_service *service; - ENTRY; + struct ptlrpc_at_array *array; + int size; + int index; + int rc; + + CFS_INIT_LIST_HEAD(&svcpt->scp_threads); + + /* rqbd and incoming request queue */ + cfs_spin_lock_init(&svcpt->scp_lock); + CFS_INIT_LIST_HEAD(&svcpt->scp_rqbd_idle); + CFS_INIT_LIST_HEAD(&svcpt->scp_rqbd_posted); + CFS_INIT_LIST_HEAD(&svcpt->scp_req_incoming); + cfs_waitq_init(&svcpt->scp_waitq); + /* history request & rqbd list */ + CFS_INIT_LIST_HEAD(&svcpt->scp_hist_reqs); + CFS_INIT_LIST_HEAD(&svcpt->scp_hist_rqbds); + + /* acitve requests and hp requests */ + cfs_spin_lock_init(&svcpt->scp_req_lock); + CFS_INIT_LIST_HEAD(&svcpt->scp_req_pending); + CFS_INIT_LIST_HEAD(&svcpt->scp_hreq_pending); + + /* reply states */ + cfs_spin_lock_init(&svcpt->scp_rep_lock); + CFS_INIT_LIST_HEAD(&svcpt->scp_rep_active); +#ifndef __KERNEL__ + CFS_INIT_LIST_HEAD(&svcpt->scp_rep_queue); +#endif + CFS_INIT_LIST_HEAD(&svcpt->scp_rep_idle); + cfs_waitq_init(&svcpt->scp_rep_waitq); + cfs_atomic_set(&svcpt->scp_nreps_difficult, 0); + + /* adaptive timeout */ + cfs_spin_lock_init(&svcpt->scp_at_lock); + array = &svcpt->scp_at_array; + + size = at_est2timeout(at_max); + array->paa_size = size; + array->paa_count = 0; + array->paa_deadline = -1; + + /* allocate memory for scp_at_array (ptlrpc_at_array) */ + OBD_ALLOC(array->paa_reqs_array, sizeof(cfs_list_t) * size); + if (array->paa_reqs_array == NULL) + return -ENOMEM; + + for (index = 0; index < size; index++) + CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]); + + OBD_ALLOC(array->paa_reqs_count, sizeof(__u32) * size); + if (array->paa_reqs_count == NULL) + goto failed; + + cfs_timer_init(&svcpt->scp_at_timer, ptlrpc_at_timer, svcpt); + /* At SOW, service time should be quick; 10s seems generous. If client + * timeout is less than this, we'll be sending an early reply. */ + at_init(&svcpt->scp_at_estimate, 10, 0); + + /* assign this before call ptlrpc_grow_req_bufs */ + svcpt->scp_service = svc; + /* Now allocate the request buffers, but don't post them now */ + rc = ptlrpc_grow_req_bufs(svcpt); + /* We shouldn't be under memory pressure at startup, so + * fail if we can't allocate all our buffers at this time. */ + if (rc != 0) + goto failed; + + return 0; + + failed: + if (array->paa_reqs_count != NULL) { + OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size); + array->paa_reqs_count = NULL; + } + + if (array->paa_reqs_array != NULL) { + OBD_FREE(array->paa_reqs_array, + sizeof(cfs_list_t) * array->paa_size); + array->paa_reqs_array = NULL; + } + + return -ENOMEM; +} - 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; - - rc = LNetSetLazyPortal(service->srv_req_portal); - LASSERT (rc == 0); - - CFS_INIT_LIST_HEAD(&service->srv_request_queue); - 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); - CFS_INIT_LIST_HEAD(&service->srv_reply_queue); - CFS_INIT_LIST_HEAD(&service->srv_free_rs_list); - cfs_waitq_init(&service->srv_free_rs_waitq); - - spin_lock_init(&service->srv_at_lock); - CFS_INIT_LIST_HEAD(&service->srv_req_in_queue); - CFS_INIT_LIST_HEAD(&service->srv_at_list); - 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; +/** + * Initialize service on a given portal. + * This includes starting serving threads , allocating and posting rqbds and + * so on. + */ +struct ptlrpc_service * +ptlrpc_register_service(struct ptlrpc_service_conf *conf, + cfs_proc_dir_entry_t *proc_entry) +{ + struct ptlrpc_service *service; + int rc; + 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); + + OBD_ALLOC_PTR(service); + if (service == NULL) + RETURN(ERR_PTR(-ENOMEM)); + + /* public members */ + cfs_spin_lock_init(&service->srv_lock); + service->srv_name = conf->psc_name; + service->srv_watchdog_factor = conf->psc_watchdog_factor; + CFS_INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */ + + /* buffer configuration */ + service->srv_nbuf_per_group = test_req_buffer_pressure ? + 1 : conf->psc_buf.bc_nbufs; + service->srv_max_req_size = conf->psc_buf.bc_req_max_size + + 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; + + ptlrpc_server_nthreads_check(conf, &service->srv_threads_min, + &service->srv_threads_max); + + service->srv_thread_name = conf->psc_thr.tc_thr_name; + service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags; + service->srv_cpu_affinity = !!conf->psc_thr.tc_cpu_affinity; + service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO; + service->srv_ops = conf->psc_ops; + + OBD_ALLOC_PTR(service->srv_part); + if (service->srv_part == NULL) + GOTO(failed, rc = -ENOMEM); + + rc = ptlrpc_service_part_init(service, service->srv_part); + if (rc != 0) + GOTO(failed, rc); + + rc = LNetSetLazyPortal(service->srv_req_portal); + LASSERT(rc == 0); + + cfs_spin_lock (&ptlrpc_all_services_lock); + cfs_list_add (&service->srv_list, &ptlrpc_all_services); + cfs_spin_unlock (&ptlrpc_all_services_lock); if (proc_entry != NULL) ptlrpc_lprocfs_register_service(proc_entry, service); @@ -386,125 +661,181 @@ ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size, CDEBUG(D_NET, "%s: Started, listening on portal %d\n", service->srv_name, service->srv_req_portal); - RETURN(service); +#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 + + RETURN(service); failed: - ptlrpc_unregister_service(service); - return NULL; + ptlrpc_unregister_service(service); + RETURN(ERR_PTR(rc)); } -static void ptlrpc_server_req_decref(struct ptlrpc_request *req) +/** + * to actually free the request, must be called without holding svc_lock. + * note it's caller's responsibility to unlink req->rq_list. + */ +static void ptlrpc_server_free_request(struct ptlrpc_request *req) { - struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd; + LASSERT(cfs_atomic_read(&req->rq_refcount) == 0); + LASSERT(cfs_list_empty(&req->rq_timed_list)); - if (!atomic_dec_and_test(&req->rq_refcount)) - return; + /* DEBUG_REQ() assumes the reply state of a request with a valid + * ref will not be destroyed until that reference is dropped. */ + ptlrpc_req_drop_rs(req); sptlrpc_svc_ctx_decref(req); - LASSERT(list_empty(&req->rq_timed_list)); - if (req != &rqbd->rqbd_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)); - } else { - struct ptlrpc_service *svc = rqbd->rqbd_service; - /* schedule request buffer for re-use. - * NB I can only do this after I've disposed of their - * reqs; particularly the embedded req */ - spin_lock(&svc->srv_lock); - list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds); - spin_unlock(&svc->srv_lock); } } -static void __ptlrpc_server_free_request(struct ptlrpc_request *req) -{ - list_del(&req->rq_list); - ptlrpc_req_drop_rs(req); - ptlrpc_server_req_decref(req); -} - -static void -ptlrpc_server_free_request(struct ptlrpc_request *req) +/** + * drop a reference count of the request. if it reaches 0, we either + * put it into history list, or free it immediately. + */ +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 (!cfs_atomic_dec_and_test(&req->rq_refcount)) + return; - if (req->rq_phase != RQ_PHASE_NEW) /* incorrect message magic */ - DEBUG_REQ(D_INFO, req, "free req"); - spin_lock(&svc->srv_at_lock); - req->rq_sent_final = 1; - list_del_init(&req->rq_timed_list); - spin_unlock(&svc->srv_at_lock); + cfs_spin_lock(&svcpt->scp_at_lock); + if (req->rq_at_linked) { + struct ptlrpc_at_array *array = &svcpt->scp_at_array; + __u32 index = req->rq_at_index; - spin_lock(&svc->srv_lock); + LASSERT(!cfs_list_empty(&req->rq_timed_list)); + cfs_list_del_init(&req->rq_timed_list); + cfs_spin_lock(&req->rq_lock); + req->rq_at_linked = 0; + cfs_spin_unlock(&req->rq_lock); + array->paa_reqs_count[index]--; + array->paa_count--; + } else + LASSERT(cfs_list_empty(&req->rq_timed_list)); - svc->srv_n_active_reqs--; - list_add(&req->rq_list, &rqbd->rqbd_reqs); + cfs_spin_unlock(&svcpt->scp_at_lock); + + /* finalize request */ + if (req->rq_export) { + class_export_put(req->rq_export); + req->rq_export = NULL; + } + + cfs_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); - /* 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); + cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds); + svcpt->scp_hist_nrqbds++; - list_del(&rqbd->rqbd_list); - svc->srv_n_history_rqbds--; + /* cull some history? + * I expect only about 1 or 2 rqbds need to be recycled here */ + while (svcpt->scp_hist_nrqbds > svc->srv_max_history_rqbds) { + rqbd = cfs_list_entry(svcpt->scp_hist_rqbds.next, + struct ptlrpc_request_buffer_desc, + rqbd_list); + + cfs_list_del(&rqbd->rqbd_list); + 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); + if (req->rq_history_seq > + svcpt->scp_hist_seq_culled) { + svcpt->scp_hist_seq_culled = + req->rq_history_seq; + } + cfs_list_del(&req->rq_history_list); + } + + cfs_spin_unlock(&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_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); - __ptlrpc_server_free_request(req); - } + cfs_spin_lock(&svcpt->scp_lock); + /* + * now all reqs including the embedded req has been + * disposed, schedule request buffer for re-use. + */ + LASSERT(cfs_atomic_read(&rqbd->rqbd_req.rq_refcount) == + 0); + cfs_list_add_tail(&rqbd->rqbd_list, + &svcpt->scp_rqbd_idle); + } + + cfs_spin_unlock(&svcpt->scp_lock); + } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) { + /* If we are low on memory, we are not interested in history */ + cfs_list_del(&req->rq_list); + cfs_list_del_init(&req->rq_history_list); + + cfs_spin_unlock(&svcpt->scp_lock); + + ptlrpc_server_free_request(req); + } else { + cfs_spin_unlock(&svcpt->scp_lock); + } +} - spin_lock(&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_history_list); - __ptlrpc_server_free_request(req); - } +/** + * to finish a request: stop sending more early replies, and release + * the request. should be called after we finished handling the request. + */ +static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt, + struct ptlrpc_request *req) +{ + ptlrpc_hpreq_fini(req); - spin_unlock(&svc->srv_lock); + cfs_spin_lock(&svcpt->scp_req_lock); + svcpt->scp_nreqs_active--; + if (req->rq_hp) + svcpt->scp_nhreqs_active--; + cfs_spin_unlock(&svcpt->scp_req_lock); + ptlrpc_server_drop_request(req); } -/* 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; - time_t oldest_time; + time_t oldest_time, new_time; ENTRY; @@ -515,9 +846,13 @@ static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay) of the list, we can be really lazy here - we don't have to evict at the exact right moment. Eventually, all silent exports will make it to the top of the list. */ - exp->exp_last_request_time = max(exp->exp_last_request_time, - cfs_time_current_sec() + extra_delay); + /* Do not pay attention on 1sec or smaller renewals. */ + new_time = cfs_time_current_sec() + extra_delay; + if (exp->exp_last_request_time + 1 /*second */ >= new_time) + RETURN_EXIT; + + exp->exp_last_request_time = new_time; CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n", exp->exp_client_uuid.uuid, exp->exp_last_request_time, exp); @@ -525,22 +860,21 @@ 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); + cfs_spin_lock(&exp->exp_obd->obd_dev_lock); - if (list_empty(&exp->exp_obd_chain_timed)) { + if (cfs_list_empty(&exp->exp_obd_chain_timed)) { /* this one is not timed */ - spin_unlock(&exp->exp_obd->obd_dev_lock); - EXIT; - return; + cfs_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); + cfs_spin_unlock(&exp->exp_obd->obd_dev_lock); if (exp->exp_obd->obd_recovering) { /* be nice to everyone during recovery */ @@ -560,8 +894,8 @@ 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() > @@ -577,28 +911,14 @@ static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay) EXIT; } -#ifndef __KERNEL__ -int lu_context_init(struct lu_context *ctx, __u32 tags) -{ - return 0; -} - -void lu_context_fini(struct lu_context *ctx) -{ -} - -void lu_context_enter(struct lu_context *ctx) -{ -} - -void lu_context_exit(struct lu_context *ctx) -{ -} - -#endif - +/** + * Sanity check request \a req. + * Return 0 if all is ok, error code otherwise. + */ static int ptlrpc_check_req(struct ptlrpc_request *req) { + int rc = 0; + if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) < req->rq_export->exp_conn_cnt)) { DEBUG_REQ(D_ERROR, req, @@ -613,44 +933,62 @@ static int ptlrpc_check_req(struct ptlrpc_request *req) 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_request *rq; - __s32 next; - - spin_lock(&svc->srv_at_lock); - if (list_empty(&svc->srv_at_list)) { - cfs_timer_disarm(&svc->srv_at_timer); - spin_unlock(&svc->srv_at_lock); - return; - } - - /* Set timer for closest deadline */ - rq = list_entry(svc->srv_at_list.next, struct ptlrpc_request, - rq_timed_list); - next = (__s32)(rq->rq_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; + + cfs_spin_lock(&svcpt->scp_at_lock); + if (array->paa_count == 0) { + cfs_timer_disarm(&svcpt->scp_at_timer); + cfs_spin_unlock(&svcpt->scp_at_lock); + return; + } + + /* Set timer for closest deadline */ + next = (__s32)(array->paa_deadline - cfs_time_current_sec() - + at_early_margin); + if (next <= 0) { + ptlrpc_at_timer((unsigned long)svcpt); + } else { + cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next)); + CDEBUG(D_INFO, "armed %s at %+ds\n", + svcpt->scp_service->srv_name, next); + } + cfs_spin_unlock(&svcpt->scp_at_lock); } /* Add rpc to early reply check list */ static int ptlrpc_at_add_timed(struct ptlrpc_request *req) { - struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service; - struct ptlrpc_request *rq; + struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt; + struct ptlrpc_at_array *array = &svcpt->scp_at_array; + struct ptlrpc_request *rq = NULL; + __u32 index; int found = 0; if (AT_OFF) @@ -662,48 +1000,53 @@ static int ptlrpc_at_add_timed(struct ptlrpc_request *req) if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0) return(-ENOSYS); - DEBUG_REQ(D_ADAPTTO, req, "add timed %lds", - req->rq_deadline - cfs_time_current_sec()); - - 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)); - /* Add to sorted list. Presumably latest rpcs will have the latest - deadlines, so search backward. */ - list_for_each_entry_reverse(rq, &svc->srv_at_list, rq_timed_list) { - if (req->rq_deadline > rq->rq_deadline) { - list_add(&req->rq_timed_list, &rq->rq_timed_list); - found++; - break; + cfs_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. */ + cfs_list_for_each_entry_reverse(rq, + &array->paa_reqs_array[index], + rq_timed_list) { + if (req->rq_deadline >= rq->rq_deadline) { + cfs_list_add(&req->rq_timed_list, + &rq->rq_timed_list); + break; + } } } - if (!found) - /* Add to front if shortest deadline or list empty */ - list_add(&req->rq_timed_list, &svc->srv_at_list); - /* Check if we're the head of the list */ - found = (svc->srv_at_list.next == &req->rq_timed_list); + /* Add the request at the head of the list */ + if (cfs_list_empty(&req->rq_timed_list)) + cfs_list_add(&req->rq_timed_list, + &array->paa_reqs_array[index]); - spin_unlock(&svc->srv_at_lock); + cfs_spin_lock(&req->rq_lock); + req->rq_at_linked = 1; + cfs_spin_unlock(&req->rq_lock); + req->rq_at_index = index; + array->paa_reqs_count[index]++; + array->paa_count++; + if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) { + array->paa_deadline = req->rq_deadline; + found = 1; + } + cfs_spin_unlock(&svcpt->scp_at_lock); - if (found) - ptlrpc_at_set_timer(svc); + if (found) + ptlrpc_at_set_timer(svcpt); - return 0; + return 0; } -static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req, - 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; - long olddl = req->rq_deadline - cfs_time_current_sec(); + cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec(); time_t newdl; int rc; ENTRY; @@ -713,48 +1056,63 @@ 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); if (olddl < 0) { - CDEBUG(D_WARNING, "x"LPU64": Already past deadline (%+lds), not" - " sending early reply. Increase at_early_margin (%d)?\n", - req->rq_xid, olddl, at_early_margin); + DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), " + "not sending early reply. Consider increasing " + "at_early_margin (%d)?", olddl, at_early_margin); + /* Return an error so we're not re-added to the timed list. */ RETURN(-ETIMEDOUT); } if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){ - CDEBUG(D_INFO, "Wanted to ask client for more time, but no AT " - "support\n"); + DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, " + "but no AT support"); RETURN(-ENOSYS); } - 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) */ - CDEBUG(D_ADAPTTO, "x"LPU64": Couldn't add any time (%ld/%ld), " - "not sending early reply\n", req->rq_xid, olddl, - newdl - cfs_time_current_sec()); - RETURN(-ETIMEDOUT); - } + 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); OBD_ALLOC(reqcopy, sizeof *reqcopy); if (reqcopy == NULL) RETURN(-ENOMEM); - OBD_ALLOC(reqmsg, req->rq_reqlen); + OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen); if (!reqmsg) { OBD_FREE(reqcopy, sizeof *reqcopy); RETURN(-ENOMEM); @@ -771,10 +1129,12 @@ 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) { - CDEBUG(D_ADAPTTO, "x"LPU64": normal reply already sent out, " - "abort sending early reply\n", req->rq_xid); - GOTO(out, rc = 0); + 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 = -EINVAL); } /* Connection ref */ @@ -812,62 +1172,91 @@ out_put: class_export_put(reqcopy->rq_export); out: sptlrpc_svc_ctx_decref(reqcopy); - OBD_FREE(reqmsg, req->rq_reqlen); + OBD_FREE_LARGE(reqmsg, req->rq_reqlen); OBD_FREE(reqcopy, sizeof *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; + cfs_list_t work_list; + __u32 index, count; + time_t deadline; time_t now = cfs_time_current_sec(); cfs_duration_t delay; 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 (list_empty(&svc->srv_at_list)) { - spin_unlock(&svc->srv_at_lock); - RETURN(0); - } - - /* The timer went off, but maybe the nearest rpc already completed. */ - rq = list_entry(svc->srv_at_list.next, struct ptlrpc_request, - rq_timed_list); - first = (int)(rq->rq_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); + cfs_spin_lock(&svcpt->scp_at_lock); + if (svcpt->scp_at_check == 0) { + cfs_spin_unlock(&svcpt->scp_at_lock); + RETURN(0); + } + delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime); + svcpt->scp_at_check = 0; + + if (array->paa_count == 0) { + cfs_spin_unlock(&svcpt->scp_at_lock); + RETURN(0); + } + + /* The timer went off, but maybe the nearest rpc already completed. */ + first = array->paa_deadline - now; + if (first > at_early_margin) { + /* We've still got plenty of time. Reset the timer. */ + cfs_spin_unlock(&svcpt->scp_at_lock); + ptlrpc_at_set_timer(svcpt); 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. */ CFS_INIT_LIST_HEAD(&work_list); - list_for_each_entry_safe(rq, n, &svc->srv_at_list, rq_timed_list) { - if (rq->rq_deadline <= now + at_early_margin) { - list_move_tail(&rq->rq_timed_list, &work_list); - counter++; - } else { + deadline = -1; + index = (unsigned long)array->paa_deadline % array->paa_size; + count = array->paa_count; + while (count > 0) { + count -= array->paa_reqs_count[index]; + cfs_list_for_each_entry_safe(rq, n, + &array->paa_reqs_array[index], + rq_timed_list) { + if (rq->rq_deadline <= now + at_early_margin) { + cfs_list_del_init(&rq->rq_timed_list); + /** + * ptlrpc_server_drop_request() may drop + * refcount to 0 already. Let's check this and + * don't add entry to work_list + */ + if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount))) + cfs_list_add(&rq->rq_timed_list, &work_list); + counter++; + array->paa_reqs_count[index]--; + array->paa_count--; + cfs_spin_lock(&rq->rq_lock); + rq->rq_at_linked = 0; + cfs_spin_unlock(&rq->rq_lock); + continue; + } + + /* update the earliest deadline */ + if (deadline == -1 || rq->rq_deadline < deadline) + deadline = rq->rq_deadline; + break; } - } - spin_unlock(&svc->srv_at_lock); + if (++index >= array->paa_size) + index = 0; + } + array->paa_deadline = deadline; + cfs_spin_unlock(&svcpt->scp_at_lock); - /* we have a new earliest deadline, restart the timer */ - ptlrpc_at_set_timer(svc); + /* we have a new earliest deadline, restart the timer */ + ptlrpc_at_set_timer(svcpt); CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early " "replies\n", first, at_extra, counter); @@ -875,60 +1264,292 @@ 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_free_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_req_decref(rq); - spin_lock(&svc->srv_at_lock); + ptlrpc_server_drop_request(rq); } - spin_unlock(&svc->srv_at_lock); - RETURN(0); + RETURN(1); /* return "did_something" for liblustre */ } -/* 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) +/** + * 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) { - struct ptlrpc_request *req; - __u32 deadline; - int rc; + int rc = 0; ENTRY; - LASSERT(svc); + if (svc->srv_ops.so_hpreq_handler) { + rc = svc->srv_ops.so_hpreq_handler(req); + if (rc) + RETURN(rc); + } + if (req->rq_export && req->rq_ops) { + /* Perform request specific check. We should do this check + * before the request is added into exp_hp_rpcs list otherwise + * it may hit swab race at LU-1044. */ + if (req->rq_ops->hpreq_check) + rc = req->rq_ops->hpreq_check(req); + + cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock); + cfs_list_add(&req->rq_exp_list, + &req->rq_export->exp_hp_rpcs); + cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock); + } - spin_lock(&svc->srv_lock); - if (list_empty(&svc->srv_req_in_queue)) { - spin_unlock(&svc->srv_lock); - RETURN(0); + RETURN(rc); +} + +/** Remove the request from the export list. */ +static void ptlrpc_hpreq_fini(struct ptlrpc_request *req) +{ + ENTRY; + if (req->rq_export && req->rq_ops) { + /* refresh lock timeout again so that client has more + * room to send lock cancel RPC. */ + if (req->rq_ops->hpreq_fini) + req->rq_ops->hpreq_fini(req); + + cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock); + cfs_list_del_init(&req->rq_exp_list); + cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock); } + EXIT; +} + +/** + * Make the request a high priority one. + * + * All the high priority requests are queued in a separate FIFO + * ptlrpc_service_part::scp_hpreq_pending list which is parallel to + * ptlrpc_service_part::scp_req_pending list but has a higher priority + * for handling. + * + * \see ptlrpc_server_handle_request(). + */ +static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service_part *svcpt, + struct ptlrpc_request *req) +{ + ENTRY; + + cfs_spin_lock(&req->rq_lock); + if (req->rq_hp == 0) { + int opc = lustre_msg_get_opc(req->rq_reqmsg); + + /* Add to the high priority queue. */ + cfs_list_move_tail(&req->rq_list, &svcpt->scp_hreq_pending); + req->rq_hp = 1; + if (opc != OBD_PING) + DEBUG_REQ(D_RPCTRACE, req, "high priority req"); + } + cfs_spin_unlock(&req->rq_lock); + EXIT; +} + +/** + * \see ptlrpc_hpreq_reorder_nolock + */ +void ptlrpc_hpreq_reorder(struct ptlrpc_request *req) +{ + struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt; + ENTRY; + + cfs_spin_lock(&svcpt->scp_req_lock); + /* It may happen that the request is already taken for the processing + * but still in the export list, or the request is not in the request + * queue but in the export list already, do not add it into the + * HP list. */ + if (!cfs_list_empty(&req->rq_list)) + ptlrpc_hpreq_reorder_nolock(svcpt, req); + cfs_spin_unlock(&svcpt->scp_req_lock); + EXIT; +} + +/** Check if the request is a high priority one. */ +static int ptlrpc_server_hpreq_check(struct ptlrpc_service *svc, + struct ptlrpc_request *req) +{ + ENTRY; + + /* Check by request opc. */ + if (OBD_PING == lustre_msg_get_opc(req->rq_reqmsg)) + RETURN(1); + + RETURN(ptlrpc_hpreq_init(svc, req)); +} + +/** Check if a request is a high priority one. */ +static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt, + struct ptlrpc_request *req) +{ + int rc; + ENTRY; + + rc = ptlrpc_server_hpreq_check(svcpt->scp_service, req); + if (rc < 0) + RETURN(rc); + + cfs_spin_lock(&svcpt->scp_req_lock); + + if (rc) + ptlrpc_hpreq_reorder_nolock(svcpt, req); + else + cfs_list_add_tail(&req->rq_list, &svcpt->scp_req_pending); + + cfs_spin_unlock(&svcpt->scp_req_lock); + + RETURN(0); +} + +/** + * 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 int ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt, + int force) +{ + if (force) + return 1; + + if (svcpt->scp_nreqs_active >= svcpt->scp_nthrs_running - 1) + return 0; + + return cfs_list_empty(&svcpt->scp_req_pending) || + svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio; +} + +static int ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt, + int force) +{ + return ptlrpc_server_allow_high(svcpt, force) && + !cfs_list_empty(&svcpt->scp_hreq_pending); +} + +/** + * 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. + * User can call it w/o any lock but need to hold + * ptlrpc_service_part::scp_req_lock to get reliable result + */ +static int ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt, + int force) +{ +#ifndef __KERNEL__ + if (1) /* always allow to handle normal request for liblustre */ + return 1; +#endif + if (force || + svcpt->scp_nreqs_active < svcpt->scp_nthrs_running - 2) + return 1; + + if (svcpt->scp_nreqs_active >= svcpt->scp_nthrs_running - 1) + return 0; + + return svcpt->scp_nhreqs_active > 0 || + svcpt->scp_service->srv_ops.so_hpreq_handler == NULL; +} - 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); +static int ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt, + int force) +{ + return ptlrpc_server_allow_normal(svcpt, force) && + !cfs_list_empty(&svcpt->scp_req_pending); +} + +/** + * 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 int +ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, int force) +{ + return ptlrpc_server_high_pending(svcpt, force) || + ptlrpc_server_normal_pending(svcpt, force); +} + +/** + * 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, int force) +{ + struct ptlrpc_request *req; + ENTRY; + + if (ptlrpc_server_high_pending(svcpt, force)) { + req = cfs_list_entry(svcpt->scp_hreq_pending.next, + struct ptlrpc_request, rq_list); + svcpt->scp_hreq_count++; + RETURN(req); + } + + if (ptlrpc_server_normal_pending(svcpt, force)) { + req = cfs_list_entry(svcpt->scp_req_pending.next, + struct ptlrpc_request, rq_list); + svcpt->scp_hreq_count = 0; + RETURN(req); + } + RETURN(NULL); +} + +/** + * 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_service *svc = svcpt->scp_service; + struct ptlrpc_request *req; + __u32 deadline; + int rc; + ENTRY; + + cfs_spin_lock(&svcpt->scp_lock); + if (cfs_list_empty(&svcpt->scp_req_incoming)) { + cfs_spin_unlock(&svcpt->scp_lock); + RETURN(0); + } + + req = cfs_list_entry(svcpt->scp_req_incoming.next, + struct ptlrpc_request, rq_list); + cfs_list_del_init(&req->rq_list); + svcpt->scp_nreqs_incoming--; + /* Consider this still a "queued" request as far as stats are + * concerned */ + cfs_spin_unlock(&svcpt->scp_lock); /* go through security check/transform */ rc = sptlrpc_svc_unwrap_request(req); @@ -944,15 +1565,18 @@ 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); @@ -963,6 +1587,13 @@ ptlrpc_server_handle_req_in(struct ptlrpc_service *svc) goto err_req; } + if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) && + lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) { + CERROR("drop incoming rpc opc %u, x"LPU64"\n", + cfs_fail_val, req->rq_xid); + goto err_req; + } + rc = -EINVAL; if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) { CERROR("wrong packet type received (type=%u) from %s\n", @@ -971,7 +1602,19 @@ ptlrpc_server_handle_req_in(struct ptlrpc_service *svc) goto err_req; } - CDEBUG(D_NET, "got req "LPD64"\n", req->rq_xid); + switch(lustre_msg_get_opc(req->rq_reqmsg)) { + case MDS_WRITEPAGE: + case OST_WRITE: + req->rq_bulk_write = 1; + break; + case MDS_READPAGE: + case OST_READ: + case MGS_CONFIG_READ: + req->rq_bulk_read = 1; + break; + } + + CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid); req->rq_export = class_conn2export( lustre_msg_get_handle(req->rq_reqmsg)); @@ -984,16 +1627,16 @@ ptlrpc_server_handle_req_in(struct ptlrpc_service *svc) "illegal security flavor,"); } - class_export_put(req->rq_export); - req->rq_export = NULL; if (rc) goto err_req; + ptlrpc_update_export_timer(req->rq_export, 0); } /* req_in handling should/must be fast */ if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5) - DEBUG_REQ(D_WARNING, req, "Slow req_in handling %lus", - cfs_time_current_sec() - req->rq_arrival_time.tv_sec); + DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s", + cfs_time_sub(cfs_time_current_sec(), + req->rq_arrival_time.tv_sec)); /* Set rpc server deadline and add it to the timed list */ deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) & @@ -1009,90 +1652,116 @@ ptlrpc_server_handle_req_in(struct ptlrpc_service *svc) ptlrpc_at_add_timed(req); /* Move it over to the request processing queue */ - spin_lock(&svc->srv_lock); - list_add_tail(&req->rq_list, &svc->srv_request_queue); - cfs_waitq_signal(&svc->srv_waitq); - spin_unlock(&svc->srv_lock); - RETURN(1); + rc = ptlrpc_server_request_add(svcpt, req); + if (rc) { + ptlrpc_hpreq_fini(req); + GOTO(err_req, rc); + } + cfs_waitq_signal(&svcpt->scp_waitq); + RETURN(1); err_req: - spin_lock(&svc->srv_lock); - svc->srv_n_queued_reqs--; - svc->srv_n_active_reqs++; - spin_unlock(&svc->srv_lock); - ptlrpc_server_free_request(req); + cfs_spin_lock(&svcpt->scp_req_lock); + svcpt->scp_nreqs_active++; + cfs_spin_unlock(&svcpt->scp_req_lock); + ptlrpc_server_finish_request(svcpt, req); - RETURN(1); + RETURN(1); } +/** + * 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) +ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt, + struct ptlrpc_thread *thread) { + struct ptlrpc_service *svc = svcpt->scp_service; struct obd_export *export = NULL; struct ptlrpc_request *request; struct timeval work_start; struct timeval work_end; long timediff; int rc; + int fail_opc = 0; ENTRY; - LASSERT(svc); - - spin_lock(&svc->srv_lock); - if (unlikely(list_empty (&svc->srv_request_queue) || - ( + cfs_spin_lock(&svcpt->scp_req_lock); #ifndef __KERNEL__ - /* !@%$# liblustre only has 1 thread */ - svc->srv_n_difficult_replies != 0 && + /* !@%$# liblustre only has 1 thread */ + if (cfs_atomic_read(&svcpt->scp_nreps_difficult) != 0) { + cfs_spin_unlock(&svcpt->scp_req_lock); + RETURN(0); + } #endif - 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); + request = ptlrpc_server_request_get(svcpt, 0); + if (request == NULL) { + cfs_spin_unlock(&svcpt->scp_req_lock); RETURN(0); } - request = list_entry (svc->srv_request_queue.next, - struct ptlrpc_request, rq_list); - list_del_init (&request->rq_list); - svc->srv_n_queued_reqs--; - svc->srv_n_active_reqs++; + 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) { + cfs_spin_unlock(&svcpt->scp_req_lock); - spin_unlock(&svc->srv_lock); + OBD_FAIL_TIMEOUT(fail_opc, 4); - do_gettimeofday(&work_start); + cfs_spin_lock(&svcpt->scp_req_lock); + request = ptlrpc_server_request_get(svcpt, 0); + if (request == NULL) { + cfs_spin_unlock(&svcpt->scp_req_lock); + RETURN(0); + } + } + } + + cfs_list_del_init(&request->rq_list); + svcpt->scp_nreqs_active++; + if (request->rq_hp) + svcpt->scp_nhreqs_active++; + + cfs_spin_unlock(&svcpt->scp_req_lock); + + ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET); + + if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG)) + libcfs_debug_dumplog(); + + cfs_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)); + svcpt->scp_nreqs_incoming); + lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR, + svcpt->scp_nreqs_active); + lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT, + at_get(&svcpt->scp_at_estimate)); } - rc = lu_context_init(&request->rq_session, LCT_SESSION); + 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; } request->rq_session.lc_thread = thread; + request->rq_session.lc_cookie = 0x5; lu_context_enter(&request->rq_session); - CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid); + 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; - request->rq_export = class_conn2export( - lustre_msg_get_handle(request->rq_reqmsg)); - if (likely(request->rq_export)) { if (unlikely(ptlrpc_check_req(request))) goto put_conn; @@ -1104,71 +1773,71 @@ ptlrpc_server_handle_request(struct ptlrpc_service *svc, The deadline is increased if we send an early reply. */ if (cfs_time_current_sec() > request->rq_deadline) { DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s" - ": deadline %ld%+lds ago\n", + ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n", libcfs_id2str(request->rq_peer), - request->rq_deadline - - request->rq_arrival_time.tv_sec, - cfs_time_current_sec() - request->rq_deadline); + cfs_time_sub(request->rq_deadline, + request->rq_arrival_time.tv_sec), + cfs_time_sub(cfs_time_current_sec(), + request->rq_deadline)); goto put_rpc_export; } - request->rq_phase = RQ_PHASE_INTERPRET; - 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), + 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)); - OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val); - - rc = svc->srv_handler(request); + if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING) + CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val); - request->rq_phase = RQ_PHASE_COMPLETE; + rc = 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: - if (likely(request->rq_export != NULL)) - class_export_put(request->rq_export); - 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 (%ld%+lds); client may timeout.", - request->rq_xid, request->rq_deadline - - request->rq_arrival_time.tv_sec, - cfs_time_current_sec() - request->rq_deadline); + "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); + cfs_gettimeofday(&work_end); timediff = cfs_timeval_sub(&work_end, &work_start, NULL); - CDEBUG(D_RPCTRACE, "request x"LPU64" opc %u from %s processed in " + 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", - 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); + cfs_curproc_comm(), + (request->rq_export ? + (char *)request->rq_export->exp_client_uuid.uuid : "0"), + (request->rq_export ? + cfs_atomic_read(&request->rq_export->exp_refcount) : -99), + lustre_msg_get_status(request->rq_reqmsg), + request->rq_xid, + libcfs_id2str(request->rq_peer), + lustre_msg_get_opc(request->rq_reqmsg), + 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); @@ -1181,58 +1850,72 @@ put_conn: } if (unlikely(request->rq_early_count)) { DEBUG_REQ(D_ADAPTTO, request, - "sent %d early replies before finishing in %lds", + "sent %d early replies before finishing in " + CFS_DURATION_T"s", request->rq_early_count, - work_end.tv_sec - request->rq_arrival_time.tv_sec); + cfs_time_sub(work_end.tv_sec, + request->rq_arrival_time.tv_sec)); } out_req: - ptlrpc_server_free_request(request); + ptlrpc_server_finish_request(svcpt, request); - RETURN(1); + RETURN(1); } +/** + * An internal function to process a single reply state object. + */ static int -ptlrpc_server_handle_reply (struct ptlrpc_service *svc) +ptlrpc_handle_rs(struct ptlrpc_reply_state *rs) { - struct ptlrpc_reply_state *rs; + 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; - spin_lock(&svc->srv_lock); - if (list_empty (&svc->srv_reply_queue)) { - spin_unlock(&svc->srv_lock); - RETURN(0); - } - - rs = list_entry (svc->srv_reply_queue.next, - struct ptlrpc_reply_state, rs_list); - exp = rs->rs_export; - obd = exp->exp_obd; LASSERT (rs->rs_difficult); LASSERT (rs->rs_scheduled); + LASSERT (cfs_list_empty(&rs->rs_list)); - list_del_init (&rs->rs_list); - - /* Disengage from notifiers carefully (lock order - irqrestore below!)*/ - spin_unlock(&svc->srv_lock); - - spin_lock (&obd->obd_uncommitted_replies_lock); - /* Noop if removed already */ - list_del_init (&rs->rs_obd_list); - spin_unlock (&obd->obd_uncommitted_replies_lock); - - spin_lock (&exp->exp_lock); + cfs_spin_lock (&exp->exp_lock); /* Noop if removed already */ - list_del_init (&rs->rs_exp_list); - spin_unlock (&exp->exp_lock); + cfs_list_del_init (&rs->rs_exp_list); + cfs_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) { + cfs_spin_lock(&exp->exp_uncommitted_replies_lock); + cfs_list_del_init(&rs->rs_obd_list); + cfs_spin_unlock(&exp->exp_uncommitted_replies_lock); + } - spin_lock(&svc->srv_lock); + cfs_spin_lock(&rs->rs_lock); been_handled = rs->rs_handled; rs->rs_handled = 1; @@ -1246,13 +1929,12 @@ ptlrpc_server_handle_reply (struct ptlrpc_service *svc) CWARN("All locks stolen from rs %p x"LPD64".t"LPD64 " o%d NID %s\n", rs, - rs->rs_xid, rs->rs_transno, - lustre_msg_get_opc(rs->rs_msg), + 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(&svc->srv_lock); + cfs_spin_unlock(&rs->rs_lock); if (!been_handled && rs->rs_on_net) { LNetMDUnlink(rs->rs_md_h); @@ -1264,90 +1946,110 @@ ptlrpc_server_handle_reply (struct ptlrpc_service *svc) ldlm_lock_decref(&rs->rs_locks[nlocks], rs->rs_modes[nlocks]); - spin_lock(&svc->srv_lock); + cfs_spin_lock(&rs->rs_lock); } rs->rs_scheduled = 0; if (!rs->rs_on_net) { /* Off the net */ - svc->srv_n_difficult_replies--; - spin_unlock(&svc->srv_lock); + cfs_spin_unlock(&rs->rs_lock); class_export_put (exp); rs->rs_export = NULL; ptlrpc_rs_decref (rs); - atomic_dec (&svc->srv_outstanding_replies); - RETURN(1); - } - - /* still on the net; callback will schedule */ - spin_unlock(&svc->srv_lock); - RETURN(1); + if (cfs_atomic_dec_and_test(&svcpt->scp_nreps_difficult) && + svc->srv_is_stopping) + cfs_waitq_broadcast(&svcpt->scp_waitq); + RETURN(1); + } + + /* still on the net; callback will schedule */ + cfs_spin_unlock(&rs->rs_lock); + RETURN(1); } #ifndef __KERNEL__ + +/** + * Check whether given service has a reply available for processing + * and process it. + * + * \param svc a ptlrpc service + * \retval 0 no replies processed + * \retval 1 one reply processed + */ +static int +ptlrpc_server_handle_reply(struct ptlrpc_service_part *svcpt) +{ + struct ptlrpc_reply_state *rs = NULL; + ENTRY; + + cfs_spin_lock(&svcpt->scp_rep_lock); + if (!cfs_list_empty(&svcpt->scp_rep_queue)) { + rs = cfs_list_entry(svcpt->scp_rep_queue.prev, + struct ptlrpc_reply_state, + rs_list); + cfs_list_del_init(&rs->rs_list); + } + cfs_spin_unlock(&svcpt->scp_rep_lock); + if (rs != NULL) + ptlrpc_handle_rs(rs); + RETURN(rs != NULL); +} + /* FIXME make use of timeout later */ int 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; + svcpt = svc->srv_part; - /* service threads can block for bulk, so this limits us - * (arbitrarily) to recursing 1 stack frame per service. - * Note that the problem with recursion is that we have to - * unwind completely before our caller can resume. */ + if (svcpt->scp_nthrs_running != 0) /* I've recursed */ + continue; - svc->srv_threads_running++; + /* service threads can block for bulk, so this limits us + * (arbitrarily) to recursing 1 stack frame per service. + * Note that the problem with recursion is that we have to + * unwind completely before our caller can resume. */ - do { - rc = ptlrpc_server_handle_req_in(svc); - rc |= ptlrpc_server_handle_reply(svc); - rc |= ptlrpc_at_check_timed(svc); - rc |= ptlrpc_server_handle_request(svc, NULL); - rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0); - did_something |= rc; - } while (rc); + svcpt->scp_nthrs_running++; - svc->srv_threads_running--; - } + do { + rc = ptlrpc_server_handle_req_in(svcpt); + rc |= ptlrpc_server_handle_reply(svcpt); + rc |= ptlrpc_at_check_timed(svcpt); + rc |= ptlrpc_server_handle_request(svcpt, NULL); + rc |= (ptlrpc_server_post_idle_rqbds(svcpt) > 0); + did_something |= rc; + } while (rc); + + svcpt->scp_nthrs_running--; + } - RETURN(did_something); + RETURN(did_something); } #define ptlrpc_stop_all_threads(s) do {} while (0) #else /* __KERNEL__ */ -/* Don't use daemonize, it removes fs struct from new thread (bug 418) */ -void ptlrpc_daemonize(char *name) -{ - struct fs_struct *fs = current->fs; - - atomic_inc(&fs->count); - cfs_daemonize(name); - exit_fs(cfs_current()); - current->fs = fs; - ll_set_fs_pwd(current->fs, init_task.fs->pwdmnt, init_task.fs->pwd); -} - 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. */ @@ -1357,38 +2059,132 @@ ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc) * space. */ if (avail <= low_water) - ptlrpc_grow_req_bufs(svc); + ptlrpc_grow_req_bufs(svcpt); - if (svc->srv_stats) - lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR, - avail); + if (svcpt->scp_service->srv_stats) { + lprocfs_counter_add(svcpt->scp_service->srv_stats, + PTLRPC_REQBUF_AVAIL_CNTR, avail); + } } static int ptlrpc_retry_rqbds(void *arg) { - struct ptlrpc_service *svc = (struct ptlrpc_service *)arg; + struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg; + + svcpt->scp_rqbd_timeout = 0; + return -ETIMEDOUT; +} + +static inline int +ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt) +{ + return svcpt->scp_nreqs_active < + svcpt->scp_nthrs_running - 1 - + (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL); +} + +/** + * allowed to create more threads + * user can call it w/o any lock but need to hold + * ptlrpc_service_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_threads_max; +} + +/** + * 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); +} + +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); - svc->srv_rqbd_timeout = 0; - return (-ETIMEDOUT); + cfs_cond_resched(); + + l_wait_event_exclusive_head(svcpt->scp_waitq, + ptlrpc_thread_stopping(thread) || + ptlrpc_server_request_incoming(svcpt) || + ptlrpc_server_request_pending(svcpt, 0) || + ptlrpc_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_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; - struct lc_watchdog *watchdog; + 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 - struct group_info *ginfo = NULL; + cfs_group_info_t *ginfo = NULL; #endif - struct lu_env env; + struct lu_env *env; int counter = 0, rc = 0; ENTRY; - ptlrpc_daemonize(data->name); + thread->t_pid = cfs_curproc_pid(); + cfs_daemonize_ctxt(thread->t_name); #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA) /* we need to do this before any per-thread allocation is done so that @@ -1396,203 +2192,369 @@ static int ptlrpc_main(void *arg) if (svc->srv_cpu_affinity) { int cpu, num_cpu; - for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) { + for (cpu = 0, num_cpu = 0; cpu < cfs_num_possible_cpus(); + cpu++) { if (!cpu_online(cpu)) continue; - if (num_cpu == thread->t_id % num_online_cpus()) + if (num_cpu == thread->t_id % cfs_num_online_cpus()) break; num_cpu++; } - set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu))); + cfs_set_cpus_allowed(cfs_current(), + node_to_cpumask(cpu_to_node(cpu))); } #endif #ifdef WITH_GROUP_INFO - ginfo = groups_alloc(0); + ginfo = cfs_groups_alloc(0); if (!ginfo) { rc = -ENOMEM; goto out; } - set_current_groups(ginfo); - put_group_info(ginfo); + cfs_set_current_groups(ginfo); + cfs_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, svc->srv_ctx_tags); + 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; + thread->t_env = env; + env->le_ctx.lc_thread = thread; + env->le_ctx.lc_cookie = 0x6; /* 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); - - watchdog = lc_watchdog_add(max_t(int, obd_timeout, AT_OFF ? 0 : - at_get(&svc->srv_at_estimate)) * - svc->srv_watchdog_factor, NULL, NULL); - - 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); - - CDEBUG(D_NET, "service thread %d (#%d)started\n", thread->t_id, - svc->srv_threads_running); - - /* XXX maintain a list of all managed devices: insert here */ - - while ((thread->t_flags & SVC_STOPPING) == 0 || - svc->srv_n_difficult_replies != 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); - - lc_watchdog_disable(watchdog); - - cond_resched(); - - l_wait_event_exclusive (svc->srv_waitq, - ((thread->t_flags & SVC_STOPPING) != 0 && - svc->srv_n_difficult_replies == 0) || - (!list_empty(&svc->srv_idle_rqbds) && - svc->srv_rqbd_timeout == 0) || - !list_empty(&svc->srv_req_in_queue) || - !list_empty(&svc->srv_reply_queue) || - (!list_empty(&svc->srv_request_queue) && - (svc->srv_n_active_reqs < - (svc->srv_threads_running - 1))) || - svc->srv_at_check, - &lwi); - - lc_watchdog_touch_ms(watchdog, max_t(int, obd_timeout, - AT_OFF ? 0 : - at_get(&svc->srv_at_estimate)) * - svc->srv_watchdog_factor); - - ptlrpc_check_rqbd_pool(svc); - - 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); - } + cfs_spin_lock(&svcpt->scp_lock); - if (!list_empty(&svc->srv_reply_queue)) - ptlrpc_server_handle_reply(svc); + LASSERT(thread_is_starting(thread)); + thread_clear_flags(thread, SVC_STARTING); - 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; - } + LASSERT(svcpt->scp_nthrs_starting == 1); + svcpt->scp_nthrs_starting--; + + /* SVC_STOPPING may already be set here if someone else is trying + * to stop the service while this new thread has been dynamically + * forked. We still set SVC_RUNNING to let our creator know that + * we are now running, however we will exit as soon as possible */ + thread_add_flags(thread, SVC_RUNNING); + svcpt->scp_nthrs_running++; + cfs_spin_unlock(&svcpt->scp_lock); + + /* wake up our creator in case he's still waiting. */ + cfs_waitq_signal(&thread->t_ctl_waitq); + + thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt), + NULL, NULL); + + cfs_spin_lock(&svcpt->scp_rep_lock); + cfs_list_add(&rs->rs_list, &svcpt->scp_rep_idle); + cfs_waitq_signal(&svcpt->scp_rep_waitq); + cfs_spin_unlock(&svcpt->scp_rep_lock); + + CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id, + svcpt->scp_nthrs_running); + + /* XXX maintain a list of all managed devices: insert here */ + while (!ptlrpc_thread_stopping(thread)) { + if (ptlrpc_wait_event(svcpt, thread)) + break; - if (svc->srv_at_check) - ptlrpc_at_check_timed(svc); + ptlrpc_check_rqbd_pool(svcpt); - /* don't handle requests in the last thread */ - if (!list_empty (&svc->srv_request_queue) && - (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 (ptlrpc_threads_need_create(svcpt)) { + /* Ignore return code - we tried... */ + ptlrpc_start_thread(svcpt, 0); } - 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); + /* Process all incoming reqs before handling any */ + if (ptlrpc_server_request_incoming(svcpt)) { + ptlrpc_server_handle_req_in(svcpt); + /* but limit ourselves in case of flood */ + if (counter++ < 100) + continue; + counter = 0; + } + + if (ptlrpc_at_check(svcpt)) + ptlrpc_at_check_timed(svcpt); + + if (ptlrpc_server_request_pending(svcpt, 0)) { + lu_context_enter(&env->le_ctx); + ptlrpc_server_handle_request(svcpt, thread); + lu_context_exit(&env->le_ctx); + } + + if (ptlrpc_rqbd_pending(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); } } - lc_watchdog_delete(watchdog); + lc_watchdog_delete(thread->t_watchdog); + thread->t_watchdog = NULL; 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); + + cfs_spin_lock(&svcpt->scp_lock); + if (thread_test_and_clear_flags(thread, SVC_STARTING)) + svcpt->scp_nthrs_starting--; + + if (thread_test_and_clear_flags(thread, SVC_RUNNING)) { + /* must know immediately */ + svcpt->scp_nthrs_running--; + } + + thread->t_id = rc; + thread_add_flags(thread, SVC_STOPPED); + + cfs_waitq_signal(&thread->t_ctl_waitq); + cfs_spin_unlock(&svcpt->scp_lock); + + return rc; +} + +struct ptlrpc_hr_args { + int thread_index; + int cpu_index; + struct ptlrpc_hr_service *hrs; +}; + +static int hrt_dont_sleep(struct ptlrpc_hr_thread *t, + cfs_list_t *replies) +{ + int result; + + cfs_spin_lock(&t->hrt_lock); + cfs_list_splice_init(&t->hrt_queue, replies); + result = cfs_test_bit(HRT_STOPPING, &t->hrt_flags) || + !cfs_list_empty(replies); + cfs_spin_unlock(&t->hrt_lock); + return result; +} + +/** + * 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); + + snprintf(threadname, sizeof(threadname), + "ptlrpc_hr_%d", hr_args->thread_index); + + cfs_daemonize_ctxt(threadname); +#if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK) + cfs_set_cpus_allowed(cfs_current(), + node_to_cpumask(cpu_to_node(hr_args->cpu_index))); +#endif + cfs_set_bit(HRT_RUNNING, &t->hrt_flags); + cfs_waitq_signal(&t->hrt_wait); + + while (!cfs_test_bit(HRT_STOPPING, &t->hrt_flags)) { + + l_wait_condition(t->hrt_wait, hrt_dont_sleep(t, &replies)); + while (!cfs_list_empty(&replies)) { + struct ptlrpc_reply_state *rs; + + rs = cfs_list_entry(replies.prev, + struct ptlrpc_reply_state, + rs_list); + cfs_list_del_init(&rs->rs_list); + ptlrpc_handle_rs(rs); + } + } + + cfs_clear_bit(HRT_RUNNING, &t->hrt_flags); + cfs_complete(&t->hrt_completion); + + return 0; +} - spin_lock(&svc->srv_lock); - svc->srv_threads_running--; /* must know immediately */ - thread->t_id = rc; - thread->t_flags = SVC_STOPPED; +static int ptlrpc_start_hr_thread(struct ptlrpc_hr_service *hr, int n, int cpu) +{ + struct ptlrpc_hr_thread *t = &hr->hr_threads[n]; + struct ptlrpc_hr_args args; + int rc; + ENTRY; - cfs_waitq_signal(&thread->t_ctl_waitq); - spin_unlock(&svc->srv_lock); + args.thread_index = n; + args.cpu_index = cpu; + args.hrs = hr; + rc = cfs_create_thread(ptlrpc_hr_main, (void*)&args, CFS_DAEMON_FLAGS); + if (rc < 0) { + cfs_complete(&t->hrt_completion); + GOTO(out, rc); + } + l_wait_condition(t->hrt_wait, cfs_test_bit(HRT_RUNNING, &t->hrt_flags)); + RETURN(0); + out: return rc; } -static void ptlrpc_stop_thread(struct ptlrpc_service *svc, - struct ptlrpc_thread *thread) +static void ptlrpc_stop_hr_thread(struct ptlrpc_hr_thread *t) { - struct l_wait_info lwi = { 0 }; + ENTRY; - spin_lock(&svc->srv_lock); - thread->t_flags = SVC_STOPPING; - spin_unlock(&svc->srv_lock); + cfs_set_bit(HRT_STOPPING, &t->hrt_flags); + cfs_waitq_signal(&t->hrt_wait); + cfs_wait_for_completion(&t->hrt_completion); - cfs_waitq_broadcast(&svc->srv_waitq); - l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED), - &lwi); + EXIT; +} - spin_lock(&svc->srv_lock); - list_del(&thread->t_link); - spin_unlock(&svc->srv_lock); +static void ptlrpc_stop_hr_threads(struct ptlrpc_hr_service *hrs) +{ + int n; + ENTRY; - OBD_FREE_PTR(thread); + for (n = 0; n < hrs->hr_n_threads; n++) + ptlrpc_stop_hr_thread(&hrs->hr_threads[n]); + + EXIT; } -void ptlrpc_stop_all_threads(struct ptlrpc_service *svc) +static int ptlrpc_start_hr_threads(struct ptlrpc_hr_service *hr) { - struct ptlrpc_thread *thread; + int rc = -ENOMEM; + int n, cpu, threads_started = 0; + ENTRY; - spin_lock(&svc->srv_lock); - while (!list_empty(&svc->srv_threads)) { - thread = list_entry(svc->srv_threads.next, - struct ptlrpc_thread, t_link); + LASSERT(hr != NULL); + LASSERT(hr->hr_n_threads > 0); - spin_unlock(&svc->srv_lock); - ptlrpc_stop_thread(svc, thread); - spin_lock(&svc->srv_lock); + for (n = 0, cpu = 0; n < hr->hr_n_threads; n++) { +#if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK) + while (!cpu_online(cpu)) { + cpu++; + if (cpu >= cfs_num_possible_cpus()) + cpu = 0; + } +#endif + rc = ptlrpc_start_hr_thread(hr, n, cpu); + if (rc != 0) + break; + threads_started++; + cpu++; + } + if (threads_started == 0) { + CERROR("No reply handling threads started\n"); + RETURN(-ESRCH); + } + if (threads_started < hr->hr_n_threads) { + CWARN("Started only %d reply handling threads from %d\n", + threads_started, hr->hr_n_threads); + hr->hr_n_threads = threads_started; } + RETURN(0); +} - spin_unlock(&svc->srv_lock); +static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt) +{ + 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); + + cfs_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); + } + + cfs_waitq_broadcast(&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; + } + cfs_spin_unlock(&svcpt->scp_lock); + + CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n", + svcpt->scp_service->srv_thread_name, thread->t_id); + l_wait_event(thread->t_ctl_waitq, + thread_is_stopped(thread), &lwi); + + cfs_spin_lock(&svcpt->scp_lock); + } + + cfs_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; } -int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc) +/** + * Stops all threads of a particular service \a svc + */ +void ptlrpc_stop_all_threads(struct ptlrpc_service *svc) +{ + ENTRY; + + if (svc != NULL && svc->srv_part != NULL) + ptlrpc_svcpt_stop_threads(svc->srv_part); + EXIT; +} + +int ptlrpc_start_threads(struct ptlrpc_service *svc) { int i, rc = 0; ENTRY; @@ -1601,256 +2563,431 @@ int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc) ptlrpc_server_handle_request */ LASSERT(svc->srv_threads_min >= 2); for (i = 0; i < svc->srv_threads_min; i++) { - rc = ptlrpc_start_thread(dev, svc); + rc = ptlrpc_start_thread(svc->srv_part, 1); /* We have enough threads, don't start more. b=15759 */ - if (rc == -EMFILE) + if (rc == -EMFILE) { + rc = 0; break; + } if (rc) { - CERROR("cannot start %s thread #%d: rc %d\n", + CERROR("cannot start %s thread #%d: rc %d\n", svc->srv_thread_name, i, rc); ptlrpc_stop_all_threads(svc); + break; } } RETURN(rc); } -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; + struct l_wait_info lwi = { 0 }; + struct ptlrpc_thread *thread; + struct ptlrpc_service *svc = svcpt->scp_service; + int rc; + ENTRY; + + LASSERT(svcpt != NULL); + + CDEBUG(D_RPCTRACE, "%s started %d min %d max %d\n", + svc->srv_name, svcpt->scp_nthrs_running, + svc->srv_threads_min, svc->srv_threads_max); + + 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_threads_min - 1)) + RETURN(-EMFILE); + + OBD_ALLOC_PTR(thread); + if (thread == NULL) + RETURN(-ENOMEM); + cfs_waitq_init(&thread->t_ctl_waitq); + + cfs_spin_lock(&svcpt->scp_lock); + if (!ptlrpc_threads_increasable(svcpt)) { + cfs_spin_unlock(&svcpt->scp_lock); + OBD_FREE_PTR(thread); + RETURN(-EMFILE); + } + + 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); + cfs_spin_unlock(&svcpt->scp_lock); + OBD_FREE_PTR(thread); + if (wait) { + CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n", + svc->srv_thread_name, svcpt->scp_thr_nextid); + cfs_schedule(); + 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); + cfs_spin_unlock(&svcpt->scp_lock); + + snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, + "%s_%02d", svc->srv_thread_name, thread->t_id); + + CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name); + /* + * CLONE_VM and CLONE_FILES just avoid a needless copy, because we + * just drop the VM and FILES in cfs_daemonize_ctxt() right away. + */ + rc = cfs_create_thread(ptlrpc_main, thread, CFS_DAEMON_FLAGS); + if (rc < 0) { + CERROR("cannot start thread '%s': rc %d\n", + thread->t_name, rc); + cfs_spin_lock(&svcpt->scp_lock); + cfs_list_del(&thread->t_link); + --svcpt->scp_nthrs_starting; + cfs_spin_unlock(&svcpt->scp_lock); + + OBD_FREE(thread, sizeof(*thread)); + RETURN(rc); + } + + if (!wait) + RETURN(0); + + l_wait_event(thread->t_ctl_waitq, + thread_is_running(thread) || thread_is_stopped(thread), + &lwi); + + rc = thread_is_stopped(thread) ? thread->t_id : 0; + RETURN(rc); +} + +int ptlrpc_hr_init(void) +{ + int i; + int n_cpus = cfs_num_online_cpus(); + struct ptlrpc_hr_service *hr; + int size; + int 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); + LASSERT(ptlrpc_hr == NULL); - OBD_ALLOC_PTR(thread); - if (thread == NULL) + size = offsetof(struct ptlrpc_hr_service, hr_threads[n_cpus]); + OBD_ALLOC(hr, size); + if (hr == 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); - } - list_add(&thread->t_link, &svc->srv_threads); - id = svc->srv_threads_started++; - spin_unlock(&svc->srv_lock); - - 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 ptlrpc_daemonize() 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); + for (i = 0; i < n_cpus; i++) { + struct ptlrpc_hr_thread *t = &hr->hr_threads[i]; - spin_lock(&svc->srv_lock); - list_del(&thread->t_link); - --svc->srv_threads_started; - spin_unlock(&svc->srv_lock); - - OBD_FREE(thread, sizeof(*thread)); - RETURN(rc); + cfs_spin_lock_init(&t->hrt_lock); + cfs_waitq_init(&t->hrt_wait); + CFS_INIT_LIST_HEAD(&t->hrt_queue); + cfs_init_completion(&t->hrt_completion); } - l_wait_event(thread->t_ctl_waitq, - thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi); + hr->hr_n_threads = n_cpus; + hr->hr_size = size; + ptlrpc_hr = hr; - rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0; + rc = ptlrpc_start_hr_threads(hr); + if (rc) { + OBD_FREE(hr, hr->hr_size); + ptlrpc_hr = NULL; + } RETURN(rc); } -#endif -int ptlrpc_unregister_service(struct ptlrpc_service *service) +void ptlrpc_hr_fini(void) { - int rc; - struct l_wait_info lwi; - struct list_head *tmp; - struct ptlrpc_reply_state *rs, *t; + if (ptlrpc_hr != NULL) { + ptlrpc_stop_hr_threads(ptlrpc_hr); + OBD_FREE(ptlrpc_hr, ptlrpc_hr->hr_size); + ptlrpc_hr = NULL; + } +} - cfs_timer_disarm(&service->srv_at_timer); +#endif /* __KERNEL__ */ - ptlrpc_stop_all_threads(service); - LASSERT(list_empty(&service->srv_threads)); +/** + * Wait until all already scheduled replies are processed. + */ +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); + } +} - spin_lock (&ptlrpc_all_services_lock); - list_del_init (&service->srv_list); - spin_unlock (&ptlrpc_all_services_lock); +static void +ptlrpc_service_del_atimer(struct ptlrpc_service *svc) +{ + struct ptlrpc_service_part *svcpt; - ptlrpc_lprocfs_unregister_service(service); + /* early disarm AT timer... */ + do { /* iterrate over multiple partitions in the future */ + svcpt = svc->srv_part; + if (svcpt == NULL || svcpt->scp_service == NULL) + break; - /* All history will be culled when the next request buffer is - * freed */ - service->srv_max_history_rqbds = 0; + cfs_timer_disarm(&svcpt->scp_at_timer); + } while (0); +} + +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; - CDEBUG(D_NET, "%s: tearing down\n", service->srv_name); + /* All history will be culled when the next request buffer is + * freed in ptlrpc_service_purge_all() */ + svc->srv_max_history_rqbds = 0; + + rc = LNetClearLazyPortal(svc->srv_req_portal); + LASSERT(rc == 0); + + do { /* iterrate over multiple partitions in the future */ + svcpt = svc->srv_part; + if (svcpt == NULL || svcpt->scp_service == NULL) + 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); + } + } while (0); + + do { /* iterrate over multiple partitions in the future */ + svcpt = svc->srv_part; + if (svcpt == NULL || svcpt->scp_service == NULL) + break; + + /* Wait for the network to release any buffers + * it's currently filling */ + cfs_spin_lock(&svcpt->scp_lock); + while (svcpt->scp_nrqbds_posted != 0) { + cfs_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); + } + cfs_spin_lock(&svcpt->scp_lock); + } + cfs_spin_unlock(&svcpt->scp_lock); + } while (0); +} - rc = LNetClearLazyPortal(service->srv_req_portal); - LASSERT (rc == 0); +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; + + do { /* iterrate over multiple partitions in the future */ + /* schedule all outstanding replies to terminate them */ + svcpt = svc->srv_part; + if (svcpt == NULL || svcpt->scp_service == NULL) + break; + + cfs_spin_lock(&svcpt->scp_rep_lock); + while (!cfs_list_empty(&svcpt->scp_rep_active)) { + rs = cfs_list_entry(svcpt->scp_rep_active.next, + struct ptlrpc_reply_state, rs_list); + cfs_spin_lock(&rs->rs_lock); + ptlrpc_schedule_difficult_reply(rs); + cfs_spin_unlock(&rs->rs_lock); + } + cfs_spin_unlock(&svcpt->scp_rep_lock); + + /* purge the request queue. NB No new replies (rqbds + * all unlinked) and no service threads, so I'm the only + * thread noodling the request queue now */ + while (!cfs_list_empty(&svcpt->scp_req_incoming)) { + req = cfs_list_entry(svcpt->scp_req_incoming.next, + struct ptlrpc_request, rq_list); + + cfs_list_del(&req->rq_list); + svcpt->scp_nreqs_incoming--; + svcpt->scp_nreqs_active++; + ptlrpc_server_finish_request(svcpt, req); + } + + while (ptlrpc_server_request_pending(svcpt, 1)) { + req = ptlrpc_server_request_get(svcpt, 1); + cfs_list_del(&req->rq_list); + svcpt->scp_nreqs_active++; + ptlrpc_hpreq_fini(req); + ptlrpc_server_finish_request(svcpt, req); + } + + 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); + } + } while (0); +} - /* 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_free(struct ptlrpc_service *svc) +{ + struct ptlrpc_service_part *svcpt; + struct ptlrpc_at_array *array; + + do { /* iterrate over multiple partitions in the future */ + svcpt = svc->srv_part; + if (svcpt == NULL || 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; + } + svcpt->scp_service = NULL; + } while (0); + + do { /* iterrate over multiple partitions in the future */ + svcpt = svc->srv_part; + if (svcpt != NULL) + OBD_FREE_PTR(svcpt); + } while (0); + + OBD_FREE_PTR(svc); +} - rc = LNetMDUnlink(rqbd->rqbd_md_h); - LASSERT (rc == 0 || rc == -ENOENT); - } +int ptlrpc_unregister_service(struct ptlrpc_service *service) +{ + ENTRY; - /* 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); + CDEBUG(D_NET, "%s: tearing down\n", service->srv_name); - if (rc == 0) - break; + service->srv_is_stopping = 1; - /* Network access will complete in finite time but the HUGE - * timeout lets us CWARN for visibility of sluggish NALs */ - lwi = LWI_TIMEOUT(cfs_time_seconds(LONG_UNLINK), NULL, NULL); - rc = l_wait_event(service->srv_waitq, - service->srv_nrqbd_receiving == 0, - &lwi); - if (rc == -ETIMEDOUT) - CWARN("Service %s waiting for request buffers\n", - service->srv_name); - } - - /* schedule all outstanding replies to terminate them */ - 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); - ptlrpc_schedule_difficult_reply(rs); - } - 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_free_request(req); - } - while (!list_empty(&service->srv_request_queue)) { - struct ptlrpc_request *req = - list_entry(service->srv_request_queue.next, - struct ptlrpc_request, - rq_list); - - list_del(&req->rq_list); - service->srv_n_queued_reqs--; - service->srv_n_active_reqs++; - - ptlrpc_server_free_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); - } - - /* wait for all outstanding replies to complete (they were - * scheduled having been flagged to abort above) */ - while (atomic_read(&service->srv_outstanding_replies) != 0) { - struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL); - - rc = l_wait_event(service->srv_waitq, - !list_empty(&service->srv_reply_queue), &lwi); - LASSERT(rc == 0 || rc == -ETIMEDOUT); + cfs_spin_lock(&ptlrpc_all_services_lock); + cfs_list_del_init(&service->srv_list); + cfs_spin_unlock(&ptlrpc_all_services_lock); - if (rc == 0) { - ptlrpc_server_handle_reply(service); - continue; - } - CWARN("Unexpectedly long timeout %p\n", service); - } + ptlrpc_lprocfs_unregister_service(service); - 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_free(service); - OBD_FREE_PTR(service); - return 0; + RETURN(0); } -/* 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) { - struct ptlrpc_request *request; - struct timeval right_now; - long timediff; - - if (svc == NULL) - return 0; - - do_gettimeofday(&right_now); - - spin_lock(&svc->srv_lock); - if (list_empty(&svc->srv_request_queue)) { - spin_unlock(&svc->srv_lock); - return 0; - } - - /* How long has the next entry been waiting? */ - 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); - } - - return 0; + struct ptlrpc_service_part *svcpt; + struct ptlrpc_request *request; + struct timeval right_now; + long timediff; + + if (svc == NULL || svc->srv_part == NULL) + return 0; + + cfs_gettimeofday(&right_now); + + svcpt = svc->srv_part; + cfs_spin_lock(&svcpt->scp_req_lock); + if (!ptlrpc_server_request_pending(svcpt, 1)) { + cfs_spin_unlock(&svcpt->scp_req_lock); + return 0; + } + + /* How long has the next entry been waiting? */ + if (cfs_list_empty(&svcpt->scp_req_pending)) { + request = cfs_list_entry(svcpt->scp_hreq_pending.next, + struct ptlrpc_request, rq_list); + } else { + request = cfs_list_entry(svcpt->scp_req_pending.next, + struct ptlrpc_request, rq_list); + } + + timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL); + cfs_spin_unlock(&svcpt->scp_req_lock); + + 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; }